Measure Theory 3: Measure Algebras - D.h. Fremlin 1t626i

[email protected]

0. Thus (A, µ ¯) is a measure algebra. 321I Definition For any measure space (X, Σ, µ) I will call (A, µ ¯), as constructed above, the measure algebra of (X, Σ, µ). 321J The Stone representation of a measure algebra Just as with Dedekind σ-complete Boolean algebras (314N), every measure algebra is obtainable from the construction above. Theorem Let (A, µ ¯) be any measure algebra. Then it is isomorphic, as measure algebra, to the measure algebra of some measure space. proof (a) We know from 314M that A is isomorphic, as Boolean algebra, to a quotient algebra Σ/M where Σ is a σ-algebra of subsets of the Stone space Z of A, and M is the ideal of meager subsets of Z. Let π : Σ/M → A be the canonical isomorphism, and set θE = πE • for each E ∈ Σ; then θ : Σ → A is a sequentially order-continuous surjective Boolean homomorphism with kernel M. (b) For E ∈ Σ, set νE = µ ¯(θE). Then (Z, Σ, ν) is a measure space. P P (i) We know already that Σ is a σ-algebra of subsets of Z. (ii) ν∅ = µ ¯(θ∅) = µ ¯0 = 0. (iii) If hEn in∈N is a dist sequence in Σ, then (because θ is a Boolean S homomorphism) hθEn in∈N is a dist sequence in A and (because θ is sequentially order-continuous) θ( n∈N En ) = supn∈N θEn ; so

321 Notes

Measure algebras

71

S P∞ P∞ ν( n∈N En ) = µ ¯(supn∈N θEn ) = n=0 µ ¯(θEn ) = n=0 νEn . Q Q (c) For E ∈ Σ, νE = 0 ⇐⇒ µ ¯(θE) = 0 ⇐⇒ θE = 0 ⇐⇒ E ∈ M. So the measure algebra of (Z, Σ, ν) is just Σ/M, with ν¯E • = νE = µ ¯(θE) = µ ¯(πE • ) for every E ∈ Σ. Thus the Boolean algebra isomorphism π is also an isomorphism between the measure algebras (Σ/M, ν¯) and (A, µ ¯), and (A, µ ¯) is represented in the required form. 321K Definition I will call the measure space (Z, Σ, ν) constructed in the proof of 321J the Stone space of the measure algebra (A, µ ¯). For later reference, I repeat the description of this space as developed in 311E, 311I, 314M and 321J. Z is a compact Hausdorff space, being the Stone space of A. A can be identified with the algebra of open-andclosed sets in Z. The ideal of ν-negligible sets coincides with the ideal of meager subsets of Z; in particular, ν is complete. The measurable sets are precisely those expressible in the form E = b a4M where a ∈ A, b a ⊆ Z is the corresponding open-and-closed set, and M is meager; in this case νE = µ ¯a and a = θE is the member of A corresponding to E. For the most important classes of measure algebras, more can be said; see 322M et seq. below. 321X Basic exercises (a) Let (A, µ ¯) be a measure algebra, and a ∈ A. Show that (Aa , µ ¯¹ Aa ) is a measure algebra, writing Aa for the principal ideal of A generated by a. (b) Let (X, Σ, µ ¯) be a measure space, and A its measure algebra. (i) Show that if T is a σ-subalgebra of Σ, then {E • : E ∈ T} is a σ-subalgebra of A. (ii) Show that if B is a σ-subalgebra of A then {E : E ∈ Σ, E • ∈ B} is a σ-subalgebra of Σ. 321Y Further exercises (a) Let (A, µ ¯) be a measure algebra, and I C A a σ-ideal. For u ∈ A/I set µ ¯u = inf{¯ µa : a ∈ A, a• = u}. Show that (A, µ ¯) is a measure algebra. 321 Notes and comments The idea behind taking the quotient Σ/N , where Σ is the algebra of measurable sets and N is the ideal of negligible sets, is just that if negligible sets can be ignored – as is the case for a very large proportion of the results of measure theory – then two measurable sets can be counted as virtually the same if they differ by a negligible set, that is, if they represent the same member of the measure algebra. The definition in 321A is designed to be an exact characterization of these quotient algebras, taking into the measures with which they are endowed. In the course of the present chapter I will work through many of the basic ideas dealt with in Volumes 1 and 2 to show how they can be translated into theorems about measure algebras, as I have done in 321B-321F. It is worth checking these correspondences carefully, because some of the ideas mutate significantly in translation. In measure algebras, it becomes sensible to take seriously the suprema and infima of uncountable sets (see 321C-321F). I should perhaps remark that while the Stone representation (321J-321K) is significant, it is not the most important method of representing measure algebras, which is surely Maharam’s theorem, to be dealt with in the next chapter. Nevertheless, the Stone representation is a canonical one, and will appear at each point that we meet a new construction involving measure algebras, just as the ordinary Stone representation of Boolean algebras can be expected to throw light on any aspect of Boolean algebra.

72

Measure algebras

§322 intro.

322 Taxonomy of measure algebras Before going farther with the general theory of measure algebras, I run through those parts of the classification of measure spaces in §211 which have expressions in of measure algebras. The most important concepts at this stage are those of ‘semi-finite’, ‘localizable’ and ‘σ-finite’ measure algebra (322Ac322Ae); these correspond exactly to the same applied to measure spaces (322B). I briefly investigate the Boolean-algebra properties of semi-finite and σ-finite measure algebras (322F, 322G), with mentions of completions and c.l.d. versions (322D), subspace measures (322I-322J), direct sums of measure spaces (322K, 322L) and subalgebras of measure algebras (322M). It turns out that localizability of a measure algebra is connected in striking ways to the properties of the canonical measure on its Stone space (322N). I end the section with a description of the ‘localization’ of a semi-finite measure algebra (322O-322P) and with some further properties of Stone spaces (322Q). 322A Definitions Let (A, µ ¯) be a measure algebra. (a) I will say that (A, µ ¯) is a probability algebra if µ ¯1 = 1. (b) (A, µ ¯) is totally finite if µ ¯1 < ∞. ¯an < ∞ for every n ∈ N and (c) (A, µ ¯) is σ-finite if there is a sequence han in∈N in A such that µ supn∈N an = 1. Note that in this case han in∈N can be taken either to be non-decreasing (consider a0n = supi
The first step is to relate these concepts to the corresponding ones for measure spaces.

Theorem Let (X, Σ, µ) be a measure space, and (A, µ ¯) its measure algebra. Then (a) (X, Σ, µ) is a probability space iff (A, µ ¯) is a probability algebra; (b) (X, Σ, µ) is totally finite iff (A, µ ¯) is; (c) (X, Σ, µ) is σ-finite iff (A, µ ¯) is; (d) (X, Σ, µ) is semi-finite iff (A, µ ¯) is; (e) (X, Σ, µ) is localizable iff (A, µ ¯) is; (f) if E ∈ Σ, then E is an atom for µ iff E • is an atom in A; (g) (X, Σ, µ) is atomless iff A is; (h) (X, Σ, µ) is purely atomic iff A is. proof (a), (b) are trivial, since µ ¯1 = µX. (c)(i) If µ is σ-finite, let hEn in∈N be a sequence of sets of finite measure covering X; then µ ¯En• < ∞ for every n, and S supn∈N En• = ( n∈N En )• = 1, so (A, µ ¯) is σ-finite. (ii) If (A, µ ¯) is σ-finite, let han in∈N be a sequence in A such ¯an < ∞ for every n and supn∈N an = 1. Sthat µ For each n, choose En ∈ Σ such that En• = an . Set E = n∈N En ; then E • = supn∈N an = 1, so E is conegligible. Now (X \ E, E0 , E1 , . . . ) is a sequence of sets of finite measure covering X, so µ is σ-finite. (d)(i) Suppose that µ is semi-finite and that a ∈ A, µ ¯a = ∞. Then there is an E ∈ Σ such that E • = a, so that µE = µ ¯a = ∞. As µ is semi-finite, there is an F ∈ Σ such that F ⊆ E and 0 < µF < ∞. Set b = F • ; then b ⊆ a and 0 < µ ¯b < ∞. (ii) Suppose that (A, µ ¯) is semi-finite and that E ∈ Σ, µE = ∞. Then µ ¯E • = ∞, so there is a • • b ⊆ E such that 0 < µ ¯b < ∞. Let F ∈ Σ be such that F = b. Then F ∩ E ∈ Σ, F ∩ E ⊆ E and (F ∩ E)• = E • ∩ b = b, so that µ(F ∩ E) = µ ¯b ∈ ]0, ∞[.

322D

Taxonomy of measure algebras

73

(e)(i) Note first that if E ⊆ Σ and F ∈ Σ, then E \ F is negligible for every E ∈ E ⇐⇒ E • ⊆ F • = 0 for every E ∈ E ⇐⇒ F • is an upper bound for {E • : E ∈ E}. So if E ⊆ Σ and H ∈ Σ, then H is an essential supremum of E in Σ, in the sense of 211G, iff H • is the supremum of A = {E • : E ∈ E} in A. P P Writing F for {F : F ∈ Σ, E \ F is negligible for every E ∈ E}, we see that B = {F • : F ∈ F } is just the set of upper bounds of A, and that H is an essential supremum of E iff H ∈ F and H • is a lower bound for B; that is, iff H • = sup A. Q Q (ii) Thus A is Dedekind complete iff every family in Σ has an essential supremum in Σ. Since we already know that (A, µ ¯) is semi-finite iff µ is, we see that (A, µ ¯) is localizable iff µ is. (f ) This is immediate from the definitions in 211I and 316L, if we always that {b : b ⊆ E • } = {F • : F ∈ Σ, F ⊆ E} (312Kb). (g), (h) follow at once from (f). 322C

I copy out the relevant parts of Theorem 211L in the new context.

Theorem (a) A probability algebra is totally finite. (b) A totally finite measure algebra is σ-finite. (c) A σ-finite measure algebra is localizable. (d) A localizable measure algebra is semi-finite. proof All except (c) are trivial; and (c) may be deduced from 211Lc-211Ld, 322Bc, 322Be and 321J, or from 316Fa and 322G below. 322D Of course not all the definitions in §211 are directly relevant to measure algebras. The concepts of ‘complete’, ‘locally determined’ and ‘strictly localizable’ measure space do not correspond in any direct way to properties of the measure algebras. Indeed, completeness is just irrelevant, as the next proposition shows. ˆ µ ˜ µ Proposition Let (X, Σ, µ) be a measure space, with completion (X, Σ, ˆ) and c.l.d. version (X, Σ, ˜) (213E). Write (A, µ ¯), (A1 , µ ¯1 ) and (A2 , µ ¯2 ) for the measure algebras of µ, µ ˆ and µ ˜ respectively. ˆ corresponds to an isomorphism between (A, µ (a) The embedding Σ ⊆ Σ ¯) and (A1 , µ ¯1 ). ˜ defines an order-continuous Boolean homomorphism π : A → A2 . Setting (b)(i) The embedding Σ ⊆ Σ Af = {a : a ∈ A, µ ¯a < ∞}, π¹ Af is a measure-preserving bijection between Af and Af2 = {c : c ∈ A2 , µ ¯2 c < ∞}. (ii) π is injective iff µ is semi-finite, and in this case µ ¯2 (πa) = µ ¯a for every a ∈ A. (iii) If µ is localizable, π is a bijection. ˆ I write F ∗ for its image in A1 ; and for G ∈ Σ, ˜ I proof For E ∈ Σ, I write E ◦ for its image in A; for F ∈ Σ, • write F for its image in A2 . (a) This is nearly trivial. The map E 7→ E ∗ : Σ → A1 is a Boolean homomorphism, being the composition ˆ and F 7→ F ∗ : Σ ˆ → A1 . Its kernel is {E : E ∈ Σ, µ of the Boolean homomorphisms E 7→ E : Σ → Σ Ê = 0} = {E : E ∈ Σ, µE = 0}, so it induces an injective Boolean homomorphism φ : A → A1 given by the formula φ(E ◦ ) = E ∗ for every E ∈ Σ (312F, 3A2G). To see that φ is surjective, take any b ∈ A1 . There is ˆ such that F ∗ = b, and there is an E ∈ Σ such that E ⊆ F and µ an F ∈ Σ ˆ(F \ E) = 0, so that π(E ◦ ) = E ∗ = F ∗ = b. Thus π is a Boolean algebra isomorphism. It is a measure algebra isomorphism because for any E ∈ Σ µ ¯1 φ(E ◦ ) = µ ¯1 E ∗ = µ Ê = µE = µ ¯E ◦ .

74

Measure algebras

322D

(b)(i) The map E 7→ E • : Σ → A2 is a Boolean homomorphism with kernel {E : E ∈ Σ, µ ˜E = 0} ⊇ {E : E ∈ Σ, µE = 0}, so induces a Boolean homomorphism π : A → A2 , defined by saying that πE ◦ = E • for every E ∈ Σ. If a ∈ Af , it is expressible as E ◦ where µE < ∞. Then µ ˜E = µE (213Fa), so πa = E • belongs to Af2 , 0 f and µ ¯2 (πa) = µ ¯a. If a, a are distinct of A , then µ ¯2 (πa 4 πa0 ) = µ ¯2 π(a 4 a0 ) = µ ¯(a 4 a0 ) > 0, so πa 6= πa0 ; thus π¹ Af is an injective map from Af to Af2 . If c ∈ Af2 , then c = G• where µ ˜G < ∞; by 213Fc, there is an E ∈ Σ such that E ⊆ G, µE = µ ˜G and µ ˜(G \ E) = 0, so that E ◦ ∈ Af and πE ◦ = E • = G• = c. As c is arbitrary, φ[Af ] = Af2 . Finally, π is order-continuous. P P Let A ⊆ A be a non-empty downwards-directed set with infimum 0, and b ∈ A2 a lower bound for π[A]. ?? If b 6= 0, then (because (A2 , µ ¯2 ) is semi-finite) there is a b0 ∈ Af2 such that 0 6= b0 ⊆ b. Let a0 ∈ A be such that πa0 = b0 . Then a0 6= 0, so there is an a ∈ A such that a 6⊇ a0 , that is, a ∩ a0 6= a0 . But now, because π¹ Af is injective, b0 = πa0 6= π(a ∩ a0 ) = πa ∩ πa0 = πa ∩ b0 , and b0 6⊆ πa, which is impossible. X X Thus b = 0, and 0 is the only lower bound of π[A]. As A is arbitrary, π is order-continuous (313L(b-ii)). Q Q (ii) (α) If µ is semi-finite, then µ ˜E = µE for every E ∈ Σ (213Hc), so µ ¯2 (πE ◦ ) = µ ¯2 E • = µ ˜E = µE = µ ¯E ◦ for every E ∈ Σ. In particular, πa = 0 =⇒ 0 = µ ¯2 (πa) = µ ¯a =⇒ a = 0, so π is injective. (β) If µ is not semi-finite, there is an E ∈ Σ such that µE = ∞ but µH = 0 whenever H ∈ Σ, H ⊆ E and µH < ∞; so that µ ˜E = 0 and E ◦ 6= 0,

πE ◦ = E • = 0.

So in this case π is not injective. ˜ there is an E ∈ Σ such that µ (iii) Now suppose that µ is localizable. Then for every G ∈ Σ ˜(E4G) = 0, ◦ • • by 213Hb; accordingly πE = E = G . As G is arbitrary, π is surjective; and we know from (ii) that π is injective, so it is a bijection, as claimed. 322E Proposition Let (A, µ ¯) be a measure algebra. (a) (A, µ ¯) is semi-finite iff it has a partition of unity consisting of elements of finite measure. (b) If (A, µ ¯) is semi-finite, a = sup{b : b ⊆ a, µ ¯b < ∞} and µ ¯a = sup{¯ µb : b ⊆ a, µ ¯b < ∞} for every a ∈ A. proof Set Af = {b : b ∈ A, µ ¯b < ∞}. (a)(i) If (A, µ ¯) is semi-finite, then Af is order-dense in A, so there is a partition of unity consisting of f of A (313K). (ii) If there is a partition of unity C ⊆ Af , and µ ¯a = ∞, then there is a c ∈ C such that a ∩ c 6= 0, and now a ∩ c ⊆ a and 0 < µ ¯(a ∩ c) < ∞; as a is arbitrary, (A, µ ¯) is semi-finite. (b) Of course Af is upwards-directed, by 321Bc, and we are supposing that its supremum is 1. If a ∈ A, then B = {b : b ∈ Af , b ⊆ a} = {a ∩ b : b ∈ Af } is upwards-directed and has supremum a (313Ba), so µ ¯a = supb∈B µ ¯b, by 321D. Remark Compare 213A. 322F Proposition If (A, µ ¯) is a semi-finite measure algebra, then A is a weakly (σ, ∞)-distributive Boolean algebra.

322J


75

proof Let hAn in∈N be a sequence of non-empty downwards-directed subsets of A, all with infimum 0. Set A = {supn∈N an : an ∈ An for every n ∈ N}. If c ∈ A \ {0}, let b ⊆ c be such that 0 < µ ¯b < ∞. For each n ∈ N, inf a∈An µ ¯(b ∩ an ) = 0, by 321F; so we may choose an ∈ An such that µ ¯(b ∩ an ) ≤ 2−n−2 µ ¯b. Set a = supn∈N an ∈ A. Then P∞ µ ¯(b ∩ a) ≤ n=0 µ ¯(b ∩ an ) < µ ¯b, so b 6⊆ a and c 6⊆ a. As c is arbitrary, inf A = 0; as hAn in∈N is arbitrary, A is weakly (σ, ∞)-distributive. 322G

Corresponding to 215B, we have the following description of σ-finite algebras.

Proposition Let (A, µ ¯) be a semi-finite measure algebra. Then the following are equiveridical: (i) (A, µ ¯) is σ-finite; (ii) A is ccc; (iii) either A = {0} or there is a functional ν¯ : A → [0, 1] such that (A, ν¯) is a probability algebra. proof (i) ⇐⇒ (ii) By 321J, it is enough to consider the case in which (A, µ ¯) is the measure algebra of a measure space (X, Σ, µ), and µ is semi-finite, by 322Bd. We know that A is ccc iff there is no uncountable dist set in Σ\N , where N is the ideal of negligible sets (316D). But 215B(iii) shows that this is equivalent to µ being σ-finite, which is equivalent to (A, µ ¯) being σ-finite, by 322Bc. (i)⇒(iii) If (A, µ ¯) is σ-finite, and A 6= {0}, let han in∈N be a dist sequence in A such ¯an < ∞ P∞that µ γ ¯an = 1. for every n and supn∈N an = 1. Then µ ¯ a > 0 for some n, so there are γ > 0 such that n n=0 n µ P∞ P∞ n ¯(a ∩ an ) for every a ∈ A; it is easy to ¯ai ).) Set νa = n=0 γn µ (Set γn0 = 2−n /(1 + µ ¯an ), γn = γn0 /( i=0 γi0 µ check that (A, ν¯) is a probability algebra. (iii)⇒(i) is a consequence of (ii)⇒(i). 322H Principal ideals If (A, µ ¯) is a measure algebra and a ∈ A, then it is easy to see (using 314Eb) that (Aa , µ ¯¹ Aa ) is a measure algebra, where Aa is the principal ideal of A generated by a. 322I Subspace measures General subspace measures give rise to complications in the measure algebra (see 322Xg, 322Yd). But subspaces with measurable envelopes (132D, 213K) are manageable. Proposition Let (X, Σ, µ) be a measure space, and A ⊆ X a set with a measurable envelope E. Let µA be the subspace measure on A, and ΣA its domain; let (A, µ ¯) be the measure algebra of (X, Σ, µ) and (AA , µ ¯A ) the measure algebra of (A, ΣA , µA ). Set a = E • and let Aa be the principal ideal of A generated by a. Then we have an isomorphism between (Aa , µ ¯¹ Aa ) and (AA , µ ¯A ) given by the formula F • 7→ (F ∩ A)◦ whenever F ∈ Σ and F ⊆ E, writing F • for the equivalence class of F in A and (F ∩ A)◦ for the equivalence class of F ∩ A in AA . proof Set ΣE = {E ∩ F : F ∈ Σ}. For F , G ∈ ΣE , F • = G• ⇐⇒ µ(F 4G) = 0 ⇐⇒ µA (A ∩ (F 4G)) = 0 ⇐⇒ (F ∩ A)◦ = (G ∩ A)◦ , because E is a measurable envelope of A. Accordingly the given formula defines an injective function from the image {F • : F ∈ ΣE } of ΣE in A to AA ; but this image is just the principal ideal Aa . It is easy to check that the map is a Boolean homomorphism from Aa to AA , and it is a Boolean isomorphism because ΣA = {F ∩ A : F ∈ ΣE }. Finally, it is measure-preserving because µ ¯F • = µF = µ∗ (F ∩ A) = µA (F ∩ A) = µ ¯A (F ∩ A)◦ for every F ∈ ΣE , again using the fact that E is a measurable envelope of A. 322J Corollary Let (X, Σ, µ) be a measure space, with measure algebra (A, µ ¯). (a) If E ∈ Σ, then the measure algebra of the subspace measure µE can be identified with the principal ideal AE • of A. (b) If A ⊆ X is a set of full outer measure (in particular, if µ∗ A = µX < ∞), then the measure algebra of the subspace measure µA can be identified with A.

76

Measure algebras

322K

322K Simple products (a) QLet h(Ai , µ ¯i )ii∈I be any indexed family ofP measure algebras. Let A be the simple product Boolean algebra i∈I Ai (315A), and for a ∈ A set µ ¯a = i∈I µ ¯i a(i). Then it is easy to check (using 315D(e-ii)) that (A, µ ¯) is a measure algebra; I will call it the simple product of the family h(Ai , µ ¯i )ii∈I . Each of the Ai corresponds to a principal ideal Aei say in A, where ei ∈ A corresponds to 1Ai ∈ Ai (315E), and the Boolean isomorphism between Ai and Aei is a measure algebra isomorphism between (Ai , µ ¯i ) and (Aei , µ ¯¹ Aei ). (b) If h(Xi , Σi , µi )ii∈I is a family of measure spaces, with direct sum (X, Σ, µ) (214K), then the measure algebra (A, µ ¯) of (X, Σ, µ) can be identified with the simple product of the measure algebras (Ai , µ ¯i ) of the (Xi , Σi , µi ). P P If, as in 214K, we set X = {(x, i) : i ∈ I, x ∈ Xi }, andQfor E ⊆ X, i ∈ I we set Ei = {x : (x, i) ∈ E}, Q then the Boolean isomorphism E 7→ hEi ii∈I : Σ → i∈I Σi induces a Boolean isomorphism from A to i∈I Ai , which is also a measure algebra isomorphism, because P P µ ¯E • = µE = i∈I µi Ei = i∈I µ ¯i Ei• for every E ∈ Σ. Q Q (c) A product of measure algebras is semi-finite, or localizable, or atomless, or purely atomic, iff every factor is. (Compare 214Jb.) (d) Let (A, µ ¯) be a localizable measure algebra.

Q ¯ ¹ Aei ) (i) If hei ii∈I is any partition of unity in A, then (A, µ ¯) is isomorphic to the product i∈I (Aei , µ of the corresponding principal ideals. P P By 315F(iii), the map a → 7 ha ∩ ei ii∈I is a Boolean isomorphism P Q µ ¯(a ∩ ei ) (321E), for ¯a = i∈I Q between A and i∈I Ai . Because hei ii∈I is dist and a = supi∈I a ∩ ei , µ ¯¹ Aei ). Q Q every a ∈ A. So a 7→ ha ∩ ei ii∈I is a measure algebra isomorphism between (A, µ ¯) and i∈I (Ai , µ (ii) In particular, since A has a partition of unity consisting of elements of finite measure (322Ea), (A, µ ¯) is isomorphic to a simple product of totally finite measure algebras. Each of these is isomorphic to the measure algebra of a totally finite measure space, so (A, µ ¯) is isomorphic to the measure algebra of a direct sum of totally finite measure spaces, which is strictly localizable. Thus every localizable measure algebra is isomorphic to the measure algebra of a strictly localizable measure space. (See also 322N below.) *322L Strictly localizable spaces The following fact is occasionally useful. Proposition Let (X, Σ, µ) be a strictly localizable measure space with µX > 0, and (A, µ ¯) its measure algebra. If hai ii∈I is a partition of unity in A, there is a dist family hXi ii∈I in Σ, with union X, such that Xi• = ai for every i ∈ I and Σ = {E : E ⊆ X, E ∩ Xi ∈ Σ ∀ i ∈ I, P µE = i∈I µ(E ∩ Xi ) for every E ∈ Σ; Q that is, the isomorphism between A and the simple product i∈I Aai of its principal ideals (315F) corresponds to an isomorphism between (X, Σ, µ) and the direct sum of the subspace measures on Xi . proof (a) Suppose to begin with that µX < ∞. In this case J = {i : S ai 6= 0} must be countable (322G). For each i ∈ J, choose Ei ∈ Σ such that Ei• = ai , and set Fi = Ei \ j∈J,j6=i Ej ; then Fi• = ai for each i ∈ J, and hFi ii∈J is dist. Because µX > 0, J is non-empty; fix some j0 ∈ J and set [ Xi = Fj0 ∪ (X \ Fj ) if i = j0 , j∈J

= Fi for i ∈ J \ {j0 }, = ∅ for i ∈ I \ J. S Then hXi ii∈I is a dist family in Σ, i∈I Xi = X and Xi• = ai for every i. Moreover, because only countably many of the Xi are non-empty, we certainly have Σ = {E : E ⊆ X, E ∩ Xi ∈ Σ ∀ i ∈ I,

322N


µE =

P i∈I

77

µ(E ∩ Xi ) for every E ∈ Σ.

(b) For the general case, start by taking a decomposition hYj ij∈J of X. We can suppose that no Yj is negligible, because there is certainly some j0 such that µYj0 > 0, and we can if necessary replace Yj0 by S Yj0 ∪ {Yj : µYj = 0}. For each j, we can identify the measure algebra of the subspace measure on Yj with the principal ideal Abj generated by bj = Yj• (322I). Now hai ∩ bjS ii∈I is a partition of unity in Abj , so by (a) • just above we can find a dist family hXji ii∈I in Σ such that i∈I Xji = Yj , Xji = ai ∩ bj for every i and Σ ∩ PYj = {E : E ⊆ Yj , E ∩ Xji ∈ Σ ∀ i ∈ I}, P µE = i∈I µ(E ∩ Xji ) for every E ∈ Σ ∩ PYj .

S Set Xi = j∈I Xji for every i ∈ I. Then hXi ii∈I is dist and covers X. Because Xi ∩ Yj = Xji is measurable for every j, Xi ∈ Σ. Because Xi• ⊇ ai ∩ bj for every j, and hbj ij∈J is a partition of unity in A (322Kb), Xi• ⊇ ai for each i; because hXi• ii∈I is dist and supi∈I ai = 1, Xi• = ai for every i. If E ⊆ X is such that E ∩ Xi ∈ Σ for every i, then E ∩ Xji ∈ Σ for all i ∈ I and j ∈ J, so E ∩ Yj ∈ Σ for every j ∈ J and E ∈ Σ. If E ∈ Σ, then X X X X µE = µ(E ∩ Yj ) = µ(E ∩ Xji ) = sup µ(E ∩ Xji ) j∈J

≤ so µE =

X i∈I

P i∈I

j∈J,i∈I

µ(E ∩ Xi ) =

sup

X

K⊆I is finite i∈K

i∈I K⊆J is finite j∈K

µ(E ∩ Xi ) ≤ µE;

µ(E ∩ Xi ). Thus hXi ii∈I is a suitable family.

322M Subalgebras: Proposition Let (A, µ ¯) be a measure algebra, and B a σ-subalgebra of A. Set ν¯ = µ ¯¹ B. (a) (B, ν¯) is a measure algebra. (b) If (A, µ ¯) is totally finite, or a probability algebra, so is (B, ν¯). (c) If (A, µ ¯) is σ-finite and (B, ν¯) is semi-finite, then (B, ν¯) is σ-finite. (d) If (A, µ ¯) is localizable and B is order-closed and (B, ν¯) is semi-finite, then (B, ν¯) is localizable. (e) If (B, ν¯) is a probability algebra, or totally finite, or σ-finite, so is (A, µ ¯). proof (a) By 314Eb, B is Dedekind σ-complete, and the identity map π : B → A is sequentially ordercontinuous; so that ν¯ = µ ¯π will be countably additive and (B, ν¯) will be a measure algebra. (b) This is trivial. (c) Use 322G. Every dist subset of B is dist in A, therefore countable, because A is ccc; so B is also ccc and (B, ν¯) (being semi-finite) is σ-finite. (d) By 314Ea, B is Dedekind complete; we are supposing that (B, ν¯) is semi-finite, so it is localizable. (e) This is elementary. 322N The Stone space of a localizable measure algebra I said above that the concepts of ‘strictly localizable’ and ‘locally determined’ measure space have no equivalents in the theory of measure algebras. But when we look at the canonical measure on the Stone space of a measure algebra, we can of course hope that properties of the measure algebra will be reflected in the properties of this measure, as happens in the next theorem. Theorem Let (A, µ ¯) be a measure algebra, Z the Stone space of A, and ν the standard measure on Z constructed by the method of 321J-321K. Then the following are equiveridical: (i) (A, µ ¯) is localizable; (ii) ν is localizable; (iii) ν is locally determined; (iv) ν is strictly localizable. proof Write Σ for the domain of ν, that is,

78

Measure algebras

322N

{E4A : E ⊆ Z is open-and-closed, A ⊆ Z is meager}, and M for the ideal of meager subsets of Z, that is, the ideal of ν-negligible sets (314M, 321K). Then a 7→ b a• : A → Σ/M is an isomorphism between (A, µ ¯) and the measure algebra of (Z, Σ, ν) (314M). Note that because any subset of a meager set is meager, ν is surely complete. (a)(i) ⇐⇒ (ii) is a consequence of 322Be. (b)(ii)⇒(iii) Suppose that ν is localizable. Of course it is semi-finite. Let V ⊆ Z be a set such that V ∩ E ∈ Σ whenever E ∈ Σ and νE < ∞. Because ν is localizable, there is a W ∈ Σ which is an essential supremum in Σ of {V ∩ E : E ∈ Σ, νE < ∞}, that is, W • = sup{(V ∩ E)• : νE < ∞} in Σ/M. I claim that W 4V is nowhere dense. P P Let G ⊆ Z be a non-empty open set. Then there is a non-zero a ∈ A such that b a ⊆ G. Because (A, µ ¯) is semi-finite, we may suppose that µ ¯a < ∞. Now (W ∩ b a)• = W • ∩ b a• = supνE<∞ (V ∩ E)• ∩ b a• = supνE<∞ (V ∩ E ∩ b a)• = (V ∩ b a)• , so (W 4V ) ∩ b a is negligible, therefore meager. But we know that A is weakly (σ, ∞)-distributive (322F), so that meager sets in Z are nowhere dense (316I), and there is a non-empty open set H ⊆ b a \ (W 4V ). Now H ⊆ G \ W 4V . As G is arbitrary, int W 4V = ∅ and W 4V is nowhere dense. Q Q But this means that W 4V ∈ M ⊆ Σ and V = W 4(W 4V ) ∈ Σ. As V is arbitrary, ν is locally determined. (c)(iii)⇒(iv) Assume that ν is locally determined. Because (A, µ ¯) is semi-finite, there is a partition of unity C ⊆ A consisting of elements of finite measure (322Ea). Set C = {b c : c ∈ C}. This is a dist family of sets of finite measure for ν. Now suppose that F ∈ Σ and νF > 0. Then there is an open-and-closed set E ⊆ Z such that F 4E is meager, and E is of the form b a for some a ∈ A. Since µ ¯a = νb a = νF > 0, there is some c ∈ C such that a ∩ c 6= 0, and now ν(F ∩ b c) = µ ¯(a ∩ c) > 0. This means that ν satisfies the conditions of 213O and must be strictly localizable. (d)(iv)⇒(ii) This is just 211Ld. b be the Dedekind completion 322O Proposition Let (A, µ ¯) be a semi-finite measure algebra, and let A b b µ of A (314U). Then there is a unique extension of µ ¯ to a functional µ ˜ on A such that (A, ˜) is a localizable b b µ measure algebra. The embedding A ⊆ A identifies the ideals {a : a ∈ A, µ ¯a < ∞} and {a : a ∈ A, ã < ∞}. b For c ∈ A, b set proof (I write the argument out as if A were actually a subalgebra of A.) µ ˜c = sup{¯ µa : a ∈ A, a ⊆ c}. b to [0, ∞] extending µ b µ Evidently µ ˜ is a function from A ¯, so µ ˜0 = 0. Because A is order-dense in A, ˜c > 0 b whenever c 6= 0, because any P such c includes a non-zero member of A. If hcn in∈N is a dist sequence in A ∞ ˜cn . P P Let A be the set of all of A expressible as a = supn∈N an with supremum c, then µ ˜c = n=0 µ where an ∈ A, an ⊆ cn for every n ∈ N. Now ∞ X

sup µ ¯a = sup{ a∈A

µ ¯an : an ∈ A, an ⊆ cn for every n ∈ N}

n=0

=

∞ X n=0

sup{¯ µan : an ⊆ cn } =

∞ X

µ ˜cn .

n=0

b cn = sup{a : a ∈ A, a ⊆ cn } for each n, and sup A, taken in A, b must be Also, because A is order-dense in A, 0 0 0 0 b c. But this means that if a ∈ A, a ⊆ c then a = supa∈A a ∩ a in A and therefore also in A; so that µ ¯a0 = supa∈A µ ¯(a0 ∩ a) ≤ supa∈A µ ¯a. Accordingly

322Q


µ ˜c = supa∈A µ ¯a =

P∞ n=0

79

µ ˜cn . Q Q

b µ This shows that (A, ˜) is a measure algebra. It is semi-finite because (A, µ ¯) is and every non-zero element b includes a non-zero element of A, which in turn includes a non-zero element of finite measure. Since A b of A b µ is Dedekind complete, (A, ¯) is localizable. If µ ¯a is finite, then surely µ ã = µ ¯a is finite. If µ ˜c is finite, then {A : a ∈ A, a ⊆ c} is upwards-directed b and supa∈A µ ¯A = µ ˜c is finite, so b = sup A is defined in A and µ ¯b = µ ˜c. Because A is order-dense in A, b = c (313K, 313O) and c ∈ A, with µ ¯c = µ ˜c. b µ 322P Definition Let (A, µ ¯) be any semi-finite measure algebra. I will call (A, ˜), as constructed above, the localization of (A, µ ¯). Of course it is unique just in so far as the Dedekind completion of A is. 322Q Further properties of Stone spaces: Proposition Let (A, µ ¯) be a semi-finite measure algebra and (Z, Σ, ν) its Stone space. (a) Meager sets in Z are nowhere dense; every E ∈ Σ is uniquely expressible as G4M where G ⊆ Z is open-and-closed and M is nowhere dense, and νE = sup{νH : H ⊆ E is open-and-closed}. (b) The c.l.d. version ν˜ of ν is strictly localizable, and has the same negligible sets as ν. (c) If (A, µ ¯) is totally finite then νE = inf{νH : H ⊇ E is open-and-closed} for every E ∈ Σ. proof (a) I have already remarked (in the proof of 322N) that A is weakly (σ, ∞)-distributive, so that meager sets in Z are nowhere dense. But we know that every member of Σ is expressible as G4M where G is open-and-closed and M is meager, therefore nowhere dense. Moreover, the expression is unique, because if G4M = G0 4M 0 then G4G0 ⊆ M ∪ M 0 is open and nowhere dense, therefore empty, so G = G0 and M = M 0. Now let a ∈ A be such that b a = G, and consider B = {b : b ∈ A, bb ⊆ E}. Then sup B = a in A. P P If b b ∈ B, then b \ b a ⊆ M is nowhere dense, therefore empty; so a is an upper bound for B. ?? If a is not the supremum of B, then there is a non-zero c ⊆ a such that b ⊆ a \ c for every b ∈ B. But now b c cannot be empty, so b c \ M is non-empty, and there is a non-zero d ∈ A such that db ⊆ b c \ M . In this case d ∈ B and d⊆ 6 a \ c. X X Thus a = sup B. Q Q It follows that νE = νG = µ ¯a = sup µ ¯b b∈B

= sup νbb ≤ sup{νH : H ⊆ E is open-and-closed} ≤ νE b∈B

and νE = sup{νH : H ⊆ E is open-and-closed}. (b) This is the same as part (c) of the proof of 322N. We have a dist family C of sets of finite measure for ν such that whenever E ∈ Σ, νE > 0 there is a C ∈ C such that µ(C ∩ E) > 0. Now if ν˜F is defined and not 0, there is an E ∈ Σ such that E ⊆ F and νE > 0 (213Fc), so that there is a C ∈ C such that ν(E ∩ C) > 0; since νC < ∞, we have ν˜(F ∩ C) ≥ ν˜(E ∩ C) = ν(E ∩ C) > 0. And of course ν˜C < ∞ for every C ∈ C. This means that C witnesses that ν˜ satisfies the conditions of 213O, so that ν˜ is strictly localizable. Any ν-negligible set is surely ν˜-negligible. If M is ν˜-negligible then it is nowhere dense. P P If G ⊆ Z is open and not empty then there is a non-empty open-and-closed set H1 ⊆ G, and now H1 ∈ Σ, so there is a non-empty open-and-closed set H ⊆ H1 such that νH is finite (because ν is semi-finite). In this case H ∩ M is ν-negligible, therefore nowhere dense, and H 6⊆ M . But this means that G 6⊆ M ; as G is arbitrary, M is nowhere dense. Q Q Accordingly M ∈ M and is ν-negligible. Thus ν and ν˜ have the same negligible sets. (c) Because νZ < ∞,

80

Measure algebras

322Q

νE = νZ − ν(Z \ E) = νZ − sup{νH : H ⊆ Z \ E is open-and-closed} = inf{ν(Z \ H) : H ⊆ Z \ E is open-and-closed} = inf{νH : H ⊇ E is open-and-closed}. 322X Basic exercises > (a) Let (A, µ ¯) be a measure algebra. Let I∞ be the set of those a ∈ A which are ‘purely infinite’, that is, µ ¯a = ∞ and µ ¯b = ∞ for every non-zero b ⊆ a. Show that I∞ is a σ-ideal of A. Show that there is a function µ ¯sf : A/I∞ → [0, ∞] defined by setting µ ¯sf a• = sup{¯ µb : b ⊆ a, µ ¯b < ∞} for every a ∈ A. Show that (A/I∞ , µ ¯sf ) is a semi-finite measure algebra. (b) Let (X, Σ, µ) be a measure space and let µsf be the ‘semi-finite version’ of µ, as defined in 213Xc. Let (A, µ ¯) be the measure algebra of (X, Σ, µ). Show that the measure algebra of (X, Σ, µsf ) is isomorphic to the measure algebra (A/I∞ , µ ¯sf ) of (a) above. ˜ µ (c) Let (X, Σ, µ) be a measure space and (X, Σ, ˜) its c.l.d. version. Let (A, µ ¯) and (A2 , µ ¯2 ) be the corresponding measure algebras, and π : A → A2 the canonical homomorphism, as in 322Db. Show that the kernel of π is the ideal I∞ , as described in 322Xa, so that A/I∞ is isomorphic, as Boolean algebra, to π[A] ⊆ A2 . Show that this isomorphism identifies µ ¯sf , as described in 322Xa, with µ ¯2 ¹π[A]. (d) Give a direct proof of 322G, not relying on 215B and 321J. > (e) Let (A, µ ¯) be any measure algebra, A a non-empty subset of A, and c ∈ A such that µ ¯c < ∞. Show that (i) c0 = sup{a ∩ c : a ∈ A} is defined in A (ii) there is a countable set B ⊆ A such that c0 = sup{a ∩ c : a ∈ B}. (f ) Let (X, Σ, µ) be a measure space and ν an indefinite-integral measure over µ (§234). Show that the measure algebra of ν can be identified, as Boolean algebra, with a principal ideal of the measure algebra of µ. (g) Let (X, Σ, µ) be a measure space and A any subset of X; let µA be the subspace measure on A and ΣA its domain. Write (A, µ ¯) for the measure algebra of (X, Σ, µ) and (AA , µ ¯A ) for the measure algebra of (A, ΣA , µA ). Show that the formula F • 7→ (F ∩ A)• defines a sequentially order-continuous Boolean homomorphism π : A → AA which has kernel I = {F • : F ∈ Σ, F ∩ A = ∅}. Show that for any a ∈ A, µ ¯A (πa) = min{¯ µb : b ∈ A, a \ b ∈ I}. (h) Let (A, µ ¯) be a measure algebra and B an order-closed subalgebra of A. Suppose that (B, µ ¯¹ B) is semi-finite. Show that (A, µ ¯) is semi-finite. (i) Let (A, µ ¯) be any measure algebra and (Z, Σ, ν) its Stone space. Show that the c.l.d. version of ν is strictly localizable. 322Y Further exercises (a) Let X be a set, Σ a σ-algebra of subsets of X, and I a σ-ideal of Σ. Set ˆ = {E4N : E ∈ Σ, N ∈ N }. N = {N : ∃ F ∈ I, N ⊆ F }. Show that N is a σ-ideal of subsets of X. Set Σ ˆ is a σ-algebra of subsets of X and that Σ/N ˆ Show that Σ is isomorphic to Σ/I. (b) Let (A, µ ¯) be a semi-finite measure algebra, and (Z, Σ, ν) its Stone space. Let ν˜ be the c.l.d. version ˜ its domain. Show that Σ ˜ is precisely the Baire property algebra {G4A : G ⊆ Z is open, A ⊆ Z of ν, and Σ ˜ is meager}, so that Σ/M can be identified with the regular open algebra of Z (314Yd) and the measure algebra of ν˜ can be identified with the localization of A. (c) Give an example of a localizable measure algebra (A, µ ¯) with a σ-subalgebra B such that (B, µ ¯¹ B) is semi-finite and atomless, but A is not atomless. (d) Let (X, Σ, µ) be a measure space and A ⊆ X a subset; let µA be the subspace measure on A, A and AA the measure algebras of µ and µA , and π : A → AA the canonical homomorphism, as described in 322Xg. (i) Show that if µA is semi-finite, then π is order-continuous. (ii) Show that if µ is semi-finite but µA is not, then π is not order-continuous.

322 Notes


81

(e) Show that if (A, µ ¯) is a measure algebra, with Stone space (Z, Σ, ν), then ν has locally determined negligible sets in the sense of 213I. (f ) Let (A, µ ¯) be a localizable measure algebra and (Z, Σ, ν) its Stone space. (i) Show that a function f : Z → R is Σ-measurable iff there is a conegligible set G ⊆ X such that f ¹G is continuous. (Hint: 316Yi.) (ii) Show that f : Z → [0, 1] is Σ-measurable iff there is a continuous function g : Z → [0, 1] such that f = g ν-a.e. 322 Notes and comments I have taken this leisurely tour through the concepts of Chapter 21 partly to recall them (or persuade you to look them up) and partly to give you practice in the elementary manipulations of measure algebras. The really vital result here is the correspondence between ‘localizability’ in measure spaces and measure algebras. Part of the object of this volume (particularly in Chapter 36) is to try to make sense of the properties of localizable measure spaces, as discussed in Chapter 24 and elsewhere, in of their measure algebras. I hope that 322Be has already persuaded you that the concept really belongs to measure algebras, and that the formulation in of ‘essential suprema’ is a dispensable expedient. I have given proofs of 322C and 322G depending on the realization of an arbitrary measure algebra as the measure algebra of a measure space, and the corresponding theorems for measure spaces, because this seems the natural approach from where we presently stand; but I am sympathetic to the view that such proofs must be inappropriate, and that it is in some sense better style to look for arguments which speak only of measure algebras (322Xd). For any measure algebra (A, µ ¯), the set Af of elements of finite measure is an ideal of A; consequently it is order-dense iff it includes a partition of unity (322E). In 322F we have something deeper: any semifinite measure algebra must be weakly (σ, ∞)-distributive when regarded as a Boolean algebra, and this has significant consequences in its Stone space, which are used in the proofs of 322N and 322Q. Of course a result of this kind must depend on the semi-finiteness of the measure algebra, since any Dedekind σ-complete Boolean algebra becomes a measure algebra if we give every non-zero element the measure ∞. It is natural to look for algebraic conditions on a Boolean algebra sufficient to make it ‘measurable’, in the sense that it should carry a semi-finite measure; this is an unresolved problem to which I will return in Chapter 39. Subspace measures, simple products, direct sums, principal ideals and order-closed subalgebras give no real surprises; I spell out the details in 322I-322M and 322Xg-322Xh. It is worth noting that completing a measure space has no effect on its measure algebra (322D, 322Ya). We see also that from the point of view of measure algebras there is no distinction to be made between ‘localizable’ and ‘strictly localizable’, since every localizable measure algebra is representable as the measure algebra of a strictly localizable measure space (322Kd). (But strict localizability does have implications for some processes starting in the measure algebra; see 322L.) It is nevertheless remarkable that the canonical measure on the Stone space of a semifinite measure algebra is localizable iff it is strictly localizable (322N). This canonical measure has many other interesting properties, which I skim over in 322Q, 322Xi, 322Yb and 322Yf. In Chapter 21 I discussed a number of methods of improving measure spaces, notably ‘completions’ (212C) and ‘c.l.d. versions’ (213E). Neither of these is applicable in any general way to measure algebras. But in fact we have a more effective construction, at least for semi-finite measure algebras, that of ‘localization’ (322O-322P); I say that it is more effective just because localizability is more important than completeness or local determinedness, being of vital importance in the behaviour of function spaces (241Gb, 243Gb, 245Ec, 363M, 364O, 365J, 367N, 369A, 369C). Note that the localization of a semi-finite measure algebra does in fact correspond to the c.l.d. b do not have the same Stone spaces, version of a certain measure (322Q, 322Yb). But of course A and A b can be effectively represented as the measure algebra of a measure on the Stone space of A. even when A b not just What is happening in 322Yb is that we are using all the open sets of Z to represent of A, the open-and-closed sets, which correspond to of A.

82

Measure algebras

§323 intro.

323 The topology of a measure algebra I take a short section to discuss one of the fundamental tools for studying totally finite measure algebras, the natural metric that each carries. The same ideas, suitably adapted, can be applied to an arbitrary measure algebra, where we have a topology corresponding closely to the topology of convergence in measure on the function space L0 . Most of the section consists of an analysis of the relations between this topology and the order structure of the measure algebra. 323A The pseudometrics ρa (a) Let (A, µ ¯) be a measure algebra. Write Af = {a : a ∈ A, µ ¯a < ∞}. f For a ∈ A and b, c ∈ A, write ρa (b, c) = µ ¯(a ∩ (b 4 c)). Then ρa is a pseudometric on A. P P (i) Because µ ¯a < ∞, ρa takes values in [0, ∞[. (ii) If b, c, d ∈ A then b 4 d ⊆ (b 4 c) ∪ (c 4 d), so ρa (b, d) = µ ¯(a ∩ (b 4 d)) ≤ µ ¯((a ∩ (b 4 c)) ∪ (a ∩ (c 4 d))) ≤µ ¯(a ∩ (b 4 c)) + µ ¯(a ∩ (c 4 d)) = ρa (b, c) + ρa (c, d). (iii) If b, c ∈ A then ρa (b, c) = µ ¯(a ∩ (b 4 c)) = µ ¯(a ∩ (c 4 b)) = ρa (c, b). Q Q (b) Now the measure-algebra topology of the measure algebra (A, µ ¯) is that generated by the family P = {ρa : a ∈ Af } of pseudometrics on A. Similarly the measure-algebra uniformity on A is that generated by P. (For a general discussion of topologies defined by pseudometrics, see 2A3F et seq. For the associated uniformities see §3A4.) (c) Note that P is upwards-directed, since ρa∪a0 ≥ max(ρa , ρa0 ) for all a, a0 ∈ Af . (d) When (A, µ ¯) is totally finite, it is more natural to work from the measure metric ρ = ρ1 , with ρ(a, b) = µ ¯(a 4 b), since this by itself defines the measure-algebra topology and uniformity. *(e) Even when (A, µ ¯) is not totally finite, we still have a metric ρ on Af defined by setting ρ(a, b) = f µ ¯(a 4 b) for all a, b ∈ A , which is sometimes useful (323Xg). Note however that the topology on Af defined from ρ is not in general the topology induced by the measure-algebra topology of A. 323B Proposition Let (A, µ ¯) be any measure algebra. Then the operations ∪ , ∩ , \ and 4 are all uniformly continuous. proof The point is that for any b, c, b0 , c0 ∈ A we have (b ∗ c) 4 (b0 ∗ c0 ) ⊆ (b 4 b0 ) ∪ (c 4 c0 ) for any of the operations ∗ = ∪ , ∩ etc.; so that if a ∈ Af then ρa (b ∗ c, b0 ∗ c0 ) ≤ ρa (b, b0 ) + ρa (c, c0 ). Consequently the operation ∗ must be uniformly continuous. 323C Proposition (a) Let (A, µ ¯) be a totally finite measure algebra. Then µ ¯ : A → [0, ∞[ is uniformly continuous. (b) Let (A, µ ¯) be a semi-finite measure algebra. Then µ ¯ : A → [0, ∞] is lower semi-continuous. (c) Let (A, µ ¯) be any measure algebra. If a ∈ A and µ ¯a < ∞, then b 7→ µ ¯(b ∩ a) : A → R is uniformly continuous. proof (a) For any a, b ∈ A, |¯ µa − µ ¯b| ≤ µ ¯(a 4 b) = ρ1 (a, b). (b) Suppose that b ∈ A and µ ¯b > α ∈ R. Then there is an a ⊆ b such that α < µ ¯a < ∞ (322Eb). If c ∈ A is such that ρa (b, c) < µ ¯a − α, then

323D

The topology of a measure algebra

83

µ ¯c ≥ µ ¯(a ∩ c) = µ ¯a − µ ¯(a ∩ (b \ c)) > α. Thus {b : µ ¯b > α} is open; as α is arbitrary, µ ¯ is lower semi-continuous. (c) |¯ µ(a 4 b) − µ ¯(a 4 c)| ≤ ρa (b, c) for all b, c ∈ A. 323D The following facts are basic to any understanding of the relationship between the order structure and topology of a measure algebra. Lemma Let (A, µ ¯) be a measure algebra. (a) Let B ⊆ A be a non-empty upwards-directed set. For b ∈ B set Fb = {c : b ⊆ c ∈ B}. (i) {Fb : b ∈ B} generates a Cauchy filter F(B ↑) on A. (ii) If sup B is defined in A, then it is a topological limit of F(B ↑); in particular, it belongs to the topological closure of B. (b) Let B ⊆ A be a non-empty downwards-directed set. For b ∈ B set Fb = {c : b ⊇ c ∈ B}. (i) {Fb : b ∈ B} generates a Cauchy filter F(B↓) on A. (ii) If inf B is defined in A, then it is a topological limit of F(B ↓); in particular, it belongs to the topological closure of B. (c)(i) Closed subsets of A are order-closed in the sense of 313D. (ii) An order-dense subalgebra of A must be dense in the topological sense. (d) Now suppose that (A, µ ¯) is semi-finite. (i) The sets {b : b ⊆ c}, {b : b ⊇ c} are closed for every c ∈ A. (ii) If B ⊆ A is non-empty and upwards-directed and e is a cluster point of F(B ↑), then e = sup B. (iii) If B ⊆ A is non-empty and downwards-directed and e is a cluster point of F(B↓), then e = inf B. proof I use the notations Af , ρa from 323A. (a)(i) (α) If b, c ∈ B then there is a d ∈ B such that b ∪ c ⊆ d, so that Fd ⊆ Fb ∩ Fc ; consequently F(B ↑) = {F : F ⊆ A, ∃ b ∈ B, Fb ⊆ F } is a filter on A. (β) Let a ∈ Af , ² > 0. Then there is a b ∈ B such that µ ¯(a ∩ c) ≤ µ ¯(a ∩ b) + 12 ² for every 0 0 0 c ∈ B, and Fb ∈ F(B ↑). If now c, c ∈ Fb , c 4 c ⊆ (c \ b) ∪ (c \ b), so ρa (c, c0 ) ≤ µ ¯(a ∩ c \ b) + µ ¯(a ∩ c0 \ b) = µ ¯(a ∩ c) + µ ¯(a ∩ c0 ) − 2¯ µ(a ∩ b) ≤ ². As a and ² are arbitrary, F(B ↑) is Cauchy. (ii) Suppose that e = sup B is defined in A. Let a ∈ Af , ² > 0. By 313Ba, a ∩ e = supb∈B a ∩ b; but {a ∩ b : b ∈ B} is upwards-directed, so µ ¯(a ∩ e) = supb∈B µ ¯(a ∩ b), by 321D. Let b ∈ B be such that µ ¯(a ∩ b0 ) ≥ µ ¯(a ∩ e) − ². Then for any c ∈ Fb , e 4 c ⊆ e \ b, so ρa (e, c) = µ ¯(a ∩ (e 4 c)) ≤ µ ¯(a ∩ (e \ b)) = µ ¯(a ∩ e) − µ ¯(a ∩ b) ≤ ². As a and ² are arbitrary, F(B ↑) → e. Because B ∈ F(B ↑), e surely belongs to the topological closure of B. (b) Either repeat the arguments above, with appropriate inversions, using 321F in place of 321D, or apply (a) to the set {1 \ b : b ∈ B}. (c)(i) This follows at once from (a) and (b) and the definition in 313D. (ii) If B ⊆ A is an order-dense subalgebra and a ∈ A, then B = {b : b ∈ B, b ⊆ a} is upwards-directed and has supremum a (313K); by (a-ii), a ∈ B ⊆ B. As a is arbitrary, B is topologically dense. (d)(i) Set F = {b : b ⊆ c}. If d ∈ A \ F , then (because (A, µ ¯) is semi-finite) there is an a ∈ Af such that δ=µ ¯(a ∩ d \ c) > 0; now if b ∈ F , ρa (d, b) ≥ µ ¯(a ∩ d \ b) ≥ δ, so that d cannot belong to the closure of F . As d is arbitrary, F is closed. Similarly, {b : b ⊇ c} is closed. (ii) (α) If b ∈ B, then e ∈ Fb , because Fb ∈ F(B ↑); but {c : b ⊆ c ∈ A} is a closed set including Fb , so contains e, and b ⊆ e. As b is arbitrary, e is an upper bound for B. (β) If d is an upper bound of B,

84

Measure algebras

323D

then {c : c ⊆ d} is a closed set belonging to F(B ↑), so contains e. As d is arbitrary, this shows that e is the supremum of B, as claimed. (iii) Use the same arguments as in (ii), but inverted. 323E Corollary Let (A, µ ¯) be a measure algebra. (a) If hbn in∈N is a non-decreasing sequence in A with supremum b, then hbn in∈N → b for the measurealgebra topology. (b) If hbn in∈N is a non-increasing sequence in A with infimum b, then hbn in∈N → b for the measure-algebra topology. proof I call this a ‘corollary’ because it is the special case of 323Da-323Db in which B is the set of of a monotonic sequence; but it is probably easier to work directly from the definition in 323A, and use 321Be or 321Bf to see that limn→∞ ρa (bn , b) = 0 whenever µ ¯a < ∞. 323F

The following is a useful calculation.

P∞ Lemma Let (A, µ ¯) be a measure algebra and hcn in∈N a sequence in A such that the sum n=0 µ ¯(cn 4 cn+1 ) is finite. Set d0 = supn∈N inf m≥n cm , d1 = inf n∈N supm≥n cm . Then d0 = d1 and, writing d for their common value, limn→∞ µ ¯(cn 4 d) = 0. P∞ proof Write αn = µ ¯(cn 4 cn+1 ), βn = k=n αk for n ∈ N; we are supposing that limn→∞ βn = 0. Set bn = supm≥n cm 4 cm+1 ; then P∞ µ ¯bn ≤ m=n µ ¯(cm 4 cm+1 ) = βn for each n. If m ≥ n, then cm 4 cn ⊆ supn≤k<m ck 4 ck+1 ⊆ bn , so cn \ bn ⊆ cm ⊆ cn ∪ bn . Consequently cn \ bn ⊆ inf k≥m ck ⊆ supk≥m ck ⊆ cn ∪ bn for every m ≥ n, and cn \ bn ⊆ d0 ⊆ d1 ⊆ cn ∪ bn , so that cn 4 d0 ⊆ bn ,

cn 4 d1 ⊆ bn ,

d1 \ d0 ⊆ bn .

As this is true for every n, limn→∞ µ ¯(cn 4 di ) ≤ limn→∞ µ ¯bn = 0 for both i, and µ ¯(d1 4 d0 ) ≤ inf n∈N µ ¯bn = 0, so that d1 = d0 . 323G The classification of measure algebras: Theorem Let (A, µ ¯) be a measure algebra, T its measure-algebra topology and U its measure-algebra uniformity. (a) (A, µ ¯) is semi-finite iff T is Hausdorff. (b) (A, µ ¯) is σ-finite iff T is metrizable, and in this case U is also metrizable. (c) (A, µ ¯) is localizable iff T is Hausdorff and A is complete under U. proof I use the notations Af , ρa from 323A. (a)(i) Suppose that (A, µ ¯) is semi-finite and that b, c are distinct of A. Then there is an a ⊆ b 4 c such that 0 < µ ¯a < ∞, and now ρa (b, c) > 0. As b and c are arbitrary, T is Hausdorff (2A3L).

323G


85

(ii) Suppose that T is Hausdorff and that b ∈ A has µ ¯b = ∞. Then b 6= 0 so there must be an a ∈ Af such that µ ¯(a ∩ b) = ρa (0, b) > 0; in which case a ∩ b ⊆ b and 0 < µ ¯(a ∩ b) < ∞. As b is arbitrary, µ ¯ is semi-finite. (b)(i) Suppose that µ ¯ is σ-finite. Let han in∈N be a non-decreasing sequence in Af with supremum 1. Set ρ(b, c) =

∞ X ρan (b, c) 1 + 2n µ ¯an n=0

for b, c ∈ A. Then ρ is a metric on A, because if ρ(b, c) = 0 then an ∩ (b 4 c) = 0 for every n, so b 4 c = 0 and b = c. If a ∈ Af and ² > 0, take n such that µ ¯(a \ an ) ≤ 12 ². If b, c ∈ A and ρ(b, c) ≤ ²/2(1 + 2n µ ¯an ), then 1 ρan (b, c) ≤ 2 ² so ρa (b, c) = ρa\an (b, c) + ρa∩an (b, c) ≤ µ ¯(a \ an ) + ρan (b, c) 1 2

≤ ² + (1 + 2n µ ¯an )ρ(b, c) ≤ ². In the other direction, given ² > 0, take n ∈ N such that 2−n ≤ 21 ²; then ρ(b, c) ≤ ² whenever ρan (b, c) ≤ ²/2(n + 1). This shows that U is the same as the metrizable uniformity defined by {ρ}; accordingly T is also defined by ρ. (ii) Now suppose that T is metrizable, and let ρ be a metric defining T. For each n ∈ N there must be an0 , . . . , ankn ∈ Af and δn > 0 such that ρani (b, 1) ≤ δn for every i ≤ kn =⇒ ρ(b, 1) ≤ 2−n . Set d = supn∈N,i≤kn ani . Then ρani (d, 1) = 0 for every n, i, so ρ(d, 1) ≤ 2−n for every n and d = 1. Thus 1 is the supremum of countably many elements of finite measure and (A, µ ¯) is σ-finite. (c)(i) Suppose that (A, µ ¯) is localizable. Then T is Hausdorff, by (a). Let F be a Cauchy filter on A. For each a ∈ Af , choose a sequence hFn (a)in∈N in F such that ρa (b, c) ≤ 2−n whenever b, c ∈ Fn (a) and n ∈ N. T Choose can ∈ k≤n Fk (a) for each n; then ρa (can , ca,n+1 ) ≤ 2−n for each n. Set da = supn∈N inf k≥n a ∩ cak . Then limn→∞ ρa (da , can ) = limn→∞ µ ¯(da 4 (a ∩ can )) = 0, by 323F. If a, b ∈ Af and a ⊆ b, then da = a ∩ db . P P For each n ∈ N, Fn (a) and Fn (b) both belong to F, so must have a point e in common; now ρa (da , db ) ≤ ρa (da , can ) + ρa (can , e) + ρa (e, cbn ) + ρa (cbn , db ) ≤ ρa (da , can ) + ρa (can , e) + ρb (e, cbn ) + ρb (cbn , db ) ≤ ρa (da , can ) + 2−n + 2−n + ρb (cbn , db ) → 0 as n → ∞. Consequently ρa (da , db ) = 0, that is, da = a ∩ da = a ∩ db . Q Q Set d = sup{db : b ∈ Af }; this is defined because A is Dedekind complete. Then F → d. P P If a ∈ Af and ² > 0, then a ∩ d = supb∈Af a ∩ db = supb∈Af a ∩ b ∩ da∪b = supb∈Af a ∩ b ∩ da = a ∩ da . So if we choose n ∈ N such that 2−n + ρa (can , da ) ≤ ², then for any e ∈ Fn (a) we shall have ρa (e, d) ≤ ρa (e, can ) + ρa (can , d) ≤ 2−n + ρa (can , da ) ≤ ². Thus

86

Measure algebras

323G

{e : ρa (d, e) ≤ ²} ⊇ Fn (a) ∈ F. As a, ² are arbitrary, F converges to d. Q Q As F is arbitrary, A is complete. (ii) Now suppose that T is Hausdorff and that A is complete under U. By (a), (A, µ ¯) is semi-finite. Let B be any non-empty subset of A, and set B 0 = {b0 ∪ . . . ∪ bn : b0 , . . . , bn ∈ B}, so that B 0 is upwardsdirected and has the same upper bounds as B. By 323Da, we have a Cauchy filter F(B 0 ↑); because A is complete, this is convergent; and because (A, µ ¯) is semi-finite, its limit must be sup B 0 = sup B, by 323Dd. As B is arbitrary, A is Dedekind complete, so (A, µ ¯) is localizable. 323H Closed subalgebras The ideas used in the proof of (c) above have many other applications, of which one of the most important is the following. You may find it helpful to read the next theorem first on the assumption that (A, µ ¯) is a probability algebra. Theorem Let (A, µ ¯) be a localizable measure algebra, and B a subalgebra of A. Then it is closed for the measure-algebra topology iff it is order-closed. proof (a) If B is closed, it must be order-closed, by 323Dc. (b) Now suppose that B is order-closed. I repeat the ideas of part (c-i) of the proof of 323G. Let e be any member of the closure of B in A. For each a ∈ Af , n ∈ N choose can ∈ B such that ρa (can , e) ≤ 2−n . Then ∞ X

µ ¯((a ∩ can ) 4 (a ∩ ca,n+1 )) =

n=0

≤

∞ X n=0 ∞ X

ρa (can , ca,n+1 ) ρa (can , e) + ρa (e, ca,n+1 ) < ∞.

n=0

So if we set ea = supn∈N inf k≥n cak , then ρa (ea , can ) = ρa (a ∩ ea , a ∩ can ) → 0 as n → ∞, by 323F, and ρa (e, ea ) = 0, that is, a ∩ ea = a ∩ e. Also, because B is order-closed, inf k≥n cak ∈ B for every n, and ea ∈ B. Because A is Dedekind complete, we can set e0a = inf{eb : b ∈ Af , a ⊆ b}; then e0a ∈ B and e0a ∩ a = inf b⊇a eb ∩ a = inf b⊇a eb ∩ b ∩ a = inf b⊇a e ∩ b ∩ a = e ∩ a. Now e0a ⊆ e0b whenever a ⊆ b, so B = {e0a : a ∈ Af } is upwards-directed, and sup B = sup{e0a ∩ a : a ∈ Af } = sup{e ∩ a : a ∈ Af } = e because (A, µ ¯) is semi-finite. Accordingly e ∈ B. As e is arbitrary, B is closed, as claimed. 323I Notation In the context of 323H, I will say simply that B is a closed subalgebra of A. 323J Proposition If (A, µ ¯) is a localizable measure algebra and B is a subalgebra of A, then the topological closure B of B in A is precisely the order-closed subalgebra of A generated by B. proof Write Bτ for the smallest order-closed subset of A including B. By 313Fc, Bτ is a subalgebra of A, and is the order-closed subalgebra of A generated by B. Being an order-closed subalgebra of A, it is topologically closed, by 323H, and must include B. On the other hand, B, being topologically closed, is order-closed (323D(c-i)), so includes Bτ . Thus B = Bτ is the order-closed subalgebra of A generated by B. 323K

I note some simple results for future reference.

Lemma If (A, µ ¯) is a localizable measure algebra and B is a closed subalgebra of A, then for any a ∈ A the subalgebra C of A generated by B ∪ {a} is closed. proof By 314Ja, C is order-closed.

323Y


87

323L Proposition Let h(Ai , µ ¯i )ii∈I be aQfamily of measure algebras with simple product (A, µ ¯) (322K). Then the measure-algebra topology on A = i∈I Ai defined by µ ¯ is just the product of the measure-algebra topologies of the Ai . proof I use the notations Af , ρa from 323A. Write T for the measure-algebra topology of A and S for the product topology. For i ∈ I, d ∈ Afi define a pseudometric ρ˜di on A by setting ρ˜di (b, c) = ρd (b(i), c(i)) whenever b, c ∈ A; then S is defined by P = {˜ ρdi : i ∈ I, a ∈ Afi } (3A3Ig). Now each ρ˜di is one of the defining pseudometrics for T, since ρ˜di (b, c) = µ ¯(d˜ ∩ (b4c)) ˜ = d, d(j) ˜ = 0 for j 6= i. So S ⊆ T. where d(i) P Now suppose that a ∈ Af and ² > 0. Then i∈I µ ¯i a(i) = µ ¯a is finite, so there is a finite set J ⊆ I such P that i∈I\J µ ¯i a(i) ≤ 12 ². For each j ∈ J, τj = ρã(j),j belongs to P, and ρa (b, c) =

X

µ ¯i (a(i) ∩ (b(i) 4 c(i))) ≤

i∈I

=

X j∈J

X j∈J

τj (b(j), c(j)) +

1 ² 2

1 2

µ ¯j (a(j) ∩ (b(j) 4 c(j))) + ²

≤²

whenever b, c are such that τj (b(j), c(j)) ≤ ²/(1 + 2#(J)) for every j ∈ J. By 2A3H, the identity map from (A, S) to (A, T) is continuous, that is, T ⊆ S. Putting these together, we see that S = T, as claimed. 323X Basic exercises (a) Let (A, µ ¯) be a semi-finite measure algebra. Show that the set {(a, b) : a ⊆ b} is a closed set in A × A. > (b) Let (X, Σ, µ) be a σ-finite measure space and (A, µ ¯) its measure algebra. (i) Show that if T is a σ-subalgebra of Σ, then {F • : F ∈ T} is a closed subalgebra of A. (ii) Show that if B is a closed subalgebra of A, then {F : F ∈ Σ, F • ∈ B} is a σ-subalgebra of Σ. (c) Let (A, µ ¯) be a localizable measure algebra, and C ⊆ A a set such that sup A, inf A belong to C for all non-empty subsets A of C. Show that C is closed for the measure-algebra topology. (d) (i) Show that if (A, µ ¯) is any measure algebra and B is a subalgebra of A, then its topological closure B is again a subalgebra. (ii) Use this fact instead of 313Fc to prove 323J. (e) Let (A, µ ¯) be a measure algebra, and e ∈ A; let Ae be the principal ideal of A generated by e, and µ ¯e its measure (322H). Show that the topology on Ae defined by µ ¯e is just the subspace topology induced by the measure-algebra topology of A. > (f ) Let (X, Σ, µ) be a measure space, and (A, µ ¯) its measure algebra. (i) Show that we have an injection χ : A → L0 (µ) (see §241) given by setting χ(E • ) = (χE)• for every E ∈ Σ. (ii) Show that χ is a homeomorphism between A and its image if A is given its measure-algebra topology and L0 (µ) is given its topology of convergence in measure (245A). (g) Let (A, µ ¯) be a measure algebra, and Af the ideal of elements of finite measure. For a, b ∈ Af set ρ(a, b) = µ ¯(a 4 b). Show that (Af , µ ¯) is a complete metric space and that the operations ∪ , ∩ , \ and 4 are uniformly continuous on Af , while µ ¯ : Af → R is also uniformly continuous. Show that the f embedding A ⊆ A is continuous for the measure-algebra topology on A. In the context of 323Xf, show that χ : Af → L0 (µ) is an isometry between Af and a subset of L1 (µ). 323Y Further exercises (a) Let (A, µ ¯) be a σ-finite measure algebra. Show that a set F ⊆ A is closed for the measure-algebra topology iff e ∈ F whenever there are non-empty sets B, C ⊆ A such that B is upwards-directed, C is downwards-directed, sup B = inf C = e and [b, c] ∩ F 6= ∅ for every b ∈ B, c ∈ C, writing [b, c] = {d : b ⊆ d ⊆ c}.

88

Measure algebras

323Yb

(b) Give an example to show that (a) is false for general localizable measure algebras. (c) Give an example of a semi-finite measure algebra (A, µ ¯) with an order-closed subalgebra which is not closed for the measure-algebra topology. (d) Let (A, µ ¯) be a probability algebra and write B for the family of closed subalgebras of A. For B, ¯(b 4 c) + supc∈C inf b∈B µ ¯(b 4 c). Show that (B, ρ) is a complete metric C ∈ B set ρ(B, C) = supb∈B inf c∈C µ space. (e) Let (A, µ ¯) be the measure algebra of Lebesgue measure on R. Show that it is separable in its measurealgebra topology. (Hint: 245Yj.) 323Z Problem Find a localizable measure algebra (A, µ ¯), a σ-subalgebra B of A and a sequence hbn in∈N in B which converges, for the measure-algebra topology, to a member of A \ B. 323 Notes and comments The message of this section is that the topology of a measure algebra is essentially defined by its order and algebraic structure; see also 324F-324H below. Of course the results are really about semi-finite measure algebras, and indeed this whole volume, like the rest of measure theory, has little of interest to say about others; they are included only because they arise occasionally and it is not absolutely essential to exclude them. We therefore expect to be able to describe such things as closed subalgebras and continuous homomorphisms in of the ordering, as in 323H and 324G. For σ-finite algebras, indeed, there is an easy description of the topology in of the order (323Ya). I think the result of this section which I shall most often wish to quote is 323I: in most contexts, there is no need to distinguish between ‘closed subalgebra’ and ‘order-closed subalgebra’. I conjecture, however, that a σ-subalgebra of a localizable measure algebra need not be topologically sequentially closed (323Z). It is also the case that the topology of a measure algebra corresponds very closely indeed to the topology of convergence in measure. A description of this correspondence is in 323Xf. Indeed all the results of this section have analogues in the theory of topological Riesz spaces. I will enlarge on the idea here in §367. For the moment, however, if you look back to Chapter 24, you will see that 323B and 323G are closely paralleled by 245D and 245E, while 323Ya is related to 245L. It is I think natural to ask whether there are any other topological Boolean algebras with the properties 323B-323D. In fact a question in this direction, the Control Measure Problem, is one of the most important questions outstanding in abstract measure theory. I will discuss it in §393; the particular form relevant to the present section is what I call ‘CM4 ’ (393J).

324 Homomorphisms In the course of Volume 2, I had occasion to remark that elementary measure theory was unusual among abstract topics in pure mathematics in not being dominated by any particular class of structure-preserving operators. We now come to what I think is one of the reasons for the gap: the most important operators of the theory are not between measure spaces at all, but between their measure algebras. In this section I run through the most elementary facts about Boolean homomorphisms between measure algebras. I start with results on the construction of such homomorphisms from functions between measure spaces (324A324E), then investigate continuity and order-continuity of homomorphisms (324F-324H) before turning to measure-preserving homomorphisms (324I-324O). 324A Theorem Let (X, Σ, µ) and (Y, T, ν) be measure spaces, and (A, µ ¯), (B, ν¯) their measure algebras. ˆ for the domain of the completion µ Write Σ ˆ of µ. Let D ⊆ X be a set of full outer measure (definition: ˆ D be the subspace σ-algebra on D induced by Σ. ˆ Let φ : D → Y be a function such that 132E), and let Σ −1 −1 ˆ D for every F ∈ T and φ [F ] is µ-negligible whenever νF = 0. Then there is a sequentially φ [F ] ∈ Σ order-continuous Boolean homomorphism π : B → A defined by the formula πF • = E • whenever F ∈ T, E ∈ Σ and (E ∩ D)4φ−1 [F ] is negligible.

324E

Homomorphisms

89

ˆ such that H ∩ D = φ−1 [F ]; now there is an E ∈ Σ such proof Let F ∈ T. Then there is an H ∈ Σ −1 that E4H is negligible, so that (E ∩ D)4φ [F ] is negligible. If E1 is another member of Σ such that (E1 ∩ D)4φ−1 [F ] is negligible, then (E4E1 ) ∩ D is negligible, so is included in a negligible member G of Σ. Since (E4E1 ) \ G belongs to Σ and is dist from D, it is negligible; accordingly E4E1 is negligible and E • = E1• in A. What this means is that the formula offered defines a map π : B → A. It is now easy to check that π is a Boolean homomorphism, because if (E ∩ D)4φ−1 [F ],

(E 0 ∩ D)4φ−1 [F 0 ]

are negligible, so are ((X \ E) ∩ D)4φ−1 [Y \ F ],

((E ∪ E 0 ) ∩ D)4φ−1 [F ∪ F 0 ].

To see that π is sequentially order-continuous, let hbn in∈N be a sequence in B. For each n we may • −1 choose S an Fn ∈ T such S that Fn = bn , and En ∈ Σ such that (En ∩ D)4φ [Fn ] is negligible; now, setting F = n∈N Fn , E = n∈N En , S (E ∩ D)4φ−1 [F ] ⊆ n∈N (En ∩ D)4φ−1 [Fn ] is negligible, so π(supn∈N bn ) = π(F • ) = E • = supn∈N En• = supn∈N πbn . (Recall that the maps E 7→ E • , F 7→ F • are sequentially order-continuous, by 321H.) So π is sequentially order-continuous (313L(c-iii)). 324B Corollary Let (X, Σ, µ) and (Y, T, ν) be measure spaces, and (A, µ ¯), (B, ν¯) their measure algebras. Let φ : X → Y be a function such that φ−1 [F ] ∈ Σ for every F ∈ T and µφ−1 [F ] = 0 whenever νF = 0. Then there is a sequentially order-continuous Boolean homomorphism π : B → A defined by the formula πF • = (φ−1 [F ])• for every F ∈ T. 324C Remarks (a) In §235 and elsewhere in Volume 2 I spent a good deal of time on functions between measure spaces which satisfy the conditions of 324A. Indeed, I take the trouble to spell 324A out in such generality just in order to catch these applications. Some of the results of the present chapter (322D, 322Jb) can also be regarded as special cases of 324A. (b) The question of which homomorphisms between the measure algebras of measure spaces (X, Σ, µ), (Y, T, ν) can be realized by functions between X and Y is important and deep; I will return to it in §§343-344. (c) In the simplified context of 324B, I have actually defined a contravariant functor; the relevant facts are the following. 324D Proposition Let (X, Σ, µ), (Y, T, ν) and (Z, Λ, λ) be measure spaces, with measure algebras ¯ Suppose that φ : X → Y and ψ : Y → Z satisfy the conditions of 324B, that is, (A, µ ¯), (B, ν¯), (C, λ). φ−1 [F ] ∈ Σ if F ∈ T,

µφ−1 [F ] = 0 if νF = 0,

ψ −1 [G] ∈ T if G ∈ Λ,

µψ −1 [G] = 0 if λG = 0.

Let πφ : B → A, πψ : C → B be the corresponding homomorphisms. Then ψφ : X → Z is another map of the same type, and πψφ = πφ πψ : C → A. proof The necessary checks are all elementary. 324E Stone spaces While in the context of general measure spaces the question of realizing homomorphisms is difficult, in the case of the Stone representation it is relatively straightforward. Proposition Let (A, µ ¯) and (B, ν¯) be measure algebras, with Stone spaces Z and W ; let µ, ν be the corresponding measures on Z and W , as described in 321J-321K, and Σ, T their domains. If π : B → A is

90

Measure algebras

324E

any order-continuous Boolean homomorphism, let φ : Z → W be the corresponding continuous function, as described in 312P. Then φ−1 [F ] ∈ Σ for every F ∈ T, µφ−1 [F ] = 0 whenever νF = 0, and (writing E ∗ for the member of A corresponding to E ∈ Σ) πF ∗ = (φ−1 [F ])∗ for every F ∈ T. proof Recall that E ∗ = a iff E4b a is meager, where b a is the open-and-closed subset of Z corresponding to a ∈ A. In particular, µE = 0 iff E is meager. Now the point is that φ−1 [F ] is nowhere dense in Z whenever F is a nowhere dense subset of W , by 313R. Consequently φ−1 [F ] is meager whenever F is meager in W , since F is then just a countable union of nowhere dense sets. Thus we see already that µφ−1 [F ] = 0 whenever νF = 0. If F is any member of T, there is an open-and-closed set F0 such that F 4F0 is meager; now φ−1 [F0 ] is open-and-closed, so φ−1 [F ] = φ−1 [F0 ]4φ−1 [F 4F0 ] belongs to Σ. Moreover, if b ∈ B is such that bb = F0 , and a = πb, then b a = φ−1 [F0 ], so πF ∗ = πb = a = (φ−1 [F0 ])∗ = (φ−1 [F ])∗ , as required. 324F

I turn now to the behaviour of order-continuous homomorphisms between measure algebras.

Theorem Let (A, µ ¯) and B, ν¯) be measure algebras and π : A → B a Boolean homomorphism. (a) π is continuous iff it is continuous at 0 iff it is uniformly continuous. (b) If (B, ν¯) is semi-finite and π is continuous, then it is order-continuous. (c) If (A, µ ¯) is semi-finite and π is order-continuous, then it is continuous. proof I use the notations Af , ρa from 323A. (a) Suppose that π is continuous at 0; I seek to show that it is uniformly continuous. Take b ∈ Bf and ² > 0. Then there are a0 , . . . , an ∈ Af and δ > 0 such that ν¯(b ∩ πc) = ρb (πc, 0) ≤ ² whenever maxi≤n ρai (c, 0) ≤ δ; setting a = supi≤n ai , ν¯(b ∩ πc) ≤ ² whenever µ ¯(a ∩ c) ≤ δ. 0

Now suppose that ρa (c, c ) ≤ δ. Then µ ¯(a ∩ (c 4 c0 )) ≤ δ, so ρb (πc, πc0 ) = ν¯(b ∩ (πc 4 πc0 )) = ν¯(b ∩ π(c 4 c0 )) ≤ ². As b, ² are arbitrary, π is uniformly continuous. The rest of the implications are elementary. (b) Let A be a non-empty downwards-directed set in A with infimum 0. Then 0 ∈ A (323D(b-ii)); because π is continuous, 0 ∈ π[A]. ?? If b is a non-zero lower bound for π[A] in B, then (because (B, ν¯) is semi-finite) there is a c ⊆ b with 0 < ν¯c < ∞; now ρc (πa, 0) = ν¯(c ∩ πa) = ν¯c > 0 for every a ∈ A, so 0 ∈ / π[A]. X X Thus inf π[A] = 0 in B; as A is arbitrary, π is order-continuous (313L(b-ii)). (c) By (a), it will be enough to show that π is continuous at 0. Let b ∈ Bf , ² > 0. ?? Suppose, if possible, that for every a ∈ Af , δ > 0 there is a c ∈ A such that µ ¯(a ∩ c) ≤ δ but ν¯(b ∩ πc) ≥ ². For each a ∈ Af , −n n ∈ N choose can such that µ ¯(a ∩ can ) ≤ 2 but ν¯(b ∩ πcan ) ≥ ². Set ca = inf n∈N supm≥n cam ; then P∞ µ ¯(a ∩ ca ) ≤ inf n∈N m=n µ ¯(a ∩ can ) = 0, so ca ∩ a = 0. On the other hand, because π is order-continuous, πca = inf n∈N supm≥n πcam , so that ν¯(b ∩ πca ) = limn→∞ ν¯(b ∩ supm≥n πcam ) ≥ ². This shows that ρb (1 \ a, 0) = ν¯(b ∩ π(1 \ a)) ≥ ν¯(b ∩ πca ) ≥ ². But now observe that A = {1 \ a : a ∈ Af } is a downwards-directed subset of A with infimum 0, because (A, µ ¯) is semi-finite. So π[A] is downwards-directed and has infimum 0, and 0 must be in the closure of π[A], by 323D(b-ii); while we have just seen that ρb (d, 0) ≥ ² for every d ∈ π[A]. X X

324K

Homomorphisms

91

Thus there must be a ∈ Af , δ > 0 such that ρb (πc, 0) = ν¯(b ∩ πc) ≤ ² whenever ρa (c, 0) = µ ¯(a ∩ c) ≤ δ. As b, ² are arbitrary, π is continuous at 0 and therefore continuous. 324G Corollary If (A, µ ¯) and (B, ν¯) are semi-finite measure algebras, a Boolean homomorphism π : A → B is continuous iff it is order-continuous. 324H Corollary If A is a Boolean algebra and µ ¯, ν¯ are two measures both rendering A a semi-finite measure algebra, then they endow A with the same uniformity (and, of course, the same topology). proof By 324G, the identity map from A to itself is continuous whichever of the topologies we place on A; and by 324F it is therefore uniformly continuous. 324I Definition Let (A, µ ¯) and (B, ν¯) be measure algebras. A Boolean homomorphism π : A → B is measure-preserving if ν¯(πa) = µ ¯a for every a ∈ A. ¯ be measure algebras, and π : A → B, θ : B → C 324J Proposition Let (A, µ ¯), (B, ν¯) and (C, λ) measure-preserving Boolean homomorphisms. Then θπ : A → C is a measure-preserving Boolean homomorphism. proof Elementary. 324K Proposition Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a measure-preserving Boolean homomorphism. (a) π is injective. (b) (A, µ ¯) is totally finite iff (B, ν¯) is, and in this case π is order-continuous, therefore continuous, and π[A] is a closed subalgebra of B. (c) If (A, µ ¯) is semi-finite and (B, ν¯) is σ-finite, then (A, µ ¯) is σ-finite. (d) If (A, µ ¯) is σ-finite and π is sequentially order-continuous, then (B, ν¯) is σ-finite. (e) If (A, µ ¯) is semi-finite and π is order-continuous, then (B, ν¯) is semi-finite. (f) If (A, µ ¯) is atomless and semi-finite, and π is order-continuous, then B is atomless. (g) If B is purely atomic and (A, µ ¯) is semi-finite, then A is purely atomic. proof (a) If a 6= 0 in A, then ν¯πa = µ ¯a > 0 so πa 6= 0. By 3A2Db, π is injective. (b) Because ν¯1B = ν¯π1A = µ ¯ 1A , (A, µ ¯) is totally finite iff (B, ν¯) is. Now suppose that A ⊆ A is downwards-directed and non-empty and that inf A = 0. Then inf a∈A ν¯πa = inf a∈A µ ¯a = 0 by 321F. So ν¯b = 0 for any lower bound b of π[A], and inf π[A] = 0. As A is arbitrary, π is order-continuous. By 324Fc, π is continuous. By 314Fa, π[A] is order-closed in B, that is, ‘closed’ in the sense of 323I. (c) I appeal to 322G. If C is a dist family in A \ {0}, then hπcic∈C is a dist family in B \ {0}, so is countable, and C must be countable, because π is injective. Thus A is ccc and (being semi-finite) (A, µ ¯) is σ-finite. (d) Let han in∈N be a sequence in A such that µ ¯an < ∞ for every n and supn∈N an = 1. Then ν¯πan < ∞ for every n and (because π is sequentially order-continuous) supn∈N πan = 1, so (B, ν¯) is σ-finite. (e) Setting Af = {a : µ ¯a < ∞}, sup Af = 1; because π is order-continuous, sup π[Af ] = 1 in B. So if ν¯b = ∞, there is an a ∈ Af such that πa ∩ b 6= 0, and now 0 < ν¯(b ∩ πa) < ∞.

92

Measure algebras

324K

(f ) Take any non-zero b ∈ B. As in (e), there is an a ∈ A such that µ ¯a < ∞ and a ∩ b 6= 0. If a ∩ b 6= b, then surely b is not an atom. Otherwise, set C = {c : c ∈ A, c ⊆ a, b ⊆ πc}. Then C is downwards-directed and contains a, so c0 = inf C is defined in A (321F), and µ ¯c0 = inf c∈C µ ¯c ≥ ν¯b > 0, so c0 6= 0. Because A is atomless, there is a d ⊆ c0 such that neither d nor c0 \ d is zero, so that neither c0 \ d nor d can belong to C. But this means that b ∩ πd and b ∩ π(c0 \ d) are both non-zero, so that again b is not an atom. As b is arbitrary, B is atomless. (g) Take any non-zero a ∈ A. Then there is an a0 ⊆ a such that 0 < µ ¯a0 < ∞. Because B is purely 0 atomic, there is an atom b of B with b ⊆ πa . Set C = {c : c ∈ A, c ⊆ a0 , b ⊆ πc}. Then C is downwards-directed and contains a0 , so c0 = inf C is defined in A, and µ ¯c0 = inf c∈C µ ¯c ≥ ν¯b > 0, so c0 6= 0. If d ⊆ c0 , then b ∩ πd must be either b or 0. If b ∩ πd = b, then d ∈ C and d = c0 . If b ∩ πd = 0, then c0 \ d ∈ C and d = 0. Thus c0 is an atom in A. As a is arbitrary, A is purely atomic. 324L Corollary Let (A, µ ¯) be a totally finite measure algebra, (B, ν¯) a measure algebra, and π : A → B a measure-preserving homomorphism. If C ⊆ A and C is the closed subalgebra of A generated by C, then π[C] is the closed subalgebra of B generated by π[C]. proof This is a special case of 314Gb. 324M Proposition Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras (A, µ ¯) and (B, ν¯). Let φ : X → Y be inverse-measure-preserving. Then we have a sequentially order-continuous measure-preserving Boolean homomorphism π : B → A defined by setting πF • = φ−1 [F ]• for every F ∈ T. proof This is immediate from 324B. 324N Proposition Let (A, µ ¯) and (B, ν¯) be measure algebras, with Stone spaces Z and W ; let µ, ν be the corresponding measures on Z and W . If π : B → A is an order-continuous measure-preserving Boolean homomorphism, and φ : Z → W the corresponding continuous function, then φ is inverse-measurepreserving. proof Use 324E. In the notation there, if F ∈ T, then νF = ν¯F ∗ = µ ¯πF ∗ = µ ¯φ−1 [F ]∗ = µφ−1 [F ]. 324O Proposition Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras, A0 a topologically dense subalgebra of A, and π : A0 → B a Boolean homomorphism such that ν¯πa = µ ¯a for every a ∈ A0 . Then π has a unique extension to a measure-preserving homomorphism from A to B. proof Let ρ, σ be the standard metrics on A, B, as in 323Ad. Then for any a, a0 ∈ A0 σ(πa, πa0 ) = ν¯(πa4πa0 ) = ν¯π(a4a0 ) = µ ¯(a4a0 ) = ρ(a, a0 ); that is, π : A0 → B is an isometry. Because A0 is dense in the metric space (A, ρ), while B is complete under σ (323Gc), there is a unique continuous function π ˆ : A → B extending π (3A4G). Now the operations (a, a0 ) 7→ π ˆ (a ∪ a0 ),

(a, a0 ) 7→ π â ∪ π ˆ a0 : A × A → B,

are continuous and agree on the dense subset A0 × A0 of A × A; because the topology of B is Hausdorff, they agree on A × A, that is, π ˆ (a ∪ a0 ) = π â ∪ π ˆ a0 for all a, a0 ∈ A (2A3Uc). Similarly, the operations a 7→ π ˆ (1 \ a),

a 7→ 1 \ π â : A → B

324X

Homomorphisms

93

are continuous and agree on the dense subset A0 of A, so they agree on A, that is, π ˆ (1 \ a) = 1 \ a for every a ∈ A. Thus π ˆ is a Boolean homomorphism. To see that it is measure-preserving, observe that a 7→ µ ¯a = ρ(a, 0),

a 7→ ν¯(ˆ π a) = σ(ˆ π a, 0) : A → R

are continuous and agree on A0 , so agree on A. Finally, π ˆ is the only measure-preserving Boolean homomorphism extending π, because any such map must be continuous (324Kb), and π ˆ is the only continuous extension of π. *324P The following fact will be useful in §386, by which time it will seem perfectly elementary; for the moment, it may be a useful exercise. Proposition Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras such that µ ¯1 = ν¯1. Suppose that A ⊆ A and φ : A → B are such that ν¯(inf i≤n φai ) = µ ¯(inf i≤n ai ) for all a0 , . . . , an ∈ A. Let C be the smallest closed subalgebra of A including A. Then φ has a unique extension to a measure-preserving Boolean homomorphism from C to B. proof (a) Let Ψ be the family of all functions ψ extending φ and having the same properties; that is, ψ is a function from a subset of A to B, and ν¯(inf i≤n ψai ) = µ ¯(inf i≤n ai ) for all a0 , . . . , an ∈ dom ψ. By Zorn’s Lemma, Ψ has a maximal member θ. Write D for the domain of θ. (b)(i) If c, d ∈ D then c ∩ d ∈ D. P P?? Otherwise, set D0 = D ∪ {c ∩ d} and extend θ to θ0 : D0 → B by 0 writing θ (c ∩ d) = θc ∩ θd. It is easy to check that θ0 ∈ Ψ, which is supposed to be impossible. X XQ Q Now ν¯(θc ∩ θd ∩ θ(c ∩ d)) = µ ¯(c ∩ d) = ν¯(θc ∩ θd) = ν¯θ(c ∩ d), so θ(c ∩ d) = θc ∩ θd. (ii) If d ∈ D then 1 \ d ∈ D. P P?? Otherwise, set D0 = D ∪ {1 \ d} and extend θ to D0 by writing θ (1 \ d) = 1 \ θd. Once again, it is easy to check that θ0 ∈ Ψ, which is impossible. X XQ Q Consequently (since D is certainly not empty, even if C is), D is a subalgebra of A (312B(iii)). 0

(iii) Since ν¯θ1 = µ ¯1 = ν¯1, θ1 = 1. If d ∈ D then ν¯θ(1 \ d) = µ ¯(1 \ d) = µ ¯1 − µ ¯d = ν¯1 − ν¯θd = ν¯(1 \ θd), while ν¯(θd ∩ θ(1 \ d)) = µ ¯(d ∩ (1 \ d)) = 0, so θd ∩ θ(1 \ d)) = 0, θ(1 \ d) ⊆ 1 \ θd and θ(1 \ d) must be equal to 1 \ θd. By 312H, θ : D → B is a Boolean homomorphism. (iv) Let D be the topological closure of D in A. Then it is an order-closed subalgebra of A (323J), so, with µ ¯¹ D, is a totally finite measure algebra in which D is a topologically dense subalgebra. By 324O, there is an extension of θ to a measure-preserving Boolean homomorphism from D to B; of course this extension belongs to Ψ, so in fact D = D is a closed subalgebra of A. (c) Since A ⊆ D, C ⊆ D and φ1 = θ¹ C is a suitable extension of φ. To see that φ1 is unique, let φ2 : C → B be any other measure-preserving Boolean homomorphism extending φ. Set C = {a : φ1 a = φ2 a}; then C is a topologically closed subalgebra of A including A, so is the whole of C, and φ2 = φ1 . 324X Basic exercises (a) Let A and B be Boolean algebras, of which A is Dedekind σ-complete, and φ : A → B a sequentially order-continuous Boolean homomorphism. Let I be an ideal of A included in the kernel of φ. Show that we have a sequentially order-continuous Boolean homomorphism π : A/I → B given by setting φ(a• ) = φa for every a ∈ A.

94

Measure algebras

324Xb

(b) Let (A, µ ¯) be a measure algebra, and B a σ-subalgebra of A. Show that provided that (B, µ ¯¹ B) is semi-finite, then the topology of B induced by µ ¯¹ B is just the subspace topology induced by the topology of A. (Hint: apply 324Fc to the embedding B ⊆ A.) ˜ µ (c) Let (X, Σ, µ) be a measure space and (X, Σ, ˜) its c.l.d. version. Let A, A2 be the corresponding measure algebras and π : A → A2 the canonical homomorphism (see 322Db). Show that π is topologically continuous. (d) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a bijective measure-preserving Boolean homomorphism. Show that π −1 : B → A is a measure-preserving homomorphism. (e) Let µ ¯ be counting measure on PN. Show that (PN, µ ¯) is a σ-finite measure algebra. Find a measurepreserving Boolean homomorphism from PN to itself which is not sequentially order-continuous. 324Y Further exercises (a) Let A and B be Boolean algebras, of which A is Dedekind complete, and φ : A → B an order-continuous Boolean homomorphism. Let I be an ideal of A included in the kernel of φ. Show that we have an order-continuous Boolean homomorphism π : A/I → B given by setting φ(a• ) = φa for every a ∈ A. (b) Let A be a Dedekind σ-complete Boolean algebra, and Z its Stone space. Write E for the algebra of open-and-closed subsets of Z, and Z for the family of nowhere dense zero sets of Z; let Zσ be the σ-ideal of subsets of Z generated by Z. Show that Σ = {E4U : E ∈ E, U ∈ Zσ } is a σ-algebra of subsets of Z, and describe a canonical isomorphism between Σ/Zσ and A. (c) Let A and B be Dedekind σ-complete Boolean algebras, with Stone spaces Z and W . Construct Zσ ⊆ Σ ⊆ PZ as in 324Yb, and let Wσ ⊆ T ⊆ PW be the corresponding structure defined from B. Let π : B → A be a sequentially order-continuous Boolean homomorphism, and φ : Z → W the corresponding continuous map. Show that if E ∗ ∈ A corresponds to E ∈ Σ, then πF ∗ = φ−1 [F ]∗ for every F ∈ T. (Hint: 313Yb.) (d) Let A be a Boolean algebra, B a ccc Boolean algebra and π : A → B an injective Boolean homomorphism. Show that A is ccc. (e) Let A be a Dedekind complete Boolean algebra, B a Boolean algebra, and π : A → B an ordercontinuous Boolean homomorphism. Show that for every atom b ∈ B there is an atom a ∈ A such that πa ⊇ b. Hence show that if A is atomless so is B, and that if B is purely atomic and π is injective then A is purely atomic. (f ) Let (A, µ ¯) and (B, ν¯) be localizable measure algebras and A0 an order-dense subalgebra of A. Suppose that π : A0 → B is an order-continuous Boolean homomorphism such that ν¯πa = µ ¯a for every a ∈ A0 . Show that π has a unique extension to a measure-preserving Boolean homomorphism from A to B. (g) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1]. (i) Show that there is an injective order-preserving function f : A → PN. (Hint: take a countable topologically dense subset D of A, and define f : A → P(D × N) by setting f (a) = {(d, q) : µ ¯(a ∩ d) ≥ q}.) (ii) Show that there is an order-preserving function h : PN → A such that h(f (a)) = a for every a ∈ A. (Hint: set h(I) = sup{a : f (a) ⊆ I}.) Compare 316Yo. (h) Let (A, µ ¯) and (B, ν¯) be probability algebras, and f : A → B an isometry for the measure metrics. Show that a 7→ f (a) 4 f (1) is a measure-preserving Boolean homomorphism. 324 Notes and comments If you examine the arguments of this section carefully, you will see that rather little depends on the measures named. Really this material deals with structures (X, Σ, I) where X is a set, Σ is a σ-ideal of subsets of X, and I is a σ-ideal of Σ, corresponding to the family of measurable negligible sets. In this abstract form it is natural to think in of sequentially order-continuous homomorphisms, as in 324Yc. I have stated 324E in of order-continuous homomorphisms just for a slight gain in

325A

Free products and product measures

95

simplicity. But in fact, when there is a difference, it is likely that order-continuity, rather than sequential order-continuity, will be the more significant condition. Note that when the domain algebra is σ-finite, the two concepts coincide, because it is ccc (316Fd, 322G). Of course I need to refer to measures when looking at such concepts as σ-finite measure algebra or measurepreserving homomorphism, but even here the real ideas involved are such notions as order-continuity and the countable chain condition, as you will see if you work through 324K. It is instructive to look at the translations of these facts into the context of inverse-measure-preserving functions; see 235Xe. 324H shows that we may speak of ‘the’ topology and uniformity of a Dedekind σ-complete Boolean algebra which carries any semi-finite measure; the topology of such an algebra is determined by its algebraic structure. Contrast this with the theory of normed spaces: two Banach spaces (e.g., `1 and `2 ) can be isomorphic as linear spaces, both being of algebraic dimension c, while they are not isomorphic as topological linear spaces. When we come to the theory of ordered linear topological spaces, however, we shall again find ourselves with operators whose algebraic properties guarantee continuity (355C, 367P).

325 Free products and product measures In this section I aim to describe the measure algebras of product measures as defined in Chapter 25. This will involve the concept of ‘free product’ set out in §315. It turns out that we cannot determine the measure algebra of a product measure from the measure algebras of the factors (325B), unless the product measure is localizable; but that there is nevertheless a general construction of ‘localizable measure algebra free product’, applicable to any pair of semi-finite measure algebras (325D), which represents the measure algebra of the product measure in the most important cases (325Eb). In the second part of the section (325I-325M) I deal with measure algebra free products of probability algebras, corresponding to the products of probability spaces treated in §254. 325A Theorem Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras (A, µ ¯) and (B, ν¯). ¯ be the corresponding measure Let λ be the c.l.d. product measure on X × Y , and Λ its domain; let (C, λ) algebra. (a)(i) The map E 7→ E × Y : Σ → Λ induces an order-continuous Boolean homomorphism from A to C. (ii) The map F 7→ X × F : T → Λ induces an order-continuous Boolean homomorphism from B to C. (b) The map (E, F ) 7→ E × F : Σ × T → Λ induces a Boolean homomorphism ψ : A ⊗ B → C. (c) ψ[A ⊗ B] is topologically dense in C. (d) For every c ∈ C, ¯ = sup{λ(c ¯ ∩ ψ(a ⊗ b)) : a ∈ A, b ∈ B, µ λc ¯a < ∞, ν¯b < ∞}. ¯ (e) If µ and ν are semi-finite, ψ is injective and λψ(a ⊗ b) = µ ¯a · µ ¯b for every a ∈ A, b ∈ B. proof (a) By 251E, E × Y ∈ Λ for every E ∈ Σ, and λ0 (E × Y ) = 0 whenever µE = 0, where λ0 is the primitive product measure described in 251A-251C; consequently λ(E × Y ) = 0 whenever µE = 0 (251F). Thus E 7→ (E × Y )• : Σ → C is a Boolean homomorphism with kernel including {E : µE = 0}, so descends to a Boolean homomorphism ε1 : A → C. To see that ε1 is order-continuous, let A ⊆ A1 be a non-empty downwards-directed set with infimum 0. ?? If there is a non-zero lower bound c of ε1 [A], express c as W • where W ∈ Λ. We have λ(W ) > 0; by the definition of λ (251F), there are G ∈ Σ, H ∈ T such that µG < ∞, νH < ∞ and λ(W ∩ (G × H)) > 0. Of course inf a∈A a ∩ G• = 0 in A, so inf a∈A µ ¯(a ∩ G• ) = 0, by 321F; let a ∈ A be such that µ ¯(a ∩ G• ) · νH < • λ(W ∩ (G × H)). Express a as E , where E ∈ Σ. Then λ(W \ (E × Y )) = 0. But this means that λ(W ∩ (G × H)) ≤ λ((E ∩ G) × H) = µ(E ∩ G) · νH = µ ¯(a ∩ G• ) · νH, contradicting the choice of a. X X Thus inf ε1 [A] = 0 in C; as A is arbitrary, ε1 is order-continuous. Similarly ε2 : B → C, induced by F 7→ X × F : T → Λ, is order-continuous. (b) Now there must be a corresponding Boolean homomorphism ψ : A ⊗ B → C such that ψ(a ⊗ b) = ε1 a ∩ ε2 b for every a ∈ A, b ∈ B, that is,

96

Measure algebras

325A

ψ(E • ⊗ F • ) = (E × Y )• ∩ (X × F )• = (E × F )• for every E ∈ Σ, F ∈ T (315I). ¯ < ∞ and ² > 0. Express c, e as U • , W • where U , W ∈ Λ. By 251Ie, there (c) Suppose that c, e ∈ C, λe S are E0 , . . . , En ∈ Σ, F0 , . . . , Fn ∈ T, all of finite measure, such that λ((U ∩ W )4 i≤n Ei × Fi ) ≤ ². Set S c1 = ( i≤n Ei × Fi )• ∈ ψ[A ⊗ B]; then

¯ ∩ (c 4 c1 )) = λ(W ∩ (U 4 S λ(e i≤n Ei × Fi )) ≤ ².

As c, e and ² are arbitrary, ψ[A ⊗ B] is topologically dense in C. (d) By the definition of λ, we have λW = sup{λ(W ∩ (E × F )) : E ∈ Σ, F ∈ T, µE < ∞, νF < ∞} for every W ∈ Λ; so all we have to do is express c as W • . (e) Now suppose that µ and ν are semi-finite. Then λ(E × F ) = µE · νF for any E ∈ Σ, F ∈ T (251J), ¯ so λψ(a ⊗ b) = µ ¯a · ν¯b for every a ∈ A, b ∈ B. To see that ψ is injective, take any non-zero c ∈ A ⊗ B; then there must be non-zero a ∈ A, b ∈ B such that a ⊗ b ⊆ c (315Jb), so that ¯ ≥ λψ(a ¯ λψc ⊗ b) = µ ¯a · ν¯b > 0 and ψc 6= 0. 325B Characterizing the measure algebra of a product space A very natural question to ask is, whether it is possible to define a ‘measure algebra free product’ of two abstract measure algebras in a way which will correspond to one of the constructions above. I give an example to show the difficulties involved. Example There are complete locally determined localizable measure spaces (X, µ), (X 0 , µ0 ), with isomorphic measure algebras, and a probability space (Y, ν) such that the measure algebras of the c.l.d. product measures on X × Y , X 0 × Y are not isomorphic. proof Let (X, Σ, µ) be the complete locally determined localizable not-strictly-localizable measure space described in 216E. Recall that, for E ∈ Σ, µE = #({γ : γ ∈ C, fγ ∈ E}) if this is finite, ∞ otherwise (216Eb), where C is a set of cardinal greater than c. The map E 7→ {γ : fγ ∈ E} : Σ → PC is surjective (216Ec), so descends to an isomorphism between A, the measure algebra of µ, and PC. Let (X 0 , Σ0 , µ0 ) be C with counting measure, so that its measure algebra (A0 , µ ¯0 ) is isomorphic to (A, µ ¯), while µ0 is of course strictly localizable. Let (Y, T, ν) be {0, 1}C with its usual measure. Let λ, λ0 be the c.l.d. product measures on X × Y , X 0 × Y ¯ (C0 , λ ¯ 0 ) the corresponding measure algebras. Then λ is not localizable (254U), so respectively, and (C, λ), ¯ (C, λ) is not localizable (322Be). On the other hand, λ0 , being the c.l.d. product of strictly localizable ¯ 0 ) is localizable, and is not isomorphic measures, is strictly localizable (251N), therefore localizable, so (C0 , λ ¯ to (C, λ). 325C Thus there can be no universally applicable method of identifying the measure algebra of a product measure from the measure algebras of the factors. However, you have no doubt observed that the example above involves non-σ-finite spaces, and conjectured that this is not an accident. In contexts in which we know that all the algebras involved are localizable, there are positive results available, such as the following. Theorem Let (X1 , Σ1 , µ1 ) and (X2 , Σ2 , µ2 ) be semi-finite measure spaces, with measure algebras (A1 , µ ¯1 ) ¯ the corresponding measure and (A2 , µ ¯2 ). Let λ be the c.l.d. product measure on X1 × X2 , and (C, λ) algebra. Let (B, ν¯) be a localizable measure algebra, and φ1 : A1 → B, φ2 : A2 → B order-continuous Boolean homomorphisms such that ν¯(φ1 (a1 ) ∩ φ2 (a2 )) = µ ¯1 a1 · µ ¯2 a2 for all a1 ∈ A1 , a2 ∈ A2 . Then there is a unique order-continuous measure-preserving Boolean homomorphism φ : C → B such that φ(ψ(a1 ⊗ a2 )) = φ1 (a1 ) ∩ φ2 (a2 ) for all a1 ∈ A1 , a2 ∈ A2 , writing ψ : A1 ⊗ A2 → C for the canonical map described in 325A.

325C


97

proof (a) Because ψ is injective, it is an isomorphism between A1 ⊗ A2 and its image in C. I trust it will cause no confusion if I abuse notation slightly and treat A1 ⊗ A2 as actually a subalgebra of C. Now the Boolean homomorphisms φ1 , φ2 correspond to a Boolean homomorphism θ : A1 ⊗ A2 → B. The point is ¯ for every c ∈ A ⊗ B. P that ν¯θc = λc P By 315Jb, every member of A1 ⊗ A2 is expressible as supi≤n ai ⊗ a0i , 0 0 where ai ∈ A1 , ai ∈ A2 and hai ⊗ ai ii≤n is dist. Now for each i we have ¯ i ⊗ a0 ), ν¯θ(ai ⊗ a0 ) = ν¯(φ1 (ai ) ∩ φ2 (a0 )) = µ ¯1 ai · µ ¯2 a0 = λ(a i

by 325Ad. So ν¯θ(c) =

i

Pn i=0

ν¯θ(ai ⊗ a0i ) =

i

Pn i=0

i

¯ Q ¯ i ⊗ a0 ) = λc. λ(a Q i

(b) The following fact will underlie many of the arguments below. If e ∈ B, ν¯e < ∞ and ² > 0, there are e1 ∈ Af1 , e2 ∈ Af2 such that ν¯(e \ θ(e1 ⊗ e2 )) ≤ ², writing Afi = {a : µ ¯i a < ∞}. P P Because (A1 , µ ¯1 ) is semi-finite, Af1 has supremum 1 in A1 ; because φ1 is order-continuous, sup{φ1 (a) : a ∈ Af1 } = 1 in B, and inf{e \ φ1 (a) : a ∈ Af1 } = 0 (313Aa). Because Af1 is upwards-directed, {e \ φ1 (a) : a ∈ Af1 } is downwards-directed, so inf{¯ ν (e \ φ(a) : a ∈ Af1 } = 0 (321F). Let e1 ∈ Af1 be such that ν¯(e \ φ1 (e1 )) ≤ 12 ². In the same way, there is an e2 ∈ Af2 such that ν¯(e \ φ2 (e2 )) ≤ 12 ². Consider e0 = e1 ⊗ e2 ∈ C. Then ν¯(e \ θe0 ) = ν¯(e \ (φ1 (e1 ) ∩ φ2 (e2 ))) ≤ ν¯(e \ φ1 (e1 )) + ν¯(e \ φ2 (e2 )) ≤ ². Q Q ¯ P (c) The next step is to check that θ is uniformly continuous for the uniformities defined by ν¯, λ. P Take ¯ 1 ⊗ e2 ) < ∞ and ν¯(e \ θ(e1 ⊗ e2 )) ≤ 1 ². Set any e ∈ Bf and ² > 0. By (b), there are e1 , e2 such that λ(e 2 ¯ 4 c0 ) ∩ e0 ) ≤ 1 ². Then e0 = e1 ⊗ e2 . Now suppose that c, c0 ∈ A1 ⊗ A2 and λ((c 2 ¯ 4 c0 ) ∩ e0 ) + 1 ² ≤ ². ν¯((θ(c) 4 θ(c0 )) ∩ e) ≤ ν¯θ((c 4 c0 ) ∩ e0 ) + ν¯(e \ θe0 ) ≤ λ((c 2

By 3A4Cc, θ is uniformly continuous for the subspace uniformity on A1 ⊗ A2 . Q Q (d) Recall that A1 ⊗ A2 is topologically dense in C (325Ab), while B is complete for its uniformity (323Gc). So there is a uniformly continuous function φ : C → B extending θ (3A4G). (e) Because θ is a Boolean homomorphism, so is φ. P P (i) The functions c 7→ φ(1 \ c), c 7→ 1 \ φ(c) are continuous and the topology of B is Hausdorff, so {c : φ(1 \ c) = 1 \ φ(c)} is closed; as it includes A1 ⊗ A2 , it must be the whole of C. (ii) The functions (c, c0 ) 7→ φ(c ∪ c0 ), (c, c0 ) 7→ φ(c) ∪ φ(c0 ) are continuous, so {(c, c0 ) : φ(c ∪ c0 ) = φ(c) ∪ φ(c0 )} is closed in C × C; as it includes (A1 ⊗ A2 ) × (A1 ⊗ A2 ), it must be the whole of C × C. Q Q (f ) Because θ is measure-preserving, so is φ. P P Take any e1 ∈ Af1 , e2 ∈ Af2 . Then the functions ¯ c 7→ λ(c ∩ (e1 ⊗ e2 )), c 7→ ν¯φ(c ∩ (e1 ⊗ e2 )) are continuous and equal on A1 ⊗ A2 , so are equal on C. The argument of (b) shows that for any b ∈ B, ν¯b = sup{¯ ν (b ∩ e) : e ∈ Bf } = sup{¯ ν (b ∩ φ(e1 ⊗ e2 )) : e1 ∈ Af1 , e2 ∈ Af2 }, so that ν¯φ(c) = sup{¯ ν φ(c ∩ (e1 ⊗ e2 )) : e1 ∈ Af1 , e2 ∈ Af2 } ¯ ∩ (e1 ⊗ e2 )) : e1 ∈ Af , e2 ∈ Af } = λc ¯ = sup{λ(c 1

2

for every c ∈ C. Q Q (g) To see that φ is order-continuous, take any non-empty downwards-directed set C ⊆ C with infimum 0. ?? If φ[C] has a non-zero lower bound b in B, let e ⊆ b be such that 0 < ν¯e < ∞. Let e0 ∈ C be such that ¯ 0 < ∞ and ν¯(e \ φ(e0 )) < ν¯e, as in (b) above, so that ν¯(e ∩ φ(e0 )) > 0. Now, because inf C = 0, there is a λe ¯ ∩ e0 ) < ν¯(e ∩ φ(e0 )). But this means that c ∈ C such that λ(c ¯ ∩ e0 ) < ν¯(e ∩ φ(e0 )) ≤ ν¯(b ∩ φ(e0 )), ν¯(b ∩ φ(e0 )) ≤ ν¯φ(c ∩ e0 ) = λ(c which is absurd. X X Thus inf φ[C] = 0 in B. As C is arbitrary, φ is order-continuous.

98

Measure algebras

325C

(h) Finally, to see that φ is unique, observe that any order-continuous Boolean homomorphism from C to B must be continuous (324Fc); so that if it agrees with φ on A1 ⊗ A2 it must agree with φ on C. 325D Theorem Let (A1 , µ ¯1 ) and (A2 , µ ¯2 ) be semi-finite measure algebras. ¯ together with order-continuous Boolean homomorphisms (a) There is a localizable measure algebra (C, λ), ψ1 : A1 → C, ψ2 : A2 → C such that whenever (B, ν¯) is a localizable measure algebra, and φ1 : A1 → B, φ2 : A2 → B are order-continuous Boolean homomorphisms and ν¯(φ1 (a1 ) ∩ φ2 (a2 )) = µ ¯1 a1 · µ ¯2 a2 for all a1 ∈ A1 , a2 ∈ A2 , then there is a unique order-continuous measure-preserving Boolean homomorphism φ : C → B such that φψj = φj for both j. ¯ ψ1 , ψ2 ) is determined up to isomorphism by this property. (b) The structure (C, λ, (c)(i) The Boolean homomorphism ψ : A1 ⊗ A2 → C defined from ψ1 and ψ2 is injective, and ψ[A1 ⊗ A2 ] is topologically dense in C. (ii) The order-closed subalgebra of C generated by ψ[A1 ⊗ A2 ] is the whole of C. ¯ (d) If j ∈ {1, 2} and (Aj , µ ¯j ) is localizable, then ψj [Aj ] is a closed subalgebra of (C, λ). proof (a)(i) We may regard (A1 , µ ¯1 ) as the measure algebra of (Z1 , Σ1 , µ1 ) where Z1 is the Stone space of A1 , Σ1 is the algebra of subsets of Z1 differing from an open-and-closed set by a meager set, and µ1 is an appropriate measure (321K). Note that in this representation, each a ∈ A1 becomes identified with b a• , where b a is the open-and-closed subset of Z1 corresponding to a. Similarly, we may think of (A2 , µ ¯2 ) as the measure algebra of (Z2 , Σ2 , µ2 ), where Z2 is the Stone space of A2 . (ii) Let λ be the c.l.d. product measure on Z1 × Z2 . The point is that λ is strictly localizable. P P By 322E, both A1 and A2 have partitions of unity consisting of elements of finite measure; let hci ii∈I , hdj ij∈J be such partitions. Then hb ci × dbj ii∈I,j∈J is a dist family of sets of finite measure in Z1 × Z2 . If W ⊆ Z1 × Z2 is such that λW > 0, there must be sets E1 , E2 of finite measure such that λ(W ∩ (E1 × E2 )) > 0. Because E1• = supi∈I E1• ∩ ci , we must have P P ci ). ¯1 (E1• ∩ ci ) = i∈I µ1 (E1 ∩ b µ1 E1 = µ ¯1 E1• = i∈I µ P Similarly, µ2 E2 = i∈J µ2 (E2 ∩ dbj ). But this means that there must be finite I 0 ⊆ I, J 0 ⊆ J such that P ci )µ2 (E2 ∩ dbj ) > µ1 E1 · µ2 E2 − λ(W ∩ (E1 × E2 )), i∈I 0 ,j∈J 0 µ1 (E1 ∩ b so that there have to be i ∈ I 0 , j ∈ J 0 such that λ(W ∩ (b ci × dbj )) > 0. Now this means that hb ci × dbj ii∈I,j∈J satisfies the conditions of 213O. Because λ is surely complete and locally determined, it is strictly localizable. Q Q ¯ to be just the measure algebra of λ. The maps ψ1 , ψ2 will be the (iii) We may therefore take (C, λ) canonical maps described in 325Aa, inducing the map ψ : A1 ⊗ B1 → C referred to in 325C; and 325C now gives the result. ¯ 0 , ψ 0 , ψ 0 ) with the same property. (b) This is nearly obvious. Suppose we had an alternative structure (C0 , λ 2 1 Then we must have an order-continuous measure-preserving Boolean homomorphism φ : C → C0 such that φψj = ψj0 for both j; and similarly we have an order-continuous measure-preserving Boolean homomorphism φ0 : C0 → C such that φ0 ψj0 = ψj for both j. Now φ0 φ : C → C is an order-continuous measure-preserving Boolean homomorphism such that φ0 ψj = ψj for both j. By the uniqueness assertion in (a), applied with B = C, φ0 φ must be the identity on C. In the same way, φφ0 is the identity on C0 . So φ and φ0 are the two halves of the required isomorphism. (c) In view of the construction for C offered in part (a) of the proof, (i) is just a consequence of 325Ac and 325Ae. Now (ii) follows by 323J. (d) If Aj is Dedekind complete then ψj [Aj ] is order-closed in C because ψj is order-continuous (314F(a-i)). ¯ together with embeddings ψ1 and ψ2 , 325E Remarks (a) We could say that a measure algebra (C, λ), as described in 325D, is a localizable measure algebra free product of (A1 , µ ¯1 ) and (A2 , µ ¯2 ); and its uniqueness up to isomorphism makes it safe, most of the time, to call it ‘the’ localizable measure algebra free product. Observe that it can equally well be regarded as the uniform space completion of the algebraic free product; see 325Yb.

*325H


99

(b) As the example in 325B shows, the localizable measure algebra free product of the measure algebras of given measure spaces need not appear directly as the measure algebra of their product. But there is one context in which it must so appear: if the product measure is localizable, 325C tells us at once that it has the right measure algebra. For σ-finite measure algebras, of course, any corresponding measure spaces have to be strictly localizable, so again we can use the product measure directly. 325F I ought not to proceed to the next topic without giving another pair of examples to show the subtlety of the concept of ‘measure algebra free product’. ¯ the measure algebra Example Let (A, µ ¯) be the measure algebra of Lebesgue measure µ on [0, 1], and (C, λ) 2 ¯ of Lebesgue measure λ on [0, 1] . Then (C, λ) can be regarded as the localizable measure algebra free product of (A, µ ¯) with itself, by 251M and 325Eb. Let ψ : A ⊗ A → C be the canonical map, as described in 325A. Then ψ[A ⊗ A] is not order-dense in C, and ψ is not order-continuous. P∞ P∞ proof (a) Let h²n in∈N be a sequence in [0, 1] such that n=0 ²n = ∞, but n=0 ²2n < 1; for instance, we 1 could take ²n = n+2 . Let hEn in∈N be a stochastically independent sequence of measurable subsets of [0, 1] such that µEn = ²n for each n. In A set an = En• , and consider cn = supi≤n ai ⊗ ai ∈ A ⊗ A for each n. (b) We have supn∈N cn = 1 in A ⊗ A. P P?? Otherwise, there is a non-zero a ∈ A ⊗ A such that a ∩ an = 0 for every n, and now there are non-zero b, b0 ∈ A such that S b ⊗ b0 ⊆ a. Set I = {n : anP ∩ b = 0}, J = {n : 0 an ∩ b } = 0. Then hEn in∈I is an independent family and µ( E ) ≤ 1 − µ ¯ b < 1, so i n∈I n∈I µEn < ∞, by P P the Borel-Cantelli lemma (273K). Similarly n∈J µEn < ∞. Because n∈N µEn = ∞, there must be some n ∈ N \ (I ∪ J). Now an ∩ b and an ∩ b0 are both non-zero, so 0 6= (an ∩ b) ⊗ (an ∩ b0 ) = (an ⊗ an ) ∩ (b ⊗ b0 ) = 0, which is absurd. X XQ Q (c) On the other hand,

P∞ ¯ P∞ P∞ µan )2 = n=0 ²2n < ∞, n=0 λψ(cn ) ≤ n=0 (¯

by the choice of the ²n . So supn∈N ψ(cn ) cannot be 1 in C. Thus ψ is not order-continuous. (d) By 313P(a-ii) and 313O, ψ[A ⊗ A] cannot be order-dense in C; alternatively, (b) shows that there can be no non-zero member of ψ[A ⊗ A] included in 1 \ supn∈N ψ(cn ). (Both these arguments rely tacitly on the fact that ψ is injective, as noted in 325Ae.) 325G Since 325F shows that the free product and the localizable measure algebra free product are very different constructions, I had better repeat an idea from §315 in the new context. ¯ the measure Example Again, let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1], and (C, λ) 2 algebra of Lebesgue measure on [0, 1] . Then there is no order-continuous Boolean homomorphism φ : C → A such that φ(a ⊗ b) = a ∩ b for all a, b ∈ A. P P Let φ : C → A be a Boolean homomorphism such that φ(a⊗b) = a ∩ b for all a, b ∈ A. For i < 2n let ani be the equivalence class in A of the interval [2−n i, 2−n (i+1)], ¯ n = 2−n for each n, so inf n∈N cn = 0 in C; and set cn = supi<2n ani ⊗ ani . Then φcn = 1 for every n, but λc thus φ cannot be order-continuous. Q Q (Compare 315P.) *325H Products of more than two factors We can of course extend the ideas of 325A, 325C and 325D to products of any finite number of factors. No new ideas are needed, so I spell the results out without proofs. (a) Let h(Ai , µ ¯i )ii∈I be a finite family of semi-finite measure algebras. Then there is a localizable mea¯ together with order-continuous Boolean homomorphisms ψi : Ai → C for i ∈ I, such sure algebra (C, λ), that whenever (B, ν¯) is a localizable measure Q algebra, and φi : Ai → B are order-continuous Boolean homomorphisms such that ν¯(inf i∈I φi (ai )) = i∈I µ ¯i ai whenever ai ∈ Ai for each i, then there is a unique order-continuous measure-preserving Boolean homomorphism φ : C → B such that φψi = φi for every i.

100

Measure algebras

325Hb

¯ hψi ii∈I ) is determined up to isomorphism by this property. (b) The structure (C, λ, (c) The Boolean homomorphism ψ : topologically dense in C.

N i∈I

Ai → C defined from the ψi is injective, and ψ[

N i∈I

Ai ] is

Nloc (d) Write c i∈I (Ai , µ ¯i ) for (a particular version of) the localizable measure algebra free product described in (a). If h(Ai , µ ¯i )ii∈I is a finite family of semi-finite measure algebras and hI(k)ik∈K is a partition of I, loc ¢ Nloc ¡N Nloc c then c (Ai , µ ¯i ) is isomorphic, in a canonical way, to c (Ai , µ ¯i ) . i∈I

k∈K

i∈I(k)

(e) Let h(Xi , Σi , µi )ii∈I be a finite family of semi-finite measure ¯i ) for the measure Q spaces, and write (Ai¯, µ algebra of (Xi , Σi , µi ). Let λ be the c.l.d. product measure on i∈I Xi (251W), and (C, λ) the corresponding ¯ into measure algebra. Then there is a canonical order-continuous measure-preserving embedding of (C, λ) the localizable measure algebra free product of the (Ai , µ ¯i ). If each µi is strictly localizable, this embedding is an isomorphism. 325I Infinite products Just as in §254, we can now turn to products of infinite families of probability algebras. Theorem Let h(Xi , Σi , µi )iQ ¯i ). Let λ i∈I be any family of probability spaces, with measure algebras (Ai , µ ¯ the corresponding measure algebra. For each i ∈ I, we have be the product measure on i∈I Xi , and (C, λ) a measure-preserving homomorphism ψi : Ai → C corresponding to the inverse-measure-preserving function x 7→ x(i) : X → Xi . Let Q(B, ν¯) be a probability algebra, and φi : Ai → B Boolean homomorphisms such ¯i ai whenever J ⊆ I is a finite set and ai ∈ Ai for every i. Then there is a that ν¯(inf i∈J φi (ai )) = i∈J µ unique measure-preserving Boolean homomorphism φ : C → B such that φψi = φi for every i ∈ I. proof (a) As remarked in 254Fb, all the maps x 7→ x(i) are inverse-measure-preserving, so correspond to measure-preserving homomorphisms ψi : Ai → C (324M). It will be helpful to use some notation from §254. Write C for the family of subsets of X expressible in the form E = {x : x ∈ X, x(i) ∈ Ei for every i ∈ J}, where J ⊆ I is finite and Ei ∈ Σi for every i ∈ J. Note that in this case E • = inf i∈J ψi (Ei• ). Set C = {E • : E ∈ C} ⊆ C, so that C is precisely the family of elements of C expressible in the form inf i∈J φi (ai ) where J ⊆ I is finite and ai ∈ Ai for each i. N The homomorphisms i∈I Ai → C (315I), which is N ψi : Ai → C define a Boolean homomorphism ψ : injective. P P If c ∈ i∈I Ai is non-zero, there must be a finite set J ⊆ I and a family hai ii∈J such that ai ∈ Ai \ {0} for each i and c ⊇ inf i∈J εi (ai ) (315Jb). Express each ai as Ei• , where Ei ∈ Σi . Then E = {x : x ∈ X, x(i) ∈ Ei for each i ∈ J} has measure λE =

Q i∈J

µEi =

Q i∈J

µ ¯ai 6= 0,

while E • = ψ(inf i∈J εi (ai )) ⊆ ψ(c), so ψ(c) 6= 0. As c is arbitrary, ψ is injective. Q Q N (b) Because ψ is injective, it is an isomorphismNbetween i∈I Ai and its image in C. I trust it will cause no confusion if I abuse notation slightly and treat i∈I NAi as actually a subalgebra of C, so that ψj : Aj → C becomes identified with the canonical map εj : Aj → i∈I Ai . Now the Boolean homomorphisms φi : Ai → N ¯ for every B correspond to a Boolean homomorphism θ : ¯θ(c) = λc i∈I Ai → B. The point is that ν

325L


101

N c ∈ i∈I Ai . P P Suppose to begin with that c ∈ C. Then we have c = E • , where E = {x : x(i) ∈ Ei ∀ i ∈ J} and Ei ∈ Σi for each i ∈ J. So Y Y ¯ = λE = λc µEi = µ ¯i Ei• = ν¯(inf φai ) i∈J

i∈J

i∈J

= ν¯(inf θψi (ai )) = ν¯θ(inf ψi (ai )) = ν¯θ(c). i∈J

i∈J

Next, any c ∈ C is expressible as the supremum of a finite dist family hck ik∈K in C (315Jb), so P P ¯ k ) = λc. ¯ Q ν¯θ(c) = k∈K ν¯θ(ck ) = k∈K λ(c Q ¯ since (c) It follows that θ is uniformly continuous for the metrics defined by ν¯, λ, ¯ 4 c0 ) ν¯(θ(c) 4 θ(c0 )) = ν¯θ(c 4 c0 ) = λ(c N for all c, c0 ∈ i∈I Ai . N (d) Next, P Let c ∈ C, ² > 0. Express c as W • . Then by 254Fe i∈I Ai is topologically dense Sin C. P there are H0 , . . . , Hk ∈ C such that λ(W 4 j≤k Hj ) ≤ ². Now cj = Hj• ∈ C for each j, so S N c0 = supj≤k cj = ( j≤k Hj )• ∈ i∈I Ai , ¯ 4 c0 ) ≤ ². Q and λ(c Q Since B is complete for its uniformity (323Gc), there is a uniformly continuous function φ : C → B extending θ (3A4G). (e) Because θ is a Boolean homomorphism, so is φ. P P (i) The functions c 7→ φ(1 \ c), 1 \ φ(c) are continuous and the topology of B is Hausdorff, so {c : φ(1 \ c) = 1 \ φ(c)} is closed; as it includes A1 ⊗ A2 , it must be the whole of C. (ii) The functions (c, c0 ) 7→ φ(c ∪ c0 ), (c, c0 ) 7→ φ(c) ∪ φ(c0 ) are continuous, so {(c, c0 ) : φ(c ∪ c0 ) = φ(c) ∪ φ(c0 )} is closed in C × C; as it includes (A1 ⊗ A2 ) × (A1 ⊗ A2 ), it must be the whole of C × C. Q Q ¯ c 7→ ν¯φ(c) are continuous and (f ) Because θ is measure-preserving, so is φ. P P The functions c 7→ λc, equal on A1 ⊗ A2 , so are equal on C. Q Q (g) Finally, to see that φ is unique, observe that any measure-preserving Boolean homomorphism from N C to B must be continuous, so that if it agrees with φ on i∈I Ai it must agree with φ on C. 325J

Of course this leads at once to a result corresponding to 325D.

Theorem Let h(Ai , µ ¯i )ii∈I be a family of probability algebras. ¯ together with order-continuous Boolean homomorphisms ψi : (a) There is a probability algebra (C, λ), Ai → C such that whenever (B, Q ν¯) is a probability algebra, and φi : Ai → B are Boolean homomorphisms ¯i ai whenever J ⊆ I is finite and ai ∈ Ai for each i ∈ J, then there is a such that ν¯(inf i∈J φi (ai )) = i∈J µ unique measure-preserving Boolean homomorphism φ : C → B such that φψj = φj for every j. ¯ hψi ii∈I ) is determined up to isomorphism by this property. (b) The structure (C, λ, N N (c) The Boolean homomorphism ψ : i∈I Ai → C defined from the ψi is injective, and ψ[ i∈I Ai ] is topologically dense in C. proof For (a) and (c), all we have to do is represent each (Ai , µ ¯i ) as the measure algebra of a probability space, and apply 325I. The uniqueness of C and the ψi follows from the uniqueness of the homomorphisms φ, as in 325D. ¯ hψi ii∈I ) is a, or the, probability algebra free 325K Definition As in 325Ea, we can say that (C, λ, product of h(Ai , µ ¯i )ii∈I . 325L Independent subalgebras If (A, µ ¯) is a probability algebra, Q we say that a family hBi ii∈I of subalgebras of A is (stochastically) independent if µ ¯(inf i∈J bi ) = i∈J µ ¯bi whenever J ⊆ I is finite and bi ∈ Bi for each i. (Compare 272Ab.) In this case the embeddings Bi ⊆ A give rise to an embedding of the probability algebra free product of h(Bi , µ ¯¹ Bi )ii∈I into A. (Compare 272J, 315Xn.)

102

Measure algebras

325M

325M We can now make a general trawl through Chapters 25 and 27 seeking results which can be expressed in the language of this section. I give some in 325Xe-325Xh. Some ideas from §254 which are thrown into sharper relief by a reformulation are in the following theorem. ¯ hψi ii∈I ) their probability algebra Theorem Let h(Ai , µ ¯i )ii∈I be a family of probability algebras and (C, λ, S free product. For J ⊆ I let CJ be the closed subalgebra of C generated by i∈J ψi [Ai ]. ¯ CJ , hψi ii∈J ) is a probability algebra free product of h(Ai , µ (a) For any J ⊆ I, (CJ , λ¹ ¯i )ii∈J . (b) For any c ∈ C, there is a unique smallest J ⊆ I such that c ∈ C , and this J is countable. J T (c) For any non-empty family J ⊆ PI, J∈J CJ = CT J . proof (a) If (B, ν¯, hφi ii∈J ) is any probability algebra free product of h(Ai , µ ¯i )ii∈J , then we have a measurepreserving homomorphism ψ : B → C such that ψφ = ψ for every i ∈ J. Because the subalgebra i i S S B0 of B generated by i∈J φi [Ai ] is topologically dense in B (325Jc), and ψ is continuous (324Kb), i∈J ψi [Ai ] is topologically dense in S ψ[B]; also ψ[B] is closed in C (324Kb again). But this means that ψ[B] is just the topological closure of i∈I ψi [Ai ] and must be CJ . Thus ψ is an isomorphism, and ¯ CJ , hψi ii∈J ) = (ψ[B], ν¯ψ −1 , hψφi ii∈I ) (CJ , λ¹ is also a probability algebra free product of h(Ai , µ ¯i )ii∈J . (b) As in 325J, we may suppose that each (Ai , µ ¯i ) is the measure algebra of a probability space (Xi , Σi , µi ), and that C is the measure algebra of their product (X, Λ, λ). Let W ∈ Λ be such that c = W • . By 254Rd, there is a unique smallest K ⊆ I such that W 4U is negligible for some U ∈ ΛK , where ΛK is the set of of Λ which are determined by coordinates in K; and K is countable. But if we look at any J ⊆ I, {x : x(i) ∈ E} ∈ ΛJ for every i ∈ J, E ∈ Σi ; so {U • : U ∈ ΛJ } is a closed subalgebra of C including ψi [Ai ] for every i ∈ J, and therefore including CJ . On the other hand, as observed in 254Ob, any member of ΛJ is approximated, in measure, by sets in the σ-algebra TJ generated by sets of the form {x : x(i) ∈ E} where i ∈ J, E S ∈ Σi . Of course TJ ⊆ ΛJ , so {W • : W ∈ ΛJ } = {W • : W ∈ TJ } is the closed subalgebra of C generated by i∈K ψi [Ai ], which is CJ . Thus K is also the unique smallest subset of I such that c ∈ CK . T T other hand, suppose (c) Of T course CK ⊆ CJ whenever K ⊆ J ⊆ I, so T J∈J CJ ⊇ C J . On the T . As c is arbitrary, C ; then by (b) there is some K ⊆ J such that c ∈ C ⊆ C that c ∈ K J J∈J J T T J∈J CJ = C J . *325N Notation In this context, I will say that an element c of C is determined by coordinates in J if c ∈ CJ . 325X Basic exercises (a) Let (A1 , µ ¯1 ), (A2 , µ ¯2 ) be two semi-finite measure algebras, and suppose that for each j we are given a closed subalgebra Bj of Aj such that (Bj , ν¯j ) is also semi-finite, where ν¯j = µ ¯j ¹ Bj . Show that the localizable measure algebra free product of (B1 , ν¯1 ) and (B2 , ν¯2 ) can be thought of as a closed subalgebra of the localizable measure algebra free product of (A1 , µ ¯1 ) and (A2 , µ ¯2 ). (b) Let (A1 , µ ¯1 ), (A2 , µ ¯2 ) be two semi-finite measure algebras, and suppose that for each j we are given a principal ideal Bj of Aj . Set ν¯j = µ ¯j ¹ Bj . Show that the localizable measure algebra free product of (B1 , ν¯1 ) and (B2 , ν¯2 ) can be thought of as a principal ideal of the localizable measure algebra free product of (A1 , µ ¯1 ) and (A2 , µ ¯2 ). > (c) Let h(Ai , µ ¯i )ii∈I and h(Bj , ν¯j )ij∈J be families of semi-finite measure algebras, with simple products b loc (B, ν¯) can be (A, µ ¯) and (B, ν¯) (322K). Show that the localizable measure algebra free product (A, µ ¯ )⊗ b identified with the simple product of the family h(Ai , µ ¯i )⊗loc (Bj , ν¯j )ii∈I,j∈J . ¯ hψi ii∈I ) their probability algebra >(d) Let h(Ai , µ ¯i )ii∈I be a family of probability algebras, and (C, λ, free product. Suppose that for each i ∈ I we are given a measure-preserving Boolean homomorphism πi : Ai → Ai . Show that there is a unique measure-preserving Boolean homomorphism π : C → C such that πψi = ψi πi for every i ∈ I.

325 Notes


103

> (e) Let (A, µ ¯) be a probability algebra. We say that a family hai ii∈I in A is (stochastically) indeQ pendent if µ ¯(inf i∈j aj ) = i∈J µ ¯ai for every non-empty finite J ⊆ I. Show that this is so iff hAi ii∈I is independent, where Ai = {0, ai , 1 \ ai , 1} for each i. (Compare 272F.) (f ) Let (A, µ ¯) be a probability algebra, and hAi ii∈I a stochastically independent family of closed subalgebras of A. Let hJ(k)ik∈K be S a dist family of subsets of I, and for each k ∈ K let Bk be the closed subalgebra of A generated by i∈J(k) Ai . Show that hBk ik∈K is independent. (Compare 272K.) (g) Let (A, µ ¯) be a probability algebra, and hAi ii∈I a stochastically independent family of closed subalS T gebras of A. For J ⊆ I let BJ be the closed subalgebra of A generated by i∈J Ai . Show that {BI\J : J is a finite subset of I} = {0, 1}. (Hint: For J ⊆ I, show that µ ¯(b ∩ c) = µ ¯b · µ ¯c for every b ∈ BI\J , c ∈ BJ . Compare 272O, 325M.) ¯ hψi ii∈I ). (h) Let h(Ai , µ ¯i )ii∈I be a family of probability algebras with probability algebra free product (C, λ, For c ∈ C let J be the smallest subset of I such that c belongs to the closed subalgebra of C generated by c S ψ [A ]. Show that if c ⊆ d in C, then there is an e ∈ C such that c ⊆ e ⊆ d and J ⊆ J ∩ J . (Hint: e c d i∈Jc i i 254R.) 325Y Further exercises (a) Let µ be counting measure on X = {0}, µ0 the countable-cocountable measure on X 0 = ω1 , and ν counting measure on Y = ω1 . Show that the measure algebras of the primitive product measures on X × Y , X 0 × Y are not isomorphic. (b) Let A be a Boolean algebra, and µ : A → [0, ∞] a function such that µ0 = 0 and µ(a ∪ b) = µa + µb whenever a, b ∈ A and a ∩ b = 0; suppose that Af = {a : µa < ∞} is order-dense in A. For e ∈ Af , a, b ∈ A set ρe (a, b) = µ(e ∩ (a 4 b)). Give A the uniformity defined by {ρe : µe < ∞}. (i) Show that the completion b of A under this uniformity has a measure µ A ˆ, extending µ, under which it is a localizable measure algebra. b µ (ii) Show that if a ∈ A, â < ∞ and ² > 0, there is a b ∈ A such that µ ˆ(a 4 b) ≤ ². (iii) Show that for every b there is a sequence han in∈N in A such that a ⊇ sup a∈A inf â = µ ˆ(supn∈N inf m≥n am ). m≥n am and µ n∈N b b (iv) In particular, the set of infima in A of sequences in A is order-dense in A. (v) Explain the relevance of this construction to the embedding A1 ⊗ A2 ⊆ C in 325D. S (c) In 325F, set W = n∈N En × En . Show that if A, B are any non-negligible subsets of [0, 1], then W ∩ (A × B) is not negligible. (d) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1]. Show that A ⊗ A is ccc but not weakly (σ, ∞)-distributive. (Hint: (i) A ⊗ A is embeddable as a subalgebra of a probability algebra (ii) in the notation of 325F, look at cmn = supm≤i≤n ei ⊗ ei .) (e) Repeat 325F-325G and 325Yc-325Yd with an arbitrary atomless probability space in place of [0, 1]. (f ) Let (A, µ ¯) be a probability algebra and hai ii∈I a (stochastically) independent family in A. Show that for any a ∈ A, ² > 0 the set {i : i ∈ I, |¯ µ(a ∩ ai ) − µ ¯a · µ ¯ai | ≥ ²} is finite, so that {i : µ ¯(a ∩ ai ) 6= µ ¯a · µ ¯ai } is countable. (Hint: 272Yd.) 325 Notes and comments 325B shows that the measure algebra of a product measure may be irregular if we have factor measures which are not strictly localizable. But two facts lead the way to the ‘localizable measure algebra free product’ in 325D-325E. The first is that every semi-finite measure algebra is embeddable, in a canonical way, in a localizable measure algebra (322N); and the second is that the Stone representation of a localizable measure algebra is strictly localizable (322M). It is a happy coincidence that we can collapse these two facts together in the construction of 325D. Another way of looking at the localizable measure algebra free product of two localizable measure algebras is to express it as the simple product of measure algebra free products of totally finite measure algebras, using 325Xc and the fact that for σ-finite measure algebras there is only one reasonable measure algebra free product, being that provided by any representation of them as measure algebras of measure spaces (325Eb).

104

Measure algebras

325 Notes

Yet a third way of approaching measure algebra free products is as the uniform space completions of algebraic free products, using 325Yb. This gives the same result as the construction of 325D because the algebraic free product appears as a topologically dense subalgebra of the localizable measure algebra free product (325Dc) which is complete as uniform space (325Dc). (I have to repeat such phrases as ‘topologically dense’ because the algebraic free product is emphatically not order-dense in the measure algebra free product (325F).) The results in 251I on approximating measurable sets for a c.l.d. product measure by combinations of measurable rectangles correspond to general facts about completions of finitely-additive measures (325Yb(ii), 325Yb(iii)). It is worth noting that the completion process can be regarded as made up of two steps; first take infima of sequences of sets of finite measure, and then take arbitrary suprema (325Yb(iv)). The idea of 325F appears S in many guises, and this is only the first time that I shall wish to call on it. The point of the set W = n∈N En × En is that it is a measurable subset of the square (indeed, by taking the En to be open sets we can arrange that W should be open), of measure strictly less than 1 (in fact, as small as we wish), such that its complement does not include any non-negligible ‘measurable rectangle’ G × H; indeed, W ∩ (A × B) is non-negligible for any non-negligible sets A, B ⊆ [0, 1] (325Yc). I believe ¨ s & Oxtoby 55; I learnt the method of 325F that the first published example of such a set was by Erdo from R.O.Davies. I include 325G as a kind of guard-rail. The relationship between preservation of measure and ordercontinuity is a subtle one, as I have already tried to show in 324K, and it is often worth considering the possibility that a result involving order-continuous measure-preserving homomorphisms has a form applying to all order-continuous homomorphisms. However, there is no simple expression of such an idea in the present context. In the context of infinite free products of probability algebras, there is a degree of simplification, since there is only one algebra which can plausibly be called the probability algebra free product, and this is produced by any realization of the algebras as measure algebras of probability spaces (325I-325K). The examples 325F-325G apply equally, of course, to this context. At this point I mention the concept of ‘(stochastically) independent’ family (325L, 325Xe) because we have the machinery to translate several results from §272 into the language of measure algebras (325Xe-325Xg). I feel that I have to use the phrase ‘stochastically independent’ here because there is the much weaker alternative concept of ‘Boolean independence’ (315Xn) also present. But I leave most of this as exercises, because the language of measure algebras offers few ideas to the probability theory already covered in Chapter 27. All it can do is formalise the ever-present principle that negligible sets often can and should be ignored.

326 Additive functionals on Boolean algebras I devote two sections to the general theory of additive functionals on measure algebras. As many readers will rightly be in a hurry to get on to the next two chapters, I remark that the only significant result needed for §§331-332 is the Hahn decomposition of a countably additive functional (326I), and that this is no more than a translation into the language of measure algebras of a theorem already given in Chapter 23. The concept of ‘standard extension’ of a countably additive functional from a subalgebra (327F-327G) will be used for a theorem in §333, and as preparation for Chapter 36. I begin with notes on the space of additive functionals on an arbitrary Boolean algebra (326A-326D), corresponding to 231A-231B, but adding a more general form of the Jordan decomposition of a bounded additive functional into positive and negative parts (326D). The next subsection (326E-326I) deals with countably additive functionals, corresponding to 231C-231F. In 326J-326P I develop a new idea, that of ‘completely additive’ functional, which does not match anything in the previous treatment. In 326Q I return to additive functionals, giving a fundamental result on the construction of additive functionals on free products. 326A Additive functionals: Definition Let A be a Boolean algebra. A functional ν : A → R is finitely additive, or just additive, if ν(a ∪ b) = νa + νb whenever a, b ∈ A and a ∩ b = 0. A non-negative additive functional is sometimes called a finitely additive measure or charge.

326D

Additive functionals on Boolean algebras

105

326B Elementary facts Let A be a Boolean algebra and ν : A → R a finitely additive functional. The following will I hope be obvious. (a) ν0 = 0 (because ν0 = ν0 + ν0). (b) If c ∈ A, then a 7→ ν(a ∩ c) is additive (because (a ∩ c) ∪ (b ∩ c) = (a ∪ b) ∩ c). (c) αν is an additive functional for any α ∈ R. If ν 0 is another finitely additive functional on A, then ν + ν 0 is additive. P (d) If hνi ii∈I is any family of finitely additive functionals such that ν 0 a = i∈I νi a is defined in R for for every a ∈ A, then ν 0 is additive. (e) If B is another Boolean algebra and π : B → A is a Boolean homomorphism, then νπ : B → R is additive. In particular, if B is a subalgebra of A, then ν¹ B : B → R is additive. (f ) ν is non-negative iff it is order-preserving – that is, νa ≥ 0 for every a ∈ A ⇐⇒ νb ≤ νc whenever b ⊆ c (because νc = νb + ν(c \ b) if b ⊆ c). 326C The space of additive functionals Let A be any Boolean algebra. From 326Bc we see that the set M of all finitely additive real-valued functionals on A is a linear space (a linear subspace of RA ). We give it the ordering induced by that of RA , so that ν ≤ ν 0 iff νa ≤ ν 0 a for every a ∈ A. This renders it a partially ordered linear space (because RA is). 326D The Jordan decomposition (I): Proposition Let A be a Boolean algebra, and ν a finitely additive real-valued functional on A. Then the following are equiveridical: (i) ν is bounded; (ii) supn∈N |νan | < ∞ for every dist sequence han in∈N in A; (iii) P limn→∞ |νan | = 0 for every dist sequence han in∈N in A; ∞ (iv) n=0 |νan | < ∞ for every dist sequence han in∈N in A; (v) ν is expressible as the difference of two non-negative additive functionals. proof (a)(i)⇒(v) Assume that ν is bounded. For each a ∈ A, set ν + a = sup{νb : b ⊆ a}. Because ν is bounded, ν + is real-valued. Now ν + is additive. P P If a, b ∈ A and a ∩ b = 0, then ν + (a ∪ b) = sup νc = c ⊆ a∪b

sup

ν(d ∪ e) =

d ⊆ a,e ⊆ b

sup

νd + νe

d ⊆ a,e ⊆ b

(because d ∩ e ⊆ a ∩ b = 0 whenever d ⊆ a, e ⊆ b) = sup νd + sup νe = ν + a + ν + b. Q Q d⊆a

e⊆b

Consequently ν − = ν + − ν is also additive (326Bc). Since 0 = ν0 ≤ ν + a,

νa ≤ ν + a

for every a ∈ A, ν + ≥ 0 and ν − ≥ 0. Thus ν = ν + − ν − is the difference of two non-negative additive functionals. (b)(v)⇒(iv) If ν is expressible as ν1 − ν2 , where ν1 and ν2 are non-negative additive functionals, and han in∈N is dist, then Pn i=0 νj ai = νj (supi≤n ai ) ≤ νj 1

106

Measure algebras

for every n, both j, so that P∞ i=0

|νai | ≤

P∞ i=0

ν1 ai +

P∞ i=0

326D

ν2 ai ≤ ν1 1 + ν2 1 < ∞.

(c)(iv)⇒(iii)⇒(ii) are trivial. (d) not-(i)⇒not-(ii) Suppose that ν is unbounded. Choose sequences han in∈N , hbn in∈N inductively, as follows. b0 = 1. Given that supa ⊆ bn |νa| = ∞, choose cn ⊆ bn such that |νcn | ≥ |νbn | + n; then |νcn | ≥ n and |ν(bn \ cn )| = |νbn − νcn | ≥ |νcn | − |νbn | ≥ n. We have ∞ = sup |νa| = sup |ν(a ∩ cn ) + ν(a \ cn )| a ⊆ bn

a ⊆ bn

≤ sup |ν(a ∩ cn )| + |ν(a \ cn )| ≤ a ⊆ bn

sup a ⊆ bn ∩cn

|νa| +

sup

|νa|,

a ⊆ bn \cn

so at least one of supa ⊆ bn ∩cn |νa|, supa ⊆ bn \cn |νa| must be infinite; take bn+1 to be one of cn , bn \ cn such that supa ⊆ bn+1 |νa| = ∞, and set an = bn \ bn+1 , so that |νan | ≥ n. Continue. On completing the induction, we have a dist sequence han in∈N such that |νan | ≥ n for every n, so that (ii) is false. Remark I hope that this reminds you of the decomposition of a function of bounded variation as the difference of monotonic functions (224D). 326E Countably additive functionals: Definition Let A be a Boolean algebra. A functional ν : P∞ A → R is countably additive or σ-additive if n=0 νan is defined and equal to ν(supn∈N an ) whenever han in∈N is a dist sequence in A and supn∈N an is defined in A. A warning is perhaps in order. It can happen that A is presented to us as a subalgebra of a larger algebra B; for instance, A might be an algebra of sets, a subalgebra of some σ-algebra Σ ⊆ PX. In this case, there may be sequences in A which have a supremum in A which is not a supremum in B (indeed, this will happen just when the embedding is not sequentially order-continuous). So we can have a countably additive functional ν : B → R such that ν¹ A is not countably additive in the sense used here. A similar phenomenon will arise when we come to the Daniell integral in Volume 4 (§434). 326F Elementary facts Let A be a Boolean algebra and ν : A → R a countably additive functional. (a) ν is finitely additive. (Setting an = 0 for every n, we see from the definition in 326E that ν0 = 0. Now, given a ∩ b = 0, set a0 = a, a1 = b, an = 0 for n ≥ 2 to see that ν(a ∪ b) = νa + νb.) (b) If han in∈N is a non-decreasing sequence in A with a supremum a ∈ A, then P∞ νa = νa0 + n=0 ν(an+1 \ an ) = limn→∞ νan . (c) If han in∈N is a non-increasing sequence in A with an infimum a ∈ A, then νa = νa0 − ν(a0 \ a) = νa0 − limn→∞ ν(a0 \ an ) = limn→∞ νan . (d) If c ∈ A, then a 7→ ν(a ∩ c) is countably additive. (For supn∈N an ∩ c = c ∩ supn∈N an whenever the right-hand-side is defined, by 313Ba.) (e) αν is a countably additive functional for any α ∈ R. If ν 0 is another countably additive functional on A, then ν + ν 0 is countably additive. (f ) If B is another Boolean algebra and π : B → A is a sequentially order-continuous Boolean homomorphism, then νπ is a countably additive functional on B. (For if hbn in∈N is a dist sequence in B with supremum b, then hπbn in∈N is a dist sequence with supremum πb.)

326I


107

(g) If A is Dedekind σ-complete and B is a σ-subalgebra of A, then ν¹ B : B → R is countably additive. (For the identity map from B to A is sequentially order-continuous, by 314Hb.) 326G Corollary Let A be a Boolean algebra and ν a finitely additive real-valued functional on A. (a) ν is countably additive iff limn→∞ νan = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0 in A. (b) If ν 0 is an additive functional on A and |ν 0 a| ≤ νa for every a ∈ A, and ν is countably additive, then 0 ν is countably additive. (c) If ν is non-negative, then ν is countably additive iff it is sequentially order-continuous. proof (a)(i) If ν is countably additive and han in∈N is a non-increasing sequence in A with infimum 0, then limn→∞ νan = 0 by 326Fc. (ii) If ν satisfies the condition, and han in∈N is a dist sequence in A with supremum a, set bn = a \ supi≤n ai for each n ∈ N; then hbn in∈N is non-increasing and has infimum 0, so Pn νa − i=0 νai = νa − ν(supi≤n ai ) = νbn → 0 P∞ as n → ∞, and νa = n=0 νan ; thus ν is countably additive. (b) If han in∈N is a dist sequence in A with supremum a, set bn = supi≤n ai for each n; then νa = limn→∞ νbn , so limn→∞ |ν 0 a − ν 0 bn | = limn→∞ |ν 0 (a \ bn )| ≤ limn→∞ ν(a \ bn ) = 0, and

P∞ n=0

ν 0 an = limn→∞ ν 0 bn = ν 0 a.

(c) If ν is countably additive, then it is sequentially order-continuous by 326Fb-326Fc. If ν is sequentially order-continuous, then of course it satisfies the condition of (a), so is countably additive. 326H The Jordan decomposition (II): Proposition Let A be a Boolean algebra and ν a bounded countably additive real-valued functional on A. Then ν is expressible as the difference of two non-negative countably additive functionals. proof Consider the functional ν + a = supb ⊆ a νb defined in the proof of 326D. If han in∈N is a dist sequence in A with supremum a, and b ⊆ a, then P∞ P∞ νb = n=0 ν(b ∩ an ) ≤ n=0 ν + an . P∞ As b is arbitrary, ν + a ≤ n=0 ν + an . But of course Pn ν + a ≥ ν + (supi≤n ai ) = i=0 ν + ai P∞ for every n ∈ N, so ν + a = n=0 ν + an . As han in∈N is arbitrary, ν + is countably additive. Now ν − = ν + − ν is also countably additive, and ν = ν + − ν − is the difference of non-negative countably additive functionals. 326I The Hahn decomposition: Theorem Let A be a Dedekind σ-complete Boolean algebra and ν : A → R a countably additive functional. Then ν is bounded and there is a c ∈ A such that νa ≥ 0 whenever a ⊆ c, while νa ≤ 0 whenever a ∩ c = 0. first proof By 314M, there are a set X and a σ-algebra Σ of subsets of X and a sequentially order-continuous Boolean homomorphism π from Σ onto A. Set ν1 = νπ : Σ → R. Then ν1 is countably additive (326Ff). So ν1 is bounded and there is a set H ∈ Σ such that ν1 F ≥ 0 whenever F ∈ Σ and F ⊆ H and ν1 F ≤ 0 whenever F ∈ Σ and F ∩ H = ∅ (231E). Set c = πH ∈ A. If a ⊆ c, then there is an F ∈ Σ such that πF = a; now π(F ∩ H) = a ∩ c = a, so νa = ν1 (F ∩ H) ≥ 0. If a ∩ c = 0, then there is an F ∈ Σ such that πF = a; now π(F \ H) = a \ c = a, so νa = ν1 (F \ H) ≤ 0. P∞ second proof (a) Note first that ν is bounded. P P If han in∈N is a dist sequence in A, then n=0 νan must exist and be equal to ν(supn∈N an ); in particular, limn→∞ νan = 0. By 326D, ν is bounded. Q Q

108

Measure algebras

326I

(b)(i) We know that γ = sup{νa : a ∈ A} < ∞. Choose a sequence han in∈N in A such that νan ≥ γ −2−n for every n ∈ N. For m ≤ n ∈ N, set bmn = inf m≤i≤n ai . Then νbmn ≥ γ − 2 · 2−m + 2−n for every n ≥ m. P P Induce on n. For n = m, this is due to the choice of am = bmm . For the inductive step, we have bm,n+1 = bmn ∩ an+1 , while surely γ ≥ ν(an+1 ∪ bmn ), so γ + νbm,n+1 ≥ ν(an+1 ∪ bmn ) + ν(an+1 ∩ bmn ) = νan+1 + νbmn ≥ γ − 2−n−1 + γ − 2 · 2−m + 2−n (by the choice of an+1 and the inductive hypothesis) = 2γ − 2 · 2−m + 2−n−1 . Subtracting γ from both sides, νbm,n+1 ≥ γ − 2 · 2−m + 2−n−1 and the induction proceeds. Q Q (ii) Set bm = inf n≥m bmn = inf n≥m an . Then νbm = limn→∞ νbmn ≥ γ − 2 · 2−m , by 326Fc. Next, hbn in∈N is non-decreasing, so setting c = supn∈N bn we have νc = limn→∞ νbn ≥ γ; since νc is surely less than or equal to γ, νc = γ. If b ∈ A and b ⊆ c, then νc − νb = ν(c \ b) ≤ γ = νc, so νb ≥ 0. If b ∈ A and b ∩ c = 0 then νc + νb = ν(c ∪ b) ≤ γ = νc so νb ≤ 0. This completes the proof. 326J Completely additive functionals: Definition Let A be a Boolean algebra. A functional ν : A → R is completely additive or τ -additive if it is finitely additive and inf a∈A |νa| = 0 whenever A is a non-empty downwards-directed set in A with infimum 0. 326K Basic facts Let A be a Boolean algebra and ν a completely additive real-valued functional on A. (a) ν is countably additive. P P If han in∈N is a non-increasing sequence in A with infimum 0, then for any infinite I ⊆ N the set {ai : i ∈ I} is downwards-directed and has infimum 0, so inf i∈I |νai | = 0; which means that limn→∞ νan must be zero. By 326Ga, ν is countably additive. Q Q (b) Let A be a non-empty downwards-directed set in A with infimum 0. Then for every ² > 0 there is an a ∈ A such that |νb| ≤ ² whenever b ⊆ a. P P?? Suppose, if possible, otherwise. Set B = {b : |νb| ≥ ², ∃ a ∈ A, b ⊇ a}. If a ∈ A there is a b0 ⊆ a such that |νb0 | > ². Now {a0 \ b0 : a0 ∈ A, a0 ⊆ a} is downwards-directed and has infimum 0, so there is an a0 ∈ A such that a0 ⊆ a and |ν(a0 \ b0 )| ≤ |νb0 | − ². Set b = b0 ∪ a0 ; then a0 ⊆ b and |νb| = |νb0 + ν(a0 \ b0 )| ≥ |νb0 | − |ν(a0 \ b0 )| ≥ ², so b ∈ B. But also b ⊆ a. Thus every member of A includes some member of B. Since every member of B includes a member of A, B is downwards-directed and has infimum 0; but this is impossible, since inf b∈B |νb| ≥ ². X XQ Q (c) If ν is non-negative, it is order-continuous. P P (i) If A is a non-empty upwards-directed set with supremum a0 , then {a0 \ a : a ∈ A} is a non-empty downwards-directed set with infimum 0, so

326M


109

supa∈A νa = νa0 − inf a∈A ν(a0 \ a) = νa0 . (ii) If A is a non-empty downwards-directed set with infimum a0 , then {a \ a0 : a ∈ A} is a non-empty downwards-directed set with infimum 0, so inf a∈A νa = νa0 + inf a∈A ν(a \ a0 ) = νa0 . Q Q (d) If c ∈ A, then a 7→ ν(a ∩ c) is completely additive. P P If A is a non-empty downwards-directed set with infimum 0, so is {a ∩ c : a ∈ A}, so inf a∈A |ν(a ∩ c)| = 0. Q Q (e) αν is a completely additive functional for any α ∈ R. If ν 0 is another completely additive functional on A, then ν + ν 0 is completely additive. P P We know from 326Bc that ν + ν 0 is additive. Let A be a non-empty downwards-directed set with infimum 0. For any ² > 0, (b) tells us that there are a, a0 ∈ A such that |νb| ≤ ² whenever b ⊆ a0 and |ν 0 b| ≤ ² whenever b ⊆ a0 . But now, because A is downwards-directed, there is a b ∈ A such that b ⊆ a ∩ a0 , which means that |νb + ν 0 b| ≤ |νb| + |ν 0 b| is at most 2². As ² is arbitrary, inf a∈A |(ν + ν 0 )(a)| = 0, and ν + ν 0 is completely additive. Q Q (f ) If B is another Boolean algebra and π : B → A is an order-continuous Boolean homomorphism, then νπ is a completely additive functional on B. P P By 326Be, νπ is additive. If B ⊆ B is a non-empty downwards-directed set with infimum 0 in B, then π[B] is a non-empty downwards-directed set with infimum 0 in A, because π is order-continuous, so inf b∈B |νπb| = 0. Q Q In particular, if B is a regularly embedded subalgebra of A, then ν¹ B is completely additive. (g) If ν 0 is another additive functional on A and |ν 0 a| ≤ νa for every a ∈ A, then ν 0 is completely additive. P P If A ⊆ A is non-empty and downwards-directed and inf A = 0, then inf a∈A |ν 0 a| ≤ inf a∈A νa = 0. Q Q 326L

I squeeze a useful fact in here.

Proposition If A is a ccc Boolean algebra, a functional ν : A → R is countably additive iff it is completely additive. proof If ν is completely additive it is countably additive, by 326Ka. If ν is countably additive and A is a non-empty downwards-directed set in A with infimum 0, then there is a (non-empty) countable subset B of A also with infimum 0 (316E). Let hbn in∈N be a sequence running over B, and choose han in∈N in A such that a0 = b0 , an+1 ⊆ an ∩ bn for every n ∈ N. Then han in∈N is a non-increasing sequence with infimum 0, so limn→∞ νan = 0 (326Fc) and inf a∈A |νa| = 0. As A is arbitrary, ν is completely additive. 326M The Jordan decomposition (III): Proposition Let A be a Boolean algebra and ν a completely additive real-valued functional on A. Then ν is bounded and expressible as the difference of two non-negative completely additive functionals. proof (a) I must first check that ν is bounded. P P Let han in∈N be a dist sequence in A. Set A = {a : a ∈ A, there is an n ∈ N such that ai ⊆ a for every i ≥ n}. Then A is closed under ∩ , and if b is any lower bound for A then b ⊆ 1 \ an ∈ A, so b ∩ an = 0, for every n ∈ N; but this means that 1 \ b ∈ A, so that b ⊆ 1 \ b and b = 0. Thus inf A = 0. By 326Kb, there is an a ∈ A such that |νb| ≤ 1 whenever b ⊆ a. By the definition of A, there must be an n ∈ N such that |νai | ≤ 1 for every i ≥ n. But this means that supn∈N |νan | is finite. As han in∈N is arbitrary, ν is bounded, by 326D(ii). Q Q (b) As in 326D and 326H, set ν + a = supb ⊆ a νb for every a ∈ A. Then ν + is completely additive. P P We know that ν + is additive. If A is a non-empty downwards-directed subset of A with infimum 0, then for every ² > 0 there is an a ∈ A such that |νb| ≤ ² whenever b ⊆ a; in particular, ν + a ≤ ². As ² is arbitrary, inf a∈A ν + a = 0; as A is arbitrary, ν + is completely additive. Q Q − + Consequently ν = ν −ν is completely additive (326Ke) and ν = ν + −ν − is the difference of non-negative completely additive functionals.

110

Measure algebras

326N

326N

I give an alternative definition of ‘completely additive’ which you may feel clarifies the concept.

Proposition Let A be a Boolean algebra, and ν : A → R a function. Then the following are equiveridical: (i) ν is completely additive; P (ii) ν1 = Pi∈I νai whenever hai ii∈I is a partition of unity in A; (iii) νa = i∈I νai whenever hai ii∈I is a dist family in A with supremum a. P proof (For notes on sums i∈I , see 226A.) (a)(i)⇒(ii) P If ν is completely additive and hai ii∈I is a partition of unity in A, then (inducing on #(J)) ν(supi∈J ai ) = i∈J νai for every finite J ⊆ I. Consider A = {1 \ supi∈J ai : J ⊆ I is finite}. Then A is non-empty and downwards-directed and has infimum 0, so for every ² > 0 there is an a ∈ A such that |νb| ≤ ² whenever b ⊆ a. Express a as 1 \ supi∈J ai where J ⊆ I is finite. If now K is another finite subset of I including J, P |ν1 − i∈K ai | = |ν(1 \ supi∈K ai )| ≤ ². P As remarked in 226Ad, this means that ν1 = i∈I νai , as claimed. (b)(ii)⇒(iii) Suppose that ν satisfies the condition (ii), and that hai ii∈I is a dist family with supremum a. Take any j ∈ / I, set J = I ∪ {j} and aj = 1 \ a; then hai ii∈J , (a, 1 \ a) are both partitions of unity, so P P ν(1 \ a) + νa = ν1 = i∈J νai = ν(1 \ a) + i∈I νai , P and νa = i∈I νai . (c)(iii)⇒(i) Suppose that ν satisfies (iii). Then ν is additive. α) ν is bounded. P (α P Let han in∈N be a dist sequence in A. Applying Zorn’s Lemma to the set C of all dist families C ⊆ A including {an : n ∈ N}, we find a partition of unity C ⊇ {an : n ∈ N}. Now P νc is defined in R, so supn∈N |νan | ≤ supc∈C |νc| is finite. By 326D, ν is bounded. Q Q c∈C β ) Define ν + from ν as in 326D. Then ν + satisfies the same condition as ν. P (β P Let hai ii∈I be a dist family in A with supremum a. Then for any b ⊆ a, we have b = supi∈I b ∩ ai , so P P νb = i∈I ν(b ∩ ai ) ≤ i∈I ν + ai . P Thus ν + a ≤ i∈I ν + ai . But of course X i∈I

X ν + ai = sup{ ν + ai : J ⊆ I is finite} i∈J +

= sup{ν (sup ai ) : J ⊆ I is finite} ≤ ν + a, i∈J

+

so ν a =

P i∈I

+

ν ai . Q Q

(γγ ) It follows that ν + is completely additive. P P If A is a non-empty downwards-directed set with infimum 0, then B = {b : ∃ a ∈ A, b ∩ a = 0} is order-dense in A, so there is a partition of unity hbi ii∈I lying in B (313K). Now if J ⊆ I is finite, there is an a ∈ A such that a ∩ supi∈J bi = 0 (because A is downwards-directed), and P ν + a + i∈J ν + bi ≤ ν + 1. P Since ν + 1 = supJ⊆I is finite i∈J ν + bi , inf a∈A ν + a = 0. As A is arbitrary, ν + is completely additive. Q Q (δδ ) Now consider ν − = ν + − ν. Of course P P P ν − a = ν + a − νa = i∈I ν + ai − i∈I νai = i∈I ν − ai whenever hai ii∈I is a dist family in A with supremum a. Because ν − is non-negative, the argument of (γ) shows that ν − = (ν − )+ is completely additive. So ν = ν + − ν − is completely additive, as required.

*326Q


111

326O For completely additive functionals, we have a useful refinement of the Hahn decomposition. I give it in a form adapted to the applications I have in mind. Proposition Let A be a Dedekind σ-complete Boolean algebra and ν : A → R a completely additive functional. Then there is a unique element of A, which I will denote [[ν > 0]], ‘the region where ν > 0’, such that νa > 0 whenever 0 6= a ⊆ [[ν > 0]], while νa ≤ 0 whenever a ∩ [[ν > 0]] = 0. proof Set C1 = {c : c ∈ A \ {0}, νa > 0 whenever 0 6= a ⊆ c}, C2 = {c : c ∈ A, νa ≤ 0 whenever a ⊆ c}. Then C1 ∪ C2 is order-dense in A. P P There is a c0 ∈ A such that νa ≥ 0 for every a ⊆ c, νa ≤ 0 whenever a ∩ c = 0 (326I). Given b ∈ A \ {0}, then b \ c0 ∈ C2 , so if b \ c0 6= 0 we can stop. Otherwise, b ⊆ c0 . If b ∈ C1 we can stop. Otherwise, there is a non-zero c ⊆ b such that νc ≤ 0; but in this case νa ≥ 0, ν(c \ a) ≥ 0 so νa = 0 for every a ⊆ c, and c ∈ C2 . Q Q There is therefore a partition of unity D ⊆ C1 ∪ C2 . Now D ∩ C1 is countable. P P If d ∈ D ∩ C1 , νd > 0. Also #({d : d ∈ D, νd ≥ 2−n }) ≤ 2n supa∈A νa is finite for each n, so D ∩ C1 is the union of a sequence of finite sets, and is countable. Q Q Accordingly D ∩ C1 has a supremum e. If 0 6= a ⊆ e then P P νa = c∈D ν(a ∩ c) = c∈D∩C1 ν(a ∩ c) ≥ 0 by 326N. Also there must be some c ∈ D ∩ C1 such that a ∩ c 6= 0, in which case ν(a ∩ c) > 0, so that νa > 0. If a ∩ e = 0, then P P νa = c∈D ν(a ∩ c) = c∈D∩C2 ν(a ∩ c) ≤ 0. Thus e has the properties demanded of [[ν > 0]]. To see that e is unique, we need observe only that if e0 has the same properties then ν(e \ e0 ) ≤ 0 (because (e \ e0 ) ∩ e0 = 0), so e \ e0 = 0 (because e \ e0 ⊆ e). Similarly, e0 \ e = 0 and e = e0 . Thus we may properly denote e by the formula [[ν > 0]]. 326P Corollary Let A be a Dedekind σ-complete Boolean algebra and µ, ν two completely additive functionals on A. Then there is a unique element of A, which I will denote [[µ > ν]], ‘the region where µ > ν’, such that µa > νa whenever 0 6= a ⊆ [[µ > ν]], µa ≤ νa whenever a ∩ [[µ > ν]] = 0. proof Apply 326O to the functional µ − ν, and set [[µ > ν]] = [[µ − ν > 0]]. *326Q Additive functionals on free products In Volume 4, when we return to the construction of measures on product spaces, the following fundamental fact will be useful. Theorem Let hAi ii∈I be a non-empty family of Boolean algebras, with free product A; write εi : Ai → A for the canonical maps, and C = {inf j∈J εj (aj ) : J ⊆ I is finite, aj ∈ Aj for every j ∈ J}. Suppose that θ : C → R is such that θc = θ(c ∩ εi (a)) + θ(c ∩ εi (1 \ a)) whenever c ∈ C, i ∈ I and a ∈ Ai . Then there is a unique finitely additive functional ν : A → R extending θ. proof (a) It will help if I note at once that θ0 = 0. P P θ0 = θ(0 ∩ εi (0)) + θ(0 ∩ εi (1)) = 2θ0

112

Measure algebras

*326Q

for any i ∈ I. Q Q (b) The key is of course the P following fact: Pnif hcr ir≤m and hds is≤n are two dist families in C with m the same supremum in A, then r=0 θcr = s=0 θds . P P Let J ⊆ I be a finite set and Bi ⊆ Ai a finite subalgebra, for each i ∈ J, such that every cr and every ds belongs to the subalgebra A0 of A generated by {εj (b) : j ∈ J, b ∈ Bj }. Next, if j ∈ J and b ∈ Bj , then Pm Pm Pm r=0 θcr = r=0 θ(cr ∩ εj (b)) + r=0 θ(cr \ εj (b)). We can therefore find a dist family hc0r ir≤m0 in C ∩ A0 such that Pm Pm0 0 supr≤m0 c0r = supr≤m cr , r=0 θcr , r=0 θcr = and whenever r ≤ m0 , j ∈ J and b ∈ Bj then either c0r ⊆ εj (b) or c0r ∩ εj (b)b = 0; that is, every c0r is either 0 or of the form inf j∈J εj (bj ) where bj is an atom of Bj for every j. Similarly, we can find hd0s is≤n0 such that Pn0 Pn 0 sups≤n0 d0s = sups≤n ds , s=0 θds = s=0 θds , and whenever s ≤ n0 , j ∈ J and b ∈ Bj then d0s is either 0 or of the form inf j∈J εj (bj ) where bj is an atom of Bj for every j. But we now have supr≤m0 c0r = sups≤n0 d0s while for any r ≤ m0 , s ≤ n0 either c0r = d0s or c0r ∩ d0s = 0. It follows that the non-zero in the finite sequence hc0r ir≤m0 are just a rearrangement of the non-zero in hd0s is≤n0 , so that Pn0 Pn Pm Pm0 0 0 s=0 θds , s=0 θds = r=0 θcr = r=0 θcr = as required. Q Q Pm (c) By 315Jb, this means that we have a functional ν : A → R such that ν(supr≤m cr ) = r=0 θcr whenever hcr ir≤m is a dist family in C. It is now elementary to check that ν is additive, and it is clearly the only additive functional on A extending θ. 326X Basic exercises (a) Let A be a Boolean algebra and ν : A → R a finitely additive functional. Show that (i) ν(a ∪ b) = νa + νb − ν(a ∩ b) (ii) ν(a ∪ b ∪ c) = νa + νb + νc − ν(a ∩ b) − ν(a ∩ c) − ν(b ∩ c) + ν(a ∩ b ∩ c) for all a, b, c ∈ A. Generalize these results to longer sequences in A. (b) Let A be a Boolean algebra and ν : A → R a finitely additive functional. Show that the following are equiveridical: (i) ν is countably additive; (ii) limn→∞ νan = νa whenever han in∈N is a non-decreasing sequence in A with supremum a. (c) Let A be a Dedekind σ-complete Boolean algebra and ν : A → R a finitely additive functional. Show that the following are equiveridical: (i) ν is countably additive; (ii) limn→∞ νan = 0 whenever han in∈N is a sequence in A and inf n∈N supm≥n am = 0; (iii) limn→∞ νan = νa whenever han in∈N is a sequence in A and a = inf n∈N supm≥n am = supn∈N inf m≥n am . (Hint: for (i)⇒(iii), consider non-negative ν first.) (d) Let X be any uncountable set, and J an infinite subset of X. Let A be the finite-cofinite algebra of X (316Yk), and for a ∈ A set νa = #(a ∩ J) if a is finite, −#(J \ a) if a is cofinite. Show that ν is countably additive and unbounded. > (e) Let A be the algebra of subsets of [0, 1] generated by the family of (closed) intervals. Show that there is a unique additive functional ν : A → R such that ν[α, β] = β − α whenever 0 ≤ α ≤ β ≤ 1. Show that ν is countably additive but not completely additive. (f ) (i) Let (X, Σ, µ) be any atomless probability space. Show that µ : Σ → R is a countably additive functional which is not completely additive. (ii) Let X be any uncountable set and µ the countablecocountable measure on X (211R). Show that µ is countably additive but not completely additive. P (g) Let A be a Boolean algebra and ν : A → R a function. (i) Show that ν is finitely additive P iff i∈I νai = ν1 for every finite partition of unity hai ii∈I . (ii) Show that ν is countably additive iff i∈I νai = ν1 for every countable partition of unity hai ii∈I .

326Yj


113

(h) Show that 326O can fail if ν is only countably additive, rather than completely additive. (Hint: 326Xf.) (i) Let A be a Boolean algebra and ν a finitely additive real-valued functional on A. Let us say that a ∈ A is a of ν if (α) νb = 0 whenever b ∩ a = 0 (β) for every non-zero b ⊆ a there is a c ⊆ b such that νc 6= 0. (i) Check that ν can have at most one . (ii) Show that if a is a for ν and ν is bounded, then the principal ideal Aa generated by a is ccc. (iii) Show that if A is Dedekind σ-complete and ν is countably additive, then ν is completely additive iff it has a , and that in the language of 326O this is [[ν > 0]] ∪ [[−ν > 0]]. (iv) Taking J = X in 326Xd, show that X is the of the functional ν there. 326Y Further exercises (a) Let A be a Boolean algebra and ν a non-negative finitely additive functional on A. Show that the following are equiveridical: (i) for every ² > 0 there is a finite partition hai ii∈I of unity in A such that νai ≤ ² for every i ∈ I; (ii) whenever ν 0 is a non-zero finitely additive functional such that 0 ≤ ν 0 ≤ ν there is an a ∈ A such that ν 0 a and ν 0 (1 \ a) are both non-zero. (Such functionals are called atomless.) (b) Let A be a Boolean algebra and ν1 , ν2 atomless non-negative additive functionals on A. Show that ν1 + ν2 , αν1 are atomless for every α ≥ 0, and that ν is atomless whenever ν is additive and 0 ≤ ν ≤ ν1 . (c) Let A be a Dedekind σ-complete Boolean algebra and ν an atomless non-negative finitely additive functional on A. Show that there is a family hat it∈[0,1] in A such that as ⊆ at and νat = tν1 whenever s ≤ t ∈ [0, 1]. (d) Let A be a Dedekind σ-complete Boolean algebra and ν0 , . . . , νn atomless non-negative finitely additive functionals on A. Show that there is an a ∈ A such that νi a = 12 νi 1 for every i ≤ n. (Hint: it is enough to consider the case ν0 ≥ ν1 . . . ≥ νn . For the inductive step, use the inductive hypothesis to construct hat it∈[0,1] such that as ⊆ at , νi at = tνi 1 if i < n, 0 ≤ s ≤ t ≤ 1. Now show that t 7→ νn (at+ 12 \ at ) is continuous on [0, 21 ].) (e) Let A be a Dedekind σ-complete Boolean algebra. Let ν : A → Rr , where r ≥ 1, be additive (in the sense that ν(a ∪ b) = νa + νb whenever a ∩ b = 0) and atomless (in the sense that for every ² > 0 there is a finite partition of unity hai ii∈I such that kνak ≤ ² whenever i ∈ I and a ⊆ ai ). Show that {νa : a ∈ A} is a convex set in Rr . (This is a version of Liapounoff ’s theorem. I am grateful to K.P.S.Bhaskara Rao for showing it to me.) (f ) Let A be a Dedekind σ-complete Boolean algebra and ν : A → [0, ∞[ a countably additive functional. Show that ν is atomless iff whenever a ∈ A and νa 6= 0 there is a b ⊆ a such that 0 < νb < νa. (g) Show that there is a finitely additive functional ν : PN → R such that ν{n} = 1 for every n ∈ N, so that ν is not bounded. (Hint: Use Zorn’s Lemma to construct a maximal linearly independent subset of `∞ including {χ{n} : n ∈ N}, and hence to construct a linear map f : `∞ → R such that f (χ{n}) = 1 for every n.) (h) Let A be any infinite Boolean algebra. Show that there is an unbounded finitely additive functional ν : A → R. (Hint: let htn in∈N be a sequence of distinct points in the Stone space of A, and set νa = ν 0 {n : tn ∈ b a} for a suitable ν 0 .) (i) Let A be a Boolean algebra, and give RA its product topology. Show that the space of finitely additive functionals on A is a closed subset of RA , but that the space of bounded finitely additive functionals is closed only when A is finite. (j) Let A be a Boolean algebra, and M the linear space of all bounded finitely additive real-valued functionals on A. For ν, ν 0 ∈ M say that ν ≤ ν 0 if νa ≤ ν 0 a for every a ∈ A. Show that (i) ν + , as defined in the proof of 326D, is just sup{0, ν} in M ; (ii) M is a Dedekind complete Riesz space (241E-241F, 353G);

114

Measure algebras

326Yj

(iii) for ν, ν 0 ∈ M , |ν| = ν ∨ (−ν), ν ∨ ν 0 , ν ∧ ν 0 are given by the formulae |ν|(a) = supb ⊆ a νb − ν(a \ b),

(ν ∨ ν 0 )(a) = supb ⊆ a νb + ν 0 (a \ b),

(ν ∧ ν 0 )(a) = inf b ⊆ a νb + ν 0 (a \ b); (iv) for any non-empty A ⊆ M , A is bounded above in M iff Pn sup{ i=0 νi ai : νi ∈ A for each i ≤ n, hai ii≤n is dist} is finite, and then sup A is defined by the formula Pn (sup A)(a) = sup{ i=0 νi ai : νi ∈ A for each i ≤ n, hai ii≤n is dist, supi≤n ai = a} for every a ∈ A; (v) setting kνk = |ν|(1), k k is an order-continuous norm on M under which M is a Banach lattice. (k) Let A be a Boolean algebra. A functional ν : A → C is finitely additive if its real and imaginary parts are. Show that P the space of bounded finitely additive functionals from A to C is a Banach space under n the norm kνk = sup{ i=0 |νai | : hai ii≤n is a partition of unity in A}. (l) Let A be a Boolean algebra, and give it the topology Tσ for which the closed sets are the sequentially order-closed sets. Show that a finitely additive functional ν : A → R is countably additive iff it is continuous for Tσ . (m) Let A be a Boolean algebra, and Mσ the set of all bounded countably additive real-valued functionals on A. Show that Mσ is a closed and order-closed linear subspace of the normed space M of all additive functionals on A (326Yj), and that |ν| ∈ Mσ whenever ν ∈ Mσ . (n) Let A be a Boolean algebra and ν a non-negative finitely additive functional on A. Set νσ a = inf{supn∈N νan : han in∈N is a non-decreasing sequence with supremum a} for every a ∈ A. Show that νσ is countably additive, and is sup{ν 0 : ν 0 ≤ ν is countably additive}. (o) Let A be a Dedekind σ-complete Boolean algebra and hνn in∈N a sequence of countably additive realvalued functionals on A such that νa = limn→∞ νn a is defined in R for every a ∈ A. Show that ν is countably additive. (Hint: use arguments from part (a) of the proof of 247C to see that limn→∞ supk∈N |νk an | = 0 for every dist sequence han in∈N in A, and therefore that limn→∞ supk∈N |νk an | = 0 whenever han in∈N is a non-increasing sequence with infimum 0.) (p) Let A be a Boolean algebra, and Mτ the set of all completely additive real-valued functionals on A. Show that Mτ is a closed and order-closed linear subspace of the normed space M of all additive functionals, and that |ν| ∈ Mτ whenever ν ∈ Mτ . (q) Let A be a Boolean algebra and ν a non-negative finitely additive functional on A. Set ντ b = inf{supa∈A νa : A is a non-empty upwards-directed set with supremum b} for every b ∈ A. Show that ντ is completely additive, and is sup{ν 0 : ν 0 ≤ ν is completely additive}. (r) Let A be a Boolean algebra, and give it the topology T for which the closed sets are the order-closed sets (313Xb). Show that a finitely additive functional ν : A → R is completely additive iff it is continuous for T. (s) Let X be a set, Σ any σ-algebra of subsets of X, and ν :P Σ → R a functional. ShowP that ν is completely ∞ ∞ additive iff there are sequences hxn in∈N , hαn in∈N such that n=0 |αn | < ∞ and νE = n=0 αn χE(xn ) for every E ∈ Σ. (t) Let A and B be Boolean algebras and µ, ν finitely additive functionals on A, B respectively. Show that there is a unique finitely additive functional λ : A ⊗ B → R such that λ(a ⊗ b) = µa · νb for all a ∈ A, b ∈ B.

326 Notes


115

N (u) Let hAi ii∈I be a family of Boolean algebras, with free product ( i∈I Ai , hεi ii∈I ), and for each i ∈ I let νi be a finitely N additive functional on Ai such that νi 1Q= 1. Show that there is a unique finitely additive functional ν : i∈I Ai → R such that ν(inf i∈J εi (ai )) = i∈J νi ai whenever J ⊆ I is non-empty and finite and ai ∈ Ai for each i ∈ J. 326 Notes and comments I have not mentioned the phrase ‘measure algebra’ anywhere in this section, and in principle this material could have been part of Chapter 31; but countably additive functionals are kissing cousins of measures, and most of the ideas here surely belong to ‘measure theory’ rather than to ‘Boolean algebra’, in so far as such divisions are meaningful at all. I have given as much as possible of the theory in a general form because the simplifications which are possible when we look only at measure algebras are seriously confusing if they are allowed too much prominence. In particular, it is important to understand that the principal properties of completely additive functionals do not depend on Dedekind completeness of the algebra, provided we take care over the definitions. Similarly, the definition of ‘countably additive’ functional for algebras which are not Dedekind σ-complete needs a moment’s attention to the phrase ‘and supn∈N an is defined in A’. It can happen that a functional is countably additive mostly because there are too few such sequences (326Xd). The formulations I have chosen as principal definitions (326A, 326E, 326J) are those which I find closest to my own intuitions of the concepts, but you may feel that 326G(i), 326Xc(iii) and 326N, or 326Yl and 326Yr, provide useful alternative patterns. The point is that countable additivity corresponds to sequential order-continuity (326Fb, 326Fc, 326Ff), while complete additivity corresponds to order-continuity (326Kc, 326Kf); the difficulty is that we must consider functionals which are not order-preserving, so that the simple definitions in 313H cannot be applied directly. It is fair to say that all the additive functionals ν we need to understand are bounded, and therefore may be studied in of their positive and negative parts ν + , ν − , which are order-preserving (326Bf); but many of the most important applications of these ideas depend precisely on using facts about ν to deduce facts about ν + and ν − . It is in 326D that we seem to start getting more out of the theory than we have put in. The ideas here have vast ramifications. What it amounts to is that we can discover much more than we might expect by looking at dist sequences. To begin with, the conditions here lead directly to 326I and 326M: every completely additive functional is bounded, and every countably additive functional on a Dedekind σ-complete Boolean algebra is bounded. (But note 326Yg-326Yh.) Naturally enough, the theory of countably additive functionals on general Boolean algebras corresponds closely to the special case of countably additive functionals on σ-algebras of sets, already treated in §§231-232 for the sake of the Radon-Nikod´ ym theorem. This should make 326E-326I very straightforward. When we come to completely additive functionals, however, there is room for many surprises. The natural map from a σ-algebra of measurable sets to the corresponding measure algebra is sequentially order-continuous but rarely order-continuous, so that there can be completely additive functionals on the measure algebra which do not correspond to completely additive functionals on the σ-algebra. Indeed there are very few completely additive functionals on σ-algebras of sets (326Ys). Of course these surprises can arise only when there is a difference between completely additive and countably additive functionals, that is, when the algebra involved is not ccc (326L). But I think that neither 326M nor 326N is obvious. I find myself generally using the phrase ‘countably additive’ in preference to ‘completely additive’ in the context of ccc algebras, where there is no difference between them. This is an attempt at -friendliness; the phrase ‘countably additive’ is the commoner one in ordinary use. But I must say that my personal inclination is to the other side. The reason why so many theorems apply to countably additive functionals in these contexts is just that they are completely additive. I have given two proofs of 326I. I certainly assume that if you have got this far you are acquainted with the Radon-Nikod´ ym theorem and the associated basic facts about countably additive functionals on σ-algebras of sets; so that the ‘first proof’ should be easy and natural. On the other hand, there are purist objections on two fronts. First, it relies on the Stone representation, which involves a much stronger form of the axiom of choice than is actually necessary. Second, the classical Hahn decomposition in 231E is evidently a special case of 326I, and if we need both (as we certainly do) then one expects the ideas to stand out more clearly if they are applied directly to the general case. In fact the two versions of the argument are so nearly identical that (as you will observe, if you have Volume 2 to hand) they can share nearly every word. You can take the ‘second proof’, therefore, as a worked example in the translation of ideas from the context of σ-algebras

116

Measure algebras

326 Notes

of sets to the context of Dedekind σ-complete Boolean algebras. What makes it possible is the fact that the only limit operations referred to involve countable families. Arguments not involving limit operations can generally, of course, be applied to all Boolean algebras; I have lifted some exercises (326Yj, 326Yn) from §231 to give you some practice in such generalizations. Almost any non-trivial measure provides an example of a countably additive functional on a Dedekind σ-complete algebra which is not completely additive (326Xf). The question of whether such a functional can exist on a Dedekind complete algebra is the ‘Banach-Ulam problem’, to which I will return in 363S. In this section I have looked only at questions which can be adequately treated in of the underlying algebras A, without using any auxiliary structure. To go much farther we shall need to study the ‘function spaces’ S(A) and L∞ (A) of Chapter 36. In particular, the ideas of 326Yg, 326Yj-326Yk and 326Ym-326Yq will make better sense when redeveloped in §362.

327 Additive functionals on measure algebras When we turn to measure algebras, we have a simplification, relative to the general context of §326, because the algebras are always Dedekind σ-complete; but there are also elaborations, because we can ask how the additive functionals we examine are related to the measure. In 327A-327C I work through the relationships between the concepts of ‘absolute continuity’, ‘(true) continuity’ and ‘countable additivity’, following 232A-232B, and adding ‘complete additivity’ from §326. These ideas provide a new interpretation of the Radon-Nikod´ ym theorem (327D). I then use this theorem to develop some machinery (the ‘standard extension’ of an additive functional from a closed subalgebra to the whole algebra, 327F-327G) which will be used in §333. 327A

I start with the following definition and theorem corresponding to 232A-232B.

Definition Let (A, µ ¯) be a measure algebra and ν : A → R a finitely additive functional. Then ν is absolutely continuous with respect to µ ¯ if for every ² > 0 there is a δ > 0 such that |νa| ≤ ² whenever µ ¯a ≤ δ. 327B Theorem Let (A, µ ¯) be a measure algebra, and ν : A → R a finitely additive functional. Give A its measure-algebra topology and uniformity (§323). (a) If ν is continuous, it is completely additive. (b) If ν is countably additive, it is absolutely continuous with respect to µ ¯. (c) The following are equiveridical: (i) ν is continuous at 0; (ii) ν is countably additive and whenever a ∈ A and νa 6= 0 there is a b ∈ A such that µ ¯b < ∞ and ν(a ∩ b) 6= 0; (iii) ν is continuous everywhere on A; (iv) ν is uniformly continuous. (d) If (A, µ ¯) is semi-finite, then ν is continuous iff it is completely additive. (e) If (A, µ ¯) is σ-finite, then ν is continuous iff it is countably additive iff it is completely additive. (f) If (A, µ ¯) is totally finite, then ν is continuous iff it is absolutely continuous with respect to µ ¯ iff it is countably additive iff it is completely additive. proof (a) If ν is continuous, and A ⊆ A is non-empty, downwards-directed and has infimum 0, then 0 ∈ A (323D(b-ii)), so inf a∈A |νa| = 0. (b) ?? Suppose, if possible, that ν is countably additive but not absolutely continuous. Then there is an ² > 0 such that for every δ > 0 there is an a ∈ A such that µ ¯a ≤ δ but |νa| ≥ ². For each n ∈ N we may choose a bn ∈ A such that µ ¯bn ≤ 2−n and |νbn | ≥ ². Consider b∗n = supk≥n bk , b = inf n∈N b∗n . Then we have P∞ µ ¯b ≤ inf n∈N µ ¯(supk≥n bk ) ≤ inf n∈N k=n 2−k = 0, so µ ¯b = 0 and b = 0. On the other hand, ν is expressible as a difference ν + − ν − of non-negative countably additive functionals (326H), each of which is sequentially order-continuous (326Gc), and

327C

Additive functionals on measure algebras

117

0 = limn→∞ (ν + + ν − )b∗n ≥ inf n∈N (ν + + ν − )bn ≥ inf n∈N |νbn | ≥ ², which is absurd. X X (c)(i)⇒(ii) Suppose that ν is continuous. Then it is completely additive, by (a), therefore countably additive. If νa 6= 0, there must be an b of finite measure such that |νd| < |νa| whenever d ∩ b = ∅, so that |ν(a \ b)| < |νa| and ν(a ∩ b) 6= 0. Thus the conditions are satisfied. (ii)⇒(iv) Now suppose that ν satisfies the two conditions in (ii). Because A is Dedekind σ-complete, ν must be bounded (326I), therefore expressible as the difference ν + − ν − of countably additive functionals. Set ν1 = ν + + ν − . Set γ = sup{ν1 b : b ∈ A, µ ¯b < ∞}, and choose a sequence hbn in∈N of elements of A of finite measure such that limn→∞ ν1 bn = γ; set b∗ = supn∈N bn . If d ∈ A and d ∩ b∗ = ∅ then νd = 0. P P If b ∈ A and µ ¯b < ∞, then |ν(d ∩ b)| ≤ ν1 (d ∩ b) ≤ ν1 (b \ bn ) = ν1 (b ∪ bn ) − ν1 bn ≤ γ − ν1 bn for every n ∈ N, so ν(d ∩ b) = 0. As b is arbitrary, the second condition tells us that νd = 0. Q Q Setting b∗n = supk≤n bk for each n, we have limn→∞ ν1 (b∗ \ b∗n ) = 0. Take any ² > 0, and (using (b) above) let δ > 0 be such that |νa| ≤ ² whenever µ ¯a ≤ δ. Let n be such that ν1 (b∗ \ b∗n ) ≤ ². Then |νa| ≤ |ν(a ∩ b∗n )| + |ν(a ∩ (b∗ \ b∗n ))| + |ν(a \ b∗ )| ≤ |ν(a ∩ b∗n )| + ν1 (b∗ \ b∗n ) ≤ |ν(a ∩ b∗n )| + ² for any a ∈ A. Now if b, c ∈ A and µ ¯((b 4 c) ∩ b∗n ) ≤ δ then |νb − νc| ≤ |ν(b \ c)| + |ν(c \ b)| ≤ |ν((b \ c) ∩ b∗ )| + |ν((c \ b) ∩ b∗ )| + 2² ≤ ² + ² + 2² = 4² because µ ¯((b \ c) ∩ b∗n ), µ ¯((c \ b) ∩ b∗n ) are both less than or equal to δ. As ² is arbitrary, ν is uniformly continuous. (iv)⇒(iii)⇒(i) are trivial. (d) One implication is covered by (a). For the other, suppose that ν is completely additive. Then it is countably additive. On the other hand, if νa 6= 0, consider B = {b : b ⊆ a, µ ¯b < ∞}. Then B is upwardsdirected and sup B = a, because µ ¯ is semi-finite (322Eb), so {a \ b : b ∈ B} is downwards-directed and has infimum 0. Accordingly inf b∈B |ν(a \ b)| = 0, and there must be a b ∈ B such that νb 6= 0. But this means that condition (ii) of (c) is satisfied, so that ν is continuous. (e) Now suppose that (A, µ ¯) is σ-finite. In this case A is ccc (322G) so complete additivity and countable additivity are the same (326L) and we have a special case of (d). (f ) Finally, suppose that µ ¯1 < ∞ and that ν is absolutely continuous with respect to µ ¯. If A ⊆ A is non-empty and downwards-directed and has infimum 0, then inf a∈A µ ¯a = 0 (321F), so inf a∈A |νa| must be 0; thus ν is completely additive. With (b) and (e) this shows that all four conditions are equiveridical. 327C Proposition Let (X, Σ, µ) be a measure space and (A, µ ¯) its measure algebra. (a) There is a one-to-one correspondence between finitely additive functionals ν¯ on A and finitely additive functionals ν on Σ such that νE = 0 whenever µE = 0, given by the formula ν¯E • = νE for every E ∈ Σ. (b) In (a), ν¯ is absolutely continuous with respect to µ ¯ iff ν is absolutely continuous with respect to µ. (c) In (a), ν¯ is countably additive iff ν is countably additive; so that we have a one-to-one correspondence between the countably additive functionals on A and the absolutely continuous countably additive functionals on Σ. (d) In (a), ν¯ is continuous for the measure-algebra topology on A iff ν is truly continuous in the sense of 232Ab. (e) Suppose that µ is semi-finite. Then, in (a), ν¯ is completely additive iff ν is truly continuous.

118

Measure algebras

327C

proof (a) This should be nearly obvious. If ν¯ : A → R is additive, then the formula defines a functional ν : Σ → R which is additive by 326Be. Also, of course, µE = 0 =⇒ E • = 0 =⇒ νE = 0. On the other hand, if ν is an additive functional on Σ which is zero on negligible sets, then, for E, F ∈ Σ, E • = F • =⇒ µ(E \ F ) = µ(F \ E) = 0 =⇒ ν(E \ F ) = ν(F \ E) = 0 =⇒ νF = νE − ν(E \ F ) + ν(F \ E) = νE, so we have a function ν¯ : A → R defined by the given formula. If E, F ∈ Σ and E • ∩ F • = 0, then ν¯(E • ∪ F • ) = ν¯(E ∪ F )• = ν(E ∪ F ) = ν(E \ F ) + νF = ν¯E • + ν¯F • because (E \ F )• = E • \ F • = E • . Thus ν¯ is additive, and the correspondence is complete. (b) This is immediate from the definitions. (c)(i) If ν is countably additive, and han in∈N is a dist sequence in A, we can express it as hEn in∈N S where hEn in∈N is a sequence in Σ. Setting Fn = En \ i
S n∈N

Fn ) =

P∞ n=0

νFn =

P∞ n=0

ν¯an .

As han in∈N is arbitrary, ν¯ is countably additive. (ii) If ν¯ is countably additive, then ν is countably additive by 326Ff. (iii) For the last remark, note that by 232Ba a countably additive functional on Σ is absolutely continuous with respect to µ iff it is zero on the µ-negligible sets. (d) The definition of ‘truly continuous’ functional translates directly to continuity at 0 in the measure algebra. But by 327Bc this is the same thing as continuity. (e) Put (d) and 327Bd together. 327D The Radon-Nikod´ ym theorem We are now ready for another look at this theorem. Theorem Let (X, Σ, µ) be a semi-finite measure space, with measure algebra (A, µ ¯). Let L1 be the space of equivalence classes of real-valued integrable functions on X (§242), and write Mτ for the set of completely additive real-valued functionals on A. Then there is an ordered linear space bijection between Mτ and L1 defined by saying that ν¯ ∈ Mτ corresponds to u ∈ L1 if ν¯a =

R

E

f whenever a = E • in A and f • = u in L1 .

• proof (a) Given ν¯ ∈ Mτ , we have a truly continuous ν : Σ → R given R by setting νE = νE for every E ∈ Σ (327Ce). Now there is an integrable function f such that νE = E f for every E ∈ Σ (232E). There is likely to be more than one such function, but any two must be equal almost everywhere (232Hd), so the corresponding equivalence class uν¯ = f • is uniquely defined.

(b) Conversely, given u ∈ L1 , we have a well-defined functional νu on Σ given by setting νu E =

R

u= E

R

E

f whenever f • = u

for every E ∈ Σ (242Ac). By 232E, νu is additive and truly continuous, and of course it is zero when µ is zero, so corresponds to a completely additive functional ν¯u on A (327Ce). (c) Clearly the maps u 7→ ν¯u and ν¯ 7→ uν¯ are now the two halves of a one-to-one correspondence. To see that it is linear, we need note only that

327F


(¯ νu + ν¯v )E • = ν¯u E • + ν¯v E • =

R E

u+

R E

v=

R E

119

u + v = ν¯u+v E •

for every E ∈ Σ, so ν¯u + ν¯v = ν¯u+v for all u, v ∈ L1 ; and similarly ν¯αu = α¯ νu for u ∈ L1 , α ∈ R. As for the 1 • • ordering, given u and v ∈ L , take integrable f , g such that u = f , v = g ; then ν¯u ≤ ν¯v ⇐⇒ ν¯u E • ≤ ν¯v E • for every E ∈ Σ Z Z ⇐⇒ u≤ v for every E ∈ Σ E E Z Z ⇐⇒ f≤ g for every E ∈ Σ E

E

⇐⇒ f ≤ g a.e. ⇐⇒ u ≤ v, using 131Ha. 327E

I slip in an elementary fact.

Proposition If (A, µ ¯) is a measure algebra, then the functional a 7→ µc a = µ ¯(a ∩ c) is completely additive whenever c ∈ A and µ ¯c < ∞. proof µc is additive because µ ¯ is additive, and by 321F inf a∈A µc a = 0 whenever A is non-empty, downwardsdirected and has infimum 0. 327F Standard extensions The machinery of 327D provides the basis of a canonical method for extending countably additive functionals from closed subalgebras, which we shall need in §333. Lemma Let (A, µ ¯) be a totally finite measure algebra and C ⊆ A a closed subalgebra. Write Mσ (A), Mσ (C) for the spaces of countably additive real-valued functionals on A, C respectively. (a) There is an operator R : Mσ (C) → Mσ (A) defined by saying that, for every ν ∈ Mσ (C), Rν is the unique member of Mσ (A) such that [[Rν > α¯ µ]] = [[ν > α¯ µ¹ C]] for every α ∈ R. (b)(i) Rν extends ν for every ν ∈ Mσ (C). (ii) R is linear and order-preserving. (iii) R(¯ µ¹ C) = µ ¯. P∞ ¯c for every (iv) If hν i Pn∞n∈N is a sequence of non-negative functionals in Mσ (C) such that n=0 νn c = µ ¯a for every a ∈ A. c ∈ C, then n=0 (Rνn )(a) = µ Remarks When saying that C is ‘closed’, I mean, indifferently, ‘topologically closed’ or ‘order-closed’; see 323H-323I. For the notation ‘[[ν > α¯ µ]]’ see 326O-326P. proof (a)(i) By 321J-321K, we may represent (A, µ ¯) as the measure algebra of a measure space (X, Σ, µ); write π for the canonical map from Σ to A. Write T for {E : E ∈ Σ, πE ∈ C}. Because C is a σ-subalgebra of C and π is a sequentially order-continuous Boolean homomorphism, T is a σ-subalgebra of Σ. (ii) For each ν ∈ Mσ (C), νπ : T → R is countably additive and zero on {F : F ∈ T, µF = 0}, so we R can choose a T-measurable function fν : X → R such that F fν d(µ¹ T) = νπF for every F ∈ T. Of course we can now think of fν as a µ-integrable function (233B), so we get a corresponding countably additive R functional Rν : A → R defined by setting (Rν)(πE) = E fν for every E ∈ Σ (327D). (In this context, of course, countably additive functionals are completely additive, by 327Bf.) For α ∈ R, set Hα = {x : fν (x) > α} ∈ T. Then for any E ∈ Σ, E ⊆ Hα , µE > 0 =⇒ E ∩ Hα = ∅ =⇒

R

R E

E

fν > αµE,

fν ≤ αµE.

Translating into of elements of A, and setting cα = πHα ∈ C, we have 0 6= a ⊆ cα =⇒ (Rν)(a) > α¯ µa, a ∩ cα = 0 =⇒ (Rν)(a) ≤ α¯ µa.

120

Measure algebras

327F

So [[Rν > α¯ µ]] = cα ∈ C. Of course we now have νc = (Rν)(c) > α¯ µc when c ∈ C, 0 6= c ⊆ cα , νc ≤ α¯ µc when c ∈ C, c ∩ cα = 0, so that cα is also equal to [[ν > µ ¯¹ C]]. Thus the functional Rν satisfies the declared formula. (iii) To see that Rν is uniquely defined, observe that if λ ∈ MσR(A) and [[λ > α¯ µ]] = [[Rν > α¯ µ]] for every α, then there is a Σ-measurable function g : X → R such that E g dµ = λπE for every E ∈ Σ; but in this case (just as in (ii)) [[λ > α¯ µ]] = πGα , where Gα = {x : g(x) > α}, for each α. So we must have πGα = πHα , that is, µ(Gα 4Hα ) = 0, for every α. Accordingly S {x : fν (x) 6= g(x)} = q∈Q Gq 4Hq R R is negligible; fν = g a.e., E fν dµ = E g dµ for every E ∈ Σ and λ = Rν. (b)(i) If ν ∈ Mσ (C), (Rν)(πF ) =

R F

fν dµ =

R F

fν d(µ¹ T) = νπF

for every F ∈ T, so Rν extends ν. (ii) If ν = ν1 + ν2 , we must have

R

F

fν = νπF = ν1 πF + ν2 πF =

R F

fν1 +

R F

fν2 =

R F

fν1 + fν2

for every F ∈ T, so fν = fν1 + fν2 a.e., and we can repeat the formulae (Rν)(πE) =

R

f = E ν

R

f + fν2 = E ν1

R

f + E ν1

R

E

fν2 = (Rν1 )(πE) + (Rν2 )(πE),

in a different order, for every E ∈ Σ, to see that Rν = Rν1 + Rν2 . Similarly, if ν ∈ Mσ (C) and γ ∈ R, fγν = γfν a.e. and R(γν) = γRν. If ν1 ≤ ν2 in Mσ (C), then

R

F

fν1 = ν1 πF ≤ ν2 πF =

R

F

fν2

for every F ∈ T, so fν1 ≤ fν2 a.e. (131Ha), and Rν1 ≤ Rν2 . Thus R is linear and order-preserving. (iii) If ν = µ ¯¹ C then

R F

fν = νπF = µF =

R F

1

for every F ∈ T, so fν = 1 a.e. and Rν = µ ¯. (iv) Now suppose that hνP n in∈N is a sequence in Mσ (C) such that, for every c ∈ C, νn c ≥ 0 for every n P ∞ n ¯c. Set gn = i=0 fνi for each n; then 0 ≤ gn ≤ gn+1 ≤ 1 a.e. for every n, and n=0 νn c = µ R Pn limn→∞ gn = limn→∞ i=0 νi 1 = µ ¯1. R R But this means that, setting g = limn→∞ gn , g ≤ 1 a.e. and g = 1, so that g = 1 a.e. and R P∞ n=0 (Rνi )(πE) = limn→∞ E gn = µE P∞ for every E ∈ Σ, so that n=0 (Rνi )(a) = µ ¯a for every a ∈ A. and

327G Definition In the context of 327F, I will call Rν the standard extension of ν to A. Remark The point of my insistence on the uniqueness of R, and on the formula in 327Fa, is that Rν really is defined by the abstract structure (A, µ ¯, C, ν), even though I have used a proof which runs through the representation of (A, µ ¯) as the measure algebra of a measure space (X, Σ, µ). 327X Basic exercises (a) Let (A, µ ¯) and (B, µ ¯0 ) be totally finite measure algebras, and π : A → B a measure-preserving Boolean homomorphism. Let C be a closed subalgebra of A, and ν a countably additive functional on the closed subalgebra π[C] of B. (i) Show that νπ is a countably additive functional on C. (ii) Show that if ν˜ is the standard extension of ν to B, then ν˜π is the standard extension of νπ to A.

327 Notes


121

(b) Let (X, Σ, µ) be a probability space, and T a σ-subalgebra of Σ. Let (A, µ ¯) be the measure algebra of (X, Σ, µ). Show that C = {F • : F ∈ T} is a closed subalgebra of A. Identify the spaces Mσ (A), Mσ (C) of countably additive functionals with L1 (µ), L1 (µ¹ T), as in 327D. Show that the conditional expectation operator P : L1 (µ) → L1 (µ¹ T) (242Jd) corresponds to the map ν 7→ ν¹ C : Mσ (A) → Mσ (C). (c) Let (A, µ ¯) be a totally finite measure algebra and ν : A → R a countably additive functional. Show that, for any a ∈ A, νa =

R∞ 0

µ ¯(a ∩ [[ν > α¯ µ]])dα −

R0

−∞

µ ¯(a \ [[ν > α¯ µ]])dα,

the integrals being taken with respect to Lebesgue measure. (Hint: take (A, µ ¯) to be the measure algebra of (X, Σ, µ); represent ν by a µ-integrable function f ; apply Fubini’s theorem to the sets {(x, t) : x ∈ E, 0 ≤ t < f (x)}, {(x, t) : x ∈ E, f (x) ≤ t ≤ 0} in X × R, where a = E • .) (d) Let (A, µ ¯) be a totally finite measure algebra, C a closed subalgebra of A and ν : C → R a countably additive functional with standard extension ν˜ : A → R. Show that, for any a ∈ A, νã =

R∞ 0

µ ¯(a ∩ [[ν > α¯ µ¹ C]])dα −

R0

−∞

µ ¯(a \ [[ν > α¯ µ¹ C]])dα.

(e) Let (A, µ ¯) be a probability algebra, and B, C stochastically independent closed subalgebras of A (definition: 325L). Let ν be a countably additive functional on C, and ν˜ its standard extension to A. Show that ν˜(b ∩ c) = µ ¯b · νc for every b ∈ B, c ∈ C. (f ) Let (X, Σ, µ) be a probability space, and T a σ-subalgebra of Σ. Let ν be a probability measure with domain T such that νE = 0 whenever E ∈ T and µE = 0. Show that there is a probability measure λ with domain Σ which extends ν. 327Y Further exercises (a) Let (A1 , µ ¯1 ) and (A2 , µ ¯2 ) be localizable measure algebras with localizable ¯ Show that if ν1 , ν2 are completely additive functionals on A1 , A2 measure algebra free product (C, λ). respectively, there is a unique completely additive functional ν : C → R such that ν(a1 ⊗ a2 ) = ν1 a1 · ν2 a2 for every a1 ∈ A1 , a2 ∈ A2 . (Hint: 253D.) (b) Let (A, µ ¯) be a totally finite measure algebra and C a closed subalgebra; let R : Mσ (C) → Mσ (A) be the standard extension operator (327G). Show (i) that R is order-continuous (ii) that R(ν + ) = (Rν)+ , kRνk = kνk for every ν ∈ Mσ (C), defining ν + and kνk as in 326Yj. (c) Let (A, µ ¯) be a totally finite measure algebra and C a closed subalgebra of A. For a countably additive functional ν on C write ν˜ for its standard extension to A. Show that if ν, hνn in∈N are countably additive functionals on C and limn→∞ νn c = νc for every c ∈ C, then limn→∞ νñ a = νã for every a ∈ A. (Hint: use ideas from §§246-247, as well as from 327G and 326Yo.) 327 Notes and comments When we come to measure algebras, it is the completely additive functionals which fit most naturally into the topological theory (327Bd); they correspond to the ‘truly continuous’ functionals which I discussed in §232 (327Cc), and therefore to the Radon-Nikod´ ym theorem (327D). I will return to some of these questions in Chapter 36. I myself regard the form here as the best expression of the essence of the Radon-Nikod´ ym theorem, if not the one most commonly applied. The concept of ‘standard extension’ of a countably additive functional (or, as we could equally well say, of a completely additive functional, since in the context of 327F the two coincide) is in a sense dual to the concept of ‘conditional expectation’. If (X, Σ, µ) is a probability space and T is a σ-subalgebra of Σ, then we have a corresponding closed subalgebra C of the measure algebra (A, µ ¯) of µ, and identifications between the spaces Mσ (A), Mσ (C) of countably additive functionals and the spaces L1 (µ), L1 (µ¹ T). Now we have a natural embedding S of L1 (µ¹ T) as a subspace of L1 (µ) (242Jb), and a natural restriction map from Mσ (A) to Mσ (C). These give rise to corresponding operators between the opposite of each pair; the standard extension operator R of 327G, and the conditional expectation operator P of 242Jd. (See 327Xb.) The fundamental fact

122

Measure algebras

327 Notes

P Sv = v for every v ∈ L1 (µ¹ T) (242Jg) is matched by the fact that Rν¹ C = ν for every ν ∈ Mσ (C). R The further identification of Rν in of integrals µ ¯(a ∩ [[ν > α¯ µ]])dα (327Xc) is relatively inessential, but is striking, and perhaps makes it easier to believe that R is truly ‘standard’ in the abstract contexts which will arise in §333 below. It is also useful in such calculations as 327Xe. The isomorphisms between Mτ spaces and L1 spaces described here mean that any of the concepts involving L1 spaces discussed in Chapter 24 can be applied to Mτ spaces, at least in the case of measure algebras. In fact, as I will show in Chapter 36, there is much more to be said here; the space of bounded additive functionals on a Boolean algebra is already an L1 space in an abstract sense, and ideas such as ‘uniform integrability’ are relevant and significant there, as well as in the spaces of countably additive and completely additive functionals. I hope that 326Yj, 326Ym-326Yn, 326Yp-326Yq and 327Yb will provide some hints to be going on with for the moment.

331B

Classification of homogeneous measure algebras

123

Chapter 33 Maharam’s theorem We are now ready for the astonishing central fact about measure algebras: there are very few of them. Any localizable measure algebra has a canonical expression as a simple product of measure algebras of easily described types. This complete classification necessarily dominates all further discussion of measure algebras; to the point that all the results of Chapter 32 have to be regarded as ‘elementary’, since however complex their formulation they have been proved by techniques not involving, nor providing, any particular insight into the special nature of measure algebras. The proof depends, of course, on developing methods of defining measure-preserving homomorphisms and isomorphisms; I give a number of results, progressively more elaborate, but all based on the same idea. These techniques are of great power, leading, for instance, to an effective classification of closed subalgebras and their embeddings. ‘Maharam’s theorem’ itself, the classification of localizable measure algebras, is in §332. I devote §331 to the definition and description of ‘homogeneous’ probability algebras. In §333 I turn to the problem of describing pairs (A, C) where A is a probability algebra and C is a closed subalgebra. Finally, in §334, I give some straightforward results on the classification of free products of probability algebras.

331 Classification of homogeneous measure algebras I embark directly on the principal theorem of this chapter (331I), split between 331B, 331D and 331I; 331B and 331D will be the basis of many of the results in later sections of this chapter. In 331E-331H I introduce the concepts of ‘Maharam type’ and ‘Maharam homogeneity’. I discuss the measure algebras of products {0, 1}κ , showing that these provide a complete set of examples of Maharam-type-homogeneous probability algebras (331J-331L). I end the section with a brief comment on ‘homogeneous’ Boolean algebras (331M-331N). 331A Definition The following idea is almost the key to the whole chapter. Let A be a Boolean algebra and B an order-closed subalgebra of A. A non-zero element a of A is a relative atom over B if every c ⊆ a is of the form a ∩ b for some b ∈ B; that is, {a ∩ b : b ∈ B} is the principal ideal generated by a. We say that A is relatively atomless over B if there are no relative atoms in A over B. (I’m afraid the phrases ‘relative atom’, ‘relatively atomless’ are bound to seem opaque at this stage. I hope that after the structure theory of §333 they will seem more natural. For the moment, note only that a is an atom in A iff it is a relative atom over the smallest subalgebra {0, 1}, and every element of A is a relative atom over the largest subalgebra A. In a way, a is a relative atom over B if its image is an atom in a kind of quotient A/B. But we are two volumes away from any prospect of making sense of this kind of quotient.) 331B

The first lemma is the heart of Maharam’s theorem.

Lemma Let (A, µ ¯) be a totally finite measure algebra and B a closed subalgebra of A such that A is relatively atomless over B. Let ν : B → R be an additive functional such that 0 ≤ νb ≤ µ ¯b for every b ∈ B. Then there is a c ∈ A such that νb = µ ¯(b ∩ c) for every b ∈ B. Remark Recall that by 323H we need not distinguish between ‘order-closed’ and ‘topologically closed’ subalgebras. proof (a) It is worth noting straight away that ν is necessarily countably additive. This is easy to check from first principles, but if you want to trace the underlying ideas they are in 313O (the identity map from B to A is order-continuous), 326Ff (so µ¹ B : B → R is countably additive) and 326Gb (therefore ν is countably additive). (b) For each a ∈ A set νa b = µ ¯(b ∩ a) for every b ∈ B; then νa is countably additive (326Fd). The key P Because idea is the following fact: for every non-zero a ∈ A there is a non-zero d ⊆ a such that νd ≤ 21 νa . P A is relatively atomless over B, there is an e ⊆ a such that e 6= a ∩ b for any b ∈ B. Consider the countably

124

Maharam’s theorem

331B

additive functional λ = νa − 2νe : B → R. By 326I, there is a b0 ∈ B such that λb ≥ 0 whenever b ∈ B, b ⊆ b0 , while λb ≤ 0 whenever b ∈ B, b ∩ b0 = 0. If e ∩ b0 6= 0, try d = e ∩ b0 . Then 0 6= d ⊆ a, and for every b ∈ B 1 2

1 2

νd b = νe (b ∩ b0 ) = (νa (b ∩ b0 ) − λ(b ∩ b0 )) ≤ νa b (because λ(b ∩ b0 ) ≥ 0) so νd ≤ 21 νa . If e ∩ b0 = 0, then (by the choice of e) e 6= a ∩ (1 \ b0 ), so d = a \ (e ∪ b0 ) 6= 0, and of course d ⊆ a. In this case, for every b ∈ B, 1 2

1 2

νd b = νa (b \ b0 ) − νe (b \ b0 ) = (λ(b \ b0 ) + νa (b \ b0 )) ≤ νa b (because λ(b \ b0 ) ≤ 0), so once again νd ≤ 21 νa . Thus in either case we have a suitable d. Q Q (c) It follows at once, by induction on n, that if a is any non-zero element of A and n ∈ N then there is a non-zero d ⊆ a such that νd ≤ 2−n νa . (d) Now let C be the set {a : a ∈ A, νa ≤ ν}. Then 0 ∈ C, so C 6= ∅. If D ⊆ C is upwards-directed and not empty, then a = sup D is defined in A, and νsup D b = µ ¯(b ∩ sup D) = µ ¯(supd∈D b ∩ d) = supd∈D µ ¯(b ∩ d) = supd∈D νd b ≤ νb using 313Ba and 321C. So a ∈ C and is an upper bound for D in C. In particular, any non-empty totally ordered subset of C has an upper bound in C. By Zorn’s Lemma, C has a maximal element c say. (e) ?? Suppose, if possible, that νc 6= ν. Then there is some b∗ ∈ B such that νc b∗ 6= νb∗ ; since νc ≤ ν, νc b∗ < νb∗ . Let n ∈ N be such that νb∗ > νc b∗ + 2−n µ ¯b∗ , and set λb = νb − νc b − 2−n µ ¯b for every b ∈ B. By 326I (for the second time), there is a b0 ∈ B such that λb ≥ 0 for b ∈ B, b ⊆ b0 , while λb ≤ 0 when b ∈ B and b ∩ b0 = 0. We have µ ¯(b0 \ c) = µ ¯b0 − µ ¯(b0 ∩ c) ≥ νb0 − νc b0 ≥ λb0 = λb∗ + λ(b0 \ b∗ ) − λ(b∗ \ b0 ) ≥ λb∗ > 0, so b0 \ c 6= 0. (This is where I use the hypothesis that ν ≤ µ ¯¹ B.) By (c), there is a non-zero d ⊆ b0 \ c such that νd ≤ 2−n νb0 \c ≤ 2−n νb0 . Now d ∩ c = 0 so c ⊂ d ∪ c. Also, for any b ∈ B, νd∪c b = νd b + νc (b ∩ b0 ) + νc (b \ b0 ) ≤ 2−n νb0 \c b + ν(b ∩ b0 ) − 2−n µ ¯(b ∩ b0 ) − λ(b ∩ b0 ) + ν(b \ b0 ) ≤ 2−n µ ¯(b ∩ b0 \ c) + ν(b ∩ b0 ) − 2−n µ ¯(b ∩ b0 ) + ν(b \ b0 ) ≤ νb. But this means that d ∪ c ∈ C and c is not maximal in C. X X Thus c is the required element of A giving a representation of ν. 331C Corollary Let (A, µ ¯) be an atomless semi-finite measure algebra, and a ∈ A. Suppose that 0≤γ≤µ ¯a. Then there is a c ∈ A such that c ⊆ a and µ ¯c = γ. proof If γ = µ ¯a, take c = a. If γ < µ ¯a, there is a d ∈ A such that d ⊆ a and γ ≤ µ ¯d < ∞ (322Eb). Apply 331B to the principal ideal Ad generated by d, with B = {0, d} and νd = γ. (The point is that because A is atomless, no non-trivial principal ideal of Ad can be of the form {c ∩ b : b ∈ B} = {0, c}.) Remark Of course this is also an easy consequence of 215D.

331Fb


125

331D Lemma Let (A, µ ¯), (B, ν¯) be totally finite measure algebras and C ⊆ A a closed subalgebra. Suppose that π : C → B is a measure-preserving Boolean homomorphism such that B is relatively atomless over π[C]. Take any a ∈ A, and let C1 be the subalgebra of A generated by C ∪ {a}. Then there is a measure-preserving homomorphism from C1 to B extending π. proof We know that π[C] is a closed subalgebra of B (324Kb), and that π is a Boolean isomorphism between C and π[C]. Consequently the countably additive functional c 7→ µ ¯(c ∩ a) : C → R is transferred to a countably additive functional λ : π[C] → R, writing λ(πc) = µ ¯(c ∩ a) for every c ∈ C. Of course λ(πc) ≤ µ ¯c = ν¯(πc) for every c ∈ C. So by 331B there is a b ∈ B such that λ(πc) = ν¯(b ∩ πc) for every c ∈ C. If c ∈ C, c ⊆ a then ν¯(b ∩ πc) = λ(πc) = µ ¯(a ∩ c) = µ ¯c = ν¯(πc), so πc ⊆ b. Similarly, if a ⊆ c ∈ C, then ν¯(b ∩ πc) = µ ¯(a ∩ c) = µ ¯(a ∩ 1) = ν¯(b ∩ π1) = ν¯b, so b ⊆ πc. It follows from 312N that there is a Boolean homomorphism π1 : C1 → B, extending π, such that π1 a = b. To see that π1 is measure-preserving, take any member of C1 . By 312M, this is expressible as e = (c1 ∩ a) ∪ (c2 \ a), where c1 , c2 ∈ C. Now ν¯(π1 e) = ν¯((πc1 ∩ b) ∪ (πc2 \ b)) = ν¯(πc1 ∩ b) + ν¯(πc2 ) − ν¯(πc2 ∩ b) =µ ¯(c1 ∩ a) + µ ¯ c2 − µ ¯(c2 ∩ a) = µ ¯e. As e is arbitrary, π1 is measure-preserving. 331E Generating sets For the sake of the next definition, we need a language a little more precise than I have felt the need to use so far. The point is that if A is a Boolean algebra and B is a subset of A, there is more than one subalgebra of A which can be said to be ‘generated’ by B, because we can look at any of the three algebras – B, the smallest subalgebra of A including B; – Bσ , the smallest σ-subalgebra of A including B; – Bτ , the smallest order-closed subalgebra of A including B. (See 313Fb.) Now I will say henceforth, in this context, that – B is the subalgebra of A generated by B, and B generates A if A = B; – Bσ is the σ-subalgebra of A generated by B, and B σ-generates A if A = Bσ ; – Bτ is the order-closed subalgebra of A generated by B, and B τ -generates or completely generates A if A = Bτ . There is a danger inherent in these phrases, because if we have B ⊆ A0 , where A0 is a subalgebra of A, it is possible that the smallest order-closed subalgebra of A0 including B might not be recoverable from the smallest order-closed subalgebra of A including B. (See 331Yb-331Yc.) This problem will not seriously interfere with the ideas below; but for definiteness let me say that the phrases ‘B σ-generates A’, ‘B τ generates A’ will always refer to suprema and infima taken in A itself, not in any larger algebra in which it may be embedded. 331F Maharam types (a) With the language of 331E established, I can now define the Maharam type or complete generation τ (A) of any Boolean algebra A; it is the smallest cardinal of any subset of A which τ -generates A. (I think that this is the first ‘cardinal function’ which I have mentioned in this treatise. All you need to know, to confirm that the definition is well-conceived, is that there is some set which τ -generates A; and obviously A τ -generates itself. For this means that the set A = {#(B) : B ⊆ A τ -generates A} is a non-empty class of cardinals, and therefore, assuming the axiom of choice, has a least member (2A1Lf). In 331Ye-331Yf I mention a further function, the ‘density’ of a topological space, which is closely related to Maharam type.) (b) A Boolean algebra A is Maharam-type-homogeneous if τ (Aa ) = τ (A) for every non-zero a ∈ A, writing Aa for the principal ideal of A generated by a.

126

Maharam’s theorem

331Fc

(c) Let (X, Σ, µ) be a measure space, with measure algebra (A, µ ¯). Then the Maharam type of (X, Σ, µ), or of µ, is the Maharam type of A; and (X, Σ, µ), or µ, is Maharam-type-homogeneous if A is. Remark I should perhaps remark that the phrases ‘Maharam type’ and ‘Maharam-type-homogeneous’, while well established in the context of probability algebras, are not in common use for general Boolean algebras. But the cardinal τ (A) is important in the general context, and is such an obvious extension of Maharam’s idea (Maharam 42) that I am happy to propose this extension of terminology. 331G For the sake of those who have not mixed set theory and algebra before, I had better spell out some basic facts. Proposition Let A be a Boolean algebra, B a subset of A. Let B be the subalgebra of A generated by B, Bσ the σ-subalgebra of A generated by B, and Bτ the order-closed subalgebra of A generated by B. (a) B ⊆ Bσ ⊆ Bτ . (b) If B is finite, so is B, and in this case B = Bσ = Bτ . (c) For every a ∈ B, there is a finite B 0 ⊆ B such that a belongs to the subalgebra of A generated by B 0 . Consequently #(B) ≤ max(ω, #(B)). (d) For every a ∈ Bσ , there is a countable B 0 ⊆ B such that a belongs to the σ-subalgebra of A generated by B 0 . (e) If A is ccc, then Bσ = Bτ . proof (a) All we need to know is that Bσ is a subalgebra of A including B, and that Bτ is a σ-subalgebra of A including B. (b) Induce on #(B), using 312M for the inductive step, to see that B is finite. In this case it must be order-closed, so is equal to Bτ . S (c)(i) For I ⊆ B, let CI be the subalgebra of A generated by I. If I, J ⊆ B then CI ∪ CJ ⊆ CI∪J . So {CI : I ⊆ B is finite} is a subalgebra of A, and must be equal to B, as claimed. (ii) To estimate the size of B, recall that the set [B]<ω of all finite subsets of B has cardinal at most max(ω, #(B)) (3A1Cd). For each I ∈ [B]<ω , CI is finite, so S #(B) = #( I∈[B]<ω CI ) ≤ max(ω, #(I), supI∈[B]<ω #(CI )) ≤ max(ω, #(B)) by 3A1Cc. (d) ForSI ⊆ B, let DI ⊆ Bσ be the σ-subalgebra of A generated by I. If I, J ⊆ B then DI ∪ DJ ⊆ DI∪J , P so B0σ = {DI : I ⊆ B is countable} is a subalgebra of A. But also it is sequentially order-closed in A. P Let han in∈N be a non-decreasing sequenceSin B0σ with supremum a in A. For each n ∈ N there is a countable I(n) ⊆ B such that an ∈ CI(n) . Set K = n∈N I(n); then K is a countable subset of B and every an belongs to DK , so a ∈ DK ⊆ B0σ . Q Q So B0σ is a σ-subalgebra of A including B and must be the whole of Bσ . (e) By 316Fb, Bσ is order-closed in A, so must be equal to Bτ . 331H Proposition Let A be a Boolean algebra. (a)(i) τ (A) = 0 iff A is either {0} or {0, 1}. (ii) τ (A) is finite iff A is finite. (b) If B is another Boolean algebra and π : A → B is a surjective order-continuous Boolean homomorphism, then τ (B) ≤ τ (A). (c) If a ∈ A then τ (Aa ) ≤ τ (A), where Aa is the principal ideal of A generated by a. (d) If A has an atom and is Maharam-type-homogeneous, then A = {0, 1}. proof (a)(i) τ (A) = 0 iff A has no proper subalgebras. (ii) If A is finite, then τ (A) ≤ #(A) is finite. If τ (A) is finite, then there is a finite set B ⊆ A which τ -generates A; by 331Gb, A is finite. (b) We know that there is a set A ⊆ A, τ -generating A, with #(A) = τ (A). Now π[A] τ -generates π[A] = B (313Mb), so τ (B) ≤ #(π[A]) ≤ #(A) = τ (A).

331I


127

(c) Apply (b) to the map b 7→ a ∩ b : A → Aa . (d) If a ∈ A is an atom, then τ (Aa ) = 0, so if A is Maharam-type-homogeneous then τ (A) = 0 and A = {0, a} = {0, 1}. 331I

We are now ready for the theorem.

Theorem Let (A, µ ¯) and (B, ν¯) be Maharam-type-homogeneous measure algebras of the same Maharam type, with µ ¯1 = ν¯1 < ∞. Then they are isomorphic as measure algebras. proof (a) If τ (A) = τ (B) = 0, this is trivial. So let us take κ = τ (A) = τ (B) > 0. In this case, because A and B are Maharam-type-homogeneous, they can have no atoms and must be infinite, so κ is infinite (331H). Let haξ iξ<κ , hbξ iξ<κ enumerate τ -generating subsets of A, B respectively. The strategy of the proof is to define a measure-preserving isomorphism π : A → B as the last of an increasing family hπξ iξ≤κ of isomorphisms between closed subalgebras Cξ , Dξ of A and B. The inductive hypothesis will be that, for some families ha0ξ iξ<κ , hb0ξ iξ<κ to be determined, Cξ is the closed subalgebra of A generated by {aη : η < ξ} ∪ {a0η : η < ξ}, Dξ is the closed subalgebra of B generated by {bη : η < ξ} ∪ {b0η : η < ξ}, πξ : Cξ → Dξ is a measure-preserving isomorphism, πξ extends πη whenever η < ξ. (Formally speaking, this will be a transfinite recursion, defining a function ξ 7→ f (ξ) = (Cξ , Dξ , πξ , a0ξ , b0ξ ) on the ordinal κ + 1 by a rule which chooses f (ξ) in of f ¹ξ, as described in 2A1B. The construction of an actual function F for which f (ξ) = F (f ¹ξ) will necessitate the axiom of choice.) (b) The induction starts with C0 = {0, 1}, D0 = {0, 1}, π0 (0) = 0, π0 (1) = 1. (The hypothesis µ ¯1 = ν¯1 is what we need to ensure that π0 is measure-preserving.) (c) For the inductive step to a successor ordinal ξ + 1, where ξ < κ, suppose that Cξ , Dξ and πξ have been defined. (i) For any non-zero b ∈ B, the principal ideal Bb of B generated by b has Maharam type κ, because B is Maharam-type-homogeneous. On the other hand, the Maharam type of Dξ is at most #({bη : η ≤ ξ} ∪ {b0η : η < ξ}) ≤ #(ξ × {0, 1}) < κ, because if ξ is finite so is ξ × {0, 1}, while if ξ is infinite then #(ξ × {0, 1}) = #(ξ) ≤ ξ < κ. Consequently Bb cannot be an order-continuous image of Dξ (331Hb). Now the map c 7→ c ∩ b : Dξ → Bb is order-continuous, because Dξ is closed, so that the embedding Dξ ⊆ B is order-continuous. It therefore cannot be surjective, and {b ∩ πξ a : a ∈ Cξ } = {b ∩ d : d ∈ Dξ } 6= Bb . This means that πξ : Cξ → Dξ satisfies the conditions of 331D, and must have an extension φξ to a measure-preserving homomorphism from the subalgebra C0ξ of A generated by Cξ ∪ {aξ } to B. We know that C0ξ is a closed subalgebra of A (314Ja), so it must be the closed subalgebra of A generated by {aη : η ≤ ξ} ∪ {a0η : η < ξ}. Also D0ξ = φξ [C0ξ ] will be the subalgebra of B generated by Dξ ∪ {b0ξ }, where b0ξ = φξ (aξ ), so is closed in B, and is the closed subalgebra of B generated by {bη : η < ξ} ∪ {b0η : η ≤ ξ}. 0 (ii) The next step is to repeat the whole of the argument above, but applying it to φ−1 ξ : Dξ → Cξ , bξ 0 in place of πξ : Cξ → Dξ and aξ . Once again, we have τ (Dξ ) < κ = τ (Aa ) for every a ∈ A, so we can use Lemma 331D to find a measure-preserving isomorphism ψξ : Dξ+1 → Cξ+1 extending φ−1 ξ , where Dξ+1 is 0 the subalgebra of B generated by Dξ ∪ {bξ }, and Cξ+1 is the subalgebra of A generated by C0ξ ∪ {a0ξ }, setting a0ξ = ψξ (bξ ). As in (i), we find that Cξ+1 is the closed subalgebra of A generated by {aη : η ≤ ξ}∪{a0η : η ≤ ξ}, while Dξ+1 is the closed subalgebra of B generated by {bη : η ≤ ξ} ∪ {b0η : η ≤ ξ}.

(iii) We can therefore take πξ+1 = ψξ−1 : Cξ+1 → Dξ+1 , and see that πξ+1 is a measure-preserving isomorphism, extending πξ , such that πξ+1 (aξ ) = b0ξ , πξ+1 (a0ξ ) = bξ . Evidently πξ+1 extends πη for every η ≤ ξ because it extends πξ and (by the inductive hypothesis) πξ extends πη for every η < ξ.

128

Maharam’s theorem

331I

(d) For the inductive step to a limit 0 < ξ ≤ κ, suppose that Cη , Dη , a0η , b0η , πη have S ordinal ξ, where ∗ ∗ been defined for η < ξ. Set Cξ = η<ξ Cξ . Then Cξ is a subalgebra of A, because it is the union of an S upwards-directed family of subalgebras; similarly, D∗ξ = η<ξ Dξ is a subalgebra of B. Next, we have a function πξ∗ : C∗ξ → D∗ξ defined by setting πξ∗ a = πη a whenever η < ξ and a ∈ Cη ; for if η, ζ < ξ and a ∈ Cη ∩ Cζ , then πη a = πmax(η,ζ) a = πζ a. Clearly S πξ∗ [C∗ξ ] = η<ξ πη [Cη ] = D∗ξ . Moreover, ν¯πξ∗ a = µ ¯a for every a ∈ C∗ξ , since ν¯πη a = µ ¯a whenever η < ξ and a ∈ Cη . Now let Cξ be the smallest closed subalgebra of A including C∗ξ , that is, the metric closure of C∗ξ in A (323J). Since Cξ is the smallest closed subalgebra of A including Cη for every η < ξ, it must be the closed subalgebra of A generated by {aη : η < ξ} ∪ {a0η : η < ξ}. By 324O, πξ∗ has an extension to a measurepreserving homomorphism πξ : Cξ → B. Set Dξ = πξ [Cξ ]; by 324Kb, Dξ is a closed subalgebra of B. Because πξ : Cξ → B is continuous (also noted in 324Kb), D∗ξ = πξ∗ [C∗ξ ] = πξ [C∗ξ ] is topologically dense in Dξ (3A3Eb), and Dξ = D∗ξ is the closed subalgebra of B τ -generated by {bη : η < ξ} ∪ {b0η : η < ξ}. Finally, if η < ξ, πξ extends πη because πξ∗ extends πη . Thus the induction continues. (e) The induction ends with ξ = κ, Cκ = A, Dκ = B and π = πκ : A → B the required measure algebra isomorphism. 331J Lemma Let κ be any infinite cardinal. Let µ be the usual measure on {0, 1}κ (254J) and (A, µ ¯) its measure algebra. Then if (B, ν¯) is any non-zero totally finite measure algebra and π : A → B is an order-continuous Boolean homomorphism, τ (B) ≥ κ. proof Set X = {0, 1}κ and write Σ for the domain of µ. (a) Set Eξ = {x : x ∈ X, x(ξ) = 1}, aξ = Eξ• for each ξ < κ. If hξn in∈N is any sequence of distinct elements of κ, T T µ( n∈N Eξn ) = limn→∞ µ( i≤n Eξn ) = limn→∞ 2−n−1 = 0, so that µ ¯(infSn∈N aξn ) = 0 and inf n∈N aξn = 0. Because π is order-continuous, inf n∈N π(aξn ) = 0 in B. Similarly, µ( n∈N Eξn ) = 1 and supn∈N π(aξn ) = 1. (b) For b ∈ B, δ > 0 set U (b, δ) = {b0 : ν¯(b0 4 b) < δ}, the ordinary open δ-neighbourhood of b. If b ∈ B, then there is a δ > 0 such that {ξ : ξ < κ, aξ ∈ U (b, δ)} is finite. P P?? Suppose, if possible, otherwise. Then there is a sequence hξn in∈N of distinct elements of κ such that ν¯(b 4 πaξn ) ≤ 2−n−2 ν¯1 for every n ∈ N. Now inf n∈N πaξn = 0, so P∞ ν¯b = ν¯(b \ inf n∈N πaξn ) ≤ n=0 ν¯(b \ πaξn ). Similarly ν¯(1 \ b) = ν¯(supn∈N πaξn \ b) ≤

P∞ n=0

ν¯(πaξn \ b).

So ν¯1 = ν¯b + ν¯(1 \ b) ≤

∞ X

ν¯(b \ πaξn ) +

n=0

=

∞ X n=0

ν¯(b 4 πaξn ) ≤

∞ X

∞ X

ν¯(πaξn \ b)

n=0

2−n−2 ν¯1 < ν¯1,

n=0

which is impossible. X XQ Q (c) Note that B is infinite; for if b ∈ B the set {ξ : πaξ = b} must be finite, and κ is supposed to be infinite. So τ (B) must be infinite. (d) Now take a set B ⊆ B, of cardinal τ (B), which τ -generates B. By (c), B is infinite. Let C be the subalgebra of B generated by B; then #(C) = #(B) = τ (B), by 331Gc, and C is topologically dense in B.

331N


129

If b ∈ B, there are c ∈ C, k ∈ N such that b ∈ U (c, 2−k ) and {ξ : πaξ ∈ U (c, 2−k )} is finite. P P By (b), there is a δ > 0 such that {ξ : πaξ ∈ U (b, δ)} is finite. Take k ∈ N such that 2 · 2−k ≤ δ, and c ∈ C ∩ U (b, 2−k ); then U (c, 2−k ) ⊆ U (b, δ) can contain only finitely many πaξ , so these c, k serve. Q Q Consider U = {U (c, 2−k ) : c ∈ C, k ∈ N, {ξ : πaξ ∈ U (c, 2−k )} is finite}. S Then #(U) ≤ max(#(C), ω) = τ (B). Also U is a cover of B. In particular, κ = U ∈U JU , where JU = {ξ : πaξ ∈ U }. But this means that κ = #(κ) ≤ max(ω, #(U ), supU ∈U #(JU )) = τ (B), as claimed. 331K Theorem Let κ be any infinite cardinal. Let µ be the usual measure on {0, 1}κ and (A, µ ¯) its measure algebra. Then A is Maharam-type-homogeneous, of Maharam type κ. proof Set X = {0, 1}κ and write Σ for the domain of µ. (a) To see that τ (A) ≤ κ, set Eξ = {x : x ∈ X, x(ξ) = 1}, aξ = Eξ• for each ξ < κ. Writing E for the algebra of subsets of X generated by {Eξ : ξ < κ}, we see that every measurable cylinder in X, as defined in 254A, belongs to E, so that every member of Σ is approximated, in measure, by of E (254Fe), that is, {E • : E ∈ E} is topologically dense in A. But this means just that the subalgebra of A generated by {aξ : ξ < κ} is topologically dense in A, so that {aξ : ξ < κ} τ -generates A, and τ (A) ≤ κ. (b) Next, if c ∈ A \ {0} and Ac is the principal ideal of A generated by c, the map a 7→ a ∩ c is an order-continuous Boolean homomorphism from A to Ac , so by 331J we must have τ (Ac ) ≥ κ. Thus κ ≤ τ (Ac ) ≤ τ (A) ≤ κ. As c is arbitrary, A is Maharam-type-homogeneous with Maharam type κ. 331L Theorem Let (A, µ ¯) be a Maharam-type-homogeneous probability algebra. Then there is exactly one κ, either 0 or an infinite cardinal, such that (A, µ ¯) is isomorphic, as measure algebra, to the measure algebra (Aκ , µ ¯κ ) of the usual measure on {0, 1}κ . proof If τ (A) is finite, it is zero, and A = {0, 1} (331Ha, 331He) so that (interpreting {0, 1}0 as {∅}) we have the case κ = 0. If κ = τ (A) is infinite, then by 331K we know that (Aκ , µ ¯κ ) is also Maharam-typehomogeneous of Maharam type κ, so 331I gives the required isomorphism. Of course κ is uniquely defined by A. 331M Homogeneous Boolean algebras Having introduced the word ‘homogeneous’, I think I ought not to leave you without mentioning its standard meaning in the context of Boolean algebras, which is connected with one of the most striking and significant consequences of Theorem 331I. Definition A Boolean algebra A is homogeneous if A is isomorphic, as Boolean algebra, to every nontrivial principal ideal of A. Remark Of course a homogeneous Boolean algebra must be Maharam-type-homogeneous, since τ (A) = τ (Ac ) whenever A is isomorphic to Ac . In general, a Boolean algebra can be Maharam-type-homogeneous without being homogeneous (331Xj, 331Yj). But for σ-finite measure algebras this doesn’t happen. 331N Proposition Let (A, µ ¯) be a Maharam-type-homogeneous σ-finite measure algebra. Then it is homogeneous as a Boolean algebra. proof If A = {0} this is trivial; so suppose that A 6= {0}. By 322G, there is a measure ν¯ on A such that (A, ν¯) is a probability algebra. Now let c be any non-zero member of A, and set γ = ν¯c, ν¯c0 = γ −1 ν¯c , where ν¯c is the restriction of ν¯ to the principal ideal Ac of A generated by c. Then (A, ν¯) and (Ac , ν¯c0 ) are Maharamtype-homogeneous probability algebras of the same Maharam type, so are isomorphic as measure algebras, and a fortiori as Boolean algebras.

130

Maharam’s theorem

331X

331X Basic exercises (a) Let (X, Σ, µ) be a probability space, T a σ-subalgebra of Σ such that for any non-negligible E ∈ Σ there in an F ∈ Σ such that F ⊆ E and µ(F 4(E ∩ H)) > 0 for every H ∈ T. Suppose R that f : X → [0, 1] is a measurable function. Show that there is an F ∈ Σ such that H f = µ(H ∩ F ) for every H ∈ T. > (b) Write out a direct proof of 331C not relying on 331B. (c) Let A be a finite Boolean algebra with n atoms. Show that τ (A) is the least k such that n ≤ 2k . > (d) Show that the measure algebra of Lebesgue measure on R is Maharam-type-homogeneous and of • Maharam type ω. (Hint: show that it is τ -generated by {]−∞, q] : q ∈ Q}.) (e) Show that the measure algebra of Lebesgue measure on Rr is Maharam-type-homogeneous, of Ma• haram type ω, for any r ≥ 1. (Hint: show that it is τ -generated by {]−∞, q] : q ∈ Qr }.) (f ) Show that the measure algebra of any Radon measure on Rr (256A) has countable Maharam type. • (Hint: show that it is τ -generated by {]−∞, q] : q ∈ Qr }.) > (g) Show that PR has Maharam type ω. (Hint: show that it is τ -generated by {]−∞, q] : q ∈ Q}.) > (h) Show that the regular open algebra of R is Maharam-type-homogeneous, of Maharam type ω. (Hint: • show that it is τ -generated by {]−∞, q] : q ∈ Q}.) (i) Let (A, µ ¯) be a totally finite measure algebra, and κ an infinite cardinal. Suppose that there is a family haξ iξ<κ in A such that inf ξ∈I aξ = 0, supξ∈I aξ = 1 for every infinite I ⊆ κ. Show that τ (Aa ) ≥ κ for every non-zero principal ideal Aa of A. (j) Let A be the measure algebra of Lebesgue measure on R, and G the regular open algebra of R. Show that the simple product A × G is Maharam-type-homogeneous, with Maharam type ω, but is not homogeneous. (Hint: A is weakly (σ, ∞)-distributive, but G is not, so they are not isomorphic.) 331Y Further exercises (a) Suppose that A is a Dedekind complete Boolean algebra, B is an orderclosed subalgebra of A and C is an order-closed subalgebra of B. Show that if a ∈ A is a relative atom in A over C, then upr(a, B) is a relative atom in B over C. So if B is relatively atomless over C, then A is relatively atomless over C. (b) Give an example of a Boolean algebra A with a subalgebra A0 and a proper subalgebra B of A0 which is order-closed in A0 , but τ -generates A. (Hint: take A to be the measure algebra AL of Lebesgue measure on R and B the subalgebra BQ of A generated by {[a, b]• : a, b ∈ Q}. Take E ⊆ R such that I ∩E, I \E have non-zero measure for every non-trivial interval I ⊆ R (134Jb), and let A0 be the subalgebra of A generated by B ∪ {E • }.) (c) Give an example of a Boolean algebra A with a subalgebra A0 and a proper subalgebra B of A0 which is order-closed in A, but τ -generates A0 . (Hint: in the notation of 331Yb, take Z to be the Stone space of AL , and set A0 = {b a : a ∈ AL }, B = {b a : a ∈ BQ }; let A be the subalgebra of PZ generated by A0 ∪ {{z} : z ∈ Z}.) (d) Let A be a Dedekind complete purely atomic Boolean algebra, and A the set of its atoms. Show that τ (A) is the least cardinal κ such that #(A) ≤ 2κ . (e) Let (A, µ ¯) be a measure algebra. Write d(A) for the smallest cardinal of any subset of A which is dense for the measure-algebra topology. Show that d(A) ≤ max(ω, τ (A)). Show that if (A, µ ¯) is localizable, then τ (A) ≤ d(A). (f ) Let (X, ρ) be a metric space. Write d(X) for the density of X, the smallest cardinal of any dense subset of S S X. (i) Show that if G is any family of open subsets of X, there is a family H ⊆ G such that H = G and #(H) ≤ max(ω, d(X)). (ii) Show that if κ > max(ω, d(X)) and hxξ iξ<κ is any family in X, then there is an x ∈ X such that #({ξ : xξ ∈ G}) > max(ω, d(X)) for every open set G containing x, and that there is a sequence hξn in∈N of distinct of κ such that x = limn→∞ xξn .

331 Notes


131

(g) Show that the algebra of open-and-closed subsets of {0, 1}N is homogeneous. b (314U). (h) Show that if A is a homogeneous Boolean algebra so is its Dedekind completion A (i) Show that the regular open algebra of R is a homogeneous Boolean algebra. (j) Let (A, µ ¯) be the simple product (322K) of c copies of the measure algebra of the usual measure on {0, 1}c . Show that A is Maharam-type-homogeneous but not homogeneous. 331 Notes and comments Maharam’s theorem belongs with the Radon-Nikod´ ym theorem, Fubini’s theorem and the strong law of large numbers as one of the theorems which make measure theory what it is. Once you have this theorem and its consequences in the next section properly absorbed, you will never again look at a measure space without classifying its measure algebra in of the types of its homogeneous principal ideals. As one might expect, a very large proportion of the important measure spaces of analysis are homogeneous, and indeed a great many are homogeneous with Maharam type ω. In this section I have contented myself with the basic statement of Theorem 331I on the isomorphism of Maharam-type-homogeneous measure algebras and the identification of representative homogeneous probability algebras (331K). The same techniques lead to an enormous number of further facts, some of which I will describe in the rest of the chapter. For the moment, it gives us a complete description of Maharamtype-homogeneous probability algebras (331L). There is the atomic algebra {0, 1}, with Maharam type 0, and for each infinite cardinal κ there is the measure algebra of {0, 1}κ , with Maharam type κ; these are all non-isomorphic, and every Maharam-type-homogeneous probability algebra is isomorphic to exactly one of them. The isomorphisms here are not unique; indeed, it is characteristic of measure algebras that they have very large automorphism groups (see Chapter 38 below), and there are correspondingly large numbers of isomorphisms between any isomorphic pair. The proof of 331I already suggests this, since we have such a vast amount of choice concerning the lists haξ iξ<κ , hbξ iξ<κ , and even with these fixed there remains a good deal of scope in the choice of ha0ξ iξ<κ and hb0ξ iξ<κ . The isomorphisms described in Theorem 331I are measure algebra isomorphisms, that is, measurepreserving Boolean isomorphisms. Obvious questions arise concerning Boolean isomorphisms which are not necessarily measure-preserving; the theorem also helps us to settle many of these (see 331M-331N). But we can observe straight away the remarkable fact that two homogeneous probability algebras which are isomorphic as Boolean algebras are also isomorphic as probability algebras, since they must have the same Maharam type. I have already mentioned certain measure space isomorphisms (254K, 255A). Of course any isomorphism between measure spaces must induce an isomorphism between their measure algebras (see 324M), and any isomorphism between measure algebras corresponds to an isomorphism between their Stone spaces (see 324N). But there are many important examples of isomorphisms between measure algebras which do not correspond to isomorphisms between the measure spaces most naturally involved. (I describe one in 343J.) Maharam’s theorem really is a theorem about measure algebras rather than measure spaces. The particular method I use to show that the measure algebra of the usual measure on {0, 1}κ is homogeneous for infinite κ (331J-331K) is chosen with a view to a question in the next section (332O). There are other ways of doing it. But I recommend study of this particular one because of the way in which it involves the topological, algebraic and order properties of the algebra B. I have extracted some of the elements of the argument in 331Xi and 331Ye-331Yf. These use the concept of ‘density’ of a topological space. This does not seem the moment to go farther along this road, but I hope you can see that there are likely to be many further ‘cardinal functions’ to provide useful measures of complexity in both algebraic and topological structures.

132

Maharam’s theorem

§332 intro.

332 Classification of localizable measure algebras In this section I present what I call ‘Maharam’s theorem’, that every localizable measure algebra is expressible as a weighted simple product of measure algebras of spaces of the form {0, 1}κ (332B). Among its many consequences is a complete description of the isomorphism classes of localizable measure algebras (332J). This description needs the concepts of ‘cellularity’ of a Boolean algebra (332D) and its refinement, the ‘magnitude’ of a measure algebra (332G). I end this section with a discussion of those pairs of measure algebras for which there is a measure-preserving homomorphism from one to the other (332P-332Q), and a general formula for the Maharam type of a localizable measure algebra (332S). 332A Lemma Let A be any Boolean algebra. Writing Aa for the principal ideal generated by a ∈ A, the set {a : a ∈ A, Aa is Maharam-type-homogeneous} is order-dense in A. proof Take any a ∈ A \ {0}. Then A = {τ (Ab ) : 0 6= b ⊆ a} has a least member; take c ⊆ a such that c 6= 0 and τ (Ac ) = min A. If 0 6= b ⊆ c, then τ (Ab ) ≤ τ (Ac ), by 331Hc, while τ (Ab ) ∈ A, so τ (Ac ) ≤ τ (Ab ). Thus τ (Ab ) = τ (Ac ) for every non-zero b ⊆ c, and Ac is Maharam-type-homogeneous. 332B Maharam’s Theorem Let (A, µ ¯) be any localizable measure algebra. Then it is isomorphic to the simple product of a family h(Ai , µ ¯i )ii∈I of measure algebras, where for each i ∈ I (Ai , µ ¯i ) is isomorphic, up to a re-normalization of the measure, to the measure algebra of the usual measure on {0, 1}κi , where κi is either 0 or an infinite cardinal. proof (a) For a ∈ A, let Aa be the principal ideal of A generated by a. Then D = {a : a ∈ A, 0 < µ ¯a < ∞, Aa is Maharam-type-homogeneous} is order-dense in A. P P If a ∈ A \ {0}, then (because (A, µ ¯) is semi-finite) there is a b ⊆ a such that 0<µ ¯b < ∞; now by 332A there is a non-zero d ⊆ b such that Ad is Maharam-type-homogeneous. By 331N, Ad is Maharam-type-homogeneous, and d ∈ D. Q Q (b) By 313K, there is a partition of unity hei ii∈I consisting of of D; by 322Kd, (A, µ ¯) is isomorphic, as measure algebra, to the simple product of the principal ideals Ai = Aei . (c) For each i ∈ I, (Ai , µ ¯i ) is a non-trivial totally finite Maharam-type-homogeneous measure algebra, writing µ ¯i = µ ¯¹ Ai . Take γi = µ ¯i (1Ai ) = µ ¯ei , and set µ ¯0i = γi−1 µ ¯i . Then (Ai , µ ¯0i ) is a Maharam-typehomogeneous probability algebra, so by 331L is isomorphic to the measure algebra (Bκi , ν¯κi ) of the usual measure on {0, 1}κi , where κi is either 0 or an infinite cardinal. Thus (Ai , µ ¯i ) is isomorphic, up to a scalar multiple of the measure, to (Bκi , ν¯κi ). Remark For the case of totally finite measure algebras, this is Theorem 2 of Maharam 42. 332C Corollary Let (A, µ ¯) be a localizable measure algebra. For any cardinal κ, write νκ for the usual measure on {0, 1}κ , and Σκ for its domain. Then we can find families hκi ii∈I , hγi ii∈I such that every κi is either 0 or an infinite cardinal, every γi is a strictly positive real number, and (A, µ ¯) is isomorphic to the measure algebra of (X, Σ, ν), where X = {(x, i) : i ∈ I, x ∈ {0, 1}κi }, Σ = {E : E ⊆ X, {x : (x, i) ∈ E} ∈ Σκi for every i ∈ I}, P νE = i∈I γi νκi {x : (x, i) ∈ E} for every E ∈ Σ. proof Take the family hκi ii∈I from the last theorem, take the γi = µ ¯ei to be the normalizing factors of the proof there, and apply 322Kb to identify the simple product of the measure algebras of ({0, 1}κi , Σκi , γi νκi ) with the measure algebra of their direct sum (X, Σ, ν).

332H

Classification of localizable measure algebras

133

332D The cellularity of a Boolean algebra In order to properly describe non-sigma-finite measure algebras, we need the following concept. If A is any Boolean algebra, write c(A) = sup{#(C) : C ⊆ A \ {0} is dist}, the cellularity of A. (If A = {0}, take c(A) = 0.) Thus A is ccc (316A) iff c(A) ≤ ω. 332E Proposition Let (A, µ ¯) be any semi-finite measure algebra, and C any partition of unity in A consisting of elements of finite measure. Then max(ω, #(C)) = max(ω, c(A)). proof Of course #(C \ {0}) ≤ c(A), because C \ {0} is dist, so max(ω, #(C)) = max(ω, #(C \ {0}) ≤ max(ω, c(A)). Now suppose that D is any dist set in A \ {0}. For c ∈ C, {d ∩ c : d ∈ D} is a dist set in the principal ideal Ac generated by c. But Ac is ccc (322G), so {d S ∩ c : d ∈ D} must be countable, and Dc = {d : d ∈ D, d ∩ c 6= 0} is countable. Because sup C = 1, D = c∈C Dc , so #(D) ≤ max(ω, #(C), supc∈C #(Dc )) = max(ω, #(C)). As D is arbitrary, c(A) ≤ max(ω, #(C)) and max(ω, c(A)) = max(ω, #(C)). 332F Corollary Let (A, µ ¯) be any semi-finite measure algebra. Then there is a dist set in A \ {0} of cardinal c(A). proof Start by taking any partition of unity C consisting of non-zero elements of finite measure. If #(C) = c(A) we can stop, because C is a dist set in A \ {0}. Otherwise, by 332E, we must have C finite and c(A) = ω. Let A be the set of atoms in A. If A is infinite, it is a dist set of cardinal ω, so we can stop. Otherwise, since there is certainly a dist set D ⊆ A \ {0} of cardinal greater than #(A), and since each member of A can meet at most one member of D, there must be a member d of D which does not include any atom. Accordingly we can choose inductively a sequence hdn in∈N such that d0 = d, 0 6= dn+1 ⊂ dn for every n. Now {dn \ dn+1 : n ∈ N} is a dist set in A \ {0} of cardinal ω = c(A). 332G Definitions For the next theorem, it will be convenient to have some special terminology. (a) The first word I wish to introduce is a variant of the idea of ‘cellularity’, adapted to measure algebras. If (A, µ ¯) is a semi-finite measure algebra, let us say that the magnitude of an a ∈ A is µ ¯a if µ ¯a is finite, and otherwise is the cellularity of the principal ideal Aa generated by a. (This is necessarily infinite, since any partition of a into sets of finite measure must be infinite.) If we take it that any real number is less than any infinite cardinal, then the class of possible magnitudes is totally ordered. I shall sometimes speak of the magnitude of the measure algebra (A, µ ¯) itself, meaning the magnitude of 1A . Similarly, if (X, Σ, µ) is a semi-finite measure space, the magnitude of (X, Σ, µ), or of µ, is the magnitude of its measure algebra. (b) Next, for any Dedekind complete Boolean algebra A, and any cardinal κ, we can look at the element eκ = sup{a : a ∈ A \ {0}, Aa is Maharam-type-homogeneous of Maharam type κ}, writing Aa for the principal ideal of A generated by a, as usual. I will call this the Maharam-type-κ component of A. Of course eκ ∩ eλ = 0 whenever λ, κ are distinct cardinals. P P a ∩ b = 0 whenever Aa , Ab are Maharam-type-homogeneous of different Maharam types, since τ (Aa∩b ) cannot be equal simultaneously to τ (Aa ) and τ (Ab ). Q Q Also {eκ : κ is a cardinal} is a partition of unity in A, because sup{eκ : κ is a cardinal} = sup{a : Aa is Maharam-type-homogeneous} = 1 by 332A. Note that there is no claim that Aeκ itself is homogeneous; but we do have a useful result in this direction. 332H Lemma Let A be a Dedekind complete Boolean algebra and κ an infinite cardinal. Let e be the Maharam-type-κ component of A. If 0 6= d ⊆ e and the principal ideal Ad generated by d is ccc, then it is Maharam-type-homogeneous with Maharam type κ.

134

Maharam’s theorem

332H

proof (a) The point is that τ (Ad ) ≤ κ. P P Set A = {a : a ∈ A \ {0}, Aa is Maharam-type-homogeneous of Maharam type κ}. Then d = sup{a ∩ d : a ∈ A}. Because Ad is ccc, there is a sequence han in∈N in A such that d = supn∈N d ∩ an (316E); set bn = d ∩ an . We have τ (Abn ) ≤ τ (Aan ) = κ for each n; let Dn be a subset of Abn , of cardinal at most κ, which τ -generates Abn . Set S D = n∈N Dn ∪ {bn : n ∈ N} ⊆ Ad . If C is the order-closed subalgebra of Ad generated by D, then C ∩ Abn is an order-closed subalgebra of Abn including Dn , so is equal to Abn , for every n. But a = supn∈N a ∩ bn for every a ∈ Ad , so C = Ad . Thus D τ -generates Ad , and τ (Ad ) ≤ #(D) ≤ max(ω, supn∈N #(Dn )) = κ. Q Q (b) If now b is any non-zero member of Ad , there is some a ∈ A such that b ∩ a 6= 0, so that κ = τ (Ab∩a ) ≤ τ (Ab ) ≤ τ (Ad ) ≤ κ. Thus we must have τ (Ab ) = κ for every non-zero b ∈ Ad , and Ad is Maharam-type-homogeneous of type κ, as claimed. 332I Lemma Let (A, µ ¯) be an atomless semi-finite measure algebra which is not totally finite. Then it has a partition of unity consisting of elements of measure 1. proof Let A be the set {a : µ ¯a = 1}, and C the family of dist subsets of A. By Zorn’s lemma, C has a maximal member C0 (compare the proof of 313K). Set D = {d : d ∈ A, d ∩ c = 0 for every c ∈ C0 }. Then D is upwards-directed. If d ∈ D, then µ ¯a 6= 1 for every a ⊆ d, so µ ¯d < 1, by 331C. So d0 = sup D is defined in A (321C); of course d0 ∈ D, so µ ¯d0 < 1. Observe that sup C0 = 1 \ d0 . Because µ ¯1 = ∞, C0 must be infinite; let han in∈N be any sequence of distinct elements of C0 . For each ¯d0 . Set ¯a0n = µ n ∈ N, use 331C again to choose an a0n ⊆ an such that µ b0 = d0 ∪ (a0 \ a00 ),

bn = a0n−1 ∪ (an \ a0n )

for every n ≥ 1. Then hbn in∈N is a dist sequence of elements of measure 1 and supn∈N bn = supn∈N an ∪ d0 . Now (C0 \ {an : n ∈ N}) ∪ {bn : n ∈ N} is a partition of unity consisting of elements of measure 1. 332J Now I can formulate a complete classification theorem for localizable measure algebras, refining the expression in 332B. Theorem Let (A, µ ¯) and (B, ν¯) be localizable measure algebras. For each cardinal κ, let eκ , fκ be the Maharam-type-κ components of A, B respectively. Then (A, µ ¯) and (B, ν¯) are isomorphic, as measure algebras, iff (i) eκ and fκ have the same magnitude for every infinite cardinal κ (ii) for every γ ∈ ]0, ∞[, (A, µ ¯) and (B, ν¯) have the same number of atoms of measure γ. proof Throughout the proof, write Aa for the principal ideal of A generated by a, and µ ¯a for the restriction of µ ¯ to Aa ; and define Bb , ν¯b similarly for b ∈ B. (a) If (A, µ ¯) and (B, ν¯) are isomorphic, then of course the isomorphism matches their Maharam-type components together and retains their magnitudes, and matches atoms of the same measure together; so the conditions are surely satisfied. (b) Now suppose that the conditions are satisfied. Set K = {κ : κ is an infinite cardinal, eκ 6= 0} = {κ : κ is an infinite cardinal, fκ 6= 0}. For γ ∈ ]0, ∞[, let Aγ be the set of atoms of measure γ in A, and set eγ = sup Aγ . Write I = K ∪ ]0, ∞[. Then hei ii∈I is a partition of unity in A, so (A, µ ¯) is isomorphic to the simple product of h(Aei , µ ¯ei )ii∈I , writing Aei for the principal ideal generated by ei and µ ¯ei for the restriction µ ¯ ¹ Aei .

332L


135

In the same way, writing Bγ for the set of atoms of measure γ in B, fγ for sup Bγ , Bfi for the principal ideal generated by fi and ν¯fi for the restriction of ν¯ fo Bfi , we have (B, ν¯) isomorphic to the simple product of h(Bfi , ν¯fi )ii∈I . (c) It will therefore be enough if I can show that (Aei , µ ¯ ei ) ∼ = (Bfi , ν¯fi ) for every i ∈ I. (i) For κ ∈ K, the hypothesis is that eκ and fκ have the same magnitude. If they are both of finite magnitude, that is, µ ¯eκ = ν¯fκ < ∞, then both (Aeκ , µ ¯eκ ) and (Bfκ , ν¯fκ ) are homogeneous and of Maharam type κ, by 332H. So 331I tells us that they are isomorphic. If they are both of infinite magnitude λ, then 332I tells us that both Aeκ , Bfκ have partitions of unity C, D consisting of sets of measure 1. So (Aeκ , µ ¯ eκ ) is isomorphic to the simple product of h(Ac , µ ¯c )ic∈C , while (Bfκ , ν¯fκ ) is isomorphic to the simple product of h(Bd , ν¯d )id∈D . But we know also that every (Ac , µ ¯c ), (Bd , ν¯d ) is a homogeneous probability algebra of Maharam type κ, by 332H again, so by Maharam’s theorem again they are all isomorphic. Since C, D and λ are all infinite, #(C) = c(Aeκ ) = λ = c(Bfκ ) = #(D) by 332E. So we are taking the same number of factors in each product and (Aeκ , µ ¯eκ ) must be isomorphic to (Bfκ , ν¯fκ ). (ii) For γ ∈ ]0, ∞[, our hypothesis is that #(Aγ ) = #(Bγ ). Now Aγ is a partition of unity in Aeγ , so (Aeγ , µ ¯eγ ) is isomorphic to the simple product of h(Aa , µ ¯a )ia∈Aγ . Similarly, (Bfγ , ν¯fγ ) is isomorphic to the simple product of h(Bb , ν¯b )ib∈Bγ . Since every (Aa , µ ¯a ), (Bb , ν¯b ) is just a simple atom of measure γ, these are all isomorphic; since we are taking the same number of factors in each product, (Aeγ , µ ¯eγ ) must be isomorphic to (Bfγ , ν¯fγ ). (iii) Thus we have the full set of required isomorphisms, and (A, µ ¯) is isomorphic to (B, ν¯). 332K Remarks (a) The partition of unity {ei : i ∈ I} of A used in the above theorem is in some sense canonical. (You might feel it more economical to replace I by K ∪ {γ : Aγ 6= ∅}.) The further partition of the atomic part into individual atoms (part (c-ii) of the proof) is also canonical. But of course the partition of the eκ of infinite magnitude into elements of measure 1 requires a degree of arbitrary choice. The value of the expressions in 332C is that the parameters κi , γi there are sufficient to identify the measure algebra up to isomorphism. For, amalgamating the language of 332C and 332J, we see that the P magnitude of eκ in 332J is just κi =κ γi if this is finite, #({i : κi = κ}) otherwise (using 332E, as usual); while the number of atoms of measure γ is #({i : κi = 0, γi = γ}). (b) The classification which Maharam’s theorem gives us is not merely a listing. It involves a real insight into the nature of the algebras, enabling us to answer a very wide variety of natural questions. I give the next couple of results as a sample of what we can expect these methods to do for us. 332L Proposition Let (A, µ ¯) be a measure algebra, and a, b ∈ A two elements of finite measure. Suppose that π : Aa → Ab is a measure-preserving isomorphism, where Aa , Ab are the principal ideals generated by a and b. Then there is a measure-preserving automorphism φ : A → A which extends π. proof The point is that Ab\a is isomorphic, as measure algebra, to Aa\b . P P Set c = a ∪ b. For each infinite cardinal κ, let eκ be the Maharam-type-κ component of Ac . Then eκ ∩ a is the Maharam-type-κ component of Aa , because if d ⊆ c and Ad is Maharam homogeneous with Maharam type κ, then Ad∩a is either {0} or again Maharam-type-homogeneous with Maharam type κ. Similarly, eκ \ a is the Maharam-type-κ component of Ac\a = Ab\a , eκ ∩ b is the Maharam-type-κ component of Ab and eκ \ b is the Maharam-type-κ component of Aa\b . Now π : Aa → Ab is an isomorphism, so π(eκ ∩ a) must be eκ ∩ b, and µ ¯(eκ \ a) = µ ¯ eκ − µ ¯(eκ ∩ a) = µ ¯ eκ − µ ¯π(eκ ∩ a) =µ ¯ eκ − µ ¯(eκ ∩ b) = µ ¯(eκ \ b). In the same way, if we write nγ (d) for the number of atoms of measure γ in Ad , then nγ (b \ a) = nγ (c) − nγ (a) = nγ (c) − nγ (b) = nγ (a \ b)

136

Maharam’s theorem

332L

for every γ ∈ ]0, ∞[. By 332J, there is a measure-preserving isomorphism π1 : Ab\a → Aa\b . Q Q If we now set φd = π(d ∩ a) ∪ π1 (d ∩ b \ a) ∪ (d \ c) for every d ∈ A, φ : A → A is a measure-preserving isomorphism which agrees with π on Aa . 332M Lemma Suppose that (A, µ ¯) and (B, ν¯) are homogeneous measure algebras, with τ (A) ≤ τ (B) and µ ¯1 = ν¯1 < ∞. Then there is a measure-preserving Boolean homomorphism from A to B. proof The case τ (A) = 0 is trivial. Otherwise, considering normalized versions of the measures, we are reduced to the case µ ¯1 = ν¯1 = 1, τ (A) = κ ≥ ω, τ (B) = λ ≥ κ, so that (A, µ ¯) is isomorphic to the measure algebra (Aκ , µ ¯κ ) of the usual measure µ ¯κ on {0, 1}κ ; and similarly (B, ν¯) is isomorphic to the measure algebra of the usual measure on {0, 1}λ . Now (identifying the cardinals κ, λ with von Neumann ordinals, as usual), κ ⊆ λ, so we have an inverse-measure-preserving map x 7→ x¹κ : {0, 1}λ → {0, 1}κ (254Oa), which induces a measure-preserving Boolean homomorphism from Aκ to Aλ (324M), and hence a measure-preserving homomorphism from A to B. 332N Lemma If (A, µ ¯) is a probability algebra and κ ≥ max(ω, τ (A)), then there is a measure-preserving homomorphism from (A, µ ¯) to the measure algebra (B, ν¯) of the usual measure ν on {0, 1}κ ; that is, (A, µ ¯) is isomorphic to a closed subalgebra of (B, ν¯). proof Let hci ii∈I be a partition of P unity in A such that every principal ideal Aci is homogeneous and no ¯ci = 1. Accordingly there is a partition of unity hdi ii∈I in Bκ ci is zero. Then I is countable and i∈I µ such that ν¯di = µ ¯ci for every i. P P Because I is countable, we may suppose that it is either N or an initial segment of N. In this case, choose hdi ii∈I inductively such that di ⊆ 1 \ supj
332P


P i∈J

ν¯πci =

P i∈J

137

µ ¯ci = µ ¯e∗v ,

the magnitude of e∗v . While if e∗v has infinite magnitude, this is #(J), by 332E, which is not greater than the magnitude of fv∗ . 332P Proposition Let (A, µ ¯), (B, ν¯) be atomless totally finite measure algebras. For each infinite cardinal κ let eκ , fκ be their Maharam-type-κ components. Then the following are equiveridical: (i) (A, µ ¯) is isomorphic to a closed subalgebra of a principal ideal of (B, ν¯); (ii) for every cardinal λ, µ ¯(supκ≥λ eκ ) ≤ ν¯(supκ≥λ fκ ). proof (a)(i)⇒(ii) Suppose that π : A → Bd is a measure-preserving isomorphism between A and a closed subalgebra of a principal ideal Bd of B. The Maharam-type-κ component of Bd is just d ∩ fκ , so 332O tells us that µ ¯(supκ≥λ eκ ) ≤ ν¯(supκ≥λ d ∩ fκ ) ≤ ν¯(supκ≥λ fκ ) for every λ. (b)(ii)⇒(i) Now suppose that the condition is satisfied. α) Let P be the set of all measure-preserving Boolean homomorphisms π from principal ideals Acπ of (α A to principal ideals Bdπ of B such that µ ¯(supκ≥λ eκ \ cπ ) ≤ ν¯(supκ≥λ ν¯fκ \ dπ ) for every cardinal λ ≥ ω. Then the trivial homomorphism from A0 to B0 belongs to P , so P is not empty. Order P by saying that π ≤ π 0 if π 0 extends π, that is, if cπ ⊆ cπ0 and π 0 a = πa for every a ∈ Acπ . Then P is a partially ordered set. β ) If Q ⊆ P is non-empty and totally ordered, it is bounded above in P . P (β P Set c∗ = supπ∈Q cπ , ∗ ∗ d = supπ∈Q dπ . For a ⊆ c set π a = supπ∈Q π(a ∩ cπ ). Because Q is totally ordered, π ∗ extends all the functions in Q. It is also easy to check that π ∗ 0 = 0, π ∗ (a ∩ a0 ) = π ∗ a ∩ π ∗ a0 and π ∗ (a ∪ a0 ) = π ∗ a ∪ π ∗ a0 for all a, a0 ∈ Ac∗ , π ∗ c∗ = d∗ and that ν¯π ∗ a = µ ¯a for every a ∈ Ac∗ ; so that π ∗ is a measure-preserving Boolean homomorphism from Ac∗ to Bd∗ . Now suppose that λ is any cardinal; then ∗

µ ¯(sup eκ \ c∗ ) = inf µ ¯(sup eκ \ cπ ) ≤ inf ν¯(sup fκ \ dπ ) = ν¯(sup fκ \ d∗ ). κ≥λ

π∈Q

κ≥λ

π∈Q

κ≥λ

κ≥λ

∗

So π ∈ P and is the required upper bound of Q. Q Q (γγ ) By Zorn’s Lemma, P has a maximal element π ˜ say. Now cπ˜ = 1. P P?? If not, then let κ0 be the least cardinal such that eκ0 \ cπ˜ 6= 0. Then 0<µ ¯(supκ≥κ0 eκ \ cπ˜ ) ≤ ν¯(supκ≥κ0 ν¯fκ \ dπ˜ ), so there is a least κ1 ≥ κ0 such that fκ1 \ dπ˜ 6= 0. Set δ = min(¯ µ(eκ0 \ cπ˜ ), ν¯(fκ1 \ dπ˜ )) > 0. Because A and B are atomless, there are a ⊆ eκ0 \ cπ˜ , b ⊆ fκ1 \ dπ˜ such that µ ¯a = ν¯b = δ (331C). Now Aa is homogeneous with Maharam type κ0 , while Bb is homogeneous with Maharam type κ1 (332H), so there is a measure-preserving Boolean homomorphism φ : Aa → Bb (332M). Set c∗ = cπ˜ ∪ a,

d∗ = dπ˜ ∪ b,

and define π ∗ : Ac∗ → Bd∗ by setting π ∗ (g) = π ˜ (g ∩ cπ˜ ) ∪ φ(g ∩ a) for every g ⊆ c∗ . It is easy to check that ∗ π is a measure-preserving Boolean homomorphism. If λ is a cardinal and λ ≤ κ0 , µ ¯(sup eκ \ c∗ ) = µ ¯(sup eκ \ cπ˜ ) − δ ≤ ν¯(sup fκ \ dπ˜ ) − δ = ν¯(sup ν¯fκ \ d∗ ). κ≥λ

If κ0 < λ ≤ κ1 ,

κ≥λ

κ≥λ

κ≥λ

138

Maharam’s theorem

332P

µ ¯(sup eκ \ c∗ ) = µ ¯(sup eκ \ cπ˜ ) ≤ µ ¯( sup eκ \ cπ˜ ) − µ ¯(eκ0 \ cπ ) κ≥λ

κ≥κ0

κ≥λ

≤µ ¯( sup eκ \ cπ˜ ) − δ ≤ ν¯( sup fκ \ dπ˜ ) − δ κ≥κ0

κ≥κ0

= ν¯( sup fκ \ dπ˜ ) − δ κ≥κ1

(by the choice of κ1 ) = ν¯( sup fκ \ d∗ ) ≤ ν¯(sup fκ \ d∗ ). κ≥κ1

κ≥λ

If λ > κ1 , µ ¯(sup eκ \ c∗ ) = µ ¯(sup eκ \ cπ˜ ) ≤ ν¯(sup fκ \ dπ˜ ) = ν¯(sup fκ \ d∗ ). κ≥λ

κ≥λ

κ≥λ

κ≥λ

∗

But this means that π ∈ P , and evidently it is a proper extension of π ˜ , which is supposed to be impossible. X XQ Q (δδ ) Thus π ˜ has domain A and is the required measure-preserving homomorphism from A to the principal ideal Bdπ˜ of B. 332Q Proposition Let (A, µ ¯) and B, ν¯) be totally finite measure algebras, and suppose that there are measure-preserving Boolean homomorphisms π1 : A → B and π2 : B → A. Then (A, µ ¯) and (B, ν¯) are isomorphic. proof Writing eκ , fκ for their Maharam-type-κ components, 332O (applied to both π1 and π2 ) tells us that µ ¯(supκ≥λ eκ ) = ν¯(supκ≥λ fκ ) for every λ. Because all these measures are finite, µ ¯ eλ = µ ¯(sup eκ ) − µ ¯(sup eκ ) κ≥λ

κ>λ

= ν¯(sup fκ − ν¯(sup fκ ) = ν¯fλ κ≥λ

κ>λ

for every λ. Similarly, writing eγ , fγ for the suprema in A, B of the atoms of measure γ, 332O tells us that µ ¯(supγ≤δ eγ ) = ν¯(supγ≤δ fγ ) for every δ ∈ ]0, ∞[, and hence that µ ¯eγ = ν¯fγ for every γ, that is, that A and B have the same number of atoms of measure γ. So (A, µ ¯) and (B, ν¯) are isomorphic, by 332J. 332R 332J tells us that if we know the magnitudes of the Maharam-type-κ components of a localizable measure algebra, we shall have specified the algebra completely, so that all its properties are determined. The calculation of its Maharam type is straightforward and useful, so I give the details. Lemma Let (A, µ ¯) be a semi-finite measure algebra. Then c(A) ≤ 2τ (A) . proof Let C ⊆ A \ {0} be a dist set, and B ⊆ A a τ -generating set of size τ (A). (a) If A is purely atomic, then for each c ∈ C choose an atom c0 ⊆ c, and set f (c) = {b : b ∈ B, c0 ⊆ b}. If c1 , c2 are distinct of C, the set {a : a ∈ A, c01 ⊆ a ⇐⇒ c02 ⊆ a} is an order-closed subalgebra of A not containing either c01 or c02 , so cannot include B, and f (c1 ) 6= f (c2 ). Thus f is injective, and #(C) ≤ #(PB) = 2τ (A) .

332T


139

(b) Now suppose that A is not purely atomic; in this case τ (A) is infinite. For each c ∈ C choose an element c0 ⊆ c of non-zero finite measure. Let B be the subalgebra of A generated by B. Then the topological closure of B is A itself (323J), and #(B) = τ (A) (331Gc). For c ∈ C set 1 2

f (c) = {b : b ∈ B, µ ¯(b ∩ c0 ) ≥ µ ¯c0 }. Then f : C → PB is injective. P P If c1 , c2 are distinct of C, then (because B is topologically dense in A) there is a b ∈ B such that µ ¯((c01 ∪ c02 ) ∩ (c01 4 b)) ≤

1 3

min(¯ µc01 , µ ¯c02 ).

But in this case 1 3

¯c01 , µ ¯(c01 \ b) ≤ µ

1 3

µ ¯(c02 ∩ b) ≤ µ ¯c02 ,

and b ∈ f (c1 )4f (c2 ), so f (c1 ) 6= f (c2 ). Q Q Accordingly #(C) ≤ 2#(B) = 2τ (A) in this case also. τ (A) As C is arbitrary, c(A) ≤ 2 . 332S Theorem Let (A, µ ¯) be a localizable measure algebra. Then τ (A) is the least cardinal λ such that (α) c(A) ≤ 2λ (β) τ (Aa ) ≤ λ for every Maharam-type-homogeneous principal ideal Aa of A. proof Fix λ as the least cardinal satisfying (α) and (β). (a) By 331Hc, τ (Aa ) ≤ τ (A) for every a ∈ A, while c(A) ≤ 2τ (A) by 332R; so λ ≤ τ (A). (b) Let C be a partition of unity in A consisting of elements of non-zero finite measure generating Maharam-type-homogeneous principal ideals (as in the proof of 332B); then #(C) ≤ c(A) ≤ 2λ , and there is an injective function f : C → Pλ. For each c ∈ C, let Bc ⊆ Ac be a τ -generating set of cardinal τ (Ac ), and fc : Bc → λ an injection. Set bξ = sup{c : c ∈ C, ξ ∈ f (c)}, b0ξ = sup{b : there is some c ∈ C such that b ∈ Bc , fc (b) = ξ} for ξ < λ. Set B = {bξ : ξ < λ} ∪ {b0ξ : ξ < λ} if λ is infinite, {bξ : ξ < λ} if λ is finite; then #(B) ≤ λ. Note that if c ∈ C and b ∈ Bc there is a b0 ∈ B such that b = b0 ∩ c. P P Since Bc 6= ∅, τ (Ac ) > 0; but this means that τ (Ac ) is infinite (see 331H) so λ is infinite and b0ξ ∈ B, where ξ = fc (b); now b = b0ξ ∩ c. Q Q Let B be the closed subalgebra of A generated by B. Then C ⊆ B. P P For c ∈ C, we surely have c ⊆ bξ if ξ ∈ f (c); but also, because C is dist, c ∩ bξ = 0 if ξ ∈ λ \ f (c). Consequently c∗ = inf ξ∈f (c) bξ ∩ inf ξ∈λ\f (c) (1 \ bξ ) includes c. On the other hand, if d is any other member of C, there is some ξ ∈ f (c)4f (d), so that d∗ ∩ c∗ ⊆ bξ ∩ (1 \ bξ ) = 0. Since sup C = 1, it follows that c = c∗ ; but c∗ ∈ B, so c ∈ B. Q Q For any c ∈ C, look at {b ∩ c : b ∈ B} ⊆ B. This is a closed subalgebra of Ac (314F(a-i)) including Bc , so must be the whole of Ac . Thus Ac ⊆ B for every c ∈ C. But sup C = 1, so a = supc∈C a ∩ c ∈ B for every a ∈ A, and A = B. Consequently τ (A) ≤ #(B) ≤ λ, and τ (A) = λ. 332T Proposition Let (A, µ ¯) be a localizable measure algebra and B a closed subalgebra of A. Then (a) there is a function ν¯ : B → [0, ∞] such that (B, ν¯) is a localizable measure algebra; (b) τ (B) ≤ τ (A). proof (a) Let D be the set of those b ∈ B such that the principal ideal Bb has Maharam type at most τ (A) and is a totally finite measure algebra when endowed with an appropriate measure. Then D is order-dense in B. P P Take any non-zero b0 ∈ B. Then there is an a ∈ A such that a ⊆ b0 and 0 < µ ¯a < ∞. Set c = inf{b : b ∈ B, a ⊆ b}; then c ∈ B and a ⊆ c ⊆ b0 . If 0 6= b ∈ Bc , then c \ b belongs to B and is properly included in c, so cannot include a; accordingly a ∩ b 6= 0. For b ∈ Bc , set ν¯b = µ ¯(a ∩ b). Because the

140

Maharam’s theorem

332T

map b 7→ a ∩ b is an injective order-continuous Boolean homomorphism, ν¯ is countably additive and strictly positive, that is, (Bc , ν¯) is a measure algebra. It is totally finite because ν¯c = µ ¯a < ∞. Let d ∈ Bc \ {0} be such that Bd is Maharam-type-homogeneous; suppose that its Maharam type is κ. The map b 7→ b ∩ a is a measure-preserving Boolean homomorphism from Bd to Aa∩d , so by 332O Aa∩d must have a non-zero Maharam-type-κ0 component for some κ0 ≥ κ; but this means that τ (Bd ) ≤ κ ≤ κ0 ≤ τ (Aa∩d ) ≤ τ (A). Thus d ∈ D, while 0 6= d ⊆ c ⊆ b0 . As b0 is arbitrary, D is order-dense. Q Q Accordingly there is a partition of unity C in B such that C ⊆ D. For each c ∈ C we have a functional ν¯c such that (B , ν¯c ) is a totally finite measure algebra of Maharam type at most τ (A); define ν¯ : B → [0, ∞] by setting ν¯b = c∈C ν¯c (b ∩ c) for every b ∈ B. It is easy to check that (B, ν¯) is a measure algebra (compare 322Ka); it is localizable because B (being order-closed in a Dedekind complete partially ordered set) is Dedekind complete. (b) The construction above ensures that every homogeneous principal ideal of B can have Maharam type at most τ (A), since it must share a principal ideal with some Bc for c ∈ C. Moreover, any dist set in B is also a dist set in A, so c(B) ≤ c(A). So 332S tells us that τ (B) ≤ τ (A). Remark I think the only direct appeal I shall make to this result will be when (A, µ ¯) is a probability algebra, in which case (a) above becomes trivial, and the proof of (b) can be shortened to some extent, though I think we still need some of the ideas of 332S. 332X Basic exercises (a) Let A be a Dedekind complete Boolean algebra. Show that it is isomorphic to a simple product of Maharam-type-homogeneous Boolean algebras. (b) Let A be a Boolean algebra of finite cellularity. Show that A is purely atomic. (c) Let A be a purely atomic Boolean algebra. Show that c(A) is the number of atoms in A. (d) Let A be any Boolean algebra, and Z its Stone space. Show that c(A) is equal to c(Z) = sup{#(G) : G is a dist family of non-empty open subsets of Z}, the cellularity of the topological space Z. (e) Let X be a topological space, and G its regular open algebra. Show that c(G) = c(X) as defined in 332Xd. (f ) Let A be a Boolean algebra, and B any subalgebra of A. Show that c(B) ≤ c(A). (g) Let hAi ii∈I be any family of Boolean algebras, with simple product A. Show that the cellularity of A is at most max(ω, #(I), supi∈I c(Ai )). Devise an elegant expression of a necessary and sufficient condition for equality. (h) Let A be any Boolean algebra, and a ∈ A; let Aa be the principal ideal generated by a. Show that c(Aa ) ≤ c(A). (i) Let (A, µ ¯) be a semi-finite measure algebra. Show that it has a partition of unity of cardinal c(A). (j) Let (A, µ ¯) and (B, ν¯) be localizable measure algebras. For each cardinal κ let eκ , fκ be their Maharamtype-κ components, and Aeκ , Bfκ the corresponding principal ideals. Show that A and B are isomorphic, as Boolean algebras, iff c(Aeκ ) = c(Bfκ ) for every κ. P(k) Let ζ be P an ordinal, and hαξ iξ<ζ , hβξ iξ<ζ two families of non-negative real numbers such that α ≤ < ∞ for every θ ≤ ζ. Show that there is a family hγξη iξ≤η<ζ of non-negative real ξ θ≤ξ<ζ θ≤η<ζ βηP P numbers such that αξ = ξ≤η<ζ γξη for every ξ < ζ, βη ≥ ξ≤η γξη for every η < ζ. (If only finitely many ´ s 79, of the αξ , βξ are non-zero, this is an easy special case of the max-flow min-cut theorem; see Bolloba §III.1 or Anderson 87, 12.3.1.); there is a statement of the theorem in 4A3M in the next volume.) Show that γξη can be chosen in such a way that if ξ < ξ 0 , η 0 < η then at least one of γξη , γξ0 η0 is zero.

332 Notes


141

(l) Use 332Xk and 332M to give another proof of 332P. (m) For each cardinal κ, write (Bκ , ν¯κ ) for the measure algebra of the usual measure on {0, 1}κ . Let (A, µ ¯) be the simple product of h(Bωn , ν¯ωn )in∈N and (B, ν¯) the simple product of (A, µ ¯) with (Bωω , ν¯ωω ). (See 3A1E for the notation ωn , ωω .) Show that there is a measure-preserving Boolean homomorphism from A to B, but that no such homomorphism can be order-continuous. (n) For each cardinal κ, write (Bκ , ν¯κ ) for the measure algebra of the usual measure on {0, 1}κ . Let (A, µ ¯) be the simple product of h(Bκn , ν¯κn )in∈N and (B, ν¯) the simple product of h(Bλn , ν¯λn )in∈N , where κn = ω for even n, ωn for odd n, while λn = ω for odd n, ωn for even n. Show that there are order-continuous measure-preserving Boolean homomorphisms from A to B and from B to A, but that these two measure algebras are not isomorphic. (o) Let C be a Boolean algebra. Show that the following are equiveridical: (i) C is isomorphic (as Boolean algebra) to a closed subalgebra of a localizable measure algebra; (ii) there is a µ ¯ such that (C, µ ¯) is itself a localizable measure algebra; (iii) C is Dedekind complete and for every non-zero c ∈ C there is a completely additive real-valued functional ν on C such that νc 6= 0. (Hint for (iii)⇒(ii): show that the set of s of non-negative completely additive functionals is order-dense in C, so includes a partition of unity.) 332Y Further exercises (a) Let (A, µ ¯), (B, ν¯) be atomless localizable measure algebras. For each infinite cardinal κ let eκ , fκ be their Maharam-type-κ components. Show that the following are equiveridical: (i) (A, µ ¯) is isomorphic to a closed subalgebra of a principal ideal of (B, ν¯); (ii) for every cardinal λ, the magnitude of supκ≥λ eκ is not greater than the magnitude of supκ≥λ fκ . b µ b νˆ) their localizations (b) Let (A, µ ¯) and (B, ν¯) be any semi-finite measure algebras, and (A, ˆ), (B, (322O-322P). Let hei ii∈I , hfj ij∈J be partitions of unity in A, B respectively into elements of finite measure generating homogeneous principal ideals Aei , Bfj . For each infinite cardinal κ set Iκ = {i : τ (Aei ) = κ}, Jκ = {j : τ (Bfj ) = κ}; for γ ∈ ]0, ∞[, set Iγ = {i : ei is an atom, µ ¯ei = γ}, Jγ = {j : fj is an atom, P P b b ¯ei = j∈Ju ν¯fj < ∞ or ν¯f = γ}. Show that (A, µ ˆ) and (B, νˆ) are isomorphic iff for each u, either i∈Iu µ P Pj ¯ei = j∈Ju ν¯fj = ∞ and #(Iu ) = #(Ju ). i∈Iu µ (c) Let (A, µ ¯) and (B, ν¯) be non-zero localizable measure algebras; let eκ , fκ be their Maharam-type-κ components. Show that the following are equiveridical: (i) A is isomorphic, as Boolean algebra, to an orderclosed subalgebra of a principal ideal of B; (ii) c(A∗λ ) ≤ c(B∗λ ) for every cardinal λ, where A∗λ , B∗λ are the principal ideals generated by supκ≥λ eκ and supκ≥λ fκ respectively. 332 Notes and comments Maharam’s theorem tells us that all localizable measure algebras – in particular, all σ-finite measure algebras – can be obtained from the basic algebra A = {0, a, 1 \ a, 1}, with µ ¯a = µ ¯(1 \ a) = 1 , by combining the constructions of probability algebra free products, scalar multiples of measures and 2 simple products. But what is much more important is the fact that we get a description of our measure algebras in sufficiently explicit to make a very wide variety of questions resolvable. The description I offer in 332J hinges on the complementary concepts of ‘Maharam type’ and ‘magnitude’. If you like, the magnitude of a measure algebra is a measure of its width, while its Maharam type is a measure of its depth. The latter is more important just because, for localizable algebras, we have such a simple decomposition into algebras of finite magnitude. Of course there is a good deal of scope for further complications if we seek to consider non-localizable semi-finite algebras. For these, the natural starting point is a proper description of their localizations, which is not difficult (332Yb). Observe that 332C gives a representation of a localizable measure algebra as the measure algebra of a measure space which is completely different from the Stone representation in 321K. It is less canonical (since there is a degree of choice about the partition hei ii∈I ) but very much more informative, since the κi , γi carry enough information to identify the measure algebra up to isomorphism (332K). ‘Cellularity’ is the second cardinal function I have introduced in this chapter. It refers as much to topological spaces as to Boolean algebras (see 332Xd-332Xe). There is an interesting question in this context. If A is an arbitrary Boolean algebra, is there necessarily a dist set in A of cardinal c(A)? This

142

Maharam’s theorem

332 Notes

is believed to be undecidable from the ordinary axioms of set theory (including the axiom of choice); see ´ sz 71, 3.1 and 6.5. But for semi-finite measure algebras we have a definite answer (332F). Juha Maharam’s classification not only describes the isomorphism classes of localizable measure algebras, but also tells us when to expect Boolean homomorphisms between them (332P, 332Yc). I have given 332P only for atomless totally finite measure algebras because the non-totally-finite case (332Ya, 332Yc) seems to require a new idea, while atoms introduce acute combinatorial complications. I offer 332T as an example of the kind of result which these methods make very simple. It fails for general Boolean algebras; in fact, there is for any κ a countably τ -generated Dedekind complete Boolean algebra A with cellularity κ (Koppelberg 89, 13.1), so that Pκ is isomorphic to an order-closed subalgebra of A, and if κ > c then τ (Pκ) > ω (332R). For totally finite measure algebras we have a kind of weak Schröder-Bernstein theorem: if we have two of them, each isomorphic to a closed subalgebra of the other, they are isomorphic (332Q). This fails for σ-finite algebras (332Xn). I call it a ‘weak’ Schröder-Bernstein theorem because it is not clear how to build the isomorphism from the two injections; ‘strong’ Schröder-Bernstein theorems include definite recipes for constructing the isomorphisms declared to exist (see, for instance, 344D below).

333 Closed subalgebras Proposition 332P tells us, in effect, which totally finite measure algebras can be embedded as closed subalgebras of each other. Similar techniques make it possible to describe the possible forms of such embeddings. In this section I give the fundamental theorems on extension of measure-preserving homomorphisms from closed subalgebras (333C, 333D); these rely on the concept of ‘relative Maharam type’ (333A). I go on to describe possible canonical forms for structures (A, µ ¯, C), where (A, µ ¯) is a totally finite measure algebra and C is a closed subalgebra of A (333K, 333N). I end the section with a description of fixed-point subalgebras (333R). 333A Definitions (a) Let A be a Boolean algebra and C a subalgebra of A. The relative Maharam type of A over C, τC (A), is the smallest cardinal of any set A ⊆ A such that A ∪ C τ -generates A. (b) In this section, I will regularly use the following notation: if A is a Boolean algebra, C is a subalgebra of A, and a ∈ A, then I will write Ca for {c ∩ a : c ∈ C}. Observe that Ca is a subalgebra of the principal ideal Aa (because c 7→ c ∩ a : C → Aa is a Boolean homomorphism); it is included in C iff a ∈ C.. (c) Still taking A to be a Boolean algebra and C to be a subalgebra of A, I will say that an element a of A is relatively Maharam-type-homogeneous over C if τCb (Ab ) = τCa (Aa ) for every non-zero b ⊆ a. 333B Evidently this is a generalization of the ordinary concept of Maharam type as used in §§331-332; if C = {0, 1} then τC (A) = τ (A). The first step is naturally to check the results corresponding to 331H. Lemma Let A be a Boolean algebra and C a subalgebra of A. (a) If a ⊆ b in A, then τCa (Aa ) ≤ τCb (Ab ). In particular, τCa (Aa ) ≤ τC (A) for every a ∈ A. (b) The set {a : a ∈ A is relatively Maharam-type-homogeneous over C} is order-dense in A. (c) If A is Dedekind complete and C is order-closed in A, then Ca is order-closed in Aa . (d) If a ∈ A is relatively Maharam-type-homogeneous over C then either Aa = Ca , so that τCa (Aa ) = 0 and a is a relative atom of A over C (definition: 331A), or τCa (Aa ) ≥ ω. (e) If D is another subalgebra of A and D ⊆ C, then τ (Aa ) = τ{0,a} (Aa ) ≥ τDa (Aa ) ≥ τCa (Aa ) ≥ τAa (Aa ) = 0 for every a ∈ A. proof (a) Let D ⊆ Ab be a set of cardinal τCb (Ab ) such that D∪Cb τ -generates Ab . Set D0 = {d ∩ a : d ∈ D}. Then D0 ∪Ca τ -generates Aa . P P Apply 313Mc to the map d 7→ d ∩ a : Ab → Aa , as in 331Hc. Q Q Consequently τCa (Aa ) ≤ #(D0 ) ≤ #(D) = τCb (Ab ),

333C

Closed subalgebras

143

as claimed. Setting b = 1 we get τCa (Aa ) ≤ τC (A). (b) Just as in the proof of 332A, given b ∈ A \ {0}, there is an a ∈ Ab \ {0} minimising τCa (Aa ), and this a must be relatively Maharam-type-homogeneous over C. (c) Ca is the image of the Dedekind complete Boolean algebra C under the order-continuous Boolean homomorphism c 7→ c ∩ a, so must be order-closed (314Fa). (d) Suppose that τCa (Aa ) is finite. Let D ⊆ Aa be a finite set such that D ∪ Ca τ -generates Aa . Then there is a non-zero b ∈ Aa such that b ∩ d is either 0 or b for every d ∈ D. But this means that Cb = {d ∩ b : d ∈ D ∪ Ca }, which τ -generates Ab ; so that τCb (Ab ) = 0. Since a is relatively Maharam-type-homogeneous over C, τCa (Aa ) must be zero, that is, Aa = Ca . (e) The middle inequality is true just because Aa will be τ -generated by D ∪Ca whenever it is τ -generated by D ∪ Da . The neighbouring inequalities are special cases of the middle one, and the outer equalities are elementary. 333C Theorem Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras, and C a closed subalgebra of A. Let φ : C → B be a measure-preserving Boolean homomorphism. (a) If, in the notation of 333A, τC (A) ≤ τφ[C]b (Bb ) for every non-zero b ∈ B, there is a measure-preserving Boolean homomorphism π : A → B extending φ. (b) If τCa (Aa ) = τφ[C]b (Bb ) for every non-zero a ∈ A, b ∈ B, then there is a measure algebra isomorphism π : A → B extending φ. proof In both parts, the idea is to use the technique of the proof of 331I to construct π as the last of an increasing family hπξ iξ≤κ of measure-preserving homomorphisms from closed subalgebras Cξ of A, where κ = τC (A). Let haξ iξ<κ be a family in A such that C ∪ {aξ : ξ < κ} τ -generates A. Write D for φ[C]; that D is a closed subalgebra of B (324L). (a)(i) In this case, we can describe the Cξ immediately; Cξ will be the closed subalgebra of A generated by C ∪ {aη : η < ξ}. The induction starts with C0 = C, π0 = φ. (ii) For the inductive step to a successor ordinal ξ + 1, where ξ < κ, suppose that Cξ and πξ have been defined. Take any non-zero b ∈ B. We are supposing that τDb (Bb ) ≥ κ > #(ξ), so Bb cannot be τ -generated by D = Db ∪ {b ∩ πξ aη : η < ξ} = πξ [C]b ∪ {b ∩ πξ aη : η < ξ} = ψ[C ∪ {aη : η < ξ}], writing ψc = b ∩ πξ c for c ∈ Cξ . As ψ is order-continuous, ψ[Cξ ] is precisely the closed subalgebra of Bb generated by D (314Gb), and is therefore not the whole of Bb . But this means that Bb 6= {b ∩ πξ c : c ∈ Cξ }. As b is arbitrary, πξ satisfies the conditions of 331D, and has an extension to a measure-preserving Boolean homomorphism πξ+1 : Cξ+1 → B, since Cξ+1 is just the closed subalgebra of A generated by C ∪ {aξ }. (iii) For the inductive step to a non-zero limit ordinal ξ ≤ κ, we can argue exactly as in part (d) of the S proof of 331I; Cξ will be the metric closure of C∗ξ = η<ξ Cη , so we can take πξ : Cξ → B to be the unique S measure-preserving homomorphism extending πξ∗ = η<ξ πη . Thus the induction proceeds, and evidently π = πκ will be a measure-preserving homomorphism from A to B extending φ. (b) (This is rather closer to the proof of 331I, being indeed a direct generalization of it.) Observe that the hypothesis (b) implies that 1A is relatively Maharam-type-homogeneous over C; so either κ = 0, in which case A = C, B = φ[C] and the result is trivial, or κ ≥ ω, by 333Bd. Let us therefore take it that κ is infinite. We are supposing, among other things, that τD (B) = κ; let hbξ iξ<κ be a family in B such that B is τ -generated by D ∪ {bξ : ξ < κ}. This time, as in 331I, we shall have to choose further families ha0ξ iξ<κ and hb0ξ iξ<κ , and Cξ will be the closed subalgebra of A generated by C ∪ {aη : η < ξ} ∪ {a0η : η < ξ}, Dξ will be the closed subalgebra of B generated by

144

Maharam’s theorem

333C

D ∪ {bη : η < ξ} ∪ {b0η : η < ξ}, πξ : Cξ → Dξ will be a measure-preserving homomorphism. The induction will start with C0 = C, D0 = D and π0 = φ, as in (a). (i) For the inductive step to a successor ordinal ξ + 1, where ξ < κ, suppose that Cξ , Dξ and πξ have been defined. α) Let b ∈ B \ {0}. Because (α τDb (Bb ) = κ > #({bη : η < ξ} ∪ {b0η : η < ξ}), Bb cannot be τ -generated by Db ∪ {b ∩ bη : η < ξ} ∪ {b ∩ b0η : η < ξ}, and cannot be equal to {b ∩ d : d ∈ Dξ }. As b is arbitrary, there is an extension of πξ to a measure-preserving homomorphism φξ from C0ξ to B, where C0ξ is the closed subalgebra of A generated by C ∪ {aη : η ≤ ξ} ∪ {a0η : η < ξ}. Setting b0ξ = φξ (aξ ), the image D0ξ = φξ [Cξ ] will be the closed subalgebra of B generated by D ∪ {bη : η < ξ} ∪ {b0η : η ≤ ξ}. β ) Now, as in 331I, we must repeat the argument of (α), applying it now to φ−1 (β : Dξ → A. If ξ a ∈ A \ {0}, τCa (Aa ) = κ > #({aη : η ≤ ξ} ∪ {a0η : η < ξ}), so that Aa cannot be {a ∩ c : c ∈ C0ξ }. As a is arbitrary, φ−1 has an extension to a measure-preserving ξ homomorphism ψξ : Dξ+1 → Cξ+1 , where Dξ+1 is the subalgebra of B generated by D0ξ ∪ {bξ }, that is, the closed subalgebra of B generated by D ∪ {bη : η ≤ ξ} ∪ {b0η : η < ξ}, and Cξ+1 is the subalgebra of A generated by C0ξ ∪ {a0ξ }, setting a0ξ = ψξ (bξ ). We can therefore take πξ+1 = ψξ−1 : Cξ+1 → Dξ+1 , as in 331I. (ii) The inductive step to a S non-zero limit ordinal ξ ≤ κ is exactly the same S as in (a) above or in 331I; Cξ is the metric closure of C∗ξ = η<ξ Cη , Dξ is the metric closure of D∗ξ = η<ξ Dη , and πξ is the unique measure-preserving homomorphism from Cξ to Dξ extending every πη for η < ξ. (iii) The induction stops, as before, with π = πκ : Cκ → Dκ , where Cκ = A, Dκ = B. 333D Corollary Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras and C a closed subalgebra of A. Suppose that τ (C) < max(ω, τ (A)) ≤ min{τ (Bb ) : b ∈ B \ {0}}. Then any measure-preserving Boolean homomorphism φ : C → B can be extended to a measure-preserving Boolean homomorphism π : A → B. proof Set κ = min{τ (Bb ) : b ∈ B \ {0}}. Then for any non-zero b ∈ B, τφ[C]b (Bb ) ≥ κ. P P There is a set C ⊆ C, of cardinal τ (C), which τ -generates C, so that C 0 = {b ∩ φc : c ∈ C} τ -generates φ[C]b . Now there is a set D ⊆ Bb , of cardinal τφ[C]b (Bb ), such that φ[C]b ∪ D τ -generates Bb . In this case C 0 ∪ D must τ -generate Bb , so κ ≤ #(C 0 ∪ D). But #(C 0 ) ≤ #(C) < κ and κ is infinite, so we must have #(D) ≥ κ, as claimed. Q Q On the other hand, τC (A) ≤ τ (A) ≤ κ. So we can apply 333C to give the result. 333E Theorem Let (C, µ ¯) be a totally finite measure algebra, κ an infinite cardinal, and (Bκ , ν¯κ ) the ¯ be the localizable measure algebra free product measure algebra of the usual measure on {0, 1}κ . Let (A, λ) of (C, µ ¯) and (Bκ , ν¯κ ), and ψ : C → A the corresponding homomorphism. Then for any non-zero a ∈ A, τψ[C]a (Aa ) = κ, in the notation of 333A above. proof Recall from 325Dd that ψ[C] is a closed subalgebra of A. (a) Let hbξ iξ<κ be the canonical independent family in Bκ of elements of measure 21 . Let ψ 0 : Bκ → A be the canonical map, and set b0ξ = ψ 0 bξ for each ξ.

333E

Closed subalgebras

145

We know that {bξ : ξ < κ} τ -generates Bκ (see part (a) of the proof of 331K). Consequently ψ[C] ∪ {b0ξ : ξ < κ} τ -generates A. P P Let A1 be the closed subalgebra of A generated by ψ[C] ∪ {b0ξ : ξ < κ}. Because 0 ψ : Bκ → A is order-continuous (325Da), ψ 0 [Bκ ] ⊆ A1 (313Mb). But this means that A1 includes ψ[C] ∪ ψ 0 [Bκ ] and therefore includes the image of C ⊗ Bκ in A; because this is topologically dense in A (325Dc), A1 = A, as claimed. Q Q (b) It follows that τψ[C]a Aa ≤ τψ[C] A ≤ κ. (c) We need to know that if ξ < κ and e belongs to the closed subalgebra Eξ of A generated by ¯ ∩ b0 ) = 1 λe. ¯ P ψ[C] ∪ {b0η : η 6= ξ}, then λ(e P Set ξ 2 E = ψ[C] ∪ {b0η : η 6= ξ},

F = {e0 ∩ . . . ∩ en : e0 , . . . , en ∈ E}.

Then every member of F is expressible in the form d = ψa ∩ inf η∈J b0η , where a ∈ C and J ⊆ κ \ {ξ} is finite. Now ¯ =µ λd ¯a · ν¯(inf η∈J bη ) = 2−#(J) µ ¯a, 1¯ ¯ 0 ∩ d) = µ λ(b ¯a · ν¯(bξ ∩ inf η∈J bη ) = 2−#(J∪{ξ}) µ ¯a = λd. ξ 2

Now consider the set ¯ ξ ∩ d) = 1 λd}. ¯ G = {d : d ∈ A, λ(b 2

We have 1A ∈ F ⊆ G, and F is closed under ∩ . Secondly, if d, d0 ∈ G and d ⊆ d0 , then ¯ ξ ∩ (d0 \ d)) = λ(b ¯ ξ ∩ d0 ) − λ(b ¯ ξ ∩ d) = 1 λd ¯ 0 − 1 λd ¯ = 1 λ(d ¯ 0 \ d), λ(b 2

2

2

0

so d \ d ∈ G. Also, if H ⊆ G is non-empty and upwards-directed, 1¯ 1¯ ¯ ξ ∩ sup H) = λ(sup ¯ ¯ λ(b d∈H bξ ∩ d) = supd∈H λ(bξ ∩ d) = supd∈H λd = λ(sup H), 2

2

so sup H ∈ G. By the Monotone Class Theorem (313Gc), G includes the order-closed subalgebra of D generated by F . But this is just Eξ . Q Q (d) The next step is to see that τψ[C]a (Aa ) > 0. P P By (a) and 323J, A is the metric closure of the 0 ¯ 0 4 a) ≤ 1 λa. ¯ subalgebra A0 generated by ψ[C] ∪ {bη : η < κ}, so there must be an a0 ∈ A0 such that λ(a 4 0 Now there is a finite J ⊆ κ such that a0 belongs to the subalgebra A1 generated by ψ[C] ∪ {bη : η ∈ J}. Take any ξ ∈ κ \ J (this is where I use the hypothesis that κ is infinite). If c ∈ C, then by (c) we have 0 ¯ ¯ ∩ (ψc 4 b0 )) ≥ λ(a ¯ 0 ∩ (ψc 4 b0 )) − λ(a ¯ 4 a0 ) λ((a ∩ ψc) 4 (a ∩ bξ )) = λ(a ξ ξ ¯ 0 ∩ b0 ) + λ(a ¯ 0 ∩ ψc) − 2λ(a ¯ 0 ∩ ψc ∩ b0 ) − λ(a ¯ 4 a0 ) = λ(a ξ

=

1¯ λa0 2

ξ

¯ 4 a0 ) − λ(a

(because both a0 and a0 ∩ ψc belong to Eξ ) 3¯ 1¯ − λ(a 4 a0 ) > 0. ≥ λa 2

2

Thus a ∩ b0ξ is not of the form a ∩ ψc for any c ∈ C, and Aa 6= ψ[C]a , so that τψ[C]a (Aa ) > 0. Q Q (e) It follows that τψ[C]a (Aa ) is infinite. P P There is a non-zero d ⊆ a which is relatively Maharam-typehomogeneous over ψ[C]. By (d), applied to d, τψ[C]d (Ad ) > 0; but now 333Bd tells us that τψ[C]d (Ad ) must be infinite, so τψ[C]a (Aa ) is infinite. Q Q

146

Maharam’s theorem

333E

(f ) If κ = ω, we can stop here. If κ > ω, we continue, as follows. Let D ⊆ Aa be any set of cardinal less than κ. Each d ∈ D ∪ {a} belongs to the closed subalgebra of A generated by C = ψ[C] ∪ {b0ξ : ξ < κ}. But because A is ccc, this is just the σ-subalgebra of A generated by C (331Ge). So d belongs to the closed subalgebra ofSA generated by some countable subset Cd of C, by 331Gd. Now Jd = {η : b0η ∈ Cd } is countable. Set J = d∈D∪{a} Jd ; then #(J) ≤ max(ω, #(D ∪ {a})) = max(ω, #(D)) < κ, so J 6= κ, and there is a ξ ∈ κ \ J. Accordingly ψ[C] ∪ D ∪ {a} is included in Eξ , as defined in (c) above, and ψ[C]a ∪ D ⊆ Eξ . As Aa ∩ Eξ is a closed subalgebra of Aa , it includes the closed subalgebra generated by ψ[C]a ∪ D. But a ∩ b0ξ surely does not belong to Eξ , since ¯ > 0, ¯ ∩ b0 ∩ b0 ) = λ(a ¯ ∩ b0 ) = 1 λa λ(a ξ ξ ξ 2

¯ ∩ b0 ∩ b0 ) 6= 1 λ(a ¯ ∩ b0 ). Thus a ∩ b0 cannot belong to the closed subalgebra of Aa generated by and λ(a ξ ξ ξ ξ 2 ψ[C]a ∪ D, and ψ[C]a ∪ D does not τ -generate Aa . As D is arbitrary, τφ[C]a (Aa ) ≥ κ. This completes the proof.

333F Corollary Let (A, µ ¯) be a totally finite measure algebra, C a closed subalgebra of A and κ an infinite cardinal. Let (Bκ , ν¯κ ) be the measure algebra of the usual measure on {0, 1}κ . ¯ be the localizable measure algebra free product of (C, µ b κ , λ) (i) Suppose that κ ≥ τC (A). Let (C⊗B ¯¹ C) b κ the corresponding homomorphism. Then there is a measure-preserving and (Bκ , ν¯κ ), and ψ : C → C⊗B b κ extending ψ. Boolean homomorphism π : A → C⊗B (ii) Suppose further that κ = τCa (Aa ) for every non-zero a ∈ A. Then π can be taken to be an isomorphism. b κ , using 333E to see that the hypothesis proof All we have to do is apply 333C with B = C⊗B τψ[C]b (Bb ) = κ for every non-zero b ∈ B is satisfied.

333G Corollary Let (C, µ ¯) be a totally finite measure algebra. Suppose that κ ≥ max(ω, τ (C)) is a ¯ be the b κ , λ) cardinal, and write (Bκ , ν¯κ ) for the measure algebra of the usual measure on {0, 1}κ . Let (C⊗B localizable measure algebra free product of (C, µ ¯) and (Bκ , ν¯κ ). Then b κ is Maharam-type-homogeneous, with Maharam type κ; (a) C⊗B (b) for every measure-preserving Boolean homomorphism φ : C → C there is a measure-preserving autob κ → C⊗B b κ such that π(c ⊗ 1) = φc ⊗ 1 for every c ∈ C, writing c ⊗ 1 for the canonical morphism π : C⊗B b κ of any c ∈ C. image in C⊗B b κ , as in 333E. proof Write A for C⊗B (a) If C ⊆ C is a set of cardinal τ (C) which τ -generates C, and B ⊆ Bκ a set of cardinal κ which τ -generates Bκ (331K), then {c ⊗ b : c ∈ C, b ∈ B} is a set of cardinal at most max(ω, τ (C), κ) = κ which τ -generates A (because the subalgebra it generates is topologically dense in A, by 325Dc). So τ (A) ≤ κ. On the other hand, if a ∈ A is non-zero, then τ (Aa ) ≥ τψ[C]a (Aa ) ≥ κ, by 333E; so A is Maharam-typehomogeneous, with Maharam type κ. (b) Writing D = {c ⊗ 1 : c ∈ C} for the canonical image of C in A, we have a measure-preserving automorphism φ1 : D → D defined by setting φ1 (c ⊗ 1) = φc ⊗ 1 for every c ∈ C. Because φ1 [D] ⊆ D, 333Be and 333E tell us that κ = τ (Aa ) ≥ τφ1 [D]a (Aa ) ≥ τDa (Aa ) = κ for every non-zero a ∈ A, so we can use 333Cb, with B = A, to see that φ1 can be extended to a measurepreserving automorphism on A.

333I

333H

Closed subalgebras

147

I turn now to the classification of closed subalgebras.

Theorem Let (A, µ ¯) be a localizable measure algebra and C a closed subalgebra of A. Then there are hµi ii∈I , hci ii∈I , hκi ii∈I such that for each i ∈ I, µi is a non-negative completely additive functional on C, ci = [[µi > 0]] ∈ C, κi is 0 or an infinite cardinal, (Cci , µi ¹ Cci ) is a totally finite measure algebra, writing Cci for the principal ideal of C generated by ci , P ¯c for every c ∈ C, i∈I µi c = µ Q b κi there is a measure-preserving isomorphism π from A to the simple product i∈I Cci ⊗B b κi of (Cci , µi ¹ Cci ), (Bκi , ν¯κi ), writing of the localizable measure algebra free products Cci ⊗B (Bκ , ν¯κ ) for the measure algebra of the usual measure on {0, 1}κ . Moreover, π may be taken such that b κi of c ∈ Cci . for every c ∈ C, πc = h(c ∩ ci ) ⊗ 1ii∈I , writing c ⊗ 1 for the image in Cci ⊗B Remark Recall that [[µi > 0]] is that element of C such that µi c > 0 whenever c ∈ C and 0 6= c ⊆ [[µi > 0]], µi c ≤ 0 whenever c ∈ C and c ∩ [[µi > 0]] = 0 (326O). proof (a) Let A be the set of those elements of A which are relatively Maharam-type-homogeneous over C (see 333Ac). By 333Bb, A is order-dense in A (compare part (a) of the proof of 332B), and consequently A0 = {a : a ∈ A, µ ¯a < ∞} is order-dense in A. So there is a partition of unity hai ii∈I in A consisting of of A0 (313K). For each i ∈ I, set µi c = µ ¯(ai ∩ c) for every c ∈ C; then µi is non-negative, and it is completely additive by 327E. Because hai ii∈I is a partition of the identity in A, P P ¯(c ∩ ai ) = i∈I µi c µ ¯c = i∈I µ for every c ∈ C. Next, (Cci , µi ¹ Cci ) is a totally finite measure algebra. P P Cci is a Dedekind σ-complete Boolean algebra because C is. µi ¹ Cci is a non-negative countably additive functional because µi is. If c ∈ Cci and µi c = 0, then c = 0 by the choice of ci . Q Q Note also that µ ¯(ai \ ci ) = µi (1 \ ci ) = 0, so that ai ⊆ ci . (b) By 333Bd, any finite κi must actually be zero. The next element we need is the fact that, for each i ∈ I, we have a measure-preserving isomorphism c 7→ c ∩ ai from (Cci , µi ¹ Cci ) to (Cai , µ ¯¹ Cai ). P P Of course this is a ring homomorphism. Because ai ⊆ ci , it is a surjective Boolean homomorphism. It is measure-preserving by the definition of µi , and therefore injective. Q Q (c) Still focusing on a particular i ∈ I, let Aai be the principal ideal of A generated by ai . Then b κi , extending the canonical homomorphism we have a measure-preserving isomorphism π ˜i : Aai → Cai ⊗B b κi . P c 7→ c ⊗ 1 : Cai → Cai ⊗B P When κi is infinite, this is just 333F(ii). But the only other case is when b κi ∼ κi = 0, that is, Cai = Aai , while Bκi = {0, 1} and Cai ⊗B Q = Cci . Q b κi and The isomorphism between (Cci , µi ¹ Cci ) and (Cai , µ ¯¹ Cai ) induces an isomorphism between Cci ⊗B b κi . So we have a measure-preserving isomorphism πi : Aai → Cci ⊗B b κi such that πi (c ∩ ai ) = c ⊗ 1 Cai ⊗B for every c ∈ Cci . Q (d) By 322Kd, we have a measure-preserving isomorphism a 7→ ha ∩ ai ii∈I : A → i∈I Aai . Q Putting this together with the isomorphisms of (c), we have a measure-preserving isomorphism π : A → b i∈I Cci ⊗Bκi , setting πa = hπi (a ∩ ai )ii∈I for a ∈ A. Observe that, for c ∈ C, πc = hπi (c ∩ ai )ii∈I = h(c ∩ ci ) ⊗ 1ii∈I , as required. 333I Remarks (a) I hope it is clear that whenever (C, µ ¯) is a Dedekind complete measure algebra, P hµi ii∈I is a family of non-negative completely additive functionals on C such that i∈I µi = µ ¯, and hκi ii∈I

148

Maharam’s theorem

333I

is a family of cardinals all infinite or zero, then the construction above can be applied to give a measure alb κi ii∈I , together with an order-continuous measure-preserving gebra (A, µ ˜), the product of the family hCci ⊗B homomorphism π : C → A; and that the partition of unity hai ii∈I in A corresponding to this product (315E) b κi , has µi c = µ ˜(ai ∩ πc) for every c ∈ C and i ∈ I, while each principal ideal Aai can be identified with Cci ⊗B so that ai is relatively Maharam-type-homogeneous over π[C]. Thus any structure (C, µ ¯, hµi ii∈I , hκi ii∈I ) of the type described here corresponds to an embedding of C as a closed subalgebra of a localizable measure algebra. (b) The obvious next step is to seek a complete classification of objects (A, µ ¯, C), where (A, µ ¯) is a localizable measure algebra and C is a closed subalgebra, corresponding to the classification of localizable measure algebras in of the magnitudes of their Maharam-type-κ components in 332J. The general case seems to be complex. But I can deal with the special case in which (A, µ ¯) is totally finite. In this case, we have the following facts. 333J Lemma Let (A, µ ¯) be a totally finite measure algebra, and C a closed subalgebra. Let A be the set of relative atoms of A over C. Then there is a unique sequence hµn in∈N of additive functionals on C such that (i) µn+1 ≤ µn for every n (ii) there is a dist sequence han in∈N in A such that supn∈N an = sup A and µn c = µ ¯(an ∩ c) for every n ∈ N, c ∈ C. Remark I hope it is plain from my wording that it is the µn which are unique, not the an . proof (a) For each a ∈ A set νa (c) = µ ¯(c ∩ a) for c ∈ C. Then νa is a non-negative completely additive real-valued functional on C (see 326Kd). The key step is I suppose in (c) below; I approach by a two-stage argument. For each b ∈ A write Ab for {a : a ∈ A, a ∩ b = 0}. (b) For every b ∈ A, non-zero c ∈ C there are a ∈ Ab , c0 ∈ C such that 0 6= c0 ⊆ c and νa (d) ≥ νe (d) whenever d ∈ C, e ∈ Ab and d ⊆ c0 . P P?? Otherwise, choose han in∈N and hcn in∈N as follows. Since 0, c won’t serve for a, c0 , there must be an a0 ∈ Ab such that νa0 (c) > 0. Let δ > 0 be such that νa0 (c) > δ µ ¯c and set c0 = c ∩ [[νa0 > δ µ ¯¹ C]]; then c0 ∈ C and 0 6= c0 ⊆ c. Given that an ∈ Ab , cn ∈ C and 0 6= cn ⊆ c, then there must be an+1 ∈ Ab , dn ∈ C such that dn ⊆ cn and νan+1 (dn ) > νan (dn ). Set cn+1 = dn ∩ [[νan+1 > νan ]], so that cn+1 ∈ C and 0 6= cn+1 ⊆ cn , and continue. There is some n ∈ N such that nδ ≥ 1. For any i < n, the construction ensures that 0 6= cn+1 ⊆ ci+1 ⊆ [[νai+1 > νai ]], so νai (cn+1 ) < νai+1 (cn+1 ); also cn+1 ⊆ c0 so Pn−1

µ ¯(ai ∩ cn+1 ) = νai (cn+1 ) ≥ νa0 (cn+1 ) > δ µ ¯cn+1 .

But this means that i=0 µ ¯(ai ∩ cn+1 ) > µ ¯cn+1 and there must be distinct j, k < n such that aj ∩ ak ∩ cn+1 is non-zero. Because aj , ak ∈ A there are d0 , d00 ∈ C such that aj ∩ ak = aj ∩ d0 = ak ∩ d00 ; set d = cn+1 ∩ d0 ∩ d00 , so that d ∈ C and aj ∩ d = aj ∩ ak ∩ cn+1 = ak ∩ d,

νaj (d) = µ ¯(aj ∩ ak ∩ cn+1 ) = νak (d).

But as 0 6= d ⊆ [[νai+1 > νai ]] for every i < n, νa0 (d) < νa1 (d) < . . . < νan (d), so this is impossible. X XQ Q (c) Now for a global, rather than local, version of the same idea. For every b ∈ A there is an a ∈ Ab such that and νa ≥ νe whenever e ∈ Ab . P P (i) By (b), the set C of those c ∈ C such that there is an a ∈ Ab such that νa ¹ Cc ≥ νe ¹ Cc for every e ∈ Ab is order-dense in C. Let hci ii∈I be a partition of unity in C consisting of of C, and for each i ∈ I choose ai ∈ Ab such that νai ¹ Cci ≥ νe ¹ Cci for every e ∈ Ab . Consider a = supi∈I ai ∩ ci . (ii) If a0 ∈ A and a0 ⊆ a, then for each i ∈ I there is a di ∈ C such that ai ∩ a0 = ai ∩ di . Set d0 = supi∈I ci ∩ di ; then (because hci ii∈I is dist) a ∩ d0 = supi∈I ai ∩ ci ∩ di = supi∈I ai ∩ ci ∩ a0 = a ∩ a0 = a0 . As a0 is arbitrary, this shows that a ∈ A. (iii) Of course a ∩ b = 0, so a ∈ Ab . Now take any e ∈ Ab , d ∈ C. Then P P νa (d) = i∈I νai (ci ∩ d) ≥ i∈I νe (ci ∩ d) = νe (d).

333K

Closed subalgebras

149

So this a has the required property. Q Q (d) Choose han in∈N inductively in A so that, for each n, an ∩ supi 0]] there is some k ∈ N such that c ∩ a0i ∩ ak 6= 0; this is because µ ¯(a0i ∩ c) = µ0i (c) > 0, so that a0i ∩ c 6= 0, while a0i ⊆ sup A = supk∈N ak . ?? Suppose, if possible, that there is some n such that µn 6= µ0n ; since the situation is symmetric, there is no loss of generality in supposing that µ0n 6≤ µn , that is, that c = [[µ0n > µn ]] 6= 0. For any i ≤ n, µ0i ≥ µ0n so c ⊆ [[µ0i > 0]]. We may therefore choose c0 , . . . , cn+1 ∈ Cc \ {0} and k(0), . . . , k(n) ∈ N such that c0 = c and, for i ≤ n, ci ∩ a0i ∩ ak(i) 6= 0 (choosing k(i), recalling that 0 6= ci ⊆ c ⊆ [[µ0i > 0]]), ci+1 ∈ C, ci+1 ⊆ ci , ci+1 ∩ a0i = ci+1 ∩ ak(i) = ci ∩ a0i ∩ ak(i) (choosing ci+1 , using the fact that a0i and ak(i) both belong to A – see the penultimate sentence in part (b) of the proof.) On reaching cn+1 , we have 0 6= cn+1 ⊆ c so µn (cn+1 ) < µ0n (cn+1 ). On the other hand, for each i ≤ n, cn+1 ∩ a0i ∩ ak(i) = cn+1 ∩ ci+1 ∩ a0i ∩ ak(i) = cn+1 ∩ a0i = cn+1 ∩ ak(i) , so µn (cn+1 ) < µ0n (cn+1 ) ≤ µ0i (cn+1 ) = µ ¯(cn+1 ∩ a0i ) = µ ¯(cn+1 ∩ ak(i) ) = µk(i) (cn+1 ), and k(i) must be less than n. There are therefore distinct i, j ≤ n such that k(i) = k(j). But in this case cn+1 ∩ a0i = cn+1 ∩ ak(i) = cn+1 ∩ ak(j) = cn+1 ∩ a0j 6= 0 because 0 6= cn+1 ⊆ [[µ0j > 0]]. So a0i , a0j cannot be dist, breaking one of the rules of the construction. X X Thus µn = µ0n for every n ∈ N. This completes the proof. 333K Theorem Let (A, µ ¯) be a totally finite measure algebra and C a closed subalgebra of A. Then there are unique families hµn in∈N , hµκ iκ∈K such that K is a countable set of infinite cardinals, P for i ∈ N ∪ K, µi is a non-negative countably additive functional on C, and i∈N∪K µi c = µ ¯c for every c ∈ C, µn+1 ≤ µn for every n ∈ N, and µκ 6= 0 for κ ∈ K, setting ei = [[µi > 0]] ∈ C, and giving the principal ideal Cei generated by ei the measure µi ¹ Cei for each i ∈ N ∪ K, and writing (Bκ , ν¯κ ) for the measure algebra of the usual measure on {0, 1}κ for κ ∈ K, we have a measure algebra isomorphism Q Q b κ π : A → n∈N Cen × κ∈K Ceκ ⊗B such that πc = (hc ∩ en in∈N , h(c ∩ eκ ) ⊗ 1iκ∈K ) b κ of c ∈ Ceκ . for each c ∈ C, writing c ⊗ 1 for the canonical image in Ceκ ⊗B proof (a) I aim to use the construction of 333H, but taking much more care over the choice of hai ii∈I in part (a) of the proof there. We start by taking han in∈N as in 333J, and setting µn c = µ ¯(an ∩ c) for every n ∈ N, c ∈ C; then these an will deal with the part in sup A, as defined in the proof of 333J.

150

Maharam’s theorem

333K

(b) The further idea required here concerns the treatment of infinite κ. Let hbi ii∈I be any partition of unity in A consisting of non-zero of A which are relatively Maharam-type-homogeneous over C, and hκi ii∈I the corresponding cardinals, so that κi = 0 iff bi ∈ A. Set I1 = {i : i ∈ I, κi ≥ ω}. Set K = {κi : i ∈ I1 }, so that K is a countable set of infinite cardinals, and for κ ∈ K set Jκ = {i : κi = κ}, aκ = supi∈Jκ bi for κ ∈ K. Now every aκ is relatively Maharam-type-homogeneous over C. P P (Compare 332H.) Jκ must be countable, because A is ccc. If 0 6= a ⊆ aκ , there is some i ∈ Jκ such that a ∩ bi 6= 0; now τCa (Aa ) ≥ τCa∩bi (Aa∩bi ) = κi = κ. At the same time, S for each i ∈ Jκ , there is a set Di ⊆ Abi such that #(Di ) = κ and Cbi ∪ Di τ -generates Abi . Set D = i∈Jκ Di ∪ {bi : i ∈ Jκ }; then #(D) ≤ max(ω, #(Jκ ), supi∈K #(Di )) = κ. Let B be the closed subalgebra of Aaκ generated by Caκ ∪ D. Then Cbi ∪ Di ⊆ {b ∩ bi : b ∈ B} = B ∩ Abi so B ⊇ Abi for each i ∈ Jκ , and B = Aaκ . Thus Caκ ∪ D τ -generates Aaκ , and τCaκ (Aaκ ) ≤ κ ≤ min06=a ⊆ aκ τCa (Aa ). This shows that aκ is relatively Maharam-type-homogeneousover C, with τCaκ (Aaκ ) = κ. Q Q Since evidently hJκ iκ∈K and haκ iκ∈K are dist, and supκ∈K aκ = supi∈I1 bi , this process yields a partition hai ii∈N∪K of unity in A. Now the arguments of 333H show that we get an isomorphism π of the kind described. (c) To see that the families hµn in∈N , hµκ iκ∈K (and therefore the ei and the (Cei , µi ¹ Cei ), but not π) are uniquely defined, argue as follows. Take families h˜ µn in∈N , h˜ µκ iκ∈K˜ which correspond to an isomorphism Q Q b κ, π ˜ : A → D = n∈N Ceñ × κ∈K˜ Ce˜κ ⊗B Q ˜ In the simple product Q b κ , we have a partition writing e˜i = [[˜ µi > 0]] for i ∈ N ∪ K. ˜ Ce˜κ ⊗B n∈N Ceñ × κ∈K of unity he∗i ii∈N∪K˜ corresponding to the product structure. Now for d ⊆ e∗i , we have τπ˜ [C]d (Dd ) = 0 if i ∈ N, ˜ = κ if i = κ ∈ K. ˜ must be So K {κ : κ ≥ ω, ∃ a ∈ A, τCa (Aa ) = κ} = K, ˜ and for κ ∈ K, π ˜ −1 e∗κ = sup{a : a ∈ A, τCa (Aa ) = κ} = aκ , so that µ ˜κ = µκ . On the other hand, h˜ π −1 e∗n in∈N must be a dist sequence with supremum sup A, and the corresponding functionals µ ñ are supposed to form a non-increasing sequence, so must be equal to the µn by 333J. 333L Remark Thus for the classification of structures (A, µ ¯, C), where (A, µ ¯) is a totally finite measure algebra and C is a closed subalgebra, it will be enough to classify objects (C, µ ¯, hµn in∈N , hµκ iκ∈K ), where (C, µ ¯) is a totally finite measure algebra, hµn in∈N is a non-increasing sequence of non-negative countably additive functionals on C, K is a countable set of infinite cardinals (possibly empty), hµκ iκ∈K is a family of non-zero non-negative countably additive functionals on C, P∞ P ¯. n=0 µn + κ∈K µκ = µ To do this we need the concept of ‘standard extension’ of a countably additive functional on a closed subalgebra of a measure algebra, treated in 327F-327G, together with the following idea.

333M

Closed subalgebras

151

333M Lemma Let (C, µ ¯) be a totally finite measure algebra and hµi ii∈I a family of countably additive functionals on C. For i ∈ I, α ∈ R set eiα = [[µi > α¯ µ]] (326P), and let C0 be the closed subalgebra of C generated by {eiα : i ∈ I, α ∈ R}. Write Σ for the σ-algebra of subsets of RI generated by sets of the form Eiα = {x : x(i) > α} as i runs through I, α runs over R. Then (a) there is a measure µ, with domain Σ, such that there is a measure-preserving isomorphism π : Σ/Nµ → • C0 for which πEiα = eiα for every i ∈ I, α ∈ R, writing Nµ for µ−1 [{0}]; (b) this formula determines both µ and π; (c) for every E ∈ Σ, i ∈ I, we have µi πE • =

R

E

x(i)µ(dx);

(d) for every i ∈ I, µi is the standard extension of µi ¹ C0 to C; (e) for every i ∈ I, µi ≥ 0 iff x(i) ≥ 0 for µ-almost every x; (f) for every i, j ∈ I, µi ≥ µj iff x(i) ≥ x(j) for µ-almost every x; (g) for every i ∈ I, µi = 0 iff x(i) = 0 for µ-almost every x. proof (a) Express (C, µ ¯) as the measure algebra of a measure space (Y, T, ν); write φ : T → C for the corresponding homomorphism. For each i ∈ I let fi : Y → R be a T-measurable, ν-integrable function such R that H fi = µi φH for every H ∈ T. Define ψ : Y → RI by setting ψ(y) = hfi (y)ii∈I ; then ψ −1 [Eiα ] ∈ Σ, and eiα = φ(ψ −1 [Eiα ]) for every i ∈ I, α ∈ R. (See part (a) of the proof of 327F.) So {E : E ⊆ RI , ψ −1 [E] ∈ T}, which is a σ-algebra of subsets of RI , contains every Eiα , and therefore includes Σ; that is, ψ −1 [E] ∈ T for every E ∈ Σ. Accordingly we may define µ by setting µE = νψ −1 [E] for every E ∈ Σ, and µ will be a measure on RI with domain Σ. The Boolean homomorphism E 7→ φψ −1 [E] : Σ → C has kernel Nµ , so descends to a homomorphism π : Σ/Nµ → C, which is measure-preserving. To see that π[Σ/Nµ ] = C0 , observe that because Σ is the σ-algebra generated by {Eiα : i ∈ I, α ∈ R}, π[Σ/Nµ ] must be the closed • subalgebra of C generated by {πEiα : i ∈ I, α ∈ R} = {eiα : i ∈ I, α ∈ R}, which is C0 . (b) Now suppose that µ0 , π 0 have the same properties. Consider A = {E : E ∈ Σ, πE • = π 0 E ◦ }, where I write E • for the equivalence class of E in Σ/Nµ , and E ◦ for the equivalence class of E in Σ/Nµ0 . Then A is a σ-subalgebra of Σ, because E 7→ πE • , E 7→ π 0 E ◦ are both sequentially order-continuous Boolean homomorphisms, and contains every Eiα , so must be the whole of Σ. Consequently µE = µ ¯πE • = µ ¯π 0 E ◦ = µ0 E for every E ∈ Σ, and µ0 = µ; it follows at once that π 0 = π. So µ and π are uniquely determined. (c) If E ∈ Σ and i ∈ I, Z

Z x(i)µ(dx) =

Z x(i)χE(x)µ(dx) =

ψ(y)(i)χE(ψ(y))ν(dy)

E

(applying 235I to the inverse-measure-preserving function ψ : Y → RI ) Z = fi (y)ν(dy) ψ −1 [E]

(by the definition of ψ) = µi φ(ψ −1 [E]) (by the choice of fi ) = µi πE • by the definition of π. (d) Because [[µi > α¯ µ]] ∈ C0 , it is equal to [[µi ¹ C0 > α¯ µ¹ C0 ]], for every α ∈ R. So µi must be the standard extension of µi ¹ C0 (327F). (e)-(g) The point is that, because the standard-extension operator is order-preserving (327F(b-ii)),

152

Maharam’s theorem

333M

µi ≥ 0 ⇐⇒ µi ¹ C0 ≥ 0 Z ⇐⇒ x(i)µ(dx) ≥ 0 for every E ∈ Σ E

⇐⇒ x(i) ≥ 0 µ-a.e., µi ≥ µj ⇐⇒ µi ¹ C0 ≥ µj ¹ C0 Z Z ⇐⇒ x(i)µ(dx) ≥ x(j)µ(dx) for every E ∈ Σ E

E

⇐⇒ x(i) ≥ x(j) µ-a.e., µi = 0 ⇐⇒ µi ¹ C0 = 0 Z ⇐⇒ x(i)µ(dx) = 0 for every E ∈ Σ E

⇐⇒ x(i) = 0 µ-a.e.. 333N A canonical form for closed subalgebras We now have all the elements required to describe a canonical form for structures (A, µ ¯, C), where (A, µ ¯) is a totally finite measure algebra and C is a closed subalgebra of A. The first step is the matching of such structures with structures (C, µ ¯, hµn in∈N , hµκ iκ∈K ), where (C, µ ¯) is a totally finite measure algebra, hµn in∈N is a non-increasing sequence of non-negative countably additive functionals on C, K is a countable set of cardinals, hµκ iκ∈K is a family of non-zero Pinfinite P ∞ ¯; this is the burden of 333K. non-negative countably additive functionals on C, and n=0 µn + κ∈K µκ = µ Next, given any structure of this second kind, we have a corresponding closed subalgebra C0 of C, a measure µ on RI , where I = N ∪ K, and an isomorphism π from the measure algebra C∗0 of µ to C0 , all uniquely defined from the family hµi ii∈I by the process of 333M. For any E belonging to the domain Σ of µ, and i ∈ I, we have µi πE • =

R

E

x(i)µ(dx)

(333Mc), so that µi ¹ C0 is fixed by π and µ. Moreover, the functionals µi can be recovered from their restrictions to C0 by the formulae of 327F (333Md). Thus from (C, µ ¯, hµi ii∈I ) we are led, by a canonical and reversible process, to the structure (C, µ ¯, C0 , I, µ, π). But the extension C of C0 = π[C∗0 ] can be described, up to isomorphism, by the same process as before; that is, it corresponds to a sequence h¯ νn in∈N and a family h¯ νκ iκ∈L of countably additive functionals on C0 satisfying the conditions of 333K. We can transfer these to C∗0 , where they correspond to families hνn in∈N , hνκ iκ∈L of absolutely continuous countably additive functionals defined on Σ, setting νj E = ν¯j πE • for E ∈ Σ, j ∈ N∪L. This process too is reversible; every absolutely continuous countably additive functional ν on Σ corresponds to countably additive functionals on C∗0 and C0 . Let me repeat that the results of 327F mean that the whole structure (C, µ ¯, hµi ii∈I ) can be recovered from (C0 , µ ¯¹ C0 , hµi ¹ C 0 ii∈I ) if we can get the P description of (C, µ ¯) right, and that the requirements µi ≥ 0, µn ≥ µn+1 , µκ 6= 0, i∈I µi = µ ¯ imposed in 333K will survive the process (327F(b-iv)). Putting all this together, a structure (A, µ ¯, C) leads, in a canonical and (up to isomorphism) reversible way, to a structure (K, µ, L, hνκ iκ∈N∪L ) such that

333P

Closed subalgebras

153

K and L are countable sets of infinite cardinals, µ is a totally finite measure on RI , where I = N ∪ K, and its domain Σ is precisely the σ-algebra of subsets of RI defined by the coordinate functionals, forPµ-almost every x ∈ RI we have x(i) ≥ 0 for every i ∈ I, x(n) ≥ x(n + 1) for every n ∈ N and i∈I x(i) = 1, for κ ∈ K, µ{x : x(κ) > 0} > 0, P (these two sections corresponding to the requirements µi ≥ 0, µn ≥ µn+1 , ¯, µκ 6= 0 – see i∈I µi = µ 333M(e)-(g)) for j ∈ J = N ∪ L, νj is a non-negative countably additive functional on Σ, P νn ≥ νn+1 for every n ∈ N, νκ 6= 0 for every κ ∈ L, j∈J νj = µ. 333O Remark I do not envisage quoting the result above very often. Indeed I do not claim that its final form adds anything to the constituent results 333K, 327F and 333M. I have taken the trouble to spell it out, however, because it does not seem to me obvious that the trail is going to end quite as quickly as it does. We need to use 333K twice, but only twice. The most important use of the ideas expressed here, I suppose, is in constructing examples to strengthen our intuition for the structures (A, µ ¯, C) under consideration, and I hope that you will experiment in this direction. 333P At the risk of tresing on the province of Chapter 38, I turn now to a special type of closed subalgebra, in which there is a particularly elegant alternative form for a canonical description. The first step is an important result concerning automorphisms of homogeneous probability algebras. Proposition Let (B, ν¯) be a homogeneous probability algebra. Then there is a measure-preserving automorphism φ : B → B such that limn→∞ ν¯(c ∩ φn (b)) = ν¯c · ν¯b for all b, c ∈ B. proof (a) The case B = {0, 1} is trivial (φ is, and must be, the identity map) so we may take it that B is the measure algebra of {0, 1}κ with its usual measure νκ , where κ is an infinite cardinal. Because #(κ × Z) = max(ω, κ) = κ, there must be a bijection θ : κ → κ such that every orbit of θ in κ is infinite (take θ to correspond to the bijection (ξ, n) 7→ (ξ, n + 1) : κ × Z → κ × Z). This induces a ˆ bijection θˆ : {0, 1}κ → {0, 1}κ through the formula θ(x) = xθ for every x ∈ {0, 1}κ , and of course θˆ is an κ automorphism of the measure space ({0, 1} , νκ ). It therefore induces a corresponding automorphism φ of B, setting φE • = (θˆ−1 [E])• for every E in the domain Σ of νκ . (b) Let Σ0 be the family of subsets E of {0, 1}κ which are determined by coordinates in finite sets, ˜ for some finite set J ⊆ κ and some E ˜ ⊆ {0, 1}J ; that is, are expressible in the form E = {x : x¹J ∈ E} equivalently, expressible as a finite union of basic cylinder sets {x : x¹J = y}. Then Σ0 is a subalgebra of Σ, so C = {E • : E ∈ Σ0 } is a subalgebra of B. (c) Now if b, c ∈ C, there is an n ∈ N such that ν¯(c ∩ φm (b)) = ν¯c · ν¯b for every m ≥ n. P P Express b, c • • ˜ ˜ as E , F where E = {x : x¹J ∈ E}, F = {x : x¹K ∈ F } and J, K are finite subsets of κ. For ξ ∈ K, all the θn (ξ) are distinct, so only finitely many of them can belong to J; as K is also finite, there is an n such that θm [J] ∩ K = ∅ for every m ≥ n. Fix m ≥ n. Then φm (b) = H • where ˜ = {x : x¹L ∈ H}, ˜ H = {x : xθm ∈ E} = {x : xθm ¹J ∈ E} ˜ = {zθ−m : z ∈ E}. ˜ So ν¯(c ∩ φm (b)) = ν(F ∩ H). But L and K are dist, because where L = θm [J] and H m ≥ n, so F and H must be independent (cf. 272K), and ν¯(c ∩ φm (b)) = νF · νH = νF · νE = ν¯c · ν¯b, as claimed. Q Q (d) Now recall that for every E ∈ Σ, ² > 0 there is an E 0 ∈ Σ0 such that ν(E4E 0 ) ≤ ² (254Fe). So, given b, c ∈ B and ² > 0, we can find b0 , c0 ∈ C such that ν¯(b 4 b0 ) ≤ ² and ν¯(c 4 c0 ) ≤ ², and in this case

154

Maharam’s theorem

333P

lim sup|¯ ν (c ∩ φn (b)) − ν¯c · ν¯b| n→∞

≤ lim sup |¯ ν (c ∩ φn (b)) − ν¯(c0 ∩ φn (b0 ))| n→∞

+ |¯ ν (c0 ∩ φn (b0 )) − ν¯c0 · ν¯b0 | + |¯ ν c · ν¯b − ν¯c0 · ν¯b0 | = lim sup |¯ ν (c ∩ φn (b)) − ν¯(c0 ∩ φn (b0 ))| + |¯ ν c · ν¯b − ν¯c0 · ν¯b0 | n→∞

≤ lim sup ν¯(c 4 c0 ) + ν¯(φn (b) 4 φn (b0 )) n→∞

+ ν¯c|¯ ν b − ν¯b0 | + |¯ ν c − ν¯c0 |¯ ν b0 ≤ ν¯(c 4 c0 ) + ν¯(b 4 b0 ) + ν¯c · ν¯(b 4 b0 ) + ν¯(c 4 c0 )¯ ν b0 ≤ 4². As ² is arbitrary, limn→∞ ν¯(c ∩ φn (b)) = ν¯c · ν¯b, as required. Remark Automorphisms of this type are called mixing (see 372P below). 333Q Corollary Let (C, µ ¯0 ) be a totally finite measure algebra and (B, ν¯) a probability algebra which is either homogeneous or purely atomic with finitely many atoms all of the same measure. Let (A, µ ¯) be the localizable measure algebra free product of (C, µ ¯0 ) and (B, ν¯). Then there is a measure-preserving automorphism π : A → A such that {a : a ∈ A, πa = a} = {c ⊗ 1 : c ∈ C}. Remark I am following 315M in using the notation c ⊗ b for the intersection in A of the canonical images of c ∈ C and b ∈ B. By 325Dc I need not distinguish between the free product C ⊗ B and its image in A. proof (a) Let me deal with the case of atomic B first. In this case, if B has n atoms b0 , . . . , bn−1 , let φ : B → B be the measure-preserving homomorphism cyclically permuting these atoms, so that φb0 = b1 , . . . , φbn−1 = b0 . Because φ is an automorphism of (B, ν¯), it induces an automorphism π of (A, µ ¯); any member of A is uniquely expressible as a = supi 0. We know that C ⊗ B is topologically dense in A (325Dc), so there is an a0 ∈ C ⊗ B such that µ ¯(a 4 a0 ) ≤ ²2 . 0 Express a as supi∈I ci ⊗ bi , where hci ii∈I is a finite partition of unity in C (315Na). Then πa0 = supi∈I ci ⊗ φ(bi ),

π n (a0 ) = supi∈I ci ⊗ φn (bi ) for every n ∈ N.

So we can get a formula for ¯(a0 ∩ π n (a0 )) = lim µ ¯(sup ci ⊗ (bi ∩ φn (bi ))) lim µ n→∞ i∈I X X = lim µ ¯0 ci · ν¯(bi ∩ φn (bi )) = µ ¯0 ci (¯ ν bi )2 .

n→∞

n→∞

i∈I

i∈I

It follows that X i∈I

µ ¯0 ci (¯ ν bi )2 = lim µ ¯(a0 ∩ π n (a0 )) n→∞

≥ lim sup µ ¯(a ∩ π n (a)) − µ ¯(a 4 a0 ) − µ ¯(π n (a) 4 π n (a0 )) n→∞ X =µ ¯a − 2¯ µ(a 4 a0 ) ≥ µ ¯a0 − 3¯ µ(a 4 a0 ) ≥ µ ¯0 ci · ν¯bi − 3²2 , i∈I

333R

Closed subalgebras

that is,

P i∈I

155

µ ¯0 ci · ν¯bi · (1 − ν¯bi ) ≤ 3²2 .

But this means that, setting J = {i : i ∈ I, ν¯bi (1 − ν¯bi ) ≥ ²}, we must have K = {i : i ∈ I, ν¯bi ≥ 1 − 2²},

L = {i : i ∈ I \ K, ν¯bi ≤ 2²},

P i∈J

µ ¯0 ci ≤ 3². Set

c = supi∈K ci .

Then I \ (K ∪ L) ⊆ J, so µ ¯(a0 4 (c ⊗ 1)) =

X

µ ¯0 ci · ν¯bi +

X

µ ¯0 ci · ν¯bi +

i∈J

≤

X i∈J

µ ¯0 ci · (1 − ν¯bi )

i∈K

i∈I\K

≤

X

X

µ ¯0 ci · ν¯bi +

i∈L

µ ¯0 ci + 2²

X i∈L

µ ¯0 ci + 2²

X

X

µ ¯0 ci · (1 − ν¯bi )

i∈K

µ ¯0 ci ≤ 3² + 2² = 5²,

i∈K

and µ ¯(a 4 (c ⊗ 1)) ≤ ²2 + 5². As ² is arbitrary, a belongs to the topological closure of C1 . But of course C1 is a closed subalgebra of A (325Dd), so must actually contain a. As a is arbitrary, π has the required property. 333R Now for the promised special type of closed subalgebra. It will be convenient to have the following temporary notation. Write Card∗ for the (proper) class of all non-zero cardinals. For infinite κ ∈ Card∗ , let (Bκ , ν¯κ ) be the measure algebra of the usual measure on {0, 1}κ . For finite n ∈ Card∗ , let Bn be the power set of {0, . . . , n − 1} and set ν¯n b = n1 #(b) for b ∈ Bn . Theorem Let (A, µ ¯) be a totally finite measure algebra and C a subset of A. Then the following are equiveridical: (i) there is some set G of measure-preserving automorphisms of A such that C = {c : c ∈ A, πc = c for every π ∈ G}; (ii) C is a closed subalgebra of A and there Q is a partition of unity hei ii∈I of C, where I is a countable b i , writing Cei for the principal ideal of C generated subset of Card∗ , such that A is isomorphic to i∈I Cei ⊗B b i for the localizable measure algebra free product of Cei and Bi by ei and endowed with µ ¯¹ Cei , and Cei ⊗B – the isomorphism being one which takes any c ∈ C to h(c ∩ ei ) ⊗ 1ii∈I , as in 333H and 333K; (iii) there is a single measure-preserving automorphism π of A such that C = {c : c ∈ A, πc = c}. α) C is a subalgebra because every π ∈ G is a Boolean homomorphism, and it is orderproof (a)(i)⇒(ii)(α closed because every π is order-continuous (324Kb). (Or, if you prefer, C is topologically closed because every π is continuous.) β ) Because C is a closed subalgebra of A, its embedding can be described in of families hµn in∈N , (β hµκ iκ∈K as in Theorem 333K. Set I 0 = K ∪N. Recall that each µi is defined by setting µi c = µ ¯(c ∩ ai ), where hai ii∈I 0 is a partition of unity in A (see the proofs of 333H and 333K). Now for κ ∈ K, aκ is the maximal element of A which is relatively Maharam-type-homogeneous over C with relative Maharam type κ (part (b) of the proof of 333K). Consequently we must have πaκ = aκ for any measure algebra automorphism of (A, µ ¯) which leaves C invariant; in particular, for every π ∈ G. Thus aκ ∈ C for every κ ∈ K. (γγ ) Now consider the relatively atomic part of A. The elements an , for n ∈ N, are not uniquely defined. However, the functionals µn and their s e0n = [[µn > 0]] are uniquely defined from the structure (A, µ ¯, C) and therefore invariant under G. Observe also that because supn∈N an = 1 \ supκ∈K aκ belongs to C, and e0n = inf{c : c ∈ C, c ⊇ an }, while e0n ⊇ e0n+1 for every n, we must have e00 = supn∈N an .

156

Maharam’s theorem

333R

Let G∗ be the set of all those automorphisms π of the measure algebra (A, µ ¯) such that πc = c for every c ∈ C. Then of course G∗ is a group including G. Now supπ∈G∗ πan must be invariant under every member of G∗ , so belongs to C; it includes an and is included in any member of C including an , so must be e0n . (δδ ) I claim now that if n ∈ N then e0n ∩ [[µ0 > µn ]] = 0. P P?? Otherwise, set c = [[µ0 > µn ]] ∩ e0n . Then µ0 c > 0 so c ∩ a0 6= 0. By the last remark in (γ), there is a π ∈ G∗ such that c ∩ a0 ∩ πan 6= 0. Now there is a c0 ∈ C such that c ∩ a0 ∩ πan = c0 ∩ a0 , and of course we may suppose that c0 ⊆ c. But this means that π(c0 ∩ an ) = c0 ∩ πan ⊇ c0 ∩ a0 ∩ πan = c0 ∩ a0 , so that µn c0 = µ ¯(c0 ∩ an ) = µ ¯π(c0 ∩ an ) ≥ µ ¯(c0 ∩ a0 ) = µ0 c0 , which is impossible, because 0 6= c0 ⊆ [[µ0 > µn ]]. X XQ Q So µ0 c ≤ µn c whenever c ∈ C and c ⊆ e0n . Because the µk have been chosen to be a non-increasing sequence, we must have µ0 c = µ1 c = . . . = µn c for every c ⊆ e0n . P 1 (²²) Recalling now that i∈I 0 µi = µ ¯¹ C, we see that µ0 c ≤ n+1 µ ¯c for every c ⊆ e0n . It follows that if ∗ 0 ∗ ∗ e = inf n∈N en , µ0 e = 0; but this must mean that e = 0. Consequently, setting I = I 0 \ {0}, en = e0n−1 \ e0n for n ≥ 1, eκ = aκ for κ ∈ K, we find that hei ii∈I is a partition of unity in C. Moreover, for n ≥ 1 and c ⊆ en , we must have P P µ ¯c = i∈I 0 µi c = k
µ ¯(ak ∩ c) = µk c = µ ¯c = µ ¯c · ν¯n {k} for all relevant k and c. Because Bn is finite, ψn is surjective. (ζζ ) Just asQ in 333H, we now see that because hei ii∈I is a partition of unity in A as well as in C, we can Q b i. identify A with i∈I Aei and therefore with i∈I Cei ⊗B Q b i , writing ψ : A → D for the canonical map. For each i ∈ I, (b)(ii)⇒(iii) Let us work in D = i∈I Cei ⊗B b i with fixed-point subalgebra {c ⊗ 1 : c ∈ Cei } we have a measure-preserving automorphism πi of Cei ⊗B (333Q). For d = hdi ii∈I ∈ D, set πd = hπi di ii∈I . Then π is a measure-preserving automorphism because every πi is. If πd = d, then for every i ∈ I there must be a ci ⊆ ei such that di = ci ⊗ 1. But this means that d = ψc, where c = supi∈I ci ∈ C. Thus the fixed-point subalgebra of π is just ψ[C]. Transferring the structure (D, ψ[C], π) back to A, we obtain a measure-preserving automorphism ψ −1 πψ of A with fixed-point subalgebra C, as required. (c)(iii)⇒(i) is trivial. 333X Basic exercises (a) Show that, in the proof of 333H, ci = upr(ai , C) (definition: 314V) for every i ∈ I. (b) In the context of Lemma 333M, show that we have a one-to-one correspondence between atoms c of C0 and points x of non-zero mass in RI , given by the formula π{x}• = c. (c) Let (A, µ ¯) be totally finite measure algebra and G a set of measure-preserving Boolean homomorphisms from A to itself such that πφ ∈ G for all π, φ ∈ G. (i) Show that a ⊆ supπ∈G πa for every a ∈ A. (Hint: if πc ⊆ c, where π ∈ G and c ∈ A, then πc = c; apply this to c = supπ∈G πa.) (ii) Set C = {c : c ∈ A, πc = c for every π ∈ G}. Show that supπ∈G πa = upr(a, C) for every a ∈ A.

333 Notes

Closed subalgebras

157

333Y Further exercises (a) Show that when I = N the algebra Σ of subsets of RI , used in 333M, is precisely the Borel σ-algebra as described in 271Ya. (b) Let (A, µ ¯) be a totally finite measure algebra, and B, C two closed subalgebras of A with C ⊆ B. Show that τC (B) ≤ τC (A). (Hint: use 333K and the ideas of 332T.) (c) Let (A, µ ¯) be a probability algebra. Show that A is homogeneous iff there is a measure-preserving automorphism of A which is mixing in the sense of 333P. (d) Let (A, µ ¯) be a totally finite measure algebra, and G a set of measure-preserving Boolean homomorphisms from A to itself. Set C = {c : c ∈ A, πc = c for every π ∈ G}. Show that C is a closed subalgebra of A of the type described in 333R. (Hint: in the language of part (a) of the proof of 333R, show that supn∈N an still belongs to C.) 333 Notes and comments I have done my best, in the first part of this section, to follow the lines already laid out in §§331-332, using what should (once you have seen them) be natural generalizations of the former definitions and arguments. Thus the Maharam type τ (A) of an algebra is just the relative Maharam type τ{0,1} (A), and A is Maharam-type-homogeneous iff it is relatively Maharam-type-homogeneous over {0, 1}. To help you trace through the correspondence, I list the code numbers: 331Fa→333Aa, 331Fb→333Ac, 331Hc→333Ba, 331Hd→333Bd, 332A→333Bb, 331I→333Cb, 331K→333E, 331L→333F(ii), 332B→333H, 332J→333K. 333D overlaps with 332P. Throughout, the principle is the same: everything can be built up from products and free products. Theorem 333Ca does not generalize any explicitly stated result, but overlaps with Proposition 332P. In the proof of 333E I have used a new idea; the same method would of course have worked just as well for 331K, but I thought it worth while to give an example of an alternative technique, which displays a different facet of homogeneous algebras, and a different way in which the algebraic, topological and metric properties of homogeneous algebras interact. The argument of 331K-331L relies (without using the term) on the fact that measure algebras of Maharam type κ have topological density at most max(κ, ω) (see 331Ye), while the the argument of 333E uses the rather more sophisticated concept of stochastic independence. Corollary 333F(i) is cruder than the more complicated results which follow, but I think that it is invaluable as a first step in forming a picture of the possible embeddings of a given (totally finite) measure algebra C in a larger algebra A. If we think of C as the measure algebra of a measure space (X, Σ, µ), then we can be sure that A is representable as a closed subalgebra of the measure algebra of X × {0, 1}κ for some κ, that is, the measure algebra of λ¹ T where λ is the product measure on X × {0, 1}κ and T is some σ-subalgebra of the domain of λ; the embedding of C in A being defined by the formula E • → (E × {0, 1}κ )• for E ∈ Σ (325A, 325D). Identifying, in our imaginations, both X and {0, 1}κ with the unit interval, we can try to picture everything in the unit square – and these pictures, although necessarily inadequate for algebras of uncountable Maharam type, already give a great deal of scope for invention. I said above that everything can be constructed from simple products and free products, judiciously combined; of course some further ideas must be mixed with these. The difference between 332B and 333H, for instance, is partly in the need for the functionals µi in the latter, whereas in the former the decomposition involves only principal ideals with the induced measures. Because the µi are completely additive, they all have s ci (326Xi) and we get measure algebras (Cci , µi ¹ Cci ) to use in the products. (I note that the ci can be obtained directly from the ai , without mentioning the functionals µi , by the process of 333Xa.) The fact that the ci can overlap means that the ‘relatively atomic’ part of the larger algebra A needs a much more careful description than before; this is the burden of 333J, and also the principal complication in the proof of 333R. The ‘relatively atomless’ part is (comparatively) straightforward, since we can use the same kind of amalgamation as before (part (c-i) of the proof of 332J, part (b) of the proof of 333K), simplified because I am no longer seeking to deal with algebras of infinite magnitude. Theorem 333K gives a canonical form for superalgebras of a given totally finite measure algebra (C, µ ¯), b taking the structure (C, µ ¯) itself for granted. I hope it is clear that while the µi amd ei and the algebra A = Q Q b b C × C ⊗B e e κ and the embedding of C in A are uniquely defined, the rest of the isomorphism n κ n∈N κ∈K b generally is not. Even when the aκ are uniquely defined the isomorphisms between Aa and π : A → A κ b Ceκ ⊗Bκ depend on choosing generating families in the Aaκ ; see the proof of 333Cb.

158

Maharam’s theorem

333 Notes

To understand the possible structures (C, hµi ii∈I ) of that theorem, we have to go rather deeper. The route I have chosen is to pick out the subalgebra C0 of C determined by hµi ii∈I and identify it with the measure algebra of a particular measure on RI . Perhaps I should apologise for not stating explicitly in the course of Lemma 333M that the measures µ here are ‘Borel measures’ (see 333Ya); but I am afraid of opening a door to an invasion of ideas which belong in Volume 4. Besides, if I were going to do anything more with these measures than observe that they are uniquely defined by the construction proposed, I would complete them and call them Radon measures. In order to validate this approach, I must show that the µi can be recovered from their restrictions to C0 ; this is 333Md, and is the motive for the discussion of ‘standard extensions’ in §327. No doubt there are other ways of doing it. One temptation which I felt it right to resist was the idea of decomposing C into its homogeneous principal ideals; this seemed merely an additional complication. Of course the subalgebra C0 has countable Maharam type (being τ -generated by the elements eiq , for i ∈ I and q ∈ Q, of 333M), so that its decomposition is relatively simple, being just a matter of picking out the atoms (333Xb). In 333P I find myself presenting an important fact about homogeneous measure algebras, rather out of context; but I hope that it will help you to believe that I have by no means finished with the insights which Maharam’s theorem provides. I give it here for the sake of 333R. For the moment, I invite you to think of 333R as just a demonstration of the power of the techniques I have developed in this chapter, and of the kind of simplification (in the equivalence of conditions (i) and (iii)) which seems to arise repeatedly in the theory of measure algebras. But you will see that the first step to understanding any automorphism will be a description of its fixed-point subalgebra, so 333R will also be basic to the theory of automorphisms of measure algebras. I note that the hypothesis (i) of 333R can in fact be relaxed (333Yd), but this seems to need an extra idea.

334 Products I devote a short section to results on the Maharam classification of the measure algebras of product measures, or, if you prefer, of the free products of measure algebras. The complete classification, even for probability algebras, is complex (334Xe, 334Ya), so I content myself with a handful of the most useful results. I start with upper bounds for the Maharam type of the c.l.d. product of two measure spaces (334A) and the localizable measure algebra free product of two semi-finite measure algebras (334B), and go on to the corresponding results for products of probability spaces and algebras (334C-334D). Finally, I show that any infinite power of a probability space is Maharam-type-homogeneous (334E). 334A Theorem Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras A, B. Let λ be ¯ its measure algebra. Then the Maharam type τ (C) is at the c.l.d. product measure on X × Y , and (C, λ) most max(ω, τ (A), τ (B)). proof Recall from 325A that we have order-continuous Boolean homomorphisms ε1 : A → C and ε2 : B → C defined by setting ε1 E • = (E × Y )• , ε2 F • = (X × F )• for E ∈ Σ, F ∈ T. Let A ⊆ A, B ⊆ B be τ -generating sets with #(A) = τ (A), #(B) = τ (B); set C = ε1 [A] ∪ ε2 [B]. Then C τ -generates C. P P Let C1 be the order-closed subalgebra of C generated by C. Because ε1 is order-continuous, ε−1 [C ] is an order-closed 1 1 subalgebra of A, and it includes A, so must be the whole of A; thus ε1 a ∈ C1 for every a ∈ A. Similarly, ε2 b ∈ C1 for every b ∈ B. This means that Λ1 = {W : W ∈ Λ, W • ∈ C1 } contains E × F for every E ∈ Σ, F ∈ T. Also Λ1 is a σ-algebra of subsets of X × Y , because C1 is N (sequentially) order-closed in C. So Λ1 ⊇ Σ c T (definition: 251D). But this means that if W ∈ Λ there is a V ∈ Λ1 such that V ⊆ W and λV = λW (251Ib); that is, if c ∈ C there is a d ∈ C1 such that d ⊆ c and ¯ = λc. ¯ Thus C1 is order-dense in C, and λd c = sup{d : d ∈ C1 , d ⊆ c} ∈ C1 for every c ∈ C. So C1 = C and C τ -generates C, as claimed. Q Q

334E

Products

159

Consequently τ (C) ≤ #(C) ≤ max(ω, τ (A), τ (B)). 334B Corollary Let (A1 , µ ¯1 ), (A2 , µ ¯2 ) be semi-finite measure algebras, with localizable measure algebra ¯ (325E). Then τ (C) ≤ max(ω, τ (A), τ (B)). free product (C, λ) proof By the construction of part (a) of the proof of 325D, C can be regarded as the measure algebra of the c.l.d. product of the Stone representations of (A1 , µ ¯1 ) and (A2 , µ ¯2 ); so the result follows at once from 334A. 334C Theorem Let h(Xi , Σi , µi )ii∈I be a family of probability spaces, with product (X, Λ, λ). Let Ai , C be the corresponding measure algebras. Then τ (C) ≤ max(ω, #(I), supi∈I τ (Ai )). proof Recall from 325I that we have order-continuous Boolean homomorphisms ψi : Ai → C corresponding to the inverse-measure-preserving maps x 7→ πi (x) =S x(i) : X → Xi . For each i ∈ I, let Ai ⊆ Ai be a set of cardinal τ (Ai ) which τ -generates Ai . Set C = i∈I ψi [Ai ]. Then C τ -generates C. P P Let C1 be the −1 order-closed subalgebra of C generated by C. Because ψi is order-continuous, ψi [C1 ] is an order-closed subalgebra of Ai , and it includes Ai , so must be the whole of Ai ; thus ψi a ∈ C1 for every a ∈ Ai , i ∈ I. This means that Λ1 = {W : W ∈ Λ, W • ∈ C1 } contains πi−1 [E] for every E ∈ Σi , i ∈ I. Also Λ1 is a σ-algebra of subsets of X, because C1 is (sequentially) N order-closed in C. So Λ1 ⊇ c i∈I Σi . But this means that if W ∈ Λ there is a V ∈ Λ1 such that V • = W • (254Ff), that is, that C1 = C, and C τ -generates C, as claimed. Q Q Consequently τ (C) ≤ #(C) ≤ max(ω, #(I), supi∈I τ (Ai )). 334D Corollary Let h(Ai , µ ¯i )ii∈I be a family of probability algebras, with probability algebra free ¯ Then product (C, λ). τ (C) ≤ max(ω, #(I), supi∈I τ (Ai )). proof See 325J-325K. 334E I come now to the question of when a probability algebra free product is homogeneous. I give just one result in detail, leaving others to the exercises. Theorem Let (X, Σ, µ) be a probability space, with measure algebra A, and I an infinite set; let C be the measure algebra of the product measure on X I . Then C is homogeneous. If τ (A) = 0 then τ (C) = 0; otherwise τ (C) = max(τ (A), #(I)). ¯ the measure on C. If proof (a) As usual, write µ ¯ for the measure of A; let λ be the measure on X I , λ τ (A) = 0, that is, A = {0, 1}, then C = {0, 1} (by 254Fe, or 325Jc, or otherwise), and in this case is surely homogeneous, with τ (C) = 0. So let us suppose hencforth that τ (A) > 0. We have τ (C) ≤ max(ω, #(I), τ (A)) = max(#(I), τ (A)), by 334C. (b) Fix on b ∈ A \ {0, 1}. For each i ∈ I, let ψi : A → C be the canonical measure-preserving homomorphism corresponding to the inverse-measure-preserving function x 7→ x(i) : X I → X. For each n ∈ N, there is a set J ⊆ I of cardinal n, and now the finite subalgebra of C generated by {ψi b : i ∈ J} has atoms of measure at most δ n , where δ = max(¯ µb, 1 − µ ¯b) < 1. Consequently C can have no atom of measure greater than δ n , for any n, and is therefore atomless.

160

Maharam’s theorem

334E

(c) Because I is infinite, there is a bijection between I and I × N; that is, there is a partition hJi ii∈I of I into countably infinite sets. Now (X I , λ) can be identified with the product of the family h(X Ji , λi )ii∈I , where λi is the product measure on X Ji (254N). By (b), every λi is atomless, so there are sets Ei ⊆ X Ji of measure 21 . The sets Ei0 = {x : x¹Ji ∈ Ei } are now stochastically independent in X. Accordingly we have an inverse-measure-preserving function f : X → {0, 1}I , endowed with its usual measure νI , defined by setting f (x)(i) = 1 if x ∈ Ei0 , 0 otherwise, and therefore a measure-preserving Boolean homomorphism π : BI → C, writing BI for the measure algebra of νI . Now if c ∈ C \ {0} and Cc is the corresponding ideal, b 7→ c ∩ πb : BI → Cc is an order-continuous Boolean homomorphism. It follows that τ (Cc ) ≥ #(I) (331J). (d) Again take any non-zero c ∈ C. For each i ∈ I, set ai = inf{a : ψi a ⊇ c}. Writing Aai for the corresponding principal ideal of A, we have an order-continuous Boolean homomorphism ψi0 : Aai → Cc , given by the formula ψi0 a = ψi a ∩ c for every a ∈ Aai . Now ψi0 is injective, so is a Boolean isomorphism between Aai and its image ψi0 [Aai ], which by 314F(a-i) is a closed subalgebra of Cc . So τ (Aai ) = τ (ψi0 [Aai ]) ≤ τ (Cc ) by 332Tb. For any finite J ⊆ I, ¯ ≤ λ(inf ¯ 0 < λc i∈J ψi ai ) =

Q i∈J

¯ i ai ) = Q ¯ai . λ(ψ i∈J µ

So for any δ < 1, {i : µ ¯ai ≤ δ} must be finite, and supi∈I µ ¯ai = 1. In particular, supi∈I ai = 1 in A. But this means that if ζ is any cardinal such that the Maharam-type-ζ component eζ of A is non-zero, then eζ ∩ ai 6= 0 for some i ∈ I, so that ζ ≤ τ (Aeζ ∩ai ) ≤ τ (Aai ) ≤ τ (Cc ). As ζ is arbitrary, τ (A) ≤ max(ω, τ (Cc )) (332S). (e) Putting (a)-(d) together, we have max(τ (A), #(I)) ≤ max(ω, τ (Cc )) = τ (Cc ) ≤ τ (C) ≤ max(τ (A), #(I)) for every non-zero c ∈ C; so C is homogeneous. 334X Basic exercises (a) Let (X, Σ, µ) and (Y, T, ν) be complete locally determined measure spaces with c.l.d. product (X × Y, Λ, λ). Let A, C be the measure algebras of µ, λ respectively. Show that if νY > 0 then τ (A) ≤ τ (C). (b) Let X be a set and µ and ν two totally finite measures on X with the same domain Σ; then λ = µ + ν ¯ for the three measure algebras. Show is also a totally finite measure. Write (A, µ ¯), (B, ν¯) and (C, λ) that (i) there is a surjective order-continuous Boolean homomorphism from C onto A; (ii) C is isomorphic to a closed subalgebra of the localizable measure algebra free product of A and B; (iii) τ (A) ≤ τ (C) ≤ max(ω, τ (B), τ (C)). ¯ > (c) Let h(Ai , µ ¯i )ii∈I be a family of probability algebras, with probability algebra free product (C, λ). Show that τ (Ai ) ≤ τ (C) for every i, and that #({i : i ∈ I, τ (Ai ) > 0}) ≤ τ (C). (d) Let X be a set and hµn in∈N a sequence of totally finite measures on X all with the same domain P∞ Σ. Let hα i be a sequence of non-negative real numbers such that αn µn X < ∞, and set λE = n n∈N n=0 P∞ ¯n ) for the measure algebra of µn and n=0 αn µn E for E ∈ Σ. Check that λ is a measure. Write (An , µ ¯ for the measure algebra of λ. Show that τ (C) ≤ max(ω, sup (C, λ) n∈N τ (An )). (e) Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces, and λ the product measure on X × Y . Show that λ is Maharam-type-homogeneous iff one of µ, ν is Maharam-type-homogeneous with Maharam type at least as great as the Maharam type of the other.

334 Notes

Products

161

(f ) Show that the product of any family of Maharam-type-homogeneous probability spaces is again Maharam-type-homogeneous. > (g) Let (X, Σ, µ) be a probability space of Maharam type κ, and I any set of cardinal at least max(ω, κ). Show that the product measure on X × {0, 1}I is Maharam-type-homogeneous, with Maharam type #(I). 334Y Further exercises (a) Let h(Xi , Σi , µi )ii∈I be an infinite family of probability spaces, with product (X, Λ, λ). Let κi be the Maharam type of µi for each i; set κ = max(#(I), supi∈I κi ). Show that either λ is Maharam-type-homogeneous, with Maharam type κ, or there are κ0 < κ, Xi0 ∈ Σi such that P 0 0 0 i∈I µi (Xi \ Xi ) < ∞ and the Maharam type of the subspace measure on Xi is at most κ for every i ∈ I 0 and either κ = 0 or #(I) < κ. 334 Notes and comments The results above are all very natural ones; I have spelt them out partly for completeness and partly for the sake of an application in §346 below. But note the second alternative in 334Ya; it is possible, even in an infinite product, for a kernel of relatively small Maharam type to be preserved.

162

Liftings

Chapter 34 The lifting theorem Whenever we have a surjective homomorphism φ : P → Q, where P and Q are mathematical structures, we can ask whether there is a right inverse of φ, a homomorphism ψ : Q → P such that φψ is the identity on Q. As a general rule, we expect a negative answer; those categories in which epimorphisms always have right inverses (e.g., the category of linear spaces) are rather special, and elsewhere the phenomenon is relatively rare and almost always important. So it is notable that we have a case of this at the very heart of the theory of measure algebras: for any complete probability space (X, Σ, µ) (in fact, for any complete strictly localizable space of non-zero measure) the canonical homomorphism from Σ to the measure algebra of µ has a right inverse (341K). This is the von Neumann-Maharam lifting theorem. Its proof, together with some essentially elementary remarks, takes up the whole of of §341. As a first application of the theorem (there will be others in Volume 4) I apply it to one of the central problems of measure theory: under what circumstances will a homomorphism between measure algebras be representable by a function between measure spaces? Variations on this question are addressed in §343. For a reasonably large proportion of the measure spaces arising naturally in analysis, homomorphisms are representable (343B). New difficulties arise if we ask for isomorphisms of measure algebras to be representable by isomorphisms of measure spaces, and here we have to work rather hard for rather narrowly applicable results; but in the case of Lebesgue measure and its closest relatives, a good deal can be done, as in 344H and 344I. Returning to liftings, there are many difficult questions concerning the extent to which liftings can be required to have special properties, reflecting the natural symmetries of the standard measure spaces. For instance, Lebesgue measure is translation-invariant; if liftings were in any sense canonical, they could be expected to be automatically translation-invariant in some sense. It seems sure that there is no canonical lifting for Lebesgue measure – all constructions of liftings involve radical use of the axiom of choice – but even so we do have many translation-invariant liftings (§345). We have less luck with product spaces; here the construction of liftings which respect the product structure is fraught with difficulties. I give the currently known results in §346.

341 The lifting theorem I embark directly on the principal theorem of this chapter (341K, ‘every non-trivial complete strictly localizable measure space has a lifting’), using the minimum of advance preparation. 341A-341B give the definition of ‘lifting’; the main argument is in 341F-341K, using the concept of ‘lower density’ (341C-341E) and a theorem on martingales from §275. In 341P I describe an alternative way of thinking about liftings in of the Stone space of the measure algebra. 341A Definition Let (X, Σ, µ) be a measure space, and A its measure algebra. By a lifting of A (or of (X, Σ, µ), or of µ) I shall mean either a Boolean homomorphism θ : A → Σ such that (θa)• = a for every a ∈ A or a Boolean homomorphism φ : Σ → Σ such that (i) φE = ∅ whenever µE = 0 (ii) µ(E4φE) = 0 for every E ∈ Σ. 341B Remarks (a) I trust that the ambiguities permitted by this terminology will not cause any confusion. The point is that there is a natural one-to-one correspondence between liftings θ : A → Σ and liftings φ : Σ → Σ given by the formula θE • = φE for every E ∈ Σ. P P (i) Given a lifting θ : A → Σ, the formula defines a Boolean homomorphism φ : Σ → Σ such that φ∅ = θ0 = ∅,

(E 4 φE)• = E • 4 (θE • )• = 0 ∀ E ∈ Σ,

341E

The lifting theorem

163

so that φ is a lifting. (ii) Given a lifting φ : Σ → Σ, the kernel of φ includes {E : µE = 0}, so there is a Boolean homomorphism θ : A → Σ such that θE • = φE for every E (3A2G), and now (θE • )• = (φE)• = E • for every E ∈ Σ, so θ is a lifting. Q Q I suppose that the word ‘lifting’ applies most naturally to functions from A to Σ; but for applications in measure theory the other type of lifting is used at least equally often. (b) Note that if φ : Σ → Σ is a lifting then φ2 = φ. P P For any E ∈ Σ, φ2 E 4 φE = φ(E 4 φE) = ∅. Q Q If φ is associated with θ : A → Σ, then φθa = θa for every a ∈ A. P P φθa = θ((θa)• ) = θa. Q Q 341C Definition Let (X, Σ, µ) be a measure space, and A its measure algebra. By a lower density of A (or of (X, Σ, µ), or of µ) I shall mean either a function θ : A → Σ such that (i) (θa)• = a for every a ∈ A (ii) θ0 = ∅ (iii) θ(a ∩ b) = θa ∩ θb for all a, b ∈ A or a function φ : Σ → Σ such that (i) φE = φF whenever E, F ∈ Σ and µ(E4F ) = 0 (ii) µ(E4φE) = 0 for every E ∈ Σ (iii) φ∅ = ∅ (iv) φ(E ∩ F ) = φE ∩ φF for all E, F ∈ Σ. 341D Remarks (a) As in 341B, there is a natural one-to-one correspondence between lower densities θ : A → Σ and lower densities φ : Σ → Σ given by the formula θE • = φE for every E ∈ Σ. (For the requirement φE = φF whenever E • = F • in A means that every φ corresponds to a function θ, and the other clauses match each other directly.) (b) As before, if φ : Σ → Σ is a lower density then φ2 = φ. If φ is associated with θ : A → Σ, then φθ = θ. (c) It will be convenient, in the course of the proofs of 341F-341H below, to have the following concept available. If (X, Σ, µ) is a measure space with measure algebra A, a partial lower density of A is a function θ : B → Σ such that (i) the domain B of θ is a subalgebra of A (ii) (θb)• = b for every b ∈ B (iii) θ0 = ∅ (iv) θ(a ∩ b) = θa ∩ θb for all a, b ∈ B. Similarly, if T is a subalgebra of Σ, a function φ : T → Σ is a partial lower density if (i) φE = φF whenever E, F ∈ T and µ(E4F ) = 0 (ii) µ(E4φE) = 0 for every E ∈ T (iii) φ∅ = ∅ (iv) φ(E∩F ) = φE∩φF for all E, F ∈ T. (d) Note that lower densities and partial lower densities are order-preserving; if a ⊆ b in A, and θ is a lower density of A, then θa = θ(a ∩ b) = θa ∩ θb ⊆ θb. (e) Of course a Boolean homomorphism from A to Σ, or from Σ to itself, is a lifting iff it is a lower density. 341E Example Let µ be Lebesgue measure on Rr , where r ≥ 1, and Σ its domain. For E ∈ Σ set φE = {x : x ∈ Rr , limδ↓0

µ(E∩B(x,δ)) µB(x,δ)

= 1}.

(Here B(x, δ) is the closed ball with centre x and radius δ.) Then φ is a lower density for µ; we may call it lower Lebesgue density. P P (You may prefer at first to suppose that r = 1, so that B(x, δ) = [x − δ, x + δ] and µB(x, δ) = 2δ.) By 261Db (or 223B, for the one-dimensional case) φE4E is negligible for every E; in particular, φE ∈ Σ for every E ∈ Σ. If E4F is negligible, then µ(E ∩ B(x, δ)) = µ(F ∩ B(x, δ)) for every x and δ, so φE = φF . If E ⊆ F , then µ(E ∩ B(x, δ)) ≤ µ(F ∩ B(x, δ)) for every x, δ, so φE ⊆ φF ; consequently φ(E ∩ F ) ⊆ φE ∩ φF for all E, F ∈ Σ. If E, F ∈ Σ and x ∈ φE ∩ φF , then

164

Liftings

341E

µ(E ∩ F ∩ B(x, δ)) = µ(E ∩ B(x, δ)) + µ(F ∩ B(x, δ)) − µ((E ∪ F ) ∩ B(x, δ)) ≥ µ(E ∩ B(x, δ)) + µ(F ∩ B(x, δ)) − µ(B(x, δ)) for every δ, so µ(E∩F ∩B(x,δ)) µB(x,δ)

≥

µ(E∩B(x,δ)) µB(x,δ)

+

µ(F ∩B(x,δ)) µB(x,δ)

−1→1

as δ ↓ 0, and x ∈ φ(E ∩ F ). Thus φ(E ∩ F ) = φE ∩ φF for all E, F ∈ Σ, and φ is a lower density. Q Q 341F The hard work of this section is in the proof of 341H below. To make it a little more digestible, I extract two parts of the proof as separate lemmas. Lemma Let (X, Σ, µ) be a probability space and A its measure algebra. Let B be a closed subalgebra of A and θ : B → Σ a partial lower density. Then for any e ∈ A there is a partial lower density θ1 , extending θ, defined on the subalgebra B1 of A generated by B ∪ {e}. proof (a) Because B is order-closed, v = upr(e, B) = inf{a : a ∈ B, a ⊇ e},

w = upr(1 \ e, B)

are defined in B (314V). Let E ∈ Σ be such that E = e. •

(b) We have a function θ1 : B1 → Σ defined by writing ¡ ¢ ¡ ¢ θ1 ((a ∩ e) ∪ (b \ e)) = θ((a ∩ v) ∪ (b \ v)) ∩ E ∪ θ((a \ w) ∪ (b ∩ w)) \ E for a, b ∈ B. P P By 312M, every element of B1 is expressible as (a ∩ e) ∪ (b \ e) for some a, b ∈ B. If a, a0 , 0 b, b ∈ B are such that (a ∩ e) ∪ (b \ e) = (a0 ∩ e) ∪ (b0 \ e), then a ∩ e = a0 ∩ e and b \ e = b0 \ e, that is, a 4 a0 ⊆ 1 \ e ⊆ w,

b 4 b0 ⊆ e ⊆ v.

This means that e ⊆ 1 \ (a 4 a0 ) ∈ B and 1 \ e ⊆ 1 \ (b 4 b0 ) ∈ B. So we also have v ⊆ 1 \ (a 4 a0 ) and w ⊆ 1 \ (b 4 b0 ). Accordingly a ∩ v = a0 ∩ v,

b ∩ w = b0 ∩ w.

a \ w = a0 \ w,

b \ v = b0 \ v.

But this means that ¡ ¢ ¡ ¢ θ((a ∩ v) ∪ (b \ v)) ∩ E ∪ θ((a \ w) ∪ (b ∩ w)) \ E ¡ ¢ ¡ ¢ = θ((a0 ∩ v) ∪ (b0 \ v)) ∩ E ∪ θ((a0 \ w) ∪ (b0 ∩ w)) \ E . Thus the formula given defines θ1 uniquely. Q Q (c) Now θ1 is a lower density. P P(i) If a, b ∈ B, ¢ ¡ ¢¢• θ((a ∩ v) ∪ (b \ v)) ∩ E ∪ θ((a \ w) ∪ (b ∩ w)) \ E ¡ ¢ ¡ ¢ = ((a ∩ v) ∪ (b \ v)) ∩ e ∪ ((a \ w) ∪ (b ∩ w)) \ e

(θ1 ((a ∩ e) ∪ (b \ e)))• =

¡¡

= (a ∩ e) ∪ (b \ e). So (θ1 c)• = c for every c ∈ B1 . (ii)

¡ ¢ ¡ ¢ θ1 (0) = θ((0 ∩ v) ∪ (0 \ v)) ∩ E ∪ θ((0 \ w) ∪ (0 ∩ w)) \ E = ∅.

(iii) If a, a0 , b, b0 ∈ B, then

341G

The lifting theorem

165

θ1 (((a ∩ e) ∪ (b \ e)) ∩ ((a0 ∩ e) ∪ (b0 \ e))) = θ1 ((a ∩ a0 ∩ e) ∪ (b ∩ b0 \ e)) ¡ ¢ ¡ ¢ = θ((a ∩ a0 ∩ v) ∪ (b ∩ b0 \ v)) ∩ E ∪ θ((a ∩ a0 \ w) ∪ (b ∩ b0 ∩ w)) \ E ¡ ¢ = θ(((a ∩ v) ∪ (b \ v)) ∩ ((a0 ∩ v) ∪ (b0 \ v))) ∩ E ¡ ¢ ∪ θ(((a \ w) ∪ (b ∩ w)) ∩ ((a0 \ w) ∪ (b0 ∩ w))) \ E ¡ ¢ = θ((a ∩ v) ∪ (b \ v)) ∩ θ((a0 ∩ v) ∪ (b0 \ v)) ∩ E ¡ ¢ ∪ θ((a \ w) ∪ (b ∩ w)) ∩ θ((a0 \ w) ∪ (b0 ∩ w)) \ E ¢ ¡ ¢¢ ¡¡ = θ((a ∩ v) ∪ (b \ v)) ∩ E ∪ θ((a \ w) ∪ (b ∩ w)) \ E ¡¡ ¢ ¡ ¢¢ ∩ θ((a0 ∩ v) ∪ (b0 \ v)) ∩ E ∪ θ((a0 \ w) ∪ (b0 ∩ w)) \ E = θ1 ((a ∩ e) ∪ (b \ e)) ∩ θ1 ((a0 ∩ e) ∪ (b0 \ e)). Q So θ1 (c ∩ c0 ) = θ1 (c) ∩ θ1 (c0 ) for all c, c0 ∈ B1 . Q (d) If a ∈ B, then θ1 (a) = θ1 ((a ∩ e) ∪ (a \ e)) ¡ ¢ ¡ ¢ = θ((a ∩ v) ∪ (a \ v)) ∩ E ∪ θ((a \ w) ∪ (a ∩ w)) \ E ¡ ¢ ¡ ¢ = θ(a) ∩ E ∪ θ(a) \ E = θa. Thus θ1 extends θ, as required. 341G Lemma Let (X, Σ, µ) be a probability space and (A, µ ¯) its measure algebra. Suppose we have a sequence hθn in∈N of partial lower densities such that, for each n, (i) the domain Bn of θn is a closed subalgebra of A (ii) Bn ⊆ Bn+1 and θn+1 extends θn . Let B be the closed subalgebra of A generated by S B . Then there is a partial lower density θ, with domain B, extending every θn . n n∈N proof (a) For each n, set Σn = {E : E ∈ Σ, E • ∈ Bn }, and set Σ∞ = {E : E ∈ Σ, E • ∈ B}. Then (because all the Bn , B are σ-subalgebras of A, and E 7→ E • is sequentially order-continuous) all the Σn , Σ∞ are is just the σ-algebra Σ∗∞ of subsets of X S σ-subalgebras of Σ. We need to know that Σ∞ S generated by n∈N Σn . P P Because Σ∞ is a σ-algebra including n∈N Σn , Σ∗∞ ⊆ Σ∞ . On the other hand, ∗ • ∗ B = {E : E ∈ Σ∞ } is a σ-subalgebra of A including Bn for every n ∈ N. Because A is ccc, B∗ is (order-)closed (316Fb), so includes B. This means that if E ∈ Σ∞ there must be an F ∈ Σ∗∞ such that E • = F • . But now (E4F )• = 0 ∈ B0 , so E4F ∈ Σ0 ⊆ Σ∗∞ , and E also belongs to Σ∗∞ . This shows that Σ∞ ⊆ Σ∗∞ and the two algebras are equal. Q Q (b) For each n ∈ N, we have the partial lower density θn : Bn → Σ. Since (θn a)• = a ∈ Bn for every a ∈ Bn , θn takes all its values in Σn . For n ∈ N, let φn : Σn → Σn be the lower density corresponding to θn (341Ba), that is, φn E = θn E • for every E ∈ Σn . (c) For a ∈ A, n ∈ N choose Ga ∈ Σ, gan such that G•a = a and gan is a conditional expectation of χGa on Σn ; that is,

R

E

gan =

R

E

χGa = µ(E ∩ Ga ) = µ ¯(E • ∩ a)

for every E ∈ Σn . As remarked in 233Db, such a function gan can always be found, and moreover we may take it to be Σn -measurable and defined everywhere on X, by 232He. Now if a ∈ B, limn→∞ gan (x) exists and is equal to χGa (x) for almost every x. P P By Lévy’s martingale theorem (275I), limn→∞ S gan is defined almost everywhere and is a conditional expectation of χGa on the σ-algebra generated by n∈N Σn . As

166

Liftings

341G

observed in (a), this is just Σ∞ ; and as χGa is itself Σ∞ -measurable, it is also a conditional expectation of itself on Σ∞ , and must be equal almost everywhere to limn→∞ gan . Q Q (d) For a ∈ B, k ≥ 1, n ∈ N set Hkn (a) = {x : x ∈ X, gan (x) ≥ 1 − 2−k } ∈ Σn , θa =

T

S k≥1

T n∈N

m≥n

˜ kn (a) = φn (Hkn (a)), H

˜ km (a). H

The rest of the proof is devoted to showing that θ : B → Σ has the required properties. ˜ kn (0) (e) G0 is negligible, so every g0n is zero almost everywhere, every Hkn (0) is negligible and every H is empty; so θ0 = ∅. P Ga \ Gb is negligible, gan ≤ gbn almost everywhere for every n, every (f ) If a ⊆ b in B, then θa ⊆ θb. P ˜ kn (a) ⊆ H ˜ kn (b) for every n and k, and θa ⊆ θb. Q Hkn (a) \ Hkn (b) is negligible, H Q (g) If a, b ∈ B then θ(a ∩ b) = θa ∩ θb. P P χGa∩b ≥ χGa + χGb − 1 a.e. so ga∩b,n ≥ gan + gbn − 1 a.e. for every n. Accordingly Hk+1,n (a) ∩ Hk+1,n (b) \ Hkn (a ∩ b) is negligible, and (because φn is a lower density) ˜ kn (a ∩ b) ⊇ φn (Hk+1,n (a) ∩ Hk+1,n (b)) = H ˜ k+1,n (a) ∩ H ˜ k+1,n (b) H for all k ≥ 1, n ∈ N. Now, if x ∈ θa ∩ θb, then, for any k ≥ 1, there are n1 , n2 ∈ N such that T ˜ ˜ k+1,m (a), x ∈ T x ∈ m≥n1 H m≥n2 Hk+1,m (b). But this means that x∈

T m≥max(n1 ,n2 )

˜ km (a ∩ b). H

As k is arbitrary, x ∈ θ(a ∩ b); as x is arbitrary, θa ∩ θb ⊆ θ(a ∩ b). We know already from (f) that θ(a ∩ b) ⊆ θa ∩ θb, so θ(a ∩ b) = θa ∩ θb. Q Q P hgan in∈N → χGa a.e., so setting (h) If a ∈ B, then θa• = a. P T S T Va = k≥1 n∈N m≥n Hkm (a) = {x : lim inf n→∞ gan (x) ≥ 1}, Va 4Ga is negligible, and Va• = a; but θa4Va ⊆

S k≥1,n∈N

˜ kn (a) Hkn (a)4H

is also negligible, so θa• is also equal to a. Q Q Thus θ is a partial lower density with domain B. P If a ∈ Bn , then Ga ∈ Σm for every m ≥ n, so gam = χGa (i) Finally, θ extends θn for every n ∈ N. P a.e. for every m ≥ n; Hkm (a)4Ga is negligible for k ≥ 1, m ≥ n; ˜ km = φm Ga = θm a = θn a H for k ≥ 1, m ≥ n (this is where I use the hypothesis that θm+1 extends θm for every m); and θa =

\ [ \

˜ km (a) H

k≥1 r∈N m≥r

=

\ [ \

k≥1 r≥n m≥r

˜ km (a) = H

\ [

θn a = θn a. Q Q

k≥1 r≥n

The proof is complete. 341H

Now for the first main theorem.

Theorem Let (X, Σ, µ) be any strictly localizable measure space. Then it has a lower density φ : Σ → Σ. If µX > 0 we can take φX = X.

341I

The lifting theorem

167

proof : Part A I deal first with the case of probability spaces. Let (X, Σ, µ) be a probability space, and (A, µ ¯) its measure algebra. (a) Set κ = #(A) and enumerate A as haξ iξ<κ . For ξ ≤ κ let Aξ be the closed subalgebra of A generated by {aη : η < ξ}. I seek to define a lower density θ : A → Σ as the last of a family hθξ iξ≤κ , where θξ : Aξ → Σ is a partial lower density for each ξ. The inductive hypothesis will be that θξ extends θη whenever η ≤ ξ ≤ κ. To start the induction, we have A0 = {0, 1}, θ0 0 = ∅, θ0 1 = X. (b) Inductive step to a successor ordinal ξ Given a succesor ordinal ξ ≤ κ, express it as ζ + 1; we are supposing that θζ : Aζ → Σ has been defined. Now Aξ is the subalgebra of A generated by Aζ ∪ {aζ } (because this is a closed subalgebra, by 323K). So 341F tells us that θζ can be extended to a partial lower density θξ with domain Aξ . (c) Inductive step to a non-zero limit ordinal ξ of countable cofinality In this case, there is a strictly increasing sequence hζ(n)in∈N with supremum ξ. Applying 341G with Bn = S Aζ(n) , we see that there is a partial lower density θξ , with domain the closed subalgebra B generated by n∈N Aζ(n) , extending every θζ(n) . Now Aζ(n) ⊆ Aξ for every ξ, so B ⊆ Aξ ; but also, if η < ξ, there is an n ∈ N such that η < ζ(n), so that aη ∈ Aζ(n) ⊆ B; as η is arbitrary, Aξ ⊆ B and Aξ = B. Again, if η < ξ, there is an n such that η ≤ ζ(n), so that θζ(n) extends θη and θξ extends θη . Thus the induction continues. S (d) Inductive step to a limit ordinal ξ of uncountable cofinality In this case, Aξ = η<ξ Aη . P P Because A is ccc, every member a of Aξ must be in the closed subalgebra of A generated by some countable subset A of {aη : η < ξ} (331Gd-e). Now A can be expressed as {aη : η ∈ I} for some countable I ⊆ ξ. As I cannot be cofinal with ξ, there is a ζ < ξ such that η < ζ for every η ∈ I, so that A ⊆ Aζ and a ∈ Aζ . Q Q But now, because θζ extends θη whenever η ≤ ζ < ξ, we have a function θξ : Aξ → Σ defined by writing θξ a = θη a whenever η < ξ and a ∈ Aη . Because the family {Aη : η < ξ} is totally ordered and every θη is a partial lower density, θξ is a partial lower density. Thus the induction proceeds when ξ is a limit ordinal of uncountable cofinality. (e) The induction stops when we reach θκ : A → Σ, which is a lower density such that θκ 1 = X. Setting φE = θκ E • , φ is a lower density such that φX = X. Part B The general case of a strictly localizable measure space follows easily. First, if µX = 0, then A = {0} and we can set φ0 = ∅. Second, if µ is totally finite but not zero, we can replace it by ν, where νE = µE/µX for every E ∈ Σ; a lower density for ν is also a lower density for µ. Third, if µ is not totally finite, let hXS i ii∈I be a decomposition of X (211E). There is surely some j such that µXj > 0; replacing Xj by Xj ∪ {Xi : i ∈ I, µXi = 0}, we may assume that µXi > 0 for every i ∈ I. For each i ∈ I, let φi : Σi → Σi be a lower density of µi , where Σi = Σ ∩ PXi and µi = µ¹Σi , such that φi Xi = Xi . Then it is easy to check that we have a lower density φ : Σ → Σ given by setting S φE = i∈I φi (E ∩ Xi ) for every E ∈ Σ, and that φX = X. 341I The next step is to give a method of moving from lower densities to liftings. I start with an elementary remark on lower densities on complete measure spaces. Lemma Let (X, Σ, µ) be a complete measure space with measure algebra A. (a) Suppose that θ : A → Σ is a lower density and θ1 : A → PX is a function such that θ1 0 = ∅, θ1 (a ∩ b) = θ1 a ∩ θ1 b for all a, b ∈ A and θ1 a ⊇ θa for all a ∈ A. Then θ1 is a lower density. If θ1 is a Boolean homomorphism, it is a lifting. (b) Suppose that φ : Σ → Σ is a lower density and φ1 : Σ → PX is a function such that φ1 E = φ1 F whenever E4F is negligible, φ1 ∅ = ∅, φ1 (E ∩ F ) = φ1 E ∩ φ1 F for all E, F ∈ Σ and φ1 E ⊇ φE for all E ∈ Σ. Then φ1 is a lower density. If φ1 is a Boolean homomorphism, it is a lifting. proof (a) All I have to check is that θ1 a ∈ Σ and (θ1 a)• = a for every a ∈ A. But θa ⊆ θ1 a, So

θ(1 \ a) ⊆ θ1 (1 \ a),

θ1 a ∩ θ1 (1 \ a) = θ1 0 = ∅.

168

Liftings

341I

θa ⊆ θ1 a ⊆ X \ θ(1 \ a). Since (θa)• = a = (X \ θ(1 \ a))• , and µ is complete, θ1 is a lower density. If it is a Boolean homomorphism, then it is also a lifting (341De). (b) This follows by the same argument, or by looking at the functions from A to Σ defined by φ and φ1 and using (a). 341J Proposition Let (X, Σ, µ) be a complete measure space such that µX > 0, and A its measure algebra. (a) If θ : A → Σ is any lower density, there is a lifting θ : A → Σ such that θa ⊇ θa for every a ∈ A. (b) If φ : Σ → Σ is any lower density, there is a lifting φ : Σ → Σ such that φE ⊇ φE for every E ∈ Σ. proof (a) For each x ∈ θ1, set Ix = {a : a ∈ A, x ∈ θ(1 \ a)}. Then Ix is a proper ideal of A. P P We have 0 ∈ Ix , because x ∈ θ1, if b ⊆ a ∈ Ix then b ∈ Ix , because x ∈ θ(1 \ a) ⊆ θ(1 \ b), if a, b ∈ Ix then a ∪ b ∈ Ix , because x ∈ θ(1 \ a) ∩ θ(1 \ b) = θ(1 \ (a ∪ b)), 1∈ / Ix because x ∈ / ∅ = θ0. Q Q For x ∈ X \ θ1, set Ix = {0}; this is also a proper ideal of A, because A 6= {0}. By 311D, there is a surjective Boolean homomorphism πx : A → {0, 1} such that πx d = 0 for every d ∈ Ix . Define θ : A → PX by setting θa = {x : x ∈ X, πx (a) = 1} for every a ∈ A. It is easy to check that, because every πx is a surjective Boolean homomorphism, θ is a Boolean homomorphism. Now for any a ∈ A, x ∈ X, x ∈ θa =⇒ 1 \ a ∈ Ix =⇒ πx (1 \ a) = 0 =⇒ πx a = 1 =⇒ x ∈ θa. Thus θa ⊇ θa for every a ∈ A. By 341I, θ is a lifting, as required. (b) Repeat the argument above, or apply it, defining θ by setting θ(E • ) = φE for every E ∈ Σ, and φ by setting φE = θ(E • ) for every E. 341K The Lifting Theorem Every complete strictly localizable measure space of non-zero measure has a lifting. proof By 341H, it has a lower density, so by 341J it has a lifting. 341L Remarks If we count 341F-341K as a single argument, it may be the longest proof, after Carleson’s theorem (§286), which I have yet presented in this treatise, and perhaps it will be helpful if I suggest ways of looking at its components. (a) The first point is that the theorem should be thought of as one about probability spaces. The shift to general strictly localizable spaces (Part B of the proof of 341H) is purely a matter of technique. I would not have presented it if I did not think that it’s worth doing, for a variety of reasons, but there is no significant idea needed, and if – for instance – the result were valid only for σ-finite spaces, it would still be one of the great theorems of mathematics. So the rest of these remarks will be directed to the ideas needed in probability spaces. (b) All the proofs I know of the theorem depend in one way or another on an inductive construction. We do not, of course, need a transfinite induction written out in the way I have presented it in 341H above. Essentially the same proof can be presented as an application of Zorn’s Lemma; if we take P to be the set of partial lower densities, then the arguments of 341G and part (A-d) of the proof of 341H can be adapted

341Lf

The lifting theorem

169

to prove that any totally ordered subset of P has an upper bound in P , while the argument of 341F shows that any maximal element of P must have domain A. I think it is purely a matter of taste which form one prefers. I suppose I have used the ordinal-indexed form largely because that seemed appropriate for Maharam’s theorem in the last chapter. (c) There are then three types of inductive step to examine, corresponding to 341F, 341G and (A-d) in 341H. The first and last are easier than the second. Seeking the one-step extension of θ : B → Σ to θ1 : B1 → Σ, the natural model to use is the one-step extension of a Boolean homomorphism presented in 312N. The situation here is rather more complicated, as θ1 is not fully specified by the value of θ1 e, and we do in fact have more freedom at this point than is entirely welcome. The formula used in the proof of 341F ¨ cker 76. is derived from Graf & Weizsa (d) At this point I must call attention to the way in which the whole proof is dominated by the choice of closed subalgebras as the domains of our partial liftings. This is what makes the inductive step to a S limit ordinal ξ of countable cofinality difficult, because Aξ will ordinarily be larger than η<ξ Aη . But it is absolutely essential in the one-step extensions. Because we are dealing with a ccc algebra A, the S requirement that the Aξ should be closed is not a problem when cf(ξ) is uncountable, since in this case η<ξ Aη is already a closed subalgebra; this is the only idea needed in (A-d) of 341H. (e) So we are left with the inductive step to ξ when cf(ξ) = ω, which is 341G. Here we actually need some measure theory, and a particularly striking bit. (You will see that the measure µ, as opposed to the algebras Σ and A and the homomorphism E 7→ E • and the ideal of negligible sets, is simply not mentioned anywhere else in the whole argument.) (i) The central idea is to use the fact that bounded martingales converge to define θa in of a sequence of conditional expectations. Because I have chosen a fairly direct assault on the problem, some of the surrounding facts are not perhaps so clearly visible as they might have been if I had used a more leisurely route. For each a ∈ A, I start by choosing a representative Ga ∈ Σ; let me emphasize that this is a crude application of the axiom of choice, and that the different sets Ga are in no way coordinated. (The theorem we are proving is that they can be coordinated, but we have not reached that point yet.) Next, I choose, arbitrarily, a conditional expectation gan of χGa on Σn . Once again, the choices are not coordinated; but the martingale theorem assures us that ga = limn→∞ gan is defined almost everywhere, and is equal almost everywhere to χGa if a ∈ B. Of course I could have gone to the gan without mentioning the Ga ; they are set up as Radon-Nikod´ ym derivatives of the countably additive functionals E 7→ µ ¯(E • ∩ a) : Σn → R. Now the gan , like the Ga , are not uniquely defined. But they are defined ‘up to a negligible set’, so that any alternative 0 0 = gan a.e. This means that the sets Hkn (a) = {x : gan (x) ≥ 1 − 2−k } are functions gan would have gan ˜ kn (a) = φn (Hkn (a)) are uniquely defined. I also defined ‘up to a negligible set’, and consequently the sets H point this out to show that it is not a complete miracle that we have formulae ˜ kn (a) ⊆ H ˜ kn (b) if a ⊆ b, H ˜ kn (a ∩ b) ⊇ H ˜ k+1,n (a) ∩ H ˜ k+1,n (b) for all a, b ∈ A H which do not ask us to turn a blind eye to any negligible sets. I note in ing that I could have defined ˜ kn (a) without mentioning the gan ; in fact the H ˜ kn (a) = θn (sup{c : c ∈ Bn , µ H ¯(a ∩ d) ≥ 1 − 2−k µ ¯d whenever d ∈ Bn , d ⊆ c}). ˜ kn (a) in hand, we can look at (ii) Now, with the sets H T S T ˜ kn (a); Vã = k≥1 n∈N m≥n H because gan → χGa a.e., Vã 4Ga is negligible and Vã• = a for every a ∈ Aξ . The rest of the argument amounts to checking that a 7→ Vã will serve for θ. (f ) The arguments above apply to all probability spaces, and show that every probability space has a lower density. The next step is to convert a lower density into a lifting. It is here that we need to assume

170

Liftings

341Lf

completeness. The point is that we can find a Boolean homomorphism θ : A → PX such that θa ⊆ θa for every a; this corresponds just to extending the ideals Ix = {a : x ∈ θ(1 \ a)} to maximal ideals (and giving a moment’s thought to x ∈ X \ θ1). In order to ensure that θa ∈ Σ and (θa)• = a, we have to observe that θa is sandwiched between θa and X \ θ(1 \ a), which differ by a negligible set; so that if µ is complete all will be well. (g) The fact that completeness is needed at only one point in the argument makes it natural to wonder whether the theorem might be true for probability spaces in general. (I will come later, in 341M, to nonstrictly-localizable spaces.) There is as yet no satisfactory answer to this. For Borel measure on R, the question is known to be undecidable from the ordinary axioms of set theory (including the axiom of choice, but not the continuum hypothesis, as usual); I will give some of the arguments in Volume 5. (For the moment, I refer you to the discussion in Fremlin 89, §4, and to Burke 93.) But I conjecture that there is an counter-example under the ordinary axioms (see 341Z below). (h) Quite apart from whether completeness is needed in the argument, it is not absolutely clear why measure theory is required. The general question of whether a lifting exists can be formulated for any triple (X, Σ, I) where X is a set, Σ is a σ-algebra of subsets of X, and I is a σ-ideal of Σ. (See 341Ya below.) S.Shelah has given an example of such a triple without a lifting in which two of the basic properties of the measure-theoretic case are satisfied: (X, Σ, I) is ‘complete’ in the sense that every subset of any member of I belongs to Σ (and therefore to I), and I is ω1 -saturated in Σ in the sense of 316C (see Shelah Sh636, Burke n96). But many other cases are known (e.g., 341Yb) in which liftings do exist. (i) It is of course possible to prove 341K without mentioning ‘lower densities’, and there are even some advantages in doing so. The idea is to follow the lines of 341H, but with ‘liftings’ instead of ‘lower densities’ throughout. The inductive step to a successor ordinal is actually easier, because we have a Boolean homomorphism θ in 341F to extend, and we can use 312N as it stands if we can choose the pair E, F = X \ E correctly. The inductive step to an ordinal of uncountable cofinality remains straightforward. But in the inductive step to an ordinal of countable cofinality, we find that in 341G we get no help from assuming that the θn are actually liftings; we are still led to to a lower density θ. So at this point we have to interpolate the argument of 341J to convert this lower density into a lifting. I have chosen the more leisurely exposition, with the extra concept, partly in order to get as far as possible without assuming completeness of the measure and partly because lower densities are an important tool for further work (see §§345-346). (j) For more light on the argument of 341G see also 363Xe and 363Yg below. 341M I remarked above that the shift from probability spaces to general strictly localizable spaces was simply a matter of technique. The question of which spaces have liftings is also primarily a matter concerning probability spaces, as the next result shows. Proposition Let (X, Σ, µ) be a complete locally determined space with µX > 0. Then it has a lifting iff it has a lower density iff it is strictly localizable. proof If (X, Σ, µ) is strictly localizable then it has a lifting, by 341K. A lifting is already a lower density, and if (X, Σ, µ) has a lower density it has a lifting, by 341J. So we have only to prove that if it has a lifting then it is strictly localizable. Let θ : A → Σ be a lifting, where A is the measure algebra of (X, Σ, µ). Let C be a partition of unity in A consisting of elements of finite measure (322Ea). Set A = {θc : c ∈ C}. Because C is dist, so is A. Because sup C = 1 in A, every set of positive measure meets some member of A in a set of positive measure. So the conditions of 213O are satisfied, and (X, Σ, µ) is strictly localizable. 341N Extension of partial liftings The following facts are obvious from the proof of 341H, but it will be useful to have them out in the open.

341P

The lifting theorem

171

Proposition Let (X, Σ, µ) be a probability space and T a σ-subalgebra of Σ. (a) Any partial lower density φ0 : T → Σ has an extension to a lower density φ : Σ → Σ. (b) Suppose now that µ is complete. If φ0 is a Boolean homomorphism, it has an extension to a lifting φ for µ. proof (a) In Part A of the proof of 341H, let Aξ be the closed subalgebra of A generated by {E • : E ∈ T} ∪ {aη : η < ξ}, and set θ0 E • = φ0 E for every E ∈ T. Proceed with the induction as before. The only difference is that we no longer have a guarantee that φX = X. (b) Suppose now that φ0 is a Boolean homomorphism. 341J tells us that there is a lifting φ : Σ → Σ such that φE ⊇ φE for every E ∈ Σ. But if E ∈ T we must have φE ⊇ φ0 E, φE \ φ0 E = φE ∩ φ0 (X \ E) ⊆ φE ∩ φ(X \ E) = ∅, so that φE = φ0 E, and φ extends φ0 . 341O Liftings and Stone spaces The arguments of this section so far involve repeated use of the axiom of choice, and offer no suggestion that any liftings (or lower densities) are in any sense ‘canonical’. There is however one context in which we have a distinguished lifting. Suppose that we have the Stone space (Z, T, ν) of a measure algebra (A, µ ¯); as in 311E, I think of Z as being the set of surjective Boolean homomorphisms from A to Z2 , so that each a ∈ A corresponds to the open-and-closed set b a = {z : z(a) = 1}. Then we have a lifting θ : A → T defined by setting θa = b a for each a ∈ A. (I am identifying A with the measure algebra of ν, as in 321J.) The corresponding lifting φ : T → T is defined by taking φE to be that unique open-and-closed set such that E4φE is negligible (or, if you prefer, meager). Generally, liftings can be described in of Stone spaces, as follows. 341P Proposition Let (X, Σ, µ) be a measure space, (A, µ ¯) its measure algebra, and (Z, T, ν) the Stone space of (A, µ ¯) with its canonical measure. (a) There is a one-to-one correspondence between liftings θ : A → Σ and functions f : X → Z such that f −1 [b a] ∈ Σ and (f −1 [b a])• = a for every a ∈ A, defined by the formula θa = f −1 [b a] for every a ∈ A. (b) If (X, Σ, µ) is complete and locally determined, then a function f : X → Z satisfies the conditions of (a) iff (α) it is inverse-measure-preserving (β) the homomorphism it induces between the measure algebras of µ and ν is the canonical isomorphism defined by the construction of Z. proof Recall that T is just the set {b a4M : a ∈ A, M ⊆ Z is meager}, and that ν(b a4M ) = µ ¯a for all such a, M ; while the canonical isomorphism π between A and the measure algebra of ν is defined by the formula πF • = a whenever F ∈ T, a ∈ A and F 4b a is meager (341K). (a) If θ : A → Σ is any Boolean homomorphism, then for every x ∈ X we have a surjective Boolean homomorphism fθ (x) : A → Z2 defined by saying that fθ (x)(a) = 1 if x ∈ θa, 0 otherwise. fθ is a function from X to Z. We can recover θ from fθ by the formula θa = {x : fθ (x)(a) = 1} = {x : fθ (x) ∈ b a} = fθ−1 [b a]. So fθ−1 [b a] ∈ Σ and, if θ is a lifting, (fθ−1 [b a])• = (θa)• = a. for every a ∈ A. Similarly, given a function f : X → Z with this property, then we can set θa = f −1 [b a] for every a ∈ A to obtain a lifting θ : A → Σ; and of course we now have f (x)(a) = 1 ⇐⇒ f (x) ∈ b a ⇐⇒ x ∈ θa, so fθ = f .

172

Liftings

341P

(b) Assume now that (X, Σ, µ) is complete and locally determined. (i) Let f : X → Z be the function associated with a lifting θ, as in (a). I show first that f is inversemeasure-preserving. P P If F ∈ T, express it as b a4M , where a ∈ A and M ⊆ Z is meager. By 322F, A is weakly (σ, ∞)-distributive, so M is nowhere dense (316I). Consider f −1 [M ]. If E ⊆ X is measurable and of finite measure, then E ∩ f −1 [M ] has a measurable envelope H (132Ed). ?? If µH > 0, then b = H • 6= 0 and bb is a non-empty open set in Z. Because M is nowhere dense, there is a non-zero a ∈ A such that b a ⊆ bb \ M . Now µ(f −1 [bb]4H) = 0, so f −1 [b a] \ H is negligible, and f −1 [b a] ∩ H is a non-negligible measurable −1 set dist from E ∩ f [M ] and included in H; which is impossible. X X Thus H and E ∩ f −1 [M ] are negligible. This is true for every measurable set E of finite measure. Because µ is complete and locally determined, f −1 [M ] ∈ Σ and µf −1 [M ] = 0. So f −1 [F ] = f −1 [b a]4f −1 [M ] is measurable, and µf −1 [F ] = µf −1 [b a] = µθa = µ ¯a = νb a = νF . As F is arbitrary, f is inverse-measure-preserving. Q Q It follows at once that for any F ∈ T, f −1 [F ]• = a = πF • where a is that element of A such that M = F 4a is meager, because in this case f −1 [b a]• = a, by (a), while −1 f [M ] is negligible. So π is the homomorphism induced by f . (ii) Now suppose that f : X → Z is an inverse-measure-preserving function such that f −1 [F ]• = πF • for every F ∈ T. Then, in particular, f −1 [b a]• = πb a• = a for every a ∈ A, so that f satisfies the conditions of (a). 341Q Corollary Let (X, Σ, µ) be a strictly localizable measure space, (A, µ ¯) its measure algebra, and Z the Stone space of A; suppose that µX > 0. For E ∈ Σ write E ∗ for the open-and-closed subset of Z corresponding to E • ∈ A. Then there is a function f : X → Z such that E4f −1 [E ∗ ] is negligible for every E ∈ Σ. If µ is complete, then f is inverse-measure-preserving. ˆ its domain. Then we can identify (A, µ proof Let µ ˆ be the completion of µ, and Σ ¯) with the measure ˆ algebra of µ ˆ (322Da). Let θ : A → Σ be a lifting, and f : X → Z the corresponding function. If E ∈ Σ then ˆ = Σ, then f E∗ = b a where a = E • , so E4f −1 [E ∗ ] = E4θE • is negligible. If µ is itself complete, so that Σ is inverse-measure-preserving, by 341Pb. 341X Basic exercises (a) Let (X, Σ, µ) be a measure space and φ : Σ → Σ a function. Show that φ is a lifting iff it is a lower density and φE ∪ φ(X \ E) = X for every E ∈ Σ. > (b) Let µ be the usual measure on X = {0, 1}N , and Σ its domain. For x ∈ X and n ∈ N set Un (x) = {y : y ∈ X, y¹n = x¹n}. For E ∈ Σ set φE = {x : limn→∞ 2n µ(E ∩ Un (x)) = 1}. Show that φ is a lower density of (X, Σ, µ). > (c) Let P be the set of all lower densities of a complete measure space (X, Σ, µ), with measure algebra A, ordered by saying that θ ≤ θ0 if θa ⊆ θ0 a for every a ∈ A. Show that any non-empty totally ordered subset of P has an upper bound in P . Show that if θ ∈ P and a ∈ A and x ∈ X \ (θa ∪ θ(1 \ a)), then θ0 : A → Σ is a lower density, where θ0 b = θb ∪ {x} if either a ⊆ b or there is a c ∈ A such that x ∈ θc and a ∩ c ⊆ b, and θ0 b = θb otherwise. Hence prove 341J. (d) Let (X, Σ, µ) and (Y, T, ν) be measure spaces and suppose that there is an inverse-measure-preserving function f : X → Y such that the associated homomorphism from the measure algebra of ν to that of µ is an isomorphism. Show that for every lifting φ of (Y, T, ν) we have a corresponding lifting ψ of (X, Σ, µ) defined uniquely by the formula ψ(f −1 [F ]) = f −1 [φF ] for every F ∈ T.

341 Notes

The lifting theorem

173

(e) Let (X, Σ, µ) be a measure space, and write L∞ (Σ) for the linear space of all bounded Σ-measurable functions from X to R. Show that for any lifting φ : Σ → Σ of µ there is a unique linear operator T : L∞ (µ) → L∞ (Σ) such that T (χE)• = χ(φE) for every E ∈ Σ and T u ≥ 0 in L∞ (Σ) whenever u ≥ 0 in L∞ (µ). Show that (i) (T u)• = u and supx∈X |(T u)(x)| = kuk∞ for every u ∈ L∞ (µ) (ii) T (u × v) = T u × T v for all u, v ∈ L∞ (µ). 341Y Further exercises (a) Let X be a set, Σ an algebra of subsets of X and I an ideal of Σ; let A be the quotient Boolean algebra Σ/I. We say that a function θ : A → Σ is a lifting if it is a Boolean homomorphism and (θa)• = a for every a ∈ A, and that θ : A → Σ is a lower density if θ0 = ∅, θ(a ∩ b) = θa ∩ θb for all a, b ∈ A, and (θa)• = a for every a ∈ A. Show that if (X, Σ, I) is ‘complete’ in the sense that F ∈ Σ whenever F ⊆ E ∈ I, and if X ∈ / I, and θ : A → Σ is a lower density, then there is a lifting θ : A → Σ such that θa ⊆ θa for every a ∈ A. (b) Let X be a non-empty Baire space, Bb the σ-algebra of subsets of X with the Baire property (314Yd) b and M the ideal of meager subsets of X. Show that there is a lifting θ from B/M to Bb such that θG• ⊇ G • for every open G ⊆ X. (Hint: in 341Ya, set θ(G ) = G for every regular open set G.) (c) Let (X, Σ, µ) be a Maharam-type-homogeneous probability space with Maharam type κ ≥ ω. Let Σ be the Baire σ-algebra of Y = {0, 1}κ , that is, the σ-algebra of subsets of Y generated by the family {{x : x(ξ) = 1} : ξ < κ}, and let ν be the restriction to Σ of the usual measure on {0, 1}κ . Show that there is an inverse-measure-preserving function f : X → Y which induces an isomorphism between their measure algebras. (d) Let (X, Σ, µ) be a complete Maharam-type-homogeneous probability space with Maharam type κ ≥ ω, and let ν be the usual measure on {0, 1}κ . Show that there is an inverse-measure-preserving function f : X → Y which induces an isomorphism between their measure algebras. (e) Let (X, Σ, µ) be a semi-finite measure space which is not purely atomic. Write L1strict for the linear space of integrable functions f : X → R. Show that there is no operator T : L1 (µ) → L1strict such that (i) (T u)• = u for every u ∈ L1 (µ) (ii) T u ≥ T v whenever u ≥ v in L1 (µ). (Hint: Suppose first that µ is the usual measure on {0, 1}N . Let F be the countable set of continuous functions f : {0, 1}N → N. Show that if T satisfies (i) then there is an x ∈ {0, 1}N such that T (f • )(x) = f (x) for every f ∈ F ; find a sequence hfn in∈N in F such that {fn• : n ∈ N} is bounded above in L1 (µ) but supn∈N fn (x) = ∞. Now transfer this argument to some atomless fragment of X.) 341Z Problems (a) Can we construct, using the ordinary axioms of mathematics (including the axiom of choice, but not the continuum hypothesis), a probability space (X, Σ, µ) with no lifting? (b) Set κ = ω3 . (There is a reason for taking ω3 here; see Volume 5, when it appears, or Fremlin 89.) Let Σ be the Baire σ-algebra of X = {0, 1}κ (as in 341Yc), and let µ be the restriction to Σ of the usual measure on {0, 1}κ . Does (X, Σ, µ) have a lifting? 341 Notes and comments Innumerable variations of the proof of 341K have been devised, as each author has struggled with the technical complications. I have discussed the reasons for my own choices in 341L. The theorem has a curious history. It was originally announced by von Neumann, but he seems never to have written his proof down, and the first published proof is that of Maharam 58. That argument is based on Maharam’s theorem, 341Xd and 341Yd, which show that it is enough to find liftings for every {0, 1}κ ; this requires most of the ideas presented above, but feels more concrete, and some of the details are slightly simpler. The argument as I have written it owes a great deal to Ionescu Tulcea & Ionescu Tulcea 69. The lifting theorem and Maharam’s theorem are the twin pillars of modern abstract measure theory. But there remains a degree of mystery about the lifting theorem which is absent from the other. The first point is that there is nothing canonical about the liftings we can construct, except in the quite exceptional case of Stone spaces (341O). Even when there is a more or less canonical lower density present (341E, 341Xb), the conversion of this into a lifting requires arbitrary choices, as in 341J. While we can distinguish some

174

The lifting theorem

341 Notes

liftings as being somewhat more regular than others, I know of no criterion which marks out any particular lifting of Lebesgue measure, for instance, among the rest. Perhaps associated with this arbitrariness is the extreme difficulty of deciding whether liftings of any given type exist. Neither positive nor negative results are easily come by (I will present a few in the later sections of this chapter), and the nature of the obstacles remains quite unclear.

342 Compact measure spaces The next three sections amount to an extended parenthesis, showing how the Lifting Theorem can be used to attack one of the fundamental problems of measure theory: the representation of Boolean homomorphisms between measure algebras by functions between appropriate measure spaces. This section prepares for the main idea by introducing the class of ‘locally compact’ measures (342Ad), with the associated concepts of ‘compact’ and ‘perfect’ measures (342Ac, 342K). These depend on the notions of ‘inner regularity’ (342Aa, 342B) and ‘compact class’ (342Ab, 342D). I list the basic permanence properties for compact and locally compact measures (342G-342I) and mention some of the compact measures which we have already seen (342J). Concerning perfect measures, I content myself with the proof that a locally compact measure is perfect (342L). I end the section with two examples (342M, 342N). 342A Definitions (a) Let (X, Σ, µ) be a measure space. If K ⊆ PX, I will say that µ is inner regular with respect to K if µE = sup{µK : K ∈ K ∩ Σ, K ⊆ E} for every E ∈ Σ. Of course µ is inner regular with respect to K iff it is inner regular with respect to K ∩ Σ. T (b) A family K of sets is a compact class if K0 6= ∅ whenever K0 ⊆ K has the finite intersection property. Note that any subset of a compact class is again a compact class. (In particular, it is convenient to allow the empty set as a compact class.) (c) A measure space (X, Σ, µ), or a measure µ, is compact if µ is inner regular with respect to some compact class of subsets of X. Allowing ∅ as a compact class, and interpreting sup ∅ as 0 in (a) above, µ is a compact measure whenever µX = 0. (d) A measure space (X, Σ, µ), or a measure µ, is locally compact if the subspace measure µE is compact whenever E ∈ Σ and µE < ∞. Remark I ought to point out that the original definitions of ‘compact class’ and ‘compact measure’ (Marczewski 53) correspond to what I call ‘countably compact class’ and ‘countably compact measure’ in Volume 4. For another variation on the concept of ‘compact class’ see condition (β) in 343B(ii)-(iii). For examples of compact measure spaces see 342J. 342B

I prepare the ground with some straightforward lemmas.

Lemma Let (X, Σ, µ) be a measure space, and K ⊆ Σ a set such that whenever E ∈ Σ and µE > 0 there is a K ∈ K such that K ⊆ E and µK > 0. Let E ∈ Σ. S (a) There is a countable dist set K1 ⊆ K such that K ⊆ E for every K ∈ K1 and µ( K1 ) = µE. S (b) If µE < ∞ then µ(E \ K1 ) = 0. (c) S In any case, there is for any γ < µE a finite dist K0 ⊆ K such that K ⊆ E for every K ∈ K0 and µ( K0 ) ≥ γ.

342F

Compact measure spaces

175

proof Set K0 = {K : K ∈ K, K ⊆ E, µK > 0}. Let K∗ be a maximal dist subfamily of K0 . If K∗ is uncountable, then there is someSn ∈ N such that {K : K ∈ K∗ , µK ≥ 2−n } is infinite, so that there is a countable K1 ⊆ K∗ such that µ( K1 ) = ∞ = µE. S If K∗ is countable, set K1 = K∗ . Then F = K1 is measurable, and F ⊆ E. Moreover, there is no member of K0 dist from F ; but this means that E \ F must be negligible. So µF = µE, and (a) is true. Now (b) and (c) follow at once, because S S µ( K1 ) = sup{µ( K0 ) : K0 ⊆ K1 is finite}. Remark This lemma can be thought of as two more versions of the principle of exhaustion; compare 215A. 342C Corollary Let (X, Σ, µ) be a measure space and K ⊆ PX a family of sets such that (α) K ∪K 0 ∈ K whenever K, K 0 ∈ K and K ∩ K 0 = ∅ (β) whenever E ∈ Σ and µE > 0, there is a K ∈ K ∩ Σ such that K ⊆ E and µK > 0. Then µ is inner regular with respect to K. proof Apply 342Bc to K ∩ Σ. 342D Lemma Let X be a set and K a family of subsets of X. (a) The following are equiveridical: (i) K is a compact class; (ii) there is a topology T on X such that X is compact and every member of K is a closed set for T. T (b) If K is a compact class, so are the families K1 = {K0 ∪ . . . ∪ Kn : K0 , . . . , Kn ∈ K} and K2 = { K0 : ∅ 6= K0 ⊆ K}. proof (a)(i)⇒(ii) Let T be the topology generated by {X \ K : K ∈ K}. Then of course every member of K is closed for T. Let F be an ultrafilter on X. Then K ∩ F has T the finite intersection property; because K is a compact class, it has non-empty intersection; take x ∈ X (K ∩ F). The family {G : G ⊆ X, either G ∈ F or x ∈ / G} is easily seen to be a topology on X, and contains X \ K for every K ∈ K (because if X \ K ∈ / F then K ∈ F and x ∈ K), so includes T; but this just means that every T-open set containing x belongs to F, that is, that F → x. As F is arbitrary, X is compact for T (2A3R). (ii)⇒(i) Use 3A3Da. (b) Let T be a topology on X such that X is compact and every member of K is closed for T; then the same is true of every member of K1 or K2 . 342E Corollary Suppose that (X, Σ, µ) is a measure space and that K is a compact class such that whenever E ∈ Σ and µE > 0 there is a K ∈ K ∩ Σ such that K ⊆ E and µK > 0. Then µ is compact. proof Set K1 = {K0 ∪ . . . ∪ Kn : K0 , . . . , Kn ∈ K}. By 342Db, K1 is a compact class, and by 342C µ is inner regular with respect to K1 . 342F Corollary A measure space (X, Σ, µ) is compact iff there is a topology on X such that X is compact and µ is inner regular with respect to the closed sets. proof (a) If µ is inner regular with respect to a compact class K, then there is a compact topology on X such that every member of K is closed; now the family F of closed sets includes K, so µ is also inner regular with respect to F. (b) If there is a compact topology on X such that µ is inner regular with respect to the family K of closed sets, then this is a compact class, so µ is a compact measure.

176

The lifting theorem

342G

342G Now I look at the standard questions concerning preservation of the properties of ‘compactness’ or ‘local compactness’ under the usual manipulations. Proposition (a) Any measurable subspace of a compact measure space is compact. (b) The completion and c.l.d. version of a compact measure space are compact. (c) A semi-finite measure space is compact iff its completion is compact iff its c.l.d. version is compact. (d) The direct sum of a family of compact measure spaces is compact. (e) The c.l.d. product of two compact measure spaces is compact. (f) The product of any family of compact probability spaces is compact. proof (a) Let (X, Σ, µ) be a compact measure space, and E ∈ Σ. If K is a compact class such that µ is inner regular with respect to K, then KE = K ∩ PE is a compact class (just because it is a subset of K) and the subspace measure µE is inner regular with respect to KE . ˇ µ (b) Let (X, Σ, µ) be a compact measure space. Write (X, Σ, ˇ) for either the completion or the c.l.d. version of (X, Σ, µ). Let K ⊆ PX be a compact class such that µ is inner regular with respect to K. Then ˇ and γ < µ µ ˇ is also inner regular with respect to K. P P If E ∈ Σ ˇE there is an E 0 ∈ Σ such that E 0 ⊆ E 0 0 and µE > γ; if µ ˇ is the c.l.d. version of µ, we may take µE to be finite. There is a K ∈ K ∩ Σ such that ˇ Q K ⊆ E 0 and µK ≥ γ. Now µ ˇK = µK ≥ γ and K ⊆ E and K ∈ K ∩ Σ. Q ˇ µ (c) Now suppose that (X, Σ, µ) is semi-finite; again write (X, Σ, ˇ) for either its completion or its c.l.d. version. We already know that if µ is compact, so is µ ˇ. IfTµ ˇ is compact, let K ⊆ PX be a compact class such that µ ˇ is inner regular with respect to K. Set K∗ = { K0 : ∅ 6= K0 ⊆ K}; then K∗ is a compact class (342Db). Now µ is inner regular with respect to K∗ . P P Take E ∈ Σ and γ < µE. Choose hEn in∈N , hKn in∈N as follows. Because µ is semi-finite, there is an E0 ⊆ E such that E0 ∈ Σ and γ < µE0 < ∞. Given En ∈ Σ ˇ such that Kn ⊆ En and µ such that µEn > γ, there is a Kn ∈ K ∩ Σ ˇKn > γ. Now there ∈ Σ such Tis an En+1 T that En+1 ⊆ Kn and µEn+1 > γ. Continue. On completing the induction, set K = n∈N Kn = n∈N En , so that K ∈ K∗ ∩ Σ and K ⊆ E and µK = limn→∞ µEn ≥ γ. As E, γ are arbitrary, µ is inner regular with respect to K∗ . Q Q As K∗ is a compact class, µ is a compact measure. (d) Let h(Xi , Σi , µi )ii∈I be a family of compact measure spaces, with direct sum (X, Σ, µ). We may suppose that each Xi is actually a subset of X, with µi the subspace measure. For S each i ∈ I let Ki ⊆ PXi be a compact class such that µi is inner regular with respect to Ki . Then K = i∈I Ki is a compact class, for if K0 ⊆ K has the finite intersection property, then K0 ⊆ Ki for some i, so has non-empty intersection. Now if E ∈ Σ, µE > 0 there is some i ∈ I such that µi (E ∩ Xi ) > 0, and we can find a K ∈ Ki ∩ Σi ⊆ K ∩ Σ such that K ⊆ E ∩ Xi and µi K > 0, in which case µK > 0. By 342E, µ is compact. (e) Let (X, Σ, µ) and (Y, T, ν) be two compact measure spaces, with c.l.d. product measure (X × Y, Λ, λ). Let T, S be topologies on X, Y respectively such that X and Y are compact spaces and µ, ν are inner regular with respect to the closed sets. Then the product topology on X × Y is compact (3A3J). The point is that λ is inner regular with respect to the family K of closed subsets of X × Y . P P Suppose that W ∈ Λ and λW > γ. Then there are E ∈ Σ, F ∈ T such that µE < ∞, νF < ∞ and λ(W ∩(E×F )) > γ (251F). Now there are sequences hEn in∈N , hFn in∈N in Σ, T respectively such that S (E × F ) \ W ⊆ n∈N En × Fn , P∞ n=0

µEn · νFn < λ((E × F ) \ W ) + λ((E × F ) ∩ W ) − γ = λ(E × F ) − γ

(251C). Set W 0 = (E × F ) \

S n∈N

E n × Fn =

T

n∈N ((E

× (F \ Fn )) ∪ ((E \ En ) × F )).

Then W 0 ⊆ W , and

S P∞ λ((E × F ) \ W 0 ) ≤ λ( n∈N En × Fn ) ≤ n=0 µEn · νFn < λ(E × F ) − γ,

so λW 0 > γ. Set ² = 41 (λW 0 − γ)/(1 + µE + µF ). For each n, we can find closed measurable sets Kn , Kn0 ⊆ X and Ln , L0n ⊆ Y such that Kn ⊆ E,

µ(E \ Kn ) ≤ 2−n ²,

342H


L0n ⊆ F \ Fn ,

ν((F \ Fn ) \ L0n ) ≤ 2−n ²,

Kn0 ⊆ E \ En ,

µ((E \ En ) \ Kn0 ) ≤ 2−n ²,

Ln ⊆ F , Set V =

T

n∈N (Kn

177

ν(F \ Ln ) ≤ 2−n ².

× L0n ) ∪ (Kn0 × Ln ) ⊆ W 0 ⊆ W .

Now W0 \ V ⊆

[

((E \ Kn ) × F ) ∪ (E × ((F \ Fn ) \ L0n ))

n∈N

∪ (((E \ En ) \ Kn0 ) × F ) ∪ (E × (F \ Ln )), so λ(W 0 \ V ) ≤

∞ X

µ(E \ Kn ) · νF + µE · ν((F \ Fn ) \ L0n )

n=0

+ µ((E \ En ) \ Kn0 ) · νF + µE · ν(F \ Ln ) ≤

∞ X

2−n ²(2µE + 2µF ) ≤ λW 0 − γ,

n=0

and λV ≥ γ. But V is a countable intersection of finite unions of products of closed measurable sets, so is itself a closed measurable set, and belongs to K ∩ Λ. Q Q Accordingly the product topology on X × Y witnesses that λ is a compact measure. (f ) The same method works. In detail: let h(Xi , Σi , µi )ii∈I be a family of compact probability spaces, with product (X, Λ, λ). For each i, let Ti be a topology on Xi such that Xi is compact and µi is inner regular with respect to the closed sets. Give X the product topology; this is compact. If W ∈ Λ and S ² > 0, let hCn in∈N be a sequence of measurable cylinder sets (in the sense of 254A) such that X \ W ⊆ n∈N Cn and Q P∞ i∈I Eni where Eni ∈ Σi for each i and Jn = {i : Eni 6= Xi } n=0 λCn ≤ λ(X \ W ) + ². Express each Cn as −n is finite. For n ∈ N, i ∈ Jn set ²ni = 2 ²/(1 + #(Jn )). Choose closed measurable sets Kni ⊆ Xi \ Eni such that µi ((Xi \ Eni ) \ Kni ) ≤ ²ni for n ∈ N, i ∈ Jn . For each n ∈ N, set S Vn = i∈Jn {x : x ∈ X, x(i) ∈ Kni }, so that Vn is a closed measurable subset of X. Observe that X \ Vn = {x : x(i) ∈ X \ Kni for i ∈ Jn } includes Cn , and that

P P λ(X \ (Vn ∪ Cn )) ≤ i∈Jn λ{x : x(i) ∈ Xi \ (Kni ∪ Eni )} ≤ i∈Jn ²ni ≤ 2−n ². T Now set V = n∈N Vn ; then V is again a closed measurable set, and S X \ V ⊆ n∈N Cn ∪ (X \ (Cn ∪ Vn ))

has measure at most

P∞ n=0

λCn + 2−n ² ≤ 1 − λW + ² + 2²,

so λV ≥ λW − 3². As W and ² are arbitrary, λ is inner regular with respect to the closed sets, and is a compact measure. 342H Proposition (a) A compact measure space is locally compact. (b) A strictly localizable locally compact measure space is compact. (c) Let (X, Σ, µ) be a measure space. Suppose that whenever E ∈ Σ and µE > 0 there is an F ∈ Σ such that F ⊆ E, µF > 0 and the subspace measure on F is compact. Then µ is locally compact. proof (a) This is immediate from 342Ga and the definition of ‘locally compact’ measure space.

178

The lifting theorem

342H

(b) Suppose that (X, Σ, µ) is a strictly localizable locally compact measure space. Let hXi ii∈I be a decomposition of X, and for each i ∈ I let µi be the subspace measure on Xi . Then µi is compact. Now µ can be identified with the direct sum of the µi , so itself is compact, by 342Gd. (c) Write F for the set of measurable sets F ⊆ X such that the subspace measures µF are compact. Take E ∈ Σ with µE < ∞. S By 342Bb, there is a countable dist 0family hFi ii∈I in F such that Fi ⊆ E 0 for each i, and F = E \ i∈I Fi is negligible; now this means that F ∈ F (342Ac), so we may take it that S E = i∈I Fi . In this case µE is isomorphic to the direct sum of the measures µFi and is compact. As E is arbitrary, µ is locally compact. 342I Proposition (a) Any measurable subspace of a locally compact measure space is locally compact. (b) A measure space is locally compact iff its completion is locally compact iff its c.l.d. version is locally compact. (c) The direct sum of a family of locally compact measure spaces is locally compact. (d) The c.l.d. product of two locally compact measure spaces is locally compact. proof (a) Trivial: if (X, Σ, µ) is locally compact, and E ∈ Σ, and F ⊆ E is a measurable set of finite measure for the subspace measure on E, then F ∈ Σ and µF < ∞, so the subspace measure on F is compact. ˇ µ (b) Let (X, Σ, µ) be a measure space, and write (X, Σ, ˇ) for either its completion or its c.l.d. version. (i) Suppose that µ is locally compact, and that µ ˇF < ∞. Then there is an E ∈ Σ such that E ⊆ F and µE = µ ˇF . Let µE be the subspace measure on E induced by the measure µ; then we are assuming that µE is compact. Let K ⊆ PE be a compact class such that µE is inner regular with respect to K. Then, as in the proof of 342Gb, the subspace measure µ ˇF on F induced by µ ˇ is also inner regular with respect to K, so µ ˇF is compact; as F is arbitrary, µ ˇ is locally compact. (ii) Now suppose that µ ˇ is locally compact, and that µE < ∞. Then the subspace measure µ ˇE is compact. But this is just the completion of the subspace measure µE , so µE is compact, by 342Gc; as E is arbitrary, µ is locally compact. (c) Put (a) and 342Hc together. (d) Let (X, Σ, µ) and (Y, T, ν) be locally compact measure spaces, with product (X × Y, Λ, λ). If W ∈ Λ and λW > 0, there are E ∈ Σ, F ∈ T such that µE < ∞, νF < ∞ and λ(W ∩ (E × F )) > 0. Now the subspace measure λE×F induced by λ on E × F is just the product of the subspace measures (251P(ii-α), so is compact, and the subspace measure λW ∩(E×F ) is therefore again compact, by 342Ga. By 342Hc, this is enough to show that λ is locally compact. 342J Examples It is time I listed some examples of compact measure spaces. (a) Lebesgue measure on Rr is compact. (Let K be the family of subsets of Rr which are compact for the usual topology. By 134Fb, Lebesgue measure is inner regular with respect to K.) (b) Similarly, any Radon measure on Rr (256A) is compact. (c) If (A, µ ¯) is any semi-finite measure algebra, the standard measure ν on its Stone space Z is compact. (By 322Qa, ν is inner regular with respect to the family of open-and-closed subsets of Z, which are all compact for the standard topology of Z, so form a compact class.) (d) The usual measure on {0, 1}I is compact, for any set I. (It is obvious that the usual measure on {0, 1} is compact; now use 342Gf.) Remark (a)-(c) above are special cases of the fact that all Radon measures are compact; I will return to this in §416. 342K One of the most important properties of (locally) compact measure spaces has been studied under the following name. Definition Let (X, Σ, µ) be a measure space. Then (X, Σ, µ), or µ, is perfect if whenever f : X → R is measurable, E ∈ Σ and µE > 0, then there is a compact set K ⊆ f [E] such that µf −1 [K] > 0.

*342N


179

342L Theorem A semi-finite locally compact measure space is perfect. proof Let (X, Σ, µ) be a semi-finite locally compact measure space, f : X → R a measurable function, and E ∈ Σ a set of non-zero measure. Because µ is semi-finite, there is an F ∈ Σ such that F ⊆ E and 0 < µF < ∞. Now the subspace measure µF is compact; let T be a topology on F such that F is compact and µF is inner regular with respect to the family K of closed sets for P T. Let h²q iq∈Q be a family of strictly positive real numbers such that q∈Q ²q < 12 µF . (For instance, you could set ²q(n) = 2−n−3 µF where hq(n)in∈N is an enumeration of Q.) For each q ∈ Q, set Eq = {x : x ∈ F, f (x) ≤ q}, Eq0 = {x : x ∈ F, f (x) > q}, and T choose Kq , Kq0 ∈ K ∩ Σ such that Kq ⊆ Eq , Kq0 ⊆ Eq0 , 0 0 µ(Eq \ Kq ) ≤ ²q and µ(Eq \ Kq ) ≤ ²q . Then K = q∈Q (Kq ∪ Kq0 ) ∈ K ∩ Σ, K ⊆ F and P µ(F \ K) ≤ q∈Q µ(Eq \ Kq ) + µ(Eq0 \ Kq0 ) < µF , so µK > 0. The point is that f ¹K is continuous. P P For any q ∈ Q, {x : x ∈ K, f (x) ≤ q} = K ∩ Kq and {x : x ∈ K, f (x) > q} = K ∩ Kq0 . If H ⊆ R is open and x ∈ K ∩ f −1 [H], take q, q 0 ∈ Q such that f (x) ∈ ]q, q 0 ] ⊆ H; then G = K \ (Kq ∪ Kq0 0 ) is a relatively open subset of K containing x and included in f −1 [H]. Thus K ∩ f −1 [H] is relatively open in K; as H is arbitrary, f is continuous. Q Q Accordingly f [K] is a continuous image of a compact set, therefore compact; it is a subset of f [E], and µf −1 [f [K]] ≥ µK > 0. As f and E are arbitrary, µ is perfect. 342M I ought to give examples to distinguish between the concepts introduced here, partly on general principles, but also because it is not obvious that the concept of ‘locally compact’ measure space is worth spending time on at all. It is easy to distinguish between ‘perfect’ and ‘(locally) compact’; ‘locally compact’ and ‘compact’ are harder to separate. Example Let X be an uncountable set and µ the countable-cocountable measure on X (211R). Then µ is perfect but not compact or locally compact. proof (a) If f : X → R is measurable and E ⊆ X is measurable, with measure greater than 0, set A = {α S : α ∈ R, {x : x ∈ X, f (x) ≤ r} is negligible}. Then α ∈ A whenever α ≤ β ∈ A. Since / A, in which case A is bounded above by X = n∈N {x : f (x) ≤ n}, there is some n such that n ∈ n. Also there is some m ∈ N such that {x : f (x) > −m} is non-negligible, in which case it must be conegligible, and −m ∈ A, so A is non-empty. Accordingly γ = sup A is defined in R. Now for any k ∈ N, {x : f (x) ≤ γ − 2−k } is negligible, so {x : f (x) < γ} is negligible. Also, for any k, {x : f (x) ≤ γ + 2−k } is non-negligible, so {x : f (x) > γ2−k } must be negligible; accordingly, {x : f (x) > γ} is negligible. But this means that {x : f (x) = γ} is conegligible and has measure 1. Thus we have a compact set K = {γ} such that µf −1 [K] = 1, and γ must belong to f [E]. As f and E are arbitrary, µ is perfect. (b) µ is not compact. P P?? Suppose, if possible, that K ⊆ PX is a compact class such that µ is inner regular with respect to K. Then for every x ∈ X there is a measurable set Kx ∈ K such that Kx ⊆ X \ {x} and µKx > 0, that is, Kx is conegligible. But this means that {Kx : x ∈ X} must have the finite intersection property; as it also has empty intersection, K cannot be a compact class. X XQ Q (c) Because µ is totally finite, it cannot be locally compact (342Hb). *342N Example There is a complete locally determined localizable locally compact measure space which is not compact. proof (a) I refer to the example of 216E. In that construction, we have a set I and a family hfγ iγ∈C in X = {0, 1}I such that for every D ⊆ C there is an i ∈ I such that D = {γ : fγ (i) = 1}; moreover, #(C) > c. The σ-algebra Σ is the family of sets E ⊆ X such that for every γ there is a countable set J ⊆ I such that {x : x¹J = fγ ¹J} is a subset of either E or X \ E; and for E ∈ Σ, µE is #({γ : fγ ∈ E}) if this is finite, ∞ otherwise. Note that any subset of X determined by a countable set of coordinates belongs to Σ. For each γ ∈ C, let iγ ∈ I be such that fγ (iγ ) = 1, fδ (iγ ) = 0 for δ 6= γ. (In 216E I took I to be PC, and iγ would be {γ}.) Set Y = {x : x ∈ X, {γ : γ ∈ C, x(iγ ) = 1} is finite}.

180

The lifting theorem

*342N

Give Y its subspace measure µY with domain ΣY . Then µY is complete, locally determined and localizable (214Id). Note that fγ ∈ Y for every γ ∈ C. (b) µY is locally compact. P P Suppose that F ∈ ΣY and µY F < ∞. If µY F = 0 then surely the subspace measure µF is compact. Otherwise, we can express F as E ∩ Y where E ∈ Σ and µE = µY F . Then D = {γ : fγ ∈ E} = {γ : fγ ∈ F } is finite. For γ ∈ D set G0γ = {x : x ∈ X, x(iγ ) = 1, x(iδ ) = 0 for every δ ∈ D \ {γ}} ∈ Σ, Kγ = {K : fγ ∈ K ⊆ F ∩ G0γ }.

S Then each Kγ is a compact class, and of different Kγ ’s are dist, so K = γ∈D Kγ is a compact class. Now suppose that H belongs to the subpsace σ-algebra ΣF and µF H > 0. Then there is a γ ∈ D such that fγ ∈ H, so that H ∩ G0γ ∈ K ∩ ΣF and µF (H ∩ G0γ ) > 0. By 342E, this is enough to show that µF is compact. As F is arbitrary, µY is locally compact. Q Q (c) µY is not compact. P P?? Suppose, if possible, that µY is inner regular with respect to a compact class K ⊆ PY . For each γ ∈ C set Gγ = {x : x ∈ X, x(iγ ) = 1}, so that fγ ∈ Gγ ∈ Σ and µY (Gγ ∩ Y ) = 1. There must therefore be a Kγ ∈ K such that Kγ ⊆ Gγ ∩ Y and µY Kγ = 1 (since µY takes no value in ]0, 1[). Express Kγ as Y ∩ Eγ , where Eγ ∈ Σ, and let Jγ ⊆ I be a countable set such that Eγ ⊇ {x : x ∈ X, x¹Jγ = fγ ¹Jγ }. At this point I call on the full strength of 2A1P. There is a set B ⊆ C, of cardinal greater than c, such that fγ ¹Jγ ∩ Jδ = fδ ¹Jγ ∩ Jδ for all γ, δ ∈ B. But this means that, for any finite set D ⊆ B, we can define x ∈ X by setting x(i) = fα (i) if α ∈ D, i ∈ Jα , [ = 0 if i ∈ I \ Jα . α∈D

It is easy to check that {γ : γ ∈ C, x(iγ ) = 1} = D, so that x ∈ Y ; but now T T x ∈ Y ∩ α∈D Eα = α∈D Kα . What this shows is that {Kα : α ∈ B} has the finite intersection property. It must therefore have non-empty intersection; say T T y ∈ α∈B Kα ⊆ α∈B Gα . But now we have a member y of Y such that {γ : y(iγ ) = 1} ⊇ B is infinite, contrary to the definition of Y . X XQ Q 342X Basic exercises > (a) Show that a measure space (X, Σ, µ) is semi-finite iff µ is inner regular with respect to {E : µE < ∞}. (b) Find a proof of 342B based on 215A. (c) Let (X, Σ, µ) be a locally compact semi-finite measure space in which all singleton sets are negligible. Show that it is atomless. (d) Let (X, Σ, µ) be a measure space, and ν an indefinite-integral measure over µ (234B). Show that ν is compact, or locally compact, if µ is. (Hint: if K satisfies the conditions of 342E with respect to µ, then it satisfies them for ν.) (e) Let f : R → R be any non-decreasing function, and νf the corresponding Lebesgue-Stieltjes measure. Show that νf is compact. (Hint: 256Xg.) (f ) Let µ be Lebesgue measure on [0, 1], ν the countable-cocountable measure on [0, 1], and λ their c.l.d. product. Show that λ is a compact measure. (Hint: let K be the family of sets K × A where A ⊆ [0, 1] is cocountable and K ⊆ A is compact.)

342 Notes


181

(g) (i) Give an example of a compact probability space (X, Σ, µ), a set Y and a function f : X → Y such that the image measure µf −1 is not compact. (ii) Give an example of a compact probability space (X, Σ, µ) and a σ-subalgebra T of Σ such that (X, T, µ¹ T) is not compact. (Hint: 342Xf.) (h) Let (X, Σ, µ) be a perfect measure space, and f : X → R a measurable function. Show that the image measure µf −1 is inner regular with respect to the compact subsets of R, so is a compact measure. (i) Let (X, Σ, µ) be a σ-finite measure space. Show that it is perfect iff for every measurable f : X → R there is a Borel set H ⊆ f [X] such that f −1 [H] is conegligible in X. (Hint: 342Xh for ‘only if’, 256C for ‘if’.) (j) Let (X, Σ, µ) be a complete totally finite perfect measure space and f : X → R a measurable function. Show that the image measure µf −1 is a Radon measure, and is the only Radon measure on R for which f is inverse-measure-preserving. (Hint: 256G.) (k) Suppose that (X, Σ, µ) is a perfect measure space. (i) Show that if (Y, T, ν) is a measure space, and f : X → Y is a function such that f −1 [F ] ∈ Σ for every F ∈ T and f −1 [F ] is µ-negligible for every ν-negligible set F , then (Y, T, ν) is perfect. (ii) Show that if T is a σ-subalgebra of Σ then (X, T, µ¹ T) is perfect. (l) Let (X, Σ, µ) be a perfect measure space such that Σ is the σ-algebra generated P∞ by a sequence of sets. Show that µ is compact. (Hint: if Σ is generated by {En : n ∈ N}, set f = n=0 3−n χEn and consider {f −1 [K] : K ⊆ f [X] is compact}.) (m) Let (X, Σ, µ) be a semi-finite measure space. Show that µ is perfect iff µ¹ T is compact for every countably generated σ-subalgebra T of Σ. (n) Show that (i) a measurable subspace of a perfect measure space is perfect (ii) a semi-finite measure space is perfect iff all its totally finite subspaces are perfect (iii) the direct sum of any family of perfect measure spaces is perfect (iv) the c.l.d. product of two perfect measure spaces is perfect (hint: put 342Xm and 342Ge together) (v) the product of any family of perfect probability spaces is perfect (vi) a measure space is perfect iff its completion is perfect (vii) the c.l.d. version of a perfect measure space is perfect (viii) any purely atomic measure space is perfect (ix) an indefinite-integral measure over a perfect measure is perfect. (o) Let µ be Lebesgue measure on R, let A be a subset of R, and let µA be the subspace measure. Show that µA is compact iff it is perfect iff A is Lebesgue measurable. (Hint: if µA is perfect, consider the image measure hµ−1 A on R, where h(x) = x for x ∈ A.) 342Y Further exercises (a) Show that the space (X, Σ, µ) of 216E and 342N is a compact measure space. (Hint: use the usual topology on X = {0, 1}I .) (b) Give an example of a compact complete locally determined measure space which is not localizable. (Hint: in 216D, add a point to each horizontal and vertical section of X, so that all the sections become compact measure spaces.) 342 Notes and comments The terminology I find myself using in this section – ‘compact’, ‘locally compact’, ‘countably compact’, ‘perfect’ – is not entirely satisfactory, in that it risks collision with the same words applied to topological spaces. For the moment, this is not a serious problem; but when in Volume 4 we come to the systematic analysis of spaces which have both topologies and measures present, it will be necessary to watch our language carefully. Of course there are cases in which a ‘compact class’ of the sort discussed here can be taken to be the family of compact sets for some familiar topology, as in 342Ja-342Jd, but in others this is not so (see 342Xf); and even when we have a familiar compact class, the topology constructed from it by the method of 342Da need not be one we might expect. (Consider, for instance, the topology on R for which the closed sets are just the sets which are compact for the usual topology.)

182

The lifting theorem

342 Notes

I suppose that ‘compact’ and ‘perfect’ measure spaces look reasonably natural objects to study; they offer to illuminate one of the basic properties of Radon measures, the fact that (at least for totally finite Radon measures on Euclidean space) the image measure of a Radon measure under a measurable function is again Radon (256G, 342Xj). Indeed this was the original impetus for the study of perfect measures (Gnedenko & Kolmogorov 54, Sazonov 66). It is not obvious that there is any need to examine ‘locally compact’ measure spaces, but actually they are the chief purpose of this section, since the main theorem of the next section is an alternative characterization of semi-finite locally compact measure spaces (343B). Of course you may feel that the fact that ‘locally compact’ and ‘compact’ coincide for strictly localizable spaces (342Hb) excuses you from troubling about the distinction at first reading. As with any new classification of measure spaces, it is worth finding out how the classes of ‘compact’ and ‘perfect’ measure spaces behave with respect to the standard constructions. I run through the basic facts in 342G-342I, 342Xd, 342Xk and 342Xn. We can also look for relationships between the new properties and those already studied. Here, in fact, there is not much to be said; 342N and 342Yb show that ‘compactness’ is largely independent of the classification in §211. However there are interactions with the concept of ‘atom’ (342Xc, 342Xn(viii)). I give examples to show that perfect measure spaces need not be locally compact, and that locally compact measure spaces need not be compact (342M, 342N). The standard examples of measure spaces which are not perfect are non-measurable subspaces (342Xo); I will return to these in the next section (343L-343M). Something which is not important to us at the moment, but is perhaps worth taking note of, is the following observation. To determine whether a measure space (X, Σ, µ) is compact, we need only the structure (X, Σ, N ), where N is the σ-ideal of negligible sets, since that is all that is referred to in the criterion of 342E. The same is true of local compactness, by 342Hc, and of perfectness, by the definition in 342K. Compare 342Xd, 342Xk, 342Xn(ix). Much of the material of this section will be repeated in Volume 4 as part of a more systematic analysis of inner regularity.

343 Realization of homomorphisms We are now in a position to make progress in one of the basic questions of abstract measure theory. In §324 I have already described the way in which a function between two measure spaces can give rise to a homomorphism between their measure algebras. In this section I discuss some conditions under which we can be sure that a homomorphism can be represented by a function. The principal theorem of the section is 343B. If a measure space (X, Σ, µ) is locally compact, then many homomorphisms from the measure algebra of µ to other measure algebras will be representable by functions into X; moreover, this characterizes locally compact spaces. In general, a homomorphism between measure algebras can be represented by widely different functions (343I, 343J). But in some of the most important cases (e.g., Lebesgue measure) representing functions are ‘almost’ uniquely defined; I introduce the concept of ‘countably separated’ measure space to describe these (343D-343H). 343A Preliminary remarks It will be helpful to establish some vocabulary and a couple of elementary facts. (a) If (X, Σ, µ) and (Y, T, ν) are measure spaces, with measure algebras A and B, I will say that a function f : X → Y represents a homomorphism π : B → A if f −1 [F ] ∈ Σ and (f −1 [F ])• = π(F • ) for every F ∈ T. (Perhaps I should emphasize here that some homomorphisms are representable in this sense, and some are not; see 343M below for examples of non-representable homomorphisms.) (b) If (X, Σ, µ) and (Y, T, ν) are measure spaces, with measure algebras A and B, f : X → Y is a function, and π : B → A is a sequentially order-continuous Boolean homomorphism, then {F : F ∈ T, f −1 [F ] ∈ Σ and f −1 [F ]• = πF • } is a σ-subalgebra of T. (The verification is elementary.)

343B

Realization of homomorphisms

183

(c) Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras A and B, and π : B → A a ˆ µ ˆ νˆ) be the Boolean homomorphism which is represented by a function f : X → Y . Let (X, Σ, ˆ), (Y, T, completions of (X, Σ, µ), (Y, T, ν); then A and B can be identified with the measure algebras of µ ˆ and ˆ µ ˆ νˆ). P νˆ (322Da). Now f still represents π when regarded as a function from (X, Σ, ˆ) to (Y, T, P If G is ν-negligible, there is a negligible F ∈ T such that G ⊆ F ; since f −1 [F ]• = πF • = 0, ˆ If G is any element of T, ˆ there is an f −1 [F ] is µ-negligible, so f −1 [E] is negligible, therefore belongs to Σ. F ∈ T such that G4F is negligible, so that ˆ f −1 [G] = f −1 [F ]4f −1 [G4F ] ∈ Σ, and f −1 [G]• = f −1 [F ]• = πF • = πG• . Q Q (d) In particular, if (X, Σ, µ) and (Y, T, ν) are measure spaces, and f : X → Y is inverse-measurepreserving, then f is still inverse-measure-preserving with respect to the completed measures µ ˆ and νˆ (apply (c) with π : B → A the homomorphism induced by f ). (See 235Hc.) 343B Theorem Let (X, Σ, µ) be a non-empty semi-finite measure space, and (A, µ ¯) its measure algebra. Let (Z, Λ, λ) be the Stone space of (A, µ ¯); for E ∈ Σ write E ∗ for the open-and-closed subset of Z corresponding to the image E • of E in A. Then the following are equiveridical: (i) (X, Σ, µ) is locally compact in the sense of 342Ad. (ii) There is a family K ⊆ Σ such that (α) whenever TE ∈ Σ and µE > 0 there is a K ∈ K such that 0 0 K⊆E T and µK > 0 (β) whenever K ⊆ K is such that µ( K0 ) > 0 for every non-empty finite set K0 ⊆ K , then K0 6= ∅. (iii) ThereTis a family K ⊆ Σ such that (α)0 µ is inner regular withTrespect to K (β) whenever K0 ⊆ K is such that µ( K0 ) > 0 for every non-empty finite set K0 ⊆ K0 , then K0 6= ∅. (iv) There is a function f : Z → X such that f −1 [E]4E ∗ is negligible for every E ∈ Σ. (v) Whenever (Y, T, ν) is a complete strictly localizable measure space, with measure algebra B, and π : A → B is an order-continuous Boolean homomorphism, then there is a g : Y → X representing π. (vi) Whenever (Y, T, ν) is a complete strictly localizable measure space, with measure algebra B, and π : A → B is an order-continuous measure-preserving Boolean homomorphism, then there is a g : Y → X representing π. proof (a)(i)⇒(ii) Because µ is semi-finite, there is a partition of unity hai ii∈I in A such that µ ¯ai < ∞ for each i. For each i ∈ I, let Ei ∈ Σ be such that Ei• = ai . Then the subspace measure µEi on Ei is S compact; let Ki ⊆ PEi be a compact class such that µ is inner regular with respect to K . Set K = Ei i i∈I Ki . If T 0 K0 ⊆ K and µ( K0 ) > 0 for every non-empty finite K ⊆ K, then K ⊆ K for some i, and surely has the 0 i T finite intersection property, so K0 6= ∅; thus K0 satisfies (β) of condition (ii). And if E ∈ Σ, µE > 0 then there must be some i ∈ I such that Ei• ∩ ai 6= 0, that is, µ(E ∩ Ei ) > 0, in which case there is a K ∈ Ki ⊆ K such that K ⊆ E ∩ Ei and µK > 0; so that K satisfies condition (α). (b)(ii)⇒(iii) Suppose that K ⊆ Σ witnesses that (ii) is true. If µX = 0 then K already witnesses that (iii) is true, so we need consider only the case µX > 0. Set L = {K0 ∪ . . . ∪ Kn : K0 , . . . , Kn ∈ K}. Then LT witnesses that (iii) is true. P P By 342Ba, µ is inner regular with respect to L. Let L0 ⊆ L be such that µ( L0 ) > 0 for every non-empty finite L0 ⊆ L0 . Then T F0 = {A : A ⊆ X, there is a finite L0 ⊆ L0 such that X ∩ L0 \ A is negligible} is a filter on X, so there is an ultrafilter F on X including F0 . T Note that every conegligible set belongs to F0 , so no negligible set can belong to F. Set K0 = K ∩ F; then T K0 belongs to F, so is not negligible, for every non-empty finite K0 ⊆ K0 . Accordingly there is some x ∈ K0 . But any member of L0 is of the form L = K0 ∪ . . . ∪ Kn where each Ki ∈ K; because F is an T ultrafilter and L ∈ F, there must be some i ≤ n such that Ki ∈ F, in which case x ∈ Ki ⊆ L. Thus x ∈ L0 . As L0 is arbitrary, L satisfies the condition (β). Q Q

184

The lifting theorem

343B

(c)(iii)⇒(iv) Let K ⊆ Σ witness that (iii) ForTany z ∈ Z, set Kz = T {K : K ∈ K, z ∈ K ∗ }. If T T is true. ∗ ∗ ∗ K0 , . . . , Kn ∈ Kz , then z ∈ i≤n Ki = ( i≤n Ki ) , so ( i≤n Ki ) 6= ∅ and µ( i≤n Ki ) > 0. By (β) of T T condition (iii), Kz 6= ∅; and even if Kz = ∅, X ∩ Kz 6= ∅ because X is non-empty. So we may choose T f (z) ∈ X ∩ Kz . This defines a function f : Z → X. Observe that, for K ∈ K and z ∈ Z, z ∈ K ∗ =⇒ K ∈ Kz =⇒ f (z) ∈ K =⇒ z ∈ f −1 [K], so that K ∗ ⊆ f −1 [K]. Now take any E ∈ Σ. Consider S S U1 = {K ∗ : K ∈ K, K ⊆ E} ⊆ {E ∗ ∩ f −1 [K] : K ∈ K, K ⊆ E} ⊆ E ∗ ∩ f −1 [E], S U2 = {K ∗ : K ∈ K, K ⊆ X \ E} ⊆ (X \ E)∗ ∩ f −1 [X \ E] = Z \ (f −1 [E] ∪ E ∗ ), so that f −1 [E]4E ∗ ⊆ Z \ (U1 ∪ U2 ). Now U1 and U2 are open subsets of Z, so M = Z \ (U1 ∪ U2 ) is closed, and in fact M is nowhere dense. P P?? Otherwise, there is a non-zero a ∈ A such that the corresponding open-and-closed set b a is included in M , and an F ∈ Σ of non-zero measure such that a = F • . At least one of F ∩ E, F \ E is non-negligible and therefore includes a non-negligible member K of K. But in this case K ∗ is a non-empty open subset of M which is included in either U1 or U2 , which is impossible. X XQ Q By the definition of λ (321J-321K), M is λ-negligible, so f −1 [E]4E ∗ ⊆ M is negligible, as required. (d)(iv)⇒(v) Now assume that f : Z → X witnesses (iv), and let (Y, T, ν) be a complete strictly localizable measure space, with measure algebra B, and π : A → B an order-continuous Boolean homomorphism. If νY = 0 then any function from Y to X will represent π, so we may suppose that νY > 0. Write W for the Stone space of B. Then we have a continuous function φ : W → Z such that φ−1 [b a] = π ca for every a ∈ A (312P), and φ−1 [M ] is nowhere dense in W for every nowhere dense M ⊆ Z (313R). It follows that φ−1 [M ] is meager for every meager M ⊆ Z, that is, φ−1 [M ] is negligible in W for every negligible M ⊆ Z. By 341Q, there is an inverse-measure-preserving function h : Y → W such that h−1 [bb]• = b for every b ∈ B. Consider g = f φh : Y → X. If E ∈ Σ, set a = E • ∈ A, so that E ∗ = b a ⊆ Z, and M = f −1 [E]4E ∗ is λ-negligible; consequently −1 φ [M ] is negligible in W . Because h is inverse-measure-preserving, g −1 [E]4h−1 [φ−1 [E ∗ ]] = h−1 [φ−1 [f −1 [E]]]4h−1 [φ−1 [E ∗ ]] = h−1 [φ−1 [M ]] is negligible. But φ−1 [E ∗ ] = π ca, so g −1 [E]• = h−1 [φ−1 [E ∗ ]]• = πa. As E is arbitrary, g induces the homomorphism π. (e)(v)⇒(vi) is trivial. (f )(vi)⇒(iv) Assume (vi). Let ν be the c.l.d. version of λ, T its domain, and B its measure algebra; then ν is strictly localizable (322Qb). The embedding Λ ⊆ T corresponds to an order-continuous measurepreserving Boolean homomorphism from A to B (322Db). By (vi), there is a function f : Z → X such that f −1 [E] ∈ T and f −1 [E]• = (E ∗ )• in B for every E ∈ Σ. But as ν and λ have the same negligible sets (322Qb), f −1 [E]4E ∗ is λ-negligible for every E ∈ Σ, as required by (iv). α) To begin with (down to the end of (γ) below) I suppose that µ is totally finite. In this (g)(iv)⇒(i)(α case we have a function g : X → Z such that E4g −1 [E ∗ ] is negligible for every E ∈ Σ (341Q again). We are supposing also that there is a function f : Z → X such that f −1 [E]4E ∗ is negligible for every E ∈ Σ. Write K for the family of sets K ⊆ E such that K ∈ Σ and there is a compact set L ⊆ Z such that f [L] ⊆ K ⊆ g −1 [L]. β ) µ is inner regular with respect to K. P (β P Take F ∈ Σ and γ < µF . Choose hVn in∈N , hFn in∈N as follows. F0 = F . Given that µFn > γ, then λ(f −1 [Fn ] ∩ Fn∗ ) = λFn∗ = µFn > γ, ∗ so there is an open-and-closed set Vn ⊆ f −1 [Fn ] ∩ Fn∗ with λVn > γ. Express Vn as Fn+1 where Fn+1 ∈ Σ; −1 ∗ ∗ −1 ∗ since Fn 4g [Fn ] is negligible, and Vn ⊆ T Fn , we may take it that T Fn+1 ⊆ g [Fn ]. Continue. At the end of the induction, set K = n∈N Fn ∈ Σ and L = n∈N Fn∗ . Because Fn+1 \ Fn ⊆ g −1 [Fn∗ ] \ Fn is negligible for each n, µK = limn→∞ µFn ≥ γ, while K ⊆ F and L is surely compact. We have

343D


L⊆ so f [L] ⊆ K. Also K⊆

T n∈N

T n∈N

Vn ⊆

T

Fn+1 ⊆

n∈N

185

f −1 [Fn ] = f −1 [K],

T n∈N

g −1 [Fn∗ ] = g −1 [L].

So K ∈ K. As F and γ are arbitrary, µ is inner regular with respect to K. Q Q (γγ ) Next, P Suppose that K0 ⊆ K has the finite intersection property. If K0 = ∅, T 0 K is a compact class. 0P of course K 6= ∅; suppose that K is non-empty. Let L be the family of closed sets L ⊆ Z such that 0 g −1 [L] includes some property, and Z is compact, so T member of K . Then L has the finite intersection there is some z ∈ L; also Z ∈ L, so z ∈ Z. For any K ∈ K0 , there is some T closed set L ⊆ Z such that f [L] ⊆ K ⊆ g −1 [L], so that L ∈ L and z ∈ L and f (z) ∈ K. Thus f (z) ∈ K0 . As K0 is arbitrary, K is a compact class. Q Q So K witnesses that µ is a compact measure. (δδ ) Now consider the general case. Take any E ∈ Σ of finite measure. If E = ∅ then surely the subspace measure µE is compact. Otherwise, we can identify the measure algebra of µE with the principal ideal AE • of A generated by E • (322Ja), and E ∗ ⊆ Z with the Stone space of AE • (312S). Take any x0 ∈ E and define f˜ : E ∗ → E by setting f˜(z) = f (z) if z ∈ E ∗ ∩ f −1 [E], x0 if z ∈ E ∗ \ f −1 [E]. Then f and f˜ agree almost everywhere on E ∗ , so f˜−1 [F ]4F ∗ is negligible for every F ∈ ΣE , that is, f˜ represents the canonical isomorphism between the measure algebras of µE and the subspace measure λE ∗ on E ∗ . But this means that condition (iv) is true of µE , so µE is compact, by (α)-(γ) above. As E is arbitrary, µ is locally compact. This completes the proof. 343C Examples (a) Let κ be an infinite cardinal. We know that the usual measure νκ on {0, 1}κ is compact (342Jd). It follows that if (X, Σ, µ) is any complete probability space such that the measure algebra Bκ of νκ can be embedded as a subalgebra of the measure algebra A of µ, there is an inversemeasure-preserving function from X to {0, 1}κ . By 332P, this is so iff every non-zero principal ideal of A has Maharam type at least κ. Of course this does not depend in any way on the results of the present chapter. If Bκ can be embedded in A, there must be a stochastically independent family hEξ iξ<κ of sets of measure 1 κ 2 ; now we get a map h : X → {0, 1} by saying that h(x)(ξ) = 1 iff x ∈ Eξ , which by 254G is inversemeasure-preserving. (b) In particular, if µ is atomless, there is an inverse-measure-preserving function from X to {0, 1}N ; since this is isomorphic, as measure space, to [0, 1] with Lebesgue measure (254K), there is an inverse-measurepreserving function from X to [0, 1]. (c) More generally, if (X, Σ, µ) is any complete atomless totally finite measure space, there is an inversemeasure-preserving function from X to the interval [0, µX] endowed with Lebesgue measure. (If µX > 0, apply (b) to the normalized measure (µX)−1 µ; or argue directly from 343B, using the fact that Lebesgue measure on [0, µX] is compact; or use the idea suggested in 343Xd.) (d) Throughout the work above – in §254 as well as in 343B – I have taken the measures involved to be complete. It does occasionally happen, in this context, that this restriction is inconvenient. Typical results not depending on completeness in the domain space X are in 343Xc-343Xd. Of course these depend not only on the very special nature of the codomain spaces {0, 1}I or [0, 1], but also on the measures on these spaces being taken to be incomplete. 343D Uniqueness of realizations The results of 342E-342J, together with 343B, give a respectable number of contexts in which homomorphisms between measure algebras can be represented by functions between measure spaces. They say nothing about whether such functions are unique, or whether we can distinguish, among the possible representations of a homomorphism, any canonical one. In fact the proof of 343B, using the Lifting Theorem as it does, strongly suggests that this is like looking for a canonical lifting, and I am sure that (outside a handful of very special cases) any such search is vain. Nevertheless, we do have a weak kind of uniqueness theorem, valid in a useful number of spaces, as follows.

186

The lifting theorem

343D

Definition A measure space (X, Σ, µ) is countably separated if there is a countable set A ⊆ Σ separating the points of X in the sense that for any distinct x, y ∈ X there is an E ∈ A containing one but not the other. (Of course this is a property of the structure (X, Σ) rather than of (X, Σ, µ).) 343E Lemma A measure space (X, Σ, µ) is countably separated iff there is an injective measurable function from X to R. proof If (X, Σ, µ) is countably separated, let A ⊆ Σ be a countable set separating the points of X. Let hEn in∈N be a sequence running over A ∪ {∅}. Set P∞ f = n=0 3−n χEn : X → R. Then f is measurable (because every En is measurable) and injective (because if x 6= y in X and n = min{i : #(Ei ∩ {x, y}) = 1} and x ∈ En , then P P P f (x) ≥ 3−n + i i>n 3−i + i
343K


187

any particular ξ ∈ c \ ω, {x : h(x¹ω) = ξ} is negligible, being either empty or of the form {x : x(n) = z(n) for every n < ω} for some z ∈ {0, 1}ω . So H = {x : h(x¹ω) ∈ J} is negligible. Now we see that for x ∈ X \ H, f (x)¹J = x¹J, so for x ∈ X \ (H ∪ (E 00 \ E 0 )), x ∈ E =⇒ x ∈ E 0 =⇒ f (x) ∈ E 0 =⇒ f (x) ∈ E, x∈ / E =⇒ x ∈ / E 00 =⇒ f (x) ∈ / E 00 =⇒ f (x) ∈ / E. Thus E4f −1 [E] ⊆ H ∪ (E 00 \ E 0 ) is negligible. Q Q 343J The split interval I introduce a construction which here will seem essentially elementary, but in other contexts is of great interest, as will appear in Volume 4. (a) Take I k to consist of two copies of each point of the unit interval, so that I k = {t+ : t ∈ [0, 1]} ∪ {t− : t ∈ [0, 1]}. For A ⊆ I k write Al = {t : t− ∈ A}, Ar = {t : t+ ∈ A}. Let Σ be the set {E : E ⊆ I k , El and Er are Lebesgue measurable and El 4Er is Lebesgue negligible}. For E ∈ Σ, set µE = µL El = µL Er where µL is Lebesgue measure on [0, 1]. It is easy to check that (I k , Σ, µ) is a complete probability space. Also it is compact. P P Take K to be the family of sets K ⊆ I k such that Kl = Kr is a compact subset of [0, 1], and check that K is a compact class and that µ is inner regular with respect to K; or use 343Xa below. Q Q The sets {t− : t ∈ [0, 1]} and {t+ : t ∈ [0, 1]} are non-measurable subsets of I k ; on both of them the subspace measures correspond exactly to µL . We have a canonical inverse-measure-preserving function h : I k → [0, 1] given by setting h(t+ ) = h(t− ) = t for every t ∈ [0, 1]; h induces an isomorphism between the measure algebras of µ and µL . I k is called the split interval or (especially when given its standard topology, as in 343Yc below) the double arrow space or two arrows space. Now the relevance to the present discussion is this: we have a map f : I k → I k given by setting f (t+ ) = t− , f (t− ) = t+ for every t ∈ [0, 1] such that f (x) 6= x for every x, but E4f −1 [E] is negligible for every E ∈ Σ, so that f represents the identity homomorphism on the measure algebra of µ. The function h : I k → [0, 1] is canonical enough, but is two-to-one, and the canonical map from the measure algebra of µ to the measure algebra of µL is represented equally by the functions t 7→ t− and t 7→ t+ , which are nowhere equal. (b) Consider the direct sum (Y, ν) of (I k , µ) and ([0, 1], µL ); for definiteness, take Y to be (I k × {0}) ∪ ([0, 1] × {1}). Setting h1 (t+ , 0) = h1 (t− , 0) = (t, 1), h1 (t, 1) = (t+ , 0), we see that h1 : Y → Y induces a measure-preserving involution of the measure algebra B of ν, corresponding to its expression as a simple product of the isomorphic measure algebras of µ and µL . But h1 is not invertible, and indeed there is no invertible function from Y to itself which induces this involution of B. P P?? Suppose, if possible, that g : Y → Y were such a function. Looking at the sets Eq = [0, q] × {1},

Fq = {(t+ , 0) : t ∈ [0, q]} ∪ {(t− , 0) : t ∈ [0, q]}

for q ∈ Q, we must have g −1 [Eq ]4Fq negligible for every q, so that we must have g(t+ , 0) = g(t− , 0) = (t, 1) for almost every t ∈ [0, 1], and g cannot be injective. X XQ Q (c) Thus even with a compact probability space, and an automorphism φ of its measure algebra, we cannot be sure of representing φ and φ−1 by functions which will be inverses of each other. 343K

342L has a partial converse.

Proposition If (X, Σ, µ) is a semi-finite countably separated measure space, it is compact iff it is locally compact iff it is perfect.

188

The lifting theorem

343K

proof We already know that compact measure spaces are locally compact and locally compact semi-finite measure spaces are perfect (342Ha, 342L). So suppose that (X, Σ, µ) is a perfect semi-finite countably separated measure space. Let f : X → R be an injective measurable function (343E). Consider K = {f −1 [L] : L ⊆ f [X], L is compact in R}. The definition of ‘perfect’ measure space states exactly that whenever E ∈ Σ and µE > 0 there is a K ∈ K such that K ⊆ E and µK > 0. And K is a compact class. P P If K0 ⊆ K has the finite intersection property, L = {f [K] : K ∈ K0 } is aTfamily of compact sets in R with the finite intersection property, and has non-empty intersection; so that K0 is also non-empty, because f is injective. Q Q By 342E, (X, Σ, µ) is compact. 343L The time has come to give examples of spaces which are not locally compact, so that we can expect to have measure-preserving homomorphisms not representable by inverse-measure-preserving functions. The most commonly arising ones are covered by the following result. Proposition Let (X, Σ, µ) be a complete locally determined countably separated measure space, and A ⊆ X a set such that the subspace measure µA is perfect. Then A is measurable. proof ?? Otherwise, there is a set E ∈ Σ such that µE < ∞ and B = A ∩ E ∈ / Σ. Let f : X → R be an injective measurable function (343E again). Then f ¹B is ΣB -measurable, where ΣB is the domain of the subspace measure µB on B. Set K = {f −1 [L] : L ⊆ f [B], L is compact in R}. Just as in the proof of 343K, K is a compact class and S µB is inner regular with respect to K. By 342Bb, there is a sequence hK i in K such that µ (B \ B n∈N K Sn ) = 0. But of course K ⊆ Σ, because f is S n n∈N X Σ-measurable, so n∈N Kn ∈ Σ. Because µ is complete, B \ n∈N Kn ∈ Σ and B ∈ Σ. X 343M Example 343L tells us that any non-measurable set X of Rr , or of {0, 1}N , with their usual measures, is not perfect, therefore not (locally) compact, when given its subspace measure. To find a non-representable homomorphism, we do not need to go through the whole apparatus of 343B. Take Y to be a measurable envelope of X (132Ed). Then the identity function from X to Y induces an isomorphism of their measure algebras. But there is no function from Y to X inducing the same isomorphism. P P?? Writing Z for Rr or {0, 1}N and µ for its measure, Z is countably separated; suppose hEn in∈N is a sequence of measurable sets in Z separating its points. For each n, (Y ∩ En )• in the measure algebra of µY corresponds to (X ∩ En )• in the measure algebra of µX . So if f : Y → X were a function representing the isomorphism of the measure algebras, (Y ∩ En )4f −1 [En ] would have to be negligible for each n, and S A = n∈N (Y ∩ En )4f −1 [En ] would be negligible. But for y ∈ Y \ A, f (y) belongs to just the same En as y does, so must be equal to y. Accordingly X ⊇ Y \ A and X is measurable. X XQ Q 343X Basic exercises (a) Let (X, Σ, µ) be a semi-finite measure space. (i) Suppose that there is a set A ⊆ X, of full outer measure, such that the subspace measure on A is compact. Show that µ is locally compact. (Hint: show that µ satisfies (ii) or (v) of 343B.) (ii) Suppose that for every non-negligible E ∈ Σ there is a non-negligible set A ⊆ E such that the subspace measure on A is compact. Show that µ is locally compact. (b) Let hXi ii∈I be a family of non-empty sets, with product X; write πi : X → Xi for the coordinate N map. Suppose we are given a σ-algebra Σi of subsets of Xi for each i; let Σ = c i∈I Σi be the corresponding σ-algebra of subsets of X generated by {πi−1 [E] : i ∈ I, E ∈ Σi }. Let µ be a totally finite measure with domain Σ, and for i ∈ I let µi be the image measure µπi−1 . Check that the domain of µi is Σi . Show that if every (Xi , Σi , µi ) is compact, then so is (X, Σ, µ). (Hint: either show that µ satisfies (v) of 343B or adapt the method of 342Gf.) (c) Let I be any set. Let T be the σ-algebra of subsets of {0, 1}I generated by the sets Fi = {z : z(i) = 1} for i ∈ I, and ν any probability measure with domain T; let B be the measure algebra of ν. Let (X, Σ, µ) be a measure space with measure algebra A, and φ : B → A an order-continuous Boolean homomorphism. Show that there is an inverse-measure-preserving function f : X → {0, 1}I representing φ. (Hint: for each i ∈ I, take Ei ∈ Σ such that Ei• = φFi• ; set f (x)(i) = 1 if x ∈ Ei , and use 343Ab.)

343Yc


189

(d) Let (X, Σ, µ) be an atomless probability space. Let µB be the restriction of Lebesgue measure to the σ-algebra of Borel subsets of [0, 1]. Show that there is a function g : X → [0, 1] which is inverseN measure-preserving P∞ −n−1 for µ and µB . (Hint: find an f : X → {0, 1} as in 343Xc, and set g = hf where h(z) = n=0 2 g(n), as in 254K; or choose Eq ∈ Σ such that µEq = q, Eq ⊆ Eq0 whenever q ≤ q 0 in [0, 1] ∩ Q, and set f (x) = inf{q : x ∈ Eq } for x ∈ E1 .) (e) Let (X, Σ, µ) be a countably separated measure space, with measure algebra A. (i) Show that {x} ∈ Σ for every x ∈ X. (ii) Show that every atom of A is of the form {x}• for some x ∈ X. (f ) Let I k be the split interval, with its usual measure µ described in 343J, and h : I k → [0, 1] the canonical surjection. Show that the canonical isomorphism between the measure algebras of µ and Lebesgue measure on [0, 1] is given by the formula ‘E • 7→ h[E]• for every measurable E ⊆ I k ’. (g) Let (X, Σ, µ) and (Y, T, ν) be measure spaces with measure algebras (A, µ ¯), (B, ν¯). Suppose that X ∩ Y = ∅ and that we have a measure-preserving isomorphism π : A → B. Set Λ = {W : W ⊆ X ∪ Y, W ∩ X ∈ Σ, W ∩ Y ∈ T, π(W ∩ X)• = (W ∩ Y )• }, and for W ∈ Λ set λW = µ(W ∩ X) = ν(W ∩ Y ). Show that (X ∪ Y, Λ, λ) is a measure space which is locally compact, or perfect, if (X, Σ, µ) is. (h) Let (X, Σ, µ) be a complete perfect totally finite measure space, (Y, T, ν) a complete countably separated measure space, and f : X → Y an inverse-measure-preserving function. Show that T = {F : F ⊆ Y, f −1 [F ] ∈ Σ}, so that a function h : Y → R is ν-integrable iff hf is µ-integrable. (Hint: if A ⊆ Y and E = f −1 [A] ∈ Σ, f ¹E is inverse-measure-preserving for the subspace measures µE , νA ; by 342Xk, νA is perfect, so by 343L A ∈ T. Now use 235L.) 343Y Further exercises (a) Let (X, Σ, µ) be a semi-finite measure space, and suppose that there is a compact class K ⊆ PX such that (α) whenever T E ∈ Σ and µE > 0 there is a non-negligible K ∈ K such that K ⊆ E (β) whenever K0 , . . . , Kn ∈ K and i≤n Ki = ∅ then there are measurable sets E0 , . . . , En T such that Ei ⊇ Ki for every i and i≤n Ei is negligible. Show that µ is locally compact. (b) (i) Show that a countably separated semi-finite measure space has magnitude at most c and Maharam type at most 2c . (ii) Show that the direct sum of c or fewer countably separated measure spaces is countably separated. (c) Let I k = {t+ : t ∈ [0, 1]} ∪ {t− : t ∈ [0, 1]} be the split interval (343J). (i) Show that the rules s− ≤ t− ⇐⇒ s+ ≤ t+ ⇐⇒ s ≤ t,

s+ ≤ t− ⇐⇒ s < t,

t− ≤ t+ for all t ∈ [0, 1] define a Dedekind complete total order on I k with greatest and least elements. (ii) Show that the intervals [0− , t− ], [t+ , 1+ ], interpreted for this ordering, generate a compact Hausdorff topology on I k for which the map h : I k → [0, 1] of 343J is continuous. (iii) Show that a subset E of I k is Borel for this topology iff the sets Er , El ⊆ [0, 1], as described in 343J, are Borel and Er 4El is countable. (iv) Show that if f : [0, 1] → R is of bounded variation then there is a continuous g : I k → R such that g = f h except perhaps at countably many points. (v) Show that the measure µ of 343J is inner regular with respect to the compact subsets of I k . (vi) Show that we have a lower density φ for µ defined by setting 1

φE = {t− : 0 < t ≤ 1, lim µ(E ∩ [(t − δ)+ , t− ]) = 1} δ↓0 δ

1

∪ {t+ : 0 ≤ t < 1, lim µ(E ∩ [t+ , (t + δ)− ]) = 1} δ↓0 δ

for measurable sets E ⊆ I k .

190

The lifting theorem

343Yd

(d) Set X = {0, 1}c , with its usual measure µ. Show that there is an inverse-measure-preserving function f : X → X such that f [X] is non-measurable but f induces the identity automorphism of the measure algebra of µ. (Hint: use the idea of 343I.) Show that under these conditions f [X], with its subspace measure, must be compact. (Hint: use 343B(iv).) (e) Let µHr be r-dimensional Hausdorff measure on Rs , where s ≥ 1 is an integer and r ≥ 0 (§264). (i) Show that µHr is countably separated. (ii) Show that the c.l.d. version of µHr is compact. (Hint: 264Yi.) (f ) Give an example of a countably separated probability space (X, Σ, µ) and a function f from X to a set Y such that the image measure µf −1 is not countably separated. (Hint: use 223B to show that if E ⊆ R is Lebesgue measurable and not negligible, then E + Q is conegligible; or use the zero-one law to show that if E ⊆ PN is measurable and not negligible for the usual measure on PN, then {a4b : a ∈ E, b ∈ [N]<ω } is conegligible.) 343 Notes and comments The points at which the Lifting Theorem impinges on the work of this section are in the proofs of (iv)⇒(i) and (iv)⇒(v) in Theorem 343B. In fact the ideas can be rearranged to give a proof of 343B which does not rely on the Lifting Theorem; I give a hint in Volume 4 (413Yc). I suppose the significant new ideas of this section are in 343B and 343K. The rest is mostly a matter of being thorough and careful. But I take this material at a slow pace because there are some potentially confusing features, and the underlying question is of the greatest importance: when, given a Boolean homomorphism from one measure algebra to another, can we be sure of representing it by a measurable function between measure spaces? The concept of ‘compact’ space puts the burden firmly on the measure space corresponding to the domain of the Boolean homomorphism, which will be the codomain of the measurable function. So the first step is to try to understand properly which spaces are compact, and what other properties they can be expected to have; which s for much of the length of §342. But having understood that many of our favourite spaces are compact, we have to come to with the fact that we still cannot count on a measure algebra isomorphism corresponding to a measure space isomorphism. I introduce the split interval (343J, 343Xf, 343Yc) as a close approximation to Lebesgue measure on [0, 1] which is not isomorphic to it. Of course we have already seen a more dramatic example: the Stone space of the Lebesgue measure algebra also has the same measure algebra as Lebesgue measure, while being in almost every other way very much more complex, as will appear in Volumes 4 and 5. As 343C suggests, elementary cases in which 343B can be applied are often amenable to more primitive methods, avoiding not only the concept of ‘compact’ measure, but also Stone spaces and the Lifting Theorem. For substantial examples in which we can prove that a measure space (X, µ) is compact, without simultaneously finding direct constructions for inverse-measure-preserving functions into X (as in 343Xc-343Xd), I think we shall have to wait until Volume 4. The concept of ‘countably separated’ measure space does not involve the measure at all, nor even the ideal of negligible sets; it belongs to the theory of σ-algebras of sets. Some simple permanence properties are in 343H and 343Yb(ii). Let us note in ing that 343Xh describes some more situations in which the ‘image measure catastrophe’, described in 235J, cannot arise. I include the variants 343B(ii), 343B(iii) and 343Ya of the notion of ‘local compactness’ because they are not obvious and may illuminate it.

344 Realization of automorphisms In 343Jb, I gave an example of a ‘good’ (compact, complete) probability space X with an automorphism φ of its measure algebra such that both φ and φ−1 are representable by functions from X to itself, but there is no such representation in which the two functions are inverses of each other. The present section is an attempt to describe the further refinements necessary to ensure that automorphisms of measure algebras can be represented by automorphisms of the measure spaces. It turns out that in the most important contexts in which this can be done, a little extra work yields a significant generalization: the simultaneous realization of countably many homomorphisms by a consistent family of functions.

344B

Realization of automorphisms

191

I will describe three cases in which such simultaneous realizations can be achieved: Stone spaces (344A), perfect complete countably separated spaces (344C) and suitable measures on {0, 1}I (344E-344G). The arguments for 344C, suitably refined, give a complete description of perfect complete countably separated strictly localizable spaces which are not purely atomic (344I, 344Xc). At the same time we find that Lebesgue measure, and the usual measure on {0, 1}I , are ‘homogeneous’ in the strong sense that two measurable subspaces (of non-zero measure) are isomorphic iff they have the same measure (344J, 344L). 344A Stone spaces The first case is immediate from the work of §§312, 313 and 321, as collected in 324E. If (Z, Σ, µ) is actually the Stone space of a measure algebra (A, µ ¯), then every order-continuous Boolean homomorphism φ : A → A corresponds to a unique continuous function fφ : Z → Z (312P) which represents φ (324E). The uniqueness of fφ means that we can be sure that fφψ = fψ fφ for all ordercontinuous homomorphisms φ and ψ; and of course fι is the identity map on Z, so that fφ−1 will have to be fφ−1 whenever φ is invertible. Thus in this special case we can consistently, and canonically, represent all order-continuous Boolean homomorphisms from A to itself. Now for two cases where we have to work for the results. 344B Theorem Let (X, Σ, µ) be a countably separated measure space with measure algebra A, and G a countable semigroup of Boolean homomorphisms from A to itself such that every member of G can be represented by some function from X to itself. Then a family hfφ iφ∈G of such representatives can be chosen in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. proof (a) Because G ∪ {ι} satisfies the same conditions as G, we may suppose from the beginning that ι belongs to G itself. Let A ⊆ Σ be a countable set separating the points of X. For each φ ∈ G take some representing function gφ : X → X; take gι to be the identity function. If φ, ψ ∈ G, then of course ((gφ gψ )−1 [E])• = (gψ−1 [gφ−1 [E]])• = ψ(gφ−1 [E])• −1 = ψφE • = (gψφ [E])•

for every E ∈ Σ. By 343F, the set Hφψ = {x : gψφ (x) 6= gφ gψ (x)} is negligible and belongs to Σ. (b) Set H=

S φ,ψ∈G

Hφψ ;

because G is countable, H is also measurable and negligible. Try defining fφ : X → X by setting fφ (x) = gφ (x) if x ∈ X \ H, fφ (x) = x if x ∈ H. Because H is measurable, fφ−1 [E] ∈ Σ for every E ∈ Σ; because H is negligible, (fφ−1 [E])• = (gφ−1 [E])• = φE • for every E ∈ Σ, and fφ represents φ, for every φ ∈ G. Of course fι = gι is the identity function on X. (c) If θ ∈ G then fθ−1 [H] = H. P P (i) If x ∈ H then fθ (x) = x ∈ H. (ii) If fθ (x) ∈ H and fθ (x) = x then of course x ∈ H. (iii) If fθ (x) = gθ (x) ∈ H then there are φ, ψ ∈ G such that gφ gψ gθ (x) 6= gψφ gθ (x). So either gψ gθ (x) 6= gθψ (x), or gφ gθψ (x) 6= gθψφ (x) or gθψφ (x) 6= gψφ gθ (x), and in any case x ∈ H. Q Q

192

The lifting theorem

344B

(d) It follows that fφ fψ = fψφ for every φ, ψ ∈ G. P P (i) If x ∈ H then fφ fψ (x) = x = fψφ (x). (ii) If x ∈ X \ H then fψ (x) ∈ / H, by (c), so fφ fψ (x) = gφ gψ (x) = gψφ (x) = fψφ (x). Q Q 344C Corollary Let (X, Σ, µ) be a countably separated perfect complete strictly localizable measure space with measure algebra A, and G a countable semigroup of order-continuous Boolean homomorphisms from A to itself. Then we can choose simultaneously, for each φ ∈ G, a function fφ : X → X representing φ, in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. In particular, if φ ∈ G is invertible, and φ−1 ∈ G, we shall have fφ−1 = fφ−1 ; so that if moreover φ and φ−1 are measure-preserving, fφ will be an automorphism of the measure space (X, Σ, µ). proof By 343K, (X, Σ, µ) is compact. So 343B(v) tells us that every member of G is representable, and we can apply 344B. Reminder: Spaces satisfying the conditions of this corollary include Lebesgue measure on Rr , the usual measure on {0, 1}N , and their measurable subspaces; see also 342J, 342Xe, 343H and 343Ye. 344D The third case I wish to present requires a more elaborate argument. I start with a kind of Schröder-Bernstein theorem for measurable spaces. Lemma Let X and Y be sets, and Σ ⊆ PX, T ⊆ PY σ-algebras. Suppose that there are f : X → Y , g : Y → X such that F = f [X] ∈ T, E = g[Y ] ∈ Σ, f is an isomorphism between (X, Σ) and (F, TF ) and g is an isomorphism between (Y, T) and (E, ΣE ), writing ΣE , TF for the subspace σ-algebras (see 121A). Then (X, Σ) and (Y, T) are isomorphic, and there is an isomorphism h : X → Y which is covered by f and g in the sense that {(x, h(x)) : x ∈ X} ⊆ {(x, f (x)) : x ∈ X} ∪ {(g(y), y) : y ∈ Y }. proof Set X0 = X, Y0 = Y , Xn+1 = g[Yn ] and Yn+1 = f [Xn ] for each n ∈ N;Tthen hXn in∈N T is a nonincreasing sequence in Σ and hYn in∈N is a non-increasing sequence in T. Set X∞ = n∈N Xn , Y∞ = n∈N Yn . Then f ¹X2k \ X2k+1 is an isomorphism between X2k \ X2k+1 and Y2k+1 \ Y2k+2 , while g¹ Y2k \ Y2k+1 is an isomorphism between Y2k \ Y2k+1 and X2k+1 \ X2k+2 ; and g¹ Y∞ is an isomorphism between Y∞ and X∞ . So the formula [ h(x) = f (x) if x ∈ X2k \ X2k+1 , k∈N

=g

−1

(x) for other x ∈ X

gives the required isomorphism between X and Y . Remark You will recognise the ordinary Schröder-Bernstein theorem (2A1G) as the case Σ = PX, T = PY . 344E Theorem Let I be any set, and let µ be a σ-finite measure on X = {0, 1}I with domain the σ-algebra B generated by the sets {x : x(i) = 1} as i runs over I; write A for the measure algebra of µ. Let G be a countable semigroup of order-continuous Boolean homomorphisms from A to itself. Then we can choose simultaneously, for each φ ∈ G, a function fφ : X → X representing φ, in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. In particular, if φ ∈ G is invertible and φ−1 ∈ G, we shall have fφ−1 = fφ−1 ; so that if moreover φ is measure-preserving, fφ will be an automorphism of the measure space (X, B, µ). proof (a) As in 344C, we may as well suppose from the beginning that ι ∈ G. The case of finite I is trivial, so I will suppose that I is infinite. For i ∈ I, set Ei = {x : x(i) = 1}; for J ⊆ I, let BJ be the σ-subalgebra • of B generated by {Ei : I ∈ J}. For i ∈ I, φ ∈ G choose Fφi ∈ B such that Fφi = φEi• . Let J be the family of those subsets J of I such that Fφi ∈ BJ for every i ∈ J, φ ∈ G.

344E


193

(b) For the purposes of this proof, I will say that a pair (J, hgφ iφ∈G ) is consistent if J ∈ J and, for each φ ∈ G, gφ is a function from X to itself such that gφ−1 [Ei ] ∈ BJ and (gφ−1 [Ei ])• = φEi• whenever i ∈ J, φ ∈ G, gφ−1 [Ei ] = Ei whenever i ∈ I \ J, φ ∈ G, gφ gψ = gψφ whenever φ, ψ ∈ G, gι (x) = x for every x ∈ X. Now the key to the proof is the following fact: if (J, hgφ iφ∈G ) is consistent, and J˜ is a member of J such ˜ h˜ that J˜ \ J is countably infinite, then there is a family h˜ gφ iφ∈G such that (J, gφ iφ∈G ) is consistent and g˜φ−1 [Ei ] = gφ−1 [Ei ] whenever i ∈ J, φ ∈ G. The construction is as follows. (i) Start by fixing on any infinite set K ⊆ J˜ \ J such that (J˜ \ J) \ K is also infinite. For z ∈ {0, 1}K , set Vz = {x : x ∈ X, x¹K = z}; then Vz ∈ BJ˜. All the sets Vz , as z runs over the uncountable set {0, 1}K , are dist, so they cannot all have non-zero measure (because µ is σ-finite), and we can choose z such that Vz is µ-negligible. (ii) Define hφ : X → X, for φ ∈ G, by setting hφ (x)(i) = gφ (x)(i) if i ∈ J, ˜ = x(i) if i ∈ I \ J, = x(i) if i ∈ J˜ \ J and x ∈ Vz , = 1 if i ∈ J˜ \ J and x ∈ Fφi \ Vz , = 0 if i ∈ J˜ \ J and x ∈ / Fφi ∪ Vz . Because Vz ∈ BJ˜ and µVz = 0, we see that −1 (α) h−1 φ [Ei ] = gφ [Ei ] ∈ BJ if i ∈ J, −1 ˜ (β) h−1 φ [Ei ] ∈ BJ˜ and hφ [Ei ]4Fφi is negligible if i ∈ J \ J, and consequently • • ˜ (γ) (h−1 φ [Ei ]) = φEi for every i ∈ J, • • (δ) (h−1 φ [E]) = φE for every E ∈ BJ˜ (by 343Ab); moreover, −1 (²) h−1 φ [E] = gφ [E] for every E ∈ BJ ,

(ζ) h−1 φ [E] ∈ BJ˜ for every E ∈ BJ˜, ˜ (η) h−1 [Ei ] = Ei if i ∈ I \ J, φ

so that (θ) h−1 φ [E] ∈ B for every E ∈ B; finally (ι) hι (x) = x for every x ∈ X. (iii) The next step is to note that if φ, ψ ∈ G then Hφ,ψ = {x : x ∈ X, hφ hψ (x) 6= hψφ (x)} belongs to BJ˜ and is negligible. P P Hφ,ψ =

S i∈I

−1 −1 h−1 ψ [hφ [Ei ]]4hψφ [Ei ].

−1 Now if i ∈ J, then h−1 φ [Ei ] = gφ [Ei ] ∈ BJ , so −1 −1 −1 −1 −1 −1 −1 h−1 ψ [hφ [Ei ]] = hψ [gφ [Ei ]] = gψ [gφ [Ei ]] = gψφ [Ei ] = hψφ [Ei ].

˜ Next, for i ∈ I \ J, −1 −1 −1 h−1 ψ [hφ [Ei ]] = hψ [Ei ] = Ei = hψφ [Ei ].

So

194

The lifting theorem

Hφ,ψ =

S ˜ i∈J\J

344E

−1 −1 h−1 ψ [hφ [Ei ]]4hψφ [Ei ].

But for any particular i ∈ J˜ \ J, Ei and h−1 φ [Ei ] belong to BJ˜, so −1 −1 −1 • • • • (h−1 ψ [hφ [Ei ]]) = ψ(hφ [Ei ]) = ψφEi = (hψφ [Ei ]) , −1 −1 and h−1 ψ [hφ [Ei ]]4hψφ [Ei ] is a negligible set, which by (ii-ζ) belongs to BJ˜. So Hφ,ψ is a countable union of sets of measure 0 in BJ˜ and is itself a negligible member of BJ˜, as claimed. Q Q

(iv) Set H=

S φ,ψ∈G

Hφ,ψ ∪

S φ∈G

h−1 φ [Vz ].

Then H ∈ BJ˜ and µH = 0. P P We know that every Hφ,ψ is negligible and belongs to BJ˜ ((iii) above), that −1 • • every h−1 [V ] belongs to B (by (ii-ζ), and that (h−1 ˜ z φ φ [Vz ]) = φVz = 0, so that hφ [Vz ] is negligible, for every J Q Also, of course, Vz = h−1 φ ∈ G (by (ii-δ)). Consequently H is negligible and belongs to BJ˜. Q ι [Vz ] ⊆ H. Next, hφ (x) ∈ / H whenever x ∈ X \ H and φ ∈ G. P P If ψ, θ ∈ G then hθψ hφ (x) = hφθψ (x) = hψ hφθ (x) = hψ hθ hφ (x), hψ hφ (x) = hφψ (x) ∈ / Vz because x∈ / Hθψ,φ ∪ Hφθ,ψ ∪ Hθ,φ ∪ Hψ,φ ∪ h−1 φψ [Vz ]; thus hφ (x) ∈ / Hψ,θ ∪ h−1 / H. Q Q ψ [Vz ]; as ψ and θ are arbitrary, hφ (x) ∈ (v) The next fact we need is that there is a bijection q : X → H such that (α) for E ⊆ H, E ∈ BJ˜ iff q −1 [E] ∈ BJ˜ (β) q(x)(i) = x(i) for every i ∈ I \ (J˜ \ J), x ∈ X. P P Fix any bijection r : J˜ \ J → J˜ \ (J ∪ K). Consider the maps p1 : X → H, p2 : H → X given by p1 (x)(i) = x(r−1 (i)) if i ∈ J˜ \ (J ∪ K), = z(i) if i ∈ K, = x(i) if i ∈ X \ (J˜ \ J), p2 (y) = y for x ∈ X, y ∈ H. Then p1 is actually an isomorphism between (X, BJ˜) and (Vz , BJ˜ ∩ PVz ). So p1 , p2 are isomorphisms between (X, BJ˜), (H, BJ˜ ∩ PH) and measurable subspaces of H, X respectively. By 344D, there is an isomorphism q between X and H such that, for every x ∈ X, either q(x) = p1 (x) or p2 (q(x)) = x. Since p1 (x)¹I \ (J˜ \ J) = x¹I \ (J˜ \ J) for every x ∈ X, and p2 (y)¹I \ (J˜ \ J) = y¹I \ (J˜ \ J) for every y ∈ H, q(x)¹I \ (J˜ \ J) = x¹I \ (J˜ \ J) for every x ∈ X. Q Q ˜ φ ∈ G then g −1 [Ei ] belongs (vi) An incidental fact which will be used below is the following: if i ∈ J, φ to BJ˜, because it belongs to BJ if i ∈ J, and otherwise is equal to Ei . Consequently gφ−1 [E] ∈ BJ˜ for every E ∈ BJ˜. (vii) I am at last ready to give a formula for g˜φ . For φ ∈ G set g˜φ (x) = hφ (x) if x ∈ X \ H, = qgφ q −1 (x) if x ∈ H. ˜ h˜ Now (J, gφ iφ∈G ) is consistent. P P ˜ (α) If i ∈ J, φ ∈ G, −1 −1 g˜φ−1 [Ei ] = (h−1 [Ei ∩ H]]] ∈ B˜J φ [Ei ] \ H) ∪ q[gφ [q −1 because H ∈ BJ˜ and h−1 [H ∩ E], gφ−1 [E] and q[E] all belong to BJ˜ for every E ∈ BJ˜. At the same φ [E], q time, because g˜φ agrees with hφ on the conegligible set X \ H, • • (˜ gφ−1 [Ei ])• = (h−1 φ [Ei ]) = φEi .

344E


195

˜ φ ∈ G, x ∈ X then (β) If i ∈ I \ J, gφ (x)(i) = hφ (x)(i) = q(x)(i) = x(i), and if x ∈ H then q −1 (x)(i) is also equal to x(i); so g˜φ (x)(i) = x(i). But this means that g˜φ−1 [Ei ] = Ei . (γ) If φ, ψ ∈ G and x ∈ X \ H, then g˜ψ (x) = hψ (x) ∈ X \ H by (iv) above. So g˜φ g˜ψ (x) = hφ hψ (x) = hψφ (x) = gψφ (x) because x ∈ / Hφ,ψ . While if x ∈ H, then g˜ψ (x) = qgψ q −1 (x) ∈ H, so g˜φ g˜ψ (x) = qgφ q −1 qgψ q −1 (x) = qgφ gψ q −1 (x) = qgψφ q −1 (x) = g˜ψφ (x). Thus g˜φ g˜ψ = g˜ψφ . (δ) Because gι (x) = hι (x) = x for every x, g˜ι (x) = x for every x. Q Q (viii) Finally, if i ∈ J and φ ∈ G, q −1 [Ei ] = Ei , so that q[Ei ∩ H] = Ei . Accordingly q(x)¹J = x¹J for every x ∈ X, while q −1 (x)¹J = x¹J for x ∈ H. So gφ q −1 (x)¹J = gφ (x)¹J for x ∈ H, and g˜φ (x)(i) = hφ (x)(i) = gφ (x)(i) if x ∈ X \ H, = qgφ q −1 (x)(i) = gφ q −1 (x)(i) = gφ (x)(i) if x ∈ X \ H. ˜ h˜ Thus (J, gφ iφ∈G ) satisfies all the required conditions. (c) The remaining idea we need is the following: there is a non-decreasingSfamily hJξ iξ≤κ in J , for some cardinal κ, such that Jξ+1 \ Jξ is countably infinite for every ξ < κ, Jξ = η<ξ Jη for every limit ordinal η < κ, and Jκ = I. P P Recall that I am already supposing that I is infinite. If I is countable, set κ = 1, J0 = ∅, J1 = I. Otherwise, set κ = #(I) and let hiξ iξ<κ be an enumeration of I. For i ∈ I, φ ∈ G let Kφi ⊆ I be a countable set such that Fφi ∈ BKφi . Choose the Jξ inductively, as follows. The inductive hypothesis must include the requirement that #(Jξ ) ≤ max(ω, #(ξ)) for every ξ. Start by setting J0 = ∅. Given S ξ<κ and Jξ ∈ J with #(Jξ ) ≤ max(ω, #(ξ)) < κ, take an infinite set L ⊆ κ \ Jξ and set Jξ+1 = Jξ ∪ n∈N Ln , where L0 = L ∪ {iξ }, S Ln+1 = i∈Ln ,φ∈G Kφi for n ∈ N, so that every Ln is countable, and Jξ+1

Fφi ∈ BLn+1 whenever i ∈ Ln , φ ∈ G S ∈ J ; since L ⊆ Jξ+1 \ Jξ ⊆ n∈N Ln , Jξ+1 \ Jξ is countably infinite, and

#(Jξ+1 ) = max(ω, #(Jξ )) ≤ max(ω, #(ξ)) = max(ω, #(ξ + 1)). S For non-zero limit ordinals ξ < κ, set Jξ = η<ξ Jη ; then #(Jξ ) ≤ max(ω, #(ξ), supη<ξ #(Jη )) ≤ max(ω, #(ξ)). Thus the induction proceeds. Observing that the construction puts iξ into Jξ+1 for every ξ, we see that Jκ will be the whole of I, as required. Q Q (d) Now put (b) and (c) together, as follows. Take hJξ iξ≤κ from (c). Set fφ0 (x) = x for every φ ∈ G, x ∈ X; then, because J0 = ∅, (J0 , hfφ0 iφ∈G ) is consistent in the sense of (b). Given that (Jξ , hfφξ iφ∈G ) is consistent, where ξ < κ, use the construction of (b) to find a family hfφ,ξ+1 iφ∈G such that (Jξ+1 , hfφ,ξ+1 iφ∈G ) is consistent and fφ,ξ+1 (x)(i) = fφξ (x)(i) for every i ∈ Jξ . At a non-zero limit ordinal ξ ≤ κ, set fφξ (x)(i) = fφη (x)(i) if x ∈ X, η < ξ, i ∈ Jη , = x(i) if i ∈ I \ Jξ .

196

The lifting theorem

344E

(The inductive hypothesis includes the requirement that fφη (x)¹Jζ = fφζ (x)¹Jζ whenever φ ∈ G, x ∈ X and ζ ≤ η < ξ.) To see that (Jξ , hfφξ iφ∈G ) is consistent, the only non-trivial point to check is that fφ,ξ fψ,ξ = fψφ,ξ for all φ, ψ ∈ G. But if i ∈ Jξ there is some η < ξ such that i ∈ Jη , and in this case −1 −1 fψ,ξ [Ei ] = fψ,η [Ei ] ∈ BJη

is determined by coordinates in Jη , so that (because fφ,ξ (x)¹Jη = fφ,η (x)¹Jη for every x) −1 −1 −1 −1 −1 −1 fφ,ξ [fψ,ξ [Ei ]] = fφ,η [fψ,η [Ei ]] = fψφ,η [Ei ] = fψφ,ξ [Ei ];

while if i ∈ I \ Jξ then −1 −1 −1 −1 −1 fψφ,ξ [Ei ] = Ei = fφ,ξ [Ei ] = fψ,ξ [Ei ] = fψ,ξ [fφ,ξ [Ei ]].

Thus −1 −1 −1 fψ,ξ [fφ,ξ [Ei ]] = fψφ,ξ [Ei ]

for every i, and fφ,ξ fψ,ξ = fφψ,ξ . On completing the induction, set fφ = fφκ for every φ ∈ G; it is easy to see that hfφ iφ∈G satisfies the conditions of the theorem. 344F Corollary Let I be any set, and let µ be a σ-finite measure on X = {0, 1}I . Suppose that µ is the completion of its restriction to the σ-algebra B generated by the sets {x : x(i) = 1} as i runs over I. Write A for the measure algebra of µ. Let G be a countable semigroup of order-continuous Boolean homomorphisms from A to itself. Then we can choose simultaneously, for each φ ∈ G, a function fφ : X → X representing φ, in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. In particular, if φ ∈ G is invertible and φ−1 ∈ G, we shall have fφ−1 = fφ−1 ; so that if moreover φ is measure-preserving, fφ will be an automorphism of the measure space (X, Σ, µ). proof Apply 344E to µ¹B; of course A is canonically isomorphic to the measure algebra of µ¹B (322Da). The functions fφ provided by 344E still represent the homomorphisms φ when re-interpreted as functions on the completed measure space ({0, 1}I , µ), by 343Ac. 344G Corollary Let I be any set, ν the usual measure on {0, 1}I , and A its measure algebra. Then any measure-preserving automorphism of A is representable by a measure space automorphism of ({0, 1}I , ν). 344H Lemma Let (X, Σ, µ) be a locally compact semi-finite measure space which is not purely atomic. Then there is a negligible subset of X of cardinal c. proof Let E be a set of non-zero finite measure not including any atom. Let K ⊆SPE be a compact class such that the subspace measure µE is inner regular with respect to K. Set S = n∈N {0, 1}n , and choose hKz iz∈S inductively , as follows. K∅ is to be any non-negligible member of K ∩ Σ included in E. Given that Kz ⊆ E and µKz > 0, where z ∈ {0, 1}n , take Fz , Fz0 ⊆ Kz to be dist non-negligible measurable sets both of measure at most 3−n ; such exist because µ is semi-finite and E does not include any atom. Choose Kza 0 ⊆ Fz , Kza 1 ⊆ Fz0 to be non-negligible of K ∩ Σ. For each w ∈ {0, 1}N , hKw¹n in∈N is a decreasing T sequence of of K all of non-zero measure, so has non-empty intersection; choose a point xw ∈ n∈N Kw¹n . Since Kza 0 ∩ Kza 1 = ∅ for every z ∈ S, all the xw are distinct, and A = {xw : w ∈ {0, 1}N } has cardinal c. Also S A ⊆ z∈{0,1}n Kz which has measure at most 2n 3−(n−1) for every n ≥ 1, so µ∗ A = 0 and A is negligible. Remark I see that in this proof I have slipped into a notation which is a touch more sophisticated than what I have used so far. See 3A1H for a note on the interpretations of ‘{0, 1}n ’, ‘{0, 1}N ’ which make sense of the formulae here. I ought to note also that the lemma is valid for all perfect spaces; see 344Yf.

344K


197

344I Theorem Let (X, Σ, µ) and (Y, T, ν) be atomless, perfect, complete, strictly localizable, countably separated measure spaces of the same non-zero magnitude. Then they are isomorphic. proof (a) The point is that the measure algebra (A, µ ¯) of µ has Maharam type ω. P P Let hEn in∈N be a sequence in Σ separating the points of X. Let Σ0 be the σ-subalgebra of Σ generated by {En : n ∈ N}, and A0 the order-closed subalgebra of A generated by {En• : n ∈ N}; then E • ∈ A0 for every E ∈ Σ0 , and (X, Σ0 , µ¹Σ0 ) is countably separated. Let f : X → R be Σ0 -measurable and injective (343E). Of course f is also Σ-measurable. If a ∈ A \ {0}, express a as E • where E ∈ Σ. Because (X, Σ, µ) is perfect, there is a compact K ⊆ R such that K ⊆ f [E] and µf −1 [K] > 0. K is surely a Borel set, so f −1 [K] ∈ Σ0 and b = f −1 [K]• ∈ A0 \ {0}. But because f is injective, we also have f −1 [K] ⊆ E and b ⊆ a. As a is arbitrary, A0 is order-dense in A; but A0 is order-closed, so must be the whole of A. Thus A is τ -generated by the countable set {En• : n ∈ N}, and τ (A) ≤ ω. Q Q On the other hand, because A is atomless, and not {0}, none of its principal ideals can have finite Maharam type, and it is Maharam-type-homogeneous, with type ω. (b) Writing (B, ν¯) for the measure algebra of ν, we see that the argument of (a) applies equally to (B, ν¯), so that (A, µ ¯) and (B, ν¯) are atomless localizable measure algebras, of Maharam type ω and the same magnitude. Consequently they are isomorphic as measure algebras, by 332J. Let φ : A → B be a measure-preserving isomorphism. By 343B, there are functions g : Y → X and f : X → Y representing φ and φ−1 , because µ and ν are compact and complete and strictly localizable. Now f g : Y → Y and gf : X → X represent the identity automorphisms on B, A, so by 343F are equal almost everywhere to the identity functions on Y , X respectively. Set E = {x : x ∈ X, gf (x) = x},

F = {y : y ∈ Y, f g(y) = y};

then both E and F are conegligible. Of course f [E] ⊆ F (since f gf (x) = f (x) for every x ∈ E), and similarly g[F ] ⊆ E; consequently f ¹E, g¹F are the two halves of a one-to-one correspondence between E and F . Because φ is measure-preserving, µf −1 [H] = νH, νg −1 [G] = µG for every G ∈ Σ, H ∈ T; accordingly f ¹E is an isomorphism between the subspace measures on E and F . (c) By 344H, applied to the subspace measure on E, there is a negligible set A ⊆ E of cardinal c. Now X and Y , being countably separated, both have cardinal at most c. (There are injective functions from X and Y to R.) Set B = A ∪ (X \ E), C = f [A] ∪ (Y \ F ). Then B and C are negligible subsets of X, Y respectively, and both have cardinal c precisely, so there is a bijection h : B → C. Set f1 (x) = f (x) if x ∈ X \ B = E \ A, = h(x) if x ∈ B. Then, because µ and ν are complete, f1 is an isomorphism between the measure spaces (X, Σ, µ) and (Y, T, ν), as required. 344J Corollary Suppose that E, F are two Lebesgue measurable subsets of Rr of the same non-zero measure. Then the subspace measures on E and F are isomorphic. 344K Corollary (a) A measure space is isomorphic to Lebesgue measure on [0, 1] iff it is an atomless countably separated compact (or perfect) complete probability space; in this case it is also isomorphic to the usual measure on {0, 1}N . (b) A measure space is isomorphic to Lebesgue measure on R iff it is an atomless countably separated compact (or perfect) σ-finite measure space which is not totally finite; in this case it is also isomorphic to Lebesgue measure on any Euclidean space Rr . (c) Let µ be Lebesgue measure on R. If 0 < µE < ∞ and we set νF = F ⊆ E, then (E, ν) is isomorphic to Lebesgue measure on [0, 1].

1 µF µE

for every measurable

198

The lifting theorem

344L

344L

The homogeneity property of Lebesgue measure described in 344J is repeated in {0, 1}I for any

I. Theorem Let I be any set, and µ the usual measure on {0, 1}I . If E, F ⊆ {0, 1}I are two measurable sets of the same non-zero finite measure, the subspace measures on E and F are isomorphic. proof Write X = {0, 1}I . (a) If I is finite, then X, E and F are all finite, and #(E) = #(F ), so the result is trivial. If I is countably infinite, then the subspace measures are perfect and complete and countably separated, so the result follows from 344I. So let us suppose that I is uncountable. (b) Let (A, µ ¯) be the measure algebra of µ. Then A is homogeneous, and µ ¯E • = µ ¯F • , so there is • • an automorphism φ : A → A such that φE = F . (Apply 333D to the finite subalgebra C generated by {E • , F • }; or argue directly from 331I.) By 344F, there is a measure space isomorphism f : X → X representing φ, so that f −1 [E]4F and f [F ]4E are negligible. (c) We can find a countably infinite set J ⊆ I and a measurable set E 0 such that E 0 is determined by coordinates in J, E 0 ⊆ E ∩ f [F ] and E \ E 0 is negligible; so that E 0 is non-empty. Take any x0 ∈ E 0 and set V = {x : x ∈ X, x¹J = x0 ¹J}; then V is a negligible subset of E and #(V ) = #(X) (because #(I \ J) = #(I) and x 7→ x¹I \ J : V → {0, 1}I\J is a bijection). Setting E1 = E ∩ f [F ], F1 = F ∩ f −1 [E], f ¹E1 is a bijection between E1 and F1 . Now #(V ∪ (E \ E1 )) = #(X) = #(f −1 [V ] ∪ (F \ F1 )), so there is a bijection h : f −1 [V ] ∪ (F \ F1 ) → V ∪ (E \ E1 ). Define g : E → F by writing g(x) = f (x) if x ∈ F1 \ f −1 [V ], = h(x) if x ∈ f −1 [V ] ∪ (F \ F1 ). Then g is a bijection, and because g = f almost everywhere on F , g −1 = f −1 almost everywhere on E, g is an isomorphism for the subspace measures. 344X Basic exercises (a) Let (X, Σ, µ) and (Y, T, ν) be measure spaces, and suppose that there are E ∈ Σ, F ∈ T such that (X, Σ, µ) is isomorphic to the subspace (F, TF , νF ), while (Y, T, ν) is isomorphic to (E, ΣE , µE ). Show that (X, Σ, µ) and (Y, T, ν) are isomorphic. (b) Let (X, Σ, µ) and (Y, T, ν) be perfect countably separated complete strictly localizable measure spaces with isomorphic measure algebras. Show that there are conegligible subsets X 0 ⊆ X, Y 0 ⊆ Y such that X 0 and Y 0 , with the subspace measures, are isomorphic. (c) Let (X, Σ, µ) and (Y, T, ν) be perfect countably separated complete strictly localizable measure spaces with isomorphic measure algebras. Suppose that they are not purely atomic. Show that they are isomorphic. (d) Give an example of two perfect countably separated complete probability spaces, with isomorphic measure algebras, which are not isomorphic. (e) Let (Z, Σ, µ) be the Stone space of a Maharam-type-homogeneous measure algebra. Show that if E, F ∈ Σ have the same non-zero finite measure, then the subspace measures on E and F are isomorphic. (f ) Let (I k , Σ, µ) be the split interval with its usual measure (343J), and A its measure algebra. (i) Show that every automorphism of A is represented by a measure space automorphism of I k . (ii) Show that if E, F ∈ Σ and µE = µF > 0 then the subspace measures on E and F are isomorphic. (g) Let I be an infinite set, and µ the usual measure on {0, 1}I . Show that if E ⊆ {0, 1}I is any set of non-zero measure, then the subspace measure on E is isomorphic to a multiple of µ.

344 Notes


199

344Y Further exercises (a) Let X be a set, Σ a σ-algebra of subsets of X, I a σ-ideal of Σ, and A the quotient Σ/I. Suppose that there is a countable set A ⊆ Σ separating the points of X. Let G be a countable semigroup of Boolean homomorphisms from A to itself such that every member of G can be represented by some function from X to itself. Show that a family hfφ iφ∈G of such representatives can be chosen in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. (b) Let A, B be Dedekind σ-complete Boolean algebras. Suppose that each is isomorphic to a principal ideal of the other. Show that they are isomorphic. (c) Let I be an infinite set, and write B for the σ-algebra of subsets of X = {0, 1}I generated by the sets {x : x(i) = 1} as i runs over I. Let µ and ν be σ-finite measures on X, both with domain B, and with measure algebras (A, µ ¯), (B, ν¯). Show that any Boolean isomorphism φ : A → B is represented by a bijection f : X → X such that f −1 represents φ−1 : B → A, and hence that (A, µ ¯) is isomorphic to (B, ν¯) iff (X, B, µ) is isomorphic to (X, B, ν). (d) Let I be any set, and write B for the σ-algebra of subsets of X = {0, 1}I generated by the sets {x : x(i) = 1} as i runs over I. Let I be an ω1 -saturated ideal of B, and write A for the quotient Boolean algebra B/I. Let G be a countable semigroup of order-continuous Boolean homomorphisms from A to itself. Show that we can choose simultaneously, for each φ ∈ G, a function fφ : X → X representing φ, in such a way that fφψ = fψ fφ for all φ, ψ ∈ G; and if the identity automorphism ι belongs to G, then we may arrange that fι is the identity function on X. In particular, if φ ∈ G is invertible and φ−1 ∈ G, fφ will be an automorphism of the structure (X, B, I). (e) Let I be any set, and write B for the σ-algebra of subsets of X = {0, 1}I generated by the sets {x : x(i) = 1} as i runs over I. Let I, J be ω1 -saturated ideals of B. Show that if the Boolean algebras B/I and B/J are isomorphic, so are the structures (X, B, I) and (X, B, J ). (f ) Show that if (X, Σ, µ) is a perfect semi-finite measure space which is not purely atomic, there is a negligible set of cardinal c. (Hint: reduce to the case in which µ is atomless and totally finite; in this case, construct a measurable function f : X → R such that the image measure ν = µf −1 is atomless, and apply 344H to ν.) 344 Notes and comments In this section and the last, I have allowed myself to drift some distance from the avowed subject of this chapter; but it seemed a suitable place for this material, which is fundamental to abstract measure theory. We find that the concepts of §§342-343 are just what is needed to characterise Lebesgue measure (344K), and the characterization shows that among non-negligible measurable subspaces of Rr the isomorphism classes are determined by a single parameter, the measure of the subspace. Of course a very large number of other spaces – indeed, most of those appearing in ordinary applications of measure theory to other topics – are perfect and countably separated (for example, those of 342Xe and 343Ye), and therefore covered by this classification. I note that it includes, as a special case, the isomorphism between Lebesgue measure on [0, 1] and the usual measure on {0, 1}N already described in 254K. In 344I, the first part of the proof is devoted to showing that a perfect countably separated measure space has countable Maharam type; I ought perhaps to note here that we must resist the temptation to suppose that all countably separated measure spaces have countable Maharam type. In fact there are countably separated probability spaces with Maharam type as high as 2c . The arguments are elementary but seem to fit better into Volume 5 than here. I have offered three contexts in which automorphisms of measure algebras are represented by automorphisms of measure spaces (344A, 344C, 344E). In the first case, every automorphism can be represented simultaneously in a consistent way. In the other two cases, there is, I am sure, no such consistent family of representations which can be constructed within ZFC; but the theorems I give offer consistent simultaneous representations of countably many homomorphisms. The question arises, whether ‘countably many’ is the true natural limit of the arguments. In fact it is possible to extend both results to families of at most ω1 automorphisms. I hope to return to this in Volume 5.

200

Liftings

344 Notes

Having successfully characterized Lebesgue measure – or, what is very nearly the same thing, the usual measure on {0, 1}N – it is natural to seek similar characterizations of the usual measures on {0, 1}κ for uncountable cardinals κ. This seems to be hard. A variety of examples (which I hope to describe in Volume 5) show that none of the most natural conjectures can be provable in ZFC. In fact the principal new ideas of this section do not belong specifically to measure theory; rather, they belong to the general theory of σ-algebras and σ-ideals of sets. In the case of the Schröder-Bernstein-type theorem 344D, this is obvious from the formulation I give. (See also 344Yb.) In the case of 344B and 344E, I offer generalizations in 344Ya-344Ye. Of course the applications of 344B here, in 344C and its corollaries, depend on Maharam’s theorem and the concept of ‘compact’ measure space. The former has no generalization to the wider context, and the value of the latter is based on the equivalences in Theorem 343B, which also do not have simple generalizations. The property described in 344J and 344L – a measure space (X, Σ, µ) in which any two measurable subsets of the same non-zero measure are isomorphic – seems to be a natural concept of ‘homogeneity’ for measure spaces; it seems unreasonable to ask for all sets of zero measure to be isomorphic, since finite sets of different cardinalities can be expected to be of zero measure. An extra property, shared by Lebesgue measure and the usual measure on {0, 1}I (and by the measure on the split interval, 344Xf) but not by counting measure, would be the requirement that measurable sets of different non-zero finite measures should be isomorphic up to a scalar multiple of the measure. All these examples have the further property, that all automorphisms of their measure algebras correspond to automorphisms of the measure spaces.

345 Translation-invariant liftings In this section and the next I complement the work of §341 by describing some important special properties which can, in appropriate circumstances, be engineered into our liftings. I begin with some remarks on translation-invariance. I restrict my attention to measure spaces which we have already seen, delaying a general discussion of translation-invariant measures on groups until Volume 4, and to results which can be proved without special axioms, delaying the use of the continuum hypothesis, in particular, until Volume 5. 345A Translation-invariant liftings In this section I shall consider two forms of translation-invariance, as follows. (a) Let µ be Lebesgue measure on Rr , and Σ its domain. A lifting φ : Σ → Σ is translation-invariant if φ(E + x) = φE + x for every E ∈ Σ, x ∈ Rr . (Recall from 134A that E + x = {y + x : y ∈ E} belongs to Σ for every E ∈ Σ, x ∈ Rr .) Similarly, writing A for the measure algebra of µ, a lifting θ : A → Σ is translation-invariant if θ(E + x)• = θE • + x for every E ∈ Σ, x ∈ Rr . It is easy to see that if θ and φ correspond to each other in the manner of 341B, then one is translationinvariant if and only it the other is. (b) Now let I be any set, and let µ be the usual measure on X = {0, 1}I , with Σ its domain and A its measure algebra. For x, y ∈ X, define x + y ∈ X by setting (x + y)(i) = x(i) +2 y(i) for every i ∈ I; that is, give X the group structure of the product group ZI2 . This makes X an abelian group (isomorphic to the additive group (PI, 4) of the Boolean algebra PI, if we match x ∈ X with {i : x(i) = 1} ⊆ I). Recall that the measure µ is a product measure (254J), being the product of copies of the fair-coin probability on the two-element set {0, 1}. If x ∈ X, then for each i ∈ I the map ² 7→ ²+2 x(i) : {0, 1} → {0, 1} is a measure space automorphism of {0, 1}, since the two singleton sets {0} and {1} have the same measure 1 2 . It follows at once that the map y 7→ y + x : X → X is a measure space automorphism. Accordingly we can again say that a lifting θ : A → Σ, or φ : Σ → Σ, is translation-invariant if θ(E + x)• = θE • + x, for every E ∈ Σ, x ∈ X.

φ(E + x) = φE + x

345C

Translation-invariant liftings

201

345B Theorem For any r ≥ 1, there is a translation-invariant lifting of Lebesgue measure on Rr . proof (a) Write µ for Lebesgue measure on Rr , Σ for its domain. Let φ : Σ → Σ be lower Lebesgue density (341E). Then φ is translation-invariant in the sense that φ(E + x) = φE + x for every E ∈ Σ, x ∈ Rr . P P φ(E + x) = {y : y ∈ Rr , lim

µ(E+x)∩B(y,δ) µ(B(y,δ))

= 1}

= {y : y ∈ Rr , lim

µ(E∩B(y−x,δ)) µ(B(y−x,δ))

= 1}

δ↓0

δ↓0

(because µ is translation-invariant) = {y + x : y ∈ Rr , lim δ↓0

µ(E∩B(y,δ)) µ(B(y,δ))

= 1}

= φE + x. Q Q (b) Let φ0 be any lifting of µ such that φ0 E ⊇ φE for every E ∈ Σ (341J). Consider φE = {y : 0 ∈ φ0 (E − y)} for E ∈ Σ. It is easy to check that φ : Σ → Σ is a Boolean homomorphism because φ0 is, so that, for instance, y ∈ φE4φF ⇐⇒ 0 ∈ φ0 (E − y)4φ0 (F − y) ⇐⇒ 0 ∈ φ0 ((E − y)4(F − y)) = φ0 ((E4F ) − y)) ⇐⇒ y ∈ φ(E4F ). (c) If µE = 0, then E − y is negligible for every y ∈ Rr , so φ0 (E − y) is always empty and φE = ∅. (d) Next, φE ⊆ φE for every E ∈ Σ. P P If y ∈ φE, then 0 = y − y ∈ φE − y = φ(E − y) ⊆ φ0 (E − y), so y ∈ φE. Q Q By 341Ib, φ is a lifting for µ. (e) Finally, φ is translation-invariant, because if E ∈ Σ and x, y ∈ Rr then y ∈ φ(E + x) ⇐⇒ 0 ∈ φ0 (E + x − y) = φ0 (E − (y − x)) ⇐⇒ y − x ∈ φE ⇐⇒ y ∈ φE + x. 345C Theorem For any set I, there is a translation-invariant lifting of the usual measure on {0, 1}I . proof I base the argument on the same programme as in 345B. This time we have to work rather harder, as we have no simple formula for a translation-invariant lower density. However, the ideas already used in 341F-341H are in fact adequate, if we take care, to produce one. (a) Since there is certainly a bijection between I and its cardinal κ = #(I), it is enough to consider the case I = κ. Write µ for the usual measure on X = {0, 1}I = {0, 1}κ and Σ for its domain. For each ξ < κ set Eξ = {x : x ∈ X, x(ξ) = 1}, and let Σξ be the σ-algebra generated by {Eη : η < ξ}. Because x + Eη is either Eη or X \ Eη , and in either case belongs to Σξ , for every η < ξ and x ∈ X, Σξ is translation-invariant. (Consider the algebra Σ0ξ = {E : E + x ∈ Σξ for every x ∈ X}; this must be Σξ .) Let Φξ be the set of partial lower densities φ : Σξ → Σ which are translation-invariant in the sense that φ(E + x) = φE + x for any E ∈ Σξ , x ∈ X. (b)(i) For ξ < κ, Σξ+1 is just the algebra of subsets of X generated by Σξ ∪ {Eξ }, that is, sets of the form (F ∩ Eξ ) ∪ (G \ Eξ ) where F , G ∈ Σξ (312M). Moreover, the expression is unique. P P Define xξ ∈ X by

202

Liftings

345C

setting xξ (ξ) = 1, xξ (η) = 0 if η 6= ξ. Then xξ + Eη = Eη for every η < ξ, so xξ + F = F for every F ∈ Σξ . If H = (F ∩ Eξ ) ∪ (G \ Eξ ) where F , G ∈ Σξ , then xξ + H = ((xξ + F ) ∩ (xξ + Eξ )) ∪ ((xξ + G) \ (xξ + Eξ )) = (F \ Eξ ) ∪ (G ∩ Eξ ), so F = (H ∩ Eξ ) ∪ ((xξ + H) \ Eξ ) = FH , G = (H \ Eξ ) ∪ ((xξ + H) ∩ Eξ ) = GH are determined by H. Q Q (ii) The functions H 7→ FH , H 7→ GH : Σξ+1 → Σξ defined above are clearly Boolean homomorphisms; moreover, if H, H 0 ∈ Σξ+1 and H4H 0 is negligible, then (FH 4FH 0 ) ∪ (GH 4GH 0 ) ⊆ (H4H 0 ) ∪ (xξ + (H4H 0 )) is negligible. It follows at once that if ξ < κ and φ ∈ Φξ , we can define φ1 : Σξ+1 → Σ by setting φ1 H = (φFH ∩ Eξ ) ∪ (φGH \ Eξ ), and φ1 will be a lower density. If H ∈ Σξ then FH = GH = H, so φ1 H = φH. (iii) To see that φ1 is translation-invariant, observe that if x ∈ X and x(ξ) = 0 then x + Eξ = Eξ , so, for any F , G ∈ Σξ , φ1 (x + ((F ∩ Eξ ) ∪ (G \ Eξ ))) = φ1 (((F + x) ∩ Eξ ) ∪ ((G + x) \ Eξ )) = (φ(F + x) ∩ Eξ ) ∪ (φ(G + x) \ Eξ ) = ((φF + x) ∩ Eξ ) ∪ ((φG + x) \ Eξ ) = x + (φF ∩ Eξ ) ∪ (φG \ Eξ ) = x + φ1 ((F ∩ Eξ ) ∪ (G \ Eξ )). While if x(ξ) = 1 then x + Eξ = X \ Eξ , so φ1 (x + ((F ∩ Eξ ) ∪ (G \ Eξ ))) = φ1 (((F + x) \ Eξ ) ∪ ((G + x) ∩ Eξ )) = (φ(F + x) \ Eξ ) ∪ (φ(G + x) ∩ Eξ ) = ((φF + x) \ Eξ ) ∪ ((φG + x) ∩ Eξ ) = x + (φF ∩ Eξ ) ∪ (φG \ Eξ ) = x + φ1 ((F ∩ Eξ ) ∪ (G \ Eξ )). So φ1 ∈ Φξ+1 . (iv) Thus every member of Φξ has an extension to a member of Φξ+1 . (c) Now suppose that hζ(n)in∈N S is a non-decreasing sequence in κ with supremum ξ < κ. Then Σξ is just the σ-algebra generated by n∈N Σζ(n) . If we have a sequence hφn in∈N such that φn ∈ Φζ(n) and φn+1 extends φn for every n, then there is a φ ∈ Φξ extending every φn . P P I repeat the ideas of 341G. (i) For E ∈ Σξ , n ∈ N choose gEn such that gEn is a conditional expectation of χE on Σζ(n) ; that is,

R

F

gEn =

R

F

χE = µ(F ∩ E)

for every E ∈ Σζ(n) . Moreover, make these choices in such a way that (α) every gEn is Σζ(n) -measurable and defined everywhere on X (β) gEn = gE 0 n for every n if E4E 0 is negligible. Now limn→∞ gEn exists and is equal to χE almost everywhere, by Lévy’s martingale theorem (275I). (ii) For E ∈ Σξ , k ≥ 1, n ∈ N set Hkn (E) = {x : x ∈ X, gEn (x) ≥ 1 − 2−k } ∈ Σζ(n) ,

˜ kn (E) = φ (Hkn (E)), H n

345C


φE =

T

S k≥1

T n∈N

m≥n

203

˜ km (E). H

˜ kn (∅) is empty; (iii) Every g∅n is zero almost everywhere, every Hkn (∅) is negligible and every H 0 0 so φ∅ = ∅. If E, E ∈ Σξ and E4E is negligible, gEn = gE 0 n for every n, Hnk (E) = Hnk (E 0 ) and ˜ nk (E) = H ˜ nk (E 0 ) for all n, k, and φE = φE 0 . H (iv) If E ⊆ F in Σξ , then gEn ≤ gF n almost everywhere for every n, every Hkn (E) \ Hkn (F ) is ˜ kn (E) ⊆ H ˜ kn (F ) for every n, k, and φE ⊆ φF . negligible, H (v) If E, F ∈ Σξ then χ(E ∩ F ) ≥ χE + χF − 1 a.e. so gE∩F,n ≥ gEn + gF n − 1 a.e. for every n. Accordingly Hk+1,n (E) ∩ Hk+1,n (F ) \ Hkn (E ∩ F ) is negligible, and (because φn is a lower density) ˜ kn (E ∩ F ) ⊇ φ (Hk+1,n (E) ∩ Hk+1,n (F )) = H ˜ k+1,n (E) ∩ H ˜ k+1,n (F ) H n for all k ≥ 1, n ∈ N. Now, if x ∈ φE ∩ φF , then, for any k ≥ 1, there are n1 , n2 ∈ N such that T ˜ k+1,m (E), x ∈ T ˜ x ∈ m≥n1 H m≥n2 Hk+1,m (F ). But this means that x∈

T m≥max(n1 ,n2 )

˜ km (E ∩ F ). H

As k is arbitrary, x ∈ φ(E ∩ F ); as x is arbitrary, φE ∩ φF ⊆ φ(E ∩ F ). We know already from (iv) that φ(E ∩ F ) ⊆ φE ∩ φF , so φ(E ∩ F ) = φE ∩ φF . (vi) If E ∈ Σξ , then gEn → χE a.e., so setting T S T V = k≥1 n∈N m≥n Hkm (E) = {x : lim supn→∞ gEn (x) ≥ 1}, VE 4E is negligible; but φE4V ⊆

S k≥1,n∈N

˜ kn (E) Hkn (E)4H

is also negligible, so φE4E is negligible. Thus φ is a partial lower density with domain Σξ . (vii) If E ∈ Σζ(n) , then E ∈ Σζ(m) for every m ≥ n, so gEm = χE a.e. for every m ≥ n; Hkm (E)4E is negligible for k ≥ 1, m ≥ n; ˜ km (E) = φ E = φ E H m n for k ≥ 1, m ≥ n; and φE = φn E. Thus φ extends every φn . (viii) I have still to check the translation-invariance of φ. If E ∈ Σξ and x ∈ X, consider gn0 , defined by setting gn0 (y) = gEn (y − x) for every y ∈ X, n ∈ N; that is, gn0 is the composition gEn ψ, where ψ(y) = y − x for y ∈ X. (I am not sure whether it is more, or less, confusing to distinguish between the operations of addition and subtraction in X. Of course y − x = y + (−x) = y + x for every y.) Because ψ is a measure space automorphism, and in particular is inverse-measure-preserving, we have

R

g0 = F +x n

R

ψ −1 [F ]

gn0 =

R

F

gEn = µ(E ∩ F )

whenever F ∈ Σζ(n) (235Ic). But because Σζ(n) is itself translation-invariant, we can apply this to F − x to get

R

F

gn0 = µ(E ∩ (F − x)) = µ((E + x) ∩ F )

for every F ∈ Σζ(n) . Moreover, for any α ∈ R, {y : gn0 (y) ≤ α} = {y : gEn (y) ≥ α} + x ∈ Σζ(n)

204

Liftings

345C

for every α, and gn0 is Σζ(n) -measurable. So gn0 is a conditional expectation of χ(E + x) on Σζ(n) , and must be equal almost everywhere to gE+x,n . This means that if we set 0 Hkn = {y : gn0 (y) ≥ 1 − 2−k } = Hkn (E) + x 0 0 for k, n ∈ N, we shall have Hkn ∈ Σζ(n) and Hkn 4Hkn (E + x) will be negligible, so 0 ˜ kn (E + x) = φ (Hkn (E + x)) = φ (Hkn ) H n n

˜ kn (E) + x. = φn (Hkn (E) + x) = φn (Hkn (E)) + x = H Consequently φ(E + x) =

\ [ \

˜ kn (E + x) H

k≥1 n∈N m≥n

=

\ [ \

˜ kn (E) + x = φE + x. H

k≥1 n∈N m≥n

Q As E and x are arbitrary, φ is translation-invariant and belongs to Φξ . Q (d) We are now ready for the proof that there is a translation-invariant lower density on X. P P Build inductively a family hφξ iξ≤κ such that (α) φξ ∈ Φξ for each ξ (β) φξ extends φη whenever η ≤ ξ ≤ κ. The induction starts with Σ0 = {∅, X}, φ0 ∅ = ∅, φ0 X = X. The inductive step to a successor ordinal is dealt with in (b), and the inductive step to a S non-zero ordinal of countable cofinality is dealt with in (c). If ξ ≤ κ has uncountable cofinality, then Σξ = η<ξ Ση , so we can (and must) take φξ to be the unique common extension of all the previous φη . The induction ends with φκ : Σκ → Σ. Note that Σκ is not in general the whole of Σ. But for every E ∈ Σ there is an F ∈ Σκ such that E4F is negligible (254Ff). So we can extend φκ to a function φ defined on the whole of Σ by setting φE = φκ F whenever E ∈ Σ, F ∈ Σκ and µ(E4F ) = 0 (the point being that φκ F = φκ F 0 if F , F 0 ∈ Σκ and µ(E4F ) = µ(E4F 0 ) = 0). It is easy to check that φ is a lower density, and it is translation-invariant because if E ∈ Σ, x ∈ X, F ∈ Σκ and E4F is negligible, then (E + x)4(F + x) = (E4F ) + x is negligible, so φ(E + x) = φκ (F + x) = φκ F + x = φE + x. Q Q (e) The rest of the argument is exactly that of parts (b)-(e) of the proof of 345B; you have to change Rr into X wherever it appears, but otherwise you can use it word for word, interpreting ‘0’ as the identity of the group X, that is, the constant function with value 0. 345D Translation-invariant liftings are of great importance, and I will return to them in §447 with a theorem dramatically generalizing the results above. Here I shall content myself with giving one of their basic properties, set out for the two kinds of translation-invariant lifting we have seen. Proposition Let (X, Σ, µ) be either Lebesgue measure on Rr or the usual measure on {0, 1}I for some set I, and let φ : Σ → Σ be a translation-invariant lifting. Then for any open set G ⊆ X we must have G ⊆ φG ⊆ G, and for any closed set F we must have int F ⊆ φF ⊆ F . proof (a) Suppose that G ⊆ X is open and that x ∈ G. Then there is an open set U such that 0 ∈ U and x+U −U = {x+y −z : y, z ∈ U } ⊆ G. P P (α) If X = Rr , take δ > 0 such that {y : ky −xk ≤ δ} ⊆ G, and set 1 U = {y : ky − xk < 2 δ}. (β) If X = {0, 1}I , then there is a finite set K ⊆ I such that {y : y¹K = x¹K} ⊆ G (3A3K); set U = {y : y(i) = 0 for every i ∈ K}. Q Q It follows that x ∈ φG. P P Consider H = x + U . Then µH = µU > 0 so H ∩ φH 6= ∅. Let y ∈ U be such that x + y ∈ φH. Then x = (x + y) − y ∈ φ(H − y) ⊆ φG

345F


205

because H − y ⊆ x + U − U ⊆ G. Q Q (b) Thus G ⊆ φG for every open set G ⊆ X. But it follows at once that if G is open and F is closed, int F ⊆ φ(int F ) ⊆ φF , G = X \ int(X \ G) ⊇ X \ φ(X \ G) = φG, F = X \ (X \ F ) ⊇ X \ φ(X \ F ) = φF . 345E I remarked in 341Lg that it is undecidable in ordinary set theory whether there is a lifting for Borel measure on R. It is however known that there can be no translation-invariant Borel lifting. The argument depends on the following fact about measurable sets in {0, 1}N . Lemma Let µ be the usual measure on X = {0, 1}N , and E any non-negligible measurable set. Then there are x, x0 ∈ E which differ at exactly one coordinate. proof By 254Fe, there is a set F , determined by coordinates in a finite set, such that µ(E4F ) ≤ 41 µE; we have µF ≥ 43 µE, so µ(E4F ) ≤ 13 µF . Suppose that F is determined by coordinates in {0, . . . , n − 1}. Then the map ψ : X → X, defined by setting (ψx)(n) = 1 − x(n), (ψx)(i) = x(i) for i 6= n, is a measure space automorphism, and µ(ψ −1 [E4F ] ∪ (E4F )) ≤ 2µ(E4F ) < µF . Take any x ∈ F \ ((E4F ) ∪ ψ −1 [E4F ]). Then x0 = ψx differs from x at exactly one coordinate; but also x0 ∈ F , by the choice of n, so both x and x0 belong to E. 345F Proposition Let µ be Borel measure on R, that is, the restriction of Lebesgue measure to the algebra B of Borel sets. Then µ is translation-invariant, but has no translation-invariant lifting. proof (a) To see that µ is translation-invariant all we have to know is that B is translation-invariant and that Lebesgue measure is translation-invariant. I have already cited 134A for the proof that Lebesgue measure is invariant, and B is invariant because G + x is open for every open set G and every x ∈ R. (b) The argument below is most easily expressed in of the geometry of the Cantor set C. Recall T that C is defined as the intersection n∈N Cn of a sequence of closed subsets of [0, 1]; each Cn consists of 2n closed intervals of length 3−n ; Cn+1 is obtained fromPCn by deleting the middle third of each interval of ∞ Cn . Any point of C is uniquely expressible as f (e) = 32 n=0 3−n e(n) for some e ∈ {0, 1}N . (See 134G.) Let ν be the usual measure of {0, 1}N . Because the map e 7→ e(n) : {0, 1}N → {0, 1} is measurable for each n, f : {0, 1}N → R is measurable. We can label the closed intervals constituting Cn as hJz iz∈{0,1}n , taking J∅ to be the unit interval [0, 1] and, for z ∈ {0, 1}n , taking Jza 0 to be the left-hand third of Jz and Jza 1 to be the right-hand third of Jz . (If the notation here seems odd to you, there is an explanation in 3A1H.) For n ∈ N, z ∈ {0, 1}n , let Jz0 be the open interval with the same centre as Jz and twice the length. Then 0 Jz \ Jz consists of two open intervals of length 3−n /2 on either side of Jz ; call the left-hand one Vz and the right-hand one Wz . Thus Vza 1 is the right-hand half of the middle third of Jz , and Wza 0 is the left-hand half of the middle third of Jz . Construct sets G, H ⊆ R as follows. G is to be the union of the intervals Vz where z takes the value 1 an even number of times, together with the intervals Wz where z takes the value 0 an odd number of times; H is to be the union of the intervals Vz where z takes the value 1 an odd number of times, together with the intervals Wz where z takes the value 0 an even number of times. ¤ 1 3£ G and H are open sets. The intervals Vz , Wz between them cover the whole of the interval ¤ 1 3−£ 2 , 2 with the exception of the set C and the countable set of midpoints of the intervals Jz ; so that − 2 , 2 \ (G ∪ H) is negligible. We have to observe that G ∩ H = ∅. P P For each z, Jz0 a 0 and Jz0 a 1 are dist subsets of Jz0 . 0 0 Consequently Jz ∩ Jw is non-empty just when one of z, w extends the other, and we need consider only

206

Liftings

345F

the intersections of the four sets Vz , Wz , Vw , Ww when w is a proper extension of z; say w ∈ {0, 1}n and z = w¹m, where m < n. (α) If in the extension (w(m), . . . , w(n − 1)) both values 0 and 1 appear, Jw0 will be a subset of Jz , and certainly the four sets will all be dist. (β) If w(i) = 0 for m ≤ i < n, then Ww ⊆ Jz is dist from the rest, while Vw ⊆ Vz ; but z and w take the value 1 the same number of times, so Vw is assigned to G iff Vz is, and otherwise both are assigned to H. (γ) Similarly, if w(i) = 1 for m ≤ i < n, Vw ⊆ Jz , Ww ⊆ Wz and z, w take the value 0 the same number of times, so Wz and Ww are assigned to the same set. Q Q The following diagram may help you to see what is supposed to be happening:

G

H

V 00

J00

W00 V 01 J01

V0 V

G

J0

H

W01

G

V 10

W0

J10

V1

W10 V 11 J11

W11

J1

J

0

H

W1 1

W

The assignment rule can be restated as follows: V = V∅ is assigned to G, W = W∅ is assigned to H; Vza 0 is assigned to the same set as Vz , and Vza 1 to the other; Wza 1 is assigned to the same set as Wz , and Wza 0 to the other. (c) Now take any n ∈ N and z ∈ {0, 1}n . Consider the two open intervals I0 = Jz0 a 0 , I1 = Jz0 a 1 . These are both of length γ = 2 · 3−n−1 and abut at the centre of Jz , so I1 is just the translate I0 + γ. I claim that I1 ∩ H = (I0 ∩ G) + γ. P P Let A be the set S m a m>n {w : w ∈ {0, 1} , w extends z 0}, and for w ∈ A let w0 be the finite sequence obtained from w by changing w(n) = 0 into w0 (n) = 1 but leaving the other values of w unaltered. Then Vw0 = Vw + γ and Ww0 = Ww + γ for every w ∈ A. Now [ I0 ∩ G = {Vw : w ∈ A, w takes the value 1 an even number of times} [ ∪ {Ww : w ∈ A, w takes the value 0 an odd number of times}, so (I0 ∩ G) + γ =

[

{Vw0 : w ∈ A, w takes the value 1 an even number of times} [ ∪ {Ww0 : w ∈ A, w takes the value 0 an odd number of times} [ = {Vw0 : w ∈ A, w0 takes the value 1 an odd number of times} [ ∪ {Ww0 : w ∈ A, w0 takes the value 0 an even number of times} = I1 ∩ H. Q Q

(d) ?? Now suppose, if possible, that φ : B → B is a translation-invariant lifting. Note first that U ⊆ φU for every open U ⊆ R. P P The argument is exactly that of 345D as applied to R = R1 . Q Q Consequently ¤ £ J∅0 = − 12 , 32 ⊆ φJ∅0 . But as J∅0 \ (G ∪ H) is negligible,

¤ £ C ⊆ − 12 , 32 ⊆ φG ∪ φH.

345Xf


207

Consider the sets E = f −1 [φG], F = {0, 1}N \ E = f −1 [φH]. Because f is measurable and φG, φH are Borel sets, E and F are measurable subsets of {0, 1}N , and at least one of them has positive measure. There must therefore be e, e0 ∈ {0, 1}N , differing at exactly one coordinate, such that either both belong to E or both belong to F (345E). Let us suppose that n is such that e(n) = 0, e0 (n) = 1 and e(i) = e0 (i) for i 6= n. Set z = e¹n = e0 ¹n. Then f (e) belongs to the open interval I0 = Jz0 a 0 , so f (e) ∈ φI0 and f (e) ∈ φG iff f (e) ∈ φ(I0 ∩ G). But now f (e0 ) = f (e) + 2 · 3−n−1 ∈ I1 = Jz0 a 1 , so

e ∈ E ⇐⇒ f (e) ∈ φG ⇐⇒ f (e) ∈ φ(I0 ∩ G) ⇐⇒ f (e0 ) ∈ φ((I0 ∩ G) + 2 · 3−n−1 ) (because φ is translation-invariant) ⇐⇒ f (e0 ) ∈ φ(I1 ∩ H) (by (c) above) ⇐⇒ f (e0 ) ∈ φH (because f (e0 ) ∈ I1 ⊆ φI1 ) ⇐⇒ e0 ∈ F. But this contradicts the choice of e. X X Thus there is no translation-invariant lifting of µ. Remark This result is due to Johnson 80; the proof here follows Talagrand 82b. For references to various generalizations see Burke 93, §3. 345X Basic exercises (a) In 345Ab I wrote ‘It follows at once that the map y 7→ y + x : X → X is a measure space automorphism’. Write the details out in full, using 254G or otherwise. (b) Let S 1 be the unit circle in R2 , and let µ be one-dimensional Hausdorff measure on S 1 (§§264-265). Show that µ is translation-invariant, if S 1 is given its usual group operation corresponding to complex multiplication (255M), and that it has a translation-invariant lifting φ. (Hint: Identifying S 1 with ]−π, π] S 0 with the group operation +2π , show that we can set φE = ]−π, π] ∩ φ ( n∈Z E + 2πn), where φ0 is any translation-invariant lifting for Lebesgue measure.) > (c) Show that there is no lifting φ of Lebesgue measure on R which is ‘symmetric’ in the sense that φ(−E) = −φE for every measurable set E, writing −E = {−x : x ∈ E}. (Hint: can 0 belong to φ([0, ∞[)?) > (d) Let µ be Lebesgue measure on X = R\{0}. Show that there is a lifting φ of µ such that φ(xE) = xφE for every x ∈ X and every measurable E ⊆ X, writing xE = {xy : y ∈ E}. (e) Let µ be the usual measure on X = {0, 1}I , for some set I, Σ its domain, and (A, µ ¯) its measure algebra. (i) Show that we can define πx (a) = a + x, for a ∈ A and x ∈ X, by the formula E • + x = (E + x)• ; and that x 7→ πx is a group homomorphism from X to the group of measure-preserving automorphisms of A. (ii) Define Σξ as in the proof of 345C, and set Aξ = {E • : E ∈ Σξ }. Say that a partial lifting θ : Aξ → Σ is translation-invariant if θ(a + x) = θa + x for every a ∈ Aξ and x ∈ X. Show that any such partial lifting can be extended to a translation-invariant partial lifting on Aξ+1 . (iii) Write out a proof of 345C in the language of 341F-341H. > (f ) Let φ be a lower density for Lebesgue measure on Rr which is translation-invariant in the sense that φ(E + x) = φE + x for every x ∈ Rr and every measurable set E. Show that φG ⊆ G for every open set G ⊆ Rr .

208

Liftings

345Xg

(g) Let µ be 1-dimensional Hausdorff measure on S 1 , as in 345Xb. Show that there is no translationinvariant lifting φ of µ such that φE is a Borel set for every E ∈ dom µ. 345Y Further exercises (a) Let (X, Σ, µ) be a complete measure space, and suppose that X has a group operation (x, y) 7→ xy (not necessarily abelian!) such that µ is left-translation-invariant, in the sense that xE = {xy : y ∈ E} ∈ Σ and µ(xE) = µE whenever E ∈ Σ and x ∈ X. Suppose that φ : Σ → Σ is a lower density which is left-translation-invariant in the sense that φ(xE) = x(φE) for every E ∈ Σ, x ∈ X. Show that there is a left-translation-invariant lifting φ : Σ → Σ such that φE ⊆ φE for every E ∈ Σ. (b) Write Σ for the σ-algebra of Lebesgue measurable subsets of R, and L0 (Σ) for the linear space of Σ-measurable functions from R to itself. Show that there is a linear operator T : L0 (µ) → L0 (Σ) such that (α) (T u)• = u for every u ∈ L0 (µ) (β) supx∈R |(T u)(x)| = kuk∞ for every u ∈ L∞ (µ) (γ) T u ≥ 0 whenever u ∈ L∞ (µ) and u ≥ 0 (δ) T is translation-invariant in the sense that T (Sx f )• = Sx T f • for every x ∈ R and f ∈ L0 (Σ), where (Sx f )(y) = f (x + y) for f ∈ L0 (Σ) and x, y ∈ R (²) T is reflection-invariant in the sense that T (Rf )• = RT f • for every f ∈ L0 (Σ), where (Rf )(x) = f (−x) for f ∈ L0 (Σ) and x ∈ R. (Hint: for f ∈ L0 (Σ), set p(f • ) = inf{α : α ∈ [0, ∞], limδ↓0

1 µ{x 2δ

: |x| ≤ δ, |f (x)| > α} = 0}.

Set V = {u : u ∈ L0 (µ), p(u) < ∞} and show that V is a linear subspace of L0 (µ) and that p¹V is a seminorm. Let h0 : V → R be a linear functional such that h0 (χR)• = 1 and h0 (u) ≤ p(u) for every u ∈ V . Extend h0 arbitrarily to a linear functional h1 : L0 (µ) → R; set h(f • ) = 12 (h1 (f • ) + h1 (Rf )• ). Set (T f • )(x) = h(S−x f )• . You will need 231C.) Show that there must be a u ∈ L1 (µ) such that u ≥ 0 but T u 6≥ 0. (c) Show that there is no translation-invariant lifting φ of the usual measure on {0, 1}N such that φE is a Borel set for every measurable set E. 345 Notes and comments I have taken a great deal of care over the concept of ‘translation-invariance’. I hope that you are already a little impatient with some of the details as I have written them out; but while it is very easy to guess at the structure of such arguments as part (e) of the proof of 345B, or (b-iii) and (c-viii) in the proof of 345C, I am not sure that one can always be certain of guessing correctly. A fair test of your intuition will be how quickly you can generate the formulae appropriate to a non-abelian group operation, as in 345Ya. Part (b) of the proof of 345C is based on the same idea as the proof of 341F. There is a useful simplification because the set Eξ in 345C, corresponding to the set E of the proof of 341F, is independent of the algebra Σξ in a very strong sense, so that the expression of an element of Σξ+1 in the form (F ∩ Eξ ) ∪ (G \ Eξ ) is unique. Interpreted in the of 341F, we have w = v = 1, so that the formula ¡ ¢ ¡ ¢ θ1 ((a ∩ e) ∪ (b \ e)) = θ((a ∩ v) ∪ (b \ v)) ∩ E ∪ θ((a \ w) ∪ (b ∩ w)) \ E used there becomes θ1 ((a ∩ e) ∪ (b \ e)) = (θa ∩ E) ∪ (θb \ E), matching the formula for φ1 in the proof of 345C. The results of this section are satisfying and natural; they have obvious generalizations, many of which are true. The most important measure spaces come equipped with a variety of automorphisms, and we can always ask which of these can be preserved by a lifting. The answers are not always obvious; I offer 345Xc and 346Xb as warnings, and 345Xd as an encouragement. 345Yb is striking (I have made it as striking as I can), but slightly off the most natural target; the sting is in the last sentence (see 341Ye).

346D

Consistent liftings

209

346 Consistent liftings I turn now to a different type of condition Q which we should naturally prefer our liftings to satisfy. If we have a product measure µ on a product X = i∈I Xi of probability spaces, then we can look for liftings φ which ‘respect coordinates’, that is, are compatible with the product structure in the sense that they factor through subproducts (346A). There seem to be obstacles in the way of the natural conjecture (346Za), and I give the partial results which are known. For Maharam-type-homogeneous spaces Xi , there is always a lifting which respects coordinates (346E), and indeed the translation-invariant liftings of §345 on {0, 1}I already have this property (346C). There is always a lower density on the product which respects coordinates, and we can ask for a little more (346G); using the full strength of 346G, we can enlarge this lower density to a lifting which respects single coordinates and initial segments of a well-ordered product (346H). In the case in which all the factors are copies of each other, we can arrange for the induced liftings on the factors to be copies also (346I, 346J, 346Yd). I end the section with an important fact about Stone spaces which is relevant here (346K-346L). 346A Definition Let h(Xi , Σi , µi )ii∈I be a family of probability spaces, with product (X, Σ, µ). I will say that a lifting φ : Σ → Σ respects coordinates if φE is determined by coordinates in J whenever E ∈ Σ is determined by coordinates in J ⊆ I. Remark Recall that a set E ⊆ X is ‘determined by coordinates in Q J’ if x0 ∈ E whenever x ∈ E, x0 ∈ X −1 0 and x ¹J = x¹J; that is, if E is expressible as πJ [F ] for some F ⊆ i∈J Xi , where πJ (x) = x¹J for every x ∈ X; that is, if E = πJ−1 [πJ [E]]. See 254M. Recall also that in this case, Q if E is measurable for the product measure on X, then πJ [E] is measurable for the product measure on i∈J Xi (254Ob). 346B Proposition Let h(Xi , Σi , µi )ii∈I be a family of probability spaces, with product (Z, Λ, λ). For J ⊆ I let (ZJ , ΛJ , λJ ) be the product of h(Xi , Σi , µi )ii∈J , and πJ : Z → ZJ the canonical map. Let φ : Λ → Λ be a lifting. If J ⊆ I is such that φW is determined by coordinates in J whenever W ∈ Λ is determined by coordinates in J, then φ induces a lifting φJ : ΛJ → ΛJ defined by the formula πJ−1 [φJ E] = φ(πJ−1 [E]) for every E ∈ ΛJ . proof If E ∈ ΛJ , then π −1 [E] and φ(πJ−1 [E]) are determined by coordinates in J, so φ(πJ−1 [E]) is of the form πJ−1 [F ] for some F ∈ ΛJ ; this defines φJ : ΛJ → ΛJ . It is now easy to see that φJ is a lifting. Remark Of course we frequently wish to use this result with a singleton set J = {j}. In this case we must that (ZJ , ΣJ , λJ ) corresponds to the completion of the probability space (Xj , Σj , µj ). 346C Theorem Let I be any set, and µ the usual measure on X = {0, 1}I . Then any translationinvariant lifting of µ respects coordinates. proof Suppose that E ⊆ X is a measurable set determined by coordinates in J ⊆ I; take x ∈ φE and x0 ∈ X such that x0 ¹J = x¹J. Set y = x0 − x; then y(i) = 0 for i ∈ J, so that E + y = y. Now x0 = x + y ∈ φE + y = φ(E + y) = φE because φ is translation-invariant. As x, x0 are arbitrary, φE is determined by coordinates in J. As E and J are arbitrary, φ respects coordinates. 346D

I describe a standard method of constructing liftings from other liftings.

Lemma Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras A, B; suppose that f : X → Y is a (Σ, T)-measurable function inducing an isomorphism F • 7→ f −1 [F ]• : B → A. Then if φ : T → T is a lifting for ν, there is a corresponding lifting φ0 : Σ → Σ given by the formula φ0 E = f −1 [φF ] whenever µ(E4f −1 [F ]) = 0. proof If we say that π : B → A is the isomorphism induced by f , then

210

Liftings

346D

φ0 E = f −1 [θ(π −1 E • )], where θ : B → T is the lifting corresponding to φ : T → T. Since θ, π −1 and F 7→ f −1 [F ] are all Boolean homomorphisms, so is φ0 , and it is easy to check that (φ0 E)• = E • for every E ∈ Σ and that φ0 E = ∅ if µE = 0. Remark Compare the construction in 341P. 346E Theorem Let h(Xi , Σi , µi )ii∈I be a family of Maharam-type-homogeneous probability spaces, with product (X, Σ, µ). Then there is a lifting of µ which respects coordinates. proof (a) Replacing each µi by its completion does not change µ (254I), so we may suppose that all the µi are complete. In this case there is for each i an isomorphism between the measure algebra (Ai , µ ¯i ) of µi and the measure algebra (Bi , ν¯i ) of some {0, 1}Ji with its usual measure νi (331L). We may suppose that the sets Ji are dist. Each νi is compact (342Jd), so the isomorphisms are represented by inverse-measurepreserving functions fi : Xi → {0, 1}Ji (343Ca). S Set K = i∈I Ji , and let ν be the usual measure on Y = {0, 1}K , T its domain. We have a natural Q bijection between i∈I {0, 1}Ji and Y , so we obtain a function f : X → Y ; literally speaking, f (x)(j) = fi (x(i))(j) for i ∈ I, j ∈ Ji , x ∈ X. (b) Now f is inverse-measure-preserving and induces an isomorphism between the measure algebras A, B of µ, ν. P P(i) If L ⊆ K is finite and z ∈ {0, 1}L , then, setting Li = LcapJi for i ∈ I, Y µ{x : x ∈ X, f (x)¹L = z} = µ( {w : w ∈ Xi , fi (w)¹Li = z¹Li }) i∈I

=

Y

µi {w : w ∈ Xi , fi (w)¹Li = z¹Li }

i∈I

=

Y

νi {v : v ∈ {0, 1}Ji , v¹Li = z¹Li }

i∈I

(because every fi is inverse-measure-preserving) Y = 2−#(Li ) = 2−#(L) = ν{y : y ∈ Y, y¹L = z}. i∈I

So µf −1 [C] = νC for every basic cylinder set C ⊆ Y . By 254G, f is inverse-measure-preserving. (ii) Accordingly f induces a measure-preserving homomorphism π : B → A. To see that π is surjective, consider Λ0 = {E : E is Σ-measurable, E • ∈ π[B]}. Because π[B] is a closed subalgebra of A (324Kb), Λ0 is a σ-subalgebra of the domain Λ of µ, and of course it contains all µ-negligible sets. If i ∈ J and G ∈ Σi , then there is an H ⊆ {0, 1}Ji such that G4fi−1 [H] is µi -negligible. Now if E = {x : x ∈ X, x(i) ∈ G} and F = {y : y ∈ Y, y¹Ji ∈ H}, E4f −1 [F ] = {x : x(i) ∈ G4fi−1 [H]} N is µ-negligible, and E ∈ Λ0 . But this means that Λ0 ⊇ c i∈I Σi , and must therefore be the whole of Λ (254Ff). Q Q (c) By 345C, there is a translation-invariant lifting φ for ν; by 346C, this respects coordinates. By 346D, we have a corresponding lifting φ0 for µ such that φ0 f −1 [F ] = f −1 [φF ] for every F ∈ T. Now suppose that E ∈ Λ is determined by coordinates in L ⊆ I. Then there is an E 0 belonging to the σ-algebra Λ0L generated by

346G

Consistent liftings

211

{{x : x(i) ∈ G} : i ∈ L, G ∈ Σi }

S such that µ(E4E ) = 0 (254Ob). Write TL for the family of sets in T determined by coordinates in i∈L Ji . Then, just as in (b-ii), every member of Λ0L differs by a negligible set from some set of the form f −1 [F ] with F ∈ TL . So there is an F ∈ TL such that E4f −1 [F ] is µ-negligible. Consequently 0

φ0 E = φ0 f −1 [F ] = f −1 [φF ]. S But φ respects coordinates, so φF is determined by coordinates in i∈L Ji . It follows at once that f −1 [φF ] is determined by coordinates in L; that is, that φ0 E is determined by coordinates in L. As E and L are arbitrary, φ0 respects coordinates, and witnesses the truth of the theorem. 346F It seems to be unknown whether 346E is true of arbitrary probability spaces (346Za); I give some partial results in this direction. The following general method of constructing lower densities will be useful. Lemma Let (X, Σ, µ) and (Y, T, ν) be complete probability spaces, with product (X × Y, Λ, λ). If φ : Λ → Λ is a lower density, then we have a lower density φ1 : Σ → Σ defined by saying that φ1 E = {x : x ∈ X, {y : (x, y) ∈ φ(E × Y )} is conegligible in Y } for every E ∈ Σ. proof For E ∈ Σ, (E × Y )4φ(E × Y ) is negligible, so that Hx = {y : (x, y) ∈ (E × Y )4φ(E × Y )} is ν-negligible for almost every x ∈ X (252D). Now E4φ1 E = {x : Hx is not negligible} is negligible, so φ1 E ∈ Σ. If E, F ∈ Σ, then φ((E ∩ F ) × Y ) = φ((E × Y ) ∩ (F × Y )) = φ(E × Y ) ∩ φ(F × Y ), so that {y : (x, y) ∈ φ((E ∩ F ) × Y )} = {y : (x, y) ∈ φ(E × Y )} ∩ {y : (x, y) ∈ φ(F × Y )} is conegligible iff both {y : (x, y) ∈ φ(E × Y )} and {y : (x, y) ∈ φ(F × Y )} are conegligible, and φ1 (E ∩ F ) = φ1 E ∩ φ1 F . The rest is easy. Of course φ(∅ × Y ) = ∅ so φ1 ∅ = ∅. If E, F ∈ Σ and E4F is negligible, then (E × Y )4(F × Y ) is negligible, φ(E × Y ) = φ(F × Y ) and φ1 E = φ1 F . So φ1 is a lower density, as claimed. 346G Theorem Let h(Xi , Σi , µi )ii∈I be a family of probability spaces with product (X, Σ, µ). For J ⊆ I let ΣJ be the set of of Σ which are determined by coordinates in J. Then there is a lower density φ : Σ → Σ such that (i) whenever J ⊆ I and E ∈ ΣJ then φE ∈ ΣJ , (ii) whenever J, K ⊆ I are dist, E ∈ ΣJ and F ∈ ΣK then φ(E ∪ F ) = φE ∪ φF . Q proof For each i ∈ I, set Yi = XiN , with the product measure Q νi ; set Y = i∈I Yi , with its product measure ν; set Zi = Xi × Yi , with its product measure λi , Q and Z = Q i∈I Zi , with its product measure λ. Then the Q natural identification of Z = i∈I Xi × Yi with i∈I Xi × i∈I Yi = X × Y makes λ correspond to the product of µ and ν (254N). Each (Zi , λi ) can be identified with an infinite power of (Xi , µi ), and is therefore Maharam-type-homogeneous (334E). Consequently there is a lifting φ : Λ → Λ which respects coordinates (346E). Regarding (Z, λ) as the product of (X, µ) and (Y, ν), we see that φ induces a lower density φ : Σ → Σ by the formula of 346F. Q If J ⊆ I and E ∈ Σ is determined by coordinates in J, then E × Y (regarded as a subset of i∈I Zi ) is determined by coordinates in J, so φ(E × Y ) also is. Now suppose that x ∈ φE, x0 ∈ X and x¹J = x0 ¹J. Then for any y ∈ Y , (x¹J, y¹J) = (x0 ¹J, y¹J), so (x, y) ∈ φ(E × Y ) iff (x0 , y) ∈ φ(E × Y ). Thus {y : (x0 , y) ∈ φ(E × Y )} = {y : (x, y) ∈ φ(E × Y )} is conegligible in Y , and x0 ∈ φE. This shows that φE is determined by coordinates in J.

212

Liftings

346G

Now suppose that J and K are dist subsets of I, that E, F ∈ Σ are determined by coordinates in J, K respectively, and that x ∈ / φE ∪ φF . Then A = {y : (x, y) ∈ / φ(E × Y )} and B = {y : (x, y) ∈ / φ(F × Y )} are non-negligible. As noted just above, φ(E × Y ) is determined by coordinates in J, so A is determined Q by coordinates in J, and can be expressed as {y : y¹J ∈ A0 }, where A0 ⊆ YJ = i∈J Yi . Because y 7→ y¹J : Y → YJ is inverse-measure-preserving, A0 cannot be negligible in YJ . Similarly, B can be expressed as {y : y¹K ∈ B 0 } for some non-negligible B 0 ⊆ YK . By 251R/251Wm, A0 × B 0 × YI\(J∪K) , regarded as a subset of Y , is non-negligible, that is, C = {y : y ∈ Y, y¹J ∈ A0 , y¹K ∈ B 0 } is non-negligible. But C = A ∩ B = {y : (x, y) ∈ / φ(E × Y ) ∪ φ(F × Y )} = {y : (x, y) ∈ / φ((E ∪ F ) × Y }. So x ∈ / φ(E ∪ F ). As x is arbitrary, φ(E ∪ F ) ⊆ φE ∪ φF ; but of course φE ∪ φF ⊆ φ(E ∪ F ), because φ is a lower density, so that φ(E ∪ F ) = φE ∪ φF , as required. Remark See Macheras Musial & Strauss p99 for an alternative proof. 346H Theorem Let ζ be an ordinal, and h(Xξ , Σξ , µξ )iξ<ζ any family of probability spaces, with product (Z, Λ, λ). For J ⊆ ζ let ΛJ be the set of those W ∈ Λ which are determined by coordinates in J. Then there is a lifting φ : Λ → Λ such that φW ∈ ΛJ whenever W ∈ ΛJ and J is either a singleton subset of ζ or an initial segment of ζ. proof (a) Let P be the set of all lower densities φ : Λ → Λ such that, for every ξ < ζ, (i) whenever E ∈ Λξ then φE ∈ Λξ (ii) whenever E ∈ Λ{ξ} then φE ∈ Λ{ξ} (iii) whenever E ∈ Λξ and F ∈ Λζ\ξ then φ(E ∪ F ) = φE ∪ φF . By 346G, P is not empty. Order P by saying that φ ≤ φ0 if φE ⊆ φ0 E for every E ∈ Λ; then P is a partially ordered set. Note that if φ ∈ P then φZ = Z (because Λ0 = {∅, Z}). (b) Any non-empty totally ordered subset Q of P has an upper bound in P . P P Define φ∗ : Λ → PX by S ∗ setting φ E = φ∈Q φE for every E ∈ Λ. (i) S φ∗ ∅ = φ∈Q ∅ = ∅. (ii) If E, F ∈ Λ and λ(E4F ) = 0 then φE = φF for every φ ∈ Q so φ∗ E = φ∗ F . (iii) If E, F ∈ Λ and E ⊆ F then φE ⊆ φF for every φ ∈ Q so φ∗ E ⊆ φ∗ F . (iv) If E, F ∈ Λ and x ∈ φ∗ E ∩ φ∗ F , then there are φ1 , φ2 ∈ Q such that x ∈ φ1 E ∩ φ2 F ; now either φ1 ≤ φ2 or φ2 ≤ φ1 , so that x ∈ (φ1 E ∩ φ1 F ) ∪ (φ2 E ∩ φ2 F ) = φ1 (E ∩ F ) ∪ φ2 (E ∩ F ) ⊆ φ∗ (E ∩ F ). Accordingly φ∗ E ∩ φ∗ F ⊆ φ∗ (E ∩ F ) and φ∗ E ∩ φ∗ F = φ∗ (E ∩ F ). (v) Taking any φ0 ∈ Q, we have φ0 E ⊆ φ∗ E for every E ∈ Λ, so (because λ is complete) φ∗ is a lower density, by 341Ib. (vi) Now suppose that J ⊆ I is either a singleton {ξ} or an initial segment ξ, and that E ∈ ΛJ . Then φE is determined by coordinates in J for every φ ∈ Q, so φ∗ E is determined by coordinates in J. (vii) Finally, suppose that ξ < ζ and that E ∈ Λξ , F ∈ Λζ\ξ . If x ∈ φ∗ (E ∪ F ) then there is a φ ∈ Q such that x ∈ φ(E ∪ F ) = φE ∪ φF ⊆ φ∗ E ∪ φ∗ F . So φ∗ (E ∪ F ) ⊆ φ∗ E ∪ φ∗ F and (using (iii) again) φ∗ (E ∪ F ) = φ∗ E ∪ φ∗ F . Thus φ∗ belongs to P and is an upper bound for Q in P . Q Q ˜ By Zorn’s Lemma, P has a maximal element φ. ˜ ∪ φ(Z ˜ \ H)), (c) For any H ∈ Λ we may define a function φH as follows. Set AH = Z \ (φH ˜ ∪ (AH ∩ φ(H ˜ ∪ E)) φH E = φE ˜ and (Z \H)4φ(Z ˜ \H) are both negligible, AH for E ∈ Λ. Then φH is a lower density. P P (i) Because H4φH • • • ˜ ˜ = ∅, is negligible and φH E is measurable and (φH E) = (φE) = E for every E ∈ Λ. (ii) Because AH ∩ φH ˜ = φF ˜ and φ(E ˜ ∪ H) = φ(F ˜ ∪ H), so φ E = φ F . φH ∅ = ∅. (iii) If E, F ∈ Λ and λ(E4F ) = 0 then φE H H ˜ and φ(E ˜ ∪ H) ⊆ φ(F ˜ ∪ H), so φ E ⊆ φ F . (v) If E, F ∈ Λ ˜ ⊆ φF (iv) If E, F ∈ Λ and E ⊆ F then φE H H and x ∈ φH E ∩ φH F , then

346H

Consistent liftings

213

(α) if x ∈ / AH , ˜ ∩ φF ˜ = φ(E ˜ ∩ F ) ⊆ φ (E ∩ F ), x ∈ φE H (β) if x ∈ AH , ˜ ∪ H) ∩ φ(F ˜ ∪ H) = φ((E ˜ x ∈ φ(E ∩ F ) ∪ H) ⊆ φH (E ∩ F ). Thus φH E ∩ φH F ⊆ φH (E ∩ F ) and φH E ∩ φH F = φH (E ∩ F ). Q Q (d) It is worth noting the following. ˜ ∪ H) = φE ˜ ∪ φH ˜ then φ E = φE. ˜ P (i) If E, H ∈ Λ and φ(E P We have H ˜ ∪ H)) = φE ˜ ∪ (AH ∩ φE) ˜ ∪ (AH ∩ φH) ˜ ˜ ˜ ∪ (AH ∩ φ(E = φE φH E = φE ˜ = ∅. Q because AH ∩ φH Q But

˜ ∪ φ(Z ˜ \ H) = Z. P ˜ we must have φ = φ. ˜ (ii) If H ∈ Λ and φH ∈ P then φH P By the maximality of φ, H ˜ \ H), AH = φH (Z \ H) \ φ(Z

˜ ∪ φ(Z ˜ \ H) = Z. Q so AH = ∅, that is, φH Q ˜ \ E) = Z, then φ(E ˜ ∪ F ) = φE ˜ ∪ φF ˜ . P ˜ ∪ φ(Z P (iii) If E, F ∈ Λ and φE ˜ ∪ F ) \ φE ˜ = φ(E ˜ ∪ F ) ∩ φ(Z ˜ \ E) = φ((E ˜ ˜ \ E) ⊆ φF ˜ , φ(E ∪ F ) ∩ (Z \ E)) = φ(F ˜ ∪ F ) ⊆ φE ˜ ∪ φF ˜ ; as the reverse inclusion is true for all E and F , we have the result. Q so φ(E Q (e) If ξ < ζ and H ∈ Λ{ξ} , then φH ∈ P . P P(i) If J ⊆ I is either a singleton or an inital segment, and E ∈ ΛJ , then ˜ and φ(E ˜ ∪ H) and AH all belong to ΛJ , so φ E ∈ ΛJ . (α) if ξ ∈ J, E ∪ H and φE H ˜ ˜ ˜ (β) If ξ ∈ / J, φ(E ∪ H) = φE ∪ φH, because there is some η such that J ⊆ η and {ξ} ⊆ ζ \ η); so ˜ ∈ ΛJ by (d-i). φH E = φE (ii) If η < ζ and E ∈ Λη , F ∈ Λζ\η , then ˜ ∪ F ∪ H) = φ(E ˜ ∪ H) ∪ φF ˜ , and if ξ < η, E ∪ H ∈ Λη so φ(E ˜ ∪ F ) ∪ (AH ∩ φ(E ˜ ∪ F ∪ H)) φH (E ∪ F ) = φ(E ˜ ∪ φF ˜ ∪ (AH ∩ φ(E ˜ ∪ H)) ∪ (AH ∩ φF ˜ ) ⊆ φ E ∪ φ F; = φE H H ˜ ∪ F ∪ H) = φ(E) ˜ ˜ ∪ H), and if η ≤ ξ, F ∪ H ∈ Λζ\η so φ(E ∪ φ(F ˜ ∪ F ) ∪ (AH ∩ φ(E ˜ ∪ F ∪ H)) φH (E ∪ F ) = φ(E ˜ ∪ φF ˜ ∪ (AH ∩ φE)) ˜ ˜ ∪ H)) ⊆ φ E ∪ φ F ; = φE ∪ (AH ∩ φ(F H H accordingly φH (E ∪ F ) = φH E ∪ φH F . Q Q By (d-ii) we have ˜ ∪ φ(Z ˜ \ H) = Z φH whenever ξ < ζ and H ∈ Λ{ξ} . ˜ is either ∅ or (f ) If ξ ≤ ζ and H ∈ Λξ , then φH ∈ P . P P Induce on ξ. For ξ = 0, H ∈ Λ0 = {∅, Z} so φH Z and the result is trivial. For the inductive step to ξ ≤ ζ, we have the following. (i) If η < ζ and E ∈ Λη , then ˜ and φ(E ˜ ∪ H) and AH all belong to Λη , so φ E ∈ Λη . (α) if ξ ≤ η, E ∪ H and φE H

214

Liftings

346H

˜ = Z \ φ(Z ˜ \ E) and φ(E ˜ ∪ H) = φE ˜ ∪ φH, ˜ (β) if η < ξ, then, by the inductive hypothesis, φE ∈ P , φE ˜ ∈ Λη by (d-i). by (d-ii) and (d-iii) above; so φH E = φE ˜ ∪ φ(Z ˜ \ E) = Z, so that φ(E ˜ ∪ H) = φE ˜ ∪ φH, ˜ by (d-iii), (ii) If η < ζ and E ∈ Λ{η} , then, by (e), φE ˜ ∈ Λ{η} , by (d-i). and φH E = φE (iii) If η < ζ and E ∈ Λη , F ∈ Λζ\η , then ˜ ∪ F ∪ H) = φ(E ˜ ∪ H) ∪ φF ˜ , and (α) if ξ ≤ η, then E ∪ H ∈ Λη and F ∈ Λζ\η , so that φ(E ˜ ∪ F ∪ H)) ˜ ∪ F ) ∪ (AH ∩ φ(E φH (E ∪ F ) = φ(E ˜ ∪ φF ˜ ∪ (AH ∩ φ(E ˜ ∪ H)) ∪ (AH ∩ φF ˜ ) ⊆ φ E ∪ φ F, = φE H H as in (e-ii) above, and accordingly φH (E ∪ F ) = φH E ∪ φH F . ˜ ∪ F ∪ H) = φE ˜ ∪ φ(F ˜ ∪ H), (β) If η < ξ then, as in (ii), using the inductive hypothesis, we have φ(E and (just as in (α)) we get φH (E ∪ F ) = φH E ∪ φH F . Thus φH ∈ P and the induction continues. Q Q (g) But the case ξ = ζ of (f) just tells us that ˜ ∪ φ(Z ˜ \ H) = Z φH for every H ∈ Λ. This means that φ˜ is actually a lifting (since it preserves intersections and complements). And the definition of P is just what is needed to ensure that it is a lifting of the right type. Remark This result is due to Macheras & Strauss 96b. 346I Theorem Let (X, Σ, µ) be a complete probability space. For any set I, write λI for the product measure on X I , ΛI for its domain and πIi (x) = x(i) for x ∈ X I , i ∈ I. Then there is a lifting ψ : Σ → Σ −1 −1 such that for every set I there is a lifting φ : ΛI → ΛI such that φ(πIi [E]) = πIi [ψE] for every E ∈ Σ, i ∈ I. proof ?? Suppose, if possible, otherwise. Let Ψ be the set of all liftings for µ. We are supposing that for every ψ ∈ Ψ there is a set Iψ for which there is no lifting for λIψ consistent with ψ in the sense above. Let κ be a cardinal greater than max(ω, #(Ψ), supψ∈Ψ #(Iψ )). Let φ0 : Λκ → Λκ be a lifting satisfying the conditions of 346H. 346B tells us −1 −1 that for every ξ < κ we have a lifting ψ for µ defined by the formula πκξ [ψE] = φ0 (πκξ [E]). For ψ ∈ Ψ set S

−1 −1 Kψ = {ξ : ξ < κ, φ0 (πκξ [E]) = πκξ [ψE] for every E ∈ Σ}.

Then ψ∈Ψ Kψ = κ, so κ ≤ max(ω, #(Ψ), supψ∈Ψ #(Kψ )) and there is some ψ ∈ Ψ such that #(Kψ ) > #(Iψ ). Take I ⊆ Kψ such that #(I) = #(Iψ ). We may regard X κ as X I × X κ\I , and in this form we can use the method of 346F to obtain a lower density φ : ΛI → ΛI from φ0 : Λκ → Λκ . Now −1 −1 φ(πIξ [E]) = πIξ [ψE] for every E ∈ Σ, ξ ∈ I. −1 −1 −1 −1 P P The point is that πIξ [E] × X κ\I corresponds to πκξ [E] ⊆ X κ , while φ0 (πκξ [E]) = πκξ [ψE] can be −1 −1 −1 κ\I [ψE]. identified with πIξ [ψE]×X . Now the construction of 346F obviously makes φ(πIξ [E]) equal to πIξ Q Q By 341Jb, there is a lifting φ : ΛI → ΛI such that φW ⊇ φW for every W ∈ ΛI . But now we must have −1 −1 −1 πIξ [ψE] = φ(πIξ [E]) ⊆ φ(πIξ [E]) −1 −1 = X I \ φ(πIξ [X \ E]) ⊆ X I \ φ(πIξ [X \ E]) −1 −1 −1 = X I \ πIξ [ψ(X \ E)] = X I \ πIξ [X \ ψE] = πIξ [ψE] −1 −1 and φ(πIξ [E]) = πIξ [ψE] for every E ∈ Σ, ξ ∈ I. But since #(I) = #(Iψ ), this must be impossible, by the choice of Iψ . X X This contradiction proves the theorem.

346K

Consistent liftings

215

346J Consistent liftings Let (X, Σ, µ) be a measure space. A lifting ψ : Σ → Σ is consistent if for every n ≥ 1 there is a lifting φn of the product measure on X n such that φn (E1 ×. . .×En ) = ψE1 ×. . .×ψEn for all E1 , . . . , En ∈ Σ. Thus 346I tells us, in part, that every complete probability space has a consistent lifting; it follows that every non-trivial complete totally finite measure space has a consistent lifting. I do not suppose you will be surprised to be told that not all liftings on probability spaces are consistent. What may be surprising is the fact that one of the standard liftings already introduced is not consistent. This depends on a general fact about Stone spaces of measure algebras which has further important applications, so I present it as a lemma. 346K Lemma Let (Z, T, ν) be the Stone space of the measure algebra of Lebesgue measure on [0, 1], and let λ be the product measure on Z × Z, with Λ its domain. Then there is a set W ∈ Λ, with λW < 1, ˜ = 1, where such that λ∗ W ˜ = S{G × H : G, H ⊆ Z are open-and-closed, (G × H) \ W is negligible}. W Remark For the sake of anybody who has already become acquainted with the alternative measures which can be put on the product of topological measure spaces, I ought to insist here that the ‘product measure’ λ is, as always in this volume, the ordinary completed product measure as defined in Chapter 25. proof (a) Let hEn in∈N be a sequence of measurable subsets of [0, 1], stochastically independent for Lebesgue 1 measure µ on [0, 1], such that µEn = n+2 for each n. Set an = En• in the measure of algebra of µ, and S ∗ an the corresponding compact open subset of Z. Set W = n∈N En∗ × En∗ . Then En = b λW ≤

P∞

2 n=0 (νEn )

=

P∞

1

n=2 n2

< 1.

˜ < 1. Then there are sequences hGn in∈N , hHn in∈N in T such that ?? Suppose, if possible, that λ∗ W S S ˜ W ⊆ n∈N Gn × Hn and λ( n∈N Gn × Hn ) < 1. Recall from 322Qc that νF = inf{νG : G is compact and open, F ⊆ G} ñ, H ˜ n such that Gn ⊆ G ˜ n , Hn ⊆ H ˜ n for for every F ∈ T. Accordingly we can find compact open sets G every n ∈ N and S P∞ P∞ ˜ ˜ n=0 ν(Hn \ Hn ) < 1 − λ( n∈N Gn × Hn ), n=0 ν(Gn \ Gn ) + S ñ × H ˜ n ) < 1. so that λ( n∈N G Let U0 be the family ˜ n : n ∈ N} ∪ {Z \ H ˜ n : n ∈ N}, {Z} ∪ {En∗ : n ∈ N} ∪ {Z \ G so that U0 is a countable subset of T. Let U be the set of finite intersections U0 ∩ U1 ∩ . . . ∩ Un where U0 , . . . , Un ∈ U0 , so that U is also a countable subset of T, and U is closed under ∩. (b) For U ∈ U, define Q(U ) as follows. If νU = 0, then Q(U ) = U . Otherwise, S Q(U ) = Z \ {En∗ : n ∈ N, ν(En∗ ∩ U ) > 0}. Then νQ(U ) is always 0. P P Of course this is true if S νU = 0, so suppose that νU > 0. Set I = {n : ν(En∗ ∩ U ) = 0}. Then we have νU 0 > 0, where U 0 = U \ n∈I En∗ , and Z \ En∗ ⊇ U 0 for every n ∈ I. Because hEn in∈N is stochastically independent for µ, hEn∗ in∈N is stochastically independent for ν, while S ν( n∈I En∗ ) ≤ 1 − νU 0 < 1. P P P∞ 1 By the Borel-Cantelli lemma (273K), n∈I νEn∗ < ∞. Consequently n∈N\I νEn∗ = ∞, because n=0 n+2 is infinite, so S ν(Z \ Q(U )) = ν( n∈N\I En∗ ) = 1, and νQ(U ) = 0. Q Q S (c) Set Q0 = U ∈U Q(U ); because U is countable, Q0 is negligible. Accordingly (Z \ Q0 )2 has measure S ñ × H ˜ n ; take (w, z) ∈ (Z \ Q0 )2 \ S ˜ ˜ 1 and cannot be included in n∈N G n∈N Gn × Hn .

216

Liftings

346K

(d) We can find sequences hCn in∈N , hDn in∈N , hUn in∈N and hVn in∈N in U such that ñ × H ˜ n ) = ∅, w ∈ Un+1 ⊆ Un , z ∈ Vn+1 ⊆ Vn , (Un+1 × Vn+1 ) ∩ (G νCn > 0, νDn > 0, Cn ⊆ Un , Dn ⊆ Vn+1 , Cn × Vn+1 ⊆ W , Un+1 × Dn ⊆ W for every n ∈ N. P P Build the sequences inductively, as follows. Start with U0 = V0 = Z. Given that ñ × H ˜ n . If w ∈ ˜ n , set Un0 = Un \ G ñ, w ∈ Un ∈ U , z ∈ Vn ∈ U , then we know that (w, z) ∈ / G / G 0 0 0 ˜ n . In either case, we have w ∈ Un0 ∈ U, z ∈ Vn0 ∈ U and Vn = Vn ; otherwise set Un = Un , Vn = Vn \ H ñ × H ˜ n ) = ∅. (Un0 × Vn0 ) ∩ (G 0 / Q(Un0 ), so ∈ U, w ∈ / Q(Un0 ). But w ∈ Un0 , so this must be because νUn0 > 0. Now z ∈ Because U n S z ∈ {Ek∗ : k ∈ N, ν(Ek∗ ∩ Un0 ) > 0}. Take some k ∈ N such that z ∈ Ek∗ and ν(Ek∗ ∩ Un0 ) > 0, and set Vn+1 = Vn0 ∩ Ek∗ ,

Cn = Ek∗ ∩ Un0 ,

so that z ∈ Vn+1 ∈ U ,

Cn ⊆ Un ,

Cn × Vn+1 ⊆ Ek∗ × Ek∗ ⊆ W , νCn > 0.

Next, z ∈ / Q(Vn+1 ) and νVn+1 > 0; also w ∈ / Q(Vn+1 ), so there is an l such that w ∈ El∗ and ν(El∗ ∩Vn+1 ) > 0. Set Un+1 = Un0 ∩ El∗ ,

Dn = El∗ ∩ Vn+1 ,

so that w ∈ Un+1 ∈ U , Dn ⊆ Vn+1 ,

Un+1 × Dn ⊆ El∗ × El∗ ⊆ W ,

νDn > 0,

ñ × H ˜ n ) = ∅, ñ × H ˜ n ) ⊆ (Un0 × Vn0 ) ∩ (G (Un+1 × Vn+1 ) ∩ (G and continue the process. Q Q S S P If m ≤ n, Dn ⊆ Vn+1 ⊆ Vm+1 , so (e) Setting C = n∈N Cn , D = n∈N Dn we see that C × D ⊆ W . P Cm × Dn ⊆ W . If m > n, Cm ⊆ Um ⊆ Un+1 , so Cm × Dn ⊆ W . Q Q Recall from 322Qa that the measurable sets of Z are precisely those of the form G4M where M is nowhere dense and negligible and G is compact and open. There must therefore be compact open sets G, H ⊆ Z such that G4C and H4D are negligible. Consequently (G × H) \ W ⊆ ((G \ C) × Z) ∪ (Z × (H \ D)) is negligible, and ˜ ⊆ G×H ⊆W

S n∈N

ñ × H ˜ n. G

ñ × H ˜ n are open, there must be some n such that But because G × H is compact (3A3J), and all the G S ˜ ˜ ˜ ˜ k ) = ∅ for every k, so G × H ⊆ k≤n Gk × Hk = S say. Now (Uk+1 × Vk+1 ) ∩ (Gk × H (Cn+2 × Dn+2 ) ∩ (G × H) ⊆ (Un+1 × Vn+1 ) ∩ S = ∅, and either Cn+2 ∩ G = ∅ or Dn+2 ∩ H = ∅. Since Cn+2 \ G ⊆ C \ G,

Dn+2 \ H ⊆ D \ H

are both negligible, one of Cn+2 , Dn+2 is negligible. But the construction took care to ensure that all the Ck , Dk were non-negligible. X X ˜ = 1, as required. (f ) Thus λ∗ W 346L Proposition Let (Z, T, ν) be the Stone space of the measure algebra of Lebesgue measure on [0, 1]. Let ψ : T → T be the canonical lifting, defined by setting ψE = G whenever E ∈ T, G is open-and-closed and E4G is negligible (341O). Then ψ is not consistent. proof ?? Suppose, if possible, that φ is a lifting on Z × Z such that φ(E × F ) = ψE × ψF for every E, F ∈ T. Let W ⊆ Z × Z be a set as in 346K, and consider φW . If G, H ⊆ Z are open-and-closed and (G × H) \ W is negligible, then

346Yd

Consistent liftings

217

G × H = ψG × ψH = φ(G × H) ⊆ φW ; ˜ ⊆ φW . But this means that that is, in the language of 346K, we must have W ˜ = 1 > λW , λ(φW ) ≥ λ∗ W which is impossible. X X Thus ψ fails the first test and cannot be consistent. 346X Basic exercises (a) Let (X, Σ, µ) be a measure space and φ a lower density for µ. Take H ∈ Σ and set A = X \ (φH ∪ φ(Z \ H)), φ0 E = φE ∪ (A ∩ φ(H ∪ E)) for E ∈ Σ. Show that φ0 is a lower density. > (b) Show that there is no lifting φ of Lebesgue measure on [0, 1]2 which is ‘symmetric’ in the sense that φ(E −1 ) = (φE)−1 for every measurable set E, writing E −1 = {(y, x) : (x, y) ∈ E}. (c) Let (X, Σ, µ) be a measure space and hφn in∈N a sequence of lower densities for µ. (i) Show that S T T E 7→ n∈N φn E and E 7→ n∈N m≥n φm E are also lower densities for µ. (ii) Show that if µ is complete S T and F is any filter on N, then E 7→ F ∈F n∈F φn E is a lower density for µ. (d) Let (X, Σ, µ) be a strictly localizable measure space, and G a countable group of measure space automorphisms from X to itself. Show that there is a lower density φ : Σ → Σ which is G-invariant in the T sense that φ(g −1 [E]) = g −1 [φE] for every E ∈ Σ, g ∈ G. (Hint: set φE = g∈G g[φ0 (g −1 [E])].) > (e) Let φ be lower Lebesgue density on R, and φ any lifting of Lebesgue measure on R such that φE ⊇ φE for every measurable set E. Show that φ is consistent. (Hint: given n ≥ 1, let φn be lower Lebesgue density on Rn . For x ∈ Rn let Ix be the ideal generated by S {W : x ∈ φn (Rn \ W )} ∪ i (g) Suppose, in 341H, that (X, Σ, µ) is a product of probability spaces, and that in the proof, instead of taking haξ iξ<κ to run over the whole measure algebra A, we take it to run over the elements of A expressible as E • where E ∈ Σ is determined by a single coordinate. Show that the resulting lower density θ respects coordinates in the sense that θE • is determined by coordinates in J whenever E ∈ Σ is determined by coordinates in J. (Compare Macheras & Strauss 95, Theorem 2.) 346Y Further exercises (a) Suppose that (X, Σ, µ) and (Y, T, ν) are complete probability spaces with product (X × Y, Λ, λ). Show that for any lifting ψ1 : Σ → Σ there are liftings ψ2 : T → T and φ : Λ → Λ such that φ(E × F ) = ψ1 E × ψ2 F for all E ∈ Σ, F ∈ T. (Hint: use the methods of §341. In the inductive construction of 341H, start with φ0 (E × Y ) = ψ1 E × Y for every E ∈ Σ. Extend each lower density φξ to the algebra generated by dom(φξ ) ∪ {X × Fξ } for some Fξ ∈ T. Make sure that φξ (X × F ) is always of the form X × F 0 , and that φξ ((E × Y ) ∪ (X × F )) = φξ (E × Y ) ∪ φξ (X × F ); adapt the construction of 341G to maintain this. Use the method of 346H to generate a lifting from the final lower density φ. See Macheras & Strauss 96a, Theorem 4.) (b) Use 346Ya and induction on ζ to prove 346H. (Macheras & Strauss 96b.) (c) Let (X1 , Σ1 , µ1 ), . . . , (Xn , Σn , µn ) be probability spaces with product (X, Σ, µ). Show that there is a lifting for µ which respects coordinates. (Burke n95.) (d) Let (X, Σ, µ) be a complete probability space. Show that there is a lifting ψ : Σ → Σ such that whenever h(Xi , Σi , µi )ii∈I is a family of probability spaces, with product measure λ, there is a lifting φ for −1 λ such that φ(πi−1 [E]) Q = πi [ψE] whenever E ∈ Σ and i ∈ I is such that (Xi , Σi , µi ) = (X, Σ, µ), writing πi (x) = x(i) for x ∈ i∈I Xi .

218

Liftings

346Ye

(e) In 346Ya, find a modification of the countable-cofinality inductive step of 341G which will ensure that the lower density obtained in the product satisfies both conditions of 346G. (f ) Let (X, Σ, µ) be a probability space, I any set, and λ the product measure on X I . Show that there is a lower density for λ which is invariant under transpositions of pairs of coordinates. 346Z Problems (a) Let h(Xi , Σi , µi )ii∈I be a family of probability spaces, with product (X, Σ, µ). Is there always a lifting for µ which respects coordinates in the sense of 346A? (b) Is there a lower density φ for the usual measure on {0, 1}N which is invariant under all permutations of coordinates? 346 Notes and comments I ought to say at once that in writing this section I have been greatly assisted by M.R.Burke. The theorem that every complete probability space has a consistent lifting (346J) is due to Talagrand 82a; it is the inspiration for the whole of the section. ‘Consistent’ liftings were devised in response to some very interesting questions (see Talagrand 84, §6) which I do not discuss here; one will be mentioned in §465 in Volume 4. My aim here is rather to suggest further ways in which a lifting on a product space can be consistent with the product structure. The labour is substantial and the results achieved are curiously partial. I offer 346Za as the easiest natural question which does not appear amenable to the methods I describe. The arguments I use are based on the fact that the translation-invariant measures of 345C already respect coordinates. Maharam’s theorem now makes it easy to show that any product of Maharam-typehomogeneous probability spaces has a lifting which respects coordinates. A kind of projection argument (346F) makes it possible to obtain a lower density which respects coordinates on any product of probability spaces (346G). In fact the methods of §341, very slightly refined, automatically produce such lower densities (346Xg). But the extra power of 346G lies in the condition (ii): if E and F are ‘fully independent’ in the sense of being determined by coordinates in dist sets, then φ(E ∪ F ) = φE ∪ φF , that is, φ is making a tentative step towards being a lifting. ( that the difference between a lifting and a lower density is mostly that a lifting preserves finite unions as well as finite intersections; see 341Xa.) This can also be achieved by a modification of the previous method, but we have to work harder at one point in the proof (346Ye). The next step is to move to liftings which continue, as far as possible, to respect coordinates. Here there seem to be quite new obstacles, and 346H is the best result I know; the lifting respects individual coordinates, and also, for a given well-ordering of the index set, initial segments of the coordinates. The treatment of initial segments makes essential use of the well-ordering, which is what leaves 346Za open. Finally, if all the factors are identical, we can seek lower densities and liftings which are invariant under permutation of coordinates. I give 345Xc and 346Xb as examples to show that we must not just assume that a symmetry in the underlying measure space can be reflected in a symmetry of a lifting. The problems there concern liftings themselves, not lower densities, since we can frequently find lower densities which share symmetries (346Xd, 346Yf). (Even for lower densities there seem to be difficulties if we are more ambitious (346Zb).) However a very simple argument (346I) shows that at least we can make each individual coordinate look more or less the same, as long as we do not investigate its relations with others. Still on the question of whether, and when, liftings can be ‘good’, note 346L/346Xf and 346Xe. The most natural liftings of Lebesgue measure are necessarily consistent; but the only example we have of a truly canonical lifting is not consistent in any non-trivial context. I have deliberately used a variety of techniques here, even though 346H (for instance) has an alternative proof based on the ideas of §341 (346Ya-346Yb). In particular, I give some of the standard methods of constructing liftings and lower densities (346D, 346F, 346B, 346Xa, 346Xc). In fact 346D was one of the elements of Maharam’s original proof of the lifting theorem (Maharam 58).

351Bc

Partially ordered linear spaces

219

Chapter 35 Riesz spaces The next three chapters are devoted to an abstract description of the ‘function spaces’ described in Chapter 24, this time concentrating on their internal structure and their relationships with their associated measure algebras. I find that any convincing of these must involve a substantial amount of general theory concerning partially ordered linear spaces, and in particular various types of Riesz space or vector lattice. I therefore provide an introduction to this theory, a kind of appendix built into the middle of the volume. The relation of this chapter to the next two is very like the relation of Chapter 31 to Chapter 32. As with Chapter 31, it is not really meant to be read for its own sake; those with a particular interest in Riesz spaces might be better served by Luxemburg & Zaanen 71, Schaefer 74, Zaanen 83 or my own book Fremlin 74a. I begin with three sections in an easy gradation towards the particular class of spaces which we need to understand: partially ordered linear spaces (§351), general Riesz spaces (§352) and Archimedean Riesz spaces (§353); the last includes notes on Dedekind (σ)-complete spaces. These sections cover the fragments of the algebraic theory of Riesz spaces which I will use. In the second half of the chapter, I deal with normed Riesz spaces (in particular, L- and M -spaces)(§354), spaces of linear operators (§355) and dual Riesz spaces (§356).

351 Partially ordered linear spaces I begin with an of the most basic structures which involve an order relation on a linear space, partially ordered linear spaces. As often in this volume, I find myself impelled to do some of the work in very much greater generality than is strictly required, in order to show more clearly the nature of the arguments being used. I give the definition (351A) and most elementary properties (351B-351L) of partially ordered linear spaces; then I describe a general representation theorem for arbitrary partially ordered linear spaces as subspaces of reduced powers of R (351M-351Q). I end with a brief note on Archimedean partially ordered linear spaces (351R). 351A Definition I repeat a definition mentioned in 241E. A partially ordered linear space is a linear space (U, +, ·) over R together with a partial order ≤ on U such that u ≤ v =⇒ u + w ≤ v + w, u ≥ 0, α ≥ 0 =⇒ αu ≥ 0 for u, v, w ∈ U and α ∈ R. 351B Elementary facts Let U be a partially ordered linear space. We have the following elementary consequences of the definition above, corresponding to the familiar rules for manipulating inequalities among real numbers. (a) For u, v ∈ U , u ≤ v =⇒ 0 = u + (−u) ≤ v + (−u) = v − u =⇒ u = 0 + u ≤ v − u + u = v, u ≤ v =⇒ −v = u + (−v − u) ≤ v + (−v − u) = −u. (b) Suppose that u, v ∈ U and u ≤ v. Then αu ≤ αv for every α ≥ 0 and αv ≤ αu for every α ≤ 0. P P (i) If α ≥ 0, then α(v − u) ≥ 0 so αv ≥ αu. (ii) If α ≤ 0 then (−α)u ≤ (−α)v so αv = −(−α)v ≤ −(−αu) = u. Q Q (c) If u ≥ 0 and α ≤ β in R, then (β − α)u ≥ 0, so αu ≤ βu. If 0 ≤ u ≤ v in U and 0 ≤ α ≤ β in R, then αu≤ βu ≤ βv.

220

Riesz spaces

351C

351C Positive cones Let U be a partially ordered linear space. (a) I will write U + for the positive cone of U , the set {u : u ∈ U, u ≥ 0}. (b) By 351Ba, the ordering is determined by the positive cone U + , in the sense that u ≤ v ⇐⇒ v − u ∈ U . +

(c) It is easy to characterize positive cones. If U is a real linear space, a set C ⊆ U is the positive cone for some ordering rendering U a partially ordered linear space iff u + v ∈ C,

αu ∈ C whenever u, v ∈ C and α ≥ 0,

0 ∈ C, P P (i) If C = U

+

u ∈ C & − u ∈ C =⇒ u = 0.

for some partially ordered linear space ordering ≤ on U , then u, v ∈ C =⇒ 0 ≤ u ≤ u + v =⇒ u + v ∈ C, u ∈ C, α ≥ 0 =⇒ αu ≥ 0, i.e., αu ∈ C, 0 ≤ 0 so 0 ∈ C, u, −u ∈ C =⇒ u = 0 + u ≤ (−u) + u = 0 ≤ u =⇒ u = 0.

(ii) On the other hand, if C satisfies the conditions, define the relation ≤ by writing u ≤ v ⇐⇒ v − u ∈ C; then u − u = 0 ∈ C so u ≤ u for every u ∈ U , if u ≤ v and v ≤ w then w − u = (w − v) + (v − u) ∈ C so u ≤ w, if u ≤ v and v ≤ u then u − v, v − u ∈ C so u − v = 0 and u = v and ≤ is a partial order; moreover, if u ≤ v and w ∈ U then (v + w) − (u + w) = v − u ∈ C and u + w ≤ v + w, if u, α ≥ 0 then αu ∈ C and αu ≥ 0, u ≥ 0 ⇐⇒ u ∈ C. So ≤ makes U a partially ordered linear space in which C is the positive cone. Q Q (d) An incidental useful fact. Let U be a partially ordered linear space, and u ∈ U . Then u ≥ 0 iff Q u ≥ −u. P P If u ≥ 0 then 0 ≥ −u so u ≥ −u. If u ≥ −u then 2u ≥ 0 so u = 21 · 2u ≥ 0. Q (e) I have called U + a ‘positive cone’ without defining the term ‘cone’. I think this is something we can by for the moment; but it will be useful to recognise that U + is always convex, for if u, v ∈ U + and α ∈ [0, 1] then αu, (1 − α)v ≥ 0 and αu + (1 − α)v ∈ U + . 351D Suprema and infima Let U be a partially ordered linear space. (a) The definition of ‘partially ordered linear space’ implies that u 7→ u+w is always an order-isomorphism; on the other hand, u 7→ −u is order-reversing, by 351Ba. (b) It follows that if A ⊆ U , v ∈ U then supu∈A v + u = v + sup A if either side is defined, inf u∈A v + u = v + inf A if either side is defined, supu∈A v − u = v − inf A if either side is defined, inf u∈A v − u = v − sup A if either side is defined.

351G


221

(c) Moreover, we find that if A, B ⊆ U and sup A and sup B are defined, then sup(A + B) is defined and equal to sup A + sup B, writing A + B = {u + v : u ∈ A, v ∈ B} as usual. P P Set u0 = sup A, v0 = sup B. Using (b), we have u0 + v0 = sup (u + v0 ) u∈A

= sup (sup (u + v)) = sup(A + B). Q Q u∈A v∈B

Similarly, if A, B ⊆ U and inf A, inf B are defined then inf(A + B) = inf A + inf B. (d) If α > 0 then u 7→ αu is an order-isomorphism, so we have sup(αA) = α sup A if either side is defined; similarly, inf(αA) = α inf A. 351E Linear subspaces If U is a partially ordered linear space, and V is any linear subspace of U , then V , with the induced linear and order structures, is a partially ordered linear space; this is obvious from the definition. 351F Positive linear operators Let U and V be partially ordered linear spaces, and write L(U ; V ) for the linear space of all linear operators from U to V . For S, T ∈ L(U ; V ) say that S ≤ T iff Su ≤ T u for every u ∈ U + . Under this ordering, L(U ; V ) is a partially ordered linear space; its positive cone is {T : T u ≥ 0 for every u ∈ U + }. P P This is an elementary verification. Q Q Note that, for T ∈ L(U ; V ), T ≥ 0 =⇒ T u ≤ T u + T (v − u) = T v whenever u ≤ v in U =⇒ 0 = T 0 ≤ T u for every u ∈ U + =⇒ T ≥ 0, so that T ≥ 0 iff T is order-preserving. In this case we say that T is a positive linear operator. Clearly ST is a positive linear operator whenever U , V and W are partially ordered linear spaces and T : U → V , S : V → W are positive linear operators (cf. 313Ia). 351G Order-continuous positive linear operators: Proposition Let U and V be partially ordered linear spaces and T : U → V a positive linear operator. (a) The following are equiveridical: (i) T is order-continuous; (ii) inf T [A] = 0 in V whenever A ⊆ U is a non-empty downwards-directed set with infimum 0 in U ; (iii) sup T [A] = T w in V whenever A ⊆ U + is a non-empty upwards-directed set with supremum w in U. (b) The following are equiveridical: (i) T is sequentially order-continuous; (ii) inf n∈N T un = 0 in V whenever hun in∈N is a non-increasing sequence in U with infimum 0 in U ; (iii) supn∈N T un = T w in V whenever hun in∈N is a non-decreasing sequence in U + with supremum w in U . proof (a)(i)⇒(iii) is trivial. (iii)⇒(ii) Assuming (iii), and given that A is non-empty, downwards-directed and has infimum 0, take any u0 ∈ A and consider A0 = {u : u ∈ A, u ≤ u0 }, B = u0 − A0 . Then A0 is non-empty, downwards-directed and has infimum 0, so B is non-empty, upwards-directed and has supremum u0 (using 351Db); by (iii), sup T [B] = T u0 and (inverting again) inf T [A0 ] = inf T [u0 − B] = inf T u0 − T [B] = T u0 − sup T [B] = 0. But (because T is positive) 0 is surely a lower bound for T [A], so it is also the infimum of T [A]. As A is arbitrary, (ii) is true. α) If A ⊆ U is non-empty, downwards-directed and has (ii)⇒(i) Suppose now that (ii) is true. (α infimum w, then A − w is non-empty, downwards-directed and has infimum 0, so

222

Riesz spaces

351G

inf T [A − w] = 0, inf T [A] = inf(T [A − w] + T w) = T w + inf T [A − w] = T w. β ) If A ⊆ U is non-empty, upwards-directed and has supremum w, then −A is non-empty, downwards(β directed and has infimum −w, so sup T [A] = − inf(−T [A]) = − inf T [−A] = −T (−w) = T w. Putting these together, T is order-continuous. (b) The arguments are identical, replacing each directed set by an appropriate sequence. 351H Riesz homomorphisms (a) For the sake of a representation theorem below (351Q), I introduce the following definition. Let U , V be partially ordered linear spaces. A Riesz homomorphism from U to V is a linear operator T : U → V such that whenever A ⊆ U is a finite non-empty set and inf A = 0 in U , then inf T [A] = 0 in V . The following facts are now nearly obvious. (b) Any Riesz homomorphism is a positive linear operator. (For if T is a Riesz homomorphism and u ≥ 0, then inf{0, u} = 0 so inf{0, T u} = 0 and T u ≥ 0.) (c) Let U and V be partially ordered linear spaces and T : U → V a Riesz homomorphism. Then inf T [A] exists = T (inf A),

sup T [A] exists = T (sup A)

whenever A ⊆ U is a finite non-empty set and inf A, sup A exist. (Apply the definition in (a) to A0 = {u − inf A : u ∈ A},

A00 = {sup A − u : u ∈ A}.)

(d) If U , V and W are partially ordered linear spaces and T : U → V , S : V → W are Riesz homomorphisms then ST : U → W is a Riesz homomorphism. 351I Solid sets Let U be a partially ordered linear space. I will say that a subset A of U is solid if S A = {v : v ∈ U, −u ≤ v ≤ u for some u ∈ A} = u∈U [−u, u] in the notation of 2A1Ab. (I should perhaps remark that while this definition is well established in the case of Riesz spaces (§352), the extension to general partially ordered linear spaces is not standard. See 351Yb for a warning.) 351J Proposition Let U be a partially ordered linear space and V a solid linear subspace of U . Then the quotient linear space U/V has a partially ordered linear space structure defined by either of the rules u• ≤ w• iff there is some v ∈ V such that u ≤ v + w, (U/V )+ = {u• : u ∈ U + }, and for this partial ordering on U/V the map u 7→ u• : U → U/V is a Riesz homomorphism. proof (a) I had better start by giving priority to one of the descriptions of the relation ≤ on U/V ; I choose the first. To see that this makes U/V a partially ordered linear space, we have to check the following. (i) 0 ∈ V and u ≤ u + 0, so u• ≤ u• for every u ∈ U . (ii) If u1 , u2 , u3 ∈ U and u•1 ≤ u•2 , u•2 ≤ u•3 then there are v1 , v2 ∈ V such that u1 ≤ u2 + v1 , u2 ≤ u3 + v2 ; in which case v1 + v2 ∈ V and u1 ≤ u3 + v1 + v2 , so u•1 ≤ u•3 . (iii) If u, w ∈ U and u• ≤ w• , w• ≤ u• then there are v, v 0 ∈ V such that u ≤ w + v, w ≤ u + v 0 . Now there are v0 , v00 ∈ V such that −v0 ≤ v ≤ v0 , −v00 ≤ v 0 ≤ v00 , and in this case v0 , v00 ≥ 0 (351Cd), so −v0 − v00 ≤ −v 0 ≤ u − w ≤ v ≤ v0 + v00 ∈ V , . Accordingly u − w ∈ V and u• = w• . Thus U/V is a partially ordered set. (iv) If u1 , u2 , w ∈ U and u•1 ≤ u•2 , then there is a v ∈ V such that u1 ≤ u2 + v, in which case u1 + w ≤ u2 + w + v and u•1 + w• ≤ u•2 + w• . (v) If u ∈ U , α ∈ R and u• ≥ 0, α ≥ 0 then there is a v ∈ V such that u + v ≥ 0; now αv ∈ V and αu + αv ≥ 0, so αu• = (αu)• ≥ 0.

351N


223

Thus U/V is a partially ordered linear space. (b) Now (U/V )+ = {u• : u ≥ 0}. P P If u ≥ 0 then of course u• ≥ 0 because 0 ∈ V and u + 0 ≥ 0. On the other hand, if we have any element p of (U/V )+ , there are u ∈ U , v ∈ V such that u• = p and u + v ≥ 0; but now p = (u + v)• is of the required form. Q Q (c) Finally, u 7→ u• is a Riesz homomorphism. P P Suppose that A ⊆ U is a non-empty finite set and that inf A = 0 in U . Then u• ≥ 0 for every u ∈ A, that is, 0 is a lower bound for {u• : u ∈ A}. Let p ∈ U/V be any other lower bound for {u• : u ∈ A}. Express p as w• where w ∈ U . For each u ∈ A, w• ≤P u• so there is 0 0 0 ∗ a vu ∈ V such that w ≤ u + vu . Next, there is a vu ∈ V such that −vu ≤ vu ≤ vu . Set v = u∈A vu0 ∈ V . Then vu ≤ vu0 ≤ v ∗ so w ≤ u + v ∗ for every u ∈ A, and w − v ∗ is a lower bound for A in U . Accordingly w − v ∗ ≤ 0, w ≤ 0 + v ∗ and p = w• ≤ 0. As p is arbitrary, inf{u• : u ∈ A} = 0; as A is arbitrary, u 7→ u• is a Riesz homomorphism. Q Q 351K Lemma Suppose that U is a partially ordered linear space, and that W , V are solid linear subspaces of U such that W ⊆ V . Then V1 = {v • : v ∈ V } is a solid linear subspace of U/W . proof (i) Because the map u 7→ u• is linear, V1 is a linear subspace of U/W . (ii) If p ∈ V1 , there is a v ∈ V such that p = v • ; because V is solid in U , there is a v0 ∈ V such that −v0 ≤ v ≤ v0 ; now v0• ∈ V1 and −v0• ≤ p ≤ v0• . (iii) If p ∈ V1 , q ∈ U/W and −p ≤ q ≤ p, take v0 ∈ V , u ∈ U such that v0• = p, u• = q. Because −v0• ≤ u• ≤ v0• , there are w, w0 ∈ W such that −v0 − w ≤ u ≤ v0 + w0 . Now −v0 − w, v0 + w0 both belong to V , which is solid, so both are mapped to 0 by the canonical Riesz homomorphism from U to U/V , and u must also be, that is, u ∈ V and q = u• ∈ V1 . (iv) Putting (ii) and (iii) together, V1 is solid. 351L Products If hUi ii∈I is any family of partially ordered linear spaces, we have a product linear Q space U = i∈I Ui ; if we set u ≤ v in U iff u(i) ≤ v(i) for every i ∈ I, U becomes a partially ordered linear space, with positive cone {u : u(i) ≥ 0 for every i ∈ I}. For each i ∈ I the map u 7→ u(i) : U → Ui is an order-continuous Riesz homomorphism (in fact, it preserves arbitrary suprema and infima). 351M Reduced powers of R (a) Let X be any set. Then RX is a partially ordered linear space if we say that f ≤ g means that f (x) ≤ g(x) for every x ∈ X, as in 351L. If now F is a filter on X, we have a corresponding set V = {f : f ∈ RX , {x : f (x) = 0} ∈ F}; it is easy to see that V is a linear subspace of RX , and is solid because f ∈ V iff |f | ∈ V . By the reduced power RX |F I shall mean the quotient partially ordered linear space RX /V . (b) Note that for f ∈ RX , f • ≥ 0 in RX |F ⇐⇒ {x : f (x) ≥ 0} ∈ F. P P (i) If f • ≥ 0, there is a g ∈ V such that f + g ≥ 0; now {x : f (x) ≥ 0} ⊇ {x : g(x) = 0} ∈ F. (ii) If {x : f (x) ≥ 0} ∈ F, then {x : (|f | − f )(x) = 0} ∈ F, so f • = |f |• ≥ 0. Q Q 351N On the way to the next theorem, the main result (in of mathematical depth) of this section, we need a string of lemmas. Lemma Let U be a partially ordered linear space. If u, v0 , . . . , vn ∈ U are such that u 6= 0 and v0 , . . . , vn ≥ 0 then there is a linear functional f : U → R such that f (u) 6= 0 and f (vi ) ≥ 0 for every i. proof The point is that at most one of u, −u can belong to the convex cone C generated by {v0 , . . . , vn }, because this is included in the convex cone set U + , and since u 6= 0 at most one of u, −u can belong to U + . Now however the Hahn-Banach theorem, in the form 3A5D, tells us that if u ∈ / C there is a linear functional f : U → R such that f (u) < 0, f (vi ) ≥ 0 for every i; while if −u ∈ / C we can get f (−u) < 0 and f (vi ) ≥ 0 for every i. Thus in either case we have a functional of the required type.

224

Riesz spaces

351O

351O Lemma Let U be a partially ordered linear space, and u0 a non-zero member of U . Then there is a solid linear subspace V of U such that u0 ∈ / V and whenever A ⊆ U is finite, non-empty and has infimum 0 then A ∩ V 6= ∅. proof (a) Let W be the family of all solid S linear subspaces of U not containing u0 . Then any non-empty totally ordered V ⊆ W has an upper bound V in W. By Zorn’s Lemma, W has a maximal element V say. This is surely a solid linear subspace of U not containing u0 . (b) Now for any w ∈ U + \ V there are α ≥ 0, v ∈ V + such that −αw − v ≤ u0 ≤ αw + v. P P Let V1 be {u : u ∈ U , there are α ≥ 0, v ∈ V + such that −αw − v ≤ u ≤ αw + v}. Then it is easy to check that V1 is a solid linear subspace of U , including V , and containing w; because w∈ / V , V1 6= V , so V1 ∈ / W and u ∈ V1 , as claimed. Q Q (c) It follows that if A ⊆ U is finite and non-empty and inf A = 0 in U then A∩V 6= ∅. P P?? Otherwise, for P every w ∈ A there must be αw ≥ 0, vw ∈ V + such that −αw w −vw ≤ u0 ≤ αw w +vw . Set α = 1+ w∈A αw , P v = w∈A vw ∈ V ; then −αw − v ≤ u0 ≤ αw + v for every w ∈ A. Accordingly α1 (u0 − v) ≤ w for every w ∈ A and α1 (u0 − v) ≤ 0, so u0 ≤ v. Similarly, − α1 (v + u0 ) ≤ w for every w ∈ A and −v ≤ u0 . But (because V is solid) this means that u0 ∈ V , which is not so. X XQ Q Accordingly V has the required properties. 351P Lemma Let U be a partially ordered linear space and u a non-zero element of U , and suppose that A0 , . . . , An are finite non-empty subsets of U such that inf Aj = 0 for every j ≤ n. Then there is a linear functional f : U → R such that f (u) 6= 0 and min f [Aj ] = 0 for every j ≤ n. proof By 351O, there is a solid linear subspace V of U such that u ∈ / V and Aj ∩ V 6= 0 for every S j ≤ n. Give the quotient space U/V its standard partial ordering (351J), and in U/V set C = {v • : v ∈ j≤n Aj }. Then C is a finite subset of (U/V )+ , while u• 6= 0, so by 351N there is a linear functional g : U/V → R such that g(u• ) 6= 0 but g(p) ≥ 0 S for every p ∈ C. Set f (v) = g(v • ) for v ∈ U ; then f : U → R is linear, f (u) 6= 0 and f (v) ≥ 0 for every v ∈ j≤n Aj . But also, for each j ≤ n, there is a vj ∈ Aj ∩ V , so that f (vj ) = 0; and this means that min f [Aj ] must be 0, as required. 351Q

Now we are ready for the theorem.

Theorem Let U be any partially ordered linear space. Then we can find a set X, a filter F on X and an injective Riesz homomorphism from U to the reduced power RX |F described in 351M. proof Let X be the set of all linear functionals f : U → R; define φ : U → RX by setting φ(u)(f ) = f (u) for every f ∈ X, u ∈ U , so that φ is linear. Let A be the family of non-empty finite sets A ⊆ U such that inf A = 0. For A ∈ A let FA be the set of those f ∈ X such that min f [A] = 0. Since 0 ∈ FA for every A ∈ A, the set T F = {F : F ⊆ X, there are A0 , . . . , An ∈ A such that F ⊇ j≤n FAj } is a filter on X. Set ψ(u) = φ(u)• ∈ RX |F for u ∈ U . The ψ : U → RX |F is an injective Riesz homomorphism. P P (i) ψ is linear because φ and h 7→ h• : RX → Rx |F are. (ii) If A ∈ A, then FA ∈ F . So, first, if v ∈ A, then {f : φ(v)(f ) ≥ 0} ∈ F, so that ψ(v) = φ(v)• ≥ 0 in RX |F (351Mb). Next, if w ∈ RX |F • and w ≤ ψ(v) for every v ∈ A, we can express w as h• where hT ≤ φ(v)• for every v ∈ A, that is, Hv = {f : h(f ) ≤ φ(v)(f )} ∈ F for every v ∈ A. But now H = FA ∩ v∈A Hv ∈ F, and for f ∈ H we have h(f ) ≤ minv∈A f (v) = 0. This means that w = h• ≤ 0. As w is arbitrary, inf ψ[A] = 0. As A is arbitrary, ψ is a Riesz homomorphism. (iii) Finally, ?? suppose, if possible, that there is a non-zero T u ∈ U such that ψ(u) = 0. Then F = {f : f (u) = 0} ∈ F, and there are A0 , . . . , An ∈ A such that F ⊇ j≤n FAj . By 351P, T there is an f ∈ j≤n FAj such that f (u) 6= 0; which is impossible. X X Accordingly ψ is injective, as claimed. Q Q

351 Notes


225

351R Archimedean spaces (a) For a partially ordered linear space U , the following are equiveridical: (i) if u, v ∈ U are such that nu ≤ v for every n ∈ N then u ≤ 0 (ii) if u ≥ 0 in U then inf δ>0 δu = 0. P P (i)⇒(ii) If (i) is true and u ≥ 0, then of course δu ≥ 0 for every δ > 0; on the other hand, if v ≤ δu for every δ > 0, then nv ≤ n · n1 u = u for every n ≥ 1, while of course 0v = 0 ≤ u, so v ≤ 0. Thus 0 is the greatest lower bound of {δu : δ > 0}. (ii)⇒(i) If (ii) is true and nu ≤ v for every n ∈ N, then 0 ≤ v and u ≤ n1 v for every n ≥ 1. If now δ > 0, then there is an n ≥ 1 such that n1 ≤ δ, so that u ≤ n1 v ≤ δv (351Bc). Accordingly u is a lower bound for {δv : δ > 0} and u ≤ 0. Q Q (b) I will say that partially ordered linear spaces satisfying the equiveridical conditions of (a) above are Archimedean. (c) Any linear subspace of an Archimedean partially ordered linear space, with the induced partially ordered linear space structure, is Archimedean. Q (d) Any product of Archimedean partially ordered linear spaces is Archimedean. P P If U = i∈I Ui is a product of Archimedean spaces, and nu ≤ v in U for every n ∈ N, then for each i ∈ I we must have nu(i) ≤ v(i) for every n, so that u(i) ≤ 0; accordingly u ≤ 0. Q Q In particular, RX is Archimedean for any set X. 351X Basic exercises > (a) Let ζ be any ordinal. The lexicographic ordering on Rζ is defined by saying that f ≤ g iff either f = g or there is a ξ < ζ such that f (η) = g(η) for η < ξ and f (ξ) < g(ξ). Show that this is a total order on Rζ which renders Rζ a partially ordered linear space. (b) Let U be a partially ordered linear space and V a linear subspace of U . Show that the formulae of 351J define a partially ordered linear space structure on the quotient U/V iff V is order-convex, that is, u ∈ V whenever v1 , v2 ∈ V and v1 ≤ u ≤ v2 . (c) Let hUi ii∈I be a family of partially ordered linear spaces with product U . Define Ti : Ui → U by setting Ti x = u where u(i) = x, u(j) = 0 for j 6= i. Show that Ti is an injective order-continuous Riesz homomorphism. > (d) Let U be a partially ordered linear space and hVi ii∈I a family of partially orderedQlinear spaces with product V . Show that L(U ; V ) can be identified, as partially ordered linear space, with i∈I L(U ; Vi ). > (e) Show that if U , V are partially ordered linear spaces and V is Archimedean, then L(U ; V ) is Archimedean. 351Y Further exercises (a) Give an example of two partially ordered linear spaces U and V and a bijective Riesz homomorphism T : U → V such that T −1 : V → U is not a Riesz homomorphism. (b) (i) Let U be a partially ordered linear space. Show that U is a solid subset of itself (on the definition 351I) iff U = U + − U + . (ii) Give an example of a partially ordered linear space U satisfying this condition with an element u ∈ U such that the intersection of the solid sets containing u is not solid. (c) Let U be a partially ordered linear space, and suppose that A, B ⊆ U are two non-empty finite sets such that (α) u ∨ v = sup{u, v} is defined for every u ∈ A, v ∈ B (β inf A and inf B and (inf A) ∨ (inf B) are defined. Show that inf{u ∨ v : u ∈ A, v ∈ B} = (inf A) ∨ (inf B). (Hint: show that this is true if U = R, if U = RX and if U = RX |F, and use 351Q.) (d) Show that a reduced power RX |F, as described in 351M, is totally ordered iff F is an ultrafilter, and in this case has a natural structure as a totally ordered field. 351 Notes and comments The idea of ‘partially ordered linear space’ is a very natural abstraction from the elementary examples of RX and its subspaces, and the only possible difficulty lies in guessing the exact boundary at which one’s standard manipulations with such familiar spaces cease to be valid in the general case. (For instance, most people’s favourite examples are Archimedean, in the sense of 351R, so it is prudent

226

Riesz spaces

351 Notes

to check your intuitions against a non-Archimedean space like that of 351Xa.) There is really no room for any deep idea to appear in 351B-351F. When I come to what I call ‘Riesz homomorphisms’, however (351H), there are some more interesting possibilities in the background. I shall not discuss the applications of Theorem 351Q to general partially ordered linear spaces; it is here for the sake of its application to Riesz spaces in the next section. But I think it is a very striking fact that not only does any partially ordered linear space U appear as a linear subspace of some reduced power of R, but the embedding can be taken to preserve any suprema and infima of finite sets which exist in U . This is in a sense a result of the same kind as the Stone representation theorem for Boolean algebras; it gives us a chance to confirm that an intuition valid for R or RX may in fact apply to arbitrary partially ordered linear spaces. If you like, this provides a metamathematical foundation for such results as those in 351B. I have to say that for partially ordered linear spaces it is generally quicker to find a proof directly from the definition than to trace through an argument relying on 351Q; but this is not always the case for Riesz spaces. I offer 351Yc as an example of a result where a direct proof does at least call for a moment’s thought, while the argument through 351Q is straightforward. ‘Reduced powers’ are of course of great importance for other reasons; I mention 351Yd as a hint of what can be done.

352 Riesz spaces In this section I sketch those fragments of the theory we need which can be expressed as theorems about general Riesz spaces or vector lattices. I begin with the definition (352A) and most elementary properties (352C-352F). In 352G-352J I discuss Riesz homomorphisms and the associated subspaces (Riesz subspaces, solid linear subspaces); I mention product spaces (352K, 352T) and quotient spaces (352Jb, 352U) and the form the representation theorem 351Q takes in the present context (352L-352M). Most of the second half of the section concerns the theory of ‘bands’ in Riesz spaces, with the algebras of complemented bands (352Q) and projection bands (352S) and a description of bands generated by upwards-directed sets (352V). I conclude with a description of ‘f -algebras’ (352W). 352A

I repeat a definition from 241E.

Definition A Riesz space or vector lattice is a partially ordered linear space which is a lattice. 352B Lemma If U is a partially ordered linear space, then it is a Riesz space iff sup{0, u} is defined for every u ∈ U . proof If U is a lattice, then of course sup{u, 0} is defined for every u. If sup{u, 0} is defined for every u, and v1 , v2 are any two of U , consider w = v1 + sup{0, v2 − v1 }; by 351Db, w = sup{v1 , v2 }. Next, inf{v1 , v2 } = − sup{−v1 , −v2 } must also be defined in U ; as v1 and v2 are arbitrary, U is a lattice. 352C Notation In any Riesz space U I will write u+ = u ∨ 0,

u− = (−u) ∨ 0 = (−u)+ ,

|u| = u ∨ (−u)

where (as in any lattice) u ∨ v = sup{u, v} (and u ∧ v = inf{u, v}). I mention immediately a term which will be useful: a family hui ii∈I in U is dist if |ui | ∧ |uj | = 0 for all distinct i, j ∈ I. Similarly, a set C ⊆ U is dist if |u| ∧ |v| = 0 for all distinct u, v ∈ C. 352D Elementary identities Let U be a Riesz space. The translation-invariance of the order, and its invariance under positive scalar multiplication, reversal under negative multiplication, lead directly to the following, which are in effect special cases of 351D: u + (v ∨ w) = (u + v) ∨ (u + w),

u + (v ∧ w) = (u + v) ∧ (u + w),

352Ec

Riesz spaces

α(u ∨ v) = αu ∨ αv,

227

α(u ∧ v) = αu ∧ αv if α ≥ 0,

−(u ∨ v) = (−u) ∧ (−v). Combining and elaborating on these facts, we get u+ − u− = (u ∨ 0) − ((−u) ∨ 0) = u + (0 ∨ (−u)) − ((−u) ∨ 0) = u, u+ + u− = 2u+ − u = (2u ∨ 0) − u = u ∨ (−u) = |u|, u ≥ 0 ⇐⇒ −u ≤ 0 ⇐⇒ u− = 0 ⇐⇒ u = u+ ⇐⇒ u = |u|, | − u| = |u|,

| |u| | = |u|,

|αu| = |α||u| (looking at the cases α ≥ 0, α ≤ 0 separately),

u ∨ v + u ∧ v = u + (0 ∨ (v − u)) + v + ((u − v) ∧ 0) = u + (0 ∨ (v − u)) + v − ((v − u) ∨ 0) = u + v, u ∨ v = u + (0 ∨ (v − u)) = u + (v − u)+ , u ∧ v = u + (0 ∧ (v − u)) = u − (−0 ∨ (u − v)) = u − (u − v)+ , 1 2

1 2

1 2

u ∨ v = (2u ∨ 2v) = (u + v + (u − v) ∨ (v − u)) = (u + v + |u − v|), 1 2

u ∧ v = u + v − u ∨ v = (u + v − |u − v|), u+ ∨ u− = u ∨ (−u) ∨ 0 = |u|,

u+ ∧ u− = u+ + u− − (u+ ∨ u− ) = 0,

|u + v| = (u + v) ∧ ((−u) + (−v)) ≤ (|u| + |v|) ∧ (|u| + |v|) =|u| + |v|, ||u| − |v|| = (|u| − |v|) ∧ (|v| − |u|) ≤ |u − v| + |v − u| = |u − v|, |u ∨ v| ≤ |u| + |v| (because −|u| ≤ u ∨ v ≤ |u| ∨ |v| ≤ |u| + |v|) for u, v, w ∈ U and α ∈ R. 352E Distributive laws Let U be a Riesz space. (a) If A, B ⊆ U have suprema a, b in U , then C = {u ∧ v : u ∈ A, v ∈ B} has supremum a ∧ b. P P Of course u ∧ v ≤ a ∧ b for all u ∈ A, v ∈ B, so a ∧ b is an upper bound for C. Now suppose that c is any upper bound for C. If u ∈ A, v ∈ B then u − (u − v)+ = u ∧ v ≤ c,

u ≤ c + (u − v)+ ≤ c + (a − v)+

(because (a − v)+ = sup{a − v, 0} ≥ sup{u − v, 0} = (u − v)+ ). As u is arbitrary, a ≤ c + (a − v)+ and a ∧ v ≤ c. Now turn the argument round: v = (a ∧ v) + (v − a)+ ≤ c + (v − a)+ ≤ c + (b − a)+ , and this is true for every v ∈ B, so b ≤ c + (b − a)+ and a ∧ b ≤ c. As c is arbitrary, a ∧ b = sup C, as claimed. Q Q (b) Similarly, or applying (a) to −A and −B, inf{u ∨ v : u ∈ A, v ∈ B} = inf A ∨ inf B whenever A, B ⊆ U and the right-hand-side is defined. (c) In particular, U is a distributive lattice (definition: 3A1Ic).

228

Riesz spaces

352F

352F Further identities and inequalities At a slightly deeper level we have the following facts. Proposition Let U be a Riesz space. (a) If u, v, w ≥ 0 in U then u ∧ (v + w) ≤ (u ∧ v) + (u ∧ P w). Pn n (b) If u0 , . . . , un ∈ U and |ui | ∧ |uj | = 0 for i 6= j, then | i=0 αi ui | = i=0 |αi ||ui | for any α0 , . . . , αn . (c) If u, v ∈ U then u+ ∧ v + ≤ (u + v)+ ≤ u+ + v + . Pm Pn (d) If u0 , . . . , um , v0 , . . . , vn ∈ U + and i=0 ui = j=0 vj , then there is a family hwij ii≤m,j≤n in U + Pm Pn such that i=0 wij = vj for every j ≤ n and j=0 wij = ui for every i ≤ m. proof (a) u ∧ (v + w) ≤ [(u + w) ∧ (v + w)] ∧ u ≤ [(u ∧ v) + w] ∧ [(u ∧ v) + u] = (u ∧ v) + (u ∧ w). α) A simple (b)(i)(α the inductive step, shows that if v0 , . . . , vm , w0 , . . . , wn are Pm induction, Pn using (a) Pmfor P n β ) Next, if u ∧ v = 0 then non-negative then i=0 vi ∧ j=0 wj ≤ i=0 j=0 vi ∧ wj . (β (u − v)+ = u − (u ∧ v) = u,

(u − v)− = (v − u)+ = v − (v ∧ u) = v,

|u − v| = (u − v)+ + (u − v)− = u + v = |u + v|, so if |u| ∧ |v| = 0 then (u+ + v + ) ∧ (u− + v − ) ≤ (u+ ∧ u− ) + (u+ ∧ v − ) + (v + ∧ u− ) + (v + ∧ v − ) ≤ 0 + (|u| ∧ |v|) + (|v| ∧ |u|) + 0 = 0 and |u + v| = |(u+ + v + ) − (u− + v − )| = u+ + v + + u− + v − = |u| + |v|. (γγ ) Finally, if |u| ∧ |v| = 0 and α, β ∈ R, |αu| ∧ |βv| = |α||u| ∧ |β||v| ≤ (|α| + |β|)|u| ∧ (|α| + |β|)|v| = (|α| + |β|)(|u| ∧ |v|) = 0. (ii) We may therefore proceed by induction. The case n = 0 is trivial. For the inductive step to n + 1, setting u0i = αi ui we have |u0i | ∧ |u0j | = 0 for all i 6= j, by (i-γ). By (i-α), Pn Pn Pn |u0n+1 | ∧ | i=0 u0i | ≤ |u0n+1 | ∧ i=0 |u0i | ≤ i=0 |u0n+1 | ∧ |u0i | = 0, so by (i-β) and the inductive hypothesis Pn+1 Pn Pn+1 | i=0 u0i | = |u0n+1 | + | i=0 u0i | = i=0 |u0i | as required. (c) By 352E, u+ ∧ v + = (u ∨ 0) ∧ (v ∨ 0) = (u ∧ v) ∨ 0. Now 1 2

1 2

u ∧ v = (u + v − |u − v|) ≤ (u + v + |u + v|) = (u + v)+ , and of course 0 ≤ (u + v)+ , so u+ ∧ v + ≤ (u + v)+ . For the other inequality we need only note that u + v ≤ u+ + v + (because u ≤ u+ , v ≤ v + ) and 0 ≤ u+ + v + . Pm Pn (d) Write w for the common value of i=0 ui and j=0 vj . Induce on k = #({(i, j) : i ≤ m, j ≤ n, ui ∧ vj > 0}). If k = 0, that is, ui ∧ vj = 0 for all i, j, then (by (a), used repeatedly) we must have w ∧ w = 0, that is, w = 0, and we can take wij = 0 for all i, j. For the inductive step to k ≥ 1, take i∗ , j ∗ such that w ˜ = ui∗ ∧ vj ∗ > 0. Set

352Ic

Riesz spaces

229

u ˜i∗ = ui∗ − w, ˜ u ˜i = ui for i 6= i∗ ,

Then

Pm i=0

u ˜i =

v˜j ∗ = vj ∗ − w, ˜ v˜j = vj for j 6= j ∗ .

Pn j=0

v˜j = w − w ˜ and u ˜i ∧ v˜j ≤ ui ∧ vj for all i, j, while u ˜i∗ ∧ v˜j ∗ = 0; so that #({(i, j) : u ˜i ∧ vj > 0}) < k.

Pn By the inductive hypothesis, there are w ˜ij ≥ 0, for i ≤ m and j ≤ n, such that u ˜i = j=0 w ˜ij for each Pm Pn ∗ ∗ i, v˜j = i=0 w ˜ij for each j. Set wi∗ j ∗ = w ˜i∗ j ∗ + w, ˜ wij = w ˜ij for (i, j) 6= (i , j ); then ui = j=0 wij , Pm vj = i=0 wij so the induction proceeds. 352G Riesz homomorphisms: Proposition Let U be a Riesz space, V a partially ordered linear space and T : U → V a linear operator. Then the following are equiveridical: (i) T is a Riesz homomorphism in the sense of 351H; (ii) (T u)+ = sup{T u, 0} is defined and equal to T (u+ ) for every u ∈ U ; (iii) sup{T u, −T u} is defined and equal to T |u| for every u ∈ U ; (iv) inf{T u, T v} = 0 in V whenever u ∧ v = 0 in U . proof (i)⇒(iii) and (i)⇒(iv) are special cases of 351Hc. For (iii)⇒(ii) we have 1 2

1 2

1 2

1 2

1 2

sup{T u, 0} = T u + sup{ T u, − T u} = T u + T |u| = T (u+ ). For (ii)⇒(i), argue as follows. If (ii) is true and u, v ∈ U , then T u ∧ T v = inf{T u, T v} = T u + inf{0, T v − T u} = T u − sup{0, T (u − v)} is defined and equal to T u − T ((u − v)+ ) = T (u − (u − v)+ ) = T (u ∧ v). Inducing on n, inf i≤n T ui = T (inf i≤n ui ) for all u0 , . . . , un ∈ U ; in particular, if inf i≤n ui = 0 then inf i≤n T ui = 0; which is the definition I gave of Riesz homomorphism. Finally, for (iv)⇒(ii), we know from (iv) that 0 = inf{T (u+ ), T (u− )}, so −T (u+ ) = inf{0, −T u} and T (u+ ) = sup{0, T u}. 352H Proposition If U and V are Riesz spaces and T : U → V is a bijective Riesz homomorphism, then T is a partially-ordered-linear-space isomorphism, and T −1 : V → U is a Riesz homomorphism. proof Use 352G(ii). If v ∈ V , set u = T −1 v; then T (u+ ) = v + so T −1 (v + ) = u+ = (T −1 v)+ . Thus T −1 is a Riesz homomorphism; in particular, it is order-preserving, so T is an isomorphism for the order structures as well as for the linear structures. 352I Riesz subspaces (a) If U is a partially ordered linear space, a Riesz subspace of U is a linear subspace V such that sup{u, v} and inf{u, v} are defined in U and belong to V for every u, v ∈ V . In this case they are the supremum and infimum of {u, v} in V , so V , with the induced order and linear structure, is a Riesz space in its own right, and the embedding map u 7→ u : V → U is a Riesz homomorphism. (b) Generally, if U is a Riesz space, V is a partially ordered linear space and T : U → V is a Riesz homomorphism, then T [U ] is a Riesz subspace of V (because, by 351Hc, T u ∨ T u0 = T (u ∨ u0 ), T (u ∧ u0 ) = T u ∧ T u0 belong to T [U ] for all u, u0 ∈ U ). (c) If U is a Riesz space and V is a linear subspace of U , then V is a Riesz subspace of U iff |u| ∈ V for every u ∈ V . P P In this case, 1 2

u ∨ v = (u + v + |u − v|), belong to V for all u, v ∈ V . Q Q

1 2

u ∧ v = (u + v − |u − v|

230

Riesz spaces

352J

352J Solid subsets (a) If U is a Riesz space, a subset A of U is solid (in the sense of 351I) iff v ∈ A whenever u ∈ A and |v| ≤ |u|. P P (α) If A is solid, u ∈ V and |v| ≤ |u|, then there is some w ∈ A such that −w ≤ u ≤ w; in this case |v| ≤ |u| ≤ w and −w ≤ v ≤ w and v ∈ A. (β) Suppose that A satisfies the condition. If u ∈ A, then |u| ∈ A and −|u| ≤ u ≤ |u|. If w ∈ A and −w ≤ u ≤ w then −u ≤ w, |u| ≤ w = |w| and u ∈ A. Thus A is solid. Q Q In particular, if A is solid, then v ∈ A iff |v| ∈ A. For any set A ⊆ U , the set {u : there is some v ∈ A such that |u| ≤ |v|} is a solid subset of U , the smallest solid set including A; we call it the solid hull of A in U . Any solid linear subspace of U is a Riesz subspace (use 352Fc). If V ⊆ U is a Riesz subspace, then the solid hull of V in U is {u : there is some v ∈ V such that |u| ≤ v} and is a solid linear subspace of U . (b) If T is a Riesz homomorphism from a Riesz space U to a partially ordered linear space V , then its kernel W is a solid linear subspace of U . P P If u ∈ W and |v| ≤ |u|, then T |u| = sup{T u, T (−u)} = 0, while −|u| ≤ v ≤ |u|, so that −0 ≤ T v ≤ 0 and v ∈ W . Q Q Now the quotient space U/W , as defined in 351J, is a Riesz space, and is isomorphic, as partially ordered linear space, to the Riesz space T [U ]. P P Because U/W is the linear space quotient of V by the kernel of the linear operator T , we have an induced linear space isomorphism S : U/W → T [U ] given by setting Su• = T u for every u ∈ U . If p ≥ 0 in U/W there is a u ∈ U + such that u• = p (351J), so that Sp = T u ≥ 0. On the other hand, if p ∈ U/W and Sp ≥ 0, take u ∈ V such that u• = p. We have T (u+ ) = (T u)+ = (Sp)+ = Sp = T u, so that T (u− ) = T u+ − T u = 0 and u− ∈ W , p = (u+ )• ≥ 0. Thus Sp ≥ 0 iff p ≥ 0, and S is a partially-ordered-linear-space isomorphism. Q Q (c) Because a subset of a Riesz space is a solid linear subspace iff it is the kernel of a Riesz homomorphism, such subspaces are sometimes called ideals. 352K Q Products If hUi ii∈I is any family of Riesz spaces, then the product partially ordered linear space U = i∈I Ui (351L) is a Riesz space, with u ∨ v = hu(i) ∨ v(i)ii∈I ,

u ∧ v = hu(i) ∧ v(i)ii∈I ,

|u| = h|u(i)|ii∈I

for all u, v ∈ U . 352L Theorem Let U be any Riesz space. Then there are a set X, a filter F on X and a Riesz subspace of the Riesz space RX |F (351M) which is isomorphic, as Riesz space, to U . proof By 351Q, we can find such X and F and an injective Riesz homomorphism T : U → RX |F. By 352K, or otherwise, RX is a Riesz space; by 352Ib, RX |F is a Riesz space (recall that it is a quotient of RX by a solid linear subspace, as explained in 351M); by 352I, T [U ] is a Riesz subspace of RX |F; and by 352H it is isomorphic to U . 352M Corollary Any identity involving the operations +, −, ∨, ∧, + , and the relation ≤, which is valid in R, is valid in all Riesz spaces.

−

, | | and scalar multiplication,

Remark I suppose some would say that a strict proof of this must begin with a formal description of what the phrase ‘any identity involving the operations. . . ’ means. However I think it is clear in practice what is involved. Given a proposed identity like Pn Pn P 0 ≤ i=0 |αi ||ui | − | i=0 αi ui | ≤ i6=j (|αi | + |αj |)(|ui | ∧ |uj |), (compare 352Fb), then to check that it is valid in all Riesz spaces you need only check (i) that it is true in R (ii) that it is true in RX (iii) that it is true in any RX |F (iv) that it is true in any Riesz subspace of RX |F; and you can hope that the arguments for (ii)-(iv) will be nearly trivial, since (ii) is generally nothing but a coordinate-by-coordinate repetition of (i), and (iii) and (iv) involve only transformations of the formula by Riesz homomorphisms which preserve its structure.

352Oc

Riesz spaces

231

352N Order-density and order-continuity Let U be a Riesz space. (a) Definition A Riesz subspace V of U is quasi-order-dense if for every u > 0 in U there is a v ∈ V such that 0 < v ≤ u; it is order-dense if u = sup{v : v ∈ V, 0 ≤ v ≤ u} for every u ∈ U + . (b) If U is a Riesz space and V is a quasi-order-dense Riesz subspace of U , then the embedding V ⊆ U is order-continuous. P P Let A ⊆ V be a non-empty set such that inf A = 0 in V . ?? If 0 is not the infimum of A in U , then there is a u > 0 such that u is a lower bound for A in U ; now there is a v ∈ V such that 0 < v ≤ u, and v is a lower bound for A in V which is strictly greater than 0. X X Thus 0 = inf A in U . As A is arbitrary, the embedding is order-continuous. Q Q (c) (i) If V ⊆ U is an order-dense Riesz subspace, it is quasi-order-dense. (ii) If V is a quasi-order-dense Riesz subspace of U and W is a quasi-order-dense Riesz subspace of V , then W is a quasi-order-dense Riesz subspace of U . (iii) If V is an order-dense Riesz subspace of U and W is an order-dense Riesz subspace of V , then W is an order-dense Riesz subspace of U . (Use (b).) (iv) If V is a quasi-order-dense solid linear subspace of U and W is a quasi-order-dense Riesz subspace of U then V ∩ W is quasi-order-dense in V , therefore in U . (d) I ought somewhere to remark that a Riesz homomorphism, being a lattice homomorphism, is ordercontinuous iff it preserves arbitrary suprema and infima; compare 313L(b-iv) and (b-v). (e) If V is a Riesz subspace of U , we say that it is regularly embedded in U if the identity map from V to U is order-continuous, that is, whenever A ⊆ V is non-empty and has infimum 0 in V , then 0 is still its greatest lower bound in U . Thus quasi-order-dense Riesz subspaces and solid linear subspaces are regularly embedded. 352O Bands Let U be a Riesz space. (a) Definition A band or normal subspace of U is an order-closed solid linear subspace. (b) If V ⊆ U is a solid linear subspace then it is a band iff sup A ∈ V whenever A ⊆ V + is a non-empty, upwards-directed subset of V with a supremum in U . P P Of course the condition is necessary; I have to show that it is sufficient. (i) Let A ⊆ V be any non-empty upwards-directed set with a supremum in V . Take any u0 ∈ A and set A1 = {u − u0 : u ∈ A, u ≥ u0 }. Then A1 is a non-empty upwards-directed subset of V + , and u0 + A1 = {u : u ∈ A, u ≥ u0 } has the same upper bounds as A, so sup A1 = sup A − u0 is defined in U and belongs to V . Now sup A = u0 + sup A1 also belongs to V . (ii) If A ⊆ V is non-empty, downwards-directed and has an infimum in U , then −A ⊆ V is upwards-directed, so inf A = sup(−A) belongs to V . Thus V is order-closed. Q Q (c) For any set A ⊆ U set A⊥ = {v : v ∈ U, |u| ∧ |v| = 0 for every u ∈ A}. Then A⊥ is a band. P P (i) Of course 0 ∈ A⊥ . (ii) If v, w ∈ A⊥ and u ∈ A, then |u| ∧ |v + w| ≤ (|u| ∧ |v|) + (|u| ∧ |w|) = 0, ⊥

⊥

so v + w ∈ A . (iii) If v ∈ A

and |w| ≤ |v| then 0 ≤ |u| ∧ |w| ≤ |u| ∧ |v| = 0

⊥

for every u ∈ A, so w ∈ A . (iv) If v ∈ A⊥ then nv ∈ A⊥ for every n, by (ii). So if α ∈ R, take n ∈ N such that |α| ≤ n; then |αv| = |α||v| ≤ n|v| ∈ A⊥ and αv ∈ A⊥ . Thus A⊥ is a solid linear subspace of U . (v) If B ⊆ (A⊥ )+ is non-empty and upwards-directed and has a supremum w in U , then |u| ∧ |w| = |u| ∧ w = supv∈B |u| ∧ v = 0 by 352Ea, so w ∈ A⊥ . Thus A⊥ is a band. Q Q

232

Riesz spaces

352Od

(d) For any A ⊆ U , A ⊆ (A⊥ )⊥ . Also B ⊥ ⊆ A⊥ whenever A ⊆ B. So A⊥⊥⊥ ⊆ A⊥ ⊆ A⊥⊥⊥ and A⊥ = A⊥⊥⊥ . (e) If W is another Riesz space and T : U → W is an order-continuous Riesz homomorphism then its kernel is a band. (For {0} is order-closed in W and the inverse image of an order-closed set by an order-continuous order-preserving function is order-closed.) 352P Complemented bands Let U be a Riesz space. A band V ⊆ U is complemented if V ⊥⊥ = V , that is, if V is of the form A⊥ for some A ⊆ U (352Od). In this case its complement is the complemented band V ⊥ . 352Q Theorem In any Riesz space U , the set C of complemented bands forms a Dedekind complete Boolean algebra, with V ∩C W = V ∩ W , 1C = U ,

V ∪C W = (V + W )⊥⊥ ,

0C = {0},

1C \C V = V ⊥ ,

V ⊆ C W ⇐⇒ V ⊆ W for V , W ∈ C. proof To show that C is a Boolean algebra, I use the identification of Boolean algebras with complemented distributive lattices (311L). (a) Of course C is partially ordered by ⊆. If V , W ∈ C then V ∩ W = V ⊥⊥ ∩ W ⊥⊥ = (V ⊥ ∪ W ⊥ )⊥ ∈ C, and V ∩ W must be inf{V, W } in C. The map V 7→ V ⊥ : C → C is an order-reversing bijection, so that V ⊆ W iff W ⊥ ⊆ V ⊥ and V ∨ W = sup{V, W } will be (V ⊥ ∩ W ⊥ )⊥ ; thus C is a lattice. Note also that V ∨ W must be the smallest complemented band including V + W , that is, it is (V + W )⊥⊥ . (b) If V1 , V2 , W ∈ C then (V1 ∨ V2 ) ∧ W = (V1 ∧ W ) ∨ (V2 ∧ W ). P P Of course (V1 ∨ V2 ) ∧ W ⊇ (V1 ∧ W ) ∨ (V2 ∧ W ). ?? Suppose, if possible, that there is a u ∈ (V1 ∨ V2 ) ∩ W \ ((V1 ∩ W ) ∨ (V2 ∩ W )). Then u∈ / ((V1 ∩ W )⊥ ∩ (V2 ∩ W )⊥ )⊥ , so there is a v ∈ (V1 ∩ W )⊥ ∩ (V2 ∩ W )⊥ such that u1 = |u| ∧ |v| > 0. Now / Vj⊥ , and there is a vj ∈ Vj such that u2 = u1 ∧|vj | > 0. / V1⊥ ∩V2⊥ ; say u1 ∈ u1 ∈ V1 ∨V2 = (V1⊥ ∩V2⊥ )⊥ so u1 ∈ ⊥ In this case we still have u2 ∈ (Vj ∩ W ) , because u2 ≤ |v|, but also u2 ∈ Vj and u2 ∈ W because u2 ≤ |u|; but this means that u2 = u2 ∧ u2 = 0, which is absurd. X X Thus (V1 ∨ V2 ) ∧ W ⊆ (V1 ∧ W ) ∨ (V2 ∧ W ) and the two are equal. Q Q (c) Now if V ∈ C, V ∧ V ⊥ = {0} is the least member of C, because if v ∈ V ∩ V ⊥ then |v| = |v| ∧ |v| = 0. By 311L, C has a Boolean algebra structure, with the Boolean relations described; by 312L, this structure is uniquely defined. (d) Finally, if V ⊆ C is non-empty, then T

V=(

S V ∈V

V ⊥ )⊥ ∈ C

and is inf V in C. So C is Dedekind complete. 352R Projection bands Let U be a Riesz space. (a) A projection band in U is a set V ⊆ U such that V + V ⊥ = U . In this case V is a complemented band. P P If v ∈ V ⊥⊥ then v is expressible as v1 + v2 where v1 ∈ V , v2 ∈ V ⊥ . Now |v| = |v1 | + |v2 | ≥ |v2 | (352Fb), so

352S

Riesz spaces

233

|v2 | = |v2 | ∧ |v2 | ≤ |v2 | ∧ |v| = 0 and v = v1 ∈ V . Thus V = V projection band.

⊥⊥

is a complemented band. Q Q Observe that U = V ⊥ + V ⊥⊥ so V ⊥ is also a

(b) Because V ∩ V ⊥ is always {0}, we must have U = V ⊕ V ⊥ for any projection band V ⊆ U ; accordingly there is a corresponding band projection PV : U → U defined by setting P (v + w) = v whenever v ∈ V , w ∈ V ⊥ . In this context I will say that v is the component of v + w in V . The kernel of P is V ⊥ , the set of values is V , and P 2 = P . Moreover, P is an order-continuous Riesz homomorphism. P P (i) P is a linear operator because V and V ⊥ are linear subspaces. (ii) If v ∈ V , w ∈ V ⊥ then |v + w| = |v| + |w|, by 352Fb, so P |v + w| = |v| = |P (v + w)|; consequently P is a Riesz homomorphism (352G). (iii) If A ⊆ U is downwardsdirected and has infimum 0, then P u ≤ u for every u ∈ A, so inf P [A] = 0; thus P is order-continuous. Q Q (c) Note that for any band projection P , and any u ∈ U , we have |P u| ∧ |u − P u| = 0, so that |u| = |P u| + |u − P u| and (in particular) |P u| ≤ |u|; consequently P [W ] ⊆ W for any solid linear subspace W of U. (d) A linear operator P : U → U is a band projection iff P u ∧ (u − P u) = 0 for every u ∈ U + . P P I remarked in (c) that the condition is satisfied for any band projection. Now suppose that P has the property. (i) For any u ∈ U + , P u ≥ 0 and u − P u ≥ 0; in particular, P is a positive linear operator. (ii) If u, v ∈ U + then u − P u ≤ (u + v) − P (u + v), so P v ∧ (u − P u) ≤ P (u + v) ∧ ((u + v) − P (u + v)) = 0 and P v ∧ (u − P u) = 0. (iii) If u, v ∈ U then |P v| ≤ P |v|, |u − P u| ≤ |u| − P |u| (because w 7→ w − P w is a positive linear operator), so |P v| ∧ |u − P u| ≤ P |v| ∧ (|u| − P |u|) = 0. (iv) Setting V = P [U ], we see that u − P u ∈ V ⊥ for every u ∈ U , so that u = u + (u − P u) ∈ V + V ⊥ for every u, and U = V + V ⊥ ; thus V is a projection band. (v) Since P u ∈ V and u − P u ∈ V ⊥ for every u ∈ U , P is the band projection onto V . Q Q 352S Proposition Let U be any Riesz space. (a) The family B of projection bands in U is a subalgebra of the Boolean algebra C of complemented bands in U . (b) For V ∈ B let PV : U → V be the corresponding projection. Then for any e ∈ U + , PV ∩W e = PV e ∧ PW e = PV PW e,

PV ∨W e = PV e ∨ PW e

for all V , W ∈ B. In particular, band projections commute. (c) If V ∈ B then the algebra of projection bands of V is just the principal ideal of B generated by V . proof (a) Of course 0C = {0} ∈ B. If V ∈ B then V ⊥ = 1C \ V belongs to B. If now W is another member of B, then (V ∩ W ) + (V ∩ W )⊥ ⊇ (V ∩ W ) + V ⊥ + W ⊥ . But if u ∈ U then we can express u as v + v 0 , where v ∈ V and v 0 ∈ V ⊥ , and v as w + w0 , where w ∈ W and w0 ∈ W ⊥ ; and as |w| ≤ |v|, we also have w ∈ V , so that u = w + v 0 + w0 ∈ (V ∩ W ) + V ⊥ + W ⊥ . This shows that V ∩ W ∈ B. Thus B is closed under intersection and complements and is a subalgebra of C. (b) If V , W ∈ B and e ∈ U + , we have e = e1 + e2 + e3 + e4 where e1 = PW PV e ∈ V ∩ W ,

e2 = P W ⊥ P V e ∈ V ∩ W ⊥ ,

234

Riesz spaces

e3 = PW PV ⊥ e ∈ V ⊥ ∩ W , e1 + e2 = PV e,

352S

e4 = PW ⊥ PV ⊥ e ∈ V ⊥ ∩ W ⊥ , e1 + e3 = PW e.

Now e2 + e3 + e4 belongs to (V ∩ W )⊥ , so e1 must be the component of e in V ∩ W ; similarly e4 is the component of e in V ⊥ ∩ W ⊥ , and e1 + e2 + e3 is the component of e in V ∨ W . But as e2 ∧ e3 = 0, we have PV ∩W e = e1 = (e1 + e2 ) ∧ (e1 + e3 ) = PV e ∧ PW e, PV ∨W e = e1 + e2 + e3 = (e1 + e2 ) ∨ (e1 + e3 ) = PV e ∨ PW e, as required. It follows that PV PW = PV ∩W = PW ∩V = PW PV . (c) If V , W ∈ B and W ⊆ V , then of course W is a band in the Riesz space V (because V is order-closed in U , so that for any set A ⊆ W its supremum in U will be its supremum in V ). For any v ∈ V , we have an expression of it as w + w0 , where w ∈ W and w0 ∈ W ⊥ , taken in U ; but as |w| + |w0 | = |w + w0 | = |v| ∈ V , w0 belongs to V , and is in WV⊥ , the band in V orthogonal to W . Thus W + WV⊥ = V and W is a projection band in V . Conversely, if W is a projection band in V , then W ⊥ (taken in U ) includes WV⊥ + V ⊥ , so that W + W ⊥ ⊇ W + WV⊥ + V ⊥ = V + V ⊥ = U and W ∈ B. Thus the algebra of projection bands in V is, as a set, equal to the principal ideal BV ; because their orderings agree, or otherwise, their Boolean algebra structures coincide. 352T Products again (a) If U = subspace

Q i∈I

Ui is a product of Riesz spaces, then for any J ⊆ I we have a

VJ = {u : u ∈ U, u(i) = 0 for all i ∈ I \ J} Q of U , canonically isomorphic to i∈J Ui . Each VJ is a projection band, its complement being VI\J ; the map J 7→ VJ is a Boolean homomorphism from PI to the algebra B of projection bands in U , and hV{i} ii∈I is a partition of unity in B. (b) Conversely, if U is a Riesz space and (V0 , . . . , Vn ) is a finite partition Pn of unity in the algebra B of projection bands in U , then every element of U is uniquely expressible as i=0 ui where ui ∈ V Qi for each i. (Induce on n.) This decomposition corresponds to a Riesz space isomorphism between U and i≤n Vi . 352U Quotient spaces (a) If U is a Riesz space and V is a solid linear subspace, then the quotient partially ordered linear space U/V (351J) is a Riesz space; if U and W are Riesz spaces and T : U → W a Riesz homomorphism, then the kernel V of T is a solid linear subspace of U and the Riesz subspace T [U ] of W is isomorphic to U/V (352Jb). (b) Suppose that U is a Riesz space and V a solid linear space. Then the canonical map from U to U/V is order-continuous iff V is a band. P P (i) If u 7→ u• is order-continuous, its kernel V is a band, by 352Oe. (ii) If V is a band, and A ⊆ U is non-empty and downwards-directed and has infimum 0, let p ∈ U/V be any lower bound for {u• : u ∈ A}. Express p as w• . Then ((u − w)− )• = (u• − w• )− = 0, that is, (w − u)+ = (u − w)− ∈ V for every u ∈ A. But this means that w+ = supu∈A (w − u)+ ∈ V ,

p+ = (w+ )• = 0,

that is, p ≤ 0. As p is arbitrary, inf u∈A u• = 0; as A is arbitrary, u 7→ u• is order-continuous. Q Q 352V Principal bands Let U be a Riesz space. Evidently the intersection of any family of Riesz subspaces of U is a Riesz subspace, the intersection of any family of solid linear subspaces is a solid linear subspace, the intersection of any family of bands is a band; we may therefore speak of the band generated by a subset A of U , the intersection of all the bands including A. Now we have the following description of the band generated by a single element.

352W

Riesz spaces

235

Lemma Let U be a Riesz space. (a) If A ⊆ U + is upwards-directed and 2w ∈ A for every w ∈ A, then an element u of U belongs to the band generated by A iff |u| = supw∈A |u| ∧ w. (b) If u ∈ U and w ∈ U + , then u belongs to the band of U generated by w iff |u| = supn∈N |u| ∧ nw. proof (a) Let W be the band generated by A and W 0 the set of elements of U satisfying the condition. (i) If u ∈ W 0 then |u| ∧ w ∈ W for every w ∈ A, because W is a solid linear subspace; because W is also order-closed, |u| and u belong to W . Thus W 0 ⊆ W . (ii) Now W 0 is a band. α) If u ∈ W 0 and |v| ≤ |u| then P P(α supw∈A |v| ∧ w = supw∈A |v| ∧ |u| ∧ w = |v| ∧ supw∈A |u| ∧ w = |v| ∧ |u| = |v| by 352Ea, so v ∈ W 0 . β ) If u, v ∈ W 0 then, for any w1 , w2 ∈ A there is a w ∈ A such that w ≥ w1 ∨ w2 . Now (β w1 + w2 ≤ 2w ∈ A, and (|u| + |v|) ∧ 2w ≥ (|u| ∧ w1 ) + (|v| ∧ w2 ). So any upper bound for {(|u|+|v|)∧w : w ∈ A} must also be an upper bound for {|u|∧w : w ∈ A}+{|v|∧w : w ∈ A} and therefore greater than or equal to sup({|u| ∧ w : w ∈ A} + {|v| ∧ w : w ∈ A}) = supw∈A |u| ∧ w + supw∈A |v| ∧ w = |u| + |v| (351Dc). But this means that supw∈A (|u| + |v|) ∧ w must be |u| + |v|, and |u| + |v| belongs to W 0 ; it follows from (i) that u + v belongs to W 0 . (γγ ) Just as in 352Oc, we now have nu ∈ W 0 for every n ∈ N, u ∈ W 0 , and therefore αu ∈ W 0 for every α ∈ R, u ∈ W 0 , since |αu| ≤ |nu| if |α| ≤ n. Thus W 0 is a solid linear subspace of U . (δδ ) Now suppose that C ⊆ (W 0 )+ has a supremum v in U . Then any upper bound of {v ∧ w : w ∈ A} must also be an upper bound of {u ∧ w : u ∈ C, w ∈ A} and greater than or equal to u = supw∈A u ∧ w for every u ∈ C, therefore greater than or equal to v = sup C. Thus v = supw∈A v ∧ w and v ∈ W 0 . As C is arbitrary, W 0 is a band (352Ob). Q Q (iii) Since A is obviously included in W 0 , W 0 must include W ; putting this together with (i), W = W 0 , as claimed. (b) Apply (a) with A = {nw : n ∈ N}. 352W f -algebras Some of the most important Riesz spaces have multiplicative structures as well as their order and linear structures. A particular class of these structures appears sufficiently often for it to be useful to develop a little of its theory. The following definition is a common approach. (a) Definition An f -algebra is a Riesz space U with a multiplication × : U × U → U such that u × (v × w) = (u × v) × w, (u + v) × w = (u × w) + (v × w),

u × (v + w) = (u × v) + (u × w),

α(u × v) = (αu) × v = u × (αv) for all u, v, w ∈ U and α ∈ R, and u × v ≥ 0 whenever u, v ≥ 0, if u ∧ v = 0 then (u × w) ∧ v = (w × u) ∧ v = 0 for every w ≥ 0. An f -algebra is commutative if u × v = v × u for all u, v.

236

Riesz spaces

352Wb

(b) Let U be an f -algebra. (i) If u ∧ v = 0 in U , then u × v = 0. P P u ∧ (u × v) = 0 so (u × v) ∧ (u × v) = 0. Q Q (ii) u × u ≥ 0 for every u ∈ U . P P (u+ − u− ) × (u+ − u− ) = u+ × u+ − u+ × u− − u− × u+ + u− × u− = u+ × u+ + u− × u− ≥ 0. Q Q (iii) If u, v ∈ U then |u × v| = |u| × |v|. P P u+ × v + , u+ × v − , u− × v + and u+ × u− are dist, so |u × v| = |u+ × v + − u+ × v − − u− × v + + u− × v − | = u+ × v + + u+ × v − + u− × v + + u− × v − = |u| × |v| by 352Fb. Q Q (iv) If v ∈ U + the maps u 7→ u × v, u 7→ v × u : U → U are Riesz homomorphisms. P P The first four clauses of the definition in (a) ensure that they are linear operators. If u ∈ U , then |u| × v = |u × v|,

v × |u| = |v × u|

by (iii), so we have Riesz homomorphisms, by 352G(iii). Q Q (c) Let hUi ii∈I be a family of f -algebras, with Riesz space product U (352K). If we set u × v = hu(i) × v(i)ii∈I for all u, v ∈ U , then U becomes an f -algebra. 352X Basic exercises > (a) Let U be any Riesz space. Show that |u+ − v + | ≤ |u − v| for all u, v ∈ U . > (b) Let U , V be a Riesz spaces and T : U → V a linear operator. Show that the following are equiveridical: (i) T is a Riesz homomorphism; (ii) T (u∨v) = T u∨T v for all u, v ∈ U ; (iii) T (u∧v) = T u∧T v for all u, v ∈ U ; (iv) |T u| = T |u| for every u ∈ U . (c) Let U be a Riesz space and V a solid linear subspace; for u ∈ U write u• for the corresponding element of U/V . Show that if A ⊆ U is solid then {u• : u ∈ A} is solid in U/W . (d) Let U and V be Riesz spaces and T : U → V a Riesz homomorphism with kernel W . Show that if W is a band in U and T [U ] is regularly embedded in V then T is order-continuous. (e) Give U = R2 its lexicographic ordering (351Xa). Show that it has a band V which is not complemented. (f ) Let U be a Riesz space, C the algebra of complemented bands in U . Show that for any V ∈ C the algebra of complemented bands of V is just the principal ideal of C generated by V . > (g) Let U = C([0, 1]) be the space of continuous functions from [0, 1] to R, with its usual linear and order structures, so that it is a Riesz subspace of R[0,1] . Set V = {u : u ∈ U, u(t) = 0 if t ≤ 12 }. Show that V is a band in U and that V ⊥ = {u : u(t) = 0 if t ≥ 21 }, so that V is complemented but is not a projection band. (h) Show that the Boolean homomorphism J 7→ VJ : PI → B of 352Ta is order-continuous. (i) Let U be a Riesz space and A ⊆ U + an upwards-directed set. Show that the band generated by A is {u : |u| = supn∈N,w∈A |u| ∧ nw}. >(j) (i) Let X be any set. Setting (u × v)(x) = u(x)v(x) for u, v ∈ RX , x ∈ X, show that RX is a commutative f -algebra. (ii) With the same definition of ×, show that `∞ (X) is an f -algebra. (iii) If X is a topological space, show that C(X), Cb (X) are f -algebras. (iv) If (X, Σ, µ) is a measure space, show that L0 (µ), L∞ (µ) (§241, §243) are f -algebras.

353A

Archimedean and Dedekind complete Riesz spaces

237

(k) Let U ⊆ RZ be the set of sequences u such that {n : u(n) 6= 0} is bounded above in Z. For u, v ∈ U (i) say that u ≤ v if either P∞ u = v or there is an n ∈ Z such that u(n) < v(n), u(i) = v(i) for every i > n (ii) say that (u ∗ v)(n) = i=−∞ u(i)v(n − i) for every n ∈ Z. Show that U is an f -algebra under this ordering and multiplication. (l) Let U be an f -algebra. (i) Show that any complemented band of U is an ideal in the ring (U, +, ×). (ii) Show that if P : U → U is a band projection, then P (u × v) = P u × P v for every u, v ∈ U . 1 γ

(m) Let U be an f -algebra with multiplicative identity e. Show that u − γe ≤ u2 for every u ∈ U , γ > 0. (Hint: (u+ − γe)2 ≥ 0.) 352 Notes and comments In this section we begin to see a striking characteristic of the theory of Riesz spaces: repeated reflections of results in Boolean algebra. Without spelling out a complete list, I mention the distributive laws (313Bc, 352Ea) and the behaviour of order-continuous homomorphisms (313Pa, 352N, 352Oe, 352Ub, 352Xd). Riesz subspaces correspond to subalgebras, solid linear subspaces to ideals and Riesz homomorphisms to Boolean homomorphisms. We even have a correspondence, though a weaker one, between the representation theorems available; every Boolean algebra is isomorphic to a subalgebra of a power of Z2 (311D-311E), while every Riesz space is isomorphic to a Riesz subspace of a quotient of a power of R (352L). It would be a closer parallel if every Riesz space were embeddable in some RX ; I must emphasize that the differences are as important as the agreements. Subspaces of RX are of great importance, but are by no means adequate for our needs. And of course the details – for instance, the identities in 352D-352F, or 352V – frequently involve new techniques in the case of Riesz spaces. Elsewhere, as in 352G, I find myself arguing rather from the opposite side, when applying results from the theory of general partially ordered linear spaces, which has little to do with Boolean algebra. In the theory of bands in Riesz spaces – corresponding to order-closed ideals in Boolean algebras – we have a new complication in the form of bands which are not complemented, which does not arise in the Boolean algebra context; but it disappears again when we come to specialize to Archimedean Riesz spaces (353B). (Similarly, order-density and quasi-order-density coincide in both Boolean algebras (313K) and Archimedean Riesz spaces (353A).) Otherwise the algebra of complemented bands in a Riesz space looks very like the algebra of order-closed ideals in a Boolean algebra (314Yh, 352Q). The algebra of projection bands in a Riesz space (352S) would correspond, in a Boolean algebra, to the algebra itself. I draw your attention to 352H. The result is nearly trivial, but it amounts to saying that the theory of Riesz spaces will be ‘algebraic’, like the theories of groups or linear spaces, rather than ‘analytic’, like the theories of partially ordered linear spaces or topological spaces, in which we can have bijective morphisms which are not isomorphisms.

353 Archimedean and Dedekind complete Riesz spaces I take a few pages over elementary properties of Archimedean and Dedekind (σ)-complete Riesz spaces. 353A Proposition Let U be an Archimedean Riesz space. Then every quasi-order-dense Riesz subspace of U is order-dense. proof Let V ⊆ U be a quasi-order-dense Riesz subspace, and u ≥ 0 in U . Set A = {v : v ∈ V, v ≤ u}. ?? Suppose, if possible, that u is not the least upper bound of A. Then there is a u1 0 in V such that v ≤ u − u1 . Now nv ≤ u1 for every n ∈ N. P P Induce on n. For n = 0 this is trivial. For the inductive step, given nv ≤ u1 , then (n + 1)v ≤ u1 + v ≤ u, so (n + 1)v ∈ A and (n + 1)v ≤ u1 . Thus the induction proceeds. Q Q But this is impossible, because v > 0 and U is supposed to be Archimedean. X X So u = sup A. As u is arbitrary, V is order-dense.

238

Riesz spaces

353B

353B Proposition Let U be an Archimedean Riesz space. Then (a) for every A ⊆ U , the band generated by A is A⊥⊥ , (b) every band in U is complemented. proof (a) Let V be the band generated by A. Then V is surely included in A⊥⊥ , because this is a band including A (352O). ?? Suppose, if possible, that V 6= A⊥⊥ . Then there is a w ∈ A⊥⊥ \ V , so that |w| ∈ / V. Set B = {v : v ∈ V, v ≤ |w|}; then B is upwards-directed and non-empty. Because V is order-closed, |w| cannot be the supremum of A, and there is a u0 > 0 such that |w|−u0 ≥ v for every v ∈ B. Now u0 ∧|w| 6= 0, so u0 ∈ / A⊥ , and there is a u1 ∈ A such that v = u0 ∧ |u1 | > 0. In this case nv ∈ B for every n ∈ N. P P Induce on n. For n = 0 this is trivial. For the inductive step, given that nv ∈ B, then nv ≤ |w| − u0 so (n + 1)v ≤ nv + u0 ≤ |w|; but also (n + 1)v ≤ nv + |u1 | ∈ V , so (n + 1)v ∈ B. Q Q But this means that nv ≤ |w| for every n, which is impossible, because U is Archimedean. X X (b) Now if V ⊆ U is any band, it is surely the band generated by itself, so is equal to V ⊥⊥ , and is complemented. Remark We may therefore speak of the band algebra of an Archimedean Riesz space, rather than the ‘complemented band algebra’ (352Q). 353C Corollary Let U be an Archimedean Riesz space and v ∈ U . Let V be the band in U generated by v. If u ∈ U , then u ∈ V iff there is no w such that 0 < w ≤ |u| and w ∧ |v| = 0. proof By 353B, V = {v}⊥⊥ . Now, for u ∈ U , u∈ / V ⇐⇒ ∃ w ∈ {v}⊥ , |u| ∧ |w| > 0 ⇐⇒ ∃ w ∈ {v}⊥ , 0 < w ≤ |u|. Turning this round, we have the condition announced. 353D Proposition Let U be an Archimedean Riesz space and V an order-dense Riesz subspace of U . Then the map W 7→ W ∩ V is an isomorphism between the band algebras of U and V . proof If W ⊆ U is a band, then W ∩ V is surely a band in V (it is order-closed in V because it is the inverse image of the order-closed set W under the embedding V ⊆ U , which is order-continuous by 352Nc and 352Nb). If W , W 0 are distinct bands in U , say W 0 6⊆ W , then W 0 6⊆ W ⊥⊥ , by 353B, so W 0 ∩ W ⊥ 6= {0}; because V is order-dense, V ∩ W 0 ∩ W ⊥ 6= {0}, and V ∩ W 0 6= V ∩ W . Thus W 7→ W ∩ V is injective. If Q ⊆ V is a band in V , then its complementary band in V is just Q⊥ ∩ V , where Q⊥ is taken in U . So (because V , like U , is Archimedean, by 351Rc) Q = (Q⊥ ∩ V )⊥ ∩ V = W ∩ V , where W = (Q⊥ ∩ V )⊥ is a band in U . Thus the map W 7→ W ∩ V is an order-preserving bijection between the two band algebras. By 312L, it is a Boolean isomorphism, as claimed. 353E Lemma Let U be an Archimedean Riesz space and V ⊆ U a band such that sup{v : v ∈ V, 0 ≤ v ≤ u} is defined for every u ∈ U + . Then V is a projection band. proof Take any u ∈ U + and set v = sup{v 0 : v 0 ∈ V + , v 0 ≤ u}, w = u − v. v ∈ V because V is a band. Also w ∈ V ⊥ . P P?? If not, there is some v0 ∈ V + such that w ∧ v0 > 0. Now for any n ∈ N we see that nv0 ≤ u =⇒ nv0 ≤ v =⇒ (n + 1)v0 ≤ v + w = u, so an induction on n shows that nv0 ≤ u for every n; which is impossible, because U is supposed to be Archimedean. X XQ Q Accordingly u = v +w ∈ V +V ⊥ . As u is arbitrary, U + ⊆ V +V ⊥ , and V is a projection band (352R). 353F Lemma Let U be an Archimedean Riesz space. If A ⊆ U is non-empty and bounded above and B is the set of its upper bounds, then inf(B − A) = 0. proof ?? If not, let w > 0 be a lower bound for B − A. If u ∈ A and v ∈ B, then v − u ≥ w, that is, u ≤ v − w; as u is arbitrary, v − w ∈ B. Take any u0 ∈ A, v0 ∈ B. Inducing on n, we see that v0 − nw ∈ B for every n ∈ N, so that v0 − nw ≥ u0 , nw ≤ v0 − u0 for every n; but this is impossible, because U is supposed to be Archimedean. X X

353J


239

353G Dedekind completeness Recall from 314A that a partially ordered set P is Dedekind (σ)complete if (countable) non-empty sets with upper and lower bounds have suprema and infima in P . For a Riesz space U , U is Dedekind complete iff every non-empty upwards-directed subset of U + with an upper bound has a least upper bound, and is Dedekind σ-complete iff every non-decreasing sequence in U + with an upper bound has a least upper bound. P P (Compare 314Bc.) (i) Suppose that any non-empty upwardsdirected order-bounded subset of U + has an upper bound, and that A ⊆ U is any non-empty set with an upper bound. Take u0 ∈ A and set B = {u0 ∨ u1 ∨ . . . ∨ un − u0 : u1 , . . . , un ∈ A}. Then B is an upwards-directed subset of U + , and if w is an upper bound of A then w − u0 is an upper bound of B. So sup B is defined in U , and in this case u0 + sup B = sup A. As A is arbitrary, U is Dedekind complete. (ii) Suppose that order-bounded non-decreasing sequences in U + have suprema, and that A ⊆ U is any countable non-empty set with an upper bound. Let hun in∈N be a sequence running over A, and set vn = supi≤n ui − u0 for each n. Then hvn in∈N is a non-decreasing order-bounded sequence in U + , and u0 + supn∈N vn = sup A. (iii) Finally, still supposing that order-bounded non-decreasing sequences in U + have suprema, if A ⊆ U is non-empty, countable and bounded below, inf A will be defined and equal to − sup(−A). Q Q 353H Proposition Let U be a Dedekind σ-complete Riesz space. (a) U is Archimedean. (b) For any v ∈ U the band generated by v is a projection band. (c) If u, v ∈ U , then u is uniquely expressible as u1 + u2 , where u1 belongs to the band generated by v and |u2 | ∧ |v| = 0. proof (a) Suppose that u, v ∈ U are such that nu ≤ v for every n ∈ N. Then nu+ ≤ v + for every n, and A = {nu+ : n ∈ N} is a countable non-empty upwards-directed set with an upper bound; say w = sup A. Since A + u+ ⊆ A, w + u+ = sup(A + u+ ) ≤ w, and u ≤ u+ ≤ 0. As u, v are arbitrary, U is Archimedean. (b) Let V be the band generated by v. Take any u ∈ U + and set A = {v 0 : v 0 ∈ V, 0 ≤ v 0 ≤ u}. Then {u ∧ n|v| : n ∈ N} is a countable set with an upper bound, so has a supremum u1 say in U . Now u1 is an upper bound for A. P P If v 0 ∈ A, then v 0 = supn∈N v 0 ∧ n|v| ≤ u1 by 352V. Q Q Since u ∧ n|v| ∈ A ⊆ V for every n, u1 ∈ V and u1 = sup A. As u is arbitrary, 353E tells us that V is a projection band. (c) Again let V be the band generated by v. Then {v}⊥⊥ is a band containing v, so {v} ⊆ V ⊆ {v}⊥⊥ ,

{v}⊥ ⊇ V ⊥ ⊇ {v}⊥⊥⊥ = {v}⊥

(352Od), and V ⊥ = {v}⊥ . Now, if u ∈ U , u is uniquely expressible in the form u1 + u2 where u1 ∈ V and u2 ∈ V ⊥ , by (b). But u2 ∈ V ⊥ ⇐⇒ u2 ∈ {v}⊥ ⇐⇒ |u2 | ∧ |v| = 0. So we have the result. 353I Proposition In a Dedekind complete Riesz space, all bands are projection bands. proof Use 353E, noting that the sets {v : v ∈ V, 0 ≤ v ≤ u} there are always non-empty, upwards-directed and bounded above, so always have suprema. 353J Proposition (a) Let U be a Dedekind σ-complete Riesz space. (i) If V is a solid linear subspace of U , then V is (in itself) Dedekind σ-complete. (ii) If W is a sequentially order-closed Riesz subspace of U then W is Dedekind σ-complete. (iii) If V is a sequentially order-closed solid linear subspace of U , the canonical map from U to V is sequentially order-continuous, and the quotient Riesz space U/V is also Dedekind σ-complete. (b) Let U be a Dedekind complete Riesz space.

240

Riesz spaces

353J

(i) If V is a solid linear subspace of U , then V is (in itself) Dedekind complete. (ii) If W ⊆ U is an order-closed Riesz subspace then W is Dedekind complete. proof (a)(i) If hun in∈N is a non-decreasing sequence in V + with an upper bound v ∈ V , then w = supn∈N un is defined in U ; but as 0 ≤ w ≤ v, w ∈ V and w = supn∈N un in V . Thus V is Dedekind σ-complete. (ii) If hun in∈N is a non-decreasing order-bounded sequence in W , then u = supn∈N un is defined in U ; but because W is sequentially order-closed, u ∈ W and u = supn∈N un in W . (iii) Let hun in∈N be a non-decreasing sequence in U with supremum u. Then of course u• is an upper bound for A = {u•n : n ∈ N} in U/V . Now let p be any other upper bound for A. Express p as v • . Then for each n ∈ N we have u•n ≤ p, so that (un − v)+ ∈ V . Because V is sequentially order-closed, (u − v)+ = supn∈N (un − v)+ ∈ V and u• ≤ p. Thus u• is the least upper bound of A. Similarly, if hun in∈N is a non-increasing sequence in U with infimum u, then u• = inf n∈N u•n in U/V . Thus u 7→ u• is sequentially order-continuous. Now suppose that hpn in∈N is a non-decreasing sequence in (U/V )+ with an upper bound p ∈ (U/V )+ . Let u ∈ U + be such that u• = p, and for each n ∈ N let un ∈ U + be such that u•n = pn . Set vn = u ∧ supi≤n ui for each n; then vn• = pn for each n, and hvn in∈N is a non-decreasing order-bounded sequence in U . Set v = supn∈N vn ; by the last paragraph, v • = supn∈N pn in U/V . As hpn in∈N is arbitrary, U/V is Dedekind σ-complete, as claimed. (b) The argument is the same as parts (i) and (ii) of the proof of (a). 353K Proposition Let U be a Riesz space and V a quasi-order-dense Riesz subspace of U which is (in itself) Dedekind complete. Then V is a solid linear subspace of U . proof Suppose that v ∈ V , u ∈ U and that |u| ≤ |v|. Consider A = {w : w ∈ V, 0 ≤ w ≤ u+ }. Then A is a non-empty subset of V with an upper bound in V (viz., |v|). So A has a supremum v0 in V . Because the embedding V ⊆ U is order-continuous (352N), v0 is the supremum of A in U . But as V is order-dense (353A), v0 = u+ and u+ ∈ V . Similarly, u− ∈ V and u ∈ V . As u, v are arbitrary, V is solid. 353L Order units Let U be a Riesz space. (a) An element e of U + is an order unit in U if U is the solid linear subspace of itself generated by e; that is, if for S every u ∈ U there is an n ∈ N such that |u| ≤ ne. (For the solid linear subspace generated by v ∈ U + is n∈N [−nv, nv].) (b) An element e of U + is a weak order unit in U if U is the principal band generated by e; that is, if u = supn∈N u ∧ ne for every u ∈ U + (352Vb). Of course an order unit is a weak order unit. (c) If U is Archimedean, then an element e of U + is a weak order unit iff {e}⊥⊥ = U (353B), that is, iff {e}⊥ = {0} (because {e}⊥ = {0} =⇒ {e}⊥⊥ = {0}⊥ = U =⇒ {e}⊥ = {e}⊥⊥⊥ = U ⊥ = {0},) that is, iff u ∧ e > 0 whenever u > 0. 353M Theorem Let U be an Archimedean Riesz space with order unit e. Then it can be embedded as an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdorff space, in such a way that e corresponds to χX; moreover, this embedding is essentially unique. Remark Here C(X) is the space of all continuous functions from X to R; because X is compact, they are all bounded, so that χX is an order unit in C(X) = Cb (X). proof (a) Let X be the set of Riesz homomorphisms x from U to R such that x(e) = 1. Define T : U → RX by setting (T u)(x) = x(u) for x ∈ X, u ∈ U ; then it is easy to check that T is a Riesz homomorphism, just because every member of X is a Riesz homomorphism, and of course T e = χX.

353M


241

(b) The key to the proof is the fact that X separates the points of U , that is, that T is injective. I choose the following method to show this. Suppose that w ∈ U and w > 0. Because U is Archimedean, there is a δ > 0 such that (w − δe)+ 6= 0. Now there is an x ∈ X such that x(w) ≥ δ. P P (i) By 351O, there is a solid linear subspace V of U such that (w − δe)+ ∈ / V and whenever u ∧ v = 0 in U then one of u, v belongs to V . (ii) Because V 6= U , e ∈ / V , so no non-zero multiple of e can belong to V . Also observe that if u, v ∈ U \ V , then one of (u − v)+ , (v − u)+ must belong to V , while neither u = u ∧ v + (u − v)+ nor v = u ∧ v + (v − u)+ does; so u ∧ v ∈ / V . (iii) For each u ∈ U set Au = {α : α ∈ R, (u − αe)+ ∈ V }. Then α ≥ β ∈ Au =⇒ 0 ≤ (u − αe)+ ≤ (u − βe)+ ∈ V =⇒ α ∈ Au . Also Au is non-empty and bounded below, because if α ≥ 0 is such that −αe ≤ u ≤ αe then α ∈ Au and −α − 1 ∈ / Au (since (u − (−α − 1)e)+ ≥ e ∈ / V ). (iv) Set x(u) = inf Au for every u ∈ U ; then α ∈ Au for every α > x(u), α ∈ / Au for every α < x(u). (v) If u, v ∈ U and α > x(u), β > x(v) then ((u + v) − (α + β)e)+ ≤ (u − αe)+ + (v − βe)+ ∈ V (352Fc), so α + β ∈ Au+v ; as α and β are arbitrary, x(u + v) ≤ x(u) + x(v). (vi) If u, v ∈ U and α < x(u), β < x(v) then ((u + v) − (α + β)e)+ ≥ (u − αe)+ ∧ (v − βe)+ ∈ / V, using (ii) of this argument and 352Fc, so α + β ∈ / Au+v . As α and β are arbitrary, x(u + v) ≥ x(u) + x(v). (vii) Thus x : U → R is additive. (viii) If u ∈ U , γ > 0 then α ∈ Au =⇒ (γu − αγe)+ = γ(u − αe)+ ∈ V =⇒ γα ∈ Aγu ; thus Aγu ⊇ γAu ; similarly, Au ⊇ γ −1 Aγu so Aγu = γAu and x(γu) = γx(u). (ix) Consequently x is linear, since we know already from (vii) that x(0u) = 0.x(u), x(−u) = −x(u). (x) If u ≥ 0 then u + αe ≥ αe ∈ /V for every α > 0, that is, −α ∈ / Au for every α > 0, and x(u) ≥ 0; thus x is a positive linear functional. (xi) If u ∧ v = 0, then one of u, v belongs to V , so min(x(u), x(v)) ≤ 0 and (using (x)) min(x(u), x(v)) = 0; thus x is a Riesz homomorphism (352G(iv)). (xii) Ae = [1, ∞[ so x(e) = 1. Thus x ∈ X. (xiii) δ ∈ / Aw so x(w) ≥ δ. Q Q (c) Thus T w 6= 0 whenever w > 0; consequently |T w| = T |w| 6= 0 whenever w 6= 0, and T is injective. I now have to define the topology of X. This is just the subspace topology on X if we regard X as a subset of RU with its product topology. To see that X is compact, observe that if for each u ∈ U we choose an Q αu such that |u| ≤ αu e, then X is a subspace of Q = u∈U [−αu , αu ]. Because Q is a product of compact spaces, it is compact, by Tychonoff’s theorem (3A3J). Now X is a closed subset of Q. P P X is just the intersection of the sets {x : x(u + v) = x(u) + x(v)}, {x : x(u+ ) = max(x(u), 0)},

{x : x(αu) = αx(u)}, {x : x(e) = 1}

as u, v run over U and α over R; and each of these is closed, so X is an intersection of closed sets and therefore itself closed. Q Q Consequently X also is compact. Moreover, the coordinate functionals x 7→ x(u) are continuous on Q, therefore on X also, that is, T u : X → R is a continuous function for every u ∈ U . Note also that because Q is a product of Hausdorff spaces, Q and X are Hausdorff (3A3Id). (d) So T is a Riesz homomorphism from U to C(X). Now T [U ] is a Riesz subspace of C(X), containing χX, and such that if x, y ∈ X are distinct there is an f ∈ T [U ] such that f (x) 6= f (y) (because there is surely a u ∈ U such that x(u) 6= y(u)). By the Stone-Weierstrass theorem (281A), T [U ] is k k∞ -dense in C(X). Consequently it is also order-dense. P P If f > 0 in C(X), set ² = 13 kf k∞ , and let u ∈ U be such that + kf − T uk∞ ≤ ²; set v = (u − ²e) . Since 0 < (f − 2²χX)+ ≤ (T u − ²χX)+ ≤ f + = f , 0 < T v ≤ f . As f is arbitrary, T [U ] is quasi-order-dense, therefore order-dense (353A). Q Q (e) I have still to show that the representation is (essentially) unique. Suppose, then, that we have another representation of U as a norm-dense Riesz subspace of C(Z), with e this time corresponding to χZ; to simplify the notation, let us suppose that U is actually a subspace of C(Z). Then for each z ∈ Z,

242

Riesz spaces

353M

we have a functional zˆ : U → R defined by setting zˆ(u) = u(z) for every u ∈ U ; of course zˆ is a Riesz homomorphism such that zˆ(e) = 1, that is, zˆ ∈ X. Thus we have a function z 7→ zˆ : Z → X. For any u ∈ U , the function z 7→ zˆ(u) = u(z) is continuous, so the function z 7→ zˆ is continuous (3A3Ib). If z1 , z2 are distinct of Z, there is an f ∈ C(Z) such that f (z1 ) 6= f (z2 ) (3A3Bf); now there is a u ∈ U such that kf − uk∞ ≤ 31 |f (z1 ) − f (z2 )|, so that u(z1 ) 6= u(z2 ) and zˆ1 6= zˆ2 . Thus z 7→ zˆ is injective. Finally, it is also surjective. P P Suppose that x ∈ X. Set V = {u : u ∈ U, x(u) = 0}; then V is a solid linear subspace of U (352Jb), not containing e. For z ∈ V + set Gv = {z : v(z) > 1}. Because e ∈ / V , Gv 6= Z. G = {Gv : v ∈ V + } is an upwards-directed family of open setsSin Z, not containing Z; consequently, because Z is compact, G cannot be an open cover of Z. Take z ∈ Z \ G. Then v(z) ≤ 1 for every v ∈ V + ; because α|v| ∈ V + whenever v ∈ V , α ≥ 0, we must have v(z) = 0 for every v ∈ V . Now, given any u ∈ U , consider v = u − x(u)e. Then x(v) = 0 so v ∈ V and v(z) = 0, that is, u(z) = (v + x(u)e)(z) = v(z) + x(u)e(z) = x(u). As u is arbitrary, zˆ = x; as x is arbitrary, we have the result. Q Q Thus z 7→ zˆ is a continuous bijection from the compact Hausdorff space Z to the compact Hausdorff space X; it must therefore be a homeomorphism (3A3Dd). This argument shows that if U is embedded as a norm-dense Riesz subspace of C(Z), where Z is compact and Hausdorff, then Z must be homeomorphic to X. But it shows also that a homeomorphism is canonically defined by the embedding; z ∈ Z corresponds to the Riesz homomorphism u 7→ u(z) in X. 353N Lemma Let U be a Riesz space, V an Archimedean Riesz space and S, T : U → V Riesz homomorphisms such that Su ∧ T u0 = 0 in V whenever u ∧ u0 = 0 in U . Set W = {u : Su = T u}. Then W is a solid linear subspace of U ; if S and T are order-continuous, W is a band. proof (a) It is easy to check that, because S and T are Riesz homomorphisms, W is a Riesz subspace of U . (b) If w ∈ W and 0 ≤ u ≤ w in U , then Su ≤ T u. P P?? Otherwise, set e = Sw = T w, and let Ve be the solid linear subspace of V generated by e, so that Ve is an Archimedean Riesz space with order unit, containing both Su and T u. By 353M (or its proof), there is a Riesz homomorphism x : Ve → R such that x(e) = 1 and x(Su) > x(T u). Take α such that x(Su) > α > x(T u), and consider u0 = (u − αw)+ , u00 = (αw − u)+ . Then x(Su0 ) = max(0, x(Su) − αx(Sw)) = max(0, x(Su) − α) > 0, x(T u00 ) = max(0, αx(T w) − x(T u)) = max(0, α − x(T u)) > 0, so x(Su0 ∧ T u00 ) = min(x(Su0 ), x(T u00 )) > 0 and Su0 ∧ T u00 > 0, while u0 ∧ u00 = 0. X XQ Q Similarly, T u ≤ Su and u ∈ W . As u and w are arbitrary, W is a solid linear subspace. (c) Finally, suppose that S and T are order-continuous, and that A ⊆ W is a non-empty upwards-directed set with supremum u in U . Then Su = sup S[A] = sup T [A] = T u and u ∈ W . As u and A are arbitrary, W is a band (352Ob). 353O f -algebras I give two results on f -algebras, intended to clarify the connexions between the multiplicative and lattice structures of the Riesz spaces in Chapter 36. Proposition Let U be an Archimedean f -algebra (352W). Then (a) the multiplication is separately order-continuous in the sense that the maps u 7→ u × w, u 7→ w × u are order-continuous for every w ∈ U + ; (b) the multiplication is commutative. proof (a) Let A ⊆ U be a non-empty set with infimum 0, and v0 ∈ U + a lower bound for {u × w : u ∈ A}. Fix u0 ∈ A. If u ∈ A and δ > 0, then v0 ∧ (u0 − 1δ u)+ ≤ δu0 × w. P P Set v = v0 ∧ (u0 − 1δ u)+ . Then

353P


243

δv ∧ (u − δu0 )+ ≤ (δu0 − u)+ ∧ (u − δu0 )+ = 0, so v ∧ (u − δu0 )+ = 0 and v ∧ ((u − δu0 )+ × w) = 0. But v ≤ v0 ≤ u × w ≤ (u − δu0 )+ × w + δu0 × w, so v ≤ ((u − δu0 )+ × w) ∧ v + (δu0 × w) ∧ v ≤ δu0 × w, by 352Fa. Q Q Taking the infimum over u, and using the distributive laws (352E), we get v0 ∧ u0 ≤ δu0 × w. Taking the infimum over δ, and using the hypothesis that U is Archimedean, v0 ∧ u0 = 0. But this means that v0 ∧ (u0 × w) = 0, while v0 ≤ u0 × w, so v0 = 0. As v0 is arbitrary, inf u∈A u × w = 0; as A is arbitrary, u 7→ u × w is order-continuous. Similarly, u 7→ w × u is order-continuous. (b)(i) Fix v ∈ U + , and for u ∈ U set Su = u × v,

T u = v × u.

Then S and T are both order-continuous Riesz homomorphisms from U to itself (352W(b-iv) and (a) above). Also, Su ∧ T u0 = 0 whenever u ∧ u0 = 0. P P 0 = (u × v) ∧ u0 = (u × v) ∧ (v × u0 ). Q Q So W = {u : u × v = v × u} is a band in U (353N). Of course v ∈ W (because Sv = T v = v 2 ). If u ∈ W ⊥ , then v ∧ |u| = 0 so Su = T u = 0 (352W(b-i)), and u ∈ W ; but this means that W ⊥ = {0} and W = W ⊥⊥ = U (353Bb). Thus v × u = u × v for every u ∈ U . (ii) This is true for every v ∈ U + . Of course it follows that v × u = u × v for every u, v ∈ U , so that multiplication is commutative. 353P Proposition Let U be an Archimedean f -algebra with multiplicative identity e. (a) e is a weak order unit in U . (b) If u, v ∈ U then u × v = 0 iff |u| ∧ |v| = 0. (c) If u ∈ U has a multiplicative inverse u−1 then |u| also has a multiplicative inverse; if u ≥ 0 then u−1 ≥ 0 and u is a weak order unit. (d) If V is another Archimedean f -algebra with multiplicative identity e0 , and T : U → V is a positive linear operator such that T e = e0 , then T is a Riesz homomorphism iff T (u × v) = T u × T v for all u, v ∈ U . proof (a) e = e2 ≥ 0 by 352W(b-iii). If u ∈ U and e ∧ |u| = 0 then |u| = (e × |u|) ∧ |u| = 0; by 353Lc, e is a weak order unit. (b) If |u| ∧ |v| = 0 then u × v = 0, by 352W(b-ii). If w = |u| ∧ |v| > 0, then w2 ≤ |u| × |v|. Let n ∈ N be such that nw 6≤ e, and set w1 = (nw − e)+ , w2 = (e − nw)+ . Then 0 6= w1 = w1 × e = w1 × w2 + w1 × (e ∧ nw) = w1 × (e ∧ nw) ≤ (nw)2 ≤ n2 |u| × |v| = n2 |u × v|, so u × v 6= 0. (c) u × u−1 = e so |u| × |u−1 | = |e| = e (352W(b-iii)), and |u−1 | = |u|−1 . (Recall that inverses in any semigroup with identity are unique, so that we need have no inhibitions in using the formulae u−1 , |u|−1 .) Now suppose that u ≥ 0. Then u−1 = |u−1 | ≥ 0. If u ∧ |v| = 0 then e ∧ |v| = (u × u−1 ) ∧ |v| = 0, so v = 0; accordingly u is a weak order unit. (d)(i) If T is multiplicative, and u ∧ v = 0 in U , then T u × T v = T (u × v) = 0 and T u ∧ T v = 0, by (b). So T is a Riesz homomorphism, by 352G.

244

Riesz spaces

353P

(ii) Accordingly I shall henceforth assume that T is a Riesz homomorphism and seek to show that it is multiplicative. If u, v ∈ U + , then T (u × v) and T u × T v both belong to the band generated by T u. P P Write W for this band. (α) For any n ≥ 1 we have (v − ne)2 ≥ 0, that is, 2nv ≤ v 2 + n2 e, so n(v − ne) ≤ 2nv − n2 e ≤ v 2 . Consequently 1 n

T (u × v) − nT u = T (u × v) − nT (u × e) = T (u × (v − ne)) ≤ T (u × v 2 ) because v 0 7→ T (u × v 0 ) is a positive linear operator; as V is Archimedean, inf n∈N (T (u × v) − nT u)+ = 0 and T (u × v) = supn∈N T (u × v) ∧ nT u belongs to W . (β) If w ∧ |T u| = 0 then w ∧ |T u × T v| = w ∧ (|T u| × |T v|) = 0; so T u × T v ∈ W ⊥⊥ = W . Q Q (iii) Fix v ∈ U + . For u ∈ U , set S1 u = T u × T v and S2 u = T (u × v). Then S1 and S2 are both Riesz homomorphisms from U to V . If u ∧ u0 = 0 in U , then S1 u ∧ S2 u0 = 0 in V , because (by (ii) just above) S1 u belongs to the band generated by T u, while S2 u0 belongs to the band generated by T u0 , and T u ∧ T u0 = T (u ∧ u0 ) = 0. By 353N, W = {u : S1 u = S2 u} is a solid linear subspace of U . Of course it contains e, since S1 e = T e × T v = e0 × T v = T v = T (e × v) = S2 e. In fact u ∈ W for every u ∈ U + . P P As noted in (ii) just above, u − ne ≤

1 2 nu

for every n ≥ 1. So

|S1 u − S2 u| = |S1 (u − ne)+ + S1 (u ∧ ne) − S2 (u − ne)+ − S2 (u ∧ ne)| 1 n

≤ S1 (u − ne)+ + S2 (u − ne)+ ≤ (S1 u2 + S2 u2 ) for every n ≥ 1, and |S1 u − S2 u| = 0, that is, S1 u = S2 u. Q Q So W = U , that is, T u × T v = T (u × v) for every u ∈ U . And this is true for every v ∈ U + . It follows at once that it is true for every v ∈ U , so that T is multiplicative, as claimed. 353X Basic exercises > (a) Let U be a Riesz space in which every band is complemented. Show that U is Archimedean. (b) A Riesz space U has the principal projection property iff the band generated by any single member of U is a projection band. Show that any Dedekind σ-complete Riesz space has the principal projection property, and that any Riesz space with the principal projection property is Archimedean. > (c) Fill in the missing part (b-iii) of 353J. (d) Let U be an Archimedean f -algebra with an order-unit which is a multiplicative identity. Show that U can be identified, as f -algebra, with a subspace of C(X) for some compact Hausdorff space X. 353Y Further exercises (a) Let U be a Riesz space in which every quasi-order-dense solid linear subspace is order-dense. Show that U is Archimedean. (b) Let X be a completely regular Hausdorff space. Show that C(X) is Dedekind complete iff Cb (X) is Dedekind complete iff X is extremally disconnected. (c) Let X be a compact Hausdorff space. Show that C(X) is Dedekind σ-complete iff G is open for every cozero set G ⊆ X. (Cf. 314Yf.) Show that in this case X is zero-dimensional. (d) Let U be an Archimedean Riesz space such that {un : n ∈ N} has a supremum in U whenever hun in∈N is a sequence in U such that um ∧un = 0 whenever m 6= n. Show that U has the principal projection property, but need not be Dedekind σ-complete.

353 Notes


245

(e) Let U be an Archimedean Riesz space. Show that the following are equiveridical: (i) U has the countable sup property (241Yd) (ii) for every A ⊆ U there is a countable B ⊆ A such that A and B have the same upper bounds; (iii) every dist subset of U + is countable. (f ) Let U be an Archimedean Riesz space with order unit e, and k ke the corresponding norm. Let Z be the unit ball of U ∗ . Show that for a linear functional f : U → R the following are equiveridical: (i) f is an extreme point of Z, that is, f ∈ Z and Z \ {f } is convex (ii) |f (e)| = 1 and one of f , −f is a Riesz homomorphism. (g) Let U be an Archimedean f -algebra. Show that an element e of U is a multiplicative identity iff e2 = e and e is a weak order unit. (Hint: start by showing that under these conditions, e × u = 0 ⇒ u = 0.) (h) Let U be an Archimedean f -algebra with a multiplicative identity. Show that if u ∈ U then u is invertible iff |u| is invertible. 353 Notes and comments As in the last section, many of the results above have parallels in the theory of Boolean algebras; thus 353A corresponds to 313K, 353G corresponds in part to remarks in 314Bc and 314Xa, and 353J corresponds to 314C-314E. Riesz spaces are more complicated; for instance, principal ideals in Boolean algebras are straightforward, while in Riesz spaces we have to distinguish between the solid linear subspace generated by an element and the band generated by the same element. Thus an ‘order unit’ in a Boolean ring would just be an identity, while in a Riesz space we must distinguish between ‘order unit’ and ‘weak order unit’. As this remark may suggest to you, (Archimedean) Riesz spaces are actually closer in spirit to arbitrary Boolean rings than to the Boolean algebras we have been concentrating on so far; to the point that in §361 below I will return briefly to general Boolean rings. Note that the standard definition of ‘order-dense’ in Boolean algebras, as given in 313J, corresponds to the definition of ‘quasi-order-dense’ in Riesz spaces (352Na); the point here being that Boolean algebras behave like Archimedean Riesz spaces, in which there is no need to make a distinction. I give the representation theorem 353M more for completeness than because we need it in any formal sense. In 351Q and 352L I have given representation theorems for general partially ordered linear spaces, and general Riesz spaces, as quotients of spaces of functions; in 368F below I give a theorem for Archimedean Riesz spaces corresponding rather more closely to the expressions of the Lp spaces as quotients of spaces of measurable functions. In 353M, by contrast, we have a theorem expressing Archimedean Riesz spaces with order units as true spaces of functions, rather than as spaces of equivalence classes of functions. All these theorems are important in forming an appropriate mental picture of ordered linear spaces, as in 352M. I give a bare-handed proof of 353M, using only the Riesz space structure of C(X); if you know a little about extreme points of dual unit balls you can approach from that direction instead, using 353Yf. The point is that (as part (d) of the proof makes clear) the space X can be regarded as a subset of the normed space dual U ∗ of U with its weak* topology. In this treatise generally, and in the present chapter in particular, I allow myself to be slightly prejudiced against normed-space methods; you can find them in any book on functional analysis, and I prefer here to develop techniques like those in part (b) of the proof of 353M, which will be a useful preparation for such theorems as 368E. There is a very close analogy between 353M and the Stone representation of Boolean algebras (311E, 311I-311K). Just as the proof of 311E looked at the set of ring homomorphisms from A to the elementary Boolean algebra Z2 , so the proof of 353M looks at Riesz homomorphisms from U to the elementary M -space R. Later on, the most important M -spaces, from the point of view of this treatise, will be the L∞ spaces of §363, explicitly defined in of Stone representations (363A). Of the two parts of 353O, it is (a) which is most important for the purposes of this book. The f -algebras we shall encounter in Chapter 36 can be seen to be commutative for different, and more elementary, reasons. The (separate) order-continuity of multiplication, however, is not always immediately obvious. Similarly, the uniferent Riesz homomorphisms we shall encounter can generally be seen to be multiplicative without relying on the arguments of 353Pd.

246

Riesz Spaces

§354 intro.

354 Banach lattices The next step is a brief discussion of norms on Riesz spaces. I start with the essential definitions (354A, 354D) with the principal properties of general Riesz norms (354B-354C) and order-continuous norms (354E). I then describe two of the most important classes of Banach lattice: M -spaces (354F-354L) and L-spaces (354M-354R), with their elementary properties. For M -spaces I give the basic representation theorem (354K-354L), and for L-spaces I give a note on uniform integrability (354P-354R). 354A Definitions (a) If U is a Riesz space, a Riesz norm or lattice norm on U is a norm k k such that kuk ≤ kvk whenever |u| ≤ |v|; that is, a norm such that k|u|k = kuk for every u and kuk ≤ kvk whenever 0 ≤ u ≤ v. (b) A Banach lattice is a Riesz space with a Riesz norm under which it is complete. Remark We have already seen many examples of Banach lattices; I list some in 354Xa below. 354B Lemma Let U be a Riesz space with a Riesz norm k k. (a) U is Archimedean. (b) The maps u 7→ |u| and u 7→ u+ are uniformly continuous. (c) For any u ∈ U , the sets {v : v ≤ u} and {v : v ≥ u} are closed; in particular, the positive cone of U is closed. (d) Any band in U is closed. (e) If V is a norm-dense Riesz subspace of U , then V + = {v : v ∈ V, v ≥ 0} is norm-dense in the positive cone U + of U . proof (a) If u, v ∈ U are such that nu ≤ v for every n ∈ N, then nu+ ≤ v + so nku+ k ≤ kv + k for every n, and ku+ k = 0, that is, u+ = 0 and u ≤ 0. As u, v are arbitrary, U is Archimedean. (b) For any u, v ∈ U , ||u| − |v|| ≤ |u − v| (352D), so k|u| − |v|k ≤ ku − vk; thus u 7→ |u| is uniformly continuous. Consequently u 7→ 12 (u + |u|) = u+ is uniformly continuous. (c) Now {v : v ≤ u} = {v : (v − u)+ = 0} is closed because the function v 7→ (v − u)+ is continuous and {0} is closed. Similarly {v : v ≥ u} = {v : (u − v)+ = 0} is closed. (d) If V ⊆ U is a band, then V = V ⊥⊥ (353Bb), that is, V = {v : |v| ∧ |w| = 0 for every w ∈ V ⊥ }. Because the function v 7→ |v| ∧ |w| = 21 (|v| + |w| − ||v| − |w||) is continuous, all the sets {v : |v| ∧ |w| = 0} are closed, and so is their intersection V . (e) Observe that V + = {v + : v ∈ V } and U + = {u+ : u ∈ U }; recall that u 7→ u+ is continuous, and apply 3A3Eb. 354C Lemma If U is a Banach lattice and hu Pn∞in∈N is a sequence in U such that supn∈N un is defined in U , with k supn∈N un k ≤ n=0 kun k.

P∞ n=0

kun k < ∞, then

proof Set vn = supi≤n ui for each n. Then 0 ≤ vn+1 − vn ≤ (un+1 − un )+ ≤ |un+1 − un | for each n ∈ N, so

P∞ n=0

kvn+1 − vn k ≤

P∞ n=0

kun+1 − un k ≤

P∞ n=0

kun+1 k + kun k

is finite, and hvn in∈N is Cauchy. Let u be its limit; because hvn in∈N is non-decreasing, and the sets {v : v ≥ vn } are all closed, u ≥ vn for each n ∈ N. On the other hand, if v ≥ vn for every n, then (u − v)+ = limn→∞ (vn − v)+ = 0, and u ≤ v. So u = supn∈N vn = supn∈N un is the required supremum.

354E

Banach lattices

To estimate its norm, observe that |vn | ≤ for the inductive step), so that

Pn i=0

|ui | for each n (induce on n, using the last item in 352D

kuk = limn→∞ kvn k ≤

354D

247

P∞ i=0

k|ui |k =

P∞ i=0

kui k.

I come now to the basic properties according to which we classify Riesz norms.

Definitions (a) A Fatou norm on a Riesz space U is a Riesz norm on U such that whenever A ⊆ U + is non-empty and upwards-directed and has a least upper bound in U , then k sup Ak = supu∈A kuk. (Observe that, once we know that k k is a Riesz norm, we can be sure that kuk ≤ k sup Ak for every u ∈ A, so that all we shall need to check is that k sup Ak ≤ supu∈A kuk.) (b) A Riesz norm on a Riesz space U has the Levi property if every upwards-directed norm-bounded set is bounded above. (c) A Riesz norm on a Riesz space U is order-continuous if inf u∈A kuk = 0 whenever A ⊆ U is a non-empty downwards-directed set with infimum 0. 354E Proposition Let U be a Riesz space with an order-continuous Riesz norm k k. (a) If A ⊆ U is non-empty and upwards-directed and has a supremum, then sup A ∈ A. (b) k k is Fatou. (c) If A ⊆ U is non-empty and upwards-directed and bounded above, then for every ² > 0 there is a u ∈ A such that k(v − u)+ k ≤ ² for every v ∈ A. (d) Any non-decreasing order-bounded sequence in U is Cauchy. (e) If U is a Banach lattice it is Dedekind complete. (f) Every order-dense Riesz subspace of U is norm-dense. proof (a) Suppose that A ⊆ U is non-empty and upwards-directed and has a least upper bound u0 . Then B = {u0 − u : u ∈ A} is downwards-directed and has infimum 0. So inf u∈A ku0 − uk = 0, and u0 ∈ A. (b) If, in (a), A ⊆ U + , then we must have ku0 k ≤ inf u∈A kuk + ku − u0 k ≤ supu∈A kuk. As A is arbitary, k k is a Fatou norm. (c) Let B be the set of upper bounds for A. Then B is downwards-directed; because A is upwardsdirected, B − A = {v − u : v ∈ B, u ∈ A} is downwards-directed. By 353F, inf(B − A) = 0. So there are w ∈ B, u ∈ A such that kw − uk ≤ ². Now if v ∈ A, (v − u)+ = (v ∨ u) − u ≤ w − u, so k(v − u)+ k ≤ ². (d) If hun in∈N is a non-decreasing order-bounded sequence, and ² > 0, then, applying (c) to {un : n ∈ N}, we find that there is an m ∈ N such that kum − un k ≤ ² whenever m ≥ n. (e) Now suppose that U is a Banach lattice. Let A ⊆ U be any non-empty set with an upper bound. Set A0 = {u0 ∨ . . . ∨ un : u0 , . . . , un ∈ A}, so that A0 is upwards-directed and has the same upper bounds as A. For each n, choose un ∈ A0 such that k(u − un )+ k ≤ 2−n for every u ∈ A0 . Set vn = supi≤n ui for each n; then vn ∈ A0 and kvm − vn k ≤ k(vm − un )+ k ≤ 2−n for all m ≥ n. So hvn in∈N is Cauchy and has a limit v say. If u ∈ A, then k(u − v)+ k = limn→∞ k(u − vn )+ k = 0, so u ≤ v; while if w is any upper bound for A, then k(v − w)+ k = limn→∞ k(vn − w)+ k = 0 and v ≤ w. Thus v = sup A and A has a supremum. (f ) If V is an order-dense Riesz subspace of U and u ∈ U + , set A = {v : v ∈ V, v ≤ u}. Then A is upwards-directed and has supremum u, so u ∈ A ⊆ V , by (a). Thus U + ⊆ V ; it follows at once that U = U+ − U+ ⊆ V .

248

Riesz Spaces

354F

354F Lemma If U is an Archimedean Riesz space with an order unit e (definition: 353L), there is a Riesz norm k ke defined on U by the formula kuke = min{α : |u| ≤ αe} for every u ∈ U . proof This is a routine verification. Because e is so always has an infimum α0 say; now |u| − α0 e ≤ |u| − α0 e ≤ 0 and |u| ≤ α0 e, so that the minimum subadditivity and homogeneity of k ke are immediate

an order-unit, {α : |u| ≤ αe} is always non-empty, δe for every δ > 0, so (because U is Archimedean) is attained. In particular, kuke = 0 iff u = 0. The from the facts that |u + v| ≤ |u| + |v|, |αu| = |α||u|.

354G Definitions (a) If U is an Archimedean Riesz space and e an order unit in U , the norm k ke as defined in 354F is the order-unit norm on U associated with e. (b) An M -space is a Banach lattice in which the norm is an order-unit norm. (c) If U is an M -space, its standard order unit is the order unit e such that k ke is the norm of U . (To see that e is uniquely defined, observe that it is sup{u : u ∈ U, kuk ≤ 1}.) 354H Examples (a) For any set X, `∞ (X) is an M -space with standard order unit χX. (As remarked in 243Xl, the completeness of `∞ (X) can be regarded as the special case of 243E in which X is given counting measure.) (b) For any topological space X, the space Cb (X) of bounded continuous real-valued functions on X is an M -space with standard order unit χX. (It is a Riesz subspace of `∞ (X) containing the order unit of `∞ (X), therefore in its own right an Archimedean Riesz space with order unit. To see that it is complete, it is enough to observe that it is closed in `∞ (X) because a uniform limit of continuous functions is continuous (3A3Nb).) (c) For any measure space (X, Σ, µ), the space L∞ (µ) is an M -space with standard order unit χX • . 354I Lemma Let U be an Archimedean Riesz space with order unit e, and V a subset of U which is dense for the order-unit norm k ke . Then for any u ∈ U there are sequences hvn in∈N , hwn in∈N in V such that vn ≤ vn+1 ≤ u ≤ wn+1 ≤ wn and kwn − vn ke ≤ 2−n for every n; so that u = supn∈N vn = inf n∈N wn in U. If V is a Riesz subspace of U , and u ≥ 0, we may suppose that vn ≥ 0 for every n. Consequently V is order-dense in U . proof For each n ∈ N, take vn , wn ∈ V such that ku −

3 e − vn ke 2n+3

≤

1 , 2n+3

ku +

3 e − wn ke 2n+3

≤

1 . 2n+3

Then u−

1 e 2n+1

≤ vn ≤ u −

1 e 2n+2

≤u≤u+

1 e 2n+2

≤ wn ≤ u +

1 e. 2n+1

Accordingly hvn in∈N is non-decreasing, hwn in∈N is non-increasing and kwn −vn ke ≤ 2−n for every n. Because U is Archimedean, supn∈N vn = inf n∈N wn = u. If V is a Riesz subspace of U , then replacing vn by vn+ if necessary we may suppose that every vn is non-negative; and V is order-dense by the definition in 352N. 354J Proposition Let U be an Archimedean Riesz space with an order unit e. Then k ke is Fatou and has the Levi property. proof This is elementary. If A ⊆ U + is non-empty, upwards-directed and norm-bounded, then it is bounded above by αe, where α = supu∈A kuke . This is all that is called for in the Levi property. If moreover sup A is defined, then sup A ≤ αe so k sup Ak ≤ α, as required in the Fatou property.

354O

Banach lattices

249

354K Theorem Let U be an Archimedean Riesz space with order unit e. Then it can be embedded as an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdorff space, in such a way that e corresponds to χX and k ke corresponds to k k∞ ; moreover, this embedding is essentially unique. proof This is nearly word-for-word a repetition of 353M. The only addition is the mention of the norms. But let X and T : U → C(X) be as in 353M. Then, for any u ∈ U , |u| ≤ kuke e, so that |T u| = T |u| ≤ kuke T e = kuke χX, and kT uk∞ ≤ kuke . On the other hand, if 0 < δ < kuke then u1 = (|u| − δe)+ > 0, so that T u1 = (|T u| − δχX)+ > 0 and kT uk∞ ≥ δ; as δ is arbitrary, kT uk∞ ≥ kuke . 354L Corollary Any M -space U is isomorphic, as Banach lattice, to C(X) for some compact Hausdorff X, and the isomorphism is essentially unique. X can be identified with the set of Riesz homomorphisms x : U → R such that x(e) = 1, where e is the standard order unit of U , with the topology induced by the product topology on RU . proof By 354K, there are a compact Hausdorff space X and an embedding of U as a norm-dense Riesz subspace of C(X) matching k ke to k k∞ . Since U is complete under k ke , its image is closed in C(X) (3A4Ff), and must be the whole of C(X). The expression is unique just in so far as the expression of 353M/354K is unique. In particular, we may, if we wish, take X to be the set of normalized Riesz homomorphisms from U to R, as in the proof of 353M. Remark If U is an M -space, then the construction of 353M represents U as C(X), where X is the set of uniferent Riesz homomorphisms from U to R; this is sometimes called the spectrum of U . 354M I come now to a second fundamental class of Banach lattices, in a strong sense ‘dual’ to the class of M -spaces, as will appear in §356. Definition An L-space is a Banach lattice U such that ku + vk = kuk + kvk whenever u, v ∈ U + . Example If (X, Σ, µ) is any measure space, then L1 (µ), with its norm k k1 , is an L-space (242D, 242F). In particular, taking µ to be counting measure on N, `1 is an L-space (242Xa). 354N Theorem If U is an L-space, then its norm is order-continuous and has the Levi property. proof (a) Both of these are consequences of the following fact: if A ⊆ U is norm-bounded and non-empty and upwards-directed, then sup A is defined in U and belongs to the norm-closure of A in U . P P For u ∈ A, set γ(u) = sup{kv − uk : v ∈ A, v ≥ u}. We surely have γ(u) ≤ kuk + supv∈A kvk < ∞. Choose a sequence hun in∈N in A such that un+1 ≥ un and kun+1 − un k ≥ 12 γ(un ) for each n. Then kun+1 − u0 k =

Pn i=0

kui+1 − ui k ≥

1 2

Pn i=0

γ(ui )

P∞

for every n, using the definition of ‘L-space’. Because A is bounded, i=0 γ(ui ) < ∞ and limn→∞ γ(un ) = 0. But kum − un k ≤ γ(un ) whenever m ≥ n, so hun in∈N is Cauchy and has a limit u∗ in U . For each n ∈ N, u∗ ≥ un because um ≥ un for every m ≥ n (see 354Bc). If u ∈ A, n ∈ N then there is a u0 ∈ A such that u0 ≥ u ∨ un ; now (u − u∗ )+ ≤ u0 − un so k(u − u∗ )+ k ≤ ku0 − un k ≤ γ(un ); as n is arbitrary, k(u − u∗ )+ k = 0 and u ≤ u∗ . Thus u∗ is an upper bound for A. But if v is any upper bound for A, then un ≤ v for every n so u∗ ≤ v. Thus u∗ is the least upper bound of A; and u∗ ∈ A because it is the norm limit of hun in∈N . Q Q (b) This shows immediately that the norm has the Levi property. But also it must be order-continuous. P P If A ⊆ U is non-empty and downwards-directed and has infimum 0, take any u0 ∈ A and consider B = {u0 − u : u ∈ A, u ≤ u0 }. Then B is upwards-directed and has supremum u0 , so u0 ∈ B and inf u∈A kuk ≤ inf v∈B ku0 − vk = 0. Q Q 354O Proposition If U is an L-space and V is a norm-closed Riesz subspace of U , then V is an L-space in its own right. In particular, any band of U is an L-space.

250

Riesz Spaces

354O

proof For any Riesz subspace V of U , we surely have ku + vk = ku| + kvk whenever u, v ∈ V + ; so if V is norm-closed, therefore a Banach lattice, it must be an L-space. But in any Banach lattice, a band is norm-closed (354Bd), so a band in an L-space is again an L-space. 354P Uniform integrability in L-spaces Some of the ideas of §246 can be readily expressed in this abstract context. Definition Let U be an L-space. A set A ⊆ U is uniformly integrable if for every ² > 0 there is a w ∈ U + such that k(|u| − w)+ k ≤ ² for every u ∈ A. 354Q Since I have already used the phrase ‘uniformly integrable’ based on a different formula, I had better check instantly that the two definitions are consistent. Proposition If (X, Σ, µ) is any measure space, then a subset of L1 = L1 (µ) is uniformly integrable in the sense of 354P iff it is uniformly integrable in the sense of 246A. proof (a) If A ⊆ L1 is uniformly integrable in the sense of 246A, then for any ² > 0 there are M ≥ 0, R E ∈ Σ such that µE < ∞ and (|u| − M χE • )+ ≤ ² for every u ∈ A; now w = M χE • belongs to (L1 )+ and k(|u| − w)+ k ≤ ² for every u ∈ A. As ² is arbitrary, A is uniformly integrable in the sense of 354P. (b) Now suppose that A is uniformly integrable in the sense of 354P. Let ² > 0. Then there is a w ∈ (L1 )+ such that k(|u| − w)+ k ≤ 21 ² for every u ∈ A. There is a simple function h : X → R such that kw − h• k ≤ 21 ² (242M); now take E = {x : h(x) 6= 0}, M = supx∈X |h(x)| (I over the trivial case X = ∅), so that h ≤ M χE and (|u| − M χE • )+ ≤ (|u| − w)+ + (w − M χE • )+ ≤ (|u| − w)+ + (w − h• )+ ,

R

(|u| − M χE • )+ ≤ k(|u| − w)+ k + kw − h• k ≤ ²

for every u ∈ A. As ² is arbitrary, A is uniformly integrable in the sense of 354P. 354R

I give abstract versions of the easiest results from §246.

Theorem Let U be an L-space. (a) If A ⊆ U is uniformly integrable, then (i) A is norm-bounded; (ii) every subset of A is uniformly integrable; (iii) for any α ∈ R, αA is uniformly integrable; (iv) there is a uniformly integrable, solid, convex, norm-closed set C ⊇ A; (v) for any other uniformly integrable set B ⊆ U , A ∪ B and A + B are uniformly integrable. (b) For any set A ⊆ U , the following are equiveridical: (i) A is uniformly integrable; (ii) limn→∞ (|un | − supi 0 there are u0 , . . . , un ∈ A such that k(|u| − supi≤n |ui |)+ k ≤ ² for every u ∈ A; (iv) A is norm-bounded and any dist sequence in the solid hull of A is norm-convergent to 0. (c) If V ⊆ U is a closed Riesz subspace then a subset of V is uniformly integrable when regarded as a subset of V iff it is uniformly integrable when regarded as a subset of U . R proof (a)(i) There must be a w ∈ U + such that (|u| − w)+ ≤ 1 for every u ∈ A; now |u| ≤ |u| − w + |w| ≤ (|u| − w)+ + |w|,

kuk ≤ k(|u| − w)+ k + kwk ≤ 1 + kwk

for every u ∈ A, so A is norm-bounded. (ii) This is immediate from the definition. (iii) Given ² > 0, we can find w ∈ U + such that |α|k(|u| − w)+ k ≤ ² for every u ∈ A; now k(|v| − |α|w)+ k ≤ ² for every v ∈ αA. (iv) If A is empty, take C = A. Otherwise, try

354R

Banach lattices

251

C = {v : v ∈ U, k(|v| − w)+ k ≤ supu∈A k(|u| − w)+ k for every w ∈ U + }. Evidently A ⊆ C, and C satisfies the definition 354M because A does. The functionals v 7→ k(|v| − w)+ k : U → R are all continuous for k k (because the operators v 7→ |v|, v 7→ v − w, v 7→ v + , v 7→ kvk are continuous), so C is closed. If |v 0 | ≤ |v| and v ∈ C, then k(|v 0 | − w)+ k ≤ k(|v| − w)+ k ≤ supu∈A k(|u| − w)+ k for every w, and v 0 ∈ C. If v = αv1 + βv2 where v1 , v2 ∈ C, α ∈ [0, 1] and β = 1 − α, then |v| ≤ α|v1 | + β|v2 |, so |v| − w ≤ (α|v1 | − αw) + (β|v2 | − βw) ≤ (α|v1 | − αw)+ + (β|v2 | − βw)+ and (|v| − w)+ ≤ α(|v1 | − w)+ + β(|v2 | − w)+ for every w; accordingly k(|v| − w)+ k ≤ αk(|v1 | − w)+ k + βk(|v2 | − w)+ k ≤ (α + β) sup k(|u| − w)+ k = sup k(|u| − w)+ k u∈A

u∈A

for every w, and v ∈ C. Thus C has all the required properties. (v) I show first that A ∪ B is uniformly integrable. P P Given ² > 0, let w1 , w2 ∈ U + be such that k(|u| − w1 )+ k ≤ ² for every u ∈ A,

k(|u| − w2 )+ k ≤ ² for every u ∈ B.

Set w = w1 ∨ w2 ; then k(|u| − w)+ k ≤ ² for every u ∈ A ∪ B. As ² is arbitrary, A ∪ B is uniformly integrable. Q Q Now (iv) tells us that there is a convex uniformly integrable set C including A ∪ B, and in this case A + B ⊆ 2C, so A + B is also uniformly integrable, using (ii) and (iii). (b)(i)⇒(ii)&(iv) Suppose that A is uniformly integrable and that hun in∈N is any sequence in the solid hull of A. Set vn = supi≤n |ui | for n ∈ N and v00 = v0 = |u0 |,

vn0 = vn − vn−1 = (|un | − supi
for n ≥ 1. Given ² > 0, there is a w ∈ U + such that k(|u| − w)+ k ≤ ² for every u ∈ A, and therefore for every u in the solid hull of A. Of course supn∈N kvn ∧ wk ≤ kwk is finite, so there is an n ∈ N such that kvi ∧ wk ≤ ² + kvn ∧ wk for every i ∈ N. But now, for m > n, 0 vm ≤ (|um | − vn )+ ≤ (|um | − |um | ∧ w)+ + ((|um | ∧ w) − vn )+

≤ (|um | − w)+ + (vm ∧ w) − (vn ∧ w), so that 0 kvm k ≤ k(|um | − w)+ k + k(vm ∧ w) − (vn ∧ w)k

= k(|um | − w)+ k + kvm ∧ wk − kvn ∧ wk ≤ 2², using the L-space property of the norm for the equality in the middle. As ² is arbitrary, limn→∞ vn0 = 0. As hun in∈N is arbitrary, condition (ii) is satisfied; but so is condition (iv), because we know from (a-i) that A is norm-bounded, and if hun in∈N is dist then vn0 = |un | for every n, so that in this case limn→∞ un = 0. (ii)⇒(iii)⇒(i) are elementary. not-(i)⇒not-(iv) Now suppose that A is not uniformly integrable. If it is not norm-bounded, we can stop. Otherwise, there is some ² > 0 such that supu∈A k(|u| − w)+ k > ² for every w ∈ U + . Consequently we shall be able to choose inductively a sequence hun in∈N in A such that k(|un | − 2n supi ² for P∞ −i every n ≥ 1. Because A is norm-bounded, i=0 2 kui k is finite, and we can set

252

Riesz Spaces

354R

P∞ vn = (|un | − 2n supi (a) Work through pPr the proofs that the following are all Banach lattices: (i) Pr 2 R with (α) kxk1 = i=1 |ξi | (β) kxk2 = i=1 |ξi | (γ) kxk∞ = maxi≤r |ξi |, where x = (ξ1 , . . . , ξr ). (ii) p ` (X), for any set X and any p ∈ [1, ∞] (242Xa, 243Xl, 244Xn). (iii) Lp (µ), for any measure space (X, Σ, µ) and any p ∈ [1, ∞] (242F, 243E, 244G). (iv) c 0 , the space of sequences convergent to 0, with the norm k k∞ inherited from `∞ . r

(b) Let hUi ii∈I be any family of Banach lattices. Write U for their Riesz space product (352K), and in U set P kuk1 = i∈I ku(i)k, V1 = {u : kuk1 < ∞}, kuk∞ = supi∈I ku(i)k,

V∞ = {u : kuk∞ < ∞}.

Show that V1 , V∞ are solid linear subspaces of U and are Banach lattices under their norms k k1 , k k∞ . (c) Let U be a Riesz space with a Riesz norm. Show that the maps (u, v) 7→ u∧v, (u, v) 7→ u∨v : U ×U → U are uniformly continuous. > (d) Let U be a Riesz space with a Riesz norm. (i) Show that any order-bounded set in U is normbounded. (ii) Show that in Rr , with any of the standard Riesz norms (354Xa(i)), norm-bounded sets are order-bounded. (iii) Show that in `1 (N) there is a sequence converging to 0 (for the norm) which is not orderbounded. (iv) Show that in c 0 any sequence converging to 0 is order-bounded, but there is a norm-bounded set which is not order-bounded. (e) Let U be a Riesz space with a Riesz norm. Show that it is a Banach lattice iff non-decreasing Cauchy sequences are convergent. (Hint: if kun+1 − un k ≤ 2−n for every n, show that hsupi≤n ui in∈N is Cauchy, and that hun in∈N converges to inf n∈N supm≥n um .) (f ) Let U be a Riesz space with a Riesz norm. Show that U is a Banach lattice iff every non-decreasing Cauchy sequence hun in∈N in U + has a least upper bound u with kuk = limn→∞ kun k. (g) Let U be a Banach lattice. Suppose that B ⊆ U is solid and supn∈N un ∈ B whenever hun in∈N is a non-decreasing sequence in B with a supremum in U . Show that B is closed. (Hint: show first that u ∈ B whenever there is a sequence hun in∈N in B ∩ U + such that ku −un k ≤ 2−n for every n; do this by considering vm = inf n≥m un .) (h) Let U be any Riesz space with a Riesz norm. Show that the Banach space completion of U (3A5Ib) has a unique partial ordering under which it is a Banach lattice. >(i) Show that the space c0 of sequences convergent to 0, with k k∞ , is a Banach lattice with an ordercontinuous norm which does not have the Levi property.

354Yc

Banach lattices

253

> (j) Show that `∞ , with k k∞ , is a Banach lattice with a Fatou norm which has the Levi property but is not order-continuous. (k) Let U be a Riesz space with a Fatou norm. Show that if V ⊆ U is a regularly embedded Riesz subspace (definition: 352Ne) then the induced norm on V is a Fatou norm. (l) Let U be a Riesz space and k k a Riesz norm on U which is order-continuous in the sense of 354Dc. Show that it is order-continuous in the sense of 313H when regarded as a function from U + to [0, ∞[. (m) Let U be a Riesz space with an order-continuous norm. Show that if V ⊆ U is a regularly embedded Riesz subspace then the induced norm on V is order-continuous. (n) Let U be a Dedekind σ-complete Riesz space with a Fatou norm which has the Levi property. Show that it is a Banach lattice. (Hint: 354Xf.) Q (o) Let hUi ii∈I be any family of Banach lattices and let V1 , V∞ be the subspaces of U = i∈I Ui as described in 354Xb. Show that V1 , V∞ have norms which are Fatou, or have the Levi property, iff every Ui has. Show that the norm of V1 is order-continuous iff the norm of every Ui is. (p) Let U be a Banach lattice with an order-continuous norm. Show that a Riesz subspace of U (indeed, any sublattice of U ) is norm-closed iff it is order-closed in the sense of 313D, and in this case is itself a Banach lattice with an order-continuous norm. > (q) Let U be an M -space and V a norm-closed Riesz subspace of U containing the standard order unit of U . (i) Show that V , with the induced norm, is an M -space. (ii) Deduce that the space c of convergent sequences is an M -space if given the norm k k∞ inherited from `∞ . (r) Show that a Banach lattice U is an M -space iff (i) its norm is a Fatou norm with the Levi property (ii) ku ∨ vk = max(kuk, kvk) for all u, v ∈ U + . >(s) Describe a topological space X such that the space c of convergent sequences (354Xq) can be identified with C(X). (t) Let D ⊆ R be any non-empty set, and PnV the space of functions f : D → R of bounded variation (§224). For f ∈ V set kf k = sup{|f (t0 )| + i=1 |f (ti ) − f (ti−1 )| : t0 ≤ t1 ≤ . . . ≤ tn in D} (224Yb). Let V + be the set of non-negative, non-decreasing functions in V . Show that V + is the positive cone of V for a Riesz space ordering under which V is an L-space. 354Y Further exercises (a) Let U be a Riesz space with a Riesz norm, and V a norm-dense Riesz subspace of U . Suppose that the induced norm on V is Fatou, when regarded as a norm on the Riesz space V . Show (i) that V is order-dense in U (ii) that the norm of U is Fatou. (Hint: for (i), show that if u ∈ U + , vn ∈ V + and ku − vn k ≤ 2−n−2 kuk for every n, then kv0 − inf i≤n vi k ≥ 41 kuk for every n, so that 0 cannot be inf n∈N vn in V .) (b) Let U be a Riesz space with a Riesz norm. Show that the following are equiveridical: (i) limn→∞ un = 0 whenever hun in∈N is a dist order-bounded sequence in U + (ii) limn→∞ un+1 − un = 0 for every orderbounded non-decreasing sequence hun in∈N in U (iii) whenever A ⊆ U + is a non-empty downwards-directed set in U + with infimum 0, inf u∈A supv∈A,v≤u ku − vk = 0. (Hint: for (i)⇒(ii), show by induction that limn→∞ un = 0 whenever hun in∈N is an order-bounded sequence such that, for some fixed k, inf i∈K ui = 0 for every K ⊆ N of size k; now show that if hun in∈N is non-decreasing and 0 ≤ un ≤ u for every n, then inf i∈K (ui+1 − ui − k1 u)+ = 0 whenever K ⊆ N, #(K) = k ≥ 1. For (iii)⇒(i), set A = {u : ∃ n, u ≥ ui ∀ i ≥ n}. See Fremlin 74a, 24H.) (c) Show that any Riesz space with an order-continuous norm has the countable sup property (definition: 241Yd).

254

Riesz Spaces

354Yd

(d) Let U be a Banach lattice. Show that the following are equiveridical: (i) the norm on U is ordercontinuous; (ii) U satisfies the conditions of 354Yb; (iii) every order-bounded monotonic sequence in U is Cauchy. (e) Let U be a Riesz space with a Fatou norm. Show that the norm on U is order-continuous iff it satisfies the conditions of 354Yb. R (f ) For f ∈ C([0, 1]), set kf k1 = |f (x)|dx. Show that k k1 is a Riesz norm on C([0, 1]) satisfying the conditions of 354Yb, but is not order-continuous. (g) Let U be a Riesz space with a Riesz norm k k. Show that (U, k k) satisfies the conditions of 354Yb iff the norm of its completion is order-continuous. (h) Let U be a Riesz space with a Riesz norm, and V ⊆ U a norm-dense Riesz subspace such that the induced norm on V is order-continuous. Show that the norm of U is order-continuous. (Hint: use 354Ya.) (i) Let U be an Archimedean Riesz space. For any e ∈ U + , let Ue be the solid linear subspace of U generated by e, so that e is an order unit in Ue , and let k ke be the corresponding order-unit norm on Ue . We say that U is uniformly complete if Ue is complete under k ke for every e ∈ U + . (i) Show that any Banach lattice is uniformly complete. (ii) Show that any Dedekind σ-complete Riesz space is uniformly complete (cf. 354Xn). (iii) Show that if U is a uniformly complete Riesz space with a Riesz norm which has the Levi property, then U is a Banach lattice. (iv) Show that if U is a Banach lattice then a set A ⊆ U is closed, for the norm topology, iff A ∩ Ue is k ke -closed for every e ∈ U + . (v) Let V be a solid linear subspace of U . Show that the quotient Riesz space U/V (352U) is Archimedean iff V ∩ Ue is k ke -closed for every e ∈ U + . (vi) Show that if U is uniformly complete and V ⊆ U is a solid linear subspace such that U/V is Archimedean, then U/V is uniformly complete. (vii) Show that U is Dedekind σ-complete iff it is uniformly complete and has the principal projection property (353Xb). (Hint: for (vii), use 353Yc.) (j) Let U be a Banach lattice such that ku + vk = kuk + kvk whenever u ∧ v = 0. Show that U is an L-space. (Hint: by 354Yd, the norm is order-continuous, so U is Dedekind complete. If u, v ≥ 0, set e = u + v, and represent Ue as C(X) where X is extremally disconnected (353Yb); now approximate u and v by functions taking only finitely many values to show that ku + vk = kuk + kvk.) (k) Let U be a uniformly complete Archimedean Riesz space (354Yi). Set UC = U × U with the complex linear structure defined by identifying (u, v) ∈ U × U as u + iv ∈ UC , so that u = Re(u + iv), v = Im(u + iv) and (α + iβ)(u + iv) = (αu − βv) + i(αv + βu). (i) Show that for w ∈ UC we can define |w| ∈ U by setting |w| = sup|ζ|=1 Re(ζw). (ii) Show that if U is a uniformly complete Riesz subspace of RX for some set X, then we can identify UC with the linear subspace of CX generated by U . (iii) Show that |w + w0 | ≤ |w| + |w0 |, |γw| = |γ||w| for all w ∈ UC , γ ∈ C. (iv) Show that if w ∈ UC and |w| ≤ u1 + u2 , where u1 , u2 ∈ U + , then w is expressible as w1 + w2 where |wj | ≤ uj for both j. (Hint: set e = u1 + u2 and represent Ue as C(X).) (v) Show that if U0 is a solid linear subspace of U , then, for w ∈ UC , |w| ∈ U0 iff Re w, Im w both belong to U0 . (vi) Show that if U has a Riesz norm then we have a norm on UC defined by setting kwk = k|w|k, and that if U is a Banach lattice then UC is a (complex) Banach space. (vii) Show that if U = Lp (µ), where (X, Σ, µ) is a measure space and p ∈ [1, ∞], then UC can be identified with LpC (µ) as defined in 242P, 243K, 244O. (We may call UC the complexification of the Riesz space U .) (l) Let (X, Σ, µ) be a measure space and V a Banach lattice. Write L1V for the space of Bochner integrable functions from conegligible subsets of X to V , and L1V for the corresponding set of equivalence classes (253Yf). (i) Show that L1V is a Banach lattice under the ordering defined by saying that f • ≤ g • iff f (x) ≤ g(x) in V for µ-almost every x ∈ X. (ii) Show that when V = L1 (ν), for some other measure space (Y, T, ν), then this ordering on L1V agrees with the ordering of L1 (λ) where λ is the (c.l.d.) product measure on X × Y and we identify L1V with L1 (λ), as in 253Yi. (iii) Show that if V has an order-continuous norm, so has L1V . (Hint: 354Yd(ii).) (iv) Show that if µ is Lebesgue measure on [0, 1] and V = `∞ , then L1V is not Dedekind σ-complete.

355A

Spaces of linear operators

255

354 Notes and comments Apart from some of the exercises, the material of this section is pretty strictly confined to ideas which will be useful later in this volume. The basic Banach lattices of measure theory are the Lp spaces of Chapter 24; these all have Fatou norms with the Levi property (244Ye-244Yf), and for p < ∞ their norms are order-continuous (244Yd). In Chapter 36 I will return to these spaces in a more abstract context. Here I am mostly concerned to establish a vocabulary in which their various properties, and the relationships between these properties, can be expressed. In normed Riesz spaces we have a very rich mixture of structures, and must take particular care over the concepts of ‘boundedness’, ‘convergence’ and ‘density’, which have more than one possible interpretation. In particular, we must scrupulously distinguish between ‘order-bounded’ and ‘norm-bounded’ sets. I have not yet formally introduced any of the various concepts of order-convergence (see §367), but I think that even so it is best to get into the habit of reminding oneself, when a convergent sequence appears, that it is convergent for the norm topology, rather than in any sense related directly to the order structure. I should perhaps warn you that for the study of M -spaces 354L is not as helpful as it may look. The trouble is that apart from a few special cases (as in 354Xs) the topological space used in the representation is actually more complicated and mysterious than the M -space it is representing. After the introduction of M -spaces, this section becomes a natural place for ‘uniformly complete’ spaces (354Yi). For the moment I leave these in the exercises. But I mention them now because they offer a straightforward route towards a theory of ‘complex Riesz spaces’ (354Yk). In large parts of functional analysis it is natural, and in some parts it is necessary, to work with normed spaces over C rather than over R, and for L2 spaces in particular it is useful to have a proper grasp of the complex case. And while the insights offered by the theory of Riesz spaces are not especially important in such areas, I think we should always seek connexions between different topics. So it is worth ing that uniformly complete Riesz spaces have complexifications. I shall have a great deal more to say about L-spaces when I come to spaces of additive functionals (§362) and to L1 spaces again (§365) and to linear operators on them (§371); and before that, there will be something in the next section on their duals, and on L-spaces which are themselves dual spaces. For the moment I just give some easy results, direct translations of the corresponding facts in §246, which have natural expressions in the language of this section, holding deeper ideas over. In particular, the characterization of uniformly integrable sets as relatively weakly compact sets (247C) is valid in general L-spaces (356Q). For an extensive treatment of Banach lattices, going very much deeper than I have space for in this volume, see Lindenstrauss & Tzafriri 79. For a careful exposition of a great deal of useful information, see Schaefer 74.

355 Spaces of linear operators We come now to a discussion of linear operators between Riesz spaces. Of course linear operators are central to any kind of functional analysis, and a feature of the theory of Riesz spaces is the way the order structure picks out certain classes of operators for special consideration. Here I concentrate on positive and order-continuous operators, with a brief mention of sequential order-continuity. It turns out, in fact, that we need to work with operators which are differences of positive operators or of order-continuous positive operators. I define the basic spaces L∼ , L× and L∼ c (355A, 355G), with their most important properties (355B, 355E, 355H-355I) and some remarks on the special case of Banach lattices (355C, 355K). At the same time I give an important theorem on extension of operators (355F) and a corollary (355J). The most important case is of course that in which the codomain is R, so that our operators become real-valued functionals; I shall come to these in the next section. 355A Definition Let U and V be Riesz spaces. A linear operator T : U → V is order-bounded if T [A] is order-bounded in V for every order-bounded A ⊆ U . I will write L∼ (U ; V ) for the set of order-bounded linear operators from U to V .

256

Riesz spaces

355B

355B Lemma If U and V are Riesz spaces, (a) a linear operator T : U → V is order-bounded iff {T u : 0 ≤ u ≤ w} is bounded above in V for every w ∈ U +; (b) in particular, any positive linear operator from U to V belongs to L∼ = L∼ (U ; V ); (c) L∼ is a linear space; (d) if W is another Riesz space and T : U → V and S : V → W are order-bounded linear operators, then ST : U → W is order-bounded. proof (a) This is elementary. If T ∈ L∼ and w ∈ U + , [0, w] is order-bounded, so its image must be orderbounded in V , and in particular bounded above. On the other hand, if T satisfies the condition, and A is order-bounded, then A ⊆ [u1 , u2 ] for some u1 ≤ u2 , and T [A] ⊆ T [u1 + [0, u2 − u1 ]] = T u1 + T [[0, u2 − u1 ]] is bounded above; similarly, T [−A] is bounded above, so T [A] is bounded below; as A is arbitrary, T is order-bounded. (b) If T is positive then {T u : 0 ≤ u ≤ w} is bounded above by T w for every w ≥ 0, so T ∈ L∼ . (c) If T1 , T2 ∈ L∼ , α ∈ R and A ⊆ U is order-bounded, then there are v1 , v2 ∈ V such that Ti [A] ⊆ [−vi , vi ] for both i. Setting v = (1 + |α|)v1 + v2 , (αT1 + T2 )[A] ⊆ [−v, v]; as A is arbitrary, αT1 + T2 belongs to L∼ ; as α, T1 , T2 are arbitrary, and since the zero operator surely belongs to L∼ , L∼ is a linear subspace of the space of all linear operators from U to V . (d) This is immediate from the definition; if A ⊆ U is order-bounded, then T [A] ⊆ V and (ST )[A] = S[T [A]] ⊆ W are order-bounded. 355C Theorem If U and V are Banach lattices then every order-bounded linear operator (in particular, every positive linear operator) from U to V is continuous. proof ?? Suppose, if possible, that T : U → V is an order-bounded linear operator which is not continuous. Then for each n ∈ N we can find a un ∈ U such that kun k ≤ 2−n but kT un k ≥ n. Now u = supn∈N |un | is defined in U (354C), and there is a v ∈ V such that −v ≤ T w ≤ v whenever −u ≤ w ≤ u; but this means that kvk ≥ kT un k ≥ n for every n, which is impossible. X X 355D Lemma Let U be a Riesz space and V any linear space over R. Then a function T : U + → V extends to a linear operator from U to V iff T (u + u0 ) = T u + T u0 ,

T (αu) = αT u

for all u, u0 ∈ U + and every α > 0, and in this case the extension is unique. proof For in this case we can, and must, set T1 u = T u1 − T u2 whenever u1 , u2 ∈ U + and u = u1 − u2 ; it is elementary to check that this defines T1 u uniquely for every u ∈ U , and that T1 is a linear operator extending T . 355E Theorem Let U be a Riesz space and V a Dedekind complete Riesz space. (a) The space L∼ of order-bounded linear operators from U to V is a Dedekind complete Riesz space; its positive cone is the set of positive linear operators from U to V . In particular, every order-bounded linear operator from U to V is expressible as the difference of positive linear operators. (b) For T ∈ L∼ , T + and |T | are defined in the Riesz space L∼ by the formulae T + (w) = sup{T u : 0 ≤ u ≤ w}, Pn Pn |T |(w) = sup{T u : |u| ≤ w} = sup{ i=0 |T ui | : i=0 |ui | = w} for every w ∈ U + . (c) If S, T ∈ L∼ then

355E


(S ∨ T )(w) = sup0≤u≤w Su + T (w − u),

257

(S ∧ T )(w) = inf 0≤u≤w Su + T (w − u)

for every w ∈ U + . (d) Suppose that A ⊆ L∼ is non-empty and upwards-directed. Then A is bounded above in L∼ iff {T u : T ∈ A} is bounded above in V for every u ∈ U + , and in this case (sup A)(u) = supT ∈A T u for every u ≥ 0. (e) Suppose that A ⊆ (L∼ )+ is non-empty and downwards-directed. Then inf A = 0 in L∼ iff inf T ∈A T u = 0 in V for every u ∈ U + . proof (a)(i) Suppose that T ∈ L∼ . For w ∈ U + set RT (w) = sup{T u : 0 ≤ u ≤ w}; this is defined because V is Dedekind complete and {T u : 0 ≤ u ≤ w} is bounded above in V . Then RT (w1 + w2 ) = RT w1 + RT w2 for all w1 , w2 ∈ U + . P P Setting Ai = [0, wi ] for each i, and A = [0, w], then of course A1 + A2 ⊆ A; but also A ⊆ A1 + A2 , because if u ∈ A then u = (u ∧ w1 ) + (u − w1 )+ , and 0 ≤ (u − w1 )+ ≤ (w − w1 )+ = w2 , so u ∈ A1 + A2 . Consequently RT w = sup T [A] = sup T [A1 + A2 ] = sup(T [A1 ] + T [A2 ]) = sup T [A1 ] + sup T [A2 ] = RT w1 + RT w2 by 351Dc. Q Q Next, it is easy to see that RT (αw) = αRT w for w ∈ U + , α > 0, just because u 7→ αu, v 7→ αv are isomorphisms of the partially ordered linear spaces U and V . It follows from 355D that we can extend RT to a linear operator from U to V . Because RT u ≥ T 0 = 0 for every u ∈ U + , RT is a positive linear operator. But also RT u ≥ T u for every u ∈ U + , so RT − T is also positive, and T = RT − (RT − T ) is the difference of two positive linear operators. (ii) This shows that every order-bounded operator is a difference of positive operators. But of course if T1 and T2 are positive, then (T1 − T2 )u ≤ T1 w whenever 0 ≤ u ≤ w in U , so that T1 − T2 is order-bounded, by the criterion in 355Ba. Thus L∼ is precisely the set of differences of positive operators. (iii) Just as in 351F, L∼ is a partially ordered linear space if we say that S ≤ T iff Su ≤ T u for every u ∈ U + . Now it is a Riesz space. P P Take any T ∈ L∼ . Then RT , as defined in (i), is an upper ∼ ∼ bound for {0, T } in L . If S ∈ L is any other upper bound for {0, T }, then for any w ∈ U + we must have Sw ≥ Su ≥ T u whenever u ∈ [0, w], so that Sw ≥ RT w; as w is arbitrary, S ≥ RT ; as S is arbitrary, RT = sup{0, T } in L∼ . Thus sup{0, T } is defined in L∼ for every T ∈ L∼ ; by 352B, L∼ is a Riesz space. Q Q (I defer the proof that it is Dedekind complete to (d-ii) below.) (b) As remarked in (a-iii), RT = T + for each T ∈ L∼ ; but this is just the formula given for T + . Now, if T ∈ L∼ and w ∈ U + , |T |(w) = 2T + w − T w = 2 sup T u − T w u∈[0,w]

= sup T (2u − w) = u∈[0,w]

sup

T u,

u∈[−w,w]

which is the first formula offered for |T |. In particular, if |u| ≤ w then T u,P−T u = T (−u) are both less than n or equal to |T |(w), so that |T u| ≤ |T |(w). So if u0 , . . . , un are such that i=0 |ui | = w, then Pn Pn i=0 |T ui | ≤ i=0 |T |(|ui |) = |T |(w). Pn Pn Thus B = { i=0 |T ui | : i=0 |ui | = w} is bounded above by |T |(w). On the other hand, if v is an upper bound for B and |u| ≤ w, then T u ≤ |T u| + |T (w − |u|)| ≤ v; as u is arbitrary, |T |(w) ≤ v; thus |T |(w) is the least upper bound for B. This completes the proof of part (ii) of the theorem. (c) We know that S ∨ T = T + (S − T )+ (352D), so that (S ∨ T )(w) = T w + (S − T )+ (w) = T w + sup (S − T )(u) 0≤u≤w

= sup T w + (S − T )(u) = sup Su + T (w − u) 0≤u≤w

0≤u≤w

258

Riesz spaces

355E

for every w ∈ U + , by the formula in (b). Also from 352D we have S ∧ T = S + T − T ∨ S, so that

(S ∧ T )(w) = Sw + T w − sup T u + S(w − u) 0≤u≤w

=

inf Sw + T w − T u − S(w − u)

0≤u≤w

(351Db) =

inf Su + T (w − u)

0≤u≤w

for w ∈ U + . (d)(i) Now suppose that A ⊆ L∼ is non-empty and upwards-directed and that {T u : T ∈ A} is bounded above in V for every u ∈ U + . In this case, because V is Dedekind complete, we may set Ru = supT ∈A T u for every u ∈ U + . Now R(u1 + u2 ) = Ru1 + Ru2 for all u1 , u2 ∈ U + . P P Set Bi = {T ui : T ∈ A} for each i, B = {T (u1 + u2 ) : T ∈ A}. Then B ⊆ B1 + B2 , so R(u1 + u2 ) = sup B ≤ sup(B1 + B2 ) = sup B1 + sup B2 = Ru1 + Ru2 . On the other hand, if vi ∈ Bi for both i, there are Ti ∈ A such that vi = Ti ui for each i; because A is upwards-directed, there is a T ∈ A such that T ≥ Ti for both i, and now R(u1 + u2 ) ≥ T (u1 + u2 ) = T u1 + T u2 ≥ T1 u1 + t2 u2 = v1 + v2 . As v1 , v2 are arbitrary, R(u1 + u2 ) ≥ sup(B1 + B2 ) = sup B1 + sup B2 = Ru1 + Ru2 . Q Q It is also easy to see that R(αu) = αRu for every u ∈ U + , α > 0. So, using 355D again, R has an extension to a linear operator from U to V . Now if we fix any T0 ∈ A, we have T0 u ≤ Ru for every u ∈ U + , so R − T0 is a positive linear operator, and R = (R − T0 ) + T0 belongs to L∼ . Again, T u ≤ Ru for every T ∈ A, u ∈ U + , so R is an upper bound for A in L∼ ; and, finally, if S is any upper bound for A in L∼ , then Su is an upper bound for {T u : T ∈ A}, and must be greater than or equal to Ru, for every u ∈ U + ; so that R ≤ S and R = sup A in L∼ . (ii) Now L∼ is Dedekind complete. P P If A ⊆ L∼ is non-empty and bounded above by S say, then A = {T0 ∨ T1 ∨ . . . ∨ Tn : T0 , . . . , Tn ∈ A} is upwards-directed and bounded above by S, so {T u : T ∈ A0 } is bounded above by Su for every u ∈ U + ; by (i) just above, A0 has a supremum in L∼ , which will also be a supremum for A. Q Q 0

(e) Suppose that A ⊆ (L∼ )+ is non-empty and downwards-directed. Then −A = {−T : T ∈ A} is non-empty and upwards-directed, so inf A = 0 ⇐⇒ sup(−A) = 0 ⇐⇒ sup (−T w) = 0 for every w ∈ U + T ∈A

⇐⇒ inf T w = 0 for every w ∈ U + . T ∈A

355F Theorem Let U and V be Riesz spaces and U0 ⊆ U a Riesz subspace which is either order-dense or a solid linear subspace. Suppose that T0 : U0 → V is an order-continuous positive linear operator such that Su = sup{T0 w : w ∈ U0 , 0 ≤ w ≤ u} is defined in V for every u ∈ U + . Then (i) T0 has an extension to an order-continuous positive linear operator T : U → V . (ii) If T0 is a Riesz homomorphism so is T . (iii) If U0 is order-dense then T is unique. (iv) If U0 is order-dense and T0 is an injective Riesz homomorphism, then so is T . proof (a) I check first that S(u + u0 ) = Su + Su0 ,

S(γu) = γSu

355F


259

for all u, u0 ∈ U + and γ > 0. P P For every u ∈ U set A(u) = {v : v ∈ U0 , 0 ≤ v ≤ u}. Note that for any u ∈ U + , v ∈ U0+ v ∧ u = supw∈A(u) v ∧ w = sup A(v ∧ u); this is because sup A(u) = u if U0 is order-dense in U , while v ∧ u ∈ A(u) if U0 is solid. (i) If v ∈ A(u + u0 ) then sup(A(v ∧ u) + A(v − u)+ ) = sup A(v ∧ u) + sup A(v − u)+ = v ∧ u + (v − u)+ = v. Now T0 v = sup T0 [A(v ∧ u) + A(v − u)+ ] = sup(T0 [A(v ∧ u)] + T0 [A(v − u)+ ]) ≤ Su + Su0 because (v − u)+ ≤ u0 . As v is arbitrary, S(u + u0 ) ≤ Su + Su0 . (ii) Next, Su + Su0 = sup T0 [A(u)] + sup T0 [A(u0 )] = sup T0 [A(u) + A(u0 )] ≤ sup T0 [A(u + u0 )] = S(u + u0 ) because A(u) + A(u0 ) ⊆ A(u + u0 ), so S(u + u0 ) = Su + Su0 . (iii) Of course A(γu) = γA(u) so S(γu) = γSu. Q Q By 355D, S has an extension to a linear operator from U to V ; call this operator T . Of course T u = Su ≥ 0 whenever u ≥ 0, so T is positive. If u ∈ U0+ then T u = Su = T0 u, so T extends T0 . (b) If B ⊆ U + is non-empty and upwards-directed and has a supremum u0 ∈ U , then of course T u ≤ T u0 for every u ∈ B, so sup T [B] ≤ T u0 . On the other hand, for any v ∈ A(u0 ) we have also

v = supu∈B v ∧ u = supu∈B supw∈A(u) v ∧ w;

S u∈B

A(u) is upwards-directed, so T0 v = sup{T0 (v ∧ w) : w ∈

S u∈B

A(u)} ≤ supu∈B T u.

As v is arbitrary, T u0 = Su0 ≤ supu∈B T u. As B is arbitrary, T is order-continuous (351Ga). (c) Now suppose that T0 is a Riesz homomorphism. If u ∈ U then, in the language of (a) above,

T u+ ∧ T u− = sup T0 [A(u+ )] ∧ T0 [A(u− )] = sup{w ∧ w0 : w ∈ T0 [A(u+ )], w0 ∈ T0 [A(u− )] (352Ea) = sup{T0 (v ∧ v 0 ) : v ∈ A(u+ ), v 0 ∈ A(u− )} (because T0 is a Riesz homomorphism) = 0. So T is a Riesz homomorphism (352G(iv)). (d) If U0 is order-dense, any order-continuous linear operator extending T0 agrees with S and T on U + , so is equal to T . (e) Finally, if U0 is order-dense and T0 is an injective Riesz homomorphism, then for any non-zero u ∈ U there is a non-zero v ∈ U0 such that |v| ≤ |u|; so that |T u| = T |u| ≥ T0 |v| > 0 because T is a Riesz homomorphism, by (c). As u is arbitrary, T is injective.

260

Riesz spaces

355G

355G Definition Let U be a Riesz space and V a Dedekind complete Riesz space. Then L× (U ; V ) will be the set of those T ∈ L∼ (U ; V ) expressible as the difference of order-continuous positive linear operators, and ∼ L∼ c (U ; V ) will be the set of those T ∈ L (U ; V ) expressible as the difference of sequentially order-continuous positive linear operators. Because a composition of (sequentially) order-continuous functions is (sequentially) order-continuous, we shall have ST ∈ L× (U ; W ) whenever S ∈ L× (V ; W ), T ∈ L× (U ; V ), ∼ ∼ ST ∈ L∼ c (U ; W ) whenever S ∈ Lc (V ; W ), T ∈ Lc (U ; V ),

for all Riesz spaces U and all Dedekind complete Riesz spaces V , W . 355H Theorem Let U be a Riesz space and V a Dedekind complete Riesz space. Then (i) L× = L× (U ; V ) is a band in L∼ = L∼ (U ; V ), therefore a Dedekind complete Riesz space in its own right (ii) a member T of L∼ belongs to L× iff |T | is order-continuous. proof There is a fair bit to check, but each individual step is easy enough. (a) Suppose that S, T are order-continuous positive linear operators from U to V . Then S + T is ordercontinuous. P P If A ⊆ U is non-empty, downwards-directed and has infimum 0, then for any u1 , u2 ∈ A there is a u ∈ A such that u ≤ u1 , u ≤ u2 , and now (S + T )(u) ≤ Su1 + T u2 . Consequently any lower bound for (S + T )[A] must also be a lower bound for S[A] + T [A]. But since inf(S[A] + T [A]) = inf S[A] + inf T [A] = 0 (351Dc), inf(S + T )[A] must also be 0; as A is arbitrary, S + T is order-continuous, by 351Ga. Q Q (b) Consequently S + T ∈ L× for all S, T ∈ L× . Since −T and αT belong to L× for every T ∈ L× and α ≥ 0, we see that L× is a linear subspace of L∼ . (c) If T : U → V is an order-continuous linear operator, S : U → V is linear and 0 ≤ S ≤ T , then S is order-continuous. P P If A ⊆ U is non-empty, downwards-directed and has infimum 0, then any lower bound of S[A] must also be a lower bound of T [A], so inf S[A] = 0; as A is arbitrary, S is order-continuous. Q Q It follows that L× is a solid linear subspace of L∼ . P P If T ∈ L× and |S| ≤ |T | in L∼ , express T as T1 − T2 where T1 , T2 are order-continuous positive linear operators. Then S + , S − ≤ |S| ≤ |T | ≤ T1 + T2 , so S + and S − are order-continuous and S = S + − S − ∈ L× . Q Q Accordingly L× is a Dedekind complete Riesz space in its own right (353J(b-i)). (d) The argument of (c) also shows that if T ∈ L× then |T | is order-continuous; so that for T ∈ L∼ , T ∈ L× ⇐⇒ |T | ∈ L× ⇐⇒ |T | is order-continuous. (e) If C ⊆ (L× )+ is non-empty, upwards-directed and has a supremum T ∈ L∼ , then T is ordercontinuous, so belongs to L× . P P Suppose that A ⊆ U + is non-empty, upwards-directed and has supremum w. Then T w = supS∈C Sw = supS∈C supu∈A Su = supu∈A T u, putting 355Ed and 351G(a-iii) together. So (using 351Ga again) T is order-continuous. Q Q Consequently L× is a band in L∼ (352Ob). This completes the proof. 355I Theorem Let U be a Riesz space and V a Dedekind complete Riesz space. Then L∼ c (U ; V ) is a band in L∼ (U ; V ), and a member T of L∼ (U ; V ) belongs to L∼ c (U ; V ) iff |T | is sequentially order-continuous. proof Copy the arguments of 355H.

355Xe


261

355J Proposition Let U be a Riesz space and V a Dedekind complete Riesz space. Let U0 ⊆ U be an order-dense Riesz subspace; then T 7→ T ¹U0 is an embedding of L× (U ; V ) as a solid linear subspace of L× (U0 ; V ). In particular, any T0 ∈ L× (U0 ; V ) has at most one extension in L× (U ; V ). proof (a) Because the embedding U0 ⊆ U is positive and order-continuous (352Nb), T ¹U0 is positive and order-continuous whenever T is; so T ¹U0 ∈ L× (U0 ; V ) whenever T ∈ L× (U ; V ). Because the map T 7→ T ¹U0 is linear, the image W of L× (U ; V ) is a linear subspace of L× (U0 ; V ). (b) If T ∈ L× (U ; V ) and T ¹U0 ≥ 0, then T ≥ 0. P P?? Suppose, if possible, that there is a u ∈ U + such × that T u 6≥ 0. Because |T | ∈ L (U ; V ) is order-continuous and A = {v : v ∈ U0 , v ≤ u} is an upwardsdirected set with supremum u, inf{|T |(u−v) : v ∈ A} = 0 and there is a v ∈ A such that T u+|T |(u−v) 6≥ 0. But T v = T u + T (v − u) ≤ T u + |T |(u − v) so T v 6≥ 0 and T ¹U0 6≥ 0. X XQ Q This shows that the map T 7→ T ¹U0 is an order-isomorphism between L× (U ; V ) and W , and in particular is injective. (c) Now suppose that S0 ∈ W and that |S| ≤ |S0 | in L× (U0 ; V ). Then S ∈ W . P P Take T0 ∈ L× (U ; V ) such that T0 ¹U0 = S0 . Then S1 = |T0 |¹U0 is a positive member of W such that S0 ≤ S1 , −S0 ≤ S1 , so S + ≤ S1 . Consequently, for any u ∈ U + , sup{S + v : v ∈ U0 , 0 ≤ v ≤ u} ≤ sup{S1 v : v ∈ U0 , 0 ≤ v ≤ u} ≤ |T0 |(u) is defined in V (recall that we are assuming that V is Dedekind complete). But this means that S + has an extension to an order-continuous positive linear operator from U to V (355F), and belongs to W . Similarly, S − ∈ W , so S ∈ W . Q Q This shows that W is a solid linear subspace of L× (U0 ; V ), as claimed. 355K Proposition Let U be a Banach lattice with an order-continuous norm. (a) If V is any Archimedean Riesz space and T : U → V is a positive linear operator, then T is ordercontinuous. (b) If V is a Dedekind complete Riesz space then L× (U ; V ) = L∼ (U ; V ). proof (a) Suppose that A ⊆ U + is non-empty and downwards-directed and has infimum 0. Then for each n ∈ N there is a un ∈ A such that kun k ≤ 4−n . By 354C, u = supn∈N 2n un is defined in U . Now T un ≤ 2−n T u for every n, so any lower bound for T [A] must also be a lower bound for {2−n T u : n ∈ N} and therefore (because V is Archimedean) less than or equal to 0. Thus inf T [A] = 0; as A is arbitrary, T is order-continuous. (b) This is now immediate from 355Ea and the definition of L× . 355X Basic exercises >(a) Let U and V be arbitrary Riesz spaces. (i) Show that the set L(U ; V ) of all linear operators from U to V is a partially ordered linear space if we say that S ≤ T whenever Su ≤ T u for every u ∈ U + . (ii) Show that if U and V are Banach lattices then the set of positive operators is closed in the normed space B(U ; V ) of bounded linear operators from U to V . > (b) If U is a Riesz space and k k, k k0 are two norms on U both rendering it a Banach lattice, show that they are equivalent, that is, give rise to the same topology. (c) Let U be a Riesz space with a Riesz norm, V an Archimedean Riesz space with an order unit, and T : U → V a linear operator which is continuous for the given norm on U and the order-unit norm on V . Show that T is order-bounded. (d) Let U be a Riesz space, V an Archimedean Riesz space, and T : U + → V + a map such that T (u1 + u2 ) = T u1 + T u2 for all u1 , u2 ∈ U + . Show that T has an extension to a linear operator from U to V. >(e) Show that if r, s ≥ 1 are integers then the Riesz space L∼ (Rr ; Rs ) can be identified with the space of real s × r matrices, saying that a matrix is positive iff every coefficient is positive, so that if T = hτij i1≤i≤s,1≤j≤r then |T |, taken in L∼ (Rr ; Rs ), is h|τij |i1≤i≤s,1≤j≤r . Show that a matrix represents a Riesz homomorphism iff each row has at most one non-zero coefficient.

262

Riesz spaces

355Xf

> (f ) Let U be a Riesz space and V a Dedekind complete Riesz space. Show that if T0 , . . . , Tn ∈ L∼ (U ; V ) then Pn Pn (T0 ∨ . . . ∨ Tn )(w) = sup{ i=0 Ti ui : ui ≥ 0 ∀ i ≤ n, i=0 ui = w} for every w ∈ U + . ∼ > (g) Let U be a Riesz space, V a Dedekind complete Riesz Pnspace, and A ⊆ L (U ; V ) a non-empty ∼ set. P Show that A is bounded above in L (U ; V ) iff Cw = { i=0 Ti ui : T0 , . . . , Tn ∈ A, u0 , . . . , un ∈ n U + , i=0 ui = w} is bounded above in V for every w ∈ U + , and in this case (sup A)(w) = sup Cw for every + w∈U .

355Y Further exercises (a) Let U and V be Banach lattices. For T ∈ L∼ = L∼ (U ; V ), set kT k∼ = supw∈U + ,kwk≤1 inf{kvk : |T u| ≤ v whenever |u| ≤ w}. Show that k k∼ is a norm on L∼ under which L∼ is a Banach space, and that the set of positive linear operators is closed in L∼ . (b) Give an example of a continuous linear operator from `2 to itself which is not order-bounded. + (c) Let PnU and V be Riesz spaces+ and Pn T : U → V a linear operator. (i) Show that for any w ∈ U , Cw = { i=0 |T ui | : u0 , . . . , un ∈ U , i=0 ui = w} is upwards-directed, and has the same upper bounds as {TP u : |u| ≤ P w}. (Hint: induce on m and n to see that if u0 , . . . , P un , u00 , . . . , u0m ∈ U + are such m n m + 0 that i=0 ui = j=0 uj , there is a family huij ii≤n,j≤m in U such that j=0 uij = ui for every i ≤ n, Pn + 0 i=0 uij = uj for every j ≤ m.) (ii) Show that if sup Cw is defined for every w ∈ U , then S = T ∨ (−T ) ∼ is defined in the partially ordered linear space L (U ; V ) and Sw = sup Cw for every w ∈ U + .

(d) Let U , V and W be Riesz spaces, of which V and W are Dedekind complete. (i) Show that for any S ∈ L× (V ; W ), the map T 7→ ST : L∼ (U ; V ) → L∼ (U ; W ) belongs to L× (L∼ (U ; V ); L∼ (U ; W )), and is a Riesz homomorphism if S is. (Hint: 355Yc.) (ii) Show that for any T ∈ L∼ (U ; V ), the map S 7→ ST : L∼ (V ; W ) → L∼ (U ; W ) belongs to L× (L∼ (V ; W ); L∼ (U ; W )). (e) Let µ be the usual measure on {0, 1}N and c the Banach lattice of convergent sequences. Find a linear operator T : L2 (µ) → c which is norm-continuous, therefore order-bounded, such that 0 and T have no common upper bound in the partially ordered linear space of all linear operators from L2 (µ) to c . (f ) Let U and V be Banach lattices. Let Lreg be the linear space of operators from U to V expressible as the difference of positive operators. For T ∈ Lreg let kT kreg be inf{kT1 + T2 k : T1 , T2 : U → V are positive, T = T1 − T2 }. Show that k kreg is a norm under which Lreg is complete. (g) Let U and V be Riesz spaces. For this exercise only, say that L× (U ; V ) is to be the set of linear operators T : U → V such that whenever A ⊆ U is non-empty, downwards-directed and has infimum 0 then {v : v ∈ V + , ∃ w ∈ A, |T u| ≤ v whenever |u| ≤ w} has infimum 0 in V . (i) Show that L× (U ; V ) is a linear space. (ii) Show that if U is Archimedean then L× (U ; V ) ⊆ L∼ (U ; V ). (iii) Show that if U is Archimedean and V is Dedekind complete then this definition agrees with that of 355G. (iv) Show that for any Riesz spaces U , V and W , ST ∈ L× (U ; W ) for every S ∈ L× (V ; W ), T ∈ L× (U ; V ). (v) Show that if U and V are Banach lattices, then L× (U ; V ) is closed in L∼ (U ; V ) for the norm k k∼ of 355Ya. (vi) Show that if V is Archimedean and U is a Banach lattice with an order-continuous norm, then L× (U ; V ) = L∼ (U ; V ). (h) Let U be a Riesz space and V a Dedekind complete Riesz space. Show that the band projection P : L∼ (U ; V ) → L× (U ; V ) is given by the formula (P T )(w) = inf{sup T u : A ⊆ U + is non-empty, upwards-directed u∈A

and has supremum w} for every w ∈ U + , T ∈ (L∼ )+ . (Cf. 362Bd.)

355 Notes


263

× (i) Show that if U is a Riesz space with the countable sup property (241Yd), then L∼ c (U ; V ) = L (U ; V ) for every Dedekind complete Riesz space V .

(j) Let U and V be Riesz spaces, of which V is Dedekind complete, and U0 a solid linear subspace of U . Show that the map T 7→ T ¹U0 is an order-continuous Riesz homomorphism from L× (U ; V ) onto a solid linear subspace of L× (U0 ; V ). (k) Let U be a uniformly complete Riesz space (354Yi) and V a Dedekind complete Riesz space. Let UC , VC be their complexifications (354Yk). Show that the complexification of L∼ (U ; V ) can be identified with the complex linear space of linear operators T : UC → VC such that BT (w) = {|T u| : |u| ≤ w} is bounded above in V for every w ∈ U + , and that now |T |(w) = sup BT (w) for every T ∈ L∼ (U ; V )C , w ∈ U + . (Hint: if u, v ∈ U and |u + iv| = w, then u and v can be simultaneously Pn approximated Pn for the order-unit norm k kw on the solid linear subspace generated by w by finite sums j=0 (cos θj )wj , j=0 (sin θj )wj where wj ∈ U + , Pn ∼ j=0 wj = w. Consequently |T (u + iv)| ≤ |T |(w) for every T ∈ LC .) 355 Notes and comments I have had to make some choices in the basic definitions of this chapter (355A, 355G). For Dedekind complete codomains V , there is no doubt what L∼ (U ; V ) should be, since the order-bounded operators (in the sense of 355A) are just the differences of positive operators (335Ea). (These are sometimes called ‘regular’ operators.) When V is not Dedekind complete, we have to choose between the two notions, as not every order-bounded operator need be regular (355Ye). In my previous book (Fremlin 74a) I chose the regular operators; I have still not encountered any really persuasive reason to settle definitively on either class. In 355G the technical complications in dealing with any natural equivalent of the larger space (see 355Yg) are such that I have settled for the narrower class, but explicitly restricting the definition to the case in which V is Dedekind complete. In the applications in this book, the codomains are nearly always Dedekind complete, so we can these questions by. The elementary extension technique in 355D may recall the definition of the Lebesgue integral (122L122M). In the same way, 351G may remind you of the theorem that a linear operator between normed spaces is continuous everywhere if it is continuous anywhere, or of the corresponding results about Boolean homomorphisms and additive functionals on Boolean algebras (313L, 326Ga, 326N). Of course 355Ea is the central fact about the space L∼ (U ; V ) for Dedekind complete V ; because we get a new Riesz space from old ones, the prospect of indefinite recursion immediately presents itself. For Banach lattices, L∼ (U ; V ) is a linear subspace of the space B(U ; V ) of bounded linear operators (355C); the question of when the two are equal will be of great importance to us. I give only the vaguest hints on how to show that they can be different (355Yb, 355Ye), but these should be enough to make it plain that equality is the exception rather than the rule. It is also very useful that we have effective formulae to describe the Riesz space operations on L∼ (U ; V ) (355E, 355Xf-355Xg, 355Yc). You may wish to compare these with the corresponding formulae for additive functionals on Boolean algebras in 326Yj and 362B. If we think of L∼ as somehow corresponding to the space of bounded additive functionals on a Boolean × correspond to the spaces of countably additive and completely additive algebra, the bands L∼ c and L functionals. In fact (as will appear in §362) this correspondence is very close indeed. For the moment, all × I have sought to establish is that L∼ are indeed bands. Of course any case in which L∼ (U ; V ) = c and L ∼ × L∼ (U ; V ) or L (U ; V ) = L (U ; V ) is of interest (355Kb, 355Yi). c c Between Banach lattices, positive linear operators are continuous (355C); it follows at once that the Riesz space structure determines the topology (355Xb), so that it is not to be wondered at that there are further connexions between the norm and the spaces L∼ and L× , as in 355K. 355F will be a basic tool in the theory of representations of Riesz spaces; if we can represent an orderdense Riesz subspace of U as a subspace of a Dedekind complete space V , we have at least some chance of expressing U also as a subspace of V . Of course it has other applications, starting with analysis of the dual spaces.

264

Riesz spaces

§356 intro.

356 Dual spaces As always in functional analysis, large parts of the theory of Riesz spaces are based on the study of linear functionals. Following the scheme of the last section, I define spaces U ∼ , Uc∼ and U × , the ‘order-bounded’, ‘sequentially order-continuous’ and ‘order-continuous’ duals of a Riesz space U (356A). These are Dedekind complete Riesz spaces (356B). If U carries a Riesz norm they are closely connected with the normed space dual U ∗ , which is itself a Banach lattice (356D). For each of them, we have a canonical Riesz homomorphism from U to the corresponding bidual. The map from U to U ×× is particularly important (356I); when this map is an isomorphism we call U ‘perfect’ (356J). The last third of the section deals with L- and M -spaces and the duality between them (356N, 356P), with two important theorems on uniform integrability (356O, 356Q). 356A Definition Let U be a Riesz space. (a) I write U ∼ for the space L∼ (U ; R) of order-bounded real-valued linear functionals on U , the orderbounded dual of U . (b) Uc∼ will be the space L∼ c (U ; R) of differences of sequentially order-continuous positive real-valued linear functionals on U , the sequentially order-continuous dual of U . (c) U × will be the space L× (U ; R) of differences of order-continuous positive real-valued linear functionals on U , the order-continuous dual of U . Remark It is easy to check that the three spaces U ∼ , Uc∼ and U × are in general different (356Xa-356Xc). But the examples there leave open the question: can we find a Riesz space U , for which Uc∼ 6= U × , and which is actually Dedekind complete, rather than just Dedekind σ-complete, as in 356Xc? This leads to unexpectedly deep water; it is yet another form of the Banach-Ulam problem. Really this is a question for Volume 5, but in 363S below I collect the relevant ideas which are within the scope of the present volume. 356B Theorem For any Riesz space U , its order-bounded dual U ∼ is a Dedekind complete Riesz space in which Uc∼ and U × are bands, therefore Dedekind complete Riesz spaces in their own right. For f ∈ U ∼ , f + and |f | ∈ U ∼ are defined by the formulae f + (w) = sup{f (u) : 0 ≤ u ≤ w},

|f |(w) = sup{f (u) : |u| ≤ w}

+

for every w ∈ U . A non-empty upwards-directed set A ⊆ U ∼ is bounded above iff supf ∈A f (u) is finite for every u ∈ U , and in this case (sup A)(u) = supf ∈A f (u) for every u ∈ U + . proof 355E, 355H, 355I. 356C Proposition Let U be any Riesz space and P a band projection on U . Then its adt P 0 : U ∼ → U ∼ , defined by setting P 0 (f ) = f P for every f ∈ U ∼ , is a band projection on U ∼ . proof (a) Because P : U → U is a positive linear operator, P 0 f ∈ U ∼ for every f ∈ U ∼ (355Bd), and P 0 is a positive linear operator from U ∼ to itself. Set Q = I − P , the complementary band projection on U ; then Q0 is another positive linear operator on U ∼ , and P 0 f + Q0 f = f for every f . Now P 0 f ∧ Q0 f = 0 for every f ≥ 0. P P For any w ∈ U + , (P 0 f − Q0 f )+ (w) = sup (P 0 f − Q0 f )(u) = sup f (P u − Qu) 0≤u≤w

0≤u≤w 0

= f (P w) = (P f )(w), so (P 0 f − Q0 f )+ = P 0 f , that is, P 0 f ∧ Q0 f = 0. Q Q By 352Rd, P 0 is a band projection. 356D Proposition Let U be a Riesz space with a Riesz norm. (a) The normed space dual U ∗ of U is a solid linear subspace of U ∼ , and in itself is a Banach lattice with a Fatou norm and has the Levi property.

356E

Dual spaces

265

(b) The norm of U is order-continuous iff U ∗ ⊆ U × . (c) If U is a Banach lattice, then U ∗ = U ∼ , so that U ∼ , U × and Uc∼ are all Banach lattices. (d) If U is a Banach lattice with order-continuous norm then U ∗ = U × = U ∼ . proof (a)(i) If f ∈ U ∗ then sup|u|≤w f (u) ≤ sup|u|≤w kf kkuk = kf kkwk < ∞ for every w ∈ U + , so f ∈ U ∼ (355Ba). Thus U ∗ ⊆ U ∼ . (ii) If f , g ∈ U ∗ and |f | ≤ |g|, then for any w ∈ U |f (w)| ≤ |f |(|w|) ≤ |g|(|w|) = sup|u|≤|w| g(u) ≤ sup|u|≤|w| kgkkuk ≤ kgkkwk. As w is arbitrary, f ∈ U ∗ and kf k ≤ kgk; as f and g are arbitrary, U ∗ is a solid linear subspace of U ∼ and the norm of U ∗ is a Riesz norm. Because U ∗ is a Banach space it is also a Banach lattice. (iii) If A ⊆ (U ∗ )+ is non-empty, upwards-directed and M = supf ∈A kf k is finite, then supf ∈A f (u) ≤ M kuk is finite for every u ∈ U + , so g = sup A is defined in U ∼ (355Ed). Now g(u) = supf ∈A f (u) for every u ∈ U + , as also noted in 355Ed, so |g(u)| ≤ g(|u|) ≤ M k|u|k = M kuk for every u ∈ U , and kgk ≤ M . But as A is arbitrary, this simultaneously proves that the norm of U ∼ is Fatou and has the Levi property. (b)(i) Suppose that the norm is order-continuous. If f ∈ U ∗ and A ⊆ U is a non-empty downwardsdirected set with infimum 0, then inf u∈A |f |(u) ≤ inf u∈A kf kkuk = 0, so |f | ∈ U

×

×

∗

and f ∈ U . Thus U ⊆ U × .

(ii) Now suppose that the norm is not order-continuous. Then there is a non-empty downwards-directed set A ⊆ U , with infimum 0, such that inf u∈A kuk = δ > 0. Set B = {v : v ≥ u for some u ∈ A}. Then B is convex. P P If v1 , v2 ∈ B and α ∈ [0, 1], there are u1 , u2 ∈ A such that vi ≥ ui for both i; now there is a u ∈ A such that u ≤ u1 ∧ u2 , so that u = αu + (1 − α)u ≤ αv1 + (1 − α)v2 , and αv1 + (1 − α)v2 ∈ B. Q Q Also inf v∈B kvk = δ > 0. By the Hahn-Banach theorem (3A5Cb), there is an f ∈ U ∗ such that inf v∈B f (v) > 0. But now inf u∈A |f |(u) ≥ inf u∈A f (u) > 0 and |f | is not order-continuous; so U ∗ 6⊆ U × . (c) By 355C, U ∼ ⊆ U ∗ , so U ∼ = U ∗ . Now U × and Uc∼ , being bands, are closed linear subspaces (354Bd), so are Banach lattices in their own right. (d) Put (b) and (c) together. 356E Biduals If you have studied any functional analysis at all, it will come as no surprise that duals-of-duals are important in the theory of Riesz spaces. I start with a simple lemma. Lemma Let U be a Riesz space and f : U → R a positive linear functional. Then for any u ∈ U + there is a positive linear functional g : U → R such that 0 ≤ g ≤ f , g(u) = f (u) and g(v) = 0 whenever u ∧ v = 0. proof Set g(v) = supα≥0 f (v ∧ αu) for every v ∈ U + . Then it is easy to see that g(βv) = βg(v) for every v ∈ U + , β ∈ [0, ∞[. If v, w ∈ U + then (v ∧ αu) + (w ∧ αu) ≤ (v + w) ∧ 2αu ≤ (v ∧ 2αu) + (w ∧ 2αu) for every α ≥ 0 (352Fa), so g(v + w) = g(v) + g(w). Accordingly g has an extension to a linear functional from U to R (355D). Of course 0 ≤ g(v) ≤ f (v) for v ≥ 0, so 0 ≤ g ≤ f in U ∼ . We have g(u) = f (u), while if u ∧ v = 0 then αu ∧ v = 0 for every α ≥ 0, so g(v) = 0.

266

Riesz spaces

356F

356F Theorem Let U be a Riesz space and V a solid linear subspace of U ∼ . For u ∈ U define u ˆ:V →R by setting u ˆ(f ) = f (u) for every f ∈ V . Then u 7→ u ˆ is a Riesz homomorphism from U to V × . proof (a) By the definition of addition and scalar multiplication in V , u ˆ is linear for every u; also αu c = αˆ u, (u1 + u2 )b= u ˆ1 + u ˆ2 for all u, u1 , u2 ∈ U and α ∈ R. If u ≥ 0 then u ˆ(f ) = f (u) ≥ 0 for every f ∈ V + , so u ˆ ≥ 0; accordingly every u ˆ is the difference of two positive functionals, and u 7→ u ˆ is a linear operator from U to V ∼ . (b) If B ⊆ V is a non-empty downwards-directed set with infimum 0, then inf f ∈B f (u) = 0 for every u ∈ U + , by 355Ee. But this means that u ˆ is order-continuous for every u ∈ U + , so that u ˆ ∈ V × for every u ∈ U. (c) If u ∧ v = 0 in U , then for any f ∈ V + there is a g ∈ [0, f ] such that g(u) = f (u), g(v) = 0 (356D). So (ˆ u ∧ vˆ)(f ) ≤ u ˆ(f − g) + vˆ(g) = f (u) − g(u) + g(v) = 0. As f is arbitrary, u ˆ ∧ vˆ = 0. As u and v are arbitrary, u 7→ u ˆ is a Riesz homomorphism (352G). 356G Lemma Suppose that U is a Riesz space such that U ∼ separates the points of U . Then U is Archimedean. proof ?? Otherwise, there are u, v ∈ U such that v > 0 and nv ≤ u for every n ∈ N. Now there is an f ∈ U ∼ such that f (v) 6= 0; but f (v) ≤ |f |(v) ≤ n1 |f |(u) for every n, so this is impossible. X X 356H Lemma Let U be an Archimedean Riesz space and f > 0 in U × . Then there is a u ∈ U such that (i) u > 0 (ii) f (v) > 0 whenever 0 < v ≤ u (iii) g(u) = 0 whenever g ∧ f = 0 in U × . Moreover, if u0 ∈ U + is such that f (u0 ) > 0, we can arrange that u ≤ u0 . proof (Because f > 0 there certainly is some u0 ∈ U such that f (u0 ) > 0.) (a) Set A = {v : 0 ≤ v ≤ u0 , f (v) = 0}. Then (v1 + v2 ) ∧ u0 ∈ A for all v1 , v2 ∈ A, so A is upwardsdirected. Because f (u0 ) > 0 = sup f [A] and f is order-continuous, u0 cannot be the least upper bound of A, and there is another upper bound u1 of A strictly less than u0 . Set u = u0 − u1 > 0. If 0 ≤ v ≤ u and f (v) = 0, then w ∈ A =⇒ w ≤ u1 =⇒ w + v ≤ u0 =⇒ w + v ∈ A; consequently nv ∈ A and nv ≤ u0 for every n ∈ N, so v = 0. Thus u has properties (i) and (ii). (b) Now suppose that g ∧ f = 0 in U × . Let ² > 0. Then for each n ∈ N there is a vn ∈ [0, u] such that f (vn ) + g(u − vn ) ≤ 2−n ² (355Ec). If v ≤ vn for every n ∈ N then f (v) = 0 so v = 0; thus inf n∈N vn = 0. Set wn = inf i≤n vi for each n ∈ N; then hwn in∈N is non-increasing and has infimum 0 so (because g is order-continuous) inf n∈N g(wn ) = 0. But Pn u − wn = supi≤n u − vi ≤ i=0 u − vi , so g(u − wn ) ≤

Pn i=0

g(u − vi ) ≤ 2²

for every n, and g(u) ≤ 2² + inf n∈N g(wn ) = 2². As ² is arbitrary, g(u) = 0; as g is arbitrary, u has the third required property. 356I Theorem Let U be any Archimedean Riesz space. Then the canonical map from U to U ×× (356F) is an order-continuous Riesz homomorphism from U onto an order-dense Riesz subspace of U ×× . If U is Dedekind complete, its image in U ×× is solid. proof (a) By 356F, u 7→ u ˆ : U → U ×× is a Riesz homomorphism. To see that it is order-continuous, take any non-empty downwards-directed set A ⊆ U with infimum 0. Then C = {ˆ u : u ∈ A} is downwards-directed, and for any f ∈ (U × )+

356K

Dual spaces

267

inf φ∈C φ(f ) = inf u∈A f (u) = 0 because f is order-continuous. As f is arbitrary, inf C = 0 (355Ee); as A is arbitrary, u 7→ u ˆ is ordercontinuous (351Ga). (b) Now suppose that φ > 0 in U ×× . By 356H, there is an f > 0 in U × such that φ(f ) > 0 and φ(g) = 0 whenever g ∧ f = 0. Next, there is a u > 0 in U such that f (u) > 0. Since U ≥ 0, u ˆ ≥ 0; since u ˆ(f ) > 0, u ˆ ∧ φ > 0. Because U ×× (being Dedekind complete) is Archimedean, inf α>0 αˆ u = 0, and there is an α > 0 such that ψ = (ˆ u ∧ φ − αˆ u)+ > 0. Let g ∈ (U × )+ be such that ψ(g) > 0 and θ(g) = 0 whenever θ ∧ ψ = 0 in U ×× . Let v ∈ U + be such that g(v) > 0 and h(v) = 0 whenever h ∧ g = 0 in U × . Because vˆ(g) = g(v) > 0, vˆ ∧ ψ > 0. As ψ ≤ u ˆ, vˆ ∧ u ˆ > 0 and vˆ ∧ αˆ u > 0. Set w = v ∧ αu; then w ˆ = vˆ ∧ αˆ u, by 356F, so w ˆ > 0. ?? Suppose, if possible, that w ˆ 6≤ φ. Then θ = (w ˆ − φ)+ > 0, so there is an h ∈ (U × )+ such that θ(h) > 0 0 0 and θ(h ) > 0 whenever 0 < h ≤ h (356H, for the fourth and last time). Now examine θ(h ∧ g) ≤ (αˆ u−φ∧u ˆ)+ (g) (because w ˆ ≤ αˆ u, φ ∧ u ˆ ≤ φ, h ∧ g ≤ g) =0 because (αˆ u−φ∧u ˆ)+ ∧ ψ = 0. So h ∧ g = 0 and h(v) = 0. But this means that θ(h) ≤ w(h) ˆ ≤ vˆ(h) = 0, which is impossible. X X ˆ of U is quasi-order-dense in U ×× , therefore order-dense Thus 0 < w ˆ ≤ φ. As φ is arbitrary, the image U (353A). ˆ . Express ψ as u (c) Now suppose that U is Dedekind complete and that 0 ≤ φ ≤ ψ ∈ U ˆ where u ∈ U , + and set A = {v : v ∈ U, v ≤ u , vˆ ≤ φ}. If v ∈ U and 0 ≤ vˆ ≤ φ, then w = v + ∧ u+ ∈ A and w ˆ = vˆ; thus ˆ . As φ and ψ are arbitrary, U ˆ is solid in U ×× . φ = sup{ˆ v : v ∈ A} = vˆ0 , where v0 = sup A. So φ ∈ U 356J Definition A Riesz space U is perfect if the canonical map from U to U ×× is an isomorphism. 356K Proposition A Riesz space U is perfect iff (i) it is Dedekind complete (ii) U × separates the points of U (iii) whenever A ⊆ U is non-empty and upwards-directed and {f (u) : u ∈ A} is bounded for every f ∈ U × , then A is bounded above in U . proof (a) Suppose that U is perfect. Because it is isomorphic to U ×× , which is surely Dedekind complete, U also is Dedekind complete. Because the map u 7→ u ˆ : U → U ×× is injective, U × separates the points of U . If A ⊆ U is non-empty and upwards-directed ad {f (u) : u ∈ A} is bounded above for every f ∈ U × , then B = {ˆ u : u ∈ A} is non-empty and upwards-directed and supφ∈B φ(f ) < ∞ for every f ∈ U × , so sup B is defined in U ×∼ (355Ed); but U ×× is a band in U ×∼ , so sup B ∈ U ×× and is of the form w ˆ for some w ∈ U . Because u 7→ u ˆ is a Riesz space isomorphism, w = sup A in U . Thus U satisfies the three conditions. (b) Suppose that U satisfies the three conditions. We know that u 7→ u ˆ is a Riesz homomorphism onto an order-dense Riesz subspace of U ×× (356I). It is injective because U × separates the points of U . If φ ≥ 0 in U ×× , set A = {u : u ∈ U + , u ˆ ≤ φ}. Then A is non-empty and upwards-directed and for any f ∈ U × supu∈A f (u) ≤ supu∈A |f |(u) ≤ supu∈A u ˆ(|f |) ≤ φ(|f |) < ∞, so by condition (iii) A has an upper bound in U . Since U is Dedekind complete, w = sup A is defined in U . Now w ˆ = supu∈A u ˆ = φ. As φ is arbitrary, the image of U includes (U ×× )+ , therefore is the whole of U ×× , and u 7→ u ˆ is a bijective Riesz homomorphism, that is, a Riesz space isomorphism.

268

Riesz spaces

356L

356L Proposition (a) Any band in a perfect Riesz space is a perfect Riesz space in its own right. (b) For any Riesz space U , U ∼ is perfect; consequently Uc∼ and U × are perfect. proof (a) I use the criterion of 356K. Let U be a perfect Riesz space and V a band in U . Then V is Dedekind complete because U is (353Jb). If v ∈ V \ {0} there is an f ∈ U × such that f (v) 6= 0; but the embedding V ⊆ U is order-continuous (352N), so g = f ¹ V belongs to V × , and g(v) 6= 0. Thus V × separates the points of V . If A ⊆ V is non-empty and upwards-directed and supv∈A g(v) is finite for every g ∈ V × , then supv∈A f (v) < ∞ for every f ∈ U × (again because f ¹ V ∈ V × ), so A has an upper bound in U ; because U is Dedekind complete, sup A is defined in U ; because V is a band, sup A ∈ V and is an upper bound for A in V . Thus V satisfies the conditions of 356K and is perfect. (b) U ∼ is Dedekind complete, by 355Ea. If f ∈ U ∼ \ {0}, there is a u ∈ U such that f (u) 6= 0; now u ˆ(f ) 6= 0, where u ˆ ∈ U ∼× (356F). Thus U ∼× separates the points of U ∼ . If A ⊆ U ∼ is non-empty and upwards-directed and supf ∈A φ(f ) is finite for every φ ∈ U ∼× , then, in particular, supf ∈A f (u) = supf ∈A u ˆ(f ) < ∞ for every u ∈ U , so A is bounded above in U ∼ , by 355Ed. Thus U ∼ satisfies the conditions of 356K and is perfect. By (a), it follows at once that U × and Uc∼ are perfect. 356M Proposition If U is a Banach lattice in which the norm is order-continuous and has the Levi property, then U is perfect. proof By 356Db, U ∗ = U × ; since U ∗ surely separates the points of U , so does U × . By 354Ee, U is Dedekind complete. If A ⊆ U is non-empty and upwards-directed and f [A] is bounded for every f ∈ U × , then A is norm-bounded, by the Uniform Boundedness Theorem (3A5Hb). Because the norm is supposed to have the Levi property, A is bounded above in U . Thus U satisfies all the conditions of 356K and is perfect. 356N L- and M -spaces I come now to the duality between L-spaces and M -spaces which I hinted at in §354. Proposition Let U be an Archimedean Riesz space with an order-unit norm. (a) U ∗ = U ∼ is an L-space. (b) If e is the standard order unit of U , then kf k = |f |(e) for every f ∈ U ∗ . (c) A linear functional f : U → R is positive iff it belongs to U ∗ and kf k = f (e). (d) If e 6= 0 there is a positive linear functional f on U such that f (e) = 1. proof (a)-(b) We know already that U ∗ ⊆ U ∼ is a Banach lattice (356Da). If f ∈ U ∼ then sup{|f (u)| : kuk ≤ 1} = sup{|f (u)| : |u| ≤ e} = |f |(e), ∗

so f ∈ U and kf k = |f |(e); thus U ∼ = U ∗ . If f , g ≥ 0 in U ∗ , then kf + gk = (f + g)(e) = f (e) + g(e) = kf k + kgk; thus U ∗ is an L-space. (c) As already remarked, if f is positive then f ∈ U ∗ and kf k = f (e). On the other hand, if f ∈ U ∗ and kf k = f (e), take any u ≥ 0. Set v = (1 + kuk)−1 u. Then 0 ≤ v ≤ e and ke − vk ≤ 1 and f (e − v) ≤ |f (e − v)| ≤ kf k = f (e). But this means that f (v) ≥ 0 so f (u) ≥ 0. As u is arbitrary, f ≥ 0. (d) By the Hahn-Banach theorem (3A5Ac), there is an f ∈ U ∗ such that f (e) = kf k = 1; by (c), f is positive. 356O Theorem Let U be an Archimedean Riesz space with order-unit norm. Then a set A ⊆ U ∗ = U ∼ is uniformly integrable iff it is norm-bounded and limn→∞ supf ∈A |f (un )| = 0 for every order-bounded dist sequence hun in∈N in U + .

356O

Dual spaces

269

proof (a) Suppose that A is uniformly integrable. Then it is surely norm-bounded (354Ra). If hun in∈N is ∗ a dist sequence in U + bounded above Pnby w, then for any ² > 0 we can find an h ≥ 0 in U such that + k(|f | − h) k ≤ ² for every f ∈ A. Now i=0 h(ui ) ≤ h(w) for every n, and limn→0 h(un ) = 0; since at the same time |f (un )| ≤ |f |(un ) ≤ h(un ) + (|f | − h)+ (un ) ≤ h(un ) + ²kun k ≤ h(un ) + ²kwk for every f ∈ A, n ∈ N, lim supn→∞ supf ∈A |f |(un ) ≤ ²kwk. As ² is arbitrary, limn→∞ supf ∈A |f |(un ) = 0, and the conditions are satisfied. (b)(i) Now suppose that A is norm-bounded but not uniformly integrable. Write B for the solid hull of A, M for supf ∈A kf k = supf ∈B kf k; then there is a dist sequence hgn in∈N in B ∩ (U ∗ )+ which is not norm-convergent to 0 (354R(b-iv)), that is, δ=

1 2

lim supn→∞ gn (e) =

1 2

lim supn→∞ kgn k > 0,

where e is the standard order unit of U . (ii) Set C = {v : 0 ≤ v ≤ e, supg∈B g(v) ≥ δ}, D = {w : 0 ≤ w ≤ e, lim supn→∞ gn (w) > δ}. Then for any u ∈ D we can find v ∈ C, w ∈ D such that v ∧ w = 0. P P Set δ 0 = lim supn→∞ gn (u), 0 η = (δ − δ)/(3 + M ) > 0; take k ∈ N so large that kη ≥ M . Because gn (u) ≥ δ 0 − η for infinitely many n, we can find P a set K ⊆ N, of size k, such that gi (u) ≥ δ 0 − η for every i ∈ K. Now we know that, for each i ∈ K, gi ∧ k j∈K,j6=i gj = 0, so there is a vi ≤ u such that P gi (u − vi ) + k j∈K,j6=i gj (vi ) ≤ η (355Ec). Now gi (vi ) ≥ gi (u) − η ≥ δ 0 − 2η, Set vi0 = (vi −

P j∈K,j6=i

gi (vj ) ≤

η k

for i, j ∈ K, i 6= j.

vj )+ for each i ∈ K; then P gi (vi0 ) ≥ gi (vi ) − j∈K,j6=i gi (vj ) ≥ δ 0 − 3η

for every i ∈ K, while vj0 ∧ vi0 = 0 for distinct i, j ∈ K. For each n ∈ N, P 1 0 i∈K gn (u ∧ vi ) ≤ gn (u) ≤ kgn k ≤ ηk, η

so there is some i(n) ∈ K such that 1 η

0 gn (u ∧ vi(n) ) ≤ η,

1 η

0 gn (u − vi(n) )+ ≥ gn (u) − η.

Since {n : gn (u) ≥ δ + 2η} is infinite, there is some m ∈ K such that J = {n : gn (u) ≥ δ + 2η, i(n) = m} is infinite. Try 0 v = (vm − ηu)+ ,

1 η

0 + ) . w = (u − vm

Then v, w ∈ [0, u] and v ∧ w = 0. Next, 0 gm (v) ≥ gm (vm ) − ηM ≥ δ 0 − 3η − ηM = δ,

so v ∈ C, while for any n ∈ J 1 η

0 )+ ≥ gn (u) − η ≥ δ + η; gn (w) = gn (u − vi(n)

since J is infinite, lim supn→∞ gn (w) ≥ δ + η > δ

270

Riesz spaces

356O

and w ∈ D. Q Q (iii) Since e ∈ D, we can choose inductively sequences hwn in∈N in D, hvn in∈N in C such that w0 = e, vn ∧ wn+1 = 0, vn ∨ wn+1 ≤ wn for every n ∈ N. But in this case hvn in∈N is a dist order-bounded sequence in [0, u], while for each n ∈ N, we can find fn ∈ A such that |fn |(vn ) > 23 δ. Now there is a un ∈ [0, vn ] such that |fn (un )| ≥ 31 δ. P P Set γ = sup0≤v≤vn |fn (v)|. Then fn+ (vn ), fn− (vn ) are both less than or equal to γ, so |fn |(vn ) ≤ 2γ and γ > 31 δ; so there is a un ∈ [0, vn ] such that |fn (un )| ≥ 31 δ. Q Q Accordingly we have a dist sequence hun in∈N in [0, e] such that supf ∈A |f (un )| ≥ 13 δ for every n ∈ N. (iv) All this is on the assumption that A is norm-bounded and not uniformly integrable. So, turning it round, we see that if A is norm-bounded and limn→∞ supf ∈A |f (un )| = 0 for every order-bounded dist sequence hun in∈N , A must be uniformly integrable. This completes the proof. 356P Proposition Let U be an L-space. (a) U is perfect. R ∗ ∼ × defined by setting R (b) U + = U − = U is an M -space; its standard order unit is the functional u = ku k − ku k for every u ∈ U . R R (c) If A ⊆ U is Rnon-empty and upwards-directed and supu∈A u is finite, then sup A is defined in U and sup A = supu∈A u. proof (a) By 354N we know that the norm on U is order-continuous and has the Levi property, so 356M tells us that U is perfect. (b) 356Dd tells us that U ∗ = U ∼ = U × . The L-space property tells us that the functional u 7→ kuk : U R+ → R is additive; of course it is also homogeneous, so by 355D it has an R R extension to a linear functional : U → R satisfying the given formula. Because u = kuk ≥ 0 for u ≥ 0, ∈ (U ∼ )+ . For f ∈ U ∼ , Z Z |f | ≤ ⇐⇒ |f |(u) ≤ u for every u ∈ U + ⇐⇒ |f (v)| ≤ kuk whenever |v| ≤ u ∈ U ⇐⇒ |f (v)| ≤ kvk for every v ∈ U ⇐⇒ kf k ≤ 1, so the norm on U ∗ = U ∼ is the order-unit norm defined from

R

, and U ∼ is an M -space, as claimed.

(c) Fix u0 ∈ A, and set B = {u+ : u ∈ A, u ≥ u0 }. Then B ⊆ U + is upwards-directed, and Z Z Z sup kvk = sup u+ = sup u + u− v∈B u∈A,u≥u0 u∈A,u≥u0 Z Z ≤ sup u + u− 0 < ∞. u∈A,u≥u0

Because k k has the Levi property, B is bounded above. But (because A is upwards-directed) every member of A is dominated by some member of B, so A also R is bounded R above. Because U is Dedekind complete, R sup A is defined in U . Finally, sup A = supu∈A u because , being a positive member of U × , is ordercontinuous. 356Q Theorem Let U be any L-space. Then a subset of U is uniformly integrable iff it is relatively weakly compact. proof (a) Let A ⊆ U be a uniformly integrable set. (i) Suppose that F is an ultrafilter on X containing A. Then A 6= ∅. Because A is norm-bounded, supu∈A |f (u)| < ∞ and φ(f ) = limu→F f (u) is defined in R for every f ∈ U ∗ (2A3Se). If f , g ∈ U ∗ then

356X

Dual spaces

271

φ(f + g) = limu→F f (u) + g(u) = limu→F f (u) + limu→F g(u) = φ(f ) + φ(g) (2A3Sf). Similarly, φ(αf ) = limu→F αf (u) = αφ(f ) ∗

∗

for every f ∈ U , α ∈ R. Thus φ : U → R is linear. Also |φ(f )| ≤ supu∈A |f (u)| ≤ kf k supu∈A kuk, so φ ∈ U

∗∗

=U

∗∼

.

(ii) Now the point of this argument is that φ ∈ U ∗× . P P Suppose that B ⊆ U ∗ is non-empty and downwards-directed and has infimum 0. Fix f0 ∈ B. Let ² > 0. Then there is a w ∈ U + such that k(|u| − w)+ k ≤ ² for every u ∈ A, which means that |f (u)| ≤ |f |(|u|) ≤ |f |(w) + |f |(|u| − w)+ ≤ |f |(w) + ²kf k for every f ∈ U ∗ and every u ∈ A. Accordingly |φ(f )| ≤ |f |(w)+²kf k for every f ∈ U ∗ . Now inf f ∈B f (w) = 0 (using 355Ee, as usual), so there is an f1 ∈ B such that f1 ≤ f0 and f1 (w) ≤ ². In this case |φ|(f1 ) = sup|f |≤f1 |φ(f )| ≤ sup|f |≤f1 |f |(w) + ²kf k ≤ f1 (w) + ²kf1 k ≤ ²(1 + kf0 k). As ² is arbitrary, inf f ∈B |φ|(f ) = 0; as B is arbitrary, |φ| is order-continuous and φ ∈ U ∗× . Q Q (iii) At this point, we recall that U ∗ = U × and that the canonical map from U to U ×× is surjective (356P). So there is a u0 ∈ U such that u ˆ0 = φ. But now we see that f (u0 ) = φ(f ) = limu→F f (u) for every f ∈ U ∗ ; which is just what is meant by saying that F → u0 for the weak topology on U (2A3Sd). Accordingly every ultrafilter on U containing A has a limit in U . But because the weak topology on U is regular (3A3Be), it follows that the closure of A for the weak topology is compact (3A3De), so that A is relatively weakly compact. (b) For the converse I use the criterion of 354R(b-iv). Suppose that A ⊆ U is relatively weakly compact. Then A is norm-bounded, by the Uniform Boundedness Theorem. Now let hun in∈N be any dist sequence in the solid hull of A. For each n, let Un be the band of U generated by un . Let Pn be the band projection from U onto Un (353H). Let vn ∈ A be such that |un | ≤ |vn |; then |un | = Pn |un | ≤ Pn |vn | = |Pn vn |, so kun k ≤ kPn vn k for each n. Let gn ∈ U ∗ be such that kgn k = 1 and gn (Pn vn ) = kPn vn k. Define T : U → RN by setting T u = hgn (Pn u)in∈N for each u ∈ U . Then T is a continuous linear operator from U to `1 . P P For m 6= n, Um ∩ Un = {0}, because |um | ∧ |un | = 0. So, for any u ∈ U , hPn uin∈N is a dist sequence in U , and Pn Pn i=0 |Pi u|k = k supi≤n |Pi u|k ≤ kuk i=0 kPi uk = k for every n; accordingly kT uk1 =

P∞ i=0

|gi Pi u| ≤

P∞ i=0

kPi uk ≤ kuk.

Since T is certainly a linear operator (because every coordinate functional gi Pi is linear), we have the result. Q Q Consequently T [A] is relatively weakly compact in `1 , because T is continuous for the weak topologies (2A5If). But `1 can be identified with L1 (µ), where µ is counting measure on N. So T [A] is uniformly integrable in `1 , by 247C, and in particular limn→∞ supw∈T [A] |w(n)| = 0. But this means that limn→∞ kun k ≤ limn→∞ |gn (Pn vn )| = limn→∞ |(T vn )(n)| = 0. As hun in∈N is arbitrary, A satisfies the conditions of 354R(b-iv) and is uniformly integrable. 356X Basic exercises (a) Show that if U = `∞ then U × = Uc∼ can be identified with `1 , and is P∞ ∼ ∼ properly included in U . (Hint: show that if f ∈ Uc then f (u) = n=0 u(n)f (en ), where en (n) = 1, en (i) = 0 for i 6= n.)

272

Riesz spaces

356Xb

(b) Show that if U = C([0, 1]) then U × = Uc∼ = {0}. (Hint: show that if f ∈ (Uc∼ )+ and hqn in∈N enumerates Q ∩ [0, 1], then for each n ∈ N there is a un ∈ U + such that un (qn ) = 1 and f (un ) ≤ 2−n .) (c) Let X be an uncountable set and µ the countable-cocountable measure on X, Σ its domain (211R). Let U be the space of bounded Σ-measurable real-valued functions on X. Show that U is a Dedekind σ-complete Banach lattice Rif given the supremum norm k k∞ . Show that U × can be identified with `1 (X) (cf. 356Xa), and that u 7→ u dµ belongs to Uc∼ \ U × . (d) Let U be a Dedekind σ-complete Riesz space and f ∈ Uc∼ . Let hun in∈N be an order-bounded sequence in U which is order-convergent to u ∈ U in the sense that u = inf n∈N supm≥n um = supn∈N inf m≥n um . Show that limn→∞ f (un ) exists and is equal to f (u). (e) Let U be any Riesz space. Show that the band projection P : U ∼ → U × is defined by the formula (P f )(u) = inf{sup f (v) : A ⊆ U is non-empty, upwards-directed v∈A

and has supremumu} for every f ∈ (U ∼ )+ , u ∈ U + . (Hint: show that the formula for P f always defines an order-continuous linear functional. Compare 355Yh, 356Yb and 362Bd.) (f ) Let U be any Riesz space. Show that the band projection P : U ∼ → Uc∼ is defined by the formula (P f )(u) = inf{supn∈N f (vn ) : hvn in∈N is a non-decreasing sequence with supremum u} for every f ∈ (U ∼ )+ , u ∈ U + . (g) Let U be a Riesz space with a Riesz norm. Show that U ∗ is perfect. (h) Let U be a Riesz space with a Riesz norm. Show that the canonical map from U to U ∗∗ is a Riesz homomorphism. (i) Let V be a perfect Riesz space and U any Riesz space. Show that L∼ (U ; V ) is perfect. (Hint: show that if u ∈ U , g ∈ V × then T 7→ g(T u) belongs to L∼ (U ; V )× .) (j) Let U be an M -space. Show that it is perfect iff it is Dedekind complete and U × separates the points of U . (k) Let U be a Banach lattice which, as a Riesz space, is perfect. Show that its norm has the Levi property. (l) Write out a proof from first principles that if hun in∈N is a sequence in `1 such that |un (n)| ≥ δ > 0 for every n ∈ N, then {un : n ∈ N} is not relatively weakly compact. (m) Let U be an L-space and A ⊆ U a non-empty set. Show that the following are equiveridical: (i) A is uniformly integrable (ii) inf f ∈B supu∈A |f (u)| for every non-empty downwards-directed set B ⊆ U × with infimum 0 (iii) inf n∈N supu∈A |fn (u)| = 0 for every non-increasing sequence hfn in∈N in U × with infimum 0 (iv) A is norm-bounded and limn→∞ supu∈A |fn (u)| = 0 for every dist order-bounded sequence hfn in∈N in U × . 356Y Further exercises (a) Let U be a Riesz space with the countable sup property. Show that U × = Uc∼ . (b) Let U be a Riesz space, and A a family of non-empty downwards-directed subsets of U + all with ∼ infimum 0. (i) Show that UA = {f : f ∈ U ∼ , inf u∈A |f |(u) = 0 for every A ∈ A} is a band in U ∼ . (ii) Set ∗ ∼ ∼ ∼ + A = {A0 + . . . + An : A0 , . . . , An ∈ A}. Sbow that UA = UA ) , and let g, h be ∗ . (iii) Take any f ∈ (U ∼ ∼ ⊥ the components of f in UA , (UA ) respectively. Show that g(u) = inf A∈A∗ supv∈A f (u − v)+ , for every u ∈ U + . (Cf. 362Xi.)

h(u) = supA∈A∗ inf v∈A f (u ∧ v)

356 Notes

Dual spaces

273

(c) Let U be a Riesz space. For any band V ⊆ U write V ◦ for {f : f ∈ U × , f (v) = 0 for every v ∈ V }. Show that V 7→ (V ⊥ )◦ is a surjective order-continuous Boolean homomorphism from the algebra of complemented bands of U onto the band algebra of U × , and that it is injective iff U × separates the points of U . (d) Let U be a Riesz space such that U ∼ separates the points of U . For any band V ⊆ U ∼ write V ◦ for {x : x ∈ U, f (x) = 0 for every f ∈ V }. Show that V 7→ (V ⊥ )◦ is a surjective Boolean homomorphism from the algebra of bands of U ∼ onto the band algebra of U , and that it is injective iff U ∼ = U × . (e) Let U be a Dedekind complete Riesz space such that U × separates the points of U and U is the solid linear subspace of itself generated by a countable set. Show that U is perfect. (f ) Let U be an L-space and hun in∈N a sequence in U such that hf (un )in∈N is Cauchy for every f ∈ U ∗ . Show that hun in∈N is convergent for the weak topology of U . (Hint: use 356Xm(iv) to show that {un : n ∈ N} is relatively weakly compact.) (g) Let U be a perfect Banach lattice with order-continuous norm and hun in∈N a sequence in U such that hf (un )in∈N is Cauchy for every f ∈ U ∗ . Show that hun in∈N is convergent for the weak topology of U . (Hint: set φ(f ) = limn→∞ fn (u). For any g ∈ (U ∗ )+ let Vg be the solid linear subspace of U ∗ generated by g, Wg = {u : g(|u|) = 0}⊥ , kukg = g(|u|) for u ∈ Wg . Show that the completion of Wg under k kg is an L-space with dual isomorphic to Vg , and hence (using 356Yf) that φ¹ Vg belongs to Vg× ; as g is arbitrary, φ ∈ V × and may be identified with an element of U .) (h) Let U be a uniformly complete Archimedean Riesz space with complexification V (354Yk). (i) Show that the complexification of U ∼ can be identified with the space of linear functionals f : V → C such that sup|v|≤u |f (v)| is finite for every u ∈ U + . (ii) Show that if U is a Banach lattice, then the complexification of U ∼ = U ∗ can be identified (as normed space) with V ∗ . (See 355Yk.) 356 Notes and comments The section starts easily enough, with special cases of results in §355 (356B). When U has a Riesz norm, the identification of U ∗ as a subspace of U ∼ , and the characterization of ordercontinuous norms (356D) are pleasingly comprehensive and straightforward. Coming to biduals, we need to think a little (356F), but there is still no real difficulty at first. In 356H-356I, however, something more substantial is happening. I have written these arguments out in what seems to be the shortest route to the main theorem, at the cost perhaps of neglecting any intuitive foundation. What I think we are really doing is matching bands in U , U × and U ×× , as in 356Yc. From now on, almost the first thing we shall ask of any new Riesz space will be whether it is perfect, and if not, which of the three conditions of 356K it fails to satisfy. For reasons which will I hope appear in the next chapter, perfect Riesz spaces are particularly important in measure theory; in particular, all Lp spaces for p ∈ [1, ∞[ are perfect (366D), as are the L∞ spaces of localizable measure spaces (365K). Further examples will be discussed in §369 and §374. Of course we have to that there are also important Riesz spaces which are not perfect, of which C([0, 1]) and c 0 are two of the simplest examples. The duality between L- and M -spaces (356N, 356P) is natural and satisfying. We are now in a position to make a determined attempt to tidy up the notion of ‘uniform integrability’. I give two major theorems. The first is yet another ‘dist-sequence’ characterization of uniformly integrable sets, to go with 246G and 354R. The essential difference here is that we are looking at dist sequences in a predual; in a sense, this means that the result is a sharper one, because the M -space U need not be Dedekind complete (for instance, it could be C([0, 1]) – this indeed is the archetype for applications of the theorem) and therefore need not have as many dist sequences as its dual. (For instance, in the dual of C([0, 1]) we have all the point masses δt , where δt (u) = u(t); these form a dist family in C([0, 1])∼ not corresponding to any dist family in C([0, 1]).) The essence of the proof is a device to extract a dist sequence in U to match approximately a subsequence of a given dist sequence in U ∼ . In the example just suggested, this would correspond, given a sequence htn in∈N of distinct points in [0, 1], to finding a subsequence htn(i) ii∈N which is discrete, so that we can find dist ui ∈ C([0, 1]) with ui (tn(i) ) = 1 for each i. The second theorem, 356Q, is a new version of a result already given in §247: in any L-space, uniform integrability is the same as relative weak compactness. I hope you are not exasperated by having been

274

Riesz spaces

356 Notes

asked, in Volume 2, to master a complex argument (one of the more difficult sections of that volume) which was going to be superseded. Actually it is worse than that. A theorem of Kakutani (369E) tells us that every L-space is isomorphic to an L1 space. So 356Q is itself a consequence of 247C. I do at least owe you an explanation for writing out two proofs. The first point is that the result is sufficiently important for it to be well worth while spending time in its neighbourhood, and the contrasts and similarities between the two arguments are instructive. The second is that the proof I have just given was not really accessible at the level of Volume 2. It does not rely on every single page of this chapter, but the key idea (that U is isomorphic to U ×× , so it will be enough if we can show that A is relatively compact in U ×× ) depends essentially on 356I, which lies pretty deep in the abstract theory of Riesz spaces. The third is an aesthetic one: a theorem about L-spaces ought to be proved in the category of normed Riesz spaces, without calling on a large body of theory outside. Of course this is a book on measure theory, so I did the measure theory first, but if you look at everything that went into it, the proof in §247 is I believe longer, in the formal sense, than the one here, even setting aside the labour of proving Kakutani’s theorem. Let us examine the ideas in the two proofs. First, concerning the proof that uniformly integrable sets are relatively compact, the method here is very smooth and natural; the definition I chose of ‘uniform integrability’ is exactly adapted to showing that uniformly integrable sets are relatively compact in the order-continuous bidual; all the effort goes into the proof that L-spaces are perfect. The previous argument depended on identifying the dual of L1 as L∞ – and was disagreeably complicated by the fact that the identification is not always valid, so that I needed to reduce the problem to the σ-finite case (part (b-ii) of the proof of 247C). After that, the Radon-Nikod´ ym theorem did the trick. Actually Kakutani’s theorem shows that the side-step to σ-finite spaces is irrelevant. It directly represents an abstract L-space as L1 (µ) for a localizable measure µ, in which case (L1 )∗ ∼ = L∞ exactly. In the other direction, both arguments depend on a dist-sequence criterion for uniform integrability (246G(iii) or 354R(b-iv)). These criteria belong to the ‘easy’ side of the topic; straightforward Riesz space arguments do the job, whether written out in that language or not. (Of course the new one in this section, 356O, lies a little deeper.) I go a bit faster this time because I feel that you ought by now to be happy with the Hahn-Banach theorem and the Uniform Boundedness Theorem, which I was avoiding in Volume 2. And then of course I quote the result for `1 . This looks like cheating. But `1 really is easier, as you will find if you just write out part (a) of the proof of 247C for this case. It is not exactly that you can dispense with any particular element of the R argument; rather it is that the formulae become much more direct when you can write u(i) in place of Fi u, and ‘cluster points for the weak topology’ become pointwise limits of subsequences, so that the key step (the ‘sliding hump’, in which uk(j) (n(k(j))) is the only significant coordinate of uk(j) ), is easier to find. We now have a wide enough variety of conditions equivalent to uniform integrability for it to be easy to find others; I give a couple in 356Xm, corresponding in a way to those in 246G. You may have noticed, in the proof of 247C, that in fact the full strength of the hypothesis ‘relatively weakly compact’ is never used; all that is demanded is that a couple of sequences should have cluster points for the weak topology. So we see that a set A is uniformly integrable iff every sequence in A has a weak cluster point. But this extra refinement is nothing to do with L-spaces; it is generally true, in any normed space U , that a set A ⊆ U is relatively weakly compact iff every sequence in A has a cluster point in U for the weak topology (‘Eberlein’s ¨ the 69, 24.2.1, or Dunford & Schwartz 57, V.6.1). theorem’; see 462D in Volume 4, Ko There is a very rich theory concerning weak compactness in perfect Riesz spaces, based on the ideas here; some of it is explored in Fremlin 74a. As a sample, I give one of the basic properties of perfect Banach lattices with order-continuous norms: they are ‘weakly sequentially complete’ (356Yg).

Chap. 36 intro.

Introduction

275

Chapter 36 Function Spaces Chapter 24 of Volume 2 was devoted to the elementary theory of the ‘function spaces’ L0 , L1 , L2 and L associated with a given measure space. In this chapter I return to these spaces to show how they can be related to the more abstract themes of the present volume. In particular, I develop constructions to demonstrate, as clearly as I can, the way in which all the function spaces associated with a measure space in fact depend only on its measure algebra; and how many of their features can (in my view) best be understood in of constructions involving measure algebras. The chapter is very long, not because there are many essentially new ideas, but because the intuitions I seek to develop depend, for their logical foundations, on technically complex arguments. This is perhaps best exemplified by §364. If two measure spaces (X, Σ, µ) and (Y, T, ν) have isomorphic measure algebras (A, µ ¯), (B, ν¯) then the spaces L0 (µ), L0 (ν) are isomorphic as topological f -algebras; and more: for any isomorphism between (A, µ ¯) and (B, ν¯) there is a unique corresponding isomorphism between the L0 spaces. The intuition involved is in a way very simple. If f , g are measurable real-valued functions on X and Y respectively, then f • ∈ L0 (µ) will correspond to g • ∈ L0 (ν) if and only if [[f • > α]] = {x : f (x) > α}• ∈ A corresponds to [[g • > α]] = {y : g(y) > α}• ∈ B for every α. But the check that this formula is consistent, and defines an isomorphism of the required kind, involves a good deal of detailed work. It turns out, in fact, that the measures µ and ν do not enter this part of the argument at all, except through their ideals of negligible sets (used in the construction of A and B). This is already evident, if you look for it, in the theory of L0 (µ); in §241, as written out, you will find that the measure of an individual set is not once mentioned, except in the exercises. Consequently there is an invitation to develop the theory with algebras A which are not necessarily measure algebras. Here is another reason for the length of the chapter; substantial parts of the work are being done in greater generality than the corresponding sections of Chapter 24, necessitating a degree of repetition. Of course this is not ‘measure theory’ in the strict sense; but for thirty years now measure theory has been coloured by the existence of these generalizations, and I think it is useful to understand which parts of the theory apply only to measure algebras, and which can be extended to other σ-complete Boolean algebras, like the algebraic theory of L0 , or even to all Boolean algebras, like the theory of L∞ . Here, then, are two of the objectives of this chapter: first, to express the ideas of Chapter 24 in ways making explicit their independence of particular measure spaces, by setting up constructions based exclusively on the measure algebras involved; second, to set out some natural generalizations to other algebras. But to justify the effort needed I ought to point to some mathematically significant idea which demands these constructions for its expression, and here I mention the categorical nature of the constructions. Between Boolean algebras we have a variety of natural and important classes of ‘morphism’; for instance, the Boolean homomorphisms and the order-continuous Boolean homomorphisms; while between measure algebras we have in addition the measure-preserving Boolean homomorphisms. Now it turns out that if we construct the Lp spaces in the natural ways then morphisms between the underlying algebras give rise to morphisms between their Lp spaces. For instance, any Boolean homomorphism from A to B produces a multiplicative norm-contractive Riesz homomorphism from L∞ (A) to L∞ (B); if A and B are Dedekind σ-complete, then any sequentially order-continuous Boolean homomorphism from A to B produces a sequentially order-continuous multiplicative Riesz homomorphism from L0 (A) to L0 (B); and if (A, µ ¯) and (B, ν¯) are measure algebras, then any measure-preserving Boolean homomorphism from A to B produces norm-preserving Riesz homomorphisms from Lp (A, µ ¯) to Lp (B, ν¯) for every p ∈ [1, ∞]. All of these are ‘functors’, that is, a composition of homomorphisms between algebras gives rise to a composition of the corresponding operators between their function spaces, and are ‘covariant’, that is, a homomorphism from A to B leads to an operator from Lp (A) to Lp (B). But the same constructions lead us to a functor which is ‘contravariant’: starting from an order-continuous Boolean homomorphism from a semi-finite measure algebra (A, µ ¯) to a measure algebra (B, ν¯), we have an operator from L1 (B, ν¯) to L1 (A, µ ¯). This last is in fact a kind of conditional expectation operator. In my view it is not possible to make sense of the theory of measure-preserving transformations without at least an intuitive grasp of these ideas. Another theme is the characterization of each construction in of universal mapping theorems: for instance, each Lp space, for 1 ≤ p ≤ ∞, can be characterized as Banach lattice in of factorizations of functions of an appropriate class from the underlying algebra to Banach lattices. ∞

276

Function spaces

Chap. 36 intro.

Now let me try to sketch a route-map for the journey ahead. I begin with two sections on the space S(A); this construction applies to any Boolean algebra (indeed, any Boolean ring), and corresponds to the space of ‘simple functions’ on a measure space. Just because it is especially close to the algebra (or ring) A, there is a particularly large number of universal mapping theorems corresponding to different aspects of its structure (§361). In §362 I seek to relate ideas on additive functionals on Boolean algebras from Chapter 23 and §§326-327 to the theory of Riesz space duals in §356. I then turn to the systematic discussion of the function spaces of Chapter 24: L∞ (§363), L0 (§364), L1 (§365) and other Lp (§366), followed by an of convergence in measure (§367). While all these sections are dominated by the objectives sketched in the paragraphs above, I do include a few major theorems not covered by the ideas of Volume 2, such as the Kelley-Nachbin characterization of the Banach spaces L∞ (A) for Dedekind complete A (363R). In the last two sections of the chapter I turn to the use of L0 spaces in the representation of Archimedean Riesz spaces (§368) and of Banach lattices separated by their order-continuous duals (§369).

361 S This is the fundamental Riesz space associated with a Boolean ring A. When A is a ring of sets, S(A) can be regarded as the linear space of ‘simple functions’ generated by the characteristic functions of of A (361L). Its most important property is the universal mapping theorem 361F, which establishes a one-toone correspondence between (finitely) additive functions on A (361B-361C) and linear operators on S(A). Simple universal mapping theorems of this type can be interesting, but do not by themselves lead to new insights; what makes this one important is the fact that S(A) has a canonical Riesz space structure, norm and multiplication (361E). From this we can deduce universal mapping theorems for many other classes of function (361G, 361H, 361I, 361Xb). (Particularly important are countably additive and completely additive real-valued functionals, which will be dealt with in the next section.) While the exact construction of S(A) (and the associated map from A to S(A)) can be varied (361D, 361L, 361M, 361Ya), its structure is uniquely defined, so homomorphisms between Boolean rings correspond to maps between their S( )-spaces (361J), and (when A is an algebra) A can be recovered from the Riesz space S(A) as the algebra of its projection bands (361K). 361A Boolean rings In this section I speak of Boolean rings rather than algebras; there are ideas in §365 below which are more naturally expressed in of the ring of elements of finite measure in a measure algebra than in of the whole algebra. I should perhaps therefore recall some of the ideas of §311, which is the last time when Boolean rings without identity were mentioned, and set out some simple facts. (a) Any Boolean ring A can be represented as the ring of compact open subsets of a zero-dimensional locally compact Hausdorff space X (311I); X is just the set of surjective ring homomorphisms from A onto Z2 (311E). (b) If A and B are Boolean rings and π : A → B is a function, then the following are equiveridical: (i) π is a ring homomorphism; (ii) π(a \ b) = πa \ πb for all a, b ∈ A; (iii) π0 = 0 and π(a ∪ b) = πa ∪ πb, π(a ∩ b) = πa ∩ πb for all a, b ∈ A. P P See 312H. To prove (ii)⇒(iii), observe that if a, b ∈ A then π(a ∩ b) = πa \ π(a \ b) = πa ∩ πb. Q Q (c) If A and B are Boolean rings and π : A → B is a ring homomorphism, then π is order-continuous iff inf π[A] = 0 whenever A ⊆ A is non-empty and downwards-directed and inf A = 0 in A; while π is sequentially order-continuous iff inf n∈N πan = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. (See 313L.) (d) The following will be a particularly important type of Boolean ring for us. If (A, µ ¯) is a measure algebra, then the ideal Af = {a : a ∈ A, µ ¯a < ∞} is a Boolean ring in its own right. Now suppose that (B, ν¯) is another measure algebra and Bf ⊆ B the corresponding ring of elements of finite measure. We

361D

S

277

can say that a ring homomorphism π : Af → Bf is measure-preserving if ν¯πa = µ ¯a for every a ∈ Af . f Now in this case π is order-continuous. P P If A ⊆ A is non-empty, downwards-directed and has infimum 0, then inf a∈A µ ¯a = 0, by 321F; but this means that inf a∈A ν¯πa = 0, and inf π[A] = 0 in Bf . Q Q 361B Definition Let A be a Boolean ring and U a linear space. A function ν : A → U is finitely additive, or just additive, if ν(a ∪ b) = νa + νb whenever a, b ∈ A and a ∩ b = 0. 361C Elementary facts We have the following immediate consequences of this definition, corresponding to 326B and 313L. Let A be a Boolean ring, U a linear space and ν : A → U an additive function. (a) ν0 = 0 (because ν0 = ν0 + ν0). (b) If a0 , . . . , am are dist in A, then ν(supj≤m aj ) =

Pm j=0

νaj .

(c) If B is another Boolean ring and π : B → A is a ring homomorphism, then νπ : B → U is additive. In particular, if B is a subring of A, then ν¹ B : B → U is additive. (d) If V is another linear space and T : U → V is a linear operator, then T ν : A → V is additive. (e) If U is a partially ordered linear space, then ν is order-preserving iff it is non-negative, that is, νa ≥ 0 for every a ∈ A. P P (α) If ν is order-preserving, then of course 0 = ν0 ≤ νa for every a ∈ A. (β) If ν is non-negative, and a ⊆ b in A, then νa ≤ νa + ν(b \ a) = νb. Q Q (f ) If U is a partially ordered linear space and ν is non-negative, then (i) ν is order-continuous iff inf ν[A] = 0 whenever A ⊆ A is a non-empty downwards-directed set with infimum 0 (ii) ν is sequentially order-continuous iff inf n∈N νan = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. P P (i) If ν is order-continuous, then of course inf ν[A] = ν0 = 0 whenever A ⊆ A is a non-empty downwardsdirected set with infimum 0. If ν satisfies the condition, and A ⊆ A is a non-empty upwards-directed set with supremum c, then {c \ a : a ∈ A} is downwards-directed with infimum 0 (313Aa), so that sup νa = sup νc − ν(c \ a) = νc − inf ν(c \ a) a∈A

a∈A

a∈A

(by 351Db) = νc. Similarly, if A ⊆ A is a non-empty downwards-directed set with infimum c, then inf a∈A νa = inf a∈A νc + ν(a \ c) = νc + inf a∈A ν(a \ c) = νc. Putting these together, ν is order-continuous. (ii) If ν is sequentially order-continuous, then of course inf n∈N νan = ν0 = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. If ν satisfies the condition, and han in∈N is a non-decreasing sequence in A with supremum c, then hc \ an in∈N is non-increasing and has infimum 0, so that supn∈N νan = supn∈N νc − ν(c \ an ) = νc − inf n∈N ν(c \ an ) = νc. Similarly, if han in∈N is a non-increasing sequence in A with infimum c, then han \ cin∈N is non-increasing and has infimum 0, so that inf n∈N νan = inf n∈N νc + ν(c \ an ) = νc + inf n∈N ν(c \ an ) = νc. Thus ν is sequentially order-continuous. Q Q 361D Construction Let A be a Boolean ring, and Z its Stone space. For a ∈ A write χa for the characteristic function of the open-and-compact subset b a of Z corresponding to a. Let S(A) be the linear subspace of RZ generated by {χa : a ∈ A}. Because χa is a bounded function for every a, S(A) is a subspace of the space `∞ (Z) of all bounded real-valued functions on Z (354Ha), and k k∞ is a norm on S(A). Because χa × χb = χ(a ∩ b) for all a, b ∈ A (writing × for pointwise multiplication of functions, as in 281B), S(A) is closed under ×.

278

361E of S(A).

Function spaces

361E

I give a portmanteau proposition running through the elementary, mostly algebraic, properties

Proposition Let A be a Boolean ring, with Stone space Z. Write S for S(A). (a) If a0 , . . . , an ∈ A, there are dist b0 , . . . , bm such that each ai is expressible as the supremum of some of the bj . Pm (b) If u ∈ S, it is expressible in the form j=0 βj χbj where b0 , . . . , bm are dist of A and βj ∈ R for each j. If all the bj are non-zero then kuk∞ = supj≤m |βj |. Pm (c) If u ∈ S is non-negative, it is expressible in the form j=0 βj χbj where b0 , . . . , bm are dist Pm of A and βj ≥ 0 for each j, and simultaneously in the form j=0 γj χcj where c0 ⊇ c1 ⊇ . . . ⊇ cm and γj ≥ 0 for every j. P m (d) If u = j=0 βj χbj where b0 , . . . , bm are dist of A and βj ∈ R for each j, then |u| = Pm j=0 |βj |χbj ∈ S. (e) S is a Riesz subspace of RZ ; in its own right, it is an Archimedean Riesz space. If A is a Boolean algebra, then S has an order unit χ1 and kuk∞ = min{α : α ≥ 0, |u| ≤ αχ1} for every u ∈ S. (f) The map χ : A → S is injective, additive, non-negative, a lattice homomorphism and order-continuous. (g) Suppose that u ≥ 0 in S and δ ≥ 0 in R. Then [[u > δ]] = sup{a : a ∈ A, (δ + η)χa ≤ u for some η > 0} is defined in A, and δχ[[u > δ]] ≤ u ≤ δχ[[u > 0]] ∨ kuk∞ [[u > δ]]. In particular, u ≤ kuk∞ χ[[u > 0]] and there is an η > 0 such that ηχ[[u > 0]] ≤ u. If u, v ≥ 0 in S then u ∧ v = 0 iff [[u > 0]] ∩ [[v > 0]] = 0. (h) Under ×, S is an f -algebra (352W) and a commutative normed algebra (2A4J). (i) For any u ∈ S, u ≥ 0 iff u = v × v for some v ∈ S. proof Write b a for the open-and-compact subset of Z corresponding to a ∈ A. (a) Induce on n. If n = 0 take m = 0, b0 = a0 . For the inductive step to n ≥ 1, take dist b0 , . . . , bm such that ai is the supremum of some of the bj for each i < n; now replace b0 , . . . , bm with b0 ∩ an , . . . , bm ∩ an , b0 \ an , . . . , bm \ an , an \ supj≤m bj to obtain a suitable string for a0 , . . . , an . Pn (b) If u = 0 set m = 0, b0 = 0, β0 = 0. Otherwise, express u as i=0 αi χai where ai ∈ A and α0 , . . . , αn are real numbers. Let b0 , . . . , bm be dist and such that every ai is expressible as the supremum of some Pk of the bj . Set γij = 1 if bj ⊆ ai , 0 otherwise, so that, because the bj are dist, χai = j=0 γij χbj for each i. Then Pn Pm Pm Pn u = i=0 αi χai = i=0 j=0 αi γij χbj = j=0 βj χbj , Pn setting βj = i=0 αi γij for each j ≤ k. The expression for kuk∞ is now obvious. (c)(i) If u ≥ 0P in (b), we must have βj = u(z) ≥ 0 whenever z ∈ bbj , so that βj ≥ 0 whenever bj 6= 0; m consequently u = j=0 |βj |χbj is in the required form. (ii) If we suppose that every βj is non-negative, and rearrange the of the sum so that β0 ≤ . . . ≤ βm , then we may set γ0 = β0 , γj = βj − βj−1 for 1 ≤ j ≤ m, cj = supi≥j bi to get Pm Pm Pm Pm Pi Pm j=0 γj χcj = j=0 i=j γj χbi = i=0 j=0 γj χbi = i=0 βi χbi = u. (d) is trivial, because bb0 , . . . , bbn are dist. (e) By (d), |u| ∈ S for every u ∈ S, so S is a Riesz subspace of RZ , and in itself is an Archimedean Riesz space. If A is a Boolean algebra, then χ1, the constant function with value 1, belongs to S, and is an order unit of S; while kuk∞ = min{α : α ≥ 0, |u(z)| ≤ α ∀ z ∈ Z} = min{α : α ≥ 0, |u| ≤ αχ1} for every u ∈ S.

361F

S

279

(f ) χ is injective because b a 6= bb whenever a 6= b. χ is additive because b a ∩ bb = ∅ whenever a ∩ b = 0. Of course it is non-negative. It is a lattice homomorphism because a 7→ b a : A → PZ and E 7→ χE : PZ → RZ are. To see that ν is order-continuous, take a non-empty downwards-directed A ⊆ A with infimum 0. ?? Suppose, if possible, that {χa : a ∈ A} does not have P infimum 0 in S. Then there is a u > 0 in S such that m u ≤ χa for every a ∈ A. Now u can be expressed as j=0 βj bj where b0 , . . . , bm are dist. There must be some z0 ∈ Z such that u(z0 ) > 0; take j such that z0 ∈ bbj , so that bj 6= 0 and βj = u(z0 ) > 0. But now, for any z ∈ bbj , a ∈ A, (χa)(z) ≥ u(z) = βj > 0 and z ∈ b a. As z is arbitrary, bbj ⊆ b a and bj ⊆ a; as a is arbitrary, bj is a non-zero lower bound for A in A. X X So inf χ[A] = 0 in S. As A is arbitrary, χ is order-continuous, by the criterion of 361C(f-i). Pm (g) Express u as j=0 βj χbj where b0 , . . . , bm are dist and every βj ≥ 0. Then given δ ≥ 0, η > 0, a ∈ A we have (δ + η)χa ≤ u iff a ⊆ sup{bj : j ≤ m, βj ≥ δ + η}. So [[u > δ]] = sup{bj : j ≤ m, βj > δ}. Writing c = [[u > δ]], d = [[u > 0]] = sup{bj : βj > 0}, we have u(z) ≤ kuk∞ if z ∈ b c, ≤ δ if z ∈ db \ b c, b = 0 if z ∈ / d. So δχc ≤ u ≤ kuk∞ χc ∨ δχd, as claimed. Taking δ = 0 we get u ≤ kuk∞ χd. Set η = min({1} ∪ {βj : j ≤ m, βj > 0}); then η > 0 and ηχd ≤ u. If u, v ∈ S + take η, η 0 > 0 such that ηχ[[u > 0]] ≤ u,

η 0 χ[[v > 0]] ≤ v.

Then min(η, η 0 )χ([[u > 0]] ∩ [[v > 0]]) ≤ u ∧ v ≤ max(kuk∞ , kvk∞ )χ([[u > 0]] ∩ [[v > 0]]). So u ∧ v = 0 =⇒ [[u > 0]] ∩ [[v > 0]] = 0 =⇒ u ∧ v = 0. (h) S is a commutative f -algebra and normed algebra just because it is a Riesz subspace of the f -algebra and commutative normed algebra `∞ (Z) and is closed under multiplication. Pm (i) If u = j=0 βj χbj where b0 , . . . , bm are dist and βj ≥ 0 for every j, then u = v × v where Pm p v = j=0 βj χbj . 361F

I now turn to the universal mapping theorems which really define the construction.

Theorem Let A be a Boolean ring, and U any linear space. Then there is a one-to-one correspondence between additive functions ν : A → U and linear operators T : S(A) → U , given by the formula ν = T χ. proof (a) The core of the proof isP the following observation. Let ν : A U be additive. If a0 , . . . , an ∈ A P→ n n and α0 , . . . , αn ∈ R are such that i=0 αi χai = 0 in S = S(A), then i=0 αi νai = 0 in U . P P As in 361E, we can find dist b0 , . . P . , bm such that eachPai is the supremum of some of the b ; set γ = 1 if bj ⊆ ai , 0 j ij Pn m m otherwise, so that χai = j=0 γij χbj , νai = j=0 γij νbj for each i. Set βj = i=0 αi γij , so that Pn Pm 0 = i=0 αi χai = j=0 βj χbj . Now βj νbj = 0 in U for each j, because either bj = 0 and νbj = 0, or there is some z ∈ bbj so that βj must be 0. Accordingly

280

Function spaces

0=

Pm j=0

βj νbj =

Pm Pn j=0

i=0

361F

αi γij νbi =

Pn i=0

αi νai . Q Q

Pn Pm (b) It follows that if u ∈ S is expressible simultaneously as i=0 αi χai = j=0 βj χbj , then Pn Pm i=0 αi χai + j=0 (−βj )χbj = 0 in S, so that

Pn i=0

and

αi νai + Pn i=0

Pm

j=0 (−βj )νbj

αi νai =

Pm j=0

= 0 in U ,

βj νbj .

We can therefore define T : S → U by setting Pn Pn T ( i=0 αi χai ) = i=0 αi νai whenever a0 , . . . , an ∈ A and α0 , . . . , αn ∈ R. (c) It is now elementary to check that T is linear, and that T χa = νa for every a ∈ A. Of course this last condition uniquely defines T , because {χa : a ∈ A} spans the linear space S. 361G Theorem Let A be a Boolean ring, and U a partially ordered linear space. Let ν : A → U be an additive function, and T : S(A) → U the corresponding linear operator. (a) ν is non-negative iff T is positive. (b) In this case, (i) if T is order-continuous or sequentially order-continuous, so is ν; (ii) if U is Archimedean and ν is order-continuous or sequentially order-continuous, so is T . (c) If U is a Riesz space, then the following are equiveridical: (i) T is a Riesz homomorphism; (ii) νa ∧ νb = 0 in U whenever a ∩ b = 0 in A; (iii) ν is a lattice homomorphism. proof Write S for S(A). (a) If T is positive, then surely νa = T χa ≥ 0 for P every a ∈ A, so ν = T χ is non-negative. If ν is m non-negative, and u ≥ 0 in S, then u is expressible as j=0 βj χbj where b0 , . . . , bm ∈ A and βj ≥ 0 for every j (361Ec), so that Pm T u = j=0 βj νbj ≥ 0. Thus T is positive. (b)(i) If T is order-continuous (resp. sequentially order-continuous) then ν = T χ is the composition of two order-continuous (resp. sequentially order-continuous) functions (361Ef), so must be order-continuous (resp. sequentially order-continuous). (ii) Assume now that U is Archimedean. α) Suppose that ν is order-continuous and that A ⊆ S is non-empty, downwards-directed and has (α infimum 0. Fix u0 ∈ A, set α = kuk∞ and a0 = [[u > 0]] (in the language of 361Eg). If α = 0 then of course inf u∈A T u = T u0 = 0. Otherwise, take any w ∈ U such that w 6≤ 0. Then there is some δ > 0 such that w 6≤ δνa0 , because U is Archimedean. Set A0 = {u : u ∈ A, u ≤ u0 }; because A is downwards-directed, A0 has the same lower bounds as A, and inf A0 = 0, while A0 is still downwards-directed. For u ∈ A0 set cu = [[u > δ]], so that δχcu ≤ u ≤ αχcu + δχ[[u > 0]] ≤ αχcu + δχa0 0

(361Eg). If u, v ∈ A and u ≤ v, then cu ⊆ cv , so C = {cu : u ∈ A0 } is downwards-directed; but if c is any lower bound for C in A, δχc is a lower bound for A0 in S, so is zero, and c = 0 in A. Thus inf C = 0 in A, and inf u∈A0 νcu = 0 in U . But this means, in particular, that α1 (w − δνa0 ) is not a lower bound for ν[C], and there is some u ∈ A0 such that α1 (w − δνa0 ) 6≤ νcu , that is, w − δνa0 6≤ ανcu , that is, w 6≤ δνa0 + ανcu . As u ≤ αχcu + δχa0 ,

361H

S

281

T u ≤ T (αχcu + δχa0 ) = ανcu + δνa0 , and w 6≤ T u. Since w is arbitrary, this means that 0 = inf T [A]; as A is arbitrary, T is order-continuous. β ) The argument for sequential order-continuity is essentially the same. Suppose that ν is se(β quentially order-continuous and that hun in∈N is a non-increasing sequence in S with infimum 0. Again set α = ku0 k, a0 = [[u0 > 0]]; again we may suppose that α > 0; again take any w ∈ U such that w 6≤ 0. As before, there is some δ > 0 such that w 6≤ δνa0 . For n ∈ N set cn = [[un > δ]], so that δχcn ≤ un ≤ αχcn + δχa0 . The sequence hcn in∈N is non-increasing because hun in∈N is, and if c ⊆ cn for every n, then δχc ≤ un for every n, so is zero, and c = 0 in A. Thus inf n∈N cn = 0 in A, and inf n∈N νcn = 0 in U , because ν is sequentially order-continuous. Replacing A0 , C in the argument above by {un : n ∈ N}, {cn : n ∈ N} we find an n such that w 6≤ T un . Since w is arbitrary, this means that 0 = inf n∈N T un ; as hun in∈N is arbitrary, T is sequentially order-continuous. (c)(i)⇒(iii) If T : S(A) → U is a Riesz homomorphism, and ν = T χ, then surely ν is a lattice homomorphism because T and χ are. (iii)⇒(ii) is trivial. Pm (ii)⇒(i) If νa ∧ νb = 0 whenever a ∩ b = 0, then for any u ∈ S(A) we have an expression of u as j=0 βj χbj , where b0 , . . . , bm ∈ A are dist. Now Pm Pm Pm |T u| = | j=0 βj νbj | = j=0 |βj |νbj = T ( j=0 |βj |χbj ) = T (|u|) by 352Fb and 361Ed. As u is arbitrary, T is a Riesz homomorphism (352G). 361H Theorem Let A be a Boolean ring and U a Dedekind complete Riesz space. Suppose that ν : A → U is an additive function and T : S = S(A) → U the corresponding linear operator. Then T ∈ L∼ = L∼ (S; U ) iff {νb : b ⊆ a} is order-bounded in U for every a ∈ A, and in this case |T | ∈ L∼ corresponds to |ν| : A → U , defined by setting |ν|(a) = sup{

n X

|νai | : a0 , . . . , an ⊆ a are dist}

j=0

= sup{νb − ν(a \ b) : b ⊆ a} for every a ∈ A. proof (a) Suppose that T ∈ L∼ and a ∈ A. Then for any b ⊆ a, we have χb ≤ χa so |νb| = |T χb| ≤ |T |(χa). Accordingly {νb : b ⊆ a} is order-bounded in U . (b) Now suppose that {νb : b ⊆ a} is order-bounded in U for every a ∈ A. Then for any a ∈ A we can define wa = sup{|νb| : b ⊆ a}; in this case, νb − ν(a \ b) ≤ 2wa whenever b ⊆ a, so θa = supb ⊆ a νb − ν(a \ b) is also defined in U . Considering b = a, b = 0 we see that θa ≥ |νa|. Next, θ : A → U is additive. P P Take a1 , a2 ∈ A such that a1 ∩ a2 = 0; set a0 = a1 ∪ a2 . For each j ≤ 2 set Aj = {ν(aj ∩ b) − ν(aj \ b) : b ∈ A} ⊆ U . Then A0 ⊆ A1 + A2 , because ν(a0 ∩ b) − ν(a0 \ b) = ν(a1 ∩ b) − ν(a1 \ b) + ν(a2 ∩ b) − ν(a2 \ b) for every b ∈ A. But also A1 + A2 ⊆ A0 , because if b1 , b2 ∈ A then ν(a1 ∩ b1 ) − ν(a1 \ b1 ) + ν(a2 ∩ b2 ) − ν(a2 \ b2 ) = ν(a0 ∩ b) − ν(a0 \ b) where b = (a1 ∩ b1 ) ∪ (a2 ∩ b2 ). So A0 = A1 + A2 , and θa0 = sup A0 = sup A1 + sup A2 = θa1 + θa2 (351Dc). Q Q

282

Function spaces

361H

We therefore Pn have a corresponding positive operator T1 : S → U such that θ = T1 χ. But we also see that θa = sup{ i=0 |νai | : a0 , . . . , an ⊆ a are dist} for every a ∈ A. P P If a0 , . . . , an are dist and included in a, then Pn Pn i=0 |νai | ≤ i=0 θai = θ(supi≤n ai ) ≤ θa. On the other hand,

Pn θa ≤ supb ⊆ a |νb| + |ν(a \ b)| ≤ sup{ i=0 |νai | : a0 , . . . , an ⊆ a are dist}. Q Q

It follows that T ∈ L∼ P Take any u ≥ 0 in S. Set a = [[u > 0]] (361Eg) and α = kuk∞ . If 0 < |v| ≤ u, P. nP then v is expressible as i=0 αi χai where a0 , . . . , an are dist and no αi nor ai is zero. Since |v| ≤ αχa, we must have |αi | ≤ α, ai ⊆ a for each i. So Pn Pn Pn |T v| = | i=0 αi χai | ≤ i=0 |αi ||νai | ≤ α i=0 |νai | = αθa. Thus {|T v| : |v| ≤ u} is bounded above by αθa. As u is arbitrary, T ∈ L∼ . Q Q (c) Thus T ∈ L∼ iff ν is order-bounded on the sets {b : b ⊆ a}, and in this case the two formulae offered for |ν| are consistent and make |ν| = θ. Finally, θ = |T |χ. P P Take a ∈ A. If a0 , . . . , an ⊆ a are dist, then Pn Pn Pn i=0 |νai | = i=0 |T χai | ≤ i=0 |T |(χai ) ≤ |T |(χa); so θa ≤ |T |(χa). On the other hand, the argument at the end of (b) above shows that |T |(χa) ≤ θa for every a. Thus |T |(χa) = θa for every a ∈ A, as required. Q Q 361I Theorem Let A be a Boolean ring, U a normed space and ν : A → U an additive function. Give S = S(A) its norm k k∞ , and let T : S → U be the linear operator corresponding to ν. Then T is a bounded linear operator iff {νa : a ∈ A} is bounded, and in this case kT k = supa,b∈A kνa − νbk. proof (a) If T is bounded, then kνa − νbk = kT (χa − χb)k ≤ kT kkχa − χbk∞ ≤ kT k for every a ∈ A, so ν is bounded and supa,b∈A kνa − νbk ≤ kT k. (b)(i) For the converse, in 361Ec. If u ∈ S and u ≥ 0 and kuk∞ ≤ 1, Pidea Pmwe need a refinement of an m P If u = 0, take n = 0, c0 = 0, γ0 = 1. then u is expressible as i=0 γi χci wherePγi ≥ 0 and i=0 γi = 1. P n Otherwise, start from an expression u = j=0 γj χcj where c0 ⊇ . . . ⊇ cn and every γj is non-negative, as in 361Ec. We may suppose that cn 6= 0, in which case Pn j=0 γj = u(z) ≤ 1 Pn for every z ∈ b cn ⊆ Z, the Stone space of A. Set m = n + 1, cm = 0 and γm = 1 − j=0 γj to get the required form. Q Q 0 , γn0 ≥ 0 (ii) The next Pm numbers: if γ0 , . . . , γm , γ0 ,0 . . . P P Pn fact0 we need is an elementary property of real n m and i=0 γi = j=0 γj , then there are δij ≥ 0 such that γi = j=0 δij for every i ≤ m and γj = i=0 δij for every j ≤ n. P P This is just the case U = R of 352Fd. Q Q

(iii) Now suppose that ν is bounded; set α0 = supa∈A kνak < ∞. Then α = supa,b∈A kνa − νbk ≤ 2α0 is also finite. If u ∈ S and kuk∞ ≤ 1, then we can express u as u+ − u− where u+ , u− are non-negative and also of norm at most 1. By (i), we can express these as Pm Pn u+ = i=0 γi χci , u− = j=0 γj0 χc0j Pm Pn where all the γi , γj0 are non-negative and i=0 γi = j=0 γj0 = 1. Take hδij ii≤m,j≤n from (ii). Set cij = ci , c0ij = c0j for all i, j, so that Pm Pn Pm Pn u+ = i=0 j=0 δij χcij , u− = i=0 j=0 δij χc0ij , u=

Pm Pn i=0

j=0 δij (χcij

− χc0ij ),

361J

S

Tu = kT uk ≤

Pm Pn

Pm Pn i=0

j=0 δij (νcij

j=0 δij kνcij

i=0

− νc0ij k ≤

283

− νc0ij ), Pm Pn i=0

j=0 δij α

= α.

As u is arbitrary, T is a bounded linear operator and kT k ≤ α, as required. 361J The last few paragraphs describe the properties of S(A) in of universal mapping theorems. The next theorem looks at the construction as a functor which converts Boolean algebras into Riesz spaces and ring homomorphisms into Riesz homomorphisms. Theorem Let A and B be Boolean rings and π : A → B a ring homomorphism. (a) We have a Riesz homomorphism Tπ : S(A) → S(B) given by the formula Tπ (χa) = χ(πa) for every a ∈ A. For any u ∈ S(A), kTπ uk∞ = min{ku0 k∞ : u0 ∈ S(A), Tπ u0 = Tπ u}; in particular, kTπ uk∞ ≤ kuk∞ . Moreover, Tπ (u × u0 ) = Tπ u × Tπ u0 for all u, u0 ∈ S(A). (b) Tπ is surjective iff π is surjective, and in this case kvk∞ = min{kuk∞ : u ∈ S(A), Tπ u = v} for every v ∈ S(B). (c) The kernel of Tπ is just the set of those u ∈ S(A) such that π[[|u| > 0]] = 0, defining [[. . . > . . . ]] as in 361Eg. (d) Tπ is injective iff π is injective, and in this case kTπ uk∞ = kuk∞ for every u ∈ S(A). (e) Tπ is order-continuous iff π is order-continuous. (f) Tπ is sequentially order-continuous iff π is sequentially order-continuous. (g) If C is another Boolean ring and φ : B → C is another ring homomorphism, then Tφπ = Tφ Tπ : S(A) → S(C). proof (a) The map χπ : A → S(B) is additive (361Cc), so corresponds to a linear operator T = Tπ : S(A) → S(B), by 361F. P χ and π are both lattice homomorphisms, so χπ also is, and T is a Riesz homomorphism n (361Gc). If u = i=0 αi χai , where a0 , . . . , an are dist, then look at I = {i : i ≤ n, πai 6= 0}. We have Pn P T u = i=0 αi χ(πai ) = i∈I αi χ(πai ) and πa0 , . . . , πan are dist, so that where u0 =

kT uk∞ = supi∈I |αi | = ku0 k∞ ≤ supai 6=0 |αi | ≤ kuk∞ ,

P i∈I

αi χai , so that T u0 = T u. If a, a0 ∈ A, then T (χa × χa0 ) = T χ(a ∩ a0 ) = χπ(a ∩ a0 ) = χπa × χπa0 = T χa × T χa0 ,

so T is multiplicative. (b) If π is surjective, then T [S(A)] must be the linear span of {T (χa) : a ∈ A} = {χ(πa) : a ∈ A} = {χb : b ∈ B}, so is the whole ofP S(B). If T is surjective, and b ∈ B, then there must be a u ∈ A such that T u = χb. We n can express u as i=0 αi χai where a0 , . . . , an are dist; now Pn χb = T u = i=0 αi χ(πai ), and πa0 , . . . , πan are dist in B, so we must have b = supi∈I πai = π(supi∈I ai ) ∈ π[A], where I = {i : αi = 1}. As b is arbitrary, π is surjective. Of course the formula for kvk∞ is a consequence of the formula for kT uk∞ in (a). (c)(i) If π[[|u| > 0]] = 0 then |u| ≤ αχa, where α = kuk∞ , a = [[|u| > 0]], so |T u| = T |u| ≤ αT (χa) = αχ(πa) = 0, Pn and T u = 0. (ii) If u ∈ S(A) and T u = 0, express u as i=0 αi χai where a0 , . . . , an are dist and every αi is non-zero (361Eb). In this case Pn 0 = |T u| = T |u| = i=0 |αi |χ(πai ),

284

Function spaces

361J

so πai = 0 for every i, and π[[|u| > 0]] = π(supi≤n ai ) = supi≤n πai = 0. (d) If T is injective and a ∈ A \ {0}, then χ(πa) = T (χa) 6= 0, so πa 6= 0; as a is arbitrary, π is injective. If π is injective then π[[|u| > 0]] 6= 0 for every non-zero u ∈ S(A), so T is injective, by (c). Now the formula in (a) shows that T is norm-preserving. (e)(i) If T is order-continuous and A ⊆ A is a non-empty downwards-directed set with infimum 0 in A, let b be any lower bound for π[A] in B. Then χb ≤ χ(πa) = T (χa) for any a ∈ A. But T χ is order-continuous, by 361Ef, so inf a∈A T (χa) = 0, and b must be 0. As b is arbitrary, inf a∈A πa = 0; as A is arbitrary, π is order-continuous. (ii) If π is order-continuous, so is χπ : A → S(B), using 361Ef again; but now by 361G(b-ii) T must be order-continuous. (f )(i) If T is sequentially order-continuous, and han in∈N is a non-increasing sequence in A with infimum 0, let b be any lower bound for {πan : n ∈ N} in B. Then χb ≤ χ(πan ) = T (χan ) for any a ∈ A. But T χ is sequentially order-continuous so inf n∈N T (χan ) = 0, and b must be 0. As b is arbitrary, inf n∈N πan = 0; as A is arbitrary, π is sequentially order-continuous. (ii) If π is sequentially order-continuous, so is χπ : A → S(B); but now T must be sequentially order-continuous. (g) We need only check that Tφπ (χa) = χ(φ(πa)) = Tφ (χ(πa)) = Tφ T (χa) for every a ∈ A. 361K Proposition Let A be a Boolean algebra. For a ∈ A write Va for the solid linear subspace of S(A) generated by χa. Then a 7→ Va is a Boolean isomorphism between A and the algebra of projection bands in S(A). proof Write S for S(A). (a) The point is that, for any a ∈ A, (i) |u| ∧ |v| = 0 whenever u ∈ Va , v ∈ V1\a , (ii) Va + V1\a = S. P P (i) is just because χa ∧ χ(1 \ a) = 0. As for (ii), if w ∈ S then w = (w × χa) + (w × χ(1 \ a)) ∈ Va + V1\a . Q Q (b) Accordingly Va + Va⊥ ⊇ Va + V1\a = U and Va is a projection band (352R). Next, any projection band U ⊆ S is of the form Va . P P We know that χ1 = u + v where u ∈ U , v ∈ U ⊥ . Since |u| ∧ |v| = 0, u and v must be the characteristic functions of complementary subsets of Z, the Stone space of A. But {z : u(z) 6= 0} = {z : u(z) ≥ 1} must be of the form b a, where a = [[u > 0]], in which case u = χa and v = χ(1 \ a). Accordingly U ⊇ Va and U ⊥ ⊇ V1\a . But this means that U must be Va precisely. Q Q (c) Thus a 7→ Va is a surjective function from A onto the algebra of projection bands in S. Now a ⊆ b ⇐⇒ χa ∈ Vb ⇐⇒ Va ⊆ Vb , so a 7→ Va is order-preserving and bijective. By 312L it is a Boolean isomorphism. 361L Proposition Let X be a set, and Σ a ring of subsets of X, that is, a subring of the Boolean ring PX. Then S(Σ) can be identified, as ordered linear space, with the linear subspace of `∞ (X) generated by the characteristic functions of of Σ, which is a Riesz subspace of `∞ (X). The norm of S(Σ) corresponds to the uniform norm on `∞ (X), and its multiplication to pointwise multiplication of functions. proof Let Z be the Stone space of Σ, and for E ∈ Σ write χE for the characteristic function of E as a subset of X, χE ˆ for the characteristic function of the open-and-compact subset of Z corresponding to E.

361Xb

S

285

Of course χ : Σ → `∞ (X) is additive, so by 361F there is a linear operator T : S → `∞ (X), writing S for S(Σ), such that T (χE) ˆ = χE for every E ∈ Σ. Pm If u ∈ S, T u ≥ 0 iff u ≥ 0. P P Express u as ˆ j where E0 , . . . , Em are dist. Then j=0 βj χE Pm T u = j=0 βj χEj , so u ≥ 0 ⇐⇒ βj ≥ 0 whenever Ej 6= ∅ ⇐⇒ T u ≥ 0. Q Q But this means (α) that T u = 0 ⇐⇒ T u ≥ 0 & T (−u) ≥ 0 ⇐⇒ u ≥ 0 & − u ≥ 0 ⇐⇒ u = 0, so that T is injective and is a linear space isomorphism between S and its image S, which must be the linear space spanned by {χE : E ∈ Σ} (β) that T is an order-isomorphism between S and S. Because χE ∧ χF = 0 whenever E, F ∈ Σ and E ∩ F = ∅, T is a Riesz homomorphism and S is a Riesz subspace of `∞ (X) (361Gc). Now kuk∞ = inf{α : |u| ≤ αχX} ˆ = inf{α : |T u| ≤ αχX} = kT uk∞ for every u ∈ S. Finally, T (χE ˆ × χF ˆ ) = T (χ(E ˆ ∩ F )) = χ(E ∩ F ) = T (χE) ˆ × T (χF ˆ ) for all E, F ∈ Σ, so S is closed under pointwise multiplication and the multiplications of S, S are identified under T . 361M Proposition Let X be a set, Σ a ring of subsets of X, and I an ideal of Σ; write A for the quotient ring Σ/I. Let S be the linear span of {χE : E ∈ Σ} in RX , and write V = {f : f ∈ S, {x : f (x) 6= 0} ∈ I}. Then V is a solid linear subspace of S. Now S(A) becomes identified with the quotient Riesz space S/V , if for every E ∈ Σ we identify χ(E • ) ∈ S(A) with (χE)• ∈ S/V . If we give S its uniform norm inherited from `∞ (X), V is a closed linear subspace of S, and the quotient norm on S/V corresponds to the norm of S(A): kf • k = min{α : {x : |f (x)| > α} ∈ I}. If we write × for pointwise multiplication on S, then V is an ideal of the ring (S, +, ×), and the multiplication induced on S/V corresponds to the multiplication of S(A). proof Use 361J and 361L. We can identify S with S(Σ). Now the canonical ring homomorphism E 7→ E • corresponds to a surjective Riesz homomorphism T from S(Σ) to S(A) which takes χE to χ(E • ). For f ∈ S, [[|f | > 0]] is just {x : f (x) 6= 0}, so the kernel of T is just the set of those f ∈ S such that {x : f (x) 6= 0} ∈ I, which is V . So S(A) = T [S] ∼ = S/V . As noted in 361Ja, T (f × g) = T f × T g for all f , g ∈ S, so the multiplications of S/V and S(A) match. As for the norms, the norm of S(A) corresponds to the norm of S/V by the formulae in 361Ja or 361Jb. (To see that V is closed in S, we need note only that if f ∈ V then kT f k∞ = inf g∈V kf + gk∞ = inf g∈V kf − gk∞ = 0, so that T f = 0 and f ∈ V .) To check the formula for kf • k, take any f ∈ S. Express it as E0 , . . . , En ∈ Σ are dist. Set I = {i : Ei ∈ / I}; then

Pn i=0

αi χEi where

kT f k∞ = maxi∈I |αi | = min{α : {x : |f (x)| > α} ∈ I}. 361X Basic exercises (a) Let A be a Boolean ring and U a linear space. Show that a function ν : A → U is additive iff ν0 = 0 and ν(a ∪ b) + ν(a ∩ b) = νa + νb for all a, b ∈ A. >(b) Let U be an algebra over R, that is, a real linear space endowed with a multiplication × such that (U, +, ×) is a ring and α(w × z) = (αw) × z = w × (αz) for all w, z ∈ U and all α ∈ R. Let A be a Boolean ring, ν : A → U an additive function and T : S(A) → U the corresponding linear operator. Show that T is multiplicative iff ν(a ∩ b) = νa × νb for all a, b ∈ A.

286

Function spaces

361Xc

> (c) Let A be a Boolean ring, and U a Dedekind complete Riesz space. Suppose that ν : A → U is an additive function such that the corresponding linear operator T : S(A) → U belongs to L∼ = L∼ (S(A); U ). Show that T + ∈ L∼ corresponds to ν + : A → U , where ν + a = supb ⊆ a νb for every a ∈ A. (d) Let A and B be Boolean algebras. Show that there is a natural one-to-one correspondence between Boolean homomorphisms π : A → B and Riesz homomorphisms T : S(A) → S(B) such that T (χ1A ) = χ1B , given by setting T (χa) = χ(πa) for every a ∈ A. (e) Let A, B be Boolean rings and T : S(A) → S(B) a linear operator such that T (u × v) = T u × T v for all u, v ∈ S(A). Show that there is a ring homomorphism π : A → B such that T (χa) = χ(πa) for every a ∈ A. (f ) Let A and B be Boolean rings. Show that any isomorphism of the algebras S(A) and S(B) (using the word ‘algebra’ in the sense of 361Xb) must be a Riesz space isomorphism, and therefore corresponds to an isomorphism between A and B. (g) Let A, B be Boolean algebras and T : S(A) → S(B) a Riesz homomorphism. Show that there are a ring homomorphism π : A → B and a non-negative v ∈ S(B) such that T (χa) = v × χ(πa) for every a ∈ A. (h) Let A be a Boolean ring. Show that for any u ∈ S(A) the solid linear subspace of S(A) generated by u is a projection band in S(A). Show that the set of such bands is an ideal in the algebra of all projection bands, and is isomorphic to A. > (i) Let X be a set, Σ a σ-algebra of subsets of X. Show that the linear span S in RX of {χE : E ∈ Σ} is just the set of Σ-measurable functions f : X → R which take only finitely many values. (j) For any Boolean ring A, we may define its ‘complex S-space’ SC (A) as the linear span in CX of the characteristic functions of open-and-compact subsets of the Stone space Z of A. State and prove results corresponding to 361Ea-361Ed, 361Eh, 361F, 361L and 361M. 361Y P Further exercises (a) Let A be a Boolean ring. Let V be the linear space of all formal sums of n the form i=0 αi ai where α0 , . . . , αn ∈ R and a0 , . . . , an ∈ A. Let W ⊆ V be the linear subspace spanned by of V of the form (a ∪ b) − a − b where a, b ∈ A are dist. Define χ0 : A → V /W by taking χ0 a to be the image in V /W of a ∈ V . Show, without using the axiom of choice, that the pair (V /W, χ0 ) has the universal mapping property of (S(A), χ) as described in 361F and that V /W has a Riesz space structure, a norm and a multiplicative structure as described in 361D-361E. Prove results corresponding to 361E-361M. (b) Let A be a Boolean ring and U a Dedekind complete Riesz space. Let A ⊆ L∼ = L∼ (S(A); U ) be a non-empty set. Suppose that T˜ = sup A is defined in L∼ , and that ν˜ = T˜χ. Show that for any a ∈ A, Pn νã = sup{ i=0 Ti (χai ) : T0 , . . . , Tn ∈ A, a0 , . . . , an ⊆ a are dist, supi≤n ai = a}. (c) Let A be a Boolean algebra. Show that the algebra of all bands of S(A) can be identified with the Dedekind completion of A (314U). (d) Let A be a Boolean ring, and U a complex normed space. Let ν : A → U be an additive function and T : SC (A) → U the corresponding linear operator (cf. 361Xj). Show that (giving SC (A) its usual norm k k∞ ) Pn kT k = sup{k j=0 ζj νaj k : a0 , . . . , an ∈ A are dist, |ζj | = 1 for every j} if either is finite. (e) Let U be a Riesz space. Show that it is isomorphic to S(A), for some Boolean algebra A, iff it has an order unit and every solid linear subspace of U is a projection band. (f ) Let hAi ii∈I be a non-empty family of Boolean algebras, with free product A; write εi : Ai → A for the canonical maps, and

§362 intro.

S∼

287

C = {inf j∈J εj (aj ) : J ⊆ I is finite, aj ∈ Aj for every j ∈ J}. Suppose that U is a linear space and θ : C → U is such that θc = θ(c ∩ εi (a)) + θ(c ∩ εi (1 \ a)) whenever c ∈ C, i ∈ I and a ∈ Ai . Show that there is a unique additive function ν : A → U extending θ. (Hint: 326Q.) 361 Notes and comments The space S(A) corresponds of course to the idea of ‘simple function’ which belongs to the very beginnings of the theory of integration (122A). All that 361D is trying to do is to set up a logically sound description of this obvious concept which can be derived from the Boolean ring A itself. To my eye, there is a defect in the construction there. It relies on the axiom of choice, since it uses the Stone space; but none of the elementary properties of S(A) have anything to do with the axiom of choice. In 361Ya I offer an alternative construction which is in a formal sense more ‘elementary’. If you work through the suggestion there you will find, however, that the technical details become significantly more complicated, and would be intolerable were it not for the intuition provided by the Stone space construction. Of course this intuition is chiefly valuable in the finitistic arguments used in 361E, 361F and 361I; and for these arguments we really need the Stone representation only for finite Boolean rings, which does not depend on the axiom of choice. It is quite true that in most of this volume (and in most of this chapter) I use the axiom of choice without scruple and without comment. I mention it here only because I find myself using arguments dependent on choice to prove theorems of a type to which the axiom cannot be relevant. The linear space structure of S(A), together with the map χ, are uniquely determined by the first universal mapping theorem here, 361F. This result says nothing about the order structure, which needs the further refinement in 361Ga. What is striking is that the partial order defined by 361Ga is actually a lattice ordering, so that we can have a universal mapping theorem for functions to Riesz spaces, as in 361Gc and 361Ja. Moreover, the same ordering provides a happy abundance of results concerning order-continuous functions (361Gb, 361Je-361Jf). When the codomain is a Dedekind complete Riesz space, so that we have a Riesz space L∼ (S; U ), and a corresponding modulus function T 7→ |T | for linear operators, there are reasonably natural formulae for |T |χ in of T χ (361H); see also 361Xc and 361Yb. The multiplicative structure of S(A) is defined by 361Xb, and the norm by 361I. The Boolean ring A cannot be recovered from the linear space structure of S(A) alone (since this tells us only the cardinality of A), but if we add either the ordering or the multiplication of S(A) then A is easy to identify (361K, 361Xf). The most important Boolean algebras of measure theory arise either as algebras of sets or as their quotients, so it is a welcome fact that in such cases the spaces S(A) have straightforward representations in of the construction of A (361L-361M). In Chapter 24 I offered a paragraph in each section to sketch a version of the theory based on the field of complex numbers rather than the field of real numbers. This was because so many of the most important applications of these ideas involve complex numbers, even though (in my view) the ideas themselves are most clearly and characteristically expressed in of real numbers. In the present chapter we are one step farther away from these applications, and I therefore relegate complex numbers to the exercises, as in 361Xj and 361Yd.

362 S ∼ The next stage in our journey is the systematic investigation of linear functionals on spaces S = S(A). We already know that these correspond to additive real-valued functionals on the algebra A (361F). My purpose here is to show how the structure of the Riesz space dual S ∼ and its bands is related to the classes of additive functionals introduced in §§326-327. The first step is just to check the identification of the linear and order structures of S ∼ and the space M of bounded finitely additive functionals (362A); all the ideas needed for this have already been set out, and the basic properties of S ∼ are covered by the general results in §356. Next, we need to be able to describe the operations on M corresponding to the Riesz space operations

288

Function spaces

§362 intro.

| |, ∨, ∧ on S ∼ , and the band projections from S ∼ onto Sc∼ and S × ; these are dealt with in 362B, with a supplementary remark in 362D. In the case of measure algebras, we have some further important bands which present themselves in M , rather than in S ∼ , and which are treated in 362C. Since all these spaces are L-spaces, it is worth taking a moment to identify their uniformly integrable subsets; I do this in 362F. While some of the ideas here have interesting extensions to the case in which A is a Boolean ring without identity, these can I think be left to one side; the work of this section will be done on the assumption that every A is a Boolean algebra. 362A Theorem Let A be a Boolean algebra. Write S for S(A). (a) The partially ordered linear space of all finitely additive real-valued functionals on A may be identified with the partially ordered linear space of all real-valued linear functionals on S. (b) The linear space of bounded finitely additive real-valued functionals on A may be identified with the L-space S ∼ of order-bounded linear functionals on S. If f ∈ S ∼ corresponds to ν : A → R, then f + ∈ S ∼ corresponds to ν + , where ν + a = supb ⊆ a νb for every a ∈ A, and kf k = supa∈A νa − ν(1 \ a). (c) The linear space of bounded countably additive real-valued functionals on A may be identified with the L-space Sc∼ . (d) The linear space of completely additive real-valued functionals on A may be identified with the L-space S×. proof By 361F, we have a canonical one-to-one correspondence between linear functionals f : S → R and additive functionals νf : A → R, given by setting νf = f χ. (a) Now it is clear that νf +g = νf + νg , ναf = ανf for all f , g and α, so this one-to-one correspondence is a linear space isomorphism. To see that it is also an order-isomorphism, we need note only that νf is non-negative iff f is, by 361Ga. (b) Recall from 356N that, because S is a Riesz space with order unit (361Ee), S ∼ has a corresponding norm under which it is an L-space. (i) If f ∈ S ∼ , then supb∈A |νf b| = supb∈A |f (χb)| ≤ sup{|f (u)| : u ∈ S, |u| ≤ χ1} is finite, and νf is bounded. (ii) Now suppose that νf is bounded and that v ∈PS + . Then there is an α ≥ 0 such that v ≤ αχ1 n (361Ee). If u ∈ S and |u| ≤ v, then we can express u as i=0 αi χai where a0 , . . . , an are dist (361Eb); now |αi | ≤ α whenever ai 6= 0, so Pn Pn |f (u)| = | i=0 αi νf ai | ≤ α i=0 |νf ai | = α(νf c1 − νf c2 ) ≤ 2α supb∈A |νf b|, setting c1 = sup{ai : i ≤ n, νf ai ≥ 0}, c2 = sup{ai : i ≤ n, νf ai < 0}. This shows that {f (u) : |u| ≤ v} is bounded. As v is arbitrary, f ∈ S ∼ (356Aa). (iii) To check the correspondence between f + and νf+ , refine the arguments of (i) and (ii) as follows. Take any f ∈ S ∼ . If a ∈ A, νf+ a = supb ⊆ a νf b = supb ⊆ a f (χb) ≤ sup{f (u) : u ∈ S, 0 ≤ u ≤ χa} = f + (χa). Pn On the other hand, if u ∈ S and 0 ≤ u ≤ χa, then we can express u as i=0 αi χai where a0 , . . . , an are dist; now 0 ≤ αi ≤ 1 whenever ai 6= 0, so Pn f (u) = i=0 αi νf ai ≤ νf c ≤ νf+ a, where c = sup{ai : i ≤ n, νf ai ≥ 0}. As u is arbitrary, f + (χa) ≤ νf+ a. This shows that νf+ = f + χ is finitely additive, and that νf+ = νf + , as claimed.

S∼

362B

289

(iv) Now, for any f ∈ S ∼ ,

kf k = |f |(χ1) (356N) = (2f + − f )(χ1) (352D) = 2νf+ 1 − νf 1 (by (iii) just above) = sup 2νf a − νf 1 = sup νf a − νf (1 \ a). a∈A

a∈A

(c) If f ≥ 0 in S ∼ , then f is sequentially order-continuous iff νf is sequentially order-continuous (361Gb), that is, iff νf is countably additive (326Gc). Generally, an order-bounded linear functional belongs to Sc∼ iff it is expressible as the difference of two sequentially order-continuous positive linear functionals (356Ab), while a bounded finitely additive functional is countably additive iff it is expressible as the difference of two non-negative countably additive functionals (326H); so in the present context f ∈ Sc∼ iff νf is bounded and countably additive. (d) If f ≥ 0 in S ∼ , then f is order-continuous iff νf is order-continuous (361Gb), that is, iff νf is completely additive (326Kc). Generally, an order-bounded linear functional belongs to S × iff it is expressible as the difference of two order-continuous positive linear functionals (356Ac), while a finitely additive functional is completely additive iff it is expressible as the difference of two non-negative completely additive functionals (326M); so in the present context f ∈ S × iff νf is completely additive. 362B Spaces of finitely additive functionals The identifications in the last theorem mean that we can relate the Riesz space structure of S(A)∼ to constructions involving finitely additive functionals. I have already set out the most useful facts as exercises (326Yj, 326Ym, 326Yn, 326Yp, 326Yq); it is now time to repeat them more formally. Theorem Let A be a Boolean algebra. Write S = S(A), and let M be the Riesz space of bounded finitely additive real-valued functionals on A, Mσ ⊆ M the space of bounded countably additive functionals, and Mτ ⊆ Mσ the space of completely additive functionals. (a) For any µ, ν ∈ M , µ ∨ ν, µ ∧ ν and |ν| are defined by the formulae (µ ∨ ν)(a) = supb ⊆ a µb + ν(a \ b), (µ ∧ ν)(a) = inf b ⊆ a µb + ν(a \ b), |ν|(a) = supb ⊆ a νb − ν(a \ b) = supb,c ⊆ a νb − νc for every a ∈ A. Setting kνk = |ν|(1) = supa∈A νa − ν(1 \ a), M becomes an L-space. (b) Mσ and Mτ are projection bands in M , therefore L-spaces in their own right. In particular, |ν| ∈ Mσ for every ν ∈ Mσ , and |ν| ∈ Mτ for every ν ∈ Mτ . (c) The band projection Pσ : M → Mσ is defined by the formula (Pσ ν)(c) = inf{supn∈N νan : han in∈N is a non-decreasing sequence with supremum c} whenever c ∈ A and ν ≥ 0 in M . (d) The band projection Pτ : M → Mτ is defined by the formula (Pτ ν)(c) = inf{supa∈A νa : A is a non-empty upwards-directed set with supremum c} whenever c ∈ A and ν ≥ 0 in M .

290

Function spaces

362B

(e) If A ⊆ M is upwards-directed, then A is bounded above in M iff {ν1 : ν ∈ A} is bounded above in R, and in this case (if A 6= ∅) sup A is defined by the formula (sup A)(a) = supν∈A νa for every a ∈ A. (f) Suppose that µ, ν ∈ M . (i) The following are equiveridical: (α) ν belongs to the band in M generated by µ; (β) for every ² > 0 there is a δ > 0 such that |νa| ≤ ² whenever |µ|a ≤ δ; (γ) limn→∞ νan = 0 whenever han in∈N is a non-increasing sequence in A such that limn→∞ |µ|(an ) = 0. (ii) Now suppose that µ, ν ≥ 0, and let ν1 , ν2 be the components of ν in the band generated by µ and its complement. Then ν1 c = supδ>0 inf µa≤δ ν(c \ a),

ν2 c = inf δ>0 supa ⊆ c,µa≤δ νa

for every c ∈ A. proof (a) Of course µ ∨ ν = ν + (µ − ν)+ , µ ∧ ν = ν − (ν − µ)+ , |ν| = ν ∨ (−ν) (352D), so the formula of 362Ab gives (µ ∨ ν)(a) = νa + sup µb − νb = sup µb + ν(a \ b), b⊆a

b⊆a

(µ ∧ ν)(a) = νa − sup νb − µb = inf µb + ν(a \ b), b⊆a

b⊆a

|ν|(a) = sup νb − ν(a \ b) ≤ sup νb − νc = sup ν(b \ c) − ν(c \ b) b⊆a

b,c ⊆ a

b,c ⊆ a

≤ sup |ν|(b \ c) + |ν|(c \ b) = sup |ν|(b 4 c) ≤ |ν|(a). b,c ⊆ a

b,c ⊆ a

The formula offered for kνk corresponds exactly to the formula in 362Ab for the norm of the associated member of S(A)∼ ; because S(A)∼ is an L-space under its norm, so is M . × (b) By 362Ac-362Ad, Mσ and Mτ may be identified with S(A)∼ c and S(A) , which are projection bands in S(A)∼ (356B); so that Mσ and Mτ are projection bands in M , and are L-spaces in their own right (354O).

(c) Take any ν ≥ 0 in M . Set νσ c = inf{supn∈N νan : han in∈N is a non-decreasing sequence with supremum c} for every c ∈ A. Then of course 0 ≤ νσ c ≤ νc for every c. The point is that νσ is countably additive. P P Let hci ii∈N be a dist sequence in A, with supremum c. Then for any ² > 0 we have non-decreasing sequences han in∈N , hain in∈N , for i ∈ N, such that supn∈N an = c,

supn∈N ain = ci for i ∈ N,

supn∈N νan ≤ νσ c + ², supn∈N νain ≤ νσ ci + 2−i ² for every i ∈ N. Set bn = supi≤n ain for each n; then hbn in∈N is non-decreasing, and supn∈N bn = supi,n∈N ain = supi∈N ci = c, so νσ c ≤ sup νbn = sup n∈N

=

∞ X

n X

νain

n∈N i=0 ∞ X

sup νain ≤

i=0 n∈N

i=0

νσ ci + 2−i ² =

∞ X

νσ ci + 2².

i=0

On the other hand, han ∩ ci in∈N is a non-decreasing sequence with supremum c ∩ ci = ci for each i, so νσ ci ≤ supn∈N ν(an ∩ ci ), and

S∼

362B

∞ X

νσ ci ≤

i=0

∞ X

291

sup ν(an ∩ ci ) = sup

i=0 n∈N

∞ X

n∈N i=0

ν(an ∩ ci )

(because han in∈N is non-decreasing) ≤ sup νan n∈N

(because hci ii∈N is dist) ≤ νσ c + ². P∞

As ² is arbitrary, νσ c = i=0 νσ ci ; as hci ii∈N is arbitrary, νσ is countably additive. Q Q Thus νσ ∈ Mσ . On the other hand, if ν 0 ∈ Mσ and 0 ≤ ν 0 ≤ ν, then whenever c ∈ A and han in∈N is a non-decreasing sequence with supremum c, ν 0 c = supn∈N ν 0 an ≤ supn∈N νan . So we must have ν 0 c ≤ νσ c. This means that νσ = sup{ν 0 : ν 0 ∈ Mσ , ν 0 ≤ ν} = Pσ ν, as claimed. (d) The same ideas, with essentially elementary modifications, deal with the completely additive part. Take any ν ≥ 0 in M . Set ντ c = inf{supa∈A νa : A is a non-empty upwards-directed set with supremum c} for every c ∈ A. Then of course 0 ≤ ντ c ≤ νc for every c. The point is that ντ is completely additive. P P Note first that if c ∈ A, ² > 0 there is a non-empty upwards-directed A, with supremum c, such that supa∈A νa ≤ ντ c + ²νc; for if νc = 0 we can take A = {c}. Now let hci ii∈I be a partition of unity in A. Then for any ² > 0 we have non-empty upwards-directed sets A, Ai , for i ∈ I, such that sup A = 1,

sup Ai = ci for i ∈ I,

supa∈A νa ≤ ντ 1 + ²ν1,

supa∈Ai νa ≤ ντ ci + ²νci for every i ∈ I. Set B = {supi∈J ai : J ⊆ I is finite, ai ∈ Ai for every i ∈ J}; then B is non-empty and upwards-directed, and sup B = sup(

S i∈I

Ai ) = 1,

so X ντ 1 ≤ sup νb = sup{ νai : J ⊆ I is finite, ai ∈ Ai ∀ i ∈ J} b∈B

≤

X

i∈J

ντ ci + ²νci ≤ ²ν1 +

i∈I

X

ντ ci .

i∈I

On the other hand, A0i = {a ∩ ci : a ∈ A} is a non-empty upwards-directed set with supremum ci for each i, so ντ ci ≤ supa∈A0i νa, and X i∈I

ντ ci ≤

X i∈I

sup ν(a ∩ ci ) = sup

a∈A

a∈A

X

ν(a ∩ ci )

i∈I

≤ sup νa ≤ ντ 1 + ²ν1. a∈A

P

As ² is arbitrary, ντ c = i∈I ντ ci ; as hci ii∈I is arbitrary, ντ is completely additive, by 326N. Q Q Thus ντ ∈ Mτ . On the other hand, if ν 0 ∈ Mτ and 0 ≤ ν 0 ≤ ν, then whenever c ∈ A and A is a non-empty upwards-directed set with supremum c,

292

Function spaces

362B

ν 0 c = supa∈A ν 0 a ≤ supa∈A νa (using 326Kc). So we must have ν 0 c ≤ ντ c. This means that ντ = sup{ν 0 : ν 0 ∈ Mτ , ν 0 ≤ ν} = Pτ ν, as claimed. (e) If A is empty, of course it is bounded above in M , and {ν1 : ν ∈ A} = ∅ is bounded above in R; so let us suppose that A is not empty. In this case, if λ0 ∈ M is an upper bound for A, then λ0 1 is an upper bound for {ν1 : ν ∈ A}. On the other hand, if supν∈A ν1 = γ is finite, γ ∗ = sup{νa : ν ∈ A, a ∈ A} is finite. P P Fix ν0 ∈ A. Set γ1 = supa∈A |ν0 a| < ∞. Then for any ν ∈ A, a ∈ A there is a ν 0 ∈ A such that ν0 ∨ ν ≤ ν 0 , so that νa ≤ ν 0 a = ν 0 1 − ν 0 (1 \ a) ≤ γ − ν0 (1 \ a) ≤ γ + γ1 . So γ ∗ ≤ γ + γ1 < ∞. Q Q Set λa = supν∈A νa for every a ∈ A. Then λ : A → R is additive. P P If a, b ∈ A are dist, then λ(a ∪ b) = sup ν(a ∪ b) = sup νa + νb = sup νa + sup νb ν∈A

ν∈A

ν∈A

ν∈A

(because A is upwards-directed) = λa + λb. Q Q Also λa ≤ γ ∗ for every a, so |λa| = max(λa, −λa) = max(λa, λ(1 \ a) − λ1) ≤ γ ∗ + |λ1| for every a ∈ A, and λ is bounded. This shows that λ ∈ M , so that A is bounded above in M . Of course λ must be actually the least upper bound of A in M . α)⇒(β β ) Suppose that ν belongs to the band in M generated by µ, that is, |ν| = supn∈N |ν| ∧ n|µ| (f )(i)(α (352Vb). Let ² > 0. Then there is an n ∈ N such that |ν|(1) ≤ 12 ² + (|ν| ∧ n|µ|)(1) ((e) above). Set 1 δ = 2n+1 ² > 0. If |µ|(a) ≤ δ, then |νa| ≤ |ν|(a) = (|ν| ∧ n|µ|)(a) + (|ν| − |ν| ∧ n|µ|)(a) 1 2

≤ n|µ|(a) + (|ν| − |ν| ∧ n|µ|)(1) ≤ nδ + ² ≤ ². So (β) is satisfied. β )⇒(α α) Suppose that ν does not belong to the band in M generated by |µ|. Then there is a ν1 > 0 (β such that ν1 ≤ |ν| and ν1 ∧ |µ| = 0 (353C). For any δ > 0, there is an a ∈ A such that ν1 (1 \ a) + |µ|(a) ≤ min(δ, 12 ν1 1) ((a) above); now |µ|(a) ≤ δ but |ν|(a) ≥ ν1 a = ν1 1 − ν1 (1 \ a) ≥ ν1 1 − 12 ν1 1 = 12 ν1 1. Thus µ, ν do not satisfy (β) (with ² = 21 ν1 1). β )⇒(γγ ) is trivial. (β α) Observe first that if hck ik∈N is a non-increasing sequence in A such that limk→∞ µck = 0, (γγ )⇒(α then limk→∞ ν + ck = 0. P P Let ² > 0. Because ν + ∧ ν − = 0, there is a b ∈ A such that ν + b + ν − (1 \ b) ≤ ², by part (a). Now hck \ bik∈N is non-increasing and limk→∞ µ(ck \ b) = 0, so limk→∞ ν(ck \ b) = 0 and lim sup ν + ck = lim sup ν + (ck ∩ b) + ν(ck \ b) + ν − (ck \ b) k→∞

k→∞

≤ ν + b + ν − (1 \ b) ≤ ². As ² is arbitrary, limk→∞ ν + ck = 0. Q Q

S∼

362C

293

?? Now suppose, if possible, that ν + does not belong to the band generated by µ. Then there is a ν1 > 0 such that ν1 ≤ ν + and ν1 ∧ |µ| = 0. Set ² = 14 ν1 1 > 0. For each n ∈ N, we can choose an ∈ A such that |µ|an + ν1 (1 \ an ) ≤ 2−n ², by part (a) again. For n ≥ k, set bkn = supk≤i≤n ai ; then Pn |µ|bkn ≤ i=k |µ|ai ≤ 2−k+1 ², and hbkn in≥k is non-decreasing. Set γk = supn≥k ν1 bkn and choose m(k) ≥ k such that ν1 bk,m(k) ≥ γk −2−k ². Setting bk = bk,m(k) , we see that bk ∪ bk+1 = bkn where n = max(m(k), m(k + 1)), so that ν1 (bk ∪ bk+1 ) ≤ γk ≤ ν1 bk + 2−k ² and ν1 (bk+1 \ bk ) ≤ 2−k ². Set ck = inf i≤k bi for each k; then ν1 (bk+1 \ ck+1 ) = ν1 (bk+1 \ ck ) ≤ ν1 (bk+1 \ bk ) + ν1 (bk \ ck ) ≤ 2−k ² + ν1 (bk \ ck ) for each k; inducing on k, we see that ν1 (bk \ ck ) ≤

Pk−1 i=0

2−i ² ≤ 2²

for every k. This means that ν + ck ≥ ν1 ck ≥ ν1 bk − 2² ≥ ν1 ak − 2² = ν1 1 − ν1 (1 \ ak ) − 2² ≥ 4² − ² − 2² = ² for every k ∈ N. On the other hand, hck ik∈N is a non-decreasing sequence and |µ|ck ≤ |µ|bk ≤ 2−k+1 ² for every k, which contradicts the paragraph just above. X X This means that ν + must belong to the band generated by µ. Similarly ν − = (−ν)+ belongs to the band generated by µ and ν = ν + + ν − also does. (ii) Take c ∈ A. Set β1 = supδ>0 inf µa≤δ ν(c \ a),

β2 = inf δ>0 supa ⊆ c,µa≤δ νa.

Then β1 = supδ>0 inf a ⊆ c,µa≤δ ν(c \ a) = νc − β2 . Take any ² > 0. Because ν1 belongs to the band generated by µ, part (i) tells us that there is a δ > 0 such that ν1 a ≤ ² whenever µa ≤ δ. In this case, if µa ≤ δ, ν(c \ a) = νc − ν(c ∩ a) ≥ νc − ² ≥ ν1 c − ²; thus β1 ≥ inf µa≤δ ν(c \ a) ≥ ν1 c − ². As ² is arbitrary, β1 ≥ ν1 c. On the other hand, given ², δ > 0, there is an a ⊆ c such that µa + ν2 (c \ a) ≤ min(δ, ²), because µ ∧ ν2 = 0 (using (a) again). In this case, of course, µa ≤ δ, while νa ≥ ν2 a = ν2 c − ν2 (c \ a) ≥ ν2 c − ². Thus supa ⊆ c,µa≤δ νa ≥ ν2 c − ². As δ is arbitrary, β2 ≥ ν2 c − ². As ² is arbitrary, β2 ≥ ν2 c; but as β1 + β2 = νc = ν1 c + ν2 c, βi = νi c for both i, as claimed. 362C The formula in 362B(f-i) has, I hope, already reminded you of the concept of ‘absolutely continuous’ additive functional from the Radon-Nikod´ ym theorem (Chapter 23, §327). The expressions in 362Bf are limited by the assumption that µ, like ν, is finite-valued. If we relax this we get an alternative version of some of the same ideas. Theorem Let (A, µ ¯) be a measure algebra. Write S = S(A), and let M be the Riesz space of bounded finitely additive real-valued functionals on A. Write Mac = {ν : ν ∈ M is absolutely continuous with respect to µ ¯} (see 327A),

294

Function spaces

362C

Mtc = {ν : ν ∈ M is continuous with respect to the measure-algebra topology on A}, Mt = {ν : ν ∈ M , |ν|1 = supµ¯a<∞ |ν|a}. Then Mac , Mtc and Mt are bands in M . proof (a)(i) It is easy to check that Mac is a linear subspace of M . (ii) If ν ∈ Mac , ν 0 ∈ M and |ν 0 | ≤ |ν| then ν 0 ∈ Mac . P P Given ² > 0 there is a δ > 0 such that |νa| ≤ 12 ² whenever µ ¯a ≤ δ. Now |ν 0 a| ≤ |ν 0 |(a) ≤ |ν|(a) ≤ 2 supc ⊆ a |νc| ≤ ² (using the formula for |ν| in 362Ba) whenever µ ¯a ≤ δ. As ² is arbitrary, ν 0 is absolutely continuous. Q Q (iii) If A ⊆ Mac is non-empty and upwards-directed and ν = sup A in M , then ν ∈ Mac . P P Let ² > 0. Then there is a ν 0 ∈ A such that ν1 ≤ ν 0 1 + 21 ² (362Be). Now there is a δ > 0 such that |νa| ≤ 12 ² whenever µ ¯a ≤ δ. If now µ ¯a ≤ δ, |νa| ≤ |ν 0 a| + (ν − ν 0 )(a) ≤ 21 ² + (ν − ν 0 )(1) ≤ ². As ² is arbitrary, ν is absolutely continuous with respect to µ ¯. Q Q Putting these together, we see that Mac is a band. (b)(i) We know that Mtc consists just of those ν ∈ M which are continuous at 0 (327Bc). Of course this is a linear subspace of M . (ii) If ν ∈ Mtc , ν 0 ∈ M and |ν 0 | ≤ |ν| then |ν| ∈ Mtc . P P Write Af = {d : d ∈ A, µ ¯d < ∞}. Given ² > 0 1 f ¯(a ∩ d) ≤ δ. Now there are d ∈ A , δ > 0 such that |νa| ≤ 2 ² whenever µ |ν 0 a| ≤ |ν 0 |(a) ≤ |ν|(a) ≤ 2 supc ⊆ a |νc| ≤ ² whenever µ ¯(a ∩ d) ≤ δ. As ² is arbitrary, ν 0 is continuous at 0 and belongs to Mtc . Q Q (iii) If A ⊆ Mtc is non-empty and upwards-directed and ν = sup A in M , then ν ∈ Mtc . P P Let ² > 0. Then there is a ν 0 ∈ A such that ν1 ≤ ν 0 1 + 21 ² (362Be). There are d ∈ Af , δ > 0 such that |νa| ≤ 21 ² whenever µ ¯(a ∩ d) ≤ δ. If now µ ¯(a ∩ d) ≤ δ, |νa| ≤ |ν 0 a| + (ν − ν 0 )(a) ≤ 21 ² + (ν − ν 0 )(1) ≤ ². As ² is arbitrary, ν is continuous at 0, therefore belongs to Mtc . Q Q Putting these together, we see that Mtc is a band. (c)(i) Mt is a linear subspace of M . P P Suppose that ν1 , ν2 ∈ Mt and α ∈ R. Given ² > 0, there are a1 , ² a2 ∈ Af such that |ν1 |(1 \ a1 ) ≤ 1+|α| , |ν2 |(1 \ a2 ) ≤ ². Set a = a1 ∪ a2 ; then µ ¯a < ∞ and |ν1 + ν2 |(1 \ a) ≤ 2²,

|αν1 |(1 \ a) ≤ ².

As ² is arbitrary, ν1 + ν2 and αν1 belong to Mt ; as ν1 , ν2 and α are arbitrary, Mt is a linear subspace of M . Q Q (ii) If ν ∈ Mt , ν 0 ∈ M and |ν 0 | ≤ |ν| then inf µ¯a<∞ |ν 0 |(1 \ a) ≤ inf µ¯a<∞ |ν|(1 \ a) = 0, so ν 0 ∈ Mt . Thus Mt is a solid linear subspace of M . (iii) If A ⊆ Mt+ is non-empty and upwards-directed and ν = sup A is defined in M , then ν ∈ Mt . P P |ν|1 = ν1 = supν 0 ∈A ν 0 1 = supν 0 ∈A,¯µa<∞ ν 0 a = supµ¯a<∞ νa. As A is arbitrary, ν ∈ Mt . Q Q Thus Mt is a band in M .

S∼

362Xb

362D

295

For semi-finite measure algebras, among others, the formula of 362Bd takes a special form.

Proposition Let A be a weakly (σ, ∞)-distributive Boolean algebra. Let M be the space of bounded finitely additive functionals on A, Mτ ⊆ M the space of completely additive functionals, and Pτ : M → Mτ the band projection, as in 362B. Then for any ν ∈ M + and c ∈ A there is a non-empty upwards-directed set A ⊆ A with supremum c such that (Pτ ν)(c) = supa∈A νa; that is, the ‘inf’ in 362Bd can be read as ‘min’. proof By 362Bd, we can find for each n a non-empty upwards-directed An , with supremum c, such that supa∈An νa ≤ (Pτ ν)(c) + 2−n . Set Bn = {c \ a : a ∈ An } for each n, so that Bn is downwards-directed and has infimum 0, and B = {b : for every n ∈ N there is a b0 ∈ Bn such that b ⊇ b0 } is also a downwards-directed set with infimum 0 (316H). Consequently A = {c \ b : b ∈ B} is upwardsdirected and has supremum c. Moreover, for any n ∈ N and a ∈ A, there is an a0 ∈ An such that a ⊆ a0 ; so, using 362Bd again and referring to the choice of An , (Pτ ν)(c) ≤ supa∈A νa ≤ supa0 ∈An νa0 ≤ (Pτ ν)(c) + 2−n . As n is arbitrary, A has the required property. 362E Uniformly integrable sets The spaces S ∼ , Sc∼ and S × of 362A, or, if you prefer, the spaces M , Mσ , Mτ , Mac , Mtc , Mt of 362B-362C, are all L-spaces, and any serious study of them must involve a discussion of their uniformly integrable ( = relatively weakly compact) subsets. The basic work has been done in 356O; I spell out its application in this context. Theorem Let A be a Boolean algebra and M the L-space of bounded finitely additive functionals on A. Then a norm-bounded set C ⊆ M is uniformly integrable iff limn→∞ supν∈C |νan | = 0 for every dist sequence han in∈N in A. proof Write C˜ for the set {f : f ∈ S(A)∼ , f χ ∈ C}. Because the map f 7→ f χ is a normed Riesz space isomorphism between S ∼ and M , C˜ is uniformly integrable in M iff C is uniformly integrable in S ∼ . (a) Suppose that C is uniformly integrable and that han in∈N is a dist sequence in A. Then hχan in∈N is a dist order-bounded sequence in S ∼ , while C˜ is uniformly integrable, so limn→∞ supf ∈C˜ |f (χan )| = 0, by 356O; but this means that limn→∞ supν∈C |νan | = 0. Thus the condition is satisfied. (b) Now suppose that C is not uniformly integrable. By 356O, in the other direction, there is a dist sequence hun in∈N in S such that 0 ≤ un ≤ χ1 for each n and lim supn→∞ supf ∈C˜ |f (un )| > 0. For each n, take cn = [[un > 0]] (361Eg); then 0 ≤ un ≤ χcn and hcn in∈N is dist. Now lim sup sup |ν|(cn ) = lim sup sup |f |(χcn ) n→∞ ν∈C

n→∞ f ∈C ˜

≥ lim sup sup |f (un )| > 0. n→∞ f ∈C ˜

So if we choose νn ∈ C such that |νn |(cn ) ≥ 21 supν∈C |ν|(cn ), we shall have lim supn→∞ |νn |(cn ) > 0. Next, for each n, we can find an ⊆ cn such that |νn an | ≥ 21 |νn |(cn ), so that lim supn∈N supν∈C |νan | ≥ lim supn→∞ |νn an | > 0. Since han in∈N , like hcn in∈N , is dist, the condition is not satisfied. This completes the proof. 362X Basic exercises (a) Let A be a Dedekind σ-complete Boolean algebra and ν1 , ν2 two countably additive functionals on A. Show that |ν1 |∧|ν2 | = 0 in the Riesz space of bounded finitely additive functionals on A iff there is a c ∈ A such that ν1 a = ν1 (a ∩ c) and ν2 a = ν2 (a \ c) for every a ∈ A. (b) Let (A, µ ¯) be a measure algebra. Write M , Mac as in 362C. Show that for any non-negative ν ∈ M , the component νac of ν in Mac is given by the formula νac c = supδ>0 inf µ¯a≤δ ν(c \ a).

296

Function spaces

362Xc

(c) Let (A, µ ¯) be a measure algebra. Write M , Mt as in 362C. (i) Show that Mt is just the set of those ν ∈ M such that νa = limb→F ν(a ∩ b) for every a ∈ A, where F is the filter on A generated by the sets {b : b ∈ Af , b ⊇ b0 } as b0 runs through the elements of A of finite measure. (ii) Show that the complementary band Mt⊥ of Mt in M is just the set of ν ∈ M such that νa = 0 whenever µ ¯a < ∞. (iii) Show that for any ν ∈ M , its component νt in Mt is given by the formula νt a = limb→F ν(a ∩ b) for every a ∈ A. (d) Let (A, µ ¯) be a measure algebra. Write M , Mσ , Mτ , Mac , Mtc and Mt as in 362B-362C. Show that (i) Mσ ⊆ Mac (ii) Mac ∩ Mt = Mtc ⊆ Mτ (iii) if (A, µ ¯) is σ-finite, then Mσ = Mt ∩ Mac . (e) Let A be a Boolean algebra. Let us say that a non-zero finitely additive functional ν : A → R is an atom if whenever a, b ∈ A and a ∩ b = 0 then at least one of νa, νb is zero. Show that for a non-zero finitely additive functional ν the following are equiveridical: (i) ν is an atom; (ii) ν is bounded and |ν| is an atom; (iii) ν is bounded and the corresponding linear functional f|ν| = |fν | ∈ S(A)∼ is a Riesz homomorphism; (iv) there are a multiplicative linear functional f : S(A) → R and an α ∈ R such that νa = αf (χa) for every a ∈ A. Show that a completely additive functional ν : A → R is an atom iff there are a ∈ A, α ∈ R \ {0} such that a is an atom in A and νb = α when a ⊆ b, 0 when a ∩ b = 0. (f ) Let A be a Boolean algebra. Let us say that a bounded finitely additive functional ν : A → R is atomless if for every ² > 0 there is a finite partition C of unity in A such that |ν|c ≤ ² for every c ∈ C (cf. 326Ya). (i) Show that the atomless functionals form a band Mc in the Riesz space M of all bounded finitely additive functionals on A. (ii) Show that the complementary band Mc⊥ consists of just those ν ∈ M P expressible as a sum i∈I νi of countably many atoms νi ∈ M . (iii) Show that if A is purely atomic then an atomless completely additive functional on A must be 0. (g) Let X be a set and Σ an algebra of subsets of X. Let M be the Riesz space of bounded finitely additive functionals on Σ, Mτ P the space of completely additive functionals and Mp the space of functionals expressible in the form νE = x∈E αx for some absolutely summable family hαx ix∈X of real numbers. (i) Show that Mp is a band in M . (ii) Show that if all singleton subsets of X belong to Σ then Mp = Mτ . (iii) Show that if Σ is a σ-algebra then every member of Mp is countably additive. (iv) Show that if X is a compact zero-dimensional Hausdorff space and Σ is the algebra of open-and-closed subsets of X then the complementary band Mp⊥ of Mp in M is the band Mc of atomless functionals described in 362Xf. (h) Let (X, Σ, µ) be a measure space. Let M be the Riesz space of bounded finitely additive functionals on Σ and Mσ the space of bounded countably additive functionals. Let Mtc , Mac be the spaces of truly continuous and bounded absolutely continuous additive functionals as defined in 232A. Show that Mtc and Mac are bands in M and that Mtc ⊆ Mσ ∩ Mac . Show that if µ is σ-finite then Mtc = Mσ ∩ Mac . (i) Let A be a Boolean algebra and M the Riesz space of bounded finitely additive functionals on A. (i) For any non-empty downwards-directed set A ⊆ A set NA = {ν : ν ∈ M, inf a∈A |ν|a = 0}. Show that NA is a band in M . (ii) For any non-empty set A of non-empty downwards-directed sets in A set MA = {ν : ν ∈ M, inf a∈A |ν|a = 0 ∀ A ∈ A}. Show that MA is a band in M . (iii) Explain how to represent as such MA the bands Mσ , Mτ , Mt , Mac , Mtc described above, and also any band generated by a single element of M . (iv) Suppose, in (ii), that A has the property that for any A, A0 ∈ A there is a B ∈ A such that for every b ∈ B there are a ∈ A, a0 ∈ A0 such that a ∪ a0 ⊆ b. Show that for any non-negative ν ∈ M , the component ν1 of ν in MA is given by the formula ν1 c = inf A∈A supa∈A ν(c \ a), so that the component ⊥ ν2 of ν in MA is given by the formula ν2 c = supA∈A inf a∈A ν(c ∩ a). (Cf. 356Yb.) 362Y Further exercises (a) Let A be a Boolean algebra. Let C be the band algebra of the Riesz space M of bounded finitely additive functionals on A (353B). Show that the bands Mσ , Mτ , Mc (362B, 362Xe, 362Xf) generate a subalgebra C0 of C with at most six atoms. Give an example in which C0 has six atoms. How many atoms can it have if (i) A is atomless (ii) A is purely atomic (iii) A is Dedekind σ-complete? (b) Let (A, µ ¯) be a measure algebra. Let C be the band algebra of the Riesz space M of bounded finitely additive functionals on A. Show that the bands Mσ , Mτ , Mc , Mac , Mtc , Mt (362B, 362C, 362Xe, 362Xf) generate a subalgebra C0 of C with at most twelve atoms. Give an example in which C0 has twelve atoms. How many atoms can it have if (i) A is atomless (ii) A is purely atomic (iii) (A, µ ¯) is semi-finite (iv) (A, µ ¯) is localizable (v) (A, µ ¯) is σ-finite (vi) (A, µ ¯) is totally finite?

362 Notes

S∼

297

(c) Give an example of a set X, a σ-algebra Σ of subsets of X, and a functional in Mp (as defined in 362Xg) which is not completely additive. (d) Let U be a Riesz space and f , g ∈ U ∼ . Show that the following are equiveridical: (α) g is in the band of U ∼ generated by f ; (β) for every u ∈ U + , ² > 0 there is a δ > 0 such that |g(v)| ≤ ² whenever 0 ≤ v ≤ u and |f |(v) ≤ δ; (γ) limn→∞ g(un ) = 0 whenever hun in∈N is a non-increasing sequence in U + and limn→∞ f (un ) = 0. (Hint: 362B(f-i).) (e) Let A be a weakly σ-distributive Boolean algebra (316Yg). Show that the ‘inf’ in the formula for Pσ ν in 362Bc can be replaced by ‘min’. (f ) Let A be any Boolean algebra and M the space of bounded finitely additive functionals on A. Let C ⊆ M be such that supν∈C |νa| < ∞ for every a ∈ A. (i) Suppose that supn∈N supν∈C |νan | is finite for every dist sequence han in∈N in A. Show that C is norm-bounded. (ii) Suppose that limn→∞ supν∈C |νan | = 0 for every dist sequence han in∈N in A. Show that C is uniformly integrable. (g) Let A be a Boolean algebra and Mτ the space of completely additive functionals on A. Let C ⊆ Mτ be such that supν∈C |νa| < ∞ for every atom a ∈ A. (i) Suppose that supn∈N supν∈C |νan | is finite for every dist sequence han in∈N in A. Show that C is norm-bounded. (ii) Suppose that limn→∞ supν∈C |νan | = 0 for every dist sequence han in∈N in A. Show that C is uniformly integrable. (h) Let A be a Dedekind σ-complete Boolean algebra and hνn in∈N a sequence of countably additive realvalued functionals on A. Suppose that νa = limn→∞ νn a is defined in R for every a ∈ A. Show that ν is countably additive and that {νn : n ∈ N} is uniformly integrable. (Hint: 246Yg.) Show that if every νn is completely additive, so is ν. (i) Let A be a Boolean algebra, M the Riesz space of bounded finitely additive functionals on A, and Mc ⊆ M the space of atomless functionals (362Xf). Show that for a non-negative ν ∈ M the component νc of ν in Mc is given by the formula Pn νc a = inf δ>0 sup{ i=0 νai : a0 , . . . , an ⊆ a are dist, νai ≤ δ for every i} for each a ∈ A. (j) Let A be a Boolean algebra and M the L-space of bounded additive real-valued functionals on A. Show that the complexification of M , as defined in 354Yk, can be identified with the Banach space of bounded additive functionals ν : A → C, writing Pn kνk = sup{ i=0 |νai | : a0 , . . . , an are dist elements of A} for such ν. 362 Notes and comments The Boolean algebras most immediately important in measure theory are of course σ-algebras of measurable sets and their quotient measure algebras. It is therefore natural to begin any investigation by concentrating on Dedekind σ-complete algebras. Nevertheless, in this section and the last (and in §326), I have gone to some trouble not to specialize to σ-complete algebras except when necessary. Partly this is just force of habit, but partly it is because I wish to lay a foundation for a further step forward: the investigation of the ways in which additive functionals on general Boolean algebras reflect the concepts of measure theory, and indeed can generate them. Some of the results in this direction can be surprising. I do not think it obvious that the condition (γ) in 362B(f-i), for instance, is sufficient in the absence of any hypothesis of Dedekind σ-completeness or countable additivity. Given a Boolean algebra A with the associated Riesz space M ∼ = S(A)∼ of bounded additive functionals on A, we now have a substantial list of bands in M : Mσ , Mτ , Mc (362Xf), and for a measure algebra the further bands Mac , Mtc and Mt ; for an algebra of sets we also have Mp (362Xg). These bands can be used to generate finite subalgebras of the band algebra of M (362Ya-362Yb), and for any such finite subalgebra we have a corresponding decomposition of M as a direct sum of the bands which are the atoms of the subalgebra (352Tb). This decomposition of M can be regarded as a recipe for decomposing its into finite sums of functionals with special properties. What I called the ‘Lebesgue decomposition’ in 232I

298

Function spaces

362 Notes

is just such a recipe. In that context I had a measure space (X, Σ, µ) and was looking at the countably additive functionals from Σ to R, that is, at Mσ in the language of this section, and the bands involved in the decomposition were Mp , Mac and Mtc . But I hope that it will be plain that these ideas can be refined indefinitely, as we refine the classification of additive functionals. At each stage, of course, the exact enumeration of the subalgebra of bands generated by the classification (as in 362Ya-362Yb) is a necessary check that we have understood the relationships between the classes we have described. These decompositions are of such importance that it is worth examining the corresponding band projections. I give formulae for the action of band projections on (non-negative) functionals in 362Bc, 362Bd, 362B(f-ii), 362Xb, 362Xc(iii), 362Xi(iv) and 362Yi. Of course these are readily adapted to give formulae for the projections onto the complementary bands, as in 362Bf and 362Xi. If we have an algebra of sets, the completely additive functionals are (usually) of relatively minor importance; in the standard examples, they correspond to functionals defined as weighted sums of point masses (362Xg(ii)). The point is that measure algebras A appear as quotients of σ-algebras Σ of sets by σ-ideals I; consequently the countably additive functionals on A correspond exactly to the countably additive functionals on Σ which are zero on I; but the canonical homomorphism from Σ to A is hardly ever order-continuous, so completely additive functionals on A rarely correspond to completely additive functionals on Σ. On the other hand, when we are looking at countably additive functionals on Σ, we have to consider the possibility that they are singular in the sense that they are carried on some member of I; in the measure algebra context this possibility disappears, and we can often be sure that every countably additive functional is absolutely continuous, as in 327Bb. For any Boolean algebra A, we can regard it as the algebra of open-and-closed subsets of its Stone space Z; the points of Z correspond to Boolean homomorphisms from A to {0, 1}, which are the fundamental ‘atoms’ in the space of additive functionals on A (362Xe, 362Xg(iv)). It is the case that all non-negative additive functionals on a Boolean algebra A can be represented by appropriate measures on its Stone space (see 416P in Volume 4), but I prefer to hold this result back until it can take its place among other theorems on representing functionals by measures and integrals. It is one of the leitmotivs of this chapter, that Boolean algebras and Riesz spaces are Siamese twins; again and again, matching results are proved by the application of identical ideas. A typical example is the pair 362B(f-i) and 362Yd. Many of us have been tempted to try to describe something which would provide a common generalization of Boolean algebras and Riesz spaces (and lattice-ordered groups). I have not yet seen any such structure which was worth the trouble. Most of the time, in this chapter, I shall be using ideas from the general theory of Riesz spaces to suggest and illuminate questions in measure theory; but if you pursue this subject you will surely find that intuitions often come to you first in the context of Boolean algebras, and the applications to Riesz spaces are secondary. In 362E I give a condition for uniform integrability in of dist sequences, following the pattern established in 246G and repeated in 354R and 356O. The condition of 362E assumes that the set is normbounded; but if you have 246G to hand, you will see that it can be done with weaker assumptions involving atoms, as in 362Yf-362Yg. I mention once again the Banach-Ulam problem: if A is Dedekind complete, can S(A)∼ c be different from S(A)× ? This is obviously equivalent to the form given in the notes to §326 above. See 363S below.

363 L∞ In this section I set out to describe an abstract construction for L∞ spaces on arbitrary Boolean algebras, corresponding to the L∞ (µ) spaces of §243. I begin with the definition of L∞ (A) (363A) and elementary facts concerning its own structure and the embedding S(A) ⊆ L∞ (A) (363B-363D). I give the basic universal mapping theorems which define the Banach lattice structure of L∞ (363E) and a description of the action of Boolean homomorphisms on L∞ spaces (363F-363G) before discussing the representation of L∞ (Σ) and L∞ (Σ/I) for σ-algebras Σ and ideals I of sets (363H). This leads at once to the identification of L∞ (µ), as defined in Volume 2, with L∞ (A), where A is the measure algebra of µ (363I). Like S(A), L∞ (A) determines the algebra A (363J). I briefly discuss the dual spaces of L∞ ; they correspond exactly to the duals of S described in §362 (363K). Linear functionals on L∞ can for some purposes be treated as ‘integrals’ (363L).

363E

L∞

299

In the second half of the section I present some of the theory of Dedekind complete and σ-complete algebras. First, L∞ (A) is Dedekind (σ-)complete iff A is (363M). The spaces L∞ (A), for Dedekind σcomplete A, are precisely the Dedekind σ-complete Riesz spaces with order unit (363N-363P). The spaces L∞ (A), for Dedekind complete A, are precisely the normed spaces which may be put in place of R in the Hahn-Banach theorem (363R). Finally, I mention some equivalent forms of the Banach-Ulam problem (363S). 363A Definition Let A be a Boolean algebra, with Stone space Z. I will write L∞ (A) for the space C(Z) = Cb (Z) of continuous real-valued functions from Z to R, endowed with the linear structure, order structure, norm and multiplication of C(Z) = Cb (Z). (Recall that because Z is compact (311I), {u(z) : z ∈ Z} is bounded for every u ∈ L∞ (A) = C(Z) (2A3N(b-iii)), that is, C(Z) = Cb (Z). Of course if A = {0}, so that Z = ∅, then C(Z) has just one member, the empty function.) 363B Theorem Let A be any Boolean algebra; write L∞ for L∞ (A). (a) L∞ is an M -space; its standard order unit is the constant function taking the value 1 at each point; in particular, it is a Banach lattice with a Fatou norm and the Levi property. (b) L∞ is a commutative Banach algebra and an f -algebra. (c) If u ∈ L∞ then u ≥ 0 iff there is a v ∈ L∞ such that u = v × v. proof (a) See 354Hb and 354J. (b)-(c) are obvious from the definitions of Banach algebra (2A4J) and f -algebra (352W) and the ordering of L∞ = C(Z). 363C Proposition Let A be any Boolean algebra. Then S(A) is a norm-dense, order-dense Riesz subspace of L∞ (A), closed under multiplication. proof Let Z be the Stone space of A. Using the definition of S = S(A) set out in 361D, it is obvious that S is a linear subspace of L∞ = L∞ (A) closed under multiplication. Because S, like L∞ , is a Riesz subspace of RX (361Ee), S is a Riesz subspace of L∞ . By the Stone-Weierstrass theorem (in either of the forms given in 281A and 281E), S is norm-dense in L∞ . Consequently it is order-dense (354I). 363D Proposition Let A be a Boolean algebra. If we regard χa ∈ S(A) (361D) as a member of L∞ (A) for each a ∈ A, then χ : A → L∞ (A) is additive, order-preserving, order-continuous and a lattice homomorphism. proof Because the embedding S = S(A) ⊆ L∞ (A) = L∞ is a Riesz homomorphism, χ : A → L∞ is additive and a lattice homomorphism (361F-361G). Because S is order-dense in L∞ (363C), the embedding S ⊆ L∞ is order-continuous (352Nb), so χ : A → L∞ is order-continuous (361Gb). 363E Theorem Let A be a Boolean algebra, and U a Banach space. Let ν : A → U be a bounded additive function. (a) There is a unique bounded linear operator T : L∞ (A) → U such that T χ = ν; in this case kT k = supa,b∈A kνa − νbk. (b) If U is a Banach lattice then T is positive iff ν is non-negative; and in this case T is order-continuous iff ν is order-continuous, and sequentially order-continuous iff ν is sequentially order-continuous. (c) If U is a Banach lattice then T is a Riesz homomorphism iff ν is a lattice homomorphism iff νa∧νb = 0 whenever a ∩ b = 0. proof Write S = S(A), L∞ = L∞ (A). (a) By 361I there is a unique bounded linear operator T0 : S → U such that T0 χ = ν, and kT0 k = sup{kνa − νbk : a, b ∈ A}. But because U is a Banach space and S is dense in L∞ , T0 has a unique extension to a bounded linear operator T : L∞ → U with the same norm (2A4I). (b)(i) If T is positive then T0 is positive so ν is non-negative, by 361Ga.

300

Function spaces

363E

(ii) If ν is non-negative then T0 is positive, by 361Ga in the other direction. But if u ∈ L∞+ and ² > 0, then by 354I there is a v ∈ S + such that ku − vk∞ ≤ ²; now kT u − T vk ≤ ²kT k. But T v = T0 v belongs to the positive cone U + of U . As ² is arbitrary, T u belongs to the closure of U + , which is U + (354Bc). As u is arbitrary, T is positive. (iii) Now suppose that ν is order-continuous as well as non-negative, and that A ⊆ L∞ is a non-empty downwards-directed set with infimum 0. Set B = {v : v ∈ S, there is some u ∈ A such that v ≥ u}. Then B is downwards-directed (indeed, v1 ∧ v2 ∈ B for every v1 , v2 ∈ B), and u = inf{v : v ∈ B, u ≤ v} for every u ∈ A (354I again), so B has the same lower bounds as A and inf B = 0 in L∞ and in S. But we know from 361Gb that T0 is order-continuous, while any lower bound for {T u : u ∈ A} in U must also be a lower bound for {T v : v ∈ B} = {T0 v : v ∈ B}, so inf u∈A T u = inf v∈B T0 v = 0 in U . As A is arbitrary, T is order-continuous (351Ga). (iv) Suppose next that ν is only sequentially order-continuous, and that hun in∈N is a non-increasing sequence in L∞ with infimum 0. For each n, k choose wnk ∈ S such that un ≤ wnk and kwnk − un k∞ ≤ 2−k (354I once more), and set wn0 = inf j,k≤n wjk for each n. Then hwn0 in∈N is a non-increasing sequence in S, and any lower bound of {wn0 : n ∈ N} is also a lower bound of {un : n ∈ N}, so 0 = inf n∈N wn0 in S and L∞ . Since T0 : S → U is sequentially order-continuous (361Gb), inf n∈N T un ≤ inf n∈N T wn0 = inf n∈N T0 wn0 = 0 in U . As hun in∈N is arbitrary, T is sequentially order-continuous. (v) On the other hand, if T is order-continuous or sequentially order-continuous, so is ν = T χ, because χ is order-continuous (363D). (c) We know that T0 : S → U is a Riesz homomorphism iff ν is a lattice homomorphism iff νa ∧ νb = 0 whenever a ∩ b = 0, by 361Gc. But T0 is a Riesz homomorphism iff T is. P P If T is a Riesz homomorphism so is T0 , because the embedding S ⊆ L∞ is a Riesz homomorphism. On the other hand, if T0 is a Riesz homomorphism, then the functions u 7→ u+ 7→ T (u+ ), u 7→ T u 7→ (T u)+ are continuous (by 354Bb) and agree on S, so agree on L∞ , and T is a Riesz homomorphism, by 352G. Q Q 363F Theorem Let A and B be Boolean algebras, and π : A → B a Boolean homomorphism. (a) There is an associated multiplicative Riesz homomorphism Tπ : L∞ (A) → L∞ (B), of norm at most 1, defined by saying that Tπ (χa) = χ(πa) for every a ∈ A. (b) For any u ∈ L∞ (A), there is a u0 ∈ L∞ (A) such that Tπ u = Tπ u0 and ku0 k∞ = kTπ uk∞ ≤ kuk∞ . (c)(i) The kernel of Tπ is the closed linear subspace of L∞ (A) generated by {χa : a ∈ A, πa = 0}. (ii) The set of values of Tπ is the closed linear subspace of L∞ (B) generated by {χ(πa) : a ∈ A}. (d) Tπ is surjective iff π is surjective, and in this case kvk∞ = min{kuk∞ : Tπ u = v} for every v ∈ L∞ (B). (e) Tπ is injective iff π is injective, and in this case kTπ uk∞ = kuk∞ for every u ∈ L∞ (A). (f) Tπ is order-continuous, or sequentially order-continuous, iff π is. (g) If C is another Boolean algebra and θ : B → C is another Boolean homomorphism, then Tθπ = Tθ Tπ : L∞ (A) → L∞ (C). proof Let Z and W be the Stone spaces of A and B. By 312P there is a continuous function φ : W → Z such that π ca = φ−1 [b a] for every a ∈ A, where b a is the open-and-closed subset of Z corresponding to a ∈ A. Write T for Tπ . (a) For u ∈ L∞ (A) = C(Z), set T u = uφ : W → R. Then T u ∈ C(W ) = L∞ (B). It is obvious, or at any rate very easy to check, that T : L∞ (A) → L∞ (B) is linear, multiplicative, a Riesz homomorphism and of norm 1 unless B = {0}, W = ∅. If a ∈ A, then T (χa) = (χa)φ = (χb a)φ = χ(φ−1 [b a]) = χ(πa), identifying χa ∈ L∞ (A) with the characteristic function χb a : Z → {0, 1} of the set b a. Of course Tπ = T is the only continuous linear operator with these properties, by 363Ea. (b) Set α = kT uk∞ , u0 (z) = max(−α, min(u(z), α)) for z ∈ Z; that is, u0 = (−αe) ∨ (u ∧ αe) in L∞ (A), where e is the standard order unit of L∞ (A). Then T e is the standard order unit of L∞ (B), so

L∞

363F

301

T u0 = (−αT e) ∨ (T u ∧ αT e) = T u, while ku0 k∞ ≤ α = kT uk∞ = kT u0 k∞ ≤ ku0 k∞ ≤ kuk∞ . (c)(i) Let U be the closed linear subspace of L∞ (A) generated by {χa : πa = 0}, and U0 the kernel of T . Because T is continuous and linear, U0 is a closed linear subspace, and T (χa) = χ0 = 0 whenever πa = 0; so U ⊆ U0 . Now take any u ∈ U0 and ² > 0. Then T (u+ ) = (T u)+ = 0, so u+ ∈ U0 . By 354I there is a 0 + + 0 0 + 0 u0 ∈ S(A) Pn such that 0 ≤ u ≤ u and ku − u k∞ ≤ ². Now 0 ≤ T u ≤ T u = 0, so T u = 0. Express 0 u as i=0 αi χai where αi ≥ 0 for each i. For each i, αi χ(πai ) = T (αi χai ) = 0, so πai = 0 or αi = 0; in either case αi χai ∈ U . Consequently u0 ∈ U . As ² is arbitrary and U is closed, u+ ∈ U . Similarly, u− = (−u)+ ∈ U and u = u+ − u− ∈ U . As u is arbitrary, U0 ⊆ U and U0 = U . (ii) Let V be the closed linear subspace of L∞ (B) generated by {χ(πa) : a ∈ A}, and V0 = T [L∞ (A)]. Then T [S(A)] ⊆ V , so V0 = T [S(A)] ⊆ T [S(A)] ⊆ V = V . On the other hand, V0 is a closed linear subspace in L∞ (B). P P It is a linear subspace because T is a linear operator. To see that it is closed, take any v ∈ V 0 . Then there is a sequence hvn in∈N in V0 such that kv − vn k∞ ≤ 2−n for every n ∈ N. Choose un ∈ L∞ (A) such that T u0 = v0 , while T un = vn − vn−1 and kun k∞ = kvn − vn−1 k∞ for n ≥ 1 (using (b) above). Then P∞ P∞ n=1 kv − vn k∞ + kv − vn−1 k∞ n=1 kun k∞ ≤ Pn is finite, so u = limn→∞ i=0 ui is defined in the Banach space L∞ (A), and Pn T u = limn→∞ i=0 T ui = limn→∞ vn = v. As v is arbitrary, V0 is closed. Q Q Since χ(πa) = T (χa) ∈ V0 for every a ∈ A, V ⊆ V0 and V = V0 , as required. (d) If π is surjective, then T is surjective, by (c-ii). If T is surjective and b ∈ B, then there is a u ∈ L∞ (A) such that T u = χb. Now there is a u0 ∈ S(A) such that ku − u0 k∞ ≤ 31 , so that kT u0 − χbk∞ ≤ 31 . Taking a ∈ A such that {z : u0 (z) ≥ 21 } = b a, we must have πa = b, since bb = {w : (T u0 )(w) ≥ 1 } = φ−1 [b a] = π ca. 2 As b is arbitrary, π is surjective. Now (b) tells us that in this case kvk∞ = min{kuk∞ : T u = v} for every v ∈ L∞ (B). (e) By (c-i), T is injective iff π is injective. In this case, for any u ∈ L∞ (A), kT uk∞ = kT |u|k∞ (because T is a Riesz homomorphism) ≥ sup{kT u0 k∞ : u0 ∈ S(A), u0 ≤ |u|} = sup{ku0 k∞ : u0 ∈ S(A), u0 ≤ |u|} (by 361Jd) = kuk∞ (by 354I) ≥ kT uk∞ , and kT uk∞ = kuk∞ . (f ) If T is (sequentially) order-continuous then π = T χ is (sequentially) order-continuous, by 363D. If π is (sequentially) order-continuous then χπ : A → L∞ (B) is (sequentially) order-continuous, so T is (sequentially) order-continuous, by 363Eb. (g) This is elementary, in view of the uniqueness of Tθπ .

302

Function spaces

363G

363G Corollary Let A be a Boolean algebra. (a) If C is a subalgebra of A, then L∞ (C) can be identified, as Banach lattice and as Banach algebra, with the closed linear subspace of L∞ (A) generated by {χc : c ∈ C}. (b) If I is an ideal of A, then L∞ (A/I) can be identified, as Banach lattice and as Banach algebra, with the quotient space L∞ (A)/V , where V is the closed linear subspace of L∞ (A) generated by {χa : a ∈ I}. proof Apply 363Fc to the identity map from C to A and the canonical map from A onto A/I. 363H Representations of L∞ (A) Much of the importance of the concept of L∞ (A) arises from the way it is naturally represented in the contexts in which the most familiar Boolean algebras appear. Proposition Let X be a set and Σ a σ-algebra of subsets of X. (a) L∞ (Σ) can be identified, as Banach algebra and Banach lattice, with the space L∞ of bounded Σmeasurable real-valued functions on X, with the norm kf k∞ = supx∈X |f (x)| for f ∈ L∞ ; this identification matches χE ∈ L∞ (Σ) with the characteristic function of E as a subset of X, for every E ∈ Σ. In particular, for any set X, L∞ (PX) can be identified with `∞ (X). (b) If A is a Dedekind σ-complete Boolean algebra and π : Σ → A is a surjective sequentially ordercontinuous Boolean homomorphism with kernel I, then L∞ (A) can be identified, as Banach algebra and Banach lattice, with L∞ /W, where W = {f : f ∈ L∞ , {x : f (x) 6= 0} ∈ I} is a closed ideal and solid linear subspace of L∞ . For f ∈ L∞ , kf • k∞ = min{α : α ≥ 0, {x : |f (x)| > α} ∈ I}. (c) In particular, if I is any σ-ideal of Σ and E 7→ E • is the canonical homomorphism from Σ onto A = Σ/I, then we have an identification of L∞ (A) with a quotient of L∞ , and for any E ∈ Σ we can identify χ(E • ) ∈ L∞ (A) with the equivalence class (χE)• ∈ L∞ /W of the characteristic function χE. proof (a) For the elementary properties of the space of Σ-measurable functions, see §121. In particular, it is easy to check that L∞ is a Riesz space, with a Riesz norm, and a normed algebra. To check that it is a Banach space, observe that if hfn in∈N is a Cauchy sequence in L∞ , then |fm (x) − fn (x)| ≤ kfm − fn k∞ → 0 as m, n → ∞, so f (x) = limn→∞ fn (x) is defined for every x ∈ X; now f is Σ-measurable, by 121Fa. Of course kf k∞ ≤ supn∈N kfn k∞ < ∞, so f ∈ L∞ , while kf − fn k∞ ≤ supm≥n kfm − fn k∞ → 0 as n → ∞, so f = limn→∞ fn in L∞ . As hfn in∈N is arbitrary, L∞ is complete. By 361L we can identify S(Σ), as Riesz space and normed algebra, with the linear span S of {χE : E ∈ Σ}, which is a subspace of L∞ . Now the point is that it is dense. P P If f ∈ L∞ and ² > P0, then for each n ∈ Z set En = {x : n² ≤ f (x) < (n + 1)²} ∈ Σ; then En = ∅ if |n| > 1 + 1² kf k∞ , so g = n∈Z n²χEn belongs to S, and of course kf − gk∞ ≤ ². Q Q Consequently the canonical normed space isomorphism between S(Σ) and S extends (uniquely) to a normed space isomorphism between L∞ (Σ) and L∞ (use 2A4I). Because the operations ∨ and × are continuous in both L∞ (Σ) and L∞ , and their actions on S(Σ) and S are identified by our isomorphism, the isomorphism between L∞ (Σ) and L∞ identifies their lattice and algebra structures. (b)(i) By 363F, we have a multiplicative Riesz homomorphism T = Tπ from L∞ (Σ) to L∞ (A) which is surjective (363Fd) and has kernel the closed linear subspace W of L∞ (Σ) generated by {χE : E ∈ I}. Now under the isomorphism described in (a), W corresponds to W. P P (α) W is a linear subspace of L∞ because {x : (f + g)(x) 6= 0} ⊆ {x : f (x) 6= 0} ∪ {x : g(x) 6= 0} ∈ I, {x : (αf )(x) 6= 0} ⊆ {x : f (x) 6= 0} ∈ I whenever f , g ∈ W and α ∈ R. (β) If hfn in∈N is a sequence in W converging to f ∈ L∞ , then S {x : f (x) 6= 0} ⊆ n∈N {x : fn (x) 6= 0} ∈ I, so f ∈ W. Thus W is a closed linear subspace of L∞ . (γ) If E ∈ I, then χE, taken in S(Σ) or L∞ (Σ), corresponds to the function χE : X → {0, 1}, which belongs to W; so that W must correspond to the closed linear span in L∞ of such characteristic functions, which is a subspace of W. (δ) On the other hand, if f ∈ W and ² > 0, set

L∞

363K

303

En = {x : n² ≤ f (x) ≤ (n + 1)²}, En0 = {x : −(n + 1)² ≤ f (x) ≤ −n²} P∞ for n ≥ 1; all these belong to I, so g = n=1 n(χEn − χEn0 ) ∈ W corresponds to a member of W , while kf − gk∞ ≤ ². As W is closed, f also must correspond to some member of W . As f is arbitrary, W and W match exactly. Q Q (ii) Because T is a multiplicative Riesz homomorphism, L∞ (A) ∼ = L∞ (Σ)/W is matched canonically, ∞ in its linear, order and multiplicative structures, with L /W. We know also that kvk∞ = inf{kuk∞ : u ∈ L∞ (Σ), T u = v} for every v ∈ L∞ (A) (363Fd), that is, that the norm of L∞ (A) corresponds to the quotient norm on L∞ (Σ)/W . As for the given formula for the norm, take any f ∈ L∞ . There is a g ∈ L∞ such that T f = T g and kT f k∞ = kgk∞ . (Here I am treating T as an operator from L∞ onto L∞ (A).) In this case {x : |f (x)| > kT f k∞ } ⊆ {x : f (x) 6= g(x)} ∈ I. On the other hand, if {x : |f (x)| > α} ∈ I, and we set h = −α1 ∨ (f ∧ 1), then T h = T f , so kT f k∞ ≤ khk∞ ≤ α. (c) Put (a) and (b) together. 363I Corollary Let (X, Σ, µ) be a measure space, with measure algebra A. Then L∞ (µ) can be identified, as Banach lattice and Banach algebra, with L∞ (A); the identification matches (χE)• ∈ L∞ (µ) with χ(E • ) ∈ L∞ (A), for every E ∈ Σ. Remark The space I called L∞ (µ) in Chapter 24 is not strictly speaking the space L∞ ∼ = L∞ (Σ) of 363H; ∞ 0 I took L (µ) ⊆ L (µ) to be the set of essentially bounded, virtually measurable functions defined almost everywhere on X, and in general this is larger. But, as remarked in the notes to §243, L∞ (µ) can equally ∞ above, because every function well be regarded as a quotient of what I there called L∞ strict , which is the L ∞ ∞ in L (µ) is equal almost everywhere to some member of Lstrict . 363J Recovering the algebra A: Proposition Let A be a Boolean algebra. For a ∈ A write Va for the solid linear subspace of L∞ (A) generated by χa. Then a 7→ Va is a Boolean isomorphism between A and the algebra of projection bands in L∞ (A). proof The proof is nearly identical to that of 361K. If a ∈ A, u ∈ Va and v ∈ V1\a , then |u| ∧ |v| = 0 because χa ∧ χ(1 \ a) = 0; and if w ∈ L∞ (A) then w = (w × χa) + (w × χ(1 \ a)) ∈ Va + V1\a because |w×χa| ≤ kwk∞ χa and |w×χ(1 \ a)| ≤ kwk∞ χ(1 \ a). So Va and V1\a are complementary projection bands in L∞ = L∞ (A). Next, if U ⊆ L∞ is a projection band, then χ1 is expressible as u + v where u ∈ U , v ∈ U ⊥ ; thinking of L∞ as the space of continuous real-valued functions on the Stone space Z of A, u and v must be the characteristic functions of complementary subsets E, F of Z, which must be open-and-closed, so that E = b a, F = 1d \ a. In this case Va ⊆ U and V1\a ⊆ U ⊥ , so U must be Va precisely. Thus a 7→ Va is surjective. Finally, just as in 361K, a ⊆ b ⇐⇒ Va ⊆ Vb , so we have a Boolean isomorphism. 363K Dual spaces of L∞ The questions treated in §362 yield nothing new in the present context. I spell out the details. Proposition Let A be a Boolean algebra. Let M , Mσ and Mτ be the L-spaces of bounded finitely additive functionals, bounded countably additive functionals and completely additive functionals on A. Then the embedding S(A) ⊆ L∞ (A) induces Riesz space isomorphisms between S(A)∼ ∼ = M and L∞ (A)∼ = L∞ (A)∗ , ∼ ∼ ∞ ∼ × ∼ ∞ × S(A)c = Mσ and L (A)c , and S(A) = Mτ and L (A) . proof Write S = S(A), L∞ = L∞ (A). (a) For the identifications S ∼ ∼ = M , Sc∼ ∼ = Mσ and S × ∼ = Mτ see 362A.

304

Function spaces

363K

(b) L∞∗ = L∞∼ either because L∞ is a Banach lattice (356Dc) or because L∞ has an order-unit norm, so that a linear functional on L∞ is order-bounded iff it is bounded on the unit ball. (c) If f is a positive linear functional on L∞ , then f ¹S is a positive linear functional. Because S is orderdense in L∞ (363C), the embedding is order-continuous (352Nb); so if f is (sequentially) order-continuous, ∼ so is f ¹S. Accordingly the restriction operator f 7→ f ¹S gives maps from L∞∼ to S ∼ , (L∞ )∼ c to Sc and ∞× × ∞∼ + ∞ L to S . If f ∈ L and f ¹S ≥ 0, then f (u ) ≥ 0 for every u ∈ S and therefore for every u ∈ L , and f ≥ 0; so all these restriction maps are injective positive linear operators. (d) I need to show that they are surjective. (i) If g ∈ S ∼ , then g is bounded on the unit ball {u : u ∈ S, kuk∞ ≤ 1}, so has an extension to a continuous linear f : L∞ → R (2A4I); thus S ∼ = {f ¹S : f ∈ L∞∼ }. This means that f 7→ f ¹S is actually a Riesz space isomorphism between L∞∼ and S ∼ . In particular, |f |¹S = |f ¹S| for any f ∈ L∞∼ . (ii) If f : L∞ → R is a positive linear operator and f ¹S ∈ Sc∼ , let hun in∈N be a non-increasing sequence in L with infimum 0. For each n, k ∈ N there is a vnk ∈ S such that un ≤ vnk ≤ un + 2−k e, where e is the standard order unit of L∞ (354I, as usual); set wn = inf i,k≤n vik ; then hwn in∈N is a non-increasing sequence in S with infimum 0, so ∞

0 ≤ inf n∈N f (un ) ≤ inf n∈N f (wn ) = 0. As hun in∈N is arbitrary, f ∈

(L∞ )∼ c .

Consequently, for general f ∈ L∞∼ ,

∞ ∼ ∼ ∼ f ∈ (L∞ )∼ c ⇐⇒ |f | ∈ (L )c ⇐⇒ |f ¹S| ∈ Sc ⇐⇒ f ¹S ∈ Sc , ∼ and the map f 7→ f ¹S : (L∞ )∼ c → Sc is a Riesz space isomorphism.

(iii) Similarly, if f ∈ L∞∼ is non-negative and f ¹S ∈ S × , then whenever A ⊆ L∞ is non-empty, downwards-directed and has infimum 0, B = {w : w ∈ S, ∃ u ∈ A, w ≥ u} has infimum 0, so inf u∈A f (u) ≤ inf w∈B f (w) ≤ 0 and f ∈ L∞× . As in (ii), it follows that f 7→ f ¹S is a surjection from L∞× onto S × . *363L Integration with respect to a finitely additive functional (a) If A is a Boolean algebra and ν : A → R is a bounded additive functional, then by 363K we have a corresponding functional fν ∈ L∞ (A)∗ defined by saying that fν (χa) R = νa for every a ∈ A. There are contexts in which it is convenient, and even helpful, to use the formula u dν in place of fν (u) for u ∈ L∞ = L∞ (A). When doing so, we must of course that we may have lost some of the standard properties of ‘integration’. But enough of our intuitions (including, for instance, the idea of stochastic independence) remain valid to make the formula a guide to interesting ideas. (b) LetR M be the L-space of bounded finitely additive functionals on A (362B). Then we have a function (u, ν) 7→ u dν : L∞ × M → R. Now this map is bilinear. P P For µ, ν ∈ M , u, v ∈ L∞ and α ∈ R, R R R R R u + v dν = u dν + v dν, αu dν = α u dν just because fν is linear. On the other side, we have (fµ + fν )(χa) = fµ (χa) + fν (χa) = µa + νa = (µ + ν)(a) = fµ+ν (χa) ∞ Rfor every a ∈ A,R so thatRfµ + fν and fµ+νR must agree on R S(A) and therefore on L . But this means that u d(µ + ν) = u dµ + u dν. Similarly, u d(αµ) = α u dµ. Q Q R (c) If ν is non-negative, we have u dν ≥ 0 whenever u ≥ 0, as in part (c) of the proof of 363K. Consequently, for any ν ∈ M and u ∈ L∞ , Z Z Z Z Z | u dν| = | u+ dν + − u− dν + − u+ dν − + u− dν − | Z Z Z Z + + − + + − ≤ u dν + u dν + u dν + u− dν − Z Z = |u|d|ν| ≤ kuk∞ χ1 d|ν| = kuk∞ |ν|(1) = kuk∞ kνk.

L∞

363Lf

305

R So (u, ν) 7→ u dν has norm (as defined in 253Ab) at most 1. If A 6= 0, the norm is exactly 1. (For this we need to know that there is a ν ∈ M + such that ν1 = 1. Take any z in the Stone space of A and set νa = 1 if z ∈ b a, 0 otherwise.) (d) We do not have any result corresponding to B.Levi’s theorem in this language, because (even if ν is non-negative andR countably additive) there is no reason to suppose that supn∈N un is defined in L∞ just because supn∈N un dν is finite. But if ν is countably additive and A is Dedekind σ-complete, we have something corresponding to Lebesgue’s Dominated Convergence Theorem (363Yh). (e) One formula which we can imitate in the present context is that of 252O, where the ordinary integral is represented in the form

R

f dµ =

R∞ 0

µ{x : f (x) ≥ t}dt.

In the context of general Boolean algebras, we cannot directly represent the set [[f ≥ t]] = {x : f (x) ≥ t} (though in the next section I will show that in Dedekind σ-complete Boolean algebras there is an effective expression of this idea, and I will use it in the principal definition of §365). But what we can say is the following. If A is any Boolean algebra, and ν : A → [0, ∞[ is a non-negative additive functional, and u ∈ L∞ (A)+ , then

R

u dν =

R∞ 0

sup{νa : tχa ≤ u}dt,

where the right-hand integral is taken with respect to Lebesgue measure. RP P (i) For t ≥ 0 set h(t) = ∞ sup{νa : tχa ≤ u}. Then h is non-increasing and zero for t > kuk∞ , so 0 h(t)dt is defined in R. If we set hn (t) = h(2−n (k + 1)) whenever k, n ∈ N and 2−n k ≤ t < 2−n (k + 1), then hhn (t)in∈N is a nondecreasing sequence which h is continuous at t, which is almost everywhere R ∞converges to h(t) Rwhenever ∞ (222A, or otherwise); so 0 h(t)dt = limn→∞ 0 hn (t)dt. Next, given n ∈ N and ² > 0, we can choose for each k ≤ k ∗ = b2n kuk∞ c an ak such that 2−n (k + 1)χak ≤ u and νak ≥ h(2−n (k + 1)) − ². In this case Pk∗ −n χak ≤ u, so k=0 2 Z

∗

∞

hn (t)dt = 2

−n

0

k X

∗

h(2

−n

(k + 1)) ≤ kuk∞ ² + 2

−n

k=0

= kuk∞ ² +

k X

νak

k=0

Z X k∗

Z 2−n χak dν ≤ kuk∞ ² +

u dν.

k=0

R∞ R As n and ² are arbitrary, 0 h(t)dt ≤ u dν. (ii) In the other direction, there is for any ² > 0 a v ∈ S(A) Pm such that v ≤ u ≤ v + ²χ1. If we express v as j=0 γj χcj where c0 ⊇ . . . ⊇ cm and γj ≥ 0 for every j Pk (361Ec), then we shall have h(t) ≥ νck whenever t ≤ j=0 γj , so

R∞

As ² is arbitrary,

R∞ 0

0

h(t)dt ≥

R

h(t)dt ≥

Pm

k=0

γk νck =

R

v dν ≥

R

u dν − ²ν1.

u dν and the two ‘integrals’ are equal. Q Q

R (f ) The formula dν is especially natural when A is an algebra of sets, so that L∞ can be directly interpreted as a space of functions (363Yf); better still, when A is actually a σ-algebra of subsets of a set X, L∞ can be identified R with the space of bounded A-measurable functions on X, as in 363Ha. So in such contexts I may write g dν when g : X → R is bounded and A-measurable, and ν : A → R is an additive functional. But I will try to take care to signal any such deviation from the normal principle that the R symbol refers to the sequentially order-continuous integral defined in §122 with the minor modifications introduced in §§133 and 135. My purpose in this paragraph has been only to indicate something of what can be done with finite additivity alone.

306

363M the last.

Function spaces

363M

Now I come to a fundamental fact underlying a number of theorems in both this volume and

Theorem Let A be a Boolean algebra. (a) A is Dedekind σ-complete iff L∞ (A) is Dedekind σ-complete. (b) A is Dedekind complete iff L∞ (A) is Dedekind complete. proof (a)(i) Suppose that A is Dedekind σ-complete. By 314M, we may identify A with a quotient Σ/M, where M is the ideal of meager subsets of the Stone space Z of A, and Σ = {E4A : E ∈ E, A ∈ M}, writing E = {b a : a ∈ A} for the algebra of open-and-closed subsets of Z. By 363H, L∞ = L∞ (A) can be identified with L∞ /V, where L∞ is the space of bounded Σ-measurable functions from Z to R, and V is the space of functions zero except on a member of I. Now suppose that hun in∈N is a sequence in L∞ with an upper bound u ∈ L∞ . Express un , u as fn• , f • where fn , f ∈ L∞ . Set g(z) = supn∈N min(fn (z), f (z)) for every z ∈ Z; then g ∈ L∞ (121F), so we have a corresponding member v = g • of L∞ . For each n ∈ N, u ≥ un so (fn − f )+ ∈ V, {z : fn (z) > g(z)} ⊆ {z : fn (z) > f (z)} ∈ M and v ≥ un . If w ∈ L for every n, so

∞

and w ≥ un for every n, then express w as h• where h ∈ L∞ ; we have (fn − h)+ ∈ V {z : g(z) > h(z)} ⊆

S

n∈N {z

: fn (z) > h(z)} ∈ M

because M is a σ-ideal, and (g − h)+ ∈ V, w ≥ v. Thus v = supn∈N un in L∞ . As hun in∈N is arbitrary, L∞ is Dedekind σ-complete (using 353G). (ii) Now suppose that L∞ is Dedekind σ-complete, and that A is a countable non-empty set in A. In this case {χa : a ∈ A} has a least upper bound u in L∞ . Take a v ∈ S(A) such that 0 ≤ v ≤ u and ku − vk∞ ≤ 13 ; set b = [[v > 13 ]], as defined in 361Eg. If a ∈ A, then k(χa − v)+ k∞ ≤ ku − vk∞ ≤ 31 , so 2 3 χa ≤ v and a ⊆ b. If c ∈ A is any upper bound for A, then v ≤ u ≤ χc so b ⊆ c. Thus b = sup A in A. As A is arbitrary, A is Dedekind σ-complete. (b)(i) For the second half of this theorem I use an argument which depends on ing the representation described in (a-i) above with the original definition of L∞ in 363A. The point is that C(Z) ⊆ L∞ , and for any f ∈ C(Z) = L∞ (A) its equivalence class f • in L∞ /V corresponds to f itself. P P Perhaps it will help to give a name T to the canonical isomorphism from L∞ /V to L∞ . Then V = {f : T f • = f } is a closed linear subspace of C(Z), because f 7→ f • and T are continuous linear operators. But if a ∈ A, then (b a)• , the • ∞ equivalence class of b a ∈ Σ in Σ/M, corresponds to a (see the proof of 314M), so (χb a) ∈ L /V corresponds to χa; that is, T (χb a)• = χb a, if we identify χa ∈ L∞ with χb a : Z → {0, 1}. So V contains χb a for every a ∈ A; because V is a linear subspace, S(A) ⊆ V ; because V is closed, L∞ ⊆ V . Q Q For a general f ∈ L∞ , g = T f • must be the unique member of C(Z) such that g • = f • , that is, such that {z : g(z) 6= f (z)} is meager. (ii) Suppose now that A is actually Dedekind complete. In this case Z is extremally disconnected (314S). Consequently every open set belongs to Σ. P P If G is open, then G is open-and-closed; but A = G \ G is a closed set with empty interior, so is meager, and G = G4A ∈ Σ. Q Q Let A ⊆ L∞ = C(Z) be any non-empty set with an upper bound in C(Z). For each z ∈ Z set g(z) = supu∈A u(z). Then S Gα = {z : g(z) > α} = u∈A {z : u(z) > α} is open for every α ∈ R (that is, g is lower semi-continuous). Thus Gα ∈ Σ for every α, so g ∈ L∞ , and v = T g • is defined in C(Z). For any u ∈ A, g ≥ u in L∞ , so v = T g • ≥ T u• = u in L∞ ; thus v is an upper bound for A in L∞ . On the other hand, if w is any upper bound for A in L∞ = C(Z), then surely w(z) ≥ u(z) for every z ∈ Z, u ∈ A, so w ≥ g and w = T w• ≥ T g • = v. This means that v is the least upper bound of A. As A is arbitrary, L∞ is Dedekind complete.

363Q

L∞

307

(iii) Finally, if L∞ is Dedekind complete, then the argument of (b-ii), applied to arbitrary non-empty subsets A of A, shows that A is also Dedekind complete. 363N

Much of the importance of L∞ spaces in the theory of Riesz spaces arises from the next result.

Proposition Let U be a Dedekind σ-complete Riesz space with an order unit. Then U is isomorphic, as Riesz space, to L∞ (A), where A is the algebra of projection bands in U . proof (a) By 353M, U is isomorphic to a norm-dense Riesz subspace of C(X) for some compact Hausdorff space X; for the rest of this argument, therefore, we may suppose that U actually is such a subspace. (b) Now U = C(X). P P If g ∈ C(X) then by 354I there are sequences hfn in∈N , hfn0 in∈N in U such that fn ≤ g ≤ gn and kgn − fn k∞ ≤ 2−n for every n. Now {fn : n ∈ N} has a least upper bound f in U ; since we must have fn ≤ f ≤ gn for every n, f = g and g ∈ U . Q Q (c) Next, X is zero-dimensional. P P Suppose that G ⊆ X is open and x ∈ G. Then there is an open set G1 such that x ∈ G1 ⊆ G1 ⊆ G (3A3Bc). There is an f ∈ C(X) such that 0 ≤ f ≤ χG1 and f (x) > 0 (also by 3A3Bc); write H for {y : f (y) > 0}. Set g = supn∈N (nf ∧ χX), the supremum being taken in U = C(X). For each y ∈ H, we must have g(y) ≥ min(1, nf (y)) for every n, so that g(y) = 1. On the other hand, if y ∈ X \ H, there is an h ∈ C(X) such that h(y) > 0 and 0 ≤ h ≤ χ(X \ H); now h ∧ f = 0 so h ∧ g = 0 and g(y) = 0. Thus χH ≤ g ≤ χH. The set {y : g(y) ∈ {0, 1}} is closed and includes H ∪ (X \ H) so must be the whole of X; thus G2 = {y : g(y) > 12 } = {y : g(y) ≥ 12 } is open-and-closed, and we have x ∈ H ⊆ G2 ⊆ H ⊆ G1 ⊆ G. As x, G are arbitrary, the set of open-and-closed subsets of X is a base for the topology of X, and X is zero-dimensional. Q Q (d) We can therefore identify X with the Stone space of its algebra E of open-and-closed sets (311J). But in this case 363A immediately identifies U = C(X) with L∞ (E). By 363J, E is isomorphic to A, so U∼ = L∞ (A). Remark Note that in part (c) of the argument above, we have to take great care over the interpretation of ‘sup’. In the space of all real-valued functions on X, the supremum of {nf ∧ χX : n ∈ N} is just χH. But g is supposed to be the least continuous function greater than or equal to nf ∧ χX for every n, and is therefore likely to be strictly greater than χH, even though sandwiched between χH and χH. 363O Corollary Let U be a Dedekind σ-complete M -space. Then U is isomorphic, as Banach lattice, to L∞ (A), where A is the algebra of projection bands of U . proof This is merely the special case of 363N in which U is known from the start to be complete under an order-unit norm. 363P Corollary Let U be any Dedekind σ-complete Riesz space and e ∈ U + . Then the solid linear subspace Ue of U generated by e is isomorphic, as Riesz space, to L∞ (A) for some Dedekind σ-complete Boolean algebra A; and if U is Dedekind complete, so is A. proof Because U is Dedekind σ-complete, so is Ue (353J(a-i)). Apply 363N to Ue to see that Ue ∼ = L∞ (A) for some A. Because Ue is Dedekind σ-complete, so is A, by 363Ma; while if U is Dedekind complete, so are Ue and A, by 353J(b-i) and 363Mb. 363Q The next theorem will be a striking characterization of the Dedekind complete L∞ spaces as normed spaces. As a warming-up exercise I give a much simpler result concerning their nature as Banach lattices. Proposition Let A be a Dedekind complete Boolean algebra. Then for any Banach lattice U , a linear operator T : U → L∞ = L∞ (A) is continuous iff it is order-bounded, and in this case kT k = k|T |k, where the modulus |T | is taken in L∼ (U ; L∞ ).

308

Function spaces

363Q

proof It is generally true that order-bounded operators between Banach lattices are continuous (355C). If T : U → L∞ is continuous, then for any w ∈ U + |u| ≤ w =⇒ kuk ≤ kwk =⇒ kT uk∞ ≤ kT kkwk =⇒ |T u| ≤ kT kkwkχ1. So T is order-bounded. As L∞ is Dedekind complete (363Mb), |T | is defined in L∼ (U ; L∞ ) (355Ea). For any w ∈ U , |T ||w| = sup{|T u| : |u| ≤ |w|} ≤ kT kkwkχ1, so k|T |(w)k ≤ kT kkwk; accordingly k|T |k ≤ kT k. On the other hand, of course, |T w| ≤ |T ||w| ≤ k|T |kkwkχ1 for every w ∈ U , so kT k ≤ k|T |k and the two norms are equal. Remark Of course what is happening here is that the spaces L∞ (A), for Dedekind complete A, are just the Dedekind complete M -spaces; this is an elementary consequence of 363N and 363M. 363R

Now for something much deeper.

Theorem Let U be a normed space over R. Then the following are equiveridical: (i) there is a Dedekind complete Boolean algebra A such that U is isomorphic, as normed space, to L∞ (A); (ii) whenever V is a normed space, V0 a linear subspace of V , and T0 : V0 → U is a bounded linear operator, there is an extension of T0 to a bounded linear operator T : V → U with kT k = kT0 k. proof For the purposes of the argument below, let us say that a normed space U satisfying the condition (ii) has the ‘Hahn-Banach property’. Part A: (i)⇒(ii) I have to show that L∞ (A) has the Hahn-Banach property for every Dedekind complete Boolean algebra A. Let V be a normed space, V0 a linear subspace of V , and T0 : V0 → L∞ = L∞ (A) a bounded linear operator. Set γ = kT0 k. Let P be the set of all functions T such that dom T is a linear subspace of V including V0 and T : dom T → U is a bounded linear operator extending T0 and with norm at most γ. Order P by saying that T1 ≤ T2 if T2 extends T1 . Then any non-empty totally ordered subset Q of P has an upper bound in P. P P S Set dom T = {dom T1 : T1 ∈ Q}, T v = T1 v whenever T1 ∈ Q and v ∈ dom T1 ; it is elementary to check that T ∈ P, so that T is an upper bound for Q in P. Q Q By Zorn’s Lemma, P has a maximal element T˜. Now dom T˜ = V . P P?? Suppose, if possible, otherwise. Write V˜ = dom T˜ and take any v˜ ∈ V \ V˜ ; let V1 be the linear span of V˜ ∪{˜ v }, that is, {v+α˜ v : v ∈ V˜ , α ∈ R}. ∞ ˜ If v1 , v2 ∈ V then, writing e for the standard order unit of L , T˜v1 + T˜v2 = T˜(v1 + v2 ) ≤ kT˜(v1 + v2 )k∞ e ≤ γkv1 + v2 ke ≤ γkv1 − v˜ke + γkv2 + v˜ke, so T˜v1 − γkv1 − v˜ke ≤ γkv2 + v˜ke − T˜v2 . Because L∞ is Dedekind complete (363Mb), u ˜ = supv1 ∈V˜ T˜v1 − γkv1 − v˜ke is defined in L∞ and u ˜ ≤ γkv2 + v˜ke − T v2 for every v2 ∈ V˜ . Putting these together, we have T˜v − u ˜ ≤ γkv − v˜ke,

T˜v + u ˜ ≤ γkv + v˜ke

for all v ∈ V˜ . Consequently, if v ∈ V˜ , then for α > 0 T˜v + α˜ u = α(T˜( 1 v) + u ˜) ≤ αγk 1 v + v˜ke = γkv + α˜ v ke, α

α

while for α < 0 T˜v + α˜ u = |α|(T˜(− α1 v) − u ˜) ≤ |α|γk − α1 v − v˜ke = γkv + α˜ v ke, and of course

363R

L∞

309

T˜v ≤ kT˜vk∞ e ≤ γkvke. So we have T˜v + α˜ u ≤ γkv + α˜ v ke for every v ∈ V˜ , α ∈ R. Define T1 : V1 → L∞ by setting T1 (v + α˜ v ) = T˜v + α˜ u for every v ∈ V˜ , α ∈ R. (This is well-defined ˜ because v˜ ∈ / V , so any member of V1 is uniquely expressible as v + α˜ v where v ∈ V˜ and α ∈ R.) Then T1 ∞ is a linear operator, extending T0 , from a linear subspace of V to L . But from the calculations above we know that T1 v ≤ γkvke for every v ∈ V1 ; since we also have T1 v = −T1 (−v) ≥ −γk − vke = −γkvke, kT1 vk∞ ≤ γkvk for every v ∈ V1 , and T1 ∈ P. But now T1 is a member of P properly extending T˜, which is supposed to be impossible. X XQ Q Accordingly T˜ : V → L∞ is an extension of T0 to the whole of V , with the same norm as T0 . As V and T0 are arbitrary, L∞ has the Hahn-Banach property. Part B: (ii)⇒(i) Now let U be a normed space with the Hahn-Banach property. If U = {0} then of course it is isomorphic to L∞ (A), where A = {0}, so henceforth I will take it for granted that U 6= {0}. (a) Let Z be the unit ball of the dual U ∗ of U , with the weak* topology. Then Z is a compact Hausdorff space (3A5F). For u ∈ U set Zu = {f : f ∈ Z, |f (u)| = kuk}; then Zu is a closed subset of Z (because f 7→ f (u) is continuous), and is non-empty, by the Hahn-Banach theorem (3A5Ab, or Part A above!) Now let P be the set of those closed sets X ⊆ Z such that X ∩ Zu 6= ∅ for every u ∈ U . If Q ⊆ P is non-empty T and totally ordered, then Q ∈ P, because for any u ∈ U {X ∩ Zu : X ∈ Q} is a downwards-directed family of non-empty compact sets, so must have non-empty intersection. By Zorn’s Lemma, upside down, P has a minimal element X; with its relative topology, X is a compact Hausdorff space. (b) We have a linear operator R : U → C(X) given by setting (Ru)(x) = x(u) for every u ∈ U , x ∈ X; because X ⊆ Z, kRuk∞ ≤ kuk, and because X ∈ P, kRuk∞ = kuk, for every u ∈ U . Moreover, if G ⊆ X is a non-empty open set (in the relative topology of X) then X \ G cannot belong to P, because X is minimal, so there is a (non-zero) u ∈ U such that |x(u)| < kuk for every x ∈ X \ G. Replacing u by kuk−1 u if need be, we may suppose that kuk = 1. What this means is that W = R[U ] is a linear subspace of C(X) which is isomorphic, as normed space, to U , and has the property that whenever G ⊆ X is a non-empty relatively open set there is an f ∈ W such that kf k∞ = 1 and |f (x)| < 1 for every x ∈ X \ G. Observe that, because X \ G is compact, there is now some α < 1 such that |f (u)| ≤ α for every f ∈ X \ G. Because W is isomorphic to U , it has the Hahn-Banach property. (c) Now consider V = `∞ (X), V0 = W , T0 : V0 → W the identity map. Because W has the Hahn-Banach property, there is a linear operator T : `∞ (X) → W , extending T0 , and of norm kT0 k = 1. (d) If h ∈ `∞ (X) and x0 ∈ X \ {x : h(x) 6= 0}, then (T h)(x0 ) = 0. P P?? Otherwise, set G = {y : y ∈ X \ {x : h(x) 6= 0}, (T h)(y) 6= 0}. This is a non-empty open set in X, so there are f ∈ W , α < 1 such that kf k∞ = 1, |f (x)| ≤ α for every x ∈ X \ G. Because kf k∞ = 1, there must be some x1 ∈ X such that |f (x1 )| = 1, and of course x1 ∈ G, so that (T h)(x1 ) 6= 0. But let δ > 0 be such that δkhk∞ ≤ 1 − α. Then, because h(x) = 0 for x ∈ G, |f (x)| + |δh(x)| ≤ 1 for every x ∈ X, and kf + δhk∞ , kf − δhk∞ are both less than or equal to 1. As T f = f and kT k = 1, this means that kf − δT hk∞ ≤ 1,

kf + δT hk∞ ≤ 1;

consequently |f (x1 )| + δ|(T h)(x1 )| = max(|(f + δT h)(x1 )|, |(f − δT h)(x1 )|) ≤ 1. But |f (x1 )| = 1 and δ(T h)(x1 ) 6= 0, so this is impossible. X XQ Q

310

Function spaces

363R

(e) It follows that T h = h for every h ∈ C(X). P P?? Suppose, if possible, otherwise. Then there is a δ > 0 such that G = {x : |(T h)(x) − h(x)| > δ} is not empty. Let f ∈ W be such that kf k = 1 but |f (x)| < 1 for every x ∈ X \ G. Then there is an x0 ∈ X such that |f (x0 )| = 1; of course x0 must belong to G. Set f1 =

h(x0 ) f, f (x0 )

so that f1 ∈ W and f1 (x0 ) = h(x0 ). Set h1 (x) = max(h(x) − δ, min(h(x) + δ, f1 (x)))

for x ∈ X. Then h1 ∈ C(X). Setting H = {x : |h(x) − h(x0 )| + |f1 (x) − f1 (x0 )| < δ}, H is an open set containing x0 and |f1 (x) − h(x)| ≤ |f1 (x0 ) − h(x0 )| + δ = δ,

h1 (x) = f1 (x)

for every x ∈ H. Consequently x0 ∈ / {x : (f1 − h1 )(x) 6= 0}, and T (f1 − h1 )(x0 ) = 0, by (d). But this means that (T h1 )(x0 ) = (T f1 )(x0 ) = f1 (x0 ) = h(x0 ), so that |h(x0 ) − (T h)(x0 )| = |T (h1 − h)(x0 )| ≤ kT (h1 − h)k∞ ≤ kh1 − hk∞ ≤ δ, which is impossible, because x0 ∈ G. X XQ Q (f ) This tells us at once that W = C(X). But (d) also tells us that X is extremally disconnected. P P Let G ⊆ X be any open set. Then χX = χG + χ(X \ G), so χX = T (χX) = h1 + h2 , where h1 = T (χG), h2 = T (χ(X \ G)). Now from (d) we see that h1 must be zero on X \ G while h2 must be zero on G. Thus we have h1 (x) = 1 for x ∈ G; as h1 is continuous, h1 (x) = 1 for x ∈ G, and h1 = χG. Of course it follows that G is open. As G is arbitrary, X is extremally disconnected. Q Q (g) Being also compact and Hausdorff, therefore regular (3A3Bc), X is zero-dimensional (3A3Bd). We may therefore identify X with the Stone space of its regular open algebra A (314S), and W = C(X) with L∞ (A). Thus R : U → C(X) is a Banach space isomorphism between U and C(X) ∼ = L∞ (A); so U is of the type declared. 363S The Banach-Ulam problem At a couple of points already (232Hc, the notes to §326) I have remarked on a problem which was early recognised as a fundamental question in abstract measure theory. I now set out some formulations of the problem which arise naturally from the work done so far. I will do this by writing down a list of statements which are equiveridical in the sense that if one of them is true, so are all the others; the ‘Banach-Ulam problem’ asks whether they are indeed true. I should remark that this is not generally counted as an ‘open’ problem. It is in fact believed by most of us that these statements are independent of the usual axioms of Zermelo-Fraenkel set theory, including the axiom of choice and even the continuum hypothesis. As such, this problem belongs to Volume 5 rather than anywhere earlier, but its manifestations will become steadily more obtrusive as we continue through this volume and the next, and I think it will be helpful to begin collecting them now. The ideas needed to show that the statements here imply each other are already accessible; in particular, they involve no set theory beyond Zorn’s Lemma. These implications constitute the following theorem, derived from Luxemburg 67a. Theorem The following statements are equiveridical. (i) There are a set X and a probability measure ν, with domain PX, such that ν{x} = 0 for every x ∈ X. (ii) There are a localizable measure space (X, Σ, µ) and an absolutely continuous countably additive functional ν : Σ → R which is not truly continuous, so has no Radon-Nikod´ ym derivative (definitions: 232Ab, 232Hf). (iii) There are a Dedekind complete Boolean algebra A and a countably additive functional ν : A → R which is not completely additive. (iv) There is a Dedekind complete Riesz space U such that Uc∼ 6= U × .

363Xd

L∞

311

proof (a)(i)⇒(ii) Let X be a set with a probability measure ν, defined on PX, such that ν{x} = 0 for every x ∈ X. Let µ be counting measure on X. Then (X, PX, µ) is strictly localizable, and ν : PX → R is countably additive; also νE = 0 whenever µE is finite, so ν is absolutely continuous with respect to µ. But if µE < ∞ then E is finite and ν(X \ E) = 1, so ν is not truly continuous, and has no Radon-Nikod´ ym derivative (232D). (b)(ii)⇒(iii) Let (X, Σ, µ) be a localizable measure space and ν : Σ → R an absolutely continuous countably additive functional which is not truly continuous. Let (A, µ ¯) be the measure algebra of µ; then we have an absolutely continuous countably additive functional ν¯ : A → R defined by setting ν¯E • = νE for every E ∈ Σ (327C). Since ν is not truly continuous, ν¯ is not completely additive (327Ce). Also A is Dedekind complete, because µ is localizable, so A and ν¯ witness (iii). (c)(iii)⇒(i) Let A be a Dedekind complete Boolean algebra and ν : A → R a countably additive functional which is not completely additive. Because ν is bounded (326I), therefore expressible as the difference of non-negative countably additive functionals (326H), there must be a non-negative countably additive functional ν 0 on A which is not completely additive. P By 326N, there is a partition of unity hai ii∈I in A such that i∈I ν 0 ai < ν 0 1. Set K = {i : i ∈ I, ν 0 ai > 0}; then K must be countable, so P ν 0 (supi∈I\K ai ) = ν 0 1 − ν 0 (supi∈K ai ) = ν 0 1 − i∈K ν 0 ai > 0. For J ⊆ I set µJ = ν 0 (supi∈J\K ai ); the supremum is always defined because A is Dedekind complete. Because ν 0 is countably additive and non-negative, so is µ; because ν 0 ai = 0 for i ∈ J \ K, µ{i} = 0 for every i ∈ I. Multiplying µ by a suitable scalar, if need be, (I, PI, µ) witnesses that (i) is true. (d)(iii)⇒(iv) If A is a Dedekind complete Boolean algebra with a countably additive functional which is not completely additive, then U = L∞ (A) is a Dedekind complete Riesz space (363Mb) and Uc∼ 6= U × , by 363K (recalling, as in (c) above, that the functional must be bounded). (e)(iv)⇒(iii) Let U be a Dedekind complete Riesz space such that U × 6= Uc∼ . Take f ∈ Uc∼ \ U × ; replacing f by |f | if need be, we may suppose that f ≥ 0 is sequentially order-continuous but not ordercontinuous (355H, 355I). Let A be a non-empty downwards-directed set in U , with infimum 0, such that inf u∈A f (u) > 0 (351Ga). Take e ∈ A, and consider the solid linear subspace Ue of U generated by e; write g for the restriction of f to Ue . Because the embedding of Ue in U is order-continuous, g ∈ (Ue )∼ c ; because A ∩ Ue is downwards-directed and has infimum 0, and inf u∈A∩Ue g(u) = inf u∈A f (u) > 0, Ue× .

g∈ / But Ue is a Riesz space with order unit e, and is Dedekind complete because U is; so it can be identified with L∞ (A) for some Boolean algebra A (363N), and A is Dedekind complete, by 363M. ∞ × Accordingly we have a Dedekind complete Boolean algebra A such that L∞ (A)∼ c 6= L (A) . By 363K, there is a (bounded) countably additive functional on A which is not completely additive, and (iii) is true. 363X Basic exercises (a) Let A be a Boolean algebra and U a Banach algebra. Let ν : A → U be a bounded additive function and T : L∞ (A) → U the corresponding bounded linear operator. Show that T is multiplicative iff ν(a ∩ b) = νa × νb for all a, b ∈ A. > (b) Let A, B be Boolean algebras and T : L∞ (A) → L∞ (A) a linear operator. Show that the following are equiveridical: (i) there is a Boolean homomorphism π : A → A such that T = Tπ (ii) T (u × v) = T u × T v for all u, v ∈ L∞ (A) (iii) T is a Riesz homomorphism and T eA = eB , where eA is the standard order unit of L∞ (A). (c) Let A, B be Boolean algebras and T : L∞ (A) → L∞ (B) a Riesz homomorphism. Show that there are a Boolean homomorphism π : A → B and a v ≥ 0 in L∞ (B) such that T u = v × Tπ u for every u ∈ L∞ (A), where Tπ is the operator associated with π (363F). (d) Let A be a Boolean algebra and C a subalgebra of A. Show that L∞ (C), regarded as a subspace of L (A) (363Ga), is order-dense in L∞ (A) iff C is order-dense in A. ∞

312

Function spaces

363Xe

> (e) Let (X, Σ, µ) be a measure space with measure algebra A, and L∞ the space of bounded Σmeasurable real-valued functions on X. A linear lifting of µ is a positive linear operator T : L∞ (A) → L∞ such that T (χ1A ) = χX and (T u)• = u for every u ∈ L∞ (A), writing f 7→ f • for the canonical map from L∞ to L∞ (A) (363H-363I). (i) Show that if θ : A → Σ is a lifting in the sense of 341A then Tθ , as defined in 363F, is a linear lifting. (ii) Show that if T : L∞ (A) → L∞ is a linear lifting, then there is a corresponding lower density θ : A → Σ defined by setting θa = {x : T (χa)(x) = 1} for each a ∈ A. (iii) Show that θ, as defined in (ii), is a lifting iff T is a Riesz homomorphism. (f ) Let U be any commutative ring with multiplicative identity 1. Show that the set A of idempotents in U (that is, elements a ∈ U such that a2 = a) is a Boolean algebra with identity 1, writing a ∩ b = ab, 1 \ a = 1 − a for a, b ∈ A. (g) Let A be a Boolean algebra. Show that A is isomorphic to the Boolean algebras of idempotents of S(A) and L∞ (A). (h) Let A be a Dedekind σ-complete Boolean algebra. (i) Show that for any u ∈ L∞ (A), α ∈ R there are elements [[u ≥ α]], [[u > α]] ∈ A, where [[u ≥ α]] is the largest a ∈ A such that u × χa ≥ αχa, and [[u > α]] = supβ>α [[u ≥ β]]. (ii) Show that in the context of 363H, if u corresponds to f • for f ∈ L∞ , then [[u ≥ α]] = {x : f (x) ≥ α}• , [[u > α]] = {x : f (x) > α}• . (iii) Show that if A ⊆ L∞ is non-empty and v ∈ L∞ , then v = sup A iff [[v > α]] = supu∈A [[u > α]] for every α ∈ R; in particular, v = u iff [[v > α]] = [[u > α]] for every α. (iv) Show that a function φ : R → A is of the form φ(α) = [[u > α]] iff (α) φ(α) = supβ>α φ(β) for every α ∈ R (β) there is an M such that φ(M ) = 0, φ(−M ) = 1. (v) Put (iii) and (iv) together to give a proof that L∞ is Dedekind σ-complete if A is. (i) Let A be a Dedekind σ-complete Boolean algebra and U ⊆ L∞ (A) a (sequentially) order-closed Riesz subspace containing χ1. Show that U can be identified with L∞ (B) for some (sequentially) order-closed subalgebra B ⊆ A. (Hint: set B = {b : χb ∈ U } and use 363N.) (j) For a Boolean algebra A, with Stone space Z, write L∞ C (A) for the Banach algebra C(Z; C) of continuous complex-valued functions on Z. Prove results corresponding to 363C, 363Ea, 363F-363I for the complex case. 363Y Further exercises (a) Let A be a Boolean algebra. Given the linear structure, ordering, multiplication and norm of S(A) as described in §361, show that a norm completion of S(A) will serve for L∞ (A) in the sense that all the results of 363B-363Q can be proved with no use of the axiom of choice except an occasional appeal to countably many choices in sequential forms of the theorems. (b) Let A be a Boolean algebra. Show that A is ccc iff L∞ (A) has the countable sup property (241Yd, 353Ye). (c) Let X be an extremally disconnected topological space, and G its regular open algebra. Show that there is a natural isomorphism between L∞ (G) and Cb (X). (d) Let X be a compact Hausdorff space. Let us say that a linear subspace U of C(X) is `∞ -complemented in C(X) if there is a linear subspace V such that C(X) = U ⊕ V and ku + vk∞ = max(kuk∞ , kvk∞ ) for all u ∈ U , v ∈ V . Show that there is a one-to-one correspondence between such subspaces U and openand-closed subsets E of X, given by setting U = {f : f ∈ C(X), f (x) = 0 ∀ x ∈ X \ E}. Hence show that if A is any Boolean algebra, there is a canonical isomorphism between A and the partially ordered set of `∞ -complemented subspaces of L∞ (A). (e) Let A be a Boolean algebra. (i) If u ∈ L∞ = L∞ (A), show that |u| = e, the standard order unit of L , iff max(ku + vk∞ , ku − vk∞ ) > 1 whenever v ∈ L∞ \ {0}. (ii) Show that if u, v ∈ L∞ then |u| ∧ |v| = 0 iff kαu + v + wk∞ ≤ max(kαu + wk∞ , kv + wk∞ ) whenever α = ±1 and w ∈ L∞ . (iii) Show that if T : L∞ → L∞ is a normed space automorphism then there are a Boolean automorphism π : A → A and a w ∈ L∞ such that |w| = e and T u = w × Tπ u for every u ∈ L∞ . ∞

363 Notes

L∞

313

(f ) Let X be a set, Σ an algebra of subsets of X, and I an ideal in Σ. (i) Show that L∞ (Σ) can be identified, as Banach lattice and Banach algebra, with the space L∞ of bounded functions f : X → R such that whenever α < β in R there is an E ∈ Σ such that {x : f (x) ≤ α} ⊆ E ⊆ {x : f (x) ≤ β}. (ii) Show that L∞ = {gφ : g ∈ C(Z)}, where Z is the Stone space of Σ and φ : X → Z is a function (to be described). (iii) Show that L∞ (Σ/I) can be identified, as Banach lattice and Banach algebra, with L∞ /V, where V is the set of those functions f ∈ L∞ such that for every ² > 0 there is a member of I including {x : |f (x)| ≥ ²}. (g) Let (X, Σ, µ) be a complete probability space with measure algebra A. Let hBn in∈N be a nondecreasing sequence of closed subalgebras of A such that A is the closed subalgebra of itself generated by S • 1 1 n∈N Bn , and set Σn = {F : F ∈ Bn } for each n. Let Pn : L (µ) → L (µ¹Σn ) be the conditional ∞ expectation operator for each n, so that Pn ¹L (µ) is a positive linear operator from L∞ (µ) ∼ = L∞ (A) to ∞ ∞ ∼ L (µ¹Σn ) = L (Bn ). Suppose that we are given for each n a lifting θn : Bn → Σn and that θn+1 b = θn b whenever n ∈ N, b ∈ Bn . Let Tn : L∞ (Bn ) → L∞ be the corresponding linear liftings (363Xe), and F any non-principal ultrafilter on N. (i) Show that for any u ∈ L∞ (A), hTn Pn uin∈N converges almost everywhere. (ii) For u ∈ L∞ (A) set (T u)(x) = limn→F (Tn Pn u)(x) for x ∈ X, u ∈ L∞ (A). Show that T is a linear lifting of µ. (iii) Use 363Xe(ii) and 341J to show that there is a lifting θ of µ extending every θn . (iv) Use this as the countable-cofinality inductive step in a proof of the Lifting Theorem (using partial liftings rather than partial lower densities, as suggested in 341Li). (h) Let A be a Dedekind σ-complete Boolean algebra and ν : A → R a countably additive func∞ tional. Suppose that hun in∈N is an order-bounded R sequence in LR (A) such that inf n∈N supm≥n um and supn∈N inf m≥n um are equal to u say. Show that u dν = limn→∞ un dν. (i) Let Σ be the family of those sets E ⊆ [0, 1] such that µ(int E) = µE, where µ is Lebesgue measure. (i) Show that Σ is an algebra of subsets of [0, 1] and that every member of Σ is Lebesgue measurable. (ii) Show that if we identify L∞ (Σ) with a set of real-valued functions on [0, 1], as in 363Yf, R then we get just the space of Riemann integrable functions. (iii) Show that if we write ν for µ¹Σ, then dν, as defined in 363L, is just the Riemann integral. (j) Show that a normed space over C has the Hahn-Banach property of 363R for complex spaces iff it is isomorphic to L∞ C (A), as described in 363Xj, for some Dedekind complete Boolean algebra A. 363 Notes and comments As with S(A), I have chosen a definition of L∞ (A) in of the Stone space of A; but as with S(A), this is optional (363Ya). By and large the basic properties of L∞ are derived very naturally from those of S. The spaces L∞ (A), for general Boolean algebras A, are not in fact particularly important; they have too few properties not shared by all the spaces C(X) for compact Hausdorff X. The point at which it becomes helpful to interpret C(X) as L∞ (A) is when C(X) is Dedekind σ-complete. The spaces X for which this is true are difficult to picture, and alternative representations of L∞ along the lines of 363H-363I can be easier on the imagination. For Dedekind σ-complete A, there is an alternative description of of L∞ (A) in of objects ‘[[u > α]]’ (363Xh); I will return to this idea in the next section. For the moment I remark only that it gives an alternative approach to 363M not necessarily depending on the representation of L∞ as a quotient L∞ /V nor on an analysis of a Stone space. I used a version of such an argument in the proof of 363M which I gave in Fremlin 74a, 43D. I spend so much time on 363M not only because Dedekind completeness is one of the basic properties of any lattice, but because it offers an abstract expression of one of the central results of Chapter 24. In 243H I showed that L∞ (µ) is always Dedekind σ-complete, and that it is Dedekind complete if µ is localizable. We can now relate this to the results of 321H and 322Be: the measure algebra of any measure is Dedekind σ-complete, and the measure algebra of a localizable measure is Dedekind complete. The ideas of the proof of 363M can of course be rearranged in various ways. One uses 353Yb: for completely regular spaces X, C(X) is Dedekind complete iff X is extremally disconnected; while for compact Hausdorff spaces, X is extremally disconnected iff it is the Stone space of a Dedekind complete algebra. With the right modification of the concept ‘extremally disconnected’ (314Yf), the same approach works for Dedekind σ-completeness.

314

Function spaces

363 Notes

363R is the ‘Nachbin-Kelley theorem’; it is commonly phrased ‘a normed space U has the Hahn-Banach extension property iff it is isomorphic, as normed space, to C(X) for some compact extremally disconnected Hausdorff space X’, but the expression in of L∞ spaces seems natural in the present context. The implication in one direction (Part A of the proof) calls for nothing but a check through one of the standard proofs of the Hahn-Banach theorem to make sure that the argument applies in the generalized form. Part B of the proof has ideas in it; I have tried to set it out in a way suggesting that if you can the construction of the set X the rest is just a matter of a little ingenuity. One way of trying to understand the multiple structures of L∞ spaces is by looking at the corresponding automorphisms. We observe, for instance, that an operator T from L∞ (A) to itself is a Banach algebra automorphism iff it is a Banach lattice automorphism preserving the standard order unit iff it corresponds to an automorphism of the algebra A (363Xb). Of course there are Banach space automorphisms of L∞ which do not respect the order or multiplicative structure; but they have to be closely related to algebra isomorphisms (363Ye). I devote a couple of exercises (363Xe, 363Yg) to indications of how the ideas here are relevant to the Lifting Theorem. If you found the formulae of the proof of 341G obscure it may help to work through the parallel argument. A lecture by W.A.J.Luxemburg on the equivalence between (i) and (iv) in 363S was one of the turning points in my mathematical apprenticeship. I introduce it here, even though the real importance of the Banach-Ulam problem lies in the metamathematical ideas it has nourished, because these formulations provide a focus for questions which arise naturally in this volume and which otherwise might prove distracting. The next group of significant ideas in this context will appear in §438.

364 L0 My next objective is to develop an abstract construction corresponding to the L0 (µ) spaces of §241. These generalized L0 spaces will form the basis of the work of the rest of this chapter and also the next; partly because their own properties are remarkable, but even more because they form a framework for the study of Archimedean Riesz spaces in general (see §368). There seem to be significant new difficulties, and I take the space to describe an approach which can be made essentially independent of the route through Stone spaces used in the last three sections. I embark directly on a definition in the new language (364A), and relate it to the constructions of §241 (364C-364E, 364J) and §§361-363 (364K). The ideas of Chapter 27 can also be expressed in this language; I make a start on developing the machinery for this in 364G, with the formula ‘[[u ∈ E]]’, ‘the region in which u belongs to E’, and some exercises (364Xd-364Xf). Following through the questions addressed in §363, I discuss Dedekind completeness in L0 (364M-364O), properties of its multiplication (364P), the expression of the original algebra in of L0 (364Q), the action of Boolean homomorphisms on L0 (364R) and product spaces (364S). In 364T-364W I describe representations of the L0 space of a regular open algebra. 364A Definition Let A be a Dedekind σ-complete Boolean algebra. I will write L0 (A) for the set of all functions α 7→ [[u > α]] : R → A such that (α) [[u > α]] = supβ>α [[u > β]] in A for every α ∈ R, (β) inf α∈R [[u > α]] = 0, (γ) supα∈R [[u > α]] = 1. 364B Remarks (a) My reasons for using the notation ‘[[u > α]]’ rather than ‘u(α)’ will I hope become clear in the next few paragraphs. For the moment, if you think of A as a σ-algebra of sets and of L0 (A) as the family of A-measurable real-valued functions, then [[u > α]] corresponds to the set {x : u(x) > α} (364Ja). (b) Some readers will recognise the formula ‘[[. . . ]]’ as belonging to the language of forcing, so that [[u > α]] could be read as ‘the Boolean value of the proposition “u > α”’. But a vocalisation closer to my intention might be ‘the region where u > α’.

L0

364C

315

(c) Note that condition (α) of 364A automatically ensures that [[u > α]] ⊆ [[u > α0 ]] whenever α0 ≤ α in R. (d) In fact it will sometimes be convenient to note that the conditions of 364A can be replaced by (α0 ) [[u > α]] = supq∈Q,q>α [[u > q]] in A for every α ∈ R, (β 0 ) inf n∈N [[u > n]] = 0, (γ 0 ) supn∈N [[u > −n]] = 1; the point being that we need look only at suprema and infima of countable subsets of A. (e) In order to make sense of this construction we need to match it with an alternative route to the same object, based on σ-algebras and σ-ideals of sets, as follows. 364C Proposition Let X be a set, Σ a σ-algebra of subsets of X, and I a σ-ideal of Σ. (a) Write L0 for the space of all Σ-measurable functions from X to R. Then L0 , with its linear structure, ordering and multiplication inherited from RX , is a Dedekind σ-complete f -algebra with multiplicative identity. (b) Set W = {f : f ∈ L0 , {x : f (x) 6= 0} ∈ I}. Then (i) W is a sequentially order-closed solid linear subspace and ideal of L0 ; (ii) the quotient space L0 /W, with its inherited linear, order and multiplicative structures, is a Dedekind σ-complete Riesz space and an f -algebra with a multiplicative identity; (iii) for f , g ∈ L0 , f • ≤ g • in L0 /W iff {x : f (x) > g(x)} ∈ I, and f • = g • in L0 /W iff {x : f (x) 6= g(x)} ∈ I. proof (Compare 241A-241H.) (a) The point is just that L0 is a sequentially order-closed Riesz subspace and subalgebra of RX . The facts we need to know – that constant functions belong to L0 , that f + g, αf , f × g, supn∈N fn belong to L0 whenever f , g, fn do and {fn : n ∈ N} is bounded above – are all covered by 121E-121F. Its multiplicative identity is of course the constant function χX. (b)(i) The necessary verifications are all elementary. (ii) Because W is a solid linear subspace of the Riesz space L0 , the quotient inherits a Riesz space structure (352Jb); because W is an ideal of the ring (L0 , +, ×), L0 /W inherits a multiplication; it is a commutative algebra because L0 is; and has a multiplicative identity e = χX • because χX is the identity of L0 . To check that L0 /W is an f -algebra it is enough to observe that, for any non-negative f , g , h ∈ L0 , f • × g • = (f × g)• ≥ 0, and if f • ∧ g • = 0 then {x : f (x) > 0} ∩ {x : g(x) > 0} ∈ I, so that {x : f (x)h(x) > 0} ∩ {x : g(x) > 0} ∈ I and (f • × h• ) ∧ g • = (h• × f • ) ∧ g • = 0. Finally, L0 /W is Dedekind σ-complete, by 353J(a-iii). (iii) For f , g ∈ L0 , f • ≤ g • ⇐⇒ (f − g)+ ∈ W ⇐⇒ {x : f (x) > g(x)} = {x : (f − g)+ (x) 6= 0} ∈ I (using the fact that the canonical map from L0 to L0 /W is a Riesz homomorphism, so that ((f − g)+ )• = (f • − g • )+ ). Similarly f • = g • ⇐⇒ f − g ∈ W ⇐⇒ {x : f (x) 6= g(x)} = {x : (f − g)(x) 6= 0} ∈ I.

316

Function spaces

364D

364D Theorem Let X be a set and Σ a σ-algebra of subsets of X. Let A be a Dedekind σ-complete Boolean algebra and π : Σ → A a surjective Boolean homomorphism, with kernel a σ-ideal I; define L0 and W as in 364C, so that U = L0 /W is a Dedekind σ-complete f -algebra with multiplicative identity. (a) We have a canonical bijection T : U → L0 = L0 (A) defined by the formula [[T f • > α]] = π{x : f (x) > α} for every f ∈ L0 , α ∈ R. (b)(i) For any u, v ∈ U , [[T (u + v) > α]] = supq∈Q [[T u > q]] ∩ [[T v > α − q]] for every α ∈ R. (ii) For any u ∈ U , γ > 0, [[T (γu) > α]] = [[T u >

α γ ]]

for every α ∈ R. (iii) For any u, v ∈ U , u ≤ v ⇐⇒ [[T u > α]] ⊆ [[T v > α]] for every α ∈ R. (iv) For any u, v ∈ U + , [[T (u × v) > α]] = supq∈Q,q>0 [[T u > q]] ∩ [[T v >

α q ]]

for every α ≥ 0. (v) Writing e = (χX)• for the multiplicative identity of U , we have [[T e > α]] = 1 if α < 1, 0 if α ≥ 1. proof (a)(i) Given f ∈ L0 , set ζf (α) = π{x : f (x) > α} for α ∈ R. Then it is easy to see that ζf satisfies the conditions (α)0 -(γ)0 of 364Bd, because π is sequentially order-continuous (313Qb). Moreover, if f • = g • in U , then ζf (α) 4 ζg (α) = π({x : f (x) > α}4{x : g(x) > α}) = 0 for every α ∈ R, because {x : f (x) > α}4{x : g(x) > α} ⊆ {x : f (x) 6= g(x)} ∈ I, and ζf = ζg . So we have a well-defined member T u of L0 defined by the given formula, for any u ∈ U . (ii) Next, given w ∈ L0 , there is a u ∈ L0 /W such that T u = w. P P For each q ∈ Q, choose Fq ∈ Σ such that πFq = [[w > q]] in A. Note that if q 0 ≥ q then π(Fq0 \ Fq ) = [[u > q 0 ]] \ [[u > q]] = 0, so Fq0 \ Fq ∈ I. Set H=

S q∈Q

Fq \

T

S n∈N

q∈Q,q≥n

Fq ∈ Σ,

and for x ∈ X set f (x) = sup{q : q ∈ Q, x ∈ Fq } if x ∈ H, = 0 otherwise. (H is chosen just to make the formula here give a finite value for every x.) We have

πH = sup [[w > q]] \ inf

sup

n∈N q∈Q,q≥n

q∈Q

[[w > q]]

= 1A \ inf [[w > n]] = 1A \ 0A = 1A , n∈N

so X \ H ∈ I. Now, for any α ∈ R,

L0

364D

[

{x : f (x) > α} =

317

Fq ∪ (X \ H) if α < 0

q∈Q,q>α

[

=

Fq \ (X \ H) if α ≥ 0,

q∈Q,q>α

and in either case belongs to Σ; so that f ∈ L0 and f • is defined in L0 . Next, for any α ∈ R, [[T f • > α]] = π{x : f (x) > α} = π(

[

Fq )

q∈Q,q>α

=

sup

[[w > q]] = [[w > α]],

q∈Q,q>α

and T f • = w. Q Q (iii) Thus T is surjective. To see that it is injective, observe that if f , g ∈ L0 , then T f • = T g • =⇒ [[T f • > α]] = [[T g • > α]] for every α ∈ R =⇒ π{x : f (x) > α} = π{x : g(x) > α} for every α ∈ R =⇒ {x : f (x) > α}4{x : g(x) > α} ∈ I for every α ∈ R [ =⇒ {x : f (x) 6= g(x)} = ({x : f (x) > q}4{x : g(x) > q}) ∈ I q∈Q

=⇒ f = g . •

•

So we have the claimed bijection. (b)(i) Let f , g ∈ L0 be such that u = f • , v = g • , so that u + v = (f + g)• . For any α ∈ R,

[[T (u + v) > α]] = π{x : f (x) + g(x) > α} [ = π( {x : f (x) > q} ∩ {x : g(x) > α − q}) q∈Q

= sup π{x : f (x) > q} ∩ π{x : g(x) > α − q} q∈Q

(because π is a sequentially order-continuous Boolean homomorphism) = sup [[T u > q]] ∩ [[T v > α − q]]. q∈Q

(ii) Let f ∈ L0 be such that f • = u, so that (γf )• = γu. For any α ∈ R, α γ

[[T (γu) > α]] = π{x : γf (x) > α} = π{x : f (x) > } = [[T u >

α γ ]].

(iii) Let f , g ∈ L0 be such that f • = u and g • = v. Then

u ≤ v ⇐⇒ {x : f (x) > g(x)} ∈ I (see 364C(b-iii)) ⇐⇒

[

{x : f (x) > q ≥ g(x)} ∈ I

q∈Q

⇐⇒ {x : f (x) > α} \ {x : g(x) > α} ∈ I for every α ∈ R ⇐⇒ π{x : f (x) > α} \ π{x : g(x) > α} = 0 for every α ⇐⇒ [[T u > α]] ⊆ [[T v > α]] for every α.

318

Function spaces

364D

(iv) Now suppose that u, v ≥ 0, so that they can be expressed as f • , g • where f , g ≥ 0 in L0 (351J), and u × v = (f × g)• . If α ≥ 0, then [[T (u × v) > α]] = π(

[

α q

{x : f (x) > q} ∩ {x : g(x) > })

q∈Q,q>0

=

α q

sup π{x : f (x) > q} ∩ π{x : g(x) > } q∈Q,q>0

=

sup [[T u > q]] ∩ [[T v > q∈Q,q>0

α q ]].

(v) This is trivial, because [[T (χX)• > α]] = π{x : (χX)(x) > α} = πX = 1 if α < 1, = π∅ = 0 if α ≥ 1. 364E Theorem Let A be a Dedekind σ-complete Boolean algebra. Then L0 = L0 (A) has the structure of a Dedekind σ-complete f -algebra with multiplicative identity e, defined by saying [[u + v > α]] = supq∈Q [[u > q]] ∩ [[v > α − q]], whenever u, v ∈ L0 and α ∈ R, [[γu > α]] = [[u >

α γ ]]

whenever u ∈ L0 , γ ∈ ]0, ∞[ and α ∈ R, u ≤ v ⇐⇒ [[u > α]] ⊆ [[v > α]] for every α ∈ R, [[u × v > α]] = supq∈Q,q>0 [[u > q]] ∩ [[v >

α q ]]

whenever u, v ≥ 0 in L0 and α ≥ 0, [[e > α]] = 1 if α < 1, 0 if α ≥ 1. proof (a) By the Loomis-Sikorski theorem (314M), we can find a set Z (the Stone space of A), a σ-algebra Σ of subsets of Z (the algebra generated by the open-and-closed sets and the ideal M of meager sets) and a surjective sequentially order-continuous Boolean homomorphism π : Σ → A (corresponding to the identification between A and the quotient Σ/M). Consequently, defining L0 and W as in 364C, we have a bijection between the Dedekind σ-complete f -algebra L0 /W and L0 (364Da). Of course this endows L0 itself with the structure of a Dedekind σ-complete f -algebra; and 364Db tells us that the description of the algebraic operations above is consistent with this structure. (b) In fact the f -algebra structure is completely defined by the description offered. For while scalar multiplication is not described for γ ≤ 0, the assertion that L0 is a Riesz space implies that 0u = 0 and that γu = (−γ)(−u) for γ < 0; so if we have formulae to describe u + v and γu for γ > 0, this suffices to define the linear structure of L0 . Note that we have an element 0 in L0 defined by setting [[0 > α]] = 0 if α ≥ 0, 1 if α < 0, and the formula for u + v shows us that [[0 + u > α]] = supq∈Q [[0 > q]] ∩ [[u > α − q]] = supq∈Q,q<0 [[u > α − q]] = [[u > α]] for every α, so that 0 is the zero of L0 . As for multiplication, if L0 is to be an f -algebra we must have [[u × v > α]] ⊇ [[0 > α]] = 1 whenever u, v ∈ (L0 )+ and α < 0, because u × v ≥ 0. So the formula offered is sufficient to determine u × v for non-negative u and v; and for others we know that

L0

364I

319

u × v = (u+ × v + ) − (u+ × v − ) − (u− × v + ) + (u− × v − ), so the whole of the multiplication of L0 is defined. 364F The rest of this section will be devoted to understanding the structure just established. I start with a pair of elementary facts. Lemma Let A be a Dedekind σ-complete Boolean algebra. (a) If u, v ∈ L0 = L0 (A) and α, β ∈ R, [[u + v > α + β]] ⊆ [[u > α]] ∪ [[v > β]]. 0

(b) If u, v ≥ 0 in L and α, β ≥ 0 in R, [[u × v > αβ]] ⊆ [[u > α]] ∪ [[v > β]]. proof (a) For any q ∈ Q, either q ≥ α and [[u > q]] ⊆ [[u > α]], or q ≤ α and [[v > α + β − q]] ⊆ [[v > β]]; thus in all cases [[u > q]] ∩ [[v > α + β − q]] ⊆ [[u > α]] ∪ [[v > β]]; taking the supremum over q, we have the result. (b) The same idea works, replacing α + β − q by αβ/q for q > 0. 364G Yet another description of L0 is sometimes appropriate, and leads naturally to an important construction (364I). Proposition Let A be a Dedekind σ-complete Boolean algebra. Then there is a bijection between L0 = L0 (A) and the set Φ of sequentially order-continuous Boolean homomorphisms from the algebra B of Borel subsets of R to A, defined by saying that u ∈ L0 corresponds to φ ∈ Φ iff [[u > α]] = φ(]α, ∞[) for every α ∈ R. proof (a) If φ ∈ Φ, then the map α 7→ φ(]α, ∞[) satisfies the conditions of 364Bd, so corresponds to an element uφ of L0 . (b) If φ, ψ ∈ Φ and uφ = uψ , then φ = ψ. P P Set A = {E : E ∈ B, φ(E) = ψ(E)}. Then A is a σ-subalgebra of B, because φ and ψ are both sequentially order-continuous Boolean homomorphisms, and contains ]α, ∞[ for every α ∈ R. Now A contains ]−∞, α] for every α, and therefore includes B (121J). But this means that φ = ψ. Q Q (c) Thus φ 7→ uφ is injective. But it is also surjective. P P As in 364E, take a set Z, a σ-algebra Σ of subsets of Z and a surjective sequentially order-continuous Boolean homomorphism π : Σ → A; let T : L0 /W → L0 be the bijection described in 364D. If u ∈ L0 , there is an f ∈ L0 such that T f • = u. Now consider φE = πf −1 [E] for E ∈ B. f −1 [E] always belongs to Σ (121Ef), so φE is always well-defined; E 7→ f −1 [E] and π are sequentially order-continuous, so φ also is; and φ(]α, ∞[) = π{z : f (z) > α} = [[u > α]] for every α, so u = uφ . Q Q Thus we have the declared bijection. 364H Definition In the context of 364G, I will write [[u ∈ E]], ‘the region where u takes values in E’, for φ(E), where φ : B → A is the homomorphism corresponding to u ∈ L0 . Thus [[u > α]] = [[u ∈ ]α, ∞[ ]]. In the same spirit I write [[u ≥ α]] for [[u ∈ [α, ∞[ ]] = inf β<α [[u > β]], [[u 6= 0]] = [[|u| > 0]] = [[u > 0]] ∪ [[u < 0]] and so on, so that (for instance) [[u = α]] = [[u ∈ {α}]] = [[u ≥ α]] \ [[u > α]] for u ∈ L0 , α ∈ R. 364I Proposition Let A be a Dedekind σ-complete Boolean algebra, E ⊆ R a Borel set, and h : E → R a Borel measurable function. Then whenever u ∈ L0 = L0 (A) is such that [[u ∈ E]] = 1, there is an element ¯ ¯ h(u) of L0 defined by saying that [[h(u) ∈ F ]] = [[u ∈ h−1 [F ]]] for every Borel set F ⊆ R. proof All we have to observe is that F → 7 [[u ∈ h−1 [F ]]] is a sequentially order-continuous Boolean homomorphism. (The condition ‘[[u ∈ E]] = 1’ ensures that [[u ∈ h−1 [R]]] = 1.)

320

Function spaces

364J

364J Examples Perhaps I should spell out the most important contexts in which we apply these ideas, even though they have in effect already been mentioned. (a) Let X be a set and Σ a σ-algebra of subsets of X. Then we may identify L0 (Σ) with the space L0 of Σ-measurable real-valued functions on X. (This is the case A = Σ of 364D.) For f ∈ L0 , [[f ∈ E]] (364H) ¯ ) (364I) is just the is just f −1 [E], for any Borel set E ⊆ R; and if h is a Borel measurable function, h(f 0 composition hf , for any f ∈ L . (b) Now suppose that I is a σ-ideal of Σ and that A = Σ/I. Then, as in 364D, we identify L0 (A) with ¯ • ) = (hf )• , for any Borel set E and any Borel a quotient L0 /W. For f ∈ L0 , [[f • ∈ E]] = f −1 [E]• , and h(f measurable function h : R → R. (c) In particular, if (X, Σ, µ) is a measure space with measure algebra A, then L0 (A) becomes identified with L0 (µ) as defined in §241. The same remarks as in 363I apply here; the space L0 (µ) of §241 is larger than the space L0 of the present section. But for every f ∈ L0 (µ) there is a g ∈ L0 such that g = f a.e. (241Bk), so that L0 (µ) can be identified with L0 /N, where N is the set of functions in L0 which are zero almost everywhere (241Yh). 364K Embedding S and L∞ in L0 : Proposition Let A be a Dedekind σ-complete Boolean algebra. (a) We have a canonical embedding of L∞ = L∞ (A) as an order-dense solid linear subspace of L0 = L0 (A); it is the solid linear subspace generated by the multiplicative identity e of L0 . Consequently S = S(A) is also embedded as an order-dense Riesz subspace and subalgebra of L0 . (b) This embedding respects the linear, lattice and multiplicative structures of L∞ and S. (c) For a ∈ A, χa, when regarded as a member of L0 , can be described by the formula [[χa > α]] = 1 if α < 0, = a if 0 ≤ α < 1, = 0 if 1 ≤ α. The function χ : A → L0 is additive, injective, order-continuous and a lattice homomorphism. (d) For every u ∈ (L0 )+ there is a non-decreasing sequence hun in∈N in S such that u0 ≥ 0 and supn∈N un = u. proof Let Z, Σ, L0 , W and π be as in the proof of 364E. I defined L∞ to be the space C(Z) of continuous realvalued functions on Z (363A); but because A is Dedekind σ-complete, there is an alternative representation as L∞ /W ∩ L∞ , where L∞ is the space of bounded Σ-measurable functions from Z to R (363Hb). Put like this, we clearly have an embedding of L∞ ∼ = L0 /W; and this embedding represents = L∞ /W ∩ L∞ in L0 ∼ ∞ 0 ∞ L as a Riesz subspace and subalgebra of L , because L is a Riesz subspace and subalgebra of L0 . L∞ becomes the solid linear subspace of L0 generated by (χZ)• = e, because L∞ is the solid linear subspace of L0 generated by χZ. To see that L∞ is order-dense in L0 , we have only to note that f = supn∈N f ∧ nχZ in L0 for every f ∈ L0 , and therefore (because the map f 7→ f • is sequentially order-continuous) u = supn∈N u∧ne in L0 for every u ∈ L0 . To identify χa, we have the formula χ(πE) = (χE)• , as in 363Hc; but this means that, if a = πE, [[χa > α]] = π{z : χE(z) > α} = πZ = 1 if α < 0, = πE = a if 0 ≤ α < 1, = π∅ = 0 if α ≥ 1, using the formula in 364Da. Evidently χ is injective. Because S is an order-dense Riesz subspace and subalgebra of L∞ (363C), the same embedding represents it as an order-dense Riesz subspace and subalgebra of L0 . (For ‘order-dense’, use 352Nc.) Because χ : A → L∞ is additive, order-continuous and a lattice homomorphism (363D), and the embedding map L∞ ⊆ L0 also is, χ : A → L0 has the same properties. Finally, if u ≥ 0 in L0 , we can represent it as f • where f ≥ 0 in L0 . For n ∈ N set

L0

364M

321

fn (z) = 2−n k if 2−n k ≤ f (z) < 2−n (k + 1) where 0 ≤ k < 4n , = 0 if f (z) ≥ 2n ; then hfn• in∈N is a non-decreasing sequence in S + with supremum u. 364L Corollary Let (A, µ ¯) be a measure algebra. Then S(Af ) can be embedded as a Riesz subspace 0 of L (A), which is order-dense iff (A, µ ¯) is semi-finite. proof (Recall that Af is the ring {a : µ ¯a < ∞}.) The embedding Af ⊆ A is an injective ring homomorphism, f so induces an embedding of S(A ) as a Riesz subspace of S(A), by 361J. Now S(Af ) is order-dense Pn in S(A) iff (A, µ ¯) is semi-finite. P P (i) If (A, µ ¯) is semi-finite and v > 0 in S(A), then v is expressible as j=0 βj χbj where βj ≥ 0 for each j and some βj χbj is non-zero; now there is a non-zero a ∈ Af such that a ⊆ bj , so that 0 < βj χa ∈ S(Af ) and βj χa ≤ v. As v is arbitrary, S(Af ) is quasi-order-dense, therefor order-dense (353A). (ii) If S(Af ) is order-dense in S(A) and b ∈ A \ {0}, there is a u > 0 in S(Af ) such that u ≤ χb; now there are α > 0, a ∈ Af \ {0} such that αχa ≤ u, in which case a ⊆ b. Q Q Now because S(Af ) ⊆ S(A) and S(A) is order-dense in L0 (A), we must have S(Af ) is order-dense in L0 (A) ⇐⇒ S(Af ) is order-dense in S(A) ⇐⇒ (A, µ ¯) is semi-finite. 364M Suprema and infima in L0 We know that any L0 (A) is a Dedekind σ-complete partially ordered set. There is a useful description of suprema for this ordering, as follows. Proposition Let A be a Dedekind σ-complete Boolean algebra and A a subset of L0 = L0 (A). (a) A is bounded above in L0 iff there is a sequence hcn in∈N in A, with infimum 0, such that [[u > n]] ⊆ cn for every u ∈ A. (b) If A is non-empty, then A has a supremum in L0 iff cα = supu∈A [[u > α]] is defined in A for every α ∈ R and inf n∈N cn = 0; and in this case cα = [[sup A > α]] for every α. (c) If A is non-empty and bounded above, then A has a supremum in L0 iff supu∈A [[u > α]] is defined in A for every α ∈ R. proof (a)(i) If A has an upper bound u0 , set cn = [[u0 > n]] for each n; then hcn in∈N satisfies the conditions. (ii) If hcn in∈N satisfies the conditions, set φ(α) = 1 if α < 0, = inf ci if n ∈ N, α ∈ [n, n + 1[ . i≤n

Then it is easy to check that φ satisfies the conditions of 364A, since inf n∈N cn = 0. So there is a u0 ∈ L0 such that φ(α) = [[u0 > α]] for each α. Now, given u ∈ A and α ∈ R, [[u > α]] ⊆ 1 = [[u0 > α]] if α < 0, ⊆

inf [[u > i]] ⊆ inf ci = [[u0 > α]] if n ∈ N, α ∈ [n, n + 1[ .

i≤n

i≤n

Thus u0 is an upper bound for A in L0 . (b)(i) Suppose that cα = supu∈A [[u > α]] is defined in A for every α, and that inf n∈N cn = 0. Then, for any α, supq∈Q,q>α cq = supu∈A,q∈Q,q>α [[u > q]] = supu∈A [[u > q]] = cα . Also, we are supposing that A contains some u0 , so that supn∈N c−n ⊇ supn∈N [[u0 > −n]] = 1. Accordingly there is a u∗ ∈ L0 such that [[u∗ > α]] = cα for every α ∈ R. But now, for v ∈ L0 ,

322

Function spaces

364M

v is an upper bound for A ⇐⇒ [[u > α]] ⊆ [[v > α]] for every u ∈ A, α ∈ R ⇐⇒ [[u∗ > α]] ⊆ [[v > α]] for every α ⇐⇒ u∗ ≤ v, so that u∗ = sup A in L0 . (ii) Now suppose that u∗ = sup A is defined in L0 . Of course [[u∗ > α]] must be an upper bound for {[[u > α]] : u ∈ A} for every α. ?? Suppose we have an α for which it is not the least upper bound, that is, there is a c ⊂ [[u∗ > α]] which is an upper bound for {[[u > α]] : u ∈ A}. Define φ : R → A by setting φ(β) = c ∩ [[u∗ > β]] if β ≥ α, = [[u∗ > β]] if β < α. It is easy to see that φ satisfies the conditions of 364A (we need the distributive law 313Ba to check that φ(β) = supγ>β φ(γ) if β ≥ α), so corresponds to a member v of L0 . But we now find that v is an upper bound for A (because if u ∈ A, β ≥ α then [[u > β]] ⊆ [[u > α]] ∩ [[u∗ > β]] ⊆ c ∩ [[u∗ > β]] = [[v > β]],) that v ≤ u∗ and that v 6= u∗ (because [[v > α]] = c 6= [[u∗ > α]]); but this is impossible, because u∗ is supposed to be the supremum of A. X X Thus if u∗ = sup A is defined in L0 , then supu∈A [[u > α]] is defined in A for every α ∈ R. Also, of course, inf n∈N supu∈A [[u > n]] = inf n∈N [[u∗ > n]] = 0. (c) This is now easy. If A has a supremum, then surely it satisfies the condition, by (b). If A satisfies the condition, then we have a family hcα iα∈R as required in (b); but also, by (a) or otherwise, there is a sequence hc0n in∈N such that cn ⊆ c0n for every n and inf n∈N c0n = 0, so inf n∈N cn is also 0, and both conditions in (b) are satisfied, so A has a supremum. 364N We do not have such a simple formula for general infima (though see 364Xl), but the following facts are useful. Proposition Let A be a Dedekind σ-complete Boolean algebra. (a) If u, v ∈ L0 = L0 (A), then [[u ∧ v > α]] = [[u > α]] ∩ [[v > α]] for every α ∈ R. (b) If A is a non-empty subset of (L0 )+ , then inf A = 0 in L0 iff inf u∈A [[u > α]] = 0 in A for every α > 0. proof (a) Take Z, L0 and π as in the proof of 364E. Express u as f • , v as g • where f , g ∈ L0 , so that u ∧ v = (f ∧ g)• , because the canonical map from L0 to L0 is a Riesz homomorphism (351J). Then [[u ∧ v > α]] = π{z : min(f (z), g(z)) > α} = π({z : f (z) > α} ∩ {z : g(z) > α}) = π{z : f (z) > α} ∩ π{z : g(z) > α} = [[u > α]] ∩ [[v > α]] for every α. (b)(i) If inf u∈A [[u > α]] = 0 for every α > 0, and v is any lower bound for A, then [[v > α]] must be 0 for every α > 0, so that [[v > 0]] = 0; now since [[0 > α]] = 0 for α ≥ 0, 1 for α < 0, v ≤ 0. As v is arbitrary, inf A = 0. (ii) If α > 0 is such that inf u∈A [[u > α]] is undefined, or not equal to 0, let c ∈ A be such that β 0 6= c ⊆ [[u > α]] for every u ∈ A, and consider v = αχc. Then [[v > β]] = [[χc > α ]] is 1 if β < 0, c if 0 ≤ β < α and 0 if β ≥ α. If u ∈ A then [[u > β]] is 1 if β < 0 (since u ≥ 0), at least [[u > α]] ⊇ c if 0 ≤ β < α, and always includes 0; so that v ≤ u. As u is arbitrary, inf A is either undefined in L0 or not 0. 364O

Now we have a reward for our labour, in that the following basic theorem is easy.

Theorem For a Dedekind σ-complete Boolean algebra A, L0 = L0 (A) is Dedekind complete iff A is. proof The description of suprema in 364Mc makes it obvious that if A is Dedekind complete, so that supu∈A [[u > α]] is always defined, then L0 must be Dedekind complete. On the other hand, if L0 is Dedekind complete, then so is L∞ (A) (by 364K and 353J(b-i)), so that A is also Dedekind complete, by 363Mb.

L0

364R

323

364P The multiplication of L0 I have already observed that L0 is always an f -algebra with identity; in particular (because L0 is surely Archimedean) the map u 7→ u × v is order-continuous for every v ≥ 0 (353Oa), and multiplication is commutative (353Ob, or otherwise). The multiplicative identity is χ1 (364E, 364Kc). By 353Pb, or otherwise, u × v = 0 iff |u| ∧ |v| = 0. There is one special feature of multiplication in L0 which I can mention here. Proposition Let A be a Dedekind σ-complete Boolean algebra. Then an element u of L0 has a multiplicative inverse in L0 iff |u| is a weak order unit in L0 iff [[|u| > 0]] = 1. proof If u is invertible, then |u| is a weak order unit, by 353Pc or otherwise. In this case, setting c = 1 \ [[|u| > 0]], we see that [[|u| ∧ χc > 0]] = [[|u| > 0]] ∩ c = 0 (364Na), so that |u| ∧ χc ≤ 0 and χc = 0, that is, c = 0; so [[|u| > 0]] must be 1. To complete the circuit, suppose that [[|u| > 0]] = 1. Let Z, Σ, L0 , π, M be as in the proof of 364E, and S : L0 → L0 the canonical map, so that [[Sh > α]] = π{z : h(z) > α} for every h ∈ L0 , α ∈ R. Express u as Sf where f ∈ L0 . Then π{z : |f (z)| > 0} = [[Sf > 0]] = 1, so {z : f (z) = 0} ∈ M. Set g(z) =

1 f (z)

if f (z) 6= 0,

g(z) = 0 if f (z) = 0.

Then {z : f (z)g(z) 6= 1} ∈ M so u × Sg = S(f × g) = S(χZ) = χ1 and u is invertible. Remark The repeated phrase ‘by 353x or otherwise’ reflects the fact that the abstract methods there can all be replaced in this case by simple direct arguments based on the construction in 364C-364E. 364Q Recovering the algebra: Proposition Let A be a Dedekind σ-complete Boolean algebra. For a ∈ A write Va for the band of L0 = L0 (A) generated by χa. Then a 7→ Va is a Boolean isomorphism between A and the algebra of projection bands of L0 (A). proof I copy from the argument for 363J, itself based on 361K. If a ∈ A, w ∈ L0 then w × χa ∈ Va . P P If v ∈ Va⊥ then |χa| ∧ |v| = 0, so χa × v = 0, so (w × χa) × v = 0, so |w × χa| ∧ |v| = 0; thus w × χa ∈ Va⊥⊥ , which is equal to Va because L0 is Archimedean (353B). Q Q Now, if a ∈ A, u ∈ Va and v ∈ V1\a , then 0 |u| ∧ |v| = 0 because χa ∧ χ(1 \ a) = 0; and if w ∈ L (A) then w = (w × χa) + (w × χ(1 \ a)) ∈ V1 + V1\a . So Va and V1\a are complementary projection bands in L0 . Next, if U ⊆ L0 is a projection band, then χ1 is expressible as u + v = u ∨ v where u ∈ U , v ∈ U ⊥ . Setting a = [[u > 12 ]], a0 = [[v > 21 ]] we must have a ∪ a0 = 1, a ∩ a0 = 0 (using 364M and 364Na), so that a0 = 1 \ a; also 21 χa ≤ u, so that χa ∈ U , and similarly χ(1 \ a) ∈ U ⊥ . In this case Va ⊆ U and V1\a ⊆ U ⊥ , so U must be Va precisely. Thus a 7→ Va is surjective. Finally, just as in 361K, a ⊆ b ⇐⇒ Va ⊆ Vb , so we have a Boolean isomorphism. 364R

I come at last to the result corresponding to 361J and 363F.

Theorem Let A and B be Boolean algebras, and π : A → B a sequentially order-continuous Boolean homomorphism. (a) We have a multiplicative sequentially order-continuous Riesz homomorphism Tπ : L0 (A) → L0 (B) defined by the formula [[Tπ u > α]] = π[[u > α]] for every α ∈ R, u ∈ L0 (A). (b) Defining χa ∈ L0 (A) as in 364K, Tπ (χa) = χ(πa) in L0 (B) for every a ∈ A. If we regard L∞ (A) and ∞ L (B) as embedded in L0 (A) and L0 (B) respectively, then Tπ , as defined here, agrees on L∞ (A) with Tπ as defined in 363F. (c) Tπ is order-continuous iff π is order-continuous, injective iff π is injective, surjective iff π is surjective.

324

Function spaces

364R

¯ π = Tπ h ¯ for (d) [[Tπ u ∈ E]] = π[[u ∈ E]] for every u ∈ L0 (A) and every Borel set E ⊆ R; consequently hT 0 ¯ every Borel measurable h : R → R, writing h indifferently for the associated maps from L (A) to itself and from L0 (B) to itself (364I). (e) If C is another Dedekind σ-complete Boolean algebra and θ : B → C another sequentially ordercontinuous Boolean homomorphism then Tθπ = Tθ Tπ : L0 (A) → L0 (C). proof I write T for Tπ . (a)(i) To see that T u is well-defined in L0 (B) for every u ∈ L0 (A), all we need to do is to check that the map α 7→ π[[u > α]] : R → B satisfies the conditions of 364Bd, and this is easy, because π preserves all countable suprema and infima. (ii) To see that T is linear and order-preserving and multiplicative, we can use the formulae of 364E. For instance, if u, v ∈ L0 (A), then [[T u + T v > α]] = sup [[T u > q]] ∩ [[T v > α − q]] = sup π[[u > q]] ∩ π[[v > α − q]] q∈Q

q∈Q

= π(sup [[u > q]] ∩ [[v > α − q]]) = π[[u + v > α]] = [[T (u + v) > α]] q∈Q

for every α ∈ R, so that T u + T v = T (u + v). In the same way, T (γu) = γT u whenever γ > 0, T u ≤ T v whenever u ≤ v, T u × T v = T (u × v) whenever u, v ≥ 0, so that, using the distributive laws, T is linear and multiplicative. To see that T is a sequentially order-continuous Riesz homomorphism, suppose that A ⊆ L0 (A) is a countable non-empty set with a supremum u∗ ∈ L0 (A); then T [A] is a non-empty subset of L0 (B) with an upper bound T u∗ , and

sup [[T u > α]] = sup π[[u > α]] = π(sup [[u > α]]) = π[[u∗ > α]] u∈A

u∈A

u∈A

(using 364M) = [[T u∗ > α]] for every α ∈ R. So using 364M again, T u∗ = supu∈A T u. Now this is true, in particular, for doubleton sets A, so that T is a Riesz homomorphism; and also for non-decreasing sequences, so that T is sequentially order-continuous. (b) The identification of T (χa) with χ(πa) is another almost trivial verification. It follows that T agrees with the map of 363F on S(A), if we think of S(A) as a subspace of L0 (A). Next, if u ∈ L∞ (A) ⊆ L0 (A), and γ = kuk∞ , then |u| ≤ γχ1A , so that |T u| ≤ γχ1B , and T u ∈ L∞ (B), with kT uk∞ ≤ γ. Thus T ¹L∞ (A) has norm at most 1. As it agrees with the map of 363F on S(A), which is k k∞ -dense in L∞ (A) (363C), and both are continuous, they must agree on the whole of L∞ (A). α) Suppose that π is order-continuous, and that A ⊆ L0 (A) is a non-empty set with a supremum (c)(i)(α 0 u ∈ L (A). Then for any α ∈ R, ∗

[[T u∗ > α]] = π[[u∗ > α]] = π(sup [[u > α]]) u∈A

(by 364M) = sup π[[u > α]] u∈A

(because π is order-continuous)

L0

364S

325

= sup [[T u > α]]. u∈A ∗

As α is arbitrary, T u = sup T [A], by 364M again. As A is arbitrary, T is order-continuous (351Ga). β ) Now suppose that T is order-continuous and that A ⊆ A is a non-empty set with supremum c (β in A. Then χc = supa∈A χa (364Kc) so χ(πc) = T (χc) = supa∈A T (χa) = supa∈A χ(πa). But now πc = [[χ(πc) > 0]] = supa∈A [[χ(πa) > 0]] = supa∈A πa. As A is arbitrary, π is order-continuous. α) If π is injective and u, v are distinct elements of L0 (A), then there must be some α such that (ii)(α [[u > α]] 6= [[v > α]], in which case [[T u > α]] 6= [[T v > α]] and T u 6= T v. β ) Now suppose that T is injective. It is easy to see that χ : A → L0 (A) is injective, so that (β T χ : A → L0 (B) is injective; but this is the same as χπ (by (b)), so π must also be injective. α) Suppose that π is surjective. Let Σ be a σ-algebra of sets such that there is a sequentially (iii)(α order-continuous Boolean surjection φ : Σ → A. Then πφ : Σ → B is surjective. So given w ∈ L0 (B), there is an f ∈ L0 (Σ) such that [[w > α]] = πφ{x : f (x) > α} for every α ∈ R (364D). But, also by 364D, there is a u ∈ L0 (A) such that [[u > α]] = φ{x : f (x) > α} for every α. And now of course T u = w. As w is arbitrary, T is surjective. β ) If T is surjective, and b ∈ B, there must be some u ∈ L0 (A) such that T u = χb. Now set (β a = [[u > 0]] and see that πa = [[χb > 0]] = b. As b is arbitrary, π is surjective. (d) The map E 7→ π[[u ∈ E]] is a sequentially order-continuous Boolean homomorphism, equal to [[T u ∈ E]] when E is of the form ]α, ∞[; so by 364G the two are equal for all Borel sets E. If h : R → R is a Borel measurable function, u ∈ L0 (A) and E ⊆ R is a Borel set, then ¯ u) ∈ E]] = [[T u ∈ h−1 [E]]] = π[[u ∈ h−1 [E]]] [[h(T ¯ ¯ = π[[h(u) ∈ E]] = [[T (h(u)) ∈ E]]. ¯ = hT ¯ . As E and u are arbitrary, T h (e) This is immediate from the definitions. 364S Products: Proposition Let hAi ii∈I be a family of Dedekind σ-complete Boolean algebras, with simple product A. If πi : A → Ai is the coordinate map for each Ti : L0 (A) → L0 (Ai ) the Q i, and 0 0 corresponding homomorphism, then u 7→ T u = hTi uii∈I : L (A) → i∈I L (Ai ) is a multiplicative Riesz Q space isomorphism, so L0 (A) may be identified with the f -algebra product i∈I L0 (Ai ) (352Wc). proof Because each πi is a surjective order-continuous Boolean homomorphism, 364R assures us that there are corresponding surjective multiplicative Riesz Q homomorphisms Ti . So all we need to check is that the multiplicative Riesz homomorphism T : L0 (A) → i∈I L0 (Ai ) is a bijection. If u, v ∈ L0 (A) are distinct, there must be some α ∈ R such that [[T u > α]] 6= [[T v > α]]. In this case there must be an i ∈ I such that πi [[T u > α]] 6= πi [[T v > α]], that is, [[Ti u > α]] 6= [[Ti v > α]]. So Ti u 6= Ti v and T u 6= T v. As u, v are arbitrary,QT is injective. If w = hwi ii∈I is any member of i∈I L0 (Ai ), then for α ∈ R set φ(α) = h[[wi > α]]ii∈I ∈ A. It is easy to check that φ satisfies the conditions of 364A, because, for instance, supβ>α πi φ(β) = supβ>α [[wi > β]] = [[wi > α]] = πi φ(α) for every i, so that supβ>α φ(β) = φ(α), for every α ∈ R; and the other two conditions are also satisfied because they are satisfied coordinate-by-coordinate. So there is a u ∈ L0 (A) such that φ(α) = [[u > α]] for every α, that is, πi [[u > α]] = [[wi > α]] for all α, i, that is, Ti u = wi for every i, that is, T u = w. As w is arbitrary, T is surjective and we are done.

326

Function spaces

*364T

*364T Regular open algebras I noted in 314P that for every topological space X there is a corresponding Dedekind complete Boolean algebra G of regular open sets. We have an identification of L0 (G) as a space of equivalence classes of functions, different in kind from the representations above, as follows. This is hard work (especially if we do it in full generality), but instructive. I start with a temporary definition. Definition Let (X, T) be a topological space, f : X → R a function. For x ∈ X write ω(f, x) = inf G∈T,x∈G supy,z∈G |f (y) − f (z)| (allowing ∞). *364U Theorem Let X be any topological space, and G its regular open algebra. Let U be the set of functions f : X → R such that {x : ω(f, x) < ²} is dense in X for every ² > 0. Then U is a Riesz subspace of RX , closed under multiplication, and we have a surjective multiplicative Riesz homomorphism T : U → L0 (G) defined by writing [[T f > α]] = supβ>α int {x : f (x) > β}, the supremum being taken in G, for every α ∈ R, f ∈ U . The kernel of T is the set W of functions f : X → R such that int{x : |f (x)| ≤ ²} is dense for every ² > 0, so L0 (G) can be identified, as f -algebra, with the quotient space U/W . α) The first thing to observe is that for any f ∈ RX , ² > 0 the set proof (a)(i)(α {x : ω(f, x) < ²} =

[ {G : G ⊆ X is open and non-empty and sup |f (y) − f (z)| < ²} y,z∈G

is open. β ) Next, it is easy to see that (β ω(f + g, x) ≤ ω(f, x) + ω(g, x), ω(γf, x) = |γ|ω(f, x), ω(|f |, x) ≤ ω(f, x), X

for all f , g ∈ R

and γ ∈ R.

(γγ ) Thirdly, it is useful to know that if f ∈ U and G ⊆ X is a non-empty open set, then there is a non-empty open set G0 ⊆ G on which f is bounded. P P Take any x0 ∈ G such that ω(f, x0 ) < 1; then there is a non-empty open set G0 containing x0 such that |f (y) − f (z)| < 1 for all y, z ∈ G0 , and we may suppose that G0 ⊆ G. But now |f (x)| ≤ 1 + |f (x0 )| for every x ∈ G0 . Q Q (ii) So if f , g ∈ U and γ ∈ R then 1 2

1 2

{x : ω(f + g, x) < ²} ⊇ {x : ω(f, x) < ²} ∩ {x : ω(g, x) < ²} is the intersection of two dense open sets and is therefore dense, while {x : ω(γf, x) < ²} ⊇ {x : ω(f, x) <

² }, 1+|γ|

{x : ω(|f |, x) < ²} ⊇ {x : ω(f, x) < ²} are also dense. As ² is arbitrary, f + g, γf and |f | all belong to U ; as f , g and γ are arbitrary, U is a Riesz subspace of RX . (iii) If f , g ∈ U then f × g ∈ U . P P Take ² > 0 and let G0 be a non-empty open subset of X. By the last remark in (i) above, there is a non-empty open set G1 ⊆ G0 such that |f | ∨ |g| is bounded on G1 ; say max(|f (x)|, |g(x)|) ≤ γ for every x ∈ G1 .

L0

*364U

Set δ = min(1,

² ) 2(1+γ)

327

> 0. Then there is an x ∈ G1 such that ω(f, x) < δ and ω(g, x) < δ. Let H, H 0

be open sets containing x such that |f (y) − f (z)| < δ for all y, z ∈ H and |g(y) − g(z)| < δ for all y, z ∈ H 0 . Consider G = G1 ∩ H ∩ H 0 . This is an open set containing x, and if y, z ∈ G then |f (y)g(y) − f (z)g(z)| ≤ |f (y) − f (z)||g(y) − g(z)| + |f (y) − f (z)||g(z)| + |f (z)||g(y) − g(z)| 2

≤ δ + δγ + γδ. Accordingly ω(f × g, x) ≤ δ(1 + 2γ) < ², while x ∈ G0 . As G0 is arbitrary, {x : ω(f × g, x) < ²} is dense; as ² is arbitrary, f × g ∈ U . Q Q Thus U is a subalgebra of RX . (b) Now, for f ∈ U , consider the map φf : R → G defined by setting φf (α) = supβ>α int {x : f (x) > β} for every α ∈ R. Then φf satisfies the conditions of 364A. P P (See 314P for the calculation of suprema and infima in G.) (i) If α ∈ R then φf (α) = sup int {x : f (x) > β} = sup int {x : f (x) > γ} β>α

γ>β>α

= sup sup int {x : f (x) > γ} = sup φf (α). β>α γ>β

β>α

(ii) If G0 ⊆ X is a non-empty open set, then there is a non-empty open set G1 ⊆ G0 such that f is bounded on G1 ; say |f (x)| < γ for every x ∈ G1 . If β > γ then G1 does not meet {x : f (x) > β}, so G1 ∩ int {x : f (x) > γ} = ∅; as β is arbitrary, G1 ∩ φf (γ) = ∅ and G0 6⊆ inf α∈R φf (α). On the other hand, G1 ⊆ {x : f (x) > −γ}, so G1 ⊆ int {x : f (x) > −γ} ⊆ φf (−γ) and G0 ∩ supα∈R φf (α) 6= ∅. As G0 is arbitrary, inf α∈R φf (α) = ∅ and supα∈R φf (α) = X. Q Q (c) Thus we have a map T : U → L0 = L0 (G) defined by setting [[T f > α]] = φf (α) for every α ∈ R, f ∈ U. It is worth noting that {x : f (x) > α + ω(f, x)} ⊆ [[T f > α]] ⊆ {x : f (x) + ω(f, x) ≥ α} for every f ∈ U , α ∈ R. P P (i) If f (x) > α + ω(f, x), set δ = 12 (f (x) − α − ω(f, x)) > 0. Then there is an open set G containing x such that |f (y) − f (z)| < ω(f, x) + δ for every y, z ∈ G, so that f (y) > α + δ for every y ∈ G, and x ∈ int{y : f (y) > α + δ} ⊆ [[T f > α]]. (ii) If f (x) + ω(f, x) < α, set δ = 12 (α − f (x) − ω(f, x)) > 0; then there is an open neighbourhood G of x such that |f (y) − f (z)| < ω(f, x) + δ for every y, z ∈ G, so that f (y) < α for every y ∈ G. Accordingly G / [[T f > α]]. Q Q does not meet {y : f (y) > β} nor {y : f (y) > β} for any β > α, G ∩ [[T f > α]] = ∅ and x ∈ (d) T is additive. P P Let f , g ∈ U and α < β ∈ R. Set δ = 15 (β − α) > 0, H = {x : ω(f, x) < δ, ω(g, x) < δ}; then H is the intersection of two dense open sets, so is itself dense and open. (i) If x ∈ H ∩ [[T (f + g) > β]], then (f + g)(x) + ω(f + g, x) ≥ β; but ω(f + g, x) ≤ 2δ (see (a-i-β) above), so f (x) + g(x) ≥ β − 2δ > α + 2δ and there is a q ∈ Q such that f (x) > q + δ ≥ q + ω(f, x),

g(x) > α − q + δ ≥ α − q + ω(g, x).

Accordingly x ∈ [[T f > q]] ∩ [[T g > α − q]] ⊆ [[T f + T g > α]].

328

Function spaces

*364U

Thus H ∩ [[T (f + g) > β]] ⊆ [[T f + T g > α]]. Because H is dense, [[T (f + g) > β]] ⊆ [[T f + T g > α]]. (ii) If x ∈ H, then

x∈

[

([[T f > q]] ∩ [[T g > β − q]])

q∈Q

=⇒ ∃ q ∈ Q, f (x) + ω(f, x) ≥ q, g(x) + ω(g, x) ≥ β − q =⇒ f (x) + g(x) + 2δ ≥ β =⇒ (f + g)(x) ≥ α + 3δ > α + ω(f + g, x) =⇒ x ∈ [[T (f + g) > α]]. Thus

S H ∩ q∈Q ([[T f > q]] ∩ [[T g > β − q]]) ⊆ [[T (f + g) > α]]. S Because H is dense and q∈Q ([[T f > q]] ∩ [[T g > β − q]]) is open, [

[[T f + T g > β]] = int

[[T f > q]] ∩ [[T g > β − q]]

q∈Q

⊆ int [[T (f + g) > α]] = [[T (f + g) > α]]. (iii) Now let β ↓ α; we have [[T (f + g) > α]] = sup [[T (f + g) > β]] ⊆ [[T f + T g > α]] β>α

= sup [[T f + T g > β]] ⊆ [[T (f + g) > α]], β>α

so [[T (f + g) > α]] = [[T f + T g > α]]. As α is arbitrary, T (f + g) = T f + T g; as f and g are arbitrary, T is additive. Q Q (e) It is now easy to see that T is linear. P P If γ > 0, f ∈ U , α ∈ R then β γ

[[T (γf ) > α]] = sup int {x : γf (x) > β} = sup int {x : f (x) > } β>α

β>α

α γ

= sup int {x : f (x) > β} = [[T f > ]] = [[γT f > α]]. β>α/γ

As α is arbitrary, T (γf ) = γT f ; because we already know that T is additive, this is enough to show that T is linear. Q Q (f ) In fact T is a Riesz homomorphism. P P If f ∈ U , α ≥ 0 then [[T (f + ) > α]] = sup int {x : f + (x) > β} = sup int {x : f (x) > β} β>α

β>α

= [[T f > α]] = [[(T f )+ > α]]. If α < 0 then [[T (f + ) > α]] = sup int {x : f + (x) > β} β>α

= X = [[(T f )+ > α]]. Q Q (g) Of course the constant function χX belongs to U , and is its multiplicative identity; and T (χX) is the multiplicative identity of L0 , because

L0

*364U

329

[[T (χX) > α]] = sup int {x : (χX)(x) > β} β>α

= X if α < 1, ∅ if α ≥ 1. By 353Pd, or otherwise, T is multiplicative. (h) The kernel of T is W . P P (i) For f ∈ U , T f = 0 =⇒ [[T |f | > 0]] = [[|T f | > 0]] = ∅ =⇒ {x : |f (x)| > ω(|f |, x)} = ∅ =⇒ int{x : |f (x)| ≤ ²} ⊇ {x : ω(|f |, x) < ²} is dense, for every ² > 0 =⇒ f ∈ W. (ii) If f ∈ W , then, first, 1 3

{x : ω(f, x) < ²} ⊇ int{x : |f (x)| ≤ ²} is dense for every ² > 0, so f ∈ U ; and next, for any β > 0, {x : |f (x)| > β} does not meet the dense open set int{x : |f (x)| ≤ β}, so [[|T f | > 0]] = [[T |f | > 0]] = supβ>0 int {x : |f (x)| > β} = ∅ and T f = 0. Q Q (i) Finally, T is surjective. P P Take any u ∈ L0 . Define f˜ : X → [−∞, ∞] by setting f˜(x) = sup{α : x ∈ [[u > α]]} for each x, counting inf ∅ as −∞. Then S {x : f˜(x) > α} = β>α [[u > β]] is open, for every α ∈ R. The set {x : f˜(x) = ∞} =

T α∈R

[[u > α]]

is nowhere dense, because inf α∈R [[u > α]] = ∅ in G; while

S {x : f˜(x) = −∞} = X \ α∈R [[u > α]]

is also nowhere dense, because supα∈R [[u > α]] = X in G. Accordingly E = int{x : f˜(x) ∈ R} is dense. Set f (x) = f˜(x) for x ∈ E, 0 for x ∈ X \ E. Let ² > 0. If G ⊆ X is a non-empty open set, there is an α ∈ R such that G 6⊆ [[u > α]], so G1 = G \ [[u > α]] 6= ∅, and f˜(x) ≤ α for every x ∈ G1 . Set α0 = supx∈G1 f˜(x) ≤ α < ∞. Because E meets G1 , α0 > −∞. Then G2 = G1 ∩ [[u > α0 − 21 ²]] is a non-empty open subset of G and α0 − 21 ² ≤ f˜(x) ≤ α0 for every x ∈ G2 . Accordingly |f (y) − f (z)| ≤ 12 ² for all y, z ∈ G2 , and ω(f, x) < ² for all x ∈ G2 . As G is arbitrary, {x : ω(f, x) < ²} is dense; as ² is arbitrary, f ∈ U . Take α < β in R, and set δ = 21 (β − α). Then H = E ∩ {x : ω(f, x) < δ} is a dense open set, and H ∩ [[T f > β]] ⊆ H ∩ {x : f (x) + ω(f, x) ≥ β} ⊆ E ∩ {x : f (x) > α} ⊆ {x : f˜(x) > α} ⊆ [[u > α]]. As H is dense, [[T f > β]] ⊆ [[u > α]]. In the other direction H ∩ [[u > β]] ⊆ H ∩ {x : f˜(x) ≥ β} = H ∩ {x : f (x) ≥ β} ⊆ {x : f (x) > α + ω(f, x)} ⊆ [[T f > α]], so [[u > β]] ⊆ [[T f > α]]. Just as in (d) above, this is enough to show that T f = u. As u is arbitrary, T is surjective. Q Q This completes the proof.

330

Function spaces

*364V

*364V Compact spaces Suppose now that X is a compact Hausdorff topological space. In this case the space U of 364U is just the space of functions f : X → R such that {x : f is continuous at x} is dense in X. P P It is easy to see that T {x : f is continuous at x} = {x : ω(f, x) = 0} = n∈N Hn where Hn = {x : ω(f, x) < 2−n T } for each n. Each Hn is an open set (see part (a-i-α) of the proof of 364U), so by Baire’s theorem (3A3G) n∈N Hn is dense iff every Hn is dense, that is, iff f ∈ U . Q Q Now W becomes {f : f ∈ U, {x : f (x) = 0} is dense}. P P (i) If f ∈ W , then T |f | = 0, so (by the formula in (c) of the proof of 364U) |f (x)| ≤ ω(|f |, x) for every x. But {x : ω(f, x) = 0} is dense, because f ∈ U , so {x : f (x) = 0} is also dense. (ii) If f ∈ U and {x : f (x) = 0} is dense, then ω(f, x) ≥ inf x∈G is open supy∈G |f (y) − f (x)| ≥ |f (x)| for every x ∈ X. So for any ² > 0, int{x : |f (x)| ≤ ²} ⊇ {x : ω(f, x) < ²} is dense, and f ∈ W . Q Q In the case of extremally disconnected spaces, we can go farther. *364W Theorem Let X be a compact Hausdorff extremally disconnected space, and G its regular open algebra. Write C ∞ = C ∞ (X) for the space of continuous functions g : X → [−∞, ∞] such that {x : g(x) = ±∞} is nowhere dense. Then we have a bijection S : C ∞ → L0 = L0 (G) defined by saying that [[Sg > α]] = {x : g(x) > α} ˙ × ˙ on C ∞ defined by for every α ∈ R. Addition and multiplication in L0 correspond to the operations +, ∞ ˙ g ×h ˙ are the unique elements of C agreeing with g + h, g × h on {x : g(x), h(x) are both saying that g +h, finite}. Scalar multiplication in L0 corresponds to the operation (γg)(x) = γg(x) for x ∈ X, g ∈ C ∞ , γ ∈ R on C ∞ (counting 0 · ∞ as 0), while the ordering of L0 corresponds to the relation g ≤ h ⇐⇒ g(x) ≤ h(x) for every x ∈ X. proof (a) For g ∈ C ∞ , set Hg = {x : g(x) ∈ R}, so that Hg is a dense open set, and define Rg : X → R by setting (Rg)(x) = g(x) if x ∈ Hg , 0 if x ∈ X \ Hg . Then Rg is continuous at every point of Hg , so belongs to the space U of 364U-364V. Set Sg = T (Rg), where T : U → L0 is the map of 364U. Then [[Sg > α]] = {x : g(x) > α} for every α ∈ R. P P (i) ω(g, x) = 0 for every x ∈ Hg , so, if β > α, Hg ∩ [[Sg > β]] ⊆ {x : x ∈ Hg , (Rg)(x) ≥ β} ⊆ {x : g(x) ≥ β} by the formula in part (c) of the proof of 364U. As [[Sg > β]] is open and Hg is dense, [[Sg > β]] ⊆ Hg ∩ [[Sg > β]] ⊆ {x : g(x) ≥ β} ⊆ {x : g(x) > α}. Now [[Sg > α]] = supβ>α [[Sg > β]] = int

S β>α

[[Sg > β]] ⊆ {x : g(x) > α}.

(ii) In the other direction, Hg ∩ {x : g(x) > α} ⊆ [[Sg > α]], by the other half of the formula in the proof of 364U. Again because {x : g(x) > α} is open and Hg is dense, {x : g(x) > α} ⊆ [[Sg > α]] = [[Sg > α]] because X is extremally disconnected (see 314S). Q Q (b) Thus we have a function S defined by the formula offered. Now if g, h ∈ C ∞ and g ≤ h, we surely have {x : g(x) > α} ⊆ {x : h(x) > α} for every α, so [[Sg > α]] ⊆ [[Sh > α]] for every α and Sg ≤ Sh. On the other hand, if g 6≤ h then Sg 6≤ Sh. P P Take x0 such that g(x0 ) > h(x0 ), and α ∈ R such that g(x0 ) > α > h(x0 ); set H = {x : g(x) > α > h(x)}; this is a non-empty open set and H ⊆ [[Sg > α]]. On the other hand, H ∩ {x : h(x) > α} = ∅ so H ∩ [[Sh > α]] = ∅. Thus [[Sg > α]] 6⊆ [[Sh > α]] and Sg 6≤ Sh. Q Q In particular, S is injective. (c) S is surjective. P P If u ∈ L0 , set

L0

364Xe

331

g(x) = sup{α : x ∈ [[u > α]]} ∈ [−∞, ∞] for every x ∈ X, taking sup ∅ = −∞. Then, for any α ∈ R, {x : g(x) > α} = other hand, S {x : g(x) < α} = β<α {x : x ∈ / [[u > β]]}

S β>α

[[u > α]] is open. On the

is also open, because all the sets [[u > β]] are open-and-closed. So g : X → [−∞, ∞] is continuous. Also S {x : g(x) > −∞} = α∈R [[u > α]], S {x : g(x) < ∞} = α∈R X \ [[u > α]] are dense, so g ∈ C ∞ . Now, for any α ∈ R, [[Sg > α]] = {x : g(x) > α} = = int

[

[

[[u > β]]

β>α

[[u > β]] = sup [[u > β]] = [[u > α]].

β>α

β>α

So Sg = u. As u is arbitrary, S is surjective. Q Q (d) Accordingly S is a bijection. I have already checked (in part (b)) that it is an isomorphism of the order structures. For the algebraic operations, observe that if g, h ∈ C ∞ then there are f1 , f2 ∈ C ∞ such that Sg + Sh = Sf1 and Sg × Sh = Sf2 , that is, T (Rg + Rh) = T Rg + T Rh = T Rf1 ,

T (Rg × Rh) = T Rg × T Rh = T Rf2 .

But this means that T (Rg + Rh − Rf1 ) = T ((Rg × Rh) − Rf2 ) = 0, so that Rg + Rh − Rf1 , (Rg × Rh) − Rf2 belong to W and are zero on dense sets (364V). Since we know also that the set G = {x : g(x), h(x) are both finite} is a dense open set, while g, h, f1 and f2 are all continuous, we must have f1 (x) = g(x) + h(x), f2 (x) = g(x)h(x) for every x ∈ G. And of course this uniquely specifies f1 and f2 as of C ∞ . ˙ × ˙ as described, rendering S additive and multiplicative. As for scalar Thus we do have operations +, multiplication, it is easy to check that R(γg) = γRg (at least, unless γ = 0, which is trivial), so that S(γg) = γSg for every g ∈ C ∞ . 364X Basic exercises (a) Let A be a Dedekind σ-complete Boolean algebra. For u, v ∈ L0 = L0 (A) set [[u < v]] = [[v > u]] = [[v − u > 0]], [[u ≤ v]] = [[v ≥ u]] = 1 \ [[v < u]], [[u = v]] = [[u ≤ v]] ∩ [[v ≤ u]]. (i) Show that ([[u < v]], [[u = v]], [[u > v]]) is always a partition of unity in A. (ii) Show that for any u, u0 , v, v 0 ∈ L0 , [[u ≤ u0 ]] ∩ [[v ≤ v 0 ]] ⊆ [[u + v ≤ u0 + v 0 ]] and [[u = u0 ]] ∩ [[v = v 0 ]] ⊆ [[u × v = u0 × v 0 ]]. (b) Let A be a Dedekind σ-complete Boolean algebra. (i) Show that if u, v ∈ L0 = L0 (A) and α, β ∈ R then [[u + v ≥ α + β]] ⊆ [[u ≥ α]] ∪ [[v ≥ β]]. (ii) Show that if u, v ∈ (L0 )+ and α, β ≥ 0 then [[u × v ≥ αβ]] ⊆ [[u ≥ α]] ∪ [[v ≥ β]]. (c) Let A be a Dedekind σ-complete Boolean algebra and u ∈ L0 (A). Show that {[[u ∈ E]] : E ⊆ R is Borel} is the σ-subalgebra of A generated by {[[u > α]] : α ∈ R}. >(d) Let (A, µ ¯) be a probability algebra. Show that for any u ∈ L0 (A) there is a unique Radon probability measure ν on R (the distribution of u) such that νE = µ ¯[[u ∈ E]] for every Borel set E ⊆ R. (Hint: 271B.) (e) Let (A, µ ¯) be a probability algebra, and hui ii∈I any family in L0 (A); for each i ∈ I let Bi be the closed subalgebra of AQgenerated by {[[ui > α]] : α ∈ R}. Show that the following are equiveridical: (i) µ ¯(inf i∈J [[ui > αi ]]) = i∈J µ ¯[[ui > αi ]] whenever J ⊆ I is finite and αi ∈ R for each i ∈ J (ii) hBi ii∈I is independent in the sense of 325L. (In this case we may call hui ii∈I (stochastically) independent.)

332

Function spaces

364Xf

(f ) Let (A, µ ¯) be a probability algebra and u, v two stochastically independent of L0 (A). Show that the distribution of their sum is the convolution of their distributions. (Hint: 272S.) > (g) Let A be a Dedekind σ-complete Boolean algebra and g, h : R → R Borel measurable functions. (i) ¯ = gh, ¯ where g¯, h ¯ : L0 → L0 are defined as in 364I. (ii) Show that g + h(u) = g¯(u) + h(u), ¯ Show that g¯h ¯ g × h(u) = g¯(u) × h(u) for every u ∈ L0 = L0 (A). (iii) Show that if hhn in∈N is a sequence of Borel ¯ n (u) = h(u) ¯ measurable functions on R and supn∈N hn = h, then supn∈N h for every u ∈ L0 . (iv) Show that ¯ ¯ if h is non-decreasing and continuous on the left, then h(sup A) = sup h[A] whenever A ⊆ L0 is a non-empty set with a supremum in L0 . (h) Let A be a Dedekind σ-complete Boolean algebra. (i) Show that S(A) can be identified (α) with the set of those u ∈ L0 = L0 (A) such that {[[u > α]] : α ∈ R} is finite (β) with the set of those u ∈ L0 such that [[u ∈ I]] = 1 for some finite I ⊆ R. (ii) Show that L∞ (A) can be identified with the set of those u ∈ L0 such that [[u ∈ [−α, α]]] = 1 for some α ≥ 0, and that kuk∞ is the smallest such α. (i) Show that if A is a Dedekind σ-complete Boolean algebra, and u ∈ L0 (A), then for any α ∈ R [[u > α]] = inf β>α sup{a : a ∈ A, u × χa ≥ βχa} (compare 363Xh). (j) Let A be a Dedekind σ-complete Boolean algebra and u ≥ 0 in L0 = L0 (A). Show that u = supq∈Q qχ[[u > q]] in L0 . (k) (i) Let A be a Dedekind σ-complete Boolean algebra and A ⊆ L0 (A) a non-empty countable set with supremum w. Show that [[w ∈ E]] ⊆ supu∈A [[u ∈ E]] for every Borel set E ⊆ R. (ii) Let (A, µ ¯) be a localizable measure algebra and A ⊆ L0 (A) a non-empty set with supremum w. Show that [[w ∈ E]] ⊆ supu∈A [[u ∈ E]] for every Borel set E ⊆ R. (l) Let A be a Dedekind σ-complete Boolean algebra and A ⊆ L0 = L0 (A) a non-empty set which is bounded below in L0 . Suppose that φ0 (α) = inf u∈A [[u > α]] is defined in A for every α ∈ R. Show that v = inf A is defined in L0 , and that [[v > α]] = supβ>α φ0 (β) for every α ∈ R. (m) Let (X, Σ, µ) be a measure space and f : X → [0, ∞[ a function; set A = {g • : g ∈ L0 (µ), g ≤ f a.e.}. (i) Show that if (X, Σ, µ) either is localizable or has the measurable envelope property (213Xl), then sup A is defined in L0 (µ). (ii) Show that if (X, Σ, µ) is complete and locally determined and w = sup A is defined in L0 (µ), then w ∈ A. (n) Let A be a Dedekind σ-complete Boolean algebra. Show that if u, v ∈ L0 = L0 (A) then the following are equiveridical: (α) [[|v| > 0]] ⊆ [[|u| > 0]] (β) v belongs to the band of L0 generated by u (γ) there is a w ∈ L0 such that u × w = v. (o) Let A be a Dedekind σ-complete Boolean algebra and a ∈ A; let Aa be the principal ideal of A generated by a. Show that L0 (Aa ) can be identified, as f -algebra, with the band of L0 (A) generated by χa. (p) Let A and B be Dedekind σ-complete Boolean algebras, and π : A → B a sequentially ordercontinuous Boolean homomorphism. Let T : L0 (A) → L0 (B) be the corresponding Riesz homomorphism (364R). Show that (i) the kernel of T is the sequentially order-closed solid linear subspace of L0 (A) generated by {χa : a ∈ A, πa = 0} (ii) the set of values of T is the sequentially order-closed Riesz subspace of L0 (B) generated by {χ(πa) : a ∈ A}. (q) Let A and B be Dedekind σ-complete Boolean algebras, and π : A → B a sequentially ordercontinuous Boolean homomorphism, with T : L0 (A) → L0 (B) the associated operator. Suppose that h is a ¯ u) = T h(u) ¯ Borel measurable real-valued function defined on a Borel subset of R. Show that h(T whenever ¯ u ∈ L0 (A) and h(u) is defined in the sense of 364I. >(r) Let X and Y be sets, Σ, T σ-algebras of subsets of X, Y respectively, and I, J σ-ideals of Σ, T. Set A = Σ/I, B = T/J . Suppose that φ : X → Y is a function such that φ−1 [F ] ∈ Σ for every F ∈ T,

364Yk

L0

333

φ−1 [F ] ∈ I for every F ∈ J . Show that there is a sequentially order-continuous Boolean homomorphism π : B → A defined by saying that πF • = φ−1 [F ]• for every F ∈ T. Let T : L0 (B) → L0 (A) be the Riesz homomorphism corresponding to π. Show that if we identify L0 (B) with L0T /WJ and L0 (A) with L0Σ /WI in the manner of 364D, then T (g • ) = (gφ)• for every g ∈ L0T . (s) Use the ideas of part (d) of the proof of 364U to show that the operator T there is multiplicative, without appealing to 353P. 364Y Further exercises (a) Show directly, without using the Stone representation, that if A is any Dedekind σ-complete Boolean algebra then the formulae of 364E define a group operation + on L0 (A), and generally an f -algebra structure. (b) Let A be a Dedekind σ-complete Boolean algebra. Show that A is ccc iff L0 (A) has the countable sup property. (c) Let A be a Dedekind σ-complete Boolean algebra and u1 , . . . , un of L0 (A). Write Bn for the algebra of Borel sets in Rn . (i) Show that there is a unique sequentially order-continuous Boolean homomorphism E 7→ [[(u1 , . . . , un ) ∈ E]] : Bn → A such that [[(u1 , . . . , un ) ∈ E]] = inf i≤n [[ui > αi ]] when Q E = i≤n ]αi , ∞[. (ii) Show that for every sequentially order-continuous Boolean homomorphism φ : Bn → A there are u1 , . . . , un ∈ L0 (A) such that φE = [[(u1 , . . . , un ) ∈ E]] for every E ∈ Bn . (d) Let A be a Dedekind σ-complete Boolean algebra, n ≥ 1 and h : Rn → R a Borel measurable ¯ : L0 (A)n → L0 (A) defined by saying that function. Show that we have a corresponding function h −1 ¯ [[h(u1 , . . . , un ) ∈ E]] = [[(u1 , . . . , un ) ∈ h [E]]] for every Borel set E ⊆ R. (e) Suppose that h1 (x, y) = x + y, h2 (x, y) = xy, h3 (x, y) = max(x, y) for all x, y ∈ R. Show that, in the ¯ 1 (u, v) = u + v, h ¯ 2 (u, v) = u × v, h ¯ 3 (u, v) = u ∨ v for all u, v ∈ L0 . language of 364Yd, h (f ) Let A and B be Dedekind σ-complete Boolean algebras, and T : L0 (A) → L0 (B) a Riesz homomorphism such that T e = e0 , where e, e0 are the multiplicative identities of L0 (A), L0 (B) respectively. Show that there is a unique sequentially order-continuous Boolean homomorphism π : A → B such that T = Tπ in the sense of 364R. (Hint: use 353Pd. Compare 375A below.) (g) Suppose, in 364U, that X = Q. (i) Show that there is an f ∈ W such that f (q) > 0 for every q ∈ Q. (ii) Show that there is a u ∈ L0 such that no f ∈ U representing u can be continuous at any point of Q. (h) Let X and Y be topological spaces and φ : X → Y a continuous function such that φ−1 [M ] is nowhere dense in X for every nowhere dense subset M of Y . (Cf. 313R.) (i) Show that we have an order-continuous Boolean homomorphism π from the regular open algebra GY of Y to the regular open algebra GX of X defined by the formula πG = int φ−1 [G] for every G ∈ GY . (ii) Show that if UX , UY are the function spaces of 364U then gφ ∈ UX for every g ∈ UY . (iii) Show that if TX : UX → L0 (GX ), TY : UY → L0 (GY ) are the canonical surjections, and T : L0 (GY ) → L0 (GX ) is the homomorphism corresponding to π, then T (TY g) = TX (gφ) for every g ∈ UY . (iv) Rewrite these ideas for the special case in which X is a dense subset of Y and φ is the identity map, showing that in this case π and T are isomorphisms. (i) Let X be a Baire space, G its algebra of regular open sets, M its ideal of meager sets, and Bb the Baire b property σ-algebra {G4A : G ⊆ X is open, A ∈ M}, so that G can be identified with B/M (314Yd). (i) Repeat the arguments of 364V in this context. (ii) Show that the space U of 364U-364V is a subspace of b and that W = U ∩ W where W = {f : f ∈ RX , {x : f (x) 6= 0} ∈ M}, so that the representations of L0 (B), 0 L (G) as U/W , L0 /W are consistent. (j) Work through the arguments of 364U and 364Yi for the case of compact Hausdorff X, seeking simplifications based on 364V. (k) Let X be an extremally disconnected compact Hausdorff space with regular open algebra G. Let U0 be the space of real-valued functions f : X → R such that int{x : f is continuous at x} is dense. Show that U0 is a Riesz subspace of the space U of 364U, and that every member of L0 (G) is represented by a member of U0 .

334

Function spaces

364Yl

(l) Let X be a Baire space. Let Q be the set of all continuous real-valued functions defined on subsets of X, and Q∗ the set of all of Q which are maximal in the sense that there is no member of Q properly extending them. (i) Show that the domain of any member of Q∗ is a dense Gδ set. (ii) Show that ˙ f ×g, ˙ γ.f we can define addition and multiplication and scalar multiplication on Q∗ by saying that f +g, are to be the unique of Q∗ extending the partially-defined functions f + g, f × g, γf , and that these definitions render Q∗ an f -algebra if we say that f ≤ g iff f (x) ≤ g(x) for every x ∈ dom f ∩ dom g. (iii) Show that every member of Q∗ has an extension to a member of U , as defined in 364U, and that these extensions define an isomorphism between Q∗ and L0 (G), where G is the regular open algebra of X. (iv) Show that if X is compact, Hausdorff and extremally disconnected, then every member of Q∗ has a unique extension to a member of C ∞ (X), as defined in 364W. (m) Let X be an extremally disconnected Hausdorff space, and Z any compact Hausdorff space. Show that if D ⊆ X is dense and f : D → Z is continuous, there is a continuous g : X → Z extending f . (n) Let A be a Dedekind σ-complete Boolean algebra. Write L0C = {u + iv : u, v ∈ L0 } for the complexification of L0 = L0 (A) as defined in 354Yk. (i) Writing B(C) for the Borel σ-algebra of C, show that there is a one-to-one correspondence between sequentially order-continuous Boolean homomorphisms φ : B(C) → A and w = u + iv of L0C defined by saying that [[u > α]] ∩ [[v > β]] = φ{z : Re z > α, Im z > β}. (ii) Show that if Σ is a σ-algebra of subsets of a set X and π : Σ → A is a surjective sequentially ordercontinuous Boolean homomorphism with kernel I, then we can identify L0C with L0C /W, where L0C is the set of Σ-measurable functions from X to C, and W is the set of those f ∈ L0C such that {x : f (x) 6= 0} ∈ I. 364 Notes and comments This has been a long section, and so far all we have is a supposedly thorough grasp of the construction of L0 spaces; discussion of their properties still lies ahead. The difficulties seem to stem from a variety of causes. First, L0 spaces have a rich structure, being linear ordered spaces with multiplications; consequently all the main theorems have to check rather a lot of different aspects. Second, unlike L∞ spaces, they are not accessible by means of the theory of normed spaces, so I must expect to do more of the work here rather than in an appendix. But this is in fact a crucial difference, because it affects the proof of the central theorem 364E. The point is that a given algebra A will be expressible in the form Σ/I for a variety of algebras Σ of sets. Consequently any definition of L0 (A) as a quotient L0 (Σ)/W must include a check that the structure produced is independent of the particular pair Σ, I chosen. The same question arises with S(A) and L∞ (A). But in the case of S, I was able to use a general theory of additive functions on A (see the proof of 361L), while in the case of L∞ I could quote the result for S and a little theory of normed spaces (see the proof of 363H). The theorems of §368 will show, among other things, that a similar approach (describing L0 as a special kind of extension of S or L∞ ) can be made to work in the present situation. I have chosen, however, an alternative route using a novel technique. The price is the time required to develop skill in the technique, and to relate it to the earlier approach (364D, 364E, 364K). The reward is a construction which is based directly on the algebra A, independent of any representation (364A), and methods of dealing with it which are complementary to those of the previous three sections. In particular, they can be used in the absence of the full axiom of choice (364Ya). I have deliberately chosen the notation [[u > α]] from the theory of Forcing. I do not propose to try to explain myself here, but I remark that much of the labour of this section is a necessary basis for understanding real analysis in Boolean-valued models of set theory. The idea is that just as a function f : X → R can be described in of the sets {x : f (x) > α}, so can an element u of L0 (A) be described in of the elements [[u > α]] of A where in some sense u is greater than α. This description is well adapted to discussion of the order struction of L0 (A) (see 364M-364O), but rather ill-adapted to discussion of its linear and multiplicative structures, which leads to a large part of the length of the work above. Once we have succeeded in describing the algebraic operations on L0 in of the values of [[u > α]], however, as in 364E, the fundamental result on the action of Boolean homomorphisms (364R) is elegant and reasonably straightforward. The concept ‘[[u > α]]’ can be dramatically generalized to the concept ‘[[(u1 , . . . , un ) ∈ E]]’, where E is a Borel subset of Rn and u1 , . . . , un ∈ L0 (A) (364H, 364Yc). This is supposed to recall the notation Pr(X ∈ E), already used in Chapter 27. If, as sometimes seems reasonable, we wish to regard a random variable as a member of L0 (µ) rather than of L0 (µ), then ‘[[u ∈ E]]’ is the appropriate translation of ‘X −1 [E]’.

L1

365B

335

The reasons why we can reach all Borel sets E here, but then have to stop, seem to lie fairly deep. We see that we have here another potential definition of L0 (A), as the set of sequentially order-continuous Boolean homomorphisms from the Borel σ-algebra of R to A. This is suitably independent of realizations of A, but makes the f -algebra structure of L0 difficult to elucidate, unless we move to a further level of abstraction in the definitions, as in 364Ye. I take the space to describe the L0 spaces of general regular open algebras in detail (364U) partly to offer a demonstration of an appropriate technique, and partly to show that we are not limited to σ-algebras of sets and their quotients. This really is a new representation; for instance, it does not meld in any straightforward way with the constructions of 364G-364I. Of course the most important examples are compact Hausdorff spaces, for which alternative methods are available (364V-364W, 364Yk, 364Yi, 364Yl); from the point of view of applications, indeed, it is worth sorting out compact Hausdorff spaces in general (364Yj). The version in 364W is derived from Vulikh 67. But I have starred everything from 364T on, because I shall not rely on this work later for anything essential.

365 L1 Continuing my programme of developing the ideas of Chapter 24 at a deeper level of abstraction, I arrive at last at L1 . As usual, the first step is to establish a definition which can be matched both with the constructions of the previous sections and with the definition of L1 (µ) in §242 (365A-365C, 365F). Next, I give what I regard as the most characteristic internal properties of L1 spaces, including versions of the Radon-Nikod´ ym theorem (365E), before turning to abstract versions of theorems in §235 (365H, 365T) and the duality between L1 and L∞ (365I-365K). As in §§361 and 363, the construction is associated with universal mapping theorems (365L-365N) which define the Banach lattice structure of L1 . As in §§361, 363 and 364, homomorphisms between measure algebras correspond to operators between their L1 spaces; but now the duality theory gives us two types of operators (365O-365Q), of which one class can be thought of as abstract conditional expectations (365R). For localizable measure algebras, the underlying algebra can be recovered from its L1 space (365S), but the measure cannot. 365A Definition Let (A, µ ¯) be a measure algebra. For u ∈ L0 (A), write kuk1 =

R∞ 0

µ ¯[[|u| > α]] dα,

the integral being with respect to Lebesgue measure on R, and allowing ∞ as a value of the integral. (Because the integrand is monotonic, it is certainly measurable.) Set L1µ¯ = L1 (A, µ ¯) = {u : u ∈ L0 (A), kuk1 < ∞}. It is convenient to note at once that if u ∈ L1 (A, µ ¯), then µ[[|u| > α]] must be finite for almost every α > 0, and therefore for every α > 0, since it is a non-increasing function of α; so that [[u > α]] also belongs to the Boolean ring Af = {a : µ ¯a < ∞} for every α > 0. 365B Theorem Let (X, Σ, µ) be a measure space with measure algebra (A, µ ¯). Then the canonical isomorphism between L0 (µ) and L0 (A), defined in 364D and 364J, matches L1 (µ) ⊆ L0 (µ), defined in §242, with L1µ¯ ⊆ L0 (A), and the standard norm of L1 (µ) with k k1 : L1µ¯ → [0, ∞[, as defined in 365A. proof Take any f ∈ L0 = L0 (Σ) (364C); write f • for its equivalence class in L0 (µ), and u for the corresponding element of L0 (A), so that [[|u| > α]] = {x : |f (x)| > α}• in A, for every α ∈ R, and kuk1 =

R∞ 0

µ{x : |f (x)| > α} dα.

Pn (a) If f is a non-negative simple function, it is expressible as i=0 αi χEi where E0 , . . . , En are dist measurable sets of finite measure and αi ≥ 0 for each i; re-enumerating if necessary, we may suppose that α0 ≥ α1 ≥ . . . ≥ αn . In this case, setting αn+1 = 0,

336

Function spaces

Z kuk1 = =

µ{x : f (x) > α} dα = n X n X

365B

j n X X ( µEi )(αj − αj+1 ) j=0 i=0 n X

µEi (αj − αj+1 ) =

i=0 j=i

Z

αi µEi =

f dµ.

i=0

(b) If f is integrable, then there is a non-decreasing sequence hfn in∈N of non-negative simple functions such that |f (x)| = supn∈N fn (x) a.e.; now µ{x : |f (x)| > α} = supn∈N µ{x : fn (x) > α} for every α ∈ R, so Z

Z Z ∞ |f | = sup fn = sup µ{x : fn (x) > α}dα n∈N n∈N 0 Z ∞ = µ{x : |f (x)| > α}dα = kuk1 . 0

1

Thus in this case u ∈ L (A, µ ¯) and kuk1 = kf • k1 . It follows that every member of L1 (µ) corresponds to a member of L1 (A, µ ¯), and that the norm of L1 (µ) 1 corresponds to k k1 on L (A, µ ¯). (c) On the other hand, if u ∈ L1 (A, µ ¯), then µ{x : |f (x)| > α} = µ ¯[[|u| > α]] is finite for every α > 0 (365A). Also, if g is any simple function with 0 ≤ g ≤ |f |,

R

g dµ =

R∞ 0

µ{x : g(x) > α}dα ≤

R∞ 0

µ{x : f (x) > α}dα = kuk1 < ∞.

So f is integrable (122J). This shows that every member of L1 (A, µ ¯) corresponds to a member of L1 (µ). 365C

Accordingly we can apply everything we know about L1 (µ) spaces to L1µ¯ spaces. For instance:

Theorem For any measure algebra (A, µ ¯), L1µ¯ = L1 (A, µ ¯) is a solid linear subspace of L0 (A), and k k1 is 1 1 a norm on Lµ¯ under which Lµ¯ is an L-space. Consequently L1µ¯ is a perfect Riesz space with an ordercontinuous norm which has the Levi property, and if hun in∈N is a non-decreasing norm-bounded sequence in L1µ¯ then it converges for k k1 to supn∈N un . proof (A, µ ¯) is isomorphic to the measure algebra of some measure space (X, Σ, µ) (321J). L1 (µ) is a solid linear subspace of L0 (µ) (242Cb), so L1µ¯ is a solid linear subspace of L0 (A). L1 (µ) is an L-space (354M), so L1µ¯ also is. The rest of the properties claimed are general features of L-spaces (354N, 354E, 356P). 365D Integration Let (A, µ ¯) be any measure algebra. (a) If u ∈ L1 = L1 (A, µ ¯), then u+ , u− belong to L1 , and we may set

R

R

u = ku+ k1 − ku− k1 =

R∞ 0

µ ¯[[u > α]] dα −

R∞ 0

µ ¯[[−u > α]] dα.

Now : L1 → R is an order-continuous positive linear functional (356P), and under the translation of 365B matches the integral on L1 (µ) as defined in 242Ab. (b) Of course kuk1 =

R

|u| ≥ |

R

u| for every u ∈ L1 .

R R (c) If u ∈ L1 , a ∈ A we may set a u = u × χa. (Compare 242Ac.) If γ > 0 and 0 6= a ⊆ [[u > γ]] then there is a δ > γ such that a0 = a ∩ [[u > δ]] 6= 0, so that

R

u= a

R∞ 0

µ ¯(a ∩ [[u > α]])dα ≥

Rγ 0

In particular, setting a = [[u > γ]], µ ¯[[u > γ]] must be finite.

µ ¯a dα +

Rδ γ

µ ¯a0 > γ µ ¯a.

L1

365E

337

R f (d) If u ∈ L1 then Af = {a : µ ¯a < ∞}, as usual. P P If u ≥ 0 R u ≥ 0 iff a u ≥ 0 for every a ∈ A , writing − then u × χa ≥ 0, a u ≥ 0 for every a ∈ A. If u 6≥ 0, then [[u > 0]] 6= 0 and there is an α > 0 such that a = [[u− > α]] 6= 0. But now µ ¯a is finite ((c) above) and

R

R

R

u × χa = − u− × χa = − µ ¯(a ∩ [[u− ≥ β]])dβ ≤ −α¯ µa < 0, R R R 1 f Q If u, v ∈ L and a u = a v for every Ra ∈ A then u = v (cf. 242Ce). If u ≥ 0 in L1 then Rso a u < 0.R Q P Of course u × χa ≤ u so a u ≤ u for every a ∈ A. On setting u = sup{ a u : a ∈ Af }. P R the other hand, R an = [[u > 2−n ]], hu × χan in∈N is a non-decreasing sequence with supremum u, so u = limn→∞ an u, while µ ¯an is finite for every n. Q Q R R (e) If u ∈ L1 , u ≥ 0 and u = 0 then u = 0 (122Rc). If u ∈ L1 , u ≥ 0 and a u = 0 then u × χa = 0, that is, a ∩ [[u > 0]] = 0. R (f )R If C ⊆ L1 is non-empty and upwards-directed and supv∈C v is finite, then sup C is defined in L1 R and sup C = supv∈C v (356P). (g) RIt will occasionally be convenient to adapt the conventions of §133 to the new context; so that I may R write u = ∞ if u ∈ L0 (A), u− ∈ L1 , u+ ∈ / L1 , while u = −∞ if u+ ∈ L1 and u− ∈ / L1 . (h) On this convention, we can restate (f) as follows: if C ⊆ (L0 )+ is non-empty R R R and upwards-directed 0 and has a supremum u in L , then u = sup v in [0, ∞]. P P For if sup v is infinite, then surely v∈C v∈C R u = ∞; while otherwise we can apply (f). Q Q 365E The Radon-Nikod´ ym theorem again (a) Let (A, µ ¯) be a semi-finite measure algebra and ν : A → R an additive functional. Then the followingRare equiveridical: (i) there is a u ∈ L1 = L1 (A, µ ¯) such that νa = a u for every a ∈ A; (ii) ν is additive and continuous for the measure-algebra topology on A; (iii) ν is completely additive. (b) Let (A, µ ¯) be any measure algebra, and ν : Af → R a function. Then the following are equiveridical: (i) ν is additive and bounded and inf a∈A |νa| = 0 whenever A ⊆ Af is downwards-directed and has infimum 0; R (ii) there is a u ∈ L1 such that νa = a u for every a ∈ Af . proof (a) The equivalence of (ii) and (iii) is 327Bd. The equivalence of (i) and (iii) is just a translation of 327D into the new context. α) Set M = supa∈Af |νa| < ∞. (b)(i)⇒(ii)(α Let D ⊆ Af be a maximal dist set. For each d ∈ D, write Ad for the principal ideal of A generated by d, and µ ¯d for the restriction of µ ¯ to Ad , so that (Ad , µ ¯d ) is a totally finite measure algebra. Set νd = ν¹ Ad ; R then νd : Ad → R is completely additive. By (a), there is a ud ∈ L1 (Ad , µ ¯d ) such that a ud = νd a = νa for every a ⊆ d. 0 0 + Now u+ ˜+ d ∈ L (Ad ) corresponds to a member u d of L (A) defined by saying + [[˜ u+ d > α]] = [[ud > α]] = [[ud > α]] if α ≥ 0,

= 1 if α ≤ 0. Set d+ = [[ud > 0]]. If a ∈ A, then

R

u ˜+ = a d

R∞

0 k˜ u+ d k1

µ ¯(a ∩ [[˜ u+ d > α]])dα =

R∞ 0

µ ¯(a ∩ [[u+ d > α]])dα =

R a∩d

u+ d;

1 = νd+ is finite, so that u ˜+ ¯. d ∈ Lµ P β ) For any finite I ⊆ D, set vI = d∈I u (β ˜+ d . Then

taking a = 1, we see that

R

vI = ν(supd∈I d+ ) ≤ M ;

consequently the upwards-directed set A = {vI : I ⊆ D is finite} is bounded above in L1µ¯ , and we can set R R R R P P + v = sup A in L1µ¯ . If a ∈ A, then a vI = d∈I a∩d u+ d∈D a∩d ud . d for each finite I ⊆ D, so a v =

338

Function spaces

365E

Applying the same arguments to −ν, there is a w ∈ L1µ¯ such that

R

w= a

for every a ∈ A. Try u = v − w; then

R

u= a

R

P

d∈D

u+ − a∩d d

R

R

P

d∈D

u− = a∩d d

u− d

a∩d

R

P d∈D

u = a∩d d

P d∈D

ν(a ∩ d)

for every a ∈ A. (γγ ) Now take any a ∈ Af . For J ⊆ D set aJ = supd∈J a ∩ d. Let ² > 0. Then there is a finite I ⊆ D such that |

R

u − νaJ | = | a

P

ν(a ∩ d) − d∈D

P

d∈J

ν(a ∩ d)| ≤ ²

whenever I ⊆ J ⊆ D and J is finite. But now consider A = {a \ aJ : I ⊆ J ⊆ D, J is finite}. Then inf A = 0, so there is a finite J such that I ⊆ J ⊆ D and |νa − νaJ | = |ν(a \ aJ )| ≤ ². Consequently |νa − As ² is arbitrary, νa =

R a

R a

u| ≤ |νa − νaJ | + |

R a

u − νaJ | ≤ 2².

u. As a is arbitrary, (ii) is proved.

R (ii)⇒(i) From where Rwe now are, this is nearly trivial. Thinking of νa as u × χa, ν is surely additive and bounded. Also |νa| ≤ |u| × χa. If A ⊆ Af is non-empty, downwards-directed and has infimum 0, the same is true of {|u| × χa : a ∈ A}, because a 7→ |u| × χa is order-continuous, so inf a∈A |νa| ≤ inf a∈A 365F

R

|u| × χa = inf a∈A k|u| × χak1 = 0.

It will be useful later to have spelt out the following elementary facts.

Lemma Let (A, µ ¯) be a measure algebra. Write S f for the intersection S(A) ∩ L1 (A, µ ¯). Then S f is a norm1 1 dense and order-dense Riesz subspace of L = L (A, µ ¯), and can be identified with S(Af ). The function f f 1 χ : A → S ⊆ L is an injective order-continuous additive lattice homomorphism. If u ≥ 0 in L1 , there is a non-decreasing sequence hun in∈N in (S f )+ such that u = supn∈N un = limn→∞ un . f 0 proof (a) As in 364L, Pn If u ∈ S, Pn we can think of S(A ) as a Riesz subspace of S = S(A), embedded in L (A). it is expressible as i=0 αi χai where a0 , . . . , an ∈ A are dist and no αi is zero. Now |u| = i=0 |αi |χai , so u ∈ L1 iff µ ¯ai < ∞ for every i, that is, iff u ∈ S(Af ); thus S f = S(Af ). Now suppose that u ≥ 0 in L1 . By 364Kd, there is a non-decreasing sequence hun in∈N in S(A)+ such that u0 ≥ 0 and u = supn∈N un in L0 . Because L1 is a solid linear subspace of L0 , every un belongs to L1 and therefore to S f . By 365C, hun in∈N is norm-convergent to u. This shows also that S f is order-dense in L1 . The map χ : Af → S f is an injective order-continuous additive lattice homomorphism; because S f is regularly embedded in L1 , χ has the same properties when regarded as a map into L1 . For general u ∈ L1 , there are sequences in S f converging to u+ and to u− , so that their difference is a sequence in S f converging to u, and u belongs to the closure of S f ; thus S f is norm-dense in L1 .

Remark Of course S f here corresponds to the space of (equivalence classes of) simple functions. 365G Semi-finite algebras: Lemma Let (A, µ ¯) be a measure algebra. (a) (A, µ ¯) is semi-finite iff L1 = L1 (A, µ ¯) is order-dense Rin L0 = L0 (A). R (b) In this case, writing S f = S(A) ∩ L1 (as in 365F), u = sup{ v : v ∈ S f , 0 ≤ v ≤ u} in [0, ∞] for every u ∈ (L0 )+ . proof (a) If (A, µ ¯) is semi-finite then S f is order-dense in L0 (364L), so L1 must also be. If L1 is order-dense 0 in L , then so is S f , by 365F and 352Nc, so (A, µ ¯) is semi-finite, by the other half of 364L.

L1

365J

339

f (b) Set C = {v : v ∈ SR f , 0 ≤ v ≤ u}. R Then C is an upwards-directed set with supremum u, because S 0 is order-dense in L . So u = supv∈C v by 365Dh.

365H Measurable transformations We have a generalization of the ideas of §235 in this abstract context. Theorem Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a sequentially order-continuous Boolean homomorphism. Let T : L0 (A) → L0 (B) be the sequentially order-continuous Riesz homomorphism associated with π (364R). R (a) Suppose that w ≥ R0 in L0 (B) is such that πa w d¯ ν=µ ¯aRwhenever a ∈ A and µ ¯a < ∞. Then for any u ∈ L1 (A, µ ¯) and a ∈ A, πa T u × w d¯ ν is defined and equal to u d¯ µ . R R R (b) Suppose that w0 ≥ 0 in L0 (A) is such that a w0 d¯ µ = ν¯(πa) for every a ∈ A. Then T u d¯ ν = u×w0 d¯ µ whenever u ∈ L0 (A) and either integral is defined in [−∞, ∞]. Remark I am using the convention of 365Dg concerning ‘∞’ as the value of an integral, and the notation R ‘ u d¯ µ’ is supposed to indicate that I am considering the integral in L1 (A, µ ¯). P n proof (a) If u ∈ S f =P L1 (A, µ ¯) ∩ S(A) then u is expressible as i=0 αi χai where a0 , . . . , an have finite n measure, so that T u = i=0 αi χ(πai ) and

R

T u × w d¯ ν=

Pn

i=0

αi

R

πai

w=

Pn

i=0

αi µ ¯ ai =

R

u d¯ µ.

If u ≥ 0 in L1 (A, µ ¯) there is a non-decreasing sequence hun in∈N in S f with supremum u, so that T u = supn∈N T un and w × T u = supn∈N w × T un in L0 (B), and

R

T u × w = supn∈N

R

T un × w = supn∈N

R

un =

R

u.

1

(365C tells us that in this context T u × w ∈ L (B, ν¯).) Finally, for general u ∈ L1 (A, µ ¯),

R

Tu × w =

R

T u+ × w −

R

T u− × w =

R

u+ −

R

u− =

R

u.

(b) The argument follows the same lines: start with u = χa for a ∈ A, then with u ∈ S(A), then with u ∈ L0 (A)+ and conclude with general u ∈ L0 (A). The point is that T is a Riesz homomorphism, so that at the last step Z Z Z Z Z + − + T u = (T u) − (T u) = T (u ) − T (u− ) Z Z Z Z Z = u+ × w0 − u− × w0 = (u × w0 )+ − (u × w0 )− = u × w0 whenever either side is defined in [−∞, ∞]. 365I The duality between L1 and L∞ Let (A, µ ¯) be a measure algebra. If we identify L∞ = L∞ (A) with the solid linear subspace of L0 = L0 (A) generated by e = χ1A (364K), then we have a bilinear map (u, v) 7→ u × v : L1 × L∞ → L1 , because |u × v| ≤ kvk∞ |u| and L1 = L1 (A, µ ¯) is a solid linear subspace of L0 . Note that ku × vk1 ≤ kuk1 kvk∞ , so that the bilinear map (u, v) → 7 u × v has norm at most 1 (253A). R Consequently we have a bilinear functional (u, v) 7→ u × v : L1 × L∞ → R, which also has norm at most 1, corresponding to linear operators S : L1 → (L∞ )∗ and T : L∞ → (L1 )∗ , both of norm at most 1. Because L1 and L∞ are both Banach lattices, we have (L1 )∗ = (L1 )∼ , (L∞ )∗ = (L∞ )∼ (356Dc). Because the norm of L1 is order-continuous, (L1 )∗ = (L1 )× (356Dd). 365J Theorem Let (A, µ ¯) be a measure algebra, and set L1 = L1 (A, µ ¯), L∞ = L∞ (A). Let S : L1 → (L∞ )∗ = (L∞ )∼ , T : L∞ → (L1 )∗ = (L1 )∼ = (L1 )× be the canonical maps defined by the duality between L1 and L∞ , as in 365I. Then (a) S and T are order-continuous Riesz homomorphisms, S[L1 ] ⊆ (L∞ )× , and S is norm-preserving; (b) (A, µ ¯) is semi-finite iff T is injective, and in this case T is norm-preserving, while S is a normed Riesz space isomorphism between L1 and (L∞ )× ;

340

Function spaces

365J

(c) (A, µ ¯) is localizable iff T is bijective, and in this case T is a normed Riesz space isomorphism between L∞ and (L1 )∗ = (L1 )× . proof Take a measure space (X, Σ, µ) such that (A, µ ¯) is isomorphic to its measure algebra. Then we can identify L1 with L1 (µ), as in 365B, and L∞ with L∞ (µ), as in 363I. Moreover, these identifications are based on the canonical embeddingsRof L1 and L∞ in L0 (µ), so that the duality described in 365I corresponds to the familiar duality (u, v) 7→ u × v already used in 243F. (a)(i) If u ≥ 0 in L1 and v ≥ 0 in L∞ then u × v ≥ 0 and (T v)(u) =

R

u × v ≥ 0.

1 ×

As u is arbitrary, T v ≥ 0 in (L ) ; as v is arbitrary, T is a positive linear operator. If v ∈ L∞ , set a = [[v > 0]] ∈ A. ( that we are identifying L0 (µ), as defined in §241, with 0 L (A), as defined in §364.) Then v + = v × χa, so for any u ≥ 0 in L1 (T v + )(u) =

R

u × v × χa = (T v)(u × χa) ≤ (T v)+ (u).

As u is arbitrary, T v + ≤ (T v)+ . On the other hand, because T is a positive linear operator, T v + ≥ T v and T v + ≥ 0, so T v + ≥ (T v)+ . Thus T v + = (T v)+ . As v is arbitrary, T is a Riesz homomorphism (352G). (ii) Exactly the same arguments show that S is a Riesz homomorphism. (iii) Given u ∈ L1 , set a = [[u > 0]]; then kSuk ≥ (Su)(χa − χ(1 \ a)) =

R a

u−

R 1\a

u=

R

|u| = kuk1 ≥ kSuk.

So S is norm-preserving. (iv) By 355Ka, S is order-continuous. (v) If A ⊆ L∞ is a non-empty downwards-directed set with infimum 0, and u ∈ (L1 )+ , then inf v∈A u × v = 0 for every u ∈ (L1 )+ , because v 7→ u × v : L0 → L0 is order-continuous. So inf v∈A (T v)(u) = inf v∈A

R

u × v = inf v∈A ku × vk1 = 0

and the only possible non-negative lower bound for T [A] in (L1 )× is 0. As A is arbitrary, T is ordercontinuous. (vi) The ideas of (v) show also that S[L1 ] ⊆ (L∞ )× . P P If u ∈ (L1 )+ and A ⊆ L∞ is non-empty, downwards-directed and has infimum 0, then inf v∈A (Su)(v) = inf v∈A

R

u × v = 0.

As A is arbitrary, Su is order-continuous. For general u ∈ L1 , Su = Su+ − Su− belongs to (L∞ )× . Q Q (b)(i) If (A, µ ¯) is not semi-finite, let a ∈ A be such that µ ¯a = ∞ and µ ¯b = ∞ whenever 0 6= b ⊆ a. If u ∈ L1 , then [[|u| > n1 ]] has finite measure for every n ≥ 1, so must be dist from a; accordingly This means that

R

a ∩ [[|u| > 0]] = supn≥1 a ∩ [[|u| >

1 n ]]

= 0.

1

u × χa = 0 for every u ∈ L . Accordingly T (χa) = 0 and T is not injective.

(ii) If (A, µ ¯) is semi-finite, take any v ∈ L∞ . Then if 0 ≤ δ < kvk∞ , a = [[|v| > δ]] 6= 0. Let b ⊆ a be such that 0 < µ ¯b < ∞. Then χb ∈ L1 , and kT vk = k|T v|k = kT |v|k ≥ (T |v|)(χb)/kχbk1 ≥ δ because |v| × χb ≥ δχb, so (T |v|)(χb) ≥ δ µ ¯b = δkχbk1 . As δ is arbitrary, kT vk ≥ kvk∞ . But we already know that kT vk ≤ kvk∞ , so the two are equal. As v is arbitrary, T is norm-preserving (and, in particular, is injective). (iii) Still supposing that (A, µ ¯) is semi-finite, S[L1 ] = (L∞ )× . P P Take any h ∈ (L∞ )× . For a ∈ A, set • νa = h(χa ). By 363K, ν : A → R is completely additive. By 365Ea, there is a u ∈ L1 such that

L1

365M

(Su)(χa) =

R

u × χa =

341

R a

u = νa = h(χa)

for every a ∈ A. Because Su and h are both linear functionals on L∞ , they must agree on S(A); because they are continuous and S(A) is dense in L∞ (363C), Su = h. As h is arbitrary, S is surjective. Q Q (c) Using (b), we know that if either T is bijective or (A, µ ¯) is localizable, then (A, µ ¯) is semi-finite, so that (X, Σ, µ) is also semi-finite (322Bd). Given this, 243G tells us that T is bijective iff (X, Σ, µ) is localizable, and in this case T is norm-preserving; but of course (X, Σ, µ) is localizable iff (A, µ ¯) is (322Be). 365K Corollary If (A, µ ¯) is a localizable measure algebra, L∞ (A) is perfect. proof By 365J(b)-(c), we can identify L∞ with (L1 )× ∼ = (L∞ )×× . 365L Theorem Let (A, µ ¯) be a measure algebra and U a Banach space. Let ν : Af → U be a function. Then the following are equiveridical: (i) there is a continuous linear operator T : L1 → U , where L1 = L1 (A, µ ¯), such that νa = T (χa) for every a ∈ Af ; (ii)(α) ν is additive (β) there is an M ≥ 0 such that kνak ≤ M µ ¯a for every a ∈ Af . Moreover, in this case, T is unique and kT k is the smallest number M satisfying the condition in (ii-β). proof (a)(i)⇒(ii) If T : L1 → U is a continuous linear operator, then χa ∈ L1 for every a ∈ Af , so ν = T χ is a function from Af to U . If a, b ∈ Af and a ∩ b = 0, then χ(a ∪ b) = χa + χb in L0 = L0 (A) and therefore in L1 , so ν(a ∪ b) = T χ(a ∪ b) = T (χa + χb) = T (χa) + T (χb) = νa + νb. If a ∈ Af then kχak1 = µ ¯a (using the formula in 365A, or otherwise), so kνak = kT (χa)k ≤ kT kkχak1 = kT k¯ µa. (b)(ii)⇒(i) Now suppose that ν : Af → U is additive and that kνak ≤ M µ ¯a for every a ∈ Af . Let S f = L1 ∩ S(A), as in 365F. Then there is a linear operator T0 : S f → U such that T0 (χa) = νa for every a Af (361F). Next, kT0 uk ≤ M kuk1 for every u ∈ S f . P P If u ∈ S f ∼ = S(Af ), then u is expressible as P∈ m f β χb where b , . . . , b ∈ A are dist (361Eb). So j 0 m j=0 j Pm Pm µbj = M kuk1 . Q Q kT0 uk = k j=0 βj νbj k ≤ M j=0 |βj |¯ There is therefore a continuous linear operator T : L1 → U , extending T0 , and with kT k ≤ kT0 k ≤ M (2A4I). Of course we still have ν = T χ. (c) The argument in (b) shows that T0 = T ¹S f and T are uniquely defined from ν. We have also seen that if T , ν correspond to each other then kνak ≤ kT k¯ µa for every a ∈ Af , kT k ≤ M whenever kνak ≤ M µ ¯a for every a ∈ Af , so that kT k = min{M : M ≥ 0, kνak ≤ M µ ¯a for every a ∈ Af }. 365M Corollary Let (X, Σ, µ) be a measure space and U any Banach space. Set Σf = {E : E ∈ Σ, µE < ∞}. Let ν : Σf → U be a function. Then the following are equiveridical: (i) there is a continuous linear operator T : L1 → U , where L1 = L1 (µ), such that νE = T (χE)• for every E ∈ Σf ; (ii)(α) ν(E ∪ F ) = νE + νF whenever E, F ∈ Σf and E ∩ F = 0 (β) there is an M ≥ 0 auch that kνEk ≤ M µE for every E ∈ Σf . Moreover, in this case, T is unique and kT k is the smallest number M satisfying the condition in (ii-β). proof This is a direct translation of 365L. The only point to note is that if ν satisfies the conditions of (ii), and E, F ∈ Σf are such that E • = F • in the measure algebra (A, µ ¯) of (X, Σ, µ), then µ(E \ F ) = µ(F \ E) = 0, so that ν(E \ F ) = ν(F \ E) = 0 (using condition (ii-β)) and

342

Function spaces

365M

νE = ν(E ∩ F ) + ν(E \ F ) = ν(E ∩ F ) + ν(F \ E) = νF . This means that we have a function ν¯ : Af → U , where Af = {a : a ∈ A, µ ¯a < ∞} = {E • : E ∈ Σf }, defined by setting ν¯E • = νE for every E ∈ Σf . Of course we now have ν¯(a ∪ b) = ν¯a + ν¯b whenever a, b ∈ Af and a ∩ b = 0 (since we can express them as a = E • , b = F • with E ∩ F = ∅), and k¯ ν ak ≤ M µ ¯a for every a ∈ Af . Thus we have a one-to-one correspondence between functions ν : Σf → U satisfying the conditions (ii) here, and functions ν¯ : Af → U satisfying the conditions (ii) of 365L. The rest of the argument is covered by the identification between L1 (µ) and L1 (A, µ ¯) in 365B. 365N Theorem Let (A, µ ¯) be a measure algebra, U a Banach lattice, and T : L1 → U a bounded linear 1 1 operator, where L = L (A, µ ¯). Let ν : Af → U be the corresponding additive function, as in 365L. (a) T is a positive linear operator iff νa ≥ 0 in U for every a ∈ Af ; in this case, T is order-continuous. (b) If U is Dedekind complete and T ∈ L∼ (L1 ; U ), then |T | : L1 → U corresponds to |ν| : Af → U , where Pn |ν|(a) = sup{ i=0 |νai | : a0 , . . . , an ⊆ a are dist} for every a ∈ Af . (c) T is a Riesz homomorphism iff ν is a lattice homomorphism. proof As in 365F, let S f be L1 ∩ S(A), identified with S(Af ). (a)(i) If T is a positive linear operator and a ∈ Af , then χa ≥ 0 in L1 , so νa = T (χa) ≥ 0 in U . 1 (ii) Now suppose that νa ≥ 0 in U for every a ∈ Af , and let u ≥ there is a Pn0 in L , ² > 0 in R. Then f v ∈ S such that 0 ≤ v ≤ u and ku − vk1 ≤ ² (365F). Express v as i=0 αi χai where ai ∈ Af , αi ≥ 0 for each i. Now

kT u − T vk ≤ kT kku − vk1 ≤ ²kT k. On the other hand, Tv =

Pn i=0

αi νai ∈ U + .

As U + is norm-closed in U , and ² is arbitrary, T u ∈ U + . As u is arbitrary, T is a positive linear operator. (iii) By 355Ka, T is order-continuous. (b) The point is that |T ¹S f | = |T |¹S f . P P (i) Because the embedding S f ⊆ L1 is positive, the map f P 7→ P ¹S is a positive linear operator from L∼ (L1 ; U ) to L∼ (S f ; U ) (see 355Bd). So |T ¹S f | ≤ |T |¹S f . (ii) There is a positive linear operator T1 : L1 → U extending |T ¹S f |, by 365M and (a) above, and now T1 ¹S f dominates both T ¹S f and −T ¹S f ; since (S f )+ is dense in (L1 )+ , T1 ≥ T and T1 ≥ −T , so that T1 ≥ |T | and |T ¹S f | = T1 ¹S f ≥ |T |¹S f . Q Q Now 361H tells us that |T |(χa) = |T ¹S f |(χa) = |ν|a for every a ∈ Af . (c)(i) If T is a lattice homomorphism, then so is ν = T χ, because χ : Af → S f is a lattice homomorphism. (ii) Now suppose that χ is a lattice homomorphism. In this case T ¹S f is a Riesz homomorphism (361Gc), that is, |T v| = T |v| for every v ∈ S f . Because S f is dense in L1 and the map u 7→ |u| is continuous both in L1 and in U (354Bb), |T u| = T |u| for every u ∈ L1 , and T is a Riesz homomorphism. 365O Theorem Let (A, µ ¯) and (B, ν¯) be measure algebras. Let π : Af → Bf be a measure-preserving ring homomorphism. (a) There is a unique order-continuous norm-preserving Riesz homomorphism Tπ : L1µ¯ → L1ν¯ such that Tπ (χa) = χ(πa) whenever a ∈ Af . We have Tπ (u × χa) = Tπ u × χ(πa) whenever a ∈ Af and u ∈ L1µ¯ . R R R R (b) Tπ u = u, πa Tπ u = a u for every u ∈ L1ν¯ , a ∈ Af .

L1

365O

343

(c) [[Tπ u > α]] = π[[u > α]] for every u ∈ L1 (A, µ), α > 0. (d) Tπ is surjective iff π is. ¯ is another measure algebra and θ : Bf → C another measure-preserving Boolean homomor(e) If (C, λ) ¯ phism, then Tθπ = Tθ Tπ : L1µ¯ → L1 (C, λ). proof Throughout the proof I will write T for Tπ and S f for S(A) ∩ L1µ¯ ∼ = S(Af ) (see 365F). (a)(i) We have a map ψ : Af → L1ν¯ defined by writing ψa = χ(πa) for a ∈ Af . Because χπ(a ∪ b) = χ(πa ∪ πb) = χπa + χπb,

kχ(πa)k1 = ν¯(πa) = µ ¯a

f

whenever a, b ∈ A and a ∩ b = 0, we get a (unique) corresponding bounded linear operator T : L1µ¯ → L1ν¯ such that T χ = χπ on Af (365L). Because π : Af → Bf and χ : Bf → L1ν¯ are lattice homomorphisms, so is ψ, and T is a Riesz homomorphism (365Nc). Pn Pn (ii) If u ∈ S f , express u as i=0 αi χai where a0 , . . . , an are dist in Af . Then T u = i=0 αi χ(πai ) and πa0 , . . . , πan are dist in Bf , so Pn Pn kT uk1 = i=0 |αi |¯ ν (πai ) = i=0 |αi |¯ µai = kuk1 . Because S f is dense in L1µ¯ and u 7→ kuk1 is continuous (in both L1µ¯ and L1ν¯ ), kT uk1 = kuk1 for every u ∈ L1µ¯ , that is, T is norm-preserving. As noted in 365Na, T is order-continuous. (iii) If a, b ∈ Af then T (χa × χb) = T (χ(a ∩ b)) = χπ(a ∩ b) = χ(πa ∩ πb) = χπa × χπb = χπa × T (χb). Because T is linear and × is bilinear, T (χa × u) = χπa × T u for every u ∈ S f . Because the maps u 7→ u × χa : L1µ¯ → L1µ¯ , T : L1µ¯ → L1ν¯ and v 7→ v × χπa : L1ν¯ → L1ν¯ are all continuous, T u × χπa = T (u × χa) for every u ∈ L1µ¯ . (iv) T is unique because the formula T (χa) = χπa defines T on the norm-dense and order-dense subspace S f . (b) Because T is positive,

R

For a ∈ Af ,

R

T u = kT u+ k1 − kT u− k1 = ku+ k1 − ku− k1 =

Tu = πa

(c) If u ∈ S f , express it as

R

Pn i=0

T u × χπa =

R

T (u × χa) =

R

R

u × χa =

u.

R a

u.

αi χai where a0 , . . . , an are dist; then

π[[u > α]] = π(supi∈I ai ) = supi∈I πai = [[T u > α]] where I = {i : i ≤ n, αi > α}. For u ∈ (L1 )+ , take a sequence hun in∈N in S f with supremum u; then supn∈N T un = T u, so π[[u > α]] = π(sup [[un > α]]) n∈N

(364Mb; [[u > α]] ∈ Af by 365A) = sup π[[un > α]] n∈N

(because π is order-continuous, see 361Ad) = sup [[T un > α]] = [[T u > α]] n∈N

because T is order-continuous. For general u ∈ L1 , π[[u > α]] = π[[u+ > α]] = [[T (u+ ) > α]] = [[(T u)+ > α]] = [[T u > α]] because T is a Riesz homomorphism.

344

Function spaces

365O

(d)(i) Suppose that T is surjective and that b ∈ Bf . Then there is a u ∈ L1µ¯ such that T u = χb. Now b = [[T u > 21 ]] = π[[u > 12 ]] ∈ π[Af ]; as b is arbitrary, π is surjective. (ii) Suppose now that π is surjective. Then T [L1µ¯ ] is a linear subspace of L1ν¯ containing χb for every b ∈ Bf , so includes S(Bf ). If v ∈ (L1ν¯ )+ there is a sequence hvn in∈N in S(Bf )+ with supremum v. For each n, choose un such that T un = vn . Setting u0n = supi≤n ui , we get a non-decreasing sequence hu0n in∈N such that vn ≤ T u0n ≤ v for every n ∈ N. So supn∈N ku0n k1 = supn∈N kT u0n k1 ≤ kvk1 < ∞ and u = supn∈N u0n is defined in L1µ¯ , with T u = supn∈N T u0n = v. Thus (L1ν¯ )+ ⊆ T [L1µ¯ ]; consequently L1ν¯ ⊆ T [L1µ¯ ] and T is surjective. (e) This is an immediate consequence of the ‘uniqueness’ assertion in (i), because for any a ∈ Af Tθ Tπ (χa) = Tθ χ(πa) = χ(θπa), so that Tθ Tπ : L1µ¯ → L1λ¯ is a bounded linear operator taking the right values at elements χa, and must therefore be equal to Tθπ . 365P Theorem Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : Af → B an order-continuous ring homomorphism. R R (a) There is a unique positive linear operator Pπ : L1ν¯ → L1µ¯ such that a Pπ v = πa v for every v ∈ L1ν¯ , a ∈ Af . (b) Pπ is order-continuous and norm-continuous, and kPπ k ≤ 1. (c) If a ∈ Af , v ∈ L1ν¯ then Pπ (v × χπa) = Pπ v × χa. (d) If π[Af ] is order-dense in B then Pπ is a norm-preserving Riesz homomorphism; in particular, Pπ is injective. (e) If (B, ν¯) is semi-finite and π is injective, then Pπ is surjective, and there is for every u ∈ L1µ¯ a v ∈ L1ν¯ such that Pπ v = u and kvk1 = kuk1 . ¯ is another measure algebra and θ : B → C an order(f) Suppose again that (B, ν¯) is semi-finite. If (C, λ) ¯ → L1 , where I write θ0 for the restriction continuous Boolean homomorphism, then Pθπ = Pπ Pθ0 : L1 (C, λ) µ ¯ of θ to Bf . proof I write P for Pπ .

R (a)-(b) For v ∈ L1ν¯ , a ∈ Af set νv (a) = πa v. Then νv : Af → R is additive, bounded (by kvk1 ) and if A ⊆ Af is non-empty, downwards-directed and has infimum 0, then inf a∈A |νv (a)| ≤ inf a∈A

R

R

|v| × χπa = 0

because a 7→ |v| × χπa is a composition of order-continuous functions, therefore order-continuous. So R R 365Eb tells us that there is a P v ∈ L1µ¯ such that a P v = νv (a) = πa v for every a ∈ Af . By 365Dd, this formula defines P v uniquely. Consequently P must be linear (since P v1 + P v2 , αP v will always have the properties defining P (v1R + v2 ), PR (αv)). If v ≥ 0 in L1ν¯ , then a P v = πa v ≥ 0 for every a ∈ Af , so P v ≥ 0 (365Dd); thus P is positive. It must therefore be norm-continuous and order-continuous (355C, 355Ka). Again supposing that v ≥ 0, we have kP vk1 = (using 365Dd). For general v ∈

R

P v = supa∈Af

R

P v = supa∈Af a

R

πa

L1ν¯ , kP vk1 = k|P v|k1 ≤ kP |v|k1 ≤ kvk1 .

(c) For any c ∈ Af ,

v ≤ kvk1

L1

365P

R c

P v × χa =

R c∩a

Pv =

R π(c∩a)

345

v=

R

v × χπa =

πc

R c

P (v × χπa).

(d) Now suppose that π[Af ] is order-dense. Take any v, v 0 ∈ L1ν¯ such that v ∧ v 0 = 0. ?? Suppose, if possible, that u = P v ∧ P v 0 > 0. Take α > 0 such that a = [[u > α]] is non-zero. Since

R

πa

v=

R

Pv ≥

a

R

a

u > 0,

b = πa ∩ [[v > 0]] 6= 0. Let c ∈ Af be such that 0 6= πc ⊆ b; then π(a ∩ c) = πc 6= 0, so a ∩ c 6= 0, and 0<

R

R

u≤

a∩c

a∩c

P v0 ≤

R

πc

v0 .

R But πc ⊆ [[v > 0]] and v ∧ v 0 = 0 so πc v 0 = 0. X X So P v ∧ P v 0 = 0. As v, v 0 are arbitrary, P is a Riesz homomorphism (352G). Next, if v ≥ 0 in L1ν¯ ,

R

P v = supa∈Af

R

R

P v = supa∈Af a

v= πa

R

v

because π[Af ] is upwards-directed and has supremum 1 in B. So, for general v ∈ L1ν¯ , kP vk1 =

R

|P v| =

R

P |v| =

R

|v| = kvk1 ,

and P is norm-preserving. (e) Next suppose that (B, ν¯) is semi-finite and that π is injective. R R (i) If u > 0 in L1µ¯ , there is a v > 0 in L1ν¯ such that P v ≤ u and P v ≥ v. P P Let δ > 0 be such that a = [[u > δ]] 6= 0. Then πa 6= 0. Beca (B, ν¯) is semi-finite, there is a non-zero b ∈ Bf such that b ⊆ πa. Set u1 = P (χb). Then u1 ≥ 0, and also a u1 = ν¯b > 0,

R

1\a

u1 = supc∈Af

R

c\a

R

u1 = supc∈Af

πc\πa

χb = 0.

So u1 × χ(1 \ a) = 0 and 0 6= [[u1 > 0]] ⊆ a. Let γ > 0 be such that [[u1 > γ]] 6= [[u1 > 0]], and set a1 = a \ [[u1 > γ]], v = δγ −1 χ(b ∩ πa1 ). Then P v = δγ −1 P (χb × χ(πa1 )) = δγ −1 P (χb) × χa1 = δγ −1 u1 × χa1 ≤ δχa ≤ u because [[u1 × χa1 > γ]] ⊆ [[u1 > γ]] ∩ a1 = 0 so u1 × χa1 ≤ γχ[[u1 > 0]] ≤ γχa. Also P v > 0 because a1 ∩ [[u1 > 0]] 6= 0, so v 6= 0; and

R

Pv ≥

R

a1

Pv =

R πa1

v=

R

v. Q Q

R R (ii) Now take any u ≥ 0 in L1µ¯ , and set B = {v : v ∈ L1ν¯ , v ≥ 0, P v ≤ u, v ≤ P v}. Let C ⊆ B be a maximal upwards-directed set (applying Zorn’s Lemma to the family P of all upwards-directed subsets of B). We have supv∈C so v0 = sup C is defined in

L1ν¯

R

v ≤ supv∈C

R

P v ≤ kuk1 ,

(365Df). Because P is order-continuous, P v0 = sup P [C] ≤ u, and

R

P v0 = supv∈C

R

P v ≤ supv∈C

R

v=

R

v0 .

if possible, that P v0 6= u. In this case, by (α), there is a v1 > 0 such that P v1 ≤ u − P v0 , R?? Suppose, R v1 ≤ P v1 . In this case, v0 + v1 ∈ B, so C 0 = C ∪ {v0 + v1 } is an upwards-directed subset of B strictly larger than C, which is impossible. X X Thus P v0 = u; also kv0 k1 =

R

v0 ≤

R

P v0 = kP v0 k1 .

(iii) Now take any u ∈ L1µ¯ . By (ii), there are non-negative v1 , v2 ∈ L1ν¯ such that P v1 = u+ , P v2 = u− , kv1 k1 ≤ ku+ k1 , kv2 k1 ≤ ku− k1 . Setting v = v1 − v2 , we have P v = u. Also we must have

346

Function spaces

365P

kvk1 ≤ kv1 k1 + kv2 k1 ≤ ku+ k1 + ku− k1 = kuk1 ≤ kvk1 , so kvk1 = kuk1 , as required. (f ) As usual, this is a consequence of the uniqueness of P . RHowever (because I do not assume that R π[Af ] ⊆ Bf ) there is an extra refinement: we need to know that b Pθ0 w = θb w for every b ∈ B, w ∈ L1λ¯ . P P Because θ is order-continuous and (B, ν¯) is semi-finite, θb = sup{θb0 : b0 ∈ Bf , b0 ⊆ b}, so if w ≥ 0 then

R

w = supb0 ∈Bf ,b0 ⊆ b θb

R

θb0

w = supb0 ∈Bf ,b0 ⊆ b

R

b0

Pθ0 w =

R

b

Pθ0 w.

Expressing w as w+ − w− , we see that the same is true for every w ∈ L1ν¯ . Q Q Now we can say that P Pθ0 is a positive linear operator from L1λ¯ to L1µ¯ such that

R

a

R

P Pθ 0 w =

πa

Pθ0 w =

R

θπa

w=

R

a

Pθπ w

whenever a ∈ Af , w ∈ L1λ¯ . 365Q Proposition Let (A, µ ¯) and (B, µ ¯) be measure algebras and π : Af → Bf a measure-preserving ring homomorphism. (a) In the language of 365O-365P above, Pπ Tπ is the identity operator on L1 (A, µ ¯). (b) If π is surjective (so that it is an isomorphism between Af and Bf ) then Pπ = Tπ−1 = Tπ−1 , Tπ = Pπ−1 = Pπ−1 . proof (a) If u ∈ L1µ¯ , a ∈ Af then

R

P T u= a π π

R

T u= πa π

R a

u.

So u = Pπ Tπ u, by 365Dd. (b) From 365Od, we know that Tπ is surjective, while Pπ Tπ is the identity, so that Pπ = Tπ−1 , Tπ = Pπ−1 . As for Tπ−1 , 365Oe tells us that Tπ−1 = Tπ−1 ; so Pπ−1 = Tπ−1 −1 = Tπ . 365R Conditional expectations It is a nearly universal rule that any investigation of L1 spaces must include a look at conditional expectations. In the present context, they take the following form. (a) Let (A, µ ¯) be a probability algebra and B a closed subalgebra; write ν¯ for the restriction µ ¯¹ B. The identity map from B to A induces operators T : L1ν¯ → L1µ¯ and P : L1µ¯ → L1ν¯ . If we take L0 (A) to be defined as the set of functions from R to A described in 364A, then L0 (B) becomes a subset of L0 (A) in the literal sense, and T is actually the identity operator associated with the subset L1ν¯ ⊆ L1µ¯ ; L1ν¯ is a norm-closed and order-closed Riesz subspace of L1µ¯ . P is a positive linear operator, while P T is the identity, so P is a projection from L1µ¯ onto L1ν¯ . P is defined by the familiar formula

R

Pu = b

R

b

u for every u ∈ L1µ¯ , b ∈ B,

so is the conditional expectation operator in the sense of 242J. (b) Just as in 233I-233J and 242K, we have a fundamental inequality concerning convex functions: if R R ¯ ¯ h : R → R is a convex function and u ∈ L1µ¯ , then h( u) ≤ h(u); and if h(u) ∈ L1µ¯ (364I), then ¯ u) ≤ P (h(u)). ¯ h(P P P I repeat the proof of 233I-233J. For each q ∈ Q, take βq ∈ R such that h(t) ≥ hq (t) = ¯ ¯ q (u) for h(q) + βq (t − q) for every t ∈ R, so that h(t) = supq∈Q hq (t) for every t ∈ R, and h(u) = supq∈Q h 0 0 every u ∈ L = L (A). (This is because S ¯ [[h(u) > α]] = [[u ∈ h−1 [ ]α, ∞[ ]]] = [[u ∈ q∈Q h−1 q [ ]α, ∞[ ]]] −1 ¯ q (u) > α]] = sup [[u ∈ h [ ]α, ∞[ ]]] = sup [[h q∈Q

q

q∈Q

¯ q (u) = h(q)e + βq (u − qe) for every u ∈ L0 , so that for every α ∈ R.) But setting e = χ1, we see that h

R

R

R

hq (u) = h(q) + βq ( u − q) = hq ( u),

L1

365T

because

R

347

¯ q (u)) = h(q)e + βq (P u − qe) = h ¯ q (P u) P (h e = 1 and P e = e. Taking the supremum over q, we get

R

R

h( u) = supq∈Q hq ( u) = supq∈Q

R

¯ q (u) ≤ h

R

¯ h(u),

¯ and if h(u) ∈ L1µ¯ then ¯ ¯ ¯ ¯ u) = sup Q h(P q∈Q hq (P u) = supq∈Q P (hq (u)) ≤ P (h(u)). Q Of course the result in this form can also be deduced from 233I-233J if we represent (A, µ ¯) as the measure algebra of a probability space (X, Σ, µ) and set T = {E : E ∈ Σ, E • ∈ B}. (c) I note here a fact which is occasionally useful. If u ∈ L1µ¯ is non-negative, then [[P u > 0]] = upr([[u > 0]], C), the upper envelope of [[u > 0]] in C as defined in 314V. P P We have only to observe that, for c ∈ C, c ∩ [[P u > 0]] = 0 ⇐⇒ χc × P u = 0 ⇐⇒

R

c

P u = 0 ⇐⇒

R

c

u = 0 ⇐⇒ c ∩ [[u > 0]] = 0.

Taking complements, c ⊇ [[P u > 0]] iff c ⊇ [[u > 0]]. Q Q 365S Recovering the algebra: Proposition (a) Let (A, µ ¯) be a localizable measure algebra. Then A is isomorphic to the band algebra of L1 = L1 (A, µ ¯). (b) Let A be a Dedekind σ-complete Boolean algebra, and µ ¯, ν¯ two measures on A such that (A, µ ¯) and (A, ν¯) are both semi-finite measure algebras. Then L1 (A, µ ¯) is isomorphic, as Banach lattice, to L1 (A, ν¯). proof (a) Because (A, µ ¯) is semi-finite, L1 is order-dense in L0 = L0 (A) (365G). Consequently, L1 and L0 have isomorphic band algebras (353D). But the band algebra of L0 is just its algebra of projection bands (because A and therefore L0 are Dedekind complete, see 364O and 353I), which is isomorphic to A (364Q). (b) Let π : A → A be the identity map. Regarding π as an order-continuous Boolean homomorphism ¯a < ∞} to (A, ν¯), we have an associated positive linear operator P = Pπ : L1ν¯ → L1µ¯ ; from Afµ¯ = {a : µ similarly, we have Q = Pπ−1 : L1µ¯ → L1ν¯ , and both P and Q have norm at most 1 (365Pb). Now 365Pf assures us that QP is the identity operator on L1ν¯ and P Q is the identity operator on L1µ¯ . So P and Q are the two halves of a Banach lattice isomorphism between L1µ¯ and L1ν¯ . 365T Having opened the question of varying measures on a single Boolean algebra, this seems an appropriate moment for a general description of how they interact. Proposition Let A be a Dedekind complete Boolean algebra, and µ ¯ : A → [0, ∞], ν¯ : A → [0, ∞] two functions such that (A, µ ¯) and (A, ν¯) are both semi-finite (therefore localizable) measure algebras. R (a) There is a unique u ∈ L0 = L0 (A) such that (if we allow ∞ as a value of the integral) a u d¯ µ = ν¯a for every a ∈ A. R R (b) For every v ∈ L1ν¯ , v d¯ ν = u × v d¯ µ. R (c) u is strictly positive (i.e., [[u > 0]] = 1) and, writing u1 for the multiplicative inverse of u, a u1 d¯ ν=µ ¯a for every a ∈ A. proof (a) Because (A, ν¯) is semi-finite, there is a partition of unity D ⊆ A such that ν¯d < ∞ for every d ∈ D. For each Rd ∈ D, the functional a 7→ ν¯(a ∩ d) : A →R R is completely additive, so there is a 1 u µ = ν¯(a ∩ d) for every a ∈ A. Because a ud d¯ µ ≥ 0 for every a, ud ≥ 0. Because R d ∈ Lµ¯ such that a ud d¯ 0 u = 0, [[ud > 0]] ⊆ d. Now u = supd∈D ud is defined in L . P P (See 368K below.) For n ∈ N, 1\d d set cn = supd∈D [[ud > n]]. If d, d0 ∈ D are distinct, then d ∩ [[ud0 > n]] = 0, so d ∩ cn = [[ud > n]]. Set c = inf n∈N cn . If d ∈ D, then d ∩ c = inf n∈N d ∩ cn = inf n∈N [[ud > n]] = 0. But c ⊆ c0 ⊆ sup D, so c = 0. By 364Ma, {ud : d ∈ D} is bounded above in L0 , so has a supremum, because L0 is Dedekind complete, by 364O. Q P Q For finite I ⊆ D set u ˜I = d∈I ud = supd∈I ud (because ud ∧ uc = 0 for distinct c, d ∈ D). Then u = sup{˜ uI : I ⊆ D, I is finite}. So, for any a ∈ A,

348

Function spaces

Z

365T

Z u d¯ µ=

sup I⊆D is finite

a

(365Dh) =

u ˜I d¯ µ a

XZ

sup I⊆D is finite

d∈I

ud d¯ µ=

a

sup I⊆D is finite

X

ν¯(a ∩ d) = ν¯a.

d∈I

Note that if a ∈ A is non-zero, then ν¯a > 0, so a ∩ [[u > 0]] 6= 0; consequently [[u > 0]] = 1. To see that u is unique, observe that if u0 has the same property then for any d ∈ D

R

a 0

(u × χd)d¯ µ = ν¯(a ∩ d) =

R

a

(u0 × χd)

for every a ∈ A, so that u × χd = u × χd; because sup D = 1 in A, u must be equal to u0 . (b) Use 365Hb, with π and T the identity maps. (c) In the same way, there is a w ∈ L0 such that observe that applying 365Hb we get

R

R

R a

w × χa × u d¯ µ=

w d¯ ν = µ ¯a for every a ∈ A. To relate u and w,

R

w × χa d¯ ν

for every a ∈ A, that is, a w × u d¯ µ=µ ¯a for every a. But from this we see that w × u × χb = χb at least when µ ¯b < ∞, so that w × u = χ1 is the multiplicative identity of L0 , and w = u1 . 365X Basic exercises (a) Let (A, µ ¯) be a measure algebra, and u ∈ L1µ¯ . Show that

R

u=

R∞ 0

µ ¯[[u > α]] dα −

R0

−∞

µ ¯(1 \ [[u > α]]) dα.

R > (b) Let (A, µ ¯) be any measure algebra, and u ∈ L1µ¯ . (i) Show that kuk1 ≤ 2 supa∈Af | a u|. (Hint: R R 246F.) (ii) Show that for any ² > 0 there is an a ∈ Af such that | u − b u| ≤ ² whenever a ⊆ b ∈ A. + > (c) Let U be an L-space. If hun in∈N is any R norm-bounded sequence in RU , show that lim inf n→∞ un = supn∈N inf m≥n um is defined in U , and that lim inf n→∞ un ≤ lim inf n→∞ un .

(d) Let U be an L-space. Let F be a filter on U + such that R{u : u ≥ 0, kuk ≤ k} belongs to F for some R k ∈ N. Show that u0 = supF ∈F inf F is defined in U , and that u0 ≤ supF ∈F inf u∈F u. iff

(e) Let (A, µ ¯) be a measure algebra and A ⊆ L1µ¯ a non-empty set. Show that A is bounded above in L1µ¯

Pn R

sup{

ui : a0 , . . . , an is a partition of unity in A, u0 , . . . , un ∈ A} R is finite, and that in this case the supremum is Rsup A. (Hint: P givenRu0 , . . . , un ∈ A, set bij = [[ui ≥ uj ]], n bi = supj6=i bij , ai = bi \ supj
ai

(f ) Let (A, µ ¯) be any measure algebra and ν : Af → R a bounded additive function. Show that the R following are equiveridical: (i) there is a u ∈ L1µ¯ such that νa = a u for every a ∈ Af ; (ii) for every ² > 0 there is a δ > 0 such that |νa| ≤ ² whenever µ ¯a ≤ δ; (iii) for every ² > 0, c ∈ Af there is a δ > 0 such that νa ≤ ² whenever a ⊆ c and µ ¯a ≤ δ; (iv) for every ² > 0 there are c ∈ Af , δ > 0 such that |νa| ≤ ² whenever f a ∈ A and µ ¯(a ∩ c) ≤ δ; (v) limn→∞ νan = 0 whenever han in∈N is a non-increasing sequence in Af with infimum 0. (g) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a sequentially order-continuous Boolean homomorphism. Let T : L0 (A) → L0R(B) be the Riesz homomorphism associated with π (364R). Suppose that w ≥ 0 inRL0 (B) is such that πa w d¯ ν = µ ¯a whenever a ∈ A. Show that for any u ∈ L0 (A, µ ¯), R T u × w d¯ ν = u d¯ µ whenever either is defined in [−∞, ∞].

365 Notes

L1

349

> (h) Let (A, µ ¯) be a measure algebra and a ∈ A; write Aa for the principal ideal it generates. Show that if π is the identity embedding of Af ∩ Aa into Af , then Tπ , as defined in 365O, identifies L1 (Aa , µ ¯¹ Aa ) with a band in L1 (A, µ ¯). > (i) Let (X, Σ, µ) and (Y, T, ν) be measure spaces, with measure algebras (A, µ ¯) and (B, ν¯). Let φ : X → Y be an inverse-measure-preserving function and π : B → A the corresponding measure-preserving homomorphism (324M). Show that Tπ : L1ν¯ → L1µ¯ (365O) corresponds to the map g • 7→ (gφ)• : L1 (ν) → L1 (µ) of 242Xf. (j) Let (A, µ ¯) and (B, ν¯) be measure algebras. Let π : Af → Bf be a ring homomorphism such that, for some γ > 0, ν¯(πa) ≤ γ µ ¯a for every a ∈ Af . (i) Show that there is a unique order-continuous Riesz 1 1 homomorphism T : Lµ¯ → Lν¯ such that T (χa) = χ(πa) whenever a ∈ Af , and that kT k ≤ γ. (ii) Show that [[T u > α]] = π[[u > α]] for every u ∈ L1 (A, µ), α > 0. (iii) Show that T is surjective iff π is, injective iff π is. (iv) Show that T is norm-preserving iff ν¯(πa) = µ ¯a for every a ∈ Af . (k) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a measure-preserving Boolean homomorphism. Let T : L1µ¯ → L1ν¯ and P : L1ν¯ → L1µ¯ be the operators corresponding to π¹ Af , as described in 365O-365P, and T˜ : L∞ (A) → L∞ (B) the operator corresponding to π, as described in 363F. (i) Show that T (u × v) = T u × T˜v for every u ∈ L1µ¯ , v ∈ L∞ (A). (ii) Show that if π is order-continuous, then R R P v × u = v × T˜u for every u ∈ L∞ (A), v ∈ L1ν¯ . > (l) Let (X, Σ, µ) be a probability space, with measure algebra (A, µ ¯), and let T be a σ-subalgebra of Σ. Set ν = µ¹ T, B = {F • : F ∈ T} ⊆ A, ν¯ = µ ¯¹ B, so that (B, ν¯) is a measure algebra. Let π : B → A be the identity homomorphism. Show that Tπ : L1ν¯ → L1µ¯ (365O) corresponds to the canonical embedding of L1 (ν) in L1 (µ) described in 242Jb, while Pπ : L1µ¯ → L1ν¯ (365P) corresponds to the conditional expectation operator described in 242Jd. b µ (m) Let (A, µ ¯) be a semi-finite measure algebra, and (A, ˆ) its localization (322P). Show that the natural b b µ embedding of A in A induces a Banach lattice isomorphism between L1µ¯ and L1 (A, ˆ), so that the band 1 b of A. algebra of Lµ¯ can be identified with the Dedekind completion A (n) Let A be a Dedekind σ-complete Boolean algebra and µ ¯, ν¯ two functions such that (A, µ ¯), (A, ν¯) are measure algebras. Show that L1µ¯ ⊆ L1ν¯ (as subsets of L0 (A)) iff there is a γ > 0 such that ν¯a ≤ γ µ ¯a for every a ∈ A. (Hint: show that the identity operator from L1µ¯ to L1ν¯ is bounded.) (o) Let (A, µ ¯) be a measure algebra, I∞ the ideal of ‘purely infinite’ elements of A, µ ¯sf the measure on B = A/I∞ (322Xa). Let π : A → B be the canonical map. Show that Tπ , as defined in 365O, is a Banach lattice isomorphism between L1 (A, µ ¯) and L1 (B, µ ¯sf ). (p) Let (X, Σ, µ) be a a semi-finite measure space. Show that L1 (µ) is separable iff µ is σ-finite and has countable Maharam type. 365Y Further exercises (a) Let (A, µ ¯) be a semi-finite measure algebra, not {0}. Show that the topological density of L1 (A, µ ¯) (331Yf) is max(ω, τ (A), c(A)), where τ (A), c(A) are the Maharam type and cellularity of A. 365 Notes and comments You should not suppose that L1 spaces appear in the second half of this chapter because they are of secondary importance. Indeed I regard them as the most important of all function spaces. I have delayed the discussion of them for so long because it is here that for the first time we need measure algebras in an essential way. The actual definition of L1µ¯ which I give is designed for speed rather than illumination; I seek only a formula, visibly independent of any particular representation of (A, µ ¯) as the measure algebra of a measure space, for which I can prove 365B. R ∞I use a relatively primitive R argument in 365B, not appealing to Fubini’s theorem; of course the formula 0 µ{x : |f (x)| > α}dα = |f (x)|dx can also be regarded as a reversal of

350

Function spaces

365 Notes

order of integration in a double integral (252N, 252O). 365C-365D and 365Ea are now elementary. In 365Eb I take a page to describe a form of the Radon-Nikod´ ym theorem which is applicable to arbitrary measure algebras, at the cost of dealing with functionals on the ring Af rather than on the whole algebra A. This is less for the sake of applications than to emphasize one of the central properties of L1 : it depends only on Af and µ ¯¹ Af . For alternative versions of the condition 365Eb(i) see 365Xf. The convergence theorems (B.Levi’s theorem, Fatou’s lemma and Lebesgue’s dominated convergence theorem) are so central to the theory of integrable functions that it is natural to look for versions in the language here. Corresponding to B.Levi’s theorem is the Levi property of a norm in an L-space; note how the abstract formulation makes it natural to speak of general upwards-directed families rather than of nondecreasing sequences, though the sequential form is so often used that I have spelt it out (365C). In the same way, the integral becomes order-continuous rather than just sequentially order-continuous (365Da). Corresponding to Fatou’s lemma we have 365Xc-365Xd. For abstract versions of Lebesgue’s theorem I will wait until §367. In 365H I have deliberately followed the hypotheses of 235A and 235T. Of course 365H can be deduced from these if we use the Stone representations of (A, µ ¯) and (B, ν¯), so that π can be represented by a function between the Stone spaces (312P). But 365H is essentially simpler, because the technical problems concerning measurability which took up so much of §235 have been swept under the carpet. In the same way, 365Xg corresponds to 235E. Here we have a fair example of the way in which the abstract expression in of measure algebras can be tidier than the expression in of measure spaces. But in my view this is because here, at least, some of the mathematics has been left out. 365J is a re-run of 243G, but with the additional refinement that I examine the action of L1 on L∞ (the operator S) as well as the action of L∞ on L1 (the operator T ). Note that (b-iii) and (c) of the proof of 365J depend on the Radon-Nikod´ ym theorem; these parts lie a step deeper than the rest. 365L-365N correspond closely to 361F-361H and 363E. Theorems 365O-365Q lie at the centre of my picture of L1 spaces, and are supposed to show their dual nature. Starting from a semi-finite measure algebra (A, µ ¯) we have two essentially different routes to the L1 -space: we can either build it up from characteristic functions of elements of finite measure, so that it is naturally embedded in L0 (A), or we can think of it as the order-continuous dual of L∞ (A). The first is a ‘covariant’ construction (signalled by the formula Tθπ = Tθ Tπ in 365Oe) and the second is ‘contravariant’ (so that Pθπ = Pπ Pθ0 in 365Pf). The first construction is the natural one if we are seeking to copy the ideas of §242, but the second arises inevitably if we follow the ordinary paths of functional analysis and study dual spaces whenever they appear. The link between them is the Radon-Nikod´ ym theorem. I have deliberately written out 365O and 365P with different hypotheses on the homomorphism π in the hope of showing that the two routes to L1 really are different, and can be expected to tell us different things about it. I use the letter P in 365P in order to echo the language of 242J; in the most important context, in which A is actually a subalgebra of B and π is the identity map, P is a kind of conditional expectation operator (365R). I note that in the proof of 365Pe I have returned to first principles, using some of the ideas of the Radon-Nikod´ ym theorem (232E), but a different approach to the exhaustion step (converting ‘for every u > 0 there is a v > 0 such that P v ≤ u’ into ‘P is surjective’). I chose the somewhat cruder method in 232E (part (c) of the proof) in order to use the weakest possible form of the axiom of choice. In the present context such scruples seem absurd. I used the words ‘covariant’ and ‘contravariant’ above; of course this distinction depends on the side of the mirror on which we are standing; if our measure-preserving homomorphism is derived (contravariantly) from an inverse-measure-preserving transformation, then the T ’s become contravariant (365Xi). An important component of this work, for me, is the fact that not all measure-preserving homomorphisms between measure algebras can be represented by inverse-measure-preserving functions (343Jb, 343M). I have already remarked (in the notes to §244) that the properties of L1 (µ) are not much affected by peculiarities in a measure space (X, Σ, µ), because (unlike L0 or L∞ ) they really depend only on Af , the ring of elements of finite measure in the measure algebra. (See 365O-365Q and 365Xm-365Xn.) Note that while the algebra A is uniquely determined (given that (A, µ ¯) is localizable, 365Sa), the measure µ ¯ is not; if A is any algebra carrying two non-isomorphic semi-finite measures, the corresponding L1 spaces are still isomorphic (365Sb). For instance, the L1 -spaces of Lebesgue measure µ on R, and the subspace measure µ[0,1] on [0, 1], are isomorphic, though their measure algebras are not. I make no attempt here to add to the results in §§246, 247, 354 and 356 concerning uniform integrability

Lp

366D

351

and weak compactness. Once we have left measure spaces behind, these ideas belong to the theory of Banach lattices, and there is little to relate them to the questions dealt with in this section. But see 373Xj and 373Xn below.

366 Lp In this section I apply the methods of this chapter to Lp spaces, where 1 < p < ∞. The constructions proceed without surprises up to 366E, translating the ideas of §244 by the methods used in §365. Turning to the action of Boolean homomorphisms on Lp spaces, I introduce a space M 0 , which can be regarded as the part of L0 that can be determined from the ring Af of elements of A of finite measure (366F), and which includes Lp whenever 1 ≤ p < ∞. Now a measure-preserving ring homomorphism from Af to Bf acts on the M 0 spaces in a way which includes injective Riesz homomorphisms from Lp (A, µ ¯) to Lp (B, ν¯) p p and surjective positive linear operators from L (B, ν¯) to L (A, µ ¯) (366H). The latter may be regarded as conditional expectation operators (366J). The case p = 2 (366K-366L) is of course by far the most important. 366A Definition Let (A, µ ¯) be a measure algebra and suppose that 1 α]] = [[|u| > α1/p ]] if α ≥ 0 = 1 if α < 0. ¯ (In the language of 364I, |u|p = h(u), where h(t) = |t|p for t ∈ R.) Set ¯) = {u : u ∈ L0 (A), |u|p ∈ L1 (A, µ ¯)}, Lpµ¯ = Lp (A, µ and for u ∈ Lpµ¯ set

R

1/p

kukp = ( |u|p )1/p = k|u|p k1 . 366B Theorem Let (X, Σ, µ) be a measure space, and (A, µ ¯) its measure algebra. Then the canonical isomorphism between L0 (µ) and L0 (A) (364Jc) makes Lp (µ), as defined in §244, correspond to Lp (A, µ ¯). proof What we really have to check is that if w ∈ L0 (µ) corresponds to u ∈ L0 (A), then |w|p , as defined in 244A, corresponds to |u|p as defined in 366A. But this was noted in 364Jb. Now, because the isomorphism between L0 (µ) and L0 (A) matches L1 (µ) with L1 (A, µ ¯) (365B), we can be sure that |w|p ∈ L1 (µ) iff |u|p ∈ L1 (A, µ ¯), and that in this case ¢1/p ¡R p ¢1/p ¡R kwkp = |w|p = |u| = kukp , as required. 366C Corollary For any measure algebra (A, µ ¯) and p ∈ ]1, ∞[, Lp = Lp (A, µ ¯) is a solid linear subspace 0 of L (A). It is a Dedekind complete Banach lattice under its norm k k . Setting q = p/(p − 1), (Lp )∗ is p R q f identified with L (A, µ ¯) by the duality (u, v) 7→ u × v. Writing A for the ring {a : a ∈ A, µ ¯a < ∞}, S(Af ) p is norm-dense in L . proof Because we can find a measure space (X, Σ, µ) such that (A, µ ¯) is isomorphic to the measure algebra of µ (321J), this is just a digest of the results in 244B, 244E, 244F, 244G, 244H and 244K. (Of course S(Af ) corresponds to the space S of equivalence classes of simple functions in 244Ha, just as in 365F.) 366D

I can add a little more, corresponding to 365C and 365J.

Theorem Let (A, µ ¯) be a measure algebra, and p ∈ ]1, ∞[. (a) The norm k kp on Lp = Lp (A, µ ¯) is order-continuous. (b) Lp has the Levi property.

352

Function spaces

366D

(c) Setting q = p/(p − 1), the canonical identification of Lq = Lq (A, µ ¯) with (Lp )∗ is a Riesz space q p ∼ p × isomorphism between L and (L ) = (L ) . (d) Lp is a perfect Riesz space. proof (a) Suppose that A ⊆ Lp is non-empty, downwards-directed and has infimum 0. For u, v ≥ 0 in Lp , u ≤ v ⇒ up ≤ v p (by the definition in 366A, or otherwise), so B = {up : u ∈ A} is downwards-directed. If 1/p v0 = inf B in L1 = L1 (A, µ ¯), then v0 (defined by the formula in 366A, or otherwise) is less than or equal to every member of A, so must be 0, and v0 = 0. Accordingly inf B = 0 in L1 . Because k k1 is order-continuous (365C), 1/p

inf u∈A kukp = inf u∈A kup k1

= (inf v∈B kvk1 )1/p = 0.

As A is arbitrary, k kp is order-continuous. (b) Now suppose that A ⊆ (Lp )+ is non-empty, upwards-directed and norm-bounded. Then B = {up : 1/p u ∈ A} is non-empty, upwards-directed and norm-bounded in L1 . So v0 = sup B is defined in L1 , and v0 p is an upper bound for A in L . (c) By 356Dd, (Lp )∗ = (Lp )∼ = (Lp )× . The extra information we need is that the identification of Lq with (LRp )∗ is an order-isomorphism. P P (α) If w ∈ (Lq )+ and u ∈ (Lp )+ then u × w ≥ 0 in L1 , so (T w)(u) = u × w ≥ 0, writing T : Lq → (Lp )∗ for the canonical bijection. As u is arbitrary, T w ≥ 0. As w is arbitrary, T is a positive linear operator. (β) If w ∈ Lq and T w ≥ 0, consider u = (w− )q/p . Then u ≥ 0 in Lp and w+ × u = 0 (because [[w+ > 0]] ∩ [[u > 0]] = [[w+ > 0]] ∩ [[w− > 0]] = 0), so 0 ≤ (T w)(u) =

R

w×u=−

R

R

w− × u = − (w− )p ≤ 0,

R and (w− )p = 0. But as (w− )p ≥ 0 in L1 , this means that (w− )p and w− must be 0, that is, w ≥ 0. As w is arbitrary, T −1 is positive and T is an order-isomorphism. Q Q (d) This is an immediate consequence of (c), since p = q/(q − 1), so that Lp can be identified with (L ) = (Lq )× . From 356M we see that it is also a consequence of (a) and (b). q ∗

366E Proposition Let (A, µ ¯) be a semi-finite measure algebra, and p ∈ [1, ∞]. Set q = p/(p − 1) if 1 < p < ∞, and q = ∞ if p = 1, q = 1 if p = ∞. Then Lq (A, µ ¯) = {u : u ∈ L0 (A), u × v ∈ L1 (A, µ ¯) for every v ∈ Lp (A, µ ¯)}. proof (a) We already know that if u ∈ Lp and v ∈ Lq then u × v ∈ L1 ; this is elementary if p ∈ {1, ∞} and otherwise is covered by 366C. (b) So suppose that u ∈ L0 \ Lp . If p = 1 then of course χ1 ∈ L∞ = Lq and u × χ1 ∈ / L1 . If p > 1 set A = {w : w ∈ S(Af ), 0 ≤ w ≤ |u|}. Because µ ¯ is semi-finite, S(Af ) is order-dense in L0 (364L), and |u| = sup A. Because the norm on Lp has the Levi property (366Db, 363Ba) and A is not bounded above in Lp , supw∈A kwkp = ∞. n q R For each n n∈ N choose wn ∈ A with kwn kp > 4 . Then there is a vn ∈ LR such that kvn kq = 1 and wn × vn ≥ 4 . P P (α) If p < ∞ this is covered by 366C, since kwn kp = sup{ wn × v : kvkq ≤ 1}. (β) If p = ∞ then [[wn > 4Rn ]] 6= 0; because µ ¯ is semi-finite, there is a b ⊆ [[wn > 4n ]] such that 0 < µ ¯b < ∞, and 1 1 n k µ¯b χbk1 = 1, while wn × µ¯b χb ≥ 4 . Q Q P∞ Because Lq is complete (363Ba, 366C), v = n=0 2−n |vn | is defined in Lq . But now

R

|u| × v ≥ 2−n

R

w n × v n ≥ 2n

for every n, so u × v ∈ / L1 . Remark This result is characteristic of perfect subspaces of L0 ; see 369C and 369J. 366F The next step is to look at the action of Boolean homomorphisms, as in 365O. It will be convenient to be able to deal with all Lp spaces at once by introducing names for a pair of spaces which include all of them.

Lp

366H

353

Definition Let (A, µ ¯) be a measure algebra. Write Mµ¯0 = M 0 (A, µ ¯) = {u : u ∈ L0 (A), µ ¯[[|u| > α]] < ∞ for every α > 0}, Mµ¯1,0 = M 1,0 (A, µ ¯) = {u : u ∈ Mµ¯0 , u × χa ∈ L1 (A, µ ¯) whenever µ ¯a < ∞}. 366G Lemma Let (A, µ ¯) be any measure algebra. Write M 0 = M 0 (A, µ ¯), etc. 0 1,0 (a) M and M are Dedekind complete solid linear subspaces of L0 which include Lp for every p ∈ [1, ∞[; moreover, M 0 is closed under multiplication. (b) If u ∈ M 0 and u ≥ 0, there is a non-decreasing sequence hun in∈N in S(Af ) such that u = supn∈N un . + (c) M 1,0 = {u : u ∈ L0 , R(|u| − ²χ1) ∈ L1 for every ² > 0} = L1 + (L∞ ∩ RM 0 ). R R 1,0 ¯a < ∞, ¯a < ∞, then u ≤ v; so if a u = a v whenever µ (d) If u, v ∈ M and a u ≤ a v whenever µ u = v. proof (a) If u, v ∈ M 0 and γ ∈ R, then for any α > 0 [[|u + v| > α]] ⊆ [[|u| > 12 α]] ∪ [[|v| > 12 α]], [[|γu| > α]] ⊆ [[|u| > [[|u × v| > α]] ⊆ [[|u| >

√

α 1+|γ| ]],

α ]] ∪ [[|v| >

√ α ]]

(364F) are of finite measure. So u + v, γu and u × v belong to M 0 . Thus M 0 is a linear subspace of L0 closed under multiplication. If u ∈ M 0 , |v| ≤ |u| and α > 0, then [[|v| > α]] ⊆ [[|u| > α]] is of finite measure; thus v ∈ M 0 and M 0 is a solid linear subspace of L0 . It follows that M 1,0 also is. If u ∈ Lp = Lp (A, µ ¯), where p < ∞, and α > 0, then [[|u| > α]] = [[|u|p > αp ]] is of finite measure, so u ∈ M 0 ; moreover, if µ ¯a < ∞, then χa ∈ Lq , where q = p/(p − 1), so u × χa ∈ L1 ; thus u ∈ M 1,0 . To see that M 0 is Dedekind complete, observe that if A ⊆ (M 0 )+ is non-empty and bounded above by u0 ∈ M 0 , and α > 0, then {[[u > α]] : u ∈ A} is bounded above by [[u0 > α]] ∈ Af , so has a supremum in A (321C). Accordingly sup A is defined in L0 (364Mc) and belongs to M 0 . Finally, M 1,0 , being a solid linear subspace of M 0 , must also be Dedekind complete. (b) If u ≥ 0 in M 0 , then there is a non-decreasing sequence hun in∈N in S = S(A) such that u = supn∈N un and u0 ≥ 0 (364Kd). But now every un belongs to S ∩ M 0 = S(Af ), just as in 365F. (c)(i) If u ∈ M 1,0 and ² > 0, then a = [[|u| > ²]] ∈ Af , so u × χa ∈ L1 = L1 (A, µ ¯); but (|u| − ²χ1)+ ≤ + 1 |u| × χa, so (|u| − ²χ1) ∈ L . (ii) Suppose that u ∈ L0 and (|u| − ²χ1)+ ∈ L1 for every ² > 0. Then, given ² > 0, v = (|u| − 21 ²χ1)+ ∈ L , and µ ¯[[v > 12 ²]] < ∞; but [[|u| > ²]] ⊆ [[v > 12 ²]], so also has finite measure. Thus u ∈ M 0 . Next, if a ∈ Af , then |u × χa| ≤ χa + (|u| − χ1)+ ∈ L1 , so u ∈ M 1,0 . 1

(iii) Of course L1 and L∞ ∩ M 0 are included in M 1,0 , so their linear sum also is. On the other hand, if u ∈ M 1,0 , then u = (u+ − χ1)+ − (u− − χ1)+ + (u+ ∧ χ1) − (u− ∧ χ1) ∈ L1 + (L∞ ∩ M 0 ). R R 0 0 (d) Take to Mµ¯1,0 , µ ¯a < ∞ and u ≤ a u0 , a R R α >0 0 and set a = [[u − u > α]]. Because both u and u belong R that is, a u − u ≤ 0; so a must be 0 (365Dc). As α is arbitrary, u − u0 ≤ 0 and u ≤ u0 . If a u = a u0 for every a ∈ Af , then u0 ≤ u so u = u0 . 366H Theorem Let (A, µ ¯) and (B, ν¯) be measure algebras. For p ∈ [1, ∞], write Lpµ¯ , Lpν¯ for Lp (A, µ ¯), 0 L (B, ν¯); similarly, write Mµ¯ for M 0 (A, µ ¯), etc. Let π : Af → Bf be a measure-preserving ring homomorphism. (a)(i) We have a unique order-continuous Riesz homomorphism T = Tπ : Mµ¯0 → Mν¯0 such that T (χa) = χ(πa) for every a ∈ Af . (ii) [[T u > α]] = π[[u > α]] for every u ∈ Mµ¯0 and α > 0. (iii) T is injective and multiplicative. p

354

Function spaces

366H

(iv)R For p ∈ [1, ∞] and u ∈ Mµ¯0 , T u ∈ Lpν¯ iff u ∈ Lpµ¯ , and in this case kT ukp = kukp . In particular, T u = u whenever u ∈ L1µ¯ . (v) For u ∈ Mµ¯0 , T u ∈ Mν¯1,0 iff u ∈ Mµ¯1,0 . (b)(i) We have a unique order-continuous positive linear operator P = Pπ : Mν¯1,0 → Mµ¯1,0 such that R R P v = πa v whenever v ∈ Mν¯1,0 and a ∈ Af . a (ii) If u ∈ Mν¯0 , v ∈ Mν¯1,0 and v × T u ∈ Mν¯1,0 , then P (v × T u) = u × P v. 0 ∞ (iii) If q ∈ [1, ∞[ and v ∈ Lqν¯ , then P v ∈ Lqµ¯ and kP vkq ≤ kvkq ; if v ∈ L∞ ν ¯ ∩ Mν ¯ , then P v ∈ Lµ ¯ and kP vk∞ ≤ kvk∞ . (iv) P T u = u for every u ∈ Mµ¯1,0 ; in particular, P [Lpν¯ ] = Lpµ¯ for every p ∈ [1, ∞[. (c) If (C, λ) is another measure algebra and θ : Bf → Cf another measure-preserving ring homomorphism, then Tθπ = Tθ Tπ : Mµ¯0 → Mλ¯0 and Pθπ = Pπ Pθ : Mλ¯1,0 → Mµ¯1,0 . (d) Now suppose that π[Af ] = Bf , so that π is a measure-preserving isomorphism between the rings Af and Bf . Then (i) T is a Riesz space isomorphism between Mµ¯0 and Mν¯0 , and its inverse is Tπ−1 . (ii) P is a Riesz space isomorphism between Mν¯1,0 and Mν¯1,0 , and its inverse is Pπ−1 . (iii) The restriction of T to Mµ¯1,0 is P −1 = Pπ−1 ; the restriction of T −1 = Tπ−1 to Mν¯1,0 is P . (iv) For any p ∈ [1, ∞[, T ¹Lpµ¯ = Pπ−1 ¹Lpµ¯ and P ¹Lpν¯ = Tπ−1 ¹Lpν¯ are the two halves of a Banach lattice isomorphism between Lpµ¯ and Lpν¯ .

R

proof (a)(i) By 361J, π induces a multiplicative Riesz homomorphism T0 : S(Af ) → S(Bf ) which is orderf continuous ) and α > 0,Pthen [[T0 u > α]] = π[[u > α]]. P P Express Pn because π is (361Ad, 361Je). If u ∈ S(A n f u as i=0 αi χai where a0 , . . . , an are dist in A ; then T0 u = i=0 αi χ(πai ), so [[T0 u > α]] = sup{πai : i ≤ n, αi > α} = π(sup{ai : i ≤ n, αi > α}) = π[[u > α]]. Q Q Now if u0 ≥ 0 in Mµ¯0 , sup{T0 u : u ∈ S(Af ), 0 ≤ u ≤ u0 } is defined in Mν¯0 . P P Set A = {u : u ∈ S(Af ), 0 ≤ u ≤ u0 }. Because u0 = sup A (366Gb), supu∈A [[T u > α]] = supu∈A π[[u > α]] = π(supu∈A [[u > α]]) = π[[u0 > α]] is defined and belongs to Bf for any α > 0. Also inf n≥1 supu∈A [[T u > n]] = π(inf n≥1 [[u0 > n]]) = 0. By 364Mb, v0 = sup T0 [A] is defined in L0 (B), and [[v0 > α]] = π[[u0 > α]] ∈ Bf for every α > 0, so v0 ∈ Mν¯0 , as required. Q Q Consequently T0 has a unique extension to an order-continuous Riesz homomorphism T : Mµ¯0 → Mν¯0 (355F). (ii) If u0 ∈ Mµ¯0 and α > 0, then [[T u0 > α]] = [[T u+ 0 > α]] (because T is a Riesz homomorphism) =

sup

[[T u > α]]

u∈S(Af ),0≤u≤u+ 0

(because T is order-continuous and S(Af ) is order-dense in Mµ¯0 ) = π[[u0 > α]] by the argument used in (i). (iii) I have already remarked, at the beginning of the proof of (i), that T (u × u0 ) = T u × T u0 for u, u ∈ S(Af ). Because both T and × are order-continuous and S(Af ) is order-dense in Mµ¯0 , 0

T (u0 × u1 ) = sup{T (u × u0 ) : u, u0 ∈ S(Af ), 0 ≤ u ≤ u0 , 0 ≤ u0 ≤ u1 } = sup T u × T u0 = T u0 × T u1 u,u0

Lp

366H

355

whenever u0 , u1 ≥ 0 in Mµ¯0 . Because T is linear and × is bilinear, it follows that T is multiplicative on Mµ¯0 . To see that it is injective, observe that if u 6= 0 in Mµ¯0 then there is some α > 0 such that a = [[|u| > α]] 6= 0, so that 0 < χπa ≤ T |u| = |T u| and T u 6= 0. α) Suppose that p ∈ [1, ∞[ and that u ∈ Lpµ¯ . Then for any α > 0, (iv)(α [[|T u|p > α]] = [[T |u| > α1/p ]] = π[[|u| > α1/p ]] = π[[|u|p > α]]. So k|T u|p k1 =

R∞ 0

R∞

ν¯[[|T u|p > α]] dα =

0

µ ¯[[|u|p > α]] dα = k|u|p k1 < ∞

and T u ∈ Lpν¯ , with kT ukp = kukp . β ) As for the case p = ∞, if u ∈ L∞ (β µ ¯ and γ = kuk∞ > 0 then [[|u| > γ]] = 0, so [[|T u| > γ]] = π[[|u| > γ]] = 0. This shows that kT uk∞ ≤ γ. On the other hand, if 0 < α < γ then a = [[|u| > α]] 6= 0, and χa ≤ |u| so χ(πa) ≤ |T u|; as πa 6= 0 (because ν¯(πa) = µ ¯a > 0), kT uk∞ > α. This shows that kT uk∞ = kuk∞ , at least when u 6= 0; but the case u = 0 is trivial. (γγ ) Now take any p ∈ [1, ∞], and suppose that u ∈ Mµ¯0 and that T u ∈ Lpν¯ . Let hun in∈N be a non-decreasing sequence in S(Af ) with supremum |u| and u0 ≥ 0 (366Gb). Then T un ≤ T u so kun kp = kT un kp ≤ kT ukp . But this means that hun in∈N is bounded above in Lpµ¯ (366Db), so that |u| and u belong to Lpµ¯ . (δδ ) If u ∈ L1µ¯ , then

R

T u = k(T u)+ k1 − k(T u)− k1 = kT u+ k1 − kT u− k1 = ku+ k1 − ku− k1 =

R

u.

P Set a = [[|u| > ²]] ∈ Af . Then (v) If u ∈ Mµ¯1,0 and ² > 0, then T (|u| ∧ ²χ1A ) = |T u| ∧ ²χ1B . P |u| ∧ ²χ1A = ²χa + |u| − |u| × χa and [[|T u| > ²]] = πa. So T (|u| ∧ ²χ1A ) = T (²χa) + T |u| − T (|u| × χa) = ²χ(πa) + |T u| − |T u| × χ(πa) = |T u| ∧ ²χ1B . Q Q Consequently T (|u| − ²χ1A )+ = T (|u| − |u| ∧ ²χ1A ) = (|T u| − ²χ1B )+ . But this means that (|u| − ²χ1A )+ ∈ L1µ¯ iff (|T u| − ²χ1B )+ ∈ L1ν¯ . Since this is true for every ² > 0, 366Gc tells us that u ∈ Mµ¯1,0 iff T u ∈ Mν¯1,0 . α) By 365Pa, we have an order-continuous positive linear operator P0 : L1ν¯ → L1µ¯ such that (b)(i)(α R P v = πa v for every v ∈ L1ν¯ and a ∈ Af . a 0

R

β ) We now find that if v0 ≥ 0 in Mν¯1,0 and B = {v : v ∈ L1ν¯ , 0 ≤ v ≤ v0 }, then P0 [B] has a (β P Because B is upwards-directed and P0 is order-preserving, supremum in L0 (A) which belongs to Mµ¯1,0 . P P0 [B] is upwards-directed. If α > 0 and v ∈ B and a = [[P0 v > α]], then α 2

v ≤ (v0 − α2 χ1B )+ + χ1B , so Z

Z

α¯ µa ≤

P0 v = Z

=

a

Z

α 2

πa

α 2

α 2

(v0 − χ1B )+ + ν¯(πa)

v≤

α 2

(v0 − χ1B )+ + µ ¯a

and µ ¯[[P0 v > α]] ≤

2 α

R

(v0 − α2 χ1B )+ .

Thus {[[P0 v > α]] : v ∈ B} is an upwards-directed set in Af with measures bounded above in R, and cα = supv∈B [[P0 v > α]] is defined in Af . Also

356

Function spaces

inf n≥1 cn ≤ inf n≥1

2 n

R

366H

(v0 − n2 χ1B )+ = 0.

So, by 364Mb, P0 [B] has a supremum u0 ∈ L0 (A), and [[u0 > α]] = cα ∈ Af for every α > 0, so u0 ∈ Mµ¯0 . If c ∈ Af , then

R

c

u0 = supv∈B

R

c

P0 v = supv∈B

R

πc

v≤

R

πc

v0 < ∞,

so u0 ∈ Mµ¯1,0 . Q Q (γγ ) Now 355F tells us that P0 has a unique extension to an order-continuous positive linear operator P : Mν¯1,0 → Mµ¯1,0 . If v0 ≥ 0 in Mν¯1,0 and a ∈ Af , then, as remarked above,

R

R

R

P v0 = sup{ a P0 v : v ∈ L1ν¯ , 0 ≤ v ≤ v0 } = sup{ πa v : v ∈ L1ν¯ , 0 ≤ v ≤ v0 } = a R R because P is linear, a P v = πa v for every v ∈ Mν¯1,0 , a ∈ Af .

R

πa

v0 ;

(δδ ) By 366Gd, P is uniquely defined by the formula

R

R

Pv = a

πa

v whenever v ∈ Mν¯1,0 , a ∈ Af .

(ii) Because Mµ¯0 is closed under multiplication, u × P v certainly belongs to Mµ¯0 . α) Suppose (α that u, v ≥ 0. Fix c ∈ Af for the moment. Suppose that u0 ∈ S(Af ). Then we can Pn 0 express u as i=0 αi χai where ai ∈ Af for every i ≤ n. Accordingly

R

u0 × P v = c

Pn

α i=0 i

R

c∩ai

Pv =

Pn

α i=0 i

R

v × χ(πai ) × χ(πc) =

R

πc

v × T u0 .

Next, we can find a non-decreasing sequence hun in∈N in S(Af )+ with supremum u, and Z Z Z sup un × P v = sup v × T un = sup v × T un n∈N c n∈N πc n∈N Z Zπc = v × sup T un = v × T u, n∈N

πc

R

πc

using the order-continuity of T , and ×. But this means that u × P v = supn∈N un × P v is integrable over R R c and that c u × P v = πc v × T u. As c is arbitrary, u × P v = P (v × T u) ∈ Mµ¯1,0 . β ) For general u, v, (β v + × T u+ + v + × T u− + v − × T u+ + v − × T u− = |v| × T |u| = |v × T u| ∈ Mν¯1,0 (because T is a Riesz homomorphism), so we may apply (α) to each of the four products; combining them, we get P (v × T u) = u × P v, as required. (iii) Because P is a positive operator, we surely have |P v| ≤ P |v|, so it will be enough to show that kP vkq ≤ kvkq for v ≥ 0 in Lqν¯ . R R α) I take the case q = 1 first. (α In this case, for any a ∈ Af , we have a P v = πa v ≤ kvk1 . In R particular, setting an = [[P v > 2−n ]], an P v ≤ kvk1 . But P v = supn∈N P v × χan , so kP vk1 = supn∈N

then

R

an

P v ≤ kvk1 .

+ β ) Next, suppose that q = ∞, so that v ∈ (L∞ (β ? If γ > 0 and a = [[P v > γ]] 6= 0, ν ¯ ) ; say kvk∞ = γ. ?

γµ ¯a <

R a

Pv =

R πa

v ≤ γ ν¯(πa) = γ µ ¯a. X X

So [[P v > γ]] = 0 and P v ∈ L∞ µ ¯ , with kP vk∞ ≤ kvk∞ , at least when kvk∞ > 0; but the case kvk∞ = 0 is trivial. (γγ ) I come at last to the ‘general’ case q ∈ ]1, ∞[, v ∈ Lqν¯ . In this case set p = q/(q − 1). If u ∈ Lpµ¯ then T u ∈ Lpν¯ so T u × v ∈ L1ν¯ and

Lp

366J

357

Z |

u × P v| ≤ ku × P vk1 = kP (T u × v)k1

(by (ii)) ≤ kT u × vk1 (by (α) just above)

Z =

|T u| × |v| ≤ kT ukp kvkq = kukp kvkq R by (a-iii) of this theorem. But this means that u 7→ u × P v is a bounded linear functional on Lpµ¯R, and R q p is therefore represented by some w ∈ Lµ¯ with kwkq ≤ kvkq . If a ∈ Af then χa ∈ Lµ¯ , so a w = a P v; accordingly P v is actually equal to w (by 366Gd) and kP vkq = kwkq ≤ kvkq , as claimed. (iv) If u ∈ Mµ¯1,0 and a ∈ Af , we must have

R

R

R

R

R

R

P T u = πa T u = T (χa) × T u = T (χa × u) = χa × u = a u, a R R using (a-iv) to see that χa × u is defined and equal to T (χa × u). As a is arbitrary, u ∈ Mµ¯1,0 and P T u = u. (c) As usual, in view of the uniqueness of Tθπ and Pθπ , all we have to check is that Tθ T (χa) = Tθ χ(πa) = χ(θπa) = Tθπ (χa),

R

P Pθ w = a

R

P w= πa θ

R

w= θπa

R a

Pθπ w

whenever a ∈ Af , w ∈ Mλ¯1,0 . (d)(i) By (c), Tπ−1 T = Tπ−1 π must be the identity operator on Mµ¯0 ; similarly, T Tπ−1 is the identity operator on Mν¯0 . Because T and Tπ−1 are Riesz homomorphisms, they must be the two halves of a Riesz space isomorphism. (ii) In the same way, P and Pπ−1 must be the two halves of an ordered linear space isomorphism between Mµ¯1,0 and Mν¯1,0 , and are therefore both Riesz homomorphisms. (iii) By (b-iv), P T u = u for every u ∈ Mµ¯1,0 , so T ¹Mµ¯1,0 must be P −1 . Similarly P = Pπ−1 −1 is the restriction of T −1 = Tπ−1 to Mν¯1,0 . (iv) Because T −1 [Lpν¯ ] = Lpµ¯ (by (a-iv)), and T is a bijection between Mµ¯0 and Mν¯0 , T ¹Lpµ¯ must be a Riesz space isomorphism between Lpµ¯ and Lpν¯ ; (a-iv) also tells us that it is norm-preserving. Now its inverse is P ¹Lpν¯ , by (iii) here. 366I Corollary Let (A, µ ¯) be a measure algebra, and B a σ-subalgebra of A. Then, for any p ∈ [1, ∞[, Lp (B, µ ¯¹ B) can be identified, as Banach lattice, with the closed linear subspace of Lp (A, µ ¯) generated by {χb : b ∈ B, µ ¯b < ∞}. proof The identity map b 7→ b : B → A induces an injective Riesz homomorphism T : L0 (B) → L0 (A) (364R) such that T u ∈ LpA = Lp (A, µ ¯) and kT ukp = kukp whenever p ∈ [1, ∞[ and u ∈ LpB = Lp (B, µ ¯¹ B) f (366H(a-iv)). Because S(B ), the linear span of {χb : b ∈ B, µ ¯b < ∞}, is dense in LpB (366C), the image of LpB in LpA must be the closure of the image of S(Bf ) in LpA , that is, the closed linear span of {χb : b ∈ Bf } interpreted as a subset of LpA . 366J Corollary If (A, µ ¯) is a probability algebra, B is a closed subalgebra of A, and P : L1 (A, µ ¯) → L1 (B, µ ¯¹ B) is the conditional expectation operator (365R), then kP ukp ≤ kukp whenever p ∈ [1, ∞] and u ∈ Lp (A, µ ¯). proof Because (A, µ ¯) is totally finite, M 0 (A, µ ¯) = L1 (A, µ ¯), so that the operator P of 366Hb can be identified with the conditional expectation operator of 365R. Now 366H(b-iii) gives the result. Remark Of course this is also covered by 244M.

358

Function spaces

366K

366K Corollary Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : Af → Bf a measure-preserving ring homomorphism. Let T : L2µ¯ → L2ν¯ and P : L2ν¯ → L2µ¯ be the corresponding operators, as in 366H. Then T P : L2ν¯ → L2ν¯ is an orthogonal projection, its range T P [L2ν¯ ] being isomorphic, as Banach lattice, to L2µ¯ . The kernel of T P is just

R

{v : v ∈ L2ν¯ ,

πa

v = 0 for every a ∈ Af }.

proof Most of this is simply because T is a norm-preserving Riesz homomorphism (so that T [L2µ¯ ] is isomorphic to L2µ¯ ), P T is the identity on L2µ¯ (so that (T P )2 = T P ) and kP k ≤ 1 (so that kT P k ≤ 1). These are enough to ensure that T P is a projection of norm at most 1, that is, an orthogonal projection. Also Z P v = 0 for every a ∈ Af T P v = 0 ⇐⇒ P v = 0 ⇐⇒ a Z ⇐⇒ v = 0 for every a ∈ Af . πa

366L Corollary Let (A, µ ¯) be a measure algebra, and π : Af → Af a measure-preserving ring automorphism. Then there is a corresponding Banach lattice isomorphism T of L2 = L2 (A, µ ¯) defined by writing T (χa) = χ(πa) for every a ∈ Af . Its inverse is defined by the formula

R

T −1 u = a

R

πa

u for every u ∈ L2 , a ∈ Af .

proof In the language of 366H, T = Tπ and T −1 = P . 366X Basic exercises (a) Let (A, µ ¯) be a measure algebra and p ∈ ]1, ∞[. Show that, for u ∈ L0 (A), R ∞ p−1 p u ∈ L (A, µ ¯) iff γ = p 0 α µ ¯[[|u| > α]] dα is finite, and that in this case kukp = γ 1/p . (Cf. 263Xa.) > (b) Let (A, µ ¯) be a localizable measure algebra and p ∈ [1, ∞]. Show that the band algebra of Lp (A, µ ¯) is isomorphic to A. (Cf. 365S.) (c) Let (A, µ ¯) be a measure algebra and p ∈ ]1, ∞[. Show that Lp (A, µ ¯) is separable iff L1 (A, µ ¯) is. (d) Let (A, µ ¯) be a measure algebra. (i) Show that L∞ (A) ∩ M 0 (A, µ ¯) and L∞ (A) ∩ M 1,0 (A, µ ¯), as defined ∞,0 in 366G, are equal. (ii) Call this intersection M (A, µ ¯). Show that it is a norm-closed solid linear subspace of L∞ (A), therefore a Banach lattice in its own right. b µ (e) Let (A, µ ¯) be a semi-finite measure algebra and (A, ˆ) its localization (322P). Show that the natural b b µ embedding of A in A induces a Banach lattice isomorphism between Lp (A, µ ¯) and Lp (A, ˆ) for every p ∈ b [1, ∞[, so that the band algebra of Lp (A, µ ¯) can be identified with A. b µ (f ) Let (A, µ ¯) be a semi-finite measure algebra which is not localizable (cf. 211Ya, 216D), and (A, ˆ) its b be the identity embedding, so that π is an order-continuous measure-preserving localization. Let π : A → A ∞ b RBooleanR homomorphism. Show that if we set v = χb where b ∈ A \ A, then there is no u ∈ L (A) such that u = v whenever µ ¯ a < ∞. a πa (g) In 366H, show that [[T u ∈ E]] = π[[u ∈ E]] whenever u ∈ Mµ¯0 and E ⊆ R is a Borel set such that 0 ∈ / E. > (h) Let (A, µ ¯) be a measure algebra and let G be the group of all measure-preserving ring automorphisms of Af . Let H be the group of all Banach lattice automorphisms of L2 (A, µ ¯). Show that the map π 7→ T of 366L is an injective group homomorphism from G to H, so that G is represented as a subgroup of H. (i) Let h(Ai , µ ¯i )ii∈I be any family of measure algebras, with simple product (A, µ ¯) (322K). Show that for any p ∈ [1, ∞[, Lp (A, µ ¯) can be identified, as normed Riesz space, with the solid linear subspace ¡P ¢ p 1/p < ∞} {u : kuk = i∈I ku(i)k Q of i∈I Lp (Ai , µ ¯i ).

366 Notes

Lp

359

(j) Let A be a Dedekind σ-complete Boolean algebra and µ ¯, ν¯ two functionals rendering A a semi-finite measure algebra. Show that for any p ∈ [1, ∞[, Lp (A, µ ¯) and Lp (A, ν¯) are isomorphic as normed Riesz spaces. (Hint: use 366Xe to reduce to the case in which A is Dedekind complete. Take w ∈ L0 (A) such that R w d¯ µ = ν¯a for every a ∈ A (365T). Set T u = w1/p × u for u ∈ Lp (A, µ ¯).) a (k) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras, and p ∈ [1, ∞[. Show that the following are equiveridical: (i) Lpµ¯ and Lpν¯ are isomorphic as Banach lattices; (ii) Lpµ¯ and Lpν¯ are isomorphic as Riesz spaces; (iii) A and B have isomorphic Dedekind completions. 366Y Further exercises (a) Let (A, µ ¯) be a measure algebra and suppose that 0 < p < 1. Write R Lp = Lp (A, µ ¯) for {u : u ∈ L0 (A), |u|p ∈ L1 (A, µ ¯)}, and for u ∈ Lp set τ (u) = |u|p . (i) Show that τ defines a Hausdorff linear space topology on Lp (see 2A5B). (ii) Show that if A ⊆ Lp is non-empty, downwardsdirected and has infimum 0 then inf u∈A τ (u) = 0. (iii) Show that if A ⊆ Lp is non-empty, upwards-directed and bounded in the linear topological space sense (see 245Yf) then A is bounded above. (iv) Show that (Lp )∼ = (Lp )× is just the set of continuous linear functionals from Lp to R, and is {0} iff A is atomless. (b) Let (A, µ ¯) be a measure algebra. Show that M 0 (A, µ ¯) has the countable sup property. (c) Let (A, µ ¯) be a measure algebra and define M ∞,0 (A, µ ¯) as in 366Xd. Show that (M ∞,0 (A, µ ¯))× can 1 be identified with L (A, µ ¯). (d) In 366H, show that if T˜ : M 0 (A, µ ¯) → M 0 (B, ν¯) is any positive linear operator such that T˜(χa) = χ(πa) for every a ∈ Af , then T˜ is order-continuous, so that it is equal to Tπ . (e) Let (A, µ ¯) be a measure algebra. (i) Show that there is a natural one-to-one correspondence between M 1,0 (A, µ ¯) and the set of additive functionals ν : Af → R such that ν << µ in the double sense that for every ² > 0 there are δ, M > 0 such that |νa| ≤ ² whenever µa ≤ δ and |νa| ≤ ²µa whenever µa ≥ M . (ii) Use this description of M 1,0 to prove 366H(b-i). (f ) In 366H, show that the following are equiveridical: (α) π[Af ] = Bf ; (β) T = Tπ is surjective; (γ) P = Pπ is injective; (δ) P is a Riesz homomorphism; (²) there is some q ∈ [1, ∞] such that kP vkq = kvkq for every v ∈ Lqν¯ ; (ζ) T P v = v for every v ∈ Mν¯1,0 . (g) Let (A, µ ¯) and (B, ν¯) be measure algebras, and suppose that π : Af → Bf is a measure-preserving ring homomorphism, as in 366H; let T : M 0 (A, µ ¯) → M 0 (B, ν¯) be the associated linear operator. Show that p 0 if 0 < p < 1 (as in 366Ya) then L (A, µ ¯) ⊆ M (A, µ ¯) and T −1 [Lp (B, ν¯)] = Lp (A, µ ¯). (h) Let (A, µ ¯) be a totally finite measure algebra. (i) For each Boolean automorphism π : A → A, let 0 0 T : L (A) → L (A) be the associated Riesz space isomorphism (364R), and let wπ ∈ L1 (A)+ be such that π R √ −1 w = µ(π a) for every a ∈ A (365Ea). Set Qπ u = Tπ u × wπ for u ∈ L0 (A). Show that kQπ uk2 = kuk2 a π for every u ∈ L2 (A). (ii) Show that if π, φ : A → A are Boolean automorphisms then Qπφ = Qπ Qφ . 366 Notes and comments The Lp spaces, for 1 ≤ p ≤ ∞, constitute the most important family of leading examples for the theory of Banach lattices, and it is not to be wondered at that their properties reflect a wide variety of general results. Thus 366Dd and 366E can both be regarded as special cases of theorems about perfect Riesz spaces (356M and 369D). In a different direction, the concept of ‘Orlicz space’ (369Xd below) generalizes the Lp spaces if they are regarded as normed subspaces of L0 invariant under measure-preserving automorphisms of the underlying algebra. Yet another generalization looks at the (non-locally-convex) spaces Lp for 0 < p < 1 (366Ya). In 366H and its associated results I try to emphasize the way in which measure-preserving homomorphisms of the underlying algebras induce both ‘direct’ and ‘dual’ operators on Lp spaces. We have already seen the phenomenon in 365P. I express this in a slightly different form in 366H, noting that we really do need the homomorphisms to be measure-preserving, for the dual operators as well as the direct operators, so we no longer have the shift in the hypotheses which appears between 365O and 365P. Of course all these refinements in the hypotheses are irrelevant to the principal applications of the results, and they make

360

Function spaces

366 Notes

substantial demands on the reader; but I believe that the demands are actually demands to expand one’s imagination, to encom the different ways in which the spaces depend on the underlying measure algebras. In the context of 366H, L∞ is set apart from the other Lp spaces, because L∞ (A) is not in general determined by the ideal Af , and the hypotheses of 366H do not look outside Af . 366H(a-iv) and 366H(b-iii) reach only the space M ∞,0 as defined in 366Xd. To deal with L∞ we need slightly stronger hypotheses. If we are given a measure-preserving Boolean homomorphism from A to B, rather than from Af to Bf , then of course the direct operator T has a natural version acting on L∞ (A) and indeed on M 1,∞ (A, µ ¯), as in 363F and 369Xm. If we know that (A, µ ¯) is localizable, then A can be recovered from Af , and the dual operator P acts on L∞ (B), as in 369Xm. But in general we can’t expect this to work (366Xf). Of course 366H can be applied to many other spaces; for reasons which will appear in §§371 and 374, the archetypes are not really Lp spaces at all, but the spaces M 1,0 (366F) and M 1,∞ (369N). I include 366L and 366Yh as pointers to one of the important applications of these ideas: the investigation of properties of a measure-preserving homomorphism in of its action on Lp spaces. The case p = 2 is the most useful because the group of unitary operators (that is, the normed space automorphisms) of L2 has been studied intensively.

367 Convergence in measure Continuing through the ideas of Chapter 24, I come to ‘convergence in measure’. The basic results of §245 all translate easily into the new language (367M-367N, 367Q). The associated concept of (sequential) order-convergence can also be expressed in abstract (367A), and I take the trouble to do this in the context of general lattices (367A-367B), since the concept can be applied in many ways (367C-367F, 367L, 367Xa-367Xm). In the particular case of L0 spaces, which are the first aim of this section, the idea is most naturally expressed by 367G. It enables us to express some of the fundamental theorems from Volumes 1 and 2 in the language of this chapter (367J-367K). In 367O and 367P I give two of the most characteristic properties of the topology of convergence in measure on L0 ; it is one of the fundamental types of topological Riesz space. Another striking fact is the way it is determined by the Riesz space structure (367T). In 367U I set out a theorem which is the basis of many remarkable applications of the concept; for the sake of a result in §369 I give one such application (367V). 367A Order*-convergence As I have remarked before, the function spaces of measure theory have three interdependent structures: they are linear spaces, they have a variety of interesting topologies, and they are ordered spaces. Ordinary elementary functional analysis studies interactions between topologies and linear structures, in the theory of normed spaces and, more generally, of linear topological spaces. Chapter 35 in this volume looked at interactions between linear and order structures. It is natural to seek to complete the triangle with a theory of topological ordered spaces. The relative obscurity of any such theory is in part due to the difficulty of finding convincing definitions; that is, isolating concepts which lead to in elegant and useful general theorems. Among the many rival ideas, however, I believe it is possible to identify one which is particularly important in the context of measure theory. In its natural home in the theory of L0 spaces, this notion of ‘order*-convergence’ has a very straightforward expression (367G). But, suitably interpreted, the same idea can be applied in other contexts, some of which will be very useful to us, and I therefore begin with a definition which is applicable in any lattice. Definition Let P be a lattice and hpn in∈N a sequence in P , p an element of P . I will say that hpn in∈N order*-converges to p, or that p is the order*-limit of hpn in∈N , if p = inf{q : ∃ n ∈ N, q ≥ (p0 ∨ pi ) ∧ p00 ∀ i ≥ n} = sup{q : ∃ n ∈ N, q ≤ p0 ∨ (pi ∧ p00 ) ∀ i ≥ n} whenever p0 ≤ p ≤ p00 in P .

367B

Convergence in measure

361

367B Lemma Let P be a lattice. (a) A sequence in P can order*-converge to at most one point. (b) A constant sequence order*-converges to its constant value. (c) Any subsequence of an order*-convergent sequence is order*-convergent, with the same limit. (d) If hpn in∈N and hp0n in∈N both order*-converge to p, and pn ≤ qn ≤ p0n for every n, then hqn in∈N order*-converges to p. (e) If hpn in∈N is an order-bounded sequence in P , then it order*-converges to p ∈ P iff p = inf{q : ∃ n ∈ N, q ≥ pi ∀ i ≥ n} = sup{q : ∃ n ∈ N, q ≥ pi ∀ i ≥ n}. (f) If P is a Dedekind σ-complete lattice (314A) and hpn in∈N is an order-bounded sequence in P , then it order*-converges to p ∈ P iff p = supn∈N inf i≥n pi = inf n∈N supi≥n pi . proof (a) Suppose that hpn in∈N is order*-convergent to both p and p˜. Set p0 = p ∧ p˜, p00 = p ∨ p˜; then p = inf{q : ∃ n ∈ N, q ≥ (p0 ∨ pi ) ∧ p00 ∀ i ≥ n} = p˜. (b) is trivial. (c) Suppose that hpn in∈N is order*-convergent to p, and that hp0n in∈N is a subsequence of hpn in∈N . Take p , p00 such that p0 ≤ p ≤ p00 , and set 0

B = {q : ∃ n ∈ N, q ≤ p0 ∨ (pi ∧ p00 ) ∀ i ≥ n}, B 0 = {q : ∃ n ∈ N, q ≤ p0 ∨ (p0i ∧ p00 ) ∀ i ≥ n}, C = {q : ∃ n ∈ N, q ≥ (p0 ∨ pi ) ∧ p00 ∀ i ≥ n}, C 0 = {q : ∃ n ∈ N, q ≥ (p0 ∨ p0i ) ∧ p00 ∀ i ≥ n}. If q ∈ B 0 and q 0 ∈ C, then for all sufficiently large i q ≤ p0 ∨ (p0i ∧ p00 ) ≤ (p0 ∨ p0i ) ∧ p00 ≤ q 0 . As p = inf C, we must have q ≤ p; thus p is an upper bound for B 0 . On the other hand, {p0i : i ≥ n} ⊆ {pi : i ≥ n} for every n, so B ⊆ B 0 and p must be the least upper bound of B 0 , since p = sup B. Similarly, p = inf C 0 . As p0 and p00 are arbitrary, hp0n in∈N order*-converges to p. (d) Take p0 , p00 such that p0 ≤ p ≤ p00 , and set B = {q : ∃ n ∈ N, q ≤ p0 ∨ (pi ∧ p00 ) ∀ i ≥ n}, B 0 = {q : ∃ n ∈ N, q ≤ p0 ∨ (qi ∧ p00 ) ∀ i ≥ n}, C = {q : ∃ n ∈ N, q ≥ (p0 ∨ p0i ) ∧ p00 ∀ i ≥ n}, C 0 = {q : ∃ n ∈ N, q ≥ (p0 ∨ qi ) ∧ p00 ∀ i ≥ n}. If q ∈ B 0 and q 0 ∈ C, then for all sufficiently large i q ≤ p0 ∨ (qi ∧ p00 ) ≤ (p0 ∨ p0i ) ∧ p00 ≤ q 0 . As p = inf C, we must have q ≤ p; thus p is an upper bound for B 0 . On the other hand, p0 ∨ (pi ∧ p00 ) ≤ p0 ∨ (qi ∧ p00 ) for every i, so B ⊆ B 0 and p = sup B 0 . Similarly, p = inf C 0 . As p0 and p00 are arbitrary, hqn in∈N order*-converges to p. (e) Set B = {q : ∃ n ∈ N, q ≤ pi ∀ i ≥ n}, C = {q : ∃ n ∈ N, q ≥ pi ∀ i ≥ n}.

362

Function spaces

367B

(i) Suppose that hpn in∈N order*-converges to p. Let p0 , p00 be such that p0 ≤ pn ≤ p00 for every n ∈ N and p0 ≤ p ≤ p00 . Then B = {q : ∃ n ∈ N, q ≤ p0 ∨ (pi ∧ p00 ) ∀ i ≥ n}, so sup B = p. Similarly, inf C = p, so the condition is satisfied. (ii) Suppose that sup B = inf C = p. Take any p0 , p00 such that p0 ≤ p ≤ p00 and set B 0 = {q : ∃ n ∈ N, q ≤ p0 ∨ (pi ∧ p00 ) ∀ i ≥ n}, C 0 = {q : ∃ n ∈ N, q ≥ (p0 ∨ pi ) ∧ p00 ∀ i ≥ n}. If q ∈ B 0 and q 0 ∈ C, then for all large enough i q ≤ p0 ∨ (pi ∧ p00 ) ≤ p0 ∨ q 0 = q 0 because p ≤ q 0 . As inf C = p, p is an upper bound for B 0 . On the other hand, if q ∈ B, then q ≤ p, so q ≤ p0 ∨ (pi ∧ p00 ) whenever q ≤ pi , which is so for all sufficiently large i, and q ∈ B 0 . Thus B 0 ⊇ B and p must be the supremum of B 0 . Similarly, p = inf C 0 ; as p0 and p00 are arbitrary, hpn in∈N order*-converges to p. (f ) This follows at once from (e). Setting B = {q : ∃ n ∈ N, q ≤ pi ∀ i ≥ n}, B 0 = {inf i≥n pi : i ∈ N}, then B 0 ⊆ B and for every q ∈ B there is a q 0 ∈ B 0 such that q ≤ q 0 ; so sup B = sup B 0 if either is defined. Similarly, inf{q : ∃ n ∈ N, q ≥ pi ∀ i ≥ n} = inf n∈N supi≥n pi if either is defined. 367C Proposition Let U be a Riesz space and hun in∈N , hvn in∈N two sequences in U order*-converging to u, v respectively. (a) If w ∈ U , hun + win∈N order*-converges to u + w, and αun order*-converges to αu for every α ∈ R. (b) hun ∨ vn in∈N order*-converges to u ∨ v. (c) If hwn in∈N is any sequence in U , then it order*-converges to w ∈ U iff h|wn − w|in∈N order*-converges to 0. (d) hun + vn in∈N order*-converges to u + v. (e) If U is Archimedean, and hαn in∈N is a sequence in R converging to α ∈ R, then hαn un in∈N order*converges to αu. (f) Again suppose that U is Archimedean. Then a sequence hwn in∈N in U + is not order*-convergent to 0 iff there is a w ˜ > 0 such that w ˜ = supi≥n w ˜ ∧ wi for every n ∈ N. proof (a)(i) hun + win∈N order*-converges to u + w because the ordering of U is translation-invariant; the map w0 7→ w0 + w is an order-isomorphism. α) If α > 0, then the map w0 7→ αw0 is an order-isomorphism, so hαun in∈N order*-converges to (ii)(α αu. β ) If α = 0 then hαun in∈N order*-converges to αu = 0 by 367Bb. (β (γγ ) If w0 ≤ −u ≤ w00 then −w00 ≤ u ≤ w0 so u = inf{w : ∃ n ∈ N, w ≥ ((−w00 ) ∨ ui ) ∧ (−w0 ) ∀ i ≥ n} = sup{w : ∃ n ∈ N, w ≤ (−w00 ) ∨ (ui ∧ (−w0 )) ∀ i ≥ n}. Turning these formulae upside down, −u = sup{w : ∃ n ∈ N, w ≤ (w00 ∧ (−ui )) ∨ w0 ∀ i ≥ n} = inf{w : ∃ n ∈ N, w ≥ w00 ∧ ((−ui ) ∨ w0 ) ∀ i ≥ n}.

367C


363

As w0 and w00 are arbitrary, h−un in∈N order*-converges to −u. (δδ ) Putting (α) and (γ) together, hαun in∈N order*-converges to αu for every α < 0. (b) Suppose that w0 ≤ u ∨ v ≤ w00 . Set B = {w : ∃ n ∈ N, w ≤ w0 ∨ ((ui ∨ vi ) ∧ w00 ) ∀ i ≥ n}, C = {w : ∃ n ∈ N, w ≥ (w0 ∨ (ui ∨ vi )) ∧ w00 ∀ i ≥ n}, B1 = {w : ∃ n ∈ N, w ≤ (w0 ∧ u) ∨ (ui ∧ w00 ) ∀ i ≥ n}, B2 = {w : ∃ n ∈ N, w ≤ (w0 ∧ v) ∨ (vi ∧ w00 ) ∀ i ≥ n}, C1 = {w : ∃ n ∈ N, w ≥ ((w0 ∧ u) ∨ ui ) ∧ w00 ∀ i ≥ n}, C2 = {w : ∃ n ∈ N, w ≥ ((w0 ∧ v) ∨ vi ) ∧ w00 ∀ i ≥ n}, If w1 ∈ B1 and w2 ∈ B2 then w1 ∨ w2 ∈ B. P P There is an n ∈ N such that w1 ≤ (w0 ∧ u) ∨ (ui ∧ w00 ) for every i ≥ n, while w2 ≤ (w0 ∧ v) ∨ (vi ∧ w00 ) for every i ≥ n. So w1 ∨ w2 ≤ (w0 ∧ u) ∨ (w0 ∧ v) ∨ (ui ∧ w00 ) ∨ (vi ∧ w00 ) = (w0 ∧ (u ∨ v)) ∨ ((ui ∨ vi ) ∧ w00 ) (by the distributive law 352Ec) = w0 ∨ ((ui ∨ vi ) ∧ w00 ) for every i ≥ n, and w1 ∨ w2 ∈ B. Q Q Similarly, if w1 ∈ C1 and w2 ∈ C2 then w1 ∨w2 ∈ C. P P There is an n ∈ N such that w1 ≥ ((w0 ∧u)∨ui )∧w00 , 0 00 w2 ≥ ((w ∧ v) ∨ vi ) ∧ w for every i ≥ n. So w1 ∨ w2 ≥ (((w0 ∧ u) ∨ ui ) ∧ w00 ) ∨ (((w0 ∧ v) ∨ vi ) ∧ w00 ) = ((w0 ∧ u) ∨ ui ∨ (w0 ∧ v) ∨ vi ) ∧ w00 = ((w0 ∧ (u ∨ v)) ∨ (ui ∨ vi )) ∧ w00 = (w0 ∨ (ui ∨ vi )) ∧ w00 for every i ≥ n, so w1 ∨ w2 ∈ C. Q Q At the same time, of course, w ≤ w ˜ whenever w ∈ B, w ˜ ∈ C, since there is some i ∈ N such that w ≤ w0 ∨ ((ui ∨ vi ) ∧ w00 ) ≤ (w0 ∨ (ui ∨ vi )) ∧ w00 ≤ w. ˜ Since sup{w1 ∨ w2 : w1 ∈ B1 , w2 ∈ B2 } = (sup B1 ) ∨ (sup B2 ) = u ∨ v, inf{w1 ∨ w2 : w1 ∈ C1 , w2 ∈ C2 } = (inf C1 ) ∨ (inf C2 ) = u ∨ v (using the generalized distributive laws in 352E), we must have sup B = inf C = u ∨ v. As w0 and w00 are arbitrary, hun ∨ vn in∈N is order*-convergent to u ∨ v. (c) The hard parts are over. (i) If hwn in∈N order*-converges to w, then hwn − win∈N , hw − wn in∈N and h|wn − w|in∈N = h(wn − w) ∨ (w − wn )in∈N all order*-converge to 0, putting (a) and (b) together. (ii) If h|wn − w|in∈N order*-converges to 0, then so do h−|wn − w|in∈N and hwn − win∈N , by (a) and 367Bd; so hwn in∈N order*-converges to 0, by (a) again. (d) h|un − u|in∈N and h|vn − v|in∈N order*-converge to 0, by (c), so h2(|un − u| ∨ |vn − v|)in∈N also order*-converges to 0, by (b) and (a). But 0 ≤ |(un + vn ) − (u + v)| ≤ |un − u| + |vn − v| ≤ 2(|un − u| ∨ |vn − v|) for every n, so h|(un + vn ) − (u + v)|in∈N order*-converges to 0, by 367Bb and 367Bd, and hun + vn in∈N order*-converges to u + v.

364

Function spaces

367C

(e) Set βn = supi≥n |αi −α| for each n. Then hβn in∈N → 0, so inf n∈N βn |u| = 0, because U is Archimedean. Consequently hβn |u|in∈N order*-converges to 0, by 367Be. But we also have β0 |un − u| order*-converging to 0, by (c) and (a), so hβ0 |un −u|+βn |u|in∈N order*-converges to 0, by (d). As |αn un −αu| ≤ β0 |un −u|+βn |u| for every n, hαn un in∈N order*-converges to αu, as required. (f )(i) Suppose that hwn in∈N is not order*-convergent to 0. Then there are w0 , w00 such that w0 ≤ 0 ≤ w00 and either B = {w : ∃ n ∈ N, w ≤ w0 ∨ (wi ∧ w00 ) ∀ i ≥ n} does not have supremum 0, or C = {w : ∃ n ∈ N, w ≥ (w0 ∨ wi ) ∧ w00 ∀ i ≥ n} does not have infimum 0. Now 0 ∈ B, because every wi ≥ 0, and every member of B is a lower bound for C; so 0 cannot be the greatest lower bound of C. Let w ˜ > 0 be a lower bound for C. Let n ∈ N, and set Cn = {w : w ≥ (w0 ∨ wi ) ∧ w00 ∀ i ≥ n} = {w : w ≥ wi ∧ w00 ∀ i ≥ n}. Because U is Archimedean, we know that inf(Cn − An ) = 0, where An = {wi ∧ w00 : i ≥ n} (353F). Now w ˜ is a lower bound for Cn , so inf (w ˜ − wi )+ ≤ inf{(w − wi )+ : w ∈ C, i ≥ n}

i≥n

≤ inf{(w − (wi ∧ w00 ))+ : w ∈ C, i ≥ n} = inf{w − (wi ∧ w00 ) : w ∈ C, i ≥ n} = inf(Cn − An ) = 0. As this is true for every n ∈ N, w ˜ has the property declared. (ii) If w ˜ > 0 is such that w ˜ = supi≥n w ˜ ∧ wi for every n ∈ N, then {w : ∃ n ∈ N, w ≥ (0 ∨ wi ) ∧ w ˜ ∀ i ≥ n} cannot have infimum 0, and hwn in∈N is not order*-convergent to 0. 367D

As an example of the use of this concept in a relatively abstract setting, I offer the following.

Proposition Let U be a Banach lattice and hun in∈N a sequence in U which is norm-convergent to u ∈ U . Then hun in∈N has a subsequence which is order-bounded and order*-convergent to u. So if hun in∈N itself is order*-convergent, its order*-limit is u. proof Let hu0n in∈N be a subsequence of hun in∈N such that ku0n − uk ≤ 2−n for each n ∈ N. Then vn = supi≥n |u0i − u| is defined in U , and kvn k ≤ 2−n+1 , for each n (354C). Because inf n∈N kvn k = 0, inf n∈N vn must be 0, while u − vn ≤ u0i ≤ u + vn whenever i ≥ n; so hu0n in∈N order*-converges to u, by 367Be. Now if hun in∈N has an order*-limit, this must be u, by 367Ba and 367Bc. 367E Proposition Let U be a Riesz space with an order-continuous norm. Then any order-bounded order*-convergent sequence is norm-convergent. proof Suppose that hun in∈N is order*-convergent to u. Then h|un − u|in∈N is order*-convergent to 0 (367Cc), so C = {v : ∃ n ∈ N, v ≥ |ui − u| ∀ i ≥ n} has infimum 0 (367Be). Because U is a lattice, C is downwards-directed, so inf v∈C kvk = 0. But inf v∈C kvk ≥ inf n∈N supi≥n kui − uk, so limn→∞ kun − uk = 0, that is, hun in∈N is norm-convergent to u.

367H


365

367F One of the fundamental obstacles to the development of any satisfying general theory of ordered topological spaces is the erratic nature of the relations between subspace topologies of order topologies and order topologies on subspaces. The particular virtue of order*-convergence in the context of function spaces is that it is relatively robust when transferred to the subspaces we are interested in. Proposition Let U be an Archimedean Riesz space and V a regularly embedded Riesz subspace. (For instance, V might be either solid or order-dense.) If hvn in∈N is a sequence in V and v ∈ V , then hvn in∈N order*-converges to v when regarded as a sequence in V , iff it order*-converges to v when regarded as a sequence in U . proof (a) Since, in either V or U , hvn in∈N order*-converges to v iff h|vn − v|in∈N order*-converges to 0 (367Cc), it is enough to consider the case vn ≥ 0, v = 0. (b) If hvn in∈N is not order*-convergent to 0 in U , then, by 367Cf, there is a u > 0 in U such that u = supi≥n u ∧ vi for every n ∈ N (the supremum being taken in U , of course). In particular, there is a k ∈ N such that u ∧ vk > 0. Now consider the set C = {w : w ∈ V, ∃ n ∈ N, w ≥ (0 ∨ vi ) ∧ vk ∀ i ≥ n}. Then for any w ∈ C, u ∧ vk = supi≥n u ∧ vi ∧ vk ≤ w, using the generalized distributive law in U , so 0 is not the greatest lower bound of C in U . But as the embedding of V in U is order-continuous, 0 is not the greatest lower bound of C in V , and hvn in∈N cannot be order*-convergent to 0 in V . (c) Now suppose that hvn in∈N is not order*-convergent to 0 in V . Because V also is Archimedean (351Rc), there is a w > 0 in V such that w = supi≥n w ∧ vi for every n ∈ N, the suprema being taken in V . Again because V is regularly embedded in U , we have the same suprema in U , so, by 367Cf in the other direction, hvn in∈N is not order*-convergent to 0 in U . 367G I now spell out the connexion between the definition above and the concepts introduced in 245C. Proposition Let X be a set, Σ a σ-algebra of subsets of X, A a Boolean algebra and π : Σ → A a sequentially order-continuous surjective Boolean homomorphism; let I be its kernel. Write L0 for the space of Σ-measurable real-valued functions on X, and let T : L0 → L0 = L0 (A) be the canonical Riesz homomorphism (364D, 364R). Then for any hfn in∈N , f in L0 , hT fn in∈N order*-converges to T f in L0 iff X \ {x : f (x) = limn→∞ fn (x)} ∈ I. proof Set H = {x : limn→∞ fn (x) exists = f (x)}; of course H ∈ Σ. Set gn (x) = |fn (x) − f (x)| for n ∈ N, x ∈ X. (a) If X \ H ∈ I, set hn (x) = supi≥n gi (x) for x ∈ H and hn (x) = 0 for x ∈ X \ H. Then hhn in∈N is a non-increasing sequence with infimum 0 in L0 , so inf n∈N T hn = 0 in L0 , because T is sequentially order-continuous (364Ra). But as \H ∈ I, T hn ≥ T gi = |T fi − T f | whenever i ≥ n, so h|T fn − T f |in∈N order*-converges to 0, by 367Be or 367Bf, and hT fn in∈N order*-converges to T f , by 367Cc. (b) Now suppose that hT fn in∈N order*-converges to T f . Set gn0 (x) = min(1, gn (x)) for n ∈ N, x ∈ X; then hT gn0 in∈N = he ∧ |T fn − T f |in∈N order*-converges to 0, where e = T (χX). By 367Bf, inf n∈N supi≥n T gi0 = 0 in L0 . But T is a sequentially order-continuous Riesz homomorphism, so T (inf n∈N supi≥n gi0 ) = 0, that is, X \ H = {x : inf n∈N supi≥n gi0 > 0} belongs to I. 367H Corollary Let A be a Dedekind σ-complete Boolean algebra. (a) Any order*-convergent sequence in L0 = L0 (A) is order-bounded. (b) If hun in∈N is a sequence in L0 , then it is order*-convergent to u ∈ L0 iff u = inf n∈N supi≥n ui = supn∈N inf i≥n ui .

366

Function spaces

367H

proof (a) We can express A as a quotient Σ/I of a σ-algebra of sets, in which case L0 can be identified with the canonical image of L0 = L0 (Σ) (364D). If hun in∈N is an order*-convergent sequence in L0 , then it is expressible as hT fn in∈N , where T : L0 → L0 is the canonical map, and 367G tells us that hfn (x)in∈N converges for every x ∈ H, where X \ H ∈ I. If we set h(x) = supn∈N |fn (x)| for x ∈ H, 0 for x ∈ X \ H, then we see that |un | ≤ T h for every n ∈ N, so that hun in∈N is order-bounded in L0 . (b) This now follows from 367Be, because L0 is Dedekind σ-complete. 367I Proposition Suppose that E ⊆ R is a Borel set and h : E → R is a continuous function. Let A be a Dedekind σ-complete Boolean algebra and set QE = {u : u ∈ L0 , [[u ∈ E]] = 1}, where L0 = L0 (A). ¯ : QE → L0 be the function defined by h (364I). Then hh(u ¯ n )in∈N order*-converges to h(u) ¯ Let h whenever hun in∈N is a sequence in QE order*-converging to u ∈ QE . proof This is an easy consequence of 367G. We can represent A as Σ/I where Σ is a σ-algebra of subsets of some set X and I is a σ-ideal of Σ (314M); let T : L0 → L0 (A) be the corresponding homomorphism (364D, 367G). Now we can find Σ-measurable functions hfn in∈N , f such that T fn = un , T f = u, as in 367G; and the hypothesis [[un ∈ E]] = 1, [[u ∈ E]] = 1 means just that, adjusting fn and f on a member of I if necessary, we can suppose that fn (x), f (x) ∈ E for every x ∈ X. (I am ing over the trivial case ¯ n ) = T (hfn ), h(u) ¯ E = ∅, X ∈ I, A = {0}.) Accordingly h(u = T (hf ), and (because h is continuous) {x : h(f (x)) 6= limn→∞ h(fn (x))} ⊆ {x : f (x) 6= limn→∞ fn (x)} ∈ I, ¯ ¯ so hh(un )in∈N order*-converges to h(u). 367J Dominated convergence We now have a suitable language in which to express an abstract version of Lebesgue’s Dominated Convergence Theorem. Theorem Let (A, µ ¯) be a measure algebra. If hun in∈N is a sequence in L1 = L1 (A, µ ¯) which R is order-bounded R and order*-convergent in L1 , then hun in∈N is norm-convergent to u in L1 ; in particular, u = limn→∞ un . R proof The norm Rof L1 is order-continuous (365C), so hun in∈N is norm-convergent to u, by 367E. As is R norm-continuous, u = limn→∞ un . 367K The Martingale Theorem In the same way, we can re-write theorems from §275 in this language. Theorem Let (A, µ ¯) be a probability algebra, and hBn in∈N a non-decreasing sequence of closed subalgebras 0 of A. For each n ∈ N let Pn : L1 = L1µ¯ → L1 ∩ L S (Bn ) be the conditional expectation operator (365R); let B be the closed subalgebra of A generated by n∈N Bn , and P the conditional expectation operator onto L1 ∩ L0 (B). (a) If hun in∈N is a norm-bounded sequence in L1 such that Pn (un+1 ) = un for every n ∈ N, then hun in∈N is order*-convergent in L1 . (b) If u ∈ L1 then hPn uin∈N is order*-convergent to P u. proof If we represent (A, µ ¯) as the measure algebra of a probability space, these become mere translations of 275G and 275I. (Note that this argument relies on the description of order*-convergence in L0 in of a.e. convergence of functions, as in 367G; so that we need to know that order*-convergence in L1 matches order*-convergence in L0 , which is what 367F is for.) 367L Some of the most important applications of these ideas concern spaces of continuous functions. I do not think that this is the time to go very far along this road, but one particular fact will be useful in §376. Proposition Let X be a locally compact Hausdorff space, and hun in∈N a sequence in C(X), the space of continuous real-valued functions on X. Then hun in∈N order*-converges to 0 in C(X) iff {x : x ∈ X, lim supn→∞ |un (x)| > 0} is meager. In particular, hun in∈N order*-converges to 0 if limn→∞ un (x) = 0 for every x.

367N


367

proof (a) The following elementary fact is worth noting: if A ⊆ C(X)+ is non-empty and inf A = 0 in S C(X), then G = u∈A {x : u(x) < ²} is dense for every ² > 0. P P?? If not, take x0 ∈ X \ G. Because X is completely regular (3A3Bb), there is a continuous function w : X → [0, 1] such that w(x0 ) = 1 and w(x) = 0 for every x ∈ G. But in this case 0 < ²w ≤ u for every u ∈ A, which is impossible. X XQ Q (b) Suppose that hun in∈N order*-converges to 0. Set vn = |un | ∧ χX, so that hvn in∈N order*-converges to 0 (using 367C, as usual). Set B = {v : v ∈ C(X), ∃ n ∈ N, vi ≤ v ∀ i ≥ n}, S so that inf B = 0 in C(X) (367Be).SFor each k ∈ N, set Gk = v∈B {x : v(x) < 2−k }; then Gk is dense, by (a), and of course is open. So H = k∈N X \ Gk is a countable union of nowhere dense sets and is meager. But this means that {x : lim sup |un (x)| > 0} = {x : lim sup vn (x) > 0} n→∞

n→∞

⊆ {x : inf v(x) > 0} ⊆ H v∈B

is meager. (c) Now suppose that hun in∈N does not order*-converge to 0. By 367Cf, there is a w > 0 in C(X) such that w = supi≥n w ∧ |ui | for every n ∈ N; that is, inf i≥n (w − |ui |)+ = 0 for every n. Set Gn = {x : inf i≥n (w − |ui |)+ (x) < 2−n } = {x : supi≥n |ui (x)| > w(x) − 2−n } for each n. Then H=

T n∈N

Gn = {x : lim supn→∞ un (x) ≥ w(x)}

is the intersection of a sequence of dense open sets, and its complement is meager. Let G be the non-empty open set {x : w(x) > 0}. Then G is not meager, by Baire’s theorem (3A3Ha); so G ∩ H cannot be meager. But {x : lim supn→∞ |un (x)| > 0} includes G ∩ H, so is also not meager. Remark Unless the topology of X is discrete, C(X) is not regularly embedded in RX , and we expect to find sequences in C(X) which order*-converge to 0 in C(X) but not in RX . But the proposition tells us that if we have a sequence in C(X) which order*-converges in RX to a member of C(X), then it order*-converges in C(X). 367M Everything above concerns a particular notion of sequential convergence. There is inevitably a suggestion that there ought to be a topological interpretation of this convergence (see 367Yc, 367Yl), but I have taken care to avoid spelling one out. I come now to something which really is a topology, and is as closely involved with order-convergence as any. Convergence in measure Let (A, µ ¯) be a measure algebra. For a ∈ Af = {a : µ ¯a < ∞} and u ∈ L0 = R 0 L (A) set τa (u) = |u| ∧ χa, τa² (u) = µ ¯(a ∩ [[|u| > ²]]). Then the topology of convergence in measure on L0 is defined either as the topology generated by the pseudometrics (u, v) 7→ τa (u − v) or by saying that G ⊆ L0 is open iff for every u ∈ G there are a ∈ Af , ² > 0 such that v ∈ G whenever τa² (u − v) ≤ ². Remark The sentences above include a number of assertions which need proving. But at this point, rather than write out any of the relevant arguments, I refer you to §245. Since we know that L0 (A) can be identified with L0 (µ) for a suitable measure space (X, Σ, µ) (321J, 364Jc), everything we know about general spaces L0 (µ) can be applied directly to L0 (A) for measure algebras (A, µ ¯); and that is what I will do for the next few paragraphs. So far, all I have done is to write τa in place of the τ¯F of 245A, and call on the remarks in 245Bb and 245F. 367N Theorem (a) For any measure algebra (A, µ ¯), the topology T of convergence in measure on L0 = L0 (A) is a linear space topology, and any order*-convergent sequence in L0 is T-convergent to the same limit. (b) (A, µ ¯) is semi-finite iff T is Hausdorff. (c) (A, µ ¯) is localizable iff T is Hausdorff and L0 is complete under the uniformity corresponding to T. (d) (A, µ ¯) is σ-finite iff T is metrizable.

368

Function spaces

367N

proof 245D, 245Cb, 245E. Of course we need 322B to assure us that the phrases ‘semi-finite’, ‘localizable’, ‘σ-finite’ here correspond to the same phrases used in §245, and 367G to identify order*-convergence in L0 with the order-convergence studied in §245. 367O Proposition Let (A, µ ¯) be a measure algebra and give L0 = L0 (A) its topology of convergence in 0 measure. If A ⊆ L is a non-empty, downwards-directed set with infimum 0, then for every neighbourhood G of 0 in L0 there is a u ∈ A such that v ∈ G whenever |v| ≤ u. R proof Let a ∈ Af , ² > 0 be such that u ∈ G whenever |u|∧ χa ≤ ² (see 245Bb). Since {u ∧ χa :Ru ∈ A} is a downwards-directed set in L1 = L1 (A, µ ¯) with infimum 0 in L1 , there must be a u ∈ A such that u∧χa ≤ ² (365Da). But now [−u, u] ⊆ G, as required. 367P Theorem Let U be a Banach lattice and (A, µ ¯) a measure algebra. Give L0 = L0 (A) its topology 0 0 of convergence in measure. If T : U → L = L (A) is a positive linear operator, then it is continuous. proof Take any open set G ⊆ L0 . ?? Suppose, if possible, that T −1 [G] is not open. Then we can find u, hun in∈N ∈ U such that T u ∈ G and kun − uk ≤ 2−n , T un ∈ / GP for every n. Set H = G −P T u; then H is an ∞ ∞ open set containing 0 but not T (un − u), for any n ∈ N. Since n=0 kun − uk < ∞, v = n=0 n|un − u| is 1 defined in U , and |T (un − u)| ≤ n T v for every n ≥ 1. But by 367N (or otherwise) we know that there is some n such that w ∈ H whenever |w| ≤ n1 T v, so that T (un − u) ∈ H for some n, which is impossible. X X 367Q Proposition Let (A, µ ¯) be a σ-finite measure algebra. (a) A sequence hun in∈N in L0 = L0 (A) converges in measure to u ∈ L0 iff every subsequence of hun in∈N has a sub-subsequence which order*-converges to u. (b) A set F ⊆ L0 is closed for the topology of convergence in measure iff u ∈ F whenever there is a sequence hun in∈N in F order*-converging to u ∈ L0 . proof 245K, 245L. 367R

It will be useful later to be able to quote the following straightforward facts.

Proposition Let (A, µ ¯) be a measure algebra. Give A its measure-algebra topology (323A) and L0 = L0 (A) the topology of convergence in measure. Then the map χ : A → L0 is a homeomorphism between A and its image in L0 . proof Of course χ is injective (364Kc). The measure-algebra topology of A is defined by the Rpseudometrics ρa (b, c) = µ ¯(a ∩ (b4c)), while the topology of L0 is defined by the pseudometrics τa (u, v) = |u − v| ∧ χa, in both cases taking a to run over elements of A of finite measure; as τa (χb, χc) is always equal to ρa (b, c), we have the result. 367S Proposition Let E ⊆ R be a Borel set, and h : E → R a continuous function. Let (A, µ ¯) be a ¯ : QE → L0 = L0 (A) the associated function, where QE = {u : u ∈ L0 , [[u ∈ E]] = 1} measure algebra, and h ¯ is continuous for the topology of convergence in measure. (364I). Then h proof (Compare 245Dd.) Express (A, µ ¯) as the measure algebra of a measure space (X, Σ, µ), and write f • for the element of L0 corresponding to any f ∈ L0 . Take any u ∈ QE , any a ∈ A such that µ ¯a < ∞, and any ² > 0. Express u as f • where f : X → R is a measurable function, and a as F • where F ∈ Σ. Then f (x) ∈ E a.e.(x). For each n ∈ N, write En for {t : t ∈ E, |h(s) − h(t)| ≤ 13 ² whenever s ∈ E and |s − t| ≤ 2−n }. Then hEn in∈N is a non-decreasing sequence of Borel sets with union E, so there is an n such that µ{x : x ∈ F , f (x) ∈ / En } ≤ 31 ². R Now suppose that v ∈ QE is such that |v − u| ∧ χa ≤ 13 ²/2n . Express v as g • where g : X → R is a measurable function. Then g(x) ∈ E for almost every x, and

R F

1 3

min(1, |g(x) − f (x)|)µ(dx) ≤ ²/2n ,

367T


369

so µ{x : x ∈ F, |f (x) − g(x)| > 2−n } ≤ 13 ², and 1 ²} 3 ⊆ {x : x ∈ F, f (x) ∈ / En } ∪ {x : g(x) ∈ / E}

{x : x ∈ F, |h(g(x)) − h(f (x))| >

∪ {x : x ∈ F, |f (x) − g(x)| > 2−n } has measure at most 23 ². But this means that

R

¯ ¯ |h(v) − h(u)| ∧ χa =

R F

min(1, |hg(x) − hf (x)|)µ(dx) ≤ ².

¯ is continuous. As u, a and ² are arbitrary, h 367T Intrinsic description of convergence in measure It is a remarkable fact that the topology of convergence in measure, not only on L0 but on its order-dense Riesz subspaces, can be described in of the Riesz space structure alone, without referring at all to the underlying measure algebra or to integration. (Compare 324H.) There is more than one way of doing this. As far as I know, none is outstandingly convincing; I present a formulation which seems to me to exhibit some, at least, of the essence of the phenomenon. Proposition Let (A, µ ¯) be a semi-finite measure algebra, and U an order-dense Riesz subspace of L0 = 0 L (A). Suppose that A ⊆ U and u∗ ∈ U . Then u∗ belongs to the closure of A for the topology of convergence in measure iff there is an order-dense Riesz subspace V of U such that for every v ∈ V + there is a non-empty downwards-directed B ⊆ U , with infimum 0, such that for every w ∈ B there is a u ∈ A such that |u − u∗ | ∧ v ≤ w. proof (a) Suppose first that u∗ ∈ A. Take V to be U ∩ L1 (A, µ ¯); then V is an order-dense Riesz subspace of L0 , by 352Nc, and is therefore order-dense in U . (This is where I use the hypothesis that (A, µ ¯) is semi-finite, so that L1 is order-dense in L0 , by 365Ga.) Take any v ∈ V + . For each n ∈ N, set an = [[v > 2−n ]] ∈ Af . Because u∗ ∈ A, there is a un ∈ A such that µ ¯bn ≤ 2−n , where bn = an ∩ [[|un − u∗ | > 2−n ]] = [[|un − u∗ | ∧ v > 2−n ]]. Set cn = supi≥n bi ; then µ ¯cn ≤ 2−n+1 for each n, so inf n∈N cn = 0 and inf n∈N wn = 0 in L0 , where wn = v × χcn + 2−n χ1. Also |un − u∗ | ∧ v ≤ wn for each n. The wn need not belong to U , so we cannot set B = {wn : n ∈ N}. But if instead we write B = {w : w ∈ U, w ≥ v ∧ wn for some n ∈ N}, then B is non-empty and downwards-directed (because hwn in∈N is non-increasing); and inf B = v − sup{v − w : w ∈ B} = v − sup{w : w ∈ U, w ≤ (v − wn )+ for some n ∈ N} = v − sup(v − wn )+ n∈N

(because U is order-dense in L0 ) = 0. Since for every w ∈ B there is an n such that v ∧ |un − u∗ | ≤ v ∧ wn ≤ w, B witnesses that the condition is satisfied. f (b) Now suppose that the condition is satisfied. Fix a ∈ A R , ² > 0. Because V is order-dense in U and 0 therefore in L , there is a v ∈ V such that 0 ≤ v ≤ χa and v ≥ µ ¯a − ². Let B be a downwards-directed

370

Function spaces

367T

set, with infimumR 0, such that for every w ∈ B there is a u ∈ A with v ∧ |u − u∗ | ≤ w. Then there is a w ∈ B such that w ∧ v ≤ ². Now there is a u ∈ A such that |u − u∗ | ∧ v ≤ w, so that

R

|u − u∗ | ∧ χa ≤ ² +

R

|u − u∗ | ∧ v ≤ ² +

R

w ∧ v ≤ 2².

∗

As a and ² are arbitrary, u ∈ A. *367U Theorem Let (A, µ ¯) be a semi-finite measure algebra; write L1 for L1 (A, µ ¯). Let P : (L1 )∗∗ → L1 be the linear operator corresponding to the band projection from (L1 )∗∗ = (L1 )×∼ onto (L1 )×× and the canonical isomorphism between L1 and (L1 )×× . For A ⊆ L1 write A∗ for the weak* closure of the image of A in (L1 )∗∗ . Then for every A ⊆ L1 P [A∗ ] ⊆ Γ(A), where Γ(A) is the convex hull of A and Γ(A) is the closure of Γ(A) in L0 = L0 (A) for the topology of convergence in measure. proof (a) The statement of the theorem includes a number of assertions: that (L1 )∗ = (L1 )× ; that (L1 )∗∗ = ((L1 )∗ )∼ ; that the natural embedding of L1 into (L1 )∗∗ = (L1 )×∼ identifies L1 with (L1 )×× ; and that (L1 )×× is a band in (L1 )×∼ . For proofs of these see 365C, 356B, 356D, 356N and 356P. Now for the new argument. First, observe that the statement of the theorem involves the measure algebra (A, µ ¯) and the space L0 only in the definition of ‘convergence in measure’; everything else depends only on the Banach lattice structure of L1 . And since we are concerned only with the question of whether of P [A∗ ], which is surely a subset of L1 , belong to Γ(A), 367T shows that this also can be answered in of the Riesz space structure of L1 . What this means is that we can suppose that (A, µ ¯) is localizable. P P b Let (A, µ ˜) be the localization of (A, µ ¯) (322P). The natural expression of A as an order-dense subalgebra of b identifies Af = {a : a ∈ A, µ b f (322O), so that L1 (A, µ b µ A ¯a < ∞} with A ¯) becomes identified with L1 (A, ˜), 1 1 b b µ by 365Od. Thus we can think of L as L (A, µ ˜), and (A, ˜) is localizable. Q Q (b) Take φ ∈ A∗ and set u0 = P φ; I have to show that u0 ∈ Γ(A). Write R for the canonical map from L to (L1 )∗∗ , so that φ belongs to the weak* closure of R[A]. Consider first the case u0 = 0. Take any c ∈ Af and ² > 0. We know that (L1 )∗ = (L1 )∼ = (L1 )× can be identified with L∞ = L∞ (A) (365Jc), so that φ ∈ (L∞ )∗ = (L∞ )∼ must be in the band orthogonal to (L∞ )× . Now we can identify (L∞ )∼ with the Riesz space M of bounded additive functionals on A, and if we do so then (L∞ )× corresponds to the space Mτ of completely additive functionals (363K). Writing Pτ : M → Mτ for the band projection, we must have Pτ (ν) = 0, where ν ∈ M is defined by setting νa = φ(χa) for each a ∈ A; consequently Pτ (|ν|) = 0 and there is an upwards-directed family C ⊆ A, with supremum 1, such that |ν|(a) = 0 for every a ∈ C (362D). Since µ ¯c = supa∈C µ ¯(a ∩ c), there is an a ∈ C such that µ ¯(c \ a) ≤ ². Consider the map Q : L1 → L1 defined by setting Qw = w × χa for every w ∈ L1 . Then its adt 0 Q : L∞ → L∞ (3A5Ed) can be defined by the same formula: Q0 v = v × χa for every v ∈ L∞ . Since |φ| ∈ (L∞ )∼ corresponds to |ν| ∈ M , we have 1

|φ(Q0 v)| ≤ kvk∞ |φ|(χa) = kvk∞ |ν|(a) = 0 for every v ∈ L∞ , and Q00 φ = 0, where Q00 : (L∞ )∗ → (L∞ )∗ is the adt of Q0 . Since Q00 is continuous for the weak* topology on (L∞ )∗ , 0 ∈ Q00 [R[A]], where Q00 R[A] is the closure for the weak* topology of (L∞ )∗ . But of course Q00 R = RQ, while the weak* topology of (L∞ )∗ corresponds, on the image R[L1 ] of L1 , to the weak topology of L1 ; so that 0 belongs to the closure of Q[A] for the weak topology of L1 . Because Q is linear, Q[Γ(A)] is convex. Since 0 belongs to the closure of Q[Γ(A)] for the weak topology of L1 , it belongs to the closure of Q[Γ(A)] for the norm topology (3A5Ee). So 0 belongs to the closure of Q[Γ(A)] for the norm topology, and there is a w ∈ Γ(A) such that kw × χak1 ≤ ²2 . But this means that µ ¯(a ∩ [[|w| ≥ ²]]) ≤ ² and µ ¯(c ∩ [[|w| ≥ ²]]) ≤ 2². Since c and ² are arbitrary, 0 ∈ Γ(A). (c) This deals with the case u0 = 0. Now the general case follows at once if we set B = A − u0 and observe that φ − Ru0 ∈ B ∗ , so 0 = P (φ − Ru0 ) ∈ Γ(B) = Γ(A) − u0 = Γ(A) − u0 .

367Xf


371

Remark This is a version of a theorem from Bukhvalov 95. *367V Corollary Let (A, µ ¯) be a localizable measure algebra. Let C be a family of convex subsets of L0 = L0 (A), all closed for the topology of convergence in measure, with the finite intersection property, and R suppose that for every non-zero a ∈ A there are a non-zero b ⊆ a and a C ∈ C such that sup |u| < ∞. u∈C b T Then C = 6 ∅. proof Because C has the finite intersection property, there is an ultrafilter F on L0 including C. Set I = {a : a ∈ A, inf F ∈F supu∈F

R

a

|u| < ∞};

because F is a filter, I is an ideal in A, and the condition on C tells us that I is order-dense. For each a ∈ I, define Qa : L0 → L0 by setting Qa u = u × χa. Then there is an F ∈ F such that Qa [F ] is a norm-bounded set in L1 , so φa = limu→F RQa u is defined in (L∞ )∗ for the weak* topology on (L∞ )∗ , writing R for the canonical map from L1 to (L∞ )∗ ∼ = (L1 )∗∗ . If P : (L∞ )∗ → L1 is the map corresponding to the band ∞ ∼ ∞ × ˜ projection P from (L ) onto (L ) , as in 367U, and C ∈ C, then 367U tells us that P (φa ) must belong to the closure of the convex set Qa [C] for the topology of convergence in measure. Moreover, if a ⊆ b ∈ I, so that Qa = Qa Qb , then P (φa ) = Qa P (φb ). P P Qa ¹L1 is a band projection on L1 , so its adt Q0a is a 1 ∼ 00 ∞ ∼ band projection on L = (L ) (356C) and Qa is a band projection on (L∞ )∗ ∼ = (L∞ )∼ . This means that 00 00 ˜ Qa will commute with P (352Sb). But also Qa is continuous for the weak* topology of (L∞ )∗ , so Q00a (φb ) = limu→F Q00a RQb u = limu→F RQa Qb u = φa , and Q P (φa ) = R−1 P˜ (φa ) = R−1 P˜ Q00a (φb ) = R−1 Q00a P˜ (φb ) = Qa R−1 P˜ (φb ) = Qa P (φb ). Q Q 0 ∼ What this means is that if we take a partition D of unity included in I (313K), so that L =

L0 (Ad ) (315F(iii), 364S), and define w ∈ L by saying that w × χd = P (φd ) for every d ∈ D, then we shall have w × χa = P (φa ) for every a ∈ I. But now, given a ∈ Af and ² > 0 and C ∈ C, there is a b ∈ I such that µ ¯(a \ b) ≤ ²; w × χb ∈ Qb [C], so there is a u ∈ C such that µ ¯(b ∩ [[|w − u| ≥ ²]]) ≤T²; and µ ¯ (a ∩ [[|w − u| ≥ ²]]) ≤ 2². As a, ² are arbitrary and C is closed, w ∈ C; as C is arbitrary, w ∈ C and T C 6= ∅. 0

d∈D

367X Basic exercises > (a) Let P be a lattice. (i) Show that if p ∈ P and hpn in∈N is a non-decreasing sequence in P , then hpn in∈N is order*-convergent to p iff p = supn∈N pn . (ii) Suppose that hpn in∈N a sequence in P order*-converging to p ∈ P . Show that p = supn∈N p ∧ pn = inf n∈N p ∨ pn . (iii) Let hpn in∈N , hqn in∈N be two sequences in P which are order*-convergent to p, q respectively. Show that if pn ≤ qn for every n then p ≤ q. (iv) Let hpn in∈N be a sequence in P . Show that hpn in∈N order*-converges to p ∈ P iff hpn ∨ pin∈N and hpn ∧ pin∈N order*-converge to p. (b) Let P and Q be lattices, and f : P → Q an order-preserving function. Suppose that hpn in∈N is an order-bounded sequence which order*-converges to p in P . Show that hf (pn )in∈N order*-converges to f (p) in Q if either f is order-continuous or P is Dedekind σ-complete and f is sequentially order-continuous. (c) Let P be either a Boolean algebra or a Riesz space. Suppose that hpn in∈N is a sequence in P such that hp2n in∈N and hp2n+1 in∈N are both order*-convergent to p ∈ P . Show that hpn in∈N is order*-convergent to p. (Hint: 313B, 352E.) >(d) Let A be a Boolean algebra and han in∈N , hbn in∈N two sequences in A order*-converging to a, b respectively. Show that han ∪ bn in∈N , han ∩ bn in∈N , han \ bn in∈N , han 4 bn in∈N order*-converge to a ∪ b, a ∩ b, a \ b and a 4 b respectively. (e) Let A be a Boolean algebra and han in∈N a sequence in A. Show that han in∈N does not order*-converge to 0 iff there is a non-zero a ∈ A such that a = supi≥n a ∧ ai for every n ∈ N. >(f ) (i) Let U be a Riesz space and hun in∈N an order*-convergent sequence in U + with limit u. Show that h(u) ≤ lim inf n→∞ h(un ) for every h ∈ (U × )+ . (ii) Let U be a Riesz space and hun in∈N an order-bounded order*-convergent sequence in U with limit u. Show that h(u) = limn→∞ h(un ) for every h ∈ U × . (Compare 356Xd.)

372

Function spaces

367Xg

> (g) Let U be a Riesz space with a Fatou norm k k. (i) Show that if hun in∈N is an order*-convergent sequence in U with limit u, then kuk ≤ lim inf n→∞ kun k. (Hint: h|un | ∧ |u|in∈N is order*-convergent to |u|.) (ii) Show P∞ that if hun in∈N is a norm-convergent sequence in U it has an order*-convergent subsequence. (Hint: if n=0 kun k < ∞ then hun in∈N order*-converges to 0.) (h) Let U and V be Archimedean Riesz spaces and T : U → V an order-continuous Riesz homomorphism. Show that if hun in∈N is a sequence in U which order*-converges to u ∈ U , then hT un in∈N order*-converges to T u in V . (i) Let A be a Boolean algebra and B an order-closed subalgebra. Show that if hbn in∈N is a sequence in B and b ∈ B, then hbn in∈N order*-converges to b in B iff it order*-converges to b in A. (j) Let A be a Dedekind σ-complete Boolean algebra and hun in∈N , hvn in∈N two sequences in L0 (A) which are order*-convergent to u, v respectively. Show that hun × vn in∈N order*-converges to u × v. Show that if −1 −1 u, un all have multiplicative inverses u−1 , u−1 . n then hun in∈N order*-converges to u (k) Let A be a Dedekind σ-complete Boolean algebra and I a σ-ideal of A. Show that for any han in∈N , a ∈ A, ha•n in∈N order*-converges to a• in A/I iff inf n∈N supm≥n am 4a ∈ I. > (l) Let A be a Dedekind σ-complete Boolean algebra, and hhn in∈N a sequence of Borel measurable ¯ n (u)in∈N functions from R to itself such that h(t) = limn→∞ hn (t) is defined for every t ∈ R. Show that hh 0 0 0 0 ¯ ¯ ¯ order*-converges to h(u) for every u ∈ L = L (A), where hn , h : L → L are defined as in 364I. (m) Let (A, µ ¯) be a measure algebra, and hun in∈N a sequence in L1 = L1 (A, µ ¯) which is order*-convergent 1 to u ∈ L . Show that hun in∈N is norm-convergent to u iff {un : n ∈ N} is uniformly integrable iff kuk1 = limn→∞ kun k1 . (Hint: 245H, 246J.) (n) Let U be an L-space and hun in∈N a norm-bounded sequence in U . Show that there are a v ∈ U and Pn 1 a subsequence hvn in∈N of hun in∈N such that h n+1 i=0 wi in∈N order*-converges to v for every subsequence hwn in∈N of hvn in∈N . (Hint: 276H.) > (o) Let (A, µ ¯) be a measure algebra and give L0 = L0 (A) its topology of convergence in measure. Show that u 7→ |u|, (u, v) 7→ u ∨ v, (u, v) 7→ u × v are continuous. (p) Let (A, µ ¯) be a measure algebra and p ∈ [1, ∞[. For v ∈ (Lp )+ = Lp (A, µ ¯)+ define ρv : L0 ×L0 → [0, ∞[ by setting ρv (u1 , u2 ) = k|u1 − u2 | ∧ vkp for all u1 , u2 ∈ U . Show that each ρv is a pseudometric and that the topology on L0 = L0 (A) defined by {ρv : v ∈ (Lp )+ } is the topology of convergence in measure. (q) Let (A, µ ¯) be a σ-finite measure algebra. Suppose we have a double sequence huij i(i,j)∈N×N in L0 = L (A) such that huij ij∈N order*-converges to ui in L0 for each i, while hui ii∈N order*-converges to u. Show that there is a strictly increasing sequence hn(i)ii∈N such that hui,n(i) ii∈N order*-converges to u. 0

(r) Let (X, Σ, µ) be a semi-finite measure space. Show that L0 (µ) is separable for the topology of convergence in measure iff µ is σ-finite and has countable Maharam type. (Cf. 365Xp.) (s) Let (A, µ ¯) be a measure algebra. (i) Show that if han in∈N is order*-convergent to a ∈ A, then han in∈N → a for the measure-algebra topology. (ii) Show that if (A, µ ¯) is σ-finite, then (α) a sequence converges to a for the topology of A iff every subsequence has a sub-subsequence which is order*-convergent to to a (β) a set F ⊆ A is closed for the topology of A iff a ∈ F whenever there is a sequence han in∈N in F which is order*-convergent to a ∈ A. (t) Let (A, µ ¯) be a measure algebra which is not σ-finite. Show that there is a set A ⊆ L0 (A) such that the limit of any order*-convergent sequence in A belongs to A, but A is not closed for the topology of convergence in measure. (u) Let U be a Banach lattice with an order-continuous norm. (i) Show that a sequence hun in∈N is norm-convergent to u ∈ U iff every subsequence has a sub-subsequence which is order-bounded and order*convergent to u. (ii) Show that a set F ⊆ U is closed for the norm topology iff u ∈ F whenever there is an order-bounded sequence hun in∈N in F order*-converging to u ∈ U .

367Yj


373

(v) Let (A, µ ¯) be a probability algebra. For u ∈ L0 = L0 (A) let νu be the distribution of u (364Xd). Show that u 7→ νu is continuous when L0 is given the topology of convergence in measure and the space of probability distributions on R is given the vague topology (274Ld). (w) Let (A, µ ¯) be a probability algebra and hun in∈N a stochastically independent sequence in L0 , all with the Cauchy distribution νC,1 with centre 0 and scale parameter 1 (285Xm). For each n let Cn be the convex hull of {ui : i ≥ n}, and Cn its closure for the topology of convergence in measure. Show that every u ∈ C0 has distribution νC,1 . (Hint: consider first u ∈ C0 .) Show that C0 is bounded for the topology of T convergence in measure. Show that n∈N Cn = ∅. (x) If U is a linear space and C ⊆ U is a convex set, a function f : C → R is convex if f (αx + (1 − α)y) ≤ αf (x) + (1 − α)f (y) whenever x, y ∈ C and α ∈ [0, 1]. Let (A, µ ¯) be a localizable measure algebra and C ⊆ L1 (A, µ ¯) a non-empty convex norm-bounded set which is closed in L0 (A) for the topology of convergence in measure. Show that any convex function f : C → R which is lower semi-continuous for the topology of convergence in measure is bounded below and attains its infimum. 367Y Further exercises (a) Give an example of an Archimedean Riesz space U and an order-bounded sequence hun in∈N in U which is order*-convergent to 0, but such that there is no non-increasing sequence hvn in∈N , with infimum 0, such that un ≤ vn for every n ∈ N. (b) Let P be a distributive lattice. Show that if hpn in∈N is a sequence in P order*-converging to p ∈ P , then hpn ∨ qin∈N , hpn ∧ qin∈N order*-converge to p ∨ q, p ∧ q respectively for any q ∈ P . (c) Let P be any lattice. (i) Show that there is a topology on P for which a set A ⊆ P is closed iff p ∈ A whenever there is a sequence in A which is order*-convergent to p. Show that any closed set for this topology is sequentially order-closed. (ii) Now let Q be another lattice, with the topology defined in the same way, and f : P → Q an order-preserving function. Show that if f is topologically continuous it is sequentially order-continuous. (d) Let us say that a lattice P is (2, ∞)-distributive if (α) whenever A, B ⊆ P are non-empty sets with infima p, q respectively, then inf{a ∨ b : a ∈ A, b ∈ B} = p ∨ q (β) whenever A, B ⊆ P are non-empty sets with suprema p, q respectively, then sup{a ∧ b : a ∈ A, b ∈ B} = p ∧ q. Show that, in this case, if hpn in∈N order*-converges to p and hqn in∈N order*-converges to q, hpn ∨ qn in∈N order*-converges to p ∨ q. (e) Give an example of a lattice P with two sequences hpn in∈N , hqn in∈N , both order*-convergent to p, such that hpn ∨ qn in∈N is not order*-convergent to p. (f ) (i) Give an example of a Riesz space U with an order-dense Riesz subspace V of U and a sequence hvn in∈N in V such that hvn in∈N order*-converges to 0 in V but does not order*-converge in U . (ii) Give an example of a Riesz space U with an order-dense Riesz subspace V of U and a sequence hvn in∈N in V , order-bounded in V , such that hvn in∈N order*-converges to 0 in U but does not order*-converge in V . (g) Let U be an Archimedean f -algebra. Show that if hun in∈N , hvn in∈N are sequences in U order*converging to u, v respectively, then hun × vn in∈N order*-converges to u × v. (h) Let A be a Dedekind σ-complete Boolean algebra and r ≥ 1. Let E ⊆ Rr be a Borel set and write QE = {(u1 , . . . , ur ) : [[(u1 , . . . , ur ) ∈ E]] = 1} ⊆ L0 (A)r (364Yc). Let h : E → R be a continuous function ¯ : QE → L0 the corresponding map (364Yd). Show that if hun in∈N is a sequence in QE which is and h ¯ n )in∈N is order*-convergent to h(u). ¯ order*-convergent to u ∈ QE (in the lattice (L0 )r ), then hh(u (i) Let X be a completely regular Baire space (definition: 314Yd), and hun in∈N a sequence in C(X). Show that hun in∈N order*-converges to 0 in C(X) iff {x : lim supn→∞ |un (x)| > 0} is meager in X. (j) (i) Give an example of a sequence hun in∈N in C([0, 1]) such that limn→∞ un (x) = 0 for every x ∈ [0, 1], but {un : n ∈ N} is not order-bounded in C([0, 1]). (ii) Give an example of an order-bounded sequence hun in∈N in C(Q) such that limn→∞ un (q) = 0 for every q ∈ Q, but supi≥n ui = χQ in C(Q) for every n ∈ N. (iii) Give an example of a sequence hun in∈N in C([0, 1]) such that hun in∈N order*-converges to 0 in C([0, 1]), but limn→∞ un (q) > 0 for every q ∈ Q ∩ [0, 1].

374

Function spaces

367Yk

(k) Write out an alternative proof of 367K/367Yi based on the fact that, for a Baire space X, C(X) can be identified with an order-dense Riesz subspace of a quotient of the space of Σ-measurable functions, where Σ is the algebra of subsets of X with the Baire property, as in 364Yi. (l) Let A be a ccc weakly (σ, ∞)-distributive Boolean algebra. Show that there is a topology on A such that the closure of any A ⊆ A is precisely the set of limits of order*-convergent sequences in A. (m) Give an example of a set X and a double sequence humn im,n∈N in RX such that limn→∞ umn (x) = um (x) exists for every m ∈ N and x ∈ X, limm→∞ um (x) = 0 for every x ∈ X, but there is no sequence hvk ik∈N in {umn : m, n ∈ N} such that limk→∞ vk (x) = 0 for every x. (n) Let U be any Banach lattice with an order-continuous norm. For v ∈ U + define ρv : U × U → [0, ∞[ by setting ρv (u1 , u2 ) = k|u1 − u2 | ∧ vk for all u1 , u2 ∈ U . Show that every ρv is a pseudometric on U , and that {ρv : v ∈ U + } defines a Hausdorff linear space topology on U . (o) Let U be any Riesz space. For h ∈ (Uc∼ )+ (356Ab), v ∈ U + define ρvh : U × U → [0, ∞[ by setting ρvh (u1 , u2 ) = h(|u1 − u2 | ∧ v) for all u1 , u2 ∈ U . Show that each ρvh is a pseudometric on U , and that {ρvh : h ∈ (Uc∼ )+ , v ∈ U + } defines a linear space topology on U . (p) Let (A, µ ¯) be a σ-finite measure algebra. Show that the function (α, u) 7→ [[u > α]] : R × L0 → A is Borel measurable when L0 = L0 (A) is given the topology of convergence in measure and A is given its measure-algebra topology. (Hint: if a ∈ A, γ ≥ 0 then {(α, u) : µ ¯(a ∩ [[u > α]]) > γ} is open.) (q) Let G be the regular open algebra of R. Show that there is no Hausdorff topology T on L0 (G) such that hun in∈N is T-convergent to u whenever hun in∈N is order*-convergent to u. (Hint: Let H be any T-open set containing 0. Enumerate Q as hqn in∈N . Find inductively a non-decreasing sequence hGn in∈N in G such that χGn ∈ H, qn ∈ Gn for every n. Conclude that χR ∈ H.) (r) Give an example of a Banach lattice with a norm which is not order-continuous, but in which every order-bounded order*-convergent sequence is norm-convergent. (s) Let A be a Dedekind σ-complete Boolean algebra and r ≥ 1. Let E ⊆ Rr be a Borel set and write QE = {(u1 , . . . , ur ) : [[(u1 , . . . , ur ) ∈ E]] = 1} ⊆ L0 (A)r (364Yc). Let h : E → R be a continuous function ¯ : QE → L0 the corresponding map (364Yd). Show that if h ¯ is continuous if L0 is given its topology and h 0 r of convergence in measure and (L ) the product topology. (t) Show that 367U is true for all measure algebras, whether semi-finite or not. 367 Notes and comments I have given a very general definition of ‘order*-convergence’. The general theory of convergence structures on ordered spaces is complex and full of traps for the unwary. I have tried to lay out a safe path to the results which are important in the context of this book. But the propositions here are necessarily full of little conditions (e.g., the requirement that U should be Archimedean, in 367F) whose significance may not be immediately obvious. In particular, the definition is very much better adapted to distributive lattices than to others (367Yb, 367Yd, 367Ye). It is useful in the study of Riesz spaces and Boolean algebras largely because these satisfy strong distributive laws (313B, 352E). The special feature which distinguishes the definition here from other definitions of order-convergence is the fact that it can be applied to sequences which are not order-bounded. For order-bounded sequences there are useful simplifications (367Be-f), but the Martingale Theorem (357J), for instance, if we want to express it in of its natural home in the Riesz space L1 , refers to sequences which are hardly ever order-bounded. The * in the phrase ‘order*-convergent’ is supposed to be a warning that it may not represent exactly the concept you expect. I think nearly any author using the phrase ‘order-convergent’ would accept sequences fulfilling the conditions of 367Bf; but beyond this no standard definitions have taken root. The fact that order*-convergent sequences in an L0 space are order-bounded (367H) is actually one of the characteristic properties of L0 . Related ideas will be important in the next section (368A, 368M). It is one of the outstanding characteristics of measure algebras in this context that they provide nontrivial linear space topologies on their L0 spaces, related in striking ways to the order structure. Not all L0

Embedding Riesz spaces in L0

368B

375

spaces have such topologies (367Yq). It is not known whether a topology corresponding to ‘convergence in measure’ can be defined on L0 (A) for any A which is not a measure algebra; this is the ‘control measure problem’, which I will discuss in §393 (see 393L). 367T shows that the topology of convergence in measure on L0 (A) is (at least for semi-finite measure algebras) determined by the Riesz space structure of L0 ; and that indeed the same is true of its order-dense Riesz subspaces. This fact is important for a full understanding of the representation theorems in §369 below. If a Riesz space U can be embedded as an order-dense subspace of any such L0 , then there is already a ‘topology of convergence in measure’ on U , independent of the embedding. It is therefore not surprising that there should be alternative descriptions of the topology of convergence in measure on the important subspaces of L0 (367Xp, 367Yn). For σ-finite measure algebras, the topology of convergence in measure is easily described in of order-convergence (367Q). For other measure algebras, the formula fails (367Xt). 367Yq shows that trying to apply the same ideas to Riesz spaces in general gives rise to some very curious phenomena. 367V enables us to prove results which would ordinarily be associated with some form of compactness. Of course compactness is indeed involved, as the proof through 367U makes clear; but it is weak* compactness in (L1 )∗∗ , rather than in the space immediately to hand. I hardly mention ‘uniform integrability’ in this section, not because it is uninteresting, but because I have nothing to add at this point to 246J and the exercises in §246. But I do include translations of Lebesgue’s Dominated Convergence Theorem (367I) and the Martingale Theorem (367J) to show how these can be expressed in the language of this chapter.

368 Embedding Riesz spaces in L0 In this section I turn to the representation of Archimedean Riesz spaces as function spaces. Any Archimedean Riesz space U can be represented as an order-dense subspace of L0 (A), where A is its band algebra (368E). Consequently we get representations of Archimedean Riesz spaces as quotients of subspaces of RX (368F) and as subspaces of C ∞ (X) (368G), and a notion of ‘Dedekind completion’ (368I-368J). Closely associated with these is the fact that we have a very general extension theorem for order-continuous Riesz homomorphisms into L0 spaces (368B). I give a characterization of L0 spaces in of lateral completeness (368M, 368Yd), and I discuss weakly (σ, ∞)-distributive Riesz spaces (368N-368S). 368A Lemma Let A be a Dedekind σ-complete Boolean algebra, and A ⊆ (L0 )+ a set with no upper bound in L0 , where L0 = L0 (A). If either A is countable or A is Dedekind complete, there is a v > 0 in L0 such that nv = supu∈A u ∧ nv for every n ∈ N. proof The hypothesis ‘A is countable or A is Dedekind complete’ ensures that cα = supu∈A [[u > α]] is defined for each α. By 364Ma, c = inf n∈N cn = inf α∈R cα is non-zero. Now for any n ≥ 1, α ∈ R [[supu∈A (u ∧ nχc) > α]] = supu∈A [[u > α]] ∩ [[χc >

α n ]]

= [[χc >

α n ]],

because if 0 ≤ α ≤ k ∈ N then supu∈A [[u > α]] ⊇ ck ⊇ c ⊇ [[χc >

α n ]],

while if α < 0 then (because A is a non-empty subset of (L0 )+ ) supu∈A [[u > α]] = 1 = [[χc >

α n ]].

So supu∈A u ∧ nχc = nχc for every n ≥ 1, and we can take v = χc. (The case n = 0 is of course trivial.) 368B Theorem Let A be a Dedekind complete Boolean algebra, U an Archimedean Riesz space, V an order-dense Riesz subspace of U and T : V → L0 = L0 (A) an order-continuous Riesz homomorphism. Then T has a unique extension to an order-continuous Riesz homomorphism T˜ : U → L0 . proof (a) The key to the proof is the following: if u ≥ 0 in U , then {T v : v ∈ V, 0 ≤ v ≤ u} is bounded above in L0 . P P?? Suppose, if possible, otherwise. Then by 368A there is a w > 0 in L0 such that

376

Function spaces

368B

nw = supv∈A nw ∧ T v for every n ∈ N, where A = {v : v ∈ V, 0 ≤ v ≤ u}. In particular, there is a v0 ∈ A such that w0 = w ∧ T v0 > 0. Because U is Archimedean, inf k≥1 k1 u = 0, so v0 = supk≥1 (v0 − k1 u)+ . Because V is order-dense in U , v0 = sup B where 1 k

B = {v : v ∈ V, 0 ≤ v ≤ (v0 − u)+ for some k ≥ 1}. Because T is order-continuous, T v0 = sup T [B] in L0 , and there is a v1 ∈ B such that w1 = w0 ∧ T v1 > 0. Let k ≥ 1 be such that v1 ≤ (v0 − k1 u)+ . Then for any m ∈ N, mv1 ∧ u ≤ (mv1 ∧ kv0 ) + (mv1 ∧ (u − kv0 )+ ) (352Fa) 1 k

≤ kv0 + (m + k)(v1 ∧ ( u − v0 )+ ) = kv0 . So for any v ∈ A, m ∈ N, mw1 ∧ T v = mw1 ∧ mT v1 ∧ T v ≤ T (mv1 ∧ v) ≤ T (mv1 ∧ u) ≤ T (kv0 ) = kT v0 . But this means that, for m ∈ N, mw1 = mw1 ∧ mw = supv∈A mw1 ∧ (mw ∧ T v) = supv∈A mw1 ∧ T v ≤ kT v0 , which is impossible because L0 is Archimedean and w1 > 0. X XQ Q (b) Because L0 is Dedekind complete, sup{T v : v ∈ V, 0 ≤ v ≤ u} is defined in L0 for every u ∈ U . By 355F, T has a unique extension to an order-continuous Riesz homomorphism from U to L0 . 368C Corollary Let A and B be Dedekind complete Boolean algebras and U , V order-dense Riesz subspaces of L0 (A), L0 (B) respectively. Then any Riesz space isomorphism between U and V extends uniquely to a Riesz space isomorphism between L0 (A) and L0 (B); and in this case A and B must be isomorphic as Boolean algebras. proof If T : U → V is a Riesz space isomorphism, then 368B tells us that we have (unique) order-continuous Riesz homomorphisms S : L0 (A) → L0 (B) and S 0 : L0 (B) → L0 (A) extending T , T −1 respectively. Now S 0 S : L0 (A) → L0 (A) is an order-continuous Riesz homomorphism agreeing with the identity on U , so must be the identity on L0 (A); similarly SS 0 is the identity on L0 (B), and S is a Riesz space isomorphism. To see that A and B are isomorphic, recall that by 364Q they can be identified with the algebras of projection bands of L0 (A) and L0 (B), which must be isomorphic. 368D Corollary Suppose that A is a Dedekind σ-complete Boolean algebra, and that U is an orderdense Riesz subspace of L0 (A) which is isomorphic, as Riesz space, to L0 (B) for some Dedekind complete Boolean algebra B. Then U = L0 (A) and A is isomorphic to B (so, in particular, is Dedekind complete). proof The identity mapping U → U is surely an order-continuous Riesz homomorphism, so by 368B extends to an order-continuous Riesz homomorphism T˜ : L0 (A) → U . Now T˜ must be injective, because if u 6= 0 in L0 (A) there is a u0 ∈ U such that 0 < u0 ≤ |u|, so that 0 < u0 ≤ |T˜u|. So we must have U = L0 (A) and T˜ the identity map. By 368C, or otherwise, A ∼ = B. 368E Theorem Let U be any Archimedean Riesz space, and A its band algebra (353B). Then U can be embedded as an order-dense Riesz subspace of L0 (A). proof (a) If U = {0} then A = {0}, L0 = L0 (A) = {0} and the result is trivial; I shall therefore suppose henceforth that U is non-trivial. Note that by 352Q A is Dedekind complete. Let C ⊆ U + \ {0} be a maximal dist set (in the sense of 352C); to obtain such a set apply Zorn’s lemma to the family of all dist subsets of U + \ {0}. Now I can write down the formula for the embedding T : U → L0 immediately, though there will be a good deal of work to do in justification: for u ∈ U and α ∈ R, [[T u > α]] will be the band in U generated by


368E

377

{e ∧ (u − αe)+ : e ∈ C}. (For once, I allow myself to use the formula [[. . . ]] without checking immediately that it represents a member of L0 ; all I claim for the moment is that [[T u > α]] is a member of A determined by u and α.) (b) Before getting down to the main argument, I make some remarks which will be useful later. (i) If u > 0 in U , then there is some e ∈ C such that u ∧ e > 0, since otherwise we ought to have added u to C. Thus C ⊥ = {0}. (ii) If u ∈ U and e ∈ C and α ∈ R, then v = e ∧ (αe − u)+ belongs to [[T u > α]]⊥ . P P If e0 ∈ C, then 0 either e 6= e so v ∧ e0 ∧ (u − αe0 )+ ≤ e ∧ e0 = 0, or e0 = e and v ∧ e0 ∧ (u − αe0 )+ ≤ (αe − u)+ ∧ (u − αe)+ = 0. Accordingly [[T u > α]] is included in the band {v}⊥ and v ∈ [[T u > α]]⊥ . Q Q (c) Now I must confirm that the formula given for [[T u > α]] is consistent with the conditions laid down in 364A. P P Take u ∈ U . (i) If α ≤ β then 0 ≤ e ∧ (u − βe)+ ≤ e ∧ (u − αe)+ ∈ [[T u > α]] so e ∧ (u − βe)+ ∈ [[T u > α]] for every e ∈ C and [[T u > β]] ⊆ [[T u > α]]. (ii) Given α ∈ R, set W = supβ>α [[T u > β]] in A, that is, the band in U generated by {e ∧ (u − βe)+ : e ∈ C, β > α}. Then for each e ∈ C, supβ>α e ∧ (u − βe)+ = e ∧ (u − inf β>α e)+ = e ∧ (u − αe)+ using the general distributive laws in U (352E), the translation-invariance of the order (351D) and the fact that U is Archimedean (to see that αe = inf β>α βe). So e ∧ (u − αe)+ ∈ W ; as e is arbitrary, [[T u > α]] ⊆ W and [[T u > α]] = W . (iii) Now set W = inf n∈N [[T u > n]]. For any e ∈ C, n ∈ N we have e ∧ (ne − u)+ ∈ [[T u > n]]⊥ ⊆ W ⊥ , so that e ∧ (e − n1 u+ )+ ≤ e ∧ (e − n1 u)+ ∈ W ⊥ for every n ≥ 1 and e = supn≥1 e ∧ (e − n1 u+ )+ ∈ W ⊥ . Thus C ⊆ W ⊥ and W ⊆ C ⊥ = {0}. So we have inf n∈N [[T u > n]] = 0. (iv) Finally, set W = supn∈N [[T u > −n]]. Then e ∧ (e − n1 u− )+ ≤ e ∧ (e + n1 u)+ ≤ e ∧ (u + ne)+ ∈ W for every n ≥ 1 and e ∈ C, so e = supn≥1 e ∧ (e − n1 u− )+ ∈ W for every e ∈ C and W ⊥ = {0}, W = U . Thus all three conditions of 364A are satisfied. Q Q (d) Thus we have a well-defined map T : U → L0 . I show next that T (u + v) = T u + T v for all u, v ∈ U . P P I rely on the formulae in 364E and 364Fa, and on partitions of unity in A, constructed as follows. Fix n ≥ 1 for the moment. Then we know that supi∈Z [[T u > So setting

i n ]]

= 1,

inf i∈Z [[T u >

i n ]]

= 0.

378

Function spaces

Vi = [[T u >

i n ]] \ [[T u

i+1 n ]]

>

368E

= [[T u >

i n ]]

∩ [[T u >

i+1 ⊥ n ]] ,

hVi ii∈Z is a partition of unity in A. Similarly, hWi ii∈Z is a partition of unity, where i n ]]

Wi = [[T v >

i+1 ⊥ n ]] .

∩ [[T v >

Now, for any i, j, k ∈ Z such that i + j ≥ k, Vi ∩ Wj ⊆ [[T u >

∩ [[T v >

j n ]]

⊆ [[T u + T v >

[[T u + T v >

k n ]]

⊇ supi+j≥k Vi ∩ Wj .

i n ]]

i+j n ]]

⊆ [[T u + T v >

k n ]];

thus

On the other hand, if q ∈ Q and k ∈ Z, there is an i ∈ Z such that [[T u > q]] ∩ [[T v >

k+1 n

− q]] ⊆ [[T u >

i n ]]

∩ [[T v >

i n

≤q<

k−i n ]]

i+1 n ,

so that

⊆ supi+j≥k Vi ∩ Wj ;

thus for any k ∈ Z [[T u + T v >

k+1 n ]]

⊆ supi+j≥k Vi ∩ Wj ⊆ [[T u + T v >

k n ]].

Also, if 0 < w ∈ Vi ∩ Wj , e ∈ C then w ∧ e ∧ (u −

i+1 + e) n

= w ∧ e ∧ (v −

j+1 + e) n

= 0,

so that w ∧ e ∧ (u + v −

i+j+2 + e) n

=0

because (u + v −

i+j+2 + e) n

by 352Fc. But this means that Vi ∩ Wj ∩ [[T (u + v) > [[T (u + v) >

i+1 + e) n

≤ (u −

k+1 n ]]

i+j+2 n ]]

+ (v −

j+1 + e) n

= {0}. Turning this round,

∩ supi+j≤k−1 Vi ∩ Wj = 0,

and because supi,j∈Z Vi ∩ Wj = U in A, [[T (u + v) >

k+1 n ]]

⊆ supi+j≥k Vi ∩ Wj .

Finally, if i + j ≥ k and 0 < w ∈ Vi ∩ Vj , then there is an e ∈ C such that w1 = w ∧ e ∧ (u − ni e)+ > 0; there is an e0 ∈ C such that w2 = w1 ∧ e0 ∧ (v − nj e0 )+ > 0; of course e = e0 , and i n

j n

i+j + e) n

0 < w2 ≤ e ∧ (u − e)+ ∧ (v − )+ ≤ e ∧ (u + v − ∈ [[T (u + v) > using 352Fc. This shows that w 6∈ [[T (u + v) >

k ⊥ n ]] ;

i+j ]] n

k n

⊆ [[T (u + v) > ]]

as w is arbitrary, Vi ∩ Wj ⊆ [[T (u + v) >

supi+j≥k Vi ∩ Wj ⊆ [[T (u + v) >

k n ]].

Putting all these four together, we see that [[T (u + v) >

k+1 n ]]

⊆ supi+j≥k Vi ∩ Wj ⊆ [[T u + T v >

k n ]],

[[T u + T v >

k+1 n ]]

⊆ supi+j≥k Vi ∩ Wj ⊆ [[T (u + v) >

k n ]]

for all n ≥ 1, k ∈ Z. But this means that we must have [[T u + T v > β]] ⊆ [[T (u + v) > α]], whenever α < β. Consequently

[[T (u + v) > β]] ⊆ [[T u + T v > α]]

k n ]];

so we get


368F

379

[[T u + T v > α]] = sup [[T u + T v > β]] ⊆ [[T (u + v) > α]] β>α

= sup [[T (u + v) > β]] ⊆ [[T u + T v > α]] β>α

and [[T u + T v > α]] = [[T (u + v) > α]] for every α, that is, T (u + v) = T u + T v. Q Q (e) The hardest part is over. If u ∈ U , γ > 0 and α ∈ R, then for any e ∈ C 1 γ

α γ

1 γ

min(1, )(e ∧ (γu − αe)+ ) ≤ e ∧ (u − e)+ ≤ max(1, )(e ∧ (γu − αe)+ ), so [[T (γu) > α]] = [[T u >

α γ ]]

= [[γT u > α]];

as α is arbitrary, γT u = T (γu); as γ and u are arbitrary, T is linear. (We need only check linearity for γ > 0 because we know from the additivity of T that T (−u) = −T u for every u.) (f ) To see that T is a Riesz homomorphism, take any u ∈ U and α ∈ R and consider the band [[T u > α]] ∪ [[−T u > α]] = [[|T u| > α]] (by 364Mb). This is the band generated by {e ∧ (u − αe)+ : e ∈ C} ∪ {e ∧ (−u − αe)+ : e ∈ C}. But this must also be the band generated by {(e ∧ (u − αe)+ ) ∨ (e ∧ (−u − αe)+ ) : e ∈ C} = {e ∧ (|u| − αe)+ : e ∈ C}, which is [[T |u| > α]]. Thus [[|T u| > α]] = [[T |u| > α]] for every α and |T u| = T |u|. As u is arbitrary, T is a Riesz homomorphism. (g) To see that T is injective, take any non-zero u ∈ U . Then there must be some e ∈ C such that |u| ∧ e 6= 0, and some α > 0 such that |u| ∧ e 6≤ αe, so that e ∧ (|u| − αe)+ 6= 0 and [[T |u| > α]] 6= {0} and T |u| 6= 0 and T u 6= 0. Thus T embeds U as a Riesz subspace of L0 . (h) Finally, I must check that T [U ] is order-dense in L0 . P P Let p > 0 in L0 . Then there is some α > 0 such that V = [[p > α]] 6= 0. Take u > 0 in V . Let e ∈ C be such that u ∧ e > 0. Then v = u ∧ αe > 0. Now e ∧ (v − αe)+ = 0; but also e0 ∧ v = 0 for every e0 ∈ C distinct from e, so that [[T v > α]] = {0}. Also v ∈ V , so e0 ∧ (v − βe0 )+ ∈ V whenever e0 ∈ C and β ≥ 0, and [[T v > β]] ⊆ V for every β ≥ 0. Accordingly we have [[T v > β]] = {0} ⊆ [[p > β]] if β ≥ α, ⊆ V ⊆ [[p > β]] if 0 ≤ β < α, = U = [[p > β]] if β < 0, and T v ≤ p. Also T v > 0, by (g). As p is arbitrary, T [U ] is order-dense in L0 . Q Q 368F Corollary A Riesz space U is Archimedean iff it is isomorphic T to a Riesz subspace of some reduced power RX |F, where X is a set and F is a filter on X such that n∈N Fn ∈ F whenever hFn in∈N is a sequence in F. proof (a) If U is an Archimedean Riesz space, then by 368E there is a space of the form L0 = L0 (A) such that U can be embedded into L0 . As in the proof of 364E, L0 is isomorphic to some space of the form L0 (Σ)/W, where Σ is a σ-algebra of subsets of some set X and W = {f : f ∈ L0 , {x : f (x) 6= T 0} ∈ I}, I being a σ-ideal of Σ. But now F = {A : A ∪ E = X for some E ∈ I} is a filter on X such that n∈N Fn ∈ F for every sequence hFn in∈N in F. (I am ing over the trivial case X ∈ I, since then U must be {0}.) And L0 /W is (isomorphic to) the image of L0 in RX |F, since W = {f : f ∈ L0 , {x : f (x) = 0} ∈ F}. Thus U is isomorphic to a Riesz subspace of RX |F. (b) On the other hand, if F is a filter on X closed under countable intersections, then W = {f : f ∈ RX , {x : f (x) = 0} ∈ F} is a sequentially order-closed solid linear subspace of the Dedekind σ-complete Riesz space RX , so that RX |F = RX /W is Dedekind σ-complete (353J(a-iii)) and all its Riesz subspaces must be Archimedean (353H, 351Rc).

380

Function spaces

368G

368G Corollary Every Archimedean Riesz space U is isomorphic to an order-dense Riesz subspace of some space C ∞ (X), where X is an extremally disconnected compact Hausdorff space. proof Let Z be the Stone space of the band algebra A of U . Because A is Dedekind complete (352Q), Z is extremally disconnected and A can be identified with the regular open algebra G of Z (314S). By 364W, L0 (G) can be identified with C ∞ (Z). So the embedding of U as an order-dense Riesz subspace of L0 (A) (368E) can be regarded as an embedding of U as an order-dense Riesz subspace of C ∞ (Z). 368H Corollary Any Dedekind complete Riesz space U is isomorphic to an order-dense solid linear subspace of L0 (A) for some Dedekind complete Boolean algebra A. proof Embed U in L0 = L0 (A) as in 368E; because U is order-dense in L0 and (in itself) Dedekind complete, it is solid (353K). 368I Corollary Let U be an Archimedean Riesz space. Then U can be embedded as an order-dense Riesz subspace of a Dedekind complete Riesz space V in such a way that the solid linear subspace of V generated by U is V itself, and this can be done in essentially only one way. If W is any other Dedekind complete Riesz space and T : U → W is an order-continuous positive linear operator, there is a unique positive linear operator T˜ : V → W extending T . proof By 368E, we may suppose that U is actually an order-dense Riesz subspace of L0 (A), where A is a Dedekind complete Boolean algebra. In this case, we can take V to be the solid linear subspace generated by U , that is, {v : |v| ≤ u for some u ∈ U }; being a solid linear subspace of the Dedekind complete Riesz space L0 (A), V is Dedekind complete, and of course U is order-dense in V . If W is any other Dedekind complete Riesz space and T : U → W is an order-continuous positive linear operator, then for any v ∈ V + there is a u0 ∈ U such that v ≤ u0 , so that T u0 is an upper bound for {T u : u ∈ U, 0 ≤ u ≤ v}; as W is Dedekind complete, supu∈U,0≤u≤v T u is defined in W . By 355F, T has a unique extension to an order-continuous positive linear operator from V to W . In particular, if V1 is another Dedekind complete Riesz space in which U can be embedded as an orderdense Riesz subspace, this embedding of U extends to an embedding of V ; since V is Dedekind complete, its copy in V1 must be a solid linear subspace, so if V1 is the solid linear subspace of itself generated by U , we get an identification between V and V1 , uniquely determined by the embeddings of U in V and V1 . 368J Definition If U is an Archimedean Riesz space, a Dedekind completion of U is a Dedekind complete Riesz space V together with an embedding of U in V as an order-dense Riesz subspace of V such that the solid linear subspace of V generated by U is V itself. 368I tells us that every Archimedean Riesz space U has an essentially unique Dedekind completion, so that we may speak of ‘the’ Dedekind completion of U . 368K

This is a convenient point at which to give a characterization of the Riesz spaces L0 (A).

Lemma Let A be a Dedekind σ-complete Boolean algebra. Suppose that A ⊆ L0 (A)+ is dist. If either A is countable or A is Dedekind complete, A is bounded above in L0 (A). proof If A = ∅, this is trivial; suppose that A is not empty. For n ∈ N, set an = supu∈A [[u > n]]; this is always defined; set a = inf n∈N an . Now a = 0. P P?? Otherwise, there must be a u ∈ A such that a0 = a ∩ [[u > 0]] 6= 0, since a ⊆ a0 . But now, for any n, and any v ∈ A \ {u}, a0 ∩ [[v > n]] ⊆ [[u > 0]] ∩ [[v > 0]] = 0, so that a0 ⊆ [[u > n]]. As n is arbitrary, inf n∈N [[u > n]] 6= 0, which is impossible. X XQ Q By 364Ma, A is bounded above. 368L Definition A Riesz space U is called laterally complete or universally complete if A is bounded above whenever A ⊆ U + is dist.


368N

381

368M Theorem Let U be an Archimedean Riesz space. Then the following are equiveridical: (i) there is a Dedekind complete Boolean algebra A such that U is isomorphic to L0 (A); (ii) U is Dedekind σ-complete and laterally complete; (iii) whenever V is an Archimedean Riesz space, V0 is an order-dense Riesz subspace of V and T0 : V0 → U is an order-continuous Riesz homomorphism, there is a positive linear operator T : V → U extending T0 . proof (a)(i)⇒(ii) and (i)⇒(iii) are covered by 368K and 368B. (b)(ii)⇒(i) Assume (ii). By 368E, we may suppose that U is actually an order-dense Riesz subspace of L0 = L0 (A) for a Dedekind complete Boolean algebra A. α) If u ∈ U + and a ∈ A then u × χa ∈ U . P (α P Set A = {v : v ∈ U, 0 ≤ v ≤ χa}, and let C ⊆ A be a maximal dist set; then w = sup C is defined in U , and is also the supremum in L0 . Set b = [[w > 0]]. As w ≤ χa, b ⊆ a. ?? If b 6= a, then χ(a \ b) > 0, and there is a v 0 ∈ U such that 0 < v 0 ≤ χ(a \ b); but now v 0 ∈ A and v 0 ∧ w = 0, so v 0 ∧ v = 0 for every v ∈ C, and we ought to have added v 0 to C. X X Thus [[w > 0]] = a. Now consider u0 = supn∈N u ∧ nw; as U is Dedekind σ-complete, u0 ∈ U . Since [[u0 > 0]] ⊆ a, u0 ≤ u × χa. On the other hand, u × χ[[w >

1 n ]]

× χ[[u ≤ n]] ≤ u ∧ n2 w ≤ u0

for every n ≥ 1, so, taking the supremum over n, u × χa ≤ u0 . Accordingly u × χa = u0 ∈ U , as required. Q Q β ) If w ≥ 0 in L0 , there is a u ∈ U such that 21 w ≤ u ≤ w. P P Set (β A = {u : u ∈ U, 0 ≤ u ≤ w}, C = {a : a ∈ A, a ⊆ [[u − 12 w ≥ 0]] for some u ∈ A}. Then sup A = w, so C is order-dense in A. (If a ∈ A \ {0}, either a ∩ [[w > 0]] = 0 and a ⊆ [[0 − 12 w ≥ 0]], so a ∈ C, or there is a u ∈ U such that 0 < u ≤ w × χa. In the latter case there is some n such that 2n u ≤ w and 2n+1 u 6≤ w, and now c = a ∩ [[2n u − 21 w ≥ 0]] is a non-zero member of C included in a.) Let D ⊆ C be a partition of unity and for each d ∈ D choose ud ∈ A such that d ⊆ [[ud − 12 w ≥ 0]]. By (α), ud × χd ∈ U for every d ∈ D, so u = supd∈D ud × χd ∈ U . Now u ≤ w, but also [[u − 21 w ≥ 0]] ⊇ d for every d ∈ D, so is equal to 1, and u ≥ 21 w, as required. Q Q (γγ ) Given w ≥ 0 in L0 , we can therefore choose hun in∈N , hvn in∈N inductively such that v0 = 0 and un ∈ U ,

1 (w 2

− vn ) ≤ un ≤ w − vn ,

vn+1 = vn + un

for every n ∈ N. Now hvn in∈N is a non-decreasing sequence in U and w − vn ≤ 2−n w for every n, so w = supn∈N vn ∈ U . As w is arbitrary, (L0 )+ ⊆ U and U = L0 is of the right form. (c)(iii)⇒(i) As in (b), we may suppose that U is an order-dense Riesz subspace of L0 . But now apply condition (iii) with V = L0 , V0 = U and T0 the identity operator. There is an extension T : L0 → U . If v ≥ 0 in L0 , T v ≥ T0 = u whenever u ∈ U and u ≤ v, so T v ≥ v, since v = sup{u : u ∈ U, 0 ≤ u ≤ v} in L0 . Similarly, T (T v − v) ≥ T v − v. But as T v ∈ U , T (T v) = T v and T (T v − v) = 0, so v = T v ∈ U . As v is arbitrary, U = L0 . 368N Weakly (σ, ∞)-distributive Riesz spaces We are now ready to look at the class of Riesz spaces corresponding to the weakly (σ, ∞)-distributive Boolean algebras of §316. Definition Let U be a Riesz space. Then U is weakly (σ, ∞)-distributive if whenever hAn in∈N is a S sequence of non-empty downwards-directed subsets of U + , each with infimum 0, and n∈N An has an upper bound in U , then {u : u ∈ U , for every n ∈ N there is a v ∈ An such that v ≤ u}

382

Function spaces

368N

has infimum 0 in U . S Remark Because the definition looks only at sequences hAn in∈N such that n∈N An is order-bounded, we can invert it, as follows: a Riesz space U is weakly (σ, ∞)-distributive iff whenever hAn in∈N is a sequence of non-empty upwards-directed subsets of U + , all with supremum u0 , then {u : u ∈ U + , for every n ∈ N there is a v ∈ An such that u ≤ v} also has supremum u0 . 368O Lemma Let U be an Archimedean Riesz space. Then the following are equiveridical: (i) U is not weakly (σ, ∞)-distributive; (ii) there are a u > 0 in U and a sequence hAn in∈N of non-empty downwards-directed sets, all with infimum 0, such that supn∈N un = u whenever un ∈ An for every n ∈ N. proof (ii)⇒(i) is immediate from the definition of ‘weakly (σ, ∞)-distributive’. For (i)⇒(ii), suppose that U is not weakly (σ, ∞)-distributive.S Then there is a sequence hAn in∈N of non-empty downwards-directed sets, all with infimum 0, such that n∈N An is bounded above, but A = {w : w ∈ U , for every n ∈ N there is a v ∈ An such that v ≤ w} does not have infimum 0. Let u > 0 be a lower bound for A, and set A0n = {u ∧ v : v ∈ An } for each n ∈ N. Then each A0n is a non-empty downwards-directed set with infimum 0. Let hun in∈N be a sequence such that un ∈ A0n for every n. Express each un as u ∧ vn where vn ∈ An . Let B be the set of upper bounds of {vn : n ∈ N}. Then inf w∈B,n∈N w − vn = 0, because U is Archimedean (353F), while B ⊆ A, so u ≤ w for every w ∈ B. If u0 is any upper bound for {un : n ∈ N}, then u − u0 ≤ u − u ∧ vn = (u − vn )+ ≤ (w − vn )+ = w − vn for every n ∈ N, w ∈ B. So u0 ≥ u. Thus u = supn∈N un . As hun in∈N is arbitrary, u and hA0n in∈N witness that (ii) is true. 368P Proposition (a) A regularly embedded Riesz subspace of an Archimedean weakly (σ, ∞)-distributive Riesz space is weakly (σ, ∞)-distributive. (b) An Archimedean Riesz space with a weakly (σ, ∞)-distributive order-dense Riesz subspace is weakly (σ, ∞)-distributive. (c) If U is a Riesz space such that U × separates the points of U , then U is weakly (σ, ∞)-distributive; in particular, U ∼ and U × are weakly (σ, ∞)-distributive for every Riesz space U . proof (a) Suppose that U is an Archimedean Riesz space and that V ⊆ U is a regularly embedded Riesz subspace which is not weakly (σ, ∞)-distributive. Then 368O tells us that there are a v > 0 in V and a sequence hAn in∈N of non-empty downwards-directed subsets of V , all with infimum 0 in V , such that supn∈N vn = v in V whenever vn ∈ An for every n ∈ N. Because V is regularly embedded in U , inf An = 0 Q in U for every n and supn∈N vn = v in U for every sequence hvn in∈N ∈ n∈N An , so U is not weakly (σ, ∞)distributive. Turning this round, we have (a). (b) Let U be an Archimedean Riesz space which is not weakly (σ, ∞)-distributive, and V an orderdense Riesz subspace of U . By 368O again, there are a u > 0 in U and a sequence hAn in∈N of non-empty downwards-directed sets in U , all with infimum 0, such that supn∈N un = u whenever un ∈ An for every n. Let v ∈ V be such that 0 < v ≤ u. Set Bn = {w : w ∈ V , there is some u ∈ An such that v ∧ u ≤ w ≤ v} for each n ∈ N. Because An is downwards-directed, w ∧ w0 ∈ Bn for all w, w0 ∈ Bn ; v ∈ Bn , so Bn 6= ∅; and inf Bn = 0 in V . P P Setting C = {w : w ∈ V + , there is some u ∈ An such that w ≤ (v − u)+ }, then (because V is order-dense) any upper bound for C in U is also an upper bound of {(v − u)+ : u ∈ An }. But supu∈An (v − u)+ = (v − inf An )+ = v,

368Q


383

so v = sup C in U and inf Bn = inf{v − w : w ∈ C} = 0 in U and in V . Q Q Now if vn ∈ Bn for every n ∈ N, we can choose un ∈ An such that v ∧ un ≤ vn ≤ v for every n, so that v = v ∧ u = v ∧ supn∈N un = supn∈N v ∧ un ≤ supn∈N vn ≤ v, and v = supn∈N vn . Thus hBn in∈N witnesses that V is not weakly (σ, ∞)-distributive. (c) Now suppose that U × separates the points of U . In this case U is surely Archimedean (356G). ?? If U is not weakly (σ, ∞)-distributive, there are a u > 0 in U and a sequence hAn in∈N of non-empty downwardsdirected sets, all with infimum 0, such that supn∈N un = u whenever un ∈ An for each n. Take f ∈ U × such that f (u) 6= 0; replacing f by |f | if necessary, we may suppose that f > 0. Set δ = f (u) > 0. For each n ∈ N, there is a un ∈ An such that f (un ) ≤ 2−n−2 δ. But in this case hsupi 0 in S and a sequence hAn in∈N of non-empty downwards-directed sets in S, all with infimum 0, such that u = supn∈N un whenever un ∈ An for every n. Let α > 0 be such that c = [[u > α]] 6= 0 (361Eg), and consider Bn = {[[v > α]] : v ∈ An } ⊆ A for each n ∈ N. Then each Bn is downwards-directed (because An is), and inf Bn = 0 in A (because if b is a lower bound of Bn , αχb ≤ v for every v ∈ An ). Because A is weakly (σ, ∞)-distributive, there must be some a ∈ A such that a 6⊇ c but there is, for every n ∈ N, a bn ∈ Bn such that a ⊇ bn . Take vn ∈ An such that bn = [[vn > α]], so that vn ≤ αχ1 ∨ kvn k∞ χbn ≤ αχ1 ∨ kuk∞ χa. Since u = supn∈N vn , u ≤ αχ1 ∨ kuk∞ χa. But in this case c = [[u > α]] ⊆ a, contradicting the choice of a. X X Thus S must be weakly (σ, ∞)-distributive if A is. (ii) Now suppose that S is weakly (σ, ∞)-distributive, and let hBn in∈N be a sequence of non-empty downwards-directed subsets of A, all with infimum 0. Set An = {χb : b ∈ Bn } for each n; then An ⊆ S is non-empty, downwards-directed and has infimum 0 in S, because χ : A → S is order-continuous (361Ef). Set A = {v : v ∈ S, for every n ∈ N there is a u ∈ An such that u ≤ v}, B = {b : b ∈ A, for every n ∈ N there is an a ∈ Bn such that a ⊆ b}. ?? If 0 is not the greatest lower bound of B, take a non-zero lower bound c. P Because S is weakly (σ, ∞)n distributive, inf A = 0, and there is a v ∈ A such that χc 6≤ v. Express v as i=0 αi χai , where hai ii≤n is dist, and set a = sup{ai : i ≤ n, αi ≥ 1}; then χa ≤ v, so c 6⊆ a. For each n there is a bn ∈ Bn such that χbn ≤ v. But in this case bn ⊆ a for each n ∈ N, so that a ∈ B; which means that c is not a lower bound for B. X X Thus inf B = 0 in A. As hBn in∈N is arbitrary, A is weakly (σ, ∞)-distributive.

384

Function spaces

368Q

(iii) Thus S is weakly (σ, ∞)-distributive iff A is. But S is order-dense in L∞ = L∞ (A) (363C), therefore regularly embedded (352Ne), so 368Pa-b tell us that S is weakly (σ, ∞)-distributive iff L∞ is. (b) In the same way, because S can be regarded as an order-dense Riesz subspace of L0 = L0 (A) (364K), L is weakly (σ, ∞)-distributive iff S is, that is, iff A is. 0

368R Corollary An Archimedean Riesz space is weakly (σ, ∞)-distributive iff its band algebra is weakly (σ, ∞)-distributive. proof Let U be an Archimedean Riesz space and A its band algebra. By 368E, U is isomorphic to an order-dense Riesz subspace of L0 = L0 (A). By 368P, U is weakly (σ, ∞)-distributive iff L0 is; and by 368Qb L0 is weakly (σ, ∞)-distributive iff A is. 368S Corollary If (A, µ ¯) is a semi-finite measure algebra, any regularly embedded Riesz subspace (in particular, any solid linear subspace and any order-dense Riesz subspace) of L0 (A) is weakly (σ, ∞)distributive. proof By 322F, A is weakly (σ, ∞)-distributive; by 368Qb, L0 (A) is weakly (σ, ∞)-distributive; by 368Pa, any regularly embedded Riesz subspace is weakly (σ, ∞)-distributive. 368X Basic exercises (a) Let X be an uncountable set and Σ the countable-cocountable σ-algebra of subsets of X. Show that there is a family A ⊆ L0 = L0 (Σ) such that u ∧ v = 0 for all distinct u, v ∈ A but A has no upper bound in L0 . Show moreover that if w > 0 in L0 then there is an n ∈ N such that nw 6= supu∈A u ∧ nw. b its Dedekind completion (314U). Show that L∞ (A) b can be (b) Let A be any Boolean algebra, and A ∞ identified with the Dedekind completions of S(A) and L (A). (c) Explain how to prove 368K from 368A. (d) Show that any product of weakly (σ, ∞)-distributive Riesz spaces is weakly (σ, ∞)-distributive. (e) Let A be a Dedekind complete weakly (σ, ∞)-distributive Boolean algebra. Show that a set A ⊆ L0 = L (A) is order-bounded iff h2−n un in∈N order*-converges to 0 in L0 whenever hun in∈N is a sequence in A. (Hint: use 368A. If v > 0 and v = supu∈A v ∧ 2−n u for every n, we can find a w > 0 and a sequence hun in∈N in A such that w ≤ 2−n un for every n.) 0

(f ) Give a direct proof of 368S, using the ideas of 322F, but not relying on it or on 368Q. 368Y Further exercises (a) (i) Use 364U-364V to show that if X is any compact Hausdorff space then C(X) can be regarded as an order-dense Riesz subspace of L0 (G), where G is the regular open algebra of X. (ii) Use 353M to show that any Archimedean Riesz space with order unit can be embedded as an order-dense Riesz subspace of some L0 (G). (iii) Let U be an Archimedean Riesz space and C ⊆ U + a maximal dist set, as in part (a) of the proof of 368E. For e ∈ C let Ue be the solid linear subspace of U generated by e, and let V be the solidQlinear subspace of U generated by C. Show that Q Q V can be embedded as an order-dense Riesz subspace of e∈C Ue and therefore in e∈C L0 (Ge ) ∼ = L0 ( e∈C Ge ) for a suitable family of regular open algebras Ge . (iv) Now use 368B to complete a proof of 368E. (b) Let U be any Archimedean Riesz space. Let V be the family of pairs (A, B) of non-empty subsets of U such that B is the set of upper bounds of A and A is the set of lower bounds of B. Show that V can be given the structure of a Dedekind complete Riesz space defined by the formulae (A1 , B1 ) + (A2 , B2 ) = (A, B) iff A1 + A2 ⊆ A, B1 + B2 ⊆ B, α(A, B) = (αA, αB) if α > 0, (A1 , B1 ) ≤ (A2 , B2 ) iff A1 ⊆ A2 . Show that u 7→ (]−∞, u] , [u, ∞[) defines an embedding of U as an order-dense Riesz subspace of V, so that V may be identified with the Dedekind completion of U .

368 Notes


385

(c) Work through the proof of 364U when X is compact, Hausdorff and extremally disconnected, and show that it is easier than the general case. Hence show that 368Yb can be used to shorten the proof of 368E sketched in 368Ya. (d) Let U be a Riesz space. Show that the following are equiveridical: (i) U is isomorphic, as Riesz space, to L0 (A) for some Dedekind σ-complete Boolean algebra A (ii) U is Dedekind σ-complete and has a weak order unit and whenever A ⊆ U + is countable and dist then A is bounded above in U . (e) Let U be a weakly (σ, ∞)-distributive Riesz space and V a Riesz subspace of U which is either solid or order-dense. Show that V is weakly (σ, ∞)-distributive. (f ) Show that C([0, 1]) is not weakly (σ, ∞)-distributive. (Compare 316K.) (g) Let A be a ccc weakly (σ, ∞)-distributive Boolean algebra. Suppose we have a double sequence haij i(i,j)∈N×N in A such that haij ij∈N order*-converges to ai in A for each i, while hai ii∈N order*-converges to a. Show that there is a strictly increasing sequence hn(i)ii∈N such that hai,n(i) ii∈N order*-converges to a. (h) Let U be a weakly (σ, ∞)-distributive Riesz space with the countable sup property. Suppose we have an order-bounded double sequence huij i(i,j)∈N×N in U such that huij ij∈N order*-converges to ui in U for each i, while hui ii∈N order*-converges to u. Show that there is a strictly increasing sequence hn(i)ii∈N such that hui,n(i) ii∈N order*-converges to u. (i) Let A be a ccc weakly (σ, ∞)-distributive Dedekind complete Boolean algebra. Show that there is a topology on L0 = L0 (A) such that the closure of any A ⊆ L0 is precisely the set of order*-limits of sequences in A. (j) Let U be a weakly (σ, ∞)-distributive Riesz space and f : U → R a positive linear functional; write fτ for the component of f in U × . (i) Show that for any u ∈ U + there is an upwards-directed A ⊆ [0, u], with supremum u, such that fτ (u) = supv∈A f (v). (See 356Xe, 362D.) (ii) Show that if f is strictly positive, so is fτ . (Compare 391D.) 368 Notes and comments 368A-368B are manifestations of a principle which will reappear in §375: Dedekind complete L0 spaces are in some sense ‘maximal’. If we have an order-dense subspace U of such an L0 , then any Archimedean Riesz space including U as an order-dense subspace can itself be embedded in L0 (368B). In fact this property characterizes Dedekind complete L0 spaces (368M). Moreover, any Archimedean Riesz space U can be embedded in this way (368E); by 368C, the L0 space (though not the embedding) is unique up to isomorphism. If U and V are Archimedean Riesz spaces, each embedded as an order-dense Riesz subspace of a Dedekind complete L0 space, then any order-continuous Riesz homomorphism from U to V extends uniquely to the L0 spaces (368B). If one Dedekind complete L0 space is embedded as an order-dense Riesz subspace of another, they must in fact be the same (368D). Thus we can say that every Archimedean Riesz space U can be extended to a Dedekind complete L0 space, in a way which respects order-continuous Riesz homomorphisms, and that this extension is maximal, in that U cannot be order-dense in any larger space. The proof of 368E which I give is long because I am using a bare-hands approach. Alternative methods shift the burdens. For instance, if we take the trouble to develop a direct construction of the ‘Dedekind completion’ of a Riesz space (368Yb), then we need prove the theorem only for Dedekind complete Riesz spaces. A more substantial aid is the representation theorem for Archimedean Riesz spaces with order unit (353M); I sketch an argument in 368Ya. The drawback to this approach is the proof of Theorem 364U, which seems to be quite as long as the direct proof of 368E which I give here. Of course we need 364U only for compact Hausdorff spaces, which are usefully easier than the general case (364V, 368Yc). 368G is a version of Ogasawara’s representation theorem for Archimedean Riesz spaces. Both this and 368F can be regarded as expressions of the principle that an Archimedean Riesz space is ‘nearly’ a space of functions. I have remarked before on the parallels between the theories of Boolean algebras and Archimedean Riesz spaces. The notion of ‘weak (σ, ∞)-distributivity’ is one of the more striking correspondences. (Compare, for

386

Function spaces

368 Notes

instance, 316Xm with 368Pa.) What is really important to us, of course, is the fact that the function spaces of measure theory are mostly weakly (σ, ∞)-distributive, by 368S. Of course this is easy to prove directly (368Xf), but I think that the argument through 368Q gives a better idea of what is really happening here. Some of the features of ‘order*-convergence’, as defined in §367, are related to weak (σ, ∞)-distributivity; in 368Yi I describe a topology which can be thought of as an abstract version of the topology of convergence in measure on the L0 space of a σ-finite measure algebra (367N).

369 Banach function spaces In this section I continue the work of §368 with results which involve measure algebras. The first step is a modification of the basic representation theorem for Archimedean Riesz spaces. If U is any Archimedean Riesz space, it can be represented as a subspace of L0 = L0 (A), where A is its band algebra (368E); now if U × separates the points of U , there is a measure rendering A a localizable measure algebra (369A, 369Xa). Moreover, we get a simultaneous representation of U × as a subspace of L0 (369C-369D), the duality between them corresponding exactly to the familiar duality between Lp and Lq . Still drawing inspiration from the classical Lp spaces, we have a general theory of ‘associated Fatou norms’ (369F-369M, 369R). I include notes on the spaces M 1,∞ , M ∞,1 and M 1,0 (369N-369Q), which will be particularly useful in the next chapter. 369A Theorem Let U be a Riesz space such that U × separates the points of U . Then U can be embedded as an order-dense Riesz subspace of L0 (A) for some localizable measure algebra (A, µ ¯). proof (a) Consider the canonical map S : U → U ×× . We know that this is a Riesz homomorphism onto an order-dense Riesz subspace of U ×× (356I). Because U × separates the points of U , S is injective. Let A be the band algebra of U ×× and T : U ×× → L0 = L0 (A) an injective Riesz homomorphism onto an order-dense Riesz subspace V of L0 , as in 368E. The composition T S : U → L0 is now an injective Riesz homomorphism, so embeds U as a Riesz subspace of L0 , which is order-dense because V is order-dense in L0 and T S[U ] is order-dense in V (352Nc). Thus all that we need to find is a measure µ ¯ on A rendering it a localizable measure algebra. (b) Note that V is isomorphic, as Riesz space, to U ×× , which is Dedekind complete (356B), so V must be solid in L0 (353K). Also V × must separate the points of V (356L). Let D be the set of those d ∈ A such that the principal ideal Ad is measurable in the sense that there is some ν¯ for which (Ad , ν¯) is a totally finite measure algebra. Then D is order-dense in A. P P Take any non-zero a ∈ A. Because V is order-dense, there is a non-zero v ∈ V such that v ≤ χa. Take h ≥ 0 in V × such that h(v) > 0. Then there is a v 0 such that 0 < v 0 ≤ v and h(w) > 0 whenever 0 < w ≤ v 0 in V (356H). Let α > 0 be such that d = [[v 0 > α]] 6= 0. Then χb ≤ α1 v 0 ∈ V whenever b ∈ Ad . Set ν¯b = h(χb) ∈ [0, ∞[ for b ∈ Ad . Because the map b 7→ χb : A → L0 is additive and order-continuous, the map b P 7→ χb : Ad → V also ∞ is, and ν¯ = hχ must be additive and order-continuous; in particular, ν¯(supn∈N bn ) = n=0 ν¯bn whenever 0 hbn in∈N is a dist sequence in Ad . Moreover, if b ∈ Ad is non-zero, then 0 < αχb ≤ v , so ν¯b = h(χb) > 0. Thus (Ad , ν¯) is a totally finite measure algebra, and d ∈ D, while 0 6= d ⊆ a. As a is arbitrary, D is order-dense. Q Q (c) By 313K, there is a partition of unity C ⊆ D. For each c ∈ C, let ν¯c : Ac → [0, ∞[ be P a functional such that (Ac , ν¯c ) is a totally finite measure algebra. Define µ ¯ : A → [0, ∞]Pby setting µ ¯a = c∈C ν¯c (a ∩ c) for every a ∈ A. Then (A, µ ¯) is a localizable measure algebra. P P (i) µ ¯0 = c∈C ν¯0 = 0. (ii) If han in∈N is a dist sequence in A with supremum a, then P P P∞ µ ¯a = c∈C ν¯c (a ∩ c) = c∈C,n∈N ν¯c (an ∩ c) = n=0 µ ¯an . (iii) If a ∈ A \ {0}, then there is a c ∈ C such that a ∩ c 6= 0, so that µ ¯a ≥ ν¯c (a ∩ c) > 0. Thus (A, µ ¯) is a measure algebra. (iv) Moreover, in (iii), µ ¯(a ∩ c) = ν¯c (a ∩ c) is finite. So (A, µ ¯) is semi-finite. (v) A is Dedekind complete, being a band algebra (352Q), so (A, µ ¯) is localizable. Q Q

369C

Banach function spaces

387

369B Corollary Let U be a Banach lattice with order-continuous norm. Then U can be embedded as an order-dense solid linear subspace of L0 (A) for some localizable measure algebra (A, µ ¯). proof By 356Dd, U × = U ∗ , which separates the points of U , by the Hahn-Banach theorem (3A5Ae). So 369A tells us that U can be embedded as an order-dense Riesz subspace of an appropriate L0 (A). But also U is Dedekind complete (354Ee), so its copy in L0 (A) must be solid, as in 368H. 369C The representation in 369A is complemented by the following result, which is a kind of generalization of 365J and 366Dc. Theorem Let (A, µ ¯) be a semi-finite measure algebra, and U ⊆ L0 = L0 (A) an order-dense Riesz subspace. Set V = {v : v ∈ L0 , v × u ∈ L1 for every u ∈ U }, writing L1 for L1 (A, µ ¯) ⊆ L0 . Then V is a solid linear subspace of L0 , and we have an order-continuous injective Riesz homomorphism T : V → U × defined by setting (T v)(u) =

R

u × v for all u ∈ U , v ∈ V .

×

The image of V is order-dense in U . If (A, µ ¯) is localizable, then T is surjective, so is a Riesz space isomorphism between V and U × . proof (a)(i) Because × : L0 × L0 → L0 is bilinear and L1 is a linear subspace of L0 , V is a linear subspace of L0 . If u ∈ U , v ∈ V , w ∈ L0 and |w| ≤ |v|, then |w × u| = |w| × |u| ≤ |v| × |u| = |v × u| ∈ L1 ; as L1 is solid, w × u ∈ L1 ; as u is arbitrary, w ∈ V ; this shows that V is solid.

R (ii) By the definition of V , (T v)(u) is defined in R for all u ∈ U , v ∈ V . Because × is bilinear and is linear, T v : U → R is linear for every v ∈ V , and T is a linear functional from V to the space of linear operators from U to R. R (iii) If u ≥ 0 in U and v ≥ 0 in V , then u × v ≥ 0 in L1 and (T v)(u) = u × v ≥ 0. This shows that T is a positive linear operator from V to U ∼ . (iv) If v ≥ 0 in V and A ⊆ U is a non-empty downwards-directed set with infimum 0 in U , then inf A = 0 in L0 , because U is order-dense (352Nb). Consequently inf u∈A u × v = 0 in L0 and in L1 (364P), and (because

R

inf u∈A (T v)(u) = inf u∈A

R

u×v =0

is order-continuous). As A is arbitrary, T v is order-continuous. As v is arbitrary, T [V ] ⊆ U × .

(v) If v ∈ V and u0 ≥ 0 in U , set a = [[v > 0]]. Then v + = v×χa. SetR A = {u : u ∈ U, 0 ≤ u ≤ u0 ×χa}. Because U is order-dense in L0 , u0 × χa = sup A in L0 . Because × and are order-continuous, Z + (T v) (u0 ) ≥ sup (T v)(u) = sup v × u u∈A u∈A Z Z = v × u0 × χa = v + × u0 = (T v + )(u0 ). As u0 is arbitrary, (T v)+ ≥ T v + . But because T is a positive linear operator, we must have T v + ≥ (T v)+ , so that T v + = (T v)+ . As v is arbitrary, T is a Riesz homomorphism. (vi) Now T is injective. P P If v 6= R 0 in V , there is a u > 0 in U such that u ≤ |v|, because U is order-dense. In this case u × |v| > 0 so u × |v| > 0. Accordingly |T v| = T |v| 6= 0 and T v 6= 0. Q Q (b) Putting (a-i) to (a-vi) together, we see that T is an injective Riesz homomorphism from V to U × . All this is easy. The point of the theorem is the fact that T [V ] is order-dense in U × . P P Take h > 0 in U × . Let U1 be the solid linear subspace of L0 generated by U . Then U is an order-dense Riesz subspace of U1 , h : U → R is an order-continuous positive linear functional, and sup{h(u) : u ∈ U, 0 ≤ ˜ of h to U1 such that h ˜ ∈ U × (355F). u ≤ v} is defined in R for every v ≥ 0 in U1 ; so we have an extension h 1

388

Function spaces

369C

Set S1 = S(A) ∩ U1 ; then S1 is an order-dense Riesz subspace of U1 , because S(A) is order-dense in L0 and U1 is solid in L0 . Note that S1 is the linear span of {χc : c ∈ I}, where I = {c : c ∈ A, χc ∈ U1 }, and that I is an ideal in A. ˜ 6= 0; there must therefore be a u0 ∈ S1 such that h(u ˜ 0 ) > 0, and a d ∈ I such that Because h 6= 0, h ˜ ˜ ˜ h(χd) > 0. For a ∈ A, set νa = hχ(d ∩ a). Because ∩ , χ and h are all order-continuous, so is ν, and ν : A → R is a non-negative completely additive functional. By 365Ea, there is a v ∈ L1 (A, µ ¯) such that

R

a

˜ v = νa = hχ(d ∩ a)

R ˜ for every a ∈ A; of course v ≥ 0. We have u × v ≤ h(u) whenever u = χa for a ∈ I, and therefore for every + + 0 0 0 u ∈ (S1 ) . If u ∈ U , then A = {u : u ∈ S1 , 0 ≤ u ≤ u} is upwards-directed, sup A = u and

R

˜ 0 ) = h(u) ˜ v × u0 ≤ supu0 ∈A h(u = h(u) R is finite, so v × u = supu∈A0 v × u0 belongs to L1 (365Df) and v × u ≤ h(u). As u is arbitrary, v ∈ V and R ˜ T v ≤ h. At the same time, d v = h(χd) > 0, so T v > 0. As h is arbitrary, T [V ] is order-dense. Q Q It follows that T is order-continuous (352Nb), as can also be easily proved by the argument of (a-iv) above. supu0 ∈A

(c) Now suppose that (A, µ ¯) is localizable, that is, that A is Dedekind complete. T −1 : T [V ] → V is a Riesz space isomorphism, so certainly an order-continuous Riesz homomorphism; because V is a solid linear subspace of L0 , T −1 is still an injective order-continuous Riesz homomorphism when regarded as a map from T [V ] to L0 . Since T [V ] is order-dense in U × , T −1 has an extension to an order-continuous Riesz homomorphism Q : U × → L0 (368B). But Q[U × ] ⊆ V . P P Take h ≥ 0 in U × and u ≥ 0 in U . Then B = {g : g ∈ T [V ],R 0 ≤ g ≤ h} is upwards-directed and has supremum h. For g ∈ B, we know that u × T −1 g ∈ L1 and u × T −1 g = g(u), by the definition of T . But this means that supg∈B

R

u × T −1 g = supg∈B g(u) = h(u) < ∞.

Since {u × T −1 g : g ∈ B} is upwards-directed, it follows that u × Qh = supg∈B u × Qg = supg∈B u × T −1 g ∈ L1 by 365Df again. As u is arbitrary, Qh ∈ V . As h is arbitrary (and Q is linear), Q[U × ] ⊆ V . Q Q Also Q is injective. P P If h ∈ U × is non-zero, there is a v ∈ V such that 0 < T v ≤ |h|, so that |Qh| = Q|h| ≥ QT v = v > 0 and Qh 6= 0. Q Q Since QT is the identity on V , Q and T must be the two halves of a Riesz space isomorphism between V and U × . 369D Corollary Let U be any Riesz space such that U × separates the points of U . Then there is a localizable measure algebra (A, µ ¯) such that the pair (U, U × ) can be represented by a pair (V, W ) of order0 dense Riesz subspaces of L = L0 (A) such that W = {w : w ∈ L0 , v × w ∈ L1 for every v ∈ V }, writing L1 for L1 (A, µ ¯). In this case, U ×× becomes represented by V˜ = {v : v ∈ L0 , v × w ∈ L1 for every w ∈ W } ⊇ V . proof Put 369A and 369C together. The construction of 369A finds (A, µ ¯) and an order-dense V which is isomorphic to U , and 369C identifies W with V × . To check that W is order-dense, take any u > 0 in L0 . There is a v ∈ V such that 0 < v ≤ u. There is an h ∈ (V × )+ such that h(v) > 0, so there is a w ∈ W + such that w × v 6= 0, that is, w ∧ v 6= 0. But now w ∧ v ∈ W , because W is solid, and 0 < w ∧ v ≤ u. Remark Thus the canonical embedding of U in U ×× (356I) is represented by the embedding V ⊆ V˜ ; U , or V , is ‘perfect’ iff V = V˜ . 369E Kakutani’s Theorem (Kakutani 41) If U is any L-space, there is a localizable measure algebra (A, µ ¯) such that U is isomorphic, as Banach lattice, to L1 = L1 (A, µ ¯). R R proof U is a perfect Riesz space, and U × = U ∗ has an order unit defined by saying that u = kuk for u ≥ 0 (356P). By 369D, we can find a localizable measure algebra (A, µ ¯) and an identification of the pair (U, U × ), as dual Riesz spaces, with a pair (V, W ) of subspaces of L0 = L0 (A); and V will be {v : v×w ∈ L1 for

369H


389

every w ∈ W }. But W , like U × , must have an order unit; call it e. Because W is order-dense, [[e > 0]] must be 1 and e must have a multiplicative inverse 1e in L0 (364P). This means that V must be {v : v × e ∈ L1 }, so that v 7→ v × e is a Riesz space isomorphism between V and L1 , which gives a Riesz space isomorphism between U and L1 . Moreover, if we write k k0 for the norm on V corresponding to the norm of U , we have kuk =

R

|u| for u ∈ U ,

kvk0 =

R

|v| × e =

R

|v × e| for v ∈ V .

1

Thus the Riesz space isomorphism between U and L is norm-preserving, and U and L1 are isomorphic as Banach lattices. 369F The Lp spaces are leading examples for a general theory of normed subspaces of L0 , which I proceed to sketch in the rest of the section. Definition Let A be a Dedekind σ-complete Boolean algebra. An extended Fatou norm on L0 = L0 (A) is a function τ : L0 → [0, ∞] such that (i) τ (u + v) ≤ τ (u) + τ (v) for all u, v ∈ L0 ; (ii) τ (αu) = |α|τ (u) for all u ∈ L0 , α ∈ R (counting 0 · ∞ as 0, as usual); (iii) τ (u) ≤ τ (v) whenever |u| ≤ |v| in L0 ; (iv) supu∈A τ (u) = τ (v) whenever A ⊆ (L0 )+ is a non-empty upwards-directed set with supremum v in 0 L ; (v) τ (u) > 0 for every non-zero u ∈ L0 ; (vi) whenever u > 0 in L0 there is a v ∈ L0 such that 0 < v ≤ u and τ (v) < ∞. 369G Proposition Let (A, µ ¯) be a Dedekind σ-complete Boolean algebra and τ an extended Fatou norm on L0 = L0 (A). Then Lτ = {u : u ∈ L0 , τ (u) < ∞} is an order-dense solid linear subspace of L0 , and τ , restricted to Lτ , is a Fatou norm under which Lτ is a Banach lattice. If hun in∈N is a non-decreasing norm-bounded sequence in (Lτ )+ , then it has a supremum in Lτ ; if A is Dedekind complete, then Lτ has the Levi property. proof (a) By (i), (ii) and (iii) of 369F, Lτ is a solid linear subspace of L0 ; by (vi), it is order-dense. Hypotheses (i), (ii), (iii) and (v) show that τ is a Riesz norm on Lτ , while (iv) shows that it is a Fatou norm. (b)(i) Suppose that hun in∈N is a non-decreasing norm-bounded sequence in (Lτ )+ . Then u = supn∈N un is defined in L0 . P P?? Otherwise, there is a v > 0 in L0 such that kv = supn∈N kv ∧un for every k ∈ N (368A). By (v)-(vi) of 369F, there is a v 0 such that 0 < v 0 ≤ v and 0 < τ (v 0 ) < ∞. Now kv 0 = supn∈N kv 0 ∧ un for every k, so kτ (v 0 ) = τ (kv 0 ) = supn∈N τ (kv 0 ∧ un ) ≤ supn∈N τ (un ) for every k, using 369F(iv), and supn∈N τ (un ) = ∞, contrary to hypothesis. X XQ Q By 369F(iv) again, τ (u) = supn∈N τ (un ) < ∞, so that u ∈ Lτ and u = supn∈N un in Lτ . (ii) It follows that Lτ is complete P Let hun in∈N be a sequence in Lτ such that τ (un+1 −un ) ≤ Pn under τ . P 2 for every n ∈ N. Set vmn = i=m |ui+1 − ui | for m ≤ n; then τ (vmn ) ≤ 2−m+1 for every n, so by (i) just above vm = supn∈N vmn is defined in Lτ , and τ (vm ) ≤ 2−m+1 . Now vm = |um+1 − um | + vm+1 for each m, so hum − vm im∈N is non-decreasing and hum + vm im∈N is non-increasing, while um − vm ≤ um ≤ um + vm for every m. Accordingly u = supm∈N um − vm is defined in Lτ and |u − um | ≤ vm for every m. But this means that limm→∞ τ (u − um ) ≤ limm→∞ τ (vm ) = 0 and u = limm→∞ um in Lτ . As hun in∈N is arbitrary, Lτ is complete. Q Q −n

(c) Now suppose that A is Dedekind complete and A ⊆ (Lτ )+ is a non-empty upwards-directed normbounded set in Lτ . By the argument of (b-i) above, using the other half of 368A, sup A is defined in L0 and belongs to Lτ . As A is arbitrary, Lτ has the Levi property. 369H Associate norms: Definition Let (A, µ ¯) be a semi-finite measure algebra, and τ an extended Fatou norm on L0 = L0 (A). Define τ 0 : L0 → [0, ∞] by setting τ 0 (u) = sup{ku × vk1 : v ∈ L0 , τ (v) ≤ 1}

390

Function spaces

369H

for every u ∈ L0 ; then τ 0 is the associate of τ . (The word suggests a symmetric relationship; it is justified by the next theorem.) 369I Theorem Let (A, µ ¯) be a semi-finite measure algebra, and τ an extended Fatou norm on L0 = L (A). Then (i) its associate τ 0 is also an extended Fatou norm on L0 ; (ii) τ is the associate of τ 0 ; (iii) ku × vk1 ≤ τ (u)τ 0 (v) for all u, v ∈ L0 . 0

proof (a) Before embarking on the proof that τ 0 is an extended Fatou seminorm on L0 , I give the greater part of the argument needed to show that τ = τ 00 , where τ 00 (u) = sup{ku × wk1 : w ∈ L0 , τ 0 (w) ≤ 1} for every u ∈ L0 . (i) Set B = {u : u ∈ L1 , τ (u) ≤ 1}. Then B is a convex set in L1 = L1 (A, µ ¯) and is closed for the norm topology of L1 . P P Suppose that u 1 belongs to the closure of B in L . Then for each n ∈ N we can choose un ∈ B such that ku − un k1 ≤ 2−n . Set vmn = inf m≤i≤n |ui | for m ≤ n, and vm = inf n≥m vmn = inf n≥m |un | ≤ |u| for m ∈ N. The sequence hvm im∈N is non-decreasing, τ (vm ) ≤ τ (um ) ≤ 1 for every m, and P∞ P∞ k|u| − vm k1 ≤ supn≥m k|u| − vmn k1 ≤ i=m k|u| − |ui |k1 ≤ i=m ku − ui k1 → 0 as m → ∞. So |u| = supm∈N vm in L0 , τ (u) = τ (|u|) = supm∈N τ (vm ) ≤ 1 and u ∈ B. Q Q (ii) Now take any u0 ∈ L0 such that τ (u0 ) > 1. Then, writing Af for {a : µ ¯a < ∞}, A = {u : u ∈ S(Af ), 0 ≤ u ≤ u0 } is an upwards-directed set with supremum u0 (this is where I use the hypothesis that (A, µ ¯) is semi-finite, so that S(Af ) is order-dense in L0 ), and supu∈A τ (u) = τ (u0 ) > 1. Take u1 ∈ A such that τ (u1 ) > 1, that is, u1 ∈ / B. By the Hahn-Banach theorem (3A5Cc), there is a continuous linear functional f : L1 → R such that f (u1 ) > 1 but f (u) ≤ 1 for every u ∈ B. Because (L1 )∗ = (L1 )∼ (356Dc), |f | is defined in (L1 )∗ , and of course |f |(u1 ) ≥ f (u1 ) > 1,

|f |(u) = sup{f (v) : |v| ≤ u} ≤ 1

whenever u ∈ B and u ≥ 0. Set c = [[u1 > 0]], so that µ ¯c < ∞, and define νa = |f |(χ(a ∩ c)) 1 for every R a ∈ A. Then ν is a completely additive real-valued functional on A, so there is a w ∈ L such that νa = a w for every a ∈ A (365Ea). Because νa ≥ 0 for every a, w ≥ 0. Now R w = |f |(χa × χc) a

for every a ∈ A, so

R

w × u = |f |(u × χc) ≤ |f |(u) ≤ 1

for every u ∈ S(A)+ ∩ B. But if τ (v) ≤ 1, then Av = {u : u ∈ S(A)+ ∩ B, u ≤ |v|} is an upwards-directed set with supremum |v|, so that kw × vk1 = supu∈Av

R

w × u ≤ 1.

369K


Thus τ 0 (w) ≤ 1. On the other hand,

R

kw × u0 k1 ≥

w × u0 ≥

R

391

w × u1 = |f |(u1 ) > 1,

so τ 00 (u0 ) > 1. (iii) This shows that, for u ∈ L0 , τ 00 (u) ≤ 1 =⇒ τ (u) ≤ 1. (c) Now I return to the proof that τ 0 is an extended Fatou norm. It is easy to check that it satisfies conditions (i)-(iv) of 369F; in effect, these depend only on the fact that k k1 is an extended Fatou norm. For (v)-(vi), take v > 0 in L0 . Then there is a u such that 0 ≤ u ≤ v and 0 < τ (u) < ∞; set α = 1/τ (u). Then τ (2αu) > 1, so that τ 00 (2αu) > 1 and there is a w ∈ L0 such that τ 0 (w) ≤ 1, k2αu × wk1 > 1. But now set v1 = v ∧ |w|; then v ≥ v1 ≥ u ∧ |w| > 0, 0

while τ (v1 ) < ∞. Also v ∧ αu 6= 0 so τ 0 (v) ≥ kv1 × αuk1 > 0. As v is arbitrary, τ 0 satisfies 369F(v)-(vi). (d) Accordingly τ 00 is also an extended Fatou norm. Now in (a) I showed that τ 00 (u) ≤ 1 =⇒ τ (u) ≤ 1. It follows easily that τ (u) ≤ τ 00 (u) for every u (since otherwise there would be some α such that τ 00 (αu) = ατ 00 (u) < 1 < ατ (u) = τ (αu).) On the other hand, we surely have τ (u) ≤ 1 =⇒ ku × vk1 ≤ 1 whenever τ 0 (v) ≤ 1 =⇒ τ 00 (u) ≤ 1, so we must also have τ 00 (u) ≤ τ (u) for every u. Thus τ 00 = τ , as claimed. (e) Of course we have ku × vk1 ≤ 1 whenever τ (u) ≤ 1 and τ 0 (v) ≤ 1. It follows easily that ku × vk1 ≤ τ (u)τ 0 (v) whenever u, v ∈ L0 and both τ (u), τ 0 (v) are non-zero. But if one of them is zero, then u × v = 0, because both τ and τ 0 satisfy (v) of 369F, so the result is trivial. 369J Theorem Let (A, µ ¯) be a semi-finite measure algebra, and τ an extended Fatou norm on L0 = L (A), with associate θ. Then 0

Lθ = {v : v ∈ L0 , u × v ∈ L1 for every u ∈ Lτ }. proof (a) If v ∈ Lθ and u ∈ Lτ , then ku × vk1 is finite, by 369I(iii), so u × v ∈ L1 . (b) P If v ∈ / Lθ then for every n ∈ N there is a un such that τ (un ) ≤ 1 and kun × vk1 ≥ 2n . Set n wn = i=0 2−i |ui | for each n. Then hwn in∈N Ris a non-decreasing sequence and τ (wn ) ≤ 2 for each n, so w = supn∈N wn is defined in Lτ , by 369G; now w × |v| ≥ n + 1 for every n, so w × v ∈ / L1 . 369K Corollary Let (A, µ ¯) be a localizable measure algebra, and τ an extended Fatou norm on L0 = L0 (A), with associate θ. Then Lθ may be identified, as normed Riesz space, with (Lτ )× ⊆ (Lτ )∗ , and Lτ is a perfect Riesz space. proof Putting 369J and 369C together, we have an identification between Lθ and (Lτ )× . Now 369I tells us that τ is the associate of θ, so that we can identify Lτ with (Lθ )× , and Lτ is perfect, as in 369D. By the definition of θ, we have, for any v ∈ Lθ , θ(v) = sup ku × vk1 τ (u)≤1

=

sup τ (u)≤1,kwk∞ ≤1

Z

Z u × v × w = sup τ (u)≤1

which is the norm of the linear functional on Lτ corresponding to v.

u × v,

392

Function spaces

369L

369L Lp I remarked above that the Lp spaces are leading examples for this theory; perhaps I should spell out the details. Let (A, µ ¯) be a semi-finite measure algebra and p ∈ [1, ∞]. Extend the functional k kp , as defined in 364K, 365A and 366A, by saying that kukp = ∞ if u ∈ L0 \ Lp . (For p = 1 this is already done by the convention in 365A.) To see that k kp is now an extended Fatou norm, as defined in 369F, we note that conditions (i)-(iii) and (v) there are true just because Lp is a solid linear subspace on which k kp is a Riesz norm, (iv) is true because k kp is a Fatou norm with the Levi property (363Ba, 365C, 366D), and (vi) is true because S(Af ) is included in Lp and order-dense in L0 (364L). As usual, set q = p/(p − 1) if 1 < p < ∞, ∞ if p = 1, and 1 if p = ∞. Then k kq is the associate extended Fatou norm of k kp . P P By 365Jb and 366C, kvkq = sup{ku × vk1 : kukp ≤ 1} for every v ∈ Lq . But as Lq 0 is order-dense in L , kvkq =

sup w∈Lq ,|w|≤v

kwkq

Z

= sup{

Z q

|u| × |w| : w ∈ L , w ≤ |v|, kukp ≤ 1} = sup{

|u| × |v| : kukp ≤ 1}

for every v ∈ L0 . Q Q 369M Proposition Let (A, µ ¯) be a semi-finite measure algebra and τ an extended Fatou norm on L0 = L0 (A, µ ¯). Then (a) the embedding Lτ ⊆ L0 is continuous for the norm topology of Lτ and the topology of convergence in measure on L0 ; (b) τ : L0 → [0, ∞] is lower semi-continuous, that is, all the balls {u : τ (u) ≤ γ} are closed for the topology of convergence in measure; (c) if hun in∈N is a sequence in L0 which is order*-convergent to u ∈ L0 (definition: 367A), then τ (u) ≤ lim inf n→∞ τ (un ). proof (a) This is a special case of 367P. (b) Set Bγ = {u : τ (u) ≤ γ}. If u ∈ L0 \ Bγ , then A = {|u| × χa : a ∈ Af } is an upwards-directed set with supremum |u|, so there is an a ∈ Af such that τ (u × χa) > γ. ?? If u is in the closure of Bγ for the topology of convergence in measure, then for every k ∈ N there is a vk ∈ Bγ such that µ ¯(a ∩ [[|u − vk | > 2−k ]]) ≤ 2−k (see the formulae in 367M). Set vk0 = |u| ∧ inf i≥k |vi | for each k, and v ∗ = supk∈N vk0 . Then τ (vk0 ) ≤ τ (vk ) ≤ γ for each k, and hvk ik∈N is non-decreasing, so τ (v ∗ ) ≤ γ. But P∞

a ∩ [[|u| − v ∗ > 2−k ]] ⊆ a ∩ supi≥k [[|u − vi | > 2−k ]]

−i has measure at most for each k, so a ∩ [[|u| − v ∗ > 0]] must be 0, that is, |u| × χa ≤ v ∗ and i=k 2 τ (|u| × χa) ≤ γ; contrary to the choice of a. X X Thus u cannot belong to the closure of Bγ . As u is arbitrary, Bγ is closed.

(c) If hun in∈N order*-converges to u, it converges in measure (367Na). If γ > lim inf n→∞ τ (un ), there is a subsequence of hun in∈N in Bγ , and τ (u) ≤ γ, by (b). As γ is arbitary, τ (u) ≤ lim inf n→∞ τ (un ). 369N contexts.

I now turn to another special case which we have already had occasion to consider in other

Definition Let (A, µ ¯) be a measure algebra. Set Mµ¯∞,1 = M ∞,1 (A, µ ¯) = L1 (A, µ ¯) ∩ L∞ (A), Mµ¯1,∞ = M 1,∞ (A, µ ¯) = L1 (A, µ ¯) + L∞ (A), and

369O


393

kuk∞,1 = max(kuk1 , kuk∞ ) 0

for u ∈ L (A). Remark I hope that the notation I have chosen here will not completely overload your short-term memory. The idea is that in M p,q the symbol p is supposed to indicate the ‘local’ nature of the space, that is, the nature of u × χa where u ∈ M p,q and µ ¯a < ∞, while q indicates the nature of |u| ∧ χ1 for u ∈ M p,q . Thus 1,∞ M is the space of u such that u × χa ∈ L1 for every a ∈ Af , |u| ∧ χ1 ∈ L∞ ; in M 1,0 we demand further that |u| ∧ χ1 ∈ M 0 (366F); while in M ∞,1 we ask that |u| ∧ χ1 ∈ L1 , u × χa ∈ L∞ for every a ∈ Af . 369O Proposition Let (A, µ ¯) be a semi-finite measure algebra. (a) k k∞,1 is an extended Fatou norm on L0 = L0 (A). (b) Its associate k k1,∞ may be defined by the formulae kuk1,∞ = min{kvk1 + kwk∞ : v ∈ L1 , w ∈ L∞ , v + w = u} Z = min{α + (|u| − αχ1)+ : α ≥ 0} Z ∞ = min(1, µ ¯[[|u| > α]])dα 0 0

1

1

for every u ∈ L , writing L = L (A, µ ¯), L∞ = L∞ (A). (c) {u : u ∈ L0 , kuk1,∞ < ∞} = M 1,∞ = M 1,∞ (A, µ ¯), {u : u ∈ L0 , kuk∞,1 < ∞} = M ∞,1 = M ∞,1 (A, µ ¯). (d) Writing Af = {a : µ ¯a < ∞}, S(Af ) is norm-dense in M ∞,1 and S(A) is norm-dense in M 1,∞ . (e) For any p ∈ [1, ∞], kuk1,∞ ≤ kukp ≤ kuk∞,1 0

for every u ∈ L . Remark By writing ‘min’ rather than ‘inf’ in the formulae of part (b) I mean to assert that the infima are attained. proof (a) This is easy; all we need to know is that k k1 and k k∞ are extended Fatou norms. (b) We have four functionals on L0 to look at; let me give them names: τ1 (u) = sup{ku × vk1 : kvk1,∞ ≤ 1}, τ2 (u) = inf{ku0 k1 + ku00 k∞ : u = u0 + u00 }, τ3 (u) = inf α≥0 (α + τ4 (u) =

R∞ 0

R

(|u| − αχ1)+ ),

min(1, µ ¯[[|u| > α]])dα.

(I write ‘inf’ here to avoid the question of attainment for the moment.) Now we have the following. (i) τ1 (u) ≤ τ2 (u). P P If kvk1,∞ ≤ 1 and u = u0 + u00 , then ku × vk1 ≤ ku0 × vk1 + ku00 × vk1 ≤ ku0 k1 kvk∞ + ku00 k∞ kvk1 ≤ ku0 k1 + ku00 k∞ . Taking the supremum over v and the infinum over u0 and u00 , τ1 (u) ≤ τ2 (u). Q Q (ii) τ2 (u) ≤ τ4 (u). P P If τ4 (u) = ∞ this is trivial. Otherwise, take w such that kwk∞ ≤ 1 and u = |u| × w. Set α0 = inf{α : µ ¯[[|u| > α]] ≤ 1}, and try u0 = w × (|u| − α0 χ1)+ , Then u = u0 + u00 ,

u00 = w × (|u| ∧ α0 χ1).

394

Function spaces

Z

369O

∞

0

µ ¯[[|u0 | > α]]dα

ku k1 = 0

Z

∞

= Z0 ∞ =

µ ¯[[|u| > α + α0 ]]dα Z ∞ µ ¯[[|u| > α]]dα = min(1, µ ¯[[|u| > α]])dα,

α0

α0

ku00 k∞ ≤ α0 =

R α0 0

min(1, [[|u| > α]])dα,

so τ2 (u) ≤ ku0 k1 + ku00 k∞ ≤ τ4 (u). Q Q (iii) τ4 (u) ≤ τ3 (u). P P For any α ≥ 0, Z α Z ∞ τ4 (u) = min(1, µ ¯[[|u| > β]])dβ + min(1, µ ¯[[|u| > β]])dβ 0 α Z ∞ ≤α+ µ ¯[[|u| > α + β]]dβ 0 Z ∞ =α+ µ ¯[[|u| > α + β]]dβ 0 Z ∞ Z =α+ µ ¯[[(|u| − αχ1)+ > β]]dβ = α + (|u| − αχ1)+ . 0

Taking the infimum over α, τ4 (u) ≤ τ3 (u). Q Q (iv) τ3 (u) ≤ τ1 (u). α) It is enough to consider the case 0 < τ1 (u) < ∞, because if τ1 (u) = 0 then u = 0 and evidently P P(α τ3 (0) = 0, while if τ1 (u) = ∞ the required inequality is trivial. Furthermore, since τ3 (u) = τ3 (|u|) and τ1 (u) = τ1 (|u|), it is enough to consider the case u ≥ 0. R 1 β ) Note next that if µ (β ¯a < ∞, then k max(1,¯ ¯a)τ1 (u). µa) χak∞,1 ≤ 1, so that a u ≤ max(1, µ (γγ ) Set c = [[u > 2τ1 (u)]]. If a ⊆ c and µ ¯a < ∞, then 2τ1 (u)¯ µa ≤

R

a

u ≤ max(1, µ ¯a)τ1 (u),

so µ ¯a ≤ 21 . As (A, µ ¯) is semi-finite, it follows that µ ¯c ≤

1 2

(322Eb).

(δδ ) I may therefore write α0 = inf{α : α ≥ 0, µ ¯[[u > α]] ≤ 1}. Now [[u > α0 ]] = supα>α0 [[u > α]], so µ ¯[[u > α0 ]] = supα>α0 µ ¯[[u > α]] ≤ 1. (²²) If α ≥ α0 then (u − α0 χ1)+ ≤ (α − α0 )χ[[u > α0 ]] + (u − αχ1)+ , so α0 +

R

(u − α0 χ1)+ ≤ α0 + (α − α0 )¯ µ[[u > α0 ]] +

R

(u − αχ1)+ ≤ α +

If 0 ≤ α < α0 then, for every β ∈ [0, α0 − α[, (u − α0 χ1)+ + β[[u > α + β]] ≤ (u − αχ1)+ , while µ ¯[[u > α + β]] > 1, so

R

(u − α0 χ1)+ + β + α ≤ α +

R

(u − αχ1)+ ;

R

(u − αχ1)+ .

369P


taking the supremum over β, α0 +

R Thus α0 + (u − α0 χ1)+ = τ3 (u).

R

(u − α0 χ1)+ ≤ α +

R

395

(u − αχ1)+ .

(ζζ ) If α0 = 0, take v = χ[[u > 0]]; then kvk∞,1 = µ ¯[[u > 0]] ≤ 1 and τ3 (u) =

R

u = ku × vk1 ≤ τ1 (u).

(ηη ) If α0 > 0, set γ = µ ¯[[u > α0 ]]. Take any β ∈ [0, α0 [. Then µ ¯([[u > β]] \ [[u > α0 ]]) > 1 − γ, so there is a b ⊆ [[u > β]] \ [[u > α0 ]] such that 1 − γ < µ ¯b < ∞. Set v = χ[[u > α0 ]] + 1−γ µ ¯ b χb. Then kvk∞,1 = 1 so τ1 (u) ≥

R

u×v ≥

R

(u − α0 χ1)+ + α0 γ + β

1−γ µ ¯b µ ¯b

= τ3 (u) − (1 − γ)(α0 − β).

As β is arbitrary, τ1 (u) ≥ τ3 (u) in this case also. Q Q (v) Thus τ1 (u) = τ2 (u) = τ3 (u) = τ4 (u) for every u ∈ L0 , and I may write kuk1,∞ for their common value; as the associate of k k∞,1 , k k1,∞ is an extended Fatou norm. As for the attainment of the infima, R the argument of (iv-²) above shows that, at least when 0 < kuk1,∞ < ∞, there is an α0 such that α0 + (|u| − α0 )+ = kuk1,∞ . This omits the cases kuk1,∞ ∈ {0, ∞}; but in either of these cases we can set α0 = 0 to see that the infimum is attained for trivial reasons. For the other infimum, observe that the argument of (ii) produces u0 , u00 such that u = u0 + u00 and ku0 k1 + ku00 k∞ ≤ τ4 (u). (c) This is now obvious from the definition of k k∞,1 and the characterization of k k1,∞ in of k k1 and k k∞ . (d) To see that S = S(A) is norm-dense in M 1,∞ , we need only note that S is dense in L∞ and S ∩ L1 is dense in L1 ; so that given v ∈ L1 , w ∈ L∞ and ² > 0 there are v 0 , w0 ∈ S such that k(v + w) − (v 0 + w0 )k1,∞ ≤ kv − v 0 k1 + kw − w0 k∞ ≤ ². As for M ∞,1 , if u ≥ 0 in M ∞,1 and r ∈ N, set vr = supk∈N 2−r kχ[[u > 2−r k]]; then each vr ∈ S fR = S(ARf ), ku − vr k∞ ≤ 2−r , and hvr ir∈N is a non-decreasing sequence with supremum u, so that limr→∞ vr = u and limr→∞ ku−vr k∞,1 = 0. Thus (S f )+ is dense in (M ∞,1 )+ . As usual, it follows that S f = (S f )+ −(S f )+ is dense in M ∞,1 = (M ∞,1 )+ − (M ∞,1 )+ . (e)(i) If p = 1 or p = ∞ this is immediate from the definition of k k∞,1 and the characterization of k k1,∞ in (b). So suppose henceforth that 1 < p < ∞. R (ii) If kuk∞,1 ≤ 1 then kukp ≤ 1. P P Because kuk∞ ≤ 1, |u|p ≤ |u|, so that |u|p ≤ kuk1 ≤ 1 and kukp ≤ 1. Q Q On considering scalar multiples of u, we see at once that kukp ≤ kuk∞,1 for every u ∈ L0 . (ii) Now set q = p/(p − 1). Then kukp = sup{ku × vk1 : kvkq ≤ 1} (369L) ≥ sup{ku × vk1 : kvk∞,1 ≤ 1} = kuk1,∞ because k k1,∞ is the associate of k k∞,1 . This completes the proof. 369P

In preparation for some ideas in §372, I go a little farther with M 1,0 , as defined in 366F.

Proposition Let (A, µ ¯) be a measure algebra. (a) M 1,0 = M 1,0 (A, µ ¯) is a norm-closed solid linear subspace of M 1,∞ = M 1,∞ (A, µ ¯). (b) The norm k k1,∞ is order-continuous on M 1,0 . (c) S(Af ) and L1 (A, µ ¯) are norm-dense and order-dense in M 1,0 . proof (a) Of course M 1,0 , being a solid linear subspace of L0 included in M 1,∞ , is a solid linear subspace of M 1,∞ . To see that it is norm-closed, take any point u of its closure. Then for any ² > 0 there is a v ∈ M 1,0

396

Function spaces

369P

such that ku − vk1,∞ ≤ ²; now (|u − v| − ²χ1)+ ∈ L1 , so [[|u − v| > 2²]] has finite measure; also [[|v| > ²]] has finite measure, so [[|u| > 3²]] ⊆ [[|u − v| > 2²]] ∪ [[|v| > ²]] (364Fa) has finite measure. As ² is arbitrary, u ∈ M 1,0 ; as u is arbitrary, M 1,0 is closed. (b) Suppose that A ⊆ M 1,0 is non-empty and downwards-directed and has infimum 0. Let ² > 0. Set B = {(u − ²χ1)+ : u ∈ A}. Then B ⊆ L1 (by 366Gc); B is non-empty and downwards-directed and has infimum 0. Because k k1 is order-continuous (365C), inf v∈B kvk1 = 0 and there is a u ∈ A such that k(u − ²χ1)+ k1 ≤ ², so that kuk1,∞ ≤ 2². As ² is arbitrary, inf u∈A kuk1,∞ = 0; as A is arbitrary, k k1,∞ is order-continuous on M 1,0 . (c) By 366Gb, S(Af ) is order-dense in M 1,0 . Because the norm of M 1,0 is order-continuous, S(Af ) is also norm-dense (354Ef). Now S(Af ) ⊆ L1 ⊆ M 1,0 , so L1 must also be norm-dense and order-dense. 369Q Corollary Let (A, µ ¯) be a localizable measure algebra. Set M 1,∞ = M 1,∞ (A, µ ¯), etc. 1,∞ × 1,0 × (a) (M ) and (M ) can both be identified with M ∞,1 . (b) (M ∞,1 )× can be identified with M 1,∞ ; M 1,∞ and M ∞,1 are perfect Riesz spaces. proof Everything is covered by 369O and 369K except the identification of (M 1,0 )× with M ∞,1 . For this I return to 369C. Of course M 1,0 is order-dense in L0 , because it includes L1 , or otherwise. Setting V = {v : v ∈ L0 , u × v ∈ L1 for every u ∈ M 1,0 }, 369C identifies V with (M 1,0 )× . Of course M ∞,1 ⊆ V just because M 1,0 ⊆ M 1,∞ . Also V ⊆ M ∞,1 . P P Let v ∈ V . (i) ?? If v ∈ / L∞ , then an = [[|v| > 4n ]] 6= 0 for everyP n. For each n, choose ∞ non-zero bn ⊆ an such thatR µ ¯bn < ∞ (using the fact that (A, µ ¯) is semi-finite). Set u = n=0 2−n (¯ µbn )−1 χbn ; 1 1,0 n 1 then u ∈ L ⊆ M , but |v| × u ≥ 2 for every n ∈ N, so |v| × u ∈ / L , which is impossible, because v ∈ V . X X ThusRv ∈ L∞ . (ii) ?? If v ∈ / L1 , then (again because (A, µ ¯) isR semi-finite, so that |v| = supa∈Af |v| × χa) supa∈Af a |v| = ∞. For each n ∈ N choose an ∈ Af such that an |v| ≥ 4n , and set u = supn∈N 2−n χan ∈ R M 1,0 ; then u×|v| ≥ 2n for each n, so again v ∈ / V.X X Thus v ∈ L1 . (iii) Putting these together, v ∈ M ∞,1 ; ∞,1 as v is arbitrary, V ⊆ M .Q Q So M ∞,1 = V can be identified with (M 1,0 )× . 369R The detailed formulae of 369O are of course special to the norms k k1 , k k∞ , but the general phenomenon is not. Theorem Let (A, µ ¯) be a localizable measure algebra, and τ1 , τ2 two extended Fatou norms on L0 = L0 (A) 0 with associates τ1 , τ20 . Then we have an extended Fatou norm τ defined by the formula τ (u) = min{τ1 (v) + τ2 (w) : v, w ∈ L0 , v + w = u} for every u ∈ L0 , and its associate τ 0 is given by the formula τ 0 (u) = max(τ10 (u), τ20 (u)) for every u ∈ L0 . Moreover, the corresponding function spaces are Lτ = Lτ1 + Lτ2 ,

0

0

0

Lτ = Lτ1 ∩ Lτ2 .

proof (a) For the moment, define τ by setting τ (u) = inf{τ1 (v) + τ2 (w) : v + w = u} for u ∈ L0 . It is easy to check that, for u, u0 ∈ L0 and α ∈ R, τ (u + u0 ) ≤ τ (u) + τ (u0 ),

τ (αu) = |α|τ (u),

τ (u) ≤ τ (u0 ) if |u| ≤ |u0 |.

(For the last, that in this case u = u0 × z where kzk∞ ≤ 1.) (b) Take any non-empty, upwards-directed set A ⊆ (L0 )+ , with supremum u0 . Suppose that γ = supu∈A τ (u) < ∞. For u ∈ A, n ∈ N set Cun = {v : v ∈ L0 , 0 ≤ v ≤ u0 , τ1 (v) + τ2 (u − v)+ ≤ γ + 2−n }.

369R


397

Then (i) every Cun is non-empty (because τ (u) ≤ γ); (ii) every Cun is convex (because if v1 , v2 ∈ Cun and α ∈ [0, 1] and v = αv1 + (1 − α)v2 , then (u − v)+ = (α(u − v1 ) + (1 − α)(u − v2 ))+ ≤ α(u − v1 )+ + (1 − α)(u − v2 )+ , so τ1 (v) + τ2 (u − v)+ ≤ ατ1 (v1 ) + (1 − α)τ1 (v2 ) + ατ2 (u − v1 )+ + (1 − α)τ2 (u − v2 )+ ≤ γ + 2−n ); (iii) if u, u0 ∈ A, m, n ∈ N and u ≤ u0 , m ≤ n then Cu0 n ⊆ Cum ; (iv) every Cun is closed for the topology of convergence in measure. P P?? Suppose otherwise. Then we can find a v in the closure of Cun for the topology of convergence in measure, but such that τ1 (v) + τ2 (u − v)+ > γ + 2−n . In this case τ1 (v) = sup{τ1 (v × χa) : a ∈ Af },

τ2 (u − v)+ = sup{τ2 ((u − v)+ × χa) : a ∈ Af },

so there is an a ∈ Af such that τ1 (v × χa) + τ2 ((u − v)+ × χa) > γ + 2−n . Now there is a sequence hvk ik∈N in Cun such that µ ¯(a ∩ [[|v − vk | ≥ 2−k ]]) ≤ 2−k for every k. Setting vk0 = inf i≥k vi ,

wk = inf i≥k (u − vi )+

we have τ1 (vk0 ) + τ2 (wk ) ≤ τ1 (vk ) + τ2 (u − vk )+ ≤ γ + 2−n for each k, and hvk0 ik∈N , hwk ik∈N are non-decreasing. So setting v ∗ = supk∈N v∧vk0 , w∗ = supk∈N (u−v)+ ∧wk , we get τ1 (v ∗ ) + τ2 (w∗ ) ≤ γ + 2−n . But v ∗ ≥ v × χa and w∗ ≥ (u − v)+ × χa, so τ1 (v × χa) + τ2 ((u − v)+ × χa) ≤ γ + 2−n , contrary to the choice of a. X XQ QT Applying 367V, we find that u∈A,n∈N Cun is non-empty. If v belongs to the intersection, then τ1 (v) + τ2 (u − v)+ ≤ γ for every u ∈ A; since {(u − v)+ : u ∈ A} is an upwards-directed set with supremum (u0 − v)+ , and τ2 is an extended Fatou norm, τ1 (v) + τ2 (u0 − v)+ ≤ γ. (c) This shows both that the infimum in the definition of τ (u) is always attained (since this is trivial if τ (u) = ∞, and otherwise we consider A = {|u|}), and also that τ (sup A) = supu∈A τ (u) whenever A ⊆ (L0 )+ is a non-empty upwards-directed set with a supremum. Thus τ satisfies conditions (i)-(iv) of 369F. Condition (vi) there is trivial, since (for instance) τ (v) ≤ τ1 (v) for every v. As for 369F(v), suppose that u > 0 in L0 . Take u1 such that 0 < u1 ≤ u and τ10 (u1 ) ≤ 1, u2 such that 0 < u2 ≤ u1 and τ20 (u2 ) ≤ 1. In this case, if u2 = v + w, we must have τ1 (v) + τ2 (w) ≥ kv × u1 k1 + kw × u2 k1 ≥ ku2 × u2 k1 ; so that τ (u) ≥ ku2 × u2 k1 > 0. Thus all the conditions of 369F are satisfied, and τ is an extended Fatou norm on L0 . (d) The calculation of τ 0 is now very easy. Since surely we have τ ≤ τi for both i, we must have τ 0 ≥ τi0 for both i. On the other hand, if u, z ∈ L0 , then there are v, w such that u = v +w and τ (u) = τ1 (v)+τ2 (w), so that

398

Function spaces

369R

ku × zk1 ≤ kv × zk1 + kw × zk1 ≤ τ1 (v)τ10 (z) + τ2 (w)τ20 (z) ≤ τ (u) max(τ10 (z), τ20 (z)); as u is arbitrary, τ 0 (z) ≤ max(τ10 (z), τ20 (z)). So τ 0 = max(τ10 , τ20 ), as claimed. (e) Finally, it is obvious that 0

0

0

Lτ = {z : τ 0 (z) < ∞} = {z : τ10 (z) < ∞, τ20 (z) < ∞} = Lτ1 ∩ Lτ2 , while the fact that the infimum in the definition of τ is always attained means that Lτ ⊆ Lτ1 + Lτ2 , so that we have equality here also. 369X Basic exercises > (a) Let A be a Dedekind σ-complete Boolean algebra. Show that the following are equiveridical: (i) there is a function µ ¯ such that (A, µ ¯) is a semi-finite measure algebra; (ii) (L∞ )× ∞ ∞ separates the points of L = L (A); (iii) for every non-zero a ∈ A there is a completely additive functional ν : A → R such that νa 6= 0; (iv) there is some order-dense Riesz subspace U of L0 = L0 (A) such that U × separates the points of U ; (v) for every order-dense Riesz subspace U of L0 there is an order-dense Riesz subspace V of U such that V × separates the points of V . (b) Let us say that a function φ : R → ]−∞, ∞] is convex if φ(αx + (1 − α)y) ≥ αφ(x) + (1 − α)φ(y) for all x, y ∈ I and α ∈ [0, 1], interpreting 0 · ∞ as 0, as usual. For any convex function φ : R → ]−∞, ∞] which is not always infinite, set φ∗ (y) = supx∈R xy − φ(x) for every y ∈ I. (i) Show that φ∗ : R → ]−∞, ∞] is convex and lower semi-continuous. (Hint: 233Xh.) (ii) Show that if φ is lower semi-continuous then φ = φ∗∗ . (Hint: It is easy to check that φ∗∗ ≤ φ. For the reverse inequality, set I = {x : φ(x) < ∞}, and consider x ∈ int I, x ∈ I \ int I and x ∈ / I separately; 233Ha is useful for the first.) > (c) For the purposes of this exercise and the next, say that a Young’s function is a non-negative non-constant lower semi-continuous convex function φ : [0, ∞[ → [0, ∞] such that φ(0) = 0 and φ(x) is finite for some x > 0. (Warning! the phrase ‘Young’s function’ has other meanings.) (i) Show that in this case φ is non-decreasing and continuous on the left and φ∗ , defined by saying that φ∗ (y) = supx≥0 xy − φ(x) for every y ≥ 0, is again a Young’s function. (ii) Show that φ∗∗ = φ. Say that φ and φ∗ are complementary. (iii) Compute φ∗ in the cases (α) φ(x) = x (β) φ(x) = max(0, x − 1) (γ) φ(x) = x2 (δ) φ(x) = xp where 1 (d) Let φ, ψ = φ∗ be complementary Young’s functions in the sense of 369Xc, and (A, µ ¯) a semi-finite measure algebra. Set R R ¯ ¯ B = {u : u ∈ L0 , φ(|u|) ≤ 1}, C = {v : v ∈ L0 , ψ(|v|) ≤ 1}. (For finite-valued φ, φ¯ : (L0 )+ → L0 is given by 364I. Devise an appropriate convention for the Rcase in which φ takes the value ∞.) (i) Show that B and C are order-closed solid convex sets, and that |u × v| ≤ 2 for all u ∈ B, v ∈ C. (Hint: for ‘order-closed’, use 364Xg(iv).) (ii) Show that there is a unique extended FatouR norm τφ on L0 for which B is the unit ball. (iii) Show that if u ∈ L0 \ B there is a v ∈ C such that |u × v| > 1. (Hint: start with the case in which u ∈ S(A)+ .) (iv) Show that τψ ≤ τφ0 ≤ 2τψ , where τψ is the extended Fatou norm corresponding to ψ and τφ0 is the associate of τφ , so that τψ and τφ0 can be interpreted as equivalent norms on the same Banach space. (U and V are complementary Orlicz spaces; I will call τφ , τψ Orlicz norms.) (e) Let U be a Riesz space such that U × separates the points of U , and suppose that k k is a Fatou norm on U . (i) Show that there is a localizable measure algebra (A, µ ¯) with an extended Fatou norm τ on L0 (A) such that U can be identified, as normed Riesz space, with an order-dense Riesz subspace of Lτ . (ii) Hence, or otherwise, show that kuk = supf ∈U × , kf k≤1 |f (u)| for every u ∈ U . (iii) Show that if U is Dedekind complete and has the Levi property, then U becomes identified with Lτ itself, and in particular is a Banach lattice (cf. 354Xn). (f ) Let (A, µ ¯) be a semi-finite measure algebra, and τ an extended Fatou norm on L0 (A). Show that τ the norm of L is order-continuous iff the norm topology of Lτ agrees with the topology of convergence in measure on any order-bounded subset of Lτ .

369Xr


399

(g) Let (A, µ ¯) be a σ-finite measure algebra of countable Maharam type, and τ an extended Fatou norm on L0 (A) such that the norm of Lτ is order-continuous. Show that Lτ is separable in its norm topology. R (h) Let (A, µ ¯) be an atomless semi-finite measure algebra. Show that kuk1,∞ = max{ a |u| : a ∈ A, µ ¯a ≤ 1} for every u ∈ L0 (A). (Hint: take a ⊇ [[|u| > α0 ]] in part (b-iv) of the proof of 369O.) (i) Let (A, µ ¯) be any semi-finite measure algebra. Show that if τφ is any Orlicz norm (369Xd), then there is a γ > 0 such that kuk1,∞ ≤ γτφ (u) ≤ γ 2 kuk∞,1 for every u ∈ L0 (A), so that Mµ¯∞,1 ⊆ Lτφ ⊆ Mµ¯1,∞ . (j) Let (A, µ ¯) be a semi-finite measure algebra. Show that the subspaces Mµ¯1,∞ , Mµ¯∞,1 of L0 (A) can be expressed as a complementary pair of Orlicz spaces, and that the norm k k∞,1 can be represented as an Orlicz norm, but k k1,∞ cannot. > (k) Let (A, µ ¯) be a measure algebra and U a Banach space. (i) Suppose that ν : A → U is an additive function such that kνak ≤ min(1, µ ¯a) for every a ∈ A. Show that there is a unique bounded linear operator 1,∞ T : Mµ¯ → U such that T (χa) = νa for every a ∈ A. (ii) Suppose that ν : Af → U is an additive function such that kνak ≤ max(1, µ ¯a) for every a ∈ Af . Show that there is a unique bounded linear operator ∞,1 T : Mµ¯ → U such that T (χa) = νa for every a ∈ Af . (l) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras, and π : Af → Bf a measure-preserving ring homomorphism, as in 366H, with associated maps T : Mµ¯0 → Mν¯0 , P : Mν¯1,0 → Mµ¯1,0 . Show that kT uk∞,1 = kuk∞,1 for every u ∈ Mµ¯∞,1 , kP vk∞,1 ≤ kvk∞,1 for every v ∈ Mν¯∞,1 . (m) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a measure-preserving Boolean homomorphism. (i) Show that there is a unique Riesz homomorphism T : Mµ¯1,∞ → Mν¯1,∞ such that T (χa) = χ(πa) for every a ∈ A and kT uk1,∞ = kuk1,∞ for every u ∈ Mµ¯1,∞ . (ii) Now suppose that (A, µ ¯) is localizable and π is order-continuous. Show that there is a unique positive linear operator P : Mν¯1,∞ → Mµ¯1,∞ R R such that a P v = πa v for every a ∈ Af , v ∈ Mν¯1,∞ , and that kP vk∞ ≤ kvk∞ for every v ∈ L∞ (B), kP vk1,∞ ≤ kvk1,∞ for every v ∈ Mν¯1,∞ . (Compare 365P.) (n) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras, and φ : [0, ∞[ → [0, ∞] a Young’s function, as in 369Xd; write τφ for the corresponding Orlicz norm on either L0 (A) or L0 (B). Let π : A → B be a measure-preserving Boolean homomorphism, with associated map T : Mµ¯1,∞ → Mν¯1,∞ , as in 369Xm. (i) Show that τφ (T u) = τφ (u) for every u ∈ Mµ¯1,∞ . (ii) Show that if (A, µ ¯) is localizable, π is order-continuous and P : Mν¯1,∞ → Mµ¯1,∞ is the map of 369Xm(ii), then τφ (P v) ≤ τφ (v) for every v ∈ Mν¯1,∞ . (Hint: 365R.) > (o) Let (A, µ ¯) be any semi-finite measure algebra and τ1 , τ2 two extended Fatou norms on L0 (A). Show that u 7→ max(τ1 (u), τ2 (u)) is an extended Fatou norm. b µ (p) Let (A, µ ¯) be a semi-finite measure algebra, and (A, ˜) its localization (322P). Show that the Dedekind 1,∞ 1,∞ b completion of M (A, µ ¯) can be identified with M (A, µ ˜). (q) Let (A, µ ¯) be a localizable measure algebra. (i) Show that if B is any closed subalgebra of A such that sup{b : b ∈ B, µ ¯b < ∞} = 1 in A, we have an order-continuous positive linear operator PB : Mµ¯1,∞ → Mµ¯1,∞ ¹B R R such that b PB u = b u whenever u ∈ Mµ¯1,∞ , b ∈ B and µ ¯b < ∞. (ii) Show that if hBn in∈N is a nondecreasing sequence of closed subalgebras of A such that sup{b : b ∈ B0 , µ ¯b < ∞} = 1 in A, and B is the S closure of n∈N Bn , then hPBn uin∈N is order*-convergent to PB u for every u ∈ Mµ¯1,∞ . (Cf. 367K.) (r) Let φ1 and φ2 be Young’s functions (369Xc) and (A, µ ¯) a semi-finite measure algebra. Set φ(x) = max(φ1 (x), φ2 (x)) for x ∈ [0, ∞[. (i) Show that φ is a Young’s function. (ii) Writing τφ1 τφ2 , τφ for the corresponding extended Fatou norms on L0 (A) (369Xd), show that τφ ≥ max(τφ1 , τφ2 ) ≥ 21 τφ , so that τ ∗ τ ∗ Lτφ = Lτφ1 ∩ Lτφ2 and Lτφ∗ = L φ1 + L φ2 , writing φ∗ for the Young’s function complementary to φ. (iii) Repeat with ψ = φ1 + φ2 in place of φ.

400

Function spaces

369Y

369Y Further exercises (a) Let (A, µ ¯) be a localizable measure algebra and A ⊆ L0 = L0 (A) a countable set. Show that the solid linear subspace U of L0 generated by A is a perfect Riesz space. (Hint: 0 + 0 + Rreduce to thencase ninR which U is order-dense. If A = {un : n ∈ N}, w ∈ (L ) \ U find vn 0∈ (L ) such that vn × u ≥ 2 ≥ 4 vn × |ui | for every i ≤ n. Show that v = supn∈N vn is defined in L and corresponds to a member of U × .) (b) Let U be a Banach lattice and suppose that p ∈ [1, ∞[ is such that ku + vkp = kukp + kvkp whenever u, v ∈ U and |u| ∧ |v| = 0. Show that U is isomorphic, as Banach lattice, to Lpµ¯ for some localizable measure algebra (A, µ ¯). (Hint: start by using 354Yd to show that the norm of U is order-continuous, as in 354Yj.) (c) Let φ : [0, ∞[ → [0, ∞[ be a strictly increasing Young’s function (369Xc) such that φ(0) = 0 and supt>0 φ(2t)/φ(t) is finite. Show that the associated Orlicz norms τφ (369Xd) are always order-continuous on their function spaces. (d) Let φ : [0, ∞[ → [0, ∞] be a Young’s function, and suppose that the corresponding Orlicz norm on L0 (AL ), where (AL , µ ¯L ) is the measure algebra of Lebesgue measure on R, is order-continuous on its function space Lτφ . Show that there is an M ≥ 0 such that φ(2t) ≤ M φ(t) for every t ≥ 0. (e) Let (A, µ ¯) be a semi-finite measure algebra and φ : [0, ∞[ → [0, ∞[ a Young’s function such that the Orlicz norm τφ is order-continuous on Lτφ . Show that if F is a filter on Lτφ , then F → u ∈ Lτφ for the norm τφ iff (i) F → u for the topology of convergence in measure (ii) lim supv→F τφ (v) ≤ τφ (u). (Compare 245Xk.) (f ) Give an example of an extended Fatou norm τ on L0 (AL ), where (AL , µ ¯L ) is the measure algebra of Lebesgue measure on [0, 1], such that (i) τ gives rise to an order-continuous norm on its function space Lτ (ii) there is a sequence hun in∈N in Lτ , converging in measure to u ∈ Lτ , such that limn→∞ τ (un ) = τ (u) but hun in∈N does not converge to u for the norm on Lτ . (g) Let (A, µ ¯) be a semi-finite measure algebra, and τ an Orlicz norm on L0 (A). Show that Lτ has the Levi property, whether or not A is Dedekind complete. (h) Let (A, µ ¯) be any measure algebra. Show that (Mµ¯1,0 )× can be identified with Mµ¯∞,1 . (Hint: show that neither M 1,0 nor M ∞,1 is changed by moving first to the semi-finite version of (A, µ ¯), as described in 322Xa, and then to its localization.) (i) Give an example to show that the result of 369R may fail if (A, µ ¯) is only semi-finite, not localizable. 369 Notes and comments The representation theorems 369A-369D give a very concrete form to the notion of ‘perfect’ Riesz space: it is just one which can be expressed as a subspace of L0 (A), for some localizable measure algebra (A, µ ¯), in such a way that it is its own second dual, where the duality here is between subspaces of L0 , taking U 0 = {v : u × v ∈ L1 for every u ∈ U }. (I see that in this expression I ought somewhere to mention that both U and U 0 are assumed to be order-dense in L0 .) Indeed I believe that the original perfect spaces were the ‘vollkommene Räume’ of G.Köthe, which were subspaces of RN , corresponding to the measure algebra PN with counting measure, so that U 0 or U × was {v : u × v ∈ `1 for every u ∈ U }. I have presented Kakutani’s theorem on the representation of L-spaces as a corollary of 369A and 369C. As usual in such things, this is a reversal of the historical relationship; Kakutani’s theorem was one of the results which led to the general theory. If we take the trouble to re-work the argument of 369A in this context, we find that the L-space condition ‘ku + vk = kuk + kvk whenever u, v ≥ 0’ can be relaxed to ‘ku + vk = kuk + kvk whenever u ∧ v = 0’ (369Yb). The complete list of localizable measure algebras provided by Maharam’s theorem (332B, 332J) now gives us a complete list of L-spaces. Just as perfect Riesz spaces come in dual pairs, so do some of the most important Banach lattices: those with Fatou norms and the Levi property for which the order-continuous dual separates the points. (Note that the dual of any space with a Riesz norm has these properties; see 356Da.) I leave the details of representing such spaces to you (369Xe). The machinery of 369F-369K gives a solid basis for studying such pairs.

369 Notes


401

Among the extended Fatou norms of 369F the Orlicz norms (369Xd, 369Yc-369Ye) form a significant subfamily. Because they are defined in a way which is to some extent independent of the measure algebra involved, these spaces have some of the same properties as Lp spaces in relation to measure-preserving homomorphisms (369Xm-369Xn). In §§373-374 I will elaborate on these ideas. Among the Orlicz spaces, we have a largest and a smallest; these are just M 1,∞ = L1 + L∞ and M ∞,1 = L1 ∩ L∞ (369N-369O, 369Xi, 369Xj). Of course these two are particularly important. There is an interesting phenomenon here. It is easy to see that k k∞,1 = max(k k1 , k k∞ ) is an extended Fatou norm and that the corresponding Banach lattice is L1 ∩L∞ ; and that the same ideas work for any pair of extended Fatou norms (369Xo). To check that the dual of L1 ∩ L∞ is precisely the linear sum L∞ + L1 a little more is needed, and the generalization of this fact to other extended Fatou norms (369Q) seems to go quite deep. In view of our ordinary expectation that properties of these normed function spaces should be reflected in perfect Riesz spaces in general, I mention that I believe I have found an example, dependent on the continuum hypothesis, of two perfect Riesz subspaces U , V of RN such that their linear sum U + V is not perfect.

402

Linear operators between function spaces

Chapter 37 Linear operators between function spaces As everywhere in functional analysis, the function spaces of measure theory cannot be properly understood without investigating linear operators between them. In this chapter I have collected a number of results which rely on, or illuminate, the measure-theoretic aspects of the theory. §371 is devoted to a fundamental property of linear operators on L-spaces, if considered abstractly, that is, of L1 -spaces, if considered in the language of Chapter 36, and to an introduction to the class T of operators which are norm-decreasing for both k k1 and k k∞ . This makes it possible to prove a version of (Birkhoff’s) Ergodic Theorem for operators which need not be positive (372D). In §372 I give various forms of this theorem, for linear operators between function spaces, for measure-preserving Boolean homomorphisms between measure algebras, and for inverse-measurepreserving functions between measure spaces, with an excursion into the theory of continued fractions. In §373 I make a fuller analysis of the class T , with a complete characterization of those u, v such that v = T u for some T ∈ T . Using this we can describe ‘rearrangement-invariant’ function spaces and extended Fatou norms (§374). Returning to ideas left on one side in §§364 and 368, I investigate positive linear operators defined on L0 spaces (§375). In the final section of the chapter (§376), I look at operators which can be defined in of kernels on product spaces.

371 The Chacon-Krengel theorem The first topic I wish to treat is a remarkable property of L-spaces: if U and V are L-spaces, then every continuous linear operator T : U → V is order-bounded, and k|T |k = kT k (371D). This generalizes in various ways to other V (371B, 371C). I apply the result to a special type of operator between M 1,0 spaces which will be conspicuous in the next section (371F-371H). 371A Lemma Let U be an L-space, V a Banach lattice and T : U → V a bounded linear operator. Take u ≥ 0 in U and set Pn Pn B = { i=0 |T ui | : u0 , . . . , un ∈ U + , i=0 ui = u} ⊆ V + . Then B is upwards-directed and supv∈B kvk ≤ kT kkuk.

Pm proof (a) Suppose that v, v 0 ∈ B. Then we have u0 , . . . , um , u00 , . . . , u0n ∈ U + such that i=0 ui = P P Pn n m 0 0 0 |T u |. Now there are v ≥ 0 in U , for i ≤ m and j ≤ n, such |T u | and v = u = u, v = ij i j i=0 j=0 j P Pm Pn Pj=0 m n 0 that ui = j=0 vij for i ≤ m and uj = i=0 vij for j ≤ n (352Fd). We have u = i=0 j=0 vij , so that Pm Pn v 00 = i=0 j=0 |T vij | ∈ B. But Pm Pm Pn Pm Pm v = i=0 |T ui | = i=0 |T ( j=0 vij )| ≤ i=0 j=0 |T vij | = v 00 , and similarly v 0 ≤ v 00 . As v and v 0 are arbitrary, B is upwards-directed. Pm Pm (b) The other part is easy. If v ∈ B is expressed as i=0 |T ui | where ui ≥ 0, i=0 ui = u then Pm Pm kvk ≤ i=0 kT ui k ≤ kT k i=0 kui k = kT kkuk because U is an L-space. 371B Theorem Let U be an L-space and V a Dedekind complete Banach lattice U with a Fatou norm. Then the Riesz space L∼ (U ; V ) = L× (U ; V ) is a closed linear subspace of the Banach space B(U ; V ) and is in itself a Banach lattice with a Fatou norm. proof (a) I start by noting that L∼ (U ; V ) = L× (U ; V ) ⊆ B(U ; V ) just because V has a Riesz norm and U is a Banach lattice with an order-continuous norm (355C, 355Kb). (b) The first new step is to check that k|T |k ≤ kT k for any T ∈ L∼ (U ; V ). P P Start with any u ∈ U + . Set Pn Pn B = { i=0 |T ui | : u0 , . . . , un ∈ U + , i=0 ui = u} ⊆ V + ,

371C

The Chacon-Krengel theorem

403

Pn Pn as in 371A. If u0 , . . . , un ≥ 0 are such that i=0 ui = u, then |T ui | ≤ |T |ui for each i, so that i=0 |T ui | ≤ P n i=0 |T |ui = |T |u; thus B is bounded above by |T |u and sup B ≤ |T |u. On the other hand, if |v| ≤ u in U , then v + + v − + (u − |v|) = u, so |T v + | + |T v − | + |T (u − |v|)| ∈ B and |T v| = |T v + + T v − | ≤ |T v + | + |T v − | ≤ sup B. As v is arbitrary, |T |u ≤ sup B and |T |u = sup B. Consequently k|T |uk ≤ k sup Bk = supw∈B kwk ≤ kT kkuk because V has a Fatou norm and B is upwards-directed. For general u ∈ U , k|T |uk ≤ k|T ||u|k ≤ kT kk|u|k = kT kkuk. This shows that k|T |k ≤ kT k. Q Q (c) Now if |S| ≤ |T | in L∼ (U ; V ), and u ∈ U , we must have kSuk ≤ k|S||u|k ≤ k|T ||u|k ≤ k|T |kk|u|k ≤ kT kkuk; as u is arbitrary, kSk ≤ kT k. This shows that the norm of L∼ (U ; V ), inherited from B(U ; V ), is a Riesz norm. (d) Suppose next that T ∈ B(U ; V ) belongs to the norm-closure of L∼ (U ; V ). For each n ∈ N choose Tn ∈ L∼ (U ; V ) such that kT − Tn k ≤ 2−n . Set Sn = |Tn+1 − Tn | ∈ L∼ (U ; V ) for each n. Then for each n, so S =

P∞ n=0

kSn k = kTn+1 − Tn k ≤ 3 · 2−n−1 Sn is defined in the Banach space B(U ; V ). But if u ∈ U + , we surely have P∞ Su = n=0 Sn u ≥ 0

in V . Moreover, if u ∈ U + and |v| ≤ u, then for any n ∈ N Pn Pn |Tn+1 v − T0 v| = | i=0 (Ti+1 − Ti )v| ≤ i=0 Si u ≤ Su, and T0 v − Su ≤ Tn+1 v ≤ T0 v + Su; letting n → ∞, we see that −|T0 |u − Su ≤ T0 v − Su ≤ T v ≤ T0 v + Su ≤ |T0 |u + Su. So |T v| ≤ |T0 |u + Su whenever |v| ≤ u. As u is arbitrary, T ∈ L∼ (U ; V ). This shows that L∼ (U ; V ) is closed in B(U ; V ) and is therefore a Banach space in its own right; putting this together with (b), we see that it is a Banach lattice. (e) Finally, the norm of L∼ (U ; V ) is a Fatou norm. P P Let A ⊆ L∼ (U ; V )+ be a non-empty, upwards∼ directed set with supremum T0 ∈ L (U ; V ). For any u ∈ U , kT0 uk = k|T0 u|k ≤ kT0 |u|k = k supT ∈A T |u|k by 355Ed. But {T |u| : T ∈ A} is upwards-directed and the norm of V is a Fatou norm, so kT0 uk ≤ supT ∈A kT |u|k ≤ supT ∈A kT kkuk. As u is arbitrary, kT0 k ≤ supT ∈A kT k. As A is arbitrary, the norm of L∼ (U ; V ) is Fatou. Q Q 371C Theorem Let U be an L-space and V a Dedekind complete Banach lattice with a Fatou norm and the Levi property. Then B(U ; V ) = L∼ (U ; V ) = L× (U ; V ) is a Dedekind complete Banach lattice with a Fatou norm and the Levi property. In particular, |T | is defined and k|T |k = kT k for every T ∈ B(U ; V ). proof (a) Let T : U → V be any bounded linear operator. Then T ∈ L∼ (U ; V ). P P Take any u ≥ 0 in U . Set Pn Pn B = { i=0 |T ui | : u0 , . . . , un ∈ U + , i=0 ui = u} ⊆ V + as in 371A. Then 371A tells us that B is upwards-directed and norm-bounded. Because V has the Levi property, B is bounded above. But just as in part (b) of the proof of 371B, any upper bound of B is also an upper bound of {T v : |v| ≤ u}. As u is arbitrary, T ∈ L∼ (U ; V ). Q Q

404


371C

(b) Accordingly L∼ (U ; V ) = B(U ; V ). By 371B, this is a Banach lattice with a Fatou norm, and equal to L× (U ; V ). To see that it also has the Levi property, let A ⊆ L∼ (U ; V ) be any non-empty norm-bounded upwards-directed set. For u ∈ U + , {T u : T ∈ A} is non-empty, norm-bounded and upwards-directed in V , so is bounded above in V . By 355Ed, A is bounded above in L∼ (U ; V ). 371D Corollary Let U and V be L-spaces. Then L∼ (U ; V ) = L× (U ; V ) = B(U ; V ) is a Dedekind complete Banach lattice with a Fatou norm and the Levi property. 371E Remarks Note that both these theorems show that L∼ (U ; V ) is a Banach lattice with properties similar to those of V whenever U is an L-space. They can therefore be applied repeatedly, to give facts about L∼ (U1 ; L∼ (U2 ; V )) where U1 , U2 are L-spaces and V is a Banach lattice, for instance. I hope that this formula will recall some of those in the theory of bilinear maps and tensor products (see 253Xa-253Xb). 371F The class T (0) For the sake of applications in the next section, I introduce now a class of operators of great intrinsic interest. Definition Let (A, µ ¯), (B, ν¯) be measure algebras. Recall that M 1,0 (A, µ ¯) is the space of those u ∈ (0) 1 ∞ L (A, µ ¯) + L (A) such that µ ¯[[|u| > α]] < ∞ for every α > 0 (366F-366G, 369P). Write T (0) = Tµ¯,¯ν for the set of all linear operators T : M 1,0 (A, µ ¯) → M 1,0 (B, ν¯) such that T u ∈ L1 (B, ν¯) and kT uk1 ≤ kuk1 for 1 ∞ every u ∈ L (A, µ ¯), T u ∈ L (B) and kT uk∞ ≤ kuk∞ for every u ∈ L∞ (A) ∩ M 1,0 (A, µ ¯). 371G Proposition Let (A, µ ¯) and (B, ν¯) be measure algebras. (0) (0) ¯); M 1,0 (B, ν¯)). If T0 : L1 (A, µ ¯) → L1 (B, ν¯) is (a) T = Tµ¯,¯ν is a convex set in the unit ball of B(M 1,0 (A, µ ∞ a linear operator of norm at most 1, and T0 u ∈ L (B) and kT0 uk∞ ≤ kuk∞ for every u ∈ L1 (A, µ ¯)∩L∞ (A), (0) then T0 has a unique extension to a member of T . (b) If T ∈ T (0) then T is order-bounded and |T |, taken in L∼ (M 1,0 (A, µ ¯); M 1,0 (B, ν¯)) = L× (M 1,0 (A, µ ¯); M 1,0 (B, ν¯)), also belongs to T (0) . (c) If T ∈ T (0) then kT uk1,∞ ≤ kuk1,∞ for every u ∈ M 1,0 (A, µ ¯). (d) If T ∈ T (0) , p ∈ [1, ∞[ and w ∈ Lp (A, µ ¯) then T w ∈ Lp (B, ν¯) and kT wkp ≤ kwkp . (0) (0) ¯ is another measure algebra then ST ∈ T (0) (e) If (C, λ) ¯ ,¯ ν and every S ∈ Tν ¯. ¯ for every T ∈ Tµ ¯ ,λ µ ¯ ,λ proof I write Mµ¯1,0 , Lpν¯ for Mµ¯1,0 , Lp (B, ν¯), etc. (a)(i) If T ∈ T (0) and u ∈ Mµ¯1,0 then there are v ∈ L1µ¯ , w ∈ L∞ µ ¯ such that u = v + w and kvk1 + kwk∞ = kuk1,∞ (369Ob); so that kT uk1,∞ ≤ kT vk1 + kT wk∞ ≤ kvk1 + kwk∞ ≤ kuk1,∞ . As u is arbitrary, T is in the unit ball of B(Mµ¯1,0 ; Mν¯1,0 ). ∞ (0) (ii) Because the unit balls of B(L1µ¯ ; L1ν¯ ) and B(L∞ . µ ¯ ; Lν ¯ ) are convex, so is T

(iii) Now suppose that T0 : L1µ¯ → L1ν¯ is a linear operator of norm at most 1 such that kT0 uk∞ ≤ kuk∞ 1 for every u ∈ L1µ¯ ∩ L∞ µ ¯ . By the argument of (i), T0 is a bounded operator for the k k1,∞ norms; since Lµ ¯ is 1,0 1,0 1,0 dense in Mµ¯ (369Pc), T0 has a unique extension to a bounded linear operator T : Mµ¯ → Mν¯ . Of course kT uk1 = kT0 uk1 ≤ kuk1 for every u ∈ L1µ¯ . 1,0 Now suppose that u ∈ L∞ ¯ ; set γ = kuk∞ . Let ² > 0, and set µ ¯ ∩ Mµ v = (u+ − ²χ1)+ − (u− − ²χ1)+ ; then |v| ≤ |u| and ku − vk∞ ≤ ² and v ∈ L1µ¯ ∩ L∞ µ ¯ . Accordingly kT u − T vk1,∞ ≤ ku − vk1,∞ ≤ ²,

kT vk∞ = kT0 vk∞ ≤ kvk∞ ≤ γ.

So if we set w = (|T u − T v| − ²χ1)+ ∈ L1ν¯ , kwk1 ≤ ²; while

371G

The Chacon-Krengel theorem

405

|T u| ≤ |T v| + w + ²χ1 ≤ (γ + ²)χ1 + w, so k(|T u| − (γ + ²)χ1)+ k1 ≤ kwk1 ≤ ². As ² is arbitrary, |T u| ≤ γχ1, that is, kT uk∞ ≤ kuk∞ . As u is arbitrary, T ∈ T (0) . (b) Because Mµ¯1,0 has an order-continuous norm (369Pb), L∼ (Mµ¯1,0 ; Mν¯1,0 ) = L× (Mµ¯1,0 ; Mν¯1,0 ) (355Kb). Take any T ∈ T (0) and consider T0 = T ¹L1µ¯ : L1µ¯ → L1ν¯ . This is an operator of norm at most 1. By 371D, T0 is order-bounded, and k|T0 |k ≤ 1, where |T0 | is taken in L∼ (L1µ¯ ; L1µ¯ ) = B(L1µ¯ ; L1ν¯ ). Now if u ∈ L1µ¯ ∩ L∞ ν ¯ , ||T0 |u| ≤ |T0 ||u| = sup|u0 |≤|u| |T0 u0 | ≤ kuk∞ χ1, so k|T0 |uk∞ ≤ kuk∞ . By (a), there is a unique S ∈ T (0) extending |T0 |. Now Su+ ≥ 0 for every u ∈ L1µ¯ , so Su+ ≥ 0 for every u ∈ Mµ¯1,0 (since the function u 7→ (Su+ )+ − Su+ : Mµ¯1,0 → Mν¯1,0 is continuous and zero on the dense set L1µ¯ ), that is, S is a positive operator; also S|u| ≥ |T u| for every u ∈ L1µ¯ , so Sv ≥ S|u| ≥ |T u| whenever u, v ∈ Mµ¯1,0 and |u| ≤ v. This means that T : Mµ¯1,0 → Mν¯1,0 is order-bounded. Because Mν¯1,0 is Dedekind complete (366Ga), |T | is defined in L∼ (Mµ¯1,0 ; Mµ¯1,0 ). If v ≥ 0 in L1µ¯ , then |T |v = sup|u|≤v T u = sup|u|≤v T0 u = |T0 |v = Sv. Thus |T | agrees with S on L1µ¯ . Because Mµ¯1,0 is a Banach lattice (or otherwise), |T | is a bounded operator, therefore continuous (2A4Fc), so |T | = S ∈ T (0) , which is what we needed to know. (c) We can express u as v + w where kvk1 + kwk∞ = kuk1,∞ ; now w = u − v ∈ Mµ¯1,0 , so we can speak of T w, and kT uk1,∞ = kT v + T wk1,∞ ≤ kT vk1 + kT wk∞ ≤ kvk1 + kwk∞ = kuk1,∞ , as required. (d) (This is a modification of 244M.) (i) Suppose that T , p, w are as described, and that in addition T is positive. The function t 7→ |t|p is convex (233Xc), so we can find families hβq iq∈Q , hγq iq∈Q of real numbers such that |t|p = supq∈Q βq +γq (t−q) for every t ∈ R (233Hb). Then |u|p = supq∈Q βq χ1 + γq (u − qχ1) for every u ∈ L0 . (The easiest way to check this is perhaps to think of L0 as a quotient of a space of functions, as in 364D; it is also a consequence of 364Xg(iii).) We know that |w|p ∈ L1µ¯ , so we may speak of T (|w|p ); while w ∈ Mµ¯1,0 (366Ga), so we may speak of T w. For any q ∈ Q, 0p ≥ βq −qγq , that is, qγq −βq ≥ 0, while γq w−|w|p ≤ (qγq −βq )χ1 and k(γq w−|w|p )+ k∞ ≤ qγq − βq . Now this means that T (γq w − |w|p ) ≤ T (γq w − |w|p )+ ≤ kT (γq w − |w|p )+ k∞ χ1 ≤ k(γq w − |w|p )+ k∞ χ1 ≤ (qγq − βq )χ1. Turning this round again, βq χ1 + γq (T w − qχ1) ≤ T (|w|p ). R R Taking the supremum over q, |T w|p ≤ T (|w|p ), so that |T w|p ≤ |w|p (because kT vk1 ≤ kvk1 for every v ∈ L1 ). Thus T w ∈ Lp and kT wkp ≤ kwkp . (ii) For a general T ∈ T (0) , we have |T | ∈ T (0) , by (b), and |T w| ≤ |T ||w|, so that kT wkp ≤ k|T ||w|kp ≤ kwkp , as required. (e) This is elementary, because kST uk1 ≤ kT uk1 ≤ kuk1 , whenever u ∈ L1µ¯ , v ∈ L∞ µ ¯ ∩

Mµ¯1,0 .

kST vk∞ ≤ kT uk∞ ≤ kuk∞

406

Linear operators between function spaces (0)

371H (0)

371H Remark In the context of 366H, Tπ ¹Mµ¯1,0 ∈ Tµ¯,¯ν , while Pπ ∈ Tν¯,¯µ . Thus 366H(a-iv), 366H(b-iii) are special cases of 371Gd. 371X Basic exercises >(a) Let U be an L-space, V a Banach lattice with an order-continuous norm and T : U → V a bounded linear operator. Let B be the unit ball of U . Show that |T |[B] ⊆ T [B]. (b) Let U and V be Banach spaces. (i) Show that the space Kw (U ; V ) of weakly compact linear operators from U to V (definition: 3A5Kb) is a closed linear subspace of B(U ; V ). (ii) Show that if U is an L-space and V is a Banach lattice with an order-continuous norm, then Kw (U ; V ) is a norm-closed Riesz subspace of L∼ (U ; V ). (c) Let (A, µ ¯) be a semi-finite measure algebra and set U = L1 (A, µ ¯). Show that L∼ (U ; U ) = B(U ; U ) is a Banach lattice with a Fatou norm and the Levi property. Show that its norm is order-continuous iff A is finite. (Hint: consider operators u 7→ u × χa, where a ∈ A.) > (d) Let U be a Banach lattice, and V a Dedekind complete M -space. Show that L∼ (U ; V ) = B(U ; V ) is a Banach lattice with a Fatou norm and the Levi property. (e) Let U and V be Riesz spaces, of which V is Dedekind complete, and let T ∈ L∼ (U ; V ). Define T ∈ L∼ (V ∼ ; U ∼ ) by writing T 0 (h) = hT for h ∈ V ∼ . (i) Show that |T |0 ≥ |T 0 | in L∼ (V ∼ ; U ∼ ). (ii) Show × + × + 0 that |T |0 h = |T 0 |h for every Pn h ∈ V . (Hint: show thatPifn u ∈ U , h ∈ (V ) then (|T |h)(u) and h(|T |u) are both equal to sup{ i=0 gi (T ui ) : |gi | ≤ h, ui ≥ 0, i=0 ui = u}.) 0

> (f ) Using 371D, but nothing about uniformly integrable sets beyond the definition (354P), show that if U and V are L-spaces, A ⊆ U is uniformly integrable in U , and T : U → V is a bounded linear operator, then T [A] is uniformly integrable in V . 371Y Further exercises (a) Let U and V be Banach spaces. (i) Show that the space K(U ; V ) of compact linear operators from U to V (definition: 3A5Ka) is a closed linear subspace of B(U ; V ). (ii) Show that if U is an L-space and V is a Banach lattice with an order-continuous norm, then K(U ; V ) is a norm-closed Riesz subspace of L∼ (U ; V ). (See Krengel 63.) (b) Let (A, µ ¯) be a measure algebra, U a Banach space, and T : L1 (A, µ ¯) → U a bounded linear operator. Show that T is a compact linear operator iff {

1 T (χa) µ ¯a

: a ∈ A, 0 < µ ¯a < ∞} is relatively compact in U .

(c) Let (A, µ ¯) be a probability algebra, and set L1 = L1 (A, µ ¯). Let han in∈N be a stochastically R R independent sequence of elements of A of measure 21 , and define T : L1 → RN by setting T u(n) = u − 2 an u for each n. Show that T ∈ B(L1 ; c 0 ) \ L∼ (L1 ; c 0 ), where c 0 is the Banach lattice of sequences converging to 0. (See 272Yd.) (d) Regarding T of 371Yc as a map from L1 to `∞ , show that |T 0 | 6= |T |0 in L∼ ((`∞ )∗ , L∞ (A)). (e) (i) In `2 define ei by setting ei (i) = 1, ei (j) = 0 if j 6= i. Show that if T ∈ L∼ (`2 ; `2 ) then (|T |ei |ej ) = |(T ei |ej )| for all i, j ∈ N. (ii) Show that for each n ∈ N there is an orthogonal (2n¶× 2n )-matrix µ An An 1 An such that every coefficient of An has modulus 2−n/2 . (Hint: An+1 = √ .) (iii) Show 2 −An An that there is a linear isometry S : `2 → `2 such that |(Sei |ej )| = 2−n/2 if 2n ≤ i, j < 2n+1 . (iv) Show that S∈ / L∼ (`2 ; `2 ). 371 Notes and comments The ‘Chacon-Krengel theorem’, properly speaking (Chacon & Krengel 64), is 371D in the case in which U = L1 (µ), V = L1 (ν); of course no new ideas are required in the generalizations here, which I have copied from Fremlin 74a. Anyone with a training in functional analysis will automatically seek to investigate properties of operators T : U → V in of properties of their adts T 0 : V ∗ → U ∗ , as in 371Xe and 371Yd. When U is

372B

The ergodic theorem

407

an L-space, then U ∗ is a Dedekind complete M -space, and it is easy to see that this forces T 0 to be orderbounded, for any Banach lattice V (371Xd). But since no important L-space is reflexive, this approach cannot reach 371B-371D without a new idea of some kind. It can however be adapted to the special case in 371Gb (Dunford & Schwartz 57, VIII.6.4). In fact the results of 371B-371C are characteristic of L-spaces (Fremlin 74b). To see that they fail in the simplest cases in which U is not an L-space and V is not an M -space, see 371Yc-371Ye.

372 The ergodic theorem I come now to one of the most remarkable topics in measure theory. I cannot do it justice in the space I have allowed for it here, but I can give the basic theorem (372D-372E) and a variety of the corollaries through which it is regularly used (372F-372K), together with brief notes on one of its most famous and characteristic applications (to continued fractions, 372M-372O) and on ‘ergodic’ and ‘mixing’ transformations (372P372R). In the first half of the section (down to 372G) I express the arguments in the abstract language of measure algebras and their associated function spaces, as developed in Chapter 36; the second half, from 372H onwards, contains translations of the results into the language of measure spaces and measurable functions, the more traditional, and more readily applicable, forms. 372A Lemma Let U be a reflexive Banach space and T : U → U a bounded linear operator of norm at most 1. Then V = {u + v − T u : u, v ∈ U, T v = v} is dense in U . proof Of course V is a linear subspace of U . ?? Suppose, if possible, that it is not dense. Then there is a non-zero h ∈ U ∗ such that h(v) = 0 for every v ∈ V (3A5Ad). Take u ∈ U such that h(u) 6= 0. Set 1 Pn i un = i=0 T u n+1

0

for each n ∈ N, taking T to be the identity operator; because kT i uk ≤ kT i kkuk ≤ kT ki kuk ≤ kuk for each i, kun k ≤ kuk for every n. Note also that T i+1 u − T i u ∈ V for every i, so that h(T i+1 u − T i u) = 0; accordingly h(T i u) = h(u) for every i, and h(un ) = u for every n. Let F be any non-principal ultrafilter on N. Because U is reflexive, v = limn→F un is defined in U for the weak topology on U (3A5Gc). Now T v = v. P P For each n ∈ N, P 1 1 n i+1 T un − un = u − T i u) = (T n+1 u − u) i=0 (T n+1

n+1

2 has norm at most n+1 kuk. So hT un − un in∈N → 0 for the norm topology U and therefore for the weak topology, and surely limn→F T un − un = 0. On the other hand (because T is continuous for the weak topology, 2A5If)

T v = limn→F T un = limn→F (T un − un ) + limn→F un = 0 + v = v, where all the limits are taken for the weak topology. Q Q But this means that v ∈ V , while h(v) = limn→F h(un ) = h(u) 6= 0, contradicting the assumption that h ∈ V ◦ . X X 372B Lemma Let (A, µ ¯) be a measure algebra, and T : L1 → L1 a positive linear operator of norm at 1 1 most 1, where L = L (A, µ ¯). Take any u ∈ L1 and m ∈ N, and set a = [[u > 0]] ∪ [[u + T u > 0]] ∪ [[u + T u + T 2 u > 0]] ∪ . . . ∪ [[u + T u + . . . + T m u > 0]].

408

Then


R a

372B

u ≥ 0.

proof Set u0 = u, u1 = u + T u, . . . , um = u + T u + . . . + T m u, v = supi≤m ui , so that a = [[v > 0]]. Consider u + T (v + ). We have T (v + ) ≥ T v ≥ T ui for every i ≤ m (because T is positive), so that u + T (v + ) ≥ u + T ui = ui+1 for i < m, and u + T (v + ) ≥ sup1≤i≤m ui . Also u + T (v + ) ≥ u because T (v + ) ≥ 0, so u + T (v + ) ≥ v. Accordingly

R

a

u≥

R

a

v−

R

a

T (v + ) =

R

v+ −

R

a

T (v + ) ≥ kv + k1 − kT v + k1 ≥ 0

because kT k ≤ 1. 372C Maximal Ergodic Theorem Let (A, µ ¯) be a measure algebra, and T : L1 → L1 a linear 1 1 operator, where L = L (A, µ ¯), P such that kT uk1 ≤ kuk1 for every u ∈ L1 and kT uk∞ ≤ kuk∞ for every n 1 1 ∞ u ∈ L ∩ L (A). Set An = n+1 i=0 T i for each n ∈ N. Then for any u ∈ L1 , u∗ = supn∈N An u is defined in L0 (A), and α¯ µ[[u∗ > α]] ≤ kuk1 for every α > 0. proof (a) To begin with, suppose that T is positive and that u ≥ 0 in L1 . Note that if v ∈ L1 ∩ L∞ , then kT i vk∞ ≤ kvk∞ for every i ∈ N, so kAn vk∞ ≤ kvk∞ for every n; in particular, An (χa) ≤ χ1 for every n and every a of finite measure. For m ∈ N and α > 0, set amα = [[supi≤m Ai u > α]]. Then α¯ µamα ≤ kuk1 . P P Set a = amα , w = u − αχa. Of course supi≤m Ai u belongs to L1 , so µ ¯a is finite 1 and w ∈ L . For any i ≤ m, Ai w = Ai u − αAi (χa) ≥ Ai u − αχ1, so [[Ai w > 0]] ⊇ [[Ai u > α]]. Accordingly a ⊆ b, where By 372B,

R b

b = supi≤m [[Ai w > 0]] = supi≤m [[w + T w + . . . + T i w > 0]]. w ≥ 0. But this means that α¯ µa = α

R

χa = b

R

u− b

R

w≤ b

R b

u ≤ kuk1 ,

as claimed. Q Q It follows that if we set cα = supn∈N anα , µ ¯cα ≤ α−1 kuk1 for every α > 0 and inf α>0 cα = 0. But ∗ this is exactly the criterion in 364Mb for u = supn∈N An u to be defined in L0 . And [[u∗ > α]] = cα , so α¯ µ[[u∗ > α]] ≤ kuk1 for every α > 0, as required. (b) Now consider the case of general T , u. In this case T is order-bounded and k|T |k ≤ 1, where |T | is the modulus of T in L∼ (L1 ; L1 ) = B(L1 ; L1 ) (371D). If w ∈ L1 ∩ L∞ , then ||T |w| ≤ |T ||w| = sup|w0 |≤|w| |T w0 | ≤ kwk∞ χ1,

Pn 1 i so k|T |wk∞ ≤ kwk∞ . Thus |T | also satisfies the conditions of the theorem. Setting Bn = n+1 i=0 |T | , ∼ 1 1 0 Bn ≥ An in L (L ; L ) and Bn |u| ≥ An u for every n. But by (a), v = supn∈N Bn |u| is defined in L and α¯ µ[[v > α]] ≤ k|u|k1 = kuk1 for every α > 0. Consequently u∗ = supn∈N An u is defined in L0 and u∗ ≤ v, so that α¯ µ[[u∗ > α]] ≤ kuk1 for every α > 0. 372D We are now ready for a very general form of the Ergodic Theorem. I express it in of the space M 1,0 from 366F and the class T (0) of operators from 371F. If these formulae are unfamiliar, you may like to glance at the statement of 372E before looking them up. The Ergodic Theorem: first form Let (A, µ ¯) be a measure algebra, and set M 1,0 = M 1,0 (A, µ ¯), T (0) = P (0) n 1 i 1,0 Tµ¯,¯µ ⊆ B(M 1,0 ; M 1,0 ) as in 371F-371G. Take any T ∈ T (0) , and set An = n+1 → M 1,0 for i=0 T : M 1,0 every n. Then for any u ∈ M , hAn uin∈N is order*-convergent (definition: 367A) and k k1,∞ -convergent to a member P u of M 1,0 . The operator P : M 1,0 → M 1,0 is a projection onto the linear subspace {u : u ∈ M 1,0 , T u = u}, and P ∈ T (0) . proof (a) It will be convenient to start with some elementary remarks. First, every An belongs to T (0) , by 371Ge and 371Ga. Next, hAn uin∈N is order-bounded in L0 = L0 (A) for any u ∈ M 1,0 ; this is because if

372D

The ergodic theorem

409

u = v + w, where v ∈ L1 = L1 (A, µ ¯) and w ∈ L∞ = L∞ (A), then hAn vin∈N and hAn (−v)in∈N are bounded above, by 372C, while hAn win∈N is norm- and order-bounded in L∞ . Accordingly I can uninhibitedly speak of P ∗ (u) = inf n∈N supi≥n Ai u and P∗ (u) = supn∈N inf i≥n Ai u for any u ∈ M 1,0 , these both being defined in L0 . (b) Write V1 for the set of those u ∈ M 1,0 such that hAn uin∈N is order*-convergent in L0 ; that is, P (u) = P∗ (u) (367Be). It is easy to see that V is a linear subspace of M 1,0 (use 367Ca and 367Cd). Also it is closed for k k1,∞ . Pn 1 i P P We know that |T |, taken in L∼ (M 1,0 ; M 1,0 ), belongs to T (0) (371Gb); set Bn = n+1 i=0 |T | for each i. Suppose that u0 ∈ V 1 . Then for any ² > 0 there is a u ∈ V1 such that ku0 − uk1,∞ ≤ ²2 . Write P u = P ∗ (u) = P∗ (u) for the order*-limit of hAn uin∈N . Express u0 − u as v + w where v ∈ L1 , w ∈ L∞ and kvk1 + kwk∞ ≤ 2²2 . Set v ∗ = supn∈N Bn |v|. Then µ ¯[[v ∗ > ²]] ≤ 2², by 372C. Next, if w∗ = supn∈N Bn |w|, we surely have ∗ 2 w ≤ 2² χ1. Now ∗

|An u0 − An u| = |An v + An w| ≤ Bn |v| + Bn |w| ≤ v ∗ + w∗ for every n ∈ N, that is, An u − v ∗ − w∗ ≤ An u0 ≤ An u + v ∗ + w∗ for every n. Because hAn uin∈N order*-converges to P u, P u − v ∗ − w∗ ≤ P∗ (u0 ) ≤ P ∗ (u0 ) ≤ P u + v ∗ + w∗ , and P ∗ (u0 ) − P∗ (u0 ) ≤ 2(v ∗ + w∗ ). On the other hand, µ ¯[[2(v ∗ + w∗ ) > 2² + 4²2 ]] ≤ µ ¯[[v ∗ > ²]] + µ ¯[[w∗ > 2²2 ]] = µ ¯[[v ∗ > ²]] ≤ 2² (using 364Fa for the first inequality). So µ ¯[[P ∗ (u0 ) − P∗ (u0 ) > 2²(1 + 2²)]] ≤ 2². Since ² is arbitrary, hAn u0 in∈N order*-converges to P ∗ (u0 ) = P∗ (u0 ), and u0 ∈ V1 . As u0 is arbitrary, V1 is closed. Q Q (c) Similarly, the set V2 of those u ∈ M 1,0 for which hAn uin∈N is norm-convergent is a linear subspace of M 1,0 , and it also is closed. P P This is a standard argument. If u0 ∈ V 2 and ² > 0, there is a u ∈ V2 such that ku0 − uk1,∞ ≤ ². There is an n ∈ N such that kAi u − Aj uk1,∞ ≤ ² for all i, j ≥ n, and now kAi u0 − Aj u0 k1,∞ ≤ 3² for all i, j ≥ n, because every Ai has norm at most 1 in B(M 1,0 ; M 1,0 ) (371Gc). As ² is arbitrary, hAi u0 in∈N is Cauchy; because M 1,0 is complete, it is convergent, and u0 ∈ V2 . As u0 is arbitrary, V2 is closed. Q Q (d) Now let V be {u+v−T u : u ∈ M 1,0 ∩L∞ , v ∈ M 1,0 , T v = v}. Then V ⊆ V1 ∩V2 . P P If u ∈ M 1,0 ∩L∞ , then for any n ∈ N An (u − T u) =

1 n+1 (u

− T n+1 u) → 0

for k k∞ , and therefore is both order*-convergent and convergent for k k1,∞ ; so u − T u ∈ V1 ∩ V2 . On the other hand, if T v = v, then of course An v = v for every n, so again v ∈ V1 ∩ V2 . Q Q (e) Consequently L2 = L2 (A, µ ¯ ) ⊆ V1 ∩ V2 . P P L2 ∩ V1 ∩ V2 is a linear subspace; but also it is closed for the 2 norm topology of L , because the identity map from L2 to M 1,0 is continuous (369Oe). We know also that T ¹L2 is an operator of norm at most 1 from L2 to itself (371Gd). Consequently W = {u + v − T u : u, v ∈ L2 , T v = v} is dense in L2 (372A). On the other hand, given u ∈ L2 and ² > 0, there is a u0 ∈ L2 ∩ L∞ such that ku−u0 k2 ≤ ² (take u0 = (u∧γχ1)∨(−γχ1) for any γ large enough), and now k(u−T u)−(u0 −T u0 )k2 ≤ 2². Thus W 0 = {u0 + v − T u0 : u0 ∈ L2 ∩ L∞ , v ∈ L2 , T v = v} is dense in L2 . But W 0 ⊆ V1 ∩ V2 , by (d) above. Thus L2 ∩ V1 ∩ V2 is dense in L2 , and is therefore the whole of L2 . Q Q (f ) L2 ⊇ S(Af ) is dense in M 1,0 , by 369Pc, so V1 = V2 = M 1,0 . This shows that hAn uin∈N is normconvergent and order*-convergent for every u ∈ M 1,0 . By 367D, the limits are the same; by 367F, hAn uin∈N is order*-convergent when regarded as a sequence in M 1,0 . Write P u for the common value of the limits.

410


372D

(g) Of course we now have kP uk∞ ≤ supn∈N kAn uk∞ ≤ kuk∞ for every u ∈ L∞ ∩ M 1,0 , while kP uk1 ≤ lim inf kAn uk1 ≤ kuk1 1

for every u ∈ L , by Fatou’s Lemma. So P ∈ T (0) . If u ∈ M 1,0 and T u = u, then surely P u = u, because An u = u for every u. On the other hand, for any u ∈ M 1,0 , T P u = P u. P P Because hAn uin∈N is norm-convergent to P u, kT P u − P uk1,∞ = limn→∞ kT An u − An uk1,∞ = limn→∞

1 kT n+1 u − uk1,∞ n+1

= 0. Q Q

Thus, writing U = {u : T u = u}, P [M 1,0 ] = U and P u = u for every u ∈ U . 372E The Ergodic Theorem: second form Let (A, µ ¯) be a measure algebra, and let T : L1 → L1 , where L1 = L1 (A, µ ¯), be a linear operator ofPnorm at most 1 such that T u ∈ L∞ = L∞ (A) and kT uk∞ ≤ n 1 i 1 1 1 kuk∞ whenever u ∈ L1 ∩ L∞ . Set An = n+1 i=0 T : L → L for every n. Then for any u ∈ L , hAn uin∈N 1 1 1 is order*-convergent to an element P u of L . The operator P : L → L is a projection of norm at most 1 onto the linear subspace {u : u ∈ L1 , T u = u}. proof By 371Ga, there is an extension of T to a member T˜ of T (0) . So 372D tells us that hAn uin∈N is order*-convergent to some P u ∈ L1 for every u ∈ L1 , and P : L1 → L1 is a projection of norm at most 1, because P is the restriction of a projection P˜ ∈ T (0) . Also we still have T P u = P u for every u ∈ L1 , and P u = u whenever T u = u, so the set of values P [L1 ] of P must be exactly {u : u ∈ L1 , T u = u}. Remark In 372D and 372E I have used the phrase ‘order*-convergent’ from §367 without always being specific about the partially ordered set in which it is to be interpreted. But, as remarked in 367F, the notion is robust enough for the omission to be immaterial here. Since both M 1,0 and L1 are solid linear subspaces of L0 , a sequence in M 1,0 is order*-convergent to a member of M 1,0 (when order*-convergence is interpreted in the partially ordered set M 1,0 ) iff it is order*-convergent to the same point (when convergence is interpreted in the set L0 ); and the same applies to L1 in place of M 1,0 . 372F Corollary Let (A, µ ¯) be a measure algebra, and π : Af → Af a measure-preserving ring homof morphism, where A = {a : µ ¯a < ∞}. Let T : M 1,0 → M 1,0 be the corresponding Riesz homomorphism, Pn 1 1,0 1,0 i where M = M (A, µ ¯) (366H, in particular part (a-v)). Set An = n+1 T for n ∈ N. Then for every i=0 u ∈ M 1,0 , hAn uin∈N is order*-convergent and k k1,∞ -convergent to some v such that T v = v. proof By 366H(a-iv), T ∈ T (0) , as defined in 371F. So the result follows at once from 372D. 372G Corollary Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean 1 homomorphism. Let T : L → L1 be the corresponding Riesz homomorphism, where L1 = L1 (A, µ ¯). Set Pn 1 i An = n+1 i=0 T for n ∈ N. Then for every u ∈ L1 , hAn uin∈N is order*-convergent and k k1 -convergent. If we set P u = limn→∞ An u for each u, P is the conditional expectation operator corresponding to the closed subalgebra C = {a : πa = a} of A. proof (a) The first part is just a special case of 372F; the point is that because (A, µ ¯) is totally finite, L∞ (A) ⊆ L1 , so M 1,0 (A, µ ¯) = L1 . Also (because µ ¯1 = 1) kuk∞ ≤ kuk1 for every u ∈ L∞ , so the norm k k1,∞ is actually equal to k k1 . (b) For the last sentence, recall that C is a closed subalgebra of A (cf. 333R). By 372D or 372E, P is a projection operator onto the subspace {u : T u = u}. Now [[T u > α]] = π[[u > α]] (365Oc), so T u = u iff [[u > α]] ∈ C for every α ∈ R, that is, iff u belongs to the canonical image of L1 (C, µ ¯¹ C) in L1 (365R). To 1 identify P u further, observe that if u ∈ L , a ∈ C then

R

Tu =

R

R

Tu = a u R i R R R R R (365Ob). Consequently a T u = a u for every i ∈ N, a An u = a u for every n ∈ N, and a P u = a u (because P u is the limit of hAn uin∈N for k k1 ). But this is enough to define P u as the conditional expectation of u on C (365R). a

πa

372K

The ergodic theorem

411

372H The Ergodic Theorem is most often expressed in of transformations of measure spaces. In the next few corollaries I will formulate such expressions. The translation is straightforward, in view of the following. Lemma Let (X, Σ, µ) be a measure space with measure algebra (A, µ ¯). For h ∈ L0 = L0 (µ) write h• for 0 0 the corresponding member of L = L (A) (364Jc). Now let φ : X → X be an inverse-measure-preserving function, π : A → A the corresponding sequentially order-continuous measure-preserving homomorphism defined by setting πE • = (φ−1 [E])• for E ∈ Σ (324M), and T : L0 → L0 the Riesz homomorphism defined by setting T (χa) = χ(πa) for a ∈ A (364R). Then T h• = (hφ)• for any h ∈ L0 . ˜ : X → R be a Σ-measurable function which is equal to h almost everywhere. Because φ−1 [E] proof Let h ˜ a.e., and hφ ˜ is measurable, so hφ ∈ L0 . For any α ∈ R, is negligible for every negligible set E, hφ = hφ ˜ • > α]] = π[[h ˜ • > α]] = π{x : h(x) ˜ [[T h• > α]] = [[T h > α}• ˜ ˜ • > α]] = [[(hφ)• > α]]. = {x : h(φ(x)) > α}• = [[(hφ) 372I Corollary Let (X, Σ, µ) be a measure space and φ : X → X an inverse-measure-preserving function. Let f be a real-valued function which is integrable over X. Then 1 Pn i g(x) = limn→∞ i=0 f (φ (x)) n+1

is defined for almost every x ∈ X, and gφ(x) = g(x) for almost every x. proof Let (A, µ ¯) be the measure algebra of (X, Σ, µ), and π : A → A, T : L0 (A) → L0 (A) the homomorphisms correspondingPto φ, as in 372H. Set u = f • in L1 (A, µ ¯P ). Then for any i ∈ N, T i u = (f φi )• (372H), n n 1 1 i • so setting An = n+1 i=0 T , An u = gn , where gn (x) = n+1 i=0 f (φi (x)) whenever this is defined. Now we know from 372E or 372F that hAn uin∈N is order*-convergent to some v such that T v = v, so hgn in∈N must be convergent almost everywhere (367G), and taking g = limn→∞ gn where this is defined, g • = v. Accordingly (gφ)• = T v = v = g • and gφ = g a.e., as claimed. 372J

The following straightforward facts will be useful in the next corollary and elsewhere.

Lemma Let (X, Σ, µ) be a measure space with measure algebra (A, µ ¯). Let φ : X → X be an inversemeasure-preserving function and π : A → A the associated homomorphism, as in 372H. Set C = {c : c ∈ A, πc = c}, T = {E : E ∈ Σ, φ−1 [E]4E is negligible} and T0 = {E : E ∈ Σ, φ−1 [E] = E}. Then T and T0 are σ-subalgebras of Σ; T0 ⊆ T, T = {E : E ∈ Σ, E • ∈ C}, and C = {E • : E ∈ T0 }. proof It is easy to see that T and T0 are σ-subalgebras of Σ and that T0 ⊆ T = {E : E • ∈ C}. So we have only to check that if c ∈ C there is an E ∈ T0 such that E • = c. P P Start with any F ∈ Σ such that F • = c. −i −i • Now F 4φ [F ] is negligible for every i ∈ N, because (φ [F ]) = π i c = c. So if we set S T E = n∈N i≥n φ−i [F ] = {x : there is an n ∈ N such that φi (x) ∈ F for every i ≥ n}, E • = c. On the other hand, it is easy to check that E ∈ T0 . Q Q 372K Corollary Let (X, Σ, µ) be a probability space and φ : X → X an inverse-measure-preserving function. Let f be a real-valued function which is integrable over X. Then 1 Pn i g(x) = limn→∞ i=0 f (φ (x)) n+1

is defined for almost every x ∈ X; gφ = g a.e., and g is a conditional expectation of f on the σ-algebra T = {E : E ∈ Σ, φ−1 [E]4E is negligible}. If either f is Σ-measurable and defined everywhere in X or φ[E] is negligible for every negligible set E, then g is a conditional expectation of f on the σ-algebra T0 = {E : E ∈ Σ, φ−1 [E] = E}. proof (a) We know by 372I that g is defined almost everywhere and that gφ = g a.e. In the language of the proof of 372I, g • = v is the conditional expectation of u = f • on the closed subalgebra C = {a : a ∈ A, πa = a} = {F • : F ∈ T} = {F • : F ∈ T0 },

412


372K

by 372G and 372J. So v must be expressible as h• where h : X → R is T0 -measurable and is a conditional expectation of f on T0 (and also on T). Since every set of measure zero belongs to T, g = h µ¹ T-a.e., and g is also a conditional expectation of f on T. (b) Suppose now that f is defined everywhere and Σ-measurable. Here I come to a technical obstruction. The definition of ‘conditional expectation’ in 233D asks for g to be µ¹ T0 -integrable, and since µ-negligible sets do not need to be µ¹ T0 -negligible we have some more checking to do, to confirm that {x : x ∈ dom g, g(x) = h(x)} is µ¹ T0 -conegligible as well as µ-conegligible. (i)T For n ∈ N, set Σn = {φ−n [E] : E ∈ Σ}; then Σn is a σ-subalgebra of Σ, including T0 . Set Σ∞ = n∈N Σn , still a σ-algebra including T0 . Now any negligible set E ∈ Σ∞ is µ¹ T0 -negligible. P P For each n ∈ N choose Fn ∈ Σ such that E = φ−n [Fn ]. Because φ is inverse-measure-preserving, every Fn is negligible, so that T S E ∗ = m∈N n∈N,j≥m φ−j [Fn ] T is negligible. Of course E = m∈N φ−m [Fm ] is included in E ∗ . Now T S T S φ−1 [E ∗ ] = m∈N n∈N,j≥m φ−j−1 [Fn ] = m≥1 n∈N,j≥m φ−j [Fn ] = E ∗ because

S n∈N,j≥1

φ−j [Fn ] ⊆

S n∈N,j≥0

φ−j [Fn ].

So E ∗ ∈ T0 and E is included in a negligible member of T0 , which is what we needed to know. Q Q P n 1 ◦ i (ii) We are assuming that f is Σ-measurable and defined everywhere, so that gn = n+1 i=0 f φ ∗ ∗ is Σ-measurable and defined everywhere. If we set g = lim supn→∞ gn , then g : X → [−∞, ∞] is Σ∞ measurable. P P For any m ∈ N, f ◦ φi is Σm -measurable for every i ≥ m, since {x : f (φi (x)) > α} = φ−m [{x : i−m f (φ (x)) > α}] for every α. Accordingly 1 Pn ◦ i g ∗ = lim supn→∞ i=m f φ n+1

∗

is Σm -measurable. As m is arbitrary, g is Σ∞ -measurable. Q Q Since h is surely Σ∞ -measurable, and h = g ∗ µ-a.e., (i) tells us that h = g ∗ µ¹ T0 -a.e. But similarly h = lim inf n→∞ gn µ¹ T0 -a.e., so we must have h = g µ¹ T0 -a.e.; and g, like h, is a conditional expectation of f on T0 . (c) Finally, suppose that φ[E] is negligible for every negligible set E. Then every µ-negligible S set isnµ¹ T0 ∗ negligible. P P If E is µ-negligible, then φ[E], φ2 [E] = φ[φ[E]], . . . are allSnegligible, so E = n∈N φ [E] is T negligible, and there is a measurable negligible set F ⊇ E ∗ . Now F∗ = m∈N n≥m φ−n [F ] is a negligible set in T0 including E, so E is µ¹ T0 -negligible. Q Q Consequently g = h µ¹ T0 -a.e., and in this case also g is a conditional expectation of f on T0 . 372L Remark Parts (b)-(c) of the proof above are dominated by the technical question of the exact definition of ‘conditional expectation of f on T0 ’, and it is natural to be impatient with such details. The kind of example I am concerned about is the following. Let C ⊆ [0, 1] be the Cantor set (134G), and φ : [0, 1] → [0, 1] a Borel measurable function such that φ[C] = [0, 1] and φ(x) = x for x ∈ [0, 1] \ C. (For instance, we could take φ agreeing with the Cantor function on C (134H).) Because C is negligible, φ is inverse-measure-preserving for Lebesgue measure µ, and if f is any Lebesgue integrable function then 1 Pn i g(x) = limn→∞ i=0 f (φ (x)) is defined and equal to f (x) for every x ∈ dom f \ C. But for x ∈ C we n+1

can, by manipulating φ, arrange for g(x) to be almost anything; and if f is undefined on C then g will also be undefined on C. On the other hand, C is not µ¹ T0 -negligible, because the only member of T0 including R C is [0, 1]. So we cannot be sure of being able to form g d(µ¹ T0 ). If instead of Lebesgue measure itself we took its restriction µB to the algebra of Borel subsets of [0, 1], then φ would still be inverse-measure-preserving for µB , but we should now have to worry about the possibility that f ¹C was non-measurable, so that g¹C came out to be non-measurable, even if everywhere defined, and g was not µB ¹ T0 -virtually measurable.

372Mc

The ergodic theorem

413

In the statement of 372K I have offered two ways of being sure that the problem does not arise: check that φ[E] is negligible whenever E is negligible (so that all negligible sets are µ¹ T0 -negligible), or check that f is defined everywhere and Σ-measurable. Even if these conditions are not immediately satisfied in a particular application, it may be possible to modify the problem so that they are. For instance, completing the measure will leave φ inverse-measure-preserving (343Ac), will not change the integrable functions but will make them all measurable (212F, 212Bc), and may enlarge T0 enough to make a difference. If our function f is measurable (because the measure is complete, or otherwise) we can extend it to a measurable function defined everywhere (121I) and the corresponding extension of g will be µ¹ T0 -integrable. Alternatively, if the difficulty seems to lie in the behaviour of φ rather than in the behaviour of f (as in the example above), it may help to modify φ on a negligible set. 372M Continued fractions A particularly delightful application of the results above is to a question which belongs as much to number theory as to analysis. It takes a bit of space to describe, but I hope you will agree with me that it is well worth knowing in itself, and that it also illuminates some of the ideas above. (a) Set X = [0, 1] \ Q. For x ∈ X, set φ(x) = < x1 >, the fractional part of x1 , and k1 (x) = x1 − φ(x), the integer part of x1 ; then φ(x) ∈ X for each x ∈ X, so we may define kn (x) = k1 (φn−1 (x)) for every n ≥ 1. The strictly positive integers k1 (x), k2 (x), k3 (x), . . . are the continued fraction coefficients of x. Of course kn+1 (x) = kn (φ(x)) for every n ≥ 1. Now define hpn (x)in∈N , hqn (x)in∈N inductively by setting p0 (x) = 0, q0 (x) = 1,

p1 (x) = 1,

pn (x) = pn−2 (x) + kn (x)pn−1 (x) for n ≥ 1,

q1 (x) = k1 (x),

qn (x) = qn−2 (x) + kn (x)qn−1 (x) for n ≥ 1.

The continued fraction approximations to x are the quotients pn (x)/qn (x). (I do not discuss rational x, because for my purposes here these are merely distracting. But if we set k1 (0) = ∞, φ(0) = 0 then the formulae above produce the conventional values for kn (x) for rational x ∈ [0, 1[. As for the pn and qn , use the formulae above until you get to x = pn (x)/qn (x), φn (x) = 0, kn+1 (x) = ∞, and then set pm (x) = pn (x), qm (x) = qn (x) for m ≥ n.) (b) The point is that the quotients rn (x) = pn (x)/qn (x) are, relatively speaking, good rational approximations to x. (See 372Yf.) We always have rn+1 (x) < x < rn (x) for every odd n ≥ 1 (372Xj). If x = π − 3, then the first few coefficients are k1 = 7, 1 7

r1 = ,

k2 = 15, r2 =

15 , 106

k3 = 1, r3 =

16 ; 113

the first and third of these corresponding to the classical approximations π l x = e − 2, we get k1 = 1, r1 = 1,

k2 = 2, 2 3

r2 = ,

k3 = 1, 3 4

r3 = ,

k4 = 1, 5 7

r4 = ,

k5 = 4, r5 =

23 , 32

22 , 7

πl

355 . 113

Or if we take

k6 = 1, r6 =

28 ; 39

17 86 note that the obvious approximations 24 , 120 derived from the series for e are not in fact as close as the 5 28 even 7 , 39 above, and involve larger numbers.

(c) Now we need a variety of miscellaneous facts about these coefficients, which I list here. (i) For any x ∈ X, n ≥ 1 we have pn−1 (x)qn (x) − pn (x)qn−1 (x) = (−1)n ,

φn (x) =

(induce on n), so x=

pn (x)+pn−1 (x)φn (x) . qn (x)+qn−1 (x)φn (x)

pn (x)−xqn (x) xqn−1 (x)−pn−1 (x)

414


372Mc

(ii) Another easy induction on n shows that for any finite string m = (m1 , . . . , mn ) of strictly positive integers the set Dm = {x : x ∈ X, ki (x) = mi for 1 ≤ i ≤ n} is an interval in X on which φn is monotonic, being strictly increasing if n is even and strictly decreasing if n is odd. (For the inductive step, note just that D(m1 ,... ,mn ) = [ m11+1 , m11 ] ∩ φ−1 [D(m2 ,... ,mn ) ].) (iii) We also need to know that the intervals Dm of (ii) are small; specifically, that if m = (m1 , . . . , mn ), the length of Dm is at most 2−n+1 . P P All the coefficients pi , qi , for i ≤ n, take constant values p∗i , qi∗ on Dm , since they are determined from the coefficients ki which are constant on Dm by definition. Now every ∗ x ∈ Dm is of the form (p∗n + tp∗n−1 )/(qn∗ + tqn−1 ) for some t ∈ X (see (i) above) and therefore lies between ∗ ∗ ∗ ∗ pn−1 /qn−1 and pn /qn . But the distance between these is ∗ ∗ ¯ ¯ p∗n qn−1 −p∗n−1 qn ¯ ¯= 1 , ∗ ∗ ∗ ∗ qn qn−1

qi∗

qn qn−1

∗ qi−1

∗ + qi−2

∗ ∗ ∗ by the first formula in (i). Next, noting that ≥ for each i ≥ 2, we see that qi∗ qi−1 ≥ 2qi−1 qi−2 ∗ for i ≥ 2, and therefore that qn∗ qn−1 ≥ 2n−1 , so that the length of Dm is at most 2−n+1 . Q Q

372N Theorem Set X = [0, 1] \ Q, and define φ : X → X as in 372M. Then for every Lebesgue integrable function f on X, limn→∞

1 n+1

Pn

f (φi (x)) = i=0

1 ln 2

R 1 f (t) 0 1+t

dt

for almost every x ∈ X. proof (a) The integral just written, and the phrase ‘almost every’, refer of course to Lebesgue measure; but the first step is to introduce another measure, so I had better give a name µL to measure on R Lebesgue 1 X. Let ν be the indefinite-integral measure on X defined by saying that νE = ln12 E 1+x µL (dx) whenever 1 this is defined. The coefficient ln12 is of course chosen to make νX = 1. Because 1+x > 0 for every x ∈ X, dom ν = dom µL and ν has just the same negligible sets as µL (234D); I can therefore safely use the ‘measurable set’, ‘almost everywhere’ and ‘negligible’ without declaring which measure I have in mind each time. (b) Now φ is inverse-measure-preserving when regarded as a function from (X, ν) to itself. P P For each h h 1 1 1 k ≥ 1, set Ik = k+1 , k . On X ∩ Ik , φ(x) = x − k. Observe that φ¹Ik : X ∩ Ik → X is bijective and differentiable relative to its domain in the sense of §262. Consider, for any measurable E ⊆ X, Z Z 1 1 µL (dy) = |φ0 (x)|µL (dx) (y+k)(y+k+1) (φ(x)+k)(φ(x)+k+1) E Ik ∩φ−1 [E] Z x2 1 = µL (dx) = ln 2 · ν(Ik ∩ φ−1 [E]), 2 Ik ∩φ−1 [E]

x+1 x

using 263D (or more primitive results, of course). But P∞ P∞ 1 1 = k=1 − k=1 (y+k)(y+k+1)

y+k

1 y+k+1

=

1 y+1

for every y ∈ [0, 1], so νE =

1 ln 2

∞ Z X

1 µL (dy) (y+k)(y+k+1) E k=1

=

∞ X

ν(Ik ∩ φ−1 [E]) = νφ−1 [E].

k=1

As E is arbitrary, ν is inverse-measure-preserving. Q Q (c) The next thing we need to know is that if E ⊆ X and φ−1 [E] = E then E is either negligible or conegligible. P P I use the sets Dm of 372M(c-ii). (i) For any string m = (m1 , . . . , mn ) of strictly positive integers, we have

372N

The ergodic theorem

x=

415

p∗n +p∗n−1 φn (x) n ∗ +q ∗ qn n−1 φ (x)

for every x ∈ Dm , where p∗n , etc., are defined from m as in 372M(c-iii). Recall also that φn is strictly monotonic on Dm . So for any interval I ⊆ [0, 1] (open, closed or half-open) with endpoints α < β, ∗ φ−n [I] ∩ Dm will be of the form X ∩ J, where J is an interval with endpoints (p∗n + p∗n−1 α)/(qn∗ + qn−1 α), ∗ (p∗n +p∗n−1 β)/(qn∗ +qn−1 β) in some order. This means that we can estimate µL (φ−n [I]∩Dm )/µL Dm , because it is the modulus of ∗ pn +p∗ n−1 β ∗ +q ∗ qn n−1 β ∗ p∗ n +pn−1 ∗ +q ∗ qn n−1

∗ p∗ n +pn−1 α ∗ +q ∗ qn n−1 α

p∗ n ∗ qn

−

−

=

∗ ∗ ∗ (β−α)qn (qn +qn−1 ) ∗ ∗ ∗ β) (q +qn−1 α)(qn +qn−1 ∗ n

≥

∗ (β−α)qn ∗ ∗ qn +qn−1

1 2

≥ (β − α).

Now look at 1 2

A = {E : E ⊆ [0, 1] is Lebesgue measurable, µL (φ−n [E] ∩ Dm ) ≥ µL E · µL Dm }. Clearly the union of two dist of A belongs to A. Because A contains every subinterval of [0, 1] it includes the algebra E of subsets of [0, 1] consisting of finite unions of intervals. Next, the union of any non-decreasing sequence in A belongs to A, and the intersection of a non-increasing sequence likewise. But this means that A must include the σ-algebra generated by E (136G), that is, the Borel σ-algebra. But also, if E ∈ A and H ⊆ [0, 1] is negligible, then 1 2

1 2

µL (φ−n [E4H] ∩ Dm ) = µL (φ−n [E] ∩ Dm ) ≥ µL E · µL Dm = µL (E4H) · µL Dm and E4H ∈ A. And this means that every Lebesgue measurable subset of [0, 1] belongs to A (134Fb). (ii) ?? Now suppose, if possible, that E is a measurable subset of X and that φ−1 [E] = E and E is neither negligible nor conegligible in X. Set γ = 21 µL E > 0. By Lebesgue’s density theorem (223B) there is some x ∈ X \ E such that limδ↓0 ψ(δ) = 0, where ψ(δ) = 1 2γ

1 µL (E 2δ

∩ [x − δ, x + δ]) for δ > 0. Take n so

−n+1

whenever 0 < δ ≤ 2 , and set mi = ki (x) for i ≤ n, so that x ∈ Dm . Taking the large that ψ(δ) < least δ such that Dm ⊆ [x − δ, x + δ], we must have δ ≤ 2−n+1 , because the length of Dm is at most 2−n+1 (372M(c-iii)), while µL Dm ≥ δ, because Dm is an interval. Accordingly µL (E ∩ Dm ) ≤ µL (E ∩ [x − δ, x + δ]) = 2δψ(δ) < γδ ≤ γµL Dm . But we also have µL (E ∩ Dm ) = µL (φ−n [E] ∩ Dm ) ≥ γµL Dm , by (i). X X This proves the result. Q Q (d) The final fact we need in preparation is that φ[E] is negligible for every negligible E ⊆ X. This is because φ is differentiable relative to its domain (see 263D(ii)). (e) Now let f be any µL -integrable function. Because consequently, using (b) above and 372K, g(x) = limn→∞

1 n+1

1 1+x

≤ 1 for every x, f is also ν-integrable (235M);

Pn i=0

f (φi (x))

is defined for almost every x ∈ X, and is a conditional expectation of f (with respect to the measure ν) on the σ-algebra T0 = {E : E is measurable, φ−1 [E] = E}. But we haveRjust seenR that T0 consists only of negligible and conegligible sets, so g must be essentially constant; since g dν = f dν, we must have 1 n→∞ n+1

lim

n X

Z f (φi (x)) =

i=0

for almost every x (using 235M to calculate

R

f dν).

f dν =

1 ln 2

Z 0

1

f (t) µL (dt) 1+t

416


372O

372O Corollary For almost every x ∈ [0, 1] \ Q, 1 n

limn→∞ #({i : 1 ≤ i ≤ n, ki (x) = k}) =

1 (2 ln(k ln 2

+ 1) − ln k − ln(k + 2))

for every k ≥ 1, where k1 (x), . . . are the continued fraction coefficients of x. 1 proof In 372N, set f = χ(X ∩ [ k+1 , k1 ]). Then (for i ≥ 1) f (φi (x)) = 1 if ki (x) = k and zero otherwise. So

lim

1

n→∞ n

#({i : 1 ≤ i ≤ n, ki (x) = k}) 1 n→∞ n

= lim = =

1 ln 2

Z

n X

1 0

1 n→∞ n+1

f (φi (x)) = lim

i=1

f (t) dt 1+t

=

1 ln 2

Z

1/k 1/k+1

n X

f (φi (x))

i=0

1 dt 1+t

1 1 1 (ln(1 + ) − ln(1 + )) ln 2 k k+1

=

1 (2 ln(k ln 2

+ 1) − ln k − ln(k + 2)),

for almost every x ∈ X. 372P Mixing and ergodic transformations This seems an appropriate moment for some brief notes on two special types of measure-preserving homomorphism or inverse-measure-preserving function. Definitions (a) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. (i) π is ergodic if {a : πa = a} = {0, 1}, that is, if the fixed-point subalgebra of π is trivial. (ii) π is mixing (sometimes strongly mixing) if limn→∞ µ ¯(π n a ∩ b) = µ ¯a · µ ¯b for all a, b ∈ A. (b) Let (X, Σ, µ) be a probability space and φ : X → X an inverse-measure-preserving function. (i) φ is ergodic (also called metrically transitive, indecomposable) if every measurable set E such that φ−1 [E] = E is either negligible or conegligible. (ii) φ is mixing if limn→∞ µ(F ∩ φ−n [E]) = µE · µF for all E, F ∈ Σ. (c) Remarks (i) The reason for introducing ‘ergodic’ homomorphisms in this section is of course 372G/372K; if π in 372G, or φ in 372K, is ergodic, then the limit P u or g must be (essentially) constant, being a conditional expectation on a trivial subalgebra. (ii) In the definition (b-i) I should perhaps emphasize that we look only at measurable sets E. We certainly expect that there will generally be many sets E for which φ−1 [E] = E, since any union of orbits of φ will have this property. (iii) Part (c) of the proof of 372N was devoted to showing that the function φ there was ergodic; see also 372Xw. For another ergodic transformation see 372Xo. For examples of mixing transformations see 333P, 372Xm, 372Xn, 372Xr, 372Xu, 372Xv. 372Q

The following facts are elementary.

Proposition (a) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. (i) If π is mixing, it is ergodic. (ii) Let T : L0 = L0 (A) → L0 be the Riesz homomorphism such that T (χa) = χπa for every a ∈ A. Then the following are equiveridical: (α) π is ergodic; (β) the only u ∈ L0 such that R T u = u are the multiples Pn 1 i 1 1 of χ1; (γ) for every u ∈ L = L (A, µ ¯), h n+1 i=0 T uin∈N order*-converges to ( u)χ1. (b) Let (X, Σ, µ) be a probability space, with measure algebra (A, µ ¯). Let φ : X → X be an inversemeasure-preserving function and π : A → A the associated homomorphism such that πE • = (φ−1 [E])• for every E ∈ Σ.

372R

The ergodic theorem

417

(i) φ is mixing iff π is, and in this case it is ergodic. (ii) The following are (α) φ is ergodic; R (β) π is ergodic; (γ) for every µ-integrable realPnequiveridical: 1 i valued function f , h n+1 f (φ (x))i converges to f for almost every x ∈ X. n∈N i=0 proof (a)(i) If π is mixing and πa = a, then 0=µ ¯(a \ a) = limn→∞ µ ¯(π n a \ a) = µ ¯a · µ ¯(1 \ a), so one of µ ¯a, µ ¯(1 \ a) must be zero, and a ∈ {0, 1}. Thus π is ergodic. α)⇒(β β ) T u = u iff π[[u > α]] = [[u > α]] for every α; if π is ergodic, this means that [[u > α]] ∈ {0, 1} (ii)(α for every α, and u must be of the form γχ1, where γ = inf{α : [[u > α]] = 0}. Pn 1 i β )⇒(γγ ) If (β) is true and u ∈ L1 , then we know from 372G that h n+1 (β i=0 T uin∈N is order*convergent and k k1 -convergent to some v such that T v = v; by (β), v is of the form γχ1; and γ=

R

v = limn→∞

1 n+1

Pn R i=0

T iu =

R

u.

α) Assuming (γ), take any a ∈ A such that πa = a, and consider u = χa. Then T i u = χa for (γγ )⇒(α every i, so χa = limn→∞

1 n+1

Pn i=0

R

T i u = ( u)χ1 = µ ¯a · χ1,

and a must be either 0 or 1. (b)(i) Simply translating the definitions, we see that π is mixing iff φ is. In this case φ is ergodic, as in (a-i). α)⇒(β β ) If πa = a there is an E such that φ−1 [E] = E and E • = a, by 372L; now µ (ii)(α ¯a = µE ∈ {0, 1}, so a ∈ {0, 1}. β )⇒(γγ ) Set u = f • ∈ L1 . In the language of (a), T i u = (f φi )• for each i, by 372H, so that (β 1 Pn 1 Pn i • i ( i=0 f φ ) = n+1 i=0 T u n+1 R R R Pn 1 i is order*-convergent to ( u)χ1 = ( f )χ1, and n+1 f a.e. i=0 f φ → α) If φ−1 [E] = E then, applying (γ) to f = χE, we see that χE = µE · χX a.e., so that E is (γγ )⇒(α either negligible or conegligible. 372R

There is a useful sufficient condition for a homomorphism or function to be mixing.

Proposition T (a) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. If n∈N π n [A] = {0, 1}, then π is mixing. (b) Let (X, Σ, µ) be a probability space, and φ : X → X an inverse-measure-preserving function. Set T = {E : for every n ∈ N there is an F ∈ Σ such that E = φ−n [F ]}. If every element of T is either negligible or conegligible, φ is mixing. proof (a) Let T : L0 = L0 (A) → L0 be the Riesz homomorphism associated with π. Take any a, b ∈ A and any non-principal ultrafilter F on N. Then hT n (χa)in∈N is a bounded sequence in the reflexive space L2µ¯ = L2 (A, µ ¯), so v = limn→F T n (χa) is defined for the weak topology of L2µ¯ . Now for each n ∈ N set n Bn = π [A]. This is a closed subalgebra of A (314F(a-i)), and contains π i a for every i ≥ n. So if we identify L2 (Bn , µ ¯¹ Bn ) with the corresponding subspace of L2µ¯ (366I), it contains T i (χa) for every i ≥ n; but also it is norm-closed, therefore weakly closed (3A5Ee), so contains v. This means that [[v > α]] must belong to T Bn for every α and every n. But in this case [[v > α]] ∈ n∈N Bn = {0, 1} for every α, and v is of the form γχ1. Also γ=

R

v = limn→F

So limn→F µ ¯(π n a ∩ b) = limn→F

R

R

T n (χa) = µ ¯a.

T n (χa) × χb =

R

v × χb = γ µ ¯b = µ ¯a · µ ¯b.

418


372R

But this is true of every non-principal ultrafilter F on N, so we must have limn→∞ µ ¯(π n a ∩ b) = µ ¯a · µ ¯b (3A3Lc). As a and b are arbitrary, π is mixing. T (b) The point is that if a ∈ n∈N π n [A], there is an E ∈ T such that E • = a. P P For each n ∈ N there is n • −n • an an ∈ A such that π an = a; say an = Fn where Fn ∈ Σ. Then φ [Fn ] = a. Set T T S S E = m∈N n≥m φ−n [Fn ], Ek = m≥k n≥m φ−(n−k) [Fn ] for each k; then E • = a and φ−k [Ek ] =

S

T m≥k

n≥m

φ−n [Fn ] =

T

S m∈N

n≥m

φ−n [Fn ] = E

for every Q T k, so E ∈ T. Q So n∈N An = {0, 1} and π and φ are mixing. 372X Basic exercisesP(a) Let U be any reflexive Banach space, and T : U → U an operator of norm n 1 i at most 1. Set An = n+1 i=0 T for each n ∈ N. Show that P u = limn→∞ An u is defined (as a limit for the norm topology) for every u ∈ U , and that P : U → U is a projection onto {u : T u = u}. (Hint: show that {u : P u is defined} is a closed linear subspace of U containing T u − u for every u ∈ U .) (This is a version of the mean ergodic theorem.) Pn (0) 1 i >(b) Let (A, µ ¯) be a measure algebra, and T ∈ Tµ¯,¯µ ; set An = n+1 i=0 T for n ∈ N. Take any p ∈ [1, ∞[ p p and u ∈ L = L (A, µ ¯). Show that hAn uin∈N is order*-convergent and k kp -convergent to some v ∈ Lp . (Hint: put 372Xa together with 372D.) (c) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. R R Let P : L1 → L1 be the operator defined as in 365P/366Hb, where L1 = L1 (A, µ ¯), so that a P u = πa u Pn 1 i 1 1 1 for u ∈ L1 , a ∈ A. Set An = n+1 i=0 P : L → L for each i. Show that for any u ∈ L , hAn uin∈N is order*-convergent and k k1 -convergent to the conditional expectation of u on the subalgebra {a : πa = a}. (d) Show that if f is any Lebesgue integrable function on R, and y ∈ R \ {0}, then Pn 1 limn→∞ n+1 k=0 f (x + ky) = 0 for almost every x ∈ R. (e) Let (X, Σ, µ) be a measure space and φ : X → X an inverse-measure-preserving function. Set T = {E : E ∈ Σ, µ(φ−1 [E]4E) = 0}, T0 = {E : E ∈ Σ, φ−1 [E] = E}. (i) Show that T = {E4F : E ∈ T0 , F ∈ Σ, µF = 0}. (ii) Show that a set A ⊆ X is µ¹ T0 -negligible iff φn [A] is µ-negligible for every n ∈ N. R > (f ) Let ν be a Radon probability measure on R such that |t|ν(dt) is finite (cf. 271C-271F). On X = RN let λ be the product measure obtained when each factor is given the measure ν. Define φ : X → X by setting φ(x)(n) = x(n + 1) for x ∈ X, R n ∈ N. (i) Show that φ is inverse-measure-preserving. (Hint: 254G. PnSee also 1 ◦ i 372Xu below.) (iii) Set γ = tν(dt), the expectation of the distribution ν. By considering n+1 i=0 f φ , Pn 1 where f (x) = x(0) for x ∈ X, show that limn→∞ n+1 i=0 x(i) = γ for λ-almost every x ∈ X. > (g) Use the Ergodic Theorem to prove Kolmogorov’s Strong Law of Large Numbers (273I), as follows. Given a complete probability space (Ω, Σ, µ) and an independent identically distributed sequence hFn in∈N of measurable functions from Ω to R, set X = RN and F (ω) = hFn (ω)in∈N for ω ∈ Ω. Show that if we give each copy of R the distribution of F0 then F is inverse-measure-preserving for µ and the product measure λ on X. Now use 372Xf. 1 (h) Show that the continued fraction coefficients of √ are 1, 2, 1, 2, . . . . 2

>(i) For x ∈ X = [0, 1] \ Q let k1 (x), k2 (x), . . . be its continued-fraction coefficients. Show that x 7→ hkn+1 (x) − 1in∈N is a bijection between X and NN which is a homeomorphism if X is given its usual topology (as a subset of R) and NN is given its usual product topology (each copy of N being given the discrete topology).

372Xt

The ergodic theorem

419

(j) For any irrational x ∈ [0, 1] let k1 (x), k2 (x), . . . be its continued-fraction coefficients and pn (x), qn (x) the numerators and denominators of its continued-fraction approximations, as described in 372M. Write rn (x) = pn (x)/qn (x). (i) Show that x lies between rn (x) and rn+1 (x) for every n ∈ N. (ii) Show that rn+1 (x) − rn (x) = (−1)n /qn (x)qn+1 (x) for every n ∈ N. (iii) Show that |x − rn (x)| ≤ 1/qn (x)2 kn (x) for every n ≥ 1. (iv) Hence show that for almost every γ ∈ R, the set {(p, q) : p ∈ Z, q ≥ 1, |γ − pq | ≤ ²/q 2 } is infinite for every ² > 0. (k) Let (A, µ ¯) be an atomless probability algebra. Show that the following are equiveridical: (i) A is homogeneous; (ii) there is an ergodic measure-preserving Boolean homomorphism π : A → A; (iii) there is a mixing measure-preserving automorphism π : A → A. (Hint: 333P.) (l) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. (i) Show that if n ≥ 1 then π is mixing iff π n is mixing. (ii) Show that if n ≥ 1 and π n is ergodic then π is ergodic. (iii) Show that if π is an automorphism then it is ergodic, or mixing, iff π −1 is. > (m) Consider the tent map φα (x) = α min(x, 1 − x) for x ∈ [0, 1], α ∈ [0, 2]. Show that φ2 is inversemeasure-preserving and mixing for Lebesgue measure on [0, 1]. (Hint: show that φn+1 (x) = φ2 (<2n x>) for 2 −n −n n ≥ 1, and hence that µ(I ∩ φ−n [J]) = µI · µJ whenever I is of the form [2 k, 2 (k + 1)] and J is an 2 interval.) (n) Consider the quadratic map ψβ (x) = βx(1 − x) for x ∈ [0, 1], β ∈ [0, 4]. Show that ψ4 p is inversemeasure-preserving and mixing for the Radon measure on [0, 1] with density function t 7→ 1/π t(1 − t). (Hint: show that the transformation t 7→ sin2 πt 2 turns it into the tent map.) Show that for almost every x, limn→∞

1 #({i n+1

: i ≤ n, ψ4i (x) ≤ α}) =

2 π

√ arcsin α

for every α ∈ [0, 1]. (o) Let (X, Σ, µ) be Lebesgue measure on [0, 1[, and fix an irrational number α ∈ [0, 1[. (i) Set φ(x) = x +1 α for every x ∈ [0, 1[, where x +1 α is whichever of x + α, x + α − 1 belongs to [0, Pn1[. Showi that φ is 1 inverse-measure-preserving. (ii) Show that if I ⊆ [0, 1[ is an interval then limn→∞ n+1 i=0 χI(φ (x)) = µI for almost every x ∈ [0, 1[. (Hint: this is Weyl’s Equidistribution Theorem (281N).) (iii) Show that φ is ergodic. (Hint: take the conditional expectation operator P of 372G, and look at P (χI • ) for intervals I.) (iv) Show that φn is ergodic for any n ∈ Z \ {0}. (v) Show that φ is not mixing. (p) Let (A, µ ¯) be a probability algebra and π : A → A a mixing measure-preserving homomorphism. Let T : L0 (A) → L0 (A) be the corresponding homomorphism. Let p, q ∈ [1, ∞] be such that p1 + 1q = 1. Show R R R that limn→∞ T n u × v = u v whenever u ∈ Lp (A, µ ¯) and v ∈ Lq (A, µ ¯). (Hint: start with u, v ∈ S(A).) (q) Let (X, Σ, µ) be a probability space and φ : X → XR a mixing inverse-measure-preserving function. R R Let p, q ∈ [1, ∞] be such that p1 + 1q = 1. Show that limn→∞ f (φn (x))g(x)dx = f g whenever f ∈ Lp (µ) and g ∈ Lq (µ). (r) Give [0, 1[ Lebesgue measure µ, and let k ≥ 2 be an integer. Define φ : [0, 1[ → [0, 1[ by setting φ(x) = , the fractional part of kx. Show that φ is inverse-measure-preserving. Show that φ is mixing. (Hint: if I = [k −n i, k −n (i + 1)[, J = [k −n j, k −n (j + 1)[ then µ(I ∩ φ−m [J]) = µI · µJ for all m ≥ n.) (s) Let (X, Σ, µ) be a probability space and φ : X → X an ergodic inverse-measure-preserving function. R Let f be a µ-virtually measurable function defined almost everywhere on X such that f dµ = ∞. Show 1 Pn i that limn→∞ i=0 f (φ (x)) = ∞ for almost every x ∈ X. (Hint: look at the corresponding limits for n+1

fm = f ∧ mχX.) (t) For irrational x ∈ [0, 1], write k1 (x), k2 (x), . . . for the continued-fraction coefficients of R x. Show that Pn limn→∞ n1 i=1 ki (x) = ∞ for almost every x. (Hint: take φ, ν as in 372N, and show that k1 dν = ∞.)

420


372Xu

(u) Let (X, Σ, µ) be any probability space, and let λ be the product measure on X N . Define φ : X N → X N by setting φ(x)(n) = x(n+1). Show that φ is inverse-measure-preserving. Show that φ satisfies the conditions of 372R, so is mixing. (v) Let (X, Σ, µ) be any probability space, and let λ be the product measure on X Z . Define φ : X Z → X Z by setting φ(x)(n) = x(n + 1). Show that φ is inverse-measure-preserving. Show that φ is mixing. (Hint: show that if C, C 0 are basic cylinder sets then µ(C ∩ φ−n [C 0 ]) = µC · µC 0 for all n large enough.) Show that φ does not satisfy the conditions of 372R. (Compare 333P.) (w) In 372N, let T1 be the family {E : for every n ∈ N there is a measurable set F ⊆ X such that φ−n [F ] = E}. Show that every member of T1 is either negligible or conegligible. (Hint: the argument of part (c) of the proof of 372N still works.) Hence show that φ is mixing for the measure ν. 372Y Further exercises (a) In 372D, show that the null space of the limit operator P is precisely the closure in M 1,0 of the subspace {T u − u : u ∈ M 1,0 }. (0)

(b) Let (A, µ ¯) be a measure algebra and T ∈ Tµ¯,¯µ . Take p ∈ ]1, ∞[ and u ∈ Lp (A, µ ¯), and set u∗ = Pn 1 i supn∈N n+1 i=0 |T u|. (i) Show that for any γ > 0, µ ¯[[u∗ > γ]] ≤

2 γ

R

[[|u|>γ/2]]

|u|. p

(Hint: apply 372C to (|u| − 12 γχ1)+ .) (ii) Show that ku∗ kp ≤ 2( )1/p kukp . (Hint: show that p−1 R∞ α¯ µ[[|u| > α]] + α µ ¯[[|u| > β]]dβ; see 365A. Use 366Xa to show that ku∗ kpp ≤ 2p

R∞ 0

γ p−2

R∞

γ/2

R [[|u|>α]]

|u| =

µ ¯[[|u| > β]]dβdγ + 2p kukpp ,

and reverse the order of integration. Compare 275Yc.) (This is Wiener’s dominated ergodic theorem.) (0)

¯) such that (c) Let (A, µ ¯) be a probability algebra and T an operator in Tµ¯,¯µ . Take u ∈ L1 = L1 (A, µ 1 ¯ is the corresponding function from L0 (A) to h(|u|) ∈ L , where h(t) = tP ln t for t ≥ 1, 0 for t ≤ 1, and h n 1 i ∗ 1 itself. Set u∗ = supn∈N n+1 i=0 |T u|. Show that u ∈ L . (Hint: use the method of 372Yb to show that R∞ R ∗ ¯ µ ¯[[u > γ]]dγ ≤ 2 h(u).) 2 (d) In 372G, suppose that A is atomless. Show that there is always an a ∈ A such that µ ¯a ≤ 12 and i inf i≤n π a 6= 0 for every n, so that (except in trivial cases) hAn (χa)in∈N will not be k k∞ -convergent. (e) Show that an irrational x ∈ ]0, 1[ has an eventually periodic sequence of continued fraction coefficients iff it is a solution of a quadratic equation with integral coefficients. (f ) In the language of 372M-372O and 372Xj, show the following. (i) For any x ∈ X, n ≥ 2, qn (x)qn−1 (x) ≥ 2n−1 , pn (x)pn+1 (x) ≥ 2n−1 , so that qn+1 (x)pn (x) ≥ 2n−1 and |1 − x/rn (x)| ≤ 2−n+1 , | ln x − ln rn (x)| ≤ 2−n+2 . Also |x − rn (x)| ≥ 1/qn (x)qn+2 (x). (ii) For any x ∈ X, n ≥ 1, pn+1 (x) = qn (φ(x)) and Qn−1 Pn−1 qn (x) i=0 rn−i (φi (x)) = 1. (iii) For any x ∈ X, n ≥ 1, | ln qn (x) + i=0 ln φi (x)| ≤ 4. (iv) For almost every x ∈ X, limn→∞

1 n

ln qn (x) = −

1 ln 2

R1

ln t

0 1+t

dt =

π2 . 12 ln 2

(Hint: 225Xi, 282Xo.) (v) For almost every x ∈ X, limn→∞ n1 ln |x − rn (x)| = −π 2 /6 ln 2. (vi) For almost every x ∈ X, 11−n ≤ |x − rn (x)| ≤ 10−n and 3n ≤ qn (x) ≤ 4n for all but finitely many n. (g) In 372M, show that for any measurable setRE ⊆ X, limn→∞ µRL φ−n [E] = νE. (Hint: recall that φ is mixing for ν (372Xw). Hence show that limn→∞ φ−n [E] g dν = νE · g dν for any integrable g. Apply this to a Radon-Nikod´ ym derivative of µL with respect to ν.) (I understand that this result is due to Gauss.)

372 Notes

The ergodic theorem

421

(h) Let h(Xi , Σi , µi )ii∈I be any family of probability spaces, with product (X, Λ, λ). Suppose that for each i ∈ I we are given an inverse-measure-preserving function φi : Xi → Xi . Show that there is a corresponding inverse-measure-preserving function φ : X → X given by setting φ(x)(i) = φi (x(i)) for x ∈ X, i ∈ I. Show that if each φi is mixing so is φ. (i) Give an example of an ergodic measure-preserving automorphism φ : [0, 1[ → [0, 1[ such that φ2 is not ergodic. (Hint: set φ(x) = 21 (1 + φ0 (2x)) for x < 12 , x − 12 for x ≥ 12 . See also 387Xg.) (j) Show that there is an ergodic φ : [0, 1] → [0, 1] such that (ξ1 , ξ2 ) 7→ (φ(ξ1 ), φ(ξ2 )) : [0, 1]2 → [0, 1]2 is not ergodic. (Hint: 372Xo.) (k) Let M be an r × r matrix with integer coefficients and non-zero determinant, where r ≥ 1. Let r r r φ : [0, 1[ → [0, 1[ be the function such that φ(x) − M x ∈ Zr for every x ∈ [0, 1[ . Show that φ is inverser measure-preserving for Lebesgue measure on [0, 1[ . 372 Notes and comments I have chosen an entirely conventional route to the Ergodic Theorem here, through the Mean Ergodic Theorem (372Xa) or, rather, the fundamental lemma underlying it (372A), and the Maximal Ergodic Theorem (372B-372C). What is not to be found in every presentation is the generality here. I speak of arbitrary T ∈ T (0) , the operators which are contractions both for k k1 and for k k∞ , not requiring T to be positive, let alone correspond to a measure-preserving homomorphism. (I do not mention T (0) in the statement of 372C, but of course it is present in spirit.) The work we have done up to this point puts this extra generality within easy reach, but as the rest of the section shows, it is not needed for the principal examples. Only in 372Xc do I offer an application not associated with a measure-preserving homomorphism or an inverse-measure-preserving function. The Ergodic Theorem is an ‘almost-everywhere pointwise convergence theorem’, like the strong law(s) of large numbers and the martingale theorem(s) (§273, §275). Indeed Kolmogorov’s form of the strong law can be derived from the Ergodic Theorem (372Xg). There are some very strong family resemblances. For instance, the Maximal Ergodic Theorem corresponds to the most basic of all the martingale inequalities (275D). Consequently we have similar results, obtained by similar methods, concerning the domination of sequences starting from of Lp (372Yb, 275Yc), though the inequalities are not identical. (Compare also 372Yc with 275Yd.) There are some tantalising reflections of these traits in results surrounding Carleson’s theorem on the pointwise convergence of square-integrable Fourier series (see §286 notes), but Carleson’s theorem seems to be much harder than the others. Other forms of the strong law (273D, 273H) do not appear to fit into quite the same pattern, but I note that here, as with the Ergodic Theorem, we begin with a study of square-integrable functions (see part (e) of the proof of 372D). After 372D, there is a contraction and concentration in the scope of the results, starting with a simple replacement of M 1,0 with L1 (372E). Of course it is almost as easy to prove 372D from 372E as the other way about; I give precedence to 372D only because M 1,0 is the space naturally associated with the class T (0) of operators to which these methods apply. Following this I turn to the special family of operators to which the rest of the section is devoted, those associated with measure-preserving homomorphisms (372F), generally on probability spaces (372G). This is the point at which we can begin to identify the limit as a conditional expectation as well as an invariant element. Next comes the translation into the language of measure spaces and inverse-measure-preserving functions, all perfectly straightforward in view of the lemmas 372H (which was an exercise in §364) and 372J. These turn 372F into 372I and 372G into the main part of 372K. In 372K-372L I find myself writing at some length about a technical problem. The root of the difficulty is in the definition of ‘conditional expectation’. Now it is generally accepted that any pure mathematician has ‘Humpty Dumpty’s privilege’: ‘When I use a word, it means just what I choose it to mean – neither more nor less’. With any privilege come duties and responsibilities; in this context, the duty to be self-consistent, and the responsibility to try to use in ways which will not mystify or mislead the unprepared reader. Having written down a definition of ‘conditional expectation’ in Volume 2, I must either stick to it, or go back and change it, or very carefully explain exactly what modification I wish to make here. I don’t wish to suggest that absolute consistency – in terminology or anything else – is supreme among mathematical virtues. Surely it is better to give local meanings to words, or tolerate ambiguities, than to suppress ideas

422


372 Notes

which cannot be formulated effectively otherwise, and among ‘ideas’ I wish to include the analogies and resonances which a suitable language can suggest. But I do say that it is always best to be conscious of what one is doing – I go farther: one of the things which mathematics is for, is to raise our consciousness of what our thoughts really are. So I believe it is right to pause occasionally over such questions. In 372M-372O (see also 372Xj, 372Xt, 372Xw, 372Yf, 372Yg) I make an excursion into number theory. This is a remarkable example of the power of advanced measure theory to give striking results in other branches of mathematics. Everything here is derived from Billingsley 65, who goes farther than I have space for, and gives references to more. Here let me point to 372Xi; almost accidentally, the construction offers a useful formula for a homeomorphism between two of the most important spaces of descriptive set theory, which will be important to us in Volume 4. I end the section by introducing two , ‘ergodic’ and ‘mixing’ transformation, not because I wish to use them for any new ideas (apart from the elementary 372R, these must wait for §§385-386) but because it may help if I immediately classify some of the inverse-measure-preserving functions we have seen (372Xm372Xo, 372Xr, 372Xt-372Xv). Of course in any application of any ergodic theorem it is of great importance to be able to identify the limits promised by the theorem, and the point about an ergodic transformation is just that our averages converge to constant limits (372Q). Actually proving that a given inverse-measurepreserving function is ergodic is rarely quite trivial (see 372N, 372Xn, 372Xo), though a handful of standard techniques cover a large number of cases, and it is usually obvious when a map is not ergodic, so that if an invariant region does not leap to the eye one has a good hope of ergodicity. I take the opportunity to mention two famous functions from [0, 1] to itself, the ‘tent’ and ‘quadratic’ maps (372Xm, 372Xn). In the formulae φα , ψβ I include redundant parameters; this is because the real importance of these functions lies in the way their behaviour depends, in bewildering complexity, on these parameters. It is only at the extreme values α = 2, β = 4 that the methods of this volume can tell us anything interesting.

373 Decreasing rearrangements I take a section to discuss operators in the class T (0) of 371F-371H and §372 and two associated classes T , T × (373A). These turn out to be intimately related to the idea of ‘decreasing rearrangement’ (373C). In 373D-373F I give elementary properties of decreasing rearrangements; then in 373G-373O I show how they may be used to characterize the set {T u : T ∈ T } for a given u. TheRargument uses a natural topology on the set T (373K). I conclude with remarks on the possible values of T u × v for T ∈ T (373P-373Q) and (0) (0) identifications between Tµ¯,¯ν , Tν¯,¯µ and Tµ¯×,¯ν (373R-373T). 373A Definition Let (A, µ ¯) and (B, ν¯) be measure algebras. Recall that Mµ¯1,∞ = L1 (A, µ ¯) + L∞ (A) 0 + is the set of those u ∈ L (A) such that (|u| − αχ1) is integrable for some α, its norm k k1,∞ being defined by the formulae kuk1,∞ = min{kvk1 + kwk∞ : v ∈ L1 , w ∈ L∞ , v + w = u} Z = min{α + (|u| − αχ1)+ : α ≥ 0} (369Ob). (a) Tµ¯,¯ν will be the space of linear operators T : Mµ¯1,∞ → Mν¯1,∞ such that kT uk1 ≤ kuk1 for every u ∈ L1µ¯ and kT uk∞ ≤ kuk∞ for every u ∈ L∞ (A). (Compare the definition of T (0) in 371F.) (b) If B is Dedekind complete, so that Mµ¯1,∞ , being a solid linear subspace of the Dedekind complete space L0 (B), is Dedekind complete, Tµ¯×,¯ν will be Tµ¯,¯ν ∩ L× (Mµ¯1,∞ ; Mµ¯1,∞ ).

373B

Decreasing rearrangements

423

373B Proposition Let (A, µ ¯) and (B, ν¯) be measure algebras. (a) T = Tµ¯,¯ν is a convex set in the unit ball of B(Mµ¯1,∞ ; Mν¯1,∞ ). (0) (b) If T ∈ T then T ¹Mµ¯1,0 belongs to Tµ¯,¯ν , as defined in 371F. So if T ∈ T , p ∈ [1, ∞[ and u ∈ Lpµ¯ then p T u ∈ Lν¯ and kT ukp ≤ kukp . (c) If B is Dedekind complete and T ∈ T , then T ∈ L∼ (Mµ¯1,∞ ; Mν¯1,∞ ) and T1 ∈ T whenever T1 ∈ L∼ (Mµ¯1,∞ ; Mν¯1,∞ ) and |T1 | ≤ |T |; in particular, |T | ∈ T . (d) If π : A → B is a measure-preserving Boolean homomorphism, then we have a corresponding operator T ∈ T defined by saying that T (χa) = χ(πa) for every a ∈ A. If π is order-continuous, then so is T . ¯ is another measure algebra and T ∈ T , S ∈ T ¯ then ST ∈ T ¯ . (e) If (C, λ) ν ¯ ,λ µ ¯ ,λ proof (a) As 371G, parts (a-i) and (a-ii) of the proof. (b) If u ∈ Mµ¯1,0 and ² > 0, then u is expressible as u0 + u00 where ku00 k∞ ≤ ² and u0 ∈ L1µ¯ . (Set u00 = (u+ ∧ ²χ1) − (u− ∧ ²χ1).) So (|T u| − ²χ1)+ ≤ |T u − T u00 | ∈ L1ν¯ . As ² is arbitrary, T u ∈ Mν¯1,0 ; as u is arbitrary, T ¹Mµ¯1,0 ∈ T (0) . Now the rest is a consequence of 371Gd. (c) Because Mν¯1,∞ is a solid linear subspace of L0 (B), which is Dedekind complete because B is, L∼ (Mµ¯1,∞ ; Mν¯1,∞ ) is a Riesz space (355Ea). Take any u ≥ 0 in Mµ¯1,∞ . Let α ≥ 0 be such that (u − αχ1)+ ∈ L1µ¯ . Because T ¹L1µ¯ belongs to B(L1µ¯ ; L1ν¯ ) = L∼ (L1µ¯ ; L1ν¯ ) (371D), w0 = sup{T v : v ∈ L1µ¯ , 0 ≤ v ≤ (u − αχ1)+ } is defined in L1ν¯ . Now if v ∈ Mµ¯1,∞ and 0 ≤ v ≤ u, we must have T v = T (v − αχ1)+ + T (v ∧ αχ1) ≤ w0 + αχ1 ∈ Mν¯1,∞ . Thus {T v : 0 ≤ v ≤ u} is bounded above in Mν¯1,∞ . As u is arbitrary, T ∈ L∼ (Mµ¯1,∞ ; Mν¯1,∞ ) (355Ba). Now take T1 such that |T1 | ≤ |T | in L∼ (Mµ¯1,∞ ; Mν¯1,∞ ). 371D also tells us that k|T ¹L1µ¯ |k = kT ¹L1µ¯ k, so that kT1 uk1 = k|T1 u|k1 ≤ k|T1 ||u|k1 ≤ k|T ||u|k1 = k sup T vk1 ≤ k sup vk1 = kuk1 |v|≤|u|

for every u ∈ then

L1µ¯

|v|≤|u|

(using one of the formulae in 355Eb for the first equality). At the same time, if u ∈ L∞ (A), |T1 u| ≤ |T1 ||u| ≤ |T ||u| = sup T v |v|≤|u|

≤ sup kT vk∞ χ1 ≤ sup kvk∞ χ1 = kuk∞ χ1, |v|≤|u|

|v|≤|u|

so kT1 uk∞ ≤ kuk∞ . Thus T1 ∈ T . (d) By 365O and 363F, we have norm-preserving positive linear operators T1 : L1µ¯ → L1ν¯ and T∞ : L (A) → L∞ (B) defined by saying that T1 (χa) = χ(πa) whenever µ ¯a < ∞ and T∞ (χa) = χ(πa) for every a ∈ A. If u ∈ S(Af ) = L1µ¯ ∩ S(A) (365F), then T1 u = T∞ u, because both T1 and T∞ are linear and they agree on {χa : µ ¯a < ∞}. If u ≥ 0 in Mµ¯∞,1 = L1µ¯ ∩ L∞ (A), there is a non-decreasing sequence hun in∈N in f S(A ) such that u = supn∈N un and ∞

limn→∞ ku − un k1 = limn→∞ ku − un k∞ = 0 (see the proof of 369Od), so that T1 u = supn∈N T1 un = supn∈N T∞ un = T∞ u. Accordingly T1 and T∞ agree on L1µ¯ ∩L∞ (A). But this means that if u ∈ Mµ¯1,∞ is expressed as v+w = v 0 +w0 , where v, v 0 ∈ L1µ¯ and w, w0 ∈ L∞ (A), we shall have

424


373B

T1 v 0 + T∞ w0 = T1 v + T∞ w + T1 (v 0 − v) − T∞ (w − w0 ) = T1 v + T∞ w, because v 0 − v = w − w0 ∈ Mµ¯∞,1 . Accordingly we have an operator T : Mµ¯1,∞ → Mν¯1,∞ defined by setting T (v + w) = T1 v + T∞ w whenever v ∈ L1µ¯ , w ∈ L∞ (A). This formula makes it easy to check that T is linear and positive, and it clearly belongs to T . To see that T is uniquely defined, observe that T ¹L1µ¯ and T ¹L∞ (A) are uniquely defined by the values T takes on S(Af ), S(A) respectively, because these spaces are dense for the appropriate norms. Now suppose that π is order-continuous. Then T1 and T∞ are also order-continuous (365Oa, 363Ff). If A ⊆ Mµ¯1,∞ is non-empty and downwards-directed and has infimum 0, take u0 ∈ A and γ > 0 such that (u0 − γχ1)+ ∈ L1µ¯ . Set A1 = {(u − γχ1)+ : u ∈ A, u ≤ u0 },

A∞ = {u ∧ γχ1 : u ∈ A}.

Then A1 ⊆ L1µ¯ and A∞ ⊆ L∞ (A) are both downwards-directed and have infimum 0, so inf T1 [A1 ] = inf T∞ [A∞ ] = 0 in L0 (B). But this means that inf(T1 [A1 ] + T∞ [A∞ ]) = 0 (351Dc). Now any w ∈ T1 [A1 ] + T∞ [A∞ ] is expressible as T (u − γχ1)+ + T (u0 ∧ γχ1) where u, u0 ∈ A; because A is downwardsdirected, there is a v ∈ A such that v ≤ u ∧ u0 , in which case T v ≤ w. Accordingly T [A] must also have infimum 0. As A is arbitrary, T is order-continuous. (e) is obvious, as usual. 373C Decreasing rearrangements The following concept is fundamental to any understanding of ¯) for the set of those u ∈ L0 (A) the class T . Let (A, µ ¯) be a measure algebra. Write Mµ¯0,∞ = M 0,∞ (A, µ such that µ ¯[[|u| > α]] is finite for some α ∈ R. (See 369N for the ideology of this notation.) It is easy to see that Mµ¯0,∞ is a solid linear subspace of L0 (A). Let (AL , µ ¯L ) be the measure algebra of Lebesgue measure , defined by setting u∗ = g • , where on [0, ∞[. For u ∈ Mµ¯0,∞ its decreasing rearrangement is u∗ ∈ Mµ¯0,∞ L g(t) = inf{α : α ≥ 0, µ ¯[[|u| > α]] ≤ t} for every t > 0. (This is always finite because inf α∈R µ ¯[[|u| > α]] = 0, by 364A(β) and 321F.) I will maintain this usage of the symbols AL , µ ¯L , u∗ for the rest of this section. 373D Lemma Let (A, µ ¯) be a measure algebra. 0,∞ (a) For any u ∈ Mµ¯ , its decreasing rearrangement u∗ may be defined by the formula [[u∗ > α]] = [0, µ ¯[[|u| > α]][ for every α ≥ 0, •

that is, µ ¯L [[u∗ > α]] = µ ¯[[|u| > α]] for every α ≥ 0. (b) If |u| ≤ |v| in Mµ¯0,∞ , then u∗ ≤ v ∗ ; in particular, |u|∗ = u∗ . Pn Pn • ∗ ¯ai [ . (c)(i) If u = P i=0 αi χ [0, µ i=0 αi χai , where a0 ⊇ a1 ⊇ . . . ⊇ an and αi ≥ 0 for each i, then u = n ∗ i=0 αi χai where Pn (ii) If u = P a0 , . . . , an are dist and |α0 | ≥ |α1 | ≥ . . . ≥ |αn |, then u = • |α |χ [β , β [ , where β = ¯ai for i ≤ n + 1. i i i+1 i i=0 j 0

Rt 0

u∗ = inf α≥0 αt +

R

(|u| − αχ1)+ .

(h) If A ⊆ (Mµ¯0,∞ )+ is non-empty and upwards-directed and has supremum u0 ∈ Mµ¯0,∞ , then u∗0 = supu∈A u∗ .

373D


425

proof (a) Set g(t) = inf{α : µ ¯[[|u| > α]] ≤ t} as in 373C. If α ≥ 0, g(t) > α ⇐⇒ µ ¯[[|u| > β]] > t for some β > α ⇐⇒ µ ¯[[|u| > α]] > t (because [[|u| > α]] = supβ>α [[|u| > β]]), so [[u∗ > α]] = {t : g(t) > α}• = [0, µ ¯[[|u| > α]][ . •

Of course this formula defines u∗ . (b) This is obvious, either from the definition in 373C or from (a) just above. Pn • (c)(i) Setting v = i=0 αi χ [0, µ ¯ai [ , we have [[v > α]] = 0 if

n X

αi ≤ α

i=0 •

= [0, µ ¯aj [ if

j−1 X

αi ≤ α <

j X

i=0

αi

i=0

•

= [0, µ ¯a0 [ if 0 ≤ α < α0 , •

and in all cases is equal to [0, µ ¯[[|u| > α]][ . (ii) A similar argument applies. (If any aj has infinite measure, then ai is irrelevant for i > j.) (d) Fix γ > 0 for the moment, and consider A = {E : E ⊆ ]γ, ∞[ is a Borel set, µ ¯L [[u∗ ∈ E]] = µ ¯[[|u| ∈ E]]}, I = {]α, ∞[ : α ≥ γ}. Then I ⊆ A (by (a)), I ∩ J ∈ I for S all I, J ∈ I, E \ F ∈ A whenever E, F ∈ A and F ⊆ E (because u ∈ Mµ¯0 , so µ ¯[[|u| ∈ E]] < ∞), and n∈N En ∈ A whenever hEn in∈N is a non-decreasing sequence in A. So, by the Monotone Class Theorem (136B), A includes the σ-algebra of subsets of ]γ, ∞[ generated by I; but this must contain E ∩ ]γ, ∞[ for every Borel set E ⊆ R. Accordingly, for any Borel set E ⊆ ]0, ∞[, µ ¯L [[u∗ ∈ E]] = supn∈N µ ¯L [[u∗ ∈ E ∩ ]2−n , ∞[ ]] = µ ¯[[|u| ∈ E]]. (e) For any α > 0, Eα = {t : h(t) > α} is a Borel subset of ]0, ∞[. If u ∈ Mµ¯0 then, using (d) above, ∗ ¯ ∗ ) > α]] = µ ¯ ¯ µ ¯L [[h(u ¯L [[u∗ ∈ Eα ]] = µ ¯[[u ∈ Eα ]] = µ ¯[[h(u) > α]] = µ ¯L [[(h(u)) > α]]. ∗ ¯ ¯ ∗ ) are equivalence classes of non-increasing functions, they must be equal. As both (h(u)) and h(u If h is continuous on the left, then Eα = ]γ, ∞[ for some γ, so we no longer need to use (d), and the argument works for any u ∈ (Mµ¯0,∞ )+ .

(f ) Apply (e) with h(β) = max(0, β − α). (g) Express u∗ as g • , where g(s) = inf{α : µ ¯[[|u| > α]] ≤ s} for every s > 0. Because g is non-increasing, it is easy to check that, for t > 0,

Rt 0

g = tg(t) +

for every α ≥ 0; so that

R∞ 0

max(0, g(s) − g(t))ds ≤ αt +

Rt 0

u∗ = minα≥0 αt +

R

R∞ 0

max(0, g(s) − α)ds

(u∗ − αχ1)+ .

426


373D

Now Z

Z

∞

(u∗ − αχ1)+ =

µ ¯L [[(u∗ − αχ1)+ > β]]dβ 0

Z

Z

∞

(|u| − αχ1)+

+

=

µ ¯[[(|u| − αχ1) > β]]dβ = 0

for every α ≥ 0, using (f) and 365A, and

Rt 0

u∗ = minα≥0 αt +

R

(|u| − αχ1)+ .

(h) µ ¯[[u0 > α]] = µ ¯(supu∈A [[u > α]]) = supu∈A µ ¯[[u > α]] for any α > 0, using 364Mb and 321D. So [[u∗0 > α]] = [0, µ ¯[[u0 > α]][ = supu∈A [0, µ ¯[[u > α]][ = [[supu∈A u∗ > α]] •

•

for every α, and u∗0 = supu∈A u∗ . R

R

u∗ × v ∗ for all u, v ∈ Mµ¯0,∞ . Pm Pn proof (a) Consider first the case u, v ≥ 0 in S(A). Then we may express u, v as i=0 αi χai , j=0 βj χbj where a0 ⊇ a1 ⊇ . . . ⊇ am , b0 ⊇ . . . ⊇ bn in A and αi , βj ≥ 0 for all i, j (361Ec). Now u∗ , v ∗ are given by Pn Pm • • ¯ bj [ ¯ai [ , v ∗ = j=0 βj χ [0, µ u∗ = i=0 αi χ [0, µ 373E Theorem Let (A, µ ¯) be a measure algebra. Then

|u × v| ≤

(373Dc). So Z u×v = =

m X n X i=0 j=0 m X n X

αi βj µ ¯(ai ∩ bj ) ≤

m X n X

αi βj min(¯ µai , µ ¯bj )

i=0 j=0

Z

αi βj µL ([0, µ ¯ai [ ∩ [0, µ ¯bj [) =

u∗ × v ∗ .

i=0 j=0

(b) For the general case, we have non-decreasing sequences hun in∈N , hvn in∈N in S(A)+ with suprema |u|, |v| respectively (364Kd), so that |u × v| = |u| × |v| = supn∈N |u| × vn = supm,n∈N um × vn = supn∈N un × vn and

R

|u × v| =

R

supn∈N un × vn = supn∈N

R

un × vn ≤ supn∈N

R

u∗n × vn∗ ≤

using 373Db. 373F Theorem Let (A, µ ¯) be a measure algebra, and u any member of Mµ¯0,∞ . (a) For any p ∈ [1, ∞], u ∈ Lpµ¯ iff u∗ ∈ Lpµ¯L , and in this case kukp = ku∗ kp . (b)(i) u ∈ Mµ¯0 iff u∗ ∈ Mµ¯0L ; (ii) u ∈ Mµ¯1,∞ iff u∗ ∈ Mµ¯1,∞ , and in this case kuk1,∞ = ku∗ k1,∞ ; L (iii) u ∈ Mµ¯1,0 iff u∗ ∈ Mµ¯1,0 ; L (iv) u ∈ Mµ¯∞,1 iff u∗ ∈ Mµ¯∞,1 , and in this case kuk∞,1 = ku∗ k∞,1 . L proof (a)(i) Consider first the case p = 1. In this case

R

|u| =

R∞ 0

µ ¯[[|u| > α]]dα =

R∞

0 ∗ p

µ ¯L [[u∗ > α]]dα =

(ii) If 1 < p < ∞, then by 373De we have (|u|p )∗ = (u ) , so that kukpp =

R

|u|p =

R

(|u|p )∗ =

R

(u∗ )p = ku∗ kpp

R

u∗ .

R

u∗ × v ∗ ,

373H


427

if either kukp or ku∗ kp is finite. (iii) As for p = ∞, kuk∞ ≤ γ ⇐⇒ [[|u| > γ]] = 0 ⇐⇒ [[u∗ > γ]] = 0 ⇐⇒ ku∗ k∞ ≤ γ. (b)(i) u ∈ Mµ¯0 ⇐⇒ µ ¯[[|u| > α]] < ∞ for every α > 0 ⇐⇒ µ ¯L [[u∗ > α]] < ∞ for every α > 0 ⇐⇒ u∗ ∈ Mµ¯0L . (ii) For any α ≥ 0,

R

(|u| − αχ1)+ =

R

(u∗ − αχ1)+

as in the proof of 373Dg. So kuk1,∞ = ku∗ k1,∞ if either is finite, by the formula in 369Ob. (iii) This follows from (i) and (ii), because M 1,0 = M 0 ∩ M 1,∞ . (iv) Allowing ∞ as a value of an integral, we have Z kuk1,∞ = min{α + (|u| − αχ1)+ : α ≥ 0} Z = min{α + (u∗ − αχ1)+ : α ≥ 0} = ku∗ k1,∞ . by 369Ob; in particular, u ∈ Mµ¯1,∞ iff u∗ ∈ Mµ¯1,∞ L 373G Lemma Let (A, µ ¯) and (B, ν¯) be measure algebras. If 1,∞ either u ∈ Mµ¯ and T ∈ Tµ¯,¯ν (0)

or u ∈ Mµ¯1,0 and T ∈ Tµ¯,¯ν , Rt then 0 (T u)∗ ≤ 0 u∗ for every t ≥ 0. Rt

proof Set T1 = T ¹L1µ¯ , so that kT1 k ≤ 1 in B(L1µ¯ ; L1ν¯ ), and |T1 | is defined in B(L1µ¯ ; L1ν¯ ), also with norm at most 1. If α ≥ 0, then we can express u as u1 + u2 where |u1 | ≤ (|u| − αχ1)R+ and |u2 | ≤ αχ1. (Let w ∈ L∞ (A) be such that kwk∞ ≤ 1, u = |u| × w; set u2 = w × (|u| ∧ αχ1).) So if (|u| − αχ1)+ < ∞, |T u| ≤ |T u1 | + |T u2 | ≤ |T1 ||u1 | + αχ1 and

R

R

R

R

(|T u| − αχ1)+ ≤ |T1 ||u1 | ≤ |u1 | ≤ (|u| − αχ1)+ . Rt Rt The formula of 373Dg now tells us that 0 (T u)∗ ≤ 0 u∗ for every t. 373H Lemma Let (A, µ ¯) be a measure algebra, and θ : Af → R an additive functional, where Af = {a : µ ¯a < ∞}. (a) The following are equiveridical: 1

(α) limt↓0 supµ¯a≤t |θa| = limt→∞ supµ¯a≤t |θa| = 0, t R 1,0 1,0 (β) there is some u ∈ M = Mµ¯ such that θa = a u for every a ∈ Af , and in this case u is uniquely defined. (b) Now suppose that (A, µ ¯) is localizable. Then the following are equiveridical: (α) limt↓0 supµ¯a≤t |θa| = 0, (β) there is some u ∈ M 1,∞ and again this u is uniquely defined.

1

lim supt→∞ supµ¯a≤t |θa| < ∞, t R 1,∞ = Mµ¯ such that θa = a u for every a ∈ Af ,

f f proof (a)(i) Assume (α). R For a, c ∈ A , setf θc (a) = θ(a ∩ c). Then for each c ∈ A , there is a unique 1 uc ∈ Lµ¯ such that θc a = a uc for every a ∈ A (365Eb). Because uc is unique we must have uc = ud × χc

428


373H

whenever c ⊆ d ∈ Af . Next, given α > 0, there is a t0 ≥ 0 such that |θa| ≤ α¯ µa whenever a ∈ Af and f + µ ¯a ≥ t0 ; so that µ ¯[[uc > α]] ≤ t0 for every c ∈ A , and e(α) = supc∈Af [[uc > α]] is defined in Af . Of course + 0 0 e(α) = [[ue(1) > α]] for every α ≥ 1, so inf α∈R e(α) = 0, and v1 = supc∈Af u+ c is defined in L = L (A) (364Mb). Because [[v1 > α]] = e(α) ∈ Af for each α > 0, v1 ∈ Mµ¯0 . For any a ∈ Af , + v1 × χa = supc∈Af u+ c × χa = ua ,

R R f so v1 ∈ M 1,0 and a v1 = a u+ a for every a ∈ A . R R − f Similarly, v2 = supc∈Af uc is defined in M 1,0 and a v2 = a u− a for every a ∈ A . So we can set 1,0 u = v1 − v2 ∈ M and get

R

a

u=

R

a

ua = θa

for every a ∈ Af . Thus (β) is true.

R + (ii) ¯a ≤ δ (365Ea), so R Assume (β). If ² > 0, there is a δ > 0 such that a (|u| − ²χ1) ≤ ²R whenever µ that | a u| ≤ ²(1 + µ ¯a) whenever µ ¯a ≤ δ. As ² is arbitrary, limt↓0 supµ¯a≤t | a u| = 0. Moreover, whenever R R t > 0 and µ ¯a ≤ t, 1t | a u| ≤ ² + 1t (|u| − ²χ1)+ . Thus lim supt→∞

1 t

supµ¯a≤t |

R

a

u| ≤ ².

As ² is arbitrary, θ satisfies the conditions in (α). (iii) The uniqueness of u is a consequence of 366Gd. (b) The argument for (b) uses the same ideas.

R (i) Assume (α). Again, for each c ∈ Af , we have uc ∈ L1 such that θc a = a uc for every a ∈ Af ; again, set e(α) = supc∈Af [[u+ c > α]], which is defined because A is supposed to be Dedekind complete. This time, there are t0 , γ ≥ 0 such that |θa| ≤ γ µ ¯a whenever a ∈ Af and µ ¯a ≥ t0 ; so that µ ¯[[uc > γ]] ≤ t0 for every c ∈ Af , and µ ¯e(γ) < ∞. Accordingly inf α≥γ e(α) = inf α≥γ [[u+ e(γ) > α]] = 0, 0 0 + 1 f and once more v1 = supc∈Af u+ c is defined in L = L (A). As before, v1 × χa = ua ∈ L for any a ∈ A , 1,∞ f 1,∞ − , with v2 ×χa = u− Because [[v1 > γ]] = e(γ) ∈ A , v1 ∈ M . Similarly, v2 = supc∈Af uc is defined in M a f 1,∞ for every a ∈ A . So u = v1 − v2 ∈ M , and

R

a

u=

R

a

ua = θa

for every a ∈ Af .

R + (ii) Take γ ≥ 0 such that β = (|u| R Assume (β). R − γχ1) is finite. If ² > 0, there is a δ > 0 such + that a (|u| − γχ1) ≤ ² whenever µ ¯a ≤ δ, so that | a u| ≤ ² + γ µ ¯a whenever µ ¯a ≤Rδ. As ² is arbitrary, R R limt↓0 supµ¯a≤t | a u| = 0. Moreover, whenever t > 0 and µ ¯a ≤ t, then 1t | a u| ≤ γ + 1t (|u| − ²χ1)+ . Thus R lim supt→∞ 1t supµ¯a≤t | a u| ≤ γ < ∞, R and the function a 7→ a u satisfies the conditions in (β). (iii) u is uniquely defined because u × χa must be ua , as defined in (i), for every a ∈ Af , and (A, µ ¯) is semi-finite. 373I Lemma Suppose that u, v, w ∈ Mµ¯0,∞ are all equivalence classes of non-negative non-increasing Rt Rt RL R functions. If 0 u ≤ 0 v for every t ≥ 0, then u × w ≤ v × w. P 4n proof For n ∈ N, i ≤ 4n set ani = [[w > 2−n i]]; set wn = i=1 2−n χani . Then each ani is of the form [0, t]• , so

R

u × wn =

P 4n

i=1

2−n

R

ani

u≤

P4n

i=1

2−n

R

ani

v=

Also hwn in∈N is a non-decreasing sequence with supremum w, so

R

u × w = supn∈N

R

u × wn ≤ supn∈N

R

v × wn =

R

R

v × wn .

v × w.

373L


429

373J Corollary Suppose that (A, µ ¯) and (B, ν¯) are measure algebras and v ∈ Mν¯0,∞ . If (0)

either u ∈ Mµ¯1,0 and T ∈ Tµ¯,¯ν then

R

or u ∈ Mµ¯1,∞ and T ∈ Tµ¯,¯ν R |T u × v| ≤ u∗ × v ∗ .

proof Put 373E, 373G and 373I together. 373K The very weak operator topology of T Let (A, µ ¯) and (B, ν¯) be two measure algebras. For u ∈ Mµ¯1,∞ , w ∈ Mν¯∞,1 set ρuw (S, T ) = |

R

Su × w −

R

T u × w| for all S, T ∈ T = Tµ¯,¯ν .

Then ρuw is a pseudometric on T . I will call the topology generated by {ρuw : u ∈ Mµ¯1,∞ , w ∈ Mν¯∞,1 } (2A3F) the very weak operator topology on T . 373L Theorem Let (A, µ ¯) be a measure algebra and (B, ν¯) a localizable measure algebra. Then T = Tµ¯,¯ν is compact in its very weak operator topology. proof Let F be an ultrafilter on T . If u ∈ Mµ¯1,∞ , w ∈ Mν¯∞,1 then |

R

T u × w| ≤

R

u∗ × w ∗ < ∞

R ∗ for every T ∈ T (373J); u × w∗ is finite because u∗ ∈ M 1,∞ and w∗ ∈ M ∞,1 (373F). R R In particular, { T u × R w : T ∈ T } is bounded. Consequently hu (w) = limT →F T u × w is defined in R (2A3Se). Because w 7→ T u × w is additive for every T ∈ T , so is hu . Also |hu (w)| ≤

R

u∗ × w∗ ≤ ku∗ k1,∞ kw∗ k∞,1 = kuk1,∞ kwk∞,1

for every w ∈ Mν¯∞,1 . Rt |hu (χb)| ≤ 0 u∗ whenever b ∈ Bf and ν¯b ≤ t. So limt↓0 supν¯b≤t |hu (χb)| ≤ limt↓0 lim supt→∞

1 t

Rt 0

u∗ = 0,

supν¯b≤t |hu (χb)| ≤ lim supt→∞

1 t

Rt 0

u∗ < ∞.

R Of course b 7→ hu (χb) is additive, so by 373H there is a unique Su ∈ Mν¯1,∞ such that hu (χb) = b Su for R every b ∈ Bf . Since both hu and w 7→ Su × w are linear and continuous on Mν¯∞,1 , and S(Bf ) is dense in Mν¯∞,1 (369Od),

R

Su × w = hu (w) = limT →F

Mν¯∞,1 .

R

Tu × w

Mµ¯1,∞ .

for every w ∈ And this is true for every u ∈ R R For any particular w ∈ Mν¯∞,1 , all the maps u 7→ T u × w are linear, so u 7→ Su × w also is; that is, S : Mµ¯1,∞ → Mν¯1,∞ is linear. Now S ∈ T . P P (α) If u ∈ L1µ¯ and b, c ∈ Bf , then Z Z Z Z T u × (χb − χc) ≤ sup T u × (χb − χc) Su − Su = lim b

c

T →F

T ∈T

≤ sup kT uk1 kχb − χck∞ ≤ kuk1 . T ∈T

But, setting e = [[Su > 0]], we have Z Z Z |Su| = Su − Su e 1\e Z Su + = sup b∈Bf ,b ⊆ e

b

sup c∈Bf ,c ⊆ 1\e

(−Su) ≤ kuk1 .

430


(β) If u ∈ L∞ (A), then |

R

Su| ≤ supT ∈T | b

R

373L

T u × χb| ≤ supT ∈T kT uk∞ ν¯b ≤ kuk∞ ν¯b

for every b ∈ Bf . So [[Su > kuk∞ ]] = [[−Su > kuk∞ ]] = 0 and kSuk∞ ≤ kuk∞ . (Note that both parts of this argument depend on knowing that (B, ν¯) is semi-finite, so that we cannot be troubled by purely infinite elements of B.) Q Q Of course we now have limT →F ρuw (T, S) = 0 for all u ∈ Mµ¯1,0 , w ∈ Mν¯∞,1 , so that S = lim F in T . As F is arbitrary, T is compact (2A3R). 373M Corollary Let (A, µ ¯) be a measure algebra and (B, ν¯) a localizable measure algebra, and u any member of Mµ¯1,∞ . Then B = {T u : T ∈ Tµ¯,¯ν } is compact in Mν¯1,∞ for the topology Ts (Mν¯1,∞ , Mν¯∞,1 ). proof The point is just that the map T 7→ T u : Tµ¯,¯ν → Mν¯1,0 is continuous for the very weak operator topology on Tµ¯,¯ν and Ts (Mν¯1,∞ , Mν¯∞,1 ). So B is a continuous image of a compact set, therefore compact (2A3Nb). 373N Corollary Let (A, µ ¯) be a measure algebra, (B, ν¯) a localizable measure algebra and u any member of Mµ¯1,∞ ; set B = {T u : T ∈ Tµ¯,¯ν }. If hvn in∈N is any non-decreasing sequence in B, then supn∈N vn is defined in Mν¯1,∞ and belongs to B. R proof By 373M, hvn in∈N must have a cluster point v ∈ B for Ts (Mν¯1,∞ , Mν¯∞,1 ). Now for any b ∈ Bf , b v R R R must be a cluster point of h n vn in∈N , because w 7→ b w is continuous for Ts (Mν¯1,∞ , MνR¯∞,1 ). But h b vnRin∈N is a non-decreasing sequence, so its only possible cluster point is its supremum; thus b v = limn→∞ b vn . Consequently v × χb must be the supremum of {vn × χb : n ∈ N} in L1 . And this is true for every b ∈ Bf ; as (B, ν¯) is semi-finite, v is the supremum of hvn in∈N in L0 (B) and in Mν¯1,∞ . 373O Theorem Let (A, µ ¯), (B, ν¯) be measure algebras and u ∈ Mµ¯1,∞ , v ∈ Mν¯1,∞ . Then the following are equiveridical: (i) there is a T ∈ Tµ¯,¯ν such that T u = v, Rt Rt (ii) 0 v ∗ ≤ 0 u∗ for every t ≥ 0. In particular, given u ∈ Mµ¯1,∞ , there are S ∈ Tµ¯,¯µL , T ∈ Tµ¯L ,¯µ such that Su = u∗ , T u∗ = u. proof (i)⇒(ii) is Lemma 373G. Accordingly I shall devote the rest of the proof to showing that (ii)⇒(i). (a) If (A, µ ¯), (B, ν¯) are measure algebras and u ∈ Mµ¯1,∞ , v ∈ Mν¯1,∞ , I will say that v 4 u if there is a T ∈ Tµ¯,¯ν such that T u = v, and that v ∼ u if v 4 u and u 4 v. (Properly speaking, I ought to write (u, µ ¯) 4 (v, ν¯), because we could in principle have two different measures on the same algebra. But I do not think any confusion is likely to arise in the argument which follows.) By 373Be, 4 is transitive and ∼ is an equivalence relation. Now we have the following facts. P There (b) If (A, µ ¯) is a measure algebra and u1 , u2 ∈ Mµ¯1,∞ are such that |u1 | ≤ |u2 |, then u1 4 u2 . P 1,∞ ∞ is a w ∈ L (A) such that u1 = w × u2 and kwk∞ ≤ 1. Set T v = w × v for for v ∈ Mµ¯ ; then T ∈ Tµ¯,¯µ and T u2 = u1 . Q Q So u ∼ |u| for every u ∈ Mµ¯1,∞ . (c) If (A, µ ¯P ) is a measure algebra and u ≥ 0 in S(A), then u 4 u∗ . P P If u = 0 this is trivial. Otherwise, n express u as i=0 αi χai where a0 , . . . , an are dist and non-zero and α0 > α1 . . . > αn > 0 ∈ R. If µ ¯ai = ∞ for any i, take m to be minimal subject to µ ¯am = ∞; otherwise, set m = n. Then u∗ = Pm Pj−1 • ¯ai for 1 ≤ j ≤ m + 1. i=0 αi χ [βi , βi+1 [ , where β0 = 0, βj = i=0 µ For i < m, and for i = m if µ ¯am < ∞, define hi : Mµ¯1,∞ → R by setting hi (v) =

1 µ ¯ai

R

ai

v

for every v ∈ Mµ¯1,∞ . If µ ¯am = ∞, then we need a different idea to define hm , as follows. Let I be {a : a ∈ A, µ ¯(a∩am ) < ∞}. Then I is an ideal of A not containing am , so there is a Boolean homomorphism π : A → {0, 1} such that πa = 0 for a ∈ I and πam = 1 (311D). This induces a corresponding k k∞ -continuous

373O


431

linear operator h : L∞ (A) → L∞ ({0, 1}) ∼ ¯a < ∞, and accordingly = R, as in 363F. Now h(χa) = 0 whenever µ h(v) = 0 whenever v ∈ Mµ¯∞,1 , since S(Af ) is dense in Mµ¯∞,1 for k k∞,1 and therefore also for k k∞ . But this means that h has a unique extension to a linear functional hm : Mµ¯1,∞ → R such that hm (v) = 0 for every v ∈ L1µ¯ , while hm (χam ) = 1 and |h(v)| ≤ kvk∞ for every v ∈ L∞ (A). Having defined hi for every i ≤ m, define T : Mµ¯1,∞ → Mµ¯1,∞ by setting L Pm • T v = i=0 hi (v)χ [βi , βi+1 [ for every v ∈ Mµ¯1,∞ . For any i ≤ m, v ∈ L1µ¯ ,

R βi+1 βi

|T v| = |hi (v)|¯ µai ≤

R ai

|v|;

summing over i, kT vk1 ≤ kvk1 . Similarly, for any i ≤ m, v ∈ L∞ (B), |hi (v)| ≤ kvk∞ , so kT vk∞ ≤ kvk∞ . Thus T ∈ Tµ¯,¯µL . Since u∗ = T u, we conclude that u∗ 4 u, as claimed. Q Q (d) If (A, µ ¯) is a measure algebra and u ≥ 0 in Mµ¯1,∞ , then u∗ 4 u. P P Let hun in∈N be a non-decreasing sequence in S(A) with u0 ≥ 0 and supn∈N un = u. Then hu∗n in∈N is a non-decreasing sequence in Mµ¯1,∞ with L supremum u∗ , by 373Db and 373Dh. Also u∗n 4 un 4 u for every n, by (b) and (c) of this proof. By 373N, u∗ 4 u. Q Q (e) If (A, µ ¯) is a measure algebra and u ≥ 0 in S(A), then u 4 u∗ . P P The argument is very similar to that of (c). Again, the result is trivial if u = 0; suppose that u > 0 and define αi , ai , m, βi as before. This Pm ˜ amd u ˜∗ = u∗ . Set time, set a0i = ai for i < m, a0m = supm≤j≤n aj , u ˜ = i=0 αi χa0i ; then u ≤ u hi (v) =

1 βi+1 −βi

R βi+1 βi

v

if i ≤ m, βi+1 < ∞ (that is, µ ¯ai < ∞) and v ∈ Mµ¯1,∞ ; and if µ ¯am = ∞, set L hm (v) = limk→F

1 k

Rk 0

v

for some non-principal ultrafilter F on N. As before, we have Rβ |hi (v)|¯ µa0i ≤ βii+1 |v|, whenever v ∈ L1µ¯L , i ≤ m, while |hi (v)| ≤ kvk∞ whenever v ∈ L∞ (AL ) and i ≤ m. So we can define Pm , and get T ∈ Tµ¯L ,¯µ by setting T v = i=0 hi (v)χa0i for every v ∈ Mµ¯1,∞ L u4u ˜ = T u ∗ 4 u∗ . Q Q (f ) If (A, µ ¯) is a measure algebra and u ≥ 0 in Mµ¯1,∞ , then u 4 u∗ . P P This time I seek to copy the ideas of (d); there is a new obstacle to circumvent, since (A, µ ¯) might not be localizable. Set α0 = inf{α : α ≥ 0, µ ¯[[u > α]] < ∞},

e = [[u > α0 ]].

−n

Then e = supn∈N [[u > α0 + 2 ]] is a countable supremum of elements of finite measure, so the principal ideal Ae , with its induced measure µ ¯e , is σ-finite. Now let hun in∈N be a non-decreasing sequence in S(A) with u0 ≥ 0 and supn∈N un = u; set u ˜ = u × χe and u ñ = un × χe, regarded as of S(Ae ), for each n. In this case u ñ 4 u ˜∗n 4 u∗ for every n. Because (Ae , µ ¯e ) is σ-finite, therefore localizable, 373N tells us that u ˜ 4 u∗ . ∗ Let S ∈ Tµ¯L ,¯µe be such that Su = u ˜. As in part (e), choose a non-principal ultrafilter F on N and set h(v) = limk→F

1 k

Rk 0

v

for v ∈ Mµ¯1,∞ . Now define T : Mµ¯1,∞ → Mµ¯1,∞ by setting L L T v = Sv + h(v)χ(1 \ e),

432


373O

here regarding Sv as a member of Mµ¯1,∞ . (I am taking it to be obvious that Mµ¯1,∞ can be identified with e {w × χe : w ∈ Mµ¯1,∞ }.) Then it is easy to see that T ∈ Tµ¯L ,¯µ . Also u ≤ T u∗ , because h(u∗ ) = inf{α : µ ¯L [[u∗ > α]] < ∞} = α0 , while u × χ(1 \ e) ≤ α0 χ(1 \ e). So we get u 4 T u∗ 4 u∗ . Q Q R Rt t (g) Now suppose that u, v ≥ 0 in Mµ¯1,∞ , that 0 u∗ ≥ 0 v ∗ for every t ≥ 0, and that v is of the form L Pn ¯L ai < ∞ for each i. Then v 4 u. i=1 αi χai where α1 > . . . > αn > 0, a1 , . . . , an ∈ AL are dist and µ P P Induce on n. If n = 0 then v = 0 and the result is trivial. For the inductive step to n ≥ 1, if v ∗ ≤ u∗ we have v ∼ v ∗ 4 u∗ ∼ u, Rt using (b), (d) and (f) above. Otherwise, look at φ(t) = 1t 0 u∗ for t > 0. We have φ(t) ≥

1 t

Rt 0

v ∗ = α1

for t ≤ β = µ ¯a1 , while limt→∞ φ(t) < α1 , because (limt→∞ φ(t))χ1 ≤ u∗ and v ∗ ≤ α1 χ1 and v ∗ 6≤ u∗ . Becaasue φ is continuoue, there is a γ ≥ β such that φ(γ) = α1 . Define T0 ∈ Tµ¯L ,¯µL by setting T0 w = (

1 γ

Rγ 0

•

. Then T0 u∗ 4 u∗ ∼ u, and for every w ∈ Mµ¯1,∞ L T0 u∗ × χ [0, γ[ = ( •

We need to know that Z

t

0

Rt 0

T0 u∗ ≥

Rt 0

•

w)χ [0, γ[ + (w × χ [γ, ∞[ )

1 γ

Rγ 0

u∗ )χ [0, γ[ = α1 χ [0, γ[ . •

•

v ∗ for every t; this is because

Z t T0 u∗ = α1 t ≥ v ∗ whenever t ≤ γ, 0 Z γ Z t Z t Z t = T0 u∗ + T0 u∗ = u∗ ≥ v ∗ whenever t ≥ γ. 0

γ

0

0

Set u1 = T0 u∗ × χ [β, ∞[ , •

v1 = v ∗ × χ [β, ∞[ . •

Then u∗1 , v1∗ are just translations of T0 u∗ , v ∗ to the left, so that

Rt

R β+t

R β+t

R β+t

R β+t

Rt

T0 u∗ = 0 T0 u∗ − α1 β ≥ 0 v ∗ − α1 β = β v ∗ = 0 v1∗ Pn Pi • ¯aj for each j. So by the inductive for every t ≥ 0. Also v1 = j=1 µ i=2 αi χ [βi−1 , βi [ where βi = hypothesis, v1 4 u1 . Let S ∈ Tµ¯L ,¯µL be such that Su1 = v1 , and define T ∈ Tµ¯L ,¯µL by setting 0

u∗1 =

β

•

•

•

T w = w × χ [0, β[ + S(w × χ [β, ∞[ ) × χ [β, ∞[

∗ ∗ ∗ ∗ for every w ∈ Mµ¯1,∞ Q µL . Then T T0 u = v , so v ∼ v 4 u ∼ u, as required. Q L ,¯ R R t t (h) We are nearly home. If u, v ≥ 0 in Mµ¯1,∞ and 0 v ∗ ≤ 0 u∗ for every t ≥ 0, then v 4 u. P P Let L f + hvn in∈N be a non-decreasing sequence in S(AL ) with supremum v. Then vn∗ ≤ v ∗ for each n, so (g) tells us that vn 4 u for every n. By 373N, for the last time, v 4 u. Q Q

(i) Finally, suppose that (A, µ ¯) and (B, ν¯) are arbitrary measure algebras and that u ∈ Mµ¯1,∞ , v ∈ Mν¯1,∞ Rt ∗ Rt ∗ are such that 0 v ≤ 0 u for every t ≥ 0. By (b), v 4 |v|; by (f), |v| 4 |v|∗ ; by 373Db, |v|∗ = v ∗ ; by (h) of this proof, v ∗ 4 u∗ ; by (d), u∗ = |u|∗ 4 |u|; and by (b) again, |u| 4 u. 373P Theorem Let (A, µ ¯) be a measure algebra and (B,Rν¯) a semi-finite algebra. Then for R ∗ measure 1,∞ 0 ∗ any u ∈ Mµ¯ and v ∈ Mν¯ , there is a T ∈ T = Tµ¯,¯ν such that T u × v = u × v .

373P


433

proof (a) It is convenient to dispose immediately of some elementary questions. R R (i) We need only find a T ∈ T such that |T u × v| ≥ u∗ × v ∗ . P P Take v0 ∈ L∞ (B) such that 1,∞ |T u × v| = v0 × T u × v and kv0 k∞ ≤ 1, and set T1 w = v0 × T w for w ∈ Mµ¯ ; then T1 ∈ T and

R

T1 u × v =

R

R

|T u × v| ≥

u∗ × v ∗ ≥

R

T1 u × v

by 373J. Q Q (ii) Consequently it will be enough to consider v ≥ 0, since of course |v|∗ = v ∗ .

R

|T u × v| =

R

|T u × |v||, while

R R (iii) It will be enough to consider u = u∗ . P P If we can find T ∈ Tµ¯L ,¯ν such that T u∗ ×v = (u∗ )∗ ×v ∗ , then we know from 373O that there is an S ∈ Tµ¯,¯µL such that Su = u∗ , so that T S ∈ T and

R

T Su × v =

R

(u∗ )∗ × v ∗ =

R

u∗ × v ∗ . Q Q

(iv) It will be enough to consider localizable (B, ν¯). P P Assuming that v ≥ 0, following (ii) above, set e = [[v > 0]] = supn∈N [[v > 2−n ]], and let ν¯e be the restriction of ν¯ to the principal ideal Be generated by e. Then if we write v˜ for the member of L0 (Be ) corresponding to v (so that [[˜ v > α]] = [[v > α]] for every ∗ ∗ α > 0), v ˜ = v . Also (B , ν ¯ ) is σ-finite, therefore localizable. Now if we can find T ∈ Tµ¯,¯νe such that e e R R T u × v˜ = u∗ × v˜∗ , then ST will belong to Tµ¯,¯ν , where S : L0 (Be ) → L0 (B) is the canonical embedding defined by the formula [[Sw > α]] = [[w > α]] if α ≥ 0, = [[w > α]] ∪ (1 \ e) if α < 0, and

R

ST u × v =

R

T u × v˜ =

R

u∗ × v˜∗ =

R

u∗ × v ∗ . Q Q

(b) So let us suppose henceforth that µ ¯ = µ ¯L , u = u∗ is the equivalence class of a non-increasing non-negative function, v ≥ 0 and (B, ν¯) is localizable. For n, i ∈ N set bni = [[v > 2−n i]],

βni = ν¯bni ,

cni = bni \ bn,i+1 ,

γni = ν¯cni = βni − βn,i+1

(because βni < ∞ if i > 0; this is really where I use the hypothesis that v ∈ M 0 ). For n ∈ N set Kn = {i : i ≥ 1, γni > 0}, Tn w =

X ¡ 1 Z βni ¢ w χcni γni βn,i+1

i∈Kn

for w ∈ Mµ¯1,∞ ; this is defined in L0 (B) because Kn is countable and hcni ii∈N is dist. Of course Tn : L 1,∞ 0 Mµ¯L → L (B) is linear. If w ∈ L∞ (AL ) then

¯

kTn wk∞ = supi∈Kn ¯ and if w ∈ L1µ¯L then kTn wk1 =

P

¯ 1 R βni ¯

i∈Kn γ ni

1 γni

¯

βn,i+1

¯

R βni

βn,i+1

w¯ν¯cni =

w¯ ≤ kwk∞ ,

P i∈Kn

¯R βni ¯

βn,i+1

¯

w¯ ≤ kwk1 ;

so Tn w ∈ Mν¯1,∞ for every w ∈ Mµ¯1,∞ , and Tn ∈ T . It will be helpful to observe that L

R

cni

Tn w =

R βni

βn,i+1

w

whenever i ≥ 1, since if i ∈ / Kn then both sides are 0. Note next that for every n, i ∈ N, bni = bn+1,2i ,

βni = βn+1,2i ,

cni = cn+1,2i ∪ cn+1,2i+1 ,

γni = γn+1,2i + γn+1,2i+1 ,

434


so that, for i ≥ 1,

R cni

R βni

Tn u =

βn,i+1

u=

R

373P

Tn+1 u.

cni

This means that if T is any cluster R point of hTn in∈N in T for the very R weak operator topology (and such a cluster point exists, by 373L), cmi T u must be a cluster point of h cmi Tn uin∈N , and therefore equal to R T u, for every m ∈ N, i ≥ 1. cmi m Consequently, if m ∈ N, Z |T u × v| ≥

∞ Z X

|T u| × v ≥ cmi

i=0

=

∞ X i=1 ∞ X

Z 2−m i|

T u| = cmi

Z −m

2

Z −m

2

i

∞ X

Z 2−m i

Tm u cmi

i=1

Z

βmi

i

|T u| cmi

i=0

(because cmi ⊆ [[v > 2−m i]]) ≥

∞ X

u≥

u × (v ∗ − 2−m χ1)+

βm,i+1

i=0

because [βm,i+1 , βmi ]• ⊆ [[v ∗ ≤ 2−m (i + 1)]] = [[(v ∗ − 2−m χ1)+ ≤ 2−m i]] for each i ∈ N. But letting m → ∞, we have

R

|T u × v| ≥ limm→∞

R

u × (v ∗ − 2−m χ1)+ =

R

u × v∗

because hu × (v ∗ − 2−m χ1)+ im∈N is a non-decreasing sequence with supremum u × v ∗ . In view of the reductions in (a) above, this is enough to complete the proof. 373Q Corollary Let (A, µ ¯) be a measure algebra, (B, ν¯) a semi-finite measure algebra, u ∈ Mµ¯1,∞ and 0,∞ v ∈ Mν¯ . Then

R

R

R

u∗ × v ∗ = sup{ |T u × v| : T ∈ Tµ¯,¯ν } = sup{ T u × v : T ∈ Tµ¯,¯ν }.

proof There is a non-decreasing sequence hcn in∈N in B such that ν¯cn < ∞ for every n and v ∗ = supn∈N (v × χcn )∗ . P P For each rational q > 0, we can find a countable non-empty set Bq ⊆ B such that b ⊆ [[|v| > q]], ν¯b < ∞ for every b ∈ Bq , supb∈Bq ν¯b = ν¯[[|v| > q]]

S (because (B, ν¯) is semi-finite). Let hbn in∈N be a sequence running over q∈Q,q>0 Bq and set cn = supi≤n bi , vn = v × χcn for each n. Then h|vn |in∈N and hvn∗ in∈N are non-decreasing and supn∈N vn∗ ≤ v ∗ in L0 (AL ). But in fact supn∈N vn∗ = v ∗ , because µ ¯L [[v ∗ > q]] = µ ¯[[|v| > q]] = supn∈N µ ¯[[vn > q]] = supn∈N µ ¯L [[vn∗ > q]] = µ ¯L [[supn∈N vn∗ > q]] for every rational q > 0, by 373Da. Q Q R R For each n ∈ N we have a Tn ∈ Tµ¯,¯ν such that Tn u × vn = u∗ × vn∗ (373P). Set Sn w = Tn w × χcn for n ∈ N, w ∈ Mµ¯1,∞ ; then every Sn belongs to Tµ¯,¯ν , so Z Z Z sup{ T u × v : T ∈ Tµ¯,¯µ } ≥ sup Sn u × v = sup Tn u × vn n∈N n∈N Z Z ∗ ∗ = sup u × vn = u∗ × v ∗ n∈N Z Z ≥ sup{ |T u × v| : T ∈ Tµ¯,¯µ } ≥ sup{ T u × v : T ∈ Tµ¯,¯µ } by 373J, as usual.

373S


435

373R Order-continuous operators: Proposition Let (A, µ ¯) be a measure algebra, (B, ν¯) a local(0) izable measure algebra, and T0 ∈ T (0) = Tµ¯,¯ν . Then there is a T ∈ T × = Tµ¯×,¯ν extending T0 . If (A, µ ¯) is semi-finite, T is uniquely defined. proof (a) Suppose first that T0 ∈ T (0) is non-negative, regarded as a member of L∼ (Mµ¯1,0 ; Mν¯1,0 ). In this case T0 has an extension to an order-continuous positive linear operator T : Mµ¯1,∞ → L0 (B) defined by saying that T w = sup{T0 u : u ∈ Mµ¯1,0 , 0 ≤ u ≤ w} for every w ≥ 0 in Mµ¯1,∞ . P P I use 355F. Mµ¯1,0 is a solid 1,∞ linear subspace of Mµ¯ . T0 is order-continuous when its codomain is taken to be Mν¯1,0 , as noted in 371Gb, and therefore if its codomain is taken to be L0 (B), because M 1,0 is a solid linear subspace in L0 , so the embedding is order-continuous. If w ≥ 0 in Mµ¯1,∞ , let γ ≥ 0 be such that u1 = (w − γχ1)+ is integrable. If u ∈ Mµ¯1,0 and 0 ≤ u ≤ w, then (u − γχ1)+ ≤ u1 , so T0 u = T0 (u − γχ1)+ + T0 (u ∧ γχ1) ≤ T0 u1 + γχ1 ∈ L0 (B). Thus {T0 u : u ∈ Mν¯1,0 , 0 ≤ u ≤ w} is bounded above in L0 (B), for any w ≥ 0 in Mµ¯1,∞ . L0 (B) is Dedekind complete, because (B, ν¯) is localizable, so sup{T0 u : 0 ≤ u ≤ w} is defined in L0 (B); and this is true for every w ∈ (Mµ¯1,∞ )+ . Thus the conditions of 355F are satisfied and we have the result. Q Q (b) Now suppose that T0 is any member of T (0) . Then T0 has an extension to a member of T × . P P |T0 |, = 21 (|T0 | + T0 ) and T0− = 12 (|T0 | − T0 ), taken in L∼ (Mµ¯1,0 ; Mν¯1,0 ), all belong to T (0) (371G), so have extensions S, S1 and S2 to order-continuous positive linear operators from Mµ¯1,∞ to L0 (B) as defined in (a). Now for any w ∈ L1µ¯ , T0+

kSwk1 = k|T0 |wk1 ≤ kwk1 , ∞

and for any w ∈ L (A), |Sw| ≤ S|w| = sup{|T0 |u : u ∈ Mµ¯1,0 , 0 ≤ u ≤ w} ≤ kwk∞ χ1, so kSwk∞ ≤ kwk∞ . Thus S ∈ T ; similarly, S1 and S2 can be regarded as operators from Mµ¯1,∞ to Mν¯1,∞ , and as such belong to T . Next, for w ≥ 0 in Mµ¯1,∞ , S1 w + S2 w = sup{T0+ u : u ∈ Mµ¯1,0 , 0 ≤ u ≤ w} + sup{T0− u : u ∈ Mµ¯1,0 , 0 ≤ u ≤ w} = sup{T0+ u + T0− u : u ∈ Mµ¯1,0 , 0 ≤ u ≤ w} = Sw. But this means that S = S1 + S2 ≥ |S1 − S2 | and T = S1 − S2 ∈ T , by 373Bc; while of course T extends T0+ − T0− = T0 . Finally, because S1 and S2 are Q order-continuous, T ∈ L× (Mµ¯1,∞ ; Mν¯1,∞ ), so T ∈ T × . Q (c) If (A, µ ¯) is semi-finite, then Mµ¯1,0 is order-dense in Mµ¯1,∞ (because it includes L1µ¯ , which is order-dense in L0 (A)); so that the extension T is unique, by 355F(iii). 373S Adts in T (0) : Theorem Let (A, µ ¯) and (B, ν¯) be measure algebras, and T any member of R R (0) Then there is a unique operator T 0 ∈ Tν¯,¯µ such that a T 0 (χb) = b T (χa) for every a ∈ Af , b ∈ Bf , R R R and now u × T 0 v = T u × v whenever u ∈ Mµ¯1,0 , v ∈ Mν¯1,0 are such that u∗ × v ∗ < ∞. (0) Tµ¯,¯ν .

proof (a) For each v ∈ Mν¯1,0 we can define T 0 v ∈ Mµ¯1,0 by the formula

R

T 0v = a

R

R

T (χa) × v

R for every a ∈ Af . P P Set θa = T (χa) × v for each a ∈ Af ; because (χa)∗ × v ∗ < ∞, the R integral is defined and finite (373J). Of course θ : Af → R is additive because χ is additive and T , × and are linear. Also limt↓0 supµ¯a≤t |θa| ≤ limt↓0 1 t

Rt 0

limt→∞ supµ¯a≤t |θa| ≤ limt→∞

v ∗ = 0,

1 t

Rt 0

v∗ = 0

436


373S

because v ∈ Mν¯1,0 , so v ∗ ∈ Mµ¯1,0 . By 373Ha, there is a unique T 0 v ∈ Mµ¯1,0 such that L f a∈A .Q Q

R a

T 0 v = θa for every (0)

(b) Because the formula uniquely determines T 0 v, we see that T 0 : Mν¯1,0 → Mµ¯1,0 is linear. Now T 0 ∈ Tν¯,¯µ . P P (i) If v ∈ L1ν¯ , then (because T 0 v ∈ Mµ¯1,0 ) |T 0 v| = supa∈Af |T 0 v| × χa, and Z Z Z Z kT 0 vk1 = |T 0 v| = sup |T 0 v| = sup ( T 0 v − T 0 v) a∈Af

Z

b,c∈Af

a

c

(χb − χc)∗ × v ∗

T (χb − χc) × v ≤ sup

= sup

b,c∈Af

b,c∈Af

Z =

b

Z

v ∗ = kvk1 .

(ii) Now suppose that v ∈ L∞ (B) ∩ Mν¯1,0 , and set γ = kvk∞ . ?? If a = [[|T 0 v| > γ]] 6= 0, then T 0 v 6= 0 so v 6= 0 and γ > 0 and µ ¯a < ∞, because T 0 v ∈ Mµ¯1,0 . Set b = [[(T 0 v)+ > γ]], c = [[(T 0 v)− > γ]]; then Z Z Z Z γµ ¯a < |T 0 v| = T 0 v − T 0 v = T (χb − χc) × v a

b

c

≤ γkT (χb − χc)k1 ≤ γkχb − χck1 = γ µ ¯a, which is impossible. X X Thus [[|T 0 v| > γ]] = 0 and kT 0 vk∞ ≤ γ = kvk∞ . (0) Q Putting this together with (i), we see that T 0 ∈ Tν¯,¯µ . Q (0)

(c) Let |T | be the modulus of T in L∼ (Mµ¯1,0 ; Mν¯1,0 ), so that |T | ∈ Tµ¯,¯µ , by 371Gb. If u ≥ 0 in Mµ¯1,0 , v ≥ 0 R sequence in S(Af )+ with supremum in Mν¯1,0 are such that u∗ × v ∗ < ∞, let R hun in∈N be a non-decreasing R u. In this case |T |u = supn∈N |T |un , so |T |u × v = supn∈N |T |un × v and |

R

as n → ∞, because

R

At the same time, because

R

Tu × v −

u × |T 0 v| ≤

R

R

|

R

R

T un × v| ≤

|T |u × v ≤

u × T 0v −

R

R

R

|T |(u − un ) × v → 0

u∗ × v ∗ < ∞.

un × T 0 v| ≤

R

(u − un ) × |T 0 v| → 0

u∗ × v ∗ < ∞. So

T u × v = limn→∞

R

T un × v = limn→∞

R

un × T 0 v =

R

u × T 0 v,

the middle equality being valid because each un is a linear of characteristic functions. R combination R Because T Rand T 0 are linear, it follows at once that u × T 0 v = T u × v whenever u ∈ Mµ¯1,0 , v ∈ Mν¯1,0 are such that u∗ × v ∗ < ∞. (d) Finally, to see that T 0 is uniquely defined by the formula in the statement of the theorem, observe that this surely defines T 0 (χb) for every b ∈ Bf , by the remarks in (a). Consequently it defines T 0 on S(Bf ). (0) Since S(Bf ) is order-dense in Mν¯1,0 , and any member of Tν¯,¯µ must belong to L× (Mν¯1,0 ; Mµ¯1,0 ) (371Gb), the restriction of T 0 to S(Bf ) determines T 0 (355J). 373T Corollary Let (A, µ ¯) and (B, ν¯) be localizable measure algebras. Then for any T ∈ Tµ¯×,¯ν there is a R R R unique T 0 ∈ Tν¯×,¯µ such that u×T 0 v = T u×v whenever u ∈ Mµ¯1,∞ , v ∈ Mν¯1,∞ are such that u∗ ×v ∗ < ∞. R (0) (0) proof The restriction T ¹Mµ¯1,0 belongs to Tµ¯,¯ν (373Bb), so there is a unique S ∈ Tν¯,¯µ such that u × Sv = R R T u × v whenever u ∈ Mµ¯1,0 , v ∈ Mν¯1,0 are such that u∗ × v ∗ < ∞ (373S). Now there is a unique T 0 ∈ Tν¯×,¯µ R R R extending S (373R). If u ≥ 0 in Mµ¯1,∞ , v ≥ 0 in Mν¯1,∞ are such that u∗ ×v ∗ < ∞, then u×T 0 v = T u×v. P P If T ≥ 0, then both are

R

sup{ u0 × T 0 v0 : u0 ∈ Mµ¯1,0 , v ∈ Mν¯1,0 , 0 ≤ u0 ≤ u, 0 ≤ v0 ≤ v}

373Xi


437

because both T and T 0 are (order-)continuous. In general, we can apply the same argument to T + and T − , taken in L∼ (Mµ¯1,∞ ; Mν¯1,∞ ), since these belong to Tµ¯×,¯ν , by 373B and 355H, and we shall surely have R R T 0 = (T + )R0 − (T − )0 . Q Q As in 373S, it follows that u × T 0 v = T u × v whenever u ∈ Mµ¯1,∞ , v ∈ Mν¯1,∞ are such that u∗ × v ∗ < ∞. 373U Corollary Let (A, µ ¯) and (B, ν¯) be localizable measure algebras, and π : A → B an ordercontinuous measure-preserving Boolean homomorphism. Then the associated map T ∈ Tµ¯×,¯ν (373Bd) has an R adt P ∈ Tν¯×,¯µ defined by the formula a P (χb) = ν¯(b ∩ πa) for a ∈ Af , b ∈ Bf . proof The adt P = T 0 must have the property that

R

a

P (χb) =

R

χa × P (χb) =

R

T (χa) × χb =

R

χ(πa) × χb = ν¯(πa ∩ b)

for every a ∈ Af , b ∈ Bf . To see that this defines P uniquely, let S ∈ Tν¯×,¯µ be any other operator with the same property. By 373Hb, S(χb) = P (χb) for every b ∈ Bf , so S and P agree on S(Bf ). Because both P and S are supposed to belong to L× (Mν¯1,∞ ; Mµ¯1,∞ ), and S(Bf ) is order-dense in Mν¯1,∞ , S = P , by 355J. 373X Basic exercises (a) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a ring homomorphism such that ν¯πa ≤ µ ¯a for every a ∈ A. (i) Show that there is a unique T ∈ Tµ¯,¯ν such that T (χa) = χ(πa) for every a ∈ A, and that T is a Riesz homomorphism. (ii) Show that T is (sequentially) order-continuous iff π is. > (b) Let (A, µ ¯) and (B, µ ¯) be measure algebras, and φ : R → R a convex function such that φ(0) ≤ 0. ¯ u) ≤ T (φ(u)) ¯ ¯ Show that if T ∈ Tµ¯,¯ν and T ≥ 0, then φ(T whenever u ∈ Mµ¯1,∞ is such that φ(u) ∈ Mµ¯1,∞ . (Hint: 233J, 365Rb.) (c) Let (A, µ ¯) be a measure algebra. Show that if w ∈ L∞ (A) and kwk∞ ≤ 1 then u 7→ u × w : Mµ¯1,∞ → Mµ¯1,∞ belongs to Tµ¯×,¯µ . (d) Let (A, µ ¯) and (B, ν¯) be measure algebras. Show that if hai ii∈I , hbi ii∈I are dist families in A, B respectively, and hTi ii∈I is any family in Tµ¯,¯ν , and either I is countable or B is Dedekind complete, then we have an operator T ∈ Tµ¯,¯ν such that T u × χbi = Ti (u × χai ) × χbi for every u ∈ Mµ¯1,∞ ,¯ ν , i ∈ I. > (e) Let I, J be sets and write µ = µ ¯, ν = ν¯ for counting measure on I, J respectively. Show that there is a natural one-to-one correspondence between Tµ¯×,¯ν and the set of matrices haij ii∈I,j∈J such that P P j∈J |aij | ≤ 1 for every i ∈ I. i∈I |aij | ≤ 1 for every j ∈ J, >(f ) Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces, with measure algebras (A,R µ ¯) and (B, ν¯), and product measure λ on X ×RY . Let h : X × Y → R be a measurable function such that |h(x, y)|dx ≤ 1 for ν-almost every y ∈ Y and |h(x, y)|dy ≤ 1 for µ-almost every x ∈ X. Show that R there is a corresponding T ∈ Tµ¯×,¯ν defined by writing T (f • ) = g • whenever f ∈ L1 (µ) + L∞ (µ) and g(y) = h(x, y)f (x)dx for almost every y. > (g) Let µ beR Lebesgue measure on R, and (A, µ ¯) its measure algebra. Show that for any µ-integrable function h with |h|dµ ≤ 1 we have a corresponding T ∈ Tµ¯×,¯µ defined by setting T (f • ) = (h ∗ f )• whenever g ∈ L1 (µ) + L∞ (µ), writing h ∗ f for the convolution of h and f (255E). Explain how this may be regarded as a special case of 373Xf. > (h) Let (Ω, Σ, µ) be a probability space and X ∈ L0 (µ) a non-negative real-valued random variable on Ω; let νX be its distribution (271C). Write u = X • ∈ L0 (µ) ∼ ¯) is the measure algebra of = L0µ¯ , where (A, µ ∗ ∗ (Ω, Σ, µ). Show that µ ¯L [[u > α]] = Pr(X > α) for every α, so that each of u , νX is uniquely determined by the other. (i) Let (A, µ ¯) and (B, ν¯) be measure algebras, and π : A → B a measure-preserving Boolean homomorphism; let T : Mµ¯1,∞ → Mν¯1,∞ be the corresponding operator (373Bd). Show that (T u)∗ = u∗ for every u ∈ Mµ¯1,∞ .

438


373Xj

(j) Let (A, µ ¯) be a totally finite measure algebra, and A a subset of L1µ¯ . Show that the following are equiveridical: (i) A is uniformly integrable; (ii) {u∗ : u ∈ A} is uniformly integrable in L1µ¯L ; (iii) Rt limt↓0 supu∈A 0 u∗ = 0. (l) Let (A, µ ¯) be a measure algebra, and A ⊆ (Mµ¯0 )+ a non-empty downwards-directed set. Show that ∗ (inf A) = inf u∈A u∗ in L0 (AL ). (k) Let (A, µ ¯) be a measure algebra. Show that kuk1,∞ =

R1 0

u∗ for every u ∈ M 1,∞ (A, µ ¯).

(m) Let (A, µ ¯) and (B, ν¯) be measure algebras, and φ a Young’s function (369Xc). Write Uφ,¯µ ⊆ L0 (A), 0 Uφ,¯ν ⊆ L (B) for the corresponding Orlicz spaces. (i) Show that if T ∈ Tµ¯,¯ν and u ∈ Uφ,¯µ , then T u ∈ Uφ,¯ν and kT ukφ ≤ kukφ . (ii) Show that u ∈ Uφ,¯µ iff u∗ ∈ Uφ,¯µL , and in this case kukφ = ku∗ kφ . > (n) Let (A, µ ¯) be a measure algebra and (B, ν¯) a totally finite measure algebra. Show that if A ⊆ L1µ¯ is uniformly integrable, then {T u : u ∈ A, T ∈ Tµ¯,¯ν } is uniformly integrable in L1ν¯ . (o) (i) Give examples of u, v ∈ L1 (AL ) such that (u + v)∗ 6≤ u∗ + v ∗ . (ii) Show that if (A, µ ¯) is any Rt Rt measure algebra and u, v ∈ Mµ¯0,∞ , then 0 (u + v)∗ ≤ 0 u∗ + v ∗ for every t ≥ 0. (p) Let (A, µ ¯) and (B, ν¯) be two measure algebras. For u ∈ Mµ¯1,0 , w ∈ Mν¯∞,1 set

R

(0)

ρuw (S, T ) = | (Su − T u) × w| for all S, T ∈ T (0) = Tµ¯,¯ν . The topology generated by the pseudometrics ρuw is the very weak operator topology on T (0) . Show that T (0) is compact in this topology. (0)

(q) Let (A, µ ¯) and (B, ν¯) be measure algebras and let u ∈ Mµ¯1,0 . (i) Show that B = {T u : T ∈ Tµ¯,¯ν } is compact for the topology Ts (Mν¯1,0 , Mν¯∞,1 ). (ii) Show that any non-decreasing sequence in B has a supremum in L0 (B) which belongs to B. (r) Let (A, µ ¯) and (B, ν¯) be measure algebras, and u ∈ Mµ¯1,0 , v ∈ Mν¯1,0 . Show that the following are Rt Rt (0) equiveridical: (i) there is a T ∈ Tµ¯,¯ν such that T u = v; (ii) 0 u∗ ≤ 0 v ∗ for every t ≥ 0. (s) Let (A, µ ¯) and (B, ν¯) be measure algebras. Suppose that u1 , u2 ∈ Mµ¯1,∞ and v ∈ Mν¯1,∞ are such Rt ∗ Rt that 0 v ≤ 0 (u1 + u2 )∗ for every t ≥ 0. Show that there are v1 , v2 ∈ Mν¯1,∞ such that v1 + v2 = v and Rt ∗ Rt ∗ v ≤ 0 ui for both i, every t ≥ 0. 0 i • • ∞ R > (t) Set g(t) = t/(t + 1) for t ≥ 0, and set v = g , u = χ[0, 1] ∈ L (AL ). Show that T u × v for every T ∈ Tµ¯L ,¯µL .

(0)

R

u∗ × v ∗ = 1 >

(0)

(u) Let (A, µ ¯) and (B, ν¯) be measure algebras, and for T ∈ Tµ¯,¯ν define T 0 ∈ Tν¯,¯µ as in 373S. Show that T = T. 00

(0)

(0)

(v) Let (A, µ ¯) and (B, ν¯) be measure algebras, and give Tµ¯,¯ν , Tν¯,¯µ their very weak operator topologies (0) (0) (373Xp). Show that the map T 7→ T 0 : Tµ¯,¯ν → Tν¯,¯µ is an isomorphism for the convex, order and topological structures of the two spaces. (By the ‘convex structure’ of a convex set C in a linear space I mean the operation (x, y, t) 7→ tx + (1 − t)y : C × C × [0, 1] → C.) 373Y Further exercises (a) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1]. Set u = f • and v = g • in L0 (A), where f (t) = t, g(t) = 1 − 2|t − 21 | for t ∈ [0, 1]. Show that u∗ = v ∗ , but that there is no measure-preserving Boolean homomorphism π : A → A such that Tπ v = u, writing Tπ : L0 (A) → L0 (A) for the operator induced by π, as in 364R. (Hint: show that {[[v > α]] : α ∈ R} does not τ -generate A.)

373 Notes


439

(b) Let (A, µ ¯) be a totally finite homogeneous measure algebra of uncountable Maharam type. Let u, v ∈ (Mµ¯1,∞ )+ be such that u∗ = v ∗ . Show that there is a measure-preserving automorphism π : A → A such that Tπ u = v. Rt Rt (c) Let u, v ∈ Mµ¯1,∞ be such that u = u∗ , v = v ∗ and 0 v ≤ 0 u for every t ≥ 0. Show that there is L Rt Rt a non-negative T ∈ Tµ¯L ,¯µL such that T u = v and 0 T w ≤ 0 w for every w ≥ 0 in M 1,∞ . Show that any such T must belong to Tµ¯×L ,¯µL . (d) Let (A, µ ¯) and (B, ν¯) be measure algebras, and u ∈ Mµ¯1,∞ . (i) Suppose that w ∈ S(Bf ). Show directly (without quoting the result of 373O, but R R possibly using some of the ideas of the proof) that for every γ < u∗ ×w∗ there is a T ∈ Tµ¯,¯ν such that T u×w ≥ γ. (ii) Suppose that (B, ν¯) is localizable and that R R v ∈ Mν¯1,∞ \ {T u : T ∈ Tµ¯,¯ν }. Show that there is a w ∈ S(Bf ) such that v × w > sup{ T u × w : T ∈ Tµ¯,¯ν }. (Hint: use 373M and the Hahn-Banach theorem in the following form: if U is a linear space with the topology Ts (U, V ) defined by a linear subspace V of L(U ; R), C ⊆ U is a non-empty closed convex set, and v ∈ U \ C, then there is an f ∈ V such that f (v) > supu∈C f (u).) (iii) Hence prove 373O for localizable (B, ν¯). (iv) Now prove 373O for general (B, ν¯). (e) (i) Define v ∈ L∞ (AL ) as in 373Xt. Show that there is no T ∈ Tµ¯×L ,¯µL such that T v = v ∗ . (ii) Set h(t) = 1 + max(0, sint t ) for t > 0, w = h• ∈ L∞ (AL ). Show that there is no T ∈ Tµ¯×L ,¯µL such that T w∗ = w. (f ) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1]. Show that Tµ¯,¯µL = Tµ¯×,¯µL can be identified, as convex ordered space, with Tµ¯×L ,¯µ , and that this is a proper subset of Tµ¯L ,¯µ . (g) Show that the adt operation of 373T is not as a rule continuous for the very weak operator topologies of Tµ¯×,¯ν , Tν¯×,¯µ . 373 Notes and comments 373A-373B are just alternative expressions of concepts already treated in 371F371H. My use of the simpler formula Tµ¯,¯ν symbolizes my view that T , rather than T (0) or T × , is the most natural vehicle for these ideas; I used T (0) in §§371-372 only because that made it possible to give theorems which applied to all measure algebras, without demanding localizability (compare 371Gb with 373Bc). The obvious examples of operators in T are those derived from measure-preserving Boolean homomorphisms, as in 373Bd, and their adts (373U). Note that the latter include conditional expectation operators. In return, we find that operators in T share some of the characteristic properties of the operators derived from Boolean homomorphisms (373Bb, 373Xb, 373Xm). Other examples are multiplication operators (373Xc), operators obtained by piecing others together (373Xd) and kernel operators of the type described in 373Xe-373Xf, including convolution operators (373Xg). (For a general theory of kernel operators, see §376 below.) Most of the section is devoted to the relationships between the classes T of operators and the ‘decreasing rearrangements’ of 373C. If you like, the decreasing rearrangement u∗ of u describes the ‘distribution’ of |u| (373Xh); but for u ∈ / M 0 it loses some information (373Xt, 373Ye). It is important to be conscious that 0 even when u ∈ L (AL ), u∗ is not necessarily obtained by ‘rearranging’ the elements of the algebra AL by a measure-preserving automorphism (which would, of course, correspond to an automorphism of the measure space ([0, ∞[ , µL ), by 344C). I will treat ‘rearrangements’ of this narrower type in the next section; for the moment, see 373Ya. Apart from this, the basic properties of decreasing rearrangements are straightforward enough (373D-373F). The only obscure area concerns the relationship between (u + v)∗ and u∗ , v ∗ (see 373Xo). In 373G I embark on results involving both decreasing rearrangements and operators in T , leading to the characterization of the sets {T u : T ∈ T } in 373O. In one direction this is easy, and is the content of 373G. In the other direction it depends on a deeper analysis, and the easiest method seems to be through studying the ‘very weak operator topology’ on T (373K-373L), even though this is an effective tool only when one of the algebras involved is localizable (373L). A functional analyst is likely to feel that the method is both natural and illuminating; but from the point of view of a measure theorist it is not perfectly satisfactory, because it is essentially non-constructive. While it tells us that there are operators T ∈ T acting in the required ways, it gives only the vaguest of hints concerning what they actually look like.

440


373 Notes

Of course the very weak operator topology is interesting in its own right; and see R t also 373Xp-373Xq. Rt The proof of 373O can be thought of as consisting of three steps. Given that 0 v ∗ ≤ 0 u∗ for every t, then I set out to show that v is expressible as T1 v ∗ (parts (c)-(d) of the proof), that v ∗ is expressible as T2 u∗ (part (g)) and that u∗ is expressible as T3 u (parts (e)-(f)), each Ti belonging to an appropriate T . In all three steps the general case follows easily from the case of u, v ∈ S(A), S(B). If we are willing to use a more sophisticated version of the Hahn-Banach theorem than those given in 3A5A and 363R, there is an alternative route (373Yd). I note that the central step above, from u∗ to v ∗ , can be performed with an order-continuous T2 (373Yc), but that in general neither of the other steps can (373Ye), so that we cannot use T × in place of T here. A companion result to 373O, in that it also shows thatR {T u : T ∈ T } is large enough to reach natural bounds, is 373P; given u and v, we can find T such that T u × v is as large as possible. In this form the result is valid only for v ∈ M (0) (373Xt). But if we do not demand that the supremum should be attained, we can deal with other v (373Q). We already know that every operator in T (0) is a difference of order-continuous operators, just because M 1,0 has an order-continuous norm (371Gb). It is therefore not surprising that of T (0) can be extended to of T × , at least when the codomain Mν¯1,∞ is Dedekind complete (373R). It is also very natural to look for a correspondence between Tµ¯,¯ν and Tν¯,¯µ , because if T ∈ Tµ¯,¯ν we shall surely have an adt operator (T ¹L1µ¯ )0 from (L1ν¯ )∗ to (L1µ¯ )∗ , and we can hope that this will correspond to some member of Tν¯,¯µ . But when we come to the details, the normed-space properties of a general member of T are not enough (373Yf), and we need some kind of order-continuity. For of T (0) this is automatically present (373S), and now the canonical isomorphism between T (0) and T × gives us an isomorphism between Tµ¯×,¯ν and Tν¯×,¯µ when µ ¯ and ν¯ are localizable (373T).

374 Rearrangement-invariant spaces As is to be expected, many of the most important function spaces of analysis are symmetric in various ways; in particular, they share the symmetries of the underlying measure algebras. The natural expression of this is to say that they are ‘rearrangement-invariant’ (374E). In fact it turns out that in many cases they have the stronger property of ‘T -invariance’ (374A). In this section I give a brief of the most important properties of these two kinds of invariance. In particular, T -invariance is related to a kind of transfer mechanism, enabling us to associate function spaces on different measure algebras (374B-374D). As for rearrangement-invariance, the salient fact is that on the most important measure algebras many rearrangement-invariant spaces are T -invariant (374K, 374M). 374A T -invariance: Definitions Let (A, µ ¯) be a measure algebra. Recall that I write Mµ¯1,∞ = L1µ¯ + L∞ (A) ⊆ L0 (A), Mµ¯∞,1 = L1µ¯ ∩ L∞ (A), Mµ¯0,∞ = {u : u ∈ L0 (A), inf α>0 µ ¯[[|u| > α]] < ∞}, (369N, 373C). (a) I will say that a subset A of Mµ¯1,∞ is T -invariant if T u ∈ A whenever u ∈ A and T ∈ T = Tµ¯,¯µ (definition: 373Aa). (b) I will say that an extended Fatou norm τ on L0 is T -invariant or fully symmetric if τ (T u) ≤ τ (u) whenever u ∈ Mµ¯1,∞ and T ∈ T . (c) As in §373, I will write (AL , µ ¯L ) for the measure algebra of Lebesgue measure on [0, ∞[, and u∗ ∈ Mµ¯0,∞ L for the decreasing rearrangement of any u belonging to any Mµ¯0,∞ (373C).

374C

374B

Rearrangement-invariant spaces

441

The first step is to show that the associate of a T -invariant norm is T -invariant.

Theorem Let (A, µ ¯) be a semi-finite measure algebra and τ a T -invariant extended Fatou norm on L0 (A). τ Let L be the Banach lattice defined from τ (369G), and τ 0 the associate extended Fatou norm (369H-369I). Then (i) Mµ¯∞,1 ⊆ Lτ ⊆ Mµ¯1,∞ ; R (ii) τ 0 is also T -invariant, and u∗ × v ∗ ≤ τ (u)τ 0 (v) for all u, v ∈ Mµ¯0,∞ . P Take any u ∈ L0 (A) \ Mµ¯0,∞ . There is surely some w > 0 in Lτ , proof (a) I check first that Lτ ⊆ Mµ¯0,∞ . P and we can suppose that w = χa for some a of finite measure. Now, for any n ∈ N, (|u| ∧ nχ1)∗ = nχ1 ≥ nw∗ in L0 (AL ), because µ ¯[[|u| > n]] = ∞. So there is a T ∈ Tµ¯,¯µ such that T (|u| ∧ nχ1) = nw, by 373O, and τ (u) ≥ τ (|u| ∧ nχ1) ≥ τ (T (|u| ∧ nχ1)) = τ (nw) = nτ (w). As n is arbitrary, τ (u) = ∞. As u is arbitrary, Lτ ⊆ Mµ¯0,∞ . Q Q R ∗ (b) Next, u × v ∗ ≤ τ (u)τ 0 (v) for every u, v ∈ Mµ¯0,∞ . P P If u ∈ Mµ¯1,∞ , then Z

Z u∗ × v ∗ = sup{

|T u × v| : T ∈ Tµ¯,¯µ }

(373Q) ≤ sup{τ (T u)τ 0 (v) : T ∈ Tµ¯,¯µ } = τ (u)τ 0 (v). Generally, setting un = |u| ∧ nχ1, hu∗n in∈N is a non-decreasing sequence with supremum u∗ (373Db, 373Dh), so

R

u∗ × v ∗ = supn∈N

R

Q u∗n × v ∗ ≤ supn∈N τ (un )τ 0 (v) = τ (u)τ 0 (v). Q

P If A = {0}, this is trivial. Otherwise, take u ∈ Lτ . There is surely (c) Consequently, Lτ ⊆ Mµ¯1,∞ . P 0 some non-zero a such that τ (χa) < ∞; now, setting v = χa,

R µ¯a 0

u∗ =

R

u∗ × v ∗ ≤ τ (u)τ 0 (v) < ∞

Q by (b) above. But this means that u∗ ∈ Mµ¯1,∞ , so that u ∈ Mµ¯1,∞ (373F(b-ii)). Q (d) Next, τ 0 is T -invariant. P P Suppose that v ∈ Mµ¯0,∞ , T ∈ Tµ¯,¯µ , u ∈ L0 (A) and τ (u) ≤ 1. Then u ∈ Mµ¯1,∞ , by (c), so

R

|u × T v| ≤

R

u∗ × v ∗ ≤ τ (u)τ 0 (v) ≤ τ 0 (v),

using 373J for the first inequality. Taking the supremum over u, we see that τ 0 (T v) ≤ τ 0 (v); as T and v are arbitrary, τ 0 is T -invariant. Q Q 0

(e) Finally, putting (d) and (c) together, Lτ ⊆ Mµ¯1,∞ , so that Lτ ⊇ Mµ¯∞,1 , using 369J and 369O. 374C For any T -invariant extended Fatou norm on L0 (AL ) there are corresponding norms on L0 (A) for any semi-finite measure algebra, as follows. Theorem Let θ be a T -invariant extended Fatou norm on L0 (AL ), and (A, µ ¯) a semi-finite measure algebra. (a) There is a T -invariant extended Fatou norm τ on L0 (A) defined by setting τ (u) = θ(u∗ ) if u ∈ Mµ¯0,∞ , = ∞ if u ∈ L0 (A) \ Mµ¯0,∞ . (b) Writing θ0 , τ 0 for the associates of θ and τ , we now have

442


374C

τ 0 (v) = θ0 (v ∗ ) if v ∈ Mµ¯0,∞ , = ∞ if v ∈ L0 (A) \ Mµ¯0,∞ . (c) If θ is an order-continuous norm on the Banach lattice Lθ , then τ is an order-continuous norm on Lτ . proof (a)(i) The argument seems to run better if I use a different formula to define τ : set

R

τ (u) = sup{ |u × T w| : T ∈ Tµ¯L ,¯µ , w ∈ L0 (AL ), θ0 (w) ≤ 1} for u ∈ L0 (A). (By 374B, w ∈ Mµ¯1,∞ whenever θ0 (w) ≤ 1, so there is no difficulty in defining T w.) Now τ (u) = θ(u∗ ) for every u ∈ Mµ¯0,∞ . P P (α) If w ∈ L0 (AL ) and θ0 (w) ≤ 1, then w ∈ Mµ¯1,∞ , by 374B(i), so L there is an S ∈ Tµ¯L ,¯µL such that Sw = w∗ (373O). Accordingly θ0 (w∗ ) ≤ θ0 (w) (because θ0 is T -invariant, by 374B); now

R

|u × T w| ≤

R

u∗ × w∗ ≤ θ(u∗ )θ0 (w∗ ) ≤ θ(u∗ )θ0 (w) ≤ θ(u∗ );

as w is arbitrary, τ (u) ≤ θ(u∗ ). (β) If w ∈ L0 (AL ) and θ0 (w) ≤ 1, then Z

Z ∗

(u∗ )∗ × w∗

|u × w| ≤ (373E)

Z

Z u × w = sup{ |u × T w| : T ∈ Tµ¯L ,¯µL } ∗

=

∗

(373Q) ≤ τ (u). But because θ is the associate of θ0 (369I(ii)), this means that θ(u∗ ) ≤ τ (u). Q Q (ii) Now τ is an extended Fatou norm on L0 (A). P P Of the conditions in 369F, (i)-(iv) are true just R because τ (u) = supv∈B |u × v| for some set B ⊆ L0 . As for (v) and (vi), observe that if u ∈ Mµ¯∞,1 then u∗ ∈ Mµ¯∞,1 (373F(b-iv)), so that τ (u) = θ(u∗ ) < ∞, by 374B(i), while also L u 6= 0 =⇒ u∗ 6= 0 =⇒ τ (u) = θ(u∗ ) > 0. ¯) is semi-finite), 369F(v)-(vi) As Mµ¯∞,1 is order-dense in L0 (A) (this is where I use the hypothesis that (A, µ are satisfied, and τ is an extended Fatou norm. Q Q (iii) τ is T -invariant. P P Take u ∈ Mµ¯1,∞ and T ∈ Tµ¯,¯µ . There are S0 ∈ Tµ¯L ,¯µ and S1 ∈ Tµ¯,¯µL such that S0 u∗ = u, S1 T u = (T u)∗ (373O); now S1 T S0 ∈ Tµ¯L ,¯µL (373Be), so τ (T u) = θ((T u)∗ ) = θ(S1 T S0 u∗ ) ≤ θ(u∗ ) = τ (u) because θ is T -invariant. Q Q (iv) We can now return to the definition of τ . I have already remarked that τ (u) = θ(u∗ ) if u ∈ Mµ¯0,∞ . For other u, we must have τ (u) = ∞ just because τ is a T -invariant extended Fatou norm (374B(i)). So the definitions in the statement of the theorem and (i) above coincide. (b) We surely have τ 0 (v) = ∞ if v ∈ L0 (A) \ Mµ¯0,∞ , by 374B, because τ 0 , like τ , is a T -invariant extended Fatou norm. So take v ∈ Mµ¯0,∞ . (i) If u ∈ L0 (A) and τ (u) ≤ 1, then

R

|v × u| ≤

R

v ∗ × u∗ ≤ θ0 (v ∗ )θ(u∗ ) = θ0 (v ∗ )τ (u) ≤ θ0 (v ∗ );

as u is arbitrary, τ 0 (v) ≤ θ0 (v ∗ ). (ii) If w ∈ L0 (AL ) and θ(w) ≤ 1, then

374D


Z

Z

443

Z

∗

∗

|v × w| ≤

∗

v × w = sup{

|v × T w| : T ∈ Tµ¯L ,¯µ }

(373Q) ≤ sup{τ 0 (v)τ (T w) : T ∈ Tµ¯L ,¯µ } = sup{τ 0 (v)θ((T w)∗ ) : T ∈ Tµ¯L ,¯µ } ≤ sup{τ 0 (v)θ(ST w) : T ∈ Tµ¯L ,¯µ , S ∈ Tµ¯,¯µL } (because, given T , we can find an S such that ST w = (T w)∗ , by 373O) ≤ sup{τ 0 (v)θ(T w) : T ∈ Tµ¯L ,¯µL } ≤ τ 0 (v). As w is arbitrary, θ0 (v ∗ ) ≤ τ 0 (v) and the two are equal. This completes the proof of (b). (c)(i) The first step is to note that Lτ ⊆ Mµ¯0 . P P?? Suppose that u ∈ Lτ \Mµ¯0 , that is, that µ ¯[[|u| > α]] = ∞ • ∗ 0 ∞ θ for some α > 0. Then u ≥ αχ1 in L (AL ), so L (AL ) ⊆ L . For each n ∈ N, set vn = χ [n, ∞[ . Then ∗ vn = v0 , so we can find a Tn ∈ Tµ¯L ,¯µL such that Tn vn = v0 (373O), and θ(vn ) ≥ θ(v0 ) for every n. But as hvn in∈N is a decreasing sequence with infimum 0, this means that θ is not an order-continuous norm. X XQ Q (ii) Now suppose that A ⊆ Lτ is non-empty and downwards-directed and has infimum 0. Then inf u∈A µ ¯[[u > α]] = 0 for every α > 0 (put 364Nb and 321F together). But this means that B = {u∗ : u ∈ A} must have infimum 0; since B is surely downwards-directed, inf v∈B θ(v) = 0, that is, inf u∈A τ (u) = 0. As A is arbitrary, τ is an order-continuous norm. 374D

What is more, every T -invariant extended Fatou norm can be represented in this way.

Theorem Let (A, µ ¯) be a semi-finite measure algebra, and τ a T -invariant extended Fatou norm on L0 (A). Then there is a T -invariant extended Fatou norm θ on L0 (AL ) such that τ (u) = θ(u∗ ) for every u ∈ Mµ¯0,∞ . proof I use the method of 374C. If A = {0} the result is trivial; assume that A 6= {0}. (a) Set

R

θ(w) = sup{ |w × T v| : T ∈ Tµ¯,¯µL , v ∈ L0 (A), τ 0 (v) ≤ 1} for w ∈ L0 (AL ). Note that

R

θ(w) = sup{ w∗ × v ∗ : v ∈ L0 (A), τ 0 (v) ≤ 1} , by 373J and 373Q again. for every w ∈ Mµ¯0,∞ L θ is an extended Fatou norm on L0 (AL ). P P As in 374C, the conditions 369F(i)-(iv) are Relementary. If w > 0 in L0 (AL ), take any v ∈ L0 (A) such that 0 < τ 0 (v) ≤ 1; then w∗ × v ∗ 6= 0 so θ(w) ≥ w∗ × v ∗ > 0. So 369F(v) is satisfied. As for 369F(vi), if w > 0 in L0 (AL ), take a non-zero a ∈ A of finite measure such that α = τ (χa) < ∞. Let β > 0, b ∈ AL be such that 0 < µ ¯L b ≤ µ ¯a and βχb ≤ w; then θ(χb) = supτ 0 (v)≤1

R

(χb)∗ × v ∗ ≤ supτ 0 (v)≤1

R

(χa)∗ × v ∗ ≤ τ (χa) < ∞

by 374B(ii). So θ(βχb) < ∞ and 369F(vi) is satisfied. Thus θ is an extended Fatou norm. Q Q (b) θ is T -invariant. P P If T ∈ Tµ¯L ,¯µL and w ∈ Mµ¯1,∞ , then L θ(T w) = supτ 0 (v)≤1

R

(T w)∗ × v ∗ ≤ supτ 0 (v)≤1

R

w∗ × v ∗ = θ(w)

by 373G and 373I. Q Q (c) θ(u∗ ) = τ (u) for every u ∈ Mµ¯0,∞ . P P We have τ (u) = supτ 0 (v)≤1

R

|u × v| ≤ supτ 0 (v)≤1

using 369I, 373E and 374B. So θ(u∗ ) = supτ 0 (v)≤1 by the remark in (a) above. Q Q

R

R

u∗ × v ∗ ≤ τ (u),

u∗ × v ∗ = τ (u)

444


374E

374E

I turn now to rearrangement-invariance. Let (A, µ ¯) be a measure algebra.

(a) I will say that a subset A of L0 = L0 (A) is rearrangement-invariant if Tπ u ∈ A whenever u ∈ A and π : A → A is a measure-preserving Boolean automorphism, writing Tπ : L0 → L0 for the isomorphism corresponding to π (364R). (b) I will say that an extended Fatou norm τ on L0 is rearrangement-invariant if τ (Tπ u) = τ (u) whenever u ∈ L0 and π : A → A is a measure-preserving automorphism. 374F Remarks (a) If (A, µ ¯) is a semi-finite measure algebra and π : A → A is a sequentially ordercontinuous measure-preserving Boolean homomorphism, then Tπ ¹Mµ¯1,∞ belongs to Tµ¯,¯µ ; this is obvious from the definition of M 1,∞ = L1 + L∞ and the basic properties of Tπ (364R). Accordingly, any T invariant extended Fatou norm τ on L0 (A) must be rearrangement-invariant, since (by 374B) we shall have τ (u) = τ (Tπ (u)) = ∞ when u ∈ / Mµ¯1,∞ . Similarly, any T -invariant subset of Mµ¯1,∞ will be rearrangementinvariant. (b) I seek to describe cases in which rearrangement-invariance implies T -invariance. This happens only for certain measure algebras; in order to shorten the statements of the main theorems I introduce a special phrase. 374G Definition I say that a measure algebra (A, µ ¯) is quasi-homogeneous if for any non-zero a, b ∈ A there is a measure-preserving Boolean automorphism π : A → A such that πa ∩ b 6= 0. 374H Proposition Let (A, µ ¯) be a semi-finite measure algebra. Then the following are equiveridical: (i) (A, µ ¯) is quasi-homogeneous; (ii) either A is purely atomic and every atom of A has the same measure or there is a κ ≥ ω such that the principal ideal Aa is homogeneous, with Maharam type κ, for every a ∈ A of non-zero finite measure. proof (i)⇒(ii) Suppose that (A, µ ¯) is quasi-homogeneous. α) Suppose that A has an atom a. In this case, for any b ∈ A \ {0} there is an automorphism π of (α (A, µ ¯) such that πa ∩ b 6= 0; now πa must be an atom, so πa = πa ∩ b and πa is an atom included in b. As b is arbitrary, A is purely atomic; moreover, if b is an atom, then it must be equal to πa and therefore of the same measure as a, so all atoms of A have the same measure. β ) Now suppose that A is atomless. In this case, if a ∈ A has finite non-zero measure, Aa is homo(β geneous. P P?? Otherwise, there are non-zero b, c ⊆ a such that the principal ideals Ab , Ac are homogeneous and of different Maharam types, by Maharam’s theorem (332B, 332H). But now there is supposed to be an automorphism π such that πb ∩ c 6= 0, in which case Ab , Aπb , Aπb∩c and Ac must all have the same Maharam type. X XQ Q Consequently, if a, b ∈ A are both of non-zero finite measure, the Maharam types of Aa , Aa∪b and Ab must all be the same infinite cardinal κ. (ii)⇒(i) Assume (ii), and take a, b ∈ A \ {0}. If a ∩ b 6= 0 we can take π to be the identity automorphism and stop. So let us suppose that a ∩ b = 0. α) If A is purely atomic and every atom has the same measure, then there are atoms a0 ⊆ a, b0 ⊆ b. (α Set πc = c if c ⊇ a0 ∪ b0 or c ∩ (a0 ∪ b0 ) = 0, = c 4 (a0 ∪ b0 ) otherwise. Then it is easy to check that π is a measure-preserving automorphism of A such that πa0 = b0 , so that πa ∩ b 6= 0. β ) If Ac is Maharam-type-homogeneous with the same infinite Maharam type κ for every non-zero (β c of finite measure, set γ = min(1, µ ¯a, µ ¯b) > 0. Because A is atomless, there are a0 ⊆ a, b0 ⊆ b with µ ¯ a0 = µ ¯b0 = γ (331C). Now Aa0 , Ab0 are homogeneous with the same Maharam type and the same

374J


445

magnitude, so by Maharam’s theorem (331I) there is a measure-preserving isomorphism π0 : Aa0 → Ab0 . Define π : A → A by setting πc = (c \ (a0 ∪ b0 )) ∪ π0 (c ∩ a0 ) ∪ π0−1 (c ∩ b0 ); then it is easy to see that π is a measure-preserving automorphism of A and that πa ∩ b 6= 0. Remark We shall return to these ideas in Chapter 38. In particular, the construction of π from π0 in the ←−−−− last part of the proof will be of great importance; in the language of 381G, π = (a0 π0 b0 ). 374I Corollary Let (A, µ ¯) be a quasi-homogeneous semi-finite measure algebra. Then (a) whenever a, b ∈ A have the same finite measure, the principal ideals Aa , Ab are isomorphic as measure algebras; (b) there is a subgroup Γ of the additive group R such that (α) µ ¯a ∈ Γ whenever a ∈ A and µ ¯a < ∞ (β) whenever a ∈ A, γ ∈ Γ and 0 ≤ γ ≤ µ ¯a then there is a c ⊆ a such that µ ¯c = γ. proof If A is purely atomic, with all its atoms of measure γ0 , set Γ = γ0 Z, and the results are elementary. If A is atomless, set Γ = R; then (a) is a consequence of Maharam’s theorem, and (b) is a consequence of 331C, already used in the proof of 374H. 374J Lemma Let (A, µ ¯) be a quasi-homogeneous semi-finite measure algebra and u, v ∈ Mµ¯0,∞ . Let Aut be the group of measure-preserving automorphisms of A. Then

R

u∗ × v ∗ = supπ∈Aut

R

|u × Tπ v|,

where Tπ : L0 (A) → L0 (A) is the isomorphism corresponding to π. f proof (a) Suppose first that u, v are non-negativeP and belong to S(AP ), where Af is the ring {a : µ ¯a < ∞}, n m β χb where α ≥ . . . αm ≥ 0, α χa , v = as usual. Then they can be expressed as u = j j 0 i i j=0 i=0 β0 ≥ . . . ≥ βn ≥ 0, a0 , . . . , am are dist and of finite measure, and b0 , . . . , bn are dist and of finite measure. Extending each list by a final term having a coefficient of 0, if need be, we may suppose that supi≤m ai = supj≤n bj . Let (t0 , . . . , ts ) enumerate in ascending order the set Pk Pk ¯bj : k ≤ n}. ¯ai : k ≤ m} ∪ { j=0 µ {0} ∪ { i=0 µ Pn Pm ¯bj . For 1 ≤ r ≤ s let k(r), ¯ai = j=0 µ Then every tr belongs to the subgroup Γ of 374Ib, and ts = i=0 µ P Pl(r) Pk(r) ¯bj . Then µ ¯ai = k(r)=i tr − tr−1 , ¯ai , tr ≤ j=0 µ l(r) be minimal subject to the requirements tr ≤ i=0 µ so (using 374Ib) we can find a dist family hcr i1≤r≤s such that cr ⊆ ak(r) and µ ¯cr = tr − tr−1 for each r. Similarly, there is a dist family hdr i1≤r≤s such that dr ⊆ bl(r) and µ ¯dr = tr − tr−1 for each r. Now the principal ideals Acr , Adr are isomorphic for every r, by 374Ia; let πr : Adr → Acr be measure-preserving isomorphisms. Define π : A → A by setting

πa = (a \ sup1≤r≤s dr ) ∪ sup1≤r≤s πr (a ∩ dr ); because supr≤s cr = supi≤m ai = supj≤n bj = supr≤s dr , π : A → A is a measure-preserving automorphism. Now Ps u = r=1 αk(r) χcr , u∗ = so

R

u × Tπ v =

Ps r=1

Ps r=1

v=

v∗ =

•

αk(r) χ [tr−1 , tr [ , αk(r) βl(r) µ ¯cr =

Ps

Ps r=1

r=1

βl(r) χdr ,

Ps r=1

•

βl(r) χ [tr−1 , tr [ ,

αk(r) βl(r) (tr − tr−1 ) =

R

u∗ × v ∗ .

(b) Now take any u0 , v0 ∈ Mµ¯0,∞ . Set A = {u : u ∈ S(Af ), 0 ≤ u ≤ |u0 |},

B = {v : v ∈ S(Af ), 0 ≤ v ≤ |v0 |}.

446


374J

Then A is an upwards-directed set with supremum |u0 |, because (A, µ ¯) is semi-finite, so {u∗ : u ∈ A} is an ∗ ∗ upwards-directed set with supremum |u0 | = u0 (373Db, 373Dh). Similarly {v ∗ : v ∈ B} is upwards-directed ∗ and has supremum v0∗ , so × v ∗ : u ∈ A, v ∈ B} is upwards-directed and supremum u∗0 × v0∗ . R {u R has ∗ ∗ ∗ Consequently, if γ < u0 × v0 , there are u ∈ A, v ∈ B such that γ ≤ u × v ∗ . Now, by (a), there is a π ∈ Aut such that γ≤

R

u × Tπ v ≤

R

|u0 | × Tπ |v0 | =

because Tπ is a Riesz homomorphism. As γ is arbitrary,

R

u∗0 × v0∗ ≤ supπ∈Aut

R

R

|u0 × Tπ v0 |

|u0 × Tπ v0 |.

But the reverse inequality is immediate from 373J. 374K Theorem Let (A, µ ¯) be a quasi-homogeneous semi-finite measure algebra, and τ a rearrangement -invariant extended Fatou norm on L0 = L0 (A). Then τ is T -invariant. proof Write τ 0 for the associate of τ . Then 374J tells us that for any u, v ∈ Mµ¯0,∞ ,

R

u∗ × v ∗ = supπ∈Aut

R

|Tπ u × v| ≤ supπ∈Aut τ (Tπ u)τ 0 (v) = τ (u)τ 0 (v),

writing u∗ , v ∗ for the decreasing rearrangements of u and v, and Aut for the group of measure-preserving automorphisms of (A, µ ¯). But now, if u ∈ Mµ¯1,∞ and T ∈ Tµ¯,¯µ , Z |T u × v| : τ 0 (v) ≤ 1}

τ (T u) = sup{ (by 369I)

Z ≤ sup{

u∗ × v ∗ : τ 0 (v) ≤ 1}

(by 373J) ≤ τ (u). As T , u are arbitrary, τ is T -invariant. 374L Lemma Let R(A, µ ¯) be a quasi-homogeneous semi-finite measure algebra. Suppose that u, v ∈ + ∗ (Mµ¯0,∞ ) are such that u × v ∗ = ∞. Then there is a measure-preserving automorphism π : A → A such R that u × Tπ v = ∞. proof I take three cases separately. (a) Suppose that A is purely atomic. Then it is surely infinite, since otherwise infinite. Let γ be the common measure of its atoms. For each n ∈ N, set αn = inf{α : α ≥ 0, µ ¯[[u > α]] ≤ 2n γ}. Then hαn in∈N is non-increasing, and µ ¯[[u > αn ]] ≤ 2n γ ≤ µ ¯[[u ≥ αn ]]. We can therefore choose a sequence h˜ an in∈N in A inductively so that [[u > αn ]] ⊆ a ñ ⊆ [[u ≥ αn ]],

µ ¯a ñ = 2n γ,

a ñ ⊆ a ñ+1 ,

for each n. Now if αn+1 ≤ α < αn , 2n γ < µ ¯[[u > α]] ≤ 2n+1 γ. So if we set £ £• • u ˜ = kuk∞ χ [0, γ[ ∨ supn∈N αn χ 2n γ, 2n+1 γ , then u∗ ≤ u ˜ in L∞ (AL ). Set an = a ñ+1 \ a ñ for each n; then han in∈N is dist and µ ¯an = 2n γ, for each n.

an ⊆ [[u ≥ αn+1 ]]

R

u∗ × v ∗ could not be

374L


447

Similarly, we can find a non-increasing sequence hβn in∈N in [0, ∞[ and a dist sequence hbn in∈N in A such that µ ¯bn = 2n γ, for each n, while

bn ⊆ [[v ≥ βn+1 ]]

£ £• • v ∗ ≤ v˜ = kvk∞ χ [0, γ[ ∨ supn∈N αn χ 2n γ, 2n+1 γ

in L∞ (AL ). R Now we are supposing that u∗ × v ∗ = ∞, so we must have ∞=

R

u ˜ × v˜ = γkuk∞ kvk∞ +

P∞ n=0

2n γαn βn .

Because 2n+1 αn+1 βn+1 ≤ 2 · 2n αn βn for each n, we must have P∞ 2n+1 α2n+1 β2n+1 = ∞. n=0 2 At this point, recall that we are dealing with a purely atomic algebra in which every S atom has measure γ. Let An , Bn be the sets of atoms included in an , bn for each n, and A = n∈N An ∪ Bn . Then #(An ) = #(Bn ) = 2n for eachSn. We therefore have S a bijection φ : A → A such that φ[B2n ] = A2n for every n. (The point is that A \ n∈N A2n and A \ n∈N B2n are both countably infinite.) Define π : A → A by setting πc = (c \ sup A) ∪ supa∈A,a ⊆ c φa. Then π is well-defined (because A is countable), and it is easy to check that it is a measure-preserving Boolean automorphism (because it is just a permutation of the atoms); and πb2n = a2n for every n. Consequently

R

u × Tπ v ≥

P∞

n=0

α2n+1 β2n+1 µ ¯a2n =

P∞

n=0

22n γα2n+1 β2n+1 = ∞.

So we have found a suitable automorphism. (b) Next, consider the case in which (A, µ ¯) is atomless and of finite magnitude γ. Of course γ > 0. For each n ∈ N set αn = inf{α : α ≥ 0, µ ¯[[u > α]] ≤ 2−n γ}. Then

£ £• u∗ ≤ supn∈N αn+1 χ 2−n−1 γ, 2−n γ .

Also µ ¯[[u > αn ]] ≤ 2−n γ ≤ µ ¯[[u ≥ αn ]] for each n, so we can choose inductively a decreasing sequence h˜ an in∈N such that [[u > αn ]] ⊆ a ñ ⊆ [[u ≥ αn ]] −n

and µ ¯a ñ = 2 γ for each n. Set an = a ñ \ a ñ+1 ; then han in∈N is dist and µ ¯an = 2−n−1 γ, an ⊆ [[u ≥ αn ]] for each n. In the same way, we can find hβn in∈N , hbn in∈N such that hbn in∈N is a£ dist sequence in A, µ ¯bn = 2−n−1 γ £ −n−1 ∗ −n • and bn ⊆ [[v ≥ βn ]] for each n, and v ≤ supn∈N βn+1 χ 2 γ, 2 γ . Now all the principal ideals Aan , Abn are homogeneous and of the same Maharam type, so there are measure-preserving isomorphisms πn : Abn → Aan . Define π : A → A by setting πc = supn∈N πn (c ∩ an ); then π is a measure-preserving automorphism of A, and πbn = an for each n. Since u × χan ≥ αn χan , v × χbn ≥ βn χbn for each n,

R

but on the other hand, So

R

R

u × Tπ v ≥

u∗ × v ∗ ≤

P∞

P∞

n=0

2−n−1 γαn βn ;

2−n−1 γαn+1 βn+1 ≤ 2 n=0

R

u × Tπ v.

u × Tπ v = ∞.

(c) Thirdly, consider the case in which A is atomless and not totally finite; take κ to be the common Maharam type of all the principal ideals Aa where 0 < µ ¯a < ∞. In this case, set

448


αn = inf{α : µ ¯[[u > α]] ≤ 2n },

374L

βn = inf{α : µ ¯[[v > α]] ≤ 2n }

for each n ∈ Z. This time

£ £• u∗ ≤ supn∈Z αn χ 2n , 2n+1 ,

£ £• v ∗ ≤ supn∈Z βn χ 2n , 2n+1 .

There are dist families han in∈Z , hbn in∈Z such that µ ¯an = µ ¯bn = 2n for each n and u ≥ supn∈Z αn+1 χan ,

v ≥ supn∈Z βn+1 χbn .

(This time, start by fixing a ˜0 such that µ ¯a ˜0 = 1 and [[u > α0 ]] ⊆ a ˜0 ⊆ [[u ≥ α0 ]], and choose a ñ+1 ⊇ a ñ for n ≥ 0, a ñ−1 ⊆ a ñ for n ≤ 0.) Set d∗ = supn∈Z an ∪ supn∈Z bn . Then d1 = d∗ \ supn∈Z a2n ,

d2 = d∗ \ supn∈Z b2n

both have magnitude ω and Maharam type κ. So there is a measure-preserving isomorphism π ˜ : Ad2 → Ad1 (332J). At the same time, for each n ∈ Z there is a measure-preserving isomorphism πn : Ab2n → Aa2n . So once again we can assemble these to form a measure-preserving automorphism π : A → A, defined by the formula πc = (c \ d∗ ) ∪ π ˜ (c ∩ d2 ) ∪ supn∈Z πn (c ∩ b2n ). Just as in (a) and (b) above, while

R

u × Tπ v ≥

R

P∞ n=−∞

u∗ × v ∗ ≤

22n α2n+1 β2n+1 ,

P∞ n=−∞

2n αn βn

is infinite. Because 22n+2 α2n+2 β2n+2 + 22n+1 α2n+1 β2n+1 ≤ 6 · 22n α2n+1 β2n+1 for every n,

R

u × Tπ v ≥

1 6

R

u∗ × v ∗ = ∞.

Thus we have a suitable π in any of the cases allowed by 374H. 374M Proposition Let (A, µ ¯) be a quasi-homogeneous localizable measure algebra, and U ⊆ L0 = L (A) a solid linear subspace which, regarded as a Riesz space, is perfect. If U is rearrangement-invariant and Mµ¯∞,1 ⊆ U ⊆ Mµ¯1,∞ , then U is T -invariant. 0

proof Set V = {v : u × v ∈ L1 for every u ∈ U }, so that V is a solid linear subspace of L0 which can be identified with U × (369C), and U becomes {u : u × v ∈ L1 for every v ∈ V }; note that Mµ¯∞,1 ⊆ V ⊆ Mµ¯1,∞ (using 369Q). R If u ∈ UR + , v ∈ V + and π : A → A is a measure-preserving automorphism, then Tπ u ∈ U , so v×Tπ u < ∞; by 374L, u∗ × v ∗ is finite. But this means that if u ∈ U , v ∈ V and T ∈ Tµ¯,¯µ ,

R

|T u × v| ≤

R

u∗ × v ∗ < ∞.

As v is arbitrary, T u ∈ U ; as T and u are arbitrary, U is T -invariant. 374X Basic exercises > (a) Let (A, µ ¯) be a measure algebra and A ⊆ Mµ¯1,∞ a T -invariant set. (i) Show that A is solid. (ii) Show that if A is a linear subspace and not {0}, then it includes Mµ¯∞,1 . (iii) Rt Rt Show that if u ∈ A, v ∈ Mµ¯0,∞ and 0 v ∗ ≤ 0 u∗ for every t > 0, then v ∈ A. (iv) Show that if (B, ν¯) is any other measure algebra, then B = {T u : u ∈ A, T ∈ Tµ¯,¯ν } and C = {v : v ∈ Mν¯1,∞ , T v ∈ A for every T ∈ Tν¯,¯µ } are T -invariant subsets of Mν¯1,∞ , and that B ⊆ C. Give two examples in which B ⊂ C. Show that if (A, µ ¯) = (AL , µ ¯L ) then B = C. >(b) Let (A, µ ¯) be a measure algebra. Show that the extended Fatou norm k kp on L0 (A) is T -invariant for every p ∈ [1, ∞]. (Hint: 371Gd.)

374Yb


449

(c) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras, and φ a Young’s function (369Xc). Let τφ , τ˜φ be the corresponding Orlicz norms on L0 (A), L0 (B). Show that τ˜φ (T u) ≤ τφ (u) for every u ∈ L0 (A), T ∈ Tµ¯,¯ν . (Hint: 369Xn, 373Xm.) In particular, τφ is T -invariant. (d) Show that if (A, µ ¯) is a semi-finite measure algebra and τ is a T -invariant extended Fatou norm on L0 (A), then the Banach lattice Lτ defined from τ is T -invariant. (e) Let (A, µ ¯) be a semi-finite measure algebra and τ a T -invariant extended Fatou norm on L0 which is an order-continuous norm on Lτ . Show that Lτ ⊆ Mµ¯1,0 . (f ) Let θ be a T -invariant extended Fatou norm on L0 (AL ) and (A, µ ¯), (B, ν¯) two semi-finite measure algebras. Let τ1 , τ2 be the extended Fatou norms on L0 (A), L0 (B) defined from θ by the method of 374C. Show that τ2 (T u) ≤ τ1 (u) whenever u ∈ Mµ¯1,∞ and T ∈ Tµ¯,¯ν . 1 R |u| µ ¯a a

>(g) Let (A, µ ¯) be a semi-finite measure algebra, not {0}, and set τ (u) = sup0<¯µa<∞ √

for

0

u ∈ L (A). Show that τ is a T -invariant extended Fatou norm. Find examples of (A, µ ¯) for which τ is, and is not, order-continuous on Lτ . (h) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras and τ a T -invariant extended Fatou norm on L0 (A). (i) Show that there is a T -invariant extended Fatou norm θ on L0 (B) defined by setting θ(v) = sup{τ (T v) : T ∈ Tν¯,¯µ } for v ∈ Mν¯1,∞ . (ii) Show that θ0 (w) = sup{τ 0 (T w) : T ∈ Tν¯,¯µ } for every w ∈ Mν¯1,∞ . (iii) Show that when (A, µ ¯) = (AL , µ ¯L ) then θ(v) = τ (v ∗ ) for every v ∈ Mν¯0,∞ . (iv) Show that when ∗ (B, ν¯) = (AL , µ ¯L ) then τ (u) = θ(u ) for every u ∈ Mµ¯0,∞ . (i) Let (A, µ ¯) be a semi-finite measure algebra and τ an extended Fatou norm on L0 = L0 (A). Suppose that Lτ is a T -invariant subset of L0 . Show that there is a T -invariant extended Fatou norm τ˜ which is 0 equivalentR to τ in the sense that, for some M > 0, τ˜(u) ≤ M τ (u) ≤ M 2 τ˜(u) for every R ∗ u ∈ ∗L . (Hint: show ∗ ∗ τ τ0 first that u × v < ∞ for every u ∈ L and v ∈ L , then that supτ (u)≤1,τ 0 (v)≤1 u × v < ∞.) (j) Suppose that τ is a T -invariant extended Fatou norm on L0 (AL ), and that 0 < w = w∗ ∈ Mµ¯1,∞ . Let L (A, µ ¯) be any semi-finite measure algebra. Show that the function u 7→ τ (w × u∗ ) extends to a T -invariant .) (When extended Fatou norm θ on L0 (A). (Hint: τ (w × u∗ ) = sup{τ (w × T u) : T ∈ Tµ¯,¯µL } for u ∈ Mµ¯1,∞ L τ = k kp these norms are called Lorentz norms; see Lindenstrauss & Tzafriri 79, p. 121.) (k) Let (A, µ ¯) be PN with counting measure. Identify L0 (A) with RN . Let U be {u : u ∈ RN , {n : u(n) 6= 0} is finite}. Show that U is a perfect Riesz space, and is rearrangement-invariant but not T -invariant. (l) Let (A, µ ¯) be an atomless quasi-homogeneous localizable measure algebra, and U ⊆ L0 (A) a rearrangement-invariant solid linear subspace which is a perfect Riesz space. Show that U ⊆ Mµ¯1,∞ and that U is T invariant. (Hint: assume U 6= {0}. Show that (i) χa ∈ U whenever µ ¯a < ∞ (ii) V = {v : v×u ∈ L1 ∀ u ∈ U } 1,∞ is rearrangement-invariant (iii) U , V ⊆ M .) 374Y Further exercises (a) Let (A, µ ¯) be a localizable measure algebra and U ⊆ Mµ¯1,∞ a non-zero T -invariant Riesz subspace which, regarded as a Riesz space, is perfect. (i) Show that U includes Mµ¯∞,1 . (ii) Show that its dual {v : v ∈ L0 , vR× u ∈ L1 ∀ u ∈ U } (which in this exercise I will denote by U × ) is also T -invariant, and is {v : v ∈ Mµ¯0,∞ , u∗ × v ∗ < ∞ ∀ u ∈ U }. (iii) Show that for any localizable measure algebra (B, ν) the set V = {v : v ∈ Mν¯1,∞ , T v ∈ U ∀ T ∈ Tν¯,¯µ } is a perfect Riesz subspace of L0 (B), and that V × = {v : v ∈ Mν¯1,∞ , T v ∈ U × ∀ T ∈ Tν¯,¯µ }. (iv) Show that if, in (i)-(iii), (A, µ ¯) = (AL , µ ¯L ), then V = {v : v ∈ M 0,∞ , v ∗ ∈ U }. (v) Show that if, in (iii), (B, ν¯) = (AL , µ ¯L ), then U = {u : u ∈ M 0,∞ , u∗ ∈ V }. (b) Let (A, µ ¯) be a semi-finite measure algebra, and suppose that 1 ≤ q ≤ p < ∞. Let wpq ∈ L0 (AL ) be the equivalence class of the function t 7→ t(q−p)/p . (i) Show that for any u ∈ L0 (A),

R

wpq × (u∗ )q = p

R∞ 0

tq−1 (¯ µ[[|u| > t]])q/p dt.

450


374Yb

(ii) Show that we have an extended Fatou norm k kp,q on L0 (A) defined by setting ¢1/q ¡ R∞ kukp,q = p 0 tq−1 (¯ µ[[|u| > t]])q/p dt 1/q

for every u ∈ L0 (A). (Hint: use 374Xj with w = wpq , k k = k kq .) (iii) Show that if (B, ν¯) is another measure algebra and T ∈ Tµ¯,¯ν , then kT ukp,q ≤ kukp,q for every u ∈ Mµ¯1,∞ . (iv) Show that k kp,q is an order-continuous norm on Lk kp,q . (c) Let (A, µ ¯) be a homogeneous measure algebra of uncountable Maharam type, and u, v ≥ 0 in Mµ¯0 ∗ such that u = v ∗ . Show that there is a measure-preserving automorphism π of A such that Tπ u = v, where Tπ : L0 (A) → L0 (A) is the isomorphism corresponding to π. (d) In L0 (AL ) let u be the equivalence class of the function f (t) = te−t . Show that there is no Boolean automorphism π of AL such that Tπ u = u∗ . (Hint: show that AL is τ -generated by {[[u∗ > α]] : α > 0}.) (e) Let (A, µ ¯) be a quasi-homogeneous semi-finite measure algebra and C ⊆ L0 (A) a solid convex orderclosed rearrangement-invariant set. Show that C ∩ Mµ¯1,∞ is T -invariant. 374 Notes and comments I gave this section the title ‘rearrangement-invariant spaces’ because it looks good on the Contents page, and it follows what has been common practice since Luxemburg 67b; but actually I think that it’s T -invariance which matters, and that rearrangement-invariant spaces are significant largely because the important ones are T -invariant. The particular quality of T -invariance which I have tried to bring out here is its transferability from one measure algebra (or measure space, of course) to another. This is what I take at a relatively leisurely pace in 374B-374D and 374Xf, and then encapsulate in 374Xh and 374Ya. The special place of the Lebesgue algebra (AL , µ ¯L ) arises from its being more or less the simplest algebra over which every T -invariant set can be described; see 374Xa. I don’t think this section is particularly easy, and (as in §373) there are rather a lot of unattractive names in it; but once one has achieved a reasonable familiarity with the concepts, the techniques used can be seen to amount to half a dozen ideas – non-trivial ideas, to be sure – from §§369 and 373. From §369 I take concepts of duality: the symmetric relationship between a perfect Riesz space U ⊆ L0 and the representation of its dual (369C-369D), and the notion of associate extended Fatou norms (369H-369K). From §373 I take the idea of ‘decreasing rearrangement’ and theorems guaranteeing the existence of useful of Tµ¯,¯ν (373O-373Q). The results of the present section all depend on repeated use of these facts, assembled in a variety of patterns. There is one new method here, but an easy one: the construction of measure-preserving automorphisms by ing isomorphisms together, as in the proofs of 374H and 374J. I shall return to this idea, in greater generality and more systematically investigated, in §381. I hope that the special cases here will give no difficulty. While T -invariance is a similar phenomenon for both extended Fatou norms and perfect Riesz spaces (see 374Xh, 374Ya), the former seem easier to deal with. The essential difference is I think in 374B(i); with a T -invariant extended Fatou norm, we are necessarily confined to M 1,∞ , the natural domain of the methods used here. For perfect Riesz spaces we have examples like RN ∼ = L0 (PN) and its dual, the space of eventually-zero sequences (374Xk); these are rearrangement-invariant but not T -invariant, as I have defined it. This problem does not arise over atomless algebras (374Xl). I think it is obvious that for algebras which are not quasi-homogeneous (374G) rearrangement-invariance is going to be of limited interest; there will be regions between which there is no communication by means of measure-preserving automorphisms, and the best we can hope for is a discussion of quasi-homogeneous components, if they exist, corresponding to the partition of unity used in the proof of 332J. There is a special difficulty concerning rearrangement-invariance in L0 (AL ): two elements can have the same decreasing rearrangement without being rearrangements of each other in the strict sense (373Ya, 374Yd). The phenomenon of 373Ya is specific to algebras of countable Maharam type (374Yc). You will see that some of the labour of 374L is because we have to make room for the pieces to move in. 374J is easier just because in that context we can settle for a supremum, rather than an actual infinity, so the rearrangement needed (part (a) of the proof) can be based on a region of finite measure.

375C

Kwapien’s theorem

451

375 Kwapien’s theorem In §368 and the first part of §369 I examined maps from various types of Riesz space into L0 spaces. There are equally striking results about maps out of L0 spaces. I start with some relatively elementary facts about positive linear operators from L0 spaces to Archimedean Riesz spaces in general (375A-375D), and then turn to a remarkable analysis, due essentially to S.Kwapien, of the positive linear operators from a general L0 space to the L0 space of a semi-finite measure algebra (375I), with a couple of simple corollaries. 375A Theorem Let A be a Dedekind σ-complete Boolean algebra and W an Archimedean Riesz space. If T : L0 (A) → W is a positive linear operator, it is sequentially order-continuous. proof (a) The first step is to observe that if hun in∈N is any non-increasing sequence in L0 = L0 (A) with infimum 0, and ² > 0, then {n(un − ²u0 ) : n ∈ N} is bounded above in L0 . P P For k ∈ N set ak = supn∈N [[n(un − ²u0 ) > k]]; set a = inf k∈N ak . ?? Suppose, if possible, that a 6= 0. Because un ≤ u0 , n(un − ²u0 ) ≤ nu0 for every n and a ⊆ a0 ⊆ [[u0 > 0]] = [[²u0 > 0]] = supn∈N [[²u0 − un > 0]]. So there is some m ∈ N such that a0 = a ∩ [[²u0 − um > 0]] 6= 0. Now, for any n ≥ m, any k ∈ N, a0 ∩ [[n(un − ²u0 ) > k]] ⊆ [[²u0 − um > 0]] ∩ [[um − ²u0 > 0]] = 0. But a0 ⊆ supn∈N [[n(un − ²u0 ) > k]], so in fact a0 ⊆ supn≤m [[n(un − ²u0 ) > k]] = [[v > k]], where v = supn≤m n(un − ²u0 ). And this means that inf k∈N [[v > k]] ⊇ a0 6= 0, which is impossible. X X Accordingly a = 0; by 364Ma, {n(un − ²u0 ) : n ∈ N} is bounded above. Q Q (b) Now suppose that hun in∈N is a non-increasing sequence in L0 with infimum 0, and that w ∈ W is a lower bound for {T un : n ∈ N}. Take any ² > 0. By (a), {n(un − ²u0 ) : n ∈ N} has an upper bound v in L0 . Because T is positive, 1 n

1 n

w ≤ T un = T (un − ²u0 ) + T (²u0 ) ≤ T ( v) + T (²u0 ) = T v + ²T u0 for every n ≥ 1. Because W is Archimedean, w ≤ ²T u0 . But this is true for every ² > 0, so (again because W is Archimedean) w ≤ 0. As w is arbitrary, inf n∈N T un = 0. As hun in∈N is arbitrary, T is sequentially order-continuous (351Gb). 375B Proposition Let A be an atomless Dedekind σ-complete Boolean algebra. Then L0 (A)× = {0}. proof ?? Suppose, if possible, that h : L0 (A) → R is a non-zero order-continuous positive linear functional. Then there is a u > 0 in L0 such that h(v) > 0 whenever 0 < v ≤ u (356H). Because A is atomless, there is a dist sequence han in∈N such that an ⊆ [[u > 0]] for each n, so that un = u × χan > 0, while um ∧ un = 0 if m 6= n. Now however v = supn∈N n(h(un ))−1 un is defined in L0 , by 368K, and h(v) ≥ n for every n, which is impossible. X X 375C Theorem Let A be a Dedekind complete Boolean algebra, W an Archimedean Riesz space, and T : L0 (A) → W an order-continuous Riesz homomorphism. Then V = T [L0 (A)] is an order-closed Riesz subspace of W . proof The kernel U of T is a band in L0 = L0 (A) (352Oe), and must be a projection band (353I), because L0 is Dedekind complete (364O). Since U + U ⊥ = L0 , T [U ] + T [U ⊥ ] = V , that is, T [U ⊥ ] = V ; since U ∩ U ⊥ = {0}, T is an isomorphism between U ⊥ and V . Now suppose that A ⊆ V is upwards-directed and has a least upper bound w ∈ W . Then B = {u : u ∈ U ⊥ , T u ∈ A} is upwards-directed and T [B] = A. The point is that B is bounded above in L0 . P P?? If not, then {u+ : u ∈ B} cannot be bounded above, so there 0 is a u0 > 0 in L such that nu0 = supu∈B nu0 ∧ u+ for every n ∈ N (368A). Since B ⊆ U ⊥ , u0 ∈ U ⊥ and T u0 > 0. But now, because T is an order-continuous Riesz homomorphism,

452


375C

nT u0 = supu∈B T (nu0 ∧ u+ ) = supv∈A nT u0 ∧ v + ≤ w+ for every n ∈ N, which is impossible. X XQ Q Set u∗ = sup B; then T u∗ = sup A = w and w ∈ V . As A is arbitrary, V is order-closed. 375D Corollary Let W be an Archimedean Riesz space and V an order-dense Riesz subspace which is isomorphic to L0 (A) for some Dedekind complete Boolean algebra A. Then V = W . proof Apply 375C to an isomorphism T : L0 (A) → V to see that V is order-closed in W . 375E I come now to the deepest result of this section, concerning positive linear operators from L0 (A) to L0 (B) where B is a measure algebra. I approach through a couple of lemmas which are striking enough in their own right. The following temporary definition will be useful. Definition Let A and B be Boolean algebras. I will say that a function φ : A → B is a σ-subhomomorphism if φ(a ∪ a0 ) = φ(a) ∪ φ(a0 ) for all a, a0 ∈ A, inf n∈N φ(an ) = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. Now we have the following easy facts. 375F Lemma Let A and B be Boolean algebras and φ : A → B a σ-subhomomorphism. (a) φ(0) = 0, φ(a) ⊆ φ(a0 ) whenever a ⊆ a0 , and φ(a) \ φ(a0 ) ⊆ φ(a \ a0 ) for every a, a0 ∈ A. (b) If µ ¯, ν¯ are measures such that (A, µ ¯) and (B, ν¯) are totally finite measure algebras, then for every ² > 0 there is a δ > 0 such that ν¯φ(a) ≤ ² whenever µ ¯a ≤ δ. proof (a) This is elementary. Set every an = 0 in the second clause of the definition 375E to see that φ(0) = 0. The other two parts are immediate consequences of the first clause. (b) (Compare 232B, 327Bb.) ?? Suppose, if possible, otherwise. Then for every n ∈ N there is an an ∈ A such that µ ¯an ≤ 2−n and ν¯φ(an ) ≥ ². Set cn = supi≥n ai for each n; then hcn in∈N is non-increasing and has infimum 0 (since µ ¯cn ≤ 2−n+1 for each n), but ν¯φ(cn ) ≥ ² for every n, so inf n∈N φcn cannot be 0. X X 375G Lemma Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras and φ : A → B a σ-subhomomorphism. Then for every non-zero b0 ∈ B there are a non-zero b ⊆ b0 and an m ∈ N such that b ∩ inf j≤m φ(aj ) = 0 whenever a0 , . . . , am ∈ A are dist. proof (a) Suppose first that A is atomless and that µ ¯1 = 1. 1 Set ² = 51 ν¯b0 and let m ≥ 1 be such that ν¯φ(a) ≤ ² whenever µ ¯a ≤ m . We need to know that 1 m 1 1 1 (1 − m ) ≤ 2 ; this is because (if m ≥ 2) ln m − ln(m − 1) ≥ m , so m ln(1 − m ) ≤ −1 ≤ − ln 2. Set C = {inf j≤m φ(aj ) : a0 , . . . , am ∈ A are dist}. ?? Suppose, if possible, that b0 ⊆ sup C. Then there are c0 , . . . , ck ∈ C such that ν¯(b0 ∩ supi≤k ci ) ≥ 4². For each i ≤ k choose dist ai0 , . . . , aim ∈ A such that ci = inf j≤m φ(aij ). Let D be the set of atoms of the finite subalgebra of A generated by {aij : i ≤ k, j ≤ m}, so that D is a finite partition of unity in A, and every aij is the of the of D it includes. Set p = #(D), and for each d ∈ D take a maximal 1 1 }, so that µ ¯(d \ sup Ed ) < pm ; set dist set Ed ⊆ {e : e ⊆ d, µ ¯e = pm S d∗ = 1 \ sup( d∈D Ed ) = supd∈D (d \ sup Ed ), 1 1 1 so that µ ¯d∗ is a multiple of pm and is less than m . Let E ∗ be a dist set of elements of measure pm with S 1 ∗ ∗ union d , and take E = E ∪ d∈D Ed , so that E is a partition of unity in A, µ ¯e = pm for every e ∈ E, and aij \ d∗ is the of the of E it includes for every i ≤ k, j ≤ m. Set

K = {K : K ⊆ E, #(K) = p},

M = #(K) =

(mp)! . p!(mp−p)!

375H

Kwapien’s theorem 1 m

For every K ∈ K, µ ¯(sup K) = R

453

so ν¯φ(sup K) ≤ ². So if we set P v = K∈K χφ(sup K),

v ≤ ²M . On the other hand, ν¯φ(d∗ ) ≤ ²,

ν¯(b0 ∩ supi≤k ci ) ≥ 4², so ν¯b1 ≥ 3², where Accordingly

R

b1 = b0 ∩ supi≤k ci \ φ(d∗ ). v ≤ 31 M ν¯b1 and b2 = b1 ∩ [[v < 12 M ]]

is non-zero. Because b2 ⊆ b1 , there is an i ≤ k such that b2 ∩ ci 6= 0. Now b2 ∩ ci ⊆ ci \ φ(d∗ ) = inf j≤m φ(aij ) \ φ(d∗ ) ⊆ inf j≤m φ(aij \ d∗ ). But every aij \ d∗ is the of the of E it includes, so b2 ∩ ci ⊆ inf φ(aij \ d∗ ) ⊆ inf φ(sup{e : e ∈ E, e ⊆ aij }) j≤m

j≤m

= inf sup{φ(e) : e ∈ E, e ⊆ aij } j≤m

= sup{ inf φ(ej ) : e0 , . . . , em ∈ E and ej ⊆ aij for every j}. j≤m

So there are e0 , . . . , em ∈ E such that ej ⊆ aij for each j and b3 = b2 ∩ inf j≤m φ(ej ) 6= 0. Because ai0 , . . . , aim are dist, e0 , . . . , em are distinct; set J = {e0 , . . . , em }. Then whenever K ∈ K and K∩J 6= ∅, b3 ⊆ φ(sup K). So let us calculate the size of K1 = {K : K ∈ K, K ∩ J 6= ∅}. This is M−

(mp−m−1)! p!(mp−p−m−1)!

¡ (mp−p)(mp−p−1)...(mp−p−m) ¢ =M 1− mp(mp−1)...(mp−m)

¡

≥M 1−(

mp−p m+1 ¢ ) mp

1 2

≥ M.

But this means that b3 ⊆ [[v ≥ 21 M ]], while also b3 ⊆ [[v < 12 M ]]; which is surely impossible. X X Accordingly b0 6⊆ sup C, and we can take b = b0 \ sup C. (b) Now for the general case. Let A be the set of atoms of A, and set d = 1 \ sup A. Then Ad is atomless, so there are a non-zero b1 ⊆ b0 and an n ∈ N such that b1 ∩ inf j≤n φ(aj ) = 0 whenever a0 , . . . , an ∈ Ad are dist. P P If µ ¯d > 0 this follows from (a), if we apply it to φ¹ Ad and (¯ µd)−1 µ ¯¹ Ad . If µ ¯d = 0 then we can just take b1 = b0 , n = 0. Q Q Let δ > 0 be such that ν¯φ(a) < ν¯b1 whenever µ ¯a ≤ δ. Let A1 ⊆ A be a finite set such that µ ¯(sup A1 ) ≥ µ ¯(sup A) − δ, and set r = #(A), d∗ = sup(A \ A1 ). Then µ ¯d∗ ≤ δ so b = b1 \ φ(d∗ ) 6= 0. Try m = n + r. If a0 , . . . , am are dist, then at most r of them can meet sup A1 , so (re-ordering if necessary) we can suppose that a0 , . . . , an are dist from sup A1 , in which case aj \ d∗ ⊆ d for each j ≤ m. But in this case (because b ∩ φ(d∗ ) = 0) b ∩ inf j≤m φ(aj ) ⊆ b ∩ inf j≤n φ(aj ) = b ∩ inf j≤n φ(aj ∩ d) = 0 by the choice of n and b1 . Thus in the general case also we can find appropriate b and m. 375H Lemma Let (A, µ ¯) and (B, ν¯) be totally finite measure algebras and φ : A → B a σ-subhomomorphism. Then for every non-zero b0 ∈ B there are a non-zero b ⊆ b0 and a finite partition of unity C ⊆ A such that a 7→ b ∩ φ(a ∩ c) is a ring homomorphism for every c ∈ C. proof By 375G, we can find b1 , m such that 0 6= b1 ⊆ b0 and b1 ∩ inf j≤m φ(aj ) = 0 whenever a0 , . . . , am ∈ A are dist. Do this with the smallest possible m. If m = 0 then b1 ∩ φ(1) = 0, so we can take

454


375H

b = b1 , C = {1}. Otherwise, because m is minimal, there must be dist c1 , . . . , cm ∈ A such that b = b1 ∩ inf 1≤j≤m φ(cj ) 6= 0. Set c0 = 1 \ sup1≤j≤m cj , C = {c0 , c1 , . . . , cm }; then C is a partition of unity in A. Set πj (a) = b ∩ φ(a ∩ cj ) for each a ∈ A, j ≤ m. Then we always have πj (a ∪ a0 ) = πj (a) ∪ πj (a0 ) for all a, a0 ∈ A, because φ is a subhomomorphism. To see that every πj is a ring homomorphism, we need only check that πj (a ∩ a0 ) = 0 whenever a ∩ a0 = 0. (Compare 312H(iv).) In the case j = 0, we actually have π0 (a) = 0 for every a, because b ∩ φ(c0 ) = b1 ∩ inf 0≤j≤m φ(cj ) = 0 by the choice of b1 and m. When 1 ≤ j ≤ m, if a ∩ a0 = 0, then πj (a) ∩ πj (a0 ) = b1 ∩ inf 1≤i≤m,i6=j φ(cj ) ∩ φ(a) ∩ φ(a0 ) is again 0, because a, a0 , c1 , . . . , cj−1 , cj+1 , . . . , cm are dist. So we have a suitable pair b, C. 375I Theorem Let A be any Dedekind σ-complete Boolean algebra and (B, ν¯) a semi-finite measure algebra. Let T : L0 (A) → L0 (B) be a positive linear operator. Then we can find B, hAb ib∈B such that B is a partition of unity in B, each Ab is a finite partition of unity in A, and u 7→ T (u × χa) × χb is a Riesz homomorphism for every b ∈ B, a ∈ Ab . proof (a) Write B ∗ for the set of potential of B; that is, the set of those b ∈ B such that there is a finite partition of unity A ⊆ A such that Tab is a Riesz homomorphism for every a ∈ A, writing Tab (u) = T (u × χa) × χb. If I can show that B ∗ is order-dense in B, this will suffice, since there will then be a partition of unity B ⊆ B ∗ . (b) So let b0 be any non-zero member of B; I seek a non-zero member of B ∗ included in b0 . Of course there is a non-zero b1 ⊆ b0 withR ν¯b1 < ∞. Let γ > 0 be such that b2 = b1 ∩ [[T (χ1) ≤ γ]] is non-zero. Define R µ : A → [0, ∞[ by setting µa = b2 T (χa) for every a ∈ A. Then µ is countably additive, because χ, T and are all additive and sequentially order-continuous (using 375A). Set N = {a : µa = 0}; then N is a σ-ideal of A, and (C, µ ¯) is a totally finite measure algebra, where C = A/N and µ ¯a• = µa for every a ∈ A (just as in 321H). (c) We have a function φ from C to the principal ideal Bb2 defined by saying that φa• = b2 ∩ [[T (χa) > 0]] for every a ∈ A. P P If a1 , a2 ∈ A are such that a•1 = a•2 in C, this means that a1 4 a2 ∈ N ; now [[T (χa1 ) > 0]] 4 [[T (χa2 ) > 0]] ⊆ [[|T (χa1 ) − T (χa2 )| > 0]] ⊆

[[T (|χa1 − χa2 |) > 0]] = [[T χ(a1 4 a2 ) > 0]]

R

is dist from b2 because b2 T χ(a1 4 a2 ) = 0. Accordingly b2 ∩ [[T (χa1 ) > 0]] = b2 ∩ [[T (χa2 ) > 0]] and we Q can take this common value for φ(a•1 ) = φ(a•2 ). Q (d) Now φ is a σ-subhomomorphism. P P (i) For any a1 , a2 ∈ A we have [[T χ(a1 ∪ a2 ) > 0]] = [[T (χa1 ) > 0]] ∪ [[T (χa2 ) > 0]] because T (χa1 ) ∨ T (χa2 ) ≤ T χ(a1 ∪ a2 ) ≤ T (χa1 ) + T (χa2 ). So φ(c1 ∪ c2 ) = φ(c1 ) ∪ φ(c2 ) for all c1 , c2 ∈ C. (ii) If hcn in∈N is a non-increasing sequence in C with infimum 0, choose an ∈ A such that a•n = cn for each n, and set a ñ = inf i≤n ai \ inf i∈N ai for each n; then a ˜•n = cn so φ(cn ) = [[T (χ˜ an ) > 0]] for each n, while h˜ an in∈N is non-increasing and inf n∈N a ñ = 0. ?? Suppose, if possible, an ) > 0]]) ≥ 2² for every n ∈ N. For each n, that b0 = inf n∈N φ(cn ) 6= 0; set ² = 21 ν¯b0 . Then ν¯(b2 ∩ [[T (χ˜ take αn > 0 such that ν¯(b2 ∩ [[T (χ˜ an ) > αn ]]) ≥ ². Then u = supn∈N nαn−1 χ˜ an is defined in L0 (A) (because −1 supn∈N [[nαn−1 χ˜ an > k]] ⊆ a ˜m if k ≥ maxi≤m iαi , so inf k∈N supn∈N [[nαn−1 χ˜ an > k]] = 0). But now ν¯(b2 ∩ [[T u > n]]) ≥ ν¯(b2 ∩ [[T (χ˜ an ) > αn ]]) ≥ ² for every n, so inf n∈N [[T u > n]] 6= 0, which is impossible. X X Thus inf n∈N φ(cn ) = 0; as hcn in∈N is arbitrary, φ is a σ-subhomomorphism. Q Q (e) By 375H, there are a non-zero b ∈ Bb2 and a finite partition of unity C ⊆ C such that d 7→ b ∩ φ(d ∩ c) is a ring homomorphism for every c ∈ C. There is a partition of unity A ⊆ A, of the same size as C, such that C = {a• : a ∈ A}. Now Tab is a Riesz homomorphism for every a ∈ A. P P It is surely a positive linear

375Xc

Kwapien’s theorem

455

operator. If u1 , u2 ∈ L0 (A) and u1 ∧ u2 = 0, set ei = [[ui > 0]] for each i, so that e1 ∩ e2 = 0. Observe that ui = supn∈N ui ∧ nχei , so that [[Tab ui > 0]] = supn∈N [[Tab (ui ∧ nχei ) > 0]] ⊆ [[Tab (χei ) > 0]] = b ∩ [[T χ(ei ∩ a) > 0]] for both i (of course Tab , like T , is sequentially order-continuous). But this means that [[Tab u1 > 0]] ∩ [[Tab u2 > 0]] ⊆ b ∩ [[T χ(e1 ∩ a) > 0]] ∩ [[T χ(e2 ∩ a) > 0]] = b ∩ φ(e•1 ∩ a• ) ∩ φ(e•2 ∩ a• ) = 0 because a• ∈ C, so d → 7 φ(d ∩ a• ) is a ring homomorphism, while e•1 ∩ e•2 = 0. So Tab u1 ∧ Tab u2 = 0. As u1 and u2 are arbitrary, Tab is a Riesz homomorphism (352G(iv)). Q Q (f ) Thus b ∈ B ∗ . As b0 is arbitrary, B ∗ is order-dense, and we’re home. 375J Corollary Let A be a Dedekind σ-complete Boolean algebra and U a Dedekind complete Riesz space such that U × separates the points of U . If T : L0 (A) → U is a positivePlinear operator, there is a ∞ 0 sequence hTn iP n∈N of Riesz homomorphisms from L (A) to U such that T = n=0 Tn , in the sense that n 0 T u = supn∈N i=0 Ti u for every u ≥ 0 in L (A). proof By 369A, U can be embedded as an order-dense Riesz subspace of L0 (B) for some localizable measure algebra (B, ν¯); being Dedekind complete, it is solid in L0 (B) (353K). Regard T as an Q operator from L0 (A) 0 0 to L (B), and take B, hAb ib∈B as in 375I. Note that L (B) can be identified with b∈B L0 (Bb ) (364S, 322K). For each b ∈ B let fb : Ab → N be an injection. If b ∈ B and n ∈ fb [Ab ], set Tnb (u) = χb × T (u × χa); otherwiseP set Tnb = 0. Then Tnb : L0 (A) → L0 (Bb ) is a Riesz homomorphism; because Ab is a finite partition ∞ of unity, n=0 Tnb u = χb × T u for every u ∈ L0 (A). But this means that if we set Tn u = hTnb uib∈B , Q ∼ L0 (B) Tn : L0 (A) → b∈B L0 (Bb ) = P∞ is a Riesz homomorphism for each n; and T = n=0 Tn . Of course every Tn is an operator from L0 (A) to U because |Tn u| ≤ T |u| ∈ U for every u ∈ L0 (A). 375K Corollary (a) If A is a Dedekind σ-complete Boolean algebra, (B, ν¯) is a semi-finite measure algebra, and there is any non-zero positive linear operator from L0 (A) to L0 (B), then there is a non-trivial sequentially order-continuous ring homomorphism from A to B. (b) If (A, µ ¯) and (B, ν¯) are homogeneous probability algebras and τ (A) > τ (B), then L∼ (L0 (A); L0 (B)) = {0}. proof (a) It is probably quickest to look at the proof of 375I: starting from a non-zero positive linear operator T : L0 (A) → L0 (B), we move to a non-zero σ-subhomomorphism φ : A/N → B and thence to a non-zero ring homomorphism from A/N to B, corresponding to a non-zero ring homomorphism from A to B, which is sequentially order-continuous because it is dominated by φ. Alternatively, quoting 375I, we have a non-zero Riesz homomorphism T1 : L0 (A) → L0 (B), and it is easy to check that a 7→ [[T (χa) > 0]] is a non-zero sequentially order-continuous ring homomorphism. (b) Use (a) and 331J. 375X Basic exercises (a) Let A be a Dedekind complete Boolean algebra and W an Archimedean Riesz space. Let T : L0 (A) → W be a positive linear operator. Show that T is order-continuous iff T χ : A → W is order-continuous. (b) Let A be an atomless Dedekind σ-complete Boolean algebra and W a Banach lattice. Show that the only order-continuous positive linear operator from L0 (A) to W is the zero operator. (c) Let A be a Dedekind complete Boolean algebra and W an Archimedean Riesz space. Let T : L0 (A) → W be an order-continuous Riesz homomorphism such that T [L0 (A)] is order-dense in W . Show that T is surjective.

456


375Xd

(d) Let A and B be Boolean algebras and φ : A → B a σ-subhomomorphism as defined in 375E. Show that φ is sequentially order-continuous. > (e) Let A be the measure algebra of Lebesgue measure on [0, 1] and G the regular open algebra of R. (i) Show that there is no non-zero positive linear operator from L0 (G) to L0 (A). (Hint: suppose T : L0 (G) → L0 (A) were such an operator. Reduce to the case T (χ1) ≤ χ1. Let an R hbn in∈N enumerate R order-dense subset of G (316Yn). For each n ∈ N take non-zero b0n ⊆ bn such that T (χb0n ) ≤ 2−n−2 T (χ1) and consider T χ(supn∈N b0n ).) (ii) Show that there is no non-zero positive linear operator from L0 (A) to L0 (G). (Hint: suppose T : L0 (A) → L0 (G) were such an operator. For each n ∈ N choose an ∈ A, αn >P0 such that µ ¯an ≤ 2−n and if bn ⊆ [[T (χ1) > 0]] then bn ∩ [[T (χan ) > αn ]] 6= 0. Consider T u where ∞ −1 u = n=0 nαn χan .) (f ) In 375J, show that for any u ∈ L0 (A) inf n∈N supm≥n [[|T u −

Pm i=0

Ti u| > 0]] = 0.

> (g) Prove directly, without quoting 375E-375K, that if A is a Dedekind σ-complete Boolean algebra then every positive linear functional from L0 (A) to R is a finite sum of Riesz homomorphisms. 375Y Further exercises (a) Show that the following are equiveridical: (i) there is a purely atomic probability space (X, Σ, µ) such that Σ = PX and µ{x} = 0 for every x ∈ X; (ii) there are a set X and a Riesz homomorphism f : RX → R which is not order-continuous; (iii) there are a Dedekind complete Boolean algebra A and a positive linear operator f : L0 (A) → R which is not order-continuous; (iv) there are a Dedekind complete Boolean algebra A and a sequentially order-continuous Boolean homomorphism π : A → {0, 1} which is not order-continuous; (v) there are a Dedekind complete Riesz space U and a sequentially order-continuous Riesz homomorphism f : U → R which is not order-continuous; *(vi) there are an atomless Dedekind complete Boolean algebra A and a sequentially order-continuous Boolean homomorphism π : A → {0, 1} which is not order-continuous. (Compare 363S.) (b) Give an example of an atomless Dedekind σ-complete Boolean algebra A such that L0 (A)∼ 6= {0}. (c) Let A, B be Dedekind σ-complete Boolean algebras of which B is weakly σ-distributive. Let T : L0 (A) → L0 (B) be a positive linear operator. Show that a 7→ [[T (χa) > 0]] : A → B is a σ-subhomomorphism. (d) Let (A, µ ¯) be a localizable measure algebra and U a Riesz space such that U × separates the points of U . Suppose that T : L0 (A) → U is an order-continuous positive linear operator. Show that T [L0 (A)] is order-closed. (e) Let A and B be Dedekind complete Boolean algebras, and φ : A → B an σ-subhomomorphism such that φ1A = 1B . Show that there is a sequentially order-continuous Boolean homomorphism π : A → B such that πa ⊆ φa for every a ∈ A. (f ) Let G be the regular open algebra of R, and L0 = L0 (G). Give an example of a non-zero positive linear operator T : L0 → L0 such that there is no non-zero Riesz homomorphism S : L0 → L0 with S ≤ T . 375Z Problem Let G be the regular open algebra of R, and L0 = L0 (G). If T : L0 → L0 is a positive linear operator, must T [L0 ] be order-closed? 375 Notes and comments Both this section, and the earlier work on linear operators into L0 spaces, can be regarded as describing different aspects of a single fact: L0 spaces are very large. The most explicit statements of this principle are 368E and 375D: every Archimedean Riesz space can be embedded into a Dedekind complete L0 space, but no such L0 space can be properly embedded as an order-dense Riesz subspace of any other Archimedean Riesz space. Consequently there are many maps into L0 spaces (368B). But by the same token there are few maps out of them (375B, 375Kb), and those which do exist have a variety of special properties (375A, 375I).

376A

Kernel operators

457

The original version of Kwapien’s theorem (Kwapien 73) was the special case of 375I in which A is the Lebesgue measure algebra. The ideas of the proof here are mostly taken from Kalton, Peck & Roberts 84. I have based my on the concept of ‘subhomomorphism’ (375E); this seems to be an effective tool when B is weakly (σ, ∞)-distributive (375Yc), but less useful in other cases. The case B = {0, 1}, L0 (B) ∼ = R is not entirely trivial and is worth working through on its own (375Xg). I mention 375C for the sake of its corollary 375D, but have to it that I am not sure it is the best possible result. For the cases of principal interest to a measure theorist, there is something a good deal stronger (375Yd). Further questions concern possible relaxations of the hypotheses of 375B. One has a straightforward resolution (375Yb); others can be reduced to the Banach-Ulam problem by the techniques of 363S (375Ya).

376 Kernel operators The theory of linear integral equations is in large part the theory of operators T defined from formulae of the type (T f )(y) =

R

k(x, y)f (x)dx

for some function k of two variables. I make no attempt to study the general theory here. However, the concepts developed in this book make it easy to discuss certain aspects of such operators defined between the ‘function spaces’ of measure theory, meaning spaces of equivalence classes of functions, and indeed allow us to do some of the work in the abstract theory of Riesz spaces, omitting all formal mention of measures (376D, 376H, 376P). I give a very brief of two theorems characterizing kernel operators in the abstract (376E, 376H), with corollaries to show the form these theorems can take in the ordinary language of integral kernels (376J, 376N). To give an idea of the kind of results we can hope for in this area, I go a bit farther with operators with domain L1 (376Mb, 376P, 376S). I take the opportunity to spell out versions of results from §253 in the language of this volume (376B376C). 376A Kernel operators To give an idea of where this section is going, I will try to describe the central idea in a relatively concrete special case. Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces; you can take them both to be X ×Y. R [0, 1] with Lebesgue measure if you like. Let λ be the product measure on 1 ∞ If k ∈ L (λ), then k(x, y)dx is defined for almost every y, by Fubini’s theorem; so if f ∈ L (µ) then R g(y) = k(x, y)f (x)dx is defined for almost every y. Also

R

g(y)dy =

R

k(x, y)f (x)dxdy

is defined, because (x, y) 7→ k(x, y)f (x) is λ-virtually measurable, defined λ-a.e. and is dominated by a multiple of the integrable function k. Thus k defines a function from L∞ (µ) to L1 (ν). Changing f on a set of measure 0 will not change g, so we can think of this as an operator from L∞ (µ) to L1 (ν); and of course we can move immediately to the equivalence class of g in L1 (ν), so getting an operator Tk from L∞ (µ) to L1 (ν). This operatorRis plainly linear; also it is easy to check that ±Tk ≤ T|k| , so that Tk ∈ L∼ (L∞ (µ); L1 (ν)), and that kTk k ≤ |k|. Moreover, changing k on a λ-negligible set does not change Tk , so that in fact we can speak of Tw for any w ∈ L1 (λ). I think it is obvious, even before investigating them, that operators representable in this way will be important. We can immediately ask what their properties will be and whether there is any straightforward way of recognising them. We can look at the properties of the map w 7→ Tw : L1 (λ) → L∼ (L∞ (µ); L1 (ν)). And we can ask what happens when L∞ (µ) and L1 (ν) are replaced by other function spaces, defined by extended Fatou norms or otherwise. Theorems 376E and 376H answer questions of this kind. R It turns out that the formula g(y) = k(x, y)f (x)dx gives rise to a variety of technical problems, and it is much easier to characterize T u in of its action on the dual. In the language of the special case above, if h ∈ L∞ (ν), then we shall have

R

k(x, y)f (x)h(y)d(x, y) =

R

g(y)h(y)dy;

458


376A

R since g • ∈ L1 (ν) is entirely determined by the integrals g(y)h(y)dy as h runs over L∞ (ν), we can define R the operator T in of the functional (f, h) 7→ k(x, y)f (x)h(y)d(x, y). This enables us to extend the results from the case of σ-finite spaces to general strictly localizable spaces; perhaps more to the point in the present context, it gives them natural expressions in of function spaces defined from measure algebras rather than measure spaces, as in 376E. Before going farther along this road, however, I give a couple of results relating the theorems of §253 to the methods of this volume. 376B The canonical map L0 × L0 → L0 : Proposition Let (A, µ ¯) and (B, ν¯) be semi-finite measure ¯ their localizable measure algebra free product (325E). Then we have a bilinear map algebras, and (C, λ) (u, v) 7→ u ⊗ v : L0 (A) × L0 (B) → L0 (C) with the following properties. (a) For any u ∈ L0 (A), v ∈ L0 (B), α ∈ R, [[u ⊗ χ1B > α]] = [[u > α]] ⊗ 1B ,

[[χ1A ⊗ v > α]] = 1A ⊗ [[v > α]]

where for a ∈ A, b ∈ B I write a ⊗ b for the corresponding member of A ⊗ B (315M), identified with a subalgebra of C (325Dc). (b)(i) For any u ∈ L0 (A)+ , the map v 7→ u ⊗ v : L0 (B) → L0 (C) is an order-continuous multiplicative Riesz homomorphism. (ii) For any v ∈ L0 (B)+ , the map u 7→ u ⊗ v : L0 (A) → L0 (C) is an order-continuous multiplicative Riesz homomorphism. (c) In particular, |u ⊗ v| = |u| ⊗ |v| for all u ∈ L0 (A), v ∈ L0 (B). (d) For any u ∈ L0 (A)+ and v ∈ L0 (B)+ , [[u ⊗ v > 0]] = [[u > 0]] ⊗ [[v > 0]]. proof The canonical maps a 7→ a ⊗ 1B , b 7→ 1A ⊗ b from A, B to C are order-continuous Boolean homomorphisms (325Da), so induce order-continuous multiplicative Riesz homomorphisms from L0 (A) and L0 (B) to L0 (C) (364R); write u ˜, v˜ for the images of u ∈ L0 (A), v ∈ L0 (B). Observe that |˜ u| = |u|∼ , ∼ ∼ ∼ |˜ v | = |v| and (χ1A ) = (χ1B ) = χ1C . Now set u ⊗ v = u ˜ × v˜. The properties listed in (a)-(c) are just a matter of putting the definition in 364Ra together with the fact that L0 (C) is an f -algebra (364E). As for [[u ⊗ v > 0]] = [[˜ u × v˜ > 0]], this is (for non-negative u, v) just [[˜ u > 0]] ∩ [[˜ v > 0]] = ([[u > 0]] ⊗ 1B ) ∩ (1A ⊗ [[v > 0]]) = [[u > 0]] ⊗ [[v > 0]]. 376C For L1 spaces we have a similar result, with additions corresponding to the Banach lattice structures of the three spaces. Theorem Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras with localizable measure algebra free product ¯ (C, λ). ¯ and (a) If u ∈ L1µ¯ = L1 (A, µ ¯) and v ∈ L1ν¯ = L1 (B, ν¯) then u ⊗ v ∈ L1λ¯ = L1 (C, λ)

R

u⊗v =

R R u

v,

ku ⊗ vk1 = kuk1 kvk1 .

(b) Let W be a Banach space and φ : L1µ¯ × L1ν¯ → W a bounded bilinear map. Then there is a unique bounded linear operator T : L1λ¯ → W such that T (u⊗v) = φ(u, v) for all u ∈ L1µ¯ and v ∈ L1ν¯ , and kT k = kφk. (c) Suppose, in (b), that W is a Banach lattice. Then (i) T is positive iff φ(u, v) ≥ 0 for all u, v ≥ 0; (ii) T is a Riesz homomorphism iff u 7→ φ(u, v0 ) : L1µ¯ → W , v 7→ φ(u0 , v) : L1ν¯ → W are Riesz homomorphisms for all v0 ≥ 0 in L1ν¯ and u0 ≥ 0 in L1µ¯ . proof (a) I refer to the proof of 325D. Let (X, Σ, µ) and (Y, T, ν) be the Stone spaces of (A, µ ¯) and (B, ν¯) ¯ can be identified with the measure algebra of the c.l.d. product measure λ on X × Y (321K), so that (C, λ) (part (a) of the proof of 325D), and L1µ¯ , L1ν¯ , L1λ¯ can be identified with L1 (µ), L1 (ν) and L1 (λ) (365B). Now ¯ if f ∈ L0 (µ) and g ∈ L0 (ν) then f ⊗ g ∈ L0 (λ) (253Cb), and it is easy to check that (f ⊗ g)• ∈ L0 (λ) corresponds to f • ⊗ g • as defined in 376B. (Look first at the cases in which one of f , g is a constant function 1 with value R1.) RBy 253E, we have a canonical map (f • , g • ) 7→ (f ⊗ g)• from L1 (µ) × L1 (ν) R R to L (λ), R with R f ⊗ g = f g (253D); so that if u ∈ L1µ¯ and v ∈ L1ν¯ we must have u ⊗ v ∈ L1λ¯ , with u ⊗ v = u v. As in 253E, it follows that ku ⊗ vk1 = kuk1 kvk1 .

376D

Kernel operators

459

(b) In view of the situation described in (a) above, this is now just a translation of the same result about L1 (µ), L1 (ν) and L1 (λ), which is Theorem 253F. (c) Identifying the algebraic free product A ⊗ B with its canonical image in C (325Dc), I write (A ⊗ B)f ¯ < ∞}, so that (A ⊗ B)f is a subring of C. Recall that any member of A ⊗ B is for {c : c ∈ A ⊗ B, λc expressible as supi≤n ai ⊗ bi where a0 , . . . , an are dist (315Na); evidently this will belong to (A ⊗ B)f iff µ ¯ai · ν¯bi is finite for every i. The next fact to lift from previous theorems is in part (e) of the proof of 253F: the linear span M of {χ(a ⊗ b) : a ∈ Af , b ∈ Bf } is norm-dense in L1λ¯ . Of course M can also be regarded as the linear span of {χc : c ∈ (A ⊗ B)f }, or S(A ⊗ B)f . (Strictly speaking, this last remark relies on 361J; the identity map from (A ⊗ B)f to C induces an injective Riesz homomorphism from S(A ⊗ B)f into S(C) ⊆ L0 (C). To see that χc ∈ M for every c ∈ (A ⊗ B)f , we need to know that c can be expressed as a dist union of of A ⊗ B, as noted above.) (i) If T is positive then of course φ(u, v) = T (u ⊗ v) ≥ 0 whenever u, v ≥ 0, since u ⊗ v ≥ 0. On the other hand, if φ is non-negative on U + × V + , then, in particular, T χ(a ⊗ b) = φ(χa, χb) ≥ 0 whenever µ ¯a· ν¯b < ∞. Consequently T (χc) ≥ 0 for every c ∈ (A⊗B)f and T w ≥ 0 whenever w ≥ 0 in M ∼ = S(A⊗B)f , as in 361Ga. Now this means that T |w| ≥ 0 whenever w ∈ M . But as M is norm-dense in L1λ¯ , w 7→ T |w| is continuous and W + is closed, it follows that T |w| ≥ 0 for every w ∈ L1λ¯ , that is, that T is positive. (ii) If T is a Riesz homomorphism then of course u 7→ φ(u, v0 ) = T (u ⊗ v0 ), v 7→ φ(u0 , v) = T (u0 ⊗ v) are Riesz homomorphisms for v0 , u0 ≥ 0. On the other hand, if all these maps are Riesz homomorphisms, then, in particular, T χ(a ⊗ b) ∧ T χ(a0 ⊗ b0 ) = φ(χa, χb) ∧ φ(χa0 , χb0 ) ≤ φ(χa, χb + χb0 ) ∧ φ(χa0 , χb + χb0 ) = φ(χa ∧ χa0 , χb + χb0 ) = 0 whenever a, a0 ∈ Af , b, b0 ∈ Bf and a ∩ a0 = 0. Similarly, T χ(a ⊗ b) ∧ T χ(a0 ⊗ b0 ) = 0 if b ∩ b0 = 0. But this means that T χc ∧ T χc0 = 0 whenever c, c0 ∈ (A ⊗ B)f and c ∩ c0 = 0. P P Express c, c0 as supi≤m ai ⊗ bi , 0 0 0 0 supj≤n aj ⊗ bj where ai , aj , bi , bj all have finite measure. Now if i ≤ m, j ≤ n, (ai ∩ a0j ) ⊗ (bi ∩ b0j ) = (ai ⊗bi ) ∩ (a0j ⊗b0j ) = 0, so one of ai ∩ a0j , bi ∩ b0j must be zero, and in either case T χ(ai ⊗bi )∧T χ(a0j ⊗b0j ) = 0. Accordingly m n X X T χc ∧ T χc0 ≤ ( T χ(ai ⊗ bi )) ∧ ( T χ(a0j × b0j )) i=0

≤

X

j=0

T χ(ai ⊗ bi ) ∧ T χ(a0j ⊗ b0j ) = 0,

i≤m,j≤n

using 352Fa for the second inequality. Q Q This implies that T ¹M must be a Riesz homomorphism (361Gc), that is, T |w| = |T w| for all w ∈ M . Again because M is dense in L1λ¯ , T |w| = |T w| for every w ∈ L1λ¯ , and T is a Riesz homomorphism. 376D Abstract integral operators: Definition The following concept will be used repeatedly in the theorems below; it is perhaps worth giving it a name. Let U be a Riesz space and V a Dedekind complete Riesz space, so that L× (U ; V ) is a Dedekind complete Riesz space (355H). If f ∈ U × and v ∈ V write Pf v u = f (u)v for each u ∈ U ; then Pf v ∈ L× (U ; V ). P P If f ≥ 0 in U × and v ≥ 0 in V × then Pf v is a positive linear operator from U to V which is order-continuous because if A ⊆ U is non-empty, downwards-directed and has infimum 0, then (as V is Archimedean) inf u∈A Pf v (u) = inf u∈A f (u)v = 0. Of course (f, g) 7→ Pf g is bilinear, so Pf v ∈ L× (U ; V ) for every f ∈ U × , v ∈ V . Q Q Now I call a linear operator from U to V an abstract integral operator if it is in the band of L× (U ; V ) generated by {Pf v : f ∈ U × , v ∈ V }.

460


376D

The first result describes these operators when U , V are expressed as subspaces of L0 (A), L0 (B) for measure algebras A, B and V is perfect. 376E Theorem Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras, with localizable measure algebra ¯ and U ⊆ L0 (A), V ⊆ L0 (B) order-dense Riesz subspaces. Write W for the set of free product (C, λ), those w ∈ L0 (C) such that w × (u ⊗ v) is integrable for every u ∈ U , v ∈ V . Then we have an operator w 7→ Tw : W → L× (U ; V × ) defined by setting Tw (u)(v) =

R

w × (u ⊗ v)

for every w ∈ W , u ∈ U and v ∈ V . The map w 7→ Tw is a Riesz space isomorphism between W and the band of abstract integral operators in L× (U ; V × ). × proof (a) The first thing to check is that the formula offered does R define a member Tw (u) of V for any w ∈ W, u ∈ U. P P Of course Tw (u) isR a linear operator because is linear and ⊗ and × are bilinear. It belongs to V ∼ because, writing g(v) = |w|×(|u|⊗v), g is a positive linear operator and |Tw (u)(v)| ≤ g(|v|) for every v. (I am here using 376Bc to see that |w × (u ⊗ v)| = |w| × (|u| ⊗ |v|).) Also g ∈ V × because R 0 0 v 7→ |u| ⊗ v, w 7→ |w| × w and are all order-continuous; so Tw (u) also belongs to V × . Q Q

(b) Next, for any given w ∈ W , the map Tw : U → V × is linear (again because ⊗ and × are bilinear). It is helpful to note that W is a solid linear subspace of L0 (C). Now if w ≥ 0 in W , then Tw ∈ L× (U ; V × ). P P If u, v ≥ 0 then u ⊗ v ≥ 0, w × (u ⊗ v) ≥ 0 and Tw (u)(v) ≥ 0; as v is arbitrary, Tw (u) ≥ 0 whenever u ≥ 0; as u is arbitrary, Tw is positive. If A ⊆ U is non-empty, downwards-directed and has infimum 0, then Tw [A] is downwards-directed, and for any v ∈ V + (inf Tw [A])(v) = inf u∈A Tw (u)(v) = inf u∈A

R

w × (u ⊗ v) = 0

because u 7→ u ⊗ v is order-continuous. So inf Tw [A] = 0; as A is arbitrary, Tw is order-continuous. Q Q For general w ∈ W , we now have Tw = Tw+ − Tw− ∈ L× (U ; V × ). (c) Ths shows that w 7→ Tw is a map from W to L× (U ; V × ). Running through the formulae once again, it is linear, positive and order-continuous; this last because, given a non-empty downwards-directed C ⊆ W with infimum 0, then for any u ∈ U + , v ∈ V + R

(because 0.

(inf w∈C Tw )(u)(v) ≤ inf w∈C

R

w × (u ⊗ v) = 0

and × are order-continuous); as v is arbitrary, (inf w∈C Tw )(u) = 0; as u is arbitrary, inf w∈C Tw =

(d) R All this is easy, being nothing but a string of applications of the elementary properties of ⊗, × and . But I think a new idea is needed for the next fact: the map w 7→ Tw : W → L× (U ; V × ) is a Riesz homomorphism. P P Write D for the set of those d ∈ C such that Tw ∧ Tw0 = 0 whenever w, w0 ∈ W + , [[w > 0]] ⊆ d and [[w0 > 0]] ⊆ 1C \ d. (i) If d1 , d2 ∈ D, w, w0 ∈ W + , [[w > 0]] ⊆ d1 ∪ d2 and [[w0 > 0]] ∩ (d1 ∪ d2 ) = 0, then set w1 = w × χd1 , w2 = w − w1 . In this case [[w1 > 0]] ⊆ d1 ,

[[w2 > 0]] ⊆ d2 ,

so Tw1 ∧ Tw0 = Tw2 ∧ Tw0 = 0,

Tw ∧ Tw0 ≤ (Tw1 ∧ Tw0 ) + (Tw2 ∧ Tw0 ) = 0.

0

As w, w are arbitrary, d1 ∪ d2 ∈ D. Thus D is closed under ∪ . (ii) The symmetry of the definition of D means that 1C \ d ∈ D whenever d ∈ D. (iii) Of course 0 ∈ D, just because Tw = 0 if w ∈ W + and [[w > 0]] = 0; so D is a subalgebra of C. (iv) If D ⊆ D is non-empty and upwards-directed, with supremum c in C, and if w, w0 ∈ W + are such that [[w > 0]] ⊆ c, [[w0 > 0]] ∩ c = 0, then consider {w × χd : d ∈ D}. This is upwards-directed, with supremum w; so Tw = supd∈D Tw×χd , because the map q 7→ Tq is order-continuous. Also Tw×χd ∧ Tw0 = 0 for every d ∈ D, so Tw ∧ Tw0 = 0. As w, w0 are arbitrary, c ∈ D; as D is arbitrary, D is an order-closed subalgebra of C. (v) If a ∈ A and w, w0 ∈ W + are such that [[w > 0]] ⊆ a ⊗ 1B , [[w0 > 0]] ∩ (a ⊗ 1B ) = 0, then any u ∈ U + is expressible as u1 + u2 where u1 = u × χa, u2 = u × χ(1A \ a). Now Tw (u2 )(v) =

R

w × (u2 ⊗ v) =

R

w × χ(a ⊗ 1B ) × (u ⊗ v) × χ((1A \ a) ⊗ 1B ) = 0

376E

Kernel operators

461

for every v ∈ V , so Tw (u2 ) = 0. Similarly, Tw0 (u1 ) = 0. But this means that (Tw ∧ Tw0 )(u) ≤ Tw (u2 ) + Tw0 (u1 ) = 0. As u is arbitrary, Tw ∧ Tw0 = 0; as w and w0 are arbitrary, a ⊗ 1B ∈ D. (vi) Now suppose that b ∈ B and that w, w0 ∈ W + are such that [[w > 0]] ⊆ 1A ⊗ b, [[w0 > 0]] ∩ (1A ⊗ b) = 0. If u ∈ U + , v ∈ V + then (Tw ∧ Tw0 )(u)(v) ≤

R

w × (u ⊗ (v × χ(1B \ b))) +

R

w0 × (u ⊗ (v × χb)) = 0.

As u, v are arbitrary, Tw ∧ Tw0 = 0; as w and w0 are arbitrary, 1A ⊗ b ∈ D. (vii) This means that D is an order-closed subalgebra of C including A ⊗ B, and is therefore the whole of C (325D(c-ii)). (viii) Now take any w, w0 ∈ W such that w ∧ w0 = 0, and consider c = [[w > 0]]. Then [[w0 > 0]] ⊆ 1C \ c and c ∈ D, so Tw ∧ Tw0 = 0. This is what we need to be sure that w 7→ Tw is a Riesz homomorphism (352G). Q Q (e) The map w 7→ Tw is injective. P P (i) If w > 0 in W , then consider A = {a : a ∈ A, ∃ u ∈ U, χa ≤ u}, 0

B = {b : b ∈ B, ∃ v ∈ V, χb ≤ v}.

0

Because U and V are order-dense in L (A) and L (B) respectively, RA and B are order-dense in A and B. R Also both are upwards-directed. So supa∈A,b∈B a ⊗ b = 1C and 0 < w = supa∈A,b∈B a⊗b w. Take a ∈ A, R b ∈ B such that a⊗b w > 0; then there are u ∈ U , v ∈ V such that χa ≤ u, χb ≤ v, so that Tw (u)(v) ≥

R

a⊗b

w>0

and Tw > 0. (ii) For general non-zero w ∈ W , we now have |Tw | = T|w| > 0 so Tw 6= 0. Q Q Thus w 7→ Tw is an order-continuous injective Riesz homomorphism. ˜ for {Tw : w ∈ W }, so that W ˜ is a Riesz subspace of L× (U ; V × ) isomorphic to W , and (f ) Write W × × c ˜ is order-dense in W c. P W for the band it generates in L (U ; V ). Then W P Suppose that S > 0 in ⊥⊥ ⊥ c=W ˜ ˜ , so there is a w ∈ W such that S ∧ Tw > 0. Set w1 = w ∧ χ1C . Then W (353Ba). Then S ∈ /W w = supn∈N w ∧ nw1 , so Tw = supn∈N Tw ∧ nTw1 and R = S ∧ Tw1 > 0. Set U1 = U ∩ L1 (A, µ ¯). Because U is an order-dense Riesz subspace of L0 (A), U1 is an order-dense Riesz 1 1 subspace of Lµ¯ = L (A, µ ¯), therefore also norm-dense. Similarly V1 = V ∩ L1 (B, ν¯) is a norm-dense Riesz subspace of L1ν¯ = L1 (B, ν¯). Define φ0 : U1 × V1 → R by setting φ0 (u, v) = R(u)(v) for u ∈ U1 , v ∈ V1 . Then φ0 is bilinear, and |φ0 (u, v)| = |R(u)(v)| ≤ |R(u)|(|v|) ≤ R(|u|)(|v|) ≤ Tw1 (|u|)(|v|) Z Z = w1 × (|u| ⊗ |v|) ≤ |u| ⊗ |v| = kuk1 kvk1 for all u ∈ U1 , v ∈ V1 , because 0 ≤ R ≤ Tw1 in L× (U ; V × ). Because U1 , V1 are norm-dense in L1µ¯ , L1ν¯ respectively, φ0 has a unique extension to a continuous bilinear map φ : L1µ¯ × L1ν¯ → R. (To reduce this to standard results on linear operators, think of R as a function from U1 to V1∗ ; since every member of V1∗ has a unique extension to a member of (L1ν¯ )∗ , we get a corresponding function R1 : U1 → (L1ν¯ )∗ which is continuous and linear, so has a unique extension to a continuous linear operator R2 : L1µ¯ → (L1ν¯ )∗ , and we set φ(u, v) = R2 (u)(v).) ¯ ∗ such that h(u ⊗ v) = φ(u, v) for every u ∈ L1 , v ∈ L1 . By 376C, there is a unique h ∈ (L1λ¯ )∗ = L1 (C, λ) µ ¯ ν ¯ ¯ Because (C, λ) is localizable, this h corresponds to a w0 ∈ L∞ (C) (365Jc), and

R

w0 × (u ⊗ v) = h(u ⊗ v) = φ0 (u, v) = R(u)(v)

for every u ∈ U1 , v ∈ V1 . Because U1 is norm-dense in L1µ¯ , U1+ is dense in (L1µ¯ )+ , and similarly V1+ is dense in (L1ν¯ )+ , so U1+ × V1+ is dense in (L1µ¯ )+ × (L1ν¯ )+ ; now φ0 is non-negative on U1+ × V1+ , so φ (being continuous) is non-negative on (L1µ¯ )+ ×(L1ν¯ )+ . By 376Cc, h ≥ 0 in (L1λ¯ )∗ and w0 ≥ 0 in L∞ (C). In the same way, because φ0 (u, v) ≤ Tw (u)(v) for u ∈ U1+ and v ∈ V1+ , w0 ≤ w1 ≤ w in L0 (C), so w0 ∈ W . We have Tw0 (u)(v) =

R

w0 × (u ⊗ v) = R(u)(v)

for all u ∈ U1 , v ∈ V1 . If u ∈ U1+ , then Tw0 (u) and R(u) are both order-continuous, so must be identical, since V1 is order-dense in V . This means that Tw0 and R agree on U1 . But as both are themselves order-continuous

462


376E

linear operators, and U1 is order-dense in U , they must be equal. ˜ is quasi-order-dense in W c , therefore order-dense Thus 0 < Tw0 ≤ S in L× (U ; V × ). As S is arbitrary, W (353A). Q Q ˜ is an injective Riesz homomorphism, we have an inverse map Q : (g) Because w 7→ Tw : W 7→ W ˜ → L0 (C), setting Q(Tw ) = w; this is a Riesz homomorphism, and it is order-continuous because W is W solid in L0 (C), so that the embedding W ⊆ L0 (C) is order-continuous. By 368B, Q has an extension to ˜ : W c → L0 (C). Because Q(S) > 0 whenever S > 0 in W ˜, an order-continuous Riesz homomorphism Q ˜ c ˜ ˜ c Q(S) > 0 whenever S > 0 in W , so Q is injective. Now Q(S) ∈ W for every S ∈ W . P P It is enough ˜ ˜ w ) : w ∈ W, Tw ≤ S} = sup C, where to look at non-negative S. In this case, Q(S) must be sup{Q(T C = {w : Tw ≤ S} ⊆ W . Take u ∈ U + , v ∈ V + . Then {w × (u ⊗ v) : w ∈ C} is upwards-directed, because C is, and supw∈C

R

w × (u ⊗ v) = supw∈C Tw (u)(v) ≤ S(u)(v) < ∞.

˜ ˜ So Q(S) × (u ⊗ v) = supw∈C w × (u ⊗ v) belongs to L1λ¯ (365Df). As u and v are arbitrary, Q(S) ∈ W. Q Q ˜ =W c and Q ˜ = Q, that is, that w 7→ Tw : W 7→ W c is a Riesz space (h) Of course this means that W isomorphism. c as the band Z of abstract integral operators in (i) I have still to check on the identification of W × × × L (U ; V ). Write Pf g (u) = f (u)g for f ∈ U , g ∈ V and u ∈ U . ×

Set U # = {u : u ∈ L0 (A), u × u0 ∈ L1µ¯ for every u0 ∈ U }, V # = {v : v ∈ L0 (B), v × v 0 ∈ L1ν¯ for every v 0 ∈ V }. R From 369C we know that if we set fu (u0 ) = u × u0 for u ∈ U # and u0 ∈ U , then fu ∈ U × for every u ∈ U # , and u 7→Rfu is an isomorphism between U # and an order-dense Riesz subspace of U × . Similarly, setting gv (v 0 ) = v × v 0 for v ∈ V # and v 0 ∈ V , v 7→ gv is an isomorphism between V # and an order-dense Riesz subspace of V × . If u ∈ U # , v ∈ V # then

R

(u ⊗ v) × (u0 ⊗ v 0 ) =

R

R

R

(u × u0 ) ⊗ (v × v 0 ) = ( u × u0 )( v × v 0 ) = fu (u0 )gv (v 0 )

for every u0 ∈ U , v 0 ∈ V , so u ⊗ v ∈ W and Tu⊗v = Pfu gv . Now take f ∈ (U × )+ and g ∈ (V × )+ . Set A = {u : u ∈ U # , u ≥ 0, fu ≤ f } and B = {v : v ∈ V # , v ≥ 0, gv ≤ g}. These are upwards-directed, so C = {u ⊗ v : u ∈ A, v ∈ B} is upwards-directed in L0 (C). Because {fu : u ∈ U # } is order-dense in U × , f = supu∈A fu ; by 355Ed, f (u0 ) = supu∈A fu (u0 ) for every u0 ∈ U + . Similarly, g(v 0 ) = supv∈B fv (v 0 ) for every v 0 ∈ V + . ?? Suppose, if possible, that C is not bounded above in L0 (C). Because C and L0 (C) are Dedekind complete, c = inf n∈N supu∈A,v∈B [[u ⊗ v ≥ n]] must be non-zero (364Ma). Because U and V are order-dense in L0 (A), L0 (B) respectively, 1A = sup{[[u0 > 0]] : u0 ∈ U }, and there are u0 ∈ U + , v 0 ∈ V + any n ∈ N,

1B = sup{[[v 0 > 0]] : v 0 ∈ V }, R such that c ∩ [[u0 > 0]] ⊗ [[v 0 > 0]] 6= 0, so that c u0 ⊗ v 0 > 0. But now, for

376E

Kernel operators

f (u0 )g(v 0 ) ≥

463

fu (u0 )gv (v 0 ) Z = sup (u ⊗ v) × (u0 ⊗ v 0 ) u∈A,v∈B Z ≥ sup ((u ⊗ v) ∧ nχc) × (u0 ⊗ v 0 ) u∈A,v∈B Z = sup ((u ⊗ v) ∧ nχc) × (u0 ⊗ v 0 ) sup

u∈A,v∈B

(because w 7→

R

u∈A,v∈B 0

0

w × (u ⊗ v ) is order-continuous) Z Z = (nχc) × (u0 ⊗ v 0 ) = n u0 ⊗ v 0 , c

which is impossible. X X Thus C is bounded above in L0 (C), and has a supremum w ∈ L0 (C). If u0 ∈ U + , v 0 ∈ V + then Z

Z 0

0

w × (u ⊗ v ) =

sup

(u ⊗ v) × (u0 ⊗ v 0 )

u∈A,v∈B

=

sup u∈A,v∈B

fu (u0 )gv (v 0 ) = f (u0 )g(v 0 ) = Pf g (u0 )(v 0 ).

Thus w ∈ W and ˜ ⊆W c. Pf g = Tw ∈ W c for every f ∈ U × , And this is true for any non-negative f ∈ U × , g ∈ V × . Of course it follows that Pf g ∈ W c is a band, it must include Z. g ∈ V × ; as W c ⊆ Z. P c ∩ Z ⊥ = {0}. Take any T > 0 in (j) Finally, W P Since Z = Z ⊥⊥ , it is enough to show that W c . There are u0 ∈ U + , v 0 ∈ V + such that T (u0 )(v 0 ) > 0. So there is a v ∈ V # such that 0 ≤ gv ≤ T (u0 ) W 0 0 0 0 R0 and gv (v00 ) > 0, that is, v × v00 > 0. Because V is order-dense in L0 (B), there is a v10 ∈ V such that 0 < v10 ≤ v00 × χ[[v > 0]], so that 0<

R

v × v10 = gv (v10 ) ≤ T (u00 )(v10 )

and [[v10 > 0]] ⊆ [[v > 0]]. Now consider the functional u0 7→ h(u0 ) = T (u0 )(v10 ) : U → R. This belongs to (U × )+ and h(u00 ) > 0, R # 0 so there is a u ∈ U such that 0 ≤ fu ≤ h and fu (u0 ) > 0. This time, u × u00 > 0 so (because U is order-dense in L0 (A)) there is a u01 ∈ U such that h(u01 ) > 0 and [[u01 > 0]] ⊆ [[u > 0]]. We can express T as Tw where w ∈ W + . In this case, we have

R

w × (u01 ⊗ v10 ) = T (u01 )(v10 ) = h(u01 ) > 0,

so 0 6= [[w > 0]] ∩ [[u01 ⊗ v10 > 0]] = [[w > 0]] ∩ ([[u01 > 0]] ⊗ [[v10 > 0]]) ⊆

[[w > 0]] ∩ ([[u > 0]] ⊗ [[v > 0]]) = [[w > 0]] ∩ [[u ⊗ v > 0]],

and w ∧ (u ⊗ v) > 0, so Tw ∧ Pfu gv = Tw ∧ Tu⊗v = Tw∧(u⊗v) > 0. c ∩ Z ⊥ = {0} and W c ⊆ Z ⊥⊥ = Z. Q Thus T ∈ / Z ⊥ . Accordingly W Q c Since we already know that Z ⊆ W , this completes the proof.

464


376F

376F Corollary Let (A, µ ¯) and (B, ν¯) be localizable measure spaces, with localizable measure algebra ¯ Let U ⊆ L0 (A), V ⊆ L0 (B) be perfect order-dense solid linear subspaces, and T : U → V free product (C, λ). a linear operator. Then the following are equiveridical: (i) T is an abstract integral operator; R R (ii) there is a w ∈ L0 (C) such that w × (u ⊗ v 0 ) is defined and equal to T u × v 0 whenever u ∈ U and v 0 ∈ L0 (B) is such that v 0 × v is integrable for every v ∈ V . proof Setting V # = {v 0 : v 0 ∈ L0 (B), v × v 0 ∈ L1 for every v ∈ V }, we know that we can identify V # with V × and V with (V # )× (369C). So the equivalence of (i) and (ii) is just 376E applied to V # in place of V . 376G Lemma Let U be a Riesz space, V an Archimedean Riesz space, T : U → V a linear operator, f ∈ (U ∼ )+ and e ∈ V + . Suppose that 0 ≤ T u ≤ f (u)e for every u ∈ U + . Then if hun in∈N is a sequence in U such that limn→∞ g(un ) = 0 whenever g ∈ U ∼ and |g| ≤ f , hT un in∈N order*-converges to 0 in V (definition: 367A). proof Let Ve be the solid linear subspace of V generated by e; then T u ∈ Ve for every u ∈ U . We can identify Ve with an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdorff space, with e corresponding to χX (353M). For x ∈ X, set gx (u) = (T u)(x) for every u ∈ U ; then 0 ≤ gx (u) ≤ f (u) for u ≥ 0, so |gx | ≤ f and limn→∞ (T un )(x) = 0. As x is arbitrary, hT un in∈N order*-converges to 0 in C(X), by 367L, and therefore in Ve , because Ve is order-dense in C(X) (367F). But Ve , regarded as a subspace of V , is solid, so 367F also tells us that hT un in∈N order*-converges to 0 in V . 376H Theorem Let U be a Riesz space and V a weakly (σ, ∞)-distributive Dedekind complete Riesz space (definition: 368N). Suppose that T ∈ L× (U ; V ). Then the following are equiveridical: (i) T is an abstract integral operator; (ii) whenever hun in∈N is an order-bounded sequence in U + and limn→∞ f (un ) = 0 for every f ∈ U × , then hT un in∈N order*-converges to 0 in V ; (iii) whenever hun in∈N is an order-bounded sequence in U and limn→∞ f (un ) = 0 for every f ∈ U × , then hT un in∈N order*-converges to 0 in V . proof For f ∈ U × , v ∈ V and u ∈ U set Pf v (u) = f (u)v. Write Z ⊆ L× (U ; V ) for the space of abstract integral operators. (a)(i)⇒(iii) Suppose that T ∈ Z + , and that hun in∈N is an order-bounded sequence in U such that limn→∞ f (un ) = 0 for every f ∈ U × . Note that {Pf v : f ∈ U ×+ , v ∈ V + } is upwards-directed, so that T = sup{T ∧ Pf v : f ∈ U ×+ , v ∈ V + } (352Va). Take u∗ ∈ U + such that |un | ≤ u∗ for every n, and set w = inf n∈N supm≥n T um , which is defined because |T un | ≤ T u∗ for every n. Now w ≤ (T − Pf v )+ (u∗ ) for every f ∈ U ×+ , v ∈ V + . P P Setting T1 = T ∧ Pf v , w0 = (T − Pf v )+ (u∗ ) we have T un − T1 un ≤ |T − T1 |(u∗ ) = (T − Pf v )+ (u∗ ) = w0 for every n ∈ N, so T un ≤ w0 + T1 un . On the other hand, 0 ≤ T1 u ≤ f (u)v for every u ∈ U + , so by 376G we must have inf n∈N supm≥n T1 um = 0. Accordingly w ≤ w0 + inf n∈N supm≥n T1 um = w0 . Q Q But as inf{(T − Pf v )+ : f ∈ U ×+ , v ∈ V + } = 0, w ≤ 0. Similarly (or applying the same argument to h−un in∈N ), supn∈N inf n∈N T un ≥ 0 and hT un in∈N order*-converges to zero. For general T ∈ Z, this shows that hT + un in∈N and hT − un in∈N both order*-converge to 0, so hT un in∈N order*-converges to 0, by 367C. As hun in∈N is arbitrary, (iii) is satisfied. (b)(iii)⇒(ii) is trivial. (c)(ii)⇒(i) ?? Now suppose, if possible, that (ii) is satisfied, but that T ∈ / Z. Because L× (U ; V ) is Dedekind complete (355H), Z is a projection band (353I), so T is expressible as T1 + T2 where T1 ∈ Z, T2 ∈ Z ⊥ and T2 6= 0. At least one of T2+ , T2− is non-zero; replacing T by −T if need be, we may suppose that T2+ > 0.

376I

Kernel operators

465

Because T2+ , like T , belongs to L× (U ; V ), its kernel U0 is a band in U , which cannot be the whole of U , and there is a u0 > 0 in U0⊥ . In this case T2+ u0 > 0; because T2+ ∧ (T2− + |T1 |) = 0, there is a u1 ∈ [0, u0 ] such that T2+ (u0 − u1 ) + (T2− + |T1 |)(u1 ) 6≥ T2+ u0 , so that T u1 ≥ T2 u1 − |T1 |(u1 ) 6≤ 0 and T u1 6= 0. Now this means that the sequence (T u1 , T u1 , . . . ) is not order*-convergent to zero, so there must be some f ∈ U × such that (f (u1 ), f (u1 ), . . . ) does not converge to 0, that is, f (u1 ) 6= 0; replacing f by |f | if necessary, we may suppose that f ≥ 0 and that f (u1 ) > 0. By 356H, there is a u2 such that 0 < u2 ≤ u1 and g(u2 ) = 0 whenever g ∈ U × and g ∧ f = 0. Because 0 < u2 ≤ u0 , u2 ∈ U0⊥ and v0 = T2+ u2 > 0. Consider Pf v0 ∈ Z. Because T2 ∈ Z ⊥ , T2+ ∧ Pf v0 = 0; set S = Pf v0 + T2− , so that T2+ ∧ S = 0. Then inf u∈[0,u2 ] T2+ (u2 − u) + Su = 0,

supu∈[0,u2 ] T2+ u − Su = v0

(use 355Ec for the first equality, and then subtract both sides from v0 ). Now Su ≥ f (u)v0 for every u ≥ 0, so that for any ² > 0 supu∈[0,u2 ],f (u)≥² T2+ u − Su ≤ (1 − ²)v0 and accordingly supu∈[0,u2 ],f (u)≤² T2+ u = v0 , since the of these two suprema is surely at least v0 , while the second is at most v0 . Note also that v0 = supu∈[0,u2 ],f (u)≤² T2+ u = sup0≤u0 ≤u≤u2 ,f (u)≤² T2 u0 = sup0≤u0 ≤u2 ,f (u0 )≤² T2 u0 . For k ∈ N set Ak = {u : 0 ≤ u ≤ u2 , f (u) ≤ 2−k }. We know that Bk = {supu∈I T2 u : I ⊆ Ak is finite} is an upwards-directed set with supremum v0 for each k. Because V is weakly (σ, ∞)-distributive, we can find a sequence hvk0 ik∈N such that vk0 ∈ Bk for every k and v1 = inf k∈N vk0 > 0. For each k let Ik ⊆ Ak be a finite set such that vk0 = supu∈Ik T2 u. Because each Ik is finite, we can build a sequence hu0n in∈N in [0, u2 ] enumerating each in turn, so that limn→∞ f (u0n ) = 0 (since f (u) ≤ 2−k if u ∈ Ik ) while supm≥n T2 u0m ≥ v1 for every n (since {u0m : m ≥ n} always includes some Ik ). Now hT2 u0n in∈N does not order*-converge to 0. However, limn→∞ g(u0n ) = 0 for every g ∈ U × . P P Express |g| as g1 + g2 where g1 belongs to the band of U × generated by f and g2 ∧ f = 0 (353Hc). Then g2 (u0n ) = g2 (u2 ) = 0 for every n, by the choice of u2 . Also g1 = supn∈N g1 ∧ nf (352Vb); so, given ² > 0, there is an m ∈ N such that (g1 − mf )+ (u2 ) ≤ ² and (g1 − mf )+ (u0n ) ≤ ² for every n ∈ N. But this means that |g(u0n )| ≤ |g|(u0n ) ≤ ² + mf (u0n ) for every n, and lim supn→∞ |g(u0n )| ≤ ²; as ² is arbitrary, limn→∞ g(u0n ) = 0. Q Q 0 Now, however, part (a) of this proof tells us that hT1 un in∈N is order*-convergent to 0, because T1 ∈ Z, while hT u0n in∈N is order*-convergent to 0, by hypothesis; so hT2 u0n in∈N = hT u0n − T1 u0n in∈N order*-converges to 0. X X This contradiction shows that every operator satisfying the condition (ii) must be in Z. 376I

The following elementary remark will be useful for the next corollary and also for Theorem 376S.

Lemma Let (X, Σ, µ) be a σ-finite measure space and U an order-dense solid linear subspace of L0 (µ). Then there is a non-decreasing sequence hXn in∈N of measurable subsets of X, with union X, such that χXn• ∈ U for every n ∈ N. proof Write A for the measure algebra of µ, so that L0 (µ) can be identified with L0 (A) (364Jc). A = {a : a ∈ A \ {0}, χa ∈ U } is order-dense in A, so includes a partition of unity hai ii∈I . Because µ is σ-finite, A is • ccc (322G) and I isScountable, so we can S take I to be a subset of N. Choose Ei ∈ Σ such that Ei = ai for i ∈ I; set E = X \ i∈I Ei , Xn = E ∪ i∈I,i≤n Ei for n ∈ N.

466


376J

376J Corollary Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces, with product measure λ on X × Y . Let U ⊆ L0 (µ), V ⊆ L0 (ν) be perfect order-dense solid linear subspaces, and T : U → V a linear operator. Write U = {f : f ∈ L0 (µ), f • ∈ U }, V# = {h : h ∈ L0 (ν), h• × v ∈ L1 for every v ∈ V }. Then the following are equiveridical: (i) T is an abstract integral operator; 0 (ii) there R is a k ∈ L (λ) such that (α) |k(x, y)f (x)h(y)|d(x, y)
R

F

g˜ =

R

T u × (χF )• =

R

F

g.

In particular, this is true whenever F ∈ T and F ⊆ Fn . So g and g˜ agree almost everywhere on Fn , for each n, and g = g˜ a.e. Thus g also represents T u, as required in (ii-β). (b)(ii)⇒(i) Set w = k • in L0 (λ). If f ∈ U and h ∈ V# the hypothesis (α) tells us that (x, y) 7→ k(x, y)f (x)h(y) is integrable (because it surely belongs to L0 (λ)). By Fubini’s theorem,

R

k(x, y)f (x)h(y)d(x, y) =

R

g(y)h(y)dy

R

where g(y) = k(x, y)f (x)dx for almost every y, so that T f • = g • , by (β). But this means that, setting u = f • , v 0 = h• ,

R

w × (u ⊗ v 0 ) =

R

T u × v0 ;

and this is true for every u ∈ U , v 0 ∈ V # . Thus T satisfies the condition 376F(ii), and is an abstract integral operator. (b)(i)⇒(iii) Because V is weakly (σ, ∞)-distributive (368S), this is covered by 376H(i)⇒(iii). (c)(iii)⇒(i) Suppose that T satisfies (iii). The point is that T + is order-continuous. P P?? Otherwise, let A ⊆ U be a non-empty downwards-directed set, with infimum 0, such that v0 = inf u∈A T + (u) > 0. Let hXn in∈N be a non-decreasing sequence of sets of finite measure covering X, and set an = Xn• for each n. For each n, inf u∈A [[u > 2−n ]] = 0, so we can find u ñ ∈ A such that µ ¯(an ∩ [[˜ un > 2−n ]]) ≤ 2−n . Set un = inf i≤n u ˜i for each n; then hun in∈N is non-increasing and has infimum 0; also, [0, un ] meets A for each n, so that v0 ≤ sup{T u : 0 ≤ u ≤ un } for each n. Because V is weakly (σ, ∞)-distributive, we can find + a sequence hIS > 0. n in∈N of finite sets such that In ⊆ [0, un ] for each n and v1 = inf n∈N supu∈In (T u) 0 0 Enumerating n∈N In as hun in∈N , as in part (d) of the proof of 376H, we see that hun in∈N is order-bounded

376M

Kernel operators

467

and limn→∞ f (u0n ) = 0 for every f ∈ U × (indeed, hu0n in∈N order*-converges to 0 in U ), while hT u0n in∈N 6→ 0 in V . X XQ Q Similarly, T − is order-continuous, so T ∈ L× (U ; V ). Accordingly T is an abstract integral operator by condition (ii) of 376H. 376K As an application of the ideas above, I give a result due to N.Dunford (376N) which was one of the inspirations underlying the theory. Following the method of Zaanen 83, I begin with a couple of elementary lemmas. Lemma Let U and V be Riesz spaces. Then there is a Riesz space isomorphism T 7→ T 0 : L× (U ; V × ) → L× (V ; U × ) defined by the formula (T 0 v)(u) = (T u)(v) for every u ∈ U , v ∈ V . If for f ∈ U × , g ∈ V × we write Pf g (u) = f (u)g for every u ∈ U , then Pf g ∈ L× (U ; V × ) and Pf0 g = Pgf in L× (V ; U × ). Consequently T is an abstract integral operator iff T 0 is. proof All the ideas involved have already appeared. For positive T ∈ L× (U ; V × ) the functional (u, v) 7→ (T u)(v) is bilinear and order-continuous in each variable separately; so (just as in the first part of the proof of 376E) corresponds to a T 0 ∈ L× (V ; U × ). The map T 7→ T 0 : L× (U ; V × )+ → L× (V ; U × )+ is evidently an additive, order-preserving bijection, so extends to an isomorphism between L× (U ; V × ) and L× (V ; U × ) given by the same formula. I remarked in part (i) of the proof of 376E that every Pf g belongs to L× (U ; V × ), and the identification Pf0 g = Pgf is just a matter of checking the formulae. Of course it follows at once that the bands of abstract integral operators must also be matched by the map T 7→ T 0 . 376L Lemma Let U be a Riesz space with an order-continuous norm. If w ∈ U + there is a g ∈ U × such that for every ² > 0 there is a δ > 0 such that kuk ≤ ² whenever 0 ≤ u ≤ w and g(u) ≤ δ. proof (a) As remarked in 356D, U ∗ = U ∼ = U × . Set A = {v : v ∈ U and there is an f ∈ (U × )+ such that f (u) > 0 whenever 0 0 whenever 0 0 whenever 0 0 whenever 0 0 in U , there is a v ∈ A such that 0 < v ≤ v0 . P P Because U × = U ∗ separates the points of U , there is a g > 0 in × U such that g(v0 ) > 0; now by 356H there is a v ∈ ]0, v0 ] such that g is strictly positive on ]0, v], so that v ∈ A. Q Q But this means that A is an order-dense solid linear subspace of U . (b) In fact w ∈ A. P P w = sup B, where B = A ∩ [0, w]. Because B is upwards-directed, w ∈ B (354Ea), and there is a sequence hu0n in∈N in B converging to w for the norm. For each n, choose fn ∈ (U × )+ such that fn (u) > 0 whenever 0 0, so that fn (u) > 0 and f (u) > 0. So f witnesses that w ∈ A. Q Q (c) Take g ∈ (U × )+ such that g(u) > 0 whenever 0 0 such that for every n ∈ N we can find a un ∈ [0, w] with g(un ) ≤ 2−n and kun k ≥ ². Set vn = supi≥n ui ; then 0 ≤ vn ≤ w, g(vn ) ≤ 2−n+1 and kvn k ≥ ² for every n ∈ N. But hvn in∈N is nondecreasing, so v = inf n∈N vn must be non-zero, while 0 ≤ v ≤ w and g(v) = 0; which is impossible. X X Q Q Thus we have found an appropriate g. 376M Theorem (a) Let U be a Banach lattice with an order-continuous norm and V a Dedekind complete M -space. Then every bounded linear operator from U to V is an abstract integral operator. (b) Let U be an L-space and V a Banach lattice with order-continuous norm. Then every bounded linear operator from U to V × is an abstract integral operator.

468


376M

proof (a) By 355Kb and 355C, L× (U ; V ) = L∼ (U ; V ) ⊆ B(U ; V ); but since norm-bounded sets in V are also order-bounded, {T u : |u| ≤ u0 } is bounded above in V for every T ∈ B(U ; V ) and u0 ∈ U + , and B(U ; V ) = L× (U ; V ). I repeat ideas from the proof of 376H. (I cannot quote 376H directly as I am not assuming that V is weakly (σ, ∞)-distributive.) ?? Suppose, if possible, that B(U ; V ) is not the band Z of abstract integral operators. In this case there is a T > 0 in Z ⊥ . Take u1 ≥ 0 such that T u1 6= 0. Let f ≥ 0 in U × be such that for every ² > 0 there is a δ > 0 such that kuk ≤ ² whenever 0 ≤ u ≤ u1 and f (u) ≤ δ (376L). Set v0 = T u1 . Then, just as in part (d) of the proof of 376H, supu∈[0,u1 ],f (u)≤δ T u = v0 for every δ > 0. But there is a δ > 0 such that kT kkuk ≤ 12 kv0 k whenever 0 ≤ u ≤ u1 and f (u) ≤ δ; in which case k supu∈[0,u1 ],f (u)≤δ T uk ≤ 12 kv0 k, which is impossible. X X Thus Z = B(U ; V ), as required. (b) Because V has an order-continuous norm, V ∗ = V × = V ∼ ; and the norm of V ∗ is a Fatou norm with the Levi property (356Da). So B(U ; V ∗ ) = L× (U ; V × ), by 371C. By 376K, this is canonically isomorphic to L× (V ; U × ). Now U × = U ∗ is an M -space (356Pb). By (a), every member of L× (V ; U × ) is an abstract integral operator; but the isomorphism between L× (V ; U × ) and L× (U ; V × ) matches the abstract integral operators in each space (376K), so every member of B(U ; V ∗ ) is also an abstract integral operator, as claimed. 376N Corollary: Dunford’s theorem Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces and T : L1 (µ) → Lp (ν) a bounded linear operator, where 1 < p ≤ ∞. Then there is a measurable function R k : X × Y → R such that T f • = gf• , where gf (y) = k(x, y)f (x)dx almost everywhere, for every f ∈ L1 (µ). p proof Set q = p−1 if p is finite, 1 if p = ∞. We can identify Lp (ν) with V × , where V = Lq (ν) ∼ = Lp (ν)× (366Dc, 365Jc) has an order-continuous norm because 1 ≤ q < ∞. By 376Mb, T is an abstract integral operator. By 376F/376J, T is represented by a kernel, as claimed.

376O

Under the right conditions, weakly compact operators are abstract integral operators.

Lemma Let U be a Riesz space, and W a solid linear subspace of U ∼ . If C ⊆ U is relatively compact for the weak topology Ts (U, W ) (3A5E), then for every g ∈ W + and ² > 0 there is a u∗ ∈ U + such that g(|u| − u∗ )+ ≤ ² for every u ∈ C. proof Let Wg be the solid linear subspace of W generated by g. Then Wg is an Archimedean Riesz space with order unit, so Wg× is a band in the L-space Wg∗ = Wg∼ (356Na), and is therefore an L-space in its own right (354O). For u ∈ U , h ∈ Wg× set (T u)(h) = h(u); then T is an order-continuous Riesz homomomorphism from U to Wg× (356F). Now Wg is perfect. P P I use 356K. Wg is Dedekind complete because it is a solid linear subspace of the Dedekind complete space U ∼ . Wg× separates the points of W because T [U ] does. If A ⊆ Wg is upwardsdirected and suph∈A φ(h) is finite for every φ ∈ Wg× , then A acts on Wg× as a set of bounded linear functionals which, by the Uniform Boundedness Theorem (3A5Ha), is uniformly bounded; that is, there is some M ≥ 0 such that suph∈A |φ(h)| ≤ M kφk for every φ ∈ Wg× . Because g is the standard order unit of Wg , we have kφk = |φ|(g) and |φ(h)| ≤ M |φ|(g) for every φ ∈ Wg× , h ∈ A. In particular, h(u) ≤ |h(u)| = |(T u)(h)| ≤ M |T u|(g) = M T u(g) = M g(u) +

for every h ∈ A, u ∈ U . But this means that h ≤ M g for every h ∈ A and A is bounded above in Wg . Thus all the conditions of 356K are satisfied and Wg is perfect. Q Q Accordingly T is continuous for the topologies Ts (U, W ) and Ts (Wg× , Wg×× ), because every element φ of Wg×× corresponds to a member of Wg ⊆ W , so 3A5Ec applies. Now we are supposing that C is relatively compact for Ts (U, W ), that is, is included in some compact set C 0 ; accordingly T [C 0 ] is compact and T [C] is relatively compact for Ts (Wg× , Wg×× ). Since Wg× is an L-space, T [C] is uniformly integrable (356Q); consequently (ignoring the trivial case C = ∅) there are φ0 , . . . , φn ∈ T [C] such that k(|φ| − supi≤n |φi |)+ k ≤ ² for every φ ∈ T [C] (354Rb), so that (|φ| − supi≤n |φi |)+ (g) ≤ ² for every φ ∈ T [C].

376R

Kernel operators

469

Translating this back into of C itself, and recalling that T is a Riesz homomorphism, we see that there are u0 , . . . , un ∈ C such that g(|u| − supi≤n |ui |)+ ≤ ² for every u ∈ C. Setting u∗ = supi≤n |ui | we have the result. 376P Theorem Let U be an L-space and V a perfect Riesz space. If T : U → V is a linear operator such that {T u : u ∈ U, kuk ≤ 1} is relatively compact for the weak topology Ts (V, V × ), then T is an abstract integral operator. proof (a) For any g ≥ 0 in V × , Mg = supkuk≤1 g(|T u|) is finite. P P By 376O, there is a v ∗ ∈ V + such ∗ + ∗ that g(|T u| − v ) ≤ 1 whenever kuk ≤ 1; now Mg ≤ g(v ) + 1. Q Q Considering kuk−1 u, we see that g(|T u|) ≤ Mg kuk for every u ∈ U . Next, we find that T ∈ L∼ (U ; V ). P P Take u ∈ U + . Set Pn Pn B = { i=0 |T ui | : u0 , . . . , un ∈ U + , i=0 ui = u} ⊆ V + . Then B is upwards-directed. (Cf. 371A.) If g ≥ 0 in V × , n X

sup g(v) = sup{ v∈B

i=0 n X

≤ sup{

n X

g(|T ui |) : Mg kui k :

i=0

i=0 n X

ui = u} ui = u} = Mg kuk

i=0

is finite. By 356K, B is bounded above in V ; and of course any upper bound for B is also an upper bound for {T u0 : 0 ≤ u0 ≤ u}. As u is arbitrary, T is order-bounded. Q Q Because U is a Banach lattice with an order-continuous norm, T ∈ L× (U ; V ) (355Kb). (b) Since we can identify L× (U ; V ) with L× (U ; V ×× ), we have an adt operator T 0 ∈ L× (V × ; U × ), as in 376K. Now if g ≥ 0 in V × and hgn in∈N is a sequence in [0, g] such that limn→∞ gn (v) = 0 for every v ∈ V , hT 0 gn in∈N order*-converges to 0 in U × . P P For any ² > 0, there is a v ∗ ∈ V + such that g(|T u| − v ∗ )+ ≤ ² whenever kuk ≤ 1; consequently kT 0 gn k = sup (T 0 gn )(u) = sup gn (T u) kuk≤1

kuk≤1

∗

≤ gn (v ) + sup gn (|T u| − v ∗ )+ kuk≤1 ∗

≤ gn (v ) + sup g(|T u| − v ∗ )+ ≤ gn (v ∗ ) + ² kuk≤1

for every n ∈ N. As limn→∞ gn (v ∗ ) = 0, lim supn→∞ kT 0 gn k ≤ ²; as ² is arbitrary, hkT 0 gn kin∈N → 0. But as U × is an M -space (356Pb), it follows that hT 0 gn in∈N order*-converges to 0. Q Q By 368Pc, V is weakly (σ, ∞)-distributive. By 376H, T 0 is an abstract integral operator, so T also is, by 376K. 376Q Corollary Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces and T : L1 (µ) → L1 (ν) a weakly compact bounded linear operator. Then there is a function k : X × Y → R such that T f • = gf• , where R gf (y) = k(x, y)f (x)dx almost everywhere, for every f ∈ L1 (µ). proof This follows from 376P and 376J, just as in 376N. 376R So far I have mentioned actual kernel functions k(x, y) only as a way of giving slightly more concrete form to the abstract kernels of 376E. But of course they can provide new structures and insights. I give one result as an example. The following lemma is useful. Lemma Let (X, Σ, µ) be a measure space, (Y, T, ν) a σ-finite measure space, and λ the c.l.d. product measure on X × Y . Suppose that k is a λ-integrable real-valued function. Then for any ² > 0 there is a finite partition E0 , . . . , En of X into measurable sets such that kk − k1 k1 ≤ ², where

470


k1 (x, y) =

1 µEi

376R

Z k(t, y)dt whenever x ∈ Ei , 0 < µEi < ∞ Ei

and the integral is defined in R, = 0 in all other cases. proof Once again I refer to the proof of 253F: there are sets H0 , . . . , Hr of finite P measure in X, sets r F0 , . . . , Fr of finite measure in Y , and α0 , . . . , αr such that kk − k2 k1 ≤ 12 ², where k2 = j=0 αi χ(Hj × Fj ). R Let E0 , . . . , En be the partition of X generated by {Hi : i ≤ r}. Then for any i ≤ n, Ei ×Y |k − k1 | is R defined and is at most 2 Ei ×Y |k − k2 |. P P If µEi = 0, this is trivial, as both are zero. If µEi = ∞, then again the result is elementary, since both k 1 and k2 are zero on Ei × Y . So let us suppose that 0 < µEi < ∞. R In this case Ei k(t, y)dt must be defined for almost every y, by Fubini’s theorem. So k1 is defined almost everywhere on Ei × Y , and

R

Ei ×Y

|k − k1 | =

R R Y

Now take some fixed y ∈ Y such that β=

1 µEi

Ei

|k(x, y) − k1 (x, y)|dxdy.

R Ei

k(t, y)dt

is defined. Then β = k1 (x, y) for every x ∈ Ei . For every x ∈ Ei , we must have k2 (x, y) = α where R P α = {αj : Ei ⊆ Hj , y ∈ Fj }. But in this case, because Ei k(x, y) − β dx = 0, we have

R

Ei

If β ≥ α,

max(0, k(x, y) − β)dx =

R Ei

if β ≤ α,

R Ei

R

max(0, k(x, y) − β)dx ≤

max(0, β − k(x, y))dx ≤

in either case, 1 2

R Ei

Ei

max(0, β − k(x, y))dx =

R Ei

R Ei

1 2

max(0, k(x, y) − α)dx ≤

max(0, α − k(x, y))dx ≤

|k(x, y) − k1 (x, y)|dx ≤

R Ei

R

Ei

R Ei

R Ei

|k(x, y) − k1 (x, y)|dx.

|k(x, y) − k2 (x, y)|dx; |k(x, y) − k2 (x, y)|dx;

|k(x, y) − k2 (x, y)|dx.

This is true for almost every y, so integrating with respect to y we get the result. Q Q Now, summing over i, we get

R

|k − k1 | ≤ 2

R

|k − k2 | ≤ ²,

as required. 376S Theorem Let (X, Σ, µ) be a complete locally determined measure space, (Y, T, ν) a σ-finite measure space, and λ the c.l.d. product measure on X × Y . Let τ be an extended Fatou norm on L0 (ν) and 0 write Lτ for {g : g ∈ L0 (ν), τ 0 (g • ) < ∞}, where τ 0 is the associate extended Fatou norm of τ (369H-369I). 0 Suppose that k ∈ L0 (λ) is such that k × (f ⊗ g) is integrable whenever f ∈ LR1 (µ) and g ∈ LRτ . Then we have a corresponding linear operator T : L1 (µ) → Lτ defined by saying that (T f • ) × g • = k × (f ⊗ g) 0 whenever f ∈ L1 (µ), g ∈ Lτ . For x ∈ X set kx (y) = k(x, y) whenever this is defined. Then kx ∈ L0 (ν) for almost every x; set vx = kx• ∈ L0 (ν) for such x. In this case x 7→ τ (vx ) is measurable and defined and finite almost everywhere, and kT k = ess supx τ (vx ). Remarks The discussion of extended Fatou norms in §369 regarded them as functionals on spaces of the form L0 (A). I trust that no-one will be offended if I now speak of an extended Fatou norm on L0 (ν), with 0 the associated function spaces Lτ , Lτ ⊆ L0 , taking for granted the identification in 364Jc. Recall that (f ⊗ g)(x, y) = f (x)g(y) for x ∈ dom f , y ∈ dom g (253B).

376S

Kernel operators

471

By ‘ess supx τ (vx )’ I mean inf{M : M ≥ 0, {x : vx is defined and τ (vx ) ≤ M } is conegligible} (see 243D). R 0 proof (a) To see that the formula (f, g) 7→ k × (f ⊗ g) gives rise to an operator in L× (U ; (Lτ )× ), it is perhaps quickest to repeat the argument of parts (a) and (b) of the proof of 376E. (We are not quite in a position to quote 376E, as stated, because the localizable measure algebra free product there might be strictly larger than the measure algebra of λ; see 325B.) R The first step, of course, is to note that changing f or g on a negligible set does not affect the integral k × (f ⊗ g), so that we have a bilinear functional on 0 L1 × Lτ ; and the other essential element is the fact that the maps f • 7→ (f ⊗ χY )• , g • 7→ (χX ⊗ g)• are order-continuous (put 325A and 364Rc together). 0 By 369K, we can identify (Lτ )× with Lτ , so that T becomes an operator in L× (U ; Lτ ). Note that it must be norm-bounded (355C). (b) By 376I, there is a non-decreasing sequence hYn in∈N of measurable sets in Y , covering Y , such that 0 χYn ∈ Lτ for every n. Set X0 = {x : x ∈ X, kx ∈ L0 (ν)}. Then X0 is conegligible in P Let E ∈ Σ be R X. P any set of finite measure. Then for any n ∈ N, k × (χE ⊗ χYn ) is integrable, that is, E×Yn k is defined and R finite; so by Fubini’s theorem Yn kx is defined and finite for almost every x ∈ E. Consequently, for almost every x ∈ E, kx × χYn ∈ L0 (ν) for every n ∈ N, that is, kx ∈ L0 (ν), that is, x ∈ X0 . Thus E \ X0 is negligible for every set E of finite measure. Because µ is complete and locally determined, X0 is conegligible. Q Q This means that vx and τ (vx ) are defined for almost every x. (c) τ (vx ) ≤ kT k for almost every x. P P Take any E ∈ Σ of finite measure, and n ∈ N. Then k × χ(E × Yn ) is integrable. For each r ∈ N, there is a finite partition Er0 , . . . , Er,m(r) of E into measurable sets such that R |k − k (r) | ≤ 2−r , where E×Yn Z 1 k (r) (x, y) = k(t, y)dt whenever y ∈ Yn , x ∈ Eri , µEri > 0 µEri

Eri

and the integral is defined in R, = 0 otherwise (r)

(r)

(376R). Now k (r) is also integrable over E × Yn , so kx ∈ L0 (ν) for almost every x ∈ E, writing kx (y) = (r) (r) (r) (r) k (r) (x, y), and we can speak of vx = (kx )• for almost every x. Note that kx = kx0 whenever x, x0 belong to the same Eri . 0 0 (r) If µEri > 0, then vx must be defined for every x ∈ Eri . If v 0 ∈ Lτ is represented by g ∈ Lτ then Z Z k × (χEri ⊗ (g × χYn )) = k(t, y)g(y)d(t, y) Eri ×Yn Z Z = µEri k (r) (x, y)g(y)dy = µEri vx(r) × v 0 for any x ∈ Eri . But this means that µEri

R

(r)

vx × v 0 =

R

• T (χEri ) × v 0 × χYn•

0

for every v 0 ∈ Lτ , so (r)

vx =

1 • T (χEri ) × χYn• , µEri

(r)

τ (vx ) ≤

1 • kT kkχEri k1 µEri (r)

= kT k

for every x ∈ Eri . RThis is true whenever µEri > 0, so in fact τ (vx ) ≤ kT k for almost every x ∈ E. P Because r∈N E×Yn |k − k (r) | < ∞, we must have k(x, y) = limr→∞ k (r) (x, y) for almost every (x, y) ∈ E × Yn . Consequently, for almost every x ∈ E, k(x, y) = limr→∞ k (r) (x, y) for almost every y ∈ Yn , that is,

472


376S

(r)

hvx ir∈N order*-converges to vx × χYn• (in L0 (ν)) for almost every x ∈ E. But this means that, for almost every x ∈ E, (r)

τ (vx × χYn• ) ≤ lim inf r→∞ τ (vx ) ≤ kT k (369Mc). Now τ (vx ) = limn→∞ τ (vx × χYn• ) ≤ kT k for almost every x ∈ E. As in (b), this implies (since E is arbitrary) that τ (vx ) ≤ kT k for almost every x ∈ X. Q Q (d) I now show that x 7→ τ (vx ) is measurable. P P Take γ ∈ [0, ∞[ and set A = {x : x ∈ X0 , τ (vx ) ≤ γ}. ˜ y) = k(x, y) when x ∈ G Suppose that µE < ∞. Let G be a measurable envelope of A ∩ E (132Ed). Set k(x, 0 and (x, y) ∈ dom k, 0 otherwise. If f ∈ L1 (µ) and g ∈ Lτ , then

R

˜ y)f (x)g(y)d(x, y) = k(x,

R

k(x, y)f (x)g(y)d(x, y) =

G×Y

R

G

f (x)

R

Y

k(x, y)g(y)dydx

is defined. R ˜ y)g(y)|dy. Then h is finite almost everywhere and measurable. For For x ∈ X0 , set h(x) = |k(x, x ∈ A ∩ E, R R ˜ y)g(y)|dy = |vx × g • | ≤ γτ 0 (g • ). |k(x, So the measurable set G0 = {x : h(x) ≤ γτ 0 (g • )} includes A ∩ E, and µ(G \ G0 ) = 0. Consequently |

R

˜ y)f (x)g(y)d(x, y)| ≤ k(x,

R

G

|f (x)|h(x)dx ≤ γkf k1 τ 0 (g • ),

0

and this is true for every f ∈ L1 (µ), g ∈ Lτ . Now we have an operator T˜ : L1 (µ) → Lτ defined by the formula

R

R

0 k˜ × (f ⊗ g) when f ∈ L1 (ν) and g ∈ Lτ , R 0 and the formula just above tells us that | T˜u × v 0 | ≤ γkuk1 τ 0 (v 0 ) for every u ∈ L1 (ν), v 0 ∈ Lτ ; that is, τ (T˜u) ≤ γkuk1 for every u ∈ L1 (µ); that is, kT˜k ≤ γ. But now (c) tells us that τ (˜ vx ) ≤ γ for almost every ˜ x ∈ X, where v˜x is the equivalence class of y 7→ k(x, y), that is, v˜x = vx for x ∈ G ∩ X0 , 0 for x ∈ X \ G. So τ (vx ) ≤ γ for almost every x ∈ G, and G \ A is negligible. But this means that A ∩ E is measurable. As E is arbitrary, A is measurable; as γ is arbitrary, x 7→ τ (vx ) is measurable. Q Q

(T˜f • ) × g • =

(e) Finally, the ideas in (d) show that kT k ≤ ess supx τ (vx ). P P Set M = ess supx τ (vx ). If f ∈ L1 (µ) τ0 and g ∈ L , then

R

|k(x, y)f (x)g(y)|d(x, y) ≤

R

|f (x)|τ (vx )τ 0 (g • )dx ≤ M kf k1 τ 0 (g • );

as g is arbitrary, τ (T f • ) ≤ M kf k1 ; as f is arbitrary, kT k ≤ M . Q Q 376X Basic exercises > (a) Let µ be Lebesgue measure on R. Let h be a µ-integrable real-valued function with khk1 ≤ 1,R and set k(x, y) = h(y − x) whenever this is defined. Show that if f is in either L1 (µ) or L∞ (µ) then g(y) = k(x, y)f (x)dx is defined for almost every y ∈ R, and that this formula gives rise to an operator T ∈ Tµ¯×,¯µ as defined in 373A. (Hint: 255H.) (b) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras with localizable measure algebra free product ¯ and take p ∈ [1, ∞]. Show that if u ∈ Lp (A, µ ¯ and (C, λ), ¯) and v ∈ Lp (B, ν¯) then u ⊗ v ∈ Lp (C, λ) ku ⊗ vkp = kukp kvkp . > (c) Let U , V , W be Riesz spaces, of which V and W are Dedekind complete, and suppose that T ∈ L× (U ; V ) and S ∈ L× (V ; W ). Show that if either S or T is an abstract integral operator, so is ST . (d) Let h be a Lebesgue integrable function on R, and f a square-integrable function. Suppose that R hfn in∈N is a sequence of measurable functions such that (α) |fn | ≤ f for every n (β) limn→∞ E fn = 0 for every measurable set E of finite measure. Show that limn→∞ (h ∗ fn )(y) = 0 for almost every y ∈ R, where h ∗ fn is the convolution of h and fn . (Hint: 376Xa, 376F.)

376Xn

Kernel operators

473

(e) Let U and V be Riesz spaces, of which V is Dedekind complete. Suppose that W ⊆ U ∼ is a solid linear subspace, and that T belongs to the band in L∼ (U ; V ) generated by operators of the form u 7→ f (u)v, where f ∈ W and v ∈ V . Show that whenever hun in∈N is an order-bounded sequence in U such that limn→∞ f (un ) = 0 for every f ∈ W , then hT un in∈N order*-converges to 0 in V . (f ) Let (A, µ ¯) be a semi-finite measure algebra and U ⊆ L0 = L0 (A) an order-dense Riesz subspace such × that U separates the points of U . Let hun in∈N be an order-bounded sequence in U . Show that the following are equiveridical: (i) limn→∞ f (|un |) = 0 for every f ∈ U × ; (ii) hun in∈N → 0 for the topology of convergence in measure on L0 . (Hint: by 367T, condition (ii) is intrinsic to U , so we can replace (A, µ ¯) by a localizable algebra and use the representation in 369D.) (g) Let U be a Banach lattice with an order-continuous norm, and V a weakly (σ, ∞)-distributive Riesz space. Show that for T ∈ L∼ (U ; V ) the following are equiveridical: (i) T belongs to the band in L∼ (U ; V ) generated by operators of the form u 7→ f (u)v where f ∈ U ∼ , v ∈ V ; (ii) hT un in∈N order*-converges to 0 in V whenever hun in∈N is an order-bounded sequence in U + which is norm-convergent to 0; (iii) hT un in∈N order*-converges to 0 in V whenever hun in∈N is an order-bounded sequence in U which is weakly convergent to 0. (h) Let (X, Σ, µ) and (Y, T, ν) be σ-finite measure spaces, with product measure λ on X × Y , and measure algebras (A, µ ¯), (B,Rν¯). Suppose that k ∈ L0 (λ). Show that the following are equiveridical: (i)(α) 1 if f ∈ L (µ) then gf (y) = k(x, y)f (x)dx is defined for almost every yR and gf ∈ L1 (ν) (β) there is an operator T ∈ Tµ¯×,¯ν defined by setting T f • = gf• for every f ∈ L1 (µ); (ii) |k(x, y)|dy ≤ 1 for almost every R x ∈ X, |k(x, y)|dx ≤ 1 for almost every y ∈ Y . > (i) Let µ be Lebesgue measureRon R. Give an example of a measurable function k : [0, 1]2 → R such that, for any f ∈ L2 (µ), gf (y) = k(x, y)f (x)dx is defined for every y and kgf k2 = kf k2 , but k is not integrable, so the linear isometry on L2 = L2 (µ) defined by k does not belong to L∼ (L2 ; L2 ). (Hint: 371Ye.) (j) Let (X, Σ, µ) be a σ-finite measure space and (Y, T, ν) a complete locally determined measure space. Let U ⊆ L0 (µ), V ⊆ L0 (ν) be solid linear subspaces, of which V is order-dense; write V # = {v : v ∈ L0 (ν), v × v 0 is integrable for every v 0 ∈ V }, U = {f : f ∈ L0 (ν), f • ∈ U }, V = {g : g ∈ L0 (ν), g • ∈ V }, V# = {h : h ∈ L0 (ν), h• ∈ V # }. Let λ be the c.l.d. product measure on X × Y , and k ∈ LR0 (λ) a function such that k×(f ⊗g) is integrable for every f ∈ U, g ∈ V. (i) Show that for any f ∈ U, hf (y) = k(x, y)f (x)dx × # is defined for almost every y ∈ Y , and that hf ∈ V# . (ii) Show R that we haveRa map T ∈ L (U ; V ) defined either by writing T f • = h•f for every f ∈ U or by writing (T f • ) × g • = k × (f ⊗ g) for every f ∈ U, g ∈ V. (k) Let (X, Σ, µ), (Y, T, ν) and (Z, Λ, λ) be σ-finite measure spaces, and U , V , W perfect order-dense solid linear subspaces of L0 (µ), L0 (ν) and L0 (λ) respectively. Suppose that T : U → V and S : V → W are abstract integral operators corresponding to kernels k1 ∈ L0 (µ × ν), k2 ∈ L0 (ν × λ), writing µ × ν for the (c.l.d. or primitive) product measure on R X × Y . Show that ST : U → W is represented by the kernel k ∈ L0 (µ × λ) defined by setting k(x, z) = k1 (x, y)k2 (y, z)dy whenever this integral is defined. (l) Let U be a perfect Riesz space. Show that a set C ⊆ U is relatively compact for Ts (U, U × ) iff for every g ∈ (U × )+ , ² > 0 there is a u∗ ∈ U such that g(|u| − u∗ )+ ≤ ² for every u ∈ C. (Hint: 376O and the proof of 356Q.) > (m) Let µ be Lebesgue measure on [0, 1], and ν counting measure on [0, 1]. Set k(x, y) = 1 if x = y, 0 otherwise. Show that 376S fails in this context (with, e.g., τ = k k∞ ). (n) Suppose, in 376Xj, that U = Lτ for some extended Fatou norm on L0 (µ) and that V = L1 (ν), so that V # = L∞ (ν). Set ky (x) = k(x, y) whenever this is defined, uy = ky• whenever ky ∈ L0 (ν). Show that uy ∈ Lτ for almost every y ∈ Y , and that the norm of T in B(Lτ ; L∞ ) is ess supy τ 0 (uy ). (Hint: do the case of totally finite Y first.)

474


376Y

376Y Further exercises (a) Let U , V and W be linear spaces (over any field F ) and φ : U × V → W a bilinear map. Let W0 be the linear subspace of W generated by φ[U × V ]. Show that the following are equiveridical: (i) for every linear space Z over F and every bilinear ψ : U × V → Z, there is a (unique) linear operator T : W0 → Z such that Pn T φ = ψ (ii) whenever u0 , . . . , un ∈ U are linearly independent and v0 , . . . , vn ∈ V are non-zero, φ(ui , vi ) 6= 0 (iii) whenever u0 , . . . , un ∈ U are non-zero and i=0P n v0 , . . . , vn ∈ V are linearly independent, i=0 φ(ui , vi ) 6= 0 (iv) for any Hamel bases hui ii∈I , hvj ij∈J of U and V , hφ(ui , vj )ii∈I,j∈J is a Hamel basis of W0 (v) for some pair hui ii∈I , hvj ij∈J of Hamel bases of U and V , hφ(ui , vj )ii∈I,j∈J is a Hamel basis of W0 . ¯ their localizable measure algebra free (b) Let (A, µ ¯), (B, ν¯) be semi-finite measure algebras, and (C, λ) 0 0 0 product. Show that ⊗ : L (A) × L (B) → L (C) satisfies the equivalent conditions of 376Ya. (c) Let (X, Σ, µ) and (Y, T, ν) be semi-finite measure spaces and λ the c.l.d. product measure on X × Y . Show that the map (f, g) 7→ f ⊗ g : L0 (µ) × L0 (ν) → L0 (λ) induces a map (u, v) 7→ u ⊗ v : L0 (µ) × L0 (ν) → L0 (λ) possessing all the properties described in 376B and 376Ya. (d) Let (A, µ ¯) be the measure algebra of {0, 1}ω1 with its usual measure, and hbξ iξ<ω1 the canonical 2 independent family of elements of measure 12 in A. Set ¯) and V = {v : v ∈ Rω1 , {ξ : v(ξ) 6= 0} is R U =L R (A, µ countable}. Define T : U → V by setting T u(ξ) = 2 bξ u − u for ξ < ω1 , u ∈ U . Show that (i) hT un in∈N order*-converges to 0 in V whenever hun in∈N is a sequence in U such that limn→∞ f (un ) = 0 for every f ∈ U × (ii) T ∈ / L∼ (U ; V ). (e) Let U be a Riesz space with the countable sup property (definition: 241Yd) such that U × separates the points of U , and hun in∈N a sequence in U . Show that the following are equiveridical: (i) limn→∞ f (v∧|un |) = 0 for every f ∈ U × , v ∈ U + ; (ii) every subsequence of hun in∈N has a sub-subsequence which is order*convergent to 0. (f ) Let U be an Archimedean Riesz space and A a weakly (σ, ∞)-distributive Dedekind complete Boolean algebra. Suppose that T : U → L0 = L0 (A) is a linear operator such that h|T un |in∈N order*-converges to 0 0 in L0 whenever hun in∈N is order-bounded and order*-convergent to 0 in U . Show that T ∈ L∼ c (U ; L ), so × 0 that if U has the countable sup property then T ∈ L (U ; L ). (g) Suppose that (Y, T, ν) is a probability space in which T = PY , ν{y} = 0 for every y ∈ Y . (See 363S.) Take X = Y and let µ be counting measure on X; let λ be the c.l.d. product measure on X × Y , and set k(x, y) = 1 if x = y, 0 otherwise. Show that we haveR an operator T : L∞ (µ) → L∞ (ν) defined by setting T f = g • whenever f ∈ L∞ (µ) ∼ = `∞ (X) and g(y) = k(x, y)f (x)dx = f (y) for every y ∈ Y . Show that T × ∞ does not belong to L (L (µ); L∞ (ν)) and in particular does not satisfy the condition (ii) of 376J. (h) Give an example of an abstract integral operator T : `2 → L1 (µ), where µ is Lebesgue measure on [0, 1], such that hT en in∈N is not order*-convergent in L1 (µ), where hen in∈N is the standard orthonormal sequence in `2 . (i) Let U be an L-space and write G for the regular open algebra of R. Show that any bounded linear operator from U to L∞ (G) is an abstract integral operator. (j) Let U be an L-space and V a Banach lattice with an order-continuous norm. Let T ∈ L∼ (U ; V ). Show that the following are equiveridical: (i) T is an abstract integral operator; (ii) T [C] is norm-compact in V whenever C is weakly compact in U . (Hint: start with the case in which C is order-bounded, and that it is weakly sequentially compact.) (k) Let (X, Σ, µ) be a semi-finite measure space. Show that 376S is valid for (X, Σ, µ) iff (X, Σ, µ) has locally determined negligible sets (213I). (l) Let (X, Σ, µ) be a complete locally determined measure space and (Y, T, ν), (Z, Λ, λ) two σ-finite measure spaces. Suppose that τ , θ are extended Fatou norms on L0 (ν), L0 (λ) respectively, and that T : L1 (µ) → Lτ is an abstract integral operator, with corresponding kernel k ∈ L0 (µ×ν), while S ∈ L× (Lτ ; Lθ ), so that ST : L1 (µ) → Lθ is an abstract integral operator (376Xc); let k˜ ∈ L0 (µ × λ) be the corresponding kernel. For x ∈ X set vx = kx• when this is defined in Lτ , as in 376S, and similarly take wx = k˜x• ∈ Lθ . Show that Svx = wx for almost every x ∈ X.

376 Notes

Kernel operators

475

P∞ (m) Set k(m, n) = 1/π(n − m + 21 ) for m, n ∈ Z. (i) Show that n=−∞ k(m, n)2 = 1 for every m ∈ Z. P∞ (Hint: 282Xo.) (ii) Show that n=−∞ k(m, n)k(m0 , n) = 0 for all distinct m, m0 ∈ Z. (Hint: look at 2 2 k(m, n) − k(m0 , n).) (iii) Show P∞ that there is a norm-preserving linear operator T from ` = ` (Z) to itself given by setting (T u)(n) = m=−∞ k(m, n)u(m) for every n. (iv) Show that T 2 is the identity operator on p P∞ `2 . (v) Show that T ∈ / L∼ (`2 ; `2 ). (Hint: consider m,n=−∞ |k(m, n)|x(m)x(n) where x(n) = 1/ |n| ln |n| for |n| ≥ 2.) (T is a form of the Hilbert transform.) (n) (i) Show that there is a compact linear operator from `2 to itself which is not in L∼ (`2 ; `2 ). (Hint: start from the operator S of 371Ye.) (ii) Show that the identity operator on `2 is an abstract integral operator. 376 Notes and comments I leave 376Yb to the exercises because I do not rely on it for any of the work here, but of course it is an essential aspect of the map ⊗ : L0 (A) × L0 (B) → L0 (C) I discuss in this section. The conditions in 376Ya are characterizations of the ‘tensor product’ of two linear spaces, a construction of great importance in abstract linear algebra (and, indeed, in modern applied linear algebra; it is by no means trivial even in the finite-dimensional case). In particular, note that conditions (ii), (iii) of 376Ya apply to arbitrary subspaces of U and V if they apply to U and V themselves. The principal ideas used in 376B-376C have already been set out in §§253 and 325. Here I do little more ¯ should have no fewer than than list the references. I remark however that it is quite striking that L1 (C, λ) three universal mapping theorems attached to it (376Ca, 376C(b-i) and 376C(b-ii)). The real work of this section begins in 376E. As usual, much of the proof is taken up with relatively straightforward verifications, as in parts (a) and (b), while part (i) is just a manoeuvre to show that it doesn’t matter if A and B aren’t Dedekind complete, because C is. But I think that parts (d), (f) and (j) have ideas in them. In particular, part (f) is a kind of application of the Radon-Nikod´ ym theorem (through ¯ ∗ with L∞ (C)). the identification of L1 (C, λ) I have split 376E from 376H because the former demands the language of measure algebras, while the latter can be put into the language of pure Riesz space theory. Asking for a weakly (σ, ∞)-distributive space V in 376H is a way of applying the ideas to V = L0 as well as to Banach function spaces. (When V = L0 , indeed, variations on the hypotheses are possible, using 376Yf.) But it is a reminder of one of the directions in which it is often possible to find generalizations of ideas beginning in measure theory. The condition ‘limn→∞ f (un ) = 0 for every f ∈ U × ’ (376H(ii)) seems natural in this context, and gives marginally greater generality than some alternatives (because it does the right thing when U × does not separate the points of U ), but it is not the only way of expressing the idea; see 376Xf and 376Ye. Note that the conditions (ii) and (iii) of 376H are significantly different. In 376H(iii) we could easily have |un | = u∗ for every n; for instance, if un = 2χan − χ1 for some stochastically independent sequence han in∈N of elements of measure 21 in a probability algebra (272Yd). If you have studied compact linear operators between Banach spaces (definition: 3A5Ka), you will have encountered the condition ‘T un → 0 strongly whenever un → 0 weakly’. The conditions in 376H and 376J are of this type. If a sequence hun in∈N in a Riesz space U is order-bounded and order*-convergent to 0, then limn→∞ f (un ) = 0 for every f ∈ U × (367Xf). Visibly this latter condition is associated with weak convergence, and order*-convergence is (in Banach lattices) closely related to norm convergence (367D367E). In the context of 376H, an abstract integral operator is one which transforms convergent sequences of a weak type into convergent sequences of a stronger type. The relationship between the classes of (weakly) compact operators and abstract integral operators is interesting, but outside the scope of this book; I leave you with 376P-376Q and 376Y, and a pair of elementary examples to guard against extravagant conjecture (376Yn). 376O belongs to an extensive general theory of weak compactness in perfect Riesz spaces, based on adaptations of the concept of ‘uniform integrability’. I give the next step in 376Xl. For more information see Fremlin 74a, chap. 8. Note that 376Mb and 376P overlap when V × in 376Mb is reflexive – for instance, when V is an Lp space for some p ∈ ]1, ∞[ – since then every bounded linear operator from L1 to V × must be weakly compact. I give 376Yi as a hint that there may be more to be said about the case in which the codomain is not perfect, indeed not weakly (σ, ∞)-distributive. For more information on the representation of operators see Dunford & Schwartz 57, particularly Table VI in the notes to Chapter VI.

476


376 Notes

As soon as we leave formulations in of the spaces L0 (A) and their subspaces, and return to the original conception of a kernel operator in of integrating functions against sections of a kernel, we are necessarily involved in the pathology of Fubini’s theorem for general measure spaces. In general, the RR RR repeated integrals k(x, y)dxdy, k(x, y)dydx need not be equal, and something has to give (376Xm). Of course this particular worry disappears if the spaces are σ-finite, as in 376J. In 376S I take the trouble to offer a more general condition, fairly near to the best possible result (376Yk), mostly as a reminder that the techniques developed in Volume 2 do enable us sometimes to go beyond the σ-finite case. Note that this is one of the many contexts in which anything we can prove about probability spaces will be true of all σ-finite spaces; but that we cannot make the next step, to all strictly localizable spaces. 376S verges on the theory of integration of vector-valued functions, which I don’t wish to enter here; but it also seems to have a natural place in the context of this chapter. It is of course a special property of L1 spaces. The formula kTk k = ess supx τ (kx• ) shows that kT|k| k = kTk k; now we know fron 376E that T|k| = |Tk |, so we get a special case of the Chacon-Krengel theorem (371D). Reversing the roles of X and Y , we find ourselves with an operator from Lτ to L∞ (376Xn), which is the other standard context in which kT k = k|T |k (371Xd). I include two exercises on L2 spaces (376Xi, 376Ym) designed to emphasize the fact that B(U ; V ) is included in L∼ (U ; V ) only in very special cases. The history of the theory here is even more confusing than that of mathematics in general, because so many of the ideas were developed in national schools in very imperfect with each other. My own gives no hint of how this material arose; I ought in particular to note that 376N is one of the oldest results, coming (essentially) from Dunford 36. For further references, see Zaanen 83, chap. 13.

381B

Automorphism groups of Boolean algebras

477

Chapter 38 Automorphism groups As with any mathematical structure, every measure algebra has an associated symmetry group, the group of all measure-preserving automorphisms. In this chapter I set out to describe some of the remarkable features of these groups. I begin with a section on automorphism groups of general Dedekind complete Boolean algebras (§381), before continuing with applications of the ideas developed there to measure algebras (§382); the principal results of these sections concern the expression of a general automorphism as a product of involutions (381N, 382D) and a description of the normal subgroups of automorphism groups of homogeneous algebras (381S, 382H). I continue with a discussion of circumstances under which these automorphism groups determine the underlying algebras and/or have few outer automorphisms (§383). One of the outstanding open problems of the subject is the ‘isomorphism problem’, the classification of automorphisms of measure algebras up to conjugacy in the automorphism group. I offer a section on ‘entropy’, the most important numerical invariant enabling us to distinguish some non-conjugate automorphisms (§384). For Bernoulli shifts on the Lebesgue measure algebra (384Q-384S), the isomorphism problem is solved by Ornstein’s theorem; I present a complete proof of this theorem in §§385-386. Finally, in §387, I give Dye’s theorem, describing the full subgroups generated by single automorphisms of measure algebras of countable Maharam type.

381 Automorphism groups of Boolean algebras My aim in this chapter is to describe the automorphism groups of measure algebras, but as usual I prefer to begin with results which can be expressed in the language of general Boolean algebras, even though on this occasion it is necessary to introduce some special terminology. I will however restrict myself to the ideas which seem necessary for the theorems in the next two sections. I begin with some elementary general results on the construction of automorphisms by piecing together fragments of others (381B-381C), a discussion of the concept of an element ing an automorphism (381D-381F, 381L), an introduction to a ‘cycle notation’ for certain automorphisms (381G-381I), and a version of Frol´ık’s theorem on the structure of automorphisms (381J). The principal theorems are 381N, giving a sufficient condition for every automorphism to be a product of involutions, and 381S, describing the normal subgroups of certain groups of automorphisms. Both depend on Dedekind completeness of the Boolean algebra and on the concept of ‘full’ subgroup of Aut A (381M); the latter also requires the group to have ‘many involutions’ (381P-381R). Both concepts are chosen with a view to the next section, where the results will be applied to groups of measure-preserving automorphisms. 381A The group Aut A For any Boolean algebra A, I write Aut A for the set of automorphisms of A, that is, the set of bijective Boolean homomorphisms π : A → A. This is a group, being a subgroup of the group of all bijections from A to itself (use 312G). Note that every member of Aut A is order-continuous; this is because it must be an isomorphism of the order structure of A, and is also a consequence of 313P(a-ii). 381B Lemma Let A be a Boolean algebra, and hai ii∈I , hbi ii∈I two partitions of unity in A. Assume either that I is finite or that I is countable and A is Dedekind σ-complete or that A is Dedekind complete. Suppose that for each i ∈ I we have an isomorphism πi : Aai → Abi between the corresponding principal ideals. The there is a unique π ∈ Aut A such that πc = πi c whenever i ∈ I and c ⊆ ai . Q Q proof By 315F, we may identify A with each of the products i∈I Aai , i∈I Abi ; now π corresponds to the isomorphism between the two products induced by the πi .

478

Automorphism groups

381C

381C Corollary Let A be a homogeneous Boolean algebra, and A, B two partitions of unity in A, neither containing 0. Let θ : A → B be a bijection. Suppose either that A, B are finite or that A, B are countable and A is Dedekind σ-complete or that A is Dedekind complete. Then there is an automorphism of A extending θ. proof For every a ∈ A, the principal ideals Aa , Aθa are isomorphic to the whole algebra A, and therefore to each other; let πa : Aa → Aθa be an isomorphism. Now apply 381B. 381D Elements ing an automorphism The following concept will be extremely useful. If A is a Boolean algebra, π ∈ Aut A and a ∈ A I will say that π is ed by a, or a s π, if πd = d for every d ⊆ 1 \ a. We have the following elementary facts. 381E Lemma Let A be a Boolean algebra. (a) If π ∈ Aut A is ed by a ∈ A then πa = a and πd ⊆ a for every d ⊆ a. (b) If π, φ ∈ Aut A are both ed by a ∈ A so are π −1 and πφ. (c) If π ∈ Aut A is ed by a ∈ A and φ ∈ Aut A then φπφ−1 is ed by φa. (d) If π, φ ∈ Aut A are ed by a, b respectively and a ∩ b = 0, then πφ = φπ. (e) Let π ∈ Aut A and let A be the set of a ∈ A ing π. Then A is non-empty and closed under arbitrary infima; also b ∈ A whenever b ⊇ a ∈ A. In particular, if π is ed by both a and b then it is ed by a ∩ b. (f) For any π ∈ Aut A, a ∈ A, a s π iff d ⊆ a whenever d ∩ πd = 0. (g) If π ∈ Aut A is ed by a ∈ A, and φ, ψ ∈ Aut A are such that φd = ψd for every d ⊆ a, then φπφ−1 = ψπψ −1 . proof (a) π(1 \ a) = 1 \ a, so πa = a, and if d ⊆ a then πd ⊆ πa = a. (b) If d ∩ a = 0 then πd = d = φd so π −1 d = d = πφd. (c) If d ∩ φa = 0 then φ−1 d ∩ a = 0 so πφ−1 d = φ−1 d and φπφ−1 d = d. (d) If d ⊆ a, then φd = d so πφd = πd. Also πd ⊆ πa = a, so φπd = πd; thus φπd = πφd. Similarly φπd = πφd if d ⊆ b. Finally, if d ⊆ 1 \ (a ∪ b), then πφd = φπd = d. By the ‘uniqueness’ assertion of 381B, or otherwise, πφ = φπ. (e) Of course 1 ∈ A, because π0 = 0; and it is also obvious that if b ⊇ a ∈ A then b ∈ A. If B ⊆ A is non-empty and c = inf B is defined in A, then for any d ⊆ 1 \ c we have d = d \ c = supb∈B d \ b, so (because π is order-continuous) πd = supb∈B π(d \ b) = supb∈B d \ b = d. So c s π. (f )(i) If a s π and d ∩ πd = 0, then d \ a = π(d \ a) ∩ d ⊆ πd ∩ d = 0, so d ⊆ a. (ii) If a does not π, there is a c ⊆ 1 \ a such that πc 6= c. So one of c \ πc, πc \ c is non-zero. If c \ πc 6= 0, take this for d; then d 6⊆ a and πd ∩ d ⊆ πc \ πc = 0. Otherwise, take d = π −1 (πc \ c); then 0 6= d ⊆ c, so d 6⊆ a, while d ∩ πd = (c \ π −1 c) ∩ (πc \ c) = 0. (g) For d ⊆ a, ψ −1 φd = ψ −1 ψd = d, so ψ −1 φ is ed by 1 \ a. By (d), πψ −1 φ = ψ −1 φπ, so φπφ−1 = ψψ −1 φπφ−1 = ψπψ −1 φφ−1 = ψπψ −1 .

381H


479

381F Corollary Let A be a Boolean algebra and π ∈ Aut A. If e ∈ A is such that there is any b ⊆ e for which πb 6= b, then there is a non-zero a ⊆ e such that a ∩ πa = 0. proof Because 1 \ e does not π, 381Ef tells us that there is a d 6⊆ 1 \ e such that d ∩ πd = 0; now set a = d ∩ e. 381G Cyclic automorphisms: Definition Let A be a Boolean algebra. (a) Suppose that a, b are dist of A and that π ∈ Aut A is such that πa = b. I will write ←−− (a π b) for the member ψ of Aut A defined by setting ψd = πd if d ⊆ a, = π −1 d if d ⊆ b, = d if d ⊆ 1 \ (a ∪ b). Observe that in this case (if a 6= 0) ψ is an involution, that is, has order 2 in the group Aut A; involutions of this type I will call exchanging involutions. (b) More generally, if a1 , . . . , an are dist elements of A and πi ∈ Aut A are such that πi ai = ai+1 for each i < n, then I will write (← a −−−a−−−−.−.−.−−−−− a−) 1 π1

2 π2

πn−1

n

for that ψ ∈ Aut A such that ψd = πi d if 1 ≤ i < n, d ⊆ ai , −1 = π1−1 π2−1 . . . πn−1 d if d ⊆ an ,

= d if d ⊆ 1 \ sup ai . i≤n

(c) It will occasionally be convenient to use the same notation when each πi is a Boolean isomorphism between the principal ideals Aai and Aai+1 , rather than an automorphism of the whole algebra A. Remark The point of this notation is that we can expect to use the standard techniques for manipulating cycles that are (I suppose) familiar to you from elementary group theory; the principal change is that we have to keep track of the subscripted automorphisms π . The following results are typical. 381H Lemma Let A be a Boolean algebra. ←−− (a) If ψ = (a π b) is an exchanging involution in Aut A, then ←−− ←−− ←−−− ψ = (a ψ b) = (b ψ a) = (b π−1 a) is ed by a ∪ b. ←−− (b) If π = (a π b) is an exchanging involution in Aut A, then for any φ ∈ Aut A, ←−−−−−−− φπφ−1 = (φa φπφ−1 φb) is another exchanging involution. ←−− ←−− (c) If π = (a π b) and φ = (c φ d) are exchanging involutions, and a, b, c, d are all dist, then π and φ ←−−−−−−− commute, and ψ = πφ = φπ is also an exchanging involution, being (a ∪ c ψ b ∪ d). proof (a) Check the action of ψ on the principal ideals Aa , Ab , A1\(a∪b) . (b) φa ∩ φb = φ(a ∩ b) = 0 and φπφ−1 φa = φπa = φb, ←−−−−−−− so ψ = (φa φπφ−1 φb) is well-defined. Now check the action of ψ on the principal ideals Aφa , Aφb , A1\φ(a∪b) . (c) Check the action of ψ on each of the principal ideals Aa , . . . , Ae , where e = 1 \ (a ∪ b ∪ c ∪ d).

480

Automorphism groups

381I

381I Remark I must emphasize that while, after a little practice, calculations of this kind become easy and safe, they are absolutely dependent on all the cycles present involving only of one list of dist elements of A. If, for instance, a, b, c are dist, then ←−− ←−− ←−−−− (a π b)(b φ c) = (a π b φ c). But if a ∩ c 6= 0 then there is no expression for the product in this language. Secondly, of course, we must −−−−− be scrupulous in checking, at every use of the notation (←− a1 π− 1 . . . an ), that a1 , . . . , an are dist and that πi ai = ai+1 for i < n. Thirdly, a significant problem can arise if the automorphisms involved don’t match. Consider for instance the product ←−− ←−− ψ = (a π b)(a φ b). Then we have ψd = π −1 φd if d ⊆ a, πφ−1 d if d ⊆ b; ψ is not necessarily expressible as a product of ‘dist’ cycles. Clearly there are indefinitely complex variations possible on this theme. A possible formal expression of a sufficient condition to avoid these difficulties is the following. Restrict yourself to calculations involving a fixed list a1 , . . . , an of dist elements of A for which you can describe a family of isomorphisms φij : Aai → Aaj such that φii is always the identity on Aai , φjk φij = φik for all i, j, k, and whenever ai π aj appears in a cycle of the calculation, then π agrees with φij on Aai . Of course this would be intolerably unwieldy if it were really necessary to exhibit all the φij every time. I believe however that it is usually easy enough to form a mental picture of the actions of the isomorphisms involved sufficiently clear to offer confidence that such φij are indeed present; and in cases of doubt, then after performing the formal operations it is always straightforward to check that the calculations are valid, by looking at the actions of the automorphisms on each relevant principal ideal. 381J

The following facts may help to put the concept of ‘exchanging involution’ in its proper place.

Lemma Let A be a Dedekind complete Boolean algebra and π any automorphism of A. Then there is a partition of unity (a0 , a00 , b0 , b00 , c, e) of A such that πa0 = b0 , πa00 = b00 , πb00 = c, π(b0 ∪ c) = a0 ∪ a00 , πd = d for every d ⊆ e. proof (a) Let P be the set {p : p ∈ A, p ∩ πp = 0}. Then P has a maximal element. P P Of course P 6= ∅, as 0 ∈ P . If Q ⊆ P is non-empty and upwards-directed, set p = sup Q, which is defined because A is Dedekind complete; then πp = sup π[Q] (since π, being an automorphism, is surely order-continuous). If q1 , q2 ∈ Q, there is a q ∈ Q such that q1 ∪ q2 ⊆ q, so q1 ∩ πq2 ⊆ q ∩ πq = 0. By 313Bc, sup Q ∩ sup π[Q] = 0, that is, p ∩ πp = 0. This means that p ∈ P and is an upper bound for Q in P . As Q is arbitrary, Zorn’s Lemma tells us that P has a maximal element. Q Q (b) Let a be a maximal element of P and set b = πa, c = πb \ a, e = 1 \ (a ∪ b ∪ c). (i) ?? Suppose, if possible, that there is some d ⊆ e such that πd 6= d. By 381F, there is a non-zero d ⊆ e such that d ∩ πd = 0. Consider a ˜ = a ∪ π −1 d; then a ⊂ a ˜. But a ˜ ∩ π˜ a = (a ∩ b) ∪ (a ∩ d) ∪ ((b ∪ d) ∩ π −1 d) = 0 because π((b ∪ d) ∩ π −1 d) ⊆ d ∩ (a ∪ c ∪ πd) = 0. So a ˜ is a member of P strictly greater than a, which is impossible. X X Thus πd = d for every d ⊆ e; in particular, πe = e. (ii) Because a ∩ b = 0, b ∩ c ⊆ π(a ∩ b) = 0, and (a, b, c, e) is a partition of unity. Because π(a ∪ b ∪ e) includes b ∪ c ∪ e, πc ⊆ 1 \ (b ∪ c ∪ e) = a. (c) Set b00 = π −1 c ⊆ b, b0 = b \ b00 , a0 = π −1 b0 , a00 = π −1 b00 . Then everything required in the statement of the lemma has been covered, with the possible exception of ‘π(b0 ∪ c) = a0 ∪ a00 ’; but as b0 ∪ c = 1 \ (a ∪ b00 ∪ e), π(b0 ∪ c) = 1 \ (πa ∪ πb00 ∪ πe) = 1 \ (b ∪ c ∪ e) = a = a0 ∪ a00 . So the proof is complete.

381N


481

381K Lemma Let A be a Dedekind complete Boolean algebra. Then every involution in Aut A is an exchanging involution in the sense of 381G. proof Let π be an involution and let a0 , a00 , b0 , b00 , c, e be a partition of unity as in 381J. Then c = π 2 a00 = a00 , c ∩ a00 = 0, so a00 = b00 = c = 0 and πd = d whenever d ⊆ 1 \ (a0 ∪ πa0 ) = e, while a0 ∩ πa0 = 0. 381L The of an automorphism For Dedekind complete Boolean algebras A, the following concept is very useful. For any π ∈ Aut A, set supp π = inf{a : π is ed by a}, the of π. Note that supp π s π (381Ee); it is the smallest element of A ing π. By 381Ef, supp π is also sup{d : d ∩ πd = 0}. Observe that in the language of 381J supp π is just a0 ∪ a00 ∪ b0 ∪ b00 ∪ c = 1 \ e. Note that supp π −1 = supp π, and that supp(ππ 0 ) ⊆ supp π ∪ supp π 0 for all π, π 0 ∈ Aut A, by 381Ed. Moreover, if π, φ ∈ Aut A, then supp(φπφ−1 ) = φ(supp π). P P By 381Ec, φ(supp π) s φπφ−1 , so supp(φπφ−1 ) ⊆ φ(supp π); but also supp π ⊆ φ−1 (supp(φπφ−1 )), so φ(supp π) ⊆ supp(φπφ−1 ). Q Q 381M In order to apply the argument of the next theorem simultaneously to the groups Aut A here and to the groups Autµ A of the next section, I introduce the following terminology. Definition If A is a Boolean algebra, a subgroup G of Aut A is full if whenever hai ii∈I is a partition of unity in A, hπi ii∈I is a family in G, and π ∈ Aut A is such that πa = πi ai whenever i ∈ I and a ⊆ ai , then π ∈ G. 381N Theorem Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A. Then every member of G is expressible as the product of at most eight involutions belonging to G. proof (a) For each n ∈ N, let In be the set of elements of G expressible as the product of n or fewer involutions belonging to G (so that I0 consists of the identity ι alone). It will be helpful to note that π −1 ∈ In , φπφ−1 ∈ In whenever π ∈ In and φ ∈ G. In particular, for instance, πψπ −1 ψ −1 = (πψπ −1 )ψ −1 ∈ I2 whenever ψ ∈ I1 and π ∈ G. − belongs to G, because it is made Note that if a ∈ A, π ∈ G and πa ∩ a = 0, then the involution (← a π−− πa) −1 up of the actions of π, π and ι on Aa , Aπa and A1\(a∪πa) respectively, and G is full. (b) If π ∈ G there are a ∈ A and involutions φ1 , φ2 ∈ G such that πa ∩ a = 0 and φ1 φ2 π is ed by a. P P Take a0 , a00 , b0 , b00 , c, e from 381J, and set a = b0 ∪ c. Then certainly πa = a0 ∪ a00 is dist from a. Set ←−−− ←−−−−−−−−−− φ1 = (b00 π c), φ2 = (a0 ∪ a00 π b0 ∪ b00 ). Then both φ1 , φ2 are involutions in G. Now trace the action of φ1 φ2 π on each of the principal ideals Aa0 ∪a00 , Ab00 , Ae ; we get φ1 φ2 πd = ιπ −1 πd = d if d ⊆ a0 ∪ a00 , φ1 φ2 πd = π −1 ιπd = d if d ⊆ b00 , φ1 φ2 πd = ιιιd = d if d ⊆ e. (I write the identity operator ι into these formulae in the hope of suggesting how the three functions act in each case.) So supp(φ1 φ2 π) ⊆ b0 ∪ c = a. Q Q (c) Suppose that π, ψ ∈ G, a ∈ A, π is ed by a and ψa ∩ a = 0. Then there are φ1 , φ2 , φ3 , ψ 0 ∈ G and ˜b ⊆ ψa such that (i) φ1 , φ2 , φ3 are all involutions ed by a ∪ ˜b (ii) ψ 0˜b ⊆ ψa \ ˜b (iii) φ1 φ2 φ3 π is ed by ˜b. P P Take a0 , a00 , b0 , b00 , c, e from 381J; then

482

Automorphism groups

a0 ∪ a00 ∪ b0 ∪ b00 ∪ c = supp π ⊆ a. Set b = b0 ∪ c,

˜b0 = ψb0 ,

c˜ = ψc,

˜b = ψb = ˜b0 ∪ c˜ ⊆ ψa.

Note that b ∩ ˜b ⊆ a ∩ ψa = 0. Set ψ 0 = ψπψ −1 , so that ψ 0˜b = ψπb = ψ(a0 ∪ a00 ) is dist from ψb = ˜b. Set

←−−− ←−−− ←−−−− π1 = (b00 π c)(a0 π b0 )(a00 π b00 )π,

so that (just as in (b) above) π1 is ed by b0 ∪ c = b. Set ←−−− ←−− φ1 = (b ψπ1 ˜b) ∈ I1 , π2 = φ1 (b ψ ˜b)π1 . Examining the action of π2 on Ab , we see that π2 d = φ1 ψπ1 d = π1−1 ψ −1 ψπ1 d = d if d ⊆ b, so that supp π2 ∩ b = 0; but as ←−− supp π2 ⊆ supp φ1 ∪ supp(b ψ ˜b) ∪ supp π1 = b ∪ ˜b, supp π2 ⊆ ˜b. Now we have an alternative expression for π2 , as follows: ←−− π2 = φ1 (b ψ ˜b)π1 ←−−− ←−− ←−−− ←−−− ←−−−− = φ1 (b0 ψ ˜b0 )(c ψ c˜)(b00 π c)(a0 π b0 )(a00 π b00 )π ←−−− ←−−− ←−− ←−−− ←−−−− = φ1 (b0 ψ ˜b0 )(a0 π b0 )(c ψ c˜)(b00 π c)(a00 π b00 )π ←−−− ←−−− ←−− (because (a0 π b0 ) commutes with (b00 π c) and (c ψ c˜)) ←−−− ←−−− ←−−−−−−−−−−−−− = φ1 (b0 ψ ˜b0 )(a0 π b0 )(c π−1 b00 π−1 a00 ψπ2 c˜)π (of course this calculation depends on the fact that c, b00 , a00 , c˜ are all dist) ←−−− ←−−− ←−−− ←−−−− ←−−−− = φ1 (b0 ψ ˜b0 )(a0 π b0 )(b00 π c)(a00 ψπ2 c˜)(b00 ψπ c˜)π ←−−− ←−−− ←−−−− ←−−− ←−−−− = φ1 (b0 ψ ˜b0 )(b00 π c)(a00 ψπ2 c˜)(a0 π b0 )(b00 ψπ c˜)π ←−−− ←−−− ←−−−− (because (a0 π b0 ) commutes with (b00 π c) and (a00 ψπ2 c˜)) = φ1 φ2 φ3 π where

←−−− ←−−− ←−−−− φ2 = (b0 ψ ˜b0 )(b00 π c)(a00 ψπ2 c˜), ←−−− ←−−−− φ3 = (a0 π b0 )(b00 ψπ c˜)

are involutions in G, each being a product of dist involutions. Now supp φ1 ⊆ b ∪ ˜b, supp φ2 ⊆ b ∪ ˜b0 ∪ b00 ∪ c ∪ a00 ∪ c˜, supp φ3 ⊆ a0 ∪ b0 ∪ b00 ∪ c˜ are all included in a ∪ ˜b, and we already know that supp(φ1 φ2 φ3 π) = supp π2 ⊆ ˜b. Q Q

381N

381O


483

(d) We are ready for the assault. Take any π ∈ G. Choose hπn in∈N , han in∈N , hbn in∈N , hφ1n in∈N , hφ2n in∈N , hφ3n in∈N , hψn in∈N as follows. Start by defining a0 , φ1 , φ2 from π as in (b), so that φ1 , φ2 ∈ I1 , πa0 ∩ a0 = 0 and φ1 φ2 π is ed by a0 . Set π0 = φ1 φ2 π, ψ0 = π and b0 = πa0 . For the inductive step, given ψn and πn in G, an ⊇ supp πn and bn = ψn an dist from an , use (c) to find φ1n , φ2n , φ3n ∈ I1 , an+1 ⊆ bn and ψn+1 ∈ G such that ψn+1 an+1 ⊆ bn \ an+1 , πn+1 = φ1n φ2n φ3n πn is ed by an+1 , and all the φin are ed by an ∪ an+1 ; set bn+1 = ψn+1 an+1 , and continue. Evidently hbn in∈N is now non-increasing, while han in∈N is dist; consequently each of the sequences ha2k ∪ a2k+1 ik∈N ,

ha2k+1 ∪ a2k+2 ik∈N

is dist. (e) For r = 0, r = 1 define θr ∈ G by setting −1 θr d = π2k+r π2k+r+1 d if k ∈ N, d ⊆ a2k+r ∪ a2k+r+1 ,

= d if d ∩ sup an = 0. n≥r

(The definition is valid because supp πn ⊆ an for every n, and produces a member of G because G is full.) Now θr ∈ I3 . P P For each n ∈ N, −1 −1 −1 πn πn+1 = φ−1 3n φ2n φ1n ,

and all these automorphisms are ed by an ∪ an+1 . So if we set φ˜ir d = φi,2k+r d if d ⊆ a2k+r ∪ a2k+r+1 , = d if d ∩ sup an = 0, n≥r

each φ˜ir will be in I1 , and ˜−1 ˜−1 θr = φ˜−1 3r φ2r φ1r belongs to I3 . Q Q (f ) Consequently θ1 θ0 ∈ I6 . But θ1 θ0 = π0 . P P If d ⊆ a0 , then θ 0 d = π0 d ⊆ a 0 ,

θ1 θ0 d = π0 d.

If d ⊆ a2k+1 , where k ∈ N, then −1 d ⊆ a2k+1 , θ0 d = π2k+1

−1 d = d = π0 d. θ1 θ0 d = π2k+1 π2k+1

θ0 d = π2k+2 d ⊆ a2k+2 ,

−1 θ1 θ0 d = π2k+2 π2k+2 d = d = π0 d;

If d ⊆ a2k+2 , then and finally, if d ∩ supn∈N an = 0, θ0 d = θ1 d = θ1 θ0 d = π0 d = d. Putting these together, π0 = θ1 θ0 . Q Q (g) Finally, −1 π = φ−1 2 φ1 π0 ∈ I8 .

As π is arbitrary, the theorem is proved. 381O Corollary Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A. Then any π ∈ G other than the identity is expressible as a product of finitely many involutions in G with s included in supp π. proof In fact the construction in the proof of 381N achieves this almost automatically. Alternatively, setting a = supp π, let Aa be the principal ideal of A generated by a, and set

484

Automorphism groups

381O

Ga = {φ¹ Aa : φ ∈ G, supp φ ⊆ a}. It is easy to check that Ga is a full subgroup of Aut Aa , so that π¹ Aa is expressible as a product of involutions belonging to Ga . These are all of the form φi ¹ Aa where φi ∈ G and supp φi ⊆ a, so that every φi is also an involution and π is the product of the φi . 381P Definition Let A be a Boolean algebra, and G a subgroup of the automorphism group Aut A. I will say that G has many involutions if for every non-zero a ∈ A there is an involution π ∈ G which is ed by a. 381Q Lemma Let A be an atomless homogeneous Boolean algebra. Then Aut A has many involutions, and in fact every non-zero element of A is the of an exchanging involution. proof If a ∈ A \ {0}, then there is a b such that 0 6= b ⊂ a. By 381C there is a π ∈ Aut A such that ←−−−− πb = a \ b; now (b π a \ b) is an involution with a. 381R Lemma Let A be a Dedekind complete Boolean algebra, and G a full subgroup of Aut A with many involutions. Then every non-zero element of A is the of an involution belonging to G. proof By the definition 381P, C = {supp π : π ∈ G is an involution} is order-dense in A. So if a ∈ A \ {0} there is a dist B ⊆ C such that sup B = a (313K). For each b ∈ B let πb ∈ G be an involution with B. Define π ∈ G by setting πd = πb d for d ⊆ b ∈ B, πd = d if d ∩ a = 0; then π ∈ G is an involution with a. 381S Theorem Let A be a Dedekind complete Boolean algebra, and G a full subgroup of Aut A with many involutions. Then a subset H of G is a normal subgroup of G iff it is of the form {π : π ∈ G, supp π ∈ I} for some ideal I C A which is G-invariant, that is, such that πa ∈ I for every a ∈ I, π ∈ G. proof (a) I deal with the easy implication first. Let I C A be a G-invariant ideal and set H = {π : π ∈ G, supp π ∈ I}. Because the of the identity automorphism ι is 0 ∈ I, ι ∈ H. If φ, ψ ∈ H and π ∈ G, then supp(φψ) ⊆ supp φ ∪ supp ψ ∈ I, supp(ψ −1 ) = supp ψ ∈ I, supp(πψπ −1 ) = π(supp ψ) ∈ I and φψ, ψ −1 , πψπ −1 all belong to H; so H C G. (b) For the rest of the proof, therefore, I suppose that H is a normal subgroup of G and seek to express it in the given form. We can in fact describe the ideal I immediately, as follows. Set J = {a : a ∈ A, π ∈ H whenever π ∈ G is an involution, supp π ⊆ a}; then 0 ∈ J and a ∈ J whenever a ⊆ b ∈ J. Also πa ∈ J whenever a ∈ J, π ∈ G. P P If φ ∈ G is an involution, supp φ ⊆ πa then φ1 = π −1 φπ is an involution in G and supp φ1 = π −1 (supp φ) ⊆ a, so φ1 ∈ H and φ = πφ1 π −1 ∈ H. As φ is arbitrary, πa ∈ J. Q Q I do not know how to prove directly that J is an ideal, so let us set I = {a1 ∪ a2 ∪ . . . ∪ an : a1 , . . . , an ∈ J}; then I C A, and πa ∈ I for every a ∈ I, π ∈ G.

381T


485

(c) If a ∈ A, ψ ∈ H and a ∩ ψa = 0 then a ∈ J. P P If a = 0, this is trivial. Otherwise, let π ∈ G be an ←−− involution with supp π ⊆ a; say π = (b π c) where b ∪ c ⊆ a. By 381R there is an involution π1 ∈ G such that ←−−−− supp π1 = b; say π1 = (b0 π1 b00 ) where b0 ∪ b00 = b. Set c0 = πb0 ,

c00 = πb00 = c \ c0 ,

←−−−− ←−−−−−−− π2 = π1 ππ1 π −1 = (b0 π1 b00 )(c0 ππ1 π−1 c00 ),

←−−− π3 = (b0 π c0 ),

φ = π2−1 ψπ2 ψ −1 ∈ H, π ¯ = π3−1 φπ3 φ−1 = π3−1 π2−1 ψπ2 ψ −1 π3 ψπ2−1 ψ −1 π2 ∈ H. Now supp(ψπ2 ψ −1 ) = ψ(supp π2 ) = ψ(b ∪ c) ⊆ ψa is dist from supp π3 = b0 ∪ c0 ⊆ a, so π3 commutes with ψπ2 ψ −1 , and π ¯ = π3−1 π2−1 π3 ψπ2 ψ −1 ψπ2−1 ψ −1 π2 = π3−1 π2−1 π3 π2 ←−−− ←−−−− ←−−−−−−− ←−−− ←−−−− ←−−−−−−− = (b0 π c0 )(b0 π1 b00 )(c0 ππ1 π−1 c00 )(b0 π c0 )(b0 π1 b00 )(c0 ππ1 π−1 c00 ) ←−−− ←−−− = (b0 π c0 )(b00 π c00 ) = π. So π ∈ H. As π is arbitrary, a ∈ J. Q Q ←−− (d) If π = (a π b) is an involution in G and a ∈ J, then π ∈ H. P P By 381R again, there is an involution ←−−−− ψ ∈ G such that supp ψ = a; because a ∈ J, ψ ∈ H. Express ψ as (a0 ψ a00 ) where a0 ∪ a00 = a. Set b0 = πa0 , ←−−− ←−−−− b00 = πa00 , so that π = (a0 π b0 )(a00 π b00 ), and ←−−−− ←−−−−−−− ψ1 = ψπψπ −1 = (a0 ψ a00 )(b0 πψπ−1 b00 ) ∈ H. ←−−− As ψ1 (a0 ∪ b0 ) = a00 ∪ b00 is dist from a0 ∪ b0 , a0 ∪ b0 ∈ J, by (c), and π1 = (a0 π b0 ) ∈ H; similarly, ←−−−− a00 ∪ b00 ∈ J, so π2 = (a00 π b00 ) ∈ H and π = π1 π2 belongs to H. Q Q ←−− (e) If π ∈ G is an involution and supp π ∈ I, then π ∈ H. P P Express π as (a π b). Let a1 , . . . , an ∈ J be such that a ∪ b ⊆ a1 ∪ . . . ∪ an . Set ←−−− cj = a ∩ aj \ supi<j ai , bj = πcj , πj = (cj π bj ) for 1 ≤ j ≤ n; then every cj belongs to J, so every πj belongs to H (by (d)) and π = π1 . . . πn ∈ H. Q Q (f ) If π ∈ G and supp π ∈ I then π ∈ H. P P By 381O, π is a product of involutions in G all with s included in supp π; by (e), they all belong to H, so π also does. Q Q (g) We are nearly home. So far we know that I is a G-invariant ideal and that π ∈ H whenever π ∈ G, supp π ∈ I. On the other hand, supp π ∈ I for every π ∈ H. P P Take a0 , a00 , b0 , b00 , c from Frol´ık’s theorem (381J). Then a0 ∩ πa0 = b0 ∩ πb0 = . . . = c ∩ πc = 0, so a0 , . . . , c all belong to J, by (c), and supp π = a0 ∪ . . . ∪ c belongs to I. Q Q So H is precisely the set of of G with s in I, as required. 381T Corollary Let A be a homogeneous Dedekind complete Boolean algebra. Then Aut A is simple.

486

Automorphism groups

381T

proof If A is {0} or {0, 1} this is trivial. Otherwise, let H be a normal subgroup of Aut A. Then by 381S and 381Q there is an invariant ideal I of A such that H = {π : supp π ∈ I}. But if H is non-trivial so is I; say a ∈ I \ {0}. If a = 1 then certainly 1 ∈ I and H = Aut A. Otherwise, there is a π ∈ Aut A such that πa = 1 \ a (as in 381C), so 1 \ a ∈ I, and again 1 ∈ I and H = Aut A. Remark I ought to remark that in fact Aut A is simple for any homogeneous Dedekind σ-complete Boolean ˇ e ˇpa ńek & Rubin 89, Theorem 5.9b. algebra; see St 381X Basic exercises > (a) Let X be a set and Σ an algebra of subsets of X containing all singleton sets. Show that Aut Σ can be identified with the group of bijections φ : X → X such that φ[E] ∈ Σ, φ−1 [E] ∈ Σ for every E ∈ Σ. > (b) Let A be a Boolean algebra and G any subgroup of Aut A. Let H be the set of those π ∈ Aut A such that for every non-zero a ∈ A there are a non-zero b ⊆ a and a φ ∈ G such that πc = φc for every c ⊆ b. Show that H is a full subgroup of Aut A, the smallest full subgroup of A including G. > (c) Let A be a Dedekind complete Boolean algebra and G any subgroup of A. Show that an element π of Aut A belongs to the full subgroup of Aut A generated by G iff there are a partition of unity hai ii∈I in A and a family hπi ii∈I in G such that πa = πi a whenever i ∈ I and a ⊆ ai . (d) For a Boolean algebra A, let us say that a subgroup G of Aut A is countably full if whenever hai ii∈I is a countable partition of unity in A, hπi ii∈I is a family in G, and π ∈ Aut A is such that πa = πi ai whenever i ∈ I and a ⊆ ai , then π ∈ G. Show that if A is a Dedekind complete Boolean algebra and G is a countably full subgroup of Aut A, then every member of G is expressible as a product of at most eight involutions belonging to G. > (e) Let X be any set. Show that any automorphism of the Boolean algebra PX is expressible as a product of at most two involutions. (f ) Recall that in any group G, a commutator in G is an element of the form ghg −1 h−1 where g, h ∈ G. Show that if A is a Dedekind complete Boolean algebra and G is a subgroup of Aut A with many involutions then every involution in G is a commutator in G, so that every element of G is expressible as a product of finitely many commutators. (g) Give an example of a Dedekind complete Boolean algebra A such that not every member of Aut A is a product of commutators in Aut A. (h) Let A be a Dedekind complete Boolean algebra, and suppose that Aut A has many involutions. Show that if H C Aut A then every member of H is expressible as the product of at most eight involutions belonging to H. (i) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions. Show that the partially ordered set H of normal subgroups of G is a distributive lattice, that is, H ∩ K1 K2 = (H ∩ K1 )(H ∩ K2 ), H(K1 ∩ K2 ) = HK1 ∩ HK2 for all H, K1 , K2 ∈ H. (j) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions. Show that if H is a the normal subgroup of G generated by a finite subset of G, then it is the normal subgroup generated by a single involution. (k) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions. Show (i) that there is an involution π ∈ G such that every member of G is expressible as a product of conjugates of π in G (ii) any proper normal subgroup of G is included in a maximal proper normal subgroup of G. (l) Let G be any group. Show that if π, φ ∈ G are involutions then πφ is conjugate to its inverse.

381 Notes


487

381Y Further exercises (a) Find a Dedekind σ-complete Boolean algebra A with an automorphism which cannot be expressed either as a product of finitely many involutions in Aut A, nor as a product of finitely many commutators in Aut A. (This seems to require a certain amount of ingenuity.) (b) Let X be a set and Σ a countably generated σ-subalgebra of subsets of X. (i) Show that if f : X → X is a bijection such that Σ = {E : E ⊆ X, f −1 [E] ∈ Σ}, then there are dist E 0 , E 00 , F 0 , F 00 , G ∈ Σ such that f −1 [E 0 ] = F 0 , f −1 [E 00 ] = F 00 , f −1 [F 00 ] = G, f −1 [F 0 ∪ G] = E 0 ∪ E 00 , and f (x) = x for every x ∈ X \ (E 0 ∪SE 00 ∪ F 0 ∪ F 00 ∪ G. (Hint: there is a sequence hEn in∈N in Σ such that En ∩ f −1 [En ] = ∅ for every n and n∈N En = {x : f (x) 6= x}.) (In particular, any member of Aut Σ has a .) (ii) Show that every involution in Aut Σ is an exchanging involution. (c) Let X be a set and Σ a countably generated σ-subalgebra of subsets of X. Show that any member of Aut Σ is expressible as a product of at most eight involutions in Aut Σ. (d) Let A be a homogeneous Boolean algebra which is isomorphic to the simple power AN . (For instance, A could be the measure algebra of Lebesgue measure on R.) Show that any automorphism of A is the ˇ e ˇpa ´ nek & Rubin 89, Corollary 5.9a(ii).) product of at most five exchanging involutions. (Cf. St 381Z Problem In 381N, is ‘eight’ best possible? ˇ e ˇpa ńek & Rubin 89 and Fathi 78. 381 Notes and comments The ideas above are adapted from St Lemma 381J is a form of what is sometimes called ‘Frol´ık’s theorem’, following Frol´ık 68. The two main results 381N and 381S, as written out above, both involve careful algebra. It seems to me that we can distinguish two essential methods. (i) There are arguments involving finitely many automorphisms, carefully pieced together from descriptions of their actions on different parts of the algebra, as in (a)-(c) of the proof of 381N, and the whole of the proof of 381S; similar ideas can be used in 381Xf. It is in these that I believe that the ‘cycle notation’ introduced in 381G-381I can be of value. Generally the hope is that we can use intuitions derived from the theory of permutation groups (that is, the case A = PX) to guide us. (ii) There is the argument in 381N involving a sequence of automorphisms, designed to express an automorphism π0 ed by an element a0 as the product of an automorphism θ0 ed by supk∈N ak with an automorphism θ1 ed by supk≥1 ak , so chosen that the actions of the θr on supk≥1 ak cancel ˇ e ˇpa ´ nek & Rubin out and we are left with π0 as the residue. (For an of the origins of this idea see St 89.) Since we know of no automorphisms except those which can be derived from the original automorphism π, the method has to be to some extent constructive. The idea is that each πn+1 is not exactly a copy on an+1 of the preceding πn , but a modification of it by involutions. At each stage of the induction we have to mention an auxiliary element ψn of G in order to be sure that there will be room (in bn = ψn an ) for the next step, safely dist from the preceding ak . When we come to build the mutually cancelling pair θ0 , θ1 we find that they incorporate the modifiers φ1n , φ2n , φ3n , which can be assembled into the involutions φ˜ir in part (f) of the proof. I note that the assumption of ‘Dedekind completeness’ (as opposed to Dedekind σ-completeness) in 381N is used only in parts (b) and (c) of the proof, when applying Frol´ık’s theorem. Consequently we have a slight generalization possible (381Xd); but we do need the full hypothesis for the theorem as stated (381Ya). There is however a very important special case, when A is a countably generated σ-algebra, for which we have a version of Frol´ık’s theorem available for different reasons (381Yb), and can get a corresponding theorem to match 381N (381Yc). A natural question arising from 381T is: does every homogeneous Boolean algebra have a simple automorphism group? This leads into deep water. As remarked after 381T, every homogeneous Dedekind σ-complete algebra has a simple automophism group. Using the continuum hypothesis, it is possible to construct a homogeneous Boolean algebra which does not have a simple automorphism group; but as far as I am aware no such construction is known which does not rely on some special axiom outside ordinary set ˇ e ˇpa ´ nek & Rubin 89, §5. theory. See St In 381Z I ask whether the number ‘eight’ appearing in 381N is actually best possible. The argument is complex enough to make it seem that there may be room for improvement – see 381Xe and 381Yd. Ornstein

488

Automorphism groups

381 Notes

& Shields 731 present examples of automorphisms in the full subgroup G = Autµ¯ A of measure-preserving automorphisms of the Lebesgue probability algebra which are not conjugate (in G) to their inverses, and therefore cannot be expressible as the product of two involutions in G.

382 Automorphism groups of measure algebras I turn now to the group of measure-preserving automorphisms of a measure algebra, seeking to apply the results of the last section. The principal theorems are 382D, which is a straightforward special case of 381N, and 382I, corresponding to 381T. I give another example of the use of 381S to describe the normal subgroups of Autµ¯ A (382J). 382A Definition Let (A, µ ¯) be a measure algebra. I will write Autµ¯ A for the set of all measurepreserving automorphisms of A. This is a group, being a subgroup of the group Aut A of all Boolean automorphisms of A. 382B Lemma Let (A, µ ¯) be a measure algebra, and hai ii∈I , hbi ii∈I two partitions of unity in A. Assume either that I is countable or that (A, µ ¯) is localizable. Suppose that for each i ∈ I we have a measure-preserving isomorphism πi : Aai → Abi between the corresponding principal ideals. Then there is a unique π ∈ Autµ¯ A such that πc = πi c whenever i ∈ I and c ⊆ ai . Q Q proof (Compare 381B.) By 322K, we may identify A with each of the simple products i∈I Aai , i∈I Abi ; now π corresponds to the isomorphism between the two products induced by the πi . 382C Corollary If (A, µ ¯) is a localizable measure algebra, then, in the language of 381M, Autµ¯ A is a full subgroup of Aut A. 382D Theorem Let (A, µ ¯) be a localizable measure algebra. Then every measure-preserving automorphism of A is expressible as the product of at most eight measure-preserving involutions. proof This is immediate from 382C and 381N. 382E Lemma If (A, µ ¯) is a homogeneous semi-finite measure algebra, it is σ-finite, therefore localizable. proof If A = {0}, this is trivial. Otherwise there is an a ∈ A such that 0 < µ ¯a < ∞. The principal ideal Aa is ccc (322G), so A also is, and (A, µ ¯) must be σ-finite (by 322G again). 382F Lemma Let (A, µ ¯) be a homogeneous semi-finite measure algebra. (a) If hai ii∈I , hbi ii∈I are partitions of unity in A with µ ¯ai = µ ¯bi for every i, there is a π ∈ Autµ¯ A such that πai = bi for each i. (b) If (A, µ ¯) is totally finite, then whenever hai ii∈I , hbi ii∈I are dist families in A with µ ¯ ai = µ ¯bi for every i, there is a π ∈ Autµ¯ A such that πai = bi for each i. proof (a) By 382E, (A, µ ¯) is σ-finite, therefore localizable. For each i ∈ I, the principal ideals Aai , Abi are homogeneous, of the same measure and the same Maharam type (being τ (A) if ai 6= 0, 0 if ai = 0). Because they are ccc, they are of the same magnitude, as defined in 332G, and there is a measure-preserving isomorphism πi : Aai → Abi (332J). By 382B there is a measure-preserving automorphism π : A → A such that πd = πi d for every i ∈ I, d ⊆ ai ; and this π serves. (b) Set a∗ = 1 \ supi∈I ai , b∗ = 1 \ supi∈I bi . We must have P P µ ¯ a∗ = µ ¯1 − i∈I µ ¯ ai = µ ¯1 − i∈I µ ¯ bi = µ ¯b∗ , so adding a∗ , b∗ to the families we obtain partitions of unity to which we can apply the result of (a). 1I

am indebted to G.Hjorth for the reference.

382J

Automorphism groups of measure algebras

489

382G Lemma (a) If (A, µ ¯) is an atomless semi-finite measure algebra, then Aut A and Autµ¯ A have many involutions. (b) If (A, µ ¯) is an atomless localizable measure algebra, then every element of A is the of some involution in Autµ¯ A. proof (a) If a ∈ A \ {0}, then by 332A there is a non-zero b ⊆ a, of finite measure, such that the principal ideal Ab is (Maharam-type-)homogeneous. Now because A is atomless, there is a c ⊆ b such that µ ¯c = 21 µ ¯b (331C), so that Ac and Ab\c are isomorphic measure algebras. If θ : Ac → Ab\c is any measure-preserving ←−−−− isomorphism, then π = (c θ b \ c) is an involution in Autµ¯ A (and therefore in Aut A) ed by a. (b) Use 382C, (a) and 381R. 382H Corollary Let (A, µ ¯) be an atomless localizable measure algebra. Then (a) the lattice of normal subgroups of Aut A is isomorphic to the lattice of Aut A-invariant ideals of A; (b) the lattice of normal subgroups of Autµ¯ A is isomorphic to the lattice of Autµ¯ A-invariant ideals of A. proof Use 381S. Taking G to be either Aut A or Autµ¯ A, and I to be the family of G-invariant ideals in A, we have a map I 7→ HI = {π : π ∈ G, supp π ∈ I} from I to the family H of normal subgroups of G. Of course this map is order-preserving; 381S tells us that it is surjective; and 382Gb tells us that it is injective and its inverse is order-preserving, since if a ∈ I \ J there is a π ∈ G with supp π = a, so that π ∈ HI \ HJ . Thus we have an order-isomorphism between H and I. 382I Normal subgroups of Aut A and Autµ¯ A 381S provides the machinery for a full description of the normal subgroups of Aut A and Autµ¯ A when (A, µ ¯) is an atomless localizable measure algebra, as we know that they correspond exactly to the invariant ideals of A. The general case is complicated. But the following are simple enough. Theorem Let (A, µ ¯) be a homogeneous semi-finite measure algebra. (a) Aut A is simple. (b) If (A, µ ¯) is totally finite, Autµ¯ A is simple. (c) If (A, µ ¯) is not totally finite, Autµ¯ A has exactly one non-trivial proper normal subgroup. proof (a) This is a special case of 381T. (b)-(c) The point is that the only possible Autµ¯ A-invariant ideals of A are {0}, Af and A. P P If A is {0} or {0, 1} this is trivial. Otherwise, A is atomless. Let I C A be an invariant ideal. (i) If I 6⊆ Af , take a ∈ I with µ ¯a = ∞. By 382E, A is σ-finite, so a has the same magnitude ω as 1. By 332I, there is a partition of unity hen in∈N in A with µ ¯en = 1 for every n; setting b = supn∈N e2n , b0 = 1 \ b, we 0 see that both b and b are of infinite measure. Similarly we can divide a into c, c0 , both of infinite measure. Now by 332J the principal ideals Ab , Ab0 , Ac , A1\c are all isomorphic as measure algebras, so that there are automorphisms π, φ ∈ Autµ¯ A such that πc = b,

φc = b0 .

But this means that both b and b0 belong to I, so that 1 = b ∪ b0 ∈ I and I = A. (ii) If I ⊆ Af and I 6= {0}, take any non-zero a ∈ I. If b is any member of A, then (because A is atomless) b can be partitioned into b0 , . . . , bn , all of measure at most µ ¯a. Then for each i there is a b0i ⊆ a 0 0 such that µ ¯bi = µ ¯bi ; since this common measure is finite, µ ¯(1 \ bi ) = µ ¯(1 \ bi ). By 332J and 382Fa, there is a πi ∈ Autµ¯ A such that πi b0i = bi , so that bi belongs to I. Accordingly b ∈ I. As b is arbitrary, I = Af . Thus the only invariant ideals of A are {0}, Af and A. Q Q By 382Hb we therefore have either one, two or three normal subgroups of Autµ¯ A, according to whether µ ¯1 is zero, finite and not zero, or infinite. Remark For the Lebesgue probability algebra, (b) is due to Fathi 78. The extension to algebras of uncountable Maharam type is from Choksi & Prasad 82. 382J

The language of §352 offers a way of describing another case.

490

Automorphism groups

382J

Proposition Let (A, µ ¯) be an atomless totally finite measure algebra. For each infinite cardinal κ, let eκ be the Maharam-type-κ component of A, and let K be {κ : eκ 6= 0}. Let H be the lattice of normal subgroups of Autµ¯ A. Then (i) if K is finite, H is isomorphic, as partially ordered set, to PK; (ii) if K is infinite, then H is isomorphic, as partially ordered set, to the lattice of solid linear subspaces of `∞ . proof (a) Let I be the family of Autµ¯ A-invariant ideals of A, so that H ∼ = I, by 382Hb. For a, b ∈ A, say that a ¹ b if there is some k ∈ N such that µ ¯(a ∩ eκ ) ≤ k µ ¯(b ∩ eκ ) for every κ ∈ K. Then an ideal I of A is Autµ¯ A-invariant iff a ∈ I whenever a ¹ b ∈ I. P P (α) Suppose that I is Autµ¯ A invariant and that b ∈ I, µ ¯(a ∩ eκ ) ≤ k µ ¯(b ∩ eκ ) for every κ ∈ K. Then for each κ we can find aκ1 , . . . , aκk such that a ∩ eκ = supi≤k aκi and µ ¯aκi ≤ µ ¯(b ∩ eκ ) for every i. Now there are measure-preserving automorphisms πκi of the principal ideal Aeκ such that πκi aκi ⊆ b. Setting πi d = supκ∈K πκi (d ∩ eκ ) for every d ∈ A, and ai = supκ∈K aκi , we have πi ∈ Autµ¯ A and πi ai ⊆ b, so ai ∈ I for each i; also a = supi≤k ai , so a ∈ I. (β) On the other hand, if a ∈ A and π ∈ Autµ¯ A, then µ ¯(πa ∩ eκ ) = µ ¯π(a ∩ eκ ) = µ ¯(a ∩ eκ ) for every κ ∈ K, because πeκ = eκ , so that πa ¹ a. So if I satisfies the condition, π[I] ⊆ I for every π ∈ Autµ¯ A and I ∈ I. Q Q (b) Consequently, for I ∈ I and κ ∈ K, eκ ∈ I iff there is some a ∈ I such that a ∩ aκ 6= 0, since in this case eκ ¹ a. (This is where I use the hypothesis that (A, µ ¯) is totally finite.) It follows that if K is finite, any I ∈ I is the principal ideal generated by sup{eκ : eκ ∈ I}. Conversely, of course, all such ideals are Autµ¯ A-invariant. Thus I is in a natural order-preserving correspondence with PK, and H ∼ = PK. (c) Now suppose that K is infinite; enumerate it as hκn in∈N . Define θ : A → `∞ by setting θa = h¯ µ(a ∩ eκn )/¯ µ(eκn )in∈N for a ∈ A; so that a ¹ b iff there is some k such that θa ≤ kθb, θa ≤ θ(a ∪ b) ≤ θa + θb ≤ 2θ(a ∪ b) for all a, b ∈ A, while θ(1A ) is the standard order unit 1 of `∞ . Let U be the family of solid linear subspaces of `∞ and define functions I 7→ VI : I → U , U 7→ JU : U → I by saying VI = {f : f ∈ `∞ , |f | ≤ kθa for some a ∈ I, k ∈ N}, JU = {a : a ∈ A, θa ∈ U }. The properties of θ just listed ensure that VI ∈ U , JU ∈ I for every I ∈ I, U ∈ U. Of course both I 7→ VI and U 7→ JU are order-preserving. If I ∈ I, then JVI = {a : ∃ b ∈ I, a ¹ b} = I, Finally, VJU = U for every U ∈ U . P P VJU = {f : ∃ a ∈ A, k ∈ N, |f | ≤ kθa ∈ U } ⊆ U because U is a solid linear subspace. But also, given g ∈ U , there is an a ∈ A such that µ ¯(a ∩ eκn ) = min(1, |g(n)|)¯ µ(eκn ) for every n (because A is atomless); in which case θa ≤ |g| ≤ max(1, kgk∞ )θa so a ∈ JU and g ∈ VJU . Thus U = VJU . Q Q So the functions I 7→ VI and U 7→ JU are the two halves of an order-isomorphism between I and U , and H ∼ =I∼ = U, as claimed. 382X Basic exercises >(a) Let (A, µ ¯) be a localizable measure algebra. For each infinite cardinal κ, let eκ be the Maharam-type-κ component of A. (i) Show that Autµ¯ A is a simple group iff either there is just one infinite cardinal κ such that eκ 6= 0, that eκ has finite measure and all the atoms of A (if any) have different measures or A is purely atomic and there is just one pair of atoms of the same measure or A is purely atomic and all its atoms have different measures. (ii) Show that Aut A is a simple group iff either

382Ye

Automorphism groups of measure algebras

491

(A, µ ¯) is σ-finite and there is just one infinite cardinal κ such that eκ 6= 0 and A has at most one atom or A is purely atomic and has at most two atoms. (b) Let (A, µ ¯) be a localizable measure algebra. (i) Show that Autµ¯ A is simple iff it is isomorphic to one of the groups {ι}, Z2 or Autν¯κ Bκ where κ is an infinite cardinal and (Bκ , ν¯κ ) is the measure algebra of the usual measure on {0, 1}κ . (ii) Show that Aut A is simple iff it is isomorphic to one of the groups {ι}, Z2 or Aut Bκ . (c) Show that if (A, µ ¯) is a semi-finite measure algebra of magnitude greater than c, its automorphism group Autµ¯ A is not simple. (d) Let (A, µ ¯) be an atomless localizable measure algebra. For each infinite cardinal κ write eκ for the Maharam-type-κ component of A. For π, ψ ∈ Autµ¯ A show that π belongs to the normal subgroup of Autµ¯ A generated by ψ iff there is a k ∈ N such that mag(eκ ∩ supp π) ≤ k mag(eκ ∩ supp ψ) for every infinite cardinal κ, writing mag a for the magnitude of a, and setting kζ = ζ if k > 0 and ζ is an infinite cardinal. > (e) Let (A, µ ¯) be the measure algebra of Lebesgue measure on R. For n ∈ N set en = [−n, n]• ∈ A. Let G ≤ Autµ¯ A be the group consisting of measure-preserving automorphisms π such that supp π ⊆ en for some n. Show that G is simple. (Hint: show that G is the union of an increasing sequence of simple subgroups.) (f ) Let (A, µ ¯) be an atomless totally finite measure algebra. Let H be the lattice of normal subgroups of Aut A. Show that H is isomorphic, as partially ordered set, to PK for some countable set K. (g) Let (A, µ ¯) be an atomless localizable measure algebra which is not σ-finite, and suppose that τ (Aa ) = τ (Ab ) whenever a, b ∈ A and 0 < µ ¯a ≤ µ ¯b < ∞. Let κ be the magnitude of A. (i) Show that the lattice H of normal subgroups of Autµ¯ A is well-ordered, with least member {ι}, next member {π : µ ¯(πa) < ∞ whenever µ ¯a < ∞}, and one member Hζ for each infinite cardinal ζ less than or equal to κ, setting Hζ = {π : π ∈ Autµ¯ A, mag(πa) ≤ ζ whenever mag a ≤ ζ}, where mag a is the magnitude of a. (ii) Show that the lattice H0 of normal subgroups of Aut A is wellordered, with least member {ι} and one member Hζ0 for each infinite cardinal ζ less than or equal to κ, setting Hζ0 = {π : π ∈ Aut A, mag(πa) ≤ ζ whenever mag a ≤ ζ}. 382Y Further exercises (a) Let (A, µ ¯) be an atomless totally finite measure algebra. Show that Autµ¯ A, Aut A have the same (cardinal) number of normal subgroups. (b) Let X be a set. Show that Aut PX has one normal subgroup if #(X) ≤ 1, two if #(X) = 2, three if #(X) = 3 or 5 ≤ #(X) ≤ ω, four if #(X) = 4 or #(X) = ω1 . (c) Let (A, µ ¯) be a totally finite measure algebra and π : A → A a measure-preserving homomorphism. Take any a ∈ A and set c = supn≥1 π n a. (i) Show that a ⊆ c and that c = supm≥n π m a for every n ∈ N. P∞ (ii) Set an = π n a \ sup1≤i
492

Automorphism groups

382Ye

and a closed subalgebra C of B such that φ[C] ⊆ C and (C, ν¯¹ C, φ¹ C) is isomorphic to (A, µ ¯, π). (Hint: Take P∞ B1 ⊆ AN to be the subalgebra consisting of sequences a = (α0 , α1 , . . . ) such that n=0 µ ¯(αn+1 4 παn ) < ∞. For a ∈ B1 set ν¯1 a = limn→∞ µ ¯αn , φ1 a = (α1 , α2 , . . . ). Let I be the ideal {a : ν¯1 a = 0} and let B be the quotient B1 /I; define φ : B → B by setting φa• = (φ1 a)• . For a ∈ A let a∗ ∈ B be the equivalence class of the sequence hπ n ain∈N ; set C = {a∗ : a ∈ A}.) (This is an abstract version of a construction known as the ‘natural extension’ of an inverse-measure-preserving function; see Petersen 83, 1.3G.) (f ) Let (X, Σ, µ) be a measure space in which Σ is countably generated as σ-algebra, and write Autµ Σ for the group of automorphisms φ : Σ → Σ such that µφ(E) = µE for every E ∈ Σ. Show that every member of Autµ Σ is expressible as a product of at most eight involutions belonging to Autµ Σ. (Hint: 381Yc.) 382 Notes and comments This section is short because there are no substantial new techniques to be developed. 382D is simply a matter of checking that the hypotheses of 381N are satisfied (and these hypotheses were of course chosen with 382D in mind), and 382I is similarly direct from 381S. 382I-382J, 382Xd and 382Xg are variations on a theme. In a general Boolean algebra A with a group G of automorphisms, we have a transitive, reflexive relation ¹G defined by saying that a ¹G b if there are π1 , . . . , πk ∈ G such that a ⊆ supi≤k πi b; the point about localizable measure algebras is that the functions ‘Maharam type’ and ‘magnitude’ enable us to describe this relation when G = Autµ¯ A, and the essence of 381S is that in that context π belongs to the normal subgroup of G generated by ψ iff supp π ¹G supp ψ. Most of the work of this chapter is focused on atomless measure algebras. There are various extra complications which appear if we allow atoms. The most striking are in the next section; here I mention only 382Xa and 382Yb.

383 Outer automorphisms Continuing with the investigation of the abstract group-theoretic nature of the automorphism groups Aut A, Autµ¯ A, I devote a section to some remarkable results concerning isomorphisms between them. Under any of a variety of conditions, any isomorphism between two groups Aut A, Aut B must correspond to an isomorphism between the underlying Boolean algebras (383E, 383F, 383J, 383M); consequently Aut A has few, or no, outer automorphisms (383G, 383K, 383O). I organise the section around a single general result (383D). 383A Lemma Let A be a Boolean algebra and G a subgroup of Aut A which has many involutions (definition: 381P). Then for every non-zero a ∈ A there is an automorphism ψ ∈ G, of order 4, which is ed by a. proof Let π ∈ G be an involution ed by a. Let b ⊆ a be such that πb 6= b. Then at least one of b \ πb, πb \ b = π(b \ πb) is non-zero, so in fact both are. Let φ be an involution ed by b \ πb. Then πφπ = πφπ −1 is an involution ed by πb \ b, so commutes with φ, and the product φπφπ is an involution. But this means that ψ = φπ has order 4, and of course it is ed by a because φ and π both are. 383B A note on s Since in this section we shall be looking at more than one automorphism group at a time, I shall need to call on the following elementary extension of a fact in 381Ec. Let A and B be Boolean algebras, and θ : A → B a Boolean isomorphism. If π ∈ Aut A is ed by a ∈ A, then θπθ−1 ∈ Aut B is ed by θa. (Use the same argument as in 381Ec.) Accordingly, if a is the of π then θa will be the of θπθ−1 , as in 381L. 383C Lemma Let A and B be two Boolean algebras, and G a subgroup of Aut A with many involutions. If θ1 , θ2 : A → B are distinct isomorphisms, then there is a φ ∈ G such that θ1 φθ1−1 6= θ2 φθ2−1 . proof Because θ1 6= θ2 , θ = θ2−1 θ1 is not the identity automorphism on A, and there is some non-zero a ∈ A such that θa ∩ a = 0. Let π ∈ G be an involution ed by a; then θπθ−1 is ed by θa, so cannot be equal to π, and θ1 πθ1−1 6= θ2 πθ2−1 .

383D

Outer automorphisms

493

383D Theorem Let A and B be Dedekind complete Boolean algebras and G and H subgroups of Aut A, Aut B respectively, both having many involutions. Let q : G → H be an isomorphism. Then there is a unique Boolean isomorphism θ : A → B such that q(φ) = θφθ−1 for every φ ∈ G. proof (a) The first half of the proof is devoted to setting up some structures in the group G. Let π ∈ G be any involution. Set Cπ = {φ : φ ∈ G, φπ = πφ}, the centralizer of π in G; Uπ = {φ : φ ∈ Cπ , φ = φ−1 , φψφψ −1 = ψφψ −1 φ for every ψ ∈ Cπ }, the set of involutions in Cπ commuting with all their conjugates in Cπ , together with the identity, Vπ = {φ : φ ∈ G, φψ = ψφ for every ψ ∈ Uφ }, the centralizer of Uπ in G, Sπ = {φ2 : φ ∈ Vπ }, Wπ = {φ : φ ∈ G, φψ = ψφ for every ψ ∈ Sπ }, the centralizer of Sπ in G. (b) The point of this list is to provide a purely group-theoretic construction corresponding to the of π in A. In the next few paragraphs of the proof (down to (f)), I set out to describe the objects just introduced in of their action on A. First, note that π is an exchanging involution (381K); express it ←−−− as (a0 π a00 ), so that the of π is aπ = a0 ∪ a00 . (c) I start with two elementary properties of Cπ : (i) φ(aπ ) = aπ for every φ ∈ Cπ . P P As remarked in 381L, the of π = φπφ−1 is φ(aπ ), so this must be aπ . Q Q (ii) If φ ∈ Cπ and φ is not ed by aπ , there is a non-zero d ⊆ 1 \ aπ such that d ∩ φd = 0, by 381Ef. (d) Now for the properties of Uπ : (i) If φ ∈ Uπ , then φ is ed by aπ . α)?? Suppose first that there is a d ⊆ 1 \ aπ such that d ∩ (φd ∪ φ2 d) = 0. Let ψ ∈ G be an involution P P(α ed by d. Then supp ψ ∩ supp π = 0, so ψ ∈ Cπ . There is a c ⊆ d such that ψc 6= c, so ψφψ −1 φc = ψφ2 c = φ2 c, because d ∩ (φc ∪ φ2 c) = 0, while φψφψ −1 c = φ2 ψ −1 c, because d ∩ φψ −1 c = 0; but this means that ψφψ −1 φc 6= φψφψ −1 c, so φ and ψφψ −1 do not commute, and φ∈ / Uπ . X X β )?? Suppose that φ2 is not ed by aπ . Then, as remarked in (b), there is a non-zero (β d ⊆ 1 \ aπ such that φ2 d ∩ d = 0. Now d 6⊆ φ2 d, so d 6⊆ φd; set d0 = d \ φd. Then d0 ∩ φd0 = d0 ∩ φ2 d0 = 0 and 0 6= d0 ⊆ 1 \ aπ ; but this is impossible, by (α). X X (γγ ) Thus φ2 d = d for every d ⊆ 1 \ aπ . ?? Suppose, if possible, that φ is not ed by aπ . Then there is a non-zero d ⊆ 1 \ aπ such that φd ∩ d = 0. By 383A, there is a ψ ∈ G, of order 4, ed by d. Because d ∩ aπ = 0, ψ ∈ Cπ . Because ψ 6= ψ −1 , there is a c ⊆ d such that ψc 6= ψ −1 c; but now φc ∩ d = φψ −1 c ∩ d = 0, so ψφψ −1 φc = ψφ2 c = ψc 6= ψ −1 c = φ2 ψ −1 c = φψφψ −1 c, and φ does not commute with its conjugate ψφψ −1 , contradicting the assumption that φ ∈ Uπ . X X So φ is ed by aπ , as claimed. Q Q (ii) If u ∈ A and πu = u, then πu ∈ Uπ , where

494

Automorphism groups

383D

πu d = πd if d ⊆ u, πu d = d if d ∩ u = 0, ←0−−−−−−00−−− that is, πu = (a ∩ u π a ∩ u). P P For any ψ ∈ Aut A, ←−−−−−−−−−−−−−−−−−− ψπu ψ −1 = (ψ(a0 ∩ u) ψπψ−1 ψ(a00 ∩ u)) (381Hb). (α) Accordingly ←−−−−−−−−− ππu π −1 = (a00 ∩ u π a0 ∩ u) = πu and πu ∈ Cπ . (β) If ψ ∈ Cπ , then ←−−−−−−−−−− ←−−−−−− π = ψπψ −1 = (ψa0 ψπψ−1 ψa00 ) = (ψa0 π ψa00 ). So

←−−−−−−−−−−−−−−−−−− ←−−−−−−−−−−−−−− ψπu ψ −1 = (ψ(a0 ∩ u) ψπψ−1 ψ(a00 ∩ u)) = (ψa0 ∩ ψu π ψa00 ∩ ψu) = πψu .

Now if πv = v then πu πv = πu4v = πv πu ; in particular, πψu πu = πu πψu . As ψ is arbitrary, πu ∈ Uπ . Q Q In particular, of course, π = π1 belongs to Uπ . (e) The two parts of (d) lead directly to the properties we need of Vπ . (i) Vπ ⊆ Cπ , because π ∈ Uπ . Consequently φaπ = aπ for every φ ∈ Vπ . (ii) If φ ∈ Vπ then φd ⊆ d ∪ πd for every d ⊆ aπ . P P?? Suppose, if possible, otherwise. Set u0 = d ∪ πd, so that πu0 = u0 , and u = φu0 \ u0 6= 0; also u ⊆ φaπ = aπ . Since πφu0 = φπu0 = φu0 , πu = u. Set v = u ∩ a0 , so that u = v ∪ πv and v 6= πv. Because u ∩ φv ⊆ φ(u0 ∩ u) = 0, πu φv = φv 6= φπv = φπu v, which is impossible. X XQ Q (iii) It follows that φ2 d = d whenever φ ∈ Vπ and d ⊆ aπ . P P Let e be the of φ. Recall that e = sup{c : c ∩ φc = 0} (381L), so that d ∩ e = sup{c : c ⊆ d, c ∩ φc = 0}. Now if c ⊆ aπ and c ∩ φc = 0, we know that φc ⊆ c ∪ πc, so in fact φc ⊆ πc. This shows that φ(d ∩ e) ⊆ π(d ∩ e). Also, because πφ = φπ, by (i), we have φ2 (d ∩ e) ⊆ φπ(d ∩ e) = πφ(d ∩ e) ⊆ π 2 (d ∩ e) = d ∩ e. Of course φ2 (d \ e) = d \ e, so φ2 d ⊆ d. This is true for every d ⊆ aπ . But as also φ2 aπ = φaπ = aπ , φ2 d = d for every d ⊆ aπ . Q Q (iv) The final thing we need to know about Vπ is that φ ∈ Vπ whenever φ ∈ G and supp φ ∩ aπ = 0; this is immediate from (d-i) above. (f ) From (e-iii), we see that if φ ∈ Sπ then supp φ ∩ aπ = 0. But we also see from (e-iv) that if 0 6= c ⊆ 1 \ aπ there is an involution in Sπ ed by c; for there is a member ψ of G, of order 4, ed by c, and now ψ ∈ Vπ so ψ 2 ∈ Sπ , while ψ 2 is an involution. (g) Consequently, Wπ is just the set of of G ed by aπ . P P (i) If supp φ ⊆ aπ and ψ ∈ Sπ , then supp ψ ∩ aπ = 0, as noted in (e), so φψ = ψφ; as ψ is arbitrary, φ ∈ Wπ . (ii) If supp φ 6⊆ aπ , then take a non-zero d ⊆ 1 \ aπ such that φd ∩ d = 0. Let ψ ∈ Sπ be an involution ed by d; then if c ⊆ d is such that ψc 6= c, φψc 6= φc = ψφc, and φψ 6= ψφ so φ ∈ / Wπ . Q Q (h) We can now return to consider the isomorphism q : G → H. If π ∈ G is an involution, then q(π) ∈ H is an involution, and it is easy to check that q[Cπ ] = Cq(π) , q[Uπ ] = Uq(π) , q[Vπ ] = Vq(π) ,

383E

Outer automorphisms

495

q[Sπ ] = Sq(π) , q[Wπ ] = Wq(π) , defining Cq(π) , . . . , Wq(π) ⊆ H as in (a) above. So we see that, for any φ ∈ G, supp φ ⊆ supp π ⇐⇒ supp q(φ) ⊆ supp q(π). (i) Define θ : A → B by writing θa = sup{supp q(π) : π ∈ G is an involution and supp π ⊆ a} for every a ∈ A. Evidently θ is order-preserving. Now if a ∈ A, π ∈ G is an involution and supp π 6⊆ a, supp q(π) 6⊆ θa. P P There is a φ ∈ G, of order 4, ed by supp π \ a. Now φ2 is an involution ed by supp π, so supp q(φ2 ) ⊆ supp q(π). On the other hand, if π 0 ∈ G is an involution ed by a, then φ ∈ Vπ0 and φ2 ∈ Sπ0 , so q(φ2 ) ∈ Sq(π0 ) and supp q(φ2 ) ∩ supp q(π 0 ) = 0. As π 0 is arbitrary, supp q(φ2 ) ∩ θa = 0; so supp q(π) \ θa ⊇ supp q(φ2 ) 6= 0. Q Q (j) In the same way, we can define θ∗ : B → A by setting θ∗ b = sup{supp q −1 (π) : π ∈ H is an involution and supp π ⊆ b} for every b ∈ B. Now θ∗ θa = a for every a ∈ A. P P (α) If 0 6= u ⊆ a, there is an involution π ∈ G ed by u. Now q(π) is an involution in H ed by θa, so u ∩ θ∗ θa ⊇ u ∩ supp q −1 q(π) = supp π 6= 0. As u is arbitrary, a ⊆ θ∗ θa. (β) If π ∈ H is an involution ed by θa, then φ = q −1 (π) is an involution in G with supp q(φ) = supp π ⊆ θa, so supp φ ⊆ a, by (i) above; as π is arbitrary, θ∗ θa ⊆ a. Q Q Similarly, θθ∗ b = b for every b ∈ B. But this means that θ and θ∗ are the two halves of an orderisomorphism between A and B. By 312L, both are Boolean homomorphisms. (k) If π ∈ G is an involution, then θ(supp π) = supp q(π). P P By the definition of θ, supp q(π) ⊆ θ(supp π). On the other hand, supp q(π) = θθ∗ (supp q(π)) ⊇ θ(supp q −1 q(π)) = θ(supp π). Q Q Similarly, if π ∈ H is an involution, θ−1 (supp π) = θ∗ (supp π) = supp q −1 (π). (l) We are nearly home. Let us confirm that q(φ) = θφθ−1 for every φ ∈ G. P P?? Otherwise, ψ = q(φ)−1 θφθ−1 is not the identity automorphism on B, and there is a non-zero b ∈ B such that ψb ∩ b = 0, that is, θφθ−1 b ∩ q(φ)b = 0. Let π ∈ H be an involution ed by b. Then q −1 (π) is ed by θ−1 b, by (j), so φθ−1 b s φq −1 (π)φ−1 and θφθ−1 b s q(φq −1 (π)φ−1 ) = q(φ)πq(φ)−1 . On the other hand, q(φ)b also s q(φ)πq(φ)−1 , which is not the identity automorphism; so these two elements of B cannot be dist. X XQ Q (m) Finally, θ is unique by 383C. Remark The ideas of the proof here are taken from Eigen 82. 383E The rest of this section may be regarded as a series of corollaries of this theorem. But I think it will be apparent that they are very substantial results. Theorem Let A and B be atomless homogeneous Boolean algebras, and q : Aut A → Aut B an isomorphism. Then there is a unique Boolean isomorphism θ : A → B such that q(φ) = θφθ−1 for every φ ∈ Aut A. b be the Dedekind completion of A (314U). Then every φ ∈ Aut A has a unique extension to proof (a) Let A b→A b (314Tb). Because the extension is unique, we must have (φψ)b= φˆψˆ a Boolean homomorphism φˆ : A −1 are inverses of each other, and φ ˆ ∈ Aut A b for each φ ∈ Aut A; for all φ, ψ ∈ Aut A; consequently, φˆ and φd ˆ moreover, φ 7→ φ is a group homomorphism. Of course it is injective, so we have a subgroup G = {φˆ : φ ∈ b which is isomorphic to Aut A. Clearly Aut A} of Aut A b φu ∈ A for every u ∈ A}. G = {φ : φ ∈ Aut A,

496

Automorphism groups

383E

b is non-zero, then there is a non-zero u ⊆ a belonging to A. Because A is atomless and homogeneous, If a ∈ A there is an involution π ∈ Aut A ed by u (381Q); now π ˆ ∈ G is an involution ed by a. As a is arbitrary, G has many involutions. b for the Dedekind completion of B, we have a subgroup H = {ψˆ : ψ ∈ Aut B} of Similarly, writing B b Aut B isomorphic to Aut B, and with many involutions. Let qˆ : G → H be the corresponding isomorphism, d for every φ ∈ Aut A. ˆ = q(φ) so that qˆ(φ) ˆ θˆ−1 for every φ ∈ G. b→B b such that qˆ(φ) = θφ By 383D, there is a Boolean isomorphism θˆ : A ˆ ∈ B. P ˆ = 0, θ1 ˆ = 1. (b) If u ∈ A, then θu P It is enough to consider the case u ∈ / {0, 1}, since surely θ0 Take any w ∈ B which is neither 0 nor 1; then there is an involution in Aut B with w (381Q again); the corresponding member π of H is still an involution with w. Its image qˆ−1 (π) in G is b of course 0 6= a 6= 1. Take non-zero u1 , u3 ∈ A such that u1 ⊂ a an involution with a = θˆ−1 w ∈ A; and u3 ⊆ 1 \ a; set u2 = 1 \ (u1 ∪ u3 ). Because A is homogeneous, there are φ, ψ ∈ G such that φu1 = u, ψu1 = u1 , ψu2 = u3 ; set φ2 = φψ. Then we have u = φu1 ⊆ φ(supp qˆ−1 (π)) = supp(φˆ q −1 (π)φ−1 ) ⊆ φ(u1 ∪ u2 ) = u ∪ φu2 , u = φ2 u1 ⊆ φ2 (supp qˆ−1 (π)) = supp(φ2 qˆ−1 (π)φ−1 2 ) ⊆ u ∪ φ2 u2 = u ∪ φu3 , so φ(supp qˆ−1 (π)) ∩ φ2 (supp qˆ−1 (π)) = u, and ˆ = θ(φ(supp ˆ ˆ 2 (supp qˆ−1 (π))) θu qˆ−1 (π))) ∩ θ(φ ˆ ˆ = θ(supp φˆ q −1 (π)φ−1 ) ∩ θ(supp φ2 qˆ−1 (π)φ−1 ) 2

ˆ ˆ = θ(supp qˆ−1 (ˆ q (φ)π qˆ(φ)−1 )) ∩ θ(supp qˆ−1 (ˆ q (φ2 )π qˆ(φ2 )−1 )) = θˆθˆ−1 (supp(ˆ q (φ)π qˆ(φ)−1 )) ∩ θˆθˆ−1 (supp(ˆ q (φ2 )π qˆ(φ2 )−1 )) = supp(ˆ q (φ)π qˆ(φ)−1 ) ∩ supp(ˆ q (φ2 )π qˆ(φ2 )−1 ) = qˆ(φ)(supp π) ∩ qˆ(φ2 )(supp π) = qˆ(φ)w ∩ qˆ(φ2 )w ∈ B because both qˆ(φ) and qˆ(φ2 ) belong to H. Q Q ˆ A is an isomorphism between A and B. Similarly, θˆ−1 v ∈ A for every v ∈ B, and θ = θ¹ We now have ˆ B = (θˆφˆθˆ−1 )¹ B = θφθ−1 q(φ) = qˆ(φ)¹ for every φ ∈ Aut A. Finally, θ is unique by 383C, as before. 383F Corollary If A and B are atomless homogeneous Boolean algebras with isomorphic automorphism groups, they are isomorphic as Boolean algebras. Remark Of course a one-element Boolean algebra {0} and a two-element Boolean algebra {0, 1} have isomorphic automorphism groups without being isomorphic. 383G Corollary If A is a homogeneous Boolean algebra, then Aut A has no outer automorphisms. proof If A = {0, 1} this is trivial. Otherwise, A is atomless, so if q is any automorphism of Aut A, there is a Boolean isomorphism θ : A → A such that q(φ) = θφθ−1 for every φ ∈ Aut A, and q is an inner automorphism. 383H Definitions Complementary to the notion of ‘many involutions’ is the following concept. (a) A Boolean algebra A is rigid if the only automorphism of A is the identity automorphism. (b) A Boolean algebra A is nowhere rigid if no non-trivial principal ideal of A is rigid.

383N

Outer automorphisms

497

383I Lemma Let A be a Boolean algebra. Then the following are equiveridical: (i) A is nowhere rigid; (ii) for every a ∈ A \ {0} there is a φ ∈ Aut A, not the identity, ed by a; (iii) for every a ∈ A \ {0} there are distinct non-zero b, c ⊆ a such that the principal ideals Ab , Ac they generate are isomorphic; (iv) the automorphism group Aut A has many involutions. proof (a)(ii)⇒(i) If a ∈ A \ {0}, let φ ∈ Aut A be a non-trivial automophism ed by a; then φ¹ Aa is a non-trivial automorphism of the principal ideal Aa , so Aa is not rigid. (b)(i)⇒(iii) There is a non-trivial automorphism ψ of Aa ; now if b ∈ Aa is such that ψb = c 6= b, Ab is isomorphic to φ[Ab ] = Ac . (c)(iii)⇒(iv) Take any non-zero a ∈ A. By (iii), there are distinct b, c ⊆ a such that Ab , Ac are isomorphic. At least one of b \ c, c \ b is non-zero; suppose the former. Let ψ : Ab → Ac be an isomorphism, ←−−− and set d = b \ c, d0 = ψ(b \ c); then d0 ⊆ c, so d0 ∩ d = 0, and φ = (d ψ d0 ) is an involution ed by a. (d)(iv)⇒(ii) is trivial. 383J Theorem Let A and B be nowhere rigid Dedekind complete Boolean algebras and q : Aut A → Aut B an isomorphism. Then there is a Boolean isomorphism θ : A → B such that q(φ) = θφθ−1 for every φ ∈ Aut A. proof Put 383I(i)⇒(iv) and 383D together. 383K Corollary Let A be a nowhere rigid Dedekind complete Boolean algebra. Then Aut A has no outer automorphisms. 383L Examples I note the following examples of nowhere rigid algebras. (a) An atomless homogeneous Boolean algebra is nowhere rigid. (b) Any principal ideal of a nowhere rigid Boolean algebra is nowhere rigid. (c) A simple product of nowhere rigid Boolean algebras is nowhere rigid. (d) Any atomless semi-finite measure algebra is nowhere rigid. (e) A free product of nowhere rigid Boolean algebras is nowhere rigid. (f) The Dedekind completion of a nowhere rigid Boolean algebra is nowhere rigid. Indeed, the difficulty is to find an atomless Boolean algebra which is not nowhere rigid; for a variety of constructions of rigid algebras, see Bekkali & Bonnet 89. 383M Theorem Let (A, µ ¯) and (B, ν¯) be atomless localizable measure algebras, and Autµ¯ A, Autν¯ B the corresponding groups of measure-preserving automorphisms. Let q : Autµ¯ A → Autν¯ B be an isomorphism. Then there is a Boolean isomorphism θ : A → B such that q(φ) = θφθ−1 for every φ ∈ Autµ¯ A. proof The point is just that Autµ¯ A has many involutions. P P Let a ∈ A \ {0}. Then there is a nonzero b ⊆ a such that the principal ideal Ab is Maharam-type-homogeneous. Take c ⊆ b, d ⊆ b \ c such that ¯b) (331C). The principal ideals Ac , Ad are now isomorphic as measure algebras (331I); µ ¯c = µ ¯d = min(1, 12 µ ←−− let ψ : Ac → Ad be a measure-preserving isomorphism. Then (c ψ d) ∈ Autµ¯ A is an involution ed by a. Q Q Similarly, Autν¯ B has many involutions, and the result follows at once from 383D. 383N

To make proper use of the last theorem we need the following result.

Proposition Let (A, µ ¯) and (B, ν¯) be localizable measure algebras and θ : A → B a Boolean isomorphism. For each infinite cardinal κ let eκ be the Maharam-type-κ component of A (332G) and for each γ ∈ ]0, ∞[ let Aγ be the set of atoms of A of measure γ. Then the following are equiveridical: (i) for every φ ∈ Autµ¯ A, θφθ−1 ∈ Autν¯ B;

498

Automorphism groups

383N

(ii)(α) for every infinite cardinal κ there is an ακ > 0 such that ν¯(θa) = ακ µ ¯a for every a ⊆ eκ (β) for every γ ∈ ]0, ∞[ there is an αγ > 0 such that ν¯(θa) = αγ µ ¯a for every a ∈ Aγ . α) Let κ be an infinite cardinal. The point is that if a, a0 ⊆ eκ and µ proof (a)(i)⇒(ii)(α ¯a = µ ¯ a0 < ∞ 0 then ν¯(θa) = ν¯(θa ). P P The principal ideals Aa , Aa0 are isomorphic as measure algebras; moreover, by 332J, the principal ideals Aeκ \a , Aeκ \a0 are isomorphic. We therefore have a φ ∈ Autµ¯ A such that φa = a0 . Consequently ψθa = θa0 , where ψ = θφθ−1 ∈ Autν¯ B, and ν¯(θa) = ν¯(θa0 ). Q Q If eκ = 0 we can take ακ = 1. Otherwise fix on some c0 ⊆ eκ such that 0 < µ ¯c0 < ∞; take b ⊆ θc0 such that 0 < ν¯b < ∞, and set c = θ−1 b, ακ = ν¯b/¯ µc. Then we shall have ν¯(θa) = ν¯(θc) = ακ µ ¯a whenever a ⊆ eκ and µ ¯a = µ ¯c. But we can find for any n ≥ 1 a partition cn1 , . . . , cnn of c into elements of measure 1 ¯c; since ν¯(θcni ) = ν¯(θcnj ) for all i, j ≤ n, we must have ν¯(θcni ) = n1 ν¯(θc) = ακ µ ¯cni for all i. So if a ⊆ eκ nµ and µ ¯a = n1 µ ¯c, ν¯(θa) = ν¯(θcn1 ) = ακ µ ¯a. Now suppose that a ⊆ eκ and µ ¯a = nk µ ¯c for some k, n ≥ 1; then a can be partitioned into k sets of measure n1 µ ¯c, so in this case also ν¯(θa) = ακ µ ¯a. Finally, for any a ⊆ eκ , set D = {d : d ⊆ a, µ ¯d is a rational multiple of µ ¯c}, 0

and let D ⊆ D be a maximal upwards-directed set. Then sup D0 = a, so θ[D0 ] is an upwards-directed set with supremum θa, and ν¯(θa) = supd∈D0 ν¯(θd) = supd∈D0 ακ µ ¯d = ακ µ ¯a. β ) Let γ ∈ ]0, ∞[. If Aγ = ∅ take αγ = 1. Otherwise, fix on any c ∈ Aγ and set αγ = ν¯(θc)/γ. If (β a ∈ Aγ then there is a φ ∈ Autµ¯ A exchanging the atoms a, c, so that θφθ−1 ∈ Autν¯ B exchanges the atoms θa, θc, and ν¯(θa) = ν¯(θc) = αγ µ ¯a. (b)(ii)⇒(i) Now suppose that the conditions (α) and (β) are satisfied, that φ ∈ Autµ¯ A and that a ∈ A. For each infinite cardinal κ, we have φeκ = eκ , so ν¯(θφ(eκ ∩ a)) = ακ µ ¯(φ(eκ ∩ a)) = ακ µ ¯(eκ ∩ a) = ν¯(θ(eκ ∩ a)). Similarly, if we write aγ = sup Aγ , then for each γ ∈ ]0, ∞[ we have φ[Aγ ] = Aγ , φaγ = aγ , and for c ⊆ aγ we have µ ¯c = γ#({e : e ∈ Aγ , e ⊆ c}); so ν¯(θφ(aγ ∩ a)) = αγ γ#({e : e ∈ Aγ , e ⊆ φa}) = αγ γ#({e : e ∈ Aγ , e ⊆ a}) X = ν¯(θe) = ν¯(θ(aγ ∩ a)). e∈Aγ ,e ⊆ a

Putting these together, X

ν¯(θφa) =

κ is an infinite cardinal

X

=


X

ν¯(θφ(eκ ∩ a)) +

ν¯(θφ(aγ ∩ a))

γ∈]0,∞[

ν¯(θ(eκ ∩ a)) +

X

γ∈]0,∞[

= ν¯(θa). But this means that ν¯(θφθ−1 b) = ν¯(θθ−1 b) = ν¯b for every b ∈ B, and θφθ−1 is measure-preserving, as required by (i).

ν¯(θ(aγ ∩ a))

383P

Outer automorphisms

499

383O Corollary If (A, µ ¯) is an atomless totally finite measure algebra, Autµ¯ A has no outer automorphisms. proof Let q : Autµ¯ A → Autµ¯ A be any automorphism. By 383M, there is a corresponding θ ∈ Aut A such that q(φ) = θφθ−1 for every φ ∈ Autµ¯ A. By 383N, there is for each infinite cardinal κ an ακ > 0 such that µ ¯(θa) = ακ µ ¯a whenever a ⊆ eκ , the Maharam-type-κ component of A. But since θeκ = eκ and µ ¯ eκ < ∞ for every κ, we must have ακ = 1 whenever eκ 6= 0; as A is atomless, X µ ¯(θa) = µ ¯(θ(a ∩ eκ )) κ is an infinite cardinal

=

X

ακ µ ¯(a ∩ eκ )


=

X

µ ¯(a ∩ eκ ) = µ ¯a


for every a ∈ A. Thus θ ∈ Autµ¯ A and q is an inner automorphism. 383P The results above are satisfying and complete in their own , but leave open a number of obvious questions concerning whether some of the hypotheses can be relaxed. Atoms can produce a variety of complications (see 383Ya-383Yd below). To show that we really do need to assume that our algebras are Dedekind complete or localizable, I offer the following. Example (a) There are an atomless localizable measure algebra (A, µ ¯) and an atomless semi-finite measure algebra (B, ν¯) such that Aut A ∼ = Aut B, Autµ¯ A ∼ = Autν¯ B but A and B are not isomorphic. proof Let (A0 , µ ¯0 ) be an atomless homogeneous probability algebra; for instance, the measure algebra 1 of Lebesgue measure on the unit interval. Let A be the simple product Boolean algebra Aω ¯ the 0 , and µ corresponding measure (322K); then (A, µ ¯) is an atomless localizable measure algebra. In A let I be the set {a : a ∈ A and the principal ideal Aa is ccc}; then I is an ideal of A, the σ-ideal generated by the elements of finite measure (cf. 322G). Set B = {a : a ∈ A, either a ∈ I or 1 \ a ∈ I}. Then B is a σ-subalgebra of A, so if we set ν¯ = µ ¯¹ B then (B, ν¯) is a measure algebra in its own right. The definition of I makes it plain that it is invariant under all Boolean automorphisms of A; so B is also invariant under all automorphisms, and we have a homomorphism φ 7→ q(φ) = φ¹ B : Aut A → Aut B. On the other hand, because B is order-dense in A, and A is Dedekind complete, every automorphism of B can be extended to an automorphism of A (see part (a) of the proof of 383E). So q is actually an isomorphism between Aut A and Aut B. Moreover, still because B is order-dense, q(φ) is measure-preserving iff φ is measure-preserving, so Autµ¯ A is isomorphic to Autν¯ B. But of course there is no Boolean isomorphism, let alone a measure algebra isomorphism, between A and B, because A is Dedekind complete while B is not. Remark Thus the hypothesis ‘Dedekind complete’ in 383D and 383J (and ‘localizable’ in 383M), and the hypothesis ‘homogeneous’ in 383E-383F, are essential. ¯ such that Aut C has an outer automorphism. (b) There is an atomless semi-finite measure algebra (C, λ) proof In fact we can take C to be the simple product of A and B above. I claim that the isomorphism between Aut A and Aut B gives rise to an outer automorphism of Aut C; this seems very natural, but I think there is a fair bit to check, so I take the argument in easy stages. (i) We may identify the Dedekind completion of C = A × B with A × A. For φ ∈ Aut C, we have a ˆ P corresponding φˆ ∈ Aut(A × A). Now B × A is invariant under φ. P Consider first φ(0, 1) = (a1 , b1 ) ∈ C. ∼ The corresponding principal ideal C(a1 ,b1 ) = Aa1 × Bb1 of C must be isomorphic to the principal ideal C(0,1) ∼ = Aa × Bb , = B; so that if (a, b) ∈ C and (a, b) ⊆ (a1 , b1 ), then just one of the principal ideals C(a,b) ∼ C(a1 \a,b1 \b) ∼ = Aa1 \a × Bb1 \b is ccc. But this can only happen if Aa1 is ccc and Bb1 is not; that is, if a1 and ˆ a) ⊆ (a1 , b1 ) belongs to B × A for every a ∈ A. We also find that 1 \ b1 belong to I. Consequently φ(0, φ(1, 0) = (1, 1) \ φ(0, 1) = (1 \ a1 , 1 \ b1 ) ∈ B × A.

500

Automorphism groups

383P

Now if b ∈ I, then Cφ(b,0) ∼ = C(b,0) ∼ = Ab is ccc and φ(b, 0) ∈ I × I ⊆ B × A; while φ(1 \ b, 0) = (1 \ a1 , 1 \ b1 ) \ φ(b, 0) ∈ B × A. This shows that φ(b, 0) ∈ B × A for every b ∈ B. So ˆ a) = φ(b, ˆ 0) ∪ φ(0, ˆ a) ∈ B × A φ(b, for every b ∈ B, a ∈ A. Q Q (ii) Let θ : A × A → A × A be the involution defined by setting θ(a, b) = (b, a) for all a, b ∈ A. Take ˆ −1 ∈ Aut(A × A). If c = (a, b) ∈ C, then θ−1 c = (b, a) ∈ B × A, so φ ∈ Aut C and consider ψ = θφθ −1 ˆ φθ c ∈ B × A, by (i), and ψc ∈ A × B = C. This shows that ψ¹ C is a homomorphism from C to itself. Of course ψ −1 = θφˆ−1 θ−1 has the same property. So we have a map q : Aut C → Aut C given by setting ˆ −1 ¹ C q(φ) = θφθ for φ ∈ Aut C. Evidently q is an automorphism. (iii) ?? Suppose, if possible, that q were an inner automorphism. Let χ ∈ Aut C be such that q(φ) = χφχ−1 for every φ ∈ Aut C. Then d = θφθ ˆ −1 χ ˆφˆχ ˆ−1 = q(φ) for every φ ∈ Aut C. Since G = {φˆ : φ ∈ Aut C} is a subgroup of Aut(A × A) with many involutions, the ‘uniqueness’ assertion of 383D tells us that χ ˆ = θ. But θ[C] = B × A 6= C = χ[C] = χ[C], ˆ so this cannot be. X X Thus q is the required outer automorphism of Aut C. Remark Thus the hypothesis ‘homogeneous’ in 383E, and the hypothesis ‘Dedekind complete’ in 383J, are necessary. 383Q Example Let µ be Lebesgue measure on R, and (A, µ ¯) its measure algebra. Then Autµ¯ A has an outer automorphism. P P Set f (x) = 2x for x ∈ R. Then E 7→ f −1 [E] = 21 E is a Boolean automorphism of the domain Σ of µ, and µ( 12 E) = 12 µE for every E ∈ Σ (263A, or otherwise). So we have a corresponding θ ∈ Aut A defined by setting θE • = ( 12 E)• for every E ∈ Σ, and µ ¯(θa) = 12 µ ¯a for every a ∈ A. By 383N, we have an automorphism q of Autµ¯ A defined by setting q(φ) = θφθ−1 for every measure-preserving automorphism φ. But q is now an outer automorphism of Autµ¯ A, because (by 383D) the only possible automorphism of A corresponding to q is θ, and θ is not measure-preserving. Q Q Thus the hypothesis ‘totally finite’ in 383O cannot be omitted. 383X Basic exercises (a) Let A be a Boolean algebra. Show that the following are equiveridical: (i) A is nowhere rigid; (ii) for every a ∈ A \ {0} and n ∈ N there are dist non-zero b0 , . . . , bn ⊆ a such that the principal ideals Abi they generate are all isomorphic; (iii) for every a ∈ A \ {0} and n ≥ 1 there is a φ ∈ Aut A, of order n, ed by a. (b) Let A be a nowhere rigid Boolean algebra. Show that its Dedekind completion is nowhere rigid. (c) Let A be an atomless homogeneous Boolean algebra and B a nowhere rigid Boolean algebra, and suppose that Aut A is isomorphic to Aut B. Show that there is an invariant order-dense subalgebra of B which is isomorphic to A.

383 Notes

Outer automorphisms

501

(d) Let A and B be nowhere rigid Boolean algebras. Show that if Aut A and Aut B are isomorphic, then b and B b are isomorphic. the Dedekind completions A (e) Find two non-isomorphic atomless totally finite measure algebras (A, µ ¯), (B, ν¯) such that Autµ¯ A and Autν¯ B are isomorphic. (This is easy.) (f ) Let (A, µ ¯) and (B, ν¯) be semi-finite measure algebras and θ : A → B a Boolean isomorphism. Show that the following are equiveridical: (i) for every φ ∈ Autµ¯ A, θφθ−1 ∈ Autν¯ B; (ii)(α) for every infinite cardinal κ there is an ακ > 0 such that ν¯(θa) = ακ µ ¯a whenever a ∈ A and the principal ideal Aa is Maharam-type-homogeneous with Maharam type κ; (β) for every γ ∈ ]0, ∞[ there is an αγ > 0 such that ν¯(θa) = αγ µ ¯a whenever a ∈ A is an atom of measure γ. (g) Let q : Aut C → Aut C be the automorphism of 383Pb. Show that q(φ) is measure-preserving whenever φ is measure-preserving, so that q¹ Autλ¯ C is an outer automorphism of Autλ¯ C. 383Y Further exercises (a) Let (A, µ ¯) and (B, ν¯) be localizable measure algebras such that Aut A ∼ = ∼ Aut B. Show that either A = B or one of A, B has just one atom and the other is atomless. ∼ Autν¯ B. Show that either (b) Let (A, µ ¯), (B, ν¯) be localizable measure algebras such that Autµ¯ A = (A, µ ¯) ∼ = (B, ν¯) or there is some γ ∈ ]0, ∞[ such that one of A, B has just one atom of measure γ and the other has none or there are γ, γ 0 ∈ ]0, ∞[ such that the number of atoms of A of measure γ is equal to the number of atoms of B of measure γ 0 , but not to the number of atoms of A of measure γ 0 . (c) Let (A, µ ¯) be a localizable measure algebra. Show that there is an outer automorphism of Aut A iff A has exactly six atoms. (d) Let (A, µ ¯) be a localizable measure algebra. For each infinite cardinal κ let eκ be the Maharam-type-κ component of A and for each γ ∈ ]0, ∞[ let Aγ be the set of atoms of A of measure γ. Show that there is an outer automorphism of Autµ¯ A iff either there is an infinite cardinal κ such that µ ¯ eκ = ∞ or there are distinct γ, δ ∈ ]0, ∞[ such that #(Aγ ) = #(Aδ ) ≥ 2 or there is a γ ∈ ]0, ∞[ such that #(Aγ ) = 6 or there are γ, δ ∈ ]0, ∞[ such that #(Aγ ) = 2 < #(Aδ ) < ω. 383 Notes and comments Let me recapitulate the results above. If A and B are Boolean algebras, any isomorphism between Aut A and Aut B corresponds to an isomorphism between A and B if either A and B are atomless and homogeneous (383E) or they are nowhere rigid and Dedekind complete (383J). If (A, µ ¯) and (B, ν¯) are atomless localizable measure algebras, then any automorphism between Autµ¯ A and Autν¯ B corresponds to an isomorphism between A and B (383M) which if µ ¯ = ν¯ is totally finite will be measure-preserving (383O). These results may appear a little less surprising if I remark that the elementary Boolean algebras PX give rise to some of the same phenomena. The automorphism group of PX can be identified with the group SX of all permutations of X, and this has no outer automorphisms unless X has just six elements. Some of the ideas of the fundamental theorem 383D can be traced through in the purely atomic case also, though of course there are significant changes to be made, and some serious complications arise, of which the most striking surround the remarkable fact that S6 does have an outer automorphism (Burnside 11, §162; Rotman 84, Theorem 7.8). I have not attempted to incorporate these in the main results. For localizable measure algebras, where the only rigid parts are atoms, the complications are superable, and I think I have listed them all (383Ya-383Yd).

502

Automorphism groups

§384 intro.

384 Entropy Perhaps the most glaring problem associated with the theory of measure-preserving homomorphisms and automorphisms is the fact that we have no generally effective method of determining when two homomorphisms are the same, in the sense that two structures (A, µ ¯, π) and (B, ν¯, φ) are isomorphic, where (A, µ ¯) and (B, ν¯) are measure algebras and π : A → A, φ : B → B are Boolean homomorphisms. Of course the first part of the problem is to decide whether (A, µ ¯) and (B, ν¯) are isomorphic; but this is solved (at least for localizable algebras) by Maharam’s theorem (see 332J). The difficulty lies in the homomorphisms. Even when we know that (A, µ ¯) and (B, ν¯) are both isomorphic to the Lebesgue measure algebra, the extraordinary variety of constructions of homomorphisms – corresponding in part to the variety of measure spaces with such measure algebras, each with its own natural inverse-measure-preserving functions – means that the question of which are isomorphic to each other is continually being raised. In this section I give the most elementary ideas associated with the concept of ‘entropy’, up to the Kolmogorov-Sinaˇı theorem (384P). This is an invariant which can be attached to any measure-preserving homomorphism on a probability algebra, and therefore provides a useful method for distinguishing non-isomorphic homomorphisms. The main work of the section deals with homomorphisms on measure algebras, but as many of the most important ones arise from inverse-measure-preserving functions on measure spaces I comment on the extra problems arising in the isomorphism problem for such functions (384T-384V). I should remark that some of the lemmas will be repeated in stronger forms in the next section. 384A Notation Throughout this section, I will use the letter q to denote the function from [0, ∞[ to R defined by saying that q(t) = −t ln t = t ln 1t if t > 0, q(0) = 0.

0

0.5

1

1.5

-0.5

The function q We shall need the following straightforward facts concerning q. (a) q is continuous on [0, ∞[ and differentiable on ]0, ∞[; q 0 (t) = −1 − ln t and q 00 (t) = − 1t for t > 0. Because q 00 is negative, q is concave, that is, −q is convex. q has a unique maximum at ( 1e , 1e ). (b) If s ≥ 0, t > 0 then q 0 (s + t) ≤ q 0 (t); consequently

Rt

q(s + t) = q(s) + 0 q 0 (s + τ )dτ ≤ q(s) + q(t) Pn Pn forPs, t ≥ 0. P It follows that q( i=0 si ) ≤ i=0 q(si ) for all s0 , . . . , sn ≥ 0 and (because q is continuous) ∞ ∞ q( i=0 si ) ≤ i=0 q(si ) for every non-negative summable series hsi ii∈N . (c) If s, t ≥ 0 then q(st) = sq(t) + tq(s); more generally, if n ≥ 1 and si ≥ 0 for i ≤ n then Qn Pn Q q( i=0 si ) = j=0 q(sj ) i6=j si . d (d) The function t 7→ q(t) + q(1 − t) has a unique maximum at ( 12 , ln 2). ( dt (q(t) + q(1 − t)) = ln 1−t t .) It 1 follows that for every ² > 0 there is a δ > 0 such that |t − 2 | ≤ ² whenever q(t) + q(1 − t) ≥ ln 2 − δ.

(e) If 0 ≤ t ≤ 21 , then q(1 − t) ≤ q(t). P P Set f (t) = q(t) − q(1 − t). Then 1 t

f 00 (t) = − +

1 1−t

=

2t−1 t(1−t)

≤0

for 0 < t ≤ 21 , while f (0) = f ( 21 ) = 0, so f (t) ≥ 0 for 0 ≤ t ≤ 12 . Q Q

384G

Entropy

503

(f ) (i) If A is a Dedekind σ-complete Boolean algebra, I will write q¯ for the function from L0 (A)+ to L (A) defined from q (364I). Note that because 0 ≤ q(t) ≤ 1 for t ∈ [0, 1], 0 ≤ q¯(u) ≤ χ1 if 0 ≤ u ≤ χ1. (ii) By (b), q¯(u + v) ≤ q¯(u) + q¯(v) for all u, v ≥ 0 in L0 (A). (Represent A as the measure algebra of a measure space, so that q¯(f • ) = (qf )• , as in 364Jb.) (iii) Similarly, if u, v ∈ L0 (A)+ , then q¯(u × v) = u × q¯(v) + v × q¯(u). 0

1 384B Lemma Let (A, µ ¯) be a probability algebra, B a closed subalgebra of A, and ¯) → R R P : L (A, µ L (A, µ ¯) the corresponding conditional expectation operator (365R). Then q¯(u) ≤ q( u) and P (¯ q (u)) ≤ q¯(P u) for every u ∈ L∞ (A)+ . 1

proof Apply the remarks in 365Rb to −q. (¯ q (u) ∈ L∞ ⊆ L1 for every u ∈ (L∞ )+ because q is bounded on every bounded interval in [0, ∞[.) 384 Definition Let (A, µ ¯) be a probability algebra. If A is a partition of unity in A, its entropy is H(A) = a∈A q(¯ µa), where q is the function defined in 384A. Remarks (a) In the definition of ‘partition of unity’ (311Gc) I allowed 0 to belong to the family. In the present context this is a mild irritant, and when convenient I shall remove 0 from the partitions of unity considered here (as in 384F below). But because q(0) = 0, it makes no difference; H(A) = H(A \ {0}) whenever A is a partition of unity. So if you wish you can read ‘partition of unity’ in this section to mean ‘partition of unity not containing 0’, if you are willing to make an occasional amendment in a formula. In important cases, in fact, A is of the form {ai : i ∈ I} or {ai : i ∈ I} \ {0}, where hai ii∈I is an indexed partition of unity, with Pai ∩ aj = 0 for i 6= j, but no restriction in the number of i with ai = 0; in this case, µai ). we still have H(A) = i∈I q(¯ (b) Many authors prefer to use log2 in place of ln. This makes sense in of one of the intuitive approaches to entropy as the ‘information’ associated with a partition. See Petersen 83, §5.1. 384D Definition Let (A, µ ¯) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A. Let P : L1 (A, µ ¯) → L1 (A, µ ¯) be the conditional expectation operator associated with B. Then the conditional entropy of A on B is R P H(A|B) = a∈A q¯(P χa), where q¯ is defined as in 384A. 384E Elementary remarks (a) In the formula R P a∈A

q¯(P χa),

we have 0 ≤ P (χa) ≤ χ1 for every a, so q¯(P χa) ≥ 0 and every term in the sum is non-negative; accordingly H(A|B) is well-defined in [0, ∞]. R (b) H(A) = H(A|{0, 1}), since if B = {0, 1} then P (χa) = µ ¯aχ1, so that q¯(P χa) = q(¯ µa). If A ⊆ B, H(A|B) = 0, since P (χa) = χa, q¯(P χa) = 0 for every a. 384F Definition If A is a Boolean algebra and A, B ⊆ A are partitions of unity, I write A ∨ B for the partition of unity {a ∩ b : a ∈ A, b ∈ B} \ {0}. (See 384Xq.) 384G Lemma Let (A, µ ¯) be a probability algebra and B a closed subalgebra. Let A ⊆ A be a partition of unity. (a) If B is another partition of unity in A, then H(A|B) ≤ H(A ∨ B|B) ≤ H(A|B) + H(B|B). (b) If B is purely atomic and D is the set of its atoms, then H(A ∨ D) = H(D) + H(A|B). (c) If C ⊆ B is another closed subalgebra of A, then H(A|C) ≥ H(A|B). In particular, H(A) ≥ H(A|B). (d) Suppose that hBn in∈N is a non-decreasing sequence of closed subalgebras of A such that B = S n∈N Bn . If H(A) < ∞ then

504

Automorphism groups

384G

H(A|B) = limn→∞ H(A|Bn ). In particular, if A ⊆ B then limn→∞ H(A|Bn ) = 0. proof Write P for the conditional expectation operator on L1 (A, µ ¯) associated with B. (a)(i) If B is infinite, enumerate it as hbj ij∈N ; if it is finite, enumerate it as hbj ij≤n and set bj = 0 for j > n. For any a ∈ A, P∞ P∞ χa = j=0 χ(a ∩ bj ), P (χa) = j=0 P χ(a ∩ bj ), n X q¯(P χa) = lim q¯( P χ(a ∩ bj )) n→∞

≤ lim

n→∞

j=0 n X

∞ X

j=0

j=0

q¯(P χ(a ∩ bj )) =

q¯(P χ(a ∩ bj ))

where all the infinite sums are to be regarded as order*-limits of the corresponding finite sums (see §367), and the middle inequality is a consequence of 384Af(ii). Accordingly Z X q¯(P χ(a ∩ b)) H(A ∨ B|B) = =

a∈A,b∈B,a∩b6=0 ∞ Z XX

XZ

a∈A j=0

a∈A

q¯(P χ(a ∩ bj )) ≥

q¯(P χai ) = H(A|B).

(ii) Suppose for the moment that A and B are both finite. For a ∈ A set ua = P (χa). If a, b ∈ A we 0 have 0 ≤ ua∩b ≤ ub inPL0 (B), so we may choose vP ab ∈ L (B) such that 0 ≤ vab ≤ χ1 P and ua∩b = vab × ub . = ub (because For any b ∈ B, a∈A vab = ub . Since a∈A χ(a ∩ b) = χb), so ub × a∈A ua∩b P [[|¯ q (ub )| > 0]] ⊆ [[ub > 0]], q¯(ub ) × a∈A vab = q¯(ub ). For any a ∈ A, q¯(ua ) = q¯(

X

ua∩b ) = q¯(

b∈B

X

ub × vab ) = q¯(P (

b∈B

X

χb × vab ))

b∈B

(because vab ∈ L0 (B) for every b, so P (χb × vab ) = P (χb) × vab ) X ≥ P (¯ q( χb × vab )) b∈B

(384B) = P(

X

χb × q¯(vab ))

b∈B

(because B is dist) =

X

ub × q¯(vab )

b∈B

(because q¯(vab ) ∈ L0 (B) for every b). Putting these together, H(A ∨ B|B) =

Z

X

q¯(ua∩b ) =

a∈A,b∈B

=

X XZ

a∈A

q¯(ub × vab )

a∈A,b∈B

Z

ub × q¯(vab ) +

a∈A,b∈B

≤

Z

X

q¯(ua ) +

XZ b∈B

X

Z vab × q¯(ub )

a∈A,b∈B

q¯(ub ) = H(A|B) + H(B|B).

384G

Entropy

505

(iii) For general partitions of unity A and B, take any finite set C ⊆ A ∨ B. Then C ⊆ {a ∩ b : a ∈ A0 , b ∈ B0 } where A0 ⊆ A and B0 ⊆ B are finite. Set A0 = A0 ∪ {1 \ sup A0 },

B 0 = B0 ∪ {1 \ sup B0 },

so that A0 and B 0 are finite partitions of unity and C ⊆ A0 ∨ B 0 . Now XZ

Z

X

q¯(P χc) ≤

q¯(P χc) = H(A0 ∨ B 0 |B) ≤ H(A0 |B) + H(B 0 |B)

c∈A0 ∨B 0

c∈C

(by (ii)) ≤ H(A0 ∨ A|B) + H(B 0 ∨ B|B) (by (i)) = H(A|B) + H(B|B). As C is arbitrary, H(A ∨ B|B) =

R

P c∈A∨B

q¯(P χc) ≤ H(A|B) + H(B|B).

P (b) It follows from 384Ab that d∈D q(¯ µ(a ∩ d)) ≥ q(¯ µa) for any a ∈ A. Now, because B is purely atomic and D is its set of atoms, P P µ ¯(a∩d) µ ¯(a∩d) P (χa) = d∈D χd, q¯(P (χa)) = d∈D q( )χd µ ¯d

µ ¯d

for every a ∈ A, H(A|B) =

P a∈A,d∈D

q(

µ ¯(a∩d) )¯ µd. µ ¯d

Putting these together, X

H(A ∨ D) =

q(¯ µ(a ∩ d)) =

a∈A,d∈D

= H(A|B) +

X a∈A,d∈D

X

q(

µ ¯(a∩d) µ ¯(a∩d) )¯ µd + q(¯ µd) µ ¯d µ ¯d

q(¯ µd) = H(A|B) + H(D).

d∈D

(c) Write PC for the conditional expectation operator corresponding to C. If a ∈ A, q¯(PC χa) = q¯(PC P χa) ≥ PC q¯(P χa) by 384B. So H(A|C) =

R

P a∈A

R R P P q¯(PC χa) ≥ a∈A PC q¯(P χa) = a∈A q¯(P χa) = H(A|B).

Taking C = {0, 1}, we get H(A) ≥ H(A|B). (d) Let Pn be the conditional expectation operator corresponding to Bn , for each n. Fix a ∈ A. Then P (χa) is the order*-limit of hPn (χa)in∈N , by Lévy’s martingale theorem (367Kb). Consequently (because q is continuous) h¯ q (Pn χa)in∈N is order*-convergent to q¯(P χa) for every a ∈ A (367I). Also, because 0 ≤ Pn χa ≤ χ1Rfor every n, R0 ≤ q¯(Pn χa) ≤ 1e χ1 for every n. By the Dominated Convergence Theorem (367J), limn→∞ q¯(Pn χa) = q¯(P χa). By 384B, we also have 0≤

R

R

for every a ∈ A and n ∈ N; since also we have |

R

q¯(Pn χa) −

R

R

q¯(Pn χa) ≤ q( Pn (χa)) = q( χa) = q(¯ µa) 0≤

R

q¯(P χa) = q(¯ µa),

q¯(P χa)| ≤ q(¯ µa) for every a ∈ A, n ∈ N.

506

Automorphism groups

384G

P Now we are supposing that H(A) is finite. Given ² > 0, we can find a finite set I ⊆ A such that µa) ≤ ², and an n0 ∈ N such that a∈A\I q(¯ R R P q¯(Pn χa) − q¯(P χa)| ≤ ² a∈I | for every n ≥ n0 ; in which case P

a∈A\I |

R

q¯(Pn χa) −

R

q¯(P χa)| ≤

P a∈A\I

q(¯ µa) ≤ ²

and |H(A|Bn ) − H(A|B)| ≤ 2² for every n ≥ n0 . As ² is arbitrary, H(A|B) = limn→∞ H(A|Bn ). 384H Corollary Let (A, µ ¯) be a probability algebra and A, B two partitions of unity in A. Then H(A) ≤ H(A ∨ B) ≤ H(A) + H(B). proof Take B = {0, 1} in 384Ga. 384I Lemma Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. If A ⊆ A is a partition of unity, then H(π[A]) = H(A). P P proof µπa) = a∈A q(¯ µa). a∈A q(¯ 384J Lemma Let (A, µ ¯) be a measure algebra. Let A be the set of its atoms. Then the following are equiveridical: (i) either A is not purely atomic or A is purely atomic and H(A) = ∞; (ii) there is a partition of unity B ⊆ A such that H(B) = ∞; (iii) for every γ ∈ R there is a finite partition of unity C ⊆ A such that H(C) ≥ γ. proof (i)⇒(ii) We need examine only the case in which A is not purely atomic. Let a ∈ A be a non-zero element such that the principal ideal Aa is atomless. By 331C we can choose inductively a dist sequence han in∈N such that an ⊆ a and µ ¯an = 2−n−1 µ ¯a. Now, for each n ∈ N, choose a dist set Bn such that n

#(Bn ) = 22 ,

n

b ⊆ an and µ ¯b = 2−2 µ ¯an for each b ∈ Bn .

Set B=

S n∈N

Bn ∪ {1 \ a}.

Then B is a partition of unity in A and H(B) ≥ =

∞ X X n=0 b∈Bn ∞ X

q(¯ µBn ) =

∞ X n=0

n

22 q

¡

µ ¯a ¢ 2n+1+2n

n ∞ ¡ 2n+1+2 ¢ X µ ¯a µ ¯a n ln ≥ 2 ln 2 = ∞. n+1 n+1 2 µ ¯a 2 n=0 n=0

(ii)⇒(iii) Enumerate B as hbi ii∈N . For each n ∈ N, Cn = {bi : i ≤ n}∪{1 \ supi≤n bi } is a finite partition of unity, and Pn limn→∞ H(Cn ) ≥ limn→∞ i=0 q(¯ µbi ) = H(B) = ∞. (iii)⇒(i) We need only consider the case in which A is purely atomic and A is its set of atoms. In this case, A ∨ C = A for every partition of unity C ⊆ A, so H(C) ≤ H(A) for every C (384H), and H(A) must be infinite. 384K Definition Let A be a Boolean algebra. If π : A → A is an order-continuous Boolean homomorphism, A ⊆ A is a partition of unity and n ≥ 1, write Dn (A, π) for the partition of unity generated by {π i a : a ∈ A, 0 ≤ i < n}, that is, {inf i
384N

Entropy

507

384L Lemma Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. Let A ⊆ A be a partition of unity. Then limn→∞ n1 H(Dn (A, π)) = inf n≥1 n1 H(Dn (A, π)) is defined in [0, ∞]. proof (a) Set α0 = 0, αn = H(Dn (A, π)) for n ≥ 1. Then αm+n ≤ αm + αn for all m, n ≥ 0. P P If m, n ≥ 1, Dm+n (A, π) = Dm (A, π) ∨ π m [Dn (A, π)]. So 384Ga tells us that H(Dm+n (A, π)) ≤ H(Dm (A, π)) + H(π m [Dn (A, π)]) = H(Dm (A, π)) + H(Dn (A, π)) because π is measure-preserving. Q Q (b) If α1 = ∞ then of course H(Dn (A, π)) ≥ H(A) = ∞ for every n, by 384H, so inf n≥1 n1 H(Dn (A, π)) = ∞ = limn→∞ n1 H(Dn (A, π)). Otherwise, αn ≤ nα1 is finite for every n. Set α = inf n≥1 n1 αn . If ² > 0 there 1 is an m ≥ 1 such that m αm ≤ α + ². Set M = maxj<m αj . Now, for any n ≥ m, there are k ≥ 1, j < m such that n = km + j, so that αn ≤ kαm + αj ,

1 αn n

k n

≤ αm +

M n

≤

1 αm m

+

M . n

Accordingly lim supn→∞ n1 αn ≤ α + ². As ² is arbitrary, 1 n

1 n

α ≤ lim inf n→∞ αn ≤ lim supn→∞ αn ≤ α and limn→∞ n1 αn = α is defined in [0, ∞]. Remark See also 384Yb and 385Nc below. 384M Definition Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. For any partition of unity A ⊆ A, set h(π, A) = inf n≥1 n1 H(Dn (A, π)) = limn→∞ n1 H(Dn (A, π)) (384L). Now the entropy of π is h(π) = sup{h(π, A) : A ⊆ A is a finite partition of unity}. Remarks (a) We always have h(π, A) ≤ H(D1 (A, π)) =H(A). (b) Observe that if π is the identity automorphism then Dn (A, π) = A \ {0} for every A and n, so that h(π) = 0. 384N Lemma Let (A, µ ¯) be a probability algebra and A, B two partitions of unity in A. Let π : A → A be a measure-preserving Boolean homomorphism. Then h(π, A) ≤ h(π, B) + H(A|B), where B is the closed subalgebra of A generated by B. proof We may suppose that 0 ∈ / B, since removing 0 from B changes neither Dn (B, π) nor B. For each n ∈ N, set An = π n [A], Bn = π n [B]. Let Bn = π n [B] be the closed subalgebra of A generated by Bn , and B∗n the closed subalgebra of A generated by Dn (B, π). Then H(An |Bn ) = H(A|B) for each n. P P The point is that, because B is purely atomic and B is its set of atoms, P µ ¯(a∩b) )¯ µb H(A|B) = a∈A,b∈B q( µ ¯b

as in the proof of 384Gb. Similarly, H(An |Bn ) = Accordingly, for any n ≥ 1,

P a∈A,b∈B

q(

µ ¯(π n a∩π n b) )¯ µ(π n b) µ ¯(π n b)

= H(A|B). Q Q

508

Automorphism groups

H(Dn (A, π)|B∗n ) ≤

n−1 X

384N

H(Ai |B∗n )

i=0

(by 384Ga) ≤

n−1 X

H(Ai |Bi )

i=0

(by 384Gc) = nH(A|B). Now

h(π, A) = lim

1

n→∞ n

1 n n→∞

H(Dn (A, π)) ≤ lim sup H(Dn (A, π) ∨ Dn (B, π))

(384Ga) 1 n

1 n

≤ lim sup H(Dn (B, π)) + H(Dn (A, π)|B∗n ) n→∞

(384Gb) ≤ h(π, B) + H(A|B). Remark Compare 385Nd below. 384O Lemma Let (A, µ ¯) be a probability algebra, π : A → A a measure-preserving Boolean homomorphism, and A ⊆ A a partition of unity such that H(A) < ∞. Then h(π, A) ≤ h(π). proof If A is finite, this is immediate from the definition of h(π); so suppose that A is infinite. Enumerate A as hai ii∈N . For each n ∈ N let Bn be the subalgebra of A generated by a0 , . . . , an ; then Bn has atoms Bn = {a0 , . . . , an , bn } where bn = supi>n ai . Now P∞ α H(A|Bn ) = i=n+1 q( i )βn βn

where αi = µ ¯ai , βn = µ ¯bn , because if Pn is the conditional expectation associated with Bn then Pn (χai ) = χai if i ≤ n, Pn (χai ) = βαni χbn if i > n. But ∞ X i=n+1

βn q(

αi ) βn

= =

because

P∞ i=0

∞ X i=n+1 ∞ X

q(αi ) −

αi q(βn ) βn

q(αi ) − q(βn ) → 0 as n → ∞

i=n+1

q(αi ) < ∞ and limn→∞ βn = 0. So by 384N and 384Gd we get h(π, A) ≤ h(π, Bn ) + H(A|Bn ) ≤ h(π) + H(A|Bn ) → h(π)

as n → ∞, and h(π, A) ≤ h(π). Remark Compare 385Nb below. 384P Theorem (Kolmogorov 58, Sinaˇı 59) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. (i) SupposeSthat A ⊆ A is a partition of unity such that H(A) < ∞ and the closed subalgebra of A generated by n∈N π n [A] is A itself. Then h(π) = h(π, A). (ii) Suppose that π is an automorphism, S and that A ⊆ A is a partition of unity such that H(A) < ∞ and the closed subalgebra of A generated by n∈Z π n [A] is A itself. Then h(π) = h(π, A).

384R

Entropy

509

proof I take the two arguments together. In both cases, by 384O, we have h(π, A) ≤ h(π), so I have to show that if B ⊆ A is S any finite partition of unity, then h(π, B) ≤ h(π, A). For (i), let An be S the partition of unity generated by 0≤j
lim

m→∞ m

H(Dm (An , π)) = lim

1

= lim

1

m→∞ m

m→∞ m

H(Dm+n (A, π)) H(Dm (A, π)).

In case (ii), we have Dm (An , π) = π −n [Dm+2n (A, π)] for every m, so that lim

1

m→∞ m

H(Dm (An , π)) = lim

1

= lim

1

m→∞ m m→∞ m

H(Dm+2n (A, π)) H(Dm (A, π)). Q Q

Let An be the purely atomic S closed subalgebra of A generated S by An ; our hypothesis is that the closed subalgebra generated by n∈N An is A itself, that is, that n∈N An is dense. But this means that limn→∞ H(B|An ) = 0 (384Gd). Since h(π, B) ≤ h(π, An ) + H(B|An ) = h(π, A) + H(B|An ) for every n (384N), we have the result. 384Q Bernoulli shifts Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. (a) π is a one-sided Bernoulli shift if there is a closed subalgebra A0 in A such that (i) hπ k [A0 ]ik∈N is Qk independent (that is, µ ¯(inf j≤k π j aj ) = j=0 µ ¯aj for all a0 , . . . , ak ∈ A0 ; see 325L) (ii) the closed subalgebra S of A generated by k∈N π k [A0 ] is A itself. In this case A0 is a root algebra for π. (b) π is a two-sided Bernoulli shift if it is an automorphism and there is a closed S subalgebra A0 in A such that (i) hπ k [A0 ]ik∈Z is independent (ii) the closed subalgebra of A generated by k∈Z π k [A0 ] is A itself. In this case A0 is a root algebra for π. It is important to be aware that a Bernoulli shift can have many, and (in the case of a two-sided shift) very different, root algebras; this is the subject of §386 below. 384R Theorem Let (A, µ ¯) be a probability algebra and π : A → B a Bernoulli shift, either one- or two-sided, with root algebra A0 . (i) If A0 is purely atomic, then h(π) = H(A), where A is the set of atoms of A0 . (ii) If A0 is not purely atomic, then h(π) = ∞. proof (a) The point is that for any partition of unity C ⊆ A0 , h(π, C) = H(C). P P For any n ≥ 1, Dn (C, π) is the partition of unity consisting of elements of the form inf j
H(Dn (C, π)) =

X

q(¯ µ( inf π j cj )) = j

c0 ,... ,cn−1 ∈C

=

X

n−1 X

c0 ,... ,cn−1 ∈C j=0

q(¯ µcj )

c0 ,... ,cn−1 ∈C

Y i6=j

(384Ac) =

n−1 XX j=0 c∈C

X

q(¯ µc) = nH(C).

µ ¯ci

n−1 Y

q(

j=0

µ ¯cj ))

510

Automorphism groups

384R

So 1 n

h(π, C) = limn→∞ H(Dn (C, π)) = H(C). Q Q (b) If A0 is purely atomic and H(A) < ∞, the result can now be read off fromS384P, becauseSthe closed subalgebra of A generated by A is A0 and the closed subalgebra of A generated by k∈N π k [A] or k∈Z π k [A] is A; so h(π) = h(π, A) = H(A). (c) Otherwise, 384J tells us that there are finite partitions of unity C ⊆ A0 such that H(C) is arbitrarily large. Since h(π) ≥ h(π, C) = H(C) for any such C, by (a) and the definition of h(π), h(π) must be infinite, as claimed. 384S Remarks (a) The standard construction of a Bernoulli shift is from a product space, as follows. If (X, Σ, µ0 ) is any probability space, write µ for the product measure on X N ; let (A, µ ¯) be the measure algebra of µ, and A0 ⊆ A the set of equivalence classes of sets of the form {x : x(0) ∈ E} where E ∈ Σ, so that (A0 , µ ¯¹ A0 ) can be identified with the measure algebra of µ0 . We have an inverse-measure-preserving function f : X N → X N defined by setting f (x)(n) = x(n + 1) for every x ∈ X N , n ∈ N, and f induces, as usual, a measure-preserving homomorphism π : A → A. Now π is a one-sided Bernoulli shift with root algebra A0 . P P (i) If a0 , . . . , ak ∈ A0 , express each aj as {x : x(0) ∈ Ej }• , where Ej ∈ Σ. Now π j aj = (f −j {x : x(0) ∈ Ej })• = {x : x(j) ∈ Ej }• for each j, so

Qk Qk T ¯aj . µ ¯(inf j≤k π j aj ) = µ( j≤k {x : x(j) ∈ Ej }) = j=0 µ0 Ej = j=0 µ S k 0 Thus hπ [A0 ]ik∈N is independent. (ii) The closed subalgebra A of A generated by k∈N π k [A0 ] must contain {x : x(k) ∈ E}• for every k ∈ N, E ∈ Σ, so must contain W • for every W in the σ-algebra generated by sets of the form {x : x(k) ∈ E}; but every set measured by µ is equivalent to such a set W . So A0 = A. Q Q (b) The same method gives us two-sided Bernoulli shifts. Again let (X, Σ, µ0 ) be a probability space, and this time write µ for the product measure on X Z ; again let (A, µ ¯) be the measure algebra of µ, and A0 ⊆ A the set of equivalence classes of sets of the form {x : x(0) ∈ E} where E ∈ Σ, so that (A0 , µ ¯¹ A0 ) can once more be identified with the measure algebra of µ0 . This time, we have a measure space automorphism f : X Z → X Z defined by setting f (x)(n) = x(n + 1) for every x ∈ X Z , n ∈ Z, and f induces a measure-preserving automorphism π : A → A. The arguments used above show that π is a two-sided Bernoulli shift with root algebra A0 . (c) I remarked above that a Bernoulli shift will normally have many root algebras. But it is important to know that, up to isomorphism, any root algebra is associated with just one Bernoulli shift of each type. P P(i) Given a probability algebra (A0 , µ ¯0 ) then we can identify it with the measure algebra of a probability space (X, Σ, µ0 ) (321J), and now the constructions of (a) and (b) provide Bernoulli shifts with root algebras isomorphic to (A0 , µ ¯0 ). (ii) Let (A, µ ¯) and (B, ν¯) be probability algebras with one-sided Bernoulli shifts π, φ with root algebras A0 , B0 , and suppose that θ0 : A0 → B0 is a measure-preserving isomorphism. Then (A, µ ¯) can be identified with the probability algebra free product of hπ k [A0 ]ik∈N (325L), while (B, ν¯) can be identified with the probability algebra free product of hπ k [B0 ]ik∈N . For each k ∈ N, φk θ0 (π k )−1 is a measure-preserving isomorphism between π k [A0 ] and φk [B0 ]. Assembling these, we have a measure-preserving isomorphism θ : A → B such that θa = φk θ0 (π k )−1 a whenever k ∈ N and a ∈ π k [A0 ], that is, θπ k a = φk θ0 a for every a ∈ A0 , k ∈ N. Of course θ extends θ0 . If we set

384T

Entropy

511

C = {a : a ∈ A, θπa = φθa}, then C is a closed subalgebra of A. If a ∈ A0 and k ∈ N, then θπ(π k a) = θπ k+1 a = φk+1 θ0 a = φ(φk θ0 a) = φθ(π k a), so π k a ∈ C. Thus φk [A0 ] ⊆ C for every k ∈ N, and C = A. This means that θ : A → B is such that φ = θπθ−1 ; θ is an isomorphism between the structures (A, µ ¯, π) and (B, ν¯, φ) extending the isomorphism θ0 from A0 to B0 . (iii) Now suppose that (A, µ ¯) and (B, ν¯) are probability algebras with two-sided Bernoulli shifts π, φ with root algebras A0 , B0 , and suppose that θ0 : A0 → B0 is a measure-preserving isomorphism. Repeating (ii) word for word, but changing each N into Z, we find that θ0 has an extension to a measure-preserving isomorphism θ : A → B such that θπ = φθ, so that once more the structures (A, µ ¯, π) and (B, ν¯, φ) are isomorphic. Q Q (d) The classic problem to which the theory of this section was directed was the following: suppose we have two two-sided Bernoulli shifts π and φ, one based on a root algebra with two atoms of measure 21 and the other on a root algebra with three atoms of measure 13 ; are they isomorphic? The Kolmogorov-Sinaˇı theorem tells us that they are not, because h(π) = ln 2 and h(φ) = ln 3 are different. The question of which Bernoulli shifts are isomorphic is addressed, and (for countably-generated algebras) solved, in §386 below. (e) We shall need to know that any Bernoulli shift (either one- or two-sided) is ergodic. In fact, it is mixing. P P Let (A, µ ¯) be a probabilitySalgebra and π : A → A a Bernoulli shift S with root algebra A0 . Let B be the subalgebra of A generated by k∈N π k [A0 ] (if π is one-sided) or by k∈Z π k [A0 ] (if π is two-sided). S If b, c ∈ B, there is some n ∈ N such that both belong to the algebra Bn generated by j≤n π j [A0 ] (if π is S one-sided) or by |j|≤n π j [A0 ] (if π is two-sided). If now k > 2n, π k b belongs to the algebra generated by S j j>n π [A0 ]. But this is independent of Bn (cf. 325Xf, 272K), so µ ¯(c ∩ π k b) = µ ¯c · µ ¯(π k b) = µ ¯c · µ ¯b. And this is true for every k ≥ n. Generally, if b, c ∈ A and ² > 0, there are b0 , c0 ∈ B such that µ ¯(b 4 b0 ) ≤ ², 0 µ ¯(c 4 c ) ≤ ², so that lim sup |¯ µ(c ∩ π k b) − µ ¯c · µ ¯b| ≤ lim sup |¯ µ(c0 ∩ π k b0 ) − µ ¯c0 · µ ¯ b0 | k→∞

k→∞

+µ ¯(c 4 c0 ) + µ ¯(π k b 4 π k b0 ) + |¯ µc · µ ¯b − µ ¯c0 · µ ¯ b0 | ≤ 0 + ² + ² + |¯ µc − µ ¯c0 | + |¯ µb − µ ¯ b0 | ≤ 4². As ², b, and c are arbitrary, π is mixing. By 372Qa, it is ergodic. Q Q (f ) The following elementary remark will be useful. If (A, µ ¯) is a probability algebra, π : A → A is a measure-preserving automorphism, and A0 ⊆ A is a closed subalgebra such that hπ k [A0 ]ik∈N is independent, then hπ k [A0 ]ik∈Z is independent. P P If J ⊆ Z is finite and haj ij∈J is a family in A0 , take n ∈ N such that −n ≤ j for every j ∈ J; then Q µ ¯(inf j∈J π j aj ) = µ ¯(inf j∈J π n+j aj ) = j∈J µ ¯ aj . Q Q 384T Isomorphic homomorphisms (a) In this section I have spoken of ‘isomorphic homomorphisms’ without offering a formal definition. I hope that my intention was indeed obvious, and that the next sentence will merely confirm what you have already assumed. If (A1 , µ ¯1 ) and (A2 , µ ¯2 ) are measure algebras, and π1 : A1 → A2 , π2 : A2 → A2 are functions, then I say that (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) are isomorphic if there is a measure-preserving isomorphism φ : A1 → A2 such that π2 = φπ1 φ−1 . In this context, using Maharam’s theorem or otherwise, we can expect to be able to decide whether (A1 , µ ¯1 ) and (A2 , µ ¯2 ) are or are not isomorphic; and if they are, we have a good hope of being able to describe a measure-preserving isomorphism θ : A1 → A2 . In this case, of course, (A2 , µ ¯2 , π2 ) will be isomorphic to (A1 , µ ¯1 , π20 ) where 0 −1 0 π2 = θ π2 θ. So now we have to decide whether (A1 , µ ¯1 , π1 ) is isomorphic to (A1 , µ ¯1 , π2 ); and when π1 , π2

512

Automorphism groups

384T

are measure-preserving Boolean automorphisms, this is just the question of whether π1 , π20 are conjugate in the group Autµ¯1 (A1 ) of measure-preserving automorphisms of A1 . Thus the isomorphism problem, as stated here, is very close to the classical group-theoretic problem of identifying the conjugacy classes in Autµ¯ (A) for a measure algebra (A, µ ¯). But we also want to look at measure-preserving homomorphisms which are not automorphisms, so there would be something left even if the conjugacy problem were solved. (In effect, we are studying conjugacy in the semigroup of all measure-preserving Boolean homomorphisms, not just in its group of invertible elements.) The point of the calculation of the entropy of a homomorphism is that it is an invariant under this kind of isomorphism; so that if π1 , π2 have different entropies then (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) cannot be isomorphic. Of course the properties of being ‘ergodic’ or ‘mixing’ (see 372P) are also invariant. (b) All the main work of this section has been done in of measure algebras; part of my purpose in this volume has been to insist that this is often the right way to proceed, and to establish a language which makes the arguments smooth and natural. But of course a large proportion of the most important homomorphisms arise in the context of measure spaces, and I take a moment to discuss such applications. Suppose that we have two quadruples (X1 , Σ1 , µ1 , f1 ), (X2 , Σ2 , µ2 , f2 ) where, for each i, (Xi , Σi , µi ) is a measure space and fi : Xi → Xi is an inverse-measure-preserving function. Then we have associated structures (A1 , µ ¯1 , π1 ), (A2 , µ ¯2 , π2 ) where (Ai , µ ¯i ) is the measure algebra of (Xi , Σi , µi ) and πi : Ai → Ai is the measure-preserving homomorphism defined by the usual formula πi E • = fi−1 [E]• . Now we can call (X1 , Σ1 , µ1 , f1 ) and (X2 , Σ2 , µ2 , f2 ) isomorphic if there is a measure space isomorphism g : X1 → X2 such that f2 = gf1 g −1 . In this case (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) are isomorphic under the obvious isomorphism φ(E • ) = g[E]• for every E ∈ Σ1 . It is not the case that if the (Ai , µ ¯i , πi ) are isomorphic, then the (Xi , Σi , µi , fi ) are; in fact we do not even need to have an isomorphism of the measure spaces (for instance, one could be Lebesgue measure, and the other the Stone space of the Lebesgue measure algebra). Even when (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) are actually identical, f1 and f2 need not be isomorphic. There are two examples in §343 of a probability space (X, Σ, µ) with a measure space automorphism f : X → X such that f (x) 6= x for every x ∈ X but the corresponding automorphism on the measure algebra is the identity (343I, 343J); writing f0 for the identity map from X to itself, (X, Σ, µ, f0 ) and (X, Σ, µ, f ) are non-isomorphic but give rise to the same (A, µ ¯, π). (c) Even with Lebesgue measure, we can have a problem in a formal sense. Take (X, Σ, µ) to be [0, 1] with Lebesgue measure, and set f (0) = 1, f (1) = 0, f (x) = x for x ∈ ]0, 1[; then f is not isomorphic to the identity function on X, but induces the identity automorphism on the measure algebra. But in this case we can sort things out just by discarding the negligible set {0, 1}, and for Lebesgue measure such a procedure is effective in a wide variety of situations. To formalize it I offer the following definition. 384U Definition Let (X1 , Σ1 , µ1 ) and (X2 , Σ2 , µ2 ) be measure spaces, and f1 : X1 → X1 , f2 : X2 → X2 two inverse-measure-preserving functions. I will say that the structures (X1 , Σ1 , µ1 , f1 ) and (X2 , Σ2 , µ2 , f2 ) are almost isomorphic if there are conegligible sets Xi0 ⊆ Xi such that fi [Xi0 ] ⊆ Xi0 for both i and the structures (Xi0 , Σ0i , µ0i , fi0 ) are isomorphic in the sense of 384Tb, where Σ0i is the algebra of relatively measurable subsets of Xi0 , µ0i is the subspace measure on Xi0 and fi0 = fi ¹Xi0 . 384V I leave the elementary properties of this notion to the exercises (384Xn-384Xp), but I spell out the result for which the definition is devised. I phrase it in the language of §§342-343; if the are not immediately familiar, start by imagining that both (Xi , Σi , µi ) are measurable subspaces of R endowed with some Radon measure (342J, 343H), or indeed that both are [0, 1] with Lebesgue measure. Proposition Let (X1 , Σ1 , µ1 ) and (X2 , Σ2 , µ2 ) be perfect, complete, strictly localizable and countably separated measure spaces, and (A1 , µ ¯1 ), (A2 , µ ¯2 ) their measure algebras. Suppose that f1 : X1 → X1 , f2 : X2 → X2 are inverse-measure-preserving functions and that π1 : A1 → A1 , π2 : A2 → A2 are the measure-preserving Boolean homomorphisms they induce. If (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) are isomorphic, then (X1 , Σ1 , µ1 , f1 ) and (X2 , Σ2 , µ2 , f2 ) are almost isomorphic. proof Because (A1 , µ ¯1 ) and (A2 , µ ¯2 ) are isomorphic, we surely have µ1 X1 = µ2 X2 . If both are zero, we can take X10 = X20 = ∅ and stop; so let us suppose that µ1 X1 > 0. Let φ : A1 → A2 be a measurepreserving automorphism such that π2 = φπ1 φ−1 . Because both µ1 and µ2 are complete and strictly

384Xi

Entropy

513

localizable and compact (343K), there are inverse-measure-preserving functions g1 : X1 → X2 , g2 : X2 → X1 representing φ−1 , φ respectively (343B). Now g1 g2 : X2 → X2 , g2 g1 : X1 → X1 , f2 g1 : X1 → X2 and g1 f1 : X1 → X2 represent, respectively, the identity automorphism on A2 , the identity automorphism on A1 , the homomorphism φ−1 π2 = π1 φ−1 : A2 → A1 and the homomorphism π1 φ−1 again. Next, because both µ1 and µ2 are countably separated, the sets E1 = {x : g2 g1 (x) = x}, H = {x : f2 g1 (x) = g1 f1 (x)} and E2 = {y : g1 g2 (y) = y} are all conegligible (343F). As in part (b) of the proof of 344I, g1 ¹E1 and g2 ¹E2 are the two halves of a bijection, a measure space isomorphism if E1 and E2 are given their subspace −1 measures. T Set G0 = E1 ∩ H, and for n ∈ N0 set Gn+1 = Gn ∩ f1 [Gn ]. Then every G0 n is conegligible, so 0 X1 = n∈N Gn is conegligible. Because X1 is a conegligible subset of E1 , h = g1 ¹X1 is a measure space isomorphism between X10 and X20 = g1 [X10 ], which is conegligible in X2 . Because f1 [Gn+1 ] ⊆ Gn for each n, f1 [X10 ] ⊆ X10 . Because X10 ⊆ H, g1 f1 (x) = f2 g1 (x) for every x ∈ X10 . Next, if y ∈ X20 , g2 (y) ∈ X10 , so f2 (y) = f2 g1 g2 (y) = g1 f1 g2 (y) ∈ g1 [f1 [X10 ]] ⊆ g1 [X10 ] = X20 . Accordingly we have f20 = hf10 h−1 , where fi0 = fi ¹Xi0 for both i. Thus h is an isomorphism between (X10 , f10 ) and (X20 , f20 ), and (X1 , Σ1 , µ1 , f1 ) and (X2 , Σ2 , µ2 , f2 ) are almost isomorphic. 384X Basic exercises (a) Let (A, µ ¯) be a probability algebra and A ⊆ A a partition of unity. Show that if #(A) = n then H(A) ≤ ln n. > (b) Let (A, µ ¯) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A, enumerated as han in∈N . Set a∗n = supi>n ai , An = {a0 , . . . , an , a∗n } for each n. Show that H(An |B) ≤ H(An+1 |B) for every n, and that H(A|B) = limn→∞ H(An |B). (c) Let (A, µ ¯) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A. Show that H(A|B) = 0 iff A ⊆ B. (d) Let (A, µ ¯) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A. Show that H(A|B) = H(A) iff µ ¯(a ∩ b) = µ ¯a · µ ¯b for every a ∈ A, b ∈ B. (Hint: for ‘only if’, start with the case B = {0, b, 1 \ b, 1} and use 384Gc.) (e) Let (A, µ ¯) be a probability algebra and A, B two partitions of unity in A. Show that H(A ∨ B) = H(A) + H(B) iff µ ¯(a ∩ b) = µ ¯a · µ ¯b for all a ∈ A, b ∈ B. Show that H(A ∨ B) = H(A) iff every member of A is included in some member of B, that is, iff A = A ∨ B. ¯ (f ) Let h(Ai , µ ¯i )ii∈I be a family of probability algebras, with probability algebra free product (C, λ) (325K). Suppose that πi : Ai → Ai is a measure-preserving Boolean homomorphism for each i ∈ P I, and that π : C → C is the measure-preserving Boolean homomorphism they induce. Show that h(π) = i∈I h(πi ). (Hint: use 384Gb and N384Gd to show that h(π) is the supremum of h(π, A) as A runs over the finite partitions of unity in i∈I Ai . Use this to reduce to the case I = {0, 1}. Now show that if Ai ⊆ Ai is a finite partition of unity for each i, and A = {a0 ⊗ a1 : a0 ∈ A0 , a1 ∈ A1 }, then H(A) = H(A0 ) + H(A1 ), so that h(π, A) = h(π0 , A0 ) + h(π1 , A1 ).) > (g) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving automorphism. Show that h(π −1 ) = h(π). (h) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. Show that h(π k ) = kh(π) for any k ∈ N. (Hint: if A ⊆ A is a partition of unity, h(π k , A) ≤ h(π k , Dk (A, π)) = kh(π, A).) >(i) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. (i) Suppose there is a partition of unity A ⊆ A such ¯(a ∩ πb) = µ ¯a · µ ¯b for every a ∈ A, b ∈ A (β) S that (α) µ A is the closed subalgebra of itself generated by n∈N π n [A]. Show that π is a one-sided Bernoulli shift, and that h(π) = H(A). (ii) Suppose that π is a one-sided Bernoulli shift of finite entropy. Show that there is a partition of unity satisfying (α) and (β).

514

Automorphism groups

384Xj

> (j) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1[. Fix an integer k ≥ 2, and define f : [0, 1[ → [0, 1[ by setting f (x) = , the fractional part of kx, for every x ∈ [0, 1[; let π : A → B be the corresponding homomorphism. (Cf. 372Xr.) Show that £ π is £a• one-sided Bernoulli shift and that h(π) = ln k. (Hint: in 384Xi, set A = {a0 , . . . , ak−1 } where ai = ki , i+1 for i < k.) k > (k) Let (A, µ ¯) be the measure algebra of Lebesgue measure on [0, 1]. Set f (x) = 2 min(x, 1 − x) for x ∈ [0, 1] (see 372Xm). Show that the corresponding homomorphism π : A → A is a one-sided Bernoulli shift and that h(π) = ln 2. (Hint: in 384Xi, set A = {a, 1 \ a} where a = [0, 12 ]• .) (l) Let (A, µ ¯) be a probability algebra and π : A → A a two-sided Bernoulli shift. Show that π −1 is a twosided Bernoulli shift and that there is a measure-preserving involution φ : A → A such that π −1 = φπφ−1 , so that π is a product of two involutions in Autµ¯ (A). ¯ their probability algebra free product. (m) Let h(Ai , µ ¯i )ii∈I be a family of probability algebras, and (C, λ) Suppose that for each i ∈ I we have a measure-preserving Boolean homomorphism πi : Ai → Ai , and that π : C → C is the measure-preserving homomorphism induced by hπi ii∈I (325Xd). (i) Show that if every πi is a one-sided Bernoulli shift so is π. (ii) Show that if every πi is a two-sided Boolean shift so is π. (n) Show that the relation ‘almost isomorphic to’ (384U) is an equivalence relation. (o) Show that the concept of ‘almost isomorphism’ described in 384U is not changed if we amend the definition to require that the subspaces X10 , X20 should be measurable. (p) Show that if (X1 , Σ1 , µ1 , f1 ) and (X2 , Σ2 , µ2 , f2 ) are almost isomorphic quadruples as described in 384U, then (A1 , µ ¯1 , π1 ) and (A2 , µ ¯2 , π2 ) are isomorphic, where for each i (Ai , µ ¯i ) is the measure algebra of (Xi , Σi , µi ) and πi : Ai → Ai is the measure-preserving Boolean homomorphism derived from fi : Xi → Xi . (q) Let (A, µ ¯) be a probability algebra, and write A for the set of partitions of unity in A not containing 0, ordered by saying that A ≤ B if every member of B is included in some member of A. (i) Show that A is a Dedekind complete lattice, and can be identified with the lattice of purely atomic closed subalgebras of A. Show that for A, B ∈ A, A ∨ B, as defined in 384F, is sup{A, B} in A. (ii) Show that H(A ∨ B) + H(A ∧ B) ≤ H(A) + H(B) for all A, B ∈ A, where ∨, ∧ are the lattice operations on A. (iii) Set A1 = {A : A ∈ A, H(A) < ∞}. For A, B ∈ A1 set ρ(A, B) = 2H(A ∨ B) − H(A) − H(B). Show that ρ is a metric on A1 (the entropy metric). (iv) Show that if π : A → A is a measure-preserving Boolean homomorphism, then |h(π, A) − h(π, B)| ≤ ρ(A, B) for all A, B ∈ A1 . (iv) Show that the lattice operations ∨, ∧ are ρ-continuous on A1 . (v) Show that H : A1 → [0, ∞[ is order-continuous. (vi) Show that if B is any closed subalgebra of A, then A 7→ H(A|B) is order-continuous on A1 . 384Y Further exercises (a) Let (A, µ ¯) be a probability algebra, and write P for the lattice of closed subalgebras of A. Show that if A is any partition of unity in A of finite entropy, then the order-preserving function B 7→ −H(A|B) : P → ]−∞, 0] is order-continuous. (b) Let (A, µ ¯) be a probability algebra, A a partition of unity in A of finite entropy, and π : A → A a measure-preserving Boolean homomorphism. Show that h(π, A) = limn→∞ H(A|Bn ), where Bn is the closed S subalgebra of A generated by 1≤i≤n π i [A]. (Hint: use 384Gb to show that H(A|Bn ) = H(Dn+1 (A, π)) − H(Dn (A, π)) and observe that limn→∞ H(A|Bn ) is defined.) (c) Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. Suppose that there is a partition of unity A of finite entropy such that the closed subalgebra of A generated S by i≥1 π i [A] is A. Show that h(π) = 0. (Hint: use 384Yb and 384Pa.) (d) Let µ be Lebesgue measure on [0, 1[, and take any α ∈ ]0, 1[. Let f : [0, 1[ → [0, 1[ be the measure space automorphism defined by saying that f (x) is to be one of x + α, x + α − 1. Let (A, µ ¯) be the measure algebra of ([0, 1[ , µ) and π : A → A the measure-preserving automorphism corresponding £to f £. Show that • h(π) = 0. (Hint: if α ∈ Q, use 384Xh; otherwise use 384Yc with A = {a, 1 \ a} where a = 0, 21 .)

385A

More about entropy

515

R 1 (e) Set X = [0, 1] \ Q, let ν be the measure on X defined by setting νE = ln12 E 1+x dx for every Lebesgue measurable set E ⊆ X, and for x ∈ X let f (x) be the fractional part < x1 > of x1 . Recall that f is inverse-measure-preserving for ν (372N). Let (A, ν¯) be the measure algebra of (X, ν) and π : A → A the homomorphism corresponding to f . Show that h(π) = π 2 /6 ln 2. (Hint: use the Kolmogorov-Sinaˇı theorem and 372Yf(v).) (f ) Consider the triplets ([0, 1[ , µ1 , f1 ) and ([0, 1], µ2 , f2 ) where µ1 , µ2 are Lebesgue measure on [0, 1[, [0, 1] respectively, f1 (x) = <2x> for each x ∈ [0, 1[, and f2 (x) = 2 min(x, 1 − x) for each x ∈ [0, 1]. Show that these structures are almost isomorphic in the sense of 384U, and give a formula for an isomorphism. (g) Let (A, µ ¯) be a probability algebra, and A1 the set of partitions of unity of finite entropy not containing 0, as in 384Xq. Show that A1 is complete under the entropy metric. (Hint: show that if hAn in∈N S is a nondecreasing sequence in A1 and supn∈N H(An ) < ∞, then the closed subalgebra of A generated by n∈N An is purely atomic.) 384 Notes and comments In preparing this section I have been heavily influenced by Petersen 83. I have taken almost the shortest possible route to Theorem 384P, the original application of the theory, ignoring both the many extensions of these ideas and their intuitive underpinning in the concept of the quantity of ‘information’ carried by a partition. For both of these I refer you to Petersen 83. The techniques described there are I think sufficiently powerful to make possible the calculation of the entropy of any of the measure-preserving homomorphisms which have yet appeared in this treatise. Of course the idea of entropy of a partition, or of a homomorphism, can be translated into the language of probability spaces and inverse-measure-preserving functions; if (X, Σ, µ) is a probability space, with measure algebra (A, µ ¯), then partitions of unity in A correspond (subject to decisions on how to treat negligible sets) to countable partitions of X into measurable sets, and an inverse-measure-preserving function from X to itself gives rise to a measure-preserving homomorphism πf : A → A; so we can define the entropy of f to be h(πf ). The whole point of the language I have sought to develop in this volume is that we can do this when and if we choose; in particular, we are not limited to those homomorphisms which are representable by inverse-measure-preserving functions. But of course a large proportion of the most important examples do arise in this way (see 384Xj, 384Xk). The same two examples are instructive from another point of view: the case k = 2 of 384Xj is (almost) isomorphic to the tent map of 384Xk. The similarity is obvious, but exhibiting an actual isomorphism is I think another matter (384Yf). I must say ‘almost’ isomorphic here because the doubling map on [0, 1[ is everywhere two-to-one, while the tent map is not, so they cannot be isomorphic in any exact sense. This is the problem grappled with in 384T-384V. In some moods I would say that a dislike of such contortions is a sign of civilized taste. Certainly it is part of my motivation for working with measure algebras whenever possible. But I have to say also that new ideas in this topic arise more often than not from actual measure spaces, and that it is absolutely necessary to be able to operate in the more concrete context.

385 More about entropy In preparation for the next two sections, I present a number of basic facts concerning measure-preserving homomorphisms and entropy. Compared with the work to follow, they are mostly fairly elementary, but the Halmos-Rokhlin-Kakutani lemma (385E) and the Shannon-McMillan-Breiman theorem (385G), in their full strengths, go farther than one might expect. 385A Periodic and aperiodic parts If X is a set and f : X → X is a bijection, then the orbits Ωx = {f n (x) : n ∈ Z} of f are described by their cardinalities, and X has a natural decomposition into the sets Xi = {x : #(Ωx ) = i} for 1 ≤ i ≤ ω. Corresponding to this is a partition of unity of a Dedekind complete Boolean algebra with an automorphism, as follows. Definition If A is a Boolean algebra, a Boolean homomorphism π : A → A is periodic, with period n ≥ 2, if A 6= {0}, π n is the identity operator and whenever b ∈ A \ {0} and 1 ≤ i < n there is a c ⊆ b such that

516

Automorphism groups

385A

π i c 6= c. π is periodic with period 1 iff it is the identity operator. π is aperiodic if for every non-zero b ∈ A, n ≥ 1 there is a c ⊆ b such that π n c 6= c. I remark immediately that if π is aperiodic, so is π n for every n ≥ 1. Note that if A = {0} then the trivial automorphism of A is counted both as aperiodic and as periodic with period 1. 385B Lemma Let A be a Dedekind complete Boolean algebra and π : A → A a Boolean homomorphism which is periodic with period n ≥ 2. Then π is a Boolean automorphism and there is an a ∈ A such that (a, πa, π 2 a, . . . , π n−1 a) is a partition of unity in A; that is (in the language of 381G) π is of the form −− (← a1−π−a−2−π−−.− .− .− π an ) where (a1 , . . . , an ) is a partition of unity in A. proof Because π n : A → A is injective and surjective, so is π, and π is a Boolean automorphism, therefore order-continuous. Set D = {d : d ∈ A, π i d ∩ d = 0 whenever 1 ≤ i < n}. Then the supremum of any upwards-directed subset of D is defined in A (because A is Dedekind complete) and belongs to D (because π i is order-continuous for every i – use 313Bc); so D has a maximal element a, by Zorn’s lemma. ?? If (a, πa, . . . , π n−1 a) is not a partition of unity in A, set b0 = 1 \ (a ∪ πa ∪ . . . ∪ π n−1 a). Then πb0 = 1 \ (πa ∪ . . . ∪ π n−1 a ∪ a) = b0 . Now we can choose b1 , . . . , bn−1 such that 0 6= bi ⊆ bi−1 and π i bi ∩ bi = 0 for 1 ≤ i < n. P P Given that bi−1 6= 0, where 1 ≤ i < n, then (by the definition of ‘periodic’ homomorphism) there is a c ⊆ bi−1 such that π i c 6= c. If c \ π i c 6= 0, take bi = c \ π i c. Otherwise, π i c \ c 6= 0 so c \ π −i c 6= 0 and we can take bi = c \ π −i c. Q Q At the end of this induction, set d = bn−1 ; then d ∩ π i d = 0 for 1 ≤ i < n, so d ∈ D; also d ∩ π i a = π i d ∩ a = 0 for 1 ≤ i < n (because π i d ⊆ π i b0 = b0 for every i), so a ∪ d ∈ D, which is supposed to be impossible. X X Thus (a, πa, . . . , π n−1 a) is a partition of unity, as required. 385C Proposition Let A be a Dedekind complete Boolean algebra and π : A → A an injective ordercontinuous Boolean homomorphism. Then there is a partition of unity hci i1≤i≤ω in A such that πci = ci for every i and π¹ Acn is periodic with period n whenever n is finite and cn 6= 0, while π¹ Acω is aperiodic. Remark As usual, I write Aa for the principal ideal of A generated by a. proof For each n ≥ 1, set Bn = {b : b ∈ A, π n c = c for every c ⊆ b},

bn = sup Bn .

Because π is a Boolean homomorphism, Bn is an ideal of A; because π is order-continuous, bn ∈ Bn , that is, Bn is the principal ideal Abn . Also πbn = bn . P P If b ∈ Bn and c ⊆ πb, then π n−1 c ⊆ π n b = b, so π n−1 (π n c) = π n (π n−1 c) = π n−1 c. But π n−1 , like π, is injective, so π n c = c. As c is arbitrary, πb ∈ Bn . As π is order-continuous, πbn = supb∈Bn πb ⊆ bn . But this means that π

i+1

i

bn ⊆ π bn for every i, so bn = π n bn ⊆ πbn ⊆ bn

and πbn = bn . Q Q Set cn = bn \ sup1≤i
cω = 1 \ supn≥1 bn = 1 \ supn≥1 cn .

Then hci i1≤i≤ω serves. P P Because πbn = bn for every n, πci = ci for every i ≤ ω. So πc ⊆ πci = ci for every c ⊆ ci , and π¹ Aci is a Boolean homomorphism from Aci to itself, for every i ≤ ω. If 1 ≤ i < ω and c ⊆ ci , then c ⊆ bi so π i c = c; accordingly (π¹ Aci )i = π¹ Aci . If 1 ≤ j < i ≤ ω and c ⊆ ci is non-zero, then c⊆ 6 bj , so there is a non-zero d ⊆ c such that π j d 6= d. This shows both that π¹ Aci is periodic with period i if 1 ≤ i < ω and that π¹ Acω is aperiodic. So we have an appropriate partition hci ii≤ω . Q Q 385D

The following elementary fact is essential.

385E

More about entropy

517

Lemma Let (A, µ ¯) be a totally finite measure algebra and π : A → A a measure-preserving Boolean homomorphism. Then π(supn∈N π n a) = supn∈N π n a,

a ⊆ supn≥1 π n a

for every a ∈ A. proof Set c = supn∈N π n a. Then πc = supn≥1 π n , because π is order-continuous (324Kb), so πc ⊆ c; as π is measure-preserving and µ ¯c < ∞, πc = c. Now a ⊆ c = πc = supn∈N π n+1 a = supn≥1 π n a. Remark See 382Yc-382Yd. 385E The Halmos-Rokhlin-Kakutani lemma Let (A, µ ¯) be a totally finite measure algebra and π : A → A a measure-preserving Boolean homomorphism. Set C = {c : πc = c}. Then C is a closed subalgebra of A, and the following are equiveridical: (i) π is aperiodic; (ii) A is relatively atomless over C (definition: 331A); (iii) whenever n ≥ 1 and 0 ≤ γ < n1 there is an a ∈ A such that a, πa, π 2 a, . . . , π n−1 a are dist and µ ¯(a ∩ c) = γ µ ¯c for every c ∈ C; (iv) whenever n ≥ 1, 0 ≤ γ < n1 and B ⊆ A is finite, there is an a ∈ A such that a, πa, π 2 a, . . . , π n−1 a are dist and µ ¯(a ∩ b) = γ µ ¯b for every b ∈ B. proof Of course C is a subalgebra of A because π is a Boolean homomorphism, and is (order-)closed because π is (order-)continuous (324Kb). (i)⇒(ii) ?? Suppose, if possible, that π is aperiodic, but there is a non-zero a ∈ A such that Aa = {a ∩ c : c ∈ C}. Let n ≥ 1 be such that b = a ∩ π n a 6= 0 (385D). If d ⊆ b there is some c ∈ C such that d = a ∩ c and π n d = π n a ∩ c ⊇ b ∩ c ⊇ d; because π is measure-preserving, π n d = d. But this means that π is not aperiodic. X X (ii)⇒(iii) Set δ = n1 ( n1 − γ) > 0. By 331B, there is a d ∈ A such that µ ¯(c ∩ d) = δ µ ¯c for every c ∈ C. Set dk = π k d \ supi
Pn−1 i=0

µ ¯(c ∩ π i a∗ ) = n¯ µ(c ∩ a∗ ),

µc = γ µ ¯c µ ¯(c ∩ a∗ \ d∗ ) ≥ µ ¯(c ∩ a∗ ) − µ ¯(c ∩ d∗ ) ≥ ( n1 − nδ)¯ for every c ∈ C.

518

Automorphism groups

385E

By 331B again (applied to the principal ideal of A generated by a∗ \ d∗ ) there is an a ⊆ a∗ \ d∗ such that µ ¯(a ∩ c) = γ µ ¯c for every c ∈ C. For 0 < i < n, π i a∗ ∩ a∗ = supk,l∈N π i dkn ∩ dln ⊆ supm≤i dm ⊆ d∗ , so π i a ∩ a = 0; accordingly a, πa, . . . , π n−1 a are all dist and (iii) is satisfied. (iii)⇒(iv) Note that A is certainly atomless, since for every k ≥ 1 we can find a c ∈ A such that µ ¯1 c, πc, . . . , π k−1 c are dist and µ ¯c = k+1 , so that we have a partition of unity consisting of sets of measure µ ¯1 0 . Let B be the set of atoms of the (finite) subalgebra of A generated by B, and m = #(B 0 ). Let δ > 0 k+1 and r, k ∈ N be such that 3δ ≤ (1 − nγ)¯ µb for every b ∈ B 0 ,

m(¯ µ1)2 < rδ 2 ,

kδ ≥ µ ¯1.

By (iii), there is a c ∈ A such that c, πc, . . . , π nr(k+1)−1 c are dist and µ ¯(supi
m

Pr−1

¯(π s(j) dj ∩ b). Then the Xj P P We can express the random variable µ ¯(d ∩ b) as X = j=0 Xj , where Xj = µ µ ¯1 are independent random variables. For each j, Xj takes values between 0 and k µ ¯c ≤ nr , and has expectation 1 0 µ ¯ (e ∩ b), where j n e0j = supi
1 ¯(e0 ∩ b) nµ

where e0 = supj
has measure n¯ µc ≤

n¯ µ1 nr(k+1)

for each j, so µ ¯(e \ e0 ) ≤ Var(X) =

Pr−1 j=0

µ ¯1 k+1

and µ ¯(1 \ e0 ) ≤ 2δ; thus E(X) ≥

Var(Xj ) ≤ r

¡ µ¯1 ¢2 nr

=

1 µb − 2δ), n (¯

while

(¯ µ1)2 . n2 r

But this means that (¯ µ1)2 n2 r

≥

¡ δ ¢2 n

¡ ¢ 1 Pr X ≤ (¯ µb − 3δ) , n

and ¡ ¢ (¯µ1)2 1 1 Pr X ≤ (¯ µb − 3δ) ≤ < 2 n

rδ

m

by the choice of r. Q Q This is true for every b ∈ B 0 , while #(B 0 ) = m. There must therefore be some choice of s(0), . . . , s(r − 1) such that, taking d∗ = supj
µb − 3δ) ≥ γ µ ¯b µ ¯(d∗ ∩ b) ≥ (¯ for every b ∈ B 0 , while d∗ , πd∗ , . . . , π n−1 d∗ are dist. Because A is atomless, there is a d ⊆ d∗ such that µ ¯(d ∩ b) = γ µ ¯b for every b ∈ B 0 . Since every member of B is a dist union of of B 0 , µ ¯(d ∩ b) = γ µ ¯b for every b ∈ B. (iv)⇒(i) If a ∈ A \ {0, 1} and n ≥ 1 then (iv) tells us that there is a b ∈ A such that b, πb, . . . , π n b are all dist and µ ¯(1 \ supi≤n π i b) < µ ¯a. Now there must be some i < n such that d = π i b ∩ a 6= 0, in which case d ∩ π n d ⊆ π i b ∩ π i+n b = π i (b ∩ π n b) = 0,

385G

More about entropy

519

and π n d 6= d. As n and a are arbitrary, π is aperiodic. 385F Corollary An ergodic measure-preserving Boolean homomorphism on an atomless totally finite measure algebra is aperiodic. proof This is (ii)⇒(i) of 385E in the case C = {0, 1}. 385G

I turn now to a celebrated result which is a kind of strong law of large numbers.

The Shannon-McMillan-Breiman theorem Let (A, µ ¯) be a probability algebra, π : A → A a measurepreserving Boolean homomorphism, and A ⊆ A a partition of unity of finite entropy. For each n ≥ 1, set 1 P 1 wn = )χd, d∈Dn (A,π) ln( n

µ ¯d

where Dn (A, π) is the partition of unity generated by {π i a : a ∈ A, i < n}, as in 384K. Then hwn in∈N is norm-convergent in L1 = L1 (A, µ ¯) to w say; moreover, hwn in∈N is order*-convergent to w (definition: 367A). If T : L0 (A) → L0 (A) is the Riesz homomorphism defined by π, so that T (χa) = χ(πa) for every a ∈ A (364R), then T w = w. proof (Petersen 83) We may suppose that 0 ∈ / A. (a) For each n ∈ N, let Bn be the subalgebra of A generated by {π i a : a ∈ A, 1 ≤ i ≤ n}, Bn the set of its atoms, and Pn the corresponding conditional expectation operator on L1 = L1 (A) (365R). Let B be the S closed subalgebra of A generated by n∈N Bn , and P the corresponding conditional expectation operator. Observe that Bn = π[Dn (A, π)] and that, in the language of 384F, Dn+1 (A, π) = A ∨ Bn . Let C be the fixed-point algebra {c : c ∈ A, πc = c} and Q the associated conditional expectation. Set L0 = L0 (A), and ¯ be the function from {v : [[v > 0]] = 1} to L0 corresponding to ln : ]0, ∞[ → R (364I). let ln (b) It will save a moment later if I note an elementary fact here: if v ∈ L1 , then h n1 T n vin≥1 is order*convergent and k k1 -convergent to 0. P P We know from the theorem (372G) that h˜ vn in∈N is order*Pnergodic 1 1 n n+1 i convergent and k k1 -convergent to Qv, where vñ = n+1 T v. Now T v = v ˜ − vñ−1 is order*n i=0 n n convergent and k k1 -convergent to Qv − Qv = 0 (using 367C for ‘order*-convergent’). Q Q (c) Set vn =

P a∈A

Pn (χa) × χa =

P a∈A,b∈Bn

µ ¯(a∩b) χ(a ∩ b). µ ¯b

By Lévy’s martingale theorem (275I, 367Kb), hvn × χain∈N = hPn (χa) × χain∈N is for every a ∈ A; consequently hvn in∈N order*-converges to v = P order*-convergent to P (χa) × χa ¯ vn in∈N order*-converges to ln ¯ v. P P (χa) × χa. It follows that h ln P The point is that, for any a ∈ A, a∈A ¯ vn is defined. Similarly, ln ¯ v is defined, n ∈ N, a ⊆ [[Pn (χa) > 0]], so that [[vn > 0]] = 1 for every n, and ln ¯ vn in∈N order*-converges to ln ¯ v by 367I. Q and hln Q As 0 ≤ vn ≤ χ1 for every n, hvn in∈N → v for k k1 , by the Dominated Convergence Theorem (367J). ¯ vn in∈N is order-bounded in L1 . P ¯ vn ≤ 0 for every n, because Pn (χa) ≤ Pn (χ1) ≤ Next, hln P Of course ln ¯ vn : n ∈ N} is bounded below in L1 , we use an idea from the χ1 for each a, so vn ≤ χ1. To see that {ln fundamental martingale inequality 275D. Set v∗ = inf n∈N vn . For α > 0, a ∈ A and n ∈ N set ban (α) = [[Pn (χa) < α]] ∩ inf i
R ban

χa = (α)

R ban (α)

Pn (χa) ≤ α¯ µ(ban (α)),

520

Automorphism groups

µ ¯(a ∩ [[v∗ < α]]) ≤ min(¯ µa,

∞ X

385G

µ ¯(a ∩ ban (α)))

n=0 ∞ X

≤ min(¯ µa, α

µ ¯ban (α)) ≤ min(¯ µa, α).

n=0

¯ v∗ is defined. Moreover, Letting α ↓ 0, µ ¯(a ∩ [[v∗ = 0]]) = 0 for every a ∈ A, so [[v∗ > 0]] = 1, and ln ¯ v∗ > − ln α]]) = µ µ ¯(a ∩ [[− ln ¯(a ∩ [[v∗ < α]]) ≤ min(¯ µa, α) for every a ∈ A, α > 0; that is, ¯ v∗ > β]]) ≤ min(¯ µ ¯(a ∩ [[− ln µa, e−β ) for every a ∈ A, β ∈ R. Accordingly Z Z ∞ XZ ¯ v∗ ) = ¯ v∗ > β]]dβ = (− ln µ ¯[[− ln 0

≤

X ¡Z

a∈A

=

X

a∈A

Z

ln(1/¯ µa)

e−β dβ

¢

ln(1/¯ µa)

ln(

ln(

∞

µ ¯a dβ +

0

X¡

a∈A

=

min(¯ µa, e−β )dβ

0

a∈A

=

∞

¯ v∗ > β]])dβ µ ¯(a ∩ [[− ln

0

a∈A

XZ

∞

1 )¯ µa + e µ ¯a

1 )¯ µa + µ ¯a

¢ ln µ ¯a

X

µ ¯a = H(A) + 1 < ∞

a∈A

¯ v∗ belongs to L1 , and of course it is a lower bound for because A has finite entropy. But this means that ln ¯ {ln vn : n ∈ N}. Q Q ¯ v ∈ L1 and hln ¯ vn in∈N → ln ¯ v for k k1 . By 367J again, ln (d) Fix n ∈ N for the moment. For each d ∈ Dn+1 (A, π) let d0 be the unique element of Bn such that d ⊆ d0 . Then X

(n + 1)wn+1 =

ln(

d∈Dn+1 (A,π)

=

X

ln(

b∈Bn

=

1 )χd − µ ¯d0

1 )χb − µ ¯b

X

ln(

d∈Dn (A,π)

X

ln(

d∈Dn+1 (A,π)

X

ln(

a∈A,b∈Bn ,a∩b6=0

µ ¯d )χd µ ¯d0

µ ¯(a∩b) )χ(a ∩ b) µ ¯b

1 ¯ vn )χ(πd) − ln µ ¯(πd)

¯ vn . = T (nwn ) − ln Inducing on n, starting from w1 =

P a∈A

ln(

1 )χa µ ¯a

¯ v0 , = − ln

we get nwn =

Pn−1 i=0

¯ vn−i−1 ), T i (− ln

wn =

1 n

Pn−1 i=0

¯ vn−i−1 ) T i (− ln

for every n ≥ 1. Pn−1 ¯ v) for n ≥ 1. By the Ergodic Theorem, hw0 in≥1 is order*-convergent and (e) Set wn0 = n1 i=0 T i (− ln n ¯ v), and T w = w. To estimate wn − w0 , set u∗ = sup ¯ ¯ k k1 -convergent to w = Q(− ln k≥n | ln vk − ln v| for n n

385I

More about entropy

521

each n ∈ N. Then hu∗n in∈N is a non-increasing sequence, u∗0 ∈ L1 (by (c) above), and inf n∈N u∗n = 0 because ¯ vn in∈N order*-converges to ln ¯ v. Now, whenever n > m ∈ N, hln |wn − wn0 | ≤ =

≤

≤

=

≤

1 n

n−1 X

¯ v − ln ¯ vn−i−1 | T i | ln

i=0

1¡ n

n−m−1 X i=0

T i u∗m +

m−1 X

i=0

1 n−m 1 n−m

¢

¢ ¯ v − ln ¯ vj | T n−1−j | ln

j=0

n−m−1 X

1 ¡ n−m

¯ v − ln ¯ vn−i−1 | T i | ln

i=n−m

n−m−1 1¡ X

n

n−1 X

¯ v − ln ¯ vn−i−1 | + T i | ln

T i u∗m

m−1 X

+

i=0

T n−1−j u∗0

¢

j=0

n−m−1 X

T i u∗m

+

i=0 n−m−1 X

T i u∗m +

i=0

1 T n−m n−m

m−1 X

T m−1−j u∗0

j=0

1 T n−m u ˜m , n−m

Pm−1

setting u ˜m = j=0 T m−1−j u∗0 . Holding m fixed and letting n → ∞, we know that 1 Pn−m−1 n−m

i=0

T i u∗m

is order*-convergent and k k1 -convergent to Qu∗m . As for the other term, and k k1 -convergent to 0, by (b). What this means is that

1 n−m u ˜m n−m T

is order*-convergent

lim supn→∞ |wn − wn0 | ≤ Qu∗m , lim supn→∞ kwn − wn0 k1 ≤ kQu∗m k1 for every m ∈ N. Since hQu∗m im∈N is surely a non-decreasing sequence with infimum 0, lim supn→∞ |wn − wn0 | = 0,

lim supn→∞ kwn − wn0 k1 = 0.

Since wn0 is order*-convergent and k k1 -convergent to w, so is wn . 385H Corollary If, in 385G, π is ergodic, then hwn in∈N is order*-convergent and k k1 -convergent to h(π, A)χ1. proof R Because the limit w in 385G has T w = w, it must be of the form γχ1, because π is ergodic. Now γ = w must be Z lim

n→∞

1 n→∞ n

wn = lim

X d∈Dn (A,π)

ln(

1 )¯ µd µ ¯d

1 n→∞ n

= lim

X

q(¯ µd)

d∈Dn (A,π)

(where q is the function of 384A) = lim

1

n→∞ n

H(Dn (A, π)) = h(π, A).

385I Definition Set p(t) = t ln t for t > 0, p(0) = 0; for any Dedekind σ-complete Boolean algebra A, let p¯ : L0 (A)+ → L0 (A) be the corresponding function, as in 364I. (Thus p = −q where q is the function of 384A.)

522

Automorphism groups

385J

´ r 67, Kullback 67) Let (A, µ 385J Lemma (Csisz a ¯) be a probability algebra, and u a member of R L1 (A, µ ¯)+ such that u = 1. Then

R

( |u − χ1|)2 ≤ 2

R

p¯(u).

R R ¯b = 1 − α and b u = 1 − β. Surely β ≤ α < 1. proof Set a = [[u < 1]], α = µ ¯a, β = a u, b = 1 \ a. Then µ If α = 0 then u = χ1 and the result is trivial; so let us suppose that 0 < α < 1. Because the function p is convex,

R

p¯(u) ≥ µ ¯a · p( a

1 µ ¯a

R

a

β α

u) = αp( ) = p(β) − β ln α,

(using 233Ib/365Rb for the first inequality), and similarly

R

b

Also

R

|u − χ1| =

R a

p¯(u) ≥ p(1 − β) − (1 − β) ln(1 − α).

(χ1 − u) +

R b

(u − χ1) = α − β + (1 − β) − (1 − α) = 2(α − β),

so Z

Z 1 p¯(u) − ( |u − χ1|)2 ≥ p(β) − β ln α + p(1 − β) − (1 − β) ln(1 − α) − 2(α − β)2 2

= φ(β) say. Now φ is continuous on [0, 1] and arbitrarily often differentiable on ]0, 1[, φ(α) = 0, φ0 (t) = ln t − ln α − ln(1 − t) + ln(1 − α) + 4(α − t) for t ∈ ]0, 1[, φ0 (α) = 0, φ00 (t) =

1 t

+

1 1−t

− 4 ≥ 0 for t ∈ ]0, 1[.

So φ(t) ≥ 0 for t ∈ [0, 1] and, in particular, φ(β) ≥ 0; but this means that

R

1 2

R

p¯(u) − ( |u − χ1|)2 ≥ 0,

R R that is, ( |u − χ1|)2 ≤ 2 p¯(u), as claimed.

385K Corollary Whenever (A, µ ¯) is a probability algebra and A, B are partitions of unity of finite entropy, X

|¯ µ(a ∩ b) − µ ¯a · µ ¯b| ≤

p 2(H(A) + H(B) − H(A ∨ B)).

a∈A,b∈B

proof Replacing A, B by A \ {0} and B \ {0} if necessary, we may suppose that neither A nor B contains ¯ be the probability algebra free product of (A, µ {0}. Let (C, λ) ¯) with itself (325E, 325K). Set P µ ¯(a∩b) χ(a ⊗ b) ∈ L0 (C); u = a∈A,b∈B µ ¯a·¯ µb

then u is non-negative and integrable and

R

u=

P a∈A,b∈B

µ ¯(a ∩ b) = 1. Now

385N

More about entropy

Z p¯(u) =

X

µ ¯(a ∩ b) ln

a∈A,b∈B

= −H(A ∨ B) −

µ ¯(a∩b) µ ¯a·¯ µb

X X

X

µ ¯(a ∩ b) ln µ ¯a −

a∈A,b∈B

= −H(A ∨ B) −

523

µ ¯a ln µ ¯a −

a∈A

X

µ ¯(a ∩ b) ln µ ¯b

a∈A,b∈B

µ ¯b ln µ ¯b

b∈B

= H(A) + H(B) − H(A ∨ B). On the other hand,

R

|u − χ1| =

P

µ ¯a · µ ¯b| a∈A,b∈B

So what we are seeking to prove is that

R

µ ¯(a∩b) µ ¯a·¯ µb

− 1| =

P a∈A,b∈B

|¯ µ(a ∩ b) − µ ¯a · µ ¯b|.

q R |u − χ1| ≤ 2 p¯(u),

which is 385J. 385L The next six lemmas are notes on some more or less elementary facts which will be used at various points in the next section. The first two are nearly trivial. Lemma Let (A, µ ¯) be a probability algebra and hai ii∈I , hbi ii∈I two partitions of unity in A. Then 1P ¯(ai 4 bi ). µ ¯(supi∈I ai ∩ bi ) = 1 − i∈I µ 2

proof µ ¯(sup ai ∩ bi ) = i∈I

X i∈I

µ ¯(ai ∩ bi ) =

X1 i∈I

=1−

2 1 2

(¯ µai + µ ¯bi − µ ¯(ai 4 bi )) X

µ ¯(ai 4 bi ).

i∈I

385M Lemma Let (A, µ ¯) be a totally finite measure algebra, hBk ik∈N a non-decreasing sequence of S subsets of A such that 0 ∈ B0 , and hci ii∈I a partition of unity in A such that ci ∈ k∈N Bk for every i ∈ I. Then limk→∞ supi∈I ρ(ci , Bk ) = 0, writing ρ(c, B) = inf b∈B µ ¯(c 4 b) for c ∈ A and non-empty B ⊆ A, as in 3A4I. proof Let ² > 0. Then J = {j : j ∈ I, µ ¯cj ≥ ²} is finite. For each j ∈ J, limk→∞ ρ(ci , Bk ) = 0, by 3A4I, while ρ(ci , Bk ) ≤ µ ¯(ci 4 0) = µ ¯ci ≤ ² for every i ∈ I \ J. So lim supk→∞ supi∈I ρ(ci , Bk ) ≤ max(², lim supk→∞ supi∈J ρ(ci , Bk )) = ². As ² is arbitrary, we have the result. 385N Lemma Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. Let A, B and C be partitions of unity in A. (a) H(A ∨ B ∨ C) + H(C) ≤ H(B ∨ C) + H(A ∨ C). (b) h(π, A) ≤ h(π, A ∨ B) ≤ h(π, A) + h(π, B) ≤ h(π, A) + H(B). (c) If H(A) < ∞,

524

Automorphism groups

385N

h(π, A) = inf H(Dn+1 (A, π)) − H(Dn (A, π)) n∈N

= lim H(Dn+1 (A, π)) − H(Dn (A, π)). n→∞

(d) If H(A) < ∞ and B is any closed subalgebra of A such that π[B] ⊆ B, then h(π, A) ≤ h(π¹ B) + H(A|B). proof (a) Let C be the closed subalgebra of A generated by C, so that C is purely atomic and C is the set of its atoms. Then H(A ∨ B ∨ C) + H(C) = H(A ∨ B|C) + 2H(C) ≤ H(A|C) + H(B|C) + 2H(C) = H(A ∨ C) + H(B ∨ C) by 384Gb and 384Ga. (b) We need only observe that Dn (A ∨ B, π) = Dn (A, π) ∨ Dn (B, π) for every n ∈ N, being the partition of unity generated by {π i a : i < n, a ∈ A} ∪ {π i b : i < n, b ∈ B}. Consequently h(π, A) = lim

1

= lim

1

≤ lim

1

n→∞ n

n→∞ n n→∞ n

H(Dn (A, π)) ≤ lim

1

n→∞ n

H(Dn (A, π) ∨ Dn (B, π))

H(Dn (A ∨ B, π)) = h(π, A ∨ B) (H(Dn (A, π) + H(Dn (B, π))) = h(π, A) + h(π, B)

≤ h(π, A) + H(B) as remarked in 384M. (c) Set γn = H(Dn+1 (A, π)) − H(Dn (A, π)) for each n ∈ N. By 384H, γn ≥ 0. From (a) we see that γn+1 = H(A ∨ π[Dn+1 (A, π)]) − H(A ∨ π[Dn (A, π)]) ≤ H(π[Dn+1 (A, π)]) − H(π[Dn (A, π]) = γn for every n ∈ N. So limn→∞ γn = inf n∈N γn ; write γ for the common value. Now 1 1 Pn−1 h(π, A) = limn→∞ H(Dn (A, π)) = limn→∞ i=0 γi = γ n

n

(273Ca). (d) Let P : L1µ¯ → L1µ¯ be the conditional expectation operator corresponding to B. Let hbk ik∈N be a sequence running over {[[P (χa) > q]] : a ∈ A, q ∈ Q}, so that bk ∈ B for every k, and for each k ∈ N let Bk ⊆ B be the subalgebra generated by {bi : i ≤ k}; let Pk be the conditional expectation operator S corresponding to Bk . Writing B∞ ⊆ B for k∈N Bk , and P∞ for the corresponding conditional expectation operator, then P (χa) ∈ L0 (B∞ ), so P∞ (χa) = P (χa), for every a ∈ A. So R P H(A|B) = a∈A q(P χa) = H(A|B∞ ) = limk→∞ H(A|Bk ), by 384Gd. For each k, let Bk be the set of atoms of Bk . Then h(π, A) ≤ h(π, Bk ) + H(A|Bk ) ≤ h(π¹ B) + H(A|Bk ) by 384N and the definition of h(π¹ B). So h(π, A) ≤ h(π¹ B) + limk→∞ H(A|Bk ) = h(π¹ B) + H(A|B).

385O

More about entropy

525

385O Lemma Let (A, µ ¯) be a probability algebra and B a closed subalgebra. (a) There is a function h : A → B such that µ ¯(a 4 h(a)) = ρ(a, B) for every a ∈ A and h(a) ∩ h(a0 ) = 0 0 whenever a ∩ a = 0. P (b) If A is a partition of unity in A, then H(A|B) ≤ a∈A q(ρ(a, B)), where q is the function of 384A. (c) If B is atomless and hai ii∈N is a partition of unity in A, then there is a partition of unity hbi ii∈N in B such that µ ¯bi = µ ¯ai and µ ¯(bi 4 ai ) ≤ 2ρ(ai , B) for every i ∈ N. proof (a) Let P : L1µ¯ → L1µ¯ be the conditional expectation operator associated with B. For any c ∈ B, Z Z Z Z Z |P (χa) − χc| = P (χa) + µ ¯c − P (χa) = χa + µ ¯c − χa 1\c

c

1\c

c

=µ ¯(a \ c) + µ ¯c − µ ¯(a ∩ c) = µ ¯(a 4 c). If a ∈ A set h(a) = [[P (χa) > 12 ]]. Then |P (χa) − χh(a)| ≤ |P (χa) − χc| for any c ∈ B, so Z ρ(a, B) = inf µ ¯(a 4 c) = inf |P (χa) − χc| c∈B c∈B Z = |P (χa) − χh(a)| = µ ¯(a 4 h(a)). If a ∩ a0 = 0, then P (χa) + P (χa0 ) = P χ(a ∪ a0 ) ≤ χ1, so h(a) ∩ h(a0 ) = [[P (χa) > 12 ]] ∩ [[P (χa0 ) > 21 ]] ⊆ [[P (χa) + P (χa0 ) > 1]] = 0, by 364D(b-i). (b) By 384Ae, q(1−t) ≤ q(t) whenever 0 ≤ t ≤ 12 . Consequently q(t) ≤ q(min(t, 1−t)) for every t ∈ [0, 1], and q¯(u) ≤ q¯(u ∧ (χ1 − u)) whenever u ∈ L0 (A) and 0 ≤ u ≤ χ1. Fix a ∈ A for the moment. We have q¯(P (χa)) ≤ q¯(P (χa) ∧ (χ1 − P (χa)) = q¯(|P (χa) − χh(a)|). Consequently Z

Z q¯(P χa) ≤

q¯(|P (χa) − χh(a)|) ≤ q

¡

Z

¢ |P (χa) − χh(a)|

(because q is concave) = q(ρ(a, B)). Summing over a, H(A|B) =

R

P a∈A

q¯(P χa) ≤

P a∈A

q(ρ(a, B)).

(c) Set b0i = h(ai ) for each i ∈ N. Then hb0i ii∈N is dist. Next, for each i ∈ N, take b00i ∈ B such that b0i and µ ¯b00i = min(¯ µai , µ ¯b00i ); then hb00i ii∈N is dist and µ ¯b00i ≤ µ ¯ai for every i. We can therefore find a 00 partition of unity hbi ii∈N such that bi ⊇ bi and µ ¯ bi = µ ¯ai for every i. (Use 331C to choose hdi ii∈N inductively so that di ⊆ 1 \ (supj µ ¯ai , then

b00i ⊆

µ ¯(ai 4 bi ) = µ ¯(ai 4 b00i ) ≤ µ ¯(ai 4 b0i ) + µ ¯(b0i 4 b00i ) =µ ¯(ai 4 b0i ) + µ ¯b0i − µ ¯ai ≤ 2¯ µ(ai 4 b0i ) = 2ρ(ai , B). If µ ¯b0i ≤ µ ¯ai , then µ ¯(ai 4 bi ) ≤ µ ¯(ai 4 b0i ) + µ ¯(b0i 4 bi ) =µ ¯(ai 4 b0i ) + µ ¯ai − µ ¯b0i ≤ 2¯ µ(ai 4 b0i ) = 2ρ(ai , B).

526

Automorphism groups

385P

385P Lemma Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving automorphism. Suppose that B ⊆ A. For k ∈ N, let Bk be the closed subalgebra of A generated by {π j b : b ∈ B, |j| ≤ k}, and let B be the closed subalgebra of A generated by {π j b : b ∈ B, j ∈ Z}. Then S (a) B is the topological closure k∈N Bk . (b) π[B] = B. (c) If C is any closed subalgebra of A such that π[C] = C, and a ∈ Bk , then P ρ(a, C) ≤ (2k + 1) b∈B ρ(b, C). proof (a) Because hBk ik∈N is non-decreasing, and must be B.

S k∈N

Bk is a subalgebra of A, so its closure also is (323J),

(b) Of course π −1 [Bk+1 ] is a closed subalgebra of A containing π j b whenever |j| ≤ k and b ∈ B, so includes Bk ; thus π[Bk ] ⊆ Bk+1 ⊆ B for every k, and S S π[B] = π[ k∈N Bk ] ⊆ k∈N π[Bk ] ⊆ B ⊆ B because π is continuous (324Kb). Similarly, π −1 [B] ⊆ B and π[B] = B. (c) For each b ∈ B, choose cb ∈ C such that µ ¯(b 4 cb ) = ρ(cb , C) (385Oa). Set e = sup|j|≤k supb∈B π j (b 4 cb ); then µ ¯e ≤ (2k + 1)

P b∈B

µ ¯(b 4 cb ) = (2k + 1)

P b∈B

ρ(b, C).

Now B0 = {d : d ∈ A, ∃ c ∈ C such that d \ e = c \ e} is a subalgebra of A. By 314Fa, applied to the order-continuous homomorphism c 7→ c \ e : C → A1\e , {c \ e : c ∈ C} is an order-closed subalgebra of the principal ideal A1\e ; by 313Id, applied to the ordercontinuous function d 7→ d \ e : A → A1\e , B0 is order-closed. If b ∈ B and |j| ≤ k, then π j b 4 π j cb ⊆ e, so π j b ∈ B0 ; accordingly B0 ⊇ Bk . Now a ∈ Bk , so there is a c ∈ C such that a 4 c ⊆ e, and P ρ(a, C) ≤ µ ¯(a 4 c) ≤ µ ¯e ≤ (2k + 1) b∈B ρ(b, C), as claimed. 385Q Lemma Let (A, µ ¯) be a probability algebra and suppose either that A is not purely atomic or that it is purely atomic and H(D0 ) = ∞, where D0 is the set of atoms of A. Then whenever A ⊆ A is a partition of unity and H(A) ≤ γ ≤ ∞, there is a partition of unity B, refining A, such that H(B) = γ. proof (a) By 384J, there is a partition of unity D1 such that H(D1 ) = ∞. Set D = D1 ∨ A; then we still have H(D) = ∞. Enumerate D as hdi ii∈N . Choose hBk ik∈N inductively, as follows. B0 = A. Given that Bk is a partition of unity, then if H(Bk ∨ {dk , 1 \ dk }) ≤ γ, set Bk+1 = Bk ∨ {dk , 1 \ dk }; otherwise set Bk+1 = Bk . S Let B be the closed subalgebra of A generated by k∈N Bk . Note that, for each d ∈ D, {c : c ∈ A, d ⊆ c or d ∩ c = 0} is a closed subalgebra of A including every Bk , so includes B. If b ∈ B \ {0}, there is surely some d ∈ D such that b ∩ d 6= 0, so b ⊇ d; thus B must be purely atomic. Let B be the set of atoms of B. Because A = B0 ⊆ B, B refines A. (b) H(B) ≤ γ. P P For each k ∈ N, let Bk be the closed subalgebra of A generated by Bk , so that S B = k∈N Bk . Suppose that b0 , . . . , bn are distinct of B. Then for each k ∈ N we can find dist b0k , . . . , bnk ∈ Bk such that µ ¯(bik 4 bi ) ≤ ρ(bi , Bk ) for every i ≤ n (385Oa). Accordingly µ ¯bi = limk→∞ µ ¯bik for each i, and Pn Pn µbi ) = limk→∞ i=0 q(¯ µbik ) ≤ supk∈N H(Bk ) ≤ γ. i=0 q(¯ As b0 , . . . , bn are arbitrary, H(B) ≤ γ. Q Q

385Yd

More about entropy

527

(c) H(B) ≥ γ. P P?? Suppose otherwise. We know that limk→∞ H({dk , 1 \ dk }) = limk→∞ q(¯ µdk ) + q(1 − µ ¯dk ) = 0. Let m ∈ N be such that H(B) + H({dk , 1 \ dk }) ≤ γ for every k ≥ m. Because B refines Bk , we must have H(Bk ∨ {dk , 1 \ dk }) ≤ H(Bk ) + H({dk , 1 \ dk }) ≤ γ, so that Bk+1 = Bk ∨ {dk , 1 \ dk } for every k ≥ m. But this means that dk ∈ B for every k ≥ m, so that P∞ γ > H(B) ≥ k=m q(¯ µdk ) = ∞, which is impossible. X XQ Q Thus B has the required properties. 385X Basic exercises > (a) Let A be a Boolean algebra, not {0}, and π : A → A an automorphism; set C = {c : πc = c}. Show that π is periodic, with period n ≥ 1, iff π¹ Ac has order n in the group Aut Ac whenever c ∈ C \ {0}. Show that π is aperiodic iff π¹ Ac has infinite order in the group Aut Ac whenever c ∈ C \ {0}. (b) In 385C, show that the family hci i1≤i≤ω is uniquely determined. > (c) Show that for a Bernoulli shift π, the Shannon-McMillan-Breiman theorem is a special case of the Pn−1 Ergodic Theorem. (Hint: wn = n1 i=0 T i w1 .) (d) Let (A, µ ¯) be a totally finite measure algebra, hBk ik∈N a non-decreasing sequence of subsets of A such that 0 ∈ B0 , and hci ii∈I a partition of unity in A. Show that P P limk→∞ i∈I ρ(ci , Bk ) = i∈I ρ(ci , B) S where B = k∈N Bk . > (e) Let (A, µ ¯) be a probability algebra, π : A → A a measure-preserving Boolean homomorphism and A a partition of unity in A. Show that h(π, Dn (A, π)) = h(π, A) = h(π, π[A]) for any n ≥ 1. (f ) Let (A, µ ¯) be a totally finite measure algebra and π : A → A a measure-preserving Boolean homomorphism. Suppose that B ⊆ A. For k ∈ N, let Bk be the closed subalgebra of A generated by {π j b : b ∈ B, j ≤ k}, and let B be the closed subalgebra of A generated by {π j b : b ∈ B, j ∈ N}. Show that S B = k∈N Bk , π[B] ⊆ B, P and that if C is any closed subalgebra of A such that π[C] ⊆ C, and a ∈ Bk , then ρ(a, C) ≤ (k+1) b∈B ρ(b, C). 385Y Further exercises (a) Give an example to show that the word ‘injective’ in the statement of 385C is essential. (b) Let (A, µ ¯) be a totally finite measure algebra and π : A → A an aperiodic measure-preserving Boolean homomorphism. Set C = {c : πc = c}. Show that whenever n ≥ 1, 0 ≤ γ < n1 and B ⊆ A is finite, there is an a ∈ A such that a, πa, π 2 a, . . . , π n−1 a are dist and µ ¯(a ∩ b ∩ c) = γ µ ¯(b ∩ c) for every b ∈ B, c ∈ C. (c) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism. Let P be the set of all closed subalgebras of A which are invariant under π, ordered by inclusion. Show that B 7→ h(π¹ B) : P → [0, ∞] is order-preserving and order-continuous on the left, in the sense that if Q ⊆ P is non-empty and upwards-directed then h(π¹ sup Q) = supB∈Q h(π¹ B). (d) Let (A, µ ¯) be a probability algebra, and π : A → A a measure-preserving Boolean homomorphism of finite entropy. Let P be the set of all π-invariant closed subalgebras of A. Show that B 7→ h(π¹ B) : P → [0, ∞[ is order-continuous. (Hint: if Q ⊆ P is non-empty and downwards-directed, then for any partition of unity A ⊆ A, H(A| inf Q) = supB∈Q H(A|B).)

528

Automorphism groups

385 Notes

385 Notes and comments I have taken the trouble to give sharp forms of the Halmos-Rokhlin-Kakutani lemma (385E) and the Cziszár-Kullback inequality (385J); while it is possible to get through the principal results of the next two sections with rather less, the formulae become better focused if we have the exact expressions available. Of course one can always go farther still (385Yb). Ornstein’s theorem in §386 (though not Sinaˇı’s, as stated there) can be deduced from the special case of the Shannon-McMillan-Breiman theorem (385G) in which the homomorphism π is a Bernoulli shift, which can be deduced from the Ergodic Theorem (385Xc). Lemma 385D is the starting point of the theory of ‘recurrence’; the next steps are in 382Yc-382Yd and 387E-387F.

386 Ornstein’s theorem I come now to the most important of the handful of theorems known which enable us to describe automorphisms of measure algebras up to isomorphism: two two-sided Bernoulli shifts (on algebras of countable Maharam type) of the same entropy are isomorphic (386I, 386K). This is hard work. It requires both delicate ²-δ analysis and substantial skill with the manipulation of measure-preserving homomorphisms. The proof is based on two difficult lemmas (386C and 386F), and includes Sinaˇı’s theorem (386E, 386L), describing the Bernoulli shifts which arise as factors of a given ergodic automorphism. 386A

The following definitions offer a language in which to express the ideas of this section.

Definitions Let (A, µ ¯) be a probability algebra and π : A → A a measure-preserving Boolean homomorphism. (a) A Bernoulli partition for π is a partition of unity hai ii∈I such that Qk ¯ai(j) µ ¯(inf j≤k π j ai(j) ) = j=0 µ whenever i(0), . . . , i(k) ∈ I. (b) If π is an automorphism, a Bernoulli partition hai ii∈I for π is (two-sidedly) generating if the closed subalgebra generated by {π j ai : i ∈ I, j ∈ Z} is A itself. (c) A factor of (A, µ ¯, π) is a triple (B, µ ¯¹ B, π¹ B) where B is a closed subalgebra of A such that π[B] = B. 386B Remarks Let (A, µ ¯) be a probability algebra, π : A → A a measure-preserving Boolean homomorphism and hai ii∈I a Bernoulli partition for π. (a) hπ k [A0 ]ik∈N is independent, where A0 is the closed subalgebra of A generated by {ai : i ∈ I}. P P Suppose that cj ∈ π j [A0 ] for j ≤ k. Then each π −j cj ∈ A0 is expressible as supi∈Ij ai for some Ij ⊆ I. Now µ ¯( inf cj ) = µ ¯( j≤k

inf π j aij )

sup

i0 ∈I0 ,... ,ik ∈Ik j≤k

=

X i0 ∈I0 ,... ,ik ∈Ik

=

k X Y j=0 i∈Ij

X

µ ¯( inf π j aij ) =

µ ¯ai =

j≤k

k Y j=0

k Y

µ ¯aij

i0 ∈I0 ,... ,ik ∈Ik j=0

µ ¯(sup ai ) = i∈Ij

k Y

µ ¯cj .

j=0

As c0 , . . . , ck are arbitrary, hπ k [A0 ]ik∈N is independent. Q Q (b) If π is an automorphism, then hπ k [A0 ]ik∈Z is independent, by 384Sf. (c) Setting A = {ai : i ∈ I} \ {0}, we have h(π, A) = H(A), as in part (a) of the proof of 384R, so h(π) ≥ H(A).

386C

Ornstein’s theorem

529

(d) If H(A) > 0, then A is atomless. P P As A contains at least two elements of non-zero measure, γ = maxa∈A µ ¯a < 1. Because hai ii∈I is a Bernoulli partition, every member of Dk (A, π) has measure at most γ k , for any k ∈ N. Thus any atom of A could have measure at most inf k∈N γ k = 0. Q Q (e) If B is any closed subalgebra of A such that π[B] ⊆ B, then h(π¹ B) ≤ h(π), just because h(π¹ B) is calculated from the action of π on a smaller set of partitions. If C+ is the closed subalgebra of A generated by {π j ai : i ∈ I, j ∈ N}, then π[C+ ] ⊆ C+ (compare 385Pb), and π¹ C+ is a one-sided Bernoulli shift with root algebra A0 and entropy H(A), so that H(A) = h(π¹ C+ ) by the Kolmogorov-Sinaˇı theorem (384P, 384R). (f ) If π is an automorphism, and C is the closed subalgebra of A generated by {π j ai : i ∈ I, j ∈ Z}, then π[C] = C (385Pb) and π¹ C is a two-sided Bernoulli shift with root algebra A0 . (g) Thus every Bernoulli partition for π gives rise to a factor of (A, µ ¯, π) which is a one-sided Bernoulli shift, and if π is an automorphism we can extend this to the corresponding two-sided Bernoulli shift. If π has a generating Bernoulli partition then it is itself a Bernoulli shift. (h) Now suppose that (B, ν¯) is another probability algebra, φ : B → B is a measure-preserving Boolean homomorphism, and hbi ii∈I is a Bernoulli partition for φ such that ν¯bi = µ ¯ai for every i. We have a unique measure-preserving Boolean homomorphism θ+ : C+ → B such that θ+ (π j ai ) = φj bi for every i ∈ I, j ∈ N. (Apply 324P.) Now θ+ π = φθ+ . (The set {a : θ+ πa = φθ+ a} is a closed subalgebra of C+ containing every π j ai .) (i) If, in (g) above, π and φ are both automorphisms, then the same arguments show that we have a unique measure-preserving Boolean homomorphism θ : C → B such that θai = bi for every i ∈ I and θπ = φθ. 386C Lemma Let (A, µ ¯) be an atomless probability algebra and π : A → A an ergodic measurepreserving automorphism. Let hai ii∈N be a partition of unity in A, of finite entropy, and hγi ii∈N a sequence of non-negative real numbers such that P∞ P∞ i=0 q(γi ) ≤ h(π), i=0 γi = 1, where q is the function of 384A. Then for any ² > 0 we can find a partition ha0i ii∈N of unity in A such that (i) {i : a0i 6= 0} is finite, P∞ (ii) i=0 |γi − µ ¯a0i | ≤ ², qP p P∞ ∞ ¯(a0i 4 ai ) ≤ ² + 6 (iii) i=0 µ µai − γi | + 2(H(A) − h(π, A)) i=0 |¯ where A = {ai : i ∈ N} \ {0}, (iv) H(A0 ) ≤ h(π, A0 ) + ² where A0 = {a0i : i ∈ N} \ {0}. p proof (a) Of course h(π, A) ≤ H(A), by 384Ma, so the square root 2(H(A) − h(π, A)) gives no difficulty. qP p ∞ 1 ²). Set β = µai − γi | + 2(H(A) − h(π, A)), δ = min( 14 , 24 i=0 |¯ P∞ There is a sequence h¯ γi ii∈N of non-negative real numbers such that {i : γ¯i > 0} is finite, i=0 γ¯i = 1, P∞ P P ∞ ∞ 2 |¯ γi −γi | ≤ 2δ 2 and i=0 q(¯ γi ) ≤ h(π). P P Take i=0P P∞k ∈ N such that i=k γi ≤ δ , and set γ¯i = γi for i < k, ∞ γ¯k = i=k γi and γ¯i = 0 for i > k; then q(¯ γk ) ≤ i=k q(γi ) (384Ab), so P∞ P∞ γi ) ≤ i=0 q(γi ) ≤ h(π), i=0 q(¯ while

P∞ i=0

|¯ γi − γi | ≤ γ¯k +

P∞ i=k

γi ≤ 2δ 2 . Q Q

P∞ P∞ Because i=0 q(¯ γi ) is finite, there is a partition of unity C in A, of finite entropy, such that i=0 q(¯ γi ) ≤ h(π, C) + 3δ; replacing C by C ∨ A if need be (note that C ∨ A still has finite entropy, by 384H), we may suppose that C refines A.

530

Automorphism groups

386C

There is a sequence hγi0 ii∈N of non-negative real numbers such that P∞ 0 2 i=0 |γi − γi | ≤ 4δ and P∞ 0 i=0 q(γi ) = h(π, C) + 3δ.

P∞ i=0

γi0 = 1, {i : γi0 > 0} is finite,

P P Take k ∈ N such that γ¯i = 0 for i > k. Take r ≥ 1 such that δ 2 ln( δr2 ) ≥ h(π, C) + 3δ and set γ˜i = (1 − δ 2 )¯ γi for i ≤ k, 1 r

= δ 2 for k + 1 ≤ i ≤ k + r, = 0 for i > k + r. Then

P∞ i=0

Pk+r i=0

|˜ γi − γ¯i | = 2δ 2 ,

P∞ i=0

|˜ γi − γi | ≤ 4δ 2 , δ2 r

r δ

q(¯ γi ) ≤ h(π, C) + 3δ ≤ δ 2 ln( 2 ) = rq( ) ≤

Now the function α 7→

Pk+r i=0

Pk+r i=0

q(˜ γi ).

q(α¯ γi + (1 − α)˜ γi ) : [0, 1] → R

is continuous, so there is some α ∈ [0, 1] such that Pk+r γi + (1 − α)˜ γi ) = h(π, C) + 3δ, i=0 q(α¯ γi + (1 − α)˜ γi for every i; of course and we can set γi0 = α¯ P∞ P∞ P∞ 0 γi − γi | ≤ 4δ 2 . Q Q γi − γi | + (1 − α) i=0 |˜ i=0 |¯ i=0 |γi − γi | ≤ α Set M = {i : γi0 6= 0}, so that M is finite. (b) Let η ∈ ]0, δ] be so small that δ

(i) |q(s) − q(t)| ≤ whenever s, t ∈ [0, 1] and |s − t| ≤ 3η, 1+#(M ) P µc, 2η)) ≤ δ, (ii) c∈C q(min(¯ 1 6

(iii) η ≤ . P µc) is finite.) (Actually, (iii) is a consequence of (i). For (ii) we must of course rely on the fact that c∈C q(¯ Let ν be the probability measure on M defined by saying that ν{i} = γi0 for every i ∈ M , and λ the product measure on M N . Define Xij : M N → {0, 1}, for i ∈ M and j ∈ N, and Yj : M N → R, for j ∈ N, by setting Xij (ω) = 1 if ω(j) = i, = 0 otherwise, 0 Yj (ω) = ln(γω(j) ) for everyω ∈ M N .

Then, for each i ∈ M , hXij ij∈N is an independent sequence of random variables, all with expectation γi0 , and hYj ij∈N is also an independent sequence of random variables, all with expectation P P∞ 0 0 0 i∈M γi ln γi = − i=0 q(γi ) = −h(π, C) − 3δ. Let n ≥ 1 be so large that (iv) µ ¯[[wn − h(π, C)χ1 ≥ δ]] < η, where 1 P 1 )χd; wn = d∈Dn (C,π) ln( n

µ ¯d

(v) ¡P ¢ 1 Pn−1 0 Pr j=0 Xij − γi | ≤ η ≥ 1 − δ, i∈M | n

386C


531

¡ 1 Pn−1 ¢ Pr | j=0 Yj + h(π, C) + 3δ| ≤ δ ≥ 1 − δ; n

(vi) enδ ≥ 2,

1 n+1

≤ η,

q(

1 n ) + q( ) n+1 n+1

≤ δ;

these will be true for all sufficiently large n, using the Shannon-McMillan-Breiman theorem (385H) for (iv) and the strong law of large numbers (in any of the forms 273D, 273H or 273I) for (v). (c) There is a family hbji ij
∞ X

X

|¯ µ(π k ai ∩ c) − γi0 µ ¯c|

i=0 c∈Dk (A,π)

≤

∞ X

X

|¯ µ(π k ai ∩ c) − µ ¯ ai · µ ¯c| +

i=0 c∈Dk (A,π)

≤

∞ X

|γi − γi0 | +

i=0

≤ 4δ 2 +

|¯ µai − γi0 |

i=0 ∞ X

|¯ µai − γi | +

i=0 ∞ X

∞ X

∞ X

X

X

µ ¯c

c∈Dk (A,π)

|¯ µ(π k ai ∩ c) − µ ¯ ai · µ ¯c|

i=0 c∈Dk (A,π)

q ¡ ¢ |¯ µai − γi | + 2 H(π k [A]) + H(Dk (A, π)) − H(Dk+1 (A, π))

i=0

(by 385K, because Dk+1 (A, π) = π k [A] ∧ Dk (A, π)) ∞ X p 2 ≤ 4δ + |¯ µai − γi | + 2(H(A) − h(π, A)) i=0

(because h(π, A) ≤ H(Dk+1 (A, π)) − H(Dk (A, π)), by 385Nc) = β 2 + 4δ 2 . Choose a partition of unity hbki ii∈M such that, for each c ∈ Dk (A, π), b ∈ Bk , i ∈ M , µ ¯(bki ∩ b ∩ c) = γi0 µ ¯(b ∩ c), if µ ¯(π k ai ∩ b ∩ c) ≥ γi0 µ ¯(b ∩ c) then bki ∩ b ∩ c ⊆ π k ai , if µ ¯(π k ai ∩ b ∩ c) ≤ γi0 µ ¯(b ∩ c) then π k ai ∩ b ∩ c ⊆ bki . (This is where I use the hypothesis that A is atomless.) Note that in these formulae we always have µ ¯(b ∩ c) = µ ¯b · µ ¯c, Consequently

µ ¯(π k ai ∩ b ∩ c) = µ ¯(π k ai ∩ c) · µ ¯b.

532

Automorphism groups

X

X

µ ¯(π k ai ∩ bki ) =

i∈M

X

X

386C

µ ¯(b ∩ c ∩ (π k ai ∩ bki ))

b∈Bk c∈Dk (A,π) i∈M

X

=

X

∞ X

min(¯ µ(b ∩ c ∩ π k ai ), γi0 µ ¯(b ∩ c))

b∈Bk c∈Dk (A,π) i=0

X

≥

X

∞ X

µ ¯(b ∩ c ∩ π k ai ) − |¯ µ(b ∩ c ∩ π k ai ) − γi0 µ ¯(b ∩ c)|


=1−

X

∞ X

X

|¯ µ(b ∩ c ∩ π k ai ) − γi0 µ ¯(b ∩ c)|


=1−

X

∞ X

X

µ ¯b · |¯ µ(c ∩ π k ai ) − γi0 µ ¯c|


=1−

X

∞ X

|¯ µ(c ∩ π k ai ) − γi0 µ ¯c| ≥ 1 − β 2 − 4δ 2 .

c∈Dk (A,π) i=0

Also we have µ ¯(bki ∩ b ∩ c) = γi0 µ ¯b · µ ¯c for every b ∈ Bk , c ∈ Dk (A, π) and i ∈ M , so the (downwards) induction proceeds. Q Q (d) Let B be the set of atoms of the algebra generated by {bji : j < n, i ∈ M }. For b ∈ B, d ∈ Dn (C, π) set Ibd = {j : j < n, ∃ i ∈ M, b ⊆ bji , d ⊆ π j ai }. Then, for any j < n, sup{b ∩ d : b ∈ B, d ∈ Dn (C, π), j ∈ Ibd } = supi∈M bji ∩ π j ai , because C refines A, so every π j ai is a supremum of of Dn (C, π). Accordingly X

#(Ibd )¯ µ(b ∩ d) =

b∈B,d∈Dn (C,π)

n−1 X

X

µ ¯(bji ∩ π j ai ) ≥ n(1 − β 2 − 4δ 2 ).

j=0 i∈M

Set e0 = sup{b ∩ d : b ∈ B, d ∈ Dn (C, π), #(Ibd ) ≥ n(1 − β − 4δ)}; then µ ¯e0 ≥ 1 − β − δ. (e) Let B 0 ⊆ B be the set of those b ∈ B such that P 1 0 µ ¯b ≤ e−n(h(π,C)+2δ) , i∈M |γi − #({j : j < n, b ⊆ bji })| ≤ η. n

0

Then µ ¯(sup B ) ≥ 1 − 2δ. P P Set B10 = {b : b ∈ B, µ ¯b ≤ e−n(h(π,C)+2δ) } = {b : b ∈ B, h(π, C) + 2δ +

1 n

ln(¯ µb) ≤ 0}

= { inf bi,i(j) : i(0), . . . , i(n − 1) ∈ M, h(π, C) + 2δ + j

Then

1 n

n−1 X j=0

0 ln γi(j) ≤ 0}.

386C


µ ¯(sup B10 ) = Pr(h(π, C) + 2δ +

1 n

≥ Pr(|h(π, C) + 3δ +

1 n

533

n−1 X

Yj ≤ 0)

j=0 n−1 X

Yj | ≤ δ) ≥ 1 − δ

j=0

by the choice of n. On the other hand, setting B20 = {b : b ∈ B,

X

1 n

|γi0 − #({j : j < n, b ⊆ bji })| ≤ η}

i∈M

X

= { inf bi,i(j) : i(0), . . . , i(n − 1) ∈ M, j

1 n

|γi0 − #({j : i(j) = i})| ≤ η},

i∈M

we have µ ¯(sup B20 ) = Pr(

P i∈M

|γi0 −

1 n

Pn−1 j=0

Xij | ≤ η) ≥ 1 − δ

by the other half of clause (b-v). Since B 0 = B10 ∩ B20 , µ ¯(sup B) ≥ 1 − 2δ. Q Q Let D00 be the set of those d ∈ Dn (C, π) such that 1 n

ln(

1 ) µ ¯d

≤ h(π, C) + δ,

i.e.,

µ ¯d ≥ e−n(h(π,C)+δ) ;

¯(sup D0 ) ≥ 1 − η. If d ∈ D0 , b ∈ B 0 by (b-iv), µ ¯(sup D00 ) > 1 − η. Let D0 ⊆ D00 be a finite set such that µ then µ ¯d ≥ e−n(h(π,C)+δ) ≥ enδ µ ¯b ≥ 2¯ µb. Since µ ¯(sup D0 ) ≤ 1 ≤ 2¯ µ(sup B 0 ) ( that δ ≤ 41 , #(D0 ) ≤ #(B 0 ). 0 Set e1 = e0 ∩ sup B , so that µ ¯e1 ≥ 1 − β − 3δ, and 1 2

D00 = {d : d ∈ D0 , µ ¯(d ∩ e1 ) ≥ µ ¯d}; then µ ¯(sup(D0 \ D00 )) ≤ 2¯ µ(1 \ e1 ) ≤ 2β + 6δ, so µ ¯(sup D00 ) ≥ 1 − 2β − 6δ − η ≥ 1 − 2β − 7δ. (f ) If d1 , . . . , dk ∈ D00 are distinct, µ ¯(sup1≤i≤k di ∩ e1 ) ≥

k 2

inf i≤k µ ¯di ≥ k supb∈B 0 µ ¯b,

and #({b : b ∈ B 0 , b ∩ e0 ∩ sup1≤i≤k di } 6= 0) ≥ k. By the Marriage Lemma (3A1K), there is an injective function f0 : D00 → B 0 such that d ∩ f0 (d) ∩ e0 6= 0 for every d ∈ D00 . Because #(D0 ) ≤ #(B 0 ), we can extend f0 to an injective function f : D0 → B 0 . (g) By the Halmos-Rokhlin-Kakutani lemma, in the strong form 385E(iv), there is an a ∈ A such that 1 µ ¯d for every d ∈ D0 ∪ {1}. Set e = sup{π −j (a ∩ d) : j < a, π −1 a, . . . , π −n+1 a are dist and µ ¯(a ∩ d) = n+1 0 −j n, d ∈ D }. Because hπ (a ∩ d)ij
(h) For i ∈ M , set a0i = sup{π −j (a ∩ d) : j < n, d ∈ D0 , f (d) ⊆ bji }.

534

Automorphism groups

386C

Then the a0i are dist. P P Suppose that i, i0 ∈ M are distinct. If j, j 0 < n and d, d0 ∈ D0 and f (d) ⊆ bji , 0 f (d ) ⊆ bj 0 i0 , then either j 6= j 0 or j = j 0 . In the former case, 0

0

π −j (a ∩ d) ∩ π −j (a ∩ d0 ) ⊆ π −j a ∩ π −j a = 0. In the latter case, bji ∩ bj 0 i0 = 0, so f (d) 6= f (d0 ) and d 6= d0 and 0

π −j (a ∩ d) ∩ π −j (a ∩ d0 ) ⊆ π −j (d ∩ d0 ) = 0. Q Q Observe that supi∈M a0i = supj
|¯ µa0i − γi0 | ≤

i∈M

X

γi0 |1 − n¯ µ(a ∩ sup D0 )| +

i∈M

X

|¯ µa0i − nγi0 µ ¯(a ∩ sup D0 )|

i∈M

n µ ¯(sup D0 ) n+1

≤1− +

X n−1 X | i∈M

j=0

X

µ ¯(π −j (a ∩ d)) − nγi0

d∈D 0 f (d) ⊆ bji

X

µ ¯(a ∩ d)|

d∈D 0

≤ 1 − (1 − η)2 X X + |¯ µ(a ∩ d) · #({j : j < n, f (d) ⊆ bji }) − nγi0 µ ¯(a ∩ d)| d∈D 0 i∈M

≤ 1 − (1 − η)2 +

X

µ ¯(a ∩ d)nη

d∈D 0

(see the definition of B 0 in (d) above) ≤ 2η + nη µ ¯a ≤ 3η. So ∞ X

¯a0m + |¯ µa0i − γi | ≤ µ

i=0

X

|¯ µa0i − γi0 | +

∞ X

i∈M

|γi0 − γi |

i=0

≤ 2η + 3η + 4δ 2 ≤ 6δ ≤ ². We shall later want to know that |¯ µa0i − γi0 | ≤ 3η for every i; for i ∈ M this is covered by the formulae 0 above, for i = m it is true because µ ¯ am = 1 − µ ¯e ≤ 2η (see (g)), and for other i it is trivial. P∞ (i) The next step is to show that i=0 µ ¯(a0i ∩ ai ) ≥ 1 − 3β − 12δ. P P It is enough to consider the case in which 3β + 12δ < 1. We know that sup a0i ∩ ai ⊇ sup{π −j (a ∩ d) : j < n, d ∈ D0 , i∈N

∃ i ∈ M such that f (d) ⊆ bji and d ⊆ π j ai } = sup{π −j (a ∩ d) : d ∈ D0 , j ∈ If (d),d } P (see (d) for the definition of Ibd ) has measure at least d∈D0 #(If (d),d )¯ µ(a ∩ d). For any d ∈ D00 , we arranged that d ∩ f (d) ∩ e0 6= 0. This means that there must be some b ∈ B and d0 ∈ Dn (C, π) such that d ∩ f (d) ∩ b ∩ f (d0 ) 6= 0 and #(Ibd0 ) ≥ n(1 − β − 4δ); of course b = f (d) and d0 = d, so that #(If (d),d ) must be at least n(1 − β − 4δ). Accordingly ∞ X i=0

µ ¯(a0i ∩ ai ) ≥

X d∈D 00

n(1 − β − 4δ)¯ µ(a ∩ d) = n(1 − β − 4δ)

1 µ ¯(sup D00 ) n+1

≥ (1 − η)(1 − β − 4δ)(1 − 2β − 7δ) ≥ 1 − 3β − 12δ. Q Q

386C


But this means that

P∞ i=0

µ ¯(a0i 4 ai ) = 2(1 −

P∞ i=0

535

µ ¯(a0i ∩ ai )) ≤ 6β + 24δ ≤ ² + 6β

(using 385L for the equality). (j) Finally, we need to estimate H(A0 ) and h(π, A0 ), where A0 = {a0i : i ∈ N} \ {0}. For the former, we have H(A0 ) ≤ h(π, C) + 4δ. P P |¯ µa0i − γi0 | ≤ 3η for every i, by (h) above. So by (b-i), P P∞ H(A0 ) = i∈M ∪{m} q(¯ µa0i ) ≤ δ + i=0 q(γi0 ) = h(π, C) + 4δ. Q Q (k) Consider the partition of unity A00 = A0 ∨ {a, 1 \ a}. Let D be the closed subalgebra of A generated by {π j c : j ∈ Z, c ∈ A00 }. (i) a ∩ d ∈ D for every d ∈ D0 . P P Of course a ∩ e ∈ D, because 1 \ e = a0m . If d0 ∈ D0 and d0 6= d, then 0 (because f is injective) f (d) 6= f (d ); there must therefore be some k < n and distinct i, i0 ∈ M such that f (d) ⊆ bki and f (d0 ) ⊆ bki0 . But this means that π −k (a ∩ d) ⊆ a0i and π −k (a ∩ d0 ) ⊆ a0i0 , so that a ∩ d ⊆ π k a0i and a ∩ d0 ∩ π k a0i = 0. What this means is that if we set d˜ = a ∩ e ∩ inf{π k a0 : k < n, i ∈ M, a ∩ d ⊆ π k a0 }, i

i

a0i

we get a member of D (because every ∈ D, and π[D] = D) including a ∩ d and dist from a ∩ d0 whenever d0 ∈ D0 and d0 6= d. But as a ∩ π −j a = 0 if 0 < j < n, a ∩ e must be sup{a ∩ d0 : d0 ∈ D0 }, and a ∩ d = d˜ belongs to D. Q Q (ii) Consequently c ∩ e ∈ D for every c ∈ C. P P We have c ∩ e = sup{c ∩ π −j (a ∩ d) : j < n, d ∈ D0 } = sup{π −j (π j c ∩ a ∩ d) : j < n, d ∈ D0 } = sup{π −j (a ∩ d) : j < n, d ∈ D0 , d ⊆ π j c} (because if d ∈ D and j < n then either d ⊆ π j c or d ∩ π j c = 0) ∈D 0

because a ∩ d ∈ D for every d ∈ D0 and π −1 [D] = D. Q Q (iii) It follows that h(π, A00 ) ≥ h(π, C) − δ. P P For any c ∈ C, 1 3

ρ(c, D) ≤ µ ¯(c 4 (c ∩ e)) = µ ¯(c \ e) ≤ min(¯ µc, 2η) ≤ . So h(π, C) ≤ h(π¹ D) + H(C|D) (385Nd, because π[D] = D) ≤ h(π, A00 ) +

X

q(ρ(c, D))

c∈C

(by the Kolmogorov-Sinaˇı theorem (384P) and 385Ob) ≤ h(π, A00 ) +

X

q(min(¯ µc, 2η))

c∈C

(because q is monotonic on [0, 31 ]) ≤ h(π, A00 ) + δ by the choice of η. Q Q (iv) Finally, h(π, A0 ) ≥ h(π, C) − 2δ. P P Using 385Nb,

536

Automorphism groups

386C

h(π, C) − δ ≤ h(π, A00 ) ≤ h(π, A0 ) + H({a, 1 \ a}) = h(π, A0 ) + q(¯ µa) + q(1 − µ ¯a) = h(π, A0 ) + q(

1 n ) + q( ) n+1 n+1

≤ h(π, A0 ) + δ

by the choice of n. Q Q (l) Putting these together, H(A0 ) ≤ h(π, C) + 4δ ≤ h(π, A0 ) + 6δ ≤ h(π, A0 ) + ², and the proof is complete.

386D Corollary Let (A, µ ¯) be an atomless probability algebra and π : A → A an ergodic measurepreserving automorphism. Let hai ii∈N be a partition of unity in A, of finite entropy, and hγi ii∈N a sequence of non-negative real numbers such that P∞ P∞ i=0 γi = 1, i=0 q(γi ) ≤ h(π). ¯a∗i = γi for every i ∈ N and Then for any ² > 0 we can find a Bernoulli partition ha∗i ii∈N for π such that µ q p P∞ P∞ ¯(a∗i 4 ai ) ≤ ² + 6 µai − γi | + 2(H(A) − h(π, A)), i=0 µ i=0 |¯ writing A = {ai : i ∈ N} \ {0}. qP p ∞ proof (a) Set β = µai − γi | + 2(H(A) − h(π, A)). Let h²n in∈N be a sequence of strictly positive i=0 |¯ real numbers such that p √ P∞ 2²n ≤ ². n=0 ²n + 6 ²n + Using 386C, we can choose inductively, for n ∈ N, partitions of unity hani ii∈N such that, for each n ∈ N, P∞ ¯ani | ≤ ²n , i=0 |γi − µ H(An ) ≤ h(π, An ) + ²n < ∞ (writing An = {ani : i ∈ N} \ {0}), P∞ i=0

while

µ ¯(an+1,i 4 ani ) ≤ ²n+1 + 6 P∞ i=0

p √ ²n + 2²n ,

µ ¯(a0i 4 ai ) ≤ ²0 + 6β.

On completing the induction, we see that √ P∞ P∞ P∞ P∞ p ¯(an+1,i 4 ani ) ≤ n=1 ²n + n=0 6 ²n + 2²n < ∞. n=0 i=0 µ P∞ In particular, given i ∈ N, n=0 µ ¯(an+1,i 4 ani ) is finite, so hani in∈N is a Cauchy sequence in the complete metric space A (323Gc), and has a limit a∗i , with µ ¯a∗i = limn→∞ µ ¯ani = γi (323C). If i 6= j, a∗i ∩ a∗j = limn→∞ ani ∩ anj = 0 (using 323B), so ha∗i ii∈N is dist; since

P∞ i=0

ha∗i ii∈N is a partition of unity. We also have

µ ¯a∗i =

P∞ i=0

γi = 1,

386F


∞ X

µ ¯(a∗i 4 ai ) ≤

∞ X

µ ¯(a0i 4 ai ) +

i=0

i=0

≤ ²0 + 6β +

∞ X

∞ X ∞ X

537

µ ¯(an+1,i 4 ani )

n=0 i=0 ∞ q X

²n +

n=1

6 ²n +

√

2²n ≤ ² + 6β.

n=0

(b) Now take any i(0), . . . , i(k) ∈ N. For each j < k, n ∈ N, H(π j [An ]) + H(Dj (An , π)) − H(Dj+1 (An , π)) ≤ H(An ) − h(π, An ) ≤ ²n (using 385Nc). But this means that ∞ X

X

|¯ µ(d ∩ π j ani ) − µ ¯d · µ ¯ani | ≤

√

2²n ,

d∈Dj (An ,π) i=0

by 385K. A fortiori, |¯ µ(d ∩ π j ani ) − µ ¯d · µ ¯ani | ≤

√

2²n

for every d ∈ Dj (An , π), i ∈ N. Inducing on r, we see that √ Qr |¯ µ(inf j≤r π j an,i(j) ) − j=0 µ ¯an,i(j) | ≤ r 2²n → 0 as n → ∞, for any r ≤ k. Because µ ¯, ∩ and π are all continuous (323C, 323B, 324Kb), µ ¯( inf π j a∗i(j) ) = lim µ ¯( inf π j an,i(j) ) n→∞ j≤k j≤k Y Y γi(j) . = lim µ ¯an,i(j) = n→∞

As i(0), . . . , i(k) are arbitrary,

ha∗i ii∈N

j≤k

j≤k

is a Bernoulli partition for π.

386E Sinaˇı’s theorem (atomic case) (Sinaˇı 62) Let (A, µ ¯) be an atomless probability algebra and π : A → A an ergodic measure-preserving automorphism. Let hγ i ii∈N be a sequence of non-negative real P∞ P∞ numbers such that i=0 γi = 1 and i=0 q(γi ) ≤ h(π). Then there is a Bernoulli partition ha∗i ii∈N for π such that µ ¯a∗i = γi for every i ∈ N. proof Apply 386D from any starting point, e.g., a0 = 1, ai = 0 for i > 0. 386F Lemma Let (A, µ ¯) be an atomless probability algebra and π a measure-preserving automorphism of A. Let hbi ii∈N , hci ii∈N be Bernoulli partitions for π, of the same finite entropy, and write B, C for the closed subalgebras of A generated by {π j bi : i ∈ N, j ∈ Z} and {π j ci : i ∈ N, j ∈ Z}. Suppose that C ⊆ B. Then for any ² > 0 we can find a Bernoulli partition hdi ii∈N for π such that (i) di ∈ C for every i ∈ N, (ii) µ ¯di = µ ¯bi for every i ∈ N, (iii) µ ¯(φci 4 ci ) ≤ ² for every i ∈ N, where φ : B → C is the measure-preserving Boolean homomorphism such that φbi = di for every i and πφ = φπ (386Bi). proof (a) Set B = {bi : i ∈ N} \ {0}, C = {ci : i ∈ N} \ {0}. If only one ci is non-zero, then H(C) = 0, so H(B) = 0 and B = {0, 1}, in which case B = C and we take di = bi and stop. Otherwise, C is atomless (386Bd). For k ∈ N, let Bk ⊆ B be the closed subalgebra of A generated by {π j bi : i ≤ k, |j| ≤ k}. Because C ⊆ B, there is an m ∈ N such that 1 4

ρ(ci , Bm ) ≤ ² for every i ∈ N (385M). Let η, ξ > 0 be such that

538

Automorphism groups

386F

P∞ √ ² ² 1 η + 6 4 2η ≤ , ξ ≤ min( , ), ¯ci )) ≤ η. i=0 q(min(2ξ, µ 4(2m+1) 4 6 P∞ (The last is achievable because i=0 q(¯ µci ) is finite.) Let r ≥ m be such that ρ(ci , Br ) ≤ ξ for every i ∈ N. Let n ≥ r be such that 2r+1 2n+2

≤ ξ,

µ ¯ci ≤ ξ for every i > n.

(b) Let hb0i ii∈N be a partition of unity in C such that µ ¯b0i = µ ¯bi for every i ∈ N. Let U be the set of j atoms of the subalgebra of B generated by {π bi : i ≤ n, |j| ≤ n} ∪ {π j ci : i ≤ n, |j| ≤ n}, and V the set of atoms of the subalgebra of C generated by {π j b0i : i ≤ n, |j| ≤ n} ∪ {π j ci : i ≤ n, |j| ≤ n}. For each v ∈ V , choose a dist family hdvu iu∈U in C such that supu∈U dvu = v and µ ¯dvu = µ ¯(v ∩ u) for every u ∈ U . By 385E(iv), there is an a ∈ C such that a, πa, . . . , π 2n a are dist and µ ¯(a ∩ dvu ) =

1 µ ¯(dvu ) 2n+2

for every

u ∈ U and v ∈ V . (π¹ C is a Bernoulli shift, therefore ergodic, by 384Se, therefore aperiodic, by 385F.) Set e = sup|j|≤n π j a, e˜ = sup|j|≤n−r π j a; then µ ¯e˜ = (2(n − r) + 1)¯ µa = 1 −

2r+1 . 2n+2

Let Ce˜ be the principal ideal of C generated by e˜. (c) The family hπ −j (a ∩ dvu )i|j|≤n,u∈U,v∈V is dist. hdvu iu∈U,v∈V and

P P All we have to note is that the families

hπ −j ai|j|≤n = hπ −n (π n+j a)i|j|≤n are dist. Q Q Consequently, if we set −j ˆbi = sup (a ∩ dvu ) ∈ C |j|≤n supv∈V supu∈U,u ⊆ π j bi π for i ∈ N, hˆbi ii∈N is dist, since a given triple (j, u, v) can contribute to at most one ˆbi . Of course ˆbi ⊆ sup|j|≤n π −j ai = e for every i. If i ≤ n, we also have µ ¯ˆbi = µ ¯e · µ ¯bi . P P For |j| ≤ n, π j bi is a union of of U , so

µ ¯ˆbi =

n X X

X

µ ¯(π −j (a ∩ dvu ))

j=−n v∈V u∈U,u ⊆ π j bi

(because hπ

−j

(a ∩ dvu )i|j|≤n,u∈U,v∈V is dist) n X X X = µ ¯(a ∩ dvu ) = j=−n v∈V u∈U,u ⊆ π j bi

1 2n+2

n X X

X

µ ¯dvu


(by the choice of a) =

1 2n+2

n X X

X

µ ¯(v ∩ u)


(by the choice of dvu ) =

1 2n+2

n X

X

µ ¯u =

j=−n u∈U,u ⊆ π j bi

1 2n+2

n X

µ ¯(π j bi )

j=−n

j

(because π bi is a dist union of of U when i ≤ n, |j| ≤ n) =

2n+1 µ ¯ bi 2n+2

=µ ¯e · µ ¯ bi . Q Q

Again because C is atomless, we can choose a partition of unity hb∗i ii∈N ∈ C such that µ ¯b∗i = µ ¯bi for every i, ∗ ∗ ˆ ˆ while bi ⊇ bi and bi ∩ e = bi for i ≤ n.

386F


539

(d) Let E be the finite subalgebra of B generated by {π j bi : i ≤ n, |j| ≤ r} ∪ {π j ci : i ≤ n, |j| ≤ r}. Define θ : E → Ce˜ by setting θb = sup|j|≤n−r supv∈V supu∈U,u ⊆ πj b π −j (a ∩ dvu ) for b ∈ E. (i) θ is a Boolean homomorphism. P P The point is that if |j| ≤ n − r and b ∈ E, then π j b belongs to k the algebra generated by {π bi : i ≤ n, |k| ≤ n} ∪ {π k ci : i ≤ n, |k| ≤ n}, so is a union of of U . Since each map b 7→ π −j (a ∩ dvu ) if u ⊆ π j b,

0 otherwise

is a Boolean homomorphism from E to the principal ideal generated by π −j (a ∩ dvu ), and hπ −j (a ∩ dvu )i|j|≤n−r,u∈U,v∈V is a partition of unity in Ce˜, θ also is a Boolean homomorphism. Q Q (ii) µ ¯(θb) ≤ µ ¯b for every b ∈ E. P P (Compare (c) above.)

µ ¯(θb) =

n−r X

X

X

µ ¯π −j (a ∩ dvu )

j=−n+r v∈V u∈U,u ⊆ π j b

=

1 2n+2

n−r X

X

X

µ ¯(v ∩ u) =

j=−n+r v∈V u∈U,u ⊆ π j b

2n−2r+1 µ ¯b 2n+2

≤µ ¯b. Q Q

P Of course π k bi ∈ E. If |j| ≤ n − r, then |j + k| ≤ n, so (iii) θ(π k bi ) = e˜ ∩ π k b∗i for i ≤ n, |k| ≤ r. P π −j a ∩ θ(π k bi ) = sup

sup

v∈V u∈U,u ⊆ π j+k bi

¡ = π k sup

π −j (a ∩ dvu )

sup

v∈V u∈U,u ⊆ π j+k bi

¢ π −j−k (a ∩ dvu )

= π (π −j−k a ∩ ˆbi ) = π −j a ∩ π k (e ∩ b∗i ) = π −j a ∩ π k b∗i k

because π −j a ⊆ π k e. Taking the supremum of these pieces we have θ(π k bi ) = sup|j|≤n−r π −j a ∩ θ(π k bi ) = sup|j|≤n−r π −j a ∩ π k b∗i = e˜ ∩ π k b∗i . Q Q It follows that θ(π k (1 \ supi≤l bi )) = e˜ ∩ π k (1 \ supi≤l b∗i )) if l ≤ n, |k| ≤ r. (iv) Finally, θci = ci ∩ e˜ for every i ≤ n. P P If |j| ≤ n − r, v ∈ V then either v ⊆ π j ci or v ∩ π j ci = 0. In the former case, dvu = v ∩ u = 0 whenever u ∈ U and u 6⊆ π j ci , so that v = supu∈U dvu = supu∈U,u ⊆ πj ci dvu ; in the latter case, dvu = v ∩ u = 0 whenever u ⊆ π j ci . So we have v ∩ π j ci = supu∈U,u ⊆ πj ci dvu for every v ∈ V , and

540

Automorphism groups

θci = =

sup sup

sup

|j|≤n−r v∈V u∈U,u ⊆ π j ci

sup π −j (a ∩ sup

=

sup

dvu )

sup π −j (a ∩ sup (v ∩ π j ci )) v∈V

|j|≤n−r

=

π −j (a ∩ dvu )

v∈V u∈U,u ⊆ π j ci

|j|≤n−r

386F

sup π

−j

(a ∩ π j ci ) = ci ∩

|j|≤n−r

sup π −j a = ci ∩ e˜. Q Q |j|≤n−r

(e) Let B∗ be the closed subalgebra of A generated by {π j b∗i : i ∈ N, |j| ∈ Z}. Then for every b ∈ Br there is a b∗ ∈ B∗ such that θb = b∗ ∩ e˜. P P The set of b for which this is true is a subalgebra of A containing π k bi for i ≤ r and |k| ≤ r, by (d-iii). Q Q It follows that ρ(ci , B∗ ) ≤ 2ξ for i ∈ N. P P If i > n this is trivial, because µ ¯ci ≤ ξ, by the choice of n. Otherwise, ci ∈ E. Take b ∈ Br such that µ ¯(b 4 ci ) = ρ(ci , Br ) ≤ ξ (385Oa). Let b∗ ∈ B∗ be such that θb = b∗ ∩ e˜. Then ρ(ci , B∗k ) ≤ µ ¯(ci 4 b∗ ) ≤ 1 − µ ¯e˜ + µ ¯(˜ e ∩ (ci 4 b∗ )) =

2r+1 2n+2

+µ ¯((˜ e ∩ ci ) 4 θb) =

=

2r+1 2n+2

+µ ¯(θ(ci 4 b)) ≤

2r+1 2n+2

+µ ¯(θci 4 θb)

(by (d-iv)) 2r+1 2n+2

+µ ¯(ci 4 b)

(by (d-ii)) ≤ 2ξ by the choice of n. Q Q P (f ) Set B ∗ = {b∗i : i ∈ N} \ {0}. Then H(B ∗ ) = h(π, C) ≤ h(π, B ∗ ) + η. P H(B ∗ ) = H(B) = H(C) ¯bi for every i, and we supposed from the beginning that H(C) = H(B)) (because µ ¯b∗i = µ = h(π, C) (because C is a Bernoulli partition, see 386Bc) ≤ h(π¹ B∗r ) + H(C|B∗r ) (385Nd) ≤ h(π¹ B∗ ) +

∞ X

q(ρ(ci , B∗r ))

i=0

(by the definition of h(π¹ B∗ ), and 385Ob) ≤ h(π, B ∗ ) +

∞ X

q(min(2ξ, µ ¯ci ))

i=0

(by the Kolmogorov-Sinaˇı theorem, 384P(ii), and (e) above, recalling that ξ ≤ 61 , so that q is monotonic on [0, 2ξ]) ≤ h(π, B ∗ ) + η by the choice of ξ. Q Q

386G


541

(g) By 386D, applied to π¹ C and the partition hb∗i ii∈N of unity in C and the sequence hγi ii∈N = h¯ µb∗i ii∈N , ∗ we have a Bernoulli partition hdi ii∈N in C such that µ ¯di = µ ¯ bi = µ ¯bi for every i ∈ N and P∞ √ ² ¯(di 4 b∗i ) ≤ η + 6 4 2η ≤ . i=0 µ 4(2m+1)

j

Let D ⊆ C be the closed subalgebra of A generated by {π di : i ∈ N, j ∈ Z}. Then (B, π¹ B, hbi ii∈N ) is isomorphic to (D, π¹ D, hdi ii∈N ), with an isomorphism φ : B → D such that φπ = πφ and φbi = di for every i ∈ N (386Bi). (h) Set e∗ = e˜ \ sup|j|≤m,i∈N π j (di 4 b∗i ). Then φ(π j bi ) ∩ e∗ = θ(π j bi ) ∩ e∗ whenever i ≤ m and |j| ≤ m. P P φ(π j bi ) ∩ e∗ = π j (φbi ) ∩ e∗ = π j di ∩ e∗ = π j b∗i ∩ e∗ = π j b∗i ∩ e˜ ∩ e∗ = θ(π j bi ) ∩ e∗ by (d-iii), because i, |j| ≤ r ≤ n. Q Q Since b 7→ φb ∩ e∗ : A → Ae∗ , b 7→ θb ∩ e∗ : E → Ae∗ are Boolean ∗ ∗ homomorphisms, φb ∩ e = θb ∩ e for every b ∈ Bm . Now µ ¯(ci 4 φci ) ≤ ² for every i ∈ N. P P If i > n then of course µ ¯(φci 4 ci ) ≤ 2¯ µci ≤ 2ξ ≤ ². If i ≤ n, then (by the choice of m) there is a b ∈ Bm such that µ ¯(ci , b) ≤ 14 ². So φci 4 ci ⊆ (φci 4 φb) ∪ (φb 4 θb) ∪ (θb 4 θci ) ∪ (θci 4 ci ) ⊆

φ(ci 4 b) ∪ (1 \ e∗ ) ∪ θ(b 4 ci )

(using the definition of e∗ and (d-iv)) has measure at most µ ¯(ci 4 b) + µ ¯(1 \ e∗ ) + µ ¯(b 4 ci ) (by (d-ii), since b and c both belong to E) ≤ 2¯ µ(ci 4 b) + µ ¯(1 \ e˜) + (2m + 1)

∞ X

µ ¯(di 4 b∗i )

i=0

≤

² 2

+

2r+1 2n+2

+

² 4

≤ ²,

as required. Q Q 386G Lemma Let (A, µ ¯) be an atomless probability algebra and π a measure-preserving automorphism of A. Let hbi ii∈N , hci ii∈N be Bernoulli partitions for π, of the same finite entropy, and write B, C for the closed subalgebras generated by {π j bi : i ∈ N, j ∈ Z} and {π j ci : i ∈ N, j ∈ Z}. Suppose that C ⊆ B. Then for any ² > 0 we can find a Bernoulli partition hdi ii∈N for π such that (i) µ ¯ di = µ ¯ci for every i ∈ N, (ii) µ ¯(di 4 ci ) ≤ ² for every i ∈ N, (iii) writing D for the closed subalgebra of A generated by {π j di : i ∈ N, j ∈ Z}, ρ(bi , D) ≤ ² for every i ∈ N. proof (a) By 386F, there is a Bernoulli partition hb∗i ii∈N for π such that b∗i ∈ C for every i ∈ N, µ ¯b∗i = µ ¯bi 1 for every i ∈ N, and µ ¯(φci 4 ci ) ≤ 4 ² for every i ∈ N, where φ : B → C is the measure-preserving Boolean homomorphism such that φbi = b∗i for every i and πφ = φπ. Note that this implies that π −1 φ = φπ −1 , and generally that π j φ = φπ −j for every j ∈ Z; so φ[B] ⊆ C is the closed subalgebra of A generated by {φπ j bi : i ∈ N, j ∈ Z} = {π j b∗i : i ∈ N, j ∈ Z} (324L), and is invariant under the action of π and π −1 . Let m ∈ N be such that ρ(ci , Bm ) ≤ 41 ² for every i ∈ N,

542

Automorphism groups

386G

where Bm is the closed subalgebra of A generated by {π j bi : i ∈ N, |j| ≤ m} (385M). Let η ∈ ]0, π] be such that P∞ (2m + 1) i=0 min(η, 2¯ µbi ) ≤ 41 ². By 386F again, applied to π¹ C, there is a Bernoulli partition hc∗i ii∈N for π such that c∗i ∈ φ[B], µ ¯c∗i = µ ¯ci ∗ ∗ and µ ¯(ψbi 4 bi ) ≤ η for every i ∈ N, where ψ : C → C is the measure-preserving Boolean homomorphism such that ψci = c∗i for every i ∈ N and ψπ = πψ. Once again, ψ[C] will be the closed subalgebra of A generated by {ψπ j ci : i ∈ N, j ∈ Z} = {π j c∗i : i ∈ N, j ∈ Z}; because every c∗i belongs to φ[B], ψ[C] ⊆ φ[B]. (b) Now µ ¯(c∗i 4 φci ) ≤ ² for every i ∈ N. P P There is a b ∈ Bm such that µ ¯(ci 4 b) ≤ 41 ². We know that φ[Bm ] is the closed subalgebra of A generated by {φπ j bi : i ∈ N, |j| ≤ m} = {π j b∗i : i ∈ N, |j| ≤ m}, and contains φb. Because ψ(φb) 4 φb ⊆ supi∈N,|j|≤m ψ(π j b∗i ) 4 π j b∗i = sup|j|≤m π j (supi∈N ψb∗i 4 b∗i ), we have µ ¯(ψφb 4 φb) ≤ (2m + 1)

∞ X

µ ¯(ψb∗i 4 b∗i )

i=0

≤ (2m + 1)

∞ X

1 4

min(η, 2¯ µbi ) ≤ ².

i=0

But this means that µ ¯(c∗i 4 φci ) = µ ¯(ψci 4 φci ) ≤ µ ¯(ψci 4 ψφb) + µ ¯(ψφb 4 φb) + µ ¯(φb 4 φci ) ≤µ ¯(ci 4 φb) + ≤

² 4

² 4

+µ ¯(b 4 ci ) ≤ µ ¯(ci 4 φci ) + µ ¯(φci 4 φb) +

+µ ¯(ci 4 b) +

² 2

² 2

≤ ². Q Q

(c) Set di = φ−1 c∗i for each i; this is well-defined because φ is injective and c∗i ∈ φ[B]. Write D for the closed subalgebra of A or of B generated by {π j di : i ∈ N, j ∈ Z} = {φ−1 ψπ j ci : i ∈ N, j ∈ Z}; then D = φ−1 [ψ[C]], by 324L again, because φ−1 : φ[B] → B is a measure-preserving homomorphism. Then ¯ci for every i ∈ N, and hdi ii∈N is a Bernoulli partition for π. P P If i(0), . . . , i(n) ∈ N, then µ ¯ di = µ ¯c∗i = µ µ ¯( inf π j di(j) ) = µ ¯( inf π j φ−1 c∗i(j) ) = µ ¯(φ( inf π j φ−1 c∗i(j) )) j≤n j≤n j≤n Y Y j ∗ ∗ =µ ¯( inf π ci(j) )) = µ ¯ci(j) = µ ¯di(j) . Q Q j≤n

j≤n

j≤n

Next, µ ¯(ci 4 di ) = µ ¯(φci 4 φdi ) = µ ¯(φci 4 c∗i ) ≤ ² for every i, by (b). Finally, if i ∈ N, then ψb∗i belongs to ψ[C], while D = φ−1 [ψ[C]], so ρ(bi , D) = ρ(φbi , ψ[C]) ≤ µ ¯(φbi 4 ψb∗i ) = µ ¯(b∗i 4 ψb∗i ) ≤ η ≤ ². This completes the proof. 386H Lemma Let (A, µ ¯) be an atomless probability algebra and π a measure-preserving automorphism of A. Let hbi ii∈N , hci ii∈N be Bernoulli partitions for π, of the same finite entropy, and write B, C for the closed subalgebras generated by {π j bi : i ∈ N, j ∈ Z} and {π j ci : i ∈ N, j ∈ Z}. Suppose that C ⊆ B. Then for any ² > 0 we can find a Bernoulli partition hdi ii∈N for π such that (i) µ ¯di = µ ¯ci for every i ∈ N, (ii) µ ¯(di 4 ci ) ≤ ² for every i ∈ N, (iii) the closed subalgebra of A generated by {π j di : i ∈ N, j ∈ Z} is B.

386H


543

proof (a) To begin with (down to the end of (c) below) suppose that A = B. Choose h²n in∈N , hδn in∈N , hrn in∈N and hhdni ii∈N in∈N inductively, as follows. Start with d0i = ci for every i, r0 = 0. Given that hdni ii∈N is a Bernoulli partition with µ ¯dni = µ ¯ci for every i, take ²n > 0 such that (2rm + 1)²n ≤ 2−n for every m ≤ n, δn > 0 such that δn ≤ 2−n−1 ²,

P∞ i=0

min(δn , 2¯ µci ) ≤ ²n ,

and use 386G to find a Bernoulli partition hdn+1,i ii∈N for π such that µ ¯dn+1,i = µ ¯ci ,

µ ¯(dn+1,i 4 dni ) ≤ δn ,

ρ(bi , D(n+1) ) ≤ 2−n−1

for every i ∈ N, where D(n+1) is the closed subalgebra of A generated by {π j dn+1,i : i ∈ N, j ∈ Z}. Let rn+1 be such that (n+1)

ρ(bi , Drn+1 ) ≤ 2−n (n+1)

for every i ∈ N, where Drn+1 is the closed subalgebra of A generated by {π j dn+1,i : i ∈ N, |j| ≤ rn+1 }. Continue. (b) For any i ∈ N,

P∞ n=0

µ ¯(dn+1,i 4 dni ) ≤

so hdni in∈N has a limit di in A. Of course µ ¯(ci 4 di ) ≤

P∞ n=0

P∞

n=0 δn

≤ ²,

µ ¯(dn+1,i 4 dni ) ≤ ²

for every i. We must have µ ¯di = limn→∞ µ ¯dni = µ ¯ ci for each i, and if i 6= j then since

P∞

di ∩ dj = limn→∞ dni ∩ dnj = 0; µ ¯ci = 1, hdi ii∈N is a partition of unity in A. For any i(0), . . . , i(k) in N, Qk Qk ¯di(j) , µ ¯(inf j≤k π j di(j) ) = limn→∞ µ ¯(inf j≤k π j dn,i(j) ) = limn→∞ j=0 µ ¯dn,i(j) = j=0 µ

i=0

so hdi ii∈N is a Bernoulli partition. (c) Let D be the closed subalgebra of A generated by {π j di : i ∈ N, j ∈ Z}. Then bj ∈ D for every j ∈ N. (m+1) (m+1) ¯(bj 4 b) ≤ 2−m . Now P P Fix m ∈ N. Then ρ(bj , Drm+1 ) ≤ 2−m , so there is a b ∈ Drm+1 such that µ ∞ X

ρ(dm+1,i , D) ≤

i=0

≤

∞ X

µ ¯(dm+1,i 4 di ) ≤

i=0 ∞ X

∞ X

∞ ∞ X X

µ ¯(dk+1,i 4 dki )

i=0 k=m+1 ∞ X

min(2¯ µci , δk ) ≤

k=m+1 i=0

²k .

k=m+1

So ρ(b, D) ≤ (2rm+1 + 1)

∞ X

ρ(dm+1,i , D)

i=0

≤

∞ X

(2rm+1 + 1)²k ≤

k=m+1

∞ X

2−k = 2−m ,

k=m+1

and ρ(bj , D) ≤ µ ¯(bj 4 b) + ρ(b, D) ≤ 2−m + 2−m = 2 · 2−m . As m is arbitrary, ρ(bj , D) = 0 and bj ∈ D. Q Q (d) This completes the proof if A = B. For the general case, apply the arguments above to (B, µ ¯¹ B, π¹ B).

544

Automorphism groups

386I

386I Ornstein’s theorem (finite entropy case) Let (A, µ ¯) and (B, ν¯) be probability algebras, and π : A → A, φ : B → B two-sided Bernoulli shifts of the same finite entropy. Then (A, µ ¯, π) and (B, ν¯, φ) are isomorphic. proof (a) Let hai ii∈N , hbi ii∈N be (two-sided) generating Bernoulli partitions in A, B respectively. By the Kolmogorov-Sinaˇı theorem, hai ii∈N and hbi ii∈N both have entropy equal to h(π) = h(φ). If this entropy is zero, then A and B are both {0, 1}, and the result is trivial; so let us assume that h(π) > 0, so that A is atomless (386Bd). (b) By Sinaˇı’s theorem (386E), there is a Bernoulli partition hci ii∈N for π such that µ ¯ci = ν¯bi for every i ∈ N. By 386H, there is a Bernoulli partition hdi ii∈N for π such that µ ¯di = µ ¯ci for every i and the closed subalgebra of A generated by {π j di : i ∈ N, j ∈ Z} is A. But now (A, µ ¯, π, hdi ii∈N ) is isomorphic to (B, ν¯, φ, hbi ii∈N ), so (A, µ ¯, π) and (B, ν¯, φ) are isomorphic. 386J Using the same methods, we can extend the last result to the case of Bernoulli shifts of infinite entropy. The first step uses the ideas of 386C, as follows. Lemma Let (A, µ ¯) be a measure algebra and π : A → A an ergodic measure-preserving automorphism. Suppose that hai ii∈I is a finite Bernoulli partition for π, with #(I) = r and µ ¯ai = 1/r for every i ∈ I, and that h(π) ≥ ln 2r. Then for any ² > 0 there is a Bernoulli partition hbij ii∈I,j∈{0,1} for π such that µ ¯(ai 4 (bi0 ∪ bi1 )) ≤ ²,

µ ¯bi0 = µ ¯bi1 =

1 2r

for every i ∈ I. proof (a) Let δ > 0 be such that

√ δ + 6 4δ ≤ ².

Let η > 0 be such that η < ln 2,

√

8η ≤ δ,

and |t − 21 | ≤ δ whenever t ∈ [0, 1] and q(t) + q(1 − t) ≥ ln 2 − 4η (384Ad). We have 1 r

H(A) = rq( ) = ln r, and µ ¯d = r−n whenever n ∈ N, d ∈ Dn (A, π). Note that A is atomless. P P?? If a ∈ A is an atom, then (because π is ergodic) supj∈N π j a = 1, and A is purely atomic, with finitely many atoms all of the same size as a; but this means that H(C) ≤ ln( µ¯1a ) for every partition of unity C ⊆ A, so that 1 n

h(π, C) = limn→∞ H(Dn (C, π), π) ≤ limn→∞

1 n

ln(

1 ) µ ¯a

=0

for every partition of unity C, and 0 = h(π) ≥ ln 2r ≥ ln 2. X XQ Q (b) There is a finite partition of unity C ⊆ A such that h(π, C) = ln 2r − η, and C refines A = {ai : i ∈ I} \ {0}. P P Because h(π) ≥ ln 2r, there is a finite partition of unity C 0 such 0 0 that h(π, C ) ≥ ln 2r − η; replacing C by C 0 ∨ A if need be, we may suppose that C 0 refines A; take such a C 0 of minimal size. Because H(C 0 ) ≥ h(π, C 0 ) > H(A), there must be distinct c0 , c1 ∈ C 0 included in the same member of A. Because A is atomless, the principal ideal generated by c1 has a closed subalgebra isomorphic, as measure algebra, to the measure algebra of Lebesgue measure on [0, 1], up to a scalar multiple of the measure; and in particular there is a family hdt it∈[0,1] such that ds ⊆ dt whenever s ≤ t, d1 = c1 and µ ¯dt = t¯ µc1 for every t ∈ [0, 1]. Let Dt be the partition of unity

386J


545

(C 0 \ {c0 , c1 }) ∪ {c0 ∪ dt , c1 \ dt } for each t ∈ [0, 1]. Then h(π, D1 ) = h(π, (C 0 \ {c0 , c1 }) ∪ {c0 ∪ c1 }) < ln 2r − η, by the minimality of #(C 0 ), while h(π, D0 ) = h(π, C 0 ) ≥ ln 2r − η. Using 384N, we also have, for any s, t ∈ [0, 1] such that |s − t| ≤ 1e , h(π, Ds ) − h(π, Dt ) ≤ H(Ds |Dt ) (where Dt is the closed subalgebra generated by Dt ) ≤ q(ρ(c0 ∪ ds , Dt )) + q(ρ(c1 \ ds , Dt )) (by 385Ob, because Ds \ Dt ⊆ {c0 ∪ ds , c1 \ ds }) ≤ q(¯ µ((c0 ∪ ds ) 4 (c0 ∪ dt ))) + q(¯ µ((c1 \ ds ) 4 (c1 \ dt ))) = 2q(¯ µ(ds 4 dt )) = 2q(|s − t|¯ µc1 ) because q is monotonic on [0, |s − t|¯ µc1 ]. But this means that t 7→ h(π, Dt ) is continuous and there must be some t such that h(π, Dt ) = ln 2r − η; take C = Dt . Q Q (c) Let ξ > 0 be such that 1 6

ξ ≤ η,

ξ≤ , P c∈C

q(2ξ) + q(1 − 2ξ) ≤ η,

q(min(2ξ, µ ¯c)) ≤ η.

Let n ∈ N be such that 1 n+1

≤ ξ,

q(

1 n ) + q( ) n+1 n+1

≤ η,

µ ¯[[wn − h(π, C)χ1 ≥ η]] ≤ ξ, where wn =

1 n

P d∈Dn (C,π)

ln(

1 )χd. µ ¯d

(The Shannon-McMillan-Breiman theorem, 385H, assures us that any sufficiently large n has these properties.) (d) Let D be the set of those d ∈ Dn (C, π) such that µ ¯d ≥ (2r)−n ,

i.e.,

1 n

ln(

1 ) µ ¯d

≤ ln 2r.

Then µ ¯(sup D) ≥ 1 − ξ, by the choice of n, because h(π, C) = ln 2r − η. Note that every member of D is included in some member of Dn (A, π), because C refines A. If b ∈ Dn (A, π), then µ ¯b = r−n , so n n #({d : d ∈ D, d ⊆ b}) ≤ 2 ; we can therefore find a function f : D → {0, 1} such that f is injective on {d : d ∈ D, d ⊆ b} for every b ∈ Dn (A, π). (e) By 385E(iv), as usual, there is an a ∈ A such that a, π −1 a, . . . , π −n+1 a are dist and µ ¯(a ∩ d) = for every d ∈ Dn (C, π). Set

1 ¯d n+1 µ

e = supd∈D,j
Pn−1 P j=0

d∈D

µ ¯(a ∩ d) =

n µ ¯(sup D) n+1

≥ (1 − ξ)2 ≥ 1 − 2ξ.

546

Automorphism groups

386J

c∗ = sup{π −j (a ∩ d) : j < n, d ∈ D, f (d)(j) = 1}. (I am identifying of {0, 1}n with functions from {0, . . . , n − 1} to {0, 1}.) Set A∗ = A ∨ {c∗ , 1 \ c∗ },

A0 = A∗ ∨ {a, 1 \ a} ∨ {e, 1 \ e},

and let A0 be the closed subalgebra of A generated by {π j a : a ∈ A0 , j ∈ Z}. Then a ∩ d ∈ A0 for every d ∈ D. P P Set d˜ = upr(a ∩ d, A0 ) = inf{c : c ∈ A0 , c ⊇ a ∩ d} ∈ A0 . Let b be the element of Dn (A, π) including d. Because a, b, e ∈ A0 , d˜ ⊆ a ∩ b ∩ e = supd0 ∈D a ∩ b ∩ d0 = sup{a ∩ d0 : d0 ∈ D, d0 ⊆ b}. Now if d0 ∈ D, d0 ⊆ b and d0 6= d, then f (d0 ) 6= f (d). Let j be such that f (d0 )(j) 6= f (d)(j); then π −j (a ∩ d) is included in one of c∗ , 1 \ c∗ and π −j (a ∩ d0 ) in the other. This means that one of π j c∗ , 1 \ π j c∗ is a member of A0 including a ∩ d and dist from a ∩ d0 , so that d˜ ∩ d0 = 0. Thus d˜ must be actually equal to a ∩ d, and a ∩ d ∈ A0 . Q Q Next, c ∩ e ∈ A0 for every c ∈ C. P P hπ −j (a ∩ d)ij
ρ(c, A0 ) ≤ µ ¯(c 4 (c ∩ e)) = µ ¯(c \ e) ≤ min(¯ µc, 2ξ) ≤ , so ln 2r − η = h(π, C) ≤ h(π¹ A0 ) + H(C|A0 ) (385Nd) ≤ h(π, A0 ) +

X

q(ρ(c, A0 ))

c∈C

(by the Kolmogorov-Sinaˇı theorem and 385Ob) ≤ h(π, A0 ) +

X

q(min(¯ µc, 2ξ)) ≤ h(π, A0 ) + η

c∈C

by the choice of ξ. Q Q Finally, h(π, A∗ ) ≥ ln 2r − 4η. P P ln 2r − 2η ≤ h(π, A0 ) ≤ h(π, A∗ ) + H({a, 1 \ a}) + H({e, 1 \ e}) (applying 385Nb twice) = h(π, A∗ ) + q(¯ µa) + q(1 − µ ¯a) + q(¯ µe) + q(1 − µ ¯e) 1 n

≤ h(π, A∗ ) + q( ) + q(

n ) + q(2ξ) + q(1 − 2ξ) n+1

≤ h(π, A∗ ) + η + η = h(π, A∗ ) + 2η. Q Q (g) It follows that H({c∗ , 1 \ c∗ }) ≥ ln 2 − 4η. P P We have ln 2r − 4η ≤ h(π, A∗ ) ≤ H(A∗ ) ≤ H(A) + H({c∗ , 1 \ c∗ }) = ln r + H({c∗ , 1 \ c∗ }), so H({c∗ , 1 \ c∗ }) ≥ ln 2 − 4η. Q Q Thus q(¯ µc∗ ) + q(1 − µ ¯c∗ ) ≥ ln 2 − 4η; by the choice of η, |¯ µc∗ − 21 | ≤ δ. Next, P 1 1 µ(ai \ c∗ ) − | ≤ 3δ. µ(ai ∩ c∗ ) − | + |¯ i∈I |¯ 2r

2r

386K


547

P P By 385K, X

1 r

1 r

|¯ µ(ai ∩ c∗ ) − µ ¯c∗ | + |¯ µ(ai \ c∗ ) − µ ¯(1 \ c∗ )|

i∈I

≤ ≤

p

2(H(A) + H({c∗ , 1 \ c∗ }) − H(A∗ )) p 2(ln r + ln 2 − ln 2r + 4η) = 8η ≤ δ.

p

So X

|¯ µ(ai ∩ c∗ ) −

i∈I

≤

1 1 | + |¯ µ(ai \ c∗ ) − | 2r 2r

X¡ 1 ∗ 1 1 |¯ µ(ai ∩ c∗ ) − µ ¯c | + |¯ µc∗ − | r

i∈I

r

2

1 r

1 r

1 2

+ |¯ µ(ai \ c∗ ) − µ ¯(1 \ c∗ )| + |¯ µ(1 \ c∗ ) − |

¢

1 2

1 2

µ(1 \ c∗ ) − | ≤ 3δ. Q Q ≤ δ + |¯ µc∗ − | + |¯ (h) Now apply 386D to the partition of unity A∗ , indexed as ha∗ij ii∈I,j∈{0,1} , where a∗i1 = ai ∩ c∗ and 1 = ai \ c∗ , and hγij ii∈I,j∈{0,1} , where γij = 2r for all i, j. We have P µa∗ij − γij | ≤ 3δ i∈I,j∈{0,1} |¯

a∗i0

by (g), while H(A∗ ) − h(π, A∗ ) ≤ ln r + ln 2 − ln 2r + 4η = 4η, so

P i∈I,j∈{0,1}

Also

|¯ µa∗ij − γij | +

p √ 2(H(A∗ ) − h(π, A∗ )) ≤ 3δ + 8η ≤ 4δ.

P i∈I,j∈{0,1}

q(γij ) = ln 2r ≤ h(π).

So 386D tells us that there is a Bernoulli partition hbij ii∈I,j∈{0,1} for π such that µ ¯b∗ij = √ P ¯(bij 4 a∗ij ) ≤ δ + 6 4δ ≤ ². i∈I,j∈{0,1} µ

1 2r

for all i, j and

Now of course X

µ ¯(ai 4 (bi0 ∪ bi1 )) ≤

i∈I

X i∈I

=

µ ¯((ai ∩ c∗ ) 4 bi1 ) + µ ¯((ai \ c∗ ) 4 bi0 )

X

µ ¯(a∗ij 4 bij ) ≤ ²,

i∈I,j∈{0,1}

as required. 386K Ornstein’s theorem (infinite entropy case) Let (A, µ ¯) be a probability algebra of countable Maharam type, and π : A → A a two-sided Bernoulli shift of infinite entropy. Then (A, µ ¯, π) is isomorphic to (B, ν¯, φ), where (B, ν¯) is the measure algebra of the usual measure on [0, 1]Z , and φ is the standard Bernoulli shift on B (384Sb). proof (a) We have to find a root algebra E for π which is isomorphic to the measure algebra of the usual measure on [0, 1]. The materials we have to start with are a root algebra A0 of A such that either A0 is not purely atomic or H(A0 ) = ∞, where A0 is the set of atoms of A0 . Because A has countable Maharam type, there is a sequence hdn in∈N in A0 such that {dn : n ∈ N} is dense in the metric of A0 .

548

Automorphism groups

386K

(b) There is a sequence hCn in∈N of partitions of unity in A0 such that Cn+1 refines Cn , H(Cn ) = n ln 2 and dn is a union of of Cn+1 for every n. P P We have ∞ = sup{H(C) : C ⊆ A0 is a partition of unity} (384J). Choose the Cn inductively, as follows. Start with C0 = {0, 1}. Given Cn with H(Cn ) = n ln 2, set Cn0 = Cn ∨ {dn , 1 \ dn }; then H(Cn0 ) ≤ H(Cn ) + H({dn , 1 \ dn }) ≤ (n + 1) ln 2. By 385Q, there is a partition of unity Cn+1 , refining Cn0 , such that H(Cn+1 ) = (n + 1) ln 2. Continue. Q Q (c) For each n ∈ N, let Cn be the closed subalgebra of A generated by {π j a : a ∈ Cn , j ∈ Z}. Then hCn in∈N is increasing, and dn ∈ Cn+1 , π[Cn ] = Cn and h(π¹ Cn ) = h(π, Cn ) = H(Cn ) = n ln 2 for every n. Choose inductively, for each n ∈ N, ²n > 0, rn ∈ N and a Bernoulli partition hbnσ iσ∈{0,1}n in Cn , as follows. Start with bn∅ = 1. (See 3A1H for the notation I am using here.) Given that hbnσ iσ∈{0,1}n is a Bernoulli partition for π which generates Cn , in the sense that Cn is the closed subalgebra of A generated by {π j bnσ : σ ∈ {0, 1}n , j ∈ Z}, and µ ¯bnσ = 2−n for every σ, take ²n > 0 such that (2rm + 1)²n ≤ 2−n for every m < n. We know that h(π¹ Cn+1 ) = (n + 1) ln 2 = ln(2 · 2n ). So we can apply 386J to (Cn+1 , π¹ Cn+1 ) to see that there is a Bernoulli partition hb0nτ iτ ∈{0,1}n+1 for π such that b0nτ ∈ Cn+1 ,

µ ¯b0nτ = 2−n−1

for every τ ∈ {0, 1}n+1 , µ ¯(bnσ 4 (b0n,σa 0 ∪ b0n,σa 1 )) ≤ 2−n ²n for every σ ∈ {0, 1}n . By 386H (with B = C = Cn+1 ), there is a Bernoulli partition hbn+1,τ iτ ∈{0,1}n+1 for π¹ Cn+1 such that the closed subalgebra generated by {π j bn+1,τ : τ ∈ {0, 1}n+1 , j ∈ Z} is Cn+1 , µ ¯bn+1,τ = 2−n−1 for every τ ∈ {0, 1}n+1 , and P ¯(bn+1,τ 4 b0nτ ) ≤ ²n . τ ∈{0,1}n+1 µ (n+1)

For each k ∈ N, let Bk be the closed subalgebra of Cn+1 generated by {π j bn+1,τ : τ ∈ {0, 1}n+1 , |j| ≤ k}. Since dm ∈ Cm+1 ⊆ Cn+1 for every m ≤ n, there is an rn ∈ N such that (n+1)

ρ(dm , Brn

) ≤ 2−n for every m ≤ n.

Continue. (d) Fix m ≤ n ∈ N for the moment. For σ ∈ {0, 1}m , set bnσ = sup{bnτ : τ ∈ {0, 1}n , τ extends σ}. (If n = m, then of course σ is the unique member of {0, 1}m extending itself, so this formula is safe.) Then µ ¯bnσ = 2−n #({τ : τ ∈ {0, 1}n , τ extends σ}) = 2−n 2n−m = 2−m . Next, if σ, σ 0 ∈ {0, 1}m are distinct, there is no member of {0, 1}n extending both, so bnσ ∩ bnσ0 = 0; thus hbnσ iσ∈{0,1}m is a partition of unity. If σ(0), . . . , σ(k) ∈ {0, 1}m , then

386K


µ ¯( inf π j bm,σ(j) ) = µ ¯( j≤k

549

inf π j bn,τ (j) )

sup

τ (0),... ,τ (k)∈{0,1}n j≤k τ (j)⊇σ(j)∀j≤k

X

=

µ ¯( inf π j bn,τ (j) ) j≤k

τ (0),... ,τ (k)∈{0,1}n τ (j)⊇σ(j)∀j≤k

X

=

(2−n )k+1

τ (0),... ,τ (k)∈{0,1} τ (j)⊇σ(j)∀j≤k

n

= (2n−m )k+1 (2−n )k+1 = (2−m )k+1 =

k Y

µ ¯bn,σ(j) ,

j=0

so hbnσ iσ∈{0,1}m is a Bernoulli partition. (e) If m ≤ n ∈ N, then

P σ∈{0,1}m

µ ¯(bnσ 4 bn+1,σ ) ≤ 2²n

whenever m ≤ n ∈ N. P P We have X

µ ¯(bnσ 4 bn+1,σ ) ≤

σ∈{0,1}m

X

µ ¯(bnτ 4 bn+1,τ )

τ ∈{0,1}n

=

X

µ ¯(bnτ 4 (bn+1,τ a 0 ∪ bn+1,τ a 1 ))

τ ∈{0,1}n

≤

X

τ ∈{0,1}n

≤

X

X

µ ¯(bnτ 4 (b0n,τ a 0 ∪ b0n,τ a 1 )) +

µ ¯(b0nυ 4 bn+1,υ )

υ∈{0,1}n+1

2−n ²n + ²n = 2²n . Q Q

τ ∈{0,1}n

(f ) In particular, for any m ∈ N and σ ∈ {0, 1}m , P∞ P∞ ¯(bnσ 4 bn+1,σ ) ≤ n=m 2²n < ∞. n=m µ So we can define bσ = limn→∞ bnσ in A. We have µ ¯bσ = limn→∞ µ ¯bnσ = 2−m ; and if σ, σ 0 ∈ {0, 1}m are distinct, then bσ ∩ bσ0 = limn→∞ bnσ ∩ bnσ0 = 0, so hbσ iσ∈{0,1}m is a partition of unity in A. If σ(0), . . . , σ(k) ∈ {0, 1}m , then µ ¯( inf π j bσ(j) ) = lim µ ¯( inf π j bn,σ(j) ) j≤k

n→∞

= lim

n→∞

j≤k

k Y

µ ¯bn,σ(j) =

j=0

k Y

µ ¯bσ(j) ,

j=0

so hbσ iσ∈{0,1}m is a Bernoulli partition for π. If σ ∈ {0, 1}m , then bσa 0 ∪ bσa 1 = limn→∞ bn,σa 0 ∪ bn,σa 1 = limn→∞ bn,σ = bσ . S (g) Let E be the closed subalgebra of A generated by m∈N {bσ : σ ∈ {0, 1}m }. Then E is atomless and countably generated, so (E, µ ¯¹ E) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. Now Qk µ ¯(inf j≤k π j ej ) = j=0 µ ¯ej for all e0 , . . . , ek ∈ E. P P Let ² > 0. For m ∈ N, let Em be the subalgebra of E

550

Automorphism groups

386K

S generated by {bσ : σ ∈ {0, 1}m }. hEm im∈N is non-decreasing, so m∈N Em is a closed subalgebra of A, and must be E. Now the function Qk (a0 , . . . , ak ) → µ ¯(inf j≤k π j aj ) − j=0 µ ¯aj : Ak+1 → R is continuous and zero on Ek+1 for every m, by 386Bb, so is zero on Ek+1 , and in particular is zero at m (e0 , . . . , ek ), as required. Q Q By 384Sf, hπ j [E]ij∈Z is independent. S (h) Let B∗ be the closed subalgebra of A generated by {π j bσ : σ ∈ m∈N {0, 1}m , j ∈ Z}; then E ⊆ B∗ , S so B∗ is the closed subalgebra of A generated by j∈Z π j [E]. It follows from (e) that, for any m ∈ N, X X ρ(bmσ , B∗ ) ≤ µ ¯(bmσ , bσ ) σ∈{0,1}m

σ∈{0,1}m ∞ X

X

≤

µ ¯(bnσ , bn+1,σ ) ≤ 2

σ∈{0,1}m n=m (m+1)

So if b ∈ Brm

∞ X

²n .

n=m

, X

ρ(b, B∗ ) ≤ (2rm + 1)

ρ(bm+1,σ , B∗ )

σ∈{0,1}m+1

≤ 2(2rm + 1)

∞ X

²n ≤ 2

n=m+1

∞ X

2−n = 2−m+1 .

n=m+1

It follows that, whenever m ≤ n in N, (n+1)

ρ(dm , B∗ ) ≤ ρ(dm , Brn

) + 2−n+1 ≤ 2−n + 2−n+1

by the choice of rn . Letting n → ∞, we see that ρ(dm , B∗ ) = 0, that is, dm ∈ B∗ , for every m ∈ N. But this means that A0 ⊆ B∗ , by the choice of hdm im∈N . Accordingly π j [A0 ] ⊆ B∗ for every j and B∗ must be the whole of A. (i) Thus π is a two-sided Bernoulli shift with root algebra E; by 384Sc, (A, µ ¯, π) is isomorphic to (B, ν¯, φ). 386L Corollary: Sinaˇı’s theorem (general case) Let (A, µ ¯) be an atomless probability algebra, and π : A → A a measure-preserving automorphism. Let (B, ν¯) be a probability algebra of countable Maharam type, and φ : B → B a one- or two-sided Bernoulli shift with h(φ) ≤ h(π). Then (B, ν¯, φ) is isomorphic to a factor of (A, µ ¯, π). proof (a) To begin with (down to the end of (b)) suppose that φ is two-sided. Let B0 be a root algebra for φ. If B0 is purely atomic, then there is a generating Bernoulli partition hbi ii∈N for φ of entropy h(φ). By 386E, there is a Bernoulli partition hci ii∈N for π such that µ ¯ci = ν¯bi for every i. Let C be the closed subalgebra of A generated by {π j ci : i ∈ N, j ∈ Z}. Now (C, µ ¯¹ C, π¹ C) is a factor of (A, µ ¯, π) isomorphic to (B, ν¯, φ). (b) If B0 is not purely atomic, then there is still a partition of unity hbi ii∈N in B0 of infinite entropy. Again, let C be the closed subalgebra of A generated by {π j ci : i ∈ N, j ∈ Z}, where hci ii∈N is a Bernoulli partition for π such that µ ¯ci = ν¯bi for every i. Now π¹ C is a Bernoulli shift of infinite entropy and C has countable Maharam type, so 386K tells us that there is a closed subalgebra C0 ⊆ C such that hπ k [C0 ]ik∈N is independent and (C0 , µ ¯¹ C0 ) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. But (B0 , ν¯¹ B0 ) is a probability algebra of countable Maharam type, so is isomorphic to a closed subalgebra k ∗ C1 of C0 (332N). S Of kcourse hπ ∗[C1 ]ik∈N is independent, so if we take C1 to be the closed subalgebra of A generated by k∈Z π [C1 ], π¹ C1 will be a two-sided Bernoulli shift isomorphic to φ (384Sf). (c) If φ is a one-sided Bernoulli shift, then 384Sa and 384Sc show that (B, ν¯, φ) can be represented in of a product measure on a space X N and the standard shift operator on X N . Now this extends naturally to the standard two-sided Bernoulli shift represented by the product measure on X Z , as described in 384Sb; so that (B, ν¯, φ) becomes represented as a factor of (B0 , ν¯0 , φ0 ) where φ0 is a two-sided Bernoulli

387A

Dye’s theorem

551

shift with the same entropy as φ (since the entropy is determined by the root algebra, by 384R). By (a)-(b), (B0 , ν¯0 , φ0 ) is isomorphic to a factor of (A, µ ¯, π), so (B, ν¯, φ) also is. Remark Thus (A, µ ¯, π) has factors which are Bernoulli shifts based on root algebras of all countablygenerated types permitted by the entropy of π. 386X Basic exercises (a) Let (A, µ ¯) be a probability algebra, and π : A → A a one- or two-sided Bernoulli shift. Show that π n isSa Bernoulli shift for any n ≥ 1. (Hint: if A0 is a root algebra for π, the closed subalgebra generated by j
387 Dye’s theorem I have repeatedly said that any satisfactory classification theorem for automorphisms of measure algebras remains elusive. There is however a classification, at least for the Lebesgue measure algebra, of the ‘orbit structures’ corresponding to measure-preserving automorphisms; in fact, they are defined by the fixed-point subalgebras, which I described in §333. We have to work hard for this result, but the ideas are instructive. 387A Full subgroups Recall the definition of ‘full’ subgroup (381M): if A is a Dedekind complete Boolean algebra, then a subgroup G of its automorphism group is full if whenever hai ii∈I is a partition of unity in A, and φ ∈ Aut A is such that for each i ∈ I there is a πi ∈ G such that φb = πi b for every b ⊆ ai , then φ ∈ G. I take the opportunity to note two facts: (i) if A is a Dedekind complete Boolean algebra and c ∈ A, then {π : π ∈ Aut A, πc = c} is full (ii) if (A, µ ¯) is a localizable measure algebra, then the set of measure-preserving automorphisms of A is a full subgroup of Aut A (382C). Evidently the intersection of any family of full subgroups of Aut A is full, so we can speak of the full subgroup Gπ generated by a given automorphism π. From the facts cited above (or otherwise) we see that (i)0 if π ∈ Aut A and C is the closed subalgebra {c : πc = c}, then φc = c for every φ ∈ Gπ and c ∈ C (ii)0 if (A, µ ¯) is a localizable measure algebra and π ∈ Aut A is measure-preserving, then every member of Gπ is measure-preserving. Now the subgroups Gπ are easy to describe, as follows.

552

Automorphism groups

387B

387B Proposition Let A be a Dedekind complete Boolean algebra and π, φ Boolean automorphisms of A. Then the following are equiveridical: (i) φ belongs to the full subgroup of Aut A generated by π; (ii) for every non-zero a ∈ A there are a non-zero b ⊆ a and an n ∈ Z such that φc = π n c for every c ⊆ b; (iii) there is a partition of unity han in∈Z in A such that φc = π n c whenever n ∈ Z and c ⊆ an . proof (i)⇒(ii) Let G be the set of those automorphisms satisfying (ii). Then G is a full subgroup of Aut A. P P (α) Suppose that ψ1 , ψ2 ∈ G and a ∈ A \ {0}. Then there are a non-zero b1 ⊆ a and an m ∈ Z such that ψ2 c = π m c for every c ⊆ b1 . Next, there are a non-zero b2 ⊆ ψ2 b1 and an n ∈ Z such that ψ1 c = π n c for every c ⊆ b2 . Set b = ψ2−1 b2 , so that 0 6= b ⊆ a and ψ1 ψ2 c = π n ψ2 c = π n π m c = π m+n c for every c ⊆ b. As a is arbitrary, ψ1 ψ2 ∈ G. (β) Suppose that ψ ∈ G and a ∈ A \ {0}. Then there are a non-zero b1 ⊆ ψ −1 a and an n ∈ Z such that ψc = π n c for every c ⊆ b1 . Set b = ψb1 , so that 0 6= b ⊆ a; then for any c ⊆ b, π −n c = π −n ψψ −1 c = ψ −1 c. As a is arbitrary, ψ −1 ∈ G. (γ) Of course π ∈ G, so G is a subgroup of Aut A. (δ) Now suppose that hai ii∈I is a partition of unity in A and that ψ ∈ Aut A is such that for every i ∈ I there is a ψi ∈ G such that ψc = ψi c for every c ⊆ ai . If a ∈ A \ {0}, then there is an i ∈ I such that a ∩ ai 6= 0, and there are a non-zero b ⊆ a ∩ ai and an n ∈ Z such that ψc = ψi c = π n c for every c ⊆ b. As a is arbitrary, ψ ∈ G. As hai ii∈I is arbitrary, G is full. Q Q Since G is a full subgroup of Aut A containing π, it includes the full subgroup generated by π and, in particular, φ ∈ G. (ii)⇒(iii) For n ∈ Z, let Bn be the set of those b ∈ A such that φc = π n c for every c ⊆ b. Set bn = sup Bn for each n; then if c ⊆ bn , φc = φ(supb∈Bn b ∩ c) = supb∈Bn φ(b ∩ c) = supb∈Bn π n (b ∩ c) = π n c. Set an = bn \ sup bi if n ∈ N, 0≤i

= bn \ sup bi if n ∈ Z \ N; i>n

then han in∈Z is dist, supn∈Z an = supn∈Z bn = sup(

S n∈Z

Bn ) = 1,

and φc = π n c for every c ⊆ an , n ∈ Z; so (iii) is true. (iii)⇒(i) is trivial. 387C Orbit structures I said that this section was directed to a classification of ‘orbit structures’, without saying what these might be. In fact what I will do is to classify the full subgroups generated by measure-preserving automorphisms of the Lebesgue measure algebra. One aspect of the relation with ‘orbits’ is the following. Proposition Let (X, Σ, µ) be a localizable countably separated measure space (definition: 343D), with measure algebra (A, µ ¯). Suppose that f and g are measure space automorphisms from X to itself, inducing measure-preserving automorphisms π, φ of A. Then the following are equiveridical: (i) φ belongs to the full subgroup of Aut A generated by π; (ii) for almost every x ∈ X, there is an n ∈ Z such that g(x) = f n (x); (iii) for almost every x ∈ X, {g n (x) : n ∈ Z} ⊆ {f n (x) : n ∈ Z}. proof (i)⇒(ii) Let hHk ik∈N be a sequence in Σ which separates the points of X; we may suppose that H0 = X. By 387B, there is a partition of unity han in∈Z in A S such that φc = π n c for every c ⊆ an , n ∈ Z. • For each n ∈ Z let En ∈ Σ be such that En = an ; then Y0 = n∈Z En is conegligible. The transformation f n induces π n , so for any k ∈ N and n ∈ Z the set

387E

Dye’s theorem

553

Fnk = {x : f n (x) ∈ En ∩ Hk , g(x) ∈ / En ∩ Hk } ∪ {x : g(x) ∈ En ∩ Hk , f n (x) ∈ / E n ∩ Hk } S is negligible, and Y = g −1 [Y0 ] \ n∈Z,k∈N Fnk is conegligible. Now, for any x ∈ Y , there is some n such that g(x) ∈ En , so that f (x) ∈ En and {k : g(x) ∈ Hk } = {k : f n (x) ∈ Hk } and g(x) = f n (x). As Y is conegligible, (ii) is satisfied. n (ii)⇒(iii) / Ωx } is negligible. S For x ∈ X, set Ωx = {f (x) : n ∈ Z}; we are supposing that A0 = {x : g(x) ∈ Set A = n∈Z g −n [A0 ], so that A is negligible and g n (x) ∈ X \ A for every x ∈ X \ A, n ∈ Z. Suppose that x ∈ X \ A and n ∈ N. Then g n (x) ∈ Ωx . P P Induce on n. Of course g 0 (x) = x ∈ Ωx . For n the inductive step to n + 1, g (x) ∈ Ωx \ A0 , so there is a k ∈ Z such that g n (x) = f k (x). At the same time, there is an i ∈ Z such that g(g n (x)) = f i (g n (x)), so that g n+1 (x) = f i+k (x) ∈ Ωx . Thus the induction continues. Q Q Consequently g −n (x) ∈ Ωx whenever x ∈ X \ A and n ∈ N. P P Since g −n (x) ∈ X \ A, there is a k ∈ Z n −n k −n −n −k such that x = g g (x) = f g (x) and g (x) = f (x) ∈ Ωx . Q Q Thus {g n (x) : n ∈ Z} ⊆ Ωx for every x in the conegligible set X \ A.

(iii)⇒(ii) is trivial. (ii)⇒(i) Set

S En = {x : g(x) = f n (x)} = X \ k∈N (g −1 Hk 4f −n [Hk ]), S S for n ∈ Z. Then (ii) tells us that n∈Z En is conegligible, so n∈Z g[En ] is conegligible. But also each En is measurable, so g[En ] also is, and we can set an = g[En ]• . Now for y ∈ g[En ], y = f n (g −1 (y)), that is, g −1 (y) = f −n (y); so φa = π n a for every a ⊆ an . Since supn∈Z an = 1 in A, φ belongs to the full subgroup generated by π. Remark Of course the requirement ‘countably separated’ is essential here; for other measure spaces we can have φ and π actually equal without g(x) and f (x) being related for any particular x (see 343I and 343J). 387D Corollary Under the hypotheses of 387C, π and φ generate the same full subgroup of Aut A iff {f n (x) : n ∈ Z} = {g n (x) : n ∈ Z} for almost every x ∈ X. 387E Induced automorphisms of principal ideals: Proposition Let (A, µ ¯) be a totally finite measure algebra and π : A → A a measure-preserving automorphism. For b ∈ A write Ab for the principal ideal of A generated by b. (a) For any b ∈ A we have a measure-preserving automorphism πb : Ab → Ab defined by saying that πb d = π n d whenever n ≥ 1 and d ⊆ b ∩ π −n b \ sup1≤i
dm ⊆ π −m b

so dm ∩ dn = 0. Also dm ⊆ b,

π n−m dn ∩ b = π n−m (dn ∩ π −(n−m) b) = 0

so π n dn ∩ π m dm = π m (π n−m dn ∩ dm ) = 0.

554

Automorphism groups

387E

Thirdly, supn≥1 dn = supn≥1 b ∩ π −n b = b by 385D, applied to π −1 . It follows that hdn in≥1 is a partition of unity in Ab . Since hπ n dn in≥1 is also a dist family in Ab , and P∞ P∞ ¯π n dn = n=1 µ ¯ dn = µ ¯b, n=1 µ it is another partition of unity. So we have an automorphism πb : Ab → Ab defined by setting πb d = π n d if d ⊆ dn (381B). Of course πb is measure-preserving, by 382C. (b) If πc = c, then π n (dn ∩ c) ⊆ c for every n, so πb (b ∩ c) ⊆ c; because µ ¯πb (b ∩ c) = µ ¯(b ∩ c), πb (b ∩ c) = b ∩ c. (c) Set D = {d : d ∈ Ac , πc d = (πb )c d}. Then D is order-dense in Ac . P P Take any non-zero a ∈ Ac . Since c ⊆ supn≥1 π −n c, there is an n ∈ N such that a0 = a ∩ π −n c \ sup1≤i
π mi+1 −mi di = di+1 ⊆ b,

while π j di = π mi +j d ⊆ 1 \ b for 1 ≤ j < mi+1 − mi ; that is, di+1 = πb di for i < k. Thus πbk d = π mk d = π n d ⊆ c, while πbi d = di = π mi d ⊆ π mi a0 ⊆ 1 \ c for every i < k, and (πb )c d = π n d = πc d, so that d ∈ D. As a is arbitrary, D is order-dense. Q Q Because πc and (πb )c are both order-continuous Boolean homomorphisms on Ac , and every member of Ac is a supremum of some subfamily of D (313K), they must be equal. (d) Induce on n. We know that there are a non-zero c ⊆ a and an i ≥ 1 such that πb d = π i d for every d ⊆ c. So if n = 1 we can just take a0 = c and m = i. For the inductive step to n + 1, take a0 and m ≥ 1 such that 0 6= a0 ⊆ πb c and πbn d = π m d for every d ⊆ a0 ; then 0 6= πb−1 a0 ⊆ c ⊆ a and πbn+1 d = π m+i d for every d ⊆ πb−1 a0 . (e) If 0 6= a ⊆ b and n ≥ 1, take a non-zero a0 ⊆ a and m ≥ 1 such that π m d = πbn d for every d ⊆ a0 ((d) above). Now (because π is aperiodic) there is a d ⊆ a0 such that π m d 6= d, so πbn d 6= d. As n and a are arbitrary, πb is aperiodic. 387F

Similar ideas lead us to the following fact.

Lemma Let (A, µ ¯) be a totally finite measure algebra, and π : A → A a measure-preserving automorphism; let C be the closed subalgebra {c : πc = c}. Let hdi ii∈I , hei ii∈I be two dist families in A such that µ ¯(c ∩ di ) = µ ¯(c ∩ ei ) for every i ∈ I and c ∈ C. Then there is a φ ∈ Gπ , the full subgroup of Aut A generated by π, such that φdi = ei for every i ∈ I. proof Adding d∗ = 1 \ supi∈I di , e∗ = 1 \ supi∈I ei to the respective families, we may suppose that hdi ii∈I , hei ii∈I are partitions of unity. Define han in∈N inductively by the formula an = supi∈I (di \ supm
387Gc

Dye’s theorem

555

Then an ∩ di ∩ am = 0 whenever m < n and i ∈ I, so han in∈N is dist. Also π n an ⊆ supi∈I ei \ supm
j∈I

n

= π (an ∩ di ) ∩ ei = π (an ∩ di ) for every i ∈ I, n ∈ N. ?? Suppose, if possible, that a = 1 \ supn∈N an is non-zero. Then there is an i ∈ I such that a ∩ di 6= 0. Set c = supn∈N π n (a ∩ di ); then πc ⊆ c so c ∈ C. Now ∞ X

µ ¯(c ∩ ei ∩ π n an ) =

n=0

=

∞ X n=0 ∞ X

µ ¯(c ∩ π n (an ∩ di )) =

∞ X

µ ¯(π n (c ∩ an ∩ di ))

n=0

µ ¯(c ∩ an ∩ di ) = µ ¯(c ∩ di \ a) < µ ¯(c ∩ di ) = µ ¯(c ∩ ei ).

n=0

So b = c ∩ ei \ supn∈N π n an is non-zero, and there is an n ∈ N such that b ∩ π n (a ∩ di ) is non-zero. But look at a0 = π −n (b ∩ π n (a ∩ di )). We have 0 6= a0 ⊆ a ∩ di , so a0 ⊆ di \ supm
556

Automorphism groups

387H

387H Example The following is the basic example of a von Neumann transformation – in a sense, the only example of a measure-preserving von Neumann transformation. Let µ be the usual measure on X = {0, 1}N , Σ its domain, and (A, µ ¯) its measure algebra. Define f : X → X by setting f (x)(n) = 1 − x(n) if x(i) = 0 for every i < n, = x(n) otherwise. Then f is a homeomorphism and a measure space automorphism. P P (i) To see that f is a homeomorphism, perhaps the easiest way is to look at g, where g(x)(n) = 1 − x(n) if x(i) = 1 for every i < n, = x(n) otherwise, and check that f and g are both continuous and that f g, gf are both the identity function. (ii) To see that f is inverse-measure-preserving, it is enough to check that µ{x : f (x)(i) = z(i) for every i ≤ n} = 2−n−1 for every n ∈ N, z ∈ X (254G). But {x : f (x)(i) = z(i) for every i ≤ n} = {x : x(i) = g(z)(i) for every i ≤ n}. (iii) Similarly, g is inverse-measure-preserving, so f is a measure space automorphism. Q Q If n ∈ N, x ∈ X then k

f 2 (x)(n) = 1 − x(n) if n ≥ k and x(i) = 0 whenever k ≤ i < n, = x(n) otherwise. (Induce on k. For the inductive step, observe that if we identify X with {0, 1} × X then f 2 (², y) = (², f (y)) for every ² ∈ {0, 1} and y ∈ X.) Let π : A → A be the corresponding automorphism, setting πE • = f −1 [E]• for E ∈ Σ. Then π is a von Neumann transformation. P P Set En = {x : x ∈ X, x(i) = 1 for every i < n}, an = En• . Then n n f −2 [En+1 ] = {x : x(i) = 1 for i < n, x(n) = 0}, so an+1 , π 2 an+1 split an for each n, and han in∈N witnesses that π is weakly von Neumann. Next, inducing on n, we find that {f −i [En ] : i < 2n } runs over the basic cylinder sets of the form {x : x(i) = z(i) for every i < n} determined by coordinates less than n. Since the equivalence classes of such sets τ -generate A (see part (a) of the proof of 331K), π is a von Neumann transformation. Q Q For another way of looking at the functions f and g, see 445Xq in Volume 4. 387I

We are now ready to approach the main results of this section.

Lemma Let (A, µ ¯) be a totally finite measure algebra and π : A → A an aperiodic measure-preserving automorphism. Let C be the closed subalgebra {c : πc = c}. Then for any a ∈ A there is a b ⊆ a such that µ ¯(b ∩ c) = 21 µ ¯(a ∩ c) for every c ∈ C and πb is a weakly von Neumann transformation, writing πb for the induced automorphism of the principal ideal Ab , as in 387E. Remark On first reading, there is something to be said for supposing here that π is ergodic, that is, that C = {0, 1}. proof Set ²n = 12 (1 + 2−n ) for each n ∈ N, so that h²n in∈N is strictly decreasing, with ²0 = 1 and limn→∞ ²n = 21 . Now there are hbn in∈N , hdni in∈N,i<2n such that, for each n ∈ N, bn+1 ⊆ bn ⊆ a,

µ ¯(bn ∩ c) = ²n µ ¯(a ∩ c) for every c ∈ C,

hdni ii<2n is dist,

supi<2n dni = bn ,

πbn dni = dn,i+1 for every i < 2n − 1, bn+1 ∩ dni = dn+1,i ∪ dn+1,i+2n for every i < 2n . P P Start with b0 = d00 = a. To construct bn+1 and hdn+1,i ii<2n+1 , given hdni ii<2n , note first that (because πbn is measure-preserving and πbn (bn ∩ c) = bn ∩ c) µ ¯(dn0 ∩ c) = µ ¯(dni ∩ c) whenever c ∈ C, i < 2n , so

387J

Dye’s theorem

557

µ ¯(dn0 ∩ c) = 2−n µ ¯(bn ∩ c) = 2−n ²n µ ¯(a ∩ c) for every c ∈ C, and n

dn0 = bn \ supi<2n −1 πbn dni = πbn dn,2n −1 = πb2n dn0 . n

Now πbn is aperiodic (387Ee) so πb2n also is (385A), and there is a dn+1,0 ⊆ dn0 such that n

πb2n dn+1,0 ∩ dn+1,0 = 0,

µ ¯(dn+1,0 ∩ c) = 2−n−1 ²n+1 µ ¯(a ∩ c) for every c ∈ C

n

(applying 385E(iii) to πb2n ¹ Adn0 , with γ = ²n+1 /2²n ). Set dn+1,j = πbjn dn+1,0 for each j < 2n+1 . Ben cause πb2n dn+1,0 ⊆ dn0 \ dn+1,0 , while hπbjn dn0 ij<2n is dist, hπbjn dn+1,0 ij<2n+1 is dist. Set bn+1 = supi<2n+1 πbi n dn+1,0 ; then bn+1 ⊆ bn and µ ¯(bn+1 ∩ c) = ²n+1 µ ¯(a ∩ c) for every c ∈ C. For j < 2n+1 , dn+1,j ⊆ dni where i is either j or j − 2n , so bn+1 ∩ dni = dn+1,i ∪ dn+1,i+2n . For i < 2n+1 − 1, πbn dn+1,i = dn+1,i+1 ⊆ bn+1 , so we must also have πbn+1 dn+1,i = (πbn )bn+1 dn+1,i = dn+1,i+1 (using 387Ec). Thus the induction continues. Q Q Set b = inf n∈N bn ,

eni = b ∩ dni for n ∈ N, i < 2n .

Because hbn in∈N is non-increasing, µ ¯(b ∩ c) = limn→∞ µ ¯(bn ∩ c) = 12 µ ¯(a ∩ c) for every c ∈ C. Next, eni = b ∩ bn+1 ∩ dni = b ∩ (dn+1,i ∪ dn+1,i+2n ) = en+1,i ∪ en+1,i+2n whenever i < 2n . If m ≤ n, j < 2m then bn ∩ dmj = sup{dni : i < 2n , i ≡ j mod 2m } (induce on n). So µ ¯(bn ∩ dmj ) = 2n−m µ ¯dn0 = 2−m ²n ; taking the limit as n → ∞, µ ¯emj = 2−m µ ¯b. Next, πbn (bn ∩ dmj ) = sup{dn,i+1 : i < 2n , i ≡ j mod 2m } = sup{dni : i < 2n , i ≡ j + 1 mod 2m } = bn ∩ dm,j+1 , here interpreting dn,2n as dn0 , dm,2m as dm0 . Consequently πb emj ⊆ em,j+1 . P P?? Otherwise, there are a non-zero e ⊆ dmj ∩ b and k ≥ 1 such that π i e ∩ b = 0 for 1 ≤ i < k and π k e ⊆ b \ dm,j+1 . Take n ≥ m so large that µ ¯e > k µ ¯(bn \ b), so that e0 = e \ sup1≤i
558

Automorphism groups

387J

proof (a) The first fact to note is that if 0 6= b0 ⊆ b, n ∈ Z and π n b0 ⊆ b, then there are m ∈ Z, b00 ⊆ b0 such that b00 6= 0 and π n d = ψ m d for every d ⊆ b00 . P P (α) If n = 0 take b00 = b0 , m = 0. (β) Next, suppose that 0 i n > 0. Take a non-zero b1 ⊆ b such that π b1 is either included in b or dist from b for every i ≤ n (e.g., an atom of the finite subalgebra generated by {b0 } ∪ {π −i b : 0 ≤ i ≤ n}). Enumerate {i : i ≤ n, π i b1 ⊆ b} in ascending order as (l0 , . . . , lk ). Then π lj d = πbj d whenever d ⊆ b1 and j ≤ k (compare part (c) of the proof of 387E); in particular, π n d = πbk d = ψbk d for every d ⊆ b1 . But now 387Ed tells us that there must be a non-zero b00 ⊆ b1 and an m ∈ N such that ψ m d = ψbk d = π n d for every d ⊆ b00 . (γ) If n < 0, then set e0 = π n b0 , so that e0 , π −n e0 ⊆ b. By (β), there are a non-zero e00 ⊆ e0 and an s ∈ N such that π −n d = ψ s d for every d ⊆ e00 . Setting b00 = π −n e00 = ψ s e00 , we have 0 6= b00 ⊆ b0 and π n d = ψ −s d for every d ⊆ b00 . Q Q (b) Now take any non-zero a ∈ A. Then there are m, n ∈ Z such that a1 = a ∩ ψ m b 6= 0, a2 = πa1 ∩ ψ n b 6= 0. Set b1 = ψ −m π −1 a2 . Because ψ ∈ Gπ , there are a non-zero b2 ⊆ b1 and a k ∈ Z such that ψ −n πψ m d = π k d for every d ⊆ b2 . Now π k b2 = ψ −n πψ m b2 ⊆ ψ −n πψ m b1 = ψ −n a2 ⊆ b. By (a), there are a non-zero b3 ⊆ b2 and an r ∈ Z such that π k d = ψ r d for every d ⊆ b3 . Consider a0 = ψ m b3 . Then 0 6= a0 ⊆ ψ m b1 = π −1 a2 ⊆ a1 ⊆ a; and, for d ⊆ a0 , ψ −m d ⊆ b3 ⊆ b2 , so that πd = ψ n (ψ −n πψ m )ψ −m d = ψ n π k ψ −m d = ψ n+r−m d. As a is arbitrary, this shows that π ∈ Gψ , so that Gπ ⊆ Gψ and the two are equal. 387K Lemma Let (A, µ ¯) be a totally finite measure algebra, π : A → A an aperiodic measurepreserving automorphism, and φ any member of the full subgroup Gπ of Aut A generated by π. Suppose that hdmi im≤n,i<2m is a finite dyadic cycle system for φ. Then there is a weakly von Neumann transformation ψ, with dyadic cycle system hd0mi im∈N,i<2m , such that Gψ = Gπ , ψa = φa whenever a ∩ dn0 = 0, and d0mi = dmi whenever m ≤ n and i < 2m . proof Write C for the closed subalgebra {c : πc = c}. By 387I there is a b ⊆ dn0 such that µ ¯(b ∩ c) = 1 µ ¯ (d ∩ c) for every c ∈ C and π : A → A is a weakly von Neumann transformation. Let he ik∈N,i<2k n0 b b b ki 2 be a dyadic cycle system for πb . If we define ψ1 ∈ Aut A by setting ψ1 d = πb d for d ⊆ b,

ψ1 d = π1\b d for d ⊆ 1 \ b,

then ψ1 ∈ Gπ . Next, for any c ∈ C, µ ¯(φ−2 n

n

+1

b ∩ c) = µ ¯φ−2

n

+1

1 2

(b ∩ c) = µ ¯(b ∩ c) = µ ¯(dn0 ∩ c) = µ ¯((dn0 \ b) ∩ c)

n

because φ−2 +1 ∈ Gπ , so φ−2 +1 c = c. By 387F, there is a ψ2 ∈ Gπ such that ψ2 (dn0 \ b) = φ−2 n n ψ3 = φ−2 +1 ψ2−1 φ−2 +1 ψ1 , so that ψ3 ∈ Gπ and ψ3 b = φ−2

n

+1

ψ2−1 φ−2

n

+1

b = φ−2 −2n +1

Thus ψ3 b and ψ2 (dn0 \ b) are dist and have union φ d00 = 1). Accordingly we can define ψ ∈ Gπ by setting

n

+1

n

+1

b. Set

(dn0 \ b).

dn0 = dn1 (if n = 0, we must read d01 as

ψd = ψ3 d if d ⊆ b, = ψ2 d if d ⊆ dn0 \ b, = φd if d ∩ dn0 = 0. Since ψdn0 = dn1 , we have ψdni = φdni for every i < 2n , and therefore ψ i dm0 = dmi whenever m ≤ n n and i < 2m . Looking at ψ 2 , we have

387L

Dye’s theorem n

n

n

ψ 2 dn0 = φ2 dn0 = dn0 , n

so that ψ 2 (dn0 \ b) = b and ψ 2

n+1

ψ 2 b = φ2

559 n

−1

ψ3 b = dn0 \ b,

b = b. Accordingly

ψ2

n+1

d = φ2

n

−1

ψ2 φ2

n

−1

ψ3 d = ψ1 d = π b d

i

for every d ⊆ b. Also supi<2n+1 ψ b = 1, so 387J tells us that Gψ = Gπ . Now define ham im∈N as follows. For m ≤ n, am = dm0 ; for m > n, am = em−n−1,0 . Then for m < n we have m

m

m

ψ 2 am+1 = ψ 2 dm+1,0 = φ2 dm+1,0 = dm+1,2m = am \ am+1 , for m = n we have n

n

n

ψ 2 an+1 = ψ 2 e00 = ψ 2 b = dn0 \ b = an \ an+1 , and for m > n we have m

ψ 2 am+1 = (ψ 2

n+1

)2

m−n−1

em−n,0 = (πb )2

m−n−1

em−n,0

= em−n,2m−n−1 = em−n−1,0 \ em−n,0 = am \ am+1 . Thus ham im∈N witnesses that ψ is a weakly von Neumann transformation. If d0mi = ψ i am for m ∈ N, i < 2m then hd0mi im∈N,i<2m will be a dyadic cycle system for ψ and d0mi = dmi for m ≤ n, as required. 387L Lemma Let (A, µ ¯) be a totally finite measure space and C a closed subalgebra of A such that A is relatively atomless over C. For a ∈ A write Ca = {a ∩ c : c ∈ C}. (a) Suppose that b ∈ A, w ∈ C and δ > 0 are such that µ ¯(b ∩ c) ≥ δ µ ¯c whenever c ∈ C and c ⊆ w. Then there is an e ∈ A such that e ⊆ b ∩ w and µ ¯(e ∩ c) = δ µ ¯c whenever c ∈ Cw . (b) Suppose that k ≥ 1 and that (b0 , . . . , br ) is a finite partition of unity in A. Then there is a partition E of unity in A such that 1 k

µ ¯(e ∩ c) = µ ¯c for every e ∈ E, c ∈ C, #({e : e ∈ E, ∃ i ≤ r, bi ∩ e ∈ / Ce }) ≤ r + 1. proof (a) Set a = b ∩ w and consider the principal ideal Aa generated by A. We know that (Aa , µ ¯¹ Aa ) is a totally finite measure algebra (322H), and that Ca is a closed subalgebra of Aa (333Bc); and it is easy to see that Aa is relatively atomless over Ca . Let θ : Cw → Ca be the Boolean homomorphism defined by setting θc = c ∩ b for c ∈ Cw . If c ∈ Cw and θc = 0, then c ∈ C and δ µ ¯c ≤ µ ¯(c ∩ b) = 0, so c = 0; thus θ is injective; since it is certainly surjective, it is a Boolean isomorphism. We can therefore define a functional ν = µ ¯θ−1 : Ca → [0, ∞[, and we shall have δνd ≤ µ ¯d for every d ∈ Ca . By 331B, there is an e ∈ Aa such that δνd = µ ¯(d ∩ e) for every d ∈ Ca , that is, δµ ¯c = µ ¯(c ∩ e) for every c ∈ Cw , as required. (b)(i) Write D for the set of all those e ∈ A such that µ ¯(c ∩ e) = k1 µ ¯c for every c ∈ C and bi ∩ e ∈ Ce for r+1 every i ≤ r. Then whenever a ∈ A and γ > k is such that µ(a ∩ c) = γµc for every c ∈ C, there is an e ∈ D such that e ⊆ a. P P For d ∈ A, c ∈ C set νd (c) = µ ¯(d ∩ c), so that νd : C → [0, ∞[ is a completely additive functional. For i ≤ r set vi = [[¯ µ¹ C > kνa∩bi ]], in the notation of 326P; so that vi ∈ C and µ ¯c ≥ kµ(a ∩ bi ∩ c) whenever c ∈ C and c ⊆ vi , while µ ¯c ≤ k µ ¯(a ∩ bi ∩ c) whenever c ∈ C and c ∩ vi = 0. Setting v = inf i≤r vi , we have Pr kγ µ ¯v = k µ ¯(a ∩ v) = i=0 kµ(a ∩ bi ∩ v) ≤ (r + 1)¯ µv. Since kγ > r + 1, v = 0. So if we now set wi = (inf j
560

Automorphism groups

µ ¯(c ∩ e) = =

r X i=0 r X i=0

µ ¯(c ∩ ei ) =

r X

387L

µ ¯(c ∩ wi ∩ ei )

i=0

1 µ ¯(c ∩ wi ) k

1 k

= µ ¯c

for every c ∈ C. So e has all the properties required. Q Q (ii) Let E0 ⊆ D be a maximal dist family, and set m = #(E0 ), a = 1 \ sup E0 . Then P m µ ¯(a ∩ c) = µ ¯c − e∈E0 µ µc ¯(c ∩ e) = (1 − )¯ k

r+1 for every c ∈ C, while a does not include any member of D. By (i), 1 − m k ≤ k , that is, m ≥ k − r − 1. 1 Applying (a) repeatedly, with w = 1 and δ = k , we can find dist d0 , . . . , dk−m−1 ⊆ a such that µ ¯(c ∩ di ) = k1 µ ¯c for every c ∈ C and i < k − m. So if we set E = E0 ∪ {di : i < k − m} we shall have a partition of unity with the properties required.

387M Lemma Let (A, µ ¯) be a totally finite measure algebra and π : A → A an aperiodic measurepreserving automorphism. Write C for the closed subalgebra {c : πc = c}. Suppose that φ is a member of the full subgroup Gπ of Aut A generated by π with a finite dyadic cycle system hdmi im≤n,i<2m , and that a ∈ A and ² > 0. Then there is a ψ ∈ Gπ such that (i) ψ has a dyadic cycle system hd0mi im≤k,i<2m , with k ≥ n and d0mi = dmi for m ≤ n, i < 2m ; (ii) ψd = φd if d ∩ dn0 = 0; (iii) there is an a0 in the subalgebra of A generated by C ∪ {d0ki : i < 2k } such that µ ¯(a 4 a0 ) ≤ ². n

proof (a) Take k ≥ n so large that 2k ² ≥ 2n 22 µ ¯1. Let D be the subalgebra of the principal ideal n Adn1 generated by {dn1 ∩ φ−j a : j < 2n }; then D has atoms b0 , . . . , br where r < 22 . (If n = 0, take d01 = d00 = 1.) Applying 387L to the closed subalgebra Cdn1 of Adn1 , we can find a partition of unity E of Adn1 such that µ ¯(e ∩ c) = 2−k µ ¯c for every e ∈ E, c ∈ C E1 = {e : e ∈ E, there is some i ≤ r such that bi ∩ e ∈ / Ce } n

has cardinal at most r + 1 ≤ 22 . Of course µ ¯e = 2−k for every e ∈ E, so #(E) = 2k−n and µ ¯(sup E1 ) ≤ −k 2n −n ∗ 2 2 ≤ 2 ². Write e for sup E1 . (b) For e ∈ E set e0 = φ2 hvi ii<2k−n . Note that

n

−1

e; then {e0 : e ∈ E} is a dist family, of cardinal 2k−n ; enumerate it as supi<2k−n vi = φ2 µ ¯(vi ∩ c) = µ ¯(φ−2

n

n

−1

+1

(sup E) = dn0 ,

vi ∩ c) = 2−k µ ¯c

for every c ∈ C, i < 2k−n . There is therefore a ψ1 ∈ Gπ such that ψ1 vi = φ−2

n

+1

vi+1 for i < 2k−n − 1,

ψ1 v2k−n −1 = φ−2

n

+1

v0

(387F). We have ψ1 dn0 = ψ1 ( sup vi ) = sup ψ1 vi = i<2k−n

= sup φ−2 i<2k−n

i<2k−n

n

+1

vi = φ−2

n

+1

sup

φ−2

n

+1

i<2k−n −1

dn0 = dn1 = φdn0 .

So we may define ψ ∈ Gπ by setting ψd = ψ1 d if d ⊆ dn0 , = φd if d ∩ dn0 = 0.

vi+1 ∪ φ−2

n

+1

v0

387N

Dye’s theorem

561

(c) For each i < 2k−n , n

ψ 2 vi = φ 2

n

−1

ψ1 vi = vi+1

(identifying v2k−n with v0 ). Moreover, ψ j vi ⊆ dnl whenever i < 2k−n and j ≡ l mod 2n . So hψ j v0 ij<2k is a partition of unity in A. What this means is that if we set d0mj = sup{ψ i v0 : i < 2k , i ≡ j mod 2m } for m ≤ k, then hd0mj im≤k,j<2m is a dyadic cycle system for ψ, with d0mj = dmj if m ≤ n, j < 2m . (d) Let B be the subalgebra of A generated by C ∪ {d0kj : j < 2k }. Recall the definition of {vi : i < 2k−n } n as {φ2 −1 e : e ∈ E}; this implies that {ψvi : i < 2k−n } = {ψ1 vi : i < 2k−n } = {φ−2

n

+1

vi : i < 2k−n } = E,

so that {ψ j+1 vi : i < 2k−n } = {φj e : e ∈ E} for j < 2n , and B ⊇ {d0kj : j < 2k } = {ψ j vi : i < 2k−n , j < 2n } = {φj e : e ∈ E, j < 2n }. Set E0 = E \ E1 . For e ∈ E0 and i ≤ r there is a cei ∈ C such that e ∩ bi = e ∩ cei . Set K = {(i, j) : 1 ≤ i ≤ r, j < 2n , bi ⊆ φ−j a}, a0 = sup{φj e ∩ cei : e ∈ E0 , (i, j) ∈ K}. Then a0 is a supremum of (finitely many) of B, so belongs to B. If (i, j) ∈ K and e ∈ E0 , then φj e ∩ cei = φj (e ∩ cei ) = φj (e ∩ bi ) ⊆ a, so a0 ⊆ a. Next, dn1 ∩ φ−j (a \ a0 ) ⊆ e∗ for each j < 2n . P P Set I = {i : i ≤ r, (i, j) ∈ K} = {i : bi ⊆ φ−j a}; then dn1 ∩ φ−j a = supi∈I bi . Now, for each i ∈ I, bi = supe∈E (bi ∩ e) ⊆ supe∈E0 (e ∩ cei ) ∪ e∗ , so that dn1 ∩ φ−j a = supi∈I bi ⊆ supe∈E0 ,i∈I (e ∩ cei ) ∪ e∗ = (dn1 ∩ φ−j a0 ) ∪ e∗ . Q Q But this means that µ ¯(dn,j+1 ∩ a \ a0 ) = µ ¯(dn1 ∩ φ−j (a \ a0 )) ≤ µ ¯e∗ ≤ 2−n ² for every j < 2n (interpreting dn,2n as dn0 , as usual), and P2n µ ¯(a 4 a0 ) = j=1 µ ¯(dnj ∩ a \ a0 ) ≤ ², so that the final condition of the lemma is satisfied. 387N Theorem Let (A, µ ¯) be a totally finite measure algebra, with Maharam type ω, and π : A → A an aperiodic measure-preserving automorphism. Then there is a relatively von Neumann transformation φ : A → A such that φ and π generate the same full subgroups of Aut A. proof (a) The idea is to construct φ as the limit of a sequence hφn in∈N of weakly von Neumann transformations such that Gφn = Gπ . Each φn will have a dyadic cycle system hdnmi im∈N,i<2m ; there will be a strictly increasing sequence hkn in∈N such that dn+1,m,i = dn,m,i whenever m ≤ kn , i < 2m , φn+1 a = φn a whenever a ∩ dn,kn ,0 = 0. Interpolated between the φn will be a second sequence hψn in∈N in Gπ , with associated (finite) dyadic cycle systems hd0nmi im≤kn0 ,i<2m .

562

Automorphism groups

387N

(b) Before starting on the inductive construction we must fix on a countable set B ⊆ A which τ -generates A, and a sequence hbn in∈N in B such that every member of B recurs cofinally often in the sequence. (For instance, take the sequence of first of an enumeration of B × N.) As usual, I write C for the closed subalgebra {c : πc = c}. The induction begins with ψ0 = π, k00 = 0, d0000 = 1. Given ψn ∈ Gπ and its dyadic cycle system hd0nmi im≤kn0 ,i<2m , use 387K to find a weakly von Neumann transformation φn , with dyadic cycle system hdnmi im∈N,i<2m , such that Gφn = Gπ , dnmi = d0nmi for m ≤ kn0 and i < 2m , and φn a = ψn a whenever a ∩ d0n,kn0 ,0 = 0. (c) Given the weakly von Neumann transformation φn , with its dyadic cycle system hdnmi im∈N,i<2m , such that Gφn = Gπ , then we have a partition of unity henj ij∈Z such that πa = φjn a whenever j ∈ Z and a ⊆ enj (387C(ii)). Take rn such that µ ¯eñ ≤ 2−n , where eñ = sup|j|>rn enj , and kn > kn0 such that 2−kn (2rn + 1) ≤ 2−n . Set e∗n = sup|j|≤rn φ−j n dn,kn ,0 , so that µ ¯e∗n ≤ 2−n+1 . 0 Now use 387M to find a ψn+1 ∈ Gπ , with a dyadic cycle system hd0n+1,m,i im≤kn+1 ,i<2m , such that 0 kn+1 ≥ kn , d0n+1,m,i = dnmi if m ≤ kn , ψn+1 a = φn a if a ∩ dn,kn ,0 = 0, and there is a b0n in the algebra 0 generated by C ∪ {d0n+1,m,i : m ≤ kn+1 , i < 2m } such that µ ¯(bn 4 b0n ) ≤ 2−n . Continue. (d) The effect of this construction is to ensure that if l < n in N then dlmi = dnmi whenever m ≤ kl , i < 2m , φn a = φl a whenever a ∩ dl,kl ,0 = 0, b0l belongs to the subalgebra generated by C ∪ {dnmi : m ≤ kn , i < 2m }, and, of course, dn,kn ,0 ⊆ dl,kl ,0 . Since hkn in∈N is strictly increasing, inf n∈N dn,kn ,0 = 0. Now, for each n ∈ N, dn,kn ,1 = φn dn,kn ,0 = φn+1 dn,kn ,0 ⊇ φn+1 dn+1,kn+1 ,0 = dn+1,kn+1 ,1 , so setting a0 = 1 \ d0,k0 ,0 ,

an+1 = dn,kn ,0 \ dn+1,kn+1 ,0 for each n,

we have φ0 a0 = 1 \ d0,k0 ,1 ,

φn+1 an+1 = dn,kn ,1 \ dn+1,kn+1 ,1 for each n,

and hφn an in∈N is a partition of unity. There is therefore a φ ∈ Aut A defined by setting φa = φn a if a ⊆ an ; because Gπ is full, φ ∈ Gπ . (e) If m ≤ n, then am ∩ dn,kn ,0 = 0, so φn a = φm a = φa for every a ⊆ am . Thus φn a = φa for every a ⊆ supm≤n am = 1 \ dn,kn ,0 . In particular, φdnmi = dn,m,i+1 whenever m ≤ kn , 1 ≤ i < 2m (counting dn,m,2m as dnm0 , as usual); so that in fact φdnmi = dn,m,i+1 whenever m ≤ kn , i < 2m . For each n, we have dnmi = d0nmi = dn+1,m,i whenever m ≤ kn and i < 2m . We therefore have a family ∗ hdmi im∈N,i<2m defined by saying that d∗mi = dnmi whenever n ∈ N, m ≤ kn and i < 2m . Now, for any m ∈ N, there is a kn > m, so that hd∗mi ii<2m = hdnmi ii<2m is a partition of unity; and d∗mi = dnmi = dn,m+1,i ∪ dn,m+1,i+2m = d∗m+1,i ∪ d∗m+1,i+2m for each i < 2m . Moreover, φd∗m,i = φn dnmi = dn,m,i+1 = d∗m,i+1 at least for 1 ≤ i < 2m (counting d∗m,2m as d∗m,0 , as usual), so that in fact φd∗mi = d∗m,i+1 for every i < 2m . Thus hd∗mi im∈N,i<2m is a dyadic cycle system for φ, and φ is a weakly von Neumann transformation. Writing B for the closed subalgebra of A generated by C ∪ {d∗mi : m ∈ N, i < 2m }, then C ∪ {d0nmi : m ≤ kn0 , i < 2m } = C ∪ {dn+1,m,i : m ≤ kn0 , i < 2m } = C ∪ {d∗mi : m ≤ kn0 , i < 2m } ⊆ B

387O

Dye’s theorem

563

for any n ∈ N. So b0n ∈ B for every n. If b ∈ B and ² > 0, there is an n ∈ N such that 2−n ≤ ² and bn = b, so that µ ¯(b 4 b0n ) ≤ ²; as every b0n belongs to B, and B is closed, b ∈ B; as b is arbitrary, and B τ -generates A, B = A. Thus φ is a von Neumann transformation. (f ) If n ∈ N and d ∩ e∗n = 0, then φj d = φjn d and φ−j d = φ−j P Induce on n d whenever 0 ≤ j ≤ rn . P j. For j = 0 the result is trivial. For the inductive step to j + 1 ≤ rn , note that if d0 ∩ dn,kn ,1 = 0 then 0 φ−1 n d ∩ dn,kn ,0 = 0, so 0 −1 0 −1 0 φ−1 d0 = φ−1 φn (φ−1 φ(φ−1 n d )=φ n d ) = φn d .

Now we have φj+1 d = φ(φjn d) = φn (φjn d) = φj+1 n d because φjn d ∩ dn,kn ,0 = φjn (d ∩ φ−j n dn,kn ,0 ) = 0, while −1 −j −j−1 φ−j−1 d = φ−1 (φ−j d n d) = φn (φn d) = φn

because −j j+1 φ−j Q n d ∩ dn,kn ,1 = φn (d ∩ φn dn,kn ,0 ) = 0. Q

Thus φj d = φjn d whenever |j| ≤ rn . (g) Finally, Gφ = Gπ . P P I remarked in (d) that φ ∈ Gπ , so that Gφ ⊆ Gπ . To see that π ∈ Gφ , take any non-zero a ∈ A. Because µ ¯(e∗n ∪ eñ ) ≤ 2−n+1 for each n, there is an n such that a0 = a \ (e∗n ∪ eñ ) 6= 0. Now there is some j ∈ Z such that a00 = a0 ∩ enj 6= 0; since a0 ∩ eñ = 0, |j| ≤ rn . If d ⊆ a00 , then πd = φjn d, by the definition of enj . But also φjn d = φj d, by (f), because d ∩ e∗n = 0. So πd = φj d for every d ⊆ a00 . As a is arbitrary, π ∈ Gφ and Gπ ⊆ Gφ . Q Q This completes the proof. 387O Theorem Let (A1 , µ ¯1 ) and (A2 , µ ¯2 ) be totally finite measure algebras of countable Maharam type, and π1 : A1 → A1 , π2 : A2 → A2 measure-preserving automorphisms. For each i, let Ci be the closed subalgebra {c : c ∈ Ai , πi c = c} and Gπi the full subgroup of Aut Ai generated by πi . If (A1 , µ ¯1 , C1 ) and (A2 , µ ¯2 , C2 ) are isomorphic, so are (A1 , µ ¯1 , Gπ1 ) and (A2 , µ ¯2 , Gπ2 ). proof (a) It is enough to consider the case in which (A1 , µ ¯1 , C1 ) and (A2 , µ ¯2 , C2 ) are actually equal; I therefore delete the subscripts and speak of a structure (A, µ ¯, C), with two automorphisms π1 , π2 of A both with fixed-point subalgebra C. (b) Suppose first that A is relatively atomless over C, that is, that both the πi are aperiodic (385E). In this case, 387N tells us that there are von Neumann transformations φ1 and φ2 of A such that Gπ1 = Gφ1 and Gπ2 = Gφ2 . But (A, µ ¯, φ1 ) and (A, µ ¯, φ2 ) are isomorphic. P P Let hdmi im∈N,i<2m and hd0mi im∈N,i<2m be dyadic cycle systems for φ1 , φ2 respectively such that C ∪ {dmi : m ∈ N, i < 2m } and C ∪ {d0mi : m ∈ N, i < 2m } both τ -generate A. Writing B1 , B2 for the subalgebras of A generated by C ∪ {dmi : m ∈ N, i < 2m } and C ∪ {d0mi : m ∈ N, i < 2m } respectively, it is easy to see that these algebras are isomorphic: we just set θ0 c = c for c ∈ C, θ0 dmi = d0mi for i < 2m to obtain a measure-preserving isomorphism θ0 : B1 → B2 . Because these are topologically dense subalgebras of A, there is a unique extension of θ0 to a measure-preserving automorphism θ : A → A (324O). Next, we see that θφ1 θ−1 c = c = φ2 c for every c ∈ C, θφ1 θ−1 d0mi = θφ1 dmi = θdm,i+1 = d0m,i+1 = φ2 d0m,i+1 for m ∈ N, i < 2m (as usual, taking dm,2m to be dm0 and d0m,2m to be d0m0 ). But this means that θφ1 θ−1 b = φ2 b for every b ∈ B2 , so (again because B2 is dense in A) θφ1 θ−1 = φ2 . Thus θ is an isomorphism between (A, µ ¯, φ1 ) and (A, µ ¯, φ2 ). Q Q Of course θ is now also an isomorphism between (A, µ ¯, Gφ1 ) = (A, µ ¯, Gπ1 ) and (A, µ ¯, Gφ2 ) = (A, µ ¯, Gπ2 ).

564

Automorphism groups

387O

(c) Next, consider the case in which π1 is periodic, with period n, for some n ≥ 1. In this case π2 ∈ Gπ1 . P P Let (d0 , . . . , dn−1 ) be a partition of unity in A such that π1 di = di+1 for i < n − 1 and π1 dn−1 = d0 (385B). If d ⊆ dj , then c = supi
π2 d = dj ∩ c = π1j−i d.

As a is arbitrary, this shows that π2 ∈ Gπ1 . Q Q Now supn∈Z π2n d0 belongs to C and includes d0 , so must be 1. Finally, if we look at the two induced automorphisms (π1 )d0 , (π2 )d0 on Ad0 , they must both be the identity, by 387Eb, because every element of Ad0 is of the form d0 ∩ c for some c ∈ C. So 387J tells us that Gπ1 = Gπ2 . (d) For the general case, we see from 385C that there is a partition of unity hci i1≤i≤ω in C such that π1 ¹ Acω is aperiodic and if i is finite and ci 6= 0 then π1 ¹ Aci is periodic with period i. For each i, let Hi be {φ¹ Aci : φ ∈ Gπ1 }; then Hi is a full subgroup of Aut Aci , and Gπ1 = {φ : φ ∈ Aut A, φ¹ Aci ∈ Hi whenever 1 ≤ i ≤ ω}. Similarly, writing

Hi0

= {φ¹ Aci : φ ∈ Gπ2 }, Gπ2 = {φ : φ ∈ Aut A, φ¹ Aci ∈ Hi0 whenever 1 ≤ i ≤ ω}.

Note also that Hi , Hi0 are the full subgroups of Aut Aci generated by π1 ¹ Aci , π2 ¹ Aci respectively. By (b) and (c), Hi = Hi0 for finite i, while there is a measure-preserving automorphism θ : Acω → Acω such that θHω θ−1 = Hω0 . Now we can define a measure-preserving automorphism θ1 : A → A by setting θ1 a = θa if a ⊆ cω , θ1 a = a if a ∩ cω = 0, and we shall have θ1 Gπ1 θ1−1 = Gπ2 . Thus (A, µ ¯, Gπ1 ) and (A, µ ¯, Gπ2 ) are isomorphic, as claimed. 387X Basic exercises > (a) Let (X, Σ, µ) be a localizable measure space, with measure algebra (A, µ ¯). Suppose that π and φ are automorphisms of A, and that π is represented (in the sense of 343A) by a measure space automorphism f : X → X. Show that the following are equiveridical: (i) φ belongs to the full subgroup of Aut A generated by π; (ii) there is a function g : X → X, representing φ, such that g(x) ∈ {f n (x) : n ∈ Z} for every x ∈ X; (iii) there is a measure space automorphism g : X → X, representing φ, such that {g n (x) : n ∈ Z} ⊆ {f n (x) : n ∈ Z} for every x ∈ X. (b) Let (X, Σ, µ) be a totally finite measure space, with measure algebra (A, µ ¯); let f : X → X be a measure space isomorphism, and π : A → A the corresponding measure-preserving automorphism. Take E ∈ Σ and b = E • ∈ A; let πb : Ab → Ab be the corresponding induced automorphism, as in 387E. (i) Set E1 = {x : x ∈ E, ∃ n ≥ 1, f n (x) ∈ E}; show that µ(E \ E1 ) = 0, so that E1• = b. (ii) For x ∈ E1 set nx = min{n : n ≥ 1, f nx (x) ∈ E}, gE (x) = f nx (x). Show that gE : E1 → E induces πb : Ab → Ab . (iii) Show that there is a measurable set G ⊆ E1 such that E1T\ G is S negligible and gE ¹G = gG is an automorphism for the induced measure on G. (Hint: try G = E ∩ n∈N m,r≥n f m [E] ∩ f −r [E].) (iv) Show n that, in (iii), {gG (x) : n ∈ Z} = G ∩ {f n (x) : n ∈ Z} for every x ∈ G. (v) Find an expression of 387Ec in this language. (c) Let (A, µ ¯) be a totally finite measure algebra, and π : A → A a measure-preserving automorphism. Let us say that a pseudo-cycle for π is a partition of unity hai ii
387 Notes

Dye’s theorem

565

(e) Let (A, µ ¯) be an atomless probability algebra of countable Maharam type, and Autµ¯ A the group of measure-preserving Boolean automorphisms of A. Let π ∈ Autµ¯ A be a von Neumann transformation. (i) Show that for any ultrafilter F on N there is a φF ∈ Autµ¯ A defined by the formula φF (a) = limn→F π n a for every a ∈ A, the limit being taken in the measure-algebra topology. (ii) Show that {φF : F is an ultrafilter on N} is a subgroup of Autµ¯ A homeomorphic to ZN 2 . (Hint: 387H.) (f ) Let A be a Boolean algebra and π : A → A a weakly von Neumann automorphism. Show that π n is a weakly von Neumann automorphism for every n ∈ Z \ {0}. (Hint: consider n = 2, n = −1, odd n ≥ 3 separately. The formula of 387H may be useful.) (g) Let A be a Boolean algebra and π : A → A a von Neumann automorphism. (i) Show that π is ergodic (in the sense that its fixed-point algebra is trivial) but π 2 is not ergodic. (ii) Show that π 2 is relatively von Neumann. (iii) Show that π n is von Neumann for every odd n ∈ Z \ {0}. 387Y Further exercises (a) Let A be a Dedekind complete Boolean algebra and G a semigroup of order-continuous Boolean homomorphisms from A to itself. Let us say that G is full if whenever φ : A → A is an order-continuous Boolean homomorphism, and there is a partition of unity hai ii∈I in A such that for every i ∈ I there is a πi ∈ G such that φa = πi a for every a ⊆ ai , then φ ∈ G. Show that if φ and π are order-continuous Boolean homomorphisms from A to itself, then the following are equiveridical: (i) φ belongs to the full semigroup generated by φ; (ii) for every non-zero a ∈ A there are a non-zero b ⊆ a and an n ∈ N such that φd = π n d for every d ⊆ b; (iii) there is a partition of unity han in∈N in A such that φa = π n a for every n ∈ N, a ⊆ an . (b) Let A be a Dedekind complete Boolean algebra and Z its Stone space. Let π, φ : A → A be Boolean automorphisms, and f , g : Z → Z the corresponding homeomorphisms. Show that the following are equiveridical: (i) φ belongs to the full subgroup of Aut A generated by π; (ii) {z : g(z) ∈ / {f n (z) : n ∈ Z}} n n is nowhere dense in Z; (iii) {z : {g (z) : n ∈ Z} 6⊆ {f (z) : n ∈ Z}} is nowhere dense in Z. (c) Let X be a set, Σ a σ-algebra of subsets of X, and I a σ-ideal of Σ such that the quotient algebra A = Σ/I is Dedekind complete and there is a countable subset of Σ separating the points of X. Suppose that f and g are automorphisms of the structure (X, Σ, I) inducing π, φ ∈ Aut A. Show that the following are equiveridical: (i) φ belongs to the full subgroup of Aut A generated by π; (ii) {x : x ∈ X, f (x) ∈ / {g n (x) : n n n ∈ Z}} ∈ I; (iii) {x : x ∈ X, {f (x) : n ∈ Z} 6⊆ {g (x) : n ∈ Z}} ∈ I. (d) Let (A, µ ¯) be a totally finite measure algebra and π : A → A a relatively von Neumann transformation. Show that π is aperiodic and has zero entropy. (e) Let (A1 , µ ¯1 ) and (A2 , µ ¯2 ) be atomless probability algebras, and (A, µ ¯) their probability algebra free product. Let π1 : A1 → A1 be a measure-preserving von Neumann automorphism and π2 : A2 → A2 a mixing measure-preserving automorphism. Let π : A → A be the measure-preserving automorphism such that π(a1 ⊗ a2 ) = π1 (a1 ) ⊗ π2 (a2 ) for all a1 ∈ A1 , a2 ∈ A2 . Show that π is an ergodic weakly von Neumann automorphism which is not a relatively von Neumann automorphism. (f ) Let µ be Lebesgue measure on [0, 1]2 , and (A, µ ¯) its measure algebra; let C be the closed subalgebra of elements expressible as (E × [0, 1])• , where E ⊆ [0, 1] is measurable. Suppose that π : A → A is a measurepreserving automorphism such that C = {c : πc = c}. Show that there is a family hfx ix∈[0,1] of ergodic measure space automorphisms of [0, 1] such that (x, y) 7→ (x, fx (y)) is a measure space automorphism of [0, 1]2 representing π. 387 Notes and comments Dye’s theorem (Dye 59) is actually Theorem 387O in the case in which π1 , π2 are ergodic, that is, in which C1 and C2 are both trivial. I take the trouble to give the generalized form here (a simplified version of that in Krieger 76) because it seems a natural target, once we have a classification of the relevant structures (A, µ ¯, C) (333R). The essential mathematical ideas are the same in both cases. You can find the special case worked out in Hajian Ito & Kakutani 75, from which I have taken the argument used here; and you may find it useful to go through the version above, to check what kind of

566

Automorphism groups

387 Notes

simplifications arise if each C is taken to be {0, 1}. Essentially the difference will be that every ‘aperiodic’ turns into ‘ergodic’ (with an occasional ‘atomless’ thrown in) and ‘331B’ turns into ‘331C’. As far as I know, there is no simplification available in the structure of the argument; of course the details become a bit easier, but with the possible exception of 387L-387M I think there is little difference. Of course modifying a general argument to give a simpler proof of a special case is a standard exercise in this kind of mathematics. What is much more interesting is the reverse process. What kinds of theorem about ergodic automorphisms will in fact be true of all automorphisms? A variety of very powerful approaches to such questions have been developed in the last half-century, and I hope to describe some of the ideas in Volumes 4 and 5. The methods used in this section are relatively straightforward and do not require any deep theoretical underpinning beyond Maharam’s lemma 331B. But an alternative approach can be found using 387Yf: in effect (at least for the Lebesgue measure algebra) any measure-preserving automorphism can be disintegrated into ergodic measure space automorphisms (the fibre maps fx of 387Yf). It is sometimes possible to guess which theorems about ergodic transformations are ‘uniformisable’ in the sense that they can be applied to such a family hfx ix∈[0,1] , in a systematic way, to provide a structure which can be interpreted on the product measure. The details tend to be complex, which is one of the reasons why I do not attempt to work through them here; but such disintegrations can be a most valuable aid to intuition. Dye’s theorem itself is less significant than many results which I have omitted; I include it partly because it is relatively easy and partly because some of the ideas on the way are very important indeed. In particular, the concept of ‘recurrence’ is vital. I have already offered a couple of relevant exercises (382Yc-382Yd). The fundamental lemmas introduced in this section are 387E and 387F. In this section I use von Neumann transformations as an auxiliary tool: the point is, first, that two von Neumann transformations are isomorphic – that is, the von Neumann transformations on a given totally finite measure algebra (A, µ ¯) (necessarily isomorphic to the Lebesgue measure algebra, since we must have A atomless and τ (A) = ω) form a conjugacy class in the group Autµ¯ A of measure-preserving automorphisms; and next, that for any ergodic measure-preserving automorphism π (on an atomless totally finite algebra of countable Maharam type) there is a von Neumann transformation φ such that Gπ = Gφ (387N). But I think they are remarkable in themselves. A (weakly) von Neumann transformation has a ‘pseudo-cycle’ (387Xc) for every power of 2. For some purposes, existence is all we need to know; but in the arguments of 387K-387N we need to keep track of named pseudo-cycles in what I call ‘dyadic cycle systems’ (387G). In this volume I have systematically preferred arguments which deal directly with measure algebras, rather than with measure spaces. I believe that such arguments can have a simplicity and clarity which repays the extra effort of dealing with more abstract structures. But undoubtedly it is necessary, if you are to have any hope of going farther in the subject, to develop methods of transferring intuitions and theorems between the two contexts. I offer 387Xb as an example. The description there of ‘induced automorphism’ requires a certain amount of manoeuvering around negligible sets, but in 387Xb(iv) gives a valuably graphic description. In the same way, 387C, 387Xa and 387Yb provide alternative ways of looking at full subgroups. There are contexts in which it is useful to know whether an element of the full subgroup generated by π actually belongs to the full semigroup generated by π (387Ya); for instance, this happens in 387F.

391C

Kelley’s theorem

567

Chapter 39 Measurable algebras In the final chapter of this volume, I present results connected with the following question: which algebras can appear as the underlying Boolean algebras of measure algebras? Put in this form, there is a trivial answer (391A). The proper question is rather: which algebras can appear as the underlying Boolean algebras of semi-finite measure algebras? This is easily reducible to the question: which algebras can appear as the underlying Boolean algebras of probability algebras? Now in one sense Maharam’s theorem (§332) gives us the answer exactly: they are the countable simple products of the measure algebras of {0, 1}κ for cardinals κ. But if we approach from another direction, things are more interesting. Probability algebras share a very large number of very special properties. Can we find a selection of these properties which will be sufficient to force an abstract Boolean algebra to be a probability algebra when endowed with a suitable functional? No fully satisfying answer to this question is known. But in exploring the possibilities we encounter some interesting and important ideas. In §391 I discuss algebras which have strictly positive additive realvalued functionals; for such algebras, weak (σ, ∞)-distributivity is necessary and sufficient for the existence of a measure; so we are led to look for conditions sufficient to ensure that there is a strictly positive additive functional. A slightly different approach lies through the concept of ‘submeasure’. Submeasures arise naturally in the theories of topological Boolean algebras, topological Riesz spaces and vector measures (see the second half of §393), and on any given algebra there is a strictly positive ‘uniformly exhaustive’ submeasure iff there is a strictly positive additive functional; this is the Kalton-Roberts theorem (392F). It is unknown whether the word ‘uniformly’ can be dropped; this is one of the forms of the Control Measure Problem, which I investigate at length in §393. In §394, I look at a characterization in of the special properties which the automorphism group of a measure algebra must have (Kawada’s theorem, 394Q). §395 complements the previous section by looking briefly at the subgroups of an automorphism group Aut A which can appear as groups of measure-preserving automorphisms.

391 Kelley’s theorem In this section I introduce the notion of ‘measurable algebra’ (391B), which will be the subject of the whole chapter once the trivial construction of 391A has been dealt with. I show that for weakly (σ, ∞)distributive algebras countable additivity can be left to look after itself, and all we need to find is a strictly positive finitely additive functional (391D). I give Kelley’s criterion for the existence of such a functional (391H-391J), and a version of Gaifman’s example of a ccc Boolean algebra without such a functional (391N). 391A Proposition Let A be any Dedekind σ-complete Boolean algebra. Then there is a function µ ¯ : A → [0, ∞] such that (A, µ ¯) is a measure algebra. proof Set µ ¯0 = 0, µ ¯a = ∞ for a ∈ A \ {0}. 391B Definition I will call a Boolean algebra A measurable if there is a functional µ ¯ : A → [0, ∞[ such that (A, µ ¯) is a totally finite measure algebra. In this case, if µ ¯ 6= 0, then it has a scalar multiple with total mass 1. So a Boolean algebra A is measurable iff either it is {0} or there is a functional µ ¯ such that (A, µ ¯) is a probability algebra. 391C Proposition Let A be a Boolean algebra. (a) The following are equiveridical: (i) there is a functional µ ¯ : A → [0, ∞] such that (A, µ ¯) is a semi-finite measure algebra; (ii) A is Dedekind σ-complete and {a : a ∈ A, Aa is measurable} is order-dense in A, writing Aa for the principal ideal generated by a. (b) The following are equiveridical: (i) there is a functional µ ¯ : A → [0, ∞] such that (A, µ ¯) is a localizable measure algebra; (ii) A is Dedekind complete and {a : a ∈ A, Aa is measurable} is order-dense in A. proof (a) (i)⇒(ii): if (A, µ ¯) is a semi-finite measure algebra, then Af = {a : µ ¯a < ∞} is order-dense in A and Aa is measurable for every a ∈ Af .

568

Measurable algebras

391C

(ii)⇒(i): setting D = {a : a ∈ A, Aa is measurable}, D is order-dense, so there is a partition of unity C ⊆P D (313K). For each c ∈ C, choose µ ¯c such that (Ac , µ ¯c ) is a totally finite measure algebra. Set µ ¯a = c∈C µ ¯c (a ∩ c) for every a ∈ A; then it is easy to check that (A, µ ¯) is a semi-finite measure algebra. (b) Follows immediately. 391D Theorem Let A be a Boolean algebra. Then the following are equiveridical: (i) A is measurable; (ii) A is Dedekind σ-complete and weakly (σ, ∞)-distributive, and there is a strictly positive finitely additive functional ν : A → [0, ∞[. Remark An additive functional ν on a Boolean algebra A is strictly positive if νa > 0 for every non-zero a ∈ A. proof (i)⇒(ii) Put the definition together with 322C(b)-(c) (for Dedekind completeness) and 322F (for weak (σ, ∞)-distributivity). (ii)⇒(i) Given that (ii) is satisfied, let µ ¯ be the completely additive part of ν, as defined in 326Yq: that is, µ ¯a = inf{supd∈D νd : D is a non-empty upwards-directed set with supremum a} for every a ∈ A. Then (A, µ ¯) is a totally finite measure algebra. α) Of course 0 ≤ µ P P(α ¯a ≤ ν1 for every a ∈ A, so µ ¯ is a function from A to [0, ∞[. β ) Suppose that han in∈N is a dist sequence in A with supremum a. Let ² > 0. For each n, (β choose an upwards-directed set Dn , with supremum an , such that supd∈Dn νd ≤ µ ¯an + 2−n ²; choose an upwards-directed set D, with supremum a, such that supd∈D νd ≤ µ ¯a + ². Set D∗ = {supi≤n di : n ∈ N, di ∈ Di for every i ≤ n}. Then D∗ is upwards-directed, d ⊆ a for every d ∈ D∗ , and any upper bound for D∗ must be an upper bound for Dn for every n, so sup D∗ = a. This means that µ ¯a ≤ sup νd = sup{ν(sup di ) : di ∈ Di for i ≤ n} d∈D ∗

i≤n

= sup{

n X

νdi : di ∈ Di for i ≤ n}

i=0

(because di ∩ dj ⊆ ai ∩ aj = 0 if i 6= j, di ∈ Di , dj ∈ Dj ) ∞ ∞ ∞ X X X −i = sup νd ≤ µ ¯ai + 2 ² = 2² + µ ¯ ai . i=0 d∈Di

i=0

i=0

Next, set Dn0 = {an ∩ d : d ∈ D} for each n. Then Dn0 is upwards-directed and has supremum an ∩ a = an (313Ba), so P∞ P∞ Pn ¯ai ≤ i=0 supd∈Di0 νd = sup{ i=0 ν(di ∩ ai ) : n ∈ N, di ∈ D for i ≤ n}. i=0 µ But, given n ∈ N and di ∈ D for i ≤ n, there is a d ∈ D such that supi≤n di ⊆ d (because D is upwardsdirected), so that (because a0 , . . . , an are dist) Pn ¯a + ². i=0 ν(di ∩ ai ) ≤ νd ≤ µ P∞ Accordingly i=0 µ ¯aP ¯a + ². i ≤µ ∞ As ² is arbitrary, i=0 µ ¯ ai = µ ¯a; as han in∈N is arbitrary, µ ¯ is countably additive. (γγ ) Now suppose that a ∈ A and µ ¯a = 0. Then for every n ∈ N there is an upwards-directed set Dn , with supremum a, such that νd ≤ 2−n for every d ∈ Dn . Set Bn = {a \ d : d ∈ Dn }, so that each Bn is downwards-directed and has infimum 0 (313A). Because A is weakly (σ, ∞)-distributive, there is a

391E

Kelley’s theorem

569

set B ⊆ A, also with infimum 0, such that for every b ∈ B, n ∈ N there is a b0 ∈ Bn with b0 ⊆ b. If B = ∅ then A = {0} and surely a = 0. Otherwise, set D = {a \ b : b ∈ B}. Then D is upwards-directed and has supremum a. But if d ∈ D, then for every n ∈ N there is a b0 ∈ Bn such that b0 ⊆ a \ d, that is, d ⊆ a \ b0 ∈ Dn , and νd ≤ 2−n . Accordingly νd = 0; but we are supposing that ν is strictly positive, so this means that D = {0} and a must be 0. This shows that µ ¯ is strictly positive. (δδ ) Thus µ ¯ is a strictly positive countably additive functional; as we are also assuming that A is Dedekind σ-complete, (A, µ ¯) is a totally finite measure algebra. Q Q Accordingly A is measurable. 391E Thus we are led naturally to the question: which Boolean algebras carry strictly positive finitely additive functionals? The Hahn-Banach theorem, suitably applied, gives some sort of answer to this question. For the sake of applications later on, I give two general results on the existence of additive functionals related to given functionals. Theorem Let A be a Boolean algebra, not {0}, and φ : A → [0, 1] a functional. Then the following are equiveridical: (i) there is a finitely additive functional ν : A → [0, 1] such P that ν1 = 1 and νa ≤ φa for every a ∈ A; (ii) whenever hai iP i∈I is a finite indexed family in A, and i∈I χai ≥ mχ1 in S = S(A) (definition: 361A), where m ∈ N, then i∈I φai ≥ m. proof (a)(i)⇒(ii) If ν : A → [0, 1] is a finitely additive functional such that ν1 = 1 and νa ≤ φa for every a ∈ A, let h : S → P R be the positive linear functional corresponding to ν (361G). Now if hai ii∈I is a finite family in A and i∈I χai ≥ mχ1, then X X X φai ≤ νai = h(χai ) i∈I

i∈I

i∈I

X = h( χai ) ≥ h(mχ1) = m. i∈I

As hai ii∈I is arbitrary, (ii) is true. (b)(ii)⇒(i) Now suppose that φ satisfies (ii). For u ∈ S, set Pn Pn p(u) = inf{ i=0 αi φai : a0 , . . . , an ∈ A, α0 , . . . , αn ≥ 0, i=0 αi χai ≥ u}. Then it is easy to check that p(u + v) ≤ p(u) + p(v) for all u, v ∈ S, and that p(αu) = αp(u) for u ∈ S, Pall n α ≥P 0. Also p(χ1) ≥ 1. P P?? If not, there are a0 , . . . , an ∈ A and α0 , . . . , αn ≥ 0 such that χ1 ≤ i=0 αi χai n but i=0 αi φai < 1. Increasing each αi slightly if necessary, we may suppose that every αi is rational; let m ≥ 1 and k0 , . . . , kn ∈ N be such that αi = ki /m for each i ≤ n. Set K = {(i, j) : 0 ≤ i ≤ n, 1 ≤ j ≤ ki }, and for (i, j) ∈ K set aij = ai . Then Pn Pn P i=0 αi χai ≥ mχ1, (i,j)∈K χaij = i=0 ki χai = m but

P (i,j)∈K

φaij =

Pn i=0

ki φai = m

Pn i=0

αi φai < m,

which is supposed to be impossible. X XQ Q By the Hahn-Banach theorem, in the form 3A5Aa, there is a linear functional h : S → R such that h(χ1) = p(χ1) ≥ 1 and h(u) ≤ p(u) for every u ∈ S. In particular, h(χa) ≤ φb whenever a ⊆ b ∈ A. Set νa = h(χa) for a ∈ A; then ν : A → [0, ∞[ is an additive functional, ν1 ≥ 1 and νa ≤ φb whenever a ⊆ b in A. We do not know whether ν is positive, but if we define ν + as in 362Ab, we shall have a non-negative additive functional such that ν + a = supb ⊆ a νb ≤ φa for every a ∈ A, and 1 ≤ ν1 ≤ ν + 1 ≤ φ1 = 1, so ν + witnesses that (i) is true.

570

Measurable algebras

391F

391F Theorem Let A be a Boolean algebra, not {0}, and ψ : A → [0, 1] a functional, where A ⊆ A. Then the following are equiveridical: (i) there is a non-negative finitely additive functional ν : A → [0, 1] such that ν1 = 1 and νa ≥ ψa for every a ∈ A; P (ii) whenever hai ii∈I is a finite indexed family in A, there is a set J ⊆ I such that #(J) ≥ i∈I ψai and inf i∈J ai 6= 0. Remark In (ii) here, we may have to interpret the infimum of the empty set in A as 1. proof We apply 391E to the functional φ, where φa = 1 − ψ(1 \ a) for a ∈ A. (a) If ν : A → [0, 1] is a non-negative finitely additive functional such that ν1 = 1, then νa ≥ ψa ⇐⇒ 1 − νa ≤ 1 − ψa ⇐⇒ ν(1 \ a) ≤ φ(1 \ a). So (i) here is true of ψ iff 391E(i) is true of φ. (b)P Suppose that (ii) here is true of ψ, and that hai ii∈I is a finite family in A such that while i∈I φai = β. Then P P i∈I ψ(1 \ ai ) = i∈I (1 − φai ) = #(I) − β,

P i∈I

χai ≥ mχ1,

so there is a set J ⊆ I such that #(J) ≥ #(I) − β and inf i∈J (1 \ ai ) = c 6= 0. Now c ∩ ai = 0 for i ∈ J, so P P mχc ≤ i∈I χ(ai ∩ c) = i∈I\J χ(ai ∩ c) ≤ #(I \ J)χc and m ≤ #(I) − #(J) ≤ β. As hai ii∈I is arbitrary, 391E(ii) is true of φ. (c) Suppose that 391E(ii) is true of φ, and that hai ii∈I is a finite family in A. Set P P β = i∈I φ(1 \ ai ) = #(I) − i∈I ψai P P and i∈I χ(1 \ ai ) 6≥ kχ1, that is, i∈I φ(1 \ ai ) < k, P let k be the least integer greater than β. Since χa ≤ 6 (#(I) − k)χ1. But this means that there must be some J ⊆ I such that #(J) > #(I) − k and i i∈I inf i∈J ai 6= 0. Now P i∈I ψai = #(I) − β ≤ #(I) − (k − 1) ≤ #(J). As hai ii∈I is arbitrary, (ii) here is true of ψ. (d) Since we know that 391E(i) ⇐⇒ 391E(ii), we can conclude that (i) and (ii) here are equivalent. 391G Corollary Let A be a Boolean algebra, B a subalgebra of A, and ν0 : B → R a non-negative finitely additive functional. Then there is a non-negative finitely additive functional ν : A → R extending ν0 . proof (a) Suppose first that ν0 1 = 1. Set ψb = ν0 b for every b ∈ B. Then ψ must satisfy the condition (ii) of 391F when regarded as a functional defined on a subset of B; but this means that it satisfies the same condition when regarded as a functional defined on a subset of A. So there is a non-negative finitely additive functional ν : A → R such that ν1 = 1 and νb ≥ ν0 b for every b ∈ B. In this case νb = 1 − ν(1 \ b) ≤ 1 − ν0 (1 \ b) = ν0 b ≤ νb for every b ∈ B, so ν extends ν0 . (b) For the general case, if ν0 1 = 0 then ν0 must be the zero functional on B, so we can take ν to be the zero functional on A; and if ν0 1 = γ > 0, we apply (a) to γ −1 ν0 . 391H Definition Let A be a Boolean algebra, and A ⊆ A \ {0} any non-empty set. The intersection number of A is the largest δ ≥ 0 such that whenever hai ii∈I is a finite family in A, with I 6= ∅, there is a J ⊆ I such that #(J) ≥ δ#(I) and inf i∈J ai 6= 0. Remarks (a) It is essential to note that in the definition above the hai ii∈I are indexed families, with repetitions allowed; see 391Xh. (b) I spoke perhaps rather glibly of ‘the largest δ such that . . . ’; you may prefer to write

391N

Kelley’s theorem

δ = inf{sup∅6=J⊆{0,... ,n},inf j∈J aj 6=0

#(J) n+1

571

: a0 , . . . , an ∈ A}.

391I Proposition Let A be a Boolean algebra and A ⊆ A \ {0} any non-empty set. Write C for the set of non-negative finitely additive functionals ν : A → [0, 1] such that ν1 = 1. Then the intersection number of A is precisely maxν∈C inf a∈A νa. proof Write δ for the intersection number of A, and δ 0 for supν∈C inf a∈A νa. (a) For any γ < δ 0 , we can find a ν ∈ C such that νa ≥ γ for every a ∈ A. So if we set ψa = γ for every a ∈ A, ψ satisfies condition (i) of 391F. But this means that if hai ii∈I is any finite family in A, there must be a J ⊆ I such that inf i∈J ai 6= 0 and #(J) ≥ γ#(I). Accordingly γ ≤ δ; as γ is arbitrary, δ 0 ≤ δ. (b) Define ψ : A → [0, 1] by setting ψa = δ for every a ∈ A. If hai ii∈I isP a finite indexed family in A, there is a J ⊆ I such that #(J) ≥ δ#(I) and inf i∈J ai 6= 0; but δ#(I) = i∈I ψai , so this means that condition (ii) of 391F is satisfied. So there is a ν ∈ C such that νa ≥ δ for every a ∈ A; and ν witnesses not only that δ 0 ≥ δ, but that the supremum is a maximum. 391J Theorem Let A be a Boolean algebra. Then the following are equiveridical: (i) there is a strictly positive finitely additive functional on A; (ii) either A = {0} or A\{0} is expressible as a countable union of sets with non-zero intersection numbers. proof (i)⇒(ii) If there is a strictly positive finitely additive functional ν on A, and A 6= {0}, set An = 1 {a : νa ≥ 2−n ν1} for every n ∈ N; then (applying S 391I to the functional ν1 ν) we see that every An has −n intersection number at least 2 , while A \ {0} = n∈N An because ν is strictly positive, so (ii) is satisfied. S (ii)⇒(i) If A \ {0} is expressible as n∈N An , where each An has intersection number δn > 0, then for each nPchoose a finitely additive functional νn on A such that νn 1 = 1, νn a ≥ δn for every a ∈ An . Setting ∞ νa = n=0 2−n νn a for every a ∈ A, ν is a strictly positive additive functional on A, and (i) is true. 391K Corollary Let A be a Boolean algebra. Then A is measurable iff it is Dedekind σ-complete and weakly (σ, ∞)-distributive and either A = {0} or A \ {0} is expressible as a countable union of sets with non-zero intersection numbers. proof Put 391D and 391J together. 391L 391J-391K are due to Kelley 59; condition (ii) of 391J is called Kelley’s criterion. It provides some sort of answer to the question ‘which Boolean algebras carry strictly positive finitely additive functionals?’, but leaves quite open the possibility that there is some more abstract criterion which is also necessary and sufficient. It is indeed a non-trivial exercise to find any ccc Boolean algebra which does not carry a strictly positive finitely additive functional. The first example published seems to have been that of Gaifman 64; I present this in a modified form, following Comfort & Negrepontis 82. First, a definition: Definition Let A be a Boolean algebra. A set A ⊆ A is linked if a ∩ b 6= 0 for all a, b ∈ A. A is σ-linked if A \ {0} is a countable union of linked sets. Remark ‘σ-linkedness’ is one of a very long list of ‘chain conditions’ which have been studied; there is an extensive investigation in Comfort & Negrepontis 82. I hope to look at some of them in Volume 5. For the moment I will introduce them one by one as the occasion arises. 391M Lemma A σ-linked Boolean algebra is ccc. proof If A is σ-linked, take a sequence hAn in∈N of linked sets with union A \ {0}. Then any dist subset of A can meet each An in at most one element, so must be countable. 391N Example There is a σ-linked Boolean algebra A such that there is no strictly positive finitely additive real-valued functional defined on A.

572

Measurable algebras

391N

proof (a) Write I for the set of half-open intervals of the form [q, q 0 [ where q, q 0 ∈ Q and q < q 0 . Then I is countable; enumerate it as hIn in∈N . For each n ∈ N let Jn ⊆ I be a dist family of subintervals of In such that #(Jn ) = (n + 1)2 . Give {0, 1}R its usual product topology, so that it is a compact Hausdorff space (3A3K); for x ∈ {0, 1}R set Qx = {t : t ∈ R, x(t) = 1}. Now set T X = n∈N {x : x ∈ {0, 1}R , #({J : J ∈ Jn , Qx ∩ J 6= ∅}) ≤ n + 1}. Take A to be the algebra of open-and-closed subsets of X. (b) I show first that there is no strictly positive finitely additive functional on A. P P?? If there were, there would be a sequence hAk ik∈N with union A \ {∅} such that every Ak had strictly positive intersection number (391J). For each t ∈ R, write et = {x : x ∈ X, x(t) = 1}, so that et ∈ A. Note that if we set xt (t) = 1, xt (s) = 0 for s 6= t, then Qx = {t} so xt ∈ X, S xt ∈ et and et 6= ∅; so there must be some k such that et ∈ Ak . Set Tk = {t : et ∈ Ak } for each k; then k∈N Tk = R. By Baire’s theorem (3A3G) there is a k ∈ N such that G = int T k 6= ∅. Let δ > 0 be the intersection number of Ak . There must be some n ≥ 1/δ such that In ⊆ G. Since I ⊆ T k , there is for every J ∈ Jn a point tJ ∈ J ∩ Tk , so that etJ ∈ Ak . Consider the family hetJ iJ∈Jn . Because the intersection number of Ak is δ, there must T be a set K ⊆ Jn , 2 of cardinal at least δ#(J ) = δ(n + 1) > n + 1, such that inf e = 6 ∅, that is, X ∩ n J∈K tJ J∈K etJ 6= ∅. But T if x ∈ J∈K etJ then Qx contains tJ for every J ∈ K, so #({J : J ∈ Jn , Qx ∩ J 6= ∅}) ≥ #(K) > n + 1 and x ∈ / X. X XQ Q Thus A does not satisfy Kelley’s criterion and there is no strictly positive finitely additive functional defined on A. (c)(i) Write F for the set of functions f such that dom f is a finite subset of R and f (t) ∈ {0, 1} for every t ∈ dom f ; for f ∈ F write Qf = {t : t ∈ dom f , f (t) = 1}, Hf = {x : x ∈ {0, 1}R , x(t) = f (t) for every t ∈ dom f }, cf = X ∩ Hf ∈ A. (ii) For k ∈ N write Kk for the set of all dist finite families K of non-empty half-open intervals with rational endpoints such that #(K) = k and whenever n ≤ 2k and J ∈ Jn then every member of K is either included in J or dist S from J. (This allows ∅ to belong to K0 .) For K ∈ Kk let FK be the set of those f ∈ F such that Qf = K ∩ dom f and Qf ∩ K has just one member for every K ∈ K. Set AK = {a : a ∈ A, there is some f ∈ FK such that ∅ 6= cf ⊆ a}. I claim that a ∩ b 6= ∅ for any a, b ∈ AK . P P There are f , g ∈ FK such that a ⊇ cf 6= ∅, b ⊇ cg 6= ∅. Now if t ∈ dom f ∩ dom g, S f (t) = 1 ⇐⇒ t ∈ K ⇐⇒ g(t) = 1; that is, f and g agree on dom f ∩dom g, and therefore there is a function h = f ∪g, with domain dom f ∪dom g, extending both. Of course h ∈ F . Define z ∈ {0, 1}R by setting z(t)S= h(t) if t ∈ dom h, 0 otherwise, so that Qz = Qh = Qf ∪ Qg and z extends both f and g. Note that Qz ⊆ K, and #(Qz ) ≤ #(Qf ) + #(Qg ) = 2k. Now we are supposing that cf is non-empty; say x ∈ cf . Then Qx ⊇ Qf , so Qx meets every member of K. If n ≤ 2k, then S {J : J ∈ Jn , Qx ∩ J 6= ∅} ⊇ {J : J ∈ Jn , J ∩ K 6= ∅} because for K ∈ K, J ∈ Jn either K ⊆ J or K ∩ J = ∅. But this means that [ #({J : J ∈ Jn , Qz ∩ J 6= ∅}) = #({J : J ∈ Jn , J ∩ K 6= ∅}) ≤ #({J : J ∈ Jn , Qx ∩ J 6= ∅}) ≤ n + 1 for every n ≤ 2k; while also

391Yb

Kelley’s theorem

573

#({J : J ∈ Jn , Qz ∩ J 6= ∅}) ≤ #(Qz ) ≤ 2k ≤ n + 1 for every n ≥ 2k. So z ∈ X. As z extends both f and g, z ∈ cf ∩ cg ⊆ a ∩ b and a ∩ b 6= ∅. Q Q Thus every AK is linked. (iii) For every non-zero a ∈ A there is some k ∈ N, K ∈ Kk such that a ∈ AK . P P Take any x ∈ a. Because a is a relatively open subset of X, there is an open G ⊆ {0, 1}R such that a = G ∩ X; now, by the definition of the topology of {0, 1}R , there is an f ∈ F such that x ∈ Hf ⊆ G, so that ∅ 6= cf ⊆ a. Set k = #(Qf ). Then we can choose a dist family hKt it∈Qf in I such that, for each t, Kt ∩dom f = {t} S and, for every J ∈ n≤2k Jn , either Kt ⊆ J or Kt ∩ J = ∅. Set K = {Kt : t ∈ Qf }; then K ∈ Kk , f ∈ FK and a ∈SAK . Q Q But k∈N Kk is countable, being a subset of the family of finite sets in the countable set I, so the AK form a countable family of linked sets covering A \ {0}, and A is σ-linked. 391X Basic exercises For this series of exercises, I will call a Boolean algebra A chargeable if there is a strictly positive finitely additive functional ν : A → [0, ∞[. (a) Show that a chargeable Boolean algebra is ccc, so is Dedekind complete iff it is Dedekind σ-complete. (b) Show (i) that any subalgebra of a chargeable Boolean algebra is chargeable (ii) that a countable simple product of chargeable Boolean algebras is chargeable (iii) that any free product of chargeable Boolean algebras is chargeable. (c) (i) Let A be a Boolean algebra with a chargeable order-dense subalgebra. Show that A is chargeable. (ii) Show that the Dedekind completion of a chargeable Boolean algebra is chargeable. (d) (i) Show that the algebra of open-and-closed subsets of {0, 1}I is chargeable for any set I. (ii) Show that the regular open algebra of R is chargeable. (e) (i) Show that any principal ideal of a chargeable Boolean algebra is chargeable. (ii) Let A be a chargeable Boolean algebra and I an order-closed ideal of A. Show that A/I is chargeable. > (f ) Show that a Boolean algebra is chargeable iff it is isomorphic to a subalgebra of a measurable algebra. (Hint: if A is a Boolean algebra and ν : A → [0, ∞[ is a strictly positive additive functional, set ρ(a, b) = ν(a 4 b) for a, b ∈ A. Show that ρ is a metric on A and that the metric completion of A under ρ is a measurable algebra under the natural operations (cf. 324O).) (g) Explain how to use the Hahn-Banach theorem to prove 391G directly, without ing through 391F. (Hint: S(B) can be regarded as a subspace of S(A).) > (h) Take X = {0, 1, 2, 3}, A = PX, A = {{0, 1}, {0, 2}, {0, 3}, {1, 2, 3}}. Show that the intersection number of A is 35 . (Hint: use 391I.) Show that if a0 , . . . , an are distinct of A then there is a set J ⊆ {0, . . . , n}, with #(J) ≥ 23 (n + 1), such that inf j∈J aj 6= 0. (i) Let A be a Boolean algebra. For non-empty A ⊆ A \ {0} write δ(A) for the intersection number of A. Show that for any non-empty A ⊆ A \ {0}, δ(A) = sup{δ(I) : I is a non-empty finite subset of A}. (j) (i) Show that any subalgebra of a σ-linked Boolean algebra is σ-linked. (ii) Show that if A is a Boolean algebra with a σ-linked order-dense subalgebra, then A is σ-linked. (iii) Show that the simple product of a countable family of σ-linked Boolean algebras is σ-linked. (iv) Show that a principal ideal of a σ-linked Boolean algebra is σ-linked. 391Y Further exercises (a) Show that in 391D and 391K we can replace ‘weakly (σ, ∞)-distributive’ by ‘weakly σ-distributive’. (b) Show that PN is σ-linked and chargeable but that the quotient algebra PN/[N]<ω is not ccc, therefore neither σ-linked nor chargeable.

574

Measurable algebras

391Yc

(c) (i) Show that if X is a separable topological space, then its regular open algebra is chargeable. (ii) Let hXi ii∈I be any family of topological spaces with chargeable regular open algebras. Show that their product has a chargeable regular open algebra. (d) Show that any σ-linked Boolean algebra can have cardinal at most c. (Hint: take a sequence hAn in∈N of linked sets covering A \ {0}. Set A∗n = {a : ∃ b ∈ An , b ⊆ a}. Show that if a 6⊆ b there is an n such that a ∈ A∗n and b ∈ / A∗n .) (e) Show that the free product of c or fewer σ-linked Boolean algebras is σ-linked. (Hint: let hAi ii∈I be a family of σ-linked Boolean algebras, where I ⊆ R. For each i ∈ I, let hAin in∈N be a sequence of linked sets with union Ai \ {0}. For q0 < q1 < . . . < qr ∈ Q and n1 , . . . , nr ∈ N let Bq,n be the set of elements expressible as inf i∈J εi (ai ) where J ⊆ I ∩ [q0 , qr [ is a set meeting each interval [qk−1 , qk [ in at most one point and ai ∈ Ai,nk if i ∈ J ∩ [qk−1 , qk [; show that Bq,n is linked.) (f ) Let X be the space of 391N. (i) Show that X is closed in {0, 1}R , therefore compact. (ii) Show that the regular open algebra of X is σ-linked but not chargeable. (g) Let A be the algebra of 391N. Show that A is not weakly (σ, ∞)-distributive. (Hint: show that for any t ∈ R there is a strictly decreasing sequence hti ii∈N , converging to t, such that x 7→ hx(ti )ii∈N is a surjection from X onto {0, 1}N , and hence that the algebra of open-and-closed subsets of {0, 1}N can be regularly embedded into A; now use 316Xq and 316Xm.) (h) If A is a Boolean algebra, a set A ⊆ A is m-linked, where m ≥ 2, if a1 ∩ a2 ∩ . . . ∩ am 6= 0 for all a1 , . . . , am ∈ A; and A is σ-m-linked if A \ {0} can be expressed as the union of a sequence of m-linked sets. (Thus ‘linked’ is ‘2-linked’ and ‘σ-linked’ is ‘σ-2-linked’.) Show that the algebra of 391N is σ-m-linked for every m ≥ 2. (Hint: in part (c) of the proof, replace each ‘2k’ by ‘mk’.) 391Z Problem Must a Dedekind complete σ-linked weakly (σ, ∞)-distributive Boolean algebra be measurable? 391 Notes and comments By the standards of this volume, this is an easy section; I note that I have hardly called on anything after Chapter 32, except for a reference to the construction S(A) in §361. I do ask for a bit of functional analysis (the Hahn-Banach theorem) in 391E. Kelley’s criterion (391J) is a little unsatisfying. It is undoubtedly useful (see part (b) of the proof of 391N, or 392F below), but at the same time the structure of the criterion – a special sequence of subsets of A – is rather close to the structure of the conclusion; after all, one is, or can be represented by, a function from A \ {0} to N, while the other is a function from A to R. Also the actual intersection number of a family A ⊆ A \ {0} can be hard to calculate; as often as not, the best method is to look at the additive functionals on A (see 391Xh). I take the trouble to show that Gaifman’s example (391N) is σ-linked in order to show that even conditions very much stronger than ‘ccc’ are not sufficient to guarantee the existence of suitable functionals. (See also 391Yh.) For other examples see Comfort & Negrepontis 82. But I note that none of the standard examples is weakly (σ, ∞)-distributive (see 391Yg), and I believe it is open, in the formal sense, whether any weakly (σ, ∞)-distributive σ-linked Boolean algebra must carry a strictly positive finitely additive functional (391Z). But I conjecture that the answer is ‘no’. Since every σ-linked algebra has cardinal at most c (391Yd), not every measurable algebra is σ-linked. In fact it is known that a measurable algebra of cardinal c or less is σ-m-linked (391Yh) for every m ≥ 2 ¯ns 93). (Dow & Stepra

392D

Submeasures

575

392 Submeasures In §391 I looked at what we can deduce if a Boolean algebra carries a strictly positive finitely additive functional. There are important contexts in which we find ourselves with a subadditive, rather than additive, functional, and this is what I wish to investigate here. It turns out that, once we have found the right hypotheses, such functionals can also provide a criterion for measurability of an algebra (392J). The argument runs through a new idea, using a result in finite combinatorics (392D). 392A Definition Let A be a Boolean algebra. A submeasure on A is a functional ν : A → [0, ∞[ such that ν(a ∪ b) ≤ νa + νb for all a, b ∈ A; νa ≤ νb whenever a ⊆ b; ν0 = 0. (In this context I do not allow ∞ as a value of a submeasure.) Any positive finitely additive functional is a submeasure (326Ba, 326Bf). 392B Definitions Let A be a Boolean algebra and ν : A → [0, ∞[ a submeasure. Then (a) ν is strictly positive if νa > 0 for every a 6= 0; (b) ν is exhaustive if limn→∞ νan = 0 for every dist sequence han in∈N in A; (c) ν is uniformly exhaustive if for every ² > 0 there is an n ∈ N such that there is no dist family a0 , . . . , an with νai ≥ ² for every i ≤ n. 392C Proposition Let A be a Boolean algebra. (a) If there is an exhaustive strictly positive submeasure on A, then A is ccc. (b) A uniformly exhaustive submeasure on A is exhaustive. (c) Any positive linear functional on A is a uniformly exhaustive submeasure. proof These are all elementary. If ν : A → [0, ∞[ is an exhaustive strictly positive submeasure, and hai ii∈I is a dist family in A \ {0}, then {i : νai ≥ 2−n } must be finite for each n, so I is countable. (Cf. 322G.) If ν : A → [0, ∞[ is a uniformly exhaustive submeasure and han in∈N is dist in A, then {i : νai ≥ 2−n } is finite for each n, so limi→∞ νai = 0. If νP: A → [0, ∞[ is a positive linear functional, and ² > 0, then take n n ≥ 1² ν1; if a0 , . . . , an are dist, then i=0 νai ≤ ν1, so mini≤n νai < ². 392D Lemma Suppose that k, l, m ∈ N are such that 3 ≤ k ≤ l ≤ m and 18mk ≤ l2 . Let L, M be sets of sizes l, m respectively. Then there is a set R ⊆ M × L such that (i) each vertical section of R has just three (ii) #(R[E]) ≥ #(E) whenever E ∈ [M ]≤k ; so that for every E ∈ [M ]≤k there is an injective function f : E → L such that (x, f (x)) ∈ R for every x ∈ E. recall that [M ]≤k = {I : I ⊆ M, #(I) ≤ k} (3A1J). proof (a) We need to know that n! ≥ 3−n nn for every n ∈ N; this is immediate from the inequality Rn Pn i=2 ln i ≥ 1 ln x dx = n ln n − n + 1 for every n ≥ 2. (b) Let Ω be the set of those R ⊆ M × L such that each vertical section of R has just three , so that ¡ l! ¢m #(Ω) = #([L]3 )m = . 3!(l−3)!

(I write [X]j for the set of subsets of X with j .) Let us regard Ω as a probability space with the uniform probability. If F ∈ [L]n , where 3 ≤ n ≤ k, and x ∈ M , then

Pr(R[{x}] ⊆ F ) = (because R[{x}] is a random member of [L]3 )

#([F ]3 ) #([L]3 )

576

Measurable algebras

=

392D

n(n−1)(n−2) l(l−1)(l−2)

n3 . l3

≤

So if E ∈ [M ]n and F ∈ [L]n , then Y

Pr(R[E] ⊆ F ) =

Pr(R[{x}] ⊆ F )

x∈E

(because the sets R[{x}] are chosen independently) ≤

n3n . l3n

Accordingly

Pr(there is an E ⊆ M such that #(R[E]) < #(E) ≤ k) ≤ Pr(there is an E ⊆ M such that #(R[E]) ≤ #(E) ≤ k) = Pr(there is an E ⊆ M such that 3 ≤ #(R[E]) ≤ #(E) ≤ k) (because if E 6= ∅ then #(R[E]) ≥ 3) ≤

k X X

X

Pr(R[E] ⊆ F ) ≤

n=3 E∈[M ]n F ∈[L]n

=

k X n=3

k X

#([M ]n )#([L]n )

n=3

l! n3n m! n!(m−n)! n!(l−n)! l3n

≤

k X mn ln n3n n=3

n!n!l

3n

≤

n3n l3n

k X mn nn 32n

l2n

n=3

(using (a)) =

k X ¡ 9mn ¢n n=3

l2

≤

k X 1 n=3

2n

< 1.

There must therefore be some R ∈ Ω such that #(R[E]) ≥ #(E) whenever E ⊆ M and #(E) ≤ k. (c) If now E ∈ [M ]≤k , the restriction RE = R ∩ (E × L) has the property that #(RE [I]) ≥ #(I) for every I ⊆ E. By Hall’s Marriage Lemma (3A1K) there is an injective function f : E → L such that (x, f (x)) ∈ RE ⊆ R for every x ∈ E. Remark Of course this argument can be widely generalized; see references in Kalton & Roberts 83. 392E Lemma Let A be a Boolean algebra and ν : A → [0, ∞[ a uniformly exhaustive submeasure. Then for any ² ∈ ]0, ν1] the set A = {a : νa ≥ ²} has intersection number greater than 0. proof (a) If ν1 = 0 this is trivial, so we may assume that ν1 > 0; since neither the hypothesis nor the conclusion is affected if we multiply ν by a positive scalar, we may suppose that ν1 = 1. Because ν is uniformly exhaustive, there is a dist family in A then P is an r ≥ 1 such that whenever hci ii∈I 1 2 δ , so that #({i : νci > 15 ²}) ≤ r, so that i∈I νci ≤ r + 51 ²#(I). Set δ = ²/5r, η = 74 δ−η ≥

1 18 (δ

− η)2 ≥

1 2 18 (δ

− 2η) = 4η.

(b) Let hai ii∈I be a non-empty finite family in A. Let m be any multiple of #(I) greater than or equal to 1/η. Then there are integers k, l such that 3η ≤

k m

≤ 4η ≤

1 (δ 18

− η)2 ,

δ−η ≤

l m

≤ δ,

in which case 3 ≤ k ≤ l ≤ m,

18mk ≤ m2 (δ − η)2 ≤ l2 .

392G

Submeasures

577

(c) Take a set M of the form I × S where #(S) = m/#(I), so that #(M ) = m. For x = (i, s) ∈ M set dx = ai . Let L be a set with l . By 392D, there is a set R ⊆ M × L such that every vertical section of R has just three and whenever E ∈ [M ]≤k there is an injective function fE : E → L such that (x, fE (x)) ∈ R for every x ∈ E. For E ⊆ M set bE = inf x∈E dx \ supx∈M \E dx , so that hbE iE⊆M is a partition of unity in A. For x ∈ M , j ∈ L set cxj = sup{bE : x ∈ E ∈ [M ]≤k , fE (x) = j}. If x, y are distinct of M and j ∈ L then cxj ∩ cyj = sup{bE : x, y ∈ E ∈ [M ]≤k , fE (x) = fE (y) = j} = 0, because every fE is injective. Set for each j ∈ L. Note that cxj We have

mj = #({x : x ∈ M, cxj 6= 0}) P = 0 if (x, j) ∈ / R, so j∈L mj ≤ #(R) = 3m. P x∈M

1 5

νcxj ≤ r + ²mj

for each j, by the choice of r; so X x∈M,j∈L

1 5

νcxj ≤ rl + ²

X

3 5

mj ≤ rl + m²

j∈L

3 5

4 5

≤ (rδ + ²)m = ²m < ²m by the choice of l and δ. There must therefore be some x ∈ M such that P ν(supj∈L cxj ) ≤ j∈L νcxj < ² ≤ νdx , and dx cannot be included in supj∈L cxj = sup{bE : x ∈ E ∈ [M ]≤k }. But as sup{bE : x ∈ E ⊆ M } is just dx , there must be an E ⊆ M , of cardinal greater than k, such that bE 6= 0. Recall now that M = I × S, and that k ≥ 3ηm = 3η#(I)#(S). The set J = {i : ∃ s, (i, s) ∈ E} must therefore have more than 3η#(I) , since E ⊆ J × S. But also d(i,s) = ai for each (i, s) ∈ E, so that inf i∈J ai ⊇ bE 6= 0. (d) As hai ii∈I is arbitrary, the intersection number of A is at least 3η > 0. 392F Theorem Let A be a Boolean algebra with a strictly positive uniformly exhaustive submeasure. Then A has a strictly positive finitely additive functional. proof If A = {0} this is trivial. Otherwise, let ν : A → [0, ∞[ be a strictly positive uniformlySexhaustive submeasure. For each n, An = {a : νa ≥ 2−n ν1} has intersection number greater than 0, and n∈N An = A \ {0} because ν is strictly positive; so A has a strictly positive finitely additive functional, by Kelley’s theorem (391J). 392G Since positive additive functionals are uniformly exhaustive submeasures, the condition of this theorem is necessary as well as sufficient. Thus we have a description, in of submeasures, of a condition equivalent to one part of the criterion for measurability of an algebra in 391D. The language of submeasures also provides a formulation of another part of this criterion, as follows.

578

Measurable algebras

392G

Definition Let A be a Boolean algebra. A Maharam submeasure or continuous outer measure on A is a submeasure ν : A → [0, ∞[ such that limn→∞ νan = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. 392H Lemma Let A be a Boolean algebra and ν a Maharam submeasure on A. (a) ν is sequentially order-continuous. (b) ν is ‘countably subadditive’, P∞ that is, whenever han in∈N is a sequence in A and a ∈ A is such that a = supn∈N a ∩ an , then νa ≤ n=0 νan . (c) If A is Dedekind σ-complete, then ν is exhaustive. proof (a) (Of course ν is an order-preserving function, by the definition of ‘submeasure’; so we can apply the ordinary definition of ‘sequentially order-continuous’ in 313Hb.) (i) If han in∈N is a non-decreasing sequence in A with supremum a, then han \ ain∈N is a non-increasing sequence with infimum 0, so limn→∞ ν(an \ a) = 0; but as νan ≤ νa ≤ νan + ν(a \ an ) for every n, it follows that νa = limn→∞ νan . (ii) If han in∈N is a non-increasing sequence in A with infimum a, then νa ≤ νan ≤ νa + ν(an \ a) → νa as n → ∞.

Pn

νai for each n (inducing on n), so that P∞ νa = limn→∞ νbn ≤ i=0 νai .

(b) Set bn = supi≤n a ∩ ai ; then νbn ≤

i=0

(c) If han in∈N is a dist sequence in A, set bn = supi≥n ai for each n; then inf n∈N bn = 0, so lim supn→∞ νan ≤ limn→∞ νbn = 0. 392I Proposition Let A be a Dedekind σ-complete Boolean algebra and ν a strictly positive Maharam submeasure on A. Then A is ccc, Dedekind complete and weakly (σ, ∞)-distributive. proof By 392Hc, ν is exhaustive; by 392Ca, A is ccc; by 316Fa, A is Dedekind complete. Now suppose that we have a sequence hAn in∈N of non-empty downwards-directed subsets of A, all with infimum 0. Let A be the set {a : a ∈ A, ∀ n ∈ N ∃ a0 ∈ An such that a0 ≤ a}. By 316Fc and 392Ha, ν is order-continuous, so inf a∈An νa = 0 for each n. Given ² > 0,Pwe can choose ∞ han in∈N such that an ∈ An and νan ≤ 2−n ² for each n; now a = supn∈N an ∈ A and νa ≤ n=0 νan ≤ 2². Thus inf a∈A νa = 0. Since ν is strictly positive, inf A = 0. As hAn in∈N is arbitrary, A is weakly (σ, ∞)distributive. 392J Theorem Let A be a Boolean algebra. Then it is measurable iff it is Dedekind σ-complete and carries a uniformly exhaustive strictly positive Maharam submeasure. proof If A is measurable, it surely satisfies the conditions, since any totally finite measure on A is also a uniformly exhaustive strictly positive Maharam submeasure. If A satisfies the conditions, then it is weakly (σ, ∞)-distributive, by 392I, and carries a strictly positive finitely additive functional, by 392F; so is measurable, by 391D. 392X Basic exercises (a) Show that the first two clauses of the definition 392A can be replaced by ‘νa ≤ ν(a ∪ b) ≤ νa + νb whenever a ∩ b = 0’. (b) Let A be any Boolean algebra and ν a submeasure on A. (i) Show that the following are equiveridical: (α) ν is order-continuous; (β) whenever A ⊆ A is non-empty, downwards-directed and has infimum 0, then S inf a∈A νa = 0. (ii) Show that in this case ν is exhaustive. (Hint: if han in∈N is dist, then n∈N {b : b ⊇ ai for every i ≥ n} has infimum 0.)

§393 intro.

The Control Measure Problem

579

> (c) Let A be the finite-cofinite algebra on an uncountable set (316Yk). (i) Set ν1 0 = 0, ν1 a = 1 for a ∈ A \ {0}. Show that ν1 is a strictly positive Maharam submeasure but is not exhaustive. (ii) Set ν2 a = 0 for finite a, 1 for cofinite a. Show that ν2 is a uniformly exhaustive Maharam submeasure but is not order-continuous. >(d) Let A be a Boolean algebra and ν a submeasure on A. Set I = {a : νa = 0}. Show that (i) I is an ideal of A (ii) there is a submeasure ν¯ on A/I defined by setting ν¯a• = νa for every a ∈ A (iii) if ν is exhaustive, so is ν¯ (iv) if ν is uniformly exhaustive, so is ν¯ (v) if ν is a Maharam submeasure, I is a σ-ideal (vi) if ν is a Maharam submeasure and A is Dedekind σ-complete, ν¯ is a Maharam submeasure. (e) Let A be a Dedekind complete Boolean algebra and ν an order-continuous submeasure on A. Show that ν has a unique a ∈ A such that ν¹ Aa is strictly positive and ν¹ A1\a is identically zero. (f ) Let A be a Dedekind σ-complete Boolean algebra and ν a uniformly exhaustive Maharam submeasure on A. Show that there is a non-negative countably additive functional µ on A such that {a : µa = 0} = {a : νa = 0}. (Hint: 392Xd(vi).) 392 Notes and comments Much of this section is a matter of generalizing earlier arguments. Thus 392C and 392H ought by now to be very easy, while 392I uses the methods of 322F-322G, and 392Xb recalls the elementary theory of τ -additive functionals. The new ideas are in the combinatorics of 392D-392E. I have cast 392D in the form of an argument in probability theory. Of course there is nothing here but simple counting, since the probability measure simply puts the same mass on each point of Ω, and every statement of the form ‘Pr(R . . . ) ≤ . . . ’ is just a matter of counting the elements R of Ω with the given property. But I think many of us find that the probabilistic language makes the calculations more natural; in particular, we can use intuitions associated with the notion of independence of events. Indeed I strongly recommend the method. It has been used to very great effect in the last fifty years in a wide variety of combinatorial problems. 392F/392J constitute the Kalton-Roberts theorem (Kalton & Roberts 83). It is not known whether every exhaustive submeasure is uniformly exhaustive; this is the Control Measure Problem, which I will treat in the next section.

393 The Control Measure Problem I come now to a discussion of a classic problem of measure theory. Its importance derives to a great extent from the variety of forms in which it appears, and this section is devoted primarily to a description of some of these forms, with the arguments to show that they are indeed all the same problem, in that a solution to one of the questions will provide solutions to all the others. The syntax of the exposition seems to be simplest if I present each formulation as a statement ‘CMn ’; the corresponding question being ‘is CMn true?’, and the proof that all the questions are really the same question becomes a proof that CMm ⇐⇒ CMn for all m, n. The propositions CM∗ are listed in 393A, 393H, 393J, 393L and 393P, while the arguments that they are equiveridical form the rest of the material down to and including 393R. CM1 -CM3B all involve submeasures in one way or another, and much of the section amounts to a theory of submeasures. On the way I mention a description of the open-and-closed algebra of {0, 1}N (393F). The propositions CM4 , CM5 and CM6 concern topologies on Boolean algebras and L0 spaces, and vector measures. They can be expressed more or less adequately with very little of the surrounding theories, but for completeness I include a basic theorem on vector measures (393S). At the end of the section I present two of the many examples of submeasures which have been described.

580

Measurable algebras

393A

393A The problem The language I introduced in the last section is already sufficient for more than one formulation of the problem. Consider the statements (CM1 ) If A is a Dedekind complete Boolean algebra and ν is a strictly positive Maharam submeasure on A, then A is measurable. (CM01 ) Let A be a Dedekind σ-complete Boolean algebra and ν : A → [0, ∞[ a Maharam submeasure. Then there is a non-negative countably additive functional µ : A → [0, ∞[ such that, for a ∈ A, µa = 0 ⇐⇒ νa = 0. (CM001 ) Let X be a set, Σ a σ-algebra of subsets of X, and ν : Σ → [0, ∞[ a Maharam submeasure. Then there is a totally finite measure µ, with domain Σ, such that, for E ∈ Σ, νE = 0 ⇐⇒ µE = 0. (CM2 ) Every exhaustive submeasure is uniformly exhaustive. (CM02 ) Every exhaustive submeasure on the algebra of open-and-closed subsets of {0, 1}N is uniformly exhaustive. The following lemmas assemble the new ideas we need in order to prove that these statements are equiveridical. 393B Lemma Let A be a Boolean algebra and ν a strictly positive submeasure on A. (a) We have a metric ρ on A defined by the formula ρ(a, b) = ν(a 4 b) for all a, b ∈ A. (b) The Boolean operations ∪ , ∩ , 4 , \ , and the function ν : A → R, are all uniformly continuous for ρ. b ρˆ) of (A, ρ) is a Boolean algebra under the natural continuous (c) The metric space completion (A, b which is again a extensions of the Boolean operations, and ν has a unique continuous extension νˆ to A strictly positive submeasure. b is Dedekind complete and ccc and νˆ is a Maharam submeasure. (d) If ν is exhaustive, then A b νˆ) is a totally finite measure algebra. (e) If ν is additive, then (A, proof (a)-(b) This is just a generalization of 323A-323B; essentially the same formulae can be used. For the triangle inequality for ρ, we have a 4 c ⊆ (a 4 b) ∪ (b 4 c), so ρ(a, c) = ν(a 4 c) ≤ ν(a 4 b) + ν(b 4 c) = ρ(a, b) + ρ(b, c). For the uniform continuity of the Boolean operations, we have (b ∗ c) 4 (b0 ∗ c0 ) ⊆ (b 4 b0 ) ∗ (c 4 c0 ) so that ρ(b ∗ c, b0 ∗ c0 ) ≤ ρ(b, b0 ) + ρ(c, c0 ) for each of the operations ∗ = ∪ , ∩ , \ and 4 . For the uniform continuity of the function ν itself, we have νb ≤ νc + ν(b \ c) ≤ νc + ρ(b, c), so that |νb − νc| ≤ ρ(b, c). b × A, b so the Boolean operations on A, regarded as uniformly continuous (c) A × A is a dense subset of A b b (3A4G). If we functions from A × A to A ⊆ A, have unique extensions to continuous binary operations on A look at A = {(a, b, c) : a 4 (b 4 c) = (a 4 b) 4 c}, b×A b × A, b because the maps (a, b, c) 7→ a 4 (b 4 c), (a, b, c) 7→ (a 4 b) 4 c are this is a closed subset of A b continuous and the topology of A is Hausdorff; since A includes the dense set A × A × A, it is the whole of b ×A b × A, b that is, a 4 (b 4 c) = (a 4 b) 4 c for all a, b, c ∈ A. b All the other identities we need to show that A b b A is a Boolean algebra can be confirmed by the same method. Of course A is now a subalgebra of A.

393C


581

b → [0, ∞[. We Because ν : A → [0, ∞[ is uniformly continuous, it has a unique continuous extension νˆ : A have νˆ0 = 0,

νâ ≤ νˆ(a ∪ b) ≤ νâ + νˆb,

νâ = ρˆ(a, 0)

b so νˆ is a submeasure on A, b and for every a, b ∈ A and therefore for every a, b ∈ A, νâ = 0 =⇒ ρˆ(a, 0) = 0 =⇒ a = 0, so νˆ is strictly positive. (d) Now suppose that ν is exhaustive. b is a Cauchy sequence for the metric ρˆ. (i) The point is that any non-increasing sequence han in∈N in A P P Let ² > 0. For each n ∈ N, choose bn ∈ A such that ρˆ(an , bn ) ≤ 2−n ², and set cn = inf i≤n bi . Then Pn ρˆ(an , cn ) = ρˆ(inf i≤n ai , inf i≤n bi ) ≤ i=0 ρˆ(ai , bi ) ≤ 2² for every n. Choose hn(k)ik∈N inductively so that, for each k ∈ N, n(k + 1) ≥ n(k) and ν(cn(k) \ cn(k+1) ) ≥ supi≥n(k) ν(cn(k) \ ci ) − ². Then hcn(k) \ cn(k+1) ik∈N is a dist sequence in A, so lim sup sup ρˆ(an(k) , ai ) ≤ 4² + lim sup sup ρˆ(cn(k) , ci ) k→∞ i≥n(k)

k→∞ i≥n(k)

= 4² + lim sup sup ν(cn(k) \ ci ) k→∞ i≥n(k)

≤ 4² + lim sup ν(cn(k) \ cn(k+1) ) + ² = 5². k→∞

As ² is arbitrary, han in∈N is Cauchy. Q Q b is Dedekind σ-complete. P b hbn in∈N = (ii) It follows that A P If han in∈N is any sequence in A, b For any k ∈ N, hinf i≤n ai in∈N is a Cauchy sequence with a limit b ∈ A. νˆ(b \ ak ) = limn→∞ νˆ(bn \ ak ) = 0, b is a lower bound for {an : n ∈ N}, we have c ⊆ bn so b ⊆ ak , because νˆ is strictly positive. While if c ∈ A for every n, so νˆ(c \ b) = limn→∞ νˆ(c \ bn ) = 0 b is Dedekind σ-complete (314Bc). Q and c ⊆ b. Thus b = inf n∈N an ; as han in∈N is arbitrary, A Q (iii) We also find that νˆ is a Maharam submeasure, because if han in∈N is a non-increasing sequence in b A with infimum 0, it must have a limit a which (as in (ii) just above) must be its infimum, that is, a = 0; consequently limn→∞ νân = νâ = 0. b is ccc (392Ca) and Dedekind complete (iv) It follows at once that νˆ is exhaustive (392Hc), so that A (316Fa). b is Dedekind complete and (e) If ν is additive, then (being finite-valued) it must be exhaustive, so that A νˆ is a Maharam submeasure. But now observe that the function b ×A b→R (a, b) 7→ νˆ(a ∪ b) + νˆ(a ∩ b) − νâ − νˆb : A b × A. b In particular, is continuous (by (b)) and zero on the dense set A × A, so is zero everywhere on A b and νˆ is additive. Since it is also countably subadditive (392Hb) νˆ(a ∪ b) = νâ + νˆb whenever a ∩ b = 0 in A, b it is countably additive, and (A, νˆ) is a totally finite measure algebra. 393C Lemma Let A be a Boolean algebra and ν an exhaustive submeasure on A. Then there are a Dedekind complete Boolean algebra B, a strictly positive Maharam submeasure λ on B, and a Boolean homomorphism π : A → B such that ν = λπ. If ν is additive, then (B, λ) is a totally finite measure algebra.

582

Measurable algebras

393C

proof (a) Set I = {a : νa = 0}. Then I C A; let B0 be the Boolean quotient algebra A/I (312K). If a1 , a2 ∈ A and a•1 ⊆ a•2 in B, then a1 \ a2 ∈ I, so νa1 ≤ νa2 + ν(a1 \ a2 ) = νa2 . So if a•1 = a•2 we must have νa1 = νa2 ; there is therefore a functional λ0 : B0 → [0, ∞[ such that λ0 a• = νa for every a ∈ A. We have just seen that λ0 a•1 ≤ λ0 a•2 if a•1 ⊆ a•2 ; now of course λ0• = ν0 = 0, λ0 (a•1 ∪ a•2 ) = λ0 (a1 ∪ a2 )• = ν(a1 ∪ a2 ) ≤ νa1 + νa2 = λ0 a•1 + λ0 a•2 for all a1 , a2 ∈ A, so λ0 is a submeasure. If b ∈ B \ {0}, then b = a• where a ∈ / I, so λ0 b = νa > 0; thus λ0 is strictly positive. If hbn in∈N is a dist sequence in B0 , express it as ha•n in∈N and set a ñ = an \ supi 0 there is a δ > 0 such that νa ≤ ² whenever ν 0 a ≤ δ. 393E Lemma Let A be a Dedekind σ-complete Boolean algebra and ν, ν 0 two Maharam submeasures on A such that νa = 0 whenever ν 0 a = 0. Then ν is absolutely continuous with respect to ν 0 . proof ?? Otherwise, we can find a sequence han in∈N in A such that ν 0 an P ≤ 2−n for every n, but ² = ∞ 0 inf n∈N νan > 0. Set bn = supi≥n ai for each n, b = inf n∈N bn . Then ν bn ≤ i=n 2−i ≤ 2−n+1 for each n (392Hb), so ν 0 b = 0; but νbn ≥ ² for each n, so νb ≥ ² (392Ha), contrary to the hypothesis. X X 393F Lemma Let B be the algebra of open-and-closed subsets of {0, 1}N . Then a Boolean algebra A is isomorphic to B iff it is atomless, countable and not {0}. proof (a) I must check that B has the declared properties. The point is that it is the subalgebra B0 of PX generated by {bi : i ∈ N}, where I write X = {0, 1}N , bi = {x : x ∈ X, x(i) = 1}. P P Of course bi and its complement {x : x(i) = 0} are open, so bi ∈ B for each i, and B0 ⊆ B. In the other direction, the open cylinder sets of X are all of the form cz = {x : x(i) = z(i) for every i ∈ J}, where J ⊆ I and z ∈ {0, 1}J ; now T S cz = X ∩ z(i)=1 bi \ z(i)=0 bi ∈ B0 . If b ∈ B then b is expressible as a union of such cylinder sets, because it is open; but also it is compact, so is the union of finitely many of them, and must belong to B0 . Thus B = B0 , as claimed. Q Q Because B = B0 is generated by a countable set, it is countable (331Gc). Next, if b ∈ B is non-empty, there are a finite J ⊆ I and a z ∈ {0, 1}J such that cz ⊆ b; now if we take any i ∈ N \ J, and look at the two extensions z0 , z1 of z to J ∪ {i}, then cz0 , cz1 are dist non-empty of B included in b. So B is atomless. (b) Now suppose that A is another algebra with the same properties. Enumerate A as han in∈N . For each n ∈ N let Bn be the finite subalgebra of B generated by {bi : i < n} (so that B0 = {0, 1}). Then hBn in∈N

393G


583

is an increasing sequence of subalgebras of B with union B; also b ∩ bn , b \ bn are non-zero for every n ∈ N, b ∈ Bn . Choose finite subalgebras An ⊆ A and isomorphisms πn : An → Bn as follows. A0 = {0, 1}, π0 0 = 0, π0 1 = 1. Given An and πn , let An be the set of atoms of An . For a ∈ An , choose a0 ∈ A such that if an ∩ a, an \ a are both non-zero, then a0 = an ∩ a; otherwise, a0 ⊆ a is any element such that a0 , a \ a0 are both non-zero. (This is where I use the hypothesis that A is atomless.) Set a0n = supa∈An a0 . Then we see that a ∩ a0n , a \ a0n are non-zero for every a ∈ An and therefore for every non-zero a ∈ An , that is, that sup{a : a ∈ An , a ⊆ a0n } = 0,

inf{a : a ∈ An , a ⊇ a0n } = 1.

Let An+1 be the subalgebra of A generated by An ∪ {a0n }. Since we have sup{b : b ∈ Bn , b ⊆ bn } = 0,

inf{b : b ∈ Bn , b ⊇ bn } = 1,

there is a (unique) extension of πn : An → Bn to a homomorphism πn+1 : An+1 → Bn+1 such that πn+1 a0n = bn (312N). Since we similarly have an extension φ of πn−1 to a homomorphism from Bn+1 to An+1 with φbn = a0n , and since φπn+1 , πn+1 φ must be the respective identity homomorphisms, πn+1 is an isomorphism, and the induction continues. Since πn+1 extends πn for each n, these isomorphisms together to give us an isomorphism S S π : n∈N An → n∈N Bn = B. 0 Observe next that the construction ensures that an ∈ An+1 S for each n, since an ∩ a is either 0 or a or an ∩ a for every a ∈ An , and in all cases belongs to An+1 . So n∈N An contains every an and (by the choice of han in∈N ) must be the whole of A. Thus π : A → B witnesses that A ∼ = B.

393G Theorem If one of CM1 , CM01 , CM001 , CM2 , CM02 is true, so are the others. proof CM1 ⇒ CM01 Assume that CM1 is true, and that A and ν are as in the statement of CM01 . Set I = {a : νa = 0}; because ν is countably subadditive, I is a σ-ideal, B = A/I is Dedekind σ-complete and the quotient map a 7→ a• is sequentially order-continuous (313Qb, 314C). Now we have a functional λ : B → [0, ∞[ defined by saying that λa• = νa for every a ∈ A, and λ is a strictly positive submeasure (as in part (a) of the proof of 393C). In fact λ is a Maharam submeasure. P P Suppose that hbn in∈N is a non-increasing sequence in B with infimum 0. Choose an ∈ A such that a•n = bn for each n, and set cn = inf i≤n ai \ inf i∈N ai ; then hcn in∈N is non-increasing and has infimum 0, while c•n = bn for each n, so limn→∞ λbn = limn→∞ νcn = 0. As hbn in∈N is arbitrary, λ is a Maharam submeasure. Q Q Now CM1 tells us that B must be a measurable algebra, that is, carries a totally finite measure µ ¯. Setting µa = µ ¯a• for a ∈ A, we see that µ is a non-negative countably additive functional, and that µa = 0 ⇐⇒ a• = 0 ⇐⇒ λa = 0, as required. CM01 ⇒ CM001 is trivial; allowing for the change in notation, CM001 is just a special case of CM01 . CM001 ⇒ CM1 Assume CM001 , and let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure ν. Then we can express A as Σ/I for some σ-algebra Σ of subsets of a set X and a σ-ideal I of Σ (314M). Set λE = νE • for E ∈ Σ; then X, Σ, λ satisfy the conditions of CM001 , so there is a totally finite measure µ on X, with domain Σ, such that µE = 0 ⇐⇒ λE = 0 ⇐⇒ νE • = 0 ⇐⇒ E • = 0. Consequently we can identify A with the measure algebra of (X, Σ, µ), and A is measurable. CM1 ⇒ CM2 Now suppose that CM1 is true, and that ν is an exhaustive submeasure on a Boolean algebra A. By 393B, we can find a Dedekind complete Boolean algebra B with a strictly positive Maharam

584

Measurable algebras

393G

submeasure λ and a Boolean homomorphism π : A → B such that νa = λπa for every a ∈ A. CM1 assures us that B is measurable; let µ ¯ be a totally finite measure on B. Let ² > 0. By 393E, λ is absolutely continuous with respect to µ ¯, so there is a δ > 0 such that λb ≤ ² whenever µ ¯a ≤ δ. Take n ≥ 1δ µ ¯1. If a0 , . . . , an are dist in A, then πa0 , . . . , πan are dist in B, so there is some i such that µ ¯πai ≤ δ, in which case νai = λπai ≤ ². As ² is arbitrary, ν is uniformly exhaustive; as A and ν are arbitrary, CM2 is true. CM2 ⇒ CM02 is trivial. CM02 ⇒ CM1 Suppose that CM02 is true, that is, every exhaustive submeasure on the algebra B of openand-closed subsets of {0, 1}N is uniformly exhaustive, and that ν is a strictly positive Maharam submeasure on a Dedekind complete Boolean algebra A. Let E be the set of atoms of A, and set a = 1 \ sup E, so that the principal ideal Aa is atomless, and ν¹ Aa is still a strictly positive Maharam submeasure. ?? Suppose, if possible, that ν¹ Aa is not uniformly exhaustive. Then there is a family hani ii≤n∈N in Aa such that hani ii≤n is dist for each n but inf i≤n∈N νani > 0. There is a countable atomless subalgebra D of Aa containing every ani . P P For each d ∈ Aa fix on a d0 ⊆ d such that d0 and d \ d0 are both non-zero. Define hDn in∈N by setting D0 = {0, a} ∪ {ani : i ≤ n ∈ N}, Dn+1 = Dn ∪ {d1 ∩ d2 : d1 , d2 ∈ Dn } ∪ {a \ d : d ∈ Dn } ∪ {d0 : d ∈ Dn }. S Then every Dn is countable, so D = n∈N Dn is countable. Because D includes D0 , it contains all the ani ; Q also d1 ∩ d2 , 1 \ d1 and d01 belong to D for every d1 , d2 ∈ D, so D is an atomless subalgebra. Q Because ν is an exhaustive submeasure on A, ν¹ D is an exhaustive submeasure on D. Because every ani belongs to D and inf i≤n∈N νani > 0, ν¹ D is not uniformly exhaustive. Let B be the algebra of open-and-closed subsets of {0, 1}N . By 393F, there is a Boolean isomorphism π : B → D. Set λb = νπb for every b ∈ B; then (B, λ) is isomorphic to (D, ν¹ D), and, in particular, λ is an exhaustive submeasure on B which is not uniformly exhaustive, which we are supposing to be impossible. X X Thus Aa has a uniformly exhaustive Maharam submeasure; since it is surely Dedekind complete, it is measurable (392J). Let µ ¯1 be a totally finite measure on Aa . Next, because limn→∞ νen = 0 for any sequence hen in∈N of distinct elements of E, and νe > 0 for every e ∈ E, E must be countable, and there is a summable P family hαe ie∈E of strictly positive real numbers. Setting µ ¯c = µ ¯1 (c ∩ a) + e∈E,e ⊆ c αe , we get a totally finite measure µ ¯ on A, and A is measurable. 393H Variations on CM2 The following modifications of CM2 are interesting because they display some general properties of exhaustive submeasures. Consider the statements (CM3A ) If A is a Boolean algebra and ν is a non-zero exhaustive submeasure on A, then there is a non-zero finitely additive functional µ on A such that 0 ≤ µa ≤ νa for every a ∈ A. (CM03A ) If A is a Boolean algebra and ν is a non-zero exhaustive submeasure on A, then there is a non-zero non-negative finitely additive functional µ on A such that µ is absolutely continuous with respect to ν. (CM3B ) If A is a Boolean algebra and ν is an exhaustive submeasure on A, then there is a nonnegative finitely additive functional µ on A such that ν is absolutely continuous with respect to µ. 393I Proposition If one of CM1 , . . . , CM02 , CM3A , CM03A , CM3B is true, so are the others. proof CM1 ⇒ CM3A & CM3B Assume CM1 , and let ν be a non-zero exhaustive submeasure on a Boolean algebra A. By 393C, there are a Dedekind complete Boolean algebra B, a strictly positive Maharam submeasure λ on B, and a Boolean homomorphism π : A → B such that ν = λπ. Since λ1 = ν1 6= 0, B 6= {0}. CM1 tells us that B is measurable; let µ ¯ be a strictly positive countably additive functional on B, and set µ1 = µ ¯π, so that µ1 is a finitely additive functional on A. By 393E, µ ¯ is absolutely continuous with respect to λ and λ is absolutely continuous with respect to µ ¯; it follows from the latter that ν is absolutely continuous with respect to µ1 .

393K


585

Set δ = 21 λ1/¯ µ1, and consider D = {b : b ∈ B, λb ≤ δ µ ¯d}. ?? If D is order-dense in B, there is a dist C ⊆ D such that sup C = 1. Because B is measurable, it is ccc, and C is countable. If C is infinite, enumerate it as hcn in∈N ; if it is finite, enumerate it as hci ii≤n and set ci = 0 for i > n. In either case, hcn in∈N is a dist sequence in C with supremum 1, so that P∞ P∞ λ1 ≤ n=0 λcn ≤ δ n=0 µ ¯cn = δ µ ¯1 < λ1 by the choice of δ, which is absurd. X X There is therefore a non-zero b ∈ B such that λc > δ µ ¯c for every non-zero c ⊆ b. Set µ2 a = δ µ ¯(b ∩ πa) ≤ λ(b ∩ πa) ≤ λπa = νa for a ∈ A; then µ2 is a non-negative finitely additive functional on A, µ2 ≤ ν, and µ2 1 = δ µ ¯b 6= 0, so µ2 is non-zero. As A and ν are arbitrary, CM3A and CM3B are both true. (The statement of CM3B does not assume ν to be non-zero, but of course the case ν = 0 is trivial.) CM3A ⇒ CM03A is trivial. CM03A ⇒ CM1 Assume CM03A , and let ν be a strictly positive Maharam submeasure on a Dedekind complete Boolean algebra A. Note that A is ccc (392Ca). Now C = {a : a ∈ A, Aa is measurable} is order-dense in A. P P Take any a ∈ A \ {0}. Set νa b = ν(a ∩ b) for b ∈ A. It is easy to check that νa is a nonzero Maharam submeasure on A; in particular, it is exhaustive. So there is a non-zero non-negative finitely additive functional µ which is absolutely continuous with respect to νa . If hbn in∈N is a non-increasing sequence in A with infimum 0, then limn→∞ νa bn = 0, so limn→∞ µbn = 0; accordingly µ is countably additive (326Ga), therefore completely additive, because A is ccc (326L). Set I = {b : µb = 0}; then I is an order-closed ideal, so contains its supremum b0 say. Since µb = 0 whenever νa b = 0, b0 ⊇ 1 \ a, and c ⊆ a, where c = 1 \ b0 . But Ac ∩ I = {0}, so µ¹ Ac is strictly positive, and witnesses that Ac is measurable. Also µc = µ1 6= 0, so c is a non-zero member of C included in a. As a is arbitrary, C is order-dense. Q Q By 391Cb, there is a function µ ¯1 such that (A, µ ¯1 ) is a localizable measure algebra. Because A is ccc, A is measurable (322G). CM3B ⇒ CM1 Assume CM3B , and let A be a Dedekind complete Boolean algebra and ν a strictly positive Maharam submeasure on A. By CM3B , there is a non-negative finitely additive functional µ on A such that ν is absolutely continuous with respect to µ. But this means that µa = 0 =⇒ νa = 0 =⇒ a = 0, that is, µ is strictly positive. Also A is Dedekind complete and weakly (σ, ∞)-distributive, by 392I. So A is measurable, by 391D. As A and ν are arbitrary, CM1 is true. 393J The first published version of the Control Measure Problem (Maharam 47), in the form ‘is CM1 true?’, was a re-formulation of a question about topologies on Boolean algebras, which I now describe. Consider the statement (CM4 ) Let A be a ccc Dedekind complete Boolean algebra with a Hausdorff topology T such that (i) the Boolean operation ∪ : A × A → A is continuous at (0, 0) (ii) if han in∈N is a non-increasing sequence in A with infimum 0, then han in∈N converges to 0 for T. Then A is measurable. 393K Proposition CM4 is true iff CM1 , . . . , CM3B are true. proof CM1 ⇒ CM4 Assume CM1 , and take A, T as in the statement of CM4 . (i) For any e ∈ A \ {0}, there is a Maharam submeasure ν on A such that νe > 0. α) Choose a sequence hGn in∈N of neighbourhoods of 0, as follows. Because T is Hausdorff, there is a P P(α neighbourhood G0 of 0 not containing e. Given Gn , choose a neighbourhood Gn+1 of 0 such that Gn+1 ⊆ Gn and a ∪ b ∪ c ∈ Gn whenever a, b, c ∈ Gn+1 . (Take neighbourhoods H, H 0 of 0 such that a ∪ b ∈ Gn for a, b ∈ H, b ∪ c ∈ H for b, c ∈ H 0 and set Gn+1 = H ∩ H 0 ∩ Gn .) Define ν0 : A → [0, 1] by setting

586

Measurable algebras

393K

ν0 a = 1 if a ∈ / G0 , = 2−n if a ∈ Gn \ Gn+1 , \ = 0 if a ∈ Gn . n∈N

Pr

Then whenever a0 , . . . , ar ∈ A, n ∈ N and i=0 ν0 ai < 2−n , supi≤r ai ∈ Gn . To see this, induce on r. If r = 0 thenP we have ν0 a0 < 2−n P so a0 ∈ Gn+1 ⊆ Gn . For the inductive step to r ≥ 1, there must be a k ≤ r such that i
Pr

ν1 a = inf{

i=0

ν0 ai : a0 , . . . , ar ∈ A, a = supi≤r ai }

for every a ∈ A. It is easy to see that ν1 (a ∪ b) ≤ ν1 a + ν1 b for all a, b ∈ A; also a ∈ Gn whenever ν1 a < 2−n , so, in particular, ν1 e ≥ 1, because e ∈ / G0 . Set νa = inf{ν1 b : a ∩ e ⊆ b ⊆ e} for every a ∈ A. Then of course 0 ≤ νa ≤ νb whenever a ⊆ b, and ν0 ≤ ν1 0 ≤ ν0 0 = 0, so ν0 = 0. If a, b ∈ A and ² > 0, there are a0 , b0 such that a ∩ e ⊆ a0 ⊆ e, b ∩ e ⊆ b0 ⊆ e, ν1 a0 ≤ νa + ² and ν1 b0 ≤ νb + ²; so that (a ∪ b) ∩ e ⊆ a0 ∪ b0 ⊆ e and ν(a ∪ b) ≤ ν1 (a0 ∪ b0 ) ≤ ν1 a0 + ν1 b0 ≤ νa + νb + 2². As ², a and b are arbitrary, ν is a submeasure. Next, if hai ii∈N is any non-increasing sequence in A with infimum 0, hai ∩ eii∈N is another, so converges to 0 for T. If n ∈ N there is an m such that ai ∩ e ∈ Gn for every i ≥ m, so that νai ≤ ν1 (ai ∩ e) ≤ ν0 (ai ∩ e) ≤ 2−n for every i ≥ m. As n is arbitrary, limi→∞ νai = 0; as hai ii∈N is arbitrary, ν is a Maharam submeasure. Finally, νe = ν1 e ≥ 1, so νe 6= 0. Q Q (ii) Write C for the set of those c ∈ A such that Ac is a measurable algebra. Then C is order-dense in A. P P Take any e ∈ A \ {0}. By (i), there is a Maharam submeasure ν such that νe > 0. Set I = {a : νa = 0}, a∗ = sup I. Because A is ccc, there is a sequence han in∈N in I such that a∗ = supn∈N an ; because ν is countably subadditive, νa∗ = 0. Consider c = e \ a∗ . Then ν¹ Ac is a strictly positive Maharam submeasure, so CM1 assures us that Ac is measurable; also c 6= 0 because νc = νe 6= 0, so 0 6= c ⊆ e and c ∈ C. As e is arbitrary, C is order-dense. Q Q Now A is a ccc, Dedekind complete Boolean algebra in which the measurable principal ideals are orderdense, so A is measurable, just as in the proof of CM03A ⇒ CM1 (393I) CM4 ⇒ CM1 Now assume that CM4 is true, and that A is a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure ν. Then the associated metric ρ on A (393B) defines a topology T satisfying the conditions of CM4 . P P The continuity of ∪ is mentioned in 393B. If han in∈N is a non-increasing sequence with infimum 0, then ρ(an , 0) = νan → 0 so an → 0 for T. Q Q By CM4 , A is measurable.

393N


587

393L My own first encounter with the Control Measure Problem was in the course of investigating topological Riesz spaces. For various questions depending on this problem, see Fremlin 75. I give the following as a sample. (CM5 ) Let A be a ccc Dedekind complete Boolean algebra, and suppose that there is a Hausdorff linear space topology T on L0 (A) such that for every neighbourhood G of 0 there is a neighbourhood H of 0 such that u ∈ G whenever v ∈ H and |u| ≤ |v|. Then A is measurable. 393M Proposition Let A be a Dedekind σ-complete Boolean algebra and ν : A → [0, ∞[ a strictly positive Maharam submeasure. For u ∈ L0 = L0 (A) set τ (u) = inf{α : α ≥ 0, ν[[|u| > α]] ≤ α}. Then τ defines a metrizable linear space topology T on L0 such that for every neighbourhood G of 0 there is a neighbourhood H of 0 such that u ∈ G whenever v ∈ H and |u| ≤ |v|. proof (a) The point is that τ (u + v) ≤ τ (u) + τ (v),

τ (αu) ≤ τ (u) if |α| ≤ 1,

limα→0 τ (αu) = 0

for every u, v ∈ L0 . P P (i) It will save a moment if we observe that whenever β > τ (u) there is an α ≤ β such that ν[[|u| > α]] ≤ α, so that ν[[|u| > β]] ≤ ν[[|u| > α]] ≤ α ≤ β. Also, because ν is sequentially order-continuous, ν[[|u| > τ (u)]] = limn→∞ ν[[|u| > τ (u) + 2−n ]] ≤ limn→∞ τ (u) + 2−n = τ (u). (ii) So

ν[[|u + v| > τ (u) + τ (v)]] ≤ ν[[|u| + |v| > τ (u) + τ (v)]]) ≤ ν([[|u| > τ (u)]] ∪ [[|v| > τ (v)]]) (364Fa) ≤ ν[[|u| > τ (u)]] + ν[[|v| > τ (v)]] ≤ τ (u) + τ (v), and τ (u + v) ≤ τ (u) + τ (v). (iii) If |α| ≤ 1 then ν[[|αu| > τ (u)]] ≤ ν[[|u| > τ (u)]] ≤ τ (u), and τ (αu) ≤ τ (u). (iv) limn→∞ ν[[|u| > n]] = 0 because h[[|u| > n]]in∈N is a non-increasing sequence with infimum 0. So if ² > 0, there is an n ≥ 1 such that ν[[|u| > n²]] ≤ ², in which case ν[[|αu| > ²]] ≤ ² whenever Q |α| ≤ n1 , so that τ (αu) ≤ ² whenever |α| ≤ n1 . As ² is arbitrary, limα→0 τ (αu) = 0. Q (b) Accordingly we have a metric (u, v) 7→ τ (u−v) which defines a linear space topology T on L0 (2A5B). Now let G be an open set containing 0. Then there is an ² > 0 such that H = {u : τ (u) < ²} is included in G. If v ∈ H and |u| ≤ |v|, then ν[[|u| > τ (v)]] ≤ ν[[|v| > τ (v)]] ≤ τ (v), so τ (u) ≤ τ (v) and u ∈ H ⊆ G. So T satisfies all the conditions. 393N Proposition CM5 is true iff CM1 , . . . , CM4 are true. proof CM4 ⇒ CM5 Assume CM4 . Let A, T be as in the statement of CM5 . Let S be the topology on A induced by T and the function χ : A → L0 ; that is, S = {χ−1 [G] : G ∈ T}. Then S satisfies the conditions of CM4 . P P (i) Because T is Hausdorff and χ is injective, S is Hausdorff. (ii) If 0 ∈ G ∈ S, there is an H ∈ T such that G = χ−1 [H]. Now 0 (the zero of L0 ) belongs to H, so there is an open set H1 containing 0 such that u ∈ H whenever v ∈ H1 and |u| ≤ |v|. Next, addition on L0 is continuous for T, so there is an open set H2 containing 0 such that u + v ∈ H1 whenever u, v ∈ H2 . Consider G0 = χ−1 [H2 ]. This is an open set in A containing 0, and if a, b ∈ G0 then |χ(a ∪ b)| ≤ χa + χb ∈ H2 + H2 ⊆ H1 ,

588

Measurable algebras

393N

so χ(a ∪ b) ∈ H and a ∪ b ∈ G. As G is arbitrary, ∪ is continuous at (0, 0). (iii) If han in∈N is a non-increasing sequence in A with infimum 0, u0 = supn∈N nχan is defined in L0 (use the criterion of 364Ma: inf m∈N supn∈N [[nχan > m]] = inf m∈N am+1 = 0.) If 0 ∈ G ∈ S, take H ∈ T such that G = χ−1 [H], and H1 ∈ T such that 0 ∈ H1 and u ∈ H whenever v ∈ H1 , |u| ≤ |v|. Because scalar multiplication is continuous for T, there is a k ≥ 1 such that k1 u0 ∈ H1 . For any n ≥ k, χan ≤ k1 u0 so χan ∈ H and an ∈ G. As G is arbitrary, han in∈N → 0 for S. As han in∈N is arbitrary, condition (ii) in the statement of CM4 is satisfied. Q Q Since CM4 is true, A must be a measurable algebra. CM5 ⇒ CM1 Assume that CM5 is true, and let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure ν. By 393M, L0 = L0 (A) has a topology satisfying the conditions of CM5 , so A is measurable. 393O The phrase ‘control measure’ derives, in fact, from none of the formulations above; it belongs to the theory of vector measures, as follows. Definitions (a) Let A be a Dedekind σ-complete Boolean algebra and U a Hausdorff linear topological space. (The idea is intended to apply, P∞ in particular, when PnA is a σ-algebra of subsets of a set.) A function θ : A → U is a vector measure if n=0 θan = limn→∞ i=0 θai is defined in U and equal to θ(supn∈N an ) for every dist sequence han in∈N in A. (b) In this case, a non-negative countably additive functional µ : A → [0, ∞[ is a control measure for θ if θa = 0 whenever µa = 0. 393P Now I can formulate the last of this string of statements equiveridical to CM1 . Consider the statement (CM6 ) If A is a Dedekind σ-complete Boolean algebra, U is a metrizable linear topological space, and θ : A → U is a vector measure, then θ has a control measure. 393Q Lemma If A is a Dedekind σ-complete Boolean algebra, U a Hausdorff linear topological space, and θ : A → U a vector measure, then limn→∞ θan = 0 whenever han in∈N is a non-increasing sequence in A with infimum 0. P∞ proof θan = i=n θ(ai \ ai+1 ) for every n. 393R Proposition CM6 is true iff CM1 , . . . , CM5 are true. proof CM01 ⇒ CM6 (i) Assume CM01 , and let A, U and θ be as in the statement of CM6 . Then there is a functional τ : U → [0, ∞[ such that, for u, v ∈ U , τ (u + v) ≤ τ (u) + τ (v), τ (αu) ≤ τ (u) whenever |α| ≤ 1, limα→0 τ (αu) = 0, τ (u) = 0 iff u = 0, and the topology of U is defined by the metric (u, v) 7→ τ (u − v) (2A5Cb). For a ∈ A set νa = supb ⊆ a min(1, τ (θb)). (ii) Consider the functional ν : A → [0, ∞[. (α) ν0 = τ (θ0) = τ (0) = 0. (To see that θ0 = 0 set an = 0 for every n in the definition 393Oa.) (β) If a ⊆ b then of course νa ≤ νb. (γ) If a, b ∈ A and c ⊆ a ∪ b, then min(1, τ (θc)) = min(1, τ (θ(c ∩ a) + θ(c \ a))) ≤ min(1, τ (θ(c ∩ a)) + τ (θ(c \ a))) ≤ νa + νb. As c is arbitrary, ν(a ∪ b) ≤ νa + νb. Thus ν is a submeasure. (δ) ?? Suppose, if possible, that there is a non-decreasing sequence han in∈N in A, with infimum 0, such that hν(an )in∈N does not converge to 0. Because hν(an )in∈N is non-increasing, ² = 31 inf n∈N νan is greater than 0. Now for each n ∈ N there are m > n, b such that b ⊆ an \ am and τ (θb) ≥ ². P P As νan ≥ 3² there is a c ⊆ an such that τ (θc) ≥ 2². Now hθ(am ∩ cim∈N converges to 0 in U , by 393Q, so there is an m > n such that

*393S


589

τ (θ(c \ am )) = τ (θc − θ(am ∩ c)) ≥ τ (θc) − τ (θ(am ∩ c)) ≥ ², and we can take b = c \ am . Q Q We may therefore choose hbk ik∈N , hn(k)ik∈N such that n(k +1) > n(k), bk ⊆ an(k) \ an(k+1) and τ (θbk ) ≥ ² P∞ for every k. But hbk ik∈N is dist, so we ought to be able to form k=0 θbk = θ(supk∈N bk ) in U , and limk→∞ θbk = 0 in U , that is, limk→∞ τ (θbk ) = 0. X X Thus limn→∞ νan = 0 for every non-increasing sequence han in∈N with infimum 0, and ν is a Maharam submeasure. (iii) By CM01 , there is a non-negative countably additive functional µ : A → [0, ∞[ such that νa = 0 whenever µa = 0. In particular, if µa = 0, then τ (θa) = 0 and θa = 0. So µ is a control measure for θ. As A, U and θ are arbitrary, CM6 is true. CM6 ⇒ CM1 Assume CM6 , and let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure ν. Give L0 = L0 (A) the topology of 393M. Then χ : A → L0 is a vector measure in the sense of 393O. P P If han in∈N is a dist sequence in A with supremum a, set bn = supi≤n ai , so that Pn χbn = i=0 χai for each n. We have ν(a \ bn ) → 0, so that τ (χa − χbn )) = min(1, ν(a \ bn )) → 0, P∞ where τ is the functional of 393M, and χa = i=0 χai in L0 . Q Q CM6 now assures us that there is a non-negative countably additive functional µ on A such that µa = 0 =⇒ χa = 0 =⇒ a = 0, so that µ is strictly positive and A is measurable. As A and ν are arbitrary, CM1 is true. *393S I must not go any farther without remarking that the generality of the phrase ‘metrizable linear topological space’ in CM6 is essential. If we look only at normed spaces we do not need to know anything about the Control Measure Problem, as the following theorem shows. Theorem Let A be a Dedekind σ-complete Boolean algebra, U a normed space and θ : A → U a vector measure. Then θ has a control measure. ˆ (3A5Ib), and as θ will still be a vector proof (a) Since U can certainly be embedded in a Banach space U ˆ measure when regarded as a map from A to U , we may assume from the beginning that U itself is complete. (b) θ is bounded (that is, supa∈A kθak is finite). P P?? Suppose, if possible, otherwise. Choose han in∈N inductively, as follows. a0 = 1. Given that supa ⊆ an kθak = ∞, choose b ⊆ an such that kθbk ≥ kθan k + 1. Then kθ(an \ b)k ≥ 1. Also supa ⊆ an kθak ≤ supa ⊆ an kθ(a ∩ b)k + kθ(a \ b)k, so at least one of supa ⊆ b kθak, supa ⊆ an \b kθak must be infinite. We may therefore take an+1 to be either b or an \ b and such that supa ⊆ an+1 kθak = ∞. Observe that in either case we shall have kθ(an \ an+1 )k ≥ 1. Continue. At the end of the P induction we shall have a dist sequence han \ an+1 in∈N such that kθ(an \ an+1 )k ≥ 1 ∞ for every n, so that n=0 θ(an \ an+1 ) cannot be defined in U ; which is impossible. X XQ Q (c) Accordingly we have a bounded linear operator T : L∞ → U , where L∞ = L∞ (A), such that T χ = θ (363Ea). Now the key to the proof is the following fact: if hun in∈N is a dist order-bounded sequence in (L∞ )+ , hT un in∈N → 0 in U . P P Let γ be such that un ≤ γχ1 for every n. Let ² > 0, and let k be the integer part of γ/². For n ∈ N, i ≤ k set ani = [[un > ²(i + 1)]]; then hani in∈N is dist for each i, and if we set Pk vn = ² i=0 χani , we get vn ≤ uP n ≤ vn + ²χ1, so kun − vn k∞ ≤ ². ∞ Because hani in∈N is dist, n=0 θani is defined in U , and hθani in∈N → 0, for each i ≤ k. Consequently Pk T vn = ² i=0 θani → 0 as n → ∞. But kT un − T vn k ≤ kT kkun − vn k∞ ≤ ²kT k

590

Measurable algebras

*393S

for each n, so lim supn→∞ kT un k ≤ ²kT k. As ² is arbitrary, limn→∞ kT un k = 0. Q Q (d) Consider the adt operator T 0 : U ∗ → (L∞ )∗ . Recall that L∞ is an M -space (363B) so that its dual is an L-space (356N). Write A = {T 0 g : g ∈ U ∗ , kgk ≤ 1} ⊆ (L∞ )∗ = (L∞ )∼ . If u ∈ L∞ , then supf ∈A |f (u)| = supkgk≤1 |(T ∗ g)(u)| = supkgk≤1 |g(T u)| = kT uk. Now A is uniformly integrable. P P I use the criterion of 356O. Of course kf k ≤ kT 0 k for every f ∈ A, so A is norm-bounded. If hun in∈N is an order-bounded dist sequence in (L∞ )+ , then supf ∈A |f (un )| = kT un k → 0 as n → ∞. So A is uniformly integrable. Q Q (e) Next, A ⊆ (L∞ )∼ P If f ∈ A, it is of the form T 0 g for some g ∈ U ∗ , that is, c . P f (χa) = (T 0 g)(χa) = gT (χa) = g(θa) for every a ∈ A. If now han in∈N is a dist sequence in A with supremum a, P∞ P∞ P∞ f (χa) = g(θ(supn∈N an )) = g( n=0 θan ) = n=0 g(θan ) = n=0 f (χan ). So f χ is countably additive. By 363K, f ∈ (L∞ )∼ Q c . Q (f ) Because A is uniformly integrable, there is for each m ∈ N an fm ≥ 0 in (L∞ )∗ such that k(|f | − fm )+ k ≤ 2−m for every f ∈ A; moreover, we can suppose that fm is of the form supi≤km |fmi | where every fmi belongs to A (354R(b-iii)), so that fm ∈ (L∞ )∼ c and µm = fm χ is countably additive. Set P∞ 1 γm = 1 + µm 1 for each m, µ = m=0 m µm ; 2 γm

then µ : A → [0, ∞[ is a non-negative countably additive functional. Now µ is a control measure for θ. P P If µa = 0, then µm a = 0, that is, fm (χa) = 0, for every m ∈ N. But this means that if g ∈ U ∗ and kgk ≤ 1, |g(θa)| = |(T 0 g)(χa)| ≤ fm (χa) + k(|T 0 g| − fm )+ k ≤ 2−m for every m, by the choice of fm ; so that g(θa) = 0. As g is arbitrary, θa = 0; as a is arbitrary, µ is a control measure for θ. Q Q 393T This concludes the list of ‘positive’ results I wish to present in this section. I now devote a few pages to significant examples of submeasures. The first example is a classic formulation (taken from Talagrand 80) which shows in clear relief some of the fundamental ways in which submeasures differ from measures. Examples (a) Fix n ≥ 1, and let I be the set {0, 1, . . . , 2n − 1}, X = [I]n , so that X is a finite set. For each i ∈ I set Ai = {a : i ∈ a ∈ X}. For E ⊆ X set νE = =

[

1 n+1

inf{#(J) : J ⊆ I, E ⊆

1 n+1

inf{#(J) : a ∩ J 6= ∅ for every a ∈ E}.

Ai }

i∈J

It is elementary to check that ν : PX → [0, ∞[ is a strictly positive submeasure, therefore (because PX is finite) a Maharam submeasure. The essential properties of ν are twofold: (i) νX = 1; (ii) for any non-negative additive functional µ such 2 . P P (i)(α) If J ⊆ I and #(J) ≤ n, there is an a ∈ [I \ J]n , that µE ≤ νE for every E ⊆ X, µX ≤ n+1 S S so that a ∈ X \ i∈J Ai and X 6⊆ i∈J Ai . This means that X cannot be covered by fewer than n + 1 of the sets Ai , so that νX must be at leastS1. (β) On the other hand, if J ⊆ I is any set of cardinal n + 1, a ∩ J 6= ∅ for every a ∈ X, so that X = i∈J Ai and νX ≤ 1. (ii) Every member of X belongs to just n of the sets Ai , so

393U


nµX = and µX ≤

2 n+1 .

P i∈I

µAi ≤

#(I) n+1

=

591

2n , n+1

Q Q

(b) This example shows at least that any proof of CM3A cannot work through any generally valid inequality of the form ‘if ν is a Maharam submeasure there is an additive functional µ with δν ≤ µ ≤ ν’. If we take a sequence of these spaces we can form a result which in one direction is stronger, as follows. For each n ≥ 1 set In =Q{0, . . . , 2n − 1} and define Xn = [In ]n , Ani = {a : i ∈ a ∈ Xn }, νn : PXn → [0, 1] ∞ as in (a) above. Set Z = n=1 Xn , K = {(n, i) : n ≥ 1, i ∈ In } and Cni = {x : x ∈ Z, x(n) ∈ Ani } for (n, i) ∈ K. Now define θ : PZ → [0, 1] by setting S P 1 : J ⊆ K, W ⊆ (n,i)∈J Cni } θW = inf{ (n,i)∈J n+1

for every W ⊆ Z. Then θ is an outer measure on Z, by arguments we have been familiar with since 114D. Also θZ = 1. P P (i) Because (for instance) X1 P is covered by the two sets A10 and A11 , Z is covered by C10 1 and C11 , so that θZ ≤ 1. (ii) If J ⊆ K and (n,i)∈J n+1 < 1, set Jn = {i : (n, i) ∈ J} for each n; then S #(Jn ) < n + 1, so we can choose an x(n) ∈ Xn \ i∈Jn Ani . This defines a sequence x ∈ Z such that x ∈ / Cni S for every (n, i) ∈ J, and Z 6= (n,i)∈J Cni . As J is arbitrary, θZ ≥ 1. Q Q Finally, if Σ is any subalgebra of PZ containing every Cni , and µ : Σ → [0, ∞[ a non-negative finitely additive functional such that µE ≤ θE for every E ∈ Σ, then µ = 0. P P For each n, every point of x belongs to n different Cni , just as in (a) above; so that P 2n nµZ ≤ i∈In µCni ≤ . n+1

As this is true for every n ≥ 1, µZ = 0. Q Q (c) A non-zero submeasure ν on a Boolean algebra A is called pathological if the only additive functional µ such that 0 ≤ µa ≤ νa for every a ∈ A is the zero functional. Thus the submeasure θ of (b) above is pathological, and CM3A can be read ‘an exhaustive submeasure cannot be pathological’. (d) It is important to note that Fubini’s theorem fails catastrophically for submeasures. As a simple example, consider the space (X, ν) of (a) above, for some fixed n ≥ 1. If we define θ : P(X × X) → [0, 1] by setting P∞ S θW = inf{ i=0 νEi · νFi : W ⊆ i∈N Ei × Fi } for each W ⊆ X × X, then we obtain an outer measure, just as if ν itself were a measure; but θ(X × X) ≤ 4n/(n + 1)2 is small compared with (νX)2 . P P Give {0, . . . , 2n − 1} = Z2n its usual group operation +2n of addition mod 2n. Then S S X × X ⊆ i<2n (Ai × Ai ) ∪ i<2n (Ai × Ai+2n 1 ), because if a, b ∈ X either a ∩ b 6= ∅ and there is some i such that a, b both belong to Ai , or b = {0, . . . , 2n − 1} \ a and there is some i such that i ∈ a, i +2n 1 ∈ b. So P 4n θ(X × X) ≤ i<2n νAi (νAi + νAi+2n 1 ) = . Q Q 2 (n+1)

393U

For the next example, I present a much deeper idea from Roberts 93.

Example Let B be the algebra of open-and-closed subsets of {0, 1}N . Then for any ² > 0 we can find a submeasure ν : B → [0, 1] such that (i) for every n ∈ N there is a dist sequence Sn0 , . . . , Snn in B such that νSni = 1 for every i ≤ n; (ii) if hEn in∈N is any dist sequence in B then lim supn→∞ νEn ≤ ². Q proof (a) For each n ∈ N let In be the finite set {0, . . . , n}, given its discrete topology; set X = n∈N In , with the product topology; let C be the algebra of subsets of X generated by sets of the form Sij = {x :

592

Measurable algebras

393U

x(i) = j}, where i ∈ N and j ≤ i. Note that X is compact and Hausdorff and that every member of C is open-and-closed (because all the Sij are). Also C is atomless, countable and non-zero, so is isomorphic to B, by 393F. It will therefore be enough if I can describe a submeasure ν : C → [0, 1] with the properties (i) and (ii) above, and this is what I will do. (b) For each n ∈ N let An be the set of non-empty of C determined by coordinates in {0, . . . , n}; note that An is finite. For k ≤ l ∈ N, say that E ∈ C is (k, l)-thin if for every A ∈ Ak there is an A0 ∈ Al such that A0 ⊆ A \ E. Note that if k 0 ≤ k ≤ l ≤ l0 then Ak0 ⊆ Ak and Al ⊆ Al0 , so any (k, l)-thin set is also (k 0 , l0 )-thin. Say that every E ∈ C is (k, 0)-small for every k ∈ N, and that for k, r ∈ N a set E ∈ C is (k, r+1)-small if there is some l ≥ k such that E is (k, l)-thin and (l, r)-small. Observe that E is (k, 1)-small iff it is (k, l)-thin for some l ≥ k, that is, there is no member of Ak included in E. Observe also that if E is (k, r)-small then it is (k 0 , r)-small for every k 0 ≤ k. Write S = {Sij : j ≤ i ∈ N}. (c) Suppose that E ∈ C and k ≤ l ≤ m are such that E is both (k, l)-thin and (l, m)-thin. Then whenever A ∈ Ak , S ∈ S and A ∩ S 6= ∅, there is an A0 ∈ Am such that A0 ⊆ A \ E and A0 ∩ S 6= ∅. P P Take S = Sni where i ≤ n. (i) If n ≤ l, then A ∩ S ∈ Al ; because E is (l, m)-thin, there is an A0 ∈ Am such that A0 ⊆ (A ∩ S) \ E. (ii) If n > l, there is an A0 ∈ Al such that A0 ⊆ A \ E, because E is (k, l)-thin; now A0 ∈ Am , and A0 ∩ S is non-empty because A0 is determined by coordinates less than n. Q Q (d) It follows that if S ∈ S, k ∈ N, A ∈ Ak , A ∩ S 6= ∅, r ∈ N and E0 , . . . , Er−1 are (k, 2r)-small, then A ∩ S is not covered by E0 , . . . , Er−1 . P P Induce on r. The case r = 0 demands only that A ∩ S should not be covered by the empty sequence, that is, A ∩ S 6= ∅, which is one of the hypotheses. For the inductive step to r + 1, we know that for each j ≤ r there are lj , mj such that k ≤ lj ≤ mj and Ej is (k, lj )-thin and (lj , mj )-thin and (mj , 2r)-small. Rearranging E0 , . . . , Er if necessary we may suppose that mr ≤ mj for every j ≤ r; set m = mr . By (c), there is an A0 ∈ Am such that A0 ∩ S 6= ∅ and A0 ⊆ A \ Er . Now S every Ej , for j < r, is (mj , 2r)-small, therefore (m, 2r)-small, so by the inductive hypothesis A0 ∩ S 6⊆ j
1 r

min{m : E ⊆ E1 ∪ . . . ∪ Em for some E1 , . . . , Em ∈ U}

if E can be covered by r or fewer of U, 1 otherwise. It is easy to check that ν : C → [0, 1] is a submeasure. By (d), no member of S can be covered by r or fewer of U, so νSni = 1 whenever i ≤ n ∈ N. By (e), if hEn in∈N is any dist sequence in C, En belongs to U for all but finitely many n, so that νEn ≤ 1r ≤ ² for all but finitely many n. Thus ν has the required properties. 393X Basic exercises (a) Let A be a Boolean algebra and ν a strictly positive Maharam submeasure on A. Show that a subset F of A is closed for the topology defined by ν iff the limit of any order*-convergent sequence in F belongs to F . (b) Let A be a Dedekind σ-complete Boolean algebra with a strictly positive Maharam submeasure ν. Show that A is complete under the metric defined from ν (393B).

393Yd


593

(c) Let A be a Dedekind complete Boolean algebra and ν, ν 0 two strictly positive Maharam submeasures on A. Show that the corresponding metrics on A (393B) are uniformly equivalent. (d) Let A be a countable Boolean algebra, not {0}. Show that A is isomorphic to an order-closed subalgebra of the algebra B of open-and-closed subsets of {0, 1}N . (Hint: show that A ⊗ B ∼ = B.) (e) Suppose that CM1 -CM6 are false. Show that there is a Dedekind complete Boolean algebra A, with a strictly positive Maharam submeasure, such that (i) the only countably additive real-valued functional on A is the zero functional (ii) τ (A) = ω. (f ) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure. Show that if τ (A) ≤ ω then the associated topology on A (393B) is separable and A is σ-linked. (g) Consider the statement (CM06 ) If X is a set, Σ a σ-algebra of subsets of X, U a metrizable linear topological space, and θ : Σ → U a vector measure, then θ has a control measure. Show that CM06 ⇐⇒ CM6 . (h) Consider the outer measures νn , θ of 393Ta-b. Give every Xn its discrete topology and Z the corresponding product topology. Let E be the algebra of open-and-closed subsets of Z. (i) Show that E is the subalgebra of PZ generated P by sets of the form {x : x ∈ Z, x(n) = a} for n ≥ 1, a ∈ Xn . (ii) Take n n ≥ 1 and set A = {a : a ∈ Xn , i∈a i is even}. Show that νn A = νn (X \ A) = n+1 . (iii) Show that for any α < 1 there is a dist sequence hEn in∈N in E such that θEn ≥ α for every n. 393Y Further exercises (a) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure. Show that if τ (A) ≤ ω then A is σ-m-linked (391Yh) for every m. (b) Let A be a Boolean algebra with just four elements, and let N + be the set of submeasures on A, N = N + − N − the set of functionals from A to R expressible as the difference of two submeasures. (i) Show that N is just the three-dimensional space of functionals ν : A → R such that ν0 = 0. (ii) Show that there is a partial order on N under which N is a partially ordered linear space with positive cone N + , but that N is now not a lattice. (c) Let A be a Dedekind complete Boolean algebra. Show that there is at most one Hausdorff topology T on A such that (i) the Boolean operations ∩ , 4 are continuous (ii) for every open set G containing 0 there is an open set H containing 0 such that a ∈ G whenever a ⊆ b ∈ H (iii) 0 ∈ A whenever A ⊆ A is non-empty and downwards-directed and has infimum 0. (Hint: Given such a topology T, show that the construction in 393K produces enough order-continuous submeasures on A to define T. Now if ν is any order-continuous submeasure, let a be its ; because Aa is ccc, there is a T-continuous submeasure with a; use 393E to see that ν is continuous.) (d) Let A be a Boolean algebra. Let T be the topology on A such that the closed sets for T are just those sets F ⊆ A such that a ∈ F whenever han in∈N is a sequence in F which order*-converges to a (367Yc). Show that a 7→ a ∪ c, a 7→ a ∩ c, a 7→ a 4 c, a 7→ a \ c, a 7→ c \ a are all T-continuous, for any c ∈ A. Write U for the family of open sets G containing 0 such that [0, a] ⊆ G for every a ∈ G; for A ⊆ A set A0 = {a ∪ b : a, b ∈ A}. Show that G ⊆ G0 ∈ U for every G ∈ U. Suppose now that A is ccc and Dedekind complete and that T is Hausdorff. Show that (i) A is weakly (σ, ∞)-distributive; (ii) whenever A ⊆ A and a ∈ A there is a sequence in A which order*-converges to a (see 367Yl); (iii) whenever 0 ∈ G ∈ T there is an H ∈ U such that H ⊆ G (hint: H = int{a : [0, a] ⊆ G} ∈ U); (iv) for any a ∈ A \ {0} there is a G ∈ U such that a ∈ / G0 ; (v) for every G ∈ U there isTan H ∈ U such that 00 0 H ⊆ G ; (vi) for any a ∈ A \ {0} there is a sequence hGn in∈N in U such that a ⊆ 6 sup( n∈N Gn ); (vii) there T is a sequence hGn in∈N in U such that n∈N Gn = {0}; (viii) for every G ∈ U there is an H ∈ U such that H 0 ⊆ G; (ix) there is a strictly positive Maharam submeasure on A (hint: 393K). ´ ski & Jech 98.) (See Balcar Glowczyn

594

Measurable algebras

393Ye

(e) Consider the statement (CM04 ) Let A be a ccc Dedekind complete Boolean algebra with a Hausdorff topology T such that if han in∈N is a sequence in A which order*-converges to a ∈ A, then han in∈N converges to 0 for the topology T. Then A is measurable. Show that CM04 is true iff CM1 is true. (f ) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure, and let S, T be the associated metrizable topologies on A, L0 (A) (393B, 393N). Write d(A), d(L0 ) for the topological densities of these spaces (331Yf). Show that max(τ (A), ω) = max(d(A), ω) = max(d(L0 ), ω). (g) Let A be a Dedekind σ-complete Boolean algebra, λ : A → [0, ∞[ a non-negative finitely additive functional, and ν a Maharam submeasure on A such that νa = 0 whenever λa = 0. Show that ν is absolutely continuous with respect to λ and therefore also with respect to the countably additive part of λ as defined in 362Bc. (h) Let A be a Dedekind σ-complete Boolean algebra, U a Hausdorff linear topological space and θ : A → U a vector measure. Suppose that there is a bounded finitely additive functional λ : A → R such that θa = 0 whenever λa = 0. Show that θ has a control measure. (i) A linear topological space U is locally convex if the topology has a base consisting of convex sets. Show that if A is a Dedekind σ-complete Boolean algebra, U is a metrizable locally convex linear topological space, and θ : A → U is a vector measure, then θ has a control measure. (h) Consider the outer S measure θ of 393Tb. (i) Show that there is a non-decreasing sequence hWn in∈N of subsets of Z such that θ( n∈N S Wn ) > supn∈N θWn . (ii) Show that if hWn in∈N is a non-decreasing sequence of open subsets of Z then θ( n∈N Wn ) = supn∈N θWn . 393 Notes and comments This section is long in pages, but I hope has been reasonably easy reading, because so many of the arguments amount to re-written versions of ideas already presented. Thus 393Bab can be thought of as an elementary extension of 323A-323B, and the completion procedure in 393Bc uses some of the same ideas as the proof of 325C. The notion of ‘absolute continuity’ for submeasures is taken directly from that for measures, and 393E is a simple translation of 232Ba. Other fragments of the argument for the Radon-Nikod´ ym theorem appear in the proof that CM1 implies CM3A (393I) and in 393S. The argument of 393K takes a fair bit of space, but is mostly a recapitulation of the proof that a uniformity can be defined from pseudometrics (Engelking 89, 8.1.10; Bourbaki 66, IX.1.4). If there is a new idea there, it is in the formula νa = inf{ν1 b : a ∩ e ⊆ b ⊆ e} in the middle of the proof. (Contrast this with the formula νa = supb ⊆ a min(1, τ (θb)) in the proof of 393R.) 393M is just a matter of picking the right description of the topology of convergence in measure to generalize. I do think it is a little surprising that CM3A and CM3B should be equiveridical; one is asking for an additive functional dominated by ν, and the other for an additive functional weakly dominating ν. Of course it is not exactly the same ν in both cases. It is tempting to look for a theory of spaces of submeasures corresponding to the theory of spaces of additive functionals (§326), but it seems that, at the very least, there are new obstacles (393Yb). The arguments of this section tend to be based on hypotheses of the form ‘CMn is true’; I hope that this will not lead to any presumption that the Control Measure Problem has a positive answer. I am myself inclined to believe that all the statements CMn are false, and that one day all this material will have to be re-written with the arguments in reverse, proving ¬CMn ⇒ ¬CMm instead of CMm ⇒ CMn , or using reductio ad absurdum at every point where it is not used above.

394A

Kawada’s theorem

595

I include CM02 because it seems to make the problem more accessible, and indeed it is significant for theoretical reasons. In this form it is possible to show that the Control Measure Problem cannot depend on any of the special axioms that have been introduced in the last thirty-five years; that is, if there is a solution using such an axiom as the continuum hypothesis, or Martin’s axiom, or the axiom of constructibility, then there must be a solution using only the ordinary axioms of Zermelo-Fraenkel set theory (including the axiom of choice). For elucidation of this remark, which I fear may be somewhat cryptic, look for ‘Shoenfield’s Absoluteness Theorem’ in a book on mathematical logic and forcing (e.g., Kunen 80). An incidental benefit from including CM02 is that it has forced me, at last, to spell out the characterization of the algebra of open-and-closed subsets of {0, 1}N (393F). I mention CM001 and CM06 (393Xg) as well as CM01 and CM6 because some natural versions of the Control Measure Problem arise in the context of σ-algebras of sets, rather than Dedekind σ-complete Boolean algebras, and a reminder that this makes no difference to the problem may be helpful. This is not a book about vector measures, but having introduced vector measures in CM6 it would be disgraceful to leave you unaware that there is no control measure problem for measures taking values in normed spaces, and since the proof is no more than an assembly of ideas already covered in their scattered locations, it seems right to set it out (393S). I have said already that my own prejudice is in favour of believing that all the statements CM∗ are false; that is, that there is a Dedekind complete Boolean algebra A, with a strictly positive Maharam submeasure, such that A is not measurable. Ordinarily such a presumption makes a problem seem less interesting; only rarely is there as much honour and profit in constructing a recondite counter-example as there is in proving a substantial theorem. But in this case I believe that it could be the starting point of a new theory of ‘submeasurable algebras’, being Dedekind complete Boolean algebras carrying strictly positive Maharam submeasures. These would be ccc weakly (σ, ∞)-distributive algebras (392I), some of them σ-m-linked for every m (393Ya), carrying metrics for which the Boolean operations were uniformly continuous and order-closed sets were closed (393B); there would be associated theories of topologies of ‘convergence in submeasure’ on algebras and L0 spaces (393B, 393M). We already have a remarkable characterization of such algebras in of order*-convergence (393Yd). But the real prize would be a new algebra for use in forcing, conceivably leading to new models of set theory. The examples in 393T and 393U are there for different purposes. In 393Ta we have submeasures on finite algebras, which are therefore necessarily uniformly exhaustive, in which dominated measures are cruelly dominated; the submeasures are ‘almost pathological’, and can readily be assembled into a submeasure which is pathological in the strict sense (393Tb), but is now very far from being exhaustive (393Xh). Of course the method of 393Tb is extraordinarily crude, but as far as I know nothing else works either (see 393Td). I called 393T a ‘classic’ construction; I am sure that whatever resolution is at last found for the Control Measure Problem, 393Ta at least will always be of interest. It is less clear that 393U will endure in the same way, but for the moment it is the best example known of a submeasure which is almost exhaustive while being far from uniformly exhaustive.

394 Kawada’s theorem I now describe a completely different characterization of (homogeneous) measurable algebras, based on the special nature of their automorphism groups. The argument depends on the notion of ‘non-paradoxical’ group of automorphisms; this is an idea of great importance in other contexts, and I therefore aim at a fairly thorough development, with proofs which are adaptable to other circumstances. 394A Definitions Let A be a Dedekind complete Boolean algebra, and G a subgroup of Aut A. For a, b ∈ A I will say that an isomorphism φ : Aa → Ab between the corresponding principal ideals belongs to the full local semigroup generated by G if there are a partition of unity hai ii∈I in Aa and a family hπi ii∈I in G such that φc = πi c whenever i ∈ I and c ⊆ ai . If such an isomorphism exists I will say that a and b are G-τ -equidecomposable. I will write a 4τG b to mean that there is a b0 ⊆ b such that a and b0 are G-τ -equidecomposable.

596

Measurable algebras

394A

For any function f with domain A, I will say that f is G-invariant if f (πa) = f (a) whenever a ∈ A and π ∈ G. 394B The notion of ‘full local semigroup’ is of course an extension of the idea of ‘full subgroup’ (381M, 387A). The word ‘semigroup’ is justified by (c) of the following lemma, and the word ‘full’ by (e). Lemma Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Write G∗τ for the full local semigroup generated by G. (a) Suppose that a, b ∈ A and that φ : Aa → Ab is an isomorphism. Then the following are equiveridical: (i) φ ∈ G∗τ ; (ii) for every non-zero c0 ⊆ a there are a non-zero c1 ⊆ c0 and a π ∈ G such that φc = πc for every c ⊆ c1 ; (iii) for every non-zero c0 ⊆ a there are a non-zero c1 ⊆ c0 and a ψ ∈ G∗τ such that φc = ψc for every c ⊆ c1 . (b) If a, b ∈ A and φ : Aa → Ab belongs to G∗τ , then φ−1 : Ab → Aa also belongs to G∗τ . (c) Suppose that a, b, a0 , b0 ∈ A and that φ : Aa → Aa0 , ψ : Ab → Ab0 belong to G∗τ . Then ψφ ∈ G∗τ ; its domain is Ac where c = φ−1 (b ∩ a0 ), and its set of values is Ac0 where c0 = ψ(b ∩ a0 ). (d) If a, b ∈ A and φ : Aa → Ab belongs to G∗τ , then φ¹ Ac ∈ G∗τ for any c ⊆ a. (e) Suppose that a, b ∈ A and that ψ : Aa → Ab is an isomorphism such that there are a partition of unity hai ii∈I in Aa and a family hφi ii∈I in G∗τ such that ψc = φi c whenever i ∈ I and c ⊆ ai . Then ψ ∈ G∗τ . proof (a) (Compare 387B.) (i)⇒(iii) is trivial, since of course G ⊆ G∗τ . (iii)⇒(ii) Suppose that φ satisfies (iii), and that 0 6= c0 ⊆ a. Then we can find a ψ ∈ G∗τ and a non-zero c1 ⊆ c0 such that φ agrees with ψ on Ac1 . Suppose that dom ψ = Ad , where necessarily d ⊇ c1 . Then there are a partition of unity hdi ii∈I in Ad and a family hπi ii∈I such that ψc = πi c whenever c ⊆ di . There is some i ∈ I such that c2 = c1 ∩ di 6= 0, and we see that φc = ψc = πi c for every c ⊆ c2 . As c0 is arbitrary, φ satisfies (ii). (ii)⇒(i) If φ satisfies (ii), set D = {d : d ⊆ a, there is some π ∈ G such that πc = φc for every c ⊆ d}. The hypothesis is that D is order-dense in A, so there is a partition of unity hai ii∈I of Aa lying within D (313K); for each i ∈ I take πi ∈ G such that φc = πi c for c ⊆ ai ; then hai ii∈I and hπi ii∈I witness that φ ∈ G∗τ . (b) This is elementary; if hai ii∈I , hπi ii∈I witness that φ ∈ G∗τ , then hφai ii∈I = hπi ai ii∈I , hπi−1 ii∈I witness that φ−1 ∈ G∗τ . (c) I ought to start by computing the domain of ψφ: d ∈ dom(ψφ) ⇐⇒ d ∈ dom φ, φd ∈ dom ψ ⇐⇒ d ⊆ a, φd ⊆ b ⇐⇒ d ⊆ φ−1 (a0 ∩ b) = c. So the domain of ψφ is indeed Ac ; now φ¹ Ac is an isomorphism between Ac and Aφc , where φc = a0 ∩ b ∈ Ab , so ψφ is an isomorphism between Ac and Aψφc = Ac0 . Let hai ii∈I , hbj ij∈J be partitions of unity in Aa , Ab respectively, and hπi ii∈I , hθj ij∈J families in G such that φd = πi d for d ⊆ ai , ψe = θj e for e ⊆ bj . Set cij = ai ∩ πi−1 bj ; then hcij ii∈I,j∈J is a partition of unity in Ac and ψφd = θj πi d for d ⊆ cij , so ψφ ∈ G∗τ (because all the θj πi belong to G). (d) This is nearly trivial; use the definition of G∗τ or the criteria of (a), or apply (c) with the identity map on Ac as one of the factors. (e) This follows at once from the criterion (a-iii) above, or otherwise. 394C Lemma Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Write G∗τ for the full local semigroup generated by G.

394D

Kawada’s theorem

597

(a) For a, b ∈ A, a 4τG b iff there is a φ ∈ G∗τ such that a ∈ dom φ and φa ⊆ b. (b)(i) 4τG is transitive and reflexive; (ii) if a 4τG b and b 4τG a then a and b are G-τ -equidecomposable. (c) G-τ -equidecomposability is an equivalence relation on A. (d) If hai ii∈I and hbi ii∈I are families in A, of which hbi ii∈I is dist, and ai 4τG bi for every i ∈ I, then supi∈I ai 4τG supi∈I bi . proof (a) This is immediate from the definition of ‘G-τ -equidecomposable’ and 394Bd. (b)(i) a 4τG a because the identity homomorphism belongs to G∗τ . If a 4τG b 4τG c there are φ, ψ ∈ G∗τ such that φa ⊆ b, ψb ⊆ c so that ψφa ⊆ c; as ψφ ∈ G∗τ (394Bc), a 4τG c. (ii) (This is of course a Schröder-Bernstein theorem, and the proof is the usual one.) Take φ, ψ ∈ G∗τ such that φa ⊆ b, ψb ⊆ a. Set a0 = a, b0 = b, an+1 = ψbn and bn+1 = φan for each n. Then han in∈N , hbn in∈N are non-increasing sequences; set a∞ = inf n∈N an , b∞ = inf n∈N bn . For each n, φ¹ Aa2n \a2n+1 : Aa2n \a2n+1 → Ab2n+1 \b2n+2 , ψ¹ Ab2n \b2n+1 : Ab2n \b2n+1 → Aa2n+1 \a2n+2 are isomorphisms, while φ¹ Aa∞ : Aa∞ → Ab∞ is another. So we can define an isomorphism θ : Aa → Ab by setting θc = φc if c ⊆ a∞ ∪ sup a2n \ a2n+1 , n∈N

= ψ −1 c if c ⊆ sup a2n+1 \ a2n+2 . n∈N

By 394Be, θ ∈ G∗τ , so a and b are G-τ -equidecomposable. (c) This is easy to prove directly from the results in 394B, but also follows at once from (b-i); any transitive reflexive relation gives rise to an equivalence relation. (d) We may suppose that I is well-ordered by a relation ≤. For i ∈ I, set a0i = ai \ supj
598

Measurable algebras

394D

(iv)⇒(i) Take φ ∈ G∗τ such that φb = a. Set ψd = φ(d ∩ b) ∪ (d \ b) for d ∈ A; then ψ ∈ G∗τ witnesses that 1 is G-τ -equidecomposable with a ∪ (1 \ b) 6= 1. 394E Definition Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. I will say that G is fully non-paradoxical if the statements of 394D are false; that is, if one of the following equiveridical statements is true: (i) if a is G-τ -equidecomposable with 1 then a = 1; (ii) there is no dist sequence han in∈N of non-zero elements of A which are all G-τ -equidecomposable; (iii) there are no non-zero G-τ -equidecomposable a, b, c ∈ A such that a ∩ b = 0 and a ∪ b ⊆ c; (iv) if a ⊆ b ∈ A and a, b are G-τ -equidecomposable then a = b. Note that if G is fully non-paradoxical, and H is a subgroup of Aut A such that H ⊇ G, then H is also fully non-paradoxical, because if a 4τG b then a 4τH b, so that a and b are H-τ -equidecomposable whenever they are G-τ -equidecomposable. 394F Proposition Let (A, µ ¯) be a totally finite measure algebra, and G = Autµ¯ A the group of measure-preserving automorphisms of A. Then G is fully non-paradoxical. proof If φ : A → Aa belongs to the full local subgroup generated by G, then we have a partition of unity hai ii∈I and a family hπi ii∈I in G such that φai = πi ai for every i; but this means that P P P ¯ai = µ ¯1. ¯πi ai = i∈I µ ¯φi ai = i∈I µ µ ¯a = i∈I µ As µ ¯1 < ∞, we can conclude that a = 1, so that G satisfies the condition (i) of 394E. 394G The fixed-point subalgebra of a group Let A be a Boolean algebra and G a subgroup of Aut A. (a) By the fixed-point subalgebra of G I mean C = {c : c ∈ A, πc = c for every π ∈ G}. (I looked briefly at this construction in 333R, and again in §387 in the special case of a group generated by a single element.) This is a subalgebra of A, and is order-closed, because every π ∈ G is order-continuous. (b) Now suppose that A is Dedekind complete. In this case C is Dedekind complete (314Ea), and we have, for any a ∈ A, an element upr(a, C) of C, defined by setting upr(a, C) = inf{c : a ⊆ c ∈ C} (314V). Now upr(a, C) = sup{πa : π ∈ G}. P P Set c1 = upr(a, C), c2 = sup{πa : π ∈ G}. (i) Because a ⊆ c1 ∈ C, πa ⊆ πc1 = c1 for every π ∈ G, and c2 ⊆ c1 . (ii) For any φ ∈ G, φc2 = supπ∈G φπa = supπ∈G πa = c2 because G = {φπ : π ∈ G}. So c2 ∈ C; since also a ⊆ c2 , c1 ⊆ c2 , and c1 = c2 , as claimed. Q Q (c) Again supposing that A is Dedekind complete, write G∗τ for the full local semigroup generated by G. Then φ(a ∩ c) = φa ∩ c whenever φ ∈ G∗τ , a ∈ dom φ and c ∈ C. P P We have φa = supi∈I πi ai , where a = supi∈I ai and πi ∈ G for every i. Now φ(a ∩ c) = supi∈I πi (ai ∩ c) = supi∈I πi ai ∩ c = φa ∩ c. Q Q Consequently upr(φa, C) = upr(a, C) whenever φ ∈ G∗τ and a ∈ dom φ. P P For c ∈ C, a ⊆ c ⇐⇒ a ∩ c = a ⇐⇒ φ(a ∩ c) = φa ⇐⇒ φa ∩ c = φa ⇐⇒ φa ⊆ c. Q Q It follows that upr(a, C) ⊆ upr(b, C) whenever a 4τG b.

394I

Kawada’s theorem

599

(d) Still supposing that A is Dedekind complete, we also find that if a 4τG b and c ∈ C then a ∩ c 4τG b ∩ c. P P There is a φ ∈ G∗τ such that φa ⊆ b; now φ(a ∩ c) = φa ∩ c ⊆ b ∩ c. Q Q Hence, or otherwise, a ∩ c and b ∩ c are G-τ -equidecomposable whenever a and b are G-τ -equidecomposable and c ∈ C. (e) Of course the case C = {0, 1} is particularly significant; when this happens I will call G ergodic. Thus an automorphism π is ‘ergodic’ in the sense of 372P iff the group {π n : n ∈ Z} it generates is ergodic. 394H

I now embark on a series of lemmas leading to the main theorem (394N).

Lemma Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A. Write C for the fixed-point subalgebra of G. Take any a, b ∈ A. Set c0 = sup{c : c ∈ C, a ∩ c 4τG b}; then a ∩ c0 4τG b and b \ c0 4τG a. proof Enumerate G as hπξ iξ<κ , where κ = #(G). Define haξ iξ<κ , hbξ iξ<κ inductively, setting aξ = (a \ supη<ξ aη ) ∩ πξ−1 (b \ supη<ξ bη ),

bξ = πξ aξ .

Then haξ iξ<κ is a dist family in Aa and hbξ iξ<κ is a dist family in Ab , and supξ<κ aξ is G-τ equidecomposable with supξ<κ bξ . Set a0 = a \ supξ<κ aξ , b0 = b \ supξ<κ bξ , c˜0 = 1 \ upr(a0 , C) = sup{c : c ∈ C, c ∩ a0 = 0}. Then a ∩ c˜0 ⊆ supξ<κ aξ 4τG b, so c˜0 ⊆ c0 . Now b0 ⊆ c˜0 . P P?? Otherwise, because c˜0 = 1 \ supξ<κ πξ a0 (394Gb), there must be a ξ < κ such that 0 0 πξ a ∩ b 6= 0. But in this case d = a0 ∩ πξ−1 b0 6= 0, and we have d ⊆ (a \ supη<ξ aη ) ∩ πξ−1 (b \ supη<ξ bη ), so that d ⊆ aξ , which is absurd. X XQ Q Consequently b \ c˜0 ⊆ supξ<κ bξ 4τG a. Now take any c ∈ C such that a ∩ c 4τG b, and consider c0 = c \ c˜0 . Then b0 ∩ c0 = 0, that is, b ∩ c0 = supξ<κ bξ ∩ c0 , which is G-τ -equidecomposable with supξ<κ aξ ∩ c0 = (a \ a0 ) ∩ c0 (394Gd). But now a ∩ c0 = a ∩ c ∩ c0 4τG b ∩ c0 4τG (a ∩ c0 ) \ (a0 ∩ c0 ); because G is fully non-paradoxical, a0 ∩ c0 must be 0, that is, c0 ⊆ c˜0 and c0 = 0. As c0 is arbitrary, c0 ⊆ c˜0 and c0 = c˜0 . So c0 has the required properties. Remark By analogy with the notation I used in discussing the Hahn decomposition of countably additive functionals (326O), we might denote c0 as ‘[[a 4τG b]]’, or perhaps ‘[[a 4τG b]]C ’, ‘the region (in C) where a 4τG b’. The same notation would write upr(a, C) as ‘[[a 6= 0]]C ’. 394I The construction I wish to use depends essentially on L0 spaces as described in §364. The next step is the following. Lemma Let A be a Dedekind complete Boolean algebra, not {0}, and G a fully non-paradoxical subgroup of Aut A. Let C be the fixed-point subalgebra of G. Suppose that a, b ∈ A and that upr(a, C) = 1. Then there are non-negative u, v ∈ L0 = L0 (C) such that [[u ≥ n]] = max{c : c ∈ C, there is a dist family hdi ii
for every n ∈ N. Moreover, we can arrange that (i) [[u ∈ N]] = [[v ∈ N]] = 1,

600

Measurable algebras

394I

(ii) [[v > 0]] = upr(b, C), (iii) u ≤ v ≤ u + χ1. Remark By writing ‘max’ in the formulae above, I mean to imply that the elements [[u ≥ n]], [[v ≤ n]] belong to the sets described. proof (a) Choose hcn in∈N , hbn in∈N as follows. Given hbi ii 0]] = c00 = upr(b, C). Because cn ⊆ c0n ⊆ cn−1 , (n + 1)χcn ≤ (n + 1)χc0n ≤ nχcn−1 + χ1 for each n ≥ 1, and u ≤ v ≤ u + χ1. (d) Now set Cn = {c : c ∈ C, there is a dist family hdi ii
bi ∩ c0 4τG a ∩ c0 4τG di ∩ c0 , while also b0n ∩ c0 4τG a ∩ c0 4τG dn ∩ c0 . Take d ⊆ dn ∩ c0 such that b0n ∩ c0 is G-τ -equidecomposable with d. Then b ∩ c0 = (b0n ∩ c0 ) ∪ supi
394K

Kawada’s theorem

601

Cn0 = {c : c ∈ C, there is a family hdi ii
then

1 \ c0n

=

max Cn0

for every n.

α) If n = 0, then of course (interpreting sup ∅ as 0) 1 \ c00 ∈ C00 because b ⊆ c00 . For each n ∈ N, set P P(α ˜bn = bn ∪ (b0 \ cn ) = (bn ∩ cn ) ∪ (b0 \ cn ). n n Because bn 4τG a and b0n \ cn 4τG a, we have bn ∩ cn 4τG a ∩ cn and b0n \ cn 4τG a \ cn , so ˜bn 4τG a (394Cd). If we look at supi
Next, there are sets J, J ⊆ A such that d 4τG a for every d ∈ J ∪ J 0 , #(J) ≤ j, #(J 0 ) ≤ k, sup J ⊇ b1 ∩ cjk and sup J 0 ⊇ b2 ∩ cjk . So sup(J ∪ J 0 ) ⊇ (b1 ∪ b2 ) ∩ cjk and J ∪ J 0 witnesses that cjk ⊆ [[db1 ∪ b2 : ae ≤ j + k]]. Q Q Accordingly cjk ⊆ [[db1 : ae + db2 : ae − db1 ∪ b2 : ae ≥ 0]]. Now as supj,k∈N cjk = 1, we must have db1 ∪ b2 : ae ≤ db1 : ae + db2 : ae.

602

Measurable algebras

394K

(d) This time, set cjk = [[bb1 : ac = j]] ∩ [[bb2 : ac = k]] for j, k ∈ N. Then cjk ⊆ [[bb1 ∪ b2 : ac ≥ j + k]] ∩ [[bb1 : ac + bb2 : ac = j + k]] for every j, k ∈ N. P P Once again, we surely have cjk ⊆ [[bb1 : ac + bb2 : ac = j + k]]. Next, we can find a family hdi ii<j+k such that hdi ii<j is dist, a ∩ cjk 4τG di ⊆ b1 for every i < k, hdi ij≤i<j+k is dist, a ∩ cjk 4τG di ⊆ b2 for j ≤ i < j + k. As b1 ∩ b2 = 0, the whole family hdi ii<j+k is dist and witnesses that cjk ⊆ [[bb1 ∪ b2 : ac ≥ j + k]]. Q Q So cjk ⊆ [[bb1 ∪ b2 : ac − bb1 : ac − bb2 : ac ≥ 0]] Since supj,k∈N cjk = 1, as before, we must have bb1 ∪ b2 : ac ≥ bb1 : ac + bb2 : ac. (e) ?? Otherwise, there must be some k ∈ N such that c0 = c ∩ [[bb : ac = k]] ∩ [[db : ae > k]] 6= 0. Let hdi ii
db : a2 e ≤ db : a1 e × da1 : a2 e.

proof I use the same method as in 394K. As usual, write G∗τ for the full local semigroup generated by G. (a) For j, k ∈ N set cj,k = [[bb : a1 c = j]] ∩ [[ba1 : a2 c = k]]. Then cj,k ⊆ [[bb : a1 c × ba1 : a2 c = jk]] ∩ [[bb : a2 c ≥ jk]]. P P Write c for cj,k . As in parts (c) and (d) of the proof of 394K, it is elementary that c is included in [[bb : a1 c × ba1 : a2 c = jk]]; what we need to check is that c ⊆ [[bb : a2 c ≥ jk]]. Again, we may suppose that c 6= 0. There are families hdi ii<j , hd∗l il
394M

Kawada’s theorem

603

Then cj,k ⊆ [[db : a1 e × da1 : a2 e = jk]] ∩ [[db : a2 e ≤ jk]]. P P Write c for cj,k . Then c ⊆ [[db : a1 e × da1 : a2 e = jk]]. There are families hdi ii<j , hd∗l il
i<j,l

l

i<j

= sup di ∩ c ⊇ b ∩ c. i<j

Q So hφi (d0i ∩ d∗l )ii<j,l 0 there is an a ∈ A such that upr(a, C) = 1 and db : ae ≤ bb : ac + ²b1 : ac for every b ∈ A. proof (a) Set B = {d : d ⊆ a, d 4τG a \ d} and let D ⊆ B be a maximal subset such that upr(d, C) ∩ upr(d0 , C) = 0 for all distinct d, d0 ∈ D. Set b = sup D. For any d ∈ D, d 4τG a \ d, so b ∩ upr(d, C) = d ∩ upr(d, C) 4τG (a \ d) ∩ upr(d, C) = (a \ b) ∩ upr(d, C) ⊆ a \ b by 394Gc. By 394H, b = b ∩ supd∈D upr(d, C) 4τG a \ b. ?? Suppose, if possible, that a0 = a \ upr(b, C) is not a relative atom over C. Let d0 ⊆ a0 be an element not expressible as a0 ∩ c for any c ∈ C; then d0 6= a ∩ upr(d0 , C) and there must be a π ∈ G such that d1 = πd0 ∩ a \ d0 is non-zero. In this case d1 4τG π −1 d1 ⊆ d0 ⊆ a \ d1 , so d1 ∈ B; but also d1 ∩ upr(d, C) ⊆ d1 ∩ upr(b, C) = 0, so upr(d1 , C) ∩ upr(d, C) = 0, for every d ∈ D, and we ought to have put d1 into D. X X Thus b has the required properties. (b)(i) For every n ∈ N we can find an ∈ A and cn ∈ C such that upr(an , C) = 1, an \ cn is a relative atom over C, and b1 : an c ≥ 2n χcn . P P Induce on n. The induction starts with a0 = c0 = 1, because b1 : 1c = χ1. For the inductive step, having found an and cn , let d ⊆ an ∩ cn be such that d 4τG an ∩ cn \ d and an ∩ cn \ upr(d, C) is a relative atom over C, as in (a). Set cn+1 = upr(d, C), an+1 = (an \ cn+1 ) ∪ d; then upr(an+1 , C) = upr(an \ cn+1 , C) ∪ upr(d, C) = (upr(an , C) \ cn+1 ) ∪ cn+1 = (1 \ cn+1 ) ∪ cn+1 = 1 by 314Vb-314Vc and the inductive hypothesis. We have cn+1 ∩ d 4τG cn+1 ∩ an \ d, so cn+1 ∩ an+1 = d ⊆ an ,

cn+1 ∩ an+1 4τG an \ d,

and ban : an+1 c ≥ 2χcn+1 ; by 394L, b1 : an+1 c ≥ b1 : an c × ban : an+1 c ≥ 2n χcn × 2χcn+1 = 2n+1 χcn+1 . If

604

Measurable algebras

394M

b ⊆ an+1 \ cn+1 = (an \ cn ) ∪ (an ∩ cn \ cn+1 ), then, because both on the right are relative atoms over C, there are c0 , c00 ∈ C such that b = (b ∩ an \ cn ) ∪ (b ∩ an ∩ cn \ cn+1 ) = (c0 ∩ an \ cn ) ∪ (c00 ∩ an ∩ cn \ cn+1 ) = c ∩ an+1 \ cn+1 where c = (c0 \ cn ) ∪ (c00 ∩ cn ) belongs to C. So an+1 \ cn+1 is a relative atom over C. Thus the induction continues. Q Q (ii) Now suppose that ² > 0. Take n such that 2−n ≤ ², and consider an , cn taken from (i) above. Let b ∈ A. Set c = [[db : an e − bb : an c − ²b1 : an c > 0]] ∈ C. Since we know that ²b1 : an c ≥ 2−n 2n χcn = χcn ,

db : an e ≤ bb : an c + χ1,

we must have c ∩ cn = 0. But this means that an ∩ c is a relative atom over C. By 394Ke, c is included in [[db : an e − bb : an c = 0]]; as also b1 : an c ≥ χ1 (394Ka), c must be zero, that is, db : an e ≤ bb : an c+²b1 : an c. 394N

We are at last ready for the theorem.

Theorem Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Then there is a unique function θ : A → L∞ (C) such that (i) θ is additive, non-negative and order-continuous; (ii) [[θa > 0]] = upr(a, C) for every a ∈ A; in particular, θa = 0 iff a = 0; (iii) θ1 = χ1; (iv) θ(a ∩ c) = θa × χc for every a ∈ A, c ∈ C; in particular, θc = χc for every c ∈ C; (v) If a, b ∈ A are G-τ -equidecomposable, then θa = θb; in particular, θ is G-invariant. proof If A = {0} this is trivial; so I suppose henceforth that A 6= {0}. (a) Set A∗ = {a : a ∈ A, upr(a, C) = 1} and for a ∈ A∗ , b ∈ A set θa (b) =

db : ae ∈ L0 = L0 (C); b1 : ac

the first thing to note is that because b1 : ac ≥ χ1, we can always do the divisions to obtain elements θa (b) of L0 (A) (364P). Set θb = inf a∈A∗ θa b 0

for b ∈ A. (Note that L (C) is Dedekind complete (364O), so the infimum is defined.) (b) The formulae of 394K tell us that, for a ∈ A∗ and b1 , b2 ∈ A, θa 0 = 0,

θa b1 ≤ θa b2 if b1 ⊆ b2 ,

θa (b1 ∪ b2 ) ≤ θa b1 + θa b2 , θa 1 ≥ χ1. It follows at once that θ0 = 0,

θb1 ≤ θb2 if b1 ⊆ b2 , θ1 ≥ χ1.

(c) For each n ∈ N there is an en ∈ A∗ such that db : en e ≤ bb : en c + 2−n b1 : en c for every b ∈ A (394M). Now θen b ≤ θa b + 2−n db : ae for every a ∈ A∗ , b ∈ A. P Pda : en e ≤ ba : en c + 2−n b1 : en c, so

394N

Kawada’s theorem

605

da : en e × b1 : ac ≤ ba : en c × b1 : ac + 2−n b1 : en c × b1 : ac ≤ b1 : en c + 2−n b1 : en c × b1 : ac (by 394L); accordingly db : en e × b1 : ac ≤ db : ae × da : en e × b1 : ac (by the other half of 394L) ≤ db : ae × b1 : en c + 2−n db : ae × b1 : en c × b1 : ac and, dividing by b1 : ac × b1 : en c, we get θen b ≤ θa b + 2−n db : ae. Q Q (d) Now θ is additive. P P Taking hen in∈N from (c), observe first that inf n∈N θen b ≤ θa b + inf n∈N 2−n db : ae = θa b for every a ∈ A∗ , b ∈ A, so that θb = inf n∈N θen b for every b. Now suppose that b1 , b2 ∈ A and b1 ∩ b2 = 0. Then, for any n ∈ N, db1 : en e + db2 : en e ≤ bb1 : en c + bb2 : en c + 2−n+1 b1 : en c ≤ bb1 ∪ b2 : en c + 2−n+1 b1 : en c (by 394Kd) ≤ db1 ∪ b2 : en e + 2−n+1 b1 : en c. Dividing by b1 : en c, we have θb1 + θb2 ≤ θen b1 + θen b2 ≤ θen (b1 ∪ b2 ) + 2−n+1 χ1. Taking the infimum over n, we get θb1 + θb2 ≤ θ(b1 ∪ b2 ). 0

∗

In the other direction, if a, a ∈ A and n ∈ N, θ(b1 ∪ b2 ) ≤ θen (b1 ∪ b2 ) ≤ θen (b1 ) + θen (b2 ) ≤ θa (b1 ) + 2−n db1 : ae + θa0 (b2 ) + 2−n db2 : a0 e. As n is arbitrary, θ(b1 ∪ b2 ) ≤ θa (b1 ) + θa0 (b2 ); as a and a0 are arbitrary, θ(b1 ∪ b2 ) ≤ θb1 + θb2 (using 351Dc). As b1 and b2 are arbitrary, θ is additive. Q Q We see also that d1 : en e ≤ (1 + 2−n )b1 : en c, so that θen 1 ≤ (1 + 2−n )χ1 for each n; since we already know that θ1 ≥ χ1, we have θ1 = χ1 exactly. (e) If c ∈ C then [[θc > 0]] ⊆ [[θ1 c > 0]] ⊆ [[dc : 1e > 0]] = upr(c, C) = c (394I(ii)). It follows that θ(b ∩ c) ≤ θb ∧ θc ≤ θb × χc for any b ∈ A, c ∈ C. Similarly, θ(b \ c) ≤ θb × χ(1 \ c); adding, we must have equality in both, and θ(b ∩ c) = θb × χc. Rather late, I point out that 0 ≤ θa ≤ θ1 = χ1 ∈ L∞ = L∞ (C) for every a ∈ A, so that θa ∈ L∞ for every a. (f ) If b ∈ A \ {0}, then [[θb > 0]] ⊆ [[θ1 b > 0]] ⊆ [[db : 1e > 0]] = upr(b, C)

606

Measurable algebras

394N

by 394I(ii) again. ?? Suppose, if possible, that [[θb > 0]] 6= upr(b, C). Set c0 = upr(b, C) \ [[θb > 0]], a0 = b ∪ (1 \ upr(b, C)) ∈ A∗ . Let k ≥ 1 be such that c1 = c0 ∩ [[d1 : a0 e ≤ k]] 6= 0. Then a0 ∩ c1 = b ∩ c1 , so θa0 × χc1 = θ(a0 ∩ c1 ) = θ(b ∩ c1 ) = θb × χc1 = 0. By 364Nb, there is an a ∈ A∗ such that c1 6⊆ [[θa a0 × χc1 ≥ k1 ]], that is, c2 = c1 ∩ [[θa a0 < k1 ]] 6= 0. Now c2 ⊆ [[b1 : ac − kda0 : ae > 0]] ⊆ [[d1 : a0 e × da0 : ae − kda0 : ae > 0]] ⊆ [[d1 : a0 e > k]], which is impossible, as c2 ⊆ c1 . X X Thus [[θb > 0]] = upr(b, C). In particular, θb = 0 iff b = 0. (g) If b, b0 ∈ A and b 4τG b0 , then θb ≤ θb0 . P P For every a ∈ A∗ , db : ae ≤ db0 : ae (394Kb) so θa b ≤ θa b0 . 0 0 0 P?? Otherwise, by 394H, there is a non-zero c0 ⊆ c Q Q So if b, b ∈ A and c = [[θb − θb > 0]], b ∩ c 4τG b. P 0 τ 0 0 such that b ∩ c 4G b . But in this case θb × χc = θ(b ∩ c0 ) ≤ θb0 and c0 ⊆ [[θb0 − θb ≥ 0]]. X XQ Q P (h) If hai ii∈I is any dist i∈I θai , where the sum is to be P family in A with supremum a, θa = interpreted as supJ⊆I is finite i∈J θai . P P Induce on #(I). If #(I) is finite, this is just finite additivity ((d) above). For the inductive step to #(I) = κ ≥ ω, we may suppose that I is actually equal to the cardinal κ. Of course P θa ≥ θ(supξ∈J aξ ) = ξ∈J θaξ P for every finite J ⊆ κ, so (because L∞ (C) is Dedekiond complete) u = ξ<κ θaξ is defined, and u ≤ θa. For ζ < κ, set bζ = supξ<ζ aξ . By the inductive hypothesis, P P θbζ = ξ<ζ θaξ = supJ⊆ζ is finite ξ∈J θaξ ≤ u. P At the same time, if J ⊆ κ is finite, there is some ζ < κ such that J ⊆ ζ, so that ξ∈J θaξ ≤ θbζ ; accordingly supζ<κ θbζ = u. ?? Suppose, if possible, that u < θa; set v = θa − u. Take δ > 0 such that c0 = [[v > δ]] 6= 0. Let ζ < κ be such that c1 = c0 \ [[u − θbζ > δ]] is non-zero (cf. 364Nb). Now u − θbζ ≤ θ(a \ bζ ), so c1 ⊆ [[θ(a \ bζ ) > 0]] = upr(a \ bζ , C), and c1 ∩ (a \ bζ ) 6= 0; there is therefore an η 0 ≥ ζ such that d = c1 ∩ aη0 6= 0. Since θd ≤ u − θbζ and c1 ⊆ [[u − θbζ ≤ δ]] ∩ [[v > δ]], [[v − θd > 0]] ⊇ c1 . Choose hdξ iξ<κ inductively, as follows. Given that hdη iη<ξ is a dist family in Aa\d such that dη is Gτ -equidecomposable with aη ∩ c1 for every η < ξ, then eξ = supη<ξ dη is G-τ -equidecomposable with bξ ∩ c1 , so that θeξ ≤ θbξ , and [[θ(a \ (d ∪ eξ )) − θaξ > 0]] = [[θa − θd − θeξ − θaξ > 0]] ⊇ [[θa − θd − θbξ − θaξ > 0]] = [[θa − θd − θbξ+1 > 0]] ⊇ [[v − θd > 0]] ⊇ c1 . aξ ∩ c1 4τG

By (g), a \ (d ∪ eξ ); take dξ ⊆ a \ (d ∪ eξ ) G-τ -equidecomposable with aξ ∩ c1 , and continue. At the end of this induction, we have a dist family hdξ iξ<κ in Aa\d such that dξ is G-τ -equidecomposable with aξ ∩ c1 for every ξ. But this means that a0 = supξ<κ dξ is G-τ -equidecomposable with a ∩ c1 , while a0 ⊆ (a \ d) ∩P c1 ; since d ∩ c1 6= 0, G cannot be fully non-paradoxical. X X Thus θa = u = ξ<κ θaξ and the induction continues. Q Q (i) It follows that P (α) If B ⊆ A is non-empty and upwards-directed and has S θ is order-continuous. P supremum e, then b∈B Ab is order-dense in Ae , so includes a partition of unity A of Ae ; now (h) tells us that P θe = a∈A θa ≤ supb∈B θb. Since of course θb ≤ θe for every b ∈ B, θe = supb∈B θb. (β) If B ⊆ A is non-empty and downwards-directed and has infimum e, then, using (α), we see that θ1 − θe = θ(1 \ e) = supb∈B θ(1 \ b) = supb∈B θ1 − θb, so that θe = inf b∈B θb. Q Q (j) I still have to show that θ is unique. Let θ0 : A → L∞ be any non-negative order-continuous G-invariant additive function such that θ0 c = χc for every c ∈ C.

394P

Kawada’s theorem

607

(i) Just as in (e) of this proof, but more easily, we see that θ0 (b ∩ c) = θ0 b × χc for every b ∈ A, c ∈ C. (ii) If hai ii∈I is a dist family in A with supremum a, then hsupi∈J ai iJ⊆I is finite is an upwardsdirected family with supremum a, so that P P θ0 a = supJ⊆I is finite θ0 (supi∈J ai ) = supJ⊆I is finite i∈J θ0 ai = i∈I θ0 ai . (iii) θ0 a = θ0 b whenever a and b are G-τ -equidecomposable. P P Take a partition hai ii∈I of a and a family hπi ii∈I in G such that hπi ai ii∈I is a partition of b. Then P P θ0 a = i∈I θ0 ai = i∈I θ0 πi ai = θ0 b. Q Q Consequently θ0 a ≤ θ0 b whenever a 4τG b. (iv) Take a ∈ A∗ , b ∈ A and for j, k ∈ N set cjk = [[b1 : ac = j]] ∩ [[db : ae = k]]. Then db : ae × χcjk ≥ θ0 b × b1 : ac × χcjk . P P If cjk = 0 this is trivial; suppose cjk 6= 0. Now we have sets I, J such that #(I) = j, #(J) ≤ k, a ∩ cjk 4τG d for every d ∈ I, e 4τG a for every e ∈ J, I is dist, and b ∩ cjk ⊆ sup J. So θ0 b × b1 : ac × χcjk = jθ0 b × χcjk = jθ0 (b ∩ cjk ) ≤ j 0

≤ jkθ (a ∩ cjk ) ≤ k

X

X

θ0 (e ∩ cjk )

e∈J 0

θ (d ∩ cjk ) ≤ kθ0 cjk

d∈I

= kχcjk = db : ae × χcjk . Q Q Summing over j and k, db : ae ≥ θ0 b × b1 : ac, that is, θa b ≥ θ0 b. Taking the infimum over a, θb ≥ θ0 b. But also θb = χ1 − θ(1 \ b) ≤ χ1 − θ0 (1 \ b) = θ0 b, so θb = θ0 b. As b is arbitrary, θ = θ0 . This completes the proof. 394O

We have reached the summit. The rest of the section is a list of easy corollaries.

Theorem Let A be a Dedekind complete Boolean algebra, not {0}, and G a fully non-paradoxical subgroup of Aut A. Then there is a G-invariant additive functional ν : A → [0, 1] such that ν1 = 1. proof Let C be the fixed-point subalgebra of G, and θ : A → L∞ (C) the function of 394N. By 311D, there is a ring homomorphism ν0 : C → {0, 1} such that ν0 1 = 1; now ν0 can also be regarded as an additive functional from C to R. Let f0 : L∞ (C) → R be the corresponding positive linear functional (363K). Set ν = f0 θ. Then ν is order-preserving and additive because f0 and θ are, ν1 = f0 (χ1) = ν0 1 > 0, and ν is G-invariant because θ is. 394P Theorem Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Then the following are equiveridical: (i) A is a measurable algebra; (ii) C is a measurable algebra; (iii) there is a strictly positive G-invariant countably additive real-valued functional on A. proof (iii)⇒(i)⇒(ii) are trivial. For (ii)⇒(iii), let θ : A → L∞ (C) be the function of 394N, and ν¯ : C → R a strictly positive countably additive functional. Let f : L∞ (C) → R be the corresponding linear operator; then f is sequentially order-continuous (363K again). Set µ ¯ = f θ. Then µ ¯ is additive and order-preserving and sequentially order-continuous because f and θ are. It is also strictly positive, because if a ∈ A \ {0} then θa > 0 (394N(ii)), that is, there is some δ > 0 such that [[θa > δ]] 6= 0, so that µ ¯a ≥ δ¯ ν [[θa > δ]] > 0. Finally, µ ¯ is G-invariant because θ is.

608

Measurable algebras

394Q

394Q Corollary: Kawada’s Theorem Let A be a Dedekind complete Boolean algebra such that Aut A is ergodic and fully non-paradoxical. Then A is measurable. proof This is the case C = {0, 1} of 394P. 394R Thus the existence of an ergodic fully non-paradoxical subgroup is a sufficient condition for a Dedekind complete Boolean algebra to be measurable. It is not quite necessary, because if a measure algebra A is not homogeneous then its automorphism group is not ergodic. But for homogeneous algebras the condition is necessary as well as sufficient, by the following result. Proposition If (A, µ ¯) is a homogeneous totally finite measure algebra, the group G = Autµ¯ A of measurepreserving automorphisms of A is ergodic. proof If A = {0, 1} this is trivial. Otherwise, A is atomless. If a ∈ A \ {0, 1}, set γ = min(¯ µa, µ ¯(1 \ a)) > 0; then there are b ⊆ a, d ⊆ 1 \ a such that µ ¯b = µ ¯d = γ. By 382Fb, there is a π ∈ G such that πb = d, so that πa 6= a. As a is arbitrary, the fixed-point subalgebra of G is {0, 1}. 394X Basic exercises (a) Re-write the section on the assumption that every group G is ergodic, that is, C is always {0, 1}, so that L0 (C) may be identified with R, the functions d. . . e and b. . . c become real-valued, the functionals θa (394N) become submeasures and θ becomes a measure. (b) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with fixed-point subalgebra C. Suppose that hci ii∈I is a partition of unity in C and that a, b ∈ A are such that a ∩ ci 4τG b for every i ∈ I. Show that a 4τG b. (c) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with fixed-point subalgebra C. Show that A is relatively atomless over C iff the full local semigroup generated by G has many involutions (definition: 381P). (d) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Show that the following are equiveridical: (i) there is a strictly positive real-valued additive functional on A; (ii) there is a strictly positive real-valued additive functional on C; (iii) there is a strictly positive G-invariant real-valued additive functional on A. (e) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Show that the following are equiveridical: (i) there is a non-zero completely additive functional on A; (ii) there is a non-zero completely additive functional on C; (iii) there is a non-zero G-invariant completely additive functional on A. (f ) Let A be a ccc Dedekind complete Boolean algebra. Show that it is a measurable algebra iff there is a fully non-paradoxical subgroup G of Aut A such that the fixed-point subalgebra of G is purely atomic. (g) Let (A, µ ¯) be a localizable measure algebra. Show that the following are equiveridical: (i) Autµ¯ A is ergodic; (ii) A is quasi-homogeneous in the sense of 374G; (iii) whenever a, b ∈ A and µa = µb then the principal ideals Aa , Ab are isomorphic. (h) Let (A, µ ¯) be a localizable measure algebra. Show that Autµ¯ A is fully non-paradoxical iff (i) for every infinite cardinal κ, the Maharam-type-κ component of A (definition: 332G) has finite measure (ii) for every γ ∈ ]0, ∞[ there are only finitely many atoms of measure γ. 394Y Further exercises (a) Let A be a Dedekind complete Boolean algebra, G a subgroup of Aut A, and G∗τ the full local semigroup generated by G. For φ, ψ ∈ G∗τ , say that φ ≤ ψ if ψ extends φ. (i) Show that every member of G∗τ can be extended to a maximal member of G∗τ . (ii) Show that G is fully non-paradoxical iff every maximal member of G∗τ is actually a Boolean automorphism of A. (b) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Show that A is ccc iff C is ccc. (Hint: if C is ccc, L∞ (C) has the countable sup property (363Yb).)

394 Notes

Kawada’s theorem

609

(c) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with fixed-point subalgebra C. Show that A is weakly (σ, ∞)-distributive iff C is. (d) Let A be a Dedekind complete Boolean algebra, G an ergodic subgroup of A, and G∗τ the full local semigroup generated by G. Suppose that there is a non-zero a ∈ A for which there is no φ ∈ G∗τ such that φa ⊂ a. Show that there is a measure µ ¯ such that (A, µ ¯) is a localizable measure algebra. (Hint: show that Aa is a measurable algebra.) 394Z Problem Suppose that A is a Dedekind complete Boolean algebra, not {0}, and G a subgroup of Aut A such that whenever hai ii≤n is a finite partition of unity in A and we are given πi , πi0 ∈ G for every i ≤ n, then the elements π0 a0 , π 0 a0 , π1 a1 , π10 a1 , . . . , πn0 an are not all dist. Must there be a non-zero non-negative G-invariant finitely additive functional θ on A? (See ‘Tarski’s theorem’ in the notes below.) 394 Notes and comments Regarded as a sufficient condition for measurability, Kawada’s theorem suffers from the obvious defect that it is going to be rather rarely that we can the existence of an ergodic fully non-paradoxical group of automorphisms without having some quite different reason for supposing that our algebra is measurable. If we think of it as a criterion for the existence of a G-invariant measure, rather than as a criterion for measurability in the abstract, it seems to make better sense. But if we know from the start that the algebra A is measurable, the argument short-circuits, as we shall see in §395. τ I take the trouble to include the ‘τ ’ in every ‘G-τ -equidecomposable’, ‘G∗τ ’ and ‘4G ’ because there are important variations on the concept, in which the partitions hai ii∈I of 394A are required to be finite or countable. Indeed Tarski’s theorem relies on one of these. I spell it out because it is close to Kawada’s in spirit, though there are significant differences in the ideas needed in the proof: Let X be a set and G a subgroup of Aut PX. Then the following are equiveridical: (i) there is a G-invariant additive functional θ : PX → [0, 1] such that θA = 1; (ii) there are no A0 , . . . , An , π0 , . . . , πn , π00 , . . . , πn0 such that A0 , . . . , An are subsets of X covering X, π0 , . . . , πn0 all belong to G, and π0 A0 , π00 A0 , π1 A1 , π10 A1 , . . . , πn0 An are all dist. For a proof, see 449I in Volume 4; for an illuminating discussion of this theorem, see Wagon 85, Chapter 9. But it seems to be unknown whether the natural translation of this result is valid in all Dedekind complete Boolean algebras (394Z). Note that we are looking for theorems which do not depend on any special properties of the group G or the Boolean algebra A. For abelian or ‘amenable’ groups, or weakly (σ, ∞)-distributive algebras, for instance, much more can be done, as described in 395Ya and §449. The methods of this section can, however, be used to prove similar results for countable groups of automorphisms on Dedekind σ-complete Boolean algebras; I will return to such questions in §448. As noted, Kawada (Kawada 44) treated the case in which the group G of automorphisms is ergodic, that is, the fixed-point subalgebra C is trivial. Under this hypothesis the proof is of course very much simpler. (You may find it useful to reconstruct the original version, as suggested in 394Xa.) I give the more general argument partly for the sake of 394O, partly to separate out the steps which really need ergodicity from those which depend only on non-paradoxicality, partly to prepare the ground for the countable version in the next volume, partly to show off the power of the construction in §364, and partly to get you used to ‘Boolean-valued’ arguments. A bolder use of language could indeed simplify some formulae slightly by writing (for instance) [[kda0 : ae < b1 : ac]] in place of [[b1 : ac − kda0 : ae > 0]] (see part (f) of the proof of 394N). As in §387, the differences involved in the extension to non-ergodic groups are, in a sense, just a matter of technique; but this time the technique is more obtrusive. The presentation here owes a good deal to Nadkarni 90 and something to Becker & Kechris 96.

610

Measurable algebras

§395 intro.

395 The Hajian-Ito theorem In the notes to the last section, I said that the argument there short-circuits if we are told that we are dealing with a measurable algebra. The point is that in this case there is a much simpler criterion for the existence of a G-invariant measure (395B(ii)), with a proof which is independent of §394 in all its non-trivial parts, which makes it easy to prove that non-paradoxicality is sufficient as well as necessary. 395A Lemma Let (A, µ ¯) be a localizable measure algebra. (a) Let π ∈ Aut A be a Boolean automorphism (not necessarily measure-preserving). Let Tπ be the corresponding Riesz homomorphism from L0 = L0 (A) to itself (364R). Then there is a unique wπ ∈ (L0 )+ R R such that wπ × v = Tπ v for every v ∈ (L0 )+ . (b) If φ, π ∈ Aut A then wπφ = wφ × Tφ−1 wπ . (c) For each π ∈ Aut A we have a norm-preserving isomorphism Uπ from L2 = L2 (A, µ ¯) to itself defined by setting √ Uπ v = Tπ v × wπ−1 for every v ∈ L2 , and Uπφ = Uπ Uφ for all π, φ ∈ Aut A. proof (a) Applying 365T with ν¯a = µ ¯(πa), we see that there is a unique wπ ∈ (L0 )+ such that for every a ∈ A. If we look at W = {v : v ∈ (L0 )+ ,

R

v × wπ =

R

R a

wπ = µ ¯(πa)

Tπ v},

0

we see that W contains χa for every a ∈ A, that v + v ∈ W and αv ∈ W whenever v, v 0 ∈ W and α ≥ 0, and that supn∈N vn ∈ W whenever hvn in∈N is a non-decreasing sequence in W which is bounded above in L0 . By 364Kd, W = (L0 )+ , as required. (b) For any v ∈ (L0 )+ , Z

Z wπφ × v =

(364Re)

Z Tπφ v =

Tπ Tφ v

Z =

(recalling that Tφ is multiplicative)

Z wπ × Tφ v =

Tφ (Tφ−1 wπ × v)

Z =

wφ × Tφ−1 wπ × v.

As v is arbitrary (and (A, µ ¯) is semi-finite), wπφ = wφ × Tφ−1 wπ . (c)(i) For any v ∈ L0 ,

R

(Tπ v ×

√

wπ−1 )2 =

R

Tπ v 2 × wπ−1 =

R

Tπ−1 Tπ v 2 =

R

v2 .

So Uπ v ∈ L2 and kUπ vk2 = kvk2 whenever v ∈ L2 , and Uπ is a norm-preserving operator on L2 . (ii) Now consider Uπφ . For any v ∈ L2 , we have √ √ Uπ Uφ v = Tπ (Tφ v × wφ−1 ) × wπ−1 p = Tπ Tφ v × Tπ wφ−1 × wπ−1 (using 364Rd) = Tπφ v × (by (b) above) = Uπφ v.

√

wφ−1 π−1

395B

The Hajian-Ito theorem

611

So Uπφ = Uπ Uφ . (iii) Writing ι for the identity operator on A, we see that Tι is the identity operator on L0 , wι = χ1 and Uι is the identity operator on L2 . Since Uπ−1 Uπ = Uπ Uπ−1 = Uι , Uπ : L2 → L2 is an isomorphism, with inverse Uπ−1 , for every π ∈ Aut A. 395B Theorem (Hajian & Ito 69) Let A be a measurable algebra and G a subgroup of Aut A. Then the following are equiveridical: (i) there is a G-invariant functional ν¯ such that (A, ν¯) is a totally finite measure algebra; (ii) whenever a ∈ A \ {0} and hπn in∈N is a sequence in G, hπn ain∈N is not dist; (iii) G is fully non-paradoxical (definition: 394E). proof (a) (i)⇒(iii) by the argument of 394F, and (iii)⇒(ii) by the criterion (ii) of 394E. So for the rest of the proof I assume that (ii) is true and seek to prove (i). (b) Let µ ¯ be such that (A, µ ¯) is a totally finite measure algebra. If a ∈ A \ {0}, then inf π∈G µ ¯(πa) > 0. P P?? Otherwise, let hπn in∈N be a sequence in G such that µ ¯πn a ≤ 2−n for each n ∈ N. Set bn = supk≥n πk a for each n; then inf n∈N bn = 0, so that inf n∈N πbn = 0,

limn→∞ µ ¯(ππn a) = 0

for every π ∈ Aut A. Choose hni ii∈N inductively so that µ ¯(πn−1 πnj a) ≤ 2−j−2 µ ¯a i whenever i < j. Set c = a \ supi<j πn−1 πnj a. i Because

P∞ Pj−1 j=1

i=0

µ ¯(πn−1 πnj a) < µ ¯a, i

c 6= 0, while πni c ∩ πnj c = 0 whenever i < j, contrary to the hypothesis (ii). X XQ Q (c) For each π ∈ G, define ∈ L0 = L0 (A) and Uπ : L2 → L2 as in 395A, where L2 = L2 (A, µ ¯). R w √π If a ∈ A \ {0}, then inf w > 0. P P ? ? Otherwise, there is a sequence hπ i in G such that π∈G a π n n∈N R √ v ≤ 4−n−2 µ ¯a for every n, where vn = wπn . In this case, µ ¯(a ∩ [[vn ≥ 2−n ]]) ≤ 2−n−2 µ ¯a for every n, so a n that b = a \ supn∈N [[vn ≥ 2−n ]] is non-zero. But now µ ¯(πn b) =

R

b

wπn =

R

b

vn2 ≤ 4−n µ ¯b → 0

as n → ∞, contradicting (b) above. X XQ Q (d) Write e = χ1 for the standard weak order unit of L0 or L2 . Let C ⊆ L2 be the convex hull of {Uπ e : π ∈ G}. Then C and its norm closure C are G-invariant in the sense that Uπ v ∈ C, Uπ v 0 ∈ C whenever v ∈ C, v 0 ∈ C and π ∈ G. By 3A5Ld, there is a unique u0 ∈ C such that ku0 k2 ≤ kuk2 for every u ∈ C. Now if π ∈ G, Uπ u0 ∈ C, while kUπ u0 k2 = ku0 k2 ; so Uπ u0 = u0 . Also, if a ∈ A \ {0}, Z

Z

Z u0 ≥ inf

(because u 7→

R a

a

u∈C

u is k k2 -continuous)

Za

= inf

π∈G

a

Z

√ Tπ e × wπ−1 a Z √ = inf wπ > 0

Uπ e = inf

= inf

by (c). So [[u0 > 0]] = 1.

u∈C

a

Z π∈G

u

u = inf

a

π∈G

√

wπ−1

π∈G

a

R (e) For a ∈ A, set ν¯a = a u20 . Because u0 ∈ L2 , ν¯ is a non-negative countably additive functional on A; because [[u20 > 0]] = [[u0 > 0]] = 1, ν¯ is strictly positive, and (A, ν¯) is a totally finite measure algebra. Finally, ν¯ is G-invariant. P P If a ∈ A and π ∈ G, then

612

Measurable algebras

Z

395B

Z

Z u20 × χ(πa) = Tπ (Tπ−1 u20 × χa) Zπa Z √ 2 = wπ × Tπ−1 u0 × χa = (Tπ−1 u0 × wπ )2 a Z Z 2 2 = (Uπ−1 u0 ) = u0 = ν¯a. Q Q u20 =

ν¯(πa) =

a

a

So (i) is true. 395C Remark If A is a Boolean algebra and G a subgroup of Aut A, a non-zero element a of A is called weakly wandering if there is a sequence hπn in∈N in G such that hπn ain∈N is dist. Thus condition (ii) of 395B may be read as ‘there is no weakly wandering element of A’. 395X Basic exercises (a) Let (A, µ ¯) be a totally finite measure algebra, and π : A → A an order0 continuous Boolean homomorphism. Let Tπ : L0 (A) → homomorphism. R L (A)Rbe the corresponding Riesz 1 Show that there is a unique wπ ∈ L (A, µ ¯) such that Tπ v = v × wπ for every v ∈ L0 (A)+ . (b) In 395A, show that the map π 7→ Uπ : Aut A → B(L2 ; L2 ) is injective. (c) Let (A, µ ¯) be a probability algebra, π ∈ Autµ¯ A an ergodic measure-preserving automorphism, and φ ∈ Aut A an automorphism. Suppose that φπφ−1 is measure-preserving. Show that φ is measure-preserving. (Hint: compare wφ , wπφ and wφπ in 395A.) (d) Let A be a measurable algebra and G a subgroup of Aut A. Suppose that there is a strictly positive G-invariant finitely additive functional on A. Show that there is a G-invariant µ ¯ such that (A, µ ¯) is a totally finite measure algebra. 395Y Further exercises (a) Let A be a weakly (σ, ∞)-distributive Dedekind complete Boolean algebra and G a subgroup of Aut A. For a, b ∈ A, say that a and b are G-equidecomposable if there are finite partitions of unity hai ii∈I in Aa and hbi ii∈I in Ab , and a family hπi ii∈I in G, such that πi ai = bi for every i ∈ I. Show that the following are equiveridical: (i) G is fully non-paradoxical in the sense of 394E; (ii) if han in∈N is a dist sequence of mutually G-equidecomposable elements of A, they must all be 0. 395 Notes and comments I have separated these few pages from §394 partly because §394 was already up to full weight and partly in order that the ideas here should not be entirely overshadowed by those of the earlier section. It will be evident that the construction of the Uπ in 395A, providing us with a faithful representation, acting on a Hilbert space, of the whole group Aut A, is a basic tool for the study of that group.

3A1E

Set Theory

613

Appendix to Volume 3 Useful Facts This volume assumes a fairly wide-ranging competence in analysis, a solid understanding of elementary set theory and some straightforward Boolean algebra. As in previous volumes, I start with a few pages of revision in set theory, but the absolutely essential material is in §3A2, on commutative rings, which is the basis of the treatment of Boolean rings in §311. I then give three sections of results in analysis: topological spaces (§3A3), uniform spaces (§3A4) and normed spaces (§3A5). Finally, I add six sentences on group theory (§3A6).

3A1 Set Theory 3A1A The axioms of set theory This treatise is based on arguments within, or in principle reducible to, ‘ZFC’, meaning ‘Zermelo-Fraenkel set theory, including the Axiom of Choice’. For discussions of this system, see, for instance, Krivine 71, Jech 78 or Kunen 80. As I remarked in §2A1, I believe that it is helpful, as a matter of general principle, to distinguish between results dependent on the axiom of choice and those which can be proved without it, or with some relatively weak axiom such as ‘countable choice’. (See 134C.) In Volumes 1 and 2, such a distinction is useful in appreciating the special features of different ideas. In the present volume, however, most of the principal theorems require something close to the full axiom of choice, and there are few areas where it seems at present appropriate to work with anything weaker. Indeed, at many points we shall approach questions which are, or may be, undecidable in ZFC; but I postpone discussion of these to Volume 5. In particular, I specifically exclude, for the time being, results dependent on such axioms as the continuum hypothesis. 3A1B Definition Let X be a set. By an enumeration of X I mean a bijection f : κ → X where κ = #(X) (2A1Kb); more often than not I shall express such a function in the form hxξ iξ<κ . In this case I say that the function f , or the family hxξ iξ<κ , enumerates X. You will see that I am tacitly assuming that #(X) is always defined, that is, that the axiom of choice is true. 3A1C Calculation of cardinalities The following formulae are basic. (a) For any sets X and Y , #(X × Y ) ≤ max(ω, #(X), #(Y )). (Enderton 77, p. 64; Jech 78, p. 42; Krivine 71, p. 33; Kunen 80, 10.13.) Qr (b) For any r ∈ N and any family hXi ii≤r of sets, #( i=0 Xi ) ≤ max(ω, maxi≤r #(Xi )). (Induce on r.) S (c) For any family hXi ii∈I of sets, #( i∈I Xi ) ≤ max(ω, #(I), supi∈I #(Xi )). (Jech 78, p. 43; Krivine 71, p. 33; Kunen 80, 10.21.) (d) For any set set [X]<ω of finite subsets of X has cardinal at most max(ω, #(X)). (There is a S X, the r surjection from r∈N X onto [X]<ω . For the notation [X]<ω see 3A1J below.) 3A1D Cardinal exponentiation For a cardinal κ, I write 2κ for #(Pκ). So 2ω = c, and κ+ ≤ 2κ for every κ. (Enderton 77, p. 132; Lipschutz 64, p. 139; Jech 78, p. 24; Krivine 71, p. 25; Halmos 60, p. 93.) 3A1E Definition The class of infinite initial ordinals, or cardinals, is a subclass of the class On of all ordinals, so is itself well-ordered; being unbounded, it is a proper class; consequently there is a unique increasing enumeration of it as hωξ iξ∈On . We have ω0 = ω, ωξ+1 = ωξ+ for every ξ (compare 2A1Fc), S ωξ = η<ξ ωη for non-zero limit ordinals ξ. (Enderton 77, pp. 213-214; Jech 78, p. 25; Krivine 71, p. 31.)

614

Appendix

3A1F

3A1F Cofinal sets (a) If P is any partially ordered set (definition: 2A1A), a subset Q of P is cofinal with P if for every p ∈ P there is a q ∈ Q such that p ≤ q. (b) If P is any partially ordered set, the cofinality of P , cf(P ), is the least cardinal of any cofinal subset of P . Note that cf(P ) = 0 iff P = ∅, and that cf(P ) = 1 iff P has a greatest element. (c) Observe that if P is upwards-directed and cf(P ) is finite, then cf(P ) is either 0 or 1; for if Q is a finite, non-empty cofinal set then it has an upper bound, which must be the greatest element of P . (d) If P is a totally ordered set of cofinality κ, then there is a strictly increasing family hpξ iξ<κ in P such that {pξ : ξ < κ} is cofinal with P . P P If κ = 0 then P = ∅ and this is trivial. Otherwise, let Q be a cofinal subset of P of cardinal κ, and {qξ : ξ < κ} an enumeration of Q. Define hpξ iξ<κ inductively, as follows. Start with p0 = q0 . Given hpη iη<ξ , where ξ < κ, then if pη < qξ for every η < ξ, take pξ = qξ ; otherwise, because #(ξ) ≤ ξ < κ, {pη : η < ξ} cannot be cofinal with P , so there is a pξ ∈ P such that pξ 6≤ pη for every η < ξ, that is, pη < pξ for every η < ξ. Note that there is some η < ξ such that qξ ≤ pη , so that qξ ≤ pξ . Continue. Now hpξ iξ<κ is a strictly increasing family in P such that qξ ≤ pξ for every ξ; it follows at once that {pξ : ξ < κ} is cofinal with P . Q Q (e) In particular, for a totally ordered set P , cf(P ) = ω iff there is a cofinal strictly increasing sequence in P . 3A1G Zorn’s Lemma In Volume 2 I used Zorn’s Lemma only once or twice, giving the arguments in detail. In the present volume I feel that continuing in such a manner would often be tedious; but nevertheless the arguments are not always quite obvious, at least until you have gained a good deal of experience. I therefore take a paragraph to comment on some of the standard forms in which they appear. The statement of Zorn’s Lemma, as quoted in 2A1M, refers to arbitrary partially ordered sets P . A large proportion of the applications can in fact be represented more or less naturally by taking P to beSa family P of sets ordered by ⊆; in such a case, it will be sufficient to check that (i) P is not empty (ii) Q ∈ P for every non-empty totally ordered Q ⊆ P. More often than not, this will in fact be true for all non-empty upwards-directed sets Q ⊆ P, and the line of the argument is sometimes clearer if phrased in this form. Within this class of partially ordered sets, we can distinguish a special subclass. If A is any set and ⊥ any relation on A, we can consider the collection P of sets I ⊆ A such that a ⊥ b for all distinct a, b ∈ I. In this case we need look no farther before declaring ‘P has S a maximal element’; for ∅ necessarily belongs S to P, and if Q is any upwards-directed subset of P, then Q ∈ P. P P If a, b are distinct elements of Q, there are I1 , I2 ∈ Q such that a ∈ I1 , b ∈ I2 ; because Q is upwards-directed, there is an I ∈ Q such that S I1 ∪ I2 ⊆ I, so that a, b are distinct of I ∈ P, and a ⊥ b. Q Q So Q is an upper bound of Q in P; as Q is arbitrary, P satisfies the conditions of Zorn’s Lemma, and must have a maximal element. Another important type of partially ordered set in this context is a family Φ of functions, ordered by saying that f ≤ g if g is an extension of f . In this case, for any non-empty upwards-directed Ψ ⊆ Φ, we shall have a function h defined by saying that S dom h = f ∈Ψ dom f , h(x) = f (x) whenever f ∈ Ψ, x ∈ dom f , and the usual attack is to seek to prove that any such h belongs to Φ. I find that at least once I wish to use Zorn’s Lemma ‘upside down’: that is, I have a non-empty partially ordered set P in which every non-empty totally ordered subset has a lower bound. In this case, of course, P has a minimal element. The point is that the definition of ‘partial order’ is symmetric, so that (P, ≥) is a partially ordered set whenever (P, ≤) is; and we can seek to apply Zorn’s Lemma to either. 3A1H Natural numbers and finite ordinals I remarked in 2A1De that the first few ordinals ∅,

{∅},

{∅, {∅}},

{∅, {∅}, {∅, {∅}}},

...

may be identified with the natural numbers 0, 1, 2, 3, . . . ; the idea being that n = {0, 1, . . . , n − 1} is a set with n elements. If we do this, then the set N of natural numbers becomes identified with the first infinite ordinal ω. This convention makes it possible to present a number of arguments in a particularly elegant

3A2A

Rings

615

form. A typical S example is in 344H. There I wish to describe an inductive construction for a family hKz iz∈S where S = n∈N {0, 1}n . If we think of n as the set of its predecessors, then z ∈ {0, 1}n becomes a function from n to {0, 1}; since the set n has just n , this corresponds well to the idea of z as the list of its n coordinates, except that it would now be natural to index them as z(0), . . . , z(n − 1) rather than as ζ1 , . . . , ζn , which was the language I favoured in Volume 2. An extension of z to a member of {0, 1}n+1 is of the form v = z a i where v(k) = z(k) for k < n and v(n) = i. If w ∈ {0, 1}N , then we can identify the initial segment (w(0), w(1), . . . , w(n − 1)) of its first n coordinates with the restriction w¹n of w to the set n = {0, . . . , n − 1}. 3A1I Definitions (a) If P and Q are lattices (2A1Ad), a lattice homomorphism from P to Q is a function f : P → Q such that f (p ∧ p0 ) = f (p) ∧ f (p0 ) and f (p ∨ p0 ) = f (p) ∨ f (p0 ) for all p, p0 ∈ P . Such a homomorphism is surely order-preserving (313H), for if p ≤ p0 in P then f (p0 ) = f (p ∨ p0 ) = f (p) ∨ f (p0 ) and f (p) ≤ f (p0 ). (b) If P is a lattice, a sublattice of P is a set Q ⊆ P such that p ∨ q and p ∧ q belong to Q for all p, q ∈ Q. (c) A lattice P is distributive if (p ∧ q) ∨ r = (p ∨ r) ∧ (q ∨ r),

(p ∨ q) ∧ r = (p ∧ r) ∨ (q ∧ r)

for all p, q, r ∈ P . 3A1J Subsets of given size The following concepts are used often enough for a special notation to be helpful. If X is a set and κ is a cardinal, write [X]κ = {A : A ⊆ X, #(A) = κ}, [X]≤κ = {A : A ⊆ X, #(A) ≤ κ}, [X]<κ = {A : A ⊆ X, #(A) < κ}. Thus [X]0 = [X]≤0 = [X]<1 = {∅}, [X]2 is the set of doubleton subsets of X, [X]<ω is the set of finite subsets of X, [X]≤ω is the set of countable subsets of X, and so on. 3A1K The next result is one of the fundamental theorems of combinatorics. In this volume it is used in the proofs of Ornstein’s theorem (§386) and the Kalton-Roberts theorem (§392). Hall’s Marriage Lemma Suppose that X and Y are finite sets and R ⊆ X × Y is a relation such that #(R[I]) ≥ #(I) for every I ⊆ X. Then there is an injective function f : X → Y such that (x, f (x)) ∈ R for every x ∈ X. Remark Recall that R[I] is the set {y : ∃ x ∈ I, (x, y) ∈ R} (1A1Bc). ´ s 79, p. 54, Theorem 7; Anderson 87, 2.2.1; Bose & Manvel 84, §10.2. proof Bolloba

3A2 Rings I give a very brief outline of the indispensable parts of the elementary theory of (commutative) rings. I assume that you have seen at least a little group theory. 3A2A Definition A ring is a triple (R, +, .) such that (R, +) is a commutative group; its identity will always be denoted 0 or 0R ; (R, .) is a semigroup, that is, ab ∈ R for all a, b ∈ R and a(bc) = (ab)c for all a, b, c ∈ R; a(b + c) = ab + ac, (a + b)c = ac + bc for all a, b, c ∈ R. A commutative ring is one in which multiplication is commutative, that is, ab = ba for all a, b ∈ R.

616

Appendix

3A2B

3A2B Elementary facts Let R be a ring. (a) a0 = 0a = 0 for every a ∈ R. P P a0 = a(0 + 0) = a0 + a0,

0a = (0 + 0)a = 0a + 0a;

because (R, +) is a group, we may subtract a0 or 0a from each side of the appropriate equation to see that 0 = a0, 0 = 0a. Q Q (b) (−a)b = a(−b) = −(ab) for all a, b ∈ R. P P ab + ((−a)b) = (a + (−a))b = 0b = 0 = a0 = a(b + (−b)) = ab + a(−b); subtracting ab from each term, we get (−a)b = −(ab) = a(−b). Q Q 3A2C Subrings If R is a ring, a subring of R is a set S ⊆ R such that 0 ∈ S and a + b, ab, −a belong to S for all a, b ∈ S. In this case S, together with the addition and multiplication induced by those of R, is a ring in its own right. 3A2D Homomorphisms (a) Let R, S be two rings. A function φ : R → S is a ring homomorphism if φ(a + b) = φ(a) + φ(b) and φ(ab) = φ(a)φ(b) for all a, b ∈ R. The kernel of φ is {a : a ∈ R, φ(a) = 0S }. (b) Note that if φ : R → S is a ring homomorphism, then it is also a group homomorphism from (R, +) to (S, +), so that φ(0R ) = 0S and φ(−a) = −φ(a) for every a ∈ R; moreover, φ[R] is a subring of S, and φ is injective iff its kernel is {0R }. (c) If R, S and T are rings, and φ : R → S, ψ : S → T are ring homomorphisms, then ψφ : R → T is a ring homomorphism, because (ψφ)(a ∗ b) = ψ(φ(a ∗ b)) = ψ(φ(a) ∗ φ(b)) = ψ(φ(a)) ∗ ψ(φ(b)) for all a, b ∈ R, taking ∗ to be either addition or multiplication. If φ is bijective, then φ−1 : S → R is a ring homomorphism, because φ−1 (c ∗ d) = φ−1 (φ(φ−1 (c)) ∗ φ(φ−1 (d))) = φ−1 φ(φ−1 (c) ∗ φ−1 (d)) = φ−1 (c) ∗ φ−1 (d) for all c, d ∈ S, again taking ∗ to be either addition or multiplication. 3A2E Ideals (a) Let R be a ring. An ideal of R is a subring I of R such that ab ∈ I and ba ∈ I whenever a ∈ I and b ∈ R. this case we write I C R. Note that R and {0} are always ideals of R. (b) If R and S are rings and φ : R → S is a ring homomorphism, then the kernel I of φ is an ideal of R. P P (i) Because φ is a group homomorphism, I is a subgroup of (R, +). (ii) If a ∈ I, b ∈ R then φ(ab) = φ(a)φ(b) = 0S φ(b) = 0S ,

φ(ba) = φ(b)φ(a) = φ(b)0S = 0S

so ab, ba ∈ I. Q Q 3A2F Quotient rings (a) Let R be a ring and I an ideal of R. A coset of I is a set of the form a + I = {a + x : x ∈ I} where a ∈ R. (Because + is commutative, we do not need to distinguish between ‘left cosets’ a + I and ‘right cosets’ I + a.) Let R/I be the set of cosets of I in R. (b) For A, B ∈ R/I, set A + B = {x + y : x ∈ A, y ∈ B},

A · B = {xy + z : x ∈ A, y ∈ B, z ∈ I}.

Then A + B, A · B both belong to R/I; moreover, if A = a + I and B = b + I, then A + B = (a + b) + I and A · B = ab + I. P P (i)

3A2G

Rings

617

A + B = (a + I) + (b + I) = {(a + x) + (b + y) : x, y ∈ I} = {(a + b) + (x + y) : x, y ∈ I} (because addition is associative and commutative) ⊆ {(a + b) + z : z ∈ I} = (a + b) + I (because I + I ⊆ I) = {(a + b) + (z + 0) : z ∈ I} ⊆ (a + I) + (b + I) = A + B because 0 ∈ I. (ii)

A · B = {(a + x)(b + y) + z : x, y, z ∈ I} = {ab + (ay + xb + z) : x, y, z ∈ I} ⊆ {ab + w : w ∈ I} = ab + I (because ay, xb ∈ I for all x, y ∈ I, and I is closed under addition) = {(a + 0)(b + 0) + w : w ∈ I} ⊆ A · B. Q Q

(c) It is now an elementary exercise to check that (R/I, +, ·) is a ring, with zero 0 + I = I and additive inverses −(a + I) = (−a) + I. (d) Moreover, the map a 7→ a + I : R → R/I is a ring homomorphism. (e) Note that for a, b ∈ R, the following are equiveridical: (i) a ∈ b+I; (ii) b ∈ a+I; (iii) (a+I)∩(b+I) 6= ∅; (iv) a+I = b+I; (v) a−b ∈ I. Thus the cosets of I are just the equivalence classes in R under the equivalence relation a ∼ b ⇐⇒ a + I = b + I; accordingly I shall generally write a• for a + I, if there seems no room for confusion. In particular, the kernel of the canonical map from R to R/I is just {a : a + I = I} = I = 0• . (f ) If R is commutative so is R/I, since a• b• = (ab)• = (ba)• = b• a• for all a, b ∈ R. 3A2G Factoring homomorphisms through quotient rings: Proposition Let R and S be rings, I an ideal of R, and φ : R → S a homomorphism such that I is included in the kernel of φ. Then we have a ring homomorphism π : R/I → S such that π(a• ) = φ(a) for every a ∈ R. π is injective iff I is precisely the kernel of φ. proof If a, b ∈ R and a• = b• in R/I, then a−b ∈ I (3A2Fe), so φ(a)−φ(b) = φ(a−b) = 0, and φ(a) = φ(b). This means that the formula offered does indeed define a function π from R/I to S. Now if a, b ∈ R and ∗ is either multiplication or addition, π(a• ∗ b• ) = π((a ∗ b)• ) = φ(a ∗ b) = φ(a) ∗ φ(b) = π(a• ) ∗ π(b• ). So π is a ring homomorphism. The kernel of π is {a• : φ(a) = 0}, which is {0} iff φ(a) = 0 ⇐⇒ a• = 0 ⇐⇒ a ∈ I.

618

Appendix

3A2H

3A2H Product rings (a) Let hRi ii∈I be any family of rings. Set R = a + b, ab ∈ R by setting (a + b)(i) = a(i) + b(i),

Q i∈I

Ri and for a, b ∈ R define

(ab)(i) = a(i)b(i)

for every i ∈ I. It is easy to check from the definition in 3A2A that R is a ring; its zero is given by the formula 0R (i) = 0Ri for every i ∈ I, and its additive inverses by the formula (−a)(i) = −a(i) for every i ∈ I. (b) Now let S be any other ring. Then it is easy to see that a function φ : S → R is a ring homomorphism iff s 7→ φ(s)(i) : S → Ri is a ring homomorphism for every i ∈ I. (c) Note that R is commutative iff Ri is commutative for every i.

3A3 General topology In §2A3, I looked at a selection of topics in general topology in some detail, giving proofs; the point was that an ordinary elementary course in the subject would surely go far beyond what we needed there, and at the same time might omit some of the results I wished to quote. It seemed therefore worth taking a bit of space to cover the requisite material, giving readers the option of delaying a proper study of the subject until a convenient opportunity arose. In the context of the present volume, this approach is probably no longer appropriate, since we need a much greater proportion of the fundamental ideas, and by the time you have reached familiarity with the topics here you will be well able to find your way about one of the many excellent textbooks on the subject. This time round, therefore, I give most of the results without proofs (as in §§2A1 and 3A1), hoping that some of the references I offer will be accessible in all senses. I do, however, give a full set of definitions, partly to avoid ambiguity (since even in this relatively mature subject, there are some awkward divergences remaining in the usage of different authors), and partly because many of the proofs are easy enough for even a novice to fill in with a bit of thought, once the meaning of the words is clear. In fact this happens so often that I will mark with a ∗ those points where a proof needs an idea not implicit in the preceding work. 3A3A Taxonomy of topological spaces I begin with the handful of definitions we need in order to classify the different types of topological space used in this volume. A couple have already been introduced in Volume 2, but I repeat them because the list would look so odd without them. Definitions Let (X, T) be a topological space. (a) X is Hausdorff or T2 if for any distinct points x, y ∈ X there are dist open sets G, H ⊆ X such that x ∈ G and y ∈ H. (b) X is regular if whenever F ⊆ X is closed and x ∈ X \ F there are dist open sets G, H ⊆ X such that x ∈ G and F ⊆ H. (Note that in this definition I do not require X to be Hausdorff, following James 87 but not Engelking 89, Bourbaki 66, Dugundji 66, Schubert 68 or Gaal 64.) (c) X is completely regular if whenever F ⊆ X is closed and x ∈ X \ F there is a continuous function f : X → [0, 1] such that f (x) = 1 and f (y) = 0 for every y ∈ F . (Note that many authors restrict the phrase ‘completely regular’ to Hausdorff spaces. The Tychonoff space and T3 21 space are also used for Hausdorff completely regular spaces.) (d) X is zero-dimensional if whenever G ⊆ X is an open set and x ∈ G then there is an open-and-closed set H such that x ∈ H ⊆ G.

3A3D

General topology

619

(e) X is extremally disconnected if the closure of every open set in X is open. (f ) X is compact if every open cover of X has a finite subcover. (g) X is locally compact if for every x ∈ X there is a set K ⊆ X such that x ∈ int K and K is compact (in its subspace topology, as defined in 2A3C). (h) If every subset of X is open, we call T the discrete topology on X. 3A3B Elementary relationships (a) A completely regular space is regular. (Engelking 89, p. 39; Dugundji 66, p. 154; Schubert 68, p. 104.) (b) A locally compact Hausdorff space is completely regular, therefore regular. ∗ (Engelking 89, 3.3.1; Dugundji 66, p. 238; Gaal 64, p. 149.) (c) A compact Hausdorff space is locally compact, therefore completely regular and regular. (d) A regular extremally disconnected space is zero-dimensional. (Engelking 89, 6.2.25.) (e) Any topology defined by pseudometrics (2A3F), in particular the weak topology of a normed space (2A5I), is completely regular, therefore regular. (Bourbaki 66, IX.1.5; Dugundji 66, p. 200.) (f ) If X is a completely regular Hausdorff space (in particular, if X is (locally) compact and Hausdorff), and x, y are distinct points in X, then there is a continuous function f : X → R such that f (x) 6= f (y). (Apply 3A3Ac with F = {y}, which is closed because X is Hausdorff.) (g) An open set in a locally compact Hausdorff space is locally compact in its subspace topology. (Engelking 89, 3.3.8; Bourbaki 66, I.9.7.) 3A3C Continuous functions Let (X, T) and (Y, S) be topological spaces. (a) If f : X → Y is a function and x ∈ X, we say that f is continuous at x if x ∈ int f −1 [H] whenever H ⊆ Y is an open set containing f (x). (b) Now a function f : X → Y is continuous iff it is continuous at every point of X. (Bourbaki 66, I.2.1; Dugundji 66, p. 80; Schubert 68, p. 24; Gaal 64, p. 183; James 87, p. 26.) (c) If f : X → Y is continuous at x ∈ X, and A ⊆ X is such that x ∈ A, then f (x) ∈ f [A]. (Bourbaki 66, I.2.1; Schubert 68, p. 23.) (d) If f : X → Y is continuous, then f [A] ⊆ f [A] for every A ⊆ X. (Engelking 89, 1.4.1; Bourbaki 66, I.2.1; Dugundji 66, p. 80; Schubert 68, p. 24; Gaal 64, p. 184; James 87, p. 27.) (e) A function f : X → Y is a homeomorphism if it is a continuous bijection and its inverse is also continuous; that is, if S = {f [G] : G ∈ T} and T = {f −1 [H] : H ∈ S}. (f ) A function f : X → [−∞, ∞] is lower semi-continuous if {x : x ∈ X, f (x) > α} is open for every α ∈ R. (Cf. 225H.) 3A3D Compact spaces Any extended series of applications of general topology is likely to involve some new features of compactness. I start with the easy bits, continuing from 2A3Nb. (a) The first is just a definition of compactness T in of closed sets instead of open sets. A family F of sets has the finite intersection property if F0 is non-empty for every finite F0 ⊆ F. Now a topological T space X is compact iff F = 6 ∅ whenever F is a family of closed subsets of X with the finite intersection property. (Engelking 89, 3.1.1; Bourbaki 66, I.9.1; Dugundji 66, p., 223; Schubert 68, p. 68; Gaal 64, p. 127.)

620

Appendix

3A3Db

(b) A marginal generalization of this is the following. Let X be a topological space and T F a family of closed subsets of X with the finite intersection property. If F contains a compact set then F 6= ∅. (Apply (a) to {K ∩ F : F ∈ F} where K ∈ F is compact.) (c) In a Hausdorff space, compact subsets are closed. (Engelking 89, 3.1.8; Bourbaki 66, I.9.4; Dugundji 66, p. 226; Schubert 68, p. 70; Gaal 64, p. 138; James 87, p. 77.) (d) If X is compact, Y is Hausdorff and φ : X → Y is continuous and injective, then φ is a homeomorphism between X and φ[X] (where φ[X] is given the subspace topology). (Engelking 89, 3.1.12; Bourbaki 66, I.9.4; Dugundji 66, p. 226; Schubert 68, p. 71; Gaal 64, p. 207.) (e) Let X be a regular topological space and A a subset of X. Then the following are equiveridical: (i) A is relatively compact in X (that is, A is included in some compact subset of X, as in 2A3Na); (ii) A is compact; (iii) every ultrafilter on X which contains A has a limit in X. P P(ii)⇒(i) is trivial, and (i)⇒(iii) is a consequence of 2A3R; neither of these requires X to be regular. Now assume (iii) and let F be an ultrafilter on X containing A. Set H = {B : B ⊆ X, there is an open set G ∈ F such that A ∩ G ⊆ B}. Then H does not contain ∅ and B1 ∩ B2 ∈ H whenever B1 , B2 ∈ H, so H is a filter on X, and it contains A. Let H∗ ⊇ H be an ultrafilter (2A1O). By hypothesis, H∗ has a limit x say. Because A ∈ H∗ , X \ A is an open set not belonging to H∗ , and cannot be a neighbourhood of x; thus x must belong to A. Let G be an open set containing x. Then there is an open set H such that x ∈ H ⊆ H ⊆ G (this is where I use the hypothesis that X is regular). Because H∗ → x, H ∈ H∗ so X \ H does not belong to H∗ and therefore does not belong to H. But X \ H is open, so by the definition of H it cannot belong to F. As F is an ultrafilter, H ∈ F and G ∈ F. As G is arbitrary, F → x. As F is arbitrary, A is compact (2A3R). Thus (iii)⇒(ii). Q Q 3A3E Dense sets Recall that a set D in a topological space X is dense if D = X, and that X is separable if it has a countable dense subset (2A3Ud). (a) If X is a topological space, D ⊆ X is dense and G ⊆ X is dense and open, then G ∩ D is dense. (Engelking 89, 1.3.6.) Consequently the intersection of finitely many dense open sets is always dense. (b) If X and Y are topological spaces, D ⊆ A ⊆ X, D is dense in A and f : X → Y is a continuous function, then f [D] is dense in f [A]. (Use 3A3Cd.) 3A3F Meager sets Let X be a topological space. (a) A set A ⊆ X is nowhere dense or rare if int A = ∅, that is, int(X \ A) = X \ A is dense, that is, for every non-empty open set G there is a non-empty open set H ⊆ G \ A. (b) A set M ⊆ X is meager, or of first category, if it is expressible as the union of a sequence of nowhere dense sets. (c) Any subset of a nowhere dense set is nowhere dense; the union of finitely many nowhere dense sets is nowhere dense. (3A3Ea.) (d) Any subset of a meager set is meager; the union of countably many meager sets is meager. (314L.) 3A3G Baire’s theorem for locally compact Hausdorff spaces T Let X be a locally compact Hausdorff space and hGn in∈N a sequence of dense open subsets of X. Then n∈N Gn is dense. ∗ (Engelking 89, 3.9.4; Bourbaki 66, IX.5.3; Dugundji 66, p. 249; Schubert 68, p. 148.) 3A3H Corollary (a) Let X be a compact Hausdorff space. Then a non-empty open subset of X cannot be meager. (Dugundji 66, p. 250; Schubert 68, p. 147.) (b) Let X be a non-empty locally compact Hausdorff space. If hAn in∈N is a sequence of sets covering X, then there is some n ∈ N such that int An is non-empty. (Dugundji 66, p. 250.)

3A3L

General topology

621

Q 3A3I Product spaces (a) Definition Let hXi ii∈I be a family of topological spaces, and X = i∈I Xi their Cartesian product. We say that a set G ⊆ X is open for the product topology if for every x ∈ G there are a finite J ⊆ I and a family hGj ij∈J such that every Gj is an open set in the corresponding Xj and {y : y ∈ X, y(j) ∈ Gj for every j ∈ J} contains x and is included in G. (Of course we must check that this does indeed define a topology; see Engelking 89, 2.3.1; Bourbaki 66, I.4.1; Schubert 68, p. 38; Gaal 64, p. 144.) (b) If hXi ii∈I is a family of topological spaces, with product X, and Y another topological space, a function φ : Y → X is continuous iff πi φ is continuous for every i ∈ I, where πi (x) = x(i) for x ∈ X, i ∈ I. (Engelking 89, 2.3.6; Bourbaki 66, I.4.1; Dugundji 66, p. 101; Schubert 68, p. 62; James 87, p. 31.) (c) Let hXi ii∈I be any family of non-empty topological spaces, with product X. If F is a filter on X and x ∈ X, then F → x iff πi [[F]] → x(i) for every i, where πi (y) = y(i) for y ∈ X, and πi [[F]] is the image filter on Xi (2A1Ib). (Bourbaki 66, I.7.6; Schubert 68, p. 61; James 87, p. 32. (d) The product of any family of Hausdorff spaces is Hausdorff. (Engelking 89, 2.3.11; Bourbaki 66, I.8.2; Schubert 68, p. 62; James 87, p. 87.) (e) Let hX Qi ii∈I be any family of topological spaces. If Di is a dense subset of Xi for each i, then is dense in i∈I Xi . (Engelking 89, 2.3.5.). (f ) Let hX Qi ii∈I be any family of topological spaces. If Fi is a closed subset of Xi for each i, then is closed in i∈I Xi . (Engelking 89, 2.3.4; Bourbaki 66I.4.3.)

Q i∈I

Q i∈I

Di Fi

(g) Let h(Xi , Ti )ii∈I be a family of topological spaces with product (X, T). Suppose that each Ti is defined by a family Pi of pseudometrics on Xi (2A3F). Then T is defined by the family P = {˜ ρi : i ∈ I, ρ ∈ Pi } of pseudometrics on X, where I write ρ˜i (x, y) = ρ(πi (x), πi (y)) whenever i ∈ I, ρ ∈ Pi and x, y ∈ X, taking πi to be the coordinate map from X to Xi , as in (b)-(c). P P (Compare 2A3Tb). (i) It is easy to check that every ρ˜i is a pseudometric on X. Write TP for the topology generated by P. (ii) If x ∈ G ∈ TP , let P0 ⊆ P and δ > 0 be such that P0 is finite and {y : τ (y, x) ≤ δ for every τ ∈ P0 } is included in G. Express P0 as {˜ ρj : j ∈ J, ρ ∈ P0j } where J ⊆ I is finite and P0j ⊆ Pj is finite for each j ∈ J. Set Gj = {t : t ∈ Xj , ρ(t, πj (x)) < δ for every ρ ∈ P0j } for every j ∈ J. Then G0 = {y : πj (y) ∈ Gj for every j ∈ J} contains x, belongs to T and is included in G. As x is arbitrary, G ∈ T; as G is arbitrary, TP ⊆ T. (iii) Now every πi is (TP , Ti )-continuous, by 2A3H; by (b) above, the identity map from X to itself is (TP , T)-continuous, that is, TP ⊆ T and TP = T, as claimed. Q Q 3A3J Tychonoff ’s theorem The product of any family of compact topological spaces is compact. proof Engelking 89, 3.2.4; Bourbaki 66, I.9.5; Dugundji 66, p. 224; Schubert 68, p. 72; Gaal 64, p. 146 and p. 272; James 87, p. 67. Q 3A3K The spaces {0, 1}I , RI For any set I, we can think of {0, 1}I as the product i∈I Xi where Xi = {0, 1} for each i. If we endow each Xi with its discrete topology, the product topology is the usual topology on {0, 1}I . Being a product of Hausdorff spaces, it is Hausdorff; by Tychonoff’s theorem, it is compact. A subset G of {0, 1}I is open iff for every x ∈ G there is a finite J ⊆ I such that {y : y ∈ {0, 1}I , y¹J = x¹J} ⊆ G. Similarly, the ‘usual topology’ of RI is the product topology when each factor is given its Euclidean topology (cf. 2A3Tc). 3A3L Cluster points of filters (a) Let X be a topological space and F a filter on X. A point x of X is a cluster point (or adherence point or accumulation point) of F if x ∈ A for every A ∈ F.

622

Appendix

3A3Lb

(b) For any topological space X, filter F on X and x ∈ X, x is a cluster point of F iff there is a filter G ⊇ F such that G → x. (Engelking 89, 1.6.8; Bourbaki 66, I.7.2; Gaal 64, p. 260; James 87, p. 22.) (c) If hαn in∈N is a sequence in R, α ∈ R and limn→H αn = α for every non-principal ultrafilter H on N (definition: 2A3Sb), then limn→∞ αn = α. P P?? If hαn in∈N 6→ α, there is some ² > 0 such that I = {n : |αn − α| ≥ ²} is infinite. Now F0 = {F : F ⊆ N, I \ F is finite} is a filter on N, so there is an ultrafilter F ⊇ F0 . But now α cannot be limn→F αn . X XQ Q 3A3M Topology bases (a) If X is a set and A is any family of subsets of X, the topology generated by A is the smallest topology on X including A. (b) If X is a set and T is a topology on X, a base for T is a set S U ⊆ T such that whenever x ∈ G ∈ T there is a U ∈ U such that x ∈ U ⊆ G; that is, such that T = { G : G ⊆ U}. In this case, of course, U generates T. (c) If X is a set and E is a family of subsets of X, then E is a base for a topology S on X iff (i) whenever E1 , E2 ∈ E and x ∈ E1 ∩ E2 then there is an E ∈ E such that x ∈ E ⊆ E1 ∩ E2 (ii) E = X. (Engelking 89, p. 12.) 3A3N Uniform convergence (a) Let X be a set, (Y, ρ) a metric space and hfn in∈N a sequence of functions from X to Y . We say that hfn in∈N converges uniformly to a function f : X → Y if for every ² > 0 there is an n0 ∈ N such that ρ(fn (x), f (x)) ≤ ² whenever n ≥ n0 and x ∈ X. (b) Let X be a topological space and (Y, ρ) a metric space. Suppose that hfn in∈N is a sequence of continuous functions from X to Y converging uniformly to f : X → Y . Then f is continuous. ∗ (Engelking 89, 1.4.7/4.2.19; Gaal 64, p. 202.) 3A3O One-point compactifications Let (X, T) be a locally compact Hausdorff space. Take any object x∞ not belonging to X and set X ∗ = X ∪ {x∞ }. Let T∗ be the family of those sets H ⊆ X ∗ such that H ∩ X ∈ T and either x∞ ∈ / H or X \ H is compact (for T). Then T∗ is the unique compact Hausdorff ∗ topology on X inducing T as the subspace topology on X; (X ∗ , T∗ ) is the one-point compactification or Alexandroff compactification of (X, T). (Engelking 89, 3.5.11; Bourbaki 66, I.9.8; Dugundji 66, p. 246.) 3A3P Miscellaneous definitions Let X be a topological space. (a) A subset F of X is a zero set or functionally closed if it is of the form f −1 [{0}] for some continuous function f : X → R. A subset G of X is a cozero set or functionally open if its complement is a zero set. (b) An isolated point of X is a point x ∈ X such that the singleton set {x} is open.

3A4 Uniformities I continue the work of §3A3 with some notes on uniformities, so as to be able to discuss completeness and the extension of uniformly continuous functions in non-metrizable contexts (3A4F-3A4H). As in §3A3, most of the individual steps are elementary; I mark exceptions with a ∗. 3A4A Uniformities (a) Let X be any set. A uniformity on X is a filter W on X × X such that (i) (x, x) ∈ W for every x ∈ X, W ∈ W; (ii) for every W ∈ W, W −1 = {(y, x) : (x, y) ∈ W } ∈ W; (iii) for every W ∈ W, there is a V ∈ W such that V ◦ V = {(x, z) : ∃ y, (x, y) ∈ V & (y, z) ∈ V } ⊆ W . It is convenient to allow the special case X = ∅, W = {∅}, even though this is not properly speaking a filter. The pair (X, W) is now a uniform space.

3A4E

Uniformities

623

(b) If W is a uniformity on a set X, the associated topology T is the set of sets G ⊆ X such that for every x ∈ G there is a W ∈ W such that W [{x}] ={y : (x, y) ∈ W } ⊆ G. (Engelking 89, 8.1.1; Bourbaki 66, II.1.2; Gaal 64, p. 48; Schubert 68, p. 115; James 87, p. 101.) (c) We say that a uniformity is Hausdorff if the associated topology is Hausdorff. (d) If U is a linear topological space, then it has an associated uniformity W = {W : W ⊆ U × U, there is an open set G containing 0 such that (u, v) ∈ W whenever u − v ∈ G}. (Schaefer 66, I.1.4.) 3A4B Uniformities and pseudometrics (a) If P is a family of pseudometrics on a set X, then the associated uniformity is the smallest uniformity on X containing all the sets W (ρ; ²) = {(x, y) : ρ(x, y) < ²} as ρ runs through P, ² through ]0, ∞[. (Engelking 89, 8.1.18; Bourbaki 66, IX.1.2.) (b) If W is the uniformity defined by a family P of pseudometrics, then the topology associated with W is the topology defined from P (2A3F). (Dugundji 66, p. 203.) (c) A uniformity W is metrizable if it can be defined by a single metric. (d) If U is a linear space with a topology defined from a family of functionals τ : U → [0, ∞[ such that τ (u + v) ≤ τ (u) + τ (v), τ (αu) ≤ τ (u) when |α| ≤ 1, and limα→0 τ (αu) = 0 (2A5B), the uniformity defined from the topology (3A4Ad) coincides with the uniformity defined from the pseudometrics ρτ (u, v) = τ (u−v). (Immediate from the definitions.) 3A4C Uniform continuity (a) If (X, W) and (Y, V) are uniform spaces, a function φ : X → Y is uniformly continuous if {(x, y) : (φ(x), φ(y)) ∈ V } ∈ W for every V ∈ V. (b) The composition of uniformly continuous functions is uniformly continuous. (Bourbaki 66, II.2.1; Schubert 68, p. 118.) (c) If uniformities W, V on sets X, Y are defined by non-empty families P, Θ of pseudometrics, then a function φ : X → Y is uniformly continuous iff for every θ ∈ Θ, ² > 0 there are ρ0 , . . . , ρn ∈ P and δ > 0 such that θ(φ(x), φ(x0 )) ≤ ² whenever x, x0 ∈ X and maxi≤n ρi (x, x0 ) ≤ δ. (Elementary verification.) (d) A uniformly continuous function is continuous for the associated topologies. (Bourbaki 66, II.2.1; Schubert 68, p. 118; James 87, p. 102.) 3A4D Subspaces (a) If (X, W) is a uniform space and Y is any subset of X, then WY = {W ∩(Y ×Y ) : W ∈ W} is a uniformity on Y ; it is the subspace uniformity. (Bourbaki 66, II.2.4; Schubert 68, p. 122.) (b) If W defines a topology T on X, then the topology defined by WY is the subspace topology on Y , as defined in 2A3C. (Schubert 68, p. 122; James 87, p. 103.) (c) If W is defined by a family P of pseudometrics on X, then WY is defined by {ρ¹Y × Y : ρ ∈ P}. (Elementary verification.) 3A4E Product uniformities (a) If (X, U) and (Y, V) are uniform spaces, the product uniformity is the smallest uniformity W on X × Y containing all sets of the form {((x, y), (x0 , y 0 )) : (x, x0 ) ∈ U, (y, y 0 ) ∈ V } as U runs through U and V through V. (Engelking 89, §8.2; Bourbaki 66, II.2.6; Schubert 68, p. 124; James 87, p. 93.)

624

Appendix

3A4Eb

(b) If U , V are defined from families P, Θ of pseudometrics, then W will be defined by the family {˜ ρ : ρ ∈ P} ∪ {θ¯ : θ ∈ Θ}, writing ¯ ρ˜((x, y), (x0 , y 0 )) = ρ(x, x0 ), θ((x, y), (x0 , y 0 )) = θ(y, y 0 ) as in 2A3Tb. (Elementary verification.) (c) If (X, U), (Y, V) and (Z, W) are uniform spaces, a map φ : Z → X × Y is uniformly continuous iff the coordinate maps φ1 : Z → X and φ2 : Z → Y are uniformly continuous. (Engelking 89, 8.2.1; Bourbaki 66, II.2.6; Schubert 68, p. 125; James 87, p. 93.) 3A4F Completeness (a) If W is a uniformity on a set X, a filter F on X is Cauchy if for every W ∈ W there is an F ∈ F such that F × F ⊆ W . Any convergent filter in a uniform space is Cauchy. (Bourbaki 66, II.3.1; Gaal 64, p. 276; Schubert 68, p. 134; James 87, p. 109.) (b) A uniform space is complete if every Cauchy filter is convergent. (c) If W is defined from a family P of pseudometrics, then a filter F on X is Cauchy iff for every ρ ∈ P and ² > 0 there is an F ∈ F such that ρ(x, y) ≤ ² for all x, y ∈ F ; equivalently, for every ρ ∈ P, ² > 0 there is an x ∈ X such that U (x; ρ; ²) ∈ F. (Elementary verification.) (d) A complete subspace of a Hausdorff uniform space is closed. (Engelking 89, 8.3.6; Bourbaki 66, II.3.4; Schubert 68, p. 135; James 87, p. 148.) (e) A metric space is complete iff every Cauchy sequence converges (cf. 2A4Db). (Schubert 68, p. 141; Gaal 64, p. 276; James 87, p. 150.) (f ) If (X, ρ) is a complete metric space, D ⊆ X a dense subset, (Y, σ) a metric space and f : X → Y is an isometry (that is, σ(f (x), f (x0 )) = ρ(x, x0 ) for all x, x0 ∈ X), then f [X] is precisely the closure of f [D] in Y . (For it must be complete, and and we use (d).) (g) If U is a linear space with a linear space topology and the associated uniformity (3A4Ad), then a filter F on U is Cauchy iff for every open set G containing 0 there is an F ∈ F such that F − F ⊆ G (cf. 2A5F). (Immediate from the definitions.) 3A4G Extension of uniformly continuous functions: Theorem If (X, W) is a uniform space, (Y, V) is a complete uniform space, D ⊆ X is a dense subset of X, and φ : D → Y is uniformly continuous (for the subspace uniformity of D), then there is a uniformly continuous φˆ : X → Y extending φ. If Y is Hausdorff, the extension is unique. ∗ (Engelking 89, 8.3.10; Bourbaki 66, II.3.6; Gaal 64, p. 300; Schubert 68, p. 137; James 87, p. 152.) In particular, if (X, ρ) is a metric space, (Y, σ) is a complete metric space, D ⊆ X is a dense subset, and φ : D → Y is an isometry, then there is a unique isometry φˆ : X → Y extending φ. 3A4H Completions (a) Theorem If (X, W) is any Hausdorff uniform space, then we can find a ˆ W) ˆ in which X is embedded as a dense subspace; moreover, any two complete Hausdorff uniform space (X, such spaces are essentially unique. ∗ (Engelking 89, 8.3.12; Bourbaki 66, II.3.7; Gaal 64, p. 297 & p. 300; Schubert 68, p. 139; James 87, p. 156.) ˆ W) ˆ is called a completion of (X, W). Because it is unique up to isomorphism as (b) Such a space (X, a uniform space, we may call it ‘the’ completion. (c) If W is the uniformity defined by a metric ρ on a set X, then there is a unique extension of ρ to a ˆ defining the uniformity W. ˆ (Bourbaki 66, IX.1.3.) metric ρˆ on X 3A4I A note on metric spaces I mention some elementary facts which are used in §386. Let (X, ρ) be a metric space. If x ∈ X and A ⊆ X is non-empty, set

3A5C

Normed spaces

625

ρ(x, A) = inf y∈A ρ(x, y). Then ρ(x, A) = 0 iff x ∈ A (2A3Kb). If B ⊆ X is another non-empty set, then ρ(x, B) ≤ ρ(x, A) + supy∈A ρ(y, B). In particular, ρ(x, A) = ρ(x, A). If hAn in∈N is a non-decreasing sequence of non-empty sets with union A, then ρ(x, A) = limn→∞ ρ(x, An ).

3A5 Normed spaces I run as quickly as possible over the results, nearly all of them standard elements of any introductory course in functional analysis, which I find myself calling on in this volume. As in the corresponding section of Volume 2 (§2A4), a large proportion of these are valid for both real and complex normed spaces, but as the present volume is almost exclusively concerned with real linear spaces I leave this unsaid, except in 3A5L, and if in doubt you should suppose that scalars belong to the field R. 3A5A The Hahn-Banach theorem: analytic forms This is one of the central ideas of functional analysis, both finite- and infinite-dimensional, and appears in a remarkable variety of forms. I list those formulations which I wish to quote, starting with those which are more or less ‘analytic’, according to the classification of Bourbaki 87. Recall that if U is a normed space I write U ∗ for the Banach space of bounded linear functionals on U . (a) Let U be a linear space and p : U → [0, ∞[ a functional such that p(u + v) ≤ p(u) + p(v) and p(αu) = αp(u) whenever u, v ∈ U and α ≥ 0. Then for any u0 ∈ U there is a linear functional f : U → R such that f (u0 ) = p(u0 ) and f (u) ≤ p(u) for every u ∈ U . Rudin 91, 3.2; Dunford & Schwartz 57, II.3.10. (b) Let U be a normed space and V a linear subspace of U . Then for any f ∈ V ∗ there is a g ∈ U ∗ , extending f , with kgk = kf k. (363R; Bourbaki 87, II.3.2; Rudin 91, 3.3; Dunford & Schwartz 57, II.3.11; Lang 93, p. 69; Wilansky 64, p. 66; Taylor 64, 3.7-B & 4.3-A.) (c) If U is a normed space and u ∈ U there is an f ∈ U ∗ such that kf k ≤ 1 and f (u) = kuk. (Bourbaki 87, II.3.2; Rudin 91, 3.3; Dunford & Schwartz 57, II.3.14; Wilansky 64, p. 67; Taylor 64, 3.7-C & 4.3-B.) (d) If U is a normed space and V ⊆ U is a linear subspace which is not dense, then there is a non-zero f ∈ U ∗ such that f (v) = 0 for every v ∈ V . (Rudin 91, 3.5; Dunford & Schwartz 57, II.3.12; Taylor 64, 4.3-D.) (e) If U is a normed space, U ∗ separates the points of U . (Rudin 91, 3.4; Lang 93, p. 70; Dunford & Schwartz 57, II.3.14.) 3A5B Cones (a) Let U be a linear space. A convex cone (with apex 0) is a set C ⊆ U such that αu + βv ∈ C whenever u, v ∈ C and α, β ≥ 0. The intersection of any family of convex cones is a convex cone, so for every subset A of U there is a smallest convex cone including A. (b) Let U be a normed space. Then the closure of a convex cone is a convex cone, and the closure of a linear subspace is a linear subspace. (Bourbaki 87, II.2.6; Dunford & Schwartz 57, V.2.1.) 3A5C Hahn-Banach theorem: geometric forms (a) Let U be a normed space and C ⊆ U a convex set such that kuk ≥ 1 for every u ∈ C. Then there is an f ∈ U ∗ such that kf k ≤ 1 and f (u) ≥ 1 for every u ∈ C. (Dunford & Schwartz 57, V.1.12.)

626

Appendix

3A5Cb

(b) Let U be a normed space and B ⊆ U a non-empty convex set such that 0 ∈ / B. Then there is an f ∈ U ∗ such that inf u∈B f (u) > 0. (Bourbaki 87, II.4.1; Rudin 91, 3.4; Lang 93, p. 70; Dunford & Schwartz 57, V.2.12.) (c) Let U be a normed space, B a closed convex subset of U containing 0, and u a point of U \ B. Then there is an f ∈ U ∗ such that f (u) > 1 and f (v) ≤ 1 for every v ∈ B. (Apply (b) to B − u to find a g ∈ U ∗ such that g(u) < inf v∈B g(v) and now set f = − α1 g where g(u) < α < inf v∈B g(v)). 3A5D Separation from finitely-generated cones Let U be a linear space over R and u, v0 , . . . , vn points of U such that u does not belong to the convex cone generated by {v0 , . . . , vn }. Then there is a linear functional f : U → R such that f (vi ) ≥ 0 for every i and f (u) < 0. proof (a) If U is finite-dimensional this is covered by Gale 60, p. 56. (b) For the general case, let V be the linear subspace of U generated by u, v0 , . . . , vn . Then there is a linear functional f0 : V → R such that f0 (u) < 0 ≤ f0 (vi ) for every i. By Zorn’s Lemma, there is a maximal linear subspace W ⊆ U such that W ∩ V = {0}. Now W + V = U (for if u ∈ / W + V , the linear subspace W 0 generated by W ∪ {u} still has trivial intersection with V ), so we have an extension of f0 to a linear functional f : U → R defined by setting f (v + w) = f0 (v) whenever v ∈ V and w ∈ W . Now f (u) < 0 ≤ mini≤n f (vi ), as required. 3A5E Weak topologies (a) Let U be any linear space over R and W a linear subspace of the space U 0 of all linear functionals from U to R. Then I write Ts (U, W ) for the linear space topology defined by the method of 2A5B from the functionals u 7→ |f (u)| as f runs through W . (Bourbaki 87, II.6.2; Rudin 91, 3.10; Dunford & Schwartz 57, V.3.2; Taylor 64, 3.81.) (b) I note that the weak topology of a normed space U (2A5Ia) is Ts (U, U ∗ ), while the weak* topology of U ∗ (2A5Ig) is Ts (U ∗ , W ) where W is the canonical image of U in U ∗∗ . (Rudin 91, 3.14.) (c) Let U and V be linear spaces over R and T : U → V a linear operator. If W ⊆ U 0 and Z ⊆ V 0 are such that gT ∈ W for every g ∈ Z, then T is continuous for Ts (U, W ) and Ts (V, Z). (Bourbaki 87, II.6.4.) (d) If U and V are normed spaces and T : U → V is a bounded linear operator then we have an adt (or conjugate, or dual) operator T 0 : V ∗ → U ∗ defined by saying that T 0 g = gT for every g ∈ V ∗ . T 0 is linear and is continuous for the weak* topologies of U ∗ and V ∗ . (Bourbaki 87, II.6.4; Dunford & Schwartz 57, §VI.2; Taylor 64, 4.5.) (e) If U is a normed space and A ⊆ U is convex, then the closure of A for the norm topology is the same as the closure of A for the weak topology of U . In particular, norm-closed convex subsets (for instance, norm-closed linear subspaces) of U are closed for the weak topology. (Rudin 91, 3.12; Lang 93, p. 88; Dunford & Schwartz 57, V.3.13.) 3A5F Weak* topologies: Theorem If U is a normed space, the unit ball of U ∗ is compact and Hausdorff for the weak* topology. (Rudin 91, 3.15; Lang 93, p. 71; Dunford & Schwartz 57, V.4.2; Taylor 64, 4.61-A.) 3A5G Reflexive spaces (a) A normed space U is reflexive if every member of U ∗∗ is of the form f 7→ f (u) for some u ∈ U . (b) A normed space is reflexive iff bounded sets are relatively weakly compact. (Dunford & Schwartz 57, V.4.8; Taylor 64, 4.61-C.) (c) If U is a reflexive space, hun in∈N is a bounded sequence in U and F is an ultrafilter on N, then limn→F un is defined in U for the weak topology. (Use (b) and the argument of 2A3Se.)

3A5Ld

Normed spaces

627

3A5H (a) Uniform Boundedness Theorem Let U be a Banach space, V a normed space, and A ⊆ B(U ; V ) a set such that {T u : T ∈ A} is bounded in V for every u ∈ U . Then A is bounded in B(U ; V ). (Rudin 91, 2.6; Dunford & Schwartz 57, II.3.21; Taylor 64, 4.4-E.) (b) Corollary If U is a normed space and A ⊆ U is such that f [A] is bounded for every f ∈ U ∗ , then A is bounded. (Wilansky 64, p. 117; Taylor 64, 4.4-AS.) Consequently any relatively weakly compact set in U is bounded. (Rudin 91, 3.18.) 3A5I Completions Let U be a normed space. (a) Recall that U has a metric ρ associated with the norm (2A4Bb), and that the topology defined by ρ is a linear space topology (2A5D, 2A5B). This topology defines a uniformity W (3A4Ad) which is also the uniformity defined by ρ (3A4Bd). The norm itself is a uniformly continuous function from U to R (because |kuk − kvk| ≤ ku − vk for all u, v ∈ U ). ˆ , W) ˆ be the uniform space completion of (U, W) (3A4H). Then addition and scalar multiplication (b) Let (U ˆ a Banach space. (Schaefer 66, I.1.5; Lang 93, p. 78.) and the norm extend uniquely to make U (c) If U and V are Banach spaces with dense linear subspaces U0 and V0 , then any norm-preserving isomorphism between U0 and V0 extends uniquely to a norm-preserving isomorphism between U and V (use 3A4G). 3A5J Normed algebras If U is a normed algebra (2A4J), its multiplication, regarded as a function from U × U to U , is continuous. (Wilansky 64, p. 259.) 3A5K Compact operators Let U and V be Banach spaces. (a) A linear operator T : U → V is compact or completely continuous if {T u : kuk ≤ 1} is relatively compact in V for the topology defined by the norm of V . (b) A linear operator T : U → V is weakly compact if {T u : kuk ≤ 1} is relatively weakly compact in V. 3A5L Hilbert spaces I mentioned the phrases ‘inner product space’, ‘Hilbert space’ briefly in 244N244O, without explanation, as I did not there rely on any of the abstract theory of these spaces. For the main result of §395 we need one of their fundamental properties, so I now skim over the definitions. R (a) An inner product space is a linear space U over R C together with an operator ( | ) : U × U → C such that

(u1 + u2 |v) = (u1 |v) + (u2 |v),

(αu|v) = α(u|v),

(u|v) = (v|u)

(the complex conjugate of (u|v)), (u|u) ≥ 0, for all u, u1 , u2 , v ∈ U and α ∈

u = 0 whenever (u|u) = 0

R C.

(b) If U is any inner product space, we have a norm on U defined by setting kuk = u ∈ U . (Taylor 64, 3.2-B.)

p

(u|u) for every

(c) A Hilbert space is an inner product space which is a Banach space under the norm of (b) above, that is, is complete in the metric defined from its norm. (d) If U is a Hilbert space, C ⊆ U is a non-empty closed convex set, and u ∈ U , then there is a unique v ∈ C such that ku − vk = inf w∈C ku − wk. (Taylor 64, 4.81-A.)

628

Appendix

§3A6 intro.

3A6 Group Theory For Chapter 38 we need four definitions and two results from elementary abstract group theory. 3A6A Definition If G is a group, I will say that an element g of G is an involution if its order is 2, that is, g 2 = e, the identity of G, but g 6= e. 3A6B Definition If G is a group, the set Aut G of automorphisms of G (that is, bijective homomorphisms from G to itself) is a group. For g ∈ G define gˆ : G → G by writing gˆ(h) = ghg −1 for every h ∈ G; then gˆ ∈ Aut G, and the map g 7→ gˆ is a homomorphism from G onto a normal subgroup J of Aut G (Rotman 84, p. 130). We call J the group of inner automorphisms of G. of (Aut G) \ J are called outer automorphisms. 3A6C Normal subgroups For any group G, the family of normal subgroups of G, ordered by ⊆, is a Dedekind complete lattice, with H ∨ K = HK and H ∧ K = H ∩ K. (Davey & Priestley 90, 2.8 & 2.19.)

Group Theory

629

References for Volume 3 Anderson I. [87] Combinatorics of Finite Sets. Oxford U.P., 1987. [332Xk, 3A1K.] Balcar B., Glówczy´ nski W. & Jech T. [98] ‘The sequential topology on complete Boolean algebras’, Fundamenta Math. 155 (1998) 59-78. [393Yd.] Becker H. & Kechris A.S. [96] The descriptive set theory of Polish group actions. Cambridge U.P., 1996 (London Math. Soc. Lecture Note Series 232). [§394 notes.] Bekkali M. & Bonnet R. [89] ‘Rigid Boolean algebras’, pp. 637-678 in Monk 89. [383L.] Bellow A. & Kölzow D. [76] Measure Theory, Oberwolfach 1975. Springer, 1976 (Lecture Notes in Mathematics 541). Billingsley P. [65] Ergodic Theory and Information. Wiley, 1965. [§372 notes.] Bollobás B. [79] Graph Theory. Springer, 1979. [332Xk, 3A1K.] Bose R.C. & Manvel B. [84] Introduction to Combinatorial Theory. Wiley, 1984. [3A1K.] Bourbaki N. [66] General Topology. Hermann/Addison-Wesley, 1968. [§393 notes, §3A3, §3A4.] Bourbaki N. [68] Theory of Sets. Hermann/Addison-Wesley, 1968. [§315 notes.] Bourbaki N. [87] Topological Vector Spaces. Springer, 1987. [§3A5.] Bukhvalov A.V. [95] ‘Optimization without compactness, and its applications’, pp. 95-112 in Huijsmans Kaashoek Luxemburg & de Pagter 95. [367U.] Burke M.R. [93] ‘Liftings for Lebesgue measure’, pp. 119-150 in Judah 93. [341Lg, 345F.] Burke M.R. [n95] ‘Consistent liftings’, privately circulated, 1995. [346Yc.] Burke M.R. [n96] ‘An example of S.Shelah on liftings for ccc algebras’, privately circulated, 1996. [341Lh.] Burnside W. [11] Theory of Groups of Finite Order. Cambridge U.P., 1911 (reprinted by Dover, 1955). [§383 notes.] Chacon R.V. & Krengel U. [64] ‘Linear modulus of a linear operator’, Proc. Amer. Math. Soc. 15 (1964) 553-559. [§371 notes.] ¨ lzow Choksi J.R. & Prasad V.S. [82] ‘Ergodic theory of homogeneous measure algebras’, pp. 367-408 of Ko & Maharam-Stone 82. [382I.] Coleman A.J. & Ribenboim P. [67] (eds.) Proceedings of the Symposium in Analysis, Queen’s University, June 1967. Queen’s University, Kingston, Ontario, 1967. Comfort W.W. & Negrepontis S. [82] Chain Conditions in Topology. Cambridge U.P., 1982. [391L, §391 notes.] Cziszár I. [67] ‘Information-type measures of difference of probability distributions and indirect observations’, Studia Scientiarum Math. Hungarica 2 (1967) 299-318. [385J.] Davey B.A. & Priestley H.A. [90] ‘Introduction to Lattices and Order’, Cambridge U.P., 1990. [3A6C.] Dow A. & Stepr¯ans J. [93] ‘The σ-linkedness of the measure algebra’, Canad. Math. Bulletin 37 (1993) 42-45. [§391 notes.] Dugundji J. [66] Topology. Allyn & Bacon, 1966. [§3A3, 3A4B.] Dunford N. [36] ‘Integration and linear operators’, Trans. Amer. Math. Soc. 40 (1936) 474-494. [§376 notes.] Dunford N. & Schwartz J.T. [57] Linear Operators I. Wiley, 1957 (reprinted 1988). [§356 notes, §371 notes, §376 notes, §3A5.] Dye H.A. [59] ‘On groups of measure preserving transformations I’, Amer. J. Math. 81 (1959) 119-159. [§387 notes.] Eigen S.J. [82] ‘The group of measure-preserving transformations of [0, 1] has no outer automorphisms’, Math. Ann. 259 (1982) 259-270. [383D.] Engelking R. [89] General Topology. Heldermann, 1989 (Sigma Series in Pure Mathematics 6). [§393 notes, §3A3, §3A4.] Enderton H.B. [77] Elements of Set Theory. Academic, 1977. [§3A1.] Erdös P. & Oxtoby J.C. [55] ‘Partitions of the plane into sets having positive measure in every non-null product set’, Trans. Amer. Math. Soc. 79 (1955) 91-102. [§325 notes.] Fathi A. [78] ‘Le groupe des transformations de [0, 1] qui préservent la mesure de Lebesgue est un groupe simple’, Israel J. Math. 29 (1978) 302-308. [§381 notes, 382I.]

630

References

Fremlin D.H. [74a] Topological Riesz Spaces and Measure Theory. Cambridge U.P., 1974. [Chap. 35 intro, 354Yb, §355 notes, §356 notes, §363 notes, §371 notes, §376 notes.] Fremlin D.H. [74b] ‘A characterization of L-spaces’, Indag. Math. 36 (1974) 270-275. [§371 notes.] Fremlin D.H. [75] ‘Inextensible Riesz spaces’, Math. Proc. Cambridge Phil. Soc. 77 (1975) 71-89. [393L.] Fremlin D.H. [84] Consequences of Martin’s Axiom. Cambridge U.P., 1984. [§316 notes.] Fremlin D.H. [89] ‘Measure algebras’, pp. 876-980 in Monk 89. [341Lg, 341Zb.] Frol´ık Z. [68] ‘Fixed points of maps of extremally disconnected spaces and complete Boolean algebras’, Bull. Acad. Polon. Sci. 16 (1968) 269-275. [§381 notes.] Gaal S.A. [64] Point Set Topology. Academic, 1964. [§3A3, §3A4.] Gaifman H. [64] ‘Concerning measures on Boolean algebras’, Pacific J. Math. 14 (1964) 61-73. [391L.] Gale D. [60] The theory of linear economic models. McGraw-Hill, 1960. [3A5D.] Gnedenko B.V. & Kolmogorov A.N. [54] Limit Distributions for Sums of Independent Random Variables. Addison-Wesley, 1954. [§342 notes.] Graf S. & Weizsäcker H.von [76] ‘On the existence of lower densities in non-complete measure spaces’, ¨ lzow 76. [341Lc.] pp. 155-158 in Bellow & Ko Hajian A. & Ito Y. [69] ‘Weakly wandering sets and invariant measures for a group of transformations’, J. of Math. and Mech. 18 (1969) 1203-1216. [395B.] Hajian A., Ito Y. & Kakutani S. [75] ‘Full groups and a theorem of Dye’, Advances in Math. 17 (1975) 48-59. [§387 notes.] Halmos P.R. [60] Naive Set Theory. Van Nostrand, 1960. [3A1D.] Huijsmans C.B., Kaashoek M.A., Luxemburg W.A.J. & de Pagter B. [95] (eds.) Operator Theory in Function Spaces and Banach Lattices. Birkhä, 1995. Ionescu Tulcea C. & Ionescu Tulcea A. [69] Topics in the Theory of Lifting. Springer, 1969. [§341 notes.] James I.M. [87] Topological and Uniform Spaces. Springer, 1987. [§3A3, §3A4.] Jech T. [78] Set Theory. Academic, 1978. [§316 notes, §3A1.] Johnson R.A. [80] ‘Strong liftings which are not Borel liftings’, Proc. Amer. Math. Soc. 80 (1980) 234-236. [345F.] Judah H. [93] (ed.) Proceedings of the Bar-Ilan Conference on Set Theory and the Reals, 1991. Amer. Math. Soc. (Israel Mathematical Conference Proceedings 6), 1993. Juhász I. [71] Cardinal Functions in Topology. North-Holland, 1975 (Math. Centre Tracts 34). [§332 notes.] Kakutani S. [41] ‘Concrete representation of abstract L-spaces and the mean ergodic theorem’, Annals of Math. 42 (1941) 523-537. [369E.] Kalton N.J., Peck N.T. & Roberts J.W. [84] ‘An F-space sampler’, Cambridge U.P., 1984. [§375 notes.] Kalton N.J. & Roberts J.W. [83] ‘Uniformly exhaustive submeasures and nearly additive set functions’, Trans. Amer. Math. Soc. 278 (1983) 803-816. [392D, §392 notes.] ¨ Kawada Y. [44] ‘Uber die Existenz der invarianten Integrale’, Jap. J. Math. 19 (1944) 81-95. [§394 notes.] Kelley J.L. [59] ‘Measures on Boolean algebras’, Pacific J. Math. 9 (1959) 1165-1177. [391L.] Kolmogorov A.N. [58] ‘New metric invariants of transitive dynamical systems and automorphisms of Lebesgue spaces’, Dokl. Akad. Nauk SSSR 119 (1958) 861-864. [384P.] Kölzow D. & Maharam-Stone D. [82] (eds.) Measure Theory Oberwolfach 1981. Springer, 1982 (Lecture Notes in Math. 945). Koppelberg S. [89] General Theory of Boolean Algebras, vol. 1 of Monk 89. [Chap. 31 intro, §332 notes.] Köthe G. [69] Topological Vector Spaces I. Springer, 1969. [§356 notes.] Kranz P. & Labuda I. [93] (eds.) Proceedings of the Orlicz Memorial Conference, 1991 , unpublished manuscript, available from first editor ( [email protected]). ¨ Krengel, U. [63] ‘Uber den Absolutbetrag stetiger linearer Operatoren und seine Anwendung auf ergodische Zerlegungen’, Math. Scand. 13 (1963) 151-187. [371Ya.] Krieger W. [76] ‘On ergodic flows and the isomorphism of factors’, Math. Ann. 223 (1976) 19-70. [§387 notes.] Krivine J.-L. [71] Introduction to Axiomatic Set Theory. D.Reidel, 1971. [§3A1.]

3A6C

Group Theory

631

Kullback S. [67] ‘A lower bound for discrimination information in of variation’, IEEE Trans. on Information Theory 13 (1967) 126-127. [385J.] Kunen K. [80] Set Theory. North-Holland, 1980. [§393 notes, §3A1.] Kwapien S. [73] ‘On the form of a linear operator on the space of all measurable functions’, Bull. Acad. Polon. Sci. 21 (1973) 951-954. [§375 notes.] Lang S. [93] Real and Functional Analysis. Springer, 1993. [§3A5.] Lindenstrauss J. & Tzafriri L. [79] Classical Banach Spaces II. Springer, 1979, reprinted in Lindenstrauss & Tzafriri 96. [§354 notes, 374Xj.] Lindenstrauss J. & Tzafriri L. [96] Classical Banach Spaces I & II. Springer, 1996. Lipschutz S. [64] Set Theory and Related Topics. McGraw-Hill, 1964 (Schaum’s Outline Series). [3A1D.] Luxemburg W.A.J. [67a] ‘Is every integral normal?’, Bull. Amer. Math. Soc. 73 (1967) 685-688. [363S.] Luxemburg W.A.J. [67b] ‘Rearrangement-invariant Banach function spaces’, pp. 83-144 in Coleman & Ribenboim 67. [§374 notes.] Luxemburg W.A.J. & Zaanen A.C. [71] Riesz Spaces I. North-Holland, 1971. [Chap. 35 intro.] Macheras N.D., Musial K. & Strauss W. [p99] ‘On products of issible liftings and densities’, J. for Analysis and its Applications 18 (1999) 651-668. [346G.] Macheras N.D. & Strauss W. [95] ‘Products of lower densities’, J. for Analysis and its Applications 14 (1995) 25-32. [346Xg.] Macheras N.D. & Strauss W. [96a] ‘On products of almost strong liftings’, J. Australian Math. Soc. (A) 60 (1996) 1-23. [346Ya.] Macheras N.D. & Strauss W. [96b] ‘The product lifting for arbitrary products of complete probability spaces’, Atti Sem. Math. Fis. Univ. Modena 44 (1996) 485-496. [346H, 346Yb.] Maharam D. [42] ‘On homogeneous measure algebras’, Proc. Nat. Acad. Sci. U.S.A. 28 (1942) 108-111. [331F, 332B.] Maharam D. [47] ‘An algebraic characterization of measure algebras’, Ann. Math. 48 (1947) 154-167. [393J.] Maharam D. [58] ‘On a theorem of von Neumann’, Proc. Amer. Math. Soc. 9 (1958) 987-994. [§341 notes, §346 notes.] Marczewski E. [53] ‘On compact measures’, Fund. Math. 40 (1953) 113-124. [342Ad.] Monk J.D. [89] (ed.) Handbook of Boolean Algebra. North-Holland, 1989. Nadkarni M.G. [90] ‘On the existence of a finite invariant measure’, Proc. Indian Acad. Sci., Math. Sci. 100 (1990) 203-220. [§394 notes.] Ornstein D.S. [74] Ergodic Theory, Randomness and Dynamical Systems. Yale U.P., 1974. [§386 notes.] Ornstein D.S. & Shields P.C. [73] ‘An uncountable family of K-automorphisms’, Advances in Math. 10 (1973) 63-88. [§381 notes.] Petersen K. [83] Ergodic Theory. Cambridge U.P., 1983. [382Ye, 384Cb, §384 notes, 385G.] Roberts J.W. [93] ‘Maharam’s problem’, in Kranz & Labuda 93. [393U.] Rotman J.J. [84] ‘An Introduction to the Theory of Groups’, Allyn & Bacon, 1984. [§383 notes, 3A6B.] Rudin W. [91] Functional Analysis. McGraw-Hill, 1991. [§3A5.] Sazonov V.V. [66] ”On perfect measures”, A.M.S. Translations (2) 48 (1966) 229-254. [§342 notes.] Schaefer H.H. [66] Topological Vector Spaces. MacMillan, 1966; reprinted with corrections Springer, 1971. [3A4Ad, 3A5Ib.] Schaefer H.H. [74] Banach Lattices and Positive Operators. Springer, 1974. [Chap. 35 intro, §354 notes.] Schubert H. [68] Topology. Allyn & Bacon, 1968. [§3A3, §3A4.] Shelah S. [Sh636] ‘The lifting problem with the full ideal’, J. Applied Analysis 4 (1998) 1-17. [341Lh.] Sikorski R. [64] Boolean Algebras. Springer, 1964. [Chap. 31 intro.] Sinaˇı Ya.G. [59] ‘The notion of entropy of a dynamical system’, Dokl. Akad. Nauk SSSR 125 (1959) 768-771. [384P.] Sinaˇı Ya.G. [62] ‘Weak isomorphism of transformations with an invariant measure’, Soviet Math. 3 (1962) 1725-1729. [386E.]

632

References

3A6C

Smorodinsky M. [71] Ergodic Theory, Entropy. Springer, 1971 (Lecture Notes in Math., 214). [§386 notes.] ˇ epánek P. & Rubin M. [89] ‘Homogeneous Boolean algebras’, pp. 679-715 in Monk 89. [381T, 381Yd, Stˇ §381 notes.] Talagrand M. [80] ‘A simple example of a pathological submeasure’, Math. Ann. 252 (1980) 97-102. [393T.] Talagrand M. [82a] ‘Closed convex hull of set of measurable functions, Riemann-measurable functions and measurability of translations’, Ann. Institut Fourier (Grenoble) 32 (1982) 39-69. [§346 notes.] Talagrand M. [82b] ‘La pathologie des relèvements invariants’, Proc. Amer. Math. Soc. 84 (1982) 379-382. [345F.] Talagrand M. [84] Pettis integral and measure theory. Mem. Amer. Math. Soc. 307 (1984). [§346 notes.] Taylor A.E. [64] Introduction to Functional Analysis. Wiley, 1964. [§3A5.] Vulikh B.C. [67] Introduction to the Theory of Partially Ordered Vector Spaces. Wolters-Noordhoff, 1967. [§364 notes.] Wagon S. [85] The Banach-Tarski Paradox. Cambridge U.P., 1985. [§394 notes.] Wilansky A. [64] Functional Analysis. Blaisdell, 1964. [§3A5.] Zaanen A.C. [83] Riesz Spaces II. North-Holland, 1983. [Chap. 35 intro., 376K, §376 notes.]

bounded

Principal topics and results

633

Index to volumes 1, 2 and 3

Principal topics and results The general index below is intended to be comprehensive. Inevitably the entries are voluminous to the point that they are often unhelpful. I am therefore preparing a shorter, better-annotated, index which will, I hope, help readers to focus on particular areas. It does not mention definitions, as the bold-type entries in the main index are supposed to lead efficiently to these; and if you draw blank here you should always, of course, try again in the main index. Entries in the form of mathematical assertions frequently omit essential hypotheses and should be checked against the formal statements in the body of the work. absolutely continuous real functions §225 —– as indefinite integrals 225E absolutely continuous additive functionals §232 —– characterization 232B —– on measure algebras 327B additive functionals on Boolean algebras §326, §327, §362 —– dominating or subordinate to a given functional 391E, 391F Archimedean Riesz spaces §353 associate norm (on a dual function space) §369 atomless measure spaces —– have elements of all possible measures 215D automorphisms and automorphism groups —– of complete Boolean algebras §381, §383 —– —– automorphisms as products of involutions 381N —– —– normal subgroups of automorphism groups 381S, 382I —– —– isomorphisms of automorphism groups 383D, 383E, 383J, 383M —– of measure algebras §382, §384, §385, §386, §387 Banach lattices §354 —– order-bounded linear operators are continuous 355C —– duals 356D Banach-Ulam problem 363S band algebra in a Riesz space 352Q Bernoulli shifts §386 —– calculation of entropy 384R —– as factors of given homomorphisms (Sinaˇı’s theorem) 386E, 386L —– specified by entropy (Ornstein’s theorem) 386I, 386K biduals of Riesz spaces 356F, 356I Boolean algebras Chap. 31 Boolean homomorphisms §312 —– between measure algebras §324 —– induced by measurable functions 324A —– continuity and uniform continuity 324F —– when measure-preserving 324K —– represented by measurable functions 343B, 344B, 344E Boolean rings §311 Borel sets in Rr 111G —– and Lebesgue measure 114G, 115G, 134F bounded variation, real functions of §224 —– as differences of monotonic functions 224D —– integrals of their derivatives 224I —– Lebesgue decomposition 226C

634

Index

Cantor

Cantor set and function 134G, 134H Carathéodory’s construction of measures from outer measures 113C Carleson’s theorem (Fourier series of square-integrable functions converge a.e.) §286 Central Limit Theorem (sum of independent random variables approximately normal) §274 —– Lindeberg’s condition 274F, 274G change R of variable in R the integral §235 —– J × gφ dµ = g dν 235A, 235E, 235L —– finding J 235O; —– —– J = | det T | for linear operators T 263A; J = | det φ0 | for differentiable operators φ 263D —– —– —– when the measures are Hausdorff measures 265B, 265E —– when φ is inverse-measure-preserving 235I R R —– gφ dµ = J × g dν 235T characteristic function of a probability distribution §285 —– sequences of distributions converge in vague topology iff characteristic functions converge pointwise 285L closed subalgebra of a measure algebra 323H, 323J —– classification theorem 333H, 333K —– defined by a group of automorphisms 333R compact measure space §342, §343 —– and representation of homomorphisms between measure algebras 343B complete measure space §212 completion of a measure 212C et seq. concentration of measure 264H conditional expectation —– of a function §233 —– as operator on L1 (µ) 242J construction of measures —– image measures 112E —– from outer measures (Carathéodory’s method) 113C —– subspace measures 131A, 214A —– product measures 251C, 251F, 251W, 254C —– as pull-backs 132G continued fractions 372M et seq. Control Measure Problem §393 convergence theorems (B.Levi, Fatou, Lebesgue) §123 convergence in measure (linear space topology on L0 ) —– on L0 (µ) §245 —– on L0 (A) §367 —– when Hausdorff/complete/metrizable 245E, 367N —– and positive linear operators 367P convex functions 233G et seq. convolution of functions r —– (on R R ) §255 R —– h × (f ∗ g) = h(x + y)f (x)g(y)dxdy 255G —– f ∗ (g ∗ h) = (f ∗ g) ∗ h 255J convolution of measures r —– (on R R ) §257 R —– h d(ν1 ∗ ν2 ) = h(x + y)ν1 (dx)ν2 (dy) 257B —– of absolutely continuous measures 257F countable sets 111F, 1A1C et seq. countable-cocountable measure 211R counting measure 112Bd

fundamental


Dedekind (σ)-complete Boolean algebras §314 —– Riesz spaces §353 Dedekind completion of a Boolean algebra 314T differentiable functions (from Rr to Rs ) §262, §263 direct sum of measure spaces 214K distribution of a random variable X §271 —– as a Radon measure 271B X ) 271J —– of φ(X distributive laws in Boolean algebras 313B —– in Riesz spaces 352E Doob’s Martingale Convergence Theorem 275G duals —– of Riesz spaces §356, §362, 369C, 369J —– of Banach lattices 356D Dye’s theorem (on the full subgroup generated by a measure-preserving automorphism) 387N, 387O entropy (of a partition or a function) §384, §385 —– calculation (Kolmogorov-Sinaˇı theorem) 384P; (Bernoulli shifts) 384R Ergodic Theorem §372 —– (Pointwise Ergodic Theorem) 372D, 372E —– (Maximal Ergodic Theorem) 372C exhaustion, principle of 215A extended real line §135 extension of order-continuous positive linear operators 355F, 368B, 368M R R Fatou’s Lemma ( lim inf ≤ lim inf for sequences of non-negative functions) 123B Féjer sums (running averages of Fourier sums) converge to local averages of f 282H —– uniformly if f is continuous 282G Fourier series §282 —– norm-converge in L2 282J —– converge at points of differentiability 282L —– converge to midpoints of jumps, if f of bounded variation 282O —– and convolutions 282Q —– converge a.e. for square-integrable function 286V Fourier transforms —– on R §283, §284 Rd ∧ —– formula for c f in of f 283F —– and convolutions 283M —– in of action on test functions 284H et seq. —– of square-integrable functions 284O, 286U —– inversion formulae for differentiable functions 283I; for functions of bounded variation 283L R∞ ∧ 2 —– f (y) = lim²↓0 √12π −∞ e−iyx e−²x f (x)dx a.e. 284M free products of Boolean algebras §315 —– universal mapping theorem 315I free products of measure algebras §325 —– universal mapping theorems 325D, 325J Fubini’s R theorem §252 RR —– f d(µ × ν) = f (x, y)dxdy 252B —– when both factors σ-finite 252C, 252H —– for characteristic functions 252D, 252F function spaces §369, §374, §376 Rx Rb d f = f (x) a.e.) §222; ( a F 0 (x)dx = F (b) − F (a)) 225E Fundamental Theorem of the Calculus ( dx a

635

636

Index

Hahn

Hahn decomposition of a countably additive functional 231E —– on a Boolean algebra 326I Hahn-Banach property see Nachbin-Kelley theorem (363R) Hardy-Littlewood Maximal Theorem 286A Hausdorff measures (on Rr ) §264 —– are topological measures 264E —– r-dimensional Hausdorff measure on Rr a multiple of Lebesgue measure 264I —– (r − 1) dimensional measure on Rr 265F-265H image measures 112E indefinite integrals —– differentiate back to original function 222E, 222H —– to construct measures §234 independent random variables §272 —– t distributions are product measures 272G —– limit theorems §273, §274 inner regularity of measures —– (with respect to compact sets) Lebesgue measure 134F integral operators §376 —– represented by functions on a product space 376E —– characterized by action on weakly convergent sequences 376H —– into M -spaces or out of L-spaces 376M —– and weak compactness 376P —– disintegrated 376S integration of real-valued functions, construction §122 —– as a positive linear functional 122O —– —– acting on L1 (µ) 242B —– by parts 225F —– characterization of integrable functions 122P, 122R —– over subsets 131D, 214E —– functions and integrals with values in [−∞, ∞] §133 invariant measures on measure algebras 394P Jensen’s inequality 233I-233J —– expressed in L1 (µ) 242K Kakutani’s theorem (on the representation of L-spaces as L1 -spaces) 369E Kolmogorov-Sinaˇı theorem (on the entropy of a measure-preserving homomorphism) 384P Komlós’ subsequence theorem 276H Kwapien’s theorem (on maps between L0 spaces) 375I Lebesgue’s Density Theorem (in R) §223 R x+h —– limh↓0 h1 x f = f (x) a.e. 223A R x+h 1 —– limh↓0 2h |f (x − y)|dy = 0 a.e. 223D x−h —– (in Rr ) 261C, 261E Lebesgue measure, construction of §114, §115 —– further properties §134 —– characterized as measure space 344I R R Lebesgue’s Dominated Convergence Theorem ( lim = lim for dominated sequences of functions) 123C, 367J R R R P PR B.Levi’s theorem ( lim = lim for monotonic sequences of functions) 123A; ( = ) 226E lifting (of a measure) —– complete strictly localizable spaces have liftings 341K —– translation-invariant liftings on Rr and {0, 1}I §345 —– respecting product structures §346

order


637

Lipschitz functions §262 —– differentiable a.e. 262Q localizable measure space —– assembling partial measurable functions 213N —– which is not strictly localizable 216E Loomis-Sikorski representation of a Dedekind σ-complete Boolean algebra 314M lower density (on a measure space) —– strictly localizable spaces have lower densities 341H —– respecting a product structure 346G Lusin’s theorem (measurable functions are almost continuous) 256F Maharam’s theorem —– (homogeneous probability algebras of Maharam type κ are isomorphic) 331I —– —– to measure algebra of {0, 1}κ ) 331L —– classification of localizable measure algebras 332J Maharam type —– (calculation of) 332S —– —– for measure algebras of qproduct spaces 334A, 334C —– relative to a subalgebra 333E martingales §275 —– L1 -bounded martingales converge a.e. 275G, 367K —– when of form E(X|Σn ) 275H, 275I measurable algebra §391 —– characterized by weak (σ, ∞)-distributivity and existence of strictly positive additive functional 391D; by existence of uniformly exhaustive strictly positive Maharam submeasure 392J —– Control Measure Problem §393 measurable envelopes —– elementary properties 132E measurable functions —– (real-valued) §121 —– —– sums, products and other operations on finitely many functions 121E —– —– limits, infima, suprema 121F measure algebras —– properties related to those of measure spaces 322B —– relationships between elementary properties 322C —– satisfying the countable chain condition 322G —– topologies and uniformities on §323, §324; uniqueness of topology and uniformity 324H measure-preserving Boolean homomorphisms —– extension from a closed subalgebra 333C Monotone Class Theorem 136B monotonic functions —– are differentiable a.e. 222A non-measurable set (for Lebesgue measure) 134B non-paradoxical group (of automorphisms of a Boolean algebra) §394 —– characterizations 394D —– invariant functionals 394O, 394P, 395B order-bounded linear operators §355 order-closed ideals and order-continuous Boolean homomorphisms 313P order-continuous linear operators §355 order*-convergence of a sequence in a lattice §367 order-dense sets in Boolean algebras 313K —– in Riesz spaces 352N

638

Index

Ornstein

Ornstein’s theorem (on isomorphism of Bernoulli shifts with the same entropy) 386I, 386K outer measures constructed from measures §132 —– elementary properties 132A outer regularity of Lebesgue measure 134F partially ordered linear spaces §351 perfect measure space 342K et seq. Plancherel Theorem (on Fourier series and transforms of square-integrable functions) 282K, 284O Poisson’s theorem (a count of independent rare events has an approximately Poisson distribution) 285Q product of two measure spaces §251 —– basic properties of c.l.d. product measure 251I —– Lebesgue measure on Rr+s as a product measure 251M —– more than two spaces 251W —– Fubini’s theorem 252B —– Tonelli’s theorem 252G —– and L1 spaces §253 —– continuous bilinear maps on L1 (µ) × L1 (ν) 253F —– conditional expectation on a factor 253H —– and measure algebras 325A et seq. product of any number of probability spaces §254 —– basic properties of the (completed) product 254F —– characterization with inverse-measure-preserving functions 254G —– products of subspaces 254L —– products of products 254N —– determination by countable subproducts 254O —– subproducts and conditional expectations 254R —– and measure algebras 325I et seq. —– having homogeneous measure algebras 334E Rademacher’s theorem (Lipschitz functions are differentiable a.e.) 262Q Radon measures —– on Rr §256 —– as completions of Borel measures 256C —– indefinite-integral measures 256E —– image measures 256G —– product measures 256K Radon-Nikod´ ym theorem (truly continuous additive set-functions have densities) 232E —– in of L1 (µ) 242I —– on a measure algebra 327D, 365E rearrangements (monotonic rearrangements of measurable functions) §373, §374 regular open algebras 314P representation —– of Boolean algebras see Stone’s theorem (311E) —– of Dedekind σ-complete Boolean algebras 314M —– of measure algebras 321J —– of homomorphisms between measure algebras §343, §344 —– of partially ordered linear spaces 351Q —– of Riesz spaces 352L, 353M, 368E et seq., 369A —– of Riesz space duals 369C —– of L-spaces 369E —– of M -spaces 354L Riesz spaces (vector lattices) §352 —– identities 352D, 352F, 352M

Weyl


639

Shannon-McMillan-Breiman theorem (entropy functions of partitions generated by a measure-preserving homomorphism converge a.e.) 385G simple product of Boolean algebras —– universal mapping theorem 315B —– of measure algebras 322K Sinaˇı’s theorem (on existence of Bernoulli partitions for a measure-preserving automorphism) 386E, 386L split interval 343J Stone space of a Boolean ring or algebra 311E et seq. —– as topological space 311I —– and Boolean homomorphisms 312P, 312Q —– of a Dedekind complete algebra 314S —– and the countable chain condition 316B —– and weak (σ, ∞)-distributivity 316I —– of a measure algebra 321J, 322N, 322Q Stone-Weierstrass theorem §281 —– for Riesz subspaces of Cb (X) 281A —– for subalgebras of Cb (X) 281E —– for *-subalgebras of Cb (X; C) 281G strictly localizable measures —– sufficient condition for strict localizability Pn 213O 1 strong law of large numbers (limn→∞ n+1 i=0 (Xi − E(Xi )) = 0 a.e.) §273 P∞ 1 —– when n=0 Var(Xn ) < ∞ 273D 2 (n+1)

—– when supn∈N E(|Xn |1+δ ) < ∞ 273H —– for identically distributed Xn 273I —– for martingale difference sequences 276C, 276F —– convergence of averages for k k1 , k kp 273N submeasure §392 subspace measures —– for measurable subspaces §131 —– for arbitrary subspaces §214 surface measure in Rr §265 tensor products of L1 spaces §253, 376C RR Tonelli’s theorem (f is integrable if |f (x, y)|dxdy < ∞) 252G uniformly exhaustive submeasures —– and additive functionals 392F uniformly integrable sets in L1 §246 —– criteria for uniform integrability 246G —– in L-spaces 354P et seq. —– and convergence in measure 246J —– and weak compactness 247C, 356Q —– in dual spaces 356O, 362E Vitali’s theorem (for coverings by intervals in R) 221A —– (for coverings by balls in Rr ) 261B weak compactness in L1 (µ) §247 —– in L-spaces 356Q weakly (σ, ∞)-distributive Boolean algebras §316 —– and measure algebras 322F, 391D —– Riesz spaces 368Q Weyl’s Equidistribution Theorem 281N

640

Index

C

c.l.d. version 213D et seq. L-space 354N et seq., §371 L0 (µ) (space of equivalence classes of measurable functions) §241 —– as Riesz space 241E L0 (A) (based on a Boolean algebra) §364, §368, §369, §375 —– as a quotient of a space of functions 364D —– algebraic operations 364E —– calculation of suprema and infima 364M —– linear operators induced by Boolean homomorphisms 364R —– *when A is a regular open algebra 364U —– —– *of a compact extremally disconnected space 364W —– as the domain of a linear operator 375A, 375C —– positive linear operators between L0 spaces can be derived from Riesz homomorphisms 375I L1 (µ) (space of equivalence classes of integrable functions) §242 —– norm-completeness 242F —– density of simple functions 242M —– (for Lebesgue measure) density of continuous functions and step functions 242O L1 (A, µ ¯) (based on a measure algebra) §365 —– related to L1 (µ) 365B —– as L-space 365C, 369E —– linear operators induced by Boolean homomorphisms 365H, 365O, 365P —– duality with L∞ 365J —– universal mapping theorems 365L, 365N Lp (µ) (space of equivalence classes of pth-power-integrable functions, where 1 < p < ∞) §244 —– is a Banach lattice 244G —– has dual Lq (µ), where p1 + 1q = 1 244K —– and conditional expectations 244M Lp (A) (based on a Boolean algebra) §366 —– related to Lp (µ) 366B —– as Banach lattice 366C —– linear operators induced by Boolean homomorphisms 366H L∞ (µ) (space of equivalence classes of bounded measurable functions) §243 —– duality with L1 243F-243G —– norm-completeness 243E —– order-completeness 243H L∞ (A) (where A is a Boolean algebra) §363 —– universal mapping theorems 363E —– linear operators induced by Boolean homomorphisms 363F —– order-completeness 363M —– Nachbib-Kelley theorem (generalized Hahn-Banach theorem) 363R L∼ (U ; V ) (space of order-bounded linear operators) §355, 371B, 371C L× (U ; V ) (space of differences of order-continuous linear operators) §355, 376E M -space 354G et seq. S(A) (space of ‘simple’ functions corresponding to a Boolean ring A) §361 —– universal mapping theorems 361F-361I —– linear operators induced by ring homomorphisms 361J —– dual spaces §362 T , T (0) , T × (classes of operators which are norm-reducing for every k kp ) 371F-371G, §372, §373, §374 σ-algebras of sets §111 —– generated by given families 136B, 136G σ-finite measures 215B

atom

General index

641

{0, 1}κ —– usual measure is (Maharam) homogeneous with Maharam type κ 331K; is homogeneous in strict sense 344L

General index This index covers the ‘full’ version of the work. References in bold type refer to definitions; references in italics are ing references. Definitions marked with > are those in which my usage is dangerously at variance with that of some other author. Abel’s theorem 224Yi absolute summability 226Ac absolutely continuous additive functional 232Aa, 232B, 232D, 232F-232I, 232Xa, 232Xb, 232Xd, 232Xf, 232Xh, 234Ce, 256J, 257Xf, 327A, 327B, 327C, 362C, 362Xh, 362Xi, 363S absolutely continuous function 225B, 225C-225G, 225K-225O, 225Xa-225Xh, 225Xn, 225Xo, 225Ya, 225Yc, 232Xb, 244Yh, 252Yj, 256Xg, 262B, 263I, 264Yp, 265Ya, 282R, 283Ci absolutely continuous submeasure 393D, 393E, 393H, 393Yg Absoluteness Theorem see Shoenfield’s Absoluteness Theorem abstract integral operator 376D, 376E, 376F, 376H, 376J, 376K, 376M, 376P, 376Xc, 376Xk, 376Yh376Yj, 376Yl, 376Yn accumulation point see cluster point (3A3La) additive functional on an algebra of sets see finitely additive (136Xg, 231A), countably additive (231C) additive function(al) on a Boolean algebra see finitely additive (326A, 361B), countably additive (326E), completely additive (326J) adherence point see cluster point (3A3La) adt operator 243Fc, 243 notes, 371Xe, 371Yd, 373S-373U, 373Xu, 373Yg, 3A5Ed algebra (over R) 361Xb, 361Xf; see also algebra of sets (231A), Baire property algebra (314Yd), Banach algebra (2A4Jb), Boolean algebra (311A), category algebra (314Yd), normed algebra (2A4J), f -algebra (352W), regular open algebra (314Q) algebra of sets 113Yi, 136E, 136F, 136G, 136Xg, 136Xh, 136Xk, 136Ya, 136Yb, 231A, 231B, 231Xa, 311Bb, 311Xb, 311Xh, 312B, 315G, 315L, 362Xg, 363Yf, 381Xa; see also Boolean algebra (311A), σ-algebra (111A) Alexandroff compactification see one-point compactification (3A3O) almost continuous function 256F almost every, almost everywhere 112Dd almost isomorphic (inverse-measure-preserving functions) 384U, 384V, 384Xn-384Xp, 384Yf almost surely 112De amenable group 394 notes analytic (complex) function 133Xc antichain 316 notes aperiodic Boolean homomorphism 385A, 385C, 385E, 385F, 385Xa, 385Yb, 387Ee, 387I, 387K, 387M, 387N, 387Yd Archimedean partially ordered linear space 351R, 351Xe, 361Gb; —– Riesz space 241F, 241Yb, 242Xc, 353A-353F, 353Ha, 353L-353P, 353Xa, 353Xb, 353Xd, 353Ya, 353Yd-353Yh, 354Ba, 354F, 354I-354K, 354Yi, 355Xd, 356G-356I, 361Ee, 367C, 367F, 367Xh, 367Ya, 368B, 368E-368G, 368I, 368J, 368M, 368O, 368P, 368R, 368Ya, 368Yb area see surface measure arrow (‘double arrow space’, ‘two arrows space’) see split interval (343J) associate extended Fatou norm 369H, 369I-369L, 369O, 369R, 369Xd, 374B, 374C, 374Xh, 376S asymptotic density 273G asymptotically equidistributed see equidistributed (281Yi) atom (in a Boolean algebra) 316L, 316M, 316Xr-316Xt, 316Xy, 316Yq, 322Bf, 324Ye, 331H, 333Xb, 343Xe, 362Xe; see also relative atom (331A)

642

Index

atom

atom (in a measure space) 211I, 211Xb, 246G, 322Bf atom (in a Riesz space) 362Xe atomic see purely atomic (211K, 316Lc) atomless Boolean algebra 316Lb, 316Mb, 316Xr, 316Xs, 316Xu-316Xw, 316Yl-316Yn, 316Yp, 316Yq, 322Bg, 322Kc, 324Kf, 324Ye, 332I, 332P, 332Ya, 375B, 375Xb, 375Yb, 383E, 383F, 383L, 383Xc, 385F, 385O, 386Cd, 386C-386H, 386L; see also relatively atomless (331A) atomless additive functional 326Ya, 326Yb-326Yd, 326Yf, 362Xf, 362Xg, 362Yi atomless measure algebra 331C, 369Xh, 374Xl, 382G, 382H, 382J, 382Xf, 382Xg, 382Ya, 383Ld, 383M, 383O, 383P, 383Xe atomless measure (space) 211J, 211Md, 211Q, 211Xb, 211Yd, 211Ye, 212G, 213He, 214Xe, 215D, 215Xe, 215Xf, 216A, 216Ya, 234Xb, 234Yd, 251Xq, 252Yp, 252Yr, 252Yt, 256Xd, 264Yg, 322Bg, 325Ye, 342Xc, 343Cb, 344I atomless vector measure 326Ye automorphism see Banach lattice automorphism, Boolean automorphism, group automorphism (3A6B), measure space automorphism automorphism group of a Boolean algebra §381, 382Yb, §383, §394; of a measure algebra 366Xh, 374J, §§382-383, 385Xa, 394F, 394R axiom see Banach-Ulam problem, choice (2A1J), countable choice Baire property (for subsets of a topological space) 314Yd Baire property algebra 314Yd, 316Yi, 341Yb, 367Yk Baire space 314Yd, 341Yb, 364Yi, 364Yl, 367Yi, 367Yk Baire’s theorem 3A3G Baire σ-algebra 341Yc, 341Zb, 343Xc, 344E, 344Ya, 344Yc-344Ye ball (in Rr ) 252Q, 252Xh; see also sphere Banach algebra > 2A4Jb, 363Xa Banach function space §369, §374 Banach lattice 242G, 242Xc, 242Yc, 242Ye, 243E, 243Xb, 326Yj, 354A, 354C, 354Ee, 354L, 354Xa, 354Xb, 354Xe-354Xj, 354Xn-354Xp, 354Yd, 354Yi-354Yl, 355C, 355K, 355Xb, 355Ya, 356D, 356M, 356Xc, 356Xk, 356Yg, 356Yh, 363E, 365N, 366C, 366Xd, 367D, 367P, 369B, 369G, 369Xe, 371B-371E, 371Xa, 371Xc, 375Xb, 376C, 376M, 376Xg, 376Yj; see also extended Fatou norm, L-space (354M), Lp , M -space (354Gb), Orlicz space Banach lattice automorphism 366Xh Banach space 231Yh, 262Yi, 2A4D, 2A4E, 2A4I, 326Yk, 354Yk, 3A5H, 3A5I; see also Banach algebra (2A4Jb), Banach lattice (242G), separable Banach space Banach-Ulam problem 232Hc, 326 notes, 363S, 376Yg band in a Riesz space §352 (352O), 353B, 353C, 353E, 353H, 353I, 354Bd, 354O, 355H, 355I, 356B, 356L, 362B, 362C, 362Xf-362Xi, 364Xn, 364Xo; see also complemented band, complement of a band, principal band, projection band band algebra (of an Archimedean Riesz space) 353B, 353D, 356Yc, 356Yd, 361Yc, 362Ya, 362Yb, 365S, 365Xm, 366Xb, 368E, 368R; see also complemented band algebra (352Q), projection band algebra (352S) band projection 352R, 352S, 352Xl, 355Yh, 356C, 356Xe, 356Xf, 356Yb, 362B, 362D, 362Ye, 362Yi base for a topology 3A3M basically disconnected topological space 314Yf Bernoulli partition 386A, 386B, 386D-386H, 386J —– shift 384Q, 384R, 384S, 384Xi-384Xm, 385Xc, 386B, 386I, 386K, 386L, 386Xa, 386Xc, 386Ya Berry-Esséen theorem 274Hc, 285 notes bidual of a normed space (U ∗∗ ) 356Xh, 3A5Ga; see also order-continuous bidual Bienaymé’s equality 272R bilinear map §253 (253A), 255Xb, 363L, 376B, 376C, 376Ya-376Yc Bochner integral 253Yf, 253Yg, 253Yi, 354Yl Bochner’s theorem 285Xr Boolean algebra Chap. 31 (311Ab), 363Xf; see also algebra of sets, complemented band algebra (352Q), Dedekind (σ-)complete Boolean algebra, measure algebra (321A), projection band algebra (352S), regular open algebra (314Q)

cellularity

General index

643

Boolean automorphism 363Ye, 366Yh, 374Yd, 381C-381G, 381J, 381L, 381Xe, 381Ya, 383A, 385B, 385Xa, 387B, 387G, 387Xa, 387Yb, 387Yc, 394Ge, 394Ya, 395A; see also involution, measure-preserving Boolean automorphism, von Neumann automorphism (387G), weakly von Neumann automorphism (387G) Boolean homomorphism 312F, 312G-312K, 312N, 312P-312S, 312Xe, 312Xg, 312Xj, 312Ya, 313L-313N, 313P-313R, 313Xp, 313Yb, 314I, 314K, 314Xh, 316Yo, §324, 326Be, 332Yc, 343A, 352Ta, 361Xd, 363F, 363Xb, 363Xc, 384K, 385A, 385C, 385Xa see also aperiodic Boolean homomorphism (385A), Boolean isomorphism, measure-preserving Boolean homomorphism, (sequentially) order-continuous Boolean homomorphism, periodic Boolean homomorphism (385A) Boolean-independent subalgebras 315Xn Boolean isomorphism 312L, 314I, 332Xj, 344Yb, 366Yh Boolean ring §311 (311Aa), §361 Boolean subalgebra see subalgebra of a Boolean algebra (312A) ‘Boolean value’ (of a proposition) 364Bb Borel algebra see Borel σ-algebra Borel-Cantelli lemma 273K Borel lifting 345F, 345Xg, 345Yc Borel measurable function 121C, 121D, 121Eg, 121H, 121K, 121Yf, 134Fd, 134Xd, 134Yt, 135Ef, 135Xc, 135Xe, 225H, 225J, 225Yf, 233Hc, 241Be, 241I, 241Xd, 256M, 364I, 364J, 364Xg, 364Yd, 367Yp Borel measure (on R) 211P, 216A, 341Lg, 343Xd, 345F Borel sets in R, Rr 111G, 111Yd, 114G, 114Yd, 115G, 115Yb, 115Yd, 121Ef, 121K, 134F, 134Xc, 135C, 136D, 136Xj, 212Xc, 212Xd, 225J, 264F, 342Xi, 364H, 364Yc Borel σ-algebra (of subsets of Rr ) 111Gd, 114Yg, 114Yh, 114Yi, 121J, 121Xd, 121Xe, 121Yb, 216A, 251L, 264E, 264Xb, 364G; —– (of other spaces) 135C, 135Xb, 256Ye, 264Yc, 271Ya, 314Ye, 333Ya bounded bilinear map 253Ab bounded linear operator 253Ab, 253F, 253Gc, 253L, 253Xc, 253Yf, 253Yj, 253Yk, 2A4F, 2A4G-2A4I, 2A5If, 355C, 3A5Ed; see also (weakly) compact linear operator (3A5K), order-bounded linear operator (355A), B(U ; V ) (2A4F) bounded set (in Rr ) 134E; (in a normed space) 2A4Bc, 3A5H; (in a linear topological space) 245Yf; see also order-bounded (2A1Ab) bounded , function with see compact bounded variation, function of §224 (224A, 224K), 225Cb, 225M, 225Oc, 225Xh, 225Xn, 225Yc, 226Bc, 226C, 264Yp, 282M, 282O, 283L, 283Xj, 283Xk, 283Xm, 283Xn, 283Xq, 284Xk, 284Yd, 343Yc, 354Xt bounding see ω ω -bounding (316Yg) Breiman see Shannon-McMillan-Breiman theorem (385G) Cantor function 134H, 134I, 222Xd, 225N, 226Cc, 262K, 264Yn Cantor measure 256Hc, 256Ia, 256Xk, 264Ym Cantor set 134G, 134H, 134I, 134Xe, 256Hc, 256Xk, 264J, 264Ym, 264Yn; see also {0, 1}N Carathéodory complete measure space see complete (211A) Carathéodory’s method (of constructing measures) 113C, 113D, 113Xa, 113Xd, 113Xg, 113Yc, 114E, 114Xa, 121Yc, 132Xc, 136Ya, 212A, 212Xg, 213C, 213Xa, 213Xf, 213Xg, 213Ya, 214H, 214Xc, 216Xb, 251C, 251Wa, 251Xd, 264C, 264K cardinal 2A1Kb, 2A1L, 3A1B-3A1E cardinal function (of a Boolean algebra) see cellularity (332D, Maharam type (331F) —– (of a topological space) see cellularity (332Xd, density (331Yf Carleson’s theorem 282K remarks, 282 notes, 284Yg, 286U, 286V category algebra (of a topological space) 314Yd Cauchy distribution 285Xm, 367Xw Cauchy filter 2A5F, 2A5G, 3A4F Cauchy’s inequality 244E Cauchy sequence 242Yc, 2A4D, 354Ed, 356Yf, 356Yg, 3A4Fe cellularity(of a topological space) 332Xd, 332Xe; (of a Boolean algebra) §332 (332D), 365Ya; see also magnitude (332Ga)

644

Index

central

Central Limit Theorem 274G, 274I-274K, 274Xc-274Xg, 285N, 285Xn, 285Ym Césaro sums 273Ca, 282Ad, 282N, 282Xn Chacon-Krengel theorem 371 notes chain condition (in Boolean algebras or topological spaces) see ccc (316A), σ-linked (391L), σ-m-linked (391Yh) chain rule for Radon-Nikod´ ym derivatives 235Xi change of variable in integration §235, 263A, 263D, 263F, 263G, 263I, 263Xc, 263Xe, 263Yc characteristic function (of a set) 122Aa —– (of a probability distribution) §285 (285Aa) —– (of a random variable) §285 (285Ab); charge 326A; see finitely additive functional chargeable Boolean algebra 391D, 391J, 391N, 391X, 391Xa-391Xf, 391Yb, 391Yc, 391Yf, 392F, 394Xd Chebyshev’s inequality 271Xa choice, axiom of 134C, 1A1G, 254Da, 2A1J, 2A1K-2A1M, 361 notes; see also countable choice choice function 2A1J Choquet capacity see capacity circle group 255M, 255Yh, 345Xb, 345Xg clopen algebra 311I, 311J, 311Xh closed convex hull 253G, 2A5E closed interval (in R or Rr ) 114G, 115G, 1A1A; (in general partially ordered spaces) 2A1Ab closed set (in a topological space) 134Fb, 134Xc, 1A2E, 1A2F, 1A2G, 2A2A, 2A3A, 2A3D, 2A3Nb; (in an ordered space) see order-closed (313D) closed subalgebra of a measure algebra 323H-323K (323I), 323Xb, 323Xd, 323Yd, 325Xa, 327F, 331B, 331D, 332P, 332T, §333, 334Xb, 364Xe, 366J, 372G, 372J, 384D, 384G, 384Xq, 385E, 385N-385P, 385Xf, 385Yc, 385Yd, 386Be, 387L; see also order-closed subalgebra closure of a set 2A2A, 2A2B, 2A3D, 2A3Kb cluster point (of a filter) 3A3L; (of a sequence) 2A3O Coarea Theorem 265 notes cofinal set (in a partially ordered set) 3A1F cofinality (of a partially ordered set) 3A1F cofinite see finite-cofinite algebra (231Xa, 316Yk) commutative (ring, algebra) 2A4Jb, 363Xf, 3A2A, 3A2Ff, 3A2Hc commutative Banach algebra 224Yb, 243D, 255Xc, 257Ya, 363B commutative group see abelian group commutator in a group 381Xf, 381Xg, 381Ya compact class of sets 342Ab, 342D compact Hausdorff space 364V, 364W, 364Yj, 364Yl, 364Ym, 3A3D, 3A3Ha, 3A3K compact linear operator 371Ya, 371Yb, 376Yn, 3A5Ka; see also weakly compact operator (3A5Kb) compact measure 342A, 342E-342H, 342J, 342M, 342N, 342Xd-342Xh, 342Xl, 342Xm, 342Xo, 342Ya, 342Yb, 343J, 343K, 343Xa, 343Xb, 343Ye; see also locally compact measure (342Ad) compact set (in a topological space) 247Xc, 2A2D, 2A2E-2A2G, > 2A3N, 2A3R, 3A3D; see also convex compact, relatively compact (2A3N), relatively weakly compact (2A5Id), weakly compact (2A5Ic) compact , function with 242O, 242Pd, 242O, 242Xh, 244H, 244Yi, 256Be, 256D, 256Xh, 262Yd262Yg compact , measure with 284Yi compact topological space > 2A3N, 3A3Dd, 3A3J see also compact Hausdorff space, locally compact (3A3Ag) compactification see one-point compactification (3A3O) complement of a band (in a Riesz space) 352P, 352Xg complementary Young’s functions 369Xc, 369Xd, 369Xr complemented band (in a Riesz space) 352P, 352Xe, 352Xg, 353B, 353Xa; see also complemented band algebra (352Q) complemented band algebra 352Q, 352S, 352Xf, 353B, 356Yc

continuity

General index

645

complete generation of a Boolean algebra see Maharam type (331F) complete lattice 314Bf complete linear topological space 245Ec, 2A5F, 2A5H, 367Nc complete locally determined measure (space) 213C, 213D, 213H, 213J, 213L, 213O, 213Xg, 213Xi, 213Xl, 213Yd, 213Ye, 214I, 214Xc, 216D, 216E, 234F, 251Yb, 252B, 252D, 252E, 252N, 252Xb, 252Yh, 252Yj, 252Yr-252Yt, 253Yj, 253Yk, 325B, 341M, 341Pb, 342N, 364Xm, 376S, 376Xj; see also c.l.d. version (213E) complete measure (space) 112Df, 113Xa, 122Ya, 211A, 211M, 211N, 211R, 211Xc, 211Xd, §212, 214I, 214J, 216A, 216C, 216Ya, 234A, 234D, 254F, 254G, 254J, 264Dc, 321K, 341J, 341K, 341Mb, 341Xc, 341Yd, 343B, 344C, 344I, 344Xb-344Xd, 384V; see also complete locally determined measure complete metric space 224Ye, 323Yd, 393Xb, 3A4Fe, 3A4Hc; see also Banach space (2A4D) complete normed space see Banach space (2A4D), Banach lattice (242G, 354Ab) complete Riesz space see Dedekind complete (241Fc) complete uniform space 323G, 3A4F, 3A4G, 3A4H; see also completion (3A4H) complete (in ‘κ-complete filter’) 368F complete see also Dedekind complete, Dedekind σ-complete, uniformly complete (354Yi) completed indefinite-integral measure see indefinite-integral measure (234B) completely additive functional on a Boolean algebra 326J, 326K-326P, 326Xe, 326Xf, 326Xh, 326Xi, 326Yp-326Ys, 327B-327E, 327Ya, 332Xo, 362Ad, 362B, 362D, 362Xe-362Xg, 363K, 363S, 365E, 394Xe; see also Mτ completely additive part (of an additive functional on a Boolean algebra) 326Yq, 362Bd; (of a linear functional on a Riesz space) 355Yh completely continuous linear operator see compact linear operator (3A5Ka) completely generate see τ -generate (331E) completely regular topological space 353Yb, 367Yi, > 3A3Ac, 3A3B completion (of a Boolean algebra) see Dedekind completion (314U) —– (of a linear topological space) , —– (of a measure (space)) §212 (212C), 213Fa, 213Xk, 214Xb, 214Xj, 232Xe, 234D, 235D, 235Xe, 241Xb, 242Xb, 243Xa, 244Xa, 245Xb, 251S, 251Wn, 251Xp, 252Ya, 254I, 256C, 322Da, 342Gb, 342Ib, 342Xn, 343Ac —– (of a metric space) 393B —– (of a normed space) 3A5Ib —– (of a normed Riesz space) 354Xh, 354Yg —– (of a uniform space) 325Ea, 325Yb, 3A4H complexification of a Riesz space 354Yk, 355Yk, 356Yh, 362Yj, 363Xj, 364Yn complex-valued function §133 component (in a Riesz space) 352Rb, 352Sb, 356Yb, 368Yj; (in a topological space) 111Ye; see also Maharam-type-κ component (332Gb) concave function 384A conditional entropy 384D, 384E, 384G, 384N, 384Xb-384Xd, 384Xq, 384Ya, 384Yb, 385N conditional expectation 233D, 233E, 233J, 233K, 233Xg, 233Yc, 235Yc, 242J, 246Ea, 253H, 253Le, 275Ba, 275H, 275I, 275K, 275Ne, 275Xi, 275Ya, 275Yk, 275Yl, 365R, 369Xq, 372K conditional expectation operator 242Jf, 242K, 242L, 242Xe, 242Yk, 243J, 244M, 244Yk, 246D, 254R, 254Xp, 275Xd, 275Xe, 327Xb, 365R, 365Xl, 366J, 372G, 384B, 384D, 384E conditionally (order)-complete see Dedekind complete (314A) cone 3A5B, 3A5D; see also positive cone (351C) conegligible set 112Dc, 214Cc connected set 222Yb conjugate operator see adt operator (3A5Ed) consistent disintegration 452A, 452C, 452E, 452I, 452K, 452Xg, 452Ya, 452Yd, 454Xf, 457R, 457Xl consistent lifting 346I, 346J, 346L, 346Xe, 346Yd continued fractions 372M-372O; (continued-fraction approximations) 372M, 372Xj, 372Yf; (continuedfraction coefficients) 372M, 372O, 372Xh-372Xj, 372Xt, 372Ye continuity, points of 224H, 224Ye, 225J

646

Index

continuous

continuous function 121D, 121Yf, 262I, 2A2C, 2A2G, 2A3B, 2A3H, 2A3Nb, 2A3Qb, 3A3C, 3A3Eb, 3A3Ib, 3A3Nb continuous at a point 3A3C continuous linear functional 284Yj; see also dual linear space (2A4H) continuous linear operator 2A4Fc; see also bounded linear operator (2A4F) continuous outer measure see Maharam submeasure (392G) continuous see also order-continuous (313H), semi-continuous (225H), uniformly continuous (3A4C) continuum see c (2A1L) control measure 393O, 393P, 393S, 393Xg, 393Yh, 393Yi Control Measure Problem §393 convergence in mean (in L1 (µ) or L1 (µ)) 245Ib >245A), 246J, 246Yc, 247Ya convergence in measure (in L0 (µ)) 232Ya, §245 (> >245A), 274Yd —– (in L0 (µ)) §245 (> —- (in L0 (A)) §367 (367M), 369M, 369Xf, 369Ye, 393M —– (of sequences) §245, 246J, 246Xh, 246Xi, 253Xe, 255Yf, 273Ba, 276Yf, 367Na, 367Q, 369Yf, 376Xf convergent almost everywhere 245C, 245K, 273Ba, 276G, 276H; see also strong law convergent filter 2A3Q, 2A3S, 2A5Ib, 3A3Ic, 3A3Lb; —– sequence 135D, 245Yi, 2A3M, 2A3Sg, 3A3Lc; see also convergent in measure (245Ad), order*convergent (367A) convex function 233G, 233H-233J, 233Xb, 233Xc, 233Xd, 233Xe, 233Xh, 233Ya, 233Yc, 242K, 242Yi, 242Yj, 242Yk, 244Xm, 244Yg, 255Yk, 275Yg, 365Rb, 367Xx, 369Xb, 369Xc, 369Xd, 373Xb; see also midconvex (233Ya) convex hull 2A5E; see also closed convex hull (2A5E) convex set 233Xd, 244Yj, 262Xh, 2A5E, 326Ye, 351Ce, 3A5C, 3A5Ee, 3A5Ld —– see also cone (3A5B), order-convex (351Xb) convex structure 373Xv, 373Yf convolution in L0 255Fc, 255Xc, 255Yf, 255Yk convolution of functions 255E, 255F-255K, 255O, 255Xa-255Xc, 255Xf-255Xj, 255Ya, 255Yb, 255Yd, 255Ye, 255Yi, 255Yl, 255Ym, 255Yn, 262Xj, 262Yd, 262Ye, 263Ya, 282Q, 282Xt, 283M, 283Wd, 283Wf, 283Wj, 283Xl, 284J, 284K, 284O, 284Wf, 284Wi, 284Xb, 284Xd, 373Xg, 376Xa, 376Xd convolution of measures §257 (257A), 272S, 285R, 285Yn, 364Xf convolution of measures and functions 257Xe, 284Xo, 284Yi convolution of sequences 255Xe, 255Yo, 282Xq, 352Xk, 376Ym countable (set) 111F, 114G, 115G, §1A1, 226Yc countable antichain condition 316 notes (see ccc (316A)) countable chain condition see ccc (316A) countable choice (axiom of) 134C countable-cocountable algebra 211R, 211Ya, 232Hb, 316Yk, 368Xa countable-cocountable measure 211R, 232Hb, 252L, 326Xf, 342M, 356Xc countable sup property (in a Riesz space) 241Yd, 242Yd, 242Ye, 244Yb, 353Ye, 354Yc, 355Yi, 356Ya, 363Yb, 364Yb, 366Yb, 368Yh, 376Ye, 376Yf countably additive functional (on a σ-algebra of sets) §231 (231C), §232, 246Yg, 246Yi, 327C, 362Xg, 363S; (on a Boolean algebra) > 326E, 326F-326I, 326L, 326Xb-326Xi, 327B, 327C, 327F, 327G, 327Xa, 327Xd, 327Xe, 362Ac, 362B, 363S, 363Yh, 392Xf; see also completely additive (326J), countably subadditive (392H), Mσ countably additive part (of an additive functional on a Boolean algebra) 326Yn, 362Bc, 362Ye, 393Yg countably full group (of automorphisms of a Boolean algebra) 381Xd countably generated σ-algebra 381Yb, 381Yc, 382Yf countably separated (measure space, σ-algebra of sets) 343D, 343E-343H, 343K, 343L, 343Xe, 343Xh, 343Yb, 343Ye, 343Yf, 344B, 344C, 344I, 344Xb-344Xd, 384V, 387C, 387D countably subadditive functional 392H counting measure 112Bd, 122Xd, 122 notes, 211N, 211Xa, 212Ya, 226A, 241Xa, 242Xa, 243Xl, 244Xi, 244Xj, 244Xn, 245Xa, 246Xc, 251Xb, 251Xh, 252K, 255Yo, 264Db, 324Xe

determined

General index

647

cover see measurable envelope (132D) covering theorem 221A, 261B, 261F, 261Xc, 261Ya, 261Yi, 261Yk cozero set 3A3Pa cycle notation (for automorphisms of Boolean algebras) 381G, 381H, 381I, 385B; see also pseudo-cycle (387Xc) cyclic automorphism 381G; see also exchanging involution (381G) cylinder (in ‘measurable cylinder’) 254Aa, 254F, 254G, 254Q, 254Xa Davies R.O. 325 notes decimal expansions 273Xf decomposable measure (space) see strictly localizable (211E) decomposition (of a measure space) 211E, 211Ye, 213O, 213Xh, 214Ia, 214K, 214M, 214Xi, 322L; see also Hahn decomposition of an additive functional (231F), Jordan decomposition of an additive functional (231F), Lebesgue decomposition of a countably additive functional (232I), Lebesgue decomposition of a function of bounded variation (§226) decreasing rearrangement (of a function) 252Yp; (of an element of M 0,∞ (A)) §373 (373C), 374B-374D, 374J, 374L, 374Xa, 374Xj, 374Ya-374Yd Dedekind complete Boolean algebra 314B, 314Ea, 314Fa, 314G, 314Ha, 314I, 314Ja, 314K, 314P, 314S, 314T, 314Va, 314Xa, 314Xb, 314Xf, 314Xg, 314Xh, 314Yd, 314Yg, 314Yh, 315F, 315Yf, 316Fa, 316K, 316Xt, 316Yi, 316Yn, 331Yd, 332Xa, 333Bc, 352Q, 363Mb, 363P-363S, 364O, 385B, 385C, 368A-368D, 368Xe, 375C, 375D, 375Xa, 375Xc, 375Ya, §381, 383D, 383J, 393A-393C, 393J, 393Xc-393Xf, 393Ya, 393Yc-393Yf Dedekind complete partially ordered set 135Ba, 314Aa, 314B, 314Ya, 315De, 343Yc, 3A6Csee also Dedekind complete Boolean algebra, Dedekind complete Riesz space Dedekind complete Riesz space 241Fc, 241G, 241Xf, 242H, 242Yc, 243H, 243Xj, 244L, 353G, 353I, 353Jb, 353K, 353Yb, 354Ee, §355, 356B, 356K, 356Xj, 356Ye, 361H, 363Mb, 363P, 363S, 364O, 366C, 366G, 368H-368J, 371B-371D, 371Xd, 371Xe, 375J, 375Ya Dedekind completion of a Boolean algebra 314T, 314U, 314Xe, 314Yh, 315Yd, 322O, 361Yc, 365Xm, 366Xk, 368Xb, 383Lf, 383Xb, 383Xd, 391Xc; see also localization (322P) Dedekind completion of a Riesz space 368I, 368J, 368Xb, 368Yb, 369Xp Dedekind σ-complete Boolean algebra 314C-314F, 314Hb, 314Jb, 314M, 314N, 314Xa, 314Xc, 314Ye, 314Yf, 315O, 316C, 316Fa, 316Hb, 316J, 316Xg, 316Yi, 316Yk, 321A, 324Yb, 324Yc, 326Fg, 326I, 326O, 326P, 326Xc, 326Xi, 326Yf, 326Yo, 344Yb, 362Xa, 363Ma, 363P, 363Xh, 363Yh, §364, 367Xk, 368Yd, 375Yb, 381B, 381C, 381T, 381Ya, 392H, 392Xd, 392Xf, 393A, 393E, 393M, 393O-393Q, 393S, 393Xb, 393Yg-393Yi Dedekind σ-complete partially ordered set 314Ab, 315De, 367Bf; see also Dedekind σ-complete Boolean algebra, Dedekind σ-complete Riesz space Dedekind σ-complete Riesz space 241Fb, 241G, 241Xe, 241Yb, 241Yh, 242Yg, 243H, 243Xb, 353G, 353H, 353J, 353Xb, 353Yc, 353Yd, 354Xn, 354Yi, 354Yl, 356Xc, 356Xd, 363M-363P, 364C, 364E delta function see Dirac’s delta function (284R) delta system see ∆-system De Morgan’s laws 311Xf; see also distributive laws dense set in a topological space 254Yc, 2A3U, 313Xj, 323Dc, 3A3E, 3A3G, 3A3Ie, 3A4Ff; see also nowhere dense (3A3Fa), order-dense (313J, 352Na), quasi-order-dense (352Na) density (of a topological space) 331Ye, 331Yf, 365Ya, 393Yf density function (of a random variable) 271H, 271I-271K, 271Xc-271Xe, 272T, 272Xd, 272Xj; see also Radon-Nikod´ ym derivative (232If, 234B) density point 223B, 223Xi, 223Yb density topology 223Yb, 223Yc, 223Yd, 261Yf density see also asymptotic density (273G), Lebesgue’s Density Theorem (223A) derivative of a function (of one variable) 222C, 222E, 222F, 222G, 222H, 222I, 222Yd, 225J, 225L, 225Of, 225Xc, 226Be, 282R; (of many variables) 262F, 262G, 262P; see also gradient, partial derivative determinant of a matrix 2A6A determined see locally determined measure space (211H), locally determined negligible sets (213I) determined by coordinates (in ‘W is determined by coordinates in J’) 254M, 254O, 254R, 254S, 254T, 254Xp, 254Xr, 311Xh, 325N, 325Xh

648

Index

Devil

Devil’s Staircase see Cantor function (134H) differentiability, points of 222H, 225J differentiable function 123D, 222A, 224I, 224Kg, 224Yc, 225L, 225Of, 225Xc, 225Xn, 233Xc, 252Yj, 255Xg, 255Xh, §262 (262F), 263D, 263Xc, 263Yc, 265E, 265Xd, 265Ya, 274E, 282L, 282Xs, 282Xk, 283I283K, 283Xm, 284Xk; see also derivative ‘differentiable relative to its domain’ 262F diffused measure see atomless measure (211J) dilation 286C dimension see Hausdorff dimension (264 notes)) Dirac’s delta function 257Xa, 284R, 284Xn, 284Xo, 285H, 285Xp direct image (of a set under a function or relation) 1A1B direct sum of measure spaces 214K, 214L, 214Xi-214Xl, 241Xg, 242Xd, 243Xe, 244Xg, 245Yh, 251Xh, 251Xi, 322Kb, 332C, 342Gd, 342Ic, 342Xn, 343Yb directed set see downwards-directed (2A1Ab), upwards-directed (2A1Ab) Dirichlet kernel 282D; see also modified Dirichlet kernel (282Xc) disconnected see extremally disconnected (3A3Ae), basically disconnected discrete topology (3A3Ah) dist family (of sets) 112Bb; (in a Boolean ring) 311Gb; (in a Riesz space) 352C dist sequence theorem 246G, 246Ha, 246Yd, 246Ye, 246Yf, 246Yj, 354Rb, 356O, 356Xm dist set (in a Boolean ring) 311Gb; (in a Riesz space) 352C dist union topology 315Xe distribution see Schwartzian distribution, tempered distribution distribution of a random variable 241Xc, 271B, 271C, 271D-271G, 272G, 272S, 272Yf, 285Ab, 285Mb, 364Xd, 364Xf, 367Xv, 373Xh; see also empirical distribution (273 notes) distribution function of a random variable > 271G, 271Xb, 271Yb, 271Yc, 272Xe, 272Yc, 273Xg, 273Xh, 285P distributive lattice 367Yb, 381Xi, 3A1Ic; see also (2, ∞)-distributive (367Yd) distributive laws in Boolean algebras 313B; see also De Morgan’s laws, weakly σ-distributive (316Yg), weakly (σ, ∞)-distributive (316G) distributive laws in Riesz spaces 352E; see also weakly σ-distributive, weakly (σ, ∞)-distributive (368N) Dominated Convergence Theorem see Lebesgue’s Dominated Convergence Theorem (123C) Doob’s Martingale Convergence Theorem 275G double arrow space see split interval (343J) doubly stochastic matrix 373Xe downwards-directed partially ordered set 2A1Ab dual linear operator see adt operator (3A5Ed) dual linear space (of a normed space) 243G, 244K, 2A4H, 356D, 356N, 356O, 356P, 356Xg, 356Yh, 365I, 365J, 366C, 366D, 369K, 3A5A, 3A5C, 3A5E-3A5H dual —– see also algebraic dual, bidual, order-bounded dual (356A), order-continuous dual (356A), sequentially order-continuous dual (356A) Dunford’s theorem 376N dyadic cycle system (for a Boolean automorphism) 387G, 387K, 387M Dye’s theorem 387O, 387 notes Dynkin class 136A, 136B, 136Xb Eberlein’s theorem 2A5J, 356 notes Egorov’s theorem 131Ya, 215Yb Egorov property (of a Boolean algebra) 316Yg, 316Yh empirical distribution 273Xh, 273 notes entropy §384; (of a partition of unity) 384C, 385G, 385K, 385N, 385Q, 386B-386D, 386F-386H; (of a measure-preserving Boolean homomorphism) 384M, 385N, 385Xe, 385Yc, 385Yd, 386C-386E, 386I-386L, 387Yd; see also conditional entropy (384D), entropy metric (384Xq) entropy metric 384Xq, 384Yg

finitely

General index

649

enumerate, enumeration 3A1B envelope see measurable envelope (132D), upr(a, C) (314V) equidecomposable (in ‘G-equidecomposable’) 395Ya; (in ‘G-τ -equidecomposable’) §394 (394A) equidistributed sequence (in a topological probability space) 281N, 281Yi, 281Yj, 281Yk Equidistribution Theorem see Weyl’s Equidistribution Theorem (281N) equivalent norms 355Xb, 369Xd equiveridical 121B, 212B, 312B ergodic automorphism (of a Boolean algebra) 387Xg, 394Ge ergodic group of automorphisms (of a Boolean algebra) 394Ge, 394Q, 394R, 394Xa, 394Xg, 394Yd ergodic inverse-measure-preserving function 372P, 372Q, 372Xo, 372Xs, 372Yi, 372Yj, 387Yf ergodic measure-preserving Boolean homomorphism (of a measure algebra) 372P, 372Q, 372Xk, 372Xl, 384Se, 385F, 385H, 386C-386E, 386J, 386Xb, 387Xg, 387Ye, 395Xc Ergodic Theorem 372D, 372E, 372Ya, 385Xc; see also Maximal Ergodic Theorem (372C), Mean Ergodic Theorem (372Xa), Wiener’s Dominated Ergodic Theorem (372Yb) essential supremum of a family of measurable sets 211G, 213K, 215B, 215C; of a real-valued function 243D, 243I, 376S, 376Xn essentially bounded function 243A Etemadi’s lemma 272U Euclidean metric (on Rr ) 2A3Fb Euclidean topology §1A2, §2A2, 2A3Ff, 2A3Tc even function 255Xb, 283Yb, 283Yc exchangeable sequence of random variables 276Xe exchanging involution 381G, 381H, 381K, 381Yb exhaustion, principle of 215A, 215C, 215Xa, 215Xb, 232E, 246Hc, 342B, 365 notes exhaustive submeasure 392Bb, 392C, 392Hc, 392Xb-392Xd, 393A-393C, 393H; see also uniformly exhaustive (392Bc) expectation of a random variable 271Ab, 271E, 271F, 271I, 271Xa, 272Q, 272Xb, 272Xi, 285Ga, 285Xo; see also conditional expectation (233D) extended Fatou norm §369 (369F), 376S, 376Xn, 376Yl; see also T -invariant extended Fatou norm (374Ab), rearrangement-invariant extended Fatou norm (374Eb) extended real line 121C, §135 extension of finitely additive functionals 391G extension of measures 132Yd, 212Xk, 327Xf; see also completion (212C), c.l.d. version (213E) extremally disconnected topological space 314S, 353Yb, 363Yc, 364W, 364Yk-364Ym, 368G, 368Yc, 3A3Ae, 3A3Bd extreme point of a convex set 353Yf factor (of an automorphism of a Boolean algebra) 386Ac, 386L, 386Xb fair-coin probability 254J Fatou’s Lemma 123B, 133K, 135G, 135Hb, 365Xc, 365Xd, 367Xf Fatou norm on a Riesz space 244Yf, 354Da, 354Eb, 354J, 354Xk, 354Xn, 354Xo, 354Ya, 354Ye, 356Da, 367Xg, 369G, 369Xe, 371B-371D, 371Xc, 371Xd; see also extended Fatou norm (369F) Féjer integral 283Xf, 283Xh-283Xj Féjer kernel 282D Féjer sums 282Ad, 282B-282D, 282G-282I, 282Yc Feller, W. Chap. 27 intro. field (of sets) see algebra (136E) filter 254Ta, 2A1I, 2A1N, 2A1O, 2A5F; see also convergent filter (2A3Q), ultrafilter (2A1N) finite-cofinite algebra 231Xa, 231Xc, 316Yk, 392Xc finite rank operators see abstract integral operators (376D) finitely additive functional on an algebra of sets 136Xg, 136Ya, 136Yb, 231A, 231B, 231Xb-231Xe, 231Ya-231Yh, 327C; see also countably additive, completely additive

650

Index

finitely

finitely additive function(al) on a Boolean algebra 326A, 326B-326D, 326Fa, 326G, 326Kg, 326Xa-326Xc, 326Xe, 326Xg, 326Xi, 326Yt-326Yd, 326Yg-326Yj, 326Yk, 326Yl-326Yn, 326Yq, 326Yr, 327B, 327C, 331B, 361Xa-361Xc, 391D-391G, 391I, 391J, 391N, 393Be, 393C, 394N, 394O; see also chargeable Boolean algebra (391X), completely additive functional (326J), countably additive functional (231C, 326E) finitely additive function(al) on a Boolean ring 361B, 361C, 361E-361I, 365Eb finitely additive measure 326A first category see meager (3A3Fb) fixed-point subalgebra (in a Boolean algebra) 394G, 394H, 394I, 394K-394N, 394P, 394Xb-394Xf, 394Yb, 394Yc Fourier coefficients 282Aa, 282B, 282Cb, 282F, 282I, 282J, 282M, 282Q, 282R, 282Xa, 282Xg, 282Xq, 282Xt, 282Ya, 283Xu, 284Ya, 284Yg Fourier’s integral formula 283Xm Fourier series 121G, §282 (282Ac) Fourier sums 282Ab, 282B-282D, 282J, 282L, 282O, 282P, 282Xi-282Xk, 282Xp, 282Xt, 282Yd, 286V, 286Xb Fourier transform 133Xd, 133Yc, §283 (283A, 283Wa), §284 (284H, 284Wd), 285Ba, 285D, 285Xd, 285Ya, 286U, 286Ya Fourier-Stieltjes transform see characteristic function (285A) Fréchet filter 2A3Sg free product of Boolean algebras §315 (315H), 316Yl, §325, 326Q, 326Yt, 326Yu, 361Yf, 383Le, 391Xb, 391Ye; see also localizable measure algebra free product (325Ea), probability algebra free product (325K) Frol´ık’s theorem 381J, 381Yb Fubini’s theorem 252B, 252C, 252H, 252R, 393Td full local semigroup (of partial automorphisms of a Boolean algebra) 394A, 394B, 394C, 394Gc, 394Xc, 394Ya, 394Yd full outer measure 132F, 132G, 132Yd, 133Yf, 134D, 134Yt, 214F, 322Jb, 324Cb, 343Xa full subgroup of Aut A 381M, 381N, 381O, 381R, 381S, 381Xb, 381Xc, 381Xi, 381Xj, 381Xk, 382C, §387; see also countably full (381Xd) —– subsemigroup of Aut A 387Ya fully non-paradoxical group of automorphisms of a Boolean algebra 394E, 394F, 394H, 394I, 394K-394R, 394Xd-394Xf, 394Xh, 394Ya-394Yc, 395B, 395Ya fully symmetric extended Fatou norm see T -invariant (374Ab) function 1A1B function space see Banach function space (§369) functionally closed see zero set (3A3Pa) functionally open see cozero set (3A3Pa) Fundamental Theorem of Calculus 222E, 222H, 222I, 225E Fundamental Theorem of Statistics 273Xh, 273 notes Gaifman’s example 391N, 391Yf-391Yh, game see infinite game, Banach-Mazur game Gamma function see Γ-function (225Xj) Gauss C.F. 372Yg Gaussian random variable, distribution see normal random variable (274Ad) generated topology 3A3Ma generated (σ-)algebra of sets 111Gb, 111Xe, 111Xf, 121J, 121Xd, 136B, 136C, 136G, 136Xc, 136X, 136Xl, 136Yb generated (σ)-subalgebra of a Boolean algebra 313F, 313M, 314G, 314Ye, 331E generating Bernoulli partition 386Ab, 386Bg generating set in a Boolean algebra 331E; see also σ-generating (331E), τ -generating (331E) Glivenko-Cantelli theorem 273 notes group 255Yn, 255Yo; see also amenable group, automorphism group, circle group, ergodic group, simple group group automorphism 3A6B; see also inner automorphism (3A6B), outer automorphism (3A6B)

independent

General index

651

Hahn-Banach theorem 363R, 373Yd, 3A5A, 3A5C Hahn decomposition of an additive functional 231F, 326I, 326O —– see also Vitali-Hahn-Saks theorem (246Yg) Hajian-Ito theorem 395B half-open interval (in R or Rr ) 114Aa, 114G, 114Xe, 114Yj, 115Ab, 115Xa, 115Xc, 115Yd Hall’s Marriage Lemma 3A1K Halmos-Rokhlin-Kakutani lemma 385E Hardy-Littlewood Maximal Theorem 286A Hausdorff dimension 264 notes Hausdorff measure §264 (264C, 264Db, 264K, 264Yo), 343Ye, 345Xb, 345Xg; see also normalized Hausdorff measure (265A) Hausdorff metric (on a space of closed subsets) 246Yb Hausdorff outer measure §264 (264A, 264K, 264Yo) Hausdorff topology 2A3E, 2A3L, 2A3Mb, 2A3S, 3A3Aa, 3A3D, 3A3Id, 3A3K, 3A4Ac, 3A4Fd Hausdorff uniformity 3A4Ac Hilbert space 244N, 244Yj, 3A5L; see also inner product space (3A5L) Hilbert transform 376Ym Hölder’s inequality 244Eb homeomorphism 3A3Ce homogeneous Boolean algebra 331M, 331N, 331Xj, 331Yg, 331Yh, 331Yi, 331Yj, 381Q, 381T, 381Yd, 383E, 383F, 383G, 383La, 383Xc; see also Maharam-type-homogeneous (331Fc), relatively Maharam-typehomogeneous (333Ac) —– measure algebra 331N, 332M, 373Yb, 374H, 374Yc, 375Kb, 382E, 382F, 382I, 394R; see also quasihomogeneous (374G) —– measure space —– probability algebra 333P, 333Yc, 334E, 372Xk homomorphic image (of a Boolean algebra) 314M homomorphism see Boolean homomorphism (312F), group homomorphism, lattice homomorphism (3A1I), ring homomorphism (3A2D) hull see convex hull (2A5E), closed convex hull (2A5E) Humpty Dumpty’s privilege 372 notes ideal in an algebra of sets 232Xc, 312C, 363Yf;see also σ-ideal (112Db) —– in a Boolean ring or algebra 311D, 312C, 312K, 312O, 312Xd, 312Xi, 312Xk, 381S, 392Xd; see also principal ideal (312D), σ-ideal (313E) —– in a Riesz space see solid linear subspace (352J) —– in a ring 352Xl, 3A2E, 3A2F, 3A2G idempotent 363Xf, 363Xg identically distributed random variables 273I, 274 notes, 276Yg, 285Xn, 285Yc, 372Xg; see also exchangeable sequence (276Xe) image filter 2A1Ib, 2A3Qb, 2A3S image measure 112E, 112F, 112Xd, 123Ya, 132G, 132Yb, 132Yf, 211Xd, 212Bd, 212Xg, 235L, 254O, 256G, 342Xh, 342Xj, 343Xb, 343Yf, 386Xd image measure catastrophe 235J, 343 notes indecomposable transformation see ergodic inverse-measure-preserving function (372Pb) indefinite integral 131Xa, 222D-222F, 222H, 222I, 222Xa-222Xc, 222Yc, 224Xg, 225E, 225Od, 225Xh, 232D, 232E, 232Yf, 232Yi; see also indefinite-integral measure indefinite-integral measure §234 (234A), 235M, 235P, 235Xi, 253I, 256E, 256J, 256L, 256Xe, 256Yd, 257F, 257Xe, 263Ya, 275Yi, 275Yj, 285Dd, 285Xe, 285Ya, 322Xf, 342Xd, 342Xn, 342Yd; see also uncompleted indefinite-integral measure independence §272 (272A) independent elements of L0 (A) 364Xe, 364Xf, 367Xw independent family in a probability algebra 325Xe, 325Yf, 371Yc

652

Index

independent

independent random variables 272Ac), 272D-272I, 272L, 272M, 272P-272U, 272Xb, 272Xd, 272Xh-272Xj, 272Ya-272Yd, 272Yf, 272Yg, 273B, 273D, 273E, 273H, 273I, 273L-273N, 273Xh, 273Xi, 273Xk, 274B-274D, 274F-274K, 274Xc, 274Xd, 274Xg, 275B, 275Yh, 276Af, 285I, 285Xf, 285Xg, 285Xm-285Xo, 285Yc, 285Yk, 285Yl independent sets 272Aa, 272Bb, 272F, 272N, 273F, 273K independent subalgebras of a probability algebra 325L, 325Xe, 325Xf, 325Xg, 327Xe, 364Xe, 384Q, 384Sf, 384Xd-384Xf, 386B; see also Boolean-independent (315Xn) independent σ-algebras 272Ab, 272B, 272D, 272F, 272J, 272K, 272M, 272O, 275Ym Individual Ergodic Theorem see Ergodic Theorem (372D, 372E) induced automorphism (on a principal ideal of a Boolean algebra) 387E, 387I, 387J, 387Xb induced topology see subspace topology (2A3C) inductive definitions 2A1B infimum in a Boolean algebra 313A-313C infinity 112B, 133A, §135 information 384Cb initial ordinal 2A1E, 2A1F, 2A1K injective ring homomorphism 3A2Da, 3A2G inner automorphism of a group 3A6B inner measure 113Yh, 212Yc, 213Xe, 213Yc, 324Cb inner product space 244N, 253Xe, 3A5L; see also Hilbert space (3A5L) inner regular measure 256A, 256B, 342Aa, 342C, 342Xa —– —– with respect to closed sets 256Ya, 342F —– —– with respect to compact sets 342Xh, 343Yc integrable function §122 (122M), 123Ya, 133B, 133Db, 133Dc, 133F, 133J, 133Xa, 135Fa, 212B, 212F, 213B, 213G; see also Bochner integrable function (253Yf), L1 (µ) (242A) integral §122 (122E, 122K, 122M), 363L; (on an L-space) 356P; see also integrable function, Lebesgue integral (122Nb), lower integral (133I), Riemann integral (134K), upper integral (133I) integral operator see kernel operator, abstract integral operator (§376) integration by parts 225F, 225Oe, 252Xi integration by substitution see change of variable in integration interior of a set 2A3D, 2A3Ka interpolation see Riesz Convexity Theorem intersection number (of a family in a Boolean algebra) 391H, 391I-391K, 391Xh, 391Xi, 392E interval see half-open interval (114Aa, 115Ab), open interval (111Xb, split interval (343J) invariant function (in ‘G-invariant function’) 394A, 394N-394P, 394Xd, 394Xe, 394Z, 395B, 395Xd —– ideal 381S —– lower density 346Xd —– see also rearrangement-invariant (374E), T -invariant (374A), translation-invariant inverse (of an element in an f -algebra) 353Pc, 353Yh inverse Fourier transform 283Ab, 283B, 283Wa, 283Xb, 284I; see also Fourier transform inverse image (of a set under a function or relation) 1A1B inverse-measure-preserving function 132G, 134Yl, 134Ym, 134Yn, 235G, 235H, 235I, 235Xe, 241Xh, 242Xf, 243Xn, 244Xo, 246Xf, 254G, 254H, 254K, 254O, 254Xc-254Xf, 254Xh, 254Yb, 324M, 324N, 341P, 341Xd, 341Yc, 341Yd, 343Ad, 343C, 343J, 343Xd, 343Xh, 343Yd, 365Xi, 372H-372L, 372Xe, 372Xf, 372Yh, 382Yd, 384S-384V, 384V; isomorphic inverse-measure-preserving functions 384Tb; see also almost isomorphic (384U); see also ergodic inverse-measure-preserving function (372P), image measure (112E), mixing inverse-measure-preserving function (372P) Inversion Theorem (for Fourier series and transforms) 282G-282I, 282L, 282O, 282P, 283I, 283L, 284C, 284M; see also Carleson’s theorem invertible element(in an f -algebra) 353Pc, 353Yh involution in a group 3A6A, 381Xl involution in Aut A 381K, 381N, 381O, 381Xd, 381Xe, 381Xf, 381Xh, 381Xk, 381Ya, 381Yc, 381Z, 382D, 382G, 382Yf, 384Xl; see also exchanging involution (381G), ‘many involutions’ (381P)

Lebesgue

General index

653

irreducible continuous function 313Ye isodiametric inequality 264H, 264 notes isolated point 316M, 316Yj, 316Yq, 3A3Pb isometry 324Yh, 3A4Ff, 3A4G isomorphic see almost isomorphic (384U) isomorphism see Boolean isomorphism, measure space isomorphism, order-isomorphism Jacobian 263Ea Jensen’s inequality 233I, 242Yi, 365Rb t distribution see distribution (271C) Jordan decomposition of an additive functional 231F, 231Ya, 232C, 326D, 326H, 326M Kakutani’s theorem 369E, 369Yb; see also Halmos-Rokhlin-Kakutani lemma (385E) Kalton-Roberts theorem 392F, 392J Kawada’s theorem 394Q Kechris A.S. see Becker-Kechris theorem, Nadkarni-Becker-Kechris theorem Kelley’s criterion 391L; see also Nachbin-Kelley theorem kernel (of a ring homomorphism) 3A2D, 3A2Eb, 3A2G; see also Dirichlet kernel (282D), Féjer kernel (282D), modified Dirichlet kernel (282Xc) kernel operator 373Xf, §376 Kirzbraun’s theorem 262C Kolmogorov’s Strong Law of Large Numbers 273I, 275Yn, 372Xg Kolmogorov-Sinaˇı theorem 384P Komlós’ theorem 276H, 276Yh, 367Xn Kronecker’s lemma 273Cb Kwapien’s theorem 375I Lacey-Thiele Lemma 286M Laplace’s central limit theorem 274Xe Laplace transform 123Xc, 123Yb, 133Xc, 225Xe laterally complete Riesz space 368L, 368M lattice 231Yc, 2A1Ad, 311L, 311Xi, 311Yb, 315Xb, 315Yg, 367A, 367B, 367Xa, 367Ye, 3A1I; see also Riesz space (=vector lattice) (352A), Banach lattice (242G, 354Ab) —– homomorphism 352Nd, 3A1Ia —– norm see Riesz norm (242Xg, 354Aa) —– of normal subgroups 381Xi, 382H, 382J, 382Xf, 382Xg, 382Ya, 3A6C law of a random variable see distribution (271C) law of large numbers see strong law (§273) Lebesgue, H. Vol. 1, intro., Chap. 27, intro. Lebesgue Covering Lemma 2A2Ed Lebesgue decomposition of a countably additive functional 232I, 232Yb, 232Yg, 256Ib Lebesgue decomposition of a function of bounded variation 226C, 226Dc, 226Ya, 232Yb Lebesgue density see lower Lebesgue density (341E) Lebesgue’s Density Theorem §223, 261C, 275Xg Lebesgue’s Dominated Convergence Theorem 123C, 133G, 363Yh, 367J, 367Xf Lebesgue extension see completion (212C) Lebesgue integrable function 122Nb, 122Yb, 122Ye, 122Yf Lebesgue integral 122Nb Lebesgue measurable function 121C, 121D, 134Xd, 225H, 262K, 262P, 262Yc Lebesgue measurable set 114E, 114F, 114G, 114Xe, 114Ye, 115E, 115F, 115G, 115Yc Lebesgue measure (on R) §114 (114E), 131Xb, 133Xc, 133Xd, 134G-134L, 212Xc, 216A, Chap. 22, 242Xi, 246Yd, 246Ye, 252N, 252O, 252Xf, 252Xg, 252Yj, 252Yp, §255, 342Xo, 343H, 344K, 345Xc, 346Xe, 372Xd, 382Xe, 383Q —– —– (on Rr ) §115 (115E), 132C, 132Ee, 133Yc, §134, 211M, 212Xd, 245Yj, 251M, 251Wi, 252Q, 252Xh, 252Yu, 254Xk, 255A, 255K, 255L, 255Xd, 255Yc, 255Yd, 256Ha, 256J-256L, 264H, 264I, 342Ja, 344K, 345B, 345D, 345Xf

654

Index

Lebesgue

—– —– (on [0, 1], [0, 1[) 211Q, 216A, 252Yq, 254K, 254Xh, 254Xj, 254Xk, 254Xl, 324Yg, 343Cb, 343J, 343Xd, 344K, 384Xj, 384Xk, 384Yd —– —– (on other subsets of Rr ) 242O, 244Hb, 244I, 244Yh, 246Yf, 246Yl, 251Q, 252Ym, 255M, 255N, 255O, 255Ye, 255Yh, 343Cc, 343H, 343M, 344J, 345Xd, 346Xb, 372Yj, 372Yk, 387Yf —– —– algebra 331Xd, 331Xe, 373C, 374C, 374D, 374Xa, 374Xf, 374Xh, 374Ya, 374Yd, 375Xe, 381Yd Lebesgue negligible set 114E, 115E, 134Yk Lebesgue outer measure 114C, 114D, 114Xc, 114Yd, 115C, 115D, 115Xb, 115Xd, 115Yb, 132C, 134A, 134D, 134Fa Lebesgue set of a function 223D, 223Xf, 223Xg, 223Xh, 223Yg, 261E, 261Ye Lebesgue-Stieltjes measure 114Xa, 114Xb, 114Yb, 114Yc, 114Yf, 131Xc, 132Xh, 134Xc, 211Xb, 212Xd, 212Xi, 225Xf, 232Xb, 232Yb, 235Xb, 235Xg, 235Xh, 252Xi, 256Xg, 271Xb, 224Yh, 342Xe left-translation-invariant see translation-invariant length of a curve 264Yl, 265Xd, 265Ya length of an interval 114Ab B.Levi’s theorem 123A, 123Xa, 133K, 135G, 135Hb, 226E, 242E, 365Df, 365Dh Levi property of a normed Riesz space 242Yb, 244Ye, 354Db, 354J, 354N, 354Xi, 354Xn, 354Xo, 354Yi, 356Da, 356L, 356Xk, 365C, 366D, 369G, 369Xe, 369Yg, 371C, 371D, 371Xc, 371Xd Lévy’s martingale convergence theorem 275I Lévy’s metric 274Ya, 285Yd lexicographic ordering 351Xa Liapounoff’s central limit theorem 274Xg Liapounoff’s vector measure theorem 326Ye lifting of a measure (algebra) Chap. 34 (341A, 341Ya), 363Xe, 363Yg; see also linear lifting (363Xe), translation-invariant lifting (345A) Lifting Theorem 341K, 363Yg limit of a filter 2A3Q, 2A3R, 2A3S limit of a sequence 2A3M, 2A3Sg limit ordinal 2A1Dd Lindeberg’s central limit theorem 274F-274H, 285Ym Lindeberg’s condition 274H linear functional 3A5D; see also positive linear functional linear lifting 341Xe, 341Ye, 345Yb, 363Xe, 363Yg linear operator 262Gc, 263A, 265B, 265C, §2A6, 355D, 3A5E; see also bounded linear operator (2A4F), (weakly) compact linear operator (3A5K), continuous linear operator, order-bounded linear operator (355A), positive linear operator (351F) linear order see totally ordered set (2A1Ac) linear space topology see linear topological space (2A5A), weak topology (2A5I), weak* topology (2A5I), Ts (U, V ) (3A5E) linear subspace (of a normed space) 2A4C; (of a partially ordered linear space) 351E, 351Rc linear topological space 245D, 284Ye, 2A5A, 2A5B, 2A5C, 2A5Eb, 2A5F, 2A5G, 2A5H, 2A5I, 367N, 367Yn, 367Yo, 393L, 393M, 393O-393Q, 393Yh, 393Yi, 3A4Ad, 3A4Bd, 3A4Fg linked set(in a Boolean algebra) 391L; (in κ-linked set, <κ-linked set) 391Yh; see also σ-linked (391L), σ-m-linked (391Yh) Lipschitz constant 262A, 262C, 264Yj Lipschitz function 225Yc, §262 (262A), 263F, 264Yj, 282Yb local convergence in measure see convergence in measure (245A) local semigroup see full local semigroup (394A) localizable measure algebra 322Ae, 322Be, 322C, 322K-322O, 323Gc, 323H, 323J, 323K, 323Xc, 323Yb, 323Z, 324Yf, 325C, 325D, 332B, 332C, 332J, 332O, 332S, 332T, 332Xo, 332Ya, 332Yc, 333H, 333Ia, 364Xk, 365J, 365K, 365S, 365T, 366Xb, 366Xf, 367Nc, 367V, 369A, 369B, 369D, 369E, 369R, 373L-373N, 373R, 373T, 373U, 375Yd, 382B-382E, 382Ga, 382Xa, 382Xb, 382Xd, 382Xg, 383M, 383N, 383Pa, 383Ya-383Yd, 391Cb, 394Xg, 394Xh, 394Yd, 395A; see also localization (322P)

Marczewski

General index

655

localizable measure algebra free product 325D, 325Ea, 325H, 325Xa, 325Xb, 325Xc, 327Ya, 333E-333H, 333K, 333Q, 333R, 334B, 334Xb, 376B, 376E, 376F, 376Xb, 376Yb; see also probability algebra free product (325K) localizable measure (space) 211G, 211L, 211Ya, 211Yb, 212G, 213Hb, 213L-213N, 213Xl, 213Xm, 214Id, 214J, 214Xa, 214Xd, 214Xf, 216C, 216E, 216Ya, 216Yb, 234F, 234G, 234Ye, 241G, 241Ya, 243G, 243H, 245Ec, 245Yf, 252Yq, 252Yr, 252Yt, 254U, 322Be, 322N, 325B, 342N, 342Yb, 363S, 364Xm; see also strictly localizable (211E) localization of a semi-finite measure algebra 322P, 322Yb, 332Yb, 365Xm, 366Xe, 366Xf, 369Xp locally compact measure (space) 342Ad, 342H, 342I, 342L-342N, 342Xc, 342Xd, 343B, 343Xa, 343Xg, 343Ya, 344H locally compact topological space 3A3Ag, 3A3B, 3A3G, 3A3Hb locally convex linear topological space 393Yi locally determined measure (space) 211H, 211L, 211Ya, 216Xb, 216Ya, 216Yb, 251Xc, 252Ya, 322N; see also complete locally determined measure locally determined negligible sets 213I, 213J-213L, 213Xk, 213Xl, 214Ib, 214Xg, 214Xh, 216Yb, 234Yb, 252Yb, 376Yk locally finite measure 256A, 256C, 256G, 256Xa, 256Ya locally integrable function 242Xi, 255Xh, 255Xi, 256E, 261Xa, 262Yg Loève, M. Chap. 27 intro. logistic map see quadratic map (372Xn) Loomis-Sikorski theorem 314M Lorentz norm 374Xj, 374Yb lower density on a measure space §341 (341C, 341Ya), 343Yc, §345, 346F, 346G, 346Xa, 346Xc, 346Xd, 346Ye, 346Yf, 346Zb, 363Xe; see also lower Lebesgue density (341E) lower integral 133I, 133J, 133Xe, 135H, 364Xm lower Lebesgue density 223Yf, 341E lower Riemann integral 134Ka lower semi-continuous function 225H, 225I, 225Xl, 225Xm, 225Yd, 225Ye, 323Cb, 367Xx Lusin’s theorem 134Yc, 256F Luxemburg, W.A.J. 363 notes magnitude of (an element in) a measure algebra 332Ga, 332J, 332O, 382Xc, 382Xd, 382Xg magnitude of a measure space 332Ga, 343Yb, 344I Maharam algebra see localizable measure algebra (322Ae) Maharam measure (space) see localizable (211G) Maharam submeasure 392G, 392H-392J, 392Xc, 392Xd, 392Xf, 393A-393C, 393E, 393M, 393Xc-393Xf, 393Ya, 393Yf, 393Yg Maharam’s theorem 331I, 332B Maharam type of a Boolean algebra 331F, 331H, 331Xc, 331Xg, 331Xh, 331Xj, 331Yd, 332H, 365Ya, 373Yb, 393Xe, 393Xf, 393Ya, 393Yf; see also relative Maharam type (333Aa) —– —– of a measure algebra 331I-331K, 331Xd-331Xf, 331Xi, 332M, 332N, 332R-332T, 333D, §334, 369Xg, 375Kb, 386K, 386L, 386Xc, 387N, 387O —– —– of a measure (space) 331Fc, 365Xp, 367Xr Maharam-type-homogeneous Boolean algebra 331Fb, 331H, 331Xh, 331Xj, 332A, 332H, 332Xa; see also homogeneous (331M) —– —– measure algebra 331I, 331K, 331L, 331N, 331Xd, 331Xe, 331Yj, 344Xe; see also relatively Maharam-type-homogeneous (333Ac) —– —– measure (space) 331Fc, 334Xe, 334Xf, 334Ya, 341Yc, 341Yd, 346E Maharam-type-κ component in a Boolean algebra 332Gb, 332H, 332J, 332O, 332P, 332Xj, 332Ya, 383N, 394Xh many involutions (group with many involutions) 381P, 381Q-381S, 381Xi-381Xk, 382G, 383A, 383C, 383D, 383I, 394Xc Marczewski functional 343E

656

Index

Markov

Markov time see stopping time (275L) Marriage Lemma see Hall’s Marriage Lemma (3A1K) martingale §275 (275A, 275Cc, 275Cd, 275Ce); see also reverse martingale martingale convergence theorems 275G-275I, 275K, 275Xf, 367K, 369Xq martingale difference sequence 276A, 276B, 276C, 276E, 276Xd, 276Ya, 276Yb, 276Ye, 276Yg martingale inequalities 275D, 275F, 275Xb, 275Yc-275Ye, 276Xb, 372 notes Max-flow Min-cut Theorem 332Xk maximal element in a partially ordered set 2A1Ab Maximal Ergodic Theorem 372C maximal theorems 275D, 275Yc, 275Yd, 276Xb, 286A, 286T, 372C, 372Yb Mazur S. see Banach-Mazur game McMillan see Shannon-McMillan-Breiman theorem (385G) meager set 314L, 314M, 314Yd, 316I, 316Yf, 316Yi, 341Yb, 3A3F, 3A3Ha mean (of a random variable) see expectation (271Ab) Mean Ergodic Theorem 372Xa —– see also convergence in mean (245Ib) measurable algebra 391B, 391C, 391D, 391K, 391Xf, 391Z, 392J, 393A, 393J, 393Ye, 394P, 394Q, 394Xf, 395B, 395Xd measurable cover see measurable envelope (132D) measurable envelope 132D, 132E, 132F, 132Xf, 132Xg, 132Xh, 134Fc, 134Xc, 213K-213M, 214G, 216Yc, 322I, 322J; see also full outer measure (132F) measurable envelope property 213Xl, 214Xl, 364Xm measurable function (taking values in R) §121 (121C), 122Ya, 212B, 212F, 213Yd, 214La, 214Ma, 235C, 235K, 252O, 252P, 256F, 256Yb, 256Yc, 316Yi, 322Yf; (taking values in Rr ) 121Yf), 256G; (taking values in other spaces) 133Da, 133E, 133Yb, 135E, 135Xd, 135Yf; (Σ, T)-measurable function 121Yb, 251Ya; see also Borel measurable, Lebesgue measurable measurable set 112A; µ-measurable set 212Cd; see also relatively measurable (121A) measurable space 111Bc measurable transformation §235, 365H; see also inverse-measure-preserving function measure 112A; see also control measure (393O), vector measure (393O) —– (in ‘µ measures E’, ‘E is measured by µ’) > 112Be measure algebra 211Yb, 211Yc, vol. 3 (321A); isomorphic measure algebras 331I, 331L, 332B, 332C, 332J, 332K, 332P, 332Q, 332Ya, 332Yb measure algebra of a measure space 321H-321K (321I), 322B, 322N, 327C; see also usual topology on a measure algebra (323A) measure-algebra topology §323 (323A), 324F, 324G, 324H, 324K, 324Xb, 324Xc, 327B, 327C, 365Ea, 367R, 367Xs, 367Yp measure-algebra uniformity 323A, 323B, 323D, 323G, 324F, 324H, 327B measure metric (on a measure algebra) 323A, 323Ca, 323Xg, 324Yh measure-preserving Boolean automorphism 332L, 333Gb, 333P-333R, 333Yc, 343Jc, 344C, 344E-344G, 344Xf, 372Xk, 372Xl, 372Yi, 373Yb, 374E, 374G, 374L, 374Yc, 382D, 382Ye, 384P, 384Sb, 384Sf, 384Xg, 384Yd, 385P, 386B-386H, 386J, 386L, 387C, 387E, 387F, 387H-387K, 387M-387O, 387Xa-387Xe, 387Yf, 395Xc; see also automorphism group of a measure algebra, two-sided Bernoulli shift (384Qb) measure-preserving Boolean homomorphism 324I, 324J-324P, 324Xd, 324Xe, 324Yf, 324Yh, 325C, 325D, 325I, 325J, 325Xd, 327Xa, 331D, 332L-332O, 332Q, 332Xm, 332Xn, 333C, 333D, 333F, 333Gb, 333Xc, 333Yd, 343B, 366Xf, 369Xm, 369Xn, 372G, 372H, 372Xc, 373Bd, 373U, 373Xi, 382Ye, 384L-384Q, 384T, 384V, 384Xf-384Xi, 384Xm, 384Xp, 384Yb-384Ye, 385D-385H, 385N, 385P, 385Xe, 385Xf, 385Yb-385Yd, 386A, 386B, 386F; isomorphic measure-preserving Boolean homomorphisms 384Ta; see also automorphism of a measure algebra, Bernoulli shift (384Q), measure-preserving ring homomorphism measure-preserving function see inverse-measure-preserving function (235G), measure space automorphism, measure space isomorphism measure-preserving ring automorphism 366L, 366Xh measure-preserving ring homomorphism 361Ad, 365O, 365Q, 366H, 366K, 366L, 366Xg, 366Xh, 366Yg, 369Xl, 372F; see also measure-preserving Boolean homomorphism

one-point

General index

657

measure space §112 (112A), 113C, 113Yi measure space automorphism 255A, 255Ca, 255N, 255Ya, 255 notes, 344C, 344E-344G, 344Xf, 345Ab, 345Xa, 346Xd, 384Sb measure space isomorphism 254K, 254Xj, 254Xk, 254Xl, 255Ca, 255Mb, 344I, 344J, 344K, 344L, 344Xa344Xc, 344Xe-344Xg, 344Yc metric 2A3F, 2A4Fb; see also entropy metric (384Xq), Euclidean metric (2A3Fb), Hausdorff metric (246Yb), Lévy’s metric (274Ya), measure metric (323Ad), pseudometric (2A3F) metric outer measure 264Xb, 264Yc metric space 224Ye, 261Yi, 316Yj, 3A4Fe, 3A4Ff, 3A4Hc; see also complete metric space, metrizable space (2A3Ff) metrically transitive transformation see ergodic inverse-measure-preserving function (372Pb) metrizable (topological) space 2A3Ff, 2A3L; see also metric space, separable metrizable space —– uniformity 3A4Bc mid-convex function 233Ya minimal element in a partially ordered set 2A1Ab mixing inverse-measure-preserving function 372P, 372Q, 372R, 372Xk, 372Xm-372Xo, 372Xq, 372Xr, 372Xu, 372Xv, 372Yh mixing measure-preserving homomorphism 333P, 333Yc, 372P, 372Q, 372R, 372Xk, 372Xl, 372Xp, 384Se, 386Xb, 387Ye modified Dirichlet kernel 282Xc modulation 286C Monotone Class Theorem 136B, 312Xc, 313G, 313Xd Monotone Convergence Theorem see B.Levi’s theorem (133A) monotonic function 121D, 222A, 222C, 222Yb, 224D, 313Xh Monte Carlo integration 273J, 273Ya multilinear map 253Xc multiplicative identity (in an f -algebra) 353P, 353Yg multiplicative Riesz homomorphism 352Xl, 353Pd, 361J, 363F, 364R Nachbin-Kelley theorem 363R, 363Yj natural extension (of an inverse-measure-preserving function) 382Ye negligible set 112D, 131Ca, 214Cb, 344H, 344Yf Neumann, J. von 341 notes non-decreasing sequence of sets 112Ce non-increasing sequence of sets 112Cf non-measurable set 134B, 134D, 134Xg norm 2A4B; (of a linear operator) 2A4F, 2A4G, 2A4I; (of a matrix) 262H, 262Ya; (norm topology) 242Xg, 2A4Bb; see also Fatou norm (354Da), Riesz norm (354Aa) normal density function 274A, 283N, 283We, 283Wf normal distribution function 274Aa, 274F-274K, 274M, 274Xe, 274Xg normal random variable 274A, 274B, 285E, 285Xm, 285Xn normal subgroup 381S, 381Xi, 381Xk, 382I, 382Xd, 382Yb; see also lattice of normal subgroups normal subspace (of a Riesz space) see band (352O) normalized Hausdorff measure 264 notes, §265 (265A) normed algebra 2A4J, 3A5J normed space 224Yf, §2A4 (2A4Ba); see also Banach space (2A4D) Novikov’s Separation Theorem see Second Separation Theorem nowhere dense set 313R, 313Yb, 315Yb, 316I, 316Yc, 316Yf, 316Yh, 3A3F nowhere rigid Boolean algebra 383H, 383I-383L, 383Xa-383Xd null set see negligible (112Da) odd function 255Xb, 283Yd Ogasawara’s representation theorem for Riesz spaces 368 notes one-point compactification 311Ya, 3A3O

658

Index

open

open interval 111Xb, 114G, 115G, 1A1A, 2A2I open map 313R open set 111Gc, 111Yc, 114Yd, 115G, 115Yb, 133Xb, 134Fa, 134Yj, 135Xa, 1A2A, 1A2B, 1A2D, 256Ye, 2A3A, 2A3G; (in R) 111Ye, 114G, 134Xc, 2A2I; see also topology (2A3A) operator topology see very weak operator topology (373K) optional time see stopping time (275L) orbit of a bijection 387C, 387Xa, 387Xb, 387Yb, 387Yc order (in a Boolean ring) 311H order-bounded dual Riesz space (U ∼ ) §356 (356Aa), 362A, 363K, it 365I, 365J, 366D, 366Ya, 368P, 371Xe, 375Yb order-bounded set (in a partially ordered space) 2A1Ab, 354Xd, 368Xe order-bounded linear operator (between Riesz spaces) §355 (355A); see also order-bounded dual (356A), L∼ (355A) order-closed set in a partially ordered space 313Da, 313Fa, 313Id, 313Xb, 313Xc, 314Ya, 316Fb, 323D, 352Oa, 354Xp; (ideal in a Boolean algebra) 313Eb, 313Pa, 313Qa, 313Xo, 314Xb, 314Yh, 316Xh, 316Xi, 316Xn, 316Xu; (Riesz subspace of a Riesz space) 353J, 354Xp, 375C; (subalgebra of a Boolean algebra) 313E-313G, 314Xg, 314Xh, 314Yc, 315Xm, 322Md, 322Xh, 323H-323J, 323Yc, 331E, 331G, 331Yb, 331Yc, 333Bc, 393Xd; see also sequentially order-closed (313Db) order-complete see Dedekind complete (241Ec, 314Aa) order-continuous bidual Riesz space (U ×× ) 356I, 356J, 369D order-continuous Boolean homomorphism 313L-313N, 313Xl-313Xn, 313Xp, 313Yf, 314Fa, 314G, 314R, 314T, 315Jc, 315P, 315Ya, 316Fd, 316Xy, 322Yd, 324E-324G, 324K, 324Ya, 324Ye, 324Yf, 325Aa, 325C, 325D, 325H, 326Kf, 331H, 331J, 332Xm, 332Xn, 334Xb, 343B, 344A, 344C, 344E, 344F, 352Xh, 363Ff, 364Rc, 364Yh, 366Xf, 369Xm, 369Xn, 373Bd, 373U, 375Ya, 385C, 395Xa order-continuous dual Riesz space (U × ) §356 (356Ac), 362A, 363K, 363S, 365I, 365J, 366D, 366Ya, 366Yc, 367Xf, 368Pc, 368Yj, 369A, 369C, 369D, 369K, 369Q, 369Xa, 369Xe, 369Yh, 371Xe, 375B, 375J, 375Yd order-continuous norm (on a Riesz space) 242Yc, 242Ye, 244Yd, 313Yd, 326Yj, 354Dc, 354E, 354N, 354Xi, 354Xj, 354Xl, 354Xm, 354Xo, 354Xp, 354Yc-354Yh, 355K, 355Yg, 356D, 356M, 356Yg, 365C, 366D, 367E, 367Xu, 367Yn, 367Yr, 369B, 369Xf, 369Xg, 369Yc-369Yf, 371Xa-371Xc, 371Ya, 374Xe, 374Xg, 374Yb, 376L, 376M order-continuous order-preserving function 313Ha, 313I, 313Xi, 313Yc, 313Yd, 315D, 315Yg, 316Fc, 326Kc, 361Cf, 361Gb, 363Eb, 363Ff, 366Yd, 367Xb, 367Yc, 384Ya, 385Yc, 385Yd, 392Xb, 394N; see also sequentially order-continuous order-continuous positive linear operator 351Ga, 355G, 355H, 355K, 364Rc, 366H, 368I, 375Xa, 375Xb, 375Yd; see also order-continuous dual (356A), order-continuous Riesz homomorphism, L× (355G) order-continuous Riesz homomorphism 327Yb, 351Xc, 352N, 352Oe, 352Rb, 352Ub, 352Xd, 353O, 353Xc, 356I, 361Je, 355F, 366H, 367Xh, 368B, 375C, 375Xc, 375Ya order-continuous ring homomorphism (between Boolean rings) 361A, 361Je, 365P >367A), 368Yg, 368Yh, order*-convergent sequence in a partially ordered set 245Xc, 356Xd, §367 (> 369Xq, 376G, 376H, 376Xe, 376Xg, 376Ye, 376Yh, 393Xa, 393Yd, 393Ye; (in C(X)) 367L, 367Yi, 367Yj, 367Yk; (in L0 (µ)) 245C, 245K, 245L, 245Xc, 245Xd, 376J; (in L0 (A)) §367, 368Yi, 372D-372G, 372Xb, 385G, 385H order-convex set in a partially ordered set 351Xb order-dense set in a Boolean algebra 313J, 313K, 313Xm, 314Yg, 332A order-dense Riesz subspace of a Riesz space 352N, 353A, 353D, 354Ef, 354I, 354Ya, 355F, 355J, 356I, 363C, 363Xd, 364L, 365F, 365G, 367Yf, 368B-368E, 368G-368I, 368M, 368P, 368S, 368Ya, 368Ye, 369A, 369B, 369C, 369D, 369G, 375D, 375Xc; see also quasi-order-dense (352Na) order-dense subalgebra of a Boolean algebra 313O, 313Xj, 313Xn, 313Ye, 314I, 314T, 314Xf, 316Xf, 316Xo, 316Xr, 316Yk, 316Yn, 323Dc, 363Xd, 391Xj order-isomorphism 312L order*-limit see order*-convergent (367A) order-preserving function 313H, 313I, 313La, 315D, 324Yg, 326Bf, 361Ce order topology (on a partially ordered set) 313Xb, 313Xj, 313Yc, 326Yr, 367Yl, 393Xa, 393Yd

porous

General index

659

order unit (in a Riesz space) 243C, 353L, 353M, 363N, 368Ya; see also order-unit norm (354Ga), standard order unit (354Gc), weak order unit (353L) order-unit norm 354F, 354G (354Ga), 354I-354K, 354Yi, 353Yf, 355Xc, 356N, 356O; see also M -space (354Gb) ordered set see partially ordered set (2A1Aa), totally ordered set (2A1Ac), well-ordered set (2A1Ae) ordinal 2A1C, 2A1D-2A1F, 2A1K ordinate set 252N, 252Yg, 252Yh Orlicz norm 369Xd, 369Xi, 369Xj, 369Xn, 369Yc, 369Yd, 369Ye, 369Yg, 374Xc Orlicz space 369Xd, 369Xi, 369Xj, 373Xm Ornstein’s theorem 386I, 386K orthogonal matrix 2A6B, 2A6C orthogonal projection in Hilbert space 244Nb, 244Yj, 244Yk, 366K orthonormal vectors 2A6B outer automorphism of a group 383G, 383O, 383Pb, 383Q, 383Yc, 383Yd, 3A6B outer measure §113 (113A), 114Xd, 132B, 132Xh, 136Ya, 212Ea, 212Xa, 212Xb, 212Xg, 213C, 213Xa, 213Xg, 213Xk, 213Ya, 251B, 251Wa, 251Xd, 254B, 254G, 254L, 254S, 254Xq, 254Yd, 264B, 264Xa, 264Ya, 264Yo, 393T, 393Xh, 393Yh; see also Lebesgue outer measure (114C, 115C), metric outer measure (264Yc), regular outer measure (132Xa), submeasure (392A) —– —– defined from a measure 113Ya, §132 (132B), 213C, 213F, 213Xa, 213Xg-213Xi, 213Xk, 213Yd, 214Cd, 215Yc, 251O, 251R, 251Wk, 251Wm, 251Xm, 251Xo, 252Yh, 254Xb, 264F, 264Yd outer regular measure 256Xi Parseval’s identity 284Qd partial derivative 123D, 252Yj, 262I, 262J, 262Xh, 262Yb, 262Yc partial lower density on a measure space 341Dc, 341N partial order see partially ordered set (2A1Aa) partially ordered linear space 231Yc, 241E, 241Yg, 326C, §351 (351A), 355Xa, 361C, 361G, 362Aa; see also Riesz space (352A) partially ordered set 2A1Aa, 313D, 313F, 313H, 313I, 313Xb, 313Xg, 313Xh, 313Yc, 315C, 315D, 315Xb, 315Yg partition of unity in a Boolean ring or algebra 311G, 313K, 313L, 313Xk, 315E, 315F, 315Xk, 315Xq, 316Xp, 322E, 332E, 332I, 332Xi, 332Yb, 352T, 375H, 375I, 381B, 381C, 381J, 382F, 384C, 384D, 384G384P, 384R, 384Xa-384Xe, 384Xi, 384Xq, 384Ya-384Yc, 384Yg, 385B, 385C, 385G, 385K, 385L, 385N, 385O, 385Q, 385Xd, 385Xe, 387Lb; see also Bernoulli partition (386A) pathological submeasure 393Tc Peano curve 134Yl, 134Ym, 134Yn, 134Yo perfect measure (space) 342K, 342L, 342M, 342Xh-342Xo, 343K-343M, 343Xg, 343Xh, 344C, 344I, 344Xb-344Xd, 344Yf, 384V perfect Riesz space 356J, 356K-356M, 356P, 356Xg, 356Xi-356Xk, 356Ye, 356Yg, 365C, 365K, 366D, 369D, 369K, 369Q, 369Ya, 374M, 374Xk, 374Xl, 374Ya, 376P, 376Xl periodic Boolean homomorphism of a Boolean algebra 385A, 385B, 385C, 385Xa periodic extension of a function on ]−π, π] 282Ae permutation group see symmetric group Plancherel Theorem (on Fourier series and transforms of square-integrable functions) 282K, 284O, 284Qd pointwise convergence (topology on a space of functions) 281Yf pointwise convergent see order*-convergent (245Cb, 367A) pointwise topology see pointwise convergence Poisson distribution 285Q, 285Xo Poisson’s theorem 285Q polar coordinates 263G, 263Xf Pólya’s urn scheme 275Xc polynomial (on Rr ) 252Yu porous set 223Ye, 261Yg, 262L

660

Index

positive

positive cone 253G, 253Xi, 253Yd, 351C positive definite function 283Xt, 285Xr positive linear operator (between partially ordered linear spaces) 351F, 355B, 355E, 355K, 355Xa, 361G, 367P, 375I-375K, 375Xa, 375Xe, 375Xg, 376Cc; see also (sequentially) order-continuous positive linear operator, Riesz homomorphism (351H) predictable sequence 276Ec presque partout 112De primitive product measure 251C, 251E, 251F, 251H, 251K, 251Wa, 251Xa-251Xc, 251Xe, 251Xf, 251Xj, 251Xl-251Xp, 252Yc, 252Yd, 252Yg, 253Ya-253Yc, 253Yg, 325Ya principal band (in a Riesz space) 352V, 353C, 353H, 353Xb, 362Yd principal ideal in a Boolean ring or algebra 312D, 312E, 312J, 312S, 312Xf, 312Yf, 314E, 314Xb, 315E, 315Xl, 316Xb, 316Xk, 316Xv, 321Xa, 322H-322J, 322Xf, 323Xe, 325Xb, 331Fb, 331H, 331M, 332A, 332L, 332P, 332Xh, 352Sc, 352Xf, 364Xo, 383Lb, 391Xe, 391Xj principal projection property (of a Riesz space) 353Xb, 353Yd, 354Yi principal ultrafilter 2A1N probability algebra 322Aa, 322Ba, 322Ca, 322G, 322Mb, §384, 385G, 385J-385L, 385N-385Q, 385Xe, 385Yc, 385Yd, §386, 391B probability algebra free product 325F, 325G, 325I-325M (325K) 325Xd, 334D, 334Xc, 384Xm, 387Ye probability density function see density (271H) probability space 211B, 211L, 211Q, 211Xb, 211Xc, 211Xd, 212G, 213Ha, 215B, 243Xi, 253H, 253Xh, §254, Chap. 27, 322Ba product Boolean algebra see simple product (315A), free product (315H) product f -algebra 352Wc, 364S product measure Chap. 25; see also c.l.d. product measure (251F, 251W), primitive product measure (251C), product probability measure (254C) product partial order 315C, 315D, 315Xb product partially ordered linear space 351L, 351Rd, 351Xc, 351Xd; see also product Riesz space product probability measure §254 (254C), 272G, 272J, 272M, 275J, 275Yi, 275Yj, 281Yk, 325I, 334C, 334E, 334Xf, 334Xg, 334Ya, 342Gf, 342Xn, 343H, §346, 372Xf, 384S product Riesz space 352K, 352T, 354Xb product ring 3A2H product topology 281Yc, 2A3T, 315Xf, 315Yb, 315Yd, 323L, 391Yc, 3A3I, 3A3J, 3A3K product uniformity 3A4E —– see also inner product space projection (on a Riesz space) see band projection (352Rb) projection band (in a Riesz space) 352R (352Ra), 352S, 352T, 352Xg, 353E, 353H, 353I, 353Xb, 361Xh, 361Ye, 362B projection band algebra 352S, 352T, 361K, 363J, 363N, 363O, 364Q pseudo-cycle (for a Boolean automorphism) 387Xc pseudometric 2A3F, 2A3G, 2A3H, 2A3I, 2A3J, 2A3K, 2A3L, 2A3Mc, 2A3S, 2A3T, 2A3Ub, 2A5B, 323A, 3A3Be, 3A4B, 3A4D-3A4F pseudo-simple function 122Ye, 133Ye pull-back measures 132G purely atomic Boolean algebra 316Lc, 316M, 316Xr-316Xx, 316Yk, 316Yl, 316Yq, 322Bh, 322Kc, 324Kg, 331Yd, 332Xb, 332Xc, 362Xf, 384J, 384Xq, 394Xf purely atomic measure (space) 211K, 211N, 211R, 211Xb, 211Xc, 211Xd, 212G, 213He, 214Xe, 234Xb, 251Xr, 322Bh, 342Xn, 375Ya purely infinite measurable set 213 notes quadratic map 372Xn quasi-homogeneous measure algebra 374G, 374H-374M, 374Xl, 374Ye, 394Xg quasi-order-dense Riesz subspace (of a Riesz space) 352N, 353A, 353K, 353Ya quasi-Radon measure (space) 256Ya, 263Ya

representation

General index

661

quasi-simple function 122Yd, 133Yd quasi-Stonian topological space 314Yf quotient Boolean algebra (or ring) 312K, 312Xk, 313Q, 313Xo, 314C, 314D, 314M, 314N, 314Yd, 316C, 316D, 316Xh, 316Xi, 316Xn, 316Xu, 316Ye, 316Yp, 361M, 363Gb, 391Xe, 391Yb, 392Xd quotient partially ordered linear space 241Yg, 351J, 351K, 351Xb; see also quotient Riesz space quotient Riesz space 241Yg, 241Yh, 242Yg, 352Jb, 352U, 354Yi, 361M, 364C, 364U quotient ring 3A2F, 3A2G quotient topology 245Ba Rademacher’s theorem 262Q Radon measure(on R or Rr ) §256 (256A), 284R, 284Yi, 342Jb, 342Xj; (on ]−π, π]) 257Yb Radon-Nikod´ ym derivative 232Hf, 232Yj, 234B, 234Ca, 234Yd, 234Yf, 235Xi, 256J, 257F, 257Xe, 257Xf, 272Xe, 272Yc, 275Ya, 275Yi, 285Dd, 285Xe, 285Ya, 363S Radon-Nikod´ ym theorem 232E, 232F, 232G, 234G, 235Xk, 242I, 244Yk, 327D, 365E, 365Xf Radon probability measure (on R or Rr ) 271B, 271C, 271Xb, 285Aa, 285M, 364Xd, 372Xf; (on other spaces) 256Ye, 271Ya Radon product measure (of finitely many spaces) 256K random variable 271Aa rapidly decreasing test function §284 (284A, 284Wa), 285Dc, 285Xd, 285Ya rare see nowhere dense (3A3Fa) rearrangement 373Ya; see also decreasing rearrangement (252Yp, 373C) rearrangement-invariant extended Fatou norm 374Eb, 374F, 374K rearrangement-invariant set 374Ea, 374F, 374M, 374Xk, 374Xl, 374Ye Recurrence Theorem (for a measure-preserving Boolean homomorphism) 382Yc, 382Yd, 385D recursion 2A1B reduced power of R 351M, 351Q, 351Yd, 352L, 352M, 368F refine 311Gd reflexive Banach space 372A, 3A5G regular embedding of Boolean algebras 313N regular measure see inner regular (256Ac) regular open algebra (of a topological space) 314P-314S (314Q), 314Yd, 315Xe, 315Yd, 316Xe, 316Yb, 316Yd, 316Yf, 316Yj, 316Yq, 332Xe, 363Yc, 364U-364W, 364Yh-364Yl, 368Ya, 391Yc, 391Yf; (of R) 316K, 316Yb, 316Yn, 316Yo, 331Xh, 331Yi, 367Yq, 375Xe, 375Yf, 375Z, 376Yi, 391Xd regular open set 314O, 314P, 314Q regular operator (between Riesz spaces) 355 notes regular outer measure 132C, 132Xa, 214Hb, 251Xm, 254Xb, 264Fb regular topological space 2A5J, 316Yf, > 3A3Ab, 3A3Ba, 3A3De; see also completely regular (3A3Ac) regularly embedded (subalgebra of a Boolean algebra) 313N, 313O, 313P, 316Xm, 316Xs, 326Kf; (Riesz subspace) 352Ne, 352Xd, 354Xk, 354Xm, 367F, 368Pa, 368S relation 1A1B relative atom in a Boolean algebra 331A relative Maharam type of a Boolean algebra over a subalgebra 333A, 333B, 333C, 333E, 333F, 333Yb relatively atomless (Boolean algebra) 331A, 385E, 387L, 394Xc relatively compact set 2A3Na, 2A3Ob, 3A3De relatively Maharam-type-homogeneous 333Ac, 333Bb relatively measurable set 121A relatively von Neumann transformation 387Ga, 387N, 387Xg, 387Yd, 387Ye relatively weakly compact set (in a normed space) 247C, 2A5I, 356Q, 356Xl, 3A5Gb, 3A5Hb, 3A5Kb; (in other linear spaces) 376O, 376P, 376Xl repeated integral §252 (252A); see also Fubini’s theorem, Tonelli’s theorem representation of homomorphisms (between Boolean algebras) 344Ya, 344Yd, 364Xr; (between measure algebras) 324A, 324B, 343A, 343B, 343G, 343J, 343M, 343Xc, 343Xf, 343Yd, 344A-344C, 344E-344G, 344Xf, 344Yc

662

Index

respects

‘respects coordinates’ (said of a lifting) 346A, 346C, 346E, 346Xg, 346Yc, 346Za reverse martingale 275K Riemann integrable function 134K, 134L, 281Yh, 281Yi Riemann integral 134K, 242 notes, 363Yi Riemann-Lebesgue lemma 282E Riesz Convexity Theorem 244 notes Riesz homomorphism (between partially ordered linear spaces) 351H, 351J, 351L, 351Q, 351Xc, 351Ya, 352G; (between Riesz spaces) 352G-352J, 352W, 352Xb, 352Xd, 353Pd, 353Yf, 355Xe, 356Xh, 361Gc, 361J, 361Xg, 362Xe, 363Ec, 363F, 363Xb, 363Xc, 364R, 365N, 375I, 375J, 375Xg, 375Ya, 376Cc; see also ordercontinuous Riesz homomorphism Riesz norm 242Xg, 354A, 354B, 354D, 354F, 354M, 354Xc-354Xf, 354Xh, 354Yb, 354Yf, 354Yk, 355Xc, 356D, 356Xg, 356Xh; see also Fatou norm (354Da), order-continuous norm (354Dc), order-unit norm (354Ga) Riesz space (= vector lattice) 241Ed, 241F, 241Yc, 241Yg, chap. 35 (352A), 361Gc, 367C, 367E, 367Xc, 367Xg, 367Yo; see also Archimedean Riesz space (§352), Banach lattice (354Ab), Riesz norm (354A) Riesz subspace (of a partially ordered linear space) 352I; (of a Riesz space) 352I, 352J, 352L, 352M, 353A, 354O, 354Rc; see also band (352O), order-dense Riesz subspace (352N), solid linear subspace (351I) rigid Boolean algebra 383Ha, 383L; see also nowhere rigid (383Hb) ring §3A2 (3A2A); see also Boolean ring (311Aa) ring homomorphism 3A2D, 3A2F-3A2H ring homomorphism between Boolean rings 311D, 312Xf, 312Xg, 312Xh, 312Yc, 312Yd, 312Ye, 312Yf, 361A, 361Cc, 361J, 361Xe, 361Xg, 375H Rokhlin see Halmos-Rokhlin-Kakutani lemma (385E) root algebra (of a Bernoulli shift) 384Q, 384R, 384S, 386B, 386Ya Saks see Vitali-Hahn-Saks theorem (246Yg) saltus function 226B, 226Db, 226Xa Schröder-Bernstein theorem 2A1G, 332 notes, 344D, 344Xa Schwartz function see rapidly decreasing test function (284A) Schwartzian distribution 284R, 284 notes; see also tempered distribution self-ing set (in a topological measure space) 256Xf semi-continuous function see lower semi-continuous (225H) semi-finite measure algebra §322 (322Ad), 323Dd, 323Ga, 323Xa, 324K, 324Xb, 325Ae, 325D, 327B, 331C, 332E, 332F, 332I, 332R, 332Xi, 332Yb, 364L, 365E, 365G, 365J, 365P, 365S, 366E, 366Xe, 366Xf, 366Xk, 367Nb, 368S, 369H, 369Xa, 371Xc, 373R, 375I, 382E, 383Ld, 383P, 383Xf, 391Ca semi-finite measure (space) 211F, 211L, 211Ya, 212G, 213A, 213B, 213Hc, 213Xc, 213Xd, 213Xj, 213Xl, 213Xm, 213Ya-213Yc 214Xe, 214Xh, 215B, 216Xa, 216Yb, 234F, 235O, 235Xd, 235Xe, 241G, 241Ya, 241Yd, 243G, 245Ea, 245J, 245Xd, 245Xj, 245Xl, 246J, 246Xh, 251J, 251Xc, 252P, 252Yf, 253Xf, 253Xg, 322Bd, 322Yd, 327C, 327D, 342L, 342Xa, 342Xc, 342Xn, 343B, 344H, 365Xp, 367Xr semi-finite version of a measure 213Xc, 213Xd, 322Xb semigroup see topological semigroup semi-martingale see submartingale (275Yf) seminorm 2A5D semi-ring of sets 115Ye separable (topological) space 2A3Ud, 316Xd, 316Yb, 316Yj, 367Xr, 391Yc, 3A3E separable Banach space 244I, 254Yc, 365Xp, 366Xc, 369Xg separable metrizable space 245Yj, 264Yb, 284Ye sequentially order-closed set in a partially ordered space 313Db, 313Xg, 313Yc, 316Fb, 353Ja, 367Yc; see also σ-ideal (313Ec), σ-subalgebra (313Ec) sequentially order-continuous additive function (on a Boolean algebra) 326Gc, 363Eb —– —– Boolean homomorphism 313Lc, 313Pb, 313Qb, 313Yb, 314Fb, 314Hb, 314Xc, 314Ye, 315Ya, 316Fd, 324A, 324B, 324Kd, 324Xa, 324Xe, 324Yc, 326Ff, 343Ab, 363Ff, 364G, 364H, 364R, 364Xp, 364Yc, 364Yf, 364Yn, 365H, 365Xg, 375Ya, 375Ye

strictly

General index

663

—– —– dual (of a Riesz space)(Uc∼ ) 356Ab, 356B, 356D, 356L, 356Xa, 356Xb, 356Xc, 356Xd, 356Xf, 356Ya, 362Ac, 363K, 363S —– —– function 313Hb, 313Ic, 313Xg, 313Yc, 315D, 316Fc, 361Cf, 361Gb, 367Xb, 367Yc, 375Xd, 392H —– —– positive linear operator or functional 351Gb, 355G, 355I, 361Gb, 363Eb, 363Ff, 364R, 375A; see also sequentially order-continuous dual (356A), L∼ c (355G) —– —– Riesz homomorphism 361Jf, 375Ya —– —– ring homomorphism between Boolean rings 361Ac, 361Jf, 375Ka Shannon-McMillan-Breiman theorem 385G, 385Xc shift operators (on function spaces based on topological groups) 286C; see also Bernoulli shift (384Q) Shoenfield’s Absoluteness Theorem §393 notes Sierpi´ nski Class Theorem see Monotone Class Theorem (136B) signed measure see countably additive functional (231C) >122A), 242M, 361D simple function §122 (> simple group 381T, 382I, 382Xa, 382Xb, 382Xc simple product of Boolean algebras §315 (315A), 316Xc, 316Xl, 316Xw, 332Xa, 332Xg, 364S, 381Yd, 391Xb, 391Xj simple product of measure algebras 322K, 323L, 325Xc, 332B, 332Xm, 332Xn, 333H, 333Ia, 333K, 333R, 366Xi, 383Lc Sinaˇı’s theorem 386E, 386L singular additive functional 232Ac, 232I, 232Yg smooth function (on R or Rr ) 242Xi, 255Xi, 262Yd, 262Ye, 262Yf, 262Yg, 284A, 284Wa smoothing by convolution 261Ye solid hull (of a subset of a Riesz space) 247Xa, 352Ja solid set (in a partially ordered linear space) 351I; (in a Riesz space) 354Xg —– linear subspace (of a partially ordered linear space) 351J, 351K, 351Yb; (of a Riesz space) 352J, 353J, 353K, 353N, 355F, 355J, 355Yj, 368Ye, 382J; see also band (352O) Souslin property see ccc (316Ab) space-filling curve 134Yl spectrum (of an M -space) 354L sphere, surface measure on 265F-265H, 265Xa-265Xc, 265Xe spherical polar coordinates 263Xf, 265F split interval (= ‘double arrow space’) 343J, 343Xf, 343Yc, 344Xf square-integrable function 244Na; see also L2 standard extension of a countably additive functional 327F, 327G, 327Xa, 327Xd, 327Xe, 327Yc standard normal distribution, standard normal random variable 274A standard order unit (in an M -space) 354Gc, 354H, 354L, 356N, 356P, 363Ba, 363Ye Steiner symmetrization 264H, 264 notes step-function 226Xa Stieltjes measure see Lebesgue-Stieltjes measure (114Xa) Stirling’s formula 252Yn stochastic see doubly stochastic matrix stochastically independent see independent (272A, 325L, 325Xe) Stone representation of a Boolean ring or algebra 311E, 352 notes; see also Stone space Stone space of a Boolean ring or algebra 311E, 311F, 311I-311K, 311Xg, 311Ya, 312O-312S, 312Xi312Xk, 312Yc-312Yf, 313C, 313R, 313Xp, 313Yb, 314M, 314S, 314T, 314Yd, 315H, 315Xc, 316B, 316I, 316Yb, 316Yc, 316Yi, 363A, 363Yf, 387Yb Stone space of a measure algebra 321J, 321K, 322N, 322Q, 322Xi, 322Yb, 322Yf, 341O, 341P, 342Jc, 343B, 344A, 344Xe, 346K, 346L, 346Xf Stone-Weierstrass theorem 281A, 281E, 281G, 281Ya, 281Yg stopping time 275L, 275M-275O, 275Xi, 275Xj strictly localizable measure (space) 211E, 211L, 211N, 211Ye, 212G, 213Ha, 213J, 213L, 213O, 213Xh, 213Xn, 213Ye, 214Ia, 214J, 215Xf, 216E, 234F, 235P, 251N, 251P, 251Xn, 252B, 252D, 252E, 252Ys, 252Yt, 322Kd, 322N, 322Qb, 322Xi, 325He, 341H, 341K, 342Hb, 343B, 344C, 344I, 344Xb, 344Xc, 346Xd, 384V

664

Index

strictly

strictly positive additive functional (on a Boolean algebra) 391D, 391J, 391N, 395Xd see also chargeable Boolean algebra (391X) —– —– submeasure (on a Boolean algebra) 392B, 392F, 392I, 392J, 392Xc, 393Xa, 393Xb, 393Yd strong law of large numbers 273D, 273H, 273I, 273Xh, 275Yn, 276C, 276F, 276Ye, 276Yg strongly mixing see mixing (372P) subalgebra of a Boolean algebra 312A, 312B, 312M, 312N, 312Xb, 312Xc, 312Xj, 313Fc, 313G, 313Xd, 313Xe, 315Xn, 315Xp, 316Xa, 331E, 331G, 332Xf, 363G, 391Xb, 391Xf, 391Xj; see also Boolean-independent subalgebras (315Xn), closed subalgebra (323I), fixed-point subalgebra (394Ga), independent subalgebras (325L), order-closed subalgebra, order-dense subalgebra, regularly embedded subalgebra (313N), σsubalgebra (233A, 313E) subhomomorphism see σ-subhomomorphism (375E) sublattice 3A1Ib submartingale 275Yf, 275Yg submeasure (on a Boolean algebra) §392 (392A), 393B, 393T, 393U, 393Yb; see also Maharam submeasure (392G), outer measure (113A), pathological submeasure (393Tc) subring 3A2C —– of a Boolean ring 311Xd, 312Xa, 312Xg subspace measure 113Yb, 214A, 214B, 214C, 214H, 214I, 214Xb-214Xh, 216Xa, 216Xb, 241Ye, 242Yf, 243Ya, 244Yc, 245Yb, 251P, 251Q, 251Wl, 251Xn, 251Yb, 254L, 254Ye, 264Yf, 322I, 322J, 322Xg, 322Yd, 343H, 343M, 343Xa; (on a measurable subset) 131A, 131B, 131C, 132Xb, 214J, 214K, 214Xa, 214Xi, 241Yf, 247A, 342Ga, 342Ia, 342Xn, 343L, 344J, 344Xa, 344Xe, 344Xf; (integration with respect to a subspace measure) 131D, 131E-131H, 131Xa-131Xc, 133Dc, 133Xa, 214D, 214E-214G, 214M subspace of a normed space 2A4C subspace topology 2A3C, 2A3J, 3A4Db subspace uniformity 3A4D subspace σ-algebra 121A, 214Ce substitution see integration by substitution successor cardinal 2A1Fc —– ordinal 2A1Dd sum over arbitrary index set 112Bd, 226A sum of measures 112Xe, 112Ya, 212Xe, 212Xh, 212Xi, 212Xj, 212Yd, 212Ye, 334Xb, 334Xd summable family of real numbers 226A, 226Xf of an additive functional on a Boolean algebra 326Xi of an automorphism of a Boolean algebra 381L, 381R, 381S, 381Yb, 383B of a measure (on a topological space) 256Xf, 257Xd of a submeasure on a Boolean algebra 392Xe see also bounded , compact ing (element in a Boolean algebra ing an automorphism) 381D, 381E, 383B; see also self-ing set (256Xf), supremum 2A1Ab —– in a Boolean algebra 313A-313C surface measure see normalized Hausdorff measure (265A) symmetric distribution 272Ye symmetric difference (in a Boolean algebra) 311G symmetrization see Steiner symmetrization Tarski’s theorem §394 notes tempered distribution 284 notes tempered function §284 (284D), 286D tempered measure 284Yi tensor product of linear spaces 253 notes, 376Ya-376Yc tent map 372Xm, 384Xk test function 242Xi, 284 notes; see also rapidly decreasing test function (284A)

upper

General index

665

thick set see full outer measure (132F) tight family of measures see uniformly tight (285Xj) Tonelli’s theorem 252G, 252H, 252R topological measure (space) 256A topological space §2A3 (2A3A), §3A3 topological vector space see linear topological space (2A5A) topology §2A2, §2A3 (2A3A); see also convergence in measure (245A, 367M), linear space topology (2A5A), measure-algebra topology (323Ab) total order see totally ordered set (2A1Ac) total variation (of an additive functional) 231Yh; (of a function) see variation (224A) totally finite measure algebra 322Ab, 322Bb, 322C, 322Mb, 322Qc, 323Ad, 323Ca, 324Kb, 324P, 327B, 331B, 331D, 332M, 332P, 332Q, 333C-333G, 333J-333N, 333Q, 333R, 366Yh, 373Xj, 373Xn, 373Yb, 375Fb, 375H, 382F, 382I, 382J, 382Xf, 382Ya, 382Yc, 383O, 383Xe, 385D-385F, 385M, 385Xd, 385Xf, 391B, 394R, 395Xa totally finite measure (space) 211C, 211L, 211Xb, 211Xc, 211Xd, 212G, 213Ha, 214Ia, 214Ja, 215Yc, 232B, 232G, 243I, 243Xk, 245Fd, 245Xe, 245Ye, 246Xi, 246Ya, 322Bb, 382Yd totally ordered set 135Ba, 2A1Ac, 3A1F trace (of a σ-algebra) see subspace σ-algebra (121A) transfinite recursion 2A1B translation-invariant inverse-measure-preserving function 386Xd translation-invariant lifting 345A, 345B-345D, 345Ya-345Yc, 346C translation-invariant lower density 345Xf, 345Ya translation-invariant measure 115Xd, 134A, 134Ye, 134Yf, 255A, 255Ba, 255Yn, 345A, 345Xb, 345Ya truly continuous additive functional 232Ab, 232B-232E, 232H, 232I, 232Xa, 232Xb, 232Xf, 232Xh, 232Ya, 232Ye, 234Ce, 327C, 362Xh, 363S Tychonoff space see completely regular (3A3Ac) Tychonoff’s theorem 3A3J type see Maharam type (331F) Ulam S. see Banach-Ulam problem ultrafilter 254Yd, 2A1N, 2A1O, 2A3R, 2A3Se, 351Yd, 3A3De, 3A3Lc; see also principal ultrafilter (2A1N) Ultrafilter Theorem 2A1O uncompleted indefinite-integral measure 234Cc uniferent homomorphism 312F Uniform Boundedness Theorem 3A5H uniformity §3A4 (3A4A), 3A5I; see also measure-algebra uniformity (323Ab), uniform space (3A4A) uniformly complete Riesz space 354Yi, 354Yk uniformly continuous function 224Xa, 255K, 3A4C, 3A4Ec, 3A4G uniformly convergent (sequence of functions) 3A3N uniformly distributed sequence see equidistributed (281Yi) uniformly exhaustive submeasure 392Bc, 392C, 392E, 392F, 392J, 392Xc, 393A >246A), 252Yp, 272Yd, 273Na, 274J, 275H, 275Xi, 275Yl, 276Xd, uniformly integrable set (in L1 ) §246 (> 276Yb; (in L1 (µ)) §246 (246A), 247C, 247D, 247Xe, 253Xd, 354Q; (in an L-space) 354P, 354Q, 354R, 356O, 356Q, 356Xm, 362E, 362Yf-362Yh, 371Xf; (in L1 (A, µ ¯)) 367Xm, 373Xj, 373Xn uniformly tight (set of measures) 285Xj, 285Xk, 285Ye, 285Yf unit ball in Rr 252Q universal mapping theorems 253F, 254G, 315B, 315I, 315 notes, 325C, 325H, 325J, 369Xk universally complete Riesz space see laterally complete (368L) up-crossing 275E, 275F upper envelope see upr(a, C) (314V) upper integral 133I, 133J, 133K, 133Xe, 133Yf, 135H, 252Ye, 252Yh, 252Yi, 253J, 253K upper Riemann integral 134Ka

666

Index

upwards

upwards-directed partially ordered set 2A1Ab usual measure on {0, 1}I 254J; see under {0, 1}I usual measure on PX 254J; see under PX vague topology (on a space of measures) 274Ld, 274Xh, 274Ya-274Yd, 285K, 285L, 285S, 285U, 285Xk, 285Xq, 285Yd, 285Yg-285Yi, 285Yn, 367Xv variance of a random variable 271Ac, 271Xa, 272R, 272Xf, 285Gb, 285Xo variation of a function §224 (224A, 224K, 224Yd, 224Ye), 226B, 226Db, 226Xc, 226Xd, 226Yb; see also bounded variation (224A) —– of a measure see total variation (231Yh) vector integration see Bochner integral (253Yf) vector lattice see Riesz space (241E, 352A) vector measure 326Ye, 361Gb, 393O, 393P, 393Q, 393S, 393Xg, 393Yh, 393Yi very weak operator topology 373K, 373L, 373Xp, 373Xv, 373Yg virtually measurable function 122Q, 122Xe, 122Xf, 212Bb, 212F, 241A, 252E Vitali cover 261Ya Vitali’s theorem 221A, 221Ya, 221Yc, 221Yd, 261B, 261Yk Vitali-Hahn-Saks theorem 246Yg, 362Yh volume 115Ac —– of a ball in Rr 252Q, 252Xh von Neumann transformation 387G, 387H, 387Xd, 387Xd, 387Xg see also relatively von Neumann (387G), weakly von Neumann (387G) Wald’s equation 272Xh wandering see weakly wandering (395C) weak operator topology see very weak operator topology (373K) weak order unit (in a Riesz space) 353L, 353P, 353Yg, 368Yd weak topology of a normed space 247Ya, 2A5I, 356Yf, 356Yg, 3A5Eb —– see also very weak operator topology (373K), (relatively) weakly compact, weakly compactly generated, weakly convergent, weakly K-countably determined weak* topology on a dual space 253Yd, 285Yg, 2A5Ig, 3A5E, 3A5F; see also vague topology (274Ld) weakly compact linear operator 371Xb, 376Q, 3A5Kb; see also compact operator (3A5Ka) weakly compact set (in a linear topological space) 247C, 247Xa, 247Xc, 247Xd, 2A5I, 376Yj; see also relatively weakly compact (2A5Id) weakly convergent sequence in a normed space 247Yb weakly von Neumann transformation 387G, 387I, 387K, 387Xf, 387Ye weakly wandering element 395C weakly θ-refinable see hereditarily weakly θ-refinable weakly σ-distributive Boolean algebra 316Yg, 316Yh, 362Ye, 375Yc, 391Ya weakly (σ, ∞)-distributive Boolean algebra 316G, 316H-316K, 316Xk-316Xq, 316Xx, 316Yf, 316Yg, 316Yi, 316Yj, 316Yp, 322F, 325Yd, 362D, 367Yl, 368Q, 368R, 368Xe, 368Yg, 368Yi, 376Yf, 391D, 391K, 391Ya, 391Yg, 391Z, 392I, 393Yd, 394Yc, 395Ya weakly (σ, ∞)-distributive Riesz space 368N, 368O-368S, 368Xd, 368Xe, 368Ye, 368Yf, 368Yh, 368Yj 376H Weierstrass’ approximation theorem 281F; see also Stone-Weierstrass theorem well-ordered set 2A1Ae, 2A1B, 2A1Dg, 2A1Ka; see also ordinal (2A1C) Well-ordering Theorem 2A1Ka Weyl’s Equidistribution Theorem 281M, 281N, 372Xo Wiener’s Dominated Ergodic Theorem 372Yb Young’s function > 369Xc, 369Xd, 369Xr, 369Yc, 369Yd, 373Xm Young’s inequality 255Ym Zermelo’s Well-ordering Theorem 2A1Ka Zermelo-Fraenkel set theory 3A1A

General index

I

667

zero-dimensional topological space 311I-311K, 315Xf, 316Xo, 316Yd, 353Yc, 3A3Ad, 3A3Bd zero-one law 254S, 272O, 272Xf, 272Xg, 325Xg zero set in a topological space 313Yb, 316Yh, 324Yb, 3A3Pa Zorn’s lemma 2A1M, 3A1G a.e. (‘almost everywhere’) 112Dd a.s. (‘almost surely’) 112De Aut (in Aut A) see automorphism group of a Boolean algebra (381A); (in Autµ A) see automorphism group of a measure algebra (382A) AL (Lebesgue measure algebra) 373C B (in B(x, δ), closed ball) 261A, 2A2B B (in B(U ; V ), space of bounded linear operators) 253Xb, 253Yj, 253Yk, 2A4F, 2A4G, 2A4H, 371B-371D, 371G, 371Xd, 371Yc, 376M, 3A5H; (B(U ; U )) 395Xb c (in c(A), where A is a Boolean algebra) see cellularity (332D) c (in c(X), where X is a topological space) see cellularity (332Xd) c (the cardinal of R or PN) 2A1H, 2A1L, 343I, 343Yb, 344H, 344Yf, 382Xc, 391Yd, 391Ye C (in C(X), where X is a topological space) 243Xo, 281Yc, 281Ye, 281Yf, 352Xj, 353M, 353Xd, 354L, 354Yf, 353Yc, 363A, 367L, 367Yi, 367Yj, 367Yk, 368Ya C([0, 1]) 242 notes, 352Xg, 356Xb, 368Yf Cb (in Cb (X), where X is a topological space) 281A, 281E, 281G, 281Ya, 281Yd, 281Yg, 285Yg, 352Xj, 354Hb C ∞ (in C ∞ (X), for extremally disconnected X) 364W, 364Yl, 368G c 354Xq, 354Xs, 355Ye c 0 354Xa, 354Xd, 354Xi, 371Yc cac (‘countable antichain condition’) 316 notes ccc Boolean algebra 316Aa, 316C-316F, 316Xa-316Xj, 316Yc-316Ye, 316Yg, 316Yp, 322G, 324Yd, 325Yd, 326L, 326Xi, 331Ge, 332D, 332H, 363Yb, 364Yb, 367Yl, 368Yg, 368Yi, 391M, 391Xa, 392Ca, 392I, 393J, 393Yd, 393Ye, 394Xf, 394Yb ccc topological space 316Ab, 316B, 316Xd, 316Xe, 316Ya, 316Yd cf (in cf P ) see cofinality (3A1Fb) c.l.d. product measure §§251-253 (251F, 251W), 254Db, 254U, 254Ye, 256K, 256L, 325A, 325B, 325C, 325H, 334A, 334Xa, 342Ge, 342Id, 342Xn, 343H, 354Yl, 376J, 376R, 376S, 376Yc c.l.d. version of a measure (space) 213E, 213F-213H, 213M, 213Xb-213Xe, 213Xg, 213Xk, 213Xn, 213Yb, 214Xf, 214Xj, 232Ye, 234Yf, 241Ya, 242Yh, 244Ya, 245Yc, 251Ic, 251S, 251Wn, 251Xd, 251Xj, 251Xk, 252Ya, 322D, 322Qb, 322Xc, 322Xi, 322Yb, 324Xc, 324Xe, 342Gb, 342Ib, 342Xn, 343H, 343Ye CM∗ (formulations of the Control Measure Problem) 393A, 393H, 393J, 393L, 393P, 393Xe, 393Xg d (in d(X)) see density (331Yf) D (in Dn (A, π), where A is a subset of a Boolean algebra, and π is a homomorphism) 384K, 384L, 384M diam (in diam A) = diameter dom (in dom f ): the domain of a function f ess sup see essential supremum (243Da) E (in E(X), expectation of a random variable) 271Ab f

(in Af ) 361Ad F (in F(B ↑), F(B↓)) 323D f -algebra 241H, 241 notes, 352W, 352Xj-352Xm, 353O, 353P, 353Xd, 353Yg, 353Yh, 361Eh, 363B, 364C-364E, 367Yg Gδ set 264Xe h (in h(π)) see entropy (384M); (in h(π, A)) 384M, 384N-384P, 384Xq, 384Yb, 386C H (in H(A)) see entropy of a partition (384C); (in H(A|B)) see conditional entropy (384D) I k see split interval (343J)

668

Index

L

`1 (in `1 (X)) 242Xa, 243Xl, 246Xd, 247Xc, 247Xd, 354Xa, 356Xc `1 (= `1 (N)) 246Xc, 354M, 354Xd, 356Xl `2 244Xn, 282K, 282Xg, 355Yb, 371Ye, 376Yh, 376Ym, 376Yn `p (in `p (X)) 244Xn, 354Xa `∞ (in `∞ (X)) 243Xl, 281B, 281D, 354Ha, 354Xa, 361D, 361L `∞ (= `∞ (N)) 243Xl, 354Xj, 356Xa, 371Yd, 382J `∞ -complemented subspace 363Yd L-space 354M, 354N-354P, 354R, 354Xt, 354Yj, 356N, 356P, 356Q, 356Xm, 356Yf, 362A, 362B, 362Yj, 365C, 365Xc, 365Xd, 367Xn, 369E, 371A-371E, 371Xa, 371Xb, 371Xf, 371Ya, 376M, 376P, 376Yi, 376Yj L0 (in L0 (µ)) 121Xb, 121Ye, §241 (241A), §245, 253C, 253Ya; (in L0 (Σ)) 345Yb, 364C, 364D, 364E, 364J, 364Yi; see also L0 (241C), L0strict (241Yh), L0C (241J) L0strict 241Yh L0C (in L0C (µ)) 241J, 253L L0 (in L0 (µ)) §241 (241A), 242B, 242J, 243A, 243B, 243D, 243Xe, 243Xj, §245, 253Xe, 253Xf, 253Xg, 271De, 272H, 323Xf, 345Yb, 352Xj, 364Jc, 376Yc; (in L0C (µ)) 241J; (in L0 (A)) §364 (364A), 368A-368E, 368H, 368K, 368M, 368Qb, 368R, 368S, 368Xa, 368Xe, 368Ya, 368Yd, 368Yi, §369, 372H, §375, 376B, 376Yb, 393M, 394I; (in L0C (A)) 364Yn; see also L0 (241A, 364C) L1 (in L1 (µ)) 122Xc, 242A, 242Da, 242Pa, 242Xb; (in L1strict (µ) 242Yg, 341Ye; (in L1C (µ) 242P, 255Yn; (in L1V (µ)) 253Yf; see also L1 , k k1 L1 (in L1 (µ)) §242 (242A), 243De, 243F, 243G, 243J, 243Xf, 243Xg, 243Xh, 245H, 245J, 245Xh, 245Xi, §246, §247, §253, 254R, 254Xp, 254Ya, 254Yc, 255Xc, 257Ya, 282Bd, 327D, 341Ye, 354M, 354Q, 354Xa, 365B, 376N, 376Q, 376S, 376Yl; (in L1V (µ)) 253Yf, 253Yi, 354Yl; (in L1 (A, µ ¯) or L1µ¯ ) §365 (365A), 366Yc, 367J, 367U, 367Yt, 369E, 369N, 369O, 369P, 371Xc, 371Yb, 371Yc, 371Yd, 372B, 372C, 372E, 372G, 372Xc, 376C, 385G, 385H, 385J; see also L1 , L1C , k k1 L1C (µ) 242P, 243K, 246K, 246Yl, 247E, 255Xc; see also convolution of functions L2 (in L2 (µ)) 244Ob, 253Yj, §286; (in L2C (µ)) 284N, 284O, 284Wh, 284Wi, 284Xi, 284Xk-284Xm, 284Yg; see also L2 , Lp , k k2 L2 (in L2 (µ)) 244N, 244Yk, 247Xe, 253Xe, 355Ye, 372 notes; (in L2C (µ)) 282K, 282Xg, 284P; (in L2 (A, µ ¯)) 366K, 366L, 366Xh, 395A, 395Xb; see also L2 , Lp , k k2 Lp (in Lp (µ)) §244 (244A), 246Xg, 252Ym, 253Xh, 255K, 255Og, 255Yc, 255Yd, 255Yl, 255Ym, 261Xa, 263Xa, 273M, 273Nb, 281Xd, 282Yc, 284Xj, 286A; see also Lp , L2 , k kp Lp (in Lp (µ), 1 < p < ∞) §244 (244A), 245G, 245Xj, 245Xk, 245Yg, 246Xh, 247Ya, 253Xe, 253Xi, 253Yk, 255Yf, 354Xa, 354Yk, 366B, 376N; (in Lp (A, µ ¯) = Lpµ¯ , 1 < p < ∞) §366 (366A), 369L, 371Gd, p 372Xp, 372Xq, 372Yb, 373Bb, 373F, 376Xb;(in LC (µ), 1 < p < ∞) 354Yk;(in Lp (A, µ ¯), 0 < p < 1) 366Ya, 366Yg;see also Lp , k kp L∞ (in L∞ (µ)) 243A, 243D, 243I, 243Xa, 243Xl, 243Xn; (in L∞ (Σ)) 341Xe, 363H;see also L∞ L∞ C 243K L∞ strict 243Xb L∞ (in L∞ (µ)) §243 (243A), 253Yd, 341Xe, 352Xj, 354Hc, 354Xa, 363I, 376Xn; (in L∞ (A)) §363 (363A), 364K, 364Xh, 365I, 365J, 365K, 365Xk, 368Q, 394N; see also L∞ , L∞ C , k k∞ L∞ 243K, 243Xm C Lτ (where τ is an extended Fatou norm) 369G, 369J, 369K, 369M, 369O, 369R, 369Xi, 374Xd, 374Xi; see also Orlicz space (369Xd), Lp , M 1,∞ (369N), M ∞,1 (369N) L (in L(U ; V ), space of linear operators) 253A, 253Xa, 351F, 351Xd, 351Xe L∼ (in L∼ (U ; V ), space of order-bounded linear operators) §355 (355A), 356Xi, 361H, 361Xc, 361Yb, 363Q, 365N, 371B-371E, 371Gb, 371Xb-371Xe, 371Ya, 371Yc-371Ye, 375Kb, 376J, 376Xe, 376Ym; see also order-bounded dual (356A) ∼ L∼ c (in Lc (U ; V )) 355G, 355I, 355Yi, 376Yf; see also sequentially order-continuous dual (356A) L× (in L× (U ; V )) 355G, 355H, 355J, 355K, 355Yg, 355Yi, 355Yj, 371B-371D, 371Gb, 376D, 376E, 376H, 376K, 376Xj, 376Yf; see also order-continuous dual (356A) lim (in lim F) 2A3S; (in limx→F ) 2A3S lim inf (in lim inf n→∞ ) §1A3 (1A3Aa), 2A3Sg; (in lim inf δ↓0 ) 2A2H; (in lim inf x→F ) 2A3S lim sup (in lim supn→∞ ) §1A3 (1A3Aa), 2A3Sg; (in lim supδ↓0 ) 2A2H, 2A3Sg; (in lim supx→F f (x)) 2A3S

General index

T

669

ln+ 275Yd M (in M (A), space of bounded finitely additive functionals) 362B, 362E, 363K M -space 354Gb, 354H, 354L, 354Xq, 354Xr, 356P, 356Xj, 363B, 363O, 371Xd, 376M; see also order-unit norm (354Ga) M 0 (in M 0 (A, µ ¯) = Mµ¯0 ) 366F, 366G, 366H, 366Yb, 366Yd, 366Yg, 373D, 373P, 373Xl 0,∞ M 252Yp M 0,∞ (in M 0,∞ (A, µ ¯) = Mµ¯0,∞ ) 373C, 373D, 373E, 373F, 373I, 373Q, 373Xo, 374B, 374J, 374L M 1,0 (in M 1,0 (A, µ ¯) = Mµ¯1,0 ) 366F, 366G, 366H, 366Ye, 369P, 369Q, 369Yh, 371F, 371G, 372D, 372Ya, 373G, 373H, 373J, 373S, 373Xp, 373Xr, 374Xe M 1,∞ (in M 1,∞ (µ)) 234Yd, 244Xl, 244Xm, 244Xo, 244Yc; (in M 1,∞ (A, µ ¯) = Mµ¯1,∞ ) 369N, 369O-369Q, 369Xi-369Xk, 369Xm, 369Xq, §373, 374A, 374B, 374M M ∞,0 (in M ∞,0 (A, µ ¯)) 366Xd, 366Yc M ∞,1 (in M ∞,1 (A, µ ¯) = Mµ¯∞,1 ) 369N, 369O, 369P, 369Q, 369Xi, 369Xj, 369Xk, 369Xl, 369Yh, 373K, 373M, 374B, 374M, 374Xa, 374Ya Mσ (in Mσ (A), space of countably additive functionals) 362B, 362Xd, 362Xh, 362Xi, 362Ya, 362Yb, 363K Mτ (in Mτ (A), space of completely additive functionals) 326Yp, 327D, 362B, 362D, 362Xd, 362Xg, 362Xi, 362Ya, 362Yb, 363K N 3A1H N × N 111Fb NN 372Xi —– see also PN N see ideal of negligibles On (the class of ordinals) 3A1E p (in p(t)) 385I, 385J P (in PX) 311Ba, 311Xe, 312B, 312C, 313Ec, 313Xf, 363S, 381Xe, 382Yb; (usual measure on PX) 254J, 254Xf, 254Xq, 254Yd PN 1A1Hb, 2A1Ha, 2A1Lb, 315O, 316Yo, 324Yg, 326Yg, 374Xk; (usual measure on) 273G, 273Xd, 273Xe P(NN ) 316Yg p.p. (‘presque partout’) 112De X ∈ E) etc. 271Ad Pr(X > a), Pr(X Q (the set of rational numbers) 111Eb, 1A1Ef, 364Yg q (in q(t)) 384A, 385O R (the set of real numbers) 111Fe, 1A1Ha, 2A1Ha, 2A1Lb, 352M RX 245Xa, 256Ye, 352Xj, 375Ya, 3A3K; see also Euclidean metric, Euclidean topology RX |F see reduced power (351M) R C 2A4A R see extended real line (§135) RO (in RO(X)) see regular open algebra (314Q) S (in S(A)) 243I, §361 (361D), 363C, 363Xg, 364K, 364Xh, 365F, 368Q, 369O; (in S f ∼ = S(Af )) 242M, ∼ ∼ × 244H, 365F, 365G, 369O, 369P; (in S(A) ) 362A; (in S(A)c ) 362Ac; (in S(A) ) 362Ad; (in SC (A)) 361Xj, 361Yd S see rapidly decreasing test function (284A) S 1 (the unit circle, as topological group) see circle group S6 (the group of permutations of six elements) 383 notes sf (in µsf ) see semi-finite version of a measure (213Xc); (in µ∗sf ) 213Xf, 213Xg, 213Xk T2 topology see Hausdorff (2A3E, 3A3Aa) (0) T (0) (in Tµ¯,¯ν ) 371F, 371G, 371H, 372D, 372Xb, 372Yb, 372Yc, 373B, 373G, 373J, 373R, 373S, 373Xp, 373Xq, 373Xr, 373Xu, 373Xv T (in Tµ¯,¯ν ) 244Xm, 244Xo, 244Yc, 246Yc, §373 (373A); see also T -invariant (374A)

670

Index

T

T × (in Tµ¯×,¯ν ) §373 (373Ab), 376Xa, 376Xh T -invariant extended Fatou norm 374Ab, 374B-374D, 374Fa, 374Xb, 374Xd-374Xj, 374Yb T -invariant set 374Aa, 374M, 374Xa, 374Xi, 374Xk, 374Xl, 374Ya, 374Ye Tm see convergence in measure (245A) Ts (in Ts (U, V )) 373M, 373Xq, 376O, 3A5E; see also weak topology (2A5I), weak* topology (2A5Ig) U (in U (x, δ)) 1A2A upr (in upr(a, C)) 314V, 314Xg, 333Xa, 365Rc, 394G, 394I, 394K-349N, Var (in Var(X)) see variance (271Ac); (in VarD f , Var f ) see variation (224A) w∗ -topology see weak* topology 2A5Ig Z (the set of integers) 111Eb, 1A1Ee; (as topological group) 255Xe Z2 (the group {0, 1}) 311Bc, 311D ZFC see Zermelo-Fraenkel set theory βr (volume of unit ball in Rr ) 252Q, 252Xh, 265F, 265H, 265Xa, 265Xb, 265Xe Γ-function 225Xj, 225Xk, 252Xh, 252Yk, 252Yn, 255Xj ∆-system 2A1Pa θ-refinable see hereditarily weakly θ-refinable µ ¯L (in §373) 373C νX see distribution of a random variable (271C) π-λ Theorem see Monotone Class Theorem (136B) σ-additive see countably additive (231C, 326E) σ-algebra of sets §111 (111A), 136Xb, 136Xi, 212Xk, 314D, 314M, 314N, 314Yd, 316D, 322Ya, 326Ys, 343D, 344D, 362Xg, 363H; see also Borel σ-algebra (111G) σ-algebra defined by a random variable 272C, 272D σ-complete see Dedekind σ-complete (241Fb, 314Ab) σ-field see σ-algebra (111A) σ-finite measure algebra 322Ac, 322Bc, 322C, 322G, 322M, 323Gb, 323Ya, 324K, 325Eb, 327B, 331N, 362Xd, 367Nd, 367Q, 367Xq, 367Xs, 369Xg, 382E σ-finite measure (space) 211D, 211L, 211M, 211Xe, 212G, 213Ha, 213Ma, 214Ia, 214Ja, 215B, 215C, 215Ya, 215Xe, 215Yb, 216A, 232B, 232F, 234F, 235O, 235R, 235Xe, 235Xk, 241Yd, 243Xi, 245Eb, 245K, 245L, 245Xe, 251K, 251Wg, 252B-252E, 252H, 252P, 252R, 252Xc, 252Yb, 252Yl, 322Bc, 342Xi, 362Xh, 365Xp, 367Xr, 376I, 376J, 376N, 376S σ-generating set in a Boolean algebra 331E σ-ideal (in a Boolean algebra) 313E, 313Pb, 313Qb, 314C, 314D, 314L, 314N, 314Yd, 316C, 316D, 316Xi, 316Ye, 321Ya, 322Ya, 392Xd —– (of sets) 112Db, 211Xc, 212Xf, 212Xk, 313Ec, 322Ya, 363H σ-linked Boolean algebra 391L, 391M, 391N, 391Xj, 391Yb, 391Yd, 391Ye, 391Z, 393Xf; σ-m-linked Boolean algebra 391Yh, 393Ya σ-order complete see Dedekind σ-complete (314Ab) σ-order-continuous see sequentially order-continuous (313H) σ-subalgebra of a Boolean algebra 313E, 313F, 313G, 313Xd, 313Xe, 314Eb, 314Fb, 314Hb, 314Jb, 314Xc, 315Yc, 321G, 321Xb, 322M, 323Z, 324Xb, 326Fg, 331E, 331G, 364Xc, 366I; see also order-closed subalgebra σ-subalgebra of sets §233 (233A), 321Xb, 323Xb σ-subhomomorphism between Boolean algebras 375E, 375F-375H, 375Xd, 375Yc, 375Ye (σ, P ∞)-distributive see weakly (σ, ∞)-distributive (316G) i∈I ai 112Bd, 222Ba, 226A

special symbols

General index

671

τ (in τ (A)) see Maharam type (331Fa); (in τC (A)) see relative Maharam type (333Aa) τ -additive functional on a Boolean algebra see completely additive (326J) τ -additive measure 256M, 256Xb, 256Xc τ -generating set in a Boolean algebra 313Fb, 313M, 331E, 331F, 331G, 331Yb, 331Yc Φ see normal distribution function (274Aa) χ (in χA, where A is a set) 122Aa; (in χa, where a belongs to a Boolean ring) 361D, 361Ef, 361L, 361M, 364K; (the function χ : A → L0 (A)) 364Kc, 367R ω (the first infinite ordinal) 2A1Fa, 3A1H; (in [X]<ω ) 3A1Cd, 3A1J ω ω -bounding Boolean algebra see weakly σ-distributive (316Yg) ω1 (the first uncountable ordinal) 2A1Fc ω1 -saturated ideal in a Boolean algebra 316C, 316D, 341Lh, 344Yd, 344Ye —– see also Pω1 ω2 2A1Fc ωξ (the ξth uncountable initial ordinal) 3A1E see closure (2A2A, 2A3Db) ∗ (in f ∗ g, u ∗ v, λ ∗ ν, ν ∗ f , f ∗ ν) see convolution (255E, 255O, 255Xe, 255Yn) * (in weak*) see weak* topology (2A5Ig);(in U ∗ = B(U ; R), linear topological space dual) see dual (2A4H)(in u∗ ) see decreasing rearrangement (373C);(in µ∗ ) see outer measure defined by a measure (132B) ∗ (in µ∗ ) see inner measure defined by a measure (113Yh) ∼ (in U ∼ ) see order-bounded dual (356A) ∼ ∼ c (in Uc ) see sequentially order-continuous dual (356A) × (in U × ) see order-continuous dual (356A); (in U ×× ) see order-continuous bidual 111C R\ (in ER \ F ,R ‘set difference’) R R (in f , f dµ, f (x)µ(dx)) R 122E, 122K, 122M, 122Nb, > 363L; (in A f )R 131D, 214D, 235Xf; R see also subspace measure; (in u) 242Ab, 242B, 242D, 363L, 365D, 365Xa; (in A u) 242Ac; (in a u) 365D, 365Xb; see also upper integral, lower integral (133I) R R see upper integral (133I) see lower integral (133I) R R see Riemann integral (134K) ¹ (in f ¹A, the restriction of a function to a set) 121Eh | | (in a Riesz space) 241Ee, 242G, §352 (352C), 354Aa, 354Bb k ke see order-unit norm (354Ga) k k1 (on L1 (µ)) §242 (242D), 246F, 253E, 275Xd, 282Ye; (on L1 (µ)) 242D, 242Yg, 273Na, 273Xi; (on 1 L (A, µ ¯)) 365A, 365B, 365C, 385G, 385H; (on the `1 -sum of Banach lattices) 354Xb, 354Xo k k2 244Da, 273Xj, 366Yh; see also L2 , k kp k kp (for 1 < p < ∞) §244 (244Da), 246Xb, 246Xh, 246Xi, 252Ym, 252Yp, 253Xe, 253Xh, 273M, 273Nb, 275Xe, 275Xf, 275Xh, 276Ya, 366A, 366C, 366D, 366H, 366J, 366Xa, 366Xi, 366Yf, 367Xp, 369Oe, 372Xb, 372Yb, 374Xb; see also Lp , Lp , k kp,q k kp,q (the Lorentz norm) 374Yb k k∞ 243D, 243Xb, 243Xo, 244Xh, 273Xk, 281B, 354Xb, 354Xo, 356Xc, 361D, 361Ee, 361I, 361J, 361L, 361M, 363A, 364Xh; see also essential supremum (243D), L∞ , L∞ , `∞ k k1,∞ 369O, 369P, 369Xh-369Xj, 371Gc, 372D, 372F, 373F, 373Xk; see also M 1,∞ , M 1,0 k k∞,1 369N, 369O, 369Xi, 369Xj, 369Xl; see also M ∞,1 ⊗ (in f ⊗ g) 253B, 253C, 253J, 253L, 253Ya, 253Yb; (in u ⊗ v) §253 (253E); (in A ⊗ B, a ⊗ b) see free product N (315M) N (in i∈I Ai ) see free product (315H) b (in Σ⊗T) b ⊗ 251D, 251K, 251L, 251Xk, 251Ya, 252P, 252Xd, 252Xe, 253C N N c (in c Σi ) 251Wb, 251Wf, 254E, 254F, 254Mc, 254Xc, 254Xi, 343Xb Q Q i∈I Q (in i∈I αi ) 254F; (in i∈I Xi ) 254Aa # (in #(X), the cardinal of X) 2A1Kb

672

Index

special symbols

S S S T (in Tn∈N En ) 111C; (in T A) 1A1F (in n∈N En ) 111C; (in E) 1A2F 4 (in E4F , ‘symmetric difference’) 111C, 311Ba ∪ , ∩ (in a Boolean ring or algebra) 311Ga, 313Xi, 323B \ , 4 (in a Boolean ring or algebra) 311Ga, 323B ⊆ , ⊇ (in a Boolean ring or algebra) 311H, 323Xa C (in I C R) see ideal (3A2Ea) −− ←−−−− −) (in (← (← a π b), (a π b φ c) etc.) see cycle notation (381G), cyclic automorphism, exchanging involution (381G) + (in κ+ , successor cardinal) 2A1Fc; (in f + , where f is a function) 121Xa, 241Ef; (in u+ , where u belongs to a Riesz space) 241Ef, 352C; (in U + , where U is a partially ordered linear space) 351C; (in F (x+ ), where F is a real function) 226Bb − (in f − , where f is a function) 121Xa, 241Ef; (in u− , in a Riesz space) 241Ef, 352C; (in F (x− ), where F is a real function) 226Bb 2 (in 2κ ) 3A1D ∨, ∧ (in a lattice) 121Xa, 2A1Ad; (in A ∨ B, where A, B are partitions of unity in a Boolean algebra) 384F ⊥ (in A⊥ , in a Boolean algebra) 313Xo; (in A⊥ , in a Riesz space) 352O, 352P, 352Q, 352R, 352Xg; (in ⊥ V , in a Hilbert space) see orthogonal complement; see also complement of a band 0 (in U 0 ) see algebraic dual; (in T 0 ) see adt operator a (in z a i) 3A1H ,

∧

∨

(in f , f ) see Fourier transform, inverse Fourier transform (283A) ¯ ¯ (in h(u), where h is a Borel function and u ∈ L0 ) 241I, 241Xd, 241Xi, 245Dd, 364I, 364J, 364Xg, 364Xq, 364Yd, 364Ye, 367I, 367S, 367Xl, 367Ys {0, 1}I (usual measure on) 254J, 254Xe, 254Yc, 272N, 273Xb, 331J-331L, 332B, 332C, 332N, 332Xm, 332Xn, 341Yc, 341Yd, 341Zb, 342Jd, 343Ca, 343I, 343Xc, 343Yd, 344G, 344L, 344Xg, 345Ab, 345C-345E, 345Xa, 346C, 382Xb; (when I = N) 254K, 254Xd, 254Xj, 256Xk, 261Yd, 341Xb, 343Cb, 343H, 343M, 345Yc, 346Zb, 387H; see also PX —– (usual topology of) 311Xh, 3A3K; (when I = N) 314Ye —– (open-and-closed algebra of) 311Xh, 315Xh, 316Xq, 316Yj, 316Ym, 331Yg, 391Xd, 393A, 393F —– (regular open algebra of) 316Yj (2, ∞)-distributive lattice 367Yd ¿ (in ν ¿ µ) see absolutely continuous (232A) 4τG (in a 4τG b) 394A, 394G, 394I, 394K, 394Ma, 394Xb ∞ see infinity [ ] (in [a, b]) see closed interval (115G, 1A1A, 2A1Ab); (in f [A], f −1 [B], R[A], R−1 [B]) 1A1B; (in [X]κ , [X]<κ , [X]≤κ ) 3A1J; (in [X]<ω ) 3A1Cd, 3A1J [[ ]] (in f [[F]]) see image filter (2A1Ib) [[ ]] (in [[u > α]], [[u ≥ α]], [[u ∈ E]] etc.) 363Xh, 364A, 364B, 364H, 364Xa, 364Xc, 364Yc; (in [[µ > ν]]) 326O, 326P [ [ (in [a, b[) see half-open interval (115Ab, 1A1A) ] ] (in ]a, b]) see half-open interval (1A1A) ] [ (in ]a, b[) see open interval (115G, 1A1A) d e (in db : ae) 394I, 394J, 394K-394M, 394Xa b c (in bb : ac) 394I, 394J, 394K-394M, 394Xa (in µ E) 234E, 235Xf ∧

∨

Measure Theory 3: Measure Algebras - D.h. Fremlin 1t626i

Overview 3e4r5l

More details w3441

Related Documents 3m3m1z

Measure Theory 3: Measure Algebras - D.h. Fremlin 1t626i

Measure Theory Fremlin Vol4 3e1q6p

Fremlin Measure Theory Volume 3 5x4t4w

Measure For Measure Isabella 5s1x1t

Measure Theory In Economics 1o3m3u

Measure Theory For Dummies 1m1g4y

More Documents from "Paulo Henrique Macedo" 614s24

Measure Theory 3: Measure Algebras - D.h. Fremlin 1t626i

Webcast Bacula 20062011 385d3b

6l6m1j

Dori Caymmi - Estrela Da Terra.pdf 6r5nm

Fenomenos De Superficie 1j1j2k

Quem Sou Eu Senhor - Vozes 2g226g