(Atlantis Studies in Mathematics 2) Alexander B. Kharazishvili-Topics in Measure Theory and Real Analysis-Atlantis Press - World Scientific (2009)

ATLANTIS STUDIES IN MATHEMATICS
VOLUME 2
SERIES EDITOR: J. VAN MILL
Atlantis Studies in Mathematics
Series Editor:
J. van Mill, VU University Amsterdam, Amsterdam, the Netherlands
(ISSN: 1875-7634)
Aims and scope of the series
The series Atlantis Studies in Mathematics (ASM) publishes monographs of high quality
in all areas of mathematics. Both research monographs and books of an expository nature
are welcome.
All books in this series are co-published with World Scientic.
For more information on this series and our other book series, please visit our website at:
www.atlantis-press.com/publications/books
AMSTERDAM PARIS
c ATLANTIS PRESS / WORLD SCIENTIFIC
Topics in Measure Theory
and Real Analysis
Alexander B. Kharazishvili
A. Razmadze Mathematical Institute
Tbilisi
Republic of Georgia
AMSTERDAM PARIS
Atlantis Press
29, avenue Laumi` ere
75019 Paris, France
For information on all Atlantis Press publications, visit our website at:
www.atlantis-press.com
Copyright
This book, or any parts thereof, may not be reproduced for commercial purposes in any
form or by any means, electronic or mechanical, including photocopying, recording or any
information storage and retrieval system known or to be invented, without prior permission
from the Publisher.
Atlantis Studies in Mathematics
Volume 1: Topological Groups and Related Structures - A. Arhangelskii, M. Tkachenko
ISBN: 978-90-78677-20-8
ISSN: 1875-7634
c 2009 ATLANTIS PRESS / WORLD SCIENTIFIC
e-ISBN: 978-94-91216-36-7
Preface
This book is concerned with questions of classical measure theory and related topics of
real analysis. At the beginning, it should be said that the choice of material included in the
present book was completely dictated by research interests and preferences of the author.
Nevertheless, we hope that this material will be of interest to a wide audience of mathe-
maticians and, primarily, to those who are working in various branches of modern math-
ematical analysis, probability theory, the theory of stochastic processes, general topology,
and functional analysis. In addition, we touch upon deep set-theoretical aspects of the top-
ics discussed in the book; consequently, set-theorists may detect nontrivial items of interest
to them and nd out new applications of set-theoretical methods to various problems of
measure theory and real analysis. It should also be noted that questions treated in this book
are related to material found in the following three monographs previously published by
the author.
1) Transformation Groups and Invariant Measures, World Scientic Publ. Co., London-
Singapore, 1998.
2) Nonmeasurable Sets and Functions, North-Holland Mathematics Studies, Elsevier, Am-
sterdam, 2004.
3) Strange Functions in Real Analysis, 2nd edition, Chapman and Hall/CRC, Boca Raton,
2006.
For the convenience of our readers, we will rst, briey and schematically, describe the
scope of this book.
In Chapter 1, we consider the general problem of extending partial real-valued functions
which, undoubtedly, is one of the most important problems in all of contemporary math-
ematics and which deserves to be discussed thoroughly. Since the satisfactory solution to
this task requires a separate monograph, we certainly do not intend on entering deeply into
v
vi Topics in Measure Theory and Real Analysis
various aspects of the problem of extending partial functions, but rather we restrict our-
selves to several examples that are important for real analysis and classical measure theory
and vividly show the fundamental character of this problem. The corresponding examples
are given in Chapter 1 and illustrate different approaches and appropriate research methods.
Notice that some of the examples presented in this chapter are considered in more details
in subsequent sections of the book.
Chapter 2 is devoted to a special, but very important, case of the extension problem for
real-valued partial functions. Namely, we discuss therein several variants of the so-called
measure extension problemand we pay our attention to purely set-theoretical, algebraic and
topological aspects of this problem. In the same chapter, the classical method of extending
measures, developed by Marczewski (see [234] and [235]), is presented. Also, a useful
theorem is proved which enables us to extend any -nite measure on a base set E to
a measure
on the same E, such that all members of a given family of pairwise disjoint
subsets of E become
-measurable (see [1] and [13]). This theorem is then repeatedly

applied in further sections of the book.
In Chapters 3 and 4 we primarily deal with those measures on E which are invariant or
quasi-invariant with respect to a certain group of transformations of E. It is widely known
that invariant and quasi-invariant measures play a central role in the theory of topological
groups, functional analysis, and the theory of dynamical systems. We discuss some gen-
eral properties of invariant and quasi-invariant measures that are helpful in various elds
of mathematics. First of all, we mean the existence and uniqueness properties of such
measures. The problem of the existence and uniqueness of an invariant measure naturally
arises for a locally compact topological group endowed with the group of all its left (right)
translations. In this way, we come to the well-known Haar measure. The theory of Haar
measure is thoroughly covered in many text-books and monographs (see, for instance, [80],
[83], [182], [202]), so we leave aside the main aspects of this theory. But we present the
classical Bogoliubov-Krylov theorem on the existence of a dynamical system for a one-
parameter group of homeomorphisms of a compact metric space E. More precisely, we
formulate and prove a signicant generalization of the Bogoliubov-Krylov statement: the
so-called xed-point theoremof Markov and Kakutani ([93], [168]) for a solvable group of
afne continuous transformations of a nonempty compact convex set in a Hausdorff topo-
logical vector space. In the same chapters, we distinguish the following two situations: the
case when a given topological space E is locally compact and the case when E is not locally
compact. The latter case involves the class of all innite-dimensional Hausdorff topolog-
Preface vii
ical vector spaces for which the problem of the existence of a nonzero -nite invariant
(respectively, quasi-invariant) Borel measure needs a specic formulation. Some results in
this direction are presented with necessary comments.
Chapter 5 is concerned with measurability properties of real-valued functions dened on
an abstract space E, when a certain class M of measures on E is determined. We introduce
three notions for a given function f acting from E into the real line R. Namely, f may be
(a) absolutely nonmeasurable with respect to M,
(b) relatively measurable with respect to M,
(c) absolutely (or universally) measurable with respect to M.
We examine these notions and show their close connections with some classical construc-
tions in measure theory. It should be pointed out that the standard concept of measurability
of f with respect to a xed measure on E is a particular case of the notions (b) and (c).
In this case, the role of M is played by the one-element class {}.
In Chapter 6 we discuss, again from the measure-theoretical point of view, some properties
of the so-called step-functions. Since step-functions are rather simple representatives of
the class of all functions (namely, the range of a step-function is at most countable), it is
reasonable to consider them in connection with the measure extension problem. It turns out
that the behavior of such functions is essentially different in the case of ordinary measures
and in the case of invariant (quasi-invariant) measures.
In Chapter 7, we introduce and investigate the class of almost measurable real-valued func-
tions on R. This class properly contains the class of all Lebesgue measurable functions
on R and has certain interesting features. A characterization of almost measurable func-
tions is given and it is shown that any almost measurable function becomes measurable
with respect to a suitable extension of the standard Lebesgue measure on R.
Chapter 8 focuses on several important facts from general topology. In particular, Kura-
towskis theorem (see, for instance, [58], [101], [149]) on closed projections is presented
with some of its applications among which we especially examine the existence of a co-
meager set of continuous nowhere differentiable functions in the classical Banach space
C[0, 1]. Also, we prove a deep theorem on the existence of Borel selectors for certain
partitions of a Polish topological space, which is essentially used in the sequel.
In Chapter 9 the concept of the weak transitivity of an invariant measure is considered
and its inuence on the existence of nonmeasurable sets is underlined. Here it is vividly
shown that some old ideas of Minkowski [173] which were successfully applied by him in
viii Topics in Measure Theory and Real Analysis
convex geometry and geometric number theory, are also helpful in constructions of various
paradoxical (e.g., nonmeasurable) sets. Actually, Minkowski had at hand all the needed
tools to prove the existence of those subsets of the Euclidean space R
n
(n 1), which are
nonmeasurable with respect to the classical Lebesgue measure
n
on R
n
.
Chapter 10 covers bad subgroups of an uncountable solvable group (G, ). The term bad,
of course, means the nonmeasurability of a subgroup with respect to a given nonzero -
nite invariant (quasi-invariant) measure on G. We establish the existence of such sub-
groups of G and, moreover, show that some of them can be applied to obtain invariant
(quasi-invariant) extensions of . So, despite their bad structural properties, certain non-
measurable subgroups of G have a positive side from the view-point of the general measure
extension problem.
The next two chapters (i.e., Chapters 11-12) are devoted to the structure of algebraic sums
of small (in a certain sense) subsets of a given uncountable commutative group (G, +).
Recall that the rst deep result in this direction was obtained by Sierpi nski in his classical
work [219] where he stated that there are two Lebesgue measure zero subsets of the real line
R, whose algebraic (i.e., Minkowskis) sum is not Lebesgue measurable. Let us stress that
in [219] the technique of Hamel bases was heavily exploited and in the sequel such an ap-
proach became a powerful research tool for further investigations. We develop Sierpi nskis
above-mentioned result and generalize it in two directions. Namely, we consider the purely
algebraic aspect of the problem and its topological aspect as well. The difference between
these two aspects is primarily caused by two distinct concepts of smallness of subsets
of R.
In Chapters 13 and 14 we turn our attention to Sierpi nski-Zygmund functions [225] and
study them from the point of view of the measure extension problem. It is well known
that the restriction of a Sierpi nski-Zygmund function to any subset of R of cardinality
continuumis discontinuous. This circumstance directly implies that a Sierpi nski-Zygmund
function is nonmeasurable in the Lebesgue sense and, moreover, is nonmeasurable with
respect to the completion of an arbitrary nonzero -nite diffused Borel measure on R. In
addition, no Sierpi nski-Zygmund function has the Baire property.
We give two new constructions of Sierpi nski-Zygmund functions.
(1) The construction of a Sierpi nski-Zygmundfunction which is absolutely nonmeasurable
with respect to the class of all nonzero -nite diffused measures on R. (Notice that
this result needs some extra set-theoretical axioms.)
Preface ix
(2) The construction of a Sierpi nski-Zygmundfunction which is relatively measurable with
respect to the class of all translation-invariant extensions of the Lebesgue measure
on R. (This result does not need any additional set-theoretical hypotheses.)
Chapters 15-17 are similar to each other in the sense that the main topics discussed therein
are connected with different constructions of nonseparable extensions of -nite measures.
Among the results presented in these chapters, let us especially mention:
(i) the construction (assuming the Continuum Hypothesis) of a nonseparable extension
of the Lebesgue measure without producing new null-sets;
(ii) the construction (also under some additional set-theoretical axioms) of nonseparable
invariant extensions of -nite invariant measures by using their nontrivial ergodic
components;
(iii) the construction (assuming again the Continuum Hypothesis) of a nonseparable non-
atomic left invariant -nite measure on any uncountable solvable group.
In Chapter 18, we consider universally measurable additive functionals. The universal
measurability is treated here in a generalized sense, namely, a real-valued functional f on
a Hilbert space E is universally measurable if and only if for any -nite Borel measure
given on E, there exists an extension
of such that f becomes measurable with

respect to
. It is established that there are universally measurable additive functionals

which are everywhere discontinuous. This result may be regarded as a counter-version to
the well-known statement (see, e.g., [97], [153], [154]), according to which any universally
measurable (in the usual sense) additive functional on E is necessarily continuous.
Chapter 19 is devoted to certain strange subsets of the Euclidean plane R
2
. We discuss
various properties of these paradoxical sets from the measure-theoretical view-point. In
particular, a subset Z of R
2
is constructed which is almost invariant under the group of all
translations of R
2
, is
2
-thick (where
2
stands for the two-dimensional Lebesgue measure
on R
2
) and, in addition, has the property that for each straight line l in R
2
, the intersection
l Z is of cardinality strictly less than the cardinality of the continuum. By using this set
Z, we dene translation-invariant extensions of
2
for which no analogue of the classical
Fubini theorem can be valid.
The nal chapter is connected with certain restrictions of functions acting from R into R.
The rst examples of those restrictions of measurable functions, which are dened on suf-
ciently large subsets of R and have various nice properties, were given in widely known
x Topics in Measure Theory and Real Analysis
statements of real analysis. For instance, in accordance with the classical Luzin theorem
(see, e.g., [16], [26], [65], [80], [161], [183], and [192]), every real-valued Lebesgue mea-
surable function restricted to a certain set of strictly positive -measure becomes continu-
ous (and an analogous purely topological result holds true in terms of the Baire property
and category). We touch upon some other results in this direction. In particular, it is proved
that every Lebesgue measurable function g : R R is monotone on a nonempty perfect
subset of R. At the same time, such a perfect set does not need to be of strictly posi-
tive -measure. The last circumstance is shown by considering Jarniks [88] continuous
nowhere approximately differentiable function whose existence is a rather deep theorem of
real analysis (cf. [33], [34], and [167]).
In general, we tried to present the material in a self-contained form completely accessible
to graduate and post-graduate students. Moreover, for the readers convenience, six Ap-
pendices are attached to the main text of this book, which can be read independently of the
material covered in the chapters.
In Appendix 1 some auxiliary set-theoretical facts and constructions are considered, which
are essential in various sections of the book. Namely, elements of innite combinatorics
(e.g., innite trees and K onigs lemma), several delicate set-theoretical statements, and the
existence of an uncountable universal measure zero subset of R are discussed.
Appendix 2 is devoted to various theorems on the existence of measurable selectors. Re-
sults of this type are important and attractive and have found applications in numerous
branches of modern mathematics. We begin with the Choquet theorem on capacities (see
[24], [52], [187]) and showits close connection with statements about measurable selectors.
Also, we present the following two fundamental results in this topic: the theorem of Kura-
towski and Ryll-Nardzewski [151] and, as one of its consequences, the Luzin-Jankov-von
Neumann theorem (see, e.g., [99]).
In Appendix 3 deep properties of -nite Borel measures on metrizable topological spaces
are examined. It is proved that if the topological weight of a metric space E is not measur-
able in the Ulam sense, then any -nite Borel measure on E admits a separable support
(cf. [192]). This important fact is essentially used in Chapter 12.
Appendix 4 contains a detailed proof of the existence of a continuous function f : R R
which is nowhere approximately differentiable. As mentioned above, the rst example of
such a function was constructed by Jarnik in his remarkable work [88]. A function of this
type is needed for considerations in Chapter 20.
Preface xi
In Appendix 5 general properties of commutative groups (primarily, innite commutative
groups) are examined and one useful theorem, due to Kulikov, on the algebraic structure
of such groups is proved. This theorem is utilized in several parts of the main text of the
book; see especially Chapters 10 and 11.
Appendix 6 presents elements of classical descriptive set theory. Namely, we touch upon
certain properties of Borel and analytic (Suslin) subsets of uncountable Polish spaces and
apply those properties to the question of measurability of sets or functions. (In this con-
text, Appendix 2 is also helpful.) The so-called separation principle, rst introduced and
extensively studied by Luzin and Sierpi nski, receives special attention. Of course, our pre-
sentation of this material is concise and supercial. The standard monographs or text-books
devoted to classical descriptive set theory are [99], [148], [150], [160], and [162]; see also
Martins article in [10].
Finally, we would like to note that all sections of the book, including the Appendices, are
provided with exercises which contain additional information concerning the questions un-
der discussion. Some of the exercises are quite easy but some of them involve difcult
mathematical facts and need intensive efforts for their solution. These more difcult exer-
cises are marked by the symbol

and we recommend that the reader solve them in order to
better understand the subject. Finally, we also state in the text several unsolved problems
which are motivated by (or closely connected with) topics presented in this book.
The Bibliography consists of 251 titles and contains only the most relevant ones. Of course,
it is far from being complete but rather provides a basic orientation to the subject in order
to stimulate further interest of our readers in various questions of measure theory and real
analysis.
A.B. Kharazishvili

Contents
Preface v
1. The problem of extending partial functions . . . . . . . . . . . . . . . . . . 1
2. Some aspects of the measure extension problem . . . . . . . . . . . . . . . 19
3. Invariant measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4. Quasi-invariant measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5. Measurability properties of real-valued functions . . . . . . . . . . . . . . . 79
6. Some properties of step-functions connected with extensions of measures . . 97
7. Almost measurable real-valued functions . . . . . . . . . . . . . . . . . . . 111
8. Several facts from general topology . . . . . . . . . . . . . . . . . . . . . . 125
9. Weakly metrically transitive measures and nonmeasurable sets . . . . . . . 145
10. Nonmeasurable subgroups of uncountable solvable groups . . . . . . . . . 159
11. Algebraic sums of measure zero sets . . . . . . . . . . . . . . . . . . . . . 177
12. The absolute nonmeasurability of Minkowskis sum of certain universal
measure zero sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
13. Absolutely nonmeasurable additive Sierpi nski-Zygmund functions . . . . . 215
14. Relatively measurable Sierpi nski-Zygmund functions . . . . . . . . . . . . 227
15. A nonseparable extension of the Lebesgue measure without new null-sets . 241
16. Metrical transitivity and nonseparable extensions of invariant measures . . 257
17. Nonseparable left invariant measures on uncountable solvable groups . . . 269
18. Universally measurable additive functionals . . . . . . . . . . . . . . . . . 281
19. Some subsets of the Euclidean plane . . . . . . . . . . . . . . . . . . . . . 297
20. Restrictions of real-valued functions . . . . . . . . . . . . . . . . . . . . . 313
Appendix 1. Some set-theoretical facts and constructions . . . . . . . . . . . . 339
Appendix 2. The Choquet theorem and measurable selectors . . . . . . . . . . 359
xiii
xiv Topics in Measure Theory and Real Analysis
Appendix 3. Borel measures on metric spaces . . . . . . . . . . . . . . . . . . 383
Appendix 4. Continuous nowhere approximately differentiable functions . . . 397
Appendix 5. Some facts from the theory of commutative groups . . . . . . . . 411
Appendix 6. Elements of descriptive set theory . . . . . . . . . . . . . . . . . 423
Bibliography 447
Subject Index 457
Chapter 1
The problem of extending partial functions
There are several general concepts and ideas in contemporary mathematics which play a
fundamental role in almost all of its branches. Among ideas of this kind, the concept of
extending a given partial function is of undoubted interest and of paramount importance for
various domains of mathematics. For instance, every working mathematician knows that
the problemof extending partial functions is considered and intensively studied in universal
algebra, general and algebraic topology, mathematical and functional analysis, as well as
other elds. Of course, this problem has specic features in any of the above-mentioned
disciplines and it frequently needs special approaches or appropriate research tools which
are suitable only for a given situation and are applicable to concrete mathematical objects,
for example, groups, topological spaces, ordered sets, differentiable manifolds, and other
structures.
However, this problem can also be examined from the abstract view-point and method-
ological conclusions of a general character can be made. Below, we touch upon different
aspects of the problem and illustrate them by relevant examples. Some of those examples
will be envisaged more thoroughly in subsequent sections of this book. The main goal
of our preliminary consideration is to demonstrate how the problem of extending partial
functions accumulates ideas from different areas of modern mathematics.
The best known example of this type is the famous Tietze-Urysohn theorem which states
that every real-valued continuous function dened on a closed subset of a normal topo-
logical space (E, T ) can be extended to a real-valued continuous function dened on the
whole space E (see, for instance, [58], [101], and [148]).
Another example of this kind is the classical Hahn-Banach theorem which states that a
continuous linear functional dened on a vector subspace of a given normed vector space
(E, [[ [[) can be extended to a continuous linear functional having the same norm and
dened on the whole E (see any text-book of functional analysis, for instance, [56], [57],
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2, 1
DOI 10.1007/978-94-91216-36-7_1, 2009 Atlantis Press/World Scientific
2 Topics in Measure Theory and Real Analysis
or [209]).
The third example of this sort, although far from the main topics of topology and analysis,
is the problem of extending a given partially recursive function to a recursive function.
As known, the latter should be dened on the set N (= ) of all natural numbers. The
existence of a partially recursive function that does not admit an extension to a recursive
function is crucial for basic statements of mathematical logic and the theory of algorithms.
It sufces to mention G odels incompleteness theorem of the formal arithmetic (see, for
instance, [10] and [215]).
Obviously, many other interesting and important examples can be pointed out in this con-
text.
The present book contains selected topics of measure theory which is a necessary part of
modern mathematical and functional analysis. As is well known, ordinary measures are
real-valued functions dened on certain classes of subsets of a given base set E and having
the countable additivity property. Of course, in contemporary mathematics the so-called
vector-valued measures and operator-valued measures are also extremely important and are
used in many questions of analysis and the theory of stochastic processes, but we do not
touch them in our further considerations. Here we would like to stress especially that topics
presented in this book are primarily concentrated around the measure extension problem
which plays a signicant role in numerous questions of real analysis, probability theory,
and set-theoretical topology. Actually, the measure extension problem will be central for
us in most sections of the book.
Consequently, it is reasonable to begin our preliminary discussion by considering several
facts from mathematical analysis, which are closely connected to extensions of partial
real-valued functions. Some of the facts listed below are fairly standard and accessible
to average-level students. But among the presented facts the reader will also encounter
those which are more important and deeper and which nd applications in various domains
of modern mathematics.
Let R denote the real line and let X be an arbitrary subset of R.
A function f : X R is called a partial function acting from R into R.
For this f , we may write f : R R saying that f is a partial function whose domain is
contained in R. As usual, we denote
dom( f ) = X.
If X = R, then we obviously have an ordinary function f : R R.
The problem of extending partial functions 3
The symbol ran( f ) denotes the range of a partial function f , i.e.,
ran( f ) = f (x) : x dom( f ).
If Y is any subset of R, then the symbol f [Y stands for the restriction of a partial function
f to Y.
As a rule, people working in classical mathematical analysis are often interested in the
following general question.
Does there exist an extension f
: R R of a partial function f : R R such that f
is
dened on the whole R and has certain nice properties?
In particular, we may require that f
should be differentiable or continuous or semicontin-

uous or monotone or convex or Borel measurable or Lebesgue measurable or should have
the Baire property. (Notice that the Baire property can be regarded as a topological version
of measurability; extensive material about this property is contained in remarkable books
[148], [176], and [192].)
An analogous question arises in a more general situation, e.g., for partial functions f acting
from subsets of an abstract set E into R, where E is assumed to be endowed with some
additional structure. In such a case an extension
f
: E R
with dom( f
) = E must preserve a given structure on E or should be compatible, in an

appropriate sense, with this structure. It is clear that questions of the above-mentioned kind
quite frequently arise in mathematical analysis, general topology, and abstract algebra.
Therefore, this topic is of interest for large groups of the working mathematicians.
Below, we have made a small list of results in this direction, have commented on each of
them or have given a necessary explanation, and have referred the reader to other related
works in which extensions of partial functions are considered more thoroughly (see, [58],
[70], [83], [101], and [148]).
For the convenience of potential readers, the material below is presented in the form of
examples of statements about extensions of real-valued partial functions. By the way, we
think that in various lecture courses for students it is useful to provide them with additional
information concerning extensions of partial functions. Such an approach essentially helps
them to see more vividly deep connections and interactions between distinct elds of con-
temporary mathematics. In addition, the students should know that the general problem of
extending partial functions is important for all mathematics because this problem almost
permanently occurs in different mathematical branches and nds numerous applications.
We start with a fundamental result of purely set-theoretical avor.
Example 1. Let X
i
: i I be an arbitrary family of nonempty sets. A partial selector of
(for) X
i
: i I is any family x
i
: i J where J is a subset of I and x
i
X
i
for all indices
i J. If J = I, then x
i
: i J is said to be a selector of (for) the given family X
i
: i I.
The natural question arises whether every partial selector of X
i
: i I can be extended to
a selector of the same family. As is well known, this question is solved positively if and
only if the Axiom of Choice (denoted by AC) is assumed.
At the same time, even in a very special case
card(I) = 2
c
, (i I)(card(X
i
) = 2),
where c denotes the cardinality of the continuum, the question of the existence of a selector
of X
i
: i I is highly nontrivial. For instance, as shown by Sierpi nski, in this case the
positive answer to the question necessarily implies the existence of a subset of R which is
not measurable in the Lebesgue sense (see [223] and references therein; cf. also [90]).
The formal Zermelo-Fraenkel set theory (denoted usually by ZF) is the standard system
of set-theoretical statements (axioms) without the Axiom of Choice (see, e.g., [10], [91],
[145], [150], and [215]). The symbol ZFC stands for the theory ZF & AC. There are
many set-theoretical assertions equivalent (within ZF) to AC, for instance, the Zorn lemma,
Zermelos theorem on the existence of a well-ordering of any set, the equality a
2
= a for
all innite cardinal numbers a, and so on. (In this connection, see especially [90], [150],
and [223].)
It is remarkable that some nontrivial and interesting equivalents of AC can be formulated
in concrete mathematical disciplines. For example, in general topology we have the fun-
damental Tychonoff theorem stating that the product of any family of quasicompact spaces
is a quasicompact space, too. As demonstrated by Kelley [100], this theorem is equivalent
to AC within ZF theory. In linear algebra we have a very important theorem stating that
every vector space (over an arbitrary eld) possesses at least one basis. As shown by Blass
[14], this theorem is also equivalent to AC within ZF theory.
Additionally, the importance of AC in classical mathematical analysis is well known (see,
e.g., the old extensive work by Sierpi nski [217]). It sufces to remind that even a proof of
the equivalence of the two standard denitions, due to Cauchy and Heine respectively, of
the continuity at a point x dom( f ) of a partial function f : RR needs some weak form
of AC.
The following simple result can be included in a beginner lecture course of real analysis.
Example 2. Let a partial function f : R R be given. Then it admits a continuous exten-
sion f
: RR dened on R if and only if for each open interval ]a, b[ R, the restriction
of f to the set dom( f ) ]a, b[ is uniformly continuous. More generally, let ]a
i
, b
i
[ : i I
be a family of open intervals in R such that
R =]a
i
, b
i
[ : i I.
We can assert that a partial function f : R R admits a continuous extension f
with
dom( f
) = R if and only if the restrictions of this f to all sets

dom( f ) ]a
i
, b
i
[ (i I)
are uniformly continuous.
The proof of this fact is quite easy. It sufces to take into account that any continuous
real-valued function dened on a closed bounded subinterval of R is uniformly continuous.
Also, it is not hard to give an example of a partial function f : RR for which there exists
a countable family [a
i
, b
i
] : i I of segments such that
R =[a
i
, b
i
] : i I
and all restrictions f [(dom( f ) [a
i
, b
i
]) are uniformly continuous, but f does not admit a
continuous extension f
with dom( f
) = R.
The next example is far from trivial, however.
Example 3. Let X and Y be two metric spaces and let Y be complete. Suppose also that
f : X Y is a continuous partial mapping. Then there exists a continuous partial mapping
f
: X Y extending f and dened on some G
-subset of X; consequently, f
can be
extended to a Borel mapping acting from the whole X into Y.
This result is due to Lavrentiev and has numerous applications in descriptive set theory and
general topology (see [58], [148], [157], and Exercise 13 for Chapter 8 of this book).
Example 4. Recall that a partial function f : R R is upper (respectively, lower) semi-
continuous if for each t R, the set
x dom( f ) : f (x) <t
(respectively, the set x dom( f ) : f (x) >t) is open in dom( f ). For f with dom( f ) = R,
this denition is equivalent to the following: f is upper (respectively, lower) semicontinu-
ous if and only if
limsup
yx
f (y) = f (x)
(respectively, liminf
yx
f (y) = f (x)) for all x R (see, e.g., [58], [101], [148], and [183]).
It is interesting to notice that every bounded upper (lower) semicontinuous partial function
admits a bounded upper (lower) semicontinuous extension dened on R. Let us formulate
a more precise result in this direction.
First, recall that a partial function g : R R is locally bounded from above (from below)
if for each point x R, there exists a neighborhoodU(x) such that g[U(x) is bounded from
above (from below).
A partial function g : R R is locally bounded if it is locally bounded from above and
from below simultaneously.
We also recall that an upper semicontinuous function can take its values from the set R
and a lower semicontinuous function can take its values from the set R+.
(These assumptions are convenient in numerous topics of mathematical analysis.)
Now, let f : R R, + be any partial function. The following two assertions are
equivalent:
(a) f admits an upper (lower) semicontinuous extension f
whose domain coincides with

the closure of dom( f );
(b) f is upper (lower) semicontinuous and locally bounded from above (from below).
The equivalence of (a) and (b) implies the validity of the next two statements.
(i) Let f : R[a, b] be a partial upper semicontinuous function. Then there exists an upper
semicontinuous function f
: R [a, b] extending f .
(ii) Let f : R [a, b] be a partial lower semicontinuous function. Then there exists a lower
semicontinuous function f
: R [a, b] extending f .
It should be mentioned that the same results hold true in a more general situation, namely,
for partial semicontinuous bounded functions acting from a normal topological space E
into the real line R. Of course, in this generalized case, the Tietze-Urysohn theorem should
be applied to E in order to obtain the required result.
The next example deals with monotone extensions of partial functions acting from R into
R.
Example 5. Let f : R R be a partial function increasing on its domain. It is easy to
show that f can always be extended to an increasing function f
dened on some maximal

(with respect to inclusion) subinterval of R. Let us denote the above-mentioned maximal
subinterval by T and let
a = inf(T), b = sup(T).
If a = and b = +, then f
is the required increasing extension of f dened on the

whole R. If a ,=, then in view of the maximality of T, we must have
inf f (t) : t T =
and, therefore, f cannot be extended to an increasing function acting from R into R. Anal-
ogously, if b ,= +, then in view of the maximality of T, we must have
sup f (t) : t T = +
and therefore f cannot be extended to an increasing function acting from R into R. We see
that in both of these cases our partial function f is not locally bounded.
A similar result holds true for any decreasing partial function f : R R. We thus obtain
a necessary and sufcient condition for extending a given partial function to a monotone
function acting from R into R. Namely, the following two assertions are equivalent:
(a) f is extendable to a monotone function f
with dom( f
) = R;
(b) f is monotone and locally bounded.
Of course, there is no problem connected with extending monotone partial functions if we
admit innite values of functions under consideration (cf. Example 4). Indeed, in such a
case any monotone partial function
f : R R, +
can be extended to a monotone function f
with dom( f
) = R.
Example 6. Let f : RR be a partial function. Suppose that f is Borel on its domain, i.e.,
for every Borel set B R, the pre-image f
1
(B) is a Borel subset of dom( f ) where dom( f )
is assumed to be endowed with the induced topology. It can be proved that f always admits
a Borel extension f
: R R with dom( f
) = R. However, the proof of this fact is far

from being easy. It needs a certain classication of all Borel partial functions acting from
R into R. This classication is due to Baire (see [3], [4], and [5]). According to it, any
Borel partial function f has its own Baire order = ( f ), where is some countable
ordinal number. For instance, the equality ( f ) = 0 simply means that f is continuous on
its domain. Taking into account the above-mentioned classication and Lavrentievs result
mentioned earlier (see Example 3), the existence of f
can be established by using the

method of transnite induction on (for more details, see [148] or Exercise 16 for Chapter
8).
Example 7. Let f : R R be a partial function. The following two assertions are equiva-
lent:
(a) f admits an extension f
dened on R and measurable in the Lebesgue sense;

(b) there exists a Lebesgue measure zero set A R such that the restriction of f to the set
dom( f ) A is a Borel function on dom( f ) A.
This fact can be proved by using the well-known Luzin criterion for the Lebesgue measur-
ability of real-valued functions (see, e.g., [161], [183], and [192]).
A parallel fact holds true for partial functions having the Baire property. Similarly to the
measurability in the Lebesgue sense, for a partial function f : R R, the following two
assertions are equivalent:
(c) f admits an extension f
dened on R and having the Baire property;

(d) there exists a rst category set B R such that the restriction of f to the set dom( f ) B
is a Borel function on dom( f ) B.
In connection with the latter fact, let us remark that if X is a second category subset of
R, then there always exists a function f : X R which cannot be extended to a function
f
: R R having the Baire property. This deep result is due to Novikov (see [188]). It
is essentially based on the Axiom of Choice and some special facts from descriptive set
theory (e.g., the separation principle for analytic sets). A detailed discussion of this result
is also given in Chapter 14 of [122].
Example 8. If g : R R is a Lebesgue measurable function (respectively, function hav-
ing the Baire property), then there exists a nonempty perfect set P R such that g[P is
monotone on P (see Chapter 20 of this book).
Sierpi nski and Zygmund proved in their joint work [225] that there exists a function
h : R R
satisfying the following condition: for any set X R of cardinality continuum, the restric-
tion h[X is not continuous.
In particular, this condition readily implies that for any set X R of cardinality continuum,
the restriction h[X is not monotone on X. Indeed, it sufces to use the fact that the set of
all discontinuity points of any monotone partial function acting from R into R is at most
countable.
Assuming the Continuum Hypothesis or, more generally, Martins Axiom (see [10], [40],
[67], [91], and [145]), we have the following two statements.
(a) If X R is of second category, then h[X cannot be extended to a function on R having
the Baire property (cf. Novikovs result mentioned above).
(b) If X R is of strictly positive outer Lebesgue measure, then h[X cannot be extended to
a function on R measurable in the Lebesgue sense.
Notice that the validity of (a) and (b) does not need the full power of Martins Axiom.
Actually, it sufces to assume that any subset of R of cardinality strictly less than c is of
rst category (respectively, of Lebesgue measure zero) in R.
Statement (a) directly implies that no Sierpi nski-Zygmund function has the Baire property.
Statement (b) directly implies that no Sierpi nski-Zygmund function is measurable in the
Lebesgue sense. Moreover, we can assert that a Sierpi nski-Zygmund function is nonmea-
surable with respect to the completion of any nonzero -nite diffused (i.e., vanishing at
all singletons) Borel measure on R (see Exercise 2 for Chapter 13).
In addition to the above, no Sierpi nski-Zygmund function is countably continuous. (A
partial function f : RR is called countably continuous if dom( f ) admits a representation
in the form dom( f ) =X
n
: n <, where all restrictions f [X
n
(n <) are continuous.)
Sierpi nski-Zygmund functions have other interesting and important properties. Many
works were devoted to functions of Sierpi nski-Zygmund type (see [7], [42], [123], [124],
[148], [185], and [199]). In the sequel, we will be dealing with Sierpi nski-Zygmund
functions possessing some additional properties which are of interest from the measure-
theoretical point of view and are closely connected with the measure extension problem
(see Chapters 13 and 14).
Example 9. Consider the set R as a vector space over the eld Q of all rational numbers.
Let f : Q Q denote the identity mapping. Since Q is a vector subspace of R (actually,
Q can be treated as a line in the innite-dimensional vector space R) and f is a partial
linear functional, it admits a linear extension
f
: R Q
with dom( f
) = R. Of course, the construction of f
is not effective because it is based on

the Axiom of Choice or on the Zorn lemma (cf. Exercise 5 for this chapter). In fact, the
obtained extension f
is a solution of the Cauchy functional equation, that is we have

f
(x +y) = f
(x) + f
(y)
for all x R and y R. At the same time, taking into account the relation
ran( f
) = Q,
we see that f
is discontinuous at all points of R or, equivalently, f
is a nontrivial solution
of the Cauchy functional equation (cf. [18], [82], [108], [143], and [176]). It is well
known that all nontrivial solutions of the Cauchy functional equation are nonmeasurable
with respect to the Lebesgue measure on R and do not possess the Baire property (see, for
instance, [143]). In our case, it can be observed that the set
X =x R : f
(x) = 0
is not Lebesgue measurable and does not have the Baire property. Notice also that among
nontrivial solutions of the Cauchy functional equation, we can encounter some Sierpi nski-
Zygmund functions (see Chapters 13 and 14). In addition, it should be pointed out that
certain nontrivial solutions of the Cauchy functional equation are successfully applied in
some deep geometrical questions concerning equidecomposability of polyhedra lying in
a nite-dimensional Euclidean space (see [18] where Hilberts third problem and related
topics are discussed in detail).
Example 10. In fact, the preceding example is purely algebraic. Another example of a
similar type is the following. Let (G, +) be a commutative group and let (H, +) be a
divisible commutative group, i.e., any equation of the form
nx = h (n N 0, h H)
has a solution in H. Suppose that a partial homomorphism : G H is given, which
means that is a homomorphism from some subgroup of G into H. Then we can assert
that there always exists a homomorphism
: G H extending (see [70], [83], [152],

or Exercise 2 in Appendix 5).
The above-mentioned result is extremely useful in all theory of commutative groups. For
instance, by applying it, we may readily prove that every commutative group can be isomor-
phically embedded in some divisible commutative group. Since all divisible commutative
groups admit a visual algebraic characterization (cf. [70] or [152]), the importance of this
result becomes quite evident.
An analogous statement concerning extensions of partial continuous homomorphisms can
be formulated in the case of a compact commutative group (G, +) and the one-dimensional
unit torus (S
1
, ) which is a canonical representative of the class of all divisible commutative
compact groups (see [83], [177], and [202]).
Example 11. We may expect that if a partial function f : R R is dened on a small
(in an appropriate sense) subset of R, then f admits extensions with nice properties as
well as extensions with bad properties. Indeed, if dom( f ) is of Lebesgue measure zero
(respectively, is of rst category), then f trivially can be extended to a Lebesgue measurable
function (respectively, to a function possessing the Baire property). Thus, in both of these
cases, we come to extensions with nice properties.
On the other hand, it is possible to present an example of an extension of a partial func-
tion which is dened on a small subset of R, with an extremely bad property from the
measure-theoretical view-point. For this purpose, let us rst introduce two auxiliary no-
tions which play a signicant role in the sequel.
We recall that a set X R is universal measure zero if there exists no nonzero -nite
diffused Borel measure on X.
Evidently, if X R is countable, then X is universal measure zero. The question of the
existence of uncountable universal measure zero subsets of R was posed many years ago
and turned out to be closely connected with some rather delicate set-theoretical construc-
tions due to Hausdorff, Mahlo, Luzin and other authors. For instance, every Luzin subset
of R is uncountable and universal measure zero (various interesting properties of Luzin
sets are considered in [143], [148], [159], [176], and [192]; see also Chapters 5, 12 and
Appendix 1).
We shall say that a function g : RRis if there exists no nonzero -nite diffused measure
on R such that g is measurable with respect to . (It is assumed in this denition that the
domain of may be an arbitrary -algebra of subsets of R, containing all singletons.)
We thus see that absolutely nonmeasurable functions (if they exist) are of more pathological
nature than well-known examples of Lebesgue nonmeasurable real-valued functions.
It can be proved that the following two assertions are equivalent:
(a) a function g : R R is absolutely nonmeasurable;
(b) the set ran(g) is universal measure zero and for each t R, the set g
1
(t) is at most
countable.
For the proof of the equivalence (a) (b), see Chapter 5.
Starting with this result, it is not hard to show that an absolutely nonmeasurable function
acting from R into R exists if and only if there exists a universal measure zero subset of R
of cardinality continuum. Also, we can infer the validity of the next statement.
Under the Continuum Hypothesis (or, more generally, under Martins Axiom), for a partial
function f : R R, the following two assertions are equivalent:
(c) f admits an absolutely nonmeasurable extension f
with dom( f
) = R;
(d) the set ran( f ) is universal measure zero and the sets f
1
(t) are countable for all t R.
Fromthe equivalence of (c) and (d) we also obtain, under the same additional set-theoretical
assumptions, that any partial function f : RR dened on a countable subset of R admits
an extension to an absolutely nonmeasurable function dened on the whole R.
Let us remark that according to Blumbergs fundamental theorem [15], any function acting
from R into R can be regarded as an extension of some continuous partial function whose
domain is a countable everywhere dense subset of R (for the proof, see [15] or Exercises
21 and 22 of Chapter 8).
The existence of a Sierpi nski-Zygmund function h : R R shows that under the Contin-
uum Hypothesis, there is no uncountable set X R such that h[X is continuous and, con-
sequently, h cannot be considered as an extension of a continuous partial function dened
on an uncountable subset of R.
On the other hand, according to a recent result of Roslanowski and Shelah [208], it is
consistent with ZFC that any function f : R R may be regarded as an extension of a
continuous function dened on a Lebesgue nonmeasurable subset of R which, obviously,
is necessarily uncountable.
Let us also notice that under Martins Axiom, there exist additive absolutely nonmeasurable
functions which simultaneously are Sierpi nski-Zygmund functions. One construction of
such functions will be given later in this book (see Chapter 13). It is based on the fact that
there exists a generalized Luzin subset of R which simultaneously is a vector space over
the eld Q of all rational numbers.
The next example is concerned with extensions of real-valued partial functions of two vari-
ables.
Example 12. Let (=
1
) denote the Lebesgue measure on the real line R (= R
1
). Con-
sider a function of two real variables
: R[0, 1] R.
Recall that satises the Carath eodory conditions if the following two relations hold:
(i) for each x R, the function (x, ) : [0, 1] R is continuous;
(ii) for each y [0, 1], the function (, y) : R R is -measurable.
Functions of this type play a prominent role in mathematical analysis, the theory of ordinary
differential equations, optimization theory, and probability theory.
It is well known that if satises the Carath eodory conditions, then is measurable
with respect to the product -algebra dom() B([0, 1]), where B([0, 1]) denotes the
-algebra of all Borel subsets of [0, 1].
Now, take a partial function of the form
F : R[0, 1] R,
i.e., suppose that dom(F) R[0, 1]. In addition, suppose that F is measurable with
respect to the product -algebra dom() B([0, 1]). Then the following two assertions
are equivalent:
(a) F admits an extension F
: R[0, 1] R satisfying the Carath eodory conditions and

such that dom(F
) = R[0, 1];
(b) for each x R, the partial function F(x, ) is uniformly continuous on its domain.
This result is rather deep. (The reader may compare it with the simplest Example 2.)
Indeed, to establish the equivalence of (a) and (b), we have to apply the Choquet theoremon
capacities and the theoremon measurable selectors due to Kuratowski and Ryll-Nardzewski
(see Appendix 2).
It should be mentioned that the same result remains true if we replace the Lebesgue measure
space (R, dom(), ) by an arbitrary -nite complete measure space (, dom(), ),
simultaneously replace the segment [0, 1] by a compact metric space Y and consider partial
functions of the form
F : Y R.
For more details, see Exercise 13 of Appendix 2. Notice also that the standard Wiener
process
W : R
[0,1]
[0, 1] R,
which is regarded as a reasonable mathematical model of the Brownian motion, yields a
good example of a function of two variables, satisfying the Carath eodory conditions. Here
R
[0,1]
stands, as usual, for the space of all real-valued functions dened on the segment
[0, 1], and this space is assumed to be equipped with the completion of the Wiener proba-
bility measure
w
. More precisely, we are dealing with a certain set R
[0,1]
such that:
(c)
w
() = 1;
(d) for any t [0, 1], the partial function W(, t) is measurable with respect to
w
;
(e) for any , the trajectory W(, ) is continuous on [0, 1].
Extensive information about stochastic processes and corresponding probability measures
may be found in [16], [26], [52], [187], [205], and [226].
Now, we would like to consider a more difcult special case of the problem of extending
partial functions. In fact, our last example will be concerned with extensions of real-valued
partial set-functions.
We recall that a set-function is any function whose domain is some family of sets. Equiva-
lently, we may say that a set-function f is a function whose domain is a subset of the power
set P(E) of some base set E. Thus, f can be treated as a partial function acting from the
set E
/
=P(E). Of course, partial functions of the form
f : P(E) R
are of prime interest in this book because they include measures and measure type func-
tionals on E, for instance, capacities in the Choquet sense (see Appendix 2).
Example 13. The Lebesgue measure is a real-valued function dened on some class of
subsets of R. It is well known that the class dom() is properly contained in P(R), i.e.,
there are -nonmeasurable sets in R. However, the proof of this fact needs uncountable
forms of the Axiom of Choice. Thus, may be regarded as a partial function acting from
the power set of R into R+, i.e., is a partial function of the form
: P(R) R+.
Several constructions of -nonmeasurable subsets of R are presented in the literature which
give us Vitali sets, Bernstein sets, nontrivial ultralters in the set N of all natural num-
bers, and other pathological sets of real numbers. Compare also Example 9, where the
-nonmeasurable set X associated with a certain nontrivial solution of the Cauchy func-
tional equation was pointed out.
It is natural to ask whether there exists a measure on R extending and dened on the
family of all subsets of R. This problem was originally posed by Banach and then was
studied by many famous mathematicians. Under some additional set-theoretical assump-
tions (e.g., the Continuum Hypothesis or Martins Axiom), the answer is negative (see, for
instance, [67], [69], [78], [79], [91], [192], [222], [224], and [238]). But it seems that this
question is undecidable within the standard system of axioms of contemporary set theory,
e.g., within the Zermelo-Fraenkel formal system ZFC.
It can be shown (see [1], [13], and Chapter 2) that for any countable disjoint family X
i
:
i I of subsets of R, there exists a measure
/
on R extending and such that
X
i
: i I dom(
/
).
We thus see that the countable additivity property can be preserved by
/
under the assump-
tion that the points X
i
(i I) onto which we extend are pairwise disjoint. Actually, it
was established in [1] and [13] that the same result remains valid for any -nite measure
given on an abstract set E and for any disjoint family Y
j
: j J of subsets of E. In
particular, we easily obtain from this result that for every nite family Z
1
, Z
2
, ..., Z
n
of
subsets of E, there exists a measure
/
extending and satisfying the relation
Z
1
, Z
2
, ..., Z
n
dom(
/
).
However, if we have an arbitrary innite sequence Z
1
, Z
2
, ..., Z
n
, ... of subsets of E, then
we cannot assert, in general, that is extendable to a measure
/
for which all these subsets
become
/
-measurable. In other words, sometimes there are countably many points in
the power set of E, which all together do not admit further extensions of a given nonzero
-nite measure . For instance, we always have such bad points Z
1
, Z
2
, ..., Z
n
, ... in the
power set of E where E is an arbitrary uncountable universal measure zero subset of R.
It should be noticed that the existence of an uncountable universal measure zero set E R
is well known and can be proved within ZFC theory. In Appendix 1 of the present book
we give a construction of E starting with the classical Sierpi nski partition of the product
set
1
1
, where
1
denotes, as usual, the least uncountable cardinal number.
In this context, the works [197] and [250] should also be mentioned, in which analogous
and stronger results are obtained stating the existence of uncountable universally small
subsets of R.
Finishing this chapter, we hope that the examples just considered and concerning extensions
of partial real-valued functions are sufciently illustrative to showthe reader the importance
of the problem of extending partial functions. We will continue our discussion of this
problemin the following sections. As already said, the measure extension problemtouched
upon in Example 13 will be of special interest in our further considerations.
EXERCISES
1. Give a proof of the statement presented in Example 1. Namely, verify that the following
two assertions are equivalent within ZF theory:
(a) the Axiom of Choice;
(b) if X
i
: i I is an arbitrary family of nonempty sets, then any partial selector of X
i
:
i I can be extended to a selector of this family.
Moreover, by using AC demonstrate that every innite set contains a countably innite
subset. (Note that this fact is deducible with the aid of some weak forms of AC but is not
provable within ZF theory.)
2. Let ]a, b[ be an open subinterval of R and let f : ]a, b[ R be a convex function on
]a, b[. Show that the following two statements are equivalent:
(1) there exists a convex extension f
: R R of f dened on the whole R;

(2) f admits a continuous extension dened on [a, b] and this extension has the right-hand
side derivative at a and the left-hand side derivative at b.
Notice that the above-mentioned derivatives are assumed to be nite.
3. Give a proof of the equivalence (a) (b), where (a) and (b) are assertions fromExample
4.
4. Give a proof of the equivalence (a) (b), where (a) and (b) are assertions fromExample
5.
5
. Show that the set X = x R : f
(x) = 0 described in Example 9 is not Lebesgue

measurable and does not possess the Baire property.
6
. Let (E, d) be a metric space and let f : E R be a partial function satisfying the
Lipschitz condition with a Lipschitz constant L, that is
[ f (x) f (y)[ Ld(x, y) (x dom( f ), y dom( f )).
Prove that there exists an extension f
: E R of f also satisfying the Lipschitz condition

with the same constant L and dened on the whole space E.
7. Let f : R R be a function and let x R. Recall that t R is a Dini right derived
number of f at x if there exists a sequence h
n
: n N of strictly positive real numbers
such that
lim
n+
h
n
= 0, t = lim
n+
( f (x +h
n
) f (x))/h
n
.
In this case, the notation t D
+
f
(x) is frequently used.
Let f : R R be a continuous function. Suppose that for any x R, this f admits a
nonnegative Dini right derived number at x. Show that f is increasing on R.
Starting with this result, prove that the following two assertions are equivalent for a contin-
uous function f acting from R into R:
(a) f satises the Lipschitz condition;
(b) there is a constant L 0 such that for any point x R, the function f has at least one
Dini right derived number whose absolute value does not exceed this L.
8. Let B
n
denote the unit ball in the Euclidean space R
n
and let S
n1
denote the boundary
of B
n
. Show that the following two assertions are equivalent:
(a) the identity embedding of S
n1
into B
n
does not admit a continuous extension dened
on the whole B
n
;
(b) B
n
has the xed-point property which means that for every continuous mapping f :
B
n
B
n
, there exists a point x B
n
such that f (x) =x.
As is well known, assertion (b) is valid and was rst proved by Brower (for the proof,
see [56], [58], and [148]). Consequently, (a) is valid, too. Browers theorem (b) is very
deep, has numerous applications, and was generalized in many directions (see, e.g., [56]
and [57]). One of important generalizations of this theorem is due to Kakutani and states
the existence of xed-points for certain set-valued mappings.
9
. Show that every partial isometry f : R

n
R
n
admits an extension to an isometry f
:
R
n
R
n
. Find necessary and sufcient conditions under which f
is uniquely determined
by f .
On the other hand, for any innite-dimensional real Hilbert space E, give an example of a
partial isometry g : E E which does not admit an extension to an isometry of the whole
E.
10. A partially ordered set (E, ) is called complete if for any set X E, there exists
sup(X) (equivalently, there exists inf(X)).
Let (E, ) be a complete partially ordered set and let : E E be a monotone mapping
(i.e., is either increasing or decreasing). Show that there exists a xed-point of ; in other
words, show that
(y E)((y) = y).
For this purpose, assume without loss of generality that is increasing and put
Y =x E : x (x), y = sup(Y).
Check that (y) = y, i.e., y is the desired xed-point of .
This result is due to Tarski and is known as Tarskis xed-point theorem.
11
. Let a function : R[0, 1] R satisfy the Carath eodory conditions (see Example
12 of this chapter). Prove that is measurable with respect to the product -algebra
dom() B([0, 1]).
12
. Let (E, S, ) be a -nite measure space and let

S
0
=X S : (X) < +.
Consider in S
0
(which is a -ring of subsets of E) the equivalence relation R(X,Y) dened
by the formula
R(X,Y) (XY) = 0,
where the symbol stands for the operation of symmetric difference of two sets, that is
XY = (X Y) (Y X).
Let S
0
/R denote the corresponding quotient set and let, for any X S
0
, the symbol [X]
denote the R-equivalence class containing X. Verify that the function
d([X], [Y]) = (XY) ([X] S
0
/R, [Y] S
0
/R)
is a metric on S
0
/R such that the metric space (S
0
/R, d) is complete.
This (S
0
/R, d) is usually called the metric space canonically associated with the given
measure . The topological weight of this space (see Appendix 3) is a certain inner char-
acteristic of . Denote it by w().
Check that if w() , then w() is equal to the topological weight of the Hilbert space
L
2
() consisting of all real-valued square integrable (with respect to ) functions on E.
A measure is called separable if w() . Otherwise, is called a nonseparable
measure.
Demonstrate that if S is a countably generated -algebra of subsets of E, then is a
separable measure. Also, give an example which shows that the converse assertion is not
valid.
Observe that the completion of a separable measure is separable, too. Infer from this fact
that the Lebesgue measure
n
on the Euclidean space R
n
is separable.
As mentioned in Example 13 of this chapter, the one-dimensional Lebesgue measure
=
1
cannot be extended (within ZFC theory) to a universal measure dened on P(R).
However, there exist nonseparable extensions of which are invariant under the group of
all translations of R (and even under the group of all isometries of R).
Nonseparable -nite measures with some specic properties will be considered in Chap-
ters 15, 16, and 17.
Chapter 2
Some aspects of the measure extension problem
In this chapter, we concisely present several versions of the general measure extension
problem which is a special case of the problem of extending partial real-valued functions,
envisaged in the preceding chapter. Undoubtedly, the measure extension problem is impor-
tant in many questions of analysis and probability theory. Among various aspects of this
problem the following three should be especially distinguished: purely set-theoretical, al-
gebraic, and topological. Below, we will touch upon each of the above-mentioned aspects
and will show their specic features. Since the measure extension problem has its origins
in real analysis, namely, in classical theory of the Lebesgue measure (=
1
) on the real
line R (= R
1
), we also schematically consider several constructions of proper extensions
of and compare such extensions to each other. However, a more detailed consideration
of those constructions will be given in subsequent sections.
Let E be a set, S be an algebra of subsets of E, and let be a nonzero -nite measure
on S. The general measure extension problem is to extend onto a maximally large class
of subsets of E. This problem was originally formulated at the end of the 19th century,
within the theory of real-valued functions. As is well known, it was partially solved by
Lebesgue [158] at the beginning of the 20th century. Namely, starting with the classical
Jordan measure, Lebesgue gave a construction of his measure for the real line R and,
in a similar way, he introduced his measure
n
for the n-dimensional Euclidean space R
n
(see [80], [158], [183], [192], and [210]). Actually, Lebesgues above-mentioned construc-
tion extends the one-dimensional (respectively, multi-dimensional) Riemann integral to the
one-dimensional (respectively, multi-dimensional) Lebesgue integral. The latter integral
allows operation with a sufciently large class of real-valued functions and turns out to be
a necessary research tool in various problems of modern analysis and probability theory.
Then Carath eodorys fundamental theorem followed, which deals with -nite measures
given on abstract measurable spaces of type (E, S). Let us recall the formulation of this
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/978-94-91216-36-7_ , 2009 Atlantis Press/World Scientific
19
2
classical result.
Theorem 1. Any -nite measure dened on an algebra S of subsets of E admits a
unique measure
/
extending and dened on the generated -algebra (S).
Here we omit the standard proof of Carath eodorys theorem(see [16], [56], [65], [80], [83],
[194], and [210]). This theoremshows that, without loss of generality, we can consider only
those -nite measures which are dened on -algebras of sets.
Afterwards, the measure extension problemfound important applications in many other do-
mains of mathematics, such as axiomatic set theory, general topology, functional analysis,
probability theory, and the theory of stochastic processes.
A sufciently general method of extending measures was suggested by Marczewski (Szpil-
rajn) in his classical works [234] and [235]. We would like to describe this method because
it will be used many times in our further considerations.
Namely, let E be a set, be a nonzero -nite complete measure on some -algebra of
subsets of E, and let I be a -ideal of subsets of E such that
(Y I)(
(Y) = 0),
where
usually stands for the inner measure associated with . Denote

S = dom()
and consider the -algebra S
/
of subsets of E, generated by S I, i.e.,
S
/
=(S I).
Obviously, any set Z S
/
can be represented in the form
Z = (X Y
1
) Y
2
,
where X S and both Y
1
and Y
2
are some members of I. Let us put
/
(Z) =
/
((X Y
1
) Y
2
) = (X).
It can easily be checked that the functional
/
is well dened on S
/
in the sense that the
value
/
(Z) does not depend on the above-mentioned representation of Z, and
/
turns out
to be a measure on S
/
extending . Also, it directly follows from the denition of
/
that
(Y I)(
/
(Y) = 0).
Clearly, the measure
/
is a proper extension of if and only if there exists a -
nonmeasurable set belonging to the -ideal I.
Some aspects of the measure extension problem 21
Observe that this construction generalizes the standard procedure of obtaining the comple-
tion of .
If we have a single set Y E with
(Y) = 0, then we can extend a given measure to a

measure
/
in such a way that
/
(Y) = 0. Indeed, it sufces to consider the -ideal I of
subsets of E, generated by the one-element family Y, that is
I =Z : Z Y,
and to apply to this I the Marczewski method described above.
The construction of
/
is, in fact, purely set-theoretical because no specic properties of
(E, S, ) are utilized here. Notice that a purely set-theoretical, or abstract, aspect of the
measure extension problemwas deeply investigated by famous representatives of the Polish
mathematical school (Banach, Kuratowski, Sierpi nski, Ulam, Tarski, and Marczewski).
They obtained a number of fundamental results in this direction which have stimulated
further remarkable investigations in the subject by other mathematicians (see, e.g., [69],
[91], [144], [150], and [231]). The structure of so-called large cardinals was deeply studied
in those investigations.
In particular, according to Ulams classical theorem [238], it is consistent with ZFC theory
that the domain of any extension
/
of a nonzero -nite diffused (i.e., vanishing at all
singletons) measure given on an uncountable set E cannot coincide with the power set
P(E) of E. Consequently, there always exists a set X E such that X ,dom(
/
). Then
/
can be extended to a measure
//
so that X becomes
//
-measurable and, in general, there
are various possibilities to construct such an extension
//
(see Exercise 1 for this chapter).
Thus, by starting with the assumption that there are no large cardinals, we can infer that
there are no maximal extensions of the original nonzero -nite diffused measure . Ex-
tensive information about large cardinals and their combinatorial properties can be found
in [10], [69], [91], [145], and [150] (see also Appendix 1).
For our purposes, we need to look at one important result on extensions of -nite mea-
sures, obtained in the works [1] and [13].
Theorem 2. Let E be a set, be a -nite measure on E and let X
i
: i I be an
arbitrary disjoint family of subsets of E. Then there exists a measure on E extending
and satisfying the relation
X
i
: i I dom().
Proof. Obviously, we may suppose (without loss of generality) that the given family X
i
:
i I is a partition of E.
Let us rst consider the case where the partition X
i
: i I is countable, i.e., we have
card(I) card(N) =.
Let t
i
: i I be an injective family of real numbers and let f : E R be a step-function
such that ran( f [X
i
) = t
i
for any i I. Clearly, it sufces to show that there exists an
extension of for which f becomes measurable.
Without loss of generality, we may assume that either I =1, 2, ..., m or I =N=. Under
this assumption, let us put:
X
/
n
= a -measurable hull of X
n
, where n I;
Y
n
= X
/
n
(X
/
k
: k < n), where n I.
Obviously, the family of sets Y
n
: n I is a disjoint covering of E. Dene a new function
f
0
: E R
by putting f
0
(x) = t
n
if and only if x Y
n
. Since I is countable and all sets Y
n
(n I) are
-measurable, it immediately follows from the denition of f
0
that f
0
is a -measurable
function. Let us show that
(x E : f (x) ,= f
0
(x)) = 0.
Suppose otherwise, that is
(x E : f (x) ,= f
0
(x)) > 0.
Then there exists an index n I such that
(Y
n
x E : f (x) ,= f
0
(x)) > 0.
On the other hand, it is easy to verify the following inclusion:
Y
n
x E : f (x) ,= f
0
(x) X
/
n
X
n
,
which gives a contradiction with the denition of the measurable hull X
/
n
of X
n
. The con-
tradiction obtained establishes the required equality
(x E : f (x) ,= f
0
(x)) = 0.
But, by virtue of the Marczewski method of extending -nite measures, the set
X =x E : f (x) ,= f
0
(x)
can be made measurable with respect to some extension of and, moreover, this proce-
dure can be made so that (X) = 0. Consequently, f becomes measurable with respect to
the same .
Let nowX
i
: i I be an arbitrary partition of E. Using the -niteness of which implies
the validity of the so-called countable chain condition, it is not difcult to show that there
exists a set J I satisfying the following two relations:
(a) card(I J) ;
(b) for any countable set J
0
J, we have
(X
j
: j J
0
) = 0.
Starting with (b) and applying again the Marczewski method, we rst make all the sets
X
j
( j J) to be measurable with respect to some extension
/
of . Notice, by the way,
that according to this method,
/
(X
j
) = 0 for all j J. Finally, we apply our previous
argument to the countable disjoint family X
i
: i I J and extend
/
to a measure
//
such that
X
i
: i I J dom(
//
).
It is clear now that =
//
turns out to be the required extension of which ends the proof
of the theorem.
Remark 1. Theorem 2 shows, in particular, that having any nite family of subsets of E,
we can always extend to a measure
/
, which makes all these subsets to be
/
-measurable
(cf. also Exercise 1). On the other hand, it is well known that an analogous assertion fails to
be true for some countable families of subsets of E, where card(E) =
1
(see Appendix 1).
Moreover, the existence of a Luzin set L R with card(L) = c readily implies that there
is a countably generated -algebra of subsets of R containing all Borel sets in R and not
admitting a nonzero -nite continuous (i.e., diffused) measure.
Recall that the existence of a Luzin set necessarily needs additional set-theoretical axioms.
In this context, it should be underlined that the existence of other small subsets of R having
cardinality
1
, can be established within the theory ZFC (see [69], [79], [113], [148],
[172], [197], and [250]). Some questions of this type will be discussed in more details in
our further considerations (see especially Appendix 1).
Another aspect of the measure extension problem has an algebraic (more precisely, group-
theoretical) avor. Namely, suppose that E is an arbitrary set and denote by Sym(E) the
family of all bijections of E onto itself. Obviously, Sym(E) becomes a group with respect
to the standard composition operation . This group is usually called the symmetric group
of E. Any subgroup G of Sym(E) is called a group of transformations of E.
We shall say that the pair (E, G) is a space equipped with a transformation group if G is a
subgroup of Sym(E).
Spaces equipped with transformation groups can be frequently met in algebra, geometry,
topology, mathematical analysis, the theory of dynamical systems, and in other elds of
mathematics.
Let G be a group of transformations of E and let L be a family of subsets of E.
We shall say that L is invariant under G or, in short, G-invariant if
(g G)(X L)(g(X) L).
We shall say that a measure on E is invariant under G or, in short, G-invariant if dom()
is a G-invariant class of subsets of E and
(g G)(X dom())((g(X)) = (X)).
More generally, we shall say that a measure on E is quasi-invariant under G or, in short,
G-quasi-invariant if dom() is a G-invariant class of subsets of E and
(g G)(X dom())((g(X)) = 0 (X) = 0).
In many situations a base set E is a group with respect to some algebraic operation , i.e., the
pair (E, ) is an abstract group. Then the group G of all left (respectively, right) translations
of E is canonically isomorphic to (E, ) and can be identied with E. So we may write
(G, ) instead of (E, ). In such a case, we deal with left invariant (left quasi-invariant)
measures on E and, analogously, with right invariant (right quasi-invariant) measures on
the same E. Actually, in many cases it sufces to consider only left invariant (left quasi-
invariant) measures on groups.
Obviously, if (E, ) is a commutative group, then the concepts of left invariant (left quasi-
invariant) and right invariant (right quasi-invariant) measures do not differ from each other.
Suppose that an uncountable group (G, ) is given and suppose that G is equipped with
a nonzero -nite left G-invariant (or, more generally, nonzero -nite left G-quasi-
invariant) measure . As stated by Kharazishvili [104] and Erd os and Mauldin [60], the
domain of such a cannot be identical with the family P(G) of all subsets of G. Notice
especially that this statement does not need any additional set-theoretical assumptions (see
proof sketched in Exercises 9, 10 and 11 of Appendix 1). In view of the above-mentioned
statement, it is natural to ask whether there exists a left G-invariant (left G-quasi-invariant)
measure
/
on G properly extending . It is consistent with ZFC theory that the an-
swer to this question is positive (see Exercise 20 for Chapter 3) but it is still unknown
whether the same result can be obtained within ZFC. However, without using additional
set-theoretical hypotheses, for some sufciently wide classes of uncountable groups this
question has a positive answer even in terms of certain subgroups of the original group. In
particular, if (G, ) is an arbitrary uncountable solvable group, then there always exists a
-nonmeasurable subgroup H of G such that can be extended to a left G-invariant (left
G-quasi-invariant) measure
/
for which we have H dom(
/
) and
/
(H) = 0. A more
detailed information about this relatively recent result can be found in Chapter 10.
Marczewskis method of extending -nite measures successfully works in the cases of
invariant and quasi-invariant measures but needs a slight modication (cf. [41], [42], [45],
[108], [115], [195], [242], [247], and [249]).
Namely, let E be a set, G be a group of transformations of E, and let be a -nite com-
plete G-invariant (G-quasi-invariant) measure on some G-invariant -algebra of subsets of
E. Suppose that a G-invariant -ideal I of subsets of E is given such that
(Y I)(
(Y) = 0).
Again, let us denote
S = dom()
and consider the -algebra S
/
of subsets of E, generated by S I. It can easily be seen
that S
/
is also a G-invariant -algebra of sets. We already know that any set Z S
/
can
be represented in the form
Z = (X Y
1
) Y
2
,
where X S and both Y
1
and Y
2
are some members of I. Let us put
/
(Z) =
/
((X Y
1
) Y
2
) = (X).
Again, it is not difcult to check that the functional
/
is well dened in the sense that
the value
/
(Z) does not depend on the above-mentioned representation of Z, and
/
turns
out to be a G-invariant (G-quasi-invariant) measure on S
/
extending . In addition to this
circumstance, the equality
/
(Y) = 0 is valid for all sets Y belonging to I.
Consider now a particular case, when we have a single set Y E with
(Y) = 0 which
satises the relation
(g G)(
(g(Y)Y) = 0),
where
stands for the outer measure associated with (any set Y E for which this
relation holds is usually called almost G-invariant with respect to ). Then we can ex-
tend a given -nite G-invariant (G-quasi-invariant) measure to a G-invariant (G-quasi-
invariant) measure
/
in such a way that
/
(Y) = 0. Indeed, it sufces to consider the
G-invariant -ideal I of subsets of E, generated by the one-element family Y, and to
apply to I the above-mentioned construction (cf. [85], [198], [234], and [235]).
The third aspect of the measure extension problem is purely topological. Let us recall
that most measures considered in mathematical analysis, general topology, and probability
theory are regular in an appropriate sense (see, e.g., [16], [26], [56], [65], [71], [80], and
[83]). For example, if a -nite measure is given on a Hausdorff topological space E
and for each set X dom(), the equality
(X) = sup(K) : K dom(), K X, K is compact
holds true, then is said to be a Radon measure.
Radon measures on locally compact topological spaces play an extremely important role in
various questions of analysis and probability theory. A wider class of measures constitute
the so-called perfect measures which were introduced by Gnedenko and Kolmogorov (see
Exercise 17 for this chapter).
There are several deep statements in topological measure theory concerning those exten-
sions of measures which preserve the regularity property. In order to present one of the
results, we need A.D. Alexandrovs theorem stating the countable additivity of any nite
nonnegative nitely additive Radon type functional given on a Hausdorff space (see, for
instance, [56]).
Theorem 3. Let E be a Hausdorff topological space and let be a nite nonnegative
nitely additive functional on some algebra of subsets of E. Suppose that is regular in
the Radon sense, i.e., for any set Z dom(), we have
(Z) = sup(K) : K dom(), K Z, K is compact.
Then is countably additive or, in other words, is a measure.
Proof. Notice rst that for every set Z dom(), we have
(Z) = inf(U) : U dom(), Z U, U is open.
This fact directly follows from our assumptions on if we apply them to the set E Z and
take into account the circumstance that every compact set in the given Hausdorff space E
is a closed subset of E.
Let now X
n
: n < be a disjoint countable family of sets which all belong to dom()
and suppose that their union
X =X
n
: n <
also belongs to dom(). The nite additivity and nonnegativity of readily imply the
inequality
(X
n
) : n < (X).
Therefore, it sufces to prove the reverse inequality
(X)
(X
n
) : n <.
Fix a strictly positive real number . Obviously, there exists a compact set K X such that
(X) (K) <.
Further, for each n <, we can nd an open set U
n
E containing X
n
and such that
(U
n
) (X
n
) </2
n
.
Since K is compact and the family U
n
: n < is an open covering of K, there exists a
natural number m satisfying the relation
K U
0
U
1
... U
m
.
This relation implies
(X) +(K) +((X
0
) +(X
1
) +... +(X
m
)) +2
3 +
(X
n
) : n <.
Since may be arbitrarily small, we obtain the required reverse inequality and, therefore,
the countable additivity of which completes the proof.
From Theorem 3 one can readily deduce the following statement due to Henry and con-
cerning extensions of nite regular measures.
Theorem 4. Let E be a Hausdorff topological space and let be any -nite Radon
measure dened on a -subalgebra of the Borel -algebra B(E). Then can be extended
to a Radon measure dened on the whole Borel -algebra B(E).
Proof. It sufces to consider the case when is a nite measure. Denote by M the family
of all those nonnegative nitely additive Radon type functionals which extend and are
dened on algebras of subsets of E. A direct verication shows that M is an inductive
partially ordered set with respect to the relation
R(
/
,
//
) :
/
is a restriction o f
//
.
By virtue of the Zorn lemma, there exists a maximal element in the above-mentioned par-
tially ordered set. Denote such a maximal element by . Using Theorem 1 and Theorem 3,
we rst derive that is a measure and dom() is a -algebra of subsets of E. Moreover,
we may assume without loss of generality that is complete. Now, we assert that dom()
contains the whole Borel -algebra B(E). Suppose otherwise, that is
B(E) dom() ,= / 0.
This assumption implies that there exists a closed subset Z of E such that Z , dom()
(because B(E) is generated by the family of all closed subsets of E). In view of Exercise
1, we may reduce our consideration to the case where
(Z) =
(E Z) = 0.
In this case, on the -algebra generated by dom() Z we dene a functional
/
by the
formula
/
((Z X) ((E Z) Y)) =(X),
where X and Y are arbitrary elements of dom(). It can readily be checked that the deni-
tion of
/
is correct and
/
is a nonnegative nitely additive Radon type functional strictly
extending . But this circumstance contradicts the maximality of . The contradiction
obtained nishes the proof.
Remark 2. Alexandrovs and Henrys theorems are important and are successfully applied
in various questions of measure theory. One typical application of Henrys theoremis given
in Exercise 10 of this chapter. The natural and interesting question arises whether any -
nite Borel measure on a complete metric space is Radon. It turns out that this question is
closely connected with some set-theoretical assumptions (for more details, see Appendix
3).
We have mentioned three aspects of the measure extension problem. But, of course, vari-
ous combinations of these aspects can be considered. For instance, in our further construc-
tions, we will be dealing with certain algebraic-topological aspects of the problem (cf. also
[249]). In particular, we will encounter some extensions of the Lebesgue measure on the
real line R, which are nonseparable (a topological property) and, simultaneously, invariant
under the group of all translations of R (an algebraic property).
In this connection, let us briey describe one useful method of extending measures by
applying those mappings whose graphs are thick in the measure-theoretical sense. This
method was originally introduced by Kodaira and Kakutani in their famous construction of
a nonseparable translation-invariant extension of (see [141] and Chapter 16).
Let (E
1
, S
1
,
1
) and (E
2
, S
2
,
2
) be two measure spaces such that
1
is -nite and
2
is
a probability measure. Let
f : E
1
E
2
be a mapping, Gr( f ) denote the graph of f , and let
1
2
be the product measure whose
corresponding factors are
1
and
2
. Suppose that the set Gr( f ) is (
1
2
)-thick in the
product space E
1
E
2
, which means that
(
1
2
)
((E
1
E
2
) Gr( f )) = 0.
For any set Z dom(
1
2
), let us dene
Z
/
=x E
1
: (x, f (x)) Z.
Further, we introduce the class of sets
S
/
1
=Z
/
: Z dom(
1
2
).
A straightforward verication shows that the following auxiliary proposition holds true.
Lemma 1. S
/
1
is a -algebra of subsets of E
1
, which contains S
1
.
Let us dene a functional
/
1
on S
/
1
by the formula
/
1
(Z
/
) = (
1
2
)(Z) (Z dom(
1
2
)).
Taking into account the thickness of the graph Gr( f ) in the product space E
1
E
2
and the
equality
2
(E
2
) = 1, we easily get the next auxiliary proposition.
Lemma 2. The functional
/
1
is well dened, countably additive, and extends the given
measure
1
.
We thus conclude that
/
1
is a measure on S
/
1
extending the original measure
1
. It can
also be readily veried that the following statement is valid.
Lemma 3. The mapping f turns out to be measurable with respect to the -algebras S
2
and S
/
1
. In other words, for any set Y S
2
, we have f
1
(Y) S
/
1
.
As already said, the proofs of Lemmas 1, 2, and 3 are not difcult and the reader may try
to prove these auxiliary propositions himself (herself). We will return to these lemmas in
the sequel and will present some of their variations.
But here we would like to give one immediate application of the above-mentioned lemmas.
For this purpose, let us consider the concrete particular case, where
E
1
= E
2
= R,
and let us put:
1
= the Lebesgue measure on R;
2
= a diffused probability Borel measure on R.
By using a Bernstein type transnite argument (see, for instance, [12], [30], [40], [99],
[148], [150], [176], [192], or Chapter 5), a function f : R R can be constructed, whose
graph is thick with respect to the product measure
1
2
and, consequently, f is nonmea-
surable with respect to (cf. Exercise 8). On the other hand, by virtue of Lemma 3, we
have the following statement.
Theorem 5. f is measurable with respect to some proper extension
/
of .
The theoremjust formulated is a typical representative of those results which are connected
with functions having thick graphs and with corresponding extensions of measures. This
approach will be developed in our further considerations (see especially Chapters 5 and 7).
Actually, Theorem 5 may be regarded as a rst step in much more complicated construc-
tions, so it plays only an illustrative role here.
For a while, let us return to the algebraic aspect of the measure extension problem. In
several sections of this book, we will need the next auxiliary proposition.
Lemma 4. Let E be a set, G be a group of transformations of E, and let be a -nite
G-quasi-invariant measure on E. If X is an arbitrary -measurable set, then there exists a
countable family g
i
: i I G such that the set
Y =g
i
(X) : i I
is almost G-invariant with respect to , i.e., we have the relation
(g G)((g(Y)Y) = 0).
The proof of this lemma is based on the method of transnite induction. A sketch of the
corresponding argument is given in Exercise 13 below. We leave to the reader to complete
all omitted details.
Let E be a set and let G be a group of transformations of E. Consider the class of all
nonzero -nite G-invariant (or, more generally, G-quasi-invariant) measures on E and
pose for them the measure extension problem. In other words, we are interested whether
any nonzero -nite G-invariant (respectively, G-quasi-invariant) measure on E admits a
proper extension belonging to the same class. Of course, in this case the problem has
specic features and its solution essentially depends on structural properties of the pair
(E, G). In our further considerations, we will be dealing with this problem many times.
For its solving within ZFC theory (in certain particular cases), we will need the notion of
absolutely negligible sets. These sets are very small from the view-point of the theory of
invariant (quasi-invariant) measures (see [108], [115], [119], and [198]). Let us give the
precise denition of such sets.
Let M be a class of -nite G-invariant (G-quasi-invariant) measures on E. We say that
a set Z E is absolutely negligible with respect to M if for every measure M, there
exists a measure M satisfying the following two conditions:
(1) is an extension of ;
(2) Z dom() and (Z) = 0.
It immediately follows from this denition that if Z is absolutely negligible with respect to
M and is an arbitrary measure from M, then
Z dom() (Z) = 0.
However, the validity of the above implication for all M is not sufcient to assert that
Z is absolutely negligible with respect to M.
Suppose now that M coincides with the class of all -nite G-invariant (G-quasi-invariant)
measures on E and let Z E. If Z is absolutely negligible with respect to this M, then we
shall say that Z is a G-absolutely negligible subset of E.
Fortunately, it is possible to give a characterization of all G-absolutely negligible subsets
of E only in terms of the pair (E, G). Namely, we have the following statement which turns
out to be rather useful in further considerations (see [108], [109], and [115]).
Theorem 6. Let (E, G) be a space equipped with a transformation group and let Z be a
subset of E. Then these two assertions are equivalent:
1) Z is a G-absolutely negligible set;
2) for any countable family g
n
: n < of transformations from the group G, there exists
a countable family h
m
: m < of transformations from the same group, such that
h
m
(g
n
(Z) : n <) : m < = / 0.
Proof. Let the given set Z satisfy 1) and suppose that 2) is not true. Then there exists a
countable family g
n
: n < of transformations from G such that for the set
Z
/
=g
n
(Z) : n <
and for an arbitrary countable family h
m
: m < of transformations from G, we have
h
m
(Z
/
) : m < ,= / 0.
Let us denote
Z
//
= E Z
/
and consider the G-invariant -ideal I of subsets of E, generated by the one-element
family Z
//
. Obviously, we can dene a complete probability G-invariant measure on E
such that I =I(), where I() stands for the -ideal of all -measure zero subsets of
E. In particular, we have the relations
Z
//
dom(), Z
/
dom(),
(Z
//
) = 0, (Z
/
) = 1.
According to 1), there exists a G-quasi-invariant extension of such that
Z dom(), (Z) = 0.
Consequently, we get
(Z
/
) =(g
n
(Z) : n <) = 0
which contradicts the relation
(Z
/
) = (Z
/
) = 1.
This contradiction establishes the validity of the implication 1) 2).
Suppose now that the set Z satises 2). Let be an arbitrary -nite G-quasi-invariant
measure on E. We shall prove that there exists a G-quasi-invariant measure on E extend-
ing and such that the set Z is measurable with respect to and (Z) = 0. In order to
show this fact, let us denote by J the G-invariant -ideal of subsets of E, generated by
the one-element family Z. Taking 2) into account and applying Lemma 4, we easily infer
that for each set X J, the equality
(X) = 0 holds true. Consequently, according to the

Marczewski method, the measure can be extended to some G-quasi-invariant measure
on E such that J dom() and (X) = 0 for all sets X J. In particular, we obtain
that Z dom() and (Z) = 0.
Finally, let us observe that if the initial measure is G-invariant, then the extended measure
can be chosen to be G-invariant, too.
We will return to absolutely negligible sets in subsequent parts of the book.
EXERCISES
1
. Let E be a set and let be a nite complete measure on E. Suppose that there exists a set
X E nonmeasurable with respect to . Show that there are continuumly many measures
/
extending for which X becomes measurable and the values
/
(X) are distinct for
different
/
.
Argue in the following manner. First, observe that
(X) <
(X). Let Y be a -
measurable kernel of X and let Z be a -measurable hull of X. Consider the sets
U = X Y, V = Z Y.
Check that the following relations hold:
V dom(), U V,
(U) =
(V U) = 0.
Keeping in mind this circumstance, taking V as E, and taking U as X, reduce the argument
to the case where
(X) =
(E X) = 0.
In the latter case, x a real number t [0, 1] and put
t
((AX) (B(E X))) =t (A) +(1 t)(B),
where A and B are arbitrary -measurable subsets of E.
Verify that
t
is a measure on the -algebra (dom() X) and that
t
is an extension
of . In addition to this fact, check that if r [0, 1] and t ,= r, then
t
(X) ,=
r
(X).
2. Let (E, S, ) be a nite measure space and let X E. We say that X has the uniqueness
property with respect to if the following two conditions hold:
(a) there exists a measure on E extending and such that X dom();
(b) for any two measures
1
and
2
on E which extend and satisfy the relation X
dom(
1
) dom(
2
), the equality
1
(X) =
2
(X) is valid.
Denote by the symbol U () the class of all those subsets of E which have the uniqueness
property with respect to . Let
/
be the completion of .
Prove that the equality U () = dom(
/
) is true.
On the other hand, show that this equality does not always hold for -nite measures on
E.
3. Let E be a Hausdorff topological space and let be a nite nonnegative nitely additive
functional on B(E) such that
(U) = sup(K) : K is compact, K U,
(X) = inf(V) : V is open, X V
for each open set U E and for each Borel set X E.
Show that is a Radon measure. In other words, demonstrate that the equality
(X) = sup(K) : K is compact, K X
holds true for every Borel set X E.
4. Give a proof of Lemma 1.
7. Let (E, S) be a measurable space equipped with a -nite measure . Let f : E R be
a -measurable function. In the product space E R consider the product measure
where as usual stands for the Lebesgue measure on R. Show that the graph of f regarded
as a subset of E R is of ( )-measure zero.
Does this result remain true if (R, ) is replaced by (F, ), where F is an arbitrary set
equipped with a -nite measure ?
8
. Let E be a topological space. We recall that a set X E is totally imperfect in E if X

contains no nonempty perfect subset of E.
A set B E is a Bernstein set in E if both B and E B are totally imperfect subsets of E.
By applying the method of transnite induction, showthat there exists a function f : RR
whose graph Gr( f ) is totally imperfect in the Euclidean plane R
2
= RR.
In particular, deduce from this property of Gr( f ) that f is nonmeasurable with respect to
the completion of any nonzero -nite diffused Borel measure on R.
For this purpose, utilize Luzins classical criterion (the so-called C-property) of the mea-
surability of real-valued functions with respect to the completion of .
9
. By using the method of transnite recursion and applying the Fubini theorem, dene a
bijective function f : RR whose graph is thick in R
2
with respect to the two-dimensional
Lebesgue measure
2
=
1
1
.
It should be noticed in this place that the rst transnite construction of a function f : R
R whose graph is thick with respect to
2
was done by Sierpi nski (cf. also [72], [144], and
[192]).
10
. Let E be a topological space. Consider the family C(E, R) of all real-valued continu-
ous functions on E. According to the standard denition, the Baire -algebra of E (denoted
by B
0
(E)) is the smallest (with respect to inclusion) -algebra S of subsets of E such that
all functions fromC(E, R) are measurable in the sense of S, i.e.,
(a R)(b R)(f C(E, R))( f
1
([a, b]) S).
Obviously, the relation B
0
(E) B(E) holds true, where B(E) denotes the Borel -
algebra of E.
Suppose that E is compact and hence Hausdorff. Prove that any nite (more generally,
-nite) measure dened on B
0
(E) is Radon.
For this purpose, rst check that B
0
(E) is generated by the family of all closed subsets of E
of type G
. Then, assuming the niteness of , consider the family L of all those subsets
X of E for which there exist two sets X
/
and X
//
satisfying the following three relations:
(a) X
/
X X
//
;
(b) X
/
is of type F
and X
//
is of type G
;
(c)
(X X
/
) =
(X
//
X) = 0.
Verify that L is a -algebra of sets and B
0
(E) L.
Then reduce the case of a nonzero -nite measure with dom() = B
0
(E) to the case
of a probability measure , replacing by an equivalent probability measure on the same
-algebra B
0
(E).
Applying Henrys theorem, deduce fromthe said above that any -nite measure on B
0
(E)
admits an extension to a Radon measure on B(E).
Formulate and prove an appropriate generalization of these results for the case of a locally
compact topological space E.
11. Let E
i
: i I be an arbitrary family of compact spaces. Show the validity of the
following equality:
B
0
(E
i
) : i I =B
0
(
E
i
: i I).
For this purpose, apply the Stone-Weierstrass theorem on approximation of continuous
real-valued functions (see, e.g., [58] and [101]).
On the other hand, give an example of a family E
i
: i I of compact spaces such that
B(E
i
) : i I ,=B(
E
i
: i I).
12. Recall that a set X R is universal measure zero if for every -nite diffused Borel
measure on R, we have
(X) = 0. Equivalently, X R is universal measure zero if

there exists no nonzero -nite diffused Borel measure on X (cf. Example 11 of Chapter
1).
Check that the family of all universal measure zero subsets of R forms a -ideal in the
Boolean algebra P(R).
Obviously, any countable subset of R is universal measure zero and, as has already been
mentioned in Chapter 1, the existence of an uncountable universal measure zero set in
R is a deep fact of classical descriptive set theory (for more details about this fact, see
Appendix 1).
Let
1
denote, as usual, the least uncountable cardinal number. Prove that the following
two assertions are equivalent:
(a) there exists an uncountable universal measure zero subset of R;
(b) there exists a countably generated -algebra S of subsets of
1
such that all singletons
in
1
belong to S, but S does not admit any nonzero -nite diffused measure.
For establishing the equivalence of (a) and (b), use Marczewskis characteristic function
(cf. the proof of Theorem 5 from Appendix 1).
13
. Let E be a set, G be a group of transformations of E and let be a -nite G-quasi-

invariant measure on E. Take any -measurable set X with (X) > 0. We shall say that a
set Y is a countable G-conguration of X if there exists a countable family g
i
: i I G
for which we have the equality
Y =g
i
(X) : i I.
Using the method of transnite recursion, dene a family Y
: <
1
of countable
congurations of X. Namely, put:
(a) Y
0
= X;
(b) let <
1
and suppose that Y
has already been dened; if Y
is such that there exists

a transformation g G for which (g(Y
)Y
) > 0, then take

Y
+1
=Y
g(Y
) g
1
(Y
);
otherwise, take Y
+1
=Y
;
(c) if <
1
is a limit ordinal, then put
Y
=Y
: <.
Prove that the family Y
: <
1
is stationary or, in other words, there exists
0
<
1
such that Y
=Y
0
for all [
0
,
1
[.
Deduce from this fact the assertion of Lemma 4.
14. Let (E, G) be a nonempty space equipped with a transformation group. Verify that
the family of all G-absolutely negligible subsets of E is a proper G-invariant ideal in the
Boolean algebra P(E) of all subsets of E. Prove also that if this ideal is not a -ideal, then
there exists a nonzero -nite G-quasi-invariant measure on E such that every G-quasi-
invariant extension
/
of is not maximal, i.e., we can nd a proper G-quasi-invariant
extension of
/
.
15
. Let E be a base set and let G

1
and G
2
be two groups of transformations of E. Showthat
the inclusion relation G
1
G
2
does not imply, in general, any inclusion relation between
the classes of all G
1
-absolutely negligible and all G
2
-absolutely negligible subsets of E.
16
. A Hausdorff topological space E is called Radon (see [26]) if every -nite Borel
measure on E is Radon. Prove that:
(a) any countable Hausdorff space is Radon;
(b) a Borel subset of a Radon space is Radon;
(c) any complete separable metric space (i.e., any Polish topological space) is Radon.
Consider the set of ordinal numbers [0,
1
] in its order topology. Check that this topological
space is compact but is not Radon.
17
. Let (E, S, ) be a -nite measure space. We say that is a perfect measure (in the
sense of Gnedenko and Kolmogorov) if for each -measurable function f : E R, there
exists a set Z B(R) such that
Z ran( f ), (E f
1
(Z)) = 0.
Verify that:
(a) every two-valued probability measure is perfect;
(b) every Radon probability measure is perfect.
In order to showthe validity of (b), check that a direct analogue of Luzins C-property holds
true for any real-valued function measurable with respect to a Radon probability measure.
Chapter 3
Invariant measures
We would like to begin this chapter with a few historical remarks. The general concept
of an invariant measure comes from two sources the rst of which is classical Euclidean
geometry where the notion of the volume of a geometric body is paramount, and this notion
presents the most natural example of an invariant functional with respect to the group of all
isometric transformations of Euclidean space. The second source is the theory of dynam-
ical systems which currently is one of the main branches in contemporary mathematics.
Actually, the extensive study of various dynamical systems was primarily dictated by the
expansion of methods of classical mathematical analysis and especially by the development
of the theory of ordinary differential equations. The most important and widely known ob-
jects in this theory are the so-called autonomous systems. They are closely connected with
certain transformation groups of a phase space. A particular case of systems of this type
are Hamiltonian systems for which the fundamental Liouville theorem holds true (see, for
instance, [196]; cf. also Exercises 1 and 2 of this chapter).
Let E be a base set, that is an abstract phase space. Any bijection of the form
g : E E
is usually called a transformation of E. Evidently, the family of all transformations of E is
a group with respect to the composition operation . As was mentioned in Chapter 2, this
group is denoted by the symbol (Sym(E), ) and is called the symmetric group of E.
Any subgroup G of Sym(E) is treated as a certain group of transformations of E. The pair
(E, G) is usually called a space equipped with a transformation group.
Let G be a group of transformations of E and let S be a -algebra of subsets of E. Let us
remember that S is invariant with respect to G (or, briey, G-invariant) if
(X S)(g G)(g(X) S).
Suppose, in addition, that is a measure dened on a G-invariant -algebra S of subsets
39
3
of E. We usually say that is invariant with respect to G (or, briey, G-invariant) if
(X S)((g(X)) = (X))
or, in other words, the values of are preserved under the action of G in E.
It should be observed that invariant measures can be met in various elds of modern math-
ematics and are intensively investigated from different points of view (see [6], [41], [45],
[46], [47], [65], [80], [81], [83], [85], [86], [93], [95], [108], [115], [141], [142], [165],
[168], [182], [191], [193], [195], [198], [229], [242], [243], [245], [246], [247], [248], and
[249]). In particular, they play a signicant role in mathematical analysis (especially, in
various topics of harmonic analysis and functional analysis), probability theory, ergodic
theory, and topological dynamics. Therefore, it is reasonable to touch upon (very briey,
of course) some general properties of such measures.
But here we do not intend to discuss the fundamental question concerning the existence
of a nonzero invariant (nite or -nite) measure for a given space (E, G) and for a G-
invariant -algebra S P(E) on which is required to be dened. Questions of this
type and related to themare thoroughly envisaged, for example, in the works by Zakrzewski
(see [245] and [246]). Our goal is rather moderate and we only wish to recall to the reader
certain classical results from the theory of invariant measures (e.g., the famous Krylov-
Bogoliubov theorem and its generalization obtained by Markov and Kakutani) and provide
the reader with essential information about -nite invariant measures on abstract spaces
endowed with transformation groups.
First of all, we would like to pay our attention to several important examples of invariant
measures. Also, we want to present certain general constructions of such measures and
demonstrate how those constructions lead to many other nontrivial examples of invariant
measures.
Undoubtedly, the best known examples of invariant measures are the standard Borel mea-
sure and the standard Lebesgue measure on the Euclidean space R
n
(n 1). Actually,
the main achievement of Borel and Lebesgue was the idea of constructing a positive (i.e.,
nonnegative) countably additive translation-invariant functional on R
n
which signicantly
extends the classical Jordan volume on the same space.
Denote by the usual symbol B(R
n
) the -algebra of all Borel subsets of R
n
. The standard
Borel measure on R
n
can be dened as a (unique) measure
n
satisfying the following three
conditions:
1) dom(
n
) =B(R
n
);
2)
n
([0, 1]
n
) = 1;
Invariant measures 41
3)
n
is invariant under the group of all translations of R
n
.
Since the space R
n
is locally compact, the measure
n
may also be regarded as a very par-
ticular case of a Haar measure on a locally compact topological group. The corresponding
denition of a Haar measure will be formulated below.
The completion of
n
is usually called the standard Lebesgue measure on the space R
n
and
is often denoted by the same symbol
n
. Obviously, the standard Lebesgue measure on R
n
can be described as a smallest (with respect to inclusion) measure
n
on R
n
satisfying the
following three relations:
(1)
n
is a complete measure;
(2) [0, 1]
n
dom(
n
) and
n
([0, 1]
n
) = 1;
(3)
n
is invariant under the group of all translations of R
n
.
It should be noticed that the above-mentioned relations directly imply the invariance of
n
under the group of all isometric transformations of R
n
. This circumstance can be easily
established by using a purely geometric argument starting with the fact that the family of
all reections (i.e., mirror symmetries with respect to afne hyperplanes in R
n
) generates
the group of all isometries of R
n
(see Exercise 3).
Let E be a locally compact topological group and let G be the group of all left (right)
translations of E. We recall that a left (right) Haar measure on E is a measure satisfying
these four conditions:
a) is not identically equal to zero;
b) dom() =B(E) and is a Radon measure on E, i.e., for any Borel set X E, we have
(X) = sup(K) : K X, K is compact;
c) is invariant with respect to G;
d) for every compact subset K of E, we have (K) < +.
Notice that condition d) is equivalent to the following condition:
d) for each point x E, there exists a neighborhood V(x) of x such that
(V(x)) < +.
In fact, the latter condition simply says that is a locally nite measure.
The famous theoremdue to Haar states that there exists at least one left (right) Haar measure
on E and, in addition, is unique in a certain sense. More precisely, any two left (right)
Haar measures on E are proportional to each other.
The theory of Haar measure is thoroughly considered in numerous books and monographs
(see, for instance, [80], [83], or [182]). In particular, the theorem mentioned above and
concerning the existence and uniqueness of Haar measure can be found in the same books
[80], [83], and [182]. A detailed proof of this theorem is also given therein as well as in
various other manuals of measure theory. In our book, we do not need many facts from this
theory.
Let us emphasize that if E is a compact topological group, then the left and right Haar mea-
sures are proportional and, therefore, can be identied. This result is not true, in general,
even for -compact locally compact topological groups. The corresponding examples of
essentially distinct left and right Haar measures on some groups of this type are due to von
Neumann and are presented, for instance, in [80] and [83].
Let us also remark that if E is a -compact locally compact topological group, then a Haar
measure on E is -nite. Conversely, it is not difcult to show that if a Haar measure
on a locally compact topological group E is -nite, then E is -compact (see Exercise 5).
The case of a Polish locally compact topological group E is of paramount importance. In
this case, taking into account the circumstance that E possesses a countable base consisting
of open sets with compact closures, we get that E is simultaneously a -compact space. In
view of the preceding remarks, we conclude that in such a situation the Haar measure
on E is -nite. In connection with this fact, let us also recall that a rather general result
due to Ulam holds true, namely, every -nite Borel measure on a Polish topological space
is necessarily Radon (see Exercise 16 from Chapter 2). The same result remains valid for
-nite Borel measures on Suslin spaces (see Appendix 2).
Example 1. Let G be the group of all isometric transformations of the n-dimensional
Euclidean space R
n
. Let us equip G with its natural topology which is induced by the
topology of the Euclidean space of dimension n
2
. Then G becomes a locally compact
Polish topological group and, moreover, it turns out to be a Lie group. Consequently, we
may consider a -nite Haar measure on G. More generally, let H be an arbitrary closed
subgroup of G. Then H is also a locally compact Polish topological group and we may
consider a -nite Haar measure on H. Actually, is isomorphic, in the purely measure-
theoretical sense, to an appropriate combination of the following canonical measures:
(a) the Haar measure on the discrete group Z of all integers;
(b) the standard Borel measure on the real line R;
(c) the Haar measure on the one-dimensional unit torus S
1
(= T);
(d) the Haar probability measure on a nite subgroup of S
1
.
Let (, ) be a topological group. We recall that is monothetic if there exists an element
g such that the subgroup of generated by the one-element set g is everywhere dense
in . Clearly, such a is necessarily commutative.
For instance, it can easily be checked that the torus S
1
considered as a commutative com-
pact topological group is monothetic. Indeed, any element of S
1
which has innite order
generates an everywhere dense subgroup of S
1
.
Example 2. Let (G, ) be an arbitrary compact metrizable monothetic topological group
equipped with the Haar probability measure . Let h be an element of G generating an
everywhere dense subgroup of G. We denote this countable everywhere dense subgroup
by H. Suppose now that is a Borel probability measure on G left invariant with respect
to H. Then it is not hard to show that coincides with the Haar measure . Indeed, let us
dene a function
h
: G G
by putting
h
(g) = h g (g G)
and consider an arbitrary real-valued continuous function p on G. Since is left invariant
with respect to H, we have the equalities
G
p((h
)
n
(g))d(g) =
G
p(h
n
g)d(g) =
G
p(g)d(g)
for every integer n. Further, let us x an element f G. Since G is metrizable, there exists
a sequence n
k
: k N of integers, such that
lim
k
h
n
k
= f .
Therefore, we obtain the following relation:
G
p( f g)d(g) =
G
lim
k
p(h
n
k
g)d(g) =
G
p(g)d(g),
which shows that for any f G, the -integrals of the functions
p : G R, p f
: G R
are equal to each other. Since p and f were taken arbitrarily, it is not difcult to deduce
from the obtained result that our measure is left invariant with respect to the whole group
G. Now, applying the uniqueness theorem to the Haar measure , we get the required
equality = .
In particular, if is a Borel probability measure on S
1
invariant under an innite cyclic
subgroup of S
1
which automatically is everywhere dense in S
1
, then coincides with the
Haar probability measure on S
1
.
Let G be a Polish topological group. As already mentioned, if G is locally compact, then
there exists a nonzero -nite Haar measure on G invariant with respect to the group of
all left (right) translations of G. Suppose now that G is not locally compact. Then the
following question naturally arises.
Does there exist a nonzero -nite Borel measure on G invariant with respect to the group
of all left (right) translations of G?
As far as we know, the question formulated above was rst raised by Gelfand (cf. the
comments given in [226]). Several authors investigated this question and it was nally
established that the answer to it is negative. Moreover, a much stronger result can be
obtained by using some simple geometrical properties of -compact subsets of a non-
locally compact Polish group G (see the next chapter devoted to -nite quasi-invariant
measures on groups).
In connection with the just said, it is also reasonable to remark that some Polish non-locally
compact topological groups admit nonzero left (right) invariant Borel measures (see, e.g.,
[6] and [191]). Of course, those measures are not -nite and, therefore, they are of less
interest for mathematical analysis, probability theory, and their applications.
As mentioned earlier, the standard Borel measure on the real line R is an essentially unique
-nite measure dened on the -algebra of all Borel subsets of R and invariant with
respect to the group of all translations of R. By using this uniqueness property, it is not
difcult to describe the class of all nonzero -nite R-quasi-invariant Borel measures on
R (see Exercise 4 for Chapter 4).
Now, let us consider some natural operations which enable us to produce a new space
with a -nite invariant measure by starting from an initial family of spaces equipped with
corresponding -nite invariant measures.
First, let us recall a purely algebraic notion of the product (respectively, weak product) of
a family of groups.
Suppose that a family of groups (G
i
,
i
) : i I is given and let e
i
denote the neutral
element of G
i
for each index i I.
As usual, the product group
(G, ) = (
iI
G
i
, )
of this family is dened by introducing the group operation in G, according to the formula
g h = (g
i
i
h
i
)
iI
for any two elements g = (g
i
)
iI
and h = (h
i
)
iI
from G =
iI
G
i
.
Further, we put
G
=g G : i I : g
i
,= e
i
is f inite.
It can easily be checked that G
is a subgroup of G. The group G
is usually called the

weak product group of a given family (G
i
,
i
) : i I. If all groups (G
i
,
i
) (i I) are
commutative, then the weak product group coincides with the direct sumof (G
i
,
i
) : i I.
We need the following easy auxiliary proposition.
Lemma 1. Let E be a set and let G be a group of transformations of E. Suppose that a
G-invariant algebra S of subsets of E is given with a -nite G-invariant measure on
S. Let S
/
= (S) denote the -algebra generated by S and let
/
denote the unique
extension of dened on S
/
. Then S
/
is a G-invariant -algebra and
/
is a G-invariant
measure.
Proof. Consider the family of sets
T =X S
/
: (g G)(g(X) S
/
).
A straightforward verication shows that T is a G-invariant -algebra of subsets of E and
S T . Consequently,
S
/
= (S) T .
At the same time, according to the denition of T , we have T S
/
. Therefore, S
/
=T
and S
/
is a G-invariant -algebra of subsets of E.
Now, let us introduce the class of sets
L =X S
/
: (g G)((g(X)) = (X)).
In a similar way, we can verify that the class L contains S and is closed under the unions
of increasing sequences of members from L and is also closed under the intersections of
decreasing sequences of members from L. This circumstance implies that L = S
/
(see
Exercise 7) and, therefore,
/
is a G-invariant measure on S
/
.
Unfortunately, we do not have an analogous proposition for -nite quasi-invariant mea-
sures (see Exercise 8).
The next simple statement concerns the behavior of invariant probability measures under
the operation of their product.
Theorem 1. Let (E
i
, G
i
) : i I be a family of spaces equipped with transformation
groups. Suppose that for each index i I, a probability measure
i
is given on the set E
i
and denote by E the product of the family of sets E
i
: i I. Let be the product measure
of the family
i
: i I and let G be the product group of the family G
i
: i I.
If all measures
i
(i I) are G
i
-invariant, then the product measure is G-invariant.
Proof. The assertion follows directly from the denition of the product measure and
from Lemma 1.
An analogous statement is true for any nite family of -nite invariant measures (see
Exercise 9).
A deeper result is contained in the next statement (see, for example, [115]).
Theorem 2. Let E be a set, G
n
: n N be a sequence of groups of transformations of
E, and let
n
: n N be a sequence of -nite measures given on E. Suppose also that
these ve conditions are satised:
1) the family G
n
: n N is increasing with respect to inclusion;
2) the family dom(
n
) : n N is increasing with respect to inclusion;
3) for each n N, the measure
n
is invariant under the group G
n
;
4) there exists a sequence Y
n
: n N of subsets of E such that for any n N, we have
Y
n
dom(
n
), Y
n
Y
n+1
,
and Y
n
is a G
n
-invariant support of
n
, i.e.,
n
(E Y
n
) = 0, (g G
n
)(g(Y
n
) =Y
n
);
5) for each n N, if a set X belongs to dom(
n+1
), then the set X Y
n
belongs to dom(
n
)
and we have the equality
n
(X Y
n
) =
n+1
(X Y
n
).
Let us denote
G =G
n
: n N
and put
S =Z E : (n N)(Z Y
n
dom(
n
)).
Then S is a G-invariant -algebra of subsets of E. Further, for each set Z S, dene
(Z) = lim
n

n
(Z Y
n
).
Then the denition of is correct, is a -nite G-invariant measure on the -algebra
S and the set
Y =Y
n
: n N
is a G-invariant support of . Furthermore, for any n N, we have
n
= [dom(
n
).
Proof. First, let us observe that if Z S, then for all natural numbers n, the relation
n
(Z Y
n
) =
n+1
(Z Y
n
)
n+1
(Z Y
n+1
)
is fullled. Consequently, there exists a limit
lim
n

n
(Z Y
n
) [0, +]
which will be denoted by (Z). In this way, we obtain a functional
: S [0, +].
It is not difcult to check the countable additivity of , so is a measure on S. Also, it
directly follows from the denition of that for every natural number n, the restriction of
to dom(
n
) coincides with the measure
n
.
All other above-mentioned properties of are easily veried and we leave the correspond-
ing details to the reader.
The measure described in Theorem 2 may be regarded as an inductive limit of the given
countable family
n
: n N of invariant measures.
Example 3. We can directly apply Theorem 2 in certain concrete situations which are
motivated by topics of classical mathematical analysis or probability theory. For instance,
let us consider the product space
E = R
N
=
R
k
: k N,
where R
k
= R for all natural indices k. Furthermore, for each k N, let us denote by
k
the standard Borel measure on R
k
and let
k
be the Borel probability measure on E whose
restriction to the unit segment [0, 1] R
k
coincides with the standard Borel measure on
[0, 1]. Finally, we put
n
= (
k
: k n) (
k
: k > n)
and denote
G
n
= R
0
... R
n
0 0 0 ...
for every natural number n. Now, it can easily be checked that the family
n
: n N
of -nite measures and the family G
n
: n N of transformation groups satisfy the
assumptions of Theorem 2. Consequently, for the set E equipped with the group
G =G
n
: n N = R
(N)
of its transformations, there exists a nonzero -nite G-invariant measure having the
properties described in Theorem 2. Also, in view of the denition, all -algebras dom(
n
)
coincide with the Borel -algebra of E. Hence we have dom() =B(E) and
([0, 1]
N
) = 1.
We thus conclude that there exists a nonzero -nite Borel measure on the innite-
dimensional Polish topological vector space R
N
, invariant with respect to the everywhere
dense vector subspace R
(N)
consisting of all eventually nite real-valued sequences.
Now, we wish to return to a nite-dimensional Euclidean space and discuss one important
geometric property of the classical Lebesgue measure given on this space. We mean here
the so-called Steinhaus property which was rst established by Steinhaus in his remarkable
paper [233].
Let
n
denote, as usual, the standard Lebesgue measure on the n-dimensional Euclidean
space R
n
and let G be the group of all isometric transformations of R
n
. Obviously, we
may treat G as a locally compact Polish topological group with respect to the natural group
operation and topology. Consider an arbitrary Lebesgue measurable subset Z of R
n
with
n
(Z) > 0. Then it is not hard to show that there exists a neighborhood U of the neutral
element of G, such that
g(Z) Z ,= / 0
for all transformations g U. This property of the set Z is usually called the Steinhaus
property of Z. Accordingly, it is frequently said that
n
possesses the Steinhaus property.
In particular, if we restrict our consideration to the group of all translations of R
n
(which is
canonically isomorphic to R
n
), then we easily obtain that there exists a neighborhoodW of
zero in R
n
such that
(h +Z) Z ,= / 0
for any translation h W. In other words, we have the relation
W Z Z =x y : x Z, y Z,
which can successfully be applied in some constructions of Lebesgue nonmeasurable sub-
sets of the space R
n
(see, for instance, [119]).
Notice that the Steinhaus property of
n
admits a direct analogue for sets measurable with
respect to a -nite Haar measure given on a -compact locally compact topological group
(see [80], [83], and Exercise 13).
It is also reasonable to point out here that the Steinhaus property has an analogue in terms
of the Baire category for arbitrary topological groups. Namely, let (G, ) be a topological
group and let Z be a subset of G having the Baire property and not belonging to the family
of all rst category subsets of G. Then the set Z Z
1
contains a neighborhood of the
neutral element of G. This fact is sometimes called the Banach-Kuratowski-Pettis theorem
(cf. [101], [148], and Exercise 15 in Appendix 3).
Now, we turn our attention to an important case of (E, G), where E is a compact metric
space and G is a certain group of homeomorphisms of E onto itself. In this situation,
several fundamental results have been obtained by famous authors, among which there
were Krylov, Bogoliubov, Markov, Kakutani, and others.
Let T be a nonempty compact metric space and let u : T T be some homeomorphism.
Since u is a certain transformation of T, the natural question arises whether there exists
a probability Borel measure on T invariant with respect to the group of transformations
generated by u. Evidently, the last group is cyclic and coincides with the family
[u] =u
m
: m Z,
where Z denotes the set of all integers. In general, [u] is innite but, sometimes, may
be nite. Here we wish to present a direct proof of the remarkable result due to Krylov
and Bogoliubov, according to which such a measure does always exist. Notice that the
argument uses some basic concepts of functional analysis (see, e.g., [47]).
First, let us make several preliminary remarks. As usual, we denote by the symbol
C(T) =C(T, R)
the Banach space of all real-valued continuous functions on T. By virtue of the classical
Riesz theorem (see, for instance, [56] and [57]), there is a canonical one-to-one correspon-
dence between Radon probability measures on T and linear normalized positive functionals
on C(T). In other words, any Radon probability measure on T can be canonically iden-
tied with some linear normalized positive (hence continuous) functional on C(T). For
the sake of convenience and brevity, this functional will be denoted by the same symbol
. Obviously, the set M of all such functionals is convex in the conjugate space (C(T))
.
Moreover, M turns out to be compact with respect to the weak topology ((C(T))
,C(T)).
The latter fact immediately follows from the classical Tychonoff theorem on products of
quasicompact spaces by taking into account the circumstance that M is a closed convex
subset of the topological product
f C(T)
[[[ f [[, [[ f [[].
In addition, since T is a compact metric space, C(T) is separable and hence M is metrizable
as well, because M can also be regarded as a closed subset of the product
f D
[[[ f [[, [[ f [[],
where D is a xed countable everywhere dense subset of C(T). So we derive that M is a
compact metrizable space which implies that any sequence of points from M contains at
least one convergent subsequence (cf. Example 2 of this chapter).
Now, take an arbitrary functional M and for each natural number n, dene a functional
n
by the formula
n
( f ) =
( f ) +( f u) +... +( f u
n
)
n +1
( f C(T)).
It can easily be seen that
n
M. Thus, we obtain the sequence
n
: n N of points
from M. Let n(k) : k N be a strictly increasing family of natural numbers, such that
the corresponding partial sequence
n(k)
: k N converges to some point M. Let us
verify that
( f ) = ( f u) ( f C(T)).
Indeed, we may write
( f ) = lim
k
n(k)
( f ),
( f u) = lim
k
n(k)
( f u).
But a simple calculation shows that
n(k)
( f )
n(k)
( f u) =
( f ) ( f u
n(k)+1
)
n(k) +1
,
which implies
[
n(k)
( f )
n(k)
( f u)[
2[[[[ [[ f [[
n(k) +1
.
Tending k to innity and taking into account the fact that n(k) k, we come to the equality
( f ) = ( f u)
for all functions f C(T). In addition, since u is a homeomorphism of T onto itself, we
have
( f u
1
) = ( f u
1
u) = ( f )
for all f C(T). So we conclude that our measure is invariant under the cyclic group of
transformations generated by u.
However, this group is rather small because it is at most countable. Kakutani [93] and
Markov [168] extended the above result to any commutative group of homeomorphisms of
T. For this purpose, they proved the following fundamental statement (cf. [22] and [57]).
Theorem 3. Let E be a Hausdorff topological vector space, K be a nonempty compact
convex subset of E and let G be a family of afne continuous mappings acting from K
into itself. As usual, denote by I
K
the identity transformation of K and suppose that the
following condition is fullled:
there exists a nite sequence G
0
, G
1
, ..., G
n
such that
I
K
= G
n
G
n1
... G
1
G
0
= G
and for each natural index i [1, n] and for any two elements u G
i1
and v G
i1
, there
is an element w G
i
satisfying the equality u v = v u w.
Then there exists at least one point of K invariant under G, that is there exists a point x K
for which the relation
(u G)(u(x) = x)
holds true.
Proof. We use induction on n. The case n = 0 is trivial.
Consider the case n =1. In this situation, we have uv =vu for any two elements u and v
from G
0
=G, i.e., all elements from G pairwise commute. Take any u G and for a natural
number r > 0, put
u
r
= (I
K
+u +u
2
+... +u
r1
)/r.
In view of the convexity of K, we get u
r
(K) K. Let G
denote the family of all nite

compositions of mappings of the form
u
r
(u G, r N 0).
It is clear that all elements of G
pairwise commute, are afne and continuous, and map K

into itself. Consider now the family of nonempty sets
v(K) : v G
.
All members of this family are compact. In addition, if v
/
G
and v
//
G
, then
(v
/
v
//
)(K) = (v
//
v
/
)(K), (v
/
v
//
)(K) v
/
(K), (v
//
v
/
)(K) v
//
(K),
which readily implies that the above-mentioned family is centered. Therefore,
v(K) : v G
,= / 0.
Choose a point x
0
v(K) : v G
. We are going to show that u(x

0
) =x
0
for any u G.
Indeed, if r is a nonzero natural number, then x
0
u
r
(K) which means that there exists
y K such that
x
0
= (y +u(y) +... +u
r1
(y))/r.
Consequently, we may write
u(x
0
) x
0
= (u
r
(y) y)/r (K K)/r.
Let C be an arbitrary star-like symmetric open neighborhood of the neutral element of E.
Clearly,
KK mC : m N 0.
Since K K is compact, we come to the inclusion K K mC (or, equivalently, to the
inclusion (K K)/m C) for all sufciently large natural numbers m and, in particular,
for sufciently large r. In other words, the relation
u(x
0
) x
0
C
is valid. Since E is Hausdorff, we infer at once that u(x
0
) x
0
= 0 and thus u(x
0
) = x
0
.
Suppose now that the assertion of the theorem has already been established for n and let us
prove it for n +1. Consider the case when a sequence
I
K
= G
n+1
G
n
... G
1
G
0
= G
is given with the corresponding properties. Applying the inductive assumption to G
1
, we
deduce that the set
K
1
=x K : (u G
1
)(u(x) = x)
is nonempty. Also, it can easily be veried that the same K
1
is convex and compact. Take
any two elements u
1
G and u
2
G. According to the condition of the theorem, there is
w G
1
such that
u
1
u
2
= u
2
u
1
w.
If x K
1
, then we have
(u
1
u
2
)(x) = (u
2
u
1
)(w(x)) = (u
2
u
1
)(x).
This circumstance means that all elements from G pairwise commute on K
1
. Finally, let us
take any u G and any w
1
G
1
G. Then, according to the same condition, we may write
the equality
w
1
u = u w
1
w
2
for some mapping w
2
G
1
. Consequently, for each x K
1
, we get
w
1
(u(x)) = u((w
1
w
2
)(x)) = u(x),
thus it follows that u(x) K
1
. In other words, all mappings from G pairwise commute and
transform K
1
into itself. So we come again to the case, where n = 1. As already shown,
there exists a point x
0
K
1
invariant under all mappings fromG which completes the proof.
Remark 1. The theorem of Markov and Kakutani is usually applied in those situations,
where G is a solvable group of continuous afne transformations of a Hausdorff topological
vector space E, which map a nonempty compact convex set K E into itself. Clearly, in
such a case, the condition of Theorem 3 is automatically satised.
In order to give an application of Theorem3 to the existence of dynamical systems, we also
need the following simple auxiliary proposition.
Lemma 2. Let T be a nonempty compact topological space and let G be a solvable group
of homeomorphisms acting from T onto itself. For each element g G, dene the mapping
g
/
: (C(T))
(C(T))
by the formula
g
/
() =
/
( (C(T))
),
where
/
( f ) = ( f g) ( f C(T)).
Then the family G
/
= g
/
: g G is a solvable group of continuous afne mappings of
(C(T))
into itself for which we have

g
/
(M) M (g
/
G
/
),
where M = M(T) denotes the compact convex set of all Radon probability measures on the
given space T.
The proof reduces to a straightforward verication. The corresponding details are not dif-
cult and are left to the reader.
Applying Theorem 3 and Lemma 2 to G
/
and to M, we obtain the following statement of
Markov and Kakutani.
Theorem 4. Let T be a nonempty compact space and let G be a solvable group of homeo-
morphisms of T onto itself. Then there exists a Radon probability measure on T invariant
under G, that is
(g(X)) = (X) (g G, X B
0
(T)).
Remark 2. Let E be a set and let be a homomorphism of the additive group R into the
symmetric group (Sym(E), ). Such a homomorphism is usually called a one-parameter
group of transformations of E. Note that the homomorphic image (R) is also frequently
called a one-parameter group of transformations of E. Obviously, any one-parameter
group is automatically commutative. Consequently, if in Theorem 4 the group G is a one-
parameter group of homeomorphisms of T onto itself, then there always exists a Radon
probability G-invariant measure on T. This result is due to Krylov and Bogoliubov.
Remark 3. Let G be a solvable compact topological group. Theorem 4 immediately im-
plies the existence of a Haar probability measure on G. In addition, by using the classical
Fubini theorem to the product space GG, it is not hard to prove the uniqueness of a Haar
probability measure on G (cf. [209] and Exercise 17).
One important application of Theorem 3 is presented in Exercise 9 of Chapter 17.
Of course, our survey of invariant measures is far from being a complete one. Further
information about various properties of invariant and quasi-invariant measures can be found
in the works which were already mentioned in this chapter (see also the next chapter).
EXERCISES
1. Any mapping f : R
n
R
n
is often called a vector eld in R
n
because f associates the
vector f (x) R
n
to every point x R
n
. This terminology is especially convenient when
the following autonomous system of rst-order ordinary differential equations is studied:
x
/
1
= f
1
(x
1
, x
2
, ..., x
n
),
x
/
2
= f
2
(x
1
, x
2
, ..., x
n
),
..................................
x
/
n
= f
n
(x
1
, x
2
, ..., x
n
),
where f = ( f
1
, f
2
, ..., f
n
). It is usually assumed that the mapping f satises the Lipschitz
condition with respect to each of its arguments (so f is necessarily continuous). Under this
assumption for any initial value (x
0
1
, x
0
2
, ..., x
0
n
) R
n
, there exists a unique local solution
(local trajectory)
(x
1
(t), x
2
(t), ..., x
n
(t)) (a t a)
of this system which has the initial value
x
1
(0) = x
0
1
, x
2
(0) = x
0
2
, ... , x
n
(0) = x
0
n
at t = 0. If the norm of f is bounded, then the local solution (local trajectory) turns out
to be global, i.e., it is dened on the whole R. Actually, we may suppose without loss
of generality that the norm of f is bounded because an appropriate replacement of f by
a vector eld with bounded norm is always possible in such a way that the trajectories
(considered as point-sets) remain the same. We will assume that this replacement is done,
so all solutions of our system are global.
Verify that the vector eld f induces a certain one-parameter group
G =g
t
: t R
of transformations of the space R
n
. Namely, for each t R, the transformation
g
t
: R
n
R
n
is dened as follows:
for any point (x
0
1
, x
0
2
, ..., x
0
n
) R
n
, we put
g
t
(x
0
1
, x
0
2
, ..., x
0
n
) = (x
1
(t), x
2
(t), ..., x
n
(t)),
where (x
1
(t), x
2
(t), ..., x
n
(t)) is the point of the above-mentioned trajectory, corresponding
to t. We shall say that the group G is associated with the vector eld f .
Conversely, suppose that a one-parameter group G =g
t
: t R of transformations of R
n
is given. Under some smoothness assumptions on G, show that this group induces a certain
vector eld f : R
n
R
n
whose corresponding autonomous system yields (by the procedure
described above) the same G.
2
. We preserve the notation of the previous exercise. Assume that a vector eld
f = ( f
1
, f
2
, ..., f
n
) : R
n
R
n
is sufciently smooth and denote
div( f ) = f
1
/x
1
+ f
2
/x
2
+... + f
n
/x
n
.
Prove that the following two assertions are equivalent:
(a) the Lebesgue measure
n
is invariant under the one-parameter group G = g
t
: t R
associated with f ;
(b) div( f )(x) = 0 for all points x R
n
.
This classical result is due to Liouville (see, for example, [196]).
In particular, take the Euclidean space R
2n
and consider the autonomous Hamiltonian sys-
tem
x
/
1
= H/y
1
, x
/
2
= H/y
2
, ..., x
/
n
= H/y
n
,
y
/
1
=H/x
1
, y
/
2
=H/x
2
, ..., y
/
n
=H/x
n
,
where H : R
2n
R is a smooth real-valued function of 2n real variables
x
1
, x
2
, ..., x
n
, y
1
, y
2
, ..., y
n
. Check that the Lebesgue measure
2n
on R
2n
is invariant under
the one-parameter group associated with this system.
3. Show that the group of all isometric transformations of the space R
n
is generated by the
family of all mirror symmetries (i.e., orthogonal reections with respect to afne hyper-
planes).
Show also that for any afne hyperplane in R
n
and for any open subset U of R
n
, it is
possible to represent U in the form
U =K
i
: i I
where K
i
: i I is a countable family of closed n-dimensional cubes whose interiors are
pairwise disjoint and, in addition, each cube K
i
(i I) has a facet parallel to .
Infer from these two geometrical facts that the Lebesgue measure
n
is invariant under the
group of all isometric transformations of R
n
(for n 1, this group is not commutative and
hence is not a one-parameter group).
4
. Let be a cardinal number and let S

1
be the unit torus in R
2
regarded as a compact
commutative group. Prove that the following three assertions are equivalent:
(a) does not exceed c;
(b) the topological product group S
1
is monothetic;
(c) the topological product group S
1
is separable which means that it contains a countable
everywhere dense subset.
For establishing the implication (a) (b), use the classical theoremof Kronecker (see, e.g.,
[23]), which states that for given real numbers
1
,
2
, ...,
n
, the following two relations are
equivalent:
(d) the system of numbers 1,
1
,
2
, ...,
n
is linearly independent over the eld Q;
(e) for any real numbers x
1
, x
2
, ..., x
n
and for every real > 0, there exist integers
p
1
, p
2
, ..., p
n
, q such that
[q
j
p
j
x
j
[ < (1 j n).
Starting with this theorem of Kronecker, consider R as a vector space over Q and x a
linearly independent system 1
j
: j J in R, where
card(J) = c.
Further, dene the element
s =s
j
: j J =(cos(2
j
), sin(2
j
)) : j J S
J
1
and verify that s generates the cyclic group which is everywhere dense in S
J
1
.
Conclude from the said above that S
1
is a nonmetrizable compact monothetic group when-
ever
1
c.
5. Let G be a locally compact topological group and let be the left Haar measure on G.
(a) is -nite;
(b) G is -compact.
For this purpose, apply Exercise 13 from Chapter 2.
Also, prove that the next two assertions are equivalent:
(c) is nite;
(d) G is compact.
6. Complete the details of Example 2.
7. Let E be a set and let L be a class of subsets of E. According to the standard denition,
L is called a monotone class if L is closed under the unions of all increasing sequences
of its members and is also closed under the intersections of all decreasing sequences of its
members (see, e.g., [80] and [187]).
Let S be an algebra of subsets of E. Show that any monotone class containing S also con-
tains the -algebra (S). Conclude fromthis fact that (S) coincides with the monotone
class generated by S.
8. Give an example of a probability quasi-invariant measure dened on an algebra of
sets, such that the unique extension
/
of dened on the generated -algebra does not
have the quasi-invariance property with respect to the same transformation group.
9. Let
i
: i I be a nite family of -nite measures such that every measure
i
(i I)
is invariant under some group G
i
. Show that the product measure
=
i
: i I
is invariant under the product group G = G
i
: i I.
10. Complete the details of the proof of Theorem 2.
11
. Show that the measure constructed in Example 3 is metrically transitive (i.e., er-
godic) with respect to R
(N)
. In other words, demonstrate that for any -measurable set Z
with (Z) > 0, there exists a countable family
g
n
: n N R
(N)
such that
(R
N
g
n
+Z : n N) = 0.
12
. We recall the notation

l
1
=(x
n
)
nN
R
N
:

nN
[x
n
[ < +,
l
2
=(x
n
)
nN
R
N
:

nN
[x
n
[
2
< +.
Observe that both l
1
and l
2
are Borel vector subspaces of the Polish topological vector
space R
N
. In addition, l
1
can be considered as a separable Banach space with respect to the
natural norm
[[x[[ =

nN
[x
n
[ (x = (x
n
)
nN
l
1
)
and l
2
can be considered as a separable Hilbert space with respect to the natural inner
product
< x, y > =

nN
x
n
y
n
(x = (x
n
)
nN
l
2
, y = (y
n
)
nN
l
2
).
Prove that there exists a nonzero -nite Borel measure on the Banach space l
1
, invariant
with respect to R
(N)
. Using this result and taking into account the relation l
1
l
2
, infer that
for every innite-dimensional separable Hilbert space E, there exists a nonzero -nite
Borel measure on E invariant with respect to some everywhere dense vector subspace of E.
13
. Let Gdenote the group of all isometric transformations of R

n
, equipped with its natural
topology. By applying the classical Lebesgue theorem on density points of Lebesgue mea-
surable sets (see, e.g., [183], [192], [210], or Appendix 4), prove that for any
n
-measurable
set Z R
n
with
n
(Z) > 0, there exists a neighborhoodV of the neutral element of G, such
that
n
(g(Z) Z) > 0
for all elements g V; in particular, we have g(Z) Z ,= / 0 for all g V.
Moreover, using the fact that the standard Borel measure on R
n
is Radon, show that
lim
ge

n
(g(Z) Z) =
n
(Z),
where e denotes the neutral element of G.
Formulate and prove an analogous result for a Haar measure on a -compact locally com-
pact topological group.
14
. Let S be a sphere in R
n
(n 2) equipped with the (n 1)-dimensional Lebesgue
probability measure which is invariant under the group G of all rotations of S about its
center. Fix a natural number k > 0. Let A be an arbitrary -measurable subset of S with
(A) < 1/k and let B be an arbitrary subset of S with card(B) k. Show that there exists
at least one rotation g G for which we have g(A) B = / 0.
15
. Describe all those (sufciently smooth) homeomorphisms g : R

2
R
2
which preserve
the two-dimensional Lebesgue measure
2
on the plane R
2
.
16. Consider the unit circle S
1
equipped with a group G of its homeomorphisms onto itself,
such that:
(a) G contains at least one rotation of innite order;
(b) G contains at least one transformation which is not an isometry of S
1
.
Show that there exists no G-invariant Borel probability measure on S
1
.
Conclude from this fact that no such group G can be solvable.
17
. By using the Fubini theorem, prove the uniqueness of a left (right) Haar probability
measure on a compact topological group (G, ). In other words, demonstrate that if is
any Borel probability left (right) invariant measure on G, then = . Show also that the
left Haar measure on G coincides with the right Haar measure on G.
Further, apply the uniqueness property of the left Haar measure on a -compact locally
compact group (H, ) and prove that is metrically transitive (ergodic) with respect to the
group of all left translations of H, i.e., prove that for every -measurable set X H with
(X) > 0, there exists a countable family h
i
: i I H such that
(H h
i
X : i I) = 0.
Obtain an analogous result for the right Haar measure on H.
18
. Let (E, G) be a space equipped with a transformation group and let be a -nite
G-invariant measure on E. Suppose that a family Z
i
: i I of subsets of E is given, which
satises the following conditions:
(a) every set Z
i
(i I) is almost G-invariant with respect to ;
(b) for any countable subset J of I and for any family Z
/
j
: j J such that
(j J)(Z
/
j
= Z
j
Z
/
j
= E Z
j
),
the intersection Z
/
j
: j J is -thick in E, i.e.,
(E (Z
/
j
: j J)) = 0.
Consider the family T of all those subsets X of E which admit a representation in the form
X =Z
f (1)
i
1
Z
f (2)
i
2
... Z
f (n)
i
n
Y
f
: f 0, 1
n
,
where n is an arbitrary natural number, i
1
, i
2
, ..., i
n
is an injective sequence of elements
from I, all sets Y
f
( f 0, 1
n
) belong to dom() and Z
f (k)
i
k
= Z
i
k
if f (k) = 0, and Z
f (k)
i
k
=
E Z
i
k
if f (k) = 1.
Verify that T is an algebra of sets and the inclusion dom() T holds true.
Further, dene a functional
/
on T by the formula
/
(X) = (1/2
n
)
(Y
f
) : f 0, 1
n
.
Check that this denition of
/
is correct in the sense that the value
/
(X) does not depend
on a representation of X in the above-mentioned form. Check also that
/
is nitely additive
and extends . Moreover, applying the K onig lemma on -trees (see Appendix 1), show
that
/
is countably additive and, therefore,
/
is a measure on T extending .
Finally, using Lemma 1 of this chapter, conclude that
/
can be extended to a G-invariant
measure dened on the -algebra (T ). For the sake of brevity, we denote this measure
by the same symbol
/
.
Observe that if the original measure is nonzero and card(I) >, then
/
is a nonseparable
extension of .
19. Let (E, G) be a space equipped with a transformation group, for which the following
conditions are satised:
(i) E is uncountable and card(G) card(E);
(ii) G acts transitively on E, i.e., for any two points x E and y E, there exists g G such
that g(x) = y.
Let denote the least ordinal number whose cardinality is equal to card(E). Prove that E
admits a representation in the form
E =X
: <
for which the following relations are valid:
(a) card(X
) card() + for every ordinal < ;

(b) X
= / 0 for any two distinct ordinals < and < ;

(c) for each set [0, [, the associated union
X() =X
:
is an almost G-invariant subset of E, i.e., we have
(g G)(card(g(X())X()) < card(E)).
In order to show the existence of X
: < , rst observe that the conditions (i) and (ii)

imply the equality
card(G) = card(E).
Then construct an -sequence G
: < of subgroups of G satisfying the relations:

(1) card(G
) card() + for any < ;

(2) G
: < = G;
(3) G
whenever < < .

Further, x a point e E and for each ordinal < , put
X
= G
(e) G
(e) : < .
Verify that the family of sets X
: < yields the desired result.

20
. Let (E, G) be a space equipped with a transformation group. Suppose that the follow-
ing relations hold:
(i) E is uncountable;
(ii) G acts transitively in E;
(iii) G acts freely in E, i.e., for any two points x E and y E, there exists at most one
g G such that g(x) = y.
Let be a nonzero -nite G-invariant (G-quasi-invariant) measure on E. Assuming that
card(E) is not measurable in the Ulam sense (see Appendix 1), prove that there exists a
G-invariant (G-quasi-invariant) measure
/
on E strictly extending .
For this purpose, consider two possible cases.
1. The conality of card(E) is equal to , i.e., E admits a representation in the form
E =X
n
: n < ,
where card(X
n
) < card(E) for all n < .
Check that X
n
,dom() for some n < and apply Marczewskis method to the G-invariant
-ideal generated by X
n
.
2. The conality of card(E) is greater than or equal to
1
.
In this case, consider a representation of E in the form
E =X
: <,
where the family of sets X
: < is as in Exercise 19. Preserving the notation of

Exercise 19 and taking into account the assumption on card(E), verify that there exists a
set [0, [ for which the associated almost G-invariant set
X() =X
:
is nonmeasurable with respect to . Then apply to X() Marczewskis method of extending
-nite G-invariant (G-quasi-invariant) measures.
In particular, let (G, ) be an arbitrary uncountable group equipped with a nonzero -nite
left invariant (left quasi-invariant) measure . If card(G) is not measurable in the Ulam
sense, then can be strictly extended to a left invariant (left quasi-invariant) measure
/
on G.
The result formulated in Exercise 20 is essentially due to Hulanicki [85] and Pkhakadze
[198]. It shows that the assertion of the strict extendability of any nonzero -nite left
invariant (left quasi-invariant) measure on an uncountable group does not contradict ZFC
theory.
At present, it is unknown whether the same assertion can be established without additional
set-theoretical hypotheses. In the sequel, we will consider an analogous statement and
prove it within ZFC for more concrete classes of uncountable groups (see Chapter 10).
Chapter 4
Quasi-invariant measures
Let E be an arbitrary set and let G be a group of transformations of E; in this case, we usu-
ally say that the pair (E, G) is a space equipped with a transformation group (see Chapters
2 and 3).
Let S be a -algebra of subsets of E. We recall that S is invariant with respect to G (or
invariant under G or, briey, G-invariant) if
(X S)(g G)(g(X) S).
Suppose that is a measure dened on a G-invariant -algebra S of subsets of E. We say
that is quasi-invariant with respect to G (or quasi-invariant under G or, briey, G-quasi-
invariant) if
(X S)(g G)((X) = 0 (g(X)) = 0).
It immediately follows from this denition that every G-invariant measure on E is neces-
sarily G-quasi-invariant. Obviously, the converse assertion is not true in general.
Similarly to invariant measures, quasi-invariant measures can be met in many elds of
mathematics and are objects of extensive investigations (see, for instance, [16], [80], [83],
[115], [117], [165], and [226]). In particular, quasi-invariant measures play an essential
role in various topics of mathematical analysis and probability theory.
In this chapter we will concisely examine certain general properties of -nite (primarily,
probability) quasi-invariant measures. We will recall several important examples of -
nite quasi-invariant measures. In addition, we will present some general constructions of
-nite quasi-invariant measures and demonstrate how those constructions lead to other
interesting examples of measures of this type.
If we are given a nonzero -nite G-quasi-invariant measure on E, then it is clear that
any -nite measure which is equivalent to turns out to be G-quasi-invariant. Indeed,
63
4
from the classical Radon-Nikodymtheorem (see, e.g., [16], [26], [56], [80], and [210]), we
have a representation
(X) =
X
f d (X dom()),
where f : E R is a strictly positive Radon-Nikodymderivative of with respect to . In
other words, f satises the equality
f = d/d.
This representation immediately yields the G-quasi-invariance of . Notice that among all
those measures which are equivalent to there always exists a probability measure.
In this context, an important problem concerning -nite invariant and probability quasi-
invariant measures arises. Its formulation is as follows:
Let E be a set, G be a group of transformations of E and let be a probability G-quasi-
invariant measure on E. Does there exist a nonzero -nite G-invariant measure on E
which is equivalent to ?
Another important problemarising in connection with nonzero -nite quasi-invariant (in-
variant) measures is the problem of their existence for a given space E equipped with a
transformation group G and with a G-invariant -algebra S of subsets of E. A particular
case which is of special interest is when we are given a group (G, ) endowed with the group
of all its left translations and with a left-invariant -algebra of its subsets. However, we
do not intend to discuss these two problems here (for more details, see [245], [246], and
[249]).
Let (G, ) be a Polish topological group. If, in addition, G is locally compact, then there
exists a nonzero -nite Haar measure on G invariant (hence, quasi-invariant) with respect
to the group of all left (right) translations of G.
Suppose now that G is not locally compact. Then the following question naturally arises.
Does there exist a nonzero -nite Borel measure on G quasi-invariant with respect to the
group of all left (right) translations of G?
As already mentioned in Chapter 3, the answer to this question is negative (in this con-
nection, see [226] and the references given therein). Moreover, as indicated in the same
chapter, a much stronger result can be obtained by starting with some simple geometrical
properties of -compact subsets of a non-locally compact Polish group G. For this purpose,
we need one auxiliary proposition.
Lemma 1. Let (G, ) be a topological group and let H and K be subsets of G. Suppose that
the following two conditions are fullled:
Quasi-invariant measures 65
1) H cannot be covered by a countable family of compact subsets of G;
2) K is -compact in G.
Then there exists an uncountable disjoint family consisting of left (right) H-translates of K.
Proof. Obviously, it is sufcient to consider only the case when the elements of H act in G
as left translations of G. As usual, let
1
denote the least uncountable ordinal number. We
are going to construct, by using the method of transnite recursion, a disjoint
1
-family
{K
: <
1
} consisting of some left H-translates of the given set K.
First of all, we put K
0
= K. Suppose now that for a nonzero ordinal number <
1
, the
partial disjoint family {K
: <} of left H-translates of K has already been constructed.

Let us dene
Z ={K
: <}.
Then, taking into account condition 2) and the inequality card() , we obtain that the
set Z is -compact in G. Consequently, the set Z
1
which is a homeomorphic image of Z
is -compact as well. Further, the set
Z Z
1
={x y
1
: x Z, y Z}
as the image of Z Z under the continuous mapping
(x, y) = x y
1
(x G, y G)
is also -compact in G. According to condition 1), we may write
H\ (Z Z
1
) = / 0.
Let h be an arbitrary element from H \ (Z Z
1
). Evidently, we have
hZ Z = / 0.
Let us put K
= hK. Then it is clear that the partial family {K
: } is disjoint and
consists of some left H-translates of K. Proceeding in such a manner, we are able to dene
the required disjoint
1
-family {K
: <
1
} of left H-translates of K. Thus, the proof
of Lemma 1 is complete.
This lemma immediately implies the following statement.
Theorem 1. Let (G, ) be a Polish topological group and let H be a subgroup of G which
cannot be covered by countably many compact subsets of G. Then there exists no nonzero
-nite left (right) H-quasi-invariant Borel measure on G.
In particular, if (G, ) is a non-locally compact Polish topological group, then there exists
no nonzero -nite left (right) G-quasi-invariant Borel measure on G.
Proof. Suppose to the contrary that there exists a nonzero -nite left (or, respectively,
right) H-quasi-invariant Borel measure on G. Since G is a Polish topological space,
must be a nonzero Radon measure. Consequently, we can nd a compact subset K of G
such that (K) > 0. Now, according to Lemma 1, there exists an uncountable disjoint
family of left (right) H-translates of K. Since is left (right) H-quasi-invariant, all sets
from the above-mentioned family have strictly positive -measure, which yields a contra-
diction with the fact that satises the countable chain condition since it is -nite. The
contradiction obtained nishes the proof of the rst part of Theorem 1.
The second part of Theorem 1 trivially follows from the remark that if a Polish topological
group G is not locally compact, then it is not -compact either; otherwise, in view of the
Baire classical theorem, at least one member of a countable covering of G by compact sets
must have nonempty interior, which easily implies the local compactness of G.
In connection with Theorem 1, it is reasonable to remark once more that some Polish non-
locally compact topological groups admit nonzero left (right) invariant Borel measures
(see [6] and [191]). But, since such measures are not -nite, they are of less interest for
purposes of mathematical analysis and probability theory.
We also wish to point out the fundamental result which is due to Mackey [165] concerning
the structure of those Borel subgroups of a Polish topological group that admit nonzero
-nite quasi-invariant Borel measures. Namely, let G be a Borel subgroup of some Polish
topological group (quite frequently, such a G is called a standard group). Then, according
to Mackeys theorem, the following two assertions are equivalent:
(i) there exists a nonzero -nite Borel measure on G which is quasi-invariant with respect
to the group of all left translations of G;
(ii) there exist a locally compact Polish topological group and a group isomorphism
: G ,
such that is simultaneously a Borel isomorphism between G and .
Moreover, if stands for some nonzero -nite Borel measure on G which is quasi-
invariant under the group of all left translations of G, then the group isomorphism can
be chosen so that the image-measure
() =
1
turns out to be equivalent to the left Haar measure on , that is the class of all ()-
measure zero sets coincides with the class of all -measure zero sets.
In particular, by starting with this deep and important result of Mackey, it is easy to demon-
strate that for any Polish topological group H, the following two relations are equivalent:
(1) H is locally compact;
(2) there exists a nonzero -nite Borel measure on H quasi-invariant with respect to the
group of all left translations of H.
For the proof of this equivalence, see [165] and [115]. Notice that the implication (1)
(2) is trivial, so the main difculty is concentrated in establishing the converse implication
(2) (1).
As mentioned earlier, the standard Borel measure on the real line R is an essentially unique
nonzero -nite measure dened on the -algebra of all Borel subsets of R and invariant
under the group of all translations of R. Actually, this fact is a very particular case of the
uniqueness theorem for a Haar measure (see Chapter 3). In the same Chapter 3 we have
pointed out that by using this uniqueness property, it is not difcult to describe the class of
all nonzero -nite R-quasi-invariant Borel measures on R (see Exercise 4).
Now, let us consider some natural operations which enable us to produce a new space
equipped with a probability quasi-invariant measure starting froman initial family of spaces
equipped with corresponding probability quasi-invariant measures.
The next simple statement concerns the behavior of the quasi-invariance property under the
operation of taking arbitrary products of probability quasi-invariant measures.
Theorem 2. Let {(E
i
, G
i
) : i I} be a family of spaces equipped with transformation
groups. Suppose that for each index i I, a probability G
i
-quasi-invariant measure
i
is
given on the set E
i
. Denote by E the product of the family of sets {E
i
: i I} and denote
by the product measure of the family {
i
: i I}, i.e., put
={
i
: i I}.
Also, let G
stand for the weak product group of the family of groups {G

i
: i I}.
Then the product measure is G
-quasi-invariant.
Proof. The statement can easily be established by using the classical Fubini theorem and
the corresponding details are left to the reader.
As an immediate application of Theorem 2, we have the following example.
Example 1. Let {
i
: i I} be a family of probability measures such that for any i I,
the measure
i
is dened on the Borel -algebra B(R) of R and is R-quasi-invariant. As
usual, we denote by R
(I)
the vector subspace of R
I
consisting of all those elements of R
I
whose supports are nite, i.e.,
R
(I)
={(x
i
)
iI
R
I
: card({i I : x
i
= 0}) <}.
Actually, R
(I)
coincides with the direct sum of a family of copies of R, whose cardinality
is equal to card(I).
Obviously, R
(I)
is everywhere dense in the topological vector space R
I
. According to
Theorem 2, the product measure of the family {
i
: i I} is dened on the canonical
cylindrical -subalgebra of B(R
I
) and is quasi-invariant with respect to the vector space
R
(I)
.
In particular, if card(I) card(N) = , then dom() coincides with B(R
I
) and, conse-
quently, turns out to be a Borel probability measure on the space R
I
, quasi-invariant
under R
(I)
.
Thus, we have a large class of Borel probability measures on the innite-dimensional Pol-
ish topological vector space R
N
, which are quasi-invariant with respect to the everywhere
dense vector subspace R
(N)
of R
N
. Notice that R
(N)
is in some sense a maximal vector
subspace of R
N
possessing the property described in Example 3 (in this connection, see
Exercise 5).
Having a Borel probability measure on R
N
quasi-invariant under an everywhere dense
vector subspace of R
N
, we can try to construct analogous measures for other innite-
dimensional topological vector spaces. In particular, the following question naturally
arises.
Does there exist for every separable Banach space (E, || ||), a Borel probability measure
on E such that is quasi-invariant with respect to some everywhere dense vector subspace
of E?
It turns out that the answer to this question is positive. In order to present the corresponding
result, we need one simple lemma from the theory of Banach spaces. First, let us recall that
l
1
usually denotes the classical Banach space consisting of all those sequences (x
n
)
nN

R
N
for which the inequality
nN
|x
n
| < +
is fullled. Of course, l
1
is equipped with its standard norm
||x|| =

nN
|x
n
| (x = (x
n
)
nN
l
1
).
The space l
1
contains R
(N)
and is separable. Moreover, it is the initial space in the class of
all innite-dimensional separable Banach spaces. More precisely, we have the following
auxiliary statement (cf. [56], [209]).
Lemma 2. Let (E, || ||) be an arbitrary separable Banach space. Then there exists a
continuous linear surjection : l
1
E.
Proof. Denote by B the closed unit ball in E, i.e., put
B ={e E : ||e|| 1}.
Since E is separable, we can choose a sequence of points
{e
n
: n N} B
which is everywhere dense in B. Now, take an arbitrary element
x = (x
n
)
nN
l
1
and dene
(x) =

nN
x
n
e
n
.
Notice that this denition of is correct because of the inequalities
nN
||x
n
e
n
||
nN
|x
n
| < +.
Also, it can easily be seen that is a linear continuous mapping acting from l
1
into E. It
remains to check that is a surjection. For this purpose, let us rst show that B (l
1
).
Let e be any point of B. We shall construct, by ordinary recursion, an injective innite
subfamily {e
n(k)
: k N} of the sequence {e
n
: n N} in such a way that
||e e
n(0)
/2
0
... e
n(k)
/2
k
|| 1/2
k+1
for each k N. Evidently, we can nd an element e
n(0)
satisfying the inequality
||e e
n(0)
|| 1/2.
Suppose now that for a natural number k, the partial family
{e
n(0)
, e
n(1)
, ... , e
n(k)
}
has already been dened. Then we may write
||2
k+1
(e e
n(0)
... e
n(k)
/2
k
)|| 1.
Since the sequence {e
n
: n N} is everywhere dense in B, we can choose an index n(k+1)
such that
n(k +1) > max{n(0), n(1), ..., n(k)}
and
||2
k+1
(e e
n(0)
... e
n(k)
/2
k
) e
n(k+1)
|| 1/2.
From the latter inequality we immediately obtain
||e e
n(0)
... e
n(k)
/2
k
e
n(k+1)
/2
k+1
|| 1/2
k+2
.
Proceeding in this manner, we come to the desired family {e
n(k)
: k N}. Now, let us put
y = (y
n
)
nN
R
N
,
where y
n
= 0 if n differs from all n(k) (k N), and y
n
= 1/2
k
if n = n(k) for some k N.
Then it is easy to see that
y = (y
n
)
nN
l
1
, (y) = e.
Consequently, we get the relation B (l
1
). Taking into account the linearity of , we
conclude that (l
1
) = E which nishes the proof of the lemma.
Theorem 3. Let (E, || ||) be an arbitrary separable Banach space. Then there exists a
Borel probability measure on E quasi-invariant with respect to some everywhere dense
vector subspace of E.
Proof. We start with a Borel probability measure on l
1
which is quasi-invariant with
respect to an everywhere dense vector subspace of l
1
. The existence of follows, e.g.,
from Exercise 6 where a much stronger result is formulated (see also Exercise 11 from
Chapter 3). Further, according to Lemma 2, there exists a linear continuous surjection
: l
1
E.
Let us put
(X) =(
1
(X)) (X B(E)).
Then, as we know, is a Borel probability measure on E since it is the image of under
. Moreover, it can easily be veried that if the original measure is quasi-invariant
with respect to an everywhere dense vector subspace U of l
1
, then the measure is quasi-
invariant with respect to the vector subspace V = (U) of E. Taking into account the fact
that is a continuous surjection, we see that V is everywhere dense in E which ends the
proof of Theorem 3.
In connection with the preceding result, we can pose the general question concerning the
existence of a Borel probability measure on a topological vector space, quasi-invariant
with respect to an everywhere dense vector subspace. More exactly, we can formulate the
following problem.
Problem. Give a characterization of all those topological vector spaces E for which there
exists at least one Borel probability measure quasi-invariant with respect to some every-
where dense vector subspace of E.
This problem still remains open for the class of all topological vector spaces. However,
for the class of all Banach spaces, we may assert that the above-mentioned problem is
completely solved.
Indeed, Theorem 3 states that every separable Banach space E admits a Borel probability
measure quasi-invariant with respect to an everywhere dense vector subspace of E. On the
other hand, if G is a nonseparable metrizable topological group whose topological weight
is not a real-valued measurable cardinal number (see Appendix 3), then G does not admit
a Borel probability measure quasi-invariant with respect to an everywhere dense subgroup
of G (see Exercise 2).
Thus, under the set-theoretical assumption that no cardinal number is real-valued measur-
able, we can infer that for a given Banach space E, these two assertions are equivalent:
1) E is a separable space;
2) there exists a Borel probability measure on E quasi-invariant with respect to some ev-
erywhere dense vector subspace of E.
Fortunately, it turns out that an additional set-theoretical assumption can be omitted here.
In this connection, see Exercise 7 in which the corresponding result is presented within the
theory ZFC.
Evidently, the problem posed above has a natural analogue for nonzero -nite invariant
Borel measures. Namely, we can formulate the following problem.
Problem. Give a characterization of all those topological vector spaces E for which there
exists at least one nonzero -nite Borel measure invariant with respect to some every-
where dense vector subspace of E.
This problem remains open, too, and we only want to recall the fact that there are innite-
dimensional topological vector spaces E which admit a nonzero -nite Borel measure
invariant under an everywhere dense vector subspace of E (see Example 3 from Chapter
3).
For a while, let us return to Lemma 2 and recall that in accordance with this lemma for every
separable Banach space E, there exists a linear continuous surjection : l
1
E. Also,
we know that there exists a nonzero -nite Borel measure on l
1
invariant with respect to
an everywhere dense vector subspace of l
1
(see Exercise 11 of Chapter 3). Unfortunately,
we cannot apply to and to an argument similar to the proof of Theorem 3 in order
to establish the existence of a nonzero -nite Borel measure on E invariant with respect
to an everywhere dense vector subspace of E. The real reason of this fact can easily be
explained. Indeed, simple examples show that a homomorphic image of a -nite measure
is not, in general, a -nite measure. So in this situation, an argument analogous to the
proof of Theorem 3 does not work.
Now, let us discuss once more an important geometric property of the classical Lebesgue
measure
n
n
, the so-called Steinhaus property (see Chapter 3).
Our goal is to show that this property completely characterizes all those measures which
are absolutely continuous with respect to
n
. We restrict our further considerations to the
case n = 1. A more general case of R
n
may be considered analogously (cf. also Exercise
10).
Let be a -nite Borel measure on the real line R. We shall say that has the Steinhaus
property if for any set X dom() with (X) > 0, the difference set
X X ={x x
: x X, x
X}
is a neighborhood of zero in R (see Chapter 3).
Below, we will give a characterization of all those -nite Borel measures on R which
have the Steinhaus property (see [179]).
First, let us describe one typical situation when the Steinhaus property fails to be true.
Theorem 4. Let be a -nite Borel measure on R satisfying the following condition:
there exist a Borel set X R with (X) > 0 and a sequence of real numbers {t
n
: n < }
such that
lim
n+
t
n
= 0, lim
n+
(X +t
n
) = 0.
Then does not possess the Steinhaus property.
Proof. We may assume, without loss of generality, that
0 < (X) < +
and that the following relations are satised:
(X +t
n
) < (X)/2
n
(n = 2, 3, 4, ...).
Let us dene
Y = X \ ({X +t
n
: 2 n <}).
Then we have the inequality
(Y) (X)/2 > 0.
Also, by virtue of the denition of Y, we get
Y (Y +t
n
) = / 0 (2 n <),
from which it immediately follows that
t
n
Y Y (2 n <).
Taking into account the equality lim
n+
t
n
= 0, we conclude that the difference set Y Y
is not a neighborhood of zero in R, so does not have the Steinhaus property.
In order to characterize all those measures on R which possess the Steinhaus property, we
need two auxiliary propositions.
Lemma 3. Let be a -nite Borel measure on R absolutely continuous with respect to
the Lebesgue measure . Then has the Steinhaus property.
This lemma is trivial by virtue of the Radon-Nikodym theorem and in view of the well-
known fact that has the Steinhaus property (see Chapter 3).
Lemma 4. Let be a -nite Borel measure on R singular with respect to . Denote by
X a Borel subset of R such that
(X) = 0, (R\ X) = 0,
and consider the function
f (t) = (X +t) (t R).
Then f (t) is equal to zero for -almost all points t R.
Proof. The argument is fairly standard. Consider the set
{(t, x) : x X +t} R
2
and denote by its characteristic function which is Borel and hence measurable with re-
spect to the product measure . Using the Fubini theorem, we may write
f (t)d(t) =
(X +t)d(t) =

(t, x)d(t)d(x) =

xX
(t)d(t))d(x) =
(X)d(x) = 0,
which immediately yields the required result.
Theorem 5. Let be a -nite Borel measure on R. The following two assertions are
equivalent:
(1) is absolutely continuous with respect to ;
(2) possesses the Steinhaus property.
Proof. The implication (1) (2) is a direct consequence of Lemma 3. Let us establish the
converse implication (2) (1).
Suppose that has the Steinhaus property but is not absolutely continuous with respect to
. Then can be represented in the form
=
1
+,
where
1
is a -nite Borel measure on R absolutely continuous with respect to and is
a nonzero -nite Borel measure on R which is singular with respect to , i.e., there exists
a Borel set X R such that
(X) > 0, (X) = 0, (R\ X) = 0.
By virtue of Lemma 4, the function
f (t) = (X +t) (t R)
is equal to zero for -almost all points t R. In particular, there exists a sequence {t
n
: n <
} R of points tending to zero such that f (t
n
) = 0 for any n <. Let us put
Z = X \ ({X +t
n
: n <}).
Then we get (Z) = (X) > 0. Further, since
(Z) =
1
(Z) +(Z) (Z),
we also have (Z) > 0. At the same time, in view of the denition of Z, it is easy to see
that
Z (Z +t
n
) = / 0 (n <),
thus it follows (by virtue of lim
n+
t
n
= 0) that does not possess the Steinhaus property,
contradicting our assumption. The contradiction obtained nishes the proof.
Example 2. A similar argument works for the left Haar measure on a Polish locally com-
pact topological group (G, ). We thus conclude that a -nite Borel measure on G has
the Steinhaus property if and only if is absolutely continuous with respect to the left Haar
measure on G (see Exercise 10 for this chapter). Moreover, the Steinhaus property im-
plies at once that if A and B are Borel subsets of (G, ) such that (A) > 0 and (B) > 0,
then the set
A B ={a b : a A, b B}
contains in itself a nonempty open subset of G (see Exercise 11 below). In this context, it
should be underlined that even in the case of the real line R and of the Lebesgue measure
on R, there are -measure zero sets C R whose difference set CC is a neighborhood
of 0. For more details and further related results, see Chapters 11 and 12.
As already mentioned in Chapter 3, the Steinhaus property admits an analogue in terms
of the Baire category for arbitrary topological groups. Namely, let (G, ) be a topological
group and let Z be a subset of G having the Baire property and not belonging to the family
of all rst category subsets of G. Then the set Z Z
1
contains a neighborhood of the neutral
element of G. This important fact also implies that if A and B are any two second category
sets in G, both having the Baire property, then A B contains a nonempty open subset of G.
EXERCISES
1. Let be a -compact locally compact topological group equipped with a Haar measure
and let
denote the completion of . Suppose also that G is some separable topological

group and let
: G
be a group homomorphism. By applying the Steinhaus property of , show that the follow-
ing two assertions are equivalent:
(a) is measurable with respect to
;
(b) is continuous.
Start with this equivalence and prove by using the Mackey result formulated in this chapter
and the classical Baire theorem on category that for any Polish topological group H, the
next two relations are equivalent:
(c) H is locally compact;
(d) there exists a Borel probability measure on H quasi-invariant with respect to the group
of all left translations of H.
2. A well-known theorem of topological measure theory states that any -nite Borel
measure given on a metric space E with non-real-valued measurable topological weight
is concentrated on a closed separable subset of E (see Appendix 3). By starting with this
theorem, establish the following result.
Let G be a metrizable topological group whose topological weight is not a real-valued
measurable cardinal. Suppose also that there exists a nonzero -nite Borel measure on G
quasi-invariant with respect to an everywhere dense (in G) group of left (right) translations
of G. Then G is a separable topological group. Consequently, if G is complete, then G is a
Polish topological group.
3. Let G be an arbitrary -compact locally compact topological group satisfying the in-
equality
card(G) > 2
c
and let be the left Haar measure on G. Then, by denition, is invariant with respect
to the group of all left translations of G. On the other hand, show that G is not a separable
topological space. In particular, suppose that
c =
1
, 2
c
=
2
,
and let ={0, 1}
2
, where {0, 1} is equipped with the discrete topology. Then is a com-
mutative (with respect to addition (modulo 2)) nonseparable compact topological group,
the weight of is equal to
2
, and according to the classical theorem of Ulam (see [238]),
2
is not a real-valued measurable cardinal.
Conclude from the said above that the assumption of the metrizability of a topological
group G is essential for the validity of the result formulated in Exercise 2.
4. Let
n
denote the standard Borel measure on the n-dimensional Euclidean space R
n
and let be an arbitrary nonzero -nite measure dened on the Borel -algebra of R
n
.
Demonstrate that these two assertions are equivalent:
(a) is quasi-invariant with respect to the group of all translations of R
n
;
(b) there exists a strictly positive Borel function : R
n
R such that
(Z) =
Z
(x)d
n
(x)
for each Borel subset Z of R
n
.
Prove an analogous result in a more general situation where the space R
n
is replaced by a
locally compact Polish topological group G and the standard Borel measure
n
is replaced
by the left Haar measure on G.
5
. Let G be a subgroup of the additive group R

N
, such that
G\ R
(N)
= / 0.
Prove that there exists a family {
n
: n N} of probability measures satisfying these two
conditions:
(a) for each natural number n, the measure
n
is dened on the Borel -algebra of the real
line R and is quasi-invariant with respect to the group of all translations of R;
(b) the product measure of the family {
n
: n N} is not quasi-invariant with respect to
G.
6. Consider the Banach space l
1
as a Borel subset of the Polish topological vector space
R
N
and show that there exists a family {
n
: n N} of probability measures satisfying
these two relations:
(a) for each natural number n, the measure
n
is dened on the Borel -algebra of R and
is quasi-invariant with respect to the group of all translations of R;
(b) for the product measure of the family {
n
: n N}, the inequality (l
1
) > 0 holds
true.
Deduce from this fact that there exists a Borel probability measure on the Banach space
l
1
, quasi-invariant with respect to the vector space R
(N)
which is everywhere dense in l
1
(cf. Exercise 11 from Chapter 3, where a much stronger result is formulated).
7. Let E be a nonseparable normed vector space, G denote the group of all translations of
E, and let H be an everywhere dense subgroup of G. Further, let B be an arbitrary ball in
E. Show that for any -nite H-quasi-invariant measure on E (not necessarily Borel),
the implication
B dom() (B) = 0
holds true. Infer from this fact that there exists no nonzero -nite Borel measure on E
quasi-invariant with respect to H.
Finally, conclude that for a Banach space F, the following two assertions are equivalent:
(a) F is separable;
(b) there exists a Borel probability measure on F quasi-invariant with respect to some
everywhere dense vector subspace of F.
8. Let be an arbitrary cardinal number. Prove that there exists a Borel probability measure
on the topological vector space R
, quasi-invariant with respect to the everywhere dense

vector subspace R
()
of R
. Also, check that if is strictly greater than the cardinality

continuum, then the space R
is not separable. Deduce from these results that for a non-

metrizable topological vector space F, the assertions (a) and (b) of the preceding exercise
can be non-equivalent.
9
. Let E be a normed vector space and let G be a vector subspace of E. We shall say
that G is admissible if it is everywhere dense in E and there exists a nonzero -nite Borel
measure on E quasi-invariant with respect to G.
Show that the following two assertions are equivalent:
(a) E is innite-dimensional;
(b) any admissible vector subspace of E is not minimal with respect to the inclusion rela-
tion.
In order to establish the equivalence (a) (b), utilize the fact that for any innite-
dimensional normed vector space F, there exists a linear functional h : F R which is
discontinuous at all points of F (cf. Exercise 7 from Chapter 18).
10. Generalize Theorem 5 of this chapter to the case of a Polish locally compact group
G equipped with its left Haar measure . In other words, show that for a -nite Borel
measure on G, these two assertions are equivalent:
(a) has the Steinhaus property;
(b) is absolutely continuous with respect to .
For this purpose, use an argument similar to the proof of Theorem 5 and apply the well-
known fact that if X is a Borel subset of G, then the equality (X) = 0 is equivalent to the
equality (X
1
) = 0 (see, e.g., [80] or [83]).
11. Let (G, ) be a -compact locally compact topological group and let denote the left
(right) Haar measure on G. By using the Steinhaus property, prove that for any two Borel
sets A G and B G with (A) > 0 and (B) > 0, the set
A B ={a b : a A, b B}
has nonempty interior.
Chapter 5
Measurability properties of real-valued functions
In this chapter, two concepts associated with measurability properties of various real-valued
functions are introduced and examined. These concepts are the notion of a relatively mea-
surable real-valued function with respect to a given class M of measures and the notion of
an absolutely nonmeasurable function with respect to M. Naturally, the usefulness of these
concepts will be illustrated below by a number of relevant examples (cf. [130]). Further,
a characterization of absolutely nonmeasurable functions will be established in the most
important case when the role of M is played by the class M
E
of all nonzero -nite dif-
fused measures on a given base set E. Also, it will be shown that the functions produced
by Vitalis classical partition of the real line R are relatively measurable with respect to the
class of all extensions of the Lebesgue measure on this line. In order to obtain the latter
result, Theorem 2 of Chapter 2 will be utilized.
We begin with the precise denitions of the above-mentioned concepts.
Let E be a set and let M be a class of measures on E. In general, we mean that the domains
of measures from M are various -algebras of subsets of E.
As usual, the real line R is assumed to be equipped with its Borel -algebra B(R), so we
permanently will be dealing with the canonical measurable space (R, B(R)).
We shall say that a function f : E R is relatively measurable with respect to M if there
exists at least one measure M such that f is measurable with respect to , i.e.,
(B B(R))( f
1
(B) dom()).
On the other hand, if there exists no measure M such that f is -measurable, then we
shall say that f is absolutely nonmeasurable with respect to the given class M.
Obviously, the introduced notion of a relatively measurable function generalizes the notion
of a real-valued measurable function with respect to a concrete measure on E. Indeed, it
we take the one-element class M ={}, then the relative measurability of a given function
DOI 10.1007/978-94-91216-36-7_ 2009 Atlantis Press/World Scientific
79
5,
f : E R with respect to M is equivalent to the ordinary denition of the -measurability
of f , and the absolute nonmeasurability of f with respect to the same M is equivalent to
the standard denition of the -nonmeasurability of f .
As soon as the notion of a relatively measurable (absolutely nonmeasurable) function is
introduced, it trivially becomes possible to dene an analogous notion of a relatively mea-
surable (absolutely nonmeasurable) subset of E.
A set X E is called relatively measurable with respect to a class M of measures on E
if the characteristic function
X
of X is relatively measurable with respect to M. On the
other hand, if there exists no measure M such that
X
is -measurable, then X is called
absolutely nonmeasurable with respect to M.
For instance, let E be a base set equipped with a group G of its transformations and let M
denote the class of all nonzero -nite G-invariant (or, more generally, G-quasi-invariant)
measures on E. If a set X E is absolutely nonmeasurable with respect to M, then we
briey say that X is G-absolutely nonmeasurable in E.
Let be a measure on E such that
(x E)({x} dom()).
As usual, we say that is diffused (or continuous) if it vanishes at all singletons in E, that
is ({x}) = 0 for each x E.
Example 1. For any set E, let M
E
be the class of all nonzero -nite diffused measures on
E. Let f : E R be a function and let for some t
0
R, the relation
card( f
1
(t
0
)) >
be satised, where denotes the rst innite cardinal number. In this case, we can assert
that f is relatively measurable with respect to M
E
. Indeed, it is not difcult to dene a
complete diffused probability measure on E such that
f
1
(t
0
) dom(), ( f
1
(t
0
)) = 1.
Consequently, for any set T R, we have ( f
1
(T)) = 1 if t
0
T, and ( f
1
(T)) = 0
if t
0
T. This circumstance, obviously, implies that f is measurable with respect to the
measure and hence is relatively measurable with respect to the given class M
E
.
We see, in particular, that if for an original set E the relation
card(E) > 2
= c
holds true, then every function f : E R is relatively measurable with respect to M
E
.
Indeed, in this case there always exists a point t
0
R for which we have card( f
1
(t
0
)) >.
Measurability properties of real-valued functions 81
Let F be a topological space all singletons of which belong to the Borel -algebra B(F).
We recall that F is a universal measure zero space if there exists no nonzero -nite dif-
fused Borel measure on F.
Nontrivial examples of such spaces are given in exercises for this chapter, namely, see
Exercises 15 and 16 where it is pointed out that some non-discrete Hausdorff universal
measure zero spaces can be of arbitrarily large cardinality.
We have already mentioned that the question is highly nontrivial whether or not there ex-
ist uncountable universal measure zero subspaces of the real line R. This question was
investigated from different points of view. By using delicate arguments, several famous
authors (Hausdorff, Luzin, Sierpi nski, Marczewski and others) have established that there
are uncountable universal measure zero subspaces of R. One classical construction of such
a subspace of R, due to Luzin and Sierpi nski, is presented in [113], [148], [150], and [172].
It is based on the canonical decomposition of a proper analytic subset of R into its Borel
components (constituents). There are also other interesting constructions of uncountable
universal measure zero subspaces of R (see, for instance, [79], [197], [250], and Appendix
1).
By taking into account Example 1 and using the notion of a universal measure zero space,
a characterization of absolutely nonmeasurable functions with respect to the class M
E
can
be obtained.
Theorem 1. A function f : E R is absolutely nonmeasurable with respect to M
E
if and
only if the following two conditions hold:
1) for each r R, the set f
1
(r) is at most countable;
2) the set ran( f ) is a universal measure zero subspace of R.
Proof. Suppose rst that f is absolutely nonmeasurable with respect to M
E
. Then the
argument given in Example 1 shows that condition 1) must be satised. Let us verify that
condition 2) is valid, too. Indeed, assuming that ran( f ) is not a universal measure zero
subset of R, consider some Borel diffused probability measure on ran( f ) and denote
S ={ f
1
(Z) : Z dom()}.
Obviously, S is a -algebra of subsets of E and the family of countable sets { f
1
(r) : r
ran( f )} forms a partition of E. We put
( f
1
(Z)) =(Z) (Z dom()).
In this manner, the probability measure on the -algebra S is well dened and, accord-
ing to its denition, the function f becomes -measurable. Clearly, the completion
of
is a diffused measure and f remains measurable with respect to
. However, this cir-

cumstance contradicts our assumption that f is absolutely nonmeasurable with respect to
the class M
E
. The contradiction obtained shows the necessity of conditions 1) and 2) for
the absolute nonmeasurability of f with respect to M
E
.
Now, assume that these two conditions are fullled for a given function f and let us es-
tablish that f is absolutely nonmeasurable with respect to M
E
. Suppose for a moment that
there exists a measure belonging to M
E
such that f is measurable with respect to . We
may assume, without loss of generality, that is a probability measure. Then, denoting by
B(ran( f )) the Borel -algebra of ran( f ) R, we may dene
(Z) = ( f
1
(Z)) (Z B(ran( f ))).
So, we get a Borel probability measure on the space ran( f ) R and we see that is
diffused in view of condition 1). But this circumstance contradicts condition 2) and the
obtained contradiction ends the proof of the theorem.
Remark 1. Let
1
denote, as usual, the least uncountable cardinal number and let E be
a set such that card(E) >
1
. It is impossible to establish within ZFC theory that there
exist functions f : E R absolutely nonmeasurable with respect to the class M
E
. Indeed,
assuming the Continuum Hypothesis (c =
1
), we have card(E) > c and, as we already
know, for such a set E there are no functions f : E R absolutely nonmeasurable with
respect to M
E
. Also, it is impossible to prove within ZFC theory the existence of functions
f : R R which are absolutely nonmeasurable with respect to the class M
R
. This fact
directly follows from Theorem 1 and the circumstance that there are models of set theory
in which the cardinality of the continuum c is strictly greater than
1
and the cardinality
of any universal measure zero subspace of R does not exceed
1
. Models of this sort were
rst constructed by Baumgartner and Laver (for more details, see [172] and the references
therein).
On the other hand, under some additional set-theoretical hypotheses, it is not difcult to
demonstrate that there exists a function
g : R R
absolutely nonmeasurable with respect to the class M
R
. For instance, the existence of such
a function follows from the existence of a Luzin subset of R whose cardinality is equal
to c (more information about Luzin sets can be found, e.g., in [40], [143], [148], [172],
[176], [192], and [222]; see also Chapter 12). The next example shows that, assuming
Martins Axiom, absolutely nonmeasurable functions can be found even among injective
homomorphisms of the additive group R into itself.
Example 2. We recall that X R is a generalized Luzin set if card(X) = c and for every
rst category set Y R, the inequality card(X Y) < c holds true. Under Martins Axiom,
there exists a generalized Luzin set L R which is simultaneously a vector space over the
eld Q of all rational numbers. The construction of such a set L is fairly standard and is
based on the method of transnite recursion (cf. Chapter 12). It is well known that every
Luzin set equipped with the induced topology is a universal measure zero space and, under
Martins Axiom, every generalized Luzin set is a universal measure zero space, too. In
particular, the above-mentioned set L is a universal measure zero subset of R. Let
g : R L
be some isomorphism between the two sets R and L which both are regarded as vector
spaces over Q. Then g can be considered as an injective group homomorphismfrom R into
R with ran(g) =L and therefore g is a nontrivial solution of the classical Cauchy functional
equation
(x +y) =(x) +(y) (x R, y R).
Furthermore, according to Theorem 1, g turns out to be absolutely nonmeasurable with
respect to the class M
R
.
In this context, let us remind that any nontrivial solution of the Cauchy functional equation
is necessarily nonmeasurable in the Lebesgue sense and does not possess the Baire property
(see Exercise 2). It should also be underlined that the latter fact does not need any additional
set-theoretical axioms which once again emphasizes the circumstance that the absolute
nonmeasurability of a function (acting from R into itself) is a much stronger property than
its nonmeasurability in the Lebesgue sense.
It is reasonable to recall in this place that the rst construction of a subset of R which
is nonmeasurable with respect to the Lebesgue measure on R, was given by Vitali in
his classical work [239]. Clearly, the characteristic function of a Vitali set yields the rst
example of a function acting from R into R and nonmeasurable in the Lebesgue sense.
Now, we would like to introduce other real-valued functions of Vitali type and examine
their measurability properties with respect to various classes of measures on R.
Let W be an equivalence relation on R all equivalence classes of which are at most count-
able.
We shall say that a mapping f : R R is a Vitali type function for W if (r, f (r)) W for
each r R and the set ran( f ) is a selector of the partition of R determined by W.
It can be shown that if V is the classical Vitali equivalence relation on R, i.e., the equiva-
lence relation dened by the formula
(x, y) V (x R & y R & x y Q),
then any Vitali type function for V is absolutely nonmeasurable with respect to the class of
all translation-invariant extensions of the Lebesgue measure on R (see Exercise 2 of this
chapter). However, the following somewhat surprising statement is valid.
Theorem 2. Let M() denote the class of all those measures on R which extend . Then
every Vitali type function for V is relatively measurable with respect to M().
Proof. Our argument is essentially based on a result discussed in Chapter 2, which states
that if E is any set, is a -nite measure on E and {X
n
: n < } is a disjoint count-
able family of subsets of E, then there always exists a measure
on E extending and
satisfying the relation
{X
n
: n <} dom(
).
We are going to apply this result to the Lebesgue measure and to a certain partition of R.
Let f : R R be any Vitali type function for V. We put
X = ran( f ).
The family of sets {X +q : q Q} forms a countable partition of R. Therefore, according
to the above-mentioned result, there exists a measure

X +q dom(
) for all q Q. We assert that f is measurable with respect to
. Indeed,
for each Borel subset B of R, the equality
f
1
(B) ={X B+q : q Q}
is easily veried. This equality can also be rewritten as
f
1
(B) ={(X +q) (B+q) : q Q}.
The right-hand side of the latter relation directly indicates that the set f
1
(B) is
-
measurable, which yields at once the measurability of f with respect to
and hence the

relative measurability of f with respect to the class M(). This conclusion ends the proof
of Theorem 2.
Example 3. Assuming Martins Axiom, it is not difcult to construct a generalized Luzin
set L R and an equivalence relation W RR such that:
1) (r R)(card(W(r)) =);
2) L is a selector of the partition {W(r) : r R} of R.
Let h : R R be a Vitali type function for W such that ran(h) = L. Then, in view of
Theorem 1, h is absolutely nonmeasurable with respect to the class M
R
and, consequently,
h is absolutely nonmeasurable with respect to the class M().
The previous example shows that the validity of Theorem 2 is essentially connected with
some special (in fact, group-theoretical) properties of the Vitali partition V.
Denote by M
() the class of all those measures on R which extend and are quasi-
invariant under the group of all translations of R. Recall that quasi-invariant measures were
dened in Chapter 4 and some of their general properties were also discussed therein. As
observed before the formulation of Theorem 2, any Vitali type function for V is absolutely
nonmeasurable with respect to the class of all translation-invariant extensions of . On the
other hand, it can be demonstrated that there exists a Vitali type function for V which is
relatively measurable with respect to M
() (see Theorem 3 below).

At the same time, we do not know an answer to the following question.
Problem. Does there exist a Vitali type function for V which is absolutely nonmeasurable
with respect to M
()?
In this context, it makes sense to consider in more details Vitalis classical construction
[239] of a non-Lebesgue measurable subset of the real line R.
Recall once more that the Vitali partition of R is determined by the following equivalence
relation V(x, y):
(x, y) V (x R & y R & x y Q).
Let Z be an arbitrary selector of this partition. It can be shown that Z is nonmeasurable with
respect to every measure on R extending and invariant under the group Q of all rational
translations of R (cf. [46], [80], [83], [108], [143], [176], [183], [192], [240], and Exercise
2 for this chapter). According to our terminology, this circumstance immediately implies
that Z is absolutely nonmeasurable with respect to the class of all translation-invariant
extensions of . For translation-quasi-invariant extensions of , the situation is essentially
different, which will be stated in the sequel.
Among the selectors of the Vitali partition (which usually are called Vitali sets), we can
encounter some subgroups of the additive group R. Indeed, if we treat R as a vector space
over the eld Q, then we may apply a well-known theoremfromlinear algebra which states
that a vector subspace Q of R admits a complemented subspace in R, i.e., we come to a
representation
R = Q+H (QH ={0}),
where H is some vector space over Q. Actually, H is a hyperplane in R complemented
to the line Q in R. Consequently, H is also a subgroup of R such that no translation-
invariant extension of can make H to be measurable. In addition, the relation
card(R/H) = card(Q) =
holds true, where the symbol R/H stands for the family of all translates of H.
There are several other constructions, which establish that there exist subgroups G of R
nonmeasurable with respect to but satisfying the inequality card(R/G) > . For in-
stance, by applying the method of transnite recursion, it is not difcult to construct a
subgroup G of R satisfying the above-mentioned inequality and the relation
(R\ G) = 0,
where
stands for the inner measure associated with . Constructions of such groups
GR can be done similarly to the classical Bernstein construction (cf. [148], [176], [192],
and Exercise 3). But the advantage of G over H is that G can be made measurable with
respect to an appropriate translation-invariant extension of . This fact is not accidental.
Its certain analogue remains valid in a more general situation, namely, in the case of an
uncountable commutative group instead of R. In this connection, see especially Chapter
10 where we give an application of one purely algebraic statement to the question of the
existence of nonmeasurable subgroups of uncountable commutative groups. In the same
chapter, we solve the problem of extending nonzero -nite invariant (quasi-invariant)
measures on such groups, by using their appropriate nonmeasurable subgroups.
Let us return to the Vitali partition of R and to Vitali sets of special type.
Let H R be as earlier, that is let H be a hyperplane in R complemented to Q, where
R is again regarded as a vector space over Q. We have already underlined that H, as
a particular case of a Vitali set, is absolutely nonmeasurable with respect to the class of
all translation-invariant extensions of . On the other hand, we can demonstrate that H
becomes measurable with respect to a suitable translation-quasi-invariant extension of .
The following statement contains a much stronger result which shows that there are many
possibilities for obtaining such translation-quasi-invariant extensions of .
Theorem 3. There exist continuumly many measures on R which extend , are quasi-
invariant under the group of all translations of R, and satisfy the relation H dom().
Proof. First, we would like to observe that H is -thick in R, i.e.,
(R\ H) = 0
which is true because H is an everywhere dense subgroup of R and H is not of -measure
zero. The details of checking the -thickness of H are left to the reader. Further, we
introduce the disjoint family of sets
{H
k
: k <} ={q +H : q Q}.
Obviously, for any h R, the family {h +H
k
: k < } coincides with the family {H
(k)
:
k <}, where is some permutation of . We thus derive that {H
k
: k <} is a countable
translation-invariant partition of R into -thick sets. Now, we introduce the class of sets
S ={
k<
(H
k
X
k
) : X
k
dom() f or all k <}.
It can easily be veried that S is a translation-invariant -algebra of subsets of R. Fix a
sequence {a
k
: k <} of strictly positive real numbers such that
k<
a
k
= 1.
Then take an arbitrary set
k<
(H
k
X
k
) from S and put
(
k<
(H
k
X
k
)) =

k<
a
k
(X
k
).
In this manner we dene a certain functional on S. Indeed, the denition of is correct
in viewof the -thickness of all sets H
k
. By reason of the same circumstance, the functional
is countably additive, so is a -nite measure on S. If X dom(), then
(X) =(
k<
(H
k
X)) =

k<
a
k
(X) =(X)
which shows that extends . Finally, if we have
(
k<
(H
k
X
k
)) =

k<
a
k
(X
k
) = 0,
then taking into account the inequalities a
k
> 0 for all k <, we get
(X
k
) = 0 (k <),
which implies for any h R that
(h +
k<
(H
k
X
k
)) =(
k<
(H
(k)
(h +X
k
))) =

k<
a
(k)
(h +X
k
) = 0,
where is again some permutation of (of course, depending on h). Therefore, the
measure is quasi-invariant under the group of all translations of R. Moreover, a similar
argument yields that is also quasi-invariant under all central symmetries of R and hence
is quasi-invariant under the group of all isometric transformations of R.
Finally, since depends on a choice of a sequence of strictly positive real numbers {a
k
: k <
} and there are continuumly many such sequences, we conclude that there exist at least
continuumly many pairwise distinct translation-quasi-invariant extensions of for which
the Vitali set H becomes measurable. Of course, it is important here that different choices
of {a
k
: k <} produce different extensions of . Theorem 3 has thus been proved.
Let E be a set equipped with a -nite measure and let f : E R be a function satisfying
the following condition:
there exists a Borel probability measure on ran( f ) such that the graph of f is a ( )-
thick subset of the product space E ran( f ).
Then, applying the standard argument (cf. [141] and Lemmas 1,2,3 of Chapter 2), it is not
hard to demonstrate that there exists a measure
on E such that:
(1)
extends ;
(2) f is measurable with respect to
.
In other words, f turns out to be relatively measurable with respect to the class M() of all
extensions of .
In particular, if E = R and a function f : R R has ( )-thick graph, then f turns
out to be measurable with respect to an appropriate extension of and, consequently, f is
relatively measurable with respect to the class M().
Notice that there are various examples of functions f : R R whose graphs are ( )-
thick subsets of the plane R
2
(see, for instance, [72], [148], [192], and Exercise 8 of Chapter
2). Moreover, the following much stronger statement is valid.
Theorem 4. There exists a function g : R R having the property that for any nonzero
-nite diffused Borel measure on R and for any -nite measure on R, the graph of
g is a ( )-thick subset of the plane R
2
.
Proof. In order to show the validity of this theorem, we start with a partition {X
t
: t R} of
R into Bernstein sets. Recall that a Bernstein set is any totally imperfect subset of R whose
complement is also totally imperfect. The existence of such a partition is well known (see,
e.g., [176], [192], and Exercise 5 for this chapter). Now, we dene the function g as follows:
for each x R, we put g(x) =t if and only if x X
t
.
It remains to check that g is the required function.
Let be an arbitrary nonzero -nite diffused Borel measure on R and let be an arbitrary
nonzero -nite measure on R. We must verify that the graph Gr(g) of g is ( )-thick
in R
2
. Indeed, if Z is any ( )-measurable set with ( )(Z) > 0, then by virtue of
the classical Fubini theorem, there exists a point t
0
R such that
({x R : (x, t
0
) Z}) > 0.
This circumstance implies that the set {x R : (x, t
0
) Z} contains a nonempty perfect
subset of R (because is a diffused Radon measure). Keeping in mind the fact that X
t
0
is
a Bernstein subset of R, we get
X
t
0
{x R : (x, t
0
) Z} = / 0,
(x)(g(x) =t
0
& (x, g(x)) Z).
Therefore, the set Gr(g) has nonempty intersection with Z and hence Gr(g) is ( )-thick
in R
2
which completes the proof of the theorem.
Notice that in our further considerations we will be dealing with some other versions of
Theorem 4 (see Chapter 18).
We would like to nish this chapter with introducing one more notion closely connected
with measurability properties of real-valued functions with respect to certain classes of
measures. It should be mentioned that for a special class of measures, this notion was rst
introduced and thoroughly considered by Marczewski.
Let E be a set and let M be a class of measures on E. We shall say that a function f : E R
is universally (or absolutely) measurable with respect to M if f turns out to be measurable
with respect to all measures from M.
In accordance with the above denition, we shall say that a set X E is universally (or
absolutely) measurable with respect to M if the characteristic function
X
of X turns out to
be universally measurable with respect to M.
The concept of universal measurability of sets and functions will be discussed more thor-
oughly in our further considerations. Here we restrict ourselves only to giving one typical
example from classical descriptive set theory.
Example 4. Let E be a Polish topological space and let M denote the class of the comple-
tions of all -nite Borel measures on E. Let A(E) stand for the class of all Suslin subsets
of E (see Chapter 8) and let (A(E)) denote the -algebra generated by this class. It is
well known that any set X (A(E)) is universally measurable with respect to M (see
[24], [26], [52], [99], [210], and Appendix 2). As shown by G odel and Novikov, an anal-
ogous fact fails to be true within ZFC theory for projective subsets of E of higher levels.
For more details, see [10], [91], [99], and [188].
EXERCISES
1. Verify that every Luzin subset of R is universal measure zero.
For this purpose, use the fact that any -nite diffused Borel measure on a separable metric
space E is concentrated on some rst category subset of E (in this connection, see Appendix
3).
2
. Let E be a set and let G be a group of transformations of E. We say that a set X E is

a G-selector if X has exactly one common point with each G-orbit in E.
Now, let G be a countable everywhere dense subgroup of the additive group R. By using an
argument similar to Vitalis [239], prove that every G-selector is absolutely nonmeasurable
with respect to the class of all G-invariant extensions of . In particular, all G-selectors are
nonmeasurable with respect to . Apply this result to the case G = Q.
Verify that any Vitali type function for the Vitali partition V of R is absolutely nonmeasur-
able with respect to the class of all translation-invariant extensions of .
Show also that any nontrivial solution f : R R of the Cauchy functional equation
(x) +(y) =(x +y) (x R, y R)
is nonmeasurable with respect to and does not have the Baire property (cf. Example 9
from Chapter 1).
3
. By using the method of transnite recursion, construct a subset G of the real line R that
satises the following three conditions:
(a) G is a vector space over Q (in particular, G is a subgroup of the additive group R);
(b) card(R/G) >;
(c)
(R\ G) = 0.
Show that any such set G is nonmeasurable with respect to but, simultaneously, is rel-
atively measurable with respect to the class of all translation-invariant extensions of .
4. Let G be an uncountable subgroup of the additive group R. Prove that every G-selector
is a G-absolutely negligible subset of R (see Chapter 2).
In particular, conclude from this fact that every G-selector is relatively measurable with
respect to the class of all G-invariant extensions of the Lebesgue measure .
5
. Applying the method of transnite recursion, construct a partition of R into continu-

umly many Bernstein sets. Moreover, prove the following general (purely set-theoretical)
statement which trivially implies the existence of such a partition of R.
Namely, let E be an arbitrary innite set and let {A
i
: i I} be a family of subsets of E
satisfying the relations
card(I) card(E), (i I)(card(A
i
) = card(E)).
Show that there exists a disjoint family {B
j
: j J} of subsets of E such that card(J) =
card(E) and
card(A
i
B
j
) = card(E)
for all indices i I and for all indices j J.
In particular, take:
E = R;
card(I) = card(R) = c;
{A
i
: i I} = the family of all nonempty perfect subsets of R.
Then the corresponding family {B
j
: j J} produces the required partition of R into con-
tinuumly many Bernstein sets.
6
. Consider R as a vector space over the eld Q. Prove that there exists a vector subspace
of R which simultaneously is a Bernstein set in R.
Also, assuming Martins Axiom, show that there exists a vector subspace of R which si-
multaneously is a generalized Luzin set in R.
7. By using Martins Axiom, give a precise construction of the set L described in Example
3.
8. Assuming Martins Axiom, prove that every generalized Luzin subset of R is universal
measure zero.
9
. Let E be a set and let G be a group of transformations of E. We say that a set Z E is

G-negligible in E if the following two conditions hold:
(a) Z is relatively measurable with respect to the class M of all nonzero -nite G-invariant
measures on E;
(b) for any measure M, we have Z dom() (Z) = 0.
Observe that every G-absolutely negligible set in E is also G-negligible.
Consider the case E = G = R
2
and applying the method of transnite induction, give an
example of a function f : R R whose graph is G-negligible but is not G-absolutely
negligible.
10. Denote by M the class of all those translation-invariant measures on R
2
which extend
the two-dimensional Lebesgue measure
2
. Let f be an arbitrary function acting from R
into R. Show that the graph Gr( f ) of f is relatively measurable with respect to M.
Moreover, prove that there exists a translation-invariant measure on R
2
which extends
2
and for which the graphs of all functions acting from R into R are -measurable.
For this purpose, use the Marczewski method of extending invariant measures described in
Chapter 2.
11. Let E be a set and let be a measure on E. Denote by M() the class of all those
measures on E which extend . Let f : E R be a function whose range is countable.
Show that f is relatively measurable with respect to M().
For this purpose, apply Theorem 2 from Chapter 2.
12
. Denote by M
R
the class of the completions of all nonzero -nite diffused Borel
measures on R. Let f : R R be a function. Show that the following two assertions are
equivalent:
(a) f is relatively measurable with respect to M
R
;
(b) f admits a representation in the form f =gh, where a function g : RR is Lebesgue
measurable and a function h : R R is a Borel isomorphism of R onto itself.
Also, verify that there exist functions belonging to the class M
R
which are not measurable
in the Lebesgue sense.
13. Let R
n
and R
m
be two Euclidean spaces such that n = 0, m = 0, n = m and let
f : R
n
R
m
be an isomorphism between the additive groups R
n
and R
m
. The existence of such an
isomorphism can be established by using Hamel bases of R
n
and R
m
. Let B
n
(respectively,
B
m
) denote the unit ball in the space R
n
(respectively, in the space R
m
). Show that:
(a) the set f (B
n
) is nonmeasurable with respect to
m
;
(b) the set f
1
(B
m
) is nonmeasurable with respect to
n
.
Infer from these facts that neither f nor f
1
are Lebesgue measurable.
14. Let E be a set whose cardinality is not real-valued measurable. Equip E with the
discrete topology. Then E becomes a locally compact topological space. Denote by E
the
Alexandrov compactication of E (see, e.g., [58] and [101]). Check that E
is a compact
non-discrete universal measure zero space.
15
. Let a be an arbitrary innite cardinal number, I be a set with card(I) = a and let
X
i
={0, 1} for any i I. We equip each set X
i
(i I) with its discrete topology. Further, in
the product set
X(a) =
{X
i
: i I}
consider the family B of all those sets Y which admit a representation
Y =
{Y
i
: i I},
where Y
i
X
i
for any i I and card({i I : Y
i
= X
i
}) < a.
Verify that B is a base of some topology T on the set X(a) and show that T satises the
following relations:
(a) (X(a), T ) is an isodyne topological space, i.e., for every nonempty open set U X(a),
we have
card(U) = card(X(a)) = 2
a
;
(b) if a is a regular cardinal number, then X(a) is an a-Baire space, i.e., no nonempty open
set U X(a) can be represented in the form
U ={Z
j
: j J},
where card(J) a and all sets Z
j
( j J) are of rst category in X(a);
(c) if a
b
= a for any nonzero cardinal b < a, then the topological weight w(X(a)) of X(a)
is equal to a.
Suppose, in addition, that the Generalized Continuum Hypothesis holds and a is a regular
cardinal number nonmeasurable in the Ulam sense (see Appendix 1). Show that X(a) is a
universal measure zero space, which means that no nonzero -nite diffused measure can
be dened on the Borel -algebra of X(a).
16
. Assume that the Generalized Continuum Hypothesis holds and let a be an innite
regular cardinal number. Consider once again the space X(a) of the previous exercise.
Prove that there exists a subset L(a) of X(a) satisfying the following three relations:
(a) for every nonempty open set U X(a), the equality
card(U L(a)) = 2
a
is valid and, in particular, L(a) is an isodyne everywhere dense subspace of X(a);
(b) if a set Z X(a) is representable in the form
Z ={Z
j
: j J},
where card(J) a and all sets Z
j
( j J) are of rst category in X(a), then card(ZL(a))
a;
(c) if a is not measurable in the Ulam sense, then L(a) is a universal measure zero space.
Also, check that the set L() can be regarded as a certain Luzin subset of the standard
Cantor space X() ={0, 1}
.
17. Prove that the following two assertions are equivalent:
(a) there exists a universal measure zero set X R with card(X) = c;
(b) any partial function f : R R whose range is universal measure zero and for which all
pre-images f
1
(t) (t R) are at most countable, can be extended to a function f
: R R
absolutely nonmeasurable with respect to the class M
R
.
18. Let E be a topological space in which all singletons are Borel subsets of E. Show that
the following two assertions are equivalent:
(a) E is universal measure zero;
(b) every subset of E is absolutely nonmeasurable with respect to the class M
E
of the
completions of all nonzero -nite diffused Borel measures on E.
Let us introduce the notation
Z
0
(E) ={Z B(E) : Z is universal measure zero}.
Let X be a subset of E. Verify that the following two assertions are equivalent:
(c) X is absolutely nonmeasurable with respect to the class M
E
;
(d) for any set Y B(E) \ Z
0
(E), the relations
Y X = / 0, Y (E \ X) = / 0
hold true.
In particular, suppose that the conditions
card(B(E) \ Z
0
(E)) card(E),
(Y B(E) \ Z
0
(E))(card(Y) = card(E))
are satised. Prove that there exists a subset of E absolutely nonmeasurable with respect to
the class M
E
.
Compare this exercise with the standard construction of Bernstein sets in uncountable Pol-
ish topological spaces.
19
. Equip Rwith the Sorgenfrey topology T whose base consists of all half-open intervals
of the form [a, b[, where a R and b R (see [58] or [101]). Check that (R, T ) is a
hereditarily Lindel of topological space whose Borel -algebra coincides with the standard
Borel -algebra B(R). Infer from this fact that any Bernstein subset of R is absolutely
nonmeasurable in the space (R, T ), with respect to the class of the completions of all
nonzero -nite diffused Borel measures on (R, T ).
Further, in the product space (R
2
, T T ) consider the set
D ={(x, y) R
2
: x +y = 1}.
Check that Dis a closed discrete subset of (R
2
, T T ). Conclude that if c is measurable in
the Ulam sense, then this product space does not contain an absolutely nonmeasurable set
with respect to the class of the completions of all nonzero -nite diffused Borel measures
on (R
2
, T T ).
20
. Assume Martins Axiomand let E be a topological space satisfying the following three
conditions:
(a) E has a countable base;
(b) all singletons in E are Borel subsets of E;
(c) every closed set in E is of type G
.
Prove that there exists a subset of E absolutely nonmeasurable with respect to the class M
E
.
More precisely, establish the same result only supposing that all those subsets of E whose
cardinalities are strictly less than c are universal measure zero.
21
. Assume Martins Axiom. Let E be a topological space such that:

(a) card(E) c;
(b) E is metrizable.
Show that there exists a subset of E absolutely nonmeasurable with respect to the class M
E
.
Conclude the same result for any topological space which is Borel isomorphic to E.
For this purpose, apply Theorem 1 from Appendix 3.
22. As shown by Roslanowski and Shelah [208], there exists a model of set theory in which
for any function f : RR, there is a set X R such that
(X) >0 and f |X is continuous.

Verify that in this model there are no absolutely nonmeasurable functions acting from R
into R.
23. Suppose that every uncountable co-analytic subset of a Polish topological space con-
tains a nonempty perfect set. As is well known, this assumption does not contradict ZFC
theory (see [10], [91], and [99]).
Let E be an uncountable Polish space and let M denote the class of the completions of all
-nite Borel measures on E. Consider any analytic set A E which is not Borel in E.
The existence of such an A is also well known (see, e.g., [99], [148], [150], [160], [162],
and Appendix 6).
Check that A cannot be represented in the form A = BX, where B B(E) and X is a
universal measure zero set in E.
Conclude from this fact that the -algebra of all those subsets of E which are universally
measurable with respect to M is not generated by the Borel -algebra B(E) and the -ideal
of all universal measure zero subsets of E.
In connection with this exercise, let us also recall that, under the Constructibility Axiom
of G odel, there exists an uncountable co-analytic subset of R which does not contain a
nonempty perfect set (see [10], [91], and [188]).
Chapter 6
Some properties of step-functions connected with
extensions of measures
In this chapter we again will be concerned with the measure extension problem and will
present one more application of Theorem 2 from Chapter 2. Primarily, we will be dealing
here with certain measurability properties of the simplest real-valued functions - the so-
called real-valued step-functions.
Let E be a nonempty base set and let f be a function acting from E into the real line R. We
recall that f is a step-function if the range of f is (at most) countable, i.e., we have
card(ran( f )) .
Clearly, every step-function f : E R produces a countable partition
{X
i
: i I} = { f
1
(t) : t ran( f )}
of E. We shall say that this partition is canonically associated with f .
Conversely, let {X
i
: i I} be an arbitrary countable partition of E. We shall say that a
step-function f : E R is associated with this partition if the following two relations are
satised:
(a) the restriction of f to any set X
i
is constant;
(b) the restriction of f to any selector of {X
i
: i I} is an injection.
It is well known that real-valued step-functions, which have some additional properties,
play an important role in different topics of mathematical analysis, especially in those
questions which are connected with various kinds of approximation. For instance, if E is
equipped with a -nite measure , then -measurable step-functions are needed for intro-
ducing the class of -integrable real-valued functions. Also, step-functions are essentially
used in some questions concerning the sup-measurability of functions of two variables (see,
for example, [120]) and in many other topics of real analysis and measure theory.
Let E be equipped with a -nite measure and let
f : E R
97
6
be a -measurable step-function. It can easily be veried that, for any subset T of R,
the pre-image f
1
(T) is a -measurable subset of E. It is natural to conjecture that this
property is closely connected with the notion of a step-function. Indeed, in the sequel it will
be shown that a similar measurability property enables us to characterize step-functions in
terms of extensions of measures (cf. [128]).
We shall say that a function f : E R is strongly measurable with respect to if for any
T R, the pre-image f
1
(T) is a -measurable subset of E.
It can readily be veried that for any step-function f : E R, the following three relations
are equivalent:
(1) f is measurable with respect to ;
(2) f is strongly measurable with respect to ;
(3) for each t R, the pre-image f
1
(t) is measurable with respect to .
For our further purposes, we need two auxiliary statements both of which have already
been considered in preceding chapters. We recall their formulations here.
Lemma 1. Let be a -nite measure on E and let {X
i
: i I} be an arbitrary disjoint
family of subsets of E. Then there exists a measure
on E extending and satisfying the

relation
{X
i
: i I} dom(
).
In particular, if the given family {X
i
: i I} is a countable partition of E, then every step-
function associated with {X
i
: i I} becomes measurable with respect to the extension
.
The lemma just formulated is a direct consequence of Theorem 2 from Chapter 2.
We also recall that a subset Y of R is universal measure zero if for any nonzero -nite
diffused Borel measure on R, we have
(Y) = 0 where
denotes the outer measure

associated with .
The next classical result is well known in descriptive set theory.
Lemma 2. There are uncountable universal measure zero subsets of R.
For the proof of Lemma 2, see Appendix 1.
Notice once more that constructions of uncountable universal measure zero subsets of R
and of other small subsets of R were presented by various authors and different ideas were
used in those constructions: the existence of a canonical decomposition of a proper analytic
set into its Borel components (constituents), Marczewskis characteristic function, Ulams
Some properties of step-functions connected with extensions of measures 99
transnite matrix, Fubini type argument, and other approaches. For more details about this
topic, see especially [79], [162], [172], [197], and [250].
Theorem 1. Let E be an uncountable set and let f : E R be a function. The following
two relations are equivalent:
(1) f is a step-function;
(2) for any -nite measure on E, there exists a measure
on E extending and such

that f is strongly measurable with respect to
.
Proof. (1) (2). Assume that (1) is valid and consider an arbitrary -nite measure on
E. Let {Y
i
: i I} denote the countable partition of E produced by f . According to Lemma
1, there exists an extension
of such that
{Y
i
: i I} dom(
).
Obviously, f is strongly measurable with respect to
, so (2) is true.
(2) (1). Assume that (2) is valid and let us show that f is a step-function. Suppose
otherwise, that is
card(ran( f ))
1
.
Clearly, we can nd a set X E such that f |X is an injection and
card(ran( f |X)) =
1
.
Consider a complete diffused probability measure on E concentrated on the set X, which
means that
(E \ X) = 0.
The existence of such a is obvious. By virtue of (2), there exists an extension
of such
that f becomes strongly measurable with respect to
. Now, for any set Z ran( f |X), let

us put
(Z) =
( f
1
(Z)).
So we get a diffused probability measure which is dened on the family of all subsets
of ran( f |X). From the existence of we readily deduce that there is no universal measure
zero subset of R whose cardinality is greater than or equal to
1
. But this circumstance
obviously contradicts Lemma 2.
Under some additional set-theoretical assumptions, Theorem 1 can be essentially strength-
ened. For instance, let us consider the following set-theoretical assertion.
(*) Any uncountable subset of R contains an uncountable universal measure zero set.
It can easily be seen that (*) is implied by the conjunction of Martins Axiom with the
negation of the Continuum Hypothesis, so (*) is consistent with ZFC theory. On the other
hand, the existence of a Sierpi nski subset of R readily implies that (*) is false (see Exercise
2 below; extensive information on Sierpi nski sets may be found in [148], [172], [176], and
[192]). So, the negation of (*) is also consistent with ZFC. We thus conclude that assertion
(*) is independent of ZFC theory.
The next statement is valid (cf. Theorem 1).
Theorem 2. Suppose that assertion (*) holds true. Let E be an uncountable set and let
f : E R be a function. The following two relations are equivalent:
(1) f is a step-function;
(2) for any -nite measure on E, there exists a measure
on E extending and such

that f is measurable with respect to
.
Proof. The argument is very similar to the proof of Theorem 1. The implication (1) (2)
does not need an additional set-theoretical assumption and can be established in the same
manner as before.
Assume now that (2) is satised and let us show that f is a step-function. Again, suppose
otherwise, i.e., suppose that
card(ran( f ))
1
.
According to (*), there exists an uncountable universal measure zero subset Y of ran( f ).
Clearly, we can nd a set X E such that ran( f |X) = Y and the restriction f |X is an
injective mapping. Consider an arbitrary complete diffused probability measure on E
which is concentrated on the set X. By virtue of (2), there exists an extension
of such
that f becomes
-measurable. Now, for every Borel subset Z of the space Y, let us put
(Z) =
( f
1
(Z)).
A straightforward verication shows that turns out to be a Borel diffused probability
measure on Y, so we obtain a contradiction with the fact that Y is a universal measure zero
subset of R which ends the proof of the theorem.
In particular, Theorem 2 implies that it is impossible to dene, within ZFC theory, a non-
step-function f : E R having the measurability property (2).
Now, let us consider step-functions for those nonzero -nite measures on E which are
invariant (or, more generally, quasi-invariant) with respect to an uncountable group G of
transformations of E.
If we suppose that G acts freely in E, then the situation differs from the case of ordinary
(i.e., non-invariant) measures.
To see this circumstance, take an arbitrary uncountable set E with
c f (card(E)) =.
It is reasonable to recall here that the symbol c f (card(E)) denotes the conality of card(E)
or, in other words, the least cardinal b such that there exists a partition {X
j
: j J} of E
into sets X
j
( j J) where card(J) = b and all card(X
j
) are strictly less than card(E).
Let G be a group of transformations of E acting freely in E and such that
card(E) = card(G).
Fix a countable partition {X
i
: i I} of E satisfying the relation
(i I)(card(X
i
) < card(E)).
Let f : E R be a step-function associated with this partition. It is easy to verify that f
cannot be measurable with respect to a nonzero -nite G-quasi-invariant measure on E.
To give deeper examples of this kind, let us recall two denitions (see Chapters 2 and 5).
Let E be a set and let G be a group of transformations of E.
We say that a set X E is G-absolutely negligible if for every -nite G-invariant (G-
quasi-invariant) measure on E, there exists a G-invariant (G-quasi-invariant) extension
of such that
(X) = 0.
We say that a set Y E is G-absolutely nonmeasurable if for any nonzero -nite G-quasi-
invariant measure on E, we have Y dom().
Some properties of G-absolutely negligible and G-absolutely nonmeasurable subsets of E
are discussed in [108], [115], and in Chapters 2, 10 and 11.
In a particular case, where E is an uncountable commutative group (identied with the
group G of all its translations), the following statement is valid.
Lemma 3. Let (G, +) be an arbitrary uncountable commutative group. Then these two
relations are satised:
(1) there exists a G-absolutely nonmeasurable subset of G;
(2) there exists a countable partition of G into G-absolutely negligible sets.
The proof of (1) is given in [107] and [119]. For the proof of (2), see [115], [119], or
Chapter 10 where an analogous statement is established in a more general case, namely, for
all uncountable solvable groups (G, ).
From Lemma 3 we directly get two nontrivial examples.
Example 1. Let X be a G-absolutely nonmeasurable subset of an uncountable commutative
group (G, +) and let f =
X
denote the characteristic function of X which trivially is a step-
function. Then f is nonmeasurable with respect to any nonzero -nite G-quasi-invariant
measure on G.
In other words, Example 1 states that there are two-valued functions absolutely nonmea-
surable with respect to the class of all nonzero -nite G-quasi-invariant measures on G.
Example 2. For an uncountable commutative group (G, +), consider its countable partition
{Y
i
: i I} into G-absolutely negligible sets. Let
f : G R
denote an arbitrary step-function associated with this partition. Then we have:
(1) f is nonmeasurable with respect to any nonzero -nite G-quasi-invariant measure on
G;
(2) for each t R, the set f
1
(t) is G-absolutely negligible.
Indeed, (2) is obvious. To see (1), let us suppose that f is measurable with respect to some
nonzero -nite G-quasi-invariant measure . Then all sets Y
i
which are the pre-images
of certain singletons in R must be -measurable. Since all of them are also G-absolutely
negligible, we must have (Y
i
) = 0. From this circumstance it follows that
(G) = ({Y
i
: i I}) =
{(Y
i
) : i I} = 0,
which yields a contradiction.
In other words, Example 2 states that there exist step-functions f on an uncountable com-
mutative group (G, +), which are absolutely nonmeasurable with respect to the class
of all nonzero -nite G-quasi-invariant measures on G, but each of the pre-images
f
1
(t) (t R) is good for extending any -nite G-invariant (G-quasi-invariant) measure
on G.
However, there are certain types of step-functions which also are good for obtaining in-
variant (respectively, quasi-invariant) extensions of -nite invariant (respectively, quasi-
invariant) measures. To describe such functions, let us return to the general situation when
an innite set E is given with some group G of its transformations. For our further pur-
poses, the notion of an almost G-invariant subset of E turns out to be helpful. It should
be mentioned that this notion was rst considered by Marczewski in connection with the
algebraic aspect of the measure extension problem (cf. [234] and [235]). Later on, it was
recognized that almost G-invariant sets play an important role in various topics of the the-
ory of invariant and quasi-invariant measures (see [81], [83], [85], [95], [108], [115], and
[198]).
There are two denitions of almost invariant sets, which are rather similar to each other (cf.
Chapter 2).
A set Z E is called almost G-invariant in E (in the set-theoretical sense) if
card(g(Z)Z) < card(E)
for each transformation g G.
If is a measure on E, then a set Z E is called almost G-invariant with respect to if
(g(Z)Z) = 0
for each transformation g G (where
as usual denotes the outer measure associated

with ).
Notice that if a set Z E is almost G-invariant with respect to , then any -measurable
hull of Z is also almost G-invariant with respect to .
The next lemma is probably well known (cf. [83], [85], [95], [108], [192], and [198]) but
for the sake of completeness, we present its short proof here.
Lemma 4. Let E be an uncountable set, G be a group of transformations of E with
card(G) card(E), and let I be a nonempty countable set. Then there exists a partition
{X
i
: i I} of E consisting of almost G-invariant subsets of E such that card(X
i
) = card(E)
for all i I.
Proof. We may assume, without loss of generality, that card(G) = card(E) and G acts
transitively in E. Let denote the least ordinal for which
card() = card(E)
and let x be a xed point of E. An increasing (by inclusion) transnite sequence {G
: <
} of subgroups of G can readily be constructed satisfying the following conditions:
(1) {G
: <} = G;
(2) card(G
) card() + for any <;

(3) G
(x) \ {G
(x) : <} = / 0 for any <.

Now, let {
i
: i I} be a partition of [0, [ such that
(i I)(card(
i
) = card()).
Putting
X
i
= {(G
(x) \ {G
(x) : <}) :
i
} (i I),
we come to the required partition {X
i
: i I} of E. The details of checking this fact are left
to the reader (cf. Exercise 19 for Chapter 3).
Lemma 5. Let be a -nite G-quasi-invariant measure on E and let f : E R be a
step-function such that for any t R, the set f
1
(t) is almost G-invariant with respect to
. Then for any g G, the functions f and f g are equivalent with respect to , that is
the equality
({x E : f (x) = ( f g)(x)}) = 0

holds true.
Proof. Denote by {X
i
: i I} the countable partition of E associated with the given step-
function f . It is clear that
{x E : f (x) = ( f g)(x)} {X
i
g
1
(X
j
) : i I, j I, i = j}.
Since the relations
(g
1
(X
j
)X
j
) = 0, X
i
X
j
= / 0 (i I, j I, i = j)
are satised, we must have
(X
i
g
1
(X
j
)) = 0.
Fromthis fact, taking into account the countability of I, we immediately obtain the required
result.
Theorem 3. Let E be a set with c f (card(E)) > , let G be a group of transformations
of E which acts freely in E and whose cardinality is equal to card(E), and let {X
i
: i I}
be a countable partition of E into almost G-invariant sets. Denote by f : E R any
step-function associated with this partition. Then for every -nite G-invariant (G-quasi-
invariant) measure on E, there exists a G-invariant (G-quasi-invariant) measure
on E
such that:
(1)
extends ;
.
Proof. The argument presented below is rather similar to the proof of Theorem 2 from
Chapter 2. However, there is also an essential difference implied by the circumstance that
here we are dealing with invariant (quasi-invariant) measures.
Since I is countable, we may suppose that either I = {1, 2, ..., n} or I = . For any i I,
denote by t
i
the value of f at some point of X
i
.
Let be an arbitrary -nite G-invariant (G-quasi-invariant) measure on E. Since
c f (card(E)) > , we may assume, without loss of generality, that every set Z E with
card(Z) < card(E) belongs to the domain of and, consequently, (Z) = 0. Actually,
in this situation the Marczewski method of extending invariant (quasi-invariant) measures
works for the G-invariant -ideal consisting of the above-mentioned sets Z.
Therefore, all sets X
i
(i I) become almost G-invariant with respect to .
For each index i I, denote by Y
i
a -measurable hull of X
i
and dene
Z
i
=Y
i
\ {Y
j
: j < i}.
Notice that all sets Z
i
are pairwise disjoint, -measurable and almost G-invariant with
respect to . Moreover, we have the equality
E = {Z
i
: i I}.
Let f
: E R be a step-function whose value on every nonempty set Z

i
is equal to t
i
. As
established in Chapter 2,
({x E : f (x) = f
(x)}) = 0,
where
stands for the inner measure associated with . But here we need a much stronger
property of the set
D = {x E : f (x) = f
(x)}.
Namely, we must show that for any countable family {g
k
: k <} G, the equality
({g
k
(D) : k <}) = 0
holds true. Indeed, the inclusion
{g
k
(D) : k <} D({{x E : ( f g
1
k
)(x) = f (x)} : k <})
({{x E : ( f
g
1
k
)(x) = f
(x)} : k <})
is readily veried. Taking into account the relation
(D) = 0 and applying Lemma 5, we

claim that
({g
k
(D) : k <}) = 0.
Thus, the set D generates a G-invariant -ideal of subsets of E, all members of which have
inner -measure zero. This circumstance enables us to extend the given measure to a
G-invariant (G-quasi-invariant) measure
on E such that
(D) =0 (cf. Chapter 2). Since

the function f
is -measurable, it is also
-measurable. In view of the equality

D = {x E : f (x) = f
(x)},
we conclude that f is
-measurable, too.
Remark 1. Actually, the preceding argument shows that if we have an arbitrary -nite
G-invariant (G-quasi-invariant) measure on E and a step-function f : E R such that
all pre-images f
1
(t) (t R) are almost G-invariant with respect to , then there exists a
G-invariant (G-quasi-invariant) extension
of for which f becomes
-measurable.
In connection with Theorems 1 and 2, the following question arises.
Does there exist a real-valued function f
1
: E R with card(ran( f
1
)) > such that every
-nite G-invariant (G-quasi-invariant) measure on E admits a G-invariant (G-quasi-
invariant) extension for which f
1
becomes measurable?
In order to give a positive answer to this question (in some natural situations), we need the
next auxiliary proposition.
Lemma 6. Let E be a set and let G be an uncountable group of transformations of E acting
freely in E. Then there exists an uncountable G-absolutely negligible subset X of E.
Proof. If card(G) >
1
, then any set X E with card(X) =
1
is G-absolutely negligible
in E (the checking this simple fact is left to the reader).
Suppose now that card(G) =
1
and x a point x E. Let {G
: <
1
} be an increasing
(by inclusion)
1
-sequence of subgroups of G satisfying the following three relations:
(a) {G
: <
1
} = G;
(b) card(G
) for each <

1
;
(c) G
(x) \ {G
(x) : <} = / 0 for each <

1
.
Let X be a selector of the family of nonempty sets
{(G
(x) \ {G
(x) : <}) : <

1
}.
Then, by using a characterization of absolutely negligible sets (see Theorem 6 of Chapter
2), it is not difcult to verify that X is a G-absolutely negligible subset of E. Since X is also
uncountable, we get the required result.
The next statement easily follows from Lemma 6.
Theorem 4. Under the assumptions of Lemma 6, there exists a function f
1
: E R such
that:
(1) card(ran( f
1
)) =
1
;
(2) for any -nite G-invariant (G-quasi-invariant) measure on E, there exists a G-
invariant (G-quasi-invariant) extension of for which f
1
becomes equivalent to zero and,
consequently, becomes measurable.
Proof. Using Lemma 6, we can nd a G-absolutely negligible subset X of E with card(X) =
1
. Let f
1
: E R be a function dened as follows:
f
1
|X is injective and f
1
|(E \ X) is equal to zero.
A straightforward verication shows that f
1
satises both relations (1) and (2).
Comparing Theorem 2 with Theorem 4, we see that the case of ordinary measures essen-
tially differs from the case of invariant (quasi-invariant) measures. Moreover, in view of
Theorem 3, the following natural question arises.
Problem. Let (E, G) be a space equipped with a transformation group. Give a characteriza-
tion of all those step-functions f on E which have the generalized measurability property in
the sense that for any -nite G-invariant (G-quasi-invariant) measure on E, there exists
a G-invariant (G-quasi-invariant) extension
of such that f becomes
-measurable?
This problem remains open.
Example 3. Let f be a step-function of Theorem 3 such that
ran( f ) {0} = / 0
and let f
1
be a function of Theorem 4. Let us put
f
2
= f + f
1
.
Then for any nonzero -nite G-invariant (G-quasi-invariant) measure on E, there exists
a G-invariant (G-quasi-invariant) extension
of such that f
2
is
-measurable and, at
the same time, f
2
never becomes equivalent to zero (with respect to
). In addition to this
fact, if card(E) =
1
and the Continuum Hypothesis holds, then the set ran( f
2
) R can be
as bad as possible (in the sense of its descriptive structure).
Remark 2. If we deal with nitely additive G-invariant normalized measure type function-
als on E, then it is reasonable to call a step-function any function f : E R whose range
is nite, i.e.,
card(ran( f )) <.
In this case, corresponding measurability properties of f essentially depend on the alge-
braic structure of G. For instance, if G is amenable and is an arbitrary nitely additive
G-invariant normalized measure type functional on E, then, according to von Neumanns
extension theorem (see, for instance, [83], [108], and [240]), every step-function on E
becomes measurable with respect to an appropriate universal nitely additive G-invariant
extension
of (as usual, the term universal means here that dom(
) coincides with
the power set of E). On the other hand, if E = G and G admits nite paradoxical de-
compositions, then it is obvious that there exist step-functions on E which are absolutely
nonmeasurable with respect to the class of all nitely additive left G-invariant normalized
measure type functionals on E (cf. Example 1). For more details about nite paradoxical
decompositions, see [155] and [240].
EXERCISES
1. Give a detailed explanation of the fact that Martins Axiom with the negation of the
Continuum Hypothesis imply assertion (*) which was introduced in this chapter.
2
. Recall that S R is a Sierpi nski set in R if S is uncountable and the inequality card(S
Y) holds true for every Lebesgue measure zero subset Y of R. It is well known that the
Continuum Hypothesis implies the existence of a Sierpi nski set (see, for instance, [148],
[172], [176], [192], [222], and Exercise 14 for Appendix 1).
Check that any uncountable subset of a Sierpi nski set S has strictly positive outer Lebesgue
measure. Using this fact, establish that S does not contain an uncountable universal measure
zero subset.
This circumstance immediately shows that the existence of a Sierpi nski subset of R implies
the negation of the above-mentioned assertion (*).
3. Let E be a set, G be a group of transformations of E and let be a -nite G-quasi-
invariant measure on E. Suppose that X E is an almost G-invariant set with respect to .
Verify that a -measurable hull of X is also almost G-invariant with respect to .
4
. Using the method of transnite induction, prove that there exists a subset Z of R satis-
fying the following conditions:
(a) card(Z) = card(R\ Z) = c;
(b)
(Z) =
(R\Z) =0, i.e., both sets Z and R\Z are thick with respect to the Lebesgue
measure on R;
(c) for any h R, the inequality card((h +Z)Z) < c holds true.
Infer from this result that, under Martins Axiom, there exists a partition of R into two
almost R-invariant subsets (with respect to ), none of which is of -measure zero.
5
. According to the classical result of Kunen, the real-valued measurability of c implies

that there exists a set X R such that card(X) <c and X is not measurable in the Lebesgue
sense (for the proof, see [69], [79], and [119]).
Deduce from this fact that under the same assumption, there exists a set Y R such that
card(Y) < c and Y is thick with respect to . Conclude that:
(a) Y is not measurable in the Lebesgue sense;
(b) Y is not universal measure zero.
6. Complete the proof of Lemma 6, by showing that the selector X described therein is a
G-absolutely negligible set in E.
7
. Let E be a set of cardinality

1
and let G be a group of transformations of E satisfying
the following conditions:
(a) card(G) =
1
;
(b) G acts freely in E.
Applying the argument presented in the proof of Lemma 6, show that there exists a count-
able covering of E by G-absolutely negligible sets. Conclude from this fact that for any
nonzero -nite G-invariant (G-quasi-invariant) measure on E, there always exists a
G-invariant (G-quasi-invariant) measure on E strictly extending .
8. Complete the argument of Example 3 and show that under the Continuum Hypothesis,
the set ran( f
2
) R can be as bad as possible from the measure-theoretical and descriptive
point of view.
Chapter 7
Almost measurable real-valued functions
In this chapter the notion of an almost measurable real-valued function is introduced and
examined. Some properties of such functions are considered and their characterization is
given by using the classical Fubini theorem and a theorem on the existence of measurable
selectors (see Appendix 2). In addition to this result, close connections of almost mea-
surable functions with extensions of the standard Lebesgue measure on the real line are
pointed out and the structure of algebraic sums of almost measurable functions is studied.
As usual, let R denote the real line and let f : RR be a function. We say that f is almost
continuous if for every open set U R
2
containing the graph Gr( f ) of f , there exists a
continuous function g : R R whose graph Gr(g) is also contained in U.
The above notion was introduced by J. Stallings [232] in connection with some xed-
point theorems. Notice that there are many works devoted to various properties of almost
continuous functions (see, e.g., [184] and the list of references therein).
The concept of almost measurable real-valued functions is relative to almost continuous
functions and seems to be of some interest from the view-point of real analysis and clas-
sical measure theory. Even the denition of almost measurable functions is very similar
to the denition of almost continuous functions. See below the denition of the almost
measurability of functions.
As usual, we denote by (=
1
) the one-dimensional Lebesgue measure on the real line
R (= R
1
). The symbol
2
stands for the two-dimensional Lebesgue measure on the Eu-
clidean plane R
2
(in fact,
2
coincides with the completion of the product measure
1
1
).
We shall say that a function f : R R is almost measurable if for every
2
-measurable set
V R
2
containing the graph Gr( f ) of f , there exists a
1
-measurable function g : R R
whose graph Gr(g) is also contained in V.
111
7
The following assertions are easy consequences of this denition.
(a) Any
1
-measurable function f : R R is almost measurable.
(b) If f : R R is almost measurable and its graph is
2
-measurable, then f is
1
-
measurable.
Indeed, (a) is trivial because for every
2
-measurable set V R
2
containing the graph of
f , we may put g = f .
To show (b), it sufces to take as V the graph of a given function f ; then any Lebesgue
measurable function g : R R whose graph is contained in V necessarily coincides with
f , which gives the required result.
In connection with (b), let us also remark that if the graph of a function
f : R R
is
2
-measurable, then it is of
2
-measure zero because in the plane R
2
there are uncount-
ably many pairwise disjoint translates of this graph. On the other hand, simple examples
enable us to claim that the equality
2
(Gr( f )) = 0 does not imply, in general, the
1
-
measurability of f .
It is well known (see, e.g., Exercise 9 fromChapter 2) that there exists a function f : RR
whose graph is a thick subset of the plane R
2
, which means that
(
2
)
(R
2
\ Gr( f )) = 0.
Clearly, such an f is not Lebesgue measurable. As already mentioned in Chapter 2, one of
the earliest examples of a function f of this type was constructed by W. Sierpi nski, with the
aid of the method of transnite recursion (see, for instance, [72], [144], [148], and [192]).
We shall establish below that all such functions f turn out to be almost measurable in the
sense of our denition and, in this respect, they cannot be treated as extremely pathological
functions.
In the sequel, we will need an auxiliary proposition concerning the existence of measurable
selectors. This proposition is a straightforward corollary of the classical theorem from
descriptive set theory, which is due to Luzin, Jankov, and von Neumann (see Appendix 2).
Lemma 1. Let A R
2
be an analytic (i.e., Suslin) set. There exists a Lebesgue measurable
function g : pr
1
(A) R whose graph is contained in A.
The proof of Lemma 1 (and of a more general statement) is presented in Appendix 2. Using
this lemma, we can obtain a certain characterization of all almost measurable functions.
Theorem 1. Let f : R R be a function and let D denote some
2
-measurable hull of the
graph of f . The following two assertions are equivalent:
Almost measurable real-valued functions 113
(1) f is almost measurable;
(2) there exists a disjoint covering {X
1
, X
2
} of R by two
1
-measurable sets such that the
restriction f |X
1
is Lebesgue measurable and for each point x X
2
, the inequality
1
({y
R : (x, y) D}) > 0 is true.
Proof. (1) (2). Suppose that (1) is satised for the given function f . Let us put:
X
1
={x R :
1
({y R : (x, y) D}) = 0},
X
2
= R\ X
1
.
Obviously, the sets X
1
and X
2
are disjoint,
1
-measurable, and cover R. Moreover, in view
of the classical Fubini theorem, we have
2
(D(X
1
R)) = 0.
Now, let us dene
D
={(x, f (x)) : x X
1
} (D(X
2
R)).
Clearly, the set D(X
2
R) is
2
-measurable. By virtue of the inclusion
{(x, f (x)) : x X
1
} D(X
1
R),
the set {(x, f (x)) : x X
1
} is of
2
-measure zero. Consequently, D
is a
2
-measurable
subset of R
2
containing the graph Gr( f ) of f . Since f is almost measurable, there exists a
Lebesgue measurable function g : R R whose graph is contained in D
. Then we get the

equality
g|X
1
= f |X
1
which shows that the restriction f |X
1
is Lebesgue measurable. At the same time, it directly
follows from the denition of X
2
that
1
({y R : (x, y) D}) > 0
for any point x X
2
. We thus conclude that the implication (1) (2) holds true.
Suppose now that (2) is satised for our function f and let {X
1
, X
2
} denote the correspond-
ing disjoint covering of R by two
1
-measurable sets. Let V R
2
be any
2
-measurable
set containing the graph of f . We may assume, without loss of generality, that V D. This
assumption implies the equality
2
(D\V) = 0
because D is a measurable hull of Gr( f ). Consider a Borel subset V
of V such that
2
(V \
V
) = 0. Clearly, we have
2
(D\V
) = 0. Further, take a Borel subset X
2
of X
2
such that
1
(X
2
\ X
2
) = 0.
Using the Fubini theorem once more, we easily deduce that
1
(X
2
\ pr
1
(V
(X
2
R))) = 0.
Applying Lemma 1 to the Borel set V
(X
2
R), we can nd a Lebesgue measurable
function
h : pr
1
(V
(X
2
R)) R
whose graph Gr(h) is contained in the set V
(X
2
R).
Let us put:
g(x) = h(x) for all points x pr
1
(V
(X
2
R));
g(x) = f (x) for all points x R\ pr
1
(V
(X
2
R)).
So we come to the function
g : R R
whose graph Gr(g) is contained inV. Finally, taking into account the Lebesgue measurabil-
ity of f |X
1
and the relation g|X
1
= f |X
1
, it is easy to see that g is also Lebesgue measurable.
This establishes the implication (2) (1) and ends the proof of the theorem.
From Theorem 1 we immediately obtain the next statement.
Theorem 2. Suppose that f : R R is a function whose graph Gr( f ) is thick with respect
to the measure
2
. Then f is almost measurable.
Proof. Indeed, using the notation of Theorem 1 and putting
D = R
2
, X
1
= / 0, X
2
= R,
we see that assertion (2) of Theorem 1 is automatically satised for f . Consequently, f is
almost measurable.
Now, we are going to show that any almost measurable function turns out to be measurable
with respect to an appropriate extension of the Lebesgue measure
1
. For this purpose, we
need the measure extension construction which is successfully applicable to a wide class
of -nite measures (cf. [141] and Chapter 2). It is reasonable to recall here some details
of the above-mentioned construction.
Let E be a set, S be a -algebra of subsets of E and let be a -nite measure dened
on S. Further, let F be another set and let be a probability measure dened on some
-algebra of subsets of F. Suppose that a function
g : E F
is given, whose graph Gr(g) is thick with respect to the product measure , i.e., we
have the equality
( )
((E F) \ Gr(g)) = 0.
For every set Z S dom(), let us put
Z
={x E : (x, g(x)) Z}

and dene
S
={Z
: Z S dom()},
(Z
) = ( )(Z) (Z
).
Taking into account the ( )-thickness of Gr(g), we can verify that:
(i) S
is a -algebra of subsets of E containing S;

(ii) the functional
is well dened and is a measure on S
extending ;
(iii) the function g is measurable with respect to the -algebras dom() and S
, which
means that for every set Y dom(), the pre-image g
1
(Y) belongs to S
.
The details of the verication of (i) - (iii) are left to the reader (cf. Lemmas 1, 2, and 3 from
Chapter 2).
The main idea of this construction will be used below in a very particular case, where we
have
E = F = R, =
1
.
Lemma 2. Let X be a
1
-measurable subset of R, let f : X R be a function, and let D
denote some
2
-measurable hull of Gr( f ). Suppose that for each point x X, the relation
0 <
1
({y R : (x, y) D}) < +
holds true. Then there exists a measure
1
on X such that:
(1)
1
extends the restriction of
1
to the -algebra of all Lebesgue measurable subsets of
X;
1
.
Proof. We may assume, without loss of generality, that X is a Borel subset of R and D is a
Borel subset of R
2
. Moreover, we may also assume that D X R. Then the function
(x) =
1
({y R : (x, y) D}) (x X)
is Borel measurable and according to our assumption, we have
0 <(x) < + (x R).
Let us put
f
1
(x) = f (x)/(x) (x X).
Since f = f
1
and is Borel measurable, it sufces to nd a measure
1
which extends
the restriction of
1
to the -algebra of all Lebesgue measurable subsets of X and for which
f
1
becomes measurable. Let us dene
(x, y) = (x, y/(x)) ((x, y) X R).
Notice that is a Borel bijection of the set X R onto itself. Putting
D
1
=(D),
we readily infer (by virtue of the Fubini theorem) that D
1
is a measurable hull of Gr( f
1
).
In addition to this fact, we have
1
({y R : (x, y) D
1
}) = 1
for all points x X. Now, we can apply the idea of the measure extension construction
which was described earlier. Let S denote the -algebra of all Lebesgue measurable
subsets of X. For any
2
-measurable set Z D
1
, let us put
Z
={x X : (x, f
1
(x)) Z}
and dene
S
={Z
: Z D
1
, Z dom(
2
)},
1
(Z
) =
2
(Z) (Z
).
A straightforward verication shows that S
is a -algebra of subsets of X containing S,

the functional
1
is well dened and turns out to be a measure on X. As before, the details
of this verication are left to the reader.
Now, take any
1
-measurable set T X and consider the set D
1
(T R). Obviously, the
relation
T ={x X : (x, f
1
(x)) (D
1
(T R))}
holds true. Keeping in mind the equality
1
({y R: (x, y) D
1
}) =1 for each point x X,
we may write
1
(T) =
2
(D
1
(T R)) =
T

1
({y R : (x, y) D
1
})dx =
1
(T)
which shows that
1
extends the restriction of
1
to the -algebra S. Finally, for any Borel
set Y R, we have
f
1
1
(Y) ={x X : (x, f
1
(x)) D
1
(X Y)} S
from which it follows that f

1
is measurable with respect to
1
. Lemma 2 has thus been
proved.
Theorem 3. Let f : R R be an arbitrary almost measurable function. Then there exists
a measure
1
on R such that:
(1)
1
extends
1
;
1
.
Proof. Suppose rst that ran( f ) [0, 1]. Denote by D a
2
-measurable hull of Gr( f ).
Obviously, we may assume that
D R[0, 1].
Since f is almost measurable, we can nd a disjoint covering {X
0
, X} of R satisfying the
(a) both sets X
0
and X are
1
-measurable;
(b) for each point x X
0
, we have
1
({y R : (x, y) D}) = 0;
(c) for each point x X, we have
1
({y R : (x, y) D}) > 0.
Notice that the restriction of f to X
0
is Lebesgue measurable (cf. the proof of Theorem 1).
At the same time, for any point x X, the relation
0 <
1
({y R : (x, y) D}) 1
is valid. So we may apply Lemma 2 to the function f |X. In this manner, we obtain an
extension
1
of
1
such that f becomes
1
-measurable.
The case of an arbitrary almost measurable function f : R R can readily be reduced to
the previous case. Indeed, it is easy to construct a homeomorphism
: R ]0, 1[
which transforms the -ideal of all Lebesgue measure zero subsets of R onto the -ideal
of all Lebesgue measure zero subsets of ]0, 1[. Let us put
(x, y) = (x, (y)) (x R, y R).
The mapping transforms the set Gr( f ) onto the set Gr( f ). The above-mentioned
property of directly implies that f is almost measurable and
ran( f ) [0, 1].
Consequently, there exists a measure
1
on R extending
1
for which f turns out to be
measurable. It immediately follows from this fact that the original function
f =
1
( f )
is measurable with respect to
1
, too, which completes the proof.
Example 1. Recall that B R is a Bernstein set in R if for every nonempty perfect set
P R, both intersections PB and P(R\B) are nonempty. Extensive information about
Bernstein sets and their properties can be found in [12], [30], [40], [72], [138], [143], [148],
[176], and [192] (cf. also Exercise 5 for Chapter 5). It is well known that the Bernstein
sets are nonmeasurable with respect to the completion of any nonzero -nite diffused (i.e.
vanishing at all singletons) Borel measure on R and, consequently, they are nonmeasurable
with respect to
1
. At the same time, if B is a Bernstein set, then there exists an extension
1
of
1
for which B becomes measurable and, as explained in Chapter 2, this fact is true
for any set X R. Denoting by
f =
B
the characteristic function of B, we conclude that f is not almost measurable but is mea-
surable with respect to
1
. In particular, this result shows that the statement converse to
Theorem 3 does not hold.
Theorem 4. Let f : R R and g : R R be two functions. Suppose that f is almost
measurable and g is Lebesgue measurable. Then their sum f +g is almost measurable.
Proof. Taking into account Theorem 1 and the fact that the sum of any two Lebesgue
measurable functions is also Lebesgue measurable, it sufces to establish the almost mea-
surability of f |X +g|X where X is a
1
-measurable subset of R and
1
({y R : (x, y) D}) > 0
for all points x X (here D denotes a
2
-measurable hull of the graph of f |X).
Without loss of generality, we may assume that X is a Borel subset of R and D is a Borel
subset of R
2
. Let be the characteristic function of D X R and let us dene
(x, y) =(x, y g(x)) (x X, y R).
By using the Fubini theorem, it is not difcult to check that turns out to be the character-
istic function of some
2
-measurable hull C of the graph of ( f +g)|X and the inequality
1
({y R : (x, y) C}) > 0
holds true for all points x X. By virtue of Theorem 1, this circumstance immediately
yields that the function
( f +g)|X = f |X +g|X
is almost measurable, which completes the proof.
The next statement shows, in particular, that the sum of two almost measurable additive
functions can be a function without the property of almost measurability.
Theorem 5. There exist two functions f : R R and g : R R satisfying the following
relations:
(1) f and g are additive, that is they are homomorphisms of the additive group R into itself;
(2) the graphs of f and g are thick with respect to
2
;
(3) ran( f +g) = Q.
Therefore, both f and g are almost measurable but their sum f +g is not almost measur-
able.
Proof. Consider R as a vector space over the eld Q. Let {e
i
: i I} be a Hamel basis of
this space. Fix an index j I and denote by L the hyperplane in R generated by the family
{e
i
: i I \ { j}}. As is well known, L is a thick subset of R with respect to the measure
1
and, consequently, L is nonmeasurable with respect to
1
. By using the standard transnite
construction, an additive function
: L L
can be dened such that its graph Gr() is thick in R
2
with respect to the measure
2
(cf.
Exercise 9 from Chapter 2).
Now, take an arbitrary x R. Then x admits a unique representation in the formx =qe
j
+e,
where q Q and e L. Let us put:
f (x) = f (qe
j
+e) = q/2 +(e),
g(x) = g(qe
j
+e) = q/2 (e).
In this way, the functions
f : R R, g : R R
are completely determined. So, it remains to check the validity of the relations (1), (2), and
(3).
The additivity of both functions f and g follows directly from their denition.
The
2
-thickness of their graphs is easily implied by the
2
-thickness of the set Gr(),
which is contained in Gr( f ) Gr(g).
Further, for any x R, consider its unique representation in the form
x = qe
j
+e,
where q Q and e L. Clearly, we have
( f +g)(x) = (q/2 +(e)) +(q/2 (e)) = q.
Consequently, the equality
ran( f +g) = Q
holds true. This equality shows that the additive function
f +g : R R
is a nontrivial solution of the Cauchy functional equation, so f +g is not Lebesgue mea-
surable. At the same time, the set Gr( f +g) is of
2
-measure zero because Gr( f +g) is
contained in the union of countably many horizontal lines in the plane R
2
. Therefore, the
function f +g is not almost measurable which completes the proof.
The functions f and g described in Theorem 5 are almost measurable, so they become
measurable with respect to appropriate extensions of the Lebesgue measure on R. It
should be noticed that those extensions can be chosen to be quasi-invariant under the group
of all translations of R (cf. Exercise 7 for this chapter).
In view of Theorem 5, the following example seems to be relevant.
Example 2. If E is an uncountable topological space, then we may introduce the class M
E
of the completions of all nonzero -nite diffused Borel measures on E. Consider more
thoroughly the particular case when E = R.
We know that a function f : R R is relatively measurable with respect to M
R
if f admits
a representation in the form
f = g ,
where g : R R is a Lebesgue measurable function and is a Borel isomorphism of R
onto itself (see Exercise 12 for Chapter 5).
The algebraic sum of two relatively measurable functions with respect to M
R
can be an ab-
solutely nonmeasurable function with respect to the same class. To show this fact, consider
two nonzero -nite diffused Borel measures
1
and
2
on R such that
1
([0, +[) = 0,
2
(] , 0]) = 0.
Let B
1
denote a Bernstein subset of the half-line [0, +[ and let B
2
denote a Bernstein
subset of the half-line ] , 0].
We dene a function f
1
: R R by putting:
f
1
(x) = 1 if x B
1
and f
1
(x) = 0 if x R\ B
1
.
Also, we dene a function f
2
: R R by putting:
f
2
(x) = 1 if x B
2
and f
1
(x) = 0 if x R\ B
2
.
It immediately follows fromthese denitions that the function f
1
is measurable with respect
to the completion of
1
and the function f
2
is measurable with respect to the completion
of
2
. At the same time, the sum f
1
+ f
2
is the characteristic function of the subset B
1
B
2
of R. But it is easy to see that B
1
B
2
is a Bernstein subset of R. Therefore, f
1
+ f
2
is
absolutely nonmeasurable with respect to M
R
.
Example 3. Let be a nonzero -nite diffused measure on R and let f : R R be a
function such that for some nonzero -nite Borel measure on ran( f ), the graph of f is
( )-thick in the product set Rran( f ). Then we can assert that f is measurable with
respect to an appropriate extension
of and hence f is not absolutely nonmeasurable.

Indeed, we may suppose without loss of generality that is a Borel probability measure on
ran( f ) and we may apply once again the method of extending -nite measures described
in this chapter (see also Chapter 2). Namely, for each set Z dom( ), we denote
Z
={x R : (x, f (x)) Z}

and introduce the family of sets
S
={Z
: Z dom( )},
which turns out to be a -algebra of subsets of R. In addition, if X dom(), then we have
X ran( f ) dom( ),
X ={x R : (x, f (x)) X ran( f )},
thus it follows that X S
. Consequently, the inclusion dom() S
holds true. Finally,

for any Z dom( ), we put
(Z
) = ( )(Z).
As a direct verication shows, the functional
is well dened by virtue of the thickness of

the graph of our function f and, moreover, this
is a measure on S
extending the initial

measure . The denition of
also implies that the given function f is measurable with

respect to
.
In particular, if the graph of a function g : R R is ( )-thick in the plane R
2
, then,
as we already know, g can be made measurable with respect to an appropriate extension of
. However, we cannot guarantee the quasi-invariance (under the group of all translations
of R) of the obtained in this way extension. For the above-mentioned quasi-invariance,
we need some additional properties of g. In particular, it sufces to suppose that g is a
homomorphism of the additive group R into itself and that the graph of g is a
2
-thick
subset of R
2
(see Exercise 7 below).
In this context, the following problem naturally arises.
Problem. Let be a measure on R extending and such that there exists a function acting
from R into R whose graph is ( )-thick in the plane R
2
. Does there exist a group
homomorphism acting from R into R whose graph is also ( )-thick in R
2
?
We do not know an answer to this question.
EXERCISES
1. Dene a function f : R R as follows:
f (x) = sin(1/x) if x = 0 and f (x) = 0 if x = 0.
Check that f is not continuous but is almost continuous. Give other concrete examples of
almost continuous functions acting from R into R which are not continuous.
Show directly that the characteristic function of Q (the so-called Dirichlet function) is not
almost continuous (cf. Exercise 4 below).
2
. Starting with the denition of almost continuous functions, demonstrate that if f : R

R is continuous and g : R R is almost continuous, then their sum f +g is almost contin-
uous.
Is it true that if f : R R and g : R R are any two almost continuous functions, then
their sum f +g is also almost continuous?
3. The notion of an almost continuous function can be generalized in the following manner
(see [232]).
Let X and Y be two topological spaces and let f : X Y be a mapping. We shall say that f
is almost continuous if for any open set U X Y containing the graph of f , there exists
a continuous mapping g : X Y whose graph is also contained in U.
Let f : X Y be almost continuous and let h : Y Z be continuous. Check that the
mapping h f : X Z is almost continuous.
We shall say that a topological space E has the xed-point property if for any continuous
mapping f : E E, there exists e E such that f (e) = e.
Suppose that a Hausdorff space E has the xed-point property (according to Browers theo-
rem, E satises this condition if it is homeomorphic to a nonempty compact convex subset
of a nite-dimensional Euclidean space). Let g : E E be an almost continuous mapping.
Show that there exists a xed point for g, i.e.,
(e E)(g(e) = e).
4
. By using the result of the previous exercise, prove that every almost continuous function
f : R R has the Darboux property or the intermediate value property, which means that
if a ran( f ) and b ran( f ), then the line segment with end-points a and b is contained in
ran( f ).
Infer from this fact that for a monotone function g : R R, the following two conditions
are equivalent:
(a) g is continuous;
(b) g is almost continuous.
Conclude that there are many Borel measurable (hence Lebesgue measurable and almost
measurable) functions which are not almost continuous.
5. Give a direct proof of Theorem 4 by starting with the denition of an almost measurable
function.
6. Give a precise construction of an additive function : L L mentioned in the proof of
Theorem 5.
7. Let h : R R be a function satisfying the following conditions:
(a) h is additive;
(b) the graph of h is
2
-thick in the plane R
2
.
Prove that h is measurable with respect to some translation-quasi-invariant extension of the
Lebesgue measure on R.
In particular, let f and g be as in Theorem 5. Show that there exist two translation-quasi-
invariant measures
1
and
2
on R both extending and such that f is measurable with
respect to
1
and g is measurable with respect to
2
.
8
. Prove that for every function h : R R, there exist two functions f : R R and
g : R R satisfying the following conditions:
(a) the graph of f is
2
-thick in R
2
;
(b) the graph of g is
2
-thick in R
2
;
(c) h = f +g.
For this purpose, construct the required f and g by using the method of transnite recursion.
Derive from the above result that under Martins Axiom the sum of two almost measurable
functions can be an absolutely nonmeasurable function (with respect to the class M
R
).
9
. Prove that for every additive function h : R R, there exist two functions f : R R
and g : R R satisfying the following relations:
(a) both f and g are additive functions;
(b) the graph of f is
2
-thick in R
2
;
(c) the graph of g is
2
-thick in R
2
;
(d) h = f +g.
For this purpose, consider Ras a vector space over Qand use again the method of transnite
recursion.
Deduce from the above result that under Martins Axiom the sum of two additive almost
measurable functions can be an additive absolutely nonmeasurable function (with respect
to the class M
R
).
Chapter 8
Several facts from general topology
It is well known that close relationships between real analysis and general topology are
very useful and fruitful for both of these mathematical disciplines. For instance, deep
analogies between the Lebesgue measure and Baire category on R were extensively studied
and underlined many times in the literature (see, e.g., [40], [148], [176], and [192]). In this
chapter, we would like to discuss some topological facts and statements which play an
auxiliary role in our further considerations.
Recall that a topological space E is quasicompact if any open covering of E contains a
nite subcovering.
We begin with a simple but remarkable theorem of Kuratowski on closed projections (see
[58] and [148]).
Theorem 1. Let X be a topological space, Y be a quasicompact topological space, and let
pr
1
: X Y X
denote the canonical projection dened by the formula
pr
1
((x, y)) = x ((x, y) X Y).
Then pr
1
is a closed mapping, i.e., for each closed subset A of X Y, the image pr
1
(A) is
closed in X.
Proof. Consider any point x X such that U(x) pr
1
(A) = / 0 for every neighborhoodU(x)
of x. We must show that x pr
1
(A). For this purpose, it sufces to establish that
({x} Y) A = / 0.
Suppose otherwise, that is ({x} Y) A = / 0. Then for each point y from Y, there exists
an open neighborhood W(x, y) of (x, y) satisfying the relation W(x, y) A = / 0. We may
assume, without loss of generality, that W(x, y) has the form
W(x, y) =U(x) V(y),
DOI 10.1007/978-94-91216-36-7_ , 2009 Atlantis Press/World Scientific 8
129
where U(x) is an open neighborhood of x and V(y) is an open neighborhood of y. Since
the space {x} Y is quasicompact, there exist a nite sequence
(x, y
1
), (x, y
2
), . . . , (x, y
n
)
of points from{x} Y and the corresponding nite sequence
U
1
(x) V
1
(y
1
), U
2
(x) V
2
(y
2
), . . . , U
n
(x) V
n
(y
n
)
of neighborhoods of these points such that
{x} Y (U
1
(x) V
1
(y
1
)) (U
2
(x) V
2
(y
2
)) ... (U
n
(x) V
n
(y
n
)).
Let us put
U(x) =U
1
(x) U
2
(x) ... U
n
(x).
Then U(x) is a neighborhood of x and it can readily be veried that
U(x) pr
1
(A) = / 0,
contradicting the denition of x. The contradiction obtained nishes the proof.
Theorem1 has many applications in general topology and real analysis. For example, it can
efciently be applied in the proof of the existence of a continuous nowhere differentiable
function f : [0, 1] R (see Exercise 4 of this chapter). Moreover, this theorem turns out to
be useful in some questions of the theory of ordinary differential equations (see [190] and
[122]).
It is interesting to notice that the statement converse to the Kuratowski theorem holds true,
too (see Exercise 3).
Let us briey discuss some other useful facts about closed mappings (i.e., those mappings
which preserve the closedness of sets) and their specic properties. For instance, it makes
sense to study the structure of the graphs of such mappings.
The following simple and well-known statement is concerned with the descriptive structure
of graphs of continuous mappings acting fromany topological space into a Hausdorff space.
Theorem 2. Let X be a topological space, Y be a Hausdorff topological space, and let
f : X Y be a continuous mapping. Then the graph Gr( f ) of f is a closed subset of the
product space X Y.
Proof. We need one easy auxiliary proposition. Namely, if Z is a topological space, Y is a
Hausdorff topological space and
g : Z Y, h : Z Y
Several facts from general topology 127
are continuous mappings, then the set {z Z : g(z) = h(z)} is closed in Z because it co-
incides with the pre-image of the closed set {(y, y) : y Y} Y Y under the continuous
mapping
(g, h) : Z Y Y.
Let now f : X Y be an arbitrary continuous mapping. Put Z = X Y and consider two
continuous mappings
g : X Y Y, h : X Y Y
dened by the formulas
g(x, y) = y, h(x, y) = f (x) ((x, y) X Y).
According to the said above, the set
{(x, y) X Y : g(x, y) = h(x, y)} ={(x, y) X Y : y = f (x)} = Gr( f )
is closed in X Y which completes the proof.
Obviously, Theorem 2 enables us to construct many examples of those continuous map-
pings which are not closed but have closed graphs. However, the following statement is
valid (cf. [58] and [101]).
Theorem 3. Let X be a regular topological space, Y be a T
1
-space and let f : X Y be a
closed continuous mapping. Then the graph Gr( f ) is a closed subset of the product space
X Y.
Proof. It sufces to show that the set (X Y) \ Gr( f ) is open in X Y. Take any point
(x, y) (X Y) \Gr( f ). We have x f
1
(y). Since f
1
(y) is closed in X, there exist open
neighborhoods U(x) and V( f
1
(y)) such that
U(x) V( f
1
(y)) = / 0.
The set Y \ f (X \V( f
1
(y))) is open and contains y. It is easy to verify that
Gr( f ) (U(x) (Y \ f (X \V( f
1
(y))))) = / 0.
This fact immediately implies that the set (X Y) \ Gr( f ) is open which ends the proof.
In general, the product of two closed continuous mappings does not need to be closed (see
Exercise 5).
On the other hand, we have the following statement.
Theorem 4. The product of nitely many mappings whose graphs are closed is a mapping
with closed graph.
Proof. It sufces to consider two mappings
f
1
: X
1
Y
1
, f
2
: X
2
Y
2
,
whose graphs are closed. The product mapping
f
1
f
2
: X
1
X
2
Y
1
Y
2
dened by the standard formula
( f
1
f
2
)(x
1
, x
2
) = ( f
1
(x
1
), f
2
(x
2
)) ((x
1
, x
2
) X
1
X
2
)
has the graph
Gr( f
1
f
2
) ={(x
1
, x
2
, y
1
, y
2
) : f
1
(x
1
) = y
1
, f
2
(x
2
) = y
2
}.
Clearly, the set Gr( f
1
f
2
) is homeomorphic to the set Gr( f
1
) Gr( f
2
) under the canon-
ical homeomorphism of X
1
X
2
Y
1
Y
2
onto X
1
Y
1
X
2
Y
2
. It remains to use the
simple fact that the product of any two closed sets is closed, too, thus the required result
immediately follows.
Recall that a topological space E is compact if E is quasicompact and Hausdorff simulta-
neously.
Theorem 5. Let X be a compact topological space, R be an equivalence relation on X and
let G = Gr(R) denote the graph of R. The following three relations are equivalent:
1) G is a closed subset of X X;
2) R is closed, i.e., the canonical surjection : X X/R is a closed mapping;
3) the quotient space X/R is Hausdorff.
Proof. 3) 1). Let us show that this implication holds true without any assumption on X.
Indeed, suppose that X/R is Hausdorff. This circumstance means that the diagonal
={z X/RX/R: pr
1
(z) = pr
2
(z)}
is closed in the product space X/RX/R. But we obviously have
G = ( )
1
().
Thus G is closed, being the pre-image of under the continuous mapping .
1) 2). Let G = Gr(R) be closed in the product space X X. We must show that : X
X/R is a closed mapping. For this purpose, consider any closed set F X. Clearly, we
may write
R(F) = pr
2
((F X) G).
The set (F X) G is closed in X X. Since X is compact, R(F) is also closed in view of
the Kuratowski theorem on closed projections. But this fact precisely means that : X
X/R is a closed mapping.
2) 3). Suppose that R is closed, i.e., : X X/R is a closed mapping. As is well known,
the compact X is also regular and, in view of the closedness of R, all of the singletons in the
quotient space X/R are closed which means that X/R is a T
1
-space. Applying Theorem 3,
we derive that the graph Gr() of is a closed subset of the product space X X/R which
implies, by virtue of Theorem 4, that the product mapping has closed graph. At the
same time, for any closed set F X X, we may write the formula
( )(F) = pr
2
((F ((X/R) (X/R))) Gr( )).
Since F is quasicompact (as a closed subset of the quasicompact product space X X), we
can apply the Kuratowski theoremand infer from it that is a closed mapping. Finally,
it is clear that
= ( )({(x, x) : x X})
and since the set {(x, x) : x X} is closed, turns out to be closed in the product space
X/RX/R. But the latter means that X/R is a Hausdorff topological space. Theorem 5
has thus been proved.
Now, we would like to present one important statement concerning the existence of Borel
selectors of certain partitions of a Polish topological space. This statement will be es-
sentially utilized in Chapter 9 where we discuss the question of measurability of various
selectors associated with countable transformation groups.
Our presentation follows Bourbakis widely known book [24] with minor changes and sim-
plications.
First of all, we need the next purely set-theoretical proposition.
Theorem 6. Let E be an arbitrary set in which a countable system of its subsets
(F
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<
(1 k <)
is given satisfying the following conditions:
1) E ={F
n
: n <};
2) F
n
1
n
2
...n
k
={F
n
1
n
2
...n
k
n
: n <} for any n
1
<, n
2
<, ..., n
k
< and for every nonzero
natural number k.
Let R be an equivalence relation in E. Dene by ordinary recursion another countable
system
(H
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<
(1 k <)
of subsets of E. Namely, put
H
n
= F
n
({E \ R(F
m
) : m < n}) (n <),
H
n
1
n
2
...n
k
n
k+1
= F
n
1
n
2
...n
k
n
k+1
H
n
1
n
2
...n
k
({E \ R(F
n
1
n
2
...n
k
m
) : m < n
k+1
}).
Then for each R-equivalence class X and for any natural number k 1, there exists a
unique set of the form H
n
1
n
2
...n
k
such that
F
n
1
n
2
...n
k
X = H
n
1
n
2
...n
k
X = / 0.
Proof. We argue by induction on k. First, consider the case when k = 1. Pick any R-
equivalence class X. Since X = / 0, there is a smallest natural number n for which X F
n
= / 0.
In accordance with the denition, we have
H
n
= F
n
({E \ R(F
m
) : m < n}).
Taking into account the relations
X = R(X), F
m
X = / 0 (m < n),
we deduce that X {E \ R(F
m
) : m < n} from which it follows
H
n
X = F
n
X = / 0.
Further, since H
m
F
m
, we readily infer that H
m
X = / 0 whenever m < n. On the other
hand, we have X R(F
n
) and if a natural number m is strictly greater than n, the denition
of H
m
yields at once that H
m
X = / 0. Thus, the assertion of the theorem holds true for
k = 1.
Suppose now that this assertion is valid for k and let us show that it remains true for k +1.
Let X be again an arbitrary R-equivalence class. According to the inductive assumption,
there exists a unique set H
n
1
n
2
...n
k
for which
H
n
1
n
2
...n
k
X = F
n
1
n
2
...n
k
X = / 0.
Using condition 2) of the theorem, we derive that there is a smallest natural number n such
that
F
n
1
n
2
...n
k
n
X = / 0.
Let us put n
k+1
= n and consider the set
H
n
1
n
2
...n
k
n
k+1
= F
n
1
n
2
...n
k
n
k+1
H
n
1
n
2
...n
k
({E \ R(F
n
1
n
2
...n
k
m
) : m < n
k+1
}).
Similarly to the case k = 1, it can readily be seen that
X {E \ R(F
n
1
n
2
...n
k
m
) : m < n
k+1
}
from which it follows that
H
n
1
n
2
...n
k
n
k+1
X = F
n
1
n
2
...n
k
n
k+1
X = / 0.
Again, according to the inductive assumption, if
(m
1
, m
2
, ..., m
k
) = (n
1
, n
2
, ..., n
k
),
then H
m
1
m
2
...m
k
X = / 0. Consequently, only the intersection of the form H
n
1
n
2
...n
k
m
X,
where m < , may turn out to be nonempty. But, as in the case k = 1, we easily conclude
that this intersection is nonempty if and only if m = n
k+1
(in view of the denitions of the
natural index n
k+1
and of the set H
n
1
n
2
...n
k
n
k+1
). Thus, by induction, the theorem is proved.
Now, we are ready to establish the next important statement.
Theorem 7. Let E be a Polish topological space and let R be an equivalence relation in E
satisfying the following conditions:
1) all R-equivalence classes are closed subsets of E;
2) for any closed subset F of E, the set R(F) is Borel in E.
Then there exists a Borel selector of the partition E/R of E.
Proof. Since E is a Polish space, we can associate to it a countable system
(F
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<
(1 k <)
of closed subsets such that the conditions 1) and 2) of Theorem 6 are satised and, in
addition, we may suppose that the diameter of each set F
n
1
n
2
...n
k
is less than 2
k
.
For the above-mentioned system of closed subsets of E, dene the corresponding countable
system of sets
(H
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<
(1 k <)
as in the formulation of Theorem 6. Using condition 2), it is easy to see that in our case all
H
n
1
n
2
...n
k
are Borel subsets of E.
Now, for any natural number k 1, let us put
B
k
={H
n
1
n
2
...n
k
: n
1
<, n
2
<, ..., n
k
<}
and consider the set
B ={B
k
: 1 k <}.
We assert that B is the required Borel selector of E/R. Obviously, B is a Borel subset of E.
Therefore, it only remains to show that the intersection of B with any R-equivalence class
is a singleton.
For this purpose, take an arbitrary R-equivalence class X. By virtue of Theorem 6, this X
uniquely determines a decreasing sequence of sets
H
n
1
H
n
1
n
2
... H
n
1
n
2
...n
k
...
such that
H
n
1
n
2
...n
k
X = F
n
1
n
2
...n
k
X = / 0
for each natural number k 1. In view of condition 1), the last relation shows, in partic-
ular, that all sets H
n
1
n
2
...n
k
X are closed in E. Since their diameters tend to zero, their
intersection has the form {x}, where x X. We thus get
BX ={x},
which nishes the proof of Theorem 7.
Some other statements concerning the existence of measurable selectors (for set-valued
mappings) are discussed in Appendix 2.
The preceding facts and constructions and especially the systems of sets described in The-
orem 6 are closely related to the so-called A-operation from classical descriptive set theory.
It makes sense to touch upon this fundamental set-theoretical operation and indicate its
important properties.
Let E be an arbitrary set in which a countable system of subsets
(F
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<
(1 k <)
is given. Let us denote
F
t
= F
t
1
F
t
1
t
2
... F
t
1
t
2
...t
k
... (t
\{0}
).
Then the set
A((F
n
1
n
2
...n
k
)
n
1
<,n
2
<,...,n
k
<,k1
) ={F
t
: t
\{0}
}
is called the result of A-operation over the above-mentioned countable system of subsets of
E.
If L is some class of subsets of E, then the symbol A(L) denotes the class of all those
sets which can be obtained as a result of A-operation over all analogous countable systems
of sets belonging to L. The members of A(L) are usually called analytic sets over L.
From the point of view of applications, the most important case is when E is a Polish
topological space and an initial class L of subsets of E coincides with the family of all
closed sets in E. In such a situation, the members of the class A(L) are frequently called
Suslin sets in E. It can be shown that every Borel subset of E is a Suslin set. The converse
assertion is not true in general. Actually, the classical theorem of Suslin states that any
uncountable Polish topological space E contains a Suslin set which is not Borel (see, e.g.,
[10], [99], [148], [150], [160], [162], or Appendix 6).
The complements of analytic (Suslin) sets in E are called co-analytic (co-Suslin) sets. An-
alytic and co-analytic sets have rather good descriptive properties. They are universally
(absolutely) measurable with respect to the class of the completions of all -nite Borel
measures on E and possess the Baire property in the restricted sense. This circumstance is
a direct consequence of the fact that the A-operation preserves the measurability and Baire
property (cf. [10], [99], [148], [150], [160], [162], and [210]).
Such a nice circumstance fails to be longer true for projective sets of higher levels. For
instance, a continuous image (in R) of a co-analytic subset of R can be nonmeasurable
in the Lebesgue sense (in some natural models of set theory). In connection with this
unpleasant fact, see [10] and [188].
EXERCISES
1. Let Y be a -quasicompact topological space, that is Y can be covered by a count-
able family of its quasicompact subspaces. Show that for every topological space X, the
canonical projection
pr
1
: X Y X
has the following property:
if A is an F
-subset of the product space X Y, then the image pr

1
(A) is an F
-subset of
X.
2. Let Y be a metrizable topological space. Verify that the following two conditions are
equivalent:
(a) Y is compact;
(b) the canonical projection pr
1
: RY R is a closed mapping.
For this purpose, use the well-known criterion of the compactness of a metric space, which
states that a metric space Y is compact if and only if it is complete and totally bounded.
3
. Let Y be a topological space having the property that for any topological space X, the
canonical projection pr
1
: X Y X is a closed mapping. Prove that Y is quasicompact.
For this purpose, suppose otherwise and take a family {F
i
: i I} of closed subsets of Y
such that {F
i
: i I} = / 0 but
F
i
1
F
i
2
... F
i
n
= / 0
for every nite subfamily {i
1
, i
2
, ..., i
n
} of I. Fix an element x
0
Y and put
X =Y {x
0
}.
Let T denote the topology on X consisting of all sets U of the form
U = Z U = Z {x
0
} (F
i
1
F
i
2
... F
i
n
),
where Z Y, n < and {i
1
, i
2
, ..., i
n
} I. It can readily be veried that T is indeed a
topology on X.
Now, in the product space X Y consider the closed set
A = cl({(y, y) : y Y}).
Observe that Y pr
1
(A) and that x
0
is an accumulation point for Y, hence for pr
1
(A) as
well. Consequently, we must have x
0
pr
1
(A) from which it follows that there exists a
point y
0
Y satisfying the relation (x
0
, y
0
) A.
Further, for any i I and for any neighborhood U(y
0
) of y
0
, the product set ({x
0
} F
i
)
U(y
0
) is a neighborhood of (x
0
, y
0
), so
{(y, y) : y Y} (({x
0
} F
i
) U(y
0
)) = / 0.
This fact implies at once that F
i
U(y
0
) = / 0 and, by virtue of the closedness of F
i
, we get
y
0
F
i
. Finally, we come to the relation
y
0
{F
i
: i I}
which contradicts the assumption that {F
i
: i I} = / 0.
This result is converse to the Kuratowski theorem on closed projections. It was obtained
by Mr owka [180].
4
. Let C[0, 1] denote the Banach space of all real-valued continuous functions on the unit
segment [0, 1] and let D be the family of all those functions from C[0, 1] which are differ-
entiable at least at one point of [0, 1].
Prove that D is a rst category subset of C[0, 1].
For this purpose, x a natural number n and consider the set
P
n
={( f , x) C[0, 1] [0, 1] : (h [x, 1 x])(| f (x +h) f (x)| n|h|)}.
Verify that P
n
is a closed subset of the product space C[0, 1] [0, 1] and that
D {pr
1
(P
n
) : n <}.
Applying Kuratowskis theoremon closed projections, check that every set pr
1
(P
n
) (n <)
is closed and, simultaneously, nowhere dense in C[0, 1], which yields the required result.
Conclude from the said above that the family of all nowhere differentiable functions from
C[0, 1] is residual (co-meager) in C[0, 1]. Consequently, there are many nowhere differen-
tiable functions in C[0, 1].
This important statement was independently established by Banach [9] and Mazurkiewicz
[171].
5. Consider two mappings f
1
: R R and f
2
: R R such that
f
1
(x) = x, f
2
(x) = 0 (x R).
Observe that both of them are closed and continuous. The product mapping
( f
1
, f
2
) : R
2
R
2
is dened by the formula
( f
1
, f
2
)(x, y) = ( f
1
(x), f
2
(y)) = (x, 0) ((x, y) R
2
)
and, in fact, coincides with the canonical projection
pr
1
: RRR.
Check that ( f
1
, f
2
) is not a closed mapping.
6. Generalize Theorem 4 to the case of an arbitrary family { f
i
: i I} of mappings with
closed graphs. Namely, show that if all f
i
(i I) have closed graphs, then the product
mapping { f
i
: i I} also has closed graph.
7
. Let X and Y be two topological spaces and let f : X Y be a continuous mapping such
that:
(a) f is surjective and closed;
(b) f
1
(y) is quasicompact for any point y Y.
Prove that the inequality w(Y) w(X) + holds true, where w(X) (respectively, w(Y))
denotes the topological weight of X (respectively, of Y).
For this purpose, choose an open base {U
i
: i I} of X such that card(I) = w(X). Assume,
without loss of generality, that I is an innite set. Let J denote the family of all nite
subsets of I. Clearly, card(J) = card(I). Verify that the family of all sets of the form
Y \ f (X \ {U
i
: i J}) (J J)
is an open base of Y. To see this circumstance, take an arbitrary point y Y and consider its
arbitrary open neighborhood W(y). According to (b), the set f
1
(y) is quasicompact and
is contained in the open set f
1
(W(y)). Consequently, there exists a set J J for which
we have
f
1
(y) {U
i
: i J} f
1
(W(y)),
thus it follows that
y Y \ f (X \ {U
i
: i J})
and the set Y \ f (X \ {U
i
: i J}) is open in Y. Also, verify that
Y \W(y) f (X \ f
1
(W(y))) f (X \ {U
i
: i J}),
which implies that
Y \ f (X \ {U
i
: i J}) W(y).
This inclusion at once yields the required inequality w(Y) w(X) +.
8. Deduce from the previous exercise that if X is a quasicompact topological space with
w(X) and Y is a Hausdorff continuous image of X, then Y is a compact metrizable
space.
In particular, a continuous image of a metrizable compact space is metrizable if and only if
it is Hausdorff.
9. Let E be a Polish topological space and let R be an equivalence relation in E such that
for any closed set F E, the set R(F) is also closed in E.
Show that the partition E/R canonically associated with R admits a Borel selector of type
F
.
10. Let E be a compact metric space and let R be an equivalence relation in E whose
graph is a closed subset of the product space E E. Deduce directly from the theorem
of Kuratowski and Ryll-Nardzewski (see [151] or Appendix 2) that there exists a Borel
selector of the partition E/R.
For this purpose, take into account the fact that the quotient space E/R is compact and
metrizable (see Exercise 8) and consider the canonical surjection
: E E/R.
Associate with this surjection a set-valued mapping
: E/R P(E)
dened by the formula
(z) =
1
(z) (z E/R).
Check that satises the assumptions of the theorem of Kuratowski and Ryll-Nardzewski
(where E/R is equipped with its Borel -algebra). Let
f : E/R E
be a Borel selector of . Keeping in mind that f is injective and applying the theorem that
an injective metrizable Borel image of a Borel subset of a Polish space is also a Borel set
(see [99], [148], [160], or Appendix 6), conclude that ran( f ) is the required Borel selector
of the partition E/R.
11. Let E be an arbitrary set. Show that the A-operation over any class L P(E) gen-
eralizes the standard operations of taking unions and intersections of countable families of
sets from L.
12
. Equip the set with its discrete topology and consider the product space
. As
usual,
is called the canonical Baire space of topological weight .

Prove that any nonempty Polish topological space E is a continuous image of
.
Moreover, apply the fact that any nonempty Suslin set in E is a continuous image of
(see Appendix 6) and infer from this fact that any Suslin set in E may be regarded as the
second projection of some closed subset of the Polish product space
E.
13
. Let E
1
be a topological space satisfying the following two conditions:
(a) each point of E
1
has a countable local base;
(b) every closed subset of E
1
is of type G
in E
1
.
Let E
2
be a complete metric space and let f : E
1
E
2
be a partial continuous mapping.
Show that there exists a partial continuous mapping f
: E
1
E
2
extending f and de-
ned on a G
-subset of E
1
(Lavrentievs theorem [157] on extensions of partial continuous
functions).
Argue in the following manner. For any point x cl(dom( f )), let B(x) denote a local base
of x and let
f
(x) = inf{diam( f (U)) : U B(x)},
where diam( f (U)) denotes the diameter of f (U).
The real number
f
(x) 0 is usually called the oscillation of f at x.
Check that for every point x dom( f ), the equality
f
(x) = 0 holds true. Further, put
X ={x cl(dom( f )) :
f
(x) = 0}
and show that the set X is of type G
in E
1
.
Now, let x X and let {x
n
: n < } dom( f ) be a sequence converging to x. Verify that
{ f (x
n
) : n < } is a Cauchy sequence in E
2
and hence converges to some point y E
2
.
Putting f
(x) = y, prove that the function

f
: X E
2
is well dened in this manner, which extends f and is continuous on X.
14
. Let E
1
and E
2
be two complete metric spaces and let f : E
1
E
2
be a partial home-
omorphism. By using the result of the previous exercise, show that there exists a partial
homeomorphism f
: E
1
E
2
extending f and such that both sets dom( f
) and ran( f
)
are of type G
in E
1
and in E
2
respectively (Lavrentievs theorem [157] on extensions of
partial homeomorphisms).
For this purpose, denote g = f
1
. In view of Exercise 13, there are two partial continuous
mappings
f
: E
1
E
2
, g
: E
2
E
1
extending f and g, respectively, and dened on some sets of type G
. In the product space

E
1
E
2
consider the set
Z ={(x, y) : y = f
(x)} {(x, y) : x = g
(y)}.
Verify that dom( f ) pr
1
(Z) and ran( f ) pr
2
(Z). Moreover, check that both sets pr
1
(Z)
and pr
2
(Z) are of type G
and the mapping f
|pr
1
(Z) establishes a homeomorphism be-
tween pr
1
(Z) and pr
2
(Z).
15. Let E
1
be a metric space, E
2
be a complete metrizable space and let f : E
1
E
2
be
a partial homeomorphism. By using the previous exercise, show that there exists a partial
homeomorphism f
: E
1
E
2
extending f and dened on some G
-subset of E
1
. Note
that, in general, we cannot assert that the set ran( f
) is of type G
in E
2
.
16
. Let E
1
be a metric space, E
2
be a Polish topological space and let f : E
1
E
2
be a
partial Borel mapping. Prove that there exists a Borel mapping f
: E
1
E
2
extending f .
For this purpose, start with Lavrentievs theorem on extensions of partial continuous func-
tions (see Exercise 13), take into account the Baire classication of Borel mappings (see
Example 6 from Chapter 1 and Appendix 6) and then use the method of transnite induc-
tion on <
1
.
17. Let E
1
and E
2
be two Polish topological spaces and let f : E
1
E
2
be a partial Borel
isomorphism. Prove that there exists a partial Borel isomorphism f
: E
1
E
2
extending
f and such that dom( f
) is a Borel subset of E
1
and ran( f
) is a Borel subset of E
2
.
For this purpose, apply the result of Exercise 16 and use an argument similar to that of
Exercise 14.
18
. Let E be an innite set and let be a family of partial injections acting from E into
itself, such that
card() = card(E), ( )(card(dom()) = card(E)).
Prove that there exists a family {X
i
: i I} of subsets of E satisfying the following condi-
tions:
(a) card(I) = 2
card(E)
;
(b) for any two distinct indices i I and j I and for each function , we have
(X
i
) \ X
j
= / 0.
In order to show this fact, denote by the least ordinal number of cardinality card(E). Let
{x
: <} be an enumeration of all points in E and let {
: <} be an enumeration
of all those partial functions which either belong to or are inverse to a member of . For
every <, dene by transnite recursion a point
p
E \ ({p
: <} {
(p
) : <, <})
and put P = {p
: < }. Further, verify that there exists a family {X

i
: i I} of subsets
of P satisfying condition (a) and such that
card(X
i
\ X
j
) = card(E) (i I, j I, i = j).
Check that this {X
i
: i I} is the required family.
19. Let E be an uncountable Polish topological space and let X and Y be two subspaces of
E. We shall say that these subspaces are incomparable in the descriptive sense if there exists
no Borel isomorphism of X onto some subset of Y and there exists no Borel isomorphism
of Y onto some subset of X.
By using the results of Exercise 17 and Exercise 18, prove that there are 2
c
subspaces of E
which are pairwise incomparable in the descriptive sense. In particular, conclude that there
are 2
c
Borel types of subspaces of E.
20. Give an example of a partial homeomorphismg : RR such that both sets dom(g) and
ran(g) are homeomorphic to the Cantor space {0, 1}
but there exists no homeomorphism

g
: R R extending g.
21
. Let f be a function acting from R into R and let x be a point of R.

We shall say that x is a pleasant point with respect to f (or, briey, f -pleasant point) if for
each real number > 0, there exists a neighborhoodV(x) =V(x, ) of x such that the set
{y R : | f (y) f (x)| <}
is categorically dense in V(x), i.e., the intersection of this set with any nonempty open
interval contained in V(x) is of second category in that interval.
In accordance with the denition above, we shall also say that a point x R is unpleasant
with respect to f (or, briey, f -unpleasant point) if x is not f -pleasant.
Show that, for any function f : RR, the set of all f -unpleasant points is of rst category
in R; consequently, the set of all f -pleasant points is co-meager in every nonempty open
subinterval of R.
For this purpose, suppose otherwise, i.e., suppose that the set
A ={x R : x is unpleasant with respect to f }
is not of rst category. For each point x A, there exists a strictly positive number (x)
such that for any neighborhood V of x, the set
{y R : | f (y) f (x)| <(x)}
is not categorically dense in V. Pick two rational numbers r(x) and s(x) satisfying the
inequalities
f (x) (x)/2 < r(x) < f (x) < s(x) < f (x) +(x)/2.
Further, for any pair (r, s) of rational numbers, put
A
r,s
={x A : r(x) = r, s(x) = s}.
Evidently, the equality
A =
(r,s)QQ
A
r,s
holds true. By the assumption, A is not of rst category, so there exists a pair (r
0
, s
0
)
QQ such that the set A
r
0
,s
0
is not of rst category, either. Consequently, there exists a
nonempty open interval ]a, b[ R such that the set A
r
0
,s
0
is categorically dense in ]a, b[.
Choose any point
x
0
]a, b[ A
r
0
,s
0
.
For this point x
0
, we may write
f (x
0
) (x
0
)/2 < r
0
< f (x
0
) < s
0
< f (x
0
) +(x
0
)/2.
Analogously, for each point y A
r
0
,s
0
, we have
f (y) (y)/2 < r
0
< f (y) < s
0
< f (y) +(y)/2.
Therefore,
| f (y) f (x
0
)| < s
0
r
0
<(x
0
).
In other words, the inclusion
A
r
0
,s
0
{y R : | f (y) f (x
0
)| <(x
0
)}
is valid, which immediately implies that the set
{y R : | f (y) f (x
0
)| <(x
0
)}
is categorically dense in ]a, b[ contradicting the denition of (x
0
). The obtained contra-
diction yields the desired result.
22
. Let f be an arbitrary function acting from R into R. Prove that there exists an every-
where dense subset X of R such that the function f |X is continuous (Blumbergs theorem
[15]).
Argue in the following manner. Since R is homeomorphic to the open unit interval ]0, 1[, it
sufces to establish Blumbergs statement for any function
f : ]0, 1[ ]0, 1[.
Let Gr( f ) denote the graph of f . Taking into account Exercise 21, recursively construct
two sequences
{Z
n
: n N}, {X
n
: n N},
(a) for each natural number n, the set Z
n
Z
n
={]a
i
, b
i
[]c
i
, d
i
[ : i I(n)},
where I(n) is a countable set, all intervals of the family {]a
i
, b
i
[ : i I(n)} are contained in
]0, 1[ and are pairwise disjoint, the set pr
1
(Z
n
Gr( f )) is categorically dense in ]0, 1[ and
for any i I(n), the length of ]c
i
, d
i
[ is strictly less than 1/(n +1);
(b) for each natural number n, the set X
n
is a nite (1/(n +1))-net of ]0, 1[, i.e., for any
point t from ]0, 1[, there exists a point x X
n
such that |t x| < 1/(n +1);
(c) the sequence of sets {Z
n
: n N} is decreasing by inclusion;
(d) the sequence of sets {X
n
: n N} is increasing by inclusion;
(e) for each natural number n, the relation X
n
pr
1
(Z
n
Gr( f )) holds and any point from
X
n
is f -pleasant.
Afterwards, put
X ={X
n
: n N}
and verify, by using the conditions (a), (c), (d), (e), that X is everywhere dense in R and the
restriction of f to X is continuous.
23. We say that a topological space E is a Blumberg space if for any function g : E R,
there exists an everywhere dense subset X of E such that the restriction g|X is continuous.
Check that any Blumberg space is necessarily a Baire topological space.
A useful survey of results concerning Blumbergs theorem, its generalizations, and Blum-
berg spaces is presented in paper [28]. See also [27], [29], and [241].
24
. We shall say that a metrizable topological space E is topologically complete if there

exists a complete metric space homeomorphic to E.
Show that any open subset U of a topologically complete metric space (E, d) is also topo-
logically complete.
For this purpose, assume that U = E and consider in the complete product space E R the
set
K ={(x, t) : td(x, E \U) = 1}.
Verify that this K satises the following relations:
(a) K is a closed subset of E R and hence K is topologically complete;
(b) K is homeomorphic to U.
Conclude from the relations (a) and (b) the topological completeness of U.
Further, suppose that {V
n
: n < } is a countable family of topologically complete sub-
spaces of a metrizable topological space E. Show that the space {V
n
: n < } is also
topologically complete.
For this purpose, consider the product space V = {V
n
: n < } which is topologically
complete and take in this space the diagonal
=V {z E
: (n <)(m <)(pr
n
(z) = pr
m
(z))}.
Verify that is a closed subset of V and the space {V
n
: n <} is homeomorphic to .
Taking into account the above-mentioned facts and Lavrentievs theorem on extensions of
homeomorphisms, prove that for a metrizable space E, the following two assertions are
equivalent:
(c) E is topologically complete;
(d) E is homeomorphic to a G
-subset of some complete metric space.

Finally, by using (d) and the topological universality of [0, 1]
in the class of all separable

metrizable spaces, showthat Polish spaces can be topologically characterized as G
-subsets
of [0, 1]
.
In this connection, recall that the equivalence between (c) and (d) was rst established by
Alexandrov (see, for instance, [58] and [148]).
25
. Suppose that the inequality 2
< 2
1
is satised (for example, this inequality is ful-
lled if the Continuum Hypothesis holds true). Let E be a separable metric space, E
be a
metrizable space, and let g : E E
be a Borel mapping.
Demonstrate that g(E) is a separable subspace of E
.
For this purpose, apply the fact that every nonseparable metrizable space contains an un-
countable discrete subspace.
Let now P be a Polish topological space or, more generally, a Suslin subset of a Polish
topological space and let f : P E
be a Borel mapping.
Demonstrate within ZFC theory that f (P) is a separable subspace of E
.
For this purpose, utilize the result of Exercise 12 and take into account the fact that any
uncountable Suslin set is of cardinality continuum(see [10], [99], [148], [160], or Appendix
6).
Chapter 9
Weakly metrically transitive measures and
nonmeasurable sets
In the previous chapters we were concerned with various extensions of -nite measures,
which are obtained by using certain types of nonmeasurable sets or nonmeasurable func-
tions. It was also mentioned that sometimes absolutely nonmeasurable sets can occur for
concrete classes of measures (e.g., Vitali sets for the class of all translation-invariant exten-
sions of the Lebesgue measure =
1
on the real line R = R
1
). Obviously, the latter sets
turn out to be useless from the point of view of the measure extension problem.
In the present chapter, we continue our consideration of absolutely nonmeasurable sets
(for nonzero -nite invariant measures) and show their connections with the notion of a
weakly metrically transitive measure. The precise denition of weakly metrically transitive
measures will be introduced later.
In 1896 Minkowski published his famous work [173] in which he suggested a geometric
approach to classical problems of number theory and extensively developed beautiful geo-
metric methods in this theory. The main role in his methods was played by convex subsets
of Euclidean spaces. Among many other important results, Minkowski proved in [173] the
following fundamental statement.
Minkowskis Theorem. If C is a compact convex body in the Euclidean space R
n
, sym-
metric with respect to the origin of R
n
, and having volume greater than or equal to 2
n
,
then C contains at least two nonzero points of the lattice Z
n
where Z denotes the set of all
integers.
This statement was then generalized in various directions (see, e.g., [76], [77] and the
references therein).
The argument used by Minkowski in the proof of his theorem is purely group-theoretical
and measure-theoretical. Actually, by applying a similar argument, it can be established
that there exists a subset of R
n
nonmeasurable in the Lebesgue sense (see Theorem 2 be-
145
9
low). In this context, it is interesting to notice that the rigorous concept of the Lebesgue
measure was introduced some years later after Minkowskis theoremwas stated, namely, in
the beginning of the twentieth century (see [158]). Moreover, it should be recalled that the
existence of subsets of R
n
which are nonmeasurable in the Lebesgue sense was rst proved
by Vitali [239] only in 1905.
An analogous astonishing situation can be observed in connection with the famous Poincar e
theorem on recurrent points for an abstract space E equipped with a nite measure and
with a single transformation g : E E which preserves , that is we have
g
1
(X) dom() & (g
1
(X)) = (X)
for all sets X dom().
The above-mentioned theorem was proved by Poincar e before the basic concepts of mea-
sure theory (e.g., the countable additivity of measures) were introduced in real analysis (cf.
[192]). The precise formulation of the Poincar e theorem is given in Exercise 2 and its proof
is also outlined therein.
The main goal of this chapter is to demonstrate the role of Minkowskis method in estab-
lishing the existence of nonmeasurable subsets of Euclidean spaces and, more generally,
in establishing the existence of nonmeasurable sets in an abstract space equipped with a
transformation group and with some measure which is invariant under that group. In the
sequel, we will introduce the concept of a weakly metrically transitive invariant measure
and will show how such a measure produces absolutely nonmeasurable sets with respect to
the class of all its invariant extensions.
We begin with several auxiliary notions concerning spaces endowed with transformation
groups.
Let E be a nonempty set, G be a group of transformations of E and let be a complete
measure dened on a G-invariant -algebra of subsets of E and invariant under all trans-
formations fromG. In this case, we briey say that is a G-invariant measure (see Chapter
2).
Let X be some subset of E.
We shall say that X is G-thick with respect to if there exists a countable subset (equiva-
lently, countable subgroup) H of G such that
(E h(X) : h H) = 0.
We shall say that X is G-thin with respect to if
(g G)(h G)(g ,= h (g(X) h(X)) = 0).
Weakly metrically transitive measures and nonmeasurable sets 147
In particular, if a group G acts freely in a space E, then every singleton is G-thin in E.
Example 1. If E = R
2
and G =0 R, then the graph Gr( f ) of any partial function
f : R R
is a G-thin subset of E. On the other hand, it also may happen that Gr( f ) is R
2
-thick
with respect to some translation-invariant extension of the two-dimensional Lebesgue
measure
2
(in this connection, see Chapter 19).
Moreover, it may happen that there exists a countable family g
n
: n < of isometric
transformations (i.e., motions) of R
2
such that
R
2
=g
n
(Gr( f )) : n <.
For more details concerning the latter intriguing fact, see [49], [51], [222], and Exercise 8
of Appendix 1.
Example 2. Suppose that is a -nite G-invariant measure on E metrically transitive
(ergodic) with respect to G, i.e., for any -measurable set Z, the relation
(g G)((g(Z)Z) = 0)
implies the disjunction
(Z) = 0 (E Z) = 0.
Then it can easily be checked that every -measurable set X with (X) > 0 is G-thick in
E and this property of is equivalent to its metrical transitivity with respect to the group
G. In this context, let us recall that the left (right) Haar measure on a -compact locally
compact topological group (H, ) is metrically transitive with respect to the group of all left
(right) translations of H (see Exercise 17 for Chapter 3).
Example 3. Suppose that a group G of transformations of a given space E is countable
and suppose, in addition, that G acts almost freely in E with respect to some G-invariant
measure on E, i.e., for any two distinct transformations g G and h G, we have
(x E : g(x) = h(x)) = 0,
where
stands for the outer measure associated with .

Consider the partition of E into all G-orbits of the points of E. Let X be an arbitrary
selector of this partition (in short, G-selector). Then it is not difcult to verify that X is
simultaneously G-thick and G-thin in E.
The following statement is essentially due to Minkowski (cf. [173]).
Theorem 1. Let (E, G) be a space equipped with a transformation group and let be a
-nite G-invariant measure on E. Let X be a G-thick subset of E and let Y be a G-thin
subset of E. If both of these sets are measurable with respect to , then (Y) (X).
Proof. Only two cases are possible.
(a) The group G is uncountable. In this case there is nothing to prove because the -
measurable G-thin subset Y of E is necessarily of -measure zero in view of the -
niteness and G-invariance of and the almost disjointness of the family g(Y) : g G.
(b) The group G is countable. In this case, we obviously have
(E g(X) : g G) = 0
and, consequently,
(Y) = (Y (g(X) : g G))
gG
(g(X) Y) =
gG
(X g(Y)) (X),
which yields the required result. Theorem 1 has thus been proved.
Let (E, G) be a space equipped with a transformation group and let be a -nite G-
invariant (or, more generally, G-quasi-invariant) measure on E.
We shall say that is weakly metrically transitive (or weakly ergodic) with respect to G
if for any > 0, there exist a -measurable set X with (X) < and a countable subset
(equivalently, countable subgroup) H of G such that
(E h(X) : h H) = 0.
It is not hard to verify that any -nite non-atomic G-invariant metrically transitive mea-
sure on E is weakly metrically transitive. The converse assertion is not true in general.
Indeed, it directly follows from the above denition that every G-invariant extension of a
G-invariant weakly metrically transitive measure is also weakly metrically transitive. An
analogous statement fails to be true for G-invariant metrically transitive measures.
Example 4. Let us consider the concrete particular case, where:
E = the real line R,
G = the group of all translations of R,
= the Lebesgue measure on R (i.e., =).
Obviously, we may identify G with the additive group R. It is well known that is metri-
cally transitive with respect to the group QR of all rational numbers.
By applying the standard transnite methods (see, e.g., [83], [85], [95], [108], [115], [176],
[192], [198], or Exercise 4 from Chapter 6), it can be proved that there exist two subsets A
and B of R satisfying the following conditions:
(1) AB = / 0 and AB = R;
(2) card(A) = card(B) = c, where c denotes the cardinality of the continuum;
(3)
(A) =
(B) = 0;
(4) both sets A and B are almost R-invariant, that is for all h R, we have
card((h +A)A) < c, card((h +B)B) < c.
By using these two sets, the original measure can be extended to an R-invariant measure
/
which fails to be metrically transitive with respect to the group R (hence, with respect
to the group QR). For this purpose, it sufces to dene on the -algebra
S =(A, B dom())
the functional
/
by the formula
/
((AX) (BY)) = (1/2)((X) +(Y)),
where X and Y are arbitrary elements from dom(). This denition of
/
is correct and
/
is a measure extending (cf. Chapter 2).
Applying Marczewskis method described in Chapter 2, we can also suppose that all subsets
of R, whose cardinalities are strictly less than c, belong to dom(
/
). Now, it is clear that
/
is R-invariant and, simultaneously, is not metrically transitive in view of the almost
invariance of the
/
-measurable sets A and B which both are of strictly positive
/
-measure;
moreover, we have
/
(A) =
/
(B) = +.
At the same time,
/
as an extension of is weakly metrically transitive.
Example 5. Suppose that a -nite G-invariant metrically transitive measure on E has
at least one atom. Then it is not difcult to describe the structure of . Namely, in this case
the space E admits a representation
E =A
i
: i I
for which the following relations are satised:
(1) the set I is at most countable and the family A
i
: i I is disjoint;
(2) each set A
i
(i I) is an atom of ;
(3) for any i I and j J, there exists a transformation g G such that
(g(A
i
)A
j
) = 0;
(4) for any transformation h G, the family h(A
i
) : i I almost coincides with the family
A
i
: i I, i.e., each member of the rst family almost coincides with some member of the
second family, and conversely.
In other words, we obtain a certain G-invariant lattice of atoms of our measure , similar
to the standard lattice z +[0, 1]
n
: z Z
n
of the Euclidean space R
n
.
Let (E, G) be a space equipped with a transformation group and let M be some class of
G-invariant (G-quasi-invariant) measures on E. We would like to stress that, in general, the
measures from M are dened on different -algebras of subsets of E.
Recall (see Chapter 5) that a set X E is absolutely nonmeasurable with respect to M if
there exists no measure from M for which X is measurable.
Now, let be a G-invariant (G-quasi-invariant) measure on E. We shall say that a set
X E is G-absolutely nonmeasurable with respect to if X is absolutely nonmeasurable
with respect to the class of all G-invariant (G-quasi-invariant) measures on E extending .
The next example highlights close relationships between convex sets and absolutely non-
measurable sets.
Example 6. Let E be an innite-dimensional separable Banach space, G be the group
of all translations of E, and let B be a bounded convex body in E (i.e., B is a closed
bounded convex set in E with nonempty interior). It can be proved that B is absolutely
nonmeasurable with respect to the class of all nonzero -nite G-invariant (more generally,
G-quasi-invariant) measures on E (see [109] and Exercise 10 for this chapter). The above
fact implies, in particular, that E does not admit a nonzero -nite G-quasi-invariant Borel
measure (cf. Chapters 3 and 4). By using an algebraic isomorphism between the additive
groups E and R, it can also be shown that there exists a subset of R which is absolutely
nonmeasurable with respect to the class of all nonzero -nite R-invariant (more generally,
R-quasi-invariant) measures on R.
Example 7. As in Example 4, let us put:
G = the group of all translations of R,
We remember that any selector of R/Q is called a Vitali subset of R (see [10], [40],
[72], [90], [143], [148], [150], [176], [183], [192], [239], [240], or Chapter 5). It is well
known that all Vitali sets are R-absolutely nonmeasurable with respect to (see the above-
mentioned works and Exercise 2 from Chapter 5).
Theorem 2. Let G be a countable group of transformations of E and let be a nonzero
-nite G-invariant weakly metrically transitive measure on E. Suppose also that G acts
almost freely in E with respect to . Then every G-selector is G-absolutely nonmeasurable
with respect to .
Proof. Take an arbitrary G-selector X in E. We should verify that for any G-invariant
measure
/
on E extending the original measure , the set X is not
/
-measurable. Suppose
otherwise, i.e., X dom(
/
) for some G-invariant extension
/
of . Then, in view of the
countability of G and of the equality
E =g(X) : g G,
we must have
/
(X) > 0. Let us denote
/
(X) = . According to our assumption, there
exist a -measurable set Y with (Y) < and a countable family H G such that
(E h(Y) : h H) = 0.
In other words, the set Y turns out to be a G-thick subset of E with respect to . Since
/
extends , the same Y is G-thick with respect to
/
as well. On the other hand, the selector
X is a G-thin subset of E with respect to and hence with respect to
/
(see Example 3).
By virtue of Theorem 1, the relation
=
/
(X)
/
(Y) = (Y) <
should be valid, which yields a contradiction. This contradiction completes the proof of
Theorem 2.
Remark 1. Clearly, Theorem 2 may be regarded as a generalization of Vitalis theorem
[239] stating the existence of subsets of R nonmeasurable in the Lebesgue sense. Indeed,
put in Theorem 2:
G = the group Q of all rational numbers,
Then we directly come to Vitalis classical result on the existence of -nonmeasurable sets
in R. Therefore, Theorem 2 may be treated as an abstract version of the above-mentioned
classical result. Some other abstract versions of Vitalis theorem are presented in [115],
[119], [195], [229], and [230] (cf. also [251]).
In order to demonstrate once more that the notion of the weak metrical transitivity is closely
connected with the existence of absolutely nonmeasurable sets, we need two auxiliary state-
ments. The rst of them deals with Borel selectors of a certain partition of a Polish topo-
logical space (see, for instance, [24] and Chapter 8).
Lemma 1. Let E be a Polish topological space and let
R(x, y) (x E, y E)
be an equivalence relation on E satisfying the following conditions:
(1) each equivalence class with respect to R(x, y) is a closed subset of E;
(2) for any closed set F E, the set
R(F) =y E : (x)(x F & (x, y) R)
is a Borel subset of E.
Let E/R denote the quotient set associated with the equivalence relation R. Then there
exists a Borel selector of E/R.
A detailed proof of Lemma 1 was given in Chapter 8.
Lemma 2. Let (G, ) be a -compact complete metrizable topological group and let H be
a closed subgroup of G. Then the family of all right (left) translates of H admits a Borel
selector.
Proof. Notice rst that the given metrizable group G is separable as a union of countably
many compact and hence separable subspaces. Moreover, applying the classical Baire
theorem, we readily infer that G is locally compact. Denote by the symbol G/H the family
of all right translates of H. Since H is a closed subgroup of G, the elements of G/H are
also closed in G. In view of Lemma 1, it remains to verify that for any closed set F G,
the set
Hx : x G, F Hx ,= / 0 = H F
is Borel in G. Indeed, the sets H and F are closed subsets of a -compact space, so they
are -compact as well. Dene a continuous mapping
: GGG
by putting
(x, y) = x y (x G, y G).
Obviously, the set H F coincides with the image of H F with respect to the mapping
, which immediately implies that H F is representable as the union of countably many
compact subsets of G. Therefore, H F is Borel in G.
Remark 2. As mentioned above, any -compact complete metrizable topological group
(G, ) is a locally compact Polish group (equivalently, is a locally compact group with a
countable base). Also, it is well known that if a Polish group (G, ) admits a nonzero
-nite Borel measure invariant under all left (right) translations of G, then G is -
compact and locally compact, and coincides with the left (right) Haar measure on G (in
this connection, see Chapters 3 and 4).
Now, we are ready to establish the following statement.
Theorem 3. Let (G, ) be an uncountable locally compact Polish group, let denote the
left Haar measure on G, and let H be a countable subgroup of G. We consider H as a
group of transformations of G. Namely, each h H is identied with the mapping
h : G G
(we preserve the same notation h) dened by the standard formula
h(x) = h x (x G).
In this way, we come to the partition of G into all left H-orbits.
The following three assertions are equivalent:
(a) is weakly metrically transitive with respect to H;
(b) every H-selector in G is absolutely nonmeasurable with respect to ;
(c) H is non-discrete in G.
Proof. The implication (a) (b) is a straightforward consequence of Theorem 2.
Let us prove the implication (b) (c). Suppose that a subgroup H of the given group G
is discrete. Then it can easily be veried that H is countable and closed in G. Applying
Lemma 2, we infer that there exists a Borel H-selector in G. This selector is -measurable
and, therefore, it cannot be absolutely nonmeasurable with respect to our . We thus have
shown that the implication (c) (b) is valid. Consequently, we also have (b) (c).
Finally, let us establish the implication (c) (a). Suppose that (c) is fullled and put
P = cl(H),
where cl() as usual denotes the closure operator. Evidently, P is an uncountable closed
subgroup of G. Again, according to Lemma 2, there exists a Borel selector X of the family
G/P of all right translates of P. Since there are uncountably many pairwise disjoint left
translates of X, we must have (X) = 0. Also, it is clear that
P X = G.
Fix > 0. Since (X) = 0, there exists an open set U G such that X U and (U) <.
Obviously,
P U = G.
Taking into account that U is open and H is everywhere dense in P, we readily deduce that
h(U) : h H = H U = G,
which yields the weak metrical transitivity of with respect to H, i.e., assertion (a) holds
true. Thus, we have shown that the implication (c) (a) is valid, which nishes the proof
of the theorem.
Remark 3. It is not difcult to verify that the implication (c) (b) remains true for an
arbitrary uncountable -compact locally compact topological group (G, ) endowed with
the left Haar measure . For the sake of completeness, we will present an easy proof
of this fact, applying the classical argument of Vitali [239] to G. Let H be a countable
non-discrete subgroup of G. Such subgroups always exist in G and a proof of this simple
auxiliary assertion is left to the reader. Pick any H-selector X in G. Suppose to the contrary
that there exists a left H-invariant extension
/
of such that X dom(
/
). Taking into
account the equality
hX : h H = G,
we must have
/
(X) > 0. Further, since G is -compact, there is a compact set K G
satisfying the relation
0 <
/
(X K)
/
(K) = (K) < +.
For the sake of brevity, denote
Y = X K.
Let V be a compact neighborhood of the neutral element of G. Since H is non-discrete,
the set HV is innite. Let us represent the last set as an injective sequence h
n
: n <.
Clearly, the following two relations are satised:
(1) h
n
Y h
m
Y = / 0 whenever n <, m < and n ,= m;
(2) h
n
Y : n < V K.
Relation (1) directly implies that
/
(h
n
Y : n <) =

n<
/
(h
n
Y) = +.
On the other hand, by virtue of the compactness of V K, relation (2) implies
/
(h
n
Y : n <)
/
(V K) = (V K) < +,
which yields an obvious contradiction. The obtained contradiction gives us the required
result.
In a similar way Theorem 2 can be applied to the n-dimensional Lebesgue measure
n
on
the Euclidean space R
n
(n 1). Indeed, we may consider
n
as an invariant measure with
respect to an arbitrary countable group H of isometric transformations (i.e., motions) of
R
n
.
Notice that some deep properties of
n
treated as an H-invariant measure on R
n
are in-
vestigated in paper [142]. In the same paper various H-invariant extensions of
n
are also
studied.
For our purpose, we need several well-known facts fromthe geometry of Euclidean spaces.
(i) Let x be a point of R
n
and let G be a family of motions of R
n
. If the set
G(x) =g(x) : g G
is relatively compact in R
n
, then the family G is relatively compact in the topological group
of all motions of R
n
. Consequently, if G is a closed subset of the group of all motions of
R
n
, then the set G(x) is closed in R
n
.
(ii) Let F be a closed subset of R
n
and let G be a closed subset of the group of all motions
of R
n
. Then the set
G(F) =g(x) : g G, x F
admits a representation in the form of a countable union of compact subsets of R
n
. There-
fore, G(F) is of type F
in R
n
and hence is Borel in R
n
.
(iii) The group of all motions of R
n
acts almost freely in R
n
with respect to
n
. Indeed, for
any two distinct motions g and h of R
n
, the set
x R
n
: g(x) = h(x)
is contained in some afne hyperplane of R
n
and, consequently, this set is of
n
-measure
zero.
(iv) Let G be an uncountable group of motions of R
n
and let X be a G-selector. Then either
X is nonmeasurable with respect to
n
or
n
(X) = 0.
Notice that (iv) readily follows from (iii). More generally, if an uncountable group G of
transformations of a base set E acts almost freely in E with respect to a -nite G-quasi-
invariant measure on E, then any G-selector is either -nonmeasurable or has -measure
zero.
Taking these facts into account, we can formulate for
n
a direct analogue of Theorem 3.
Namely, we have
Theorem 4. Let H be a countable group of motions of the Euclidean space R
n
. The
following three assertions are equivalent:
(a)
n
is weakly metrically transitive with respect to H;
(b) every H-selector in R
n
is absolutely nonmeasurable with respect to
n
;
(c) H is non-discrete in the group of all motions of R
n
.
The proof of this statement is similar to the proof of Theorem 3, and all corresponding
details are left to the reader.
For further information closely connected with the topic of this chapter, see [39], [40], [46],
[60], [83], [84], [104], [107], [115], [119], [164], [176], [229], [230], [249], and [251].
EXERCISES
1
. Prove Minkowskis theoremon convex bodies, formulated in the beginning of the chap-
ter.
For this purpose, consider the set C/2 where C is a given compact convex body in the
Euclidean space R
n
, whose volume
n
(C) is strictly greater than 2
n
, and verify that C/2 is
not Z
n
-thin in R
n
.
Then reduce the case of a compact convex body C R
n
with
n
(C) = 2
n
to the previous
case by using an appropriate limit process.
2
. Let E be a set equipped with a nite measure and let g : E E be a transformation

of E preserving , i.e., for any set X dom(), we have
g
1
(X) dom(), (g
1
(X)) = (X).
Let k be a natural number and let A be any -measurable set. We denote
R
k
(A) =x A : (n [k, [)( f
n
(x) A).
Prove that
(A R
k
(A)) = 0.
Deduce from this fact that
(A R
k
(A) : k <) = 0.
Conclude that -almost all points of A innitely many times return to A via appropriate
iterations of g.
This result is the classical Poincar e theorem on recurrence.
3. Let (E
1
, G
1
) and (E
2
, G
2
) be two spaces equipped with groups of transformations G
1
and G
2
respectively. Let
1
be a -nite G
1
-quasi-invariant measure on E
1
and let
2
be a
-nite G
2
-quasi-invariant measure on E
2
. Suppose that
1
is weakly metrically transitive
with respect to G
1
and
2
is weakly metrically transitive with respect to G
2
. Show that the
product measure
1
2
is weakly metrically transitive with respect to the product group
G
1
G
2
.
6
. Let E be a Polish topological space and let Z

i
: i I be a countable family of Borel
subsets of E. Prove that the following two assertions are equivalent:
(a) Z
i
: i I separates the points in E, i.e., for any two distinct points x E and y E,
there exists an index i I such that
card(Z
i
x, y) = 1;
(b) the -algebra generated by Z
i
: i I coincides with the Borel -algebra B(E) of E.
For establishing this equivalence, use Marczewskis characteristic function of the given
family Z
i
: i I and the classical result from descriptive set theory stating that if f : E
E
/
is a Borel injection into a Polish space E
/
, then f (X) is Borel in E
/
whenever X is a
Borel subset of E (see [99], [148], and Appendix 6).
7. Let (E, [[ [[) be a real Banach space, X be a bounded convex body in E and let G be
a countable everywhere dense subgroup of the additive group E. Applying the result of
the preceding exercise, show that the countable family g(X) : g G generates the Borel
-algebra B(E).
8. Let (E, [[ [[) be an innite-dimensional Banach space and let K be a -compact subset of
E. Prove that there exists a nonzero vector e E such that every straight line in E parallel
to e intersects K in at most one point (in other words, K is uniform in direction e). Deduce
from this fact that K is an E-absolutely negligible subset of E.
9. Let (E, [[ [[) be an innite-dimensional separable Banach space. Using the previous
exercise, infer that there is no nonzero -nite E-quasi-invariant Borel measure on E.
For this purpose, apply the fact that any -nite Borel measure on E is Radon (cf. also
Chapter 4 where a more general result is presented).
10
. Let E be an innite-dimensional separable Banach space.

Prove that any bounded convex body in E is an E-absolutely nonmeasurable subset of E.
Prove the same fact for an arbitrary open bounded convex subset of E.
For this purpose, take into account the results of Exercises 7 and 9.
11. Give detailed proofs of the statements (i)-(iv) of this chapter.
12. By using the statements (i)-(iv), show the validity of Theorem 4.
13. Let G be a metrizable topological group. Prove that the following three assertions are
equivalent:
(a) G is topologically complete and -compact;
(b) G is separable and locally compact;
(c) G is Polish and locally compact.
Check that the topological vector space R
(N)
of all eventually nite real-valued sequences
is -compact and metrizable, but the same vector space R
(N)
is not complete with respect
to any metrizable topology compatible with the vector structure of this space.
For this purpose, rst observe that R
(N)
is representable as the union of countably many
nite-dimensional vector subspaces, all of which are nowhere dense in R
(N)
. Then take
into account the Baire theorem on category for complete metrizable topological spaces.
Chapter 10
Nonmeasurable subgroups of uncountable
solvable groups
It is well known that basic ideas and concepts of modern theory of dynamical systems
come from the classical theory of ordinary differential equations. Indeed, for a given au-
tonomous system of rst-order ordinary differential equations in a nite-dimensional phase
space (e.g., in R
n
), the famous Liouville theorem establishes necessary and sufcient con-
ditions, under which the corresponding phase ow preserves the Lebesgue measure (see,
for instance, [196] and Exercise 2 of Chapter 3). Those conditions are trivially fullled
for a Hamiltonian system in the same phase space. Consequently, we obtain a natural ex-
ample of a dynamical system which is systematically used in studies of various physical
(primarily, evolutional) processes.
As is known, an abstract dynamical system can be treated as a triplet (E, G, ), where E
is a nonempty phase space, G is some group of transformations of E, and is a nonzero
nite (or -nite) G-invariant measure dened on some G-invariant -algebra of subsets
of E.
Many deep and important results were established in terms of (E, G, ). Among them,
let us especially recall the Poincar e recurrence theorem (see Exercise 2 for Chapter 9),
the Birkhoff ergodic theorem (see [47] and [81]), the Krylov-Bogoliubov theorem on the
existence of a dynamical system for a one-parameter group G of homeomorphisms of a
nonempty compact metric space E (see Chapter 3). The last result was essentially gener-
alized by Markov and Kakutani to the case of a solvable group G of homeomorphisms of
a nonempty compact space E (for more details, see again Chapter 3). This generalization
shows, in particular, the remarkable role of solvable transformation groups in the theory of
dynamical systems and, equivalently, in the theory of invariant measures. Additionally, the
classical results of Banach and von Neumann should be mentioned, which state that if G is
a solvable group of isometric transformations of the Euclidean space R
n
, then there always
exists a nonnegative nitely additive G-invariant functional dened on the family P(R
n
)
DOI 10.1007/978-94-91216-36-7_1 2009 Atlantis Press/World Scientific
159
0,
of all subsets of R
n
and extending the Lebesgue measure
n
(see [22], [57], [83], [155],
and [240]).
The main goal of the present chapter is to discuss the measure extension problem for
uncountable solvable groups equipped with nonzero -nite left invariant (left quasi-
invariant) measures. Namely, it will be demonstrated below that special algebraic prop-
erties of an uncountable solvable group (G, ) enable us to extend a given nonzero -nite
left invariant (left quasi-invariant) measure on G by using some subgroups of G which
are nonmeasurable with respect to .
Let (G, ) be an arbitrary group. As usual, in our further considerations, we will identify G
with the group of all left translations of G.
We have already mentioned in Chapter 2 that if is a nonzero -nite left G-invariant (left
G-quasi-invariant) measure on an uncountable group (G, ), then the domain of differs
from the family of all subsets of G (see, for instance, [60], [104], and Exercise 11 of
Appendix 1). In this connection, the following natural problem arises.
Does there exist a left G-invariant (left G-quasi-invariant) measure
on G strictly extend-
ing ?
In the classical case, when G coincides with the n-dimensional Euclidean space R
n
(n 1)
and is a translation-invariant extension of the standard Lebesgue measure on R
n
, this
problem was originally posed by Sierpi nski (see [234] and [235]). Numerous works were
devoted to related problems concerning more general situations (see [41], [45], [85], [105],
[108], [115], [119], [135], [142], [195], [198], [242], [247], [249], and Exercise 20 from
Chapter 3).
There are various aspects of the above-mentioned Sierpi nski problemand of its generalized
version for an arbitrary uncountable group (G, ). In the present chapter, we are going to
discuss a certain aspect of this problem which is closely connected with the concept of
absolutely negligible sets. We will try to demonstrate that the notion of an absolutely
negligible set is of undoubted interest from the measure-theoretical and group-theoretical
points of view and plays a signicant role in questions concerning extensions of invariant
(quasi-invariant) measures (see [105], [108], [115], and [198]).
Let X be a subset of (G, ). We recall (cf. Chapter 2) that X is G-absolutely negligible
(in G) if for any -nite left G-invariant (respectively, left G-quasi-invariant) measure
on G, there exists a left G-invariant (respectively, left G-quasi-invariant) measure
on G
extending and satisfying the relation
(X) = 0.
This denition directly implies that:
Nonmeasurable subgroups of uncountable solvable groups 161
1) if X is a G-absolutely negligible set in G and X belongs to the domain of a given -nite
left G-invariant (left G-quasi-invariant) measure , then (X) = 0;
2) if X is a G-absolutely negligible set in G and X does not belong to the domain of a -
nite left G-invariant (left G-quasi-invariant) measure on G, then is strictly extendable,
by using this X, to some left G-invariant (left G-quasi-invariant) measure
such that
(X) = 0.
The properties 1) and 2) naturally lead to the following method of extending . Denote by
the rst innite cardinal and suppose that a given group G admits a countable covering
{X
n
: n <} such that all sets X
n
(n <) are G-absolutely negligible in G. If our measure
is not identically equal to zero, then there always exists at least one index n < for
which the set X
n
does not belong to dom(). Consequently, our can be strictly extended
with the aid of this X
n
(see Chapter 2).
So we come to the following problem essentially motivated by the generalized version of
Sierpi nskis problem mentioned above.
Problem. Give a characterization of all those uncountable groups (G, ) which admit a
countable covering by G-absolutely negligible sets.
This problemstill remains open. The main goal of the present chapter is to demonstrate that
any uncountable solvable group (G, ) can be covered by a countable family of G-absolutely
negligible sets. Moreover, a much stronger result will be obtained below in terms of certain
subgroups of (G, ).
For this purpose, we need several auxiliary propositions the rst of which was already
stated earlier in Chapter 2.
Lemma 1. Let (G, ) be an arbitrary group and let X be a subset of G. The following two
1) X is G-absolutely negligible;
2) for any countable family {g
i
: i I} G, there exists a countable family { f
j
: j J} G
such that
jJ
f
j
(
iI
g
i
X) = / 0.
A detailed proof of Lemma 1 was given in Chapter 2 where even a more general result
was presented concerning an analogous characterization of absolutely negligible sets in an
abstract space E equipped with some group of its transformations.
Obviously, Lemma 1 yields a purely algebraic characterization of all G-absolutely negli-
gible sets in (G, ). In particular, it immediately follows from this lemma that if X is a
G-absolutely negligible subset of G and is an arbitrary -nite left G-quasi-invariant
measure on G, then
({g
i
X : i I}) = 0
for every countable family {g
i
: i I} of elements of G (cf. Exercise 13 from Chapter 2).
Here the symbol
as usual denotes the inner measure associated with .

Lemma 2. Let Gand H be two groups and let : GH be an epimorphism(i.e., surjective
homomorphism). Suppose that Y is an H-absolutely negligible subset of H. Then the set
X =
1
(Y) is G-absolutely negligible in G.
Proof. Take any countable family {g
i
: i I} of elements of G and denote h
i
= (g
i
) for
each i I. Consider the set
Y
={h
i
Y : i I}.
Since Y is H-absolutely negligible, there exists a countable family
{h
j
: j J} H
such that
{h
j
Y
: j J} = / 0.
Further, since is a surjection, we can nd a countable family {g
j
: j J} G such that
(g
j
) = h
j
for all j J. Now, a straightforward verication shows that the equality
jJ
g
j
(
iI
g
i
X) = / 0
holds true. We leave the corresponding details to the reader. Finally, in view of Lemma
1, we can conclude that the set X is G-absolutely negligible in G which ends the proof of
Lemma 2.
Remark 1. There are several notions in the theory of invariant and quasi-invariant mea-
sures, which are connected with the concept of a G-absolutely negligible set. Among them
let us especially mention the notion of a G-negligible set and the notion of a G-absolutely
nonmeasurable set (for precise denitions and more details, see [108], [115], [119], and
Chapters 2 and 5). We would like to stress at this point that direct analogues of Lemma 2
are valid for those sets, too.
Namely, let G and H be again two groups, : G H be an epimorphism and let Y be a
subset of H. Then the following two assertions are valid:
a) if Y is H-negligible, then
1
(Y) is G-negligible;
b) if Y is H-absolutely nonmeasurable, then
1
(Y) is G-absolutely nonmeasurable.
These facts are useful in some constructions of G-negligible and G-absolutely nonmeasur-
able sets (cf. [108], [115], and [119]).
Lemma 3. Let (G, ) be a group and let H be a normal subgroup of G such that
card(G/H) . Fix a selector {g
k
: k K} of the countable disjoint family G/H. Suppose
that a set Y H is H-absolutely negligible in H. Then the set
X ={g
k
Y : k K}
turns out to be G-absolutely negligible in G.
Proof. We may assume without loss of generality that if g
k
H, then g
k
coincides with the
neutral element of G. Consider an arbitrary countable family {g
i
: i I} of elements from
G. Obviously, for each index i I, we can write
g
i
= g
k(i)
h
i
,
where h
i
H. Further, since H is a normal subgroup of G, we have the inclusion
{g
i
X : i I} {g
k
Z : k K},
where
Z ={h
i
Y : i I}
for some countable family {h
i
: i I} of elements of H. Keeping in mind that Y is H-
absolutely negligible, we derive that there exists a countable subgroup F of H for which
{f Z : f F} = / 0.
Now, consider the family
U ={g
k
f : k K, f F}
of elements of G. This family is countable and we assert that the equality
uU
u(
kK
g
k
Z) = / 0
is fullled. Indeed, suppose otherwise, i.e., suppose that there exists an element
t
uU
u(
kK
g
k
Z).
Clearly, we can represent t in the form t = g
k
0
h
0
for some elements k
0
K and h
0
H.
Consequently, we have the relation
t = g
k
0
h
0

f F
g
k
0
f (
kK
g
k
Z)
from which it readily follows that
h
0

f F
f (
kK
g
k
Z)
or, equivalently,
Fh
0
{g
k
Z : k K}.
Notice now that if g
k
does not belong to H, then
g
k
Z H = / 0
and, therefore,
Fh
0
g
k
Z = / 0.
This circumstance implies at once the inclusion Fh
0
Z and hence the relation
h
0
{f Z : f F},
which contradicts the above-mentioned equality {f Z : f F} = / 0. The obtained contra-
diction shows that
uU
u(
kK
g
k
Z) = / 0
and, consequently,
uU
u(
iI
g
i
X) = / 0.
The last equality establishes the G-absolute negligibility of X. Lemma 3 has thus been
proved.
Remark 2. All G-absolutely negligible sets may be regarded as small subsets of a given
uncountable group (G, ) because they form an ideal in the power set P(G) (see Exercise
14 fromChapter 2). However, the group operation applied to such subsets sometimes yields
a non-small set. More precisely, it can be demonstrated that for every uncountable solvable
group (G, ), there exist two G-absolutely negligible sets A and B satisfying the equality
A B = G (in this connection, see Exercise 20 for Chapter 11).
In what follows the symbol (G, ) will denote an arbitrary group and (G, +) will denote
an arbitrary commutative group. We do not suppose that G is equipped with a topology
compatible with its algebraic structure or, equivalently, we may assume that G is endowed
with its discrete topology.
Let (G, ) be an uncountable group. As previously shown, if is a nonzero -nite left
G-quasi-invariant measure on G, then there exists a subset of G nonmeasurable with respect
to (see [60], [104], and Exercise 11 for Appendix 1). This result is based on the classical
theorem of Ulam stating that the rst uncountable cardinal number
1
is not real-valued
measurable (see [238]).
For an uncountable commutative group (G, +) equipped with a nonzero -nite G-quasi-
invariant measure , a much stronger result can be established. Namely, it can be demon-
strated the existence of a subgroup of (G, +) which is nonmeasurable with respect to . In
general, the latter strong result does not hold for an uncountable non-commutative group
(H, ) endowed with a nonzero -nite left H-quasi-invariant measure. In this respect,
commutative groups have some advantage because for any uncountable commutative group
(G, +) equipped with a nonzero -nite G-quasi-invariant measure , there are uncount-
ably many subgroups of (G, +) nonmeasurable with respect to (see [116] and [119]).
Here we are interested in analogous statements closely connected with the existence of
those nonmeasurable subgroups of (G, +) which can be used for obtaining G-invariant
(G-quasi-invariant) extensions of initial nonzero -nite G-invariant (G-quasi-invariant)
measures on G. In order to obtain such subgroups, we need some helpful facts from the
general theory of commutative groups.
The algebraic structure of an arbitrary commutative group is investigated more or less thor-
oughly. The following statement yields a description of this structure and will be crucial in
further constructions.
Kulikovs Theorem. Any commutative group (G, +) admits a representation in the form
G ={G
n
: n <},
where {G
n
: n < } is an increasing (by inclusion) sequence of subgroups of G and, for
each n <, the group G
n
is a direct sum of cyclic groups (nite or innite).
For the proof of this important statement, see [70] and [152] (cf. also Appendix 5).
Suppose now that our group (G, +) is uncountable. Then we may assume, without loss of
generality, that all G
n
in the above-mentioned representation are also uncountable. Let us
express G
n
in the form of a direct sum
G
n
=

iI
n
G
i
,
where all G
i
are cyclic groups (nite or innite) and
card(G
i
) 2 (i I).
According to our assumption, we have card(I
n
) >. Further, let us represent every set I
n
in the form
I
n
={I
n,k
: k <},
where {I
n,k
: k < } is a partition of I
n
and card(I
n,k
) > for each k < . Finally, let us
put
G
n,k
=

iI
n,0
I
n,1
...I
n,k
G
i
and consider the countable family
{G
n,k
: n <, k <}
of subgroups of G. It is not difcult to verify the validity of the following two relations:
(1) card(G
n
/G
n,k
) > for any n < and k <;
(2) {G
n,k
: k <} = G
n
for every n <.
Relation (1) immediately implies that
(n <)(k <)(card(G/G
n,k
) >).
Relation (2) implies at once that
G ={G
n,k
: n <, k <}.
Summarizing these two consequences, we obtain the following auxiliary assertion.
Lemma 4. Any uncountable commutative group (G, +) admits a representation
G ={G
j
: j J},
where J is a countable set, G
j
is a subgroup of G for each j J, and the inequality
card(G/G
j
) > holds true.
We have already mentioned Serpi nskis problem which deals with various translation-
invariant extensions of the Lebesgue measure
n
n
. Lemma 4
enables us to give a solution of generalized Sierpi nskis problem concerning extensions of
nonzero -nite translation-invariant (respectively, translation-quasi-invariant) measures
on uncountable commutative groups by using the existence of certain nonmeasurable sub-
groups.
For this purpose, we need the next simple auxiliary proposition.
Lemma 5. Let (G, +) be an uncountable commutative group, be a -nite G-invariant
(G-quasi-invariant) measure on G and let G
0
be a subgroup of G such that card(G/G
0
) >
. Then there exists a G-invariant (G-quasi-invariant) measure
on G which extends
and satises the equality
(G
0
) = 0.
Proof. We apply a fairly standard argument based on Marczewskis method (see Chapter
2). Namely, let us introduce the following family of sets in G:
I ={X : X can be covered by countably many translates o f G
0
}.
Observe that I is a G-invariant -ideal of subsets of G. Moreover, if X I, then for
some countable family {g
i
: i I} G, we have
X {g
i
+G
0
: i I}.
Therefore, X (h +X) = / 0 for each element h G not belonging to the set
{g
i
g
j
+G
0
: i I, j I}.
This fact implies by easy transnite induction that there exists an uncountable family {h
:
<
1
} of elements of G such that the family
{h
+X : <
1
}
consists of pairwise disjoint sets. We thus conclude that
(X) = 0 in view of the -

niteness and G-quasi-invariance of .
Now, we can utilize the method of extending described in Chapter 2. Namely, we suppose
without loss of generality that is complete and introduce the -algebra
S
={YX : Y dom(), X I},

where denotes the symmetric difference of sets. Further, we put
(YX) = (Y)
for any set YX S
(here Y dom() and X I). A simple verication shows that
is well dened and is a complete G-invariant (G-quasi-invariant) measure extending .

According to the denition,
(X) = 0 for all members X of the -ideal I. In particular,

we have
(G
0
) = 0 which ends the proof of the lemma.
Remark 3. In fact, the preceding argument establishes that G
0
is a G-absolutely negligible
subset of G. Observe that this argument essentially exploits the commutativity of a given
group G. The reader can check it himself (herself).
From the above lemmas we readily obtain the following result.
Theorem 1. Let (G, +) be an uncountable commutative group. There exists a countable
family {G
j
: j J} of subgroups of G such that:
1) G ={G
j
: j J};
2) for any nonzero -nite G-invariant (G-quasi-invariant) measure on G, at least one
subgroup G
j
is nonmeasurable with respect to ;
3) if G
j
is nonmeasurable with respect to , then there exists a G-invariant (G-quasi-
invariant) measure
on G extending and satisfying the relation
(G
j
) = 0.
Proof. Let {G
j
: j J} be as in Lemma 4 and let be an arbitrary nonzero -nite
G-invariant (G-quasi-invariant) measure on G. By virtue of the equality
G ={G
j
: j J},
we may write
0 < (G)
jJ
(G
j
)
where
denotes the outer measure associated with . Consequently, there exists an index
j J such that
(G
j
) > 0. In view of the relation
card(G/G
j
) >
and of the G-quasi-invariance of , we easily infer that G
j
must be nonmeasurable with
respect to . Finally, applying Lemma 5, we conclude that for the same G
j
, there exists
a G-invariant (G-quasi-invariant) measure
on G extending and satisfying the equality
(G
j
) = 0 which completes the proof.
Theorem 1 may be regarded as the rst step in solving (within ZFC theory) the measure
extension problem for nonzero -nite invariant (quasi-invariant) measures. By using a
more complicated argument, this theoremcan be generalized to the class of all uncountable
solvable groups. For obtaining the desired result, we need one more notion.
Let (G, ) be an arbitrary group. We shall say that a family {G
j
: j J} of subgroups of G
is admissible if:
(i) card(J) ;
(ii) there exists a countable family {g
k
: k K} of elements of G such that
G ={g
k
({G
j
: j J}) : k K};
(iii) every group G
j
( j J) is G-absolutely negligible in G.
It can easily be seen that if {G
j
: j J} is an admissible family for G and is an arbitrary
nonzero -nite left G-invariant (left G-quasi-invariant) measure on G, then at least one
group G
j
is nonmeasurable with respect to . Therefore, can be extended to a left G-
invariant (left G-quasi-invariant) measure
on G such that
(G
j
) = 0 (cf. Lemma 5).
Recall that a group (G, ) is solvable if there exists a nite sequence
(
0
,
1
, ...,
n
)
of subgroups of G such that:
(a)
0
= G and card(
n
) = 1;
(b) for each integer k [0, n 1], we have
k+1

k
, the group
k+1
is normal in
k
and
the quotient group
k
/
k+1
is commutative.
Lemma 6. For every uncountable solvable group (G, ), there exists an admissible family
of its subgroups.
Proof. Let (
0
,
1
, ...,
n
) be a nite family of subgroups of G satisfying the above-
mentioned conditions (a) and (b). We use induction on n. Let
:
0

0
/
1
denote the canonical epimorphism. Only two cases are possible.
1. card(
0
/
1
) . In this case, we obviously have card(
1
) >. By inductive assump-
tion, there exists an admissible family {G
j
: j J} of subgroups of
1
. Taking into account
Lemma 3 and the inequality card(
0
/
1
) , we derive that the same family {G
j
: j J}
is admissible for
0
= G.
2. card(
0
/
1
) >. In view of Theorem 1, there exists an admissible family {G
j
: j J}
of subgroups of the uncountable commutative group
0
/
1
. Now, it is not difcult to verify
that the family of groups {
1
(G
j
) : j J} is admissible for the original group
0
= G.
The details of this verication are left to the reader.
The two preceding lemmas immediately imply an analogue of Theorem 1 for an arbitrary
uncountable solvable group (G, ), in terms of nonzero -nite left G-invariant (left G-
quasi-invariant) measures on G. Namely, we have the next statement.
Theorem 2. Let (G, ) be an uncountable solvable group. There exists a family {G
j
: j J}
of subgroups of G satisfying the following relations:
1) {G
j
: j J} is admissible;
2) for any nonzero -nite left G-invariant (left G-quasi-invariant) measure on G, at
least one subgroup G
j
is nonmeasurable with respect to ;
3) if G
j
is nonmeasurable with respect to , then there exists a left G-invariant (left G-
quasi-invariant) measure
on G extending and such that
(G
j
) = 0.
In connection with Theorem 2, we would like to formulate the following open problem.
Problem. Give a characterization of all those uncountable groups (G, ) for which there
exists at least one admissible family of subgroups of G.
Remark 4. Let (G, ) be a connected -compact locally compact topological group and let
D be a subgroup of G. Denote by the completion of the left Haar measure on G. Then,
by using the Steinhaus property (see Chapters 3 and 4), it can be shown that
(D) = 0
(G\ D) = 0.
In the case of the equality
(G\D) =0, the group D is usually called thick with respect to

(or, briey, -thick). Evidently, a -thick subgroup is everywhere dense in the original
group G. Moreover, if a -thick subgroup D is proper (i.e., G = D), then D is necessarily
nonmeasurable with respect to . Indeed, supposing for a while that a proper -thick
subgroup D is -measurable (hence, is of full -measure) and taking any element g G\D,
we obviously get gDD = / 0, thus it follows that
0 =(G\ D) (gD), 0 =(gD) =(D) > 0,
and we come to a contradiction.
Analogously, if D is an everywhere dense subgroup of an arbitrary -compact locally com-
pact topological group (G, ) equipped with the completion of the left Haar measure, then
either (D) = 0 or D is -thick in G.
Notice that there exists an uncountable commutative non-discrete locally compact topo-
logical group (, +) which does not admit proper everywhere dense subgroups (see [46],
[92], [103], [203], and Exercise 6 for Appendix 5). This fact also implies that for the same
(, +), there are no proper subgroups thick with respect to the completion of the Haar
measure on , i.e., there are no thick subgroups nonmeasurable with respect to .
On the other hand, the existence of a -nonmeasurable subgroup of follows directly
from Theorem 2 and it must be underlined that no topological technique is needed for
establishing this result.
It should also be mentioned that every uncountable non-discrete locally compact group G
is a resolvable topological space, i.e., G can be represented in the form
G = AB,
where A and B are some two disjoint everywhere dense subsets of G. More generally, the
well-known theorem due to Hewitt states that every locally compact topological space E
without isolated points is resolvable (see Exercise 7 for this chapter).
In various natural situations the topological structure of a given topological group does not
admit a nonzero -nite quasi-invariant measure.
For instance, it can be proved that if G is a proper uncountable analytic (or co-analytic)
subgroup of R, then there exists no nonzero -nite G-quasi-invariant Borel measure on G
(see [115] and [117]).
Similarly, if E is an arbitrary innite-dimensional Banach space, then there does not exist a
nonzero -nite E-quasi-invariant Borel measure on E (see, for instance, Chapter 4 of this
book; cf. also [226] where the case of an innite-dimensional Hilbert space E is considered
in detail).
Taking into account examples of this kind, it makes sense sometimes to ignore the topolog-
ical concepts and pose the following general question.
Let (G, ) be an uncountable solvable group. Does there exist a nonzero -nite measure
dened on a sufciently large -algebra of subsets of G and invariant under the group of
all left translations of G?
Theorem 2 shows that any such measure is dened on a proper subclass of the family
of all subsets of G and that can be strictly extended by using some -nonmeasurable
subgroup of G. However, the method of extending measures described in Theorems 1 and
2 does not essentially change the class of measurable sets. For instance, if the original
measure on G is separable (see Exercise 11 for Chapter 1), then the extended measure
produced by this method remains separable because all
-measurable sets are obtained

with the aid of members of a certain G-invariant -ideal of subsets of G which does not
change the metrical structure of . Therefore, it is reasonable to ask whether G admits
a nonseparable -nite left G-invariant measure. It turns out that the answer to the latter
question is positive under an additional set-theoretic assumption. In fact, the following
statement can be proved within ZFC theory.
(*) Let (G, ) be a solvable group whose cardinality is greater than or equal to the cardinality
of the continuum c. Then there exists a non-atomic nonseparable -nite left G-invariant
measure on G.
As an immediate consequence of (*), we have the next statement.
(**) Let (G, ) be an uncountable solvable group. Assuming the Continuum Hypothesis,
there always exists a non-atomic nonseparable -nite left G-invariant measure on G.
We do not know whether CH is necessary for the validity of (**).
A detailed proof of the statements (*) and (**) is given in Chapter 17. Notice that an
essential role in the proof is played by the following purely algebraic fact.
If (H, +) is an uncountable commutative group, then there exists a homomorphic image
(H
1
, +) of (H, +) such that card(H
1
) =
1
.
Simple examples show that this fact fails to be true for uncountable non-commutative
groups.
Also, it is unknown whether any uncountable non-commutative group (G, ) admits a
non-atomic nonseparable -nite left G-invariant measure. This property is fullled if
card(G) 2
and, simultaneously, card(G) is not conal with . However, if these condi-

tions do not hold for G, the question remains open.
In this context, it should be pointed out that if (G, ) is an uncountable -compact locally
compact topological group, then the equality
card(G) = (card(G))
is always valid. Therefore, in this case the above-mentioned conditions are satised auto-
matically.
Finishing this chapter, let us recall that the rst two (absolutely different) constructions
of nonseparable translation-invariant extensions of the Lebesgue measure on R were
done by Kakutani and Oxtoby [95], and Kodaira and Kakutani [141], respectively. Their
constructions admit a straightforward generalization to a certain class of -compact locally
compact groups (cf. [83]). In Chapter 16, we will consider one more method of obtaining
nonseparable invariant extensions of nonzero -nite invariant measures, which essentially
utilizes their nontrivial ergodic components.
More information concerning extensions of measures invariant under various transforma-
tion groups can be found in [39], [41], [42], [45], [46], [69], [83], [85], [86], [95], [108],
[115], [119], [141], [142], [195], [198], [234], [235], [242], [247], and [249].
EXERCISES
1. Prove the assertions a) and b) of Remark 1.
2. Verify that the group G of all isometric transformations of R
n
, where n 2, is solvable.
This algebraic fact implies that for n 2, there exists a functional
: P(R
n
) [0, +[
which is nitely additive, G-invariant and extends the Lebesgue measure
n
(see [22], [57],
[83], [108], and [240]). On the other hand, such a does not exist for n 3 (see [155] and
[240]).
3
. Recall that a topological space E is isodyne if for every nonempty open set U E, the
relation card(U) = card(E) is valid (equivalently, all nonempty open sets in E have one
and the same cardinality).
Suppose that the Generalized Continuum Hypothesis holds and let E be an isodyne Haus-
dorff topological space of second category. Prove that the relation (card(E))
= card(E)
is satised.
For this purpose, use the fact that if X is an innite subspace of E, then
card(X) 2
2
card(Y)
,
where Y is any everywhere dense subset of X. Check also that this inequality is precise,
i.e., for any innite cardinal number a, there exist a compact Hausdorff space X and its
everywhere dense subset Y such that
card(Y) = a, card(X) = 2
2
a
.
On the other hand, take an arbitrary uncountable set F and equip F with the topology
T ={X F : card(F \ X) }.
Verify that (F, T ) is an isodyne T
1
-space of second category. This circumstance shows that
the assumption on E that it should be Hausdorff is essential for the validity of the equality
(card(E))
= card(E).
4. Let (G, ) be a non-discrete -compact locally compact topological group. Infer from
the result of the previous exercise that under the Generalized Continuum Hypothesis, the
relation (card(G))
= card(G) holds true.

In connection with the last fact, it must be mentioned that for the same G, the equality
card(G) = 2
w(G)
is satised, where the symbol w(G) denotes the topological weight of G
(see, e.g., [46]). Actually, this result does not need any additional set-theoretical assump-
tions and trivially implies the equality (card(G))
= card(G).
5
. Let E be a topological space satisfying the following conditions:

(a) the smallest cardinality of nonempty open subsets of E is innite;
(b) there exists a base B of open sets in E such that card(B) .
Prove that E is a resolvable space.
For this purpose, use the method of transnite recursion and a Bernstein type construction
(cf. [148] and [192]).
6. Let E be a topological space. Check that for every nonempty open set U E, there
exists a nonempty isodyne open set V U. In other words, the family of all isodyne open
subspaces of E forms a pseudo-base of open sets in E.
Verify that if all members of some pseudo-base of open sets in E are resolvable, then E is
also resolvable.
For this purpose, consider a maximal (with respect to the inclusion relation) disjoint family
of elements of this pseudo-base and show that the union of such a family is everywhere
dense in E (cf. Exercise 12 for Appendix 3).
7
. Let E be an arbitrary locally compact topological space. Show that there exists a base
of E whose cardinality does not exceed card(E), i.e., the inequality w(E) card(E) holds
true.
By using the result of Exercise 5, infer from this fact that if a topological space Y is innite,
locally compact, and isodyne, then Y is resolvable.
Finally, taking into account Exercise 6, conclude that if X is a locally compact topological
space without isolated points, then X is resolvable (Hewitts theorem).
In particular, any non-discrete locally compact topological group is resolvable.
8. Give an example of an uncountable group (G, ) such that no homomorphic image of
(G, ) has cardinality
1
.
9. Let (G, +) be an arbitrary uncountable commutative group.
Show that there exists a countable family {G
j
: j J} of subgroups of G satisfying the
(a) G ={G
j
: j J};
(b) card(G/G
j
) = card(G) for any j J.
For this purpose, utilize an argument similar to that presented after the formulation of
Kulikovs theorem.
10
. Let (G, +) be an uncountable commutative -compact locally compact group and let
denote the completion of the Haar measure on G.
Prove that there exists a countable family {H
j
: j J} of subgroups of G satisfying the
(a) (G\ {H
j
: j J}) = 0;
(b) card(G/H
j
) = card(G) for any j J;
(c) each group H
j
( j J) is -thick in G;
(d) for every G-invariant extension
of , there exists an index j J such that H

j
is
nonmeasurable with respect to
;
(e) if some H
j
is nonmeasurable with respect to
, then
can be extended to a G-invariant

measure
such that
(H
j
) = 0.
In order to demonstrate the existence of {H
j
: j J} with properties (a)-(e), use the previous
exercise and the remark made after Exercise 4.
The result just formulated shows that in the case of an uncountable commutative -compact
locally compact group (G, +) equipped with the completion of its Haar measure, the
corresponding measure extension problemcan be solved with the aid of a certain countable
family of -thick subgroups of G.
Chapter 11
Algebraic sums of measure zero sets
This chapter and the next one are devoted to the following general problem concerning the
behavior of small subsets of uncountable groups under the group-theoretical operation.
Suppose that a group (G, ) is given with an ideal I of subsets of G and, as a rule, this I
is assumed to be left G-invariant. We will be interested in those situations, where there are
two sets A I and B I such that A B , I. In other words, we will be dealing with
those ideals of subsets of G, which are not preserved under the group-theoretical operation.
Naturally, we will especially pay our attention to the case of a group (G, ) equipped with
a nonzero -nite complete G-invariant (G-quasi-invariant) measure. In this case, the role
of I is played by the -ideal of all measure zero subsets of G.
Analogously, if (G, ) is a second category topological group, then the role of I is usually
played by the -ideal of all rst category subsets of G. This case is also of certain interest
from a topological view-point.
In addition, we especially want to examine the situations where the group-theoretical oper-
ation is commutative. In such a case, it is more convenient to denote it by +.
Let (G, +) be an arbitrary uncountable commutative group and let be a nonzero -nite
complete G-invariant or, more generally, G-quasi-invariant measure on G. As usual, we
denote by dom() the -algebra of all -measurable sets and by I() the -ideal of all
-measure zero sets in G.
In particular, let G coincide with the real line R and let coincide with the standard
Lebesgue measure on R, that is =. In connection with the said above, the well-known
elementary fact should be mentioned, stating that there are two subsets A and B of R for
which the relations
(A) =(B) = 0, A+B = R
hold true. For more details, see Exercises 2, 3, and 4 of this chapter. Also, it must be
noticed that by starting with the above-mentioned fact and applying the techique of Hamel
177
11
bases, Sierpi nski [219] was able to establish the existence of two subsets A
/
and B
/
of R
such that
(A
/
) =(B
/
) = 0, A
/
+B
/
, dom().
Analogous results are valid for the n-dimensional Euclidean space R
n
(n 2) and for the
Lebesgue measure
n
on this space. Moreover, by using deep theorems on the structure
of locally compact topological groups (see, for instance, [83], [177], [201], [202]) and the
fundamental Mackey theorem [165], Sierpi nskis result can be extended to a wide class of
uncountable topological groups equipped with -nite left invariant (left quasi-invariant)
Borel measures (in this connection, cf. also [37], [44], [116], [118], [121], [131], [132],
[137], and [236]).
It is reasonable to ask whether similar statements hold true for more general situations when
no topology is considered on a given group. Namely, it is natural to pose the following
question.
Let (G, +) be an uncountable commutative group equipped with a nonzero -nite com-
plete G-invariant (G-quasi-invariant) measure . Do there exist two sets A I() and
B I() whose algebraic sum A+B does not belong to dom()?
As already mentioned, the old theorem of Sierpi nski [219] states that if G coincides with
the real line R (or, more generally, with the Euclidean space R
n
where n 1) and if
coincides with the standard Lebesgue measure (
n
) on G, then the answer to this question
is positive.
Obviously, for an arbitrary uncountable commutative group (G, +) and for a nonzero -
nite complete G-invariant (G-quasi-invariant) measure on G, we do not have a direct
analogue of this theorem because the -ideal I() can be very poor. For example, I()
may coincide with the family of all countable subsets of G and, in this case, we readily get
I() +I() I() dom().
Various other and less trivial examples were constructed describing the situations, where
these inclusions are valid. Therefore, the formulation of the question posed above should
be replaced by another one. Here we suggest a slightly changed formulation in terms
of extensions of a given measure, which seems to be sufciently natural. Namely, the
following problem is of interest from the measure-theoretical point of view.
Problem. Let (G, ) be an uncountable group and let be a nonzero -nite left G-
invariant (respectively, left G-quasi-invariant) measure on G. Does there exist a left G-
invariant (respectively, left G-quasi-invariant) complete measure
/
on G extending and
Algebraic sums of measure zero sets 179
such that for some sets A I(
/
) and B I(
/
), the relation A B ,dom(
/
) is satised?
Our aim in the present chapter is to establish that the answer to this question is positive for
all uncountable commutative groups (G, +). Notice that for uncountable non-commutative
groups, the situation is still unclear. It should also be pointed out that various aspects con-
cerning the behavior of sets with a nice descriptive structure, under the operation of alge-
braic sum (i.e., Minkowskis sum) in a vector space, have been discussed and investigated
in many works (see, for instance, [62], [207], [227], [228], and Chapter 12).
We begin our considerations with the case of an uncountable commutative group (G, +)
and then we will briey touch upon the case of uncountable non-commutative groups.
Suppose rst that (G, +) is representable in the form of a direct sum of cyclic groups, i.e.,
G =
G
i
: i I,
where each subgroup G
i
(i I) is cyclic and is generated by a certain element g
i
G
i
0
i
(consequently, card(G
i
) 2). In this situation, we shall say that the set g
i
: i I is a basis
for G.
In some sense, g
i
: i I may be regarded as an analogue of a Hamel basis of any vector
space over the eld Q of all rational numbers. More precisely, every element g G admits
a unique representation
g = m
1
g
i
1
+m
2
g
i
2
+... +m
k
g
i
k
,
where i
1
, i
2
, ..., i
k
are pairwise distinct elements of I and all summands on the right-hand
side of this equality differ from zero (= the neutral element of G).
Let us denote:
X =g
i
: i I,
p(g) = k (g G),
X
1
= Z X =mx : m Z, x X,
X
n+1
= X
n
+X
1
(n = 1, 2, ...).
In the above notation, the symbol Z as usual stands for the ring of all integers.
Obviously, the following relations are valid:
(a) p(x) n for all elements x X
n
;
(b) X
1
X
2
... X
n
...;
(c) G =X
n
: 1 n <.
These simple relations will be used below.
Let now (G, ) be an arbitrary uncountable group. We remember from Chapter 2 that a
set Y G is G-absolutely negligible if for any -nite left G-invariant (respectively, left
G-quasi-invariant) measure on G, there exists a left G-invariant (respectively, left G-
quasi-invariant) measure
/
on G extending and satisfying the equality
/
(Y) = 0.
Various properties of absolutely negligible sets are considered in the monographs [108] and
[115] (see also Chapters 2 and 10). Among those properties, let us especially emphasize a
purely algebraic characterization of absolutely negligible sets which was mentioned several
times earlier and which will be very useful in our further constructions.
Lemma 1. Let (G, ) be an arbitrary uncountable group and let Y be a subset of G. The
following two assertions are equivalent:
1) Y is G-absolutely negligible in G;
2) for any countable family g
n
: n < of elements from G, there exists a countable family
f
m
: m < of elements from G such that
m<
f
m
(
n<
(g
n
Y)) = / 0.
Recall that a detailed proof of Lemma 1 was given in Chapter 2 (in fact, a more general
proposition was proved in the same chapter).
Lemma 2. Let (G, +) be a commutative group, H be a subgroup of G, and let Y be an
H-absolutely negligible subset of H. Then Y is a G-absolutely negligible subset of G.
Proof. Only two cases are possible.
1. card(G/H) > . In this case, by applying Lemma 1 it is not difcult to show that the
subgroup H is a G-absolutely negligible set in G (see, for instance, Lemma 5 and Remark
3 from Chapter 10). Since Y H, we easily infer that Y is G-absolutely negligible in G,
too. Actually, here we do not need the assumption that Y is H-absolutely negligible in H
because the inclusion Y H directly yields the required result.
2. card(G/H) . In this case, denote by g
j
: j J a selector of the countable family
G/H and dene
Y
/
=g
j
+Y : j J.
Taking into account the fact that Y is H-absolutely negligible in H, it can be demonstrated
that Y
/
is G-absolutely negligible in G (see Lemma 3 from Chapter 10). This circumstance
immediately implies that each set g
j
+Y ( j J) is also G-absolutely negligible in G. Con-
sequently, the set
Y = (g
j
+Y) g
j
turns out to be G-absolutely negligible, too, which ends the proof of Lemma 2.
We also need the following auxiliary proposition.
Lemma 3. Let (G, +) be an uncountable commutative group representable in the form of a
direct sum of cyclic groups and let X be a basis for G. Then X is a G-absolutely negligible
subset of G.
The proof of Lemma 3 can be carried out by the same method as in paper [106] where
vector spaces over Q, instead of commutative groups, are under consideration. Actually,
the argument presented in [106] works for uncountable commutative groups as well. Here
we establish a more general statement containing Lemma 3 as a very particular case.
Theorem 1. Let (G, +) be an uncountable commutative group representable in the form
of a direct sum of cyclic groups and let X be a basis for G. Then all sets X
n
(n = 1, 2, ...)
dened earlier are G-absolutely negligible in G.
Proof. Fix a natural number n 1 and consider the corresponding set X
n
. Take any count-
able family f
j
: j J of elements from G. Since G is uncountable, its basis X is uncount-
able, too, so there exists an element f G such that
p( f
i
f
j
f ) 2n +1
for all indices i J and j J. This circumstance implies that
( f +f
j
+X
n
: j J) (f
j
+X
n
: j J) = / 0.
Indeed, suppose to the contrary that
( f +f
j
+X
n
: j J) (f
j
+X
n
: j J) ,= / 0.
Then for some elements x
/
X
n
and x
//
X
n
and for some indices j J and i J, we must
have
x
/
+ f
j
+ f = x
//
+ f
i
.
Consequently, we get
x
/
x
//
= f
i
f
j
f ,
2n p(x
/
x
//
) = p( f
i
f
j
f ) 2n +1,
which yields a contradiction. Now, in view of Lemma 1, we conclude that the set X
n
is
G-absolutely negligible in G.
Remark 1. At present, it is known that every uncountable commutative group (G, +) can
be covered by a countable family of G-absolutely negligible sets (see, e.g., Chapter 10).
By virtue of Theorem 1, we are able to give another proof of this fact. Indeed, according to
Kulikovs theorem, we have a representation
G =G
k
: k <,
where all subgroups G
k
(k < ) are direct sums of cyclic groups (see [70], [152], and
Appendix 5). For any natural number k, denote by the symbol X
k
a basis of the group G
k
and for any natural number n > 0, denote by the symbol X
k,n
the corresponding set in G
k
.
Then we may write the equality
G 0 =X
k,n
: k <, 0 < n <
from which, taking into account Lemma 2 and Theorem 1, the required result immediately
follows, i.e., we come to a covering of G with countably many G-absolutely negligible
subsets.
For all uncountable non-commutative groups an analogous result is still unknown. In other
words, we do not know whether any uncountable group (G, ) admits a countable covering
with G-absolutely negligible sets. This open question seems to be of interest from the
view-point of the algebraic aspect of the general measure extension problem (cf. Chapter
2).
The next auxiliary proposition plays the key role for obtaining the main result of this chap-
ter.
Lemma 4. Let (G, +) be a commutative group, be a -nite complete G-quasi-invariant
measure on G and let H be a subgroup of G such that
(H) > 0. Then H cannot be

represented in the form
H = AY,
where (A) = 0 and Y is an H-absolutely negligible subset of H.
Proof. Suppose to the contrary that H admits a representation in the above-mentioned
form. Clearly, we may assume without loss of generality that
AY = / 0.
Further, since
(H) > 0, we may consider the nonzero -nite complete measure on H

which is produced by the original measure , i.e., is dened on the -algebra of sets
S =Z H : Z dom()
and for any set Z H S, we have by denition
(Z H) =
(Z H) = (Z H
/
),
where H
/
denotes a -measurable hull of H. It can easily be veried that is an H-quasi-
invariant measure on H. In addition, we get
A dom(), (A) = 0.
Therefore, the set Y = H A is -measurable and the relation
(Y) =(H A) =(H) > 0
must be satised which contradicts the assumption that Y is an H-absolutely negligible
subset of H. The contradiction obtained nishes the proof.
Remark 2. It is essential in the formulation of the previous lemma that H is a subgroup of
G. In this context, let us point out that if (G, +) is an arbitrary uncountable commutative
group and is a nonzero -nite G-quasi-invariant measure on G, then there always exists
a G-absolutely negligible subset Y of G with
(Y) > 0 (cf. Remark 1 above) and, in

particular, Y turns out to be nonmeasurable with respect to . Moreover, if G coincides
with the real line R and coincides with the standard Lebesgue measure on R, then
there are even -thick G-absolutely negligible sets in G (see, for instance, Exercise 6). The
existence of sets of this type can be established by the method of transnite recursion using
Lemma 1 and an argument similar to the classical Bernstein construction (cf. [107]).
Now, we are ready to prove the following statement.
Theorem 2. Let (G, +) be an uncountable commutative group and let be a nonzero
-nite G-invariant (respectively, G-quasi-invariant) measure on G. There exists a G-
invariant (respectively, G-quasi-invariant) complete measure
/
on G such that:
1)
/
extends ;
2) for some two sets A I(
/
) and B I(
/
), the relation A+B , dom(
/
) is satised.
Proof. According to Kulikovs theorem (see [70], [152], or Appendix 5), the given group
G can be represented in the form
G =H
k
: k <,
where H
k
: k < is an increasing (by inclusion) sequence of subgroups of G and each
H
k
(k <) is a direct sum of cyclic groups. Since is not identically equal to zero, there
exists a natural number k such that
(H
k
) > 0. Obviously, we may assume without loss of
generality that H
k
is uncountable. Let us put H = H
k
and let X be a basis for H. By virtue
of Theorem 1, the set X
1
is H-absolutely negligible in H. In view of Lemma 2, the same X
1
is a G-absolutely negligible subset of the original group G. Denote by
/
the minimal with
respect to the inclusion relation, complete, G-invariant (G-quasi-invariant) extension of
such that
/
(X
1
) = 0. By applying Lemma 4, it is not difcult to show that H remains to
be a set of strictly positive outer measure with respect to
/
, i.e., H does not belong to the
-ideal I(
/
). Further, since the equality
H =X
n
: 0 < n <
holds, there exists a smallest natural number n for which the set X
n
is also of strictly positive
outer
/
-measure. Notice that n > 1 because of the relation
/
(X
1
) = 0. Consequently, we
can write
X
n
= X
n1
+X
1
,
/
(X
n1
) =
/
(X
1
) = 0.
Now, using Theorem 1 once more, we see that X
n
is H-absolutely negligible in H and,
simultaneously, is not of
/
-measure zero. This fact immediately implies that X
n
is not
measurable with respect to
/
. Therefore, putting
A = X
1
, B = X
n1
,
we nally obtain
A I(
/
), B I(
/
), A+B = X
n
, dom(
/
),
which yields the required result.
Remark 3. In paper [118] several related questions are considered for an arbitrary un-
countable vector space E over the eld Q and for an arbitrary nonzero -nite E-invariant
(E-quasi-invariant) measure given on E. In this case, some stronger results for are
obtained by starting with the well-known fact that any vector subspace of E is a direct sum-
mand in E. The same property is true for any divisible subgroup of a commutative group
(G, +).
Remark 4. It would be interesting to generalize the main statement, that is Theorem 2 of
the present chapter, to a maximally wide class of uncountable groups (G, ) (not necessarily
commutative). For some partial results in this direction, see [131], [132], and [139].
Let (G, +) be again an uncountable commutative group and let be a nonzero -nite
complete G-invariant or, more generally, G-quasi-invariant measure dened on some -
algebra of subsets of G. In general, we cannot assert that there are two subsets A and B of
G satisfying the relations
(A) = (B) = 0, A+B = G.
For instance, the -ideal I() can coincide with the family of all countable subsets of G
and, consequently, I() can be closed under the operation of algebraic sum. In such a
case, it follows that A+B ,= G for any sets A I() and B I().
However, the question naturally arises whether there exists a G-invariant (G-quasi-
invariant) measure
/
on G which extends and for which there are two sets A and B
in G such that
/
(A) =
/
(B) = 0, A+B = G.
We are going to show that the answer to this question is always positive and, in fact, the
required sets A and B can be chosen not depending on an initial measure . For this
purpose, we need several auxiliary propositions. Note that G-absolutely negligible subsets
of G play again a key role here.
Remark 5. Let (G, ) be an innite group and let Z be a subset of G. We recall that Z is
almost left G-invariant in G if for any g G, the relation
card((gZ)Z) < card(G)
is valid. Suppose now that the cardinality of G is equal to
1
, and let X be a subset of G.
By using Lemma 1, it can be demonstrated that the following two assertions are equivalent:
(a) X is not G-absolutely negligible in G;
(b) there exists a countable family g
i
: i I G such that the set
Y =g
i
X : i I
contains an uncountable almost left G-invariant subset.
Let (G, ) be an uncountable group and let H be a subgroup of G. As usual, we denote by
G/H the family of all left H-orbits in G.
By using Lemma 1, we can prove the next auxiliary statement (cf. [121]).
Lemma 5. Let (G, ) be an uncountable group and suppose that a set Z G has the
following property:
for any countable subgroup H of G, the relation
card(T G/H : card(T Z) 2) < card(G)
is satised. Then Z is a G-absolutely negligible subset of G.
Proof. Take any countable family f
i
: i I G and denote by F the subgroup of G
generated by this family. Since
card(F) , card(G) >,
we can choose an element h GF. Further, denote by H the subgroup of G generated by
h f
i
: i I. Obviously, card(H) . According to our assumption, we have
card(T G/H : card(T Z) > 1) < card(G).
Let us put
P =T G/H : card(T Z) 1,
Z
/
= Z P.
Then we can write
card(Z Z
/
) < card(G)
and in view of Lemma 1, it is sufcient to demonstrate that
gH
g(
f F
f Z
/
) = / 0.
Suppose to the contrary that there exists an element
z
gH
g(
f F
f Z
/
).
Taking into account the denition of Z
/
, we infer that there exists a unique element z
/
Z
/
for which the inclusion H z F z
/
is valid. Consequently, we have
z F z
/
, F z = F z
/
, H z F z.
The last inclusion implies at once that
h z = f z
for some element f F. Therefore, we get h = f and h F, which contradicts the choice
of h. The obtained contradiction ends the proof of Lemma 5.
Lemma 6. Let (H, ) be a group of cardinality
1
. There exist two H-absolutely negligible
sets A and B in H such that A B = H.
The proof of Lemma 3 is sketched in Exercise 9 for this chapter.
Notice that a more general result is also valid. Namely, if card(H) is an uncountable regular
cardinal, then there exist H-absolutely negligible sets A and B in H such that A B =H (see
Exercise 10).
If (H, +) is an uncountable commutative group, then the assertion of Lemma 6 holds true
without any assumption on card(H) (see Theorem 3 below).
Lemma 7. Let (G
1
, ) and (G
2
, ) be two groups, : G
1
G
2
be a surjective homomor-
phism and let Y be a G
2
-absolutely negligible subset of G
2
. Then the set X =
1
(Y) is
G
1
-absolutely negligible in G
1
.
This auxiliary statement was proved in Chapter 10 (see Lemma 2 therein).
Lemma 8. Let (G, +) be an arbitrary uncountable commutative group. There exists a
commutative group (H, +) such that:
1) card(H) =
1
;
2) H is a homomorphic image of G.
Proof. Since G is uncountable, it contains a subgroup G
1
of cardinality
1
. According
to a well-known theorem from the theory of groups (see, e.g., [70], [152], and Appendix
5), there exists a divisible commutative group G
2
containing G
1
as a subgroup. We may
assume, without loss of generality, that
card(G
2
) =
1
.
Denote by
0
the identity embedding of G
1
into G
2
. Since G
2
is divisible,
0
can be
extended to a homomorphism
: G G
2
.
Finally, putting H =(G), we see that the group H satises both relations 1) and 2) of the
lemma.
A natural generalization of Lemma 5 is given in Exercise 18 for this chapter.
Theorem 3. For any uncountable commutative group (G, +), there exist two G-absolutely
negligible sets A and B in G such that A+B = G.
Proof. Applying Lemma 8, we can nd a commutative group (H, +) of cardinality
1
which is a homomorphic image of G. Fix an epimorphism
: G H.
In view of Lemma 6, there are two H-absolutely negligible sets A
/
and B
/
in H for which
A
/
+B
/
= H. Let us put
A =
1
(A
/
), B =
1
(B
/
).
According to Lemma 7, the sets A and B are G-absolutely negligible in G. Moreover, we
can write
A+B =
1
(A
/
) +
1
(B
/
) =
1
(A
/
+B
/
) =
1
(H) = G,
which ends the proof of Theorem 3.
Remark 6. It immediately follows from Theorem 3 that if (G, +) is an arbitrary uncount-
able commutative group, then there exists a G-absolutely negligible subset C of G such
that
C+C = G.
Indeed, take A and B as in Theorem 3 and put C = AB. Then the set C is G-absolutely
negligible in G and we have
C+C = (AB) +(AB) A+B = G,
which implies the required equality C+C = G.
Let Z be a subset of a given uncountable group (G, ). We recall (see Chapter 5) that Z is G-
absolutely nonmeasurable in G if for any nonzero -nite left G-quasi-invariant measure
on G, we have Z , dom().
At present, it is known that every uncountable commutative and, more generally, every
uncountable solvable group (G, ) contains a G-absolutely nonmeasurable subset (see [107]
or [119]). In this connection, the following problem is of certain interest.
Problem. Let (G, +) be an uncountable commutative group. Do there exist two G-
absolutely negligible sets A and B in G such that their algebraic sum A+B is a G-absolutely
nonmeasurable subset of G?
The problemjust formulated remains open. The concrete particular case, where G=R, will
be discussed later in the present chapter. In this context, we would like to recall once more
a weaker result stating that for every uncountable commutative group (G, +) and for any
nonzero -nite G-invariant (respectively, G-quasi-invariant) measure on G, there exists
a G-invariant (respectively, G-quasi-invariant) measure
/
on G extending and such that
/
(X) =
/
(Y) = 0, X +Y , dom(
/
)
for some G-absolutely negligible subsets X and Y of G (see Theorem 2 of this chapter).
Remark 7. Suppose that for a given uncountable group (G, ), there exist two G-absolutely
negligible sets X G and Y G and a G-absolutely nonmeasurable set Z G satisfying
the relation
X Y = Z.
Then we can assert that there are also two G-absolutely negligible sets A and B in G for
which
A B = G.
Indeed, since Z is G-absolutely nonmeasurable, we have the equality
g
i
Z : i I = G
for some countable family g
i
: i I G (see Exercise 12). Putting
A =g
i
X : i I, B =Y,
we get the required G-absolutely negligible sets A and B in G.
Theorem 4. Let (G, +) be an uncountable commutative group and let be a -nite
G-invariant (respectively, G-quasi-invariant) measure on G. There exists a G-invariant
(respectively, G-quasi-invariant) extension
/
of such that
/
(A) =
/
(B) = 0, A+B = G
for some G-absolutely negligible subsets A and B of G which do not depend on an initial
measure .
Proof. The desired result easily follows from Theorem 3. Indeed, let A and B be as in The-
orem 3. First, we extend our measure to a G-invariant (respectively, G-quasi-invariant)
measure
1
on G for which
A dom(
1
),
1
(A) = 0.
Then we extend the measure
1
to a G-invariant (respectively, G-quasi-invariant) measure
2
on G for which
B dom(
2
),
2
(B) = 0.
Finally, we put
/
=
2
and conclude that
/
is the required measure on G.
It is natural to ask whether a result similar to Theorem 3 holds true for uncountable non-
commutative groups. The following statement shows that a direct analogue of Theorem
3 is valid for a wide class of such groups. However, to prove this analogue, we need an
essentially different argument.
Theorem 5. Let (G, ) be an uncountable group such that
(card(G))
= card(G).
Then there exist two G-absolutely negligible sets A and B in G for which the equality A B =
G holds true.
Proof. Denote by the least ordinal number whose cardinality is equal to card(G), and let
g
: < be an enumeration of all elements of G. Taking into account the equality

(card(G))
= card(G),
we may introduce an -sequence H
: < whose members are all countable subgroups

of G. Now, we dene recursively two -sequences
a
: <, b
: <
of elements from G in the following manner.
Suppose that for an ordinal <, the partial families a
: < and b
: < have
already been dened. Then we denote:
G
= the group generated by H
: <;
C
=a
: < b
: <.
Obviously, we have the inequalities
card(G
) card() + < card(G).

Consequently, there exist two elements a and b from G such that
(G
a) (G
) = / 0, (G
b) (G
) = / 0, ab = g
.
We put a
=a and b
=b. Proceeding in this way, we are able to construct the -sequences

a
: < and b
: <. It immediately follows from our construction that the sets

A =a
: <, B =b
: <
satisfy the relations
card(T G/H : card(AT) 2) < card(G),
card(T G/H : card(BT) 2) < card(G)
for any countable subgroup H of G. In view of Lemma 5, both sets A and B turn out to be
G-absolutely negligible in G. Finally, by virtue of the relation
g
= a
A B ( <),
we obtain that G = A B which completes the proof of Theorem 5.
As soon as Theorem 5 is established, we may formulate a direct analogue of Theorem4 for
those uncountable groups (G, ) which satisfy the equality
(card(G))
= card(G).
We leave the formulation and proof of this analogue to the reader.
Let us return to almost invariant subsets of groups (see Chapters 2, 3, and Remark 5). By
applying some algebraic properties of these subsets, the next statement can be deduced.
Theorem 6. Let (G, ) be an arbitrary uncountable group whose cardinality is not co-
nal with and let be an arbitrary -nite left G-invariant (respectively, left G-quasi-
invariant) measure on G. Then there exist an almost left G-invariant set A in G and a left
G-invariant (respectively, left G-quasi-invariant) measure
/
on G which extends and for
which we have
/
(A) =
/
(A
1
) = 0, A A = A A
1
= G.
Proof. By using an argument similar to that of the classical Sierpi nski construction (cf.
[176], [192], [222], or Exercise 19 for Chapter 3), it is not difcult to dene a family
A
i
: i I of subsets of G satisfying the following relations:
(a) card(I) = card(G) and card(A
i
) = card(G) for all i I;
(b) for each i I, the set A
i
is almost left G-invariant and is almost invariant with respect
to the symmetry of G;
(c) A
i
A
j
= / 0 for any two distinct indices i I and j I.
Since G is an uncountable group and is a -nite measure, we can nd an index i I
for which
(A
i
) = 0, where
as usual denotes the inner measure associated with . Let

us put A = A
i
. Applying the standard method of extending invariant measures (see [234],
[235], and Chapter 2), we can extend to a left G-invariant (respectively, left G-quasi-
invariant) measure
/
on G for which the equalities
/
(A) =
/
(A
1
) = 0
are valid. Now, take any g G. Since A is almost left G-invariant, we have
card((gA)A) < card(G).
Keeping in mind that card(A) = card(G), we easily infer that gAA ,= / 0, which yields the
relation g A A
1
and, therefore,
G = A A
1
= A A.
The last formula ends the proof of Theorem 6.
Now, we wish to turn our attention to those algebraic sums of two absolutely negligible
sets which are extremely bad from the measure-theoretical point of view. Namely, we are
interested in the situation where A and B are xed small subsets of (G, +) whose algebraic
sum A+B is a set with utterly bad properties from the view-point of the theory of invariant
(quasi-invariant) measures. The precise notion of smallness which plays a signicant role in
our further considerations is the same concept of absolute negligibility discussed in Chapter
2. The precise notion of ultimate badness is the concept of absolute nonmeasurability which
was introduced earlier in Chapter 5.
Let us mention that the structure of absolutely nonmeasurable sets can be rather simple in
some innite-dimensional vector spaces considered as commutative groups. Namely, the
following proposition is valid.
Lemma 9. Let E be an innite-dimensional separable Hilbert space (over R) and let K be
an arbitrary open ball in E. Then K is an E-absolutely nonmeasurable subset of E.
The proof of Lemma 9 can easily be deduced from the result of Exercise 10 of Chapter
9. The assertion of this lemma readily implies the well-known fact that E does not admit
a nonzero -nite Borel measure quasi-invariant under the group of all translations of E
(see, e.g., [226] and Chapter 4).
Lemma 10. Suppose that (G
1
, +) and (G
2
, +) are two isomorphic commutative groups.
Then the following relations are equivalent:
1) there exist G
1
-absolutely negligible subsets X and Y of G
1
whose algebraic sum X +Y
is G
1
-absolutely nonmeasurable in G
1
;
2) there exist G
2
-absolutely negligible subsets A and B of G
2
whose algebraic sum A+B is
G
2
-absolutely nonmeasurable in G
2
.
We omit the trivial proof of Lemma 10.
Now, we are able to establish the following statement.
Theorem 7. There exist two R-absolutely negligible subsets of R such that their algebraic
sum is an R-absolutely nonmeasurable set in R.
Proof. Fix an innite-dimensional separable Hilbert space (E, [[ [[) and denote
K =e E : [[e[[ < 2.
By virtue of Lemma 9, the open ball K is an E-absolutely nonmeasurable subset of E.
Taking into account Lemma 10 and the fact that E and R are isomorphic as commutative
groups, it is sufcient to show that there exist two E-absolutely negligible sets X and Y in
E for which the equality
X +Y = K
holds true. We are going to dene the required sets X and Y by using the method of
transnite induction.
Let be the least ordinal number of cardinality continuum, k
: < be an enumeration
of all elements from K and let H
: < be an enumeration of all countable subgroups

of the additive group E. For any < , denote by G
the subgroup of E generated by the

set H
: <. Now, construct by transnite recursion two -sequences

x
: <, y
: <
of elements from E satisfying the following conditions:
(1) [[x
[[ < 1 and [[y
[[ < 1 for each <;

(2) x
+y
= k
for each <;

(3) (G
+x
) (G
+x
: <) = / 0 for any <;

(4) (G
+y
) (G
+y
: <) = / 0 for any <.

Suppose that for an ordinal <, the partial -sequences x
: < and y
: <
have already been constructed. Let us put
Z
=x
: < y
: <,
K
=e E : [[e k
[[ < 1,
D =e E : [[e[[ < 1.
Notice that
K
= D+k
,
card(G
+Z
) card() + < card(E),

card(K
D) = card(E).
Consequently, there are two points x D and y D such that
(G
+x) (G
+Z
) = / 0,
(G
+y) (G
+Z
) = / 0,
x +y = k
.
Let us dene x
= x and y
= y.
Proceeding in this manner, we are able to construct the required -sequences x
: <
and y
: < with properties (1) - (4). Now, putting

X =x
: <,
Y =y
: <,
we easily deduce that X +Y = K in view of (1) and (2). We also conclude that both X and
Y are E-absolutely negligible subsets of E in view of (3), (4), and Lemma 5. Theorem 7
Actually, the preceding argument yields a much stronger result. Namely, we can assert that
there exists an E-absolutely negligible set C E such that C+C = K. Indeed, it sufces
to put
C = X Y,
where X and Y are the above-mentioned E-absolutely negligible subsets of E. Therefore,
taking into account an algebraic isomorphism between the additive groups E and R, we
claim that there exists an R-absolutely negligible set Z R such that the algebraic sum
Z +Z is R-absolutely nonmeasurable.
Theorem 8. There are two subsets A and B of R having the following property:
for every nonzero -nite R-invariant (R-quasi-invariant) measure on R, there exists an
R-invariant (R-quasi-invariant) measure
/
/
(A) =
/
(B) = 0,
A+B , dom(
/
).
Proof. Indeed, it sufces to take as A and B any two R-absolutely negligible subsets of
R whose algebraic sum A+B is R-absolutely nonmeasurable in R. The existence of such
subsets is stated by Theorem 7.
Let (G
1
, +) and (G
2
, +) be commutative groups and let : G
1
G
2
be a surjective homo-
morphism. We already know that:
(a) if a set Y G
2
is G
2
-absolutely negligible, then the set X =
1
(Y) is G
1
-absolutely
negligible;
(b) if a set Y G
2
is G
2
-absolutely nonmeasurable, then the set X =
1
(Y) is G
1
-
absolutely nonmeasurable.
From the relations (a), (b), and Theorem 7 we can easily derive (under CH) that in every
uncountable vector space E over the eld Q of all rational numbers there exist two E-
absolutely negligible sets whose algebraic sumis E-absolutely nonmeasurable (cf. Exercise
17 below). It would be interesting to extend this result to the case of all uncountable
commutative groups.
EXERCISES
1. Show that there exists a totally disconnected compact subset Z of the Euclidean plane
R
2
= RR such that pr
1
(Z) is a non-degenerate segment of the line R0.
Notice that the rst example of a subset of R
2
of this type was constructed by Baire.
2. Let C denote the Cantor set on the real line R. Show that the equalities
C+C = [0, 2],
CC =C+(C) = [1, 1]
hold true. Conclude from these facts that the ideal of all Lebesgue measure zero subsets
of R and the ideal of all nowhere dense subsets of R are not closed under the operation of
algebraic sum (vector sum).
3. Deduce from the previous exercise that there exists a set X R such that:
(a) X is of Lebesgue measure zero;
(b) X is of rst category;
(c) X +X = X X = R.
4
. By using Exercise 3, show that there exists a subset D of the Euclidean space R
n
, where
n 1, satisfying the following conditions:
(a) D is of
n
-measure zero;
(b) D is of rst category;
(c) D+D = DD = R
n
.
Starting with this result and applying the method of transnite recursion, construct a Hamel
basis in R
n
which is of rst category and of
n
-measure zero.
5. Let (G, +) be a commutative group and let H be an arbitrary subgroup of G. Verify that
there exists a nonzero -nite G-invariant measure on G such that H dom().
Conclude from this fact that H is not absolutely nonmeasurable with respect to the class of
all nonzero -nite G-invariant measures on G.
6. By using the method of transnite recursion, construct an R-absolutely negligible subset
X of R which is thick with respect to the Lebesgue measure on R, i.e.,
(R X) = 0.
7. Let (G, ) be a group and let (H, ) be an arbitrary homomorphic image of (G, ). Suppose
that H admits a representation H = A B where A and B are some H-absolutely negligible
subsets of H. Show that G admits an analogous representation G = A
/
B
/
where A
/
and B
/
are some G-absolutely negligible subsets of G.
8. Prove the equivalence of assertions (a) and (b) of Remark 5.
9
. Give a detailed proof of Lemma 6.

For this purpose, take any group (H, ) of cardinality
1
and consider an
1
-sequence
H
: <
1
of subgroups of H such that:
(a) card(H
) for any ordinal <

1
;
(b) <
1
H
;
(c) H
: < ,= / 0 for each ordinal <

1
;
(d) H =H
: <
1
.
Verify that any partial selector of the family
(H
: <) : <
1
is an H-absolutely negligible subset of H.

Let h
: <
1
be an enumeration of all elements of H. By applying the method of
transnite recursion, construct two families
a
: <
1
, b
: <
1
of elements of H satisfying the following conditions:

(e) a
: <
1
and b
: <
1
are partial selectors of the above-mentioned family
(H
: <) : <
1
;
(f) a
= h
for any ordinal <

1
.
Then, putting A =a
: <
1
and B =b
: <
1
, demonstrate that A and B are the
required H-absolutely negligible sets in H.
10. Let (H, ) be a group such that card(H) is an uncountable regular cardinal. Show that
there exist two H-absolutely negligible sets A and B in H such that A B = H.
For this purpose, use an argument similar to the argument of Exercise 9. Try to establish
an analogous result for all those uncountable groups (H, ) which satisfy the equality
(card(H))
= card(H).
11. Formulate and prove a direct analogue of Theorem 4 for all those uncountable groups
(G, ) which satisfy the equality
(card(G))
= card(G).
12. Let E be a set, G be a group of transformations of E, and let X be a G-absolutely
nonmeasurable subset of E. Show that there exists a countable family g
i
: i I G such
that
E =g
i
(X) : i I.
For this purpose, suppose otherwise and consider the G-invariant -ideal I of subsets of
E generated by X. Since this I is proper, there exists a probability G-invariant measure
on E such that
(Z I)((Z) = 0).
In particular, the relation (X) = 0 holds true which contradicts the G-absolute nonmea-
surability of X. The contradiction obtained yields the desired result.
13. Let us preserve the notation of the proof of Theorem 6. Give a detailed construction of
the family A
i
: i I of subsets of G satisfying the relations (a), (b), and (c).
14. Let (E, [[ [[) be a Banach space over R. Prove that:
(a) no closed half-space in E is E-absolutely nonmeasurable;
(b) if E is innite-dimensional and separable, then for every nonzero -nite E-quasi-
invariant measure on E, there exists a closed half-space in E which is nonmeasurable
with respect to .
15. By using the technique of Hamel bases, show that R and any vector space E over R
with card(E) = c are isomorphic as commutative groups.
In particular, an innite-dimensional separable Hilbert space H is isomorphic to R in the
group-theoretical sense. However, any isomorphism f between H and R is very bad from
the point of view of descriptive set theory and general topology. For instance, check that f
does not possess the Baire property.
16. Prove that there is an R-absolutely negligible set C having the following property:
for every -nite R-invariant (R-quasi-invariant) measure on R, there exists an R-
invariant (R-quasi-invariant) extension
/
of such that
/
(C) = 0, C+C , dom(
/
).
Moreover, generalize this result to the case of any vector space E over Q with card(E) c.
17. Assuming the Continuum Hypothesis prove that for every uncountable vector space
E over Q, there exists an E-absolutely negligible set Z such that Z +Z is E-absolutely
nonmeasurable.
18
. Let (G, +) be an arbitrary innite commutative group. Showthat for any innite cardi-
nal a card(G), there exists a homomorphic image (H, +) of (G, +) such that card(H) =a.
Prove this fact by applying two essentially different methods. Namely, the rst one is
similar to the argument used in the proof of Lemma 8 of this chapter and the second one
is based on the structural representation of any commutative group, i.e., exploits Kulikovs
theorem.
19. Suppose that every commutative group (H, +) with card(H) =
1
contains an H-
absolutely negligible set X such that X +X is H-absolutely nonmeasurable. In this case,
show that any uncountable commutative group (G, +) contains a G-absolutely negligible
set Y such that Y +Y is G-absolutely nonmeasurable.
20
. Let (G, ) be an uncountable solvable group. Show that there exist two G-absolutely
negligible sets A and B in G satisfying the equality A B = G.
For this purpose, consider a composition series
e = G
n
G
n1
... G
1
G
0
= G
for G and use induction on n. Observe that if n = 1, then G is commutative, so Theorem 3
of this chapter is applicable to G.
Chapter 12
The absolute nonmeasurability of Minkowskis
sum of certain universal measure zero sets
As mentioned in the previous chapter, there are examples which show that when the opera-
tion of algebraic sum is applied to subsets of the real line R or, more generally, to subsets of
a given topological vector space E, then the descriptive structure of the obtained sets some-
times essentially differs from the structure of summands. We shall see below further vivid
realizations of this phenomenon. Actually, the main goal of this chapter is to demonstrate
that the algebraic sum (frequently called Minkowskis sum) of two very small subsets of a
topological vector space E can be absolutely nonmeasurable with respect to the class of the
completions of all nonzero -nite continuous Borel measures on E.
In preceding considerations we have been concerned with some interesting statements on
certain algebraic and topological properties of the classical Cantor set on the real line R(see
Exercises 2 and 3 for Chapter 11). In this connection, we would like to recall once more a
simple but beautiful fact, which quite often is cited in various text-books of mathematical
analysis and elementary topology (see, e.g., [72] and [176]). Namely, if C denotes the
Cantor set on R, then the algebraic sum
C+C ={x +y : x C, y C}
coincides with the closed interval [0, 2] and the difference set
CC ={x y : x C, y C}
coincides with the closed interval [1, 1].
The proof of both these assertions is not difcult and may be found in many sources (see,
for instance, [72] or [176]). Moreover, in [72] and [176] a nice geometrical interpretation
of this fact is given (cf. Exercise 1 for Chapter 11). As is well known, C is nowhere dense
in R and its Lebesgue measure equals zero. Consequently, we infer that the algebraic sum
of two small subsets of R (in the sense of Lebesgue measure or Baire category) can have
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2, 00
DOI 10.1007/978-94-91216-36-7_1 2009 Atlantis Press/World Scientific 2,
2
nonempty interior, so can be a non-small subset of R. Furthermore, putting
D ={nC : n is integer},
we easily get
D+D = DD= R,
where the set D is of rst category and of Lebesgue measure zero.
We thus conclude that the algebraic sum of two small sets in R can coincide with R. It
should be emphasized that here the smallness of sets is meant both in the sense of Baire
category and Lebesgue measure.
These examples are the simplest and historically earliest which show that the operation of
an algebraic sum applied to small subsets of the real line leads to big sets. In particular, we
see that certain descriptive properties of summands are not preserved under this operation.
Among many other examples of this kind, let us mention the following two facts closely
connected with the assertions formulated above.
Example 1. There exist two Borel subsets X and Y of R such that the set X +Y is not Borel
in R.
As far as we know, this statement was rst established in [227] and [228]. Further results
in this direction were obtained by various authors (see, for instance, [37], [62], and [207]).
Notice that if A and B are any two Borel subsets of R, then the set A +B which is a
continuous image of the Borel subset A B of the plane R
2
is necessarily an analytic
(Suslin) set in R. Consequently, A +B has the Baire property and is measurable in the
Lebesgue sense. Moreover, A+B is universally measurable with respect to the class of the
completions of all -nite Borel measures on R (see Appendix 2).
Example 2. There exist two Lebesgue measure zero subsets X and Y of R such that the set
X +Y is not Lebesgue measurable.
We have already pointed out that this remarkable result is due to Sierpi nski [219]. An
analogous result is valid in terms of Baire category, and the arguments in both cases are
similar. More precisely, in both cases the proofs of these results use the technique of Hamel
bases and, of course, rely on an uncountable form of the Axiom of Choice.
The phenomenon described in Example 2 is not accidental. It remains true for a large class
of measures on the real line R. To give a formulation of the corresponding assertion, let us
rst recall that an afne transformation of R is any mapping
h : R R
The absolute nonmeasurability of Minkowskis sum of certain universal measure zero sets 201
representable in the form
h(x) = ax +b (x R),
where a and b are some xed real numbers and, in addition, a = 0. The family of all afne
transformations of R is a group with respect to the standard composition operation. Let H
denote this group. As usual, H is called the afne group of R and is signicantly larger
than the group of all isometric transformations of R.
We recall that a measure dened on some H-invariant -algebra of subsets of R is quasi-
invariant under H (or, in short, H-quasi-invariant) if for each h H and for every -measure
zero set X, we have (h(X)) = 0.
Obviously, the standard Lebesgue measure on R is H-quasi-invariant but is not H-
invariant. There are various extensions of which also are quasi-invariant under H. In
other words, the class of such measures is sufciently large. Notice that there exist even
nonseparable measures belonging to this class (cf. [115]).
The following statement may be regarded as a generalization of Example 2.
Theorem 1. Let be a nonzero -nite complete measure on R quasi-invariant under the
group H. Suppose that there are two -measure zero sets A and B such that
A+B = R.
Then there exists a -measure zero set X such that X +X is not measurable with respect to
.
For the proof, see [116]. The technique of Hamel bases is also useful for establishing
Theorem 1 but essentially new ideas are needed, too. Observe that Theorem 1 signicantly
generalizes Example 2.
The Cantor set C which is of Lebesgue measure zero is not an absolutely small subset of
R. Indeed, C carries a Borel probability measure whose completion is isomorphic with
the restriction of to the -algebra of all Lebesgue measurable subsets of the closed unit
interval [0, 1]. Actually, coincides with the product measure
nN
n
where for each
natural number n, the measure
n
is dened on the family of all subsets of {0, 1} and
n
({0}) =
n
({1}) = 1/2.
We thus see that from the stand-point of the denition of , the set C is big because
(C) = 1 = 0.
However, it was repeatedly underlined in preceding sections that there exist uncountable
subsets of R, which are extremely small from the measure-theoretical view-point.
Let us recall the precise denition of small subsets of R in the sense of topological measure
theory (see Chapters 2 and 5).
Let T be a topological space such that all singletons in T are Borel subsets of T. A set
X T is said to be universal measure zero in T if for every -nite diffused Borel measure
on T, we have
(X) = 0.
As mentioned earlier, several constructions of uncountable universal measure zero subsets
of uncountable Polish topological spaces are known in the literature (see, for instance, [69],
[79], [113], [148], [172], and Appendix 1). More delicate constructions of universally
small sets are presented in [197] and [250]. The existence of a universal measure zero set
of cardinality continuum cannot be established within ZFC theory (see [172]). However,
assuming the ContinuumHypothesis, we easily get the existence of such sets. For example,
as well known, all Luzin subsets of R are universal measure zero. By assuming Martins
Axiom, the same result can be obtained for all generalized Luzin sets in R. Notice that
Martins Axiom is much weaker than CH and does not bound from above the cardinality
of the continuum c.
Let us recall the denition of generalized Luzin subsets of Polish topological spaces.
Let T be an uncountable Polish topological space. A set X T is a generalized Luzin
subset of T if card(X) = c and for each rst category set Z T, we have card(X Z) < c.
Under Martins Axiom, there are generalized Luzin sets in T and any such set is universal
measure zero (see Exercise 14 for Appendix 1). In fact, the construction of a generalized
Luzin set imitates the classical construction of Luzin or the construction of Mahlo, where
the Continuum Hypothesis is assumed instead of Martins Axiom (see [143], [148], [159],
[176], [192], and [222]). By modifying these constructions, the existence of generalized
Luzin sets with additional algebraic properties can be shown. For example, several authors
proved that under Martins Axiom there exists a generalized Luzin set on the real line,
which is simultaneously a vector space over the eld Q of all rational numbers (cf. [40],
[59], [172], and Chapter 5).
In this connection, it should be pointed out that a much stronger statement was proved by
Erd os, Kunen and Mauldin in their joint paper [59]. Namely, they have established the
following statement.
Theorem 2. Suppose that Martins Axiom holds. Then there exist two sets X R and
Y R such that:
1) both X and Y are generalized Luzin sets in R;
2) both X and Y are vector spaces over the eld Q;
3) X +Y = R and X Y ={0}; in other words, R is a direct sum of its vector subspaces X
and Y.
Proof. Let us make several preliminary remarks. In fact, we are able to give a proof
of Theorem 2 which also works in a more general case, namely, for a Polish topological
vector space E = {0} instead of R (cf. Lemma 1 below). Notice that our argument differs
slightly from the one presented in [59] and, in some technical details, is easier.
First, let us recall that if E is a vector space over an arbitrary eld and U, V are vector
subspaces of E, then there exists a vector space V
V such that
U +V =U +V
,
U V
={0}.
Indeed, U V is a vector subspace of V, so we can take as V
a vector subspace of V
complemented to U V.
Consequently, it sufces to prove the existence of sets X and Y satisfying the relations 1),
2), and such that
X +Y = R.
Let denote the least ordinal number of cardinality continuum and let {z
: < } be
an enumeration of all points of R. Further, let {F
: < } be an enumeration of all

rst category F
-subsets of R. By using transnite recursion, we shall construct two -

sequences
{X
: <}, {Y
: <}
of subsets of R, fullling the following relations:
a) for each <, we have card(X
) = card(Y
) = card() +;
b) if < <, then X
and Y
;
c) if <, then X
and Y
are vector spaces over Q;

d) if <, then z
+Y
;
e) if < <, then F
= F
and F
= F
.
Suppose that for an ordinal <, the partial -sequences {X
: <} and {Y
: <}
have already been constructed. Let us denote:
X
={X
: <},
Y
={Y
: <}.
Let z be a point of R dened as follows:
z = z
if z
does not belong to X
+Y
, and z is some point of R\(X
+Y
) if z
belongs to
X
+Y
.
Keeping in mind the relations
card(X
) card() + < c,
card(Y
) card() + < c
and taking into account the fact that {F
: <} is a rst category set in R (see Remark

1 below), it is not difcult to show that there exist two points x R\X
and y R\Y
such
that:
x +y = z,
x X
+Q({F
: <}),
y Y
+Q({F
: <}).
We then put
X
= Qx +X
, Y
= Qy +Y
.
Proceeding in this way, we get the -sequences {X
: < } and {Y
: < }. Finally,
we dene
X ={X
: <},
Y ={Y
: <}.
It follows from the above construction that X and Y are vector spaces over Q, are also
generalized Luzin sets, and
X +Y = R
which ends the proof of the theorem.
We immediately obtain from Theorem 2 that under Martins Axiom, the algebraic sum of
two universal measure zero sets can coincide with R. Moreover, putting
A = X Y,
where X and Y are as in Theorem 2, we get a generalized Luzin set A (which is universal
measure zero as well) such that
A+A = R.
In particular, we infer that the algebraic sum A+A, where A is universal measure zero in
R, can coincide with the whole R in certain models of set theory.
In this connection, it makes sense to consider the question whether the algebraic sumof two
universal measure zero sets can be very bad from the measure-theoretical point of view.
First, let us remember the precise denition of extremely bad sets in the sense of topological
measure theory (see Chapter 5).
Let T be a topological space such that all singletons in T are Borel subsets of T. We say
that a set Y T is absolutely nonmeasurable in T if for every nonzero -nite continuous
Borel measure on T, we have Y dom(
), where
denotes the completion of .

It turns out that absolutely nonmeasurable sets in an uncountable Polish space T admit a
purely topological characterization. For this purpose, we need one notion from general
topology.
Let Z be a subset of a topological space T. Recall that Z is a Bernstein set if for each
nonempty perfect set P T, both intersections PZ and P(T \ Z) are nonempty.
This denition directly shows that Z T is a Bernstein set if and only if T \Z is a Bernstein
set.
The well-known argument from Bernstein which is based on the method of transnite in-
duction establishes the existence of a Bernstein set in an uncountable Polish topological
space T (cf. [12], [148], [176], [192], and Exercise 5 for Chapter 5). Moreover, any
construction of such a set needs an uncountable form of the Axiom of Choice. Various
properties of Bernstein sets are discussed in [12], [40], [148], [176], and [192] with their
applications in general topology, measure theory, and real analysis.
We would like to mention a simple fact concerning the cardinality of Bernstein sets.
Namely, as known, every Bernstein subset of R is necessarily of cardinality continuum.
This fact immediately follows from the existence of a disjoint family {P
j
: j J} of
nonempty perfect subsets of R, where card(J) = c. To get such a family {P
j
: j J}, it
sufces to take some Peano type mapping
f : [0, 1] [0, 1]
2
,
which is a continuous surjection and consider the family of all pre-images
f
1
({x} [0, 1]) (x [0, 1]).
These pre-images are pairwise disjoint and each of them is an uncountable closed set in R.
So each of them contains a nonempty perfect subset.
The next statement shows that in uncountable Polish topological spaces Bernstein sets are
identical with absolutely nonmeasurable sets (cf. Exercise 18 from Chapter 5).
Theorem3. Let E be an uncountable Polish space and let Z be a subset of E. The following
two assertions are equivalent:
1) Z is absolutely nonmeasurable in E;
2) Z is a Bernstein set in E.
Proof. 1) 2). Assume that for a set Z E, relation 1) holds and let us show that Z is a
Bernstein set. Supposing to the contrary that Z is not a Bernstein set, we deduce that for
some nonempty perfect set P E, we have either PZ = / 0 or P(E \ Z) = / 0. We may
assume, without loss of generality, that PZ = / 0. According to a well-known theorem of
descriptive set theory (see, e.g., [40], [99], [148], or Appendix 6), the topological spaces
P and [0, 1] are Borel isomorphic. Let
0
denote the restriction of the Lebesgue measure
to the Borel -algebra of [0, 1]. Using any Borel isomorphism between [0, 1] and P, we
can transfer
0
to P. In this manner, we obtain a Borel diffused probability measure on
E concentrated on P, i.e.,
(P) = 1, (E \ P) = 0
and hence
(Z) = 0. Consequently, Z turns out to be measurable with respect to the com-

pletion of . This circumstance contradicts 1) and, therefore, establishes the implication 1)
2).
2) 1). Let Z be a Bernstein set and let be an arbitrary nonzero -nite diffused
Borel measure on E. Denote by
the completion of . We have to verify that Z is

not measurable with respect to
. Suppose for a while that Z dom(
). Then either
(Z) > 0 or
(E \ Z) > 0. We may assume, without loss of generality, that
(Z) > 0.
Since
is a Radon measure, there exists a compact set K Z such that
(K) > 0. Since
is diffused, K must be uncountable. So K contains a nonempty perfect subset, which

contradicts the assumption that Z is a Bernstein set. Theorem 3 has thus been proved.
The following auxiliary statement is a direct analogue of Theorem 2.
Lemma 1. Suppose that Martins Axiom holds. Let E ={0} be a Polish topological vector
space. Then there exist two subsets L
1
and L
2
of E such that:
1) both L
1
and L
2
are generalized Luzin sets in E;
2) both L
1
and L
2
are vector spaces over Q;
3) L
1
+L
2
= E and L
1
L
2
={0}.
Consequently, under Martins Axiom, there exist two universal measure zero sets in E
whose algebraic sum coincides with E.
The proof of this lemma is completely analogous to the proof of Theorem 2. The corre-
sponding details are left to the reader.
Now, we are going to determine some connection of Bernstein sets with the fact described
in Example 2. The following result can be treated as a certain stronger version of Example
2. However, the proof of this version needs an additional set-theoretical assumption.
Theorem 4. Suppose again that Martins Axiom holds. Let E = {0} be a Polish topo-
logical vector space. Then there exist two generalized Luzin subsets of E such that their
Minkowskis sum is a Bernstein set in E.
Consequently, under Martins Axiom, there exist two universal measure zero sets in E
whose Minkowskis sum is absolutely nonmeasurable.
Proof. According to Lemma 1, we have a representation
E = X +Y (X Y ={0}),
where X and Y are generalized Luzin sets and both of them are vector spaces over Q. Let
denote again the least ordinal number of cardinality continuum and let {P
: < }
denote the family of all nonempty perfect subsets of E. We may assume, without loss of
generality, that each of the partial families
{P
: <, is an even ordinal}, {P
: <, is an odd ordinal}

also contains all nonempty perfect subsets of E. By using the method of transnite recur-
sion, we shall construct an injective -sequence {y
: <} of points of E such that:

a) {y
: <} Y;
b) for every ordinal <, we have (X +y
) P
= / 0.
Suppose that for an ordinal < , the partial -sequence {y
: < } has already been

dened. Consider the set P
and equip this set with a Borel probability diffused measure

(cf. the proof of Theorem 3). Since all sets X +y
( <) are universal measure zero, we

get
(P
(X +y
)) = 0 ( <).
In view of Martins Axiom (see Remark 1 below), we come to the relation
({P
(X +y
) : <}) = 0.
Therefore, we can write
P
\ {X +y
: <} = / 0.
Let z be a point of P
\ {X +y
: <}. By virtue of the equality

E = X +Y,
there exists a point y Y such that z X +y. We put y
= y.
Proceeding in this manner, we will be able to construct the desired -sequence of points
{y
: <}. Afterwards, we dene

L
1
= X,
L
2
={y
: <, is odd},
L
3
={y
: <, is even}
and we assert that L
1
and L
2
are the required sets. Indeed, according to the denition,
L
1
= X is a generalized Luzin set. Notice that L
2
which is a subset of Y with card(L
2
) = c
is a generalized Luzin set, too. Further, for any odd ordinal <, we have the relation
P
(X +L
2
) P
(X +y
) = / 0.
Analogously, for any even ordinal <, we may write
P
(X +L
3
) P
(X +y
) = / 0.
At the same time, taking into account the equality X Y ={0}, we easily obtain
(X +L
2
) (X +L
3
) = / 0.
Thus both X +L
2
and X +L
3
are Bernstein subsets of E which completes the proof of
Theorem 4.
By slightly changing the above argument, we can also prove that if Martins Axiomis valid,
then there exists a generalized Luzin set L such that L+L is a Bernstein subset of E (in this
context, cf. [44]).
By applying a similar technique, it can be established that under Martins Axiomthere exist
two generalized Luzin sets L
1
and L
2
which are vector spaces over Q and whose algebraic
sum L
1
+L
2
is a Bernstein set. In this case, we necessarily have L
1
= L
2
.
Remark 1. It is easy to see that the proof of Theorem 2 (as well as the proof of Theorem
4) does not need the full power of Martins Axiom. Indeed, in the proof of Theorem 2 we
only use the following consequence of this axiom:
if E is a second category topological space with a countable base, then the -ideal of all
rst category subsets of E is c-additive.
Analogously, in the proof of Theorem 4 we use the following consequence of Martins
Axiom:
if E is a topological space with a countable base, in which all closed subsets are of type
G
, and is a nonzero -nite Borel measure on E, then the -ideal generated by all
-measure zero subsets of E is c-additive
For more details, see e.g. [10], [40], [67], and [216].
Remark 2. It should be noticed that an additional set-theoretical assumption in the formu-
lation of Theorem4 cannot be omitted. Indeed, as mentioned before, there is a model of set
theory in which the Continuum Hypothesis fails to be true (i.e.,
1
< c) and the cardinality
of any universal measure zero subset of R does not exceed
1
(see [172]). If U
1
and U
2
are
any two universal measure zero subsets of R in such a model, then
card(U
1
+U
2
)
1

1
=
1
and, consequently, U
1
+U
2
cannot be a Bernstein set because the cardinality of every Bern-
stein set is equal to c. Recall that one purely algebraic version of Theorem 4 was discussed
in Chapter 11 (see Theorem 7 therein). Notice that Theorem 7 from Chapter 11 has some
advantage over Theorem 4 of this chapter because it does not need any additional set-
theoretical hypotheses.
Remark 3. It would be interesting to extend the preceding results to a more general case
where E is an uncountable Borel subgroup (nontrivial Borel vector subspace) of a Polish
topological commutative group (Polish topological vector space). Namely, in connection
with Lemma 1, the following two questions naturally arise for E assuming of course the
Continuum Hypothesis or Martins Axiom.
(a) Do there exist two universal measure zero subgroups X and Y of E such that X +Y =E?
(b) Do there exist two universal measure zero subsets X
and Y
of E such that X
+Y
is
absolutely nonmeasurable?
As far as we know, these questions remain open.
Also, it is natural to ask whether Lemma 1 admits a generalization to the case of all com-
plete metrizable topological vector spaces of cardinality c. A canonical nonseparable rep-
resentative of this class is the Banach space
l
={(x
n
)
n<
R
: sup
n<
|x
n
| < +},
which plays an important role in many questions of functional analysis and probability
theory. It turns out that the corresponding generalization of Lemma 1 is possible under the
Continuum Hypothesis. However, the argument should be essentially changed in order to
obtain the desired result. We need the following auxiliary statement which is important in
topological measure theory (see, e.g., [192] or Appendix 3).
Lemma 2. Let T be a metrizable topological space whose topological weight is not a
real-valued measurable cardinal. Let be an arbitrary -nite Borel measure on T. Then
possesses a separable support, i.e., there exists a closed separable set F T such that
(T \ F) = 0.
Adetailed proof of Lemma 2 is given in Appendix 3. Also, some applications of this lemma
are presented therein.
Remark 4. As known, Martins Axiom implies that the cardinality of the continuum c
is not real-valued measurable (see, for instance, [10], [40], and [67]). This fact is a direct
consequence of the existence of generalized Luzin sets on the real line R and of the circum-
stance that all generalized Luzin sets are universal measure zero (under Martins Axiom).
Lemma 3. Assume Martins Axiom. Let E be a complete metrizable topological vector
space whose topological weight is equal to c. There exist two universal measure zero
sets U
1
E and U
2
E such that each of them is a vector space over Q and U
1
+U
2
is
absolutely nonmeasurable in E.
Proof. We only sketch the proof leaving all corresponding technical details to the reader.
The argument is based on the fact that the family of all separable subspaces of E generates
a c-additive ideal of subsets of E. Therefore, in view of Lemma 2, any Luzin type set for
this ideal (cf. Exercise 14 for Appendix 1) turns out to be universal measure zero. Further,
by using the method of transnite induction, it can be proved that there exist two universal
measure zero sets U E and V E which simultaneously are vector spaces over Q and
satisfy the relations
U V ={0}, U +V = E.
Applying again the method of transnite induction, we can show that there exists a vector
subspace V
of V such that U +V
is a Bernstein set in E (cf. the proof of Theorem 4).

Notice that V
is a subset of a universal measure zero set and is also universal measure

zero. Finally, it follows from Theorem 3 and Lemma 2 that any Bernstein subset of E is
absolutely nonmeasurable in E, so we may put
U
1
=U, U
2
=V
,
which completes the proof.
The following statement is a straightforward consequence of Lemma 1 and Lemma 3 for-
mulated above.
Theorem 5. Suppose that the ContinuumHypothesis holds. Let E be an arbitrary complete
metrizable topological vector space with card(E) = c. There exist two universal measure
zero sets U
1
E and U
2
E which are vector spaces over Q and whose Minkowskis sum
U
1
+U
2
is absolutely nonmeasurable in E.
Remark 5. We do not know whether the assertion of Theorem 5 remains valid under
Martins Axiom.
Some relatively recent works were devoted to algebraic sums of small sets (see, for in-
stance, [37], [44], [131], and [132]).
Let us point out that in papers [131] and [132] similar questions are considered for an un-
countable commutative group (G, +) which is not assumed to be endowed with a topology
but is equipped with a nonzero -nite complete G-invariant or, more generally, G-quasi-
invariant measure (cf. also Chapter 11). Obviously, for this , the question can be posed
whether there are two -measure zero sets in G whose algebraic sum is nonmeasurable
with respect to . In this context, the following open problem seems to be of interest (cf.
the formulation of Theorem 1).
Problem. Let (G, +) be an uncountable commutative group and let be a nonzero -nite
complete G-invariant (G-quasi-invariant) measure on G such that A+B = G for some two
-measure zero sets A and B. Do there exist two -measure zero sets X and Y such that
X +Y is nonmeasurable with respect to ?
In Chapter 11 we were also concerned with a related problem which was formulated in
terms of G-absolutely negligible sets and G-absolutely nonmeasurable sets. In view of the
said above, it is reasonable to recall its formulation once again.
Problem. Let (G, +) be an uncountable commutative group. Do there exist two G-
absolutely negligible sets A and B in G such that their algebraic sum A+B is G-absolutely
nonmeasurable?
We know that if G is a vector space over Q with card(G) c, then the latter problem
has a positive solution. In particular, under CH, this problem is positively solved for all
uncountable vector spaces over Q and among them are R
n
(n 1), R
, c
0
, l
1
, l
2
, and l
.
EXERCISES
1
. Let E = {0} be a normed vector space or, more generally, locally convex Hausdorff
topological vector space. Prove that there exist two rst category subsets X and Y of E
such that X +Y does not have the Baire property in E.
For this purpose, use the Hahn-Banach theorem and reduce the argument to the particular
case E = R.
2. Let E ={0} be a Polish topological vector space. Show that if Martins Axiom is valid,
then there exists a generalized Luzin set L E such that L+L is a Bernstein subset of E.
3. Let E = {0} be a Polish topological vector space. Assuming Martins Axiom, demon-
strate that there exist two generalized Luzin sets L
1
and L
2
which are vector spaces over
Q and whose algebraic sum L
1
+L
2
is a Bernstein set. In this case, we obviously have
L
1
= L
2
.
4. Verify the validity of the assertion of Remark 1 concerning the fact that the results
presented in this chapter do not need the full power of Martins Axiom.
5
. Let E be a complete metric space of cardinality continuum. Suppose that E does not
contain isolated points. Show that there exists at least one Bernstein subset of E.
For this purpose, take into account the circumstance that the family of all closed separable
subspaces of E is also of cardinality continuum and that every nonempty perfect subset of
E contains a topological copy of the Cantor space {0, 1}
which trivially is separable.

Moreover, show that the following two assertions are equivalent:
(a) c is not real-valued measurable;
(b) if E is any complete metric space of cardinality c, then the class of all Bernstein subsets
of E coincides with the class of all absolutely nonmeasurable subsets of E.
In order to establish the equivalence (a) (b), apply the special case of Lemma 2 stating
that if c is not measurable in the Ulamsense, then for every metric space E with card(E) =c
and for any -nite Borel measure on E, there exists a separable support of .
6. Complete the details of the proof of Lemma 3.
7. Let X
1
and X
2
be any two universal measure zero topological spaces. Show that their
topological product X
1
X
2
is universal measure zero, too. Infer from this circumstance
that the topological product of a nite family of universal measure zero spaces is also
universal measure zero.
On the other hand, give an example of a countable family of universal measure zero spaces
whose topological product is not universal measure zero.
8
. Assuming Martins Axiom, prove that there exists a universal measure zero subset of
R whose image under some Borel mapping acting from R into R is not universal measure
zero.
For this purpose, use the previous exercise and take into account the existence of two gen-
eralized Luzin subsets L
1
and L
2
of R such that L
1
+L
2
= R.
9
. Let X be a universal measure zero subset of a Polish topological space E and let
f : E E
be a Borel mapping acting from E into a Polish topological space E
. Suppose that
(y E
)(card( f
1
(y)) ).
Show that the set f (X) is universal measure zero.
10
. Demonstrate that an injective continuous Hausdorff image of a universal measure zero

topological space is not, in general, universal measure zero.
For this purpose, take into account the fact that the rst uncountable cardinal
1
equipped
with the discrete topology is universal measure zero in view of the Ulam theorem [238] but
the same
1
equipped with its order topology is not universal measure zero (cf. Exercise 8
from Appendix 3).
Chapter 13
Absolutely nonmeasurable additive
Sierpi nski-Zygmund functions
Here we will focus our consideration on functions acting from R into R and having bad
structural properties (cf. Chapters 1 and 5). First of all, we mean those properties of
functions which are bad from the measure-theoretical point of view. Notice that from the
stand-point of general topology, bad functions are those which do not possess the Baire
property. Notice also that, in many respects, the Baire property can be treated as a natural
topological analogue of measurability (see, for instance, [40], [143], [148], [176], and
[192]).
Let E be a nonempty set and let M be a class of measures on E. In general, the domains of
measures from M may be different -algebras of subsets of E. Let f : E R be a function.
We recall that f is absolutely nonmeasurable with respect to M if f is nonmeasurable with
respect to any measure from M (see Chapter 5).
In connection with this notion, it makes sense to remember the following two typical ex-
amples.
Example 1. Let E = R = R
1
and let M denote the class of all translation-invariant exten-
sions of the Lebesgue measure =
1
on R. Let X be an arbitrary Vitali subset of R. Then
the characteristic function f =
X
of X is absolutely nonmeasurable with respect to M (see
Exercise 2 of Chapter 5, where a more general result is formulated).
Example 2. Let E again coincide with R and let M be the class of the completions of all
nonzero -nite diffused Borel measures on R. Denote by X a Bernstein subset of R and
let g =
X
be the characteristic function of X. As we know from Theorem 3 of Chapter 12,
the function g is absolutely nonmeasurable with respect to M. Below, another interesting
and important example of a function f : R R will be given which also turns out to be
absolutely nonmeasurable with respect to the same class M.
Now, denote by M
E
the class of all nonzero -nite diffused measures on E. Let us em-
215
3,
phasize once more that the domains of measures from this class are various -algebras of
subsets of E.
We recall that a function f : E R is absolutely nonmeasurable if f is absolutely nonmea-
surable with respect to M
E
.
Also, remember that a set X R is universal measure zero if for every -nite diffused
Borel measure on R, the equality
(X) = 0 is satised, where
as usual denotes the

outer measure associated with .
It was repeatedly mentioned in preceding sections that there exist uncountable universal
measure zero subsets of R (see, e.g., [69], [79], [148], [172], and Appendix 1). A much
stronger version of this result is contained in [197] and [250] where delicate examples of
uncountable universally small subsets of R are constructed.
Let us recall a characterization of absolutely nonmeasurable functions which was obtained
in terms of universal measure zero sets and pre-images of singletons (see Chapter 5).
Namely, for a function f : E R, the following two relations are equivalent:
(1) f is absolutely nonmeasurable;
(2) the range of f is a universal measure zero subset of R and
card( f
1
(t))
for each point t R.
A detailed proof of the equivalence (1) (2) was given in Chapter 5. In the same chapter,
the question was considered whether Vitali type functions are absolutely nonmeasurable
with respect to certain classes of extensions of . In this context, let us remember several
notions.
Let Q denote the set of all rational numbers.
A function f : R R is of Vitali type if the following two conditions hold:
a) f (x) x Q for every x R;
b) ran( f ) is a selector of the partition R/Q.
It was shown in Chapter 5 that if f : R R is an arbitrary Vitali type function, then the
Lebesgue measure can be extended to a measure on R such that f turns out to be
measurable with respect to . In other words, f is relatively measurable with respect to the
class of all extensions of .
Let c as usual denote the cardinality of the continuum. We already know that the following
three assertions are direct consequences of the equivalence between the relations (1) and
(2) and of the existence of an uncountable universal measure zero set in R.
I. If card(E) >c, then there are no absolutely nonmeasurable functions f with dom( f ) =E.
Absolutely nonmeasurable additive Sierpi nski-Zygmund functions 217
II. There exists an injective absolutely nonmeasurable function f : E R, where E is an
arbitrary set with card(E) =
1
.
III. If the cardinality of any universal measure zero subset of R is strictly less than c, then
there are no absolutely nonmeasurable functions f such that dom( f ) = R.
In particular, the latter consequence shows that absolutely nonmeasurable functions f :
RR exist if and only if there exists a universal measure zero set X R with card(X) =c.
More generally, absolutely nonmeasurable functions
f : E R
exist if and only if there exists a universal measure zero subset X of R with card(X) =
card(E).
This circumstance also shows that the existence of an absolutely nonmeasurable function
acting from R into R cannot be established within the theory ZFC. Indeed, there are
models of set theory in which the negation of the Continuum Hypothesis holds and the
cardinality of any universal measure zero subset of R does not exceed the rst uncountable
cardinal
1
(see, for example, [172]). Clearly, in such a model we do not have an absolutely
nonmeasurable real-valued function dened on R.
In this context, notice that the nonexistence of absolutely nonmeasurable functions act-
ing from R into R can be deduced by using some other set-theoretical assumptions. For
instance, it is easy to see that if c is a real-valued measurable cardinal, then no function act-
ing from R into R is absolutely nonmeasurable because in this case R carries a probability
continuous measure with dom() = P(R) and, therefore, all functions g : R R turn
out to be measurable with respect to .
Recall that a function f : R R is additive (or satises the Cauchy functional equation) if
f (x +y) = f (x) + f (y) (x R, y R).
In other words, f is a solution of the Cauchy functional equation if and only if f is a ho-
momorphism of the additive group R into itself. There are many works devoted to the
Cauchy functional equation (see, e.g., [143] and references therein). Obviously, all contin-
uous solutions of the Cauchy functional equation are representable in the standard linear
form
x ax (x R),
where a R. Any such function is called a trivial solution of the Cauchy functional equa-
tion. By using the technique of Hamel bases, we can get many nontrivial solutions of
this equation. Actually, the family of all such solutions has cardinality 2
c
and all of them
are nonmeasurable in the Lebesgue sense and do not possess the Baire property (for more
details, see [40], [108], [143], [176], or Chapters 1 and 5). It should be pointed out that
nontrivial solutions of the Cauchy functional equation have found interesting applications
in the classical theory of equidecomposability of polyhedra (see, e.g., [18]).
As mentioned earlier, within ZFC it is impossible to establish the existence of absolutely
nonmeasurable solutions of the Cauchy functional equation. On the other hand, by assum-
ing Martins Axiom and applying some properties of generalized Luzin subsets of R, it can
be proved that there are additive absolutely nonmeasurable functions acting from R into R.
In other words, under Martins Axiom, there exist absolutely nonmeasurable solutions of
the Cauchy functional equation as seen in Chapter 5. Our main goal here is to strengthen
this result in a certain direction. Namely, assuming the same Martins Axiom, we intend to
show the existence of an additive absolutely nonmeasurable function acting from R into R,
which is also utterly bad from the topological point of view.
Sierpi nski and Zygmund constructed in their remarkable work [225] a function
f : R R
having the following property:
for each subset Y of R with card(Y) = c, the restriction f |Y is not continuous on Y.
This classical result of Sierpi nski and Zygmund was essentially motivated by a theorem of
Blumberg [15] stating that for any function g : R R, there exists an everywhere dense
subset Dof R such that the restriction g|Dis continuous on D (see Exercise 22 fromChapter
8). Obviously, the set D is innite because it is everywhere dense in R. Consequently, the
existence of a Sierpi nski-Zygmund function f : R R shows that within ZFC theory we
cannot assert the uncountability of D.
Let us briey sketch the standard construction of Sierpi nski-Zygmund type functions. The
construction is not effective, i.e., it appeals to the Axiom of Choice which is a necessary
component in constructions of this sort.
Let be the least ordinal number of cardinality continuum, let {x
: <} denote an in-

jective family of all points of R, and let {g
: <} be the family of all partial continuous

functions acting from R into R whose domains are uncountable G
-subsets of R.
By using the method of transnite recursion, we dene a function
f : R R.
Suppose that for an ordinal <, all values f (x
) ( <) have already been dened and

consider the point x
. Since the set

{g
(x
) : <, x
dom(g
)}
is of cardinality strictly less than c, we may pick a point
y R\ {g
(x
) : <, x
dom(g
)}
and then put f (x
) = y. Proceeding in this manner, the desired function f will be com-

pletely determined.
It immediately follows from the above denition of f that
card({x R : f (x) = g(x)}) < c
for any partial continuous function g : R R whose domain is an uncountable G
-subset
of R. But the last circumstance implies that we also have
card({x R : f (x) = h(x)}) < c
for any partial continuous mapping h : R R whose domain is of cardinality continuum.
Indeed, according to Lavrentievs theorem (see Exercise 13 from Chapter 8), there exists a
partial continuous mapping h
: RR extending h and dened on a G
-subset of R. Since
card(dom(h)) = c, we claim that card(dom(h
)) = c. Consequently,
card({x R : f (x) = h
(x)}) < c,
which immediately yields
card({x R : f (x) = h(x)}) < c.
From the said above, we readily conclude that for any subset X of R with card(X) = c, the
restriction f |X is not a continuous mapping. Thus, f is a Sierpi nski-Zygmund function.
Various works are devoted to Sierpi nski-Zygmund functions (see, e.g., [7], [42], [123],
[124], [185], and [199]). In those works different constructions are presented which yield
further examples of Sierpi nski-Zygmund functions with additional properties important
from the view-point of real analysis.
Let M
R
denote the class of the completions of all nonzero -nite diffused Borel mea-
sures on the real line R. It is not difcult to prove that every Sierpi nski-Zygmund function
f : R R turns out to be absolutely nonmeasurable with respect to M
R
(see Exercise 2 of
the present chapter). In particular, f is nonmeasurable with respect to the Lebesgue mea-
sure on R. At the same time, as mentioned above, we cannot assert that f is absolutely
nonmeasurable. Moreover, it may happen that the graph Gr( f ) of f is a ( )-thick
subset of the plane R
2
and, in this case, f becomes measurable with respect to a suitable
extension of (cf. Chapter 7). A much stronger result will be obtained in the next chap-
ter. Namely, it will be demonstrated that there exists a Sierpi nski-Zygmund function mea-
surable with respect to some translation-invariant extension of and a certain transnite
construction of such a function will be done within ZFC theory.
The following statement shows that under an appropriate set-theoretical assumption, there
are absolutely nonmeasurable additive functions acting from R into R which are not
Sierpi nski-Zygmund functions.
Theorem 1. Assume Martins Axiom. There exists a function
f : R R
1) f is injective and additive;
2) f is absolutely nonmeasurable;
3) the restriction of f to some set K R with card(K) = c is the identical transformation
of K and, consequently, f is not a Sierpi nski-Zygmund function.
Proof. We recall that a set L R is a generalized Luzin subset of R if L is of cardinality
continuumand for every rst category set P R, the inequality card(PL) < c holds true.
It is well known that under Martins Axiom, there exists a generalized Luzin subset L of
the real line, which is simultaneously a vector space over the eld Q of all rational numbers
(see, e.g., [40], [59], [172], or Chapter 12 of this book). Let H
L
denote a Hamel basis for
L. Clearly, we have
card(H
L
) = c.
Extend H
L
to a Hamel basis H of R. Let
f
0
: H H
L
be a bijection such that for some set K H
L
with card(K) = c, the restriction of f
0
to K
is the identical transformation of K. The bijection f
0
can be extended to an isomorphism
f (linear over Q) between R and L. Obviously, we may consider f as an injective group
homomorphism from R into R whose range coincides with L. Since L is a universal mea-
sure zero subset of R (see Exercises 1 and 8 of Chapter 5 or Exercise 14 of Appendix 1)
and since the function f is injective, we infer that f is absolutely nonmeasurable. At the
same time, the restriction f |K coincides with the restriction f
0
|K. Consequently, f |K is the
identical transformation of K. This circumstance immediately implies that our f cannot be
a Sierpi nski-Zygmund function, which completes the proof of Theorem 1.
The next statement establishes within ZFC that for any GR which is a vector space over
Q and has cardinality c, there is an injective additive Sierpi nski-Zygmund function acting
from R into G.
Theorem 2. Let G be as above. There exists a function
f : R R
1) f is an injection;
2) f is a nontrivial solution of the Cauchy functional equation;
3) ran( f ) G;
4) f is a Sierpi nski-Zygmund function.
Proof. We shall construct the required function f by using the method of transnite re-
cursion. Consider c as an initial ordinal number equipped with its natural well-ordering
which will be denoted by . Of course, this notation means that for any ordinal < =c,
the relation card() < c is valid.
Let be a well-ordering of R isomorphic to .
Let {h
: < } be an enumeration of all Borel mappings acting from uncountable Borel

subsets of R into R.
Under this notation, we are going to dene three -sequences
{x
: <}, {V
: <}, { f
: <}
satisfying the following relations:
(a) {x
: <} is a Hamel basis of R;

(b) for each ordinal < , the set V
coincides with the vector subspace of R, over the

eld Q of all rational numbers, generated by {x
: };
(c) for each ordinal < , the function f
is a group monomorphism acting from V
into
G;
(d) if < <, then f
extends f
;
(e) if <, then we have
f
(qx
+v) = h
(qx
+v)
for all
<, q Q\ {0}, v {V
: <}
such that
qx
+v dom(h
).
Suppose that for an ordinal <, the partial -sequences
{x
: <}, {V
: <}, { f
: <}
have already been constructed. Let us put:
V
={V
: <},
f
={f
: <}.
Applying (c) and (d), we claim that f
is a group monomorphismacting fromV
into G and
card(V
) card() + < c.
Let x denote the least element (with respect to ) of the set R\V
. We put x
= x and
denote by V
the vector space over Q generated by V
{x}. Further, we dene

D ={(1/q)(h
(qx +v) f
(v)) : q Q\ {0}, <, v V
, qx +v dom(h
)}.
Since card(D) < c and card(ran( f
)) < c, we may choose an element

y G\ (Dran( f
)).
Clearly, there exists a unique group monomorphism
f
: V
G
extending f
and such that f
(x) = y.
Proceeding in this manner, we get the required families
{x
: <}, {V
: <}, { f
: <}.
Finally, denote
f ={f
: <}.
It easily follows from our construction that {x
: <} is a Hamel basis of R and, there-

fore,
dom( f ) ={V
: <} = R.
Since all f
( < ) are group monomorphisms, f is an injective homomorphism acting

from R into G. Also, it can readily be veried that for any Borel function h
( < ), we
have
card({z dom(h
) : f (z) = h
(z)}) card() + < c.

This relation shows that f is a Sierpi nski-Zygmund function (cf. [148] and [225]) and
completes the proof.
From Theorem 2 we easily obtain the following result.
Theorem 3. Assuming Martins Axiom, there exists an injective additive absolutely non-
measurable Sierpi nski-Zygmund function.
Proof. Let L denote again a generalized Luzin subset of R which simultaneously is a vector
space over Q (see Chapter 5 or Chapter 12). Let in Theorem 2 the role of G be played by
L. Then we get a group monomorphism
f : R L
such that f is also a Sierpi nski-Zygmund function. Now, keeping in mind the fact that L is
universal measure zero and taking into account a characterization of absolutely nonmeasur-
able functions, we see that f is an absolutely nonmeasurable Sierpi nski-Zygmund function.
Theorem 3 has thus been proved.
From the said above we can conclude that there are Sierpi nski-Zygmund functions with
ultimately bad measure-theoretical properties. On the other hand, in the next chapter the
existence of a Sierpi nski-Zygmund function will be established which is relatively measur-
able with respect to the class of all translation-invariant extensions of the Lebesgue measure
on R.
Finally, it is reasonable to recall in this place that under Martins Axiom, there exist in-
jective additive absolutely nonmeasurable functions f : R R, which are not Sierpi nski-
Zygmund functions (see Theorem 1 of the present chapter).
EXERCISES
1. Show that no Sierpi nski-Zygmund function has the Baire property.
2. Let M
R
denote the class of the completions of all nonzero -nite diffused Borel mea-
sures on R. Prove that every Sierpi nski-Zygmund function is absolutely nonmeasurable
with respect to M
R
.
For this purpose, consider any Sierpi nski-Zygmund function f : R R and x a nonzero
-nite diffused Borel measure on R. Suppose that f is measurable with respect to the
completion
of . Since R is a Radon space,
is a Radon measure. Consequently,

there exists a compact set K R with
(K) = (K) > 0 such that the restriction f |K is

continuous. Take into account the diffusedness of and infer that card(K) > fromwhich
it follows that card(K) = c. So a contradiction is obtained with the assumption that f is a
Sierpi nski-Zygmund function.
3. A function f : R R is called weakly continuous at a point x R if there exist two
sequences {t
n
: n < } and {r
n
: n < } of strictly positive real numbers tending to zero,
such that
lim
n
f (x t
n
) = lim
n
f (x +r
n
) = f (x).
Verify that every function f acting fromR into R is weakly continuous at all points of some
co-countable subset D of R. Also, check that D essentially depends on f .
In particular, any Sierpi nski-Zygmund function is weakly continuous on a co-countable
subset of R.
4. By using the method of transnite recursion, construct a Sierpi nski-Zygmund function
f : R R whose graph Gr( f ) is
2
-thick in R
2
.
Formulate and prove an analogous result in terms of the Baire property.
5. Show that there exists a Sierpi nski-Zygmund function f : R R whose graph Gr( f ) is
a
2
-measure zero subset of the plane R
2
.
6. Show that there exists a Sierpi nski-Zygmund function f : R R whose graph Gr( f ) is
a rst category subset of R
2
.
7
. Prove that there exists a family of functions

f
i
: R R (i I)
(a) card(I) > c;
(b) every f
i
(i I) is a Sierpi nski-Zygmund function;
(c) this family of functions is almost disjoint, i.e., the inequality
card(Gr( f
i
) Gr( f
j
)) < c
holds true for any two distinct indices i I and j I.
8
. Assuming Martins Axiom, give an example of a Sierpi nski-Zygmund function f : R

R such that for some set X R of cardinality c, the restriction f |X is a semicontinuous
function on X.
9. Let E and E
be two innite sets and let be a family of partial functions acting from E
into E
. Suppose that
card(E) = card(E
) = card().
Show that there exists a mapping
f : E E
such that for any set X E with card(X) =card(E) and for any partial function , the
relation f |X =|X is satised.
In particular, take:
E = a separable metric space of cardinality c;
E
= the real line R;

= the family of all partial continuous functions : E R such that dom() is a G
-set
in E of cardinality c.
Then the above-mentioned mapping f : E R turns out to be a Sierpi nski-Zygmund type
function on E.
10. By using the result of the previous exercise and of Exercise 16 from Chapter 8, prove
that there exists a function
f : R R
for every set X R with card(X) = c, the restriction of f to X is not a Borel mapping.
Conclude that any such f turns out to be a Sierpi nski-Zygmund function but the converse
assertion is not true in general (see Exercise 8 above).
11
. A set D R is called nowhere Lebesgue measure zero if for each non-degenerate

interval J R, the intersection J D has strictly positive outer Lebesgue measure, i.e.,
(J D) > 0.
Show that there exists a function f : RR such that for every set DR which is nowhere
Lebesgue measure zero, the restriction f |D is discontinuous at all points of D.
In order to demonstrate this fact, consider a partition {X
n
: n < } of R such that X
0
is an
everywhere dense G
-set in R of -measure zero and all sets X

n
(0 < n <) are nowhere
dense in R. Further, dene a function f : R R as follows:
f (x) = n if and only if x X
n
.
Verify that f is the required one.
For this purpose, take an arbitrary nowhere Lebesgue measure zero subset D of R and pick
an arbitrary point x in R. Only two cases are possible.
(a) There exists a neighborhood U(x) of x which intersects only nitely many sets from
the family {X
n
: n < }. In this case, utilizing the fact that all sets X
n
(0 < n < ) are
nowhere dense in R, we can nd a nonempty open interval J U(x) such that J X
0
.
Since (X
0
) = 0, the interval J must also have Lebesgue measure zero. But this yields an
obvious contradiction, so we claim that the situation just described is impossible.
(b) Any neighborhood U(x) of x intersects innitely many sets from the family {X
n
: n <
}. In this case, it follows from the denitions of f and D that if x D, then f |D is
unbounded on every neighborhood of x. The last circumstance directly implies the discon-
tinuity of f |D at all points of D.
The above-mentioned result is due to Brown [27]. It can easily be seen that this result
does not need uncountable forms of the Axiom of Choice which are necessary in various
constructions of Sierpi nski-Zygmund type functions on R.
12. Let E be a set of cardinality c and let {T
i
: i I} be a family of topologies on E
satisfying the following two conditions:
(a) card(I) c;
(b) each topology T
i
(i I) has a countable base.
Prove that there exists a function f : E R such that for any topology T
i
(i I), the
corresponding mapping
f : (E, T
i
) R
is a Sierpi nski-Zygmund type function for the space (E, T
i
).
Chapter 14
Relatively measurable Sierpi nski-Zygmund
functions
In this chapter, we again will be dealing with Sierpi nski-Zygmund functions. Our main
goal remains the same. Namely, we consider structural properties of Sierpi nski-Zygmund
type functions from the measure-theoretical point of view and associate those properties
with the measure extension problem (see Chapter 2 where three different aspects of this
problem are discussed).
The reader may think that the title of the present chapter appears misleading because all
Sierpi nski-Zygmund functions are nonmeasurable with respect to the Lebesgue measure
on the real line R (in this connection, see Exercise 2 from Chapter 13 where a more general
result is formulated). But here we are going to prove that there exist some Sierpi nski-
Zygmund functions which are measurable with respect to certain translation-invariant ex-
tensions of .
Let us remember several facts concerning absolutely nonmeasurable and relatively measur-
able real-valued functions.
Let E be a nonempty set and let f : E R be a function.
We recall that f is absolutely nonmeasurable if f is nonmeasurable with respect to any
nonzero -nite diffused (i.e., continuous) measure dened on a -algebra of subsets of
E.
Recall also that a set X R is universal measure zero if for every -nite diffused Borel
measure on R, the equality
(X) = 0 is satised, where
as usual denotes the outer

measure associated with .
As already said in preceding parts of the book, there exist uncountable universal measure
zero subsets of R (see [69], [79], [148], [172], and Appendix 1).
In addition, we have a characterization of absolutely nonmeasurable functions in terms of
universal measure zero sets and pre-images of singletons. Namely, as shown in Chapter 5,
227
4,
for a given function
f : E R,
the following two assertions are equivalent:
(1) f is absolutely nonmeasurable;
(2) the range of f is a universal measure zero subset of R and the inequality card( f
1
(t))
holds for each point t R.
For the proof of the equivalence (1) (2), see again Chapter 5.
Let c denote the cardinality of the continuum. We would like to mention the following two
straightforward consequences of the equivalence between the assertions (1) and (2).
I. If card(E) > c, then there are no absolutely nonmeasurable functions f : E R;
II. If the cardinality of any universal measure zero subset of R is strictly less than c, then
there are no absolutely nonmeasurable functions f : R R.
The latter consequence shows us that absolutely nonmeasurable functions f : R R exist
if and only if there exists a universal measure zero set X R with card(X) = c. More
generally, absolutely nonmeasurable functions
f : E R
do exist if and only if there exists a universal measure zero subset X of R such that
card(X) = card(E).
This circumstance also implies that the existence of an absolutely nonmeasurable function
acting from R into R cannot be established within the theory ZFC. But there exists an
injective absolutely nonmeasurable function f : E R, where E is an arbitrary set of
cardinality
1
, and the latter fact holds true within ZFC theory (see Appendix 1).
In this context it makes sense to recall that by assuming Martins Axiom and applying
some properties of generalized Luzin subsets of R, it can be proved that there are injective
additive absolutely nonmeasurable functions acting from R into R. In other words, under
Martins Axiom, there exist absolutely nonmeasurable solutions of the Cauchy functional
equation (see Chapters 5 and 12). Some further extensions of this result are closely con-
nected with Sierpi nski-Zygmund type functions (cf. Theorem 3 of the previous chapter).
As pointed out in Chapters 1 and 13, Sierpi nski and Zygmund constructed in their classical
work [225] a function
f : R R
Relatively measurable Sierpi nski-Zygmund functions 229
for each subset Y of R with card(Y) = c, the restriction f [Y is not continuous on Y.
It was also mentioned in the same chapters that the result of Sierpi nski and Zygmund may
be regarded as a counter-version of Blumbergs remarkable theorem [15] stating that for
any function
g : R R,
there exists an everywhere dense subset D of R such that the restriction g[D is continuous
on D (see Exercise 22 from Chapter 8). Obviously, the set D must be innite. On the
other hand, the existence of a Sierpi nski-Zygmund function shows that we cannot assert, in
general, the uncountability of D.
There are many works devoted to Sierpi nski-Zygmund functions. In those works vari-
ous constructions are presented, which yield further interesting examples of Sierpi nski-
Zygmund functions with additional properties important from the view-point of real anal-
ysis, measure theory, and general topology (see [42] and references therein).
It is known that every Sierpi nski-Zygmund function f turns out to be nonmeasurable with
respect to the completion of any nonzero -nite diffused Borel measure on R (see Exer-
cise 2 for Chapter 13). In other words, every Sierpi nski-Zygmund function is bad from the
view-point of topological measure theory.
At the same time, we cannot assert that any Sierpi nski-Zygmund function f is absolutely
nonmeasurable, because it may happen (e.g., in those models of set theory where the Con-
tinuum Hypothesis fails to be true) that the inequalities
c > card( f
1
(t)) >
hold for some point t R. In this case, by the equivalence of assertions (1) and (2), f must
be measurable with respect to a certain nonzero -nite diffused measure on R. The present
chapter is devoted to some much stronger measurability properties of certain Sierpi nski-
Zygmund functions. As already said in the beginning of this chapter, the main goal is
to demonstrate that there exist Sierpi nski-Zygmund functions which are measurable with
respect to appropriate translation-invariant extensions of the standard Lebesgue measure
on R.
For this purpose, we need several auxiliary propositions.
Lemma 1. Let f : R R be a function satisfying the following condition:
for each subset Y of R with card(Y) = c, the restriction f [Y is not a Borel mapping.
Then f is a Sierpi nski-Zygmund function.
This lemma is trivial; however, it plays an essential role below.
It is reasonable to point out that in Exercise 9 from Chapter 13 a general statement is for-
mulated concerning the existence of Sierpi nski-Zygmund type functions. It readily yields
the existence of a function f : R R which satises the condition of Lemma 1. Notice
also that under Martins Axiom, there are Sierpi nski-Zygmund functions for which this
condition is not fullled (see Exercise 8 of the same Chapter 13).
As usual, we denote by T (=S
1
) the one-dimensional unit torus in the Euclidean plane R
2
.
Recall that T is a commutative compact topological group canonically isomorphic to the
group of all rotations of the plane R
2
= RR about its origin (0, 0).
Lemma 2. Let : R T be the canonical surjective continuous group homomorphism
dened by
(x) = (cos(x), sin(x)) (x R).
There exists a mapping : T R such that:
1) coincides with the identity transformation of T;
2) is discontinuous only at one point of T (hence, is a Borel mapping).
We omit an easy proof of Lemma 2. Actually, in the sequel we only need the fact that is a
Borel mapping. So we may treat this simple lemma as a direct consequence of the theorem
of Kuratowski and Ryll-Nardzewski [151] on measurable selectors (see Appendix 2).
The torus T regarded as a compact topological group carries the Haar probability measure
. Obviously, the completion of coincides with the Lebesgue probability measure on the
torus T, which is invariant under the group all translations of T.
As usual, we denote by the product measure whose multipliers are and . For our
further purposes, it will be convenient to denote by the same symbol the completion
of the above-mentioned product measure.
We recall that a function f : R T has thick graph with respect to if every ( )-
measurable set Z with ( )(Z) > 0 intersects the graph Gr( f ) of f ; in other words, f
has thick graph if
( )
((RT) Gr( f )) = 0.
We need the following lemma (cf. Chapters 2 and 7).
Lemma 3. Let f : R T be a group homomorphism whose graph Gr( f ) is thick with
respect to . For any set Z dom( ), let us put
Z
/
=x R : (x, f (x)) Z
and introduce the class of sets
S
/
=Z
/
: Z dom( ).
Finally, dene a functional on S
/
by the formula
(Z
/
) = ( )(Z) (Z
/
S
/
).
Then the following relations hold:
1) S
/
is a -algebra of subsets of R containing dom() and invariant under the group of
all isometric transformations of R;
2) the functional is well dened;
3) is a measure on S
/
extending and invariant under the same group of all isometries
of R;
4) the original homomorphism f is measurable with respect to .
Proof. We use an argument similar to that given in the paper [141] (cf. also [115] or
Chapters 2 and 7).
If Z dom( ) and Z
i
dom( ) for each index i from a countable set I, then we
can write
R Z
/
=x R : (x, f (x)) (RT) Z,
Z
/
i
: i I =x R : (x, f (x)) Z
i
: i I.
Now, taking into account the relations
(RT) Z dom( ), Z
i
: i I dom( ),
we derive that S
/
is a -algebra of subsets of R.
The invariance of S
/
under the group of all isometries of R follows directly from the
invariance of dom( ) under all translations of the product group RT and under the
symmetry of this group.
If for a set Z
/
S
/
, we have simultaneously
Z
/
=x R : (x, f (x)) Z
1
and
Z
/
=x R : (x, f (x)) Z
2
,
where Z
1
and Z
2
are some ( )-measurable sets, then
x R : (x, f (x)) Z
1
Z
2
= / 0.
In view of the ( )-thickness of the graph Gr( f ) of f , we infer that
( )(Z
1
Z
2
) = 0,
which also implies that
( )(Z
1
) = ( )(Z
2
).
This equality establishes the correctness of the denition of .
In a similar way, we prove the countable additivity of and its invariance under the group
of all isometries of R.
Further, if X is an arbitrary -measurable set, then we can write
X =x R : (x, f (x)) X T,
X S
/
,
(X) = ( )(X T) = (X) (T) = (X)
and, therefore, the measure extends .
Finally, for any Borel set B T, we have
f
1
(B) =x R : (x, f (x)) RB,
RB dom( ),
f
1
(B) S
/
,
which shows that f is measurable with respect to and completes the proof of Lemma 3.
Lemma 4. Let (G, +) be an innite commutative divisible group such that, for any nonzero
natural number n and for any element g G, the equation
nx = g
has at most countably many solutions in G. Let A be an arbitrary subset of G. There exists
a subgroup [A] of G satisfying the following conditions:
1) A [A];
2) card([A]) = card(A) +;
3) for each nonzero natural number n and for each element g [A], all solutions of the
equation nx = g belong to [A] and, in particular, [A] is a divisible subgroup of G.
Proof. Let B be any countably innite subset of G. Construct by recursion an innite
sequence
(G
0
, G
1
, ..., G
k
, ...)
of subgroups of G. First of all, put:
G
0
= the group generated by AB.
Suppose now that the subgroup G
k
has already been dened satisfying the following equal-
ity:
card(G
k
) = card(A) +.
Denote by X
k
the set of all solutions of the equations
nx = g (n = 1, 2, ..., g G
k
).
Taking into account the assumption of the lemma, we get
card(X
k
) card(G
k
) + = card(A) +.
Let us put:
G
k+1
= the group generated by the set G
k
X
k
.
Proceeding in this way, we obtain the increasing sequence
G
0
G
1
... G
k
...
of subgroups of G. Finally, denote
[A] =G
k
: k < .
A straightforward verication shows that [A] is the required group.
Remark 1. Lemma 4 has a direct analogue for innite non-commutative divisible groups
(G, ) and the proof of that analogue can be carried out similarly to the proof of Lemma 4.
Lemma 5. There exists a group homomorphism
f : R T
possessing the following properties:
1) the graph of f is thick with respect to the product measure ;
2) for any uncountable Borel subset Y of R and for any Borel mapping h : Y T, we have
the inequality
card(x R : f (x) = h(x)) < c.
Proof. We shall construct the required homomorphism f by using the method of transnite
recursion (cf. the proof of Theorem 2 from Chapter 13).
Let denote the least ordinal number of cardinality c and let _ be an arbitrary well-
ordering of R isomorphic to .
Let h
: < be an enumeration of all Borel mappings acting from uncountable Borel

subsets of R into T.
Let Z
: < be an enumeration of all those Borel subsets of RT whose ( )-

measure is strictly positive. We may assume, without loss of generality, that the range of
the partial family
Z
: < , is an odd ordinal

coincides with the range of the entire family Z
: < .
Under this notation, we are going to dene four -sequences
x
: <, y
: < ,
V
: < , f
: <
(a) x
: < is a Hamel basis of R;

(b) for each ordinal < , the set V
is the vector subspace of R, over the eld Q of all

rational numbers, generated by x
: ;
(c) for each ordinal < , we have the group homomorphism
f
: V
T
such that f
(x
) = y
;
(d) if < < , then f
extends f
;
(e) for each odd ordinal < , we have (x
, y
) Z
;
(f) if < , then we have
f
(qx
+v) ,= h
(qx
+v)
for all
< , q Q 0, v V
: <
such that
qx
+v dom(h
).
Suppose that for an ordinal < , the partial -sequences
x
: < , y
: < ,
V
: < , f
: <
have already been constructed. Let us put:
V
/
=V
: < ,
f
/
=f
: < .
Applying (d), we may assert that f
/
is a group homomorphismacting fromV
/
into T. Now,
consider two possible cases.
1. The ordinal is even. In this case, let x be the least element (with respect to _) of RV
/
and let
A =h
(qx +v) : < , q Q 0, v V

/
, qx +v dom(h
) + f
/
(V
/
).
According to Lemma 4, we have
card([A]) = card(A) + card() + < c.
Choose an element y T [A] and put
x
= x, y
= y.
Taking into account the fact that T is a divisible group, we can extend f
/
to a group homo-
morphism
f
: V
T
satisfying the condition f
(x
) = y
.
2. The ordinal is odd. In this case, we apply the classical Fubini theorem to the set Z
and choose an element x RV

/
such that
(t T : (x, t) Z
) > 0.
Further, we choose an element
y t T : (x, t) Z
[A],
where A is determined as in the previous case, and put again
x
= x, y
= y.
Obviously, we have (x
, y
) Z
. As above, using the divisibility of T, we dene a group

homomorphism
f
: V
T
extending f
/
and satisfying the condition f
(x
) = y
.
Proceeding in this manner, we will be able to construct the required four -sequences.
Then we put
f =f
: < .
In view of (a), this f is a group homomorphismacting from R into T.
In view of (e), the graph of f is thick in the product space RT (with respect to the product
measure ).
Finally, relation (f) implies that the inequality
card(x R : f (x) = h
(x)) < c
is satised for any ordinal <. This circumstance also yields that for any set X R with
card(X) =c, the restriction of f to X is not a Borel mapping. To show the last fact, suppose
otherwise, i.e., suppose that f [X is Borel. Then, according to a well-known statement
from classical descriptive set theory, the function f [X can be extended to a T-valued Borel
function f
dened on a Borel subset of R (see, e.g., [58], [99], [148], or Exercise 16 from
Chapter 8). Since the relation
card(dom( f
)) = c
holds true, we conclude that f
= h
for some ordinal < . But this circumstance is

impossible in view of the above-mentioned inequality. Lemma 5 has thus been proved.
Theorem 1. There exists a Sierpi nski-Zygmund function
: R R
measurable with respect to some extension of which is invariant under all isometric
transformations of R.
Proof. Let f be as in Lemma 5. By virtue of Lemma 3, f is measurable with respect to a
certain measure on R which extends and is invariant under the group of all isometries
of R. Let us put
= f ,
where : T R denotes the Borel mapping of Lemma 2.
We assert that is the required Sierpi nski-Zygmund function.
Indeed, is a -measurable function as the composition of two functions, rst of which is
-measurable and the second one is Borel.
Let X be a subset of R with card(X) =c. It sufces to show that the restriction [X is not a
Borel mapping. Suppose otherwise, i.e., suppose that [X is Borel. Let : R T denote
the canonical epimorphism dened by
(x) = (cos(x), sin(x)) (x R).
By virtue of Lemma 2, the composition coincides with the identity transformation of
T. Therefore, the mapping
[X = f [X = f [X
must be Borel, which is impossible in view of Lemma 5. The contradiction obtained ends
the proof of our theorem.
Remark 2. We cannot assert that the function of Theorem 1 is additive. In this connec-
tion, the following question seems to be interesting.
Does there exist an additive Sierpi nski-Zygmund function acting from R into R and mea-
surable with respect to some translation-invariant extension of ?
We do not know an answer to the posed question. In this context, notice that by applying a
method analogous to the described above, it is possible to prove that there exists an additive
Sierpi nski-Zygmund function acting from R into R and measurable with respect to some
translation-quasi-invariant extension of (see Exercise 3 below).
Replacing R by a Polish topological vector space equipped with a nonzero -nite diffused
Borel measure, we come to a certain analogue of Theorem 1.
Theorem 2. Let E ,= 0 be a Polish topological vector space and let be a -nite
diffused Borel measure on E. There exist a measure
/
on E extending and a Sierpi nski-
Zygmund function : E R measurable with respect to
/
.
Moreover, if is invariant (quasi-invariant) under a group G E, then
/
can also be
chosen invariant (quasi-invariant) under the same G.
The proof of this statement is similar to the proof of Theorem 1 and is left to the reader.
Remark 3. It is well known that if E is an innite-dimensional Polish topological vector
space, then there exists no nonzero -nite Borel measure on E quasi-invariant under the
group of all translations of E (see Chapter 4). On the other hand, it should be mentioned
that, for a certain E, it is possible to construct a nonzero -nite Borel measure on E in-
variant under an everywhere dense vector subspace of E (see, e.g., [110], [115], or Chapter
3).
We would like to nish this chapter with a statement which once again underlines bad
behavior of all Sierpi nski-Zygmund functions from the points of view of the Baire property
and Lebesgue measurability.
Theorem 3. Let h : R R be an arbitrary Sierpi nski-Zygmund function. The following
two assertions are valid:
1) if every set in R of cardinality strictly less than c is of rst category, then for any sec-
ond category set X R, the restriction h[X cannot be extended to a function h
: R R
possessing the Baire property;
2) if every set in R of cardinality strictly less than c is of Lebesgue measure zero, then for
any set Y R of strictly positive outer Lebesgue measure, the restriction h[Y cannot be
extended to a function h
: R R measurable in the Lebesgue sense.

The proof of this statement is left to the reader (cf. Examples 7 and 8 from Chapter 1).
EXERCISES
1. Deduce a weaker formof Lemma 2 (requiring only the Borel measurability of ) directly
from the theorem of Kuratowski and Ryll-Nardzewski stating the existence of measurable
selectors of multi-valued functions (see Appendix 2).
2. Prove the assertion of Remark 1, that is establish an analogue of Lemma 4 for innite
non-commutative divisible groups (G, ).
3. Show that there exists a function
f : R R
(a) f is additive;
(b) f is a Sierpi nski-Zygmund function;
(c) f is measurable with respect to some extension of which is quasi-invariant under
the group of all isometries of R.
For this purpose, equip R with a probability measure equivalent to and dene f by the
scheme described in the proof of Lemma 5. Further, for checking the R-quasi-invariance
of , apply to the product measure Theorem 2 from Chapter 4.
4. Assuming Martins Axiom, show that there exists a function
g : R R
(a) g is a Sierpi nski-Zygmund function;
(b) for each set X R with card(X) = c, the restriction g[X is relatively measurable with
respect to the class of all nonzero -nite diffused measures on R;
(c) g is measurable with respect to some extension of which is invariant under the
group of all isometries of R.
For this purpose, use an argument similar to the proof of Theorem 1 of this chapter and
try to construct the required g in such a way that ran(g) would be a generalized Sierpi nski
subset of R (see Exercise 14 of Appendix 1).
5. Give a proof of Theorem 2 of this chapter by applying an argument analogous to the
proof of Theorem 1 and taking into account Theorem 2 from Chapter 4.
6. Give a proof of Theorem 3.
In connection with the last exercise, notice once more that according to a result of Novikov
[188], for any second category set X R, there exists a function f : X R which does not
admit an extension f
: R R having the Baire property and with dom( f
) = R.
The assertion of Theorem 3 is much stronger and shows that, under Martins Axiom, there
exists a universal function h : R R of Sierpi nski-Zygmund type such that for any second
category set X R, the restriction h[X cannot be extended to a function acting from R into
R and having the Baire property. However, certain additional set-theoretical assumptions
are needed for the validity of Theorem 3.
Chapter 15
A nonseparable extension of the Lebesgue
measure without new null-sets
Let E be a set, S be a -algebra of subsets of E containing all singletons in E, and let
be a nonzero -nite continuous, i.e., vanishing at the singletons measure on S.
We recall from Chapter 2 that the general measure extension problem is to extend to a
maximally large class of subsets of E.
According to Ulams theorem (see, e.g., [69], [91], [192], [238], or Appendix 1), it is
consistent with the axioms of contemporary set theory that the domain of any extension
/
of cannot coincide with the power set P(E) of E.
For instance, this circumstance holds if card(E) is strictly smaller than the rst inaccessible
cardinal number. Much stronger results of purely set-theoretical avor can be found in [10],
[69], [91], [150], and [231].
Consequently, there always exists a set X E such that X , dom(
/
). Then an easy ar-
gument shows that
/
can be extended to a measure
//
so that X becomes
//
-measurable
(see Chapter 2 or Example 2 below).
Thus, assuming that there are no large uncountable cardinals, we can conclude that there
are no maximal extensions of the original measure .
In this context, a result on extensions of measures should be recalled, which states that if
X
j
: j J is an arbitrary partition of E, then there exists a measure on E extending
and satisfying the relation
X
j
: j J dom().
For the proof of this result, see [1], [13], or Chapter 2. Notice that the measures under
consideration in [1] and [13] are assumed to be probability ones but the same argument
works for arbitrary -nite measures (see the proof of Theorem 2 from Chapter 2).
Actually, we do not need the above result in our further considerations but would like to
mention one of its immediate corollaries. Namely, if a nite family of subsets of E is
241
5,
given, then we can always extend to a measure which measures all these subsets. Indeed,
it sufces to consider the nite partition of E generated by all members of the given family.
An analogous assertion fails to be true for arbitrary countable families of subsets of E,
i.e., if we have a countable family X
n
: n N of subsets of E, then we cannot assert, in
general, that there exists a measure on E extending and satisfying the inclusion
X
n
: n N dom().
Moreover, the existence of a Luzin set L on the real line R having the cardinality of the
continuum(see Chapters 5 and 12) implies that there is a countably generated -algebra of
subsets of R containing all singletons and not admitting any nonzero -nite continuous
measure. Indeed, it is not difcult to show that every -nite continuous Borel measure
given on a separable metric space is concentrated on some rst category subset of the space
(see Lemma 2 of Appendix 3). Since all rst category subsets of L are at most countable,
it follows that L does not admit a nonzero -nite continuous measure on the countably
generated Borel -algebra of L. Now, using a one-to-one correspondence between L and
R, we obtain the required -algebra of subsets of R.
We know that the same L is a small subset of R because it has outer measure zero with
respect to any -nite continuous Borel measure on R. In other words, L turns out to be a
universal measure zero subspace of R. Recall that the existence of Luzin sets needs some
additional set-theoretical axioms. For instance, the Continuum Hypothesis readily implies
that there are Luzin subsets of R. Similarly to this circumstance, Martins Axiom implies
that there are so-called generalized Luzin subsets of R (see Chapter 12 and Exercise 14 of
Appendix 1).
In this connection, it should be underlined that the existence of analogous small subsets
of R having cardinality
1
, where
1
is the least uncountable cardinal, can be established
within the theory ZFC (see [69], [79], [148], [172], [197], [250], or Appendix 1). This fact
implies that there exists a certain countably generated -algebra of subsets of
1
separating
all points in
1
and not admitting a nonzero -nite diffused measure on it. Once again, we
would like to stress that the latter important result is valid within ZFC theory (see Theorem
5 of Appendix 1).
If E is a nonempty set, S is a -algebra of subsets of E and is a measure on E with
dom() =S, then we usually say that the triplet (E, S, ) is a measure space.
All measures considered below are assumed to be nonzero, -nite, and continuous.
If the need arises and without loss of generality, we can additionally suppose that a measure
under consideration is also complete. Indeed, if necessary we may always replace a given
A nonseparable extension of the Lebesgue measure without new null-sets 243
measure by its completion.
For any nonzero complete measure on a set E, we denote by I() the -ideal of all
-measure zero sets. For the sake of brevity, we shall say in the sequel that all members of
I() are -null-sets in E.
The symbol
(respectively,
) denotes the outer (respectively, inner) measure associated

with .
The symbol (=
1
) stands for the Lebesgue measure on R (= R
1
). Recall that is a
complete measure; however, sometimes it is convenient to denote by the same symbol
the restriction of the Lebesgue measure to the Borel -algebra B(R) of R.
As usual, the symbols (= N) and c denote, respectively, the least innite cardinal and
the cardinality of the continuum.
In this chapter we will be dealing with certain nonseparable extensions of the Lebesgue
measure . For the denition of separable and nonseparable measure spaces, see Exercise
12 of Chapter 1.
First of all, let us briey discuss several typical examples concerning extensions of various
-nite measures.
Example 1. A well-known method of extending a given -nite complete measure is
based on adding to I() some new sets which are nonmeasurable with respect to and
whose inner -measure is equal to zero (cf. [234], [235], and Chapter 2).
Proceeding in this way, we come to a larger -ideal I
/
which properly contains I() and
all of whose elements are of inner -measure zero.
Then we consider the -algebra S
/
generated by dom() and I
/
. As seen in Chapter 2,
any element U of this -algebra admits a representation
U = (X Y) Z,
where X dom(), Y I
/
and Z I
/
. We dene a functional
/
on S
/
by the formula
/
(U) =
/
((X Y) Z) = (X).
A straightforward verication shows that the functional
/
is well dened in the sense that
the value
/
(U) does not depend on a representation of U in the above-mentioned formand
/
is also a measure on S
/
extending the initial measure . In addition, all members of I
/
become sets of
/
-measure zero.
Example 2. The method of extending measures described in Example 1 has a weak side.
Indeed, fromthe view-point of the theory of Boolean algebras, the complete measure and
its extension
/
obtained in the previous example are the same. Nevertheless, by slightly
changing the above method, we can achieve some difference between a measure and its
extension if both of them are considered on the corresponding quotient Boolean algebras.
For this purpose, let us take any set T E nonmeasurable with respect to a complete
measure . Obviously, we must have
(T) <
(T).
If T
0
denotes a -measurable kernel of T and T
1
stands for a -measurable hull of T, then
(T
1
T
0
) > 0 and the set T T
0
regarded as a subset of T
1
T
0
satises the equalities
(T T
0
) =
((T
1
T
0
) (T T
0
)) = 0.
So we may assume (replacing, if necessary, T by T T
0
and E by T
1
T
0
) that
(T) =
(E T) = 0.
Let S
/
denote the -algebra of all those subsets U of E which admit a representation in
the form
U = (X T) (Y (E T)),
where X dom() and Y dom(). Dene a functional
/
on S
/
by the equality
/
(U) = (1/2)((X) +(Y)).
As earlier,
/
is well dened and turns out to be a measure on S
/
extending . Also, since
T is
/
-measurable,
/
strictly extends . Moreover, we see that the quotient Boolean
algebra associated with is properly contained in the quotient Boolean algebra associated
with
/
. At the same time, we have the relation
I(
/
) =I(),
i.e., the measure
/
does not produce new null-sets.
Further, considering more thoroughly the extensions of described in Example 1 and
Example 2, we readily observe that both of these constructions do not essentially change
the metrical structure of . Indeed, we can see that the metric space associated with a
measure and the metric space associated with its extension
/
obtained by using any of
the two described constructions have the same topological weight. Therefore, if the original
measure is separable (i.e., if its metric space is separable), then the extended measure
/
is separable, too.
In this context, we would like to recall that Kakutani and Oxtoby [95] gave a construction
of a certain nonseparable extension of the Lebesgue measure on the real line R. Another
construction of this kind was presented by Kodaira and Kakutani [141]. It is remarkable
that both those extensions of are invariant under the group of all isometries of R.
It should be observed that the topological weight of the metric space canonically associ-
ated with the extension of obtained by Kakutani and Oxtoby in [95] is equal to 2
c
. This
extension necessarily yields new null-sets. In other words, there appear null-sets which are
not of Lebesgue measure zero because the method of Kakutani and Oxtoby uses an un-
countable -independent family of subsets of R such that the intersection of any countable
subfamily of this family is a Lebesgue nonmeasurable set. Actually, the above-mentioned
intersection becomes a null-set with respect to the extended measure (for more details, see
[83], [108], or Exercise 4 of this chapter).
The topological weight of the metric space canonically associated with the extension of
obtained by Kodaira and Kakutani in [141] is equal to c. By applying their method, we can
also obtain new null-sets (see, for instance, Theorem 2 below).
In this connection, the following question seems to be interesting.
Does there exist a nonseparable extension of all null-sets of which are precisely the -
null-sets?
The question is of interest in view of the following circumstance.
In many topics of real analysis, measure theory, and probability theory the inner structure of
a measure under consideration does not play any role and only the induced concept almost
everywhere is essential. The standard example of this type is the well-known Lebesgue
theorem stating that a bounded function
f : [a, b] R
is integrable in the Riemann sense if and only if f is continuous almost everywhere on
[a, b]. In such a case, f is also -measurable because Luzins C-property automatically
holds for f .
The just mentioned Lebesgue theorem does not need the notion of the Lebesgue measure.
For its proof, it completely sufces to apply the notion of a null-set in the Lebesgue sense.
Of course, numerous other examples of this kind can be pointed out.
In this chapter, our goal is to demonstrate that under the ContinuumHypothesis there exists
a nonseparable extension of which yields no new null-sets.
It should be noticed that our argument may be regarded as a certain combination of the
method of Kodaira and Kakutani [141] with the method of Luzin [159] by means of which
he proved the existence of his set on the real line R (see, e.g., Chapter 12 or Exercise 14
from Appendix 1).
We begin with some preliminary considerations which were already utilized in preceding
parts of the book.
Let (E
1
, S
1
,
1
) and (E
2
, S
2
,
2
) be two measure spaces such that
1
is -nite and
2
is
a probability measure. Let f : E
1
E
2
be a mapping and let Gr( f ) denote the graph of f .
Suppose that Gr( f ) is (
1
2
)-thick in the product set E
1
E
2
, i.e., suppose that
(
1
2
)
((E
1
E
2
) Gr( f )) = 0.
As in preceding sections, for any set Z dom(
1
2
), let us put
Z
/
=x E
1
: (x, f (x)) Z.
Further, introduce the class of sets
S
/
1
=Z
/
: Z dom(
1
2
)
and dene the functional
/
1
(Z
/
) = (
1
2
)(Z) (Z dom(
1
2
)).
Then S
/
1
is a -algebra of subsets of E
1
and the functional
/
1
is a measure on S
/
1
extending
1
(cf. Chapter 2). In addition to this circumstance, the original mapping f turns out to be
measurable with respect to the -algebras S
/
1
and S
2
, i.e., we have
(Y S
2
)( f
1
(Y) S
/
1
).
The remarks just made are rather simple and, in fact, are familiar to the reader from Chap-
ters 2 and 5. Actually, these remarks are a starting point for our further constructions.
Lemma 1. Under the previous notation, if a measure
1
is nonzero and a measure
2
is
nonseparable, then the measure
/
1
is nonseparable, too.
Proof. Take any set X dom(
1
) with 0 <
1
(X) <+. The nonseparability of
2
implies
that for some > 0, there exists an uncountable family Y
i
: i I of
2
-measurable sets,
such that
2
(Y
i
Y
j
) > (i I, j I, i ,= j).
We leave the checking this fact to the reader (cf. also Exercises 1 and 2 from Chapter 17).
Further, we may write
/
1
((X Y
i
)
/
(X Y
j
)
/
) = (
1
2
)(X (Y
i
Y
j
)) =
1
(X)
2
(Y
i
Y
j
) >
1
(X) (i I, j I, i ,= j),
/
1
is nonseparable which ends the proof of Lemma 1.
Let J be a xed set with card(J) = c.
For any index j J, denote by
j
the restriction of the Lebesgue measure to the Borel
-algebra of the closed unit interval [0, 1]. Thus,
j
is a Borel diffused probability measure
on [0, 1].
Let stand for the product of the family of measures
j
: j J, i.e.,
=
jJ
j
.
The following two auxiliary propositions are straightforward but, for the sake of complete-
ness, we provide the short proofs here.
Lemma 2. The cardinality of dom() is equal to c and the topological weight of the metric
space canonically associated with is also equal to c. In particular, is a nonseparable
measure.
Proof. First, let us remark that any set Z dom() can be represented in the form B
[0, 1]
JJ
0
, where J
0
is a countable subset of J (certainly, depending on Z) and B is a Borel
subset of [0, 1]
J
0
.
Taking into account the equality c
= c, this remark readily implies that

card(dom()) = c.
Further, for each index j J, let us denote
Z
j
= [0, 1/2]
j
[0, 1]
J j
.
Then we have
(Z
j
Z
k
) = 1/2 ( j J, k J, j ,= k),
thus it follows that the topological weight of the metric space canonically associated with
is equal to c which ends the proof of Lemma 2.
Lemma 3. Let a be an innite cardinal number satisfying the relation
a
= a,
and let (E
1
, S
1
) and (E
2
, S
2
) be two measurable spaces such that
card(S
1
) a, card(S
2
) a.
Then the cardinality of the product -algebra S
1
S
2
does not exceed a either.
Proof. The product -algebra S
1
S
2
is generated by the family T of all sets Z having
the form Z = X Y, where X S
1
and Y S
2
. The cardinality of this family does not
exceed a a = a. Consequently,
card(S
1
S
2
) = card((T )) a
.
Taking into account the equality a
= a, we come to the required result.

The next lemma plays a key role in our further consideration.
Lemma 4. Assume the Continuum Hypothesis. Let J be a set whose cardinality is equal to
c and let
=
j
: j J,
where
j
( j J) is the restriction of to the Borel -algebra of [0, 1].
There exists a mapping
f : R [0, 1]
J
(1) the graph Gr( f ) of f is thick with respect to the product measure ;
(2) for any ( )-measure zero set Z, the set
Z
/
=x : (x, f (x)) Z
is of -measure zero.
Proof. The required mapping f will be constructed by transnite recursion.
In what follows the symbol
0
stands for the restriction of the Lebesgue measure on R to
the Borel -algebra of R.
Let denote the least ordinal number of cardinality c. According to our assumption,
=
1
and hence card() for each ordinal <.
Let _ denote a well-ordering of R which is isomorphic to .
Applying Lemmas 2 and 3, we easily derive the equality
card(dom(
0
)) = c.
Let Z
: < be the family of all those (

0
)-measurable sets whose measure is
strictly positive. Without loss of generality, we may suppose that the range of this family
coincides with the range of the family
Z
: <, is odd.
Let T
: < be an enumeration of all those (

0
)-measurable sets whose measure
is equal to zero.
We are going to construct an -sequence (x
, y
) : < of points of the product space

R[0, 1]
J
.
Suppose that for an ordinal < , the partial -sequence (x
, y
) : < has already

been constructed. Consider two possible cases.
(a) The ordinal is even.
In this case, denote by x the _-least element of the set Rx
: < and for each ordinal

<, dene
T
(x) =y : (x, y) T
.
Also, introduce a set by the formula
= : < & (T
(x)) = 0.
Obviously, we have the equality
(T
(x) : ) = 0.
Consequently,
[0, 1]
J
T
(x) : ,= / 0.
Choose a point
y [0, 1]
J
T
(x) :
and put (x
, y
) = (x, y).
(b) The ordinal is odd.
In this case, for each ordinal <, dene
T
0
=x R : (T
(x)) > 0.
Since (
0
)(T
) = 0, we have (T
0
) = 0. Consider the set Z
. Taking into account the

circumstance that (
0
)(Z
) > 0 and applying the Fubini theorem, we can nd a point

x R T
0
: <
satisfying the relation (Z
(x)) > 0. Choose a point

y Z
(x) T
(x) : <
and put (x, y) = (x
, y
).
Thus, in both cases (a) and (b) we have determined the point (x
, y
) in the product space

R[0, 1]
J
.
Proceeding in this manner, we are able to construct the desired -sequence (x
, y
) : <
of points of R[0, 1]
J
. Now, let us put
f (x
) = y
( <).
So, a certain partial function
f : R [0, 1]
J
is dened. We assert that f is the required mapping. Indeed, the equality
R =x
: <
holds true because the well-ordering _ is isomorphic to . Therefore, the domain of f
coincides with R. The thickness of Gr( f ) is a straightforward consequence of the relation
(x
, y
) Z
( <, is odd).
Finally, let us show that the inclusion
x : (x, f (x)) T
T
0
:
is valid for each ordinal < . Take any element (x, f (x)) T
. According to our con-

struction, we have
(x, f (x)) = (x
, y
)
for some ordinal number <. Consequently,
(x
, y
) T
, y
(x
).
If , then there is nothing to prove. Suppose now that < and consider two cases.
(i) (T
(x
)) = 0. If is even, then our construction yields y
, T
(x
). If is odd, then
our construction also yields y
, T
(x
). Therefore, this case is impossible for <.

(ii) (T
(x
)) > 0. This relation immediately implies x
T
0
which yields at once the

desired result.
The proof of Lemma 4 is thus completed.
Theorem 1. Under the Continuum Hypothesis, there exists a nonseparable measure
/
on
the real line R extending and such that
I(
/
) =I().
Proof. In view of Lemma 4, there exists a mapping
f : R [0, 1]
J
such that the following two relations are satised:
(1) the graph of f is thick in the product space R[0, 1]
J
with respect to the product
measure ;
(2) for each ( )-measure zero set Z, the set
Z
/
=x : (x, f (x)) Z
is of -measure zero.
Now, for any Z dom(
0
), we put
Z
/
=x : (x, f (x)) Z,
/
(Z
/
) = (
0
)(Z).
As said earlier, the functional
/
is well dened and is a measure extending
0
. Obviously,
the completion of
/
extends the Lebesgue measure on R. We preserve the same notation
for the completion of
/
. By virtue of Lemmas 1 and 2,
/
is a nonseparable measure.
More precisely, we can assert that the topological weight of the metric space canonically
associated with
/
is equal to c. Finally, relation (2) yields at once the desired equality
I(
/
) =I(),
which ends the proof of the theorem.
Remark 1. Theorem 1 was established by assuming the Continuum Hypothesis. It must
be noticed that the existence of a nonseparable extension of without new null-sets cannot
be proved within ZFC theory (cf. Exercise 3 below).
In this connection, it should also be underlined that within ZFC there is a proper separable
extension of whose null-sets are identical with the -null-sets (see Example 2 of this
chapter).
Remark 2. As pointed out at the end of the proof of Theorem 1, the topological weight of
the metric space canonically associated with
/
is equal to c. We do not know whether it is
possible to show (at least, in some models of set theory) the existence of an extension
//
of such that the topological weight of the metric space canonically associated with
//
is
strictly greater than c and I(
//
) =I().
Remark 3. Also, we do not know within ZFC theory whether there exists a translation-
invariant proper extension of such that I() =I().
Let (E, S, ) be a probability measure space. The argument presented before Lemma 1
shows that if a mapping
g : R E
is given whose graph is ( )-thick in RE, then this mapping produces the measure
g
which is an extension of the Lebesgue measure on R.
Theorem 2. Suppose that is diffused and card(S) c. Then there exists a mapping
h : R E
(1) the graph Gr(h) of h is ( )-thick in the product space RE;
(2) some set of
h
-measure zero is thick in R with respect to ; in particular, we have
I(
h
) ,=I().
Proof. Let B be a Bernstein subset of R (see [12], [30], [40], [148], [176], [192], or
Chapters 2 and 5). We recall that according to the denition of Bernstein sets, both B and
RB are totally imperfect subsets of R and this circumstance readily implies the equalities
card(B) = card(R B) = c,
(B) =
(R B) = 0.
In particular, the sets B and RB are not measurable in the Lebesgue sense and, moreover,
both of them are nonmeasurable with respect to a wide class of nonzero -nite diffused
measures on R (in this connection, see Theorem 3 from Chapter 12). Choose a point y E
and take the constant mapping
h
0
: B y.
By using the method of transnite recursion, a mapping
h
1
: R B E y
can easily be constructed such that the graph Gr(h
1
) is ( )-thick in the product space
RE (cf. the proof of Theorem 1). Let h stand for the common extension of h
0
and h
1
.
Clearly, the graph of h is also ( )-thick in RE. Consider the set Ry. Since the
measure is -nite and (y) = 0, we get
( )(Ry) = 0.
Consequently, we must have
h
(x : (x, h(x)) Ry) = 0.
Therefore,
x : (x, h(x)) Ry I(
h
).
But it is easy to check the validity of the relation
B =x : (x, h(x)) Ry
from which it follows that the set x : (x, h(x)) Ry is thick with respect to . More-
over, the above-mentioned set is a Bernstein subset of R and hence is thick with respect to
any -nite continuous Borel measure on R, which completes the proof of Theorem 2.
In exercises below we give some additional information about nonseparable extensions of
-nite measures.
This topic will be continued in subsequent sections but from another point of view (see
especially the next two chapters).
EXERCISES
1. Let E be a metrizable topological space and let be a -nite Borel measure on E such
that (U) > 0 for every nonempty open set U E. Show that the following two assertions
are equivalent:
(a) E is a separable space;
(b) is a separable measure.
2
. Investigate the question whether the assertion of Theorem 1 remains valid under Mar-
tins Axiom instead of the Continuum Hypothesis.
3
. Let (E, S, ) be a -nite complete measure space. We introduce the following nota-
tion (cf. [40]):
add(I()) = the smallest cardinal number of a subfamily of I() whose union does not
belong to I();
cov(I()) = the smallest cardinal number of a covering of E with members of I().
Let T (= S
1
) as usual denote the unit torus. Consider the compact topological group T
1
equipped with the completion of its Haar probability measure. Prove that
add(I()) =
1
.
Deduce from this equality that the conjunction of Martins Axiom and the negation of the
Continuum Hypothesis implies the relation
add(I()) < add(I()) = c.
It can be shown that the inequality
cov(I()) < cov(I())
also does not contradict ZFC theory. The latter circumstance enables to demonstrate that
the assertion of Theorem 1 of this chapter cannot be established within ZFC.
4
. Verify that the method of extending -nite invariant measures to nonseparable ones,
due to Kakutani and Oxtoby (see [83], [95], or Exercise 18 for Chapter 3), necessarily
produces new null-sets.
5. Suppose that E is a hereditarily Lindel of topological space, i.e., for any open set V E,
if V is represented in the form
V =V
i
: i I,
where all V
i
(i I) are also open in E, then there exists a countable set I
/
I such that
V =V
i
: i I
/
.
Further, suppose that is a nite Borel measure on E which is outer regular, i.e., for any
Borel set X E, we have
(X) = inf(V) : X V, V is open in E.
Show that the topological weight of the metric space canonically associated with does
not exceed w(E) +, where w(E) as usual denotes the topological weight of E.
For this purpose, x a base V
j
: j J of open sets in the given space E such that card(J) =
w(E) and consider the family L of the unions of all nite subfamilies of V
j
: j J. Verify
that the canonical image of L is everywhere dense in the metric space associated with .
6. Let E be a compact topological space with w(E) and let be a nite Radon
measure on the Borel -algebra B(E). Prove that the topological weight of the metric
space canonically associated with does not exceed w(E).
7. Let E be a topological space and let DE be an everywhere dense subset of E. Denote
by C
b
(E, R) the Banach space (with respect to the norm of uniform convergence) of all
real-valued bounded continuous functions on E. Show that the inequality
card(C
b
(E, R)) 2
card(D)+
holds true. Infer from this fact that
card(C
b
(E, R)) 2
w(E)+
,
where w(E) as earlier stands for the topological weight of E.
8. Let E be a completely regular topological space and let C
b
(E, R) be as in the previous
exercise. Prove that
w(E) w(C
b
(E, R)).
Argue in the following manner. Take any family f
j
: j J C
b
(E, R) which is every-
where dense in C
b
(E, R) and for which card(J) = w(C
b
(E, R)). Check that this family
separates the points and closed subsets of E. Deduce from this circumstance that E can be
topologically embedded in the Tychonoff cube [0, 1]
card(J)
and conclude that
w(E) card(J) = w(C
b
(E, R)).
Infer from the said above that if E is a compact space, then the following two assertions are
equivalent:
(a) the Banach space C(E, R) of all real-valued continuous functions on E is separable;
(b) E has a countable base (equivalently, E is metrizable).
9
. Let E be a nonempty topological space equipped with a nonzero nite Borel measure
such that (U) > 0 for every nonempty open set U E. Suppose, in addition, that
card(C
b
(E, R)) > c.
Prove that there exists a strictly positive integer n and an uncountable family f
j
: j J
C
b
(E, R) such that
E
( f
j
(x) f
i
(x))
2
d(x) 1/n
for any two distinct indices j J and i J. Deduce from this fact that the Hilbert space
L
2
() of all real-valued square integrable -measurable functions is nonseparable and,
consequently, the measure is nonseparable, too.
For this purpose, use the Erd os-Rado combinatorial theorem of Ramsey type (see [61] or
[144]).
10. Let E be a completely regular topological space and let be a nonzero nite Borel
measure on E such that (U) > 0 for every nonempty open set U E. Show that if
w(E) > c, then is a nonseparable measure.
For this purpose, apply the result of the previous exercise.
In particular, if (E, ) is a compact topological group satisfying the relation w(E) > c, then
the Haar probability measure on E is nonseparable.
11
. Let (E, S, ) be a probability measure space such that:

(a) is universal, i.e., dom() =P(E);
(b) is a non-atomic measure.
Prove that cannot be a perfect measure.
In order to establish this fact, utilize Exercise 17 for Appendix 1.
In particular, conclude from the said above that no universal extension of the Lebesgue
measure can be perfect.
In connection with this result, a deep statement due to Gitik and Shelah should be men-
tioned, according to which the so-called Maharam type of any universal extension of is
necessarily strictly greater than c (see [73]; cf. also [96]).
Chapter 16
Metrical transitivity and nonseparable
extensions of invariant measures
In this chapter we continue our consideration of nonseparable extensions of -nite mea-
sures. However, here we will be dealing with those nonseparable extensions of -nite
measures which are invariant with respect to a xed group of transformations of a base set
E, assuming that the cardinality of E is equal to the cardinality of the continuum.
As is widely known, metrical transitivity (i.e., ergodicity) plays an extremely important
role in various questions of the theory of invariant and quasi-invariant measures. Here
we would like to conrm this circumstance by demonstrating that under the Continuum
Hypothesis, any nonzero -nite metrically transitive left (right) invariant measure on a
group of cardinality continuum admits a nonseparable left (right) invariant extension. An
application of this result to the left (right) Haar measure on a -compact locally compact
topological group of the same cardinality is also presented.
Various left (right) invariant extensions of the left (right) Haar measure given on a Polish
locally compact group are known in the literature. For instance, some nonseparable in-
variant extension of the Haar probability measure on an uncountable compact metrizable
commutative group is thoroughly considered in the monograph by Hewitt and Ross [83].
The construction of this extension imitates the remarkable construction of Kakutani and
Oxtoby [95] which was done in a particular case, namely, for the Lebesgue measure on
the one-dimensional unit torus T = S
1
.
By using slight modications, the same construction works for the Lebesgue measure
n
on the n-dimensional Euclidean space R
n
, where n 1, and, more generally, for the left
Haar measure on an uncountable locally compact Polish topological group (cf. [108]).
It seems at rst sight that specic topological properties of the Haar (respectively,
Lebesgue) measure play an important role in the above-mentioned constructions. How-
ever, we will demonstrate in the sequel that the situation is different. The main goal of
this chapter is to show that topological concepts are inessential for constructing nonsepa-
257
6,
rable left (right) invariant extensions of a nonzero -nite left (right) invariant metrically
transitive measure given on a group of cardinality continuum. In fact, we will show that a
dominant role in constructions of such a kind is played by the metrical transitivity of this
measure.
The notion of metrical transitivity for measures invariant or, more generally, quasi-invariant
under transformation groups was already introduced in preceding sections of this book.
See, e.g., Chapter 3 or Chapter 9 where the related notion of weak metrical transitivity is
also discussed.
All measures considered in the present chapter are assumed to be diffused (continuous),
i.e., all of them vanish at the singletons.
This assumption is necessary for our further considerations (cf. Example 2 in the end of
this chapter).
Let E be a set, G be a group of transformations of E and let be a nonzero -nite
G-invariant (G-quasi-invariant) measure on E. Recall that is metrically transitive (i.e.,
ergodic) if for every -measurable set X E with (X) >0, there exists a countable family
{g
i
: i I} G such that
(E \ {g
i
(X) : i I}) = 0.
It is well known that the metrical transitivity of an invariant measure is closely connected
with its uniqueness property. Indeed, if has the uniqueness property, then is necessarily
metrically transitive. Conversely, if G is uncountable and acts freely in E and is complete,
then the metrical transitivity of turns out to be equivalent to the uniqueness property (for
more details, see [105], [108], [115], and Exercise 1 of this chapter).
Example 1. The left Haar measure on a -compact locally compact group (G, ) is
metrically transitive. More precisely, let H be an arbitrary everywhere dense subgroup of
G. Then the same considered as a left H-invariant measure is also metrically transitive
(see, for instance, [83] and [115]).
Recall that a set Y E is almost G-invariant with respect to if the equality
(g(Y)Y) = 0
holds true for all transformations g G.
Clearly, Y E is almost G-invariant with respect to if and only if E \Y is almost G-
invariant with respect to .
Metrical transitivity and nonseparable extensions of invariant measures 259
Moreover, the union of a countable family of almost G-invariant sets with respect to is
almost G-invariant with respect to , too.
Therefore, the class of all almost G-invariant sets with respect to forms a G-invariant
-algebra of subsets of E.
Recall also that a set Z E is -thick in E if
(E \ Z) = 0.
We need two auxiliary propositions.
Lemma 1. Let E be a set, G be a group of transformations of E and let be a nonzero
-nite complete G-quasi-invariant metrically transitive measure on E. Suppose that a set
Y E is almost G-invariant with respect to . Then at least one of the following three
relations is valid:
1) (Y) = 0;
2) (E \Y) = 0;
3) both sets Y and E \Y are thick with respect to and, consequently, both of them are
nonmeasurable with respect to .
Proof. Suppose that 3) does not hold, i.e., at least one of the sets Y and E \Y is not thick
with respect to . Without loss of generality, we may assume that Y is not thick with
respect to . Then there exists a -measurable set T of strictly positive -measure, such
that
T Y = / 0.
Since is metrically transitive, we can nd a countable family {g
i
: i I} G satisfying
the relation
(E \ {g
i
(T) : i I}) = 0.
Since Y is almost G-invariant with respect to , we may write
(Y({g
i
(Y) : i I})) = 0,
(Y ({g
i
(T) : i I})) = 0
and, consequently, (Y) = 0 which ends the proof of Lemma 1.
As usual, we denote by the rst innite cardinal, by
1
the rst uncountable cardinal,
and by c the cardinality of the continuum. Moreover, (respectively,
1
and c) can be
identied with the least ordinal number whose cardinality is equal to (respectively, to
1
and c).
For our further purposes, we need the notion of Ulams matrix which is very important
in combinatorial set theory, measure theory and set-theoretic topology (see [144], [148],
[192], and [238]).
Let U be a set of cardinality
1
. Recall that a double family
{U
n,
: n <, <
1
}
of subsets of U is an Ulam matrix over U if the following two conditions are satised:
(*) for any n <, the family {U
n,
: <
1
} is disjoint;
(**) for any <
1
, the set U \ {U
n,
: n <} is at most countable.
The existence of Ulam matrices was rst proved by Ulam in his classical work [238] (see
also [144], [192], and Exercise 18 for Appendix 1). Ulams result directly implies that
1
is not a real-valued measurable cardinal. Some applications of Ulam matrices to the theory
of invariant and quasi-invariant measures are presented in [108], [115], and [119].
The next lemma is crucial for our further considerations.
Lemma 2. Let E be a set with card(E) =
1
and let G be a group of transformations of E
such that card(G)
1
. Suppose that is a nonzero -nite G-quasi-invariant metrically
transitive measure on E. Then there exists a disjoint family {Y
: <
1
} of subsets of E
such that:
1) each set Y
( <
1
) is thick with respect to ;
2) for any
1
, the set {Y
: } is almost G-invariant with respect to .

Proof. We may assume, without loss of generality, that the given measure is complete.
Consider the partition {O
i
: i I} of E into G-orbits. Only two cases are possible.
1. All G-orbits are of -measure zero.
In this case, we necessarily have the equality card(I) =
1
because our measure is
nonzero. Let
{I
n,
: n <, <
1
}
denote an Ulam matrix over I. We dene the sets
Z
n,
={O
i
: i I
n,
} (n <, <
1
).
Taking into account the conditions (*) and (**), we easily get the following relations:
(a) for each n <, the family {Z
n,
: <
1
} is disjoint;
(b) for each <
1
, the set E \ {Z
n,
: n <} is of -measure zero.
It follows from(a) and (b) that there exist a natural number n and an uncountable set
1
such that the inequality
(Z
n,
) > 0 is satised for all ordinals . Clearly, we also
have
Z
n,
Z
n,
= / 0 ( , , =).
Further, since our measure is metrically transitive and each set Z
n,
( ) is G-
invariant, we can conclude (in view of Lemma 1) that all sets Z
n,
( ) are thick with
respect to . Let
:
1

be any bijection acting from
1
onto . Putting
{Y
: <
1
} ={Z
n,()
: <
1
},
we come to the required family of subsets of E.
2. There exists at least one index i
0
I for which we have
(O
i
0
) > 0.
In this case, we obviously obtain
card(O
i
0
) =
1
, card(G) =
1
.
Also, by Lemma 1, the orbit O
i
0
is a thick set with respect to our measure . Let us denote
V = O
i
0
and let us x a point x V. Clearly, the group G acts transitively on V, i.e.,
V = G(x) ={g(x) : g G}.
Let us represent the group G in the form
G ={G
: <
1
},
where {G
: <
1
} is an increasing (by inclusion) family of subgroups of G satisfying
the following relations:
(c) G
={G
: <} for all <

1
;
(d) card(G
) for all <

1
.
It is not difcult to show the existence of such a representation of G (cf. Exercise 19 from
Chapter 3 or Exercise 9 from Chapter 11). Evidently, we may write
V = G(x) ={G
(x) : <
1
}.
Now, putting
V
= G
(x) \ {G
(x) : <}
for any <
1
, we see that {V
: <
1
} is a disjoint covering of V by countable sets. In
addition, for every subset of
1
, the set {V
: } is almost G-invariant with respect

to the given measure .
Further, consider an Ulam matrix
{U
n,
: n <, <
1
}
over the set U =
1
and dene
Z
n,
={V
: U
n,
} (n <, <
1
).
Again, taking into account the conditions (*) and (**), we derive that:
(e) for each n <, the family of sets {Z
n,
: <
1
} is disjoint;
(f) for each <
1
, the set V \{Z
n,
: n <} is at most countable and, consequently, the
set {Z
n,
: n <} is thick with respect to .
It follows from the relations (e) and (f) that there exist a natural number n and an un-
countable set
1
such that the inequality
(Z
n,
) > 0 is valid for all ordinals .
Using once more the metrical transitivity of and applying Lemma 1, we infer that all sets
Z
n,
( ) are thick with respect to . Let
:
1

be any bijection acting from
1
onto . Putting again
{Y
: <
1
} ={Z
n,()
: <
1
},
we obtain the required family of subsets of E. Lemma 2 has thus been proved.
The next auxiliary proposition is a direct consequence of the previous lemma.
Lemma 3. Let denote the least ordinal number of cardinality c and let E be an arbitrary
set with card(E) = c. Let G be a group of transformations of E such that card(G) c.
Suppose that is a nonzero -nite G-invariant metrically transitive measure on E. Then
assuming the Continuum Hypothesis, there exists a disjoint family {Y
: < } of subsets
of E such that:
1) each set Y
( <) is thick with respect to ;

2) for any , the set {Y
: } is almost G-invariant with respect to .

Now, we are able to establish the following statement.
Theorem 1. Under the assumptions of Lemma 3, there exists a nonseparable G-invariant
extension
of . More precisely,
can be chosen so that the Hilbert dimension of the

space L
2
(
) is equal to 2
c
.
Proof. We preserve the notation of Lemma 3. Take a family {Y
: <} as in that lemma.

According to a result of Tarski (see [83], [108], [150], or Exercise 3 for this chapter), there
exists a family {
j
: j J} of subsets of such that:
(1) card(J) = 2
c
;
(2) for any injective sequence { j
k
: k <} J, we have
card({
j
k
: k <}) = c,
where
j
k
=
j
k
or
j
k
= c \
j
k
for all indices j
k
(k <).
Now, for each index j J, let us dene
A
j
={Y
:
j
}.
In this way we come to the family {A
j
: j J} of almost G-invariant subsets of E which
are mutually -independent (in the generalized sense) with respect to . The last phrase
means that for any injective sequence { j
k
: k <} J, the equality
(E \ {A
j
k
: k <}) = 0
holds true, where A
j
k
= A
j
k
or A
j
k
= E \ A
j
k
. Now, applying the Kakutani-Oxtoby method
of extending invariant measures (see [83], [95], or Exercise 18 from Chapter 3), we can
extend our to a G-invariant measure
so that all sets from the family {A

j
: j J}
become
-measurable. Moreover, the above-mentionedmethod guarantees that the Hilbert

dimension of the space L
2
(
) is equal to 2
c
(see again [83] and [95]) which completes the
proof of Theorem 1.
Unfortunately, the obtained extension
is not metrically transitive because all sets A

j
( j
J) are almost G-invariant with respect to
. It would be interesting to nd a method

of extending measures by means of which a given invariant (respectively, quasi-invariant)
metrically transitive measure can be extended to a nonseparable invariant (respectively,
quasi-invariant) metrically transitive measure. More precisely, we would like to formulate
the following question.
Problem. Let E be a set with card(E) = c and let G be a group of transformations of E
such that card(G) c. Suppose that is a nonzero -nite G-invariant metrically transitive
measure on E. Does there exist a nonseparable G-invariant metrically transitive measure
on E extending ?
We do not know an answer to this question.
In this context, notice that the method described in the paper by Kodaira and Kakutani
[141] preserves the property of metrical transitivity of the Lebesgue measure, but we do
not know any purely measure-theoretical version of their method.
Theorem 1 directly implies the following statement.
Theorem 2. Assume the Continuum Hypothesis. Let G be a -compact locally compact
group of cardinality c and let be a left invariant metrically transitive extension of the
left Haar measure on G. Then there exists a left invariant extension
of such that the

Hilbert dimension of the space L
2
(
) is equal to 2
c
.
Remark 1. It should be mentioned that a measure in Theorem 2 can be nonseparable
itself (in this connection, see [115]). This circumstance also indicates that Theorem 2
cannot be established only by using the methods developed in [83] and [95].
Remark 2. As pointed out in the beginning of this chapter, the construction of Kakutani and
Oxtoby [95] as well as its generalization presented in the monograph [83] by Hewitt and
Ross is based on special topological properties of the Lebesgue measure (respectively, Haar
measure). On the other hand, Theorem1 shows that replacing those properties by the metri-
cal transitivity, we can get a much stronger assertion. However, the Continuum Hypothesis
was essentially used in our argument and, at this moment, we do not know whether addi-
tional set-theoretical assumptions are necessary for obtaining the above-mentionedstronger
result.
Example 2. Let E be a set of cardinality c and let E
be an innite countable subset of E.

Dene G as the group of all those transformations of E whose restrictions to the set E \ E
coincide with the identical transformation of E \ E
. Obviously, we have
card(G) = card(E) = c.
Further, dene a measure on the power set P(E) of E by putting:
(Z) = card(Z E
) if card(Z E
) <;
(Z) = + if card(Z E
) .
It can easily be veried that is a -nite, G-invariant, metrically transitive and separable
measure. At the same time, admits no proper extensions because the domain of is
the whole power set P(E). In particular, does not admit a nonseparable G-invariant
extension.
This simple example shows that the assumption of the diffusedness of a measure in the
formulation of Theorem 1 is rather essential for the validity of this theorem.
Example 3. Let E and G be as in Example 2. In a similar way we can dene a probability
measure
1
on the power set P(E), which is G-quasi-invariant, metrically transitive, and
separable. Analogously to the measure from the previous example, this
1
has no proper
extensions and, consequently, does not admit a nonseparable G-quasi-invariant extension.
EXERCISES
1
. Let E be a set, G be a group of transformations of E and let be a -nite G-invariant

complete measure on E. Suppose that G acts almost freely in E with respect to , i.e., for
any two distinct transformations g G and h G, we have
({x E : g(x) = h(x)}) = 0.

(a) has the uniqueness property, i.e., for every -nite G-invariant measure with
dom() = dom(), there exists t [0, +[ such that =t ;
(b) is metrically transitive (ergodic).
For this purpose, use the fact that any -measurable set X E with
(X) > 0 contains a

-nonmeasurable subset (see Exercises 9, 10 and 11 of Appendix 1).
Also, show that (a) always implies (b) (without the assumption of the completeness of ).
2. Deduce Lemma 3 from Lemma 2.
3
. Prove that for any innite set T such that (card(T))
= card(T), there exists a family

{T
j
: j J} of subsets of T satisfying the following conditions:
(a) card(J) = 2
card(T)
;
(b) {T
j
: j J} is -independent in the set-theoretical sense, i.e., for every injective se-
quence { j
n
: n N} J and for every function f : N {0, 1}, we have
card(T
f (0)
j
0
T
f (1)
j
1
... T
f (n)
j
n
...) = card(T),
where T
f (n)
j
n
= T
j
n
if f (n) = 0 and T
f (n)
j
n
= T \ T
j
n
if f (n) = 1.
4. Let (E, G) be a space equipped with a transformation group and let be a -nite
G-invariant (G-quasi-invariant) measure on E.
Let Y be a -measurable almost G-invariant subset of E with (Y) >0. Then Y determines
a nonzero -nite G-invariant (G-quasi-invariant) measure
Y
by the formula
Y
(X) = (X Y) (X dom()).
If
Y
is ergodic (i.e., metrically transitive), then we say that
Y
is an ergodic component of
.
Obviously, is ergodic if and only if
E
is an ergodic component of .
Show that the nonseparable translation-invariant extension of obtained by the method of
Kakutani and Oxtoby (see [83], [95], [108], or Exercise 18 from Chapter 3) has no ergodic
components.
5. Let (E, G) be a space equipped with a transformation group and let be a -nite
G-invariant (G-quasi-invariant) measure on E.
We say that a family {X
i
: i I} of nonempty subsets of E is admissible for if the follow-
ing relations are satised:
(a) all sets from {X
i
: i I} are pairwise disjoint;
(b)
({X
i
: i I}) > 0;
(c) for each index i I, we have (X
i
) = 0;
(d) for any subset J of I, the set {X
i
: i J} is almost G-invariant with respect to .
Supposing that there exists at least one admissible family for , it is convenient to introduce
the notation:
a() = inf{card(I) : there exists an admissible f amily {X
i
: i I} f or }.
Verify that a() is an uncountable regular cardinal and a() card(E).
6
. Prove that there exists a translation-invariant extension of the Lebesgue measure

on R such that a() =
1
.
More generally, suppose that E is a vector space over the eld Q of all rational numbers
with card(E)
1
and is a nonzero -nite E-invariant (E-quasi-invariant) measure on
E. Show that there exists an E-invariant (E-quasi-invariant) extension
of such that
a(
) =
1
.
7. Suppose that (E, G) is a space equipped with a transformation group and satisfying the
following two relations:
(a) card(E) = card(G);
(b) G acts transitively in E.
Let be a nonzero -nite G-invariant (G-quasi-invariant) measure on E such that
(Z) =
0 for all sets Z E with card(Z) < card(E).
Check that if c f (card(E)) >, then can be extended to a G-invariant (G-quasi-invariant)
measure on E for which there exists at least one admissible family of subsets of E.
8
. Let E be a set, G be a group of transformations of E and let be a nonzero -nite

G-invariant measure on E. In addition, suppose that an uncountable disjoint family {X
t
:
t T} of -thick subsets of E is given such that for each T
0
T, the set {X
t
: t T
0
} is
almost G-invariant with respect to .
Prove that there exists a G-invariant extension
of such that the Hilbert dimension of

L
2
(
) is strictly greater than card(T).

For this purpose, apply the results of the preceding exercises.
9
. Assume that the Generalized ContinuumHypothesis holds. Let E be a set, G be a group

of transformations of E, and let be a nonzero -nite G-invariant measure on E. Suppose
that these ve conditions are satised:
(a) c f (card(E)) >;
(b) card(G) = card(E) and G acts transitively in E;
(c) has at least one ergodic component;
(d) a() is strictly less than the rst uncountable inaccessible cardinal number;
(e)
(Z) = 0 for all sets Z E with card(Z) < card(E).

Prove that there exists a G-invariant extension

the space L
2
(
) is greater than or equal to 2

a()
.
10. Assume the Generalized Continuum Hypothesis. Let E be a set, G be a group of
transformations of E and let be a nonzero -nite G-invariant measure on E. Suppose
also that:
(a) the cardinality of E is strictly less than the rst uncountable inaccessible cardinal num-
ber and c f (card(E)) >;
(b) card(G) = card(E) and G acts transitively in E;
(c) has at least one ergodic component;
(d)
(Z) = 0 for all sets Z E with card(Z) < card(E).

Show that there exists a G-invariant extension

L
2
(
) is greater than or equal to 2

c
and, in particular,
is a nonseparable extension of .
11. Prove that there exists a nonseparable translation-invariant extension of the Lebesgue
measure on R, which does not have the Steinhaus property (for the denition of this
property, see Chapters 3 and 4).
12. Let E be a set equipped with a nonseparable -nite measure . Show that there exists
a -measurable set Y E with (Y) > 0 such that the restriction of to the -algebra
S
Y
={Y X : X dom()}
of subsets of Y is a non-atomic measure.
In order to demonstrate this assertion, take into account the fact that any maximal disjoint
family of atoms of a -nite measure is at most countable.
13
. Let T as usual denote the one-dimensional unit torus. Consider the product space
G = T
c
as a compact commutative group equipped with its Haar probability measure .
Prove that there exists a subgroup H of G satisfying the following relations:
(a) card(H) = c < 2
c
= card(G);
(b) H is -thick in G.
Infer from (a) and (b) that H is nonmeasurable with respect to the completion of .
In connection with the last exercise, it should be mentioned that the existence of a subset
of R which is nonmeasurable with respect to and whose cardinality is strictly less than
card(R) = c cannot be established within ZFC theory. Indeed, as well known, if Martins
Axiom holds true, then any subset X of R with card(X) < c is of -measure zero and,
consequently, -measurable.
Chapter 17
Nonseparable left invariant measures on
uncountable solvable groups
In the previous chapter we were concerned with certain nonseparable extensions of invari-
ant measures given on a space E which is equipped with a transformation group G such
that
card(E) = card(G) = c,
where c denotes the cardinality of the continuum.
In connection with the result presented in that chapter, the following question arises.
What can be said in the situation when card(E) = card(G) > c?
In particular, a priori it is not clear whether there are -nite invariant (quasi-invariant)
nonseparable measures in the above-mentioned more general case. However, under some
purely set-theoretical assumptions and rather natural assumptions on a group G, the exis-
tence of such measures can be proved. Notice that among those assumptions, the General-
ized Continuum Hypothesis was heavily exploited (see Exercises 9 and 10 for the previous
chapter).
Assuming the Continuum Hypothesis, we are going to establish that every uncountable
solvable group admits a nonseparable non-atomic -nite left invariant measure. Several
related questions concerning properties of such measures will be discussed, too.
Let E be a nonempty set, S be a -algebra of subsets of E and let be a measure on E
with dom() =S. As usual, the triple (E, S, ) is called a measure space.
For any two sets X S and Y S having nite -measure, we may put
d(X,Y) = (XY).
As is well known, the function d is a quasi-metric and after the corresponding factorization
it yields a complete metric space canonically associated with (cf. Exercise 12 of Chapter
1).
269
7,
We recall that a measure is nonseparable if the above-mentioned metric space is nonsep-
arable.
It is also known that the following two assertions are equivalent:
(1) is nonseparable;
(2) there exist a real > 0 and an uncountable family {Y
i
: i I} of -measurable subsets
of E such that
d(Y
i
,Y
j
) > (i I, j I, i = j).
Nonseparable probability measures naturally appear when we deal with uncountable prod-
ucts of probability measure spaces. For example, if T denotes the one-dimensional unit
torus endowed with the standard Lebesgue probability measure, then the product group of
uncountably many copies of T carries the product measure which is nonseparable and, in
fact, coincides with the Haar probability measure on this product group.
Uncountable direct sums of probability measure spaces also yield examples of nonsepara-
ble measures. However, in the latter case the obtained measures fail to be -nite, and this
circumstance is a weak side of the operation of taking direct sums of probability measure
spaces.
Many years ago, a problemwas formulated about the existence of nonseparable translation-
invariant extensions of the standard Lebesgue measure on the real line R and about
the existence of nonseparable translation-invariant extensions of the Lebesgue probability
measure on the torus T (see [234] and [235]).
This problem was intensively investigated by several authors and was solved positively in
the classical works [95] and [141] by Kakutani, Kodaira and Oxtoby.
Moreover, some generalized versions of the above-mentioned problemwere considered for
a Haar measure on a -compact locally compact topological group (see [83], [129], and
the previous chapter).
For further information on methods of extending invariant and, more generally, quasi-
invariant measures which are given on various spaces equipped with their transformation
groups, see [41], [42], [45], [46], [83], [85], [86], [95], [105], [108], [115], [119], [124],
[129], [133], [141], [142], [193], [195], [198], [234], [235], [242], [247], [249], and Chap-
ters 2 and 3 of this book.
Let (G, ) be an arbitrary uncountable group. In our consideration below we do not assume
that G is equipped with a topology compatible with its algebraic structure.
There are trivial examples of probability left-invariant measures dened on some -
algebras of subsets of G. For instance, we may introduce the probability measure
0
dened
Nonseparable left invariant measures on uncountable solvable groups 271
on the family of all countable and co-countable subsets of G and vanishing at all singletons
in G. Obviously, the measure
0
is invariant under the group of all left (right) translations
of G. In this context, the following natural question arises.
Does there exist a nonseparable non-atomic -nite left-invariant measure on G?
Using the Continuum Hypothesis, our goal is to show that if G is solvable, then the answer
to this question is positive. Moreover, it turns out that if G is uncountable and commutative,
then there always exists a nonseparable non-atomic invariant probability measure on G.
In order to demonstrate this fact, we need several auxiliary propositions.
The rst lemma is purely algebraic in nature and is well known.
Lemma 1. Let (G, +) be an uncountable commutative group. There exists a commutative
group (H, +) such that:
1) card(H) =
1
;
2) H is a homomorphic image of G.
A detailed proof of this statement was given in Chapter 11 (see Lemma 8 therein).
Remark 1. One application of the above lemma has already been presented in the same
Chapter 11 where algebraic sums of so-called absolutely negligible sets in uncountable
commutative groups were discussed. Also, as demonstrated in preceding sections, these
sets play an essential role in some questions concerning extensions of nonzero -nite left
invariant (left quasi-invariant) measures on groups.
Recall that if a commutative group (G, +) is given and is a left invariant measure on G,
then automatically is a right invariant measure on G. In this case, we simply say that is
an invariant (or G-invariant) measure on G.
Starting with Lemma 1, it is not difcult to obtain the following statement.
Lemma 2. Suppose that every commutative group of cardinality
1
admits a nonseparable
non-atomic -nite invariant measure. Then every uncountable commutative group admits
a nonseparable non-atomic -nite invariant measure.
Proof. Let (G, +) be an arbitrary uncountable commutative group. In view of Lemma 1,
there exists a surjective homomorphism
: G H,
where H is some commutative group of cardinality
1
. According to our assumption, H
admits a nonseparable non-atomic -nite invariant measure . Consider the family of
sets
S ={
1
(Y) : Y dom()}
and dene a functional on S by putting
(
1
(Y)) = (Y) (Y dom()).
It can easily be veried that:
(a) the denition of is correct because is a surjection;
(b) S is a G-invariant -algebra of subsets of G;
(c) is a non-atomic -nite measure on S.
Let us check that is also nonseparable and G-invariant.
Indeed, since the measure is nonseparable, there exists an uncountable family {Y
i
: i I}
of -measurable subsets of H such that
d(Y
i
,Y
j
) > (i I, j I, i = j)
for some strictly positive real number , where d denotes the quasi-metric canonically
associated with . Then we may write
(
1
(Y
i
)
1
(Y
j
)) > (i I, j I, i = j),
which shows the nonseparability of .
Take now an arbitrary set X S and an arbitrary element g G. Then X =
1
(Y) for
some set Y dom(). Denoting
h =(g),
we easily come to the equality
g +X =
1
(h +Y)
which yields
(g +X) = (h +Y) = (Y) =(
1
(Y)) =(X),
i.e., turns out to be a G-invariant measure. Lemma 2 has thus been proved.
Lemma 3. Suppose again that every commutative group of cardinality
1
admits a non-
separable non-atomic -nite invariant measure. Then every uncountable solvable group
admits a nonseparable non-atomic -nite left invariant measure.
Proof. Let (G, ) be an arbitrary uncountable solvable group. There exists a composition
series
{e} = G
0
G
1
... G
n1
G
n
= G
of subgroups of G, where e denotes the neutral element of G and for each natural number
i [0, n1], the group G
i
is normal in G
i+1
and the quotient group G
i+1
/G
i
is commutative.
We must show that the group G = G
n
admits a nonseparable non-atomic -nite left in-
variant measure on some -algebra of its subsets. For this purpose, we use the induction
method on n.
If n = 1, then the quotient group G = G
1
/G
0
is commutative and we may directly apply
Lemma 2.
Suppose now that our assertion has already been proved for uncountable solvable groups
whose composition series contains at most n subgroups, and let us establish the validity of
our assertion for those uncountable solvable groups G whose composition series consists
of n +1 subgroups, i.e., G = G
n
.
Let : G
n
G
n
/G
n1
denote the canonical epimorphism.
Consider two possible cases.
1. The commutative group G
n
/G
n1
is uncountable.
In this case we apply again Lemma 2 and equip the group G
n
/G
n1
with a nonseparable
non-atomic -nite left invariant measure . Further, as in the proof of Lemma 2, we
dene
S ={
1
(Y) : Y dom()},
(
1
(Y)) = (Y) (Y dom()).
In this manner we obtain the -algebra S of subsets of G
n
= G and a nonseparable non-
atomic -nite left invariant measure on S (cf. the proof of Lemma 2). Therefore,
we conclude that our group G admits a nonseparable non-atomic -nite left invariant
measure.
2. The commutative group G
n
/G
n1
is at most countable.
In this case, taking into account the uncountability of our group G
n
, we rst derive that the
subgroup G
n1
is uncountable, too. According to the inductive assumption, G
n1
admits
some nonseparable non-atomic -nite left invariant measure .
Denote by {g
k
: k K} a selector of the countable family G
n
/G
n1
. For each index k K,
we have a canonical bijection between the sets G
n1
and g
k
G
n1
. By using this bijection,
we transfer the measure to the set g
k
G
n1
and denote the obtained measure by
k
. Let
us underline that one and only one of the measures
k
(k K) coincides with the original
measure .
Now, let us put
S
= the family of all subsets X of G

n
such that (X g
k
G
n1
) dom(
k
) for each k K.
It can easily be seen that S
is a -algebra of sets in G
n
. Further, for any set X S
, we
dene
(X) =

kK
k
(X g
k
G
n1
).
A direct verication shows that is a non-atomic -nite left invariant measure on the
group G
n
= G.
Finally, is nonseparable since it contains as a direct summand the nonseparable measure
.
We thus conclude that G admits a nonseparable non-atomic -nite left invariant measure
which completes the proof of Lemma 3.
Lemma 4. Assume the ContinuumHypothesis. Then every commutative group (H, +) with
card(H) =
1
admits a nonseparable non-atomic invariant probability measure.
Proof. It is known (see, e.g., Exercise 19 from Chapter 3) that there exists a partition
{H
: <
1
} of H satisfying the following relations:
(a) card(H
) for any ordinal <

1
;
(b) for each subset of
1
, the set
Z() ={H
: }
is almost H-invariant in H, i.e., we have
(h H)(card((h +Z())Z()) ).
Also, according to an old result on -independent families of sets due to Tarski, if E is a
set of cardinality continuum, then there exists an uncountable family {E
i
: i I} of subsets
of E such that
(c) for any injective sequence {i
n
: n <} I, we have
card({E
i
n
: n <}) = c,
where
E
i
n
= E
i
n
E
i
n
= E \ E
i
n
(n <).
Notice that the existence of {E
i
: i I} can be established directly by using the method
of transnite induction and without appealing to the above-mentioned result of Tarski (cf.
Exercise 3 for Chapter 16).
In view of our assumption c =
1
, we immediately obtain an analogous uncountable family
{
i
: i I} of subsets of
1
, i.e., the following relation holds:
(d) for any injective sequence {i
n
: n <} I, we have
card({
i
n
: n <}) = c,
where
i
n
=
i
n

i
n
=
1
\
i
n
(n <).
Now, for every index i I, let us put
A
i
={H
:
i
}
and consider the family {A
i
: i I} of subsets of H. This family is -independent, too, and
each set A
i
(i I) is almost H-invariant in H by virtue of relation (b).
Let
0
denote the probability H-invariant measure on H which is dened for all countable
and co-countable subsets of H and vanishes at all singletons in H. Applying the Kakutani-
Oxtoby method of extending invariant measures (cf. [83], [95], [108], or Exercise 18 from
Chapter 3), we can extend
0
to a non-atomic H-invariant measure on H such that
{A
i
: i I} dom(),
d(A
i
, A
j
) = 1/2 (i I, j I, i = j).
Since the set I is uncountable, the last relation yields that the measure is nonseparable.
Moreover, we may take I so that card(I) = 2
c
and, consequently, the topological weight of
the metric space canonically associated with will attain its maximum value. Lemma 4
Remark 2. Unfortunately, the role of the Continuum Hypothesis in Lemma 4 is still un-
clear. In other words, we do not know whether it is possible to prove this lemma within the
theory ZFC.
Remark 3. Lemma 4 can be easily generalized to all those innite groups (H, ) which
satisfy the equality
(card(H))
= card(H),
but are not necessarily commutative or solvable. The proof can be carried out by the same
method of -independent families of sets. However, this method does not work for those
uncountable groups (H, ) whose cardinality is conal with . Also, it is reasonable to
point out in this place that under the Generalized Continuum Hypothesis, the following
two assertions are equivalent for an innite group H:
(1) (card(H))
= card(H);
(2) card(H) is not conal with .
Finally, let us recall that if G is an uncountable -compact locally compact topological
group, then the equality
(card(G))
= card(G)
is automatically valid (see, for instance, [46], [83], [94], or Exercise 4 from Chapter 10).
Taking into account Lemmas 1 - 4, we come to the following statement.
Theorem 1. If the Continuum Hypothesis holds, then every uncountable solvable group
admits a nonseparable non-atomic -nite left invariant measure.
For commutative groups, we also have a certain analogue of this theorem within the theory
ZFC.
Theorem 2. Any commutative group (G, +) with card(G) c admits a nonseparable non-
atomic G-invariant probability measure.
The proof of Theorem 2 follows from the corresponding analogues of Lemmas 1, 2, and 4.
In connection with the results presented in this chapter, the following two open questions
seem to be of interest.
Problem. Is it true that every uncountable group (G, ) admits a nonseparable non-atomic
-nite left invariant measure? More generally, is it true that any nonzero -nite left
invariant measure given on G admits a nonseparable left invariant extension?
Problem. Is it true that every uncountable solvable group (G, ) admits a nonseparable non-
atomic -nite left invariant measure having the uniqueness property? More generally, is
it true that any nonzero -nite left invariant measure on G with the uniqueness property
admits a nonseparable left invariant extension with the same property?
Rich information about the uniqueness property for invariant measures can be found in
[47], [80], [81], [83], [105], [108], [115], [182], [198], and [248].
Notice that a nonseparable invariant extension of the Lebesgue measure constructed by
Kakutani and Oxtoby [95] does not possess this property. On the other hand, a nonseparable
invariant extension of the Lebesgue measure constructed by Kodaira and Kakutani [141]
has the uniqueness property (for more details, see [115]).
EXERCISES
1. Let be a -nite measure. Show that the following two assertions are equivalent:
(a) is nonseparable;
(b) there exist a real > 0 and an uncountable family {Y
i
: i I} of -measurable subsets
of E such that
d(Y
i
,Y
j
) > (i I, j I, i = j),
where d denotes the quasi-metric canonically associated with .
2. Let (E
1
, S
1
,
1
) and (E
2
, S
2
,
2
) be two -nite measure spaces and suppose that a
mapping
f : E
1
E
2
satises the relations
(Y S
2
)( f
1
(Y) S
1
),
(Y S
2
)(
1
( f
1
(Y)) =
2
(Y)),
which simply means that
2
is a homomorphic image of
1
under f .
Show that if
2
is nonseparable, then
1
is nonseparable, too.
Formulate and prove a more general result in terms of the topological weights of the metric
spaces canonically associated with
1
and
2
, respectively.
3. Let (G, +) and (H, +) be two commutative groups and let : G H be a surjective
homomorphism. Suppose that H is equipped with a -nite measure . Show that the
pre-image
1
() has the same weight as .
4. Generalize Lemma 4 to the case of all those uncountable groups (H, ) which satisfy the
equality
(card(H))
= card(H).
For this purpose, use Tarskis result on -independent families of sets and Exercises 18 and
19 from Chapter 3.
5. Assuming the Generalized Continuum Hypothesis, show the equivalence of assertions
(1) and (2) of Remark 3.
6. Prove that the nonseparable translation-invariant extension of obtained by the Kodaira-
Kakutani method [141] has the uniqueness property.
7. Let E be a set and let G be a group of transformations of E. Show that the following two
(a) there exists a nite (-nite) nonseparable non-atomic G-invariant measure on E;
(b) there exists a nite (-nite) nonseparable G-invariant measure on E.
8. Let (E, S, ) be a -nite measure space and let be an arbitrary -nite measure
equivalent to .
Is it true that is separable if and only if is separable?
More generally, is it true that the topological weight of the metric space canonically asso-
ciated with is equal to the topological weight of the metric space canonically associated
with ?
9
. Let (G, ) be an uncountable solvable group and let P(G) as usual denote the family of
all subsets of G. Prove that there exists a functional
: P(G) [0, 1]
(a) (G) = 1;
(b) is nitely additive, i.e., (X Y) =(X) +(Y) for any two disjoint sets X G and
Y G;
(c) is right invariant, i.e., (Xg) =(X) for each set X G and for each element g G.
The functional is called a right invariant mean for G (cf. [83] and [240]).
In order to establish the existence of , consider the family F = F(G) of all real-valued
bounded functions on G, equipped with the norm of uniform convergence, that is
|| f || = sup
gG
| f (g)| ( f F).
For any function f : G R and for any element g G, dene the function f
g
: G R by
the formula
f
g
(h) = f (h g) (h G).
Further, let = F
denote the conjugate space of F endowed with the weak topology

(F
, F). For each and for each g G, dene

g
by the relation
g
( f ) =( f
g
) ( f F).
In this manner, every g G produces the continuous afne mapping
T
g
: ,
determined by the formula
T
g
() =
g
( ).
Check that the obtained family {T
g
: g G} satises the assumptions of Markov-Kakutanis
theorem from Chapter 3 and transforms the compact convex set
K ={ : is positive, (
G
) = 1}
into itself (here
G
denotes the characteristic function of G, which is identically equal to
1).
Conclude from these facts that there exists K invariant under all mappings from {T
g
:
g G}, i.e., is a linear positive normalized right invariant functional on F(G). Finally,
for every set X G, put
(X) =(
X
)
where
X
denotes the characteristic function of X, and check that is the required func-
tional.
10
. Let E be a set, S be an algebra of subsets of E and let

: S [0, +]
be a nitely additive functional on S, i.e.,
(AB) = (A) +(B)
for any two disjoint sets A S and B S.
Prove that there exists a functional
: P(E) [0, +],

which extends and is nitely additive on P(E).
For this purpose, use the Zorn lemma and an argument similar to the method of extending
measures described in Chapter 2.
11
. Let E be a set, G be a group of transformations of E and let S be a G-invariant algebra

of subsets of E. Suppose that a functional
: S [0, +]
is given such that:
(a) is nitely additive, i.e., (AB) = (A) +(B) for any two disjoint sets A S and
B S;
(b) is G-invariant.
Assuming that G is a solvable group, prove that there exists a functional
: P(E) [0, +]
which extends and is also nitely additive and G-invariant.
In order to demonstrate the existence of
, rst extend to a functional
: P(E) [0, +]
which is nitely additive (see Exercise 10). Then consider a linear positive normalized
right invariant functional
: F(G) R
for the given group G (see Exercise 9). Further, for any set A E, dene the function
f
A
: G R
by putting
f
A
(g) =
(g(A)) (g G).
In general, f
A
is an unbounded real-valued function, so we cannot assert that f
A
F(G).
However, for each natural number n, the relation
min( f
A
, n) F(G)
holds true. Now, dene
(A) = lim
n+
(min( f
A
, n))
and check that
is the required functional.

For this purpose, observe that if A and B are any two disjoint subsets of E, then
min( f
A
+ f
B
, n) min( f
A
, n) +min( f
B
, n) min( f
A
+ f
B
, 2n),
whenever n <. This relation readily implies the nite additivity of
.
Furthermore, if g G, A E and n <, then
(min( f
g(A)
, n)) =(min(( f
A
)
g
, n)) =((min( f
A
, n))
g
) =(min( f
A
, n))
from which the G-invariance of
easily follows.
In particular, infer from the stated above that there exists a nonnegative nitely additive
functional which is dened on the family of all subsets of the plane R
2
, is invariant under
the group of all isometric transformations of R
2
and extends the two-dimensional Lebesgue
measure
2
(Banachs classical theorem).
In order to show the validity of Banachs theorem, take into account the fact that the group
of all motions of R
2
is solvable and apply the general result presented in this exercise.
Chapter 18
Universally measurable additive functionals
This chapter is devoted to those additive functionals on a real Hilbert space H which have
good, in some generalized sense, measurability properties with respect to a certain class of
measures on H but, simultaneously, are bad from the topological point of view, that is are
discontinuous at all points of H.
Primarily, additive functionals on innite-dimensional vector spaces (over R) will be under
consideration in this chapter. In particular, for an innite-dimensional separable Hilbert
space H, the problem of measurability of additive functionals
f : H R
with respect to various extensions of -nite diffused Borel measures on H will be dis-
cussed (cf. Chapters 2 and 5).
More precisely, it will be shown below that there exists a real-valued additive everywhere
discontinuous functional f on H such that for any -nite diffused Borel measure on
H, this f can be made measurable with respect to an appropriate extension of . Special
attention will be paid to the case where is invariant or quasi-invariant under a certain
subgroup of H (in this connection, see also Chapters 3 and 4).
Let E be a topological space such that all singletons in E are Borel subsets of E. For
example, if every singleton in E is closed (i.e., E is a T
1
-space), then E automatically
satises the above-mentioned condition.
Recall that a Borel measure on E is diffused (or continuous) if (x) =0 for each point
x E.
As in preceding sections, we denote by the symbol M
/
E
the class of the completions of all
nonzero -nite Borel measures on E.
A set X E is called universally measurable (or absolutely measurable) with respect to M
/
E
if for any measure M
/
E
, we have X dom() (see Chapter 5).
281
8,
It is well known that if E is a Polish space, then all analytic (also, co-analytic) subsets of
E are universally measurable with respect to M
/
E
. For the proof, see, e.g., [24], [52], [99],
[187], or Appendix 2.
The family of all universally measurable sets in E forms a -algebra of subsets of E. A
straightforward verication of this simple fact is left to the reader.
Recall that a functional f : E R is universally measurable with respect to the class M
/
E
if
f is measurable with respect to any measure M
/
E
(cf. Chapter 5).
Let H denote an innite-dimensional separable Hilbert space (over the eld R of all real
numbers) and let
f : H R
be a universally measurable additive functional with respect to M
/
E
. Then f turns out to
be continuous (cf. [66], [97], [153], and [154]). In fact, f is continuous whenever it
is universally measurable with respect to the class of the completions of all probability
diffused Borel measures on H.
To demonstrate this circumstance, take such an f and suppose to the contrary that f is not
continuous. Then, for every natural number n, there exists an element h
n
H satisfying the
relations
[[h
n
[[ = 1, [ f (h
n
)[ > n 4
n
.
Let us put e
n
= h
n
/4
n
. Obviously, we have
[[e
n
[[ = 1/4
n
, [ f (e
n
)[ > n (n <).
Further, consider the Cantor space C = 0, 1
as a compact commutative group with re-

spect to the product topology and addition operation (modulo 2). EquipC with the comple-
tion
/
of the Haar probability measure , which is isomorphic to the standard Lebesgue
measure on the segment [0, 1] and, consequently, these two measures can be identied.
Let us introduce a mapping
: C H
by the formula
(
A
) =

nA
e
n
(A ),
where
A
denotes the characteristic function of a set A . It can easily be veried that:
(i) is injective and additive;
(ii) is continuous;
Universally measurable additive functionals 283
(iii) [( f )(
n
)[ =[ f (e
n
)[ > n for each n <.
Let denote the distribution in H of the random variable . Clearly, is a Borel diffused
probability measure on H and, according to our assumption, f is measurable with respect
to the completion of . This circumstance implies that the composition f is measurable
with respect to
/
. Since this composition is also additive on the power set P(), we
conclude (see Exercise 1 for this chapter) that the set
( f )(
n
) : n <
must be bounded in R which is impossible in view of relation (iii). The contradiction
obtained yields the required result.
Obviously, the same argument works for an arbitrary innite-dimensional Banach space
(E, [[ [[) instead of H.
We thus see that only continuous linear functionals on H can be universally measurable
with respect to the class of the completions of all -nite (equivalently, probability) dif-
fused Borel measures on H.
For a nite-dimensional Hilbert space over R, that is for Euclidean space, we have a much
stronger result which states that if an additive functional
f : R
n
R
is measurable with respect to the standard Lebesgue measure
n
on R
n
, then f is continuous
(see, e.g., [108] and [143]). An analogous assertion is valid for any homomorphism f acting
from a -compact locally compact topological group (, ) into the additive group R and
measurable with respect to the completion of the Haar measure on . Notice that the so-
called Steinhaus property of the Haar measure plays a signicant role in the proof (see [80],
[83], or Exercise 3 of this chapter).
Below, we shall introduce an essentially different notion of universal measurability of func-
tionals. Also, it will be shown in the sequel that there exist universally measurable additive
functionals on H (in the sense of our notion) which are everywhere discontinuous. Then
we will demonstrate how such universally measurable additive functionals can be applied
to some questions connected with the measure extension problem.
Let E be an arbitrary nonempty set and let M be a class of measures on E. We do not
assume, in general, that the measures from M are dened on one and the same -algebra
of subsets of E.
We recall once more from Chapter 5 that a functional
f : E R
is universally (absolutely) measurable with respect to M if f is measurable with respect to
all measures from M. But here we need a modied notion of universal measurability of
functionals.
We shall say that a functional f : E R is universally measurable with respect to M (in the
generalized sense) if for each measure M, there exists a measure
/
on E extending
and such that f becomes measurable with respect to
/
.
Evidently, this denition is more general than the denition given in the beginning of the
chapter.
Let us also remark that in both these denitions we may replace the real line R by an
uncountable Polish space or, more generally, by an uncountable Borel subset of a Polish
space since, according to the well-known result of descriptive set theory, any two uncount-
able Borel subsets of a Polish space are Borel isomorphic.
Recall that a subset X of a topological space E is a Bernstein set in E if for each nonempty
perfect set P E, we have
PX ,= / 0 & P(E X) ,= / 0.
Various properties of Bernstein sets are discussed in [30], [40], [72], [143], [148], [172],
[176], and [192]. In particular, any uncountable Polish space contains a Bernstein subset
and the cardinality of such a subset is equal to the cardinality of the continuum which as
usual is denoted by c.
For our further purposes, we need two lemmas.
Lemma 1. A separable Hilbert space H ,=0 can be represented in the form
H = X
1
+X
2
(X
1
X
2
=0),
where X
1
and X
2
are Bernstein subsets of H and, simultaneously, they are vector spaces
over the eld Q of all rational numbers.
Proof. The argument is fairly standard (cf. Exercise 6 for Chapter 5). Denote by the
least ordinal number of cardinality continuum and let P
: < be an enumeration of
all nonempty perfect subsets of H. By using the method of transnite recursion, it is not
difcult to construct two -sequences
x
: < H, y
: < H
(a) the family x
: < y
: < is linearly independent over Q;

(b) for any ordinal <, we have x
and y
.
Now, let us take as X
1
the vector space (over Q) generated by x
: <. Further, let X

2
denote a maximal vector subspace V of H (over Q) such that
V X
1
=0, y
: < V.
The existence of V is evident and it is also easy to see that X
1
and X
2
are the required
Bernstein subsets of H.
Remark 1. The proof given above works for an arbitrary Banach space (E, [[ [[) of cardi-
nality continuum, not necessarily separable (see Exercise 4 for this chapter).
Lemma 2. Let H ,=0 be a separable Hilbert space. There exists an additive functional
f : H R
for any -nite diffused Borel measure on H and for any -nite measure on R, the
graph Gr( f ) of f is a ( )-thick subset of the product space HR, that is this graph
intersects every ( )-measurable set of strictly positive measure.
Proof. First, applying Lemma 1, we represent H in the form
H = X
1
+X
2
(X
1
X
2
=0),
where X
1
and X
2
are Bernstein subsets of H and, simultaneously, they are vector spaces
over Q. Observe now that in view of the equalities
card(X
1
) = c, card(R) = c,
the vector spaces X
1
and R (again over Q) are isomorphic to each other. Let
: X
1
R
denote some isomorphism between these two spaces. We dene a functional f as follows.
Take any x H. This x admits a unique representation in the form
x = x
1
+x
2
,
where x
1
X
1
and x
2
X
2
. Let us put
f (x) = f (x
1
+x
2
) =(x
1
).
Obviously, f is an additive functional on H. Further, x a -nite diffused Borel measure
on H and a -nite measure on R. We assert that the graph of f is ( )-thick
in H R. Indeed, if Z is an arbitrary ( )-measurable set with ( )(Z) > 0, then,
according to the Fubini theorem, there exists a point t R such that (Z(t)) > 0 where
Z(t) =x H : (x, t) Z.
Consider the point
x
1
=
1
(t) X
1
.
Since X
2
is a Bernstein subset of H and Z(t) is an uncountable Borel subset of H, we have
X
2
(Z(t) x
1
) ,= / 0.
Choose a point x
2
from the set X
2
(Z(t) x
1
) and dene
x = x
1
+x
2
.
Then we get
x = x
1
+x
2
Z(t), (x, t) Z,
(x, f (x)) = (x, t), t = f (x), (x, f (x)) Z,
Let f : H R be as in Lemma 2. Notice that such an f is everywhere discontinuous on
H because the graph of any measurable (in particular, continuous) functional is always
of measure zero, so it cannot be thick in the product space H R. Fix a Borel diffused
probability measure on R. For instance, we may take as an arbitrary Borel probability
measure equivalent to the standard Lebesgue measure on R assuming that the latter
measure is restricted to the Borel -algebra of R. Now, let be any -nite diffused Borel
measure on H. For each ( )-measurable set Z H R, we denote
Z
/
=x H : (x, f (x)) Z.
Further, we put
S
/
=Z
/
: Z dom( ).
It can easily be veried that S
/
is a -algebra of subsets of H containing the Borel -
algebra of H. Finally, we dene a functional
/
on S
/
by putting
/
(Z
/
) = ( )(Z) (Z dom( )).
A straightforward verication shows that the denition of
/
is correct in view of the (
)-thickness of the graph of f (see Chapter 2). Also,
/
turns out to be a measure on S
/
which extends the original measure .
Keeping in mind all these facts, we are able to prove the following statement.
Theorem 1. The additive functional f : H R just described is universally measurable
(in the generalized sense) with respect to the class of all -nite diffused Borel measures
on H.
Proof. It sufces to show that for any -nite diffused Borel measure on H, the func-
tional f is measurable with respect to the measure
/
. For this purpose, take an arbitrary
Borel subset B of R. Clearly, we may write
f
1
(B) =x H : f (x) B =x H : (x, f (x)) HB S
/
which establishes the measurability of f with respect to
/
and, therefore, nishes the
proof.
Let us indicate some other properties of f nice from the measure-theoretical point of view.
It is well known that if H is an innite-dimensional separable Hilbert space, then there
exists no nonzero -nite Borel measure on H invariant (quasi-invariant) under the group
of all translations of H (see, for instance, Chapter 4). At the same time, there exist various
nonzero -nite Borel measures on H which are invariant under everywhere dense vector
subspaces of H (see Chapter 3). Let be any such measure on H. We already know that
f is measurable with respect to an appropriate extension
/
of . So it is natural to ask
whether the measure
/
can be chosen to be invariant under the same everywhere dense
vector subspace of H. It turns out that the answer to this question is always positive.
In order to demonstrate this fact, let us slightly change the construction presented above (cf.
[141] and Chapter 14). We replace the real line R by the one-dimensional unit torus T re-
garded as a compact commutative group. The torus T is equipped with the Haar probability
measure which, actually, coincides with the Lebesgue probability measure on T invariant
under all translations of T. We denote the latter measure by the same symbol .
The following statement is true.
Lemma 3. Let H ,=0 be a separable Hilbert space. There exists a group homomorphism
f : H T
having the property that for any Borel -nite diffused measure on H, the graph of f is
a ( )-thick subset of the product space HT.
Proof. The argument given below is very similar to that of the proof of Lemma 2. Again,
let us represent H in the form
H = X
1
+X
2
(X
1
X
2
=0),
where both X
1
and X
2
are Bernstein subsets of H and, simultaneously, they are vector
spaces over the eld Q of all rational numbers. Taking into account the facts that X
1
and
R are isomorphic as vector spaces over Q and that there is a canonical surjective group
homomorphism from R onto T, it is easy to dene a surjective group homomorphism
: X
1
T.
Now, if x is an arbitrary element of H, then x admits a unique representation in the form
x = x
1
+x
2
(x
1
X
1
, x
2
X
2
).
So, we may dene
f (x) = f (x
1
+x
2
) =(x
1
).
In this way, we obtain a group homomorphism f : H T. Let be an arbitrary -nite
diffused Borel measure on H. Take any set Z dom( ) with ( )(Z) > 0. Accord-
ing to the Fubini theorem, (Z(t)) > 0 for some t T. Since is a surjection, we can nd
x
1
X
1
such that (x
1
) =t. Keeping in mind that X
2
is a Bernstein subset of H, we get
X
2
(Z(t) x
1
) ,= / 0.
Choose an element x
2
from the set X
2
(Z(t) x
1
) and put
x = x
1
+x
2
.
As in the proof of Lemma 2, we easily conclude that (x, f (x)) Z, which shows the (
)-thickness of the graph of f in the product space HT.
Starting with the homomorphism f : H T described in the preceding lemma, we can
extend any -nite diffused Borel measure on H to the measure
/
on the same H (by
using the scheme indicated before).
Theorem 2. Let H ,= 0 be a separable Hilbert space and let be a -nite Borel
measure on H invariant (quasi-invariant) under some subgroup G of H. Then the measure
/
is also invariant (quasi-invariant) under G.
Proof. It sufces to consider the case of an invariant measure (for quasi-invariant mea-
sures the argument is absolutely analogous). Take any set Z
/
from the domain of
/
. Ac-
cording to the denition, we may write
Z
/
=x H : (x, f (x)) Z,
where Z belongs to the domain of . For each element g G, we have
Z
/
+g =x H : (x g, f (x g)) Z =
x H : (x, f (x)) Z +(g, f (g)).
Since the product measure is (GT)-invariant (see Exercise 9 from Chapter 3), we
may write
( )(Z) = ( )(Z +(g, f (g)))
and, consequently,
/
(Z
/
) =
/
(Z
/
+g)
which completes the proof of the theorem.
Obviously, for any nonzero -nite Borel measure on H, the measure
/
obtained by
using the group homomorphism f : H T is a proper extension of . But the -algebra
dom(
/
) is not signicantly larger than the -algebra dom(). Indeed, it can easily be
seen that dom(
/
) is a countably generated -algebra and, therefore, the measure
/
is
separable. If we wish to extend to a nonseparable measure, then we must replace the torus
T by the innite-dimensional torus T
c
which is the topological product of continuumly
many copies of T. Let now denote the Haar probability measure on T
c
. This measure is
nonseparable. Moreover, there exists a family of sets
Z
i
: i I dom()
having the following properties:
(1) card(I) = c;
(2) (i I)((Z
i
) > 0);
(3) for every -measurable set Z with (Z) > 0, there is a set Z
i
contained in Z.
The proof of the existence of Z
i
: i I with the above-mentioned properties can be found,
e.g., in [83] (cf. also Exercise 11 from Chapter 2).
Lemma 4. There exists a group homomorphism
f : H T
c
such that for any -nite diffused Borel measure on H, the graph of f is a ( )-thick
subset of HT
c
.
Proof. Fix a family of sets Z
i
: i I dom() satisfying the relations (1) - (3) formulated
above. Without loss of generality, we may suppose that the set I of indices is well-ordered
and its order type coincides with the least ordinal number whose cardinality is equal to
c. In other words, we may put
Z
i
: i I =Z
: <.
As before, we represent our Hilbert space H in the form
H = X
1
+X
2
(X
1
X
2
=0),
where X
1
and X
2
are some Bernstein subsets of H and, simultaneously, they are vector
spaces over Q. Now, it is not difcult to construct two injective -sequences
y
: < X
1
, z
: < T
c
(a) the family y
: < is linearly independent over Q;

(b) for each <, we have z
.
Further, we put
(y
) = z
( <).
Taking into account condition (a) and the fact that T
c
is a divisible group, we can extend
the mapping to a group homomorphism
: X
1
T
c
.
In view of condition (b), the set (X
1
) is -thick in T
c
.
Now, take any x H. We have a unique representation
x = x
1
+x
2
,
where x
1
X
1
and x
2
X
2
. Dene
f (x) = f (x
1
+x
2
) =(x
1
).
In this manner, we get a group homomorphism
f : H T
c
.
Let us show that f is the required homomorphism. Consider any -nite diffused Borel
measure on H. Let Z be an arbitrary ( )-measurable set such that ( )(Z) >
0. Applying again the Fubini theorem and the -thickness of (X
1
) in T
c
, we see that
(Z(t)) > 0 for some t (X
1
). Choose an element x
1
X
1
such that (x
1
) =t. Since X
2
is a Bernstein set in H, we have
X
2
(Z(t) x
1
) ,= / 0.
Consequently, we may pick some point x
2
X
2
(Z(t) x
1
). Then we put x = x
1
+x
2
. An
easy verication shows that (x, f (x)) Z (cf. the nal part of the proof of Lemma 2). Thus,
the graph of f is ( )-thick in the product space HT
c
and the proof is complete.
Applying the previous lemma, we readily come to the following result.
Theorem 3. Let H ,=0 be a separable Hilbert space and let a group homomorphism
f : H T
c
be as in Lemma 4. For every -nite diffused Borel measure , denote by
/
the extension
of obtained by using this group homomorphism. Then we have:
1)
/
is a nonseparable measure;
2) if is invariant (quasi-invariant) under a group G H, then
/
is invariant (quasi-
invariant) under the same G;
3) f is measurable with respect to
/
.
This statement also indicates that f is universally measurable in the generalized sense, i.e.,
for any -nite diffused Borel measure on H, there exists an extension
/
of such that
f is
/
-measurable.
Remark 2. Obviously, the preceding results are valid in the particular case where the n-
dimensional Euclidean space R
n
(n 1) is taken as H. Thus, we may assert that there
exists an additive functional
f : R
n
R
which is discontinuous at all points of R
n
and is universally measurable (in the generalized
sense) with respect to the class of all -nite diffused Borel measures on R
n
.
Remark 3. It should be noticed that if a vector space E (over the eld Q) is of cardinality
continuum, then, assuming Martins Axiom, there exists an injective additive functional
g : E R
which is absolutely nonmeasurable with respect to the class of all nonzero -nite diffused
measures on E. Indeed, the existence of such a g is implied by the fact that there exists
a generalized Luzin subset of R which simultaneously is a vector space over Q (for more
details, see [59] or Chapters 5 and 12). Consequently, under Martins Axiom, there are
real-valued absolutely nonmeasurable additive functionals on a separable Hilbert space
H ,=0.
Remark 4. As has already been mentioned, the standard notion of universal measurability
of functionals on a Hilbert space yields nothing new for the class of additive functionals
because every universally measurable additive functional turns out to be continuous. More-
over, it was shown in [66] that even the mid-point convexity of a given functional with its
universal measurability (in the standard sense) implies its continuity. This result can be
regarded as an analogue of Sierpi nskis old theorem stating that any Lebesgue measurable
mid-point convex function dened on a nite-dimensional Euclidean space is necessarily
continuous (see [220]; cf. also [122] and [143]).
In this context, we should also indicate paper [97] where the universal measurability (in the
standard sense) of linear functionals, with respect to the class of all Gaussian measures on
a Hilbert space H, is envisaged.
EXERCISES
1
. Consider the standard Cantor space

C = 2
=0, 1
.
This C is a compact commutative group with respect to the addition operation (modulo
2). The same space may be regarded as the family P() of all subsets of in which the
symmetric difference of sets is considered as a basic operation +, i.e., we have
X +Y = XY = (X Y) (Y X) (X , Y ).
Actually, these two algebraic structures are canonically isomorphic, so they can be treated
as identical to each other by identifying any subset X of with its characteristic function
X
.
Let : P() R be a function. We shall say that is additive if
(X +Y) =(X) +(Y)
for any two disjoint subsets X and Y of .
We shall say that : P() R is measurable in the Luzin sense if it is measurable with
respect to the completion
/
of the Haar probability measure on C. Recall that
/
is
isomorphic to the Lebesgue measure on the closed unit interval [0, 1].
Prove that if is additive and measurable in the Luzin sense, then
lim
n+
(n) = 0.
In particular, conclude that the sequence (n) : n < is necessarily bounded in R.
Do it in the following manner. Suppose otherwise, i.e., suppose that the above-mentioned
equality fails to be true. This fact means that there exist a neighborhood V of the neutral
element in C and a strictly increasing sequence n
k
: k < of natural numbers such that
(n
k
) ,V (k <).
Since is
/
-measurable and
/
is Radon, there exists a compact set P C satisfying the
relation
/
(P) > 1/2 and such that the restriction [P is continuous in view of Luzins
classical theorem on the structure of real-valued Lebesgue measurable functions.
The invariance of the Haar measure implies the relation
P+P =C.
This relation means that each subset Z of admits a representation Z =XY, where X P
and Y P. Applying this fact to the sequence n
k
: k <, infer that
n
k
= A
k
B
k
(k <),
where all sets A
k
and B
k
belong to P. Deduce from these relations that, without loss of
generality, we may suppose
A
k
=n
k
B
k
, n
k
, B
k
(k <).
Further, taking into account the compactness and metrizability of P and passing if necessary
to an appropriate subsequence, assume that
lim
k+
A
k
= A, lim
k+
B
k
= B
for some A P and B P. Now, keeping in mind that
A
k
= B
k
+n
k
(k <)
and lim
k+
n
k
= / 0, conclude that A = B.
Since the restriction [P is continuous, we can write
lim
k+
(A
k
) =(A) =(B) = lim
k+
(B
k
),
from which it follows that
lim
k+
(n
k
) = lim
k+
((A
k
) (B
k
)) = 0.
But this circumstance contradicts the assumption that (n
k
) , V for all k < . The
contradiction obtained yields the required result.
2
. Let H be an innite-dimensional separable Hilbert space. By using the result of the

previous exercise, prove the continuity or, equivalently, boundedness of an arbitrary uni-
versally measurable (in the usual sense) additive functional f : H R.
Also, consider the question whether an analogous result can be established for an arbitrary
universally measurable (in the usual sense) additive functional g : H T.
3. Let (, ) be a -compact locally compact topological group equipped with its Haar
measure =
. Suppose that an additive functional

f : R
is given which is measurable with respect to the completion of . Verify that f is continu-
ous.
For this purpose, apply the Steinhaus property of (see Chapter 3) and show that f is
bounded on some neighborhood of the neutral element of G, fromwhich the required result
easily follows.
4
. Prove an analogue of Lemma 1 for an arbitrary Banach space (E, [[ [[) of cardinality
continuum.
For this purpose, take into account the fact that the family of all separable perfect subsets
of E is also of cardinality c.
5. Give a detailed proof of Theorem 3.
6
. A function f : R
n
R is called mid-point convex if the relation
f ((x +y)/2) ( f (x) + f (y))/2
holds true for all points x and y from R
n
.
Prove Sierpi nskis theorem stating that any mid-point convex Lebesgue measurable func-
tion
f : R
n
R
is continuous and, therefore, is convex in the usual sense.
This theorem strengthens the well-known result of Fr echet which states that if a solution of
the Cauchy functional equation is measurable in the Lebesgue sense, then it is continuous
(i.e., turns out to be a trivial solution of this equation).
7
. Recall that if E is a vector space, then the symbol dim(E) denotes the algebraic dimen-
sion of E, i.e., dim(E) stands for the cardinality of any basis of E (this cardinality does not
depend on a choice of a basis).
As usual, we say that a vector space E is nite-dimensional if dim(E) <. Otherwise, we
say that E is innite-dimensional.
Let E be an innite-dimensional topological vector space such that the neutral element 0
of E admits a fundamental system (i.e., local base) V
i
: i I of its open neighborhoods,
satisfying the inequality card(I) dim(E).
Show that there exists a linear functional
f : E R
which is discontinuous at all points of E.
For this purpose, denote by the least ordinal number whose cardinality is equal to card(I).
Let V
: < be an enumeration of all sets from the local base V

i
: i I. By using
transnite induction on <, dene an -sequence e
: < of linearly independent

elements of E such that e
for each <. Then put

f (e
) = 1 ( <)
and extend this partial function to a linear functional f : E R. Verify that such an f is
discontinuous at 0 and hence it is discontinuous at all points of E.
Conclude fromthis result that if E is any innite-dimensional metrizable topological vector
space, then there exists a linear functional f : E R everywhere discontinuous on E.
On the other hand, if E is an arbitrary vector space and the symbol F denotes the family
of all linear functionals on E, then the so-called weak topology (E, F) on E is dened in
such a way that each functional f F becomes continuous on the topological vector space
(E, (E, F)). This fact shows, in particular, that the cardinality of any local base of the
neutral element of (E, (E, F)) is strictly greater than dim(E).
8. Let H be a Hilbert space over R (not necessarily separable) and let
f : H R
be an additive functional universally measurable with respect to the class M
/
H
. Verify that
f is continuous.
Chapter 19
Some subsets of the Euclidean plane
Here we would like to consider those subsets of the Euclidean plane R
2
which possess
rather strange geometrical and measure-theoretical properties. In particular, we will be
dealing with thick (massive) subsets of R
2
having small linear sections in all possible di-
rections, and we will investigate these sets from the point of view of the measurability with
respect to translation-invariant extensions of the classical Lebesgue measure
2
on R
2
.
Many examples of paradoxical sets in a nite-dimensional Euclidean space that have small
sections by certain afne hyperplanes in this space are known in the literature. Some of
them have already been discussed in this book (cf. Chapters 2, 5, and 7). In this context,
it is reasonable to remind that one of the earliest examples of this sort is due to Sierpi nski
who was able to construct a bijective function
f : R R
such that its graph Gr( f ) is
2
-thick in R
2
. Obviously, the
2
-thickness of Gr( f ) directly
implies that Gr( f ) is nonmeasurable with respect to
2
and hence f is not measurable in the
Lebesgue sense. At the same time, any straight line in R
2
parallel to the axis of abscissae
or to the axis of ordinates meets the graph of f in exactly one point.
Further, it should be mentioned in this context that Mazurkiewicz constructed a subset X of
R
2
having the property that for each straight line l in R
2
, the set l X consists of precisely
two points (see [170]). Notice that X can be chosen to be
2
-thick and, consequently,
nonmeasurable with respect to
2
(see Exercise 1 for this chapter). Later, various examples
of sets with small sections, but being large in a certain sense, were given by other authors
(cf. [72], [105], [119], [176], and [192]).
In the present chapter, we will consider some sets of this type in connection with the fol-
lowing natural question.
How small are such sets from the point of view of translation-invariant extensions of the
Lebesgue measure
2
?
297
9,
Namely, we are going to demonstrate in our further considerations that there are translation-
invariant extensions of
2
concentrated on sets with small linear sections in all directions.
Actually, a more stronger result will be obtained stating that the corresponding sets can
have small sections by all analytic curves lying in the plane.
Let
n
denote the standard n-dimensional Lebesgue measure on the Euclidean space R
n
,
where n 1.
Below, we need the following auxiliary statement, which also turns out to be rather useful
in many other situations (cf. [108], [181], and [240]).
Lemma 1. Let Z be a
n
-measurable subset of the Euclidean space R
n
with
n
(Z) >0 and
let {M
i
: i I} be a family of analytic manifolds in R
n
, such that:
1) card(I) is strictly less than the cardinality of the continuum;
2) for each index i I, the dimension of M
i
is strictly less than n.
Then the relation
n
(Z \ {M
i
: i I}) > 0
is satised. In particular, we always have Z \ {M
i
: i I} = / 0.
The proof of this lemma is not difcult and can be carried out by induction on n (here the
classical Fubini theorem plays a dominant role). For more details, see [108], [181], [240],
and Exercise 3 of this chapter.
Some applications of Lemma 1 to certain questions of the geometry of Euclidean spaces
may be found in [108] and [240]. Notice also that this lemma does not hold longer true
for topological manifolds in R
n
. For example, there are models of set theory in which the
Euclidean plane can be covered by a family of Jordan curves, whose cardinality is strictly
less than the cardinality of the continuum (cf. Exercise 5 of this chapter).
As usual, we denote by c the cardinality of the continuum. Starting with the above-
mentioned lemma, we are able to establish the following result.
Theorem 1. Let G be the group of all analytic diffeomorphisms of R
n
onto itself. Then
there exists a subset X of R
n
such that:
1) card(X) = c and X is almost G-invariant, i.e.,
(g G)(card(g(X)X) < c);
2) X is
n
-thick in R
n
;
3) for any analytic manifold M in R
n
with dim(M) < n, we have
card(MX) < c.
Some subsets of the Euclidean plane 299
Proof. We use the method rst suggested and successfully applied by Sierpi nski (see [222];
compare also [83], [95], [108], [115], [176], [192], [234], and [235]). This method has
already been exploited in preceding sections of the present book (see, for instance, Chapters
3 and 16).
Let denote the least ordinal number of cardinality continuum. Since the equality
card(G) = c holds, we may write
G = {G
: < },
where {G
: < } is some -sequence of subgroups of G which satises these two

conditions:
(a) for each ordinal < , we have the inequality
card(G
) card() +;
(b) the family {G
: < } is increasing by inclusion.

Also, we may consider the family {Y
: <} consisting of all those Borel subsets of R

n
which are of strictly positive
n
-measure. Finally, we denote by {M
: < } the family

of all those analytic manifolds in R
n
whose dimensions are strictly less than n.
Now, by using the method of transnite recursion, we dene an injective -sequence
{x
: < } of points in R
n
(i) for any < , the point x
belongs to Y
;
(ii) for any < , we have
x
({M
: } {x
: < }).
Notice that Lemma 1 guarantees, at each -step of our recursion, the existence of a needed
point x
. So the recursion can be continued up to . In this way, we will be able to construct

the required -sequence of points {x
: < }. Finally, putting

X = {G
(x
) : < },
we can readily check that this X is the desired set, which nishes the proof of Theorem 1.
The next statement is a trivial consequence of the above theorem.
Theorem 2. Let G denote the group of all isometric transformations (i.e., motions) of R
n
.
Then there exists a subset X of R
n
such that:
1) card(X) = c and X is almost G-invariant;
2) X is
n
-thick in R
n
;
3) for each analytic manifold M in R
n
card(X M) < c.
In particular, for any afne hyperplane L in R
n
, the inequality card(X L) < c is true.
Moreover, if the Continuum Hypothesis holds, then according to Theorem 2, for each ana-
lytic manifold M in R
n
card(X M) .
Let us point out one application of Theorem 2 to a particular case of the measure extension
problem.
Dealing with various invariant extensions of the Lebesgue measure
n
, the following ques-
tion naturally arises.
Does there exist a measure
n
on R
n
extending
n
, invariant under the group of all isometric
transformations of R
n
, and concentrated on some subset X of R
n
with small sections by all
afne hyperplanes?
Of course, here the smallness of sections means that for any afne hyperplane L in R
n
, the
cardinality of LX is strictly less than the cardinality of the continuum.
Theorem 2 immediately yields a positive answer to this question.
Indeed, let G be again the group of all isometric transformations of R
n
. Consider the
G-invariant -ideal of subsets of R
n
generated by R
n
\ X, where X is the set indicated
in Theorem 2. We denote this -ideal by J. Then for any set Z J, the inner
n
-
measure of Z is equal to zero since X is almost G-invariant and
n
-thick in R
n
. Taking this
fact into account and applying the standard methods of extending invariant measures (see,
for instance, [234], [235], or Chapter 2), we derive that there exists a measure
n
on R
n
1)
n
extends
n
;
2)
n
is invariant under the group G;
3) J dom(
n
);
4) for each set Z J, we have
n
(Z) = 0.
By virtue of relation 4), we also have
n
(R
n
\ X) = 0,
that is our measure
n
is concentrated on X. At the same time, we know that X has small
sections by all afne hyperplanes in R
n
and, moreover, by all analytic manifolds in R
n
whose dimensions are strictly less than n.
In particular, for n =2, we obtain that there exists a measure
2
on the Euclidean plane R
2
,
such that:
(1)
2
is an extension of the two-dimensional Lebesgue measure
2
;
(2)
2
is invariant under the group of all isometric transformations of the plane R
2
;
(3)
2
is concentrated on some subset X of R
2
having the property that all linear sections of
X are small (i.e., they are of cardinality strictly less than the cardinality of the continuum).
Notice that the question on the existence of a measure
2
on the plane, satisfying conditions
(1) - (3), was formulated by Mabry (oral communication). Moreover, we see that the above-
mentioned support X of
2
has a stronger property, namely, for each analytic curve l in R
2
,
the cardinality of l X is strictly less than c.
In this context, let us point out that in [105] an analogous question was considered for
those extensions of
2
which are invariant under the group of all translations of R
2
. More
precisely, it was demonstrated in [105] that there exists a translation-invariant extension of
2
concentrated on a subset of R
2
all linear sections of which are at most countable.
Remark 1. If the Continuum Hypothesis holds, then we also can conclude that the above-
mentioned subset X of the plane has the property that all its linear sections are at most
countable and, obviously, the same is true for sections of X by all analytic curves lying in
the plane. Thus, in the case of the validity of CH, the measure
2
is concentrated on a set
all linear sections of which are countable. In this connection, the following question seems
to be natural.
Does there exist a measure
2
on the plane R
2
, extending
2
, invariant under the group of
all isometric transformations of R
2
, and concentrated on a set all linear sections of which
are nite?
It turns out that the answer to this question is negative (cf. Exercise 7).
Remark 2. By assuming some additional set-theoretical axioms, a result much more
stronger than Theorem 1 can be established. Namely, let us suppose that for any cardi-
nal < c, the space R
n
cannot be covered by a -sequence of
n
-measure zero sets. In
fact, it sufces to assume this property of R
n
only for n = 1, i.e., that the real line R cannot
be covered by a family of
1
-measure zero sets, whose cardinality is strictly less than c.
For example, it is well known that this hypothesis follows directly from Martins Axiom
(see, e.g., [10], [40], [67], [145], and [216]).
We say that a group G of transformations of R
n
is admissible if each element g from G
preserves the -ideal of all
n
-measure zero subsets of R
n
.
For instance, if G coincides with the group of all afne transformations of R
n
, then G is
admissible.
Let us x an admissible group G with card(G) = c. Then, applying an argument similar to
the proof of Theorem1, we easily obtain the next statement which strengthens this theorem.
There exists a subset X of R
n
1) card(X) = c and X is almost G-invariant;
2) X is
n
-thick in R
n
;
3) for any
n
-measure zero subset Z of R
n
, we have card(Z X) < c.
Condition 3) shows, in particular, that the above-mentionedset X is a generalized Sierpi nski
subset of R
n
. For the denition and various properties of Sierpi nski sets, see [40], [148],
[176], [192], and [222] where the dual objects to the Sierpi nski sets (i.e., Luzin sets) are
discussed as well. The general denition of Luzin type sets for a given ideal of subsets of
an abstract space E is formulated in Exercise 14 of Appendix 1.
Recall that from the measure-theoretical view-point and, more exactly, from the point of
view of topological measure theory, Luzin sets in R (respectively, in R
n
) are very small
because they turn out to be universal measure zero subsets of R (respectively, of R
n
).
Here we would like to give one more construction of a subset of R
2
of Luzin type. This
construction is in connection with a certain innite version of a well-known problem from
combinatorial geometry, rst considered by Erd os and Szekeres.
Let E be a vector space (over R) and let X be a subset of E. We recall that X is said to be
convexly independent in E if for each point x X, the relation
x conv(X \ {x})
is valid, where conv() as usual denotes the operation of taking convex hulls of subsets of
E.
In such a case, it is also said that the points of X are in convex position.
In their joint paper [63] Erd os and Szekeres proved the following statement.
Erd os-Szekeres Theorem. For any natural number m 3, there exists a smallest natural
number N(m) possessing the following property:
every set consisting of N(m) points in general position in the plane R
2
contains a subset
of m points which is convexly independent, i.e., those m points are in convex position and,
consequently, they are the vertices of some convex polygon.
In addition, Erd os and Szekeres posed the problem to determine the precise value of N(m)
and in connection with this problem they conjectured that
N(m) = 2
m2
+1.
The above-mentioned intriguing problem was then investigated by many mathematicians.
However, it still remains unsolved. An extensive survey about this problem and closely
related questions of combinatorial geometry can be found in [178] where a further list of
references is also presented.
As said above, in this chapter we would like to discuss some innite variant of the Erd os-
Szekeres problem. The main attention will be paid to those innite sets of points, which
are either countable or are of cardinality continuum.
Recall that the basic technical tool utilized by Erd os and Szekeres in [63] is the classical
combinatorial theorem of Ramsey [204]. By applying this theorem, we easily get an upper
bound for N(m). The same method successfully works for point-sets lying in Euclidean
spaces of higher dimension (see, e.g., [178]).
Here we only need a countably innite version of the Ramsey theorem which is formulated
as follows.
Ramsey Theorem. Let i be a nonzero natural number, A be an arbitrary innite set, F
i
(A)
denote the family of all i-element subsets of A and let {L
1
, L
2
, ..., L
k
} be a nite covering
of F
i
(A). Then there exist a nonzero natural number r k and an innite subset B of A
such that F
i
(B) L
r
.
It is well known that from this innite version of the Ramsey theorem its more popular
nite version can be derived with the aid of the K onig lemma (see, for instance, [10]) or
with the aid of the existence of a nontrivial ultralter on the set of all natural numbers.
Further, by applying the above-mentioned innite version of the Ramsey theorem, the fol-
lowing result can easily be obtained.
Theorem 3. Let n 2 be a natural number and let A be an innite subset of the Euclidean
space R
n
, such that card(Al) < for all straight lines l lying in R
n
. Then there exists
an innite convexly independent set B A. In particular, if card(A) = , then we have
card(B) = card(A).
Proof. The assumption of the theorem implies that there are afne linear manifolds L in
R
n
satisfying the relation card(AL) . Let L
be such a manifold whose dimension is

minimal. Observe that dim(L
) > 1 and every afne hyperplane H in L
has the property

that card(HA) <. By using an easy induction, we can dene an innite set A
AL
of points in general position in L
. Now, applying to A
the above-mentioned innite version

of the Ramsey theorem(namely, for i =dim(L
) +2 and k =2), we get the required innite

convexly independent set B A
A which ends the proof.

Remark 3. It should be noticed that the use of the Ramsey theorem for obtaining the asser-
tion of Theorem 3 is not necessary. Indeed, let us rst consider the case of the Euclidean
plane R
2
. Suppose that A is an innite subset of R
2
such that card(A l) < for any
straight line l lying in R
2
. Then A contains an innite countable subset A
of points in
general position. Only three cases are possible:
1) there exist a ray p R
2
and an innite set {a
n
: n < } A
, such that {a
n
: n <
} p = / 0, the sequence of points {a
n
: n < } converges to the end-point o of p and the
rays oa
n
(n < ) converge to p;
2
n
: n < } A
, such that
{a
n
: n < } p = / 0, lim
n
||a
n
|| = +, lim
n
dist(a
n
, p) = 0;
2
n
: n < } A
, such that
{a
n
: n < } p = / 0, lim
n
||a
n
|| = +, lim
n
dist(a
n
, p) = +
and the rays oa
n
(n < ) converge to p, where o is again the end-point of p.
In each of these cases, it is not difcult to dene (by induction) a polygonal convex curve
P with innitely many sides whose all vertices belong to the set {a
n
: n < } and hence
to the set A, as well. In the rst case, P can be chosen in such a way that the lengths of
its sides will be tending to zero. In the second and third cases, P can be chosen so that the
lengths of its sides will be tending to innity.
Suppose now that A is an innite subset of the Euclidean space R
n
, where n > 2, and sup-
pose that card(Al) < for all straight lines l lying in R
n
. Again, there exists an innite
countable set A
A of points in general position in some afne linear submanifold of R

n
.
Without loss of generality, we may assume that the above-mentioned manifold coincides
with R
n
. Further, since A
is countable, we can nd a plane L in R

n
such that the orthog-
onal projection of R
n
onto L, restricted to A
, is injective and transforms A
to a set of
points in general position in L. As we already know, the latter set contains an innite con-
vexly independent subset. Now, we can readily verify that the pre-image with respect to the
above-mentioned projection of this subset yields the required innite convexly independent
subset of A
and hence of A, as well.

It would be interesting to investigate the question whether it is possible to derive the Erd os
Szekeres result directly from Theorem 3 without using the nite version of the Ramsey
theorem.
Now, let us discuss the case when uncountable point-sets in the plane R
2
are under consid-
eration and examine an analogue of the Erd os-Szekeres problem for such sets.
Let A be an uncountable subset of the plane. Assuming that all points of A are in general
position, the following natural question arises.
Does there exist an uncountable convexly independent subset of A?
We shall demonstrate below that, in general, the existence of such a subset of A cannot be
guaranteed (within ZFC theory).
Let denote the least ordinal number of cardinality continuum. Consider the family of all
convex curves in R
2
. Obviously, we can represent this family in the form of a transnite
sequence {C
: < }. By using the method of transnite recursion, let us construct a

family {a
: < } of points in general position in the plane. Suppose that, for an ordinal
< , the partial family of points {a
: < } has already been dened. Take the curves

C
( < ). Since every curve C
is convex, the family of all those straight lines whose

intersection with C
contains more than two elements is at most countable. Hence there

exists a straight line l R
2
such that
( < )(card(l C
) 2).
Clearly, we can nd a point a
l not belonging to the set

{C
: < } {a
: < }.
Moreover, by slightly modifying the above argument, we can always choose a point a
l
in such a way that all the points a
( ) will be in general position in the plane R

2
.
Proceeding in this manner, we are able to construct the required family of points {a
: <
}. Finally, we put
A = {a
: < }.
Then for A, the following statement is true.
Theorem 4. The set A possesses these three properties:
1) card(A) = c;
2) the points of A are in general position in R
2
;
3) no subset of A of cardinality continuum is convexly independent.
Proof. The properties 1) and 2) follow immediately from our construction. Let us check
the validity of 3). Suppose to the contrary that some set B A of cardinality continuum is
convexly independent and put
B
= cl(conv(B)), C
= bd(B
),
where cl() and bd() denote the operations of taking the closure and boundary, respec-
tively. Then B
is a closed convex subset of R

2
with nonempty interior and C
as the bound-
ary of B
is a convex curve. Notice that all points of B must belong to C
. This fact can

easily be deduced, e.g., from the well-known Steinitz theorem which says that a point x be-
longs to the interior of the convex hull of a set X R
n
if and only if there exists a set Y X
with card(Y) 2n, such that x int(conv(Y)) (see, e.g., [76] and [77]). Consequently, we
get
card(AC
) card(B) = card() = c.
On the other hand, there exists an ordinal < such that C
=C
. Taking into account the

denition of A, we may write
AC
{a
: }
and hence
card(AC
) = card(AC
) card() + < c.
The obtained contradiction nishes the proof.
Remark 4. It should be observed that we were able to construct the set A of Theorem 4
because of a rather nice structure of convex curves lying in the plane R
2
. More precisely,
all convex curves in the plane turn out to be so thin in R
2
that any family {C
i
: i I} of such
curves, with card(I) < c, does not form a covering of R
2
. Indeed, this moment is crucial
for the transnite construction which was described before the formulation of Theorem
4. Notice that for more general classes of curves lying in the plane, the corresponding
situation can be signicantly different. Namely, there are models of set theory in which the
existence of a family {P
i
: i I} of homeomorphic images of [0, 1], satisfying the relations
card(I) < c, {P
i
: i I} = R
2
,
is possible (see, e.g., Exercises 5 and 6 below). In this situation, no construction analogous
to the presented above can be carried out.
Remark 5. Theorem 4 admits a direct generalization to the case of the Euclidean space
R
n
, where n 2. In other words, there exists a set A R
n
of cardinality continuum whose
points are in general position in R
n
but no subset of A of the same cardinality is convexly
independent. The proof of this result is rather similar to the proof of Theorem 4. However,
some insignicant technical details occur in the case of R
n
.
Remark 6. In various topics of real analysis, measure theory, and set-theoretical topol-
ogy analogous constructions are known for obtaining those sets which almost avoid the
members of a given family of sets (cf. [148], [159], [176], [192], and [225]). In most
situations, the given family of sets forms an ideal of subsets of an original space. For ex-
ample, we have the ideal of all measure zero sets in a nonzero complete measure space
and the ideal of all rst category sets in a nonempty Baire topological space. As a rule,
the corresponding transnite constructions need some additional set-theoretical axioms. In
this connection, see [148], [176] or [192] where by assuming the Continuum Hypothesis
the classical constructions of so-called Luzin sets and Sierpi nski sets are discussed in detail
(see also Exercise 14 for Appendix 1).
EXERCISES
1
. We say that X R
2
is a Mazurkiewicz subset of R
2
if for every straight line l R
2
, we
have card(X l) = 2.
Show that if a Mazurkiewicz set X is
2
-measurable, then
2
(X) = 0. Formulate and prove
an analogous assertion in terms of the Baire property and Baire category. In the latter case,
use the topological analogue of the Fubini theorem due to Kuratowski and Ulam (see [148]
or [192]).
Further, by applying the method of transnite induction, prove that the following two state-
ments are valid:
(a) there exists a Mazurkiewicz subset of R
2
which is thick with respect to
2
;
(b) there exists a Mazurkiewicz subset of R
2
which is of
2
-measure zero and nowhere
dense in R
2
.
2. Let J be a set of cardinality continuum and let {l
j
: j J} be an injective family of all
straight lines in R
2
. Fix any family {n
j
: j J} of natural numbers such that n
j
2 for
each index j J. By applying the method of transnite induction, show that there exists a
set Y R
2
such that
(j J)(card(l
j
Y) = n
j
).
In particular, if n
j
= 2 for all j J, then Y is a Mazurkiewicz subset of R
2
.
3. Give a proof of Lemma 1 by using the Fubini theorem. Moreover, generalize this
lemma to the case of analytic submanifolds {M
i
: i I} of a given analytic manifold M
equipped with an appropriate analogue of the Lebesgue measure, where card(I) < c and
dim(M
i
) < dim(M) for all i I.
4
. Prove the following statement due to Sierpi nski.

If n 3, then there exists a free group G of rotations of the space R
n
about its origin, such
that card(G) = c.
For this purpose, keep in mind Lemma 1 and construct the required G by means of the
method of transnite recursion.
5
. As known, there is a model of set theory (e.g., Cohens original model) in which a
certain subset of R having cardinality strictly less than c is of strictly positive outer
1
-
measure and hence is not measurable with respect to
1
. Show that in such a model there
exists a covering {X
i
: i I} of R by nowhere dense compact subsets X
i
(i I), for which
card(I) < c.
Deduce fromthis fact that in the same model there is a covering {Z
i
: i I} of R
2
by Jordan
curves Z
i
(i I), where again card(I) < c.
Recall that a Jordan curve is usually dened as a homeomorphic image of the one-
dimensional unit torus S
1
(= T).
6
. Infer the main result of the previous exercise from the following set-theoretical hypoth-
esis:
c is measurable in the Ulam sense.
In other words, prove that if c is a real-valued measurable cardinal, then there exists a
covering {X
i
: i I} of R such that card(I) <c and all X
i
(i I) are nowhere dense compact
sets in R.
7
. Show that if a set Z R

2
is nite in a direction e R
2
, then Z is necessarily R
2
-
negligible. For the denition of negligible sets, see Exercise 9 of Chapter 5.
Derive from this fact that there exists no subset Z of R
2
such that:
(a) Z is nite in a xed direction e R
2
;
(b) (R
2
\ Z) = 0 for some nonzero -nite translation-invariant measure on R
2
.
8. Give a proof of the statement formulated in Remark 2. In other words, showthe existence
of a subset X of R
n
satisfying conditions 1), 2) and 3) formulated therein.
9. Let us identify the plane R
2
with the eld C of all complex numbers and consider two
mappings
: C C, : C C
dened as follows:
(z) = z +1, (z) = e
i
z (z C),
where, as usual,
i
2
= 1, e
i
= cos(1) +isin(1).
Obviously, is a translation of R
2
and is a rotation of R
2
about its origin (0, 0).
Denote by G the semigroup of motions of R
2
generated by {, } and put
X = G((0, 0)).
Taking into account the fact that e
i
is a transcendental number, show that
(X) (X) = / 0, (X) (X) = X.
In other words, the countable set X R
2
admits a decomposition (i.e., partition) into two
subsets such that each of them is G-congruent to X.
This result is due to Mazurkiewicz and Sierpi nski (cf. [240]) and does not need any form
of the Axiom of Choice.
Show also that if Y is a nonempty bounded subset of R
2
, then there exists no partition
{Y
1
,Y
2
} of Y such that the sets Y, Y
1
and Y
2
are pairwise congruent.
10
. Let T
2
denote the group of all translations of R
2
and let g be a rotation of R
2
such
that all g
n
(1 n < ) are distinct from the identical transformation of R
2
. Denote by G
the group of transformations generated by T
2
{g}. Prove that there exists a set X R
2
(a) X is T
2
-negligible;
(b) X is G-absolutely nonmeasurable.
In connection with the result presented in Exercise 10, let us point out that it is unknown
whether there exists a set Y R
2
which is T
2
-absolutely negligible and, simultaneously,
M
2
-absolutely nonmeasurable where M
2
stands for the group of all isometric transforma-
tions (i.e., motions) of R
2
.
11. Consider in detail the cases 1) - 3) of Remark 3 and, for each of them, construct the
required polygonal convex curve P with innitely many sides.
12. Generalize Theorem 4 to the case of the Euclidean space R
n
, where n 2. In other
words, prove that there exists a set X R
n
(a) card(X) = c and all points of X are in general position;
(b) no subset Y of X with card(Y) = c is convexly independent.
Conclude from this fact that, under the Continuum Hypothesis, there exists an uncountable
set X R
n
of points in general position such that no uncountable subset of X is convexly
independent.
13. Let k 3 be a natural number. By using the method of transnite recursion, demon-
strate that there exists a set Y R
2
(a) Y is
2
-thick in R
2
;
(b) every circumference in R
2
has exactly k common points with Y.
Condition (b) shows that Y may be regarded as a certain analogue of the Mazurkiewicz set
for the family of all circumferences in the plane.
14
. Let k > 0 be a natural number and let {X

j
: j J} be a family of subsets of R
2
. We
shall say that this family is a k-homogeneous covering of R
2
if every point in R
2
belongs
exactly to k sets from the family.
Show that there exists no 1-homogeneous covering of R
2
by Jordan curves. On the other
hand, use the method of transnite recursion and prove that for k 2, there exists a k-
homogeneous covering of R
2
by pairwise congruent circumferences.
Moreover, for any even k > 0, give a direct construction (within ZF theory) of a k-
homogeneous covering of R
2
by pairwise congruent circumferences.
Taking into account the result of Exercise 13, conclude that the following two dual state-
ments are equivalent within ZF theory and are valid within ZFC theory:
(a) there exists a subset X of R
2
such that every circumference of radius 1 in R
2
meets X
in precisely three points;
(b) there exists a family of circumferences of radius 1 in R
2
such that every point of R
2
belongs to exactly three circumferences of the family.
Finally, give an example of a set X R
2
which has the property formulated in (a) but
simultaneously is such that card(CX) = 4 for some circumference C R
2
.
15
. Let Z be a subset of R
2
satisfying the following two conditions:
(a) every straight line in R
2
meets Z in at least two points;
(b) Z contains a simple arc, that is Z contains a set homeomorphic to [0, 1].
Demonstrate that there exists a straight line l R
2
such that
card(l Z) 3.
Conclude from this fact that no Mazurkiewicz set (see Exercise 1) contains a simple arc.
Show also that no Mazurkiewicz set can be an F
-subset of R
2
.
For this purpose, suppose otherwise and infer that some Mazurkiewicz set must contain a
simple arc which contradicts the said above.
16
. Let Y be a subset of R
2
. We shall say that Y is an ot-set if for any three pairwise
distinct points y
1
, y
2
, y
3
fromY, the triangle [y
1
, y
2
, y
3
] is non-degenerate and has an obtuse
angle.
Demonstrate the validity of the following assertions:
(a) every ot-set is contained in a maximal (with respect to inclusion) ot-set;
(b) no nite ot-set is maximal;
(c) there exists a countable maximal ot-set.
In addition to these facts, verify that
{(cos(t), sin(t)) : 0 t < }
is a maximal ot-set in R
2
whose cardinality is equal to c.
17
. Let f : R
2
R
2
be a partial homeomorphism such that both sets dom( f ) and ran( f )
are homeomorphic to the Cantor space {0, 1}
.
Prove that there exists a homeomorphism f
: R
2
R
2
extending f (cf. Exercise 20 from
Chapter 8).
It should be noticed that the result just formulated cannot be extended to those partial
homeomorphisms f : R
3
R
3
whose domains and ranges are homeomorphic to the Can-
tor space. In order to verify this circumstance, we need to apply some properties of the
so-called Antoine necklace (see, for instance, [19] where this necklace is thoroughly con-
sidered in connection with a wild topological embedding of the sphere S
2
into R
3
).
18
. Within ZF theory, show that there are two subsets X and Y of R

2
satisfying the fol-
lowing relations:
(a) card(X) = card(Y) = c;
(b) for any two isometric transformations f and g of R
2
, the set f (X) g(Y) contains
exactly two points.
Formulate and prove an analogous result replacing 2 by an even natural number 2k > 2.
Applying the method of transnite recursion, construct subsets X
and Y
of R
2
such that:
(a) card(X
) = card(Y
) = c;
(b) for any two isometric transformations f and g of R
2
, the set f (X
)g(Y
) is a singleton.
It should be mentioned that the existence of X
and Y
satisfying both relations (a) and (b)

was rst established by Sierpi nski.
19
. Let E be a set and let {X

i
: i I} be a family of subsets of E.
Recall that {X
i
: i I} is said to be independent (in the set-theoretical sense) if for each
nite subset J of I and for every function f : J {0, 1}, we have
{X
f ( j)
j
: j J} = / 0
where X
f ( j)
j
= X
j
if f ( j) = 0, and X
f ( j)
j
= E \ X
j
if f ( j) = 1.
It is well known that for any innite set E, there exists an independent family {X
i
: i I}
of subsets of E such that card(I) = 2
card(E)
(in this connection, see [150]). The fact just
indicated easily implies that the family of all nontrivial ultralters in E has cardinality
2
2
card(E)
.
Suppose now that E = R
2
.
Construct an innite independent family of convex polygons in R
2
.
On the other hand, show that there exists no uncountable independent family of simple
polygons in E.
20
. We say that a compact set P R

2
is a quasi-polygon in R
2
if P has nonempty interior
and the boundary of P can be represented as the union of countably many line segments.
Observe that any polygon in R
2
is a quasi-polygon and, obviously, the converse assertion
fails to be true.
Prove that there exists an uncountable independent family of convex quasi-polygons in R
2
.
Some other results concerning combinatorial and descriptive properties of subsets of the
Euclidean space R
n
(in particular, of the plane R
2
) can be found in [2], [48], [49], [51],
[54], [55], [76], [77], [105], [156], [169], [222], and [223].
Chapter 20
Restrictions of real-valued functions
This last chapter may be regarded as dual to Chapter 1 in which various extensions of
partial functions were considered.
In 1897 Poincar e wrote that on the intuitive level every function f : RR is closely related
to some continuous function (see [200]). This phrase can be expressed in the following
form: certain (of course, nontrivial) restrictions of f turn out to be continuous on their
domains.
Here we intend to consider those restrictions of real-valued functions which have rather
good descriptive properties. Let us formulate the main problem which will be touched in
the present chapter.
Let f : RR be a function. Does there exist a large (or, at least, non-small) subset X of R
such that the restriction f [X possesses nice structural properties?
For instance, we may require that f [X should be continuous (continuous at some points of
X) or should belong to a given Baire class or should be measurable in the Lebesgue sense
or should be with the Baire property (with the Baire property in the restricted sense).
We will give only several results in this direction which seem to be interesting from the
point of view of real analysis. To illustrate very briey the said above, let us begin with the
simplest fact of this type.
Let g : R R be a function. It is natural to ask whether there exists a set X R with
card(X) = card(R) = c
such that the restriction g[X is bounded. The answer to this question is trivially positive.
Indeed, consider the sets
X
n
= g
1
([n, n]) (n <).
Then, in view of the equality R = X
n
: n < , at least one set X
n
0
is of cardinality
continuum. Putting X = X
n
0
, we see that
(x X)([ f (x)[ n)
313
20
and so we come to the required result.
Evidently, our argument was based on the relation c f (c) > which can be proved only
with the aid of the Axiom of Choice. In a more general situation, if we have a set E with
card(E) , then we may assert that the following two statements are equivalent:
1) c f (card(E)) >;
2) for any function f : E R, there exists a set X E such that card(X) = card(E) and
the restriction f [X is bounded.
We thus see that the simple question posed above which, at rst sight, is of purely analytic
avor turns out to be closely connected with set-theoretical characteristics of the contin-
uum. In the sequel, several other (less trivial) examples of this kind will be pointed out and
discussed in detail.
Now, let us recall the classical theorem of Blumberg [15] which is rather surprising be-
cause no assumptions on functions under consideration are required in its formulation (cf.
Poincar es prediction at the beginning of this chapter).
Theorem 1. Let f : R R be an arbitrary function. Then there exists a subset D of R
(a) D is everywhere dense in R;
(b) the restricted function f [D is continuous.
The proof of this theorem was sketched in Exercises 21 and 22 for Chapter 8. There are
many works devoted to various versions of Blumbergs theorem (see [8], [27], [28], [29],
[89], [206], and [241]). In view of the existence of Sierpi nski-Zygmund type functions
(see Chapters 1, 13 and 14), it is impossible to assert that such a D can be chosen to be
uncountable. So the class of all functions acting from R into R should be replaced by a
narrower class if we want to get good restrictions of functions to some non-small (e.g., in
the sense of cardinality) subsets of R.
It seems natural to treat all nonempty perfect sets in R as non-small subsets of R. We now
turn our attention to those real-valued functions whose restrictions to nonempty perfect
subsets of R are not totally discontinuous.
Let E be a topological space and let f : E R be a function. We shall say that f is of Baire
zero class if f is continuous at all points of E, i.e., f is continuous on E.
The family of all continuous functions acting from E into R is usually denoted by the
symbol C(E, R). In accordance with the denition above, the notation Ba
0
(E, R) will be
used for the same family of functions. Thus, we have
Ba
0
(E, R) =C(E, R).
Restrictions of real-valued functions 315
The next important family of functions is the class of all Baire 1 functions.
By the standard denition due to Baire (see [3], [4], and [5]), a function
f : E R
is of Baire 1 class (or of rst Baire class) if there exists a sequence f
n
: n < of functions
from Ba
0
(E, R) such that lim
n+
f
n
(x) = f (x) for each point x E. In other words,
f : E R belongs to the rst Baire class if and only if f can be represented as the pointwise
limit of a sequence of functions belonging to Ba
0
(E, R).
It is well known that the functions of rst Baire class play a signicant role in various
topics of real analysis. The following simple but important example emphasizes this cir-
cumstance.
Example 1. Let E = R and let f : E R be a derivative. This assumption means that
there exists a continuous function g : E R such that g
/
(x) = f (x) for all x R. Dene a
sequence g
n
: 1 n < of real-valued functions on R by the formula
g
n
(x) = n(g(x +1/n) g(x)) (x R, n = 1, 2, ...).
Obviously, we have
lim
n+
g
n
(x) = f (x) (x R).
Since each g
n
(1 n < ) is a continuous function on R, we conclude that f belongs to
the rst Baire class.
For any topological space E, the family of all functions f : E R belonging to the rst
Baire class will be denoted by the symbol Ba
1
(E, R).
Notice that the other Baire classes Ba
(E, R) of real-valued functions on E are naturally

introduced by iterating the limit process and using the method of transnite recursion on
<
1
(in this connection, see [5], [99], [148], [162], [183], and Appendix 6).
Here our main interest is concentrated on functions from the class Ba
1
(E, R). The follow-
ing simple properties are easily implied by the denition of Ba
1
(E, R).
(i) Ba
1
(E, R) is a linear algebra over the eld R; in other words, if we have f Ba
1
(E, R),
g Ba
1
(E, R), a R and b R, then
a f +bg Ba
1
(E, R), f g Ba
1
(E, R).
(ii) If f Ba
1
(E, R), g Ba
1
(E, R) and g(x) ,= 0 for all x E, then
f /g Ba
1
(E, R).
(iii) If f Ba
1
(E, R) and : ]a, b[ R is a continuous function such that ran( f ) ]a, b[,
then
f Ba
1
(E, R).
Below, we will consider some other properties of the class Ba
1
(E, R), which are not so
trivial.
Lemma 1. Let E be a topological space, a
n
: 1 n < be a sequence of strictly positive
real numbers such that
1n<
a
n
< +,
and let f
n
: 1 n < Ba
1
(E, R) be a sequence of functions such that
[ f
n
(x)[ < a
n
(x E, 1 n <).
Dene a function f : E R by the formula
f (x) = f
1
(x) + f
2
(x) +... + f
n
(x) +... (x E).
Then f also belongs to Ba
1
(E, R).
Proof. First of all, let us notice that the function f is well dened since the series
f
1
(x) + f
2
(x) +... + f
n
(x) +... (x E)
converges uniformly with respect to x E (in view of the relations [ f
n
(x)[ < a
n
for any
x E and n 1). Further, since f
n
Ba
1
(E, R) for each natural number n 1, we can
write
f
n
(x) = lim
k+
f
n,k
(x) (x E),
where f
n,k
(k = 1, 2, ...) are some real-valued continuous functions on E. Without loss of
generality, we may assume that
[ f
n,k
(x)[ a
n
(x E, n = 1, 2, ..., k = 1, 2, ...).
Now, let us put
h
k
(x) = f
1,k
(x) + f
2,k
(x) +... + f
k,k
(x) (x E, k = 1, 2, ...).
Clearly, the functions h
k
are continuous on E. It sufces to show that
f (x) = lim
k+
h
k
(x) (x E).
For this purpose, take an arbitrary real number >0. There exists a natural number m such
that
a
m+1
+a
m+2
+... +a
i
+... </3.
Consequently, we have
[ f
m+1
(x)[ +[ f
m+2
(x)[ +... +[ f
i
(x)[ +... </3 (x E),
[ f
m+1,k
(x)[ +[ f
m+2,k
(x)[ +... +[ f
i,k
(x)[ +... /3 (x E, k = 1, 2, ...).
For any x E and any k > m, we can write the inequalities
[ f (x) h
k
(x)[ [ f
1
(x) f
1,k
(x)[ +[ f
2
(x) f
2,k
(x)[ +... +[ f
m
(x) f
m,k
(x)[+
[ f
m+1
(x)[ +[ f
m+2
(x)[ +... +[ f
i
(x)[ +... +
[ f
m+1,k
(x)[ +[ f
m+2,k
(x)[ +... +[ f
k,k
(x)[
[ f
1
(x) f
1,k
(x)[ +[ f
2
(x) f
2,k
(x)[ +... +[ f
m
(x) f
m,k
(x)[ +2/3.
If x E is xed, then we can nd k
0
< so large that for all natural numbers k > k
0
, the
relation
[ f
1
(x) f
1,k
(x)[ +[ f
2
(x) f
2,k
(x)[ +... +[ f
m
(x) f
m,k
(x)[ </3
will be satised. But this relation immediately yields the inequality
[ f (x) h
k
(x)[ <
for every integer k > k
0
. Therefore, we get
lim
k+
h
k
(x) = f (x) (x E),
We need this lemma in order to prove the following result due to Baire (cf. [5], [99], [148],
[162], and [183]).
Theorem 2. Let E be a topological space. Suppose that a sequence
f
n
: n < Ba
1
(E, R)
is given, uniformly converging to a function f : E R. Then f also belongs to the class
Ba
1
(E, R).
Proof. According to our assumption, for any natural number k, there exists a natural num-
ber n
k
such that
[ f (x) f
n
k
(x)[ <
1
2
k+1
(x E).
Evidently, we may assume that n
0
< n
1
< ... < n
k
< ... . Let us consider the series of
functions
( f
n
1
f
n
0
) +( f
n
2
f
n
1
) +... +( f
n
k+1
f
n
k
) +... .
Since the inequalities
[ f
n
k+1
(x) f
n
k
(x)[ [ f
n
k+1
(x) f (x)[ +[ f
n
k
(x) f (x)[ <
1
2
k+2
+
1
2
k+1
<
1
2
k
hold true for all x E, we can apply Lemma 1 to this series. So we obtain that the function
g = ( f
n
1
f
n
0
) +( f
n
2
f
n
1
) +... +( f
n
k+1
f
n
k
) +...
belongs to Ba
1
(E, R). But it is easy to see that
g = lim
k+
f
n
k+1
f
n
0
= f f
n
0
.
According to property (i), we nally get
f = g + f
n
0
Ba
1
(E, R).
The theorem has thus been proved.
Remark 1. Theorem 2 implies that for E ,= / 0, the family of all bounded functions from
B
1
(E, R) becomes a Banach space with respect to the norm of uniform convergence, i.e.,
with respect to the standard sup-norm
[[ f [[ = sup
xE
[ f (x)[.
Remark 2. Theorem2 can be directly generalized to the case of the Baire class Ba
(E, R),
where is an arbitrary ordinal number strictly less than
1
. The proof essentially remains
the same as above (cf. [148]).
Lemma 2. Let E be a topological space and let g Ba
1
(E, R). Then for every t R, the
pre-images
g
1
(] , t[) =x E : g(x) <t, g
1
(]t, +[) =x E : g(x) >t
are F
-subsets of E.
Proof. Take any t R. Because of g Ba
1
(E, R), there exists a sequence g
n
: n <
Ba
0
(E, R) such that
lim
n+
g
n
(x) = g(x) (x E).
It is not difcult to verify the following two relations:
g(x) <t (k <)(n <)(m [n, [)(g
m
(x) t 1/k),
g(x) >t (k <)(n <)(m [n, [)(g
m
(x) t +1/k).
These relations yield at once that the above-mentioned sets
x E : g(x) <t, x E : g(x) >t
are of type F
in E, and the lemma is proved.

Lemma 3. Let E be a normal topological space, g : E R be a function and suppose that
ran(g) =t
1
, t
2
, ...,t
k
.
If for any integer i [1, k], the set
X
i
=x E : g(x) =t
i
= g
1
(t
i
)
is an F
-subset of E, then g Ba
1
(E, R).
Proof. Obviously, we can write
E = X
1
X
2
... X
k
,
X
i
= F
i,0
F
i,1
... F
i,n
... (i = 1, 2, ..., k),
where all F
i,n
(1 i k, n <) are closed subsets of E and
F
i,0
F
i,1
... F
i,n
... .
Let us put
F
n
= F
1,n
F
2,n
... F
k,n
(n = 0, 1, 2, ...)
and dene a function
g
n
: F
n
R (n = 0, 1, 2, ...)
by the formula:
g
n
(x) =t
i
if and only if x F
i,n
.
Because the nite family of closed sets F
1,n
, F
2,n
, ..., F
k,n
is disjoint, the function g
n
is
continuous on the closed set F
n
. By the Tietze-Urysohn theorem, g
n
admits a continuous
extension g
n
: E R. Now, it is easy to check that
lim
n+
g
n
(x) = g(x)
for all points x E and, therefore, g Ba
1
(E, R) which nishes the proof of the lemma.
Lemma 4. Let E be a topological space in which every open set is of type F
(or, equiv-
alently, in which every closed set is of type G
). Let a set X E be representable in the

form
X = A
1
A
2
... A
k
,
where all A
j
( j = 1, 2, ..., k) are F
-subsets of E. Then X is representable in the form

X = B
1
B
2
... B
k
,
where B
j
A
j
( j = 1, 2, ..., k), all sets B
j
are also of type F
and, in addition, they are

pairwise disjoint.
Proof. Obviously, we have the equality
X = F
1
F
2
... F
i
...,
where all sets F
i
(1 i <) are closed in E and each F
i
is contained in some set A
j(i)
. Let
us put
C
1
= F
1
, C
2
= F
2
F
1
, ..., C
i
= F
i
(F
1
... F
i1
), ... .
The family of sets C
i
: 1 i < is disjoint and in view of our assumption on E, all C
i
are of type F
in E. Moreover, we have
X =C
1
C
2
... C
i
... .
Now, for any natural number j [1, k], dene the set B
j
by the formula
B
j
=C
i
: j is the least number such that C
i
A
j
.
Clearly, the family B
1
, B
2
, ..., B
k
is disjoint, all sets B
j
( j =1, 2, ..., k) are of type F
in E
and
B
j
A
j
( j = 1, 2, ..., k),
X = B
1
B
2
... B
k
.
Lemma 4 has thus been proved.
Recall that a topological space E is perfectly normal if E is normal and each open set in
E is of type F
. For such spaces, the next important statement due to Lebesgue is true (cf.
[5], [148], and [183]).
Theorem 3. Let E be a perfectly normal space and let f : E R be a function. The
following three assertions are equivalent:
(1) f Ba
1
(E, R);
(2) for any t R, both sets x E : f (x) <t and x E : f (x) >t are of type F
in E;
(3) for any open set U R, the pre-image f
1
(U) is an F
-subset of E.
Proof. The equivalence (2) (3) is trivial. The implication (1) (2) was established by
Lemma 2 even for an arbitrary topological space E. Consequently, it remains to prove the
implication (2) (1). Suppose that (2) is valid and suppose rst that ran( f ) ]0, 1[ or,
equivalently,
0 < f (x) < 1 (x E).
For any integer n 1, consider the sequence t
0
, t
1
, ...,t
n
of points of R determined by the
conditions
t
0
= 0, t
j+1
t
j
= 1/n ( j = 0, ..., n 1).
In fact, we have t
j
= j/n for j = 0, ..., n. Further, introduce the following sets:
A
0
=x E : f (x) <t
1
,
A
n
=x E : f (x) >t
n1
,
A
j
=x E : t
j1
< f (x) <t
j+1
( j = 1, ..., n 1).
Obviously, we have the equality
E = A
0
A
1
... A
n
and all A
j
( j = 0, 1, ..., n) are F
-subsets of E. Applying Lemma 4, we get another repre-

sentation
E = B
0
B
1
... B
n
,
where all B
j
( j = 0, 1, ..., n) are also F
-subsets of E, are pairwise disjoint, and

B
j
A
j
( j = 0, 1, ..., n).
Further, dene a function f
n
: E R by putting: f
n
(x) =t
j
if and only if x B
j
. According
to Lemma 3, the function f
n
belongs to Ba
1
(E, R).
Now, take an arbitrary x E. Then x B
j
for some integer j [0, n].
If j = 0, then we have
t
0
< f (x) <t
1
, f
n
(x) =t
0
, [ f (x) f
n
(x)[ < 1/n.
If j = n, then we have
t
n1
< f (x) <t
n
, f
n
(x) =t
n
, [ f (x) f
n
(x)[ < 1/n.
Finally, if 1 j n 1, then
t
j1
< f (x) <t
j+1
, f
n
(x) =t
j
, [ f (x) f
n
(x)[ < 2/n.
All these relations show that lim
n+
f
n
(x) = f (x) uniformly with respect to x E. By
virtue of Theorem 2, we obtain that f Ba
1
(E, R).
Suppose now that f : E R is an arbitrary function satisfying (2). Fix any increasing
homeomorphism : R ]0, 1[ and consider the function f . This function also satises
(2) and
ran( f ) ]0, 1[.
As demonstrated above, f Ba
1
(E, R). Consequently, in view of (iii),
f =
1
( f ) Ba
1
(E, R)
which completes the proof of Theorem 3.
Example 2. Let E be a perfectly normal space and let X be a subset of E. Denote by
X
the characteristic function of X. It is easy to check that if X is closed in E, then relation
(2) of Theorem 2 is satised for f =
X
. Therefore, according to this theorem, we have
X
Ba
1
(E, R). Now, if Y is an open subset of E, then taking into account the equality
Y
= 1
EY
,
we see that
Y
Ba
1
(E, R), too (by virtue of property (i)).
The above-mentioned facts also directly follow from Lemma 3.
Example 3. Let E be a subinterval of R and let f : E R be a monotone function. It is
easy to verify that for any t R, both sets
x E : f (x) <t, x E : f (x) >t
are some subintervals of E. Because each interval in E is an F
-subset of E, we claim (in

view of Theorem 3) that f Ba
1
(E, R).
The same conclusion is true for those f : E R which are of nite variation on E. Indeed,
such functions are representable in the form of the difference of two increasing functions
on E (see, e.g., [183] or [210]) and it sufces to apply property (i) of the class Ba
1
(E, R).
Example 4. Let E be again a perfectly normal space and let f : E R be an upper
semicontinuous function. According to the denition of upper semicontinuous functions,
for any t R, the set x E : f (x) <t is open in E and hence is an F
-subset of E. At the
same time, the set
x E : f (x) >t =
n<
x E : f (x) t +1/(n +1)
is the union of countably many closed sets, i.e., is also an F
-subset of E. Using Theorem

3, we deduce that f Ba
1
(E, R).
From this fact it immediately follows again by virtue of property (i) that g Ba
1
(E, R) for
any lower semicontinuous function g : E R.
Actually, the characteristic function
X
of a closed set X E (see Example 2) is upper
semicontinuous.
Recall that a topological space E is Baire if no nonempty open subset of E is of rst
category in E. For such an E, an important result due to Baire is well known which yields
an essential information about the structure of the set D( f ) of all discontinuity points of an
arbitrary function f Ba
1
(E, R).
Theorem 4. Let E be a Baire space and let f Ba
1
(E, R). Then the set D( f ) of all
discontinuity points of f is of rst category in E. Consequently, we have
C( f ) U ,= / 0
for any nonempty open set U E (where C( f ) as usual denotes the set of all continuity
points of f ).
Proof. For each point x E, let us put
f
(x) = infdiam( f (V)) : V V (x),
where V (x) is the lter of all neighborhoods of x and diam( f (V)) denotes the diameter of
f (V). The real number
f
(x) 0 is usually called the oscillation of f at x (cf. Exercise 13
from Chapter 8).
As is well known (see Exercise 3 for this chapter), the equality
C( f ) =
1n<
x E :
f
(x) < 1/n
is valid, where all sets x E :
f
(x) < 1/n are open in E. So, it sufces to demonstrate
that all these sets are everywhere dense in E. In other words, it sufces to show that for any
real number > 0 and for any nonempty open set U E, there exists a nonempty open set
W U such that
(x W)(y W)([ f (x) f (y)[ <).
Taking into account the fact that f belongs to Ba
1
(E, R), choose a sequence f
k
: k <
Ba
0
(E, R) satisfying the relation
f (x) = lim
k+
f
k
(x) (x E).
Further, for any natural number k, dene the set
X
k
=x E : (i k)(j k)([ f
i
(x) f
j
(x)[ /3).
Obviously, all sets X
k
(k <) are closed in E and
(k <)(X
k
X
k+1
), E =X
k
: k <.
Consequently, we have
U = (U X
0
) (U X
1
) ... (U X
k
) ... .
Since E is a Baire space, there exists a natural number n such that
int(U X
n
) ,= / 0.
Let V U X
n
be a nonempty open set. If x is an arbitrary point of V, then
(i n)(j n)([ f
i
(x) f
j
(x)[ /3).
Putting j = n and tending i to +, we get
(x V)([ f (x) f
n
(x)[ /3).
Therefore, we can write
[ f (y) f (x)[ [ f (y) f
n
(y)[ +[ f
n
(y) f
n
(x)[ +[ f
n
(x) f (x)[
2/3 +[ f
n
(y) f
n
(x)[
for any two points x and y fromV. Finally, because f
n
is a continuous function, there exists
a nonempty open set W V such that
(x W)(y W)([ f
n
(y) f
n
(x)[ </3).
This relation implies at once that
(x W)(y W)([ f (y) f (x)[ <)
which ends the proof of the Baire theorem.
Remark 3. The proof presented above is fairly standard and, actually, is based on the
classical argument originally due to Baire (cf., for instance, [183] and [192]). More general
versions of the Baire theorem with further information about it can be found in [58] and
[148]. In fact, the main part of the Baire theorem remains true for an arbitrary topological
space E (see Exercise 6 of this chapter).
The next classical result is widely known as the Cantor-Baire stationarity principle (see
[58], [148], and [183]).
Theorem 5. Let E be a hereditarily Lindel of topological space and let
F
0
F
1
... F
... ( <
1
)
be a decreasing (with respect to the inclusion relation)
1
-sequence of closed subsets of E.
Then there exists an ordinal <
1
such that
( [,
1
[)(F
= F
).
Proof. For each ordinal <
1
, let us denote G
= E F
and put
G =G
: <
1
.
Then the family G
: <
1
is an open covering of an open set G E. Since E is
hereditarily Lindel of, there exists a countable set [0,
1
[ such that
G =G
: .
In view of the regularity of
1
, the set is bounded from above in [0,
1
[, i.e., there exists
an ordinal <
1
for which we have sup() < . Now, it is easy to verify that F
= F
for all ordinals [,

1
[.
Remark 4. A certain analogue of the Cantor-Baire stationarity principle was obtained by
Luzin for decreasing
1
-sequences of F
-subsets of R and this analogue was successfully

applied to some deep problems of descriptive set theory (in this connection, see [160],
[162], and [188]).
The next example yields a good illustration of the usefulness of the Cantor-Baire stationar-
ity principle.
Example 5. Suppose again that E is a hereditarily Lindel of topological space. Take an
arbitrary closed subset X of E and dene by transnite recursion an
1
-sequence X
: <
1
in the following manner:
(a) X
0
= X;
(b) X
+1
= (X
)
/
for any <
1
, where (X
)
/
denotes the set of all accumulation points of
X
;
(c) X
=X
: < for any limit ordinal <

1
.
Applying the Cantor-Baire stationarity principle to X
: <
1
, we infer that there exists
a smallest ordinal
0
<
1
for which we have
X
= X
0
(
0
<
1
).
Since E is hereditarily Lindel of, all sets X
X
+1
( <
1
) are at most countable. Taking
into account the equality
X = (X
X
+1
: <
0
) X
0
,
we conclude that X admits a representation in the form
X =Y Z (Y Z = / 0),
where Y = X
0
is a perfect subset of E and Z is at most countable.
Recall that this representation was rst obtained by Cantor and Bendixson (the Cantor-
Bendixson theorem).
Remark 5. According to the classical Baire theorem (see [5], [148], and [183]), for any
function g : R R, the following two assertions are equivalent:
(1) g is of rst Baire class;
(2) for every nonempty closed set F R, there are points x F at which the function g[F
is continuous.
In other words, (2) yields a characterization of all functions acting from R into R and
belonging to the rst Baire class. This characterization is given in terms of certain restric-
tions of functions. Unfortunately, we do not have an analogous nice characterization of the
derivatives on R, which form an important proper subclass of Ba
1
(R, R) (see Example 1).
Notice that the proof of the equivalence (1) (2) is outlined in Exercise 9 for this chapter.
The Cantor-Baire stationarity principle is heavily exploited in that proof.
According to the denition introduced by Luzin, a topological space X is always of rst
category (or X is perfectly meager) if each nonempty dense in itself subset of X is a rst
category space.
Let E be a topological space and let X be a subspace of E such that, for every perfect set
P E, the set X P is of rst category in P. Then it is not difcult to show that the space
X is always of rst category (see, e.g., [148]).
Luzin proved that there exists an uncountable subspace X of R which is always of rst
category (see [148] or [162]). Other constructions of uncountable universally small sets
can be found in [197] and [250] (cf. also Appendices 1 and 6).
Here we wish to give one construction of such a kind. It needs Martins Axiom because
a perfectly meager set obtained by means of this construction possesses deep additional
properties. As usual, we denote by the standard Lebesgue measure on the real line R.
Theorem 6. Suppose that Martins Axiom holds. Then there exists a set X R satisfying
the following conditions:
(1) for every nonempty perfect set P R, the intersection X P is of rst category in P;
(2) for each Lebesgue measurable set Y R with (Y) > 0, the intersection X Y is
nonempty;
(3) X is a generalized Sierpi nski subset of R.
Proof. Let denote the least ordinal number whose cardinality is equal to c (actually, we
may identify with c).
Let Z
: < denote the family of all those Borel subsets of R which have strictly
positive Lebesgue measure, i.e.,
Z
: < =B(R) I(),

and let T
: < denote the family of all those Borel subsets of R which have Lebesgue
measure zero, i.e.,
T
: < =B(R) I().

For any ordinal <, we x a partition Z
0
, Z
1
of Z
such that Z
0
is a Lebesgue measure
zero set and Z
1
is a rst category subset of Z
. Notice that the existence of such a partition

follows directly from Lemma 2 of Appendix 3. Now, we dene an injective -sequence
x
: < of real numbers. Suppose that < and that the partial -sequence x
:
< has already been determined. Let us consider the set
D
= (Z
0
: ) x
: < (T
: ).
Martins Axiom implies that this D
is also of Lebesgue measure zero. Hence we have

Z
,= / 0. So we can choose a point x
from the above-mentioned nonempty difference

of two sets. In this way, we are able to dene the entire -sequence x
: < of points
of R. Now, put
X =x
: <.
We are going to show that X is the required set. Let P be any nonempty perfect subset of
R. If (P) = 0, then for some ordinal < , we may write P = T
. Consequently, from
the method of construction of the set X, we immediately obtain
card(PX) < c.
Applying Martins Axiom once more, we see that the intersection PX is of rst category
in P. Suppose now that (P) > 0. Then for some ordinal < , we can write P = Z
.
Therefore,
PX Z
1
: <.
Taking account of the fact that the set P does not have isolated points, we obtain from the
last inclusion that PX is again of rst category in P. Hence condition (1) is satised for
our set X. Furthermore, since x
for each ordinal < , we conclude that condi-

tion (2) holds for the set X, too. The validity of condition (3) follows directly from our
construction. Theorem 6 has thus been proved.
Lemma 5. Let E be a hereditarily Lindel of T
1
-space always of rst category and let f :
E R be a function. Then for each subspace X of E, the set D( f [X) is of rst category in
X.
Proof. The argument is very easy, so we only sketch it. Namely, we begin with establishing
the fact that X admits a representation in the form
X =Y Z,
where Y is dense in itself, Z is at most countable, andY Z = / 0 (cf. Example 5 and Exercise
8). Then we verify the inclusion
D( f [X) Y (X
/
Z),
where X
/
denotes the set of all accumulation points of X (in E). Finally, observe that both
sets Y and X
/
Z are of rst category in X which completes the proof.
Theorem 7. Let E be a subspace of R satisfying the following relations:
(a) card(E) = c;
(b) E is always of rst category.
Then there exists a function f : E R having the following properties:
(c) f is not Borel (consequently, f does not belong to the class Ba
1
(E, R));
(d) for any subspace Z of E, the set D( f [Z) is of rst category in Z.
Proof. Since the equality card(E) = c holds, the family B(E, R) of all Borel mappings
acting from E into R is also of cardinality continuum. But the family of all mappings
acting from E into R is of cardinality 2
c
. So we can choose a function f : E R which
is not Borel on E, i.e., such an f has property (c). In view of Lemma 5, the same f has
property (d) which ends the proof.
Remark 6. Recall that the existence of a subspace E of R satisfying the relations (a) and (b)
was demonstrated by using Martins Axiom (see Theorem 6 of this chapter). The classical
result formulated in Theorem7 is due to Luzin (see [148] and [162]). In particular, it shows
that property (d) of f does not imply the relation f Ba
1
(E, R).
Theorem 8. Let f : R R be a Lebesgue measurable function (or let f : R R have the
Baire property). Then there exists a nonempty perfect set P R such that the restriction
f [P is monotone on P.
Proof. The argument given below is fairly standard and is usually applied in many similar
situations (cf. [122]).
As is well known, there exists a nonempty perfect subset T of R such that the restriction
f [T is continuous. This fact is true for all Lebesgue measurable functions and for all those
functions which possess the Baire property. Of course, we may assume that diam(T) < 1
where the symbol diam(T) stands for the diameter of T. Let us denote g = f [T and suppose
that for every nonempty perfect set Q T, the restriction g[Q is not decreasing. Then by
using ordinary induction, we can construct a dyadic system
(T
i
1
i
2
...i
k
)
i
1
0,1,i
2
0,1,...,i
k
0,1
(k <)
of nonempty perfect subsets of T satisfying the following conditions:
(a) T
/ 0
= T;
(b) T
i
1
i
2
...i
k
i
k+1
T
i
1
i
2
...i
k
;
(c) T
i
1
i
2
...i
k
0
T
i
1
i
2
...i
k
1
= / 0;
(d) diam(T
i
1
i
2
...i
k
) < 1/2
k
;
(e) if (i
1
, i
2
, ..., i
k
) ( j
1
, j
2
, ..., j
k
), then x < y and g(x) < g(y) for all points x T
i
1
i
2
...i
k
and y T
j
1
j
2
... j
k
, where _ denotes the standard lexicographical ordering in the set of all
k-sequences whose terms belong to 0, 1.
The details of the above-mentioned construction are left to the reader.
Now, we dene
P =
k<
(T
i
1
i
2
...i
k
: (i
1
, i
2
, ...i
k
) 0, 1
k
).
It follows from the construction of P that g[P = f [P is an increasing function on P which
completes the proof of Theorem 8.
The natural question arises whether the preceding theorem admits further generalizations.
In particular, can we assert that it is possible to choose the above-mentioned nonempty
perfect set P in such a way that (P) > 0?
We will show that the answer to the posed question is negative. For this purpose, we need
the existence of a continuous function f : RR which is nowhere approximately differen-
tiable. This fact is much deeper than the existence of a continuous nowhere differentiable
function. The denition of an approximate derivative is given in Appendix 4 with all other
auxiliary notions needed for our purposes.
The following classical result was rst obtained by Jarnik (see [88]).
Theorem 9. There exist continuous bounded functions acting from R into R which are
nowhere approximately differentiable.
Actually, Jarnik proved that almost all (in the sense of the Baire category) functions from
the Banach space C[0, 1] are nowhere approximately differentiable. Notice that this re-
sult signicantly strengthens the corresponding result of Banach and Mazurkiewicz for the
usual differentiability (see Exercise 4 of Chapter 8).
Remark 7. In [122] the existence of a nowhere approximately differentiable function
is applied to the question concerning some relationships between the notions of sup-
measurability and weak sup-measurability of functions acting from RR into R.
Theorem 10. There exists a continuous function f : R R such that for every perfect set
P R with (P) > 0, the restriction f [P is not monotone on P.
Proof. Let f : RR be a continuous, bounded, and nowhere approximately differentiable
function (see Theorem 9 above). We are going to show that f is the required function.
Take any perfect set P R whose Lebesgue measure is strictly positive. We assert that
f [P cannot be monotone. Suppose otherwise, i.e., f [P is either increasing or decreasing.
Without loss of generality, we may assume that f [P is increasing. Denote by
f
: R R
some increasing function extending f [P (see Example 5 from Chapter 1). According to the
classical result of Lebesgue, f
is differentiable almost everywhere. Consequently, there

exists x P such that x is a density point of P and f
is differentiable at x. By virtue of
the relation f
[P = f [P, this circumstance immediately implies the fact that the original
function f is approximately differentiable at x. But the latter contradicts the denition of
f . The contradiction obtained nishes the proof.
Remark 8. Actually, the function f of Theorem 10 is such that for every set X R with
(X) > 0, the restriction f [X is not monotone; the proof remains the same as before. It
should be underlined that this result holds true within ZFC theory. The question naturally
arises whether in ZFC theory there exists a function g : R R such that for every set
X R with
(X) > 0, the restriction g[X is not continuous. As shown by Roslanowski

and Shelah [208], the existence of g cannot be proved within ZFC. On the other hand, as-
suming Martins Axiom, any Sierpi nski-Zygmund function can play the role of the above-
mentioned g.
In view of Theorems 8 and 10, it makes sense to introduce the following denitions.
Let f : R R be a function and let L be a family of subsets of R.
We say that f is relatively monotone with respect to L if there exists at least one set X L
for which the restriction f [X is monotone.
We say that f is absolutely nonmonotone with respect to L if f is not relatively monotone
with respect to L (i.e., there exists no set X L for which f [X is monotone).
Example 6. Any function f : RR turns out to be relatively monotone with respect to the
class of all countably innite subsets of R. This fact can easily be deduced from the innite
version of the Ramsey theorem (see Chapter 19) and also admits a direct simple proof.
In some sense, the above-mentioned result cannot be strengthened because, under the Con-
tinuum Hypothesis, any Sierpi nski-Zygmund function is absolutely nonmonotone with re-
spect to the class of all uncountable subsets of R.
In connection with Example 6, the following statement seems to be of interest.
Theorem 11. Assuming the Continuum Hypothesis, there exists a function g : R R
which is not a Sierpi nski-Zygmund function but is absolutely nonmonotone with respect to
the family of all uncountable subsets of R.
Proof. Let us take an arbitrary continuous nowhere differentiable function g
1
: R R and
an arbitrary Sierpi nski-Zygmund function g
2
: R R. Let L be a Luzin subset of R, the
existence of which is implied by the Continuum Hypothesis (see Exercise 14 for Appendix
1).
Dene the function g : R R as follows:
g(x) = g
1
(x) if x L and g(x) = g
2
(x) if x R L.
Let us check that g is the required function. First of all, g is not a Sierpi nski-Zygmund
function because g[L =g
1
[L is continuous. Let now X be an uncountable subset of R. Only
two cases are possible.
1. card(X L) . In this case, we have card(X (R L)) >. Since g
2
is a Sierpi nski-
Zygmund function, the restriction
g[(X (R L)) = g
2
[(X (R L))
is not monotone (see Example 6 of this chapter or Example 8 of Chapter 1).
2. card(X L) > . Suppose for a while that g[(X L) = g
1
[(X L) is monotone. Then
the continuous function g
1
should be monotone on the closure of X L. Further, since g
1
is nowhere differentiable, the set cl(X L) is necessarily nowhere dense in R. Taking into
account the fact that L is a Luzin set, we get
card(Lcl(X L)) ,
which contradicts the inclusion X L Lcl(X L) and the uncountability of X L. The
obtained contradiction nishes the proof.
Some other results concerning restrictions of various functions to non-small subsets of the
real line can be found in the works [8], [27], [28], [29], [31], [35], [42], [89], [148], [206],
[208], and [214].
EXERCISES
1. Suppose that E is a separable topological space (i.e., E contains a countable everywhere
dense subset). Show that card(Ba
1
(E, R)) c. In fact, show that if E is nonempty and
separable, then card(Ba
1
(E, R)) = c.
2. Verify that the set Q R of all rational numbers is not a G
-subset of R. Deduce from

this circumstance that the characteristic function
Q
(the so-called Dirichlet function) does
not belong to Ba
1
(R, R). Actually,
Q
is of second Baire class.
More generally, prove that if E is an uncountable Polish space without isolated points and
X is a countable everywhere dense subset of E, then the characteristic function
X
does not
belong to Ba
1
(E, R).
3. Let E be a topological space and let f : E R be a function. As usual, denote by
D( f ) the set of all discontinuity points of f and, respectively, denote by C( f ) the set of all
continuity points of f . Verify that:
(a)
f
(x) = 0 if and only if f is continuous at x (equivalently,
f
(x) > 0 if and only if x is
a discontinuity point of f );
(b) for any t R, the set x E :
f
(x) t is closed in E;
(c) the set D( f ) is representable in the form
D( f ) = E
1
E
2
... E
n
...,
where E
n
=x E :
f
(x) 1/n for each integer n 1.
Conclude from (b) and (c) that D( f ) is an F
-subset of E and therefore the set C( f ) of all

continuity points of f is a G
-subset of E.
Let B be a base of open sets in R and let
F =Y R : RY B.
Prove the equality
D( f ) =cl( f
1
(Y)) f
1
(Y) : Y F.
Generalize these results to an arbitrary function f : E E
/
where E
/
is a given metric
space.
4
. Let E be a nonempty resolvable topological space (see Chapter 10) and let X be an
F
-subset of E.
Show that there exists a function f : E R such that X coincides with the set D( f ) of all
discontinuity points of f .
For this purpose, rst represent X in the form
X = F
1
F
2
... F
n
...,
where all sets F
n
are closed in E and F
n
F
n+1
for each natural number n 1. Further, put
F
0
= / 0 and for any integer n 1, dene a function
f
n
: E 0, 1
satisfying the relations:
(a) f
n
is equal to zero at all points of the set E F
n
;
(b)
f
n
(x) = 1 for each point x F
n
.
Now, take a sequence a
n
: n 1 of strictly positive real numbers, such that
a
n+1
+a
n+2
+... +a
k
+... < a
n
(n = 1, 2, ...).
For example, it sufces to put
a
n
=
1
3
n
(n 1).
Finally, consider the function
f = a
1
f
1
+a
2
f
2
+... +a
n
f
n
+... .
This function is well dened because the series on the right-hand side of the above equality
converges uniformly on E. Verify that:
(c) f is continuous at all points of the set E X;
(d) for any integer n 1 and for all points x F
n
F
n1
, we have
f
(x) a
n
k>n
a
k
> 0.
Conclude from (c) and (d) that D( f ) = X.
Some further results closely connected with Exercise 4 can be found in [17] and [189].
5. Let E be a perfectly normal topological space and let f : E R be a function whose
graph is closed in the product space E R. Applying the Kuratowski theorem on closed
projections (see Chapter 8), show that f is of rst Baire class.
Give an example of a function f : R R satisfying the following two conditions:
(a) the graph Gr( f ) is closed in RR;
(b) D( f ) is a nonempty perfect subset of R.
6. Let E be an arbitrary topological space and let f Ba
1
(E, R). Demonstrate that the set
D( f ) is of rst category in E (for this purpose, use Exercise 3 and Lemma 2).
Another way to show this fact is based on the Banach statement (see Exercise 12 for Ap-
pendix 3) which leads to a representation of E in the form
E = E
0
E
1
,
where E
0
is an open Baire subspace of E and E
1
is a rst category closed subset of E.
Applying Theorem 4 to the set D( f [E
0
), we get the required result.
Deduce from this result that if f Ba
1
(E, R) and X is an arbitrary subspace of E, then the
set D( f [X) is of rst category in X.
In particular, conclude that if E is a complete metric space, f Ba
1
(E, R) and X is a
nonempty closed subspace of E, then there exist points in X at which f [X is continuous.
7. Give another proof of the Cantor-Bendixson theorem (see Example 5) that does not use
the method of transnite induction.
For this purpose, take an arbitrary closed set X in a hereditarily Lindel of space E, consider
the set of all condensation points of X and take it as Y. Then dene the set Z by the equality
Z = X Y and check that Z is at most countable.
8
. Let E be a topological space. We say that E is scattered if E does not contain a nonempty
dense in itself subset.
Demonstrate that the following two assertions are equivalent:
(a) E is scattered;
(b) E can be represented in the form of an injective -sequence
E =e
: <,
where is some ordinal number and, for any <, the element e
is an isolated point of
the set e
: <.
Notice that the implication (b) (a) is trivial. Supposing now that (a) is valid, use the
method of transnite recursion for obtaining the required representation of E.
Finally, demonstrate that every topological space X can be expressed in the form X =Y Z
where Y is a perfect subset of X, Z is a scattered subset of X, and Y Z = / 0. This classical
result is due to Cantor.
9
. Let E be a separable metric space and let g : E R be a function such that for every
nonempty closed set F E, there exists a point x F at which the restricted function g[F
is continuous. Prove that g is of rst Baire class. This very important statement is due to
Baire and yields a characterization of all functions belonging to Ba
1
(E, R).
The following argument enables to establish the above-mentioned result.
Take any a R and b R such that a < b, and denote
A =x E : g(x) > a, B =x E : g(x) < b.
Clearly, we have the equality E = A B. Further, construct by transnite recursion an
1
-sequence
F
0
F
1
... F
... ( <
1
)
of closed subsets of E. Put F
0
= E. Suppose that for a given <
1
, the partial family
F
: < has already been dened.

If is a limit ordinal, then we put
F
=F
: <.
If = +1, consider the set F
. Only two cases are possible.

(i) F
= / 0. In this case, we dene F
= F
= / 0.
(ii) F
,= / 0. In this case, there exists a point x F
at which the function g[F
is continuous.
Consequently, there exists an open neighborhood V(x) of x such that
F
V(x) A F
V(x) B.
Then we dene F
= F
V(x).
Proceeding in this way, we will be able to construct the sets F
( <
1
).
Observe now that for each ordinal <
1
, we have
F
F
+1
A F
F
+1
B
and according to the Cantor-Baire stationarity principle, for some <
1
, the equalities
F
= / 0 ( <
1
)
are valid. Deduce from these facts that there exist two sets A
/
and B
/
such that
A
/
A, B
/
B, A
/
B
/
= E, A
/
B
/
= / 0
and both A
/
and B
/
are F
-subsets of E.
Now, let b
n
: n < be a strictly decreasing sequence of real numbers satisfying the
relation lim
n+
b
n
= a. Put again
A =x E : g(x) > a
and for each n <, dene
B
n
=x E : g(x) < b
n
.
As above, show the existence of F
-sets A
/
n
and B
/
n
such that
A
/
n
A, B
/
n
B
n
, A
/
n
B
/
n
= E, A
/
n
B
/
n
= / 0.
Finally, denote X =A
/
n
: n < and verify that
X =x E : g(x) > a = A.
Hence A is an F
-subset of E.
By using a similar argument, demonstrate that B =x E : g(x) < b is an F
-subset of E,
too.
Conclude, in view of the Lebesgue theorem (i.e., Theorem 3 of this chapter) that the func-
tion g is of rst Baire class.
In connection with Exercise 9, we would like to mention that more general versions of
the result presented above can be found in monograph [148] where a different argument is
applied. Namely, it is proved there that if E is a complete metric space and g : E R is a
function, then the following two assertions are equivalent:
(1) g Ba
1
(E, R);
(2) for any nonempty closed set F E, there exists a point x F at which the restricted
function g[F is continuous.
The Baire theorem (i.e., Theorem 4 of this chapter) and the preceding exercise establish
the equivalence (1) (2) in the case of a Polish space E. For a nonseparable complete
metric space E, the proof of (1) (2) relies on properties of the Montgomery operation
(see [148] and Exercise 4 from Appendix 3). Notice that this operation needs uncountable
forms of the Axiom of Choice.
In Luzins opinion, the Baire characterization of all functions of rst class is a jewel of real
analysis (see [162]).
10. Let E be a Polish topological space and let f : E R be a function whose set of
discontinuity points is at most countable. Show, by applying Exercise 9, that f Ba
1
(E, R).
Infer from this fact that any function
g : [a, b] R
of nite variation on a segment [a, b] belongs to the rst Baire class. Another way to
establish this result was indicated in Example 3.
11. Let E be a topological space and let X be a subset of E such that for every perfect set
P E, the intersection X P is of rst category in P.
Show that X is always of rst category.
12
. A function f : RR is called nowhere constant if for any non-degenerate subinterval

U of R, the restriction f [U is not constant.
Let F denote the family of all those continuous functions acting from R into itself which
are nowhere constant.
Prove under the Continuum Hypothesis that there exists a set X R having the following
property:
if f F and g F are any functions from F such that f (X) = g(X), then f = g.
Every set X with this property is called a magic set for the family F (cf. [11] and [42]).
Argue by using the method of transnite induction. Let f
: <
1
be an injective
enumeration of all functions from F. Suppose that for an ordinal number <
1
, the two
families
x
,n
: <, n <, y
,
: < <
of points of R are dened in such a way that for < <, the relation
y
,
f
(W) f
(W)
holds, where
W =x
,n
: <, n <.
Further, denoting W
/
=y
,
: < <, put
U =f
(W) : ,
V = (f
1
(U) : ) (f
1
(W
/
) : ).
Observe that the set U is countable and the set V is of rst category in R. Then represent the
family f
: < in the form of an -sequence g

n
: n < and construct by ordinary
recursion an injective -sequence x
,n
: n < of points in R satisfying the following
relations:
(a) x
,n
,V for all n <;
(b) f
(x
,n
) ,= g
n
(x
,m
) for all n < and m <.
Afterwards, dene an appropriate family y
,
: < of points in R.
Finally, put
X =x
,n
: <
1
, n <
and check that X is a magic set for the family F.
Obtain the same result under Martins Axiom (instead of the Continuum Hypothesis).
A similar statement on the existence of a magic set for those Lebesgue measurable func-
tions which are not constant on subsets of R of strictly positive -measure can be found in
[35].
In connection with Exercise 12, the result of Ciesielski and Shelah should also be men-
tioned which states that there exists a model of ZFC where there are no magic sets for the
family F (see [42]).
13
. Assuming Martins Axiom, prove that there exists a partial continuous function f :
R R satisfying the following two conditions:
(a) the set dom( f ) is -thick in R;
(b) for any set Y dom( f ) with card(Y) = c, the restriction f [Y is not monotone.
For this purpose, consider a suitable restriction of a continuous nowhere approximately
differentiable function whose existence is guaranteed by Theorem 9 of the present chapter
(see also Appendix 4).
14. Let (E, ) be an innite, Dedekind complete, dense, linearly ordered set endowed with
its order topology and let
f : E R
be a function having the Baire property. Formulate and prove for f an analogue of Theorem
8. Apply the obtained result to the particular case where (E, ) is a Suslin line (see Exercise
16 from Appendix 3).
Appendix 1
Some set-theoretical facts and constructions
The main goal of this Appendix is to present several purely set-theoretical statements and
constructions which are useful in various elds of modern mathematics and, therefore, have
a number of nontrivial and, quite often, unexpected applications.
Primarily, we will be concerned with fundamental facts of innite combinatorics. The
reader will see in the sequel that some of them are closely connected with the measure
extension problem (cf. Chapters 2, 16) and, actually, were essentially inspired by this
problem.
We begin with the standard denition of a tree which is a special (but extremely important)
case of a partially ordered set.
Let (E, _) be a partially ordered set. We say that (E, _) is a tree if the following conditions
are satised:
(a) (E, _) has a least (smallest) element;
(b) for any e E, the set
E(e) =x E : x e
is well-ordered.
It trivially follows from condition (a) that any tree is nonempty.
The least element of a tree (E, _) is called its root.
For each element e E, the ordinal type of the above-mentioned well-ordered set E(e) is
called the height of e and is usually denoted by o(e).
In particular, the height of the root is 0.
If is an ordinal, then the set E
of all those elements from E whose heights are equal to

can be considered. This set is called the -th level of E. In other words, we have
E
=e E : o(e) =.
Notice that all elements of the -th level are pairwise incomparable (because no well-
ordered set is isomorphic to its proper initial subinterval).
DOI 10.1007/978-94-91216-36-7, 2009 Atlantis Press/World Scientific
339
We also dene
o(E) = supo(e) +1 : e E.
The ordinal o(E) is called the height of a given tree (E, _).
It can readily be seen that the height of E coincides with the least element of the class of
all those ordinals for which the -th level of (E, _) is empty.
A tree (E, _) is called an -tree if its height is equal to .
In particular, an -tree is a tree all whose n-th levels are nonempty, where n < , but its
-th level is empty.
We say that a set X is a branch in a tree (E, _) if X is a maximal linearly ordered (by the
induced ordering) subset of (E, _).
Actually, any branch is a maximal well-ordered subset of (E, _) (this fact follows directly
from the denition of trees).
Let be an ordinal number. We say that a set X E is an -branch in (E, _) if X is a
branch with height equal to .
We say that (E, _) is a K onig tree if (E, _) is an -tree all levels of which are nite.
The following important statement is due to K onig (cf. [91], [144], [145], and [150]).
Theorem 1. Let (E, _) be a K onig tree. Then there exists an -branch in (E, _).
Proof. For every natural number n, a subset X of E will be called a partial n-branch in
(E, _) if X is linearly ordered by the induced ordering, has nonempty intersection with all
k-th levels of (E, _) where k < n, but does not intersect the k-th levels of (E, _), where
k n.
Obviously, every partial n-branch contains exactly n elements.
An element e E will be called admissible if for any natural number n > 0, there exists a
partial n-branch containing e.
It can readily be proved (by ordinary induction) that there are admissible elements in every
nonempty level of (E, _). In particular, the root of (E, _) is admissible. Moreover, suppose
that a nite sequence
(e
0
, e
1
, . . . , e
n
)
of admissible elements is constructed such that the set e
0
, e
1
, . . . , e
n
is linearly ordered
and each e
k
belongs to the k-th level (i.e., this set is a partial (n +1)-branch in (E, _)).
Then, by using the niteness of the (n +1)-th level of (E, _), it is not difcult to infer
that there exists an admissible element e
n+1
from the (n +1)-th level such that e
n
e
n+1
.
Continuing in this manner, we will obtain an -branch
(e
0
, e
1
, . . . , e
n
, . . .)
Appendix 1. Some set-theoretical facts and constructions 341
of the given K onig tree (E, _). This completes the proof.
As already mentioned, the assertion just proved is due to K onig and is known in the litera-
ture as the K onig lemma. It yields useful consequences in graph theory, combinatorics and
many other areas of mathematics.
For example, by applying this lemma it can be established that if any nite subgraph of a
given countable graph is planar (i.e., has an isomorphic copy in the Euclidean plane R
2
),
then the graph is planar, too.
Notice also that the proof of the K onig lemma needs some weak form of the Axiom of
Choice. In other words, Theorem1 is not provable within ZF theory (see Exercise 2 below).
It is natural to ask whether certain analogues of the K onig lemma are valid for those trees
whose height is uncountable but all levels have relatively small sizes in comparison with
the height.
First, let us introduce a relevant notion.
We shall say that the branch property holds for a given tree (E, _) if there exists at least
one branch in (E, _), whose height is equal to the height of (E, _).
The K onig lemma states that all K onig trees have the branch property. So it is reasonable
to ask whether an analogous result is true for
1
-trees with countable levels. However,
the answer to the question just posed is negative. The corresponding counterexample was
originally constructed by Aronszajn (cf. [91], [144], and [145]).
Theorem 2. There exists an
1
-tree (E, _) such that:
(1) all levels of (E, _) are countable;
(2) (E, _) does not have the branch property.
Proof. Recall that the symbol
<
1
denotes the family of all functions f such that dom( f )
is a proper initial subinterval of
1
and ran( f ) is a subset of . Using this notation, let us
put
F = f
<
1
: f is an injection.
For any two functions f F and g F, dene f g if and only if
dom( f ) = dom(g), card( dom( f ) : f () ,= g()) <.
Applying the method of transnite recursion, we will construct an
1
-sequence
f
: <
1
F
(a) dom( f
) = [0, [ for each ordinal <

1
;
(b) f
[[0, [ f
for any two ordinals <

1
and <;
(c) card( ran( f
)) = for each ordinal <

1
.
Notice rst that if for some <
1
, the function f
is already dened, then the role of f

+1
can be played by any function from F which extends f
and whose domain coincides with

[0, ] = [0, +1[.
The case of limit countable ordinals needs a more delicate argument.
Let <
1
be a limit ordinal and suppose that the partial family of functions f
: <
satisfying the relations (a), (b), (c) has already been dened. Take an arbitrary strictly
increasing sequence of ordinal numbers
n
: n < such that
lim
n+
n
=,
and construct (by ordinary recursion) a sequence
f
/
n
: n < F
with the following properties:
(i) f
/
n
f
n
for each n <;
(ii) f
/
n+1
[[0,
n
[ = f
/
n
for each n <.
Simple details of such a construction are left to the reader.
Now, let f
/
denote the unique function dened on [0, [ and extending all functions f
/
n
(n <
). Obviously, the existence of f
/
follows directly from property (ii).

Also, it is easy to verify that f
/
F and f
/
[[0, [ f
for all ordinals <.

Further, it is not hard to show that an injective mapping
f
: [0, [
can be dened satisfying the next two conditions:
(iii) f
() = f
/
() for any ordinal [0, [

n
: n <;
(iv) the set ran( f
) is innite.
A straightforward verication shows that f
is the required function, i.e., the relations (a),

(b), (c) are valid for the partial family f
: .
Finally, having the
1
-sequence of functions
f
: <
1
F
with properties (a), (b), and (c), we put
T = f F : ( <
1
)( f f
)
and equip T with its standard ordering _ (in other words, f _g if and only if g extends f ).
Summarizing all the said above, we readily conclude that the partially ordered set (T, _)
is an
1
-tree, all levels of (T, _) are countable, and (T, _) does not possess the branch
property because there exists no injective function acting from
1
into . Theorem 2 has
thus been proved.
(T, _) is usually called an Aronszajn tree.
At rst sight, Aronszajn trees seem to be useless in measure theory and real analysis. But
it turns out that such trees can successfully be applied to concrete problems in the above-
mentioned elds of mathematics. For example, in connection with the general measure
extension problem Ryll-Nardzewski posed the following natural question.
Let (E, S, ) be a probability measure space and let L be a family of subsets of E. Sup-
pose that for every countable subfamily L
/
of L, there exists a measure
/
extending
and dened on the -algebra generated by S L
/
. Does there exist a measure also
extending and dened on the -algebra generated by S L?
It turns out that the answer to this question is negative. The corresponding counterexample
was constructed by K.P.S. Bhaskara Rao with the aid of an Aronszajn tree.
Theorem 2 shows that within ZFC theory the K onig lemma does not admit a natural gen-
eralization to all uncountable cardinals and we have seen that even the rst uncountable
cardinal provides us with a counterexample. Nevertheless, the branch property holds for
some trees of large cardinality.
For instance, cardinals which are real-valued measurable (i.e., measurable in the Ulam
sense) may be treated as a certain kind of large cardinals.
We recall that a cardinal a is real-valued measurable if there exists a nonzero -nite
(equivalently, probability) diffused measure on a with dom() = P(a), where P(a)
denotes the power set of a.
The existence of real-valued measurable cardinals cannot be established within ZFC theory
(see [91], [145], [238], or Exercise 18). For the least of such cardinals, the next statement
is valid.
Theorem 3. Let (E, _) be a tree satisfying the following two conditions:
(a) o(E) is the rst real-valued measurable cardinal;
(b) all levels of E are of cardinality strictly less than o(E).
Then (E, _) has the branch property.
Proof. In fact, the argument given below shows that a more general result is true.
Namely, if is a cardinal number with a -additive diffused probability measure on P()
and (E, _) is a -tree all levels of which are of cardinality strictly less than , then (E, _)
has the branch property.
Of course, in this formulation is identied with the least ordinal of cardinality and we
recall that some approaches to the notion of a cardinal number automatically yield such an
identication.
To establish the above-mentioned generalized result, let us dene
E
(x) =y E : x _y
for each x E and let E
be the -th level of E for any ordinal < . Denote by a

-additive diffused probability measure on P(E) and put
t
0
(x) = (E
(x)),
t
1
() = supt
0
(x) : x E
.
Since the equality
(E
(x) : x E
) = 1
holds for any <, we infer in view of the -additivity of that
( <)(t
1
() > 0).
Furthermore, if <, then we have
t
1
() t
1
().
Therefore, there exist an ordinal < and a real number r > 0 such that
( [, [)(t
1
() = r).
Now, we introduce the set
S =x : ( )(x E
& t
0
(x) > (1/2)r).
Obviously, this S satises the relation
S E
,= / 0
for all those which are strictly smaller than . In particular, we have
card(S) =.
The set S possesses also the following property:
(y S)(x S)(x E
& x _y).
Indeed, to verify this fact, take any y S. Then y E
for some and t

0
(y) > (1/2)r.
Pick an element x _y from the level E
. Clearly, the inclusion

E
(y) E
(x)
holds from which it follows that t
0
(x) > (1/2)r and hence x S, as well. Taking into
account the circumstance that card(E
) <, we deduce that

card(S E
(z)) =
for some z E
. Finally, let us demonstrate that any two distinct elements u and v from
SE
(z) are comparable with respect to _. Supposing to the contrary that u ,_v and v ,_u,
we get
E
(u) E
(v) = / 0,
E
(u) E
(v) E
(z),
t
0
(u) > (1/2)r, t
0
(v) > (1/2)r
and, consequently,
t
0
(z) t
0
(u) +t
0
(v) > r,
which is impossible in view of the relation
t
0
(z) t
1
() = r.
The obtained contradiction shows that SE
(z) is a linearly ordered subset of E whose car-

dinality equals . Evidently, this subset can be included in some -branch of E. Theorem
3 has thus been proved.
Combining Theorems 2 and 3, we readily obtain the following result which is due to Ulam
but was established by himin absolutely another manner (see [148], [176], [192], and [238]
where a transnite matrix of a certain combinatorial type is essentially exploited; compare
also with Exercise 18 for this Appendix).
Theorem 4.
1
Proof. Indeed, suppose to the contrary that
1
is real-valued measurable. The unique
innite cardinal strictly less than
1
is , which obviously is not real-valued measurable.
So we conclude that
1
necessarily is the rst real-valued measurable cardinal. According
to Theorem 3, all
1
-trees with countable levels should have the branch property. But this
circumstance contradicts the existence of an Aronszajn tree (see Theorem 2).
Theorem 4 is of special interest to us because it is closely connected with the general
measure extension problem (see Chapter 2). It is also connected with one classical result
of Sierpi nski [218] (cf. [222], [223]), which will be proved below. To present this result
with some related ones, we need several auxiliary propositions.
Lemma 1. Let E be a set and let f : E R be a partial function. There exists a countably
generated -algebra S of subsets of E such that f is measurable with respect to B(R)
and S, i.e., we have
(X B(R))( f
1
(X) S).
Proof. Let V
n
: n < be a countable topological base of R. For example, we may take
the family of all open intervals in R with rational end-points. Putting
S =( f
1
(V
n
) : n <),
we get the required countably generated -algebra of subsets of E.
Lemma 2. Let E be a subset of R and let f : E E be a partial function. Then there exists
a countably generated -algebra S of subsets of E such that
Gr( f ) S S,
where Gr( f ) as usual denotes the graph of f .
This statement can readily be deduced from Lemma 1 (see Exercise 6). In addition to this
fact, it should be noticed that we may assume without loss of generality that S contains
all singletons in E. Indeed, it sufces to consider the -algebra (S B(E)), instead of
S, and take into account the fact that if S is countably generated, then (S B(E)) is
also countably generated.
Lemma 3. Let Z be any subset of the product set
1
1
. There exists a countably
generated -algebra S of subsets of
1
such that Z S S.
Proof. In the product set
1
1
consider the following two subsets:
A =(, ) : <, B =(, ) : .
Obviously, we have the equalities
AB =
1
1
, AB = / 0.
In other words, A, B is a partition of the product set
1
1
, which is frequently called
the Sierpi nski partition of
1
1
. Additionally, the sets A and B satisfy the following two
relations:
(a) card(A(
1
)) for every <
1
;
(b) card(B(
1
)) for every <
1
.
Evidently, we can write the equality
Z = (Z A) (Z B).
This equality and the relations (a) and (b) show that Z admits a representation in the form
Z =Z
n
: n <,
where for each n <, either Z
n
or Z
1
n
is the graph of a partial function acting from
1
into
1
. Since
1
does not exceed the cardinality of the continuum, we can apply Lemma 2 to
a set E R with card(E) =
1
and, in fact, here we identify E with
1
. According to the
above-mentioned lemma, every set Z
n
belongs to S
n
S
n
, where S
n
is a certain countably
generated -algebra of subsets of
1
. Now, it directly follows from this circumstance that
Z belongs to S S where the -algebra
S =(S
n
: n <)
is countably generated, too. Lemma 3 has thus been proved.
The next auxiliary proposition is a stronger version of Theorem 4 and is closely connected
with the Sierpi nski partition of
1
1
.
Lemma 4. There exists a -algebra S of subsets of
1
(1) S is countably generated;
(2) S contains all singletons in
1
;
(3) there is no nontrivial -nite diffused measure on S.
Proof. In view of Lemma 3, there exists a countably generated -algebra S of subsets of
1
such that
A S S,
where A is the rst member of the Sierpi nski partition A, B of
1
1
. We may also
assume that S contains all singletons in
1
(cf. the remark made after Lemma 2). It
remains to show that S does not admit a nonzero -nite diffused measure. Indeed, let
be a -nite diffused measure with dom() =S. Consider the product measure
= .
Keeping in mind the relations (a) and (b) of the proof of Lemma 3 and using the Fubini
theorem, we readily infer that
(A) = 0, (B) = 0.
Consequently, the equalities
((
1
))
2
=(
1
1
) =(AB) =(A) +(B) = 0, (
1
) = 0
must be valid which completes the proof.
Now, we are able to establish within ZFC theory the existence of an uncountable universal
measure zero subset of the real line R. Taking into account the fact that all uncountable
Polish topological spaces are Borel isomorphic (see Appendix 6), it sufces to show the
existence of an uncountable universal measure zero set in some uncountable Polish topo-
logical space E.
We will show the existence of such a set for the Cantor discontinuum, that is for
E = 2
=0, 1
.
Namely, we have the following classical statement (see, e.g., [69], [79], [113], [148], [162],
and [172]).
Theorem5. There exists an uncountable universal measure zero subset of the Cantor space
2
.
Proof. Let
n
: n < be a countable family of subsets of
1
satisfying the following
conditions:
(a)
n
: n < separates the points in
1
, i.e., for any two distinct ordinals <
1
and
<
1
, there is
n
such that
card(
n
, ) = 1;
(b) the -algebra (
n
: n <) does not admit a nonzero -nite diffused measure.
Notice that the existence of
n
: n < is directly implied by Lemma 4.
Now, we dene a mapping
:
1
0, 1
by putting
() = (
n
())
n<
( <
1
),
where
n
() = 1 if
n
and
n
() = 0 if ,
n
.
In other words, we associate with the family of sets
n
: n < its characteristic function
which is frequently called Marczewskis function (see [148] and [150]).
Since
n
: n < separates the points in
1
, the mapping is injective. Denote
Z =(
1
).
Since is an injection, we have the equality
card(Z) =
1
.
Further, from the denition of it immediately follows that
n
=
1
(Z t 2
: t
n
= 1)
for all n < . Suppose for a moment that there exists a nonzero -nite diffused Borel
measure on Z. Then in view of the preceding relation, we come to a nonzero -nite
diffused measure dened on (
n
: n <) by the formula
(
1
(Y)) = (Y) (Y B(Z)).
But this gives a contradiction with condition (b). Therefore, Z is a universal measure zero
subspace of 2
which completes the proof.

There are several other constructions of uncountable universal measure zero subspaces of
the real line, which use essentially different ideas. For more details, see [197] and [250].
Once again, we would like to underline that the existence of an uncountable universal
measure zero subset of R directly implies the nonmeasurability (in the Ulam sense) of
1
.
The nonmeasurability in the Ulam sense of
1
has a nontrivial application in the general
theory of -nite invariant (quasi-invariant) measures (see Exercises 9, 10, and 11 below).
There are also applications of Ulams transnite matrix to some questions concerning the
existence of those subsets of a given second category topological space, which do not
possess the Baire property (see [148] and Exercise 19 for this Appendix).
EXERCISES
1. Let (E, _) be a tree and let (X, _) be a branch in (E, _). Verify that, for any element
x X, the set y X : y _x coincides with y E : y _x.
2
. Prove that the following two assertions are equivalent within ZF theory:
(a) the K onig lemma;
(b) for any family X
n
: n < of nonempty nite sets, there exists its selector.
3
. Show that if a countable graph is such that all its nite subgraphs are realizable in the
Euclidean plane R
2
, then the entire graph is also realizable in the plane R
2
.
Give an example of a graph of cardinality
1
, all countable subgraphs of which are realiz-
able in R
2
, but the graph cannot be realized in R
2
.
4. Verify that the K onig lemma is essentially used in the Kakutani-Oxtoby construction of
a nonseparable translation-invariant extension of the Lebesgue measure on the real line
R (see Exercise 18 for Chapter 3; cf. also Chapters 16 and 17).
5
. Prove an analogue of the K onig lemma for those innite trees all levels of which are
nite.
For this purpose, keep in mind the fact that an arbitrary innite set admits an appropriate
nontrivial ultralter of its subsets and use an argument similar to the proof of Theorem 3.
6. Give a detailed proof of Lemma 2.
For this purpose, consider any partial function
f : E E,
where E is a subset of R, and introduce the following two sets associated with this partial
function:
f
=(x, y) E E : x dom( f ), y f (x),
f
=(x, y) E E : x dom( f ), y f (x).
Observe that
Gr( f ) =
f

f
.
By virtue of Lemma 1, there exists a countably generated -algebra T of subsets of E
such that f becomes measurable with respect to the -algebras B(E) and T . Then show
that
f
T B(E),
f
T B(E)
and, consequently, Gr( f ) T B(E). To prove these relations, rst consider the case
where f is a partial step-function and then approximate an arbitrary partial function acting
from E into itself by a sequence of partial step-functions.
Conclude from the said above that S =(T B(E)) is the required countably generated
-algebra of subsets of E and, moreover, all singletons in E belong to S.
7. Let X
: <
1
be a family of subsets of
1
. Show that there exists a countably
generated -algebra S of subsets of
1
such that
X
: <
1
S.
For this purpose, apply Lemma 3 to the set
Z = X
: <
1

which is contained in the product set
1
1
.
8
. Assuming the Continuum Hypothesis, prove that there exist two subsets Z
1
and Z
2
of
the Euclidean plane R
2
, satisfying the following conditions:
(a) Z
1
is uniform in the direction parallel to the line 0 R;
(b) Z
2
is uniform in the direction parallel to the line R0;
(c) there is a countable family h
n
: n < of translations of R
2
such that
h
n
+(Z
1
Z
2
) : n < = R
2
.
Conversely, derive the Continuum Hypothesis from the assumption that there exist two
subsets Z
1
and Z
2
of R
2
satisfying the above-mentioned relations (a), (b), and (c).
Observe that relation (c) implies the existence of a function f : R R such that
R
2
=g
n
(Gr( f )) : n <
for some countable family g
n
: n < of isometric transformations of R
2
.
The results just formulated are essentially due to Sierpi nski (cf. [105], [218], [222], and
[223]).
9
. Let E be a set, G be a group of transformations of E and let be a nonzero -nite

G-quasi-invariant measure on E. Assume that G is uncountable and acts almost freely in E
with respect to , i.e., for any two distinct transformations g G and f G, we have
(x E : g(x) = f (x)) = 0.
Prove that dom() ,=P(E); in other words, there exists a subset of E nonmeasurable with
respect to .
For this purpose, suppose otherwise, i.e., suppose that dom() =P(E) and assume, with-
out loss of generality, that is a probability G-quasi-invariant measure on E. Denote by H
a subgroup of G with card(H) =
1
and let X be an arbitrary H-selector in E. First, check
that:
(a) h(X) : h H = E;
(b) (h(X) f (X)) = 0 for any two distinct transformations h H and f H.
Then derive from relation (b) that
(c) (h(X)) = 0 for each transformation h H.
Taking into account both relations (a) and (c), construct by transnite recursion an
1
-
sequence X
: <
1
of subsets of E such that:
(d) X
: <
1
= E;
(e) X
= / 0 whenever <
1
, <
1
and ,=;
(f) (X
) = 0 for all <

1
.
Finally, for each set
1
, put
() = (X
: )
and verify that is a probability diffused measure on
1
whose domain coincides with
P(
1
). This circumstance yields a contradiction with the fact that the cardinal
1
is not
real-valued measurable (see Theorem 4).
10. Preserving the notation and assumptions of Exercise 9, strengthen the result obtained
therein by proving that any set Y E with
(Y) > 0 contains a subset nonmeasurable

with respect to .
Reduce this case to the situation of Exercise 9 by using the following argument. Suppose
otherwise, i.e., P(Y) dom() and consider a set Y
/
E which is almost G-invariant
with respect to and is representable in the form
Y
/
=g
k
(Y) : k <,
where g
k
: k < is some countable family of transformations from G. The existence of
Y
/
follows from Exercise 13 of Chapter 2.
Check that P(Y
/
) dom(). Further, put
(Z) = (Z Y
/
) (Z E)
and verify that is a nonzero -nite G-quasi-invariant measure on E with dom() =
P(E) which contradicts the result of Exercise 9.
11. Deduce from the two preceding exercises that if E is a set and G is an uncountable
group of transformations of E acting freely in E, then for every nonzero -nite G-quasi-
invariant measure on E, the relation dom() ,= P(E) holds true. More generally, any
set Y E with
(Y) > 0 contains a subset nonmeasurable with respect to .

12
. Let X and Y be two sets and suppose that two injective functions
f : X Y, g : Y X
are given. Prove (within ZF theory) that there exist four sets A, B, A
/
, B
/
satisfying the
(a) AA
/
= / 0 and AA
/
= X;
(b) BB
/
= / 0 and BB
/
=Y;
(c) ran( f [A) = B;
(d) ran(g[B
/
) = A
/
.
In order to show the existence of such sets, denote
Z = X g(Y), = g f
and then put
A = Z (Z)
2
(Z) ...
n
(Z) ...,
A
/
= X A, B = f (A), B
/
=Y B.
Finally, verify that g(B
/
) = A
/
.
This result is due to Banach (cf. [150]). It may be regarded as a strengthened form of the
well-known Cantor-Bernstein theorem. Indeed, we can dene a function
h : X Y
by putting h(x) = f (x) if x A and h(x) =g
1
(x) if x A
/
. Clearly, h is a bijection between
the given sets X and Y.
Moreover, if the original functions f and g have nice structural properties, then, in many
cases, h also has the same properties.
For instance, let X and Y be two topological spaces such that X contains a Borel subset
which is Borel isomorphic to Y and, conversely, Y contains a Borel subset which is Borel
isomorphic to X. Check that there exists a Borel isomorphism between X and Y.
Analogously, let X and Y be two topological spaces such that X contains a topological
copy of Y and Y contains a topological copy of X. Show that there exist two sets A X
and B Y such that A is homeomorphic to B and X A is homeomorphic to Y B.
13. Infer the above-mentioned result of Banach from Tarskis xed-point theorem (see
Exercise 10 for Chapter 1).
14
. Let E be a set with card(E) and let I be an ideal of subsets of E such that x
I for all x E. Obviously, this condition is equivalent to the relation E =X : X I.
Let us denote
cov(I) = the smallest cardinality of a covering of E by sets belonging to I.
Recall that a base of I is any family BI such that, for every set X I, there exists a
set from B containing X. We denote
co f (I) = the smallest cardinality of a base of I.
Suppose that the equalities
cov(I) = co f (I) = card(E)
are fullled. Prove that there exists a subset D of E such that:
(a) card(D) = card(E);
(b) card(X D) < card(E) for all sets X I.
Argue in the following manner. Let be the least ordinal number whose cardinality is equal
to card(E). In view of the equality co f (I) = card(E), there exists a base X
: < of
I. Since cov(I) = card(E), we have
card(E X
: <) = card(E)
for each <. Dene by transnite recursion an -sequence x
: < of points of E
such that
x
E (x
: < (X
: <))
for any <. Verify that D =x
: < is the required set.

In particular, take E = R and put:
I
1
= the -ideal of all rst category subsets of R;
I
2
= the -ideal of all -measure zero subsets of R, where as usual denotes the Lebesgue
measure on R.
Under the Continuum Hypothesis, the equalities
cov(I
1
) = co f (I
1
) = card(R) =
1
,
cov(I
2
) = co f (I
2
) = card(R) =
1
are automatically satised. Then the set D
1
for I
1
is called a Luzin subset of R and the set
D
2
for I
2
is called a Sierpi nski subset of R.
Under Martins Axiom, we have
cov(I
1
) = co f (I
1
) = card(R) = c,
cov(I
2
) = co f (I
2
) = card(R) = c.
In this case, the set D
1
for I
1
is called a generalized Luzin subset of R and the set D
2
for
I
2
is called a generalized Sierpi nski subset of R.
Suppose that the ContinuumHypothesis holds. Let L be a Luzin set and let S be a Sierpi nski
set. Show that:
(a) L is universal measure zero;
(b) the outer -measure of any uncountable subset of S is strictly positive (hence, any
uncountable subset of S is nonmeasurable with respect to ).
Suppose that Martins Axiom holds. Let L be a generalized Luzin set and let S be a gener-
alized Sierpi nski set. Show that:
(a) L is universal measure zero;
(b) the outer -measure of any subset X of S with card(X) =c is strictly positive and hence
X is nonmeasurable with respect to .
15. Prove that under Martins Axiomthere exists a countable family X
n
: n < of subsets
of R such that the Lebesgue measure on R cannot be extended to a measure dened on
the -algebra (dom() X
n
: n <).
In connection with this result, let us notice that the following assertion is consistent with
ZFC theory:
for any countable family Y
n
: n < of subsets of R, there exists a measure extending
and dened on (dom() Y
n
: n <).
The consistency of this assertion with the standard axioms of set theory was rst established
by Carlson [36].
16
. A result much stronger than Theorem5 is known. Namely, there exists an uncountable
universal measure zero subset X of R which simultaneously is a vector space over the eld
Q of all rational numbers (see, for example, [197] or Exercise 17 from Appendix 6).
By using the existence of such an X, dene an additive function
f : R R
which is absolutely nonmeasurable with respect to the class of all nonzero -nite
translation-quasi-invariant measures on R (cf. Chapter 5).
17
. Let E be an innite set and let X

i
: i I be a family of subsets of E such that
card(I) card(E), (i I)(card(X
i
) = card(E)).
Prove that there exists a family Y
j
: j J of subsets of E satisfying the following relations:
(a) card(J) > card(E);
(b) card(Y
j
Y
k
) < card(E) for any two distinct indices j J and k J (this circumstance
means that Y
j
: j J is an almost disjoint family of subsets of E);
(c) card(X
i
Y
j
) = card(E) for all indices i I and j J.
For this purpose, construct the required family Y
j
: j J by using the method of trans-
nite recursion.
Putting E = R and taking the family of all nonempty perfect subsets of R as X
i
: i
I, infer from the above result that there exists an almost disjoint family Y
j
: j J of
Bernstein sets in R for which card(J) > c.
18
. For every ordinal number <

1
, dene an injective function
f
: [0, [
and introduce a double family of sets U
,n
: <
1
, n < where
U
,n
= : < <
1
, f
() = n.
Verify the following properties of this family:
(a) U
,n
U
,n
= / 0 for any n < and for any two distinct ordinals <
1
and <
1
;
(b)
1
U
,n
: n < [0, ] for each ordinal <
1
.
Any family
U
,n
: <
1
, n <
of subsets of
1
, which satises the relations (a) and (b), is usually called Ulams (
1
, )-
matrix (over
1
).
Infer from the existence of Ulams matrix U
,n
: <
1
, n < that there is no nonzero
-nite diffused measure on
1
such that
U
,n
: <
1
, n < dom()
and, consequently,
1
In a similar way, construct Ulams (
+1
,
)-matrix of subsets of
+1
, where is an
arbitrary ordinal number, and show that if the cardinal
is not real-valued measurable,

then
+1
is not real-valued measurable, either.
Further, verify that if a
i
: i I is a family of cardinal numbers such that card(I) and all
a
i
(i I) are not real-valued measurable, then the cardinal a
i
: i I is not real-valued
measurable, either.
Derive from these facts that all those cardinal numbers which are strictly less than the rst
weakly inaccessible cardinal are not real-valued measurable.
Under the Generalized Continuum Hypothesis, this result also yields that all those cardinal
numbers which are strictly less than the rst strongly inaccessible cardinal are not real-
valued measurable.
It is well known that the existence of weakly (strongly) inaccessible cardinals cannot be
proved within ZFC theory (see, e.g., [91], [145], [150], and [215]). In view of the said
above, the assumption that there are no real-valued measurable cardinal numbers does not
contradict ZFC theory.
All results of the previous exercise are due to Ulam (see [238]). They are closely connected
with the general measure extension problem (see Chapter 2) and essentially inspired the
development of combinatorial set theory and the theory of large cardinals.
19
. Let E be a topological space satisfying the following conditions:

(a) card(E) =
1
;
(b) every singleton in E is of rst category;
(c) E is of second category;
(d) E has the Suslin property, i.e., any disjoint family of nonempty open sets in E is at most
countable.
Prove that there exists a disjoint family X
: <
1
of subsets of E none of which has
the Baire property.
For this purpose, use Ulams transnite matrix and Exercise 12 from Appendix 3.
20. Let
2
as usual denote the two-dimensional Lebesgue measure on the Euclidean plane
R
2
. Show that if the cardinality of the continuum is real-valued measurable, then there
exists a countably additive functional dened on dom(
2
) such that:
(a) ,=
2
;
(b) (T) = 1 for every square T in R
2
congruent to [0, 1[
2
.
For this purpose, utilize the fact that the real-valued measurability of c implies the existence
of a measure on R extending , with dom() = P(R). By virtue of the classical Vitali
theorem, cannot be invariant under all translations of R. Hence, there are a set X
dom() and an element h R such that
(X +h) ,=(X).
We may assume, without loss of generality, that X is a bounded subset of R and, conse-
quently, X is contained in some closed interval [a, b] R. Now, identifying the real line
with the subset R0 of the plane R
2
, dene a functional on dom(
2
) by the formula
(Z) =
2
(Z) +(Z [a, b]) ((Z[a, b]) +h) (Z dom(
2
)).
Then verify that is countably additive,
2
(X) = 0, (X) =(X) (X +h) ,= 0
and (T) =1 for all those squares T in the plane R
2
which are congruent to the unit square
of this plane.
Derive from the said above that is the required functional (this result is due to Jacab and
Laczkovich [87]).
Conversely, it was proved that the existence of a countably additive functional on
dom(
2
) with properties (a) and (b) implies the real-valued measurability of c (for more
details, see [87]). Therefore, we can conclude that if c is not real-valued measurable, then
every countably additive functional on dom(
2
) satisfying condition (b) necessarily co-
incides with
2
. An easy induction on n yields that the same result remains true for the
Euclidean space R
n
, where n > 2.
On the other hand, it is not difcult to show for R
1
= R that there are many countably
additive functionals on dom(
1
) satisfying the following relations:
(a) ,=
1
;
(b) (T) = 1 for every interval T R whose length is equal to 1.
21. Let L be a Luzin subset of R and let a function f : L R have the Baire property.
Prove that the set f (L) is universal measure zero.
Derive from this fact that for any nonempty perfect set P R, there exists a nonempty
perfect set P
/
P such that P
/
f (L) = / 0; in particular, f (L) is totally imperfect in R.
22. Let S be a Sierpi nski subset of R and let f : RR be a Lebesgue measurable function.
Prove that for any nonempty perfect set P R, there exists a nonempty perfect set P
/
P
such that P
/
f (S) = / 0; in particular, f (S) is totally imperfect in R.
Appendix 2
The Choquet theorem and measurable selectors
Let X and Y be any two sets and let F : X P(Y) be a multi-valued (set-valued) mapping,
that is for any x X, we have F(x) Y. In fact, this situation is equivalent to giving a binary
relation G
F
X Y which is dened as follows:
(x X)(y Y)((x, y) G
F
y F(x)).
This binary relation is usually called the graph of the multi-valued mapping F.
A function f : X Y is called a selector of F if f (x) F(x) for all x X.
The Axiom of Choice states that if F(x) ,= / 0 for any x X, then there always exists a se-
lector of F. However, only this fact is not satisfactory in those situations where some nice
additional structural properties of selectors are needed, for example, continuity, semiconti-
nuity, measurability, Baire property, certain algebraic properties and so on.
In this Appendix, we will envisage the problem of the existence of nice selectors (in the
above-mentioned sense) and will establish several theorems stating that under some natural
assumptions, such selectors always exist. Recall that in Chapter 8 we were concerned with
a rather particular case of this problem. Namely, we proved therein the existence of Borel
selectors for certain partitions of a Polish topological space X. Of course, from the point
of view of classical descriptive set theory, those selectors may be regarded as subsets of X
with good structural properties (see Appendix 6).
Since we are primarily interested in measurability properties of various sets and func-
tions, we will restrict ourselves to the statements on the existence of measurable selectors.
The statements of such a kind are extremely important in analysis, probability theory and
stochastic processes, optimization theory, and the theory of differential equations. There
are many works devoted to this topic (see, e.g., [38], [52], [68], [99], [149], [151], [162],
[188], and especially the excellent book by J.E. Jayne and C.A. Rogers, Selectors, Prince-
ton University Press, Princeton, 2002).
359
We shall concentrate our attention on the following two classical results: Luzin-Jankov-von
Neumann theorem (see, e.g., [99]) and the theorem of Kuratowski and Ryll-Nardzewski
[151] which have a number of applications in different domains of modern mathematics.
Notice that the rst result can easily be deduced from the second one (for more details, see
Theorems 8 and 11 below).
But rst of all, we wish to discuss the well-known Choquet theorem on capacities because
this theorem is closely connected with measurable selectors. We shall formulate and prove
a general version of the Choquet theorem. Then we will present its applications to some
problems concerning measurability of various sets and functions (in particular, selectors).
Let E be a base set and let L be some class of subsets of E. Suppose also that a function
: P(E) R
is given. We shall say that is a capacity on E for the class L (or with respect to L) if
the following three conditions hold:
1) if X Y E, then (X) (Y);
2) if a sequence X
n
: n < of subsets of E is increasing by the inclusion relation, then
(X
n
: n <) = lim
n+
(X
n
);
3) if a sequence X
n
: n < L is decreasing by the inclusion relation, then
(X
n
: n <) = lim
n+
(X
n
).
The next example vividly shows that capacities can be frequently met in measure theory
and related elds of mathematics, such as real analysis, abstract harmonic analysis, and
probability theory.
Example 1. Let (E, S, ) be a measure space equipped with a nite measure . We take
as L any class of sets contained in the -algebra S. Let us dene the function by the
formula
(X) =
(X) (X E),
where
denotes the outer measure associated with . It is easy to check that the function
is a capacity with respect to the class L. As a rule, the class L is taken in such a way
that S =(L); in other words, it is usually assumed that L generates S.
Let L be a class of subsets of E. In our further considerations, we denote by A(L) the
analytic class generated by L (see Chapter 8 or Appendix 6). The symbol L
will denote
the class of the intersections of all countable families of sets from L.
Appendix 2. The Choquet theorem and measurable selectors 361
The following fundamental result is due to Choquet (cf. [24], [40], [52], and [187]).
Theorem 1. Let E be a base set and let L be some class of subsets of E, closed under nite
unions and nite intersections. Furthermore, let be any capacity on E with respect to L.
Then for every set X from the analytic class A(L), the following relation holds:
(X) = sup(Y) : Y L
&Y X.
Proof. As usual, we denote by
<
the family of all nite sequences of natural numbers.
Take any set X A(L). Obviously, we can write
X =
k<
F
t
0
...t
k
,
where F
t
: t
<
is some countable system of sets fromthe given class L (see Chapter
8).
Since L is closed under nite intersections, we can assume (without loss of generality)
that the above-mentioned system of sets is regular, i.e.,
F
t
0
...t
k
t
k+1
F
t
0
...t
k
(k <).
For any innite sequence
t = (t
0
, t
1
, . . . ,t
k
, ...)
,
let us put
F
t
= F
t
0
F
t
0
t
1
... F
t
0
t
1
...t
k
... .
It should be mentioned that F
t
may not belong to L. At the same time, it is clear that the
following equality holds:
X =
n<
(
t:t
0
n
F
t
).
Now, taking into account the properties 1) and 2) of the capacity , for any real number
> 0, we can nd a natural index n
0
such that
(X) <(G
0
),
where the set G
0
is dened by the formula
G
0
=
t:t
0
n
0
F
t
.
Starting with this G
0
, we continue our construction by ordinary recursion. Suppose that a
nite sequence (n
0
, n
1
, . . . , n
k
) of natural numbers and a nite sequence (G
0
, G
1
, . . . , G
k
) of
sets are constructed in such a way that
G
r
=
t:t
0
n
0
,...,t
r
n
r
F
t
(r = 0, . . . , k)
and the relations
(X) <(G
r
) (r = 0, . . . , k)
are satised. Let us consider the set G
k
. Obviously, we can write
G
k
=
n<
(
t:t
0
n
0
,...,t
k
n
k
,t
k+1
n
F
t
).
From the inequality
(X) <(G
k
)
and the properties 1) and 2) of it follows again that there exists a natural index n
k+1
such
that for the set
G
k+1
=
t:t
0
n
0
,...,t
k
n
k
,t
k+1
n
k+1
F
t
,
the inequality
(X) <(G
k+1
)
holds true, too. Proceeding in this way, we will be able to construct the required innite
sequences
(n
0
, n
1
, . . . , n
k
, ...), (G
0
, G
1
, . . . , G
k
, ...).
Further, for every natural number k, we put
H
k
=
t
0
n
0
,...,t
k
n
k
F
t
0
...t
k
.
Obviously, in the above formula we use the union of nitely many sets. Therefore, H
k
L.
Moreover, it is easy to check that the sequence H
k
: k < is decreasing with respect to
the inclusion relation and
G
k
H
k
(k = 0, 1, 2, ...).
Condition 3) for implies that
(X) inf(G
k
) : k < lim
k+
(H
k
) =(H
k
: k <).
Now, applying the regularity of the systemF
t
: t
<
and the K onig lemma on -trees
with nite levels (see Appendix 1), we readily get the inclusion
H
k
: k < X.
So, denoting H =H
k
: k <, we conclude that the set H belongs to the class L
and
H X, (X) (H).
Since > 0 was taken arbitrarily, we obtain the desired result.
The theorem just proved can be formulated as follows.
All analytic sets over the original class L (i.e., all sets from A(L)) are capacitable with
respect to any capacity for L.
Example 2. Let (E, S, ) be a measure space with a nite (or, more generally, -nite)
measure and let L be any class of subsets of E, which generates the -algebra S.
According to Example 1, the function
is a capacity for the class L. Therefore, by the

Choquet theorem, any analytic set over L is capacitable with respect to
. But, as it is
easy to check, this fact means the following:
any analytic set over L is measurable with respect to the completion
/
of the measure .
It is also clear that if the original measure is complete, then we simply obtain the -
measurability of any analytic set over the class L.
Now, let E be an arbitrary Polish topological space, be an arbitrary nite (or, more
generally, -nite) Borel measure on E and let L be the family of all closed subsets of
E. Applying the above considerations to this particular case, we see that every analytic
(Suslin) subset of E is
/
-measurable.
If E = R, then we immediately obtain the old classical result which says that all analytic
subsets of R are Lebesgue measurable. Hence, all complements of the analytic subsets of
R are also Lebesgue measurable. Furthermore, we readily derive that the sets belonging
to the -algebra generated by the family of all analytic subsets of R are good from the
measure-theoretical point of view. Notice that the same sets are good from the topological
point of view because they have the Baire property and, moreover, they have the Baire
property in the restricted sense (in this connection, see [99] or [148]).
Let us give some other applications of the Choquet theorem.
Theorem 2. Let (E, S, ) be a complete probability space (or, more generally, a complete
measure space with a -nite measure ). Let K be a nonempty locally compact topolog-
ical space with a countable base and let (K, B(K)) be the measurable space canonically
associated with K (i.e., K is equipped with its Borel -algebra B(K)). Denote by
pr
1
: E K E
the canonical projection of E K onto E. Then for any set Z belonging to the product
-algebra S B(K), the set pr
1
(Z) belongs to the -algebra S.
Proof. We may assume, without loss of generality, that is a probability measure.
Let us consider the family Comp(K) of all compact subsets of the given topological space
K. Evidently, the class Comp(K) is closed under nite unions and nite intersections. It is
also clear that the Borel -algebra B(K) is generated by Comp(K).
In the product set E K consider the family of all sets of the form
X
i
Y
i
: 1 i m,
where m is an arbitrary natural number, X
i
are elements of the -algebra S and Y
i
are
members of Comp(K). We denote the class of all sets described in this manner by the
symbol L. The class L is closed under nite unions and nite intersections, too. It
is also obvious that the -algebra generated by L coincides with the product -algebra
S B(K). Let us dene a real-valued function on the family of all subsets of E K by
the formula
(Z) =
(pr
1
(Z)) (Z E K).
We are going to verify that the function is a capacity on E K with respect to the class
L. Indeed, it is clear that
Z
1
Z
2
(Z
1
) (Z
2
)
for any two sets Z
1
E K and Z
2
E K. Moreover, if Z
n
: n < is an increasing
sequence of subsets of E K and Z =Z
n
: n <, then we have
pr
1
(Z) =pr
1
(Z
n
) : n <,
(Z) =
(pr
1
(Z)) = lim
n+
(pr
1
(Z
n
)) = lim
n+
(Z
n
).
The latter formula immediately follows from standard properties of the outer measure
associated with the given probability measure .

Further, we must check that if Z
n
: n < is a decreasing sequence of sets from the class
L, then
(Z
n
: n <) = lim
n+
(Z
n
).
For this purpose, let us rst prove the equality
pr
1
(Z
n
: n <) =pr
1
(Z
n
) : n < ().
In fact, it is sufcient to establish the inclusion
pr
1
(Z
n
) : n < pr
1
(Z
n
: n <)
because the converse inclusion does always hold. Let x be an arbitrary element of
pr
1
(Z
n
) : n < . This circumstance means that for any natural number n, there ex-
ists a point y
n
such that
(x, y
n
) X
n,i(n)
Y
n,i(n)
,
where X
n,i(n)
Y
n,i(n)
is some component of the representation of the set Z
n
in the form
Z
n
=X
n,i
Y
n,i
: 1 i m(n).
In particular, we have y
n
Y
n,i(n)
and hence Y
n,i(n)
,= / 0. Moreover, since the sequence of
sets Z
n
: n < is decreasing by the inclusion relation, we may assume without loss of
generality that the families of sets
X
n+1,i
: 1 i m(n +1), Y
n+1,i
: 1 i m(n +1)
are respectively inscribed in the families
X
n,i
: 1 i m(n), Y
n,i
: 1 i m(n).
Now, applying the K onig lemma on -trees with nite levels (see Appendix 1), we can
nd a decreasing sequence of sets
X
n,i(n)
Y
n,i(n)
: n <
such that all the sets Y
n,i(n)
are nonempty and the relation x X
n,i(n)
holds true for every
n <. Since all Y
n,i(n)
are also compact and decrease by inclusion, we obviously have
Y
n,i(n)
: n < ,= / 0.
Let y be any element of Y
n,i(n)
: n <. Then it is clear that
(x, y) Z
n
: n <
and, consequently,
x pr
1
(Z
n
: n <).
Thus, the required equality () is proved.
Now, taking into account the -measurability of all sets from the family
pr
1
(Z
n
) : n <,
we can write
(Z
n
: n <) =
(pr
1
(Z
n
: n <)) = (pr
1
(Z
n
) : n <) =
lim
n+
(pr
1
(Z
n
)) = lim
n+
(Z
n
).
In this way, we have checked that the function is a capacity with respect to the class L.
The Choquet theorem implies that any set from the class A(L) is capacitable with respect
to . Since K is a locally compact topological space with a countable base, the inclusion
S B(K) A(L) holds (see Appendix 6). Consequently, for any set Z S B(K),
we have
(Z) = sup(D) : D L
& D Z.
In other words, we obtain the equality
(pr
1
(Z)) = sup(pr
1
(D)) : D L
& D Z.
This equality immediately yields the measurability of the set pr
1
(Z) with respect to the
original measure which ends the proof.
The statement just established is sometimes called the theorem on measurable projection.
Now, we are ready to consider an application of the Choquet theorem to certain questions
about the existence of measurable selectors. First, we introduce one denition and prove
an auxiliary assertion.
Let (E, S, ) be a complete probability space or, more generally, a complete measure
space with a -nite measure , let K be a locally compact space with a countable base,
and let B(K) as usual denote the Borel -algebra of K.
Let Z be a subset of the Cartesian product E K. We say that Z is a measurable functional
graph if the following two conditions are satised:
(a) Z S B(K);
(b) for every e E, the corresponding section
Z(e) =k K : (e, k) Z
contains at most one point.
The next result gives a characterization of measurable functional graphs in terms of mea-
surable partial mappings acting from E into K (cf. [52]).
Theorem 3. Let Z E K. The following two relations are equivalent:
1) Z is a measurable functional graph in E K;
2) there exist a set X S and a measurable (with respect to the -algebras B(K) and S)
mapping g : X K such that the equality
Z =(x, y) X K : g(x) = y
holds true.
Proof. First, let us prove the implication 2) 1). Let relation 2) be valid. Consider a
mapping
: X K K K
(x, y) = (g(x), y) (x X, y K).
Observe that is measurable with respect to the product -algebras
B(K) B(K) =B(K K), (S[X) B(K),
where S[X denotes the restriction of the -algebra S to the set X, i.e.,
S[X =X Y : Y S.
Analogously, the diagonal (y, y) : y K of the product space K K is measurable with
respect to the -algebra B(K) B(K). Hence, the pre-image
1
((y, y) : y K) is
measurable with respect to (S[X) B(K). But it is obvious that
1
((y, y) : y K) =(x, y) X K : g(x) = y = Z.
These equalities immediately yield that Z is a measurable graph in the product set E K.
Now, we are going to prove the implication 1) 2). Suppose that relation 1) holds. Let us
put X = pr
1
(Z). The preceding theoremon measurable projection says that X is measurable
with respect to the -algebra S. Let us dene a mapping g : X K in the following way:
for every x X, the value g(x) is equal to the unique point from the intersection Z(x
K).
Let us check that this mapping is measurable with respect to the -algebras B(K) and
S[X. For this purpose, take any Borel subset Y of K. Then we can write
g
1
(Y) = pr
1
(Z (E Y)).
But it is clear that the set Z (E Y) is measurable with respect to the product -algebra
S B(K). Using the theorem on measurable projection once more, we obtain that the set
g
1
(Y) is measurable with respect to the -algebra S, so the measurability of our mapping
g is established which completes the proof of the theorem (cf. Exercise 6 from Appendix
1).
Among various situations in which the statements just presented may be efciently applied,
the most important is the case when a given topological space K coincides either with the
real line R or with the positive half-line
R
+
=x R : x 0.
It should be noticed that the last case can be frequently met in the theory of stochastic
processes (see, for example, [52], [187], and [205]). It will be convenient to formulate
some results below in terms of the half-line R
+
. Taking into account the existence of
a Borel isomorphism between R
+
and any uncountable Polish topological space K (see
Appendix 6), it can easily be shown that analogous results remain valid for K, too. As
usual, we assume that the half-line R
+
is equipped with its Borel -algebra B(R
+
).
Consider any complete probability space (E, S, ) or, more generally, a complete measure
space with a -nite measure . Let Z be a subset of the Cartesian product E R
+
such
that pr
1
(Z) = E.
In the sequel, we shall say that the function
d
Z
: E R
+
d
Z
(e) = inft R
+
: (e, t) Z (e E)
is the debut of Z.
The following easy proposition is true.
Theorem 4. Let (E, S, ) be a complete probability space (or, more generally, a complete
measure space with a -nite measure ) and let a subset Z of E R
+
be measurable with
respect to the product -algebra S B(R
+
). Suppose also that pr
1
(Z) = E. Then the
debut d
Z
of Z is a real-valued function measurable with respect to the -algebras B(R
+
)
and S.
Proof. Indeed, for any real number t 0, the pre-image
d
1
Z
([0, t[) =e E : d
Z
(e) <t
coincides with the set pr
1
(Z (E [0, t[)). The set Z (E [0, t[) is measurable with
respect to the -algebra S B(R
+
). By the theorem on measurable projection, the set
d
1
Z
([0, t[) is measurable with respect to S. Hence, the function d
Z
is measurable with
respect to B(R
+
) and S.
The next result is a typical representative froma large group of the so-called uniformization
theorems or, equivalently, theorems on measurable selectors (see [52]).
Theorem 5. Let (E, S, ) be any complete probability space (or, more generally, a com-
plete measure space with a -nite measure ) and let a set Z E R
+
be measurable
with respect to the product -algebra S B(R
+
). Suppose also that pr
1
(Z) = E. Then
there exists a function g : E R
+
such that:
1) g is measurable with respect to the -algebras B(R
+
) and S;
2) the graph Gr(g) of g is contained in the given set Z.
Proof. It sufces to show that for every real number > 0, there exist a set X
S and a
measurable function g
: X R
+
such that the following two conditions hold:
(a) (E X
) <;
(b) the graph of g
is contained in Z.
Indeed, as soon as this fact is established, the next steps of the proof become evident: after
nding a set X
and a function g
, we move to the measure space

(E X
, S[(E X
), [(S[(E X
)))
and to the set
Z ((E X
) R
+
),
and continue analogous process for the real number /2, the new measure space and the
new set. After countably many steps, we will ll the whole space E except, perhaps, a
-measure zero set with a disjoint countable family of the constructed -measurable sets.
Then we obtain the required function g as a common extension of all functions of type g
and some function dened on a -measure zero subset of E.

So, let be any strictly positive real number. Consider the class Comp(R
+
) of all compact
subsets of the half-line R
+
. Moreover, let us introduce the class L of all those sets Z
E R
+
which can be represented in the form
Z =X
i
Y
i
: 1 i m,
where m is an arbitrary natural number, X
i
are members of S and Y
i
are members of
Comp(R
+
). Earlier, we have already mentioned that the class L generates the product
-algebra S B(R
+
) and that L is closed under nite unions and nite intersections.
Further, it is easy to see that
S B(R
+
) A(L).
Using the same argument as in the proof of the theorem on measurable projection, we can
nd for the given and Z, a set Z
Z belonging to the class L
and such that

(pr
1
(Z) pr
1
(Z
)) = (E pr
1
(Z
)) <.
For the set Z
, we may consider its debut

d
Z
: pr
1
(Z
) R
+
which is dened on the set pr
1
(Z
). Notice that for each element x pr

1
(Z
), the section
Z
(x) =t R
+
: (x, t) Z
is a nonempty compact subset of R

+
because it coincides with the intersection of some
countable family of nonempty compact subsets of R
+
. Consequently, we have
d
Z
(x) = inf(Z
(x)) Z
(x).
From this fact it immediately follows that the graph of the function d
Z
is contained in the
set Z
, so it is contained in the original set Z as well. It remains to take into account that, in
view of Theorem 4, the function d
Z
is measurable with respect to the -algebras B(R
+
)
and S. Therefore, we can put
X
= pr
1
(Z
), g
= d
Z
,
and Theorem 5 has thus been proved.
It is obvious that the notion of the debut of a given set lying in the Cartesian product
E R
+
can be generalized to those subsets of E R
+
whose rst projections are not
necessarily the whole space E. After this generalization it becomes clear that if a given
set Z E R
+
is measurable with respect to the product -algebra S B(R
+
), then its
debut is a measurable function acting from the set pr
1
(Z) into R
+
. This fact also indicates
how we can generalize Theorems 4 and 5 to the case of any set Z E R
+
measurable
with respect to the product -algebra S B(R
+
). Therefore, using the same argument as
in the proof of Theorem 5, we obtain the following classical result.
Theorem 6. Let E be a Polish topological space and let Z be an analytic subset of the
topological product E R
+
. Suppose that a Borel nite (or, more generally, -nite)
measure on E is given and let
/
denote the completion of . Then there exists a function
g : pr
1
(Z) R
+
such that:
1) g is measurable with respect to the -algebras B(R
+
) and dom(
/
);
2) the graph of g is contained in the set Z.
A detailed proof of this statement is left to the reader.
Now, let P be an arbitrary uncountable Polish topological space. As is well known (see,
e.g., [40], [99], [148], or Appendix 6), there exists a Borel isomorphism
h : (R
+
, B(R
+
)) (P, B(P)).
It is not difcult to check that for this isomorphism, the equality
h(A(R
+
)) =A(P)
holds true. Taking into account this trivial remark, we can immediately deduce from The-
orem 6 the following result.
Theorem 7. Let E
1
and E
2
be two Polish topological spaces and let Z be an analytic subset
of the topological product E
1
E
2
. Suppose also that a Borel nite (or -nite) measure
on E
1
is given and let
/
denote the completion of . Then there exists a mapping
g : pr
1
(Z) E
2
such that:
1) g is measurable with respect to the -algebras B(E
2
) and dom(
/
);
2) the graph of g is contained in the set Z.
In many cases, the theorem formulated above is sufcient for various applications in mea-
sure theory and real analysis. But, sometimes, we need more subtle formulations of results
of this type (see, for instance, the theorem of Kuratowski and Ryll-Nardzewski below).
Now, let us turn our attention to set-valued mappings with those measurability properties
which are not directly connected with the notion of measure.
First, let us recall that the pair (X, S) is an abstract measurable space if X is a base set and
S is some -algebra of subsets of X.
For instance, if E is an arbitrary topological space, then the pair (E, B(E)) is a canonical
example of a measurable space. In this context, it should be mentioned that there are
some uncountable subspaces E of R for which B(E) does not admit any nonzero -nite
diffused measure (see Appendix 1).
Suppose that (X, S) is a measurable space.
Let Y be a topological space and let F : X P(Y) be a set-valued mapping such that F(x)
is a nonempty closed subset of Y for every point x X. In the sequel, only those set-valued
functions F will be considered which satisfy this condition.
We say that F is a weakly measurable set-valued mapping if for any open subset U of Y,
we have
x X : F(x) U ,= / 0 S.
It is not difcult to check the validity of the following auxiliary proposition.
Lemma 1. Let (Y, d) be a separable metric space. For a set-valued mapping F : X
P(Y), these two assertions are valid:
1) F is weakly measurable if and only if for each point y Y, the function g
y
: X R
g
y
(x) = d(F(x), y) (x X)
is measurable with respect to the -algebras B(Y) and S;
2) if F is weakly measurable, then the graph of F is a measurable subset of the product
space (X Y, S B(Y)).
The proof of Lemma 1 is left to the reader.
Let f be a mapping acting fromX into Y. We recall that f is a selector of a given set-valued
mapping F : X P(Y) if for any element x X, we have f (x) F(x). In view of the
relation
(x X)(F(x) ,= / 0),
the Axiom of Choice immediately yields the existence of a selector of F. But if we want to
get a measurable selector, then we must use a more delicate additional argument.
For our further purposes, we also need the next simple auxiliary proposition.
Lemma 2. Let (X, d) be an arbitrary metric space. Then there exist a Banach space
(E, [[ [[) and an isometric embedding
: (X, d) (E, [[ [[).
In particular, if (X, d) is a complete metric space, then the image (X) is a closed subset
of the space E. Moreover, the space E can be chosen in such a way that w(E) w(X) +,
where w(E) (respectively, w(X)) denotes the topological weight of E (respectively, of X).
Proof. Without loss of generality, we may assume that X is a nonempty space. Let t be a
xed element of X. Further, for each element y X, let us consider a mapping f
y
: X R
f
y
(x) = d(x, y) d(x, t) (x X).
It is easy to see that
[ f
y
(x)[ d(y, t) (x X).
Hence the function f
y
is bounded. Now, we can consider the family E of all bounded
functions f acting from X into R. Equip E with the standard norm
[[ f [[ = sup
xX
[ f (x)[.
Obviously, E is a Banach space with respect to this norm. Moreover, it is not difcult
to check that [[ f
y
f
z
[[ = d(y, z) for any two elements y and z from X. Consequently, a
mapping : X E dened by the formula
(y) = f
y
(y X)
is an isometric embedding of X into E. A priori, the topological weight of E may be
signicantly greater than the topological weight of X. But if we take the closed vector
subspace of E generated by the set (X), then the weight of this subspace does not exceed
w(X) +. Lemma 2 is thus proved.
Notice that the same argument leads to the following classical result:
any metric space (X, d) can be isometrically embedded into some complete metric space E
in such a way that X becomes everywhere dense in E.
Let us recall that this result was rst obtained by Hausdorff (see, e.g., [58], [101], and
[148]). We also wish to remark that in the theory of metric spaces there are theorems
much stronger than Lemma 2. For instance, if (X, d) is an arbitrary separable metric space,
then there exists an isometric embedding of X into the separable Banach space C[0, 1] of
all continuous real-valued functions dened on the unit segment [0, 1]. In other words,
the separable Banach space C[0, 1] turns out to be universal for the class of all separable
metric spaces. Recall that this important result is due to Banach and Mazur (see, e.g., [56]).
However, for our further considerations, Lemma 2 is completely sufcient.
Now, we can formulate and prove a famous theoremconcerning the existence of measurable
selectors. This important theorem was established by Kuratowski and Ryll-Nardzewski in
[151].
Theorem 8. Let (X, S) be a measurable space, let Y be a Polish space and let F : X
P(Y) be a set-valued mapping satisfying the following conditions:
1) for each element x X, the set F(x) is nonempty and closed in Y;
2) F is weakly measurable.
Then there exists a mapping f : X Y such that:
a) f is a selector of F;
b) f is a measurable mapping acting from (X, S) into (Y, B(Y)).
Proof. Taking into account Lemma 2, we may assume without loss of generality that Y is
a separable Banach space. Let d be a complete metric on Y agreed with the norm of Y; in
other words, for any two points y and z fromY, we put
d(y, z) =[[y z[[.
Furthermore, denote by y
k
: 1 k < a countable family of points of Y which is every-
where dense in this space.
Using the ordinary method of mathematical recursion, we shall construct a sequence f
n
:
1 n < of mappings acting from X into Y and satisfying the following relations:
(1) ran( f
n
) y
k
: 1 k < for any natural number n 1;
(2) d( f
n
(x), f
n1
(x)) < 1/2
n1
for each natural number n 2 and for each element x X;
(3) d(F(x), f
n
(x)) < 1/2
n1
for any natural number n 1 and for any element x X;
(4) f
n
is a measurable mapping acting from (X, S) into (Y, B(Y)) for any natural number
n 1.
Let x be an arbitrary point of the space X. Then we put f
1
(x) = y
k
where k is the smallest
natural number for which d(F(x), y
k
) < 1. Evidently, we obtain a certain mapping
f
1
: X Y.
Since F is a weakly measurable set-valued mapping, it is easy to check that f
1
is a measur-
able mapping acting from (X, S) into (Y, B(Y)).
Suppose now that a partial nite sequence of functions
f
1
, f
2
, ... , f
n
satisfying the above relations (1) - (4) has already been dened. Fix again an arbitrary
element x X and consider the following two sets:
V =y Y : d(y, f
n
(x)) < 1/2
n
,
W =y Y : d(y, F(x)) < 1/2
n
.
Clearly, V and W are open subsets of the space Y. Moreover, since
d(F(x), f
n
(x)) < 1/2
n1
,
there exists a point z F(x) such that d(z, f
n
(x)) < 1/2
n1
. Now, it can easily be seen that
the point
y = (z + f
n
(x))/2
belongs to the set V W. Hence V W is a nonempty open subset of Y. Let k be the
smallest natural number for which y
k
V W. Let us put
f
n+1
(x) = y
k
.
Taking into account the fact that x is an arbitrary element of X, we have a certain mapping
f
n+1
: X Y.
Since our set-valued mapping F is weakly measurable, it is not difcult to check that f
n+1
is a measurable mapping acting from (X, S) into (Y, B(Y)). Also, it is not hard to show
that all relations (1) - (4) remain true for the partial sequence
f
1
, f
2
, ... , f
n
, f
n+1
.
Proceeding in such a way, we are able to construct the required innite sequence of func-
tions
f
1
, f
2
, ... , f
n
, ... .
Finally, we put
f (x) = lim
n+
f
n
(x) (x X).
Notice that this denition of f is correct because, in view of relation (2), the sequence of
functions f
n
: 1 n < uniformly converges on X.
Thus, we have a mapping f acting from X into Y. Now, using relation (3), we can readily
conclude that f is a measurable selector of the original set-valued mapping F. The proof
of Theorem 8 is completed.
Some applications of Theorem8 may be found in [99], [149], and [151] (in this connection,
see also [38] and [68]). Here we wish to present only two statements which can be obtained
with the aid of Theorem 8. The rst of these statements will be formulated in terms of
topological measure theory. The second one will be formulated in terms of Baire category.
Theorem 9. Let E
1
and E
2
be two Polish topological spaces and let A be an analytic subset
of the product space E
1
E
2
. Suppose that E
1
is equipped with a -nite Borel measure
and suppose that for almost all (with respect to ) points x pr
1
(A), the section
A(x) =y E
2
: (x, y) A
is an uncountable subset of E
2
. Then there exists a set B A such that:
1) B is a Borel subset of the space E
1
E
2
;
2) for almost all (with respect to ) points x pr
1
(A), the section
B(x) =y E
2
: (x, y) B
is a nonempty perfect compact subset of E
2
.
Theorem 10. Let E
1
and E
2
be two Polish topological spaces and let A be an analytic
subset of the product space E
1
E
2
. Suppose that for almost all (in the Baire category
sense) points x pr
1
(A), the section A(x) is an uncountable subset of E
2
. Then there exists
a set B A such that:
1) B is a Borel subset of the space E
1
E
2
;
2) for almost all (in the Baire category sense) points x pr
1
(A), the section B(x) is a
nonempty perfect compact subset of E
2
.
As mentioned above, the proofs of Theorems 9 and 10 are based on the theorem of Kura-
towski and Ryll-Nardzewski (for more details, see Exercises 12 and 13 of this chapter).
It is easy to check that both Theorems 9 and 10 may be treated as some parameterized
versions of the classical result of Alexandrov and Hausdorff stating that any uncountable
analytic set (in a Polish topological space) contains a nonempty perfect compact subset (see
[99], [148], [160], [162], or Appendix 6).
Furthermore, from the theorem of Kuratowski and Ryll-Nardzewski we can easily obtain
another classical result concerning the uniformization of an analytic subset of the product
of two Polish topological spaces. This result is due to Luzin, Jankov and von Neumann
(cf. [99]). It essentially strengthens Theorem 7, which was obtained by using the Choquet
theorem.
Theorem 11. Let E
1
and E
2
be two Polish spaces, let A be an analytic subset of the product
space E
1
E
2
and let S be the -algebra of subsets of pr
1
(A), generated by the family of
all analytic subsets of pr
1
(A). Then there exists a mapping
f : pr
1
(A) E
2
1) the graph Gr( f ) of f is contained in the set A;
2) f is a measurable mapping acting from (pr
1
(A), S) into (E
2
, B(E
2
)).
In particular, f is measurable with respect to the completion of an arbitrary -nite Borel
measure given on the Suslin space pr
1
(A) and f has the Baire property in the restricted
sense.
Proof. Let us equip the set of all natural numbers with the discrete topology and take the
product space
which usually is called the canonical Baire space of countable topologi-

cal weight. Since our set A is analytic, there exists a continuous function
g :
E
1
E
2
such that g(
) = A (see Exercise 12 from Chapter 8).

Denote by G the graph of g, i.e., put
G =(t, x, y)
E
1
E
2
: g(t) = (x, y).
Obviously, G is a closed subset of the product space
E
1
E
2
(see Theorem 2 from
Chapter 8). Now, we dene a set-valued mapping
: pr
1
(A) P(
E
1
E
2
)
by the formula
(x) = G(
x E
2
) (x pr
1
(A)).
It is not difcult to check that satises the assumptions of the theorem of Kuratowski and
Ryll-Nardzewski. According to this theorem, there exists a selector
: pr
1
(A)
E
1
E
2
of which is measurable with respect to the -algebras B(
E
1
E
2
) and S. It
remains to put
f = pr
3

and to verify that f is the required mapping. The corresponding easy details are left to the
reader.
The assumption that Y is a Polish topological space is very essential in the formulation
of the theorem of Kuratowski and RyllNardzewski. However, in some particular cases,
the result of this theorem remains true if we replace Y by a nonseparable complete metric
space Y
/
. Moreover, sometimes we do not need even the completeness of the space Y
/
, for
example, in those cases where all sets F(x) are compact and nonempty.
Various situations closely connected with the theorem of Kuratowski and Ryll-Nardzewski
are discussed in monograph [68] where it is shown that some additional set-theoretical
axioms are necessary for appropriate generalizations of this theorem (see also [38]).
EXERCISES
1. Let E be any base set and let L be some class of subsets of E.
We say that L is a quasicompact class (in the sense of Marczewski) if for every countable
family Z
n
: n < L, the relation
Z
n
: n < = / 0
implies that there exists a nite subfamily Z
n
1
, Z
n
2
, . . . , Z
n
k
of Z
n
: n < such that
Z
n
1
Z
n
2
... Z
n
k
= / 0.
For example, if E is an arbitrary Hausdorff topological space and L is the class of all
compact subsets of E, then L is a quasicompact class of sets.
Now, let L be any quasicompact class of subsets of a base set E and let L
denote the
class of all nite unions of elements fromL. Prove that L
is also a quasicompact class.

For this purpose, apply the K onig lemma on -trees with nite levels (see Appendix 1.)
2. Give generalizations of some results presented in this Appendix to the case when a
locally compact topological space K with a countable base is replaced by a set F equipped
with a quasicompact class L P(F).
3. Give a detailed proof of Theorem 6 of this chapter and deduce Theorem 7 from it.
4
. Starting with the fact that there are two disjoint co-Suslin sets in R which cannot be
separated by any two disjoint Borel subsets of R (see [99], [148], [160], and [162]), prove
that there exists a Borel set Z R
2
such that:
(a) pr
1
(Z) = R;
(b) there is no Borel function f : R R whose graph is contained in Z.
This deep result is due to Luzin and Novikov (cf. [162], [188]) and shows that for the class
of all Borel subsets of the plane R
2
, the uniformization problem has a negative solution if a
selector is required to be Borel. This fact also explains why in Theorem 5 of this Appendix
we should consider measurable (but not Borel) functions. As follows from the theorem of
Kuratowski and Ryll-Nardzewski, we only have the guaranty that a selector can be chosen
to be measurable with respect to the -algebras B(R) and S, where S is generated by
the family of all analytic subsets of R.
5. Let (E, S) be a measurable space and let X be a subset of E. We say that X is absolutely
(or universally) measurable with respect to the -algebra S if for every -nite measure
dened on S, the relation X dom(
/
) holds, where
/
Deduce from the Choquet theorem that all members of the analytic class A(S) are abso-
lutely measurable with respect to the -algebra S.
In particular, if E is an arbitrary topological space, then all analytic sets in E (i.e., the
members of A(F) where F denotes the class of all closed subsets of E) are absolutely
measurable with respect to the Borel -algebra B(E).
Moreover, let us consider the particular case
(E, S) = (R, B(R)).
In this case, give an example of a Lebesgue measurable subset of R which is not absolutely
measurable with respect to the Borel -algebra B(R).
6. Let E be an arbitrary Hausdorff topological space and let X be a subset of E which is a
Radon space with respect to the induced topology. Show that X is absolutely measurable
with respect to the Borel -algebra B(E).
Let E be an arbitrary Radon topological space and let X be a subset of E which is absolutely
measurable with respect to the Borel -algebra B(E). Show that the set X equipped with
the induced topology is a Radon topological space.
Conclude that every analytic subset of a Polish topological space P is a Radon space. The
same is true for the complement of any analytic subset of P and, more generally, for each
member of the -algebra generated by the family of all analytic sets in P.
7. Assuming Martins Axiom, show that there exists a subset of the real line R, which
is absolutely measurable with respect to the Borel -algebra B(R) but does not have the
Baire property in R.
Conversely, give an example (within ZFC theory) of a rst category subset of R which is
not absolutely measurable with respect to B(R).
8
. Let us consider the measure space (R, dom(), ) where is the standard Lebesgue
measure on R. Let P be an arbitrary projective subset of R and let B(P) denote the Borel
-algebra of the topological space P. Let us put
S = dom() B(P)
and consider the canonical projection
pr
1
: RP R.
Assuming that all projective subsets of R are Lebesgue measurable, prove that
(X S)(pr
1
(X) dom()).
Similarly, let us put
L =Ba(R) B(P),
where Ba(R) is the -algebra of all those subsets of R which have the Baire property.
Assuming that all projective subsets of R have the Baire property, prove that
(X L)(pr
1
(X) Ba(R)).
Recall that the Lebesgue measurability and the Baire property of all projective subsets of
R follow from the Axiom of Projective Determinacy (see, e.g., [10], [91], and [99]).
9
. Let X be an analytic subset of a Polish topological space, let E be a metric space

equipped with a -nite Borel measure , and let
f : X E
be a Borel mapping. Show that for every real number > 0, there exists a subset Y of X
such that:
(a) Y is a compact subspace of X;
(b) the restriction of f to Y is a homeomorphismbetween the spaces Y and f (Y);
(c)
/
( f (X) f (Y)) <, where
/
10. Prove Lemma 1 of this chapter.
11
. Give a proof of Theorem 9 from this chapter.

For this purpose, consider the family C = C(E
2
) of all nonempty compact subsets of E
2
and endow C with the Vietoris topology (see [58], [101], [149]). Verify that the obtained
topological space is Polish. Then dene an appropriate analytic subset D of the product
space E
1
C and apply to D the Luzin-Jankov-von Neumann theorem.
12
. Give a proof of Theorem 10 from this chapter.

Use the same hint as in the previous exercise.
13
. This exercise deals with extensions of Carath eodory type partial functions which are
important in various questions of real analysis, measure theory, the theory of ordinary dif-
ferential equations, optimization, and probability theory.
Let X be a set equipped with a -nite complete measure and let Y be a compact metric
space. Suppose that a partial mapping
f
0
: X Y R
is measurable with respect to the -algebras B(R) and dom() B(Y).
By using the theorem on measurable projection and the theorem of Kuratowski and Ryll-
Nardzewski, prove that the following two relations are equivalent:
(a) for each x X, the partial mapping f
0
(x, ) : Y R is uniformly continuous;
(b) there exists a mapping f : X Y R extending f
0
and satisfying Carath eodorys con-
ditions, i.e., for each x X, the mapping f (x, ) : Y R is continuous and for each y Y,
the mapping f (, y) : X R is -measurable.
In order to establish the equivalence (a) (b), consider the separable Banach space C(Y) =
C(Y, R) of all real-valued continuous functions on Y and dene an appropriate set-valued
weakly measurable mapping
F : X P(C(Y))
associated with the given mapping f
0
. Then apply to F the theorem of Kuratowski and
Ryll-Nardzewski.
14
. Let E be a Suslin space and let S be a countably generated subalgebra of the Borel
-algebra B(E). Suppose that is a probability measure with dom() =S. Demonstrate
that there exists a Borel measure
/
on E extending .
For this purpose, use an appropriate theorem on the existence of measurable selectors and
Marczewskis characteristic function (cf. the proof of Theorem 5 from Appendix 1).
The result of Exercise 14 is due to Ershov (see [64]). It can be shown that, in general, the
same result fails to be true for a co-Suslin subspace E of a Polish topological space.
15. Assuming Martins Axiom, prove that there exists a subset Z of the Euclidean plane
R
2
, satisfying the following relations:
(a) for each point x R, the cardinality of the corresponding section Z(x) =y R: (x, y)
Z is equal to c;
(b) every selector of the set-valued mapping : R P(R) dened by
(x) = Z(x) (x R)
is absolutely nonmeasurable with respect to the class of all nonzero -nite diffused mea-
sures on R.
For this purpose, consider a disjoint family L
x
: x R of generalized Luzin sets in R such
that L
x
: x R is also a generalized Luzin subset of R, and put
Z =x L
x
: x R.
Appendix 3
Borel measures on metric spaces
This Appendix is devoted to some topological aspects of the theory of -nite measures
which are given on the Borel -algebras of metric spaces. Namely, here we are going to
describe the topological structure of supports of such measures. Since every nonzero -
nite measure is equivalent to a probability measure (dened on the same -algebra), we
may restrict our further considerations to the case of probability measures.
Let E be an arbitrary topological space and let be a -nite Borel measure on E. We
recall that a Borel set X E is a support of (or, in other words, is concentrated on X)
if
(E X) = 0.
Obviously, for a Borel probability measure on E, the following two assertions are equiv-
alent:
1) X is a support of ;
2) (X) = (E) = 1.
For our purposes, we need one important result concerning supports of -nite Borel mea-
sures given on a metric space (E, d) whose topological weight is not measurable in the
Ulam sense, i.e., is not real-valued measurable.
We recall that the topological weight of E is the smallest cardinality among the cardinalities
of all open bases of E. The topological weight of E is usually denoted by the symbol w(E).
For example, taking the classical topological vector spaces
R
N
= R
=(x
n
)
n<
: (n <)(x
n
R),
c
0
=(x
n
)
n<
R
: lim
n+
x
n
= 0,
l
=(x
n
)
n<
R
: sup
n<
[x
n
[ < +,
383
we have respectively
w(R
N
) = w(c
0
) = card(N) =,
w(l
) = card(R) = c.
Recall that an innite cardinal number a is real-valued measurable or measurable in the
Ulamsense if there exists a nonzero -nite measure dened on the -algebra of all subsets
of a and vanishing at all one-element subsets of a.
Obviously, in this denition we may consider any set A with card(A) =a, instead of a. Also,
it is clear that a is real-valued measurable if and only if there exists a diffused probability
measure on a with dom() =P(a).
Extensive information about real-valued measurable cardinals and related topics of com-
binatorial set theory is presented in [69], [91], [144], [145], [150], and [231]. Here we
only wish to recall the important fact that the existence of real-valued measurable cardinals
cannot be established within ZFC theory. Consequently, the assumption that all cardinals
are not real-valued measurable is consistent with ZFC (cf. Exercise 18 for Appendix 1).
So, in order to manipulate with real-valued measurable cardinals, we are forced to appeal
to additional set-theoretical axioms.
Let us observe that if a cardinal a is measurable in the Ulam sense and b is another cardinal
greater than a, then b is also measurable in the Ulam sense. This fact follows directly from
the denition. Indeed, if A and B are any two sets such that
card(A) = a, card(B) = b, A B,
and is a diffused probability measure with dom() =P(A), then a functional
: P(B) [0, 1]
(X) = (X A) (X B)
is a diffused probability measure with dom() =P(B) and, in fact, this is concentrated
on the set A.
First, we would like to formulate and prove the following auxiliary statement from the
theory of metric spaces (see [148], [174], and [192]).
Lemma 1. Let (E, d) be a metric space, I be a set of indices linearly ordered by some
relation _, and let
A
i
: i I, B
i
: i I, C
i
: i I
Appendix 3. Borel measures on metric spaces 385
be three families of subsets of E such that:
1) all sets A
i
(i I) are open in E;
2) for each index i I, we have
B
i
= A
i
A
j
: j i;
3) for each index i I, the set C
i
is closed in E and C
i
B
i
.
Then the set
C =C
i
: i I
can be represented as the union of some countable family of closed subsets of E (in other
words, C is of type F
in E).
Proof. For any index i I and for any natural number n 1, let us put
C
n
i
=z : z C
i
, d(z, E A
i
) 1/n.
Evidently, we have the equality
C
n
i
: 1 n < =C
i
.
Furthermore, each set C
n
i
is closed in E. Now, if
j i, x C
n
i
, y C
n
j
,
then d(x, y) 1/n. Consequently, we get
(n)(1 n < d(C
n
i
,C
n
j
) 1/n)
for any two distinct indices i and j from I (here we use the assumption that the set I is
linearly ordered by _). Thus for each natural number n 1, the set
C
(n)
=C
n
i
: i I
turns out to be closed in E. It remains to observe that
C =C
i
: i I =C
(n)
: 1 n <
which shows that C is an F
-set in E. Lemma 1 has thus been proved.

Now, we are able to prove the following important statement (cf. [192]).
Theorem 1. Let (E, d) be a metric space whose topological weight is not real-valued
measurable, and let be an arbitrary -nite Borel measure on E. Then there exists a
separable support of , i.e., there exists a closed separable subset F = F
of E such that
(E F) = 0.
In particular, is a separable measure.
Proof. Since every nonzero -nite measure is equivalent to a probability measure, we
may assume, without loss of generality, that our is a Borel probability measure on E. Fix
a natural number k and consider a covering
U
k,i
: i I(k)
of E by open balls whose diameters are less than or equal to 1/(k +1). Obviously, we can
suppose that
card(I(k)) w(E).
Moreover, applying the Zermelo theorem, we can also suppose that the set I(k) is well-
ordered by some ordering relation _. Further, for each index i I(k), we denote
Z
k,i
=U
k,i
U
k, j
: j i.
Clearly, the family of sets
Z
k,i
: i I(k)
is disjoint and, simultaneously, is a covering of E. In addition to this circumstance, each
set Z
k,i
(as a difference of two open sets) admits a representation in the form
Z
k,i
=Z
k,i,m
: m <,
where all sets Z
k,i,m
are closed in E. Since is a probability measure, the set of all those
indices i I(k) for which (Z
k,i
) > 0, is at most countable. Let us denote this set by T(k)
and put
J(k) = I(k) T(k).
We are going to demonstrate that
(Z
k,i
: i J(k)) = 0.
Suppose otherwise, that is
(Z
k,i
: i J(k)) > 0.
Then for any subset J of J(k), we may write
Z
k,i
: i J =Z
k,i,m
: i J, m <.
According to Lemma 1, the set
Z
k,m
=Z
k,i,m
: i J
is Borel in E and hence the set
Z
k,i
: i J =Z
k,m
: m <
turns out to be Borel in E, too. Now, putting
(J) = (Z
k,i
: i J)
for every set J J(k), we come to the nonzero nite measure dened on the -algebra
of all subsets of J(k) and vanishing at all one-element subsets of J(k). In other words, we
derive that card(J(k)) is a real-valued measurable cardinal. But this is impossible because
of the relation
card(J(k)) card(I(k)) w(E)
and in view of our assumption that w(E) is not real-valued measurable. The contradiction
obtained shows that
(Z
k,i
: i J(k)) = 0
which implies, in particular, that our measure is concentrated on the set
U(k) =U
k,i
: i T(k)
for each natural number k. Now, it can easily be checked that the set
X =U(k) : k <
is a separable subspace of E and is concentrated on X as well. Thus, it sufces to put
F = cl(X) where cl(X) denotes the closure of X. The proof of Theorem 1 is complete.
Some applications of this theorem will be given later (see especially Exercises 2 and 3).
Remark 1. The assumption that a metric space E in Theorem 1 has non-real-valued mea-
surable topological weight is very essential for the validity of this theorem. Indeed, suppose
that a real-valued measurable cardinal exists and take such a cardinal a. Further, choose any
set E with card(E) = a and equip E with the discrete topology (or, equivalently, with the
discrete metric). Then E becomes a metrizable space whose topological weight is equal
to a. Let be a nonzero -nite measure dened on the family of all subsets of E and
vanishing at all one-element subsets of E. Then, obviously, is a Borel measure on E
which does not possess a separable support because any separable subset of E is at most
countable. In other words, the assertion of Theorem 1 fails to be true for such a .
In addition, let us notice that for the validity of Theorem 1, the assumption of metrizability
of E is essential, too (cf. Exercise 8 below).
We thus see that under some natural assumption which does not contradict the axioms of
set theory, any -nite Borel measure on a metric space E is concentrated on a separa-
ble subspace of E. Furthermore, if we deal with a Borel probability measure dened on
a separable metric space, then the question arises whether there exists a support of this
measure with more or less nice geometrical properties. Since every separable metric space
is isometrically contained in a separable Banach space and, more precisely, according to
the classical Banach-Mazur theorem, every separable metric space can be realized as an
isometric copy of some subset of the function space C[0, 1], it is reasonable to ask whether
each Borel probability measure on a separable Banach space admits a support with good
geometrical properties. This and similar questions are studied in many works but we do
not intend to discuss them here. We only touch upon the topological structure of supports
of Borel probability measures.
The next auxiliary proposition shows certain connections between -nite Borel measures
on separable metric spaces and sets of rst category in these spaces.
Lemma 2.. Let E be a topological space satisfying the following conditions:
1) there exists a countable set X E which is everywhere dense in E;
2) for any point x X, the one-element set x is a G
-subset of E.
Let be a -nite Borel measure on E vanishing on X (equivalently, for each x X, we
have (x) = 0). Then there exists a set P E such that:
(a) P is an F
-subset of E;
(b) P is of rst category in E;
(c) P is a support of .
In particular, if E is an arbitrary separable metric space, then any -nite diffused Borel
measure on E is concentrated on some rst category subset of E.
Proof. If is identically equal to zero, then there is nothing to prove. So we may assume
that ,= 0. Moreover, since every -nite measure is equivalent to a probability measure,
we may also suppose that (E) = 1. Actually, in our further consideration we only use the
assumption that is a nite measure.
Fix an enumeration x
n
: n < of all points of X. For each natural index n < , let
U
n,m
: m < be a countable family of open subsets of E such that
U
n,m
: m < =x
n
.
Taking into account the upper semicontinuity of the nite measure , we infer that for any
two natural numbers n and k, there exists a natural number m(n, k) satisfying the relation
(U
n,m(n,k)
) < 1/2
n+k+1
.
Let us dene
U(k) =U
n,m(n,k)
: n <.
Obviously, the set U(k) is open and everywhere dense in E. Also, the relation
(U(k))
1/2
n+k+1
: n < = 1/2
k
holds true. Consequently,
Q =U(k) : k <
is a G
-subset of E and the equality (Q) = 0 is valid. Denoting

P = E Q,
we see that the set P satises the relations (a), (b), and (c) which ends the proof of the
lemma.
From Theorem 1 and Lemma 2 we easily obtain the following important statement (see
[192]).
Theorem 2. Let (E, d) be a metric space whose topological weight is not real-valued
measurable and let be a -nite diffused Borel measure on E. Then there exists a disjoint
covering A, B of E such that (A) = 0 and B is a rst category set in E.
Proof. In view of Theorem 1, the measure admits a closed separable support C. By
virtue of Lemma 2, the set C can be represented in the form
C =C
/
C
//
(C
/
C
//
= / 0),
where C
/
is of -measure zero and C
//
is of rst category in C (hence, C
//
is of rst category
in E as well). Putting
A =C
/
(E C), B =C
//
,
we get the required result.
EXERCISES
1. Verify the validity of the relations
w(R
N
) = w(c
0
) =, w(l
) = c.
2. Let (E, d) be a complete metric space whose topological weight is not real-valued mea-
surable, and let be a Borel probability measure on E. Show that for each Borel set X E
and for any real number > 0, there exists a compact set K X satisfying the relation
(X K) <.
Deduce from this fact that if is an arbitrary -nite Borel measure on E and X is any
Borel set in E, then
(X) = sup(K) : K X, K is compact.
Thus, there exists a set X
/
X such that:
(a) (X X
/
) = 0;
(b) X
/
can be represented as the union of a countable family of compact subsets of E.
Infer from these results that every -nite Borel measure on E is Radon and hence E is a
Radon space.
Finally, taking into account the said above, conclude that the following statement holds
true:
under the assumption that there are no cardinals measurable in the Ulam sense, every com-
plete metric space is Radon.
3. Let (E, d) be a metric space whose topological weight is not real-valued measurable,
and let be a -nite diffused Borel measure on E. Demonstrate that there exists a set
X E satisfying the following relations:
(a) X is a separable F
-subset of E;
(b) X is of rst category in E;
(c) is concentrated on X, i.e., X turns out to be a support of .
4. Let E be a topological space, be an ordinal number and let U
: < be an
-sequence of open subsets of E such that
E =U
: <.
For each ordinal <, denote
M
=U
: <.
Let X
: < be an arbitrary -sequence of subsets of E. We put

M(X
: <) =M
: <
and we say that the set M(X
: < ) is obtained from X
: < by applying
the Montgomery operation (in short, M-operation) whose basis is the given -sequence
U
: <.
Verify that:
(a) if for any <, we have X
=X
,i
: i I, then
M(X
: <) =M(X
,i
: <) : i I;
(b) if for any <, we have X
=X
,i
: i I, then
M(X
: <) =M(X
,i
: <) : i I;
(c) if for any <, we have X
=Y
, then
M(X
: <) = M(Y
: <) M(Z
: <).
5. Let E be a topological space. For each ordinal <
1
, dene by transnite recursion
the family F
(E) of subsets of E. Namely, rst put

F
0
(E) = the family of all closed subsets of E.
If is odd, then let F
(E) coincide with the family of all countable unions of members

from F
(E) : <.
If > 0 is even, then let F
(E) coincide with the family of all countable intersections of

members from F
(E) : <.
Finally, put
F(E) =F
(E) : <
1
.
Verify the validity of the inclusion
F(E) B(E),
where B(E) as usual denotes the Borel -algebra of E.
Show that if a topological space E is such that every open set in E is of type F
or, equiva-
lently, every closed set in E is of type G
, then
F(E) =B(E).
In particular, this equality holds true for any metrizable topological space E.
For an ordinal <
1
, we say that a set X F(E) is of order if
X F
(E) F
(E) : <.
The classical theorem due to Lebesgue states that if E is an uncountable Polish space, then
for each <
1
, there are subsets of E of order (see, e.g., [99], [148], [160], [162], [183],
or Appendix 6).
6. Let E be a metric space, be a nonzero countable ordinal number and let X be a subset
of E. Suppose that X is locally Borel of order , i.e., for any point x E, there exists a
neighborhood U(x) such that the set U(x) X is Borel of order . Show that X is Borel
of order .
For this purpose, starting with the result of Exercise 4 and Lemma 1, use the method of
transnite induction on <
1
.
7. Give an example of a metric space E such that:
(a) the topological weight of E is equal to
1
;
(b) there exists a subset of E which is locally Borel but is not Borel in E.
8
. Equip the least uncountable ordinal

1
with its order topology and consider the Borel
-algebra B(
1
). Prove that there exists a probability measure on B(
1
) such that:
(a) is diffused and two-valued;
(b) (F) = 1 for any uncountable closed subset F of
1
.
Taking into account the circumstance that every separable subspace of
1
is at most count-
able, show that is not Radon and does not admit a separable support.
Conclude from this fact that the assumption of metrizability of a space E in the formulation
of Theorem 1 is essential.
Show also that the same assumption of metrizability is essential for the validity of Theorem
2.
The probability measure just described is usually called the Dieudonn e measure on
1
(see [53], [71], [80], and [101]).
Verify that is invariant under the semigroup of all those mappings
f :
1

1
which are representable in the form
f () = + ( <
1
),
where = ( f ) <
1
. In other words, f coincides with the left shift of
1
corresponding
to .
Notice that
1
is a locally compact topological space, every point of which has a countable
local base. Dene an analogue of for the compact space
1
+1 and check that this
analogue turns out to be a Borel probability measure which is not Radon and does not
admit a separable support.
9
. Let E be a separable Hilbert space (over R) equipped with the inner product <, >.
We recall that a set H E is a (closed) half-space in E if there exist a vector e E 0
and a real number t such that
H =x E : < e, x > t.
Let be a Borel probability measure on E. Prove that is completely determined by its
values on the half-spaces of E. This phrase means that if is any Borel probability measure
on E such that (H) =(H) for all half-spaces H E, then =.
Deduce from this fact that any Borel probability measure on E is completely determined
by its values on all balls of E.
In connection with Exercise 9, it should be mentioned that an analogous result does not
longer hold for more general separable metric spaces. The rst example of a separable
metric space E, in which some Borel probability measure is not determined by its values
on all balls of E, was given by Davies (see [50]).
10. Let (E, S, ) be a space equipped with a nonzero complete measure. We shall say that
a partition E
i
: i I of E is admissible for if the following two relations are satised:
(a) E
i
dom() for all i I;
(b) for every set X dom(), there exists a countable set J I such that
(X E
i
: i J) = 0.
Assume the Continuum Hypothesis. Let be a nonzero complete measure such that there
exists a family Z
i
: i I dom() satisfying the following conditions:
(c) card(I) c;
(d) for any Z dom() and for any real >0, there is an index i I such that (ZZ
i
) <.
Prove that there exists an admissible partition for .
11. Let (E, d) be a metric space whose topological weight is not measurable in the Ulam
sense, and let be a -nite Borel measure on E. Assuming Martins Axiom, show that
the completion
/
of is c-additive, i.e., for every disjoint family X
i
: i I dom(
/
)
with card(I) < c, the equality
/
(X
i
: i I) =
/
(X
i
) : i I
holds true.
12
. Let E be a topological space and let U

i
: i I be a family of open subsets of E such
that each U
i
(i I) is of rst category in E. Prove that the open set
U =U
i
: i I
is also of rst category in E. This important statement is due to Banach.
In order to show the validity of Banachs statement, consider a maximal (with respect to
the inclusion relation) disjoint family V
j
: j J of open sets in E such that for any index
j J, the set V
j
is contained in some set U
i
, where i = i( j). Verify that:
(a) the set V =V
j
: j J is of rst category in E;
(b) U cl(V) =V bd(V), where cl(V) denotes the closure of V and bd(V) stands for the
boundary of V.
Taking into account the fact that bd(V) is nowhere dense in E, conclude that U is of rst
category in E.
Infer from the said above that an arbitrary topological space E admits a representation in
the form
E = E
1
E
2
(E
1
E
2
= / 0),
where E
1
is a closed rst category subspace of E and E
2
is an open Baire subspace of E.
13. Let E be a topological space and let X be a subset of E which locally is of rst category
in E, i.e., for any point e E, there exists a neighborhood U(e) such that the set U(e) X
is of rst category in E.
Show that X is of rst category in E.
For this purpose, apply the result of Exercise 12.
14. Let (G, ) be an arbitrary topological group. Prove that the disjunction of the following
two assertions is valid:
(a) G is a rst category topological space;
(b) G is a Baire topological space.
For this purpose, apply again the result of Exercise 12.
15. Let (G, ) be a topological group and let A and B be two second category subsets of G
such that both of them have the Baire property. Show that the set
A B =a b : a A, b B
contains a nonempty open subset of G (the Banach-Kuratowski-Pettis theorem).
16
. Anonempty linearly ordered set (S, ) is called a Suslin line if it satises the following
conditions:
(a) S has neither least element, nor greatest element;
(b) S is dense in itself, i.e., for any two elements x S and y S such that x <y, there exists
z S such that x < z < y;
(c) S is complete in the Dedekind sense, i.e., for every nonempty bounded from above set
X S, there exists sup(X) in S;
(d) any disjoint family of nonempty open subintervals of S is at most countable.
(e) (S, ) is not isomorphic to the real line R equipped with its standard order.
The existence of a Suslin line does not contradict the axioms of ZFC theory. On the other
hand, it follows from Martins Axiom and the negation of the Continuum Hypothesis that
there is no Suslin line (for more details, see [40], [91], [144], and [145]).
Verify that the next two relations (e) and (e) are equivalent to condition (e):
(e) S does not contain a countable subset having common elements with every nonempty
open subinterval of S;
(e) S is not metrizable with respect to its order topology.
Let us endow(S, ) with its order topology and let be an arbitrary -nite Borel measure
on S. Prove that has a separable support in S, i.e., there exists a separable closed set F S
such that (S F) = 0.
It is useful to compare the last exercise with Exercise 8.
Appendix 4
Continuous nowhere approximately
differentiable functions
In Exercise 4 for Chapter 8, an application of Kuratowskis theorem on closed projection
to the question of the existence of nowhere differentiable functions in the Banach space
C[0, 1] was given. It was also underlined that such pathological functions are typical, i.e.,
they constitute a residual (co-meager) set in C[0, 1]. Recall that this classical result is due
to Banach [9] and Mazurkiewicz [171].
Analogous questions on the existence of continuous nowhere differentiable functions with
respect to other concepts of derivative are much deeper and more complicated. Among
those concepts the notion of an approximate derivative is of especial interest and impor-
tance (cf., for instance, [32], [162], [183], and [210]).
As mentioned at the end of Chapter 20, the rst example of a nowhere approximately
differentiable function from the Banach space C[0, 1] is due to Jarnik (see [88]). Moreover,
he showed that such functions also are typical in the space C[0, 1].
In this Appendix we present one precise construction of a continuous function acting from
R into R, which is nowhere approximately differentiable. This construction is due to Mal y
[167]. It is not difcult and, at the same time, is rather vivid from the geometrical point of
view.
We begin with some preliminary notions and facts.
Let denote the standard Lebesgue measure on R and let X be an arbitrary -measurable
subset of R. We recall that x R is a density point for (of) X if
lim
h0, h>0
(X [x h, x +h])/2h = 1.
According to the classical theorem of Lebesgue (see [32], [158], [161], [183], and [210]),
almost all points of X are its density points.
The notion of a density point turned out to be rather deep and fruitful not only for real
analysis but also for general topology, probability theory and some other domains of math-
397
ematics. For example, by using this notion the important concept of the density topology on
R was introduced and examined. Then various properties of the density topology were ex-
tensively investigated by many authors (Pauc, Goffman, Waterman, Nishiura, Neugebauer,
Tall, and others). This topology and its further generalizations were studied from different
points of view (see [75], [147], [166], [194], [205], and [237]). We shall touch upon the
density topology in several exercises of this Appendix.
Now, let f : R R be a function and let x R. We recall that f is approximately
continuous at x if there exists a -measurable set X such that:
(i) x is a density point of X;
(ii) the function f [(X x) is continuous at x.
Exercises 3 and 4 for this Appendix show that all Lebesgue measurable functions can be
described in terms of approximate continuity.
Let f : RR be a function and let x R. We recall that f is approximately differentiable
at x if there exists a Lebesgue measurable set Y R for which x is a density point and for
which there is a limit
lim
yx, yY, y,=x
f (y) f (x)
y x
.
This limit is denoted by f
/
ap
(x) and is usually called an approximate derivative of f at
the point x. It should be noticed that f
/
ap
(x) is uniquely determined by f and x (in this
connection, see Exercise 7).
For our purposes below, we need two simple auxiliary propositions.
Lemma 1. Let f : R R be a Lebesgue measurable function, let x be a point of R and
suppose that f is approximately differentiable at x. Then for any real number t
1
> f
/
ap
(x),
we have
lim
h0+
(y [x h, x +h] x : ( f (y) f (x))/(y x) t
1
)/2h = 0.
Similarly, for any real number t
2
< f
/
ap
(x), we have
lim
h0+
(y [x h, x +h] x : ( f (y) f (x))/(y x) t
2
)/2h = 0.
Proof. Since the argument in both cases is completely analogous, we shall consider only
the case of t
1
> f
/
ap
(x). There exists a -measurable set X such that x is a density point of
X and
lim
yx,y,=x,yX
( f (y) f (x))/(y x) = f
/
ap
(x).
Appendix 4. Continuous nowhere approximately differentiable functions 399
Fix > 0 for which
f
/
ap
(x) + <t
1
.
Then there exists a real number >0 such that for any strictly positive real h <, we have
(y X [x h, x +h] x)(( f (y) f (x))/(y x) f
/
ap
(x) +).
But if > 0 is sufciently small, then
(X [x h, x +h])/2h 1
for all strictly positive reals h <. So we obtain the relation
(y [x h, x +h] x : ( f (y) f (x))/(y x) t
1
)/2h ,
and the lemma is proved.
Actually, in our further considerations we need only the following auxiliary assertion which
is an immediate consequence of Lemma 1.
Lemma 2. Let f : R R be a Lebesgue measurable function, let x be a point of R, and
suppose that for every strictly positive real number t, the relation
liminf
h0+
(y [x h, x +h] x : [ f (y) f (x)[/[y x[ t)
2h
> 0
holds true. Then f is not approximately differentiable at x.
In particular, suppose that two sequences
h
k
: k N, t
k
: k N
of real numbers are given and satisfy the following conditions:
(1) h
k
> 0 and t
k
> 0 for all integers k 0;
(2) lim
k+
h
k
= 0 and lim
k+
t
k
= +;
(3) the lower limit
liminf
k+
(y [x h
k
, x +h
k
] x :
[ f (y) f (x)[/[y x[ t
k
)/2h
k
is strictly positive.
Then, in view of Lemma 2, we can assert that the function f is not approximately differen-
tiable at x.
After these simple preliminary remarks, we are able to begin the construction of a nowhere
approximately differentiable function.
First of all, let us put
f
1
(0/9) = 0, f
1
(1/9) = 1/3, f
1
(2/9) = 0, f
1
(3/9) = 1/3,
f
1
(4/9) = 2/3, f
1
(5/9) = 1/3, f
1
(6/9) = 2/3, f
1
(7/9) = 3/3,
f
1
(8/9) = 2/3, f
1
(9/9) = 3/3
and extend (uniquely) this partial function to a continuous function
f
1
: [0, 1] [0, 1]
in such a way that f
1
becomes afne on each segment [k/9, (k+1)/9], where k =0, 1, . . . , 8.
We shall start with this function f
1
. In our further construction, we also need an analogous
function g acting from the segment [0, 9] into the segment [0, 3]. Namely, we put
g(x) = 3 f
1
(x/9) (x [0, 9]).
Obviously, g is continuous and afne on each segment [k, k+1], where k =0, 1, . . . , 8. Also,
another function similar to g will be useful in our construction. Namely, we denote by g
the function acting from [0, 9] into [0, 3] whose graph is symmetric to the graph of g, with
respect to the straight line
(x, y) RR : y = 3/2.
In other words, we put
g
(x) = 3 g(x)
for all x [0, 9]. Suppose now that for a natural number n 1, the function
f
n
: [0, 1] [0, 1]
has already been dened, such that:
(a) f
n
is continuous;
(b) for each segment of the form [k/9
n
, (k +1)/9
n
], where
k 0, 1, . . . , 9
n
1,
the function f
n
is afne on [k/9
n
, (k +1)/9
n
] and the image of this segment with respect to
f
n
is some segment of the form [ j/3
n
, ( j +1)/3
n
], where
j 0, 1, . . . , 3
n
1.
Let us construct a function
f
n+1
: [0, 1] [0, 1].
For this purpose, it sufces to dene f
n+1
on any segment [k/9
n
, (k +1)/9
n
], where k
0, 1, . . . , 9
n
1. Here only two cases are possible.
1. f
n
is increasing on [k/9
n
, (k +1)/9
n
]. In this case, let us consider the following two sets
of points of the plane:
(0, 0), (0, 3), (9, 3), (9, 0),
(k/9
n
, f
n
(k/9
n
)), (k/9
n
, f
n
((k +1)/9
n
)),
((k +1)/9
n
, f
n
((k +1)/9
n
)), ((k +1)/9
n
, f
n
(k/9
n
)).
Because we have here the vertices of two rectangles, there exists a unique afne transfor-
mation
h : R
2
R
2
satisfying the conditions
h(0, 0) = (k/9
n
, f
n
(k/9
n
)), h(0, 3) = (k/9
n
, f
n
((k +1)/9
n
)),
h(9, 3) = ((k +1)/9
n
, f
n
((k +1)/9
n
)), h(9, 0) = ((k +1)/9
n
, f
n
(k/9
n
)).
Let the graph of the restriction of f
n+1
to the segment [k/9
n
, (k +1)/9
n
] coincide with the
image of Gr(g) under action of h.
2. f
n
is decreasing on [k/9
n
, (k +1)/9
n
]. In this case, let us consider the following two sets
of points of the plane:
(0, 0), (0, 3), (9, 3), (9, 0),
(k/9
n
, f
n
((k +1)/9
n
)), (k/9
n
, f
n
(k/9
n
)),
((k +1)/9
n
, f
n
(k/9
n
)), ((k +1)/9
n
, f
n
((k +1)/9
n
)).
Here we also have the vertices of two rectangles, so there exists a unique afne transfor-
mation
h
: R
2
R
2
h
(0, 0) = (k/9
n
, f
n
((k +1)/9
n
)), h
(0, 3) = (k/9
n
, f
n
(k/9
n
)),
h
(9, 3) = ((k +1)/9

n
, f
n
(k/9
n
)), h
(9, 0) = ((k +1)/9

n
, f
n
((k +1)/9
n
)).
Let the graph of the restriction of f
n+1
to the segment [k/9
n
, (k +1)/9
n
] coincide with the
image of Gr(g
) under action of h
.
The function f
n+1
has thus been determined. It immediately follows from the above
construction that the corresponding analogues of the conditions (a) and (b) hold true
for f
n+1
, too. In other words, f
n+1
is continuous and for each segment of the form
[k/9
n+1
, (k +1)/9
n+1
], where
k 0, 1, . . . , 9
n+1
1,
the function f
n+1
is afne on [k/9
n+1
, (k +1)/9
n+1
] and the image of this segment with
respect to f
n+1
is some segment of the form [ j/3
n+1
, ( j +1)/3
n+1
], where
j 0, 1, . . . , 3
n+1
1.
Moreover, our construction directly yields the relation
(x [0, 1])([ f
n+1
(x) f
n
(x)[ 1/3
n
).
In addition, let
[u, v] = [k/9
n
, (k +1)/9
n
]
be an arbitrary segment on which f
n
is afne. Then it is not hard to check that
f
n+1
([u, (2u +v)/3]) = f
n
([u, (2u +v)/3]),
f
n+1
([(2u +v)/3, (2v +u)/3]) = f
n
([(2u +v)/3, (2v +u)/3]),
f
n+1
([(u +2v)/3, v]) = f
n
([(u +2v)/3, v]).
Proceeding in this way, we come to the sequence of functions
( f
1
, f
2
, . . . , f
n
, . . . )
which are uniformly converging to some continuous function f . Obviously, f also acts
from [0, 1] into [0, 1].
We assert that f is nowhere approximately differentiable on the segment [0, 1]. In order to
demonstrate this fact, let us take an arbitrary point x [0, 1] and x a natural number n 1.
Clearly, there exists a number k 0, 1, . . . , 9
n
1 such that
x [k/9
n
, (k +1)/9
n
].
Therefore, we have
f
n
(x) [ j/3
n
, ( j +1)/3
n
]
for some number j 0, 1, . . . , 3
n
1. For the sake of simplicity, denote
[u, v] = [k/9
n
, (k +1)/9
n
], [p, q] = [ j/3
n
, ( j +1)/3
n
].
From the remarks made above it immediately follows that for all natural numbers m > n,
we have
f
m
(x) [p, q]
and, consequently, f (x) [p, q], too. Further, we may assume without loss of generality that
f
n
is increasing on [u, v] because the case when f
n
is decreasing on [u, v] can be considered
completely analogously.
Suppose rst that f (x) (p +q)/2 and put
D
1
= [(2v +u)/3, v].
Then for each point y D
1
, we may write
f (y) [(2q + p)/3, q].
Hence, we get
( f (y) f (x))/(y x) ((2q + p)/3 (p +q)/2)/(v u) = (1/6)(3
n
).
Suppose now that f (x) (p +q)/2 and denote
D
2
= [u, (2u +v)/3].
In this case, for any point y D
2
, we may write
f (y) [p, (2p +q)/3].
Hence, we get
( f (x) f (y))/(x y) ((p +q)/2 (2p +q)/3)/(v u) = (1/6)(3
n
).
Thus, in both of these cases we have the inequality
(y [x 1/9
n
, x +1/9
n
] x :
[ f (y) f (x)[/[y x[ (1/6)(3
n
)) (1/3)(1/9
n
)
or, equivalently,
(y [x 1/9
n
, x +1/9
n
] x :
[ f (y) f (x)[/[y x[ (1/6)(3
n
)) (1/6)([x 1/9
n
, x +1/9
n
]).
The latter relation immediately implies that our function f is not approximately differen-
tiable at x (see Lemma 2 and the comments after it).
Remark 1. The function f constructed above has a number of other interesting properties
(in this connection, see [167]).
Now, starting with an arbitrary continuous nowhere approximately differentiable function
acting from [0, 1] into [0, 1], we can easily get an analogous function for R. We thus come
to the following classical result (rst obtained by Jarnik in 1934).
Theorem 1. There exist continuous bounded functions acting from R into R, which are
nowhere approximately differentiable.
Remark 2. As already mentioned, Jarnik established a much stronger theorem. Let us
remember once again the precise formulation. Namely, Jarnik proved that almost all (in the
sense of the Baire category) functions from the Banach space C[0, 1] are nowhere approx-
imately differentiable. Clearly, this result essentially strengthens the corresponding result
of Banach and Mazurkiewicz obtained by them for the usual differentiability. Further in-
vestigations showed that analogous statements hold true for many kinds of generalized
derivatives. In this connection, we refer the reader to the fundamental paper [33] where
category analogues of Theorem 1 for generalized derivatives are discussed in detail (see
also [34]).
In Chapter 15 of [122], a nontrivial application of nowhere approximately differentiable
functions is demonstrated to the question concerning some relationships between the sup-
measurability and weak sup-measurability of functions acting from RR into R.
The concept of an approximate derivative relies essentially on the notion of a density point,
so it is reasonable to touch upon the density topology T
d
on R and to consider several
elementary properties of this topology. A precise denition of the density topology and
some of its interesting features are given in Exercises 9, 10, 11 for this Appendix.
One of the most important facts concerning the density topology states that (R, T
d
) is a
completely regular topological space (see, for instance, [192], [194], and [237]). This
property of T
d
implies highly nontrivial consequences in real analysis. To illustrate the
just said, we shall sketch a proof of the existence of an everywhere differentiable function
: R R
which is not monotone on any non-degenerate subinterval of R (notice that the construction
presented below is due to Goffman [74]).
Consider any two disjoint countable sets
A =a
n
: n N R,
B =b
n
: n N R
such that both of them are everywhere dense in R. Taking into account Exercise 10 and
the fact that (R, T
d
) is completely regular, we can nd for each natural number n, an
approximately continuous function
f
n
: R [0, 1]
0 f
n
(x) 1 (x R),
f
n
(a
n
) = 1, (x B)( f
n
(x) = 0).
Analogously, for each natural number n, there exists an approximately continuous function
g
n
: R [0, 1]
such that
0 g
n
(x) 1 (x R),
g
n
(b
n
) = 1, (x A)(g
n
(x) = 0).
Now, dene a function
h : R R
by the formula
h =

nN
(1/2
n
)( f
n
g
n
).
It is easy to check that:
a) h is bounded and approximately continuous;
b) h(a) > 0 for all a A;
c) h(b) < 0 for all b B.
Let denote an indenite integral of h. Keeping in mind the above relations a), b) and c),
it is not difcult to show that:
i) is everywhere differentiable on R and
/
(x) = h(x) for each x R;
ii) is not monotone on any nonempty open subinterval of R.
We see that with the aid of the density topology T
d
, it is possible to prove this very non-
trivial fact of mathematical analysis (another, more classical proof is presented in [98]).
Remark 3. The density topology on R can be regarded as a particular case of the so-called
von Neumann topology. Let (E, S, ) be a space with a complete probability measure
or, more generally, with a complete nonzero -nite measure. Then, in accordance with
a deep theorem of von Neumann and Maharam (see [166], [192], [194], and [205]), there
exists a topology T =T () on E such that:
(1) (E, T ) is a Baire space satisfying the Suslin condition;
(2) the family of all those subsets of (E, T ) which have the Baire property in (E, T )
coincides with the -algebra S;
(3) a set X E is of -measure zero if and only if X is of rst category in (E, T ).
We say that T = T () is a von Neumann topology associated with the given measure
space (E, S, ). Notice that T is not unique in general. This circumstance is not so
surprising because the proof of the existence of T , similarly to the proof of the well-known
Hahn-Banach theorem, is essentially based on uncountable forms of the Axiom of Choice
which are not needed in the case of T
d
. There are many nontrivial applications of a von
Neumann topology in various branches of contemporary mathematics. For instance, some
applications of this topology to the general theory of stochastic processes can be found in
[205].
Remark 4. There exist translation-invariant extensions of the Lebesgue measure for
which an analogue of the classical Lebesgue theorem on density points does not hold. For
example, there are a measure on R and a -measurable set X R, such that:
(1) is an extension of ;
(2) is invariant under the group of all isometric transformations of R;
(3) there is only one -density point for X, i.e., there exists a unique point x R for which
we have
lim
h0+
(X [x h, x +h])
2h
= 1.
A more detailed account of the measure and its other extraordinary properties can be
found in [115].
EXERCISES
1. Let (t
n
)
nN
be a sequence of strictly positive real numbers such that
lim
n+
t
n
= 0, lim
n+
t
n
/t
n+1
= 1.
Let X be a Lebesgue measurable subset of R and let x R. Prove that the following two
(a) x is a density point of X;
(b) lim
n+
(X [x t
n
, x +t
n
])/2t
n
= 1.
2. Let X be a Lebesgue measurable subset of R and let x R. Show that the following two
(a) x is a density point of X;
(b) lim
h0+, k0+
(X [x h, x +k])/(h +k) = 1.
3
. Let g : R R be a function, let be a xed strictly positive real number and suppose
that for any -measurable set X with (X) >0, there exists a -measurable set Y X with
(Y) > 0, such that
(x Y)(y Y)([g(x) g(y)[ <).
Demonstrate that there exists a -measurable function
h : R R
for which we have
(x R)([g(x) h(x)[ <).
Infer fromthis fact that if a function g : RR satises the above condition for every >0,
then g is measurable in the Lebesgue sense.
4. Let f : R R be a function. By applying the result of Exercise 3 and utilizing Luzins
theorem on the structure of Lebesgue measurable functions (i.e., the so-called C-property),
show that the following two assertions are equivalent:
(a) the function f is measurable in the Lebesgue sense;
(b) for almost all (with respect to ) points x R, the function f is approximately contin-
uous at x.
5. Let f : R R be a locally bounded Lebesgue measurable function and let
F(x) =
x
0
f (t)dt (x R).
Prove that for any point x R at which the function f is approximately continuous, we
have F
/
(x) = f (x).
Check also that the local boundedness of f is essential for the validity of this assertion.
6. Verify that if a function f : R R is approximately differentiable at x R, then f is
also approximately continuous at x.
7. Check that an approximate derivative of a function
f : R R
at a point x R is uniquely determined, i.e., it does not depend on the choice of a Lebesgue
measurable set Y for which x is a density point and for which the corresponding limit exists.
Check also that the family of all functions acting from R into R and approximately differ-
entiable at a point x R forms a vector space over R.
8. Demonstrate that if a function f : R R is differentiable in the usual sense at a point
x R, then f is approximately differentiable at x and f
/
ap
(x) = f
/
(x).
Give an example which shows that the converse assertion is not true.
9. For any Lebesgue measurable subset X of R, let us denote
d(X) =x R : x is a density point f or X.
Further, denote by T
d
the family of all those Lebesgue measurable sets Y R for which
Y d(Y). Verify that:
(a) T
d
is a topology on R strictly extending the standard Euclidean topology of R;
(b) (R, T
d
) is a Baire topological space and satises the Suslin condition, i.e., no nonempty
open set in (R, T
d
) is of rst category and any disjoint family of nonempty open sets in
(R, T
d
) is at most countable;
(c) every rst category set in (R, T
d
) is nowhere dense and closed; consequently, the family
of all those subsets of (R, T
d
) which have the Baire property in (R, T
d
), coincides with the
Borel -algebra of (R, T
d
);
(d) a set X R is Lebesgue measurable if and only if X has the Baire property in (R, T
d
);
(e) a set X R is of Lebesgue measure zero if and only if X is a rst category subset of
(R, T
d
);
(f) the space (R, T
d
) is not separable.
The above-mentioned topology T
d
is called the density topology on R. In a similar way,
the density topology can be introduced for the Euclidean space R
n
(n 2) equipped with
the n-dimensional Lebesgue measure
n
.
10. Let f : R R be a function and let x R. Prove that the following two assertions are
equivalent:
(a) f is approximately continuous at x;
(b) f regarded as a mapping acting from (R, T
d
) into R is continuous at x.
11
. By starting with the result of the previous exercise, show that the topological space
(R, T
d
) is connected.
For this purpose, suppose to the contrary that there exists a partition A, B of R into two
nonempty sets A T
d
and B T
d
. Then dene a function
f : R R
by putting f (x) = 1 for all x A, and f (x) = 1 for all x B. Obviously, f is a bounded
continuous mapping acting from (R, T
d
) into R and according to Exercise 10, f is approx-
imately continuous at each point of R. Further, dene
F(x) =
x
0
f (t)dt (x R).
By applying Exercise 5 of this Appendix, demonstrate that the function F is differentiable
everywhere on R and
F
/
(x) = 1 F
/
(x) =1
for each x R. This circumstance yields a contradiction with the Darboux property of any
derivative (cf. [32] and [183]).
12
. Supposing that the Continuum Hypothesis holds, prove that there exists a function
f : R R satisfying the following relations:
(a) for each uncountable set X R, the restriction f [X is not monotone;
(b) there are a set Y R with card(Y) = c and an everywhere differentiable function g :
R R such that f [Y = g[Y; consequently, f is not a Sierpi nski-Zygmund function.
Formulate and prove an analogous result assuming Martins Axiom instead of the Contin-
uum Hypothesis.
Appendix 5
Some facts from the theory of commutative
groups
The fundamental role of group theory in many elds of contemporary mathematics, such
as the theory of algebraic equations, geometry, topology, analysis, and the theory of differ-
ential equations, is widely known. In the class of all groups the subclass of all commutative
(i.e., Abelian) groups is of special interest. For instance, as demonstrated in Chapters 3, 10
and 17, certain properties of commutative and solvable groups are important in studies of
invariant and quasi-invariant measures.
In this Appendix, we would like to recall some material from the general theory of innite
commutative groups. This theory is thoroughly considered, for example, in remarkable
monographs [70], [83], and [152].
We shall present only those statements about commutative groups, which were essentially
utilized in the preceding chapters of this book. For instance, the reader could see in Chap-
ters 3, 10 and 17 how specic features of a given commutative transformation group G im-
ply the corresponding properties of those measures which are invariant (or quasi-invariant)
with respect to G.
Here we do not touch upon deep relationships between commutative groups and modules.
Obviously, any commutative group may be regarded as a module over the ring Z of all
integers. Numerous facts of the theory of commutative groups can be treated from a more
general view-point, as facts of the theory of modules. Such an approach is fruitful for both
of these theories and very often leads to important results.
Concerning nite commutative groups, we may say that there is no problemconnected with
their structural description. Indeed, the classical theorem says that any nitely generated
(in particular, nite) commutative group G is representable as a direct sum of cyclic groups.
The proof can be carried out by induction on card(X), where X is a nite set of generators
of G (in this connection, see [70] or [152]).
For innite commutative groups, we do not have such a nice result and the corresponding
411
theory becomes more complicated. Moreover, there are many innite commutative groups
which do not admit a nontrivial representation in the form of a direct sum (some simple
examples will be given below). It is also known that there exist commutative groups with
the above-mentioned property whose cardinalities are arbitrarily large.
Let (G, +) be a commutative group and let 0 denote the neutral element of G.
We recall that Gis free if it contains at least one subset X satisfying the following condition:
for any commutative group H and for any mapping
: X H,
there exists a unique homomorphism
/
: G H extending .
Actually, this notion is a particular case of the denition of a free object for a given abstract
algebraic structure. A set X in the denition above is usually called a basis of the group G.
A nite commutative group G is free if and only if G is trivial (in other words, G =0).
The group (Z, +) of all integers is a natural example of an innite free commutative group.
The one-element set 1 is a basis of this group (as well as the one-element set 1).
It can easily be veried that Z is not decomposable; this phrase means that Z does not admit
a representation in the form of a direct sum of its two proper subgroups.
Evidently, the direct sum of a family of free commutative groups is also a free commutative
group. More precisely, if G
i
: i I is a family of free commutative groups and X
i
is a
basis of G
i
for each i I, then the set
X =X
i
: i I
is a basis of the free commutative group
G =
iI
G
i
.
Of course, we identify here each set X
j
( j I) with its image under the canonical monomor-
phism from G
j
into
iI
G
i
.
It immediately follows from the said above that any commutative group is a homomor-
phic image of some free commutative group. This fact is a particular case of the general
statement from the theory of algebraic systems.
We say that a commutative group (G, +) is projective if for any two commutative groups
(H, +) and (H
/
, +) satisfying the relation (H, +) (H
/
, +) and for an arbitrary homomor-
phism
: G H
/
/H,
Appendix 5. Some facts from the theory of commutative groups 413
there exists a homomorphism
/
: G H
/
such that
=
/
,
where : H
/
H
/
/H is the canonical epimorphism.
The following statement yields a characterization of free commutative groups.
Theorem 1. Let (G, +) be a commutative group. Then these three assertions are equiva-
lent:
1) G is a free group;
2) G can be represented as a direct sum of a family of groups, all of which are isomorphic
to Z;
3) G is a projective group.
For the proof of this statement, see [70] or [152]. Thus, if (G, ) is a free commutative
group, then we may write
G =
iI
Z
i
,
where I is some set of indices and, for each i I, the group Z
i
is isomorphic to Z. In other
words, we have
Z
i
=ke
i
: k Z,
where e
i
denotes a certain element of Z
i
generating Z
i
and such that
ke
i
,= le
i
(k Z, l Z, k ,= l).
Clearly, the set e
i
: i I can be regarded as a basis of G =
iI
Z
i
.
Let us mention an important statement, according to which any subgroup of a free commu-
tative group is free, too. This statement remains true for non-commutative groups as well
(see, for instance, [152]).
We say that a commutative group (G, +) is divisible if for every natural number n > 0 and
for any element g G, the equation nx = g has at least one solution in G.
Evidently, the product of any family of divisible commutative groups is a divisible com-
mutative group; however, the standard proof of this simple fact is based on the Axiom of
Choice.
Example 1. The additive group (Q, +) of all rational numbers is divisible. All nonzero
elements in Q are of innite order. Similarly to Z, the group Q is not decomposable. The
commutative group (S
1
, ) is divisible, too. Recall that S
1
denotes the unit circumference
in R
2
equipped with the natural group operation which is the restriction to S
1
of the usual
multiplication operation of complex numbers (here R
2
is identied with the eld C of all
complex numbers). Moreover, for every nonzero n N, there are elements in S
1
of order
n. Clearly, there are also continuumly many elements in S
1
of innite order.
Example 2. Let p be an arbitrary prime natural number. Fix a natural number k and denote
by G
k
the subgroup of (S
1
, ) consisting of all those elements z S
1
for which the equality
z
p
k
= 1
is valid. In this way, we obtain a strictly increasing (by inclusion) sequence of nite com-
mutative groups
G
0
G
1
... G
k
... .
Let us dene
p
=G
k
: k <.
Then
p
is also a subgroup of S
1
and card(
p
) =.
Any group isomorphic to
p
is usually called a quasi-cyclic group of type p
.
Actually, the group
p
may be considered as the inductive limit of the family of groups
(G
k
)
k<
with respect to their canonical embeddings
k,k+1
: G
k
G
k+1
(k <).
Let us verify that
p
is divisible. First, let us observe that it sufces to establish the follow-
ing fact:
for any prime natural number q and for any element s
p
, there exists an element z
p
such that z
q
= s.
Indeed, assume that this fact is true. If n is an arbitrary nonzero natural number, then we
can represent n in the form
n = q
1
q
2
q
k
for some nite sequence (q
1
, q
2
, . . . , q
k
) of prime numbers. According to our assumption,
the equalities
s = z
q
1
1
, z
1
= z
q
2
2
, . . . , z
k2
= z
q
k1
k1
, z
k1
= z
q
k
k
are satised for certain elements z
1
, z
2
, . . . , z
k1
, z
k
of
p
, thus it immediately follows that
s = z
q
1
q
2
q
k
k
= z
n
k
.
Now, let q be a prime natural number and let s be an arbitrary element from
p
. Then there
exists a natural number k such that
s
p
k
= 1.
Only two cases are possible.
1. q ,= p. In this case, the numbers p
k
and q are co-prime. According to a well-known
theorem from elementary number theory, we may write
mp
k
+lq = 1
for some two integers m and l. Therefore, putting z = s
l
, we obtain
z
q
= s
lq
= s
1mp
k
= s.
Thus, z belongs to
p
and is a solution of the equation x
q
= s.
2. q = p. In this case, consider an element z S
1
such that z
p
= s. Since
z
p
k+1
= s
p
k
= 1,
we conclude that z belongs to
p
and is a solution of the equation x
q
= s.
Finally, let us point out the following interesting property of the group
p
:
every proper subgroup of
p
is necessarily nite.
We leave to the reader the checking this fact which also yields that
p
is not decomposable.
A commutative group (G, +) is called injective if for any two commutative groups (H, +)
and (H
/
, +) satisfying the relation (H, +) (H
/
, +) and for an arbitrary homomorphism
: H G,
there exists a homomorphism
/
: H
/
G extending .
The next statement yields a characterization of divisible commutative groups.
Theorem 2. Let (G, +) be a commutative group. Then the following three assertions are
equivalent:
1) G is a divisible group;
2) G can be represented as the direct sum of a family of groups, all of which are isomorphic
either to Q or to a group of type p
, where p is an arbitrary prime number;

3) G is an injective group.
The proof of this theorem may be found in [70] and [152].
From assertion 2) we infer at once that if (G, +) is a divisible commutative group whose
all nonzero elements are of innite order, then G is a vector space over Q.
Notice that assertion 3) easily implies the next useful fact.
If G is a divisible subgroup of a commutative group H, then G is a direct summand in H;
in other words, we have a representation
H = G+G
/
(GG
/
=0)
for some subgroup G
/
of H.
To see this circumstance, let us consider the identity mapping
0
: G G
which obviously is an isomorphism of G onto itself. Applying the injectivity of G, we
claim that there exists a homomorphism
: H G
extending
0
. Let us put
G
/
= ker() =
1
(0).
Then G
/
is a subgroup of H, and it can readily be veried that H is a direct sum of G and
G
/
.
This property characterizes all divisible commutative groups. In other words, if (G, +) is a
direct summand in every larger commutative group, then G is divisible.
Let us mention, in addition to the said above, that any direct summand in a divisible com-
mutative group is divisible itself. Indeed, suppose that (G, +) is a divisible commutative
group and let
G = H +H
/
(HH
/
=0)
be a representation of G in the form of the direct sum of its two subgroups H and H
/
.
Let us demonstrate that H is divisible, too. Take any natural number n > 0 and an arbitrary
element h H. Since G is divisible, there exists an element x G such that nx =h. Clearly,
we may write x = y +z, where y H and z H
/
. Further, we have
nx = n(y +z) = ny +nz = h, h ny = nz, h ny H, nz H
/
.
Consequently, h ny = nz = 0 and h = ny which shows that H is divisible.
Theorem 2 also implies that for any commutative group (G, +), there are sufciently many
homomorphisms acting from G into (S
1
, ). Indeed, take any nonzero element g G and
consider the group [g] generated by g. In view of Example 1, the group S
1
contains a
subgroup T isomorphic to [g]. Let
: [g] T
be an isomorphism between these two subgroups and let
/
: G S
1
be a homomorphism extending . Obviously, we have
/
(g) ,= e, where e stands for the
neutral element of S
1
.
It immediately follows fromthe said above that if x G, y G and x ,=y, then there always
exists a homomorphism
x,y
: G S
1
such that (x) ,= (y). In other words, the family of all homomorphisms acting from
G into S
1
separates the elements of G. Keeping in mind this circumstance and applying
the standard argument, we see that every commutative group G can be embedded in the
corresponding product group S
1
where is some cardinal number. More precisely, we
may take equal to card(GG).
From this fact we readily derive that for any commutative group (G, +), there exists a di-
visible commutative group (G
/
, +) containing G as a subgroup. Moreover, we may assume
without loss of generality that
card(G
/
) card(G) +.
Let us point out another important consequence of the above-mentioned fact. Namely,
every innite commutative group G can be endowed with a non-discrete Hausdorff topol-
ogy compatible with the group operation in G. This result does not hold for innite non-
commutative groups. For instance, assuming CH, Shelah constructed a group of cardinal-
ity continuum, which does not admit a non-discrete Hausdorff group topology (see [213]).
Later, Olshanskii gave an example (without using extra axioms) of an innite group which
also does not admit a non-discrete Hausdorff group topology.
We have the following important statement.
Theorem 3. Let (G, +) be a commutative group and let H be a subgroup of G. If G can be
represented in the form of a direct sum of cyclic groups, then H can be represented in the
same form.
A detailed proof of Theorem 3 is given in [152]. It is useful to compare this theorem with
the result mentioned earlier and stating that any subgroup of a free commutative group is
free, too.
Example 3. It can easily be veried that the additive group Q is not a direct sum of cyclic
groups.
However, there exists an increasing (by the inclusion relation) countable family H
n
: n <
of subgroups of Q, such that each H
n
is representable as a direct sum of cyclic groups
and
Q =H
n
: n <.
Indeed, for n <, it sufces to put
H
n
= the subgroup of Q generated by 1/n!.
An analogous representation is trivially valid for any group of type p
where p is an arbi-
trary prime natural number (see Example 2).
Finally, we are able to establish the following statement which is due to Kulikov and was
essentially exploited in this book.
Theorem4. Let (G, +) be an arbitrary commutative group. Then there exists an increasing
(by inclusion) countable family G
n
: n < of subgroups of G such that:
1) G
n
: n < = G;
2) each group G
n
(n <) is a direct sum of cyclic groups.
Proof. We know that G can be embedded in a divisible commutative group G
/
. In view of
Theorem 2, this G
/
is the direct sum of a family of groups which are isomorphic either to Q
or to a group of type p
, where p is a prime natural number. Taking into account Example

3, we see that G
/
G
/
=G
/
n
: n <,
where G
/
n
: n < is an increasing (by inclusion) countable family of subgroups of G
/
and each G
/
n
is a direct sum of cyclic groups. Now, let us put
G
n
= GG
/
n
(n <).
Clearly, G
n
is a subgroup of G for any n <. Moreover, we have
(n <)(G
n
G
n+1
), G
n
: n < = G.
Finally, since G
n
is a subgroup of G
/
n
, we may apply Theorem 3 and conclude that G
n
is a
direct sum of cyclic groups. Theorem 4 has thus been proved.
The last theorem played a key role in our considerations concerning the existence of non-
measurable sets and the measure extension problem (see, for instance, Chapter 10). This
theoremalso implies that the family of all subgroups of an uncountable commutative group
is rather large from the measure-theoretical view-point. For example, it is known that if G
is a commutative group of cardinality
1
, then there always exists a subgroup of G non-
measurable with respect to a given nonzero -nite diffused measure on G (see [119] or
Exercise 7 of this Appendix). For non-commutative groups of the same cardinality, an anal-
ogous fact does not longer hold because there are groups G with card(G) =
1
but without
proper uncountable subgroups (see [213]).
Let us underline once more that many important statements concerning the structure of
innite commutative groups are presented in fundamental monographs [70], [83], [152].
Extensive information about locally compact commutative groups can be found in [46],
[83], [177], and [202].
Notice that there are rather natural questions concerning innite commutative groups,
which are not decidable within ZFC theory. Of course, the famous Whitehead problem
is of special interest among questions of such a kind. Since the topological formulation of
the Whitehead problem looks much more attractive than its purely algebraic formulation,
we will give below the topological version of this problem.
Let be any cardinal number. Consider the topological product G = S
1
. This G may
be regarded as a compact, commutative, and arcwise-connected group (the last property is
easily veried).
Whiteheads question can be formulated as follows.
Is any compact commutative arcwise-connected group isomorphic to the product of some
copies of S
1
?
As demonstrated by Shelah, this question is undecidable within ZFC theory (see his re-
markable works [211] and [212]).
EXERCISES
1
. Let p be a prime natural number. Show that the group

p
has the property that any its
proper subgroup is a cyclic group of order p
k
for some natural number k. In particular,
p
does not contain proper innite subgroups.
For this purpose, consider a representation of
p
in the form
p
= G
0
G
1
G
2
G
k
,
where each subgroup G
k
is cyclic of order p
k
, and verify that G
k
is a unique cyclic subgroup
of
p
whose order is p
k
.
Let a
0
= 1 and, consequently, G
0
= [a
0
], and let G
k
= [a
k
] for k > 0. Suppose that G is a
proper subgroup of
p
. Then there exists a smallest natural number n such that a
n+1
, G.
Check that G = G
n
. The inclusion G
n
G is trivial because of the relation a
n
G. To
show the reverse inclusion G G
n
, suppose otherwise, i.e., suppose that g G G
n
for
some element g
p
. There exists a smallest k > n such that g G
k
= [a
k
] and, therefore,
g , G
k1
= [a
k1
]. The group [g] is nite cyclic and its order cannot be less than p
k1
because the only cyclic subgroup of
p
with order p
k1
is G
k1
. Consequently, the order
of [g] is equal to p
k
and [g] must coincide with G
k
. This fact immediately implies that
a
k
G and hence a
n+1
G contradicting the denition of n.
In connection with Exercise 1, let us also remark that the groups
p
are unique (with ex-
actness to an isomorphism) innite commutative groups having the property that all their
proper subgroups are nite (see [152]).
2
. Give a direct proof of the fact that if (G, +) is an arbitrary commutative group, (H, +)
is a divisible commutative group and : G H is a partial homomorphism, then admits
an extension
: G H to a group homomorphism.
For this purpose, consider any partial homomorphism
/
: GH extending . Let g be an
arbitrary element of Gdom(
/
). Show that
/
can be extended to a partial homomorphism
//
: G H such that g dom(
//
). Indeed, only two cases are possible.
1. kg dom(
/
) for some natural number k >0. We may assume, without loss of generality,
that k is the least natural number with this property. Since H is divisible, there exists
h H satisfying the equality kh =
/
(kg). Put
//
(g) = h and prove that the latter equality
correctly determines the partial homomorphism
//
: G H whose domain coincides with
the group generated by dom(
/
) g.
2. kg , dom(
/
) for all natural numbers k >0. In this case, pick an arbitrary element h H
and put
//
(g) = h. Verify that this relation also correctly determines a partial homomor-
phism
//
: G H whose domain coincides with the group generated by dom(
/
) g.
Finally, apply the Zorn lemma for obtaining the required homomorphism
: G H ex-
tending the initial partial homomorphism.
3. Show that if a commutative group (G, +) has the property that it is a direct summand in
every commutative group containing G, then G is divisible.
4
. Check that any uncountable commutative group (G, +) contains a proper subgroup H
with card(H) = card(G).
In connection with this exercise, a deep result of Shelah [213] should be mentioned stating
the existence of a group (, ) of cardinality
1
, which does not contain proper uncountable
subgroups. The existence of such a (, ) solves a famous problem posed by Kurosh [152]
many years ago.
5
. Let E be a compact topological group and let X be an innite subgroup of E. Verify

that the topology on X induced by the topology of E is not discrete.
Deduce fromthis fact that any innite commutative group (G, +) is isomorphic to an every-
where dense subgroup of some closed subgroup of S
1
, where is an appropriate cardinal
number. Taking this circumstance into account, conclude that G admits a non-discrete
Hausdorff topology which is totally bounded and compatible with the algebraic structure
of G.
The above result was rst obtained in [102]. In connection with Exercise 5, it should also
be noticed that any innite commutative group admits a non-discrete metrization whose
metric is invariant under all translations of the group (see [70]).
6
. Consider the two-element multiplicative discrete group 1, 1 as a subgroup of the

countable multiplicative group
D =cos(2n/2
m
) +isin(2n/2
m
) : n Z, m Z.
The commutative group D is 2-divisible, i.e., the equation x
2
= d is solvable in D for
any d D. Let us introduce the product group G = D
which also is 2-divisible. Then

H = 1, 1
is a compact subgroup of G. By using H, we dene a topology T on

G. First of all, the set H remains in its product topology T
1
and we equip any translate
gH (g G) with the topology which is obtained fromT
1
via the left translation g. Further,
a set U G is open in the new topology T if and only if for every g G, the intersection
U gH is open in gH.
Verify that the pair (G, T ) is a commutative non-discrete locally compact topological
group.
Show that G does not contain proper everywhere dense subgroups.
For this purpose, consider any everywhere dense subgroup G
0
of G. Let g be an arbitrary
element of G. Since the set gH is open, we must have
gHG
0
,= / 0.
So we may take x H such that gx G
0
. Since G is 2-divisible, there exists y G such
that y
2
= x. Again, the set yH is open, so
yHG
0
,= / 0.
We may choose z H such that yz G
0
. Observe now that the order of any element from
H does not exceed 2. Therefore, denoting by e the neutral element of G, we can write
x = y
2
= y
2
e = y
2
z
2
= (yz)
2
(G
0
)
2
= G
0
,
gx G
0
, g G
0
x
1
= G
0
,
which yields the required equality G
0
= G.
In connection with the previous exercise, see also [46], [92], [103], and [203].
7
. Let (G, ) be a commutative group of cardinality

1
and let be a nonzero -nite
diffused measure on G.
Prove that there exists a subgroup of G nonmeasurable with respect to .
For this purpose, use an appropriate modication of Ulams (
1
)-matrix and Kulikovs
theorem on the structure of commutative groups.
Appendix 6
Elements of descriptive set theory
As is widely known, general set theory is concerned with abstract sets and various relations
between them. Descriptive set theory primarily deals with those subsets of the real line
R or, more generally, of a Polish topological space E, which have rather good structure
and can be described more or less effectively, e.g., without the aid of uncountable forms
of the Axiom of Choice. Here we would like to recall the beginnings of this beautiful and
important branch of mathematics.
From the view-point of real analysis and general topology, the most interesting properties
of a subset of R are its measurability in the Lebesgue sense and the so-called Baire property
which may be regarded as a certain topological analogue of measurability (see [40], [148],
[176], and [192]).
Recall that a set X in a topological space E has the Baire property if it admits a representa-
tion in the form
X = (U Y) Z
where U is an open subset of E and Y and Z are some rst category subsets of E.
The family of all sets in E having the Baire property is denoted by Ba(E). In fact, Ba(E)
coincides with the -algebra generated by the Borel -algebra B(E) and the family of all
rst category subsets of E. If E is of second category, then this family forms a -ideal of
subsets of E.
Let E
/
be a topological space and let g : E E
/
be a mapping. We say that g has the Baire
property if for each open set V E
/
, the pre-image g
1
(V) has the Baire property in E.
Observe that if there exists a rst category set X E such that g[(E X) is continuous,
then g necessarily has the Baire property. The converse assertion is also true under some
assumption on E
/
. Namely, suppose that the topology of E
/
is countably generated and a
mapping g : E E
/
has the Baire property. Let V
n
: n < denote a countable base of
423
E
/
. For each n <, we may write
g
1
(V
n
) = (U
n
Y
n
) Z
n
where U
n
is an open subset of E and Y
n
and Z
n
are some rst category sets in E. Let us put
X =Y
n
Z
n
: n <.
Obviously, X is a rst category subset of E and it can readily be veried that the restriction
g[(E X) is continuous.
For extensive information about the Baire property and the Baire property in the restricted
sense, see [40], [148], [176], and [192].
Recall that the Borel -algebra B(E) is generated by the family of all open (equivalently,
by the family of all closed) subsets of E. The following simple auxiliary proposition turns
out to be useful in many questions of descriptive set theory.
Lemma 1. Let E be a topological space such that any open set in E is of type F
(or,
equivalently, any closed set in E is of type G
). Then B(E) coincides with the monotone

class generated by the family of all open sets in E (equivalently, by the family of all closed
sets in E).
An easy proof is left to the reader (cf. Exercise 5 for Appendix 3).
Let E be a topological space and let f : E R be a function.
We say (see Chapter 20) that f is of Baire zero class if f is continuous at all points of E,
i.e., f is continuous on E.
The family of all continuous functions acting from E into R is usually denoted by the
symbol C(E, R) (or by C(E)). In accordance with the denition above, we use the notation
Ba
0
(E, R) for the same family of functions. Thus, we have
Ba
0
(E, R) =C(E, R).
Suppose now that for an ordinal <
1
, the Baire classes Ba
(E, R) ( <) have already

been determined. We say that a function f : E R belongs to the class Ba
(E, R) if there
exists a sequence of functions
f
n
: n < Ba
(E, R) : <
which pointwise converges to f , i.e.,
lim
n+
f
n
(x) = f (x) (x E).
Appendix 6. Elements of descriptive set theory 425
By proceeding in this way, it becomes possible to dene the classes Ba
(E, R) for all

ordinals <
1
. Clearly, these classes increase by the inclusion relation. Further, putting
Ba(E, R) =Ba
(E, R) : <
1
,
we obtain the class of all Baire functions acting from E into R (see [5], [99], [148], and
[162]).
In viewof the regularity of
1
, this class is closed with respect to the pointwise convergence
of sequences of functions.
We say that a function f Ba(E, R) is of Baire order <
1
if
f Ba
(E, R) Ba
(E, R) : <.
In Chapter 20 we were concerned with some important properties of the rst Baire class
Ba
1
(E, R). Recall that if E = R, then all monotone functions and all derivatives belong to
Ba
1
(R, R).
The following three simple assertions are valid for Ba
(E, R):
(1) Ba
(E, R) is a linear algebra over the eld R; in other words, if we have f Ba
(E, R),
g Ba
(E, R), a R and b R, then

a f +bg Ba
(E, R), f g Ba
(E, R);
(2) if f Ba
(E, R), g Ba
(E, R) and g(x) ,= 0 for all x E, then

f /g Ba
(E, R);
(3) if f Ba
(E, R) and Ba
(R, R), then f Ba

+
(E, R).
In fact, the above-mentionedrelations (1), (2), and (3) can readily be checked by the method
of transnite induction.
Let us also formulate a less trivial property of the class Ba
(E, R). Namely, if f

n
: n <
is a sequence of functions from this class, which uniformly converges to a function f : E
R, then f also belongs to Ba
(E, R) (cf. Theorem 2 from Chapter 20).

Recall that the symbol B(E, R) denotes the family of all Borel functions acting from E into
R. The following statement shows a close connection between Baire and Borel functions.
Theorem 1. If E is a perfectly normal topological space, then the equality
Ba(E, R) = B(E, R)
holds true. In particular, this equality is satised for any metric space E.
Keeping in mind Lemma 1, the Tietze-Urysohn theorem on extensions of real-valued con-
tinuous functions and the relation
Ba(E, R) =Ba
(E, R) : <
1
,
we can prove Theorem 1 by using the method of transnite induction on <
1
. We leave
the corresponding technical details to the reader (see Exercise 3).
Recall that
as usual denotes the canonical Baire space of countable topological weight.

We also recall that the symbol
<
denotes the family of all nite sequences of natural
numbers. In the sequel, we do not assume that this family is equipped with a topology. For
any s
<
, we put
lh(s) = card(dom(s)).
Actually, lh(s) coincides with the length of s. Further, if
s =s
0
, s
1
, ..., s
k

<
and n <, then the symbol s n denotes the extended sequence
s
0
, s
1
, ..., s
k
, s
k+1

<
where s
k+1
= n. In our further considerations, it will be convenient to identify each natural
number k with the set 0, 1, ..., k 1. In fact, this assumption is redundant since in mod-
ern set theory the natural numbers and, more generally, the ordinal numbers are usually
introduced in the von Neumann sense (see [10], [91], [145], and [150]).
In Chapter 8, the notion of A-operation was briey considered over appropriately indexed
countable families of subsets of a base set E. We would like to recall that if a countable
family of sets
X
s
: s
<
P(E)
is given, then the set
A(X
s
: s
<
) =
(X
[k
: k <)
is the result of A-operation applied to this family (see [99], [148], [150], [160], and [162]).
In many cases we may assume, without loss of generality, that an initial family X
s
: s
<
is regular, i.e.,
(n <)(s
<
)(X
sn
X
s
).
Indeed, if necessary, we may replace this family by Y
s
: s
<
where
Y
n
0
n
1
...n
k
= X
n
0
X
n
0
n
1
... X
n
0
n
1
...n
k
for any nite sequence s = (n
0
, n
1
, ..., n
k
) of natural numbers. Clearly, we have
A(X
s
: s
<
) = A(Y
s
: s
<
).
Also, suppose for a while that an initial regular family X
s
: s
<
satises the following
condition:
(s
<
)(t
<
)((s ,=t & lh(s) = lh(t)) X
s
X
t
= / 0).
Then we come to the equality
A(X
s
: s
<
) =
k<
(
s
<
, lh(s)=k
X
s
),
which is easily veried. Consequently, in this case, the A-operation is reduced to the oper-
ations of countable unions and countable intersections.
If a class L of subsets of E is given, then the symbol A(L) stands for the class of all those
sets which can be represented in the form A(X
s
: s
<
), where X
s
: s
<
L.
The members from A(L) are usually called analytic sets over the original class L.
It readily follows from the denition of the class A(L) that if L is closed under nite
unions and nite intersections, then the same property holds true for A(L).
Moreover, under the assumption of the closedness of L with respect to nite unions and -
nite intersections, we can assert that A(L) is closed under countable unions and countable
intersections. Indeed, consider an arbitrary countable family of sets
W
i
: i < A(L).
Then for each i <, we have
W
i
= A(X
i
s
: s
<
)
where X
i
s
: s
<
L is a regular family of sets. Let us dene
Y
/ 0
= E, Y
in
0
n
1
...n
k
= X
i
n
0
n
1
...n
k
for any i < and for any (n
0
, n
1
, ..., n
k
)
k+1
. Applying the A-operation to the obtained
family Y
s
: s
<
L, we easily infer that
W
i
: i < A(L).
Further, let us dene
Z
n
0
n
1
...n
k
=
ik
(X
i
m
0
m
1
...m
k
: m
0
+m
1
+... +m
k
n
0
+n
1
+... +n
k
)
for any (n
0
, n
1
, ..., n
k
)
k+1
. Applying the A-operation to the obtained family Z
s
: s
<
L and taking into account the K onig lemma on -trees with nite levels (see
Appendix 1), we derive that
W
i
: i < A(L).
From the point of view of numerous applications, the most important case is when the role
of E is played by an uncountable Polish topological space and the A-operation is applied
to appropriately indexed countable families of closed subsets of E, i.e., the role of L is
played by the class of all closed subsets of E. In this way we come to the Suslin subsets of
E. The next statement shows that they admit a nice topological description (see [99], [148],
[160], and [162]).
Theorem 2. Let E be a Polish space and let X be a nonempty Suslin subset of E. Then
there exists a continuous surjection
f :
X.
Conversely, any continuous image (in E) of the space
is a nonempty Suslin subset of E.

Proof. For every s
<
, denote
U(s) =z
: (i)(i < lh(s) z(i) = s(i)).

It is obvious that the family of sets U(s) : s
<
forms a base of
consisting of
closed-open sets. Since the given Polish space E is nonempty, it is not difcult to construct
a regular system F
s
: s
<
of nonempty closed subsets of E such that:
(1) lim
lh(s)+
diam(F
s
) = 0;
(2) E = A(F
s
: s
<
).
Actually, relation (2) shows that E is a continuous image of the Baire space
. Indeed,
for any z
, we may put
(z) = F
z
0
F
z
0
z
1
... F
z
0
z
1
...z
k
... .
It is easy to verify that this formula determines the continuous surjection
:
E.
Since the given nonempty set X E is Suslin, we can write
X = A(X
s
: s
<
)
where X
s
: s
<
is also a regular system of closed subsets of E. Taking into account
the above relations (1) and (2), we may assume without loss of generality that
lim
lh(s)+
diam(X
s
) = 0.
Now, let us put
Z =z
: X
z[k
: k < ,= / 0.
It can readily be checked that Z is a nonempty closed subset of
. Consequently, Z is a
nonempty Polish space as well. Let us dene a mapping
g : Z E
in the following way:
for each z Z, the value g(z) is equal to the unique point fromX
z[k
: k <.
An easy verication shows that g is continuous and g(Z) = X. Keeping in mind the same
relations (1) and (2) and applying their analogues to Z, we infer that there exists a continu-
ous surjection
:
Z.
In this manner we come to the required continuous surjection f = g of
onto X.
Conversely, let Y E be such that there exists a continuous surjection
h :
Y.
Then for each z
, we may write
h(z) h(U(z[k)) : k < cl(h(U(z[k))) : k <.
By virtue of the continuity of h, we get
lim
k+
diam(cl(h(U(z[k)))) = 0
from which it immediately follows that
h(z) =cl(h(U(z[k))) : k <.
Finally, denoting
P
s
= cl(h(U(s))) (s
<
),
we conclude that for the given set Y, the equality
Y = A(P
s
: s
<
)
holds true, which completes the proof of Theorem 2.
By using Theorem2, it can readily be shown that a metrizable continuous image of a Suslin
space is also Suslin and the topological product of a countable family of Suslin spaces is
Suslin, too.
If E is an arbitrary Polish or, more generally, perfectly normal topological space, then we
obviously have the inclusion B(E) A(E). Therefore, Theorem 2 directly yields the fact
that any nonempty Borel subset of a Polish space E can be considered as a continuous
image of the Baire space
. In this context, it should be noticed that for Borel subsets of

Polish spaces, a more precise result can be established. Namely, we have the next important
statement essentially due to Luzin (see [99], [148], [160], and [162]).
Theorem 3. Let E be a Polish space. Every Borel subset of E may be regarded as a
bijective continuous image of some Polish space.
Proof. Denote by L the class of all those subsets of E which are bijective continuous
images of Polish spaces.
Since any open subset U of E is a Polish space (see Exercise 24 for Chapter 8), we obtain
at once that U belongs to L.
Let X
i
: i I be an arbitrary disjoint countable family of subsets of E belonging to the
class L. For each index i I, denote by P
i
a Polish space such that there exists a continuous
bijection f
i
: P
i
X
i
. Let P be the topological sum of the family of spaces P
i
: i I.
Clearly, P is a Polish space, too. Without loss of generality, we may treat each P
i
(i I) as
a subset of P. Let
f : P X
i
: i I
denote the common extension of all mappings f
i
(i I). It can easily be seen that f is a
continuous bijection, thus it immediately follows that
X
i
: i I L.
Let now Y
i
: i I be an arbitrary countable family of subsets of E belonging to the
class L. Again, for each index i I, denote by Q
i
a Polish space such that there exists a
continuous bijection h
i
: Q
i
Y
i
. Both topological products Q
i
: i I and E
I
are Polish
spaces. Dene a continuous mapping
h :
Q
i
: i I E
I
by putting
h(q) = (h
i
(q
i
))
iI
(q = (q
i
)
iI

Q
i
: i I).
Obviously, h is injective and continuous. Further, denote
=z E
I
: (i I)(j I)(pr
i
(z) = pr
j
(z)),
Q = h
1
().
Since Q is a closed subset of Q
i
: i I, we see that Q is a Polish space as well. Addi-
tionally, the restricted mapping
h[Q : Q (
Y
i
: i I)
is a continuous bijection. Taking into account the simple fact that the set Y
i
: i I
is homeomorphic to the set Y
i
: i I, we infer that Y
i
: i I belongs to the class L.
All the said above enables us to conclude that L contains the Borel -algebra B(E) (cf.
Exercise 1) which completes the proof.
In connection with Theorem 3, see also Exercise 9 of this Appendix.
The following statement is a direct consequence of Theorem 3:
any uncountable Borel subset of a Polish space has cardinality continuum.
It is remarkable that the same statement is valid for all uncountable Suslin sets. In order to
show this fact, we need one auxiliary proposition.
Lemma 2. Let E be an arbitrary separable metric space and let g be a continuous mapping
acting from E onto some uncountable metric space. Then there exist two open balls U E
and V E such that:
(1) both sets g(cl(U)) and g(cl(V)) are uncountable;
(2) g(cl(U)) g(cl(V)) = / 0.
The proof of this lemma is quite simple and is left to the reader.
Now, we are able to formulate and prove the famous Alexandrov-Hausdorff theorem con-
cerning the perfect subset property of Suslin sets in a Polish space (see [99], [148], [160],
and [162]).
Theorem 4. Let E be a Polish topological space, Y be an uncountable Suslin subset of E
and let g : E Y be a continuous surjection. Then there exists a set X E homeomorphic
to the Cantor space 0, 1
such that the restriction

g[X : X g(X)
is a bijection (hence, in view of the compactness of X, this restriction is also a homeomor-
phism between X and g(X)).
Proof. By using Lemma 2 and ordinary recursion, it is not difcult to construct a dyadic
system
F
s
: s 0, 1
<
of closed balls in E such that for any s 0, 1

<
, the following relations are satised:
(a) F
s0
F
s1
F
s
;
(b) g(F
s0
) g(F
s1
) = / 0;
(c) g(F
s
) is an uncountable set;
(d) diam(F
s
) < 1/(lh(s) +1).
Having this system of balls, we may put
X(k) =F
s
: s 0, 1
<
, lh(s) = k (k <),
X =X(k) : k <.
A direct verication shows that the set X is homeomorphic to the Cantor space by virtue
of the relations (a), (b), and (d). In addition to this circumstance, relation (b) implies that
g[X is an injection. Keeping in mind the compactness of X, we conclude that g[X is a
homeomorphismbetween the sets X and g(X).
Remark 1. As an immediate consequence of Theorem 4, we obtain that the cardinality of
any uncountable Suslin (hence Borel) subset of a Polish space is equal to c. We thus see
that Suslin sets in some sense realize the Continuum Hypothesis: they are either countable
or they are of cardinality continuum. Unfortunately, the same statement fails to be true for
co-Suslin sets (in this connection, see [10], [91], [99], [150], [162], and [188]; cf. also
Exercise 14).
Let E be an arbitrary topological set and let X and Y be two subsets of E. According to
Luzins classical denition (see [160], [162]), we say that X and Y can be separated by
Borel sets if there exist two Borel sets X
/
E and Y
/
E such that
X X
/
, Y Y
/
, X
/
Y
/
= / 0.
The following auxiliary statement is valid.
Lemma 3. Let E be a topological space, let X and Y be two subsets of E and suppose that
X =X
n
: n <, Y =Y
n
: n <.
If the given sets X and Y cannot be separated by Borel sets, then there exists a pair (m, n)
of natural numbers such that the sets X
m
and Y
n
also cannot be separated by Borel sets.
Proof. The argument is quite easy. Indeed, suppose to the contrary that for any pair (m, n)
of natural numbers, the sets X
m
and Y
n
can be separated by some Borel sets. This assump-
tion means that there exists a Borel set P
mn
E satisfying the relation
X
m
P
mn
E Y
n
.
Now, putting
B
m
=P
mn
: n < (m <),
we see that
X B
m
: m <, Y (B
m
: m <) = / 0,
which contradicts the condition that X and Y cannot be separated by Borel sets. The ob-
tained contradiction completes the proof.
The next classical result is due to Luzin and is usually called the separation principle for
analytic (Suslin) sets (see [160] and [162]).
Theorem 5. Let E be a Polish topological space and let X and Y be two disjoint Suslin
subsets of E. Then X and Y can be separated by Borel sets.
Proof. Suppose to the contrary that X and Y cannot be separated by Borel sets. Let
f :
X, g :
Y
be two continuous surjections (see Theorem 2). Using the notation from the proof of The-
orem 2, we may write
X =f (U(n)) : n <, Y =g(U(m)) : m <.
By virtue of Lemma 3, there exist natural numbers n
0
and m
0
such that the sets f (U(n
0
))
and g(U(m
0
)) cannot be separated by Borel sets. Proceeding by ordinary recursion, we
will be able to construct two innite sequences
a = (n
0
, n
1
, ..., n
k
, ...)
, b = (m
0
, m
1
, ..., m
k
, ...)
such that for every k < , the corresponding sets f (U(a[k)) and g(U(b[k)) cannot be sep-
arated by Borel sets. On the other hand, keeping in mind the continuity of both mappings
f and g, we readily infer that for a sufciently large natural index k, the sets f (U(a[k)) and
g(U(b[k)) are contained in some disjoint open neighborhoods of the two distinct points
f (a) X and g(b) Y. We thus come to a contradiction, which ends the proof of Theorem
5.
Remark 2. Let E be an arbitrary uncountable Polish space. As shown by Luzin and
Novikov (see, e.g., [160], [162]), there are two disjoint co-Suslin sets X and Y in E, which
cannot be separated by Borel sets. This important fact immediately yields that there exists
a Suslin subset of E, which is not Borel (see also Exercise 13).
The separation principle for analytic sets implies many useful consequences.
For instance, let E be a Polish space and let X E be an analytic set such that its comple-
ment E X is also analytic. Then we can assert that X is a Borel subset of E. This classical
result is due to Suslin and follows directly from Luzins separation principle for analytic
sets.
Let us give another example of an application of this principle. Suppose that E and F are
any two Polish spaces and h is a mapping acting from E into F such that the graph Gr(h)
of h is a Suslin subset of the product space E F. Then we can assert that h is a Borel
mapping and, therefore, Gr(h) is a Borel subset of E F. To see this circumstance, take
any Borel set B F. Clearly, we may write
h
1
(B) = pr
1
(Gr(h) (E B)),
h
1
(F B) = pr
1
(Gr(h) (E (F B))),
which implies that h
1
(B) and h
1
(F B) are Suslin sets in E. In addition, we have
h
1
(B) h
1
(F B) = / 0, h
1
(B) h
1
(F B) = E.
By virtue of the separation principle, both sets h
1
(B) and h
1
(F B) must be Borel, which
shows that h is a Borel mapping.
The next auxiliary proposition may be regarded as a natural generalization of Theorem 5
to the case of a countable disjoint family of Suslin sets.
Lemma 4. Let E be a Polish space and let X
n
: n < be a countable disjoint family of
Suslin sets in E. Then there exists a countable disjoint family Y
n
: n < of Borel sets in
E such that X
n
Y
n
for all n <.
Proof. According to Theorem5, for any pair (n, m) of distinct natural numbers, there exists
a Borel set B
nm
such that
X
n
B
nm
E X
m
.
Now, we put
Y
0
=B
0m
: 1 m <
and for n 0, dene by recursion
Y
n
= (B
nm
: m <, m ,= n) (Y
0
Y
1
... Y
n1
).
It is easy to check that Y
n
: n < is the required family of Borel sets.
We also have the following important statement (see [99], [148], and [160]).
Theorem 6. Let E
1
and E
2
be two Polish spaces, X be a Borel subset of E
1
, and let a
mapping g : X E
2
be injective and continuous. Then the image g(X) is a Borel subset of
E
2
.
Proof. In view of Exercises 8 and 9, it sufces to show that for every Polish space E and
for any injective continuous mapping g :
E, the set ran(g) is Borel in E. In order to

establish this fact, x a natural number k and consider the corresponding disjoint countable
family of Suslin sets
g(U(s)) (s
k
).
It follows from the separation principle formulated in Lemma 4 that there are pairwise
disjoint Borel sets Y
s
(s
k
) in E for which the relations
g(U(s)) Y
s
(s
k
)
are satised. Now, we dene by recursion the following Borel sets:
Y
n
=Y
n
cl(g(U(n))) (n <),
Y
sn
=Y
sn
cl(g(U(s n))) Y
s
(s
<
, n <).
Using the method of induction on lh(s), it is not difcult to check the inclusions
Y
s
Y
s
, g(U(s)) Y
s
cl(g(U(s))) (s
<
).
Consequently, for any z
<
, we get the equality
g(z) =Y
z[k
: k <,
which directly implies that
g(
) =
k<
Y
z[k
).
Keeping in mind the circumstance that for each k < , the family Y
s
: s
k
of Borel
sets in E is disjoint, we come to the relation
g(
) =
k<
(
s
<
, lh(s)=k
Y
s
)
from which it follows that g(
) is a Borel subset of E. Theorem 6 has thus been proved.

Theorems 3 and 6 yield that the Borel subsets of Polish topological spaces coincide with the
bijective continuous images of G
-subsets of the canonical Baire space
(cf. Exercise
9). In addition to this result, it directly follows from Theorem 6 that if X and Y are two
Borel subsets of Polish spaces and
f : X Y
is a continuous bijection, then f
1
: Y X is a Borel bijection and, consequently, f turns
out to be a Borel isomorphism between X and Y.
Remark 3. The assertion of Theorem 6 remains valid in the case of an injective Borel
mapping g : X E
2
(see Exercise 10). In this context, it should be also mentioned that
a much stronger result holds true. Namely, let E and E
/
be two Polish spaces and let
f : E E
/
be a partial Borel mapping dened on a Borel subset of E and such that
card( f
1
(y))
for every point y E
/
. Then for any Borel set B E, the image f (B) is a Borel subset of
E
/
. Moreover, the set dom( f ) admits a representation
dom( f ) =X
n
: n <,
where all X
n
(n <) are pairwise disjoint Borel subsets of E and for each index n <, the
restriction f [X
n
is a Borel isomorphism between X
n
and f (X
n
) (see [99], [148], [160], and
[162]).
So far, we were concerned with various properties of Borel and analytic subsets of a Polish
space. As said earlier, these classes of sets are important from the point of view of nu-
merous applications. However, descriptive set theory also deals with more general classes
of sets, which are of paramount interest from the point of view of logical foundations of
contemporary mathematics. For example, one such class is formed by the so-called pro-
jective subsets of the real line, which were introduced and thoroughly studied by Luzin
and Sierpi nski (see [10], [91], [99], [148], [150], [160], [162], and [188]). Here we do
not intend to discuss deep properties of projective sets. Notice that some of them (e.g., the
measurability in the Lebesgue sense) turn out to be independent of ZFC theory. However,
they become provable under the assumption of the existence of large cardinals of a certain
type.
Below, we only present the classical denition of projective sets (cf. [148] and [150]).
Let E be an arbitrary Polish space. We dene the classes of sets
Pr
0
(E), Pr
1
(E), . . . , Pr
n
(E), . . .
by ordinary recursion. First, let us put
Pr
0
(E) =B(E).
Suppose now that for a natural number n 1, the classes Pr
k
(E), where k < n, have
already been determined.
If n is odd, then we dene Pr
n
(E) as the class of all continuous images (in E) of the sets
from Pr
n1
(E).
If n is even, then we dene Pr
n
(E) as the class of all complements of the sets from
Pr
n1
(E). Finally, we put
Pr(E) =Pr
n
(E) : n <.
By denition, any member of Pr(E) is called a projective subset of E. We thus see that
Pr
0
(E) =B(E), Pr
1
(E) =A(E),
i.e., Borel and analytic sets are only the rst two steps in the classical construction of the
Luzin-Sierpi nski projective hierarchy.
EXERCISES
Actually, deduce this lemma from a more general fact stating that if R is an algebra of
subsets of a base set E, then the -algebra generated by R coincides with the monotone
class generated by R.
Further, let L be a family of subsets of E such that:
(a) R L;
(b) L is closed under the unions of all countable disjoint families of its members;
(c) L is closed under the intersections of all countable families of its members.
Verify that L contains the -algebra generated by R.
2. Let <
1
. Show that the Baire class Ba
(E, R) is closed with respect to the uniform

convergence of sequences of its elements (cf. Theorem 2 from Chapter 20).
3
. Let E be an arbitrary topological space. By using the method of transnite induction,

prove that:
(a) Ba(E, R) B(E, R);
(b) if E is perfectly normal, then Ba(E, R) = B(E, R).
To show the validity of (b), argue in the following manner. First, reduce the argument to the
case of the characteristic function of any Borel subset X of E, i.e., prove that the function
X
belongs to Ba(E, R). For this purpose, use the Tietze-Urysohn theorem, Lemma 1 and
the method of transnite induction on the order of X (cf. Exercise 5 from Appendix 3).
4. Let E and E
/
be two Polish spaces and let f : E E
/
be a Borel mapping. Show that for
any Suslin set X E, its image f (X) is a Suslin subset of E
/
.
Obviously, the result just formulated generalizes Theorem 2 of this Appendix.
6. Let E be a nonempty complete metric space without isolated points. Show that there
exists a subset X of E which is homeomorphic to the Cantor space 0, 1
.
For this purpose, construct a dyadic system of closed balls in E similar to the system of
balls indicated in the proof of Theorem 4.
7. Let E be a zero-dimensional Polish topological space and let Z be a subset of E such
that:
(a) Z is of type G
;
(b) Z is everywhere dense in E;
(c) E Z is everywhere dense in E.
Prove that Z is homeomorphic to the canonical Baire space
.
For this purpose, show that Z can be expressed as the result of A-operation applied to some
regular system F
s
: s
<
of nonempty closed-open sets in E satisfying the relations:
(i) lim
lh(s)+
diam(F
s
) = 0;
(ii) F
s
F
t
= / 0 for any two distinct s
<
and t
<
such that lh(s) = lh(t).
8. Prove that an arbitrary uncountable zero-dimensional Polish space E can be represented
in the form
E = Z D (Z D = / 0),
where Z is homeomorphic to
and D is at most countable.

In order to show this fact, denote by P the set of all condensation points of E and observe
that P is an uncountable perfect subset of E whose complement is countable. Let D
0
be a
countable everywhere dense subset of P. Put
Z = P D
0
, D = (E P) D
0
and, keeping in mind the result of Exercise 7, check that these Z and D yield the required
decomposition of E.
9. Demonstrate that any Borel subset of a Polish space can be regarded as a bijective
continuous image of some Polish subspace of the Baire space
.
For this purpose, take into account the existence of a canonical continuous bijection acting
from the space
onto the interval ]0, 1], which implies the existence of a continuous
bijection of
onto the product space ]0, 1]
. Further, use the fact that ]0, 1]
contains topological copies of all Polish spaces and apply Theorem 3.

10. Let E
1
and E
2
be two Polish spaces, X be a Borel subset of E
1
and let g : X E
2
be an
injective Borel mapping. Prove that g(X) is a Borel subset of E
2
.
Argue in the following manner. In the product set X E
2
consider the graph of g, i.e.,
consider the set
G =(x, y) X E
2
: g(x) = y.
First, show that G is a Borel subset of X E
2
. For this purpose, denote by d any metric on
E
2
compatible with the topology of E
2
and introduce a function
: X E
2
[0, +[
(x, y) = d(g(x), y) (x X, y E
2
).
Check that is a Borel mapping and G =
1
(0), which yields that G is Borel. Finally,
consider the injective continuous mapping
pr
2
[G : G E
2
and apply to it Theorem 6 of this Appendix.
11
. Let X and X
/
be any two uncountable Borel subsets of Polish spaces E and E
/
respec-
tively. Demonstrate that there exists a Borel isomorphism
f : X X
/
.
For this purpose, rst show that the Cantor space 0, 1
is Borel isomorphic to the Hilbert

cube [0, 1]
. Then, applying the topological universality of [0, 1]
for the class of all sep-

arable metrizable spaces, the Alexandrov-Hausdorff theorem from this Appendix and Ba-
nachs theorem on two injections from Exercise 12 of Appendix 1, obtain the required
result.
12
. In this exercise, we will be dealing with the extended real line

[, +] = R, +
which is equipped with the standard ordering and topology. So [, +] becomes isomor-
phic to the closed interval [1, 1].
Let E be an uncountable Polish space. Prove the classical Lebesgue theorem stating that
for any ordinal <
1
, there exists a mapping
: E [0, 1] [, +]
(a)
is Borel;
(b) for each function f Ba
(E, R), there is a point t [0, 1] such that
(, t) = f .
In order to demonstrate this important statement, rst observe that in view of Exercise 11,
it sufces to reduce the argument to the case E = [0, 1].
Consider any Peano type mapping
= (
k
)
k<
: [0, 1] [0, 1]
.
This phrase simply means that is a continuous surjection. Let Q
k
: k < denote the
sequence of all polynomials on [0, 1] with rational coefcients. Show that we can take
0
(x, t) = limsup
k+
Q
k
(x)
k
(t) (x [0, 1], t [0, 1]).
If for <
1
, the mapping
is already dened, then we can put
+1
(x, t) = limsup
k+
(x,
k
(t)) (x [0, 1], t [0, 1]).
Finally, if <
1
is a limit ordinal, then take a strictly increasing sequence
k
: k < of
ordinals such that lim
k+
k
= and put
(x, t) = limsup
k+
k
(x,
k
(t)) (x [0, 1], t [0, 1]).
This procedure enables us to dene the required mapping
for any ordinal number <
1
.
By starting with the existence of
for each <

1
, derive that
Ba
([0, 1], R) Ba
([0, 1], R) : < ,= / 0

and that the same holds true for an arbitrary uncountable Polish space E instead of [0, 1].
Argue in the following manner. Suppose to the contrary that
Ba
([0, 1], R) Ba
([0, 1], R) : < = / 0

and deduce from this assumption that
Ba([0, 1], R) = Ba
([0, 1], R).

Then, by using the mapping
, construct a Borel mapping

: [0, 1] [0, 1] [0, 1]
such that for any Borel function f : [0, 1] [0, 1], there exists a point t [0, 1] for which
(, t) = f .
Further, dene a Borel mapping by the formula
(x, t) = lim
n+
n(x, t)
1 +n(x, t)
(x [0, 1], t [0, 1])
and check that:
(a) ran() =0, 1;
(b) for any Borel function g : [0, 1] 0, 1, there exists a point t [0, 1] such that (, t) =
g.
Finally, put
h(x) = 1 (x, x) (x [0, 1])
and verify that:
(c) the function h is Borel;
(d) ran(h) 0, 1.
Now, keeping in mind relation (b), take a point t [0, 1] for which h =(, t) and infer the
equalities
h(t) = 1 (t, t) =(t, t), (t, t) = 1/2,
contradicting relation (a).
Thus, the obtained contradiction yields the existence of Baire functions of any order <
1
and, consequently, the existence of Borel sets in [0, 1] of the same order .
13
. Let E be an arbitrary uncountable Polish space. Prove the classical Suslin theorem
(see [99], [148], [160], and [162]) stating that there exists a Suslin subset of E which is not
Borel (consequently, there exists a co-analytic subset of E which is not analytic).
Argue in the following manner. By virtue of Exercise 11, it sufces to establish this fact
for some concrete Polish space E, e.g., in the case where
E =0, 1
.
Consider the product space K = 0, 1
[0, 1] and denote by Comp(K) the family of all

compact subsets of K. Endowed with the standard Hausdorff metric, this family becomes
a compact metric space (see [58], [101]). Consequently, there exists a Peano type mapping
g : 0, 1
Comp(K),
i.e., g is a continuous surjection. Let I denote the set of all irrational numbers from [0, 1].
As widely known, this set is homeomorphic to the Baire space
. Dene a set-valued
mapping
G : 0, 1
P(0, 1
)
by the formula
G(t) = pr
1
((0, 1
I) g(t)) (t 0, 1
).
Verify the validity of the following relation:
ran(G) =A(0, 1
).
Further, in the Cantor space 0, 1
consider the set

X =t 0, 1
: t , G(t).
Suppose for a while that X is a Suslin set in 0, 1
. Then for some t

0
0, 1
, we must
have G(t
0
) = X. It easily follows from this equality that
t
0
X t
0
, X,
which yields a contradiction. Therefore, X cannot be a Suslin subset of 0, 1
. On the
other hand, show that the set
0, 1
X =t 0, 1
: t G(t)
is Suslin. For this purpose, check the validity of the relation
0, 1
X = pr
1
(D),
where the Borel set D K is dened as follows:
D =(t, r) K : (t, r) g(t) & r I.
Infer from the said above the required result.
14
. Let E be a base set, L be a class of subsets of E and let X

s
: s
<
be a countable
system of sets from L.
By using the method of transnite recursion on <
1
, dene the following sets:
(a) X
0
s
= X
s
for all s
<
;
(b) X
+1
s
= X
s
(X
sn
: n <) for any <
1
and for all s
<
;
(c) X
s
=X
s
: < for a limit ordinal < and for each s
<
.
Check (by using transnite induction) that if <
1
, then
X
s
X
s
(s
<
).
Further, for any <
1
, put:
Y
=X
n
: n <;
T
=(X
s
X
+1
s
) : s
<
;
Z
=Y
.
Verify that all the sets
X
s
, Y
, T
, Z
( <
1
, s
<
)
belong to the -ring generated by the given class L.
Prove the following important equalities due to Sierpi nski:
A(X
s
: s
<
) =Y
: <
1
=Z
: <
1
.
In order to establish these equalities, argue in the following manner. First, show the validity
of the inclusion
(i) Z
: <
1
A(X
s
: s <).
Take an arbitrary z Z
: <
1
. Then for some ordinal <
1
, we must have z Y
and z , T
. Consequently, for some natural number n

0
, the relations
z X
n
0
, z , X
n
0
X
+1
n
0
are satised, thus it follows that z X
+1
n
0
. Keeping in mind the inclusion
X
+1
n
0
X
n
0
n
: n <,
obtain that z X
n
0
n
1
for some natural number n
1
. Continuing this process by ordinary
recursion, construct an innite sequence
= (n
0
, n
1
, n
2
, ...)
such that
z X
[k
: k < X
[k
: k <
and, therefore, z A(X
s
: s
<
). This result obviously yields (i).
Now, check the validity of the inclusion
(ii) A(X
s
: s
<
) Y
: <
1
.
For this purpose, show that if <
1
, r < and
, then the relation

X
[k
: k < X
[r
holds true (use the method of transnite induction on <
1
). Consequently,
X
[k
: k < X
(0)
X
n
: n < =Y
,
which immediately yields (ii).
Finally, verify the validity of the relation
(iii) T
: <
1
= / 0.
To show this fact, suppose otherwise, i.e., suppose that there exists an element x T
:
<
1
. Then for any ordinal <
1
, we can nd a sequence s = s()
<
such that
x X
s
X
+1
s
.
Since
<
is countable and
1
is uncountable, there are two distinct ordinals <
1
and
<
1
for which s = s() = s(). We may assume, without loss of generality, that <.
But in this case the relation x X
s
X
+1
s
must be true, thus a contradiction follows.
Taking into account the relations (i), (ii) and (iii) established above, derive the validity of
the Sierpi nski equalities.
Infer from these equalities that any analytic set over the class L can be represented as the
intersection of an
1
-sequence of elements from the -ring generated by L, and as the
union of an
1
-sequence of elements from the same -ring.
Conclude that if D is a co-Suslin subset of a Polish topological space, then
card(D) 0, 1, ..., n, ..., ,
1
, c.
The same relation remains valid if D is any Borel image (contained in a Polish space) of a
co-Suslin set.
15. We preserve the notation of Exercise 14.
Suppose that a base set E is equipped with a topology satisfying the Suslin condition. Show
that if all sets from X
s
: s
<
have the Baire property, then the set X = A(X
s
: s
<
) has the Baire property, too.
For this purpose, take s
<
and consider the sets
X
s
X
+1
s
( <
1
),
which are pairwise disjoint. Assuming without loss of generality that E is a Baire space,
derive that there exists an ordinal (s) <
1
such that all sets X
s
X
+1
s
are of rst category
in E whenever (s) <
1
. Further, put
= sup(s) : s
<
and conclude that the set T
is of rst category in E. Taking into account the inclusions

X Z
,
obtain the desired result.
Suppose now that E is equipped with a -nite complete measure . Show that if all sets
from a family X
s
: s
<
are -measurable, then the set X = A(X
s
: s
<
) is
-measurable, too.
For this purpose, apply a method analogous to the described above.
In fact, the second part of Exercise 15 directly follows from its rst part. In order to
establish this implication, it sufces to consider a von Neumann topology associated with
a given -nite complete measure .
In addition, the invariance of the Baire property under the A-operation remains valid for an
arbitrary topological space E (see [148]).
16
. Prove that there exists a nonempty perfect subset P of R which is linearly independent
over the eld Q of all rational numbers.
By using the Zorn lemma, extend P to a Hamel basis of R and claim that there are Hamel
bases in R which contain nonempty perfect subsets.
On the other hand, show that no Hamel basis in R can be a Suslin subset of R (Sierpi nskis
theorem).
In connection with the result of Exercise 16, let us remark that von Neumann [186] was
able to construct a nonempty perfect set of algebraically independent real numbers.
17
. Let E and E
/
be two Polish spaces and let f : E E
/
be a Borel mapping such that
(y E
/
)(card( f
1
(y)) ).
If X E is a universal measure zero set, then its image f (X) is universal measure zero, too
(see Exercise 9 from Chapter 12).
Taking into account the above-mentioned fact and using Remark 3 of this Appendix, prove
that there exists an uncountable universal measure zero subset of R which is simultaneously
a vector space over the eld Q.
For this purpose, start with the existence of a nonempty perfect set P R linearly indepen-
dent over Q and pick an uncountable universal measure zero set X P. Denote by Y the
linear hull of X (over the same Q) and check that Y is a universal measure zero subset of
R.
18
. Show that there exists a Borel isomorphism f : [0, 1] [0, 1] which does not preserve
the Baire property of subsets of [0, 1].
On the other hand, let E
1
and E
2
be two uncountable Polish spaces without isolated points.
Prove that there exists a Borel isomorphism g : E
1
E
2
satisfying the following condition:
a set X E
1
is of rst category in E
1
if and only if its image g(X) is of rst category in E
2
.
Derive from this condition that both mappings g and g
1
preserve the Baire property.
In order to establish the existence of the required Borel isomorphism g, use Exercise 8 of
this Appendix.
19. Let X R be an analytic non-Borel set (such a set exists in view of the result of
Exercise 13). According to one of Sierpi nskis equalities, we have
X =Z
: <
1
where all Z
( <
1
) are some Borel subsets of R (see Exercise 14). Verify that
Z
: < ,= / 0
for uncountably many ordinals <
1
. By starting with this fact and taking into account the
argument presented in Exercise 15, give one more construction of an uncountable universal
measure zero set in R.
Bibliography
1. A. Ascherl, J. Lehn, Two principles of extending probability measures, Manuscripta Math., vol.
21, 1977, pp. 4350.
2. F. Bagemihl, P. Erd os, Intersections of prescribed power, type, or measure, Fund. Math., vol. 41,
1957, pp. 5767.
3. R. Baire, Sur les fonctions de variables r eelles, Ann. di Math., vol. 3, no. 3, 1899.
4. R. Baire, Sur la repr esentation des fonctions discontinues, Acta Math., vol. 30, 1905.
5. R. Baire, Lecons sur les Fonctions Discontinues, Gauthier-Villars, Paris, 1905.
6. R. Baker, Lebesgue measure on R
, Proc. Amer. Math. Soc., vol. 113, no. 4, 1991, pp. 1023
1029.
7. M. Balcerzak, K. Ciesielski, T. Natkaniec, Sierpi nskiZygmund functions that are Darboux,
almost continuous, or have a perfect road, Arch. Math. Logic, vol. 37, 1997, pp. 2935.
8. S. Baldwin, Martins Axiom implies a stronger version of Blumbergs theorem, Real Analysis
Exchange, vol. 16, 1990-1991, pp. 6773.
9. S. Banach,

Uber die Bairesche Kategorie gewisser Funktionenmengen, Studia Math., vol. 3,
1931, pp. 174179.
10. J. Barwise (editor), Handbook of Mathematical Logic, Elsevier, Amsterdam, 1977.
11. A. Berarducci, D. Dikranjan, Uniformly approachable functions and UA spaces, Rend. Ist.
Matematico, Univ. di Trieste, vol. 25, 1993, pp. 2356.
12. F. Bernstein, Zur Theorie der trigonometrischen Reihen, Sitzungsber. S achs. Akad. Wiss.
Leipzig. Math.-Natur. Kl., vol. 60, 1908, pp. 325338.
13. D. Bierlein,

Uber die Fortsetzung von Wahrscheinlichkeitsfeldern, Z. Wahr. Verw. Gebiete, vol.
1, 1962, pp. 2846.
14. A. Blass, Existence of bases implies the Axiom of Choice, Contemporary Mathematics, vol. 31,
1984, pp. 3133.
15. H. Blumberg, New properties of all real functions, Trans. Amer. Math. Soc., vol. 24, 1922, pp.
113128.
16. V.I. Bogachev, Foundations of Measure Theory, vol. 1, vol. 2, Dynamics, Moscow, 2003 (Rus-
sian).
17. R. Bolstein, Sets of points of discontinuity, Proc. Amer. Math. Soc., vol. 38, 1973, pp. 193197.
18. V.G. Boltjanskii, Hilberts Third Problem, Izd. Nauka, Moscow, 1977 (Russian).
19. V.G. Boltjanskii, V.A. Efremovich, Intuitive Topology, Bibliot. Kvant, Moscow, 1982 (Rus-
sian).
20. K. Bouhjar, J.J. Dijkstra, R.D. Mauldin, No n-point set is -compact, Proc. Amer. Math. Soc.,
vol. 129, 2000, pp. 621622.
21. K. Bouhjar, J.J. Dijkstra, J. van Mill, Three-point sets, Topology and its Applications, vol. 112,
2001, pp. 215227.
447
22. N. Bourbaki, Espaces Vectoriels Topologiques, Hermann et Cie, Paris, 19531955.
23. N. Bourbaki, Topologie G en erale, Ch. 7, Hermann et Cie, Paris, 1960.
24. N. Bourbaki, Topologie G en erale, Ch. 9, Hermann et Cie, Paris, 1960.
25. N. Bourbaki, Theory of Sets, Hermann et Cie, Paris, 1968.
26. N. Bourbaki, Integration, Izd. Nauka, Moscow, 1977 (Russian, translation from the French).
27. J.B. Brown, A measure-theoretic variant of Blumbergs theorem, Proc. Amer. Math. Soc., vol.
66, 1977, pp. 266268.
28. J.B. Brown, Variations on Blumbergs theorem, Real Analysis Exchange, vol. 9, 1983-1984, pp.
123137.
29. J.B. Brown, Restriction theorems in real analysis, Real Analysis Exchange, vol. 20, no. 1, 1994-
1995, pp. 510526.
30. J.B. Brown, G.V. Cox, Classical theory of totally imperfect spaces, Real Analysis Exchange, vol.
7, 1982, pp. 139.
31. J.B. Brown, K. Prikry, Variations on Lusins theorem, Trans. Amer. Math. Soc., vol. 302, 1987,
pp. 7786.
32. A. Bruckner, Differentiation of Real Functions, Springer-Verlag, Berlin, 1978.
33. A. Bruckner, J. Haussermann, Strong porosity features of typical continuous functions, Acta
Math. Hung., vol. 45, no. 1-2, 1985, pp. 713.
34. A.M. Bruckner, R.J. OMalley, B.S. Thomson, Path derivatives: a unied view of certain gener-
alized derivatives, Trans. Amer. Math. Soc., vol. 283, no. 1, 1984, pp. 97125.
35. M.R. Burke, K. Ciesielski, Sets on which measurable functions are determined by their range,
Canadian Journal of Mathematics, vol. 49, no. 6, 1997, pp. 10891116.
36. T. Carlson, Extending Lebesgue measure by innitely many sets, Pacic Journal of Mathematics,
vol. 115, 1984, pp. 3345.
37. J. Cicho n, A. Jasi nski, A note on algebraic sums of subsets of the real line, Real Analysis Ex-
change, vol. 28, no. 2, 2002-2003, pp. 493500.
38. J. Cicho n, A.B. Kharazishvili, On ideals with projective bases, Georgian Mathematical Journal,
vol. 9, no. 3, 2002, pp. 461472.
39. J. Cicho n, A. Kharazishvili, B. Weglorz, On sets of Vitalis type, Proc. Amer. Math. Soc., vol.
118, no. 4, 1993, pp. 12211228.
40. J. Cicho n, A.B. Kharazishvili, B. Weglorz, Subsets of the Real Line, L od z University Press,
L od z, 1995.
41. K. Ciesielski, Algebraically invariant extensions of -nite measures on Euclidean spaces,
Trans. Amer. Math. Soc., vol. 318, 1990, pp. 261273.
42. K. Ciesielski, Set-theoretic real analysis, Journal of Applied Analysis, vol. 3, no. 2, 1997, pp.
143190.
43. K. Ciesielski, Measure zero sets whose complex sum is nonmeasurable, Real Analysis Exchange,
vol. 26, no. 2, 2000-2001, pp. 919922.
44. K. Ciesielski, H. Fejzi c, C. Freiling, Measure zero sets with nonmeasurable sum, Real Analysis
Exchange, vol. 27, no. 2, 2001-2002, pp. 783794.
45. K. Ciesielski, A. Pelc, Extensions of invariant measures on Euclidean spaces, Fund. Math., vol.
125, 1985, pp. 110.
46. W.W. Comfort, Topological groups, in: Handbook of Set-Theoretic Topology, North-Holland
Publ. Co., Amsterdam, 1984, pp. 11431263.
47. I.P. Cornfeld, J.G. Sinaj, S.V. Fomin, Ergodic Theory, Izd. Nauka, Moscow, 1980 (Russian).
48. R.O. Davies, The power of the continuum and some propositions of plane geometry, Fund. Math.,
vol. LII, 1963, pp. 277281.
49. R.O. Davies, Covering the plane with denumerably many curves, Journal of London Math. Soc.,
vol. 38, 1963, pp. 433438.
50. R.O. Davies, Measures not approximable or not speciable by means of balls, Mathematika, vol.
Bibliography 449
18, 1971, pp. 157160.
51. R.O. Davies, Covering space with denumerably many curves, Bull. London Math. Soc., vol. 6,
1974, pp. 189190.
52. C. Dellacherie, Capacit es et Processus Stochastiques, Springer-Verlag, Berlin, 1972.
53. J. Dieudonn e, Un exemple despace normal non susceptible dune structure uniforme de espace
complet, C.R. Acad. Sci. Paris, vol. 209, 1939, pp. 145147.
54. J.J. Dijkstra, J. van Mill, Two point set extensions - a counterexample, Proc. Amer. Math. Soc.,
vol. 125, no. 8, 1997, pp. 25012502.
55. J.J. Dijkstra, J. van Mill, On sets that meet every hyperplane in n-space in at most n points, Bull.
London Math. Soc., vol. 34, 2002, pp. 361368.
56. N. Dunford, J.T. Schwartz, Linear Operators, Part 1: General Theory, Interscience Publishers,
New York, 1958.
57. R.E. Edwards, Functional Analysis, Holt, Rinehart and Winston, New York, 1965.
58. R. Engelking, General Topology, PWN, Warszawa, 1985.
59. P. Erd os, K. Kunen, R.D. Mauldin, Some additive properties of sets of real numbers, Fund. Math.,
vol. CXIII, no. 3, 1981, pp. 187199.
60. P. Erd os, R.D. Mauldin, The nonexistence of certain invariant measures, Proc. Amer. Math. Soc.,
vol. 59, 1976, pp. 321322.
61. P. Erd os, R. Rado, A partition calculus in set theory, Bull. Amer. Math. Soc., vol. 62, 1956, pp.
427489.
62. P. Erd os, A.H. Stone, On the sum of two Borel sets, Proc. Amer. Math. Soc., vol. 25, no. 2, 1970,
pp. 304306.
63. P. Erd os, G. Szekeres, A combinatorial problem in geometry, Compos. Math., vol. 2, 1935, pp.
463470.
64. M.P. Ershov, Measure extensions and stochastic equations, Probability Theory and its Applica-
tions, vol. 19, no. 3, 1974 (Russian).
65. H. Federer, Geometric Measure Theory, Springer-Verlag, New York, 1969.
66. K. Fischer, Z. Slodkowski, Christensen zero sets and measurable convex functions, Proc. Amer.
Math. Soc., vol. 79, no. 3, 1980, 449453.
67. D.H. Fremlin, Consequences of Martins Axiom, Cambridge University Press, Cambridge, 1984.
68. D.H. Fremlin, Measure additive coverings and measurable selectors, Dissertationes Mathemat-
icae, vol. CCLX, 1987.
69. D.H. Fremlin, Real-valued-measurable cardinals, Israel Math. Conf. Proc., vol. 6, 1993, pp.
151304.
70. L. Fuchs, Innite Abelian Groups, vol. 1, Academic Press, New York-London, 1970.
71. R.J. Gardner, W.F. Pfeffer, Borel measures, in: Handbook of Set-Theoretic Topology, North-
Holland Publ. Co., Amsterdam, 1984, pp. 9611043.
72. B.R. Gelbaum, J.M.H. Olmsted, Counterexamples in Analysis, Holden-Day, San Francisco,
1964.
73. M. Gitik, S. Shelah, Forcing with ideals and simple forcing notions, Israel Journal of Mathemat-
ics, vol. 68, no. 2, 1989, pp. 129160.
74. C. Goffman, Everywhere differentiable functions and the density topology, Proc. Amer. Math.
Soc., vol. 51, 1975, pp. 250251.
75. C. Goffman, C.J. Neugebauer, T. Nishiura, Density topology and approximate continuity, Duke
Math. Journal, vol. 28, 1961, pp. 497505.
76. P.M. Gruber, Aspects of convexity and its applications, Expositiones Mathematicae, vol. 2, 1984,
pp. 4783.
77. P.M. Gruber, J.M. Wills (editors), Handbook of Convex Geometry, vol. 1, vol. 2, North-Holland
Publ. Co., Amsterdam, 1990-1991.
78. E. Grzegorek, Remarks on -elds without continuous measures, Colloq. Math., vol. 39, 1978,
pp. 7375.
79. E. Grzegorek, Solution of a problem of Banach on -elds without continuous measures, Bull.
Acad. Polon. Sci., Ser. Sci. Math., vol. 28, 1980, pp. 710.
80. P.R. Halmos, Measure Theory, D. Van Nostrand, New York, 1950.
81. P.R. Halmos, Lectures on Ergodic Theory, The Mathematical Society of Japan, Tokyo, 1956.
82. G. Hamel, Eine Basis aller Zahlen und die unstetigen L osungen der Funktionalgleichung: f (x+
y) = f (x) + f (y), Math. Ann., vol. 60, 1905, pp. 459462.
83. E. Hewitt, K.A. Ross, Abstract Harmonic Analysis, vol. I, Springer-Verlag, Berlin-New York,
1963.
84. E. Hewitt, K. Stromberg, Some examples of nonmeasurable sets, Journ. Austral. Math. Soc., vol.
18, 1974, pp. 236238.
85. A. Hulanicki, Invariant extensions of the Lebesgue measure, Fund. Math., vol. 51, 1962, pp.
111115.
86. G.L. Itzkowitz, Extensions of Haar measure for compact connected Abelian groups, Indag.
Math., vol. 27, 1965, pp. 190207.
87. Zs. Jacab, M. Laczkovich, A characterization of Jordan and Lebesgue measures, Colloq. Math.,
vol. XL, no. 1, 1978, pp. 3952.
88. V. Jarnik, Sur la d erivabilit e des fonctions continues, Spisy Privodov, Fak. Univ. Karlovy, vol.
129, 1934, pp. 39.
89. J. Jasi nski, I. Reclaw, Restrictions to continuous and pointwise discontinuous functions, Real
Analysis Exchange, vol. 23, no. 1, 1997-1998, pp. 161174.
90. T. Jech, The Axiom of Choice, North-Holland Publ. Co., Amsterdam, 1973.
91. T. Jech, Set Theory, Academic Press, New York-London, 1978.
92. M.I. Kabenyuk, Dense subgroups of locally compact abelian groups, Siberian Math. J., vol. 21,
1980, pp. 902903.
93. S. Kakutani, Two xed-point theorems concerning bicompact convex sets, Proc. Imp. Acad.
Tokyo, vol. 14, 1938, pp. 242245.
94. S. Kakutani, On cardinal numbers related with a compact Abelian group, Proc. Imperial Acad.
Tokyo, vol. 19, 1943, pp. 366372.
95. S. Kakutani, J.C. Oxtoby, Construction of a nonseparable invariant extension of the Lebesgue
measure space, Ann. Math., vol. 52, 1950, pp. 580590.
96. A. Kamburelis, A new proof of the Gitik-Shelah theorem, Israel Journal of Mathematics, vol. 72,
1990, pp. 373380.
97. M.P. Kats, The continuity of universally measurable linear mappings, Sib. Mat. Zhurnal, vol. 23,
1982, 8390 (Russian).
98. Y. Katznelson, K. Stromberg, Everywhere differentiable, nowhere monotone functions, Amer.
Math. Monthly, vol. 81, 1974, pp. 349354.
99. A.S. Kechris, Classical Descriptive Set Theory, Springer-Verlag, New York, 1995.
100. J.L. Kelley, The Tychonoff product theorem implies the Axiom of Choice, Fund. Math., vol. 37,
1950, pp. 7576.
101. J.L. Kelley, General Topology, D. Van Nostrand, New York, 1955.
102. A. Kert esz, T. Szele, On the existence of non-discrete topologies in innite abelian groups,
Publ. Math. Debrecen, vol. 3, 1953, pp. 187189.
103. M.A. Khan, Chain conditions on subgroups of LCA groups, Pacic Journal of Mathematics,
vol. 86, 1980, pp. 517534.
104. A.B. Kharazishvili, On certain types of invariant measures, Dokl. Akad. Nauk SSSR, vol. 222,
no. 3, 1975, pp. 538540 (Russian).
105. A.B. Kharazishvili, Questions in the Theory of Sets and in Measure Theory, Izd. Tbil. Gos.
Univ., Tbilisi, 1978 (Russian).
106. A.B. Kharazishvili, On a property of Hamel bases, Bull. Acad. Sci. GSSR, vol. 95, no. 2, 1979
Bibliography 451
(Russian).
107. A.B. Kharazishvili, Absolutely nonmeasurable sets in Abelian groups, Bull. Acad. Sci. GSSR,
vol. 97, no. 3, 1980 (Russian).
108. A.B. Kharazishvili, Invariant Extensions of the Lebesgue Measure, Izd. Tbil. Gos. Univ., Tbil-
isi, 1983 (Russian).
109. A.B. Kharazishvili, Topological Aspects of Measure Theory, Naukova Dumka, Kiev, 1984
(Russian).
110. A.B. Kharazishvili, On invariant measures in Hilbert space, Bull. Acad. Sci. GSSR, vol. 114,
no. 1, 1984 (Russian).
111. A.B. Kharazishvili, Some properties of isodyne topological spaces, Bull. Acad. Sci. GSSR, vol.
127, no. 2, 1987, pp. 261264 (Russian).
112. A.B. Kharazishvili, On cardinalities of isodyne topological spaces, Bull. Acad. Sci. GSSR, vol.
137, no. 1, 1990, pp. 3336 (Russian).
113. A.B. Kharazishvili, Selected Topics in General Topology, Izd. Tbil. Gos. Univ., Tbilisi, 1990
(Russian).
114. A.B. Kharazishvili, On countably generated invariant -algebras which do not admit measure
type functionals, Real Analysis Exchange, vol. 23, no. 1, 1997-1998, pp. 287294.
115. A.B. Kharazishvili, Transformation Groups and Invariant Measures, World Scientic Publ.
Co., London-Singapore, 1998.
116. A.B. Kharazishvili, Applications of Point Set Theory in Real Analysis, Kluwer Academic Pub-
lishers, Dordrecht, 1998.
117. A.B. Kharazishvili, Some remarks on quasi-invariant and invariant measures, Real Analysis
Exchange, vol. 24, no. 1, 1998-1999, pp. 427434.
118. A.B. Kharazishvili, On vector sums of measure zero sets, Georgian Mathematical Journal, vol.
8, no. 3, 2001, pp. 493498.
119. A.B. Kharazishvili, Nonmeasurable Sets and Functions, North-Holland Mathematics Studies,
Elsevier, Amsterdam, 2004.
120. A.B. Kharazishvili, On generalized step-functions and superposition operators, Georgian
Mathematical Journal, vol. 11, no. 4, 2004, pp. 753758.
121. A.B. Kharazishvili, The algebraic sum of two absolutely negligible sets can be an absolutely
nonmeasurable set, Georgian Mathematical Journal, vol. 12, no. 3, 2005, pp. 455460.
122. A.B. Kharazishvili, Strange Functions in Real Analysis, (2nd edition), Chapman & Hall/CRC,
Boca Raton, 2006.
123. A.B. Kharazishvili, On additive absolutely nonmeasurable SierpinskiZygmund functions, Real
Analysis Exchange, vol. 31, no. 2, 2006.
124. A.B. Kharazishvili, On measurable Sierpi nskiZygmund functions, Journal of Applied Analy-
sis, no. 12, no. 2, 2006, pp. 283292.
125. A.B. Kharazishvili, Some aspects of the measure extension problem, Proc. A. Razmadze Math.
Inst., vol. 143, 2007, pp. 7378.
126. A.B. Kharazishvili, A nonseparable extension of the Lebesgue measure without new null-sets,
Real Analysis Exchange, vol. 33, no. 1, 2008.
127. A.B. Kharazishvili, On bad descriptive structure of Minkowskis sum of certain small sets in a
topological vector space, Theory of Stochastic Processes, vol. 14 (30), issue 2, 2008.
128. A.B. Kharazishvili, Some properties of step-functions connected with extensions of measures,
Acta Universitatis Carolinae, Ser. Phys-Mat., vol. 49, 2009.
129. A.B. Kharazishvili, Metrical transitivity and nonseparable extensions of invariant measures,
Taiwanese Journal of Mathematics, vol. 13, no. 3, 2009.
130. A.B. Kharazishvili, A.P. Kirtadze, On the measurability of functions with respect to certain
classes of measures, Georgian Mathematical Journal, vol. 11, 2004, pp. 489494.
131. A.B. Kharazishvili, A.P. Kirtadze, On algebraic sums of measure zero sets in uncountable com-
mutative groups, Proceedings of A.Razmadze Mathematical Institute, vol. 135, 2004, pp.
97103.
132. A.B. Kharazishvili, A.P. Kirtadze, On algebraic sums of absolutely negligible sets, Proc. A.
Razmadze Math. Institute, vol. 136, 2004, pp. 5561.
133. A.B. Kharazishvili, A.P. Kirtadze, Nonseparable left-invariant measures on uncountable solv-
able groups, Proc. A. Razmadze Math. Institute, vol. 139, 2005, pp. 4552.
134. A.B. Kharazishvili, A.P. Kirtadze, On weakly metrically transitive measures and nonmeasur-
able sets, Real Analysis Exchange, vol. 32, no. 2, 2006-2007, pp. 553562.
135. A.B. Kharazishvili, A.P. Kirtadze, On nonmeasurable subgroups of uncountable solvable
groups, Georgian Mathematical Journal, vol. 14, no. 3, 2007, pp. 435444.
136. A.B. Kharazishvili, A.P. Kirtadze, On extensions of partial functions, Expositiones Mathemat-
icae, vol. 25, no. 4, 2007, pp. 345353.
137. A.B. Kharazishvili, A.P. Kirtadze, On measurability of algebraic sums of small sets, Studia
Scientiarum Mathematicarum Hungarica, vol. 45, no. 3, 2008, pp. 433442.
138. B. King, Some remarks on difference sets of Bernstein sets, Real Analysis Exchange, vol. 19,
no. 2, 1993-1994, pp. 478490.
139. A.P. Kirtadze, On the method of direct products in the theory of quasi-invariant measures,
Georgian Mathematical Journal, vol. 12, no. 1, 2005, pp. 115120.
140. M. Kneser, Summenmengen in lokalkompakten Abelschen Gruppen, Math. Z., vol. 66, 1956,
pp. 88110.
141. S. Kodaira, S. Kakutani, A nonseparable translation-invariant extension of the Lebesgue mea-
sure space, Ann. Math., vol. 52, 1950, pp. 574579.
142. A. Krawczyk, P. Zakrzewski, Extensions of measures invariant under countable groups of
transformations, Trans. Amer. Math. Soc., vol. 326, no. 1, 1991, pp. 211226.
143. M. Kuczma, An Introduction to the Theory of Functional Equations and Inequalities: Cauchys
Equation and Jensens Inequality, PWN, Warszawa-Katowice, 1985.
144. K. Kunen, Combinatorics, in: Handbook of Mathematical Logic, North-Holland Publ. Co.,
Amsterdam, 1977.
145. K. Kunen, Set Theory, North-Holland Publ. Co., Amsterdam, 1980.
146. K. Kunen, Random and Cohen reals, in: Handbook of Set-Theoretic Topology, North-Holland
Publ. Co., Amsterdam, 1984.
147. K. Kunen, J.E. Vaughan (editors), Handbook of set-theoretic topology, North-Holland Publ.
Co., Amsterdam, 1984.
148. K. Kuratowski, Topology, vol. 1, Academic Press, New York-London, 1966.
149. K. Kuratowski, Topology, vol. 2, Academic Press, New York-London, 1968.
150. K. Kuratowski, A. Mostowski, Set Theory, North-Holland Publ. Co., Amsterdam, 1967.
151. K. Kuratowski, C. Ryll-Nardzewski, A general theorem on selectors, Bull. Acad. Polon. Sci.,
ser. Math., vol. 13, 1965, pp. 397402.
152. A.G. Kurosh, The Theory of Groups, Izd. Nauka, Moscow, 1967 (Russian).
153. I. Labuda, Universal measurability and summable families in topological vector spaces, Indag.
Math., vol. 82, 1979, pp. 2734.
154. I. Labuda, R.D. Mauldin, Problem 24 of the Scottish Book concerning additive functionals,
Colloq. Math., vol. XLVIII, no. 1, 1984, 8991.
155. M. Laczkovich, Paradoxes in measure theory, in: Handbook of Measure Theory, Elsevier,
Amsterdam, 2002, pp. 83123.
156. D.G. Larman, A problem of incidence, Journ. London Math. Soc., vol. 43, 1968, pp. 407409.
157. M. Lavrentieff, Contribution a la th eorie des ensembles hom eomorphes, Fund. Math., vol. 6,
1924.
158. H. Lebesgue, Lecons sur lint egration et la Recherche des Fonctions Primitives, Paris, 1904.
159. N. Lusin, Sur une probl eme de M. Baire, C. R. Acad. Sci. Paris, vol. 158, 1914, p. 1259.
Bibliography 453
160. N. Lusin, Lecons sur les Ensembles Analytiques et leurs Applications, Gauthier-Villars, Paris,
1930.
161. N.N. Luzin, Integral and Trigonometric Series, GITTL, Moscow, 1956 (Russian).
162. N.N. Luzin, Collected Works, vol. 1, vol. 2, vol. 3, Izd. Akad. Nauk SSSR, Moscow, 1958
(Russian).
163. R.D. Mabry, Sets which are well-distributed and invariant relative to all isometry invariant
total extensions of Lebesgue measure, Real Analysis Exchange, vol. 16, 1990-1991, pp.
425459.
164. R.D. Mabry, Some remarks concerning the uniformly gray sets of G. Jacopini, Rend. Accad.
Naz. Sci., Mem. Mat. Appl., vol. 22, no. 5, 1998, pp. 4349.
165. G.W. Mackey, Borel structures in groups and their duals, Trans. Amer. Math. Soc., vol. 85,
1957, pp. 134169.
166. D. Maharam, On a theorem of von Neumann, Proc. Amer. Math. Soc., vol. 9, 1958, pp. 987
994.
167. J. Mal y, The Peano curve and the density topology, Real Analysis Exchange, vol. 5, 1979-1980,
pp. 326329.
168. A.A. Markov, Some theorems on Abelian sets, DAN SSSR, vol. 1, 1936, pp. 299302 (Russian).
169. R.D. Mauldin, On sets which meet each line in exactly two points, Bull. London Math. Soc.,
vol. 30, 1998, pp. 397403.
170. S. Mazurkiewicz, Sur un ensemble plan qui a avec chaque droite deux et seulement deux points
communs, C. R. Varsovie, vol. 7, 1914, pp. 382384.
171. S. Mazurkiewicz, Sur les fonctions non-derivables, Studia Math., vol. 3, 1931, pp. 9294.
172. A.W. Miller, Special subsets of the real line, in: Handbook of Set-Theoretic Topology, North-
Holland Publ. Co., Amsterdam, 1984.
173. H. Minkowski, Geometrie der Zahlen, Leipzig, 1896.
174. M.D. Montgomery, Nonseparable metric spaces, Fund. Math., vol. 25, 1935, pp. 527533.
175. M.D. Montgomery, L. Zippin, Topological Transformation Groups, Interscience Publishers,
New York, 1955.
176. J.C. Morgan II, Point Set Theory, Marcel Dekker, Inc., New York-Basel, 1990.
177. S.A. Morris, Pontryagin Duality and the Structure of Locally Compact Abelian Groups, Cam-
bridge University Press, Cambridge, 1977.
178. W. Morris, V. Soltan, The Erd os-Szekeres problem on points in convex position - a survey, Bull.
Amer. Math. Soc., vol. 37, no. 4, 2000-2001, pp. 437458.
179. Y.V. Mospan, A converse to a theorem of Steinhaus, Real Analysis Exchange, vol. 31, no. 1,
2005/2006, pp. 291294.
180. S. Mr owka, Compactness and product spaces, Colloq. Math., vol. 7, 1959-1960, pp. 1922.
181. J. Mycielski, Independent sets in topological algebras, Fund. Math., vol. 55, 1964, pp. 139
147.
182. L. Nachbin, The Haar Integral, D. Van Nostrand, Princeton, NJ, 1965.
183. I.P. Natanson, The Theory of Functions of a Real Variable, Izd. Nauka, Moscow, 1957 (Rus-
sian).
184. T. Natkaniec, Almost Continuity, WSP, Bydgoszcz, 1992.
185. T. Natkaniec, H. Rosen, An example of an additive almost continuous Sierpi nskiZygmund
function, Real Analysis Exchange, vol. 30, 2004-2005, pp. 261266.
186. J. (von) Neumann, Ein System algebraisch unabh angiger Zahlen, Math. Ann., vol. 99, 1928,
pp. 134141.
187. J. Neveu, Bases Math ematiques du Calcul des Probabilit es, Masson et Cie, Paris, 1964.
188. P.S. Novikov, Selected Works, Izd. Nauka, Moscow, 1979 (in Russian).
189. B. Novotn y, On subcontinuity, Real Analysis Exchange, vol. 31, no. 2, 2005-2006, pp. 535
545.
190. W. Orlicz, Zur Theorie der Differentialgleichung y
/
= f (x, y), Polska Akademia Umiejetnosci,
Krakow, 1932, pp. 221228.
191. J.C. Oxtoby, Invariant measures on groups which are not locally compact, Trans. Amer. Math.
Soc., vol. 60, 1946, pp. 215237.
192. J.C. Oxtoby, Measure and Category, Springer-Verlag, New York-Berlin, 1971.
193. G. Pantsulaia, An application of independent families of sets to the measure extension problem,
Georgian Mathematical Journal, vol. 11, no. 2, 2004, pp. 379390.
194. E. Pap (editor), Handbook of Measure Theory, vol. 1, vol. 2, Elsevier, Amsterdam, 2002.
195. A. Pelc, Invariant measures and ideals on discrete groups, Dissertationes Mathematicae, vol.
255, 1986.
196. I.G. Petrovskii, Lectures on the Theory of Ordinary Differential Equations, Izd. Nauka,
Moscow, 1984 (Russian).
197. W.F. Pfeffer, K. Prikry, Small spaces, Proc. London Math. Soc., vol. 58, 1989, pp. 417438.
198. Sh.S. Pkhakadze, The theory of Lebesgue measure, Proc. A. Razmadze Math. Institute, vol. 25,
1958, pp. 3272 (in Russian).
199. K. Potka, Sum of Sierpi nskiZygmund and Darboux like functions, Topology and its Applica-
tions, vol. 122, 2002, pp. 547564.
200. H. Poincar e, Les rapports de lanalyse et de la physique math ematique, Revue g en. des sci.
pures et appl., vol. 8, 1897, pp. 857861.
201. L. Pontrjagin, The theory of topological commutative groups, Ann. Math., vol. 35, no. 2, pp.
361388.
202. L. Pontrjagin, Topological Groups, Princeton University Press, Princeton, NJ, 1939.
203. M. Rajagopalan, H. Subrahmanian, Dense subgroups of locally compact groups, Colloq. Math.,
vol. 35, 1976, pp. 289292.
204. F.P. Ramsey, On a problem of formal logic, Proc. London Math. Soc., ser. 2, vol. 30, 1930, pp.
264286.
205. M.M. Rao, Stochastic Processes and Integration, Sijthoff and Noordhoff, Alphen aan den Rijn,
1979.
206. I. Reclaw, Restrictions to continuous functions and Boolean algebras, Proc. Amer. Math. Soc.,
vol. 118, no. 3, 1993, pp. 791796.
207. C.A. Rogers, A linear Borel set whose difference set is not a Borel set, Bull. London Math.
Soc., vol. 2, 1970, pp. 4142.
208. A. Roslanowski, S. Shelah, Measured creatures, Israel Journal of Mathematics, vol. 151, 2006,
pp. 61110.
209. W. Rudin, Functional Analysis, (2nd edition), McGraw-Hill Science, New York, 1991.
210. S. Saks, Theory of the Integral, Warszawa-Lw ow, 1937.
211. S. Shelah, Innite Abelian groups - Whitehead problem and some constructions, Israel Journal
of Mathematics, vol. 18, 1974, pp. 243256.
212. S. Shelah, A compactness theorem for singular cardinals, free algebras, Whitehead problem
and transversals, Israel Journal of Mathematics, vol. 21, 1975, pp. 319349.
213. S. Shelah, On a problem of Kurosh, J onsson groups, and applications, in: Word Problems, II,
North-Holland Publ. Co., Amsterdam, 1980, pp. 373394.
214. S. Shelah, Possibly every real function is continuous on a non-meagre set, Publ. Inst. Mat.
Beograd (N.S.), vol. 57, no. 71, 1995, pp. 4760.
215. J.R. Shoeneld, Mathematical Logic, Addison-Wesley, New York, 1967.
216. J.R. Shoeneld, Martins Axiom, Amer. Math. Monthly, vol. 82, no. 6, 1975, pp. 610617.
217. W. Sierpi nski, Laxiome de M. Zermelo et son role dans la th eorie des ensembles et lanalyse,
Bull. Intern. Acad. Sci., Cracovie, Ser. A, 1918, pp. 97152.
218. W. Sierpi nski, Sur un th eor eme equivalent a lhypoth ese du continu, Bull. Intern. Acad. Sci.
Cracovie, Ser. A, 1919, pp. 13.
Bibliography 455
219. W. Sierpi nski, Sur la question de la mesurabilite de la base de M.Hamel, Fund. Math., vol. 1,
1920, pp. 105111.
220. W. Sierpi nski, Sur les fonctions convexes mesurables, Fund. Math., vol. 1, 1920, pp. 125128.
221. W. Sierpi nski, Sur un probleme concernant les fonctions semi-continues, Fund. Math., vol. 28,
1937, pp. 16.
222. W. Sierpi nski, Hypoth ese du Continu, (second edition), Chelsea Publ. Co., 1956.
223. W. Sierpi nski, Cardinal and Ordinal Numbers, PWN, Warszawa, 1958.
224. W. Sierpi nski, E. Szpilrajn (E. Marczewski), Remarque sur le probleme de la mesure, Fund.
Math., vol. 26, 1936, pp. 256261.
225. W. Sierpi nski, A. Zygmund, Sur une fonction qui est discontinue sur tout ensemble de puissance
du continu, Fund. Math., vol. 4, 1923, pp. 316318.
226. A.V. Skorokhod, Integration in Hilbert Space, Springer-Verlag, Berlin, 1975.
227. B.S. Sodnomov, On arithmetical sums of sets, Dokl. Akad. Nauk SSSR, vol. 80, 1951, pp.
173175 (Russian).
228. B.S. Sodnomov, An example of two G
-sets whose arithmetical sum is not Borel measurable,

Dokl. Akad. Nauk SSSR, vol. 99, 1954, pp. 507510 (Russian).
229. S. Solecki, On sets nonmeasurable with respect to invariant measures, Proc. Amer. Math. Soc.,
vol. 119, no. 1, 1993, pp. 115124.
230. S. Solecki, Measurability properties of sets of Vitalis type, Proc. Amer. Math. Soc., vol. 119,
no. 3, 1993, pp. 897902.
231. R.M. Solovay, Real-valued measurable cardinals, Proc. of Symposia in Pure Mathematics, vol.
13, part 1, 1971, pp. 397428.
232. J. Stallings, Fixed point theorems for connectivity maps, Fund. Math., vol. 47, 1959, pp. 249
263.
233. H. Steinhaus, Sur les distances des points des ensembles de mesure positive, Fund. Math., vol.
1, 1920, pp. 93104.
234. E. Szpilrajn (E. Marczewski), Sur lextension de la mesure lebesguienne, Fund. Math., vol. 25,
1935, pp. 551558.
235. E. Szpilrajn (E. Marczewski), On problems of the theory of measure, Uspekhi Mat. Nauk, vol.
1, no. 2 (12), 1946, pp. 179188 (Russian).
236. M. Talagrand, Sommes vectorielles densembles de mesure nulle, C.R. Acad. Sci. Paris, vol.
280, 1975, pp. 853855.
237. F. Tall, The density topology, Pacic Journal of Mathematics, vol. 62, 1976, pp. 275284.
238. S.Ulam, Zur Masstheorie in der allgemeinen Mengenlehre, Fund. Math., vol. 16, 1930, pp.
140150.
239. G. Vitali, Sul problema della misura dei gruppi di punti di una retta, Bologna, Italy, 1905.
240. S. Wagon, The Banach-Tarski Paradox, Cambridge University Press, Cambridge, 1985.
241. H.E. White Jr., Topological spaces in which Blumbergs theorem holds, Proc. Amer. Math. Soc.,
vol. 44, 1974, pp. 454462.
242. P. Zakrzewski, Extensions of isometrically invariant measures on Euclidean spaces, Proc.
Amer. Math. Soc., vol. 110, no. 2, 1990, pp. 325331.
243. P. Zakrzewski, Paradoxical decompositions and invariant measures, Proc. Amer. Math. Soc.,
vol. 111, no. 2, 1991, pp. 533539.
244. P. Zakrzewski, When do equidecomposable sets have equal measures?, Proc. Amer. Math. Soc.,
vol. 113, no. 3, 1991, pp. 831837.
245. P. Zakrzewski, The existence of invariant -nite measures for a group of transformations,
Israel Journal of Mathematics, vol. 83, 1993, pp. 275287.
246. P. Zakrzewski, The existence of invariant probability measures for a group of transformations,
Israel Journal of Mathematics, vol. 83, 1993, pp. 343352.
247. P. Zakrzewski, Extending invariant measures on topological groups, Ann. New York Acad. of
Sciences, 1996, pp. 218222.
248. P. Zakrzewski, The uniqueness of Haar measure and set theory, Colloq. Math., vol. 74, no. 1,
1997, pp. 109121.
249. P. Zakrzewski, Measures on algebraic-topological structures, in: Handbook of Measure The-
ory, North-Holland Publ. Co., 2002, pp. 10911130.
250. P. Zakrzewski, On a construction of universally small sets, Real Analysis Exchange, vol. 28,
2002-2003, pp. 221226.
251. P. Zakrzewski, On nonmeasurable selectors of countable group actions, (to appear).
Subject Index
Absolutely measurable function, 89
absolutely negligible set, 31
absolutely nonmeasurable function, 11
absolutely nonmeasurable set, 80
A.D. Alexandrovs theorem, 26
admissible family of sets for a measure,
266
admissible family of subgroups, 168
admissible partition for a measure, 393
admissible vector subspace, 78
Alexandrov-Hausdorff theorem, 431
algebraic aspect of the measure extension
problem, 19
algebraic sum of sets, 177
almost continuous function, 111
almost disjoint family of sets, 355
almost invariant set, 25
almost measurable function, 111
analytic manifold, 298
analytic set, 132
A-operation, 132
approximate derivative, 398
approximately continuous function, 398
approximately differentiable function,
398
Aronszajn tree, 343
autonomous system of ordinary
differential equations, 54
Axiom of Choice, 4
Baire classes of functions, 315
Baire order, 7
Baire property, 3
Banach-Kuratowski-Pettis theorem, 49,
394
Banach-Mazur theorem, 373
Banach theorem on two injections, 352,
353
Bernstein construction, 86, 91
Bernstein set, 34, 88, 91, 205
Blumbergs theorem, 11, 141
Borel measure, 40
Borel partial function, 7
Borel selector, 129, 131
branch of a tree, 340
branch property, 341
Browers theorem, 16
Canonical Baire space, 377, 426
Cantor-Baire stationarity principle, 325
Cantor-Bendixson theorem, 326, 334
457
Cantor set on the real line, 195
capacity in the Choquet sense, 14
Caratheodory conditions, 12
Caratheodorys theorem, 19
cardinal measurable in the Ulam sense,
61
Cauchy functional equation, 9
Choquet theorem, 359, 361
closed equivalence relation, 128
closed mapping, 125
co-analytic set, 96
commutative group, 10
complete partially ordered set, 17
completely regular topological space,
255
continuous nowhere approximately
differentiable function, 397
continuous nowhere differentiable
function, 126, 135
Continuum Hypothesis, 8
convex body, 145
convex polygon, 312
convexly independent set, 302
countably continuous partial function, 9
Darboux property, 123
debut of a set, 368
decreasing partial function, 7
density point, 397
density topology, 404
Dieudonne measure, 392
diffused measure, 9
Dini right derived number, 16
Dirichlet function, 122
divisible commutative group, 10
domain of a partial function, 2
dynamical system, 159
Erdos-Szekeres theorem, 302
ergodic measure, 147
Family of nonempty sets, 4
xed-point property, 16
free commutative group, 412
Generalized Luzin set, 84, 202
generalized Sierpinski set, 327
global solution of an autonomous system,
55
G-negligible set, 91
Godels incompleteness theorem, 2
G-thick set, 146
G-thin set, 146
Haar measure, 41
Hahn-Banach theorem, 1
Hamel basis, 92
Hamiltonian system, 56
height of a tree, 339
Henrys theorem, 27
Hewitts theorem, 170, 174
homogeneous covering, 310
Increasing partial function, 6
independent family of sets, 311
inductive limit of a family of groups, 414
inductive limit of a family of invariant
measures, 47
injective commutative group, 415
intermediate value property, 123
invariant measure, 24
Subject Index 459
isodyne topological space, 93, 172
Jarniks theorem, 329, 404
Jordan curve, 308
Konig lemma, 60, 340
Konig tree, 340
Kroneckers theorem, 56
Krylov-Bogoliubov theorem, 40
Kulikovs theorem, 165
Kuratowskis theorem on closed
projection, 125
Lavrentievs theorem, 137, 138
Lebesgue integral, 19
Lebesgue measurable function, 8
Lebesgue measure, 14, 41
Lebesgue theorem on density points, 58
level of a tree, 339
Liouville theorem, 39
Lipschitz condition, 16
locally compact topological group, 41
local solution of an autonomous system,
55
locally nite measure, 41
lower semicontinuous partial function, 5
Luzin-Jankov-von Neumann theorem,
112, 376
Luzins C-property, 37
Luzin set, 11
Mackeys theorem, 66
Marczewskis characteristic function,
157, 348
Marczewskis method of extending
measures, 20
Markov-Kakutani theorem, 40, 51
Martins Axiom, 8
Mazurkiewicz set, 307
measurable functional graph, 366
measurable selector, 13, 359
measure extension problem, 19
metric space canonically associated with a
measure, 18
metrically transitive measure, 147
mid-point convex function, 294
Minkowskis sum of sets, 199
Minkowskis theorem, 145
monothetic topological group, 42
monotone class of sets, 57
monotone partial function, 7
Montgomery operation, 390
Negligible set, 91
nonmeasurable set, 14
nonseparable extension of an invariant
measure, 257
nonseparable measure, 18
normal topological space, 1
normed vector space, 1
nowhere constant function, 337
nowhere Lebesgue measure zero set, 225
null-set, 241
One-dimensional unit torus, 42
one-parameter group of transformations,
55
oscillation of a function at a point, 137
ot-set, 310
Partial function, 1
partial isometry, 17
partial selector of a family of sets, 4
partially ordered set, 17
partially recursive function, 2
Peano type mapping, 205
perfect measure, 37
perfectly meager topological space, 326
perfectly normal topological space, 320
pleasant point with respect to a function,
140
Poincare theorem on recurrence, 157
points in general position, 302
Polish locally compact group, 42
problem of extending partial functions, 1
product of a family of groups, 44
projective commutative group, 412
projective set, 437
Quasi-compact class of sets, 378
quasicompact topological space, 4
quasi-cyclic group, 414
quasi-invariant measure, 24
quasi-polygon, 312
quotient Boolean algebra, 244
quotient space, 128
Radon measure, 26
Radon-Nikodym derivative, 64
Radon-Nikodym theorem, 64
Radon topological space, 37
Ramsey theorem, 303
range of a partial function, 3
real-valued measurable cardinal, 61
recursive function, 2
relatively measurable function, 79
relatively measurable set, 80
resolvable topological space, 170
Riemann integral, 19
Riesz theorem, 49
right invariant mean, 278
root of a tree, 339
Scattered topological space, 334
selector of a family of sets, 4
separable measure, 18
separation principle for analytic sets, 8
set-function, 13
set-theoretical aspect of the measure
extension problem, 19
set-valued mapping, 16
Sierpinski partition, 15, 346
Sierpinski set, 100
Sierpinski-Zygmund function, 9, 215
solvable group, 159, 168
Sorgenfrey topology, 94
Steinhaus property, 48
step-function, 22
step-function associated with a countable
partition, 97
support of a measure, 383
Suslin line, 338, 394
Suslin set, 89
Suslin theorem, 441
symmetric group, 23
Tarskis xed-point theorem, 17
theorem of Kuratowski and
Ryll-Nardzewski, 13, 373
theorem on measurable projection, 366
thick set, 29
Tietze-Urysohn theorem, 1
Subject Index 461
topological aspect of the measure
extension problem, 19
topologically complete metrizable space,
142
totally imperfect set, 34
transformation group, 23
tree, 339
Tychonoff theorem, 4
Ulams theorem, 21
Ulam transnite matrix, 260
universal measure zero set, 11
universally measurable function, 89
universally measurable functional in the
generalized sense, 284
unpleasant point with respect to a
function, 140
upper semicontinuous partial function, 5
Vector eld, 54
Vietoris topology, 380
Vitali equivalence relation, 84
Vitali partition, 85
Vitali set, 83
Vitali type function, 84
von Neumann topology, 406
Weak product of a family of groups, 44
weakly continuous function, 224
weakly measurable set-valued mapping,
371
weakly metrically transitive measure,
145, 148
Whitehead problem, 419
Wiener measure, 13
Wiener stochastic process, 13
Zermelo-Fraenkel set theory, 4
Zermelos theorem, 4
Zorn lemma, 4

(Atlantis Studies in Mathematics 2) Alexander B. Kharazishvili-Topics in Measure Theory and Real Analysis-Atlantis Press - World Scientific (2009)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Atlantis Studies in Mathematics 2) Alexander B. Kharazishvili-Topics in Measure Theory and Real Analysis-Atlantis Press - World Scientific (2009)

Uploaded by

Copyright:

Available Formats

ATLANTIS STUDIES IN MATHEMATICS

-measurable (see [1] and [13]). This theorem is then repeatedly

of such that f becomes measurable with

. It is established that there are universally measurable additive functionals

: R R of a partial function f : R R such that f

should be differentiable or continuous or semicontin-

) = E must preserve a given structure on E or should be compatible, in an

) = R if and only if the restrictions of this f to all sets

: X Y extending f and dened on some G

whose domain coincides with

dened on some maximal

is the required increasing extension of f dened on the

) = R. However, the proof of this fact is far

can be established by using the

dened on R and measurable in the Lebesgue sense;

dened on R and having the Baire property;

) = R. Of course, the construction of f

is not effective because it is based on

is a solution of the Cauchy functional equation, that is we have

is discontinuous at all points of R or, equivalently, f

: G H extending (see [70], [83], [152],

: R[0, 1] R satisfying the Carath eodory conditions and

: R R of f dened on the whole R;

. Show that the set X = x R : f

(x) = 0 described in Example 9 is not Lebesgue

: E R of f also satisfying the Lipschitz condition

. Show that every partial isometry f : R

. Let (E, S, ) be a -nite measure space and let

usually stands for the inner measure associated with . Denote

(Y) = 0, then we can extend a given measure to a

(X) = 0 holds true. Consequently, according to the

. Let E be a topological space. We recall that a set X E is totally imperfect in E if X

(X) = 0. Equivalently, X R is universal measure zero if

. Let E be a set, G be a group of transformations of E and let be a -nite G-quasi-

has already been dened; if Y

is such that there exists

) > 0, then take

. Let E be a base set and let G

is a subgroup of G. The group G

is usually called the

denote the family of all nite

pairwise commute, are afne and continuous, and map K

. We are going to show that u(x

into itself for which we have

. Let be a cardinal number and let S

. We recall the notation

. Let Gdenote the group of all isometric transformations of R

. Describe all those (sufciently smooth) homeomorphisms g : R

) card() + for every ordinal < ;

= / 0 for any two distinct ordinals < and < ;

: < , rst observe that the conditions (i) and (ii)

: < of subgroups of G satisfying the relations:

) card() + for any < ;

whenever < < .

: < yields the desired result.

: < is as in Exercise 19. Preserving the notation of

: <} of left H-translates of K has already been constructed.

= hK. Then it is clear that the partial family {K

stand for the weak product group of the family of groups {G

denote the completion of . Suppose also that G is some separable topological

. Let G be a subgroup of the additive group R

, quasi-invariant with respect to the everywhere dense

. Also, check that if is strictly greater than the cardinality