VOLUME 2
SERIES EDITOR: J. VAN MILL
Atlantis Studies in Mathematics
Series Editor:
J. van Mill, VU University Amsterdam, Amsterdam, the Netherlands
(ISSN: 18757634)
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 copublished with World Scientic.
For more information on this series and our other book series, please visit our website at:
www.atlantispress.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.atlantispress.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: 9789078677208
ISSN: 18757634
c 2009 ATLANTIS PRESS / WORLD SCIENTIFIC
eISBN: 9789491216367
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 settheoretical aspects of the top
ics discussed in the book; consequently, settheorists may detect nontrivial items of interest
to them and nd out new applications of settheoretical 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, NorthHolland 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 realvalued 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
realvalued partial functions. Namely, we discuss therein several variants of the socalled
measure extension problemand we pay our attention to purely settheoretical, 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
is
dened on the whole R and has certain nice properties?
In particular, we may require that f
: E R
with dom( 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
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
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)
6 Topics in Measure Theory and Real Analysis
(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
: 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 TietzeUrysohn 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
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 preimage 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 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.
The problem of extending partial functions 9
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 nskiZygmund function has the Baire property.
Statement (b) directly implies that no Sierpi nskiZygmund function is measurable in the
Lebesgue sense. Moreover, we can assert that a Sierpi nskiZygmund 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 nskiZygmund 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 nskiZygmund functions have other interesting and important properties. Many
works were devoted to functions of Sierpi nskiZygmund type (see [7], [42], [123], [124],
[148], [185], and [199]). In the sequel, we will be dealing with Sierpi nskiZygmund
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 innitedimensional vector space R) and f is a partial
linear functional, it admits a linear extension
f
: R Q
with dom( 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 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
10 Topics in Measure Theory and Real Analysis
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 nitedimensional 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
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 settheoretical
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
12 Topics in Measure Theory and Real Analysis
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 nskiZygmund 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 nskiZygmund 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 realvalued 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:
The problem of extending partial functions 13
(a) F admits an extension 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 RyllNardzewski
(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 realvalued 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 realvalued
partial setfunctions.
We recall that a setfunction is any function whose domain is some family of sets. Equiva
lently, we may say that a setfunction 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
14 Topics in Measure Theory and Real Analysis
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 realvalued 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 settheoretical 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 ZermeloFraenkel 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
The problem of extending partial functions 15
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 realvalued 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
. 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
:
R
n
R
n
. Find necessary and sufcient conditions under which f
is uniquely determined
by f .
On the other hand, for any innitedimensional 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 xedpoint 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 xedpoint of .
This result is due to Tarski and is known as Tarskis xedpoint 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
(Y) = 0),
where
/
(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 abovementioned 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
(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)
Some aspects of the measure extension problem 23
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 socalled 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 settheoretical 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.
24 Topics in Measure Theory and Real Analysis
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, Ginvariant if
(g G)(X L)(g(X) L).
We shall say that a measure on E is invariant under G or, in short, Ginvariant if dom()
is a Ginvariant class of subsets of E and
(g G)(X dom())((g(X)) = (X)).
More generally, we shall say that a measure on E is quasiinvariant under G or, in short,
Gquasiinvariant if dom() is a Ginvariant 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 quasiinvariant)
measures on E and, analogously, with right invariant (right quasiinvariant) 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 quasiinvariant) 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 Ginvariant (or, more generally, nonzero nite left Gquasi
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 settheoretical assumptions (see
proof sketched in Exercises 9, 10 and 11 of Appendix 1). In view of the abovementioned
statement, it is natural to ask whether there exists a left Ginvariant (left Gquasiinvariant)
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
Some aspects of the measure extension problem 25
settheoretical 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 Ginvariant (left
Gquasiinvariant) 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 quasiinvariant 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 Ginvariant (Gquasiinvariant) measure on some Ginvariant algebra of subsets of
E. Suppose that a Ginvariant 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 Ginvariant 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 abovementioned representation of Z, and
/
turns
out to be a Ginvariant (Gquasiinvariant) 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 Ginvariant with respect to ). Then we can ex
tend a given nite Ginvariant (Gquasiinvariant) measure to a Ginvariant (Gquasi
invariant) measure
/
in such a way that
/
(Y) = 0. Indeed, it sufces to consider the
26 Topics in Measure Theory and Real Analysis
Ginvariant ideal I of subsets of E, generated by the oneelement family Y, and to
apply to I the abovementioned 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 socalled 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 <
Some aspects of the measure extension problem 27
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 abovementioned par
tially ordered set. Denote such a maximal element by . Using Theorem 1 and Theorem 3,
28 Topics in Measure Theory and Real Analysis
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 settheoretical 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 algebraictopological 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 measuretheoretical sense. This
method was originally introduced by Kodaira and Kakutani in their famous construction of
a nonseparable translationinvariant extension of (see [141] and Chapter 16).
Some aspects of the measure extension problem 29
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 abovementioned lemmas.
For this purpose, let us consider the concrete particular case, where
E
1
= E
2
= R,
30 Topics in Measure Theory and Real Analysis
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
Gquasiinvariant 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 Ginvariant 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 Ginvariant (or, more generally, Gquasiinvariant) measures on E and
pose for them the measure extension problem. In other words, we are interested whether
any nonzero nite Ginvariant (respectively, Gquasiinvariant) 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
Some aspects of the measure extension problem 31
(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 viewpoint of the theory of
invariant (quasiinvariant) measures (see [108], [115], [119], and [198]). Let us give the
precise denition of such sets.
Let M be a class of nite Ginvariant (Gquasiinvariant) 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 Ginvariant (Gquasiinvariant)
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 Gabsolutely negligible subset of E.
Fortunately, it is possible to give a characterization of all Gabsolutely 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 Gabsolutely 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.
32 Topics in Measure Theory and Real Analysis
Let us denote
Z
//
= E Z
/
and consider the Ginvariant ideal I of subsets of E, generated by the oneelement
family Z
//
. Obviously, we can dene a complete probability Ginvariant 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 Gquasiinvariant 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 Gquasiinvariant
measure on E. We shall prove that there exists a Gquasiinvariant 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 Ginvariant ideal of subsets of E, generated by
the oneelement family Z. Taking 2) into account and applying Lemma 4, we easily infer
that for each set X J, the equality
. 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
Some aspects of the measure extension problem 33
/
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.
34 Topics in Measure Theory and Real Analysis
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.
5. Give a proof of Lemma 2.
6. Give a proof of Lemma 3.
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
. 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 twodimensional
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 realvalued 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).
Some aspects of the measure extension problem 35
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 StoneWeierstrass theorem on approximation of continuous
realvalued 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
: <
1
of countable
congurations of X. Namely, put:
(a) Y
0
= X;
(b) let <
1
and suppose that Y
)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 Gabsolutely negligible subsets of E is a proper Ginvariant 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 Gquasiinvariant measure on E such that every Gquasi
invariant extension
/
of is not maximal, i.e., we can nd a proper Gquasiinvariant
extension of
/
.
Some aspects of the measure extension problem 37
15
. 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 twovalued 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 Cproperty holds
true for any realvalued 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 socalled 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, Ginvariant) if
(X S)(g G)(g(X) S).
Suppose, in addition, that is a measure dened on a Ginvariant algebra S of subsets
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific
39
3
40 Topics in Measure Theory and Real Analysis
of E. We usually say that is invariant with respect to G (or, briey, Ginvariant) 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 translationinvariant 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 abovementioned 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
42 Topics in Measure Theory and Real Analysis
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 ndimensional
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 onedimensional 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 oneelement set g is everywhere dense
Invariant measures 43
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 realvalued 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
.
44 Topics in Measure Theory and Real Analysis
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 quasiinvariant
measures on groups).
In connection with the just said, it is also reasonable to remark that some Polish nonlocally
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 Rquasiinvariant 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
.
Invariant measures 45
Further, we put
G
=g G : i I : g
i
,= e
i
is f inite.
It can easily be checked that G
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 Ginvariant 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 Ginvariant measure on the algebra
S and the set
Y =Y
n
: n N
Invariant measures 47
is a Ginvariant 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 abovementioned 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)
48 Topics in Measure Theory and Real Analysis
of its transformations, there exists a nonzero nite Ginvariant 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 realvalued sequences.
Now, we wish to return to a nitedimensional Euclidean space and discuss one important
geometric property of the classical Lebesgue measure given on this space. We mean here
the socalled Steinhaus property which was rst established by Steinhaus in his remarkable
paper [233].
Let
n
denote, as usual, the standard Lebesgue measure on the ndimensional 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).
Invariant measures 49
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 BanachKuratowskiPettis 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 realvalued continuous functions on T. By virtue of the classical
Riesz theorem (see, for instance, [56] and [57]), there is a canonical onetoone 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 [[].
50 Topics in Measure Theory and Real Analysis
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.
Invariant measures 51
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
.
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 abovementioned family is centered. Therefore,
v(K) : v G
,= / 0.
52 Topics in Measure Theory and Real Analysis
Choose a point x
0
v(K) : v G
(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))
. 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
.
56 Topics in Measure Theory and Real Analysis
Prove that the following two assertions are equivalent:
(a) the Lebesgue measure
n
is invariant under the oneparameter 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 realvalued 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 oneparameter 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 ndimensional 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 oneparameter group).
4
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:
Invariant measures 57
(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.
Prove that the following two assertions are equivalent:
(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 quasiinvariant measure dened on an algebra of
sets, such that the unique extension
/
of dened on the generated algebra does not
have the quasiinvariance property with respect to the same transformation group.
58 Topics in Measure Theory and Real Analysis
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
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
. 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.
60 Topics in Measure Theory and Real Analysis
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
Ginvariant 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 Ginvariant 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 abovementioned 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 Ginvariant
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.
Invariant measures 61
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
:
is an almost Ginvariant subset of E, i.e., we have
(g G)(card(g(X())X()) < card(E)).
In order to show the existence of X
: < = G;
(3) G
= G
(e) G
(e) : < .
Verify that the family of sets X
. 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 Ginvariant (Gquasiinvariant) measure on E. Assuming that
card(E) is not measurable in the Ulam sense (see Appendix 1), prove that there exists a
Ginvariant (Gquasiinvariant) measure
/
on E strictly extending .
For this purpose, consider two possible cases.
62 Topics in Measure Theory and Real Analysis
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 Ginvariant
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 nonmeasurable with respect to . Then apply to X() Marczewskis method of extending
nite Ginvariant (Gquasiinvariant) measures.
In particular, let (G, ) be an arbitrary uncountable group equipped with a nonzero nite
left invariant (left quasiinvariant) measure . If card(G) is not measurable in the Ulam
sense, then can be strictly extended to a left invariant (left quasiinvariant) 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 quasiinvariant) measure on an uncountable group does not contradict ZFC
theory.
At present, it is unknown whether the same assertion can be established without additional
settheoretical 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
Quasiinvariant 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, Ginvariant) if
(X S)(g G)(g(X) S).
Suppose that is a measure dened on a Ginvariant algebra S of subsets of E. We say
that is quasiinvariant with respect to G (or quasiinvariant under G or, briey, Gquasi
invariant) if
(X S)(g G)((X) = 0 (g(X)) = 0).
It immediately follows from this denition that every Ginvariant measure on E is neces
sarily Gquasiinvariant. Obviously, the converse assertion is not true in general.
Similarly to invariant measures, quasiinvariant 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, quasiinvariant 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) quasiinvariant measures. We will recall several important examples of 
nite quasiinvariant measures. In addition, we will present some general constructions of
nite quasiinvariant measures and demonstrate how those constructions lead to other
interesting examples of measures of this type.
If we are given a nonzero nite Gquasiinvariant measure on E, then it is clear that
any nite measure which is equivalent to turns out to be Gquasiinvariant. Indeed,
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific
63
4
64 Topics in Measure Theory and Real Analysis
from the classical RadonNikodymtheorem (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 RadonNikodymderivative of with respect to . In
other words, f satises the equality
f = d/d.
This representation immediately yields the Gquasiinvariance 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 Gquasi
invariant measure on E. Does there exist a nonzero nite Ginvariant measure on E
which is equivalent to ?
Another important problemarising in connection with nonzero nite quasiinvariant (in
variant) measures is the problem of their existence for a given space E equipped with a
transformation group G and with a Ginvariant 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 leftinvariant 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, quasiinvariant) 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 quasiinvariant 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 nonlocally 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:
Quasiinvariant 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) Htranslates 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 Htranslates 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
: <}.
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
: } is disjoint and
consists of some left Htranslates of K. Proceeding in such a manner, we are able to dene
the required disjoint
1
family {K
: <
1
} of left Htranslates 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) Hquasiinvariant Borel measure on G.
66 Topics in Measure Theory and Real Analysis
In particular, if (G, ) is a nonlocally compact Polish topological group, then there exists
no nonzero nite left (right) Gquasiinvariant Borel measure on G.
Proof. Suppose to the contrary that there exists a nonzero nite left (or, respectively,
right) Hquasiinvariant 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) Htranslates of K. Since is left (right) Hquasiinvariant, all sets
from the abovementioned 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 quasiinvariant 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 quasiinvariant 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 imagemeasure
() =
1
Quasiinvariant measures 67
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 quasiinvariant 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 Rquasiinvariant 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 quasiinvariant measure starting froman initial family of spaces
equipped with corresponding probability quasiinvariant measures.
The next simple statement concerns the behavior of the quasiinvariance property under the
operation of taking arbitrary products of probability quasiinvariant 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
quasiinvariant 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
quasiinvariant.
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.
68 Topics in Measure Theory and Real Analysis
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 Rquasiinvariant. 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 quasiinvariant 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
, quasiinvariant
under R
(I)
.
Thus, we have a large class of Borel probability measures on the innitedimensional Pol
ish topological vector space R
N
, which are quasiinvariant 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
quasiinvariant 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 quasiinvariant 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
).
Quasiinvariant measures 69
The space l
1
contains R
(N)
and is separable. Moreover, it is the initial space in the class of
all innitedimensional 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.
70 Topics in Measure Theory and Real Analysis
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 quasiinvariant with respect to some everywhere dense
vector subspace of E.
Proof. We start with a Borel probability measure on l
1
which is quasiinvariant 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 quasiinvariant
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.
Quasiinvariant measures 71
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, quasiinvariant
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 quasiinvariant 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 abovementioned problem is
completely solved.
Indeed, Theorem 3 states that every separable Banach space E admits a Borel probability
measure quasiinvariant 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 realvalued measurable cardinal number (see Appendix 3), then G does not admit
a Borel probability measure quasiinvariant with respect to an everywhere dense subgroup
of G (see Exercise 2).
Thus, under the settheoretical assumption that no cardinal number is realvalued 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 quasiinvariant with respect to some ev
erywhere dense vector subspace of E.
Fortunately, it turns out that an additional settheoretical 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).
72 Topics in Measure Theory and Real Analysis
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
on the Euclidean space R
n
, the socalled 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, ...).
Quasiinvariant measures 73
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 RadonNikodym 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) =
74 Topics in Measure Theory and Real Analysis
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
Quasiinvariant measures 75
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
;
(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 quasiinvariant with respect to the group
of all left translations of H.
76 Topics in Measure Theory and Real Analysis
2. A wellknown theorem of topological measure theory states that any nite Borel
measure given on a metric space E with nonrealvalued 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 realvalued
measurable cardinal. Suppose also that there exists a nonzero nite Borel measure on G
quasiinvariant 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 realvalued 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 ndimensional 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 quasiinvariant 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.
Quasiinvariant measures 77
5
. 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 quasiinvariant with respect to G.
Show that the following two assertions are equivalent:
(a) E is innitedimensional;
(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 realvalued functions
In this chapter, two concepts associated with measurability properties of various realvalued
functions are introduced and examined. These concepts are the notion of a relatively mea
surable realvalued 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 abovementioned 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 realvalued measurable function with respect to a concrete measure on E. Indeed, it
we take the oneelement class M ={}, then the relative measurability of a given function
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ 2009 Atlantis Press/World Scientific
79
5,
80 Topics in Measure Theory and Real Analysis
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 Ginvariant (or, more generally, Gquasiinvariant)
measures on E. If a set X E is absolutely nonmeasurable with respect to M, then we
briey say that X is Gabsolutely 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 realvalued 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 nondiscrete 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
82 Topics in Measure Theory and Real Analysis
is a diffused measure and f remains measurable with respect to
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 abovementioned result, there exists a measure
. 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 righthand 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
() the class of all those measures on R which extend and are quasi
invariant under the group of all translations of R. Recall that quasiinvariant 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 translationinvariant 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
()?
In this context, it makes sense to consider in more details Vitalis classical construction
[239] of a nonLebesgue 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 translationinvariant
extensions of . For translationquasiinvariant 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 wellknown theoremfromlinear algebra which states
86 Topics in Measure Theory and Real Analysis
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 abovementioned 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 translationinvariant 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 (quasiinvariant)
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 translationinvariant extensions of . On the other hand, we can demonstrate that H
becomes measurable with respect to a suitable translationquasiinvariant extension of .
The following statement contains a much stronger result which shows that there are many
possibilities for obtaining such translationquasiinvariant extensions of .
Measurability properties of realvalued functions 87
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
translationinvariant 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 translationinvariant 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,
88 Topics in Measure Theory and Real Analysis
where is again some permutation of (of course, depending on h). Therefore, the
measure is quasiinvariant under the group of all translations of R. Moreover, a similar
argument yields that is also quasiinvariant under all central symmetries of R and hence
is quasiinvariant 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 translationquasiinvariant 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
.
Measurability properties of realvalued functions 89
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 realvalued 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
90 Topics in Measure Theory and Real Analysis
[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
. 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 translationinvariant extensions of .
4. Let G be an uncountable subgroup of the additive group R. Prove that every Gselector
is a Gabsolutely negligible subset of R (see Chapter 2).
In particular, conclude from this fact that every Gselector is relatively measurable with
respect to the class of all Ginvariant extensions of the Lebesgue measure .
Measurability properties of realvalued functions 91
5
. 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
. 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 realvalued 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
nondiscrete 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
Measurability properties of realvalued functions 93
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 aBaire 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;
94 Topics in Measure Theory and Real Analysis
(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
preimages 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 halfopen 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
Measurability properties of realvalued functions 95
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
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
).
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
.
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 stepfunction. 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
( 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 settheoretical assumptions, Theorem 1 can be essentially strength
ened. For instance, let us consider the following settheoretical assertion.
100 Topics in Measure Theory and Real Analysis
(*) 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 stepfunction;
(2) for any nite measure on E, there exists a measure
.
Proof. The argument is very similar to the proof of Theorem 1. The implication (1) (2)
does not need an additional settheoretical 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 stepfunction. 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
stepfunction f : E R having the measurability property (2).
Now, let us consider stepfunctions for those nonzero nite measures on E which are
invariant (or, more generally, quasiinvariant) with respect to an uncountable group G of
transformations of E.
Some properties of stepfunctions connected with extensions of measures 101
If we suppose that G acts freely in E, then the situation differs from the case of ordinary
(i.e., noninvariant) 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 stepfunction associated with this partition. It is easy to verify that f
cannot be measurable with respect to a nonzero nite Gquasiinvariant 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 Gabsolutely negligible if for every nite Ginvariant (G
quasiinvariant) measure on E, there exists a Ginvariant (Gquasiinvariant) extension
of such that
(X) = 0.
We say that a set Y E is Gabsolutely nonmeasurable if for any nonzero nite Gquasi
invariant measure on E, we have Y dom().
Some properties of Gabsolutely negligible and Gabsolutely 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 Gabsolutely nonmeasurable subset of G;
(2) there exists a countable partition of G into Gabsolutely 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.
102 Topics in Measure Theory and Real Analysis
Example 1. Let X be a Gabsolutely 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 Gquasiinvariant
measure on G.
In other words, Example 1 states that there are twovalued functions absolutely nonmea
surable with respect to the class of all nonzero nite Gquasiinvariant measures on G.
Example 2. For an uncountable commutative group (G, +), consider its countable partition
{Y
i
: i I} into Gabsolutely negligible sets. Let
f : G R
denote an arbitrary stepfunction associated with this partition. Then we have:
(1) f is nonmeasurable with respect to any nonzero nite Gquasiinvariant measure on
G;
(2) for each t R, the set f
1
(t) is Gabsolutely negligible.
Indeed, (2) is obvious. To see (1), let us suppose that f is measurable with respect to some
nonzero nite Gquasiinvariant measure . Then all sets Y
i
which are the preimages
of certain singletons in R must be measurable. Since all of them are also Gabsolutely
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 stepfunctions f on an uncountable com
mutative group (G, +), which are absolutely nonmeasurable with respect to the class
of all nonzero nite Gquasiinvariant measures on G, but each of the preimages
f
1
(t) (t R) is good for extending any nite Ginvariant (Gquasiinvariant) measure
on G.
However, there are certain types of stepfunctions which also are good for obtaining in
variant (respectively, quasiinvariant) 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 Ginvariant 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
Some properties of stepfunctions connected with extensions of measures 103
recognized that almost Ginvariant sets play an important role in various topics of the the
ory of invariant and quasiinvariant 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 Ginvariant in E (in the settheoretical 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 Ginvariant with respect to if
(g(Z)Z) = 0
for each transformation g G (where
: <
} of subgroups of G can readily be constructed satisfying the following conditions:
(1) {G
: <} = G;
(2) card(G
(x) \ {G
(x) \ {G
(x) : <}) :
i
} (i I),
104 Topics in Measure Theory and Real Analysis
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 Gquasiinvariant measure on E and let f : E R be a
stepfunction such that for any t R, the set f
1
(t) is almost Ginvariant with respect to
. Then for any g G, the functions f and f g are equivalent with respect to , that is
the equality
(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 Ginvariant sets. Denote by f : E R any
stepfunction associated with this partition. Then for every nite Ginvariant (Gquasi
invariant) measure on E, there exists a Ginvariant (Gquasiinvariant) measure
on E
such that:
(1)
extends ;
(2) f is measurable with respect to
.
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 (quasiinvariant) 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
.
Some properties of stepfunctions connected with extensions of measures 105
Let be an arbitrary nite Ginvariant (Gquasiinvariant) 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 (quasiinvariant) measures
works for the Ginvariant ideal consisting of the abovementioned sets Z.
Therefore, all sets X
i
(i I) become almost Ginvariant 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 Ginvariant with
respect to . Moreover, we have the equality
E = {Z
i
: i I}.
Let f
({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
({g
k
(D) : k <}) = 0.
Thus, the set D generates a Ginvariant ideal of subsets of E, all members of which have
inner measure zero. This circumstance enables us to extend the given measure to a
Ginvariant (Gquasiinvariant) measure
on E such that
is measurable, it is also
(x)},
106 Topics in Measure Theory and Real Analysis
we conclude that f is
measurable, too.
Remark 1. Actually, the preceding argument shows that if we have an arbitrary nite
Ginvariant (Gquasiinvariant) measure on E and a stepfunction f : E R such that
all preimages f
1
(t) (t R) are almost Ginvariant with respect to , then there exists a
Ginvariant (Gquasiinvariant) extension
measurable.
In connection with Theorems 1 and 2, the following question arises.
Does there exist a realvalued function f
1
: E R with card(ran( f
1
)) > such that every
nite Ginvariant (Gquasiinvariant) measure on E admits a Ginvariant (Gquasi
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 Gabsolutely negligible subset X of E.
Proof. If card(G) >
1
, then any set X E with card(X) =
1
is Gabsolutely 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
(x) \ {G
(x) \ {G
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 (quasiinvariant) 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 stepfunctions f on E which have the generalized measurability property in
the sense that for any nite Ginvariant (Gquasiinvariant) measure on E, there exists
a Ginvariant (Gquasiinvariant) extension
measurable?
This problem remains open.
Example 3. Let f be a stepfunction 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 Ginvariant (Gquasiinvariant) measure on E, there exists
a Ginvariant (Gquasiinvariant) 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 Ginvariant normalized measure type function
als on E, then it is reasonable to call a stepfunction 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
Ginvariant normalized measure type functional on E, then, according to von Neumanns
extension theorem (see, for instance, [83], [108], and [240]), every stepfunction on E
becomes measurable with respect to an appropriate universal nitely additive Ginvariant
extension
) coincides with
108 Topics in Measure Theory and Real Analysis
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 stepfunctions on E which are absolutely
nonmeasurable with respect to the class of all nitely additive left Ginvariant 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 abovementioned assertion (*).
3. Let E be a set, G be a group of transformations of E and let be a nite Gquasi
invariant measure on E. Suppose that X E is an almost Ginvariant set with respect to .
Verify that a measurable hull of X is also almost Ginvariant 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 Rinvariant subsets (with respect to ), none of which is of measure zero.
5
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.
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific
111
7
112 Topics in Measure Theory and Real Analysis
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 realvalued 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
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
114 Topics in Measure Theory and Real Analysis
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
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 gX
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 abovementioned construction.
Almost measurable realvalued functions 115
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
={Z
: Z S dom()},
(Z
) = ( )(Z) (Z
).
Taking into account the ( )thickness of Gr(g), we can verify that:
(i) 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 preimage 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;
116 Topics in Measure Theory and Real Analysis
(2) f is measurable with respect to
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
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.
Almost measurable realvalued functions 117
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
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
;
(2) f is measurable with respect to
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
following relations:
(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[
118 Topics in Measure Theory and Real Analysis
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 abovementioned
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 +gX 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).
Almost measurable realvalued functions 119
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 +gX
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),
120 Topics in Measure Theory and Real Analysis
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 quasiinvariant 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 ,
Almost measurable realvalued functions 121
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 halfline [0, +[ and let B
2
denote a Bernstein
subset of the halfline ] , 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
={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 )},
122 Topics in Measure Theory and Real Analysis
thus it follows that X S
(Z
) = ( )(Z).
As a direct verication shows, the functional
is a measure on S
.
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 quasiinvariance (under the group of all translations
of R) of the obtained in this way extension. For the abovementioned quasiinvariance,
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 socalled Dirichlet function) is not
almost continuous (cf. Exercise 4 below).
2
. 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 endpoints 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 translationquasiinvariant extension of the
Lebesgue measure on R.
124 Topics in Measure Theory and Real Analysis
In particular, let f and g be as in Theorem 5. Show that there exist two translationquasi
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),
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific 8
129
126 Topics in Measure Theory and Real Analysis
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 wellknown 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 preimage 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.
128 Topics in Measure Theory and Real Analysis
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 preimage 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).
Several facts from general topology 129
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 settheoretical 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 <)
130 Topics in Measure Theory and Real Analysis
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 Requivalence 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 Requivalence 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
}).
Several facts from general topology 131
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 Requivalence 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 abovementioned 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 <}.
132 Topics in Measure Theory and Real Analysis
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 Requivalence class
is a singleton.
For this purpose, take an arbitrary Requivalence 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 setvalued
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 socalled Aoperation from classical descriptive set theory.
It makes sense to touch upon this fundamental settheoretical 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 Aoperation over the abovementioned 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 Aoperation over all analogous countable systems
of sets belonging to L. The members of A(L) are usually called analytic sets over L.
Several facts from general topology 133
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 coanalytic (coSuslin) sets. An
alytic and coanalytic 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 Aoperation 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 coanalytic 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
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 wellknown 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.
134 Topics in Measure Theory and Real Analysis
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 realvalued 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) nh)}.
Several facts from general topology 135
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 (comeager) 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)
136 Topics in Measure Theory and Real Analysis
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 RyllNardzewski (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 setvalued mapping
: E/R P(E)
Several facts from general topology 137
dened by the formula
(z) =
1
(z) (z E/R).
Check that satises the assumptions of the theorem of Kuratowski and RyllNardzewski
(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 Aoperation 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,
.
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}
138 Topics in Measure Theory and Real Analysis
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 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
(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
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
.
Several facts from general topology 139
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 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
: R R extending g.
21
. 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},
satisfying the following conditions:
(a) for each natural number n, the set Z
n
can be represented in the form
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;
142 Topics in Measure Theory and Real Analysis
(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 gX 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
: (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 abovementioned facts and Lavrentievs theorem on extensions of
homeomorphisms, prove that for a metrizable space E, the following two assertions are
equivalent:
Several facts from general topology 143
(c) E is topologically complete;
(d) E is homeomorphic to a 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
< 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 translationinvariant 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 grouptheoretical
and measuretheoretical. 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
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific
145
9
146 Topics in Measure Theory and Real Analysis
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 abovementioned 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 Ginvariant algebra of subsets of E and invariant under all trans
formations fromG. In this case, we briey say that is a Ginvariant measure (see Chapter
2).
Let X be some subset of E.
We shall say that X is Gthick 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 Gthin 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 Gthin 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 Gthin subset of E. On the other hand, it also may happen that Gr( f ) is R
2
thick
with respect to some translationinvariant extension of the twodimensional 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 Ginvariant 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 Gthick 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 Ginvariant
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
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, Gquasiinvariant) 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 nonatomic Ginvariant 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 Ginvariant extension of a
Ginvariant weakly metrically transitive measure is also weakly metrically transitive. An
analogous statement fails to be true for Ginvariant 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.
Weakly metrically transitive measures and nonmeasurable sets 149
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 Rinvariant, 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 Rinvariant 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 Rinvariant 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 Ginvariant 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 ;
150 Topics in Measure Theory and Real Analysis
(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 Ginvariant 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
Ginvariant (Gquasiinvariant) 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 Ginvariant (Gquasiinvariant) measure on E. We shall say that a set
X E is Gabsolutely nonmeasurable with respect to if X is absolutely nonmeasurable
with respect to the class of all Ginvariant (Gquasiinvariant) measures on E extending .
The next example highlights close relationships between convex sets and absolutely non
measurable sets.
Example 6. Let E be an innitedimensional 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 Ginvariant (more generally,
Gquasiinvariant) 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 Gquasiinvariant 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 Rinvariant (more generally,
Rquasiinvariant) measures on R.
Example 7. As in Example 4, let us put:
E = the real line R,
G = the group of all translations of R,
= the Lebesgue measure on R (i.e., =).
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
Weakly metrically transitive measures and nonmeasurable sets 151
known that all Vitali sets are Rabsolutely 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 Ginvariant weakly metrically transitive measure on E. Suppose also that G acts
almost freely in E with respect to . Then every Gselector is Gabsolutely nonmeasurable
with respect to .
Proof. Take an arbitrary Gselector X in E. We should verify that for any Ginvariant
measure
/
on E extending the original measure , the set X is not
/
measurable. Suppose
otherwise, i.e., X dom(
/
) for some Ginvariant 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 Gthick subset of E with respect to . Since
/
extends , the same Y is Gthick with respect to
/
as well. On the other hand, the selector
X is a Gthin 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:
E = the real line R,
G = the group Q of all rational numbers,
= the Lebesgue measure on R (i.e., =).
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 abovementioned
classical result. Some other abstract versions of Vitalis theorem are presented in [115],
[119], [195], [229], and [230] (cf. also [251]).
152 Topics in Measure Theory and Real Analysis
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).
Weakly metrically transitive measures and nonmeasurable sets 153
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 Horbits.
The following three assertions are equivalent:
(a) is weakly metrically transitive with respect to H;
(b) every Hselector in G is absolutely nonmeasurable with respect to ;
(c) H is nondiscrete 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 Hselector 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
154 Topics in Measure Theory and Real Analysis
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
nondiscrete subgroup of G. Such subgroups always exist in G and a proof of this simple
auxiliary assertion is left to the reader. Pick any Hselector X in G. Suppose to the contrary
that there exists a left Hinvariant 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 nondiscrete,
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.
Weakly metrically transitive measures and nonmeasurable sets 155
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 ndimensional 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 Hinvariant measure on R
n
are in
vestigated in paper [142]. In the same paper various Hinvariant extensions of
n
are also
studied.
For our purpose, we need several wellknown 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.
156 Topics in Measure Theory and Real Analysis
(iv) Let G be an uncountable group of motions of R
n
and let X be a Gselector. 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 Gquasi
invariant measure on E, then any Gselector 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 Hselector in R
n
is absolutely nonmeasurable with respect to
n
;
(c) H is nondiscrete 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
(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
quasiinvariant measure on E
1
and let
2
be a
nite G
2
quasiinvariant 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
.
4. Complete the details of Example 4.
5. Complete the details of Example 5.
6
on G strictly extend
ing ?
In the classical case, when G coincides with the ndimensional Euclidean space R
n
(n 1)
and is a translationinvariant 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 abovementioned 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 measuretheoretical and grouptheoretical
points of view and plays a signicant role in questions concerning extensions of invariant
(quasiinvariant) measures (see [105], [108], [115], and [198]).
Let X be a subset of (G, ). We recall (cf. Chapter 2) that X is Gabsolutely negligible
(in G) if for any nite left Ginvariant (respectively, left Gquasiinvariant) measure
on G, there exists a left Ginvariant (respectively, left Gquasiinvariant) 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 Gabsolutely negligible set in G and X belongs to the domain of a given nite
left Ginvariant (left Gquasiinvariant) measure , then (X) = 0;
2) if X is a Gabsolutely negligible set in G and X does not belong to the domain of a 
nite left Ginvariant (left Gquasiinvariant) measure on G, then is strictly extendable,
by using this X, to some left Ginvariant (left Gquasiinvariant) 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 Gabsolutely 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 Gabsolutely 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 Gabsolutely
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
assertions are equivalent:
1) X is Gabsolutely 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 Gabsolutely negli
gible sets in (G, ). In particular, it immediately follows from this lemma that if X is a
162 Topics in Measure Theory and Real Analysis
Gabsolutely negligible subset of G and is an arbitrary nite left Gquasiinvariant
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
={h
i
Y : i I}.
Since Y is Habsolutely 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 Gabsolutely negligible in G which ends the proof of
Lemma 2.
Remark 1. There are several notions in the theory of invariant and quasiinvariant mea
sures, which are connected with the concept of a Gabsolutely negligible set. Among them
let us especially mention the notion of a Gnegligible set and the notion of a Gabsolutely
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 Hnegligible, then
1
(Y) is Gnegligible;
Nonmeasurable subgroups of uncountable solvable groups 163
b) if Y is Habsolutely nonmeasurable, then
1
(Y) is Gabsolutely nonmeasurable.
These facts are useful in some constructions of Gnegligible and Gabsolutely 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 Habsolutely negligible in H. Then the set
X ={g
k
Y : k K}
turns out to be Gabsolutely 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)
164 Topics in Measure Theory and Real Analysis
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 abovementioned 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 Gabsolute negligibility of X. Lemma 3 has thus been
proved.
Remark 2. All Gabsolutely 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 nonsmall set. More precisely, it can be demonstrated that for every uncountable solvable
group (G, ), there exist two Gabsolutely 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
Gquasiinvariant measure on G, then there exists a subset of G nonmeasurable with respect
Nonmeasurable subgroups of uncountable solvable groups 165
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 realvalued
measurable (see [238]).
For an uncountable commutative group (G, +) equipped with a nonzero nite Gquasi
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 noncommutative group
(H, ) endowed with a nonzero nite left Hquasiinvariant measure. In this respect,
commutative groups have some advantage because for any uncountable commutative group
(G, +) equipped with a nonzero nite Gquasiinvariant 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 Ginvariant
(Gquasiinvariant) extensions of initial nonzero nite Ginvariant (Gquasiinvariant)
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 abovementioned 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 <},
166 Topics in Measure Theory and Real Analysis
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
on the Euclidean space R
n
. Lemma 4
enables us to give a solution of generalized Sierpi nskis problem concerning extensions of
nonzero nite translationinvariant (respectively, translationquasiinvariant) 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 Ginvariant
(Gquasiinvariant) measure on G and let G
0
be a subgroup of G such that card(G/G
0
) >
. Then there exists a Ginvariant (Gquasiinvariant) 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
}.
Nonmeasurable subgroups of uncountable solvable groups 167
Observe that I is a Ginvariant 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
(YX) = (Y)
for any set YX S
(G
0
) = 0 which ends the proof of the lemma.
Remark 3. In fact, the preceding argument establishes that G
0
is a Gabsolutely 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 Ginvariant (Gquasiinvariant) 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 Ginvariant (Gquasi
invariant) measure
(G
j
) = 0.
168 Topics in Measure Theory and Real Analysis
Proof. Let {G
j
: j J} be as in Lemma 4 and let be an arbitrary nonzero nite
Ginvariant (Gquasiinvariant) 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 Gquasiinvariance 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 Ginvariant (Gquasiinvariant) measure
(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 (quasiinvariant) 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 Gabsolutely 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 Ginvariant (left Gquasiinvariant) 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 Gquasiinvariant) 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:
Nonmeasurable subgroups of uncountable solvable groups 169
(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 Ginvariant (left G
quasiinvariant) 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 Ginvariant (left Gquasiinvariant) 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 Ginvariant (left G
quasiinvariant) measure
(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,
170 Topics in Measure Theory and Real Analysis
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
is always valid. Therefore, in this case the abovementioned conditions are satised auto
matically.
Finishing this chapter, let us recall that the rst two (absolutely different) constructions
of nonseparable translationinvariant 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, Ginvariant 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).
Nonmeasurable subgroups of uncountable solvable groups 173
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 nondiscrete 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).
5
. 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 nondiscrete 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
following relations:
(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
following relations:
(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 Ginvariant extension
;
(e) if some H
j
is nonmeasurable with respect to
, then
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.
Nonmeasurable subgroups of uncountable solvable groups 175
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 grouptheoretical 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 Ginvariant. 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 grouptheoretical operation.
Naturally, we will especially pay our attention to the case of a group (G, ) equipped with
a nonzero nite complete Ginvariant (Gquasiinvariant) 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 viewpoint.
In addition, we especially want to examine the situations where the grouptheoretical 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 Ginvariant or, more generally, Gquasiinvariant 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 wellknown
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 abovementioned fact and applying the techique of Hamel
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_ , 2009 Atlantis Press/World Scientific
177
11
178 Topics in Measure Theory and Real Analysis
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 ndimensional 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 quasiinvariant)
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 Ginvariant (Gquasiinvariant) 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 Ginvariant (Gquasiinvariant) 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 measuretheoretical point of view.
Problem. Let (G, ) be an uncountable group and let be a nonzero nite left G
invariant (respectively, left Gquasiinvariant) measure on G. Does there exist a left G
invariant (respectively, left Gquasiinvariant) 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 noncommutative
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 noncommutative 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 righthand
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
...;
180 Topics in Measure Theory and Real Analysis
(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 Gabsolutely negligible if for any nite left Ginvariant (respectively, left
Gquasiinvariant) measure on G, there exists a left Ginvariant (respectively, left G
quasiinvariant) 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 Gabsolutely 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
Habsolutely negligible subset of H. Then Y is a Gabsolutely 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 Gabsolutely negligible set in G (see, for instance, Lemma 5 and Remark
3 from Chapter 10). Since Y H, we easily infer that Y is Gabsolutely negligible in G,
too. Actually, here we do not need the assumption that Y is Habsolutely 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 Habsolutely negligible in H, it can be demonstrated
that Y
/
is Gabsolutely negligible in G (see Lemma 3 from Chapter 10). This circumstance
Algebraic sums of measure zero sets 181
immediately implies that each set g
j
+Y ( j J) is also Gabsolutely negligible in G. Con
sequently, the set
Y = (g
j
+Y) g
j
turns out to be Gabsolutely 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 Gabsolutely 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 Gabsolutely 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,
182 Topics in Measure Theory and Real Analysis
which yields a contradiction. Now, in view of Lemma 1, we conclude that the set X
n
is
Gabsolutely negligible in G.
Remark 1. At present, it is known that every uncountable commutative group (G, +) can
be covered by a countable family of Gabsolutely 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 Gabsolutely negligible
subsets.
For all uncountable noncommutative groups an analogous result is still unknown. In other
words, we do not know whether any uncountable group (G, ) admits a countable covering
with Gabsolutely negligible sets. This open question seems to be of interest from the
viewpoint 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 Gquasiinvariant
measure on G and let H be a subgroup of G such that
(Z H) = (Z H
/
),
where H
/
denotes a measurable hull of H. It can easily be veried that is an Hquasi
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 Habsolutely 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 Gquasiinvariant measure on G, then there always exists
a Gabsolutely negligible subset Y of G with
(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 Habsolutely negligible in H. In view of Lemma 2, the same X
1
is a Gabsolutely negligible subset of the original group G. Denote by
/
the minimal with
respect to the inclusion relation, complete, Ginvariant (Gquasiinvariant) 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 Habsolutely 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 Einvariant
(Equasiinvariant) measure given on E. In this case, some stronger results for are
obtained by starting with the wellknown 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].
Algebraic sums of measure zero sets 185
Let (G, +) be again an uncountable commutative group and let be a nonzero nite
complete Ginvariant or, more generally, Gquasiinvariant 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 Ginvariant (Gquasi
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 Gabsolutely 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 Ginvariant 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 Gabsolutely 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 Ginvariant 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 Horbits 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:
186 Topics in Measure Theory and Real Analysis
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 Gabsolutely 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 Habsolutely negligible
sets A and B in H such that A B = H.
Algebraic sums of measure zero sets 187
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 Habsolutely 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 wellknown 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.
Now, we are ready to prove the following statement.
Theorem 3. For any uncountable commutative group (G, +), there exist two Gabsolutely
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.
188 Topics in Measure Theory and Real Analysis
In view of Lemma 6, there are two Habsolutely 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 Gabsolutely 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 Gabsolutely 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 Gabsolutely
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 Gquasiinvariant 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 Gabsolutely 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 Gabsolutely
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 Ginvariant (respectively, Gquasiinvariant) measure on G, there exists
a Ginvariant (respectively, Gquasiinvariant) measure
/
on G extending and such that
/
(X) =
/
(Y) = 0, X +Y , dom(
/
)
Algebraic sums of measure zero sets 189
for some Gabsolutely 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 Gabsolutely
negligible sets X G and Y G and a Gabsolutely nonmeasurable set Z G satisfying
the relation
X Y = Z.
Then we can assert that there are also two Gabsolutely negligible sets A and B in G for
which
A B = G.
Indeed, since Z is Gabsolutely 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 Gabsolutely negligible sets A and B in G.
Theorem 4. Let (G, +) be an uncountable commutative group and let be a nite
Ginvariant (respectively, Gquasiinvariant) measure on G. There exists a Ginvariant
(respectively, Gquasiinvariant) extension
/
of such that
/
(A) =
/
(B) = 0, A+B = G
for some Gabsolutely 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 Ginvariant (respectively, Gquasiinvariant)
measure
1
on G for which
A dom(
1
),
1
(A) = 0.
Then we extend the measure
1
to a Ginvariant (respectively, Gquasiinvariant) 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
190 Topics in Measure Theory and Real Analysis
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 Gabsolutely 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
= card(G),
we may introduce an sequence H
: <, 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
: <;
C
=a
: < b
: <.
Obviously, we have the inequalities
card(G
a) (G
) = / 0, (G
b) (G
) = / 0, ab = g
.
We put a
=a and b
: < and b
: <, 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
Gabsolutely negligible in G. Finally, by virtue of the relation
g
= a
A B ( <),
Algebraic sums of measure zero sets 191
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 Ginvariant (respectively, left Gquasi
invariant) measure on G. Then there exist an almost left Ginvariant set A in G and a left
Ginvariant (respectively, left Gquasiinvariant) 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 Ginvariant 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
/
(A) =
/
(A
1
) = 0
are valid. Now, take any g G. Since A is almost left Ginvariant, 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.
192 Topics in Measure Theory and Real Analysis
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 measuretheoretical 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 viewpoint of the theory of invariant
(quasiinvariant) 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 innitedimensional vector spaces considered as commutative groups. Namely, the
following proposition is valid.
Lemma 9. Let E be an innitedimensional separable Hilbert space (over R) and let K be
an arbitrary open ball in E. Then K is an Eabsolutely 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 wellknown fact that E does not admit
a nonzero nite Borel measure quasiinvariant 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 Rabsolutely negligible subsets of R such that their algebraic
sum is an Rabsolutely nonmeasurable set in R.
Proof. Fix an innitedimensional separable Hilbert space (E, [[ [[) and denote
K =e E : [[e[[ < 2.
By virtue of Lemma 9, the open ball K is an Eabsolutely nonmeasurable subset of E.
Taking into account Lemma 10 and the fact that E and R are isomorphic as commutative
Algebraic sums of measure zero sets 193
groups, it is sufcient to show that there exist two Eabsolutely 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
: <, y
: <
of elements from E satisfying the following conditions:
(1) [[x
+y
= k
+x
) (G
+x
+y
) (G
+y
: < 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
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
.
194 Topics in Measure Theory and Real Analysis
Let us dene x
= x and y
= y.
Proceeding in this manner, we are able to construct the required sequences x
: <
and y
: <,
Y =y
: <,
we easily deduce that X +Y = K in view of (1) and (2). We also conclude that both X and
Y are Eabsolutely negligible subsets of E in view of (3), (4), and Lemma 5. Theorem 7
has thus been proved.
Actually, the preceding argument yields a much stronger result. Namely, we can assert that
there exists an Eabsolutely negligible set C E such that C+C = K. Indeed, it sufces
to put
C = X Y,
where X and Y are the abovementioned Eabsolutely negligible subsets of E. Therefore,
taking into account an algebraic isomorphism between the additive groups E and R, we
claim that there exists an Rabsolutely negligible set Z R such that the algebraic sum
Z +Z is Rabsolutely nonmeasurable.
Theorem 8. There are two subsets A and B of R having the following property:
for every nonzero nite Rinvariant (Rquasiinvariant) measure on R, there exists an
Rinvariant (Rquasiinvariant) measure
/
on R extending and such that
/
(A) =
/
(B) = 0,
A+B , dom(
/
).
Proof. Indeed, it sufces to take as A and B any two Rabsolutely negligible subsets of
R whose algebraic sum A+B is Rabsolutely 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;
Algebraic sums of measure zero sets 195
(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 Eabsolutely 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 nondegenerate 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 Ginvariant measure on G such that H dom().
196 Topics in Measure Theory and Real Analysis
Conclude from this fact that H is not absolutely nonmeasurable with respect to the class of
all nonzero nite Ginvariant measures on G.
6. By using the method of transnite recursion, construct an Rabsolutely 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 Habsolutely negligible
subsets of H. Show that G admits an analogous representation G = A
/
B
/
where A
/
and B
/
are some Gabsolutely negligible subsets of G.
8. Prove the equivalence of assertions (a) and (b) of Remark 5.
9
: <
1
of subgroups of H such that:
(a) card(H
;
(c) H
: <
1
.
Verify that any partial selector of the family
(H
: <) : <
1
: <
1
be an enumeration of all elements of H. By applying the method of
transnite recursion, construct two families
a
: <
1
, b
: <
1
: <
1
and b
: <
1
are partial selectors of the abovementioned family
(H
: <) : <
1
;
(f) a
= h
: <
1
and B =b
: <
1
, demonstrate that A and B are the
required Habsolutely negligible sets in H.
10. Let (H, ) be a group such that card(H) is an uncountable regular cardinal. Show that
there exist two Habsolutely 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).
Algebraic sums of measure zero sets 197
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 Gabsolutely
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 Ginvariant ideal I of subsets of
E generated by X. Since this I is proper, there exists a probability Ginvariant measure
on E such that
(Z I)((Z) = 0).
In particular, the relation (X) = 0 holds true which contradicts the Gabsolute 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 halfspace in E is Eabsolutely nonmeasurable;
(b) if E is innitedimensional and separable, then for every nonzero nite Equasi
invariant measure on E, there exists a closed halfspace 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 innitedimensional separable Hilbert space H is isomorphic to R in the
grouptheoretical 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 Rabsolutely negligible set C having the following property:
for every nite Rinvariant (Rquasiinvariant) measure on R, there exists an R
invariant (Rquasiinvariant) extension
/
of such that
/
(C) = 0, C+C , dom(
/
).
198 Topics in Measure Theory and Real Analysis
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 Eabsolutely negligible set Z such that Z +Z is Eabsolutely
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 Habsolutely nonmeasurable. In this case,
show that any uncountable commutative group (G, +) contains a Gabsolutely negligible
set Y such that Y +Y is Gabsolutely nonmeasurable.
20
. Let (G, ) be an uncountable solvable group. Show that there exist two Gabsolutely
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 textbooks 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 10.1007/9789491216367_1 2009 Atlantis Press/World Scientific 2,
2
200 Topics in Measure Theory and Real Analysis
nonempty interior, so can be a nonsmall 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 Hinvariant algebra of subsets of R is quasi
invariant under H (or, in short, Hquasiinvariant) 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 Hquasiinvariant but is not H
invariant. There are various extensions of which also are quasiinvariant 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 quasiinvariant 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 standpoint 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 measuretheoretical viewpoint.
202 Topics in Measure Theory and Real Analysis
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;
The absolute nonmeasurability of Minkowskis sum of certain universal measure zero sets 203
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
: <}, {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
+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
: <}.
204 Topics in Measure Theory and Real Analysis
Let z be a point of R dened as follows:
z = z
if z
+Y
+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
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.
The absolute nonmeasurability of Minkowskis sum of certain universal measure zero sets 205
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 measuretheoretical 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
). Then either
(Z) > 0 or
(Z) > 0.
Since
: < }
denote the family of all nonempty perfect subsets of E. We may assume, without loss of
generality, that each of the partial families
{P
: <} Y;
b) for every ordinal <, we have (X +y
) P
= / 0.
Suppose that for an ordinal < , the partial sequence {y
(P
(X +y
)) = 0 ( <).
In view of Martins Axiom (see Remark 1 below), we come to the relation
({P
(X +y
) : <}) = 0.
208 Topics in Measure Theory and Real Analysis
Therefore, we can write
P
\ {X +y
: <} = / 0.
Let z be a point of P
\ {X +y
= y.
Proceeding in this manner, we will be able to construct the desired sequence of points
{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:
The absolute nonmeasurability of Minkowskis sum of certain universal measure zero sets 209
if E is a second category topological space with a countable base, then the ideal of all
rst category subsets of E is cadditive.
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 cadditive
For more details, see e.g. [10], [40], [67], and [216].
Remark 2. It should be noticed that an additional settheoretical 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
 < +},
210 Topics in Measure Theory and Real Analysis
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
realvalued 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 realvalued 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 cadditive 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
,
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 Ginvariant or, more generally, Gquasi
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 Ginvariant (Gquasiinvariant) 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 Gabsolutely negligible sets and Gabsolutely 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 Gabsolutely
nonmeasurable?
212 Topics in Measure Theory and Real Analysis
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 HahnBanach 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}
. 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
. Suppose that
(y E
)(card( f
1
(y)) ).
Show that the set f (X) is universal measure zero.
10
subsets of R.
By using the method of transnite recursion, we dene a function
f : R R.
Absolutely nonmeasurable additive Sierpi nskiZygmund functions 219
Suppose that for an ordinal <, all values f (x
(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
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
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 nskiZygmund function.
Various works are devoted to Sierpi nskiZygmund functions (see, e.g., [7], [42], [123],
[124], [185], and [199]). In those works different constructions are presented which yield
further examples of Sierpi nskiZygmund functions with additional properties important
from the viewpoint 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 nskiZygmund 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
220 Topics in Measure Theory and Real Analysis
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 nskiZygmund function mea
surable with respect to some translationinvariant 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 settheoretical assumption, there
are absolutely nonmeasurable additive functions acting from R into R which are not
Sierpi nskiZygmund functions.
Theorem 1. Assume Martins Axiom. There exists a function
f : R R
satisfying the following conditions:
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 nskiZygmund 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
Absolutely nonmeasurable additive Sierpi nskiZygmund functions 221
identical transformation of K. This circumstance immediately implies that our f cannot be
a Sierpi nskiZygmund 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 nskiZygmund function acting
from R into G.
Theorem 2. Let G be as above. There exists a function
f : R R
satisfying the following conditions:
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 nskiZygmund 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 wellordering
which will be denoted by . Of course, this notation means that for any ordinal < =c,
the relation card() < c is valid.
Let be a wellordering of R isomorphic to .
Let {h
: <}, {V
: <}, { f
: <}
satisfying the following relations:
(a) {x
: };
(c) for each ordinal < , the function f
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
: <}
222 Topics in Measure Theory and Real Analysis
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
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
(qx +v) f
, qx +v dom(h
)}.
Since card(D) < c and card(ran( f
)).
Clearly, there exists a unique group monomorphism
f
: V
G
extending 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
: <} = R.
Since all f
( < ), we
have
card({z dom(h
) : f (z) = h
R
denote the class of the completions of all nonzero nite diffused Borel mea
sures on R. Prove that every Sierpi nskiZygmund function is absolutely nonmeasurable
with respect to M
R
.
For this purpose, consider any Sierpi nskiZygmund function f : R R and x a nonzero
nite diffused Borel measure on R. Suppose that f is measurable with respect to the
completion
be two innite sets and let be a family of partial functions acting from E
into E
. Suppose that
card(E) = card(E
) = card().
Absolutely nonmeasurable additive Sierpi nskiZygmund functions 225
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
set
in E of cardinality c.
Then the abovementioned mapping f : E R turns out to be a Sierpi nskiZygmund 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
having the following property:
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 nskiZygmund function but the converse
assertion is not true in general (see Exercise 8 above).
11
(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
((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
Relatively measurable Sierpi nskiZygmund functions 231
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.
232 Topics in Measure Theory and Real Analysis
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
, ...)
Relatively measurable Sierpi nskiZygmund functions 233
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 noncommutative 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 .
234 Topics in Measure Theory and Real Analysis
Let h
: < .
Under this notation, we are going to dene four sequences
x
: <, y
: < ,
V
: < , f
: <
satisfying the following relations:
(a) 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
: <
Relatively measurable Sierpi nskiZygmund functions 235
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
) + 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
) > 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
: V
T
extending f
/
and satisfying the condition f
(x
) = y
.
236 Topics in Measure Theory and Real Analysis
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 wellknown statement
from classical descriptive set theory, the function f [X can be extended to a Tvalued 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
: 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.
The assertion of Theorem 3 is much stronger and shows that, under Martins Axiom, there
exists a universal function h : R R of Sierpi nskiZygmund 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 settheoretical assumptions
are needed for the validity of Theorem 3.
Chapter 15
A nonseparable extension of the Lebesgue
measure without new nullsets
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 settheoretical 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
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_1 2009 Atlantis Press/World Scientific
241
5,
242 Topics in Measure Theory and Real Analysis
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 onetoone 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 settheoretical 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 socalled 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 nullsets 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 nullsets in E.
The symbol
(respectively,
/
(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 abovementioned 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 viewpoint of the theory of Boolean algebras, the complete measure and
its extension
/
obtained in the previous example are the same. Nevertheless, by slightly
244 Topics in Measure Theory and Real Analysis
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 nullsets.
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
A nonseparable extension of the Lebesgue measure without new nullsets 245
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 nullsets. In other words, there appear nullsets 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 abovementioned
intersection becomes a nullset 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 nullsets (see, for instance, Theorem 2 below).
In this connection, the following question seems to be interesting.
Does there exist a nonseparable extension of all nullsets of which are precisely the 
nullsets?
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 wellknown 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 Cproperty 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 nullset 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 nullsets.
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).
246 Topics in Measure Theory and Real Analysis
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),
A nonseparable extension of the Lebesgue measure without new nullsets 247
thus it follows that
/
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
= 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.
248 Topics in Measure Theory and Real Analysis
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
: <, is odd.
A nonseparable extension of the Lebesgue measure without new nullsets 249
Let T
, y
, y
(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
: <
satisfying the relation (Z
(x) T
(x) : <
and put (x, y) = (x
, y
).
Thus, in both cases (a) and (b) we have determined the point (x
, y
, 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 wellordering _ 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
, 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
, T
(x
). If is odd, then
our construction also yields y
, T
(x
(x
T
0
/
(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 nullsets 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 nullsets are identical with the nullsets (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
252 Topics in Measure Theory and Real Analysis
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
satisfying the following conditions:
(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.
A nonseparable extension of the Lebesgue measure without new nullsets 253
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 abovementioned 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.
254 Topics in Measure Theory and Real Analysis
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 nullsets.
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
realvalued 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)+
,
A nonseparable extension of the Lebesgue measure without new nullsets 255
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 realvalued 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 realvalued square integrable measurable functions is nonseparable and,
consequently, the measure is nonseparable, too.
For this purpose, use the Erd osRado 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
(g(Y)Y) = 0
holds true for all transformations g G.
Clearly, Y E is almost Ginvariant 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 Ginvariant sets with respect to is
almost Ginvariant with respect to , too.
Therefore, the class of all almost Ginvariant sets with respect to forms a Ginvariant
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 Gquasiinvariant metrically transitive measure on E. Suppose that a set
Y E is almost Ginvariant 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 Ginvariant with respect to , we may write
(Y({g
i
(Y) : i I})) = 0,
thus it follows that
(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).
260 Topics in Measure Theory and Real Analysis
For our further purposes, we need the notion of Ulams matrix which is very important
in combinatorial set theory, measure theory and settheoretic 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 realvalued measurable cardinal. Some applications of Ulam matrices to the theory
of invariant and quasiinvariant 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 Gquasiinvariant 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
(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
(x) : <
1
}.
Now, putting
V
= G
(x) \ {G
(x) : <}
262 Topics in Measure Theory and Real Analysis
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
: 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 Ginvariant 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
of . More precisely,
) is equal to 2
c
.
Metrical transitivity and nonseparable extensions of invariant measures 263
Proof. We preserve the notation of Lemma 3. Take a family {Y
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 Ginvariant 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 KakutaniOxtoby method
of extending invariant measures (see [83], [95], or Exercise 18 from Chapter 3), we can
extend our to a Ginvariant measure
) is equal to 2
c
(see again [83] and [95]) which completes the
proof of Theorem 1.
Unfortunately, the obtained extension
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 measuretheoretical version of their method.
264 Topics in Measure Theory and Real Analysis
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
) 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 settheoretical assumptions are necessary for obtaining the abovementionedstronger
result.
Example 2. Let E be a set of cardinality c and let 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, Ginvariant, 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 Ginvariant
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 Gquasiinvariant, metrically transitive, and
Metrical transitivity and nonseparable extensions of invariant measures 265
separable. Analogously to the measure from the previous example, this
1
has no proper
extensions and, consequently, does not admit a nonseparable Gquasiinvariant extension.
EXERCISES
1
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 .
266 Topics in Measure Theory and Real Analysis
Show that the nonseparable translationinvariant 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
Ginvariant (Gquasiinvariant) 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 Ginvariant 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
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 Ginvariant (Gquasiinvariant) 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 Ginvariant (Gquasiinvariant)
measure on E for which there exists at least one admissible family of subsets of E.
8
is a nonseparable extension of .
11. Prove that there exists a nonseparable translationinvariant 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 nonatomic 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.
268 Topics in Measure Theory and Real Analysis
13
. Let T as usual denote the onedimensional 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 (quasiinvariant)
nonseparable measures in the abovementioned more general case. However, under some
purely settheoretical 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 nonatomic 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 quasimetric and after the corresponding factorization
it yields a complete metric space canonically associated with (cf. Exercise 12 of Chapter
1).
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_1 2009 Atlantis Press/World Scientific
269
7,
270 Topics in Measure Theory and Real Analysis
We recall that a measure is nonseparable if the abovementioned 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 onedimensional 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 translationinvariant 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 abovementioned 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 leftinvariant 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 cocountable 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 nonatomic nite leftinvariant 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 nonatomic 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 socalled 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 quasiinvariant) 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 Ginvariant) 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
nonatomic nite invariant measure. Then every uncountable commutative group admits
a nonseparable nonatomic 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 nonatomic nite invariant measure . Consider the family of
272 Topics in Measure Theory and Real Analysis
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 Ginvariant algebra of subsets of G;
(c) is a nonatomic nite measure on S.
Let us check that is also nonseparable and Ginvariant.
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 quasimetric 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 Ginvariant measure. Lemma 2 has thus been proved.
Lemma 3. Suppose again that every commutative group of cardinality
1
admits a non
separable nonatomic nite invariant measure. Then every uncountable solvable group
admits a nonseparable nonatomic 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
Nonseparable left invariant measures on uncountable solvable groups 273
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 nonatomic 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
nonatomic 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 nonatomic 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 nonatomic 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
274 Topics in Measure Theory and Real Analysis
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 nonatomic 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 nonatomic 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 nonatomic 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
: }
is almost Hinvariant 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 abovementioned 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:
Nonseparable left invariant measures on uncountable solvable groups 275
(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 Hinvariant in H by virtue of relation (b).
Let
0
denote the probability Hinvariant measure on H which is dened for all countable
and cocountable 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 nonatomic Hinvariant 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
has thus been proved.
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);
276 Topics in Measure Theory and Real Analysis
(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 nonatomic 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 Ginvariant 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 nonatomic
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
Nonseparable left invariant measures on uncountable solvable groups 277
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 quasimetric 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
preimage
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 translationinvariant 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
assertions are equivalent:
278 Topics in Measure Theory and Real Analysis
(a) there exists a nite (nite) nonseparable nonatomic Ginvariant measure on E;
(b) there exists a nite (nite) nonseparable Ginvariant 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]
satisfying the following conditions:
(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 realvalued
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
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
( ).
Nonseparable left invariant measures on uncountable solvable groups 279
Check that the obtained family {T
g
: g G} satises the assumptions of MarkovKakutanis
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
: P(E) [0, +]
280 Topics in Measure Theory and Real Analysis
which extends and is also nitely additive and Ginvariant.
In order to demonstrate the existence of
: 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 realvalued 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
.
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 Ginvariance 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 twodimensional 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 innitedimensional vector spaces (over R) will be under
consideration in this chapter. In particular, for an innitedimensional 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 realvalued 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 quasiinvariant 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 abovementioned 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).
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_1 2009 Atlantis Press/World Scientific
281
8,
282 Topics in Measure Theory and Real Analysis
It is well known that if E is a Polish space, then all analytic (also, coanalytic) 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 innitedimensional 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
: < 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
satisfying the following conditions:
(a) the family x
: < y
and y
.
Universally measurable additive functionals 285
Now, let us take as X
1
the vector space (over Q) generated by x
: < 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
having the following property:
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.
286 Topics in Measure Theory and Real Analysis
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,
which completes the proof of Lemma 2.
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.
Universally measurable additive functionals 287
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 measuretheoretical point of view.
It is well known that if H is an innitedimensional separable Hilbert space, then there
exists no nonzero nite Borel measure on H invariant (quasiinvariant) 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 onedimensional 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.
288 Topics in Measure Theory and Real Analysis
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 (quasiinvariant) under some subgroup G of H. Then the measure
/
is also invariant (quasiinvariant) under G.
Proof. It sufces to consider the case of an invariant measure (for quasiinvariant 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)))
Universally measurable additive functionals 289
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 innitedimensional 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 abovementioned 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 wellordered
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),
290 Topics in Measure Theory and Real Analysis
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
satisfying the following conditions:
(a) the family y
.
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
Universally measurable additive functionals 291
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 (quasiinvariant) 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
realvalued 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 midpoint 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
midpoint convex function dened on a nitedimensional Euclidean space is necessarily
continuous (see [220]; cf. also [122] and [143]).
292 Topics in Measure Theory and Real Analysis
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
=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 abovementioned
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 realvalued Lebesgue measurable functions.
Universally measurable additive functionals 293
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
. 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 midpoint 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 midpoint convex Lebesgue measurable func
tion
f : R
n
R
is continuous and, therefore, is convex in the usual sense.
This theorem strengthens the wellknown 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 nitedimensional if dim(E) <. Otherwise, we
say that E is innitedimensional.
Let E be an innitedimensional 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.
Universally measurable additive functionals 295
For this purpose, denote by the least ordinal number whose cardinality is equal to card(I).
Let V
) = 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 innitedimensional 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 socalled 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 measuretheoretical 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 translationinvariant extensions of the classical Lebesgue measure
2
on R
2
.
Many examples of paradoxical sets in a nitedimensional 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 translationinvariant extensions of the
Lebesgue measure
2
?
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367_1 2009 Atlantis Press/World Scientific
297
9,
298 Topics in Measure Theory and Real Analysis
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 ndimensional 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 Ginvariant, 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
) card() +;
(b) the family {G
: < } of points in R
n
satisfying the following relations:
(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
(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 Ginvariant;
2) X is
n
thick in R
n
;
300 Topics in Measure Theory and Real Analysis
3) for each analytic manifold M in R
n
with dim(M) < n, we have
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
with dim(M) < n, we have
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
Ginvariant 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 Ginvariant 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
satisfying the following relations:
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.
Some subsets of the Euclidean plane 301
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 twodimensional 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 translationinvariant 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 settheoretical 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
302 Topics in Measure Theory and Real Analysis
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
satisfying the following conditions:
1) card(X) = c and X is almost Ginvariant;
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 abovementionedset 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 measuretheoretical viewpoint 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 wellknown 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 osSzekeres 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.
Some subsets of the Euclidean plane 303
The abovementioned 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 pointsets 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 ielement 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 abovementioned 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
AL
. Now, applying to 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 endpoint o of p and the
rays oa
n
(n < ) converge to p;
2) there exist a ray p R
2
and an innite set {a
n
: n < } A
, such that
{a
n
: n < } p = / 0, lim
n
a
n
 = +, lim
n
dist(a
n
, p) = 0;
3) there exist a ray p R
2
and an innite set {a
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 endpoint 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
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 preimage with respect to the
abovementioned projection of this subset yields the required innite convexly independent
subset of A
: < } of points in general position in the plane. Suppose that, for an ordinal
< , the partial family of points {a
) 2).
Clearly, we can nd a point a
: < } {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
: <
}. 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
as the bound
ary of B
) card(B) = card() = c.
On the other hand, there exists an ordinal < such that C
=C
{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 settheoretical 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
Some subsets of the Euclidean plane 307
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 settheoretical axioms. In
this connection, see [148], [176] or [192] where by assuming the Continuum Hypothesis
the classical constructions of socalled 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
. 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 settheoretical hypoth
esis:
c is measurable in the Ulam sense.
In other words, prove that if c is a realvalued 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
. 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
satisfying the following relations:
(a) X is T
2
negligible;
(b) X is Gabsolutely 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
satisfying the following relations:
(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
satisfying the following conditions:
310 Topics in Measure Theory and Real Analysis
(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 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 otset if for any three pairwise
distinct points y
1
, y
2
, y
3
fromY, the triangle [y
1
, y
2
, y
3
] is nondegenerate and has an obtuse
angle.
Demonstrate the validity of the following assertions:
Some subsets of the Euclidean plane 311
(a) every otset is contained in a maximal (with respect to inclusion) otset;
(b) no nite otset is maximal;
(c) there exists a countable maximal otset.
In addition to these facts, verify that
{(cos(t), sin(t)) : 0 t < }
is a maximal otset 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
socalled 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
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
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 realvalued 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.
Restrictions of realvalued functions 317
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),
which completes the proof of Lemma 1.
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).
318 Topics in Measure Theory and Real Analysis
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 supnorm
[[ 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
preimages
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),
Restrictions of realvalued functions 319
g(x) >t (k <)(n <)(m [n, [)(g
m
(x) t +1/k).
These relations yield at once that the abovementioned sets
x E : g(x) <t, x E : g(x) >t
are of type 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 TietzeUrysohn 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
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 preimage 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
Restrictions of realvalued functions 321
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
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 abovementioned 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. 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
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 <.
324 Topics in Measure Theory and Real Analysis
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 CantorBaire 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
)
Restrictions of realvalued functions 325
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
: <
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
: <
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
,
326 Topics in Measure Theory and Real Analysis
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 CantorBaire 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;
Restrictions of realvalued functions 327
(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
: < denote the family of all those Borel subsets of R which have Lebesgue
measure zero, i.e.,
T
, Z
1
of Z
such that Z
0
is a Lebesgue measure
zero set and Z
1
: < 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
: < 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
328 Topics in Measure Theory and Real Analysis
our set X. Furthermore, since x
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 abovementioned 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.
330 Topics in Measure Theory and Real Analysis
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 supmeasurability of functions acting from RR into R.
Now, we are ready to prove the following statement.
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 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
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.
Restrictions of realvalued functions 333
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 righthand 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.
334 Topics in Measure Theory and Real Analysis
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 CantorBendixson 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.
Restrictions of realvalued functions 335
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 abovementioned 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
=F
: <.
If = +1, consider the set F
= F
= / 0.
(ii) 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 CantorBaire 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
336 Topics in Measure Theory and Real Analysis
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.
Restrictions of realvalued functions 337
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
: <
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
(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 <
338 Topics in Measure Theory and Real Analysis
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 settheoretical facts and constructions
The main goal of this Appendix is to present several purely settheoretical 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 wellordered.
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 abovementioned wellordered 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
=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).
A.B. Kharazishvili, Topics in Measure Theory and Real Analysis, Atlantis Studies in Mathematics 2,
DOI 10.1007/9789491216367, 2009 Atlantis Press/World Scientific
339
340 Topics in Measure Theory and Real Analysis
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 nth 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 wellordered 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 nbranch in
(E, _) if X is linearly ordered by the induced ordering, has nonempty intersection with all
kth levels of (E, _) where k < n, but does not intersect the kth levels of (E, _), where
k n.
Obviously, every partial nbranch contains exactly n elements.
An element e E will be called admissible if for any natural number n > 0, there exists a
partial nbranch 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 kth 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 settheoretical 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
342 Topics in Measure Theory and Real Analysis
satisfying the following three relations:
(a) dom( f
[[0, [ 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
/
F and f
/
[[0, [ f
: [0, [
can be dened satisfying the next two conditions:
(iii) f
() = f
/
) is innite.
A straightforward verication shows that 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
)
Appendix 1. Some settheoretical facts and constructions 343
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 RyllNardzewski 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 realvalued 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 realvalued 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 realvalued 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 realvalued 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()
344 Topics in Measure Theory and Real Analysis
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 abovementioned generalized result, let us dene
E
(x) =y E : x _y
for each x E and let 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).
Appendix 1. Some settheoretical facts and constructions 345
Indeed, to verify this fact, take any y S. Then y 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
(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
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:
Appendix 1. Some settheoretical facts and constructions 347
(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
abovementioned 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
satisfying the following conditions:
(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.
348 Topics in Measure Theory and Real Analysis
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
).
Appendix 1. Some settheoretical facts and constructions 349
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
. 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
.
350 Topics in Measure Theory and Real Analysis
4. Verify that the K onig lemma is essentially used in the KakutaniOxtoby construction of
a nonseparable translationinvariant 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 stepfunction and then approximate an arbitrary partial function acting
from E into itself by a sequence of partial stepfunctions.
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
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 abovementioned 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
(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 Gquasiinvariant measure on E. Denote by H
a subgroup of G with card(H) =
1
and let X be an arbitrary Hselector 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 ,=;
352 Topics in Measure Theory and Real Analysis
(f) (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
realvalued 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
. 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
following relations:
(a) AA
/
= / 0 and AA
/
= X;
(b) BB
/
= / 0 and BB
/
=Y;
(c) ran( f [A) = B;
(d) ran(g[B
/
) = A
/
.
Appendix 1. Some settheoretical facts and constructions 353
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
wellknown CantorBernstein 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 abovementioned result of Banach from Tarskis xedpoint 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:
354 Topics in Measure Theory and Real Analysis
(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
. 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
translationquasiinvariant measures on R (cf. Chapter 5).
17
: [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
is not a realvalued measurable cardinal.
In a similar way, construct Ulams (
+1
,
)matrix of subsets of
+1
, where is an
arbitrary ordinal number, and show that if the cardinal
: <
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 twodimensional Lebesgue measure on the Euclidean plane
R
2
. Show that if the cardinality of the continuum is realvalued 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 realvalued 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 realvalued measurability of c (for more
details, see [87]). Therefore, we can conclude that if c is not realvalued measurable, then
358 Topics in Measure Theory and Real Analysis
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 multivalued (setvalued) 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 multivalued 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
abovementioned 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
360 Topics in Measure Theory and Real Analysis
We shall concentrate our attention on the following two classical results: LuzinJankovvon
Neumann theorem (see, e.g., [99]) and the theorem of Kuratowski and RyllNardzewski
[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 wellknown 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 abovementioned 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
.
362 Topics in Measure Theory and Real Analysis
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 <).
Appendix 2. The Choquet theorem and measurable selectors 363
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
. 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
measuretheoretical 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
364 Topics in Measure Theory and Real Analysis
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 realvalued 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
(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]).
Appendix 2. The Choquet theorem and measurable selectors 367
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
dened by the formula
(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 preimage
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
368 Topics in Measure Theory and Real Analysis
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 halfline
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 halfline 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 halfline 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
+
dened by the formula
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 realvalued function measurable with respect to the algebras B(R
+
)
and S.
Proof. Indeed, for any real number t 0, the preimage
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
Appendix 2. The Choquet theorem and measurable selectors 369
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 socalled 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
, 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
)) = (E pr
1
(Z
)) <.
For the set Z
) R
+
which is dened on the set pr
1
(Z
), the section
Z
(x) =t R
+
: (x, t) 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.
Appendix 2. The Choquet theorem and measurable selectors 371
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 RyllNardzewski below).
Now, let us turn our attention to setvalued 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 setvalued mapping such that F(x)
is a nonempty closed subset of Y for every point x X. In the sequel, only those setvalued
functions F will be considered which satisfy this condition.
We say that F is a weakly measurable setvalued mapping if for any open subset U of Y,
we have
x X : F(x) U ,= / 0 S.
372 Topics in Measure Theory and Real Analysis
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 setvalued 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
dened by the formula
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 setvalued
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
dened by the formula
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).
Appendix 2. The Choquet theorem and measurable selectors 373
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 realvalued 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 RyllNardzewski 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 setvalued 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)).
374 Topics in Measure Theory and Real Analysis
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 setvalued 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
Appendix 2. The Choquet theorem and measurable selectors 375
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 setvalued 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 setvalued 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
;
376 Topics in Measure Theory and Real Analysis
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 RyllNardzewski (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 RyllNardzewski 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
satisfying the following relations:
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.
Appendix 2. The Choquet theorem and measurable selectors 377
Proof. Let us equip the set of all natural numbers with the discrete topology and take the
product space
E
1
E
2
such that g(
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 setvalued 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
RyllNardzewski. 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 RyllNardzewski
are discussed in monograph [68] where it is shown that some additional settheoretical
axioms are necessary for appropriate generalizations of this theorem (see also [38]).
EXERCISES
378 Topics in Measure Theory and Real Analysis
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
. Starting with the fact that there are two disjoint coSuslin 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 RyllNardzewski, 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
/
denotes the completion of .
Deduce from the Choquet theorem that all members of the analytic class A(S) are abso
lutely measurable with respect to the algebra S.
Appendix 2. The Choquet theorem and measurable selectors 379
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),
380 Topics in Measure Theory and Real Analysis
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
. 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;
Appendix 2. The Choquet theorem and measurable selectors 381
(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 realvalued continuous functions on Y and dene an appropriate setvalued
weakly measurable mapping
F : X P(C(Y))
associated with the given mapping f
0
. Then apply to F the theorem of Kuratowski and
RyllNardzewski.
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 coSuslin 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 setvalued 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 realvalued 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
384 Topics in Measure Theory and Real Analysis
we have respectively
w(R
N
) = w(c
0
) = card(N) =,
w(l
) = card(R) = c.
Recall that an innite cardinal number a is realvalued 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 oneelement 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 realvalued measurable if and only if there exists a diffused probability
measure on a with dom() =P(a).
Extensive information about realvalued 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 realvalued measurable cardinals
cannot be established within ZFC theory. Consequently, the assumption that all cardinals
are not realvalued measurable is consistent with ZFC (cf. Exercise 18 for Appendix 1).
So, in order to manipulate with realvalued measurable cardinals, we are forced to appeal
to additional settheoretical 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]
dened by the formula
(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
of E such that
(E F) = 0.
386 Topics in Measure Theory and Real Analysis
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
Appendix 3. Borel measures on metric spaces 387
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 oneelement subsets of J(k). In other words, we
derive that card(J(k)) is a realvalued 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 realvalued 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 nonrealvalued mea
surable topological weight is very essential for the validity of this theorem. Indeed, suppose
that a realvalued 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 oneelement 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).
388 Topics in Measure Theory and Real Analysis
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 BanachMazur 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 oneelement 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
.
Appendix 3. Borel measures on metric spaces 389
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
) = c.
2. Let (E, d) be a complete metric space whose topological weight is not realvalued 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) <.
390 Topics in Measure Theory and Real Analysis
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 realvalued 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
: <) =M
: <
and we say that the set M(X
: < by applying
the Montgomery operation (in short, Moperation) 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;
Appendix 3. Borel measures on metric spaces 391
(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) : <.
If > 0 is even, then let 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
.
392 Topics in Measure Theory and Real Analysis
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
. Let E be a separable Hilbert space (over R) equipped with the inner product <, >.
We recall that a set H E is a (closed) halfspace 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 halfspaces of E. This phrase means that if is any Borel probability measure
on E such that (H) =(H) for all halfspaces H E, then =.
Appendix 3. Borel measures on metric spaces 393
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 cadditive, 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
. 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]).
Appendix 3. Borel measures on metric spaces 395
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 (comeager) 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
398 Topics in Measure Theory and Real Analysis
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.
400 Topics in Measure Theory and Real Analysis
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].
Appendix 4. Continuous nowhere approximately differentiable functions 401
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
satisfying the relations
h
(0, 0) = (k/9
n
, f
n
((k +1)/9
n
)), h
(0, 3) = (k/9
n
, f
n
(k/9
n
)),
h
) 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
]
Appendix 4. Continuous nowhere approximately differentiable functions 403
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
]).
404 Topics in Measure Theory and Real Analysis
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 supmeas