Documente Academic
Documente Profesional
Documente Cultură
Volume 3
D.H.Fremlin
MEASURE THEORY
Volume 3
Measure Algebras
D.H.Fremlin
Research Professor in Mathematics, University of Essex
This book may be ordered from the publisher at the address below. For price and means of payment see the authors Web page http://www.essex.ac.uk/maths/staff/fremlin/mtsales.htm,
or enquire from fremdh@essex.ac.uk.
The right of D.H.Fremlin to be identied as author of this work has been asserted in accordance with
the Copyright, Designs and Patents Act 1988. This work is issued under the terms of the Design Science
License as published in http://dsl.org/copyleft/dsl.txt. For the source les see http://www.essex.
ac.uk/maths/staff/fremlin/mt3.2004/index.htm.
Library of Congress classication QA312.F72
AMS 2000 classication 28A99
ISBN 0-9538129-3-6
Typeset by AMS-TEX
Contents
General Introduction
10
Introduction to Volume 3
11
14
14
Boolean rings and algebras; ideals and ring homomorphisms to Z2 ; Stones theorem; the operations
, , , \ and the relation ; topology of the Stone space; Boolean algebras as complemented
distributive lattices.
312 Homomorphisms
22
Subalgebras; ideals; Boolean homomorphisms; the ordering determines the ring structure; quotient algebras; extension of homomorphisms; homomorphisms and Stone spaces.
313 Order-continuity
31
General distributive laws; order-closed sets; order-closures; Monotone Class Theorem; order-preserving
functions; order-continuity; order-dense sets; order-continuous Boolean homomorphisms; regularly embedded subalgebras.
314 Order-completeness
41
52
Simple product of Boolean algebras; free product of Boolean algebras; algebras of sets and their quotients.
61
The countable chain condition; weak (, )-distributivity; Stone spaces; atomic and atomless Boolean
algebras.
69
69
Measure algebras; elementary properties; the measure algebra of a measure space; Stone spaces.
72
Totally finite, -finite, semi-finite and localizable measure algebras; relation to corresponding types of
measure space; completions and c.l.d. versions of measures; semi-finite measure algebras are weakly
(, )-distributive; subspace measures; simple products of measure algebras; Stone spaces of localizable
measure algebras; localizations of semi-finite measure algebras.
82
Defining a topology and uniformity on a measure algebra; continuity of algebraic operations; order-closed
sets; Hausdorff and metrizable topologies, complete uniformities; closed subalgebras.
324 Homomorphisms
89
96
The measure algebra of a product measure; the localizable measure algebra free product of two semi-finite
measure algebras; the measure algebra of a product of probability measures; the probability algebra free
product of probability algebras; factorizing through subproducts.
105
Additive, countably additive and completely additive functionals; Jordan decomposition; Hahn decomposition.
116
123
123
132
Any localizable measure algebra is isomorphic to a simple product of homogeneous totally finite algebras;
exact description of isomorphism types; closed subalgebras.
142
Relative Maharam types; extension of measure-preserving Boolean homomorphisms; complete classification of closed subalgebras of probability algebras as triples (A,
, C).
334 Products
158
162
162
Liftings and lower densities; strictly localizable spaces have lower densities; construction of a lifting from
a density; complete strictly localizable spaces have liftings; liftings and Stone spaces.
174
Inner regular measures; compact classes; compact and locally compact measures; perfect measures.
182
Representing homomorphisms between measure algebras by functions; possible when target measure
space is locally compact; countably separated measures and uniqueness of representing functions; the
split interval; perfect measures.
190
200
208
Liftings of product measures which respect the product structure; translation-invariant liftings on {0, 1}I ;
products of Maharam-type-homogeneous probability spaces; lower densities respecting product structures; consistent liftings; the Stone space of Lebesgue measure.
219
219
Partially ordered linear spaces; positive cones; suprema and infima; positive linear operators; ordercontinuous linear operators; Riesz homomorphisms; quotient spaces; reduced powers; representation of
p.o.l.ss as subspaces of reduced powers of R; Archimedean spaces.
226
Riesz spaces; identities; general distributive laws; Riesz homomorphisms; Riesz subspaces; order-dense
subspaces and order-continuous operators; bands; the algebra of complemented bands; the algebra of
projection bands; principal bands; f -algebras.
237
Order-dense subspaces; bands; Dedekind ()-complete spaces; order-closed Riesz subspaces; order units;
f -algebras.
246
Riesz norms; Fatou norms; the Levi property; order-continuous norms; order-unit norms; M -spaces; are
isomorphic to C(X) for compact Hausdorff X; L-spaces; uniform integrability in L-spaces.
255
Order-bounded linear operators; the space L (U ; V ); order-continuous operators; extension of ordercontinuous operators; the space L (U ; V ); order-continuous norms.
264
275
276
Additive functions on Boolean rings; the space S(A); universal mapping theorems for linear operators
on S; the map T : S(A)S(B) induced by a ring homomorphism : A B; projection bands in S(A);
identifying S(A) when A is a quotient of an algebra of sets.
362 S
288
Bounded additive functionals on A identified with order-bounded linear functionals on S(A); the Lspace S and its bands; countably additive, completely additive, absolutely continuous and continuous
functionals; uniform integrability in S .
363 L
298
The space L (A), as an M -space and f -algebra; universal mapping theorems for linear operators on L ;
T : L (A)L (B); representing L when A is a quotient of an algebra of sets; integrals with respect
to finitely additive functionals; projection bands in L ; (L ) and its bands; Dedekind completeness of
A and L ; representing -complete M -spaces; the generalized Hahn-Banach theorem; the Banach-Ulam
problem.
364 L0
314
The space L0 (A); representing L0 when A is a quotient of a -algebra of sets; algebraic operations on
L0 ; action of Borel measurable functions on L0 ; identifying L0 (A) with L0 () when A is a measure
algebra; embedding S and L in L0 ; suprema and infima in L0 ; Dedekind completeness in A and L0 ;
multiplication in L0 ; projection bands; T : L0 (A)L0 (B); simple products; *regular open algebras;
*the space C (X).
365 L1
335
366 Lp
351
R
R
ym theorem again; wT u d
=
The space L1 (A,
); identification with L1 (); a u; the Radon-Nikod
R
1
1
u d
; additive functions on A and linear operators on L ; the duality between L and L ; T :
L1 (A,
)L1 (B, ) and P : L1 (B, )L1 (A,
); conditional expectations; bands in L1 ; varying
.
360
376
387
Riesz spaces separated by their order-continuous duals; representing U when U L0 ; Kakutanis repre
sentation of L-spaces as L1 spaces; extended Fatou norms; associate norms; L
= (L ) ; Fatou norms
,1
1,
2
1
and convergence in measure; M
and M
, k k,1 and k k1, ; L + L .
403
403
(0)
408
(0)
The Maximal Ergodic Theorem and the Ergodic Theorem for operators in T,
; for inverse-measurepreserving functions : XX; limit operators as conditional expectations; applications to continued
fractions; mixing and ergodic transformations.
R
R
The classes T , T ; the space M 0, ; decreasing rearrangements u ; ku kp = kukp ; |T uv| u v
if T T ; the very weak operator topology and compactness of T ; v is expressible as T u, where T T ,
R
R
R
R
(0)
iff 0t v 0t u for every t; finding T such that T uv = u v ; the adjoint operator from T,
to
(0)
T, .
424
442
452
Linear operators on L0 spaces; if B is measurable, a positive linear operator from L0 (A) to L0 (B) can
be assembled from Riesz homomorphisms.
458
Kernel operators; free products of measure algebras and tensor products of L0 spaces; tensor products
of L1 spaces; abstract integral operators (i) as a band in L (U ; V ) (ii) represented by kernels belonging
b
to L0 (AB)
(iii) as operators converting weakly convergent sequences into order*-convergent sequences;
operators into M -spaces or out of L-spaces; disintegrations.
478
478
Assembling an automorphism; elements supporting an automorphism; periodic and aperiodic parts; full
and countably full subgroups; recurrence; induced automorphisms of principal ideals; Stone spaces; cyclic
automorphisms.
491
506
511
If G Aut A, H Aut B have many involutions, any isomorphism between G and H arises from an
isomorphism between A and B; if A is nowhere rigid, Aut A has no outer automorphisms; applications
to localizable measure algebras.
385 Entropy
520
Entropy of a partition of unity in a probability algebra; conditional entropy; entropy of a measurepreserving homomorphism; calculation of entropy (Kolmogorov-Sina theorem); Bernoulli shifts; isomorphic homomorphisms and conjugacy classes in Aut A; almost isomorphic inverse-measure-preserving
functions.
534
546
Bernoulli partitions; finding Bernoulli partitions with elements of given measure (Sinas theorem); adjusting Bernoulli partitions; Ornsteins theorem (Bernoulli shifts of the same finite entropy are isomorphic); Ornsteins and Sinas theorems in the case of infinite entropy.
569
Orbits of inverse-measure-preserving functions; von Neumann transformations; von Neumann transformations generating a given full subgroup; classification of full subgroups generated by a single automorphism.
583
583
Measurable algebras; strictly positive additive functionals and weak (, )-distributivity; additive functionals subordinate to or dominating a given functional; intersection numbers; existence of strictly
positive additive functionals; -linked Boolean algebras; Gaifmans example.
392 Submeasures
590
Submeasures; exhaustive, uniformly exhaustive and Maharam submeasures; the Kalton-Roberts theorem
(a strictly positive uniformly exhaustive submeasure provides a strictly positive additive functional);
order*-convergence in Boolean algebras; the order-sequential topology on a Boolean algebra.
598
613
627
Appendix to Volume 3
Introduction
3A1 Set theory
631
631
Calculation of cardinalities; cofinal sets, cofinalities; notes on the use of Zorns Lemma; the natural
numbers as finite ordinals; lattice homomorphisms; the Marriage Lemma.
3A2 Rings
633
Rings; subrings, ideals, homomorphisms, quotient rings, the First Isomorphism Theorem; products.
636
Hausdorff, regular, completely regular, zero-dimensional, extremally disconnected, compact and locally
compact spaces; continuous functions; dense subsets; meager sets; Baires theorem for locally compact
spaces; products; Tychonoffs theorem; the usual topologies on {0, 1}I , R I ; cluster points of filters;
topology bases; uniform convergence of sequences of functions; one-point compactifications.
3A4 Uniformities
641
Uniform spaces; and pseudometrics; uniform continuity; subspaces; product uniformities; Cauchy filters
and completeness; extending uniformly continuous functions; completions.
643
The Hahn-Banach theorem in analytic and geometric forms; cones and convex sets; weak and weak*
topologies; reflexive spaces; Uniform Boundedness Theorem; completions; normed algebras; compact
linear operators; Hilbert spaces.
3A6 Groups
646
Concordance
647
648
652
660
10
General introduction
In this treatise I aim to give a comprehensive description of modern abstract measure theory, with
some indication of its principal applications. The rst two volumes are set at an introductory level; they are
intended for students with a solid grounding in the concepts of real analysis, but possibly with rather limited
detailed knowledge. As the book proceeds, the level of sophistication and expertise demanded will increase;
thus for the volume on topological measure spaces, familiarity with general topology will be assumed. The
emphasis throughout is on the mathematical ideas involved, which in this subject are mostly to be found in
the details of the proofs.
My intention is that the book should be usable both as a rst introduction to the subject and as a reference
work. For the sake of the rst aim, I try to limit the ideas of the early volumes to those which are really
essential to the development of the basic theorems. For the sake of the second aim, I try to express these ideas
in their full natural generality, and in particular I take care to avoid suggesting any unnecessary restrictions
in their applicability. Of course these principles are to to some extent contradictory. Nevertheless, I nd that
most of the time they are very nearly reconcilable, provided that I indulge in a certain degree of repetition.
For instance, right at the beginning, the puzzle arises: should one develop Lebesgue measure rst on the
real line, and then in spaces of higher dimension, or should one go straight to the multidimensional case? I
believe that there is no single correct answer to this question. Most students will nd the one-dimensional
case easier, and it therefore seems more appropriate for a rst introduction, since even in that case the
technical problems can be daunting. But certainly every student of measure theory must at a fairly early
stage come to terms with Lebesgue area and volume as well as length; and with the correct formulations, the
multidimensional case diers from the one-dimensional case only in a denition and a (substantial) lemma.
So what I have done is to write them both out (114-115). In the same spirit, I have been uninhibited,
when setting out exercises, by the fact that many of the results I invite students to look for will appear in
later chapters; I believe that throughout mathematics one has a better chance of understanding a theorem
if one has previously attempted something similar alone.
As I write this Introduction (April 2004), the plan of the work is as follows:
Volume
Volume
Volume
Volume
Volume
1:
2:
3:
4:
5:
Volume 1 is intended for those with no prior knowledge of measure theory, but competent in the elementary
techniques of real analysis. I hope that it will be found useful by undergraduates meeting Lebesgue measure
for the rst time. Volume 2 aims to lay out some of the fundamental results of pure measure theory
(the Radon-Nikod
ym theorem, Fubinis theorem), but also gives short introductions to some of the most
important applications of measure theory (probability theory, Fourier analysis). While I should like to
believe that most of it is written at a level accessible to anyone who has mastered the contents of Volume 1,
I should not myself have the courage to try to cover it in an undergraduate course, though I would certainly
attempt to include some parts of it. Volumes 3 and 4 are set at a rather higher level, suitable to postgraduate
courses; while Volume 5 will assume a wide-ranging competence over large parts of analysis and set theory.
There is a disclaimer which I ought to make in a place where you might see it in time to avoid paying for
this book. I make no attempt to describe the history of the subject. This is not because I think the history
uninteresting or unimportant; rather, it is because I have no condence of saying anything which would not
be seriously misleading. Indeed I have very little condence in anything I have ever read concerning the
history of ideas. So while I am happy to honour the names of Lebesgue and Kolmogorov and Maharam in
more or less appropriate places, and I try to include in the bibliographies the works which I have myself
consulted, I leave any consideration of the details to those bolder and better qualied than myself.
The work as a whole is not yet complete; and when it is nished, it will be far too long to be printed as a
single volume in any reasonable format. I am therefore publishing it one part at a time. However, drafts of
most of the rest are available on the Internet; see http://www.essex.ac.uk/maths/staff/fremlin/mt.htm
for detailed instructions. For the time being, at least, printing will be in short runs. I hope that readers will
be energetic in commenting on errors and omissions, since it should be possible to correct these relatively
promptly. An inevitable consequence of this is that paragraph references may go out of date rather quickly.
Introduction to Volume 3
11
I shall be most attered if anyone chooses to rely on this book as a source for basic material; and I am
willing to attempt to maintain a concordance to such references, indicating where migratory results have
come to rest for the moment, if authors will supply me with copies of papers which use them.
I mention some minor points concerning the layout of the material. Most sections conclude with lists of
basic exercises and further exercises, which I hope will be generally instructive and occasionally entertaining. How many of these you should attempt must be for you and your teacher, if any, to decide, as no
two students will have quite the same needs. I mark with a > those which seem to me to be particularly
important. But while you may not need to write out solutions to all the basic exercises, if you are in any
doubt as to your capacity to do so you should take this as a warning to slow down a bit. The further
exercises are unbounded in diculty, and are unied only by a presumption that each has at least one
solution based on ideas already introduced. Occasionally I add a nal problem, a question to which I do
not know the answer and which seems to arise naturally in the course of the work.
The impulse to write this book is in large part a desire to present a unied account of the subject.
Cross-references are correspondingly abundant and wide-ranging. In order to be able to refer freely across
the whole text, I have chosen a reference system which gives the same code name to a paragraph wherever
it is being called from. Thus 132E is the fth paragraph in the second section of the third chapter of
Volume 1, and is referred to by that name throughout. Let me emphasize that cross-references are supposed
to help the reader, not distract her. Do not take the interpolation (121A) as an instruction, or even a
recommendation, to lift Volume 1 o the shelf and hunt for 121. If you are happy with an argument as it
stands, independently of the reference, then carry on. If, however, I seem to have made rather a large jump,
or the notation has suddenly become opaque, local cross-references may help you to ll in the gaps.
Each volume will have an appendix of useful facts, in which I set out material which is called on
somewhere in that volume, and which I do not feel I can take for granted. Typically the arrangement of
material in these appendices is directed very narrowly at the particular applications I have in mind, and is
unlikely to be a satisfactory substitute for conventional treatments of the topics touched on. Moreover, the
ideas may well be needed only on rare and isolated occasions. So as a rule I recommend you to ignore the
appendices until you have some direct reason to suppose that a fragment may be useful to you.
During the extended gestation of this project I have been helped by many people, and I hope that my
friends and colleagues will be pleased when they recognise their ideas scattered through the pages below.
But I am especially grateful to those who have taken the trouble to read through earlier drafts and comment
on obscurities and errors.
Introduction to Volume 3
One of the rst things one learns, as a student of measure theory, is that sets of measure zero are
frequently negligible in the straightforward sense that they can safely be ignored. This is not quite a
universal principle, and one of my purposes in writing this treatise is to call attention to the exceptional
cases in which negligible sets are important. But very large parts of the theory, including some of the topics
already treated in Volume 2, can be expressed in an appropriately abstract language in which negligible sets
have been factored out. This is what the present volume is about. A measure algebra is a quotient of
an algebra of measurable sets by a null ideal; that is, the elements of the measure algebra are equivalence
classes of measurable sets. At the cost of an extra layer of abstraction, we obtain a language which can give
concise and elegant expression to a substantial proportion of the ideas of measure theory, and which oers
insights almost everywhere in the subject.
It is here that I embark wholeheartedly on pure measure theory. I think it is fair to say that the
applications of measure theory to other branches of mathematics are more often through measure spaces
rather than measure algebras. Certainly there will be in this volume many theorems of wide importance
outside measure theory; but typically their usefulness will be in forms translated back into the language of
the rst two volumes. But it is also fair to say that the language of measure algebras is the only reasonable
way to discuss large parts of a subject which, as pure mathematics, can bear comparison with any.
In the structure of this volume I can distinguish seven working and two accessory chapters. The
accessory chapters are 31 and 35. In these I develop the theories of Boolean algebras and Riesz spaces (=
vector lattices) which are needed later. As in Volume 2 you have a certain amount of choice in the order in
which you take the material. Everything except Chapter 35 depends on Chapter 31, and everything except
12
Introduction to Volume 3
Chapters 31 and 35 depends on Chapter 32. Chapters 33, 34 and 36 can be taken in any order, but Chapter
36 relies on Chapter 35. (I do not mean that Chapter 33 is never referred to in Chapter 34, nor even that the
later chapters do not rely on results from Chapter 33. What I mean is that their most important ideas are
accessible without learning the material of Chapter 33 properly.) Chapter 37 depends on Chapters 35 and
36. Chapter 38 would be dicult to make sense of without some notion of what has been done in Chapter
33. Chapter 39 uses fragments of Chapters 35 and 36.
The rst half of the volume follows almost the only line permitted by the structure of the subject. If we
are going to study measure algebras at all, we must know the relevant facts about Boolean algebras (Chapter
31) and how to translate what we know about measure spaces into the new language (Chapter 32). Then
we must get a proper grip on the two most important theorems: Maharams theorem on the classication of
measure algebras (Chapter 33) and the von Neumann-Maharam lifting theorem (Chapter 34). Since I am
now writing for readers who are committed I hope, happily committed to learning as much as they can
about the subject, I take the space to push these ideas as far as they can easily go, giving a full classication
of closed subalgebras of probability algebras, for instance (333), and investigating special types of lifting
(345-346). I mention here three sections interpolated into Chapter 34 (342-344) which attack a subtle
and important question: when can we expect homomorphisms between measure algebras to be realizable in
terms of transformations between measure spaces, as discussed briey in 235 and elsewhere.
Chapters 36 and 37 are devoted to re-working the ideas of Chapter 24 on function spaces in the more
abstract context now available, and relating them to the general Riesz spaces of Chapter 35. I am concerned
here not to develop new structures, nor even to prove striking new theorems, but rather to oer new ways
of looking at the old ones. Only in the Ergodic Theorem (372) do I come to a really important new result.
Chapter 38 looks at two questions, both obvious ones to ask if you have been trained in twentieth-century
pure mathematics: what does the automorphism group of a measure algebra look like, and inside such an
automorphism group, what do the conjugacy classes look like? (The second question is a fancy way of asking
how to decide, given two automorphisms of one of the structures considered in this volume, whether they
are really dierent, or just copies of each other obtained by looking at the structure a dierent way up.)
Finally, in Chapter 39, I discuss what is known about the question of which Boolean algebras can appear
as measure algebras.
Concerning the prerequisites for this volume, we certainly do not need everything in Volume 2. The
important chapters there are 21, 23, 24, 25 and 27. If you are approaching this volume without having
read the earlier parts of this treatise, you will need the Radon-Nikod
ym theorem and product measures
(of arbitrary families of probability spaces), for Maharams theorem; a simple version of the martingale
theorem, for the lifting theorem; and an acquaintance with Lp spaces (particularly, with L0 spaces) for
Chapter 36. But I would recommend the results-only versions of Volumes 1 and 2 in case some reference
is totally obscure. Outside measure theory, I call on quite a lot of terms from general topology, but none
of the ideas needed are dicult (Baires and Tychonos theorems are the deepest); they are sketched in
3A3 and 3A4. We do need some functional analysis for Chapters 36 and 39, but very little more than was
already used in Volume 2, except that I now call on versions of the Hahn-Banach theorem (3A5).
In this volume I assume that readers have substantial experience in both real and abstract analysis, and
I make few concessions which would not be appropriate when addressing active researchers, except that
perhaps I am a little gentler when calling on ideas from set theory and general topology than I should be
with my own colleagues, and I continue to include all the easiest exercises I can think of. I do maintain
my practice of giving proofs in very full detail, not so much because I am trying to make them easier, but
because one of my purposes here is to provide a complete account of the ideas of the subject. I hope that
the result will be accessible to most doctoral students who are studying topics in, or depending on, measure
theory.
Note on second printing
For the second printing of this volume I have taken the opportunity to strengthen a theorem in 381
concerning automorphisms of Boolean algebras; I am grateful to B.D.Miller for allowing me to incorporate
his ideas into what is now 382. This has forced substantial changes in the rest of the chapter. I have
also added a discussion of the order-sequential topology on a Boolean algebra in 392, following Balcar
ski & Jech 98 and Balcar Jech & Paza
k p03, expanding some ideas previously in 393. A
Glowczyn
couple of results on the topology of convergence in measure have been added to 367, and a handful of new
Introduction to Volume 3
13
denitions are included. Besides these changes, there are many minor corrections (once again, I should thank
T.D.Austin for his help) and new exercises, and two of the problems (323Z and 381Z) mentioned in the
rst printing have been solved. Details may be found in http://www.essex.ac.uk/maths/staff/fremlin
/mterr3.02.ps.
14
Boolean algebras
Chapter 31
Boolean algebras
The theory of measure algebras naturally depends on certain parts of the general theory of Boolean
algebras. In this chapter I collect those results which will be useful later. Since many students encounter
the formal notion of Boolean algebra for the rst time in this context, I start at the beginning; and indeed I
include in the Appendix (3A2) a brief account of the necessary part of the theory of rings, as not everyone
will have had time for this bit of abstract algebra in an undergraduate course. But unless you nd the
algebraic theory of Boolean algebras so interesting that you wish to study it for its own sake in which case
you should perhaps turn to Sikorski 64 or Koppelberg 89 I do not think it would be very sensible to
read the whole of this chapter before proceeding to the main work of the volume in Chapter 32. Probably
311 is necessary to get an idea of what a Boolean algebra looks like, and a glance at the statements of
the theorems in 312 would be useful, but the later sections can wait until you have need of them, on the
understanding that apparently innocent formal manipulations may depend on concepts which take some
time to master. I hope that the cross-references will be suciently well-targeted to make it possible to read
this material in parallel with its applications.
311D
15
Boolean algebras
A X = X A = A for every A X,
so that (PX, , ) is a Boolean algebra and X is its identity. Q
Q
(b) Recall that an algebra of subsets of X (136E) is a family PX such that , X \ E for
every E , and E F for all E, F . In this case (, , ) is a Boolean algebra with zero and
identity X. P
P If E, F , then
E F = X \ ((X \ E) (X \ F )) ,
EF = (E (X \ F )) (F (X \ E)) .
Because and X = X \ both belong to , we can work through the identities in (a) above to see that ,
like PX, is a Boolean algebra. Q
Q
(c) Consider the ring Z2 = {0, 1}, with its ring operations +2 , given by setting
0 +2 0 = 1 +2 1 = 0,
0 +2 1 = 1 +2 0 = 1,
0 0 = 0 1 = 1 0 = 0,
1 1 = 1.
I leave it to you to check, if you have not seen it before, that this is a ring. Because 0 0 = 0 and 1 1 = 1,
it is a Boolean algebra.
311C Proposition Let A be a Boolean ring.
(a) a + a = 0, that is, a = a, for every a A.
(b) ab = ba for all a, b A.
proof (a) If a A, then
a + a = (a + a)(a + a) = a2 + a2 + a2 + a2 = a + a + a + a,
so we must have 0 = a + a.
(b) Now for any a, b A,
a + b = (a + b)(a + b) = a2 + ab + ba + b2 = a + ab + ba + b,
so
0 = ab + ba = ab + ab
and ab = ba.
311D Lemma Let A be a Boolean ring, I an ideal of A (3A2E), and a A \ I. Then there is a ring
homomorphism : A Z2 such that a = 1 and d = 0 for every d I.
proof (a) Let I be the family of those ideals J of A which include I and do not contain a. Then I has a
maximal element K say. P
PSApply Zorns lemma. Since I I, I =
6 . If J is a non-empty totally ordered
subset of I, then set J = J . If b, c J and d A, then there are J1 , J2 J such that b J1 and
c J2 ; now J = J1 J2 is equal to one of J1 , J2 , so belongs to J , and 0, b + c, bd all belong to J, so all
belong to J . Thus J A; of course I J and a
/ J , so J I and is an upper bound for J in I. As
J is arbitrary, the hypotheses of Zorns lemma are satised and I has a maximal element. Q
Q
(b) For b A set Kb = {d : d A, bd K}. The following are easy to check:
(i) K Kb for every b A, because K is an ideal.
(ii) Kb A for every b A. P
P 0 K Kb . If d, d Kb and c A then
b(d + d ) = bd + bd ,
b(dc) = (bd)c
belong to K, so d + d , dc Kb . Q
Q
(iii) If b A and a
/ Kb , then Kb I so Kb = K.
(iv) Now a2 = a
/ K, so a
/ Ka and Ka = K.
(v) If b A \ K then b
/ Ka , that is, ba = ab
/ K, and a
/ Kb ; consequently Kb = K.
16
Boolean algebras
311D
(vi) If b, c A \ K then c
/ Kb so bc
/ K.
(vii) If b, c A \ K then
bc(b + c) = b2 c + bc2 = bc + bc = 0 K,
(ii) If b K, c A \ K then
so b + c
/ K, while bc K, so
(bc) = 0 = b c.
c = (b + b) + c = b + (b + c)
/K
(b + c) = 1 = b +2 c,
(bc) = 0 = b c.
(b + c) = 1 = b +2 c,
(bc) = 0 = b c
(iii) Similarly,
if b A \ K and c K.
(iv) If b, c A \ K, then by (b-vi) and (b-vii) we have b + c K, bc
/ K so
(b + c) = 0 = b +2 c,
(bc) = 1 = b c.
Thus a 7 b
a is a ring homomorphism.
(b) If a A and a 6= 0, then by 311D, with I = {0}, there is a z Z such that z(a) = 1, that is, z b
a;
so that b
a 6= . This shows that the kernel of a 7 b
a is {0}, so that the homomorphism is injective (3A2Db).
(c) If A is a Boolean algebra, and z Z, then there is some a A such that z(a) = 1, so that
z(1A )z(a) = z(1A a) 6= 0 and z(1A ) 6= 0; thus b
1A = Z.
311F Remarks (a) For any Boolean ring A, I will say that the Stone space of A is the set Z of non-zero
ring homomorphisms from A to Z2 , and the canonical map a 7 b
a : A PZ is the Stone representation.
(b) Because the map a 7 b
a : A PZ is an injective ring homomorphism, A is isomorphic, as Boolean
ring, to its image E = {b
a : a A}, which is a subring of PZ. Thus the Boolean rings PX of 311Ba are
leading examples in a very strong sense.
(c) I have taken the set Z of the Stone representation to be actually the set of homomorphisms from A
onto Z2 . Of course we could equally well take any set which is in a natural one-to-one correspondence with
Z; a popular choice is the set of maximal ideals of A, since a subset of A is a maximal ideal i it is the
kernel of a member of Z, which is then uniquely dened.
311I
17
Boolean algebras
, \,
(a) Using the Stone representation, we can see that the elementary operations , , \, of set theory
all correspond to operations on A. If we set
a b = a + b + ab,
for a, b A, then we see that
a b = ab,
a \ b = a + ab,
ab = a+b
ad
b = b
abb(b
a bb) = b
a bb,
ad
b = b
a bb,
ac
\b = b
a \ bb,
ad
b = b
abb.
Consequently all the familiar rules for manipulation of , , etc. will apply also to
for instance,
a (b c) = (a b) (a c),
, ,
a (b c) = (a b) (a c)
as c is arbitrary, C renes E. Q
Q
311H The order structure of a Boolean ring Again treating a Boolean ring A as an algebra of sets,
we have a natural ordering on it, setting a b if ab = a, so that a b i b
a bb. This translation makes
it obvious that is a partial ordering on A, with least element 0, and with greatest element 1 i A is a
Boolean algebra. Moreover, A is a lattice (denition: 2A1Ad), with a b = sup{a, b}, a b = inf{a, b} for
all a, b A. Generally, for a0 , . . . , an A,
supin ai = a0 . . .
an ,
inf in ai = a0 . . .
an ;
suprema and inma of nite subsets A correspond to unions and intersections of the corresponding families
in the Stone space. (But suprema and inma of infinite subsets of A are a very dierent matter; see 313
below.)
It may be obvious, but it is nevertheless vital to recognise that when A is a ring of sets then agrees
with .
311I The topology of a Stone space: Theorem Let Z be the Stone space of a Boolean ring A, and
let T be
{G : G Z and for every z G there is an a A such that z b
a G}.
Then T is a topology on Z, under which Z is a locally compact zero-dimensional Hausdor space, and
E = {b
a : a A} is precisely the set of compact open subsets of Z. A is a Boolean algebra i Z is compact.
18
311I
Boolean algebras
S
proof (a) Because E is closed under , and E = Z (recall that Z is the set of surjective homomorphisms
from A to Z2 , so that every z Z is somewhere non-zero and belongs to some b
a), E is a topology base, and
T is a topology.
(b) T is Hausdor. P
P Take any distinct z, w Z. Then there is an a A such that z(a) 6= w(a);
let us take it that z(a) = 1, w(a) = 0. There is also a b A such that w(b) = 1, so that w(b + ab) =
w(b) +2 w(a)w(b) = 1 and w (b + ab)b; also
a(b + ab) = ab + a2 b = ab + ab = 0,
so
b
a (b + ab)b = (a(b + ab))b = b
0 = ,
and b
a, (b + ab)b are disjoint members of T containing z, w respectively. Q
Q
(c) If a A then b
a is compact. P
P Let F be an ultralter on Z containing b
a. For each b A,
z0 (b) = limzF z(b) must be dened in Z2 , since one of the sets {z : z(b) = 0}, {z : z(b) = 1} must belong
to F. If b, c A, then the set
F = {z : z(b) = z0 (b), z(c) = z0 (c), z(b + c) = z0 (b + c), z(bc) = z0 (bc)}
G = {b
a : a A, b
a E}.
S
Then G is a family of open subsets
S of Z and G = E, because E is open. But E is also compact, so there
is a nite G0 G such that E = G0 . Express G0 as {b
a0 , . . . , b
an }. Then
E=b
a0 . . . b
an = (a0 . . .
an )b.
Proposition Let X be a locally compact zero-dimensional Hausdor space. Then the set A of open-andcompact subsets of X is a subring of PX. If Z is the Stone space of A, there is a unique homeomorphism
: Z X such that b
a = 1 [a] for every a A.
proof (a) Because X is Hausdor, all its compact sets are closed, so every member of A is closed. Consequently a b, a \ b, a b and ab belong to A for all a, b A, and A is a subring of PX.
It will be helpful to know that A is a base for the topology of X. P
P If G X is open and x G,
then (because X is locally compact) there is a compact set K X such that x int K; now (because X is
zero-dimensional) there is an open-and-closed set a X such that x a G int K; because a is a closed
subset of a compact subset of X, it is compact, and belongs to A, while x a G. Q
Q
311L
Boolean algebras
19
P
P (i) If z Z and x, x X are distinct, then, because X is Hausdor, there is an open set G X
containing x and not containing x ; because A is a base for the topology of X, there is an a A such that
x a G, so that x
/ a. Now either z(a) = 1 and (z, x )
/ R, or z(a) = 0 and (z, x)
/ R. Thus R is the
graph of a function with domain included in Z and taking values in X.
so x a0 \ a and x
/ a. Thus (z, x) R and (z) = x is dened. As z is arbitrary, the domain of is the
whole of Z.
(iii) If x X, dene z : A Z2 by setting z(a) = 1 if x a, 0 otherwise. It is elementary to check
that z is a ring homomorphism form A to Z2 . To see that it takes the value 1, note that because A is a
base for the topology of X there is an a A such that x a, so that z(a) = 1. So z Z, and of course
(z, x) R. As x is arbitrary, is surjective.
(iv) If z, z Z and (z) = (z ), then, for any a A,
so z = z . Thus is injective. Q
Q
(c) For any a A,
It follows that
P (i) If G X is open, then (because A is a base for the topology
S is a homeomorphism. P
of X) G = {a : a A, a G} and
S
S
1 [G] = {1 [a] : a A, a G} = {b
a : a A, a G}
is an open subset of Z. As G is S
arbitrary, is continuous. (ii) On theSother hand, if G X and 1 [G] is
1
open, then [G] is of the form aA b
a for some A A, so that G = A is an open set in X. Accordingly
is a homeomorphism. Q
Q
(d) Finally, I must check the uniqueness of . But of course if : Z X is any function such that
1
[a] = b
a for every a A, then the graph of must be R, so = .
311K Remark Thus we have a correspondence between Boolean rings and zero-dimensional locally
compact Hausdor spaces which is (up to isomorphism, on the one hand, and homeomorphism, on the
other) one-to-one. Every property of Boolean rings which we study will necessarily correspond to some
property of zero-dimensional locally compact Hausdor spaces.
311L Complemented distributive lattices I have introduced Boolean algebras through the theory
of rings; this seems to be the quickest route to them from an ordinary undergraduate course in abstract
algebra. However there are alternative approaches, taking the order structure rather than the algebraic
operations as fundamental, and for the sake of an application in Chapter 35 I give the details of one of these.
Proposition Let A be a lattice such that
(i) (a b) c = (a c) (b c) for all a, b, c A;
(ii) there is a bijection a 7 a : A A which is order-reversing, that is, a b i b a , and
such that a = a for every a;
(iii) A has a least element 0 and a a = 0 for every a A.
Then A has a Boolean algebra structure for which a b i a b.
20
311L
Boolean algebras
Similarly, a b = (a b) . If a, b, c A then
for a, b A.
ab = a b
(c)(i) If a, b A then
(a + b) = (a b) (a b ) = (a a) (a b ) (b a) (b b )
= 0 (a b ) (b a) = (a b ) (a b).
So if a, b, c A then
(a + b) + c = ((a + b) c ) ((a + b) c)
= (a b c ) (a b c ) (a b c) (a b c);
as this last formula is symmetric in a, b and c, it is also equal to a + (b + c). Thus addition is associative.
(ii) For any a A,
a + 0 = 0 + a = (a 0) (a 0 ) = 0 (a 1) = a,
a + a = (a a ) (a a) = 0 0 = 0
so each element of A is its own additive inverse, and (A, +) is a group. It is abelian because , are
commutative.
(d) Because is associative and commutative, (A, ) is a commutative semigroup; also 1 is its identity,
because a 1 = a for every a A. As for the distributive law in A,
ab + ac = (a b (a c) ) ((a b) a c)
= (a b (a c )) ((a b ) a c)
= (a b a ) (a b c ) (a a c) (b a c)
= (a b c ) (b a c)
a b ab = a a b = a a b,
Remark It is the case that the Boolean algebra structure of A is uniquely determined by its order structure,
but I delay the proof to the next section (312L).
311X Basic exercises (a) Let A0 , . . . , An be sets. Show that
A0 . . . An = {x : #({i : i n, x Ai }) is odd}.
311 Notes
Boolean algebras
21
(b) Let X be a set, and PX. Show that the following are equiveridical: (i) is an algebra of subsets
of X; (ii) is a subring of PX (that is, contains and is closed under and ) and contains X; (iii) ,
X \ E for every E , and E F for all E, F .
(c) Let A be any Boolean ring. Let a 7 a be any bijection between A and a set B disjoint from A. Set
B = A B, and extend the addition and multiplication of A to form binary operations on B by using the
formulae
a + b = a + b = (a + b) ,
a b = b + ab,
a + b = a + b,
ab = a + ab, a b = (a + b + ab) .
a (b c) = (a b) (a c)
for all a, b, c A directly from the denitions in 311G, without using Stones theorem.
> (g) Let A be any Boolean ring. Show that if we regard the Stone space Z of A as a subset of {0, 1}A ,
then the topology of Z (311I) is just the subspace topology induced by the ordinary product topology of
{0, 1}A .
(h) Let I be any set, and set X = {0, 1}I with its usual topology (3A3K). Show that for a subset E of
X the following are equiveridical: (i) E is open-and-compact; (ii) E is determined by coordinates in a nite
subset of I (denition: 254M); (iii) E belongs to the algebra of subsets of X generated by {Ei : i I},
where Ei = {x : x(i) = 1} for each i.
(i) Let (A, ) be a lattice such that () A has a least element 0 and a greatest element 1 () for every
a, b, c A, a (b c) = (a b) (a c) and a (b c) = (a b) (a c) () for every a A there is an
a A such that a a = 1 and a a = 0. Show that there is a Boolean algebra structure on A for which
agrees with .
311Y Further exercises (a) Let A be a Boolean ring, and B the Boolean algebra constructed by the
method of 311Xc. Show that the Stone space of B can be identied with the one-point compactication
(3A3O) of the Stone space of A.
(b) Let (A, , , 0, 1) be such that (i) (A, ) is a commutative semigroup with identity 0 (ii) (A, ) is a
commutative semigroup with identity 1 (iii) a (b c) = (a b) (a c), a (b c) = (a b) (a c) for all
a, b, c A (iv) a a = a a = a for every a A (v) for every a A there is an a A such that a a = 1,
a a = 0. Show that there is a Boolean algebra structure on A for which = , = .
311 Notes and comments My aim in this section has been to get as quickly as possible to Stones theorem,
since this is surely the best route to a picture of general Boolean algebras; they are isomorphic to algebras
of sets. This means that all their elementary algebraic properties indeed, all their rst-order properties
can be eectively studied in the context of elementary set theory. In 311G-311H I describe a few of the
ways in which the Stone representation suggests algebraic properties of Boolean algebras.
You should not, however, come too readily to the conclusion that Boolean algebras will never be able to
surprise you. In this book, in particular, we shall need to work a good deal with suprema and inma of
innite sets in Boolean algebras, for the ordering of 311H; and even though
this corresponds to the ordering
S
of ordinary sets, we nd that (sup A)b is suciently dierent from aA b
a to need new kinds of intuition.
22
Boolean algebras
311 Notes
S
(The point is that aA b
a is an open set in the Stone space, but need not be compact if A is innite, so
may well be smaller than (sup A)b, even when sup A is dened in A.) There is also the fact that Stones
theorem depends crucially on a fairly strong form of the axiom of choice (employed through Zorns Lemma
in the argument of 311D). Of course I shall be using the axiom of choice without scruple throughout this
volume. But it should be clear that such results as 312B-312C in the next section cannot possibly need the
axiom of choice for their proofs, and that to use Stones theorem in such a context is slightly misleading.
Nevertheless, it is so useful to be able to regard a Boolean algebra as an algebra of sets especially when
dealing with only nitely many elements of the algebra at a time that henceforth I will almost always use
the symbols , for the addition and multiplication of a Boolean ring, and will use , \ , without
further comment, just as if I were considering , \ and in the Stone space. (In 311Gb I have given a
denition of disjointness in a Boolean algebra based on the same idea.) Even without the axiom of choice
this approach can be justied, once we have observed that nitely-generated Boolean algebras are nite
(311Xd), since relatively elementary methods show that any nite Boolean algebra is isomorphic to PX for
some nite set X.
I have taken a Boolean algebra to be a particular kind of commutative ring with identity. Of course
there are other approaches. If we wish to think of the order relation as primary, then 311L and 311Xi are
reasonably natural. Other descriptions can be based on a list of the properties of the binary operations ,
and the complementation operation a 7 a = 1 \ a, as in 311Yb. I give extra space to 311L only because
this is well adapted to an application in 352Q below.
312 Homomorphisms
I continue the theory of Boolean algebras with a section on subalgebras, ideals and homomorphisms.
From now on, I will relegate Boolean rings which are not algebras to the exercises; I think there is no need
to set out descriptions of the triing modications necessary to deal with the extra generality. The rst part
of the section (312A-312K) concerns the translation of the basic concepts of ring theory into the language
which I propose to use for Boolean algebras. 312L shows that the order relation on a Boolean algebra denes
the algebraic structure, and in 312M-312N I give a fundamental result on the extension of homomorphisms.
I end the section with results relating the previous ideas to the Stone representation of a Boolean algebra
(312O-312S).
312A Subalgebras Let A be a Boolean algebra. I will use the phrase subalgebra of A to mean a
subring of A containing its multiplicative identity 1 = 1A .
312B Proposition Let A be a Boolean algebra, and B a subset of A. Then the following are equiveridical,
that is, if one is true so are the others:
(i) B is a subalgebra of A;
(ii) 0 B, a b B for all a, b B, and 1 \ a B for all a B;
(iii) B 6= , a b B for all a, b B, and 1 \ a B for all a B.
proof (a)(i)(iii) If B is a subalgebra of A, and a, b B, then of course we shall have
0, 1 B, so B 6= ,
a b B,
1 \ a = 1 a B.
0 = b0 (1 \ b0 ) B.
a b = 1 \ ((1 \ a) (1 \ b)) B.
312G
Homomorphisms
23
a b = 1 \ ((1 \ a) (1 \ b)) B,
a b = (a (1 \ b)) (b (1 \ a)) B,
so (because also 0 B) B is a subring of A, and
1 = 1 \ 0 B,
so B is a subalgebra.
Remark Thus an algebra of subsets of a set X, as dened in 136E or 311Bb, is just a subalgebra of the
Boolean algebra PX.
312C Ideals in Boolean algebras: Proposition If A is a Boolean algebra, a set I A is an ideal of
A i 0 I, a b I for all a, b I, and a I whenever b I and a b.
proof (a) Suppose that I is an ideal. Then of course 0 I. If a, b I then a b I so a b =
(a b) (a b) I. If b I and a b then a = a b I.
(b) Now suppose that I satises the conditions proposed. If a, b I then
ab
ab I
a I,
so b a = a b I; thus I is an ideal.
312D Principal ideals Of course, while an ideal I in a Boolean algebra A is necessarily a subring, it is
not as a rule a subalgebra, except in the special case I = A. But if we say that a principal ideal of A is
the ideal Aa generated by a single element a of A, we have a special phenomenon.
312E Proposition Let A be a Boolean algebra, and a any element of A. Then the principal ideal Aa of
A generated by a is just {b : b A, b a}, and (with the inherited operations Aa Aa , Aa Aa ) is
a Boolean algebra in its own right, with multiplicative identity a.
proof b
a i b a = a, so that
Aa = {b : b
a} = {b a : b A}
is an ideal of A, and of course it is the smallest ideal of A containing a. Being an ideal, it is a subring; the
idempotent relation b b = b is inherited from A, so it is a Boolean ring; and a is plainly its multiplicative
identity.
312F Boolean homomorphisms Now suppose that A and B are two Boolean algebras. I will use the
phrase Boolean homomorphism to mean a function : A B which is a ring homomorphism (that is,
(a b) = a b, (a b) = a b for all a, b A) which is uniferent, that is, (1A ) = 1B .
312G Proposition Let A, B and C be Boolean algebras.
(a) If : A B is a Boolean homomorphism, then [A] is a subalgebra of B.
(b) If : A B and : B C are Boolean homomorphisms, then : A C is a Boolean
homomorphism.
(c) If : A B is a bijective Boolean homomorphism, then 1 : B A is a Boolean homomorphism.
proof These are all immediate consequences of the corresponding results for ring homomorphisms (3A2D).
24
312H
Boolean algebras
312H Proposition Let A and B be Boolean algebras, and : A B a function. Then the following
are equiveridical:
(i) is a Boolean homomorphism;
(ii) (a b) = a b and (1A \ a) = 1B \ a for all a, b A;
(iii) (a b) = a b and (1A \ a) = 1B \ a for all a, b A;
(iv) (a b) = a b and a b = 0B whenever a, b A and a b = 0A , and (1A ) = 1B .
proof (i)(iv) If is a Boolean homomorphism then of course (1A ) = 1B ; also, given that a b = 0 in A,
a b = (a b) = (0A ) = 0B ,
(a b) = (a b) = a b = a b.
(iv)(iii) Assume (iv), and take a, b A. Then
a = (a b) (a \ b),
b = (a b) (b \ a),
so
(a b) = a (b \ a) = (a b) (a \ b) (b \ a) = a b.
Taking b = 1 \ a, we must have
1B = (1A ) = a (1A \ a),
0B = a (1A \ a),
(a b) = a b ,
(a b) = a b ,
(a \ b) = a \ b
312M
25
Homomorphisms
b a \ b I.
{u : u a } = {b : b
a}.
for every a A, A/I is a Boolean ring; because 1 is a multiplicative identity, it is a Boolean algebra, and
a 7 a is a Boolean homomorphism. The formulae given are now elementary.
(b) We have
a
Now
{u : u
b a \ b = 0 a \ b I.
a } = {u a : u A/I} = {(b a) : b A} = {b : b
a}.
312L The above results are both repetitive and nearly trivial. Now I come to something with a little
more meat to it.
Proposition If A and B are Boolean algebras and : A B is a bijection such that a b whenever
a b, then is a Boolean algebra isomorphism.
proof (a) Because is surjective, there must be c0 , c1 A such that c0 = 0B , c1 = 1B ; now (0A )
c1 (1A ), so we must have (0A ) = 0B , (1A ) = 1B .
c0 ,
c a = 0B ,
(c \ a)
c (1A \ a) = 0B ;
c a = 0B ,
(c b)
c b = 0B ,
26
312M
Boolean algebras
(a c) (b \ c) (a c) (b \ c) = ((a a ) c) ((b b ) \ c) A1
so A1 includes A0 {c}; and nally it is clear that any subalgebra of A including A0 {c}, being closed under
, and complementation, must include A1 , so that A1 is the subalgebra of A generated by A0 {c}.
312N Lemma Let A and B be Boolean algebras, A0 a subalgebra of A, : A0 B a Boolean
homomorphism, and c A. If v B is such that a v b whenever a, b A0 and a c b, then there
is a unique Boolean homomorphism 1 from the subalgebra A1 of A generated by A0 {c} such that 1
extends and 1 c = v.
proof (a) The basic fact we need to know is that if a, a , b, b A0 and
(a c) (b \ c) = d = (a c) (b \ c),
then
(a v) (b \ v) = (a v) (b \ v).
P
P We have
a c = d c = a c.
Accordingly (a a ) c = 0 and c 1 \ (a a ). Consequently (since a a surely belongs to A0 )
v
(1 \ (a a )) = 1 \ (a a ),
and
a v = a v.
Similarly,
b \ c = d \ c = b \ c,
so
(b b ) \ c = 0,
b b
c,
(b b ) v
and
b \ v = b \ v.
Putting these together, we have the result. Q
Q
(b) Consequently, we have a function 1 dened by writing
1 ((a c) (b \ c)) = (a v) (b \ v)
for all a, b A0 ; and 312M tells us that the domain of 1 is just A1 . Now 1 is a Boolean homomorphism.
P
P This amounts to running through the proof of 312M again.
(i) If a, b A0 , then
1 (1 \ ((a c) (b \ c))) = 1 (((1 \ a) c) ((1 \ b) \ c))
= ((1 \ a) v) ((1 \ b) \ v)
= ((1 \ a) v) ((1 \ b) \ v)
= 1 \ ((a v) (b \ v)) = 1 \ 1 ((a c) (b \ c)).
312P
Homomorphisms
27
So 1 (1 \ d) = 1 \ 1 d for every d A1 .
(ii) If a, b, a , b A0 , then
1 ((a c) (b \ c) (a c) (b \ c)) = 1 (((a a ) c) ((b b ) \ c))
= ((a a ) v) ((b b ) \ v)
= ((a a ) v) ((b b ) \ v)
= (a v) (b \ v) (a v) (b \ v)
So 1 (d d ) = 1 d 1 d for all d, d A1 .
By 312H(iii), 1 is a Boolean homomorphism. Q
Q
(c) If a A0 , then
1 a = 1 ((a c) (a \ c)) = (a v) (a \ v) = a,
proof (a) Recall that I have constructed Z, W as the sets of Boolean homomorphisms from A, B to
Z2 (311E). So if : A B is any Boolean homomorphism, and w W , (w) = w is a Boolean
homomorphism from A to Z2 (312Gb), and belongs to Z. Now 1 [b
a] =
ca for every a A. P
P
28
Boolean algebras
312P
1 [b
a] = {w : (w) b
a} = {w : w b
a} = {w : w(a) = 1} = {w : w
ca}. Q
Q
S
Consequently is continuous. P
P Let G be any open subset of Z. Then G = {b
a:b
a G}, so
S
S 1
1
a] : b
a G} = {c
a : b
a G}
[G] = { [b
is open. As G is arbitrary, is continuous. Q
Q
so (w) = w .
Now is a Boolean homomorphism. P
P (i) If a, b A then
c
(a b)b = 1 [(a b)b] = 1 [b
a bb] = 1 [b
a] 1 [bb] = d
a b = ( a b)b,
so (a b) = a b. (ii) If a A, then
P For a A,
(c) For any Boolean homomorphism : A B, = . P
Q
so a = a. Q
a] =
ca,
( a)b = 1 [b
d = ((a))b = 1 [c
a
a] = 1 [1 [b
a]] = ()1 [b
a].
312R Proposition Let A and B be Boolean algebras, with Stone spaces Z and W , and : A B a
Boolean homomorphism, with associated continuous function : W Z. Then
(a) is injective i is surjective;
(b) is surjective i is injective.
proof (a) If a A, then
b
a [W ] = (w)
/b
a for every w W
312Xf
Homomorphisms
29
is not surjective Z \ [W ] 6=
(3A2Db).
(b)(i) If is surjective and w, w are distinct members of W , then there is a b B such that w bb
and w
/ bb. Now b = a for some a A, so (w) b
a and (w )
/b
a, and (w) 6= (w ). As w and w are
arbitrary, is injective.
(ii) If is injective and b B,Sthen K = [bb], L = [W \ bb] are disjoint compact subsets of Z. Consider
I = {a : a A, L b
a = }. Then aI b
a = Z \ L K. Because K is compact and every b
a is open, there
is a nite family a0 , . . . , an I such that K b
a0 . . . b
an . Set a = a0 . . . an . Then b
a=b
a0 . . . b
an
includes K and is disjoint from L. So
ca = 1 [b
a] includes bb and is disjoint from W \ bb; that is,
ca = bb and
a = b. As b is arbitrary, is surjective.
312S Principal ideals If A is a Boolean algebra and a A, we have a natural surjective Boolean
homomorphism b 7 b a : A Aa , the principal ideal generated by a (312J). Writing Z for the Stone space
of A and Za for the Stone space of Aa , this homomorphism must correspond to an injective continuous
function : Za Z (312Rb). Because Za is compact and Z is Hausdor, must be a homeomorphism
between Za and its image [Za ] Z (3A3Dd). To identify [Za ], note that it is compact, therefore closed,
and that
Z \ [Za ] =
=
{bb : b A, bb [Za ] = }
[
{bb : 1 [bb] = } = {bb : b a = 0} = Z \ b
a,
so that [Za ] = b
a. It is therefore natural to identify Za with the open-and-closed set b
a Z.
312T Fixed-point subalgebras For future reference I introduce the following idea. If A is a Boolean
algebra and : A A is a Boolean homomorphism, then {a : a A, a = a} is a subalgebra of A (put
312A and 312F together); I will call it the fixed-point subalgebra of .
312X Basic exercises (a) Let A be a Boolean ring, and B a subset of A. Show that B is a subring of
A i 0 B and a b, a \ b B for all a, b B.
(b) Let A be a Boolean algebra and B a subset of A. Show that B is a subalgebra of A i 1 B and
a \ b B for all a, b B.
(c) Let A be a Boolean algebra. Suppose that I A A are such that 1 A, a b I for all a, b I
and a \ b A whenever a, b A and b a. Show that A includes the subalgebra of A generated by I.
(Hint: 136Xf.)
(d) Show that if A is a Boolean ring, a set I A is an ideal of A i 0 I, a b I for all a, b I, and
a I whenever b I and a b.
(e) Let A and B be Boolean algebras, and : A B a function such that (i) (a) (b) whenever
a b (ii) (a) (b) = 0B whenever a b = 0A (iii) (a) (b) (c) = 1B whenever a b c = 1A . Show
that is a Boolean homomorphism.
(f ) Let A be a Boolean ring, and a any member of A. Show that the map b 7 ab is a ring homomorphism
from A onto the principal ideal Aa generated by a.
30
Boolean algebras
312Xg
(g) Let A1 and A2 be Boolean rings, and let B1 , B2 be the Boolean algebras constructed from them
by the method of 311Xc. Show that any ring homomorphism from A1 to A2 has a unique extension to a
Boolean homomorphism from B1 to B2 .
(h) Let A and B be Boolean rings, A0 a subalgebra of A, : A0 B a ring homomorphism, and c A.
Show that if v B is such that a \ v = b v = 0 whenever a, b A0 and a \ c = b c = 0, then there is
a unique ring homomorphism 1 from the subring A1 of A generated by A0 {c} such that 1 extends 0
and 1 c = v.
(i) Let A be a Boolean ring, and Z its Stone space. Show that there
S is a one-to-one correspondence
between ideals I of A and open sets G Z, given by the formulae G = aI b
a, I = {a : b
a G}.
(j) Let A be a Boolean algebra, and suppose that A is the subalgebra of itself generated by A0 {c},
where A0 is a subalgebra of A and c A. Let Z be the Stone space of A and Z0 the Stone space of A0 . Let
: Z Z0 be the continuous surjection corresponding to the embedding of A0 in A. Show that b
c and
Z \ b
c are injective.
Now let B be another Boolean algebra, with Stone space W , and : A0 B a Boolean homomorphism,
with corresponding function : W Z0 . Show that there is a continuous function 1 : W Z such that
1 = i there is an open-and-closed set V W such that [V ] [b
c] and [W \ V ] [Z \ b
c].
(k) Let A be a Boolean algebra, with Stone space Z, and I an ideal of A, corresponding to an open set
G Z. Show that the Stone space of the quotient algebra A/I may be identied with Z \ G.
312Y Further exercises (a) Find a function : P{0, 1, 2} Z2 such that (1 \ a) = 1 \ a for every
a P{0, 1, 2} and (a) (b) whenever a b, but is not a Boolean homomorphism.
(b) Let A be the Boolean ring of nite subsets of N. Show that there is a bijection : A A such that
a b whenever a b but is not a ring homomorphism.
(c) Let A, B be Boolean rings, with Stone spaces Z, W . Show that we have a one-to-one correspondence
between ring homomorphisms : A B and continuous functions : H Z, where H W is an open set,
such that 1 [K] is compact for every compact set K Z, given by the formula a = b 1 [b
a] = bb.
(d) Let A, B, C be Boolean rings, with Stone spaces Z, W and V . Let : A B and : B C be
ring homomorphisms, with corresponding continuous functions : H Z and : G W . Show that the
ring homomorphism : A C corresponds to the continuous function : 1 [H] Z.
(e) Let A and B be Boolean rings, with Stone spaces Z and W , and : A B a ring homomorphism,
with associated continuous function : H Z. Show that is injective i [H] is dense in Z, and that
is surjective i is injective and H = W .
(f ) Let A be a Boolean ring and a A. Show that the Stone space of the principal ideal Aa of A generated
by a can be identied with the compact open set b
a in the Stone space of A. Show that the identity map is
a ring homomorphism from Aa to A, and corresponds to the identity function on b
a.
312 Notes and comments The denitions of subalgebra and Boolean homomorphism (312A, 312F),
like that of Boolean algebra, are a trie arbitrary, but will be a convenient way of mandating appropriate
treatment of multiplicative identities. I run through the work of 312A-312J essentially for completeness;
once you are familiar with Boolean algebras, they should all seem obvious. 312L has a little bit more to it.
It shows that the order structure of a Boolean algebra denes the ring structure, in a fairly strong sense.
I call 312N a lemma, but actually it is the most important result in this section; it is the basic tool we have
for extending a homomorphism from a subalgebra to a slightly larger one, and with Zorns Lemma (another
lemma which deserves a capital L) will provide us with general methods of constructing homomorphisms.
In 312O-312S I describe the basic relationships between the Boolean homomorphisms and continuous
functions on Stone spaces. 312P-312Q show that, in the language of category theory, the Stone representation
313B
Order-continuous homomorphisms
31
provides a contravariant functor from the category of Boolean algebras with Boolean homomorphisms to the
category of topological spaces with continuous functions. Using 311I-311J, we know exactly which topological
spaces appear, the zero-dimensional compact Hausdor spaces; and we know also that the functor is faithful,
that is, that we can recover Boolean algebras and homomorphisms from the corresponding topological spaces
and continuous functions. There is an agreeable duality in 312R. All of this can be done for Boolean rings,
but there are some extra complications (312Yc-312Yf).
To my mind, the very essence of the theory of Boolean algebras is the fact that they are abstract rings,
but at the same time can be thought of locally as algebras of sets. Consequently we can bring two
quite separate kinds of intuition to bear. 312N gives an example of a ring-theoretic problem, concerning the
extension of homomorphisms, which has a resolution in terms of the order relation, a concept most naturally
described in terms of algebras-of-sets. It is very much a matter of taste and habit, but I myself nd that a
Boolean homomorphism is easiest to think of in terms of its action on nite subalgebras, which are directly
representable as PX for some nite X (311Xe); the corresponding continuous map between Stone spaces is
less helpful. I oer 312Xj, the Stone-space version of 312N, for you to test your own intuitions on.
(a0 \ e) (e \ c)
e \ a0 .
Remark In the arguments above I repeatedly encourage you to treat , , \ , as if they were the
corresponding operations and relation of basic set theory. This is perfectly safe so long as we take care that
every manipulation so justied has only nitely many elements of the Boolean algebra in hand at once.
313B General distributive laws: Proposition Let A be a Boolean algebra.
(a) If e A and A A is a non-empty set such that sup A is dened in A, then sup{e a : a A} is
dened and equal to e sup A.
(b) If e A and A A is a non-empty set such that inf A is dened in A, then inf{e a : a A} is
dened and equal to e inf A.
(c) Suppose that A, B A are non-empty and sup A, sup B are dened in A. Then sup{a b : a A, b
B} is dened and is equal to sup A sup B.
32
313B
Boolean algebras
(d) Suppose that A, B A are non-empty and inf A, inf B are dened in A. Then inf{a b : a A, b B}
is dened and is equal to inf A inf B.
proof (a) Set
Using 313A, we have
B = {e \ a : a A},
C = {e \ b : b B} = {e a : a A}.
inf B = e \ sup A,
as required.
(b) Set a0 = inf A, B = {e a : a A}. Then e a0 e a for every a A, so e a0 is a lower bound
for B. If c is any lower bound for B, then c \ e a for every a A, so c \ e a0 and c e a0 ; thus e a0
is the greatest lower bound for B, as claimed.
(c) By (a), we have
a sup B = supbB a b
for every a A, so
b
a bb
b
a bb
aA
aA
S
because bb is certainly closed in Z. It follows at once that if b
a0 is actually equal to aA b
a then a0 must be
S
S
ab
a0 . ?? If b
a0 6= aA b
a,
the least upper bound of A in A. On the other hand, if a0 = sup A, then aA b
S
a is a non-empty open set in Z, so includes bb for some non-zero b A; now b
ab
a0 \ bb, so
then b
a0 \ aA b
a a0 \ S
b for every a A, and a0 \ b is an upper bound for A strictly less than a0 . X
X Thus b
a0 must be
exactly aA b
a.
(b) Take complements: setting a1 = 1 \ a0 , we have
a0 = inf A a1 = sup 1 \ a
aA
(by 313A)
b
a1 =
(c) Since
a
aA b
aA
b
a0 = Z \
Z \b
a
[
aA
Z \b
a = int
aA
b
a.
is surely a closed set, it is nowhere dense i it has empty interior, that is, i 0 = inf A.
313E
Order-continuous homomorphisms
33
313D I started the section with the results above because they are easily stated and of great importance.
But I must now turn to some new denitions, and I think it may help to clarify the ideas involved if I give
them in their own natural context, even though this is far more general than we have any immediate need
for here.
Definitions Let P be a partially ordered set and C a subset of P .
(a) C is order-closed if sup A C whenever A is a non-empty upwards-directed subset of C such that
sup A is dened in P , and inf A C whenever A is a non-empty downwards-directed subset of C such that
inf A is dened in P .
(b) C is sequentially order-closed if supnN pn C whenever hpn inN is a non-decreasing sequence in
C such that supnN pn is dened in P , and inf nN pn C whenever hpn inN is a non-increasing sequence in
C such that inf nN pn is dened in P .
Remark I hope it is obvious that an order-closed set is sequentially order-closed.
313E Order-closed subalgebras and ideals Of course, in the very special cases of a subalgebra
or ideal of a Boolean algebra, the concepts order-closed and sequentially order-closed have expressions
simpler than those in 313D. I spell them out.
(a) Let B be a subalgebra of a Boolean algebra A.
(i) The following are equiveridical:
() B is order-closed in A;
() sup B B whenever B B and sup B is dened in A;
( ) inf B B whenever B B and inf B is dened in A;
() sup B B whenever B B is non-empty and upwards-directed and sup B is dened in A;
( ) inf B B whenever B B is non-empty and downwards-directed and inf B is dened in A.
P
P Of course () (). If () is true and B B is any set with a supremum in A, then B =
{0} {b0 . . . bn : b0 , . . . , bn B} is a non-empty upwards-directed set with the same upper bounds as
B, so sup B = sup B B. Thus () () and (), () are equiveridical. Next, if () is true and B B is
a set with an inmum in A, then B = {1 \ b : b B} B and sup B = 1 \ inf B is dened, so sup B and
inf B belong to B . Thus () ( ). In the same way, ( ) ( ) () and (), ( ), (), ( ) are all
equiveridical. But since we also have () ()&( ), () is equiveridical with the others. Q
Q
Replacing the sets B above by sequences, the same arguments provide conditions for B to be sequentially
order-closed, as follows.
A;
A.
(b) Now suppose that I is an ideal of A. Then if A I is non-empty all lower bounds of A necessarily
belong to I; so that
I is order-closed i sup A I whenever A I is non-empty, upwards-directed and has a
supremum in A;
I is sequentially order-closed i supnN an I whenever han inN is a non-decreasing sequence
in I with a supremum in A.
Moreover, because I is closed under ,
I is order-closed i sup A I whenever A I has a supremum in A;
34
Boolean algebras
313E
(c) If A = PX is a power set, then a sequentially order-closed subalgebra of A is just a -algebra of sets,
while a sequentially order-closed ideal of A is a what I have called a -ideal of sets (112Db). If A is itself a
-algebra of sets, then a sequentially order-closed subalgebra of A is a -subalgebra in the sense of 233A.
Accordingly I will normally use the phrases -subalgebra, -ideal for sequentially order-closed subalgebras and ideals of Boolean algebras.
313F Order-closures and generated sets (a) It is an immediate consequence
T of the denitions that
(i) if S is any non-empty family of subalgebras of a Boolean algebra A, then S is a subalgebra
T of A;
(ii) if F is any non-empty family of order-closed subsets of a partially ordered set P , then F is an
order-closed subset of P ;
T (iii) if F is any non-empty family of sequentially order-closed subsets of a partially ordered set P , then
F is a sequentially order-closed subset of P .
(b) Consequently, given any Boolean algebra A and a subset B of A, we have a smallest subalgebra B of
A including B, being the intersection of all the subalgebras of A which include B; a smallest -subalgebra
B of A including B, being the intersection of all the -subalgebras of A which include B; and a smallest
order-closed subalgebra B of A including B, being the intersection of all the order-closed subalgebras of
A which include B. We call B, B and B the subalgebra, -subalgebra and order-closed subalgebra
generated by B. (I shall return to this in 331E.)
(c) If A is a Boolean algebra and B any subalgebra of A, then the smallest order-closed subset B of A
which includes B is again a subalgebra of A (so is the order-closed subalgebra of A generated by B). P
P (i)
The set {b : 1 \ b B} is order-closed (use 313A) and includes B, so includes B; thus 1 \ b B for every
b B. (ii) If c B, the set {b : b c B} is order-closed (use 313Bb) and includes B, so includes B; thus
b c B whenever b B and c B. (iii) If c B, the set {b : b c B} is order-closed and includes B
(by (ii)), so includes B; thus b c B whenever b, c B. (iv) By 312B, B is a subalgebra of A. Q
Q
313G This is a convenient moment at which to spell out an abstract version of the Monotone Class
Theorem (136B).
Lemma Let A be a Boolean algebra.
(a) Suppose that 1 I A A and that
a b I for all a, b I,
b \ a A whenever a, b A and a b.
P
P If c B, then of course c b B A for every b B, because B P. If c A \ B, consider
J = B {c b : b B}.
(c b1 ) b2 = b1 (c b2 ) = (c b1 ) (c b2 ) = c (b1 b2 ) J,
313I
Order-continuous homomorphisms
35
c. P
P If a B, then b a, c a B A and
(c \ b) a = (c a) \ (b a) A
so 1 \ b B. If a, b B, then
1 I B,
a b = 1 \ ((1 \ a) (1 \ b)) B.
36
313I
Boolean algebras
(p) B and p 1 [B]. Similarly, if A 1 [B] is non-empty and downwards-directed, and inf A is
dened in P , then (inf A) = inf [A] B and inf A 1 [B]. Thus 1 [B] is order-closed; as B is
arbitrary, satises the condition.
(ii) Now suppose that 1 [B] is order-closed in P whenever B Q is order-closed in Q. Let A P
be a non-empty upwards-directed subset of P with a supremum p P . Then (p) is an upper bound of
[A]. Let q be any upper bound of [A] in Q. Consider B = {r : r q}; then B Q is upwards-directed
and order-closed, so 1 [B] is order-closed. Also A 1 [B] is non-empty and upwards-directed and has
supremum p, so p 1 [B] and (p) B, that is, (p) q. As q is arbitrary, (p) = sup [A]. Similarly,
(inf A) = inf [A] whenever A P is non-empty, downwards-directed and has an inmum in P ; so is
order-continuous.
313J It is useful to introduce here the following notion.
Definition Let A be a Boolean algebra. A set D A is order-dense if for every non-zero a A there is
a non-zero d D such that d a.
Remark Many authors use the simple word dense where I have insisted on the phrase order-dense. In
the work of this treatise it will be important to distinguish clearly between this concept of dense set and
the topological concept (2A3U); typically, in those contexts in which both appear, an order-dense set can
be in some sense much smaller than a topologically dense set.
313K Lemma If A is a Boolean algebra and D A is order-dense, then for any a A there is a disjoint
C D such that sup C = a; in particular, a = sup{d : d D, d a} and there is a partition of unity
C D.
proof Set Da = {d : d D, d a}. Applying Zorns lemma to the family C of disjoint sets C Da , we
have a maximal C C. Now if b A and b 6 a, there is a d D such that 0 6= d a \ b. Because C is
maximal, there must be a c C such that c d 6= 0, so that c 6 b. Turning this round, any upper bound of
C must include a, so that a = sup C. It follows at once that a = sup Da .
Taking a = 1 we obtain a partition of unity included in D.
313L Proposition Let A and B be Boolean algebras and : A B a Boolean homomorphism.
(a) is order-preserving.
(b) The following are equiveridical:
(i) is order-continuous;
(ii) whenever A A is non-empty and downwards-directed and inf A = 0 in A, then inf [A] = 0 in B;
(iii) whenever A A is non-empty and upwards-directed and sup A = 1 in A, then sup [A] = 1 in B;
(iv) whenever A A and sup A is dened in A, then (sup A) = sup [A] in B;
(v) whenever A A and inf A is dened in A, then (inf A) = inf [A] in B;
(vi) whenever C A is a partition of unity, then [C] is a partition of unity in B.
(c) The following are equiveridical:
(i) is sequentially order-continuous;
(ii) whenever han inN is a non-increasing sequence in A and inf nN an = 0 in A, then inf nN an = 0
in B;
(iii) whenever A A is countable and sup A is dened in A, then (sup A) = sup [A] in B;
(iv) whenever A A is countable and inf A is dened in A, then (inf A) = inf [A] in B;
(v) whenever C A is a countable partition of unity, then [C] is a partition of unity in B.
proof (a) This is 312I.
(b)(i)(ii) is trivial, as 0 = 0.
(ii)(iv) Assume (ii), and let A be any subset of A such that c = sup A is dened in A. If A = ,
then c = 0 and sup [A] = 0 = c. Otherwise, set
A = {a0 . . .
an
: a0 , . . . , an A},
C = {c \ a : a A }.
313M
37
Order-continuous homomorphisms
Then A is upwards-directed and has the same upper bounds as A, so c = sup A and 0 = inf C, by 313Aa.
Also C is downwards-directed, so inf [C] = 0 in B. But now
[C] = {c \ a : a A } = {c \ b : b [A ]},
[A ] = {a0 . . .
an
: a0 , . . . , an A} = {b0 . . .
bn
: b0 , . . . , bn [A]},
because is a Boolean homomorphism. Again using 313Aa and the fact that b
get
c for every b [A ], we
Then D is order-dense in A. P
P If e A \ {0}, then there is an a A such that e 6 a, so that e \ a is a
non-zero member of D included in e. Q
Q Consequently there is a partition of unity C D, by 313K. But
now if b is any lower bound for [A] in B, we must have b d = 0 for every d D, so c 1 \ b for every
c C, and 1 \ b = 1, b = 0. Thus inf [A] = 0. As A is arbitrary, (ii) is satised.
(v)&(iv)(i) is trivial.
(c) We can use nearly identical arguments, remembering only to interpolate the word countable from
time to time. I spell out the new version of (ii)(iii), even though it requires no more than an adaptation
of the language. Assume (ii), and let A be a countable subset of A with a supremum c A. If A = , then
c = 0 so c = 0 = sup [A]. Otherwise, let han inN be a sequence running over A; set an = a0 . . . an and
cn = c \ an for each n. Then han inN is non-decreasing, with supremum c, and hcn inN is non-increasing,
with inmum 0; so inf nN cn = 0 and
supnN an = supnN an = c.
For (v)(ii), however, a dierent idea is involved. Assume (v), and suppose that han inN is a nonincreasing sequence in A with inmum 0. Set c0 = 1 \ a0 , cn = an1 \ an for n 1; then C = {cn : n N}
is a partition of unity in A (because if c cn = 0 for every n, then c an for every n), so [C] is a partition
of unity in B. Now if b an for every n, b cn for every n, so b = 0; thus inf nN an = 0. As han inN is
arbitrary, (ii) is satised.
313M The following result is perfectly elementary, but it will save a moment later on to have it spelt
out.
Lemma Let A and B be Boolean algebras and : A B an order-continuous Boolean homomorphism.
(a) If D is an order-closed subalgebra of B, then 1 [D] is an order-closed subalgebra of A.
(b) If C is the order-closed subalgebra of A generated by C A, then the order-closed subalgebra D of
B generated by [C] includes [C].
(c) Now suppose that is surjective and that C A is such that the order-closed subalgebra of A
generated by C is A itself. Then the order-closed subalgebra of B generated by [C] is B.
proof (a) Setting C = 1 [D]: if a, a C then (a b) = a b, (a a ) = a a D, so a a ,
a a C; 1 = 1 D so 1 C; thus C is a subalgebra of A. By 313Id, C is order-closed.
38
Boolean algebras
313M
313Xh
Order-continuous homomorphisms
39
(i) is order-continuous;
(ii) 1 [M ] is nowhere dense in W for every nowhere dense set M Z;
(iii) int [H] 6= for every non-empty open set H W .
proof (a)(i)(iii) Suppose that is order-continuous. ?? Suppose, if possible, that H W is a non-empty
open set and int [H] = . Let b B \ {0} be such that bb H. Then [bb] has empty interior; but also it
S
is a closed set, so its complement is dense. Set A = {a : a A, b
a [bb] = }. Then aA b
a = Z \ [bb] is a
dense open set, so sup A = 1 in A (313Ca). Because is order-continuous, sup [A] = 1 in B (313L(b-iii)),
and there is an a A such that a b 6= 0. But this means that bb 1 [b
a] 6= and [bb] b
a 6= , contrary
to the denition of A. X
X
Thus there is no such set H, and (iii) is true.
(b)(iii)(ii) Now assume (iii). If M Z is nowhere dense, set N = 1 [M ], so that N W is a closed
set. If H = int N , then int [H] int M = , so (iii) tells us that H is empty; thus N and 1 [M ] are
nowhere dense, as required by (ii).
T
(c)(ii)(i) Assume (ii), and let A A be a non-empty set such that inf A = 0 in A. Then M = aA b
a
has empty interior in Z (313Cb), so (being closed) is nowhere dense, and 1 [M ] is also nowhere dense. If
b B \ {0}, then
T
1
bb 6 1 [M ] = T
[b
a] = aA
ca,
aA
so b is not a lower bound for [A]. This shows that inf [A] = 0 in B. As A is arbitrary, is order-continuous
(313L(b-ii)).
313X Basic exercises (a) Use 313C to give alternative proofs of 313A and 313B.
(b) Let P be a partially ordered set. Show that there is a topology on P for which the closed sets are
just the order-closed sets.
(c) Let P be a partially ordered set, Q P an order-closed set, and R a subset of Q which is order-closed
in Q when Q is given the partial ordering induced by that of P . Show that R is order-closed in P .
> (d) Let A be a Boolean algebra. Suppose that 1 I A and that a b I for all a, b I. (i) Let
B be the intersection of all those subsets A of A such that I A and b \ a A whenever a, b A and
a b. Show that B is a subalgebra of A. (ii) Let B be the intersection of all those subsets A of A such
that I A, b \ a A whenever a, b A and a b and supnN bn A whenever hbn inN is a non-decreasing
sequence in A with a supremum in A. Show that B is a -subalgebra of A. (iii) Let B be the intersection
of all those subsets A of A such that I A, b \ a A whenever a, b A and a b and sup B A whenever
B is a non-empty upwards-directed subset of A with a supremum in A. Show that B is an order-closed
subalgebra of A. (iv) Hence give a proof of 313G not relying on Zorns Lemma or any other use of the axiom
of choice.
(e) Let A be a Boolean algebra, and B a subalgebra of A. Let B be the smallest sequentially order-closed
subset of A including B. Show that B is a subalgebra of A.
> (f ) Let X be a set, and A a subset of PX. Show that A is an order-closed subalgebra of PX i it is
of the form {f 1 [F ] : F Y } for some set Y , function f : X Y .
(g) Let P and Q be partially ordered sets, and : P Q an order-preserving function. Show that is
sequentially order-continuous i 1 [C] is sequentially order-closed in A for every sequentially order-closed
C B.
(h) For partially ordered sets P and Q, let us call a function : P Q monotonic if it is either
order-preserving or order-reversing. State and prove denitions and results corresponding to 313H, 313I and
313Xg for general monotonic functions.
40
Boolean algebras
313Xi
> (i) Let A be a Boolean algebra. Show that the operations (a, b) 7 a b and (a, b) 7 a b are ordercontinuous operations from AA to A, if we give AA the product partial order, saying that (a, b) (a , b )
i a a and b b .
(j) Let A be a Boolean algebra. Show that if a subalgebra of A is order-dense then it is dense in the
topology of 313Xb.
> (k) Let A be a Boolean algebra and A A any disjoint set. Show that there is a partition of unity in
A including A.
> (l) Let A, B be Boolean algebras and 1 , 2 : A B two order-continuous Boolean homomorphisms.
Show that {a : 1 a = 2 a} is an order-closed subalgebra of A.
(m) Let A and B be Boolean algebras and 1 , 2 : A B two Boolean homomorphisms. Suppose that
1 and 2 agree on some order-dense subset of A, and that one of them is order-continuous. Show that they
are equal. (Hint: if 1 is order-continuous, 2 a 1 a for every a.)
(n) Let A and B be Boolean algebras, A0 an order-dense subalgebra of A, and : A B a Boolean
homomorphism. Show that is order-continuous i A0 : A0 B is order-continuous.
> (o) Let A be a Boolean algebra. For A A set A = {b : a b = 0 a A}. (i) Show that A is an
order-closed ideal of A. (ii) Show that a set A A is an order-closed ideal of A i A = A . (iii) Show
that if I A is an order-closed ideal then {a : a I } is an order-dense ideal in the quotient algebra A/I.
(p) Let A and B be Boolean algebras, with Stone spaces Z and W ; let : A B be a Boolean
homomorphism, and : W Z the corresponding continuous function. Show that the following are
equiveridical: (i) is order-continuous; (ii) int 1 [F ] = 1 [int F ] for every closed F Z (iii) 1 [G] =
1 [G] for every open G Z.
(q) Let A be a Boolean algebra and : A A a Boolean homomorphism with xed-point subalgebra C
(312T). (i) Show that if is sequentially order-continuous then C is a -subalgebra of A. (ii) Show that if
is order-continuous then C is order-closed.
313Y Further exercises (a) Prove 313A-313C for general Boolean rings.
(b) Let A and B be Boolean algebras, with Stone spaces Z and W , and : A B a Boolean homomorphism, with associated continuous function : W Z. Show that is sequentially order-continuous
i 1 [M ] is nowhere dense for every nowhere dense zero set M Z.
(c) Let P be any partially ordered set, and let T be the topology of 313Xb. (i) Show that a sequence
hpn inN in P is T-convergent to p P i every subsequence of hpn inN has a monotonic sub-subsequence
with supremum or inmum equal to p. (ii) Show that a subset A of P is sequentially order-closed, in the
sense of 313Db, i the T-limit of any T-convergent sequence in A belongs to A. (iii) Suppose that A is an
upwards-directed subset of P with supremum p0 P . For a A set Fa = {p : a p A}, and let F be the
lter on P generated by {Fa : a A}. Show that F converges to p0 for T. (iv) Show that if Q is another
partially ordered set, endowed with a topology S in the same way, then a monotonic function : P Q is
order-continuous i it is continuous for the topologies T and S, and is sequentially order-continuous i it is
sequentially continuous for these topologies.
(d) Let U be a Banach lattice (242G, 354Ab). Show that its norm is order-continuous in the sense of
242Yg (or 354Dc) i its restriction to {u : u 0} is order-continuous in the sense of 313Ha.
(e) Let A and B be Boolean algebras with Stone spaces Z and W respectively, : A B a Boolean
homomorphism and : W Z the corresponding continuous function. Show that [A] is order-dense in B
i is irreducible, that is, [F ] 6= [W ] for any proper closed subset F of W .
314 intro.
Order-completeness
41
(f ) Let A and B be Boolean algebras with Stone spaces Z and W respectively, : A B a Boolean
homomorphism and : W Z the corresponding continuous function. Show that the following are
equiveridical: (i) is injective and order-continuous; (ii) for M Z, M is nowhere dense i 1 [M ] is
nowhere dense.
313 Notes and comments I give elementary proofs of 313A-313B because I believe that they help to
exhibit the relevant aspects of the structure of Boolean algebras; but various abbreviations are possible,
notably if we allow ourselves to use the Stone representation (313Xa). 313A and 313Ba-b can be expressed
by saying that the Boolean operations , and \ are (separately) order-continuous. Of course, \ is orderreversing, rather than order-preserving, in the second variable; but the natural symmetry in the concept of
partial order means that the ideas behind 313H-313I can be applied equally well to order-reversing functions
(313Xh). In fact, and can be regarded as order-continuous functions on the product space (313Bc-d,
313Xi). Clearly 313Bc-d can be extended into forms valid for any nite sequence A0 , . . . , An of subsets of
A in place of A, B. But if we seek to go to innitely many subsets of A we nd ourselves saying something
new; see 316G-316I below.
Proposition 313C, and its companions 313R, 313Xp and 313Yb, are worth studying not only as a useful
technique, but
S also in order to understand the dierence between sup A, where A is a set in a Boolean
algebra, and A, where A is a family of sets. Somehow sup A can be larger, and inf A smaller, than ones
rst intuition might suggest, corresponding to the fact that not every subset of the Stone space corresponds
to an element of the Boolean algebra.
I should like to use the words order-closed and sequentially order-closed to mean closed, or sequentially
closed, for some more or less canonical topology. The diculty is that while a great many topologies can be
dened from a partial order (one is described in 313Xb and 313Yc, and another in 367Yb), none of them has
such pre-eminence that it can be called the order-topology. Accordingly there is a degree of arbitrariness
in the language I use here. Nevertheless (sequentially) order-closed subalgebras and ideals are of such
importance that they seem to deserve a concise denotation. The same remarks apply to (sequential) ordercontinuity. Concerning the term order-dense in 313J, this has little to do with density in any topological
sense, but the word dense, at least, is established in this context.
With all these denitions, there is a good deal of scope for possible interrelations. The most important
to us is 313Q, which will be used repeatedly (typically, with A an algebra of sets), but I think it is worth
having the expanded version in 313P available.
I take the opportunity to present an abstract form of an important lemma on -algebras generated by
families closed under (136B, 313Gb). This time round I use the Zorns Lemma argument in the text
and suggest the alternative, elementary method in the exercises (313Xd). The two methods are opposing
extremes in the sense that the Zorns Lemma argument looks for maximal subalgebras included in A (which
are not unique, and have to be picked out using the axiom of choice) and the other approach seeks minimal
subalgebras including I (which are uniquely dened, and can be described without the axiom of choice).
Note that the concept of order-closed algebra of sets is not particularly useful; there are too few orderclosed subalgebras of PX and they are of too simple a form (313Xf). It is in abstract Boolean algebras that
the idea becomes important. In the most important partially ordered sets of measure theory, the sequentially
order-closed sets are the same as the order-closed sets (see, for instance, 316Fb below), and most of the
important order-closed subalgebras dealt with in this chapter can be thought of as -subalgebras which are
order-closed because they happen to lie in the right kind of algebra.
314 Order-completeness
The results of 313 are valid in all Boolean algebras, but of course are of most value when many suprema
and inma exist. I now set out the most useful denitions which guarantee the existence of suprema and
inma (314A) and work through their elementary relationships with the concepts introduced so far (314C314J). I then embark on the principal theorems concerning order-complete Boolean algebras: the extension
theorem for homomorphisms to a Dedekind complete algebra (314K), the Loomis-Sikorski representation of
a Dedekind -complete algebra as a quotient of a -algebra of sets (314M), the characterization of Dedekind
complete algebras in terms of their Stone spaces (314S), and the idea of Dedekind completion of a Boolean
algebra (314T-314U). On the way I describe regular open algebras (314O-314Q).
42
Boolean algebras
314A
P
P Because A has a least element 0 and a greatest element 1, every subset of A has upper and lower
bounds; so the two one-sided conditions together are equivalent to saying that A is Dedekind -complete.
I therefore have to show that they are equiveridical. Now if A A is a non-empty countable set, so is
B = {1 \ a : a A}, and
inf A = 1 \ sup B,
sup A = 1 \ inf B
whenever the right-hand-sides are dened (313A). So if the existence of a supremum (resp. inmum) of B
is guaranteed, so is the existence of an inmum (resp. supremum) of A. Q
Q
The real point here is of course that (A, ) is isomorphic to (A, ).
(d) Most specialists in Boolean algebra speak of complete, or -complete, Boolean algebras. I prefer
the longer phrases Dedekind complete and Dedekind -complete because we shall be studying metrics on
Boolean algebras and shall need the notion of metric completeness as well as that of order-completeness.
(e) I have had to make some rather arbitrary choices in the denition here. The principal examples of
partially ordered set to which we shall apply these denitions are Boolean algebras and Riesz spaces, which
are all lattices. Consequently it is not possible to distinguish in these contexts between the property of
Dedekind completeness, as dened above, and the weaker property, which we might call monotone ordercompleteness,
(i) whenever A P is non-empty, upwards-directed and bounded above then A has a least
upper bound in P (ii) whenever A P is non-empty, downwards-directed and bounded below
then A has a greatest lower bound in P .
(See 314Xa below. Monotone order-completeness is the property involved in 314Ya, for instance.) Nevertheless I am prepared to say, on the basis of my own experience of working with other partially ordered sets,
that Dedekind completeness, as I have dened it, is at least of sucient importance to deserve a name.
Note that it does not imply that P is a lattice, since it allows two elements of P to have no common upper
bound.
314F
Order-completeness
43
(f ) The phrase complete lattice is sometimes used to mean a Dedekind complete lattice with greatest
and least elements; equivalently, a Dedekind complete partially ordered set with greatest and least elements.
Thus a Dedekind complete Boolean algebra is a complete lattice in this sense, but R is not.
(g) The most important Dedekind complete Boolean algebras (at least from the point of view of measure
theory) are the measure algebras of the next chapter. I shall not pause here to give other examples, but
will proceed directly with the general theory.
314C Proposition Let A be a Dedekind -complete Boolean algebra and I a -ideal of A. Then the
quotient Boolean algebra A/I is Dedekind -complete.
proof I use the description in 314Bc. Let B A/I be a non-empty countable set. For each u B, choose
an au A such that u = au = au + I. Then c = supuB au is dened in A; consider v = c in A/I. Because
the map a 7 a is sequentially order-continuous (313Qb), v = sup B. As B is arbitrary, A/I is Dedekind
-complete.
314D Corollary Let X be a set, a -algebra of subsets of X, and I a -ideal of subsets of X. Then
I is a -ideal of the Boolean algebra , and / I is Dedekind -complete.
S
proof Of course is Dedekind -complete, because if hEn inN is any sequence in then nN En is
the least
upper bound of {En : n N} in . It is also easy to see that I is a -ideal of , since
S
F nN En I whenever F and hEn inN is a sequence in I. So 314C gives the result.
314E Proposition Let A be a Boolean algebra.
(a) If A is Dedekind complete, then all its order-closed subalgebras and principal ideals are Dedekind
complete.
(b) If A is Dedekind -complete, then all its -subalgebras and principal ideals are Dedekind -complete.
proof All we need to note is that if C is either an order-closed subalgebra or a principal ideal of A, and
B C is such that b = sup B is dened in A, then b C (see 313E(a-i-)), so b is still the supremum of B
in C; while the same is true if C is a -subalgebra and B C is countable, using 313E(a-ii-).
314F I spell out some further connexions between the concepts order-closed set, order-continuous
function and Dedekind complete Boolean algebra which are elementary without being quite transparent.
Proposition Let A and B be Boolean algebras and : A B a Boolean homomorphism.
(a)(i) If A is Dedekind complete and is order-continuous, then [A] is order-closed in B.
(ii) If B is Dedekind complete and is injective and [A] is order-closed then is order-continuous.
(b)(i) If A is Dedekind -complete and is sequentially order-continuous, then [A] is a -subalgebra of
B.
(ii) If B is Dedekind -complete and is injective and [A] is a -subalgebra of B then is sequentially
order-continuous.
proof (a)(i) If B [A], then a0 = sup( 1 [B]) is dened in A; now
a0 = sup([ 1 [B]]) = sup B
in B (313L(b-iv)), and of course a0 [A]. By 313E(a-i-) again, this is enough to show that [A] is
order-closed in B.
(ii) Suppose that A A and inf A = 0 in A. Then [A] has an inmum b0 in B, which belongs to
[A] because [A] is an order-closed subalgebra of B (313E(a-i- )). Now if a0 A is such that a0 = b0 ,
we have
(a a0 ) = a a0 = a
for every a A, so (because is injective) a a0 = a0 and a0 a for every a A. But this means that
a0 = 0 and b0 = 0 = 0. As A is arbitrary, is order-continuous (313L(b-ii)).
(b) Use the same arguments, but with sequences in place of the sets B, A above.
44
314G
Boolean algebras
314G Corollary (a) If A is a Dedekind complete Boolean algebra and B is an order-closed subalgebra
of A, then B is regularly embedded in A (denition: 313N).
(b) Let A be a Dedekind complete Boolean algebra, B a Boolean algebra and : A B an ordercontinuous Boolean homomorphism. If C A and C is the order-closed subalgebra of A generated by C,
then [C] is the order-closed subalgebra of B generated by [C].
proof (a) Apply 314F(a-ii) to the identity map from B to A.
(b) Let D be the order-closed subalgebra of B generated by [C]. By 313Mb, [C] D. On the other
hand, the identity homomorphism : C A is order-continuous, by (a), so : C B is order-continuous,
and [C] = [C] is order-closed in B, by 314F(a-i). But since [C] is surely included in [C], D is also
included in [C]. Accordingly [C] = D, as claimed.
314H Corollary Let A be a Boolean algebra and B a subalgebra of A.
(a) If A is Dedekind complete, then B is order-closed i it is Dedekind complete in itself and is regularly
embedded in A.
(b) If A is Dedekind -complete, then B is a -subalgebra i it is Dedekind -complete in itself and the
identity map from B to A is sequentially order-continuous.
proof Put 314E and 314F together.
314I Corollary (a) If A is a Boolean algebra and B is an order-dense subalgebra of A which is Dedekind
complete in itself, then B = A.
(b) If A is a Dedekind complete Boolean algebra, B is a Boolean algebra, : A B is an injective
Boolean homomorphism and [A] is order-dense in B, then is an isomorphism.
proof (a) Being order-dense, B is regularly embedded in A (313O), so this is a special case of 314Ha.
(b) Because [A] is order-dense, it is regularly embedded in B; also, the kernel of is {0}, which is
surely order-closed in A, so 313P(a-ii) tells us that is order-continuous. By 314F(a-i), [A] is order-closed
in B; being order-dense, it must be the whole of B (313K). Thus is surjective; being injective, it is an
isomorphism.
314J When we come to applications of the extension procedure in 312N, the following will sometimes
be needed.
Lemma Let A be a Boolean algebra and A0 a subalgebra of A. Take any c A, and set
A1 = {(a c) (b \ c) : a, b A0 },
d},
B = {b : b A0 , b \ c E}.
Because A is Dedekind complete, a = sup A and b = sup B are dened in A; because A0 is order-closed,
both belong to A0 , so d = (a c) (b \ c) belongs to A1 .
Now if d D, it is expressible as (a c) (b \ c) for some a, b A0 ; since a A and b B, we have
a a and b b , so d d . Thus d is an upper bound for D. On the other hand, if d is any other upper
bound for D in A, it is also an upper bound for E, so we must have
a c = supaA a c d ,
b \ c = supbB b \ c d ,
and d d . Thus d = sup D. This shows that the supremum of any subset of A1 belongs to A1 , so that
A1 is order-closed.
(b) The argument is the same, except that we replace D by a sequence hdn inN , and A, B by sequences
han inN , hbn inN in A0 such that dn = (an c) (bn \ c) for every n.
314M
Order-completeness
45
314K Extension of homomorphisms The following is one of the most striking properties of Dedekind
complete Boolean algebras.
Theorem Let A be a Boolean algebra and B a Dedekind complete Boolean algebra. Let A0 be a Boolean
subalgebra of A and 0 : A0 B a Boolean homomorphism. Then there is a Boolean homomorphism
1 : A B extending 0 .
proof (a) Let P be the set of all Boolean homomorphisms such that dom is a Boolean subalgebra of
A including A0 and extends 0 . Identify each member of P with its graph, which is a subset of A B,
and order P by inclusion, so that means just that extends . Then any non-empty totally ordered
subset Q of P has an upper bound in P . P
P Let be the simple union of these graphs. (i) If (a, b) and
(a, b ) both belong to , then there are , Q such that a = b, a = b ; now either or ;
in either case, = Q, so that
b = a = a = a = b .
for some Q; thus extends 0 (and, in particular, 0 dom ). (iii) Now suppose that a, a dom( ).
Then there are , Q such that a dom , a dom ; once again, = Q, so that a, a dom ,
and
a a dom dom ,
1 \ a dom dom ,
(a a ) = (a a ) = a a = a a ,
(1 \ a) = (1 \ a) = 1 \ a = 1 \ a.
(iv) This shows that dom is a subalgebra of A and that is a Boolean homomorphism, that is, that
P ; and of course is an upper bound for Q in P . Q
Q
(b) By Zorns Lemma, P has a maximal element 1 say.
?? Suppose, if possible, that A1 = dom 1 is not the whole of A; take c A \ A1 . Set A = {a : a
A1 , a c}. Because B is Dedekind complete, d = sup 1 [A] is dened in B. If a A and c a , then of
course a a and 1 a 1 a whenever a A, so that 1 a is an upper bound for 1 [A], and d 1 a .
But this means that there is an extension of 1 to a Boolean homomorphism on the Boolean subalgebra
of A generated by A1 {c} (312N). And this must be a member of P properly extending 1 , which is
supposed to be maximal. X
X
Thus dom 1 = A and 1 is an extension of 0 to A, as required.
314L The Loomis-Sikorski representation of a Dedekind -complete Boolean algebra The
construction in 314D is not only the commonest way in which new Dedekind -complete Boolean algebras
appear, but is adequate to describe them all. I start with an elementary general fact.
Lemma Let X be any topological space, and write M for the family of meager subsets of X. Then M is a
-ideal of subsets of X.
proof The point is that if A X is nowhere dense, so is every subset of A; this is obvious, since if B A
then B A so int B int A = . So if B A M, let hAn inN be a sequence of nowhere dense sets with
union A; then hB An inN is a sequence of nowhere dense sets with union B, so B M. If hAn inN is a
sequence in M with union A, then for each n we may choose a sequence hAnm imN of nowhere dense sets
with union An ; then the countable family hAnm in,mN may be re-indexed as a sequence of nowhere dense
sets with union A, so A M. Finally, is nowhere dense, so belongs to M.
314M Theorem Let A be a Dedekind -complete Boolean algebra, and Z its Stone space. Let E be the
algebra of open-and-closed subsets of Z, and M the -ideal of meager subsets of Z. Then = {EA : E
E, A M} is a -algebra of subsets of Z, M is a -ideal of , and A is isomorphic, as Boolean algebra, to
/M.
46
314M
Boolean algebras
and G = int G.
314P Theorem Let X be any topological space, and write RO(X) for the set of regular open sets in X.
Then RO(X) is a Dedekind complete Boolean algebra, with 1RO(X) = X and 0RO(X) = , and with Boolean
operations given by
G RO H = G H,
with Boolean ordering given by
G RO H = int GH,
G RO H = int G H,
G RO H G H,
sup H = int
G \RO H = G \ H,
H,
inf H = int
H = int
RO
RO
\RO
RO
314P
Order-completeness
47
is insuciently marked.
proof I base the proof on the study of an auxiliary algebra of sets which involves some of the ideas already
used in 314M.
(a) Let I be the family of nowhere dense subsets of X. Then I is an ideal of subsets of X. P
P Of course
I. If A B I then int A int B = . If A, B I and G is a non-empty open set, then G \ A is
a non-empty open set and (G \ A) \ B is non-empty; accordingly G cannot be a subset of A B = A B.
Q
This shows that int A B = , so that A B I. Q
(b) For any set A X, write A for the boundary of A, that is, A \ int A. Set
= {E : E X, E I}.
So if E , (X \ E) = E I and X \ E . Q
Q
If A I, then of course A = A I, so A ; accordingly I is an ideal in the Boolean algebra , and
we can form the quotient /I.
It will be helpful to note that every open set belongs to , since if G is open then G = G \ G cannot
include any non-empty open set (since any open set meeting G must meet G).
(c) For each E , set VE = int E; then VE is the unique member of RO(X) such that EVE I.
P
P (i) Being the interior of a closed set, VE RO(X). Since int E VE E, EVE E I. (ii) If
G RO(X) is such that EG I, then
G \ VE G \ VE (GE) (VE E) I,
so G \ VE , being open, must be actually empty, and G VE ; but this means that G int VE = VE .
Similarly, VE G and VE = G. This shows that VE is unique. Q
Q
(d) It follows that the map G 7 G : RO(X) /I is a bijection, and we have a Boolean algebra
structure on RO(X) dened by the Boolean algebra structure of /I. What this means is that for each of the
binary Boolean operations RO , RO , RO , \RO and for G, H RO(X) we must have GRO H = int G H,
writing RO for the operation on the algebra RO(X) and for the corresponding operation on or PX.
(e) Before working through the identications, it will be helpful to observe that if H is any non-empty
T
T
T
P Set G = int H. For every H H, G H so G int H = H;
subset of RO(X), then int H = int H. P
thus
T
T
G int H int H = G,
T
T
Q Consequently int H, being the interior of a closed set, belongs to RO(X).
so G = int H. Q
(f )(i) If G, H RO(X) then their intersection in the algebra RO(X) is
G RO H = int G H = int(G H) = G H,
using (d) for the rst equality and (e) for the second.
(v) If G, H RO(X), then G RO H = int G H and G RO H = int GH, by the remarks in (d).
48
Boolean algebras
314P
G RO H G RO H = G G H = G G H;
that is, the ordering of the Boolean algebra RO(X) is just the partial ordering induced on RO(X) by the
Boolean ordering of PX or .
S
T
(h) If H is any non-empty subset of RO(X), consider G0 = int H and G1 = int H.
G0 = inf H in RO(X). P
P By (e), G0 RO(X). Of course G0 H for every H H, so T
G0 is a lower
bound for H. If G is any
lower
bound
for
H
in
RO(X),
then
G
H
for
every
H
H,
so
G
H; but also
T
G is open, so G int H = G0 . Thus G0 is the greatest lower bound for H. Q
Q
G1 = sup H in RO(X). P
P Being the interior of a closed set, G1 RO(X), and of course
S
H = int H int H = G1
for every H H, so G1 is an upper bound for H in RO(X). If G is any upper bound for H in RO(X), then
S
G = int G int H = G1 ;
thus G1 is the least upper bound for H in RO(X). Q
Q
This shows that every non-empty H RO(X) has a supremum and an inmum in RO(X); consequently
RO(X) is Dedekind complete, and the proof is nished.
314Q Remarks (a) RO(X) is called the regular open algebra of the topological space X.
(b) Note that the map E 7 VE : RO(X) of part (c) of the proof above is a Boolean homomorphism,
if RO(X) is given its Boolean algebra structure. Its kernel is of course I; the induced map E 7 VE :
/I RO(X) is just the inverse of the isomorphism G 7 G : RO(X) /I.
*314R I interpolate a lemma corresponding to 313R.
Lemma Let X and Y be topological spaces, and f : X Y a continuous function such that f 1 [M ]
is nowhere dense in X for every nowhere dense M Y . Then we have an order-continuous Boolean
homomorphism from the regular open algebra RO(Y ) of Y to the regular open algebra RO(X) of X
dened by setting H = int f 1 [H] for every H RO(Y ).
proof (a) By the remark in 314O, the formula for H always denes a member of RO(X); and of course
is order-preserving.
Observe that if H RO(Y ), then f 1 [H] is open, so f 1 [H] H. It will be convenient to note straight
away that if V Y is a dense open set then f 1 [V ] is dense in X. P
P M = Y \ V is nowhere dense, so
f 1 [M ] is nowhere dense and its complement f 1 [V ] is dense. Q
Q
(b) If H1 , H2 RO(Y ) then (H1 H2 ) = H1 H2 . P
P Because is order-preserving, (H1 H2 )
H1 H2 . ?? Suppose, if possible, that they are not equal. Then (because (H1 H2 ) is a regular open
set) G = H1 H2 \ (H1 H2 ) is non-empty. Set M = f [G]. Then f 1 [M ] G is not nowhere dense,
so H = int M must be non-empty. Now G H1 f 1 [H1 ], so
f [G] f [f 1 [H1 ]] f [f 1 [H1 ]] H 1 ,
which is impossible. X
XQ
Q
(c) If H RO(Y ) and H = Y \ H is its complement in RO(Y ) then H = X \ H is the complement
of H in RO(X). P
P By (b), H and H are disjoint. Now H H is a dense open subset of Y , so
H H f 1 [H] f 1 [H ] = f 1 [H H ]
is dense in X, and the regular open set H must include the complement of H in RO(X). Q
Q
314T
Order-completeness
49
Putting (b) and (c) together, we see that the conditions of 312H(ii) are satised, so that is a Boolean
homomorphism.
(d)STo see that it is order-continuous, let H RO(Y ) be a non-empty set with supremum Y . Then
H0 = H is a dense open subset of Y (see the formula in 314P). So
S
S
1
[H] = f 1 [H0 ]
HH H
HH f
314S It is now easy to characterize the Stone spaces of Dedekind complete Boolean algebras.
Theorem Let A be a Boolean algebra, and Z its Stone space; write E for the algebra of open-and-closed
subsets of Z, and RO(Z) for the regular open algebra of Z. Then the following are equiveridical:
(i) A is Dedekind complete;
(ii) Z is extremally disconnected (denition: 3A3Ae);
(iii) E = RO(Z).
proof (i)(ii) If A is Dedekind
complete, let G be any open set in Z. Set A = {a : a A, b
a G},
S
a0 = sup A. Then G = {b
a : a A}, because E is a base for the topology of Z, so b
a0 = G, by 313Ca.
Consequently G is open. As G is arbitrary, Z is extremally disconnected.
(ii)(iii) If E E, then of course E = E = int E, so E is a regular open set. Thus E RO(Z). On the
other hand, suppose that G Z is a regular open set. Because Z is extremally disconnected, G is open; so
G = int G = G is open-and-closed, and belongs to E. Thus E = RO(Z).
(iii)(i) Since RO(Z) is Dedekind complete (314P), E and A are also Dedekind complete Boolean
algebras.
Remark Note that if the conditions above are satised, either 312L or the formulae in 314P show that the
Boolean structures of E and RO(Z) are identical.
314T I come now to a construction of great importance, both as a foundation for further constructions
and as a source of insight into the nature of Dedekind completeness.
Theorem Let A be a Boolean algebra, with Stone space Z; for a A let b
a be the corresponding open-andb be the regular open algebra of Z (314P).
closed subset of Z. Let A
(a) The map a 7 b
a is an injective order-continuous Boolean homomorphism from A onto an order-dense
b
subalgebra of A.
(b) If B is any Dedekind complete Boolean algebra and : A B is an order-continuous Boolean
b B such that 1 b
a = a
homomorphism, there is a unique order-continuous Boolean homomorphism 1 : A
for every a A.
proof (a)(i) Setting E = {b
a : a A}, every member of E is open-and-closed, so is surely equal to the
b for every a A. The formulae in 314P tell us
interior of its closure, and is a regular open set; thus b
aA
b
b
that if a, b A, then b
a b, taken in A, is just the set-theoretic intersection b
a bb = (a b) b ; while 1 \ b
a,
b is
taken in A,
a = Z \b
a = (1 \ a) b .
Z \b
b is not empty, then there is a non-empty member of E included in it, by the denition of
(iii) If G A
b
the topology of Z (311I). So E is an order-dense subalgebra of A.
50
314T
Boolean algebras
(b) Now suppose that B is a Dedekind complete Boolean algebra and : A B is an order-continuous
b is an isomorphism between A and the
Boolean homomorphism. Write a = b
a for a A, so that : A A
1
b
order-dense subalgebra E of A. Accordingly : E B is an order-continuous Boolean homomorphism,
being the composition of the order-continuous Boolean homomorphisms and 1 . By 314K, it has an
b B, and 1 = , that is, 1 b
extension to a Boolean homomorphism 1 : A
a = a for every a A. Now
b has supremum 1 in A.
b Set
1 is order-continuous. P
P Suppose that H A
b
Because E is order-dense in A,
sup( )[H ] = 1. Since any upper bound for 1 [H] must also be an upper bound for 1 [H ], sup 1 [H] = 1
in B. As H is arbitrary, 1 is order-continuous (313L(b-iii)). Q
Q
b
If 1 : A B is any other Boolean homomorphism such that 1 b
a = a for every a A, then 1 and 1
b G is the
agree on E, and the argument just above shows that 1 is also order-continuous. But if G A,
b of F = {E : E E, E G}, so
supremum (in A)
1 G = supEF 1 E = supEF 1 E = 1 G.
c}
if the inmum is dened in C. (The most important cases are when A is Dedekind complete and C is orderclosed in A, so that C is Dedekind complete (314E) and upr(a, C) is dened for every a A; but others also
arise.) upr(a, C) is sometimes called the projection of a on C.
(b) If A A is such that upr(a, C) is dened for every a A, a0 = sup A is dened in A and c0 =
supaA upr(a, C) is dened in C, then c0 = upr(a0 , C). P
P If c C then
c0
c upr(a, C)
c for every a A
a c for every a A a0
c. Q
Q
c (1 \ c) c upr(a, C)
c . Q
Q
314X Basic exercises (a) Let A be a Boolean algebra. (i) Show that the following are equiveridical:
() A is Dedekind complete () every non-empty upwards-directed subset of A with an upper bound has a
least upper bound () every non-empty downwards-directed subset of A with a lower bound has a greatest
lower bound. (ii) Show that the following are equiveridical: () A is Dedekind -complete () every nondecreasing sequence in A with an upper bound has a least upper bound () every non-increasing sequence
in A with a lower bound has a greatest lower bound.
314Ye
Order-completeness
51
(b) Let A be a Boolean algebra. Show that any principal ideal of A is order-closed. Show that A is
Dedekind complete i every order-closed ideal is principal.
> (c) Let A be a Dedekind -complete Boolean algebra, B a Boolean algebra and : A B a sequentially
order-continuous Boolean homomorphism. If C A and C is the -subalgebra of A generated by C, show
that [C] is the -subalgebra of B generated by [C].
(d) Let A be a Dedekind complete Boolean algebra, B an order-closed subalgebra of A, and a A; let
Aa be the principal ideal of A generated by a. Show that {a b : b B} is an order-closed subalgebra of
Aa .
(e) Find a proof of 314Tb which does not appeal to 314K.
> (f ) Let A be a Dedekind complete Boolean algebra and B an order-dense subalgebra of A. Show that
A is isomorphic to the Dedekind completion of B.
(g) Let A be a Dedekind complete Boolean algebra and C an order-closed subalgebra of A. Show that
an element a of A belongs to C i upr(1 \ a, C) = 1 \ upr(a, C) i upr(1 \ a, C) upr(a, C) = 0.
> (h) Let A be a Dedekind complete Boolean algebra, C an order-closed subalgebra of A, and a0 A,
c0 C. Show that the following are equiveridical: (i) there is a Boolean homomorphism : A C such
that c = c for every c C and a0 = c0 (ii) 1 \ upr(1 \ a0 , C) c0 upr(a0 , C).
(i) Let X be a topological space and RO(X) its regular open algebra. Show that if G RO(X) then the
regular open algebra of G is just the principal ideal RO(X) PG of RO(X) generated by G.
(j) Let X and Y be extremally disconnected compact Hausdor spaces, RO(X) and RO(Y ) their regular
open algebras, and : X Y a continuous surjection. Show that the following are equiveridical: (i) the
Boolean homomorphism V 7 [V ] from RO(Y ) to RO(X) (312P, 314S) is order-continuous; (ii) [U ] is
open-and-closed in Y for every open-and-closed set U X; (iii) [G] is open in Y for every open set G X.
314Y Further exercises (a) Let P be a Dedekind complete partially ordered set. Show that a set
Q P is order-closed i sup R, inf R belong to Q whenever R Q is a totally ordered subset of Q with
upper and lower bounds in P . (Hint: show by induction on that if A Q is upwards-directed and bounded
above and #(A) then sup A Q.)
(b) Let P be a lattice. Show that P is Dedekind complete i every non-empty totally ordered subset of
P with an upper bound in P has a least upper bound in P . (Hint: if A P is non-empty and bounded
below in P , let B be the set of lower bounds of A and use Zorns Lemma to nd a maximal element of B.)
(c) Give an example of a Boolean algebra A with an order-closed subalgebra A0 and an element c such
that the subalgebra generated by A0 {c} is not order-closed.
(d) Let X be any topological space. Let M be the -ideal of meager subsets of X, and set
Bb = {GA : G X is open, A M}.
b
(i) Show that Bb is a -algebra of subsets of X, and that B/M
is Dedekind complete. (Members of Bb are
said to be the subsets of X with the Baire property; Bb is the Baire property algebra of X.) (ii) Show
S
b is dense, then A B.
b (iii) Show that there is a
that if A X and {G : G X is open, A G B}
largest open set V M. (iv) Let RO(X) be the regular open algebra of X. Show that the map G 7 G
b
is an order-continuous Boolean homomorphism from RO(X) onto B/M,
so induces a Boolean isomorphism
b
b
between the principal ideal of RO(X) generated by X \ V and B/M.
(B/M
is the category algebra of X;
b
it is a Dedekind complete Boolean algebra. X is called a Baire space if V = ; in this case RO(X)
= B/M.)
(e) Let A be a Dedekind -complete Boolean algebra, and han inN any sequence in A. For n N set
En = {x : x {0, 1}N , x(n) = 1}, and let B be the -algebra of subsets of {0, 1}N generated by {En : n N}.
(B is the Borel -algebra of {0, 1}N ; see 4A3E in Volume 4.) Show that there is a unique sequentially ordercontinuous Boolean homomorphism : B A such that (En ) = an for every n N. (Hint: dene a suitable
function from the Stone space Z of A to {0, 1}N , and consider {E : E {0, 1}N , 1 [E] has the Baire
property in Z}.) Show that [B] is the -subalgebra of A generated by {an : n N}.
52
Boolean algebras
314Yf
(f ) Let A be a Boolean algebra, and Z its Stone space. Show that A is Dedekind -complete i G is open
whenever G is a cozero set in Z. (Such spaces are called basically disconnected or quasi-Stonian.)
(g) Let A, B be Dedekind complete Boolean algebras and D A \ {0} an order-dense set. Suppose that
: D B is such that (i) [D] is order-dense in B (ii) for all d, d D, d d = 0 i d d = 0. Show
that has a unique extension to a Boolean isomorphism from A to B.
(h) Let A be any Boolean algebra. Let J be the family of order-closed ideals in A. Show that (i) J
is a Dedekind complete Boolean algebra with operations dened by the formulae I J = I J, 1 \ J =
{a : a b = 0 for every b J} (ii) the map a 7 Aa , the principal ideal generated by a, is an injective
order-continuous Boolean homomorphism from A onto an order-dense subalgebra of J (iii) J is isomorphic
to the Dedekind completion of A.
314 Notes and comments At the risk of being tiresomely long-winded, I have taken the trouble to spell out
a large proportion of the results in this section and the last in their sequential as well as their unrestricted
forms. The point is that while (in my view) the underlying ideas are most clearly and dramatically expressed
in terms of order-closed sets, order-continuous functions and Dedekind complete algebras, a large proportion
of the applications in measure theory deal with sequentially order-closed sets, sequentially order-continuous
functions and Dedekind -complete algebras. As a matter of simple technique, therefore, it is necessary
to master both, and for the sake of later reference I generally give the statements of both versions in full.
Perhaps the points to look at most keenly are just those where there is a dierence in the ideas involved, as
in 314Bb, or in which there is only one version given, as in 314M and 314T.
If you have seen the Hahn-Banach theorem (3A5A), it may have been recalled to your mind by Theorem
314K; in both cases we use an order relation and a bit of algebra to make a single step towards an extension
of a function, and Zorns lemma to turn this into the extension we seek. A good part of this section has
turned out to be on the borderland between the theory of Boolean algebra and general topology; naturally
enough, since (as always with the general theory of Boolean algebra) one of our rst concerns is to establish
connexions between algebras and their Stone spaces.
I think 314T is the rst substantial universal mapping theorem in this volume; it is by no means the
b is not just that we obtain a Dedekind complete Boolean algebra in
last. The idea of the construction A
which A is embedded as an order-dense subalgebra, but that we simultaneously obtain a theorem on the
b of order-continuous Boolean homomorphisms dened on A. This characterization
canonical extension to A
b a 7 b
b becomes
is enough to dene the pair (A,
a) up to isomorphism, so the exact method of construction of A
of secondary importance. The one used in 314T is very natural (at least, if we believe in Stone spaces), but
there are others (see 314Yh), with dierent virtues.
314K and 314T both describe circumstances in which we can nd extensions of Boolean homomorphisms.
Clearly such results are fundamental in the theory of Boolean algebras, but I shall not attempt any systematic
presentation here. 314Ye can also be regarded as belonging to this family of ideas.
315F
53
a a = ha(i) a(i)iiI = a
for every a A; and it is a Boolean algebra because if we set 1A = h1Ai iiI , then 1A a = a for every a A.
I will call A the simple product of the family hAi iiI .
I should perhaps remark that when I = then A becomes {}, to be interpreted as the singleton Boolean
algebra.
(b) The Boolean operations on A are now dened by the formulae
for all a, b A.
a b = ha(i) b(i)iiI ,
a \ b = ha(i) \ b(i)iiI
315B Theorem Let hAi iiI be a family of Boolean algebras, and A their simple product.
(a) The maps a 7 i (a) = a(i) : A Ai are all Boolean homomorphisms.
(b) If B is any other Boolean algebra, then a map : B A is a Boolean homomorphism i i : B Ai
is a Boolean homomorphism for every i I.
proof Verication of these facts amounts just to applying the denitions with attention.
315C Products of partially ordered sets (a) It is perhaps worth spelling out the following
elementary
Q
denition. If hPi iiI is any family of partially ordered sets, its product is the set P = iI Pi ordered by
saying that p q i p(i) q(i) for every i I; it is easy to check that P is now a partially ordered set.
(b) The point is that if A is the simple product of a family hAi iiI of Boolean algebras, then the ordering
of A is just the product partial order:
a b a b = a a(i) b(i) = a(i) i I a(i)
b(i) i I.
Now we have the following elementary, but extremely useful, general facts about products of partially
ordered sets.
315D Proposition Let hPi iiI be a family of non-empty partially ordered sets with product P .
(a) For any non-empty set A P and q P ,
(i) sup A = q in P i suppA p(i) = q(i) in Pi for every i I,
(ii) inf A = q in P i inf pA p(i) = q(i) in Pi for every i I.
(b) The coordinate maps p 7 i (p) = p(i) : P Pi are all order-preserving and order-continuous.
(c) For any partially ordered set Q and function : Q P , is order-preserving i i is order-preserving
for every i I.
(d) For any partially ordered set Q and order-preserving function : Q P ,
(i) is order-continuous i i is order-continuous for every i,
(ii) is sequentially order-continuous i i is sequentially order-continuous for every i.
(e)(i) P is Dedekind complete i every Pi is Dedekind complete.
(ii) P is Dedekind -complete i every Pi is Dedekind -complete.
proof All these are elementary verications. Of course parts (b), (d) and (e) rely on (a).
315E Factor algebras as principal ideals Because Boolean algebras have least elements, we have a
second type of canonical homomorphism associated with their products. If hAi iiI is a family of Boolean
algebras with simple product A, dene i : Ai A by setting (i a)(i) = a, (i a)(j) = 0Aj if i I, a Ai and
j I \ {i}. Each i is a ring homomorphism, and is a Boolean isomorphism between Ai and the principal
ideal of A generated by i (1Ai ). The family hi (1Ai )iiI is a partition of unity in A.
Associated with these embeddings is the following important result.
315F Proposition Let A be a Boolean algebra and hei iiI a partition of unity in A. Suppose
either (i) that I is nite
or (ii) that I is countable and A is Dedekind -complete
or (iii) that A is Dedekind complete.
54
315F
Boolean algebras
iI
proof The given map is a Boolean homomorphism because each of the maps a 7 a ei : A Aei is (312J).
It is injective because supiI ei = 1,Qso if a A \ {0} there is an i such that a ei 6= 0. It is surjective
because hei iiI is disjoint and if c iI Aei then a = supiI c(i) is dened in A and
a ej = supiI c(i) ej = c(j)
for every j I (using 313Ba). The three alternative versions of the hypotheses of this proposition are
designed to ensure that the supremum is always well-dened in A.
315G Algebras of sets and their quotients The Boolean algebras of measure theory are mostly
presented as algebras of sets or quotients of algebras of sets, so it is perhaps worth spelling out the ways in
which the product construction applies to such algebras.
Proposition Let hXi iiI be
Q a family of sets, and i an algebra of subsets of Xi for each i.
(a) The simple product iI i may be identied with the algebra
= {E : E X, {x : (x, i) E} i for every i I}
for each E .
Q
(b) Now suppose that Ji is an ideal of i for each i. Then iI i /Ji may be identied with /J , where
J = {E : E , {x : (x, i) E} Ji for every i I},
(c) The topological space Z may be identied with the Stone space of the Boolean algebra A. P
P By
Tychonos theorem (3A3J), Z is compact. If z Z and G is an open subset of Z containing z, then there
are J, hGj iiJ such that J is a nite subset of I, Gj is an open subset of Zj for each j J, and
z {w : w Z, w(j) Gj for every j J} G.
315J
55
315I Theorem Let hAi iiI be a family of Boolean algebras, with free product A.
(a) The canonical map i : Ai A is a Boolean homomorphism for every i I.
(b) For any Boolean algebra B and any family hi iiI such that i is a Boolean homomorphism from Ai
to B for every i, there is a unique Boolean homomorphism : A B such that i = i for each i.
proof
These are both consequences of 312P-312Q. As in 315H, write Zi for the Stone space of A, and Z for
Q
iI Zi , identied with the Stone space of A, as observed in 315Hc. The maps i : Ai A are dened as
the homomorphisms corresponding to the continuous maps z 7 i (z) = z(i) : Z Zi , so (a) is surely true.
Now suppose that we are given a Boolean homomorphism i : Ai B for each i I. Let W be the
Stone space of B, and let i : W Zi be the continuous function corresponding to i . By 3A3Ib, the map
w 7 (w)
= hi (w)iiI : W Z is continuous, and corresponds to a Boolean homomorphism : A B;
i = i for each i. Moreover, is the only Boolean homomorphism with this property,
because i = i ,
because if : A B is a Boolean homomorphism such that i = i for every i, then corresponds to
a continuous function : W Z, and we must have i = i for each i, so that = and = . This
proves (b).
315J Of course 315I is the dening property of the free product (see 315Xg below). I list a few further
basic facts.
Proposition Let hAi iiI be a family of Boolean algebras, and A their free product; write i : Ai A for
the canonical homomorphisms.
S
(a) A is the subalgebra of itself generated by iI i [Ai ].
(b) Write C for the set of those members of A expressible in the form inf jJ j (aj ), where J I is nite
and aj Aj for every j. Then every member of A is expressible as the supremum of a disjoint nite subset
of C. In particular, C is order-dense in A.
(c) Every i is order-continuous.
(d) A = {0A } i there is some i I such that Ai = {0Ai }.
(e) Now suppose that Ai 6= {0Ai } for every i I.
(i) i is injective for every i I.
(ii) If J I is nite and aj is a non-zero member of Aj for each j J, then inf jJ j (aj ) 6= 0.
(iii) If i, j are distinct members of I and a Ai , b Aj , then i (a) = j (b) i either a = 0Ai and
b = 0Aj or a = 1Ai and b = 1Aj .
Q
proof As usual, write Zi for the Stone space of Ai , and Z = iI Zi , identied with the Stone space of A
(315Hc).
S
(a) Write A for the subalgebra of A generated by iI i [Ai ]. Then i : Ai A is a Boolean homomorphism for each i, so by 315Ib there is a Boolean homomorphism : A A such that i = i for each
i. Now, regarding as a Boolean homomorphism from A to itself, the uniqueness assertion of 315Ib (with
B = A) shows that must be the identity, so that A = A.
(b) Write D for the set of nite partitions of unity in A consisting of members of C, and A for the set of
members of A expressible in the form sup D where D is a subset of a member of D. Then A is a subalgebra
of A. P
P (i) 1A C (set J = in the denition of members of C) so {1A } D and 0A , 1A A. (ii) Note that
if c, d C then c d C. (iii) If a, b A, express them as sup D , sup E where D D D, E E D.
Then
so
F = {d e : d D, e E} D,
1A \ a = sup D \ D A,
a b = sup{f : f F, f
a b} A. Q
Q
Q
Also, i [Ai ] A for each i I. P
P If a Ai , then {i (a), i (1Ai \ a)} D, so i (a) A. Q
So (a) tells us that A = A, and every member of A is a nite disjoint union of members of C.
(c) If i I and A Ai and inf A = 0 in Ai , take any non-zero c A. By (b), we can nd a nite J I
and a family haj ijJ such that c = inf jJ j (aj ) c and c 6= 0. Regarding c as a subset of Z, we have
56
Boolean algebras
315J
a point z c . Adding i to J and setting ai = 1Ai if necessary, we may suppose that i J. Now c 6= 0A
ai \ b
a. In this case, setting w(i) = t,
so ai 6= 0Ai and there is an a A such that ai 6 a, so there is a t b
w(j) = z(j) for j 6= i, we have w c \ i (a), and c , c are not included in i (a). As c is arbitrary, this shows
that inf i [A] = 0. As A is arbitrary, i is order-continuous.
(d) The point is that A = {0A } i Z = , which is so i some Zi is empty.
(e)(i) Because no Zi is empty, all the coordinate maps from Z to Zi are surjective, so the corresponding
homomorphisms i are injective (312Ra).
(ii) Because J is nite,
is not empty.
(iii) If i (a) = j (b) = 0A then (using (i)) a = 0Ai and b = 0Aj ; if i (a) = j (b) = 1A then a = 1Ai and
a and u Zj \ bb. Now there is a z Z such
b = 1Aj . ?? If i (a) = j (b) A \ {0A , 1A }, then there are t b
that z(i) = t and z(j) = u, so that z i (a) \ j (b). X
X
315K Proposition Let hAi iiI be any family of Boolean algebras, and hJk ikK any partition (that is,
disjoint cover) of I. Then the free product A of hAi iiI is isomorphic to the free product B of hBk ikK ,
where each Bk is the free product of hAi iiJk .
proof Write i : Ai A, i : A Bk and k : Bk B for the canonical homomorphisms when k K,
i Jk . Then the homomorphisms k i : Ai B correspond to a homomorphism : A B such that
i = k i whenever i Jk . Next, for each k, the homomorphisms i : Ai A, for i Jk , correspond to
a homomorphism k : Bk A such that k i = i for i Jk ; and the family hk ikK corresponds to a
homomorphism : B A such that k = k for k K. Consequently
i = k i = k i = i
whenever k K, i Jk . Once again using the uniqueness assertion in 315Ib, is the identity homomorphism on A. On the other hand, if we look at : B B, then we see that
k i = k i = i = k i
whenever
k K, i Jk . Now, for given k, {b : b Bk , k b = k b} is a subalgebra of Bk including
S
[A
[B
k
k ], and is the whole of B. Thus is the identity on B and , are the two halves of an
kK
isomorphism between A and B.
315L Algebras of sets and their quotients Once again I devote a paragraph to spelling out the
application of the construction to the algebras most important to us.
Proposition Let hXi iiI
Nbe a family of sets, and i an algebra of subsets of Xi for eachQi.
(a) The free product iI i may be identied with the algebra of subsets of X = iI Xi generated
by the set {i (E) : i I, E i }, where i (E) = {x : x X,Nx(i) E}.
(b) Now suppose that Ji is an ideal of i for each i. Then iI i /Ji may be identied with /J , where
J is the ideal of generated by {i (E) : i I, E Ji }; the corresponding canonical maps i : i /Ji /J
being dened by the formula i (E ) = (i (E)) for i I, E i .
proof I start by proving (b) in detail; the argument for (a) is then easy to extract. Write Ai = i /Ji ,
A = /J .
(i) Fix i I for the moment. By the denition of , i (E) for E i , and it is easy to check that
i : i is a Boolean homomorphism. Again, because i (E) J whenever E Ji , the kernel of the
homomorphism E 7 (i (E)) : i A includes Ji , so the formula for i denes a homomorphism from Ai
to A.
N
Now let C = iI Ai be the free product, and write i : Ai C for the canonical homomorphisms. By
315I, there is a Boolean homomorphism : C A such that i = i for each i. The set
315N
57
{E : E , E [C]}
is a subalgebra of including i [i ] for every i, so is itself, and is surjective.
(ii) We need a simple description of the ideal J , as follows: a set ES
belongs to J i there are a nite
K I and a family hFk ikK such that Fk Jk for each k and E kK k (Fk ). For evidently such sets
have to belong to J , since the k (Fk ) will be in J , while the family of all these sets is an ideal containing
i (F ) whenever i I, F Ji .
(iii) Now we can see that : C QA is injective. P
P Take any non-zero c C. By 315Jb, we can nd a
nite J I and a family haj ijJ in jJ Aj such that 0 6= inf jJ j aj c. Express each aj as Ej , where
T
Ej j , and consider E = X jJ j (Ej ) . Then
E = inf jJ j aj = (inf jJ j aj )
(c).
j aj
6= 0, Ej
/ Jj for each j. But it follows that E
/ J , because if K I is nite and
Also, because
Fk Jk for each k K, set Ei = Xi for i S
I \ J, Fi = for i I \ K; then there is an x X such that
x(i) Ei \ Fi for each i I, so that x E \ kK Fk . By the criterion of (ii), E
/ J . So
0 6= E
(c).
jJ
iI,jJ
So c c C for all c, c C. (ii) If hai iiI is a nite partition of unity in A and bi B for each i, then
1 \ supiI ai bi = (supiI ai 1) \ (supiI ai bi ) = supiI ai (1 \ bi ) C.
58
Boolean algebras
315O
so that neither aj {n} nor bj {n} is empty, that is, n aj bj . But this means that m aj , so that
(aj bj ) ({m} {n}) = (aj {m}) (bj {n}) 6= 0,
(e) Let hXi iiI be any family of topological spaces. Let X be their disjoint union {(x, i) : i I, x Xi },
with the disjoint union topology; that is, a set G X is open in X i {x : (x, i) G} is open in Xi for
every i I. (i) Show that the algebra of open-and-closed subsets of X can be identied, as Boolean algebra,
with the simple product of the algebras of open-and-closed sets of the Xi . (ii) Show that the regular open
algebra of X can be identied, as Boolean algebra, with the simple product of the regular open algebras of
the Xi .
(f ) Show that the topological product of any family of zero-dimensional spaces is zero-dimensional.
(g) Let hAi iiI be any family of Boolean algebras, with free product A, and i : Ai A the canonical
homomorphisms. Suppose we have another Boolean algebra A , with homomorphisms i : Ai A , such
that for every Boolean algebra B and every family hi iiI of homomorphisms from the Ai to B there is a
unique homomorphism : A B such that i = i for every i. Show that there is a unique isomorphism
: A A such that i = i for every i I.
315Ye
59
(h) Let I be any set, and let A be the algebra of open-and-closed sets of {0, 1}I ; for each i I set
ai = {x : x {0, 1}I , x(i) = 1} A. Show that for any Boolean algebra B, any family hbi iiI in B there is
a unique Boolean homomorphism : A B such that (ai ) = bi for every i I.
(i) Let hAi iiI , Q
hBj ijJ be
Q of Boolean algebras. Show that there is a natural injective
Qtwo families
homomorphism : iI Ai jJ Bj iI,jJ Ai Bj dened by saying that
for a
Ai , b
(a b) = ha(i) b(j)iiI,jJ
jJ
(k) Let A and B be Boolean algebras with partitions of unity hai iiI , hbj ijJ . Show that hai bj iiI,jJ
is a partition of unity in A B.
(l) Let A and B be Boolean algebras and a A, b B. Write Aa , Bb for the corresponding principal
ideals. Show that there is a canonical isomorphism between Aa Bb and the principal ideal of A B
generated by a b.
N
(m) Let hAi iiI be any family of Boolean algebras, with free product
iI Ai , and i : Ai A the
canonical maps. Show that i [Ai ] is an order-closed subalgebra of A for every i.
(n) Let A be a Boolean algebra. Let us say that a family hAi iiI of subalgebras of A is Booleanindependent if inf jJ aj 6= 0 wheneverSJ I is nite and aj Aj \ {0} for every
Nj J. Show that in this
case the subalgebra of A generated by iI Ai is isomorphic to the free product iI Ai .
(o) Let hAi iiI and hBi iiI be two families of Boolean algebras, and suppose that for each i I we are
given a Boolean homomorphism
N
Ni : Ai Bi with kernel Ki Ai . Show that the i induce a Boolean
homomorphism : iI Ai iI Bi with kernel generated by the images of the Ki . Show that if every
i is surjective, so is .
(p) Let hAi iiI be any family
of non-trivial Boolean algebras. Show that if J
N
NI and Bj is a subalgebra
of Aj for each j J, then jJ Bj is canonically embedded as a subalgebra of iI Ai .
(q) Let A and B be Boolean algebras, neither {0}. Show that any element of AB is uniquely expressible
as supiI ai bi where hai iiI is a partition of unity in A, with no ai equal to 0, and bi 6= bj in B for i 6= j.
315Y Further exercises (a) Let hAi iiI and hBi iiI be two families of Boolean
algebras,
N and suppose
N
that we are given Boolean homomorphisms i : Ai Bi for each i; let :
iI Bi be the
iI Ai
induced homomorphism. (i) Show that if every i is order-continuous, so is . (ii) Show that if every i is
sequentially order-continuous, so is .
(b) Let hZi iiI be any family of topological spaces with product Z. For i I, z Z set i (z) = z(i).
Show that if M Zi is nowhere dense in Zi then 1
i [M ] is nowhere dense in Z. Use this to prove 315Jc.
(c) Let hAi iiI be
N a family of Boolean
N algebras, and suppose that we are given subalgebras Bi of Ai
for each i; set A = iI Ai and B = iI Bi , and let : B A be the homomorphism induced by the
embeddings Bi Ai . (i) Show that if every Bi is order-closed in Ai , then [B] is order-closed in A. (ii)
Show that if every Bi is a -subalgebra of Ai , then [B] is a -subalgebra in A.
(d) Let hXi iiI be a family of topological spaces, with product X. Let RO(Xi ), RO(X) be the corresponding
regular open algebras. Show that RO(X) can be identied with the Dedekind completion of
N
iI RO(Xi ).
(e) Use the ideas of 315Xh and 315L to give an alternative construction of free product, for which 315I
and 315J(e-ii) are true, which does not depend on the concept of Stone space nor on any other use of the
60
Boolean algebras
315Ye
axiom of choice. (Hint: show that for any Boolean algebra A there is a canonical surjection from the algebra
EA onto A, where EJ is the algebra of subsets of {0, 1}J generated by sets of the form {x : x(j) = 1}; show
that for such
N algebras EJ , at least, the method of 315H-315I can be used; now apply the method of 315L to
describe iI Ai as a quotient of EJ where J = {(a, i) : i I, a Ai }. Finally check that if no Ai is trivial,
then nor is the free product.)
(f ) Let A and B be Boolean algebras. Show that A B is Dedekind complete i either A = {0} or
B = {0} or A is nite and B is Dedekind complete or B is nite and A is Dedekind complete.
(g) Let hPi iiI be any family of partially ordered spaces. (i) Give a construction of a partially ordered
space P , together with a family of order-preserving maps i : Pi P , such that whenever Q is a partially
ordered set and i : Pi Q is order-preserving for every i I, there is a unique order-preserving map
: P Q such that i = i for every i. (ii) Show that will be order-continuous i every i is. (iii)
Show that P will be Dedekind complete i every Pi is, but (except in trivial cases) is not a lattice.
315 Notes and comments In this section I nd myself asking for slightly more sophisticated algebra than
seems necessary elsewhere. The point is that simple products and free products are best regarded as dened
by the properties described in 315B and 315I. That is, it is sometimes right to think of a simple product
of a family hAi iiI of Boolean algebras as being a structure (A, hi iiI ) where A is a Boolean algebra,
i : A Ai is a homomorphism for every i I, and every family of homomorphisms from a Boolean algebra
B to the Ai can be uniquely represented by a single homomorphism from B to A. Similarly, reversing the
direction of the homomorphisms, we can speak of a free product (it would be natural to say coproduct)
(A, hi iiI ) of hAi iiI . On such denitions, it is elementary that any two simple products, or free products,
are isomorphic in the obvious sense (315Xa, 315Xg), and very general arguments from abstract algebra, not
restricted to Boolean algebras (see Bourbaki 68, IV.3.2), show that they exist. (But in order to prove such
basic facts as that the i are surjective, or that the i are, except when the construction collapses altogether,
injective, we do of course have to look at the special properties of Boolean algebras.) Now in the case of
simple products, the Cartesian product construction is so direct and so familiar that there seems no need
to trouble our imaginations with any other. But in the case of free products, things are more complicated.
I have given primacy to the construction in terms of Stone spaces because I believe that this is the fastest
route to eective mental pictures. But in some ways this approach seems to be inappropriate. If you take
what in my view is a tenable position, and say that a Boolean algebra is best regarded as the limit of its
nite subalgebras, then you might prefer a construction of a free product as a limit of free products of nitely
many nite subalgebras. Or you might feel that it is wrong to rely on the axiom of choice to prove a result
which certainly does not need it (see 315Ye).
Because I believe that the universal mapping theorem 315I is the right basis for the study of free products,
I am naturally led to use it as the starting point for proofs of theorems about free products, as in 315K.
But 315J(e-ii) seems to lie deeper. (Note, for instance,
that in 315L we do need the axiom of choice, in part
Q
(c) of the proof, since without it the product iI Xi could be empty.)
Both simple product and free product are essentially algebraic constructions involving the category of
Boolean algebras and Boolean homomorphisms, and any relationships with such concepts as order-continuity
must be regarded as accidental. 315Cb and 315D show that simple products behave very straightforwardly
when the homomorphisms involved are order-continuous. 315P, 315Xm and 315Ya-315Yc show that free
products are much more complex and subtle.
For nite products, we have a kind of distributivity; (A B) C can be identied with (A C) (B C)
(315Xi, 315Xj). There are contexts in which this makes it seem more natural to write A B in place
of A B, and indeed I have already spoken of a direct sum of measure spaces (214K) in terms which
correspond closely to the simple product of algebras of sets described in 315Ga. Generally, the simple
product corresponds to disjoint unions of Stone spaces (315Xc) and the free product to products of Stone
spaces. But the simple product is indeed the product Boolean algebra, in the ordinary category sense;
the universal mapping theorem 315B is exactly of the type we expect from products of topological spaces
(3A3Ib) or partially ordered sets (315Dc), etc. It is the free product which is special to Boolean algebras.
The nearest analogy that I know of elsewhere is with the concept of tensor product of linear spaces (cf.
253).
316E
61
Further topics
d = sup< a a I,
c = a \ d A \ I.
c c c a c d = 0
aA {d
:d
a}.
Applying Zorns lemma to the family C of disjoint subsets of D, we have a maximal C0 C. For each c C0
choose a bc A such that c bc , and set B0 = {bc : c C0 }. Because A is ccc, C0 is countable, so B0 is also
countable. ?? If there is an upper bound e for B0 which is not an upper bound for A, take a A such that
c = a \ e 6= 0; then c D and c c = c bc = 0 for every c C0 , so C0 {c } C; but C0 was supposed
to be maximal in C. X
X Thus every upper bound for B0 is also an upper bound for A.
62
316E
Boolean algebras
(b) Similarly, there is a countable set B1 A = {1 \ a : a A} such that every upper bound for B1 is
an upper bound for A . Set B1 = {1 \ b : b B1 }; then B1 is a countable subset of A and every lower bound
for B1 is a lower bound for A. Try B = B0 B1 . Then B is a countable subset of A and every upper (resp.
lower) bound for B is an upper (resp. lower) bound for A; so that B must have exactly the same upper and
lower bounds as A has.
316F Corollary Let A be a ccc Boolean algebra.
(a) If A is Dedekind -complete it is Dedekind complete.
(b) If A A is sequentially order-closed it is order-closed.
(c) If Q is any partially ordered set and : A Q is a sequentially order-continuous order-preserving
function, it is order-continuous.
(d) If B is another Boolean algebra and : A B is a sequentially order-continuous Boolean homomorphism, it is order-continuous.
proof (a) If A is any subset of A, let B A be a countable set with the same upper bounds as A; then
sup B is dened in A and must be sup A.
(b) Suppose that B A is non-empty and upwards-directed and has a supremum a in A. Then there is
a non-empty countable set C B with the same upper bounds as B. Let hcn inN be a sequence running
over C. Because B is upwards-directed, we can choose hbn inN inductively such that
b0 = c0 ,
Now any upper bound for {bn : n N} must also be an upper bound for {cn : n N} = C, so is an upper
bound for the whole set B. But this means that a = supnN bn . As hbn inN is a non-decreasing sequence in
A, and A is sequentially order-closed, a A.
In the same way, if B A is downwards-directed and has an inmum in A, this is also the inmum of
some non-increasing sequence in B, so must belong to A. Thus A is order-closed.
(c)(i) Suppose that A A is a non-empty upwards-directed set with supremum a0 A. As in (b), there
is a non-decreasing sequence hcn inN in A with supremum a0 . Because is sequentially order-continuous,
a0 = supnN cn in Q. But this means that a0 must be the least upper bound of [A].
(ii) Similarly, if A A is a non-empty downwards-directed set with inmum a0 , there is a non-increasing
sequence hcn inN in A with inmum a0 , so that
inf [A] = inf nN cn = a0 .
(iii) whenever hAn inN is a sequence of upwards-directed subsets of A, each with a supremum cn = sup An ,
and
B = {b : b A, for every n N there is an a An such that b
a},
then inf{cn \ b : n N, b B} = 0;
(iv) whenever hAn inN is a sequence of upwards-directed subsets of A, each with a supremum cn = sup An ,
and inf nN cn = c is dened, then c = sup B, where
316I
63
Further topics
a}.
proof (i)(iii) Suppose that A is weakly (, )-distributive, and that hAn inN is a sequence of upwardsdirected subsets of A, each with a supremum cn = sup An . For each n N,
S
Dn = {d : d 1 \ cn } aAn {d : d a}
(iv)(ii) Suppose that (iv) is true and that An and B are as in the statement of (ii). Set An = {1 \ a :
a An }, so that An is an upwards-directed set with supremum 1 for each n, and
then
a} = {1 \ b : b B};
so that Cn is downwards-directed and has inmum 0. Set E = {e : for every n N there is a c Cn such
that c e}; then inf E = 0. So B0 = {b : b e = 0 for some e E} is order-dense in A and includes a
partition B of unity. If n N and b B, take e E such that b e = 0, c Cn such that c e, and a nite
set D An such that c = 1 \ sup D; then
b
1 \ e 1 \ c sup D
(b) Suppose
S that A is weakly (, )-distributive and that M is a meager subset of Z. Then M can be
expressed as nN Mn where each Mn is nowhere dense. For each n N, let An be a partition of unity such
that Mn b
a = for every a An . Let B be a partition of unity such that {a : a An , a b 6= 0} is nite
for every n N and b B. Now Mn bb = for every n N and b B. P
P C = {a : a An , b a 6= 0} is
S
nite. So F = aC b
a is closed and G = bb \ F is open. But G b
a = for every a A, so G is empty and
bb F Z \ Mn . Q
Q Accordingly M bb = for every b B and M is nowhere dense.
64
Boolean algebras
316I
An = int{qi : i n} = .
for every G A, and A {R}; which means that inf A 6= in RO(R), and 316G cannot be satised.
316K Atoms in Boolean algebras (a) If A is a Boolean algebra, an atom in A is a non-zero a A
such that the only elements included in a are 0 and a.
(b) A Boolean algebra is atomless if it has no atoms.
(c) A Boolean algebra is purely atomic if every non-zero element includes an atom.
316L Proposition Let A be a Boolean algebra, with Stone space Z.
(a) There is a one-to-one correspondence between atoms a of A and isolated points z Z, given by the
formula b
a = {z}.
(b) A is atomless i Z has no isolated points.
(c) A is purely atomic i the isolated points of Z form a dense subset of Z.
proof (a)(i) If z is an isolated point in Z, then {z} is an open-and-closed subset of Z, so is of the form b
a
for some a A; now if b a, bb must be either or {z}, so b must be either a or 0, and a is an atom.
(ii) If a A and b
a has two points z and w, then (because Z is Hausdor, 311I) there is an open set G
containing z but not w. Now there is a c A such that z b
c G, so that a c must be dierent from both
0 and a, and a is not an atom.
(b) This follows immediately from (a).
316Xu
Further topics
65
(b) Show that any principal ideal of a ccc Boolean algebra is ccc.
(c) Let hAi iiI be a family of Boolean algebras, with simple product A. Show that A is ccc i every Ai
is ccc and {i : Ai 6= {0}} is countable.
> (d) Let X be a separable topological space. Show that X is ccc.
> (e) Show that the regular open algebra of a topological space X is ccc i X is ccc, so that, in particular,
RO(R) is ccc.
(f ) Show that if A is a Boolean algebra and B is an order-dense subalgebra of A, then A is ccc i B is.
(g) Let A be a Dedekind -complete Boolean algebra. Show that it is ccc i there is no family ha i<1
in A such that a a whenever < < 1 .
(h) Let A be any Boolean algebra and I an order-closed ideal of A. Show that A/I is ccc i there is no
uncountable disjoint family in A \ I.
(i) Let A be a ccc Boolean algebra. Show that if I is a -ideal of A, then it is order-closed, and A/I is
ccc.
(j) Let A be a Boolean algebra. Show that the following are equiveridical: (i) A is ccc; (ii) every -ideal
of A is order-closed; (iii) every -subalgebra of A is order-closed; (iv) every sequentially order-continuous
Boolean homomorphism from A to another Boolean algebra is order-continuous. (Hint: 313Q.)
(k) Show that any principal ideal of a weakly (, )-distributive Boolean algebra is a weakly (, )distributive Boolean algebra.
(l) Let hAi iiI be a family of Boolean algebras, with simple product A. Show that A is weakly (, )distributive i every Ai is.
> (m) Show that if A is a weakly (, )-distributive Boolean algebra and B is a subalgebra of A which
is regularly embedded in A, then B is weakly (, )-distributive.
(n) Show that if A is a weakly (, )-distributive Boolean algebra and I is an order-closed ideal of A,
then A/I is weakly (, )-distributive.
> (o) (i) Show that if A is a Boolean algebra and B is an order-dense subalgebra of A, then A is weakly
(, )-distributive i B is. (ii) Let X be a zero-dimensional compact Hausdor space. Show that the
regular open algebra of X is weakly (, )-distributive i the algebra of open-and-closed subsets of X is.
> (p) Show that the algebra of open-and-closed subsets of {0, 1}N , with its usual topology, is not weakly
(, )-distributive.
(q) Let A be a Boolean algebra and B an order-dense subalgebra of A. Show that A and B have the
same atoms, so that A is atomless, or purely atomic, i B is.
(r) Let A be a Boolean algebra and B a regularly embedded subalgebra of A. Show that (i) every atom
of A is included in an atom of B (ii) if A is purely atomic, so is B (iii) if B is atomless, so is A.
> (s) Let A be a Dedekind complete purely atomic Boolean algebra. Show that it is isomorphic to PA,
where A is the set of atoms of A.
(t) Let A be a Boolean algebra and I an order-closed ideal of A. Show that (i) if A is atomless, so is
A/I (ii) if A is purely atomic, so is A/I.
(u) Let A be a Boolean algebra. Show that (i) if A is atomless, so is every principal ideal of A (ii) if A
is purely atomic, so is every principal ideal of A.
66
Boolean algebras
316Xv
(v) Let hAi iiI be a family of Boolean algebras with simple product A. Show that (i) A is purely atomic
i every Ai is (ii) A is atomless i every Ai is.
> (w) Show that any purely atomic Boolean algebra is weakly (, )-distributive.
(x) Let A be a Boolean algebra. Show that there is a one-to-one correspondence between atoms a of A
and order-continuous Boolean homomorphisms : A Z2 , dened by saying that corresponds to a i
(a) = 1.
316Y Further exercises (a) Let I be any set. Show that {0, 1}I , with its usual topology, is ccc. (Hint:
show that if E {0, 1}I is a non-empty open-and-closed set, then E > 0, where is the usual measure on
{0, 1}I .)
(b) Show that the Stone space of RO(R) is separable. More generally, show that if a topological space
X is separable so is the Stone space of its regular open algebra.
(c) Let A be a Boolean algebra and Z its Stone space. Show that A is ccc i every nowhere dense subset
of Z is included in a nowhere dense zero set.
(d) Let X be a zero-dimensional topological space. Show that X is ccc i the regular open algebra of X
is ccc i the algebra of open-and-closed subsets of X is ccc.
(e) Set X = {0, 1}1 , and for < 1 set E = {x : x X, x() = 1}. Let be the algebra of subsets
of X generated by {E : < 1 } {{x} : x X}, and I the -ideal of generated by {E E : < <
1 } {{x} : x X}. Show that /I is not ccc, but that there is no uncountable disjoint family in \ I.
(f ) Let X be a regular topological space and RO(X) its regular open algebra. Show that RO(X) is
weakly (, )-distributive i every meager set in X is nowhere dense.
(g) Let A be a Boolean algebra. A is weakly -distributive if whenever hAn inN is a sequence of
countable partitions of unity in A then there is a partition B of unity such that {a : a An , a b 6= 0}
is nite for every b B and n N. (Dedekind complete weakly -distributive algebras are also called
-bounding.) A has the Egorov property if whenever hamn im,nN is a double sequence in A such that
hamn inN is non-increasing and has inmum 0 for every m N, then there is a non-increasing sequence
ham imN such that inf mN am = 0 and for every m N there is an n N such that am amn . (i) Show that
if A has the Egorov property it is weakly -distributive. (ii) Show that if A is ccc then it is weakly (, )distributive i it has the Egorov property i it is weakly -distributive. (iii) Show that P(N N ) does not
have the Egorov property, even though it is weakly (, )-distributive. (Hint: try amn = {f : f (m) n}.)
(h) Let A be a Boolean algebra and Z its Stone space. (i) Show that A is weakly -distributive i the
union of any sequence of nowhere dense zero sets in Z is nowhere dense. (ii) Show that A has the Egorov
property i the union of any sequence of nowhere dense zero sets in Z is included in a nowhere dense zero
set.
(i) Let A be a Dedekind -complete weakly (, )-distributive Boolean algebra, Z its Stone space, E the
algebra of open-and-closed subsets of Z, M the -ideal of meager subsets of Z, and the Baire property
algebra {EM : E E, M M}, as in 314M. (i) Suppose that f : Z R is a -measurable function.
Show that there is a dense open set G Z such that f G is continuous. (ii) Now suppose that A is Dedekind
complete. Show that if f : Z R is a function such that f G is continuous for some dense open set G Z,
then f is -measurable; and that if f is also bounded, there is a continuous function g : Z R such that
{z : f (z) 6= g(z)} is meager. (Hint: the graph of g will be the closure of the graph of f G; because Z is
extremally disconnected, this is the graph of a function.)
(j) (i) Let X be a non-empty separable Hausdor space without isolated points. Show that its regular
open algebra is not weakly (, )-distributive. (ii) Let (X, ) be a non-empty metric space without isolated
points. Show that its regular open algebra is not weakly (, )-distributive. (iii) Let I be any innite set.
Show that the algebra of open-and-closed subsets of {0, 1}I is not weakly (, )-distributive. Show that the
regular open algebra of {0, 1}I is not weakly (, )-distributive.
316 Notes
Further topics
67
(i) Show that CX is an algebra of subsets of X (the finite-cofinite algebra). (ii) Show that a Boolean
algebra is purely atomic i it has an order-dense subalgebra isomorphic to the nite-conite algebra of
some set. (iii) Show that a Dedekind -complete Boolean algebra is purely atomic i it has an order-dense
subalgebra isomorphic to the countable-cocountable algebra of some set (211R).
(l) Let hAi iiI be a family of Boolean algebras, none of them {0}, with free product A. (i) Show that A
is purely atomic i every Ai is purely atomic and {i : Ai 6= {0, 1}} is nite. (ii) Show that A is atomless i
either some Ai is atomless or {i : Ai 6= {0, 1}} is innite.
(m) Show that a Boolean algebra is isomorphic to the algebra of open-and-closed subsets of {0, 1}N i it
is countable, atomless and not {0}.
(n) Show that a Boolean algebra is isomorphic to RO(R) i it is atomless, Dedekind complete, has a
countable order-dense subalgebra and is not {0}.
(o) Write [N]< for the ideal of PN consisting of the nite subsets of N. Show that PN/[N]< is atomless,
weakly (, )-distributive and not ccc.
(p) Let X be a Hausdor space and RO(X) its regular open algebra. (i) Show that the atoms of RO(X)
are precisely the sets {x} where x is an isolated point in X. (ii) Show that RO(X) is atomless i X has no
isolated points. (iii) Show that RO(X) is purely atomic i the set of isolated points in X is dense in X.
316 Notes and comments The phrase countable chain condition is perhaps unfortunate, since the
disjoint sets to which the denition 316A refers could more naturally be called antichains; but there is in
fact a connexion between countable chains and countable antichains (316Xg). While some authors speak of
the countable antichain condition or cac, the term ccc has become solidly established. In the Boolean
algebra context, it could equally well be called the countable sup property (316E).
The countable chain condition can be thought of as a restriction on the width of a Boolean algebra; it
means that the algebra cannot spread too far laterally (see 316Xc), though it may be indenitely complex
in other ways. Generally it means that in a wide variety of contexts we need look only at countable families
and monotonic sequences, rather than arbitrary families and directed sets (316E, 316F, 316Yg). Many of
the ideas of 316B-316F have already appeared in 215B; see 322G below.
I remarked in the notes to 313 that the distributive laws described in 313B have important generalizations, of which weak (, )-distributivity and its cousin weak -distributivity (316Yg) are two. They are
characteristic of the measure algebras which are the chief subject of this volume. The Egorov property
(316Yg) is an alternative formulation applicable to ccc spaces.
Of course every property of Boolean algebras has a reection in a topological property of their Stone
spaces; happily, the concepts of this section correspond to reasonably natural topological expressions (316B,
316I, 316L, 316Yh).
With four new properties (ccc, weakly (, )-distributive, atomless, purely atomic) to incorporate into
the constructions of the last few sections, a very large number of questions can be asked; most are elementary.
Any subalgebra of a ccc algebra is ccc (316Xa). All four properties are inherited by order-dense subalgebras
and principal ideals (316Xb, 316Xf, 316Xk, 316Xo, 316Xr, 316Xu); with the exception of the countable chain
condition (316Xc), they are inherited by simple products (316Xl, 316Xv); with the exception of atomlessness,
they are inherited by regularly embedded subalgebras (316Xm, 316Xr), and, in particular, by order-closed
subalgebras of Dedekind complete algebras. As for quotient algebras (equivalently, homomorphic images),
all four properties are inherited by order-continuous images (316Xi, 316Xn, 316Xt). The countable chain
condition is so important that it is worth noting that a sequentially order-continuous image of a ccc algebra
is ccc (316Xi), and that there is a useful necessary and sucient condition for a sequentially order-continuous
image of a -complete algebra to be ccc (316C, 316D, 316Xh; but see also 316Ye). To see that sequentially
order-continuous images do not inherit weak (, )-distributivity, recall that the regular open algebra of R
is isomorphic to the quotient of the Baire-property algebra Bb of R by the meager ideal M (314Yd); but that
68
Boolean algebras
316 Notes
Bb is purely atomic (since it contains all singletons), therefore weakly (, )-distributive (316Xw). Similarly,
PN/[N]< is a non-ccc image of a ccc algebra (316Yo).
The denitions here provide a language in which a remarkably interesting question can be asked: is
the free product of ccc Boolean algebras always ccc? equivalently, is the product of ccc topological spaces
always ccc? What is special about this question is that it cannot be answered within the ordinary rules of
mathematics (even including the axiom of choice); it is undecidable, in the same way that the continuum
hypothesis is. I will deal with a variety of undecidable questions in Volume 5; this particular one is treated
in Jech 78 and Fremlin 84. Note that the free product of two weakly (, )-distributive algebras need
not be weakly (, )-distributive (325Yd).
I have taken the opportunity to mention three of the most important of all Boolean algebras: the algebra
of open-and-closed subsets of {0, 1}N (316Ym), the regular open algebra of R (316J, 316Yn) and the quotient
PN/[N]< (316Yo). A fourth algebra which belongs in this company is the Lebesgue measure algebra, which
is atomless, ccc and weakly (, )-distributive (so that every countable subset of its Stone space Z is nowhere
dense, and Z is a non-separable ccc space); but for this I wait for the next chapter.
321Be
69
Measure algebras
Chapter 32
Measure algebras
I now come to the real work of this volume, the study of the Boolean algebras of equivalence classes of
measurable sets. In this chapter I work through the elementary theory, dening this to consist of the parts
which do not depend on Maharams theorem or the lifting theorem or non-trivial set theory.
321 gives the denition of measure algebra, and relates this idea to its origin as the quotient of a algebra of measurable sets by a -ideal of negligible sets, both in its elementary properties (following those of
measure spaces treated in 112) and in an appropriate version of the Stone representation. 322 deals with
the classication of measure algebras according to the scheme already worked out in 211 for measure spaces.
323 discusses the canonical topology and uniformity of a measure algebra. 324 contains results concerning
Boolean homomorphisms between measure algebras, with the relationships between topological continuity,
order-continuity and preservation of measure. 325 is devoted to the measure algebras of product measures,
and their abstract characterization. Finally, 326-327 address the properties of additive functionals on
Boolean algebras, generalizing the ideas of Chapter 23.
321 Measure algebras
I begin by dening measure algebra and relating this concept to the work of Chapter 31 and to the
elementary properties of measure spaces.
321A Definition A measure algebra is a pair (A,
), where A is a Dedekind -complete Boolean
algebra and
: A [0, ] is a function such that
0 = 0;
P
whenever han inN is a disjoint sequence in A,
(supnN an ) = n=0
an ;
(a b) =
(supnN an ) = n=0
an =
a +
b. Q
Q
a
a +
(b \ a) =
b. Q
Q
(a b) =
a +
(b \ a)
a +
b. Q
Q
P
(d) If han inN is any sequence in A, then
(supnN an ) n=0
an . P
P For each n, set bn = an \ supi<n ai .
Inducing on n, we see that supin ai = supin bi for each n, so supnN an = supnN bn and
P
P
(supnN an ) =
(supnN bn ) = n=0
bn n=0
an
because hbn inN is disjoint. Q
Q
(sup an ) =
(sup bn ) =
nN
nN
= lim
n
X
i=0
bn
n=0
bi = lim
(sup bi ) = lim
an . Q
Q
n
in
70
321Bf
Measure algebras
a =
ak
(ak \ a) =
ak lim
bn
n
b = limn
bn ,
an
b for every n N,
so
b = .
If a A, then for every n N there is an an A such that a an
an , so that
(a \ b)
(a \ an )
(an \ an ) =
an
an
an 2n .
(sup A) = aA
a.
(sup A) =
(sup B) = supbB
b = aA
(inf A) =
a0
(sup B) =
a0 sup
b
bB
= inf
(a0 \ b) = inf
(a0 a) = inf
a.
bB
aA
aA
321J
Measure algebras
71
Proposition Let (X, , ) be a measure space, and N the null ideal of . Let A be the Boolean algebra
quotient / N . Then we have a functional
: A [0, ] dened by setting
E = E for every E ,
and (A,
) is a measure algebra. The canonical map E 7 E : A is sequentially order-continuous.
E F + (E \ F ) = F E + (F \ E) = E
0 =
= = 0.
If hanS
inN is a disjoint sequence in A, choose for each n N an En such that En = an . Set Fn =
En \ i<n Ei ; then
Fn = En \ supi<n Ei = an \ supi<n ai = an
S
S
for each n, so
an = Fn for each n. Now set E = nN En = nN Fn ; then E = supnN Fn = supnN an .
So
P
P
E =
(E).
E = 0
(E) = 0 E = 0 E M.
E = E =
(E) =
(E )
for every E . Thus the Boolean algebra isomorphism is also an isomorphism between the measure
algebras (/M, ) and (A,
), and (A,
) is represented in the required form.
72
Measure algebras
321K
321K Definition I will call the measure space (Z, , ) constructed in the proof of 321J the Stone
space of the measure algebra (A,
).
For later reference, I repeat the description of this space as developed in 311E, 311I, 314M and 321J. Z
is a compact Hausdor space, being the Stone space of A. A can be identied with the algebra of openand-closed sets in Z. The null ideal of coincides with the ideal of meager subsets of Z; in particular, is
complete. The measurable sets are precisely those expressible in the form E = b
aM where a A, b
a Z is
the corresponding open-and-closed set, and M is meager; in this case E =
a and a = E is the member
of A corresponding to E.
For the most important classes of measure algebras, more can be said; see 322N et seq. below.
321X Basic exercises (a) Let (A,
) be a measure algebra, and a A. Show that (Aa ,
Aa ) is a
measure algebra, writing Aa for the principal ideal of A generated by a.
(b) Let (X, ,
) be a measure space, and A its measure algebra. (i) Show that if T is a -subalgebra
of , then {E : E T} is a -subalgebra of A. (ii) Show that if B is a -subalgebra of A then {E : E
, E B} is a -subalgebra of .
321Y Further exercises (a) Let (A,
) be a measure algebra, and I A a -ideal. For u A/I set
u = inf{
a : a A, a = u}. Show that (A/I,
) is a measure algebra.
321 Notes and comments The idea behind taking the quotient /N , where is the algebra of measurable
sets and N is the null ideal, 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 dier by a negligible set, that is, if they represent the same member of the measure algebra.
The denition in 321A is designed to be an exact characterization of these quotient algebras, taking into
account 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 signicantly in translation. In measure algebras, it becomes sensible to
take seriously the suprema and inma of uncountable sets (see 321C-321F).
I should perhaps remark that while the Stone representation (321J-321K) is signicant, it is not the most
important method of representing measure algebras, which is surely Maharams 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.
322B
73
a such that
b < .
(e) (A,
) is localizable or Maharam if it is semi-nite and the Boolean algebra A is Dedekind complete.
322B The rst step is to relate these concepts to the corresponding ones for measure spaces.
Theorem Let (X, , ) be a measure space, and (A,
) its measure algebra. Then
(a) (X, , ) is a probability space i (A,
) is a probability algebra;
(b) (X, , ) is totally nite i (A,
) is;
(c) (X, , ) is -nite i (A,
) is;
(d) (X, , ) is semi-nite i (A,
) is;
(e) (X, , ) is localizable i (A,
) is;
(f) if E , then E is an atom for i E is an atom in A;
(g) (X, , ) is atomless i A is;
(h) (X, , ) is purely atomic i A is.
proof (a), (b) are trivial, since
1 = X.
(c)(i) If is -nite, let hEn inN be a sequence of sets of nite measure covering X; then
En < for
every n, and
S
supnN En = ( nN En ) = 1,
so (A,
) is -nite.
(ii) If (A,
) is -nite, let han inN be a sequence in A such
an < for every n and supnN an = 1.
Sthat
For each n, choose En such that En = an . Set E = nN En ; then E = supnN an = 1, so E is
conegligible. Now (X \ E, E0 , E1 , . . . ) is a sequence of sets of nite measure covering X, so is -nite.
(d)(i) Suppose that is semi-nite and that a A,
a = . Then there is an E such that E = a,
so that E =
a = . As is semi-nite, there is an F such that F E and 0 < F < . Set
b = F ; then b a and 0 <
b < .
(ii) Suppose that (A,
) is semi-nite and that E , E = . Then
E = , so there is a
E \ F = 0 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 i H F and H is a lower bound for B; that is, i H = sup A. Q
Q
74
322B
Measure algebras
(ii) Thus A is Dedekind complete i every family in has an essential supremum in . Since we
already know that (A,
) is semi-nite i is, we see that (A,
) is localizable i is.
(f ) This is immediate from the denitions in 211I and 316K, if we remember always that {b : b
{F : F , F E} (312Kb).
E} =
Thus is a Boolean algebra isomorphism. It is a measure algebra isomorphism because for any E
1 (E ) =
1 E =
E = E =
E .
2 (a a ) =
2 (a a ) =
(a a ) > 0,
so a 6= a ; thus Af is an injective map from Af to Af2 . If c Af2 , then c = G where
G < ; by
213Fc, there is an E such that E G, E =
G and
(G \ E) = 0, so that E Af and
E = E = G = c.
322F
75
Finally, is order-continuous. P
P Let A A be a non-empty downwards-directed set with inmum 0,
and b A2 a lower bound for [A]. ?? If b 6= 0, then (because (A2 ,
2 ) is semi-nite) there is a b0 Af2 such
that 0 6= b0 b. Let a0 A be such that a0 = 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 Af is injective,
b0 = a0 6= (a a0 ) = a a0 = a b0 ,
2 (E ) =
2 E =
E = E =
E
a = 0 = 0 =
2 (a) =
a = a = 0,
E = E = 0.
(b a) = 0, by 321F; so we may
If c A \ {0}, let b c be such that 0 <
b < . For each n N, inf aAn
choose an An such that
(b an ) 2n2
b. Set a = supnN an A. Then
P
(b an ) <
b,
(b a) n=0
76
Measure algebras
322G
a
>
0
for
some
n,
so
there
are
>
0
such
that
an = 1.
n
n
n
n=0
P
P
an ), n = n /( i=0 i
(Set n = 2n /(1 +
ai ).) Set a = n=0 n
(a an ) for every a A; it is easy to
check that (A, ) is a probability algebra.
(iii)(i) is a consequence of (ii)(i).
322H Principal ideals If (A,
) is a measure algebra and a A, then it is easy to see (using 314Eb)
that (Aa ,
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, , ) be a measure space, and A X a set with a measurable envelope E. Let A be
the subspace measure on A, and A its domain; let (A,
) be the measure algebra of (X, , ) and (AA ,
A )
the measure algebra of (A, A , A ). Set a = E and let Aa be the principal ideal of A generated by a. Then
we have an isomorphism between (Aa ,
Aa ) and (AA ,
A ) given by the formula
F 7 (F A)
whenever F 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 E = {E F : F }. For F , G E ,
F = G (F G) = 0 A (A (F G)) = 0 (F A) = (G A) ,
because E is a measurable envelope of A. Accordingly the given formula denes an injective function from
the image {F : F E } of E 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
A = {F A : F E }. Finally, it is measure-preserving because
F = F = (F A) = A (F A) =
A (F A)
322J Corollary Let (X, , ) be a measure space, with measure algebra (A,
).
(a) If E , then the measure algebra of the subspace measure E can be identied with the principal
ideal AE of A.
(b) If A X is a set of full outer measure (in particular, if A = X < ), then the measure algebra
of the subspace measure A can be identied with A.
322K Simple products (a) Q
Let h(Ai ,
i )iiI be any indexed family of P
measure algebras. Let A be the
simple product Boolean algebra iI Ai (315A), and for a A set
a = iI
i a(i). Then it is easy to
check (using 315D(e-ii)) that (A,
) is a measure algebra; I will call it the simple product of the family
h(Ai ,
i )iiI . 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
Aei ).
between (Ai ,
i ) and (Aei ,
*322L
77
(b) If h(Xi , i , i )iiI is a family of measure spaces, with direct sum (X, , ) (214K), then the measure
algebra (A,
) of (X, , ) can be identied with the simple product of the measure algebras (Ai ,
i ) of
the (Xi , i , i ). 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 iiI : iI i induces a Boolean
isomorphism from A to iI Ai , which is also a measure algebra isomorphism, because
P
P
E = E = iI i Ei = iI
i Ei
for every E . Q
Q
(c) A product of measure algebras is semi-nite, or localizable, or atomless, or purely atomic, i every
factor is. (Compare 214Jb.)
(d) Let (A,
) be a localizable measure algebra.
Q
Aei )
(i) If hei iiI is any partition of unity in A, then (A,
) is isomorphic to the product iI (Aei ,
of the corresponding
principal
ideals.
P
P
By
315F(iii),
the
map
a
7
ha
ei iiI is a Boolean isomorphism
P
Q
(a ei ) (321E), for
a = iI Q
between A and iI Ai . Because hei iiI is disjoint and a = supiI a ei ,
every a A. So a 7 ha ei iiI is a measure algebra isomorphism between (A,
) and iI (Ai ,
Q
Aei ). Q
(ii) In particular, since A has a partition of unity consisting of elements of nite measure (322Ea),
(A,
) is isomorphic to a simple product of totally nite measure algebras. Each of these is isomorphic to
the measure algebra of a totally nite measure space, so (A,
) is isomorphic to the measure algebra of a
direct sum of totally nite 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, , ) be a strictly localizable measure space with X > 0, and (A,
) its measure
algebra. If hai iiI is a partition of unity in A, there is a partition hXi iiI of X into members of such that
Xi = ai for every i I and
= {E : E X, E Xi i I},
P
E = iI (E Xi ) for every E ;
Q
that is, the isomorphism between A and the simple product iI Aai of its principal ideals (315F) corresponds to an isomorphism between (X, , ) and the direct sum of the subspace measures on Xi .
proof (a) Suppose to begin with that X < . In this case J = {i : S
ai 6= 0} must be countable (322G).
For each i J, choose Ei such that Ei = ai , and set Fi = Ei \ jJ,j6=i Ej ; then Fi = ai for each
i J, and hFi iiJ is disjoint. Because X > 0, J is non-empty; x some j0 J and set
[
Xi = Fj0 (X \
Fj ) if i = j0 ,
jJ
= Fi for i J \ {j0 },
= for i I \ J.
S
Then hXi iiI is a disjoint family in , iI Xi = X and Xi = ai for every i. Moreover, because only
countably many of the Xi are non-empty, we certainly have
= {E : E X, E Xi i I},
P
E = iI (E Xi ) for every E .
(b) For the general case, start by taking a decomposition hYj ijJ of X. We can suppose that no Yj is
negligible,
because there is certainly some j0 such that Yj0 > 0, and we can if necessary replace Yj0 by
S
Yj0 {Yj : Yj = 0}. For each j, we can identify the measure algebra of the subspace measure on Yj with
78
Measure algebras
*322L
just above we can nd a disjoint family hXji iiI in such that iI Xji = Yj , Xji
= ai bj for every i and
PYj = {E : E Yj , E Xji i I},
P
E = iI (E Xji ) for every E PYj .
S
Set Xi = jI Xji for every i I. Then hXi iiI is disjoint and covers X. Because Xi Yj = Xji is
measurable for every j, Xi . Because Xi ai bj for every j, and hbj ijJ is a partition of unity in A
(322Kb), Xi ai for each i; because hXi iiI is disjoint and supiI ai = 1, Xi = ai for every i. If E X is
such that E Xi for every i, then E Xji for all i I and j J, so E Yj for every j J
and E . If E , then
X
X
X
X
E =
(E Yj ) =
(E Xji ) =
sup
(E Xji )
jJ
so E =
iI
X
iI
jJ,iI
(E Xi ) =
sup
KI is finite iK
iI KJ is finite jK
(E Xi ) E;
and M for the ideal of meager subsets of Z, that is, the null ideal of (314M, 321K). Then a 7 b
a : A
/M is an isomorphism between (A,
) and the measure algebra of (Z, , ) (314M). Note that because any
subset of a meager set is meager, is surely complete.
322O
79
so (W V ) b
a is negligible, therefore meager. But we know that A is weakly (, )-distributive (322F), so
that meager sets in Z are nowhere dense (316I), and there is a non-empty open set H b
a \ (W V ). Now
Q
H G \ W V . As G is arbitrary, int W V = and W V is nowhere dense. Q
But this means that W V M and V = W (W V ) . As V is arbitrary, is locally
determined.
(c)(iii)(iv) Assume that is locally determined. Because (A,
) is semi-nite, there is a partition of
unity C A consisting of elements of nite measure (322Ea). Set C = {b
c : c C}. This is a disjoint family
of sets of nite measure for . Now suppose that F and F > 0. Then there is an open-and-closed set
E Z such that F E is meager, and E is of the form b
a for some a A. Since
a = b
a = F > 0,
(F b
c) =
(a c) > 0.
This means that satises 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,
) be a semi-nite measure algebra, and let A
b
b
of A (314U). Then there is a unique extension of
to a functional
on A such that (A,
) is a localizable
b identies the ideals {a : a A,
b
measure algebra. The embedding A A
a < } and {a : a A,
a < }.
b For c A,
b set
proof (I write the argument out as if A were actually a subalgebra of A.)
c = sup{
a : a A, a c}.
b to [0, ] extending
b
Evidently
is a function from A
, so
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 inN is a disjoint sequence in A
n=0
n=0
an : an A, an
sup{
an : an
cn }
cn
for every n N}
cn .
n=0
b
c. But this means that if a A, a c then a = supaA a a in A and therefore also in A; so that
Accordingly
a = supaA
(a a) supaA
a.
c = supaA
a =
n=0
cn . Q
Q
b
This shows that (A,
) is a measure algebra. It is semi-nite because (A,
) is and every non-zero element
b
b
of A includes a non-zero element of A, which in turn includes a non-zero element of nite measure. Since A
b
is Dedekind complete, (A,
) is localizable.
80
Measure algebras
322O
If
a is nite, then surely
a =
a is nite. If
c is nite, then {a : a A, a c} is upwards-directed
b b=c
and supaA
a =
c is nite, so b = sup A is dened in A and
b =
c. Because A is order-dense in A,
(313K, 313O) and c A, with
c =
c.
b
322P Definition Let (A,
) be any semi-nite measure algebra. I will call (A,
), as constructed above,
the localization of (A,
). 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,
) be a semi-nite measure algebra
and (Z, , ) its Stone space.
(a) Meager sets in Z are nowhere dense; every E is uniquely expressible as GM where G Z is
open-and-closed and M is nowhere dense, and E = sup{H : H E is open-and-closed}.
(b) The c.l.d. version of is strictly localizable, and has the same negligible sets as .
(c) If (A,
) is totally nite then E = inf{H : H E is open-and-closed} for every E .
proof (a) I have already remarked (in the proof of 322N) that A is weakly (, )-distributive, so that
meager sets in Z are nowhere dense. But we know that every member of is expressible as GM where G
is open-and-closed and M is meager, therefore nowhere dense. Moreover, the expression is unique, because
if GM = G M then GG M M is open and nowhere dense, therefore empty, so G = G and
M = M .
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
E = G =
a = sup
b
bB
(b) This is the same as part (c) of the proof of 322N. We have a disjoint family C of sets of nite measure
for such that whenever E , E > 0 there is a C C such that (C E) > 0. Now if F is dened
and not 0, there is an E such that E F and E > 0 (213Fc), so that there is a C C such that
(E C) > 0; since C < , we have
(F C) (E C) = (E C) > 0.
And of course C < for every C C. This means that C witnesses that satises the conditions of 213O,
so that is strictly localizable.
Any -negligible set is surely -negligible. If M is -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 , so there is a
non-empty open-and-closed set H H1 such that H is nite (because is semi-nite). In this case H M
is -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 -negligible.
Thus and have the same negligible sets.
(c) Because Z < ,
E = Z (Z \ E) = Z sup{H : H Z \ E is open-and-closed}
= inf{(Z \ H) : H Z \ E is open-and-closed}
= inf{H : H E is open-and-closed}.
322Yf
81
a,
b < } for
(b) Let (X, , ) be a measure space and let sf be the semi-nite version of , as dened in 213Xc.
Let (A,
) be the measure algebra of (X, , ). Show that the measure algebra of (X, , sf ) is isomorphic
to the measure algebra (A/I ,
sf ) of (a) above.
(c) Let (X, , ) be a measure space and (X, ,
) its c.l.d. version. Let (A,
) and (A2 ,
2 ) be the
corresponding measure algebras, and : A A2 the canonical homomorphism, as in 322Db. Show that
the kernel of is the ideal I , as described in 322Xa, so that A/I is isomorphic, as Boolean algebra, to
[A] A2 . Show that this isomorphism identies
sf , as described in 322Xa, with
2 [A].
(d) Give a direct proof of 322G, not relying on 215B and 321J.
> (e) Let (A,
) be any measure algebra, A a non-empty subset of A, and c A such that
c < .
Show that (i) c0 = sup{a c : a A} is dened in A (ii) there is a countable set B A such that
c0 = sup{a c : a B}.
> (f ) Let (X, , ) be a measure space and an indenite-integral measure over (234). Show that the
measure algebra of can be identied, as Boolean algebra, with a principal ideal of the measure algebra of
.
(g) Let (X, , ) be a measure space and A any subset of X; let A be the subspace measure on A and
A its domain. Write (A,
) for the measure algebra of (X, , ) and (AA ,
A ) for the measure algebra
of (A, A , A ). Show that the formula F 7 (F A) denes a sequentially order-continuous Boolean
homomorphism : A AA which has kernel I = {F : F , F A = }. Show that for any a A,
A (a) = min{
b : b A, a \ b I}.
(h) Let (A,
) be a measure algebra and B an order-closed subalgebra of A. Suppose that (B,
B) is
semi-nite. Show that (A,
) is semi-nite.
(i) Let (A,
) be any measure algebra and (Z, , ) its Stone space. Show that the c.l.d. version of is
strictly localizable.
322Y Further exercises (a) Let X be a set, a -algebra of subsets of X, and I a -ideal of . Set
= {EN : E , N N }.
N = {N : F I, N F }. Show that N is a -ideal of subsets of X. Set
is a -algebra of subsets of X and that /N
Show that
is isomorphic to /I.
(b) Let (A,
) be a semi-nite measure algebra, and (Z, , ) its Stone space. Let be the c.l.d. version
is meager}, so that /M can be identied with the regular open algebra of Z (314Yd) and the measure
algebra of can be identied with the localization of A.
(c) Give an example of a localizable measure algebra (A,
) with a -subalgebra B such that (B,
B)
is semi-nite and atomless, but A is not atomless.
(d) Let (X, , ) be a measure space and A X a subset; let A be the subspace measure on A, A
and AA the measure algebras of and A , and : A AA the canonical homomorphism, as described in
322Xg. (i) Show that if A is semi-nite, then is order-continuous. (ii) Show that if is semi-nite but
A is not, then is not order-continuous.
(e) Show that if (A,
) is a semi-nite measure algebra, with Stone space (Z, , ), then has locally
determined negligible sets in the sense of 213I.
(f ) Let (A,
) be a localizable measure algebra and (Z, , ) its Stone space. (i) Show that a function
f : Z R is -measurable i there is a conegligible set G X such that f G is continuous. (Hint: 316Yi.)
(ii) Show that f : Z [0, 1] is -measurable i there is a continuous function g : Z [0, 1] such that
f = g -a.e.
82
Measure algebras
322 Notes
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 terms 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 terms 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,
), the set Af of elements of nite measure is an ideal of A; consequently
it is order-dense i it includes a partition of unity (322E). In 322F we have something deeper: any seminite measure algebra must be weakly (, )-distributive when regarded as a Boolean algebra, and this has
signicant 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-niteness of the measure algebra, since any Dedekind -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 sucient to make it measurable, in the sense that it
should carry a semi-nite 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 eect 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 seminite measure algebra is localizable i 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 eective
construction, at least for semi-nite measure algebras, that of localization (322O-322P); I say that it is
more eective 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, 365M, 367N,
369A, 369C). Note that the localization of a semi-nite 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
even when A can be eectively represented as the measure algebra of a measure on the Stone space of A.
b not just
What is happening in 322Yb is that we are using all the open sets of Z to represent members of A,
the open-and-closed sets, which correspond to members of A.
323C
83
a (b, d) =
(a (b d))
((a (b c)) (a (c d)))
(a (b c)) +
(a (c d)) = a (b, c) + a (c, d).
(iii) If b, c A then
a (b, c) =
(a (b c)) =
(a (c b)) = a (c, b). Q
Q
(a b) for all a, b Af , which is sometimes useful (323Xg). I will call the corresponding topology on Af
the strong measure-algebra topology.
323B Proposition Let (A,
) be any measure algebra. Then the operations
uniformly continuous.
, , \
and
are all
(b b ) (c c )
a (b c, b c ) a (b, b ) + a (c, c ).
|
a
b|
(a b) = 1 (a, b).
c
(a c) =
a
(a (b \ c)) > .
Thus {b :
b > } is open; as is arbitrary,
is lower semi-continuous.
(c) |
(a b)
(a c)| a (b, c) for all b, c A.
84
Measure algebras
323D
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,
) 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 lter F(B ) on A.
(ii) If sup B is dened 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 lter F(B) on A.
(ii) If inf B is dened 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,
) is semi-nite.
(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 , a from 323A.
(a)(i) () 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 }
a (c, c )
(a c \ b) +
(a c \ b) =
(a c) +
(a c ) 2
(a b) .
(a b)
(a e) . Then for any c Fb , e c e \ b, so
a (e, c) =
(a (e c))
(a (e \ b)) =
(a e)
(a b) .
323G
85
bn m=n
(cm cm+1 ) = n
for each n. If m n, then
cm cn
supnk<m ck ck+1
bn ,
so
cn \ bn
cm cn bn .
Consequently
cn \ bn
for every m n, and
inf km ck
cn \ bn
d0
supkm ck
d1
cn bn
cn bn ,
so that
cn d0
bn ,
cn d1
bn ,
d1 \ d0
bn .
(d1 d0 ) inf nN
bn = 0,
so that d1 = d0 .
323G The classification of measure algebras: Theorem Let (A,
) be a measure algebra, T its
measure-algebra topology and U its measure-algebra uniformity.
(a) (A,
) is semi-nite i T is Hausdor.
(b) (A,
) is -nite i T is metrizable, and in this case U is also metrizable.
(c) (A,
) is localizable i T is Hausdor and A is complete under U.
proof I use the notations Af , a from 323A.
(a)(i) Suppose that (A,
) is semi-nite and that b, c are distinct members of A. Then there is an
a b c such that 0 <
a < , and now a (b, c) > 0. As b and c are arbitrary, T is Hausdor (2A3L).
(ii) Suppose that T is Hausdor and that b A has
b = . Then b 6= 0 so there must be an a Af
such that
(a b) = a (0, b) > 0; in which case a b b and 0 <
(a b) < . As b is arbitrary,
is
semi-nite.
86
Measure algebras
322G
X
an (b, c)
1
+ 2n
an
n=0
+ (1 + 2n
an )(b, c) .
In the other direction, given > 0, take n N such that 2n 12 ; then (b, c) whenever an (b, c)
/2(n + 1).
This shows that U is the same as the metrizable uniformity dened by {}; accordingly T is also dened
by .
(ii) Now suppose that T is metrizable, and let be a metric dening T. For each n N there must be
an0 , . . . , ankn Af and n > 0 such that
ani (b, 1) n for every i kn = (b, 1) 2n .
Set d = supnN,ikn ani . Then ani (d, 1) = 0 for every n, i, so (d, 1) 2n for every n and d = 1. Thus 1
is the supremum of countably many elements of nite measure and (A,
) is -nite.
(c)(i) Suppose that (A,
) is localizable. Then T is Hausdor, by (a). Let F be a Cauchy lter on A. For
each a Af , choose
a
sequence
hFn (a)inN in F such that a (b, c) 2n whenever b, c Fn (a) and n N.
T
Choose can kn Fk (a) for each n; then a (can , ca,n+1 ) 2n for each n. Set da = supnN inf kn a cak .
Then
limn a (da , can ) = limn
(da (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
a (da , db ) a (da , can ) + a (can , e) + a (e, cbn ) + a (cbn , db )
a (da , can ) + a (can , e) + b (e, cbn ) + b (cbn , db )
Thus
{e : a (d, e) } Fn (a) F.
As a, are arbitrary, F converges to d. Q
Q As F is arbitrary, A is complete.
323K
87
(ii) Now suppose that T is Hausdor and that A is complete under U. By (a), (A,
) is semi-nite.
Let B be any non-empty subset of A, and set B = {b0 . . . bn : b0 , . . . , bn B}, so that B is upwardsdirected and has the same upper bounds as B. By 323Da, we have a Cauchy lter F(B ); because A is
complete, this is convergent; and because (A,
) is semi-nite, its limit must be sup B = sup B, by 323Dd.
As B is arbitrary, A is Dedekind complete, so (A,
) 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 nd it helpful to read the next theorem rst on
the assumption that (A,
) is a probability algebra.
Theorem Let (A,
) be a localizable measure algebra, and B a subalgebra of A. Then it is closed for the
measure-algebra topology i 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 a (can , e) 2n .
Then
n=0
n=0
n=0
a (can , ca,n+1 )
a (can , e) + a (e, ca,n+1 ) < .
as n , by 323F, and a (e, ea ) = 0, that is, a ea = a e. Also, because B is order-closed, inf kn cak B
for every n, and ea B.
Because A is Dedekind complete, we can set
then ea B and
Now ea
eb
ea = inf{eb : b Af , a
b};
because (A,
) is semi-nite. 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,
) 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 B for the smallest order-closed subset of A including B. By 313Fc, B 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 B . Thus B = B is the order-closed subalgebra of A generated by B.
323K I note some simple results for future reference.
Lemma If (A,
) 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.
88
323L
Measure algebras
iI
jJ
X
jJ
1
2
whenever b, c are such that j (b(j), c(j)) /(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,
) be a semi-nite measure algebra. Show that the set {(a, b) : a b}
is a closed set in A A.
> (b) Let (X, , ) be a -nite measure space and (A,
) its measure algebra. (i) Show that if T is a
-subalgebra of , then {F : F T} is a closed subalgebra of A. (ii) Show that if B is a closed subalgebra
of A, then {F : F , F B} is a -subalgebra of .
(c) Let (A,
) 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,
) 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,
) be a measure algebra, and e A; let Ae be the principal ideal of A generated by e, and
e
its measure (322H). Show that the topology on Ae dened by
e is just the subspace topology induced by
the measure-algebra topology of A.
> (f ) Let (X, , ) be a measure space, and (A,
) its measure algebra. (i) Show that we have an
injection : A L0 () (see 241) given by setting (E ) = (E) for every E . (ii) Show that is a
homeomorphism between A and its image if A is given its measure-algebra topology and L0 () is given its
topology of convergence in measure (245A).
(g) Let (A,
) be a measure algebra and the measure metric on the ideal Af of elements of nite
measure. (i) Show that (Af , ) is a complete metric space and that the operations , , \ and are
uniformly continuous on Af , while
: Af R is also uniformly continuous. (ii) Show that the embedding
f
A A is continuous for the measure-algebra topology on A. (iii) In the context of 323Xf, show that
: Af L0 () is an isometry between Af and a subset of L1 ().
323Y Further exercises (a) Let (A,
) be a -nite measure algebra. Show that a set F A is closed
for the measure-algebra topology i 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}.
324A
Homomorphisms
89
(b) Give an example to show that (a) is false for general localizable measure algebras.
(c) Give an example of a semi-nite measure algebra (A,
) with an order-closed subalgebra which is not
closed for the measure-algebra topology.
(d) Let (A,
) be a probability algebra and write B for the family of closed subalgebras of A. For B,
C B set (B, C) = supbB inf cC
(b c) + supcC inf bB
(b c). Show that (B, ) is a complete metric
space.
(e) Let (A,
) be the measure algebra of Lebesgue measure on R. Show that it is separable in its
measure-algebra topology. (Hint: 245Yj.)
323 Notes and comments The message of this section is that the topology of a measure algebra is
essentially dened by its order and algebraic structure; see also 324F-324H below. Of course the results are
really about semi-nite 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 terms of the ordering, as in 323H and 324G. For -nite
algebras, indeed, there is an easy description of the topology in terms of the order (323Ya). I think the result
of this section which I shall most often wish to quote is 323H: in most contexts, there is no need to distinguish
between closed subalgebra and order-closed subalgebra. However a -subalgebra of a localizable measure
algebra need not be topologically sequentially closed; I present an example in Fremlin n02.
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-324P).
324A Theorem Let (X, , ) and (Y, T, ) be measure spaces, and (A,
), (B, ) their measure algebras.
90
Measure algebras
324A
. Since (EE1 ) \ G belongs to and is disjoint from D, it is negligible; accordingly EE1 is negligible
and E = E1 in A.
What this means is that the formula oered denes a map : B A. It is now easy to check that is
a Boolean homomorphism, because if
are negligible, so are
(E D)1 [F ],
((X \ E) D)1 [Y \ F ],
(E D)1 [F ]
((E E ) D)1 [F F ].
To see that is sequentially order-continuous, let hbn inN be a sequence in B. For each n we may
choose
that Fn = bn , and En such that (En D)1 [Fn ] is negligible; now, setting
S an Fn T such S
F = nN Fn and E = nN En ,
S
(E D)1 [F ] nN (En D)1 [Fn ]
is negligible, so
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, , ),
(Y, T, ) can be realized by functions between X and Y is important and deep; I will return to it in 343-344.
(c) In the simplied context of 324B, I have actually dened a contravariant functor; the relevant facts
are the following.
324D Proposition Let (X, , ), (Y, T, ) and (Z, , ) be measure spaces, with measure algebras (A,
),
Suppose that : X Y and : Y Z satisfy the conditions of 324B, that is,
(B, ), (C, ).
1 [F ] if F T,
1 [F ] = 0 if F = 0,
1 [G] T if G ,
1 [G] = 0 if G = 0.
324F
Homomorphisms
91
as required.
324F I turn now to the behaviour of order-continuous homomorphisms between measure algebras.
Theorem Let (A,
) and (B, ) be measure algebras and : A B a Boolean homomorphism.
(a) is continuous i it is continuous at 0 i it is uniformly continuous.
(b) If (B, ) is semi-nite and is continuous, then it is order-continuous.
(c) If (A,
) is semi-nite and is order-continuous, then it is continuous.
proof I use the notations Af , a from 323A.
(a) Suppose that is continuous at 0; I seek to show that it is uniformly continuous. Take b Bf and
> 0. Then there are a0 , . . . , an Af and > 0 such that
setting a = supin ai ,
b (c, c ) = (b (c c )) = (b (c c )) .
As b and are arbitrary, is uniformly continuous. The rest of the implications are elementary.
(b) Let A be a non-empty downwards-directed set in A with inmum 0. Then 0 A (323D(b-ii)); because
is continuous, 0 [A]. ?? If b is a non-zero lower bound for [A] in B, then (because (B, ) is semi-nite)
there is a c b with 0 < c < ; now
c (a, 0) = (c a) = c > 0
for every a A, so 0
/ [A]. X
X
Thus inf [A] = 0 in B; as A is arbitrary, is order-continuous (313L(b-ii)).
(c) By (a), it will be enough to show that is continuous at 0. Let b Bf , > 0. ?? Suppose, if possible,
that for every a Af , > 0 there is a c A such that
(a c) but (b c) . For each a Af ,
n
n N choose can such that
(a can ) 2
but (b can ) . Set ca = inf nN supmn cam ; then
P
(a ca ) inf nN m=n
(a can ) = 0,
so ca a = 0. On the other hand, because is order-continuous, ca = inf nN supmn cam , so that
b (c, 0) = (b c)
92
Measure algebras
324F
a (c, 0) =
(a c) .
324O
Homomorphisms
93
c0 = inf cC
c 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 d and b (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 a
atomic, there is an atom b of B with b a . Set
C = {c : c A, c
a , b c}.
c0 = inf cC
c b > 0,
F = F =
F =
1 [F ] = 1 [F ].
that is, : A0 B is an isometry. Because A0 is dense in the metric space (A, ), while B is complete
under (323Gc), there is a unique continuous function
: A B extending (3A4G). Now the operations
(a, a ) 7
(a a ),
(a, a ) 7
a
a : A A B,
are continuous and agree on the dense subset A0 A0 of A A; because the topology of B is Hausdor,
they agree on A A, that is,
(a a ) =
a
a for all a, a A (2A3Uc). Similarly, the operations
a 7
(1 \ a),
a 7 1 \
a : A B
are continuous and agree on the dense subset A0 of A, so they agree on A, that is,
(1 \ a) = 1 \ a for every
a A. Thus
is a Boolean homomorphism. To see that it is measure-preserving, observe that
a 7
a = (a, 0),
a 7 (
a) = (
a, 0) : A R
94
Measure algebras
*324P
*324P The following fact will be useful in 387, by which time it will seem perfectly elementary; for the
moment, it may be a useful exercise.
Proposition Let (A,
) and (B, ) be totally nite measure algebras such that
1 = 1. Suppose that
A A and : A B are such that (inf in ai ) =
(inf in ai ) for all a0 , . . . , an A. Let C be
the smallest closed subalgebra of A including A. Then has a unique extension to a measure-preserving
Boolean homomorphism from C to B.
proof (a) Let be the family of all functions extending and having the same properties; that is, is
a function from a subset of A to B, and (inf in ai ) =
(inf in ai ) for all a0 , . . . , an dom . By Zorns
Lemma, has a maximal member . Write D for the domain of .
(b)(i) If c, d D then c d D. P
P?? Otherwise, set D = D {c d} and extend to : D B by
(iii) Since
1 =
1 = 1,
1 = 1. If d D then
(1 \ d) =
(1 \ d) =
1
d = 1 d = (1 \ d),
while
(d (1 \ d)) =
(d (1 \ d)) = 0,
so d (1 \ d)) = 0, (1 \ d) 1 \ d and (1 \ d) must be equal to 1 \ d.
By 312H(ii), : 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
D, is a totally nite measure algebra in which D is a topologically dense subalgebra. By 324O, there
is an extension of to a measure-preserving Boolean homomorphism from D to B; of course this extension
belongs to , so in fact D = D is a closed subalgebra of A.
(c) Since A D, C D and 1 = C is a suitable extension of .
To see that 1 is unique, let 2 : C B be any other measure-preserving Boolean homomorphism
extending . Set C = {a : 1 a = 2 a}; then C is a topologically closed subalgebra of A including A, so is
the whole of C, and 2 = 1 .
324X Basic exercises (a) Let A and B be Boolean algebras, of which A is Dedekind -complete, and
: A B a sequentially order-continuous Boolean homomorphism. Let I be an ideal of A included in the
kernel of . Show that we have a sequentially order-continuous Boolean homomorphism : A/I B given
by setting (a ) = a for every a A.
(b) Let (A,
) be a measure algebra, and B a -subalgebra of A. Show that provided that (B,
B) is
semi-finite, then the topology of B induced by
B is just the subspace topology induced by the topology
of A. (Hint: apply 324Fc to the embedding B A.)
(c) Let (X, , ) be a measure space and (X, ,
) its c.l.d. version. Let A, A2 be the corresponding
measure algebras and : A A2 the canonical homomorphism (see 322Db). Show that is topologically
continuous.
324 Notes
Homomorphisms
95
96
Measure algebras
325 intro.
(c c1 )) = (W (U S
(e
in Ei Fi )) .
(e) Now suppose that and are semi-nite. Then (E F ) = E F for any E , F T (251J),
so (a
b) =
a b for every a A, b B.
325C
97
To see that 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
(a
c
b) =
a b > 0
and c 6= 0.
325B Characterizing the measure algebra of a product space A very natural question to ask is,
whether it is possible to dene 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 diculties involved.
Example There are complete locally determined localizable measure spaces (X, ), (X , ), with isomorphic
measure algebras, and a probability space (Y, ) such that the measure algebras of the c.l.d. product measures
on X Y , X Y are not isomorphic.
proof Let (X, , ) be the complete locally determined localizable not-strictly-localizable measure space
described in 216E. Recall that, for E , E = #({ : C, f E}) if this is nite, otherwise
(216Eb), where C is a set of cardinal greater than c. The map E 7 { : f E} : PC is surjective
(216Ec), so descends to an isomorphism between A, the measure algebra of , and PC. Let (X , , ) be
C with counting measure, so that its measure algebra (A ,
) is isomorphic to (A,
), while is of course
strictly localizable.
Let (Y, T, ) be {0, 1}C with its usual measure. Let , be the c.l.d. product measures on X Y , X Y
(C ,
) the corresponding measure algebras. Then is not localizable (254U), so
respectively, and (C, ),
is not localizable (322Be). On the other hand, , being the c.l.d. product of strictly localizable
(C, )
) is localizable, and is not isomorphic
measures, is strictly localizable (251N), therefore localizable, so (C ,
to (C, ).
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--nite 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 , 1 , 1 ) and (X2 , 2 , 2 ) be semi-nite measure spaces, with measure algebras (A1 ,
1 )
the corresponding measure
and (A2 ,
2 ). Let be the c.l.d. product measure on X1 X2 , and (C, )
algebra. Let (B, ) be a localizable measure algebra, and 1 : A1 B, 2 : A2 B order-continuous
Boolean homomorphisms such that (1 (a1 ) 2 (a2 )) =
1 a1
2 a2 for all a1 A1 , a2 A2 . Then there is a
unique order-continuous measure-preserving Boolean homomorphism : C B such that ((a1 a2 )) =
1 (a1 ) 2 (a2 ) for all a1 A1 , a2 A2 , writing : A1 A2 C for the canonical map described in 325A.
proof (a) Because 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 1 , 2 correspond to a Boolean homomorphism : A1 A2 B. The point is
for every c A B. P
that c = c
P By 315Jb, every member of A1 A2 is expressible as supin ai ai ,
where ai A1 , ai A2 and hai ai iin is disjoint. Now for each i we have
i a ),
(ai a ) = (1 (ai ) 2 (a )) =
1 ai
2 a = (a
i
by 325Ad. So
(c) =
Pn
i=0
(ai ai ) =
Pn
i=0
i a ) = c.
Q
(a
Q
i
(b) The following fact will underlie many of the arguments below. If e B, e < and > 0, there
are e1 Af1 , e2 Af2 such that (e \ (e1 e2 )) , writing Afi = {a :
i a < }. P
P Because (A1 ,
1 )
f
f
is semi-nite, A1 has supremum 1 in A1 ; because 1 is order-continuous, sup{1 (a) : a A1 } = 1 in
B, and inf{e \ 1 (a) : a Af1 } = 0 (313Aa). Because Af1 is upwards-directed, {e \ 1 (a) : a Af1 } is
downwards-directed, so inf{
(e \ (a)) : a Af1 } = 0 (321F). Let e1 Af1 be such that (e \ 1 (e1 )) 21 .
In the same way, there is an e2 Af2 such that (e \ 2 (e2 )) 12 . Consider e = e1 e2 C. Then
(e \ e ) = (e \ (1 (e1 ) 2 (e2 ))) (e \ 1 (e1 )) + (e \ 2 (e2 )) . Q
Q
98
325C
Measure algebras
P
(c) The next step is to check that is uniformly continuous for the uniformities dened by , .
P Take
1 e2 ) < and (e \ (e1 e2 )) 1 . Set
any e Bf and > 0. By (b), there are e1 , e2 such that (e
2
c ) e ) 1 . Then
e = e1 e2 . Now suppose that c, c A1 A2 and ((c
2
1
2
c ) e ) + .
(((c) (c )) e) ((c c ) e ) + (e \ e ) ((c
By 3A4Cc, 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 : C B extending (3A4G).
(e) Because is a Boolean homomorphism, so is . P
P (i) The functions c 7 (1 \ c), c 7 1 \ (c) are
continuous and the topology of B is Hausdor, so {c : (1 \ c) = 1 \ (c)} is closed; as it includes A1 A2 ,
it must be the whole of C. (ii) The functions (c, c ) 7 (c c ), (c, c ) 7 (c) (c ) are continuous, so
{(c, c ) : (c c ) = (c) (c )} is closed in C C; as it includes (A1 A2 ) (A1 A2 ), it must be the whole
of C C. Q
Q
(f ) Because is measure-preserving, so is . P
P Take any e1 Af1 , e2 Af2 . Then the functions
(e1 e2 )), c 7 (c (e1 e2 )) are continuous and equal on A1 A2 , so are equal on C. The
c 7 (c
argument of (b) shows that for any b B,
b = sup{
(b e) : e Bf }
= sup{
(b (e1 e2 )) : e1 Af1 , e2 Af2 },
so that
(c) = sup{
(c (e1 e2 )) : e1 Af1 , e2 Af2 }
(e1 e2 )) : e1 Af , e2 Af } = c
= sup{(c
1
for every c C. Q
Q
(g) To see that is order-continuous, take any non-empty downwards-directed set C C with inmum
0. ?? If [C] has a non-zero lower bound b in B, let e b be such that 0 < e < . Let e C be such that
< and (e \ (e )) < e, as in (b) above, so that (e (e )) > 0. Now, because inf C = 0, there is a
e
e ) < (e (e )). But this means that
c C such that (c
e ) < (e (e )) (b (e )),
(b (e )) (c e ) = (c
which is absurd. X
X Thus inf [C] = 0 in B. As C is arbitrary, is order-continuous.
(h) Finally, to see that is unique, observe that any order-continuous Boolean homomorphism from C
to B must be continuous (324Fc); so that if it agrees with on A1 A2 it must agree with on C.
325D Theorem Let (A1 ,
1 ) and (A2 ,
2 ) be semi-nite measure algebras.
together with order-continuous Boolean homomorphisms
(a) There is a localizable measure algebra (C, ),
1 : A1 C, 2 : A2 C such that whenever (B, ) is a localizable measure algebra, and 1 : A1 B,
2 : A2 B are order-continuous Boolean homomorphisms and (1 (a1 ) 2 (a2 )) =
1 a1
2 a2 for all
a1 A1 , a2 A2 , then there is a unique order-continuous measure-preserving Boolean homomorphism
: C B such that j = j for both j.
1 , 2 ) is determined up to isomorphism by this property.
(b) The structure (C, ,
(c)(i) The Boolean homomorphism : A1 A2 C dened from 1 and 2 is injective, and [A1 A2 ]
is topologically dense in C.
(ii) The order-closed subalgebra of C generated by [A1 A2 ] is the whole of C.
325F
99
where b
a is the open-and-closed subset of Z1 corresponding to a. Similarly, we may think of (A2 ,
2 ) as the
measure algebra of (Z2 , 2 , 2 ), where Z2 is the Stone space of A2 .
(ii) Let be the c.l.d. product measure on Z1 Z2 . The point is that is strictly localizable. P
P By
322E, both A1 and A2 have partitions of unity consisting of elements of nite measure; let hci iiI , hdj ijJ be
such partitions. Then hb
ci dbj iiI,jJ is a disjoint family of sets of nite measure in Z1 Z2 . If W Z1 Z2
is such that W > 0, there must be sets E1 , E2 of nite measure such that (W (E1 E2 )) > 0. Because
E1 = supiI E1 ci , we must have
P
P
1 E1 =
1 E1 = iI
1 (E1 ci ) = iI 1 (E1 b
ci ).
P
Similarly, 2 E2 = iJ 2 (E2 dbj ). But this means that there must be nite I I, J J such that
P
ci )2 (E2 dbj ) > 1 E1 2 E2 (W (E1 E2 )),
iI ,jJ 1 (E1 b
so that there have to be i I , j J such that (W (b
ci dbj )) > 0.
b
Now this means that hb
ci dj iiI,jJ satises the conditions of 213O. Because is surely complete and
locally determined, it is strictly localizable. Q
Q
100
325F
Measure algebras
b
<
1,
so
nI i P
nI En < , by
P
the Borel-Cantelli lemma (273K). Similarly nJ En < . Because nN En = , there must be some
n N \ (I J). Now an b and an b are both non-zero, so
which is absurd. X
XQ
Q
(c) On the other hand,
P
P
P
an )2 = n=0 2n < 1,
n=0 (cn )
n=0 (
(d) By 313P(a-ii) and 313O, [A A] cannot be order-dense in C; alternatively, (b) shows that there can
be no non-zero member of [A A] included in 1 \ supnN (cn ). (Both these arguments rely tacitly on the
fact that is injective, as noted in 325Ae.)
325G Since 325F shows that the free product and the localizable measure algebra free product are very
dierent constructions, I had better repeat an idea from 315 in the new context.
the measure
Example Again, let (A,
) be the measure algebra of Lebesgue measure on [0, 1], and (C, )
algebra of Lebesgue measure on [0, 1]2 . Then there is no order-continuous Boolean homomorphism : C A
such that (a b) = a b for all a, b A. P
P Let : C A be a Boolean homomorphism such that
(ab) = a b for all a, b A. For i < 2n let ani be the equivalence class in A of the interval [2n i, 2n (i+1)],
n = 2n for each n, so inf nN cn = 0 in C;
and set cn = supi<2n ani ani . Then cn = 1 for every n, but c
thus 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 nite number of factors. No new ideas are needed, so I spell the results out without
proofs.
(a) Let h(Ai ,
i )iiI be a nite family of semi-nite measure algebras. Then there is a localizable mea together with order-continuous Boolean homomorphisms i : Ai C for i I, such
sure algebra (C, ),
that whenever (B, ) is a localizable measure
Q algebra, and i : Ai B are order-continuous Boolean homomorphisms such that (inf iI i (ai )) = iI
i ai whenever ai Ai for each i, then there is a unique
order-continuous measure-preserving Boolean homomorphism : C B such that i = i for every i.
hi iiI ) is determined up to isomorphism by this property.
(b) The structure (C, ,
N
N
(c) The Boolean homomorphism :
iI Ai ] is
iI Ai C dened from the i is injective, and [
topologically dense in C.
Nloc
i ) for (a particular version of) the localizable measure algebra free product described
(d) Write c iI (Ai ,
in (a). If h(Ai ,
i )iiI is a nite family of semi-nite measure algebras and hI(k)ikK is a partition of I,
loc
Nloc
Nloc N
c
then c (Ai ,
i ) is isomorphic, in a canonical way, to c
(Ai ,
i ) .
iI
kK
iI(k)
325I
101
Theorem Let h(Xi , i , i )iiI beQany family of probability spaces, with measure algebras (Ai ,
i ). Let
the corresponding measure algebra. For each i I,
be the product measure on X = iI Xi , and (C, )
we have a measure-preserving homomorphism i : Ai C corresponding to the inverse-measure-preserving
function x 7 x(i) : X Xi .Q
Let (B, ) be a probability algebra, and i : Ai B Boolean homomorphisms
i ai whenever J I is a nite set and ai Ai for every i. Then there is
such that (inf iJ i (ai )) = iJ
a unique measure-preserving Boolean homomorphism : C B such that i = i 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 i : 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},
Set
C = {E : E C} C,
so that C is precisely the family of elements of C expressible in the form inf iJ i (ai ) where J I is nite
and ai Ai for each i.
N
The homomorphisms
iI Ai C (315I), which is
N i : Ai C dene a Boolean homomorphism :
injective. P
P If c iI Ai is non-zero, there must be a nite set J I and a family hai iiJ such that
ai Ai \ {0} for each i and c inf iJ i (ai ) (315Jb). Express each ai as Ei , where Ei i . Then
E = {x : x X, x(i) Ei for each i J}
has measure
E =
while
iJ
Ei =
iJ
E = (inf iJ i (ai ))
ai 6= 0,
(c),
N
(b) Because is injective, it is an isomorphismNbetween iI Ai and its image in C. I trust it will cause
no confusion if I abuse notation slightly and treat iI
NAi as actually a subalgebra of C, so that j : Aj C
becomes identied with the canonical map j : Aj iI Ai . Now the Boolean homomorphisms i : Ai
N
for every
B correspond
to a Boolean homomorphism :
(c) = c
iI Ai B. The point is that
N
iJ
Ei =
i Ei = (inf ai )
iJ
iJ
iJ
Next, any c C is expressible as the supremum of a nite disjoint family hck ikK in C (315Jb), so
P
P
k ) = c.
Q
Q
(c) = kK (ck ) = kK (c
since
(c) It follows that is uniformly continuous for the metrics dened by , ,
c )
((c) (c )) = (c c ) = (c
N
for all c, c iI Ai .
N
P Let c C, > 0. Express c as W . Then by 254Fe
(d) Next,
iI Ai is topologically dense
Sin C. P
there are H0 , . . . , Hk C such that (W jk Hj ) . Now cj = Hj C for each j, so
S
N
c = supjk cj = ( jk Hj ) iI Ai ,
102
Measure algebras
325I
c ) . Q
and (c
Q
Since B is complete for its uniformity (323Gc), there is a uniformly continuous function : C B
extending (3A4G).
(e) Because is a Boolean homomorphism, so is . P
P (i) The functions c 7 (1 \ c), 1 \ (c) are
continuous and the topology of B is Hausdor, so {c : (1 \ c) = 1 \ (c)} is closed; as it includes A1 A2 ,
it must be the whole of C. (ii) The functions (c, c ) 7 (c c ), (c, c ) 7 (c) (c ) are continuous, so
{(c, c ) : (c c ) = (c) (c )} is closed in C C; as it includes (A1 A2 ) (A1 A2 ), it must be the whole
of C C. Q
Q
I
such
that
c
C
,
and
this
J
is
countable.
J
T
(c) For any non-empty family J PI, JJ CJ = CT J .
proof (a) If (B, , hi iiJ ) is any probability algebra free product of h(Ai ,
i )iiJ , then we have a measurepreserving homomorphism
:
B
C
such
that
for
every
i
J.
Because the subalgebra
i
i
S B0 of B
S
generated by iJ i [Ai ] is topologically dense in B (325Jc), and is continuous (324Kb), iJ i [Ai ] is
topologically dense in S
[B]; also [B] is closed in C (324Kb again). But this means that [B] is just the
topological closure of iI i [Ai ] and must be CJ . Thus is an isomorphism, and
CJ , hi iiJ ) = ([B], 1 , hi iiI )
(CJ ,
325Y
103
where i J, E
S i . Of course TJ J , so {W : W J } = {W : W TJ } is the closed subalgebra of
C generated by iK i [Ai ], which is CJ . Thus K is also the unique smallest subset of I such that c CK .
T
T
(c) Of T
course CK CJ whenever K J I, so
other hand, suppose
T JJ CJ C J . On the
T . As c is arbitrary,
C
;
then
by
(b)
there
is
some
K
J
such
that
c
C
that
c
J
K
J
JJ
T
T
JJ 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 ,
1 ), (A2 ,
2 ) be two semi-nite measure algebras, and suppose that
for each j we are given a closed subalgebra Bj of Aj such that (Bj , j ) is also semi-nite, where j =
j Bj .
Show that the localizable measure algebra free product of (B1 , 1 ) and (B2 , 2 ) can be thought of as a closed
1 ) and (A2 ,
2 ).
subalgebra of the localizable measure algebra free product of (A1 ,
(b) Let (A1 ,
1 ), (A2 ,
2 ) be two semi-nite measure algebras, and suppose that for each j we are given
a principal ideal Bj of Aj . Set j =
j Bj . Show that the localizable measure algebra free product of
(B1 , 1 ) and (B2 , 2 ) can be thought of as a principal ideal of the localizable measure algebra free product
of (A1 ,
1 ) and (A2 ,
2 ).
> (c) Let h(Ai ,
i )iiI and h(Bj , j )ijJ be families of semi-nite measure algebras, with simple products
b loc (B, ) can be
(A,
) and (B, ) (322K). Show that the localizable measure algebra free product (A,
)
b
identied with the simple product of the family h(Ai ,
i )loc (Bj , j )iiI,jJ .
104
Measure algebras
325Yb
(b) Let A be a Boolean algebra, and : A [0, ] a function such that 0 = 0 and (a b) = a + b
whenever a, b A and a b = 0; suppose that Af = {a : a < } is order-dense in A. For e Af , a, b A
set e (a, b) = (e (a b)). Give A the uniformity dened by {e : e < }. (i) Show that the completion
b of A under this uniformity has a measure
A
, extending , under which it is a localizable measure algebra.
b
(ii) Show that if a A,
a < and > 0, there is a b A such that
(a b) . (iii) Show that for every
b there is a sequence han inN in A such that a sup
aA
inf
a =
(supnN inf mn am ).
mn am and
nN
b
b
(iv) In particular, the set of inma 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 = nN 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,
) be the measure algebra of Lebesgue measure on [0, 1]. Show that A A is ccc but not
weakly (, )-distributive. (Hint: (i) A A is embeddable as a subalgebra of a probability algebra (ii) in
the notation of 325F, look at cmn = supmin ei ei .)
(e) Repeat 325F-325G and 325Yc-325Yd with an arbitrary atomless probability space in place of [0, 1].
(f ) Let (A,
) be a probability algebra and hai iiI a (stochastically) independent family in A. Show that
for any a A, > 0 the set {i : i I, |
(a ai )
a
ai | } is nite, so that {i :
(a ai ) 6=
a
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 rst is that every semi-nite 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 nite measure algebras, using 325Xc and the fact that for -nite
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).
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 nitely-additive measures (325Yb(ii),
325Yb(iii)). It is worth noting that the completion process can be regarded as made up of two steps; rst
take inma of sequences of sets of nite measure, and then take arbitrary suprema (325Yb(iv)).
The idea of 325F appears
S in many guises, and this is only the rst time that I shall wish to call on it.
The point of the set W = nN 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 rst 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 innite free products of probability algebras, there is a degree of simplication, since there
is only one algebra which can plausibly be called the probability algebra free product, and this is produced
326C
105
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 oers 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.
(because c = b + (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 nitely additive real-valued functionals on A is a linear space (a linear subspace of R A ). We
give it the ordering induced by that of R A , so that i a a for every a A. This renders it a
partially ordered linear space (because R A is).
106
326D
Measure algebras
326D The Jordan decomposition (I): Proposition Let A be a Boolean algebra, and a nitely
additive real-valued functional on A. Then the following are equiveridical:
(i) is bounded;
(ii) supnN |an | < for every disjoint sequence han inN in A;
(iii) P
limn |an | = 0 for every disjoint sequence han inN in A;
(iv) n=0 |an | < for every disjoint sequence han inN in A;
(v) is expressible as the dierence of two non-negative additive functionals.
proof (a)(i)(v) Assume that is bounded. For each a A, set
+ a = sup{b : b
Because is bounded,
is real-valued. Now
+ (a b) = sup c =
c ab
(because d e a b = 0 whenever d
a, e
a}.
is additive. P
P If a, b A and a b = 0, then
sup
d a,e b
(d e) =
sup
d + e
d a,e b
b)
= sup d + sup e = + a + + b. Q
Q
da
Consequently
Since
eb
i=0
|ai |
i=0
1 ai +
i=0
2 ai 1 1 + 2 1 < .
(d) not-(i)not-(ii) Suppose that is unbounded. Choose sequences han inN , hbn inN inductively, as
follows. b0 = 1. Given that supa bn |a| = , choose cn bn such that |cn | |bn | + n; then |cn | n
and
We have
a bn
sup
a bn cn
|a| +
sup
a bn \cn
|a|,
so at least one of supa bn cn |a|, supa bn \cn |a| must be innite; take bn+1 to be one of cn , bn \ cn such
that supa bn+1 |a| = , and set an = bn \ bn+1 , so that |an | n. Continue.
On completing the induction, we have a disjoint sequence han inN such that |an | 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
dierence of monotonic functions (224D).
326E Countably additive functionals: Definition
Let A be a Boolean algebra. A functional :
P
A R is countably additive or -additive if n=0 an is dened and equal to (supnN an ) whenever
han inN is a disjoint sequence in A and supnN an is dened in A.
326G
107
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 -algebra 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 : B R such that 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 (437).
326F Elementary facts Let A be a Boolean algebra and : A R a countably additive functional.
(a) is nitely additive. (Setting an = 0 for every n, we see from the denition in 326E that 0 = 0.
Now, given a b = 0, set a0 = a, a1 = b, an = 0 for n 2 to see that (a b) = a + b.)
(b) If han inN is a non-decreasing sequence in A with a supremum a A, then
P
a = a0 + n=0 (an+1 \ an ) = limn an .
(c) If han inN is a non-increasing sequence in A with an inmum a A, then
is countably additive.
(c) If is non-negative, then is countably additive i it is sequentially order-continuous.
proof (a)(i) If is countably additive and han inN is a non-increasing sequence in A with inmum 0, then
limn an = 0 by 326Fc. (ii) If satises the condition, and han inN is a disjoint sequence in A with
supremum a, set bn = a \ supin ai for each n N; then hbn inN is non-increasing and has inmum 0, so
Pn
a i=0 ai = a (supin ai ) = bn 0
P
as n , and a = n=0 an ; thus is countably additive.
(b) If han inN is a disjoint sequence in A with supremum a, set bn = supin ai for each n; then a =
limn bn , so
and
n=0
an = limn bn = a.
108
Measure algebras
326H
326H The Jordan decomposition (II): Proposition Let A be a Boolean algebra and a bounded
countably additive real-valued functional on A. Then is expressible as the dierence of two non-negative
countably additive functionals.
proof Consider the functional + a = supb a b dened in the proof of 326D. If han inN is a disjoint
sequence in A with supremum a, and b a, then
P
P
b = n=0 (b an ) n=0 + an .
P
As b is arbitrary, + a n=0 + an . But of course
Pn
+ a + (supin ai ) = i=0 + ai
P
for every n N, so + a = n=0 + an . As han inN is arbitrary, + is countably additive.
Now = + is also countably additive, and = + is the dierence of non-negative countably
additive functionals.
326I The Hahn decomposition: Theorem Let A be a Dedekind -complete Boolean algebra and
: A R a countably additive functional. Then is bounded and there is a c A such that a 0
whenever a c, while a 0 whenever a c = 0.
first proof By 314M, there are a set X and a -algebra of subsets of X and a sequentially order-continuous
Boolean homomorphism from onto A. Set 1 = : R. Then 1 is countably additive (326Ff).
So 1 is bounded and there is a set H such that 1 F 0 whenever F and F H and 1 F 0
whenever F and F H = (231Eb). Set c = H A. If a c, then there is an F such that
F = a; now (F H) = a c = a, so a = 1 (F H) 0. If a c = 0, then there is an F such that
F = a; now (F \ H) = a \ c = a, so a = 1 (F \ H) 0.
P
second proof (a) Note rst that is bounded. P
P If han inN is a disjoint sequence in A, then n=0 an
must exist and be equal to (supnN an ); in particular, limn an = 0. By 326D, is bounded. Q
Q
(b)(i) We know that = sup{a : a A} < . Choose a sequence han inN in A such that an 2n
for every n N. For m n N, set bmn = inf min ai . Then bmn 2 2m + 2n 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 (an+1 bmn ), so
+ bm,n+1 (an+1 bmn ) + (an+1 bmn )
= 2 2 2m + 2n1 .
Subtracting from both sides, bm,n+1 2 2m + 2n1 and the induction proceeds. Q
Q
(ii) Set
c b = (c \ b) = c,
c + b = (c b) = c
326L
109
B = {b : |b| , a A, b a}.
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 inmum 0; but this is impossible, since
inf bB |b| . X
XQ
Q
(c) If 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 inmum 0, so
supaA a = a0 inf aA (a0 \ a) = a0 .
(e) is a completely additive functional for any R. If is another completely additive functional
on A, then + is completely additive. P
P We know from 326Bc that + is additive. Let A be a
non-empty downwards-directed set with inmum 0. For any > 0, (b) tells us that there are a, a A such
that |b| whenever b a and | b| whenever b a . But now, because A is downwards-directed,
there is a b A such that b a a , which means that |b + b| |b| + | b| is at most 2. As is arbitrary,
inf aA |( + )(a)| = 0, and + is completely additive. Q
Q
(f ) If B is another Boolean algebra and : B A is an order-continuous Boolean homomorphism,
then is a completely additive functional on B. P
P By 326Be, is additive. If B B is a non-empty
downwards-directed set with inmum 0 in B, then [B] is a non-empty downwards-directed set with inmum
0 in A, because is order-continuous, so inf bB |b| = 0. Q
Q In particular, if B is a regularly embedded
subalgebra of A, then B is completely additive.
(g) If is another additive functional on A and | a| a for every a A, then is completely additive.
P
P If A A is non-empty and downwards-directed and inf A = 0, then inf aA | a| inf aA a = 0. Q
Q
326L I squeeze a useful fact in here.
Proposition If A is a ccc Boolean algebra, a functional : A R is countably additive i it is completely
additive.
proof If is completely additive it is countably additive, by 326Ka. If is countably additive and A is
a non-empty downwards-directed set in A with inmum 0, then there is a (non-empty) countable subset B
of A also with inmum 0 (316E). Let hbn inN be a sequence running over B, and choose han inN in A such
that a0 = b0 , an+1 an bn for every n N. Then han inN is a non-increasing sequence with inmum 0,
so limn an = 0 (326Fc) and inf aA |a| = 0. As A is arbitrary, is completely additive.
110
326M
Measure algebras
326M The Jordan decomposition (III): Proposition Let A be a Boolean algebra and a completely
additive real-valued functional on A. Then is bounded and expressible as the dierence of two non-negative
completely additive functionals.
proof (a) I must rst check that is bounded. P
P Let han inN be a disjoint sequence in A. Set
A = {a : a A, there is an n N such that ai
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 |b| 1 whenever b a. By the denition of A, there must be an n N such that
|ai | 1 for every i n. But this means that supnN |an | is nite. As han inN is arbitrary, is bounded,
by 326D(ii). Q
Q
(b) As in 326D and 326H, set + a = supb a b for every a A. Then + is completely additive. P
P
+
We know that is additive. If A is a non-empty downwards-directed subset of A with inmum 0, then for
every > 0 there is an a A such that |b| whenever b a; in particular, + a . As is arbitrary,
inf aA + a = 0; as A is arbitrary, + is completely additive. Q
Q
Consequently = + is completely additive (326Ke) and = + is the dierence of non-negative
completely additive functionals.
326N I give an alternative denition of completely additive which you may feel claries the concept.
Proposition Let A be a Boolean algebra, and : A R a function. Then the following are equiveridical:
(i) is completely
additive;
P
(ii) 1 = PiI ai whenever hai iiI is a partition of unity in A;
(iii) a = iI ai whenever hai iiI is a disjoint family in A with supremum a.
P
proof (For notes on sums iI , see 226A.)
(a)(i)(ii) P
If is completely additive and hai iiI is a partition of unity in A, then (inducing on #(J))
(supiJ ai ) = iJ ai for every nite J I. Consider
A = {1 \ supiJ ai : J I is nite}.
Then A is non-empty and downwards-directed and has inmum 0, so for every > 0 there is an a A such
that |b| whenever b a. Express a as 1 \ supiJ ai where J I is nite. If now K is another nite
subset of I including J,
P
|1 iK ai | = |(1 \ supiK ai )| .
P
As remarked in 226Ad, this means that 1 = iI ai , as claimed.
(b)(ii)(iii) Suppose that satises the condition (ii), and that hai iiI is a disjoint family with supremum a. Take any j
/ I, set J = I {j} and aj = 1 \ a; then hai iiJ , (a, 1 \ a) are both partitions of unity,
so
P
P
(1 \ a) + a = 1 = iJ ai = (1 \ a) + iI ai ,
P
and a = iI ai .
(c)(iii)(i) Suppose that satises (iii). Then is additive.
) is bounded. P
(
P Let han inN be a disjoint sequence in A. Applying Zorns Lemma to the set C
of
all
disjoint
families
C
A including {an : n N}, we nd a partition of unity C {an : n N}. Now
P
c
is
dened
in
R,
so supnN |an | supcC |c| is nite. By 326D, is bounded. Q
Q
cC
326P
X
iI
111
X
+ ai = sup{
+ ai : J I is nite}
iJ
+
so + a =
iI
+ ai . Q
Q
whenever hai iiI is a disjoint family in A with supremum a. Because is non-negative, the argument of
() shows that = ( )+ is completely additive. So = + is completely additive, as required.
326O For completely additive functionals, we have a useful renement of the Hahn decomposition. I
give it in a form adapted to the applications I have in mind.
Proposition Let A be a Dedekind -complete Boolean algebra and : A R a completely additive
functional. Then there is a unique element of A, which I will denote [[ > 0]], the region where > 0, such
that a > 0 whenever 0 6= a [[ > 0]], while a 0 whenever a [[ > 0]] = 0.
proof Set
C1 = {c : c A \ {0}, a > 0 whenever 0 6= a c},
C2 = {c : c A, a 0 whenever a c}.
Then C1 C2 is order-dense in A. P
P There is a c0 A such that a 0 for every a c0 , a 0 whenever
a c0 = 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 c 0; but in this case a 0, (c \ a) 0 so
a = 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 , d > 0.
Also
#({d : d D, d 2n }) 2n supaA a
is nite for each n, so D C1 is the union of a sequence of nite sets, and is countable. Q
Q
Accordingly D C1 has a supremum e. If 0 6= a e then
P
P
a = cD (a c) = cDC1 (a c) 0
by 326N. Also there must be some c D C1 such that a c 6= 0, in which case (a c) > 0, so that a > 0.
If a e = 0, then
P
P
a = cD (a c) = cDC2 (a c) 0.
Thus e has the properties demanded of [[ > 0]]. To see that e is unique, we need observe only that if
e has the same properties then (e \ e ) 0 (because (e \ e ) e = 0), so e \ e = 0 (because e \ e e).
Similarly, e \ e = 0 and e = e . Thus we may properly denote e by the formula [[ > 0]].
326P Corollary Let A be a Dedekind -complete Boolean algebra and , two completely additive
functionals on A. Then there is a unique element of A, which I will denote [[ > ]], the region where > ,
such that
112
Measure algebras
326P
c = (c i (a)) + (c i (1 \ a))
suprm cr = suprm cr ,
r=0 cr ,
r=0 cr =
and whenever r m , j J and b Bj then either cr j (b) or cr j (b) = 0; that is, every cr is either 0
or of the form inf jJ j (bj ) where bj is an atom of Bj for every j. Similarly, we can nd hds isn such that
Pn
Pn
supsn ds = supsn ds ,
s=0 ds =
s=0 ds ,
and whenever s n , j J and b Bj then ds is either 0 or of the form inf jJ j (bj ) where bj is an atom
of Bj for every j. But we now have suprm cr = supsn ds while for any r m , s n either cr = ds or
cr ds = 0. It follows that the non-zero terms in the nite sequence hcr irm are just a rearrangement of
the non-zero terms in hds isn , so that
Pn
Pn
Pm
Pm
s=0 ds =
s=0 ds ,
r=0 cr =
r=0 cr =
as required. Q
Q
Pm
(c) By 315Jb, this means that we have a functional : A R such that (suprm cr ) =
r=0 cr
whenever hcr irm is a disjoint family in C. It is now elementary to check that is additive, and it is clearly
the only additive functional on A extending .
326X Basic exercises (a) Let A be a Boolean algebra and : A R a nitely additive functional. Show
that (i) (a b) = a + b (a b) (ii) (a b c) = a + b + c (a b) (a c) (b c) + (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 : A R a nitely additive functional. Show that the following
are equiveridical: (i) is countably additive; (ii) limn an = a whenever han inN is a non-decreasing
sequence in A with supremum a.
326Ye
113
(c) Let A be a Dedekind -complete Boolean algebra and : A R a nitely additive functional. Show
that the following are equiveridical: (i) is countably additive; (ii) limn an = 0 whenever han inN is a
sequence in A and inf nN supmn am = 0; (iii) limn an = a whenever han inN is a sequence in A and
a = inf nN supmn am = supnN inf mn am . (Hint: for (i)(iii), consider non-negative rst.)
(d) Let X be any uncountable set, and J an innite subset of X. Let A be the nite-conite algebra of
X (316Yk), and for a A set a = #(a J) if a is nite, #(J \ a) if a is conite. Show that 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 : A R such that [, ] = whenever 0 1. Show
that is countably additive but not completely additive.
(f ) (i) Let (X, , ) be any atomless probability space. Show that : R is a countably additive
functional which is not completely additive. (ii) Let X be any uncountable set and the countablecocountable measure on X (211R). Show that is countably additive but not completely additive.
P
(g) Let A be a Boolean algebra and : A R a function. (i) Show that is nitely additive
P i iI ai =
1 for every nite partition of unity hai iiI . (ii) Show that is countably additive i iI ai = 1 for
every countable partition of unity hai iiI .
(h) Show that 326O can fail if is only countably additive, rather than completely additive. (Hint:
326Xf.)
(i) Let A be a Boolean algebra and a nitely additive real-valued functional on A. Let us say that
a A is a support of if () b = 0 whenever b a = 0 () for every non-zero b a there is a c b such
that c 6= 0. (i) Check that can have at most one support. (ii) Show that if a is a support for and
is bounded, then the principal ideal Aa generated by a is ccc. (iii) Show that if A is Dedekind -complete
and is countably additive, then is completely additive i it has a support, and that in the language of
326O this is [[ > 0]] [[ > 0]]. (iv) Taking J = X in 326Xd, show that X is the support of the functional
there.
326Y Further exercises (a) Let A be a Boolean algebra and a non-negative nitely additive functional
on A. Show that the following are equiveridical: (i) for every > 0 there is a nite partition hai iiI of unity
in A such that ai for every i I; (ii) whenever is a non-zero nitely additive functional such that
0 there is an a A such that a and (1 \ a) are both non-zero. (Such functionals are called
atomless.)
(b) Let A be a Boolean algebra and 1 , 2 atomless non-negative additive functionals on A. Show that
1 + 2 , 1 are atomless for every 0, and that is atomless whenever is additive and 0 1 .
(c) Let A be a Dedekind -complete Boolean algebra and an atomless non-negative nitely additive
functional on A. Show that there is a family hat it[0,1] in A such that as at and at = t1 whenever
s t [0, 1].
(d) Let A be a Dedekind -complete Boolean algebra and 0 , . . . , n atomless non-negative nitely
additive functionals on A. Show that there is an a A such that i a = 12 i 1 for every i n. (Hint: it
is enough to consider the case 0 1 . . . n . For the inductive step, use the inductive hypothesis to
construct hat it[0,1] such that as at , i at = ti 1 if i < n, 0 s t 1. Now show that t 7 n (at+ 21 \ at )
is continuous on [0, 12 ].)
(e) Let A be a Dedekind -complete Boolean algebra. Let : A R r , where r 1, be additive (in the
sense that (a b) = a + b whenever a b = 0) and atomless (in the sense that for every > 0 there is a
nite partition of unity hai iiI such that kak whenever i I and a ai ). Show that {a : a A} is
a convex set in R r . (This is a version of Liapounoff s theorem. I am grateful to K.P.S.Bhaskara Rao for
showing it to me.)
114
Measure algebras
326Yf
(f ) Let A be a Dedekind -complete Boolean algebra and : A [0, [ a countably additive functional.
Show that is atomless i whenever a A and a 6= 0 there is a b a such that 0 < b < a.
(g) Show that there is a nitely additive functional : PN R such that {n} = 1 for every n N, so
that is not bounded. (Hint: Use Zorns Lemma to construct a maximal linearly independent subset of
including {{n} : n N}, and hence to construct a linear map f : R such that f ({n}) = 1 for every
n.)
(h) Let A be any innite Boolean algebra. Show that there is an unbounded nitely additive functional
: A R. (Hint: let htn inN be a sequence of distinct points in the Stone space of A, and set a = {n :
tn b
a} for a suitable .)
(i) Let A be a Boolean algebra, and give RA its product topology. Show that the space of nitely additive
functionals on A is a closed subset of R A , but that the space of bounded nitely additive functionals is closed
only when A is nite.
(j) Let A be a Boolean algebra, and M the linear space of all bounded nitely additive real-valued
functionals on A. For , M say that if a a for every a A. Show that
(i) + , as dened in the proof of 326D, is just sup{0, } in M ;
(ii) M is a Dedekind complete Riesz space (241E-241F, 353G);
(iii) for , M , || = (), , are given by the formulae
||(a) = supb a b (a \ b),
for every a A;
(v) setting kk = ||(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 : A C is finitely additive if its real and imaginary
parts are. Show that P
the space of bounded nitely additive functionals from A to C is a Banach space under
n
the norm kk = sup{ i=0 |ai | : hai iin is a partition of unity in A}.
(l) Let A be a Boolean algebra, and give it the topology T for which the closed sets are the sequentially
order-closed sets. Show that a nitely additive functional : A R is countably additive i it is continuous
for T .
(m) Let A be a Boolean algebra, and M the set of all bounded countably additive real-valued functionals
on A. Show that M is a closed and order-closed linear subspace of the normed space M of all additive
functionals on A (326Yj), and that || M whenever M .
(n) Let A be a Boolean algebra and a non-negative nitely additive functional on A. Set
a = inf{supnN an : han inN is a non-decreasing sequence with supremum a}
for every a A. Show that is countably additive, and is sup{ : is countably additive}.
(o) Let A be a Dedekind -complete Boolean algebra and hn inN a sequence of countably additive realvalued functionals on A such that a = limn n a is dened in R for every a A. Show that is countably
additive. (Hint: use arguments from part (a) of the proof of 247C to see that limn supkN |k an | = 0 for
every disjoint sequence han inN in A, and therefore that limn supkN |k an | = 0 whenever han inN is a
non-increasing sequence with inmum 0.)
326 Notes
115
(p) Let A be a Boolean algebra, and M the set of all completely additive real-valued functionals on A.
Show that M is a closed and order-closed linear subspace of the normed space M of all additive functionals,
and that || M whenever M .
(q) Let A be a Boolean algebra and a non-negative nitely additive functional on A. Set
b = inf{supaA a : A is a non-empty upwards-directed set with supremum b}
for every b A. Show that is completely additive, and is sup{ : 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 nitely additive functional : A R is completely additive i it is continuous
for T.
(s) Let X be a set, any -algebra of subsets of X, and :P
R a functional. ShowP
that is completely
additive i there are sequences hxn inN , hn inN such that n=0 |n | < and E = n=0 n E(xn ) for
every E .
(t) Let A and B be Boolean algebras and , nitely additive functionals on A, B respectively. Show
that there is a unique nitely additive functional : A B R such that (a b) = a b for all a A,
b B.
N
(u) Let hAi iiI be a family of Boolean algebras, with free product ( iI Ai , hi iiI ), and for each i I
let i be a nitely
N additive functional on Ai such that i 1Q= 1. Show that there is a unique nitely additive
functional : iI Ai R such that (inf iJ i (ai )) = iJ i ai whenever J I is non-empty and nite
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 simplications 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 denitions. Similarly, the denition of countably additive
functional for algebras which are not Dedekind -complete needs a moments attention to the phrase and
supnN an is dened 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 denitions (326A, 326E, 326J) are those which I nd 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 diculty is that we must consider functionals which are not order-preserving, so that the simple
denitions in 313H cannot be applied directly. It is fair to say that all the additive functionals we need
to understand are bounded, and therefore may be studied in terms of their positive and negative parts + ,
, which are order-preserving (326Bf); but many of the most important applications of these ideas depend
precisely on using facts about to deduce facts about + and .
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 ramications. What it amounts to is that we can discover much more than we might expect by looking
at disjoint 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 -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 -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
116
Measure algebras
326 Notes
a -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 -algebra. Indeed there are very few completely
additive functionals on -algebras of sets (326Ys). Of course these surprises can arise only when there
is a dierence 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 nd myself generally using the phrase countably additive in preference to completely additive in the
context of ccc algebras, where there is no dierence between them. This is an attempt at user-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 -algebras
of sets; so that the rst 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 -algebras
of sets to the context of Dedekind -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
-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 terms 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.
a .
327B
117
b inf nN
(supkn bk ) inf nN k=n 2k = 0,
so
b = 0 and b = 0. On the other hand, is expressible as a dierence + of non-negative countably
additive functionals (326H), each of which is sequentially order-continuous (326Gc), and
0 = limn ( + + )bn inf nN ( + + )bn inf nN |bn | ,
which is absurd. X
X
(c)(i)(ii) Suppose that is continuous. Then it is completely additive, by (a), therefore countably
additive. If a 6= 0, there must be an b of nite measure such that |d| < |a| whenever d b = , so that
|(a \ b)| < |a| and (a b) 6= 0. Thus the conditions are satised.
(ii)(iv) Now suppose that satises the two conditions in (ii). Because A is Dedekind -complete,
must be bounded (326I), therefore expressible as the dierence + of countably additive functionals.
Set 1 = + + . Set
= sup{1 b : b A,
b < },
and choose a sequence hbn inN of elements of A of nite measure such that limn 1 bn = ; set b =
supnN bn . If d A and d b = then d = 0. P
P If b A and
b < , then
|(d b)| 1 (d b) 1 (b \ bn ) = 1 (b bn ) 1 bn 1 bn
for any a A.
Now if b, c A and
((b c) bn ) then
|b c| |(b \ c)| + |(c \ b)|
|((b \ c) b )| + |((c \ b) b )| + 2 + + 2 = 4
because
((b \ c) bn ),
((c \ b) bn ) are both less than or equal to . As is arbitrary, is uniformly
continuous.
118
Measure algebras
327B
On the other hand, if is an additive functional on which is zero on negligible sets, then, for E, F ,
E = F = (E \ F ) = (F \ E) = 0
= (E \ F ) = (F \ E) = 0
= F = E (E \ F ) + (F \ E) = E,
so we have a function : A R dened by the given formula. If E, F and E F = 0, then
(E F ) = (E F ) = (E F )
= (E \ F ) + F = E + F
for each n. So
S
P
P
(supnN an ) = ( nN Fn ) = n=0 Fn = n=0 an .
327F
119
(d) The denition 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, , ) be a semi-nite measure space, with measure algebra (A,
). Let L1 be the space of
equivalence classes of real-valued integrable functions on X (242), and write M for the set of completely
additive real-valued functionals on A. Then there is an ordered linear space bijection between M and L1
dened by saying that M corresponds to u L1 if
a =
f whenever a = E in A and f = u in L1 .
u=
f whenever f = u
for every E (242Ac). By 232E, u is additive and truly continuous, and of course it is zero when is
zero, so corresponds to a completely additive functional u on A (327Ce).
(c) Clearly the maps u 7 u and 7 u are now the two halves of a one-to-one correspondence. To see
that it is linear, we need note only that
(
u + v )E = u E + v E =
u+
v=
u + v = u+v E
u
v for every E
E
E
Z
Z
g for every E
f
f a.e. g u v,
using 131Ha.
327E I slip in an elementary fact.
Proposition If (A,
) is a measure algebra, then the functional a 7 c a =
(a c) is completely additive
whenever c A and
c < .
proof c is additive because
is additive, and by 321F inf aA c a = 0 whenever A is non-empty, downwardsdirected and has inmum 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,
) be a totally nite measure algebra and C A a closed subalgebra. Write M (A), M (C)
for the spaces of countably additive real-valued functionals on A, C respectively.
(a) There is an operator R : M (C) M (A) dened by saying that, for every M (C), R is the
unique member of M (A) such that [[R >
]] = [[ >
C]] for every R.
(b)(i) R extends for every M (C).
(ii) R is linear and order-preserving.
120
327F
Measure algebras
(iii) R(
C) =
.
P
(iv) If h
c for every
PninN is a sequence of non-negative functionals in M (C) such that n=0 n c =
c C, then n=0 (Rn )(a) =
a for every a A.
Remarks When saying that C is closed, I mean, indierently, topologically closed or order-closed; see
323H-323I.
For the notation [[ >
]] see 326O-326P.
proof (a)(i) By 321J-321K, we may represent (A,
) as the measure algebra of a measure space (X, , );
write for the canonical map from to A. Write T for {E : E , E C}. Because C is a -subalgebra
of C and is a sequentially order-continuous Boolean homomorphism, T is a -subalgebra of .
(ii) For each M (C), : T R is countably additive
and zero on {F : F T, F = 0}, so we
R
can choose a T-measurable function f : X R such that F f d( T) = F for every F T. Of course
we can now think of f as a -integrable function (233B),
so we get a corresponding countably additive
R
functional R : A R dened by setting (R)(E) = E f for every E (327D). (In this context, of
course, countably additive functionals are completely additive, by 327Bf.)
For R, set H = {x : f (x) > } T. Then for any E ,
E H , E > 0 =
E H = =
f > E,
f E.
So [[R >
]] = c C. Of course we now have
c = (R)(c) >
c when c C, 0 6= c c ,
c
c when c C, c c = 0,
(R)(F ) =
for every F T, so R extends .
f d =
f = F = 1 F + 2 F =
f d( T) = F
f1 +
f2 =
f =
E
f + f2 =
E 1
f +
E 1
f1 + f2
f1 = 1 F 2 F =
f2
327Y
121
f = F = F =
n
c. Set gn = i=0 fi for each n; then 0 a.e. gn a.e. gn+1 a.e. 1 for every n, and
n=0 n c =
R
Pn
1.
limn gn = limn i=0 i 1 =
R
R
But this means that, setting g = limn gn , g a.e. 1 and g = 1, so that g =a.e. 1 and
R
P
n=0 (Ri )(E) = limn E gn = E
P
a for every a A.
for every E , so that n=0 (Ri )(a) =
and
327G Definition In the context of 327F, I will call R the standard extension of to A.
Remark The point of my insistence on the uniqueness of R, and on the formula in 327Fa, is that R really
is dened by the abstract structure (A,
, C, ), even though I have used a proof which runs through the
representation of (A,
) as the measure algebra of a measure space (X, , ).
327X Basic exercises (a) Let (A,
) and (B,
) be totally nite measure algebras, and : A B a
measure-preserving Boolean homomorphism. Let C be a closed subalgebra of A, and a countably additive
functional on the closed subalgebra [C] of B. (i) Show that is a countably additive functional on C. (ii)
Show that if is the standard extension of to B, then is the standard extension of to A.
(b) Let (X, , ) be a probability space, and T a -subalgebra of . Let (A,
) be the measure algebra
of (X, , ). Show that C = {F : F T} is a closed subalgebra of A. Identify the spaces M (A), M (C)
of countably additive functionals with L1 (), L1 ( T), as in 327D. Show that the conditional expectation
operator P : L1 () L1 ( T) (242Jd) corresponds to the map 7 C : M (A) M (C).
(c) Let (A,
) be a totally nite measure algebra and : A R a countably additive functional. Show
that, for any a A,
a =
R
0
(a [[ >
]])d
R0
(a \ [[ >
]])d,
the integrals being taken with respect to Lebesgue measure. (Hint: take (A,
) to be the measure algebra
of (X, , ); represent by a -integrable function f ; apply Fubinis 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,
) be a totally nite measure algebra, C a closed subalgebra of A and : C R a countably
additive functional with standard extension : A R. Show that, for any a A,
a =
R
0
(a [[ >
C]])d
R0
(a \ [[ >
C]])d.
122
Measure algebras
327Yb
P Sv = v for every v L1 ( T)
331B
123
Chapter 33
Maharams 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 classication 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 dening 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 eective classication of closed subalgebras and their embeddings.
Maharams theorem itself, the classication of localizable measure algebras, is in 332. I devote 331
to the denition 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 classication of free products of probability algebras.
124
331B
Maharams theorem
If e b0 6= 0, try d = e b0 . Then 0 6= d
1
2
d b = e (b b0 ) = (a (b b0 ) (b b0 )) a b
(because (b b0 ) 0) so d 21 a .
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
case, for every b B,
1
2
a. In this
1
2
d b = a (b \ b0 ) e (b \ b0 ) = ((b \ b0 ) + a (b \ b0 )) a b
(because (b \ b0 ) 0), so once again d 21 a .
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 d 2n a .
(d) Now let C be the set
{a : a A, a }.
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 Zorns Lemma, C has a maximal element c say.
(e) ?? Suppose, if possible, that c 6= . Then there is some b B such that c b 6= b ; since c ,
c b < b . Let n N be such that b > c b + 2n
b , and set
b = b c b 2n
b
for every b B. By 326I (for the second time), there is a b0 B such that b 0 for b B, b
b 0 when b B and b b0 = 0. We have
b0 ,
while
(b0 \ c) =
b0
(b0 c) b0 c b0
b0 = b + (b0 \ b ) (b \ b0 ) b > 0,
b0 \ c
such
d 2n b0 \c 2n b0 .
dc b = d b + c (b b0 ) + c (b \ b0 )
2n b0 \c b + (b b0 ) 2n
(b b0 ) (b b0 ) + (b \ b0 )
2n
(b b0 \ c) + (b b0 ) 2n
(b b0 ) + (b \ b0 )
b.
331Fb
125
so c b. Similarly, if a c C, then
(b c) =
(a c) =
(a 1) = (b 1) = b,
so b c. It follows from 312N that there is a Boolean homomorphism 1 : C1 B, extending , such that
1 a = b.
To see that 1 is measure-preserving, take any member of C1 . By 312M, this is expressible as e =
(c1 a) (c2 \ a), where c1 , c2 C. Now
(1 e) = ((c1 b) (c2 \ b)) = (c1 b) + (c2 ) (c2 b)
=
(c1 a) +
c2
(c2 a) =
e.
As e is arbitrary, 1 is measure-preserving.
331E Generating sets For the sake of the next denition, 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;
B , the smallest -subalgebra of A including B;
B , 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;
B is the -subalgebra of A generated by B, and B -generates A if A = B ;
B is the order-closed subalgebra of A generated by B, and B -generates or completely generates
A if A = B .
There is a danger inherent in these phrases, because if we have B A , where A is a subalgebra of
A, it is possible that the smallest order-closed subalgebra of A 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 deniteness let me say that the phrases B -generates A, B generates A will always refer to suprema and inma 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 dene the Maharam
type or complete generation (A) of any Boolean algebra A; it is the smallest cardinal of any subset of
A which -generates A.
(I think that this is the rst cardinal function which I have mentioned in this treatise. All you need
to know, to conrm that the denition is well-conceived, is that there is some set which -generates A;
and obviously A -generates itself. For this means that the set A = {#(B) : B A -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 (Aa ) = (A) for every non-zero a A,
writing Aa for the principal ideal of A generated by a.
126
Maharams theorem
331Fc
by 3A1Cc.
331I
127
This means that : C D satises the conditions of 331D, and must have an extension to
a measure-preserving homomorphism from the subalgebra C of A generated by C {a } to B. We
know that C is a closed subalgebra of A (314Ja), so it must be the closed subalgebra of A generated
by {a : } {a : < }. Also D = [C ] will be the subalgebra of B generated by D {b }, where
b = (a ), so is closed in B, and is the closed subalgebra of B generated by {b : < } {b : }.
(ii) The next step is to repeat the whole of the argument above, but applying it to 1
: D C , b
in place of : C D and a . Once again, we have (D ) < = (Aa ) for every a A, so we can use
Lemma 331D to nd a measure-preserving isomorphism : D+1 C+1 extending 1
, where D+1 is
the subalgebra of B generated by D {b }, and C+1 is the subalgebra of A generated by C {a }, setting
a = (b ). As in (i), we nd that C+1 is the closed subalgebra of A generated by {a : }{a : },
while D+1 is the closed subalgebra of B generated by {b : } {b : }.
(iii) We can therefore take +1 = 1 : C+1 D+1 , and see that +1 is a measure-preserving
isomorphism, extending , such that +1 (a ) = b , +1 (a ) = b . Evidently +1 extends for every
because it extends and (by the inductive hypothesis) extends for every < .
128
331I
Maharams theorem
been dened for < . Set C = < C . Then C is a subalgebra of A, because it is the union of an
S
upwards-directed family of subalgebras; similarly, D = < D is a subalgebra of B. Next, we have a
function : C D dened by setting a = a whenever < and a C ; for if , < and
a C C , then a = max(,) a = a. Clearly
S
[C ] = < [C ] = D .
Moreover, a =
a for every a C , since a =
a whenever < and a C .
Now let C be the smallest closed subalgebra of A including C , that is, the metric closure of C in A
(323J). Since C is the smallest closed subalgebra of A including C for every < , it must be the closed
subalgebra of A generated by {a : < } {a : < }. By 324O, has an extension to a measurepreserving homomorphism : C B. Set D = [C ]; by 324Kb, D is a closed subalgebra of B.
Because : C B is continuous (also noted in 324Kb),
D = [C ] = [C ]
(a) Set E = {x : x X, x() = 1}, b = E for each < . If hn inN is any sequence of distinct
elements of ,
T
T
( nN En ) = limn ( in En ) = limn 2n1 = 0,
(a bn ) 2n2
P
(a \ bn ).
a =
(a \ inf nN bn ) n=0
Similarly
So
(1 \ a) =
(supnN bn \ a)
1 =
a +
(1 \ a)
=
n=0
n=0
(a \ bn ) +
n=0
(a bn )
(bn \ a).
n=0
(bn \ a)
2n2
1 <
1,
n=0
which is impossible. X
XQ
Q
(c) Note that A is innite; for if a A the set { : b = a} must be nite, and is supposed to be
innite. So (A) must be innite.
(d) Now take a set A A, of cardinal (A), which -generates A. By (c), A is innite. Let C be the
subalgebra of A generated by A; then #(C) = #(B) = (A), by 331Gc, and C is topologically dense in A.
331N
129
as claimed.
331K Theorem Let be any innite cardinal. Let be the usual measure on {0, 1} and (B , ) its
measure algebra. Then B is Maharam-type-homogeneous, with Maharam type .
proof Set X = {0, 1} and write T for the domain of .
(a) To see that (B ) , set E = {x : x X, x() = 1}, b = E for each < . Writing E for the
algebra of subsets of X generated by {E : < }, we see that every measurable cylinder in X, as dened
in 254A, belongs to E, so that every member of T is approximated, in measure, by members of E (254Fe),
that is, {E : E E} is topologically dense in B . But this means just that the subalgebra of B generated
by {b : < } is topologically dense in B , so that {b : < } -generates B , and (B ) .
(b) Next, if c B \ {0} and (B )c is the principal ideal of B generated by c, the map a 7 a c is an
order-continuous Boolean homomorphism from B to (B )c , so by 331J we must have ((B )c ) . Thus
((B )c ) (B ) .
proof If (A) is nite, it is zero, and A = {0, 1} (331Ha, 331He) so that (interpreting {0, 1}0 as {}) we
have the case = 0. If = (A) is innite, then by 331K we know that (B , ) is also Maharam-typehomogeneous of Maharam type , so 331I gives the required isomorphism. Of course is uniquely dened
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 signicant 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 (A) =
(Ac ) whenever A is isomorphic to Ac . In general, a Boolean algebra can be Maharam-type-homogeneous
without being homogeneous (331Xj, 331Yj). But for -nite measure algebras this doesnt happen.
331N Proposition Let (A,
) be a Maharam-type-homogeneous -nite 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 on A such that
(A, ) is a probability algebra. Now let c be any non-zero member of A, and set = c, c = 1 c , where c
is the restriction of to the principal ideal Ac of A generated by c. Then (A, ) and (Ac , c ) are Maharamtype-homogeneous probability algebras of the same Maharam type, so are isomorphic as measure algebras,
and a fortiori as Boolean algebras.
130
Maharams theorem
331X
331X Basic exercises (a) Let (X, , ) be a probability space, T a -subalgebra of such that for any
non-negligible E there in an F such that F E and (F (E H)) > 0 for every
H T. Suppose
R
that f : X [0, 1] is a measurable function. Show that there is an F such that H f = (H F ) for
every H T.
> (b) Write out a direct proof of 331C not relying on 331B.
(c) Let A be a nite Boolean algebra with n atoms. Show that (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 has
Maharam type , for any r 1. (Hint: show that it is -generated by {], q] : q Qr }.)
(f ) Show that the measure algebra of any Radon measure on R r (256A) has countable Maharam type.
331 Notes
131
(f ) Let (X, ) 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(, d(X)). (ii) Show that if > max(, d(X)) and hx i< is any family in X,
then there is an x X such that #({ : x G}) > max(, d(X)) for every open set G containing x, and
that there is a sequence hn inN of distinct members of such that x = limn xn .
(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.
132
Maharams theorem
332 intro.
D = {a : a A, 0 <
a < , Aa is Maharam-type-homogeneous}
is order-dense in A. P
P If a A \ {0}, then (because (A,
) is semi-nite) there is a b a such that
0<
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 iiI consisting of members of D; by 322Kd, (A,
) is
isomorphic, as measure algebra, to the simple product of the principal ideals Ai = Aei .
(c) For each i I, (Ai ,
i ) is a non-trivial totally nite Maharam-type-homogeneous measure algebra,
i . Then (Ai ,
i ) is a Maharam-typeei , and set
i = i1
writing
i =
Ai . Take i =
i (1Ai ) =
homogeneous probability algebra, so by 331L is isomorphic to the measure algebra (Bi , i ) of the usual
measure on {0, 1}i , where i is either 0 or an innite cardinal. Thus (Ai ,
i ) is isomorphic, up to a scalar
multiple of the measure, to (Bi , i ).
Remark For the case of totally nite measure algebras, this is Theorem 2 of Maharam 42.
332C Corollary Let (A,
) be a localizable measure algebra. For any cardinal , write for the usual
measure on {0, 1} , and T for its domain. Then we can nd families hi iiI , hi iiI such that every i is
either 0 or an innite cardinal, every i is a strictly positive real number, and (A,
) is isomorphic to the
measure algebra of (X, , ), where
X = {(x, i) : i I, x {0, 1}i },
for every E .
proof Take the family hi iiI from the last theorem, take the i =
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}i , Ti , i i )
with the measure algebra of their direct sum (X, , ).
332H
133
332D The cellularity of a Boolean algebra In order to properly describe non-sigma-nite measure
algebras, we need the following concept. If A is any Boolean algebra, write
c(A) = sup{#(C) : C A \ {0} is disjoint},
the cellularity of A. (If A = {0}, take c(A) = 0.) Thus A is ccc (316A) i c(A) .
332E Proposition Let (A,
) be any semi-nite measure algebra, and C any partition of unity in A
consisting of elements of nite measure. Then max(, #(C)) = max(, c(A)).
proof Of course #(C \ {0}) c(A), because C \ {0} is disjoint, so
Now suppose that D is any disjoint set in A \ {0}. For c C, {d c : d D} is a disjoint 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 = cC Dc , so
#(D) max(, #(C), supcC #(Dc )) = max(, #(C)).
writing Aa for the principal ideal of A generated by a, as usual. I will call this the Maharam-type-
component of A. Of course e e = 0 whenever , are distinct cardinals. P
P a b = 0 whenever Aa , Ab
are Maharam-type-homogeneous of dierent Maharam types, since (Aab ) cannot be equal simultaneously
to (Aa ) and (Ab ). Q
Q
Also {e : is a cardinal} is a partition of unity in A, because
sup{e : is a cardinal} = sup{a : Aa is Maharam-type-homogeneous} = 1
by 332A. Note that there is no claim that Ae itself is homogeneous; but we do have a useful result in this
direction.
332H Lemma Let A be a Dedekind complete Boolean algebra and an innite cardinal. Let e be the
Maharam-type- 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 .
134
Maharams theorem
332H
Then d = sup{a d : a A}. Because Ad is ccc, there is a sequence han inN in A such that d = supnN d an
(316E); set bn = d an . We have (Abn ) (Aan ) = for each n; let Dn be a subset of Abn , of cardinal at
most , which -generates Abn . Set
S
D = nN Dn {bn : n N} Ad .
Thus we must have (Ab ) = for every non-zero b Ad , and Ad is Maharam-type-homogeneous of type ,
as claimed.
332I Lemma Let (A,
) be an atomless semi-nite measure algebra which is not totally nite. Then it
has a partition of unity consisting of elements of measure 1.
proof Let A be the set {a :
a = 1}, and C the family of disjoint subsets of A. By Zorns 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
a 6= 1 for every a d, so
d < 1, by 331C. So d0 = sup D is dened
in A (321C); of course d0 D, so
d0 < 1. Observe that sup C0 = 1 \ d0 .
Because
1 = , C0 must be innite; let han inN be any sequence of distinct elements of C0 . For each
d0 . Set
an =
n N, use 331C again to choose an an an such that
b0 = d0 (a0 \ a0 ),
bn = an1 (an \ an )
for every n 1. Then hbn inN is a disjoint sequence of elements of measure 1 and supnN bn = supnN an d0 .
Now
(C0 \ {an : n N}) {bn : n N}
332J Now I can formulate a complete classication theorem for localizable measure algebras, rening
the expression in 332B.
Theorem Let (A,
) and (B, ) be localizable measure algebras. For each cardinal , let e , f be the
Maharam-type- components of A, B respectively. Then (A,
) and (B, ) are isomorphic, as measure
algebras, i (i) e and f have the same magnitude for every innite cardinal (ii) for every ]0, [,
(A,
) and (B, ) have the same number of atoms of measure .
proof Throughout the proof, write Aa for the principal ideal of A generated by a, and
a for the restriction
of
to Aa ; and dene Bb , b similarly for b B.
(a) If (A,
) and (B, ) 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 satised.
(b) Now suppose that the conditions are satised. Set
K = { : is an innite cardinal, e 6= 0} = { : is an innite cardinal, f 6= 0}.
For ]0, [, let A be the set of atoms of measure in A, and set e = sup A . Write I = K ]0, [.
ei )iiI ,
Then hei iiI is a partition of unity in A, so (A,
) is isomorphic to the simple product of h(Aei ,
Aei .
ei for the restriction
writing Aei for the principal ideal generated by ei and
332L
135
In the same way, writing B for the set of atoms of measure in B, f for sup B , Bfi for the principal
ideal generated by fi and fi for the restriction of fo Bfi , we have (B, ) isomorphic to the simple product
of h(Bfi , fi )iiI .
ei )
(c) It will therefore be enough if I can show that (Aei ,
= (Bfi , fi ) for every i I.
(i) For K, the hypothesis is that e and f have the same magnitude. If they are both of nite
e ) and (Bf , f ) are homogeneous and of Maharam
magnitude, that is,
e = f < , then both (Ae ,
type , by 332H. So 331I tells us that they are isomorphic. If they are both of innite magnitude , then
e )
332I tells us that both Ae , Bf have partitions of unity C, D consisting of sets of measure 1. So (Ae ,
is isomorphic to the simple product of h(Ac ,
c )icC , while (Bf , f ) is isomorphic to the simple product
of h(Bd , d )idD . But we know also that every (Ac ,
c ), (Bd , d ) is a homogeneous probability algebra of
Maharam type , by 332H again, so by Maharams theorem again they are all isomorphic. Since C, D and
are all innite,
#(C) = c(Ae ) = = c(Bf ) = #(D)
e ) must be isomorphic
by 332E. So we are taking the same number of factors in each product and (Ae ,
to (Bf , f ).
(ii) For ]0, [, our hypothesis is that #(A ) = #(B ). Now A is a partition of unity in Ae ,
a )iaA . Similarly, (Bf , f ) is isomorphic to
e ) is isomorphic to the simple product of h(Aa ,
so (Ae ,
a ), (Bb , b ) is just a simple atom of measure ,
the simple product of h(Bb , b )ibB . Since every (Aa ,
e ) must
these are all isomorphic; since we are taking the same number of factors in each product, (Ae ,
be isomorphic to (Bf , f ).
(iii) Thus we have the full set of required isomorphisms, and (A,
) is isomorphic to (B, ).
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 { : A 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 e of innite magnitude into elements of measure 1 requires a degree of arbitrary choice.
The value of the expressions in 332C is that the parameters i , i there are sucient to identify the
measure algebra up to isomorphism.
For, amalgamating the language of 332C and 332J, we see that the
P
magnitude of e in 332J is just i = i if this is nite, #({i : i = }) otherwise (using 332E, as usual);
while the number of atoms of measure is #({i : i = 0, i = }).
(b) The classication which Maharams 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,
) be a measure algebra, and a, b A two elements of nite measure. Suppose
that : 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 : A A which extends .
proof The point is that Ab\a is isomorphic, as measure algebra, to Aa\b . P
P Set c = a b. For each innite
cardinal , let e be the Maharam-type- component of Ac . Then e a is the Maharam-type- component of
Aa , because if d c and Ad is Maharam homogeneous with Maharam type , then Ada is either {0} or again
Maharam-type-homogeneous with Maharam type . Similarly, e \ a is the Maharam-type- component of
Ac\a = Ab\a , e b is the Maharam-type- component of Ab and e \ b is the Maharam-type- component
of Aa\b . Now : Aa Ab is an isomorphism, so (e a) must be e b, and
(e \ a) =
e
(e a) =
e
(e a)
=
e
(e b) =
(e \ b).
In the same way, if we write n (d) for the number of atoms of measure in Ad , then
n (b \ a) = n (c) n (a) = n (c) n (b) = n (a \ b)
136
Maharams theorem
332L
e
v };
then ev = supiJ ci .
Now the point is that if i J then ci fv . P
P We need to consider two cases. (i) If ci is an atom,
then v ]0, [ and
ci v. So we need only observe that 1 \ fv is just the supremum in B of the atoms
of measure greater than v, none of which can meet ci , since this has measure at most v. (ii) Now suppose
that Aci is atomless, with (Aci ) = v. If 0 6= b ci , then a 7 b a : Aci Bb is an order-continuous
Boolean homomorphism, while Aci is isomorphic (as Boolean algebra) to the measure algebra of {0, 1} ,
so 331J tells us that (Bb ) . This means, rst, that b cannot be an atom, so that ci cannot meet
sup]0,[ f ; and also that b cannot be included in f for any innite < , so that ci cannot meet
sup < f . Thus ci must be included in sup f = fv . Q
Q
Of course hci iiJ is disjoint. So if ev has nite magnitude, the magnitude of fv is at least
332P
137
iJ
ci =
iJ
ci =
ev ,
the magnitude of ev . While if ev has innite magnitude, this is #(J), by 332E, which is not greater than
the magnitude of fv .
332P Proposition Let (A,
), (B, ) be atomless totally nite measure algebras. For each innite
cardinal let e , f be their Maharam-type- components. Then the following are equiveridical:
(i) (A,
) is isomorphic to a closed subalgebra of a principal ideal of (B, );
(ii) for every cardinal ,
(sup e ) (sup f ).
proof (a)(i)(ii) Suppose that : A Bd is a measure-preserving isomorphism between A and a closed
subalgebra of a principal ideal Bd of B. The Maharam-type- component of Bd is just d f , so 332O tells
us that
(sup e \ c ) (sup f \ d )
for every cardinal . Then the trivial homomorphism from A0 to B0 belongs to P , so P is not empty.
Order P by saying that if extends , that is, if c c and a = a for every a Ac . Then P
is a partially ordered set.
) If Q P is non-empty and totally ordered, it is bounded above in P . P
(
P Set c = supQ c ,
d = supQ d . For a c set a = supQ (a c ). Because Q is totally ordered, extends all the
functions in Q. It is also easy to check that 0 = 0, (a a ) = a a and (a a ) = a a for
all a, a Ac , c = d and that a =
a for every a Ac ; so that is a measure-preserving Boolean
homomorphism from Ac to Bd .
Now suppose that is any cardinal; then
(sup e \ c ) = inf
(sup e \ c ) inf (sup f \ d ) = (sup f \ d ).
d = d b,
c .
(sup e \ c ) =
(sup e \ c ) (sup f \ d ) = (sup f \ d ).
If 0 < 1 ,
138
332P
Maharams theorem
(sup e \ c ) =
(sup e \ c )
( sup e \ c )
(e0 \ c )
( sup e \ c ) ( sup f \ d )
0
= ( sup f \ d )
1
If > 1 ,
(sup e \ c ) =
(sup e \ c ) (sup f \ d ) = (sup f \ d ).
(sup e ) = (sup f )
for every . Because all these measures are nite,
e =
(sup e )
(sup e )
>
= (sup f (sup f ) = f
>
for every .
Similarly, writing e , f for the suprema in A, B of the atoms of measure , 332O tells us that
(sup e ) = (sup f )
for every ]0, [, and hence that
e = f for every , that is, that A and B have the same number of
atoms of measure .
So (A,
) and (B, ) are isomorphic, by 332J.
332R 332J tells us that if we know the magnitudes of the Maharam-type- components of a localizable
measure algebra, we shall have specied 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,
) be a semi-nite measure algebra. Then c(A) 2 (A) .
proof Let C A \ {0} be a disjoint set, and B A a -generating set of size (A).
(a) If A is purely atomic, then for each c C choose an atom c
If c1 , c2 are distinct members of C, the set
{a : a A, c1
a c2
c1
or
c2 ,
b}.
a}
so cannot include B, and f (c1 ) 6= f (c2 ).
332T
139
(b) Now suppose that A is not purely atomic; in this case (A) is innite. For each c C choose
an element c c of non-zero nite measure. Let B be the subalgebra of A generated by B. Then the
topological closure of B is A itself (323J), and #(B) = (A) (331Gc). For c C set
1
2
c }.
f (c) = {b : b B,
(b c )
Then f : C PB is injective. P
P If c1 , c2 are distinct members of C, then (because B is topologically
dense in A) there is a b B such that
1
3
c2 ).
min(
c1 ,
c1 ,
(c1 \ b)
1
3
(c2 b)
c2 ,
for < . Set B = {b : < } {b : < } if is innite, {b : < } if is nite; then #(B) . Note
that if c C and b Bc there is a b B such that b = b c. P
P Since Bc 6= , (Ac ) > 0; but this means
Q
that (Ac ) is innite (see 331H) so is innite and b B, where = fc (b); now b = b c. Q
Let B be the closed subalgebra of A generated by B. Then C B. P
P For c C, we surely have c b
if f (c); but also, because C is disjoint, c b = 0 if \ f (c). Consequently
c = inf f (c) b inf \f (c) (1 \ b )
includes c. On the other hand, if d is any other member of C, there is some f (c)f (d), so that
d c
b (1 \ b )
= 0.
140
Maharams theorem
332T
332 Notes
141
142
Maharams 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-nite measure algebras we have a denite answer (332F).
Juha
Maharams classication 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 nite measure algebras because the non-totally-nite case (332Ya, 332Yc) seems
to require a new idea, while atoms introduce acute combinatorial complications.
I oer 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 a countably -generated Dedekind complete Boolean algebra
A with cellularity (Koppelberg 89, 13.1), so that P is isomorphic to an order-closed subalgebra of A,
and if > c then (P) > (332R).
For totally nite measure algebras we have a kind of weak Schroder-Bernstein theorem: if we have two
of them, each isomorphic to a closed subalgebra of the other, they are isomorphic (332Q). This fails for
-nite algebras (332Xn). I call it a weak Schroder-Bernstein theorem because it is not clear how to build
the isomorphism from the two injections; strong Schroder-Bernstein theorems include denite recipes for
constructing the isomorphisms declared to exist (see, for instance, 344D below).
proof (a) Let D Ab be a set of cardinal Cb (Ab ) such that DCb -generates Ab . Set D = {d a : d D}.
Then D Ca -generates Aa . P
P Apply 313Mc to the map d 7 d a : Ab Aa , as in 331Hc. Q
Q Consequently
Ca (Aa ) #(D ) #(D) = Cb (Ab ),
333C
Closed subalgebras
143
(b) Just as in the proof of 332A, given b A \ {0}, there is an a Ab \ {0} minimising Ca (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 Ca (Aa ) is nite. Let D Aa be a nite set such that D Ca -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 -generates Ab ; so that Cb (Ab ) = 0. Since a is relatively Maharam-typehomogeneous over C, Ca (Aa ) must be zero, that is, Aa = Ca .
(e) The middle inequality is true just because Aa will be -generated by D Ca whenever it is -generated
by D Da . The neighbouring inequalities are special cases of the middle one, and the outer equalities are
elementary.
333C Theorem Let (A,
) and (B, ) be totally nite measure algebras, and C a closed subalgebra of
A. Let : C B be a measure-preserving Boolean homomorphism.
(a) If, in the notation of 333A, C (A) [C]b (Bb ) for every non-zero b B, there is a measure-preserving
Boolean homomorphism : A B extending .
(b) If Ca (Aa ) = [C]b (Bb ) for every non-zero a A, b B, then there is a measure algebra isomorphism
: A B extending .
proof In both parts, the idea is to use the technique of the proof of 331I to construct as the last of an
increasing family h i of measure-preserving homomorphisms from closed subalgebras C of A, where
= C (A). Let ha i< be a family in A such that C {a : < } -generates A. Write D for [C];
remember that D is a closed subalgebra of B (324L).
(a)(i) In this case, we can describe the C immediately; C will be the closed subalgebra of A generated
by C {a : < }. The induction starts with C0 = C, 0 = .
(ii) For the inductive step to a successor ordinal + 1, where < , suppose that C and have
been dened. Take any non-zero b B. We are supposing that Db (Bb ) > #(), so Bb cannot be
-generated by
D = Db {b a : < } = [C]b {b a : < } = [C {a : < }],
144
Maharams theorem
333C
a A \ {0},
Ca (Aa ) = > #({a : } {a : < }),
homomorphism : D+1 C+1 , where D+1 is the subalgebra of B generated by D {b }, that is,
the closed subalgebra of B generated by D {b : } {b : < }, and C+1 is the subalgebra of A
generated by C {a }, setting a = (b ).
We can therefore take +1 = 1 : C+1 D+1 , as in 331I.
(ii) The inductive step to a S
non-zero limit ordinal is exactly the same
S as in (a) above or in 331I;
C is the metric closure of C = < C , D is the metric closure of D = < D , and is the unique
measure-preserving homomorphism from C to D extending every for < .
(iii) The induction stops, as before, with = : C D , where C = A, D = B.
333D Corollary Let (A,
) and (B, ) be totally nite measure algebras and C a closed subalgebra of
A. Suppose that
(C) < max(, (A)) min{ (Bb ) : b B \ {0}}.
P
P There is a set C C, of cardinal (C), which -generates C, so that C = {b c : c C} -generates
[C]b . Now there is a set D Bb , of cardinal [C]b (Bb ), such that [C]b D -generates Bb . In this case
C D must -generate Bb , so #(C D). But #(C ) #(C) < and is innite, so we must have
#(D) , as claimed. Q
Q
On the other hand, C (A) (A) . So we can apply 333C to give the result.
333E Theorem Let (C,
) be a totally nite measure algebra, an innite cardinal, and (B , ) the
be the localizable measure algebra free product
measure algebra of the usual measure on {0, 1} . Let (A, )
of (C,
) and (B , ), and : C A the corresponding homomorphism. Then for any non-zero a A,
[C]a (Aa ) = ,
333E
145
Closed subalgebras
We know that {b : < } -generates B (see part (a) of the proof of 331K). Consequently [C] {b :
< } -generates A. P
P Let A1 be the closed subalgebra of A generated by [C] {b : < }. Because
2
E = [C] {b : 6= },
F = {e0 . . .
en
: e0 , . . . , en E}.
G = {d : d A, (b
2
Secondly, if d, d G and d
d , then
(d \ d)) = (b
d ) (b
d) = 1 d
1 d
= 1 (d
\ d),
(b
2
so d
\d
1
1
sup H) = (sup
(b
dH b d) = supdH (b d) = supdH d = (sup H),
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 E . Q
Q
P By (a) and 323J, A is the metric closure of the
(d) The next step is to see that [C]a (Aa ) > 0. P
0 a) 1 a.
Now there is a nite J such that a0 belongs to the subalgebra A1 generated by [C] {b : J}.
Take any \ J (this is where I use the hypothesis that is innite). If c C, then by (c) we have
(c b )) (a
0 (c b )) (a
a0 )
((a
c) (a b )) = (a
0 b ) + (a
0 c) 2(a
0 c b ) (a
a0 )
= (a
1
a0
2
a0 )
(a
Q
Thus a b is not of the form a c for any c C, and Aa 6= [C]a , so that [C]a (Aa ) > 0. Q
P There is a non-zero d a which is relatively Maharam-type(e) It follows that [C]a (Aa ) is innite. P
homogeneous over [C]. By (d), applied to d, [C]d (Ad ) > 0; but now 333Bd tells us that [C]d (Ad ) must
Q
be innite, so [C]a (Aa ) is innite. Q
146
333E
Maharams theorem
(f ) If = , we can stop here. If > , we continue, as follows. Let D Aa be any set of cardinal
less than . Each d D {a} belongs to the closed subalgebra of A generated by C = [C] {b : < }.
But because A is ccc, this is just the -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 = { : b Cd } is
countable. Set J = dD{a} Jd ; then
#(J) max(, #(D {a})) = max(, #(D)) < ,
b b ) 6= 1 (a
b ). Thus a b cannot belong to the closed subalgebra of Aa generated by
and (a
2
[C]a D, and [C]a D does not -generate Aa . As D is arbitrary, [C]a (Aa ) .
This completes the proof.
is satised.
(a) If C C is a set of cardinal (C) which -generates C, and B B a set of cardinal which
-generates B (331K), then {c b : c C, b B} is a set of cardinal at most max(, (C), ) = which
-generates A (because the subalgebra it generates is topologically dense in A, by 325Dc). So (A) .
On the other hand, if a A is non-zero, then (Aa ) [C]a (Aa ) , by 333E; so A is Maharam-typehomogeneous, with Maharam type .
(b) Writing D = {c 1 : c C} for the canonical image of C in A, we have a measure-preserving
automorphism 1 : D D dened by setting 1 (c 1) = c 1 for every c C. Because 1 [D] D, 333Be
and 333E tell us that
= (Aa ) 1 [D]a (Aa ) Da (Aa ) =
for every non-zero a A, so we can use 333Cb, with B = A, to see that 1 can be extended to a measurepreserving automorphism on A.
333I
Closed subalgebras
147
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
A = {a : a A,
a < } is order-dense in A. So there is a partition of unity hai iiI in A consisting of
members of A (313K). For each i I, set i c =
(ai c) for every c C; then i is non-negative, and it is
completely additive by 327E. Because hai iiI is a partition of unity in A,
P
P
c = iI
(c ai ) = iI i c
P Cci is a Dedekind -complete
for every c C. Next, (Cci , i Cci ) is a totally nite measure algebra. P
Boolean algebra because C is. i Cci is a non-negative countably additive functional because i is. If c Cci
and i c = 0, then c = 0 by the choice of ci . Q
Q Note also that
(ai \ ci ) = i (1 \ ci ) = 0,
so that ai
ci .
(b) By 333Bd, any nite i must actually be zero. The next element we need is the fact that, for
P
Cai ). P
each i I, we have a measure-preserving isomorphism c 7 c ai from (Cci , i Cci ) to (Cai ,
Of course this is a ring homomorphism. Because ai ci , it is a surjective Boolean homomorphism. It is
measure-preserving by the denition of i , 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 i , extending the canonical homomorphism
we have a measure-preserving isomorphism
i : Aai Cai B
b i . P
P When i is innite, this is just 333F(ii). But the only other case is when
c 7 c 1 : Cai Cai B
b i
Q
i = 0, that is, Cai = Aai , while Bi = {0, 1} and Cai B
= Cci . Q
b i and
Cai ) induces an isomorphism between Cci B
The isomorphism between (Cci , i Cci ) and (Cai ,
b
b
Cai Bi . So we have a measure-preserving isomorphism i : Aai Cci Bi such that i (c ai ) = c 1
for every c Cci .
Q
(d) By 322Kd, we have a measure-preserving isomorphism a 7 ha ai iiI : A iI Aai .
Q Putting this together with the isomorphisms of (c), we have a measure-preserving isomorphism : A
b
iI Cci Bi , setting a = hi (a ai )iiI for a A. Observe that, for c C,
c = hi (c ai )iiI = h(c ci ) 1iiI ,
as required.
333I Remarks (a) I hope it is clear that whenever (C,
) is a Dedekind complete
measure algebra,
P
hi iiI is a family of non-negative completely additive functionals on C such that iI i =
, and hi iiI
148
Maharams theorem
333I
is a family of cardinals all innite or zero, then the construction above can be applied to give a measure alb i iiI , together with an order-continuous measure-preserving
gebra (A,
), the product of the family hCci B
homomorphism : C A; and that the partition of unity hai iiI in A corresponding to this product (315E)
b i ,
has i c =
(ai c) for every c C and i I, while each principal ideal Aai can be identied with Cci B
so that ai is relatively Maharam-type-homogeneous over [C]. Thus any structure (C,
, hi iiI , hi iiI ) 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 classication of objects (A,
, C), where (A,
) is a
localizable measure algebra and C is a closed subalgebra, corresponding to the classication of localizable
measure algebras in terms of the magnitudes of their Maharam-type- components in 332J. The general
case seems to be complex. But I can deal with the special case in which (A,
) is totally nite. In this case,
we have the following facts.
333J Lemma Let (A,
) be a totally nite 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 hn inN of additive functionals on C such
that (i) n+1 n for every n (ii) there is a disjoint sequence han inN in A such that supnN an = sup A
and n c =
(an c) for every n N, c C.
Remark I hope it is plain from my wording that it is the n which are unique, not the an .
proof (a) For each a A set a (c) =
(c a) for c C. Then a 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 , c C such that 0 6= c c and a (d) e (d)
whenever d C, e Ab and d c . P
P?? Otherwise, choose han inN and hcn inN as follows. Since 0, c wont
c and set
serve for a, c , there must be an a0 Ab such that a0 (c) > 0. Let > 0 be such that a0 (c) >
C]]; then c0 C and 0 6= c0 c. Given that an Ab , cn C and 0 6= cn c, then there
c0 = c [[a0 >
must be an+1 Ab , dn C such that dn cn and an+1 (dn ) > an (dn ). Set cn+1 = dn [[an+1 > an ]], so
that cn+1 C and 0 6= cn+1 cn , and continue.
There is some n N such that n 1. For any i < n, the construction ensures that
0 6= cn+1
c0
ci+1
so
cn+1 .
But as 0 6= d
XQ
Q
[[ai+1 > ai ]] for every i < n, a0 (d) < a1 (d) < . . . < an (d), so this is impossible. X
(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 a e whenever e Ab . P
P (i) By (b), the set C of those c C such that there is an a Ab such
that a Cc e Cc for every e Ab is order-dense in C. Let hci iiI be a partition of unity in C consisting
of members of C, and for each i I choose ai Ab such that ai Cci e Cci for every e Ab . Consider
a = supiI ai ci . (ii) If a A and a a, then for each i I there is a di C such that ai a = ai di .
Set d = supiI ci di ; then (because hci iiI is disjoint)
a d = supiI ai ci di = supiI ai ci a = a a = a .
333K
Closed subalgebras
149
e = e (1) an (1) =
But as han inN is disjoint, this means that e = 0, that is, a supnN an . As a is arbitrary, sup A supnN an .
Q
Q
(e) Thus we have a sequence hn inN of the required type, witnessed by han inN . To see that it is
unique, suppose that hn inN , han inN are another pair of sequences with the same properties. Note
rst that if c C and 0 6= c [[i > 0]] there is some k N such that c ai ak 6= 0; this is because
(ai c) = i (c) > 0, so that ai c 6= 0, while ai sup A = supkN ak . ?? Suppose, if possible, that there is
some n such that n 6= n ; since the situation is symmetric, there is no loss of generality in supposing that
n 6 n , that is, that c = [[n > n ]] 6= 0. For any i n, i n so c [[i > 0]]. We may therefore choose
c0 , . . . , cn+1 Cc \ {0} and k(0), . . . , k(n) N such that c0 = c and, for i n,
(choosing k(i), recalling that 0 6= ci
ci ai ak(i) 6= 0
ci+1 C, ci+1
ci , ci+1 ai
(choosing ci+1 , using the fact that ai 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 n (cn+1 ) < n (cn+1 ). On the other hand, for
each i n,
cn+1 ai ak(i) = cn+1 ci+1 ai ak(i) = cn+1 ai = cn+1 ak(i) ,
so
(cn+1 ai ) =
(cn+1 ak(i) ) = k(i) (cn+1 ),
n (cn+1 ) < n (cn+1 ) 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 ai = cn+1 ak(i) = cn+1 ak(j) = cn+1 aj 6= 0
X
because 0 6= cn+1 [[j > 0]]. So ai , aj cannot be disjoint, breaking one of the rules of the construction. X
b of c Ce .
for each c C, writing c 1 for the canonical image in Ce B
proof (a) I aim to use the construction of 333H, but taking much more care over the choice of hai iiI in
part (a) of the proof there. We start by taking han inN as in 333J, and setting n c =
(an c) for every
n N, c C; then these an will deal with the part in sup A, as dened in the proof of 333J.
150
Maharams theorem
333K
(b) The further idea required here concerns the treatment of innite . Let hbi iiI be any partition
of unity in A consisting of non-zero members of A which are relatively Maharam-type-homogeneous over
C, and hi iiI the corresponding cardinals, so that i = 0 i bi A. Set I1 = {i : i I, i }. Set
K = {i : i I1 }, so that K is a countable set of innite cardinals, and for K set J = {i : i = },
P (Compare
a = supiJ bi for K. Now every a is relatively Maharam-type-homogeneous over C. P
332H.) J must be countable, because A is ccc. If 0 6= a a , there is some i J such that a bi 6= 0; now
Ca (Aa ) Cabi (Aabi ) = i = .
Cbi Di {b bi : b B} = B Abi
Q
This shows that a is relatively Maharam-type-homogeneousover C, with Ca (Aa ) = . Q
Since evidently hJ iK and ha iK are disjoint, and supK a = supiI1 bi , this process yields a
partition hai iiNK of unity in A. Now the arguments of 333H show that we get an isomorphism of the
kind described.
(c) To see that the families hn inN , h iK (and therefore the ei and the (Cei , i Cei ), but not ) are
uniquely dened, argue as follows. Take families h
n inN , h
iK which correspond to an isomorphism
Q
Q
b ,
: A D = nN Cen K Ce B
Q
Q
In the simple product
b , we have a partition
writing ei = [[
i > 0]] for i N K.
Ce B
nN Cen
K
of unity hei iiNK corresponding to the product structure. Now for d ei , we have
[C]d (Dd ) = 0 if i N,
= if i = K.
must be
So K
and for K,
{ : , a A, Ca (Aa ) = } = K,
1 e = sup{a : a A, Ca (Aa ) = } = a ,
so that
= . On the other hand, h
1 en inN must be a disjoint sequence with supremum sup A, and
the corresponding functionals
n are supposed to form a non-increasing sequence, so must be equal to the
n by 333J.
333L Remark Thus for the classication of structures (A,
, C), where (A,
) is a totally nite measure
algebra and C is a closed subalgebra, it will be enough to classify objects (C,
, hn inN , h iK ), where
(C,
) is a totally nite measure algebra,
hn inN is a non-increasing sequence of non-negative countably additive functionals on C,
K is a countable set of innite cardinals (possibly empty),
333M
151
Closed subalgebras
C0 for which Ei
= ei for every i I, R, writing N for 1 [{0}];
(b) this formula determines both and ;
(c) for every E , i I, we have
i E =
x(i)(dx);
where I write E for the equivalence class of E in /N , and E for the equivalence class of E in /N .
Then A is a -subalgebra of , because E 7 E , E 7 E are both sequentially order-continuous Boolean
homomorphisms, and contains every Ei , so must be the whole of . Consequently
E =
E =
E = E
for every E , and = ; it follows at once that = . So and are uniquely determined.
(c) If E and i I,
Z
x(i)(dx) =
x(i)E(x)(dx) =
(y)(i)E((y))(dy)
152
Maharams theorem
333M
(e)-(g) The point is that, because the standard-extension operator is order-preserving (327F(b-ii)),
i 0 i C0 0
Z
x(i) 0 -a.e.,
i j i C0 j C0
Z
Z
x(j)(dx) for every E
x(i)(dx)
i = 0 i C0 = 0
Z
x(i) = 0 -a.e..
333N A canonical form for closed subalgebras We now have all the elements required to describe
a canonical form for structures
(A,
, C),
where (A,
) is a totally nite measure algebra and C is a closed subalgebra of A. The rst step is the
matching of such structures with structures
(C,
, hn inN , h iK ),
where (C,
) is a totally nite measure algebra, hn inN is a non-increasing sequence of non-negative countably additive functionals on C, K is a countable set of
cardinals,
h iK is a family of non-zero
Pinnite
P
x(i)(dx)
(333Mc), so that i C0 is xed by and . Moreover, the functionals i can be recovered from their
restrictions to C0 by the formulae of 327F (333Md). Thus from (C,
, hi iiI ) we are led, by a canonical and
reversible process, to the structure
(C,
, C0 , I, , ).
[C0 ]
for E , j NL. This process too is reversible; every absolutely continuous countably additive functional
on corresponds to countably additive functionals on C0 and C0 . Let me repeat that the results of 327F
mean that the whole structure (C,
, hi iiI ) can be recovered from (C0 ,
C0 , hi C
0 iiI ) if we can get the
P
description of (C,
) right, and that the requirements i 0, n n+1 , 6= 0, iI i =
imposed in
333K will survive the process (327F(b-iv)).
Putting all this together, a structure (A,
, C) leads, in a canonical and (up to isomorphism) reversible
way, to a structure
(K, , L, h iNL )
333P
Closed subalgebras
153
such that
K and L are countable sets of innite cardinals,
is a totally nite measure on R I , where I = N K, and its domain is precisely the
-algebra of subsets of R I dened by the coordinate functionals,
forP-almost every x R I we have x(i) 0 for every i I, x(n) x(n + 1) for every n N
and iI x(i) = 1,
for K, {x : x() > 0} > 0,
P
, 6= 0 see
(these two sections corresponding to the requirements i 0, n n+1 ,
iI i =
333M(e)-(g))
for j J = N L, j is a non-negative countably additive functional on ,
P
n n+1 for every n N, 6= 0 for every L, jJ j = .
333O Remark I do not envisage quoting the result above very often. Indeed I do not claim that its nal
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,
, C) under consideration, and
I hope that you will experiment in this direction.
333P At the risk of trespassing 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 rst
step is an important result concerning automorphisms of homogeneous probability algebras.
Proposition Let (B, ) be a homogeneous probability algebra. Then there is a measure-preserving automorphism : B B such that
for all b, c B.
limn (c n (b)) = c b
proof (a) The case B = {0, 1} is trivial ( is, and must be, the identity map) so we may take it that
B is the measure algebra of {0, 1} with its usual measure , where is an innite cardinal. Because
#( Z) = max(, ) = , there must be a bijection : such that every orbit of in is
innite (take to correspond to the bijection (, n) 7 (, n + 1) : Z Z). This induces a
H = {x : xm E} = {x : xm J E}
= {zm : z E}.
So (c m (b)) = (F H). But L and K are disjoint, because
where L = m [J] and H
m n, so F and H must be independent (cf. 272K), and
as claimed. Q
Q
(c m (b)) = F H = F E = c b,
(d) Now recall that for every E , > 0 there is an E 0 such that (EE ) (254Fe). So,
given b, c B and > 0, we can nd b , c C such that (b b ) and (c c ) , and in this case
154
333P
Maharams theorem
lim sup|
(c n (b)) c b|
n
lim sup |
(c n (b)) (c n (b ))|
n
+ |
(c n (b )) c b | + |
c b c b |
= lim sup |
(c n (b)) (c n (b ))| + |
c b c b |
n
+ c|
b b | + |
c c |
b
(c c ) + (b b ) + c (b b ) + (c c )
b 4.
As is arbitrary,
limn (c n (b)) = c b,
as required.
Remark Automorphisms of this type are called mixing (see 372P below).
333Q Corollary Let (C,
0 ) be a totally nite measure algebra and (B, ) a probability algebra which
is either homogeneous or purely atomic with nitely many atoms all of the same measure. Let (A,
) be
the localizable measure algebra free product of (C,
0 ) and (B, ). Then there is a measure-preserving
automorphism : A A such that
{a : a A, a = 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 rst. In this case, if B has n atoms b0 , . . . , bn1 , let
: B B be the measure-preserving homomorphism cyclically permuting these atoms, so that b0 =
b1 , . . . , bn1 = b0 . Because is an automorphism of (B, ), it induces an automorphism of (A,
);
any member of A is uniquely expressible as a = supi<n ci bi , and now a = supi<n ci bi+1 , if we set
bn = b0 . So a = a i ci = ci+1 for i < n 1 and cn1 = c0 , that is, i all the ci are the same and
a = supi<n c bi = c 1 for some c C.
Express a as supiI ci bi , where hci iiI is a nite partition of unity in C (315Na). Then
a = supiI ci (bi ),
lim
(a n (a )) = lim
(sup ci (bi n (bi )))
n
iI
X
X
= lim
0 ci (bi n (bi )) =
0 ci (
bi )2 .
iI
iI
It follows that
X
iI
0 ci (
bi )2 = lim
(a n (a ))
n
lim sup
(a n (a))
(a a )
( n (a) n (a ))
n
X
=
a 2
(a a )
a 3
(a a )
0 ci bi 32 ,
iI
333R
155
Closed subalgebras
that is,
iI
0 ci bi (1 bi ) 32 .
Then I \ (K L) J, so
(a (c 1)) =
iI\K
X
iJ
X
iJ
L = {i : i I \ K, bi 2},
0 ci bi +
0 ci bi +
0 ci + 2
iK
X
iL
X
iL
iJ
0 ci 3. Set
c = supiK ci .
0 ci (1 bi )
0 ci bi +
0 ci + 2
iK
iK
0 ci (1 bi )
0 ci 3 + 2 = 5,
and
(a (c 1)) 2 + 5.
As 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, 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 innite Card , let
(B , ) be the measure algebra of the usual measure on {0, 1} . For nite n Card , let Bn be the power
set of {0, . . . , n 1} and set n b = n1 #(b) for b Bn .
Theorem Let (A,
) be a totally nite 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, c = c for every G};
(ii) C is a closed subalgebra of A and there
Q is a partition of unity hei iiI 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 iI Cei B
b
by ei and endowed with
Cei , and Cei Bi for the localizable measure algebra free product of Cei and Bi
the isomorphism being one which takes any c C to h(c ei ) 1iiI , as in 333H and 333K;
(iii) there is a single measure-preserving automorphism of A such that
C = {c : c A, c = c}.
) C is a subalgebra because every G is a Boolean homomorphism, and it is orderproof (a)(i)(ii)(
closed because every is order-continuous (324Kb). (Or, if you prefer, C is topologically closed because
every is continuous.)
) Because C is a closed subalgebra of A, its embedding can be described in terms of families hn inN ,
(
h iK as in Theorem 333K. Set I = K N. Recall that each i is dened by setting i c =
(c ai ), where
hai iiI is a partition of unity in A (see the proofs of 333H and 333K). Now for K, a is the maximal
element of A which is relatively Maharam-type-homogeneous over C with relative Maharam type (part
(b) of the proof of 333K). Consequently we must have a = a for any measure algebra automorphism of
(A,
) which leaves C invariant; in particular, for every G. Thus a C for every K.
( ) Now consider the relatively atomic part of A. The elements an , for n N, are not uniquely
dened. However, the functionals n and their supports en = [[n > 0]] are uniquely dened from the
structure (A,
, C) and therefore invariant under G. Observe also that because supnN an = 1 \ supK a
belongs to C, and en = inf{c : c C, c an }, while en en+1 for every n, we must have e0 = supnN an .
156
333R
Maharams theorem
Let G be the set of all those automorphisms of the measure algebra (A,
) such that c = c for every
c C. Then of course G is a group including G. Now supG an 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 en .
( ) I claim now that if n N then en [[0 > n ]] = 0. P
P?? Otherwise, set c = [[0 > n ]] en . Then
0 c > 0 so c a0 6= 0. By the last remark in (), there is a G such that c a0 an 6= 0. Now there is
a c C such that c a0 an = c a0 , and of course we may suppose that c c. But this means that
(c an ) = c an c a0 an = c a0 ,
so that
n c =
(c an ) =
(c an )
(c a0 ) = 0 c ,
e = inf nN en , 0 e = 0; but this must mean that e = 0. Consequently, setting I = I \ {0}, en = en1 \ en
for n 1, e = a for K, we nd that hei iiI is a partition of unity in C.
Moreover, for n 1 and c en , we must have
P
P
c = iI i c = k<n k c = n0 c,
so that k c = n1 c for every k < n. But this means that we have a measure-preserving homomorphism
b n given by setting
n : Aen Cen B
n (ak c) = c {k}
ek ,
1
n
c =
c n {k}
(ak c) = k c =
for all relevant k and c. Because Bn is nite, n is surjective.
( ) Just asQ
in 333H, we now see that because
hei iiI is a partition of unity in A as well as in C, we can
Q
b i.
identify A with iI Aei and therefore with iI Cei B
Q
b i , writing : A D for the canonical map. For each i I,
(b)(ii)(iii) Let us work in D = iI Cei B
b i with xed-point subalgebra {c 1 : c Cei }
we have a measure-preserving automorphism i of Cei B
(333Q). For d = hdi iiI D, set
d = hi di iiI .
Then is a measure-preserving automorphism because every i is. If d = d, then for every i I there
must be a ci ei such that di = ci 1. But this means that d = c, where c = supiI ci C. Thus
the xed-point subalgebra of is just [C]. Transferring the structure (D, [C], ) back to A, we obtain a
measure-preserving automorphism 1 of A with xed-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) (denition: 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 R I , given by the formula {x} = c.
(c) Let (A,
) be totally nite measure algebra and G a set of measure-preserving Boolean homomorphisms
from A to itself such that G for all , G. (i) Show that a supG a for every a A. (Hint: if
c c, where G and c A, then c = c; apply this to c = supG a.) (ii) Set C = {c : c A, c = c
for every G}. Show that supG a = upr(a, C) for every a A.
333 Notes
Closed subalgebras
157
333Y Further exercises (a) Show that when I = N the algebra of subsets of R I , used in 333M, is
precisely the Borel -algebra as described in 271Ya.
(b) Let (A,
) be a totally nite measure algebra, and B, C two closed subalgebras of A with C B.
Show that C (B) C (A). (Hint: use 333K and the ideas of 332T.)
(c) Let (A,
) be a probability algebra. Show that A is homogeneous i there is a measure-preserving
automorphism of A which is mixing in the sense of 333P.
(d) Let (A,
) be a totally nite measure algebra, and G a set of measure-preserving Boolean homomorphisms from A to itself. Set C = {c : c A, c = c for every 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 supnN an
still belongs to C.)
333 Notes and comments I have done my best, in the rst 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
denitions and arguments. Thus the Maharam type (A) of an algebra is just the relative Maharam type
{0,1} (A), and A is Maharam-type-homogeneous i it is relatively Maharam-type-homogeneous over {0, 1}.
To help you trace through the correspondence, I list the code numbers: 331Fa333Aa, 331Fb333Ac,
331Hc333Ba, 331Hd333Bd, 332A333Bb, 331I333Cb, 331K333E, 331L333F(ii), 332B333H,
332J333K. 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 dierent
facet of homogeneous algebras, and a dierent 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 have topological density at most max(, ) (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 rst step in forming a picture of the possible embeddings of a given (totally nite) measure algebra C
in a larger algebra A. If we think of C as the measure algebra of a measure space (X, , ), then we can be
sure that A is representable as a closed subalgebra of the measure algebra of X {0, 1} for some , that
is, the measure algebra of T where is the product measure on X {0, 1} and T is some -subalgebra
of the domain of ; the embedding of C in A being dened by the formula E (E {0, 1} ) for E
(325A, 325D). Identifying, in our imaginations, both X and {0, 1} 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 dierence between 332B and 333H,
for instance, is partly in the need for the functionals i in the latter, whereas in the former the decomposition
involves only principal ideals with the induced measures. Because the i are completely additive, they all
have supports ci (326Xi) and we get measure algebras (Cci , i Cci ) to use in the products. (I note that the
ci can be obtained directly from the ai , without mentioning the functionals i , 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), simplied
because I am no longer seeking to deal with algebras of innite magnitude.
Theorem 333K gives a canonical form for superalgebras of a given totally nite measure algebra (C,
),
b=
taking the structure (C,
) itself for granted. I hope it is clear that while the i amd ei and the algebra A
Q
Q
b
b
K Ce B and the embedding of C in A are uniquely dened, the rest of the isomorphism
nN Cen
b
: A A generally is not. Even when the a are uniquely dened the isomorphisms between Aa and
b depend on choosing generating families in the Aa ; see the proof of 333Cb.
Ce B
158
Maharams theorem
333 Notes
To understand the possible structures (C, hi iiI ) 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 hi iiI and identify it with the measure
algebra of a particular measure on R I . Perhaps I should apologise for not stating explicitly in the course of
Lemma 333M that the measures 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 dened by the construction proposed, I would complete them
and call them Radon measures. In order to validate this approach, I must show that the i 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 -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 nd 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 nished with the insights which
Maharams 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 simplication (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 rst step to understanding any automorphism will
be a description of its xed-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 classication of the measure algebras of product
measures, or, if you prefer, of the free products of measure algebras. The complete classication, 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-nite measure algebras (334B), and go on to
the corresponding results for products of probability spaces and algebras (334C-334D). Finally, I show that
any innite power of a probability space is Maharam-type-homogeneous (334E).
334A Theorem Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras A, B. Let be
its measure algebra. Then the Maharam type (C) is at
the c.l.d. product measure on X Y , and (C, )
most max(, (A), (B)).
proof Recall from 325A that we have order-continuous Boolean homomorphisms 1 : A C and 2 : B C
dened by setting 1 E = (E Y ) , 2 F = (X F ) for E , F T. Let A A, B B be -generating
sets with #(A) = (A), #(B) = (B); set C = 1 [A] 2 [B]. Then C -generates C. P
P Let C1 be the
order-closed subalgebra of C generated by C. Because 1 is order-continuous, 1
1 [C1 ] is an order-closed
subalgebra of A, and it includes A, so must be the whole of A; thus 1 a C1 for every a A. Similarly,
2 b C1 for every b B.
This means that
1 = {W : W , W C1 }
c} C1
334E
Products
159
(b) Fix on b A \ {0, 1}. For each i I, let i : 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 nite subalgebra of C generated by {i b : i J} has atoms of
measure at most n , where = max(
b, 1
b) < 1. Consequently C can have no atom of measure greater
than n , for any n, and is therefore atomless.
(c) Because I is innite, there is a bijection between I and I N; that is, there is a partition hJi iiI of
I into countably innite sets. Now (X I , ) can be identied with the product of the family h(X Ji , i )iiI ,
160
Maharams theorem
334E
where i is the product measure on X Ji (254N). By (b), every i is atomless, so there are sets Ei X Ji
of measure 21 . The sets Ei = {x : xJi 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 I , dened
by setting f (x)(i) = 1 if x Ei , 0 otherwise, and therefore a measure-preserving Boolean homomorphism
: BI C, writing BI for the measure algebra of I .
Now if c C \ {0} and Cc is the corresponding ideal, b 7 c b : BI Cc is an order-continuous Boolean
homomorphism. It follows that (Cc ) #(I) (331J).
(d) Again take any non-zero c C. For each i I, set ai = inf{a : i a c}. Writing Aai for the
corresponding principal ideal of A, we have an order-continuous Boolean homomorphism i : Aai Cc ,
given by the formula
i a = i a c for every a Aai .
Now i is injective, so is a Boolean isomorphism between Aai and its image i [Aai ], which by 314F(a-i) is
a closed subalgebra of Cc . So
(Aai ) = (i [Aai ]) (Cc )
by 332Tb.
For any nite J I,
Q
Q
(inf
ai .
0 < c
iJ i ai ) =
iJ
iJ (i ai ) =
max( (A), #(I)) max(, (Cc )) = (Cc ) (C) max( (A), #(I))
334X Basic exercises (a) Let (X, , ) and (Y, T, ) be complete locally determined measure spaces
with c.l.d. product (X Y, , ). Let A, C be the measure algebras of , respectively. Show that if Y > 0
then (A) (C).
(b) Let X be a set and and two totally nite measures on X with the same domain ; then = +
for the three measure algebras. Show
is also a totally nite measure. Write (A,
), (B, ) and (C, )
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) (A) (C)
max(, (B), (C)).
E
for
E
.
Check
that
is
a
measure.
Write
(A
,
)
for
the measure algebra of n and
n
n
n=0 n n
for the measure algebra of . Show that (C) max(, sup
(C, )
(A
)).
n
nN
(e) Let (X, , ) and (Y, T, ) be -nite measure spaces, and the product measure on X Y . Show
that is Maharam-type-homogeneous i one of , 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, , ) be a probability space of Maharam type , and I any set of cardinal at least max(, ).
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 , i , i )iiI be an innite family of probability spaces, with
product (X, , ). Let i be the Maharam type of i for each i; set = max(#(I), supiI i ). Show that
either is Maharam-type-homogeneous, with Maharam type , or there are < , Xi i such that
P
iI i (Xi \ Xi ) < and the Maharam type of the subspace measure on Xi is at most for every i I
162
Liftings
Chapter 34
The lifting theorem
Whenever we have a surjective homomorphism : P Q, where P and Q are mathematical structures,
we can ask whether there is a right inverse of , a homomorphism : Q P such that 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, , ) (in fact, for any complete strictly
localizable space of non-zero measure) the canonical homomorphism from to the measure algebra of 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 rst 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 diculties 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 dicult questions concerning the extent to which liftings can be
required to have special properties, reecting 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 diculties. I give the currently
known results in 346.
P
P (i) Given a lifting : A , the formula denes a Boolean homomorphism : such that
= 0 = ,
(E E) = E (E ) = 0 E ,
so that is a lifting. (ii) Given a lifting : , the kernel of includes {E : E = 0}, so there is a
Boolean homomorphism : A such that E = E for every E (3A2G), and now
341E
163
(E ) = (E) = E
for every E , so is a lifting. Q
Q
I suppose that the word lifting applies most naturally to functions from A to ; but for applications in
measure theory the other type of lifting is used at least equally often.
(b) Note that if : is a lifting then 2 = . P
P For any E ,
2 E E = (E E) = . Q
Q
b in A, and is a
b.
(EB(x,))
B(x,)
= 1}.
(Here B(x, ) is the closed ball with centre x and radius .) Then is a lower density for ; we may call it
lower Lebesgue density. P
P (You may prefer at rst to suppose that r = 1, so that B(x, ) = [x , x + ]
and B(x, ) = 2.) By 261Db (or 223B, for the one-dimensional case) EE is negligible for every E;
in particular, E for every E . If EF is negligible, then (E B(x, )) = (F B(x, )) for
every x and , so E = F . If E F , then (E B(x, )) (F B(x, )) for every x, , so E F ;
consequently (E F ) E F for all E, F . If E, F and x E F , then
164
341E
Liftings
for every , so
(EF B(x,))
B(x,)
(EB(x,))
B(x,)
(F B(x,))
B(x,)
11
w = upr(1 \ e, B)
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, a ,
1 \ e w,
b b
e v.
b w = b w.
a \ w = a \ w,
1 \ (a a ) and
b \ v = b \ v.
(1 ((a e) (b \ e))) =
= (a e) (b \ e).
So (1 c) = c for every c B1 .
(ii)
(iii) If a, a , b, b B, then
341G
165
= 1 ((a a e) (b b \ e))
Q
So 1 (c c ) = 1 (c) 1 (c ) for all c, c B1 . Q
(d) If a B, then
1 (a) = 1 ((a e) (a \ e))
= (a) E (a) \ E = a.
n = {E : E , E Bn },
and set
= {E : E , E B}.
Then (because all the Bn , B are -subalgebras of A, and E 7 E is sequentially order-continuous) all
the n , are
is just the -algebra of subsets of X
S -subalgebras of . We need to know that S
P Because is a -algebra including nN n , . On the other hand,
generated by nN n . P
g =
E an
Ga = (E Ga ) =
(E a)
for every E n . As remarked in 233Db, such a function gan can always be found, and moreover we may
take it to be n -measurable and dened everywhere on X. Now if a B, limn gan (x) exists and is
equal to Ga (x) for almost every x. P
P By Levys martingale theorem (275I), limn
S gan is dened almost
everywhere and is a conditional expectation of Ga on the -algebra generated by nN n . As observed
166
341G
Liftings
in (a), this is just ; and as Ga is itself -measurable, it is also a conditional expectation of itself on
, and must be equal almost everywhere to limn gan . Q
Q
(d) For a B, k 1, n N set
k1
nN
mn
km (a).
H
The rest of the proof is devoted to showing that : B 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 0 = .
P Ga \ Gb is negligible, gan gbn almost everywhere for every n, every
(f ) If a b in B, then a b. P
kn (a) H
kn (b) for every n and k, and a b. Q
Hkn (a) \ Hkn (b) is negligible, H
Q
(g) If a, b B then (a b) = a b. P
P Gab a.e. Ga + Gb 1 so gab,n a.e. gan + gbn 1 for
every n. Accordingly
Hk+1,n (a) Hk+1,n (b) \ Hkn (a b)
is negligible, and (because n is a lower density)
kn (a b) n (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 a b, then, for any k 1, there are n1 , n2 N such that
T
k+1,m (a), x T
x mn1 H
mn2 Hk+1,m (b).
But this means that
mmax(n1 ,n2 )
km (a b).
H
aVa
k1,nN
kn (a)
Hkn (a)H
for k 1, m n (this is where I use the hypothesis that m+1 extends m for every m); and
a =
\ [ \
km (a)
H
k1 rN mr
\ [ \
k1 rn mr
km (a) =
H
\ [
Q
n a = n a. Q
k1 rn
341I
167
proof : Part A I deal rst with the case of probability spaces. Let (X, , ) be a probability space, and
(A,
) its measure algebra.
(a) Set = #(A) and enumerate A as ha i< . For let A be the closed subalgebra of A generated
by {a : < }. I seek to dene a lower density : A as the last of a family h i , where : A
is a partial lower density for each . The inductive hypothesis will be that extends whenever .
To start the induction, we have A0 = {0, 1}, 0 0 = , 0 1 = X.
(b) Inductive step to a successor ordinal Given a succesor ordinal , express it as + 1; we are
supposing that : A has been dened. Now A is the subalgebra of A generated by A {a }
(because this is a closed subalgebra, by 323K). So 341F tells us that can be extended to a partial lower
density with domain A .
(c) Inductive step to a non-zero limit ordinal of countable cofinality In this case, there is a strictly
increasing sequence h(n)inN with supremum . Applying 341G with Bn = S
A(n) , we see that there is a
partial lower density , with domain the closed subalgebra B generated by nN A(n) , extending every
(n) . Now A(n) A for every , so B A ; but also, if < , there is an n N such that < (n),
so that a A(n) B; as is arbitrary, A B and A = B. Again, if < , there is an n such that
(n), so that (n) extends and extends . Thus the induction continues.
S
P Because
(d) Inductive step to a limit ordinal of uncountable cofinality In this case, A = < A . P
A is ccc, every member a of A must be in the closed subalgebra of A generated by some countable subset A
of {a : < } (331Gd-e). Now A can be expressed as {a : I} for some countable I . As I cannot
be conal with , there is a < such that < for every I, so that A A and a A . Q
Q
But now, because extends whenever < , we have a function : A dened by writing
a = a whenever < and a A . Because the family {A : < } is totally ordered and every is a
partial lower density, is a partial lower density.
Thus the induction proceeds when is a limit ordinal of uncountable conality.
(e) The induction stops when we reach : A , which is a lower density such that 1 = X. Setting
E = E , is a lower density such that X = X.
Part B The general case of a strictly localizable measure space follows easily. First, if X = 0, then
A = {0} and we can set 0 = . Second, if is totally nite but not zero, we can replace it by , where
E = E/X for every E ; a lower density for is also a lower density for . Third, if is not totally
nite, let hXS
i iiI be a decomposition of X (211E). There is surely some j such that Xj > 0; replacing
Xj by Xj {Xi : i I, Xi = 0}, we may assume that Xi > 0 for every i I. For each i I, let
i : i i be a lower density of i , where i = PXi and i = i , such that i Xi = Xi . Then it is
easy to check that we have a lower density : given by setting
S
E = iI i (E Xi )
for every E , and that X = 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, , ) be a complete measure space with measure algebra A.
(a) Suppose that : A is a lower density and 1 : A PX is a function such that 1 0 = ,
1 (a b) = 1 a 1 b for all a, b A and 1 a a for all a A. Then 1 is a lower density. If 1 is a
Boolean homomorphism, it is a lifting.
(b) Suppose that : is a lower density and 1 : PX is a function such that 1 E = 1 F
whenever EF is negligible, 1 = , 1 (E F ) = 1 E 1 F for all E, F and 1 E E for all
E . Then 1 is a lower density. If 1 is a Boolean homomorphism, it is a lifting.
proof (a) All I have to check is that 1 a and (1 a) = a for every a A. But
a 1 a,
So
(1 \ a) 1 (1 \ a),
1 a 1 (1 \ a) = 1 0 = .
168
Liftings
341I
a 1 a X \ (1 \ a).
Since
(a) = a = (X \ (1 \ a)) ,
and is complete, 1 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 dened by and 1
and using (a).
341J Proposition Let (X, , ) be a complete measure space such that X > 0, and A its measure
algebra.
(a) If : A is any lower density, there is a lifting : A such that a a for every a A.
(b) If : is any lower density, there is a lifting : such that E E for every E .
proof (a) For each x 1, set
Ix = {a : a A, x (1 \ a)}.
for every a A. It is easy to check that, because every x is a surjective Boolean homomorphism, is a
Boolean homomorphism. Now for any a A, x X,
x a = 1 \ a Ix = x (1 \ a) = 0 = x a = 1 = x a.
(b) Repeat the argument above, or apply it, dening by setting (E ) = E for every E , and
by setting E = (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 Carlesons
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 rst 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 its worth doing, for a variety of reasons, but there is no signicant
idea needed, and if for instance the result were valid only for -nite 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 transnite induction written out in the way I have presented it in 341H above.
Essentially the same proof can be presented as an application of Zorns 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
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
Maharams 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 rst and last are easier than the second. Seeking the one-step extension of : B to
1 : B1 , 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 1 is not fully specied by the value of 1 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 of countable conality dicult, because A will ordinarily be larger than < A . But it is
absolutely essential in the one-step extensions.
Because we are dealing with a ccc algebra A, the
S requirement that the A should be closed is not a
problem when cf() is uncountable, since in this case < A 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 when cf() = , which is 341G. Here we actually need
some measure theory, and a particularly striking bit. (You will see that the measure , as opposed to the
algebras 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 dene a in terms 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 ; let me emphasize that this is a crude
application of the axiom of choice, and that the dierent 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 Ga on n . Once again, the choices are not coordinated; but
the martingale theorem assures us that ga = limn gan is dened almost everywhere, and is equal almost
everywhere to Ga 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
(E a) : n R. Now the
gan , like the Ga , are not uniquely dened. But they are dened up to a negligible set, so that any alternative
=a.e. gan . This means that the sets Hkn (a) = {x : gan (x) 1 2k } are
would have gan
functions gan
kn (a) = n (Hkn (a)) are uniquely dened. I
also dened 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 passing that I could have dened
kn (a) without mentioning the gan ; in fact
the H
kn (a) = n (sup{c : c Bn ,
(a d) 1 2k
d whenever d Bn , d
H
c}).
because gan Ga a.e., Va Ga is negligible and Va = a for every a A . The rest of the argument
amounts to checking that a 7 Va will serve for .
(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 nd a Boolean homomorphism : A PX such that a a for
every a; this corresponds just to extending the ideals Ix = {a : x (1 \ a)} to maximal ideals (and giving
a moments thought to x X \ 1). In order to ensure that a and (a) = a, we have to observe that
a is sandwiched between a and X \ (1 \ a), which dier by a negligible set; so that if 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, , I) where X is a set, is a -algebra of subsets of X, and I is a -ideal of . (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 satised: (X, , I) is complete in the sense that every subset of any member of
I belongs to (and therefore to I), and I is 1 -saturated in 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 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 conality remains straightforward. But in the
inductive step to an ordinal of countable conality, we nd that in 341G we get no help from assuming that
the n are actually liftings; we are still led to to a lower density . 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, , ) be a complete locally determined space with X > 0. Then it has a lifting i it
has a lower density i it is strictly localizable.
proof If (X, , ) is strictly localizable then it has a lifting, by 341K. A lifting is already a lower density,
and if (X, , ) 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 : A be a lifting, where A is the measure algebra of (X, , ). Let C be a partition of unity in
A consisting of elements of nite measure (322Ea). Set A = {c : c C}. Because C is disjoint, 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 satised, and (X, , ) 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
171
(b) If (X, , ) is complete and locally determined, then a function f : X Z satises the conditions of
(a) i () it is inverse-measure-preserving () the homomorphism it induces between the measure algebras
of and is the canonical isomorphism dened by the construction of Z.
proof Recall that T is just the set {b
aM : a A, M Z is meager}, and that (b
aM ) =
a for all such
a, M ; while the canonical isomorphism between A and the measure algebra of is dened by the formula
F = a whenever F T, a A and F b
a is meager
(341K).
(a) If : A is any Boolean homomorphism, then for every x X we have a surjective Boolean
homomorphism f (x) : A Z2 dened by saying that f (x)(a) = 1 if x a, 0 otherwise. f is a function
from X to Z. We can recover from f by the formula
a].
a = {x : f (x)(a) = 1} = {x : f (x) b
a} = f1 [b
a] and, if is a lifting,
So f1 [b
a]) = (a) = a.
(f1 [b
for every a A.
Similarly, given a function f : X Z with this property, then we can set a = f 1 [b
a] for every a A to
obtain a lifting : A ; and of course we now have
so f = f .
f (x)(a) = 1 f (x) b
a x a,
172
Liftings
341P
As F is arbitrary, f is inverse-measure-preserving. Q
Q
It follows at once that for any F T,
f 1 [F ] = a = F
algebra of
(322Da). Let : A be a lifting, and f : X Z the corresponding function. If E then
= , then f
E = b
a where a = E , so Ef 1 [E ] = EE is negligible. If is itself complete, so that
is inverse-measure-preserving, by 341Pb.
341X Basic exercises (a) Let (X, , ) be a measure space and : a function. Show that is
a lifting i it is a lower density and E (X \ E) = X for every E .
> (b) Let be the usual measure on X = {0, 1}N , and its domain. For x X and n N set
Un (x) = {y : y X, yn = xn}. For E set E = {x : limn 2n (E Un (x)) = 1}. Show that is a
lower density of (X, , ).
> (c) Let P be the set of all lower densities of a complete measure space (X, , ), with measure algebra
A, ordered by saying that if a a for every a A. Show that any non-empty totally ordered
subset of P has an upper bound in P . Show that if P and a A and x X \ (a (1 \ a)), then
: A is a lower density, where b = b {x} if either a b or there is a c A such that x c and
a c b, and b = b otherwise. Hence prove 341J.
(d) Let (X, , ) and (Y, T, ) 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 to that of
is an isomorphism. Show that for every lifting of (Y, T, ) we have a corresponding lifting of (X, , )
dened uniquely by the formula
(f 1 [F ]) = f 1 [F ] for every F T.
341 Notes
173
(e) Let (X, , ) be a measure space, and write L () for the linear space of all bounded -measurable
functions from X to R. Show that for any lifting : of there is a unique linear operator
T : L () L () such that T (E) = (E) for every E and T u 0 in L () whenever u 0 in
L (). Show that (i) (T u) = u and supxX |(T u)(x)| = kuk for every u L () (ii) T (u v) = T u T v
for all u, v L ().
> (f ) Let A be a Boolean algebra, I an ideal of A, and B a countable subalgebra of the quotient algebra
A/I. Show that there is a Boolean homomorphism : B A such that (b) = b for every b B. (Hint:
let hbn inN run over B; let Bn be the subalgebra of B generated by {bi : i < n}; given Bn , show that
there is an an A such that an = bn and b an b whenever b , b Bn and b bn b .)
341Y Further exercises (a) Let X be a set, an algebra of subsets of X and I an ideal of ; let
A be the quotient Boolean algebra /I. We say that a function : A is a lifting if it is a Boolean
homomorphism and (a) = a for every a A, and that : A is a lower density if 0 = ,
(a b) = a b for all a, b A, and (a) = a for every a A.
Show that if (X, , I) is complete in the sense that F whenever F E I, and if X
/ I, and
: A is a lower density, then there is a lifting : A such that a a for every a A.
(b) Let X be a non-empty Baire space, Bb the -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 from B/M
to Bb such that G G
for every open G X. (Hint: in 341Ya, set (G ) = G for every regular open set G.)
(c) Let (X, , ) be a Maharam-type-homogeneous probability space with Maharam type . Let
Ba be the Baire -algebra of Y = {0, 1} , that is, the -algebra of subsets of Y generated by the family
{{x : x() = 1} : < }, and let be the restriction to Ba of the usual measure on {0, 1} . Show that there
is an inverse-measure-preserving function f : X Y which induces an isomorphism between their measure
algebras.
(d) Let (X, , ) be a complete Maharam-type-homogeneous probability space with Maharam type ,
and give Y = {0, 1} its usual measure. Show that there is an inverse-measure-preserving function f : X Y
which induces an isomorphism between their measure algebras.
(e) Let (X, , ) be a semi-nite 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 () L1strict such that (i)
(T u) = u for every u L1 () (ii) T u T v whenever u v in L1 (). (Hint: Suppose rst that 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 satises (i) then there is an x {0, 1}N such that T (f )(x) = f (x) for every f F ; nd a sequence
hfn inN in F such that {fn : n N} is bounded above in L1 () but supnN 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, , ) with no lifting?
(b) Set = 3 . (There is a reason for taking 3 here; see Volume 5, when it appears, or Fremlin 89.)
Let Ba be the Baire -algebra of X = {0, 1} (as in 341Yc), and let be the restriction to Ba of the usual
measure on {0, 1} . Does (X, Ba, ) 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 rst published proof is that of Maharam 58. That argument is based
on Maharams theorem, 341Xd and 341Yd, which show that it is enough to nd liftings for every {0, 1} ;
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.
174
341 Notes
The lifting theorem and Maharams 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 rst 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
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 diculty 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.
for every E .
Of course is inner regular with respect to K i it is inner regular with respect to K .
T
(b) A family K of sets is a compact class if K 6= whenever K K has the nite 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, , ), or a measure , is compact if 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, is a compact measure whenever
X = 0.
(d) A measure space (X, , ), or a measure , is locally compact if the subspace measure E is
compact whenever E and E < .
Remark I ought to point out that the original denitions 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 () in 343B(ii)-(iii).
For examples of compact measure spaces see 342J.
342B I prepare the ground with some straightforward lemmas.
Lemma Let (X, , ) be a measure space, and K a set such that whenever E and E > 0 there
is a K K such that K E and K > 0. Let E .
S
(a) There is a countable disjoint
set K1 K such that K E for every K K1 and ( K1 ) = E.
S
(b) If E < then (E \ K1 ) = 0.
342F
175
(c) In any case, there is for any < E a nite disjoint K0 K such that K E for every K K0 and
S
( K0 ) .
176
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-nite measure space is compact i its completion is compact i 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, , ) be a compact measure space, and E . If K is a compact class such that 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 E is inner regular with respect to KE .
(b) Let (X, , ) be a compact measure space. Write (X, ,
) for either the completion or the c.l.d.
version of (X, , ). Let K PX be a compact class such that is inner regular with respect to K. Then
and <
and E > ; if
is the c.l.d. version of , we may take E to be nite. There is a K K such that
Q
K E and K . Now
K = K and K E and K K .
Q
(c) Now suppose that (X, , ) is semi-nite; again write (X, ,
) for either its completion or its c.l.d.
version. We already know that if is compact, so is
. IfT
is compact, let K PX be a compact class
such that
is inner regular with respect to K. Set K = { K : =
6 K K}; then K is a compact class
n=0
(251C). Set
W = (E F ) \
Then W W , and
nN
En F n =
nN ((E
(F \ Fn )) ((E \ En ) F )).
S
P
((E F ) \ W ) ( nN En Fn ) n=0 En Fn < (E F ) ,
so W > .
Set = 41 (W )/(1 + E + F ). For each n, we can nd closed measurable sets Kn , Kn X and
Ln , Ln Y such that
Kn E,
(E \ Kn ) 2n ,
342H
Ln F \ Fn ,
((F \ Fn ) \ Ln ) 2n ,
Kn E \ En ,
((E \ En ) \ Kn ) 2n ,
Ln F ,
Set
V =
Now
W \ V
nN (Kn
nN
177
(F \ Ln ) 2n .
Ln ) (Kn Ln ) W W .
((E \ Kn ) F ) (E ((F \ Fn ) \ Ln ))
(((E \ En ) \ Kn ) F ) (E (F \ Ln )),
so
(W \ V )
n=0
n=0
(E \ Kn ) F + E ((F \ Fn ) \ Ln )
+ ((E \ En ) \ Kn ) F + E (F \ Ln )
2n (2E + 2F ) W ,
and V . But V is a countable intersection of nite unions of products of closed measurable sets, so is
itself a closed measurable set, and belongs to K . Q
Q
Accordingly the product topology on X Y witnesses that is a compact measure.
(f ) The same method works. In detail: let h(Xi , i , i )iiI be a family of compact probability spaces,
with product (X, , ). For each i, let Ti be a topology on Xi such that Xi is compact and i is inner
regular with respect to the closed sets. Give X the product topology; this is compact. If W S
and > 0,
let hCn inN be a sequence of measurable cylinders (in the sense of 254A) such that X \ W nN Cn and
P
Q
iI Eni where Eni i for each i and Jn = {i : Eni 6= Xi }
n=0 Cn (X \ W ) + . Express each Cn as
is nite. For n N, i Jn set ni = 2n /(1 + #(Jn )). Choose closed measurable sets Kni Xi \ Eni such
that i ((Xi \ Eni ) \ Kni ) ni for n N, i Jn . For each n N, set
S
Vn = iJn {x : x X, x(i) Kni },
so that Vn is a closed measurable subset of X. Observe that
P
P
(X \ (Vn Cn )) iJn {x : x(i) Xi \ (Kni Eni )} iJn ni 2n .
T
Now set V = nN Vn ; then V is again a closed measurable set, and
S
X \ V nN Cn (X \ (Cn Vn ))
n=0
Cn + 2n 1 W + + 2,
so V W 3. As W and are arbitrary, 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, , ) be a measure space. Suppose that whenever E and E > 0 there is an F such
that F E, F > 0 and the subspace measure on F is compact. Then is locally compact.
proof (a) This is immediate from 342Ga and the denition of locally compact measure space.
178
342H
(b) Suppose that (X, , ) is a strictly localizable locally compact measure space. Let hXi iiI be a
decomposition of X, and for each i I let i be the subspace measure on Xi . Then i is compact. Now
can be identied with the direct sum of the i , so itself is compact, by 342Gd.
(c) Write F for the set of measurable sets F X such that the subspace measures F are compact.
Take E with E < .
S By 342Bb, there is a countable disjoint family hFi iiI in F such that Fi E
for each
i,
and
F
=
E
\
iI Fi is negligible; now this means that F F (342Ac), so we may take it that
S
E = iI Fi . In this case E is isomorphic to the direct sum of the measures Fi and is compact. As E is
arbitrary, 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 i its completion is locally compact i 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, , ) is locally compact, and E , and F E is a measurable set of nite
measure for the subspace measure on E, then F and F < , so the subspace measure on F is
compact.
(b) Let (X, , ) be a measure space, and write (X, ,
) for either its completion or its c.l.d. version.
(i) Suppose that is locally compact, and that
F < . Then there is an E such that E F
and E =
F . Let E be the subspace measure on E induced by the measure ; then we are assuming that
E is compact. Let K PE be a compact class such that E is inner regular with respect to K. Then, as
in the proof of 342Gb, the subspace measure
F on F induced by
is also inner regular with respect to K,
so
F is compact; as F is arbitrary,
is locally compact.
(ii) Now suppose that
is locally compact, and that E < . Then the subspace measure
E is
compact. But this is just the completion of the subspace measure E , so E is compact, by 342Gc; as E is
arbitrary, is locally compact.
(c) Put (a) and 342Hc together.
(d) Let (X, , ) and (Y, T, ) be locally compact measure spaces, with product (X Y, , ). If W
and W > 0, there are E , F T such that E < , F < and (W (E F )) > 0. Now the
subspace measure EF induced by on E F is just the product of the subspace measures (251P(ii-),
so is compact, and the subspace measure W (EF ) is therefore again compact, by 342Ga. By 342Hc, this
is enough to show that is locally compact.
342J Examples It is time I listed some examples of compact measure spaces.
(a) Lebesgue measure on R r is compact. (Let K be the family of subsets of R r which are compact for
the usual topology. By 134Fb, Lebesgue measure is inner regular with respect to K.)
(b) Similarly, any Radon measure on R r (256A) is compact.
(c) If (A,
) is any semi-nite measure algebra, the standard measure on its Stone space Z is compact.
(By 322Qa, 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, , ) be a measure space. Then (X, , ), or , is perfect if whenever f : X R is
measurable, E and E > 0, then there is a compact set K f [E] such that f 1 [K] > 0.
*342N
179
so K > 0.
The point is that f K is continuous. P
P For any q Q, {x : x K, f (x) q} = K Kq and
{x : x K, f (x) > q} = K Kq . If H R is open and x K f 1 [H], take q, q Q such that
f (x) ]q, q ] H; then G = K \ (Kq Kq ) 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
f 1 [f [K]] K > 0. As f and E are arbitrary, 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 the countable-cocountable measure on X (211R). Then 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 = {
S : R, {x : x X, f (x) r} is negligible}. Then A whenever A. Since
X = nN {x : f (x) n}, there is some n such that n
/ A, in which case A is bounded above by
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 = sup A is dened in R. Now for any k N,
{x : f (x) 2k } is negligible, so {x : f (x) < } is negligible. Also, for any k, {x : f (x) + 2k } is
non-negligible, so {x : f (x) > 2k } must be negligible; accordingly, {x : f (x) > } is negligible. But this
means that {x : f (x) = } is conegligible and has measure 1. Thus we have a compact set K = {} such
that f 1 [K] = 1, and must belong to f [E]. As f and E are arbitrary, is perfect.
(b) is not compact. P
P?? Suppose, if possible, that K PX is a compact class such that 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 Kx > 0, that is, Kx is conegligible. But this means that {Kx : x X} must have the nite intersection
property; as it also has empty intersection, K cannot be a compact class. X
XQ
Q
(c) Because is totally nite, 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 hf iC in
X = {0, 1}I such that for every D C there is an i I such that D = { : f (i) = 1}; moreover, #(C) > c.
The -algebra is the family of sets E X such that for every there is a countable set J I such that
{x : xJ = f J} is a subset of either E or X \ E; and for E , E is #({ : f E}) if this is nite,
otherwise. Note that any subset of X determined by a countable set of coordinates belongs to .
For each C, let i I be such that f (i ) = 1, f (i ) = 0 for 6= . (In 216E I took I to be PC,
and i would be {}.) Set
Y = {x : x X, { : C, x(i ) = 1} is nite}.
180
*342N
Give Y its subspace measure Y with domain Y . Then Y is complete, locally determined and localizable
(214Id). Note that f Y for every C.
S
Then each K is a compact class, and members of dierent K s are disjoint, so K = D K is a compact
class.
Now suppose that H belongs to the subpsace -algebra F and F H > 0. Then there is a D such
that f H, so that H G K F and F (H G ) > 0. By 342E, this is enough to show that F is
compact. As F is arbitrary, Y is locally compact. Q
Q
(c) Y is not compact. P
P?? Suppose, if possible, that Y is inner regular with respect to a compact class
K PY . For each C set G = {x : x X, x(i ) = 1}, so that f G and Y (G Y ) = 1.
There must therefore be a K K such that K G Y and Y K = 1 (since Y takes no value in ]0, 1[).
Express K as Y E , where E , and let J I be a countable set such that
E {x : x X, xJ = f J }.
At this point I call on the full strength of 2A1P. There is a set B C, of cardinal greater than c, such
that f J J = f J J for all , B. But this means that, for any nite set D B, we can dene
x X by setting
x(i) = f (i) if D, i J ,
[
= 0 if i I \
J .
D
What this shows is that {K : B} has the nite intersection property. It must therefore have
non-empty intersection; say
T
T
y B K B G .
But now we have a member y of Y such that { : y(i ) = 1} B is innite, contrary to the denition of Y .
X
XQ
Q
342X Basic exercises > (a) Show that a measure space (X, , ) is semi-nite i is inner regular
with respect to {E : E < }.
(b) Find a proof of 342B based on 215A.
(c) Let (X, , ) be a locally compact semi-nite measure space in which all singleton sets are negligible.
Show that it is atomless.
(d) Let (X, , ) be a measure space, and an indenite-integral measure over (234B). Show that is
compact, or locally compact, if is. (Hint: if K satises the conditions of 342E with respect to , then it
satises them for .)
(e) Let f : R R be any non-decreasing function, and f the corresponding Lebesgue-Stieltjes measure.
Show that f is compact. (Hint: 256Xg.)
(f ) Let be Lebesgue measure on [0, 1], the countable-cocountable measure on [0, 1], and their c.l.d.
product. Show that 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, , ), a set Y and a function f : X Y such
that the image measure f 1 is not compact. (ii) Give an example of a compact probability space (X, , )
and a -subalgebra T of such that (X, T, T) is not compact. (Hint: 342Xf.)
(h) Let (X, , ) be a perfect measure space, and f : X R a measurable function. Show that the
image measure f 1 is inner regular with respect to the compact subsets of R, so is a compact measure.
(i) Let (X, , ) be a -nite measure space. Show that it is perfect i 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, , ) be a complete totally nite perfect measure space and f : X R a measurable function.
Show that the image measure f 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, , ) is a perfect measure space. (i) Show that if (Y, T, ) is a measure space,
and f : X Y is a function such that f 1 [F ] for every F T and f 1 [F ] is -negligible for every
-negligible set F , then (Y, T, ) is perfect. (ii) Show that if T is a -subalgebra of then (X, T, T) is
perfect.
(l) Let (X, , ) be a perfect measure space such that is the -algebra generated
P by a sequence of sets.
Show that is compact. (Hint: if is generated by {En : n N}, set f = n=0 3n En and consider
{f 1 [K] : K f [X] is compact}.)
(m) Let (X, , ) be a semi-nite measure space. Show that is perfect i T is compact for every
countably generated -subalgebra T of .
(n) Show that (i) a measurable subspace of a perfect measure space is perfect (ii) a semi-nite measure
space is perfect i all its totally nite 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 i 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 indenite-integral measure over a perfect measure is
perfect.
(o) Let be Lebesgue measure on R, let A be a subset of R, and let A be the subspace measure. Show
that A is compact i it is perfect i A is Lebesgue measurable. (Hint: if A is perfect, consider the image
measure h1
A on R, where h(x) = x for x A.)
342Y Further exercises (a) Show that the space (X, , ) 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.)
(c) Let U be a Banach space such that there is a linear operator T : U U , of norm at most 1, such
that T u
= u for every u U , writing u
for the member of U corresponding to u. Show that the family of
closed balls in U is a compact class.
342 Notes and comments The terminology I nd 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,
182
342 Notes
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.)
I suppose that compact and perfect measure spaces look reasonably natural objects to study; they oer
to illuminate one of the basic properties of Radon measures, the fact that (at least for totally nite 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-nite 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 rst reading.
As with any new classication of measure spaces, it is worth nding 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 classication 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, , ) is compact, we need only the
structure (X, , N ), where N is the -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 denition 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.
343B
Realization of homomorphisms
183
If G is any element of T,
there is an
f 1 [F ] is -negligible, so f 1 [E] is negligible, therefore belongs to .
F T such that GF is negligible, so that
f 1 [G] = f 1 [F ]f 1 [GF ] ,
and
f 1 [G] = f 1 [F ] = F = G . Q
Q
343B Theorem Let (X, , ) be a non-empty semi-nite measure space, and (A,
) its measure algebra. Let (Z, , ) be the Stone space of (A,
); for E 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, , ) is locally compact in the sense of 342Ad.
(ii) There is a family K such that () whenever TE and E > 0 there is a K K such that
KE
K > 0 () whenever K K is such that ( K0 ) > 0 for every non-empty nite set K0 K ,
T and
then K 6= .
(iii) ThereTis a family K such that () is inner regular withTrespect to K () whenever K K is
such that ( K0 ) > 0 for every non-empty nite set K0 K , then K 6= .
(iv) There is a function f : Z X such that f 1 [E]E is negligible for every E .
(v) Whenever (Y, T, ) is a complete strictly localizable measure space, with measure algebra B, and
: A B is an order-continuous Boolean homomorphism, then there is a g : Y X representing .
(vi) Whenever (Y, T, ) is a complete strictly localizable measure space, with measure algebra B, and
: A B is an order-continuous measure-preserving Boolean homomorphism, then there is a g : Y X
representing .
proof (a)(i)(ii) Because is semi-nite, there is a partition of unity hai iiI in A such that
ai < for
each i. For each i I, let Ei be such that Ei = ai . Then the subspace measure Ei on Ei is
S compact;
is
inner
regular
with
respect
to
K
.
Set
K
=
let Ki PEi be
a
compact
class
such
that
i
Ei
iI Ki . If
T
K,
then
K
K
for
some
i,
and
surely
has the
0
i
T
nite intersection property, so K 6= ; thus K satises () of condition (ii). And if E , E > 0 then
there must be some i I such that Ei ai 6= 0, that is, (E Ei ) > 0, in which case there is a K Ki K
such that K E Ei and K > 0; so that K satises condition ().
(b)(ii)(iii) Suppose that K witnesses that (ii) is true. If X = 0 then K already witnesses that
(iii) is true, so we need consider only the case X > 0. Set L = {K0 . . . Kn : K0 , . . . , Kn K}. Then
LT
witnesses that (iii) is true. P
P By 342Ba, is inner regular with respect to L. Let L L be such that
( L0 ) > 0 for every non-empty nite L0 L . Then
T
F0 = {A : A X, there is a nite L0 L such that X L0 \ A is negligible}
184
343B
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 denes 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 . 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]E 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 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 denition of (321J-321K), M is -negligible, so f 1 [E]E M is negligible, as required.
(d)(iv)(v) Now assume that f : Z X witnesses (iv), and let (Y, T, ) be a complete strictly localizable measure space, with measure algebra B, and : A B an order-continuous Boolean homomorphism.
If Y = 0 then any function from Y to X will represent , so we may suppose that Y > 0. Write W for
the Stone space of B. Then we have a continuous function : W Z such that 1 [b
a] =
ca for every
a A (312P), and 1 [M ] is nowhere dense in W for every nowhere dense M Z (313R). It follows that
1 [M ] is meager for every meager M Z, that is, 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 h1 [bb] = b for every b B.
Consider g = f h : Y X.
If E , set a = E A, so that E = b
a Z, and M = f 1 [E]E is -negligible; consequently
1
[M ] is negligible in W . Because h is inverse-measure-preserving,
g 1 [E]h1 [1 [E ]] = h1 [1 [f 1 [E]]]h1 [1 [E ]] = h1 [1 [M ]]
is negligible. But 1 [E ] =
ca, so
g 1 [E] = h1 [1 [E ]] = a.
where Fn+1 ;
so there is an open-and-closed set Vn f 1 [Fn ] Fn with Vn > . Express Vn as Fn+1
1
g
[F
].
Continue.
n+1
n
n
T
T
At the end of the induction, set K = nN Fn and L = nN Fn . Because Fn+1 \ Fn g 1 [Fn ] \ Fn
is negligible for each n, K = limn Fn , while K F and L is surely compact. We have
T
T
L nN Vn nN f 1 [Fn ] = f 1 [K],
343D
so f [L] K. Also
Realization of homomorphisms
nN
Fn+1
nN
185
g 1 [Fn ] = g 1 [L].
( ) Next,
P Suppose that K K has the nite intersection property. If K = ,
T K is a compact class. P
of course K 6= ; suppose that K is non-empty. Let L be the family of closed sets L Z such that
( ) Now consider the general case. Take any E of nite measure. If E = then surely the subspace
measure E is compact. Otherwise, we can identify the measure algebra of E 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
dene 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 f1 [F ]F is negligible for every F E , that is, f represents the canonical
isomorphism between the measure algebras of E and the subspace measure E on E . But this means
that condition (iv) is true of E , so E is compact, by ()-() above. As E is arbitrary, is locally compact.
This completes the proof.
343C Examples (a) Let be an innite cardinal. We know that the usual measure on {0, 1}
is compact (342Jd). It follows that if (X, , ) is any complete probability space such that the measure
algebra B of can be embedded as a subalgebra of the measure algebra A of , there is an inversemeasure-preserving function from X to {0, 1} . By 332P, this is so i every non-zero principal ideal of A has
Maharam type at least . Of course this does not depend in any way on the results of the present chapter.
If B can be embedded in A, there must be a stochastically independent family hE i< of sets of measure
1
2 ; now we get a map h : X {0, 1} by saying that h(x)() = 1 i x E , which by 254G is inversemeasure-preserving.
(b) In particular, if 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, , ) is any complete atomless totally nite measure space, there is an inversemeasure-preserving function from X to the interval [0, X] endowed with Lebesgue measure. (If X > 0,
apply (b) to the normalized measure (X)1 ; or argue directly from 343B, using the fact that Lebesgue
measure on [0, X] 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.
Definition A measure space (X, , ) is countably separated if there is a countable set A 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, ) rather than of (X, , ).)
186
343E
343E Lemma A measure space (X, , ) is countably separated i there is an injective measurable
function from X to R.
proof If (X, , ) is countably separated, let A be a countable set separating the points of X. Let
hEn inN be a sequence running over A {}. Set
P
f = n=0 3n En : 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) 3n + i<n 3i Ei (x) > i>n 3i + i<n 3i Ei (y) f (y).)
Remark The construction of the function f from the sequence hEn inN in the proof above is a standard
trick; such f are sometimes called Marczewski functionals.
343F Proposition Let (X, , ) be a countably separated measure space and (Y, T, ) any measure space.
Let f , g : Y X be two functions such that f 1 [E] and g 1 [E] both belong to T, and f 1 [E]g 1 [E] is
-negligible, for every E . Then f = g -almost everywhere, and {y : y Y, f (y) 6= g(y)} is measurable
as well as negligible.
proof Let A be a countable set separating the points of X. Then
S
{y : f (y) 6= g(y)} = EA f 1 [E]g 1 [E]
is measurable and negligible.
343G Corollary If, in 343B, (X, , ) is countably separated, then the functions f : Y X of 343B(v)(vi) are almost uniquely dened in the sense that if f , g both represent the same homomorphism from A to
B then f =a.e. g.
343H Examples Leading examples of countably separated spaces are
(i) R (take A = {], q] : q Q});
(ii) {0, 1}N (take A = {En : n N}, where En = {x : x(n) = 1});
(iii) subspaces (measurable or not) of countably separated spaces;
(iv) nite products of countably separated spaces;
(v) countable products of countably separated probability spaces;
(vi) completions and c.l.d. versions of countably separated spaces.
As soon as we move away from these elementary ideas, however, some interesting diculties arise.
343I Example Let c be the usual measure on X = {0, 1}c , where c = #(R), and Tc its domain. Then
there is a function f : X X such that f (x) 6= x for every x X, but Ef 1 [E] is negligible for every
E Tc . P
P The set c \ is still of cardinal c, so there is an injection h : {0, 1} c \ . (As usual, I am
identifying the cardinal number c with the corresponding initial ordinal. But if you prefer to argue without
the full axiom of choice, you can express all the same ideas with R in the place of c and N in the place of
.) For x X, set
f (x)() = 1 x() if = h(x),
= x() otherwise .
Evidently f (x) 6= x for every x. If E X is measurable, then we can nd a countable set J c and sets E ,
E , both determined by coordinates in J, such that E E E and E \E is negligible (254Oc). Now for
any particular c \ , {x : h(x) = } is negligible, being either empty or of the form {x : x(n) = z(n) for
every n < } for some z {0, 1} . So H = {x : h(x) J} is negligible. Now we see that for x X \ H,
f (x)J = xJ, so for x X \ (H (E \ E )),
x E = x E = f (x) E = f (x) E,
343K
Realization of homomorphisms
187
x
/ E = x
/ E = f (x)
/ E = f (x)
/ E.
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 be the set
{E : E I k , El and Er are Lebesgue measurable and El Er is Lebesgue negligible}.
For E , set
E = L El = L Er
where L is Lebesgue measure on [0, 1]. It is easy to check that (I k , , ) 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 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 L . 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 and L .
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 Ef 1 [E] is negligible for every E , so that f represents the
identity homomorphism on the measure algebra of . The function h : I k [0, 1] is canonical enough,
but is two-to-one, and the canonical map from the measure algebra of to the measure algebra of L is
represented equally by the functions t 7 t and t 7 t+ , which are nowhere equal.
(b) Consider the direct sum (Y, ) of (I k , ) and ([0, 1], L ); for deniteness, take Y to be (I k {0})
([0, 1] {1}). Setting
h1 (t+ , 0) = h1 (t , 0) = (t, 1),
for q Q, we must have g 1 [Eq ]Fq 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 of its measure algebra, we
cannot be sure of representing and 1 by functions which will be inverses of each other.
343K 342L has a partial converse.
Proposition If (X, , ) is a semi-nite countably separated measure space, it is compact i it is locally
compact i it is perfect.
proof We already know that compact measure spaces are locally compact and locally compact semi-nite
measure spaces are perfect (342Ha, 342L). So suppose that (X, , ) is a perfect semi-nite 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}.
188
343K
The denition of perfect measure space states exactly that whenever E and E > 0 there is a K K
such that K E and K > 0. And K is a compact class. P
P If K K has the nite intersection
property, L = {f [K] : K K } is aTfamily of compact sets in R with the nite intersection property, and
has non-empty intersection; so that K is also non-empty, because f is injective. Q
Q By 342E, (X, , ) 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, , ) be a complete locally determined countably separated measure space, and A X
a set such that the subspace measure A is perfect. Then A is measurable.
proof ?? Otherwise, there is a set E such that E < and B = A E
/ . Let f : X R be an
injective measurable function (343E again). Then f B is B -measurable, where B is the domain of the
subspace measure B on B. Set
K = {f 1 [L] : L f [B], L is compact in R}.
(B
\
B
nN K
S n nN
Sn ) = 0. But of course K , because f is
-measurable, so nN Kn . Because is complete, B \ nN Kn and B . X
X
343M Example 343L tells us that any non-measurable set X of R r , or of {0, 1}N , with their usual
measures, is not perfect, therefore not (locally) compact, when given its subspace measure.
To nd 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 (132Ee). 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 R r or {0, 1}N and for its measure, Z is countably separated; suppose hEn inN is a
sequence of measurable sets in Z separating its points. For each n, (Y En ) in the measure algebra of
Y corresponds to (X En ) in the measure algebra of X . So if f : Y X were a function representing
the isomorphism
of the measure algebras, (Y En )f 1 [En ] would have to be negligible for each n, and
S
A = nN (Y En )f 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, , ) be a semi-nite 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 is locally
compact. (Hint: show that satises (ii) or (v) of 343B.) (ii) Suppose that for every non-negligible E
there is a non-negligible set A E such that the subspace measure on A is compact. Show that is locally
compact.
(b) Let hXi iiI be a family of non-empty sets, with product X; write i : X Xi for the coordinate
N
map. Suppose we are given a -algebra i of subsets of Xi for each i; let = c iI i be the corresponding
-algebra of subsets of X generated by {i1 [E] : i I, E i }. Let be a totally nite measure with
domain , and for i I let i be the image measure i1 . Check that the domain of i is i . Show that if
every (Xi , i , i ) is compact, then so is (X, , ). (Hint: either show that satises (v) of 343B or adapt
the method of 342Gf.)
(c) Let I be any set. Let Ba be the -algebra of subsets of {0, 1}I generated by the sets Fi = {z : z(i) = 1}
for i I, and any probability measure with domain Ba; let B be the measure algebra of . Let (X, , )
be a measure space with measure algebra A, and : B A an order-continuous Boolean homomorphism.
Show that there is an inverse-measure-preserving function f : X {0, 1}I representing . (Hint: for each
i I, take Ei such that Ei = Fi ; set f (x)(i) = 1 if x Ei , and use 343Ab.)
(d) Let (X, , ) be an atomless probability space. Let B be the restriction of Lebesgue measure to
the -algebra of Borel subsets of [0, 1]. Show that there is a function g : X [0, 1] which is inverse-
343Yc
Realization of homomorphisms
189
measure-preserving
for and B . (Hint: nd an f : X {0, 1}N as in 343Xc, and set g = hf where
P
h(z) = n=0 2n1 g(n), as in 254K; or choose Eq such that Eq = q, Eq Eq whenever q q in
[0, 1] Q, and set f (x) = inf{q : x Eq } for x E1 .)
(e) Let (X, , ) be a countably separated measure space, with measure algebra A. (i) Show that {x}
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 described in 343J, and h : I k [0, 1] the
canonical surjection. Show that the canonical isomorphism between the measure algebras of and Lebesgue
measure on [0, 1] is given by the formula E 7 h[E] for every measurable E I k .
(g) Let (X, , ) and (Y, T, ) be measure spaces with measure algebras (A,
), (B, ). Suppose that
X Y = and that we have a measure-preserving isomorphism : A B. Set
= {W : W X Y, W X , W Y T, (W X) = (W Y ) },
(b) (i) Show that a countably separated semi-nite 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,
dene 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 Hausdor 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 i the
sets Er , El [0, 1], as described in 343J, are Borel and Er El 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 of 343J is inner regular with respect to the compact subsets of
I k . (vi) Show that we have a lower density for dened by setting
1
190
343Yd
(d) Set X = {0, 1}c , with its usual measure c . 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 c . (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 Hr be r-dimensional Hausdor measure on R s , where s 1 is an integer and r 0 (264). (i)
Show that Hr is countably separated. (ii) Show that the c.l.d. version of Hr is compact. (Hint: 264Yi.)
(f ) Give an example of a countably separated probability space (X, , ) and a function f from X to a
set Y such that the image measure f 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 {ab : a E, b [N]< } 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 signicant 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 rmly on the measure
space corresponding to the domain of the Boolean homomorphism, which will be the codomain of the
measurable function. So the rst step is to try to understand properly which spaces are compact, and what
other properties they can be expected to have; which accounts for much of the length of 342. But having
understood that many of our favourite spaces are compact, we have to come to terms 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, ) is compact, without simultaneously nding 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 null
ideal; it belongs to the theory of -algebras of sets. Some simple permanence properties are in 343H and
343Yb(ii). Let us note in passing 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.
344B
Realization of automorphisms
191
arguments for 344C, suitably rened, give a complete description of perfect complete countably separated
strictly localizable spaces which are not purely atomic (344I, 344Xc). At the same time we nd 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 i they have the same measure (344J, 344L).
344A Stone spaces The rst case is immediate from the work of 312, 313 and 321, as collected
in 324E. If (Z, , ) is actually the Stone space of a measure algebra (A,
), then every order-continuous
Boolean homomorphism : A A corresponds to a unique continuous function f : Z Z (312P)
which represents (324E). The uniqueness of f means that we can be sure that f = f f for all ordercontinuous homomorphisms and ; and of course f is the identity map on Z, so that f1 will have to
be f1 whenever 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, , ) 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 hf iG of such representatives can be chosen
in such a way that f = f f for all , G; and if the identity automorphism belongs to G, then we
may arrange that f is the identity function on X.
proof (a) Because G {} satises the same conditions as G, we may suppose from the beginning that
belongs to G itself. Let A be a countable set separating the points of X. For each G take some
representing function g : X X; take g to be the identity function. If , G, then of course
((g g )1 [E]) = (g1 [g1 [E]]) = (g1 [E])
1
= E = (g
[E])
H = {x : g (x) 6= g g (x)}
,G
H ;
because G is countable, H is also measurable and negligible. Try dening f : X X by setting f (x) =
g (x) if x X \ H, f (x) = x if x H. Because H is measurable, f1 [E] for every E ; because H
is negligible,
(f1 [E]) = (g1 [E]) = E
for every E , and f represents , for every G. Of course f = g is the identity function on X.
(c) If G then f1 [H] = H. P
P (i) If x H then f (x) = x H. (ii) If f (x) H and f (x) = x then
of course x H. (iii) If f (x) = g (x) H then there are , G such that g g g (x) 6= g g (x). So
either
g g (x) 6= g (x),
or
g g (x) 6= g (x)
or
and in any case x H. Q
Q
g (x) 6= g g (x),
192
344B
f f (x) = x = f (x).
(ii) If x X \ H then f (x)
/ H, by (c), so
344C Corollary Let (X, , ) 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 G, a function f : X X representing
, in such a way that f = f f for all , G; and if the identity automorphism belongs to G, then
we may arrange that f is the identity function on X. In particular, if G is invertible, and 1 G, we
shall have f1 = f1 ; so that if moreover and 1 are measure-preserving, f will be an automorphism
of the measure space (X, , ).
proof By 343K, (X, , ) 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 R r , 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
Schroder-Bernstein theorem for measurable spaces.
Lemma Let X and Y be sets, and PX, T PY -algebras. Suppose that there are f : X Y ,
g : Y X such that F = f [X] T, E = g[Y ] , f is an isomorphism between (X, ) and (F, TF ) and
g is an isomorphism between (Y, T) and (E, E ), writing E , TF for the subspace -algebras (see 121A).
Then (X, ) 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 inN T
is a nonincreasing sequence in and hYn inN is a non-increasing sequence in T. Set X = nN Xn , Y = nN Yn .
Then f X2k \ X2k+1 is an isomorphism between X2k \ X2k+1 and Y2k+1 \ Y2k+2 , while g Y2k \ Y2k+1 is an
isomorphism between Y2k \ Y2k+1 and X2k+1 \ X2k+2 ; and g Y is an isomorphism between Y and X .
So the formula
[
h(x) = f (x) if x
X2k \ X2k+1 ,
kN
=g
= Ei . Let J be the
of Ba generated by {Ei : I J}. For i I, G choose Fi Ba such that Fi
family of those subsets J of I such that Fi BJ for every i J, G.
344E
Realization of automorphisms
193
(b) For the purposes of this proof, I will say that a pair (J, hg iG ) is consistent if J J and, for
each G, g is a function from X to itself such that
g1 [Ei ] BJ and (g1 [Ei ]) = Ei whenever i J, G,
g1 [Ei ] = Ei whenever i I \ J, G,
g g = g whenever , G,
g (x) = x for every x X.
Now the key to the proof is the following fact: if (J, hg iG ) is consistent, and J is a member of J such
h
that J \ J is countably innite, then there is a family h
g iG such that (J,
g iG ) is consistent and
1
1
g [Ei ] = g [Ei ] whenever i J, G. The construction is as follows.
(i) Start by xing on any innite set K J \ J such that (J \ J) \ K is also innite. For z {0, 1}K ,
set Vz = {x : x X, xK = z}; then Vz BJ. All the sets Vz , as z runs over the uncountable set {0, 1}K ,
are disjoint, so they cannot all have non-zero measure (because is -nite), and we can choose z such that
Vz is -negligible.
(ii) Dene h : X X, for G, by setting
h (x)(i) = g (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 Fi \ Vz ,
= 0 if i J \ J and x
/ Fi Vz .
() h1
[Ei ] BJ and h [Ei ]Fi is negligible if i J \ J,
and consequently
() (h1
[Ei ]) = Ei for every i J,
() (h1
[E]) = E for every E BJ
(by 343Ab); moreover,
1
() h1
[E] = g [E] for every E BJ ,
() h1
[E] BJ for every E BJ,
() h1 [Ei ] = Ei if i I \ J,
so that
nally
() h1
[E] Ba for every E Ba;
() h (x) = x for every x X.
H, = {x : x X, h h (x) 6= h (x)}
H, =
iI
1
1
h1
[h [Ei ]]h [Ei ].
1
Now if i J, then h1
[Ei ] = g [Ei ] BJ , so
1
1
1 1
1 1
1
h1
[h [Ei ]] = h [g [Ei ]] = g [g [Ei ]] = g [Ei ] = h [Ei ].
Next, for i I \ J,
So
1
1
1
h1
[h [Ei ]] = h [Ei ] = Ei = h [Ei ].
194
344E
H, =
iJ\J
1
1
h1
[h [Ei ]]h [Ei ].
1
1
1
(h1
[h [Ei ]]) = (h [Ei ]) = Ei = (h [Ei ]) ,
1
1
and h1
[h [Ei ]]h [Ei ] is a negligible set, which by (ii-) belongs to BJ. So H, 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=
,G
H,
h1
[Vz ].
Then H BJ and H = 0. P
P We know that every H, is negligible and belongs to BJ ((iii) above), that
1
every h1
[V
]
belongs
to
B
(by
(ii-), and that (h1
z
because
x
/ H, H, H, H, h1
[Vz ];
thus h (x)
/ H, h1
/ H. Q
Q
[Vz ]; as and are arbitrary, h (x)
(v) The next fact we need is that there is a bijection q : X H such that () for E H, E BJ i
P Fix any bijection r : J \ J J \ (J K).
q 1 [E] BJ () q(x)(i) = x(i) for every i I \ (J \ J), x X. P
Consider the maps p1 : X H, p2 : H X given by
p1 (x)(i) = x(r1 (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)I \ (J \ J) = xI \ (J \ J) for every x X, and p2 (y)I \ (J \ J) = yI \ (J \ J) for every y H,
q(x)I \ (J \ J) = xI \ (J \ J) for every x X. Q
Q
to BJ, because it belongs to BJ if i J, and otherwise is equal to Ei . Consequently g1 [E] BJ for every
E BJ.
(vii) I am at last ready to give a formula for g . For G set
g (x) = h (x) if x X \ H,
= qg q 1 (x) if x H.
h
Now (J,
g iG ) is consistent. P
P
G,
() If i J,
1 1
g1 [Ei ] = (h1
[Ei H]]] BJ
[Ei ] \ H) q[g [q
1
because H BJ and h1
[H E], g1 [E] and q[E] all belong to BJ for every E BJ. At the same
[E], q
time, because g agrees with h on the conegligible set X \ H,
(
g1 [Ei ]) = (h1
[Ei ]) = Ei .
344E
Realization of automorphisms
195
G, x X then
() If i I \ J,
and if x H then q (x)(i) is also equal to x(i); so g (x)(i) = x(i). But this means that g1 [Ei ] = Ei .
() If , G and x X \ H, then
1
g (x) = h (x) X \ H
by (iv) above. So
g (x) = qg q 1 (x) H,
h
Thus (J,
g iG ) satises all the required conditions.
(c) The remaining idea we need is the following: there is a non-decreasingSfamily hJ i in J , for some
cardinal , such that J+1 \ J is countably innite for every < , J = < J for every limit ordinal
< , and J = I. P
P Recall that I am already supposing that I is innite. If I is countable, set = 1,
J0 = , J1 = I. Otherwise, set = #(I) and let hi i< be an enumeration of I. For i I, G let Ki I
be a countable set such that Fi BKi . Choose the J inductively, as follows. The inductive hypothesis
must include the requirement that #(J ) max(, #()) for every . Start by setting J0 = . Given
S <
and J J with #(J ) max(, #()) < , take an innite set L \ J and set J+1 = J nN Ln ,
where
L0 = L {i },
S
Ln+1 = iLn ,G Ki
Fi BLn+1 whenever i Ln , G
S
J ; since L J+1 \ J nN Ln , J+1 \ J is countably innite, and
Thus the induction proceeds. Observing that the construction puts i into J+1 for every , we see that J
will be the whole of I, as required. Q
Q
(d) Now put (b) and (c) together, as follows. Take hJ i from (c). Set f0 (x) = x for every G,
x X; then, because J0 = , (J0 , hf0 iG ) is consistent in the sense of (b). Given that (J , hf iG ) is
consistent, where < , use the construction of (b) to nd a family hf,+1 iG such that (J+1 , hf,+1 iG )
is consistent and f,+1 (x)(i) = f (x)(i) for every i J . At a non-zero limit ordinal , set
f (x)(i) = f (x)(i) if x X, < , i J ,
= x(i) if i I \ J .
196
344E
(The inductive hypothesis includes the requirement that f (x)J = f (x)J whenever G, x X
and < .) To see that (J , hf iG ) is consistent, the only non-trivial point to check is that
f, f, = f,
for all , G. But if i J there is some < such that i J , and in this case
1
1
f,
[Ei ] = f,
[Ei ] BJ
while if i I \ J then
1 1
1
1
1
[f, [Ei ]].
[Ei ] = f,
[Ei ] = f,
[Ei ] = Ei = f,
f,
Thus
1 1
1
f,
[f, [Ei ]] = f,
[Ei ]
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
Realization of automorphisms
344I Theorem Let (X, , ) and (Y, T, ) 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,
) of has Maharam type . P
P Let hEn inN be a
sequence in separating the points of X. Let 0 be the -subalgebra of generated by {En : n N},
and A0 the order-closed subalgebra of A generated by {En : n N}; then E A0 for every E 0 , and
(X, 0 , 0 ) is countably separated. Let f : X R be 0 -measurable and injective (343E). Of course f
is also -measurable. If a A \ {0}, express a as E where E . Because (X, , ) is perfect, there is a
compact K R such that K f [E] and f 1 [K] > 0. K is surely a Borel set, so f 1 [K] 0 and
b = f 1 [K] A0 \ {0}.
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 E, gF are the two halves of a one-to-one correspondence between E and F .
Because is measure-preserving, f 1 [H] = H, g 1 [G] = G for every G , H T; accordingly f E
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 and are complete, f1 is an isomorphism between the measure spaces (X, , ) and
(Y, T, ), as required.
344J Corollary Suppose that E, F are two Lebesgue measurable subsets of R r 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] i 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 i it is an atomless countably separated
compact (or perfect) -nite measure space which is not totally nite; in this case it is also isomorphic to
Lebesgue measure on any Euclidean space R r .
(c) Let be Lebesgue measure on R. If 0 < E < and we set F =
1
F
E
198
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 I the usual measure on {0, 1}I . If E, F {0, 1}I are two measurable sets
of the same non-zero nite measure, the subspace measures on E and F are isomorphic.
proof Write X = {0, 1}I .
(a) If I is nite, then X, E and F are all nite, and #(E) = #(F ), so the result is trivial. If I is
countably innite, 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 (BI , I ) be the measure algebra of I . Then BI is homogeneous, and I E = I F , so there is
an automorphism : BI BI such that E = F . (Apply 333D to the nite subalgebra C generated
by {E , F }; or argue directly from 331I.) By 344F, there is a measure space isomorphism f : X X
representing , so that f 1 [E]F and f [F ]E are negligible.
(c) We can nd a countably innite set J I and a measurable set E such that E is determined by
coordinates in J, E E f [F ] and E \ E is negligible; so that E is non-empty. Take any x0 E
and set V = {x : x X, xJ = x0 J}; then V is a negligible subset of E and #(V ) = #(X) (because
#(I \ J) = #(I) and x 7 xI \ J : V {0, 1}I\J is a bijection). Setting E1 = E f [F ], F1 = F f 1 [E],
f E1 is a bijection between E1 and F1 . Now
#(V (E \ E1 )) = #(X) = #(f 1 [V ] (F \ F1 )),
= h(x) if x f 1 [V ] (F \ F1 ).
344 Notes
Realization of automorphisms
199
344Y Further exercises (a) Let X be a set, a -algebra of subsets of X, I a -ideal of , and
A the quotient /I. Suppose that there is a countable set A 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 hf iG of such representatives can be
chosen in such a way that f = f f for all , G; and if the identity automorphism belongs to G,
then we may arrange that f is the identity function on X.
(b) Let A, B be Dedekind -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 innite set, and write Ba for the -algebra of subsets of X = {0, 1}I generated by the
sets {x : x(i) = 1} as i runs over I. Let and be -nite measures on X, both with domain Ba, and
with measure algebras (A,
), (B, ). Show that any Boolean isomorphism : A B is represented by a
bijection f : X X such that f 1 represents 1 : B A, and hence that (A,
) is isomorphic to (B, )
i (X, Ba, ) is isomorphic to (X, Ba, ).
(d) Let I be any set, and write Ba for the -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 1 -saturated ideal of Ba, 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 G, a function f : X X representing , in such
a way that f = f f for all , G; and if the identity automorphism belongs to G, then we may
arrange that f is the identity function on X. In particular, if G is invertible and 1 G, f will be
an automorphism of the structure (X, Ba, I).
(e) Let I be any set, and write Ba for the -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 1 -saturated ideals of Ba. Show that if the Boolean algebras
Ba/I and Ba/J are isomorphic, so are the structures (X, Ba, I) and (X, Ba, J ).
(f ) Show that if (X, , ) is a perfect semi-nite measure space which is not purely atomic, there is a
negligible set of cardinal c. (Hint: reduce to the case in which is atomless and totally nite; in this case,
construct a measurable function f : X R such that the image measure = f 1 is atomless, and apply
344H to .)
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 nd 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 R r 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 classication. 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 rst 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
t better into Volume 5 than here.
I have oered three contexts in which automorphisms of measure algebras are represented by automorphisms of measure spaces (344A, 344C, 344E). In the rst 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 oer 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 1
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} for
uncountable cardinals . 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 specically to measure theory; rather, they
belong to the general theory of -algebras and -ideals of sets. In the case of the Schroder-Bernstein-type
theorem 344D, this is obvious from the formulation I give. (See also 344Yb.) In the case of 344B and
344E, I oer generalizations in 344Ya-344Ye. Of course the applications of 344B here, in 344C and its
corollaries, depend on Maharams 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, , ) 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 nite sets of dierent
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 dierent non-zero nite 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.
(E + x) = E + x
for every E , x X.
345B Theorem For any r 1, there is a translation-invariant lifting of Lebesgue measure on R r .
345C
201
Translation-invariant liftings
proof (a) Write for Lebesgue measure on R r , for its domain. Let : be lower Lebesgue density
P
(341E). Then is translation-invariant in the sense that (E + x) = E + x for every E , x R r . P
(E + x) = {y : y R r , lim
0
(E+x)B(y,)
(B(y,))
= 1}
= {y : y R r , lim
(EB(yx,))
(B(yx,))
= 1}
(because is translation-invariant)
= {y + x : y R r , lim
0
(EB(y,))
(B(y,))
= 1}
Q
= E + x. Q
(b) Let 0 be any lifting of such that 0 E E for every E (341J). Consider
E = {y : 0 0 (E y)}
for E . It is easy to check that : is a Boolean homomorphism because 0 is, so that, for
instance,
y EF 0 0 (E y)0 (F y)
so y E. Q
Q By 341Ib, is a lifting for .
y E + 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 = #(I), it is enough to consider the
case I = . Write for the usual measure on X = {0, 1}I = {0, 1} and T for its domain. For each <
set E = {x : x X, x() = 1}, and let be the -algebra generated by {E : < }. Because x + E is
either E or X \ E , and in either case belongs to , for every < and x X, is translation-invariant.
(Consider the algebra
= {E : E + x for every x X};
this must be .) Let be the set of partial lower densities : T which are translation-invariant
in the sense that (E + x) = E + x for any E , x X.
(b)(i) For < , +1 is just the algebra of subsets of X generated by {E }, that is, sets of the
form (F E ) (G \ E ) where F , G (312M). Moreover, the expression is unique. P
P Dene x X by
202
345C
Liftings
so
F = (H E ) ((x + H) \ E ) = FH ,
G = (H \ E ) ((x + H) E ) = GH
are determined by H. Q
Q
(ii) The functions H 7 FH , H 7 GH : +1 dened above are clearly Boolean homomorphisms;
moreover, if H, H +1 and HH is negligible, then
(FH FH ) (GH GH ) (HH ) (x + (HH ))
= ((F + x) E ) ((G + x) \ E )
= x + (F E ) (G \ E )
= x + 1 ((F E ) (G \ E )).
= x + (F E ) (G \ E )
= x + 1 ((F E ) (G \ E )).
So 1 +1 .
(iv) Thus every member of has an extension to a member of +1 .
(c) Now suppose that h(n)inN
S is a non-decreasing sequence in with supremum < . Then is
just the -algebra generated by nN (n) . If we have a sequence hn inN such that n (n) and n+1
extends n for every n, then there is a extending every n . P
P I repeat the ideas of 341G.
(i) For E , n N choose gEn such that gEn is a conditional expectation of E on (n) ; that is,
gEn =
E = (F E)
for every E (n) . Moreover, make these choices in such a way that () every gEn is (n) -measurable
and dened everywhere on X () gEn = gE n for every n if EE is negligible. Now limn gEn exists and
is equal to E almost everywhere, by Levys martingale theorem (275I).
(ii) For E , k 1, n N set
345C
Translation-invariant liftings
E =
k1
nN
mn
203
km (E).
H
kn () is empty;
(iii) Every gn is zero almost everywhere, every Hkn () is negligible and every H
k+1,n (E) H
k+1,n (F )
kn (E F ) (Hk+1,n (E) Hk+1,n (F )) = H
H
n
for all k 1, n N. Now, if x E F , then, for any k 1, there are n1 , n2 N such that
T
k+1,m (E), x T
x mn1 H
mn2 Hk+1,m (F ).
But this means that
mmax(n1 ,n2 )
km (E F ).
H
VE E is negligible; but
EV
k1,nN
kn (E)
Hkn (E)H
for every y X, n N; that is, gn is the composition gEn , where (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 is a measure space automorphism, and in
particular is inverse-measure-preserving, we have
g
F +x n
g
1 [F ] n
gEn = (E F )
whenever F (n) (235Ic). But because (n) is itself translation-invariant, we can apply this to F x to
get
gn = (E (F x)) = ((E + x) F )
for every , and gn is (n) -measurable. So gn is a conditional expectation of (E + x) on (n) , and must
be equal almost everywhere to gE+x,n .
204
Liftings
345C
Hkn
= {y : gn (y) 1 2k } = Hkn (E) + x
kn (E) + x.
= n (Hkn (E) + x) = n (Hkn (E)) + x = H
Consequently
(E + x) =
\ [ \
kn (E + x)
H
k1 nN mn
\ [ \
kn (E) + x = E + x.
H
k1 nN mn
Q
As E and x are arbitrary, is translation-invariant and belongs to . 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 h i such that () for each () extends whenever . The
induction starts with 0 = {, X}, 0 = , 0 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 conality is dealt with in (c). If
has uncountable conality, then = < , so we can (and must) take to be the unique common
extension of all the previous .
The induction ends with : T . Note that is not in general the whole of T . But for every
E T there is an F such that EF is negligible (254Ff). So we can extend to a function
dened on the whole of T by setting
E = F whenever E T , F and (EF ) = 0
(the point being that F = F if F , F and (EF ) = (EF ) = 0). It is easy to check that
is a lower density, and it is translation-invariant because if E T , x X, F and EF is negligible,
then (E + x)(F + x) = (EF ) + x is negligible, so
(E + x) = (F + x) = F + x = E + 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 R r
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, , ) be either Lebesgue measure on R r or the usual measure on {0, 1}I for some
set I, and let : be a translation-invariant lifting. Then for any open set G X we must have
G G G, and for any closed set F we must have int F F 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 () If X = Rr , take > 0 such that {y : ky xk } G, and set
1
U = {y : ky xk < 2 }. () If X = {0, 1}I , then there is a nite set K I such that {y : yK = xK} G
(3A3K); set U = {y : y(i) = 0 for every i K}. Q
Q
It follows that x G. P
P Consider H = x + U . Then H = U > 0 so H H 6= . Let y U be such
that x + y H. Then
because
x = (x + y) y (H y) G
H y x + U U G. Q
Q
345F
Translation-invariant liftings
205
(b) Thus G G for every open set G X. But it follows at once that if G is open and F is closed,
int F (int F ) F ,
G = X \ int(X \ G) X \ (X \ G) = G,
F = X \ (X \ F ) X \ (X \ F ) = F .
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}I .
Lemma Let I be an innite set, I the usual measure on X = {0, 1}I , and E X any non-negligible
measurable set. Then there are x, x E which dier at exactly one coordinate.
proof By 254Fe, there is a set F , determined by coordinates in a nite set J I, such that I (EF )
1
3
1
4 I E; we have I F 4 I E, so I (EF ) 3 I F . Take any i I \ J. Then the map : X X, dened
by setting (x)(i) = 1 x(i), (x)(j) = x(j) for j 6= i, is a measure space automorphism, and
I ( 1 [EF ] (EF )) 2I (EF ) < I F .
Take any x F \ ((EF ) 1 [EF ]). Then x = x diers from x at exactly one coordinate; but also
x F , by the choice of i, so both x and x belong to E.
345F Proposition Let be Borel measure on R, that is, the restriction of Lebesgue measure to the
Borel -algebra B. Then is translation-invariant, but has no translation-invariant lifting.
proof (a) To see that 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 terms of the geometry of the Cantor set C. Recall
T
that C is dened as the intersection nN Cn of a sequence of closed subsets of [0, 1]; each Cn consists of
2n closed intervals of length 3n ; Cn+1 is obtained from P
Cn by deleting the middle third of each interval of
Cn . Any point of C is uniquely expressible as f (e) = 23 n=0 3n e(n) for some e {0, 1}N . (See 134Gb.)
Give {0, 1}N its usual measure. 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 Jz be the open interval with the same centre as Jz and twice the length. Then
Jz \ Jz consists of two open intervals of length 3n /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, Jz a 0 and Jz a 1 are disjoint subsets of Jz .
Consequently Jz Jw is non-empty just when one of z, w extends the other, and we need consider only
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 = wm, where m < n. () If in the extension (w(m), . . . , w(n 1)) both values 0 and 1 appear, Jw will be
a subset of Jz , and certainly the four sets will all be disjoint. () If w(i) = 0 for m i < n, then Ww Jz
is disjoint from the rest, while Vw Vz ; but z and w take the value 1 the same number of times, so Vw is
206
345F
Liftings
assigned to G i Vz is, and otherwise both are assigned to H. () 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:
V 00
J00
W00 V 01 J01
V0
V
J0
W01
V10
W0
J10
V1
W11
J1
W1
1
and for w A let w be the nite sequence obtained from w by changing w(n) = 0 into w (n) = 1 but
leaving the other values of w unaltered. Then Vw = Vw + and Ww = Ww + 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) + =
(d) ?? Now suppose, if possible, that : B B is a translation-invariant lifting. Note rst that U U
for every open U R. P
P The argument is exactly that of 345D as applied to R = R 1 . Q
Q Consequently
1 3
J = 2 , 2 J .
But as J \ (G H) is negligible,
C 21 , 23 G H.
Consider the sets E = f 1 [G], F = {0, 1}N \ E = f 1 [H]. Because f is measurable and G, H 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, e {0, 1}N , diering 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, e (n) = 1 and e(i) = e (i) for i 6= n.
345Y
Translation-invariant liftings
207
Set z = en = e n. Then f (e) belongs to the open interval I0 = Jz a 0 , so f (e) I0 and f (e) G i
f (e) (I0 G). But now
so
f (e ) = f (e) + 2 3n1 I1 = Jz a 1 ,
f (e ) ((I0 G) + 2 3n1 )
f (e ) (I1 H)
f (e ) H
e F.
with the group operation +2 , show that we can set E = ], ] ( nZ E + 2n), where is any
translation-invariant lifting for Lebesgue measure.)
> (c) Show that there is no lifting of Lebesgue measure on R which is symmetric in the sense that
(E) = E for every measurable set E, writing E = {x : x E}. (Hint: can 0 belong to ([0, [)?)
> (d) Let be Lebesgue measure on X = R \ {0}. Show that there is a lifting of such that
(xE) = xE for every x X and every measurable E X, writing xE = {xy : y E}.
(e) Let I be the usual measure on X = {0, 1}I , for some set I, TI its domain, and (BI , I ) its measure
algebra. (i) Show that we can dene x (a) = a + x, for a BI and x X, by the formula E + x = (E + x) ;
and that x 7 x is a group homomorphism from X to the group of measure-preserving automorphisms of
BI . (ii) Dene as in the proof of 345C, and set A = {E : E }. Say that a partial lifting : A TI
is translation-invariant if (a + x) = a + x for every a A and x X. Show that any such partial lifting
can be extended to a translation-invariant partial lifting on A+1 . (iii) Write out a proof of 345C in the
language of 341F-341H.
> (f ) Let be a lower density for Lebesgue measure on R r which is translation-invariant in the sense
that (E + x) = E + x for every x R r and every measurable set E. Show that G G for every open
set G R r .
(g) Let be 1-dimensional Hausdor measure on S 1 , as in 345Xb. Show that there is no translationinvariant lifting of such that E is a Borel set for every E dom .
345Y Further exercises (a) Let (X, , ) be a complete measure space, and suppose that X has a
group operation (x, y) 7 xy (not necessarily abelian!) such that is left-translation-invariant, in the sense
that xE = {xy : y E} and (xE) = E whenever E and x X. Suppose that : is a
lower density which is left-translation-invariant in the sense that (xE) = x(E) for every E , x X.
Show that there is a left-translation-invariant lifting : such that E E for every E .
208
345Yb
Liftings
(b) Write for the -algebra of Lebesgue measurable subsets of R, and L0 () for the linear space of
-measurable functions from R to itself. Show that there is a linear operator T : L0 () L0 () such that
() (T u) = u for every u L0 () () supxR |(T u)(x)| = kuk for every u L () () T u 0 whenever
u L () and u 0 () T is translation-invariant in the sense that T (Sx f ) = Sx T f for every x R and
f L0 (), where (Sx f )(y) = f (x + y) for f L0 () and x, y R () T is reection-invariant in the sense
that T (Rf ) = RT f for every f L0 (), where (Rf )(x) = f (x) for f L0 () and x R. (Hint: for
f L0 (), set
p(f ) = inf{ : [0, ], lim0
1
{x
2
Set V = {u : u L0 (), p(u) < } and show that V is a linear subspace of L0 () and that pV is a
seminorm. Let h0 : V R be a linear functional such that h0 (R) = 1 and h0 (u) p(u) for every
u V . Extend h0 arbitrarily to a linear functional h1 : L0 () R; set h(f ) = 21 (h1 (f ) + h1 (Rf ) ). Set
(T f )(x) = h(Sx f ) . You will need 223C.) Show that there must be a u L1 () such that u 0 but
T u 6 0.
(c) Show that there is no translation-invariant lifting of the usual measure on {0, 1}N such that E 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 simplication
because the set E in 345C, corresponding to the set E of the proof of 341F, is independent of the algebra
in a very strong sense, so that the expression of an element of +1 in the form (F E ) (G \ E ) is
unique. Interpreted in the terms of 341F, we have w = v = 1, so that the formula
346E
Consistent liftings
209
346A Definition Let h(Xi , i , i )iiI be a family of probability spaces, with product (X, , ). I will
say that a lifting : respects coordinates if E is determined by coordinates in J whenever E
is determined by coordinates in J I.
Remark Recall that a set E X is determined by coordinates in Q
J if x E whenever x E, x X
1
and x J = xJ; that is, if E is expressible as J [F ] for some F iJ Xi , where J (x) = xJ for every
x X; that is, if E = J1 [J [E]]. See 254M. Recall also that in this case,
Q if E is measurable for the product
measure on X, then J [E] is measurable for the product measure on iJ Xi (254Ob).
346B Proposition Let h(Xi , i , i )iiI be a family of probability spaces, with product (Z, , ). For
J I let (ZJ , J , J ) be the product of h(Xi , i , i )iiJ , and J : Z ZJ the canonical map. Let
: be a lifting. If J I is such that W is determined by coordinates in J whenever W is
determined by coordinates in J, then induces a lifting J : J J dened by the formula
J1 [J E] = (J1 [E]) for every E J .
proof If E J , then 1 [E] and (J1 [E]) are determined by coordinates in J, so (J1 [E]) is of the
form J1 [F ] for some F J ; this denes J : J J . It is now easy to see that J is a lifting.
Remark Of course we frequently wish to use this result with a singleton set J = {j}. In this case we must
remember that (ZJ , J , J ) corresponds to the completion of the probability space (Xj , j , j ).
346C Theorem Let I be any set, and I the usual measure on X = {0, 1}I . Then any translationinvariant lifting of I respects coordinates.
proof Suppose that E X is a measurable set determined by coordinates in J I; take x E and
x X such that x J = xJ. Set y = x x; then y(i) = 0 for i J, so that E + y = y. Now
x = x + y E + y = (E + y) = E
210
Liftings
346E
S
Set K = iI Ji , and let K be the usual measure on Y = {0, 1}K , TK its domain. We have a natural
Q
bijection between iI {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,
BK of and K .
P
P(i) If L K is nite and z {0, 1}L , then, setting Li = LcapJi for i I,
Y
{x : x X, f (x)L = z} = ( {w : w Xi , fi (w)Li = zLi })
iI
Y
iI
Y
iI
i {w : w Xi , fi (w)Li = zLi }
Ji {v : v {0, 1}Ji , vLi = zLi }
Because [BK ] is a closed subalgebra of A (324Kb), is a -subalgebra of the domain of , and of course
it contains all -negligible sets. If i J and G i , then there is an H {0, 1}Ji such that Gfi1 [H] is
i -negligible. Now if E = {x : x X, x(i) G} and F = {y : y Y, yJi H},
such
that (EE ) = 0 (254Ob). Write TL for the family of sets in TK determined by coordinates in
S
iL Ji . Then, just as in (b-ii), every member of L diers by a negligible set from some set of the form
E = f 1 [F ] = f 1 [F ].
S
But respects coordinates, so F is determined by coordinates in iL Ji . It follows at once that f 1 [F ]
is determined by coordinates in L; that is, that E is determined by coordinates in L. As E and L are
arbitrary, 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.
346G
Consistent liftings
211
Lemma Let (X, , ) and (Y, T, ) be complete probability spaces, with product (X Y, , ). If :
is a lower density, then we have a lower density 1 : dened by saying that
1 E = {x : x X, {y : (x, y) (E Y )} is conegligible in Y }
for every E .
proof For E , (E Y )(E Y ) is negligible, so that
Hx = {y : (x, y) (E Y )(E Y )}
is -negligible for almost every x X (252D). Now E1 E = {x : Hx is not negligible} is negligible, so
1 E . If E, F , then
((E F ) Y ) = ((E Y ) (F Y )) = (E Y ) (F Y ),
so that
{y : (x, y) ((E F ) Y )} = {y : (x, y) (E Y )} {y : (x, y) (F Y )}
is conegligible i both {y : (x, y) (E Y )} and {y : (x, y) (F Y )} are conegligible, and 1 (E F ) =
1 E 1 F .
The rest is easy. Of course ( Y ) = so 1 = . If E, F and EF is negligible, then
(E Y )(F Y ) is negligible, (E Y ) = (F Y ) and 1 E = 1 F . So 1 is a lower density, as claimed.
346G Theorem Let h(Xi , i , i )iiI be a family of probability spaces with product (X, , ). For J I
let J be the set of members of which are determined by coordinates in J. Then there is a lower density
: such that
(i) whenever J I and E J then E J ,
(ii) whenever J, K I are disjoint, E J and F K then (E F ) = E F .
Q
proof For each i I, set Yi = XiN , with the product measure
Q i ; set Y = iI Yi , with its product measure
measure i , Q
and Z = Q
; set Zi = Xi Yi , with its product
iI Zi , with its product measure . Then the
Q
natural identication of Z = iI Xi Yi with iI Xi iI Yi = X Y makes correspond to the
product of and (254N).
Each (Zi , i ) can be identied with an innite power of (Xi , i ), and is therefore Maharam-type-homogeneous (334E). Consequently there is a lifting : which respects coordinates (346E). Regarding
(Z, ) as the product of (X, ) and (Y, ), we see that induces a lower density : by the formula
of 346F.
Q
If J I and E is determined by coordinates in J, then E Y (regarded as a subset of iI Zi ) is
determined by coordinates in J, so (E Y ) also is. Now suppose that x E, x X and xJ = x J.
Then for any y Y , (xJ, yJ) = (x J, yJ), so (x, y) (E Y ) i (x , y) (E Y ). Thus
{y : (x , y) (E Y )} = {y : (x, y) (E Y )}
is non-negligible. But
C = A B = {y : (x, y)
/ (E Y ) (F Y )} = {y : (x, y)
/ ((E F ) Y }.
So x
/ (E F ). As x is arbitrary, (E F ) E F ; but of course E F (E F ), because is
a lower density, so that (E F ) = E F , as required.
212
Liftings
346G
= ,
is negligible and H E is measurable and (H E) = (E) = E for every E . (ii) Because AH H
= F
and (E
H) = (F
H), so E = F .
H = . (iii) If E, F and (EF ) = 0 then E
H
H
and (E
H) (F
H), so E F . (v) If E, F
F
(iv) If E, F and E F then E
H
H
and x H E H F , then
() if x
/ AH ,
= (E
F ) (E F ),
F
x E
H
() if x AH ,
H) (F
H) = ((E
x (E
F ) H) H (E F ).
Q
Thus H E H F H (E F ) and H E H F = H (E F ). Q
(d) It is worth noting the following.
P
H
then E = E.
H) = E
P We have
(i) If E, H and (E
H
(AH (E
H)) = E
(AH E)
(AH H)
H E = E
= E
= . Q
Q
because AH H
346H
But
Consistent liftings
213
(Z
\ H) = Z. P
we must have = .
(ii) If H and H P then H
P By the maximality of ,
H
\ H),
AH = H (Z \ H) \ (Z
\ H) = Z. Q
(Z
Q
so AH = , that is, H
(Z
\ E) = Z, then (E
F ) = E
F
. P
(iii) If E, F and E
P
= (E
F ) (Z
\ E) = ((E
\ E) F
,
F ) \ E
F ) (Z \ E)) = (F
(E
F ) E
F
; as the reverse inclusion is true for all E and F , we have the result. Q
so (E
Q
(e) If < and H {} , then H P .
P
P(i) If J I is either a singleton or an inital segment, and E J , then
and (E
H) and AH all belong to J , so E J .
() if J, E H and E
H
() If
/ J, (E H) = E H, because there is some such that J and {} \ ); so
J by (d-i).
H E = E
(ii) If < and E , F \ , then
F H) = (E
H) F
, and
if < , E H so (E
F ) (AH (E
F H))
H (E F ) = (E
(AH (E
H)) (AH F
) E F;
F
= E
H
H
H), and
F H) = (E)
(F
if , F H \ so (E
F ) (AH (E
F H))
H (E F ) = (E
(AH E))
H)) E F ;
F
(AH (F
= E
H
H
accordingly H (E F ) = H E H F . Q
Q
By (d-ii) we have
whenever < and H {} .
(Z
\ H) = Z
H
is either or
(f ) If and H , then H P . P
P Induce on . For = 0, H 0 = {, Z} so H
Z and the result is trivial. For the inductive step to , we have the following.
(i) If < and E , then
and (E
H) and AH all belong to , so E .
() if , E H and E
H
= Z \ (Z
\ E) and (E
H) = E
H,
(AH (E
H)) (AH F
) E F,
F
= E
H
H
214
Liftings
346H
Q
Thus H P and the induction continues. Q
(g) But the case = of (f) just tells us that
(Z
\ H) = Z
H
for every H . This means that is actually a lifting (since it preserves intersections and complements).
And the denition 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, , ) be a complete probability space. For any set I, write I for the product
measure on X I , I for its domain and Ii (x) = x(i) for x X I , i I. Then there is a lifting :
1
1
such that for every set I there is a lifting : I I such that (Ii
[E]) = Ii
[E] for every E ,
i I.
proof ?? Suppose, if possible, otherwise.
Let be the set of all liftings for . We are supposing that for every there is a set I for
which there is no lifting for I consistent with in the sense above. Let be a cardinal greater than
max(, #(), sup #(I )). Let 0 : be a lifting satisfying the conditions of 346H. 346B tells us
1
1
that for every < we have a lifting for dened by the formula
[E] = 0 (
[E]). For set
1
1
K = { : < , 0 (
[E]) =
[E] for every E }.
S
Then K = , so max(, #(), sup #(K )) and there is some such that #(K ) >
#(I ). Take I K such that #(I) = #(I ).
We may regard X as X I X \I , and in this form we can use the method of 346F to obtain a lower
density : I I from 0 : . Now
1
1
(I
[E]) = I
[E] for every E , I.
1
1
1
1
[E] can be
[E]) =
[E] X , while 0 (
[E] X \I corresponds to
P
P The point is that I
1
1
1
\I
identied with I [E]X . Now the construction of 346F obviously makes (I [E]) equal to I
[E].
Q
Q
By 341Jb, there is a lifting : I I such that W W for every W I . But now we must have
1
1
1
[E])
[E]) (I
[E] = (I
I
1
1
= X I \ (I
[X \ E]) X I \ (I
[X \ E])
1
1
1
[E]
[X \ E] = I
[(X \ E)] = X I \ I
= X I \ I
1
1
[E] for every E , I. But since #(I) = #(I ), this must be impossible, by the
[E]) = I
and (I
choice of I . X
X
This contradiction proves the theorem.
346J Consistent liftings Let (X, , ) be a measure space. A lifting : is consistent if for
every n 1 there is a lifting n of the product measure on X n such that n (E1 . . .En ) = E1 . . .En
for all E1 , . . . , En . Thus 346I tells us, in part, that every complete probability space has a consistent
lifting; it follows that every non-trivial complete totally nite 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, ) be the Stone space of the measure algebra of Lebesgue measure on [0, 1],
and let be the product measure on Z Z, with its domain. Then there is a set W , with W < 1,
= 1, where
such that W
= S{G H : G, H Z are open-and-closed, (G H) \ W is negligible}.
W
346K
215
Consistent liftings
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
is, as always in this volume, the ordinary completed product measure as dened in Chapter 25.
proof (a) Let hEn inN be a sequence of measurable subsets of [0, 1], stochastically independent for Lebesgue
1
measure on [0, 1], such that En = n+2
for each n. Set an = En in the measure of algebra of , and
S
2
n=0 (En )
n=2 n2
< 1.
< 1. Then there are sequences hGn inN , hHn inN in T such that
?? Suppose, if possible, that W
S
S
n, H
n such that Gn G
n , Hn H
n for
for every F T. Accordingly we can nd compact open sets G
every n N and
P
P
S
n=0 (Gn \ Gn ) +
n=0 (Hn \ Hn ) < 1 ( nN Gn Hn ),
S
n H
n ) < 1.
so that ( nN G
Let U0 be the family
n : n N} {Z \ H
n : n N},
{Z} {E : n N} {Z \ G
n
n H
n ; take (w, z) (Z \ Q0 )2 \ S
1 and cannot be included in nN G
nN Gn Hn .
(d) We can nd sequences hCn inN , hDn inN , hUn inN and hVn inN in U such that
n H
n ) = ,
w Un+1 Un , z Vn+1 Vn , (Un+1 Vn+1 ) (G
Cn > 0, Dn > 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
n H
n . If w
n , set Un = Un \ G
n,
w Un U, z Vn U, then we know that (w, z)
/ G
/ G
Vn+1 = Vn Ek , Cn = Ek Un ,
216
346K
Liftings
z Vn+1 U, Cn Un ,
Cn Vn+1 Ek Ek W ,
Cn > 0.
Next, z
/ Q(Vn+1 ) and Vn+1 > 0; also w
/ Q(Vn+1 ), so there is an l such that w El and (El Vn+1 ) > 0.
Set
Un+1 = Un El ,
so that
w Un+1 U,
Dn Vn+1 ,
Dn = El Vn+1 ,
Un+1 Dn El El W , Dn > 0,
n H
n ) (Un Vn ) (G
n H
n ) = ,
(Un+1 Vn+1 ) (G
(G H) \ W ((G \ C) Z) (Z (H \ D))
S
GH W
nN Gn Hn .
n 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 kn Gk Hk = S say. Now (Uk+1 Vk+1 ) (Gk H
(Cn+2 Dn+2 ) (G H) (Un+1 Vn+1 ) S = ,
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
(f ) Thus W = 1, as required.
346L Proposition Let (Z, T, ) be the Stone space of the measure algebra of Lebesgue measure on [0, 1].
Let : T T be the canonical lifting, dened by setting E = G whenever E T, G is open-and-closed
and EG is negligible (341O). Then is not consistent.
proof ?? Suppose, if possible, that is a lifting on Z Z such that (E F ) = E F for every E,
F T. Let W Z Z be a set as in 346K, and consider W . If G, H Z are open-and-closed and
(G H) \ W is negligible, then
G H = G H = (G H) W ;
346X Basic exercises (a) Let (X, , ) be a measure space and a lower density for . Take H
and set A = X \ (H (Z \ H)), E = E (A (H E)) for E . Show that is a lower density.
> (b) Show that there is no lifting of Lebesgue measure on [0, 1]2 which is symmetric in the sense that
(E 1 ) = (E)1 for every measurable set E, writing E 1 = {(y, x) : (x, y) E}.
(c) Let (X, , ) be a measure space and hn inN a sequence of lower densities for . (i) Show that
S
T
T
E 7 nN n E and E 7 nN mn m E are also lower densities for . (ii) Show that if is complete
S
T
and F is any lter on N, then E 7 F F nF n E is a lower density for .
346Zb
Consistent liftings
217
(d) Let (X, , ) 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 : which is G-invariant in the
T
sense that (g 1 [E]) = g 1 [E] for every E , g G. (Hint: set E = gG g[0 (g 1 [E])].)
> (e) Let be lower Lebesgue density on R, and any lifting of Lebesgue measure on R such that
E E for every measurable set E. Show that is consistent. (Hint: given n 1, let n be lower
Lebesgue density on R n . For x R n let Ix be the ideal generated by
S
{W : x n (R n \ W )} i<n {i1 [E] : x(i) (R \ E)};
show that R n
/ Ix , so that we can use the method of 341J to construct a lifting of Lebesgue measure on
R n .)
(f ) Show that Lemma 346K is valid for any (Z, T, ) which is the Stone space of an atomless probability
space.
> (g) Suppose, in 341H, that (X, , ) is a product of probability spaces, and that in the proof, instead of
taking ha i< to run over the whole measure algebra A, we take it to run over the elements of A expressible
as E where E is determined by a single coordinate. Show that the resulting lower density respects
coordinates in the sense that E is determined by coordinates in J whenever E is determined by
coordinates in J. (Compare Macheras & Strauss 95, Theorem 2.)
346Y Further exercises (a) Suppose that (X, , ) and (Y, T, ) are complete probability spaces with
product (X Y, , ). Show that for any lifting 1 : there are liftings 2 : T T and :
such that (E F ) = 1 E 2 F for all E , F T. (Hint: use the methods of 341. In the inductive
construction of 341H, start with 0 (E Y ) = 1 E Y for every E . Extend each lower density to
the algebra generated by dom( ) {X F } for some F T. Make sure that (X F ) is always of the
form X F , and that ((E Y ) (X F )) = (E Y ) (X F ); adapt the construction of 341G to
maintain this. Use the method of 346H to generate a lifting from the nal lower density . See Macheras
& Strauss 96a, Theorem 4.)
(b) Use 346Ya and induction on to prove 346H. (Macheras & Strauss 96b.)
(c) Let (X1 , 1 , 1 ), . . . , (Xn , n , n ) be probability spaces with product (X, , ). Show that there is a
lifting for which respects coordinates. (Burke n95.)
(d) Let (X, , ) be a complete probability space. Show that there is a lifting : such that
whenever h(Xi , i , i )iiI is a family of probability spaces, with product measure , there is a lifting for
1
such that (i1 [E])
Q = i [E] whenever E and i I is such that (Xi , i , i ) = (X, , ), writing
i (x) = x(i) for x iI Xi .
(e) In 346Ya, nd a modication of the countable-conality inductive step of 341G which will ensure
that the lower density obtained in the product satises both conditions of 346G.
(f ) Let (X, , ) be a probability space, I any set, and the product measure on X I . Show that there
is a lower density for which is invariant under transpositions of pairs of coordinates.
346Z Problems (a) Let h(Xi , i , i )iiI be a family of probability spaces, with product (X, , ). Is
there always a lifting for which respects coordinates in the sense of 346A?
(b) Is there a lower density for the usual measure on {0, 1}N which is invariant under all permutations
of coordinates?
218
Liftings
346 Notes
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 oer 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. Maharams 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 rened, 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 disjoint sets, then (E F ) = E F , that is, is making a
tentative step towards being a lifting. (Remember that the dierence between a lifting and a lower density
is mostly that a lifting preserves nite unions as well as nite intersections; see 341Xa.) This can also be
achieved by a modication 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 reected in a symmetry of a lifting. The problems
there concern liftings themselves, not lower densities, since we can frequently nd lower densities which share
symmetries (346Xd, 346Yf). (Even for lower densities there seem to be diculties 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 Maharams original proof of the lifting theorem (Maharam 58).
351Bc
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 nd that any convincing account 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 ()-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).
for u, v, w U and R.
u 0, 0 = u 0
351B Elementary facts Let U be a partially ordered linear space. We have the following elementary
consequences of the denition 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 u v for every 0 and v u for every 0. P
P
(i) If 0, then (v u) 0 so v u. (ii) If 0 then ()u ()v so
v = ()v (u) = u. Q
Q
220
Riesz spaces
351C
(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 i
u + v C,
u C whenever u, v C and 0,
0 C,
u C & u C = u = 0.
P
P (i) If C = U + for some partially ordered linear space ordering on U , then
u, v C = 0 u u + v = u + v C,
u C, 0 = u 0, i.e., u 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 satises the conditions, dene 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
351G
221
(c) Moreover, we nd that if A, B U and sup A and sup B are dened, then sup(A + B) is dened 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 )
uA
Similarly, if A, B U and inf A, inf B are dened then inf(A + B) = inf A + inf B.
(d) If > 0 then u 7 u is an order-isomorphism, so we have sup(A) = sup A if either side is dened;
similarly, inf(A) = 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
denition.
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 i 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 verication. 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,
But (because T is positive) 0 is surely a lower bound for T [A], so it is also the inmum of T [A]. As A is
arbitrary, (ii) is true.
) If A U is non-empty, downwards-directed and has
(ii)(i) Suppose now that (ii) is true. (
inmum w, then A w is non-empty, downwards-directed and has inmum 0, so
222
Riesz spaces
inf T [A w] = 0,
351G
(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 denition. 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 nite 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),
whenever A U is a nite non-empty set and inf A, sup A exist. (Apply the denition in (a) to
A = {u inf A : u A},
A = {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} = uA [u, u]
in the notation of 2A1Ab. (I should perhaps remark that while this denition 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 dened by either of the rules
u w i 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 rst. 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 u1 u2 , u2 u3 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 u1 u3 .
(iii) If u, w U and u w , w u then there are v, v V such that u w + v, w u + v . Now
there are v0 , v0 V such that v0 v v0 , v0 v v0 , and in this case v0 , v0 0 (351Cd), so
v0 v0 v u w v v0 + v0 V ,
.
Accordingly u w V and u = w . Thus U/V is a partially ordered set.
(iv) If u1 , u2 , w U and u1 u2 , then there is a v V such that u1 u2 + v, in which case
u1 + w u2 + w + v and u1 + w u2 + w .
(v) If u U , R and u 0, 0 then there is a v V such that u + v 0; now v V and
u + v 0, so u = (u) 0.
351N
223
it is easy to see that V is a linear subspace of RX , and is solid because f V i |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.
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 nite, non-empty and has inmum
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 Zorns 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 0, v V + such that w v u0 w + v. P
P Let V1 be
{u : u U , there are 0, v V + such that w v u w + 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 nite and non-empty and inf A = 0 in U then AV 6= . P
P?? Otherwise,
for
P
every w A there must be w 0, vw V + such that w w vw u0 w w +vw . Set = 1+ wA w ,
P
v = wA vw V ; then w v u0 w + v for every w A. Accordingly 1 (u0 v) w for every
w A and 1 (u0 v) 0, so u0 v. Similarly, 1 (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 nite 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 jn Aj }.
Then C is a nite 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 jn 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 nd a set X, a lter 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; dene : U RX by setting (u)(f ) = f (u)
for every f X, u U , so that is linear. Let A be the family of non-empty nite 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 jn FAj }
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 >0 u = 0. P
P
(i)(ii) If (i) is true and u 0, then of course u 0 for every > 0; on the other hand, if v u for
every > 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 {u : > 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 > 0, then there is an n 1 such that n1 , so that u n1 v v (351Bc).
Accordingly u is a lower bound for {v : > 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 = iI 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 be any ordinal. The lexicographic ordering on R is dened by
saying that f g i either f = g or there is a < such that f () = g() for < and f () < g(). Show
that this is a total order on R which renders R 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 dene a partially ordered linear space structure on the quotient U/V i V is order-convex, that is,
u V whenever v1 , v2 V and v1 u v2 .
(c) Let hUi iiI be a family of partially ordered linear spaces with product U . Dene 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 iiI a family of partially orderedQlinear spaces
with product V . Show that L(U ; V ) can be identied, as partially ordered linear space, with iI 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 denition
351I) i 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 nite sets
such that () u v = sup{u, v} is dened for every u A, v B () inf A and inf B and (inf A) (inf B)
are dened. 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 i F is an ultralter, and
in this case has a natural structure as a totally ordered eld.
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 diculty lies in guessing the exact
boundary at which ones standard manipulations with such familiar spaces cease to be valid in the general
case. (For instance, most peoples favourite examples are Archimedean, in the sense of 351R, so it is prudent
226
351 Notes
Riesz spaces
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 inma of nite 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 conrm that an intuition valid for R or R X 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 nd a proof directly from the denition
than to trace through an argument relying on 351Q; but this is not always the case for Riesz spaces. I oer
351Yc as an example of a result where a direct proof does at least call for a moments 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.
u = (u) 0 = (u)+ ,
|u| = u (u)
u + (v w) = (u + v) (u + w),
(u v) = u v, (u v) = u v if 0,
352F
227
Riesz spaces
(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|,
|u| = |||u|
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 = u + v u v = (u + v |u v|),
u+ u = u (u) 0 = |u|,
u+ u = u+ + u (u+ u ) = 0,
u c + (u v)+ c + (a v)+
228
Riesz spaces
352F
u+ v + (u + v)+ u+ + v + .
Pn
Pm
(d) If u0 , . . . , um , v0 , . . . , vn U + and i=0 ui = j=0 vj , then there is a family hwij iim,jn 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
) A simple
(b)(i)(
the inductive step, shows that if v0 , . . . , vm , w0 , . . . , wn are
Pm induction,
Pn using (a)
Pmfor P
n
) Next, if u v = 0 then
non-negative then i=0 vi j=0 wj i=0 j=0 vi wj . (
(u v)+ = u (u v) = u, (u v) = (v u)+ = v (v u) = v,
|u v| = (u v)+ + (u v) = u + v = |u + v|,
and
|u + v| = |(u+ + v + ) (u + v )| = u+ + v + + u + v = |u| + |v|.
|u| |v| = |||u| |||v| (|| + ||)|u| (|| + ||)|v| = (|| + ||)(|u| |v|) = 0.
(ii) We may therefore proceed by induction. The case n = 0 is trivial. For the inductive step to n + 1,
setting ui = i ui we have |ui | |uj | = 0 for all i 6= j, by (i-). By (i-),
Pn
Pn
Pn
|un+1 | | i=0 ui | |un+1 | i=0 |ui | i=0 |un+1 | |ui | = 0,
as required.
(c) By 352E,
u+ v + = (u 0) (v 0) = (u v) 0.
Now
1
2
1
2
u
i = ui w,
u
i = ui for i 6= i ,
352Ic
Then
229
Riesz spaces
Pm
i=0
u
i =
vj = vj for j 6= j .
vj = vj w,
Pn
j=0
vj = w w
and u
i vj ui vj for all i, j, while u
i vj = 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, vj = 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 dened and equal to T (u+ ) for every u U ;
(iii) sup{T u, T u} is dened 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
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
352J
Riesz spaces
352J Solid subsets (a) If U is a Riesz space, a subset A of U is solid (in the sense of 351I) i v A
whenever u A and |v| |u|. P
P () 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. () Suppose that A satises
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 i |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}
(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 dened 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 U such that u = p. We have
T (u+ ) = (T u)+ = (Sp)+ = Sp = T u,
u v = hu(i) v(i)iiI ,
u v = hu(i) v(i)iiI ,
|u| = h|u(i)|iiI
352L Theorem Let U be any Riesz space. Then there are a set X, a lter F on X and a Riesz subspace
of the Riesz space RX |F (denition: 351M) which is isomorphic, as Riesz space, to U .
proof By 351Q, we can nd such X and F and an injective Riesz homomorphism T : U RX |F. By
352K, or otherwise, RX is a Riesz space; by 352Jb, RX |F is a Riesz space (recall that it is a quotient of RX
by a solid linear subspace, as explained in 351M); by 352Ib, 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.
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 |i ||ui | | i=0 i ui | i6=j (|i | + |j |)(|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
so v + w A . (iii) If v A
by 352Ea, so w A . Thus A
232
Riesz spaces
352Od
and A = A .
W = V W,
0C = {0},
1C = U ,
V
for V , W C.
W = (V + W ) ,
1C \C V = V ,
W V W
proof To show that C is a Boolean algebra, I use the identication 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,
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 dened.
(d) Finally, if V C is non-empty, then
T
S
V = ( V V V ) C
352S
233
Riesz spaces
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
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 = PW PV e V W ,
234
352S
Riesz spaces
e3 = PW PV e V W ,
e1 + e2 = PV e,
e4 = PW PV e V W ,
e1 + e3 = PW e.
as required.
It follows that
PV W e = e1 + e2 + e3 = (e1 + e2 ) (e1 + e3 ) = PV e PW e,
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 + w , where w W and w W , taken in U ; but as |w| + |w | = |w + w | = |v| V ,
w 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
iI
p+ = (w+ ) = 0,
352W
235
Riesz spaces
proof (a) Let W be the band generated by A and W the set of elements of U satisfying the condition.
(i) If u W 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 W .
(ii) Now W is a band.
supwA |v| w = supwA |v| |u| w = |v| supwA |u| w = |v| |u| = |v|
by 352Ea, so v W .
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}) = sup |u| w + sup |v| w
wA
wA
= |u| + |v|
(351Dc). But this means that supwA (|u| + |v|) w must be |u| + |v|, and |u| + |v| belongs to W ; it follows
from () that u + v belongs to W .
( ) Just as in 352Oc, we now have
nu W for every n N, u W ,
and therefore u W for every R, u W , since |u| |nu| if || n. Thus W is a solid linear
subspace of U .
( ) Now suppose that C (W )+ 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 = supwA u w for
every u C, therefore greater than or equal to v = sup C. Thus v = supwA v w and v W . As C is
arbitrary, W is a band (352Ob). Q
Q
(iii) Since A is obviously included in W , W must include W ; putting this together with (i), W = W ,
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 suciently often for it to be
useful to develop a little of its theory. The following denition 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),
for all u, v, w U and R, and
u (v + w) = (u v) + (u w),
(u v) = (u) v = u (v)
u v 0 whenever u, v 0,
236
Riesz spaces
352W
(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
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 rst four
clauses of the denition in (a) ensure that they are linear operators. If u U , then
|u| v = |u v|,
v |u| = |v u|
353A
237
(j) Let U be a Riesz space and A U + an upwards-directed set. Show that the band generated by A is
{u : |u| = supnN,wA |u| nw}.
> (k) (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 denition 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, , ) is a measure space, show that
L0 () and L () (241, 243) are f -algebras.
(l) Let U R Z 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.
(m) 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
(n) Let U be an f -algebra with multiplicative identity e. Show that u e u2 for every u U , > 0.
352 Notes and comments In this section we begin to see a striking characteristic of the theory of Riesz
spaces: repeated reections 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, 313Qa,
352N, 352Oe, 352Ub, 352Xe). 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 dierences are as important as the agreements. Subspaces of R X 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 nd 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.
238
Riesz spaces
353A
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 i 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|.
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 are distinct bands in U , say W 6 W , then W 6 W , by 353B, so W W 6= {0};
because V is order-dense, V W W 6= {0}, and V W 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 dened for every u U + . Then V is a projection band.
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.
353I
239
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
353G Dedekind completeness Recall that a partially ordered set P is Dedekind ()-complete if
(countable) non-empty sets with upper and lower bounds have suprema and inma in P (314A). For a Riesz
space U , U is Dedekind complete i every non-empty upwards-directed subset of U + with an upper bound
has a least upper bound, and is Dedekind -complete i 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 upwards-directed
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}.
by 352Vb. 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} ,
240
Riesz spaces
353J
proof (a)(i) If hun inN is a non-decreasing sequence in V + with an upper bound v V , then w = supnN un
is dened in U ; but as 0 w v, w V and w = supnN un in V . Thus V is Dedekind -complete.
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 Hausdor space, in such a
way that e corresponds to X; 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 X is an order unit in C(X) = Cb (X).
353M
241
proof (a) Let X be the set of Riesz homomorphisms x from U to R such that x(e) = 1. Dene 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 = X.
(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
> 0 such that (w e)+ 6= 0. Now there is an x X such that x(w) . P
P (i) By 351O, there is a solid
linear subspace V of U such that (w e)+
/ 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 = { : R, (u e)+ V }. Then
Au = 0 (u e)+ (u e)+ V = Au .
Also Au is non-empty and bounded below, because if 0 is such that e u e then Au and
1
/ Au (since (u ( 1)e)+ e
/ V ). (iv) Set x(u) = inf Au for every u U ; then Au for
every > x(u),
/ Au for every < x(u). (v) If u, v U and > x(u), > x(v) then
((u + v) ( + )e)+ (u e)+ + (v e)+ V
(352Fc), so + Au+v ; as and are arbitrary, x(u + v) x(u) + x(v). (vi) If u, v U and < x(u),
< x(v) then
((u + v) ( + )e)+ (u e)+ (v e)+
/ V,
{x : x(u) = x(u)},
{x : x(e) = 1}
as u, v run over U and 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 Hausdor spaces, Q and X are Hausdor (3A3Id).
(d) So T is a Riesz homomorphism from U to C(X). Now T [U ] is a Riesz subspace of C(X), containing
X, 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 = 31 kf k , and let u U be such that
kf T uk ; set v = (u e)+ . Since
0 < (f 2X)+ (T u X)+ f + = f ,
242
Riesz spaces
353M
(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
Z; to simplify the notation, let us suppose that U is actually a subspace of C(Z). Then for each z Z,
we have a functional z : U R dened by setting z(u) = u(z) for every u U ; of course z is a Riesz
homomorphism such that z(e) = 1, that is, z X. Thus we have a function z 7 z : Z X. For any
u U , the function z 7 z(u) = u(z) is continuous, so the function z 7 z is continuous (3A3Ib). If z1 , z2
are distinct members 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 z1 6= z2 . Thus z 7 z 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
|v| V + whenever v V , 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).
353P
243
v v0 u w (u u0 )+ w + u0 w,
so
v ((u u0 )+ w) v + (u0 w) v u0 w,
by 352Fa. Q
Q
Taking the inmum over u, and using the distributive laws (352E), we get
v0 u0 u0 w.
Taking the inmum over , and using the hypothesis that U is Archimedean,
v0 u0 = 0.
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 u = 0 whenever u u = 0. P
P
0 = (u v) u = (u v) (v u ). Q
Q
so u v 6= 0.
244
353P
Riesz spaces
(c) u u1 = e so |u| |u1 | = |e| = e (352W(b-iii)), and |u1 | = |u|1 . (Recall that inverses in any
semigroup with identity are unique, so that we need have no inhibitions in using the formulae u1 , |u|1 .)
Now suppose that u 0. Then u1 = |u1 | 0. If u |v| = 0 then
e |v| = (u u1 ) |v| = 0,
Consequently
1
n
T (u v) nT u = T (u v) nT (u e) = T (u (v ne)) T (u v 2 )
because v 7 T (u v ) is a positive linear operator; as V is Archimedean, inf nN (T (u v) nT u)+ = 0
and T (u v) = supnN T (u v) nT u belongs to W . () If w |T u| = 0 then
so T u T v W
w |T u T v| = w (|T u| |T v|) = 0;
= 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 u = 0 in U , then S1 u S2 u = 0 in V , because (by (ii) just
above) S1 u belongs to the band generated by T u, while S2 u belongs to the band generated by T u , and
T u T u = T (u u ) = 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 = e T v = T v = T (e v) = S2 e.
1 2
nu
for every n 1. So
353 Notes
245
(b) Let X be a completely regular Hausdor space. Show that C(X) is Dedekind complete i Cb (X) is
Dedekind complete i X is extremally disconnected.
(c) Let X be a compact Hausdor space. Show that C(X) is Dedekind -complete i 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 inN 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 -complete.
(e) Let U be an Archimedean Riesz space. Show that the following are equiveridical: (i) U has the
countable sup property (241Ye); (ii) for every A U with an upper bound there is a countable B A such
that A and B have the same upper bounds; (iii) every order-bounded disjoint subset of U + is countable.
(f ) Let U be an Archimedean f -algebra. Show that an element e of U is a multiplicative identity i
e2 = e and e is a weak order unit. (Hint: start by showing that under these conditions, e u = 0 u = 0.)
(g) Let U be an Archimedean f -algebra with a multiplicative identity. Show that if u U then u is
invertible i |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 briey to general Boolean rings.
Note that the standard denition of order-dense in Boolean algebras, as given in 313J, corresponds to
the denition 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 354Yj. The
point is that (as part (d) of the proof of 353M 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 nd 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 dened in terms of Stone representations (363A).
Of the two parts of 353O, it is (a) which is more important for the purposes of this book. The f -algebras
we shall encounter in Chapter 36 can be seen to be commutative for dierent, 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
354 intro.
Riesz Spaces
n=0
n=0
kvn+1 vn k
n=0
kun+1 un k
n=0
kun+1 k + kun k
is nite, and hvn inN is Cauchy. Let u be its limit; because hvn inN 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
and u v. So
is the required supremum.
354E
247
Banach lattices
Pn
i=0
i=0
k|ui |k =
i=0
kui k.
354D 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 = supuA 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 supuA 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 uA kuk = 0 whenever A U is a
non-empty downwards-directed set with inmum 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 > 0 there is a
u A such that k(v u)+ k 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 inmum 0. So inf uA ku0 uk = 0, and u0 A.
(b) If, in (a), A U + , then we must have
(v u)+ = (v u) u w u,
(d) If hun inN is a non-decreasing order-bounded sequence, and > 0, then, applying (c) to {un : n N},
we nd that there is an m N such that kum un k 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
A = {u0 . . . un : u0 , . . . , un A}, so that A is upwards-directed and has the same upper bounds as A.
For each n, choose un A such that k(u un )+ k 2n for every u A . Set vn = supin ui for each n;
then vn A and kvm vn k k(vm un )+ k 2n for all m n. So hvn inN 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
354F
Riesz Spaces
354F Lemma If U is an Archimedean Riesz space with an order unit e (denition: 353L), there is a
Riesz norm k ke dened on U by the formula
kuke = min{ : |u| e}
for every u U .
354G Definitions (a) If U is an Archimedean Riesz space and e an order unit in U , the norm k ke as
dened 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 dened, 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 X. (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 X. (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, , ), the space L () is an M -space with standard order unit X .
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 inN , hwn inN in V such
that vn vn+1 u wn+1 wn and kwn vn ke 2n for every n; so that u = supnN vn = inf nN 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
uu+
1
e
2n+2
wn u +
1
e.
2n+1
Accordingly hvn inN is non-decreasing, hwn inN is non-increasing and kwn vn ke 2n for every n. Because
U is Archimedean, supnN vn = inf nN 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 denition in 352Na.
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 e, where = supuA kuke . This is all that is called for in the Levi property. If moreover sup A
is dened, then sup A e so k sup Ak , as required in the Fatou property.
354O
249
Banach lattices
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 Hausdor space, in such a
way that e corresponds to X 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 X,
and kT uk kuke . On the other hand, if 0 < < kuke then u1 = (|u| e)+ > 0, so that T u1 =
(|T u| X)+ > 0 and kT uk ; as is arbitrary, kT uk kuke .
354L Corollary Any M -space U is isomorphic, as Banach lattice, to C(X) for some compact Hausdor
X, and the isomorphism is essentially unique. X can be identied 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 R U .
proof By 354K, there are a compact Hausdor 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, , ) is any measure space, then L1 (), with its norm k k1 , is an L-space (242D, 242F). In
particular, taking 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 dened in U and belongs to the norm-closure of A in U . P
P For u A,
set (u) = sup{kv uk : v A, v u}. We surely have (u) kuk + supvA kvk < . Choose a sequence
hun inN in A such that un+1 un and kun+1 un k 21 (un ) for each n. Then
kun+1 u0 k =
Pn
i=0
kui+1 ui k
1 Pn
(ui )
2 i=0
P
for every n, using the denition of L-space. Because A is bounded, i=0 (ui ) < and limn (un ) = 0.
But kum un k (un ) whenever m n, so hun inN 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 u A such that u u un ; now (u u )+ u un so k(u u )+ k ku un k (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 inN . 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 inmum 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 uA kuk inf vB ku0 vk = 0. Q
Q
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 > 0 there is a w U +
such that k(|u| w)+ k for every u A.
354Q Since I have already used the phrase uniformly integrable based on a dierent formula, I had
better check instantly that the two denitions are consistent.
Proposition If (X, , ) is any measure space, then a subset of L1 = L1 () is uniformly integrable in the
sense of 354P i 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 > 0 there are M 0,
R
E such that E < and (|u| M E )+ for every u A; now w = M E belongs to (L1 )+ and
k(|u| w)+ k for every u A. As 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 > 0. Then there is a w (L1 )+
such that k(|u| w)+ k 12 for every u A. There is a simple function h : X R such that kw h k 12
(242Mb); now take E = {x : h(x) 6= 0}, M = supxX |h(x)| (I pass over the trivial case X = ), so that
h M E and
(|u| M E )+ (|u| w)+ + (w M E )+ (|u| w)+ + (w h )+ ,
354R
251
Banach lattices
Evidently A C, and C satises the denition 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 | |v| and v C, then
k(|v | w)+ k k(|v| w)+ k supuA k(|u| w)+ k
for every w, and v C. If v = v1 + v2 where v1 , v2 C, [0, 1] and = 1 , then |v| |v1 | + |v2 |,
so
|v| w (|v1 | w) + (|v2 | w) (|v1 | w)+ + (|v2 | w)+
and
uA
for n 1. Given > 0, there is a w U + such that k(|u| w)+ k for every u A, and therefore for
every u in the solid hull of A. Of course supnN kvn wk kwk is nite, so there is an n N such that
kvi wk + kvn wk for every i N. But now, for m > n,
vm
(|um | vn )+ (|um | |um | w)+ + ((|um | w) vn )+
so that
kvm
k k(|um | w)+ k + k(vm w) (vn w)k
using the L-space property of the norm for the equality in the middle. As is arbitrary, limn vn = 0. As
hun inN is arbitrary, condition (ii) is satised; but so is condition (iv), because we know from (a-i) that A
is norm-bounded, and if hun inN is disjoint then vn = |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 > 0 such that supuA k(|u| w)+ k > for every w U + . Consequently
we shall be able to choose inductively a sequence
hun inN in A such that k(|un | 2n supi<n |ui |)+ k > for
P i
every n 1. Because A is norm-bounded, i=0 2 kui k is nite, and we can set
252
Riesz Spaces
354R
P
vn = (|un | 2n supi<n |ui | i=n+1 2i |ui |)+
P
Pm
for each n. (The sum i=n+1 2i |ui | is dened because h i=n+1 2i |ui |imn+1 is a Cauchy sequence. We
have vm |um |,
vm vn (|um | 2n |un |)+ (|un | 2n |um |)+
whenever m < n, so hvn inN is a disjoint sequence in the solid hull of A; while
P
kvn k k(|un | 2n supi<n |ui |)+ k i=n+1 2i kui k 2n supuA kuk
(c) Now this follows at once, because conditions (b-ii) and (b-iv) are satised in V i they are satsied
in U .
r
354X Basic exercises > (a) Work through
pPr the proofs that the following are all Banach lattices: (i) R
Pr
2
with () kxk1 = i=1 |i | () kxk2 =
i=1 |i | () kxk = maxir |i |, where x = (1 , . . . , r ). (ii)
p (X), for any set X and any p [1, ] (242Xa, 243Xl, 244Xn). (iii) Lp (), for any measure space (X, , )
and any p [1, ] (242F, 243E, 244G). (iv) c 0 , the space of sequences convergent to 0, with the norm k k
inherited from .
(b) Let hUi iiI be any family of Banach lattices. Write U for their Riesz space product (352K), and in
U set
P
kuk1 = iI ku(i)k, V1 = {u : kuk1 < },
kuk = supiI 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 R r , 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 i non-decreasing Cauchy
sequences are convergent. (Hint: if kun+1 un k 2n for every n, show that hsupin ui inN is Cauchy,
and that hun inN converges to inf nN supmn um .)
(f ) Let U be a Riesz space with a Riesz norm. Show that U is a Banach lattice i every non-decreasing
Cauchy sequence hun inN 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 supnN un B whenever hun inN is a
non-decreasing sequence in B with a supremum in U . Show that B is closed. (Hint: show rst that u B
whenever there is a sequence hun inN in B U + such that ku un k 2n for every n; do this by considering
vm = inf nm 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 c 0 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 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 -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 iiI be any family of Banach lattices and let V1 , V be the subspaces of U = iI Ui as
described in 354Xb. Show that V1 , V have norms which are Fatou, or have the Levi property, i every Ui
has. Show that the norm of V1 is order-continuous i 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 i 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 i (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
identied with C(X).
(t) Let D R be any non-empty set, and
Pn V 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 (ti1 )| : 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 2n2 kuk for every n, then kv0 inf in vi k 14 kuk for every n, so that 0 cannot
be inf nN 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 inN is a disjoint order-bounded sequence in U + (ii) limn un+1 un = 0 for every orderbounded non-decreasing sequence hun inN in U (iii) whenever A U + is a non-empty downwards-directed
set in U + with inmum 0, inf uA supvA,vu ku vk = 0. (Hint: for (i)(ii), show by induction that
limn un = 0 whenever hun inN is an order-bounded sequence such that, for some xed k, inf iK ui = 0
for every K N of size k; now show that if hun inN is non-decreasing and 0 un u for every n, then
inf iK (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 (denition:
241Ye).
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 satises 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 i it satises
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) satises the conditions of 354Yb i
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 -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, i 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 is Archimedean i 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 -complete i it is uniformly complete and
has the principal projection property (353Xb). (Hint: for (vii), use 353Yc.)
(j) Let U be an Archimedean Riesz space with an order unit, endowed with its order-unit 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.
(k) 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 nitely many values to show that ku + vk = kuk + kvk.)
(l) Let U be a uniformly complete Archimedean Riesz space (354Yi). Set UC = U U with the complex
linear structure dened by identifying (u, v) U U as u + iv UC , so that u = Re(u + iv), v = Im(u + iv)
and ( + i)(u + iv) = (u v) + i(v + u). (i) Show that for w UC we can dene |w| U by setting
|w| = sup||=1 Re(w). (ii) Show that if U is a uniformly complete Riesz subspace of R X for some set X,
then we can identify UC with the linear subspace of CX generated by U . (iii) Show that |w + w | |w| + |w |,
|w| = |||w| for all w UC , 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 i Re w, Im w both belong
to U0 . (vi) Show that if U has a Riesz norm then we have a norm on UC dened 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 (), where
(X, , ) is a measure space and p [1, ], then UC can be identied with LpC () as dened in 242P, 243K
and 244O. (We may call UC the complexification of the Riesz space U .)
(m) Let (X, , ) 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 dened by saying that f g i
f (x) g(x) in V for -almost every x X. (ii) Show that when V = L1 (), for some other measure space
(Y, T, ), then this ordering on L1V agrees with the ordering of L1 () where is the (c.l.d.) product measure
355A
255
on X Y and we identify L1V with L1 (), as in 253Yi. (iii) Show that if V has an order-continuous norm,
so has L1V . (Hint: 354Yd(ii).) (iv) Show that if is Lebesgue measure on [0, 1] and V = , then L1V is
not Dedekind -complete.
354 Notes and comments Apart from some of the exercises, the material of this section is pretty strictly
conned 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 oer a
straightforward route towards a theory of complex Riesz spaces (354Yl). 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 oered by the theory of Riesz spaces are not especially important in such areas, I think we should
always seek connexions between dierent topics. So it is worth remembering that uniformly complete Riesz
spaces have complexications.
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.
256
355B
Riesz spaces
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 , 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 + ||)v1 + v2 , (T1 + T2 )[A] [v, v]; as A is arbitrary, T1 + T2 belongs to L ;
as , 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 denition; 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 nd a un U such that kun k 2n but kT un k n. Now u = supnN |un | is
dened 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 i
T (u + u ) = T u + T u ,
for all u, u U
T (u) = T u
it is elementary to check that this denes 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 dierence of positive linear operators.
(b) For T L , T + and |T | are dened 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
257
for every w U + .
(d) Suppose that A L is non-empty and upwards-directed. Then A is bounded above in L i
{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 i 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 dened 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 (w) = RT w for w U + and > 0, just because u 7 u,
v 7 v 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 dierence of two positive linear operators.
(ii) This shows that every order-bounded operator is a dierence 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 dierences of positive operators.
(iii) Just as in 351F, L is a partially ordered linear space if we say that S T i Su T u for
every u U + . Now it is a Riesz space. P
P Take any T L . Then RT , as dened 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 dened 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 rst formula oered for |T |. In particular, if |u| w then T u,PT 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 of 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)
0uw
0uw
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)
0uw
inf Sw + T w T u S(w u)
0uw
(351Db)
=
inf Su + T (w u)
0uw
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,
It is also easy to see that R(u) = Ru for every u U + and > 0. So, using 355D again, R has an
extension to a linear operator from U to V .
Now if we x 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 and u U + , so R is an upper bound
for A in L ; and, nally, 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 .
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 dened 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 rst that
S(u + u ) = Su + Su , S(u) = Su
355F
259
uB
A(u) is upwards-directed, so
T0 v = sup{T0 (v w) : w
uB
A(u)} supuB T u.
= 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
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 dierence of order-continuous positive linear operators, and
L
c (U ; V ) will be the set of those T L (U ; V ) expressible as the dierence 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 i |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 inmum 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 T belong to L for every T L and
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 inmum 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 ,
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 U0 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 U0 is positive and
order-continuous whenever T is; so T U0 L (U0 ; V ) whenever T L (U ; V ). Because the map T 7 T U0
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 U0 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 |(uv) : v A} = 0 and there is a v A such that T u+|T |(uv) 6 0.
But T v = T u + T (v u) T u + |T |(u v) so T v 6 0 and T U0 6 0. X
XQ
Q
This shows that the map T 7 T U0 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 U0 = S0 . Then S1 = |T0 |U0 is a positive member of W such that S0 S1 and 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 dened 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 inmum 0. Then for
each n N there is a un A such that kun k 4n . By 354C, u = supnN 2n un is dened in U . Now
T un 2n T u for every n, so any lower bound for T [A] must also be a lower bound for {2n 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 denition 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 k 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 (R r ; R s ) can be identied with the
space of real s r matrices, saying that a matrix is positive i every coecient is positive, so that if
T = hij i1is,1jr then |T |, taken in L (R r ; R s ), is h|ij |i1is,1jr . Show that a matrix represents a
Riesz homomorphism i each row has at most one non-zero coecient.
262
355Xf
Riesz spaces
> (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 + .
set. P
Show that A is bounded above in L (U ; V ) i 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
+
wU .
355Y Further exercises (a) Let U and V be Banach lattices. For T L = L (U ; V ), set kT k =
supwU + ,kwk1 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
and
Pn T : U V a linear operator. (i) Show that for any w U ,
PnU and V be Riesz spaces
+
Cw = { i=0 |T ui | : u0 , . . . , un U , i=0 ui = w} is upwards-directed, and has the same upper bounds
as {T u : |u| w}. (Hint: 352Fd.) (ii) Show that if sup Cw is dened for every w U + , then S = T (T )
is dened 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 be the usual measure on {0, 1}N and c the Banach lattice of convergent sequences. Find a
linear operator T : L2 () 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 () 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 dierence of positive operators. For T Lreg let kT kreg be
inf{kT1 + T2 k : T1 , T2 : U V are positive, T = T1 T2 }.
(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 inmum 0 then
{v : v V + , w A, |T u| v whenever |u| w} has inmum 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 denition 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 ) and 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
uA
(i) Show that if U is a Riesz space with the countable sup property (241Ye), then L
c (U ; V ) = L (U ; V )
for every Dedekind complete Riesz space V .
355 Notes
263
(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 U0 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 complexications (354Yl). Show that the complexication of L (U ; V ) can be identied 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 and 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 nite sums j=0 (cos j )wj , j=0 (sin j )wj where
Pn
wj U + , j=0 wj = w. Consequently |T (u + iv)| |T |(w) for every T L
C .)
355 Notes and comments I have had to make some choices in the basic denitions 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 dierences of positive operators (355Ea).
(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 denitively 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 denition 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 pass these questions by.
The elementary extension technique in 355D may recall the denition 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 indenite 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 dierent (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 eective 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
264
356 intro.
Riesz spaces
for every w U + . A non-empty upwards-directed set A U is bounded above i supf A f (u) is nite 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 adjoint P : U
U , dened by setting P (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 f U for every f U (355Bd), and P is
a positive linear operator from U to itself. Set Q = I P , the complementary band projection on U ; then
Q is another positive linear operator on U , and P f + Q f = f for every f . Now P f Q f = 0 for every
f 0. P
P For any w U + ,
(P f Q f )+ (w) = sup (P f Q f )(u) = sup f (P u Qu)
0uw
0uw
= f (P w) = (P f )(w),
so (P f Q f )+ = P f , that is, P f Q f = 0. Q
Q By 352Rd, P 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.
(b) The norm of U is order-continuous i U U .
356E
Dual spaces
265
(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
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 nite, then supf A f (u)
M kuk is nite for every u U + , so g = sup A is dened 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 inmum 0, then
inf uA |f |(u) inf uA 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 inmum 0, such that inf uA kuk = > 0. Set
B = {v : v u for some u A}.
Then B is convex. P
P If v1 , v2 B and [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 = u + (1 )u v1 + (1 )v2 ,
and v1 + (1 )v2 B. Q
Q Also inf vB kvk = > 0. By the Hahn-Banach theorem (3A5Cb), there is an
f U such that inf vB f (v) > 0. But now
inf uA |f |(u) inf uA f (u) > 0
(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 dualsof-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) = sup0 f (v u) for every v U + . Then it is easy to see that g(v) = g(v) for every
v U + , [0, [. If v, w U + then
(v u) + (w u) (v + w) 2u (v 2u) + (w 2u)
for every 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 u v = 0 for every 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 dene u
:V R
by setting u
(f ) = f (u) for every f V . Then u 7 u
is a Riesz homomorphism from U to V .
proof (a) By the denition of addition and scalar multiplication in V , u
is linear for every u; also u
c =
u,
(u1 + u2 )b= u
1 + u
2 for all u, u1 , u2 U and R. If u 0 then u
(f ) = f (u) 0 for every f V + , so
u
0; accordingly every u
is the dierence of two positive functionals, and u 7 u
is a linear operator from
U to V .
(b) If B V is a non-empty downwards-directed set with inmum 0, then inf f B f (u) = 0 for every
u U + , by 355Ee. But this means that u
is order-continuous for every u U + , so that u
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
(
u v)(f ) u
(f g) + v(g) = f (u) g(u) + g(v) = 0.
As f is arbitrary, u
v = 0. As u and v are arbitrary, u 7 u
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 > 0. Then for each n N there is a vn [0, u] such that
f (vn ) + g(u vn ) 2n (355Ec). If v vn for every n N then f (v) = 0 so v = 0; thus inf nN vn = 0.
Set wn = inf in vi for each n N; then hwn inN is non-increasing and has inmum 0 so (because g is
order-continuous) inf nN g(wn ) = 0. But
Pn
u wn = supin u vi i=0 u vi ,
so
g(u wn )
for every n, and
Pn
i=0
g(u vi ) 2
356K
Dual spaces
267
Let g (U )+ be such that (g) > 0 and (g) = 0 whenever = 0 in U . Let v U + be such that
g(v) > 0 and h(v) = 0 whenever h g = 0 in U .
Because v(g) = g(v) > 0, v > 0. As u
, v u
> 0 and v
u > 0. Set w = v u; then w
= v
u,
by 356F, so w
> 0.
?? Suppose, if possible, that w
6 . Then = (w
)+ > 0, so there is an h (U )+ such that (h) > 0
and (h ) > 0 whenever 0 < h h (356H, for the fourth and last time). Now examine
(because w
u, u
, h g g)
(h g) (
uu
)+ (g)
=0
because (
uu
)+ = 0. So h g = 0 and h(v) = 0. But this means that
(h) w(h)
v(h) = 0,
which is impossible. X
X
of U is quasi-order-dense in U , therefore order-dense
Thus 0 < w
. As is arbitrary, the image U
(353A).
. Express as u
(c) Now suppose that U is Dedekind complete and that 0 U
where u U ,
+
and set A = {v : v U, v u , v }. If v U and 0 v , then w = v + u+ A and w
= v; thus
. As and are arbitrary, U
is solid in U .
= sup{
v : v A} = v0 , where v0 = sup A. So 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 i (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
: 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 = {
u : u A} is non-empty and upwards-directed and supB (f ) < for every f U , so sup B is
dened in U (355Ed); but U is a band in U , so sup B U and is of the form w
for some w U .
Because u 7 u
is a Riesz space isomorphism, w = sup A in U . Thus U satises the three conditions.
(b) Suppose that U satises the three conditions. We know that u 7 u
is a Riesz homomorphism onto
an order-dense Riesz subspace of U (356I). It is injective because U separates the points of U . If 0
in U , set A = {u : u U + , u
}. Then A is non-empty and upwards-directed and for any f U
supuA f (u) supuA |f |(u) supuA u
(|f |) (|f |) < ,
so by condition (iii) A has an upper bound in U . Since U is Dedekind complete, w = sup A is dened in U .
Now
w
= supuA u
= .
As is arbitrary, the image of U includes (U )+ , therefore is the whole of U , and u 7 u
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 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 supvA g(v) is nite for every g V ,
then supvA f (v) < for every f U (again because f V V ), so A has an upper bound in U ; because
U is Dedekind complete, sup A is dened in U ; because V is a band, sup A V and is an upper bound for
A in V . Thus V satises 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
(f ) 6= 0, where u
U (356F). Thus U separates the points of U . If A U is non-empty and
upwards-directed and supf A (f ) is nite for every U , then, in particular,
supf A f (u) = supf A u
(f ) <
for every u U , so A is bounded above in U , by 355Ed. Thus U satises 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 satises 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 i 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.
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).
(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 i it is norm-bounded and limn supf A |f (un )| = 0 for every order-bounded disjoint
sequence hun inN in U + .
356O
269
Dual spaces
proof (a) Suppose that A is uniformly integrable. Then it is surely norm-bounded (354Ra). If hun inN is
1
2
1
2
gi (vi ) gi (u) 2,
jK,j6=i
gi (vj )
for i, j K, i 6= j.
) ,
gn (u vi(n)
gn (u vi(n)
)+ gn (u) .
u)+ ,
v = (vm
+
w = (u vm
) .
) M 3 M = ,
gm (v) gm (vm
gn (w) = gn (u vi(n)
)+ gn (u) + ;
since J is innite,
lim supn gn (w) + >
270
Riesz spaces
356O
and w D. Q
Q
(iii) Since e D, we can choose inductively sequences hwn inN in D, hvn inN in C such that w0 = e,
vn wn+1 = 0, vn wn+1 wn for every n N. But in this case hvn inN is a disjoint order-bounded
sequence in [0, u], while for each n N, we can nd fn A such that |fn |(vn ) > 32 . Now there is a
P Set = sup0vvn |fn (v)|. Then fn+ (vn ), fn (vn ) are both less than
un [0, vn ] such that |fn (un )| 13 . P
Q
or equal to , so |fn |(vn ) 2 and > 31 ; so there is a un [0, vn ] such that |fn (un )| 31 . Q
1
Accordingly we have a disjoint sequence hun inN in [0, e] such that supf A |f (un )| 3 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 disjoint
sequence hun inN , A must be uniformly integrable.
This completes the proof.
356P Proposition Let U be an L-space.
(a) U is perfect.
R
dened 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 supuA u is nite, then sup A is dened in U and
sup A = supuA 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,
R
so the norm on U = U is the order-unit norm dened from , 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
vB
uA,uu0
uA,uu0
Z
Z
sup
u + u
0 < .
uA,uu0
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 dened in U . Finally, sup A = supuA 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 i it is relatively
weakly compact.
proof (a) Let A U be a uniformly integrable set.
(i) Suppose that F is an ultralter on X containing A. Then A 6= . Because A is norm-bounded,
supuA |f (u)| < and (f ) = limuF f (u) is dened in R for every f U (2A3Se).
If f , g U then
356X
271
Dual spaces
=U
for every f U and every u A. Accordingly |(f )| |f |(w)+kf 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) . In this case
||(f1 ) = sup|f |f1 |(f )| sup|f |f1 |f |(w) + kf k f1 (w) + kf1 k (1 + kf0 k).
(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
0 = . But now we see that
f (u0 ) = (f ) = limuF 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 ultralter 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 inN be any disjoint 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.
Dene T : U R N by setting T u = hgn (Pn u)inN 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 uinN is a
disjoint sequence in U , and
Pn
Pn
i=0 kPi uk = k
i=0 |Pi u|k = k supin |Pi u|k kuk
kT uk1 =
i=0
|gi Pi u|
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 identied with L1 (), where is counting measure on N. So T [A] is uniformly
integrable in 1 , by 247C, and in particular limn supwT [A] |w(n)| = 0. But this means that
limn kun k limn |gn (Pn vn )| = limn |(T vn )(n)| = 0.
As hun inN is arbitrary, A satises the conditions of 354R(b-iv) and is uniformly integrable.
356X Basic exercises (a) Show that if U = then U
Uc can be identied with 1 , and is properly
P=
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
356Xb
Riesz spaces
(b) Show that if U = C([0, 1]) then U = Uc = {0}. (Hint: show that if f (Uc )+ and hqn inN
enumerates Q [0, 1], then for each n N there is a un U + such that un (qn ) = 1 and f (un ) 2n .)
(c) Let X be an uncountable set, the countable-cocountable measure on X and its domain (211R).
Let U be the space of bounded -measurable real-valued functions on X. Show that U is a Dedekind
-complete Banach lattice Rif given the supremum norm k k . Show that U can be identied with 1 (X)
(cf. 356Xa), and that u 7 u d belongs to Uc \ U .
(d) Let U be a Dedekind -complete Riesz space and f Uc . Let hun inN be an order-bounded sequence
in U which is order-convergent to u U in the sense that u = inf nN supmn um = supnN inf mn 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 dened by the formula
(P f )(u) = inf{sup f (v) : A U is non-empty, upwards-directed
vA
for every f (U ) , u U . (Hint: show that the formula for P f always denes 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 dened by the formula
(P f )(u) = inf{supnN f (vn ) : hvn inN 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 i 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 rst principles that if hun inN is a sequence in 1 such that |un (n)| > 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 supuA |f (u)| for every non-empty downwards-directed set B U with
inmum 0 (iii) inf nN supuA |fn (u)| = 0 for every non-increasing sequence hfn inN in U with inmum 0
(iv) A is norm-bounded and limn supuA |fn (u)| = 0 for every disjoint order-bounded sequence hfn inN
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
) , and let g, h be
A = {A0 + . . . + An : A0 , . . . , An A}. Show that UA
= UA
. (iii) Take any f (U
the components of f in UA
, (UA
) respectively. Show that
+
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 i 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 i 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 inN a sequence in U such that hf (un )inN is Cauchy for every f U .
Show that hun inN 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 inN a sequence in U such
that hf (un )inN is Cauchy for every f U . Show that hun inN is convergent for the weak topology of U .
(Hint: set (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 Vg belongs to Vg ; as g is arbitrary,
V and may be identied with an element of U .)
(h) Let U be a uniformly complete Archimedean Riesz space with complexication V (354Yl). (i) Show
that the complexication of U can be identied with the space of linear functionals f : V C such that
sup|v|u |f (v)| is nite for every u U + . (ii) Show that if U is a Banach lattice, then the complexication
of U = U can be identied (as normed space) with V . (See 355Yk.)
(i) Let U be a Banach lattice with an order-continuous norm with the Levi property. Show that the
family of closed balls in U is a compact class. (Hint: 342Yc.)
356 Notes and comments The section starts easily enough, with special cases of results in 355 (356B).
When U has a Riesz norm, the identication 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 diculty at rst. 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 rst 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 (365N). Further
examples will be discussed in 369 and 374. Of course we have to remember 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 rst is yet another disjoint-sequence characterization of uniformly integrable sets, to go with 246G
and 354R. The essential dierence here is that we are looking at disjoint 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 disjoint sequences as its dual. (For instance, in the dual of C([0, 1]) we have all
the point masses t , where t (u) = u(t); these form a disjoint family in C([0, 1]) not corresponding to any
disjoint family in C([0, 1]).) The essence of the proof is a device to extract a disjoint sequence in U to match
approximately a subsequence of a given disjoint sequence in U . In the example just suggested, this would
274
Riesz spaces
356 Notes
correspond, given a sequence htn inN of distinct points in [0, 1], to nding a subsequence htn(i) iiN which is
discrete, so that we can nd disjoint 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
asked, in Volume 2, to master a complex argument (one of the more dicult 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 rst point is that the result is suciently 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
rst, 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 Kakutanis 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 denition I chose of uniform
integrability is exactly adapted to showing that uniformly integrable sets are relatively compact in the
order-continuous bidual; all the eort 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
identication is not always valid, so that I needed to reduce the problem to the -nite case (part (b-ii) of
the proof of 247C). After that, the Radon-Nikod
ym theorem did the trick. Actually Kakutanis theorem
shows that the side-step to -nite spaces is irrelevant. It directly represents an abstract L-space as L1 ()
for a localizable measure , in which case (L1 )
= L exactly.
In the other direction, both arguments depend on a disjoint-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 nd 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 signicant
coordinate of uk(j) ), is easier to nd.
We now have a wide enough variety of conditions equivalent to uniform integrability for it to be easy to
nd 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 i every sequence in A has a weak cluster point. But this extra
renement 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 i every sequence in A has a cluster point in U for the weak topology (Eberleins
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 terms 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 exemplied by 364. If two measure spaces (X, , ) and (Y, T, ) have isomorphic measure algebras
(A,
), (B, ) then the spaces L0 (), L0 () are isomorphic as topological f -algebras; and more: for any
isomorphism between (A,
) and (B, ) 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 () will correspond to g L0 () if and only if [[f > ]] = {x : f (x) > } A
corresponds to [[g > ]] = {y : g(y) > } B for every . But the check that this formula is consistent,
and denes an isomorphism of the required kind, involves a good deal of detailed work. It turns out, in fact,
that the measures and do not enter this part of the argument at all, except through their null ideals
(used in the construction of A and B). This is already evident, if you look for it, in the theory of L0 (); in
241, as written out, you will nd 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 forty 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 -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: rst, 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 eort needed I ought to point to some mathematically signicant 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
-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,
) and
(B, ) are measure algebras, then any measure-preserving Boolean homomorphism from A to B produces
norm-preserving Riesz homomorphisms from Lp (A,
) to Lp (B, ) 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-nite measure
algebra (A,
) to a measure algebra (B, ), we have an operator from L1 (B, ) to L1 (A,
). 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 terms of universal mapping theorems: for
instance, each Lp space, for 1 p , can be characterized as Banach lattice in terms of factorizations of
functions of an appropriate class from the underlying algebra to Banach lattices.
Now let me try to sketch a route-map for the journey ahead. I begin with two sections on the space
276
Function spaces
Chap. 36 intro.
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 dierent 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 account
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 members of
A (361L). Its most important property is the universal mapping theorem 361F, which establishes a one-toone correspondence between (nitely) 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
dened, 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 terms of the ring of elements of nite measure in a measure
algebra than in terms 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 Hausdor 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 : A B is a function, then the following are equiveridical:
(i) is a ring homomorphism; (ii) (a \ b) = a \ b for all a, b A; (iii) 0 = 0 and (a b) = a b,
(a b) = a b for all a, b A. P
P See 312H. To prove (ii)(iii), observe that if a, b A then
(a b) = a \ (a \ b) = a b. Q
Q
(c) If A and B are Boolean rings and : A B is a ring homomorphism, then is order-continuous
i inf [A] = 0 whenever A A is non-empty and downwards-directed and inf A = 0 in A; while is
sequentially order-continuous i inf nN an = 0 whenever han inN is a non-increasing sequence in A with
inmum 0. (See 313L.)
(d) The following will be a particularly important type of Boolean ring for us. If (A,
) is a measure
algebra, then the ideal Af = {a : a A,
a < } is a Boolean ring in its own right. Now suppose that
(B, ) is another measure algebra and Bf B the corresponding ring of elements of nite measure. We
can say that a ring homomorphism : Af Bf is measure-preserving if a =
a for every a Af .
f
Now in this case is order-continuous. P
P If A A is non-empty, downwards-directed and has inmum 0,
then inf aA
a = 0, by 321F; but this means that inf aA a = 0, and inf [A] = 0 in Bf . Q
Q
361D
277
361B Definition Let A be a Boolean ring and U a linear space. A function : A U is finitely
additive, or just additive, if (a b) = a + b whenever a, b A and a b = 0.
361C Elementary facts We have the following immediate consequences of this denition, corresponding
to 326B and 313L. Let A be a Boolean ring, U a linear space and : A U an additive function.
(a) 0 = 0 (because 0 = 0 + 0).
(b) If a0 , . . . , am are disjoint in A, then (supjm aj ) =
Pm
j=0
aj .
aA
aA
aA
(by 351Db)
= c.
Similarly, if A A is a non-empty downwards-directed set with inmum c, then
Similarly, if han inN is a non-increasing sequence in A with inmum c, then han \ cinN is non-increasing
and has inmum 0, so that
inf nN an = inf nN c + (c \ an ) = c + inf nN (c \ an ) = c.
Thus is sequentially order-continuous. Q
Q
361D Construction Let A be a Boolean ring, and Z its Stone space. For a A write a 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 {a : a A}. Because a 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
a b = (a b) for all a, b A (writing for pointwise multiplication of functions, as in 281B), S(A) is
closed under .
278
Function spaces
361E
361E I give a portmanteau proposition running through the elementary, mostly algebraic, properties of
S(A).
Proposition Let A be a Boolean ring, with Stone space Z. Write S for S(A).
(a) If a0 , . . . , an A, there are disjoint 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 j bj where b0 , . . . , bm are disjoint members of A and
j R for each j. If all the bj are non-zero then kuk = supjm |j |.
Pm
(c) If u S is non-negative, it is expressible in the form j=0 j bj where b0 , . . . , bm are disjoint members
Pm
of A and j 0 for each j, and simultaneously in the form j=0 j cj where c0 c1 . . . cm and j 0
for every j. P
m
(d) If u = j=0 j bj where b0 , . . . , bm are disjoint members of A and j R for each j, then |u| =
Pm
j=0 |j |bj 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 1 and kuk = min{ : 0, |u| 1} for every u S.
(f) The map : A S is injective, additive, non-negative, a lattice homomorphism and order-continuous.
(g) Suppose that u 0 in S and 0 in R. Then
is dened in A, and
In particular, u kuk [[u > 0]] and there is an > 0 such that [[u > 0]] u. If u, v 0 in S then
u v = 0 i [[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 i u = v v for some v S.
proof Write b
a for the open-and-compact subset of Z corresponding to a A.
(c)(i) If u 0P
in (b), we must have j = u(z) 0 whenever z bbj , so that j 0 whenever bj 6= 0;
m
consequently u = j=0 |j |bj is in the required form.
(ii) If we suppose that every j is non-negative, and rearrange the terms of the sum so that 0 . . .
m , then we may set 0 = 0 , j = j j1 for 1 j m, cj = supij bi to get
Pm
Pm Pi
Pm Pm
Pm
i=0 i bi = u.
j=0 j bi =
i=0
i=j j bi =
j=0
j=0 j cj =
(d) is trivial, because bb0 , . . . , bbn are disjoint.
(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 1, the constant function with value 1, belongs to S, and is an order
unit of S; while
for every u S.
361F
279
(f ) is injective because b
a 6= bb whenever a 6= b. 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 E : PZ RZ
are. To see that is order-continuous, take a non-empty downwards-directed A A with inmum 0. ??
Suppose, if possible, that {a : a A} does not have P
inmum 0 in S. Then there is a u > 0 in S such that
m
u a for every a A. Now u can be expressed as j=0 j bj where b0 , . . . , bm are disjoint. There must
be some z0 Z such that u(z0 ) > 0; take j such that z0 bbj , so that bj 6= 0 and j = u(z0 ) > 0. But now,
for any z bbj , a A,
(a)(z) u(z) = j > 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 [A] = 0 in S. As A is arbitrary, is order-continuous, by the criterion of 361C(f-i).
Pm
(g) Express u as j=0 j bj where b0 , . . . , bm are disjoint and every j 0. Then given 0, > 0,
a A we have ( + )a u i a sup{bj : j m, j + }. So [[u > ]] = sup{bj : j m, j > }.
Writing c = [[u > ]], d = [[u > 0]] = sup{bj : j > 0}, we have
u(z) kuk if z b
c,
b
if z d \ b
c,
b
= 0 if z
/ d.
So
c u kuk c d,
Then
min(, )([[u > 0]] [[v > 0]]) u v max(kuk , kvk )([[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 j bj where b0 , . . . , bm are disjoint and j 0 for every j, then u = v v where
p
Pm
v = j=0 j bj .
361F I now turn to the universal mapping theorems which really dene 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 : A U and linear operators T : S(A) U , given by the formula = T .
proof (a) The core of the proof isP
the following observation. Let : A
U be additive. If a0 , . . . , an A
P
n
n
and 0 , . . . , n R are such that i=0 i ai = 0 in S = S(A), then i=0 i ai = 0 in U . P
P As in 361E,
we can nd disjoint b0 , . . P
. , bm such that eachPai is the supremum of some of the
b
;
set
=
1 if bj ai , 0
ij
Pn j
m
m
otherwise, so that ai = j=0 ij bj , ai = j=0 ij bj for each i. Set j = i=0 i ij , so that
Pn
Pm
0 = i=0 i ai = j=0 j bj .
Now j bj = 0 in U for each j, because either bj = 0 and bj = 0, or there is some z bbj so that j must
be 0. Accordingly
280
361F
Function spaces
0=
Pm
j=0
j bj =
Pm Pn
j=0
i=0
i ij bi =
Pn
i=0
i ai . Q
Q
Pn
Pm
(b) It follows that if u S is expressible simultaneously as i=0 i ai = j=0 j bj , then
Pn
Pm
i=0 i ai +
j=0 (j )bj = 0 in S,
so that
Pn
i=0
and
i ai +
Pn
i=0
Pm
j=0 (j )bj
i ai =
Pm
j=0
= 0 in U ,
j bj .
whenever a0 , . . . , an A and 0 , . . . , n R.
(c) It is now elementary to check that T is linear, and that T a = a for every a A. Of course this
last condition uniquely denes T , because {a : a A} spans the linear space S.
361G Theorem Let A be a Boolean ring, and U a partially ordered linear space. Let : A U be an
additive function, and T : S(A) U the corresponding linear operator.
(a) is non-negative i T is positive.
(b) In this case,
(i) if T is order-continuous or sequentially order-continuous, so is ;
(ii) if U is Archimedean and 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) a b = 0 in U whenever a b = 0 in A;
(iii) is a lattice homomorphism.
proof Write S for S(A).
(a) If T is positive, then surely a = T a 0 for P
every a A, so = T is non-negative. If is
m
non-negative, and u 0 in S, then u is expressible as j=0 j bj where b0 , . . . , bm A and j 0 for
every j (361Ec), so that
Pm
T u = j=0 j bj 0.
Thus T is positive.
361H
281
T u T (cu + a0 ) = cu + a0 ,
and w 6 T u. Since w is arbitrary, this means that 0 = inf T [A]; as A is arbitrary, T is order-continuous.
) The argument for sequential order-continuity is essentially the same. Suppose that is se(
quentially order-continuous and that hun inN is a non-increasing sequence in S with inmum 0. Again set
= ku0 k, a0 = [[u0 > 0]]; again we may suppose that > 0; again take any w U such that w 6 0. As
before, there is some > 0 such that w 6 a0 . For n N set cn = [[un > ]], so that
cn un cn + a0 .
The sequence hcn inN is non-increasing because hun inN is, and if c cn for every n, then c un for
every n, so is zero, and c = 0 in A. Thus inf nN cn = 0 in A, and inf nN cn = 0 in U , because is
sequentially order-continuous. Replacing A , C in the argument above by {un : n N}, {cn : n N} we
nd an n such that w 6 T un . Since w is arbitrary, this means that 0 = inf nN T un ; as hun inN is arbitrary,
T is sequentially order-continuous.
(c)(i)(iii) If T : S(A) U is a Riesz homomorphism, and = T , then surely is a lattice homomorphism because T and are.
(iii)(ii) is trivial.
Pm (ii)(i) If a b = 0 whenever a b = 0, then for any u S(A) we have an expression of u as
j=0 j bj , where b0 , . . . , bm A are disjoint. Now
Pm
Pm
Pm
|T u| = | j=0 j bj | = j=0 |j |bj = T ( j=0 |j |bj ) = T (|u|)
361H Theorem Let A be a Boolean ring and U a Dedekind complete Riesz space. Suppose that
: A U is an additive function and T : S = S(A) U the corresponding linear operator. Then
T L = L (S; U ) i {b : b a} is order-bounded in U for every a A, and in this case |T | L
corresponds to || : A U , dened by setting
n
X
|ai | : a0 , . . . , an
||(a) = sup{
j=0
= sup{b (a \ b) : b
a are disjoint}
a}
for every a A.
a, we have b a so
|b| = |T b| |T |(a).
a} is order-bounded in U .
(b) Now suppose that {b : b a} is order-bounded in U for every a A. Then for any a A we can
dene wa = sup{|b| : b a}; in this case, b (a \ b) 2wa whenever b a, so a = supb a b (a \ b)
is also dened in U . Considering b = a, b = 0 we see that a |a|. Next, : A U is additive. P
P Take
a1 , a2 A such that a1 a2 = 0; set a0 = a1 a2 . For each j 2 set
Then A0 A1 + A2 , because
Aj = {(aj b) (aj \ b) : b A} U .
282
361H
Function spaces
We therefore
Pn have a corresponding positive operator T1 : S U such that = T1 . But we also see that
a = sup{ i=0 |ai | : a0 , . . . , an a are disjoint} for every a A. P
P If a0 , . . . , an are disjoint and included
in a, then
Pn
Pn
i=0 |ai |
i=0 ai = (supin ai ) a.
On the other hand,
Pn
a supb a |b| + |(a \ b)| sup{ i=0 |ai | : a0 , . . . , an
a are disjoint}. Q
Q
It follows that T L
P Take any u 0 in S. Set a = [[u > 0]] (361Eg) and = kuk . If 0 < |v| u,
P. nP
then v is expressible as i=0 i ai where a0 , . . . , an are disjoint and no i nor ai is zero. Since |v| a,
we must have |i | , ai a for each i. So
Pn
Pn
Pn
|T v| = | i=0 i ai | i=0 |i ||ai | i=0 |ai | = a.
(c) Thus T L i is order-bounded on the sets {b : b a}, and in this case the two formulae oered
for || are consistent and make || = . Finally, = |T |. P
P Take a A. If a0 , . . . , an a are disjoint,
then
Pn
Pn
Pn
i=0 |ai | =
i=0 |T ai |
i=0 |T |(ai ) |T |(a);
so a |T |(a). On the other hand, the argument at the end of (b) above shows that |T |(a) a for
every a. Thus |T |(a) = a for every a A, as required. Q
Q
361I Theorem Let A be a Boolean ring, U a normed space and : A U an additive function. Give
S = S(A) its norm k k , and let T : S U be the linear operator corresponding to . Then T is a bounded
linear operator i {a : a A} is bounded, and in this case kT k = supa,bA ka bk.
proof (a) If T is bounded, then
ka bk = kT (a b)k kT kka bk kT k
(iii) Now suppose that is bounded; set 0 = supaA kak < . Then
= supa,bA ka bk 20
is also nite. 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
Pn
Pm
u+ = i=0 i ci , u = j=0 j cj
Pm
Pn
where all the i , j are non-negative and i=0 i = j=0 j = 1. Take hij iim,jn from (ii). Set cij = ci ,
cij = cj for all i, j, so that
Pm Pn
Pm Pn
u+ = i=0 j=0 ij cij , u = i=0 j=0 ij cij ,
u=
Pm P n
i=0
j=0 ij (cij
cij ),
361J
283
Tu =
kT uk
Pm Pn
Pm Pn
i=0
j=0 ij (cij
j=0 ij kcij
i=0
cij k
cij ),
Pm Pn
i=0
j=0 ij
= .
proof (a) The map : A S(B) is additive (361Cc), so corresponds to a linear operator T = T : S(A)
S(B), by 361F. P
and are both lattice homomorphisms, so also is, and T is a Riesz homomorphism
n
(361Gc). If u = i=0 i ai , where a0 , . . . , an are disjoint, then look at I = {i : i n, ai 6= 0}. We have
Pn
P
T u = i=0 i (ai ) = iI i (ai )
and a0 , . . . , an are disjoint, so that
where u =
iI
i ai , so that T u = T u. If a, a A, then
T (a a ) = T (a a ) = (a a ) = a a = T a T a ,
so T is multiplicative.
where I = {i : i = 1}. As b is arbitrary, is surjective. Of course the formula for kvk is a consequence
of the formula for kT uk in (a).
(c)(i) If [[|u| > 0]] = 0 then |u| a, where = kuk , a = [[|u| > 0]], so
284
Function spaces
361J
for any a A. But T is order-continuous, by 361Ef, so inf aA T (a) = 0, and b must be 0. As b is arbitrary,
inf aA a = 0; as A is arbitrary, is order-continuous. (ii) If is order-continuous, so is : 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 inN is a non-increasing sequence in A with inmum
0, let b be any lower bound for {an : n N} in B. Then
b (an ) = T (an )
w = (w a) + (w (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 1 = 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 = a and
v = (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 a Vb Va Vb ,
361Xb
285
Of course : (X) is additive, so by 361F there is a linear operator T : S (X), writing S for
S(), such that T (E)
= E for every E .
Pm
If u S, T u 0 i u 0. P
P Express u as
j where E0 , . . . , Em are disjoint. Then
j=0 j E
Pm
T u = j=0 j Ej , so
But this means () that
u 0 j 0 whenever Ej 6= T u 0. Q
Q
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 {E : E } () that T is an order-isomorphism between S and S.
Because E F = 0 whenever E, F and E F = , T is a Riesz homomorphism and S is a Riesz
subspace of (X) (361Gc). Now
kuk = inf{ : |u| X}
= inf{ : |T u| X} = kT uk
T (E
F
) = T ((E
F )) = (E F ) = T (E)
T (F
)
for all E, F , so S is closed under pointwise multiplication and the multiplications of S, S are identied
under T .
361M Proposition Let X be a set, a ring of subsets of X, and I an ideal of ; write A for the
quotient ring /I. Let S be the linear span of {E : E } 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 identied with the quotient Riesz space S/V , if
for every E we identify (E ) S(A) with (E) 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{ : {x : |f (x)| > } 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(). Now the canonical ring homomorphism E 7 E
corresponds to a surjective Riesz homomorphism T from S() to S(A) which takes E to (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 gV kf + gk = inf gV kf gk = 0,
Pn
i=0
i Ei where
361X Basic exercises (a) Let A be a Boolean ring and U a linear space. Show that a function : A U
is additive i 0 = 0 and (a b) + (a b) = a + b 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 (w z) = (w) z = w (z) for all w, z U and all R. Let A be a
Boolean ring, : A U an additive function and T : S(A) U the corresponding linear operator. Show
that T is multiplicative i (a b) = a b 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 : 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 + : A U , where + a = supb a b 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 : A B and Riesz homomorphisms T : S(A) S(B) such that T (1A ) = 1B ,
given by setting T (a) = (a) 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 : A B such that T (a) = (a) 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 : A B and a non-negative v S(B) such that T (a) = v (a) 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, a -algebra of subsets of X. Show that the linear span S in RX of {E : E }
is just the set of -measurable functions f : X R which take only nitely many values.
(j) For any Boolean ring A, we may dene 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.
(k) Let A be a Boolean algebra, : A A a Boolean homomorphism and T : S(A) S(A) the
associated Riesz homomorphism. Set C = {a : a = a}. Show that S(C) may be identied with the linear
subspace of S(A) generated by {c : c C}, and that this is {u : u S(A), T u = u}.
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 i ai where 0 , . . . , n R and a0 , . . . , an A. Let W V be the linear subspace spanned
by members of V of the form (a b) a b where a, b A are disjoint. Dene : A V /W by taking a
to be the image in V /W of a V . Show, without using the axiom of choice, that the pair (V /W, ) has the
universal mapping property of (S(A), ) 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 dened in L , and that = T. Show that for any a A,
Pn
a = sup{ i=0 Ti (ai ) : T0 , . . . , Tn A, a0 , . . . , an a are disjoint, supin ai = a}.
(c) Let A be a Boolean algebra. Show that the algebra of all bands of S(A) can be identied with the
Dedekind completion of A (314U).
(d) Let A be a Boolean ring, and U a complex normed space. Let : 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 j aj k : a0 , . . . , an A are disjoint, |j | = 1 for every j}
if either is nite.
361 Notes
287
(e) Let U be a Riesz space. Show that it is isomorphic to S(A), for some Boolean algebra A, i it has
an order unit and every solid linear subspace of U is a projection band.
(f ) Let hAi iiI be a non-empty family of Boolean algebras, with free product A; write i : Ai A for
the canonical maps, and
C = {inf jJ j (aj ) : J I is nite, aj Aj for every j J}.
c = (c i (a)) + (c i (1 \ a))
288
Function spaces
362 intro.
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 rst step is just to check the identication of the linear
and order structures of S and the space M of bounded nitely 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
| |, , 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 nitely additive real-valued functionals on A may be identied
with the partially ordered linear space of all real-valued linear functionals on S.
(b) The linear space of bounded nitely additive real-valued functionals on A may be identied with the
L-space S of order-bounded linear functionals on S. If f S corresponds to : A R, then f + S
corresponds to + , where
+ a = supb a b
for every a A, and
kf k = supaA a (1 \ a).
(c) The linear space of bounded countably additive real-valued functionals on A may be identied with
the L-space Sc .
(d) The linear space of completely additive real-valued functionals on A may be identied 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 f : A R, given by setting f = f .
(a) Now it is clear that f +g = f + g , f = f for all f , g and , 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 f is
non-negative i 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
(iii) To check the correspondence between f + and f+ , rene the arguments of (i) and (ii) as follows.
Take any f S . If a A,
362B
289
(356N)
(by (iii) just above)
aA
M becomes an L-space.
(b) M and M are projection bands in M , therefore L-spaces in their own right. In particular, || M
for every M , and || M for every M .
(c) The band projection P : M M is dened by the formula
(P )(c) = inf{supnN an : han inN is a non-decreasing sequence with supremum c}
290
362B
Function spaces
whenever c A and 0 in M .
(d) The band projection P : M M is dened by the formula
whenever c A and 0 in M .
(e) If A M is upwards-directed, then A is bounded above in M i {1 : A} is bounded above in R,
and in this case (if A 6= ) sup A is dened by the formula
(sup A)(a) = supA a for every a A.
0.
(ii) Now suppose that , 0, and let 1 , 2 be the components of in the band generated by and
its complement. Then
1 c = sup>0 inf a (c \ a),
for every c A.
proof (a) Of course = + ( )+ , = ( )+ , || = () (352D), so the formula of
362Ab gives
( )(a) = a + sup b b = sup b + (a \ b),
ba
ba
ba
b,c a
b,c a
b,c a
The formula oered for kk 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 .
in S(A) (356B); so that M and M are projection bands in M , and are L-spaces in their own right (354O).
for every c A. Then of course 0 c c for every c. The point is that is countably additive. P
P Let
hci iiN be a disjoint sequence in A, with supremum c. Then for any > 0 we have non-decreasing sequences
han inN , hain inN , for i N, such that
supnN an = c,
supnN an c + ,
supnN ain ci + 2i for every i N.
Set bn = supin ain for each n; then hbn inN is non-decreasing, and
so
362B
c sup bn = sup
nN
n
X
i=0 nN
ain
nN i=0
sup ain
291
ci + 2i =
ci + 2.
i=0
i=0
On the other hand, han ci inN is a non-decreasing sequence with supremum c ci = ci for each i, so
ci supnN (an ci ), and
X
i=0
ci
i=0 nN
nN i=0
(an ci )
sup an
nN
c + .
P
As is arbitrary, c = i=0 ci ; as hci iiN is arbitrary, is countably additive. Q
Q
Thus M . On the other hand, if M and 0 , then whenever c A and han inN is a
non-decreasing sequence with supremum c,
c = supnN an supnN an .
= sup{ : M , } = P ,
as claimed.
(d) The same ideas, with essentially elementary modications, deal with the completely additive part.
Take any 0 in M . Set
c = inf{supaA a : A is a non-empty upwards-directed set with supremum c}
for every c A. Then of course 0 c c for every c. The point is that is completely additive.
P
P Note rst that if c A, > 0 there is a non-empty upwards-directed A, with supremum c, such that
supaA a c + c; for if c = 0 we can take A = {c}. Now let hci iiI be a partition of unity in A. Then
for any > 0 we have non-empty upwards-directed sets A, Ai , for i I, such that
sup A = 1,
sup Ai = ci for i I,
supaA a 1 + 1,
Set
S
sup B = sup( iI Ai ) = 1,
so
X
1 sup b = sup{
ai : J I is nite, ai Ai i J}
bB
X
iI
iJ
ci + ci 1 +
ci .
iI
On the other hand, Ai = {a ci : a A} is a non-empty upwards-directed set with supremum ci for each i,
so ci supaAi a, and
292
362B
Function spaces
X
iI
ci
X
iI
sup (a ci ) = sup
aA
aA
X
iI
(a ci )
sup a 1 + 1.
aA
P
As is arbitrary, c = iI ci ; as hci iiI is arbitrary, is completely additive, by 326N. Q
Q
Thus M . On the other hand, if M and 0 , then whenever c A and A is a non-empty
upwards-directed set with supremum c,
c = supaA a supaA a
= sup{ : M , } = P ,
as claimed.
So
+ 1 < . Q
Q
(because A is upwards-directed)
= a + b. Q
Q
Also a for every a, so
362B
293
)( ) is trivial.
(
) Observe rst that if hck ikN is a non-increasing sequence in A such that limk ck = 0,
( )(
then limk + ck = 0. P
P Let > 0. Because + = 0, there is a b A such that + b + (1 \ b) ,
by part (a). Now hck \ bikN is non-increasing and limk (ck \ b) = 0, so limk (ck \ b) = 0 and
lim sup + ck = lim sup + (ck b) + (ck \ b) + (ck \ b)
k
+ b + (1 \ b) .
As is arbitrary, limk + ck = 0. Q
Q
?? Now suppose, if possible, that + does not belong to the band generated by . Then there is a 1 > 0
such that 1 + and 1 || = 0. Set = 14 1 1 > 0. For each n N, we can choose an A such that
||an + 1 (1 \ an ) 2n , by part (a) again. For n k, set bkn = supkin ai ; then
Pn
||bkn i=k ||ai 2k+1 ,
and hbkn ink is non-decreasing. Set k = supnk 1 bkn and choose m(k) k such that 1 bk,m(k) k 2k .
Setting bk = bk,m(k) , we see that bk bk+1 = bkn where n = max(m(k), m(k + 1)), so that
1 (bk bk+1 ) k 1 bk + 2k
Pk1
i=0
2i 2
+ ck 1 ck 1 bk 2 1 ak 2 = 1 1 1 (1 \ ak ) 2 4 2 =
for every k N. On the other hand, hck ikN is a non-decreasing sequence and
||ck ||bk 2k+1
Then
1 = sup>0 inf a c,a (c \ a) = c 2 .
Take any > 0. Because 1 belongs to the band generated by , part (i) tells us that there is a > 0 such
that 1 a whenever a . In this case, if a ,
(c \ a) = c (c a) c 1 c ;
thus
1 inf a (c \ a) 1 c .
294
Function spaces
362C
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 , like , is nite-valued. If we relax this we get an alternative version of
some of the same ideas.
Theorem Let (A,
) be a measure algebra. Write S = S(A), and let M be the Riesz space of bounded
nitely additive real-valued functionals on A. Write
(see 327A),
a . If now
a ,
|a| | a| + ( )(a) 12 + ( )(1) .
(b)(i) We know that Mtc consists just of those M which are continuous at 0 (327Bc). Of course this
is a linear subspace of M .
(ii) If Mtc , M and | | || then || Mtc . P
P Write Af = {d : d A,
d < }. Given > 0
1
f
there are d A , > 0 such that |a| 2 whenever
(a d) . Now
| a| | |(a) ||(a) 2 supc a |c|
whenever
(a d) . As is arbitrary, is continuous at 0 and belongs to Mtc . Q
Q
(iii) If A Mtc is non-empty and upwards-directed and = sup A in M , then Mtc . P
P Let > 0.
Then there is a A such that 1 1 + 12 (362Be). There are d Af , > 0 such that |a| 21
whenever
(a d) . If now
(a d) ,
|a| | a| + ( )(a) 21 + ( )(1) .
|1 |(1 \ a) .
362Xb
295
P
(iii) If A Mt+ is non-empty and upwards-directed and = sup A is dened in M , then Mt . P
||1 = 1 = sup A 1 = sup A,a< a = supa< a.
As A is arbitrary, Mt . Q
Q Thus Mt is a band in M .
362D For semi-nite measure algebras, among others, the formula of 362Bd takes a special form.
Proposition Let A be a weakly (, )-distributive Boolean algebra. Let M be the space of bounded nitely
additive functionals on A, M M the space of completely additive functionals, and P : M M the
band projection, as in 362B. Then for any M + and c A there is a non-empty upwards-directed set
B A with supremum c such that (P )(c) = supbB b; that is, the inf in 362Bd can be read as min.
proof By 362Bd, we can nd for each n a non-empty upwards-directed An , with supremum c, such that
supaAn a (P )(c) + 2n . Set
B = {b : for every n N there is an a Bn such that b
a};
then sup B = c (316H(iv)); of course B is upwards-directed. Using 362Bd again and referring to the choice
of An ,
(P )(c) supbB b supaAn a (P )(c) + 2n
362E Uniformly integrable sets The spaces S , Sc and S of 362A, or, if you prefer, the spaces
M , M , M , 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 nitely additive functionals on A.
Then a norm-bounded set C M is uniformly integrable i limn supC |an | = 0 for every disjoint
sequence han inN in A.
proof Write C for the set {f : f S(A) , f C}. Because the map f 7 f is a normed Riesz space
isomorphism between S and M , C is uniformly integrable in M i C is uniformly integrable in S .
(a) Suppose that C is uniformly integrable and that han inN is a disjoint sequence in A. Then han inN
is a disjoint order-bounded sequence in S , while C is uniformly integrable, so limn supf C |f (an )| = 0,
by 356O; but this means that limn supC |an | = 0. Thus the condition is satised.
(b) Now suppose that C is not uniformly integrable. By 356O, in the other direction, there is a disjoint
sequence hun inN in S such that 0 un 1 for each n and lim supn supf C |f (un )| > 0. For each n,
take cn = [[un > 0]] (361Eg); then 0 un cn and hcn inN is disjoint. Now
lim sup sup ||(cn ) = lim sup sup |f |(cn )
n C
n f C
n f C
So if we choose n C such that |n |(cn ) supC ||(cn ), we shall have lim supn |n |(cn ) > 0. Next,
for each n, we can nd an cn such that |n an | 12 |n |(cn ), so that
lim supnN supC |an | lim supn |n an | > 0.
Since han inN , like hcn inN , is disjoint, the condition is not satised. This completes the proof.
362X Basic exercises (a) Let A be a Dedekind -complete Boolean algebra and 1 , 2 two countably
additive functionals on A. Show that |1 ||2 | = 0 in the Riesz space of bounded nitely additive functionals
on A i there is a c A such that 1 a = 1 (a c) and 2 a = 2 (a \ c) for every a A.
(b) Let (A,
) be a measure algebra. Write M , Mac as in 362C. Show that for any non-negative M ,
the component ac of in Mac is given by the formula
ac c = sup>0 inf a (c \ a).
296
Function spaces
362Xc
2 of in MA
is given by the formula 2 c = supAA inf aA (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 nitely additive functionals on A (353B). Show that the bands M , M , 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 -complete?
(b) Let (A,
) be a measure algebra. Let C be the band algebra of the Riesz space M of bounded nitely
additive functionals on A. Show that the bands M , M , 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,
) is semi-nite (iv) (A,
)
is localizable (v) (A,
) is -nite (vi) (A,
) is totally nite?
362 Notes
297
(c) Give an example of a set X, a -algebra of subsets of X, and a functional in Mp (as dened in
362Xg) which is not completely additive.
(d) Let U be a Riesz space and f , g U . Show that the following are equiveridical: () g is in the
band of U generated by f ; () for every u U + , > 0 there is a > 0 such that |g(v)| whenever
0 v u and |f |(v) ; () limn g(un ) = 0 whenever hun inN is a non-increasing sequence in U + and
limn f (un ) = 0. (Hint: 362B(f-i).)
(e) Let A be a weakly -distributive Boolean algebra (316Yg). Show that the inf in the formula for
P in 362Bc can be replaced by min.
(f ) Let A be any Boolean algebra and M the space of bounded nitely additive functionals on A. Let
C M be such that supC |a| < for every a A. (i) Suppose that supnN supC |an | is nite for every
disjoint sequence han inN in A. Show that C is norm-bounded. (ii) Suppose that limn supC |an | = 0
for every disjoint sequence han inN in A. Show that C is uniformly integrable.
(g) Let A be a Boolean algebra and M the space of completely additive functionals on A. Let C M
be such that supC |a| < for every atom a A. (i) Suppose that supnN supC |an | is nite for every
disjoint sequence han inN in A. Show that C is norm-bounded. (ii) Suppose that limn supC |an | = 0
for every disjoint sequence han inN in A. Show that C is uniformly integrable.
(h) Let A be a Dedekind -complete Boolean algebra and hn inN a sequence of countably additive
real-valued functionals on A. Suppose that a = limn n a is dened in R for every a A. Show that
is countably additive and that {n : n N} is uniformly integrable. (Hint: 246Yg.) Show that if every n
is completely additive, so is .
(i) Let A be a Boolean algebra, M the Riesz space of bounded nitely additive functionals on A, and
Mc M the space of atomless functionals (362Xf). Show that for a non-negative M the component c
of in Mc is given by the formula
Pn
c a = inf >0 sup{ i=0 ai : a0 , . . . , an a are disjoint, ai 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 complexication of M , as dened in 354Yl, can be identied with the Banach space of bounded
additive functionals : A C, writing
Pn
kk = sup{ i=0 |ai | : a0 , . . . , an are disjoint elements of A}
for such .
362 Notes and comments The Boolean algebras most immediately important in measure theory are of
course -algebras of measurable sets and their quotient measure algebras. It is therefore natural to begin any
investigation by concentrating on Dedekind -complete algebras. Nevertheless, in this section and the last
(and in 326), I have gone to some trouble not to specialize to -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 reect 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 () in 362B(f-i), for instance, is sucient in the absence of any
hypothesis of Dedekind -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 : M , M , 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 nite subalgebras of the band algebra of M (362Ya-362Yb), and for any such nite 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 members
into nite 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, , ) and was looking at the countably
additive functionals from to R, that is, at M 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
rened indenitely, as we rene the classication of additive functionals. At each stage, of course, the exact
enumeration of the subalgebra of bands generated by the classication (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 dened as weighted sums of point masses
(362Xg(ii)). The point is that measure algebras A appear as quotients of -algebras of sets by -ideals I;
consequently the countably additive functionals on A correspond exactly to the countably additive functionals on which are zero on I; but the canonical homomorphism from to A is hardly ever order-continuous,
so completely additive functionals on A rarely correspond to completely additive functionals on . On the
other hand, when we are looking at countably additive functionals on , 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 416Q 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 nd that intuitions often come to you rst in the context of Boolean
algebras, and the applications to Riesz spaces are secondary.
In 362E I give a condition for uniform integrability in terms of disjoint 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 dierent 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 () spaces of 243. I begin with the denition 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 dene 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 () and
L (/I) for -algebras and ideals I of sets (363H). This leads at once to the identication of L (), as
dened in Volume 2, with L (A), where A is the measure algebra of (363I). Like S(A), L (A) determines
the algebra A (363J). I briey 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
299
In the second half of the section I present some of the theory of Dedekind complete and -complete
algebras. First, L (A) is Dedekind (-)complete i A is (363M). The spaces L (A), for Dedekind complete A, are precisely the Dedekind -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 i there is a v L such that u = v v.
proof (a) See 354Hb and 354J.
(b)-(c) are obvious from the denitions 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 denition 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 a S(A) (361D) as a member of
L (A) for each a A, then : 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, : 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 : A L is order-continuous (361Gb).
363E Theorem Let A be a Boolean algebra, and U a Banach space. Let : A U be a bounded
additive function.
(a) There is a unique bounded linear operator T : L (A) U such that T = ; in this case kT k =
supa,bA ka bk.
(b) If U is a Banach lattice then T is positive i is non-negative; and in this case T is order-continuous
i is order-continuous, and sequentially order-continuous i is sequentially order-continuous.
(c) If U is a Banach lattice then T is a Riesz homomorphism i is a lattice homomorphism i ab = 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 = , and kT0 k =
sup{ka bk : 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 is non-negative, by 361Ga.
300
Function spaces
363E
(ii) If is non-negative then T0 is positive, by 361Ga in the other direction. But if u L+ and > 0,
then by 354I there is a v S + such that ku vk ; now kT u T vk kT k. But T v = T0 v belongs to
the positive cone U + of U . As 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 is order-continuous as well as non-negative, and that A L is a non-empty
downwards-directed set with inmum 0. Set
B = {v : v S, there is some u A such that v u}.
proof Let Z and W be the Stone spaces of A and B. By 312P there is a continuous function : W Z
such that
ca = 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 T .
(b) Set = kT uk , u (z) = max(, min(u(z), )) for z Z; that is, u = (e) (u e) 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
363F
while
301
T u = (T e) (T u T e) = T u,
ku k = kT uk = kT u k ku k kuk .
(c)(i) Let U be the closed linear subspace of L (A) generated by {a : a = 0}, and U0 the kernel of T .
Because T is continuous and linear, U0 is a closed linear subspace, and T (a) = 0 = 0 whenever a = 0;
so U U0 . Now take any u U0 and > 0. Then T (u+ ) = (T u)+ = 0, so u+ U0 . By 354I there is a
+
+
u S(A)
Pn such that 0 u u and ku u k . Now 0 T u T u = 0, so T u = 0. Express
(d) If 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 = b. Now there is a u S(A) such that ku u k 31 , so that kT u bk 13 . Taking
a, we must have a = b, since
a A such that {z : u (z) 21 } = b
bb = {w : (T u )(w) 1 } = 1 [b
a] =
ca.
2
As b is arbitrary, 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 i is injective. In this case, for any u L (A),
kT uk = kT |u|k
and kT uk = kuk .
(f ) If T is (sequentially) order-continuous then = T is (sequentially) order-continuous, by 363D.
If is (sequentially) order-continuous then : 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 T .
302
Function spaces
363G
(c) In particular, if I is any -ideal of and E 7 E is the canonical homomorphism from onto
A = /I, then we have an identication of L (A) with a quotient of L , and for any E we can
identify (E ) L (A) with the equivalence class (E) L /W of the characteristic function E.
proof (a) For the elementary properties of the space of -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 inN 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 dened for every x X; now f is -measurable, by 121Fa. Of
course kf k supnN kfn k < , so f L , while
kf fn k supmn kfm fn k 0
363K
303
for every v L (A) (363Fd), that is, that the norm of L (A) corresponds to the quotient norm on
L ()/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.
L () of 363H;
Remark The space I called L () in Chapter 24 is not strictly speaking the space L =
0
I took L () L () to be the set of essentially bounded, virtually measurable functions dened almost
everywhere on X, and in general this is larger. But, as remarked in the notes to 243, L () can equally
because |wa| kwk a and |w(1 \ a)| kwk (1 \ a). So Va and V1\a are complementary projection
bands in L = L (A). Next, if U L is a projection band, then 1 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 , M and M be the L-spaces of bounded nitely 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) ,
M
and
L
(A)
.
,
and
S(A)
M
and
L
(A)
S(A)
=
c
c =
proof Write S = S(A), L = L (A).
(a) For the identications S
= M see 362A.
= M and S
= M , Sc
304
363K
Function spaces
(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 i it is bounded on the unit ball.
(c) If f is a positive linear functional on L , then f S 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,
L
to S . If f L
and f S 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 S : f L }. This means that f 7 f S is actually
a Riesz space isomorphism between L and S . In particular, |f |S = |f S| for any f L .
(ii) If f : L R is a positive linear operator and f S Sc , let hun inN be a non-increasing sequence
in L with inmum 0. For each n, k N there is a vnk S such that un vnk un + 2k e, where e is the
standard order unit of L (354I, as usual); set wn = inf i,kn vik ; then hwn inN is a non-increasing sequence
in S with inmum 0, so
(L )
c .
f (L )
c |f | (L )c |f S| Sc f S Sc ,
u + v d =
u d +
v d,
u d = u d
for
R every a A,R so thatRf + f and f+R must agree on
R S(A) and therefore on L . But this means that
u d( + ) = u d + u d. Similarly, u d() = u d. Q
Q
R
(c) If is non-negative, we have u d 0 whenever u 0, as in part (c) of the proof of 363K.
Consequently, for any M and u L ,
Z
Z
Z
Z
Z
+
+
+
+
| u d| = | u d u d u d + u d |
Z
Z
Z
Z
u+ d + + u d + + u+ d + u d
Z
Z
= |u|d|| kuk 1 d|| = kuk ||(1) = kuk kk.
363Lf
305
R
So (u, ) 7 u d has norm (as dened in 253Ab) at most 1. If A 6= 0, the norm is exactly 1. (For this we
need to know that there is a M + such that 1 = 1. Take any z in the Stone space of A and set a = 1
if z b
a, 0 otherwise.)
(d) We do not have any result corresponding to B.Levis theorem in this language, because (even if
is non-negative and
R countably additive) there is no reason to suppose that supnN un is dened in L just
because supnN un d is nite. But if is countably additive and A is Dedekind -complete, we have
something corresponding to Lebesgues 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
f d =
R
0
{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 -complete Boolean algebras there is an eective
expression of this idea, and I will use it in the principal denition of 365). But what we can say is the
following. If A is any Boolean algebra, and : A [0, [ is a non-negative additive functional, and
u L (A)+ , then
u d =
R
0
sup{a : ta u}dt,
sup{a : ta u}. Then h is non-increasing and zero for t > kuk , so 0 h(t)dt is dened in R. If
n
n
n
we set hn (t) = h(2 (k + 1)) whenever k, n N and 2 k t < 2 (k + 1), then hhn (t)inN 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 > 0, we can choose for
each k k = 2n kuk an ak such that 2n (k + 1)ak u and ak h(2n (k + 1)) . In this case
Pk n
ak u, so
k=0 2
Z
hn (t)dt = 2
k
X
h(2
k=0
= kuk +
(k + 1)) kuk + 2
Z X
k
k=0
ak d kuk +
k
X
ak
k=0
u d.
R
R
As n and are arbitrary, 0 h(t)dt u d. (ii) In the other direction, there is for any > 0 a v S(A)
Pm
such that v u v + 1. If we express v as j=0 j cj where c0 . . . cm and j 0 for every j
Pk
(361Ec), then we shall have h(t) ck whenever t j=0 j , so
R
R
R
Pm
h(t)dt k=0 k ck = v d u d 1.
0
R
R
As is arbitrary, 0 h(t)dt u d and the two integrals are equal. Q
Q
R
(f ) The formula d 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 -algebra of subsets of a set
X, L can be identied with
the space Rof bounded A-measurable functions on X, as in 363Ha. So in
R
such contexts I may write g d or even g(x)(dx) when g : X R is bounded and A-measurable, and
: A R is an additive functional.
But I will try to take care to signal any such deviation from the normal
R
principle that the symbol
refers to the sequentially order-continuous integral dened in 122 with the
minor modications 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
Function spaces
363M
363M Now I come to a fundamental fact underlying a number of theorems in both this volume and the
last.
Theorem Let A be a Boolean algebra.
(a) A is Dedekind -complete i L (A) is Dedekind -complete.
(b) A is Dedekind complete i L (A) is Dedekind complete.
proof (a)(i) Suppose that A is Dedekind -complete. By 314M, we may identify A with a quotient /M,
where M is the ideal of meager subsets of the Stone space Z of A, and = {EA : 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
identied with L /V, where L is the space of bounded -measurable functions from Z to R, and V is the
space of functions zero except on a member of I.
Now suppose that hun inN is a sequence in L with an upper bound u L . Express un , u as fn , f
where fn , f L . Set g(z) = supnN 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 and w un for every n, then express w as h where h L ; we have (fn h)+ V
for every n, so
S
{z : g(z) > h(z)} nN {z : fn (z) > h(z)} M
because M is a -ideal, and (g h)+ V, w v. Thus v = supnN un in L . As hun inN is arbitrary, L
is Dedekind -complete (using 353G).
(ii) Now suppose that L is Dedekind -complete, and that A is a countable non-empty set in A.
In this case {a : a A} has a least upper bound u in L . Take a v S(A) such that 0 v u and
ku vk 31 ; set b = [[v > 31 ]], as dened in 361Eg. If a A, then k(a v)+ k ku vk 13 , so
2
3 a v and a b. If c A is any upper bound for A, then v u c so b c. Thus b = sup A in A. As
A is arbitrary, A is Dedekind -complete.
(b)(i) For the second half of this theorem I use an argument which depends on joining the representation
described in (a-i) above with the original denition 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 in /M, corresponds to a (see the proof of 314M), so (b
a) L /V corresponds
to a; that is, T (b
a) = b
a, if we identify a L with b
a : Z {0, 1}. So V contains b
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 . 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 = GA . 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) = supuA u(z). Then
S
G = {z : g(z) > } = uA {z : u(z) > }
is open for every R (that is, g is lower semi-continuous). Thus G for every , so g L , and
v = T g is dened 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
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 -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 Hausdor
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 inN , hfn inN in U such that
fn g gn and kgn fn k 2n 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 G1 and f (x) > 0 (also
by 3A3Bc); write H for {y : f (y) > 0}. Set g = supnN (nf X), 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 (X \ H); now h f = 0 so h g = 0 and
g(y) = 0. Thus H g H. 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 identies 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 X : n N} is just H.
But g is supposed to be the least continuous function greater than or equal to nf X for every n, and is
therefore likely to be strictly greater than H, even though sandwiched between H and H.
363O Corollary Let U be a Dedekind -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 -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 -complete
Boolean algebra A; and if U is Dedekind complete, so is A.
proof Because U is Dedekind -complete, so is Ue (353J(a-i)). Apply 363N to Ue to see that Ue
= L (A)
for some A. Because Ue is Dedekind -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 i 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 kkwk1.
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 = 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 . 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 Zorns 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+
v : v V , R}.
If v1 , v2 V then, writing e for the standard order unit of L ,
Tv1 + Tv2 = T(v1 + v2 ) kT(v1 + v2 )k e
so
u
= supv1 V Tv1 kv1 vke
is dened in L and u
kv2 + vke T v2 for every v2 V . Putting these together, we have
Tv u
kv vke,
Tv + u
kv + vke
Tv +
u = ||(T( 1 v) u
) ||k 1 v vke = kv +
v ke,
363R
So we have
309
Tv kTvk e kvke.
Tv +
u kv +
v ke
for every v V , R.
Dene T1 : V1 L by setting T1 (v +
v ) = Tv +
u for every v V , R. (This is well-dened
because v
/ V , so any member of V1 is uniquely expressible as v +
v where v V and 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 kvke for every v V1 ; since we also have
T1 v = T1 (v) k vke = kvke,
kT1 vk kvk 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 Hausdor
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 Zorns
Lemma, upside down, P has a minimal element X; with its relative topology, X is a compact Hausdor
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 kuk1 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 < 1 such that |f (u)| 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.
P?? Otherwise, set G = {y : y
(d) If h (X) and x0 X \ {x : h(x) 6= 0}, then (T h)(x0 ) = 0. P
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 , < 1 such that
kf k = 1, |f (x)| 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 > 0 be such that khk 1 . Then, because h(x) = 0 for x G,
|f (x)| + |h(x)| 1 for every x X, and kf + hk , kf hk are both less than or equal to 1. As T f = f
and kT k = 1, this means that
consequently
kf T hk 1,
kf + T hk 1;
310
363R
Function spaces
h(x0 )
f,
f (x0 )
h1 (x) = f1 (x)
(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 X = G + (X \ G), so
X = T (X) = h1 + h2 ,
where h1 = T (G), h2 = T ((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 = G.
Q
Of course it follows that G is open. As G is arbitrary, X is extremally disconnected. Q
(g) Being also compact and Hausdor, therefore regular (3A3Bc), X is zero-dimensional (3A3Bd). We
may therefore identify X with the Stone space of its regular open algebra RO(X) (314S), and W = C(X) with
L (RO(X)). Thus R : U C(X) is a Banach space isomorphism between U and C(X)
= L (RO(X));
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 Zorns 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 , with domain PX, such that {x} = 0 for every x X.
(ii) There are a localizable measure space (X, , ) and an absolutely continuous countably additive
functional : R which is not truly continuous, so has no Radon-Nikod
ym derivative (denitions:
232Ab, 232Hf).
(iii) There are a Dedekind complete Boolean algebra A and a countably additive functional : A R
which is not completely additive.
(iv) There is a Dedekind complete Riesz space U such that Uc 6= U .
363Xd
311
proof (a)(i)(ii) Let X be a set with a probability measure , dened on PX, such that {x} = 0 for
every x X. Let be counting measure on X. Then (X, PX, ) is strictly localizable, and : PX R
is countably additive; also E = 0 whenever E is nite, so is absolutely continuous with respect to .
But if E < then E is nite and (X \ E) = 1, so is not truly continuous, and has no Radon-Nikod
ym
derivative (232D).
(b)(ii)(iii) Let (X, , ) be a localizable measure space and : R an absolutely continuous
countably additive functional which is not truly continuous. Let (A,
) be the measure algebra of ; then
we have an absolutely continuous countably additive functional : A R dened by setting E = E
for every E (327C). Since is not truly continuous, is not completely additive (327Ce). Also A is
Dedekind complete, because is localizable, so A and witness (iii).
(c)(iii)(i) Let A be a Dedekind complete Boolean algebra and : A R a countably additive
functional which is not completely additive. Because is bounded (326I), therefore expressible as the
dierence of non-negative countably additive functionals (326H), there must be a non-negative countably
additive functional on A which is not completely additive.P
By 326N, there is a partition of unity hai iiI in A such that iI ai < 1. Set K = {i : i I, ai > 0};
then K must be countable, so
P
(supiI\K ai ) = 1 (supiK ai ) = 1 iK ai > 0.
For J I set J = (supiJ\K ai ); the supremum is always dened because A is Dedekind complete.
Because is countably additive and non-negative, so is ; because ai = 0 for i J \ K, {i} = 0 for
every i I. Multiplying by a suitable scalar, if need be, (I, PI, ) 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 inmum 0, such that
inf uA 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 inmum 0, and
inf uAUe g(u) = inf uA f (u) > 0,
g
/ Ue . But Ue is a Riesz space with order unit e, and is Dedekind complete because U is; so it can be
identied with L (A) for some Boolean algebra A (363N), and A is Dedekind complete, by 363M.
312
Function spaces
363Xe
> (e) Let (X, , ) be a measure space with measure algebra A, and L the space of bounded measurable real-valued functions on X. A linear lifting of is a positive linear operator T : L (A) L
such that T (1A ) = X 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 : A is a lifting in the sense of 341A then T , as dened 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 : A dened by setting a = {x : T (a)(x) = 1} for each a A. (iii) Show that , as
dened in (ii), is a lifting i T is a Riesz homomorphism i T is multiplicative.
(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 -complete Boolean algebra. (i) Show that for any u L (A), R there
are elements [[u ]], [[u > ]] A, where [[u ]] is the largest a A such that u a a, and
[[u > ]] = sup> [[u ]]. (ii) Show that in the context of 363H, if u corresponds to f for f L , then
[[u ]] = {x : f (x) } , [[u > ]] = {x : f (x) > } . (iii) Show that if A L is non-empty and v L ,
then v = sup A i [[v > ]] = supuA [[u > ]] for every R; in particular, v = u i [[v > ]] = [[u > ]] for
every . (iv) Show that a function : R A is of the form () = [[u > ]] i () () = sup> () for
every R () there is an M such that (M ) = 0, (M ) = 1. (v) Put (iii) and (iv) together to give a
proof that L is Dedekind -complete if A is.
(i) Let A be a Dedekind -complete Boolean algebra and U L (A) a (sequentially) order-closed Riesz
subspace containing 1. Show that U can be identied with L (B) for some (sequentially) order-closed
subalgebra B A. (Hint: set B = {b : b 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 i L (A) has the countable sup property (241Ye,
353Ye).
(c) Let X be an extremally disconnected topological space, and RO(X) its regular open algebra. Show
that there is a natural isomorphism between L (RO(X)) and Cb (X).
(d) Let X be a compact Hausdor 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 , i max(ku + vk , ku vk ) > 1 whenever v L \ {0}. (ii) Show that if u, v L then |u| |v| = 0
i ku + v + wk max(ku + wk , kv + wk ) whenever = 1 and w L . (iii) Show that if
T : L L is a normed space automorphism then there are a Boolean automorphism : A A and a
w L such that |w| = e and T u = w T u for every u L .
363 Notes
313
(f ) Let X be a set, an algebra of subsets of X, and I an ideal in . (i) Show that L () can be
identied, as Banach lattice and Banach algebra, with the space L of bounded functions f : X R such
that whenever < in R there is an E such that {x : f (x) } E {x : f (x) }. (ii) Show that
L = {g : g C(Z)}, where Z is the Stone space of and : X Z is a function (to be described). (iii)
Show that L (/I) can be identied, as Banach lattice and Banach algebra, with L /V, where V is the
set of those functions f L such that for every > 0 there is a member of I including {x : |f (x)| }.
(g) Let (X, , ) be a complete probability space with measure algebra A. Let hBn inN be a nondecreasing
sequence of closed subalgebras of A such that A is the closed subalgebra of itself generated by
S
1
1
nN Bn , and set n = {F : F Bn } for each n. Let Pn : L () L (n ) be the conditional
L (n ) = L (Bn ). Suppose that we are given for each n a lifting n : Bn n and that n+1 b = n b
whenever n N, b Bn . Let Tn : L (Bn ) L be the corresponding linear liftings (363Xe), and F any
non-principal ultralter on N. (i) Show that for any u L (A), hTn Pn uinN converges almost everywhere.
(ii) For u L (A) set (T u)(x) = limnF (Tn Pn u)(x) for x X, u L (A). Show that T is a linear lifting
of . (iii) Use 363Xe(ii) and 341J to show that there is a lifting of extending every n . (iv) Use this as
the countable-conality 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 -complete Boolean algebra and : A R a countably additive func
tional. Suppose that hun inN is an order-bounded
R sequence in LR (A) such that inf nN supmn um and
supnN inf mn um are equal to u say. Show that u d = limn un d.
(i) Let be the family of those sets E [0, 1] such that (int E) = E, where is Lebesgue measure.
(i) Show that is an algebra of subsets of [0, 1] and that every member of is Lebesgue measurable. (ii)
Show that if we identify L () 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 for , then d, as dened 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 i it is
isomorphic to L
C (A), as described in 363Xj, for some Dedekind complete Boolean algebra A.
(k) Find a Boolean algebra A and an automorphism : A A such that (i) is ergodic in the sense that
{a : a = a} = {0, 1} (ii) there is a non-constant u L = L (A) such that T u = u, where T : L L
is the Riesz homomorphism associated with .
363 Notes and comments As with S(A), I have chosen a denition of L (A) in terms 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 Hausdor X. The
point at which it becomes helpful to interpret C(X) as L (A) is when C(X) is Dedekind -complete. The
spaces X for which this is true are dicult to picture, and alternative representations of L along the lines
of 363H-363I can be easier on the imagination.
For Dedekind -complete A, there is an alternative description of members of L (A) in terms of objects
[[u > ]] (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 oers an abstract expression of one of the central results of Chapter 24. In 243H
I showed that L () is always Dedekind -complete, and that it is Dedekind complete if is localizable.
We can now relate this to the results of 321H and 322Be: the measure algebra of any measure is Dedekind
-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 i X is extremally disconnected; while for compact
314
Function spaces
363 Notes
Hausdor spaces, X is extremally disconnected i it is the Stone space of a Dedekind complete algebra.
With the right modication of the concept extremally disconnected (314Yf, 353Yc), the same approach
works for Dedekind -completeness.
363R is the Nachbin-Kelley theorem; it is commonly phrased a normed space U has the Hahn-Banach
extension property i it is isomorphic, as normed space, to C(X) for some compact extremally disconnected
Hausdor space X, but the expression in terms 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 remember 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 i it is a Banach lattice automorphism preserving the standard order unit i 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 signicant ideas in this context will appear in 438.
364 L0
My next objective is to develop an abstract construction corresponding to the L0 () 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 signicant new diculties, 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 denition 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 terms 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 -complete Boolean algebra. I will write L0 (A) for the set of all
functions 7 [[u > ]] : R A such that
() [[u > ]] = sup> [[u > ]] in A for every R,
() inf R [[u > ]] = 0,
() supR [[u > ]] = 1.
364B Remarks (a) My reasons for using the notation [[u > ]] rather than u() will I hope become
clear in the next few paragraphs. For the moment, if you think of A as a -algebra of sets and of L0 (A)
as the family of A-measurable real-valued functions, then [[u > ]] corresponds to the set {x : u(x) > }
(364Ja).
(b) Some readers will recognise the formula [[. . . ]] as belonging to the language of forcing, so that [[u > ]]
could be read as the Boolean value of the proposition u > . But a vocalisation closer to my intention
might be the region where u > .
L0
364C
R.
315
(c) Note that condition () of 364A automatically ensures that [[u > ]] [[u > ]] whenever in
(d) In fact it will sometimes be convenient to note that the conditions of 364A can be replaced by
( ) [[u > ]] = supqQ,q> [[u > q]] in A for every R,
( ) inf nN [[u > n]] = 0,
( ) supnN [[u > n]] = 1;
the point being that we need look only at suprema and inma 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 -algebras and -ideals of sets, as follows.
364C Proposition Let X be a set, a -algebra of subsets of X, and I a -ideal of .
(a) Write L0 for the space of all -measurable functions from X to R. Then L0 , with its linear structure,
ordering and multiplication inherited from RX , is a Dedekind -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
-complete Riesz space and an f -algebra with a multiplicative identity;
(iii) for f , g L0 , f g in L0 /W i {x : f (x) > g(x)} I, and f = g in L0 /W i {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, f , f g, supnN 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 X.
(b)(i) The necessary verications are all elementary.
(ii) Because W is a solid linear subspace of the Riesz space L0 , the quotient inherits a Riesz space
structure (351J, 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 = X because X 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.
(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
364D
Function spaces
364D Theorem Let X be a set and a -algebra of subsets of X. Let A be a Dedekind -complete
Boolean algebra and : A a surjective Boolean homomorphism, with kernel a -ideal I; dene L0 and
W as in 364C, so that U = L0 /W is a Dedekind -complete f -algebra with multiplicative identity.
(a) We have a canonical bijection T : U L0 = L0 (A) dened by the formula
[[T f > ]] = {x : f (x) > }
for every f L0 , R.
(b)(i) For any u, v U ,
for every R.
(ii) For any u U , > 0,
for every R.
(iii) For any u, v U ,
+
]]
q ]]
for every 0.
(v) Writing e = (X) for the multiplicative identity of U , we have
[[T e > ]] = 1 if < 1, 0 if 1.
proof (a)(i) Given f L0 , set f () = {x : f (x) > } for R. Then it is easy to see that f satises
the conditions () -() of 364Bd, because is sequentially order-continuous (313Qb). Moreover, if f = g
in U , then
f () g () = ({x : f (x) > }{x : g(x) > }) = 0
for every R, because
and f = g . So we have a well-dened member T u of L0 dened 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
such that Fq = [[w > q]] in A. Note that if q q then
so Fq \ Fq I. Set
H=
Fq \
nN
qQ,qn
Fq ,
f (x) = sup{q : q Q, x Fq } if x H,
= 0 otherwise.
(H is chosen just to make the formula here give a nite value for every x.) We have
sup
nN qQ,qn
L0
364D
{x : f (x) > } =
qQ,q>
qQ,q>
317
Fq (X \ H) if < 0
Fq \ (X \ H) if 0,
and in either case belongs to ; so that f L0 and f is dened in L0 . Next, for any R,
[[T f > ]] = {x : f (x) > } = (
Fq )
qQ,q>
sup
qQ,q>
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 > ]] = [[T g > ]] for every R
= 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 R,
[[T (u + v) > ]] = {x : f (x) + g(x) > }
[
= ( {x : f (x) > q} {x : g(x) > q})
qQ
]].
(see 364C(b-iii))
318
364D
Function spaces
(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 0, then
[[T (u v) > ]] = (
qQ,q>0
qQ,q>0
qQ,q>0
q ]].
whenever u, v L and R,
]]
whenever u, v 0 in L0 and 0,
q ]]
proof (a) By the Loomis-Sikorski theorem (314M), we can nd a set Z (the Stone space of A), a -algebra
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 : A (corresponding to the
identication between A and the quotient /M). Consequently, dening L0 and W as in 364C, we have
a bijection between the Dedekind -complete f -algebra L0 /W and L0 (364Da). Of course this endows L0
itself with the structure of a Dedekind -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 dened by the description oered. For while scalar
multiplication is not described for 0, the assertion that L0 is a Riesz space implies that 0u = 0 and that
u = ()(u) for < 0; so if we have formulae to describe u + v and u for > 0, this suces to dene
the linear structure of L0 . Note that we have an element 0 in L0 dened by setting
[[0 > ]] = 0 if 0, 1 if < 0,
and the formula for u + v shows us that
[[0 + u > ]] = supqQ [[0 > q]] [[u > q]] = supqQ,q<0 [[u > q]] = [[u > ]]
for every , so that 0 is the zero of L0 . As for multiplication, if L0 is to be an f -algebra we must have
0 +
L0
364I
319
u v = (u+ v + ) (u+ v ) (u v + ) + (u v ),
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 -complete Boolean algebra.
(a) If u, v L0 = L0 (A) and , R,
(b) If u, v 0 in L0 and , 0 in R,
proof (a) For any q Q, either q and [[u > q]] [[u > ]], or q and [[v > + q]]
in all cases
[[u > q]] [[v > + q]] [[u > ]] [[v > ]];
for every , so u = u . 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 (E), where : B A is the homomorphism corresponding to u L0 . Thus [[u > ]] = [[u ], [ ]].
In the same spirit I write [[u ]] for [[u [, [ ]] = inf < [[u > ]], [[u 6= 0]] = [[|u| > 0]] = [[u > 0]] [[u < 0]]
and so on, so that (for instance) [[u = ]] = [[u {}]] = [[u ]] \ [[u > ]] for u L0 , R.
364I Proposition Let A be a Dedekind -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 dened by saying that [[h(u)
F ]] = [[u h1 [F ]]] for every Borel set F R.
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 eect already been mentioned.
(a) Let X be a set and a -algebra of subsets of X. Then we may identify L0 () with the space L0
of -measurable real-valued functions on X. (This is the case A = 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
composition hf , for any f L0 .
(b) Now suppose that I is a -ideal of and that A = /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, , ) is a measure space with measure algebra A, then L0 (A) becomes identied
with L0 () as dened in 241.
The same remarks as in 363I apply here; the space L0 () of 241A is larger than the space L0 of the
present section. But for every f L0 () there is a g L0 such that g =a.e. f (241Bk), so that L0 () can
be identied with L0 /N, where N is the set of functions in L0 which are zero almost everywhere (241Yc).
364K Embedding S and L in L0 : Proposition Let A be a Dedekind -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, a, when regarded as a member of L0 , can be described by the formula
[[a > ]] = 1 if < 0,
= a if 0 < 1,
= 0 if 1 .
= = 0 if 1,
L0
364M
321
364M Suprema and infima in L0 We know that any L0 (A) is a Dedekind -complete partially ordered
set. There is a useful description of suprema for this ordering, as follows.
Proposition Let A be a Dedekind -complete Boolean algebra and A a subset of L0 = L0 (A).
(a) A is bounded above in L0 i there is a sequence hcn inN in A, with inmum 0, such that [[u > n]] cn
for every u A.
(b) If A is non-empty, then A has a supremum in L0 i c = supuA [[u > ]] is dened in A for every
R and inf nN cn = 0; and in this case c = [[sup A > ]] for every .
(c) If A is non-empty and bounded above, then A has a supremum in L0 i supuA [[u > ]] is dened in
A for every R.
proof (a)(i) If A has an upper bound u0 , set cn = [[u0 > n]] for each n; then hcn inN satises the conditions.
(ii) If hcn inN satises the conditions, set
() = 1 if < 0,
= inf ci if n N, [n, n + 1[ .
in
Then it is easy to check that satises the conditions of 364A, since inf nN cn = 0. So there is a u0 L0
such that () = [[u0 > ]] for each . Now, given u A and R,
[[u > ]] 1 = [[u0 > ]] if < 0,
in
in
Accordingly there is a u L such that [[u > ]] = c for every R. But now, for v L0 ,
0
322
Function spaces
364M
so that u = sup A in L0 .
u v,
(ii) Now suppose that u = sup A is dened in L0 . Of course [[u > ]] must be an upper bound for
{[[u > ]] : u A} for every . ?? Suppose we have an for which it is not the least upper bound, that is,
there is a c [[u > ]] which is an upper bound for {[[u > ]] : u A}. Dene : R A by setting
() = c [[u > ]] if ,
= [[u > ]] if < .
It is easy to see that satises the conditions of 364A (we need the distributive law 313Ba to check that
() = sup> () if ), so corresponds to a member v of L0 . But we now nd that v is an upper
bound for A (because if u A, then
[[u > ]] [[u > ]] [[u > ]] c [[u > ]] = [[v > ]],)
that v u and that v 6= u (because [[v > ]] = c 6= [[u > ]]); but this is impossible, because u is
supposed to be the supremum of A. X
X Thus if u = sup A is dened in L0 , then supuA [[u > ]] is dened
in A for every R. Also, of course,
inf nN supuA [[u > n]] = inf nN [[u > n]] = 0.
(c) This is now easy. If A has a supremum, then surely it satises the condition, by (b). If A satises
the condition, then we have a family hc iR as required in (b); but also, by (a) or otherwise, there is a
sequence hcn inN such that cn cn for every n and inf nN cn = 0, so inf nN cn is also 0, and both conditions
in (b) are satised, so A has a supremum.
364N We do not have such a simple formula for general inma (though see 364Xl), but the following
facts are useful.
Proposition Let A be a Dedekind -complete Boolean algebra.
(a) If u, v L0 = L0 (A), then [[u v > ]] = [[u > ]] [[v > ]] for every R.
(b) If A is a non-empty subset of (L0 )+ , then inf A = 0 in L0 i inf uA [[u > ]] = 0 in A for every > 0.
proof (a) Take Z, L0 and 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 > ]] = {z : min(f (z), g(z)) > } = ({z : f (z) > } {z : g(z) > })
= {z : f (z) > } {z : g(z) > } = [[u > ]] [[v > ]]
for every .
(b)(i) If inf uA [[u > ]] = 0 for every > 0, and v is any lower bound for A, then [[v > ]] must be 0 for
every > 0, so that [[v > 0]] = 0; now since [[0 > ]] = 0 for 0, 1 for < 0, v 0. As v is arbitrary,
inf A = 0.
(ii) If > 0 is such that inf uA [[u > ]] is undened, or not equal to 0, let c A be such that
]] is 1 if < 0, c if
0 6= c [[u > ]] for every u A, and consider v = c. Then [[v > ]] = [[c >
0 < and 0 if . If u A then [[u > ]] is 1 if < 0 (since u 0), at least [[u > ]] c if 0 < ,
and always includes 0; so that v u. As u is arbitrary, inf A is either undened 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 -complete Boolean algebra A, L0 = L0 (A) is Dedekind complete i A is.
proof The description of suprema in 364Mc makes it obvious that if A is Dedekind complete, so that
supuA [[u > ]] is always dened, 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 1 (364E,
364Kc). By 353Pb, or otherwise, u v = 0 i |u| |v| = 0. There is one special feature of multiplication in
L0 which I can mention here.
Proposition Let A be a Dedekind -complete Boolean algebra. Then an element u of L0 has a multiplicative
inverse in L0 i |u| is a weak order unit in L0 i [[|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| c > 0]] = [[|u| > 0]] c = 0
(364Na), so that |u| c 0 and c = 0, that is, c = 0; so [[|u| > 0]] must be 1. To complete the circuit,
suppose that [[|u| > 0]] = 1. Let Z, , L0 , , M be as in the proof of 364E, and S : L0 L0 the canonical
map, so that [[Sh > ]] = {z : h(z) > } for every h L0 , R. Express u as Sf where f L0 . Then
{z : |f (z)| > 0} = [[Sf > 0]] = 1, so {z : f (z) = 0} M. Set
g(z) =
Then {z : f (z)g(z) 6= 1} M so
and u is invertible.
1
f (z)
if f (z) 6= 0,
g(z) = 0 if f (z) = 0.
u Sg = S(f g) = S(Z) = 1
Remark The repeated phrase by 353x or otherwise reects 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 -complete Boolean algebra. For
a A write Va for the band of L0 = L0 (A) generated by a. 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 a Va . P
P If
v Va then |a| |v| = 0, so a v = 0, so (w a) v = 0, so |w a| |v| = 0; thus w a Va ,
which is equal to Va because L0 is Archimedean (353B). Q
Q Now, if a A, u Va and v V1\a , then
|u| |v| = 0 because a (1 \ a) = 0; and if w L0 (A) then
w = (w a) + (w (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 1
is expressible as u + v = u v where u U , v U . Setting a = [[u > 12 ]], a = [[v > 12 ]] we must have
a a = 1, a a = 0 (using 364M and 364Na), so that a = 1 \ a; also 21 a u, so that a U , and
similarly (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 : A B a sequentially order-continuous Boolean
homomorphism.
(a) We have a multiplicative sequentially order-continuous Riesz homomorphism T : L0 (A) L0 (B)
dened by the formula
[[T u > ]] = [[u > ]]
0
L (B) as embedded in L0 (A) and L0 (B) respectively, then T , as dened here, agrees on L (A) with T
as dened in 363F.
(c) T is order-continuous i is order-continuous, injective i is injective, surjective i is surjective.
324
364R
Function spaces
for
= T h
(d) [[T u E]] = [[u E]] for every u L0 (A) and every Borel set E R; consequently hT
indierently for the associated maps from L0 (A) to itself and
every Borel measurable h : R R, writing h
from L0 (B) to itself (364I).
(e) If C is another Dedekind -complete Boolean algebra and : B C another sequentially ordercontinuous Boolean homomorphism then T = T T : L0 (A) L0 (C).
proof I write T for T .
(a)(i) To see that T u is well-dened in L0 (B) for every u L0 (A), all we need to do is to check that
the map 7 [[u > ]] : R B satises the conditions of 364Bd, and this is easy, because preserves all
countable suprema and inma.
(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 > ]] = sup [[T u > q]] [[T v > q]] = sup [[u > q]] [[v > q]]
qQ
= (sup [[u > q]] [[v > q]]) = [[u + v > ]] = [[T (u + v) > ]]
qQ
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 > ]] = sup [[u > ]] = (sup [[u > ]]) = [[u > ]]
uA
uA
uA
(using 364M)
= [[T u > ]]
for every R. So using 364M again, T u = supuA 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 identication of T (a) with (a) is another almost trivial verication. 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 = kuk , then |u| 1A , so that |T u| 1B , and T u L (B), with kT uk . Thus T L (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).
) Suppose that is order-continuous, and that A L0 (A) is a non-empty set with a supremum
(c)(i)(
0
u L (A). Then for any R,
(by 364M)
= sup [[u > ]]
uA
(because is order-continuous)
L0
364S
325
) Suppose that is surjective. Let be a -algebra of sets such that there is a sequentially
(iii)(
order-continuous Boolean surjection : A. Then : B is surjective. So given w L0 (B), there
is an f L0 () such that [[w > ]] = {x : f (x) > } for every R (364D). But, also by 364D, there
is a u L0 (A) such that [[u > ]] = {x : f (x) > } for every . And now of course T u = w. As w is
arbitrary, T is surjective.
) If T is surjective, and b B, there must be some u L0 (A) such that T u = b. Now set
(
a = [[u > 0]] and see that a = [[b > 0]] = b. As b is arbitrary, is surjective.
(d) The map E 7 [[u E]] is a sequentially order-continuous Boolean homomorphism, equal to [[T u E]]
when E is of the form ], [; 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 h1 [E]]] = [[u h1 [E]]]
[[h(T
= [[h(u)
E]] = [[T (h(u))
E]].
= hT
.
As E and u are arbitrary, T h
proof Because each i 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) iI L0 (Ai ) is a bijection.
If u, v L0 (A) are distinct, there must be some R such that [[T u > ]] 6= [[T v > ]]. In this case
there must be an i I such that i [[T u > ]] 6= i [[T v > ]], that is, [[Ti u > ]] 6= [[Ti v > ]]. So Ti u 6= Ti v
and T u 6= T v. As u, v are arbitrary,QT is injective.
If w = hwi iiI is any member of iI L0 (Ai ), then for R set
() = h[[wi > ]]iiI A.
It is easy to check that satises the conditions of 364A, because, for instance,
sup> i () = sup> [[wi > ]] = [[wi > ]] = i ()
for every i, so that sup> () = (), for every R; and the other two conditions are also satised
because they are satised coordinate-by-coordinate. So there is a u L0 (A) such that () = [[u > ]] for
every , that is, i [[u > ]] = [[wi > ]] for all , 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
*364T
Function spaces
*364T Regular open algebras I noted in 314P that for every topological space X there is a corresponding Dedekind complete Boolean algebra RO(X) of regular open sets. We have an identication of
L0 (RO(X)) as a space of equivalence classes of functions, dierent 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
denition.
Definition Let (X, T) be a topological space, f : X R a function. For x X write
(f, x) = inf GT,xG supy,zG |f (y) f (z)|
(allowing ).
*364U Theorem Let X be any topological space, and RO(X) its regular open algebra. Let U be the
set of functions f : X R such that {x : (f, x) < } is dense in X for every > 0. Then U is a Riesz
subspace of RX , closed under multiplication, and we have a surjective multiplicative Riesz homomorphism
T : U L0 (RO(X)) dened by writing
[[T f > ]] = sup> int {x : f (x) > },
the supremum being taken in RO(X), for every R, f U . The kernel of T is the set W of functions
f : X R such that int{x : |f (x)| } is dense for every > 0, so L0 (RO(X)) can be identied, as
f -algebra, with the quotient space U/W .
) The rst thing to observe is that for any f RX , > 0 the set
proof (a)(i)(
[
{x : (f, x) < } = {G : G X is open and non-empty
is open.
) Next, it is easy to see that
(
(f + g, x) (f, x) + (g, x),
(f, x) = ||(f, x),
(|f |, x) (f, x),
( ) 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 G G on which f is bounded. P
P Take any x0 G such that (f, x0 ) < 1; then there
is a non-empty open set G containing x0 such that |f (y) f (z)| < 1 for all y, z G , and we may suppose
that G G. But now |f (x)| 1 + |f (x0 )| for every x G . Q
Q
(ii) So if f , g U and R then
1
2
1
2
},
1+||
are also dense. As is arbitrary, f + g, f and |f | all belong to U ; as f , g and are arbitrary, U is a Riesz
subspace of RX .
(iii) If f , g U then f g U . P
P Take > 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)|) for every x G1 .
L0
*364U
Set = min(1,
)
2(1+)
327
> 0. Then there is an x G1 such that (f, x) < and (g, x) < . Let H, H
be open sets containing x such that |f (y) f (z)| < for all y, z H and |g(y) g(z)| < for all y, z H .
Consider G = G1 H H . 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)|
2
+ + .
Accordingly
(f g, x) (1 + 2) < ,
>>
>
(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)| < for every x G1 . If > then G1 does not meet {x : f (x) > }, so
G1 int {x : f (x) > } = ; as is arbitrary, G1 f () = and G0 6 inf R f (). On the other hand,
G1 {x : f (x) > }, so
G1 int {x : f (x) > } f ()
for every f U , R. P
P (i) If f (x) > + (f, x), set = 21 (f (x) (f, x)) > 0. Then there is an
open set G containing x such that |f (y) f (z)| < (f, x) + for every y, z G, so that f (y) > + for
every y G, and
x int{y : f (y) > + } [[T f > ]].
(ii) If f (x) + (f, x) < , set = 21 ( f (x) (f, x)) > 0; then there is an open neighbourhood G of x
such that |f (y) f (z)| < (f, x) + for every y, z G, so that f (y) < for every y G. Accordingly G
does not meet {y : f (y) > } nor {y : f (y) > } for any > , G [[T f > ]] = and x
/ [[T f > ]]. Q
Q
(d) T is additive. P
P Let f , g U and < R. Set = 15 ( ) > 0, H = {x : (f, x) < , (g, x) <
}; then H is the intersection of two dense open sets, so is itself dense and open.
(i) If x H [[T (f + g) > ]], then (f + g)(x) + (f + g, x) ; but (f + g, x) 2 (see (a-i-)
above), so f (x) + g(x) 2 > + 2 and there is a q Q such that
f (x) > q + q + (f, x),
Accordingly
x [[T f > q]] [[T g > q]] [[T f + T g > ]].
328
*364U
Function spaces
Thus H [[T (f + g) > ]] [[T f + T g > ]]. Because H is dense, [[T (f + g) > ]] [[T f + T g > ]].
(ii) If x H, then
x
= (f + g)(x) + 3 > + (f + g, x)
S
H qQ ([[T f > q]] [[T g > q]]) [[T (f + g) > ]].
S
Because H is dense and qQ ([[T f > q]] [[T g > q]]) is open,
[
[[T (f ) > ]] = sup int {x : f (x) > } = sup int {x : f (x) > }
>
>
As is arbitrary, T (f ) = T 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 , 0 then
[[T (f + ) > ]] = sup int {x : f + (x) > } = sup int {x : f (x) > }
>
>
L0
*364U
329
= X if < 1, if 1.
= f W.
1
3
[[|T f | > 0]] = [[T |f | > 0]] = sup>0 int {x : |f (x)| > } =
T
{x : f(x) = } = R [[u > ]]
is also nowhere dense, because supR [[u > ]] = X in RO(X). Accordingly E = int{x : f(x) R} is dense.
Set f (x) = f(x) for x E, 0 for x X \ E.
Let > 0. If G X is a non-empty open set, there is an R such that G 6 [[u > ]], so G1 =
G \ [[u > ]] 6= , and f(x) for every x G1 . Set
= supxG1 f(x) < .
Because E meets G1 , > . Then G2 = G1 [[u > 12 ]] is a non-empty open subset of G and
12 f(x) for every x G2 . Accordingly |f (y) f (z)| 12 for all y, z G2 , and (f, x) < for
all x G2 . As G is arbitrary, {x : (f, x) < } is dense; as is arbitrary, f U .
Take < in R, and set = 21 ( ). Then H = E {x : (f, x) < } is a dense open set, and
H [[T f > ]] H {x : f (x) + (f, x) } E {x : f (x) > }
{x : f(x) > } [[u > ]].
so [[u > ]] [[T f > ]]. 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
*364V
Function spaces
*364V Compact spaces Suppose now that X is a compact Hausdor 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 : (f, x) = 0} = nN Hn
for every x X. So for any > 0, int{x : |f (x)| } {x : (f, x) < } 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 Hausdor extremally disconnected space, and RO(X) 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 (RO(X)) dened by saying
that
[[Sg > ]] = {x : g(x) > }
on C dened by
for every 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,
nite}. Scalar multiplication in L0 corresponds to the operation
(g)(x) = g(x) for x X, g C , R
proof (a) For g C , set Hg = {x : g(x) R}, so that Hg is a dense open set, and dene 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 > ]] = {x : g(x) > }
for every R. P
P (i) (g, x) = 0 for every x Hg , so, if > ,
by the formula in part (c) of the proof of 364U. As [[Sg > ]] is open and Hg is dense,
[[Sg > ]] Hg [[Sg > ]] {x : g(x) } {x : g(x) > }.
Now
[[Sg > ]] = sup> [[Sg > ]] = int
>
(ii) In the other direction, Hg {x : g(x) > } [[Sg > ]], by the other half of the formula in the proof of
364U. Again because {x : g(x) > } is open and Hg is dense,
{x : g(x) > } [[Sg > ]] = [[Sg > ]]
because X is extremally disconnected (see 314S). Q
Q
(b) Thus we have a function S dened by the formula oered. Now if g, h C and g h, we surely
have {x : g(x) > } {x : h(x) > } for every , so [[Sg > ]] [[Sh > ]] for every 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 R such that
g(x0 ) > > h(x0 ); set H = {x : g(x) > > h(x)}; this is a non-empty open set and H [[Sg > ]]. On
the other hand, H {x : h(x) > } = so H [[Sh > ]] = . Thus [[Sg > ]] 6 [[Sh > ]] and Sg 6 Sh. Q
Q
In particular, S is injective.
L0
364Xe
331
(c) S is surjective. P
P If u L0 , set
>
is also open, because all the sets [[u > ]] are open-and-closed. So g : X [, ] is continuous. Also
S
{x : g(x) > } = R [[u > ]],
S
{x : g(x) < } = R X \ [[u > ]]
are dense, so g C . Now, for any R,
[[u > ]]
>
>
>
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 ,
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 nite} 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 species
f1 and f2 as members 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(g) = Rg (at least, unless = 0, which is trivial), so that
S(g) = Sg for every g C .
364X Basic exercises > (a) Let A be a Dedekind -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, u , v, v L0 ,
[[u u ]] [[v v ]] [[u + v u + v ]] and [[u = u ]] [[v = v ]] [[u v = u v ]].
(b) Let A be a Dedekind -complete Boolean algebra. (i) Show that if u, v L0 = L0 (A) and ,
R then [[u + v + ]] [[u ]] [[v ]]. (ii) Show that if u, v (L0 )+ and , 0 then
[[u v ]] [[u ]] [[v ]].
(c) Let A be a Dedekind -complete Boolean algebra and u L0 (A). Show that {[[u E]] : E R is
Borel} is the -subalgebra of A generated by {[[u > ]] : R}.
> (d) Let (A,
) be a probability algebra. Show that for any u L0 (A) there is a unique Radon probability
measure on R (the distribution of u) such that E =
[[u E]] for every Borel set E R. (Hint: 271B.)
(e) Let (A,
) be a probability algebra, and hui iiI any family in L0 (A); for each i I let Bi be the
closed subalgebra of AQgenerated by {[[ui > ]] : R}. Show that the following are equiveridical: (i)
[[ui > i ]] whenever J I is nite and i R for each i J (ii) hBi iiI is
332
Function spaces
364Xf
(f ) Let (A,
) be a probability algebra and u, v two
-independent members of L0 (A). Show that the
distribution of their sum is the convolution of their distributions. (Hint: 272S.)
> (g) Let A be a Dedekind -complete Boolean algebra and g, h : R R Borel measurable functions.
= gh,
where g, h
: L0 L0 are dened as in 364I. (ii) Show that g + h(u) = g(u) + h(u),
(i) Show that gh
0
0
g h(u) = g(u) h(u) for every u L = L (A). (iii) Show that if hhn inN is a sequence of Borel
n (u) = h(u)
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 -complete Boolean algebra. (i) Show that S(A) can be identied () with the
set of those u L0 = L0 (A) such that {[[u > ]] : R} is nite () with the set of those u L0 such that
[[u I]] = 1 for some nite I R. (ii) Show that L (A) can be identied with the set of those u L0 such
that [[u [, ]]] = 1 for some 0, and that kuk is the smallest such .
(i) Show that if A is a Dedekind -complete Boolean algebra, and u L0 (A), then for any R
(compare 363Xh).
(j) Let A be a Dedekind -complete Boolean algebra and u 0 in L0 = L0 (A). Show that u =
supqQ q[[u > q]] in L0 .
(k) (i) Let A be a Dedekind -complete Boolean algebra and A L0 (A) a non-empty countable set with
supremum w. Show that [[w E]] supuA [[u E]] for every Borel set E R. (ii) Let (A,
) be a localizable
measure algebra and A L0 (A) a non-empty set with supremum w. Show that [[w E]] supuA [[u E]]
for every Borel set E R.
(l) Let A be a Dedekind -complete Boolean algebra and A L0 = L0 (A) a non-empty set which is
bounded below in L0 . Suppose that 0 () = inf uA [[u > ]] is dened in A for every R. Show that
v = inf A is dened in L0 , and that [[v > ]] = sup> 0 () for every R.
(m) Let (X, , ) be a measure space and f : X [0, [ a function; set A = {g : g L0 (), g a.e. f }.
(i) Show that if (X, , ) either is localizable or has the measurable envelope property (213Xl), then sup A
is dened in L0 (). (ii) Show that if (X, , ) is complete and locally determined and w = sup A is dened
in L0 (), then w A.
(n) Let A be a Dedekind -complete Boolean algebra. Show that if u, v L0 = L0 (A) then the following
are equiveridical: () [[|v| > 0]] [[|u| > 0]] () v belongs to the band of L0 generated by u () there is a
w L0 such that u w = v.
(o) Let A be a Dedekind -complete Boolean algebra and a A; let Aa be the principal ideal of A
generated by a. Show that L0 (Aa ) can be identied, as f -algebra, with the band of L0 (A) generated by a.
(p) Let A and B be Dedekind -complete Boolean algebras, and : 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 {a : a A, a = 0} (ii) the set of values of T is the sequentially order-closed Riesz subspace of L0 (B)
generated by {(a) : a A}.
(q) Let A and B be Dedekind -complete Boolean algebras, and : 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 dened on a Borel subset of R. Show that h(T
whenever
0
364Yh
L0
333
334
Function spaces
364Yi
(i) Let X be a Baire space, RO(X) its algebra of regular open sets, M its ideal of meager sets, and Bb
b
the Baire property -algebra {GA : G X is open, A M}, so that RO(X) can be identied with B/M
(314Yd). (i) Repeat the arguments of 364V in this context. (ii) Show that the space U of 364U-364V is
b and that W = U W where W = {f : f RX , {x : f (x) 6= 0} M}, so that the
a subspace of L0 (B),
representations of L0 (RO(X)) as U/W , L0 /W are consistent.
(j) Work through the arguments of 364U and 364Yi for the case of compact Hausdor X, seeking
simplications based on 364V.
(k) Let X be an extremally disconnected compact Hausdor space with regular open algebra RO(X).
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 (RO(X)) is represented
by a member of U0 .
(l) Let X be a Baire space. Let Q be the set of all continuous real-valued functions dened on subsets
of X, and Q the set of all members 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 G set. (ii) Show that
f g,
.f
we can dene addition and multiplication and scalar multiplication on Q by saying that f +g,
are to be the unique members of Q extending the partially-dened functions f + g, f g, f , and that
these denitions render Q an f -algebra if we say that f g i 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 dened in 364U, and that these
extensions dene an isomorphism between Q and L0 (RO(X)), where RO(X) is the regular open algebra of
X. (iv) Show that if X is compact, Hausdor and extremally disconnected, then every member of Q has
a unique extension to a member of C (X), as dened in 364W.
(m) Let X be an extremally disconnected Hausdor space, and Z any compact Hausdor 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 -complete Boolean algebra. Write L0C = {u + iv : u, v L0 } for the
complexication of L0 = L0 (A) as dened in 354Yl. (i) Writing B(C) for the Borel -algebra of C, show
that there is a one-to-one correspondence between sequentially order-continuous Boolean homomorphisms
: B(C) A and members w = u + iv of L0C dened by saying that [[u > ]] [[v > ]] = {z : Re z >
, Im z > }. (ii) Show that if is a -algebra of subsets of a set X and : A is a surjective
sequentially order-continuous Boolean homomorphism with kernel I, then we can identify L0C with L0C /W,
where L0C is the set of -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 diculties 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 dierent 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 dierence, because it aects
the proof of the central theorem 364E. The point is that a given algebra A will be expressible in the form
/I for a variety of algebras of sets. Consequently any denition of L0 (A) as a quotient L0 ()/W must
include a check that the structure produced is independent of the particular pair , 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).
L1
365B
335
I have deliberately chosen the notation [[u > ]] 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 terms of the sets {x : f (x) > }, so can an element u of L0 (A) be described in terms of
the elements [[u > ]] of A where in some sense u is greater than . 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 terms of the values of [[u > ]], however, as in
364E, the fundamental result on the action of Boolean homomorphisms (364R) is elegant and reasonably
straightforward.
The concept [[u > ]] can be dramatically generalized to the concept [[(u1 , . . . , un ) E]], where E is
a Borel subset of R n 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 () rather than of L0 (), then [[u E]] is the appropriate translation of X 1 [E].
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 denition of L0 (A), as the set of sequentially order-continuous Boolean
homomorphisms from the Borel -algebra of R to A. This is suitably independent of realizations of A, but
makes the f -algebra structure of L0 dicult to elucidate, unless we move to a further level of abstraction
in the denitions, as in 364Ye.
I take the space to describe the L0 spaces of general regular open algebras in detail (364U) partly to oer
a demonstration of an appropriate technique, and partly to show that we are not limited to -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 Hausdor
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 Hausdor 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 rst step is to establish a denition which can be matched both with the
constructions of the previous sections and with the denition of L1 () 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 (365L-365N). As in 361 and 363, the construction is associated with
universal mapping theorems (365I-365K) which dene 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,
) be a measure algebra. For u L0 (A), write
kuk1 =
R
0
[[|u| > ]] d,
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 L1 = L1 (A,
) = {u : u L0 (A), kuk1 < }.
1
It is convenient to note at once that if u L , then [[|u| > ]] must be nite for almost every > 0,
and therefore for every > 0, since it is a non-increasing function of ; so that [[u > ]] also belongs to the
Boolean ring Af = {a :
a < } for every > 0.
336
365B
Function spaces
365B Theorem Let (X, , ) be a measure space with measure algebra (A,
). Then the canonical
isomorphism between L0 () and L0 (A) (364Jc) matches L1 () L0 (), dened in 242, with L1 (A,
)
L0 (A), and the standard norm of L1 () with k k1 : L1 (A,
) [0, [, as dened in 365A.
proof Take any -measurable function f : X R (364C); write f for its equivalence class in L0 (), and
u for the corresponding element of L0 (A). Then [[|u| > ]] = {x : |f (x)| > } in A for every R, and
kuk1 =
R
0
{x : |f (x)| > } d =
|f |d
proof (A,
) is isomorphic to the measure algebra of some measure space (X, , ) (321J). L1 () is a solid
linear subspace of L0 () (242Cb), so L1 is a solid linear subspace of L0 (A). L1 () is an L-space (354M), so
L1 also is. The rest of the properties claimed are general features of L-spaces (354N, 354E, 356P).
365D Integration Let (A,
) be any measure algebra.
(a) If u L1 = L1 (A,
), then u+ and u belong to L1 , and we may set
R
u = ku+ k1 ku k1 =
R
0
[[u > ]] d
R
0
[[u > ]] d.
Now
: L R is an order-continuous positive linear functional (356Pc), and under the translation of
365B matches the integral on L1 () as dened in 242Ab. Note that if a Af then
a =
R
0
R1
0
a d =
a,
is totally nite then the integral here agrees with that of 363L on L (A). I will sometimes write
Rso that if
u d
if it seems helpful to indicate the measure.
R
R
(b) Of course kuk1 = |u| | u| for every u L1 .
R
R
(c) If u L1 , a A we may set a u = u a. (Compare 242Ac.) If > 0 and 0 6= a [[u > ]] then
there is a > such that a = a [[u > ]] 6= 0, so that
u=
R
0
R
0
a d +
a >
a.
u < 0. Q
Q
u a = u a =
(a [[u ]])d
a < 0,
R
u = a v for every a Af then u = v (cf. 242Ce).
R
R
R
P Of course u a u so a u u for every a A.
(iii) If u 0 in L1 then u = sup{ a u : a Af }. P
n
OnRthe other hand,
R setting an = [[u > 2 ]], hu an inN is a non-decreasing sequence with supremum u,
an is nite for every n. Q
Q
so u = limn an u, while
(ii) If u, v L1 and
R
1
1
R (e) If u L , u 0 and u = 0 then u = 0 (put 365B and 122Rc together). If u L , u 0 and
u = 0 then u a = 0, that is, a [[u > 0]] = 0.
a
L1
365E
337
(fR) If C L1 is non-empty
and upwards-directed and supvC
R
and sup C = supvC v (356Pc).
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.
= 1 if < 0.
If a A, then
=
u
+ d
a d
(a [[
u+
d > ]])d =
0
k
u+
d k1
R
0
d (a [[u+
d > ]])d =
ad
d ;
u+
d d
1
+
= ku+
d L .
d k1 = [[ud > 0]] is nite, so that u
P
) For any nite I D, set vI = dI u
+
(
d . Then
1
consequently the upwards-directed
D is nite} is bounded above
in
R we can set
R
R set A
PL , and
P= {vRI : I
+
1
v = sup A in L . If a A, then a vI = dI ad ud for each nite I D, so a v = dD ad u+
d.
Applying the same arguments to , there is a w L1 such that
R
R
P
w
=
u
dD
a
ad d
u
=
u
u
=
u
=
d
dD
dD
dD (a d)
d
d
a
ad
ad
ad
for every a A.
( ) Now take any a Af . For J D set aJ = supdJ a d. Let > 0. Then there is a nite I D
such that
R
P
P
| a u aJ | = | dD (a d) dJ (a d)|
338
365E
Function spaces
A = {a \ aJ : I J D, J is nite}.
|a aJ | = |(a \ aJ )| .
Consequently
As is arbitrary, a =
|a
a
u| |a aJ | + |
a
u aJ | 2.
R
(ii)(i) From where Rwe now are, this is nearly trivial. Thinking of a as u a, is surely additive
and bounded. Also |a| |u| a. If A Af is non-empty, downwards-directed and has inmum 0, the
same is true of {|u| a : a A}, because a 7 |u| a is order-continuous, so
inf aA |a| inf aA
365F It will be useful later to have spelt out the following elementary facts.
Lemma Let (A,
) be a measure algebra. Write S f for the intersection S(A) L1 (A,
). Then S f is a
1
norm-dense and order-dense Riesz subspace of L (A,
), and can be identied with S(Af ). The function
f
f
1
: A S L (A,
) is an injective order-continuous additive lattice homomorphism. If u 0 in
L1 (A,
), there is a non-decreasing sequence hun inN in (S f )+ such that u = supnN un = limn un .
proof As in 364L,Pwe can think of S(Af ) as a Riesz subspace of S = S(A), embedded in L0 (A).
Pn If u S,
n
it is expressible as i=0 i ai where a0 , . . . , an A are disjoint and no i is zero. Now |u| = i=0 |i |ai ,
so u L1 i
ai < for every i, that is, i u S(Af ); thus S f = S(Af ).
Now suppose that u 0 in L1 . By 364Kd, there is a non-decreasing sequence hun inN in S(A)+ such
that u0 0 and u = supnN 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 inN is norm-convergent to u. This shows also that S f is order-dense in
L1 .
The map : Af S f is an injective order-continuous additive lattice homomorphism; because S f is
regularly embedded in L1 (352Ne), 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 dierence 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, as in
242Mb.
365G Semi-finite algebras: Lemma Let (A,
) be a measure algebra.
(a) (A,
) is semi-nite i L1 = L1 (A,
) 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,
) is semi-nite 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,
) is semi-nite, by the other half of 364L.
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 = supvC v by 365Dh.
365H Measurable transformations We have a generalization of the ideas of 235 in this abstract
context.
Theorem Let (A,
) and (B, ) be measure algebras, and : A B a sequentially order-continuous
Boolean homomorphism. Let T : L0 (A) L0 (B) be the sequentially order-continuous Riesz homomorphism
associated with (364R).
R
=
aRwhenever a A and
a < . Then for any
(a) Suppose that w R0 in L0 (B) is such that a w d
.
is dened and equal to a u d
u L1 (A,
) and a A, a T u w d
L1
365I
339
R
R
0
w
d
(a)
for
every
a
A.
Then
T u d
=
(b)
Suppose
that
w
0
in
L
(A)
is
such
that
a
R
0
u w d
whenever u L (A) and either integral is dened in [, ].
T u w = supnN
T un w = supnN
un =
u.
Tu w =
T u+ w
T u w =
u+
u =
u.
(b) The argument follows the same lines: start with u = a 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+ w u w = (u w )+ (u w ) = u w
whenever either side is dened in [, ].
365I Theorem Let (A,
) be a measure algebra and U a Banach space. Let : Af U be a function.
Then the following are equiveridical:
(i) there is a continuous linear operator T from L1 = L1 (A,
) to U such that a = T (a) for
every a Af ;
(ii)() is additive
() there is an M 0 such that kak M
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-).
If a A then kak1 =
a (using the formula in 365A, or otherwise), so
(c) The argument in (b) shows that T0 = T S f and T are uniquely dened from . We have also seen
that if T , correspond to each other then
kak kT k
a for every a Af ,
340
Function spaces
so that
365I
kT k M whenever kak M
a for every a Af ,
kT k = min{M : M 0, kak M
a for every a Af }.
365J Corollary Let (X, , ) be a measure space and U any Banach space. Set f = {E : E , E <
}. Let : f U be a function. Then the following are equiveridical:
(i) there is a continuous linear operator T : L1 () U such that E = T (E) for every
E f ;
(ii)() (E F ) = E + F whenever E, F f and E F = 0 () there is an M 0 auch
that kEk M E for every E f .
Moreover, in this case, T is unique and kT k is the smallest number M satisfying the condition in (ii-).
proof This is a direct translation of 365I. The only point to note is that if satises the conditions of (ii), and
E, F f are such that E = F in the measure algebra (A,
) of (X, , ), then (E \ F ) = (F \ E) = 0,
so that (E \ F ) = (F \ E) = 0 (using condition (ii-)) and
E = (E F ) + (E \ F ) = (E F ) + (F \ E) = F .
Af = {a : a A,
a < } = {E : E f },
1
(ii) Now suppose that a 0 in U for every a Af , and let u
Pn0 in L , > 0 in R. Then there is a
v S f such that 0 v u and ku vk1 (365F). Express v as i=0 i ai where ai Af , i 0 for
each i. Now
kT u T vk kT kku vk1 kT k.
Tv =
Pn
i=0
i ai U + .
L1
365M
341
|T S f | = T1 S f |T |S f . Q
Q
for every a Af .
(ii) Now suppose that is a lattice homomorphism. In this case T S f is a Riesz homomorphism
(361Gc), that is, |T v| = T |v| for every v S f . Because S f is norm-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.
365L The duality between L1 and L Let (A,
) be a measure algebra, and set L1 = L1 (A,
),
L = L (A). If we identify L with the solid linear subspace of L0 = L0 (A) generated by e = 1A (364K),
then we have a bilinear map (u, v) 7 u v : L1 L L1 , because |u v| kvk |u| and L1 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
R
at most 1 (253A). 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, dened by the formula
(Su)(v) = (T v)(u) =
u v for u L1 , v L .
Because L1 and L are both Banach lattices, we have (L1 ) = (L1 ) and (L ) = (L ) (356Dc).
Because the norm of L1 is order-continuous, (L1 ) = (L1 ) (356Dd).
365M Theorem Let (A,
) be a measure algebra, and set L1 = L1 (A,
), L = L (A). Let S : L1
(L ) = (L ) , T : L (L1 ) = (L1 ) = (L1 ) be the canonical maps dened by the duality between
L1 and L , as in 365L. Then
(a) S and T are order-continuous Riesz homomorphisms, S[L1 ] (L ) , S is norm-preserving and
T [L ] is order-dense in (L1 ) ;
(b) (A,
) is semi-nite i T is injective, and in this case T is norm-preserving, while S is a normed Riesz
space isomorphism between L1 and (L ) ;
(c) (A,
) is localizable i T is bijective, and in this case T is a normed Riesz space isomorphism between
L and (L1 ) = (L1 ) = (L1 ) .
proof (a)(i) If u 0 in L1 and v 0 in L then u v 0 and
(T v)(u) =
u v 0.
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)(a (1 \ a)) =
So S is norm-preserving.
1\a
u=
u v = inf vA ku vk1 = 0
342
365M
Function spaces
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 inmum 0, then
inf vA (Su)(v) = inf vA
u v = 0.
v = a = h(a) > 0,
1
n ]]
a. If
= 0.
(ii) If (A,
) is semi-nite, take any v L . Then if 0 < kvk , a = [[|v| > ]] 6= 0. Let b a be
such that 0 <
b < . Then b L1 , and
because |v| b b, so
As 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,
) is semi-nite, S[L1 ] = (L ) . P
P Take any h (L ) . For a A, set
u a =
u = a = h(a)
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,
) is localizable, then (A,
) is semi-nite.
Given this, if T is bijective, then it is a Riesz space isomorphism between L and (L1 ) , which is Dedekind
complete (356B); so 363Mb tells us that A is Dedekind complete and (A,
) is localizable. In the other
direction, if (A,
) is localizable, then L is Dedekind complete. As T is injective, T [L ] is, in itself,
Dedekind complete; being an order-dense Riesz subspace of (L1 ) (by (a) here) it must be solid (353K); as
it contains T (1), which is the standard order unit of the M -space (L1 ) , it is the whole of (L1 ) , and T
is bijective.
L1
365O
343
then T = T T : L1 (A,
) L1 (C, ).
proof Throughout the proof I will write T for T and S f for S(A) L1
= S(Af ) (see 365F).
(a)(i) We have a map : Af L1 dened by writing a = (a) for a Af . Because
(a b) = (a b) = a + b,
k(a)k1 = (a) =
a
Because S f is dense in L1 and u 7 kuk1 is continuous (in both L1 and L1 ), kT uk1 = kuk1 for every u L1 ,
that is, T is norm-preserving. As noted in 365Ka, T is order-continuous.
(iii) If a, b Af then
For a Af ,
T u = kT u+ k1 kT u k1 = ku+ k1 ku k1 =
Tu =
a
(c) If u S f , express it as
Pn
i=0
T u a =
T (u a) =
u a =
u.
u.
where I = {i : i n, i > }. For u (L1 )+ , take a sequence hun inN in S f with supremum u; then
supnN T un = T u, so
[[u > ]] = (sup [[un > ]])
nN
344
Function spaces
365O
[[u > ]] = [[u+ > ]] = [[T (u+ ) > ]] = [[(T u)+ > ]] = [[T u > ]]
(ii) Suppose now that is surjective. Then T [L1 ] is a linear subspace of L1 containing b for every
b Bf , so includes S(Bf ). If v (L1 )+ there is a sequence hvn inN in S(Bf )+ with supremum v. For each
n, choose un such that T un = vn . Setting un = supin ui , we get a non-decreasing sequence hun inN such
that vn T un v for every n N. So
supnN kun k1 = supnN kT un k1 kvk1 <
T u = supnN T un = v.
Thus (L1 )+ T [L1 ]; consequently L1 T [L1 ] and T is surjective.
(e) This is an immediate consequence of the uniqueness assertion in (i), because for any a Af
T T (a) = T (a) = (a),
L1
so that T T :
therefore be equal
L1
to T .
is a bounded linear operator taking the right values at elements a, and must
R
(a)-(b) For v L1 , a Af set v (a) = a v. Then v : Af R is additive, bounded (by kvk1 ) and if
A Af is non-empty, downwards-directed and has inmum 0, then
R
|v| a = 0
L1
365P
345
R
R
If v 0 in L1 , then a P v = a v 0 for every a Af , so P v 0 (365D(d-i)); thus P is positive. It
must therefore be norm-continuous and order-continuous (355C, 355Ka).
Again supposing that v 0, we have
kP vk1 =
P v a =
c
P v = supaAf
L1 ,
P v = supaAf
a
v kvk1
Pv =
ca
v=
(ca)
v a =
c
P (v a).
(d) Now suppose that [Af ] is order-dense. Take any v, v L1 such that v v = 0. ?? Suppose, if
possible, that u = P v P v > 0. Take > 0 such that a = [[u > ]] is non-zero. Since
v=
a
Pv
a
u > 0,
P v = supaAf
a
u
ac
P v
ac
v .
X
But c [[v > 0]] and v v = 0 so c v = 0. X
So P v P v = 0. As v, v are arbitrary, P is a Riesz homomorphism (352G).
Next, if v 0 in L1 ,
P v = supaAf
v=
a
|P v| =
P |v| =
|v| = kvk1 ,
u = supcAf
1\a 1
u = supcAf
c\a 1
c\a
b = 0.
So u1 (1 \ a) = 0 and 0 6= [[u1 > 0]] a. Let > 0 be such that [[u1 > ]] 6= [[u1 > 0]], and set a1 =
P v = P (b (a1 )) = P (b) a1 = u1 a1 a u,
because
[[u1 a1 > ]] [[u1 > ]] a1 = 0
so
Pv
a1
Pv =
a1
v=
v. Q
Q
R
R
(ii) Now take any u 0 in L1 , and set B = {v : v L1 , v 0, P v u, v R P v}.
R B is not empty
because it contains 0. If C B is non-empty and upwards-directed, then supvC v u is nite, so C
has a supremum in L1 (365Df). Because P is order-continuous, P (sup C) = sup P [C] u; also
sup C = supvC
v supvC
Pv
P (sup C).
346
365P
Function spaces
Thus sup C B. As C is arbitrary, B satises the conditions of Zorns Lemma, and has a maximal element
v0 say.
R
R
?? Suppose, if possible, that P v0 6= u. By (), there is a v1 > 0 such that P v1 u P v0 , v1 P v1 .
In this case, v0 < v0 + v1 B, which is impossible. X
X Thus P v0 = u; also
kv0 k1 =
v0
P v0 = kP v0 k1 .
(iii) Finally, take any u L1 . By (ii), there are non-negative v1 , v2 L1 such that P v1 = u+ ,
P v2 = u , kv1 k1 ku+ k1 and kv2 k1 ku k1 . Setting v = v1 v2 , we have P v = u. Also we must have
kvk1 kv1 k1 + kv2 k1 ku+ k1 + ku k1 = kuk1 kP kkvk1 = kvk1 ,
w = supb Bf ,b b
w = supb Bf ,b b
P w =
P w.
Q
Expressing w as w+ w , we see that the same is true for every w L1 . Q
Now we can say that P P is a positive linear operator from L1 to L1 such that
f
whenever a A and w
L1 ,
P P w =
a
P w =
a
w=
a
P w
P T u =
T u =
u.
(b) From 365Od, we know that T is surjective, while P T is the identity, so that P = T1 , T = P1 .
As for T1 , 365Oe tells us that T1 = T1 ; so
P1 = T1
1 = T .
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,
) be a probability algebra and B a closed subalgebra; write for the restriction
B. The
identity map from B to A induces operators T : L1 (B, ) L1 (A,
) and P : L1 (A,
) L1 (B, ). If
we take L0 (A) to be dened 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
L1 (B, ) L1 (A,
); L1 (B, ) is a norm-closed and order-closed Riesz subspace of L1 (A,
). P is a positive
linear operator, while P T is the identity, so P is a projection from L1 (A,
) onto L1 (B, ). P is dened by
the familiar formula
Pu =
L1
365T
347
R
R
u) P (h(u)).
h( u) h(u);
and if h(u)
L1 (A,
), then h(P
P
P I repeat the proof of 233I-233J. For each
q Q, take q R such that h(t) hq (t) = h(q) + q (t q) for every t R, so that h(t) = supqQ hq (t) for
[[h(u)
> ]] = [[u h1 [ ], [ ]]] = [[u qQ h1
q [ ], [ ]]]
1
q (u) > ]]
= sup [[u h [ ], [ ]]] = sup [[h
qQ
because
and if h(u)
L1 then
h( u) = supqQ hq ( u) = supqQ
q (u)
h
h(u),
u) = sup
Q
h(P
qQ hq (P u) = supqQ 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,
) as the measure
algebra of a probability space (X, , ) and set T = {E : E , E B}.
(c) I note here a fact which is occasionally useful. If u L1 (A,
) is non-negative, then [[P u > 0]] =
upr([[u > 0]], C), the upper envelope of [[u > 0]] in C as dened in 314V. P
P We have only to observe that, for
c C,
Z
c [[P u > 0]] = 0 c P u = 0
Pu = 0
c
Z
348
365T
Function spaces
1
u
1
d
a u
=
a
proof (a) Because (A, ) is semi-nite, there is a partition of unity D A such that d < for every
1
d D. For
R each d D, the functional a 7 (a d) : RA R is completely additive, so there isR a ud L
0 for every a, ud 0. Because 1\d ud = 0,
= (a d) for every a A. Because a ud d
such that a ud d
[[ud > 0]] d. Now u = supdD ud is dened in L0 . P
P (This is a special case of 368K below.) For n N,
set cn = supdD [[ud > n]]. If d, d D are distinct, then d [[ud > n]] = 0, so d cn = [[ud > n]]. Set
c = inf nN cn . If d D, then
d c = inf nN d cn = inf nN [[ud > n]] = 0.
u d
=
sup
ID is finite
(365Dh)
=
sup
ID is finite
u
I d
XZ
dI
ud d
=
sup
ID is finite
X
dI
(a d) = a.
Note that if a A is non-zero, then a > 0, so a [[u > 0]] 6= 0; consequently [[u > 0]] = 1.
To see that u is unique, observe that if u has the same property then for any d D
u d d
= (a d) =
a
u d d
w a u d
=
w d
=
a for every a A. To relate u and w,
w a d
=
a for every a. But from this we see that w u b = b at least
for every a A, that is, a w u d
when
b < , so that w u = 1 is the multiplicative identity of L0 , and w = u1 .
365X Basic exercises (a) Let (A,
) be a measure algebra, and u L1 . Show that
u=
R
0
[[u > ]] d
R0
R
> (b) Let (A,
) be any measure algebra, and u L1 . (i) Show that kuk1 2 supaAf | a u|. (Hint:
R
R
246F.) (ii) Show that for any > 0 there is an a Af such that | u b u| whenever a b A.
+
> (c) Let U be an L-space. If hun inN is any
R norm-bounded sequence in RU , show that lim inf n un =
supnN inf mn um is dened in U , and that lim inf n un lim inf n un .
(d) Let U be an L-space. Let F be a lter 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 dened in U , and that u0 supF F inf uF u.
Pn R
ui : a0 , . . . , an is a partition of unity in A, u0 , . . . , un A}
R
is nite, 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<i bj , and show that supin ui = i=0 ai ui .)
sup{
i=0
ai
365Xp
L1
349
(f ) Let (A,
) be any measure algebra and : Af R a bounded
additive functional. Show that the
R
following are equiveridical: (i) there is a u L1 such that a = a u for every a Af ; (ii) for every > 0
there is a > 0 such that |a| whenever
a ; (iii) for every > 0, c Af there is a > 0 such that
a whenever a c and
a ; (iv) for every > 0 there are c Af , > 0 such that |a| whenever
f
a A and
(a c) ; (v) limn an = 0 whenever han inN is a non-increasing sequence in Af with
inmum 0.
(g) Let (A,
) and (B, ) be measure algebras, and : A B a sequentially order-continuous Boolean
homomorphism. Let T : L0 (A) L0R(B) be the Riesz homomorphism associated with (364R). Suppose
=
a whenever a A. Show that for any u L0 (A,
),
that
w 0 inRL0 (B) is such that a w d
R
T u w d
= u d
whenever either is dened in [, ].
> (h) Let (A,
) be a measure algebra and a A; write Aa for the principal ideal it generates. Show that
if is the identity embedding of Af Aa into Af , then T , as dened in 365O, identies L1 (Aa ,
Aa ) with
a band in L1 .
> (i) Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras (A,
) and (B, ). Let :
X Y be an inverse-measure-preserving function and : B A the corresponding measure-preserving
homomorphism (324M). Show that T : L1 L1 (365O) corresponds to the map g 7 (g) : L1 ()
L1 () of 242Xd.
(j) Let (A,
) and (B, ) be measure algebras. Let : Af Bf be a ring homomorphism such that,
for some > 0, (a)
a for every a Af . (i) Show that there is a unique order-continuous Riesz
1
1
homomorphism T : L L such that T (a) = (a) whenever a Af , and that kT k . (ii) Show that
[[T u > ]] = [[u > ]] for every u L1 , > 0. (iii) Show that T is surjective i is, injective i is. (iv)
Show that T is norm-preserving i (a) =
a for every a Af .
(k) Let (A,
) and (B, ) be measure algebras, and : A B a measure-preserving Boolean homomorphism. Let T : L1 L1 and P : L1 L1 be the operators corresponding to Af , as described
in 365O-365P, and T : L (A) L (B) the operator corresponding to , as described in 363F. (i) Show
that T (u v) = T u Tv for every u L1 , v L (A). (ii) Show that if is order-continuous, then
R
R
P v u = v Tu for every u L (A), v L1 .
> (l) Let (X, , ) be a probability space, with measure algebra (A,
), and let T be a -subalgebra of
. Set = T, B = {F : F T} A, =
B, so that (B, ) is a measure algebra. Let : B A be
the identity homomorphism. Show that T : L1 L1 (365O) corresponds to the canonical embedding of
L1 () in L1 () described in 242Jb, while P : L1 L1 (365P) corresponds to the conditional expectation
operator described in 242Jd.
b
(m) Let (A,
) be a semi-nite measure algebra, and (A,
) its localization (322P). Show that the natural
b
embedding of A in A induces a Banach lattice isomorphism between L1 and L1 , so that the band algebra
b of A.
of L1 can be identied with the Dedekind completion A
350
Function spaces
365Y
Lp
366B
351
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 () are not much aected by
peculiarities in a measure space (X, , ). In this section I oer an explanation: unlike L0 or L , L1 really
depends only on Af , the ring of elements of nite measure in the measure algebra. (See 365O-365Q, 365Xm
and 365Xo.) Note that while the algebra A is uniquely determined (given that (A,
) is localizable, 365Sa),
the measure
is not; if A is any algebra carrying two non-isomorphic semi-nite measures, the corresponding
L1 spaces are still isomorphic (365Sb). For instance, the L1 -spaces of Lebesgue measure on R, and the
subspace measure [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
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 nite 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,
) to Lp (B, )
p
p
and surjective positive linear operators from L (B, ) to L (A,
) (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,
) be a measure algebra and suppose that 1 < p < . For u L0 (A), dene
0
|u| L (A) by setting
p
Lp = Lp (A,
) = {u : u L0 (A), |u|p L1 (A,
)},
1/p
proof What we really have to check is that if w L0 () corresponds to u L0 (A), then |w|p , as dened
in 244A, corresponds to |u|p as dened in 366A. But this was noted in 364Jb.
Now, because the isomorphism between L0 () and L0 (A) matches L1 () with L1 (365B), we can be sure
that |w|p L1 () i |u|p L1 , and that in this case
1/p R p 1/p
R
kwkp =
|w|p
|u|
=
= kukp ,
as required.
352
366C
Function spaces
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 dened 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 identication of
L with (LRp ) is an order-isomorphism. P
P () If w (Lq )+ and u (Lp )+ then u w 0 in L1 , so
q
(T w)(u) = u w 0, writing T : L (Lp ) for the canonical bijection. As u is arbitrary, T w 0. As w
is arbitrary, T is a positive linear operator. () 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
q
0 (T w)(u) =
w u = 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 identied with
(L ) = (Lq ) . From 356M we see that it is also a consequence of (a) and (b).
q
Because
is semi-nite, 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 , supwA kwkp = .
Lp
366G
353
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 () If p < this is covered by 366C, since kwn kp = sup{ wn v : kvkq 1}. () If
p = then [[wn > 4Rn ]] 6= 0; because
is semi-nite, there is a b [[wn > 4n ]] such that 0 <
b < , and
Q
k 1b bk1 = 1, while wn 1b b 4n . Q
P
Because Lq is complete (363Ba, 366C), v = n=0 2n |vn | is dened in Lq . But now
for every n, so u v
/ L1 .
|u| v 2n
wn vn 2n
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.
Definition Let (A,
) be a measure algebra. Write
) = {u : u L0 (A),
[[|u| > ]] < for every > 0},
M0 = M 0 (A,
M1,0 = M 1,0 (A,
) = {u : u M0 , u a L1 (A,
) whenever
a < }.
366G Lemma Let (A,
) be any measure algebra. Write M 0 = M 0 (A,
), 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 inN in S(Af ) such that u = supnN un .
+
(c) M 1,0 = {u : u L0 , R(|u| 1)
L1 for every > 0} = L1 + (L RM 0 ). R
R
1,0
a < ,
a < , then u v; so if a u = a v whenever
(d) If u, v M
and a u a v whenever
u = v.
proof (a) If u, v M 0 and R, then for any > 0
1+|| ]],
]] [[|v| >
]]
(ii) Suppose that u L0 and (|u| 1)+ L1 for every > 0. Then, given > 0, v = (|u| 21 1)+
L1 , and
[[v > 21 ]] < ; but [[|u| > ]] [[v > 12 ]], so also has nite measure. Thus u M 0 . Next, if a Af ,
then |u a| a + (|u| 1)+ L1 , so u M 1,0 .
(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
354
Function spaces
366G
R
R
u a u ,
(d) Take
to M1,0 ,
a < and
a R
R
R > 0 and set a = [[u u > ]]. Because both u and u belong
that is, a u u 0; so a must be 0 (365Dc). As is arbitrary, u u 0 and u u . If a u = a u for
every a Af , then u u so u = u .
366H Theorem Let (A,
) and (B, ) be measure algebras. Let : Af Bf be a measure-preserving
ring homomorphism.
(a)(i) We have a unique order-continuous Riesz homomorphism T = T : M 0 (A,
) M 0 (B, ) such
f
that T (a) = (a) for every a A .
(ii) [[T u > ]] = [[u > ]] for every u M 0 (A,
) and > 0.
(iii) T is injective and multiplicative.
0
(iv) For pR [1, ]
), T u Lp (B, ) i u Lp (A,
), and in this case kT ukp = kukp .
R and u M (A,
In particular, T u = u whenever u L1 (A,
).
(v) For u M 0 (A,
), T u M 1,0 (B, ) i u M 1,0 (A,
).
(b)(i) We
have
a
unique
order-continuous
positive
linear
operator P = P : M 1,0 (B, ) M 1,0 (A,
)
R
R
1,0
f
such that a P v = a v whenever v M (B, ) and a A .
(ii) If u M 0 (A,
), v M 1,0 (B, ) and v T u M 1,0 (B, ), then P (v T u) = u P v.
(iii) If q [1, [ and v Lq (B, ), then P v Lq (A,
) and kP vkq kvkq ; if v L (B) M 0 (B, ),
P Set A = {u : u S(Af ), 0
Now if u0 0 in M0 , sup{T0 u : u S(Af ), 0 u u0 } is dened in M0 . P
u u0 }. Because u0 = sup A (366Gb),
supuA [[T u > ]] = supuA [[u > ]] = (supuA [[u > ]]) = [[u0 > ]]
Lp
366H
355
sup
[[T u > ]]
uS(Af ),0uu+
0
[[|T u|p > ]] = [[T |u| > 1/p ]] = [[|u| > 1/p ]] = [[|u|p > ]].
So
k|T u|p k1 =
R
0
R
0
) As for the case p = , if u L (A) and = kuk > 0 then [[|u| > ]] = 0, so [[|T u| > ]] =
(
[[|u| > ]] = 0. This shows that kT uk . On the other hand, if 0 < < then a = [[|u| > ]] 6= 0,
and a |u| so (a) |T u|; as a 6= 0 (because (a) =
a > 0), kT uk > . This shows that
kT uk = kuk , at least when u 6= 0; but the case u = 0 is trivial.
( ) Now take any p [1, ], and suppose that u M0 and that T u Lp . Let hun inN 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 inN is bounded above in Lp (366Db), so that |u| and u belong
to Lp .
( ) If u L1 , then
u.
= (a) + |T u| |T u| (a) = |T u| 1B . Q
Q
Consequently
T (|u| 1A )+ = T (|u| |u| 1A ) = (|T u| 1B )+ .
But this means that (|u| 1A )+ L1 i (|T u| 1B )+ L1 . Since this is true for every > 0, 366Gc
tells us that u M1,0 i T u M1,0 .
R
356
366H
Function spaces
v (v0 2 1B )+ + 1B ,
so
a
=
P0 v =
Za
(v0
1B )+
2
+
a
and
[[P0 v > ]]
(v0 1B )+ + (a)
(v0 2 1B )+ .
(v0 n2 1B )+ = 0.
Thus {[[P0 v > ]] : v B} is an upwards-directed set in Af with measures bounded above in R, and
c = supvB [[P0 v > ]] is dened in Af . Also
inf n1 cn inf n1
2
n
So, by 364Mb, P0 [B] has a supremum u0 L0 (A), and [[u0 > ]] = c Af for every > 0, so u0 M0 . If
c Af , then
so u0 M1,0 . Q
Q
u = supvB
c 0
P v = supvB
c 0
v
c
v0 < ,
( ) Now 355F tells us that P0 has a unique extension to an order-continuous positive linear operator
P : M1,0 M1,0 . If v0 0 in M1,0 and a Af , then, as remarked above,
P v0 = sup{ a P0 v : v L1 , 0 v v0 } = sup{ a v : v L1 , 0 v v0 } =
R
R
because P is linear, a P v = a v for every v M1,0 , a Af .
a
v0 ;
Pv =
v whenever v M1,0 , a Af .
P
v
=
P
v
=
(a
)
(c)
=
v T u .
i
i
i
i=0
i=0
c
ca
c
i
Next, we can nd a non-decreasing sequence hun inN in S(Af )+ with supremum u, and
Z
Z
Z
v T un =
sup un P v = sup
sup v T un
nN c
nN c
c nN
Z
Z
=
v sup T un =
v T u,
R
nN
using the order-continuity of T , and . But this means that u P v = supnN un P v is integrable over
R
R
c and that c u P v = c v T u. As c is arbitrary, u P v = P (v T u) M1,0 .
) For general u, v,
(
Lp
366H
357
R
R
f
) I take the case q = 1 rst.
v kvk1 . In
P
v
=
(
In
this
case,
for
any
a
A
,
we
have
a
a
R
n
particular, setting an = [[P v > 2 ]], an P v kvk1 . But P v = supnN P v an , so
kP vk1 = supnN
0, then
an
P v kvk1 .
) Next, suppose that q = , so that v L (B)+ ; say kvk = . ?? If > 0 and a = [[P v > ]] 6=
(
a <
Pv =
a
v (a) =
a. X
X
So [[P v > ]] = 0 and P v L (A), with kP vk kvk , at least when kvk > 0; but the case kvk = 0
is trivial.
( ) I come at last to the general case q ]1, [, v Lq . In this case set p = q/(q 1). If u Lp
then T u Lp so T u v L1 and
|
u P v| ku P vk1
= kP (T u v)k1
(by (ii))
kT u vk1
R
by (a-iii) of this theorem. But this means that u 7 u P v is a bounded linear functional
on LpR, and
R
q
p
is therefore represented by some w L with kwkq kvkq . If a Af then a L , 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 M1,0 and a Af , we must have
P T u = a T u = T (a) T u = T (a u) = a u = a u,
R
R
using (a-iv) to see that a u is dened and equal to T (a u). As a is arbitrary, u M1,0 and
P T u = u.
a
(c) As usual, in view of the uniqueness of T and P , all we have to check is that
T T (a) = T (a) = (a) = T (a),
whenever a Af , w M1,0 .
P P w =
P w =
w=
P w
358
Function spaces
366I
v = 0 for every a Af }.
proof Most of this is simply because T is a norm-preserving Riesz homomorphism (so that T [L2 ] is
isomorphic to L2 ), P T is the identity on L2 (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
T P v = 0 P v = 0
P v = 0 for every a Af
a
Z
v = 0 for every a Af .
a
T 1 u =
u for every u L2 , a Af .
366Yf
Lp
359
b
(e) Let (A,
) be a semi-nite measure algebra and (A,
) its localization (322P). Show that the natural
b
embedding of A in A induces a Banach lattice isomorphism between Lp and Lp for every p [1, [, so that
b
the band algebra of Lp can be identied with A.
b
(f ) Let (A,
) be a semi-nite measure algebra which is not localizable (cf. 211Ye, 216D), and (A,
) its
b
localization. Let : A A be the identity embedding, so that is an order-continuous measure-preserving
b
Boolean
R homomorphism. Show that if we set v = b where b A \ A, then there is no u L (A) such that
R
a < .
u = a v whenever
a
(g) In 366H, show that [[T u E]] = [[u E]] whenever u M0 and E R is a Borel set such that
0
/ E.
P
p 1/p
< }
{u : kuk =
iI ku(i)k
Q
p
of iI Li .
360
Function spaces
366Yg
L
(A) be the associated Riesz space isomorphism (364R), and let w (L1 )+ be such that
1
w = ( a) for every a A (365Ea). Set Q u = T u w for u L0 (A). Show that kQ uk2 = kuk2
a
2
for every u L . (ii) Show that if , : A A are Boolean automorphisms then Q = Q Q .
(i) Let (A,
) be a measure algebra, and : Af Af a measure-preserving Boolean homomorphism,
with associated linear operator T : M0 M0 . Show that the following are equiveridical: (i) there is some
p [1, [ such that {Tn Lp : n N} is relatively compact in B(Lp ; Lp ) for the strong operator topology; (ii)
for every p [1, [, {Tn Lp : n N} is relatively compact in B(Lp ; Lp ) for the strong operator topology;
(iii) { n a : n N} is relatively compact in Af , for the strong measure-algebra topology, for every a Af .
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
reect 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 dierent 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 dierent 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
renements in the hypotheses are irrelevant to the principal applications of the results, and they make
substantial demands on the reader; but I believe that the demands are actually demands to expand ones
imagination, to encompass the dierent 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 dened 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 M1, , as in 363F and
369Xm. If we know that (A,
) 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 cant 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, .
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 terms 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.
367B
Convergence in measure
361
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 diculty of nding convincing denitions; that is, isolating concepts which lead to 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 denition which is applicable in any lattice.
Definition Let P be a lattice and hpn inN a sequence in P , p an element of P . I will say that hpn inN
order*-converges to p, or that p is the order*-limit of hpn inN , if
p = inf{q : n N, q (p pi ) p i n}
= sup{q : n N, q p (pi p ) i n}
whenever p p p in P .
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 inN and hpn inN both order*-converge to p, and pn qn pn for every n, then hqn inN
order*-converges to p.
(e) If hpn inN is an order-bounded sequence in P , then it order*-converges to p P i
p = inf{q : n N, q pi i n}
= sup{q : n N, q pi i n}.
(f) If P is a Dedekind -complete lattice (314A) and hpn inN is an order-bounded sequence in P , then it
order*-converges to p P i
p = supnN inf in pi = inf nN supin pi .
proof (a) Suppose that hpn inN is order*-convergent to both p and p. Set p = p p, p = p p; then
p = inf{q : n N, q (p pi ) p i n} = p.
(b) is trivial.
(c) Suppose that hpn inN is order*-convergent to p, and that hpn inN is a subsequence of hpn inN . Take
p , p such that p p p , and set
B = {q : n N, q p (pi p ) i n},
B = {q : n N, q p (pi p ) i n},
C = {q : n N, q (p pi ) p i n},
362
Function spaces
367B
C = {q : n N, q (p pi ) p i n}.
q p (pi p ) (p pi ) p q .
As p = inf C, we must have q p; thus p is an upper bound for B . On the other hand, {pi : i n} {pi :
i n} for every n, so B B and p must be the least upper bound of B , since p = sup B.
Similarly, p = inf C . As p and p are arbitrary, hpn inN order*-converges to p.
(d) Take p , p such that p p p , and set
B = {q : n N, q p (pi p ) i n},
B = {q : n N, q p (qi p ) i n},
C = {q : n N, q (p pi ) p i n},
C = {q : n N, q (p qi ) p i n}.
q p (qi p ) (p pi ) p q .
As p = inf C, we must have q p; thus p is an upper bound for B . On the other hand, p (pi p )
p (qi p ) for every i, so B B and p = sup B . Similarly, p = inf C . As p and p are arbitrary, hqn inN
order*-converges to p.
(e) Set
B = {q : n N, q pi i n},
C = {q : n N, q pi i n}.
(i) Suppose that hpn inN order*-converges to p. Let p , p be such that p pn p for every n N
and p p p . Then
B = {q : n N, q p (pi p ) i n},
q p (pi p ) p q = q
B = {inf in pi : i N},
then B B and for every q B there is a q B such that q q ; so sup B = sup B if either is dened.
Similarly,
inf{q : n N, q pi i n} = inf nN supin pi
if either is dened.
367C
Convergence in measure
363
367C Proposition Let U be a Riesz space and hun inN , hvn inN two sequences in U order*-converging
to u, v respectively.
(a) If w U , hun + winN order*-converges to u + w, and un order*-converges to u for every R.
(b) hun vn inN order*-converges to u v and hun vn inN order*-converges to u v.
(c) If hwn inN is any sequence in U , then it order*-converges to w U i h|wn w|inN order*-converges
to 0.
(d) hun + vn inN order*-converges to u + v.
(e) If U is Archimedean, and hn inN is a sequence in R converging to R, then hn un inN order*converges to u.
(f) Again suppose that U is Archimedean. Then a sequence hwn inN in U + is not order*-convergent to
0 i there is a w
> 0 such that w
= supin w
wi for every n N.
proof (a)(i) hun + winN order*-converges to u + w because the ordering of U is translation-invariant; the
map w 7 w + w is an order-isomorphism.
u.
( ) If w u w then w u w so
u = inf{w : n N, w ((w ) ui ) (w ) i n}
B = {w : n N, w w ((ui vi ) w ) i n},
C = {w : n N, w (w (ui vi )) w i n},
B1 = {w : n N, w (w u) (ui w ) i n},
B2 = {w : n N, w (w v) (vi w ) i n},
C1 = {w : n N, w ((w u) ui ) w i n},
C2 = {w : n N, w ((w v) vi ) w i n},
If w1 B1 and w2 B2 then w1 w2 B. P
P There is an n N such that w1 (w u) (ui w ) for
= (w (u v)) ((ui vi ) w )
= w ((ui vi ) w )
364
Function spaces
367C
w1 w2 (((w u) ui ) w ) (((w v) vi ) w )
= ((w u) ui (w v) vi ) w
= ((w (u v)) (ui vi )) w
= (w (ui vi )) w
for every i n, so w1 w2 C. Q
Q
At the same time, of course, w w
whenever w B and w
C, since there is some i N such that
w w ((ui vi ) w ) (w (ui vi )) w w.
Since
(using the generalized distributive laws in 352E), we must have sup B = inf C = u v. As w and w are
arbitrary, hun vn inN is order*-convergent to u v.
Putting this together with (a), we see that hun vn inN = h((un ) (vn ))inN order*-converges to
((u) (v)) = u v.
(c) The hard parts are over. (i) If hwn inN order*-converges to w, then hwn winN , hw wn inN and
h|wn w|inN = h(wn w) (w wn )inN all order*-converge to 0, putting (a) and (b) together. (ii) If
h|wn w|inN order*-converges to 0, then so do h|wn w|inN and hwn winN , by (a) and 367Bd; so
hwn inN order*-converges to 0, by (a) again.
(d) h|un u|inN and h|vn v|inN order*-converge to 0, by (c), so h2(|un u| |vn v|)inN 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)|inN order*-converges to 0, by 367Bb and 367Bd, and hun + vn inN
order*-converges to u + v.
(e) Set n = supin |i | for each n. Then hn inN 0, so inf nN n |u| = 0, because U is Archimedean.
Consequently hn |u|inN order*-converges to 0, by 367Be. But we also have 0 |un u| order*-converging to
0, by (c) and (a), so h0 |un u|+n |u|inN order*-converges to 0, by (d). As |n un u| 0 |un u|+n |u|
for every n, hn un inN order*-converges to u, as required.
(f )(i) Suppose that hwn inN is not order*-convergent to 0. Then there are w , w such that w 0 w
and either
B = {w : n N, w w (wi w ) i n}
C = {w : n N, w (w wi ) w i n}
does not have inmum 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 (w wi ) w i n} = {w : w wi w i n}.
in
367F
Convergence in measure
365
(ii) If w
> 0 is such that w
= supin w
wi for every n N, then
{w : n N, w (0 wi ) w
i n}
367D As an example of the use of this concept in a relatively abstract setting, I oer the following.
Proposition Let U be a Banach lattice and hun inN a sequence in U which is norm-convergent to u U .
Then hun inN has a subsequence which is order-bounded and order*-convergent to u. So if hun inN itself is
order*-convergent, its order*-limit is u.
proof Let hun inN be a subsequence of hun inN such that kun uk 2n for each n N. Then
vn = supin |ui u| is dened in U , and kvn k 2n+1 , for each n (354C). Because inf nN kvn k = 0,
inf nN vn must be 0, while u vn ui u + vn whenever i n; so hun inN order*-converges to u, by
367Be.
Now if hun inN 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 inN is order*-convergent to u. Then h|un u|inN is order*-convergent to 0
(367Cc), so
C = {v : n N, v |ui u| i n}
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 inN is a sequence in V and v V , then hvn inN
order*-converges to v when regarded as a sequence in V , i it order*-converges to v when regarded as a
sequence in U .
proof (a) Since, in either V or U , hvn inN order*-converges to v i h|vn v|inN order*-converges to 0
(367Cc), it is enough to consider the case vn 0, v = 0.
(b) If hvn inN is not order*-convergent to 0 in U , then, by 367Cf, there is a u > 0 in U such that
u = supin 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
Then for any w C,
C = {w : w V, n N, w vi vk i n}.
u vk = supin 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 inN cannot
be order*-convergent to 0 in V .
(c) Now suppose that hvn inN is not order*-convergent to 0 in V . Because V also is Archimedean (351Rc),
there is a w > 0 in V such that w = supin 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 inN is not order*-convergent to 0 in U .
366
Function spaces
367G
367G I now spell out the connexion between the denition above and the concepts introduced in 245C.
Proposition Let X be a set, a -algebra of subsets of X, A a Boolean algebra and : A a
sequentially order-continuous surjective Boolean homomorphism; let I be its kernel. Write L0 for the
space of -measurable real-valued functions on X, and let T : L0 L0 = L0 (A) be the canonical Riesz
homomorphism (364D, 364R). Then for any hfn inN and f in L0 , hT fn inN order*-converges to T f in L0
i X \ {x : f (x) = limn fn (x)} I.
proof Set H = {x : limn fn (x) exists = f (x)}; of course H . Set gn (x) = |fn (x) f (x)| for n N
and x X.
(a) If X \ H I, set hn (x) = supin gi (x) for x H and hn (x) = 0 for x X \ H. Then hhn inN
is a non-increasing sequence with inmum 0 in L0 , so inf nN T hn = 0 in L0 , because T is sequentially
order-continuous (364Ra). But as X \ H I, T hn T gi = |T fi T f | whenever i n, so h|T fn T f |inN
order*-converges to 0, by 367Be or 367Bf, and hT fn inN order*-converges to T f , by 367Cc.
(b) Now suppose that hT fn inN order*-converges to T f . Set gn (x) = min(1, gn (x)) for n N, x X; then
hT gn inN = he |T fn T f |inN order*-converges to 0, where e = T (X). By 367Bf, inf nN supin T gi = 0
in L0 . But T is a sequentially order-continuous Riesz homomorphism, so T (inf nN supin gi ) = 0, that is,
belongs to I.
proof (a) We can express A as a quotient /I of a -algebra of sets, in which case L0 can be identied
with the canonical image of L0 = L0 () (364D). If hun inN is an order*-convergent sequence in L0 , then
it is expressible as hT fn inN , where T : L0 L0 is the canonical map, and 367G tells us that hfn (x)inN
converges for every x H, where X \ H I. If we set h(x) = supnN |fn (x)| for x H, 0 for x X \ H,
then we see that |un | T h for every n N, so that hun inN is order-bounded in L0 .
(b) This now follows from 367Bf, because L0 is Dedekind -complete.
367I Proposition Suppose that E R is a Borel set and h : E R is a continuous function. Let A
be a Dedekind -complete Boolean algebra and set QE = {u : u L0 , [[u E]] = 1}, where L0 = L0 (A).
: QE L0 be the function dened by h (364I). Then hh(u
n )inN order*-converges to h(u)
Let h
whenever
hun inN is a sequence in QE order*-converging to u QE .
proof This is an easy consequence of 367G. We can represent A as /I where is a -algebra of subsets
of some set X and I is a -ideal of (314M); let T : L0 L0 (A) be the corresponding homomorphism
(364D, 367G). Now we can nd -measurable functions hfn inN , 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 passing over the trivial case
n ) = T (hfn ) and h(u)
order*-converges to h(u).
367J Dominated convergence We now have a suitable language in which to express an abstract version
of Lebesgues Dominated Convergence Theorem.
Theorem Let (A,
) be a measure algebra. If hun inN is a sequence in L1 = L1 (A,
) which
R is order-bounded
R
and order*-convergent in L1 , then hun inN is norm-convergent to u in L1 ; in particular, u = limn un .
R
proof The norm Rof L1 is order-continuous
(365C), so hun inN is norm-convergent to u, by 367E. As is
R
norm-continuous, u = limn un .
367L
Convergence in measure
367
367K The Martingale Theorem In the same way, we can re-write theorems from 275 in this language.
Theorem Let (A,
) be a probability algebra, and hBn inN a non-decreasing sequence of closed subalgebras
of A. For each n N let Pn : L1 = L1 (A,
) L1 LS0 (Bn ) be the conditional expectation operator (365R);
let B be the closed subalgebra of A generated by nN Bn , and P the conditional expectation operator
onto L1 L0 (B).
(a) If hun inN is a norm-bounded sequence in L1 such that Pn (un+1 ) = un for every n N, then hun inN
is order*-convergent in L1 .
(b) If u L1 then hPn uinN is order*-convergent to P u.
proof If we represent (A,
) 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 terms 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 Hausdor space, and hun inN a sequence in C(X), the space
of continuous real-valued functions on X. Then hun inN order*-converges to 0 in C(X) i {x : x
X, lim supn |un (x)| > 0} is meager. In particular, hun inN order*-converges to 0 if limn un (x) = 0
for every x.
proof (a) The following
elementary fact is worth noting: if A C(X)+ is non-empty and inf A = 0 in
S
P?? If not, take x0 X \ G. Because X is
C(X), then G = uA {x : u(x) < } is dense for every > 0. P
completely regular (3A3Bb), there is a continuous function w : X [0, 1] such that w(x0 ) = 1 and w(x) = 0
XQ
Q
for every x G. But in this case 0 < w u for every u A, which is impossible. X
(b) Suppose that hun inN order*-converges to 0. Set vn = |un | X, so that hvn inN 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 = vB {x : v(x) < 2k }; then Gk is dense, by
(a), and of course is open. So H = kN 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
is meager.
(c) Now suppose that hun inN does not order*-converge to 0. By 367Cf, there is a w > 0 in C(X) such
that w = supin w |ui | for every n N; that is, inf in (w |ui |)+ = 0 for every n. Set
Gn = {x : inf in (w |ui |)+ (x) < 2n } = {x : supin |ui (x)| > w(x) 2n }
nN
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 Baires 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 nd
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).
368
Function spaces
367M
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 367Yb, 367Yk, 3A3P),
but I have taken care to avoid spelling one out at this stage; I will return to the point in 392. 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,
) be a measure algebra. For a Af = {a :
a < } and u L0 =
R
0
L (A) set a (u) = |u| a, a (u) =
(a [[|u| > ]]). Then the topology of convergence in measure
on L0 is dened either as the topology generated by the pseudometrics (u, v) 7 a (u v) or by saying that
G L0 is open i for every u G there are a Af and > 0 such that v G whenever a (u v) .
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 identied
with L0 () for a suitable measure space (X, , ) (321J, 364Jc), everything we know about general spaces
L0 () can be applied directly to L0 (A) for measure algebras (A,
); and that is what I will do for the next
few paragraphs. So far, all I have done is to write a in place of the F of 245Ac, and call on the remarks
in 245Bb and 245F.
367N Theorem (a) For any measure algebra (A,
), 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,
) is semi-nite i T is Hausdor.
(c) (A,
) is localizable i T is Hausdor and L0 is complete under the uniformity corresponding to T.
(d) (A,
) is -nite i T is metrizable.
proof 245D, 245Cb, 245E. Of course we need 322B to assure us that the phrases semi-nite, localizable,
-nite 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,
) be a measure algebra and give L0 = L0 (A) its topology of convergence in
measure.
(a) If A L0 is a non-empty, downwards-directed set with inmum 0, then for every neighbourhood G
of 0 in L0 there is a u A such that v G whenever |v| u.
(b) If U L0 is an order-dense Riesz subspace, it is topologically dense.
(c) In particular, S(A) and L (A) are topologically dense.
R
proof (a) Let a Af , > 0 be such that u G whenever |u| a (see 245Bb). Since {u Ra : u A}
is a downwards-directed set in L1 = L1 with inmum 0 in L1 , there must be a u A such that u a
(365Da). But now [u, u] G, as required.
P If v (L0 )+ , then {u : u U , u v} is an
(b) Write U for the closure of U . Then (L0 )+ U . P
upwards-directed set with supremum u, so A = {v u : u U , u v} is a downwards-directed set with
inmum 0 (351Db). By (a), every neighbourhood of 0 meets A, and (because subtraction is continuous)
Q
every neighbourhood of v meets U , that is, v U . Q
0
Since U is a linear subspace of L (2A5Ec), it includes (L0 )+ (L0 )+ = L0 (352D).
(c) By 364Ka, S(A) and L (A) are order-dense Riesz subspaces of L0 .
367P Theorem Let U be a Banach lattice and (A,
) a measure algebra. Give L0 = L0 (A) its topology
0
of convergence in measure. If T : U L 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 nd u,
hun inN U such that T u G and kun uk 2n , T un
/ G for
u; then H is an
Pevery n. Set H = G T P
open set containing 0 but not T (un u), for any n N. Since n=0 nkun uk < , v = n=0 n|un u|
is dened in U , and |T (un u)| n1 T v for every n 1. But by 367Oa (or otherwise) we know that there is
X
some n such that w H whenever |w| n1 T v, so that T (un u) H for some n, which is impossible. X
367T
Convergence in measure
369
{x : x F, f (x)
/ En } {x : g(x)
/ E}
|h(v)
h(u)|
a = F min(1, |hg(x) hf (x)|)(dx) 2 + F = .
is continuous.
As u, a and 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
terms 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.
370
367T
Function spaces
proof (a) Suppose rst that u A. Take V to be U L1 ; 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,
) is semi-nite,
so that L1 is order-dense in L0 , by 365Ga.)
Take any v V + . For each n N, set an = [[v > 2n ]] Af . Because u A, there is a un A such
that
bn 2n , where
bn = an [[|un u | > 2n ]] = [[|un u | v > 2n ]].
(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
satised.
(b) Now suppose that the condition is satised. Fix a ARf and > 0. Because V is order-dense in U
and therefore in L0 , there is a v V such that 0 v a and v
a . Let B be a downwards-directed
set, with inmumR 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 . Now there is a u A such that |u u | v w, so that
|u u | a +
|u u | v +
w v 2.
canonical isomorphism between L and (L ) . For A L write A for the weak* closure of the image of
A in (L1 ) . Then for every A L1
P [A ] (A),
where (A) is the convex hull of A and (A) is the closure of (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 ) identies L1 with (L1 ) ; and that (L1 )
is a band in (L1 ) . For proofs of these see 365C, 356D and 356B.
Now for the new argument. First, observe that the statement of the theorem involves the measure algebra
(A,
) and the space L0 only in the denition of convergence in measure; everything else depends only on
*367V
371
Convergence in measure
the Banach lattice structure of L1 . And since we are concerned only with the question of whether members
of P [A ], which is surely a subset of L1 , belong to (A), 367T shows that this also can be answered in terms
of the Riesz space structure of L1 . What this means is that we can suppose that (A,
) is localizable. P
P
b
Let (A,
) be the localization of (A,
) (322P). The natural expression of A as an order-dense subalgebra of
b identies Af = {a : a A,
b f (322O), so that L1 becomes identied with L1 , by 365Od.
A
a < } with A
b
Thus we can think of L1 as L1 , and (A,
) is localizable. Q
Q
(b) Take A and set u0 = P ; I have to show that u0 (A). Write R for the canonical map from
L to (L1 ) , so that belongs to the weak* closure of R[A].
Consider rst the case u0 = 0. Take any c Af and > 0. We know that (L1 ) = (L1 ) = (L1 ) can
be identied with L = L (A) (365Mc), so that (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 M of completely additive functionals (363K). Writing
P : M M for the band projection, we must have P () = 0, where M is dened by setting
a = (a) for each a A; consequently P (||) = 0 and there is an upwards-directed family C A, with
supremum 1, such that ||(a) = 0 for every a C (362D). Since
c = supaC
(a c), there is an a C
such that
(c \ a) .
Consider the map Q : L1 L1 dened by setting Qw = w a for every w L1 . Then its adjoint
A
there
are
a
non-zero
b
a
and
a
C
C
such
that
sup
uC b
T
Then C =
6 .
proof Because C has the nite intersection property, there is an ultralter F on L0 including C. Set
I = {a : a A, inf F F supuF
|u| < };
because F is a lter, I is an ideal in A, and the condition on C tells us that I is order-dense. For each a I,
dene Qa : L0 L0 by setting Qa u = u a. Then there is an F F such that Qa [F ] is a norm-bounded
set in L1 , so a = limuF RQa u is dened 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 (a ) 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 (a ) = Qa P (b ). P
P Qa L1 is a band projection on L1 , so its adjoint Qa is a
1
band projection on L = (L ) (356C) and Qa is a band projection on (L )
= (L ) . This means that
Qa will commute with P (352Sb). But also Qa is continuous for the weak* topology of (L ) , so
372
Function spaces
*367V
Qa P (b ) = Qa Qb P (b ) = Qab P (b ) = P (ab ) = Qb P (a ).
Q
What this means is that if we take a partition D of unity included in I (313K), so that L0
= dD L0 (Ad )
(315F(iii), 364S), and dene w L0 by saying that w d = P (d ) for every d D, then we shall have
w a d = P (a ) d whenever a I and d D, so w a = P (a ) for every a I. But now, given
a Af and > 0 and C C, there is a b I such that
(a \ b) ; w b Qb [C], so there is a u C
such that
(b [[|w u| ]]) T
; and T
(inf
[[u
>
]])
=
6
[[ui > i ]]. Set ai = [[ui > i ]] for each i. Let
iJ
i
i
iJ
Q
(inf iJ ai )| 2#(J) and |i
ai | 2 for every i J.
> 0 be such that 6= iJ i whenever |
Let ]0, 1] be such that
[[ui > i + 2]]
ai for every i J.
Now if hvi iiI (L0 )I and
[[|vi ui | > ]] for each i J, hvi iiI is not independent. P
P For each
i J, consider bi = [[vi > i + ]], ai = [[ui > i + 2]]. We have
(364Fa), and
so
has measure at most 2. It follows that (inf iJ bi ) (inf iJ ai ) has measure at most 2#(J), and |
(inf iJ bi )
(inf iJ bi ) 6= iJ
bi , and hvi iiI is not independent. Q
Q
This shows that the set of non-independent families is open in (L0 )I , so that the set of independent
families is closed, as claimed.
367X Basic exercises > (a) Let P be a lattice. (i) Show that if p P and hpn inN is a non-decreasing
sequence in P , then hpn inN is order*-convergent to p i p = supnN pn . (ii) Suppose that hpn inN is a
sequence in P order*-converging to p P . Show that p = supnN p pn = inf nN p pn . (iii) Let hpn inN ,
hqn inN 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 inN be a sequence in P . Show that hpn inN order*-converges to p P i
hpn pinN and hpn pinN order*-converge to p.
(b) Let P and Q be lattices, and f : P Q an order-preserving function. Suppose that hpn inN is an
order-bounded sequence which order*-converges to p in P . Show that hf (pn )inN order*-converges to f (p)
in Q if either f is order-continuous or P is Dedekind -complete and f is sequentially order-continuous.
367Xq
Convergence in measure
373
(c) Let P be either a Boolean algebra or a Riesz space. Suppose that hpn inN is a sequence in P such
that hp2n inN and hp2n+1 inN are both order*-convergent to p P . Show that hpn inN is order*-convergent
to p. (Hint: 313B, 352E.)
> (d) Let A be a Boolean algebra and han inN , hbn inN two sequences in A order*-converging to a, b
respectively. Show that han bn inN , han bn inN , han \ bn inN , han bn inN order*-converge to a b, a b,
a \ b and a b respectively.
(e) Let A be a Boolean algebra and han inN a sequence in A. Show that han inN does not order*-converge
to 0 i there is a non-zero a A such that a = supin a ai for every n N.
> (f ) (i) Let U be a Riesz space and hun inN 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 inN an orderbounded order*-convergent sequence in U with limit u. Show that h(u) = limn h(un ) for every h U .
(Compare 356Xd.)
> (g) Let U be a Riesz space with a Fatou norm k k. (i) Show that if hun inN is an order*-convergent
sequence in U with limit u, then kuk lim inf n kun k. (Hint: h|un | |u|inN is order*-convergent to
|u|.) (ii) Show
P that if hun inN is a norm-convergent sequence in U it has an order*-convergent subsequence.
(Hint: if n=0 kun k < then hun inN 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 inN is a sequence in U which order*-converges to u U , then hT un inN order*-converges
to T u in V .
(i) Let A be a Boolean algebra and B a regularly embedded subalgebra. Show that if hbn inN is a
sequence in B and b B, then hbn inN order*-converges to b in B i it order*-converges to b in A.
(j) Let A be a Dedekind -complete Boolean algebra and hun inN , hvn inN two sequences in L0 (A) which
are order*-convergent to u, v respectively. Show that hun vn inN order*-converges to u v. Show that if
1
1
u, un all have multiplicative inverses u1 , u1
.
n then hun inN order*-converges to u
(k) Let A be a Dedekind -complete Boolean algebra and I a -ideal of A. Show that for any han inN ,
a A, han inN order*-converges to a in A/I i inf nN supmn am a I.
> (l) Let A be a Dedekind -complete Boolean algebra, and hhn inN a sequence of Borel measurable
n (u)inN
functions from R to itself such that h(t) = limn hn (t) is dened for every t R. Show that hh
0
0
0
0
374
Function spaces
367Xr
(r) Let (X, , ) be a semi-nite measure space. Show that L0 () is separable for the topology of
convergence in measure i is -nite and has countable Maharam type. (Cf. 365Xp.)
(s) Let (A,
) be a measure algebra. (i) Show that if han inN is order*-convergent to a A, then
han inN a for the measure-algebra topology. (ii) Show that if (A,
) is -nite, then () a sequence
converges to a for the topology of A i every subsequence has a sub-subsequence which is order*-convergent
to to a () a set F A is closed for the topology of A i a F whenever there is a sequence han inN in F
which is order*-convergent to a A.
(t) Let (A,
) be a measure algebra which is not -nite. 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 inN is
norm-convergent to u U i 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 i u F whenever there is an
order-bounded sequence hun inN in F order*-converging to u U .
(v) Let (A,
) be a probability algebra. For u L0 = L0 (A) let u be the distribution of u (364Xd).
Show that u 7 u 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,
) be a probability algebra and hun inN a stochastically independent sequence in L0 (A), all
with the Cauchy distribution C,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 C,1 . (Hint: consider rst u C0 .) Show that C0 is bounded for the topology of
T
convergence in measure. Show that nN Cn = .
367Y Further exercises (a) Give an example of an Archimedean Riesz space U and an order-bounded
sequence hun inN in U which is order*-convergent to 0, but such that there is no non-increasing sequence
hvn inN , with inmum 0, such that un vn for every n N.
(b) Let P be any lattice. (i) Show that there is a topology on P for which a set A P is closed i
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 dened in the
same way, and f : P Q an order-preserving function. Show that if f is topologically continuous it is
sequentially order-continuous.
(c) Give an example of a distributive lattice P with p, q P and a sequence hpn inN , order*-convergent
to p, such that hpn qinN is not order*-convergent to p q.
(d) Let us say that a lattice P is (2, )-distributive if () whenever A, B P are non-empty sets with
inma p, q respectively, then inf{a b : a A, b B} = p q () 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 inN
order*-converges to p and hqn inN order*-converges to q, hpn qn inN order*-converges to p q.
367Yr
Convergence in measure
375
(e) (i) Give an example of a Riesz space U with an order-dense Riesz subspace V of U and a sequence
hvn inN in V such that hvn inN 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 inN in V ,
order-bounded in V , such that hvn inN order*-converges to 0 in U but does not order*-converge in V .
(f ) Let U be an Archimedean f -algebra. Show that if hun inN , hvn inN are sequences in U order*converging to u, v respectively, then hun vn inN order*-converges to u v.
(g) Let A be a Dedekind -complete Boolean algebra and r 1. Let E R r 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 = L0 (A) the corresponding map (364Yd). Show that if hun inN is a sequence in QE which
and h
n )inN is order*-convergent to h(u).
(l) Give an example of a set X and a double sequence humn im,nN 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 ikN in {umn : m, n N} such that limk vk (x) = 0 for every x.
(m) Let U be a Banach lattice with an order-continuous norm. For v U + dene v : U U [0, [
by setting v (u1 , u2 ) = k|u1 u2 | vk for all u1 , u2 U . Show that every v is a pseudometric on U , and
that {v : v U + } denes a Hausdor linear space topology on U .
(n) Let U be any Riesz space. For h (Uc )+ (356Ab), v U + dene vh : U U [0, [ by setting
vh (u1 , u2 ) = h(|u1 u2 | v) for all u1 , u2 U . Show that each vh is a pseudometric on U , and that
{vh : h (Uc )+ , v U + } denes a linear space topology on U .
(o) Let (A,
) be a -nite measure algebra. Show that the function (, u) 7 [[u > ]] : 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, 0 then {(, u) :
(a [[u > ]]) > } is open.)
(p) Let G be the regular open algebra of R. Show that there is no Hausdor topology T on L0 (G) such
that hun inN is T-convergent to u whenever hun inN is order*-convergent to u. (Hint: Let H be any T-open
set containing 0. Enumerate Q as hqn inN . Find inductively a non-decreasing sequence hGn inN in G such
that Gn H, qn Gn for every n. Conclude that R H.)
(q) 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.
(r) Let A be a Dedekind -complete Boolean algebra and r 1. Let E R r 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 = L0 (A) the corresponding map (364Yd). Show that if h
is continuous if L0 is given its
and h
0 r
topology of convergence in measure and (L ) the product topology.
376
Function spaces
367Ys
(s) Show that 367U is true for all measure algebras, whether semi-nite or not.
367 Notes and comments I have given a very general denition 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 signicance may not be immediately obvious. In particular, the denition is very much better adapted
to distributive lattices than to others (367Yc, 367Yd). 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 denition here from other denitions 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 simplications
(367Be-f), but the Martingale Theorem (367K), for instance, if we want to express it in terms 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
fullling the conditions of 367Bf; but beyond this no standard denitions 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
spaces have such topologies (367Yp). It is not known whether a topology corresponding to convergence in
measure can be dened 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-nite 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 (367Xo, 367Ym).
For -nite measure algebras, the topology of convergence in measure is easily described in terms of
order-convergence (367Q). For other measure algebras, the formula fails (367Xt). 367Yp 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 Lebesgues
Dominated Convergence Theorem (367J) and the Martingale Theorem (367K) to show how these can be
expressed in the language of this chapter.
368C
377
368A Lemma Let A be a Dedekind -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 = supuA u nv for every n N.
proof The hypothesis A is countable or A is Dedekind complete ensures that c = supuA [[u > ]] is
dened for each . By 364Ma, c = inf nN cn = inf R c is non-zero. Now for any n 1, R
[[supuA (u nc) > ]] = supuA [[u > ]] [[c >
because if 0 k N then
n ]]
= [[c >
n ]],
n ]],
n ]].
So supuA u nc = nc for every n 1, and we can take v = c. (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
nw = supvA 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 k1 k1 u = 0, so v0 = supk1 (v0 k1 u)+ . Because
V is order-dense in U , v0 = sup B where
1
k
(352Fa)
1
k
378
Function spaces
368D
368D Corollary Suppose that A is a Dedekind -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 u U such that 0 < u |u|, so that 0 < u |Tu|. 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 disjoint set (in the sense of 352C); to obtain such a set apply Zorns
lemma to the family of all disjoint 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 justication: for u U and
R, [[T u > ]] will be the band in U generated by
{e (u e)+ : 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 > ]] is a member of A determined by u and .)
(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}.
either e 6= e so
or e = e and
v e (u e )+ e e = 0,
v e (u e )+ (e u)+ (u e)+ = 0.
Accordingly [[T u > ]] is included in the band {v} and v [[T u > ]] . Q
Q
(c) Now I must conrm that the formula given for [[T u > ]] is consistent with the conditions laid down
in 364A. P
P Take u U .
(i) If then
so e (u e)+ [[T u > ]] for every e C and [[T u > ]] [[T u > ]].
(ii) Given R, set W = sup> [[T u > ]] in A, that is, the band in U generated by {e (u e)+ :
e C, > }. Then for each e C,
sup> e (u e)+ = e (u inf > e)+ = e (u e)+
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 e = inf > e). So e (u e)+ W ; as e is arbitrary, [[T u > ]] W
and [[T u > ]] = W .
(iii) Now set W = inf nN [[T u > n]]. For any e C, n N we have
so that
for every n 1 and
e (e n1 u+ )+ e (e n1 u)+ W
368E
379
e = supn1 e (e n1 u+ )+ W .
Thus C W and W C = {0}. So we have inf nN [[T u > n]] = 0.
(iv) Finally, set W = supnN [[T u > n]]. Then
e (e n1 u )+ e (e + n1 u)+ e (u + ne)+ W
e = supn1 e (e n1 u )+ W
for every e C and W = {0}, W = U . Thus all three conditions of 364A are satised. Q
Q
(d) Thus we have a well-dened 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
i
n ]]
= 1,
i
n ]]
= 0.
So setting
Vi = [[T u >
i
n ]] \ [[T u
i+1
n ]]
>
= [[T u >
i
n ]]
[[T u >
i+1
n ]] ,
hVi iiZ is a partition of unity in A. Similarly, hWi iiZ is a partition of unity, where
i
n ]]
Wi = [[T v >
i+1
n ]] .
[[T v >
[[T v >
j
n ]]
[[T u + T v >
[[T u + T v >
k
n ]]
supi+jk Vi Wj .
i
n ]]
i+j
n ]]
[[T u + T v >
k
n ]];
thus
k+1
n
[[T u + T v >
k+1
n ]]
i
n ]]
[[T v >
i
n
q<
ki
n ]]
i+1
n ,
so that
supi+jk Vi Wj ;
k
n ]].
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
(u
k+1
n ]]
i+1 +
e)
n
i+j+2
n ]]
+ (v
j+1 +
e)
n
supi+jk1 Vi Wj = 0,
k+1
n ]]
supi+jk 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 e C such that w2 = w1 e (v nj e )+ > 0; of course e = e , and
380
368E
Function spaces
j
n
i
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) > ]]
k
n ]];
so we get
k
n ]].
k+1
n ]]
k
n ]],
[[T u + T v >
k+1
n ]]
k
n ]]
]]
as is arbitrary, T u = T (u); as and u are arbitrary, T is linear. (We need only check linearity for > 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 R and consider the band
[[T u > ]] [[T u > ]] = [[|T u| > ]] (by 364Mb). This is the band generated by {e (u e)+ : e
C} {e (u e)+ : e C}. But this must also be the band generated by
{(e (u e)+ ) (e (u e)+ ) : e C} = {e (|u| e)+ : e C},
which is [[T |u| > ]]. Thus [[|T u| > ]] = [[T |u| > ]] for every 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 > 0 such that |u| e 6 e, so that e (|u| e)+ 6= 0 and [[T |u| > ]] 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 > 0
such that V = [[p > ]] 6= 0. Take u > 0 in V . Let e C be such that u e > 0. Then v = u e > 0. Now
e (v e)+ = 0; but also e v = 0 for every e C distinct from e, so that [[T v > ]] = {0}. Also v V ,
so e (v e )+ V whenever e C and 0, and [[T v > ]] V for every 0. Accordingly we have
[[T v > ]] = {0} [[p > ]] if ,
368K
381
382
368K
Function spaces
368K This is a convenient point at which to give a characterization of the Riesz spaces L0 (A).
Lemma Let A be a Dedekind -complete Boolean algebra. Suppose that A L0 (A)+ is disjoint. 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 = supuA [[u > n]]; this
is always dened; set a = inf nN an . Now a = 0. P
P?? Otherwise, there must be a u A such that
a = a [[u > 0]] 6= 0, since a a0 . But now, for any n, and any v A \ {u},
a [[v > n]] [[u > 0]] [[v > 0]] = 0,
so that a [[u > n]]. As n is arbitrary, inf nN [[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 disjoint.
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 -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.
) If u U + and a A then u a U . P
(
P Set A = {v : v U, 0 v a}, and let C A be
a maximal disjoint set; then w = sup C is dened in U , and is also the supremum in L0 . Set b = [[w > 0]].
As w a, b a. ?? If b 6= a, then (a \ b) > 0, and there is a v U such that 0 < v (a \ b); but
now v A and v w = 0, so v v = 0 for every v C, and we ought to have added v to C. X
X Thus
[[w > 0]] = a.
Now consider u = supnN u nw; as U is Dedekind -complete, u U . Since [[u > 0]] a, u u a.
On the other hand,
u [[w >
1
n ]]
[[u n]] u n2 w u
as required. Q
Q
C = {a : a A, 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 a. 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 12 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 (), ud d U
for every d D, so u = supdD ud d U . Now u w, but also [[u 21 w 0]] d for every d D, so is
equal to 1, and u 12 w, as required. Q
Q
( ) Given w 0 in L0 , we can therefore choose hun inN , hvn inN inductively such that v0 = 0 and
un U ,
1
(w
2
vn ) u n w v n ,
vn+1 = vn + un
for every n N. Now hvn inN is a non-decreasing sequence in U and w vn 2n w for every n, so
w = supnN vn U .
368P
383
does not have inmum 0. Let u > 0 be a lower bound for A, and set An = {u v : v An } for each n N.
Then each An is a non-empty downwards-directed set with inmum 0. Let hun inN be a sequence such that
un An 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 wB,nN w vn = 0, because U is Archimedean (353F), while B A, so u w for
every w B. If u is any upper bound for {un : n N}, then
u u u u vn = (u vn )+ (w vn )+ = w vn
for every n N, w B. So u u. Thus u = supnN un . As hun inN is arbitrary, u and hAn inN witness
that (ii) is true.
368P Proposition (a) A regularly embedded Riesz subspace of an Archimedean weakly (, )-distributive Riesz space is weakly (, )-distributive.
(b) An Archimedean Riesz space with a weakly (, )-distributive order-dense Riesz subspace is weakly
(, )-distributive.
(c) If U is a Riesz space such that U separates the points of U , then U is weakly (, )-distributive; in
particular, U and U are weakly (, )-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 (, )-distributive. Then 368O tells us that there are a v > 0 in V and
a sequence hAn inN of non-empty downwards-directed subsets of V , all with inmum 0 in V , such that
supnN vn = v in V whenever vn An for every n N. Because V is regularly embedded in U , inf An = 0
384
368P
Function spaces
nN
An , so U is not weakly (, )-
(b) Let U be an Archimedean Riesz space which is not weakly (, )-distributive, and V an orderdense Riesz subspace of U . By 368O again, there are a u > 0 in U and a sequence hAn inN of non-empty
downwards-directed sets in U , all with inmum 0, such that supnN 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}
then (because V is order-dense) any upper bound for C in U is also an upper bound of {(v u)+ : u An }.
But
supuAn (v u)+ = (v inf An )+ = v,
and v = supnN vn . Thus hBn inN witnesses that V is not weakly (, )-distributive.
(c) Now suppose that U separates the points of U . In this case U is surely Archimedean (356G). ?? If U
is not weakly (, )-distributive, there are a u > 0 in U and a sequence hAn inN of non-empty downwardsdirected sets, all with inmum 0, such that supnN 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 = f (u) > 0. For each
n N, there is a un An such that f (un ) 2n2 . But in this case hsupi<n ui inN is a non-decreasing
sequence with supremum u, so
P
f (u) = limn f (supin ui ) i=0 f (ui ) 12 < f (u),
which is absurd. X
X Thus U is weakly (, )-distributive.
For any Riesz space U , U acts on U as a subspace of U (356F); as U surely separates the points of
U , so does U . So U is weakly (, )-distributive. Now U is a band in U (356B), so is regularly
embedded, and must also be weakly (, )-distributive, by (a) above.
368Q Theorem (a) For any Boolean algebra A, A is weakly (, )-distributive i S(A) is weakly (, )distributive i L (A) is weakly (, )-distributive.
(b) For a Dedekind -complete Boolean algebra A, L0 (A) is weakly (, )-distributive i A is weakly
(, )-distributive.
proof (a)(i) ?? Suppose, if possible, that A is weakly (, )-distributive but S = S(A) is not. By 368O, as
usual, we have a u > 0 in S and a sequence hAn inN of non-empty downwards-directed sets in S, all with
inmum 0, such that u = supnN un whenever un An for every n. Let > 0 be such that c = [[u > ]] 6= 0
(361Eg), and consider
Bn = {[[v > ]] : 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 , b v for every v An ). Because A is weakly (, )-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 (316H). Take vn An
such that bn = [[vn > ]], so that
vn 1 kvn k bn 1 kuk a.
c = [[u > ]] a,
368Y
385
b}.
(b) In the same way, because S can be regarded as an order-dense Riesz subspace of L0 = L0 (A) (364K),
L0 is weakly (, )-distributive i S is, that is, i A is.
368R Corollary An Archimedean Riesz space is weakly (, )-distributive i its band algebra is weakly
(, )-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 (, )-distributive i L0 is; and by 368Qb
L0 is weakly (, )-distributive i A is.
368S Corollary If (A,
) is a semi-nite measure algebra, any regularly embedded Riesz subspace (in
particular, any solid linear subspace and any order-dense Riesz subspace) of L0 (A) is weakly (, )-distributive.
proof By 322F, A is weakly (, )-distributive; by 368Qb, L0 (A) is weakly (, )-distributive; by 368Pa,
any regularly embedded Riesz subspace is weakly (, )-distributive.
368X Basic exercises (a) Let X be an uncountable set and the countable-cocountable -algebra of
subsets of X. Show that there is a family A L0 = L0 () 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= supuA u nw.
b its Dedekind completion (314U). Show that L (A)
b can be
(b) Let A be any Boolean algebra, and A
386
Function spaces
368Y
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 (RO(X)). (iii) Let U be an Archimedean Riesz space and C U +
a maximal disjoint 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 VQcan be embedded
Q
as an order-dense Riesz subspace of eC Ue and therefore in eC L0 (RO(Xe ))
= L0 ( eC RO(Xe )) for
a suitable family of regular open algebras RO(Xe ). (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 dened by the formulae
(A1 , B1 ) + (A2 , B2 ) = (A, B) i A1 + A2 A, B1 + B2 B,
(A, B) = (A, B) if > 0,
(A1 , B1 ) (A2 , B2 ) i A1 A2 .
Show that u 7 (], u] , [u, [) denes an embedding of U as an order-dense Riesz subspace of V, so that
V may be identied with the Dedekind completion of U .
(c) Work through the proof of 364U when X is compact, Hausdor 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 -complete Boolean algebra A (ii) U is Dedekind -complete and has a weak
order unit and whenever A U + is countable and disjoint then A is bounded above in U .
(e) Let U be a weakly (, )-distributive Riesz space and V a Riesz subspace of U which is either solid
or order-dense. Show that V is weakly (, )-distributive.
(f ) Show that C([0, 1]) is not weakly (, )-distributive. (Compare 316J.)
(g) Let A be a ccc weakly (, )-distributive Boolean algebra. Suppose we have a double sequence
haij i(i,j)NN in A such that haij ijN order*-converges to ai in A for each i, while hai iiN order*-converges
to a. Show that there is a strictly increasing sequence hn(i)iiN such that hai,n(i) iiN order*-converges to a.
(h) Let U be a weakly (, )-distributive Riesz space with the countable sup property. Suppose we have
an order-bounded double sequence huij i(i,j)NN in U such that huij ijN order*-converges to ui in U for
each i, while hui iiN order*-converges to u. Show that there is a strictly increasing sequence hn(i)iiN such
that hui,n(i) iiN order*-converges to u.
(i) Let A be a ccc weakly (, )-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 (, )-distributive Riesz space and f : U R a positive linear functional; write f
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 f (u) = supvA f (v). (See 356Xe, 362D.) (ii) Show that if f is strictly positive, so
is f . (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
369A
387
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 Hausdor spaces, which are usefully easier than the general case (364V, 368Yc).
368G is a version of Ogasawaras 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 (, )-distributivity is one of the more striking correspondences. (Compare, for
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 (, )-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 dened in 367, are related to weak (, )-distributivity
(compare 367Yi, 367Yp); 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 -nite measure algebra (367N).
388
369A
Function spaces
Let > 0 be such that d = [[v > ]] 6= 0. Then b 1 v V whenever b Ad . Set b = h(b) [0, [ for
b Ad . Because the map b 7 b : A L0 is additive and order-continuous, the map b P
7 b : Ad V also
is, and = h must be additive and order-continuous; in particular, (supnN bn ) = n=0 bn whenever
hbn inN is a disjoint sequence in Ad . Moreover, if b Ad is non-zero, then 0 < b v , so b = h(b) > 0.
Thus (Ad , ) is a totally nite 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 c : Ac [0, [ be
P a functional
such that (Ac , c ) is a totally nite measure algebra. Dene
: A [0, ]Pby setting
a = cC c (a c)
for every a A. Then (A,
) is a localizable measure algebra. P
P (i)
0 = cC 0 = 0. (ii) If han inN is a
disjoint sequence in A with supremum a, then
P
P
P
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,
).
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 365M and 366Dc.
Theorem Let (A,
) be a semi-nite measure algebra, and U L0 = L0 (A) an order-dense Riesz subspace.
Set
V = {v : v L0 , v u L1 for every u U },
u v for all u U , v V .
R
(ii) By the denition of V , (T v)(u) is dened 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 inmum 0 in U , then
inf A = 0 in L0 , because U is order-dense (352Nb). Consequently inf uA u v = 0 in L0 and in L1 (364P),
and
(because
uv =0
369C
389
(v) If v V and u0 0 in U , set a = [[v > 0]]. Then v + = va. SetR A = {u : u U, 0 u u0 a}.
Because U is order-dense in L0 , u0 a = sup A in L0 . Because and are order-continuous,
Z
+
(T v) (u0 ) sup (T v)(u) = sup v u
uA
uA
Z
Z
= v u0 a = 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 dened in R for every v 0 in U1 ; so we have an extension h
1
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 {c : c I}, where I = {c : c A, c 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
v = a = h(d
a)
R
) = h(u)
v u supu A h(u
= h(u)
R
is nite, so v u = supuA v u belongs to L1 (365Df) and v u h(u). As u is arbitrary, v V and
R
u Qh = supgB u Qg = supgB u T 1 g L1
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 .
390
369D
Function spaces
369D Corollary Let U be any Riesz space such that U separates the points of U . Then there is a
localizable measure algebra (A,
) such that the pair (U, U ) can be represented by a pair (V, W ) of orderdense Riesz subspaces of L0 = L0 (A) such that W = {w : w L0 , v w L1 for every v V }, writing L1
for L1 (A,
). 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 nds (A,
) and an order-dense V which is
isomorphic to U , and 369C identies W with V and 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 i V = V .
369E Kakutanis theorem (Kakutani 41) If U is any L-space, there is a localizable measure algebra
(A,
) such that U is isomorphic, as Banach lattice, to L1 =L1 (A,
).
R
R
proof U is a perfect Riesz space, and U = U has an order unit dened by saying that u = kuk for
u 0 (356P). By 369D, we can nd a localizable measure algebra (A,
) and an identication 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 : vw L1 for
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 k for the norm on V corresponding to the norm of U , we have
kuk =
|u| for u U ,
kvk =
|v| e =
|v e| for v V .
Thus the Riesz space isomorphism between U and L1 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 -complete Boolean algebra. An extended Fatou norm on L0 = L0 (A)
is a function : L0 [0, ] such that
(i) (u + v) (u) + (v) for all u, v L0 ;
(ii) (u) = || (u) for all u L0 , R (counting 0 as 0, as usual);
(iii) (u) (v) whenever |u| |v| in L0 ;
(iv) supuA (u) = (v) whenever A (L0 )+ is a non-empty upwards-directed set with supremum v in
0
L ;
(v) (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 (v) < .
369G Proposition Let A be a Dedekind -complete Boolean algebra and an extended Fatou norm
on L0 = L0 (A). Then L = {u : u L0 , (u) < } is an order-dense solid linear subspace of L0 , and
, restricted to L , is a Fatou norm under which L is a Banach lattice. If hun inN is a non-decreasing
norm-bounded sequence in (L )+ , then it has a supremum in L ; if A is Dedekind complete, then L has
the Levi property.
proof (a) By (i), (ii) and (iii) of 369F, L is a solid linear subspace of L0 ; by (vi), it is order-dense.
Hypotheses (i), (ii), (iii) and (v) show that is a Riesz norm on L , while (iv) shows that it is a Fatou
norm.
(b)(i) Suppose that hun inN is a non-decreasing norm-bounded sequence in (L )+ . Then u = supnN un
is dened in L0 . P
P?? Otherwise, there is a v > 0 in L0 such that kv = supnN kv un for every k N (368A).
By (v)-(vi) of 369F, there is a v such that 0 < v v and 0 < (v ) < . Now kv = supnN kv un for
every k, so
369I
391
(c) Now suppose that A is Dedekind complete and A (L )+ is a non-empty upwards-directed normbounded set in L . By the argument of (b-i) above, using the other half of 368A, sup A is dened in L0 and
belongs to L . As A is arbitrary, L has the Levi property.
369H Associate norms: Definition Let (A,
) be a semi-nite measure algebra, and an extended
Fatou norm on L0 = L0 (A). Dene : L0 [0, ] by setting
(u) = sup{ku vk1 : v L0 , (v) 1}
for every u L0 ; then is the associate of . (The word suggests a symmetric relationship; it is justied
by the next theorem.)
369I Theorem Let (A,
) be a semi-nite measure algebra, and an extended Fatou norm on L0 =
L (A). Then
(i) its associate is also an extended Fatou norm on L0 ;
(ii) is the associate of ;
(iii) ku vk1 (u) (v) for all u, v L0 .
0
proof (a) Before embarking on the proof that is an extended Fatou seminorm on L0 , I give the greater
part of the argument needed to show that = , where
for every u L0 .
) Set
(
B = {u : u L1 , (u) 1},
for m N. The sequence hvm imN is non-decreasing, (vm ) (um ) 1 for every m, and
P
P
k|u| vm k1 supnm k|u| vmn k1 i=m k|u| |ui |k1 i=m ku ui k1 0
as m . So |u| = supmN vm in L0 ,
and u B. Q
Q
) Now take any u0 L0 such that (u0 ) > 1. Then, writing Af for {a :
(
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,
) is semi-nite,
so that S(Af ) is order-dense in L0 ), and supuA (u) = (u0 ) > 1. Take u1 A such that (u1 ) > 1, that
is, u1
/ B. By the Hahn-Banach theorem (3A5Cc), there is a continuous linear functional f : L1 R such
392
Function spaces
369I
that f (u1 ) > 1 but f (u) 1 for every u B. Because (L1 ) = (L1 ) (356Dc), |f | is dened in (L1 ) , and
of course
|f |(u1 ) f (u1 ) > 1,
1
for every
R a A. Then is a completely additive real-valued functional on A, so there is a w L such that
a = a w for every a A (365Ea). Because a 0 for every a, w 0. Now
for every a A, so
w = |f |(a c)
w u = |f |(u c) |f |(u) 1
Av = {u : u S(A)+ B, u |v|}
w u0
w u 1.
w u1 = |f |(u1 ) > 1,
(u) 1 = (u) 1.
(c) Now I return to the proof that is an extended Fatou norm. It is easy to check that it satises
conditions (i)-(iv) of 369F; in eect, 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 < (u) < ; set = 1/ (u). Then
(2u) > 1, so that (2u) > 1 and there is a w L0 such that (w) 1, k2u wk1 > 1. But now set
v1 = v |w|; then
v v1 u |w| > 0,
(d) Accordingly is also an extended Fatou norm. Now in (a) I showed that
(u) 1 = (u) 1.
It follows easily that (u) (u) for every u (since otherwise there would be some such that
(u) = (u) < 1 < (u) = (u)).
(e) Of course we have ku vk1 1 whenever (u) 1 and (v) 1. It follows easily that ku vk1
(u) (v) whenever u, v L0 and both (u), (v) are non-zero. But if one of them is zero, then u v = 0,
because both and satisfy (v) of 369F, so the result is trivial.
369M
393
(b) P
If v
/ L then for every n N there is a un such that (un ) 1 and kun vk1 2n . Set
n
wn = i=0 2i |ui | for each n. Then hwn inN Ris a non-decreasing sequence and (wn ) 2 for each n, so
w = supnN wn is dened in L , by 369G; now w |v| n + 1 for every n, so w v
/ L1 .
sup
(u)1,kwk 1
u v w = sup
(u)1
u v,
Z
= sup{ |u| |v| : kukp 1}
for every v L0 . Q
Q
369M Proposition Let (A,
) be a semi-nite measure algebra and an extended Fatou norm on
L0 = L0 (A). Then
(a) the embedding L L0 is continuous for the norm topology of L and the topology of convergence
in measure on L0 ;
(b) : L0 [0, ] is lower semi-continuous, that is, all the balls {u : (u) } are closed for the
topology of convergence in measure;
(c) if hun inN is a sequence in L0 which is order*-convergent to u L0 (denition: 367A), then (u)
lim inf n (un ).
proof (a) This is a special case of 367P.
(b) Set B = {u : (u) }. If u L0 \ B , then
A = {|u| a : a Af }
394
Function spaces
369M
for each k, and v = supkN vk . Then (vk ) (vk ) for each k, and hvk ikN is non-decreasing, so
(v ) . But
a [[|u| v > 2k ]] a supik [[|u vi | > 2k ]]
P i
has measure at most
for each k, so a [[|u| v > 0]] must be 0, that is, |u| a v and
i=k 2
(|u| a) ; contrary to the choice of a. X
X Thus u cannot belong to the closure of B . As u is arbitrary,
B is closed.
(c) If hun inN order*-converges to u, it converges in measure (367Na). If > lim inf n (un ), there is
a subsequence of hun inN in B , and (u) , by (b). As is arbitary, (u) lim inf n (un ).
369N I now turn to another special case which we have already had occasion to consider in other
contexts.
Definition Let (A,
) be a measure algebra. Set
M,1 = M ,1 (A,
) = L1 (A,
) L (A),
) = L1 (A,
) + L (A),
M1, = M 1, (A,
and
kuk,1 = max(kuk1 , kuk )
for u L0 (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 a where u M p,q and
a < , while q indicates the nature of |u| 1 for u M p,q . Thus
1,
M
is the space of u such that u a L1 for every a Af and |u| 1 L ; in M 1,0 we demand
further that |u| 1 M 0 (366F); while in M ,1 we ask that |u| 1 L1 and that u a L for every
a Af .
369O Proposition Let (A,
) be a semi-nite measure algebra.
(a) k k,1 is an extended Fatou norm on L0 = L0 (A).
(b) Its associate k k1, may be dened by the formulae
kuk1, = min{kvk1 + kwk : v L1 , w L , v + w = u}
Z
= min{ + (|u| 1)+ : 0}
Z
min(1,
[[|u| > ]])d
=
0
(d) Writing Af = {a :
a < }, S(Af ) is norm-dense in M ,1 and S(A) is norm-dense in M 1, .
(e) For any p [1, ],
0
for every u L .
369O
395
Remark By writing min rather than inf in the formulae of part (b) I mean to assert that the inma 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:
1 (u) = sup{ku vk1 : kvk,1 1},
2 (u) = inf{ku k1 + ku k : u = u + u },
3 (u) = inf 0 ( +
4 (u) =
R
0
(|u| 1)+ ),
min(1,
[[|u| > ]])d.
(I write inf here to avoid the question of attainment for the moment.) Now we have the following.
(i) 1 (u) 2 (u). P
P If kvk,1 1 and u = u + u , then
Taking the supremum over v and the innum over u and u , 1 (u) 2 (u). Q
Q
u = w (|u| 0 1)+ ,
u = w (|u| 0 1).
ku k1 =
min(1,
[[|u| > ]])d,
ku k 0 =
so
R 0
0
2 (u) ku k1 + ku k 4 (u). Q
Q
(iii) 4 (u) 3 (u). P
P For any 0,
Z
Z
min(1,
[[|u| > ]])d
min(1,
[[|u| > ]])d +
4 (u) =
0
Z
u max(1,
a)1 (u),
396
369O
Function spaces
1
2
(322Eb).
so
0 +
(u 0 1) 0 + ( 0 )
[[u > 0 ]] +
Z
+ (u 1)+ .
(u 1)+
while
[[u > + ]] > 1, so
(u 0 1)+ + + +
0 +
R
Thus 0 + (u 0 1)+ = 3 (u).
(u 0 1)+ +
(u 1)+ ;
(u 1)+ .
u = ku vk1 1 (u).
( ) If 0 > 0, set =
[[u > 0 ]]. Take any [0, 0 [. Then
([[u > ]] \ [[u > 0 ]]) > 1 , so there
is a b [[u > ]] \ [[u > 0 ]] such that 1 <
b < . Set v = [[u > 0 ]] + 1
b b. Then kvk,1 = 1 so
1 (u)
uv
(u 0 1)+ + 0 +
= 3 (u) (1 )(0 ).
As for M ,1 , if u 0 in M ,1 and r N, set vr = supkN 2r k[[u > 2r k]]; then each vr belongs
to S f =R S(Af ),R ku vr k 2r , and hvr irN 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 )+ .
369R
397
(e)(i) If p = 1 or p = this is immediate from the denition 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
(369L)
(b) Suppose that A M 1,0 is non-empty and downwards-directed and has inmum 0. Let > 0. Set
B = {(u 1)+ : u A}. Then B L1 (by 366Gc); B is non-empty and downwards-directed and
has inmum 0. Because k k1 is order-continuous (365C), inf vB kvk1 = 0 and there is a u A such that
k(u 1)+ k1 , so that kuk1, 2. As is arbitrary, inf uA 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,
) be a localizable measure algebra. Set M 1, = M 1, (A,
), etc.
1,
1,0
(a) (M
) and (M ) can both be identied with M ,1 .
(b) (M ,1 ) can be identied with M 1, ; M 1, and M ,1 are perfect Riesz spaces.
proof Everything is covered by 369O and 369K except the identication 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 },
Also V M ,1 . P
P Because L1 M 1,0 and k k is the associate of k k1 , V L
R . ?? If there is a
1
v V \ L , then (because (A,
) isR semi-nite, so that |v| = supaAf |v| a) supaAf a |v|
R = . For each
n N choose an Af such that an |v| 4n , and set u = supnN 2n an M 1,0 ; then u |v| 2n for
each n, so again v
/ V. X
X Thus V L1 and V M ,1 . Q
Q
,1
So M
= V can be identied 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.
398
369R
Function spaces
L = L1 L2 .
(u + u ) (u) + (u ),
(u) = || (u),
(u) (u ) if |u| |u |.
(For the last, remember that in this case u = u z where kzk 1.)
(b) Take any non-empty, upwards-directed set A (L0 )+ , with supremum u0 . Suppose that =
supuA (u) < . For u A and n N set
Then
so
1 (v) + 2 (u v)+ 1 (v1 ) + (1 )1 (v2 ) + 2 (u v1 )+ + (1 )2 (u v2 )+
+ 2n );
wk = inf ik (u vi )+
we have
1 (vk ) + 2 (wk ) 1 (vk ) + 2 (u vk )+ + 2n
for each k, and hvk ikN , hwk ikN are non-decreasing. So setting v = supkN vvk , w = supkN (uv)+ wk ,
we get
1 (v ) + 2 (w ) + 2n .
1 (v a) + 2 ((u v)+ a) + 2n ,
369Xd
399
1 (v) + 2 (u v)+
for every u A; since {(u v)+ : u A} is an upwards-directed set with supremum (u0 v)+ , and 2 is an
extended Fatou norm,
1 (v) + 2 (u0 v)+ .
(c) This shows both that the inmum in the denition of (u) is always attained (since this is trivial if
(u) = , and otherwise we consider A = {|u|}), and also that (sup A) = supuA (u) whenever A (L0 )+
is a non-empty upwards-directed set with a supremum. Thus satises conditions (i)-(iv) of 369F. Condition
(vi) there is trivial, since (for instance) (v) 1 (v) for every v. As for 369F(v), suppose that u > 0 in L0 .
Take u1 such that 0 < u1 u and 1 (u1 ) 1, u2 such that 0 < u2 u1 and 2 (u2 ) 1. In this case, if
u2 = v + w, we must have
1 (v) + 2 (w) kv u1 k1 + kw u2 k1 ku2 u2 k1 ;
so that
(u) ku2 u2 k1 > 0.
Thus all the conditions of 369F are satised, and is an extended Fatou norm on L0 .
(d) The calculation of is now very easy. Since surely we have i for both i, we must have i
for both i. On the other hand, if u, z L0 , then there are v, w such that u = v +w and (u) = 1 (v)+2 (w),
so that
ku zk1 kv zk1 + kw zk1 1 (v)1 (z) + 2 (w)2 (z) (u) max(1 (z), 2 (z));
while the fact that the inmum in the denition of is always attained means that L L1 + L2 , so that
we have equality here also.
369X Basic exercises > (a) Let A be a Dedekind -complete Boolean algebra. Show that the following
are equiveridical: (i) there is a function
such that (A,
) is a semi-nite measure algebra; (ii) (L )
separates the points of L = L (A); (iii) for every non-zero a A there is a completely additive functional
: A R such that a 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 : R ], ] is convex if (x + (1 )y) (x) + (1 )(y)
for all x, y R and [0, 1], interpreting 0 as 0, as usual. For any convex function : R ], ]
which is not always innite, set (y) = supxR xy (x) for every y R. (i) Show that : R ], ]
is convex and lower semi-continuous and not always innite. (Hint: 233Xh.) (ii) Show that if is lower
semi-continuous then = . (Hint: It is easy to check that . For the reverse inequality, set
I = {x : (x) < }, and consider x int I, x I \ int I and x
/ I separately; 233Ha is useful for the rst.)
> (c) For the purposes of this exercise and the next, say that a Youngs function is a non-negative
non-constant lower semi-continuous convex function : [0, [ [0, ] such that (0) = 0 and (x) is nite
for some x > 0. (Warning! the phrase Youngs function has other meanings.) (i) Show that in this case
is non-decreasing and continuous on the left and , dened by saying that (y) = supx0 xy (x) for
every y 0, is again a Youngs function. (ii) Show that = . Say that and are complementary.
(iii) Compute in the cases () (x) = x () (x) = max(0, x 1) () (x) = x2 () (x) = xp where
1 < p < .
> (d) Let , = be complementary Youngs functions in the sense of 369Xc, and (A,
) a semi-nite
measure algebra. Set
400
369Xd
Function spaces
B = {u : u L0 ,
(|u|)
1},
C = {v : v L0 ,
(|v|)
1}.
(For nite-valued , : (L0 )+ L0 is given by 364I. Devise an appropriate convention for the Rcase in which
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 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 2 , where
is the extended Fatou norm corresponding to and is the associate of , so that and can be
interpreted as equivalent norms on the same Banach space.
(U and V are complementary Orlicz spaces; I will call , 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,
) with an extended Fatou norm on
L0 (A) such that U can be identied, as normed Riesz space, with an order-dense Riesz subspace of L . (ii)
Hence, or otherwise, show that kuk = supf U , kf k1 |f (u)| for every u U . (iii) Show that if U is Dedekind
complete and has the Levi property, then U becomes identied with L itself, and in particular is a Banach
lattice (cf. 354Xn).
(f ) Let (A,
) be a semi-nite measure algebra, and an extended Fatou norm on L0 (A). Show that
the norm of L is order-continuous i the norm topology of L agrees with the topology of convergence in
measure on any order-bounded subset of L .
(g) Let (A,
) be a -nite measure algebra of countable Maharam type, and an extended Fatou norm
on L0 (A) such that the norm of L is order-continuous. Show that L is separable in its norm topology.
R
a
(h) Let (A,
) be an atomless semi-nite measure algebra. Show that kuk1, = max{ a |u| : a A,
1} for every u L0 (A). (Hint: take a [[|u| > 0 ]] in part (b-iv) of the proof of 369O.)
(i) Let (A,
) be any semi-nite measure algebra. Show that if is any Orlicz norm on L0 = L0 (A), then
there is a > 0 such that kuk1, (u) 2 kuk,1 for every u L0 , so that M,1 L M1, .
(j) Let (A,
) be a semi-nite measure algebra. Show that the subspaces M1, , M,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,
) be a measure algebra and U a Banach space. (i) Suppose that : A U is an additive
function such that kak min(1,
a) for every a A. Show that there is a unique bounded linear operator
T : M1, U such that T (a) = a for every a A. (ii) Suppose that : Af U is an additive function
such that kak max(1,
a) for every a Af . Show that there is a unique bounded linear operator
,1
T : M U such that T (a) = a for every a Af .
(l) Let (A,
) and (B, ) be semi-nite measure algebras, and : Af Bf a measure-preserving ring
homomorphism, as in 366H, with associated maps T : M0 M0 and P : M1,0 M1,0 . Show that
kT uk,1 = kuk,1 for every u M,1 , kP vk,1 kvk,1 for every v M,1 .
(m) Let (A,
) and (B, ) be measure algebras, and : A B a measure-preserving Boolean homomorphism. (i) Show that there is a unique Riesz homomorphism T : M1, M1, such that T (a) = (a)
for every a A and kT uk1, = kuk1, for every u M1, . (ii) Now suppose that (A,
) is localizable
and is order-continuous. Show that there is a unique positive linear operator P : M1, M1, such
R
R
that a P v = a v for every a Af and v M1, , and that kP vk kvk for every v L (B),
kP vk1, kvk1, for every v M1, . (Compare 365P.)
(n) Let (A,
) and (B, ) be semi-nite measure algebras, and : [0, [ [0, ] a Youngs function;
write for the corresponding Orlicz norm on either L0 (A) or L0 (B). Let : A B be a measurepreserving Boolean homomorphism, with associated map T : M1, M1, , as in 369Xm. (i) Show that
(T u) = (u) for every u M1, . (ii) Show that if (A,
) is localizable, is order-continuous and
P : M1, M1, is the map of 369Xm(ii), then (P v) (v) for every v M1, . (Hint: 365R.)
369Yi
401
402
Function spaces
369 Notes
369 Notes and comments The representation theorems 369A-369D give a 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,
), in such a way that it is its own second dual, where the duality here is between
subspaces of L0 , taking U = {v : uv L1 for every u U }. (I see that in this expression I ought somewhere
to mention that both U and U are assumed to be order-dense in L0 .) Indeed I believe that the original
perfect spaces were the vollkommene Raume of G.Kothe, which were subspaces of R N , corresponding to
the measure algebra PN with counting measure, so that U or U was {v : u v 1 for every u U }.
I have presented Kakutanis 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; Kakutanis 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 nd 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 Maharams 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.
Among the extended Fatou norms of 369F the Orlicz norms (369Xd, 369Yc-369Ye) form a signicant
subfamily. Because they are dened 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 (369R) seems to
go quite deep. In view of our ordinary expectation that properties of these normed function spaces should
be reected 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 R N such that their linear sum U + V
is not perfect.
371B
403
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 Birkhos 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
dened on L0 spaces (375). In the nal section of the chapter (376), I look at operators which can be
dened in terms of kernels on product spaces.
Pm
proof
(a) Suppose
that v, v B. Then
we have u0 , . . . , um , u0 , . . . , un U + such that
i=0 ui =
Pn
P
P
m
n
u, v = i=0 |T ui | and v = j=0 |T uj |. Now there are vij 0 in U , for i m and j n, such
j=0 uj =P
Pm
Pm Pn
n
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
P
P
m
n
v = 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 ,
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 rst 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 + ,
404
371B
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 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 2n . Set Sn = |Tn+1 Tn | L (U ; V ) for each n. Then
for each n, so S =
n=0
371G
405
(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,
), (B, ) be measure algebras. Recall that M 1,0 (A,
) is the space of those u
(0)
1
L (A,
) + L (A) such that
[[|u| > ]] < for every > 0 (366F-366G, 369P). Write T (0) = T, for
1,0
1,0
1
the set of all linear operators T : M (A,
) M (B, ) such that T u L (B, ) and kT uk1 kuk1 for
every u L1 (A,
), T u L (B) and kT uk kuk for every u L (A) M 1,0 (A,
).
371G Proposition Let (A,
) and (B, ) be measure algebras.
(0)
(a) T (0) = T, is a convex set in the unit ball of B(M 1,0 (A,
); M 1,0 (B, )). If T0 : L1 (A,
) L1 (B, ) is
a linear operator of norm at most 1, and T0 u L (B) and kT0 uk kuk for every u L1 (A,
)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,
); M 1,0 (B, )) = L (M 1,0 (A,
); M 1,0 (B, )),
,
proof I write M1,0 , Lp for M1,0 , Lp (B, ), etc.
(a)(i) If T T (0) and u M1,0 then there are v L1 , w L
(0)
(ii) Because the unit balls of B(L1 ; L1 ) and B(L
.
; L
) are convex, so is T
(iii) Now suppose that T0 : L1 L1 is a linear operator of norm at most 1 such that kT0 uk kuk
1
for every u L1 L
is
. By the argument of (i), T0 is a bounded operator for the k k1, norms; since L
1,0
1,0
1,0
dense in M (369Pc), T0 has a unique extension to a bounded linear operator T : M M . Of course
kT uk1 = kT0 uk1 kuk1 for every u L1 .
1,0
Now suppose that u L
; set = kuk . Let > 0, and set
M
v = (u+ 1)+ (u 1)+ ;
. Accordingly
kT u T vk1, ku vk1, ,
+
So if we set w = (|T u T v| 1)
L1 ,
kT vk = kT0 vk kvk .
kwk1 ; while
406
371G
|T u| |T v| + w + 1 ( + )1 + w,
so
k(|T u| ( + )1)+ k1 kwk1 .
,
||T0 |u| |T0 ||u| = sup|u ||u| |T0 u | kuk 1,
so k|T0 |uk kuk . By (a), there is a unique S T (0) extending |T0 |. Now Su+ 0 for every u L1 , so
Su+ 0 for every u M1,0 (since the function u 7 (Su+ )+ Su+ : M1,0 M1,0 is continuous and zero
on the dense set L1 ), that is, S is a positive operator; also S|u| |T u| for every u L1 , so Sv S|u| |T u|
whenever u, v M1,0 and |u| v. This means that T : M1,0 M1,0 is order-bounded. Because M1,0 is
Dedekind complete (366Ga), |T | is dened in L (M1,0 ; M1,0 ).
If v 0 in L1 , then
|T |v = sup|u|v T u = sup|u|v T0 u = |T0 |v = Sv.
Thus |T | agrees with S on L1 . Because M1,0 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 M1,0 , 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 modication 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 nd families hq iqQ , hq iqQ of real numbers such that |t|p = supqQ q +q (tq)
for every t R (233Hb). Then |u|p = supqQ q 1 + q (u q1) 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 L1 , so we may speak of T (|w|p ); while w M1,0 (366Ga), so we may
speak of T w.
For any q Q, 0p q qq , that is, qq q 0, while q w|w|p (qq q )1 and k(q w|w|p )+ k
qq q . Now this means that
T (q w |w|p ) T (q w |w|p )+ kT (q w |w|p )+ k 1
k(q w |w|p )+ k 1 (qq q )1.
1,0
whenever u L1 , v L
.
kST vk kT uk kuk
371 Notes
407
(0)
371H Remark In the context of 366H, T M1,0 T, , while P T, . 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 (denition: 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,
) be a semi-nite measure algebra and set U = L1 (A,
). 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 i A is
nite. (Hint: consider operators u 7 u a, 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 ). Dene
T L (V ; U ) by writing T (h) = hT for h V . (i) Show that |T | |T | in L (V ; U ). (ii) Show
+
+
>(f ) Using 371D, but nothing about uniformly integrable sets beyond the denition (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 (denition: 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,
) be a measure algebra, U a Banach space, and T : L1 (A,
) U a bounded linear operator.
1
T (a)
: a A, 0 <
a < } is relatively compact in U .
An A n
1
An such that every coecient of An has modulus 2n/2 . (Hint: An+1 =
.) (iii) Show
2 An An
that there is a linear isometry S : 2 2 such that |(Sei |ej )| = 2n/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 (), V = L1 (); 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 terms of properties of their adjoints T : V U , as in 371Xe and 371Yd. When U is
408
371 Notes
an L-space, then U is a Dedekind complete M -space, and it is easy to see that this forces T to be orderbounded, for any Banach lattice V (371Xd). But since no important L-space is reexive, 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.
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
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 ultralter on N. Because U is reexive, v = limnF un is dened 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
u T i u) =
(T n+1 u u)
T u n un =
i=0 (T
n+1
n+1
2
n+1 kuk.
where all the limits are taken for the weak topology. Q
Q
But this means that v V , while
u 0.
[[u + T u + . . . + T m u
> 0]].
372D
409
T (v + ) =
v+
T (v + ) kv + k1 kT v + k1 0
u L L (A). Set An = n+1 i=0 T i for each n N. Then for any u L1 , u = supnN An u is dened
in L0 (A), and
[[u > ]] kuk1 for every > 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 (a) 1 for every n
and every a of nite measure.
For m N and > 0, set
am = [[supim Ai u > ]].
Then
am kuk1 . P
P Set a = am , w = u a. Of course supim Ai u belongs to L1 , so
a is nite
1
and w L . For any i m,
Ai w = Ai u Ai (a) Ai u 1,
a =
a =
b
u
b
w
b
u kuk1 ,
as claimed. Q
Q
It follows that if we set c = supnN an ,
c 1 kuk1 for every > 0 and inf >0 c = 0. But
this is exactly the criterion in 364Mb for u = supnN An u to be dened in L0 . And [[u > ]] = c , so
Pn
1
i
so k|T |wk kwk . Thus |T | also satises 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 = supnN Bn |u| is dened in L and
[[v > ]] k|u|k1 = kuk1 for every > 0. Consequently u = supnN An u is dened in L0 and u v, so
that
[[u > ]] kuk1 for every > 0.
372D We are now ready for a very general form of the Ergodic Theorem. I express it in terms 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,
) be a measure algebra, and set M 1,0 = M 1,0 (A,
), T (0) =
P
(0)
n
1
i
1,0
T, 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 uinN is order*-convergent (denition: 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 uinN is order-bounded in L0 = L0 (A) for any u M 1,0 ; this is because if
410
372D
u = v + w, where v L1 = L1 (A,
) and w L = L (A), then hAn vinN and hAn (v)inN are bounded
above, by 372C, while hAn winN is norm- and order-bounded in L . Accordingly I can uninhibitedly speak
of P (u) = inf nN supin Ai u and P (u) = supnN inf in Ai u for any u M 1,0 , these both being dened in
L0 .
(b) Write V1 for the set of those u M 1,0 such that hAn uinN 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 > 0 there is a u V1 such that ku0 uk1, 2 . Write
P u = P (u) = P (u) for the order*-limit of hAn uinN . Express u0 u as v + w where v L1 , w L and
kvk1 + kwk 22 .
Set v = supnN Bn |v|. Then
[[v > ]] 2, by 372C. Next, if w = supnN Bn |w|, we surely have
2
w 2 1. Now
P u v w P (u0 ) P (u0 ) P u + v + w ,
[[2(v + w ) > 2 + 42 ]]
[[v > ]] +
[[w > 22 ]] =
[[v > ]] 2
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,
) V1 V 2 . 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 L2 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 > 0, there is a u L2 L such
that kuu k2 (take u = (u1)(1) for any large enough), and now k(uT u)(u T u )k2 2.
Thus W = {u + v T u : u L2 L , v L2 , T v = v} is dense in L2 . But W 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 uinN is normconvergent and order*-convergent for every u M 1,0 . By 367D, the limits are the same; by 367F, hAn uinN
is order*-convergent when regarded as a sequence in M 1,0 . Write P u for the common value of the limits.
372G
411
for every u L
1,0
1
kT n+1 u uk1,
n+1
= 0. Q
Q
(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 > ]] = [[u > ]] (365Oc), so T u = u i
[[u > ]] C for every R, that is, i u belongs to the canonical image of L1 (C,
C) in L1 (365R). To
1
identify P u further, observe that if u L , a C then
Tu = a u
R
R
R
R
R
R i
(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 uinN for k k1 ). But this is enough to dene P u as the conditional expectation
of u on C (365R).
a
Tu =
412
372H
372H The Ergodic Theorem is most often expressed in terms 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, , ) be a measure space with measure algebra (A,
). For h L0 = L0 () write h for
0
0
the corresponding member of L = L (A) (364Jc). Now let : X X be an inverse-measure-preserving
function, : A A the corresponding sequentially order-continuous measure-preserving homomorphism
dened by setting E = (1 [E]) for E (324M), and T : L0 L0 the Riesz homomorphism dened
by setting T (a) = (a) for a A (364R). Then T h = (h) for any h L0 .
is dened for almost every x X, and g(x) = g(x) for almost every x.
so setting An = n+1 i=0 T , An u = gn , where gn (x) = n+1 i=0 f (i (x)) whenever this is dened. Now
we know from 372E or 372F that hAn uinN is order*-convergent to some v such that T v = v, so hgn inN
must be convergent almost everywhere (367G), and taking g = limn gn where this is dened, g = v.
Accordingly (g) = T v = v = g and g =a.e. g, as claimed.
372J The following straightforward facts will be useful in the next corollary and elsewhere.
Lemma Let (X, , ) be a measure space with measure algebra (A,
). Let : X X be an inversemeasure-preserving function and : A A the associated homomorphism, as in 372H. Set C = {c : c
A, c = c}, T = {E : E , 1 [E]E is negligible} and T0 = {E : E , 1 [E] = E}. Then T and T0
are -subalgebras of ; T0 T, T = {E : E , E C}, and C = {E : E T0 }.
proof It is easy to see that T and T0 are -subalgebras of 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 such that F = c.
i
i
is dened for almost every x X; g =a.e. g, and g is a conditional expectation of f on the -algebra
T = {E : E , 1 [E]E is negligible}. If either f is -measurable and dened everywhere in X
or [E] is negligible for every negligible set E, then g is a conditional expectation of f on the -algebra
T0 = {E : E , 1 [E] = E}.
proof (a) We know by 372I that g is dened almost everywhere and that g =a.e. g. 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, a = a} = {F : F T} = {F : F T0 },
372L
413
nN,j1
j [Fn ]
nN,j0
j [Fn ].
(c) Finally, suppose that [E] is negligible for every negligible set E. Then every -negligible
S set isn T0
negligible. P
P If E is -negligible, then [E], 2 [E] = [[E]], . . . are allSnegligible,
so
E
=
nN [E] is
T
negligible, and there is a measurable negligible set F E . Now F = mN nm n [F ] is a negligible
set in T0 including E, so E is T0 -negligible. Q
Q Consequently g = h 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
denition 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
: [0, 1] [0, 1] a Borel measurable function such that [C] = [0, 1] and (x) = x for x [0, 1] \ C.
(For instance, we could take agreeing with the Cantor function on C (134H).) Because C is negligible,
is inverse-measure-preserving for Lebesgue measure , and if f is any Lebesgue integrable function then
1 Pn
i
g(x) = limn
i=0 f ( (x)) is dened and equal to f (x) for every x dom f \ C. But for x C we
n+1
can, by manipulating , arrange for g(x) to be almost anything; and if f is undened on C then g will also
be undened on C. On the other hand, C is not 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( T0 ).
If instead of Lebesgue measure itself we took its restriction B to the algebra of Borel subsets of [0, 1], then
would still be inverse-measure-preserving for B , but we should now have to worry about the possibility
that f C was non-measurable, so that gC came out to be non-measurable, even if everywhere dened, and
g was not B T0 -virtually measurable.
414
372L
In the statement of 372K I have oered two ways of being sure that the problem does not arise: check
that [E] is negligible whenever E is negligible (so that all negligible sets are T0 -negligible), or check
that f is dened everywhere and -measurable. Even if these conditions are not immediately satised in a
particular application, it may be possible to modify the problem so that they are. For instance, completing
the measure will leave inverse-measure-preserving (235Hc), will not change the integrable functions but will
make them all measurable (212F, 212Bc), and may enlarge T0 enough to make a dierence. If our function
f is measurable (because the measure is complete, or otherwise) we can extend it to a measurable function
dened everywhere (121I) and the corresponding extension of g will be T0 -integrable. Alternatively, if
the diculty seems to lie in the behaviour of rather than in the behaviour of f (as in the example above),
it may help to modify 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 (x) = < x1 >, the fractional part of x1 , and k1 (x) = x1 (x), the
integer part of x1 ; then (x) X for each x X, so we may dene kn (x) = k1 (n1 (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 ((x)) for every n 1. Now dene hpn (x)inN , hqn (x)inN inductively by setting
p0 (x) = 0,
q0 (x) = 1,
p1 (x) = 1,
q1 (x) = k1 (x),
r1 = ,
k2 = 15,
r2 =
15
,
106
k3 = 1,
r3 =
16
;
113
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
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 terms 7 , 39 above, and involve larger numbers.
(c) Now we need a variety of miscellaneous facts about these coecients, which I list here.
(i) For any x X, n 1 we have
pn1 (x)qn (x) pn (x)qn1 (x) = (1)n ,
n (x) =
(induce on n), so
x=
pn (x)xqn (x)
xqn1 (x)pn1 (x)
372N
415
(ii) Another easy induction on n shows that for any nite 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 n 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 ] 1 [D(m2 ,... ,mn ) ].)
(iii) We also need to know that the intervals Dm of (ii) are small; specically, that if m = (m1 , . . . , mn ),
the length of Dm is at most 2n+1 . P
P All the coecients pi , qi , for i n, take constant values pi , qi on
Dm , since they are determined from the coecients ki which are constant on Dm by denition. Now every
pn qn1
pn1 qn
= 1 ,
qn qn1
qi
qn qn1
qi1
+ qi2
372N Theorem Set X = [0, 1]\Q, and dene : X X as in 372M. Then for every Lebesgue integrable
function f on X,
limn
for almost every x X.
1 Pn
f (i (x))
i=0
n+1
1
ln 2
R 1 f (t)
0 1+t
dt
proof (a) The integral just written, and the phrase almost every, refer of course to Lebesgue measure;
but the rst step is to introduce another measure, so I had better give a name L to
measure on
R Lebesgue
1
L (dx) whenever
X. Let be the indenite-integral measure on X dened by saying that E = ln12 E 1+x
1
> 0 for every x X,
this is dened. The coecient ln12 is of course chosen to make X = 1. Because 1+x
dom = dom L and has just the same negligible sets as L (234D); I can therefore safely use the terms
measurable set, almost everywhere and negligible without declaring which measure I have in mind each
time.
(b) Now is inverse-measure-preserving
when regarded as a function from (X, ) to itself. P
P For each
h
h
1
1
1
k 1, set Ik = k+1 , k . On X Ik , (x) = x k. Observe that Ik : X Ik X is bijective and
dierentiable relative to its domain in the sense of 262Fb. Consider, for any measurable E X,
Z
Z
1
1
L (dy) =
| (x)|L (dx)
(y+k)(y+k+1)
((x)+k)((x)+k+1)
1
I [E]
E
Zk
x2 1
=
L (dx) = ln 2 (Ik 1 [E]),
2
Ik 1 [E]
x+1 x
E =
1
ln 2
Z
X
1
L (dy)
(y+k)(y+k+1)
k=1 E
1
y+k
k=1
1
y+k+1
1
y+1
As E is arbitrary, is inverse-measure-preserving. Q
Q
(c) The next thing we need to know is that if E X and 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
416
372N
x=
pn +pn1 n (x)
+q
n
qn
n1 (x)
for every x Dm , where pn , etc., are dened from m as in 372M(c-iii). Recall also that n is strictly
monotonic on Dm . So for any interval I [0, 1] (open, closed or half-open) with endpoints < ,
),
n [I] Dm will be of the form X J, where J is an interval with endpoints (pn + pn1 )/(qn + qn1
n
(pn +pn1 )/(qn +qn1 ) in some order. This means that we can estimate L ( [I]Dm )/L Dm , because
it is the modulus of
p
n +pn1
+q
qn
n1
p
n +pn1
+q
qn
n1
p
n +pn1
+q
qn
n1
p
n
qn
()qn
(qn
+qn1
)
+q
+q
(qn
)(q
n1
n1 )
n
()qn
+q
qn
n1
1
2
( ).
Now look at
1
2
1
2
1
L (E
2
large that () < 21 whenever 0 < 2n+1 , and set mi = ki (x) for i n, so that x Dm . Taking the
least such that Dm [x , x + ], we must have 2n+1 , because the length of Dm is at most 2n+1
(372M(c-iii)), while L Dm , because Dm is an interval. Accordingly
L (E Dm ) L (E [x , x + ]) = 2() < L Dm .
1
1+x
1 Pn
f (i (x))
n+1 i=0
is dened for almost every x X, and is a conditional expectation of f (with respect to the measure ) on
the -algebra T0 = {E : E is measurable, 1 [E] = E}. But we haveRjust seenR that T0 consists only of
negligible and conegligible sets, so g must be essentially constant; since g d = f d, we must have
1
n n+1
lim
n
X
f (i (x)) =
i=0
f d).
f d =
1
ln 2
f (t)
L (dt)
1+t
0
372Q
417
1
(2 ln(k
ln 2
+ 1) ln k ln(k + 2))
1
, k1 ]). Then (for i 1) f (i (x)) = 1 if ki (x) = k and zero otherwise. So
proof In 372N, set f = (X [ k+1
lim
n n
= lim
n
X
1
n n+1
f (i (x)) = lim
i=1
f (t)
dt
1+t
0
1
ln 2
f (i (x))
i=0
1/k
1
dt
1+t
1/k+1
1
ln 2
1
1
1
(ln(1 + ) ln(1 +
))
ln 2
k
k+1
n
X
1
(2 ln(k
ln 2
+ 1) ln k ln(k + 2)),
418
372Q
0=
(a \ a) = limn
( n a \ a) =
a
(1 \ a),
so one of
a,
(1 \ a) must be zero, and a {0, 1}. Thus is ergodic.
)(
) T u = u i [[u > ]] = [[u > ]] for every ; if is ergodic, this means that [[u > ]] {0, 1}
(ii)(
for every , and u must be of the form 1, where = inf{ : [[u > ]] = 0}.
Pn
1
i
)( ) If () is true and u L1 , then we know from 372G that h n+1
(
i=0 T uinN is order*convergent and k k1 -convergent to some v such that T v = v; by (), v is of the form 1; and
R
R
1 Pn R i
= v = limn
T u = u.
i=0
n+1
) Assuming (), take any a A such that a = a, and consider u = a. Then T i u = a for
( )(
every i, so
a = limn
and a must be either 0 or 1.
1 Pn
T iu
i=0
n+1
= ( u)1 =
a 1,
(b)(i) Simply translating the denitions, we see that is mixing i is. In this case is ergodic, as in
(a-i).
)(
) If a = a there is an E such that 1 [E] = E and E = a, by 372J; now
(ii)(
a = E {0, 1},
so a {0, 1}.
proof (a) Let T : L0 = L0 (A) L0 be the Riesz homomorphism associated with . Take any a, b A
and any non-principal ultralter F on N. Then hT n (a)inN is a bounded sequence in the reexive space
L2 = L2 (A,
), so v = limnF T n (a) is dened for the weak topology of L2 . Now for each n N set
n
Bn = [A]. This is a closed subalgebra of A (314F(a-i)), and contains i a for every i n. So if we identify
L2 (Bn ,
Bn ) with the corresponding subspace of L2 (366I), it contains T i (a) for every i n; but also
it is norm-closed, therefore weakly closed (3A5Ee), so contains
v. This means that [[v > ]] must belong to
T
Bn for every and every n. But in this case [[v > ]] nN Bn = {0, 1} for every , and v is of the form
1. Also
=
So
v = limnF
limnF
( n a b) = limnF
T n (a) =
a.
T n (a) b =
v b =
b =
a
b.
372Xi
419
k [Ek ] =
mk
nm
n [Fn ] =
for every
Q
T k, so E T. Q
So nN An = {0, 1} and and are mixing.
mN
nm
n [Fn ] = E
372X Basic exercisesP(a) Let U be any reexive 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 dened (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 dened} 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,
) be a measure algebra, and T T, ; set An = n+1
i=0 T for n N. Take any
p
p
p [1, [ and u L = L (A,
). Show that hAn uinN is order*-convergent and k kp -convergent to some
v Lp . (Hint: put 372Xa together with 372D.)
(c) Let (A,
) be a probability algebra, and : A A a measure-preserving BooleanR homomorphism.
R
Let P : L1 L1 be the operator dened as in 365P/366Hb, where L1 = L1 , so that a P u = a u for
P
n
1
i
1
1
1
u L1 , a A. Set An = n+1
i=0 P : L L for each i. Show that for any u L , hAn uinN is
order*-convergent and k k1 -convergent to the conditional expectation of u on the subalgebra {a : a = 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
> (g) Use the Ergodic Theorem to prove Kolmogorovs Strong Law of Large Numbers (273I), as follows.
Given a complete probability space (, , ) and an independent identically distributed sequence hFn inN
of measurable functions from to R, set X = R N and F () = hFn ()inN for . Show that if we give
each copy of R the distribution of F0 then F is inverse-measure-preserving for and the product measure
on X. Now use 372Xf.
(h) Show that the continued fraction coecients of
are 1, 2, 1, 2, . . . .
> (i) For x X = [0, 1] \ Q let k1 (x), k2 (x), . . . be its continued-fraction coecients. Show that x 7
hkn+1 (x) 1inN is a bijection between X and N N which is a homeomorphism if X is given its usual
topology (as a subset of R) and N N is given its usual product topology (each copy of N being given the
discrete topology).
420
372Xj
(j) For any irrational x [0, 1] let k1 (x), k2 (x), . . . be its continued-fraction coecients 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 R, the set {(p, q) : p Z, q 1, | pq | /q 2 } is
innite for every > 0.
(k) Let (A,
) be an atomless probability algebra. Show that the following are equiveridical: (i) A is
homogeneous; (ii) there is an ergodic measure-preserving Boolean homomorphism : A A; (iii) there is
a mixing measure-preserving automorphism : A A. (Hint: 333P.)
(l) Let (A,
) be a probability algebra, and : A A a measure-preserving Boolean homomorphism. (i)
Show that if n 1 then is mixing i n is mixing. (ii) Show that if n 1 and n is ergodic then is
ergodic. (iii) Show that if is an automorphism then it is ergodic, or mixing, i 1 is.
> (m) Consider the tent map (x) = min(x, 1 x) for x [0, 1], [0, 2]. Show that 2 is inverse(x) = 2 (<2n x>) for
measure-preserving and mixing for Lebesgue measure on [0, 1]. (Hint: show that n+1
2
n
n
n
n 1, and hence that (I 2 [J]) = I J whenever I is of the form [2 k, 2 (k + 1)] and J is an
interval.)
(n) Consider the quadratic map (x) = x(1 x) for x [0, 1], [0, 4]. Show that 4 p
is inversemeasure-preserving and mixing for the Radon measure on [0, 1] with density function t 7 1/ t(1 t).
(Hint: show that the transformation t 7 sin2 t
2 turns it into the tent map.) Show that for almost every x,
limn
1
#({i
n+1
: i n, 4i (x) }) =
arcsin
fm = f mX.)
(t) For P
irrational x [0, 1], write k1 (x), k2 (x), . . . for the continued-fraction coecients of
R x. Show that
n
limn n1 i=1 ki (x) = for almost every x. (Hint: take , as in 372N, and show that k1 d = .)
372Yg
421
(u) Let (X, , ) be any probability space, and let be the product measure on X N . Dene : X N X N
by setting (x)(n) = x(n+1). Show that is inverse-measure-preserving. Show that satises the conditions
of 372R, so is mixing.
(v) Let (X, , ) be any probability space, and let be the product measure on X Z . Dene : X Z X Z
by setting (x)(n) = x(n + 1). Show that is inverse-measure-preserving. Show that is mixing. (Hint:
show that if C, C are basic cylinder sets then (C n [C ]) = C C for all n large enough.) Show that
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
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 is mixing for the measure .
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)
[[u > ]]
[[|u|>/2]]
|u|.
p
[[|u| > ]] +
[[|u| > ]]d; see 365A. Use 366Xa to show that
ku kpp 2p
R
0
p2
/2
[[|u|>]]
|u| =
and reverse the order of integration. Compare 275Yc.) (This is Wieners dominated ergodic theorem.)
(0)
1
itself. Set u = supnN n+1
i=0 |T u|. Show that u L . (Hint: use the method of 372Yb to show that
R
R
1
n
ln qn (x) =
1
ln 2
R1
ln t
0 1+t
dt =
2
.
12 ln 2
(Hint: 225Xi, 282Xo.) (v) For almost every x X, limn n1 ln |x rn (x)| = 2 /6 ln 2. (vi) For almost
every x X, 11n |x rn (x)| 10n and 3n qn (x) 4n for all but nitely many n.
(g) In 372M, show that for any measurable setR E X, limn RL n [E] = E. (Hint: recall that is
mixing for (372Xw). Hence show that limn n [E] g d = E g d for any integrable g. Apply this
to a Radon-Nikod
ym derivative of L with respect to .) (I understand that this result is due to Gauss.)
422
372Yh
(h) Let h(Xi , i , i )iiI be any family of probability spaces, with product (X, , ). Suppose that for each
i I we are given an inverse-measure-preserving function i : Xi Xi . Show that there is a corresponding
inverse-measure-preserving function : X X given by setting (x)(i) = i (x(i)) for x X, i I. Show
that if each i is mixing so is .
(i) Give an example of an ergodic measure-preserving automorphism : [0, 1[ [0, 1[ such that 2 is not
ergodic. (Hint: set (x) = 21 (1 + 0 (2x)) for x < 12 , x 21 for x 12 . See also 388Xe.)
(j) Show that there is an ergodic : [0, 1] [0, 1] such that (1 , 2 ) 7 ((1 ), (2 )) : [0, 1]2 [0, 1]2 is
not ergodic. (Hint: 372Xo.)
(k) Let M be an r r matrix with integer coecients and non-zero determinant, where r 1. Let
r
r
r
: [0, 1[ [0, 1[ be the function such that (x) M x Zr for every x [0, 1[ . Show that is inverser
measure-preserving for Lebesgue measure on [0, 1[ .
(l) (i) Let (X, , ) and (Y, T, ) be probability spaces, with c.l.d. product (X Y, , ). Suppose that
: X X is a mixing inverse-measure-preserving function and : Y Y is an ergodic inverse-measurepreserving function. Dene : X Y X Y by setting (x, y) = ((x), (y)) for all x, y. Show that
is an ergodic inverse-measure-preserving function. (ii) Let (A,
) and (B, ) be probability algebras, with
Suppose that : A A is a mixing measure-preserving Boolean
probability algebra free product (C, ).
homomorphism and : B B is an ergodic measure-preserving Boolean homomorphism. Let : C C
be the measure-preserving Boolean homomorphism such that (a b) = a b for all a A and b B
(325Xd). Show that is ergodic.
(m) Let (X, , ) be a measure space with measure algebra (A,
). Let be a family of inverse-measurepreserving functions from X to itself, and for let : A A be the associated homomorphism. Set
C = {c : c A, c = c for every }, T = {E : E , 1 [E]E is negligible for every } and
T0 = {E : E , 1 [E] = E for every }. Show that (i) T and T0 are -subalgebras of (ii) T0 T
(iii) T = {E : E , E C} (iv) if is countable and = for all , , then C = {E : E T0 }.
(n) (i) Let (A,
) be a probability algebra, : A A a mixing measure-preserving Boolean homomorphism, and T = T : L1 L1 the corresponding linear operator (365O). Show that if u L1 is such that
{T n u : n N} is relatively compact for the norm topology, then u = 1 for some . (ii) Let be Lebesgue
measure on [0, 1[, (A,
) its measure algebra, [0, 1[ an irrational number, (x) = x +1 for x [0, 1[ (as
in 372Xo), and T : L1 () L1 () the linear operator dened by setting T g = (g) for g L1 (). Show
that {T n : n Z} is relatively compact for the strong operator topology on B(L1 (); L1 ()).
(o) Let U be a Banach space, (A,
) a semi-nite measure algebra and hTn inN a sequence of continuous
linear operators from U to L0 = L0 (A) with its topology of convergence in measure. Suppose that supnN Tn u
is dened in L0 for every u U . Show that {u : u U , hTn uinN is order*-convergent in L0 } is a norm-closed
linear subspace of U .
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 oer 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 Kolmogorovs 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
372 Notes
423
(275D). Consequently we have similar results, obtained by similar methods, concerning the domination of
sequences starting from members of Lp (372Yb, 275Yc), though the inequalities are not identical. (Compare also 372Yc with 275Yd.) There are some tantalising reections of these traits in results surrounding
Carlesons theorem on the pointwise convergence of square-integrable Fourier series (see 286 notes), but
Carlesons theorem seems to be much harder than the others. Other forms of the strong law (273D, 273H)
do not appear to t 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 nd myself writing at some length about a technical problem. The root of the diculty is
in the denition of conditional expectation. Now it is generally accepted that any pure mathematician has
Humpty Dumptys 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 terms in ways which will not mystify or mislead the unprepared reader.
Having written down a denition of conditional expectation in Volume 2, I must either stick to it, or go
back and change it, or very carefully explain exactly what modication I wish to make here. I dont 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
which cannot be formulated eectively 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
oers 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 terms, ergodic and mixing transformation, not because I wish to
use them for any new ideas (apart from the elementary 372R, these must wait for 386-387) but because it
may help if I immediately classify some of the inverse-measure-preserving functions we have seen (372Xm372Xo, 372Xr, 372Xu, 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 , 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 = 2, = 4 that the methods of this volume can tell us
anything interesting.
424
373 intro.
every u L (A,
) and kT uk kuk for every u L (A). (Compare the denition of T (0) in 371F.)
(b) If B is Dedekind complete, so that M 1, (A,
), being a solid linear subspace of the Dedekind complete
); M 1, (A,
)).
space L0 (B), is Dedekind complete, T, will be T, L (M 1, (A,
373B Proposition Let (A,
) and (B, ) be measure algebras.
(a) T = T, is a convex set in the unit ball of B(M 1, (A,
); M 1, (B, )).
(0)
(b) If T T then T M 1,0 (A,
) belongs to T, , as dened in 371F. So if T T , p [1, [ and
p
p
u L (A,
) then T u L (B, ) and kT ukp kukp .
(c) If B is Dedekind complete and T T , then T L (M 1, (A,
); M 1, (B, )) and T1 T whenever
1,
1,
T1 L (M
(A,
); M
(B, )) and |T1 | |T |; in particular, |T | T .
(d) If : A B is a measure-preserving Boolean homomorphism, then we have a corresponding operator
T T dened by saying that T (a) = (a) for every a A. If is order-continuous, then so is T .
is another measure algebra and T T , S T then ST T .
(e) If (C, )
,
proof (a) As 371G, parts (a-i) and (a-ii) of the proof.
(b) If u M1,0 and > 0, then u is expressible as u + u where ku k and u L1 . (Set
u = (u+ 1) (u 1).)
So
(|T u| 1)+ |T u T u | L1 .
As is arbitrary, T u M1,0 ; as u is arbitrary, T M1,0 T (0) . Now the rest is a consequence of 371Gd.
(c) Because M1, is a solid linear subspace of L0 (B), which is Dedekind complete because B is,
L (M1, ; M1, ) is a Riesz space (355Ea).
Take any u 0 in M1, . Let 0 be such that (u 1)+ L1 . Because T L1 belongs to
B(L1 ; L1 ) = L (L1 ; L1 ) (371D), w0 = sup{T v : v L1 , 0 v (u 1)+ } is dened in L1 . Now if
v M1, and 0 v u, we must have
T v = T (v 1)+ + T (v 1) w0 + 1 M1, .
373C
Decreasing rearrangements
425
|v||u|
(d) By 365O and 363F, we have norm-preserving positive linear operators T1 : L1 L1 and T :
L (A) L (B) dened by saying that T1 (a) = (a) whenever
a < and T (a) = (a) for every
a A. If u S(Af ) = L1 S(A) (365F), then T1 u = T u, because both T1 and T are linear and they
agree on {a :
a < }. If u 0 in M,1 = L1 L (A), there is a non-decreasing sequence hun inN in
S(Af ) such that u = supnN un and
limn ku un k1 = limn ku un k = 0
T1 u = supnN T1 un = supnN T un = T u.
Accordingly T1 and T agree on L1 L (A). But this means that if u M1, is expressed as v+w = v +w ,
where v, v L1 and w, w L (A), we shall have
T1 v + T w = T1 v + T w + T1 (v v) T (w w ) = T1 v + T w,
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 dened, observe that T L1 and T L (A) are uniquely dened by the values T
takes on S(Af ), S(A) respectively, because these spaces are dense for the appropriate norms.
Now suppose that is order-continuous. Then T1 and T are also order-continuous (365Oa, 363Ff). If
A M1, is non-empty and downwards-directed and has inmum 0, take u0 A and > 0 such that
(u0 1)+ L1 . Set
L1
A1 = {(u 1)+ : u A, u u0 },
A = {u 1 : u A}.
and A L (A) are both downwards-directed and have inmum 0, so inf T1 [A1 ] =
Then 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 1)+ + T (u 1) where u, u A; because A is downwardsdirected, there is a v A such that v u u , in which case T v w. Accordingly T [A] must also have
inmum 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 the
class T . Let (A,
) be a measure algebra. Write M0, = M 0, (A,
) for the set of those u L0 (A) such
that
[[|u| > ]] is nite for some R. (See 369N for the ideology of this notation.) It is easy to see that
M 0, (A,
) is a solid linear subspace of L0 (A). Let (AL ,
L ) be the measure algebra of Lebesgue measure
on [0, [. For u M 0, (A,
) its decreasing rearrangement is u M 0, (AL ,
L ), dened by setting
u = g , where
g(t) = inf{ : 0,
[[|u| > ]] t}
426
373D
that is,
L [[u > ]] =
[[|u| > ]] for every 0.
(c)(i) If u = P
ai [ .
i=0 i ai , where a0 a1 . . . an and i 0 for each i, then u =
i=0 i [0,
n
i=0 i ai where
Pn (ii) If u =
P a0 , . . . , an are disjoint and |0 | |1 | . . . |n |, then u =
ai for i n + 1.
i=0 |i | [i , i+1 [ , where i =
j<i
(d) If E ]0, [ is any Borel set, and u M 0 (A,
), then
L [[u E]] =
[[|u| E]].
for the
(e) Let h : [0, [ [0, [ be a non-decreasing function such that h(0) = 0, and write h
0
+
0
+
corresponding functions on L (A) and L (AL ) (364I). Then (h(u)) = h(u ) whenever u 0 in M 0 (A,
).
) whenever u 0 in M 0, (A,
If h is continuous on the left, (h(u))
= h(u
).
(f) If u M 0, (A,
) and 0, then
(u 1)+ = ((|u| 1)+ ) .
(g) If u M 0, (A,
), then for any t > 0
Rt
0
u0
u = inf 0 t +
(|u| 1)+ .
(h) If A (M 0, (A,
))+ is non-empty and upwards-directed and has supremum u0 M 0, (A,
), then
= supuA u .
g(t) = inf{ :
[[|u| > ]] t}
g(t) >
[[|u| > ]] > t for some >
[[|u| > ]] > t
(b) This is obvious, either from the denition in 373C or from (a) just above.
Pn
n
X
i=0
= [0,
aj [ if
j1
X
i=0
i <
j
X
i=0
= [0,
a0 [ if 0 < 0 ,
373E
427
Decreasing rearrangements
by the Monotone Class Theorem (136B), A includes the -algebra of subsets of ], [ generated by I; but
this must contain E ], [ for every Borel set E R.
Accordingly, for any Borel set E ]0, [,
(e) For any > 0, E = {t : h(t) > } is a Borel subset of ]0, [. If u M0 then, using (d) above,
) > ]] =
L [[h(u
L [[u E ]] =
[[u E ]] =
[[h(u)
> ]] =
L [[(h(u))
> ]].
Rt
0
g = tg(t) +
R
0
Rt
0
Now
Z
u = min0 t +
(u 1) =
=
(h)
u = min0 t +
(u 1)+ .
Rt
(|u| 1)+
(|u| 1)+ .
[[u0 > ]] =
(supuA [[u > ]]) = supuA
[[u > ]]
for any > 0, using 364Mb and 321D. So
[[u0 > ]] = [0,
[[u0 > ]][ = supuA [0,
[[u > ]][ = [[supuA u > ]]
u = i=0 i [0,
ai [ , v = j=0 j [0,
bj [
373E Theorem Let (A,
) be a measure algebra. Then
|u v|
(373Dc). So
uv =
=
n
m X
X
i=0 j=0
n
m X
X
i=0 j=0
i j
(ai bj )
n
m X
X
i j min(
ai ,
bj )
i=0 j=0
i j L ([0,
ai [ [0,
bj [) =
u v .
428
373E
(b) For the general case, we have non-decreasing sequences hun inN , hvn inN in S(A)+ with suprema |u|,
|v| respectively (364Kd), so that
|u v| = |u| |v| = supnN |u| vn = supm,nN um vn = supnN un vn
and
using 373Db.
|u v| =
supnN un vn = supnN
un vn supnN
un vn
u v ,
1,0
(iii) u M (A,
) i u M (AL ,
L );
(iv) u M ,1 (A,
) i u M ,1 (AL ,
L ), and in this case kuk,1 = ku k,1 .
proof (a)(i) Consider rst the case p = 1. In this case
|u| =
R
0
0
p
|u|p =
(|u|p ) =
u .
(u )p = ku kpp
(b)(i)
u M0
[[|u| > ]] < for every > 0
L [[u > ]] < for every > 0 u M0L .
(|u| 1)+ =
(u 1)+
as in the proof of 373Dg. So kuk1, = ku k1, if either is nite, by the formula in 369Ob.
(iii) This follows from (i) and (ii), because M 1,0 = M 0 M 1, .
or u M 1,0 (A,
) and T T, ,
Rt
proof Set T1 = T L1 , so that kT1 k 1 in B(L1 ; L1 ), and |T1 | is dened in B(L1 ; L1 ), also with norm
at most 1. If 0, then we can express u as u1 + u2 where |u1 | (|u| 1)R+ and |u2 | 1. (Let
w L (A) be such that kwk 1, u = |u| w; set u2 = w (|u| 1).) So if (|u| 1)+ < ,
|T u| |T u1 | + |T u2 | |T1 ||u1 | + 1
373H
and
429
Decreasing rearrangements
a < }.
(a) The following are equiveridical:
1
a t0 ; so that
[[uc > ]] t0 for every c A , and e() = supcAf [[uc > ]] is dened in Af . Of course
+
0
0
e() = [[ue(1) > ]] for every 1, so inf R e() = 0, and v1 = supcAf u+
c is dened in L = L (A)
(364Mb). Because [[v1 > ]] = e() Af for each > 0, v1 M0 . For any a Af ,
+
v1 a = supcAf u+
c a = ua ,
R
R
f
so v1 M1,0 and a v1 = a u+
a for every a A .
R
R
f
u=
a
ua = a
R
+
a (365Ea), so
(ii)
R Assume (). If > 0, there is a > 0 such that a (|u| 1) R whenever
a) whenever
a . As is arbitrary, limt0 supat | a u| = 0. Moreover, whenever
that | a u| (1 +
R
R
t > 0 and
a t, 1t | a u| + 1t (|u| 1)+ . Thus
lim supt
1
t
supat |
u| .
R
(i) Assume (). Again, for each c Af , we have uc L1 such that c a = a uc for every a Af ;
again, set e() = supcAf [[u+
c > ]], which is dened because A is supposed to be Dedekind complete. This
time, there are t0 , 0 such that |a|
a whenever a Af and
a t0 ; so that
[[uc > ]] t0 for
f
every c A , and
e() < . Accordingly
inf e() = inf [[u+
e() > ]] = 0,
1
f
0
0
+
and once more v1 = supcAf u+
c is dened in L = L (A). As before, v1 a = ua L
for any a A ,
1,
1,
, with v2 a = u
Because [[v1 > ]] = e() Af , v1 M . Similarly, v2 = supcAf u
a
c is dened in M
1,
f
for every a A . So u = v1 v2 M , and
u=
ua = a
430
373H
for every a Af .
R
+
(ii)
Take 0 such that = (|u|
R 1) is nite. If > 0, there is a > 0 such
R Assume ().
+
a whenever
a R. As is arbitrary,
whenever
a , so that | a u| +
that a (|u| 1)
R
R
a t, then 1t | a u| + 1t (|u| 1)+ . Thus
limt0 supat | a u| = 0. Moreover, whenever t > 0 and
and the function a 7
lim supt
1
t
supat |
u| < ,
(iii) u is uniquely dened because u a must be ua , as dened in (i), for every a Af , and (A,
) is
semi-nite.
373I Lemma Suppose that u, v, w M 0, (AL ,
L ) are all equivalence classes of non-negative nonRt
Rt
R
R
increasing functions. If 0 u 0 v for every t 0, then u w v w.
P4 n
proof For n N, i 4n set ani = [[w > 2n i]]; set wn = i=1 2n ani . Then each ani is of the form [0, t] ,
so
R
R
R
R
P4n
P4n
u wn = i=1 2n a u i=1 2n a v = v wn .
ni
ni
u w = supnN
u wn supnN
v wn =
v w.
then
or u M 1, (A,
) and T T,
R
|T u v| u v .
uw (S, T ) = | Su w
T u w| for all S, T T = T, .
| T u w|
u w <
lim supt
1
t
Rt
0
u = 0,
1 t
t 0
u < .
373O
431
Decreasing rearrangements
R
Of course b 7 hu (b) is additive, so by 373H there is a unique Su M1, such that hu (b) = b Su for
R
every b Bf . Since both hu and w 7 Su w are linear and continuous on M,1 , and S(Bf ) is dense in
M,1 (369Od),
M,1 .
Su w = hu (w) = limT F
M1, .
Tu w
for every w
And this is true for every u
R
R
For any particular w M,1 , all the maps u 7 T u w are linear, so u 7 Su w also is; that is,
S : M1, M1, is linear.
Now S T . P
P () If u L1 and b, c Bf , then
Z
Z
Z
Z
Su Su = lim
T u (b c) sup T u (b c)
b
T F
T T
() If u L (A), then
sup
cBf ,c 1\e
(Su) kuk1 .
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, ) is semi-nite, so that we cannot be troubled by purely innite
elements of B.) Q
Q
Of course we now have limT F uw (T, S) = 0 for all u M1,0 , w M,1 , so that S = lim F in T . As
F is arbitrary, T is compact (2A3R).
373M Corollary Let (A,
) be a measure algebra and (B, ) a localizable measure algebra, and u
any member of M 1, (A,
). Then B = {T u : T T, } is compact in M 1, (B, ) for the topology
Ts (M 1, (B, ), M ,1 (B, )).
proof The point is just that the map T 7 T u : T, M1,0 is continuous for the very weak operator
topology on T, and Ts (M1, , M,1 ). So B is a continuous image of a compact set, therefore compact
(2A3Nb).
373N Corollary Let (A,
) be a measure algebra, (B, ) a localizable measure algebra and u any member
of M 1, (A,
); set B = {T u : T T, }. If hvn inN is any non-decreasing sequence in B, then supnN vn is
dened in M 1, (B, ) and belongs to B.
R
proof By 373M, hvn inN must have a cluster point v B for Ts (M1, , M,1 ). Now for any b Bf , b v
R
R
R
must be a cluster point of h n vn inN , because w 7 b w is continuous for Ts (M1, , MR,1 ). But h b vnRinN
is a non-decreasing sequence, so its only possible cluster point is its supremum; thus b v = limn b vn .
Consequently v b must be the supremum of {vn b : n N} in L1 . And this is true for every b Bf ;
as (B, ) is semi-nite, v is the supremum of hvn inN in L0 (B) and in M1, .
373O Theorem Let (A,
), (B, ) be measure algebras and u M 1, (A,
), v M 1, (B, ). Then the
following are equiveridical:
(i) there is a T T, such that T u = v,
Rt
Rt
(ii) 0 v 0 u for every t 0.
In particular, given u M 1, (A,
), there are S T,L , T TL , such that Su = u , T u = u.
432
373O
proof (i)(ii) is Lemma 373G. Accordingly I shall devote the rest of the proof to showing that (ii)(i).
(a) If (A,
), (B, ) are measure algebras and u M1, , v M1, , I will say that v 4 u if there is
a T T, such that T u = v, and that v u if v 4 u and u 4 v. (Properly speaking, I ought to write
(u,
) 4 (v, ), because we could in principle have two dierent 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.
(b) If (A,
) is a measure algebra and u1 , u2 M1, are such that |u1 | |u2 |, then u1 4 u2 . P
P There
1,
ai for 1 j m + 1.
i=0 i [i , i+1 [ , where 0 = 0, j =
i=0
For i < m, and for i = m if
am < , dene hi : M1, R by setting
hi (v) =
1 R
ai ai
R i+1
i
|T v| = |hi (v)|
ai
ai
|v|;
summing over i, kT vk1 kvk1 . Similarly, for any i m, v L (B), |hi (v)| kvk , so kT vk kvk .
Q
Thus T T,L . Since u = T u, we conclude that u 4 u, as claimed. Q
(d) If (A,
) is a measure algebra and u 0 in M1, , then u 4 u. P
P Let hun inN be a non-decreasing
with
sequence in S(A) with u0 0 and supnN un = u. Then hun inN is a non-decreasing sequence in M1,
L
supremum u , by 373Db and 373Dh. Also un 4 un 4 u for every n, by (b) and (c) of this proof. By 373N,
u 4 u. Q
Q
(e) If (A,
) 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
dene
i , ai , m, i as before. This
Pm
= i=0 i ai ; then u u
time, set ai = ai for i < m, am = supmjn aj , u
amd u
= u . Set
hi (v) =
R i+1
1
i+1 i i
; and if
am = , set
if i m, i+1 < (that is,
ai < ) and v M1,
L
hm (v) = limkF
1R k
v
k 0
R i+1
i
|v|,
373O
433
Decreasing rearrangements
whenever v L1L , i m, while |hi (v)| kvk whenever v L (AL ) and i m. So we can dene
Pm
, and get
T TL , by setting T v = i=0 hi (v)ai for every v M1,
L
u4u
= T u 4 u . Q
Q
(f ) If (A,
) is a measure algebra and u 0 in M1, , 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,
) might not be localizable. Set
0 = inf{ : 0,
[[u > ]] < },
Then e = supnN [[u > 0 + 2n ]] is a countable supremum of elements of nite measure, so the principal
ideal Ae , with its induced measure
e , is -nite. Now let hun inN be a non-decreasing sequence in S(A)
with u0 0 and supnN un = u; set u
= u e and u
n = un e, regarded as members of S(Ae ), for each
n. In this case
u
n 4 u
n 4 u
for every n. Because (Ae ,
e ) is -nite, therefore localizable, 373N tells us that u
4 u .
. As in part (e), choose a non-principal ultralter F on N and set
Let S TL ,e be such that Su = u
h(v) = limkF
for v M1,
. Now dene T : M1,
M1, by setting
L
L
1R k
v
k 0
T v = Sv + h(v)(1 \ e),
M1, .
1R t
v
t 0
= 1
for t =
a1 , while limt (t) < 1 , because (limt (t))1 u and v 1 1 and v 6 u .
Becaasue is continuoue, there is a such that () = 1 . Dene T0 TL ,L by setting
T0 w = (
1R
0
. Then T0 u 4 u u, and
for every w M1,
L
T0 u [0, [ = (
Rt
0
T0 u
1R
0
u ) [0, [ = 1 [0, [ .
Z t
v whenever t ,
T0 u = 1 t
0
Z t
Z t
Z t
Z
v whenever t .
u
T0 u =
T0 u +
=
0
Set
Rt
w) [0, [ + (w [, [ )
434
373O
u1 = T0 u [, [ ,
v1 = v [, [ .
Rt
u1 =
R +t
R +t
R +t
R +t
Rt
T0 u = 0 T0 u 1 0 v 1 = v = 0 v1
0
Pn
Pi
T w = w [0, [ + S(w [, [ ) [, [
hvn inN 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
by 373J. Q
Q
T1 u v =
|T u v|
u v
T1 u v
|T u v| =
|T u |v||, while
R
R
(iii) It will be enough to consider u = u . P
P If we can nd T TL , such that T u v = (u ) v ,
then we know from 373O that there is an S T,L such that Su = u , so that T S T and
T Su v =
(u ) v =
u v . Q
Q
>
0),
v
=
v
.
Also
(B
,
)
is
-nite,
therefore
localizable.
Now
if
we
can
nd T T,e such that
e e
R
R
T u v = u v , then ST will belong to T, , where S : L0 (Be ) L0 (B) is the canonical embedding
dened by the formula
[[Sw > ]] = [[w > ]] if 0,
and
ST u v =
T u v =
u v =
u v . Q
Q
ni = bni ,
ni = cni = ni n,i+1
373Q
435
Decreasing rearrangements
(because ni < if i > 0; this is really where I use the hypothesis that v M 0 ). For n N set
Kn = {i : i 1, ni > 0},
Tn w =
X 1 Z ni
w cni
ni n,i+1
iKn
; this is dened in L0 (B) because Kn is countable and hcni iiN is disjoint. Of course Tn :
for w M1,
L
1,
0
ML L (B) is linear. If w L (AL ) then
1 R ni
kwk ,
kTn wk = supiKn
w
ni
kTn wk1 =
iKn
ni
n,i+1
R ni
wcni =
n,i+1
iKn
R ni
n,i+1
w kwk1 ;
cni
Tn 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 ,
ni = n+1,2i ,
so that, for i 1,
cni
R ni
n,i+1
R ni
Tn u =
n,i+1
u=
cni
ni = n+1,2i + n+1,2i+1 ,
Tn+1 u.
|T u v|
Z
X
|T u| v
i=0
cmi
2m i|
i=1
2m i
i=0
X
i=0
T u| =
cmi
2m i
i=1
mi
m,i+1
2m i
cmi
|T u|
Tm u
cmi
u (v 2m 1)+
because
[m,i+1 , mi ]
|T u v| limm
u (v 2m 1)+ =
u v
u v = sup{ |T u v| : T T, } = sup{ T u v : T T, }.
436
373Q
proof There is a non-decreasing sequence hcn inN in B such that cn < for every n and v = supnN (v
cn ) . P
P For each rational q > 0, we can nd a countable non-empty set Bq B such that
b
S
(because (B, ) is semi-nite). Let hbn inN be a sequence running over qQ,q>0 Bq and set cn = supin bi ,
vn = v cn for each n. Then h|vn |inN and hvn inN are non-decreasing and supnN vn v in L0 (AL ).
But in fact supnN vn = v , because
by 373J, as usual.
373R Order-continuous operators: Proposition Let (A,
) be a measure algebra, (B, ) a localizable
(0)
) is semi-nite,
measure algebra, and T0 T (0) = T, . Then there is a T T = T, extending T0 . If (A,
T is uniquely dened.
proof (a) Suppose rst that T0 T (0) is non-negative, regarded as a member of L (M1,0 ; M1,0 ). In this
case T0 has an extension to an order-continuous positive linear operator T : M1, L0 (B) dened by
P I use 355F. M1,0 is a solid
saying that T w = sup{T0 u : u M1,0 , 0 u w} for every w 0 in M1, . P
linear subspace of M1, . T0 is order-continuous when its codomain is taken to be M1,0 , 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 M1, , let 0 be such that u1 = (w 1)+ is integrable. If
u M1,0 and 0 u w, then (u 1)+ u1 , so
T0 u = T0 (u 1)+ + T0 (u 1) T0 u1 + 1 L0 (B).
Thus {T0 u : u M1,0 , 0 u w} is bounded above in L0 (B), for any w 0 in M1, . L0 (B) is Dedekind
complete, because (B, ) is localizable, so sup{T0 u : 0 u w} is dened in L0 (B); and this is true for
every w (M1, )+ . Thus the conditions of 355F are satised 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 |,
= 12 (|T0 | + T0 ) and T0 = 12 (|T0 | T0 ), taken in L (M1,0 ; M1,0 ), all belong to T (0) (371G), so have
extensions S, S1 and S2 to order-continuous positive linear operators from M1, to L0 (B) as dened in (a).
Now for any w L1 ,
T0+
so kSwk kwk . Thus S T ; similarly, S1 and S2 can be regarded as operators from M1, to M1, ,
and as such belong to T . Next, for w 0 in M1, ,
373S
437
Decreasing rearrangements
and T = S1 S2 T , by 373Bc; while of course T extends T0+ T0 = T0 . Finally, because S1 and S2 are
order-continuous, T L (M1, ; M1, ), so T T . Q
Q
(c) If (A,
) is semi-nite, then M1,0 is order-dense in M1, (because it includes L1 , which is order-dense
0
in L (A)); so that the extension T is unique, by 355F(iii).
373S Adjoints in T (0) : Theorem Let (A,
) and (B, ) be measure algebras, and T any member of
R
R
(0)
f
f
Then
there
is
a
unique
operator
T
,
such that a T (b) = b T (a) for every
R
R
R a A , b B ,
1,0
1,0
and now u T v = T u v whenever u M (A,
), v M (B, ) are such that u v < .
(0)
T, .
proof (a) For each v M1,0 we can dene T v M1,0 by the formula
T v =
T (a) v
R
for every a Af . P
P Set a = T (a) v for each a Af ; because (a) v < , the
R integral is dened
and nite (373J). Of course : Af R is additive because is additive and T , and are linear. Also
limt0 supat |a| limt0
1
t
Rt
0
v = 0,
1 t
t 0
v = 0
T v = a for every
(0)
(b) Because the formula uniquely determines T v, we see that T : M1,0 M1,0 is linear. Now T T, .
P
P (i) If v L1 , then (because T v M1,0 ) |T v| = supaAf |T v| a, and
Z
Z
Z
Z
kT vk1 = |T v| = sup
|T v| = sup ( T v T v)
= sup
b,cAf
aAf
b,cAf
T (b c) v sup
b,cAf
(b c) v
v = kvk1 .
(ii) Now suppose that v L (B) M1,0 , and set = kvk . ?? If a = [[|T v| > ]] 6= 0, then T v 6= 0 so
v 6= 0 and > 0 and
a < , because T v M1,0 . Set b = [[(T v)+ > ]], c = [[(T v) > ]]; then
Z
Z
Z
Z
a < |T v| = T v T v = T (b c) v
a
kT (b c)k1 kb ck1 =
a,
which is impossible. X
X Thus [[|T v| > ]] = 0 and kT vk = kvk .
(0)
Putting this together with (i), we see that T T, . Q
Q
(0)
| Tu v
T un v|
|T |(u un ) v 0
438
as n , because
At the same time,
because
373S
u |T v|
|T |u v
| u T v
u v < . So
T u v = limn
u v < .
un T v|
(u un ) |T v| 0
T un v = limn
un T v =
u T v,
(d) Finally, to see that T is uniquely dened by the formula in the statement of the theorem, observe
that this surely denes T (b) for every b Bf , by the remarks in (a). Consequently it denes T on S(Bf ).
(0)
Since S(Bf ) is order-dense in M1,0 , and any member of T, must belong to L (M1,0 ; M1,0 ) (371Gb), the
restriction of T to S(Bf ) determines T (355J).
373T Corollary Let (A,
) and (B, ) be localizable measure algebras. Then for any T T, there is a
R
R
u v < .
R
(0)
(0)
proof The restriction T M1,0 belongs to T, (373Bb), so there is a unique S T, such that u Sv =
R
R
T u v whenever u M1,0 , v M1,0 are such that u v < (373S). Now there is a unique T T,
R
R
R
extending S (373R). If u 0 in M1, , v 0 in M1, are such that u v < , then uT v = T uv.
P
P If T 0, then both are
because both T and T are (order-)continuous. In general, we can apply the same argument to T + and
T , taken in L (M1, ; M1, ), since these belong to T, , by 373B and 355H, and we shall surely have
R
R
T = (T + )R (T ) . Q
Q As in 373S, it follows that u T v = T u v whenever u M1, , v M1, are
such that u v < .
373U Corollary Let (A,
) and (B, ) be localizable measure algebras, and : A B an ordercontinuous measure-preserving Boolean homomorphism. Then the associated map T T, (373Bd) has an
R
adjoint P T, dened by the formula a P (b) = (b a) for a Af , b Bf .
proof The adjoint P = T must have the property that
P (b) =
a
a P (b) =
T (a) b =
(a) b = (a b)
for every a A , b B . To see that this denes P uniquely, let S T, be any other operator with the
same property. By 373Hb, S(b) = P (b) for every b Bf , so S and P agree on S(Bf ). Because both P
and S are supposed to belong to L (M1, ; M1, ), and S(Bf ) is order-dense in M1, , S = P , by 355J.
f
373Xq
439
Decreasing rearrangements
> (f ) Let (X, , ) and (Y, T, ) be -nite measure spaces, with measure algebras (A,
) and (B, ), and
R
product measure on X RY . Let h : X Y R be a measurable function such that |h(x, y)|dx 1 for
-almost every y Y and |h(x, y)|dy 1 for -almost every x X. Show that
R there is a corresponding
T T, dened by writing T (f ) = g whenever f L1 () + L () and g(y) = h(x, y)f (x)dx for almost
every y.
R1
0
(0)
440
373Xr
R > (t) Set g(t) = t/(t + 1) for t 0, and set v = g , u = [0, 1] L (AL ). Show that u v = 1 >
T u v for every T TL ,L .
(0)
(0)
(0)
(0)
0
and v = g in L (A), where f (t) = t, g(t) = 1 2|t 12 | for t [0, 1]. Show that u = v , but that there is
no measure-preserving Boolean homomorphism : A A such that T v = u, writing T : L0 (A) L0 (A)
for the operator induced by , as in 364R. (Hint: show that {[[v > ]] : R} does not -generate A.)
(b) Let (A,
) be a totally nite homogeneous measure algebra of uncountable Maharam type. Let u,
1, +
v (M ) be such that u = v . Show that there is a measure-preserving automorphism : A A such
that T u = v.
Rt
Rt
(c) Let u, v M1,
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 TL ,L 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 TL ,L .
(d) Let (A,
) and (B, ) be measure algebras, and u M1, . (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 < u w there is a T T, such that T uw . (ii) Suppose that (B, ) is localizable and that
R
R
v M1, \ {T u : T T, }. Show that there is a w S(Bf ) such that v w > sup{ T u w : T T, }.
(Hint: use 373M and the Hahn-Banach theorem in the following form: if U is a linear space with the
topology Ts (U, V ) dened 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) > supuC f (u).) (iii) Hence prove 373O for localizable
(B, ). (iv) Now prove 373O for general (B, ).
(e) (i) Dene v L (AL ) as in 373Xt. Show that there is no T TL ,L 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 TL ,L such that T w = w.
(f ) Let (A,
) be the measure algebra of Lebesgue measure on [0, 1]. Show that T,L = T,L can be
identied, as convex ordered space, with TL , , and that this is a proper subset of TL , .
(g) Show that the adjoint operation of 373T is not as a rule continuous for the very weak operator
topologies of T, , T, .
373 Notes and comments 373A-373B are just alternative expressions of concepts already treated in 371F371H. My use of the simpler formula T, 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).
373 Notes
Decreasing rearrangements
441
The obvious examples of operators in T are those derived from measure-preserving Boolean homomorphisms, as in 373Bd, and their adjoints (373U). Note that the latter include conditional expectation operators. In return, we nd 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, [ , L ), 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 eective 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.
Of course the very weak operator topology is interesting in its own right; and see
373Xp-373Xq.
R t also
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 nd 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 dierence of order-continuous operators, just because
1,0
M
has an order-continuous norm (371Gb). It is therefore not surprising that members of T (0) can be
extended to members of T , at least when the codomain M1, is Dedekind complete (373R). It is also very
natural to look for a correspondence between T, and T, , because if T T, we shall surely have an
adjoint operator (T L1 ) from (L1 ) to (L1 ) , and we can hope that this will correspond to some member
of T, . 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 members of T (0) this is automatically
present (373S), and now the canonical isomorphism between T (0) and T gives us an isomorphism between
and are localizable (373T).
T, and T, when
442
374 intro.
(369N, 373C).
Let L be the Banach lattice dened from (369G), and the associate extended Fatou norm (369H-369I).
Then
(i) M,1 L M1, ;
R
(ii) is also T -invariant, and u v (u) (v) for all u, v M0, .
in L0 (AL ), because
[[|u| > n]] = . So there is a T T, such that T (|u| n1) = nw, by 373O, and
(u) (|u| n1) (T (|u| n1)) = (nw) = n (w).
Z
u v = sup{ |T u v| : T T, }
(373Q)
sup{ (T u) (v) : T T, } = (u) (v).
Generally, setting un = |u| n1, hun inN is a non-decreasing sequence with supremum u (373Db, 373Dh),
so
374C
Rearrangement-invariant spaces
u v = supnN
443
R a
0
u =
by (b) above. But this means that u M1, , so that u M1, (373F(b-ii)). Q
Q
(d) Next, is T -invariant. P
P Suppose that v M0, , T T, , u L0 (A) and (u) 1. Then
1,
u M , by (c), so
|u T v|
using 373J for the rst inequality. Taking the supremum over u, we see that (T v) (v); as T and v are
arbitrary, is T -invariant. Q
Q
(e) Finally, putting (d) and (c) together, L M1, , so that L M,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-nite measure algebra, as follows.
Theorem Let be a T -invariant extended Fatou norm on L0 (AL ), and (A,
) a semi-nite measure algebra.
(a) There is a T -invariant extended Fatou norm on L0 (A) dened by setting
(u) = (u ) if u M0, ,
= if u L0 (A) \ M0, .
= if v L0 (A) \ M0, .
|u T w|
u w (u ) (w ) (u ) (w) (u );
|u w|
(373E)
=
(u ) w
Z
u w = sup{ |u T w| : T TL ,L }
(373Q)
(u).
But because is the associate of (369I(ii)), this means that (u ) (u). Q
Q
444
374C
As M,1 is order-dense in L0 (A) (this is where I use the hypothesis that (A,
) is semi-nite), 369F(v)-(vi)
are satised, and is an extended Fatou norm. Q
Q
(iii) is T -invariant. P
P Take u M1, and T T, . There are S0 TL , and S1 T,L such that
because is T -invariant. Q
Q
(iv) We can now return to the denition of . I have already remarked that (u) = (u ) if u M0, .
For other u, we must have (u) = just because is a T -invariant extended Fatou norm (374B(i)). So the
denitions in the statement of the theorem and (i) above coincide.
(b) We surely have (v) = if v L0 (A) \ M0, , by 374B, because , like , is a T -invariant extended
Fatou norm. So take v M0, .
(i) If u L0 (A) and (u) 1, then
|v u|
as u is arbitrary, (v) (v ).
v u (v )(u ) = (v ) (u) (v );
|v w|
Z
v w = sup{ |v T w| : T TL , }
(373Q)
sup{ (v) (T w) : T TL , } = sup{ (v)((T w) ) : T TL , }
[[|u| > ]] =
P?? Suppose that u L \M0 , that is, that
(c)(i) The rst step is to note that L M0 . P
for some > 0. Then u 1 in L (AL ), so L (AL ) L . For each n N, set vn = [n, [ . Then
vn = v0 , so we can nd a Tn TL ,L such that Tn vn = v0 (373O), and (vn ) (v0 ) for every n. But as
hvn inN is a decreasing sequence with inmum 0, this means that is not an order-continuous norm. X
XQ
Q
(ii) Now suppose that A L is non-empty and downwards-directed and has inmum 0. Then
inf uA
[[u > ]] = 0 for every > 0 (put 364Nb and 321F together). But this means that B = {u : u A}
must have inmum 0; since B is surely downwards-directed, inf vB (v) = 0, that is, inf uA (u) = 0. As A
is arbitrary, is an order-continuous norm.
374D What is more, every T -invariant extended Fatou norm can be represented in this way.
374H
445
Rearrangement-invariant spaces
by 374B(ii). So (b) < and 369F(vi) is satised. Thus is an extended Fatou norm. Q
Q
, then
(b) is T -invariant. P
P If T TL ,L and w M1,
L
(T w) = sup (v)1
by 373G and 373I. Q
Q
(T w) v sup (v)1
|u v| sup (v)1
(u ) = sup (v)1
by the remark in (a) above. Q
Q
w v = (w)
u v (u),
u v = (u)
446
374H
) Suppose that A has an atom a. In this case, for any b A \ {0} there is an automorphism of
(
(A,
) such that a b 6= 0; now a must be an atom, so a = a b and a 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 a and therefore of the
same measure as a, so all atoms of A have the same measure.
) Now suppose that A is atomless. In this case, if a A has nite non-zero measure, Aa is homo(
geneous. P
P?? Otherwise, there are non-zero b, c a such that the principal ideals Ab , Ac are homogeneous
and of dierent Maharam types, by Maharams theorem (332B, 332H). But now there is supposed to be an
automorphism such that b c 6= 0, in which case Ab , Ab , Abc and Ac must all have the same Maharam
type. X
XQ
Q
Consequently, if a, b A are both of non-zero nite measure, the Maharam types of Aa , Aab and Ab
must all be the same innite cardinal .
(ii)(i) Assume (ii), and take a, b A \ {0}. If a b 6= 0 we can take to be the identity automorphism
and stop. So let us suppose that a b = 0.
) If A is purely atomic and every atom has the same measure, then there are atoms a0
(
Set
a, b0
b.
c = c if c a0 b0 or c (a0 b0 ) = 0,
= c (a0 b0 ) otherwise.
Then it is easy to check that is a measure-preserving automorphism of A such that a0 = b0 , so that
a b 6= 0.
) If Ac is Maharam-type-homogeneous with the same innite Maharam type for every non-zero
(
c of nite measure, set = min(1,
a,
b) > 0. Because A is atomless, there are a0 a, b0 b with
a0 =
b0 = (331C). Now Aa0 , Ab0 are homogeneous with the same Maharam type and the same
magnitude, so by Maharams theorem (331I) there is a measure-preserving isomorphism 0 : Aa0 Ab0 .
Dene : A A by setting
c = (c \ (a0 b0 )) 0 (c a0 ) 01 (c b0 );
last part of the proof will be of great importance; in the language of 381R, = (a0 0 b0 ).
374I Corollary Let (A,
) be a quasi-homogeneous semi-nite measure algebra. Then
(a) whenever a, b A have the same nite measure, the principal ideals Aa , Ab are isomorphic as measure
algebras;
(b) there is a subgroup of the additive group R such that ()
a whenever a A and
a < ()
whenever a A, and 0
a then there is a c a such that
c = .
proof If A is purely atomic, with all its atoms of measure 0 , set = 0 Z, and the results are elementary.
If A is atomless, set = R; then (a) is a consequence of Maharams theorem, and (b) is a consequence of
331C, already used in the proof of 374H.
u v = supAut
|u T v|,
f
proof (a) Suppose rst that u, v are non-negativeP
and belong to S(AP
), where Af is the ring {a :
a < },
m
n
as usual. Then they can be expressed as u =
a
,
v
=
b
where
.
.
.
m 0,
i
i
j
j
0
i=0
j=0
0 . . . n 0, a0 , . . . , am are disjoint and of nite measure, and b0 , . . . , bn are disjoint and of nite
measure. Extending each list by a nal term having a coecient of 0, if need be, we may suppose that
supim ai = supjn bj .
374K
447
Rearrangement-invariant spaces
because
suprs cr = supim ai = supjn bj = suprs dr ,
: A A is a measure-preserving automorphism.
Now
Ps
u = r=1 k(r) cr ,
so
u =
Ps
k(r) [tr1 , tr [ ,
u T v =
Ps
k(r) l(r)
cr =
v=
r=1
Ps
r=1
v =
r=1
Ps
r=1
l(r) dr ,
Ps
r=1
l(r) [tr1 , tr [ ,
u v .
u T v
|u0 | T |v0 | =
u0 v0 supAut
|u0 T v0 |
|u0 T v0 |.
proof Write for the associate of . Then 374J tells us that for any u, v M0, ,
u v = supAut
writing u , v for the decreasing rearrangements of u and v, and Aut for the group of measure-preserving
automorphisms of (A,
). But now, if u M1, and T T, ,
Z
(T u) = sup{ |T u v| : (v) 1}
(by 369I)
(by 373J)
Z
sup{ u v : (v) 1}
448
374K
(u).
As T , u are arbitrary, is T -invariant.
374L Lemma Let (A,
) be a quasi-homogeneous semi-nite measure algebra. Suppose that u, v
R
+
)
are
such
that
u
v = . Then there is a measure-preserving automorphism : A A such
(M0,
R
that u T v = .
(a) Suppose that A is purely atomic. Then it is surely innite, since otherwise
innite. Let be the common measure of its atoms. For each n N, set
u v could not be
n = inf{ : 0,
[[u > ]] 2n }.
[[u > n ]] 2n
[[u n ]].
a
n = 2n ,
[[u n ]],
a
n
a
n+1 ,
n+1
. So if we set
u
= kuk [0, [ supnN n 2n , 2n+1 ,
then u u
in L (AL ). Set an = a
n+1 \ a
n for each n; then han inN is disjoint and
an = 2n ,
an
[[u n+1 ]]
for each n.
Similarly, we can nd a non-increasing sequence hn inN in [0, [ and a disjoint sequence hbn inN in A
such that
bn = 2n ,
for each n, while
bn
[[v n+1 ]]
in L (AL ).
R
Now we are supposing that u v = , so we must have
R
P
= u
v = kuk kvk + n=0 2n n n .
At this point, recall that we are dealing with a purely atomic algebra in which every
S atom has measure
. Let An , Bn be the sets of atoms included in an , bn for each n, and A = nN An Bn . Then
#(An ) = #(Bn ) = 2n for eachSn. We therefore have
S a bijection : A A such that [B2n ] = A2n for
every n. (The point is that A \ nN A2n and A \ nN B2n are both countably innite.) Dene : A A
by setting
c = (c \ sup A) supaA,a c a.
Then is well-dened (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 b2n = a2n for every n. Consequently
R
P
P
u T v n=0 2n+1 2n+1
a2n = n=0 22n 2n+1 2n+1 = .
So we have found a suitable automorphism.
374L
449
Rearrangement-invariant spaces
Then
Also
[[u > n ]] 2n
[[u n ]]
for each n, so we can choose inductively a decreasing sequence h
an inN such that
[[u > n ]] a
n
[[u n ]]
and
a
n = 2n for each n. Set an = a
n \ a
n+1 ; then han inN is disjoint and
an = 2n1 , an [[u n ]]
for each n.
In the same way, we can nd hn inN , hbn inN such that
hbn inN is a disjoint sequence in A,
bn = 2n1
n1
n
and bn [[v n ]] for each n, and v supnN n+1 2
, 2 .
Now all the principal ideals Aan , Abn are homogeneous and of the same Maharam type, so there are
measure-preserving isomorphisms n : Abn Aan . Dene : A A by setting c = supnN n (c an );
then is a measure-preserving automorphism of A, and bn = an for each n. Since u an n an ,
v bn n bn for each n,
R
P
u T v n=0 2n1 n n ;
but on the other hand,
So
u T v = .
u v
n=0
u T v.
(c) Thirdly, consider the case in which A is atomless and not totally nite; take to be the common
Maharam type of all the principal ideals Aa where 0 <
a < . In this case, set
n = inf{ :
[[u > ]] 2n },
u supnZ n 2n , 2n+1 ,
n = inf{ :
[[v > ]] 2n }
v supnZ n 2n , 2n+1 .
There are disjoint families han inZ , hbn inZ such that
an =
bn = 2n for each n and
u supnZ n+1 an ,
v supnZ n+1 bn .
d2 = d \ supnZ b2n
while
is innite. Because
u T v
n=
u v
n=
2n n n
450
374L
u T v
1R
6
u v = .
proof Set V = {v : u v L1 for every u U }, so that V is a solid linear subspace of L0 which can be
identied with U (369C), and U becomes {u : u v L1 for every v V }; note that M,1 V M1,
(using 369Q).
R
If u UR + , v V + and : A A is a measure-preserving automorphism, then T u U , so vT u < ;
by 374L, u v is nite. But this means that if u U , v V and T T, ,
|T u v|
u v < .
a a
for
u L0 (A). Show that is a T -invariant extended Fatou norm. Find examples of (A,
) for which is, and
is not, order-continuous on L .
(h) Let (A,
) and (B, ) be semi-nite measure algebras and a T -invariant extended Fatou norm on
L0 (A). (i) Show that there is a T -invariant extended Fatou norm on L0 (B) dened by setting (v) =
sup{ (T v) : T T, } for v M1, . (ii) Show that (w) = sup{ (T w) : T T, } for every w M1, .
(iii) Show that when (A,
) = (AL ,
L ) then (v) = (v ) for every v M0, . (iv) Show that when
(B, ) = (AL ,
L ) then (u) = (u ) for every u M0, .
(i) Let (A,
) be a semi-nite measure algebra and an extended Fatou norm on L0 = L0 (A). Suppose
that L is a T -invariant subset of L0 . Show that there is a T -invariant extended Fatou norm which is
0
equivalentR to in the sense that, for some M > 0, (u) M (u) M 2 (u) for every
R u L . (Hint: show
rst that u v < for every u L and v L , then that sup (u)1, (v)1 u v < .)
374 Notes
Rearrangement-invariant spaces
451
(j) Suppose that is a T -invariant extended Fatou norm on L0 (AL ), and that 0 < w = w M1,
. Let
L
(A,
) be any semi-nite measure algebra. Show that the function u 7 (w u ) extends to a T -invariant
.) (When
extended Fatou norm on L0 (A). (Hint: (w u ) = sup{ (w T u) : T T,L } for u M1,
L
= k kp these norms are called Lorentz norms; see Lindenstrauss & Tzafriri 79, p. 121.)
(k) Let (A,
) be PN with counting measure. Identify L0 (A) with R N . Let U be {u : u R N , {n : u(n) 6=
0} is nite}. Show that U is a perfect Riesz space, and is rearrangement-invariant but not T -invariant.
(l) Let (A,
) 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 M1, and that U is T invariant. (Hint: assume U 6= {0}. Show that (i) a U whenever
a < (ii) V = {v : vu L1 u U }
1,
is rearrangement-invariant (iii) U , V M
.)
374Y Further exercises (a) Let (A,
) be a localizable measure algebra and U M1, a non-zero
T -invariant Riesz subspace which, regarded as a Riesz space, is perfect. (i) Show that U includes M,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 M0, , u v < u U }. (iii) Show that for any localizable measure
algebra (B, ) the set V = {v : v M1, , T v U T T, } is a perfect Riesz subspace of L0 (B), and
that V = {v : v M1, , T v U T T, }. (iv) Show that if, in (i)-(iii), (A,
) = (AL ,
L ), then
V = {v : v M 0, , v U }. (v) Show that if, in (iii), (B, ) = (AL ,
L ), then U = {u : u M 0, , u V }.
(b) Let (A,
) be a semi-nite measure algebra, and suppose that 1 q p < . Let wpq L0 (AL ) be
the equivalence class of the function t 7 t(qp)/p . (i) Show that for any u L0 (A),
wpq (u )q = p
R
0
tq1 (
[[|u| > t]])q/p dt.
(ii) Show that we have an extended Fatou norm k kp,q on L0 (A) dened by setting
1/q
R
kukp,q = p 0 tq1 (
[[|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, ) is another
measure algebra and T T, , then kT ukp,q kukp,q for every u M1, . (iv) Show that k kp,q is an
order-continuous norm on Lk kp,q .
such that u = v . Show that there is a measure-preserving automorphism of A such that T u = v, where
T : L0 (A) L0 (A) is the isomorphism corresponding to .
(d) In L0 (AL ) let u be the equivalence class of the function f (t) = tet . Show that there is no Boolean
automorphism of AL such that T u = u . (Hint: show that AL is -generated by {[[u > ]] : > 0}.)
(e) Let (A,
) be a quasi-homogeneous semi-nite measure algebra and C L0 (A) a solid convex orderclosed rearrangement-invariant set. Show that C M1, 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 its T -invariance which matters, and that rearrangement-invariant spaces are signicant
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 ,
L ) arises from its being more or less
the simplest algebra over which every T -invariant set can be described; see 374Xa.
I dont 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
452
374 Notes
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 members of T,
(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 joining 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
diculty.
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 dierence is I think in 374B(i);
with a T -invariant extended Fatou norm, we are necessarily conned to M 1, , the natural domain of the
methods used here. For perfect Riesz spaces we have examples like R N
= L0 (PN) and its dual, the space of
eventually-zero sequences (374Xk); these are rearrangement-invariant but not T -invariant, as I have dened
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 diculty 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 specic 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 innity, so the rearrangement needed
(part (a) of the proof) can be based on a region of nite measure.
proof (a) The rst step is to observe that if hun inN is any non-increasing sequence in L0 = L0 (A)
with inmum 0, and > 0, then {n(un u0 ) : n N} is bounded above in L0 . P
P For k N set
ak = supnN [[n(un u0 ) > k]]; set a = inf kN ak . ?? Suppose, if possible, that a 6= 0. Because un u0 ,
n(un u0 ) nu0 for every n and
a
a0
[[u0 > 0]] = [[u0 > 0]] = supnN [[u0 un > 0]].
So there is some m N such that a = a [[u0 um > 0]] 6= 0. Now, for any n m, any k N,
But a
where v = supnm n(un u0 ). And this means that inf kN [[v > k]] a 6= 0, which is impossible. X
X
Accordingly a = 0; by 364Ma, {n(un u0 ) : n N} is bounded above. Q
Q
(b) Now suppose that hun inN is a non-increasing sequence in L0 with inmum 0, and that w W is a
lower bound for {T un : n N}. Take any > 0. By (a), {n(un u0 ) : n N} has an upper bound v in L0 .
Because T is positive,
375F
453
Kwapiens theorem
1
n
1
n
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 disjoint sequence han inN such that an [[u > 0]] for each n, so that un = u an > 0, while um un = 0
if m 6= n. Now however
v = supnN n(h(un ))1 un
proof (a) This is elementary. Set every an = 0 in the second clause of the denition 375E to see that
(0) = 0. The other two parts are immediate consequences of the rst clause.
(b) (Compare 232Ba, 327Bb.) ?? Suppose, if possible, otherwise. Then for every n N there is an an A
such that
an 2n and (an ) . Set cn = supin ai for each n; then hcn inN is non-increasing and has
inmum 0 (since
cn 2n+1 for each n), but (cn ) for every n, so inf nN cn cannot be 0. X
X
454
375G
?? Suppose, if possible, that b0 sup C. Then there are c0 , . . . , ck C such that (b0 supik ci ) 4.
For each i k choose disjoint ai0 , . . . , aim A such that ci = inf jm (aij ). Let D be the set of atoms of
the nite subalgebra of A generated by {aij : i k, j m}, so that D is a nite partition of unity in A,
and every aij is the join of the members of D it includes. Set p = #(D), and for each d D take a maximal
1
1
}, so that
(d \ sup Ed ) < pm
; set
disjoint set Ed {e : e d,
e = pm
S
d = 1 \ sup( dD Ed ) = supdD (d \ sup Ed ),
1
1
1
so that
d is a multiple of pm
and is less than m
. Let E be a disjoint set of elements of measure pm
with
S
1
K = {K : K E, #(K) = p},
1
m
For every K K,
(sup K) =
R
Accordingly
is non-zero.
Because b2
(d ) ,
(b0 supik ci ) 4,
b1 = b0 supik ci \ (d ).
v 13 M b1 and
b1 ,
(mp)!
.
p!(mpp)!
so (sup K) . So if we set
P
v = KK (sup K),
so b1 3, where
R
M = #(K) =
b2 = b1 [[v < 21 M ]]
ci \ (d )
= inf jm (aij ) \ (d )
inf jm (aij \ d ).
inf (aij \ d )
jm
inf (sup{e : e E, e
jm
= inf sup{(e) : e E, e
jm
aij })
aij }
So there are e0 , . . . , em E such that ej aij for each j and b3 = b2 inf jm (ej ) 6= 0. Because
ai0 , . . . , aim are disjoint, e0 , . . . , em are distinct; set J = {e0 , . . . , em }. Then whenever K K and KJ 6= ,
b3 (sup K).
So let us calculate the size of K1 = {K : K K, K J 6= }. This is
M
(mpm1)!
p!(mppm1)!
(mpp)(mpp1)...(mppm)
=M 1
mp(mp1)...(mpm)
M 1(
mpp m+1
)
mp
1
2
M.
375I
Kwapiens theorem
455
X
But this means that b3 [[v 21 M ]], while also b3 [[v < 12 M ]]; which is surely impossible. 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 jn (aj ) = 0 whenever a0 , . . . , an Ad are
disjoint. P
P If
d > 0 this follows from (a), if we apply it to Ad and (
d)1
Ad . If
d = 0 then we can
just take b1 = b0 , n = 0. Q
Q
Let > 0 be such that (a) < b1 whenever
a . Let A1 A be a nite set such that
(sup A1 )
proof (a) Write B for the set of potential members of B; that is, the set of those b B such that
there is a nite partition of unity A A such that Tab is a Riesz homomorphism for every a A, writing
Tab (u) = T (u a) b. If I can show that B is order-dense in B, this will suce, 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 b1 < . Let > 0 be such that b2 = b1 [[T (1) ]] is non-zero. Dene
R
: A [0, [ by setting a = b2 T (a) for every a A. Then is countably additive, because , T and
are all additive and sequentially order-continuous (using 375A). Set N = {a : a = 0}; then N is a -ideal
of A, and (C,
) is a totally nite measure algebra, where C = A/N and
a = a for every a A (just as
in 321H).
(c) We have a function from C to the principal ideal Bb2 dened by saying that a = b2 [[T (a) > 0]]
for every a A. P
P If a1 , a2 A are such that a1 = a2 in C, this means that a1 a2 N ; now
[[T (a1 ) > 0]] [[T (a2 ) > 0]] [[|T (a1 ) T (a2 )| > 0]]
456
375I
R
is disjoint from b2 because b2 T (a1 a2 ) = 0. Accordingly b2 [[T (a1 ) > 0]] = b2 [[T (a2 ) > 0]] and we
can take this common value for (a1 ) = (a2 ). Q
Q
(d) Now is a -subhomomorphism. P
P (i) For any a1 , a2 A we have
[[T (a1 a2 ) > 0]] = [[T (a1 ) > 0]] [[T (a2 ) > 0]]
because
T (a1 ) T (a2 ) T (a1 a2 ) T (a1 ) + T (a2 ).
So (c1 c2 ) = (c1 ) (c2 ) for all c1 , c2 C. (ii) If hcn inN is a non-increasing sequence in C with inmum
0, choose an A such that an = cn for each n, and set a
n = inf in ai \ inf iN ai for each n; then a
n = cn so
(cn ) = [[T (
an ) > 0]] for each n, while h
an inN is non-increasing and inf nN a
n = 0. ?? Suppose, if possible,
that b = inf nN (cn ) 6= 0; set = 12 b . Then (b2 [[T (
an ) > 0]]) 2 for every n N. For each n,
take n > 0 such that (b2 [[T (
an ) > n ]]) . Then u = supnN nn1
an is dened in L0 (A) (because
1
1
1
an > k]] = 0). But now
supnN [[nn
an > k]] a
m if k maxim ii , so inf kN supnN [[nn
(b2 [[T u > n]]) (b2 [[T (
an ) > n ]])
for both i (of course Tab , like T , is sequentially order-continuous). But this means that
[[Tab u1 > 0]] [[Tab u2 > 0]] b [[T (e1 a) > 0]] [[T (e2 a) > 0]]
= b (e1 a ) (e2 a ) = 0
because a C, so d
7 (d a ) is a ring homomorphism, while e1 e2 = 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 were home.
375J Corollary Let A be a Dedekind -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
nN of Riesz homomorphisms from L (A) to U such that T =
n=0 Tn , in the sense that
n
T u = supnN i=0 Ti u for every u 0 in L0 (A).
proof By 369A, U can be embedded as an order-dense Riesz subspace of L0 (B) for some localizable measure
algebra (B, ); 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 ibB as in 375I. Note that L (B) can be identied with bB 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) = b T (u a);
otherwiseP
set Tnb = 0. Then Tnb : L0 (A) L0 (Bb ) is a Riesz homomorphism; because Ab is a nite partition
of unity, n=0 Tnb u = b T u for every u L0 (A). But this means that if we set Tn u = hTnb uibB ,
Q
Tn : L0 (A) bB L0 (Bb )
= L0 (B)
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 -complete Boolean algebra, (B, ) is a semi-nite 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.
375Yc
Kwapiens theorem
457
(b) If (A,
) and (B, ) are homogeneous probability algebras and (A) > (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 -subhomomorphism : 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 . 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 (a) > 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 i T : A W
is order-continuous.
(b) Let A be an atomless Dedekind -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.
(d) Let A and B be Boolean algebras and : A B a -subhomomorphism as dened in 375E. Show
that 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 (1) 1. Let
an
R
R hbn inN enumerate
order-dense subset of G (316Yn). For each n N take non-zero bn bn such that T (bn ) 2n2 T (1)
and consider T (supnN bn ).) (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,
n >P0 such that
an 2n and if bn [[T (1) > 0]] then bn [[T (an ) > n ]] 6= 0. Consider T u where
1
u = n=0 nn an .)
(f ) In 375J, show that for any u L0 (A)
Pm
i=0
Ti u| > 0]] = 0.
> (g) Prove directly, without quoting 375E-375K, that if A is a Dedekind -complete Boolean algebra
then every positive linear functional from L0 (A) to R is a nite sum of Riesz homomorphisms.
375Y Further exercises (a) Show that the following are equiveridical: (i) there is a purely atomic
probability space (X, , ) such that = PX and {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
: 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 : A {0, 1} which is not order-continuous. (Compare 363S.)
(b) Give an example of an atomless Dedekind -complete Boolean algebra A such that L0 (A) 6= {0}.
(c) Let A, B be Dedekind -complete Boolean algebras of which B is weakly -distributive. Let
T : L0 (A) L0 (B) be a positive linear operator. Show that a 7 [[T (a) > 0]] : A B is a -subhomomorphism.
458
375Yd
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 dened 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 account 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, , ) and (Y, T, ) be -nite measure spaces; you can
take them both to be
X Y.
R [0, 1] with Lebesgue measure if you like. Let be the product measure on
1
If k L
(),
then
k(x,
y)dx
is
dened
for
almost
every
y,
by
Fubinis
theorem;
so
if
f
L
()
then
R
g(y) = k(x, y)f (x)dx is dened for almost every y. Also
376C
Kernel operators
g(y)dy =
459
is dened, because (x, y) 7 k(x, y)f (x) is -virtually measurable, dened -a.e. and is dominated by a
multiple of the integrable function k. Thus k denes a function from L () to L1 (). Changing f on a set
of measure 0 will not change g, so we can think of this as an operator from L () to L1 (); and of course we
can move immediately to the equivalence class of g in L1 (), so getting an operator Tk from L () to L1 ().
This operatorRis plainly linear; also it is easy to check that Tk T|k| , so that Tk L (L (); L1 ()), and
that kTk k |k|. Moreover, changing k on a -negligible set does not change Tk , so that in fact we can
speak of Tw for any w L1 ().
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 () L (L (); L1 ()).
And we can ask what happens when L () and L1 () are replaced by other function spaces, dened 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 terms of its action on the dual. In the language of the special case
above, if h L (), then we shall have
where for a A, b B I write a b for the corresponding member of A B (315M), identied 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 (1A ) = (1B ) = 1C . Now set u v = u
v. The properties listed in (a)-(c) are just a
matter of putting the denition 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.
460
376C
(C, ).
and
) and v L1 = L1 (B, ) then u v L1 = L1 (C, )
(a) If u L1 = L1 (A,
uv =
R R
u v,
(b) Let W be a Banach space and : L1 L1 W a bounded bilinear map. Then there is a unique
bounded linear operator T : L1 W such that T (uv) = (u, v) for all u L1 and v L1 , and kT k = kk.
(c) Suppose, in (b), that W is a Banach lattice. Then
(i) T is positive i (u, v) 0 for all u, v 0;
(ii) T is a Riesz homomorphism i u 7 (u, v0 ) : L1 W , v 7 (u0 , v) : L1 W are Riesz
homomorphisms for all v0 0 in L1 and u0 0 in L1 .
proof (a) I refer to the proof of 325D. Let (X, , ) and (Y, T, ) be the Stone spaces of (A,
) and (B, )
can be identied with the measure algebra of the c.l.d. product measure on X Y
(321K), so that (C, )
(part (a) of the proof of 325D), and L1 , L1 , L1 can be identied with L1 (), L1 () and L1 () (365B). Now
corresponds to f g as dened in 376B. (Look rst 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 () L1 ()
R
R to L (),
R with
R
1
1
1
f g = f g (253D); so that if u L and v L we must have u v L , with u v = u v.
As in 253E, it follows that ku vk1 = kuk1 kvk1 .
(b) In view of the situation described in (a) above, this is now just a translation of the same result about
L1 (), L1 () and L1 (), 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, c
expressible as supin ai bi where a0 , . . . , an are disjoint (315Na); evidently this will belong to (A B)f
i
ai bi is nite 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
{(a b) : a Af , b Bf } is norm-dense in L1 . Of course M can also be regarded as the linear span of
{c : 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
c M for every c (A B)f , we need to know that c can be expressed as a disjoint union of members of
A B, as noted above.)
(i) If T is positive then of course (u, v) = T (u v) 0 whenever u, v 0, since u v 0. On
the other hand, if is non-negative on U + V + , then, in particular, T (a b) = (a, b) 0 whenever
(a, b + b ) (a , b + b )
= (a a , b + b ) = 0
376E
461
Kernel operators
n
m
X
X
T (aj bj ))
T (ai bi )) (
T c T c (
j=0
i=0
im,jn
T (ai bi ) T (aj bj ) = 0,
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 rst thing to check is that the formula oered does
R dene 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
(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 inmum 0, then Tw [A]
is downwards-directed, and for any v V +
(inf Tw [A])(v) = inf uA Tw (u)(v) = inf uA
w (u v) = 0
(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 inmum 0, then for any u U + , v V +
(because
0.
w (u v) = 0
462
376E
(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 Tw = 0 whenever w,
w W + , [[w > 0]] d and [[w > 0]] 1C \ d. (i) If d1 , d2 D, w, w W + , [[w > 0]] d1 d2 and
[[w > 0]] (d1 d2 ) = 0, then set w1 = w d1 , w2 = w w1 . In this case
[[w1 > 0]] d1 ,
so
Tw1 Tw = Tw2 Tw = 0,
Tw Tw (Tw1 Tw ) + (Tw2 Tw ) = 0.
As w, w are arbitrary, d1 d2 D. Thus D is closed under . (ii) The symmetry of the denition 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, w W + are such that [[w > 0]] c, [[w > 0]] c = 0, then consider {w d : d D}. This is
upwards-directed, with supremum w; so Tw = supdD Twd , because the map q 7 Tq is order-continuous.
Also Twd Tw = 0 for every d D, so Tw Tw = 0. As w, w are arbitrary, c D; as D is arbitrary,
D is an order-closed subalgebra of C. (v) If a A and w, w W + are such that [[w > 0]] a 1B ,
[[w > 0]] (a 1B ) = 0, then any u U + is expressible as u1 + u2 where u1 = u a, u2 = u (1A \ a).
Now
Tw (u2 )(v) =
w (u2 v) =
w (a 1B ) (u v) ((1A \ a) 1B ) = 0
w (u (v (1B \ b))) +
w (u (v b)) = 0.
B = {b : b B, v V, b v}.
Because U and V are order-dense in L0 (A) and L0 (B) respectively, RA and B are order-dense
in A and B.
R
Also both are upwards-directed. So supaA,bB a b = 1C and 0 < w = supaA,bB ab w. Take a A,
R
b B such that ab w > 0; then there are u U , v V such that a u, b v, so that
Tw (u)(v)
ab
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 for the band it generates in L (U ; V ). Then W
is order-dense in W
c. P
W
P Suppose that S > 0 in
W =W
(353Ba). Then S
/ W , so there is a w W such that S Tw > 0. Set w1 = w 1C . Then
w = supnN w nw1 , so Tw = supnN Tw nTw1 and R = S Tw1 > 0.
Set U1 = U L1 (A,
). Because U is an order-dense Riesz subspace of L0 (A), U1 is an order-dense Riesz
1
1
subspace of L = L (A,
), therefore also norm-dense. Similarly V1 = V L1 (B, ) is a norm-dense Riesz
1
1
subspace of L = L (B, ). Dene 0 : U1 V1 R by setting 0 (u, v) = R(u)(v) for u U1 , v V1 . Then
0 is bilinear, and
376E
Kernel operators
463
for every u U1 , v V1 .
Because U1 is norm-dense in L1 , U1+ is dense in (L1 )+ , and similarly V1+ is dense in (L1 )+ , so U1+ V1+
is dense in (L1 )+ (L1 )+ ; now 0 is non-negative on U1+ V1+ , so (being continuous) is non-negative on
(L1 )+ (L1 )+ . By 376Cc, h 0 in (L1 ) and w 0 in L (C). In the same way, because 0 (u, v) Tw (u)(v)
for u U1+ and v V1+ , w w1 w in L0 (C), so w W . We have
Tw (u)(v) =
w (u v) = R(u)(v)
for all u U1 , v V1 . If u U1+ , then Tw (u) and R(u) are both order-continuous, so must be identical, since
V1 is order-dense in V . This means that Tw and R agree on U1 . But as both are themselves order-continuous
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 < Tw 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
0
c , so Q
is injective. Now Q(S)
c. P
Q(S)
> 0 whenever S > 0 in W
W for every S W
P It is enough
So Q(S)
(u v) = supwC w (u v) belongs to L1 (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.
464
(u v) (u v ) =
376E
(u u ) (v v ) = ( u u )( v v ) = fu (u )gv (v )
must be non-zero (364Ma). Because U and V are order-dense in L0 (A), L0 (B) respectively,
1A = sup{[[u > 0]] : u U },
f (u )g(v )
(because w 7
fu (u )gv (v )
Z
= sup
(u v) (u v )
uA,vB
Z
sup
((u v) nc) (u v )
uA,vB
Z
=
sup ((u v) nc) (u v )
sup
uA,vB
uA,vB
w (u v ) is order-continuous)
Z
Z
= (nc) (u v ) = n u v ,
c
which is impossible. X
X
Thus C is bounded above in L0 (C), and has a supremum w L0 (C). If u U + , v V + then
Z
Z
w (u v ) = sup
(u v) (u v )
uA,vB
sup
uA,vB
Thus w W and
W
c.
Pf g = T w 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
g V ; as W is a band, it must include Z.
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
so there is a u U such that 0 fu h and fu (u0 ) > 0. This time, u u0 > 0 so (because U is
order-dense in L0 (A)) there is a u1 U such that h(u1 ) > 0 and [[u1 > 0]] [[u > 0]].
We can express T as Tw where w W + . In this case, we have
376H
Kernel operators
so
465
0 6= [[w > 0]] [[u1 v1 > 0]] = [[w > 0]] ([[u1 > 0]] [[v1 > 0]])
and w (u v) > 0, so
[[w > 0]] ([[u > 0]] [[v > 0]]) = [[w > 0]] [[u v > 0]],
Tw Pfu gv = Tw Tuv = Tw(uv) > 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.
proof Setting V # = {v : v L0 (B), v v 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 inN is a sequence
in U such that limn g(un ) = 0 whenever g U and |g| f , hT un inN order*-converges to 0 in V
(denition: 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 Hausdor space,
with e corresponding to X (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 inN 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 inN order*-converges to 0 in V .
376H Theorem Let U be a Riesz space and V a weakly (, )-distributive Dedekind complete Riesz
space (denition: 368N). Suppose that T L (U ; V ). Then the following are equiveridical:
(i) T is an abstract integral operator;
(ii) whenever hun inN is an order-bounded sequence in U + and limn f (un ) = 0 for every f U ,
then hT un inN order*-converges to 0 in V ;
(iii) whenever hun inN is an order-bounded sequence in U and limn f (un ) = 0 for every f U , then
hT un inN 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 inN 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 nN supmn T um , which is dened 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
466
376H
w w0 + inf nN supmn T1 um = w0 . Q
Q
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 rst equality, and then subtract both sides from v0 ). Now Su f (u)v0 for every u 0,
so that for any > 0
supu[0,u2 ],f (u) T2+ u Su (1 )v0
and accordingly
supu[0,u2 ],f (u) T2+ u = v0 ,
since the join 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) T2+ u = sup0u uu2 ,f (u) T2 u = sup0u u2 ,f (u ) T2 u .
For k N set Ak = {u : 0 u u2 , f (u) 2k }. We know that
Bk = {supuI T2 u : I Ak is nite}
is an upwards-directed set with supremum v0 for each k. Because V is weakly (, )-distributive, we can
nd a sequence hvk ikN such that vk Bk for every k and v1 = inf kN vk > 0. For each k let Ik Ak be a
nite set such that vk = supuIk T2 u.
Because each Ik is nite, we can build a sequence hun inN in [0, u2 ] enumerating each in turn, so that
limn f (un ) = 0 (since f (u) 2k if u Ik ) while supmn T2 um v1 for every n (since {um : m n}
always includes some Ik ). Now hT2 un inN does not order*-converge to 0.
P Express |g| as g1 + g2 where g1 belongs to the band
However, limn g(un ) = 0 for every g U . P
of U generated by f and g2 f = 0 (353Hc). Then g2 (un ) = g2 (u2 ) = 0 for every n, by the choice of u2 .
Also g1 = supnN g1 nf (352Vb); so, given > 0, there is an m N such that (g1 mf )+ (u2 ) and
(g1 mf )+ (un ) for every n N. But this means that
|g(un )| |g|(un ) + mf (un )
Q
for every n, and lim supn |g(un )| ; as is arbitrary, limn g(un ) = 0. Q
Now, however, part (a) of this proof tells us that hT1 un inN is order*-convergent to 0, because T1 Z,
while hT un inN is order*-convergent to 0, by hypothesis; so hT2 un inN = hT un T1 un inN order*-converges
to 0. X
X
This contradiction shows that every operator satisfying the condition (ii) must be in Z.
376J
467
Kernel operators
376I The following elementary remark will be useful for the next corollary and also for Theorem 376S.
Lemma Let (X, , ) be a -nite measure space and U an order-dense solid linear subspace of L0 (). Then
there is a non-decreasing sequence hXn inN of measurable subsets of X, with union X, such that Xn U
for every n N.
proof Write A for the measure algebra of , so that L0 () can be identied with L0 (A) (364Jc). A = {a :
a A \ {0}, a U } is order-dense in A, so includes a partition of unity hai iiI . Because is -nite, A is
376J Corollary Let (X, , ) and (Y, T, ) be -nite measure spaces, with product measure on X Y .
Let U L0 (), V L0 () be perfect order-dense solid linear subspaces, and T : U V a linear operator.
Write U = {f : f L0 (), f U }, V# = {h : h L0 (), 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 () such that
() |k(x, y)f (x)h(y)|d(x, y) <R for every f U, h V# ,
() if f U and we set g(y) = k(x, y)f (x)dx wherever this is dened, then g L0 () and T f = g ;
(iii) T L (U ; V ) and whenever hun inN is an order-bounded sequence in U + and limn h(un ) = 0 for
every h U , then hT un inN order*-converges to 0 in V .
Remark I write d(x, y) above to indicate integration with respect to the product measure . Recall that
in the terminology of 251, can be taken to be either the primitive or c.l.d. product measure (251K).
proof The idea is of course to identify L0 () and L0 () with L0 (A) and L0 (B), where (A,
) and (B, )
are the measure algebras of and , so that their localizable measure algebra free product can be identied
with the measure algebra of (325Eb), while V # = {h : h V# } can be identied with V .
R
R
(a)(i)(ii) By 376F, there is a w L0 () such that w(uv ) is dened andR equal to T uv for every
0
#
Ru U , v V . Express w as k where k L (). If f U and h V then |k(x, y)f (x)h(y)|d(x, y) =
|w (f h | is nite, so (ii-) is satised.
R
Now take any f U, and set g(y) = k(x, y)f (x)dx whenever this is dened in R. Write F for the set
#
of those F T such that F
R V . Then for any F F, g is dened almost everywhere on F and gF is
-virtually
P k(x, y)f (x)F (y)d(x, y) is dened in R, so by Fubinis theorem (252B, 252C)
R measurable. P
gF (y) = k(x, y)f (x)F (y)dx is dened for almost every y, and is -virtually measurable; now gF = gF F .
Q
Q Next, there is a sequence hFn inN in F with union Y , by 376I, because V is perfect and order-dense, so
V # must also be order-dense in L0 ().
For each n N, thereSis a measurable set Fn Fn dom g such that gFn is measurable and Fn \ Fn is
negligible. Setting G = nN Fn , G is conegligible and gG is measurable, so g L0 ().
If g L0 () represents T u L0 (), then for any F F
g =
T u (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 =a.e. g. Thus g also represents T u, as required in (ii-).
(b)(ii)(i) Set w = k in L0 (). If f U and h V# the hypothesis () tells us that (x, y) 7
k(x, y)f (x)h(y) is integrable (because it surely belongs to L0 ()). By Fubinis theorem,
R
g(y)h(y)dy
where g(y) = k(x, y)f (x)dx for almost every y, so that T f = g , by (). But this means that, setting
u = f , v = h ,
w (u v ) =
T u v ;
468
376J
Enumerating nN In as hun inN , as in part (d) of the proof of 376H, we see that hun inN is order-bounded
and limn f (un ) = 0 for every f U (indeed, hun inN order*-converges to 0 in U ), while hT un inN 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 : L (U ; V )
L (V ; U ) dened by the formula
(T v)(u) = (T u)(v) for every u U , v V .
Then v A whenever |v | |v| A and v + v A for all v, v A (if f (u) > 0 whenever 0 < u |v| and
f (u) > 0 whenever 0 < u |v |, then (f + f )(u) > 0 whenever 0 < u |v + v |); moreover, if v0 > 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 hun inN in B converging to w for the norm. For each n, choose fn (U )+ such
that fn (u) > 0 whenever 0 < u un . Set
P
1
fn
f = n=0 n
2 (1+kfn k)
in U = U . Then whenever 0 < u w there is some n N such that u un > 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 < u w. This g serves. P
P?? Otherwise, there
is some > 0 such that for every n N we can nd a un [0, w] with g(un ) 2n and kun k . Set
vn = supin ui ; then 0 vn w, g(vn ) 2n+1 and kvn k for every n N. But hvn inN is nondecreasing, so v = inf nN vn must be non-zero, while 0 v w and g(v) = 0; which is impossible. X
X
Q
Q
376O
Kernel operators
469
p
proof Set q = p1
if p is nite, 1 if p = . We can identify Lp () with V , where V = Lq ()
= Lp ()
(366Dc, 365Mc) 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 > 0 there is a u U + such that
g(|u| u )+ 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 suphA (h) is nite for every 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 suphA |(h)| M kk for every Wg . Because g is the standard order unit of Wg , we have
kk = ||(g) and |(h)| M ||(g) for every 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 satised and Wg is perfect. Q
Q
470
376O
Accordingly T is continuous for the topologies Ts (U, W ) and Ts (Wg , Wg ), because every element of
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 ; accordingly T [C ] 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 0 , . . . , n
T [C] such that k(|| supin |i |)+ k for every T [C] (354Rb), so that (|| supin |i |)+ (g) for
every T [C].
Translating this back into terms of C itself, and recalling that T is a Riesz homomorphism, we see that
there are u0 , . . . , un C such that g(|u| supin |ui |)+ for every u C. Setting u = supin |ui | we
have the result.
Wg
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 = supkuk1 g(|T u|) is nite. 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 kuk1 u, we see that
g(|T u|) Mg kuk for every u U .
Next, we nd 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 + .
n
n
X
X
ui = u}
g(|T ui |) :
sup g(v) = sup{
vB
i=0
i=0
n
n
X
X
ui = u} = Mg kuk
Mg kui k :
sup{
i=0
i=0
is nite. By 356K, B is bounded above in V ; and of course any upper bound for B is also an upper bound
for {T u : 0 u 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 adjoint operator T L (V ; U ), as
in 376K. Now if g 0 in V and hgn inN is a sequence in [0, g] such that limn gn (v) = 0 for every v V ,
hT gn inN order*-converges to 0 in U . P
P For any > 0, there is a v V + such that g(|T u| v )+
whenever kuk 1; consequently
kT gn k = sup (T gn )(u) = sup gn (T u)
kuk1
kuk1
gn (v ) + sup gn (|T u| v )+
kuk1
gn (v ) + sup g(|T u| v )+ gn (v ) +
kuk1
376S
471
Kernel operators
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, , ) be a measure space, (Y, T, ) a -nite measure space, and the c.l.d. product
measure on X Y . Suppose that k is a -integrable real-valued function. Then for any > 0 there is a
nite partition E0 , . . . , En of X into measurable sets such that kk k1 k1 , where
k1 (x, y) =
1
Ei
Ei
Ei Y
|k k1 | =
R R
Y
Ei
1
Ei
Ei
k(t, y)dt
is dened.
Then = k1 (x, y) for every x Ei . For every
x Ei , we must have k2 (x, y) = where
R
P
= {j : Ei Hj , y Fj }. But in this case, because Ei k(x, y) dx = 0, we have
If ,
if ,
Ei
Ei
Ei
in either case,
1
2
Ei
Ei
Ei
Ei
Ei
1R
2 Ei
Ei
Ei
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
as required.
|k k1 | 2 |k k2 | ,
376S Theorem Let (X, , ) be a complete locally determined measure space, (Y, T, ) a -nite measure
space, and the c.l.d. product measure on X Y . Let be an extended Fatou norm on L0 () and write L
0
for {g : g L (), (g ) < }, where is the associate extended Fatou norm of (369H-369I). Suppose
f L1 (), g L .
472
376S
For x X set kx (y) = k(x, y) whenever this is dened. Then kx L0 () for almost every x; set
vx = kx L0 () for such x. In this case x 7 (vx ) is measurable and dened and nite almost everywhere,
and kT k = ess supx (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 oended if I now speak of an extended Fatou norm on L0 (), with
the associated function spaces L , L L0 , taking for granted the identication in 364Jc.
Recall that (f g)(x, y) = f (x)g(y) for x dom f , y dom g (253B).
By ess supx (vx ) I mean
inf{M : M 0, {x : vx is dened and (vx ) M } is conegligible}
(see 243D).
R
proof (a) To see that the formula (f, g) 7 k (f g) gives rise to an operator in L (U ; (L ) ), 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 ; see 325B.)
R The rst step, of course, is to note that changing
f or g on a negligible set does not aect the integral k (f g), so that we have a bilinear functional on
L1 L ; and the other essential element is the fact that the maps f 7 (f Y ) , g 7 (X g) are
order-continuous (put 325A and 364Rc together).
Eri
(r)
(376R). Now k (r) is also integrable over E Yn , so kx L0 () 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 = kx whenever x, x belong
to the same Eri .
(r)
If Eri > 0, then vx must be dened for every x Eri . If v L is represented by g L then
Z
Z
k (Eri (g Yn )) =
k(t, y)g(y)d(t, y)
Eri Yn
Z
Z
= Eri k (r) (x, y)g(y)dy = Eri vx(r) v
for any x Eri . But this means that
Eri
for every v L , so
(r)
vx v =
T (Eri
) v Yn
376Xc
473
Kernel operators
(r)
vx =
T (Eri
) Yn ,
Eri
(r)
(vx )
kT kkEri
k1
Eri
= kT k
(r)
for every x
PEri . RThis is true whenever Eri > 0, so in fact (vx ) kT k for almost every x E.
Because rN EYn |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,
(r)
hvx irN order*-converges to vx Yn (in L0 ()) for almost every x E. But this means that, for almost
every x E,
(r)
(369Mc). Now
y)f (x)g(y)d(x, y) =
k(x,
GY
f (x)
k(x, y)g(y)dydx
is dened.
R
y)g(y)|dy. Then h is nite almost everywhere and measurable. For
For x X0 , set h(x) = |k(x,
x A E,
R
R
y)g(y)|dy = |vx g | (g ).
|k(x,
So the measurable set G = {x : h(x) (g )} includes A E, and (G \ G ) = 0. Consequently
|
|f (x)|h(x)dx kf k1 (g ),
k (f g) when f L1 () and g L ,
R
and the formula just above tells us that | Tu v | kuk1 (v ) for every u L1 (), v L ; that is,
(Tu) kuk1 for every u L1 (); that is, kTk . But now (c) tells us that (
vx ) for almost every
x X, where vx is the equivalence class of y 7 k(x, y), that is, vx = vx for x G X0 , 0 for x X \ G.
So (vx ) 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 is arbitrary, x 7 (vx ) is measurable. Q
Q
(Tf ) g =
(e) Finally, the ideas in (d) show that kT k ess supx (vx ). P
P Set M = ess supx (vx ). If f L1 ()
and g L , then
as g is arbitrary, (T f ) M kf k1 ; as f is arbitrary, kT k M . Q
Q
376X Basic exercises > (a) Let be Lebesgue measure on R. Let h be a -integrable real-valued
function with khk1 1,R and set k(x, y) = h(y x) whenever this is dened. Show that if f is in either L1 ()
or L () then g(y) = k(x, y)f (x)dx is dened for almost every y R, and that this formula gives rise to
an operator T T, as dened in 373A. (Hint: 255H.)
(b) Let (A,
) and (B, ) be semi-nite measure algebras with localizable measure algebra free product
and take p [1, ]. Show that if u Lp (A,
and
(C, ),
) and v Lp (B, ) then u v Lp (C, )
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 .
474
376Xd
(j) Let (X, , ) be a -nite measure space and (Y, T, ) a complete locally determined measure space.
Let U L0 (), V L0 () be solid linear subspaces, of which V is order-dense; write V # = {v : v
L0 (), v v is integrable for every v V }, U = {f : f L0 (), f U }, V = {g : g L0 (), g V },
V# = {h : h L0 (), h V # }. Let be the c.l.d. product measure on X Y , and k LR0 () 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 dened for almost every y Y , and that hf V# . (ii) Show
we haveRa map T L (U ; V # ) dened
R that
(l) Let U be a perfect Riesz space. Show that a set C U is relatively compact for Ts (U, U ) i for
every g (U )+ , > 0 there is a u U such that g(|u| u )+ for every u C. (Hint: 376O and the
proof of 356Q.)
> (m) Let be Lebesgue measure on [0, 1], and 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., = k k ).
376Yj
Kernel operators
475
(n) Suppose, in 376Xj, that U = L for some extended Fatou norm on L0 () and that V = L1 (), so
that V # = L (). Set ky (x) = k(x, y) whenever this is dened, uy = ky whenever ky L0 (). Show that
uy L for almost every y Y , and that the norm of T in B(L ; L ) is ess supy (uy ). (Hint: do the case
of totally nite Y rst.)
376Y Further exercises (a) Let U , V and W be linear spaces (over any eld F ) and : U V W
a bilinear map. Let W0 be the linear subspace of W generated by [U V ]. Show that the following are
equiveridical: (i) for every linear space Z over F and every bilinear : U V Z, there is a (unique)
linear operator T : W0 Z such that
Pn T = (ii) whenever u0 , . . . , un U are linearly independent
and v0 , . . . , vn V are non-zero,
(ui , vi ) 6= 0 (iii) whenever u0 , . . . , un U are non-zero and
i=0P
n
v0 , . . . , vn V are linearly independent, i=0 (ui , vi ) 6= 0 (iv) for any Hamel bases hui iiI , hvj ijJ of U
and V , h(ui , vj )iiI,jJ is a Hamel basis of W0 (v) for some pair hui iiI , hvj ijJ of Hamel bases of U and
V , h(ui , vj )iiI,jJ is a Hamel basis of W0 .
their localizable measure algebra free
(b) Let (A,
), (B, ) be semi-nite measure algebras, and (C, )
0
0
0
product. Show that : L (A) L (B) L (C) satises the equivalent conditions of 376Ya.
(c) Let (X, , ) and (Y, T, ) be semi-nite measure spaces and the c.l.d. product measure on X Y .
Show that the map (f, g) 7 f g : L0 () L0 () L0 () induces a map (u, v) 7 u v : L0 () L0 ()
L0 () possessing all the properties described in 376B and 376Ya.
(d) Let (A,
) be the measure algebra of {0, 1}1 with its usual measure, and hb i<1 the canonical
2
) and V = {v : v R 1 , { : v() 6= 0} is
independent family of elements of measure 21 in A. Set
R U = LR (A,
countable}. Dene T : U V by setting T u() = 2 b u u for < 1 , u U . Show that (i) hT un inN
order*-converges to 0 in V whenever hun inN 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 (denition: 241Ye) such that U separates the
points of U , and hun inN 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 inN has a sub-subsequence which is order*convergent to 0.
(f ) Let U be an Archimedean Riesz space and A a weakly (, )-distributive Dedekind complete Boolean
algebra. Suppose that T : U L0 = L0 (A) is a linear operator such that h|T un |inN order*-converges to 0
0
in L0 whenever hun inN 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, ) is a probability space in which T = PY , {y} = 0 for every y Y . (See 363S.)
Take X = Y and let be counting measure on X; let 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 () L () dened by setting
T f = g whenever f L ()
= (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 (); L ()) and in particular does not satisfy the condition (ii) of 376J.
(h) Give an example of an abstract integral operator T : 2 L1 (), where is Lebesgue measure on
[0, 1], such that hT en inN is not order*-convergent in L1 (), where hen inN 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
remember that it is weakly sequentially compact.)
476
376Yk
(k) Let (X, , ) be a semi-nite measure space. Show that 376S is valid for (X, , ) i (X, , ) has
locally determined negligible sets (213I).
(l) Let (X, , ) be a complete locally determined measure space and (Y, T, ), (Z, , ) two -nite
measure spaces. Suppose that , are extended Fatou norms on L0 (), L0 () respectively, and that T :
L1 () L is an abstract integral operator, with corresponding kernel k L0 (), while S L (L ; L ),
so that ST : L1 () L is an abstract integral operator (376Xc); let k L0 ( ) be the corresponding
kernel. For x X set vx = kx when this is dened in L , as in 376S, and similarly take wx = kx L .
Show that Svx = wx for almost every x X.
P
(m) Set k(m, n) = 1/(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(m , n) = 0 for all distinct m, m Z. (Hint: look at
2
2
k(m, n) k(m , n).) (iii) Show
P that there is a norm-preserving linear operator2T from = (Z) to itself
given by setting (T u)(n) = m= k(m, n)u(m) for every n. (iv) Show that T 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 nite-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, )
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 verications, as in parts (a) and (b), while part (i) is just a manoeuvre to show that it
doesnt matter if A and B arent 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 identication of L1 (C, )
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 (, )-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 nd 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 signicantly dierent. In 376H(iii) we could easily have |un | = u for
every n; for instance, if un = 2an 1 for some stochastically independent sequence han inN of elements
of measure 21 in a probability algebra (272Yd).
If you have studied compact linear operators between Banach spaces (denition: 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 inN 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
376 Notes
Kernel operators
477
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 reexive 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 (, )-distributive. For more information on the representation of operators see
Dunford & Schwartz 57, particularly Table VI in the notes to Chapter VI.
As soon as we leave formulations in terms of the spaces L0 (A) and their subspaces, and return to the
original conception of a kernel operator in terms of integrating functions against sections of a kernel, we
are necessarily involved
in the pathology
of Fubinis 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 -nite, as in 376J. In 376S I take the trouble
to oer 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 -nite case. Note that
this is one of the many contexts in which anything we can prove about probability spaces will be true of all
-nite 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 dont 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 (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 nd ourselves with an operator from L 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 contact with each other. My own
account 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.
478
Automorphisms
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 elementary results on automorphisms of general Boolean algebras
(381), introducing denitions for the rest of the chapter. In 382 I describe a general theorem on the
expression of an automorphism as the product of involutions (382M), with a description of the normal
subgroups of certain groups of automorphisms (382R). Applications of these ideas to measure algebras are
in 383. I continue with a discussion of circumstances under which these automorphism groups determine
the underlying algebras and/or have few outer automorphisms (384).
One of the outstanding open problems of the subject is the isomorphism problem, the classication of
automorphisms of measure algebras up to conjugacy in the automorphism group. I oer two sections on
entropy, the most important numerical invariant enabling us to distinguish some non-conjugate automorphisms (385-386). For Bernoulli shifts on the Lebesgue measure algebra (385Q-385S), the isomorphism
problem is solved by Ornsteins theorem; I present a complete proof of this theorem in 386-387. Finally,
in 388, I give Dyes theorem, describing the full subgroups generated by single automorphisms of measure
algebras of countable Maharam type.
381E
479
(f ) If A is a Boolean algebra, a subgroup G of Aut A is countably full if whenever hai iiI is a countable
partition of unity in A, hi iiI is a family in G, and Aut A is such that d = i d whenever i I and
d ai , then G.
(g) If A is a Boolean algebra, a A and : A A is a Boolean homomorphism, I say that is recurrent
on a if for every non-zero b a there is a k 1 such that a k b 6= 0. If Aut A and and 1 are
both recurrent on a, I say that is doubly recurrent on a.
381C Before setting out to explore the concepts just listed, I give a fundamental result on piecing
automorphisms together from fragments.
Lemma Let A be a Boolean algebra, and hai iiI , hbi iiI two partitions of unity in A. Assume
either that I is nite
or that I is countable and A is Dedekind -complete
or that A is Dedekind complete.
Suppose that for each i I we have an isomorphism i : Aai Abi between the corresponding principal
ideals. The there is a unique Aut A such that d = i d whenever i I and d ai .
Q
Q
proof By 315F, we may identify A with each of the products iI Aai , iI Abi ; now corresponds to
the isomorphism between the two products induced by the i .
381D Corollary Let A be a homogeneous Boolean algebra, and A, B two partitions of unity in A,
neither containing 0. Let : A B be a bijection. Suppose
either that A, B are nite
or that A, B are countable and A is Dedekind -complete
or that A is Dedekind complete.
Then there is an automorphism of A extending .
proof For every a A, the principal ideals Aa , Aa are isomorphic to the whole algebra A, and therefore
to each other; let a : Aa Aa be an isomorphism. Now apply 381C.
381E Lemma Let A be a Boolean algebra, and , , : A A Boolean homomorphisms of which is
injective.
(a) If a A supports then a = a and d a for every d a.
(b) If a A supports both and then it supports .
(c) Let A be the set of elements of A supporting . Then A is non-empty and closed under ; also b A
whenever b a A. If is order-continuous, then inf B A whenever B A has an inmum in A.
(d) If a A supports , then a supports .
(e) If commutes with , and a A is such that a supports , then a supports .
(f) If is supported by a and is supported by b, where a b = 0, then = .
(g) For any n 1 and a A, a supports n i a supports n . Consequently (supp n ) = supp n if
n
has a support.
(h) If Aut A and a A supports , then a supports 1 .
(i) If Aut A and a A, then a supports i d a whenever d d = 0 i d d 6= 0 whenever
0 6= d 1 \ a.
(j) If Aut A and a A supports , then a supports 1 .
(k) If a A supports , and 1 , 2 Aut A agree on Aa , then 1 11 = 2 21 .
proof (a) (1 \ a) = 1 \ a, so a = a, and if d
a then d a = a.
(b) If d a = 0 then d = d = d so d = d.
(c) Of course 1 A, because 0 = 0; and it is also obvious that if b a A then b A. If a, b A and
d a b = 0, then d = (d \ a) (d \ b) = d. If is order-continuous, B A is non-empty and c = inf B
is dened in A, then for any d 1 \ c we have
d = d \ c = supbB d \ b,
480
381E
Automorphisms
and
d = supbB (d \ b) = supbB d \ b = d.
So in this case c supports .
(d) If d a = 0 then d a = d a = 0, so d = d and (because is injective) d = d.
(e) If d a = 0 then d a = 0, so d = d = d and d = d.
d
(f ) If d a then d a and d = d so d = d = d; if d
1 \ (a b) then d = d = d.
b then d = d = d; and if
d d = 0,
a.
) If d
(
a whenever d d = 0, and 0 6= d
1 \ a, then of course d d 6= 0.
a, 21 1 d = 21 2 d = d, so 21 1 is supported by 1 \ a. By (f), 21 1 = 21 1 , so
1 11 = 2 21 1 11 = 2 21 1 11 = 2 21 .
381F Corollary If A is a Dedekind complete Boolean algebra, then every order-continuous Boolean
homomorphism : A A has a support.
proof By 381Ec, inf{a : a A supports } is the support of .
381G Corollary Let A be a Boolean algebra, and suppose that Aut A has a support e.
(a) e = e.
(b) e = sup{d : d A, d d = 0}.
(c) e is the support of 1 .
(d) For any Aut A, e is the support of 1 .
proof (a) 381Ea.
(b) By 381Ei, d
d d d \ e 6= 0.
e whenever d d = 0.
(c) 381Eh.
(d) By 381Ej, e supports 1 . At the same time, if a A supports 1 , then 1 a supports , so
e 1 a and a e. Thus e is the smallest element of A supporting 1 and is the support of 1 .
381H Proposition Let A be a Dedekind -complete Boolean algebra and : A A an injective Boolean
homomorphism such that n has a support for every n N. Then there is a partition of unity hci i1i in
A such that ci ci for every i and Acn is periodic with period n whenever n is nite and cn 6= 0, while
Ac is aperiodic.
381I
481
proof Set
c1 = 1 \ supp ,
c = inf supp n .
nN
482
Automorphisms
381I
Suppose that G and a A \ {0}. Then there are a non-zero b1 1 a and an n Z such that
c = n c for every c b1 . Set b = b1 , so that 0 =
6 b a; then for any c b,
n c = n 1 c = 1 c.
= bn \ sup bi if n Z \ N;
i>n
S
supnZ an = supnZ bn = sup( nZ Bn ) = 1,
and c = n c for every c an , n Z. Thus satises condition (ii) of (a) and belongs to the countably
full subgroup generated by .
(iii)(i) is trivial.
(ii)(iv) The point is that, for n Z and a A,
c = n c for every c a n c = c for every c a
a supp( n ) = 0.
So (ii) amounts to saying that supnZ 1 \ supp( n ) = 0, that is, that inf nZ supp( n ) = 0.
381J Lemma Let A be a Boolean algebra, and Aut A. Suppose that belongs to the full subgroup
of Aut A generated by .
(a) If c A is such that c = c, then c = c.
(b) If a A supports then it supports .
proof (a) Let G be the set of all Aut A such that c = c. Then G is a subgroup of Aut A containing
. Also G is full. P
P If hai iiI is a partition of unity in A, hi iiI is a family in G, and Aut A is such
that that d = i d whenever d ai , then
c = supiI (c ai ) = supiI i (c ai ) = supiI i c i ai = supiI c i ai = c.
So G; as is arbitrary, G is full. Q
Q So G and c = c, as claimed.
(b) If c a = 0 then c = c so c = c.
381K Lemma Let A be a Dedekind -complete Boolean algebra and : A A a sequentially ordercontinuous Boolean homomorphism.
(a) If a A and a = inf kN supik i a, then a = a .
(b) If a A is such that a supi1 i a, then supik i a = supiN i a for every k N.
381M
483
(313L(c-iii))
= inf sup i a = inf sup i a = inf sup i a = a .
kN ik+1
k1 ik
kN ik
(b) Induce on k. For k = 0 the result is just the hypothesis. For the inductive step to k + 1, because
is sequentially order-continuous, so is k (313Ic), so
sup i a = sup k i a = k (sup i a)
i1
ik+1
i1
iN
ik
dn = a n a \ sup1i<n i a.
If 1 m < n then
dn
n a \ m a,
dm
m a
so dm dn = 0. Also
dm
a,
nm dn a = nm (dn (nm) a) = 0
so
n dn m dm = m ( nm dn dm ) = 0.
Finally, supn1 dn = a supn1 n a = a, because is recurrent on a (using (a)).
484
381M
Automorphisms
It follows that hdn in1 is a partition of unity in Aa . Since h n dn in1 is also a disjoint family in Aa , and
sup n dn = sup( n a a \ sup ni a)
n1
n1
1i<n
1i<n
n1
(b) Induce on n. For n = 0 we can take b0 = a and bi = 0 for i > 0. For the inductive step to n + 1, let
hbi iin be a partition of unity in Aa such that an b = i b for b bi . Then ha1 bi iin and hdk a1 bi ik1,in
are partitions of unity in Aa . If b dk a1 bi , then a b = k b bi , so an+1 b = k+i b. This means that
if we set bj = supk1,in,k+i=j dk a1 bi for j n + 1, hbj ijn+1 will be a partition of unity in Aa , and
an+1 b = j b whenever b bj . So the induction continues.
(c) Induce on n. If n = 1 then b a 1 a so a d = d for every d b and we can take b = b, j = 1.
For the inductive step to n 2, if b i a = 0 for 1 i < n then we can take b = b and j = n. Otherwise,
take the rst i 1 such that b1 = b i a 6= 0. Then a d = i d for every d b1 . Also ni i b1 a, so, by
the inductive hypothesis, there are a non-zero c i b1 and a j such that 1 j n i and ni d = aj d for
every d c. Setting b = i c, we have 0 6= c b and
n d = i ni d = a aj d = aj+1 d
whenever d
b .
(d) Again induce on n. If 1 n < m then a n a = 0 and the result is trivial. If n = m, then
a n a = dn and a d = n d for every d dn , so we can set bm1 = dm . For the inductive step to n > m,
we have
381N
485
a n a = dn
mk<n
sup dk n a
= dn
mk<n
sup dk k (a (nk) a) = dn
sup
mknm
1j(nk)/m
dk k bnk,j
n
,
m
(e) Applying (b) and (d) to and 1 , and using 381L and (a), we see that is doubly recurrent on b
i a is doubly recurrent on b.
In this case, set D = {d : d Ab , b d = (a )b d}. Then D is order-dense in Ab . P
P Take any non-zero
c Ab . Since b supn1 n b, there is an n 1 such that c = c n b \ sup1i<n i b is non-zero.
Next, there is a non-zero d c such that for every m n either d m a or d m a = 0. Enumerate
{m : m n, d m a} in ascending order as (m0 , . . . , mk ) (note that as c a n a, we must have
m0 = 0 and mk = n). Set di = mi d for i k, so that
mi+1 mi di = di+1
d0 = d,
a,
while
j di = mi +j d
1\a
ak d = mk d = n d
b,
mi c
while
ai d = di = mi d
1\b
a c n (a c) \ sup1i<n i (a c) = c a n a \ sup1i<n i a,
a; it follows that b 2 b = (b b) = 0. If c = b 2 b,
2 b b = 2 c = a c a b
is disjoint from b and must be 0. So b, b and 2 b are all disjoint.
(i) By 381C, the formula denes an automorphism
a . Setting d0 = 1 \ a, hdn inN is a partition of unity
in A and
a d = n d for d dn , so
a belongs to the countably full subgroup of Aut A generated by .
486
Automorphisms
381O
381O Lemma Let A be a Boolean algebra and : A A a Boolean homomorphism. Then the following
are equiveridical:
(i) is recurrent on every a A;
(ii) for every non-zero a A there is a k 1 such that a k a 6= 0;
(iii) a = supk1 a k a for every a A.
proof (i)(ii) If (i) is true, and a A \ {0}, then taking b = a in the denition 381Bg we see that there is
a k 1 such that a k a 6= 0.
(ii)(iii) Suppose (ii) is true. ?? If a A is not the supremum of {a k a : k 1}, let b a be non-zero
and disjoint from k a for every k 1. Then b k b = 0 for every k 1, which is impossible. X
X
(iii)(i) Suppose (iii) is true. If 0 6= b a then b = supk1 b k b, so there is certainly some k 1 such
that b k b 6= 0, in which case a k b 6= 0. As b is arbitrary, is recurrent on a; as a is arbitrary, (i) is
true.
Remark The condition recurrent on every a A looks, and is, very restrictive; but it is satised by the
homomorphisms we care about most (386A).
381P Proposition Let A be a Boolean algebra and : A A a Boolean homomorphism which is
recurrent on every a A. Then is aperiodic i A is relatively atomless (denition: 331A) over the
xed-point algebra C = {c : c A, c = c}. In particular, if is ergodic, it is aperiodic i A is atomless.
proof It is elementary to check that C is a subalgebra of A.
(a) Suppose that is not aperiodic. Then there is a least n 1 such that 1 is not the support of n ;
that is, there is a non-zero a A such that n d = d for every d a. Now if 0 6= b a and 1 i < n there
is a non-zero b b such that b i b = 0. P
P We are supposing that the support of i is 1, so there is a
i
i
d b such that d 6= d. If d \ d 6= 0, take b = d \ i d. Otherwise, try b = d \ ni d; then
i b = i d \ n d = i d \ d 6= 0,
so b 6= 0, while b i b d \ n d = 0. Q
Q
We can therefore nd a non-zero b a such that b i b = 0 whenever 1 i < n. Now b is a relative
atom of A over C. P
P If d b, set c = sup0i<n i d. Then c = sup1in i d = c, so c C, while b i d = 0
for 1 i < n, so d = b c. Q
Q Thus a n a witnesses that A is not relatively atomless over C.
(ii) Suppose that A is not relatively atomless over C. Then there is a relative atom a A; as is
recurrent on a, there is a rst n 1 such that a n a 6= 0. Then n b = b for every b a n a. P
P Because a
is a relative atom over C, there is a c C such that b = a c. Now n b = n a c b. Set b1 = sup0i<n i b;
then b1 = sup1in i b b1 . So b1 = b1 , by (i), and n b supi<n i b. Next,
n b i b = i ( ni b b)
for 0 < i < n, so n b
i ( ni a a) = 0
b and n b = b. Q
Q Thus b witnesses that is not aperiodic.
(c) Finally, if is ergodic, then C = {0, 1}, so that relatively atomless over C becomes atomless.
381Q As far as possible I will express the ideas of this chapter in pure Boolean algebra terms, without
shifting to measure spaces or Stone spaces. However there is a crucial argument in 382 for which the Stone
representation is an invaluable aid, and anyone studying the subject has to be able to use it.
Proposition Let A be a Boolean algebra and Z its Stone space. For a A let b
a be the corresponding
open-and-closed subset of Z; recall that b
a can be identied with the Stone space of Aa (312S). For a Boolean
a] for every a A
homomorphism : A A let f : Z Z be the continuous function such that
ca = f1 [b
(312P).
(a) If a, b A and : Aa Ab is a Boolean homomorphism represented by a continuous function
g : bb b
a, then Aut A agrees with on Aa i f agrees with g on bb.
381Q
487
nZ
n1
fn [b
a].
(e) Suppose that Aut A is recurrent on a A, and that a is the induced automorphism on Aut Aa
autohomeomorphism
of b
a. For k 1, set Gk = {z : z b
a, f k (z) b
a,
(381M). Let fa be the corresponding
S
S
i
k
f (z)
/b
a for 1 i < k}. Then k1 Gk = b
a k1 f [b
a] is a dense open subset of b
a and fa (z) = fk (z)
whenever k 1 and z Gk .
proof Recall that f = f f for all Boolean homomorphisms , : A A (312Q).
(a) The point is that {db : d a} is a base for the Hausdor topology of of b
a. So if g 6= f bb, there are a
b
b
b
z b such that f (z) 6= g(z) and a d a such that g(z) d and f (z)
/ d. In this case,
b \ f 1 [d]
b = d
c \ d,
c
z g 1 [d]
a, and = Aa .
b = g 1 [d]
b = d
c
c = f 1 [d]
d
(b)
a A supports agrees with the identity on 1 \ a
\
f (z) = z for every z (1
\ a) = Z \ b
a
b
a {z : f (z) 6= z}
b
a {z : f (z) 6= z}.
So the smallest such a, if there is one, must have b
a = {z : f (z) 6= z}.
(c) If G, let han inZ be a partition of unity in A such that b = n b whenever n Z and b
dn ((a) above). As sup
(381Ia). Then g(z) = fn (z) whenever z a
nZ an = 1 in A,
S
S
n
d
nZ an
nZ int{x : f (z) = f (z)}
an
is dense (313Ca).
S
If nZ int{x : f (z) = fn (z)} is dense, it is a dense open subset of {z : f (z) {fn (z) : n Z}, so the
latter is comeager.
If {z : f (z) {fn (z) : n Z} is comeager,
set Fn = {z : S
f (z) = fn (z)} for each n. Then Fn \ int Fn
S
is nowhere dense for each n, and Z \ nZ Fn is meager, so nZ int Fn is comeager, therefore dense (by
c int Fn 6= and a b A
Baires theorem, 3A3G). If a A is non-zero, there are an n Z such that a
1
n
b
c
such that =
6 b a Fn , in which case 0 6= b a and c = c for every c b. By 381I(b-ii), G.
So the cycle is complete.
488
381Q
Automorphisms
(d)
is recurrent on a whenever 0 6= b
whenever 0 6= b
whenever 0 6= b
b
a
b
a
k1
such that b
a (fk )1 [bb] 6=
a there is a k 1
such that fk [b
a] bb 6=
fk [b
a] is dense in b
a
fk [b
a].
k1
k1
1 c
[
[
k dk = a dk = fa [dk ] = Gk .
Gk is dense in b
a.
381R Cyclic automorphisms I end the section by describing a notation which is often useful.
Definition Let A be a Boolean algebra.
(a) Suppose that a, b are disjoint members of A and that Aut A is such that a = b. I will write
a,
d if d
= d if d
b,
1 \ (a b).
Observe that in this case (if a 6= 0) is an involution, that is, has order 2 in the group Aut A; I will call
such a an exchanging involution, and say that it exchanges a with b.
(b) More generally, if a1 , . . . , an are disjoint elements of A and i Aut A are such that i ai = ai+1 for
each i < n, then I will write
a)
(
a a...
1 1
2 2
n1
11 21
= d if d
1
. . . n1
d
ai ,
if d
an ,
1 \ sup ai .
in
(c) It will occasionally be convenient to use the same notation when each i 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 . The following results are typical.
381Xc
489
= (a b) = (b a) = (b 1 a)
is supported by a b.
1 = (a 1 b)
is another exchanging involution.
(c) If = (a b) and = (c d) are exchanging involutions, and a, b, c, d are all disjoint, then and
(a b)(b c) = (a b c).
But if a c 6= 0 then there is no expression for the product in this language. Secondly, of course, we must
..
.
an ), that a1 , . . . , an are disjoint and that
be scrupulous in checking, at every use of the notation (
a1
1
i ai = ai+1 for i < n. Thirdly, a signicant problem can arise if the automorphisms involved dont match.
Consider for instance the product
= (a b)(a b).
Then we have d = 1 d if d a, 1 d if d b; is not necessarily expressible as a product of disjoint
cycles. Clearly there are indenitely complex variations possible on this theme. A possible formal expression
of a sucient condition to avoid these diculties is the following. Restrict yourself to calculations involving a
xed list a1 , . . . , an of disjoint elements of A for which you can describe a family of isomorphisms ij : Aai
Aaj such that ii is always the identity on Aai , jk ij = ik for all i, j, k, and whenever ai aj appears in
a cycle of the calculation, then agrees with ij on Aai . Of course this would be intolerably unwieldy if it
were really necessary to exhibit all the ij every time. I believe however that it is usually easy enough to
form a mental picture of the actions of the isomorphisms involved suciently clear to oer condence that
such ij 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.
381X Basic exercises (a) Let X be a set and an algebra of subsets of X containing all singleton
sets. Show that Aut can be identied with the group of bijections f : X X such that f [E] and f 1 [E]
belong to for every E .
(b) Let A and B be Boolean algebras, and hai iiI , hbi iiI partitions of unity in A, B respectively. Assume
either that I is nite or that I is countable and B is Dedekind -complete or that B is Dedekind complete.
Suppose that for each i I we have a Boolean homomorphism i : Aai Bbi . (i) Show that there is a
Boolean homomorphism : A B extending every i . (ii) Show that is injective i every i is. (iii)
Show that is surjective i every i is.
(c) Let A be a Boolean algebra. Show that if Aut A and k Z \ {0}, then is aperiodic i k is.
490
Automorphisms
381Xd
(d) In 381H, show that the family hci i1i is uniquely determined.
> (e) Let A be a Boolean algebra and G any subgroup of Aut A. Let H be the set of those Aut A
such that for every non-zero a A there are a non-zero b a and a G such that c = c for every c b.
Show that H is a full subgroup of Aut A, the smallest full subgroup of A including G.
(f ) Let A be a Dedekind complete Boolean algebra and G any subgroup of A. Show that an element
of Aut A belongs to the full subgroup of Aut A generated by G i there are a partition of unity hai iiI in A
and a family hi iiI in G such that a = i a whenever i I and a ai .
(g) Let A be a Boolean algebra. Let us say that a subgroup G of Aut A is finitely full if whenever hai iiI
is a nite partition of unity in A, hi iiI is a family in G, and Aut A is such that a = i ai whenever
i I and a ai , then G. Show that if , Aut A then belongs to the nitely full subgroup of
Aut A generated by i there are an n N and a partition of unity hai inin in A such that d = i d
whenever |i| n and d ai .
(h) Let A be a Boolean algebra and : A A a Boolean homomorphism which is recurrent on a A.
Show that for any non-zero b a and any n N there is a k n such that a k b 6= 0.
(i) Let A be a Boolean algebra, : A A a Boolean homomorphism, and a A. Show that the following
are equiveridical: (i) is recurrent on every b a; (ii) for every non-zero b a there is an n 1 such that
b n b 6= 0; (iii) b = supn1 b n b for every b a.
(j) Let A be a Boolean algebra, not {0}, and : A A an automorphism with xed-point subalgebra C.
Show that is periodic, with period n 1, i Ac has order n in the group Aut Ac whenever c C \ {0}.
Show that is aperiodic i Ac has innite order in the group Aut Ac whenever c C \ {0}.
(k) For a Boolean algebra A, a Boolean homomorphism : A A is nowhere aperiodic if inf{a : a A,
a supports n for some n 1} = 0. Show that if A is Dedekind -complete and Aut A is nowhere
aperiodic and doubly recurrent on a A, then the induced automorphism a is nowhere aperiodic.
(l) Let A be a Dedekind -complete Boolean algebra, Aut A an automorphism and C the xed-point
subalgebra of . Suppose that is doubly recurrent on a A and that a is the induced automorphism on
Aa . Show that the xed-point subalgebra of a is {c a : c C}; in particular, if is ergodic, so is a .
(m) Let A be a Boolean algebra with Stone space Z, and : A A a Boolean homomorphism corresponding to f : Z Z. (i) Show that is periodic, with period n 1, i f n (z) = z for every z Z and
{z : f i (z) = z} is nowhere dense whenever 1 i < n. (ii) Show that is aperiodic i {z : f n (z) = z,
f n (w) 6= z for every w 6= z} is nowhere dense for every n 1.
> (n) Let (X, , ) be a countably separated measure space (denition: 343D), and A its measure algebra,
f : X X an inverse-measure-preserving function and : A A the induced homomorphism (343A). (i)
Show that the support of is {x : x X, f (x) 6= x} . (ii) Show that is periodic, with period n 1, i
f n (x) = x for almost every x and {x : f i (x) = x} is negligible for 1 i < n.
(o) Let (X, , ) be a localizable measure space, with measure algebra (A,
). Suppose that and are
automorphisms of A, and that is represented by a measure space automorphism f : X X. Show that
the following are equiveridical: (i) belongs to the full subgroup of Aut A generated by ; (ii) there is a
function g : X X, representing , such that g(x) {f n (x) : n Z} for every x X. (Hint: for (ii)(i),
consider measurable envelopes of sets F g[An ], where An = {x : g(x) = f n (x)} and F < .)
> (p) Let (X, , ) be a measure space, A its measure algebra, f : X X a measure space automorphism,
and the
automorphism
of A. (i) Show that if E then is doubly recurrent on a = E
S corresponding
S
n
n
i E \ n1 f [E] and E \ n1 f [E] are negligible. (ii) Show that in this case there is a measurable
F E such that E \ F is negligible and {n : n Z, f n (x) F } is unbounded above and below in Z for
every x F . (iii) For x F let k(x) = min{n : n 1, f n (x) F }. Show that x 7 f k(x) (x) : F F
represents the induced automorphism a on the principal ideal Aa .
382B
Factorization of automorphisms
491
381Y 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 : A A
is an order-continuous Boolean homomorphism, and there is a partition of unity hai iiI in A such that for
every i I there is a i G such that a = i a for every a ai , then G. Show that if and
are order-continuous Boolean homomorphisms from A to itself, then the following are equiveridical: (i)
belongs to the full semigroup generated by ; (ii) for every non-zero a A there are a non-zero b a and an
n N such that d = n d for every d b; (iii) there is a partition of unity han inN in A such that a = n a
for every n N, a an .
(b) Give an example to show that the word injective in the statement of 381H is essential.
(c) Give an example of a Dedekind -complete Boolean algebra Aut A and an automorphism of A such
that the countably full subgroup generated by is not full.
(d) Let A be a Dedekind complete Boolean algebra, and let G be the countably full subgroup of Aut A
generated by a subset A of Aut A. Show that if either A is countable or A is ccc, then G is full.
381 Notes and comments There are no long individual proofs in this section, and in so far as there is any
delicacy in the arguments it is as often as not because (as in 381E) I am taking facts which are easy to prove
for automorphisms of Dedekind complete algebras and separating out the parts which happen to be true in
greater generality. However the parts are numerous enough for the sum to be not entirely predictable. The
most important ideas are surely in 381M-381N.
In 381Q I give indications, including the minimum necessary for an application in the next section, of
how to express the concepts here in terms of continuous functions on Stone spaces. When we come, in 383
and onwards, to look specically at measure algebras, many of our homomorphisms will be derived from
inverse-measure-preserving functions, and the results will be more eective if we can display them in terms
of functions on measure spaces. Some appropriate translations are in 381Xn-381Xp. But these I will avoid
in the proofs of the main theorems because not all automorphisms of measure algebras can be represented
by automorphisms of the measure spaces we start from (343Jc). Of course Lebesgue measure is dierent, in
ways explored in 344, and classical ergodic theory has not needed to make a clear distinction here. One of
my purposes in this volume is to set out a framework in which transformations of measure spaces take their
proper place as an inspiration for the theory rather than a foundation.
382B Lemma Let A be a Boolean algebra and Aut A. If every power of has a separator and n
is the identity, where n 1, then has a transversal.
492
382B
Automorphisms
proof (a) For 0 j < n let aj A be a separator for j . Let B be the subalgebra of A generated by
A = { i aj : 0 i, j < n}. Because [A] = A, [B] = B. (The set {a : a B, a B, 1 a B} is a
subalgebra of A including A, so must be B.) Because A is nite, so is B; let B be the set of atoms of B.
Then B is a bijection from the nite set B to itself.
(b) Let C be the set of orbits of B, that is, the family of sets of the form { k b : k Z} for b B.
If b C C, set m = #(C); then d = m d for every d b. P
P If m = n this is trivial. Otherwise, b is
either disjoint from, or included in, i am whenever 0 i < n, and therefore for every i Z. But we have
am m am = 0, so i am i+m am = 0 for every i, and b i am must be 0. By the other clause in the
denition of separator, m d = d for every d b. Q
Q
(c) For each C C of B, choose bc C. Set c = supCC bC . Then c is a transversal for . P
P If C C,
we have n bC = bC , so kC = #(C) is a factor of n. Now
S
sup0k<n k c = supCC,0k<n k bC = supCC sup C = sup( C) = sup B = 1.
e.
proof (i)(ii) Suppose that a is a separator for . Set a+ = supn1 n a, a = supn1 n a; we are
supposing that a a = 0 and that a a+ a supports . For n N set an = n a \ sup0i<n i a, so that
han inN is disjoint and has supremum a a+ . Set b1 = supnN a2n \ 1 a. Since a 1 a = 1 (a a) = 0,
a b1 a a+ . For any n N,
(a2n \ 1 a) = 2n+1 a \ (a sup1i2n i a) = a2n+1 ,
so b1 b1 = 0. Note that b1
Set c = a \ a+ . Then
a+ , while a+ \ 1 a b1 b1 .
i c j c = j (c ij c)
j (a \ ij a+ )
j (a \ ij ji a) = 0
n1
n1
i>n
n1
= sup n a \ (a a+ ) = a \ (a a+ ).
n1
i<n
382E
493
Factorization of automorphisms
If k 1 and i 0 then
k c i a = k (c i+k a) k (c a+ ) = 0;
k1
j,k1
0 0 (sup 2k1 c b1 ) 0 = 0.
k1
Since
b b 1 b b1 b1 1 a sup n c
n1
aa
(a
\ (a a
)) = a a+ a ,
Next,
sup k a =
kZ
sup
nN,kZ
sup
nN,kZ
cn k an \ k1 an
cn k+1 an \ k an = sup cn bn
nN
nN
nN
(b c) = (1 \ (a b e)) = 1 \ (a b e) = 1 \ (a c e) = a = a a .
(iv)(ii) If a , a , b , b , c, e witness (iv), then a = a a witnesses (ii).
382E Corollary (a) If A is a Dedekind -complete Boolean algebra and Aut A has a separator, then
has a support.
(b) If A is a Dedekind complete Boolean algebra then every Aut A has a separator.
proof (a) Taking a A such that a a = 0 and e = a a 2 a supports , we see that e must actually
be the support of (381Ei).
(b) If A is Dedekind complete and Aut A, let P be the set {d : d A, d d = 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 a = sup Q,
494
382E
Automorphisms
which is dened because A is Dedekind complete; then a = sup [Q] (since , being an automorphism, is
surely order-continuous). If d1 , d2 Q, there is a d Q such that d1 d2 d, so d1 d2 d d = 0. By
313Bc, a a = 0. This means that a P and is an upper bound for Q in P . As Q is arbitrary, Zorns
Lemma tells us that P has a maximal element. Q
Q
Let b P be maximal. Then b b = 0. Set e = b a 1 b. ?? If e does not support , let d 1 \ e
be such that d d = 0 (381Ei). Then d b d e = 0, while also b d ( 1 b d) (e d) = 0; so
(b d) (b d) = 0, and b b d P , which is impossible. X
X So if we set a = 1 b we have a witness of
382D(ii), and has a separator.
Remark 382Eb and 382D(i)(ii) together amount to Frolks theorem (Frolik 68).
382F Corollary Let A be a Dedekind complete Boolean algebra.
(a) Every involution in Aut A is an exchanging involution.
(b) If Aut A is periodic with period n 2, there is an a A such that (a, a, 2 a, . . . , n1 a)
a1 a2.
.
.
is a partition of unity in A; that is (in the language of 381R) is of the form (
an ) where
(a1 , . . . , an ) is a partition of unity in A.
proof (a) By 382Eb, every involution has a separator; now use 382C.
(b) Again because every automorphism has a separator, 382C tells us that has a transversal a. In this
case, a k a must be disjoint from the support of k for every k Z; since supp k = 1 for 0 < k < n,
a k a = 0 for 0 < k < n; of course it follows that i a j a = i (a ji a) = 0 if 0 i < j < n. So
a, a, . . . , n1 a are disjoint; since sup0i<n i a = supiZ i a = 1, they constitute a partition of unity.
382G Lemma Let A be a Dedekind -complete Boolean algebra and Aut A.
(a) Suppose that han inN is a family in A such that an = an and Aan has a transversal for every n.
Set a = supnN an ; then a = a and Aa has a transversal.
(b) If a is a transversal for it is a transversal for 1 .
(c) Suppose that a A. Set
a = supnZ ( n a \ supi>n i a),
proof (a) Of course a = supnN an = a, so we can speak of Aa . For each n N, let bn be a transversal
P Of course b Aa . Now
for Aan . Set b = supnN (bn \ supi<n ai ). Then b is a transversal for Aa . P
sup k b = sup sup( k bn \ sup k ai ) = sup sup( k bn \ sup ai )
i<n
kZ nN
kZ
i<n
nN kZ
nN kZ
nN
i<n
nN
m,nN
j<n
m,nN
j<n
nN
i<m
i>n
i>n+1
nZ
i>n
382H
495
Factorization of automorphisms
m (supiZ i bi nm a) = 0.
nj
( i am j an ) 6= 0,
For any n 1,
h 2j j a0 ijZ = h j a0 ijZ ,
h 12j j a0 ijZ = h 1j a0 ijZ
j a0
or if 0 j < n and d
2 d = 12j d if j Z and d
j a0
or if 0 j < n and d
j an ,
j an .
Note that if n 1 and k Z is arbitrary, then we have k an = j an where 0 j < n and j k mod n, so
if d k an then
1 d = 2j d = 2k d,
because n d = d. So if d
2 d = 12j d = 12k d
j an and
2 1 d = 12(j) 2j d = d.
Because supnN,jZ j an = 1, 2 1 = . Of course both 1 and 2 belong to the countably full subgroup
generated by . Next, 1 exchanges
496
382H
Automorphisms
sup j a0
j an ,
sup
j1
n2
0<j(n1)/2
sup j a0
j an ,
sup
j1
n2
(n1)/2j<0
sup
j an ,
n2
1jn/2
sup j a0 sup an
j0
n2
sup
j an ,
n2
n/2<j0
so it too is either the identity or an exchanging involution. Thus we have a factorization of the desired type.
382I Lemma Let A be a Dedekind -complete Boolean algebra, and G a countably full subgroup of
Aut A such that every member of G has a separator.
(a) Every member of G has a support.
(b) Suppose G and n 1 are such that n is the identity. Then has a transversal.
(c) Let G, and set e = inf n1 supp( n ). Then A1\e has a transversal.
(d) If e A is such that e = e for every G, then { Ae : G} is a countably full subgroup of
Aut Ae , and Ae has a separator for every G.
proof (a) 382Ea.
(b) Induce on n. If n = 1 then 1 is a transversal for . For the inductive step to n > 1, let a A be
such that a a = 0 and b = b whenever b i a = 0 for every i Z. Let B be the (nite) subalgebra of
A generated by { i a : 0 i < n}. Then n a = a B, so {b : b B} is a subalgebra of A containing i a
whenever i < n, and includes B; thus b B for every b B. As is injective, B Aut B. Let E be
the set of atoms of B; then E is a bijection from E to itself.
Let C E be an orbit of . Then (sup C) = sup C, and Asup C has a transversal. P
P Take e C,
k = #(C). Then i e C \ {e}, so e i e = 0, whenever 1 i < k. As n is the identity, k is a factor
of n. If k = 1, then e itself is a transversal for Asup C = Ae . If k > 1, dene Aut A by setting
d = k (e d) (d \ e) for every d A. Then G, because G is countably full, and n/k is the identity.
By the inductive hypothesis, has a transversal c A. There is some m Z such that e = e m c 6= 0.
Now
supiZ ki e = supiZ i e = supiZ (e m+i c) = e supiZ i c = e,
so
supjZ j e = sup0j<k j (supiZ ki e ) = sup0j<k j e = sup C.
Also, if 0 j < k and i Z and
0 6= d
e ki+j e
e ki+j e = e j e,
382K
497
Factorization of automorphisms
i1
j0
i1
j0
(381Nb)
= sup i an = 1.
i1
i1
j0
a
) = sup i an = 1,
sup i (sup aj
n n+1
i1
j0
i1
for every n N, writing en Aut Aen for the induced automorphism. Let G be the countably full subgroup
of Aut A generated by . Then there is a G such that is either the identity or an exchanging involution
and inf n1 supp()n = 0.
n G for the corresponding automorphism of A, as in
proof (a) For each n N, write n for en and
381Ni. Set
498
382K
Automorphisms
un = n1 en+1 ,
un = n en+1 .
en em+1
um um
is disjoint from
1
m
(em+1 ) m (em+1 ). Q
Q
n d if n N, d
un ,
= n1 n+1 n1 d =
n1
n+1
n1 d if n N, d
= d if d sup(un un ) = 0;
un ,
nN
is the identity and 1 exchanges supnN un with supnN un , so is either the identity or an
1 G and
exchanging involution. Set c0 = inf k1 supp(1 )k and c1 = supiZ i c0 , so that c1 = c1 and 1 c1 = c1
(381J).
21
vl = l c0 \ supl<il i c0 ,
vl = l c0 \ supli<l i c0 .
d if d
so that 2 G,
22
supl1 vl vl .
1 \ c1
c1 .
It is easy to check that is either the identity or an exchanging involution. Set c2 = inf n1 supp()n .
(f ) I wish to show that c2 = 0. At this point I see no practical alternative to moving to the Stone
space representation. Let Z be the Stone space of A, and f , g1 , g2 and g the autohomeomorphisms of Z
corresponding to , 1 , 2 and ; write b
a Z for the open-and-closed set corresponding to a A. For each
n N, let fn : ec
e
c
be
the
autohomeomorphism
corresponding to en . Since
n
n
k
supp = 1 for every k 1,
supik i en = supik i en = 1 for every n N, k Z (381L),
c0 = inf k1 supp(1 )k ,
c1 = supiZ i c0 ,
382K
Factorization of automorphisms
499
(f ) Fix x X cb1 for the time being. Because f k (x) 6= x for any k 1, the map i 7 f i (x) : Z X
is injective. Because gk (f i (z)) {f i+j (z) : j Z} for every i Z and both k {1, 2}, we can dene g1x ,
g2x : Z Z by saying that gkx (i) = j if gk (f i (x)) = f j (x). Similarly, f is represented on {f i (x) : i Z} by
s, where s(i) = i + 1 for every i Z.
(i) For n N, set
En = {i : i Z, f i (x) ec
n },
cn }, Un = {i : f i (x) u
cn }.
Un = {i : f i (x) u
S
Because x ik f i [c
en ] ik f i [c
en ] for every k, En is unbounded above and below. If i En , then
i
k+i
fn (f (x)) = f (x) where k 1 is the rst such that f k+i (x) ec
n (381Qe), that is, such that k + i En .
Turning this round, fn1 (f i (x)) = f j (x) where j is the greatest member of En less than i. In particular,
i Un i i is the next point of En above a point of En+1 , and i Un i i is the next point of En below
a point of En+1 . If i Un , then fn1 f i (x) = f j (x) where j En+1 is the next point of En below i, and
fn+1 fn1 f i (x) = f k (x) where k is the next point of En+1 above j. Since g1 must agree with fn1 fn+1 fn1 on
cn (381Qa), g1 f i (x) = fn1 fn+1 fn1 fi (x) = f l (x) where l is the next point of En below f k (x). This means
u
that g1x exchanges pairs i < l exactly when i, l En are the rst and last points in En ]j, k[ where j, k
are successive points of En+1 . In this case, there is no point of En+1 in the interval [i, l]. Accordingly, if
i < l and g1x exchanges i and l and either i or l is in ]i, l[, we must have i , l Em for some m < n; and
as the interval [i , l ] cannot meet Em+1 En , it is included in ]i, l[. Thus g1x xes ]i, l[ in the sense that if
i < i < l then g1x (i ) = l for some l ]i, l[. It follows that g1x s xes [i, l[. In this case, of course, every point
of [i, l[ must be xed by some power of g1x s.
The following diagram attempts to show how g1x links pairs of integers. The points of En , as n increases,
are shown as progressively multiplied circles.
S
500
382K
Automorphisms
S
Because X does not meet cb1 \ iZ f i [cb0 ], C0 is not empty. Now C0 has no greatest member. P
P Let
j0 C0 . Then j0
/ [i, l[ for any pair i, l exchanged by g1x . If j0 + 1 C0 we can stop. Otherwise, there are
/ C0 there are i1 , l1 exchanged by g1x such that
i0 , l0 exchanged by g1x such that i0 j0 + 1 < l0 . ?? If l0
i1 l0 < l1 . But in this case i1 j0 < l1 . X
X Thus j0 < l0 C0 and j0 cannot be the greatest member of
C0 . Q
Q
Similarly, C0 has no least member. P
P If j0 C0 but j0 1
/ C0 , take i0 , l0 exchanged by g1x such
that i0 j0 1 < l0 . ?? If i0 1
/ C0 , take i1 , l1 exchanged by C0 such that i1 i0 1 < l1 ; then
i1 j0 = l0 < l1 . X
X So i0 1 is a member of C0 less than j0 . Q
Q
Thus C0 is unbounded above and below.
(iii) For l 1,
S
vbl = f l [cb0 ] \ li<l f i [cb0 ],
so setting
Vl = {i : f i (x) vbl },
we see that
S
i
l
vc
l<il f [cb0 ];
l = f [cb0 ] \
Vl = {i : f i (x) vc
l },
Vl = {i + l : i C0 , i + j
/ C0 if l j < l},
Vl = {i : i l C0 , i + j
/ C0 if l < j l};
kj1
,
2
then
if 1 l l0 we have exactly one i Vl [j, k] and exactly one i Vl [j, k] and i < i , while if l > l0
then neither Vl nor Vl meets [j, k].
} = V (V + j) \
Dlj
= {i : f i (x) dc
l
l
lj
= {i : i
Vl ,
Vl + i
1i<j
i j = max(Vl ], i[},
} = V (V + j) \
Dlj
= {i : f i (x) dc
l
l
lj
= {i : i
Vl ,
Vl + i
1i<j
Dlj = {i : f i (x) dc
lj } = Dlj (Dlj + j).
that is, whenever i + j Dlj . This means that g2 exchanges pairs i < i exactly when, for some l, i is the
greatest member of Vl less than i and i is the least member of Vl greater than i . Since X does not meet
S
x
\
supp
2 \ l1 (vbl vc
l ), g2 does not move any other i.
But, starting from any l 1 and i Vl , let i be the greatest element of Vl less than i. Then i l and
i + l belong to C0 , and if k, k are any successive members of C0 such that i < k < k < i then there is
no member of Vl in [k, k ] and therefore no member of Vl in [k, k ]. So i is the least member of Vl greater
than i , and g2x (i ) = i . Similarly, every member of every Vl is moved by g2x .
382M
Factorization of automorphisms
501
At the same time we see that if i Vl and i Vl are exchanged by g2x , and m > l, then there can be
no interval of C0 of length 2m + 1 or greater between i and i , so there is no point of Vm Vm in [i , i ].
For the same reason, if m < l then no pair of points in Vm Vm exchanged by g2x can bracket either i or
i . So g2x leaves the interval [i , i ] invariant. Accordingly g2 s leaves [i , i [ invariant.
The next diagram attempts to illustrate g2x . Members of C0 are shown as multiple circles1 .
of C0 such that k 0 < k . Take l such that k k 2l. Let i be the least member of Vl greater than 0,
and i the greatest member of Vl less than 0; since neither Vl nor Vl meets [k, k ], i and i are exchanged
by g2x , while 0 [i , i [. Q
Q This means that there is a k 1 such that (g2x s)k (0) = 0, that is, (g2 f )k (x) = x.
1
(vi) We know that g agrees with g2 on d
[cb1 ] = cb1 , (gf )k (x) = x.
2 c1 = cb1 . Since x cb1 and f
k
Because X does not meet cb2 {z : (gf ) (x) 6= x for every k 1}, x
/ cb2 .
This is true for every x X cb1 . Since X is dense in Z, cb1 cb2 is empty, that is, c1 c2 = 0.
(g) Since agrees with 1 on A1\c1 , and c1 = c1 , supp()k \ c1 = supp(1 )k \ c1 for every k, and
c2 \ c1 = inf k1 supp(1 )k \ c1
inf k1 supp(1 )k \ c0 = 0.
502
382N
Automorphisms
382N Corollary If A is a Dedekind complete Boolean algebra and G is a full subgroup of Aut A, every
G expressible as the product of at most three involutions all belonging to G and all supported by supp .
proof We may suppose that G is the full subgroup of Aut A generated by . By 382Eb, every member of
G has a separator. By 382M, is the product of at most three involutions all belonging to G; by 381Jb,
they are all supported by supp .
382O 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 G which is
supported by a.
382P 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 support of an exchanging involution.
proof If a A \ {0}, then there is a b such that 0 6= b a. Let : Ab Aa\b be an isomorphism; dene
Aut A to agree with on Ab , with 1 on Aa\b , and with the identity on A1\c . Then is an exchanging
involution with support a.
382Q 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 support of an involution belonging to G.
proof By the denition 382O,
C = {supp : G is an involution}
is order-dense in A. So if a A \ {0} there is a disjoint B C such that sup B = a (313K). For each b B
let b G be an involution with support b. Dene G by setting d = b d for d b B, d = d if
d a = 0; then G is an involution with support a.
382R 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 i it is of the form
{ : G, supp I}
for some ideal I A which is G-invariant, that is, such that a I for every a I, G.
proof (a) I deal with the easy implication rst. Let I A be a G-invariant ideal and set H = { : G,
supp I}. Because the support of the identity automorphism is 0 I, H. If , H and G,
then
supp() supp supp I,
supp( 1 ) = supp I,
supp( 1 ) = (supp ) I
(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, H whenever G is an involution and supp
a};
an
: a0 , . . . , an J};
382S
503
Factorization of automorphisms
involution with supp a; say = (b c) where b c a. By 382Q there is an involution 1 G such that
c = b = c \ c ,
2 = 1 1 1 = (b 1 b )(c 1 1 c ),
3 = (b c ),
= 21 2 1 H,
= 31 3 1 = 31 21 2 1 3 21 1 2 H.
Now
supp(2 1 ) = (supp 2 ) = (b c)
is disjoint from
supp 3 = b c
a,
= 31 21 3 2 1 21 1 2
= 31 21 3 2
= (b c )(b 1 b )(c 1 1 c )(b c )(b 1 b )(c 1 1 c )
= (b c )(b c )
= .
So H. As is arbitrary, a J. Q
Q
Q
a b J, so 2 = (a b ) H and = 1 2 belongs to H. Q
cj = a aj \ supi<j ai , bj = cj , j = (cj bj )
for j n; then every cj belongs to J, so every j belongs to H (by (d)) and = 0 . . . n H. Q
Q
(f ) If G and supp I then H. P
P By 382N, is a product of involutions in G all with supports
included in supp ; by (e), they all belong to H, so also does. Q
Q
(g) We are nearly home. So far we know that I is a G-invariant ideal and that H whenever G,
supp I. On the other hand, supp I for every H. P
P By 382Eb, has a separator; take a , a ,
504
Automorphisms
382S
proof If A is {0} or {0, 1} this is trivial. Otherwise, let H be a normal subgroup of Aut A. Then by 382R
and 382P there is an invariant ideal I of A such that H = { : supp 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 Aut A such that
a = 1 \ a (as in 381D), 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 -complete Boolean
e
pa
nek & Rubin 89, Theorem 5.9b.
algebra; see St
382X Basic exercises (a) Let A be a Boolean algebra and Z its Stone space. Suppose that Aut A
is represented by f : Z Z. For z Z, write Orb (z) = {f n (z) : n Z}. (i) Show that a A is a
separator for i f1 [b
a] b
a is empty and {z : Orb (z) b
a} 6= } is dense in {z : f (z) 6= z}. (ii) Show that
a A is a transversal for i {z : Orb (z) b
a 6= } is dense in Z and #(Orb (z) b
a) 1 for every z.
(b) Show that the word aperiodic, in the statement of Lemma 382K, is redundant.
> (c) Let X be any set. Show that any automorphism of the Boolean algebra PX is expressible as a
product of at most two involutions.
>(d) (Miller 04) Let X be a set and a -algebra of subsets of X. Suppose that (X, ) is countably
separated in the sense that there is a countable subset of separating the points of X (cf. 343D). Let G
be the group of bijections f : X X such that = {f 1 [E] : E }. Show that every automorphism of
the Boolean algebra has a separator, so that every member of G is expressible as the product of at most
three involutions belonging to G.
(e) Recall that in any group G, a commutator in G is an element of the form ghg 1 h1 where g,
h G. Show that if A is a Dedekind complete Boolean algebra and G is a full 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 three commutators.
(f ) 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.
(g) Let A be a Dedekind complete Boolean algebra, and suppose that Aut A has many involutions. Show
that if H Aut A then every member of H is expressible as the product of at most three involutions
belonging to H.
(h) 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.
(i) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions.
Show that if H is the normal subgroup of G generated by a nite subset of G, then it is the normal subgroup
generated by a single involution.
(j) 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 G such that every member of G is expressible as a product of
conjugates of in G (ii) any proper normal subgroup of G is included in a maximal proper normal subgroup
of G.
(k) Let A be an innite Dedekind -complete Boolean algebra. Show that if : A A is an ergodic
automorphism it has no transversal.
(l) Show that if A is a Dedekind -complete Boolean algebra with countable Maharam type (denition:
331F), then every Aut A has a separator. (Hint: show that if b A, then {a : a a b} is an
order-closed subalgebra.)
382 Notes
Factorization of automorphisms
505
382Y Further exercises (a) Find a Dedekind -complete Boolean algebra with an involution which is
not an exchanging involution.
(b) Let B be the algebra of Borel subsets of R. Show that Aut B has exactly three proper normal
subgroups. (Hint: re-work the proof of 382R, paying particular attention to calls on Lemma 382Q. You will
need to know that if E B is uncountable then the subspace -algebra on E is isomorphic to B; see 424
in Volume 4.)
(c) Find a Dedekind -complete Boolean algebra A with an automorphism which cannot be expressed
either as a product of nitely many involutions in Aut A, nor as a product of nitely many commutators in
Aut A. (This seems to require a certain amount of ingenuity.)
382 Notes and comments The ideas of 382A and 382G-382N are adapted from Miller 04, and (most
conspicuously in part (f) of the proof of 382K) betray their origin in a study of Borel automorphisms of R (see
382Xd). The magic number of three involutions appears in Ryzhikov 93 and Truss 89. The idea of the
method presented here is to shift from a separator to a transversal. Since there are many automorphisms
without transversals (382Xk), something quite surprising has to happen. The diagrams in the proof of 382K
are supposed to show the two steps involved in the argument. We are trying to draw non-overlapping links
to build a function g x such that every point of Z will belong to a nite orbit of g x s. This must be done by
some uniform, translation-invariant, process based on congurations already present; in particular, we are
not permitted to single out any point of Z as a centre for the construction. The rst attempt is based on the
sequence hEn inN of sets corresponding to the decreasing sequence hen inN . Of course the construction of
such a sequence (382J) requires that there be many separators, which is why these results cannot be applied
to all Boolean algebras, or even to all homogeneous ones. If this rst attempt fails, however, the points not
recurrent under g1x s provide a set C0 with arbitrarily large gaps both to left and to right, from which the
second method can build an adequate family of links.
Of course the search for these factorizations was inspired by the well-known corresponding fact for algebras
PX (382Xc). In those algebras we can use the axiom of choice unscrupulously to pick out a point of each
orbit, thereby forming a transversal in one step without considering separators, and then apply 382H in its
original simple form. Perhaps the principal psychological barrier we need to overcome in 382K is raised in
the phrase x x X cb1 . What I could have said is x an orbit of f meeting cb1 , and order it by the
transitive closure of the relation f ; because the whole point of the subsequent argument is that we do not
have a marker to work from.
This volume is concerned with measure algebras, and all the most important measure algebras are Dedekind complete. I take the trouble to express the ideas down to Theorem 382M in terms of -complete
algebras partly because this is the natural boundary of the arguments given and partly because in Volume
4 I will look at Borel automorphisms, as in 382Xd, and 382M as stated may then be illuminating. But
note that in 382N -completeness is insucient (382Yc). In 382S I allow myself for once to present a result
with a stronger hypothesis than is required for the conclusion; the point being that homogeneous semi-nite
measure algebras are necessarily Dedekind complete (383E), and the arguments for the more general case
do not seem to tell us anything which we can use elsewhere in this treatise.
It is natural to ask whether the number three in 382M is best possible (cf. 382Xc). It seems to be quite
dicult to exhibit an automorphism requiring three involutions; examples may be found in Anzai 51 and
Ornstein & Sheilds 732 .
Just as well-known facts about symmetry groups lead us to the factorization theorem 382M, they suggest
that automorphism groups of Boolean algebras may often have few normal subgroups; and once again we
nd that the form of the theorem changes signicantly. However the root of the phenomenon remains
e
pa
nek & Rubin 89 and
the fact that our groups are multiply transitive. 382O-382S are derived from St
Fathi 78. An obvious question arising from 382S is: does every homogeneous Boolean algebra have a simple
automorphism group? This leads into deep water. As remarked after 382S, every homogeneous Dedekind
-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
2I
506
Automorphisms
382 Notes
I am aware no such construction is known which does not rely on some special axiom outside ordinary set
e
pa
nek & Rubin 89, 5.
theory. See St
383 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 383D, which is a straightforward special case of 382N,
and 383I, corresponding to 382S. I give another example of the use of 382R to describe the normal subgroups
of Aut A (383J). I conclude with an important fact about conjugacy in Aut A and Aut A (383L).
383A Definition Let (A,
) be a measure algebra. I will write Aut A for the set of all measure-preserving
automorphisms of A. This is a group, being a subgroup of the group Aut A of all Boolean automorphisms
of A.
383B Lemma Let (A,
) be a measure algebra, and hai iiI , hbi iiI two partitions of unity in A. Assume
either that I is countable
or that (A,
) is localizable.
Suppose that for each i I we have a measure-preserving isomorphism i : Aai Abi between the
corresponding principal ideals. Then there is a unique Aut A such that c = i c whenever i I and
c ai .
Q
Q
proof (Compare 381C.) By 322K, we may identify A with each of the simple products iI Aai , iI Abi ;
now corresponds to the isomorphism between the two products induced by the i .
383C Corollary If (A,
) is a localizable measure algebra, then, in the language of 381Be, Aut A is a
full subgroup of Aut A.
383D Theorem Let (A,
) be a localizable measure algebra. Then every measure-preserving automorphism of A is expressible as the product of at most three measure-preserving involutions.
proof This is immediate from 383C and 382N.
383E Lemma If (A,
) is a homogeneous semi-nite measure algebra, it is -nite, therefore localizable.
proof If A = {0}, this is trivial. Otherwise there is an a A such that 0 <
a < . The principal ideal
Aa is ccc (322G), so A also is, and (A,
) must be -nite, by 322G in the opposite direction.
383F Lemma Let (A,
) be a homogeneous semi-nite measure algebra.
(a) If hai iiI , hbi iiI are partitions of unity in A with
ai =
bi for every i, there is a Aut A such
that ai = bi for each i.
(b) If (A,
) is totally nite, then whenever hai iiI , hbi iiI are disjoint families in A with
ai =
bi for
every i, there is a Aut A such that ai = bi for each i.
proof (a) By 383E, (A,
) is -nite, therefore localizable. For each i I, the principal ideals Aai , Abi
are homogeneous, of the same measure and the same Maharam type (being (A) if ai 6= 0, 0 if ai = 0).
Because they are ccc, they are of the same magnitude, as dened in 332G, and there is a measure-preserving
isomorphism i : Aai Abi (332J). By 383B there is a measure-preserving automorphism : A A such
that d = i d for every i I, d ai ; and this serves.
(b) Set a = 1 \ supiI ai , b = 1 \ supiI bi . We must have
P
P
a =
1 iI
ai =
1 iI
bi =
b ,
so adding a , b to the families we obtain partitions of unity to which we can apply the result of (a).
383G Lemma (a) If (A,
) is an atomless semi-nite measure algebra, then Aut A and Aut A have
many involutions.
383J
507
(b) If (A,
) is an atomless localizable measure algebra, then every element of A is the support of some
involution in Aut A.
proof (a) If a A \ {0}, then by 332A there is a non-zero b a, of nite measure, such that the principal
b
ideal Ab is (Maharam-type-)homogeneous. Now because A is atomless, there is a c b such that
c = 12
(331C), so that Ac and Ab\c are isomorphic measure algebras. If : Ac Ab\c is any measure-preserving
we see that both b and b are of innite measure. Similarly we can divide a into c and c , both of innite
measure. Now by 332J the principal ideals Ab , Ab , Ac , A1\c are all isomorphic as measure algebras, so that
there are automorphisms , Aut A such that
c = b,
c = b .
such that
bi =
bi ; since this common measure is nite,
(1 \ bi ) =
(1 \ bi ). By 332J and 383Fa, there is
a i Aut A such that i bi = 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 383Hb we therefore have either one, two or three normal subgroups of Aut A, according to whether
508
Automorphisms
383J
(a e ) =
(a e ) =
(a e )
for every K, because e = e , so that a a. So if I satises the condition, [I] I for every
Aut A and I I. Q
Q
(b) Consequently, for I I and K, e I i there is some a I such that a a 6= 0, since in this
case e a. (This is where I use the hypothesis that (A,
) is totally nite.) It follows that if K is nite,
any I I is the principal ideal generated by sup{e : e I}. Conversely, of course, all such ideals are
Aut A-invariant. Thus I is in a natural order-preserving correspondence with PK, and H
= PK.
(c) Now suppose that K is innite; enumerate it as hn inN . Dene : A by setting a =
(en )inN for a A; so that
h
(a en )/
a b i there is some k such that a kb,
a (a b) a + b 2(a b)
for all a, b A, while (1A ) is the standard order unit 1 of . Let U be the family of solid linear subspaces
of and dene functions I 7 VI : I U, U 7 JU : U I by saying
VI = {f : f , |f | ka for some a I, k N},
JU = {a : a A, a U }.
The properties of just listed ensure that VI U and 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 b} = I.
P
Finally, VJU = U for every U U. P
VJU = {f : a A, k N, |f | ka U } U
because U is a solid linear subspace. But also, given g U , there is an a A such that
(a en ) =
min(1, |g(n)|)
(en ) for every n (because A is atomless); in which case
a |g| max(1, kgk )a
383Xg
509
(c c) > 0.
(c \ c) =
(c)
(c c) =
c
(c c) =
(c \ c) =
Next,
c =
(c) =
(1 c) = c,
so we also have (c \ c) = (c \ c). But now observe that
(c \ c)
(c \ c),
(c \ c) >
(c \ c
writing mag a for the magnitude of a, and setting k = if k > 0 and is an innite cardinal.
> (f ) Let (A,
) be the measure algebra of Lebesgue measure on R. For n N set en = [n, n] A. Let
G Aut A be the group consisting of measure-preserving automorphisms such that supp en for some
n. Show that G is simple. (Hint: show that G is the union of an increasing sequence of simple subgroups.)
(g) Let (A,
) be an atomless totally nite 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.
510
Automorphisms
383Xh
where mag a is the magnitude of a. (ii) Show that the lattice H of normal subgroups of Aut A is wellordered, with least member {} and one member H for each uncountable cardinal less than or equal to
+ , setting
H = { : Aut A, mag(supp ) < }.
(i) Let (A,
) be the measure algebra of Lebesgue measure on [0, 1]. Give an example of two measurepreserving automorphisms of A which are conjugate in Aut A but not in Aut A.
383Y Further exercises (a) Let (A,
) be an atomless totally nite measure algebra. Show that Aut A
and 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) , four if #(X) = 4 or #(X) = , ve if #(X) = 1 .
(c) Let (A,
) be a totally nite measure algebra and : A A a measure-preserving Boolean homomorphism. (i) Show that there are a measure algebra (B,
S ), a measure-preserving automorphism : B B
and a closed subalgebra C of B such that [C] C, nZ n [C] is topologically dense in B, (C, C, C) is
isomorphicPto (A,
, ). (Hint: Take B1 AN to be the subalgebra consisting of sequences a = (a0 , a1 , . . . )
a A let a B be the equivalence class of the sequence h ainN ; set C = {a : a A}.) (ii) Show
that (B, , C, ) are unique up to isomorphism. (This is an abstract version of a construction known as the
natural extension of an inverse-measure-preserving function; see Petersen 83, 1.3G.)
383 Notes and comments This section is short because there are no substantial new techniques to be
developed. 383D is simply a matter of checking that the hypotheses of 382N are satised (and these hypotheses were of course chosen with 383D in mind), and 383I is similarly direct from 382R. 383I-383J, 383Xe
and 383Xh are variations on a theme. In a general Boolean algebra A with a group G of automorphisms,
we have a transitive, reexive relation G dened by saying that a G b if there are 1 , . . . , k G such
that a supik i b; the point about localizable measure algebras is that the functions Maharam type and
magnitude enable us to describe this relation when G = Aut A, and the essence of 382R is that in that
context belongs to the normal subgroup of G generated by i supp G supp .
Some of the most interesting questions concerning automorphism groups of measure algebras can be
expressed in the form how can we determine when a given pair of automorphisms are conjugate? Generally,
people have concentrated on conjugacy in Aut A. But the same question can be asked in Aut A. In
particular, it is possible for two members of Aut A to be conjugate in Aut A but not in Aut A (383Xi).
However this phenomenon does not occur for ergodic automorphisms, or even for ergodic measure-preserving
Boolean homomorphisms (383K-383L).
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 383Xb and 383Yb.
384D
Outer automorphisms
511
proof Because 1 6= 2 , = 21 1 is not the identity automorphism on A, and there is some non-zero a A
such that a a = 0. Let G be an involution supported by a; then 1 is supported by a, so cannot
be equal to , and 1 11 6= 2 21 .
384D 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 : A B such that q() = 1 for every G.
proof (a) The rst half of the proof is devoted to setting up some structures in the group G. Let G be
any involution. Set
the centralizer of in G;
C = { : G, = },
U = { : C , = 1 , 1 = 1 for every C },
the set of involutions in C commuting with all their conjugates in C , together with the identity,
the centralizer of U in G,
V = { : G, = for every U },
S = {2 : V },
the centralizer of S in G.
W = { : G, = for every S },
(b) The point of this list is to provide a purely group-theoretic construction corresponding to the support
of in A. In the next few paragraphs of the proof (down to (f)), I set out to describe the objects just
introduced in terms of their action on A. First, note that is an exchanging involution (382Fa); express it
512
384D
Automorphisms
1 \ a such that d d = 0, by
because d (c 2 c) = 0, while
1 c = 2 1 c
because d 1 c = 0; but this means that 1 c 6= 1 c, so and 1 do not commute, and
/ U . X
X
)?? Suppose that 2 is not supported by a . Then, as remarked in (c-ii), there is a non-zero
(
d 1 \ a such that 2 d d = 0. Now d 6 2 d, so d 6 d; set d = d \ d. Then d d = d 2 d = 0 and
0 6= d 1 \ a ; but this is impossible, by (). X
X
( ) Thus 2 d = d for every d 1 \ a . ?? Suppose, if possible, that is not supported by a . Then
there is a non-zero d 1 \ a such that d d = 0. By 384A, there is a G, of order 4, supported
by d. Because d a = 0, C . Because 6= 1 , there is a c d such that c 6= 1 c; but now
c d = 1 c d = 0, so
1 c = 2 c = c 6= 1 c = 2 1 c = 1 c,
and does not commute with its conjugate 1 , contradicting the assumption that U . X
X
So is supported by a , as claimed. Q
Q
(ii) If u A and u = u, then u U , where
u d = d if d u, u d = d if d u = 0,
u 1 = ((a u) 1 (a u))
(381Sb). () Accordingly
u 1 = (a u a u) = u
and u C . () If C , then
= 1 = (a 1 a ) = (a a ).
So
u 1 = ((a u) 1 (a u)) = (a u a u) = u .
384D
513
Outer automorphisms
which is impossible. X
XQ
Q
(iii) It follows that 2 d = d whenever V and d a . P
P Let e be the support of . Recall that
e = sup{c : c c = 0} (381Gb), so that d e = sup{c : c d, c c = 0}. Now if c a and c c = 0, we
know that c c c, so in fact c c. This shows that (d e) (d e). Also, because = , by
(i), we have
2 (d e)
Of course 2 (d \ e) = d \ e, so 2 d
for every d a . Q
Q
(d e) = (d e)
2 (d e) = d e.
a . But as also 2 a = a = a , 2 d = d
(iv) The nal thing we need to know about V is that V whenever G and supp a = 0;
this is immediate from (d-i) above.
(f ) From (e-iii), we see that if S then supp a = 0. But we also see from (e-iv) that if
0 6= c 1 \ a there is an involution in S supported by c; for there is a member of G, of order 4,
supported by c, and now V so 2 S , while 2 is an involution.
(g) Consequently, W is just the set of members of G supported by a . P
P (i) If supp a and S ,
then supp a = 0, as noted in (e), so = ; as is arbitrary, W . (ii) If supp 6 a , then take a
non-zero d 1 \ a such that d d = 0. Let S be an involution supported by d; then if c d is such
that c 6= c,
c 6= c = c,
and 6= so
/ W . Q
Q
(h) We can now return to consider the isomorphism q : G H. If G is an involution, then q() H
is an involution, and it is easy to check that
q[C ] = Cq() ,
q[U ] = Uq() ,
q[V ] = Vq() ,
q[S ] = Sq() ,
q[W ] = Wq() ,
dening Cq() , . . . , Wq() H as in (a) above. So we see that, for any G,
supp supp supp q() supp q().
a}
b}
514
384D
Automorphisms
(supp ).
(l) We are nearly home. Let us conrm that q() = 1 for every G. P
P?? Otherwise, =
q()1 1 is not the identity automorphism on B, and there is a non-zero b B such that b b = 0,
that is, 1 b q()b = 0. Let H be an involution supported by b. Then q 1 () is supported by 1 b,
by (j), so 1 b supports q 1 ()1 and 1 b supports q(q 1 ()1 ) = q()q()1 . On the other
hand, q()b also supports q()q()1 , which is not the identity automorphism; so these two elements of B
cannot be disjoint. X
XQ
Q
(m) Finally, is unique by 384C.
Remark The ideas of the proof here are taken from Eigen 82.
384E 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 : A B such that q() = 1 for every Aut A.
b be the Dedekind completion of A (314U). Then every Aut A has a unique extension to
proof (a) Let A
bA
b (314Tb). Because the extension is unique, we must have ()b=
a Boolean homomorphism : A
1 are inverses of each other, and
Aut A
b for each Aut A;
for all , Aut A; consequently, and d
1 ) = supp q()
(supp
) = supp(
B. P
= 0 and
(b) If u A, then u
P It is enough to consider the case u
/ {0, 1}, since surely 0
1 = 1. Take any w B which is neither 0 nor 1; then there is an involution in Aut B with support w (382P
again); the corresponding member of H is still an involution with support w. Its image q1 () in G is
b of course 0 6= a 6= 1. Take non-zero u1 , u3 A such that u1 a
an involution with support a = 1 w A;
and u3 1 \ a; set u2 = 1 \ (u1 u3 ). Because A is homogeneous, there are , G such that u1 = u,
u1 = u1 , u2 = u3 ; set 2 = . Then we have
so
u = u1
u = 2 u 1
u 2 u2 = u u3 ,
384I
515
Outer automorphisms
2 (supp q1 ()))
u
q1 ())) (
= (supp
q 1 ()1 ) (supp
2 q1 ()1 )
2
= (supp
q1 (
q () q()1 )) (supp
q1 (
q (2 ) q(2 )1 ))
= supp(
q () q()1 ) supp(
q (2 ) q(2 )1 )
(see the last sentence of (a) above)
516
Automorphisms
384J
384J Theorem Let A and B be nowhere rigid Dedekind complete Boolean algebras and q : Aut A
Aut B an isomorphism. Then there is a unique Boolean isomorphism : A B such that q() = 1
for every Aut A.
proof Put 384I(i)(iv) and 384D together.
384K Corollary Let A be a nowhere rigid Dedekind complete Boolean algebra. Then Aut A has no
outer automorphisms.
384L Examples I note the following examples of nowhere rigid algebras.
(a) A non-trivial 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-nite 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 diculty is to nd an atomless Boolean algebra which is not nowhere rigid; for a variety of
constructions of rigid algebras, see Bekkali & Bonnet 89.
384M Theorem Let (A,
) and (B, ) be atomless localizable measure algebras, and Aut A, Aut B the
corresponding groups of measure-preserving automorphisms. Let q : Aut A Aut B be an isomorphism.
Then there is a unique Boolean isomorphism : A B such that q() = 1 for every Aut A.
proof The point is just that Aut 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
c =
d = min(1, 21
b) (331C). The principal ideals Ac , Ad are now isomorphic as measure algebras (331I);
then (a) = (a ). P
P The principal ideals Aa , Aa are isomorphic as measure algebras; moreover, by
332J, the principal ideals Ae \a , Ae \a are isomorphic. We therefore have a Aut A such that a = a .
Consequently a = a , where = 1 Aut B, and (a) = (a ). Q
Q
If e = 0 we can take = 1. Otherwise x on some c0 e such that 0 <
c0 < ; take b c0 such
that 0 < b < , and set c = 1 b, = b/
c. Then we shall have (a) = (c) =
a whenever
a e and
a =
c. But we can nd for any n 1 a partition cn1 , . . . , cnn of c into elements of measure
1
c; since (cni ) = (cnj ) for all i, j n, we must have (cni ) = n1 (c) =
cni for all i. So if a e
n
c, (a) = (cn1 ) =
a. Now suppose that a e and
a = nk
c for some k, n 1; then
and
a = n1
a can be partitioned into k elements of measure n1
c, so in this case also (a) =
a. Finally, for any
a e , set
D = {d : d
a,
d is a rational multiple of
c},
384P
517
Outer automorphisms
c = #({e : e A , e c});
so
((a a)) = #({e : e A , e a})
= #({e : e A , e a})
X
=
(e) = ((a a)).
eA ,e a
is an infinite cardinal
is an infinite cardinal
((e a)) +
((e a)) +
]0,[
]0,[
((a a))
(a) =
a whenever a e , the Maharam-type- component of A. But since e = e and
e <
for every , we must have = 1 whenever e 6= 0; as A is atomless,
X
(a) =
((a e ))
is an infinite cardinal
is an infinite cardinal
is an infinite cardinal
(a e )
(a e ) =
a
518
Automorphisms
384P
of complications (see 384Ya-384Yd below). To show that we really do need to assume that our algebras are
Dedekind complete or localizable, I oer the following.
Example (a) There are an atomless localizable measure algebra (A,
) and an atomless semi-nite measure
algebra (B, ) such that Aut A
= Aut B but A and B are not isomorphic.
= Aut B, Aut A
proof Let (A0 ,
0 ) be an atomless homogeneous probability algebra; for instance, the measure algebra of
Lebesgue measure on the unit interval. Let (A,
) be the simple product measure algebra (A0 ,
0 )1 (322K);
then (A,
) 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 -ideal generated by the elements of nite measure (cf. 322G). Set
B = {a : a A, either a I or 1 \ a I}.
Then B is a -subalgebra of A, so if we set =
B then (B, ) is a measure algebra in its own right.
The denition 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 7 q() = 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 384E). So q is actually an isomorphism
between Aut A and Aut B. Moreover, still because B is order-dense, q() is measure-preserving i is
measure-preserving, so Aut A is isomorphic to Aut 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 384D and 384J (and localizable in 384M), and the
hypothesis homogeneous in 384E-384F, are essential.
such that Aut C has an outer automorphism.
(b) There is an atomless semi-nite measure algebra (C, )
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 Aut C, we have a
P
corresponding Aut(A A). Now B A is invariant under .
P Consider rst (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
a) (a1 , b1 ) belongs to B A for every a A. We also nd that
1 \ b1 belong to I. Consequently (0,
Now if b I, then
is ccc and
(b, 0) I I
B A;
while
(1 \ b, 0) = (1 \ a1 , 1 \ b1 ) \ (b, 0) B A.
a) = (b,
0) (0,
a) B A
(b,
(ii) Let : A A A A be the involution dened by setting (a, b) = (b, a) for all a, b A. Take
1 Aut(A A). If c = (a, b) C, then 1 c = (b, a) B A, so
Aut C and consider =
1
384Yb
Outer automorphisms
519
1 C
q() =
for Aut C. Evidently q is an automorphism.
(iii) ?? Suppose, if possible, that q were an inner automorphism. Let Aut C be such that q() =
1 for every Aut C. Then
d =
1
1 = q()
for every Aut C. Since G = { : Aut C} is a subgroup of Aut(A A) with many involutions, the
uniqueness assertion of 384D tells us that
= . But
[C] = B A 6= C = [C] = [C],
Remark Thus the hypothesis homogeneous in 384E, and the hypothesis Dedekind complete in 384J, are
necessary.
384Q Example Let be Lebesgue measure on R, and (A,
) its measure algebra. Then Aut 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 of , and ( 21 E) = 12 E for every E (263A, or otherwise). So we have a corresponding
(a) = 21
a for every a A. By
Aut A dened by setting E = ( 21 E) for every E , and
384N, we have an automorphism q of Aut A dened by setting q() = 1 for every measure-preserving
automorphism . But q is now an outer automorphism of Aut A, because (by 384D) the only possible
automorphism of A corresponding to q is , and is not measure-preserving. Q
Q
Thus the hypothesis totally nite in 384O cannot be omitted.
384X 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 disjoint 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
Aut A, of order n, supported by a.
(b) 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.
(c) 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
(d) Find two non-isomorphic atomless totally nite measure algebras (A,
), (B, ) such that Aut A and
Aut B are isomorphic. (This is easy.)
Aut B. Show that either A = B or one of A, B has just one atom and the other is atomless.
(b) Let (A,
), (B, ) be localizable measure algebras such that Aut A
= Aut B. Show that either
(A,
) = (B, ) or there is some ]0, [ such that one of A, B has just one atom of measure and the
other has none or there are , ]0, [ such that the number of atoms of A of measure is equal to the
number of atoms of B of measure , but not to the number of atoms of A of measure .
520
Automorphisms
384Yc
385 Entropy
Perhaps the most glaring problem associated with the theory of measure-preserving homomorphisms and
automorphisms is the fact that we have no generally eective method of determining when two homomorphisms are the same, in the sense that two structures (A,
, ) and (B, , ) are isomorphic, where (A,
)
and (B, ) are measure algebras and : A A, : B B are Boolean homomorphisms. Of course the
rst part of the problem is to decide whether (A,
) and (B, ) are isomorphic; but this is solved (at least
for localizable algebras) by Maharams theorem (see 332J). The diculty lies in the homomorphisms. Even
when we know that (A,
) and (B, ) 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-Sina theorem (385P). 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 (385T-385V). I should remark that some of
the lemmas will be repeated in stronger forms in the next section.
385A Notation Throughout this section, I will use the letter q to denote the function from [0, [ to R
dened by saying that q(t) = t ln t = t ln 1t if t > 0, q(0) = 0.
385C
521
Entropy
0.5
1.5
-0.5
The function q
We shall need the following straightforward facts concerning q.
(a) q is continuous on [0, [ and dierentiable on ]0, [; q (t) = 1 ln t and q (t) = 1t for t > 0.
Because q is negative, q is concave, that is, q is convex. q has a unique maximum at ( 1e , 1e ).
(b) If s 0 and t > 0 then q (s + t) q (t); consequently
Rt
q(s + t) = q(s) + 0 q (s + )d 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 iiN .
(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
(q(t) + q(1 t)) = ln 1t
(d) The function t 7 q(t) + q(1 t) has a unique maximum at ( 12 , ln 2). ( dt
t .)
1
It follows that for every > 0 there is a > 0 such that |t 2 | whenever q(t) + q(1 t) ln 2 .
f (t) = +
1
1t
2t1
t(1t)
1
385B Lemma Let (A,
) be a probability algebra, B a closed subalgebra
)
R of A, and
R P : L (A,
L (A,
) the corresponding conditional expectation operator (365R). Then q(u) q( u) and P (
q (u))
q(P u) for every u L (A)+ .
1
Remarks (a) In the denition 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 385F below). But because q(0) = 0, it makes no dierence; 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 iiI is an indexed
522
385C
Automorphisms
q(P a),
we have 0 P (a) 1 for every a, so q(P a) 0 and every term in the sum is non-negative; accordingly
H(A|B) is well-dened in [0, ].
(b) H(A) = H(A|{0, 1}), since if B = {0, 1} then P (a) =
a1, so that
H(A|B) = 0, since P (a) = a, q(P a) = 0 for every a.
q(P a) = q(
a). If A B,
385F 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 385Xq.)
385G Lemma Let (A,
) 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 a smaller closed subalgebra of A, then H(A|C) H(A|B). In particular, H(A) H(A|B).
(d) Suppose that hBn inN is a non-decreasing sequence of closed subalgebras of A such that B =
S
nN Bn . If H(A) < then
H(A|B) = limn H(A|Bn ).
lim
j=0
n
X
j=0
q(P (a bj )) =
X
j=0
q(P (a bj ))
where all the innite sums are to be regarded as order*-limits of the corresponding nite sums (see 367),
and the middle inequality is a consequence of 385A(f-ii). Accordingly
385G
523
Entropy
H(A B|B) =
aA,bB,ab6=0
Z
XX
q(P (a b))
q(P (a bj ))
aA j=0
XZ
q(P a) = H(A|B).
aA
(ii) Suppose for the moment that A and B are both nite. For a A set ua = P (a). If a, b A we
0
have 0 uab ub inPL0 (B), so we may choose vP
ab L (B) such that 0 vab 1
P and uab = vab ub .
For any b B,
= ub (because
aA uab P
aA (a b) = b), so ub
aA vab = ub . Since
[[|
q (ub )| > 0]] [[ub > 0]], q(ub ) aA vab = q(ub ).
For any a A,
q(ua ) = q(
uab ) = q(
bB
bB
ub vab ) = q(P (
bB
b vab ))
(385B)
= P(
bB
(because B is disjoint)
=
bB
b q(vab ))
ub q(vab )
H(A B|B) =
aA,bB
aA,bB
(385A(f-iii))
XZ
aA
Z
Z
q(uab ) =
aA,bB
ub q(vab ) +
q(ua ) +
XZ
q(ub vab )
aA,bB
vab q(ub )
bB
(iii) For general partitions of unity A and B, take any nite set C A B. Then C {a b : a
A0 , b B0 } where A0 A and B0 B are nite. Set
A = A0 {1 \ sup A0 },
B = B0 {1 \ sup B0 },
cC
(by (ii))
(by (i))
q(P c)
cA B
524
385G
Automorphisms
= H(A|B) + H(B|B).
As C is arbitrary,
H(A B|B) =
cAB
P
(a d)) q(
a) for any a A.
(b) It follows from 385Ab that dD q(
Now, because B is purely atomic and D is its set of atoms,
P
P
(ad)
(ad)
P (a) = dD
d, q(P (a)) = dD q(
)d
for every a A,
H(A|B) =
Putting these together,
H(A D) =
aA,dD
aA,dD
q(
(a d)) =
(385Ac)
= H(A|B) +
q(
(ad)
)
d.
aA,dD
q(
(ad)
(ad)
)
d +
q(
d)
q(
d) = H(A|B) + H(D).
dD
by 385B. So
H(A|C) =
aA
q(PC a)
aA
R
P
PC q(P a) = aA q(P a) = H(A|B).
(d) Let Pn be the conditional expectation operator corresponding to Bn , for each n. Fix a A.
Then P (a) is the order*-limit of hPn (a)inN , by Levys martingale theorem (367Kb). Consequently
(because q is continuous) h
q (Pn a)inN is order*-convergent to q(P a) for every a A (367I). Also, because
0 Pn a 1Rfor every n, R0 q(Pn a) 1e 1 for every n. By the Dominated Convergence Theorem
(367J), limn q(Pn a) = q(P a).
By 385B, we also have
0
q(Pn a) q( Pn (a)) = q( a) = q(
a)
q(P a) = q(
a),
a)
,
and
an
n
N
such
that
0
aA\I
R
R
P
(Pn a) q(P a)|
aI | q
aA\I |
q(Pn a)
P
q(P a)| aA\I q(
a)
385L
525
Entropy
#(Bn ) = 22 ,
Set
B=
Then B is a partition of unity in A and
H(B)
=
X
X
n=0 bBn
an and
b = 22
an for each b Bn .
S
nN
Bn {1 \ a}.
q(
Bn ) =
22 q
n=0
n+1+2n
a
ln
n+1
2
a
n=0
a
2n+1+2n
a n
2 ln 2 = .
n+1
2
n=0
(ii)(iii) Enumerate B as hbi iiN . For each n N, Cn = {bi : i n}{1 \ supin bi } is a nite partition
of unity, and
Pn
limn H(Cn ) limn i=0 q(
bi ) = H(B) = .
(iii)(i) We need only consider the case in which A is purely atomic. In this case, A C = A for every
partition of unity C A, so H(C) H(A) for every C (385H), and H(A) must be innite.
385K Definition Let A be a Boolean algebra. If : A A is an order-continuous Boolean homomorphism, A A is a partition of unity and n 1, write Dn (A, ) for the partition of unity generated by
{ i a : a A, 0 i < n}, that is, {inf i<n i ai : ai A for every i < n} \ {0}. It will occasionally be
convenient to take D0 (A, ) = {1} (or in the trivial case A = {0}). Observe that D1 (A, ) = A \ {0} and
for every n N.
because is measure-preserving. Q
Q
(b) If 1 = then of course H(Dn (A, )) H(A) = for every n 1, by 385H, so inf n1 n1 H(Dn (A, ))
= = limn n1 H(Dn (A, )). Otherwise, n n1 is nite for every n. Set = inf n1 n1 n . If > 0
1
there is an m 1 such that m
m + . Set M = maxj<m j . Now, for any n m, there are k 1,
j < m such that n = km + j, so that
526
385L
Automorphisms
n km + j ,
1
n
n
k
n
m +
M
n
1
m
m
M
.
n
1
n
(ab)
H(A|B) = aA,bB q(
)
b
H(An |Bn ) =
Accordingly, for any n 1,
aA,bB
q(
( n a n b)
)
( n b)
( n b)
= H(A|B). Q
Q
n1
X
H(Ai |Bn )
n1
X
H(Ai |Bi )
i=0
(by 385Ga)
i=0
(by 385Gc)
= nH(A|B).
Now
h(, A) = lim
n n
(385Ga)
1
n
n
385Qa
527
Entropy
1
n
1
n
(385Gb)
h(, B) + H(A|B).
385O Lemma Let (A,
) be a probability algebra, : A A a measure-preserving Boolean homomorphism, and A A a partition of unity such that H(A) < . Then h(, A) h().
proof If A is nite, this is immediate from the denition of h(); so suppose that A is innite. Enumerate
S
A as hai iiN . For each n N let Bn be the subalgebra of A generated by a0 , . . . , an ; set B = nN Bn .
Then A B, so
limn H(A|Bn ) = H(AB) = 0
proof I take the two arguments together. In both cases, by 385O, we have h(, A) h(), so I have to
show that if B A is S
any nite partition of unity, then h(, B) h(, A). For (i), let An be
S the partition
of unity generated by 0j<n j [A]; for (ii), let An be the partition of unity generated by nj<n j [A].
Then h(, An ) = h(, A) for every n. P
P In case (i), we have Dm (An , ) = Dm+n (A, ) for every m, so that
lim
m m
m m
= lim
m m
H(Dm+n (A, ))
H(Dm (A, )).
m m
m m
= lim
m m
H(Dm+2n (A, ))
H(Dm (A, )). Q
Q
528
385Qb
Automorphisms
q(
( inf cj )) =
n1
X
j<n
c0 ,... ,cn1 C
q(
cj )
q(
c0 ,... ,cn1 C
n1
Y
cj ))
j=0
ci
i6=j
(385Ac)
=
n1
XX
q(
c) = nH(C).
j=0 cC
So
1
n
Q
h(, C) = limn H(Dn (C, )) = H(C). Q
(b) If A0 is purely atomic and H(A) < , the result can now be read o fromS385P, becauseSthe closed
subalgebra of A generated by A is A0 and the closed subalgebra of A generated by kN k [A] or kZ k [A]
is A; so h() = h(, A) = H(A).
(c) Otherwise, 385J tells us that there are nite partitions of unity C A0 such that H(C) is arbitrarily
large. Since h() h(, C) = H(C) for any such C, by (a) and the denition of h(), h() must be innite,
as claimed.
385S Remarks (a) The standard construction of a Bernoulli shift is from a product space, as follows.
If (X, , 0 ) is any probability space, write for the product measure on X N ; let (A,
) be the measure
algebra of , and A0 A the set of equivalence classes of sets of the form {x : x(0) E} where E , so
that (A0 ,
A0 ) can be identied with the measure algebra of 0 . We have an inverse-measure-preserving
function f : X N X N dened by setting
f (x)(n) = x(n + 1) for every x X N , n N,
j aj = {x : (f j (x))(0) Ej } = {x : x(j) Ej }
Qk
Qk
T
aj .
385Se
Entropy
529
(b) The same method gives us two-sided Bernoulli shifts. Again let (X, , 0 ) be a probability space,
and this time write for the product measure on X Z ; again let (A,
) be the measure algebra of , and
A0 A the set of equivalence classes of sets of the form {x : x(0) E} where E , so that (A0 ,
A0 ) can
once more be identied with the measure algebra of 0 . This time, we have a measure space automorphism
f : X Z X Z dened by setting
f (x)(n) = x(n + 1) for every x X Z , n Z,
and f induces a measure-preserving automorphism : A A. The arguments used above show that 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 probability algebra is the root algebra of just one Bernoulli shift of
each type.
P
P(i) Given a probability algebra (A0 ,
0 ) then we can identify it with the measure algebra of a probability
space (X, , 0 ) (321J), and now the constructions of (a) and (b) provide Bernoulli shifts with root algebras
isomorphic to (A0 ,
0 ).
(ii) Let (A,
) and (B, ) be probability algebras with one-sided Bernoulli shifts , with root algebras
A0 , B0 , and suppose that 0 : A0 B0 is a measure-preserving isomorphism. Then (A,
) can be identied
with the probability algebra free product of h k [A0 ]ikN (325L), while (B, ) can be identied with the
probability algebra free product of h k [B0 ]ikN . For each k N, k 0 ( k )1 is a measure-preserving
isomorphism between k [A0 ] and k [B0 ]. Assembling these, we have a measure-preserving isomorphism
: A B such that a = k 0 ( k )1 a whenever k N and a k [A0 ], that is, k a = k 0 a for every
a A0 , k N. Of course extends 0 .
If we set
C = {a : a A, a = a},
(c k b) =
c
( k b) =
c
b.
And this is true for every k n. Generally, if b, c A and > 0, there are b , c B such that
(b b )
and
(c c ) , so that
530
Automorphisms
385Se
lim sup |
(c k b)
c
b| lim sup |
(c k b )
c
b |
k
+
(c c ) +
( k b k b ) + |
c
b
c
b |
0 + + + |
c
c | + |
b
b | 4.
(inf jJ j aj ) =
(inf jJ n+j aj ) = jJ
aj . Q
Q
385T Isomorphic homomorphisms (a) In this section I have spoken of isomorphic homomorphisms
without oering a formal denition. I hope that my intention was indeed obvious, and that the next
sentence will merely conrm what you have already assumed. If (A1 ,
1 ) and (A2 ,
2 ) are measure algebras,
and 1 : A1 A2 , 2 : A2 A2 are functions, then I say that (A1 ,
1 , 1 ) and (A2 ,
2 , 2 ) are isomorphic
if there is a measure-preserving isomorphism : A1 A2 such that 2 = 1 1 . In this context, using
Maharams theorem or otherwise, we can expect to be able to decide whether (A1 ,
1 ) and (A2 ,
2 ) are or
are not isomorphic; and if they are, we have a good hope of being able to describe a measure-preserving
isomorphism : A1 A2 . In this case, of course, (A2 ,
2 , 2 ) will be isomorphic to (A1 ,
1 , 2 ) where
1
1 , 1 ) is isomorphic to (A1 ,
1 , 2 ); and when 1 , 2
2 = 2 . So now we have to decide whether (A1 ,
are measure-preserving Boolean automorphisms, this is just the question of whether 1 , 2 are conjugate in
the group Aut1 (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 Aut (A)
for a measure algebra (A,
). 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 eect,
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 1 , 2 have dierent entropies then (A1 ,
1 , 1 ) and (A2 ,
2 , 2 ) 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 terms 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 , 1 , 1 , f1 ) and (X2 , 2 , 2 , f2 ) where, for each i, (Xi , i , i ) is
a measure space and fi : Xi Xi is an inverse-measure-preserving function. Then we have associated
structures (A1 ,
1 , 1 ) and (A2 ,
2 , 2 ) where (Ai ,
i ) is the measure algebra of (Xi , i , i ) and i : Ai Ai
is the measure-preserving homomorphism dened by the usual formula i E = fi1 [E] . Now we can call
(X1 , 1 , 1 , f1 ) and (X2 , 2 , 2 , f2 ) isomorphic if there is a measure space isomorphism g : X1 X2 such
that f2 = gf1 g 1 . In this case (A1 ,
1 , 1 ) and (A2 ,
2 , 2 ) are isomorphic under the obvious isomorphism
(E ) = g[E] for every E 1 .
It is not the case that if the (Ai ,
i , i ) are isomorphic, then the (Xi , i , i , 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 ,
1 , 1 ) and (A2 ,
2 , 2 ) are
actually identical, f1 and f2 need not be isomorphic. There are two examples in 343 of a probability space
(X, , ) 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 for the identity
map from X to itself, (X, , , ) and (X, , , f ) are non-isomorphic but give rise to the same (A,
, ).
(c) Even with Lebesgue measure, we can have a problem in a formal sense. Take (X, , ) 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
385Xd
Entropy
531
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 eective in a wide variety of situations. To formalize it I oer the following denition.
385U Definition Let (X1 , 1 , 1 ) and (X2 , 2 , 2 ) be measure spaces, and f1 : X1 X1 , f2 : X2 X2
two inverse-measure-preserving functions. I will say that the structures (X1 , 1 , 1 , f1 ) and (X2 , 2 , 2 , f2 )
are almost isomorphic if there are conegligible sets Xi Xi such that fi [Xi ] Xi for both i and
the structures (Xi , i , i , fi ) are isomorphic in the sense of 385Tb, where i is the algebra of relatively
measurable subsets of Xi , i is the subspace measure on Xi and fi = fi Xi .
385V I leave the elementary properties of this notion to the exercises (385Xn-385Xp), but I spell out
the result for which the denition is devised. I phrase it in the language of 342-343; if the terms are not
immediately familiar, start by imagining that both (Xi , i , i ) 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 , 1 , 1 ) and (X2 , 2 , 2 ) be perfect, complete, strictly localizable and countably
separated measure spaces, and (A1 ,
1 ), (A2 ,
2 ) their measure algebras. Suppose that f1 : X1 X1 ,
f2 : X2 X2 are inverse-measure-preserving functions and that 1 : A1 A1 , 2 : A2 A2 are the
measure-preserving Boolean homomorphisms they induce. If (A1 ,
1 , 1 ) and (A2 ,
2 , 2 ) are isomorphic,
then (X1 , 1 , 1 , f1 ) and (X2 , 2 , 2 , f2 ) are almost isomorphic.
proof Because (A1 ,
1 ) and (A2 ,
2 ) are isomorphic, we surely have 1 X1 = 2 X2 . If both are zero, we
can take X1 = X2 = and stop; so let us suppose that 1 X1 > 0. Let : A1 A2 be a measure-preserving
automorphism such that 2 = 1 1 . Because both 1 and 2 are complete and strictly localizable
and compact (343K), there are inverse-measure-preserving functions g1 : X1 X2 and g2 : X2 X1
representing 1 , 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 1 2 = 1 1 : A2 A1 and the homomorphism 1 1 again. Next, because
both 1 and 2 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 E1 and
g2 E2 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 N set Gn+1 = Gn f1 [Gn ]. Then every G n is conegligible, so
532
Automorphisms
385Xe
(f ) Let h(Ai ,
i )iiI be a family of probability algebras, with probability algebra free product (C, )
(325K). Suppose that i : Ai Ai is a measure-preserving Boolean homomorphism for each i P
I, and that
: C C is the measure-preserving Boolean homomorphism they induce. Show that h() = iI h(i ).
(Hint: use 385Gb and
N385Gd to show that h() is the supremum of h(, A) as A runs over the nite
partitions of unity in iI Ai . Use this to reduce to the case I = {0, 1}. Now show that if Ai Ai is a
nite partition of unity for each i, and A = {a0 a1 : a0 A0 , a1 A1 }, then H(A) = H(A0 ) + H(A1 ), so
that h(, A) = h(0 , A0 ) + h(1 , A1 ).)
> (g) Let (A,
) be a probability algebra and : A A a measure-preserving automorphism. Show that
h( 1 ) = h().
(h) Let (A,
) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
Show that h( k ) = kh() for any k N. (Hint: if A A is a partition of unity, h( k , A) h( k , Dk (A, )) =
kh(, A).)
>(i) Let (A,
) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
(i) Suppose there is a partition of unity A A such
(a b) =
a
b for every a A, b A ()
S that ()
A is the closed subalgebra of itself generated by nN n [A]. Show that is a one-sided Bernoulli shift, and
that h() = H(A). (ii) Suppose that is a one-sided Bernoulli shift of nite entropy. Show that there is a
partition of unity satisfying () and ().
> (j) Let (A,
) be the measure algebra of Lebesgue measure on [0, 1[. Fix an integer k 2, and dene
f : [0, 1[ [0, 1[ by setting f (x) = <kx>, the fractional part of kx, for every x [0, 1[; let : A A be the
corresponding homomorphism. (Cf. 372Xr.) Show that
is a one-sided Bernoulli shift and that h() = ln k.
(Hint: in 385Xi, set A = {a0 , . . . , ak1 } where ai = ki , i+1
for i < k.)
k
385Yh
Entropy
533
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 (A, B) = 2H(A B) H(A) H(B). Show that
is a metric on A1 (the entropy metric). (iv) Show that if : A A is a measure-preserving Boolean
homomorphism, then |h(, A) h(, B)| (A, B) for all A, B A1 . (iv) Show that the lattice operations
and are -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 .
(r) Let (A,
) be a probability algebra and B a topologically dense subalgebra of A. (i) Show that if
hai iin is a partition of unity in A and > 0, there is a partition hbi iin of unity in B such that
(ai bi )
for every i n. (ii) Show that if A is a nite partition of unity in A and > 0 then there is a nite partition
of unity D B such that H(A D) H(A) + . (iii) Show that if : A A is a measure-preserving
Boolean homomorphism, then h() = sup{h(, D) : D B is a nite partition of unity}. (Hint: 385N,
385Gb.)
(s) Let (A,
) be a probability algebra and : A A an ergodic measure-preserving Boolean homomorphism. Show that if h() > 0 then A is atomless.
385Y Further exercises (a) Let (A,
) 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 nite entropy, then the order-preserving
function B 7 H(A|B) : P ], 0] is order-continuous.
(b) Let (A,
) be a probability algebra, A a partition of unity in A of nite entropy, and : A A a
measure-preserving Boolean homomorphism.
Show that h(, A) = limn H(A|Bn ), where Bn is the closed
S
subalgebra of A generated by 1in i [A]. (Hint: use 385Gb to show that H(A|Bn ) = H(Dn+1 (A, ))
H(Dn (A, )) and observe that limn H(A|Bn ) is dened.)
(c) Let (A,
) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
Suppose
that
there
is a partition of unity A of nite entropy such that the closed subalgebra of A generated
S
by i1 i [A] is A. Show that h() = 0. (Hint: use 385Yb and 385Pa.)
(d) Let be Lebesgue measure on [0, 1[, and take any ]0, 1[. Let f : [0, 1[ [0, 1[ be the measure
space automorphism dened by saying that f (x) is to be one of x + , x + 1. Let (A,
) be the measure
algebra of ([0, 1[ , ) and : A A the measure-preserving automorphism corresponding to f . Show that
h() = 0. (Hint: if Q, use 385Xh; otherwise use 385Yc with A = {a, 1 \ a} where a = 0, 12 .)
R 1
dx for every
(e) Set X = [0, 1] \ Q, let be the measure on X dened by setting E = ln12 E 1+x
1
1
Lebesgue measurable set E X, and for x X let f (x) be the fractional part < x > of x . Recall that f is
inverse-measure-preserving for (see 372N). Let (A, ) be the measure algebra of (X, ) and : A A the
homomorphism corresponding to f . Show that h() = 2 /6 ln 2. (Hint: use the Kolmogorov-Sina theorem
and 372Yf(v).)
(f ) Let (A,
) be a probability algebra, and : A A a one-sided Bernoulli shift. Show that the
procedure of 383Yc extends to a two-sided Bernoulli shift.
(g) Consider the triplets ([0, 1[ , 1 , f1 ) and ([0, 1], 2 , f2 ) where 1 , 2 are Lebesgue measure on [0, 1[,
[0, 1] respectively, f1 (x) = <2x> for each x [0, 1[, and f2 (x) = 2 min(x, 1x) for each x [0, 1]. Show that
these structures are almost isomorphic in the sense of 385U, and give a formula for an almost-isomorphism.
(h) Let (A,
) be a probability algebra, and A1 the set of partitions of unity of nite entropy not
containing 0, as in 385Xq. Show that A1 is complete under the entropy metric. (Hint: show that if hAn inN
is a non-decreasing sequence in A1 and supnN H(An ) < , then the closed subalgebra of A generated by
S
nN An is purely atomic.)
534
Automorphisms
385 Notes
385 Notes and comments In preparing this section I have been heavily inuenced by Petersen 83.
I have taken almost the shortest possible route to Theorem 385P, 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 suciently 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, , ) is a probability space, with measure
algebra (A,
), 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 f : X X
gives rise to a measure-preserving homomorphism f : A A; so we can dene the entropy of f to be h(f ).
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 inversemeasure-preserving functions. But of course a large proportion of the most important examples do arise in
this way (see 385Xj, 385Xk). The same two examples are instructive from another point of view: the case
k = 2 of 385Xj is (almost) isomorphic to the tent map of 385Xk. The similarity is obvious, but exhibiting
an actual isomorphism is I think another matter (385Yg).
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 385T-385V. 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.
proof By 386A and 381O, a supk1 k a. Set a = supkN k a; by 381Kb, a = supkn k a for every n;
by 381Ka, a C. Also, of course, a c whenever a c C, so a = upr(a, C).
(a c) =
c for every c C;
386C
535
() If i + j < k, then
i dj dk
i+j d dk = 0. Q
Q
d c , c = 1.
Set a = supmN dmn (the mn here is a product, not a double subscript!), d = supi<n di = supi<n i d.
Then
Pn1 i
Pn1
(c i d) = i=0
(c d) = n
(c d) = n
c
(c d ) i=0
for every c C. Next, i dmn dmn+i for all m and i, so
supi<n i a = supiN di = 1.
Consequently
Pn1
i=0
(c i a ) = n
(c a ),
c =
c
(c a \ d )
(c a )
(c d ) ( n1 n)
for every c C.
By 331B again (applied to the principal ideal of A generated by a \ d ) there is an a a \ d such that
(a c) =
c for every c C. For 0 < i < n,
i a a = supk,lN i dkn dln
supmi dm
d ,
1
c, c, . . . , k1 c are disjoint and
c = k+1
, so that we have a partition of unity consisting of sets of measure
k+1 . Let B be the set of atoms of the (nite) subalgebra of A generated by B, and m = #(B ). Let > 0
and r, k N be such that
3 (1 n)
b for every b B ,
m(
1)2 < r 2 ,
k
1.
1
1
Pr
(d b) (
b 3) < .
n
P
P We can express the random variable
(d b) as X =
Pr1
j=0
Xj , where Xj =
( s(j) dj b). Then the Xj
are independent random variables. For each j, Xj takes values between 0 and
dj = k
c
1
,
nr
and has
536
expectation
386C
Automorphisms
1
(ej b),
n
where
ej = supi<n i dj = supl<k,i<n n(k+1)j+nl+i c.
So X has expectation
1
(e b)
n
has measure n
c
n
1
nr(k+1)
for each j, so
(e \ e )
Var(X) =
Pr1
j=0
and
and
(1 \ e ) 2; thus E(X)
Var(Xj ) r
2
n
1
k+1
1 2
nr
1
b 2),
n (
while
(
1)2
.
n2 r
1
b 3) ,
Pr X (
n
1
b 3)
Pr X (
n
(
1)2
r 2
<
1
m
by the choice of r. Q
Q
This is true for every b B , while #(B ) = m. There must therefore be some choice of s(0), . . . , s(r 1)
such that, taking d = supj<r s(j) dj ,
1
n
b 3)
b
(d b) (
for every b B , while d , d , . . . , n1 d are disjoint. Because A is atomless, there is a d d such that
(d b) =
b for every b B . Since every member of B is a disjoint union of members of B ,
(d b) =
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, b, . . . , n b are
all disjoint and
(1 \ supin i b) <
a. Now there must be some i < n such that d = i b a 6= 0, in which
case
d n d
i b i+n b = i (b n b) = 0,
where Dn (A, ) is the partition of unity generated by { a : a A, i < n}, as in 385K. Then hwn inN is
norm-convergent in L1 =L1 (A,
) to w say; moreover, hwn inN is order*-convergent to w (denition: 367A).
If T : L0 (A) L0 (A) is the Riesz homomorphism dened by , so that T (a) = (a) 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 { i a : a A, 1 i n}, Bn the set of
its atoms, and Pn the corresponding
conditional expectation operator on L1 (365R). Let B be the closed
S
subalgebra of A generated by nN Bn , and P the corresponding conditional expectation operator. Observe
386E
537
that Bn = [Dn (A, )] and that, in the language of 385F, Dn+1 (A, ) = A Bn . Let C be the xed-point
be the function
subalgebra of and Q the associated conditional expectation. Set L0 = L0 (A), and let ln
from {v : [[v > 0]] = 1} to L0 corresponding to ln : ]0, [ R (364I).
(b) It will save a moment later if I note a simple fact here: if v L1 , then h n1 T n vin1 is order*-convergent
and k k1 -convergent to 0. P
P We know from the
theorem (372G) that h
vn inN is order*-convergent
Pnergodic
1
n+1
1 n
i
and k k1 -convergent to Qv, where vn = n+1
T
v
=
v
vn1 is order*-convergent
T
v.
Now
n
i=0
n
n
and k k1 -convergent to Qv Qv = 0 (using 367C for order*-convergent). Q
Q
(c) Set
vn =
aA
Pn (a) a =
aA,bBn
(ab)
(a b).
P
(a)
a.
It
follows
that
h
ln
v
i
order*-converges
to
ln
v.
P
P
The point is that, for any a A
n nN
aA
v is dened,
and n N, a [[Pn (a) > 0]], so that [[vn > 0]] = 1 for every n, and ln vn is dened. Similarly, ln
1 for each a, so vn 1. To see that {ln vn : n N} is bounded below in L1 , we use an idea from the
fundamental martingale inequality 275D. Set v = inf nN vn . For > 0, a A and n N set
ban () = [[Pn (a) < ]] inf i<n [[Pi (a) ]],
so that
[[v < ]] = supaA,nN a ban ().
Now ban () Bn , so
(a ban ()) =
and
ban ()
a =
ban ()
(a ban ()))
n=0
min(
a,
Pn (a)
(ban ()),
n=0
v is dened. Moreover,
Letting 0,
(a [[v = 0]]) = 0 for every a A, so [[v > 0]] = 1, and ln
v > ln ]]) =
(a [[ ln
(a [[v < ]]) min(
a, )
for every a A, > 0; that is,
(a [[ ln
a, e )
538
386E
Automorphisms
v ) =
( ln
XZ
X Z
aA
a d +
v > ]])d
(a [[ ln
ln(
ln(
e d
ln(1/
a)
min(
a, e )d
ln(1/
a)
aA
aA
aA
aA
XZ
v > ]]d =
[[ ln
1
)
a + eln a
1
)
a +
aA
a = H(A) + 1 <
{ln vn : n N}. Q
Q
v L1 and hln
vn inN ln
v for k k1 .
By 367J again, ln
(d) Fix n N for the moment. For each d Dn+1 (A, ) let d be the unique element of Bn such that
d d . Then
(n + 1)wn+1 =
ln(
dDn+1 (A,)
ln(
bBn
1
)d
1
)b
ln(
dDn (A,)
ln(
dDn+1 (A,)
ln(
aA
bBn
ab6=0
d
)d
(ab)
)(a b)
1
vn
)(d) ln
(d)
vn .
= T (nwn ) ln
aA
ln(
1
)a
v0 ,
= ln
we get
nwn =
Pn1
i=0
vni1 ),
T i ( ln
wn =
1 Pn1 i
vni1 )
T ( ln
n i=0
for every n 1.
Pn1
v) for n 1. By the Ergodic Theorem, hw in1 is order*-convergent and
(e) Set wn = n1 i=0 T i ( ln
n
vk ln
v| for
k k1 -convergent to w = Q( ln v), and T w = w. To estimate wn wn , set un = supkn | ln
each n N. Then hun inN is a non-increasing sequence, u0 L (by (c) above), and inf nN un = 0 because
vn inN order*-converges to ln
v. Now, whenever n > m N,
hln
386G
539
|wn wn |
=
1
n
n1
X
i=0
v ln
vni1 |
T i | ln
1
n
nm1
X
1
n
nm1
X
i=0
v ln
vni1 | +
T | ln
T i um +
m1
X
j=0
i=0
1
nm
n1
X
nm1
X
T i um +
i=nm
v ln
vj |
T n1j | ln
m1
X
T n1j u0
j=0
i=0
1
nm
nm1
X
T i um +
1
nm
nm1
X
T i um +
1
T nm
nm
i=0
m1
X
T m1j u0
j=0
1
T nm u
m ,
nm
i=0
Pm1
setting u
m = j=0 T m1j u0 .
Holding m xed and letting n , we know that
1 Pnm1
nm
v ln
vni1 |
T i | ln
i=0
T i um
1
nm
u
m
nm T
is order*-convergent
for every m N. Since hQum imN is surely a non-decreasing sequence with inmum 0,
lim supn |wn wn | = 0,
386F Corollary If, in 386E, is ergodic, then hwn inN is order*-convergent and k k1 -convergent to
h(, A)1.
proof Because
R the limit w in 386E has T w = w, it must be of the form 1, because is ergodic (372Q(aii)). Now = w must be
lim
1
n n
wn = lim
dDn (A,)
ln(
1
)
d
1
n n
= lim
q(
d)
dDn (A,)
n n
386G Definition Set p(t) = t ln t for t > 0, p(0) = 0; for any Dedekind -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
385A.)
540
Automorphisms
386H
( |u 1|)2 2 p(u).
R
R
b = 1 and b u = 1 . Surely < 1.
proof Set a = [[u < 1]], =
a, = a u, b = 1 \ a. Then
If = 0 then u = 1 and the result is trivial; so let us suppose that 0 < < 1. Because the function p is
convex,
p(u)
a p(
1 R
u)
a a
= p( ) = p() ln ,
Also
so
Z
|u 1| =
(1 u) +
a
(u 1) = + (1 ) (1 ) = 2( ),
Z
1
p(u) ( |u 1|)2 p() ln + p(1 ) (1 ) ln(1 ) 2( )2
2
= ()
say. Now is continuous on [0, 1] and arbitrarily often dierentiable on ]0, 1[,
() = 0,
(t) = ln t ln ln(1 t) + ln(1 ) + 4( t) for t ]0, 1[,
() = 0,
(t) =
1
t
1
1t
1
2
p(u) ( |u 1|)2 0,
R
R
that is, ( |u 1|)2 2 p(u), as claimed.
aA,bB
|
(a b)
a
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,
{0}. Let (C, )
) with itself (325E, 325K). Set
P
(ab)
u = aA,bB
(a b) L0 (C);
a
b
u=
aA,bB
(a b) = 1. Now
386L
541
p(u) =
aA,bB
(a b) ln
= H(A B)
= H(A B)
(ab)
a
b
aA,bB
aA
(a b) ln
a
a ln
a
aA,bB
(a b) ln
b
b ln
b
bB
|u 1| =
aA,bB
a
b|
(ab)
a
b
which is 386H.
|u 1|
1| =
q R
p(u),
aA,bB
|
(a b)
a
b|.
386J The next six lemmas are notes on more or less elementary facts which will be used at various
points in the next section. The rst two are nearly trivial.
Lemma Let (A,
) be a probability algebra and hai iiI , hbi iiI two partitions of unity in A. Then
1P
(supiI ai bi ) = 1
(ai bi ).
iI
2
proof
(sup ai bi ) =
iI
X
iI
(ai bi ) =
X1
iI
=1
2
1
2
(
ai +
bi
(ai bi ))
X
iI
(ai bi ).
(ci , Bk )
(ci 0) =
ci
542
386L
Automorphisms
(d) If H(A) < and B is any closed subalgebra of A such that [B] B, then h(, A) h( 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)
n n
= lim
n n
lim
n n
n n
H(Dn (A B, )) = h(, A B)
(H(Dn (A, ) + H(Dn (B, ))) = h(, A) + h(, B)
h(, A) + H(B)
as remarked in 385M.
(c) Set n = H(Dn+1 (A, )) H(Dn (A, )) for each n N. By 385H, n 0. From (a) we see that
n+1 = H(A [Dn+1 (A, )]) H(A [Dn (A, )])
H([Dn+1 (A, )]) H([Dn (A, ]) = n
for every n N. So limn n = inf nN n ; write for the common value. Now
1 Pn1
1
h(, A) = limn H(Dn (A, )) = limn
i=0 i =
n
(273Ca).
(d) Let P : L1 L1 be the conditional expectation operator corresponding to B. Let hbk ikN be a
sequence running over {[[P (a) > 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 kN Bk , and P for the corresponding conditional expectation
operator, then P (a) L0 (B ), so P (a) = P (a), for every a A. So
R
P
H(A|B) = aA q(P a) = H(A|B ) = limk H(A|Bk ),
by 385Gd.
For each k, let Bk be the set of atoms of Bk . Then
386M
543
1\b
=
(a \ b) +
b
(a b) =
(a b).
If a A set h(a) = [[P (a) > 21 ]]. Then |P (a) h(a)| |P (a) b| for any b B, so
Z
(a, B) = inf
(a b) = inf
|P (a) b|
bB
bB
Z
= |P (a) h(a)| =
(a h(a)).
If a a = 0, then
P (a) + P (a ) = P (a a ) 1,
so
h(a) h(a ) = [[P (a) > 12 ]] [[P (a ) > 21 ]] [[P (a) + P (a ) > 1]] = 0,
by 364D(b-i).
(b) By 385Ae, q(t) q(1t) whenever 21 t 1. Consequently q(t) q(min(t, 1t)) for every t [0, 1],
and q(u) q(u (1 u)) whenever u L0 (A) and 0 u 1. Fix a A for the moment. We have
Consequently
q(P a)
(because q is concave)
|P (a) h(a)|
= q((a, B)).
Summing over a,
H(A|B) =
aA
q(P a)
aA
q((a, B)).
(c) Set bi = h(ai ) for each i N. Then hbi iiN is disjoint. Next, for each i N, take bi B such that
bi and
bi = min(
ai ,
bi ); then hbi iiN is disjoint and
bi
ai for every i. We can therefore nd a
Take any i N. If
bi >
ai , then
bi
(ai bi ) =
(ai bi )
(ai bi ) +
(bi bi )
If
bi
ai , then
=
(ai bi ) +
bi
ai 2
(ai bi ) = 2(ai , B).
(ai bi )
(ai bi ) +
(bi bi )
=
(ai bi ) +
ai
bi 2
(ai bi ) = 2(ai , B).
544
Automorphisms
386N
kN
then
Now
P
P
e (2k + 1) bB
(b cb ) = (2k + 1) bB (b, C).
B = {d : d A, c C such that d \ e = c \ e}
c or d c = 0}
386 Notes
545
(c) H(B) . P
P?? Suppose otherwise. We know that
Let m N be such that H(B) + H({dk , 1 \ dk }) for every k m. Because B renes Bk , we must have
H(Bk {dk , 1 \ dk }) H(Bk ) + H({dk , 1 \ dk }) ,
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
> H(B) k=m q(
dk ) = ,
which is impossible. X
XQ
Q
Thus B has the required properties.
546
387 intro.
Automorphisms
(b) If is an automorphism, a Bernoulli partition hai iiI for is (two-sidedly) generating if the
closed subalgebra generated by { j ai : i I, j Z} is A itself.
(c) A factor of (A,
, ) is a triple (B,
B, B) where B is a closed subalgebra of A such that
[B] = B.
387B Remarks Let (A,
) be a probability algebra, : A A a measure-preserving Boolean homomorphism and hai iiI a Bernoulli partition for .
(a) h k [A0 ]ikN is independent, where A0 is the closed subalgebra of A generated by {ai : i I}. P
P
Suppose that cj j [A0 ] for j k. Then each j cj A0 is expressible as supiIj ai for some Ij I. Now
( inf cj ) =
(
jk
inf j aij )
sup
i0 I0 ,... ,ik Ik jk
( inf j aij ) =
i0 I0 ,... ,ik Ik
k X
Y
ai =
jk
k
Y
j=0
j=0 iIj
k
Y
aij
(sup ai ) =
iIj
k
Y
cj .
j=0
387C
Ornsteins theorem
547
p
proof (a) Of course h(,
in 385M, so the square root 2(H(A) h(, A)) gives
qPA) H(A), as remarked
p
1
ai i | + 2(H(A) h(, A)), = min( 14 , 24
).
no diculty. Set =
i=0 |
P
There is a sequence h
i iiN of non-negative real numbers such that
{i : i > 0} is nite, i=0 i = 1,
P
P
P
2
|
i i | 2 2 and i=0 q(
i ) h(). P
P Take
i=0P
Pk N such that i=k i , and set i = i for i < k,
i=0
|
i i | k +
i=k
i 2 2 . Q
Q
P
P
Because i=0 q(
i ) is nite, there is a partition of unity C in A, of nite entropy, such that i=0 q(
i )
h(, C) + 3; replacing C by C A if need be (note that C A still has nite entropy, by 385H), we may
suppose that C renes A.
P
There is a sequence hi iiN of non-negative real numbers such that i=0 i = 1, {i : i > 0} is nite,
P
2
i=0 |i i | 4 and
P
= 2 for k + 1 i k + r,
= 0 for i > k + r.
548
387C
Automorphisms
Then
i=0
Pk+r
i=0
|
i i | = 2 2 ,
i=0
|
i i | 4 2 ,
2
r
q(
i ) h(, C) + 3 2 ln( 2 ) = rq( )
7
Pk+r
i=0
Pk+r
i=0
q(
i ).
q(
i + (1 )
i ) : [0, 1] R
i i | + (1 ) i=0 |
i i | 4 2 . Q
Q
i=0 |i i |
i=0 |
Set M = {i : i 6= 0}, so that M is nite.
(iii) .
P
c) is nite.)
(Actually, (iii) is a consequence of (i). For (ii) we must of course rely on the fact that cC q(
Let be the probability measure on M dened by saying that {i} = i for every i M , and the
product measure on M N . Dene Xij : M N {0, 1}, for i M and j N, and Yj : M N R, for j N, by
setting
Xij () = 1 if (j) = i,
= 0 otherwise,
Yj () = ln((j)
) for every M N .
Then, for each i M , hXij ijN is an independent sequence of random variables, all with expectation i ,
and hYj ijN is also an independent sequence of random variables, all with expectation
P
P
iM i ln i =
i=0 q(i ) = h(, C) 3.
Let n 1 be so large that
(iv)
[[wn h(, C)1 ]] < , where
1P
wn =
dDn (C,) ln(
n
1
)d;
(v)
P
1 Pn1
Pr
iM |
j=0 Xij i | 1 ,
n
1 Pn1
Pr |
j=0 Yj + h(, C) + 3| 1 ;
n
(vi) en 2,
1
n+1
q(
1
n
) + q(
)
n+1
n+1
these will be true for all suciently large n, using the Shannon-McMillan-Breiman theorem (386E) 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<n,iM such that
() for each j < n, hbji iiM is a partition of unity in A,
Qn1
()
(inf j<n bj,i(j) ) = j=0 i(j)
for every i(0), . . . , i(n 1) M ,
387C
549
Ornsteins theorem
()
iM
P
P Construct hbji iiM for j = n 1, n 2, . . . , 0, as follows. Given bji , for k < j < n, such that
Qn1
|
( k ai c) i
c|
i=0
|
( k ai c)
ai
c| +
|i i | +
4 2 +
X
i=0
X
i=0
|
ai i | +
X
i=0
|
ai i |
cDk (A,)
|
( k ai c)
ai
c|
|
ai i | + 2 H( k [A]) + H(Dk (A, )) H(Dk+1 (A, ))
X
p
|
ai i | + 2(H(A) h(, A))
4 2 +
i=0
Choose a partition of unity hbki iiM such that, for each c Dk (A, ), b Bk and i M ,
(bki b c) = i
(b c),
if
( k ai b c) i
(b c) then bki b c k ai ,
if
( k ai b c) i
(b c) then k ai b c bki .
(This is where I use the hypothesis that A is atomless.) Note that in these formulae we always have
(b c) =
b
c,
Consequently
( k ai b c) =
( k ai c)
b.
550
387C
Automorphisms
iM
( k ai bki ) =
min(
(b c k ai ), i
(b c))
(b c k ai ) |
(b c k ai ) i
(b c)|
=1
=1
=1
(b c ( k ai bki ))
|
(b c k ai ) i
(b c)|
b |
(c k ai ) i
c|
|
(c k ai ) i
c| 1 2 4 2 .
Also we have
(bki b c) = i
b
c =
(bki b)
c
for every b Bk , c Dk (A, ) 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 and
d Dn (C, ) set
Ibd = {j : j < n, i M, b bji , d
j ai }.
bB,dDn (C,)
#(Ibd )
(b d) =
n1
X
j=0 iM
(bji j ai ) n(1 2 4 2 ).
Set
e0 = sup{b d : b B, d Dn (C, ), #(Ibd ) n(1 4)};
then
e0 1 .
(e) Let B B be the set of those b B such that
P
1
b en(h(,C)+2) ,
iM |i #({j : j < n, b bji })| .
n
Then
(sup B ) 1 2. P
P Set
B1 = {b : b B,
b en(h(,C)+2) }
= {b : b B, h(, C) + 2 +
1
n
ln(
b) 0}
Then
1
n
n1
X
j=0
ln i(j)
0}.
387C
551
Ornsteins theorem
(sup B1 ) = Pr(h(, C) + 2 +
1
n
Pr(|h(, C) + 3 +
1
n
n1
X
j=0
Yj 0)
n1
X
j=0
Yj | ) 1
iM
1
n
iM
1
n
we have
P
1 Pn1
(sup B2 ) = Pr( iM |i
j=0 Xij | ) 1
n
B1
B2 ,
(sup B) 1 2. Q
Q
Let D0 be the set of those d Dn (C, ) such that
1
n
ln(
1
)
h(, C) + ,
i.e.,
d en(h(,C)+) ;
by (b-iv),
(sup D0 ) > 1 . Let D D0 be a nite set such that
(sup D ) 1 . If d D and b B
then
d en(h(,C)+) en
b 2
b.
Since
(sup D ) 1 2
(sup B ) (remember that 41 ), #(D ) #(B ).
D = {d : d D ,
(d e1 )
d};
then
(sup(D \ D )) 2
(1 \ e1 ) 2 + 6,
so
(sup D ) 1 2 6 1 2 7.
(f ) If d1 , . . . , dk D are distinct,
(sup1ik di e1 )
k
2
inf ik
di k supbB
b,
and
#({b : b B , b e0 sup1ik di } 6= 0) k.
By the Marriage Lemma (3A1K), there is an injective function f0 : D B such that d f0 (d) e0 6= 0
for every d D . Because #(D ) #(B ), we can extend f0 to an injective function f : D B .
(g) By the Halmos-Rokhlin-Kakutani lemma, in the strong form 386C(iv), there is an a A such that
1
j
n, d D }. Because h (a d)ij<n,dD is disjoint,
Pn1 P
n P
d (1 )2 1 2.
(a d) =
e = j=0 dD
dD
n+1
bji }.
552
387C
Automorphisms
bji ,
j (a d) j (a d ) j a j a = 0.
j (a d) j (a d )
j (d d ) = 0. Q
Q
Observe that
supiM ai = supj<n,dD j (a d) = e
because if j < n and d D there must be some i M such that f (d) bji . Take any m N \ M and set
am = 1 \ e, ai = 0 for i N \ (M {m}); then hai iiN is a partition of unity. Now
X
iM
|
ai i |
iM
i |1 n
(a sup D )| +
iM
|
ai ni
(a sup D )|
(sup D )
n+1
1
+
X
X n1
|
iM
j=0
dD
f (d) bji
( j (a d)) ni
dD
1 (1 )2
X X
+
|
(a d) #({j : j < n, f (d)
dD iM
1 (1 )2 +
dD
(a d)|
bji }) ni
(a d)|
(a d)n
X
i=0
|
ai i |
am +
iM
|
ai i | +
X
i=0
|i i |
2 + 3 + 4 6 .
bji
and d
j ai }
= sup{ j (a d) : d D , j If (d),d }
P
(see (d) for the denition of Ibd ) has measure at least dD #(If (d),d )
(a d).
For d D , we arranged that d f (d) e0 6= 0. This means that there must be some b B and
d Dn (C, ) such that d f (d) b f (d ) 6= 0 and #(Ibd ) n(1 4); of course b = f (d) and d = d
(because f is injective), so that #(If (d),d ) must be at least n(1 4). Accordingly
X
i=0
(ai ai )
dD
n(1 4)
(a d) = n(1 4)
(sup D )
n+1
387C
553
Ornsteins theorem
i=0
(ai ai ) = 2(1
i=0
(ai ai )) 6 + 24 + 6
(j) Finally, we need to estimate H(A ) and h(, A ), where A = {ai : i N} \ {0}. For the former, we
have H(A ) h(, C) + 4. P
P |
ai i | 3 for every i, by (h) above. So by (b-i),
P
P
Q
ai ) + i=0 q(i ) = h(, C) + 4. Q
H(A ) = iM {m} q(
(k) Consider the partition of unity
A = A {a, 1 \ a}.
ai
j
whenever d D and d 6= d. But as a a = 0 if 0 < j < n, a e must be sup{a d : d D }, 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 j (a d) : j < n, d D }
= sup{ j ( j c a d) : j < n, d D }
= sup{ j (a d) : j < n, d D , d
j c or d j c = 0)
j c}
(c, D)
(c (c e)) =
(c \ e) min(
c, 2) .
So
h(, C) h( D) + H(C|D)
h(, A ) +
q((c, D))
q(min(
c, 2))
cC
cC
554
387C
Automorphisms
n
1
) + q(
)
n+1
n+1
h(, A ) +
by the choice of n. Q
Q
(l) Putting these together,
H(A ) h(, C) + 4 h(, A ) + 6 h(, A ) + ,
and the proof is complete.
P
2n .
n=0 n + 6 n +
Using 387C, we can choose inductively, for n N, partitions of unity hani iiN such that, for each n N,
P
ani | n ,
i=0 |i
H(An ) h(, An ) + n <
i=0
while
i=0
(a0i ai ) 0 + 6.
P P
P
P p
(an+1,i ani ) n=1 n + n=0 6 n + 2n < .
n=0
i=0
P
In particular, given i N, n=0
(an+1,i ani ) is nite, so hani inN is a Cauchy sequence in the complete
metric space A (323Gc), and has a limit ai , with
ai = limn
ani = i
(323C). If i 6= j,
i=0
ai =
i=0
i = 1,
387F
555
Ornsteins theorem
X
i=0
(ai ai )
X
i=0
(a0i ai ) +
0 + 6 +
X
X
n +
n=1
(an+1,i ani )
n=0 i=0
q
X
n +
n=0
2n + 6.
|
(d j ani )
d
ani |
by 386I. A fortiori,
|
(d j ani )
d
ani |
2n ,
2n
Qr
|
(inf jr j an,i(j) ) j=0
an,i(j) | r 2n 0
as n , for any r k. Because
,
= lim
jk
k
Y
j=0
an,i(j) =
k
Y
i(j) .
j=0
556
387F
Automorphisms
, min( , ),
ci )) .
+ 6 4 2
i=0 q(min(2,
4(2m+1)
4 6
P
ci ) is nite.) Let r m be such that
(The last is achievable because i=0 q(
(ci , Br ) for every i N.
2r+1
2n+2
(dvu )
2n+2
for every
u U and v V . ( C is a Bernoulli shift, therefore ergodic, by 385Se, therefore aperiodic, by 386D.) Set
e = sup|j|n j a, e = sup|j|nr j a; then
e = (2n + 1)
a =
2n+1
,
2n+2
e = (2(n r) + 1)
a = 1
2r+1
.
2n+2
P
P All we have to note is that the families
are disjoint. Q
Q Consequently, if we set
j
bi = sup
(a dvu ) C
|j|n supvV supuU,u j bi
for i N, hbi iiN is disjoint, since a given triple (j, u, v) can contribute to at most one bi .
Of course bi sup|j|n j a = e for every i. If i n, we also have
bi =
e
bi . P
P For |j| n, j bi is
a union of members of U , so
bi =
n
X
X
j=n vV uU,u j bi
( j (a dvu ))
(a dvu ) =
=
j=n vV uU,u j bi
1
2n+2
1
2n+2
n
X
X
j=n vV uU,u j bi
n
X
n
X
X
dvu
j=n vV uU,u j bi
(v u)
u =
j=n uU,u j bi
1
2n+2
1
2n+2
n
X
( j bi )
j=n
2n+1
bi
2n+2
=
e
bi . Q
Q
Again because C is atomless, we can choose a partition of unity hbi iiN in C such that
bi =
bi for every
387F
557
Ornsteins theorem
for b E.
(i) is a Boolean homomorphism. P
P The point is that if |j| n r and b E, then j b belongs to
k
the algebra generated by { bi : i n, |k| n} { k ci : i n, |k| n}, so is a union of members of U .
Since each map
b 7 j (a dvu ) if u
j b,
0 otherwise
(b) =
nr
X
j=n+r vV
1
2n+2
nr
X
uU,u j b
j (a dvu )
j=n+r vV uU,u j b
(v u) =
2n2r+1
b
2n+2
b. Q
Q
sup
vV uU,u j+k bi
= k sup
j (a dvu )
sup
vV uU,u j+k bi
jk (a dvu )
= k ( jk a bi ) = j a k (e bi ) = j a k bi
because j a k e. Taking the supremum of these pieces we have
( k bi ) = sup|j|nr j a ( k bi ) = sup|j|nr j a k bi = e k bi . Q
Q
It follows that
( k (1 \ supil bi )) = e k (1 \ supil bi ))
if l n and |k| r.
(iv) Finally, ci = ci e for every i n. P
P If |j| nr and v V then either v
In the former case,
dvu = v u = 0 whenever u U and u 6 j ci ,
so that
v = supuU dvu = supuU,u j ci dvu ;
in the latter case, dvu = v u = 0 whenever u
j ci . So we have
v j ci = supuU,u j ci dvu
for every v V , and
j ci or v j ci = 0.
558
387F
Automorphisms
ci =
=
sup sup
sup
|j|nr vV uU,u j ci
sup j (a sup
dvu )
sup j (a sup (v j ci ))
vV
|j|nr
sup
vV uU,u j ci
|j|nr
j (a dvu )
sup
|j|nr
(a j ci ) = ci
sup j a = ci e. Q
Q
|j|nr
(e) Let B be the closed subalgebra of A generated by { j bi : i N, |j| Z}. Then for every b Br
there is a b B such that b = b e. P
P The set of b for which this is true is a subalgebra of A containing
k bi for i r and |k| r, by (d-iii). Q
Q It follows that
(ci , B ) 2 for i N.
P
P If i > n this is trivial, because
ci , by the choice of n. Otherwise, ci E. Take b Br such that
2r+1
2n+2
+
((
e ci ) b) =
2r+1
2n+2
+
((ci b))
2r+1
2n+2
+
(ci b)
(by (d-iv))
2r+1
2n+2
+
(ci b)
(by (d-ii))
2
by the choice of n. Q
Q
(f ) Set B = {bi : i N} \ {0}. Then H(B ) = h(, C) h(, B ) + . P
P
H(B ) = H(B) = H(C)
(because
bi =
bi for every i, and we supposed from the beginning that H(C) = H(B))
= h(, C)
(because C is a Bernoulli partition, see 387Bc)
h( B ) + H(C|B )
(386Ld)
h( B ) +
q((ci , B ))
q(min(2,
ci ))
i=0
(386Mb)
h(, B ) +
i=0
(by the Kolmogorov-Sina theorem, 385P(ii), and (e) above, recalling that 16 , so that q is monotonic on
[0, 2])
h(, B ) +
by the choice of . Q
Q
Note also that H(B ) = h(, C) h().
387G
559
Ornsteins theorem
(g) By 387D, applied to C and the partition hbi iiN of unity in C and the sequence hi iiN = h
bi iiN ,
we have a Bernoulli partition hdi iiN in C such that
di =
bi =
bi for every i N and
P
(di bi ) + 6 4 2
.
i=0
4(2m+1)
Let D C be the closed subalgebra of A generated by { j di : i N, j Z}. Then (B, B, hbi iiN ) is
isomorphic to (D, D, hdi iiN ), with an isomorphism : B D such that = and bi = di for every
i N (387Bi).
(h) Set
e = e \ sup|j|m,iN j (di bi ).
Then ( j bi ) e = ( j bi ) e whenever i m and |j| m. P
P
( j bi ) e = j (bi ) e = j di e
= j bi e = j bi e e = ( j bi ) e
(ci ci ) 2
ci 2 .
(ci b) (b b) (b ci ) (ci ci )
(ci b) (1 \ e ) (b ci )
(ci b) +
(1 \ e ) +
(b ci )
(by (d-ii), since b and ci both belong to E)
2
(ci b) +
(1 \ e) + (2m + 1)
2r+1
2n+2
X
i=0
(di bi )
as required. Q
Q
387G Lemma Let (A,
) be an atomless probability algebra and a measure-preserving automorphism
of A. Let hbi iiN and hci iiN be Bernoulli partitions for , of the same nite entropy, and write B, C for
the closed subalgebras generated by { j bi : i N, j Z} and { j ci : i N, j Z}. Suppose that C B.
Then for any > 0 we can nd a Bernoulli partition hdi iiN for such that
(i)
di =
ci for every i N,
(ii)
(di ci ) for every i N,
(iii) writing D for the closed subalgebra of A generated by { j di : i N, j Z}, (bi , D) for every
i N.
proof (a) By 387F, there is a Bernoulli partition hbi iiN for such that bi C for every i N,
bi =
bi
1
for every i N, and
(ci ci ) 4 for every i N, where : B C is the measure-preserving Boolean
homomorphism such that bi = bi for every i and = . Note that this implies that 1 = 1 ,
and generally that j = j for every j Z; so [B] C is the closed subalgebra of A generated by
{ j bi : i N, j Z} = { j bi : i N, j Z} (324L), and is invariant under the action of and 1 .
Let m N be such that
(ci , Bm ) 14 for every i N,
560
387G
Automorphisms
where Bm is the closed subalgebra of A generated by { j bi : i N, |j| m} (386K). Let ]0, ] be such
that
P
(2m + 1) i=0 min(, 2
bi ) 41 .
By 387F again, applied to C, there is a Bernoulli partition hci iiN for such that ci [B],
ci =
ci
and
(bi bi ) for every i N, where : C C is the measure-preserving Boolean homomorphism
such that ci = ci for every i N and = . Once again, [C] will be the closed subalgebra of A
generated by { j ci : i N, j Z} = { j ci : i N, j Z}; because every ci belongs to [B], [C] [B].
(b) Now
(ci ci ) for every i N. P
P There is a b Bm such that
(ci b) 41 . We know that
j
[Bm ] is the closed subalgebra of A generated by { bi : i N, |j| m} = { j bi : i N, |j| m}, and
contains b. Because
(b) b supiN,|j|m ( j bi ) j bi = sup|j|m j (supiN bi bi ),
we have
(b b) (2m + 1)
(bi bi )
(2m + 1)
min(, 2
bi ) .
i=0
i=0
1
4
(ci ci ) =
(ci ci )
(ci b) +
(b b) +
(b ci )
(ci b) +
+
(b ci )
(ci ci ) +
(ci b) +
+
(ci b) +
. Q
Q
(c) Set di = 1 ci for each i; this is well-dened because is injective and ci [B]. Write D for
the closed subalgebra of A generated by { j di : i N, j Z} = {1 j ci : i N, j Z}; note that
D = 1 [[C]], by 324L again, because 1 : [B] B is a measure-preserving homomorphism. Then
di =
ci =
ci for every i N, and hdi iiN is a Bernoulli partition for . P
P If i(0), . . . , i(n) N, then
( inf j di(j) ) =
( inf j 1 ci(j) ) =
(( inf j 1 ci(j) ))
jn
jn
=
( inf j ci(j) )) =
jn
jn
n
Y
j=0
ci(j) =
n
Y
di(j) . Q
Q
j=0
Next,
(ci di ) =
(ci di ) =
(ci ci )
387H
561
Ornsteins theorem
proof (a) To begin with (down to the end of (c) below) suppose that A = B. Choose hn inN , hn inN ,
hrn inN and hhdni iiN inN inductively, as follows. Start with r0 = 0 and d0i = ci for every i. Given that
hdni iiN is a Bernoulli partition with
dni =
ci for every i, take n , n > 0 such that
(2rm + 1)n 2n for every m n,
P
n 2n1 ,
ci ) n ,
i=0 min(n , 2
and use 387G to nd a Bernoulli partition hdn+1,i iiN for such that
(dn+1,i dni ) n ,
dn+1,i =
ci ,
for every i N, where D(n+1) is the closed subalgebra of A generated by { j dn+1,i : i N, j Z}. Let rn+1
be such that
(n+1)
(bi , Drn+1 ) 2n
(n+1)
for every i N, where Drn+1 is the closed subalgebra of A generated by { j dn+1,i : i N, |j| rn+1 }.
Continue.
(b) For any i N,
n=0
(dn+1,i dni )
(ci di )
n=0
n=0 n
(dn+1,i dni )
di = limn
dni =
ci
for each i, and if i 6= j then
since
i=0
(c) Let D be the closed subalgebra of A generated by { j di : i N, j Z}. Then bj D for every j N.
(m+1)
(m+1)
(bj b) 2m . Now
P
P Fix m N. Then (bj , Drm+1 ) 2m , so there is a b Drm+1 such that
X
i=0
(dm+1,i , D)
(dm+1,i di )
i=0
k=m+1 i=0
X
X
i=0 k=m+1
min(2
ci , k )
(dk+1,i dki )
k .
k=m+1
So
(b, D) (2rm+1 + 1)
(dm+1,i , D)
i=0
(386Nc)
(2rm+1 + 1)k
k=m+1
2k = 2m ,
k=m+1
and
(bj , D)
(bj b) + (b, D) 2m + 2m = 2m+1 .
(d) This completes the proof if A = B. For the general case, apply the arguments above to (B,
B, B).
562
387I
Automorphisms
bi0 =
bi1 =
1
2r
for every i I.
proof (a) Let > 0 be such that
+ 6 4 .
< ln 2,
and
(385Ad). We have
H(A) = rq( ) = ln r,
and
d = rn whenever n N, d Dn (A, ).
Note that A is atomless. P
P?? If a A is an atom, then (because is ergodic) supjN j a = 1, and A is
purely atomic, with nitely many atoms all of the same size as a; but this means that H(C) ln( 1a ) for
every partition of unity C A, so that
1
n
1
n
ln(
1
)
=0
h(, C) = ln 2r ,
that h(, C ) ln 2r ; replacing C by C A if need be, we may suppose that C renes A; take such
a C of minimal size. Because H(C ) h(, C ) > H(A), there must be distinct c0 , c1 C 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
dt = t
c1 for every t [0, 1]. Let Dt be the partition of unity
387J
563
Ornsteins theorem
(C \ {c0 , c1 }) {c0 dt , c1 \ dt }
h(, D1 ) = h(, (C \ {c0 , c1 }) {c0 c1 }) < ln 2r ,
h(, D0 ) = h(, C ) ln 2r .
q(
((c0 ds ) (c0 dt ))) + q(
((c1 \ ds ) (c1 \ dt )))
= 2q(
(ds dt )) = 2q(|s t|
c1 )
1
n+1
q(
q(2) + q(1 2) ,
1
n
) + q(
)
n+1
n+1
cC
q(min(2,
c)) .
where
wn =
1
1P
ln( )d.
n dDn (C,)
(The Shannon-McMillan-Breiman theorem, 386E, assures us that any suciently large n has these properties.)
(d) Let D be the set of those d Dn (C, ) such that
d (2r)n ,
i.e.,
1
n
ln(
1
)
ln 2r.
Then
(sup D) 1 , by the choice of n, because h(, C) = ln 2r . Note that every member of
D is included in some member of Dn (A, ), because C renes A. If b Dn (A, ), then
b = rn , so
n
n
#({d : d D, d b}) 2 ; we can therefore nd a function f : D {0, 1} such that f is injective on
{d : d D, d b} for every b Dn (A, ).
(e) By 386C(iv), as usual, there is an a A such that a, 1 a, . . . , n+1 a are disjoint and
(a d) =
for every d Dn (C, ). Set
1
d
n+1
e = supdD,j<n j (a d);
then
e =
Pn1 P
j=0
dD
(a d) =
(sup D)
n+1
(1 )2 1 2.
(f ) Set
c = sup{ j (a d) : j < n, d D, f (d)(j) = 1}.
(I am identifying members of {0, 1}n with functions from {0, . . . , n 1} to {0, 1}.) Set
A = A {c , 1 \ c },
564
387J
Automorphisms
and let A be the closed subalgebra of A generated by { j a : a A , j Z}. Then a d A for every
d D. P
P Set
d = upr(a d, A ) = inf{c : c A , c a d} A .
Let b be the element of Dn (A, ) including d. Because a, b, e A ,
d a b e = supd D a b d = sup{a d : d D, d
b}.
Now if d D, d b and d 6= d, then f (d ) 6= f (d). Let j be such that f (d )(j) 6= f (d)(j); then j (a d)
is included in one of c , 1 \ c and j (a d ) in the other. This means that one of j c , 1 \ j c is a member
of A including a d and disjoint from a d , so that d d = 0. Thus d must be actually equal to a d, and
a d A . Q
Q
Next, c e A for every c C. P
P h j (a d)ij<n, dD is a disjoint family in A with supremum e. But
whenever d D and j < n we must have d j c for some c C, so either d j c or d j c = 0; thus
j (a d) must be either included in c or disjoint from it. Accordingly
c e = sup{ j (a d) : j < n, d D, d j c} A . Q
Q
Consequently h(, A ) ln 2r 2. P
P For any c C,
1
3
(c, A )
(c (c e)) =
(c \ e) min(
c, 2) ,
so
ln 2r = h(, C) h( A ) + H(C|A )
(386Ld)
h(, A ) +
(by the Kolmogorov-Sina theorem and 386Mb)
h(, A ) +
q((c, A ))
q(min(
c, 2)) h(, A ) +
cC
cC
by the choice of . Q
Q
Finally, h(, A ) ln 2r 4. P
P
ln 2r 2 h(, A ) h(, A ) + H({a, 1 \ a}) + H({e, 1 \ e})
= h(, A ) + q(
a) + q(1
a) + q(
e) + q(1
e)
1
n
h(, A ) + q( ) + q(
n
) + q(2) + q(1 2)
n+1
h(, A ) + + = h(, A ) + 2. Q
Q
(g) We have
ln 2r 4 h(, A ) H(A )
so
q(
c ) + q(1
c ) = H({c , 1 \ c }) ln 2 4.
By the choice of , |
c 21 | .
Next,
P
iI
|
(ai c )
1
1
| + |
(ai \ c ) |
2r
2r
3.
387K
565
Ornsteins theorem
P
P By 386I,
X
iI
1
r
1
r
|
(ai c )
c | + |
(ai \ c )
(1 \ c )|
So
X
iI
|
(ai c )
X
iI
1
1
| + |
(ai \ c ) |
2r
2r
1
r
1
r
1
2
c | + |
c |
|
(ai c )
1
1
1
+ |
(ai \ c )
(1 \ c )| + |
(1 \ c ) |
r
1
2
1
2
(1 \ c ) | 3. Q
Q
+ |
c | + |
(h) Now apply 387D to the partition of unity A , indexed as haij iiI,j{0,1} , where ai1 = ai c and
1
= ai \ c , and hij iiI,j{0,1} , where ij = 2r
for all i, j. We have
P
aij ij | 3
iI,j{0,1} |
ai0
by (g), while
H(A ) h(, A ) ln 2r ln 2r + 4 = 4,
so
iI,j{0,1}
Also
|
aij ij | +
P
iI,j{0,1}
2(H(A ) h(, A )) 3 +
8 4.
q(ij ) = ln 2r h().
So 387D tells us that there is a Bernoulli partition hbij iiI,j{0,1} for such that
bij =
P
(bij aij ) + 6 4 .
iI,j{0,1}
1
2r
Now of course
X
iI
X
iI
((ai c ) bi1 ) +
((ai \ c ) bi0 )
iI,j{0,1}
(aij bij ) ,
as required.
387K Ornsteins theorem (infinite entropy case) Let (A,
) be a probability algebra of countable
Maharam type, and : A A a two-sided Bernoulli shift of innite entropy. Then (A,
, ) is isomorphic
to (B, , ), where (B, ) is the measure algebra of the usual measure on [0, 1]Z , and is the standard
two-sided Bernoulli shift on B (385Sb).
proof (a) We have to nd a root algebra E for 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 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 inN in A0 such that {dn : n N} is
dense in the metric of A0 .
566
Automorphisms
387K
(b) There is a sequence hCn inN of partitions of unity in A0 such that Cn+1 renes Cn , H(Cn ) = n ln 2
and dn is a union of members of Cn+1 for every n. P
P We have
sup{H(C) : C A0 is a partition of unity} =
(385J). Choose the Cn inductively, as follows. Start with C0 = {0, 1}. Given Cn with H(Cn ) = n ln 2, set
Cn = Cn {dn , 1 \ dn }; then
H(Cn ) H(Cn ) + H({dn , 1 \ dn }) (n + 1) ln 2.
By 386O, there is a partition of unity Cn+1 , rening Cn , 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 { j a : a Cn , j Z}. Then
hCn inN is increasing. For each n, [Cn ] = Cn ; because Cn A0 , Cn is a Bernoulli shift with generating
partition Cn . Accordingly
h( Cn ) = h(, Cn ) = H(Cn ) = n ln 2.
Of course dn Cn+1 for every n.
Choose inductively, for each n N, n > 0, rn N and a Bernoulli partition hbn i{0,1}n in Cn , as
follows. Start with b0 = 1. (See 3A1H for the notation I am using here.) Given that hbn i{0,1}n is a
Bernoulli partition for which generates Cn , in the sense that Cn is the closed subalgebra of A generated
by { j bn : {0, 1}n , j Z}, and
bn = 2n for every , take n > 0 such that
(2rm + 1)n 2n for every m < n.
We know that
h( Cn+1 ) = (n + 1) ln 2 = ln(2 2n ).
So we can apply 387J to (Cn+1 , Cn+1 ) to see that there is a Bernoulli partition hbn i {0,1}n+1 for such
that
bn Cn+1 ,
for every {0, 1}n+1 ,
bn = 2n1
for every {0, 1}n . By 387H (with B = C = Cn+1 ), there is a Bernoulli partition hbn+1, i {0,1}n+1 for
Cn+1 such that the closed subalgebra generated by { j bn+1, : {0, 1}n+1 , j Z} is Cn+1 ,
bn+1, =
2n1 for every {0, 1}n+1 , and
P
(bn+1, bn ) n .
{0,1}n+1
(n+1)
(dm , Brn
) 2n for every m n.
Continue.
(d) Fix m n N for the moment. For {0, 1}m , set
(If n = m, then of course is the unique member of {0, 1}m extending itself, so this formula is safe.) Then
Next, if , {0, 1}m are distinct, there is no member of {0, 1}n extending both, so bn bn = 0; thus
hbn i{0,1}m is a partition of unity. If (0), . . . , (k) {0, 1}m , then
387K
567
Ornsteins theorem
( inf j bn,(j) ) =
(
jk
sup
(0),... , (k){0,1}n jk
(j)(j)jk
(0),... , (k){0,1}
(j)(j)jk
(0),... , (k){0,1}
(j)(j)jk
jk
(2n )k+1
k
Y
bn,(j) ,
j=0
P
P We have
X
{0,1}m
(bn bn+1, )
=
{0,1}m
(bn bn+1, ) 2n .
(bn
bn+1, )
(bn
(bn+1, a 0 bn+1, a 1 ))
(bn
(bn, a 0 bn, a 1 )) +
{0,1}n
{0,1}n
{0,1}n
{0,1}n+1
(bn bn+1, )
2n n + n = 2n . Q
Q
{0,1}n
b = limn
bn = 2m ;
and if , {0, 1}m are distinct, then
b b = limn bn bn = 0,
= lim
jk
k
Y
bn,(j) =
j=0
k
Y
b(j) ,
j=0
S
(g) Let E be the closed subalgebra of A generated by mN {b : {0, 1}m }. Then E is atomless and
countably generated, so (E,
E) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. Now
Qk
(inf jk j ej ) = j=0
ej for all e0 , . . . , ek E. P
P Let > 0. For m N, let Em be the subalgebra of E
568
387K
Automorphisms
S
generated by {b : {0, 1}m }. hEm imN is non-decreasing, so mN Em is a closed subalgebra of A, and
must be E. Now the function
Qk
(a0 , . . . , ak )
(inf jk j aj ) j=0
aj : Ak+1 R
(bm b )
(bm , B )
{0,1}m
{0,1}m
(m+1)
So if b Brm
{0,1}m n=m
(bn bn+1, ) 2
n .
n=m
,
(b, B ) (2rm + 1)
(bm+1, , B )
{0,1}m+1
(386Nc)
2(2rm + 1)
It follows that, whenever m n in N,
n=m+1
(n+1)
n 2
2n = 2m+1 .
n=m+1
) + 2n+1 2n + 2n+1
by the choice of rn . Letting n , we see that (dm , B ) = 0, that is, dm B , for every m N. But
this means that A0 B , by the choice of hdm imN . Accordingly j [A0 ] B for every j and B must be
the whole of A.
(i) Thus is a two-sided Bernoulli shift with root algebra E; by 385Sc, (A,
, ) is isomorphic to (B, , ).
387L Corollary: Sinas theorem (general case) Let (A,
) be an atomless probability algebra, and
: A A a measure-preserving automorphism. Let (B, ) be a probability algebra of countable Maharam
type, and : B B a one- or two-sided Bernoulli shift with h() h(). Then (B, , ) is isomorphic to
a factor of (A,
, ).
proof (a) To begin with (down to the end of (b)) suppose that is two-sided. Let B0 be a root algebra
for . If B0 is purely atomic, then there is a generating Bernoulli partition hbi iiN for of entropy h().
By 387E, there is a Bernoulli partition hci iiN for such that
ci = bi for every i. Let C be the closed
subalgebra of A generated by { j ci : i N, j Z}. Now (C,
C, C) is a factor of (A,
, ) isomorphic to
(B, , ).
(b) If B0 is not purely atomic, then there is still a partition of unity hbi iiN in B0 of innite entropy.
Again, let C be the closed subalgebra of A generated by { j ci : i N, j Z}, where hci iiN is a Bernoulli
partition for such that
ci = bi for every i. Now C is a Bernoulli shift of innite entropy and C has
countable Maharam type, so 387K tells us that there is a closed subalgebra C0 C such that h k [C0 ]ikN
is independent and (C0 ,
C0 ) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. But
(B0 , B0 ) is a probability algebra of countable Maharam type, so is isomorphic to a closed subalgebra
k
C1 of C0 (332N).
S Of kcourse h [C1 ]ikN is independent, so if we take C1 to be the closed subalgebra of A
generated by kZ [C1 ], C1 will be a two-sided Bernoulli shift isomorphic to (385Sf).
(c) If is a one-sided Bernoulli shift, then 385Sa and 385Sc show that (B, , ) can be represented
in terms of a product measure on a space X N and the standard shift operator on X N . Now this extends
388A
Dyes theorem
569
naturally to the standard two-sided Bernoulli shift represented by the product measure on X Z , as described
in 385Sb (cf. 385Yf); so that (B, , ) becomes represented as a factor of (B , , ) where is a two-sided
Bernoulli shift with the same entropy as (since the entropy is determined by the root algebra, by 385R).
By (a)-(b), (B , , ) is isomorphic to a factor of (A,
, ), so (B, , ) also is.
Remark Thus (A,
, ) has factors which are Bernoulli shifts based on root algebras of all countablygenerated types permitted by the entropy of .
387X Basic exercises (a) Let (A,
) be a probability algebra, and : A A a one- or two-sided
Bernoulli shift. Show that n isSa Bernoulli shift for any n 1. (Hint: if A0 is a root algebra for , the
closed subalgebra generated by j<n j [A0 ] is a root algebra for n .)
for E N. Set X = N Z and let , be the product measures on X derived from 0 and 0 . Show that there
is a bijection f : X X such that is precisely the image measure f 1 and f is translation-invariant,
that is, f (x) = f (x) for every x X, where (n) = n + 1 for every n Z.
387Y Further exercises (a) Suppose that (A,
, ) and (B, , ) are probability algebras with onesided Bernoulli shifts, and that they are isomorphic. Show that they have isomorphic root algebras. (Hint:
apply the results of 333 to (A,
, [A]).)
387 Notes and comments The arguments here are expanded from Smorodinsky 71 and Ornstein 74.
I have sought the most direct path to 387I and 387K; of course there is a great deal more to be said (387Xc
is a hint), and, in particular, extensions of the methods here provide powerful theorems enabling us to show
that automorphisms are Bernoulli shifts. (See Ornstein 74.)
570
388A
Automorphisms
proof (i)(ii) Let hHk ikN be a sequence in which separates the points of X; we may suppose that
H0 = X. By 381I, there is a partition of unity han inZ in A S
such that c = n c for every c an , n Z.
For each n Z let En be such that En = an ; then Y0 = nZ En is conegligible. The transformation
f n induces n , so for any k N and n Z the set
Fnk = {x : f n (x) En Hk , g(x)
/ En H k }
{x : g(x) En Hk , f n (x)
/ En H k }
S
is negligible, and Y = g 1 [Y0 ] \ nZ,kN Fnk is conegligible. Now, for any x Y , there is some n such
that g(x) En , so that f n (x) En and {k : g(x) Hk } = {k : f n (x) Hk } and g(x) = f n (x). As Y is
conegligible, (ii) is satised.
n
(ii)(iii)
are supposing that A0 = {x : g(x)
/ x } is negligible.
S Fornx X, set x = {f (x) : n Z}; we
Set A = nZ g [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) x . P
P Induce on n. Of course g 0 (x) = x x . For
n
the inductive step to n + 1, g (x) x \ 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) x . Thus the induction
continues. Q
Q
Consequently g n (x) x 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) x . Q
Q
Thus {g n (x) : n Z} x 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 \ kN (g 1 [Hk ]f n [Hk ]),
S
S
for n Z. Then (ii) tells us that nZ En is conegligible, so nZ 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 a = n a for every a an . Since supnZ an = 1 in A, belongs to the full subgroup
generated by .
Remark Of course the requirement countably separated is essential here; for other measure spaces we can
have and actually equal without g(x) and f (x) being related for any particular x (see 343I and 343J).
388B Corollary Under the hypotheses of 388A, and generate the same full subgroup of Aut A i
{f n (x) : n Z} = {g n (x) : n Z} for almost every x X.
388C Extending some ideas from 381M-381N, we have the following fact.
Lemma Let (A,
) be a totally nite measure algebra, and : A A a measure-preserving automorphism;
let C be its xed-point subalgebra {c : c = c}. Let hdi iiI , hei iiI be two disjoint families in A such that
(c di ) =
(c ei ) for every i I and c C. Then there is a G , the full subgroup of Aut A generated
by , such that di = ei for every i I.
proof Adding d = 1 \ supiI di , e = 1 \ supiI ei to the respective families, we may suppose that hdi iiI ,
hei iiI are partitions of unity. Dene han inN inductively by the formula
an = supiI (di \ supm<n am ) n (ei \ supm<n m am ).
supiI ei \ supm<n m am
ej
for each j,
jI
= (an di ) ei = (an di )
388E
571
Dyes theorem
n=0
(c ei n an ) =
=
n=0
n=0
(c n (an di )) =
n=0
( n (c an di ))
(c an di ) =
(c di \ a) <
(c di ) =
(c ei ).
ei \
supm<n m am .
572
388E
Automorphisms
and check that f and g are both continuous and that f g and gf are both the identity function. (ii) To see
that f is inverse-measure-preserving, it is enough to check that {x : f (x)(i) = z(i) for every i n} = 2n1
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}.
(Induce on k. For the inductive step, observe that if we identify X with {0, 1} X then f 2 (, y) = (, f (y))
for every {0, 1} and y X.)
Let : A A be the corresponding automorphism, setting E = f 1 [E] for E . Then is
a von Neumann automorphism. P
P Set En = {x : x X, x(i) = 1 for every i < n}, an = En . Then
n
2n
[En+1 ] = {x : x(i) = 1 for i < n, x(n) = 0}, so an+1 and 2 an+1 split an for each n, and han inN
f
witnesses that is weakly von Neumann. Next, inducing on n, we nd 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 -generate A (see part (a) of the proof of 331K), is a von
Neumann automorphism. Q
Q
f is sometimes called the odometer transformation. For another way of looking at the functions f
and g, see 445Xp in Volume 4.
388F We are now ready to approach the main results of this section.
Lemma Let (A,
) be a totally nite measure algebra and : A A an aperiodic measure-preserving
automorphism. Let C be its xed-point subalgebra. Then for any a A there is a b a such that
(b c) =
1
(a
c)
for
every
c
C
and
is
a
weakly
von
Neumann
automorphism,
writing
for
the
induced
b
b
2
automorphism of the principal ideal Ab , as in 381M.
Remark On rst reading, there is something to be said for supposing here that is ergodic, that is, that
C = {0, 1}.
proof I should remark straight away that is doubly recurrent on every b A (386A), so we have an
induced automorphism b : Ab Ab for every b A (381M).
(a) Set n = 21 (1 + 2n ) for each n N, so that hn inN is strictly decreasing, with 0 = 1 and
limn n = 12 . Now there are hbn inN , hdni inN,i<2n such that, for each n N,
bn+1
bn
a,
(bn c) = n
(a c) for every c C,
supi<2n dni = bn ,
(dn0 c) = 2n
(bn c) = 2n n
(a c)
388G
573
Dyes theorem
n
(applying 386C(iii) to b2n Adn0 , with = n+1 /2n ). Set dn+1,j = bjn dn+1,0 for each j < 2n+1 . Ben
cause b2n dn+1,0 dn0 \ dn+1,0 , while hbjn dn0 ij<2n is disjoint, hbjn dn+1,0 ij<2n+1 is disjoint. Set bn+1 =
(bn+1 c) = n+1
(a c) for every c C. For j < 2n+1 ,
supi<2n+1 bi n dn+1,0 ; then bn+1 bn and
n
dn+1,j dni where i is either j or j 2 , so bn+1 dni = dn+1,i dn+1,i+2n .
For i < 2n+1 1,
bn dn+1,i = dn+1,i+1
bn+1 ,
b = inf nN bn ,
Because hbn inN is non-increasing,
1
2
(a c)
(b c) = limn
(bn c) =
for every c C. Next,
whenever i < 2 .
If m n, j < 2m then
(induce on n). So
bn ,
and
bn e = k e
1 \ dm,j+1 .
574
388G
Automorphisms
a,
b2
b3 b2 ,
n m b1 = n a2
k
b.
m b1 = 1 a2
a1
b3 .
Consider a = m b3 .
a;
so that
d = n ( n m ) m d = n k m d = n+rm d.
As a is arbitrary, this shows that G , so that G G and the two are equal.
388H Lemma Let (A,
) be a totally nite measure algebra, : A A an aperiodic measure-preserving
automorphism, and any member of the full subgroup G of Aut A generated by . Suppose that
hdmi imn,i<2m is a nite dyadic cycle system for . Then there is a weakly von Neumann automorphism ,
with dyadic cycle system hdmi imN,i<2m , such that G = G , a = a whenever a dn0 = 0, and dmi = dmi
whenever m n and i < 2m .
proof Write C for the closed subalgebra {c : c = c}. By 388F there is a b dn0 such that
(b c) =
1
(d
c)
for
every
c
C
and
:
A
A
is
a
weakly
von
Neumann
automorphism.
Let
he
i
k be
n0
b
b
b
ki
kN,i<2
2
a dyadic cycle system for b .
If we dene 1 Aut A by setting
1 d = b d for d
b,
1 d = 1\b d for d
(2
n
+1
b c) =
2
+1
1 \ b,
1
2
(dn0 c) =
((dn0 \ b) c)
(b c) =
(b c) =
+1
21 2
+1
b = 2
2 +1
+1
+1
b. Set
(dn0 \ b).
d = 3 d if d
b,
= 2 d if d
dn0 \ b,
= d if d dn0 = 0.
Since dn0 = dn1 , we have dni = dni for every i < 2n , and therefore i dm0 = dmi whenever m n
n
and i < 2m . Looking at 2 , we have
n
n+1
3 b = dn0 \ b,
b = b. Accordingly
2
i
2 b = 2
n+1
d = 2
2 2
3 d = 1 d = b d
388I
575
Dyes theorem
m
2 am+1 = ( 2
n+1
)2
mn1
emn,0 = (b )2
mn1
emn,0
(e c) =
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 ,
Aa ) is
a totally nite 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 : Cw Ca be the Boolean homomorphism dened by setting c = c b for c Cw . If c Cw and
c = 0, then c C and
c
(c b) = 0, so c = 0; thus is injective; since it is certainly surjective, it is
a Boolean isomorphism. We can therefore dene a functional =
1 : Ca [0, [, and we shall have
d
d for every d Ca . By 331B, there is an e Aa such that d =
(d e) for every d Ca , that is,
c =
(c e) for every c Cw , as required.
c for every c C and bi e Ce for
(b)(i) Write D for the set of all those e A such that
(c e) = k1
r+1
every i r. Then whenever a A and > k is such that (a c) = c for every c C, there is an e D
such that e a. P
P For d A, c C set d (c) =
(d c), so that d : C [0, [ is a completely additive
c k(a bi c)
functional. For i r set vi = [[
C > kabi ]], in the notation of 326P; so that vi C and
whenever c C and c vi , while
c k
(a bi c) whenever c C and c vi = 0. Setting v = inf ir vi ,
we have
Pr
k
v = k
(a v) = i=0 k(a bi v) (r + 1)
v.
(c e) =
Pr
i=0
(c ei ) =
Pr
i=0
(c wi ei ) =
Pr
(c wi )
i=0 k
1
k
=
c
(ii) Let E0 D be a maximal disjoint family, and set m = #(E0 ), a = 1 \ sup E0 . Then
P
m
(c e) = (1 )
(a c) =
c eE0
c
k
576
388I
Automorphisms
r+1
for every c C, while a does not include any member of D. By (i), 1 m
k k , that is, k m r 1.
Applying (a) repeatedly, with w = 1 and = k1 , we can nd disjoint d0 , . . . , dkm1 a such that
(c di ) = k1
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.
(e c) = 2k
c for every e E, c C,
E1 = {e : e E, there is some i r such that bi e
/ Ce }
n
(vi c) =
(2
+1
(sup E) = dn0 ,
vi c) = 2k
c
+1
1 v2kn 1 = 2
+1
v0
(388C). We have
1 dn0 = 1 ( sup vi ) = sup 1 vi =
i<2kn
= sup 2
i<2kn
n
+1
vi = 2
+1
sup
+1
i<2kn 1
vi+1 2
+1
v0
i<2kn
dn0 ,
= d if d dn0 = 0.
(c) For each i < 2kn ,
n
2 vi = 2
1 vi = vi+1
(identifying v2kn with v0 ). Moreover, vi dnl whenever i < 2kn and j l mod 2n . So h j v0 ij<2k is a
partition of unity in A. What this means is that if we set
dmj = sup{ i v0 : i < 2k , i j mod 2m }
for m k, then hdmj imk,j<2m is a dyadic cycle system for , with dmj = dmj if m n, j < 2m .
(d) Let B be the subalgebra of A generated by C {dkj : j < 2k }. Recall the denition of {vi : i < 2kn }
n
as {2 1 e : e E}; this implies that
388K
577
Dyes theorem
+1
vi : i < 2kn } = E,
so that
{ j+1 vi : i < 2kn } = {j e : e E}
j a},
so a
a. Next, dn1 j (a \ a )
I = {i : i r, (i, j) K} = {i : bi
bi = supeE (bi e)
j a};
supeE0 (e cei ) e ,
so that
dn1 j a = supiI bi
(dn,j+1 a \ a ) =
(dn1 j (a \ a ))
e 2n
(a a ) = j=1
578
Automorphisms
388K
Now use 388J to nd a n+1 G , with a dyadic cycle system hdn+1,m,i imkn+1
,i<2m , such that kn+1
kn , dn+1,m,i = dnmi if m kn , n+1 a = n a if a dn,kn ,0 = 0, and there is a bn in the algebra generated by
C {dn+1,m,i : m kn+1
, i < 2m } such that
(bn bn ) 2n . Continue.
(d) The eect of this construction is to ensure that if l < n in N then
dlmi = dnmi whenever m kl , i < 2m ,
n a = l a whenever a dl,kl ,0 = 0,
bl belongs to the subalgebra generated by C {dnmi : m kn , i < 2m },
dl,kl ,0 . Since hkn inN is strictly increasing, inf nN dn,kn ,0 = 0. Now, for each n N,
we have
0 a0 = 1 \ d0,k0 ,1 ,
= C {dmi : m kn , i < 2m } B
for any n N. So bn B for every n. If b B and > 0, there is an n N such that 2n and bn = b,
so that
(b bn ) ; as every bn belongs to B, and B is closed, b B; as b is arbitrary, and B -generates
A, B = A. Thus is a relatively von Neumann automorphism.
(f ) If n N and d en = 0, then j d = jn d and j d = 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 d dn,kn ,1 = 0 then
1
n d dn,kn ,0 = 0, so
1
(1
1 d = 1 n (1
n d ) = n d .
n d )=
Now we have
j+1 d = (jn d) = n (jn d) = j+1
n d
388L
Dyes theorem
579
because
jn d dn,kn ,0 = jn (d j
n dn,kn ,0 ) = 0,
while
j1
1 j
d
j1 d = 1 (j
n d) = n (n d) = n
because
j
j+1
j
Q
n d dn,kn ,1 = n (d n dn,kn ,0 ) = 0. Q
2 d = dj c = 1ji d.
580
Automorphisms
388L
Now supnZ 2n d0 belongs to C and includes d0 , so must be 1. Finally, the two induced automorphisms
P If 0 6= d d0 there are a non-zero d d and an m 1
(1 )d0 , (2 )d0 on Ad0 are both the identity. P
m
such that (2 )d0 d = 2 d for every d d . As 2m G1 , there are a non-zero d d and a k Z such that
2m d = 1k d. Now 1k d d0 so k is a multiple of n and (2 )d0 d = 1k d = d. This shows that {d : (2 )d0 d = d}
is order-dense in Ad0 and must be the whole of Ad0 . As for 1 , we have (1 )d0 d = 1n d = d for every d d0 .
Q
Q
So 388G tells us that G1 = G2 .
(d) For the general case, we see from 381H that there is a partition of unity hci i1i in C such that
1 Ac is aperiodic and if i is nite and ci 6= 0 then 1 Aci is periodic with period i. For each i, let Hi be
{ Aci : G1 }; then Hi is a full subgroup of Aut Aci , and
Similarly, writing
Hi
= { Aci : G2 },
Note also that Hi , Hi are the full subgroups of Aut Aci generated by 1 Aci , 2 Aci respectively. By (b)
and (c), Hi = Hi for nite i, while there is a measure-preserving automorphism : Ac Ac such that
H 1 = H . Now we can dene a measure-preserving automorphism 1 : A A by setting 1 a = a
, G2 ) are
, G1 ) and (A,
if a c , 1 a = a if a c = 0, and we shall have 1 G1 11 = G2 . Thus (A,
isomorphic, as claimed.
388X Basic exercises > (a) Let (A,
) be a totally nite measure algebra, and : A A a measurepreserving automorphism. Let us say that a pseudo-cycle for is a partition of unity hai ii<n , where
n 1, such that ai = ai+1 for i < n 1 (so that an1 = a0 ). (i) Show that if we have pseudo-cycles
hai ii<n and hbj ij<m , where m is a multiple of n, then we have a pseudo-cycle hcj ij<m with c0 a0 , so that
ai = sup{cj : j < m, j i mod n} for every i < n. (ii) Show that is weakly von Neumann i it has a
pseudo-cycle of length 2n for any n N.
(b) Let (A1 ,
1 ) and (A2 ,
2 ) be probability algebras, and 1 : A1 A2 and 2 : A2 A2 measurepreserving von Neumann automorphisms. Show that there is a measure-preserving Boolean isomorphism
: A1 A2 such that 2 = 2 1 .
(c) Let (A,
) be an atomless probability algebra of countable Maharam type, and Aut A the group of
measure-preserving Boolean automorphisms of A. Let Aut A be a von Neumann automorphism. (i)
Show that for any ultralter F on N there is a F Aut A dened by the formula F (a) = limnF n a for
every a A, the limit being taken in the measure-algebra topology. (ii) Show that {F : F is an ultralter
on N} is a subgroup of Aut A homeomorphic to ZN
2 . (Hint: 388E.)
(d) Let A be a Boolean algebra and : A A a weakly von Neumann automorphism. Show that 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 388E may be useful.)
(e) Let A be a Boolean algebra and : A A a von Neumann automorphism. (i) Show that is ergodic
but 2 is not ergodic. (ii) Show that 2 is relatively von Neumann. (iii) Show that n is von Neumann for
every odd n Z.
388Y Further exercises (a) Let X be a set, a -algebra of subsets of X, and I a -ideal of such
that the quotient algebra A = /I is Dedekind complete and there is a countable subset of separating
the points of X. Suppose that f and g are automorphisms of the structure (X, , I) inducing , Aut A.
Show that the following are equiveridical: (i) belongs to the full subgroup of Aut A generated by ; (ii)
{x : x X, f (x)
/ {g n (x) : n Z}} I; (iii) {x : x X, {f n (x) : n Z} 6 {g n (x) : n Z}} I.
(b) Let (A,
) be a totally nite measure algebra and : A A a relatively von Neumann automorphism.
Show that is aperiodic and has zero entropy.
388 Notes
Dyes theorem
581
582
Automorphisms
388 Notes
between the two contexts. I oer 381Xp as an example. The description there of induced automorphism
requires a certain amount of manoeuvering around negligible sets, but gives a valuably graphic description.
In the same way, 381Xo, 388A and 381Qc 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
actually belongs to the full semigroup generated by (381Ya); for instance, this happens in 388C.
391C
Kelleys theorem
583
Chapter 39
Measurable algebras
In the nal 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-nite 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 Maharams theorem (332) gives us
the answer exactly: they are the countable simple products of the measure algebras of {0, 1} for cardinals
. But if we approach from another direction, things are more interesting. Probability algebras share a very
large number of very special properties. Can we nd a selection of these properties which will be sucient
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 (, )-distributivity is necessary and sucient for the existence
of a measure; so we are led to look for conditions sucient to ensure that there is a strictly positive
additive functional. A slightly dierent 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 i 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 terms of the special
properties which the automorphism group of a measure algebra must have (Kawadas theorem, 394Q). 395
complements the previous section by looking briey at the subgroups of an automorphism group Aut A
which can appear as groups of measure-preserving automorphisms.
584
391C
Measurable algebras
a = cC
c (a c) for every a A; then it is easy to check that (A,
) is a semi-nite 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 -complete and weakly (, )-distributive, and there is a strictly positive nitely
additive functional : A [0, [.
Remark An additive functional on a Boolean algebra A is strictly positive if a > 0 for every non-zero
a A.
proof (i)(ii) Put the denition together with 322C(b)-(c) (for Dedekind completeness) and 322F (for
weak (, )-distributivity).
(ii)(i) Given that (ii) is satised, let M be the L-space of bounded additive functionals on A, M M
the band of completely additive functionals, and P : M M the band projection (362B). Set
= P ().
Then
is strictly positive. P
P If c A is non-zero, there is an upwards-directed set A, with supremum c,
such that
c = supaA a (362D); as is strictly positive and A contains a non-zero element,
c > 0. Q
Q Of
course
is countably additive, so it witnesses that 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 : A [0, 1] a functional. Then the following are
equiveridical:
(i) there is a nitely additive functional : A [0, 1] such
P that 1 = 1 and a a for every a A;
(ii) whenever hai iP
iI is a nite indexed family in A, and
iI ai m1 in S = S(A) (denition: 361A),
where m N, then iI ai m.
proof (a)(i)(ii) If : A [0, 1] is a nitely additive functional such that 1 = 1 and a a for every
a A, let h : S
P R be the positive linear functional corresponding to (361G). Now if hai iiI is a nite
family in A and iI ai m1, then
X
X
X
ai
ai =
h(ai )
iI
iI
iI
X
= h(
ai ) h(m1) = m.
iI
Then it is easy to check that p(u + v) p(u) + p(v) for all u, v S, and that p(u) = p(u) for
u S,
Pall
n
P
0. Also p(1) 1. P
P?? If not, there are a0 , . . . , an A and 0 , . . . , n 0 such that 1 i=0 i ai
n
but i=0 i ai < 1. Increasing each i slightly if necessary, we may suppose that every i is rational; let
m 1 and k0 , . . . , kn N be such that i = 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
P
Pn
Pn
(i,j)K aij =
i=0 ki ai = m
i=0 i ai m1,
but
(i,j)K
aij =
Pn
i=0
ki ai = m
Pn
i=0
i ai < m,
391G
585
Kelleys theorem
1 1 + 1 1 = 1,
391F Theorem Let A be a Boolean algebra, not {0}, and : A [0, 1] a functional, where A A.
Then the following are equiveridical:
(i) there is a non-negative nitely additive functional : A [0, 1] such that 1 = 1 and a a for
every a A;
P
(ii) whenever hai iiI is a nite indexed family in A, there is a set J I such that #(J) iI ai and
inf iJ ai 6= 0.
Remark In (ii) here, we may have to interpret the inmum of the empty set in A as 1.
proof We apply 391E to the functional , where a = 1 (1 \ a) for a A.
(a) If : A [0, 1] is a non-negative nitely additive functional such that 1 = 1, then
a a 1 a 1 a (1 \ a) (1 \ a).
iI
ai m1,
so there is a set J I such that #(J) #(I) and inf iJ (1 \ ai ) = c 6= 0. Now c ai = 0 for i J, so
P
P
mc iI (ai c) = iI\J (ai c) #(I \ J)c
(c) Suppose that 391E(ii) is true of , and that hai iiI is a nite family in A. Set
P
P
= iI (1 \ ai ) = #(I) iI ai
P
P
and
iI (1 \ ai ) < k,
iI (1 \ ai ) 6 k1, that is,
P let k be the least integer greater than . Since
iI ai 6 (#(I) k)1. But this means that there must be some J I such that #(J) > #(I) k and
inf iJ ai 6= 0. Now
P
iI ai = #(I) #(I) (k 1) #(J).
(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 0 : B R a non-negative
nitely additive functional. Then there is a non-negative nitely additive functional : A R extending
0 .
proof (a) Suppose rst that 0 1 = 1. Set b = 0 b for every b B. Then must satisfy the condition (ii)
of 391F when regarded as a functional dened on a subset of B; but this means that it satises the same
condition when regarded as a functional dened on a subset of A. So there is a non-negative nitely additive
functional : A R such that 1 = 1 and b 0 b for every b B. In this case
586
391G
Measurable algebras
b = 1 (1 \ b) 1 0 (1 \ b) = 0 b b
(b) For the general case, if 0 1 = 0 then 0 must be the zero functional on B, so we can take to be the
zero functional on A; and if 0 1 = > 0, we apply (a) to 1 0 .
391H Definition Let A be a Boolean algebra, and A A \ {0} any non-empty set. The intersection
number of A is the largest 0 such that whenever hai iiI is a nite family in A, with I 6= , there is a
J I such that #(J) #(I) and inf iJ ai 6= 0.
Remarks (a) It is essential to note that in the denition above the hai iiI are indexed families, with
repetitions allowed; see 391Xh.
(b) I spoke perhaps rather glibly of the largest such that . . . ; you may prefer to write
= inf{sup6=J{0,... ,n},inf jJ aj 6=0
#(J)
n+1
: 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 nitely additive functionals : A [0, 1] such that 1 = 1. Then the intersection number
of A is precisely maxC inf aA a.
proof Write for the intersection number of A, and for supC inf aA a.
(a) For any < , we can nd a C such that a for every a A. So if we set a = for every
a A, satises condition (i) of 391F. But this means that if hai iiI is any nite family in A, there must
be a J I such that inf iJ ai 6= 0 and #(J) #(I). Accordingly ; as is arbitrary, .
(b) Dene : A [0, 1] by setting a = for every a A. If hai iiI isP
a nite indexed family in A,
there is a J I such that #(J) #(I) and inf iJ ai 6= 0; but #(I) = iI ai , so this means that
condition (ii) of 391F is satised. So there is a C such that a for every a A; and witnesses not
only that , 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 nitely 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 nitely additive functional on A, and A 6= {0}, set An =
1
{a : a 2n 1} for every n N; then (applying
S 391I to the functional 1 ) we see that every An has
n
intersection number at least 2 , while A \ {0} = nN An because is strictly positive, so (ii) is satised.
S
(ii)(i) If A \ {0} is expressible as nN An , where each An has intersection number n > 0, then for
each nPchoose a nitely additive functional n on A such that n 1 = 1, n a n for every a An . Setting
a = n=0 2n n a for every a A, is a strictly positive additive functional on A, and (i) is true.
391K Corollary Let A be a Boolean algebra. Then A is measurable i it is Dedekind -complete and
weakly (, )-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 Kelleys criterion. It provides some sort of answer to the question which Boolean algebras carry strictly positive nitely additive
functionals?, but leaves quite open the possibility that there is some more abstract criterion which is also
necessary and sucient. It is indeed a non-trivial exercise to nd any ccc Boolean algebra which does not
carry a strictly positive nitely additive functional. The rst example published seems to have been that of
Gaifman 64; I present this in a modied form, following Comfort & Negrepontis 82. First, a denition:
391N
587
Kelleys theorem
(b) I show rst that there is no strictly positive nitely additive functional on A. P
P?? If there were,
there would be a sequence hAk ikN 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 kN Tk = R. By Baires theorem (3A3G) there is a
k N such that G = int T k 6= .
Let > 0 be the intersection number of Ak . There must be some n 1/ 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 iJJn . Because the intersection number of Ak is , there must
T be a set K Jn ,
2
=
6
,
that
is,
X
of cardinal
at
least
#(J
)
=
(n
+
1)
>
n
+
1,
such
that
inf
e
n
JK
t
J
JK etJ 6= . But
T
if x JK 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 Kelleys criterion and there is no strictly positive nitely additive functional
dened on A.
(c)(i) Write F for the set of functions f such that dom f is a nite 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},
a}.
588
Measurable algebras
f (t) = 1 t
391N
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 . Dene 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
#({J : J Jn , Qz J 6= }) #(Qz ) 2k n + 1
391 Notes
Kelleys theorem
589
(i) Let A be a Boolean algebra. For non-empty A A \ {0} write (A) for the intersection number of A.
Show that for any non-empty A A \ {0}, (A) = sup{(I) : I is a non-empty nite subset of A}.
(j) (i) Show that any subalgebra of a -linked Boolean algebra is -linked. (ii) Show that if A is a Boolean
algebra with a -linked order-dense subalgebra, then A is -linked. (iii) Show that the simple product of
a countable family of -linked Boolean algebras is -linked. (iv) Show that a principal ideal of a -linked
Boolean algebra is -linked.
391Y Further exercises (a) Show that in 391D and 391K we can replace weakly (, )-distributive
by weakly -distributive.
(b) Show that PN is -linked and chargeable but that the quotient algebra PN/[N]< is not ccc, therefore
neither -linked nor chargeable.
(c) (i) Show that if X is a separable topological space, then its regular open algebra is chargeable. (ii) Let
hXi iiI 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 -linked Boolean algebra can have cardinal at most c. (Hint: take a sequence hAn inN
of linked sets covering A \ {0}. Set An = {a : b An , b a}. Show that if a 6 b there is an n such that
/ An .)
a An and b
(e) Show that the free product of c or fewer -linked Boolean algebras is -linked. (Hint: let hAi iiI be
a family of -linked Boolean algebras, where I R. For each i I, let hAin inN 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 iJ i (ai ) where J I [q0 , qr [ is a set meeting each interval [qk1 , qk [ in at most one point
and ai Ai,nk if i J [qk1 , 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 -linked but not chargeable.
(g) Let A be the algebra of 391N. Show that A is not weakly (, )-distributive. (Hint: show that for
any t R there is a strictly decreasing sequence hti iiN , converging to t, such that x 7 hx(ti )iiN 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 316Xp 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 -m-linked if A \ {0} can be expressed as the union of a sequence of m-linked
sets. (Thus linked is 2-linked and -linked is -2-linked.) Show that the algebra of 391N is -m-linked
for every m 2. (Hint: in part (c) of the proof, replace each 2k by mk.)
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. Kelleys 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 Gaifmans example (391N) is -linked in order to show that even conditions
very much stronger than ccc are not sucient to guarantee the existence of suitable functionals. (See also
391Yh.) For other examples see Comfort & Negrepontis 82. But none of the standard examples is
weakly (, )-distributive (see 391Yg), for reasons which I hope to return to in Volume 5.
Since every -linked algebra has cardinal at most c (391Yd), not every measurable algebra is -linked.
In fact it is known that a measurable algebra of cardinal c or less is -m-linked (391Yh) for every m 2
ns 93).
(Dow & Stepra
590
Measurable algebras
392 intro.
392 Submeasures
In 391 I looked at what we can deduce if a Boolean algebra carries a strictly positive nitely additive
functional. There are important contexts in which we nd 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 nite combinatorics (392D).
On any lattice, we can dene a topology from the notion of order*-convergence of sequences introduced in
367 (392K). It turns out that on Boolean algebras these topologies are intimately connected with Maharam
submeasures (392O).
392A Definition Let A be a Boolean algebra. A submeasure on A is a functional : A [0, [ such
that
(a b) a + b for all a, b A;
a b whenever a b;
0 = 0.
(In this context I do not allow as a value of a submeasure.) Any positive nitely additive functional is
a submeasure (326Ba, 326Bf).
392B Definitions Let A be a Boolean algebra and : A [0, [ a submeasure. Then
(a) is strictly positive if a > 0 for every a 6= 0;
(b) is exhaustive if limn an = 0 for every disjoint sequence han inN in A;
(c) is uniformly exhaustive if for every > 0 there is an n N such that there is no disjoint family
a0 , . . . , an with ai 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 : A [0, [ is an exhaustive strictly positive submeasure, and hai iiI
is a disjoint family in A \ {0}, then {i : ai 2n } must be nite for each n, so I is countable. (Cf. 322G.)
If : A [0, [ is a uniformly exhaustive submeasure and han inN is disjoint in A, then {i : ai 2n } is
nite for each n, so limi ai = 0. If P: A [0, [ is a positive linear functional, and > 0, then take
n
n 1 1; if a0 , . . . , an are disjoint, then i=0 ai 1, so minin ai < .
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 members (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! 3n 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 be the set of those R M L such that each vertical section of R has just three members, so
that
l! m
#() = #([L]3 )m =
.
3!(l3)!
(I write [X] for the set of subsets of X with j members.) Let us regard as a probability space with the
uniform probability.
If F [L]n , where 3 n k, and x M , then
392E
591
Submeasures
Pr(R[{x}] F ) =
#([F ]3 )
#([L]3 )
n(n1)(n2)
l(l1)(l2)
n3
.
l3
xE
Pr(R[{x}] F )
n3n
.
l3n
Accordingly
k
X
X
n=3
Pr(R[E] F )
l!
n3n
m!
n!(mn)! n!(ln)! l3n
k
X
n=3
k
X
#([M ]n )#([L]n )
n=3
mn ln n3n
n!n!l3n
n3n
l3n
k
X
mn nn 32n
n=3
l2n
(using (a))
=
k
X
9mn n
n=3
l2
k
X
n=3
1
2n
< 1.
There must therefore be some R 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 Halls 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 : A [0, [ a uniformly exhaustive submeasure. Then
for any ]0, 1] the set A = {a : a } has intersection number greater than 0.
proof (a) If 1 = 0 this is trivial, so we may assume that 1 > 0; since neither the hypothesis nor
the conclusion is aected if we multiply by a positive scalar, we may suppose that 1 = 1. Because
is uniformly exhaustive, there
is a disjoint family in A then
P is an r 1 such that whenever hci iiI
1 2
, so that
#({i : ci > 51 }) r, so that iI ci r + 15 #(I). Set = /5r, = 74
1
18 (
)2
1
2
18 (
2) = 4.
(b) Let hai iiI be a non-empty nite family in A. Let m be any multiple of #(I) greater than or equal
to 1/. Then there are integers k, l such that
592
392E
Measurable algebras
k
m
1
(
18
)2 ,
l
m
in which case
18mk m2 ( )2 l2 .
3 k l m,
(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 members. By 392D, there is a set R M L such that every vertical section
of R has just three members 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 xE dx \ supxM \E dx ,
xM,jL
xM
1
5
cxj r + mj
1
5
cxj rl +
(r +
jL
3
5
mj rl + m
3
)m
5
4
5
= m < m
The set J = {i : s, (i, s) E} must therefore have more than 3#(I) members, since E J S. But also
d(i,s) = ai for each (i, s) E, so that inf iJ ai bE 6= 0.
(d) As hai iiI is arbitrary, the intersection number of A is at least 3 > 0.
392F Theorem Let A be a Boolean algebra with a strictly positive uniformly exhaustive submeasure.
Then A has a strictly positive nitely additive functional.
proof If A = {0} this is trivial. Otherwise, let : A [0, [ be a strictly positive uniformlySexhaustive
submeasure. For each n, An = {a : a 2n 1} has intersection number greater than 0, and nN An =
A \ {0} because is strictly positive; so A has a strictly positive nitely additive functional, by Kelleys
theorem (391J).
392J
Submeasures
593
392G Since positive additive functionals are uniformly exhaustive submeasures, the condition of this
theorem is necessary as well as sucient. Thus we have a description, in terms 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.
Definition Let A be a Boolean algebra. A Maharam submeasure or continuous outer measure on A
is a submeasure : A [0, [ such that limn an = 0 whenever han inN is a non-increasing sequence
in A with inmum 0.
392H Lemma Let A be a Boolean algebra and a Maharam submeasure on A.
(a) is sequentially order-continuous.
(b) is countably subadditive,
P that is, whenever han inN is a sequence in A and a A is such that
a = supnN a an , then a n=0 an .
(c) If A is Dedekind -complete, then is exhaustive.
proof (a) (Of course is an order-preserving function, by the denition of submeasure; so we can apply the
ordinary denition of sequentially order-continuous in 313Hb.) (i) If han inN is a non-decreasing sequence in
A with supremum a, then han \ ainN is a non-increasing sequence with inmum 0, so limn (an \ a) = 0;
but as
an a an + (a \ an )
for every n, it follows that a = limn an . (ii) If han inN is a non-increasing sequence in A with inmum
a, then
a an a + (an \ a) a
as n .
(b) Set bn = supin a ai ; then bn
Pn
(c) If han inN is a disjoint sequence in A, set bn = supin ai for each n; then inf nN bn = 0, so
lim supn an limn bn = 0.
392I Proposition Let A be a Dedekind -complete Boolean algebra and a strictly positive Maharam
submeasure on A. Then A is ccc, Dedekind complete and weakly (, )-distributive, and is ordercontinuous.
proof By 392Hc, is exhaustive; by 392Ca, A is ccc; by 316Fa, A is Dedekind complete.
Now suppose that we have a sequence hAn inN of non-empty downwards-directed subsets of A, all with
inmum 0. Let A be the set
{a : a A, n N a An such that a a}.
By 316Fc and 392Ha, is order-continuous, so inf aAn a = 0 for each n. Given > 0,Pwe can choose
han inN such that an An and an 2n for each n; now a = supnN an A and a n=0 an 2.
Thus inf aA a = 0. Since is strictly positive, inf A = 0. As hAn inN is arbitrary, A is weakly (, )distributive(316H).
392J Theorem Let A be a Boolean algebra. Then it is measurable i it is Dedekind -complete and
carries a uniformly exhaustive strictly positive Maharam submeasure.
proof If A is measurable, it surely satises the conditions, since any totally nite measure on A is also
a uniformly exhaustive strictly positive Maharam submeasure. If A satises the conditions, then it is
weakly (, )-distributive, by 392I, and carries a strictly positive nitely additive functional, by 392F; so is
measurable, by 391D.
594
392K
Measurable algebras
392K The problem of identifying the Boolean algebras which are measurable in the sense of 391B seems
to be hard. If we look for a strictly positive Maharam submeasure instead of a measure, then (at least for
Dedekind -complete spaces) there are interesting criteria in terms of the order*-convergence introduced in
367.
Definition Let P be a lattice, and consider the relation hpn inN order*-converges to p as a relation between
P N and P . By 367Bc, this satises the hypotheses of 3A3Pa, so there is a unique T1 topology on P for
which a set F P is closed i a F whenever han inN is a sequence in F which order*-converges to a in
P . I will call this topology the order-sequential topology of P .
ski &
Warning! For the next few paragraphs I shall be closely following the papers Balcar Glowczyn
k p03. I should therefore note explicitly that if A is a Boolean algebra
Jech 98 and Balcar Jech & Paza
which is neither Dedekind -complete nor ccc, my order-sequential topology on A may not be identical to
theirs.
392L Lemma Let A be a Boolean algebra.
(a) A sequence han inN order*-converges to a A i there is partition B of unity in A such that {n : n N,
(an a) b 6= 0} is nite for every b B.
(b) If han inN order*-converges to a and c A, then han cinN , han cinN and han cinN order*converge to a c, a c and a c respectively.
(c) For any c A, the maps a 7 a c, a 7 a c and a 7 a c are continuous for the order-sequential
topology.
proof (a) Let han inN be a sequence in A and a A; set
also has inmum 0 (313A, 313B). So there is a partition B of unity such that for every b B there is an
e E such that b e = 0. Now, given b B, there are c C and d D such that b d \ c = 0; there are
n1 , n2 N such that c an for n n1 and an d for n n2 ; so that {n : (an a) b 6= 0} is bounded
above by max(n1 , n2 ) and is nite. So B witnesses that the condition is satised.
(ii) Now suppose that B is a partition of unity such that {n : (an a) b 6= 0} is nite for every b B.
Then a (1 \ b) D for every b B, because {n : an 6 a (1 \ b)} {n : (an a) b 6= 0} is nite. So any
lower bound for D is also a lower bound for {a (1 \ b) : b B} and is included in a. Similarly, any upper
bound for C includes a; as c d whenever c C and d D, sup C = inf D = a and han inN a.
(b) These are all immediate from (a), because
(an c) (a c)
an a,
(an c) (a c)
an a,
(an c) (a c) = an a
for every n.
(c) By (b), we can apply 3A3Pb to each of these three functions.
392M Lemma Let A be a ccc weakly (, )-distributive Boolean algebra, endowed with its ordersequential topology.
(a) If hamn im,nN , ham imN and a are such that hamn inN order*-converges to am for each m, while
ham imN order*-converges to a, then there is a sequence hk(m)imN in N such that ham,k(m) imN order*converges to a.
(b) If A A and a A, there is a sequence in A which order*-converges to a.
392N
Submeasures
595
(b) Let A be the set of order*-limits of sequences in A. Of course A must be included in A. But from
(a) we see that the limit of any order*-convergent sequence in A belongs to A . So A is closed and is
equal to A. Turning this round, we see that A is just the set of order*-limits of sequences in A, as claimed.
(c) Set D = {d : d A, [0, d] 6 G}, H = A \ D. Since D A \ G, H is an open subset of G.
?? If 0 D, then (b) tells us that there is a sequence hdn inN in D order*-converging to 0. Now there
is for each n N a cn dn such that cn
/ G. By 367Bb or 392La, hcn inN order*-converges to 0, and
X Thus 0 H and H is a neighbourhood of 0.
0 A \ G; but G is supposed to be a neighbourhood of 0. X
?? If a H and b [0, a] \ H, then b D, so there is a sequence hdn inN in D order*-converging to b. In
this case, hdn ainN order*-converges to b a = a, by 392Lb. But also [0, dn a] [0, dn ] is not included
X Thus [0, a] H for every a H, and H
in G, so dn a D for each n, and a D; which is impossible. X
has the properties declared.
392N Lemma (Maharam 47) Let A be a ccc Boolean algebra with a T1 topology T such that (i)
: A A A is continuous at (0, 0) (ii) whenever han inN is a non-decreasing sequence in A with inmum
0, then han inN 0 for T. Then A has a strictly positive Maharam submeasure.
proof (a) For any e A \ {0}, there is a Maharam submeasure on A such that e > 0.
P
P(i) Choose a sequence hGn inN of neighbourhoods of 0, as follows. Because T is T1 , G0 = A \ {e} is a
neighbourhood 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 of 0 such that a b Gn for a,
b H, b c H for b, c H and set Gn+1 = H H Gn .) Dene 0 : A [0, 1] by setting
0 a = 1 if a
/ G0 ,
= 2n if a Gn \ Gn+1 ,
\
Gn .
= 0 if a
nN
Pr
596
392N
Measurable algebras
Pr
1 a = inf{ i=0 0 ai : a0 , . . . , ar A, a = supir ai }
for every a A. It is easy to see that 1 (a b) 1 a + 1 b for all a, b A; also a Gn whenever 1 a < 2n ,
so, in particular, 1 e 1, because e
/ G0 .
Set
a = inf{1 b : a e
e}
e, b e
e, 1 a a + and
(a b) 1 (a b ) 1 a + 1 b a + b + 2.
As , a and b are arbitrary, is a submeasure. Next, if hai iiN is any non-increasing sequence in A with
inmum 0, hai eiiN 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
ai 1 (ai e) 0 (ai e) 2n
e = 1 e 1,
(b) Write C for the set of those c A such that there is a strictly positive Maharam submeasure on
the principal ideal Ac . Then C is order-dense in A. P
P Take any e A \ {0}. By (a), there is a Maharam
submeasure such that e > 0. Set A = {e \ a : a = 0}. Because is a submeasure, A is downwardsdirected. ?? If inf A = 0 then, because A is ccc, there is a non-increasing sequence han inN in A with inmum
0; because is a Maharam submeasure,
e inf nN an + (e \ an ) = inf nN an = 0. X
X
Thus A has a non-zero lower bound c, and Ac is a strictly positive Maharam submeasure, while c e. Q
Q
(c) Because A is ccc, there is a sequence hcn inN in C with supremum 1. For each n, let n be a strictly
n
positive Maharam submeasure on Acn ; multiplying by a scalar
P if necessary, we may suppose that n cn 2 .
We can therefore dene : A [0, 2] by setting a = n=0 n (a cn ) for every a A, and it is easy to
check that is a strictly positive Maharam submeasure on A.
k p03) Let A be a Dedekind -complete Boolean algebra. Then
392O Theorem (Balcar Jech & Paza
the following are equiveridical:
(i) there is a strictly positive Maharam submeasure on A;
(ii) A is ccc and weakly (, )-distributive and {0} is a G set for the order-sequential topology of A;
(iii) A is ccc and there is a T1 topology on A such that () : A A A is continuous at (0, 0) ()
whenever han inN is a non-decreasing sequence in A with inmum 0, then han inN 0.
proof (a)(i)(ii) Suppose that A has a strictly positive Maharam submeasure . By 392I, A is ccc and
weakly (, )-distributive. As in 323A (se also 393B below), we have a metric on A dened by setting
(a, b) = (a b) for all a, b A. Now every -open set G A is open for the order-sequential topology.
P
P Let han inN be an order*-convergent sequence in A \ G with limit a. Set cn = supin (ai a) for each n;
then hcn inN is non-increasing and has inmum 0, so
lim supn (an , a) = lim supn (an a) lim supn cn = 0.
392Xf
Submeasures
597
(b)(ii)(iii) Now suppose that the conditions in (ii) are satised. In the following argument all topological terms will refer to the order-sequential topology on A.
T
) There is a non-increasing sequence hGn inN of open neighbourhoods of 0 such that nN Gn = {0}.
(
P
P Let hUn inN be a sequence of open sets with intersection {0}. Set G0 = A, and for n N choose an
open neighbourhood
Gn+1 of 0, included in Un Gn , such that [0, a] Gn+1 for every a Gn+1 (392Mc).
T
?? If 0 6= d nN Gn , then for each n N we can nd a sequence hani iiN in Gn order*-converging to
d (392Mb). By 392Ma, there is a sequence hk(n)inN in N such that han,k(n) inN order*-converges to d.
Now d = supnN inf in ai,k(i) (367Bf), so there is an n N such that c = inf in ai,k(i) is non-zero. But in
this case we must have c ai,k(i) Gi and c Gi Uj whenever i max(n, j + 1), so c = 0. X
X Thus
T
G
=
{0},
as
required.
Q
Q
nN n
(b) Let A be any Boolean algebra and a submeasure on A. (i) Show that the following are equiveridical:
() is order-continuous; () whenever A A is non-empty, downwards-directed and has inmum
0, then
S
inf aA a = 0. (ii) Show that in this case is exhaustive. (Hint: if han inN is disjoint, then nN {b : b ai
for every i n} has inmum 0.)
> (c) Let A be the nite-conite algebra on an uncountable set (316Yk). (i) Set 1 0 = 0, 1 a = 1
for a A \ {0}. Show that 1 is a strictly positive Maharam submeasure but is not exhaustive. (ii) Set
2 a = 0 for nite a, 1 for conite a. Show that 2 is a uniformly exhaustive Maharam submeasure but is
not order-continuous.
> (d) Let A be a Boolean algebra and a submeasure on A. Set I = {a : a = 0}. Show that (i) I is
an ideal of A (ii) there is a submeasure on A/I dened by setting a = a for every a A (iii) if is
exhaustive, so is (iv) if is uniformly exhaustive, so is (v) if is a Maharam submeasure, I is a -ideal
(vi) if is a Maharam submeasure and A is Dedekind -complete, is a Maharam submeasure.
(e) Let A be a Dedekind complete Boolean algebra and an order-continuous submeasure on A. Show
that has a unique support a A such that Aa is strictly positive and A1\a is identically zero.
(f ) Let A be a Dedekind -complete Boolean algebra and a uniformly exhaustive Maharam submeasure
on A. Show that there is a non-negative countably additive functional on A such that {a : a = 0} = {a :
a = 0}. (Hint: 392Xd(vi).)
598
Measurable algebras
392Xg
(g) Let A be an atomless Dedekind -complete Boolean algebra and a strictly positive Maharam
submeasure on A. (i) Show that if a A and 0 a then there is a b a such that b = . (ii) Show
that for every > 0 there is a nite partition of unity A A such that a for every a A.
(h) Let RO(R) be the regular open algebra of R, with its order-sequential topology. (i) Show that if U ,
V are open sets in RO(R) containing and R respectively, then U V 6= . (Hint: enumerate Q as hqn inN ,
and show that there is a sequence hGn inN in U such that Gn+1 Gn {qn } for every n.) (ii) Show that
if U is an open set in RO(R) containing then there are G, H U such that H = R \ G. (iii) Show that
{} is a G set in RO(R). (Hint: if q < q in Q, then {G : ]q, q [ G} is closed in RO(R).)
(i) Let (A,
) be a semi-nite measure algebra. Write Tos for the order-sequential topology on A and Tma
for the measure-algebra topology. Show that Tos Tma , with equality i (A,
) is -nite.
392Y Further exercises (a) Let A be any Boolean algebra with a strictly positive Maharam submeasure.
Show that A is weakly -distributive.
(b) Let G be the regular open algebra of R. (i) Show that there is no non-zero Maharam submeasure on
G. (ii) Show that there is no non-zero countably additive functional on G.
392 Notes and comments Much of the rst part 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 -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 , and every statement of the form Pr(R . . . ) . . . is just
a matter of counting the elements R of with the given property. But I think many of us nd 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 eect in the last fty 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.
In 367 I examined order*-convergence in Riesz spaces, without explicitly discussing the associated topology, and in 392L-392O here I look at Boolean algebras. In both cases the usefulness of the idea arises largely
from the fact that the algebraic operations are separately continuous (367C, 392L), which is itself a consequence of the strong distributive laws in 313A-313B and 352E. It is remarkable that the concept of weak
(, )-distributivity, important for apparently unrelated reasons, should be associated with fundamental
properties of the order-sequential topology (392M).
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 CM4A , CM4B , 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.
393B
599
393A The problem The language I introduced in the last section is already sucient for more than
one formulation of the problem. Consider the statements
(CM1 ) If A is a Dedekind complete Boolean algebra and is a strictly positive Maharam submeasure
on A, then A is measurable.
(CM1 ) Let A be a Dedekind -complete Boolean algebra and : A [0, [ a Maharam submeasure. Then there is a non-negative countably additive functional : A [0, [ such that, for
a A, a = 0 a = 0.
, , \
and
b c + (b \ c) c + (b, c),
bA
b A,
b because the maps (a, b, c) 7 a (b c), (a, b, c) 7 (a b) c are
this is a closed subset of A
b is Hausdor; since A includes the dense set A A A, it is the whole of
continuous and the topology of A
b
b
b
b All the other identities we need to show that
A A A, that is, a (b c) = (a b) c for all a, b, c A.
b
b
A is a Boolean algebra can be conrmed by the same method. Of course A is now a subalgebra of A.
600
393B
Measurable algebras
b [0, [. We
Because : A [0, [ is uniformly continuous, it has a unique continuous extension : A
have
0 = 0,
a (a b) a + b,
a = (a, 0)
b so is a submeasure on A,
b and
for every a, b A and therefore for every a, b A,
a = 0 = (a, 0) = 0 = a = 0,
so is strictly positive.
(d) Now suppose that is exhaustive.
b is a Cauchy sequence for the metric .
(i) The point is that any non-increasing sequence han inN in A
P
P Let > 0. For each n N, choose bn A such that (an , bn ) 2n , and set cn = inf in bi . Then
Pn
(an , cn ) = (inf in ai , inf in bi ) i=0 (ai , bi ) 2
for every n. Choose hn(k)ikN inductively so that, for each k N, n(k + 1) n(k) and
(cn(k) \ cn(k+1) ) supin(k) (cn(k) \ ci ) .
k in(k)
b is Dedekind -complete. P
b hbn inN =
(ii) It follows that A
P If han inN is any sequence in A,
b
hinf in ai inN is a Cauchy sequence with a limit b A. For any k N,
(b \ ak ) = limn (bn \ ak ) = 0,
(iii) We also nd that is a Maharam submeasure, because if han inN is a non-increasing sequence in
b with inmum 0, it must have a limit a which (as in (ii) just above) must be its inmum, that is, a = 0;
A
consequently
limn an = a = 0.
b is ccc (392Ca) and Dedekind complete
(iv) It follows at once that is exhaustive (392Hc), so that A
(316Fa).
b A.
b In particular,
is continuous (by (b)) and zero on the dense set A A, so is zero everywhere on A
b
(a b) = a + b whenever a b = 0 in A, and is additive. Since it is also countably subadditive (392Hb)
b ) is a totally nite measure algebra.
it is countably additive, and (A,
393C Lemma Let A be a Boolean algebra and an exhaustive submeasure on A. Then there are a
Dedekind complete Boolean algebra B, a strictly positive Maharam submeasure on B, and a Boolean
homomorphism : A B such that = . If is additive, then (B, ) is a totally nite measure algebra.
393F
601
proof (a) Set I = {a : a = 0}. Then I A; let B0 be the Boolean quotient algebra A/I (312K). If a1 ,
a2 A and a1 a2 in B, then a1 \ a2 I, so
a1 a2 + (a1 \ a2 ) = a2 .
for every a A.
(c) If is additive, then (a1 a2 )+(a1 a2 ) = a1 +a2 for all a1 , a2 A, so 0 (b1 b2 )+0 (b1 b2 ) =
0 b1 + 0 b2 for all b1 , b2 B0 , and 0 is additive; by 393Bd, (B, ) is a totally nite measure algebra. This
completes the proof.
393D Definition Let A be a Boolean algebra and , two submeasures on A. Then is absolutely
continuous with respect to if for every > 0 there is a > 0 such that a whenever a .
393E Lemma Let A be a Dedekind -complete Boolean algebra and , two Maharam submeasures
on A such that a = 0 whenever a = 0. Then is absolutely continuous with respect to .
proof ?? Otherwise, we can nd a sequence han inN in A such that an P
2n for every n, but =
inf nN an > 0. Set bn = supin ai for each n, b = inf nN bn . Then bn i=n 2i 2n+1 for each n
proof (a) I must check that B has the declared properties. The point is that it is the subalgebra B 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 B 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 B .
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 nitely many of them, and must belong to B . Thus B = B , as claimed. Q
Q
Because B = B is generated by a countable set, it is countable (331Gc). Next, if b B is non-empty,
there are a nite 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 disjoint non-empty members 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 inN . For each
n N let Bn be the nite subalgebra of B generated by {bi : i < n} (so that B0 = {0, 1}). Then hBn inN
602
Measurable algebras
393F
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 nite subalgebras An A and isomorphisms n : An Bn as follows. A0 = {0, 1}, 0 0 = 0,
0 1 = 1. Given An and n , let An be the set of atoms of An . For a An , choose a A such that
if an a, an \ a are both non-zero, then a = an a;
otherwise, a a is any element such that a , a \ a are both non-zero.
(This is where I use the hypothesis that A is atomless.) Set an = supaAn a . Then we see that a an , a \ an
are non-zero for every a An and therefore for every non-zero a An , that is, that
sup{a : a An , a an } = 0,
inf{a : a An , a an } = 1.
sup{b : b Bn , b bn } = 0,
inf{b : b Bn , b bn } = 1,
393G Theorem If one of CM1 , CM1 , CM1 , CM2 , CM2 is true, so are the others.
proof CM1 CM1 Assume that CM1 is true, and that A and are as in the statement of CM1 . Set
I = {a : a = 0}; because is countably subadditive, I is a -ideal, B = A/I is Dedekind -complete
and the quotient map a 7 a is sequentially order-continuous (313Qb, 314C). Now we have a functional
: B [0, [ dened by saying that a = a for every a A, and is a strictly positive submeasure
(as in part (a) of the proof of 393C). In fact is a Maharam submeasure. P
P Suppose that hbn inN is a
non-increasing sequence in B with inmum 0. Choose an A such that an = bn for each n, and set
cn = inf in ai \ inf iN ai ;
then hcn inN is non-increasing and has inmum 0, while cn = bn for each n, so
limn bn = limn cn = 0.
as required.
CM1 CM1 is trivial; allowing for the change in notation, CM1 is just a special case of CM1 .
CM1 CM1 Assume CM1 , and let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure . Then we can express A as /I for some -algebra of subsets of a set X and a
-ideal I of (314M). Set E = E for E ; then X, , satisfy the conditions of CM1 , so there is a
totally nite measure on X, with domain , such that
E = 0 E = 0 E = 0 E = 0.
Consequently we can identify A with the measure algebra of (X, , ), and A is measurable.
CM1 CM2 Now suppose that CM1 is true, and that is an exhaustive submeasure on a Boolean
algebra A. By 393B, we can nd a Dedekind complete Boolean algebra B with a strictly positive Maharam
393I
603
submeasure and a Boolean homomorphism : A B such that a = a for every a A. CM1 assures
us that B is measurable; let
be a totally nite measure on B.
Let > 0. By 393E, is absolutely continuous with respect to
, so there is a > 0 such that b
whenever
a . Take n 1
1. If a0 , . . . , an are disjoint in A, then a0 , . . . , an are disjoint in B,
so there is some i such that
ai , in which case ai = ai . As is arbitrary, is uniformly
exhaustive; as A and are arbitrary, CM2 is true.
CM2 CM2 is trivial.
CM2 CM1 Suppose that CM2 is true, that is, every exhaustive submeasure on the algebra B of openand-closed subsets of {0, 1}N is uniformly exhaustive, and that 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 Aa is still a strictly positive Maharam submeasure.
?? Suppose, if possible, that Aa is not uniformly exhaustive. Then there is a family hani iinN in Aa
such that hani iin is disjoint for each n but inf inN ani > 0. There is a countable atomless subalgebra D
of Aa containing every ani . P
P For each d Aa x on a d d such that d and d \ d are both non-zero.
Dene hDn inN by setting
D0 = {0, a} {ani : i n N},
Dn+1 = Dn {d1 d2 : d1 , d2 Dn } {a \ d : d Dn } {d : d Dn }.
S
Then every Dn is countable, so D = nN Dn is countable. Because D includes D0 , it contains all the ani ;
also d1 d2 , 1 \ d1 and d1 belong to D for every d1 , d2 D, so D is an atomless subalgebra. Q
Q
Because is an exhaustive submeasure on A, D is an exhaustive submeasure on D. Because every ani
belongs to D and inf inN ani > 0, 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
: B D. Set b = b for every b B; then (B, ) is isomorphic to (D, D), and, in particular, 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
1 be a totally nite measure on Aa . Next, because limn en = 0 for any sequence
hen inN of distinct elements of E, and e > 0 for every e E, E must be countable,
and there is a summable
P
family he ieE of strictly positive real numbers. Setting
c =
1 (c a) + eE,e c e , we get a totally
nite measure
on A, and A is measurable.
393H Variations on CM2 The following modications of CM2 are interesting because they display
some general properties of exhaustive submeasures. Consider the statements
(CM3A ) If A is a Boolean algebra and is a non-zero exhaustive submeasure on A, then there is a
non-zero nitely additive functional on A such that 0 a a for every a A.
(CM3A ) If A is a Boolean algebra and is a non-zero exhaustive submeasure on A, then there is
a non-zero non-negative nitely additive functional on A such that is absolutely continuous
with respect to .
(CM3B ) If A is a Boolean algebra and is an exhaustive submeasure on A, then there is a nonnegative nitely additive functional on A such that is absolutely continuous with respect to
.
393I Proposition If one of CM1 , . . . , CM2 , CM3A , CM3A , CM3B is true, so are the others.
proof CM1 CM3A & CM3B Assume CM1 , and let 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 on B, and a Boolean homomorphism : A B such that = . Since 1 = 1 6= 0,
B 6= {0}. CM1 tells us that B is measurable; let
be a strictly positive countably additive functional on B,
and set 1 =
, so that 1 is a nitely additive functional on A. By 393E,
is absolutely continuous with
respect to and is absolutely continuous with respect to
; it follows from the latter that is absolutely
continuous with respect to 1 .
604
Measurable algebras
393I
1, and consider D = {b : b B, b
d}. ?? If D is order-dense in B, there is a disjoint
Set = 21 1/
C D such that sup C = 1. Because B is measurable, it is ccc, and C is countable. If C is innite,
enumerate it as hcn inN ; if it is nite, enumerate it as hci iin and set ci = 0 for i > n. In either case,
hcn inN is a disjoint sequence in C with supremum 1, so that
P
P
1 n=0 cn n=0
cn =
1 < 1
by the choice of , which is absurd. X
X
There is therefore a non-zero b B such that c >
c for every non-zero c b. Set
2 a =
(b a) (b a) a = a
CM3A CM1 Assume CM3A , and let 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 a b = (a b) for b A. It is easy to check that a is a nonzero Maharam submeasure on A; in particular, it is exhaustive. So there is a non-zero non-negative nitely
additive functional which is absolutely continuous with respect to a . If hbn inN is a non-increasing
sequence in A with inmum 0, then limn a bn = 0, so limn bn = 0; accordingly is countably
additive (326Ga), therefore completely additive, because A is ccc (326L). Set I = {b : b = 0}; then I is an
order-closed ideal, so contains its supremum b0 say. Since b = 0 whenever a b = 0, b0 1 \ a, and c a,
where c = 1 \ b0 . But Ac I = {0}, so Ac is strictly positive, and witnesses that Ac is measurable. Also
c = 1 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
1 such that (A,
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 a strictly
positive Maharam submeasure on A. By CM3B , there is a non-negative nitely additive functional on A
such that is absolutely continuous with respect to . But this means that
a = 0 = a = 0 = a = 0,
that is, is strictly positive. Also A is Dedekind complete and weakly (, )-distributive, by 392I. So A is
measurable, by 391D. As A and are arbitrary, CM1 is true.
393J The rst 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
(CM4A ) Let A be a ccc Dedekind complete Boolean algebra with a T1 topology T such that (i)
the Boolean operation : A A A is continuous at (0, 0) (ii) if han inN is a non-increasing
sequence in A with inmum 0, then han inN converges to 0 for T. Then A is measurable.
k p03:
Associated with this is a version from Balcar Jech & Paza
(CM4B ) Let A be a ccc weakly (, )-distributive Dedekind complete Boolean algebra such that
{0} is a G set in the order-sequential topology of A as described in 392K. Then A is measurable.
393K Proposition If one of CM1 , . . . , CM3B , CM4A , CM4B is true, so are the others.
proof The three statements CM1 , CM4A and CM4A all amount to saying that if a ccc Dedekind complete
Boolean algebra satises one of the conditions of 392O, then it is measurable.
393L My own rst 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.
393N
605
(CM5 ) Let A be a ccc Dedekind complete Boolean algebra, and suppose that there is a Hausdor 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 -complete Boolean algebra and : A [0, [ a strictly
positive Maharam submeasure. For u L0 = L0 (A) set
(u) = inf{ : 0, [[|u| > ]] }.
Then denes 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
(u + v) (u) + (v),
0
(u) (u) if || 1,
lim0 (u) = 0
for every u, v L . P
P (i) It will save a moment if we observe that whenever > (u) there is an
such that [[|u| > ]] , so that
[[|u| > ]] [[|u| > ]] .
(364Fa)
[[|u| > (u)]] = limn [[|u| > (u) + 2n ]] limn (u) + 2n = (u).
[[|u + v| > (u) + (v)]] [[|u| + |v| > (u) + (v)]]) ([[|u| > (u)]] [[|v| > (v)]])
[[|u| > (u)]] + [[|v| > (v)]] (u) + (v),
and (u) (u). (iv) limn [[|u| > n]] = 0 because h[[|u| > n]]inN is a non-increasing sequence with
inmum 0. So if > 0, there is an n 1 such that [[|u| > n]] , in which case [[|u| > ]] whenever
Q
|| n1 , so that (u) whenever || n1 . As is arbitrary, lim0 (u) = 0. Q
(b) Accordingly we have a metric (u, v) 7 (uv) which denes a linear space topology T on L0 (2A5B).
Now let G be an open set containing 0. Then there is an > 0 such that H = {u : (u) < } is included in
G. If v H and |u| |v|, then
[[|u| > (v)]] [[|v| > (v)]] (v),
606
Measurable algebras
393N
0 = (0) = (0) = 0.
(To see that 0 = 0 set an = 0 for every n in the denition 393Oa.) () If a b then of course a b. ()
If a, b A and c a b, then
min(1, (c)) = min(1, ((c a) + (c \ a)))
min(1, ((c a)) + ((c \ a))) a + b.
As c is arbitrary, (a b) a + b. Thus is a submeasure. () ?? Suppose, if possible, that there is
a non-decreasing sequence han inN in A, with inmum 0, such that h(an )inN does not converge to 0.
Because h(an )inN is non-increasing, = 31 inf nN an is greater than 0. Now for each n N there are
m > n, b such that b an \ am and (b) . P
P As an 3 there is a c an such that (c) 2. Now
h(am cimN converges to 0 in U , by 393Q, so there is an m > n such that
((c \ am )) = (c (am c)) (c) ((am c)) ,
*393S
607
We may therefore choose hbk ikN , hn(k)ikN such that n(k +1) > n(k), bk an(k) \ an(k+1) and (bk )
P
for every k. But hbk ikN is disjoint, so we ought to be able to form k=0 bk = (supkN bk ) in U , and
limk bk = 0 in U , that is, limk (bk ) = 0. X
X
Thus limn an = 0 for every non-increasing sequence han inN with inmum 0, and is a Maharam
submeasure.
(iii) By CM1 , there is a non-negative countably additive functional : A [0, [ such that a = 0
whenever a = 0. In particular, if a = 0, then (a) = 0 and a = 0. So is a control measure for . As
A, U and are arbitrary, CM6 is true.
CM6 CM1 Assume CM6 , and let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure . Give L0 = L0 (A) the topology of 393M. Then : A L0 is a vector measure in
the sense of 393O. P
P If han inN is a disjoint sequence in A with supremum a, set bn = supin ai , so that
Pn
bn = i=0 ai for each n. We have (a \ bn ) 0, so that
(a bn )) = min(1, (a \ bn )) 0,
P
where is the functional of 393M, and a = i=0 ai in L0 . Q
Q
CM6 now assures us that there is a non-negative countably additive functional on A such that
a = 0 = a = 0 = a = 0,
so that is strictly positive and A is measurable. As A and 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 -complete Boolean algebra, U a normed space and : A U a vector
measure. Then has a control measure.
(3A5Ib), and as will still be a vector
proof (a) Since U can certainly be embedded in a Banach space U
measure when regarded as a map from A to U , we may assume from the beginning that U itself is complete.
(b) is bounded (that is, supaA kak is nite). P
P?? Suppose, if possible, otherwise. Choose han inN
inductively, as follows. a0 = 1. Given that supa an kak = , choose b an such that kbk kan k + 1.
Then k(an \ b)k 1. Also
supa an kak supa an k(a b)k + k(a \ b)k,
so at least one of supa b kak, supa an \b kak must be innite. We may therefore take an+1 to be either b
or an \ b and such that supa an+1 kak = . Observe that in either case we shall have k(an \ an+1 )k 1.
Continue.
At the end of the P
induction we shall have a disjoint sequence han \ an+1 inN such that k(an \ an+1 )k 1
for every n, so that n=0 (an \ an+1 ) cannot be dened 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 =
(363Ea).
Now the key to the proof is the following fact: if hun inN is a disjoint order-bounded sequence in (L )+ ,
hT un inN 0 in U . P
P Let be such that un 1 for every n. Let > 0, and let k be the integer
part of /. For n N, i k set ani = [[un > (i + 1)]]; then hani inN is disjoint for each i, and if we set
Pk
vn = i=0 ani , we get vn un vn + 1, so kun vn k .
P
Because hani inN is disjoint, n=0 ani is dened in U , and hani inN 0, for each i k. Consequently
Pk
T vn = i=0 ani 0
as n . But
kT un T vn k kT kkun vn k kT k
608
*393S
Measurable algebras
(d) Consider the adjoint operator T : U (L ) . Recall that L is an M -space (363B) so that its
dual is an L-space (356N). Write
If u L , then
A = {T g : g U , kgk 1} (L ) = (L ) .
supf A |f (u)| = supkgk1 |(T g)(u)| = supkgk1 |g(T u)| = kT uk.
as n . So A is uniformly integrable. Q
Q
(e) Next, A (L )
P If f A, it is of the form T g for some g U , that is,
c . P
f (a) = (T g)(a) = gT (a) = g(a)
1
n+1
inf{#(J) : J I, E
1
n+1
iJ
Ai }
393U
nX =
and X
609
2
n+1 .
Q
Q
iI
Ai
#(I)
n+1
2n
,
n+1
(b) This example shows at least that any proof of CM3A cannot work through any generally valid
inequality of the form if is a Maharam submeasure there is an additive functional with . 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 dene Xn = [In ]n , Ani = {a : i a Xn }, n : PXn [0, 1]
for every W Z. Then is an outer measure on Z, by arguments we have been familiar with since 114D.
Also Z = 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 Z 1. (ii) If J K and (n,i)J n+1
< 1, set Jn = {i : (n, i) J} for each n; then
S
A
.
This
denes a sequence x Z such that x
/ Cni
#(Jn ) < n + 1, so we can choose
an
x(n)
X
\
ni
n
iJn
S
for every (n, i) J, and Z 6= (n,i)J Cni . As J is arbitrary, Z 1. Q
Q
Finally, if is any subalgebra of PZ containing every Cni , and : [0, [ a non-negative nitely
additive functional such that E E for every E , then = 0. P
P For each n, every point of x belongs
to n dierent Cni , just as in (a) above; so that
P
2n
nZ iIn Cni
.
n+1
(c) A non-zero submeasure on a Boolean algebra A is called pathological if the only additive functional
such that 0 a a for every a A is the zero functional. Thus the submeasure of (b) above is
pathological, and CM3A can be read an exhaustive submeasure cannot be pathological.
(d) It is important to note that Fubinis theorem fails catastrophically for submeasures. As a simple
example, consider the space (X, ) of (a) above, for some xed n 1. If we dene : P(X X) [0, 1] by
setting
P
S
W = inf{ i=0 Ei Fi : W iN Ei Fi }
for each W X X, then we obtain an outer measure, just as if itself were a measure; but (X X)
4n/(n + 1)2 is small compared with (X)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 ),
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 > 0 we can nd a
submeasure : B [0, 1] such that
(i) for every n N there is a disjoint sequence Sn0 , . . . , Snn in B such that Sni = 1 for every i n;
(ii) if hEn inN is any disjoint sequence in B then lim supn En .
Q
proof (a) For each n N let In be the nite set {0, . . . , n}, given its discrete topology; set X = nN In ,
with the product topology; let C be the algebra of subsets of X generated by sets of the form Sij = {x :
610
Measurable algebras
393U
x(i) = j}, where i N and j i. Note that X is compact and Hausdor 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 : 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 members of C determined by coordinates in {0, . . . , n};
note that An is nite. For k l N, say that E C is (k, l)-thin if for every A Ak there is an A Al
such that A A \ E. Note that if k k l l then Ak Ak and Al Al , so any (k, l)-thin set is also
(k , l )-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 i 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 , r)-small for every k 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 A Am such that A A \ E and A 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 A Am such that
A (A S) \ E. (ii) If n > l, there is an A Al such that A A \ E, because E is (k, l)-thin; now
A Am , and A S is non-empty because A 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 , . . . , Er1 are (k, 2r)-small, then
A S is not covered by E0 , . . . , Er1 . 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 A Am such that A S 6= and A A \ Er . Now
S every
Ej , for j < r, is (mj , 2r)-small, therefore (m, 2r)-small, so by the inductive hypothesis A S 6 j<r Ej .
S
Accordingly A S 6 jr Ej and the induction continues. Q
Q
(e) Now suppose that hEn inN is a disjoint sequence inSC. Then for any k N there are l, n N such
that En is (k, l)-thin for every n n . P
P Consider Gn = jn Ej for each n N. Then every Gn is open
T
and nN Gn = . If A Ak , then A, with its subspace topology, is compact, so Baires theorem (3A3G)
tells us that there is an nA such that GnA A is not dense in A; let lA be such that A \ GnA includes a
member of AlA . Set n = max{nA : A Ak }, l = max{lA : A Ak }. If n n , A Ak there is an
A AlA Al such that
A A \ GnA A \ En .
1
r
393Ye
611
(c) Let A be a Dedekind complete Boolean algebra and , 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) (A) = .
(f ) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure. Show
that if (A) then the associated topology on A (393B) is separable and A is -linked.
(g) Consider the statement
(CM6 ) If X is a set, a -algebra of subsets of X, U a metrizable linear topological space, and
: U a vector measure, then has a control measure.
Show that CM6 CM6 .
(h) Consider the outer measures n , 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
. (iii) Show that for
n 1 and set A = {a : a Xn , ia i is even}. Show that n A = n (X \ A) = n+1
any < 1 there is a disjoint sequence hEn inN in E such that En for every n.
(i) Let (A,
) be a -nite measure algebra and : A [0, [ a non-zero submeasure such that a
a
for every a A. Show that there is a non-zero additive functional : A [0, [ such that a a for
every a A. (Hint: if > 0 and
c c, then {a : a <
a} cannot be order-dense in A.)
393Y Further exercises (a) Let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure. Show that if (A) then A is -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 dierence of two submeasures. (i) Show
that N is just the three-dimensional space of functionals : A R such that 0 = 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 Hausdor topology T
on A such that (i) the Boolean operations , 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 inmum 0. (Hint: Given such a topology T, show that the construction
in 393K produces enough order-continuous submeasures on A to dene T. Now if is any order-continuous
submeasure, let a be its support; because Aa is ccc, there is a T-continuous submeasure with support a; use
393E to see that is continuous.)
(d) Consider the statement
(CM4A ) Let A be a ccc Dedekind complete Boolean algebra with a Hausdor topology T such that
if han inN is a sequence in A which order*-converges to a A, then han inN converges to 0 for
the topology T. Then A is measurable.
Show that CM4A is true i CM1 is true.
(e) 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( (A), ) = max(d(A), ) = max(d(L0 ), ).
612
Measurable algebras
393Yf
394B
Kawadas theorem
613
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 (, )-distributive algebras (392I), some of them -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 terms of order*-convergence (392O). 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 dierent purposes. In 393Ta we have submeasures on nite
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.
614
394B
Measurable algebras
(iii)(ii) Suppose that satises (iii), and that 0 6= c0 a. Then we can nd a G and a non-zero
c1 c0 such that agrees with on Ac1 . Suppose that dom = Ad , where necessarily d c1 . Then there
are a partition of unity hdi iiI in Ad and a family hi iiI such that c = i c whenever c di . There is
some i I such that c2 = c1 di 6= 0, and we see that c = c = i c for every c c2 . As c0 is arbitrary,
satises (ii).
(ii)(i) If satises (ii), set
D = {d : d
The hypothesis is that D is order-dense in A, so there is a partition of unity hai iiI of Aa lying within D
(313K); for each i I take i G such that c = i c for c ai ; then hai iiI and hi iiI witness that
G .
(b) This is elementary; if hai iiI , hi iiI witness that G , then hai iiI = hi ai iiI , hi1 iiI witness
that 1 G .
(c) I ought to start by computing the domain of :
d dom() d dom , d dom
d
a, d b d
1 (a b) = c.
(ii) (This is of course a Schroder-Bernstein theorem, and the proof is the usual one.) Take , G
such that a b, b a. Set a0 = a, b0 = b, an+1 = bn and bn+1 = an for each n. Then han inN ,
hbn inN are non-increasing sequences; set a = inf nN an , b = inf nN bn . For each n,
Aa2n \a2n+1 : Aa2n \a2n+1 Ab2n+1 \b2n+2 ,
Ab2n \b2n+1 : Ab2n \b2n+1 Aa2n+1 \a2n+2
are isomorphisms, while
Aa : Aa Ab
394F
615
Kawadas theorem
= 1 c if c
By 394Be,
G ,
nN
(c) This is easy to prove directly from the results in 394B, but also follows at once from (b-i); any
transitive reexive relation gives rise to an equivalence relation.
(d) We may suppose that I is well-ordered by a relation . For i I, set ai = ai \ supj<i aj . Set
a = supiI ai = supiI ai , b = supiI bi . For each i I, we have a bi bi and a i G such that i ai = bi .
Set b = supiI bi b; then we have an isomorphism : Aa Ab dened by setting d = i d if d ai ,
and G , so a and b are G- -equidecomposable and a 4G b.
394D Theorem Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Then the
following are equiveridical:
(i) there is an a 6= 1 such that a is G- -equidecomposable with 1;
(ii) there is a disjoint sequence han inN of non-zero elements of A which are all G- -equidecomposable;
(iii) there are non-zero G- -equidecomposable a, b, c A such that a b = 0 and a b c;
(iv) there are G- -equidecomposable a, b A such that a b.
proof Write G for the full local semigroup generated by G.
(i)(ii) Assume (i). There is a G such that 1 = a. Set an = n (1 \ a) for each n N; because
every n belongs to G (counting 0 as the identity operator on A, and using 394Bc), with dom n = A, an
is G- -equidecomposable with a0 = 1 \ a for every n. Also an = n 1 \ n+1 1 for each n, while hn 1inN is
non-increasing, so han inN is disjoint. Thus (ii) is true.
(ii)(iii) Assume (ii). Set a = supnN a2n , b = supnN a2n+1 , c = supnN an , so that a b = 0 and
a b = c. For each n we have a n G such that n a0 = an . So if we set
d = supnN n 1
2n (d a2n ) for d
a,
(using 394B) and witnesses that a and c are G- -equidecomposable. Similarly, b and c are
belongs to
G- -equidecomposable, so (iii) is true.
(iii)(iv) is trivial.
(iv)(i) Take G such that b = a. Set
d = (d b) (d \ b)
for d A; then
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- -equidecomposable with 1 then a = 1;
(ii) there is no disjoint sequence han inN of non-zero elements of A which are all G- -equidecomposable;
(iii) there are no non-zero G- -equidecomposable a, b, c A such that a b = 0 and a b c;
(iv) if a b A and a, b are G- -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 4H b then a 4G b, so that a and b are G- -equidecomposable whenever they
are H- -equidecomposable.
394F Proposition Let (A,
) be a totally nite measure algebra, and G = Aut A the group of measurepreserving automorphisms of A. Then G is fully non-paradoxical.
616
394F
Measurable algebras
proof If : A Aa belongs to the full local subgroup generated by G, then we have a partition of unity
hai iiI and a family hi iiI in G such that ai = i ai for every i; but this means that
P
P
P
ai =
1.
i ai = iI
i ai = iI
a = iI
As
1 < , we can conclude that a = 1, so that G satises 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, c = c for every G}.
(I looked briey at this construction in 333R, and at various points in Chapter 38 in the special case of a
group generated by a single element.) This is a subalgebra of A, and is order-closed, because every 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, dened by setting
upr(a, C) = inf{c : a c C}
c2 , c1 c2 ,
and c1 = c2 , as claimed. Q
Q
(c) Again supposing that A is Dedekind complete, write G for the full local semigroup generated by
G. Then (a c) = a c whenever G , a dom and c C. P
P We have a = supiI i ai , where
a = supiI ai and i G for every i. Now
(a c) = supiI i (ai c) = supiI i ai c = a c. Q
Q
P For c C,
Consequently upr(a, C) = upr(a, C) whenever G and a dom . P
a c a c = a (a c) = a a c = a a c. Q
Q
upr(b, C) whenever a 4G b.
(d) Still supposing that A is Dedekind complete, we also nd that if a 4G b and c C then a c 4G b c.
Q Hence, or otherwise, a c and b c
P
P There is a G such that a b; now (a c) = a c b c. Q
are G- -equidecomposable whenever a and b are G- -equidecomposable and c C.
(e) Of course the case C = {0, 1} is particularly signicant; when this happens I will call G ergodic.
Thus an automorphism is ergodic in the sense of 372Pa i the group { 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 xed-point subalgebra of G. Take any a, b A. Set c0 = sup{c : c C, a c 4G b}; then
a c0 4G b and b \ c0 4G a.
proof Enumerate G as h i< , where = #(G). Dene ha i< , hb i< inductively, setting
a = (a \ sup< a ) 1 (b \ sup< b ),
b = a .
Then ha i< is a disjoint family in Aa and hb i< is a disjoint family in Ab , and sup< a is G- equidecomposable with sup< b . Set a = a \ sup< a , b = b \ sup< b ,
Then
394I
617
Kawadas theorem
a c0
sup< a 4G b,
so c0 c0 .
Now b c0 . P
P?? Otherwise, because c0 = 1 \ sup< a (394Gb), there must be a < such that
(a \ sup< a ) 1 (b \ sup< b ),
a , which is absurd. X
XQ
Q Consequently
b \ c0
sup< b 4G a.
Now take any c C such that a c 4G b, and consider c = c \ c0 . Then b c = 0, that is, b c =
sup< b c , which is G- -equidecomposable with sup< a c = (a \ a ) c (394Gd). But now
a c = a c c 4G b c 4G (a c ) \ (a c );
because G is fully non-paradoxical, a c must be 0, that is, c
and c0 = c0 . So c0 has the required properties.
c
0
and c = 0. As c is arbitrary, c0
c
0
Remark By analogy with the notation I used in discussing the Hahn decomposition of countably additive
functionals (326O), we might denote c0 as [[a 4G b]], or perhaps [[a 4G b]]C , the region (in C) where a 4G 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 xed-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 disjoint family hdi ii<n
such that c a 4G di
upr(bn , C).
618
394I
Measurable algebras
(c) Consider next c0 = upr(b, C), cn = cn1 upr(bn , C) for n 1. Then hcn inN is a non-increasing
sequence with zero inmum, so again we can dene v L0 by setting v = supnN (n + 1)cn . Once again,
[[v N]] = 1, and [[v n]] = 1 \ cn for each n.
Of course [[v > 0]] = c0 = upr(b, C). Because cn cn cn1 ,
(n + 1)cn (n + 1)cn ncn1 + 1
cn1
...
c0 ,
) Suppose that c Cn+1 ; let hdi iin be a disjoint family such that c a 4G di
(
c = c \ cn . For each i < n, bi 4G a, so
bi c 4G a c 4G di c ,
while also
bn c 4G a c 4G dn c .
Take d
b c .
Because G is fully non-paradoxical, d supi<n (di c ) must be exactly b c , so d must be the whole of
dn c , and
a c 4G dn c = d 4G bn .
But this means that c
Accordingly
cn .
for n 1. For n = 0 we have [[u 0]] = 1 = max C0 . So [[u n]] = max Cn for every n, as required.
(e) Similarly, if we set
Cn = {c : c C, there is a family hdi ii<n
then
1 \ cn
max Cn
sup di }
i<n
for every n.
c0 .
b \ supi<n bi
bn cn1 cn ,
) Now take any c Cn and a corresponding family hdi ii<n such that di 4G a for every i < n and
(
b c supi<n di .
394K
619
Kawadas theorem
ci .
c b,
P
P We may suppose that cjk 6= 0. Of course
cjk
[[b1 : a + b2 : a = j + k]].
Next, there are sets J, J A such that d 4G a for every d J J , #(J) j, #(J ) k, sup J b1 cjk
and sup J b2 cjk . So sup(J J ) (b1 b2 ) cjk and J J witnesses that cjk [[b1 b2 : a j + k]].
Q
Q
Accordingly
cjk
[[b1 : a + b2 : a b1 b2 : a 0]].
(d) This time, set cjk = [[b1 : a = j]] [[b2 : a = k]] for j, k N. Then
cjk
for every j, k N. P
P Once again, we surely have
cjk
[[b1 : a + b2 : a = j + k]].
b1
b2
for j i < j + k.
As b1 b2 = 0, the whole family hdi ii<j+k is disjoint and witnesses that cjk
So
cjk
[[b1 b2 : a b1 : a b2 : a 0]]
[[b1 b2 : a j + k]]. Q
Q
620
394K
Measurable algebras
Let hdi ii<k be a disjoint family in Ab such that a c0 4G di for each i; cutting the di down if necessary, we
may suppose that a c0 is G- -equidecomposable with di for each i. As c0 6 [[b : a k]], b c0 6 supi<k di ;
set d = b c0 \ supi<k di 6= 0. If c C is non-zero and c c0 , then a c 4G di for every i < k, while
c
6 [[b : a k + 1]], so a c 64G d; by 394H, d c0 4G a and d = d c0 4G a c0 . There is therefore a
non-zero a
a c0 such that a
4G d. But now remember that a c is supposed to be a relative atom over
C, so a
= a c for some c C such that c c0 . In this case, a c 4G di for every i < k and also a c 4G d,
so 0 6= c [[b : a k + 1]], which is absurd. X
X
394L Lemma Let A be a Dedekind complete Boolean algebra, not {0}, and G a fully non-paradoxical
subgroup of Aut A with xed-point subalgebra C. Suppose that a1 , a2 , b A and that upr(a1 , C) =
upr(a2 , C) = 1. Then
b : a2 b : a1 a1 : a2 ,
b : a2 b : a1 a1 : a2 .
proof I use the same method as in 394K. As usual, write G for the full local semigroup generated by G.
(a) For j, k N set
Then
cj,k
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
[[b : a1 a1 : a2 = jk]]; what we need to check is that c [[b : a2 jk]]. Again, we may suppose that
c 6= 0. There are families hdi ii<j , hdl il<k such that
4G
dl c
4G
i (dl c)
i (a1 c) di
b.
Also hi (dl c)ii<j,l<k is disjoint because hi (a1 c)ii<j and hdl il<k are, so witnesses that c is included in
[[b : a2 jk]]. Q
Q
Now, just as in 394K, it follows from the fact that supj,kN cj,k = 1 that b : a1 a1 : a2 b : a2 .
(b) For j, k N set
Then
cj,k
P
P Write c for cj,k . Then c [[b : a1 a1 : a2 = jk]]. There are families hdi ii<j , hdl il<k such that
di 4G a1 for every i < j, dl 4G a2 for every l < k, b c supi<j di and a1 c supl<k dl . For each i < j,
let di a1 be G- -equidecomposable with di , and take i G such that i di = di . Then
i (di dl ) 4G dl 4G a2 for every i < j, l < k,
i<j,l<k
i<j
l<k
= sup di c b c.
i<j
i<j
394M
621
Kawadas theorem
394M Lemma Let A be a Dedekind complete Boolean algebra, not {0}, and G a subgroup of Aut A
with xed-point subalgebra C.
(a) For any a A, there is a b a such that b 4G a \ b and a = a \ upr(b, C) is a relative atom over C.
(b) Now suppose that G is fully non-paradoxical. Then for any > 0 there is an a A such that
upr(a, C) = 1 and b : a b : a + 1 : a for every b A.
proof (a) Set B = {d : d a, d 4G a \ d} and let D B be a maximal subset such that upr(d, C) upr(d , C)
= 0 for all distinct d, d D. Set b = sup D. For any d D, d 4G a \ d, so
b upr(d, C) = d upr(d, C) 4G (a \ d) upr(d, C) = (a \ b) upr(d, C)
a\b
by 394Gc. By 394H,
b = b supdD upr(d, C) 4G a \ b.
?? Suppose, if possible, that a = a \ upr(b, C) is not a relative atom over C. Let d0 a be an element
not expressible as a c for any c C; then d0 6= a upr(d0 , C) and there must be a G such that
d1 = d0 a \ d0 is non-zero. In this case
d1 4G 1 d1
so d1 B; but also
d1 upr(d, C)
d0
a \ d1 ,
d1 upr(b, C) = 0,
an ,
cn+1 an+1 4G an \ d,
If
b
then, because both terms on the right are relative atoms over C, there are c , c C such that
b = (b an \ cn ) (b an cn \ cn+1 )
b : an b : an + 1,
622
394M
Measurable algebras
we must have c cn = 0. But this means that an c is a relative atom over C. By 394Ke, c is included in
[[b : an b : an = 0]]; as also 1 : an 1 (394Ka), c must be zero, that is, b : an b : an +1 : an .
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 xed-point subalgebra C. Then there is a unique function : A L (C) such that
(i) is additive, non-negative and order-continuous;
(ii) [[a > 0]] = upr(a, C) for every a A; in particular, a = 0 i a = 0;
(iii) 1 = 1;
(iv) (a c) = a c for every a A, c C; in particular, c = c for every c C;
(v) If a, b A are G- -equidecomposable, then a = b; in particular, 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
a (b) =
b : a
L0 = L0 (C);
1 : a
the rst thing to note is that because 1 : a 1, we can always do the divisions to obtain elements a (b)
of L0 (A) (364P). Set
b = inf aA a b
for b A. (Note that L0 (C) is Dedekind complete (364O), so the inmum is dened.)
(b) The formulae of 394K tell us that, for a A and b1 , b2 A,
a 0 = 0,
a b1 a b2 if b1
b2 ,
a (b1 b2 ) a b1 + a b2 ,
a 1 1.
b1 b2 if b1
b2 ,
1 1.
(c) For each n N there is an en A such that b : en b : en + 2n 1 : en for every b A (394M).
Pa : en a : en + 2n 1 : en , so
Now en b a b + 2n b : a for every a A , b A. P
a : en 1 : a a : en 1 : a + 2n 1 : en 1 : a
1 : en + 2n 1 : en 1 : a
b : en 1 : a b : a a : en 1 : a
b : a 1 : en + 2n b : a 1 : en 1 : a
Q
and, dividing by 1 : a 1 : en , we get en b a b + 2n b : a. Q
(d) Now is additive. P
P Taking hen inN from (c), observe rst that
inf nN en b a b + inf nN 2n b : a = a b
for every a A , b A, so that b = inf nN en b for every b. Now suppose that b1 , b2 A and b1 b2 = 0.
Then, for any n N,
394N
Kawadas theorem
623
b1 : en + b2 : en b1 : en + b2 : en + 2n+1 1 : en
b1 b2 : en + 2n+1 1 : en
(by 394Kd)
b1 b2 : en + 2n+1 1 : en .
Dividing by 1 : en , we have
b1 + b2 en b1 + en b2 en (b1 b2 ) + 2n+1 1.
b1 + b2 (b1 b2 ).
a (b1 ) + 2n b1 : a + a (b2 ) + 2n b2 : a .
As n is arbitrary, (b1 b2 ) a (b1 ) + a (b2 ); as a and a are arbitrary, (b1 b2 ) b1 + b2 (using 351Dc).
As b1 and b2 are arbitrary, is additive. Q
Q
We see also that 1 : en (1 + 2n )1 : en , so that en 1 (1 + 2n )1 for each n; since we already
know that 1 1, we have 1 = 1 exactly.
(e) If c C then
[[c > 0]] [[1 c > 0]] [[c : 1 > 0]] = upr(c, C) = c
for any b A, c C. Similarly, (b \ c) b (1 \ c); adding, we must have equality in both, and
(b c) = b c.
Rather late, I point out that
0 a 1 = 1 L = L (C)
[[b > 0]] [[1 b > 0]] [[b : 1 > 0]] = upr(b, C)
by 394I(ii) again. ?? Suppose, if possible, that [[b > 0]] 6= upr(b, C). Set c0 = upr(b, C) \ [[b > 0]], a0 =
b (1 \ upr(b, C)) A . Let k 1 be such that c1 = c0 [[1 : a0 k]] 6= 0. Then a0 c1 = b c1 , so
a0 c1 = (a0 c1 ) = (b c1 ) = b c1 = 0.
By 364Nb, there is an a A such that c1 6 [[a a0 c1 k1 ]], that is, c2 = c1 [[a a0 < k1 ]] 6= 0. Now
c2
[[1 : a ka0 : a > 0]] [[1 : a0 a0 : a ka0 : a > 0]] [[1 : a0 > k]],
which is impossible, as c2 c1 . X
X
Thus [[b > 0]] = upr(b, C). In particular, b = 0 i b = 0.
624
394N
Measurable algebras
a (supJ a ) =
P
for every nite J , so (because L (C) is Dedekiond complete) u = < a is dened, and u a.
For < , set b = sup< a . By the inductive hypothesis,
P
P
b = < a = supJ is finite J a u.
P
At the same time, if J is nite, there is some < such that J , so that J a b ; accordingly
sup< b = u.
?? Suppose, if possible, that u < a; set v = a u. Take > 0 such that c0 = [[v > ]] 6= 0. Let <
be such that c1 = c0 \ [[u b > ]] is non-zero (cf. 364Nb). Now u b (a \ b ), so
c1
a c1 4G
By (g),
a \ (d e ); take d a \ (d e ) G- -equidecomposable with a c1 , and continue.
At the end of this induction, we have a disjoint family hd i< in Aa\d such that d is G- -equidecomposable with a c1 for every . But this means that a = sup< d is G- -equidecomposable with a c1 ,
while a (a \ d) P
c1 ; since d c1 6= 0, G cannot be fully non-paradoxical. X
X
Thus a = u = < a and the induction continues. Q
Q
1 e = (1 \ e) = supbB (1 \ b) = supbB 1 b,
(j) I still have to show that is unique. Let : A L be any non-negative order-continuous G-invariant
additive function such that c = c for every c C.
(i) Just as in (e) of this proof, but more easily, we see that (b c) = b c for every b A, c C.
(ii) If hai iiI is a disjoint family in A with supremum a, then hsupiJ ai iJI is finite is an upwardsdirected family with supremum a, so that
P
P
a = supJI is finite (supiJ ai ) = supJI is finite iJ ai = iI ai .
(iv) Take a A , b A and for j, k N set cjk = [[1 : a = j]] [[b : a = k]]. Then
b : a cjk b 1 : a cjk .
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 4G d for every d I, e 4G a for every e J, I is disjoint, and b cjk sup J. So
394R
625
Kawadas theorem
jk (a cjk ) k
X
dI
X
eJ
(e cjk )
(d cjk ) k cjk
= kcjk = b : a cjk . Q
Q
Summing over j and k, b : a b 1 : a, that is, a b b. Taking the inmum over a, b b. But
also
b = 1 (1 \ b) 1 (1 \ b) = b,
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 : A [0, 1] such that 1 = 1.
proof Let C be the xed-point subalgebra of G, and : A L (C) the function of 394N. By 311D, there
is a ring homomorphism 0 : C {0, 1} such that 0 1 = 1; now 0 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
= f0 . Then is order-preserving and additive because f0 and are, 1 = f0 (1) = 0 1 > 0, and is
G-invariant because is.
394P Theorem Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup
of Aut A with xed-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 : A L (C) be the function of 394N, and : 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
= f . Then
is additive and order-preserving
and sequentially order-continuous because f and are. It is also strictly positive, because if a A \ {0}
then a > 0 (394N(ii)), that is, there is some > 0 such that [[a > ]] 6= 0, so that
a
[[a > ]] > 0.
Finally,
is G-invariant because is.
394Q Corollary: Kawadas 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 sucient 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 sucient, by the following result.
Proposition If (A,
) is a homogeneous totally nite measure algebra, the group G = Aut 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 = min(
a,
(1 \ a)) > 0;
then there are b a, d 1 \ a such that
b =
d = . By 383Fb, there is a G such that b = d, so that
a 6= a. As a is arbitrary, the xed-point subalgebra of G is {0, 1}.
626
Measurable algebras
394X
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 identied with R, the functions . . . and . . . become real-valued,
the functionals a (394N) become submeasures and becomes a measure.
(b) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with xed-point subalgebra
C. Suppose that hci iiI is a partition of unity in C and that a, b A are such that a ci 4G b for every
i I. Show that a 4G b.
(c) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with xed-point subalgebra
C. Show that A is relatively atomless over C i the full local semigroup generated by G has many involutions
(denition: 382O).
(d) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with
xed-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 xed-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 i there is
a fully non-paradoxical subgroup G of Aut A such that the xed-point subalgebra of G is purely atomic.
(g) Let (A,
) be a localizable measure algebra. Show that the following are equiveridical: (i) Aut A is
ergodic; (ii) A is quasi-homogeneous in the sense of 374G; (iii) whenever a, b A and a = b then the
principal ideals Aa , Ab are isomorphic.
(h) Let (A,
) be a localizable measure algebra. Show that Aut A is fully non-paradoxical i (i) for
every innite cardinal , the Maharam-type- component of A (denition: 332G) has nite measure (ii) for
every ]0, [ there are only nitely many atoms of measure .
394Y Further exercises (a) Let A be a Dedekind complete Boolean algebra, G a subgroup of Aut A,
and G the full local semigroup generated by G. For , G , say that if extends . (i) Show that
every member of G can be extended to a maximal member of G . (ii) Show that G is fully non-paradoxical
i every maximal member of G 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 xed-point subalgebra C. Show that A is ccc i C is ccc. (Hint: if C is ccc, L (C) has the countable
sup property (363Yb).)
(c) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A
with xed-point subalgebra C. Show that A is weakly (, )-distributive i C is.
(d) Let A be a Dedekind complete Boolean algebra, G an ergodic subgroup of A, and G the full local
semigroup generated by G. Suppose that there is a non-zero a A for which there is no G such that
a a. Show that there is a measure
such that (A,
) 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 iin is a nite partition of unity in A and we are given i , i G for every
i n, then the elements 0 a0 , 0 a0 , 1 a1 , 1 a1 , . . . , n an are not all disjoint. Must there be a non-zero
non-negative G-invariant nitely additive functional on A?
(See Tarskis theorem in the notes below.)
395A
627
394 Notes and comments Regarded as a sucient condition for measurability, Kawadas theorem suers
from the obvious defect that it is going to be rather rarely that we can verify the existence of an ergodic
fully non-paradoxical group of automorphisms without having some quite dierent 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.
I take the trouble to include the in every G- -equidecomposable, G and 4G because there are
important variations on the concept, in which the partitions hai iiI of 394A are required to be nite or
countable. Indeed Tarskis theorem relies on one of these. I spell it out because it is close to Kawadas in
spirit, though there are signicant dierences 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 : PX [0, 1] such that A = 1; (ii) there are no A0 , . . . , An ,
0 , . . . , n , 0 , . . . , n such that A0 , . . . , An are subsets of X covering X, 0 , . . . , n all belong
to G, and 0 [A0 ], 0 [A0 ], 1 [A1 ], 1 [A1 ], . . . , n [An ] are all disjoint.
For a proof, see 449J 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
(, )-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 -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 xed-point subalgebra C is trivial. Under this hypothesis the proof is of course very much
simpler. (You may nd 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 o 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) [[ka0 : a < 1 : a]] in place of [[1 : a ka0 : a > 0]] (see part (f) of the proof of
394N). As in 388, the dierences 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.
U v = T v w1
for every v L2 , and U = U U for all , Aut A.
is
sequentially
n
nN
P
order-continuous) a = supnN an , so a = n=0 an ; if a 6= 0 then a 6= 0 so a > 0. Thus (A, ) is a
628
395A
Measurable algebras
v w =
T T v
T v},
w v =
(364Re)
T v =
w T v =
w T1 w v.
T (T1 w v)
(T v
w1 )2 =
T v 2 w1 =
T1 T v 2 =
v2 .
(using 364Rd)
(by (b) above)
U U v = T (T v w1 ) w1
p
= T T v T w1 w1
= T v
w1 1
= U v.
So U = U U .
(iii) Writing for the identity operator on A, we see that T is the identity operator on L0 , w = 1
and U is the identity operator on L2 . Since U1 U = U U1 = U , U : L2 L2 is an isomorphism,
with inverse U1 , for every 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 such that (A, ) is a totally nite measure algebra;
(ii) whenever a A \ {0} and hn inN is a sequence in G, hn ainN is not disjoint;
(iii) G is fully non-paradoxical (denition: 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
be such that (A,
) is a totally nite measure algebra. If a A \ {0}, then inf G
(a) > 0.
P
P?? Otherwise, let hn inN be a sequence in G such that
n a 2n for each n N. Set bn = supkn k a
for each n; then inf nN bn = 0, so that
395C
inf nN bn = 0,
629
limn
(n a) = 0
a
nj a) 2j2
(n1
i
P Pj1
j=1
i=0
a,
nj a) <
(n1
i
XQ
Q
c 6= 0, while ni c nj c = 0 whenever i < j, contrary to the hypothesis (ii). X
(c) For each G, dene
L0 = L0 (A) and U : L2 L2 as in 395A, where L2 = L2 (A,
).
R w
w
>
0.
P
P
?
?
Otherwise,
there
is
a
sequence
h
i
in
G
such
that
If
a
A
\
{0},
then
inf
n
nN
G
a
R
v 4n2
a for every n, where vn = wn . In this case,
(a [[vn 2n ]]) 2n2
a for every n, so
a n
n
that b = a \ supnN [[vn 2 ]] is non-zero. But now
(n b) =
as n , contradicting (b) above. X
XQ
Q
w =
b n
b 0
vn2 4n
(d) Write e = 1 for the standard weak order unit of L0 or L2 . Let C L2 be the convex hull of
{U e : G}. Then C and its norm closure C are G-invariant in the sense that U v C, U v C
whenever v C, v C and G. By 3A5Ld, there is a unique u0 C such that ku0 k2 kuk2 for every
u C. Now if G, U u0 C, while kU u0 k2 = ku0 k2 ; so U u0 = u0 . Also, if a A \ {0},
(because u 7
u0 inf
uC
u is k k2 -continuous)
= inf
= inf
u = inf
U e = inf
Za
uC
T e w1
a
Z
w > 0
= inf
w1
Z
u20 (a) = T (T1 u20 a)
Za
Z
2
= w T1 u0 a = (T1 u0 w )2
a
Z
Z
2
2
= (U1 u0 ) = u0 = a. Q
Q
(a) =
u20 =
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 hn inN in G such that hn ainN is disjoint. Thus condition (ii)
of 395B may be read as there is no weakly wandering element of A.
630
Measurable algebras
395X
(c) Let A be a measurable algebra and G a subgroup of Aut A. Suppose that there is a strictly positive
G-invariant nitely additive functional on A. Show that there is a G-invariant
such that (A,
) is a totally
nite measure algebra.
395Y Further exercises (a) Let A be a weakly (, )-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 iiI in Aa and hbi iiI in Ab , and a family hi iiI in G, such that i 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 inN is a disjoint 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 U 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
631
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).
3A1D Cardinal exponentiation For a cardinal , I write 2 for #(P). So 2 = c, and + 2 for
every . (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 innite 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 h iOn . We have 0 = , +1 = + for every (compare 2A1Fc),
S
= < for non-zero limit ordinals . (Enderton 77, pp. 213-214; Jech 78, p. 25; Krivine 71, p.
31.)
632
3A1F
Appendix
3A1F Cofinal sets (a) If P is any partially ordered set (denition: 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 conal subset
of P . Note that cf(P ) = 0 i P = , and that cf(P ) = 1 i P has a greatest element.
(c) Observe that if P is upwards-directed and cf(P ) is nite, then cf(P ) is either 0 or 1; for if Q is a
nite, non-empty conal set then it has an upper bound, which must be the greatest element of P .
(d) If P is a totally ordered set of conality , then there is a strictly increasing family hp i< in P such
that {p : < } is conal with P . P
P If = 0 then P = and this is trivial. Otherwise, let Q be a conal
subset of P of cardinal , and {q : < } an enumeration of Q. Dene hp i< inductively, as follows.
Start with p0 = q0 . Given hp i< , where < , then if p < q for every < , take p = q ; otherwise,
because #() < , {p : < } cannot be conal with P , so there is a p P such that p 6 p for
every < , that is, p < p for every < . Note that there is some < such that q p , so that
q p . Continue.
Now hp i< is a strictly increasing family in P such that q p for every ; it follows at once that
{p : < } is conal with P . Q
Q
(e) In particular, for a totally ordered set P , cf(P ) = i there is a conal strictly increasing sequence
in P .
3A1G Zorns Lemma In Volume 2 I used Zorns 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 Zorns 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 sucient 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 members of I P, and a b. Q
Q So Q is an upper bound of Q in P;
as Q is arbitrary, P satises the conditions of Zorns Lemma, and must have a maximal element.
Another important type of partially ordered set in this context is a family of functions, ordered by
saying that f g if g is an extension of f . In this case, for any non-empty upwards-directed , we
shall have a function h dened by saying that
S
dom h = f dom f , h(x) = f (x) whenever f , x dom f ,
and the usual attack is to seek to prove that any such h belongs to .
I nd that at least once I wish to use Zorns 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 denition of partial order is symmetric, so that (P, ) is
a partially ordered set whenever (P, ) is; and we can seek to apply Zorns Lemma to either.
3A1H Natural numbers and finite ordinals I remarked in 2A1De that the rst few ordinals
,
{},
{, {}},
{, {}, {, {}}},
...
may be identied 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 identied with the rst innite
ordinal . This convention makes it possible to present a number of arguments in a particularly elegant
3A2A
Rings
633
form. A typical
S example is in 344H. There I wish to describe an inductive construction for a family hKz izS
where S = nN {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 members, 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
1 , . . . , n , 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 rst n coordinates with the restriction wn 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 p ) = f (p) f (p ) and f (p p ) = f (p) f (p ) for all p, p P . Such
a homomorphism is surely order-preserving (313H), for if p p in P then f (p ) = f (p p ) = f (p) f (p )
and f (p) f (p ).
(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
for all p, q, r P .
(p q) r = (p r) (q r),
(p q) r = (p r) (q r)
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 is a cardinal, write
[X] = {A : A X, #(A) = },
[X] = {A : A X, #(A) },
Thus
[X]2 is the set of doubleton subsets of X, [X]< is the set of nite subsets of X, [X] 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 Ornsteins theorem (387) and the Kalton-Roberts theorem (392).
Halls Marriage Lemma Suppose that X and Y are nite 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.
634
Appendix
3A2B
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
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}.
3A2G
Rings
635
A + B = (a + I) + (b + I)
= {(a + x) + (b + y) : x, y I}
= {(a + b) + (x + y) : x, y I}
{(a + b) + z : z I} = (a + b) + 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
= {(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) ab 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 : R S a homomorphism such that I is included in the kernel of . Then we have a
ring homomorphism : R/I S such that (a ) = (a) for every a R. is injective i I is precisely the
kernel of .
proof If a, b R and a = b in R/I, then ab I (3A2Fe), so (a)(b) = (ab) = 0, and (a) = (b).
This means that the formula oered does indeed dene a function from R/I to S. Now if a, b R and
is either multiplication or addition,
(a b ) = ((a b) ) = (a b) = (a) (b) = (a ) (b ).
So is a ring homomorphism.
The kernel of is {a : (a) = 0}, which is {0} i (a) = 0 a = 0 a I.
636
3A2H
Appendix
3A2H Product rings (a) Let hRi iiI be any family of rings. Set R =
a + b, ab R by setting
(a + b)(i) = a(i) + b(i),
iI
(ab)(i) = a(i)b(i)
for every i I. It is easy to check from the denition in 3A2A that R is a ring; its zero is given by the
formula
and its additive inverses by the formula
3A3D
General topology
637
638
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 nite 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 Hausdor 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 Hausdor and : X Y is continuous and injective, then is a homeomorphism
between X and [X] (where [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 ultralter 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
ultralter on X containing A. Set
H = {B : B X, there is an open set G F such that A G B}.
3A3L
639
General topology
Q
3A3I Product spaces (a) Definition Let hXi iiI be a family of topological spaces, and X = iI 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 nite J I and a family hGj ijJ such that every Gj is an open set in the corresponding Xj and
{y : y X, y(j) Gj for every j J}
(f ) Let hX
Qi iiI be any family of topological spaces. If Fi is a closed subset of Xi for each i, then
is closed in iI Xi . (Engelking 89, 2.3.4; Bourbaki 66, I.4.3.)
iI
iI
Di
Fi
(g) Let h(Xi , Ti )iiI be a family of topological spaces with product (X, T). Suppose that each Ti is dened
by a family Pi of pseudometrics on Xi (2A3F). Then T is dened by the family P = {
i : i I, Pi } of
pseudometrics on X, where I write i (x, y) = (i (x), i (y)) whenever i I, Pi and x, y X, taking
i 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 i is a pseudometric on X. Write TP for the topology generated by P. (ii) If x G TP , let P P
and > 0 be such that P is nite and {y : (y, x) for every P } is included in G. Express P as
{
j : j J, Pj } where J I is nite and Pj Pj is nite for each j J. Set
Gj = {t : t Xj , (t, j (x)) < for every Pj }
for every j J. Then G = {y : j (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 i 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 , R I For any set I, we can think of {0, 1}I as the product iI 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 Hausdor spaces, it is Hausdor; by Tychonos theorem, it is
compact. A subset G of {0, 1}I is open i for every x G there is a nite J I such that {y : y
{0, 1}I , yJ = xJ} G.
Similarly, the usual topology of R I 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 lter 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.
640
Appendix
3A3Lb
(b) For any topological space X, lter F on X and x X, x is a cluster point of F i there is a lter
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 hn inN is a sequence in R, R and limnH n = for every non-principal ultralter H
on N (denition: 2A3Sb), then limn n = . P
P?? If hn inN 6 , there is some > 0 such that
I = {n : |n | } is innite. Now F0 = {F : F N, I \ F is nite} is a lter on N, so there is an
ultralter F F0 . But now cannot be limnF n . 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 i (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, ) a metric space and hfn inN a sequence of
functions from X to Y . We say that hfn inN converges uniformly to a function f : X Y if for every
> 0 there is an n0 N such that (fn (x), f (x)) whenever n n0 and x X.
(b) Let X be a topological space and (Y, ) a metric space. Suppose that hfn inN 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 Hausdor 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 Hausdor
3A4Bb
Uniformities
641
(ii) Now suppose that hxn inN x. ?? If hxn inN does not converge topologically to x, then there
is an open set G containing x such that {n : xn
/ G} is innite, that is, there is a subsequence hxn inN
of hxn inN such that xn
/ G for every n. Now hxn inN x; but X \ G is supposed to be closed under
*-convergence, so this is impossible. X
X
(iii) The condition in (iii) says just that if F = {x} is a singleton, then x is the only possible -limit
of any sequence in F , so F is closed.
(b) Let H Y be open; set F = Y \H and let hxn inN be a sequence in f 1 [F ] such that hxn inN X x.
Then hf (xn )inN Y f (x), so f (x) F and x f 1 [F ], As hxn inN and x are arbitrary, f 1 [F ] is closed
and f 1 [H] is open; as H is arbitrary, f is continuous.
3A3Q 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. A subset E of X is a G set if it is expressible as the intersection of a sequence of open sets.
(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 lter 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 lter.
The pair (X, W) is now a uniform space.
(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 Hausdor.
(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
642
Appendix
3A4Bc
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.)
(b) If U, V are dened from families P, of pseudometrics, then W will be dened by the family
{
: P} { : }, writing
3A5 intro.
Normed spaces
643
(d) A complete subspace of a Hausdor 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 i every Cauchy sequence converges (cf. 2A4Db). (Schubert 68, p. 141;
Gaal 64, p. 276; James 87, p. 150.)
(f ) If (X, ) is a complete metric space, D X a dense subset, (Y, ) a metric space and f : X Y is
an isometry (that is, (f (x), f (x )) = (x, x ) for all x, x 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
lter F on U is Cauchy i for every open set G containing 0 there is an F F such that F F G (cf.
2A5F). (Immediate from the denitions.)
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 : D Y is uniformly continuous
(for the subspace uniformity of D), then there is a uniformly continuous : X Y extending . If Y
is Hausdor, 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, ) is a metric space, (Y, ) is a complete metric space, D X is a dense subset, and
: D Y is an isometry, then there is a unique isometry : X Y extending .
3A4H Completions (a) Theorem If (X, W) is any Hausdor uniform space, then we can nd a
W)
in which X is embedded as a dense subspace; moreover, any two
complete Hausdor 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.)
W)
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 dened by a metric on a set X, then there is a unique extension of to a
dening the uniformity W.
(Bourbaki 66, IX.1.3.)
metric on X
3A4I A note on metric spaces I mention some elementary facts which are used in 387. Let (X, )
be a metric space. If x X and A X is non-empty, set
(x, A) = inf yA (x, y).
In particular, (x, A) = (x, A). If hAn inN is a non-decreasing sequence of non-empty sets with union A,
then
(x, A) = limn (x, An ).
644
Appendix
3A5A
3A5A The Hahn-Banach theorem: analytic forms This is one of the central ideas of functional
analysis, both nite- and innite-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
classication 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(u) = p(u) whenever u, v U and 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
u + v C whenever u, v C and , 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. (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.)
(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 uB 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 nd a g U
such that g(u) < inf vB g(v) and now set f = 1 g where g(u) < < inf vB 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 nite-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 Zorns 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 generated by W {u} still has trivial intersection with V ), so we have an extension of f0 to
a linear functional f : U R dened by setting f (v + w) = f0 (v) whenever v V and w W . Now
f (u) < 0 minin f (vi ), as required.
3A5Ic
Normed spaces
645
3A5E Weak topologies (a) Let U be any linear space over R and W a linear subspace of the space U
of all linear functionals from U to R. Then I write Ts (U, W ) for the linear space topology dened 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 and Z V 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 adjoint
(or conjugate, or dual) operator T : V U dened by saying that T g = gT for every g V . T
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
Hausdor 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 reexive i bounded sets are relatively weakly compact. (Dunford & Schwartz
57, V.4.8; Taylor 64, 4.61-C.)
(c) If U is a reexive space, hun inN is a bounded sequence in U and F is an ultralter on N, then
limnF un is dened in U for the weak topology. (Use (b) and the argument of 2A3Se.)
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 associated with the norm (2A4Bb), and that the topology dened by
is a linear space topology (2A5D, 2A5B). This topology denes a uniformity W (3A4Ad) which is also the
uniformity dened by (3A4Bd). The norm itself is a uniformly continuous function from U to R (because
|kuk kvk| ku vk for all u, v U ).
, W)
be the uniform space completion of (U, W) (3A4H). Then addition and scalar multi(b) Let (U
a Banach space. (Schaefer 66, I.1.5; Lang 93, p.
plication and the norm extend uniquely to make U
78.)
(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).
646
3A5J
Appendix
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 dened by the norm of V .
V.
3A5L Hilbert spaces I mentioned the phrases inner product space, Hilbert space briey 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 denitions.
(a) An inner product space is a linear space U over
such that
(u1 + u2 |v) = (u1 |v) + (u2 |v),
R
C
(u|v) = (u|v),
R
C
(u|v) = (v|u)
(u|u) 0,
u = 0 whenever (u|u) = 0
R
C.
(b) If U is any inner product space, we have a norm on U dened 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 dened 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 wC ku wk. (Taylor 64, 4.81-A.)
3A5M Strong operator topology If U and V are normed spaces, the strong operator topology on
B(U ; V ) is that dened by the seminorms T 7 kT uk as u runs over U . If A B(U ; V ), then A is relatively
compact for the strong operator topology i {T u : T A} is relatively compact in V for every u U . (Put
3A5Ha and 2A3R together.)
3A3P
Concordance
647
Concordance
I list here the section and paragraph numbers which have (to my knowledge) appeared in print in references
to this volume, and which have since been changed.
316J Weakly (, )-distributive algebras The reference to 316J in the 2003 edition of Volume 4
should be changed to 316H.
354Yk Complexifications of normed Riesz spaces This exercise, referred to in the 2003 edition of
Volume 4, is now 354Yl.
3A3P Definitions This paragraph, referred to in the 2003 edition of Volume 4, is now 3A3Q.
648
References
References
649
Erdos 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 preservent la mesure de Lebesgue est un groupe
simple, Israel J. Math. 29 (1978) 302-308. [382 notes, 383I.]
Fremlin D.H. [74a] Topological Riesz Spaces and Measure Theory. Cambridge U.P., 1974. [Chap. 35 intro,
354Yb, 355 notes, 356 notes, 363 notes, 365 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 Martins Axiom. Cambridge U.P., 1984. [316 notes.]
Fremlin D.H. [89] Measure algebras, pp. 876-980 in Monk 89. [341Lg, 341Zb.]
Fremlin D.H. [n02] Problem EZ, note of 10.11.02; http://www.essex.ac.uk/maths/staff/fremlin/
probEZ.ps. [323 notes.]
Frolk Z. [68] Fixed points of maps of extremally disconnected spaces and complete Boolean algebras,
Bull. Acad. Polon. Sci. 16 (1968) 269-275. [382E.]
Gaal S.A. [64] Point Set Topology. Academic, 1964. [3A3, 3A4.]
Gaifman H. [64] Concerning measures on Boolean algebras, Pacic 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. & Weizsacker 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. [388 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. Birkhauser, 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.
Juhasz 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.]
650
References
Kranz P. & Labuda I. [93] (eds.) Proceedings of the Orlicz Memorial Conference, 1991 , unpublished
manuscript, available from rst editor (mmkranz@olemiss.edu).
References
651
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.]
Sina Ya.G. [59] The notion of entropy of a dynamical system, Dokl. Akad. Nauk SSSR 125 (1959)
768-771. [385P.]
Sina Ya.G. [62] Weak isomorphism of transformations with an invariant measure, Soviet Math. 3 (1962)
1725-1729. [387E.]
Smorodinsky M. [71] Ergodic Theory, Entropy. Springer, 1971 (Lecture Notes in Math., 214). [387
notes.]
epanek P. & Rubin M. [89] Homogeneous Boolean algebras, pp. 679-715 in Monk 89. [382S, 382
St
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 rel`evements 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.]
Truss J.K. [89] Innite permutation groups I: products of conjugacy classes, J. Algebra 120 (1989)
454-493. [382 notes.]
Vulikh B.C. [67] Introduction to the Theory of Partially Ordered Vector Spaces. Wolters-Noordho, 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.]
652
Index
Dedekind
653
654
Index
differentiable
Rd
formula for c f in terms of f 283F
and convolutions 283M
in terms of action on test functions 284H et seq.
of square-integrable functions 284O, 286U
inversion formulae for dierentiable functions 283I; for functions of bounded variation 283L
R
2
f (y) = lim0 12 eiyx ex 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
Fubinis
R theorem 252
RR
f d( ) =
f (x, y)dxdy 252B
when both factors -nite 252C, 252H
for characteristic functions 252D, 252F
function spaces 369, 374, 376
Rx
Rb
d
Fundamental Theorem of the Calculus ( dx
f = f (x) a.e.) 222; ( a F (x)dx = F (b) F (a)) 225E
a
Hahn decomposition of a countably additive functional 231E
on a Boolean algebra 326I
Hahn-Banach property see Nachbin-Kelley theorem (363R)
Lipschitz
655
656
Index
localizable
simple
657
658
Index
Sina
Sinas theorem (on existence of Bernoulli partitions for a measure-preserving automorphism) 387E, 387L
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 (, )-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
sucient 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)
special symbols
659
660
Index
General index
General index
References in bold type refer to denitions; references in italics are passing references. Denitions marked
with > are those in which my usage is dangerously at variance with that of some other author or authors.
Abels theorem 224Yi
Abel-Poisson kernel 282Yg
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, 233Xc, 244Yh, 252Yj, 256Xg, 262Bc, 263I, 264Yp, 265Ya, 282R, 282Yf, 283Ci
absolutely continuous submeasure 393D, 393E, 393H, 393Yf
Absoluteness Theorem see Shoenelds 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 nitely additive (136Xg, 231A), countably additive (231C)
additive function(al) on a Boolean algebra see nitely additive (326A, 361B), countably additive (326E),
completely additive (326J)
adherence point see cluster point (3A3La)
adjoint 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), -algebra
(111A)
Alexandro compactication see one-point compactication (3A3O)
almost continuous function 256F
almost every, almost everywhere 112Dd
almost isomorphic (inverse-measure-preserving functions) 385U, 385V, 385Xn-385Xp, 385Yf
almost surely 112De
amenable group 394 notes
analytic (complex) function 133Xc
angelic topological space 2A5J
antichain 316 notes
aperiodic Boolean homomorphism 381Ad, 381H, 381Ng, 381P, 381Xc, 381Xj, 381Xm, 382J, 386C, 386D,
386Ya, 388F, 388H, 388J, 388K, 388Yb; see also nowhere aperiodic (381Xk)
Archimedean partially ordered linear space 351R, 351Xe, 361Gb
Archimedean Riesz space 241F, 241Yb, 242Xc, 353A-353F, 353Ha, 353L-353P, 353Xa, 353Xb, 353Xd,
353Ya, 353Yd-353Yg, 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) 316K, 316L, 316Xq-316Xs, 316Xx, 316Yp, 322Bf, 324Ye, 331H, 333Xb,
343Xe, 362Xe; see also relative atom (331A)
atom (in a measure space) 211I, 211Xb, 246G, 322Bf
atom (in a Riesz space) 362Xe
atomic see purely atomic (211K, 316Kc)
Boolean
General index
661
atomless Boolean algebra 316Kb, 316Lb, 316Xq, 316Xr, 316Xt-316Xv, 316Yl-316Yn, 316Yo, 316Yp,
322Bg, 322Kc, 324Kf, 324Ye, 331Xk, 332I, 332P, 332Ya, 375B, 375Xb, 375Yb, 381P, 382Q, 384E, 384F,
384Xb, 386D, 386M, 387Cd, 387C-387H, 387L, 392Xg; see also relatively atomless (331A)
atomless additive functional 326Ya, 326Yb-326Yd, 326Yf, 362Xf, 362Xg, 362Yi
atomless measure algebra 331C, 369Xh, 374Xl, 383G, 383H, 383J, 383Xg, 383Xh, 383Ya, 384Ld, 384M,
384O, 384P, 384Xd
atomless measure (space) 211J, 211Md, 211Q, 211Xb, 211Yd, 211Ye, 212G, 213He, 214Xe, 215D, 215Xe,
215Xf, 216A, 216Ya, 234Ff, 234Yd, 251T, 251Wo, 251Xq, 251Xs, 251Yc, 252Yp, 252Yr, 252Yt, 254V, 254Yf,
256Xd, 264Yg, 322Bg, 325Ye, 342Xc, 343Cb, 343Xi, 344I
atomless vector measure 326Ye
automorphism see Banach lattice automorphism, Boolean automorphism,cyclic automorphism,ergodic automorphism, group automorphism (3A6B), induced automorphism (381M), inner automorphism (3A6B),
measure-preserving Boolean automorphism, measure-preserving ring automorphism, measure space automorphism, outer automorphism (387B)
automorphism group of a Boolean algebra 381A, 381B, 381Xa, 382O-382S, 382Xg, 382Yb, 383Xh, 383Ya,
383Yb, 384, 394; of a measure algebra 366Xh, 374J, 381Xj, 383-384, 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
Baires theorem 3A3G
Baire -algebra 254Xs, 341Yc, 341Zb, 343Xc, 344E, 344Ya, 344Yc-344Ye
ball (in R r ) 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, 354Yk-354Ym, 355C, 355K, 355Xb, 355Ya, 356D, 356M,
356Xc, 356Xk, 356Yg, 356Yh, 356Yi, 363E, 365K, 366C, 366Xd, 367D, 367P, 369B, 369G, 369Xe, 371B371E, 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, 354Yl, 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, 353Yc
Bernoulli partition 387A, 387B, 387D-387H, 387J
shift 385Q, 385R, 385S, 385Xi-385Xm, 386Xc, 387B, 387I, 387K, 387L, 387Xa, 387Xc, 387Ya
Berry-Esseen theorem 274Hc, 285 notes
bidual of a normed space (U ) 342Yc, 356Xh, 3A5Ga; see also order-continuous bidual
Bienaymes equality 272R
bilinear map 253 (253A), 255Xb, 363L, 376B, 376C, 376Ya-376Yc
Bochner integral 253Yf, 253Yg, 253Yi, 354Ym
Bochners theorem 285Xr
>311Ab), 363Xf; see also algebra of sets, complemented band algebra (352Q),
Boolean algebra Chap. 31 (>
Dedekind (-)complete Boolean algebra, measure algebra (321A), projection band algebra (352S), regular
open algebra (314Q)
662
Index
Boolean
Boolean automorphism 363Ye, 366Yh, 374Yd, 381C-381E, 381G, 381I, 381J, 381L, 381Q, 381Xc, 381Xf,
381Xg, 381Xo, 381Xp, 381Yc, 382A, 382B, 382D, 382E, 382G-382M, 382Xa, 382Xc, 382Xd, 384A, 388D,
388Ya, 394Ge, 394Ya, 395A; see also cyclic automorphism (381R), induced automorphism (381M), involution, measure-preserving Boolean automorphism, periodic automorphism (381Bc), von Neumann automorphism (388D)
Boolean homomorphism 312F, 312G-312K, 312N, 312P-312T, 312Xe, 312Xg, 312Xj, 312Ya, 313L-313N,
313P-313R, 313Xp, 313Yb, 314I, 314K, 314Xh, 324F, 324G, 324Xc, 324Yd, 326Be, 332Yc, 343A, 352Ta,
361Xd, 363F, 363Xb, 363Xc, 381B, 381H, 381Xj, 385K; see also aperiodic Boolean homomorphism (381Bd),
Boolean isomorphism, measure-preserving Boolean homomorphism, (sequentially) order-continuous Boolean
homomorphism, periodic Boolean homomorphism (381Bc)
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 -algebra (111Gd, 135C, 256Ye)
Borel-Cantelli lemma 273K
Borel lifting 345F, 345Xg, 345Yc
Borel measurable function 121C, 121D, 121Eg, 121H, 121K, 121Yf, 134Fd, 134Xg, 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, R r 111G, 111Yd, 114G, 114Yd, 115G, 115Yb, 115Yd, 121Ef, 121K, 134F, 134Xd, 135C,
136D, 136Xj, 212Xc, 212Xd, 254Xs, 225J, 264E, 264F, 264Xb, 342Xi, 364H, 364Yc
Borel -algebra (of subsets of Rr ) 111Gd, 114Yg, 114Yh, 114Yi, 121J, 121Xd, 121Xe, 121Yb, 216A,
251L, 364G, 382Yb;
(of other spaces) 135C, 135Xb, 256Ye, 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 R r ) 134E; (in a normed space) 2A4Bc, 3A5H; (in a linear topological space) 245Yf; see
also order-bounded (2A1Ab)
bounded support, function with see compact support
bounded variation, function of 224 (224A, 224K), 225Cb, 225M, 225Oc, 225Xh, 225Xn, 225Yc, 226Bc,
226C, 264Yp, 264Yb, 282M, 282O, 283L, 283Xj, 283Xk, 283Xm, 283Xn, 283Xq, 284Xk, 284Yd, 343Yc,
354Xt
bounding see -bounding (316Yg)
Breiman see Shannon-McMillan-Breiman theorem (386E)
Cantor function 134H, 134I, 222Xd, 225N, 226Cc, 262K, 264Xf, 264Yn
Cantor measure 256Hc, 256Ia, 256Xk, 264Ym
Cantor set 134G, 134H, 134I, 134Xf, 256Hc, 256Xk, 264J, 264Ym, 264Yn; see also {0, 1}I
Caratheodory complete measure space see complete (211A)
Caratheodorys method (of constructing measures) 113C, 113D, 113Xa, 113Xd, 113Xg, 113Yc, 114E,
114Xa, 121Yc, 132Xc, 136Ya, 212A, 212Xg, 213C, 213Xa, 213Xb, 213Xd, 213Xf, 213Xg, 213Xj, 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 partially ordered set) see conality (3A1F)
(of a topological space) see cellularity (332Xd), density (331Yf)
Carlesons theorem 282K remarks, 282 notes, 284Yg, 286U, 286V
category algebra (of a topological space) 314Yd
Cauchy distribution 285Xm, 367Xw
Cauchy lter 2A5F, 2A5G, 3A4F
compact
General index
663
664
Index
compactification
countably
General index
665
666
Index
countably
countably additive part (of an additive functional on a Boolean algebra) 326Yn, 362Bc, 362Ye, 393Yf
countably full group (of automorphisms of a Boolean algebra) 381Af, 381I, 381Ni, 381Yc, 382H-382M
countably separated (measure space, -algebra of sets) 343D, 343E-343H, 343K, 343L, 343Xe, 343Xh,
343Xi, 343Yb, 343Ye, 343Yf, 344B, 344C, 344I, 344Xb-344Xd, 381Xn, 382Xd, 383Xa, 385V, 388A, 388B
countably subadditive functional 392H
counting measure 112Bd, 122Xd, 122 notes, 211N, 211Xa, 213Xb, 226A, 241Xa, 242Xa, 243Xl, 244Xi,
244Xn, 245Xa, 246Xc, 251Xb, 251Xh, 252K, 255Yo, 264Db, 324Xe
cover see measurable envelope (132D)
covering theorem 221A, 261B, 261F, 261Xc, 261Ya, 261Yi, 261Yk
cozero set 3A3Qa
cycle notation (for automorphisms of Boolean algebras) 381R, 381S, 381T, 382Fb; see also pseudo-cycle
(388Xa)
cyclic automorphism 381R; see also exchanging involution (381R)
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, 316J,
316Xs, 316Yi, 316Yn, 331Yd, 332Xa, 333Bc, 352Q, 363Mb, 363P-363S, 364O, 368A-368D, 368K, 368M,
368Xe, 375C, 375D, 375Xa, 375Xc, 375Ya, 381C, 381D, 381F, 381Ib, 381Xb, 381Xf, 381Xk, 381Yd, 382E,
382F, 382N, 382Q-382S, 382Xe-382Xj, 384D, 384J, 393A-393C, 393J, 393Xc-393Xf, 393Ya, 393Yc-393Ye
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-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, 384Lf, 384Xc, 391Xc; see also localization (322P)
Dedekind completion of a Riesz space 368I, 368J, 368Xb, 368Yb, 369Xp
Dedekind -complete Boolean algebra 314C-314F, 314Hb, 314Jb, 314M, 314N, 314Xa, 314Xc, 314Ye,
314Yf, 315O, 316C, 316Fa, 316H, 316Xg, 316Yi, 316Yk, 321A, 324Yb, 324Yc, 326Fg, 326I, 326O, 326P,
326Xc, 326Xi, 326Yf, 326Yo, 344Yb, 362Xa, 363Ma, 363P, 363Xh, 363Yh, 364, 367Xk, 368Yd, 375Yb,
381C, 381D, 381Ia, 381K-381N, 381Xb, 381Xk, 381Xl, 381Yc, 382D, 382E, 382G-382M, 382S, 382Xk, 382Xl,
382Ya, 382Yc, 392H, 392Xd, 392Xf, 392Xg, 393A, 393E, 393M, 393O-393Q, 393S, 393Xb, 393Yf-393Yh
Dedekind -complete partially ordered set 314Ab, 315De, 367Bf; see also Dedekind -complete Boolean
algebra, Dedekind -complete Riesz space
Dedekind -complete Riesz space 241Fb, 241G, 241Xe, 241Yb, 241Yh, 242Yg, 243H, 243Xb, 353G,
353H, 353Ja, 353Xb, 353Yc, 353Yd, 354Xn, 354Yi, 354Ym, 356Xc, 356Xd, 363M-363P, 364C, 364E
delta function see Diracs delta function (284R)
delta system see -system
De Morgans laws 311Xf; see also distributive laws
dense set in a topological space 242Mb, 242Ob, 242Pd, 242Xi, 243Ib, 244H, 244Ob, 244Yi, 254Xo, 281Yc,
2A3U, 2A4I, 313Xj, 323Dc, 367O, 3A3E, 3A3G, 3A3Ie, 3A4Ff; see also nowhere dense (3A3Fa), orderdense (313J, 352Na), quasi-order-dense (352Na)
density (of a topological space) 331Ye, 331Yf, 365Ya, 393Ye
dual
General index
667
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), Lebesgues Density Theorem (223A), lower density (341C)
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 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-254T, 254Xp,
254Xr, 311Xh, 325N, 325Xh
Devils Staircase see Cantor function (134H)
dierentiability, points of 222H, 225J
dierentiable function (of one variable) 123D, 222A, 224I, 224Kg, 224Yc, 225L, 225Of, 225Xc, 225Xn,
233Xc, 252Yj, 255Xg, 255Xh, 262Xk, 265Xd, 274E, 282L, 282Xs, 283I-283K, 283Xm, 284Xc, 284Xk; (of
many variables) 262Fa, 262Gb, 262I, 262Xg, 262Xi, 262Xj; see also derivative
dierentiable relative to its domain 262Fb, 262I, 262M-262Q, 262Xd-262Xf, 262Yc, 263D, 263Xc,
263Xd, 263Yc, 265E, 282Xk
diused measure see atomless measure (211J)
dilation 286C
dimension see Hausdor dimension (264 notes)
Diracs 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 modied Dirichlet kernel (282Xc)
disconnected see extremally disconnected (3A3Ae), basically disconnected
discrete topology (3A3Ah)
disjoint family (of sets) 112Bb; (in a Boolean ring) 311Gb; (in a Riesz space) 352C
disjoint sequence theorem 246G, 246Ha, 246Yd, 246Ye, 246Yf, 246Yj, 354Rb, 356O, 356Xm
disjoint set (in a Boolean ring) 311Gb; (in a Riesz space) 352C
disjoint union topology 315Xe
distribution see Schwartzian distribution, tempered distribution
distribution of a random variable 241Xc, 271E, 271F, 271Ga, 272G, 272S, 272Xe, 272Yc, 272Yf, 272Yg,
285H, 285Xg; (of an element of L0 (A,
)) 364Xd, 364Xf, 367Xv, 373Xh; see also Cauchy distribution
(285Xm), empirical distribution (273 notes), Poisson distribution (285Xo)
of a nite family of random variables 271B, 271C, 271D-271G, 272G, 272Ye, 272Yf, 285Ab, 285C,
285Mb
distribution function of a random variable > 271G, 271L, 271Xb, 271Yb, 271Yc, 271Yd, 272Xe, 272Yc,
273Xg, 273Xh, 274F-274L, 274Xd, 274Xg, 274Xh, 274Ya, 274Yc, 285P
distributive lattice 367Yb, 382Xh, 3A1Ic; see also (2, )-distributive (367Yd)
distributive laws in Boolean algebras 313B; see also De Morgans laws, weakly -distributive (316Yg),
weakly (, )-distributive (316G)
distributive laws in Riesz spaces 352E; see also weakly -distributive, weakly (, )-distributive (368N)
Dominated Convergence Theorem see Lebesgues Dominated Convergence Theorem (123C)
Doobs Martingale Convergence Theorem 275G
double arrow space see split interval (343J)
doubly recurrent Boolean automorphism 381Ag, 381M, 381N, 381Xk, 381Xl, 381Xp, 382J
doubly stochastic matrix 373Xe
downwards-directed partially ordered set 2A1Ab
dual linear operator see adjoint operator (3A5Ed)
dual normed space 243G, 244K, 2A4H, 356D, 356N, 356O, 356P, 356Xg, 356Yh, 365L, 365M, 366C,
366D, 369K, 3A5A, 3A5C, 3A5E-3A5H
668
Index
dual
see also bidual, order-bounded dual (356A), order-continuous bidual, order-continuous dual (356A),
sequentially order-continuous dual (356A)
Dunfords theorem 376N
dyadic cycle system (for a Boolean automorphism) 388D, 388H, 388J
Dyes theorem 388L
Dynkin class 136A, 136B, 136Xb
Eberleins theorem 2A5J, 356 notes
Egorovs theorem 131Ya, 215Yb
Egorov property (of a Boolean algebra) 316Yg, 316Yh
embedding see regular embedding (313N)
empirical distribution 273Xh, 273 notes
entropy 385; (of a partition of unity) 385C, 386E, 386I, 386L, 386O, 387B-387D, 387F-387H; (of a
measure-preserving Boolean homomorphism) 385M, 386L, 386Xe, 386Yb, 386Yc, 387C-387E, 387I-387L,
388Yb; see also conditional entropy (385D), entropy metric (385Xq)
entropy metric 385Xq, 385Yg
enumerate, enumeration 3A1B
envelope see measurable envelope (132D), upr(a, C) (314V)
equidecomposable (in G-equidecomposable) 395Ya; (in G- -equidecomposable) 394 (394A)
equidistributed sequence (in a topological probability space) 281N, 281Yi, 281Yj, 281Yk
Equidistribution Theorem see Weyls Equidistribution Theorem (281N)
equivalent norms 355Xb, 369Xd
equiveridical 121B, 212B, 312B
ergodic automorphism (of a Boolean algebra) 363F, 381Xl, 382Xk, 385Se, 385Ta, 387C, 387D, 387E,
387J, 387Xb, 388Xe, 388Yc, 394Ge
ergodic Boolean homomorphism 372Pa, 381P, 386D, 386F
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, 372Yl, 388Yd
ergodic measure-preserving Boolean homomorphism (of a measure algebra) 372Q, 372Xk, 372Xl, 372Yl,
385Se, 386D, 386F, 387C-387E, 387J, 387Xb, 388Yc, 395Xc
Ergodic Theorem 372D, 372E, 372Ya, 386Xc; see also Maximal Ergodic Theorem (372C), Mean Ergodic
Theorem (372Xa), Wieners 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
Etemadis lemma 272U
Euclidean metric (on R r ) 2A3Fb
Euclidean topology 1A2, 2A2, 2A3Ff, 2A3Tc
even function 255Xb, 283Yb, 283Yc
exchangeable sequence of random variables 276Xe
exchanging involution 381Ra, 381S, 382C, 382E, 382H, 382K, 382L, 382M, 382P, 382Ya
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 nitely additive functionals 391G
extension of measures 132Yd, 212Xk, 327Xf; see also completion (212C), c.l.d. version (213E)
extremally disconnected topological space 314S, 314Xj, 353Yb, 363Yc, 364W, 364Yk-364Ym, 368G,
368Yc, 3A3Ae, 3A3Bd
fully
General index
669
670
Index
function
function 1A1B
function space see Banach function space (369)
functionally closed see zero set (3A3Qa)
functionally open see cozero set (3A3Qa)
Fundamental Theorem of Calculus 222E, 222H, 222I, 225E
Fundamental Theorem of Statistics 273Xh, 273 notes
Gaifmans example 391N, 391Yf-391Yh,
Gamma function see -function (225Xj)
Gauss C.F. 372Yg
Gaussian distribution see standard normal distribution (274Aa)
Gaussian random variable see normal random variable (274Ad)
generated topology 3A3Ma
generated (-)algebra of sets 111Gb, 111Xe, 111Xf, 121J, 121Xd, 136B, 136C, 136G, 136Xc, 136X,
136Xl, 136Yb
generated ()-subalgebra of a Boolean algebra 313F, 313M, 314G, 314Ye, 331E
generating Bernoulli partition 387Ab, 387Bg
generating set in a Boolean algebra 331E; see also -generating (331E), -generating (331E)
Glivenko-Cantelli theorem 273 notes
graph of a function 264Xf, 265Yb
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)
Hahn-Banach theorem 363R, 373Yd, 3A5A, 3A5C
Hahn decomposition of an additive functional 231Eb, 231F, 326I, 326O
see also Vitali-Hahn-Saks theorem (246Yg)
Hajian-Ito theorem 395B
half-open interval (in R or R r ) 114Aa, 114G, 114Xe, 114Yj, 115Ab, 115Xa, 115Xc, 115Yd
Halls Marriage Lemma 3A1K
Halmos-Rokhlin-Kakutani lemma 386C
Hardy-Littlewood Maximal Theorem 286A
Hausdor dimension 264 notes
Hausdor measure 264 (264C, 264Db, 264K, 264Yo), 265Yb, 343Ye, 345Xb, 345Xg; see also normalized Hausdor measure (265A)
Hausdor metric (on a space of closed subsets) 246Yb
Hausdor outer measure 264 (264A, 264K, 264Yo)
Hausdor topology 2A3E, 2A3L, 2A3Mb, 2A3S, 3A3Aa, 3A3D, 3A3Id, 3A3K, 3A4Ac, 3A4Fd
Hausdor uniformity 3A4Ac
Hilbert space 244N, 244Yj, 3A5L; see also inner product space (3A5L)
Hilbert transform 376Ym
Holders inequality 244Eb
homeomorphism 3A3Ce
homogeneous Boolean algebra 331M, 331N, 331Xj, 331Xk, 331Yg-331Yj, 381D, 382Q, 382S, 384E384G, 384La, 384Xb; see also Maharam-type-homogeneous (331Fc), relatively Maharam-type-homogeneous
(333Ac)
measure algebra 331N, 331Xk, 332M, 373Yb, 374H, 374Yc, 375Kb, 383E, 383F, 383I, 394R; see also
quasi-homogeneous (374G)
measure space 344L
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)
integral
General index
671
672
Index
integration
limit
General index
673
674
Index
limit
measurable
General index
675
lower semi-continuous function 225H, 225I, 225Xl, 225Xm, 225Yd, 225Ye, 323Cb, 367Xx
Lusins theorem 134Yd, 256F
Luxemburg, W.A.J. 363 notes
magnitude of (an element in) a measure algebra 332Ga, 332J, 332O, 383Xd, 383Xe, 383Xh
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, 392N, 392O, 392Xc, 392Xd, 392Xf, 392Xg, 392Ya, 392Yb,
393A-393C, 393E, 393M, 393Xa-393Xc, 393Xe, 393Xf, 393Ya, 393Ye, 393Yf
Maharams theorem 331I, 332B
Maharam type of a Boolean algebra 331F, 331H, 331Xc, 331Xg, 331Xh, 331Xj, 331Yd, 332H, 365Ya,
373Yb, 382Xl, 393Xe, 393Xf, 393Ya, 393Ye; see also relative Maharam type (333Aa)
of a measure algebra 331I-331K, 331Xd-331Xf, 331Xi, 332M, 332N, 332R-332T, 333D, 334,
369Xg, 375Kb, 387K, 387L, 387Xc, 388K, 388L
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- component in a Boolean algebra 332Gb, 332H, 332J, 332O, 332P, 332Xj, 332Ya, 384N,
394Xh
many involutions (group with many involutions) 382O, 382P-382R, 382Xe, 382Xg-382Xj, 383G, 384A,
384C, 384D, 384I, 394Xc
Marczewski functional 343E
Markov time see stopping time (275L)
Marriage Lemma see Halls Marriage Lemma (3A1K)
martingale 275 (275A, 275Cc, 275Cd, 275Ce); see also reverse martingale
martingale convergence theorems 275G-275I, 275K, 275Xf, 367K, 369Xq
martingale dierence sequence 276A, 276B, 276C, 276E, 276Xd, 276Ya, 276Yb, 276Ye, 276Yg
martingale inequalities 275D, 275F, 275Xb, 275Yc-275Ye, 276Xb, 372 notes
Max-ow 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
McMillan see Shannon-McMillan-Breiman theorem (386E)
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, 392J, 393A, 393J, 393Yd, 394P, 394Q, 394Xf, 395B,
395Xd
measurable cover see measurable envelope (132D)
measurable envelope 132D, 132E, 132F, 132Xf, 132Xg, 134Fc, 134Xd, 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 R r ) 121Yf), 256G
(taking values in other spaces) 133Da, 133E, 133Yb, 135E, 135Xd, 135Yf
((, T)-measurable function) 121Yb, 235Xc, 251Ya, 251Yc, 343A
see also Borel measurable, Lebesgue measurable
measurable set 112A; -measurable set 212Cd; see also relatively measurable (121A)
676
Index
measurable
order-continuous
General index
677
678
Index
order-continuous
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, 381Ec, 381F, 395Xasee also
regular embedding (313N)
order-continuous dual Riesz space (U ) 356 (356Ac), 362A, 363K, 363S, 365L, 365M, 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, 356Yi, 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, 385Ya, 386Yb, 386Yc, 392I, 392Xb, 394N; see
also sequentially order-continuous
order-continuous positive linear operator 351Ga, 355G, 355H, 355K, 364Rc, 365P, 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, 353Oa,
353Xc, 356I, 361Je, 355F, 365O, 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, 392K, 392L, 393Xa, 393Yd; (in C(X)) 367L, 367Yi,
367Yj, 367Yk; (in L0 ()) 245C, 245K, 245L, 245Xc, 245Xd, 376J; (in L0 (A)) 367, 368Yi, 372D-372G,
372Xb, 372Yo, 386E, 386F
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, 367Ob, 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, 316Xq, 316Yk, 316Yn, 323Dc, 363Xd, 391Xj
order-isomorphism 312L
order*-limit see order*-convergent (367A)
order-preserving function 313H, 313I, 313La, 315D, 326Bf, 361Ce
order-sequential topology (on a lattice) 367Yk, > 392K, 392L, 392M, 392O, 392Xh, 392Xi, 393Xa
order topology (on a partially ordered set) 313Xb, 313Xj, 313Yc, 326Yr; see also order-sequential topology
(392K)
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, 354I-354K, 354Yi, 354Yj, 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
Ornsteins theorem 387I, 387K
orthogonal matrix 2A6B, 2A6C
orthogonal projection in Hilbert space 244Nb, 244Yj, 244Yk, 366K
orthonormal vectors 2A6B
outer automorphism of a group 384G, 384O, 384Pb, 384Q, 384Yc, 384Yd, 3A6B
outer measure 113 (113A), 114Xd, 132B, 132Xg, 136Ya, 212Ea, 212Xa, 212Xb, 212Xg, 213C, 213Xa,
213Xg, 213Xk, 213Ya, 251B, 251Wa, 251Xd, 254B, 264B, 264Xa, 264Ya, 264Yo, 393T, 393Xh, 393Yi;
see also Lebesgue outer measure (114C, 115C), metric outer measure (264Yc), regular outer measure
(132Xa), submeasure (392A)
principal
General index
679
dened from a measure 113Ya, 132 (132B), 213C, 213F, 213Xa, 213Xg-213Xj, 213Xk, 213Yd,
214Cd, 215Yc, 235Ya, 251O, 251R, 251Wk, 251Wm, 251Xm, 251Xo, 252Yh, 254G, 254L, 254S, 254Xb,
254Xq, 254Yd, 264Fb, 264Ye
outer regular measure 134Fa, 134Xe, 256Xi
Parsevals 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 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 1A1J
partition of unity in a Boolean ring or algebra 311G, 313K, 313L, 313Xk, 315E, 315F, 315Xk, 315Xq,
316H, 322E, 332E, 332I, 332Xi, 332Yb, 352T, 375H, 375I, 381C, 381D, 381H, 381Ia, 381N, 381Xf, 381Xg,
382Fb, 383F, 385C, 385D, 385G-385P, 385R, 385Xa-385Xe, 385Xi, 385Xq, 385Xr, 385Ya-385Yc, 385Yg,
386E, 386I, 386J, 386L, 386M, 386O, 386Xd, 386Xe, 388Ib, 392Xg; see also Bernoulli partition (387A)
pathological submeasure 393Tc
Peano curve 134Yl-134Yo
perfect measure (space) 342K, 342L, 342M, 342Xh-342Xo, 343K-343M, 343Xg, 343Xh, 344C, 344I,
344Xb-344Xd, 344Yf, 385V
perfect Riesz space 356J, 356K-356M, 356P, 356Xg, 356Xi-356Xk, 356Ye, 356Yg, 365C, 365N, 366D,
369D, 369K, 369Q, 369Ya, 374M, 374Xk, 374Xl, 374Ya, 376P, 376Xl
periodic Boolean automorphism 381Ac, 381H, 381R, 381Xj, 381Xm, 381Xn, 382F
periodic extension of a function on ], ] 282Ae
Plancherel Theorem (on Fourier series and transforms of square-integrable functions) 282K, 284O, 284Qd
point-supported measure 112Bd, 112Xg, 211K, 211O, 211Qb, 211Rc, 211Xb, 211Xf, 213Xo, 234Xc,
215Xr, 256Hb, 343Xi
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
Poissons theorem 285Q
see also Abel-Poisson kernel (282Yg)
polar coordinates 263G, 263Xf
Polyas urn scheme 275Xc
polynomial (on R r ) 252Yu
porous set 223Ye, 261Yg, 262L
positive cone 253G, 253Xi, 253Yd, 351C
positive denite 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, 314Xi,
315E, 315Xl, 316Xb, 316Xk, 316Xu, 321Xa, 322H-322J, 322Xf, 323Xe, 325Xb, 331Fb, 331H, 331M, 332A,
332L, 332P, 332Xh, 352Sc, 352Xf, 364Xo, 384Lb, 391Xe, 391Xj
principal projection property (of a Riesz space) 353Xb, 353Yd, 354Yi
principal ultralter 2A1N
680
Index
probability
probability algebra 322Aa, 322Ba, 322Ca, 322G, 322Mb, 385, 386E, 386H-386J, 386L-386O, 386Xe,
386Yb, 386Yc, 387, 391B
probability algebra free product 325F, 325G, 325I-325M (325K), 325Xd, 334D, 334Xc, 372Yl, 385Xm,
388Yc
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, 385S; see also {0, 1}I
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 subalgebra of a Boolean algebra) see upper envelope (314Va); (on a Riesz space) see
band projection (352Rb)
projection band (in a Riesz space) 352R (352Ra), 352S, 352T, 352Xg, 353E, 353Hb, 353I, 353Xb, 361Xh,
361Ye, 362B
projection band algebra 352S, 352T, 361K, 363J, 363N, 363O, 364Q
pseudo-cycle (for a Boolean automorphism) 388Xa
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 316Kc, 316L, 316Xq-316Xw, 316Yk, 316Yl, 316Yp, 322Bh, 322Kc, 324Kg,
331Yd, 332Xb, 332Xc, 362Xf, 385J, 385Xq, 394Xf
purely atomic measure (space) 211K, 211N, 211R, 211Xb, 211Xc, 211Xd, 212G, 213He, 214Xe, 234Xb,
251Xr, 322Bh, 342Xn, 343Xi, 375Ya
purely innite measurable set 213 notes
push-forward measure see image measure (112F)
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
quasi-simple function 122Yd, 133Yd
quasi-Stonian topological space 314Yf, 353Yc
quotient Boolean algebra (or ring) 312K, 312Xk, 313Q, 313Xo, 314C, 314D, 314M, 314N, 314Yd, 316C,
316D, 316Xh, 316Xi, 316Xn, 316Xt, 316Ye, 316Yo, 341Xf, 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
Rademachers theorem 262Q
Radon measure(on R or R r ) 256 (256A), 284R, 284Yi, 342Jb, 342Xj; (on ], ]) 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-232G, 234G, 235Xk, 242I, 244Yk, 327D, 365E, 365Xf
Riesz
General index
681
Radon probability measure (on R or R r ) 271B, 271C, 271Xb, 285Aa, 285M, 364Xd, 372Xf; (on other
spaces) 256Ye, 271Ya
Radon product measure (of nitely 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) 386Xa, 386Xb, 386A
recurrent Boolean homomorphism 381Ag, 381L, 381O, 381P, 381Q, 381Xh, 381Xi
recursion 2A1B
reduced power of R 351M, 351Q, 351Yd, 352L, 352M, 368F
rene 311Gd
reexive 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), 314Xi, 314Xj, 314Yd, 315Xe, 315Yd,
316Xe, 316Yb, 316Yd, 316Yf, 316Yj, 316Yp, 332Xe, 363Yc, 364U-364W, 364Yh-364Yl, 368Ya, 391Yc,
391Yf; (of R) 316J, 316Yb, 316Yn, 331Xh, 331Yi, 367Yq, 375Xe, 375Yf, 375Z, 376Yi, 391Xd, 392Xh,
392Yb
regular open set 314O, 314P, 314Q
regular operator (between Riesz spaces) 355 notes
regular outer measure 132C, 132Xa, 213C, 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, 316Xr, 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, 381P, 386C, 388I, 394Xc
relatively compact set 2A3Na, 2A3Ob, 3A3De, 3A5M
relatively Maharam-type-homogeneous 333Ac, 333Bb
relatively measurable set 121A
relatively von Neumann automorphism 388Da, 388K, 388Xe, 388Yb, 388Yc, 388Yf
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 Fubinis theorem, Tonellis theorem
representation of homomorphisms (between Boolean algebras) 344Ya, 344Yd, 364Xr; (between measure
algebras) 324A, 324B, 324N, 343A, 343B, 343G, 343J, 343M, 343Xc, 343Xf, 343Yd, 344A-344C, 344E-344G,
344Xf, 344Yc
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, 354Yj, 355Xe, 356Xh, 361Gc, 361J,
361Xg, 362Xe, 363Ec, 363F, 363Xb, 363Xc, 364R, 365K, 375I, 375J, 375Xg, 375Ya, 376Cc; see also ordercontinuous Riesz homomorphism
Riesz norm 242Xg, 354A, 354B, 354D, 354F, 354M, 354Xc-354Xf, 354Xh, 354Yb, 354Yf, 354Yl, 355Xc,
356D, 356Xg, 356Xh; see also Fatou norm (354Da), order-continuous norm (354Dc), order-unit norm
(354Ga)
682
Index
Riesz
Riesz space (= vector lattice) 231Yc, 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 384Ha, 384L; see also nowhere rigid (384Hb)
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 (386C)
root algebra (of a Bernoulli shift) 385Q, 385R, 385S, 387B, 387Ya
Saks see Vitali-Hahn-Saks theorem (246Yg)
saltus function 226B, 226Db, 226Xa
Schroder-Bernstein theorem 2A1G, 332 notes, 344D, 344Xa
Schwartz function see rapidly decreasing test function (284A)
Schwartzian distribution 284R, 284 notes; see also tempered distribution (284 notes)
self-supporting set (in a topological measure space) 256Xf
semi-continuous function see lower semi-continuous (225H)
semi-nite measure algebra 322 (322Ad), 323Dd, 323Ga, 323Xa, 324K, 324Xb, 325Ae, 325D, 327B,
331C, 331Xl, 332E, 332F, 332I, 332R, 332Xi, 332Yb, 364L, 365E, 365G, 365M, 365P, 365Sb, 366E, 366Xe,
366Xf, 366Xk, 367Nb, 368S, 369H, 369Xa, 371Xc, 373R, 375I, 383E, 383F, 383Ga, 383I, 384Ld, 384P, 384Xe,
391Ca
semi-nite measure (space) 211F, 211L, 211Xf, 211Ya, 212G, 213A, 213B, 213Hc, 213Xc, 213Xd, 213Xj,
213Xl, 213Xm, 213Ya-213Yc 214Xe, 214Xh, 215B, 216Xa, 216Yb, 234Fa, 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-nite version of a measure 213Xc, 213Xd, 322Xb
semigroup see full local semigroup (394A)
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
separatedsee countably separated (343D)
separator (for a Boolean automorphism) 382Aa, 382B-382E, 382I, 382J, 382L, 382M, 382Xa, 382Xd,
382Xl
sequentially order-closed set in a partially ordered space 313Db, 313Xg, 313Yc, 316Fb, 353Ja, 364Xm,
367Yc; see also -ideal (313Ec), -subalgebra (313Ec)
sequentially order-continuous additive function (on a Boolean algebra) 326Gc, 363Eb
Boolean homomorphism 313Lc, 313Pb, 313Qb, 313Xq, 313Yb, 314Fb, 314Hb, 314Xc, 314Ye,
315Ya, 316Fd, 324A, 324B, 324Kd, 324Xa, 324Xe, 324Yc, 326Ff, 343Ab, 363Ff, 364G, 364H, 364R, 364Xp,
364Xv, 364Yc, 364Yf, 364Yn, 365H, 365Xg, 375Ya, 375Ye, 381K
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 386E, 386Xc
strongly
General index
683
shift operators (on function spaces based on topological groups) 286C; see also Bernoulli shift (385Q)
Shoenelds 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 382S, 383I, 383Xb-383Xd
simple product of Boolean algebras 315 (315A), 316Xc, 316Xl, 316Xv, 332Xa, 332Xg, 364S, 384Lc,
391Xb, 391Xj
of measure algebras 322K, 323L, 325Xc, 332B, 332Xm, 332Xn, 333H, 333Ia, 333K, 333R, 366Xi
Sinas theorem 387E, 387L
singular additive functional 232Ac, 232I, 232Yg
singular measures 231Yf, 232Yg, 362Xa
smooth function (on R or R r ) 242Xi, 255Xi, 262Yd-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, 383J; see also band (352O)
Souslin property see ccc (316Ab)
space-lling 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
step-function 226Xa
Stieltjes measure see Lebesgue-Stieltjes measure (114Xa)
Stirlings 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, 381Q, 381Xm, 382Xa
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, 211Xf, 211Ye, 212G, 213Ha, 213J, 213L, 213O,
213Xa, 213Xh, 213Xn, 213Ye, 214Ia, 214J, 215Xf, 216E, 234Fd, 235P, 251N, 251P, 251Xn, 252B, 252D,
252E, 252Ys, 252Yt, 322Kd, 322N, 322Qb, 322Xi, 325He, 341H, 341K, 342Hb, 343B, 343Xi, 344C, 344I,
344Xb, 344Xc, 346Xd, 385V
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, 392Xg, 392Ya, 393Xa, 393Xb
strong law of large numbers 273D, 273H, 273I, 273Xh, 275Yn, 276C, 276F, 276Ye, 276Yg
strong measure-algebra topology 323Ae, 323Xg, 366Yi
strong operator topology 366Yi, 372Yn, 388Yf, 3A5M
strongly mixing see mixing (372P)
684
Index
subalgebra
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), xed-point subalgebra (394Ga), independent subalgebras (325L), order-closed subalgebra, order-dense subalgebra, regularly embedded subalgebra (313N),subalgebra (233A, 313E)
subhomomorphism see -subhomomorphism (375E)
sublattice 3A1Ib
submartingale 275Yf, 275Yg
submeasure (on a Boolean algebra) 392 (392A), 393B, 393T, 393U, 393Xi, 393Yb; see also Maharam
submeasure (392G), outer measure (113A), pathological submeasure (393Tc), strictly positive submeasure
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, 254La, 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, 214Xm
subspace of a normed space 2A4C
subspace topology 2A3C, 2A3J, 3A4Db
subspace uniformity 3A4D
subspace -algebra 121A, 214Ce
substitution see change of variable in integration
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
support of an additive functional on a Boolean algebra 326Xi
support of a Boolean homomorphism 381B, 381G, 381H, 381Xn, 382E, 382I, 382K, 382N, 382P-382R,
384B
support of a topological measure 256Xf, 257Xd
support of a submeasure on a Boolean algebra 392Xe
support see also bounded support, compact support
supported see point-supported (112Bd)
supporting (element in a Boolean algebra supporting an automorphism) 381Ba, 381E, 381Jb, 381Qb,
381Sa, 381Xk, 382D, 382N, 382O, 384B; see also self-supporting set (256Xf), support
supremum 2A1Ab
in a Boolean algebra 313A-313C
surface measure see normalized Hausdor measure (265A)
symmetric distribution 272Ye
symmetric dierence (in a Boolean algebra) 311G
symmetrization see Steiner symmetrization
Tarskis 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, 385Xk
test function 242Xi, 284 notes; see also rapidly decreasing test function (284A)
thick set see full outer measure (132F)
tight see uniformly tight (285Xj)
Tonellis theorem 252G, 252H, 252R
upper
General index
685
686
Index
upwards
General index
687
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, 3A3Qa
Zorns 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 Aut A) see automorphism
group of a measure algebra (383A)
AL (Lebesgue measure algebra) 373C
B (in B(x, ), 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, 383Xd, 391Yd, 391Ye
C = the set of complex numbers; (in R
C ) 2A1A
C (in C(X), where X is a topological space) 243Xo, 281Yc, 281Ye, 281Yf, 352Xj, 353M, 353Xd, 353Yc,
354L, 354Yf, 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 (the space of convergent sequences) 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, 316Yo, 322G, 324Yd, 325Yd,
326L, 326Xi, 331Ge, 332D, 332H, 363Yb, 364Yb, 367Yl, 368Yg, 368Yi, 381Yd, 391M, 391Xa, 392Ca, 392I,
392M-392O, 393J, 393Yd, 394Xf, 394Yb
ccc topological space 316Ab, 316B, 316Xd, 316Xe, 316Ya, 316Yd
cf (in cf P ) see conality (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, 354Ym, 376J, 376R, 376S, 376Yc
c.l.d. version of a measure (space) 213E, 213F-213H, 213M, 213Xb-213Xe, 213Xg, 213Xj, 213Xk, 213Xn,
213Xo, 213Yb, 214Xf, 214Xj, 232Ye, 234Yf, 241Ya, 242Yh, 244Ya, 245Yc, 251Ic, 251S, 251Wn, 251Xd,
251Xj, 251Xk, 252Ya, 322D, 322Qb, 322Xc, 322Xi, 322Yb, 324Xc, 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, ), where A is a subset of a Boolean algebra, and is a homomorphism) 385K, 385L, 385M
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, 353Yf, 353Yg, 361Eh, 363B,
364C-364E, 367Yg
688
Index
General index
689
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 limxF ) 2A3S
lim inf (in lim inf n ) 1A3 (1A3Aa), 2A3Sg; (in lim inf 0 ) 2A2H; (in lim inf xF ) 2A3S
lim sup (in lim supn ) 1A3 (1A3Aa), 2A3Sg; (in lim sup0 ) 2A2H, 2A3Sg; (in lim supxF f (x))
2A3S
ln+ 275Yd
M (in M (A), space of bounded nitely 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,
) = M0 ) 366F, 366G, 366H, 366Yb, 366Yd, 366Yg, 373D, 373P, 373Xk
0,
M
252Yp
M 0, (in M 0, (A,
) = M0, ) 373C, 373D, 373E, 373F, 373I, 373Q, 373Xo, 374B, 374J, 374L
M 1,0 (in M 1,0 (A,
) = M1,0 ) 366F, 366G, 366H, 366Ye, 369P, 369Q, 369Yh, 371F, 371G, 372D, 372Ya,
373G, 373H, 373J, 373S, 373Xp, 373Xr, 374Xe
M 1, (in M 1, ()) 234Yd, 244Xl, 244Xm, 244Xo, 244Yc; (in M 1, (A,
) = M1, ) 369N, 369O-369Q,
369Xi-369Xk, 369Xm, 369Xq, 373, 374A, 374B, 374M
M ,0 (in M ,0 (A,
)) 366Xd, 366Yc
M ,1 (in M ,1 (A,
) = M,1 ) 369N, 369O, 369P, 369Q, 369Xi, 369Xj, 369Xk, 369Xl, 369Yh, 373K,
373M, 374B, 374M, 374Xa, 374Ya
M (in M (A), space of countably additive functionals) 362B, 362Xd, 362Xh, 362Xi, 362Ya, 362Yb, 363K
M (in M (A), space of completely additive functionals) 326Yp, 327D, 362B, 362D, 362Xd, 362Xg, 362Xi,
362Ya, 362Yb, 363K
534D, 534E-534H, 534L-534Q, 534T, 534Xc, 534Xd, 534Xl, 534Xf, 534Xg, 534Xk, 534Yb, 534Z
med (in med(u, v, w)) 352Xn
N 3A1H; see also PN
N N 111Fb
N N 372Xi
On (the class of ordinals) 3A1E
p (in p(t)) 386G, 386H
P (in PX) 311Ba, 311Xe, 312B, 312C, 313Ec, 313Xf, 363S, 382Xc, 383Yb;(usual measure on PX) 254J,
254Xf, 254Xq, 254Yd
PN 1A1Hb, 2A1Ha, 2A1Lb, 315O, 326Yg, 374Xk; (usual measure on) 273G, 273Xd, 273Xe
P(N N ) 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)) 385A, 386M
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, 367Oc, 368Q, 369O; (in S f
= S(Af ))
242M, 244H, 365F, 365Gb, 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
690
Index
sf (in sf ) see semi-nite version of a measure (213Xc); (in sf ) 213Xf, 213Xg, 213Xk
T1 topology 3A3Aa, 3A3Pa
T2 topology see Hausdor (2A3E, 3A3Aa)
(0)
T (0) (in T, ) 371F, 371G, 371H, 372D, 372Xb, 372Yb, 372Yc, 373B, 373G, 373J, 373R, 373S, 373Xp,
373Xq, 373Xr, 373Xu, 373Xv
T (in T, ) 244Xm, 244Xo, 244Yc, 246Yc, 373 (373A); see also T -invariant (374A)
T (in T, ) 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, )) 1A2A
upr (in upr(a, C)) 314V, 314Xg, 333Xa, 365Rc, 386B, 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
r (volume of unit ball in R r ) 252Q, 252Xh, 265F, 265H, 265Xa, 265Xb, 265Xe
-function 225Xj, 225Xk, 252Xh, 252Yk, 252Yn, 255Xj
-system 2A1Pa
G (standard normal distribution) 274Aa
special symbols
General index
691
(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
=a.e. 112Dg, 112Xh, 222E, 241C
a.e. 112Dg, 112Xh, 212B
a.e. 112Dg, 112Xh, 233I
692
Index
special symbols
R
see lower integral (133I)
R
R see Riemann integral (134K)
(in f A, 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 ()) 242 (242D), 246F, 253E, 275Xd, 282Ye; (on L1 ()) 242D, 242Yg, 273Na, 273Xi; (on
1
L (A,
)) 365A, 365B, 365C, 386E, 386F; (on the 1 -sum of Banach lattices) 354Xb, 354Xo
k k2 244Da, 273Xj, 282Yf, 366Yh; see also L2 , k kp
k kp (for 1 < p < ) 244 (244Da), 245Xm, 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, 373Xl; 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 iI Ai ) see free product (315H)
b (in T)
b
(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 2 ) 3A1D
(in A , in a Boolean algebra) 313Xo; (in A , in a Riesz space) 352O, 352P, 352Q, 352R, 352Xg; see
also complement of a band (352B)
a
(in z a i) 3A1H
special symbols
General index
693