1

LECTURE NOTES

IN MEASURE THEORY

Christer Borell
Matematik
Chalmers och Göteborgs universitet
412 96 Göteborg
(Version: Feb 05)
2

PREFACE

These are lecture notes on integration theory for a ten-week course at the
Chalmers University of Technology and the Göteborg University. The parts
de…ning the course essentially lead to the same results as the …rst three
chapters in the Folland book [F ] ; which is used as a text book on the course.
The proofs in the lecture notes sometimes di¤er from those given in [F ] : Here
is a brief description of the di¤erences to simplify for the reader.
In Chapter 1 we introduce so called - and -systems, which are substi-
tutes for monotone classes of sets [F ]. Besides we prefer to emphasize metric
outer measures instead of so called premeasures. Throughout the course, a
variety of important measures are obtained as image measures of the linear
measure on the real line. In Section 1.6 positive measures in R induced by
increasing right continuous mappings are constructed in this way.
Chapter 2 deals with integration and is very similar to [F ] and most
other texts.
Chapter 3 starts with some standard facts about metric spaces and relates
the concepts to measure theory. For example Ulam’s Theorem is included.
The existence of product measures is based on properties of - and -systems.
Chapter 4 deals with di¤erent modes of convergence and is mostly close
to [F ] : Here we include a section about orthogonality since many students
have seen parts of this theory before.
The Lebesgue Decomposition Theorem and Radon-Nikodym Theorem
in Chapter 5 are proved using the von Neumann beautiful L2 -proof.
To illustrate the power of abstract integration these notes contain several
sections, which do not belong to the course but may help the student to a
better understanding of measure theory. The corresponding parts are set
between the symbols
###
and
"""
respectively.
Finally I would like to express my deep gratitude to the students in
my classes for suggesting a variety of improvements and a special thank
 3

to Jonatan Vasilis who has provided numerous comments and corrections in

my original text.

Göteborg 2006
Christer Borell
4

CONTENT
1 Measures
1.1 -Algebras and Measures
1.2 Measure Determining Classes
1.3 Lebesgue Measure
1.4 Carathéodory’s Theorem
1.5 Existence of Linear Measure

2 Integration
2.1 Integration of Functions with Values in [0; 1]
2.2 Integration of Functions with Arbitrary Sign
2.3 Comparison of Riemann and Lebesgue Integrals

3 Further Construction Methods of Measures

3.1 Metric Spaces
3.2 Linear Functionals and Measures
3.3 q-Adic Expansions of Numbers in the Unit Interval
3.4 Product Measures
3.5 Change of Variables in Volume Integrals
3.6 Independence in Probability

4 Modes of Convergence
4.1 Convergence in Measure, in L1 ( ); and in L2 ( )
4.2 Orthogonality
4.3 The Haar Basis and Wiener Measure

5 Decomposition of Measures
5.1 Complex Measures
5.2 The Lebesgue Decomposition and the Radon-Nikodym Theorem
5.3 The Wiener Maximal Theorem and Lebesgue Di¤erentiation Theorem
 5

5.4 Absolutely Continuous Functions and Functions of Bounded Variation

5.5 Conditional Expectation

6 Complex Integration
6.1 Complex Integrand
6.2 The Fourier Transform
6.3 Fourier Inversion
6.4 Non-Di¤erentiability of Brownian Paths

References
6

CHAPTER 1
MEASURES

Introduction

The Riemann integral, dealt with in calculus courses, is well suited for com-
putations but less suited for dealing with limit processes. In this course we
will introduce the so called Lebesgue integral, which keeps the advantages of
the Riemann integral and eliminates its drawbacks. At the same time we will
develop a general measure theory which serves as the basis of contemporary
analysis and probability.
In this introductory chapter we set forth some basic concepts of measure
theory, which will open for abstract Lebesgue integration.

1.1. -Algebras and Measures

Throughout this course

N = f0; 1; 2; :::g (the set of natural numbers)

Z = f:::; 2; 1; 0; 1; ; 2; :::g (the set of integers)
Q = the set of rational numbers
R = the set of real numbers
C = the set of complex numbers.

If A R; A+ is the set of all strictly positive elements in A:

If f is a function from a set A into a set B; this means that to every x 2 A
there corresponds a point f (x) 2 B and we write f : A ! B: A function is
often called a map or a mapping. The function f is injective if

(x 6= y) ) (f (x) 6= f (y))
 7

and surjective if to each y 2 B; there exists an x 2 A such that f (x) = y:

An injective and surjective function is said to be bijective.
A set A is …nite if either A is empty or there exist an n 2 N+ and a
bijection f : f1; :::; ng ! A: The empty set is denoted by : A set A is said
to be denumerable if there exists a bijection f : N+ ! A: A subset of a
denumerable set is said to be at most denumerable.
Let X be a set. For any A X; the indicator function A of A relative
to X is de…ned by the equation

1 if x 2 A
A (x) =
0 if x 2 Ac :
The indicator function A is sometimes written 1A : We have the following
relations:
Ac =1 A

A\B = min( A; B) = A B

and
A[B = max( A; B) = A + B A B:

De…nition 1.1.1. Let X be a set.

a) A collection A of subsets of X is said to be an algebra in X if A has
the following properties:

(i) X 2 A:
(ii) A 2 A )Ac 2 A; where Ac is the complement of A relative to X:
(iii) If A; B 2 A then A [ B 2 A:

(b) A collection M of subsets of X is said to be a -algebra in X if M

is an algebra with the following property:

If An 2 M for all n 2 N+ , then [1

n=1 An 2 M:
8

If M is a -algebra in X; (X; M) is called a measurable space and the

members of M are called measurable sets. The so called power set P(X),
that is the collection of all subsets of X, is a -algebra in X: It is simple to
prove that the intersection of any family of -algebras in X is a -algebra. It
follows that if E is any subset of P(X); there is a unique smallest -algebra
(E) containing E; namely the intersection of all -algebras containing E:
The -algebra (E) is called the -algebra generated by E: The -algebra
generated by all open intervals in R is denoted by R. It is readily seen that
the -algebra R contains every subinterval of R. Before we proceed, recall
that a subset E of R is open if to each x 2 E there exists an open subinterval
of R contained in E and containing x; the complement of an open set is said
to be closed. We claim that R contains every open U subset of R: To see
this suppose x 2 U and let x 2 ]a; b[ U; where 1 < a < b < 1: Now
pick r; s 2 Q such that a < r < x < s < b: Then x 2 ]r; s[ U and it follows
that U is the union of all bounded open intervals with rational boundary
points contained in U: Since this family of intervals is at most denumberable
we conclude that U 2 R: In addition, any closed set belongs to R since its
complements is open. It is by no means simple to grasp the de…nition of R at
this stage but the reader will successively see that the -algebra R has very
nice properties. At the very end of Section 1.3, using the so called Axiom of
Choice, we will exemplify a subset of the real line which does not belong to
R. In fact, an example of this type can be constructed without the Axiom
of Choice (see Dudley’s book [D]).
In measure theory, inevitably one encounters 1: For example the real
line has in…nite length. Below [0; 1] = [0; 1[ [ f1g : The inequalities x y
and x < y have their usual meanings if x; y 2 [0; 1[. Furthermore, x 1
if x 2 [0; 1] and x < 1 if x 2 [0; 1[ : We de…ne x + 1 = 1 + x = 1 if
x; y 2 [0; 1] ; and

0 if x = 0
x 1=1 x=
1 if 0 < x 1:

Sums and multiplications of real numbers are de…ned in the usual way.
If An X; n 2 N+ , and Ak \ An = if k 6= n, the sequence (An )n2N+ is
called a disjoint denumerable collection. If (X; M) is a measurable space, the
collection is called a denumerable measurable partition of A if A = [1 n=1 An
and An 2 M for every n 2 N+ : Some authors call a denumerable collection
of sets a countable collection of sets.
 9

De…nition 1.1.2. (a) Let A be an algebra of subsets of X: A function

: A ! [0; 1] is called a content if

(i) ( ) = 0
(ii) (A [ B) = (A) + (B) if A; B 2 A and A \ B = :

(b) If (X; M) is a measurable space a content de…ned on the -algebra M

is called a positive measure if it has the following property:

For any disjoint denumerable collection (An )n2N+ of members of M

([1
n=1 An ) =
1
n=1 (An ):

If (X; M) is a measurable space and the function : M ! [0; 1] is a

positive measure, (X; M; ) is called a positive measure space. The quantity
(A) is called the -measure of A or simply the measure of A if there is
no ambiguity. Here (X; M; ) is called a probability space if (X) = 1; a
…nite positive measure space if (X) < 1; and a -…nite positive measure
space if X is a denumerable union of measurable sets with …nite -measure.
The measure is called a probability measure, …nite measure, and -…nite
measure, if (X; M; ) is a probability space, a …nite positive measure space,
and a -…nite positive measure space, respectively. A probability space is
often denoted by ( ; F; P ): A member A of F is called an event.
As soon as we have a positive measure space (X; M; ), it turns out to
be a fairly simple task to de…ne a so called -integral
Z
f (x)d (x)
X

as will be seen in Chapter 2.

10

The class of all …nite unions of subintervals of R is an algebra which is

denoted by R0 : If A 2 R0 we denote by l(A) the Riemann integral
Z 1
A (x)dx
1

and it follows from courses in calculus that the function l : R0 ! [0; 1] is a

content. The algebra R0 is called the Riemann algebra and l the Riemann
content. If I is a subinterval of R, l(I) is called the length of I: Below we
follow the convention that the empty set is an interval.
If A 2 P(X), cX (A) equals the number of elements in A, when A is a
…nite set, and cX (A) = 1 otherwise. Clearly, cX is a positive measure. The
measure cX is called the counting measure on X:
Given a 2 X; the probability measure a de…ned by the equation a (A) =
A (a); if A 2 P(X); is called the Dirac measure at the point a: Sometimes
we write a = X;a to emphasize the set X:
If and are positive measures de…ned on the same -algebra M, the
sum + is a positive measure on M: More generally, + is a positive
measure for all real ; 0: Furthermore, if E 2 M; the function (A) =
(A \ E); A 2 M; is a positive measure. Below this measure will be
denoted by E and we say that E is concentrated on E: If E 2 M; the class
ME = fA 2 M; A Eg is a -algebra of subsets of E and the function
(A) = (A), A 2 ME ; is a positive measure. Below this measure will be
denoted by jE and is called the restriction of to ME :
Let I1 ; :::; In be subintervals of the real line. The set

I1 ::: In = f(x1 ; :::; xn ) 2 Rn ; xk 2 Ik ; k = 1; :::; ng

is called an n-cell in Rn ; its volume vol(I1 ::: In ) is, by de…nition, equal

to
n
vol(I1 ::: In ) = k=1 l(Ik ):

If I1 ; :::; In are open subintervals of the real line, the n-cell I1 ::: In is
called an open n-cell. The -algebra generated by all open n-cells in Rn is
denoted by Rn : In particular, R1 = R. A basic theorem in measure theory
states that there exists a unique positive measure vn de…ned on Rn such that
the measure of any n-cell is equal to its volume. The measure vn is called the
volume measure on Rn or the volume measure on Rn : Clearly, vn is -…nite.
The measure v2 is called the area measure on R2 and v1 the linear measure
on R:
 11

Theorem 1.1.1. The volume measure on Rn exists.

Theorem 1.1.1 will be proved in Section 1.5 in the special case n = 1. The
general case then follows from the existence of product measures in Section
3.4. An alternative proof of Theorem 1.1.1 will be given in Section 3.2. As
soon as the existence of volume measure is established a variety of interesting
measures can be introduced.
Next we prove some results of general interest for positive measures.

Theorem 1.1.2. Let A be an algebra of subsets of X and a content

de…ned on A. Then,
(a) is …nitely additive, that is

(A1 [ ::: [ An ) = (A1 ) + ::: + (An )

if A1 ; :::; An are pairwise disjoint members of A:

(b) if A; B 2 A;

(A) = (A n B) + (A \ B):

Moreover, if (A \ B) < 1; then

(A [ B) = (A) + (B) (A \ B)

(c) A B implies (A) (B) if A; B 2 A:

(d) …nitely sub-additive, that is

(A1 [ ::: [ An ) (A1 ) + ::: + (An )

if A1 ; :::; An are members of A:

If (X; M; ) is a positive measure space

12

(e) (An ) ! (A) if A = [n2N+ An ; An 2 M; and

A1 A2 A3 ::: :

(f) (An ) ! (A) if A = \n2N+ An ; An 2 M;

A1 A2 A3 :::

and (A1 ) < 1:

(g) is sub-additive, that is for any denumerable collection (An )n2N+ of
members of M,
([1n=1 An )
1
n=1 (An ):

PROOF (a) If A1 ; :::; An are pairwise disjoint members of A;

([nk=1 Ak ) = (A1 [ ([nk=2 Ak ))

= (A1 ) + ([nk=2 Ak )
and, by induction, we conclude that is …nitely additive.

(b) Recall that

A n B = A \ Bc:
Now A = (A n B) [ (A \ B) and we get

(A) = (A n B) + (A \ B):

Moreover, since A [ B = (A n B) [ B;

(A [ B) = (A n B) + (B)
and, if (A \ B) < 1; we have

(A [ B) = (A) + (B) (A \ B).

(c) Part (b) yields (B) = (B n A) + (A \ B) = (B n A) + (A); where

the last member does not fall below (A):
 13

(d) If (Ai )ni=1 is a sequence of members of A de…ne the so called disjunction

(Bk )nk=1 of the sequence (Ai )ni=1 as

B1 = A1 and Bk = Ak n [ki=11 Ai for 2 k n:

Then Bk Ak ; [ki=1 Ai = [ki=1 Bi ; k = 1; ::; n; and Bi \Bj = if i 6= j: Hence,
by Parts (a) and (c),

([nk=1 Ak ) = n
k=1 (Bk ) n
k=1 (Ak ):

(e) Set B1 = A1 and Bn = An n An 1 for n 2: Then An = B1 [ :::: [ Bn ;

Bi \ Bj = if i 6= j and A = [1
k=1 Bk : Hence

n
(An ) = k=1 (Bk )

and
1
(A) = k=1 (Bk ):
Now e) follows, by the de…nition of the sum of an in…nite series.

(f) Put Cn = A1 n An ; n 1: Then C1 C2 C3 :::;

A1 n A = [1
n=1 Cn

and (A) (An ) (A1 ) < 1: Thus

(Cn ) = (A1 ) (An )

and Part (e) shows that

(A1 ) (A) = (A1 n A) = lim (Cn ) = (A1 ) lim (An ):

n!1 n!1

This proves (f).

(g) The result follows from Parts d) and e).

This completes the proof of Theorem 1.1.2.
14

The hypothesis ” (A1 ) < 1 ”in Theorem 1.1.2 ( f) is not super‡uous. If

cN+ is the counting measure on N+ and An = fn; n + 1; :::g ; then cN+ (An ) =
1 for all n but A1 A2 :::: and cN+ (\1 1
n=1 An ) = 0 since \n=1 An = :
If A; B X; the symmetric di¤erence A B is de…ned by the equation

A B =def (A n B) [ (B n A):

Note that
A B =j A B j:
Moreover, we have
A B = Ac B c
and
([1 1
i=1 Ai ) ([i=1 Bi ) [1
i=1 (Ai Bi ):

Example 1.1.1. Let be a …nite positive measure on R: We claim that

to each set E 2 R and " > 0; there exists a set A; which is …nite union of
intervals (that is, A belongs to the Riemann algebra R0 ), such that

(E A) < ":

To see this let S be the class of all sets E 2 R for which the conclusion
is true. Clearly 2 S and, moreover, R0 S: If A 2 R0 , Ac 2 R0 and
therefore E c 2 S if E 2 S: Now suppose Ei 2 S; i 2 N+ : Then to each " > 0
and i there is a set Ai 2 R0 such that (Ei Ai ) < 2 i ": If we set

E = [1
i=1 Ei

then
(E ([1
i=1 Ai ))
1
i=1 (Ei Ai ) < ":
Here
E ([1 1 c c 1
i=1 Ai ) = fE \ (\i=1 Ai )g [ fE \ ([i=1 Ai )g

and Theorem 1.1.2 (f) gives that

(fE \ (\ni=1 Aci )g [ f(E c \ ([1

i=1 Ai )g) < "

if n is large enough (hint: \i2I (Di [ F ) = (\i2I Di ) [ F ): But then

(E [ni=1 Ai ) = (fE \ (\ni=1 Aci )g [ fE c \ ([ni=1 Ai )g) < "

 15

if n is large enough we conclude that the set E 2 S: Thus S is a -algebra

and since R0 S R it follows that S = R:

Exercises

1. Prove that the sets N N = f(i; j); i; j 2 Ng and Q are denumerable.

2. Suppose A is an algebra of subsets of X and and two contents on A

such that and (X) = (X) < 1: Prove that = :

3. Suppose A is an algebra of subsets of X and a content on A with

(X) < 1: Show that

(A [ B [ C) = (A) + (B) + (C)

(A \ B) (A \ C) (B \ C) + (A \ B \ C):

4. A collection C of subsets of X is an algebra with the following property:

If An 2 C; n 2 N+ and Ak \ An = if k 6= n, then [1 n=1 An 2 C. Prove that
C is a -algebra.

5. Let (X; M) be a measurable space and ( k )1k=1 a sequence of positive

measures on M such that 1 2 3 ::: . Prove that the set function

(A) = lim k (A); A2M

k!1

is a positive measure.

6. Let (X; M; ) be a positive measure space. Show that

q
n n n
(\k=1 Ak ) k=1 (Ak )
16

for all A1 ; :::; An 2 M:

7. Let (X; M; ) be a -…nite measure space with (X) = 1: Show that for
any r 2 [0; 1[ there is some A 2 M with r < (A) < 1:

8. Show that the symmetric di¤erence of sets is associative:

A (B C) = (A B) C:

9. (X; M; ) is a …nite positive measure space. Prove that

j (A) (B) j (A B):

10. Let E = 2N: Prove that

cN (E A) = 1

if A is a …nite union of intervals.

11. Suppose (X; P(X); ) is a …nite positive measure space such that (fxg) >
0 for every x 2 X: Set

d(A; B) = (A B); A; B 2 P(X):

Prove that
d(A; B) = 0 , A = B;
d(A; B) = d(B; A)
and
d(A; B) d(A; C) + d(C; B):

12. Let (X; M; ) be a …nite positive measure space. Prove that

([ni=1 Ai ) n
i=1 (Ai ) 1 i<j n (Ai \ Aj )
 17

for all A1 ; :::; An 2 M and integers n 2:

1.2. Measure Determining Classes

Suppose and are probability measures de…ned on the same -algebra M,

which is generated by a class E: If and agree on E; is it then true that
and agree on M? The answer is in general no. To show this, let

X = f1; 2; 3; 4g

and
E = ff1; 2g ; f1; 3gg :
Then (E) = P(X): If = 41 cX and

1 1 1 1
= X;1 + X;2 + X;3 + X;4
6 3 3 6
then = on E and 6= :
In this section we will prove a basic result on measure determining classes
for -…nite measures. In this context we will introduce so called -systems
and -systems, which will also be of great value later in connection with the
construction of so called product measures in Chapter 3.

De…nition 1.2.1. A class G of subsets of X is a -system if A \ B 2 G

for all A; B 2 G:

The class of all open n-cells in Rn is a -system.

De…nition 1.2.2. A class D of subsets of X is a -system if the following

properties hold:
(a) X 2 D:
(b) If A; B 2 D and A B; then B n A 2 D:
18

(c) If (An )n2N+ is a disjoint denumerable collection of members of the

class D; then [1n=1 An 2 D:

Theorem 1.2.1. If a class M is both a -system and -system, then M is

a -algebra.

PROOF. If A 2 M; then Ac = X n A 2 M since X 2 M and M is a

-system. Moreover, if (An )n2N+ is a denumerable collection of members of
M;
A1 [ ::: [ An = (Ac1 \ ::: \ Acn )c 2 M
for each n; since M is a -system and a -system. Let (Bn )1 n=1 be the
disjungation of (An )1
n=1 : Then (B )
n n2N+ is a disjoint denumerable collection
of members of M and De…nition 1.2.2(c) implies that [n=1 An = [1
1
n=1 Bn 2
M:

Theorem 1.2.2. Let G be a -system and D a -system such that G

D: Then (G) D:

PROOF. Let M be the intersection of all -systems containing G: The class

M is a -system and G M D. In view of Theorem 1.2.1 M is a -
algebra, if M is a -system and in that case (G) M: Thus the theorem
follows if we show that M is a -system.
Given C X; denote by DC be the class of all D X such that D \ C 2
M.

CLAIM 1. If C 2 M; then DC is a -system.

PROOF OF CLAIM 1. First X 2 DC since X \ C = C 2 M: Moreover, if

A; B 2 DC and A B; then A \ C; B \ C 2 M and
(B n A) \ C = (B \ C) n (A \ C) 2 M:
Accordingly from this, BnA 2 DC : Finally, if (An )n2N+ is a disjoint denumer-
able collection of members of DC , then (An \ C)n2N+ is disjoint denumerable
collection of members of M and
([n2N+ An ) \ C = [n2N+ (An \ C) 2 M:
 19

Thus [n2N+ An 2 DC :

CLAIM 2. If A 2 G; then M DA :

PROOF OF CLAIM 2. If B 2 G; A \ B 2 G M: Thus B 2 DA : We

have proved that G DA and remembering that M is the intersection of all
-systems containing G Claim 2 follows since DA is a -system.

To complete the proof of Theorem 1.2.2, observe that B 2 DA if and only

if A 2 DB : By Claim 2, if A 2 G and B 2 M; then B 2 DA that is A 2 DB :
Thus G DB if B 2 M. Now the de…nition of M implies that M DB if
B 2 M: The proof is almost …nished. In fact, if A; B 2 M then A 2 DB
that is A \ B 2 M: Theorem 1.2.2 now follows from Theorem 1.2.1.

Theorem 1.2.3. Let and be positive measures on M = (G), where

G is a -system, and suppose (A) = (A) for every A 2 G:
(a) If and are probability measures, then = :
(b) Suppose there exist En 2 G; n 2 N+ ; such that X = [1
n=1 En ;
E1 E2 :::; and

(En ) = (En ) < 1; all n 2 N+ :

Then = :

PROOF. (a) Let

D = fA 2 M; (A) = (A)g :
It is immediate that D is a -system and Theorem 1.2.2 implies that M =
(G) D since G D and G is a -system.

(b) If (En ) = (En ) = 0 for all all n 2 N+ , then

(X) = lim (En ) = 0

n!1
20

and, in a similar way, (X) = 0: Thus = : If (En ) = (En ) > 0; set

1 1
n (A) = (A \ En ) and n (A) = (A \ En )
(En ) (En )
for each A 2 M: By Part (a) n = n and we get

(A \ En ) = (A \ En )

for each A 2 M: Theorem 1.1.2(e) now proves that = :

Theorem 1.2.3 implies that there is at most one positive measure de…ned
on Rn such that the measure of any open n-cell in Rn equals its volume.
Next suppose f : X ! Y and let A X and B Y: The image of A
and the inverse image of B are

f (A) = fy; y = f (x) for some x 2 Ag

and
1
f (B) = fx; f (x) 2 Bg
respectively. Note that
1
f (Y ) = X
and
1 1
f (Y n B) = X n f (B):
Moreover, if (Ai )i2I is a collection of subsets of X and (Bi )i2I is a collection
of subsets of Y

f ([i2I Ai ) = [i2I f (Ai )

and
1 1
f ([i2I Bi ) = [i2I f (Bi ):
Given a class E of subsets of Y; set
1 1
f (E) = f (B); B 2 E :
1
If (Y; N ) is a measurable space, it follows that the class f (N ) is a -algebra
in X: If (X; M) is a measurable space
1
B 2 P(Y ); f (B) 2 M
 21

is a -algebra in Y . Thus, given a class E of subsets of Y;

1 1
(f (E)) = f ( (E)):

De…nition 1.2.3. Let (X; M) and (Y; N ) be measurable spaces. The func-
tion f : X ! Y is said to be (M; N )-measurable if f 1 (N ) M. If we say
that f : (X; M) ! (Y; N ) is measurable this means that f : X ! Y is an
(M; N )-measurable function.

Theorem 1.2.4. Let (X; M) and (Y; N ) be measurable spaces and suppose
E generates N : The function f : X ! Y is (M; N )-measurable if
1
f (E) M:

PROOF. The assumptions yield

1
(f (E)) M:

Since
1 1 1
(f (E)) = f ( (E)) = f (N )
we are done.

Corollary 1.2.1. A function f : X ! R is (M; R)-measurable if and only

if the set f 1 (] ; 1[) 2 M for all 2 R:

If f : X ! Y is (M; N )-measurable and is a positive measure on M,

the equation
(B) = (f 1 (B)), B 2 N
de…nes a positive measure on N : We will write = f 1 ; = f ( ) or
= f : The measure is called the image measure of under f and f is
said to transport to : Two (M; N )-measurable functions f : X ! Y and
g : X ! Y are said to be -equimeasurable if f ( ) = g( ):
22

As an example, let a 2 Rn and de…ne f (x) = x + a if x 2 Rn : If B Rn ;

1
f (B) = fx; x + a 2 Bg = B a:

Thus f 1 (B) is an open n-cell if B is, and Theorem 1.2.4 proves that f is
(Rn ; Rn )-measurable. Now, granted the existence of volume measure vn ; for
every B 2 Rn de…ne

(B) = f (vn )(B) = vn (B a):

Then (B) = vn (B) if B is an open n-cell and Theorem 1.2.3 implies that
= vn : We have thus proved the following

Theorem 1.2.5. For any A 2 Rn and x 2 Rn

A + x 2 Rn

and
vn (A + x) = vn (A):

Suppose ( ; F; P ) is a probability space. A measurable function de…ned

on is called a random variable and the image measure P is called the
probability law of : We sometimes write

L( ) = P :

Here are two simple examples.

If the range of a random variable consists of n points S = fs1 ; :::; sn g
(n 1) and P = n1 cS ; is said to have a uniform distribution in S. Note
that
1 n
P = s :
n k=1 k
Suppose > 0 is a constant. If a random variable has its range in N
and n
P = 1 n=0 e n
n!
then is said to have a Poisson distribution with parameter :
 23

Exercises

1. Let f : X ! Y , A X; and B Y: Show that

1 1
f (f (B)) B and f (f (A)) A:

2. Let (X; M) be a measurable space and suppose A X: Show that the

function A is (M; R)-measurable if and only if A 2 M:

3. Suppose (X; M) is a measurable space and fn : X ! R; n 2 N; a

sequence of (M; R)-measurable functions such that

lim fn (x) exists and = f (x) 2 R

n!1

for each x 2 X: Prove that f is (M; R)-measurable.

4. Suppose f : (X; M) ! (Y; N ) and g : (Y; N ) ! (Z; S) are measurable.

Prove that g f is (M; S)-measurable.

5. Granted the existence of volume measure vn , show that vn (rA) = rn vn (A)

if r 0 and A 2 R:

6. Let be the counting measure on Z2 and f (x; y) = x; (x; y) 2 Z2 : The

measure is -…nite. Prove that the image measure f ( ) is not -…nite.

7. Let ; : R ! [0; 1] be two positive measures such that (I) = (I) < 1
for each open subinterval of R: Prove that = :
24

8. Suppose has a Poisson distribution with parameter : Show that P [2N] =

e cosh :

9. Find a -system which is not a -algebra.

1.3. Lebesgue Measure

Once the problem about the existence of volume measure is solved the exis-
tence of the so called Lebesgue measure is simple to establish as will be seen
in this section. We start with some concepts of general interest.
If (X; M; ) is a positive measure space, the zero set Z of is, by
de…nition, the set at all A 2 M such that (A) = 0: An element of Z is
called a null set or -null set. If

(A 2 Z and B A) ) B 2 M

the measure space (X; M; ) is said to be complete. In this case the measure
is also said to be complete. The positive measure space (X; f ; Xg ; );
where X = f0; 1g and = 0; is not complete since X 2 Z and f0g 2 = f ; Xg :

Theorem 1.3.1 If (En )1

n=1 is a denumerable collection of members of Z
1
then [n=1 En 2 Z :

PROOF We have

0 ([1
n=1 En )
1
n=1 (En ) = 0

which proves the result.

Granted the existence of linear measure v1 it follows from Theorem 1.3.1

that Q 2 Zv1 since Q is countable and fag 2 Zv1 for each real number a.
Suppose (X; M; ) is an arbitrary positive measure space. It turns out
that is the restriction to M of a complete measure. To see this suppose
M is the class of all E X is such that there exist sets A; B 2 M such that
 25

A E B and B n A 2 Z : It is obvious that X 2 M since M M : If

E 2 M ; choose A; B 2 M such that A E B and B n A 2 Z : Then
c c
B E A and A n B = B n A 2 Z and we conclude that E c 2 M : If
c c c
1
(Ei )i=1 is a denumerable collection of members of M ; for each i there exist
sets Ai ; Bi 2 M such that Ai E Bi and Bi n Ai 2 Z : But then
[1
i=1 Ai [1
i=1 Ei [1
i=1 Bi

where [1 1 1 1
i=1 Ai ; [i=1 Bi 2 M. Moreover, ([i=1 Bi ) n ([i=1 Ai ) 2 Z since

([1 1
i=1 Bi ) n ([i=1 Ai ) [1
i=1 (Bi n Ai ):

Thus [1
i=1 Ei 2 M and M is a -algebra.
If E 2 M; suppose Ai ; Bi 2 M are such that Ai E Bi and Bi n Ai 2
Z for i = 1; 2: Then for each i; (B1 \ B2 ) n Ai 2 Z and
(B1 \ B2 ) = ((B1 \ B2 ) n Ai ) + (Ai ) = (Ai ):
Thus the real numbers (A1 ) and (A2 ) are the same and we de…ne (E) to
be equal to this common number. Note also that (B1 ) = (E): It is plain
that ( ) = 0: If (Ei )1
i=1 is a disjoint denumerable collection of members
of M; for each i there exist sets Ai ; Bi 2 M such that Ai Ei Bi and
Bi n Ai 2 Z : From the above it follows that
([1 1
i=1 Ei ) = ([i=1 Ai ) =
1
n=1 (Ai ) = 1
n=1 (Ei ):
We have proved that is a positive measure on M . If E 2 Z the
de…nition of shows that any set A E belongs to the -algebra M : It
follows that the measure is complete and its restriction to M equals :
The measure is called the completion of and M is called the com-
pletion of M with respect to :

De…nition 1.3.1 The completion of volume measure vn on Rn is called

Lebesgue measure on Rn and is denoted by mn : The completion of Rn with
respect to vn is called the Lebesgue -algebra in Rn and is denoted by Rn :
A member of the class Rn is called a Lebesgue measurable set in Rn or a
Lebesgue set in Rn : A function f : Rn ! R is said to be Lebesgue measurable
if it is (Rn ; R)-measurable. Below, m1 is written m if this notation will not
lead to misunderstanding. Furthermore, R1 is written R .
26

Theorem 1.3.2. Suppose E 2 Rn and x 2Rn : Then E + x 2 Rn and

mn (E + x) = mn (E):

PROOF. Choose A; B 2 Rn such that A E B and B n A 2 Zvn : Then,

by Theorem 1.2.5, A + x; B + x 2 Rn ; vn (A + x) = vn (A) = mn (E); and
(A + x) n (B + x) = (A n B) + x 2 Zvn : Since A + x E + x B + x the
theorem is proved.

The Lebesgue -algebra in Rn is very large and contains each set of

interest in analysis and probability. In fact, in most cases, the -algebra Rn is
su¢ ciently large but there are some exceptions. For example, if f : Rn ! Rn
is continuous and A 2 Rn , the image set f (A) need not belong to the class
Rn (see e.g. the Dudley book [D]). To prove the existence of a subset of the
real line, which is not Lebesgue measurable we will use the so called Axiom
of Choice.

Axiom of Choice. If (Ai )i2I is a non-empty collection of non-empty sets,

there exists a function f : I ! [i2I Ai such that f (i) 2 Ai for every i 2 I:

Let X and Y be sets. The set of all ordered pairs (x; y); where x 2 X
and y 2 Y is denoted by X Y: An arbitrary subset R of X Y is called a
relation. If (x; y) 2 R , we write x s y: A relation is said to be an equivalence
relation on X if X = Y and

(i) x s x (re‡exivity)
(ii) x s y ) y s x (symmetry)
(ii) (x s y and y s z) ) x s z (transitivity)

The equivalence class R(x) =def fy; y s xg : The de…nition of the equiv-
alence relation s implies the following:

(a) x 2 R(x)
 27

(b) R(x) \ R(y) 6= ) R(x) = R(y)

(c) [x2X R(x) = X:

An equivalence relation leads to a partition of X into a disjoint collection

of subsets of X:
1 1
Let X = ; and de…ne an equivalence relation for numbers x; y in X
2 2
by stating that x s y if x y is a rational number. By the Axiom of Choice
it is possible to pick exactly one element from each equivalence class. Thus
there exists a subset N L of X which contains exactly one element from each
equivalence class.
If we assume that N L 2 R we get a contradiction as follows. Let (ri )1 i=1
be an enumeration of the rational numbers in [ 1; 1]. Then

X [1
i=1 (ri + N L)

and it follows from Theorem 1.3.1 that ri + N L 2 = Zm for some i: Thus, by

Theorem 1.3.2, N L 2 = Zm :
Now assume (ri + N L) \ (rj + N L) 6= : Then there exist a0 ; a00 2 N L
such that ri + a0 = rj + a00 or a0 a00 = rj ri : Hence a0 s a00 and it follows
that a0 and a00 belong to the same equivalence class. But then a0 = a00 : Thus
ri = rj and we conclude that (ri + N L)i2N+ is a disjoint enumeration of
Lebesgue sets. Now, since

3 3
[1
i=1 (ri + N L) ;
2 2

it follows that
3 m([1
i=1 (ri + N L)) =
1
n=1 m(N L):

But then N L 2 Zm ; which is a contradiction. Thus N L 2

=R :

In the early 1970’Solovay [S] proved that it is consistent with the usual
axioms of Set Theory, excluding the Axiom of Choice, that every subset of
R is Lebesgue measurable.
From the above we conclude that the Axiom of Choice implies the exis-
tence of a subset of the set of real numbers which does not belong to the class
R: Interestingly enough, such an example can be given without any use of
28

the Axiom of Choice and follows naturally from the theory of analytic sets.
The interested reader may consult the Dudley book [D] :

Exercises

1. (X; M; ) is a positive measure space. Prove or disprove: If A E B

and (A) = (B) then E belongs to the domain of the completion :

2. Prove or disprove: If A and B are not Lebesgue measurable subsets of

R; then A [ B is not Lebesgue measurable.

3. Let (X; M; ) be a complete positive measure space and suppose A; B 2

M, where B n A is a -null set. Prove that E 2 M if A E B (stated
otherwise M = M).

4. Suppose E R and E 2
= R . Show there is an " > 0 such that

m(B n A) "

for all A; B 2 R such that A E B:

5. Suppose (X; M; ) is a positive measure space and (Y; N ) a measurable

space. Furthermore, suppose f : X ! Y is (M; N )-measurable and let
= f 1 ; that is (B) = (f 1 (B)); B 2 N : Show that f is (M ; N )-
measurable, where M denotes the completion of M with respect to and
N the completion of N with respect to :

1.4. Carathéodory’s Theorem

 29

In these notes we exhibit two famous approaches to Lebesgue measure: One

is based on the Carathéodory Theorem, which we present in this section, and
the other one, due to F. Riesz, is a representation theorem of positive linear
functionals on spaces of continuous functions in terms of positive measures.
The latter approach, is presented in Chapter 3. Both methods depend on
topological concepts such as compactness.

De…nition 1.4.1. A function : P(X) ! [0; 1] is said to be an outer

measure if the following properties are satis…ed:

(i) ( ) = 0:
(ii) (A) (B) if A B:
(iii) for any denumerable collection (An )1
n=1 of subsets of X

([1
n=1 An )
1
n=1 (An ):

Since
E = (E \ A) [ (E \ Ac )
an outer measure satis…es the inequality

(E) (E \ A) + (E \ Ac ):

If is an outer measure on X we de…ne M( ) as the set of all A X

such that
(E) = (E \ A) + (E \ Ac ) for all E X
or, what amounts to the same thing,

(E) (E \ A) + (E \ Ac ) for all E X:

The next theorem is one of the most important in measure theory.

30

Theorem 1.4.1. (Carathéodory’s Theorem) Suppose is an outer

measure. The class M( ) is a -algebra and the restriction of to M( ) is
a complete measure.

PROOF. Clearly, 2 M( ) and Ac 2 M( ) if A 2 M( ): Moreover, if

A; B 2 M( ) and E X;

(E) = (E \ A) + (E \ Ac )

= (E \ A \ B) + (E \ A \ B c )
+ (E \ Ac \ B) + (E \ Ac \ B c ):
But
A [ B = (A \ B) [ (A \ B c ) [ (Ac \ B)
and
Ac \ B c = (A [ B)c
and we get
(E) (E \ (A [ B)) + (E \ (A [ B)c ):
It follows that A [ B 2 M( ) and we have proved that the class M( ) is an
algebra. Now if A; B 2 M( ) are disjoint

(A [ B) = ((A [ B) \ A) + ((A [ B) \ Ac ) = (A) + (B)

and therefore the restriction of to M( ) is a content.

Next we prove that M( ) is a -algebra. Let (Ai )1i=1 be a disjoint denu-
merable collection of members of M( ) and set for each n 2 N

Bn = [1 i n Ai and B = [1
i=1 Ai

(here B0 = ). Then for any E X;

(E \ Bn ) = (E \ Bn \ An ) + (E \ Bn \ Acn )

= (E \ An ) + (E \ Bn 1 )
and, by induction,
n
(E \ Bn ) = i=1 (E \ Ai ):
But then
(E) = (E \ Bn ) + (E \ Bnc )
 31

n
i=1 (E \ Ai ) + (E \ B c )
and letting n ! 1;
1
(E) i=1 (E \ Ai ) + (E \ B c )

([1 c
i=1 (E \ Ai )) + (E \ B )

= (E \ B) + (E \ B c ) (E):
All the inequalities in the last calculation must be equalities and we conclude
that B 2 M( ) and, choosing E = B; results in
1
(B) = i=1 (Ai ):

Thus M( ) is a -algebra and the restriction of to M( ) is a positive

measure.
Finally we prove that the the restriction of to M( ) is a complete
measure. Suppose B A 2 M( ) and (A) = 0: If E X;

(E) (E \ B) + (E \ B c ) (E \ B c ) (E)

and so B 2 M( ): The theorem is proved.

Exercises

1. Suppose i : P(X) ! [0; 1[ ; i = 1; 2; are outer measures. Prove that

= max( 1 ; 2 ) is an outer measure.

2. Suppose a; b 2 R and a 6= b: Set = max( a ; b ): Prove that

fag ; fbg 2
= M( ):

1.5. Existence of Linear Measure

32

The purpose of this section is to show the existence of linear measure on R

using the Carathéodory Theorem and a minimum of topology.
First let us recall the de…nition of in…mum and supremum of a non-
empty subset of the extended real line. Suppose A is a non-empty subset
of [ 1; 1] = R[ f 1; 1g : We de…ne 1 x and x 1 for all x 2
[ 1; 1] : An element b 2 [ 1; 1] is called a majorant of A if x b for all
x 2 A and a minorant if x b for all x 2 A: The Supremum Axiom states
that A possesses a least majorant, which is denoted by sup A. From this
follows that if A is non-empty, then A possesses a greatest minorant, which
is denoted by inf A. (Actually, the Supremum Axiom is a theorem in courses
where time is spent on the de…nition of real numbers.)

Theorem 1.5.1. (The Heine-Borel Theorem; weak form) Let [a; b] be

a closed bounded interval and (Ui )i2I a collection of open sets such that

[i2I Ui [a; b] :

Then
[i2J Ui [a; b]
for some …nite subset J of I:

PROOF. Let A be the set of all x 2 [a; b] such that

[i2J Ui [a; x]

for some …nite subset J of I: Clearly, a 2 A since a 2 Ui for some i: Let

c = sup A: There exists an i0 such that c 2 Ui0 : Let c 2 ]a0 ; b0 [ Ui0 ; where
a0 < b0 : Furthermore, by the very de…nition of least upper bound, there
exists a …nite set J such that

[i2J Ui [a; (a0 + c)=2] :

Hence
[i2J[fi0 g Uk [a; (c + b0 )=2]
and it follows that c 2 A and c = b. The lemma is proved.
 33

A subset K of R is called compact if for every family of open subsets Ui ;

i 2 I; with [i2I Ui K we have [i2J Ui K for some …nite subset J of I:
The Heine-Borel Theorem shows that a closed bounded interval is compact.
If x; y 2 R and E; F R; let

d(x; y) =j x yj

be the distance between x and y; let

d(x; E) = inf d(x; u)

u2E

be the distance from x to E; and let

d(E; F ) = inf d(u; v)

u2E;v2F

be the distance between E and F (here the in…mum of the emty set equals
1): Note that for any u 2 E;

d(x; u) d(x; y) + d(y; u)

and, hence
d(x; E) d(x; y) + d(y; u)
and
d(x; E) d(x; y) + d(y; E):
By interchanging the roles of x and y and assuming that E 6= ; we get

j d(x; E) d(y; E) j d(x; y):

Note that if F R is closed and x 2

= F; then d(x; F ) > 0:
An outer measure : P(R) ! [0; 1] is called a metric outer measure if

(A [ B) = (A) + (B)

for all A; B 2 P(R) such that d(A; B) > 0:

Theorem 1.5.2. If : P(R) ! [0; 1] is a metric outer measure, then

R M( ):
34

PROOF. Let F 2 P(R) be closed. It is enough to show that F 2 M( ): To

this end we choose E X with (E) < 1 and prove that

(E) (E \ F ) + (E \ F c ):

Let n 1 be an integer and de…ne

1
An = x 2 E \ F c ; d(x; F ) :
n

Note that An An+1 and

E \ F c = [1
n=1 An :

Moreover, since is a metric outer measure

(E) ((E \ F ) [ An ) = (E \ F ) + (An )

and, hence, proving

(E \ F c ) = lim (An )
n!1

we are done.
Let Bn = An+1 \ Acn : It is readily seen that

1
d(Bn+1 ; An )
n(n + 1)

since if x 2 Bn+1 and

1
d(x; y) <
n(n + 1)
then
1 1 1
d(y; F ) d(y; x) + d(x; F ) < + = :
n(n + 1) n + 1 n
Now
(A2k+1 ) (B2k [ A2k 1 ) = (B2k ) + (A2k 1 )
k
::: i=1 (B2i )
and in a similar way
k
(A2k ) i=1 (B2i 1 ):
 35

But (An ) (E) < 1 and we conclude that

1
i=1 (Bi ) < 1:

We now use that

E \ F c = An [ ([1
i=n Bi )

to obtain
(E \ F c ) (An ) + 1
i=n (Bi ):
Now, since (E \ F c ) (An ),

(E \ F c ) = lim (An )
n!1

and the theorem is proved.

PROOF OF THEOREM 1.1.1 IN ONE DIMENSION. Suppose > 0: If

A R; de…ne

1
(A) = inf k=1 l(Ik )

the in…mum being taken over all open intervals Ik with l(Ik ) < such that

A [1
k=1 Ik :

Obviously, ( ) = 0 and (A) (B) if A B: Suppose (An )1

n=1 is a
denumerable collection of subsets of R and let " > 0: For each n there exist
intervals Ikn ; k 2 N+ ; such that l(Ikn ) < ;

An [1
k=1 Ikn

and
1 n
k=1 l(Ikn ) (An ) + "2 :
Then
A =def [1
n=1 An [1
k;n=1 Ikn

and
1 1
k;n=1 l(Ikn ) n=1 (An ) + ":
Thus
1
(A) n=1 (An ) + "
36

and, since " > 0 is arbitrary,

1
(A) n=1 (An ):

It follows that is an outer measure.

If I is an open interval it is simple to see that

(I) l(I):

To prove the reverse inequality, choose a closed bounded interval J I: Now,

if
I [1 k=1 Ik

where each Ik is an open interval of l(Ik ) < ; it follows from the Heine-Borel
Theorem that
J [nk=1 Ik
for some n: Hence
n 1
l(J) k=1 l(Ik ) k=1 l(Ik )

and it follows that

l(J) (I)
and, accordingly from this,
l(I) (I):
Thus, if I is an open interval, then

(I) = l(I).

Note that 1 2 if 0 < 1 2: We de…ne

0 (A) = lim (A) if A R:

!0

It obvious that 0 is an outer measure such that 0 (I) =l(I); if I is an open

interval.
To complete the proof we show that 0 is a metric outer measure. To this
end let A; B R and d(A; B) > 0: Suppose 0 < < d(A; B) and

A[B [1
k=1 Ik

where each Ik is an open interval with l(Ik ) < : Let

= fk; Ik \ A 6= g
 37

and
= fk; Ik \ B 6= g :
Then \ = ;
A [k2 Ik
and
B [k2 Ik
and it follows that
1
k=1 l(Ik ) k2 l(Ik ) + k2 l(Ik )

(A) + (B):
Thus
(A [ B) (A) + (B)
and by letting ! 0 we have

0 (A [ B) 0 (A) + 0 (B)

and
0 (A [ B) = 0 (A) + 0 (B):

Finally by applying the Carathéodory Theorem and Theorem 1.5.2 it

follows that the restriction of 0 to R equals v1 .

We end this section with some additional results of great interest.

Theorem 1.5.3. For any > 0; = 0: Moreover, if A R

1
0 (A) = inf k=1 l(Ik )

the in…mum being taken over all open intervals Ik ; k 2 N+ ; such that
[1k=1 Ik A:
38

PROOF. It follows from the de…nition of 0 that 0 : To prove the

reverse inequality let A R and choose open intervals Ik ; k 2 N+ ; such that
[1k=1 Ik A: Then
1 1
0 (A) 0 ([k=1 Ik ) k=1 0 (Ik )

1
= k=1 l(Ik ):

Hence
1
0 (A) inf k=1 l(Ik )

the in…mum being taken over all open intervals Ik ; k 2 N+ ; such that
[1k=1 Ik A: Thus 0 (A) (A); which completes the proof of Theorem
1.5.3.

Theorem 1.5.4. If A R;

0 (A) = inf 0 (U ):
U A
U open

Moreover, if A 2 M( 0 );

0 (A) = sup 0 (K):

K A
K closed bounded

PROOF. If A U, 0 (A) 0 (U ): Hence

0 (A) inf 0 (U ):
U A
U open

Next let " > 0 be …xed and choose open intervals Ik ; k 2 N+ ; such that
[1
k=1 Ik A and
1
k=1 l(Ik ) 0 (A) + "

(here observe that it may happen that 0 (A) = 1). Then the set U =def
[1k=1 Ik is open and

1 1
0 (U ) k=1 0 (Ik ) = k=1 l(Ik ) 0 (A) + ":

Thus
inf 0 (U ) 0 (A)
U A
U open
 39

and we have proved that

0 (A) = inf 0 (U ):
U A
U open

If K A; 0 (K) 0 (A) and, accordingly from this,

sup 0 (K) 0 (A):

K A
K closed bounded

To prove the reverse inequality we …rst assume that A 2 M( 0 ) is bounded.

Let " > 0 be …xed and suppose J is a closed bounded interval containing A:
Then we know from the …rst part of Theorem 1.5.4 already proved that there
exists an open set U J r A such that

0 (U ) < 0 (J r A) + ":

But then

0 (J) 0 (J r U) + 0 (U ) < 0 (J r U) + 0 (J r A) + "

and it follows that

0 (A) "< 0 (J n U ):
Since J r U is a closed bounded set contained in A we conclude that

0 (A) sup 0 (K):

K A
K closed bounded

If A 2 M( 0 ) let An = A \ [ n; n] ; n 2 N+ : Then given " > 0 and n 2

N+ ; let Kn be a closed bounded subset of An such that 0 (Kn ) > 0 (An ) ":
Clearly, there is no loss of generality to assume that K1 K2 K3 :::
and by letting n tend to plus in…nity we get

lim 0 (Kn ) 0 (A) ":

n!1

Hence
0 (A) = sup 0 (K):
K A
K compact

and Theorem 1.5.4 is completely proved.

40

Theorem 1.5.5. Lebesgue measure m1 equals the restriction of 0 to M( 0 ):

PROOF. Recall that linear measure v1 equals the restriction of 0 to R and

m1 = v1 : First suppose E 2 R and choose A; B 2 R such that A E B
and B rA 2 Zv1 : But then 0 (E rA) = 0 and E = A[(E rA) 2 M( 0 ) since
the Carathéodory Theorem gives us a complete measure. Hence m1 (E) =
v1 (A) = 0 (E).
Conversely suppose E 2 M( 0 ): We will prove that E 2 R and m1 (E) =
0 (E). First assume that E is bounded. Then for each positive integer n there
exist open Un E and closed bounded Kn E such that
n
0 (Un ) < 0 (E) +2
and

n
0 (Kn ) > 0 (E) 2 :
The de…nitions yield A = [1 1
1 Kn ; B = \1 Un 2 R and

0 (E) = 0 (A) = 0 (B) = v1 (A) = v1 (B) = m1 (E):

It follows that E 2 R and 0 (E) = m1 (E):
In the general case set En = E \ [ n; n] ; n 2 N+ : Then from the above
En 2 R and 0 (En ) = m1 (En ) for each n and Theorem 1.5.5 follows by
letting n go to in…nity.

The Carathéodory Theorem can be used to show the existence of volume

measure on Rn but we do not go into this here since its existence follows by
several other means below. By passing, let us note that the Carathéodory
Theorem is very e¢ cient to prove the existence of so called Haussdor¤ mea-
sures (see e.g. [F ]); which are of great interest in Geometric Measure Theory.

Exercises

1. Prove that a subset K of R is compact if and only if K is closed and

bounded.
 41

2. Suppose A 2 R and m(A) < 1: Set f (x) = m(A \ ] 1; x]); x 2 R:

Prove that f is continuous.

3. Suppose A 2 Zm and B = fx3 ; x 2 Ag : Prove that B 2 Zm :

4. Let A be the set of all real numbers x such that

p 1
jx j
q q3
for in…nitely many pairs of positive integers p and q: Prove that A 2 Zm :

5. Let I1 ; :::; In be open subintervals of R such that

Q\ [0; 1] [nk=1 Ik :
n
Prove that k=1 m(Ik ) 1:

6. If E 2 R and m(E) > 0; for every 2 ]0; 1[ there is an interval I such

that m(E \ I) > m(I). (Hint: m(E) = inf 1 k=1 m(Ik ), where the in…mum
1
is taken over all intervals such that [k=1 Ik E:)

7. If E 2 R and m(E) > 0; then the set E E = fx y; x; y 2 Eg contains

an open non-empty interval centred at 0:(Hint: Take an interval I with
m(E\I) 34 m(I): Set " = 12 m(I): If j x j "; then (E\I)\(x+(E\I)) 6= :)

1
8. Let be the restriction of the positive measure k=1 R; k1 to R: Prove that

inf (U ) > (A)

U A
U open

if A = f0g :
42

1.6. Positive Measures Induced by Increasing Right Continuous

Functions

Suppose F : R ! [0; 1[ is a right continuous increasing function such that

lim F (x) = 0:
x! 1

Set
L = lim F (x):
x!1

We will prove that there exists a unique positive measure : R ! [0; L] such
that
(] 1; x]) = F (x); x 2 R:
The special case L = 0 is trivial so let us assume L > 0 and introduce

H(y) = inf fx 2 R; F (x) yg ; 0 < y < L:

The de…nition implies that the function H increases.

Suppose a is a …xed real number. We claim that

fy 2 ]0; L[ ; H(y) ag = ]0; F (a)] \ ]0; L[ :

To prove this …rst suppose that y 2 ]0; L[ and H(y) a: Then to each
n
positive integer n; there is an xn 2 [H(y); H(y) + 2 [ such that F (xn ) y:
Then xn ! H(y) as n ! 1 and we obtain that F (H(y)) y since F is right
continuous. Thus, remembering that F increases, F (a) y: On the other
hand, if 0 < y < L and 0 < y F (a); then, by the very de…nition of H(y);
H(y) a:
We now de…ne
= H(v1j]0;L[ )
and get
(] 1; x]) = F (x); x 2 R:
The uniqueness follows at once from Theorem 1.2.3. Note that the measure
is a probability measure if L = 1:
 43

Exercises

1. Suppose F : R ! R is a right continuous increasing function. Prove that

there is a unique positive measure on R such that

(]a; x]) = F (x) F (a); if a; x 2 R and a < x:

2. Suppose F : R ! R is an increasing function. Prove that the set of

all discontinuity points of F is at most denumerable. (Hint: Assume …rst
that F is bounded and prove that the set of all points x 2 R such that
F (x+) F (x ) > " is …nite for every " > 0.)

3. Suppose is a -…nite positive measure on R: Prove that the set of all

x 2 R such that (fxg) > 0 is at most denumerable.

4. Suppose is a -…nite positive measure on Rn : Prove that there is an at

most denumerable set of hyperplanes of the type

xk = c (k = 1; :::; n; c 2 R)

with positive -measure.

5. Construct an increasing function f : R ! R such that the set of discon-

tinuity points of f equals Q.
44

CHAPTER 2
INTEGRATION
Introduction

In this chapter Lebesgue integration in abstract positive measure spaces is

introduced. A series of famous theorems and lemmas will be proved.

2.1. Integration of Functions with Values in [0; 1]

Recall that [0; 1] = [0; 1[ [ f1g : A subinterval of [0; 1] is de…ned in the

natural way. We denote by R0;1 the -algebra generated by all subintervals
of [0; 1] : The class of all intervals of the type ] ; 1] ; 0 < 1; (or of
the type [ ; 1] ; 0 < 1) generates the -algebra R0;1 and we get the
following

Theorem 2.1.1. Let (X; M) be a measurable space and suppose f : X !

[0; 1] :
(a) The function f is (M; R0;1 )-measurable if f 1 (] ; 1]) 2 M for
every 0 < 1:
(b) The function f is (M; R0;1 )-measurable if f 1 ([ ; 1]) 2 M for
every 0 < 1:

Note that the set ff > g 2 M for all real if f is (M; R0;1 )-measurable.
If f; g : X ! [0; 1] are (M; R0;1 )-measurable, then min(f; g); max(f; g),
and f + g are (M; R0;1 )-measurable, since, for each 2 [0; 1[ ;

min(f; g) , (f and g )

max(f; g) , (f or g )
 45

and [
ff + g > g = (ff > qg \ fg > qg):
q2Q

Given functions fn : X ! [0; 1] ; n = 1; 2; :::; f = supn 1 fn is de…ned

by the equation
f (x) = sup ffn (x); n = 1; 2; :::g :
Note that
1
f (] ; 1]) = [1 1
n=1 fn (] ; 1])

for every real 0 and, accordingly from this, the function supn 1 fn is
(M; R0;1 )-measurable if each fn is (M; R0;1 )-measurable. Moreover, f =
inf n 1 fn is given by

f (x) = inf ffn (x); n = 1; 2; :::g :

Since
1
f ([0; [) = [1 1
n=1 fn ([0; [)

for every real 0 we conclude that the function f = inf n 1 fn is (M; R0;1 )-
measurable if each fn is (M; R0;1 )-measurable.
Below we write
fn " f
if fn ; n = 1; 2; :::; and f are functions from X into [0; 1] such that fn fn+1
for each n and fn (x) ! f (x) for each x 2 X as n ! 1:
An (M; R0;1 )-measurable function ' : X ! [0; 1] is called a simple
measurable function if '(X) is a …nite subset of [0; 1[ : If it is neccessary to
be more precise, we say that ' is a simple M-measurable function.

Theorem 2.1.2. Let f : X ! [0; 1] be (M; R0;1 )-measurable. There exist

simple measurable functions 'n ; n 2 N+ ; on X such that 'n " f :

PROOF. Given n 2 N+ , set

1 i 1 i
Ein = f ( ; n ); i 2 N+
2n 2
46

and
X
1
i 1
n = Ein +1 f 1 (f1g) :
i=1
2n
It is obvious that n f and that n n+1 : Now set 'n = min(n; n) and
we are done.

Let (X; M; ) be a positive measure space and ' : X ! [0; 1[ a simple

measurable function: If 1 ; :::; n are the distinct values of the simple function
', and if Ei = ' 1 (f i g); i = 1; :::; n; then
n
'= i=1 i Ei :

Furthermore, if A 2 M we de…ne
Z
(A) = 'd = nk=1 i (Ei \ A) = n
k=1 i
Ei
(A):
A

Clearly, is a positive measure since each term in the right side is a positive
measure as a function of A. Note that
Z Z
'd = 'd if 0 <1
A A

and Z
%d = a (A)
A

if a 2 [0; 1[ and % is a simple measurable function such that % = a on A:

If is another simple measurable function and ' ;
Z Z
'd d :
A A

1
To see this, let 1 ; :::; p be the distinct values of and Fj = ( j );
j = 1; :::; p: Now, putting Bij = Ei \ Fj ;
Z
'd = ([ij (A \ Bij ))
A
Z Z
= ij (A \ Bij ) = ij 'd = ij id
A\Bij A\Bij
 47
Z Z
ij jd = d :
A\Bij A

In a similar way one proves that

Z Z Z
(' + )d = 'd + d :
A A A

From the above it follows that

Z Z
n
' Ad = i=1 i Ei \A d
A A
Z
n n
= i=1 i Ei \A d = i=1 i (Ei \ A)
A
and Z Z
' Ad = 'd :
A A

If f : X ! [0; 1] is an (M; R0;1 )-measurable function and A 2 M, we

de…ne
Z Z
f d = sup 'd ; 0 ' f; ' simple measurable
A A
Z
= sup 'd ; 0 ' f; ' simple measurable and ' = 0 on Ac :
A

The left member in this equation is called the Lebesgue integral of f over A
with respect to the measure : Sometimes we also speek of the -integral of f
over A: The two de…nitions of the -integral of a simple measurable function
' : X ! [0; 1[ over A agree.
From now on in this section, an (M; R0;1 )-measurable function f : X !
[0; 1] is simply called measurable.
The following properties are immediate consequences of the de…nitions.
The functions and sets occurring in the equations are assumed to be mea-
surable.

R R
(a) If f; g 0 and f g on A; then A
fd A
gd :
48
R R
(b) A
fd = X Af d :

R R
(c) If f 0 and 2 [0; 1[, then A
fd = A
fd :

R
(d) A
f d = 0 if f = 0 and (A) = 1:

R
(e) A
f d = 0 if f = 1 and (A) = 0:

If f : X ! [0; 1] is measurable and 0 < < 1; then f f 1 ([ ;1]) =

ff g and Z Z Z
fd ff gd = ff gd :
X X X

This proves the so called Markov Inequality

Z
1
(f ) fd
X

where we write (f ) instead of the more precise expression (ff g):

Example 2.1.1. Suppose f : X ! [0; 1] is measurable and

Z
f d < 1:
X

We claim that
1
ff = 1g = f (f1g) 2 Z :
To prove this we use the Markov Inequality and have
Z
1
(f = 1) (f ) fd
X
 49

for each 2 ]0; 1[ : Thus (f = 1) = 0:

Example 2.1.2. Suppose f : X ! [0; 1] is measurable and

Z
f d = 0:
X

We claim that
1
ff > 0g = f (]0; 1]) 2 Z :
To see this, note that

1 1
f (]0; 1]) = [1
n=1 f
1
( ;1 )
n
Furthermore, for every …xed n 2 N+ ; the Markov Inequality yields
Z
1
(f > ) n fd = 0
n X

and we get ff > 0g 2 Z since a countable union of null sets is a null set.

We now come to one of the most important results in the theory.

Theorem 2.1.3. (Monotone Convergence Theorem) Let fn : X !

[0; 1] , n = 1; 2; 3; ::::; be a sequence of measurable functions and suppose
that fn " f; that is 0 f1 f2 ::: and

fn (x) ! f (x) as n ! 1, for every x 2 X:

Then f is measurable and

Z Z
fn d ! f d as n ! 1:
X X

PROOF. The function f is measurable since f = supn 1 fn :

50
R R R
The inequalities fn fn+1 f yield X fn d f d
X n+1 X
f d and
we conclude that there exists an 2 [0; 1] such that
Z
fn d ! as n ! 1
X

and Z
fd :
X
To prove the reverse inequality, let ' be any simple measurable function
such that 0 ' f , let 0 < < 1 be a constant, and de…ne, for …xed
n 2 N+ ;
An = fx 2 X; fn (x) '(x)g :
If 1 ; :::; p are the distinct values of ';
An = [pk=1 (fx 2 X; fn (x) kg \ f' = k g)

and it follows that An is measurable. Clearly, A1 A2 ::: . Moreover, if

f (x) = 0; then x 2 A1 and if f (x) > 0; then '(x) < f (x) and x 2 An for
all su¢ ciently large n. Thus [1
n=1 An = X: Now
Z Z
fn d 'd
An An

and we get Z
'd
X
R
since the map A ! A
'd is a positive measure on M: By letting " 1,
Z
'd
X

and, hence Z
fd :
X
The theorem follows.

Theorem 2.1.4. (a) Let f; g : X ! [0; 1] be measurable functions. Then

Z Z Z
(f + g)d = fd + gd :
X X X
 51

(b) (Beppo Levi’s Theorem) If fk : X ! [0; 1] , k = 1; 2; ::: are mea-

surable, Z Z
1 1
k=1 fk d = k=1 fk d
X X

PROOF. (a) Let ('n )1 1

n=1 and ( n )n=1 be sequences of simple and measurable
functions such that 0 'n " f and 0 n " g: We proved above that
Z Z Z
('n + n )d = 'n d + nd
X X X

and, by letting n ! 1; Part (a) follows from the Monotone Convergence

Theorem.

(b) Part (a) and induction imply that

Z Z
n n
k=1 fk d = k=1 fk d
X X

and the result follows from monotone convergence.

Theorem 2.1.5. Suppose w : X ! [0; 1] is a measurable function and

de…ne Z
(A) = wd ; A 2 M:
A
Then is a positive measure and
Z Z
fd = f wd ; A 2 M
A A

for every measurable function f : X ! [0; 1] :

PROOF. Clearly, ( ) = 0. Suppose (Ek )1 k=1 is a disjoint denumerable col-

1
lection of members of M and set E = [k=1 Ek : Then
Z Z Z
1 1
([k=1 Ek ) = wd = E wd = k=1 Ek wd
E X X
52

where, by the Beppo Levi Theorem, the right member equals

Z Z
1 1
k=1 Ek wd = k=1 wd = 1 k=1 (Ek ):
X Ek

This proves that is a positive measure.

Let A 2 M. To prove the last part in Theorem 2.1.5 we introduce the
class C of all measurable functions f : X ! [0; 1] such that
Z Z
fd = f wd :
A A

The indicator function of a measurable set belongs to C and from this we

conclude that every simple measurable function belongs to C: Furthermore, if
fn 2 C; n 2 N; and fn " f ; the Monotone Convergence Theorem proves that
f 2 C: Thus in view of Theorem 2.1.2 the class C contains every measurable
function f : X ! [0; 1] : This completes the proof of Theorem 2.1.5.

The measure in Theorem 2.1.5 is written

=w

or
d = wd :
Let ( n )1
n=1 be a sequence in [ 1; 1] : First put k = inf f k ; k+1 ; k+2 ; :::g
and = sup f 1 ; 2 ; 3 ; ::g = limn!1 n : We call the lower limit of ( n )1n=1
and write
= lim inf n :
n!1

Note that
= lim n
n!1

if the limit exists. Now put k = sup f k ; k+1 ; k+2 ; :::g and = inf f 1; 2; 3 ; ::g =
1
limn!1 n : We call the upper limit of ( n )n=1 and write

= lim sup n:
n!1

Note that
= lim n
n!1
 53

if the limit exists.

Given measurable functions fn : X ! [0; 1] ; n = 1; 2; :::; the function
lim inf n!1 fn is measurable. In particular, if

f (x) = lim fn (x)

n!1

exists for every x 2 X; then f is measurable.

Theorem 2.1.6. (Fatou’s Lemma) If fn : X ! [0; 1] ; n = 1; 2; :::; are

measurable Z Z
lim inf fn d lim inf fn d :
X n!1 n!1 X

PROOF. Introduce
gk = inf fn :
n k

The de…nition gives that gk " lim inf n!1 fn and, moreover,
Z Z
gk d fn d ; n k
X X

and Z Z
gk d inf fn d :
X n k X
The Fatou Lemma now follows by monotone convergence.

Below we often write Z

f (x)d (x)
E
instead of Z
fd :
E

Example 2.1.3. Suppose a 2 R and f : (R;R ) ! ([0; 1] ; R0;1 ) is

measurable. We claim that
Z Z
f (x + a)dm(x) = f (x)dm(x):
R R
54

First if f = A; where A 2 R ,
Z Z
f (x + a)dm(x) = A a (x)dm(x) = m(A a) =
R R
Z
m(A) = f (x)dm(x):
R
Next it is clear that the relation we want to prove is true for simple mea-
surable functions and …nally, we use the Lebesgue Dominated Convergence
Theorem to deduce the general case.

Exercises

1. Suppose fn : X ! [0; 1] ; n = 1; 2; :::; are measurable and

1
k=1 (fn > 1) < 1:

Prove that
lim sup fn > 1 2Z .
n!1

2. Set fn = n2 [0; 1 ] ; n 2 N+ : Prove that

n

Z Z
lim inf fn dm = 0 < 1 = lim inf fn dm
R n!1 n!1 R

(the inequality in the Fatou Lemma may be strict).

3. Suppose f : (R;R ) ! ([0; 1] ; R0;1 ) is measurable and set

1
g(x) = k=1 f (x + k); x 2 R:

Show that Z
gdm < 1 if and only if ff > 0g 2 Zm :
R
 55

4. Let (X; M; ) be a positive measure space and f : X ! [0; 1] an

(M; R0;1 )-measurable function such that

f (X) N

and Z
f d < 1:
X
For every t 0, set

F (t) = (f > t) and G(t) = (f t):

Prove that Z
1 1
fd = n=0 F (n) = n=1 G(n):
X

2.2. Integration of Functions with Arbitrary Sign

As usual suppose (X; M; ) is a positive measure space. In this section when

we speak of a measurable function f : X ! R it is understood that f is an
(M; R)-measurable function, if not otherwise stated. If f; g : X ! R are
measurable, the sum f + g is measurable since
[
ff + g > g = (ff > qg \ fg > qg)
q2Q

for each real : Besides the function f and the di¤erence f g are mea-
surable. It follows that a function f : X ! R is measurable if and only if
the functions f + = max(0; f ) and f = max(0; f ) are measurable since
f = f+ f :
We write f 2 L1 ( ) if f : X ! R is measurable and
Z
jf jd <1
X
and in this case we de…ne
Z Z Z
+
fd = f d f d :
X X X
56

Note that Z Z
j fd j jf jd
X X

since j f j= f + + f : Moreover, if E 2 M we de…ne

Z Z Z
+
fd = f d f d
E E E

and it follows that Z Z

fd = Ef d :
E X

Note that Z
f d = 0 if (E) = 0:
E

Sometimes we write Z
f (x)d (x)
E

instead of Z
fd :
E

If f; g 2 L1 ( ); setting h = f + g;
Z Z Z
jhjd jf jd + jgjd <1
X X X

and it follows that h + g 2 L1 ( ): Moreover,

h+ h = f+ f + g+ g

and the equation

h+ + f + g = f + + g + + h
gives
Z Z Z Z Z Z
+ + +
h d + f d + g d = f d + g d + h d :
X X X X X X

Thus Z Z Z
hd = fd + gd :
X X X
 57

Moreover, Z Z
fd = fd
X X
for each real : The case 0 follows from (c) in Section 2.1. The case
= 1 is also simple since ( f )+ = f and ( f ) = f + :

Theorem 2.2.1. (Lebesgue’s Dominated Convergence Theorem)

Suppose fn : X ! R; n = 1; 2; :::; are measurable and

f (x) = lim fn (x)

n!1

exists for every x 2 X: Moreover, suppose there exists a function g 2 L1 ( )

such that
j fn (x) j g(x); all x 2 X and n 2 N+ :
Then f 2 L1 ( ), Z
lim j fn f jd =0
n!1 X
and Z Z
lim fn d = fd
n!1 X X
Proof. Since j f j g, the function f is real-valued and measurable since
f + and f are measurable. Note here that

f (x) = lim fn (x); all x 2 X:

n!1

We now apply the Fatous Lemma to the functions 2g j fn f j; n =

1; 2; :::; and have
Z Z
2gd lim inf (2g j fn f j)d
X n!1 X
Z Z
= 2gd lim sup j fn f jd :
X n!1 X
R
But X
2gd is …nite and we get
Z
lim j fn f j d = 0:
n!1 X
58

Since
Z Z Z Z
j fn d f d j=j (f fn )d j jf fn j d
X X X X

the last part in Theorem 2.2.1 follows from the …rst part. The theorem is
proved.

Example 2.2.1. Suppose f : ]a; b[ X ! R is a function such that f (t; ) 2

L1 ( ) for each t 2 ]a; b[ and, moreover, assume @f
@t
exists and
@f
j (t; x) j g(x) for all (t; x) 2 ]a; b[ X
@t
where g 2 L1 ( ). Set
Z
F (t) = f (t; x)d (x) if t 2 ]a; b[ :
X

We claim that F is di¤erentiable and

Z
0 @f
F (t) = (t; x)d (x):
X @t

To see this let t 2 ]a; b[ be …xed and choose a sequence (tn )1

n=1 in ]a; b[ n
ft g which converges to t : De…ne
f (tn ; x) f (t ; x)
hn (x) = if x 2 X:
tn t
Here each hn is measurable and
@f
lim hn (x) = (t ; x) for all x 2 X:
n!1 @t
Furthermore, for each …xed n and x there is a n;x 2 ]tn ; t [ such that hn (x) =
@f
@t
( n;x ; x) and we conclude that j hn (x) j g(x) for every x 2 X: Since
Z
F (tn ) F (t )
= hn (x)d (x)
tn t X

the claim above now follows from the Lebesgue Dominated Convergence The-
orem.
 59

Suppose S(x) is a statement, which depends on x 2 X: We will say that

S(x) holds almost (or -almost) everywhere if there exists an N 2 Z such
that S(x) holds at every point of X n N: In this case we write ”S holds a.e.
” or ”S holds a.e. [ ]”. Sometimes we prefer to write ”S(x) holds a.e.”
or ”S(x) holds a.e. [ ]”: If the underlying measure space is a probability
space, we often say ”almost surely”instead of almost everywhere. The term
”almost surely”is abbreviated a.s.
Suppose f : X ! R; is an (M; R)-measurable functions and g : X ! R:
If f = g a.e. [ ] there exists an N 2 Z such that f (x) = g(x) for every
x 2 X n N: We claim that g is (M ; R)-measurable. To see this let 2 R
and use that

fg > g = [ff > g \ (X n N )] [ [fg > g \ N ] :

Now if we de…ne
A = ff > g \ (X n N )
the set A 2 M and
A fg > g A [ N:
Accordingly from this fg > g 2 M and g is (M ; R)-measurable since
is an arbitrary real number.
Next suppose fn : X ! R; n 2 N+ ; is a sequence of (M; R)-measurable
functions and f : X ! R a function. Recall if

lim fn (x) = f (x); all x 2 X

n!1

then f is (M; R)-measurable since

1
ff > g = [k;l2N+ \n k fn > +l ; all 2 R:

If we only assume that

lim fn (x) = f (x); a.e. [ ]

n!1

then f need not be (M; R)-measurable but f is (M ; R)-measurable. To

see this suppose N 2 Z and

lim fn (x) = f (x); all x 2 X n N:

n!1
60

Then
lim XnN (x)fn (x) = XnN (x)f (x)
n!1

and it follows that the function XnN f is (M; R)-measurable. Since f =

XnN f a.e. [ ] it follows that f is (M ; R)-measurable. The next example
shows that f need not be (M; R)-measurable.

Example 2.2.2. Let X = f0; 1; 2g ; M = f ; f0g ; f1; 2g ; Xg ; and (A) =

A (0); A 2 M: Set fn = f1;2g ; n 2 N+ ; and f (x) = x; x 2 X: Then each
fn is (M; R)-measurable and

lim fn (x) = f (x) a.e. [ ]

n!1

since n o
x 2 X; lim fn (x) = f (x) = f0; 1g
n!1

and N = f1; 2g is a -null set. The function f is not (M; R)-measurable.

Suppose f; g 2 L1 ( ): The functions f and g are equal almost everywhere

with respect to if and only if ff 6= gg 2 Z : This is easily seen to be an
equivalence relation and the set of all equivalence classes is denoted by L1 ( ):
Moreover, if f = g a.e. [ ] ; then
Z Z
fd = gd
X X

since
Z Z Z Z Z
fd = fd + fd = fd = gd
X ff =gg ff 6=gg ff =gg ff =gg

and, in a similar way, Z Z

gd = gd :
X ff =gg

Below we consider the elements of L1 ( ) as members of L1 ( ) and two mem-

bers of L1 ( ) are identi…ed if they are equal a.e. [ ] : From this convention
 61

it is straight-forward to de…ne f + g and f for all f; g 2 L1 ( ) and 2 R:

Moreover, we get
Z Z Z
(f + g)d = fd + gd if f; g 2 L1 ( )
X X X

and Z Z
fd = f d if f 2 L1 ( ) and 2 R:
X X
Next we give two theorems where exceptional null sets enter. The …rst
one is a mild variant of Theorem 2.2.1 and needs no proof.

Theorem 2.2.2. Suppose (X; M; ) is a positive complete measure space

and let fn : X ! R; n 2 N+ ; be measurable functions such that

sup j fn (x) j g(x) a.e. [ ]

n2N+

where g 2 L1 ( ): Moreover, suppose f : X ! R is a function and

f (x) = lim fn (x) a.e. [ ] :

n!1

Then, f 2 L1 ( ), Z
lim j fn f jd =0
n!1 X
and Z Z
lim fn d = fd :
n!1 X X

Theorem 2.2.3. Suppose (X; M; ) is a positive measure space.

(a) If f : (X; M ) ! ([0; 1] ; R0;1 ) is measurable there exists a measur-
able function g : (X; M) ! ([0; 1] ; R0;1 ) such that f = g a.e. [ ] :
(b) If f : (X; M ) ! (R; R) is measurable there exists a measurable
function g : (X; M) ! (R; R) such that f = g a.e. [ ] :

PROOF. Since f = f + f it is enough to prove Part (a): There exist simple

M -measurable functions 'n ; n 2 N+ ; such that 0 'n " f: For each …xed
62

n suppose 1n ; :::; kn n are the distinct values of 'n and choose for each …xed
i = 1; :::; kn a set Ain 'n 1 (f in g) such that Ain 2 M and 'n 1 ( in ) n Ain
2 Z . Set
kn
n = i=1 in Ain :
kn
Clearly n (x) " f (x) if x 2 E =def \1
n=1 ([i=1 Ain ) and (X n E) = 0: We
now de…ne g(x) = f (x); if x 2 E; and g(x) = 0 if x 2 X n E: The theorem
is proved.

Exercises

1. Suppose f and g are real-valued measurable functions. Prove that f 2 and

f g are measurable functions.

2. Suppose f 2 L1 ( ): Prove that

Z
lim j f j d = 0:
!1 jf j
R R
(Here jf j
means fjf j g
.)

3. Suppose f 2 L1 ( ): Prove that to each " > 0 there exists a > 0 such
that Z
jf jd <"
E

whenever (E) < :

4. Let (fn )1
n=1 be a sequence of (M; R)-measurable functions. Prove that
the set of all x 2 R such that the sequence (fn (x))1
n=1 converges to a real
limit belongs to M:

5. Let (X; M; R) be a positive measure space such that (A) = 0 or 1 for

every A 2 M: Show that f 2 L1 ( ) if and only if f (x) = 0 a.e. [ ] :
 63

6. Let (X; M; ) be a positive measure space and suppose f and g are

non-negative measurable functions such that
Z Z
fd = gd ; all A 2 M:
A A

(a) Prove that f = g a.e. [ ] if is -…nite.

(b) Prove that the conclusion in Part (a) may fail if is not -…nite.

7. Let (X; M; ) be a …nite positive measure space and suppose the functions
fn : X ! R; n = 1; 2; :::; are measurable. Show that there is a sequence
( n )1
n=1 of positive real numbers such that

lim n fn = 0 a.e. [ ] :
n!1

8. Let (X; M; ) be a positive measure space and let fn : X ! R; n = 1; 2; :::;

be a sequence in L1 ( ) which converges to f a.e. [ ] as n ! 1: Suppose
f 2 L1 ( ) and Z Z
lim j fn j d = jf jd :
n!1 X X
Show that Z
lim j fn f j d = 0:
n!1 X

9. Let (X; M; ) be a …nite positive measure space and suppose f 2 L1 ( )

is a bounded function such that
Z Z Z
2 3
f d = f d = f 4d :
X X X

Prove that f = A for an appropriate A 2 M:

10. Let (X; M; ) be a …nite positive measure space and f : X ! R a

measurable function. Prove that f 2 L1 ( ) if and only if
1
k=1 (j f j k) < 1:
64

11. Suppose f 2 L1 (m): Prove that the series 1

k= 1 f (x + k) converges for
m-almost all x:

R
12. a) Suppose f : R ! [0; 1[ is Lebesgue measurable and R
f dm < 1:
Prove that
lim m(f ) = 0:
!1

b) Find a Lebesgue measurable function f : R ! [0; 1[ such that f 2

=
1
L (m); m(f > 0) < 1; and

lim m(f ) = 0:
!1

2.3 Comparison of Riemann and Lebesgue Integrals

In this section we will show that the Lebesgue integral is a natural general-
ization of the Riemann integral. For short, the discussion is restricted to a
closed and bounded interval.
Let [a; b] be a closed and bounded interval and suppose f : [a; b] ! R is
a bounded function. For any partition

: a = x0 < x1 < ::: < xn = b

of [a; b] de…ne
n
S f= i=1 ( sup f )(xi xi 1 )
]xi 1 ;xi ]

and
n
s f= k=1 ( inf f )(xi xi 1 ):
]xi 1 ;xi ]

The function f is Riemann integrable if

inf S f = sup s f

Rb
and the Riemann integral a
f (x)dx is, by de…nition, equal to this common
value.
 65

Below an ((R )[a;b] ; R)-measurable function is simply called Lebesgue

measurable. Furthermore, we write m instead of mj[a;b] :

Theorem 2.3.1. A bounded function f : [a; b] ! R is Riemann integrable

if and only if the set of discontinuity points of f is a Lebesgue null set.
Moreover, if the set of discontinuity points of f is a Lebesgue null set, then
f is Lebesgue measurable and
Z b Z
f (x)dx = f dm:
a [a;b]

PROOF. A partition 0 : a = x00 < x01 < ::: < x0n0 = b is a re…nement of a
partition : a = x0 < x1 < ::: < xn = b if each xk is equal to some x0l and in
0
this case we write : The de…nitions give S f S 0 f and s f s 0 f
0
if : We de…ne, mesh( ) = max1 i n (xi xi 1 ):
First suppose f is Riemann integrable. For each partition let
n
G = f (a) fag + i=1 ( sup f ) ]xi 1 ;xi ]
]xi 1 ;xi ]

and
n
g = f (a) fag + i=1 ( inf f ) ]xi 1 ;xi ]
]xi 1 ;xi ]

and note that Z

G dm = S f
[a;b]

and Z
g dm = s f:
[a;b]

Suppose k; k = 1; 2; :::; is a sequence of partitions such that k k+1 ,

Z b
S kf # f (x)dx
a

and Z b
s k
f" f (x)dx
a
66

as k ! 1: Let G = limk!1 G k and g = limk!1 g k : Then G and g are

(R[a;b] ; R)-measurable, g f G; and by dominated convergence
Z Z Z b
Gdm = gdm = f (x)dx:
[a;b] [a;b] a

But then Z
(G g)dm = 0
[a;b]

so that G = g a.e. [m] and therefore G = f a.e. [m] : In particular, f is

Lebesgue measurable and
Z b Z
f (x)dx = f dm:
a [a;b]

Set
N = fx; g(x) < f (x) or f (x) < G(x)g :
We proved above that m(N ) = 0: Let M be the union of all those points which
belong to some partition k : Clearly, m(M ) = 0 since M is denumerable.
We claim that f is continuous o¤ N [ M: If f is not continuous at a point
c2= N [ M , there is an " > 0 and a sequence (cn )1 n=1 converging to c such
that
j f (cn ) f (c) j " all n:
Since c 2
= M , c is an interior point to exactly one interval of each partition
k and we get
G k (c) g k (c) "
and in the limit
G(c) g(c) ":
But then c 2 N which is a contradiction.
Conversely, suppose the set of discontinuity points of f is a Lebesgue null
set and let ( k )1
k=1 is an arbitrary sequence of partitions of [a; b] such that
k k+1 and mesh( k ) ! 0 as k ! 1: By assumption,

lim G k
(x) = lim g k
(x) = f (x)
k!1 k!1
 67

at each point x of continuity of f . Therefore f is Lebesgue measurable and

dominated convergence yields
Z Z
lim G k dm = f dm
k!1 [a;b] [a;b]

and Z Z
lim g k
dm = f dm:
k!1 [a;b] [a;b]

Thus f is Riemann integrable and

Z b Z
f (x)dx = f dm:
a [a;b]

In the following we sometimes write

Z
f (x)dx (A 2 R )
A

instead of Z
f dm (A 2 R ):
A
In a similar way we often prefer to write
Z
f (x)dx (A 2 Rn )
A

instead of Z
f dmn (A 2 Rn ):
A
Rb R
Furthermore, a f dm means [a;b] f dm: Here, however, a warning is moti-
vated. It is simple to …nd a real-valued function f on [0; 1[, which is bounded
on each bounded subinterval of [0; 1[ ; such that the generalized Riemann
integral Z 1
f (x)dx
0
is convergent, that is Z b
lim f (x)dx
b!1 0
68

exists and the limit is a real number, while the Riemann integral
Z 1
j f (x) j dx
0

is divergent (take e.g. f (x) = sinx x ): In this case the function f does not
belong to L1 with respect to Lebesgue measure on [0; 1[ since
Z Z b
j f j dm = lim j f (x) j dx = 1:
[0;1[ b!1 0

Exercises

1. Let fn : [0; 1] ! [0; 1], n 2 N; be a sequence of Riemann integrable

functions such that

lim fn (x) exists = f (x) all x 2 [0; 1] :

n!1

Show by giving an example that f need not be Riemann integrable.

2 njxj
2. Suppose
R fn (x) = n j x j e ; x 2 R; n 2 N+ : Compute limn!1 fn and
limn!1 R fn dm.

3. Compute the following limits and justify the calculations:

a) Z 1
sin(ex )
lim dx:
n!1 0 1 + nx2
b) Z n
x n
lim (1 + ) cos xdx:
n!1 0 n
c) Z n
x n 2x
lim (1 + ) e dx:
n!1 0 n
d) Z 1
x n x
lim (1 + ) exp( (1 + )n )dx:
n!1 0 n n
 69

e) Z n
x n x
lim (1 ) e 2 dx:
n!1 0 n

f) Z n
x n 1 + nx
lim (1 ) cos xdx
n!1 0 n n+x
g) Z 1
x n2 nx
lim (1 + ) e dx:
n!1 0 n

4. Let (rn )1
n=1 be an enumeration of Q and de…ne

1 n
f (x) = n=1 2 '(x rn )
1
where '(x) = x 2 if 0 < x < 1 and '(x) = 0 if x 0 or x 1: Show that
a) Z 1
f (x)dx = 2:
1

b)
Z b
f 2 (x)dx = 1 if a < b:
a
c)
f < 1 a.s. [m] :
d)
sup f (x) = +1 if a < b:
a<x<b

5. Suppose Z 1
tx ln(1
+ x)
f (t) = e dx; t > 0:
0 1+x
R1
a) Show that 0 f (t)dt < 1:
b) Show that f is in…nitely many times di¤erentiable.

6. Suppose
Z 1 2
xe x
f (t) = dx; t > 0:
0 x2 + t2
70

Compute Z 1
lim f (t) and f (t)dt:
t!0+ 0
Finally, prove that f is di¤erentiable.

2.4. Expectation

Suppose ( ; F; P ) is a probability space and : ( ; F) ! (S; S) a random

variable. Recall that the probability law of is given by the image measure
P : By de…nition, Z Z
Bd = B( )dP
S
for every B 2 S; and, hence
Z Z
'd = '( )dP
S

for each simple S-measurable function ' on S (we sometimes write f g=

f (g)): By monotone convergence, we get
Z Z
fd = f ( )dP
S

for every measurable f : S ! [0; 1] : Thus if f : S ! R is measurable,

f 2 L1 ( ) if and only if f ( ) 2 L1 (P ) and in this case
Z Z
fd = f ( )dP:
S

In the special case when is real-valued and 2 L1 (P );

Z Z
xd (x) = dP:
R

The integral in the right-hand side is called the expectation of and is

denoted by E [ ] :
 71

CHAPTER 3
Further Construction Methods of Measures

Introduction

In the …rst section of this chapter we collect some basic results on metric
spaces, which every mathematician must know about. Section 3.2 gives a
version of the Riesz Representation Theorem, which leads to another and
perhaps simpler approach to Lebesgue measure than the Carathéodory The-
orem. A reader can skip Section 3.2 without losing the continuity in this
paper. The chapter also treats so called product measures and Stieltjes in-
tegrals.

3.1. Metric Spaces

The construction of our most important measures requires topological con-

cepts. For our purpose it will be enough to restrict ourselves to so called
metric spaces.
A metric d on a set X is a mapping d : X X ! [0; 1[ such that

(a) d(x; y) = 0 if and only if x = y

(b) d(x; y) = d(y; x) (symmetry)
(c) d(x; y) d(x; z) + d(z; y) (triangle inequality).

Here recall, if A1 ; :::; An are sets,

A1 ::: An = f(x1 ; :::; xn ); xi 2 Ai for all i = 1; :::; ng

A set X equipped with a metric d is called a metric space. Sometimes we

write X = (X; d) to emphasize the metric d: If E is a subset of the metric
72

space (X; d); the function djE E (x; y) = d(x; y); if x; y 2 E; is a metric on
E: Thus (E; djE E ) is a metric space.
The function '(t) = min(1; t); t 0; satis…es the inequality

'(s + t) '(s) + '(t):

Therefore, if d is a metric on X, min(1; d) is a metric on X: The metric

min(1; d) is a bounded metric.
The set R equipped with the metric d1 (x; y) =j x y j is a metric space.
More generally, Rn equipped with the metric

dn (x; y) = dn ((x1 ; :::; xn ); (y1 ; :::; yn )) = max j xk yk j

1 k n

is a metric space. If not otherwise stated, it will always be assumed that Rn

is equipped with this metric.
Let C [0; T ] denote the vector space of all real-valued continuous functions
on the interval [0; T ] ; where T > 0: Then

d1 (x; y) = max j x(t) y(t) j

0 t T

is a metric on C [0; T ] :
If (Xk ; ek ); k = 1; :::; n, are metric spaces,

d(x; y) = max ek (xk ; yk ); x = (x1 ; :::; xn ) ; y = (y1 ; :::; yn )

1 k n

is a metric on X1 ::: Xn : The metric d is called the product metric on

X1 ::: Xn :
If X = (X; d) is a metric space and x 2 X and r > 0; the open ball with
centre at x and radius r is the set B(x; r) = fy 2 X; d(y; x) < rg : If E X
and E is contained in an appropriate open ball in X it is said to be bounded.
The diameter of E is, by de…nition,

diam E = sup d(x; y)

x;y2E

and it follows that E is bounded if and only if diam E < 1. A subset of X

which is a union of open balls in X is called open. In particular, an open
ball is an open set. The empty set is open since the union of an empty family
of sets is empty. An arbitrary union of open sets is open. The class of all
 73

open subsets of X is called the topology of X: The metrics d and min(1; d)

determine the same topology. A subset E of X is said to be closed if its
complement E c relative to X is open. An intersection of closed subsets of
X is closed. If E X, E denotes the largest open set contained in E and
E (or E) the smallest closed set containing E: E is the interior of E and
E its closure. The -algebra generated by the open sets in X is called the
Borel -algebra in X and is denoted by B(X). A positive measure on B(X)
is called a positive Borel measure.
A sequence (xn )1n=1 in X converges to x 2 X if

lim d(xn ; x) = 0:
n!1

If, in addition, the sequence (xn )1

n=1 converges to y 2 X; the inequalities

0 d(x; y) d(xn ; x) + d(xn ; y)

imply that y = x and the limit point x is unique.
If E X and x 2 X; the following properties are equivalent:

(i) x 2 E :
(ii) B(x; r) \ E 6= ; all r > 0:
(iii) There is a sequence (xn )1
n=1 in E which converges to x:

If B(x; r) \ E = , then B(x; r)c is a closed set containing E but not x:

Thus x 2 = E ; since E c is
= E : This proves that (i))(ii). Conversely, if x 2
open there exists an open ball B(y; s) such that x 2 B(y; s) E c E c : Now
choose r = s d(x; y) > 0 so that B(x; r) B(y; s): Then B(x; r) \ E = :
This proves (ii))(i).
If (ii) holds choose for each n 2 N+ a point xn 2 E with d(xn ; x) < n1
and (iii) follows. If there exists an r > 0 such that B(x; r) \ E = ; then
(iii) cannot hold. Thus (iii))(ii).
If E X, the set E nE is called the boundary of E and is denoted by
@E:
A set A X is said to be dense in X if A = X: The metric space X is
called separable if there is an at most denumerable dense subset of X: For
example, Qn is a dense subset of Rn : The space Rn is separable.
74

Theorem 3.1.1. B(Rn ) = Rn :

PROOF. The -algebra Rn is generated by the open n-cells in Rn and an

open n-cell is an open subset of Rn : Hence Rn B(Rn ): Let U be an open
subset in Rn and note that an open ball in Rn = (Rn ; dn ) is an open n-cell.
If x 2 U there exist an a 2Qn \ U and a rational number r > 0 such that
x 2 B(a; r) U: Thus U is an at most denumerable union of open n-cells
and it follows that U 2 Rn : Thus B(Rn ) Rn and the theorem is proved.

Let X = (X; d) and Y = (Y; e) be two metric spaces. A mapping f :

X ! Y (or f : (X; d) ! (Y; e) to emphasize the underlying metrics) is said
to be continuous at the point a 2 X if for every " > 0 there exists a > 0
such that
x 2 B(a; ) ) f (x) 2 B(f (a); "):
Equivalently this means that for any sequence (xn )1 n=1 in X which converges
to a in X; the sequence (f (xn ))1
n=1 converges to f (a) in Y: If f is continuous
at each point of X, the mapping f is called continuous. Stated otherwise
this means that
f 1 (V ) is open if V is open
or
1
f (F ) is closed if F is closed.
The mapping f is said to be Borel measurable if
1
f (B) 2 B(X) if B 2 B(Y )

or, what amounts to the same thing,

1
f (V ) 2 B(X) if V is open.

A Borel measurable function is sometimes called a Borel function. A

continuous function is a Borel function.

Example 3.1.1. Let f : (R;d1 )! (R;d1 ) be a continuous strictly increasing

function and set (x; y) =j f (x) f (y) j; x; y 2 R: Then is a metric on R.
 75

De…ne j(x) = x; x 2 R: The mapping j : (R;d1 )! (R; ) is continuous. We

claim that the map j : (R; ) ! (R;d1 ) is continuous: To see this, let a 2 R
and suppose the sequence (xn )1 n=1 converges to a in the metric space (R; );
that is j f (xn ) f (a) j! 0 as n ! 1: Let " > 0: Then

f (xn ) f (a) f (a + ") f (a) > 0 if xn a+"

and
f (a) f (xn ) f (a) f (a ") > 0 if xn a ":
Thus xn 2 ]a "; a + "[ if n is su¢ ciently large. This proves that he map
j : (R; ) ! (R;d1 ) is continuous.
The metrics d1 and determine the same topology and Borel subsets of
R:

A mapping f : (X; d) ! (Y; e) is said to be uniformly continuous if for

each " > 0 there exists a > 0 such that e(f (x); f (y)) < " as soon as
d(x; y) < :
If x 2 X and E; F X; let

d(x; E) = inf d(x; u)

u2E

be the distance from x to E and let

d(E; F ) = inf d(u; v)

u2E;v2F

be the distance between E and F: Note that d(x; E) = 0 if and only if x 2 E:

If x; y 2 X and u 2 E;

d(x; u) d(x; y) + d(y; u)

and, hence
d(x; E) d(x; y) + d(y; u)
and
d(x; E) d(x; y) + d(y; E):
Next suppose E 6= : Then by interchanging the roles of x and y; we get

j d(x; E) d(y; E) j d(x; y)

76

and conclude that the distance function d(x; E); x 2 X; is continuous. In

fact, it is uniformly continuous. If x 2 X and r > 0; the so called closed ball
B(x; r) = fy 2 X; d(y; x) rg is a closed set since the map y ! d(y; x);
y 2 X; is continuous.
If F X is closed and " > 0, the continuous function
X 1
F;" = max(0; 1 d( ; F ))
"
X
ful…ls 0 F;" 1 and XF;" = 1 on F: Furthermore,
X
F;" (a) > 0 if and only
if a 2 F" =def fx 2 X; d(x; F ) < "g : Thus
X
F F;" F" :

Let X = (X; d) be a metric space. A sequence (xn )1 n=1 in X is called

a Cauchy sequence if to each " > 0 there exists a positive integer p such
that d(xn ; xm ) < " for all n; m p: If a Cauchy sequence (xn )1n=1 contains a
convergent subsequence (xnk )1 k=1 it must be convergent. To prove this claim,
1
suppose the subsequence (xnk )k=1 converges to a point x 2 X: Then
d(xm ; x) d(xm ; xnk ) + d(xnk ; x)
can be made arbitrarily small for all su¢ ciently large m by choosing k su¢ -
ciently large. Thus (xn )1
n=1 converges to x:
A subset E of X is said to be complete if every Cauchy sequence in E
converges to a point in E: If E X is closed and X is complete it is clear
that E is complete. Conversely, if X is a metric space and a subset E of X
is complete, then E is closed:
It is important to know that R is complete equipped with its standard
metric. To see this let (xn )1
n=1 be a Cauchy sequence. There exists a positive
integer such that j xn xm j< 1 if n; m p: Therefore
j xn j j xn xp j + j xp j 1+ j xp j
for all n p: We have proved that the sequence (xn )1
n=1 is bounded (the
reader can check that every Cauchy sequence in a metric space has this
property). Now de…ne
a = sup fx 2 R; there are only …nitely many n with xn xg :
The de…nition implies that there exists a subsequence (xnk )1
k=1 ; which con-
verges to a (since for any r > 0; xn 2 B(a; r) for in…nitely many n). The
 77

original sequence is therefore convergent and we conclude that R is complete

(equipped with its standard metric d1 ): It is simple to prove that the product
of n complete spaces is complete and we conclude that Rn is complete.
Let E X: A family (Vi )i2I of subsets of X is said to be a cover of E
if [i2I Vi E and E is said to be covered by the Vi0 s: The cover (Vi )i2I is
said to be an open cover if each member Vi is open. The set E is said to be
totally bounded if, for every " > 0; E can be covered by …nitely many open
balls of radius ": A subset of a totally bounded set is totally bounded.
The following de…nition is especially important.

De…nition 3.1.1. A subset E of a metric space X is said to be compact if

to every open cover (Vi )i2I of E, there is a …nite subcover of E, which means
there is a …nite subset J of I such that (Vi )i2J is a cover of E:

If K is closed, K E; and E is compact, then K is compact. To see this,

let (Vi )i2I be an open cover of K: This cover, augmented by the set X n K
is an open cover of E and has a …nite subcover since E is compact. Noting
that K \ (X n K) = ; the assertion follows.

Theorem 3.1.2. The following conditions are equivalent:

(a) E is complete and totally bounded.
(b) Every sequence in E contains a subsequence which converges to a
point of E:
(c) E is compact.

PROOF. (a))(b). Suppose (xn )1 n=1 is a sequence in E: The set E can be

covered by …nitely many open balls of radius 2 1 and at least one of them
must contain xn for in…nitely many n 2 N+ : Suppose xn 2 B(a1 ; 2 1 ) if
n 2 N1 N0 =def N+ ; where N1 is in…nite. Next E \ B(a1 ; 2 1 ) can be
covered by …nitely many balls of radius 2 2 and at least one of them must
contain xn for in…nitely many n 2 N1 : Suppose xn 2 B(a2 ; 2 1 ) if n 2 N2 ;
where N2 N1 is in…nite. By induction, we get open balls B(aj ; 2 j ) and
in…nite sets Nj Nj 1 such that xn 2 B(aj ; 2 j ) for all n 2 Nj and j 1:
78

Let n1 < n2 < :::, where nk 2 Nk ; k = 1; 2; ::: . The sequence (xnk )1

k=1 is a
Cauchy sequence, and since E is complete it converges to a point of E .

(b))(a). If E is not complete there is a Cauchy sequence in E with no

limit in E: Therefore no subsequence can converge in E; which contradicts
(b). On the other hand if E is not totally bounded, there is an " > 0 such
that E cannot be covered by …nitely many balls of radius ": Let x1 2 E
be arbitrary. Having chosen x1 ; :::; xn 1 ; pick xn 2 En [ni=11 B(xi ; "); and
so on. The sequence (xn )1n=1 cannot contain any convergent subsequence as
d(xn ; xm ) " if n 6= m; which contradicts (b).

f(a) and (b)g )(c). Let (Vi )i2I be an open cover of E: Since E is totally
bounded it is enough to show that there is an " > 0 such that any open
ball of radius " which intersects E is contained in some Vi : Suppose on the
contrary that for every n 2 N+ there is an open ball Bn of radius 2 n
which intersects E and is contained in no Vi : Choose xn 2 Bn \ E and
assume without loss of generality that (xn )1
n=1 converges to some point x in
E by eventually going to a subsequence. Suppose x 2 Vi0 and choose r > 0
such that B(x; r) Vi0 : But then Bn B(x; r) Vi0 for large n, which
contradicts the assumption on Bn :

(c))(b). If (xn )1
n=1 is a sequence in E with no convergent subsequence in
E, then for every x 2 E there is an open ball B(x; rx ) which contains xn for
only …nitely many n: Then (B(x; rx ))x2E is an open cover of E without a
…nite subcover.

Corollary 3.1.1. A subset of Rn is compact if and only if it is closed and

bounded.

PROOF. Suppose K is compact. If xn 2 K and xn 2 = B(0; n) for every

n 2 N+ ; the sequence (xn )1
n=1 cannot contain a convergent subsequence.
Thus K is bounded. Since K is complete it is closed.
 79

Conversely, suppose K is closed and bounded. Since Rn is complete and

K is closed, K is complete. We next prove that a bounded set is totally
bounded. It is enough to prove that any n-cell in Rn is a union of …nitely
many n-cells I1 ::: In where each interval I1 ; :::; In has a prescribed positive
length. This is clear and the theorem is proved.

Corollary 3.1.2. Suppose f : X ! R is continuous and X compact:

(a) There exists an a 2 X such that maxX f = f (a) and a b 2 X
such that minX f = f (b):
(b) The function f is uniformly continuous:

PROOF. (a) For each a 2 X; let Va = fx 2 X : f (x) < 1 + f (a)g : The open
cover (Va )a2K of X has a …nite subcover and it follows that f is bounded. Let
(xn )1
n=1 be a sequence in X such that f (xn ) ! supK f as n ! 1: Since X is
compact there is a subsequence (xnk )1 k=1 which converges to a point a 2 X:
Thus, by the continuity of f; f (xnk ) ! f (a) as k ! 1:
The existence of a minimum is proved in a similar way.

(b) If f is not uniformly continuous there exist " > 0 and sequences
(xn )1 1
n=1 and (yn )n=1 such that j f (xn ) f (yn ) j " and j xn yn j< 2 n
for every n 1: Since X is compact there exists a subsequence (xnk )1 k=1 of
(xn )1
n=1 which converges to a point a 2 X: Clearly the sequence (y 1
nk )k=1
converges to a and therefore

j f (xnk ) f (ynk ) j j f (xnk ) f (a) j + j f (a) f (ynk ) j! 0

as k ! 1 since f is continuous. But j f (xnk ) f (ynk ) j " and we have got

a contradiction. The corollary is proved.

Example 3.1.2. Suppose X = ]0; 1] and de…ne 1 (x; y) = d1 (x; y) and

1 1
2 (x; y) =j x y
j; x; y 2 X: As in Example 3.1.1 we conclude that the metrics
1 and 2 determine the same topology of subsets of X: The space (X; 1 )
80

totally bounded but not complete. However, the space (X; 2 ) is not totally
bounded but it is complete. To see this, let (xn )1
n=1 be a Cauchy sequence in
(X; 2 ): As a Cauchy sequence it must be bounded and therefore there exists
an " 2 ]0; 1] such that xn 2 ["; 1] for all n: But then, by Corollary 3.1.1,
(xn )1
n=1 contains a convergent subsequence in (X; 1 ) and, accordingly from
this, the same property holds in (X; 2 ): The space (X; 2 ) is not compact,
since (X; 1 ) is not compact, and we conclude from Theorem 3.1.2 that the
space (X; 2 ) cannot be totally bounded.

Example 3.1.3. Set R̂=R[ f 1; 1g and

^ y) =j arctan x
d(x; arctan y j
if x; y 2 R̂: Here
arctan 1 = and arctan 1 = :
2 2
Example 3.1.1 shows that the standard metric d1 and the metric d^jR R
determine the same topology.
We next prove that the metric space R̂ is compact. To this end, consider
a sequence (xn )1
n=1 in R̂. If there exists a real number M such that j xn j M
1
for in…nitely many n; the sequence (xn )n=1 contains a convergent subsequence
since the interval [ M; M ] is compact. In the opposite case, for each positive
real number M , either xn M for in…nitely many n or xn M for
in…nitely many n: Suppose xn M for in…nitely many n for every M 2
^ n ; 1) =j arctan xn
N+ : Then d(x j! 0 as k ! 1 for an appropriate
k k 2
1
subsequence (xnk )k=1 :
The space R̂= (R̂,d)^ is called a two-point compacti…cation of R.
It is an immediate consequence of Theorem 3.1.2 that the product of
…nitely many compact metric spaces is compact. Thus R ^ n equipped with
the product metric is compact.
We will …nish this section with several useful approximation theorems.

Theorem 3.1.3. Suppose X is a metric space and positive Borel measure

in X: Moreover, suppose there is a sequence (Un )1
n=1 of open subsets of X
such that
X = [1 n=1 Un
 81

and
(Un ) < 1; all n 2 N+ :
Then for each A 2 B(X) and " > 0; there are a closed set F A and an
open set V A such that
(V n F ) < ":
In particular, for every A 2 B(X);

(A) = inf (V )
V A
V open

and
(A) = sup (F )
F A
F closed

If X = R and (A) = 1 n=1 n1 (A) ; A 2 R; then (f0g) = 0 and (V ) =

1 for every open set containing f0g : The hypothesis that the sets Un ; n 2
N+ ; are open (and not merely Borel sets) is very important in Theorem 3.1.3.

PROOF. First suppose that is a …nite positive measure.

Let A be the class of all Borel sets A in X such that for every " > 0
there exist a closed F A and an open V A such that (V n F ) < ": If F
is a closed subset of X and Vn = x; d(x; F ) < n1 ; then Vn is open and, by
Theorem 1.1.2 (f), (Vn ) # v(F ) as n ! 1. Thus F 2 A and we conclude
that A contains all closed subsets of X:
Now suppose A 2 A: We will prove that Ac 2 A: To this end, we choose
" > 0 and a closed set F A and an open set V A such that (V nF ) < ":
Then V c Ac F c and, moreover, (F c n V c ) < " since

V n F = F c n V c:

If we note that V c is closed and F c open it follows that Ac 2 A:

Next let (Ai )1
i=1 be a denumerable collection of members of A. Choose
" > 0: By de…nition, for each i 2 N+ there exist a closed Fi Ai and an
open Vi Ai such that (Vi n Fi ) < 2 i ": Set

V = [1
i=1 Vi :
82

Then
(V n ([1
i=1 Fi )) ([1
i=1 (Vi n Fi ))
1
i=1 (Vi n Fi ) < ":
But
V n ([1 1 n
i=1 Fi ) = \n=1 fV n ([i=1 Fi )g

and since is a …nite positive measure

(V n ([1 n
i=1 Fi )) = lim (V n ([i=1 Fi )):
n!1

Accordingly, from these equations

(V n ([ni=1 Fi )) < "

if n is large enough. Since a union of open sets is open and a …nite union of
closed sets is closed, we conclude that [1 i=1 Ai 2 A: This proves that A is a
-algebra. Since A contains each closed subset of X; A = B(X):
We now prove the general case. Suppose A 2 B(X): Since Un is a …nite
positive measure the previous theorem gives us an open set Vn A \ Un such
that Un (Vn n (A \ Un )) < "2 n : By eventually replacing Vn by Vn \ Un we can
assume that Vn Un : But then (Vn n (A \ Un )) = Un (Vn n (A \ Un )) < "2 n :
Set V = [1 n=1 Vn and note that V is open. Moreover,

V nA [1
n=1 (Vn n (A \ Un ))

and we get
1
(V n A) n=1 (Vn n (A \ Un )) < ":
By applying the result already proved to the complement Ac we conclude
there exists an open set W Ac such that

(A n W c ) = (W n Ac ) < ":

Thus if F =def W c it follows that F A V and (V n F ) < 2": The

theorem is proved.

If X is a metric space C(X) denotes the vector space of all real-valued

continuous functions f : X ! R: If f 2 C(X); the closure of the set of
 83

all x where f (x) 6= 0 is called the support of f and is denoted by suppf:

The vector space of all all real-valued continuous functions f : X ! R with
compact support is denoted by Cc (X):

Corollary 3.1.3. Suppose and are positive Borel measures in Rn such

that
(K) < 1 and (K) < 1
for every compact subset K of Rn : If
Z Z
f (x)d (x) = f (x)d (x); all f 2 Cc (Rn )
Rn Rn

then = :

PROOF. Let F be closed. Clearly (B(0; i)) < 1 and (B(0; i)) < 1 for
every positive integer i: Hence, by Theorem 3.1.3 it is enough to show that
(F ) = (F ): Now …x a positive integer i and set K = B(0; i) \ F: It is
enough to show that (K) = (K): But
Z Z
Rn Rn
K;2 j (x)d (x) = K;2 j (x)d (x)
Rn Rn

for each positive integer j and letting j ! 1 we are done.

A metric space X is called a standard space if it is separable and com-

plete. Standard spaces have a series of very nice properties related to measure
theory; an example is furnished by the following

Theorem 3.1.4. (Ulam’s Theorem) Let X be a standard space and

suppose is a …nite positive Borel measure on X: Then to each A 2 B(X)
and " > 0 there exist a compact K A and an open V A such that
(V n K) < ":
84

PROOF. Let " > 0: We …rst prove that there is a compact subset K of X
such that (K) > (X) ": To this end, let A be a dense denumerable subset
of X and let (ai )1 i=1 be an enumeration of A: Now for each positive integer
j; [1
i=1 B(a i ; 2 j
") = X; and therefore there is a positive integer nj such that
n
j
([i=1 B(ai ; 2 j ")) > (X) 2 j ":
Set
n
j
Fj = [i=1 B(ai ; 2 j ")
and
L = \1
j=1 Fj :

The set L is totally bounded. Since X is complete and L closed, L is complete.

Therefore, the set L is compact and, moreover

(K) = (X) (Lc ) = (X) ([1 c

j=1 Fj )

1
(X) j=1 (Fjc ) = (X) 1
j=1 ( (X) (Fj ))
1 j
(X) j=1 2 " = (X) ":
Depending on Theorem 3.1.3 to each A 2 B(X) there exists a closed
F A and an open V A such that (V n F ) < ": But

V n (F \ L) = (V n F ) [ (F n L)

and we get

(V n (F \ L)) (V n F ) + (X n K) < 2":

Since the set F \ L is compact Theorem 3.1.4 is proved.

Two Borel sets in Rn are said to be almost disjoint if their intersection

has volume measure zero.

Theorem 3.1.5. Every open set U in Rn is the union of an at most denu-

merable collection of mutually almost disjoint cubes.
 85

Before the proof observe that a cube in Rn is the same as a closed ball
in Rn equipped with the metric dn .

PROOF. For each, k 2 N+ ; let Qk be the class of all cubes of side length 2 k
whose vertices have coordinates of the form i2 k ; i 2 Z: Let F1 be the union
of those cubes in Q1 which are contained in U: Inductively, for k 1; let
Fk be the union of those cubes in Qk which are contained in U and whose
interiors are disjoint from [kj=11 Fj : Since d(x;Rn n U ) > 0 for every x 2 U it
follows that U = [1 j=1 Fj :

Exercises

1. Suppose f : (X; M) ! (Rd ; Rd ) and g : (X; M) ! (Rn ; Rn ) are measur-

able. Set h(x) = (f (x); g(x)) 2 Rd+n if x 2 X. Prove that h : (X; M) !
(Rd+n ; Rd+n ) is measurable.

2. Suppose f : (X; M) ! (R; R) and g : (X; M) ! (R; R) are measurable.

Prove that f g is (M; R)-measurable.

3. The function f : R ! R is a Borel function. Set g(x; y) = f (x); (x; y) 2

R2 : Prove that g : R2 ! R is a Borel function.

4. Suppose f : [0; 1] ! R is a continuous function and g : [0; 1] ! [0; 1] a

Borel function. Compute the limit
Z 1
lim f (g(x)n )dx:
n!1 0

5. Suppose X and Y are metric spaces and f : X ! Y a continuous mapping.

Show that f (E) is compact if E is a compact subset of X.
86

6. Suppose X and Y are metric spaces and f : X ! Y a continuous bijection.

Show that the inverse mapping f 1 is continuous if X is compact.

7. Construct an open bounded subset V of R such that m(@V ) > 0:

8. The function f : [0; 1] !R has a continuous derivative. Prove that the

set f (K) 2 Zm if K = (f 0 ) 1 (f0g):

9. Let P denote the class of all Borel probability measures on [0; 1] and L
the class of all functions f : [0; 1] ! [ 1; 1] such that

j f (x) f (y) j j x y j; x; y 2 [0; 1] :

For any ; 2 P; de…ne

Z Z
( ; ) = sup j fd fd j :
f 2L [0;1] [0;1]

(a) Show that (P; ) is a metric space. (b) Compute ( ; ) if is linear

measure on [0; 1] and = n1 nk=01 k ; where n 2 N+ (linear measure on [0; 1]
n
is Lebesgue measure on [0; 1] restricted to the Borel sets in [0; 1]).

10. Suppose is a …nite positive Borel measure on Rn : (a) Let (Vi )i2I be a
family of open subsets of Rn and V = [i2I Vi . Prove that

(V ) = sup (Vi1 [ ::: [ Vik ):

i1 ;:::;ik 2I
k2N+

(b) Let (Fi )i2I be a family of closed subsets of Rn and F = \i2I Fi . Prove
that
(F ) = inf (Fi1 \ ::: \ Fik ):
i1 ;:::;ik 2I
k2N+
 87

###
3.2. Linear Functionals and Measures

Let X be a metric space. A mapping T : Cc (X) ! R is said to be a linear

functional on Cc (X) if

T (f + g) = T f + T g; all f; g 2 Cc (X)

and
T ( f ) = T f; all 2 R; f 2 Cc (X):
If in addition T f 0 for all f 0; T is called a positive linear functional
on Cc (X): In this case T f T g if f g since g f 0 and T g T f =
T (g f ) 0: Note that Cc (X) = C(X) if X is compact.
The main result in this section is the following

Theorem 3.2.1. (The Riesz Representation Theorem) Suppose X is

a compact metric space and let T be a positive linear functional on C(X):
Then there exists a unique …nite positive Borel measure in X with the
following properties:
(a) Z
Tf = f d ; f 2 C(X):
X

(b) For every E 2 B(X)

(E) = sup (K):

K E
K compact

(c) For every E 2 B(X)

(E) = inf (V ):
V E
V open
88

The property (c) is a consequence of (b), since for each E 2 B(X) and
" > 0 there is a compact K X n E such that

(X n E) < (K) + ":

But then
(X n K) < (E) + "
and X n K is open and contains E: In a similar way, (b) follows from (c)
since X is compact.
The proof of the Riesz Representation Theorem depends on properties of
continuous functions of independent interest. Suppose K X is compact
and V X is open. If f : X ! [0; 1] is a continuous function such that

f V and suppf V

we write
f V
and if
K f V and suppf V
we write
K f V:

Theorem 3.2.2. Let K be compact subset X.

(a) Suppose K V where V is open. There exists a function f on X
such that
K f V:

(b) Suppose X is compact and K V1 [ ::: [ Vn ; where K is compact and

V1 ; :::; Vn are open. There exist functions h1 ; :::; hn on X such that

hi Vi ; i = 1; :::; n

and
h1 + ::: + hn = 1 on K:
 89

PROOF. (a) Suppose " = 12 minK d( ; V c ): By Corollary 3.1.2, " > 0: The
continuous function f = X K;" satis…es K f K" ; that is K f K" :
Part (a) follows if we note that the closure (K" ) of K" is contained in V:

(b) For each x 2 K there exists an rx > 0 such that B(x; rx ) Vi for some
i. Let Ux = B(x; 21 rx ): It is important to note that (Ux ) Vi and (Ux )
is compact since X is compact. There exist points x1 ; :::; xm 2 K such that
[mj=1 Uxi K: If 1 i n; let Fi denote the union of those (Uxj ) which are
contained in Vi : By Part (a), there exist continuous functions fi such that
Fi fi Vi ; i = 1; :::; n: De…ne

h1 = f1
h2 = (1 f1 )f2
::::
hn = (1 f1 ):::(1 fn 1 )fn :

Clearly, hi Vi ; i = 1; :::; n: Moreover, by induction, we get

h1 + ::: + hn = 1 (1 f1 ):::(1 fn 1 )(1 fn ):

Since [ni=1 Fi K we are done.

The uniqueness in Theorem 3.2.1 is simple to prove. Suppose 1 and

2 are two measures for which the theorem holds. Fix " > 0 and compact
K X and choose an open set V so that 2 (V ) 2 (K) + ": If K f V;
Z Z
1 (K) = Kd 1 fd 1 = Tf
X X
Z Z
= fd 2 Vd 2 = 2 (V ) 2 (K) + ":
X X

Thus 1 (K) 2 (K): If we interchange the roles of the two measures, the
opposite inequality is obtained, and the uniqueness of follows.
To prove the existence of the measure in Theorem 3.2.1; de…ne for every
open V in X,
(V ) = sup T f:
f V
90

Here ( ) = 0 since the supremum over the empty set, by convention, equals
0: Note also that (X) = T 1: Moreover, (V1 ) (V2 ) if V1 and V2 are open
and V1 V2 : Now set

(E) = inf (V ) if E 2 B(X):

V E
V open

Clearly, (E1 ) (E2 ); if E1 E2 and E1; E2 2 B(X): We therefore say

that is increasing.

Lemma 3.2.1. (a) If V1 ; :::; Vn are open,

([ni=1 Vi ) n
i=1 (Vi ):

(b) If E1 ; E2 ; ::: 2 B(X);

([1
i=1 Ei )
1
i=1 (Ei ):

(c) If K1 ; :::; Kn are compact and pairwise disjoint,

([ni=1 Ki ) = n
i=1 (Ki ):

PROOF. (a) It is enough to prove (a) for n = 2: To this end …rst choose
g V1 [ V2 and then hi Vi , i = 1; 2; such that h1 + h2 = 1 on supp g: Then

g = h1 g + h2 g

and it follows that

T g = T (h1 g) + T (h2 g) (V1 ) + (V2 ):

Thus
(V1 [ V2 ) (V1 ) + (V2 ):
 91

(b) Choose " > 0 and for each i 2 N+ , choose an open Vi Ei such (Vi ) <
(Ei ) + 2 i ": Set V = [1
i=1 Vi and choose f V: Since suppf is compact,
f V1 [ ::: [ Vn for some n: Thus, by Part (a),
n 1
Tf (V1 [ ::: [ Vn ) i=1 (Vi ) i=1 (Ei ) + "

and we get
1
(V ) i=1 (Ei )
since " > 0 is arbitrary. But [1
i=1 Ei V and it follows that

([1
i=1 Ei )
1
i=1 (Ei ):

(c) It is enough to treat the special case n = 2: Choose " > 0: Set =
d(K1 ; K2 ) and V1 = (K1 ) =2 and V2 = (K2 ) =2 : There is an open set U
K1 [ K2 such that (U ) < (K1 [ K2 ) + " and there are functions fi U \ Vi
such that T fi > (U \ Vi ) " for i = 1; 2: Now, using that increases

(K1 ) + (K2 ) (U \ V1 ) + (U \ V2 )

T f1 + T f2 + 2" = T (f1 + f2 ) + 2":

Since f1 + f2 U;

(K1 ) + (K2 ) (U ) + 2" (K1 [ K2 ) + 3"

and, by letting " ! 0;

(K1 ) + (K2 ) (K1 [ K2 ):

The reverse inequality follows from Part (b). The lemma is proved.

Next we introduce the class

8 9
< =
M = E 2 B(X); (E) = sup (K)
: K E ;
K compact
92

Since is increasing M contains every compact set. Recall that a closed

set in X is compact, since X is compact. Especially, note that and X 2 M.

COMPLETION OF THE PROOF OF THEOREM 3.2.1:

CLAIM 1. M contains every open set.

PROOF OF CLAIM 1. Let V be open and suppose < (V ): There exists

an f V such that < T f: If U is open and U K =def suppf; then f U;
and hence T f (U ): But then T f (K): Thus < (K) and Claim 1
follows since K is compact and K V:

CLAIM 2. Let (Ei )1

i=1 be a disjoint denumerable collection of members of
1
M and put E = [i=1 Ei : Then
1
(E) = i=1 (Ei )

and E 2 M:

PROOF OF CLAIM 2. Choose " > 0 and for each i 2 N+ , choose a compact
Ki Ei such that (Ki ) > (Ei ) 2 i ": Set Hn = K1 [ ::: [ Kn : Then, by
Lemma 3.2.1 (c),
n n
(E) (Hn ) = i=1 (Ki ) > i=1 (Ei ) "

and we get
1
(E) i=1 (Ei ):
Thus, by Lemma 3.2.1 (b), (E) = 1 i=1 (Ei ). To prove that E 2 M; let "
be as in the very …rst part of the proof and choose n such that
n
(E) i=1 (Ei ) + ":
 93

Then
(E) < (Hn ) + 2"
and this shows that E 2 M:

CLAIM 3. Suppose E 2 M and " > 0: Then there exist a compact K and
an open V such that K E V and (V n K) < ":

PROOF OF CLAIM 3. The de…nitions show that there exist a compact K

and an open V such that
" "
(V ) < (E) < (K) + :
2 2
The set V n K is open and V n K 2 M by Claim 1. Thus Claim 2 implies
that
(K) + (V n K) = (V ) < (K) + "
and we get (V n K) < ":

CLAIM 4. If A 2 M; then X n A 2 M:

PROOF OF CLAIM 4. Choose " > 0: Furthermore, choose compact K A

and open V A such that (V n K) < ": Then

X nA (V n K) [ (X n V ):

Now, by Lemma 3.2.1 (b),

(X n A) " + (X n V ):

Since X n V is a compact subset of X n A; we conclude that X n A 2 M:

Claims 1, 2 and 4 prove that M is a -algebra which contains all Borel

sets. Thus M = B(X):
94

We …nally prove (a). It is enough to show that

Z
Tf fd
X

for each f 2 C(X): For once this is known

Z Z
Tf = T( f) fd fd
X X

and (a) follows.

Choose " > 0: Set f (X) = [a; b] and choose y0 < y1 < ::: < yn such that
y1 = a, yn 1 = b; and yi yi 1 < ": The sets
1
Ei = f ([yi 1 ; yi [); i = 1; :::; n

constitute a disjoint collection of Borel sets with the union X: Now, for each i;
pick an open set Vi Ei such that (Vi ) (Ei )+ n" and Vi f 1 (] 1; yi [):
By Theorem 3.2.2 there are functions hi Vi ; i = 1; :::; n; such that ni=1 hi =
1 on suppf and hi f yi hi for all i: From this we get
n n n
Tf = i=1 T (hi f ) i=1 yi T hi i=1 yi (Vi )
n " n
i=1 yi (Ei ) + i=1 yi
n
n
i=1 (yi ") (Ei ) + " (X) + (b + ")"
Z
n
i=1 f d + " (X) + (b + ")"
Ei
Z
= f d + " (X) + (b + ")":
X
Since " > 0 is arbitrary, we get
Z
Tf fd :
X

This proves Theorem 3.2.1.

It is now simple to show the existence of volume measure in Rn : For

pedagogical reasons we …rst discuss the so called volume measure in the unit
cube Q = [0; 1]n in Rn :
 95

The Riemann integral Z

f (x)dx;
Q

is a positive linear functional as a function of f 2 C(Q): Moreover, T 1 = 1

and the Riesz Representation Theorem gives us a Borel probability measure
in Q such that Z Z
f (x)dx = fd :
Q Q

Suppose A Q is a closed n-cell and i 2 N+ : Then

Z
Q
vol(A) A;2 i
(x)dx vol(A2 i )
Q

and
Q
A;2 i (x) ! A (x) as i ! 1
for every x 2Rn : Thus
(A) = vol(A):
The measure is called the volume measure in the unit cube. In the special
case n = 2 it is called the area measure in the unit square and if n = 1 it is
called the linear measure in the unit interval.

PROOF OF THEOREM 1.1.1. Let R̂=R[ f 1; 1g be the two-point com-

^ n denote the product
pacti…cation of R introduced in Example 3.1.3 and let R
^ Clearly,
of n copies of the metric space R:
n o
n n
B(R ) = A \ R ; A 2 B(R ) :^ n

Moreover, let w : Rn ! ]0; 1[ be a continuous map such that

Z
w(x)dx = 1:
Rn

Now we de…ne Z
Tf = ^ n ):
f (x)w(x)dx; f 2 C(R
Rn
96

^ n)
Note that T 1 = 1. The function T is a positive linear functional on C(R
and the Riesz Representation Theorem gives us a Borel probability measure
^ n such that
on R
Z Z
f (x)w(x)dx = f d ; f 2 C(R ^ n ):
Rn ^n
R

As above we get Z
w(x)dx = (A)
A
for each compact n-cell in Rn : Thus
Z
n
(R ) = lim w(x)dx = 1
i!1 [ i;i]n

and we conclude that is concentrated on Rn : Set 0 (A) = (A); A 2

B(Rn ); and
1
dmn = d :
w 0
Then, if f 2 Cc (Rn );
Z Z
f (x)w(x)dx = fd 0
Rn Rn

and by replacing f by f =w;

Z Z
f (x)dx = f dmn :
Rn Rn

From this mn (A) =vol(A) for every compact n-cell A and it follows that mn
is the volume measure on Rn . Theorem 1.1.1 is proved.

"""

3.3 q-Adic Expansions of Numbers in the Unit Interval

To begin with in this section we will discuss so called q-adic expansions of

real numbers and give some interesting consequences. As an example of an
 97

application, we construct a one-to-one real-valued Borel map f de…ned on

a proper interval such that the range of f is a Lebesgue null set. Another
example exhibits an increasing continuous function G on the unit interval
with the range equal to the unit interval such that the derivative of G is
equal to zero almost everywhere with respect to Lebesgue measure. In the
next section we will give more applications of q-adic expansions in connection
with in…nite product measures.
To simplify notation let ( ; P; F) = ([0; 1[ ; v1j[0;1[ ; B([0; 1[)). Furthermore,
let q 2 be an integer and de…ne a function h : R ! f0; 1; 2; :::; q 1g of
period one such that
k k+1
h(x) = k; x< ; k = 0; :::; q 1:
q q
Furthermore, set for each n 2 N+ ;

n (!) = h(q n 1 !); 0 ! < 1:

Then
1
P[ n = k] = ; k = 0; :::; q 1:
q
Moreover, if k1 ; :::; kn 2 f0; 1; 2; :::; q 1g ; it becomes obvious on drawing a
…gure that
q 1
P 1 = k1 ; :::; n 1 = kn 1 = i=0 P 1 = k1 ; :::; n 1 = kn 1 ; n =i

where each term in the sum in the right-hand side has the same value. Thus

P 1 = k1 ; :::; n 1 = kn 1 = qP 1 = k1 ; :::; n 1 = kn 1 ; n = kn

and

P 1 = k1 ; :::; n 1 = kn 1 ; n = kn = P 1 = k1 ; :::; n 1 = kn 1 P[ n = kn ] :

By repetition,
n
P 1 = k1 ; ::: n 1 = kn 1 ; n = kn = i=1 P [ i = ki ] :

From this we get

n
P 1 2 A1 ; ::: n 1 2 An 1 ; n 2 An = i=1 P [ i 2 Ai ]
98

for all A1 ; :::; An f0; 1; 2; :::; q 1g :

Note that each ! 2 [0; 1[ has a so called q-adic expansion

1 i (!)
!= i=1 :
qi

If necessary, we write n = (q) n to indicate q explicitly.

Let k0 2 f0; 1; 2; :::; q 1g be …xed and consider the event A that a num-
ber in [0; 1[ does not have k0 in its q-adic expansion. The probability of A
equals

P [A] = P [ i 6= k0 ; i = 1; 2; :::] = lim P [ i 6= k0 ; i = 1; 2; :::; n]

n!1

n q 1
= lim i=1 P [ i 6= k0 ] = lim ( )n = 0:
n!1 n!1 q
In particular, if
n o
(3)
Dn = ! 2 [0; 1[ ; i 6= 1; i = 1; :::; n :

then, D = \1
n=1 Dn is a P -zero set.
Set
(2)
2 i (!)
f (!) = 1 i=1 ; 0 ! < 1:
3i
We claim that f is one-to-one. If 0 !; ! 0 < 1 and ! 6= ! 0 let n be the
(2) (2)
least i such that i (!) 6= i (! ); we may assume that (2)
0
n (!) = 0 and
(2) 0
n (! ) = 1: Then

(2) 0 (2) 0
0 n 2 i (! ) n 1 2 i (! ) 2
f (! ) i=1 i
= i=1 i
+
3 3 3n
(2) (2)
n 1 2 i (!) 1 4 1 2 i (!)
= i=1 + i=n+1 i > i=1 = f (!):
3i 3 3i
Thus f is one-to-one. We next prove that f ( ) = D: To this end choose
(3)
y 2 D: If i (y) = 2 for all i 2 N+ ; then y = 1 which is a contradiction. If
(3) (3)
k 1 is …xed and k (y) = 0 and i (y) = 2; i k + 1; then it is readily
(3)
seen that k (y) = 1 which is a contradiction. Now de…ne
1 (3)
1 2 i (y)
!= i=1
2i
 99

and we have f (!) = y:

Let Cn = Dn ; n 2 N+ : The set C = \1 n=1 Cn ; is called the Cantor set.
The Cantor set is a compact Lebesgue zero set. The construction of the
Cantor set may alternatively be described as follows. First C0 = [0; 1]. Then
trisect C0 and remove the middle interval 13 ; 23 to obtain C1 = C0 n 13 ; 23 =
0; 31 [ 23 ; 1 : At the second stage subdivide each of the closed intervals
of C1 into thirds and remove from each one the middle open thirds. Then
C2 = C1 n ( 19 ; 29 [ 79 ; 89 ): What is left from Cn 1 is Cn de…ned above. The
set [0; 1] n Cn is the union of 2n 1 intervals numbered Ikn ; k = 1; :::; 2n 1;
where the interval Ikn is situated to the left of the interval Iln if k < l:
Suppose n is …xed and let Gn : [0; 1] ! [0; 1] be the unique monotone in-
creasing continuous function, which satis…es Gn (0) = 0; Gn (1) = 1; Gn (x) =
k2 n for x 2 Ikn and which is a¢ ne on each interval of Cn: It is clear that
Gn = Gn+1 on each interval Ikn , k = 1; :::; 2n 1: Moreover, j Gn Gn+1 j
2 n 1 and thus
n+k n
j Gn Gn+k j k=n j Gk Gk+1 j 2 :
Let G(x) = limn!1 Gn (x); 0 x 1: The continuous and increasing func-
tion G is constant on each removed interval and it follows that G0 = 0 a.e.
with respect to linear measure in the unit interval.The function G is called
the Cantor function or Cantor-Lebesgue function.
Next we introduce the following convention, which is standard in Lebesgue
integration. Let (X; M; ) be a positive measure space and suppose A 2 M
and (Ac ) = 0: If two functions g; h 2 L1 ( ) agree on A;
Z Z
gd = hd :
X X

If a function f : A ! R is the restriction to A of a function g 2 L1 ( ) we

de…ne Z Z
fd = gd :
X X
Now suppose F : R ! R is a right continuous increasing function and
let be the unique positive Borel such that
(]a; x]) = F (x) F (a) if a; x 2 R and a < x:
If h 2 L1 ( ) and E 2 R; the so called Stieltjes integral
Z
h(x)dF (x)
E
100

is by de…nition equal to Z
hd :
E
If a; b 2 R, a < b; and F is continuous at the points a and b; we de…ne
Z b Z
h(x)dF (x) = hd
a I

where I is any interval with boundary points a and b:

The reader should note that the integral
Z
h(x)dF (x)
R

in general is di¤erent from the integral

Z
h(x)F 0 (x)dx:
R

For example, if G is the Cantor function and G is extended so that G(x) = 0

for negative x and G(x) = 1 for x larger than 1, clearly
Z
h(x)G0 (x)dx = 0
R

since G0 (x) = 0 a.e. [m] : On the other hand, if we choose h = [0;1] ;

Z
h(x)dG(x) = 1:
R

3.4. Product Measures

Suppose (X; M) and (Y; N ) are two measurable spaces. If A 2 M and

B 2 N ; the set A B is called a measurable rectangle in X Y: The product
-algebra M N is, by de…nition, the -algebra generated by all measurable
rectangles in X Y: If we introduce the projections

X (x; y) = x; (x; y) 2 X Y

and
Y (x; y) = y; (x; y) 2 X Y;
 101

the product -algebra M N is the least -algebra S of subsets of X Y ,

which makes the maps X : (X Y; S) ! (X; M) and Y : (X Y; S) !
(Y; N ) measurable, that is M N = ( X1 (M)[ Y 1 (N )):.
Suppose E generates M; where X 2 E; and F generates N ; where Y 2 F.
We claim that the class

E F = fE F ; E 2 E and F 2 Fg

generates the -algebra M N : First it is clear that

(E F) M N:

Moreover, the class

1
fE 2 M; E Y 2 (E F) g = M\ E X; X (E) 2 (E F)

is a -algebra, which contains E and therefore equals M. Thus A Y 2

(E F) for all A 2 M and, in a similar way, X B 2 (E F) for all
B 2 N and we conclude that A B = (A Y ) \ (X B) 2 (E F) for
all A 2 M and all B 2 N : This proves that

M N (E F)

and it follows that

(E F) = M N:
Thus
(E F) = (E) (F) if X 2 E and Y 2 F:
Since the -algebra Rn is generated by all open n-cells in Rn , we conclude
that
Rk+n = Rk Rn :
Given E X Y; de…ne

Ex = fy; (x; y) 2 Eg if x 2 X

and
E y = fx; (x; y) 2 Eg if y 2 Y:
If f : X Y ! Z is a function and x 2 X; y 2 Y , let

fx (y) = f (x; y); if y 2 Y

102

and
f y (x) = f (x; y); if x 2 X:

Theorem 3.4.1 (a) If E 2 M N ; then Ex 2 N and E y 2 M for every

x 2 X and y 2 Y:
(b) If f : (X Y; M N ) ! (Z; O) is measurable, then fx is (N ; O)-
measurable for each x 2 X and f y is (M; O)-measurable for each y 2 Y:

Proof. (a) Let

S = fE 2 M N ; Ex 2 N and E y 2 M for every x 2 X and y 2 Y g :

Clearly, X Y 2 S: Furthermore, if E; E1; E2 ; ::: 2 S; (E c )x = (Ex )c 2 N

and ([1 1 c y y c
i=1 Ei )x = [i=1 (Ei )x 2 N for every x in X and (E ) = (E ) 2 M
1 y 1 y
and ([i=1 Ei ) = [i=1 (Ei ) 2 M for every y in Y: It follows that S is a
-algebra. Furthermore, if A 2 M and B 2 N ; (A B)x = B 2 N if x 2 A
and (A B)x = 2 N if x 2 = A and (A B)y = A 2 M if y 2 B and
y
(A B) = 2 M if y 2 = B: Thus A B 2 S and, accordingly from this,
S = M N:
(b) For any set V 2 O;
1
(f (V ))x = (fx ) 1 (V )

and
1
(f (V ))y = (f y ) 1 (V ):
Part (b) now follows from (a).

Below an (M; R0;1 )-measurable or (M; R)-measurable function is simply

called M-measurable.

Theorem 3.4.2. Suppose (X; M; ) and (Y; N ; ) are positive -…nite

measurable spaces and suppose E 2 M N . If

f (x) = (Ex ) and g(y) = (E y )

 103

for every x 2 X and y 2 Y; then f is M-measurable, g is N -measurable,

and Z Z
fd = gd :
X Y

Proof. We …rst assume that (X; M; ) and (Y; N ; ) are …nite positive
measure spaces.
Let D be the class of all sets E 2 M N for which the conclusion of
the theorem holds. It is clear that the class G of all measurable rectangles
in X Y is a subset of D and G is a -system. Furthermore, the Beppo
Levi Theorem shows that D is a -system. Therefore, using Theorem 1.2.2,
M N = (G) D and it follows that D = M N :
In the general case, choose a denumerable disjoint collection (Xk )1k=1 of
1
members of M and a denumerable disjoint collection (Yn )n=1 of members of
N such that
[1 1
k=1 Xk = X and [n=1 Yn = Y:

Set
k = Xk , k = 1; 2; :::
and
n = Yn , n = 1; 2; ::: .
Then, by the Beppo Levi Theorem, the function
Z
1
f (x) = n=1 E (x; y) Yn (y)d (y)
X
Z
1 1
= n=1 E (x; y) Yn (y)d (y) = n=1 n (Ex )
X
is M-measurable. Again, by the Beppo Levi Theorem,
Z Z
1
f d = k=1 fd k
X X

and
Z Z Z
1 1 1
fd = k=1 ( n=1 n (Ex )d k (x)) = k;n=1 n (Ex )d k (x):
X X X

In a similar way, the function g is N -measurable and

Z Z Z
1 1 y 1 y
gd = n=1 ( k=1 k (E )d n (y)) = k;n=1 k (E )d n (y):
Y Y Y
104

Since the theorem is true for …nite positive measure spaces, the general case
follows.

De…nition 3.4.1. If (X; M; ) and (Y; N ; ) are positive -…nite measur-

able spaces and E 2 M N , de…ne
Z Z
( )(E) = (Ex )d (x) = (E y )d (y):
X Y

The function is called the product of the measures and :

Note that Beppo Levi’s Theorem ensures that is a positive measure.

Before the next theorem we recall the following convention. Let (X; M; )
be a positive measure space and suppose A 2 M and (Ac ) = 0: If two
functions g; h 2 L1 ( ) agree on A;
Z Z
gd = hd :
X X

If a function f : A ! R is the restriction to A of a function g 2 L1 ( ) we

de…ne Z Z
fd = gd :
X X

Theorem 3.4.3. Let (X; M; ) and (Y; N ; ) be positive -…nite measur-

able spaces.

(a) (Tonelli’s Theorem) If h : X Y ! [0; 1] is (M N )-measurable

and Z Z
f (x) = h(x; y)d (y) and g(y) = h(x; y)d (x)
Y X

for every x 2 X and y 2 Y; then f is M-measurable, g is N -measurable,

and Z Z Z
fd = hd( )= gd
X X Y Y

(b) (Fubini’s Theorem)

 105

(i) If h : X Y ! R is (M N )-measurable and

Z Z
( j h(x; y) j d (y))d (x) < 1
X Y

then h 2 L1 ( ): Moreover,
Z Z Z Z Z
( h(x; y)d (y))d (x) = hd( ) = ( h(x; y)d (x))d (y)
X Y X Y Y X

(ii) If h 2 L1 (( ) ); then hx 2 L1 ( ) for -almost all x and

Z Z Z
hd( )= ( h(x; y)d (y))d (x)
X Y X Y

(iii) If h 2 L1 (( ) ); then hy 2 L1 ( ) for -almost all y and

Z Z Z
hd( ) = ( h(x; y)d (x))d (y)
X Y Y X

PROOF. (a) The special case when h is a non-negative (M N )-measurable

simple function follows from Theorem 3.4.2. Remembering that any non-
negative measurable function is the pointwise limit of an increasing sequence
of simple measurable functions, the Lebesgue Monotone Convergence Theo-
rem implies the Tonelli Theorem.

(b) PART (i) : By Part (a)

Z Z Z
+
1> ( h (x; y)d (y))d (x) = h+ d( )
X Y X Y
Z Z
= ( h+ (x; y)d (x))d (y)
Y X

and Z Z Z
1> ( h (x; y)d (y))d (x) = h d( )
X Y X Y
Z Z
= ( h (x; y)d (x))d (y):
Y X
106

Let
A = x 2 X; (h+ )x ; (h )x 2 L1 ( ) :
Then Ac is a -null set and we get
Z Z Z
+
( h (x; y)d (y))d (x) = h+ d( )
A Y X Y

and Z Z Z
( h (x; y)d (y))d (x) = h d( ):
A Y X Y

Thus Z Z Z
( h(x; y)d (y))d (x) = hd( )
A Y X Y

and, hence,
Z Z Z
( h(x; y)d (y))d (x) = hd( ):
X Y X Y

The other case can be treated in a similar way. The theorem is proved.

PART (ii) : We …rst use Theorem 2.2.3 and write h = ' + where ' 2
L1 ( ); is (M N ) -measurable and = 0 a.e. [ ] : Set

A = x 2 X; ('+ )x ; (' )x 2 L1 ( ) :

Furthermore, suppose E f(x; y); (x; y) 6= 0g ; E 2 M N and

( )(E) = 0:

Then, by Tonelli’s Theorem

Z
0= (Ex )d (x):
X

Let B = fx 2 X; (Ex ) 6= 0g and note that B 2 M: Moreover (B) = 0

and if x 2
= B, then x = 0 a.e. [ ] that is hx = 'x a.e. [ ] : Now, by Part (i)
Z Z Z Z
hd( ) = 'd( ) = ( '(x; y)d (y))d (x)
X Y X Y A Y
 107
Z Z Z Z
= ( '(x; y)d (y))d (x) = ( h(x; y)d (y))d (x)
A\B c Y A\B c Y
Z Z
= ( h(x; y)d (y))d (x):
X Y

Part (iii) is proved in the same manner as Part (ii): This concludes the
proof of the theorem.

If (Xi ; Mi ); i = 1; :::; n; are measurable spaces, the product -algebra

M1 ::: Mn is, by de…nition, the -algebra generated by all sets of the
form
A1 ::: An
where Ai 2 Mi ; i = 1; :::; n: Now assume (Xi ; Mi ; i ); i = 1; :::; n; are -…nite
positive measure spaces. By induction, we de…ne 1 = 1 and k = k 1 k;
k = 1; 2; :::; n: The measure, n is called the product of the measures 1 ; :::; n
and is denoted by 1 ::: n : It is readily seen that

Rn = R1 ::: R1 (n factors)

and
vn = v1 ::: v1 (n factors):
Moreover,

Rn (R1 )n =def R1 ::: R1 (n factors):

If A 2 P(R) n R1 ; by the Tonelli Theorem, the set A f0; :::; 0g (n 1

zeros) is an mn -null set, which, in view of Theorem 3.4.1, cannot belong to
the -algebra (R1 )n : Thus the Axiom of Choice implies that

Rn 6= (R1 )n :

Clearly, the completion of the measure m1 ::: m1 (n factors) equals

mn :
Sometimes we prefer to write
Z
f (x1 ; :::; xn )dx1 :::dxn
A1 ::: An
108

instead of Z
f (x)dmn (x)
A1 ::: An
or Z
f (x)dx:
A1 ::: An

Moreover, the integral

Z Z
::: f (x1 ; :::; xn )dx1 :::dxn
A1 An

is the same as Z
f (x1 ; :::; xn )dx1 :::dxn :
A1 ::: An

De…nition 3.4.2. (a) The measure

Z
x2 dx
1 (A) = e 2 p ; A2R
A 2
is called the canonical Gauss measure in R:

(b) The measure

n = 1 ::: 1 (n factors)

is called the canonical Gauss measure in Rn : Thus, if

q
j x j= x21 + ::: + x2n ; x = (x1 ; :::; xn ) 2 Rn

we have Z
jxj2 dx
n (A) = e 2 p ; A 2 Rn :
A 2

(c) A Borel measure in R is said to be a centred Gaussian measure if

= f ( 1 ) for some linear map f : R ! R:
 109

(d) A real-valued random variable is said to be a centred Gaussian

random variable if its probability law is a centred Gaussian measure in R:
Stated otherwise, is a real-valued centred Gaussian random variable if either

L( ) = 0 (abbreviated 2 N (0; 0))

or there exists a > 0 such that

L( ) = 1 (abbreviated 2 N (0; )).

(e) A family ( t )t2T of real-valued random variables is said to be a centred

real-valued Gaussian process if for all t1 ; :::; tn 2 T; 1 ; :::; n 2 R and every
n 2 N+ ; the sum
= nk=1 k tk
is a centred Gaussian random variable:

Exercises

1. Let (X; M; ) and (Y; N ; ) be two -…nite measure spaces. Let f 2 L1 ( )

and g 2 L1 ( ) and de…ne h(x; y) = f (x)g(y); (x; y) 2 X Y: Prove that
h 2 L1 ( ) and
Z Z Z
hd( )= fd gd :
X Y X Y

2. Let (X; M; ) be a -…nite measure space and f : X ! [0; 1[ a measur-

able function. Prove that
Z
fd = ( m) f(x; y); 0 < y < f (x); x 2 Xg :
X

3. Let (X; M; ) be a -…nite measure space and f : X ! R a measurable

function. Prove that ( m)(f(x; f (x)); x 2 X g) = 0:
110

1
4. Let E 2 R2 and E [0; 1] [0; 1] : Suppose m(Ex ) 2
for m-almost all
x 2 [0; 1] : Show that

1
m(fy 2 [0; 1] ; m(E y ) = 1g) :
2

5. Let c be the counting measure on R restricted to R and

D = f(x; x); x 2 Rg :

De…ne for every A 2 (R R) [ fDg ;

Z Z
(A) = ( A (x; y)dv1 (x))dc(y)
R R

and Z Z
(A) = ( A (x; y)dc(y))dv1 (x):
R R

(a) Prove that and agree on R R:

(b) Prove that (D) 6= (D):

6. Let I = ]0; 1[ and

x2 y 2
h(x; y) = ; (x; y) 2 I I:
(x2 + y 2 )2

Prove that Z Z
( h(x; y)dy)dx = ;
I I 4
Z Z
( h(x; y)dx)dy =
I I 4
and Z
j h(x; y) j dxdy = 1:
I I

7. For t > 0 and x 2 R let

1 x2
g(t; x) = p e 2t
2 t
 111

and
@g
h(t; x) = :
@t
Given a > 0; prove that
Z 1 Z 1
( h(t; x)dt)dx = 1
1 a
and Z 1 Z 1
( h(t; x)dx)dt = 0
a 1
and conclude that Z
j h(t; x) j dtdx = 1:
[a;1[ R

(Hint: First prove that Z 1

g(t; x)dx = 1
1
and
@g 1 @2g
= :)
@t 2 @x2

8. Given f 2 L1 (m); let

Z x+1
1
g(x) = f (t)dt; x 2 R:
2 x 1

Prove that Z Z
j g(x) j dx j f (x) j dx:
R R
9. Let I = [0; 1] and suppose f : I ! R is a Lebesgue measurable function
such that Z
j f (x) f (y) j dxdy < 1:
I I
Prove that Z
j f (x) j dx < 1:
I
10. Suppose A 2 R and f 2 L1 (m): Set
Z
d(y; A)f (y)
g(x) = dy; x 2 R:
R j x y j2
112

Prove that Z
j g(x) j dx < 1:
A

11. Suppose that the functions f; g : R ! [0; 1[ are Lebesgue measurable

and introduce = f m and = gm: Prove that the measures and are
-…nite and
Z
( )(E) = f (x)g(y)dxdy if E 2 R R :
E

12. Suppose is a …nite positive Borel measure on Rn and f : Rn ! R a

Borel measurable function. Set g(x; y) = f (x) f (y); x; y 2 Rn : Prove that
f 2 L1 ( ) if and only if g 2 L1 ( ):

13. A random variable is non-negative and

R 1possesses the distribution func-
tion F (x) = P [ x] : Prove thatE [ ] = 0 (1 F (x))dx:

14. Let (X; d) be a metric space and suppose Y 2 B(X): Then Y equipped
with the metric djY Y is a metric space. Prove that

B(Y ) = fA \ Y ; A 2 B(X)g :

15. The continuous bijection f : (X; d) ! (Y; e) has a continuous inverse.

Prove that f (A) 2 B(Y ) if A 2 B(X)

16. A real-valued function f (x; y); x; y 2 R; is a Borel function of x for every

…xed y and a continuous function of y for every …xed x: Prove that f is a
Borel function. Is the same conclusion true if we only assume that f (x; y) is
a real-valued Borel function in each variable separately?
 113

17. Suppose a > 0 and

X
1
an
a
a =e n
n=0
n!
where n (A) = A (n) if n 2 N = f0; 1; 2; :::g and A N: Prove that
1
( a b )s = a+b

for all a; b > 0; if s(x; y) = x + y; x; y 2 N:

18. Suppose
Z 1 2
xe x
f (t) = dx; t > 0:
0 x2 + t2
Compute Z 1
lim f (t) and f (t)dt:
t!0+ 0
Finally, prove that f is di¤erentiable.

3.5 Change of Variables in Volume Integrals

If T is a non-singular n by n matrix with real entries, we claim that

1
T (vn ) = vn
j det T j

(here T is viewed as a linear map of Rn into Rn ). Remembering Corollary

3.1.3 this means that the following linear change of variables formula holds,
viz. Z Z
1
f (T x)dx = f (x)dx all f 2 Cc (Rn ):
Rn j det T j Rn
The case n = 1 is obvious. Moreover, by Fubini’s Theorem the linear change
of variables formula is true for arbitrary n in the following cases:
(a) T x = (x (1) ; :::; x (n) ); where is a permutation of the numbers 1; :::; n:
(b) T x = ( x1 ; x2 ; :::; xn ); where is a non-zero real number.
114

(c) T x = (x1 + x2 ; x2 ; :::; xn ):

Recall from linear algebra that every non-singular n by n matrix T can be
row-reduced to the identity matrix, that is T can by written as the product
of …nitely many transformations of the types in (a),(b), and (c). This proves
the above linear change of variables formula.
Our main objective in this section is to prove a more general change
of variable formula. To this end let and be open subsets of Rn and
G : ! a C 1 di¤eomorphism, that is G = (g1 ; :::; gn ) is a bijective
@gi
continuously di¤erentiable map such that the matrix G0 (x) = ( @x j
(x))1 i;j n
is non-singular for each x 2 : The inverse function theorem implies that
G 1 : ! is a C 1 di¤eomorphism [DI] :

Theorem 3.5.1. If f is a non-negative Borel function in ; then

Z Z
f (x)dx = f (G(x)) j det G0 (x) j dx:

The proof of Theorem 3.5.1 is based on several lemmas.

Throughout, Rn is equipped with the metric

dn (x; y) = max j xk yk j :
1 k n

Let K be a compact convex subset of : Then if x; y 2 K and 1 i n;

Z 1
d
gi (x) gi (y) = gi (y + t(x y))dt
0 dt
Z 1
n @gi
= k=1 (y + t(x y))(xk yk )dt
0 @xk
and we get
dn (G(x); G(y)) M (G; K)dn (x; y)
where
n @gi
M (G; K) = max k=1 max j (z) j :
1 i n z2K @xk
Thus if B(a; r) is a closed ball contained in K;

G(B(a; r)) B(G(a); M (G; K)r):

 115

Lemma 3.5.1. Let (Qk )1

k=1 be a sequence of closed balls contained in such
that
Qk+1 Qk
and
diam Qk ! 0 as k ! 1:
Then, there is a unique point a belonging to each Qk and
vn (G(Qk ))
lim sup j det G0 (a) j :
n!1 vn (Qk )
PROOF. The existence of a point a belonging to each Qk is an immediate
consequence of Theorem 3.1.2. The uniqueness is also obvious since the
diameter of Qk converges to 0 as k ! 1: Set T = G0 (a) and F = T 1 G:
Then, if Qk = B(xk ; rk );
1 1
vn (G(Qk )) = vn (T (T G(Qk ))) =j det T j vn (T G(B(xk ; rk )))
1 1 1
j det T j vn (B(T G(xk ); M (T G; Qk )rk ) =j det T j M (T G; Qk )n vn (Qk ):
Since
1
lim M (T G; Qk ) = 1
k!1

the lemma follows at once.

Lemma 3.5.2. Let Q be a closed ball contained in : Then

Z
vn (G(Q)) j det G0 (x) j dx:
Q

PROOF. Suppose there is a closed ball Q contained in such that

Z
vn (G(Q)) > j det G0 (x) j dx:
Q

This will lead us to a contradiction as follows.

Choose " > 0 such that
Z
vn (G(Q)) (1 + ") j det G0 (x) j dx:
Q
116

n
Let Q = [21 Bk where B1 ; :::; B2n are mutually almost disjoint closed balls
with the same volume. If
Z
vn (G(Bk )) < (1 + ") j det G0 (x) j dx; k = 1; :::; 2n
Bk

we get
2n
vn (G(Q)) k=1 vn (G(Bk ))
Z Z
2n 0
< k=1 (1 + ") j det G (x) j dx = (1 + ") j det G0 (x) j dx
Bk Q

which is a contradiction. Thus

Z
vn (G(Bk )) (1 + ") j det G0 (x) j dx
Bk

for some k: By induction we obtain a sequence (Qk )1

k=1 of closed balls con-
tained in such that
Qk+1 Qk ;
diam Qk ! 0 as k ! 1
and Z
vn (G(Qk )) (1 + ") j det G0 (x) j dx:
Qk

But applying Lemma 3.5.1 we get a contradiction.

PROOF OF THEOREM 3.5.1. Let U be open and write U = [1

i=1 Qi
0
where the Qi s are almost disjoint cubes as in Theorem 3.1.5. Then
Z
1 1
vn (G(U )) i=1 vn (G(Qi )) i=1 j det G0 (x) j dx
Qi
Z
= j det G0 (x) j dx:
U

Using Theorem 3.1.3 we now have that

Z
vn (G(E)) j det G0 (x) j dx
E
 117

for each Borel set E : But then

Z Z
f (x)dx f (G(x)) j det G0 (x) j dx

for each simple Borel measurable function f 0 and, accordingly from this
and monotone convergence, the same inequality holds for each non-negative
Borel function f: But the same line of reasoning applies to G replaced by
G 1 and f replaced by f (G) j det G0 j, so that
Z Z
0
f (G(x)) j det G (x) j dx f (x) j det G0 (G 1 (x)) jj det(G 1 )0 (x) j dx

Z
= f (x)dx:

This proves the theorem.

Example 3.5.1. If f : R2 ! [0; 1] is (R2 ; R0;1 )-measurable and 0 < " <
R < 1, the substitution

G(r; ) = (r cos ; r sin )

yields
Z Z R Z 2
p f (x1 ; x2 )dx1 dx2 = f (r cos ; r sin )rdrd
"< x21 +x22 <R " 0

and by letting " ! 0 and R ! 1; we have

Z Z 1 Z 2
f (x1 ; x2 )dx1 dx2 = f (r cos ; r sin )rdrd :
R2 0 0

The purpose of the example is to show an analogue formula for volume

measure in Rn :
Let S n 1 = fx 2 Rn ; j x j= 1g be the unit sphere in Rn : We will de…ne a
so called surface area Borel measure n 1 on S n 1 such that
Z Z 1Z
f (x)dx = f (r!)rn 1 drd n 1 (!)
Rn 0 Sn 1
118

for any (Rn ; R0;1 )-measurable function f : Rn ! [0; 1] : To this end de…ne
G : Rn n f0g ! ]0; 1[ S n 1 by setting G(x) = (r; !); where
x
r =j x j and ! = :
jxj
1
Note that G : ]0; 1[ Sn 1
! Rn n f0g is given by the equation

G 1 (r; !) = r!:

Moreover,

G 1 (]0; a] E) = aG 1 (]0; 1] E) if a > 0 and E S n 1:

If E 2 B(S n 1 ) we therefore have that

vn (G 1 (]0; a] E)) = an vn (G 1 (]0; 1] E)):

We now de…ne

n 1 (E) = nvn (G 1 (]0; 1] E)) if E 2 B(S n 1 )

and Z
(A) = rn 1 dr if A 2 B(]0; 1[):
A

Below, by abuse of language, we write vnjRn nf0g = vn : Then, if 0 < a

b < 1 and E 2 B(S n 1 );

G(vn )(]0; a] E) = (]0; a]) n 1 (E)

and
G(vn )(]a; b] E) = (]a; b]) n 1 (E):

Thus, by Theorem 1.2.3,

G(vn ) = n 1

and the claim above is immediate.

To check the normalization constant in the de…nition of n 1 ; …rst note
that Z RZ
Rn
vn (j x j< R) = rn 1 drd (!) = n 1 (S
n 1
)
0 S n 1 n
 119

and we get
d
vn (j x j< R) = Rn 1
n 1 (S
n 1
):
dR

Exercises

1. Extend Theorem 3.5.1 to Lebesgue measurable functions.

R
2. The function f : R ! [0; 1[ is Lebesgue measurable
R and R f dm = 1:
Determine all non-zero real numbers such that R hdm < 1; where
1 n
h(x) = n=0 f ( x + n); x 2 R:

###
3.6. Independence in Probability

Suppose ( ; F; P ) is a probability space. The random variables k : ( ;P) !

(Sk ; Sk ); k = 1; :::; n are said to be independent if
n
P( 1 ;:::; n )
= k=1 P k
:

A family ( i )i2I of random variables is said to be independent if i1 ; :::; in

are independent for any i1 ; :::in 2 I with ik 6= il if k 6= l: A family of
events (Ai )i2I is said to be independent if ( Ai )i2I is a family of independent
random variables. Finally a family (Ai )i2I of sub- -algebras of F is said to
be independent if, for any Ai 2 Ai ; i 2 I; the family (Ai )i2I is a family of
independent events.

Example 3.6.1. Let q 2 be an integer. A real number ! 2 [0; 1[ has a

q-adic expansion
(q)
1 k
!= k=1 k :
q
120

(q)
The construction of the Cantor set shows that ( k )1 k=1 is a sequence of
independent random variables based on the probability space

([0; 1[ ; v1j[0;1[ ; B([0; 1[)):

Example 3.6.2. Let (X; M; ) be a positive measure space and let Ai 2 M;

i 2 N+ ; be such that
1
i=1 (Ai ) < 1:

The …rst Borel-Cantelli Lemma asserts that -almost all x 2 X lie in Ai

for at most …nitely many i: This result is an immediate consequence of the
Beppo Levi Theorem since
Z Z
1 1
i=1 Ai d = i=1 Ai d < 1
X X

implies that
1
i=1 Ai < 1 a.e. [ ] :
Suppose ( ; F; P ) is a probability space and let (Ai )1
i=1 be independent
events such that
1
i=1 P [Ai ] = 1:

The second Borel-Cantelli Lemma asserts that almost surely Ai happens for
in…nitely many i:
To prove this, we use the inequality

1+x ex ; x 2 R

to obtain
P \k+n c
i=k Ai =
k+n
i=k P [Aci ]
k+n
k+n k+n P [Ai ] i=k P [Ai ]
= i=k (1 P [Ai ]) i=k e =e :
By letting n ! 1;
P [\1 c
i=k Ai ] = 0

or

P [[1
i=k Ai ] = 1:
 121

But then
P [\1 1
k=1 [i=k Ai ] = 1

and the second Borel-Cantelli Lemma is proved.

Theorem 3.6.1. Suppose 1 ; :::; n are independent random variables and

k 2 N (0; 1); k = 1; :::; n: If 1 ; :::; n 2 R; then

n n 2
k=1 k k 2 N (0; k=1 k)

PROOF. The case 1 ; :::; n = 0 is trivial so assume k 6= 0 for some k: We

have for each open interval A;
Z
n
P [ k=1 k k 2 A] = d 1 (x1 ):::d 1 (xn )
n
k=1 k xk 2A

Z
1 1
(x21 +:::+x2n )
p ne 2 dx1 :::dxn :
n
k=1 k xk 2A
2
p
2 2
Set = 1 + ::: + n and let y = Gx be an orthogonal transformation
such that
1
y1 = ( 1 x1 + ::: + n xn ):

Then, since det G = 1;

Z
1 1 1 2 2)
(y +:::+yn
P[ k=1 k k 2 A] = p ne 2 1 dy1 :::dyn
y1 2A 2
Z
1 1 2
y
= p e 2 1 dy1
y1 2A 2
where we used Fubini’s theorem in the last step. The theorem is proved.

Finally, in this section, we prove a basic result about the existence of

in…nite product measures. Let k ; k 2 N+ be Borel probability measures
in R. The space RN+ is, by de…nition, the set of all sequences x = (xk )1
k=1
122

of real numbers. For each k 2 N+ ; set k (x) = xk : The -algebra RN+

is the least -algebra S of subsets of RN+ which makes all the projections
N+
k : (R ; S) ! (R; R); k 2 N+ ; measurable. Below, ( 1 ; :::; n ) denotes
the mapping of RN+ into Rn de…ned by the equation

( 1 ; :::; n )(x) = ( 1 (x); :::; n (x)):

Theorem 3.6.1. There is a unique probability measure on RN+ such that

( 1 ;:::; n )
= 1 ::: n

for every n 2 N+ :

The measure in Theorem 3.6.1 is called the product of the measures

k ; k 2 N+ ; and is often denoted by

1
k=1 k :

PROOF OF THEOREM 3.6.1. Let ( ; P; F) = ([0; 1[ ; v1j[0;1[ ; B([0; 1[) and

set
(2)
k (!)
(!) = 1 k=1 ; !2 :
2k
We already know that P = P: Now suppose (ki )1 i=1 is a strictly increasing
sequence of positive integers and introduce
(2)
0 1 ki (!)
= i=1 i
; !2 :
2
(2)
Note that for each …xed positive integer n; the Rn -valued maps ( 1 ; :::; (2)
n )
(2) (2)
and ( k1 ; :::; kn ) are P -equimeasurable. Thus, if f : ! R is continuous,
Z Z (2)
n k (!)
f ( )dP = lim f( k=1 )dP (!)
n!1 2k
Z (2) Z
n ki (!)
= lim f( i=1 i
)dP (!) = f ( 0 )dP
n!1 2
and it follows that P 0 = P = P:
 123

By induction, we de…ne for each k 2 N+ an in…nite subset Nk of the set

N+ n [ki=11 Ni such that the set N+ n [ki=1 Ni contains in…nitely many elements
and de…ne
(2)
1 nik
(!)
k = i=1
2i
where (nik )1i=1 is an enumeration of Nk : The map

(!) = ( k (!))1
k=1

is a measurable map of ( ; F) into (RN+ ; RN+ ) and

1
P = k=1 i

where i = P for each i 2 N+ :

For each i 2 N+ there exists a measurable map 'i of ( ; F) into (R; R)
such that P'i = i (see Section 1.6). The map

(x) = ('i (xi ))1

i=1

is a measurable map of (RN+ ; RN+ ) into itself and we get = (P ) . This

completes the proof of Theorem 3.6.1.

"""
124

CHAPTER 4
MODES OF CONVERGENCE

Introduction

In this chapter we will treat a variety of di¤erent sorts of convergence notions

in measure theory. So called L2 -convergence is of particular importance.

4.1. Convergence in Measure, in L1 ( ); and in L2 ( )

Let (X; M; ) be a positive measure space and denote by F(X) the class of
measurable functions f : (X; M) ! (R; R). For any f 2 F(X); set
Z
k f k1 = j f (x) j d (x)
X

and sZ
k f k2 = f 2 (x)d (x):
X

The Cauchy-Schwarz inequality states that

Z
j fg j d k f k2 k g k2 if f; g 2 F(X):
X

To prove this, without loss of generality, it can be assumed that

0 <k f k2 < 1 and 0 <k g k2 < 1:

We now use the inequality

1 2 2
( + ); ; 2R
2
 125

to obtain
Z Z
jf j jgj 1 f2 g2
d ( + )d = 1
X k f k2 k g k2 2 k f k22 k g k22
and the Cauchy-Schwarz inequality is immediate.
If not otherwise stated, in this section p is a number equal to 1 or 2: If it
is important to emphasize the underlying measure k f kp is written k f kp; :
We now de…ne

Lp ( ) = ff 2 F(X); k f kp < 1g :

The special case p = 1 has been introduced earlier. We claim that the
following so called triangle inequality holds, viz.

k f + g kp k f kp + k g kp if f; g 2 Lp ( ):

The case p = 1; follows by -integration of the relation

jf +g j jf j+jg j:

To prove the case p = 2; we use the Cauchy-Schwarz inequality and have

k f + g k22 kj f j + j g jk22
Z
2
=k f k2 +2 j f g j d + k g k22
X

kf k22 +2 k f k2 k g k2 + k g k22 = (k f k2 + k g k2 )2
and the triangle inequality is immediate.
Suppose f; g 2 Lp ( ): The functions f and g are equal almost everywhere
with respect to if ff 6= gg 2 Z : This is easily seen to be an equivalence
relation and the set of all equivalence classes is denoted by Lp ( ): Below
we consider the elements of Lp ( ) as members of Lp ( ) and two members
of Lp ( ) are identi…ed if they are equal a.e. [ ] : From this convention it is
straight-forward to de…ne f + g and f for all f; g 2 Lp ( ) and 2 R and
the function d(p) (f; g) =k f g kp is a metric on Lp ( ): Convergence in the
metric space Lp ( ) = (Lp ( ),d(p) ) is called convergence in Lp ( ): A sequence
(fk )1
k=1 in F(X) converges in measure to a function f 2 F(X) if

lim (j fk f j> ") = 0 all " > 0:

k!1
126

If the sequence (fk )1

k=1 in F(X) converges in measure to a function f
2 F(X) as well as to a function g 2 F(X); then f = g a.e. [ ] since
n "o n "o
fj f g j> "g j f fk j> [ j fk g j>
2 2
and
" "
(j f g j> ") (j f fk j> ) + (j fk g j> )
2 2
1
for every " > 0 and positive integer k: A sequence (fk )k=1 in F(X) is said
to be Cauchy in measure if for every " > 0;

(j fk fn j> ") ! 0 as k; n ! 1:

By the Markov inequality, a Cauchy sequence in Lp ( ) is Cauchy in measure.

p
Example 4.1.1. (a) If fk = k [0; 1 ] ; k 2 N+ ; then
k

1
k fk k2;m = 1 and k fk k1;m = p :
k

Thus fk ! 0 in L1 (m) as k ! 1 but fk 9 0 in L2 (m) as k ! 1:

(b) L1 (m) * L2 (m) since

1
[1;1[ (x) 2 L2 (m) n L1 (m)
jxj

and L2 (m) * L1 (m) since

1
]0;1] (x) p 2 L1 (m) n L2 (m):
jxj

Theorem 4.1.1. Suppose p = 1 or 2:

(a) Convergence in Lp ( ) implies convergence in measure:
 127

(b) If (X) < 1; then L2 ( ) L1 ( ) and convergence in L2 ( ) implies

convergence in L1 ( ):

Proof. (a) Suppose the sequence (fn )1 p

n=1 converges to f in L ( ) and let
" > 0: Then, by the Markov inequality,
Z
1 p 1
(j fn f j ") j fn f j d = k fn f kpp
"p X "p

and (a) follows at once.

(b) The Cauchy-Schwarz inequality gives for any f 2 F(X);

Z Z Z
2 2
( j f j 1d ) f d 1d
X X X

or p
k f k1 k f k2 (X)
and Part (b) is immediate.

Theorem 4.1.2. Suppose fn 2 F(X); n 2 N+ :

(a) If (fn )1
n=1 is Cauchy in measure, there is a measurable function f :
X ! R such that fn ! f in measure as n ! 1 and a strictly increasing
sequence of positive integers (nj )1
j=1 such that fnj ! f a.e. [ ] as j ! 1.
(b) If is a …nite positive measure and fn ! f 2 F(X) a.e. [ ] as
n ! 1; then fn ! f in measure.
(c) (Egoro¤’s Theorem) If is a …nite positive measure and fn !
f 2 F(X) a.e. [ ] as n ! 1; then for every " > 0 there exists E 2 M such
that (E) < " and

sup j fk (x) f (x) j! 0 as n ! 1:

k n
x2E c

PROOF. (a) For each positive integer j; there is a positive integer nj such
that
(j fk fl j> 2 j ) < 2 j ; all k; l nj :
128

There is no loss of generality to assume that n1 < n2 < ::: : Set

j
Ej = j fnj fnj+1 j> 2

and
Fk = [1
j=k Ej :

If x 2 Fkc and i j k
X
j fni (x) fnj (x) j j fnl+1 (x) fnl (x) j
j l<i

X
l j+1
2 <2
j l<i

and we conclude that (fnj (x))1 c

j=1 is a Cauchy sequence for every x 2 Fk : Let
1 c
G = [k=1 Fk and note that for every …xed positive integer k;
X
1
c j k+1
(G ) (Fk ) < 2 =2 :
j=k

Thus Gc is a -null set. We now de…ne f (x) = limj!1 fnj (x) if x 2 G and
f (x) = 0 if x 2
= G:
We next prove that the sequence (fn )1
n=1 converges to f in measure. If
c
x 2 Fk and j k we get
j+1
j f (x) fnj (x) j 2 :

Thus, if j k
j+1 k+1
(j f fnj j> 2 ) (Fk ) < 2 :
Since
" "
(j fn f j> ") (j fn fnj j> ) + (j fnj f j> )
2 2
if " > 0; Part (a) follows at once.

(b) For each " > 0;

Z
(j fn f j> ") = ]";1[ (j fn f j)d
X
 129

and Part (c) follows from the Lebesgue Dominated Convergence Theorem.

(c) Set for …xed k; n 2 N+ ;

1
Ekn = [1
j=n j fj f j> :
k
We have
\1
n=1 Ekn 2 Z

and since is a …nite measure

(Ekn ) ! 0 as n ! 1:

Given " > 0 pick nk 2 N+ such that (Eknk ) < "2 k : Then, if E = [1
k=1 Eknk ,
(E) < ". Moreover, if x 2
= E and j nk
1
j fj (x) f (x) j :
k
The theorem is proved.

Corollary 4.1.1. The spaces L1 ( ) and L2 ( ) are complete.

PROOF. Suppose p = 1 or 2 and let (fn )1 p

n=1 be a Cauchy sequence in L ( ):
We know from the previous theorem that there exists a subsequence (fnj )1j=1
which converges pointwise to a function f 2 F(X) a.e. [ ] : Thus, by Fatou’s
Lemma, Z Z
p
j f fk j d lim inf j fnj fk jp d
X j!1 X
and it follows that f fk 2 L ( ) and, hence f = (f fk ) + fk 2 Lp ( ):
p

Moreover, we have that k f fk kp ! 0 as k ! 1: This concludes the proof

of the theorem.

Corollary 4.1.2. Suppose n 2 N (0; 2n ); n 2 N+ ; and n ! in L2 (P ) as

n ! 1: Then is a centred Gaussian random variable.
130
q
PROOF. We have that k n k2 = E 2n = n and k n k2 !k k2 =def
as n ! 1:
Suppose f is a bounded continuous function on R. Then, by dominated
convergence,
Z Z
E [f ( n )] = f ( n x)d 1 (x) ! f ( x)d 1 (x)
R R

1
as n ! 1. Moreover, there exists a subsequence ( nk )k=1 which converges
to a.s. Hence, by dominated convergence

E f( nk ) ! E [f ( )]

as k ! 1 and it follows that

Z
E [f ( )] = f ( x)d 1 (x):
R

By using Corollary 3.1.3 the theorem follows at once.

Theorem 4.1.3. Suppose X is a standard space and a positive -…nite

Borel measure on X. Then the spaces L1 ( ) and L2 ( ) are separable.

PROOF. Let (Ek )1 k=1 be a denumerable collection of Borel sets with …nite
-measures and such that Ek Ek+1 and [1 k=1 Ek = X: Set k = Ek and
…rst suppose that the set Dk is at most denumerable and dense in Lp ( k )
for every k 2 N+ : Without loss of generality it can be assumed that each
member of Dk vanishes o¤ Ek : By monotone convergence
Z Z
f d = lim f d k , f 0 measurable,
X k!1 X

and it follows that the set [1 p

k=1 Dk is at most denumerable and dense in L ( ):
From now on we can assume that is a …nite positive measure. Let A
be an at most denumerable dense subset of X and and suppose the subset
frn ; n 2 N+ ; g of ]0; 1[ is dense in ]0; 1[ : Furthermore, denote by U the
 131

class of all open sets which are …nite unions of open balls of the type B(a; rn );
a 2 A; n 2 N+ . If U is any open subset of X

U = [ [V : V U and V 2 U]

and, hence, by the Ulam Theorem

(U ) = sup f (V ); V 2 U and V Ug :

Let K be the class of all functions which are …nite sums of functions of
the type U ; where is a positive rational number and U 2 U. It follows
that K is at most denumerable.
Suppose " > 0 and that f 2 Lp ( ) is non-negative. There exists a
sequence of simple measurable functions ('i )1i=1 such that

0 'i " f a.e. [ ] :

Since j f 'i jp f p ; the Lebesgue Dominated Convergence Theorem shows

that k f 'k kp < 2" for an appropriate k: Let 1 ; :::; l be the distinct
positive values of 'k and set
l
C =1+ k=1 k:

Now for each …xed j 2 f1; :::; lg we use Theorem 3.1.3 to get an open
Uj 'k 1 (f j g) such that k Uj "
'k 1 (f j g) kp < 4C and from the above we
"
get a Vj 2 U such that Vj Uj and k Uj Vj kp < 4C : Thus

"
k Vj 'k 1 (f j g)
kp <
2C
and
l
kf k=1 j Vj kp < "
Now it is simple to …nd a 2 K such that k f kp < ": From this we
deduce that the set
K K = fg h; g; h 2 Kg
is at most denumerable and dense in Lp ( ):
132

The set of all real-valued and in…nitely many times di¤erentiable functions
de…ned on Rn is denoted by C (1) (Rn ) and

Cc(1) (Rn ) = f 2 C (1) (Rn ); suppf compact :

Recall that the support suppf of a real-valued continuous function f de…ned

on Rn is the closure of the set of all x where f (x) 6= 0: If
Y
n
f (x) = f'(1 + xk )'(1 xk )g ; x = (x1 ; :::; xn ) 2 Rn
k=1

where '(t) = exp( t 1 ); if t > 0; and '(t) = 0; if t 0; then f 2 Cc1 (Rn ) .

The proof of the previous theorem also gives Part (a) of the following

Theorem 4.1.4. Suppose is a positive Borel measure in Rn such that

(K) < 1 for every compact subset K of Rn : The following sets are dense
in L1 ( ); and L2 ( ) :
(a) the linear span of the functions

I; I open bounded n-cell in Rn ;

(1)
(b) Cc (Rn ):

PROOF. a) The proof is almost the same as the proof of Theorem 4.1.3.
First the Ek :s can be chosen to be open balls with their centres at the origin
since each bounded set in Rn has …nite -measure. Moreover, as in the proof
of Theorem 4.1.3 we can assume that is a …nite measure. Now let A be an
at most denumerable dense subset of Rn and for each a 2 A let

R(a) = fr > 0; (fx 2 X; j xk ak j= rg) > 0 for some k = 1; :::; ng :

Then [a2A R(a) is at most denumerable and there is a subset frn ; n 2 N+ g

of ]0; 1[ n [a2A R(a) which is dense in ]0; 1[ : Finally, let U denote the class
of all open sets which are …nite unions of open balls of the type B(a; rn );
a 2 A; n 2 N+ ; and proceed as in the proof of Theorem 4.1.3. The result
follows by observing that the characteristic function of any member of U
equals a …nite sum of characteristic functions of open bounded n-cells a.e.
[ ]:
 133

Part (b) in Theorem 4.1.4 follows from Part (a) and the following

Lemma 4.1.1. Suppose K U Rn , where K is compact and U is open.

(1)
Then there exists a function f 2 Cc (Rn ) such that

K f U

that is
K f U and suppf U:

PROOF. Suppose 2 Cc1 (Rn ) is non-negative, supp B(0; 1); and

Z
dmn = 1:
Rn

Moreover, let " > 0 be …xed. For any g 2 L1 (vn ) we de…ne

Z
n
f" (x) = " g(y) (" 1 (x y))dy:
Rn

Since
@ k1 +:::+kn
j g j max j k1
j2 L1 (vn ); all k1 ; :::; kn 2 N
Rn
@x1 :::@xn k n

the Lebesgue Dominated Convergent Theorem shows that f" 2 C 1 (Rn ):

Here f" 2 Cc1 (Rn ) if g vanishes o¤ a bounded subset of Rn : In fact,

supp f" (supp g)" :

1
Now choose a positive number " 2
d(K; U c ) and de…ne g = K" : Since
Z
f" (x) = g(x "y) (y)dy
Rn

we also have that f" (x) = 1 if x 2 K: The lemma is proved.

134

Example 4.1.2. Suppose f 2 L1 (mn ) and let g : Rn ! R be a bounded

Lebesgue measurable function. Set
Z
h(x) = f (x y)g(y)dy; x 2 Rn :
Rn

We claim that h is continuous.

To see this …rst note that
Z
h(x + x) h(x) = (f (x + x y) f (x y))g(y)dy
Rn

and
Z
j h(x + x) h(x) j K j f (x + x y) f (x y) j dy
Rn
Z
=K j f ( x + y) f (y) j dy
Rn

if j g(x) j K for every x 2 R : Now …rst choose " > 0 and then ' 2 Cc (Rn )
n

such that
k f ' k1 < ":
Using the triangle inequality, we get
Z
j h(x + x) h(x) j K(2 k f ' k1 + j '( x + y) '(y) j dy)
Rn
Z
K(2" + j '( x + y) '(y) j dy)
Rn

where the right hand side is smaller than 3K" if j x j is su¢ ciently small.
This proves that h is continuous.

Example 4.1.3. Suppose A 2 Rn and mn (A) > 0: We claim that the set

A A = fx x; x 2 Ag

contains a neighbourhood of the origin.

To show this there is no loss of generality to assume that mn (A) < 1:
Set
f (x) = mn (A \ (A + x)); x 2 Rn :
 135

Note that Z
f (x) = A (y) A (y x)dy
Rn
and Example 4.1.2 proves that f is continuous. Since f (0) > 0 there exists a
> 0 such that f (x) > 0 if j x j< : In particular, A \ (A + x) 6= if j x j< ;
which proves that
B(0; ) A A:

The following three examples are based on the Axiom of Choice.

Example 4.1.4. Let N L be the non-Lebesgue measurable set constructed

in Section 1.3. Furthermore, assume A R is Lebesgue measurable and
A N L: We claim that m(A) = 0: If not, there exists a > 0 such that
B(0; ) A A N L N L: If 0 < r < and r 2 Q, there exist a; b 2 N L
such that
a = b + r:
But then a 6= b and at the same time a and b belong to the same equivalence
class, which is a contradiction. Accordingly from this, m(A) = 0:

1 1
Example 4.1.5. Suppose A ; is Lebesgue measurable and m(A) >
2 2
0: We claim there exists a non-Lebesgue measurable subset of A: To see this
note that
A = [1i=1 ((ri + N L) \ A)

where (ri )1
i=1 is an enumeration of the rational numbers in the interval [ 1; 1] :
If each set (ri + N L) \ A; is Lebesgue measurable
1
m(A) = i=1 m((ri + N L) \ A)
and we conclude that m((ri + N L) \ A) > 0 for an appropriate i: But then
m(N L \ (A ri )) > 0 and N L \ (A ri ) N L; which contradicts Example
4.1.4. Hence (ri + N L) \ A is non-Lebesgue measurable for an appropriate i:
If A is a Lebesgue measurable subset of the real line of positive Lebesgue
measure, we conclude that A contains a non-Lebesgue measurable subset.
136

Example 4.1.6. Set I = [0; 1] : We claim there exist a continuous function

f : I ! I and a Lebesgue measurable set L I such that f 1 (L) is not
Lebesgue measurable.
First recall from Section 3.3 the construction of the Cantor set C and the
Cantor function G. First C0 = [0; 1]. Then trisect C0 and remove the middle
interval 31 ; 23 to obtain C1 = C0 n 13 ; 23 = 0; 13 [ 23 ; 1 : At the second
stage subdivide each of the closed intervals of C1 into thirds and remove
from each one the middle open thirds. Then C2 = C1 n ( 19 ; 29 [ 79 ; 89 ): We
repeat the process and what is left from Cn 1 is Cn . The set [0; 1] n Cn is the
union of 2n 1 intervals numbered Ikn ; k = 1; :::; 2n 1; where the interval Ikn
is situated to the left of the interval Iln if k < l: The Cantor set C = \1n=1 Cn :
Suppose n is …xed and let Gn : [0; 1] ! [0; 1] be the unique the monotone
increasing continuous function, which satis…es Gn (0) = 0; Gn (1) = 1; Gn (x) =
k2 n for x 2 Ikn and which is linear on each interval of Cn: It is clear that
Gn = Gn+1 on each interval Ikn , k = 1; :::; 2n 1: The Cantor function is
de…ned by the limit G(x) = limn!1 Gn (x); 0 x 1:
Now de…ne
1
h(x) = (x + G(x)); x 2 I
2
where G is the Cantor function. Since h : I ! I is a strictly increasing and
continuous bijection, the inverse function f = h 1 is a continuous bijection
from I onto I: Set

A = h(I n C)
and
B = h(C):
Recall from the de…nition of G that G is constant on each removed interval
Ikn and that h takes each removed interval onto an interval of half its length.
Thus m(A) = 12 and m(B) = 1 m(A) = 12 :
By the previous example there exists a non-Lebesgue measurable subset
M of B: Put L = h 1 (M ): The set L is Lebesgue measurable since L C
and C is a Lebesgue null set. However, the set M = f 1 (L) is not Lebesgue
measurable.

Exercises
 137

1. Let (X; M; ) be a …nite positive measure space and suppose '(t) =

min(t; 1); t 0: Prove that fn ! f in measure if and only if '(j fn f j) ! 0
in L1 ( ):

2. Let = mj[0;1] : Find measurable functions fn : [0; 1] ! [0; 1] ; n 2 N+ ;

such that fn ! 0 in L2 ( ) as n ! 1;

lim inf fn (x) = 0 all x 2 [0; 1]

n!1

and
lim sup fn (x) = 1 all x 2 [0; 1] :
n!1

3. If f 2 F(X) set

k f k0 = inf f 2 [0; 1] ; (j f j> ) g:

Let
L0 ( ) = ff 2 F(X); k f k0 < 1g
and identify functions in L0 ( ) which agree a.e. [ ] :
(a) Prove that d(0) =k f g k0 is a metric on L0 ( ) and that the corre-
sponding metric space is complete.

(b) Show that F(X) = L0 ( ) if is a …nite positive measure.

4. Suppose Lp (X; M; ) is separable, where p = 1 or 2: Show that Lp (X; M ; )

is separable.

5. Suppose g is a real-valued, Lebesgue measurable, and bounded function

of period one. Prove that
Z 1 Z 1 Z 1
lim f (x)g(nx)dx = f (x)dx g(x)dx
n!1 1 1 0
138

for every f 2 L1 (m):

6. Let hn (t) = 2 + sin nt; 0 t 1; and n 2 N+ : Find real constants and

such that Z Z
1 1
lim f (t)hn (t)dt = f (t)dt
n!1 0 0
and Z Z
1 1
f (t)
lim dt = f (t)dt
n!1 0 hn (t) 0

for every real-valued Lebesgue integrable function f on [0; 1] :

7. If k = (k1 ; :::; kn ) 2 Nn+ ; set ek (x) = ni=1 sin ki xi ; x = (x1 ; :::; xn ) 2 Rn ;

1
and j k j= ( ni=1 ki2 ) 2 : Prove that
Z
lim f ek dmn = 0
jkj!1 Rn

for every f 2 L1 (mn ):

8. Suppose f 2 L1 (mn ); where mn denotes Lebesgue measure on Rn . Com-

pute the following limit and justify the calculations:
Z
lim j f (x + h) f (x) j dx:
jhj!1 Rn

4.2 Orthogonality

Suppose (X; M; ) is a positive measure space. If f; g 2 L2 ( ); let

Z
hf; gi =def f gd
X

be the so called scalar product of f and g: The Cauchy-Schwarz inequality

j hf; gi j k f k2 k g k2
 139

shows that the map f ! hf; gi of L2 ( ) into R is continuous. Observe that

k f + g k22 =k f k22 +2hf; gi+ k g k22
and from this we get the so called Parallelogram Law
k f + g k22 + k f g k22 = 2(k f k22 + k g k22 ):
We will say that f and g are orthogonal (abbr. f ? g) if hf; gi = 0: Note
that
k f + g k22 =k f k22 + k g k22 if and only if f ? g:
Since f ? g implies g ? f; the relation ? is symmetric. Moreover, if
f ? h and g ? h then ( f + g) ? h for all ; 2R. Thus h? =def
ff 2 L2 ( ); f ? hg is a subspace of L2 ( ); which is closed since the map
f ! hf; hi; f 2 L2 ( ) is continuous. If M is a subspace of L2 ( ); the set
M ? =def \h2M h?
is a closed subspace of L2 ( ): The function f = 0 if and only if f ? f:
If M is a subspace of L2 ( ) and f 2 L2 ( ) there exists at most one point
g 2 M such that f g 2 M ? : To see this, let g0 ; g1 2 M be such that
f gk 2 M ? ; k = 0; 1: Then g1 g0 = (f g0 ) (f g1 ) 2 M ? and hence
g1 g0 ? g1 g0 that is g0 = g1 :

Theorem 4.2.1. Let M be a closed subspace in L2 ( ) and suppose f 2

L2 ( ): Then there exists a unique point g 2 M such that
kf g k2 k f h k2 all h 2 M:
Moreover,
f g 2 M ?:

The function g in Theorem 4.2.1 is called the projection of f on M and

is denoted by ProjM f:

PROOF OF THEOREM 4.2.1. Set

d =def d(2) (f; M ) = inf k f g k2 :
g2M
140

and let (gn )1

n=1 be a sequence in M such that

d = lim k f gn k2 :
n!1

Then, by the Parallelogram Law

k (f gk )+(f gn ) k22 + k (f gk ) (f gn ) k22 = 2(k f gk k22 + k f gn k22 )

that is
1
4kf (gk + gn ) k22 + k gn gk k22 = 2(k f gk k22 + k f gn k22 )
2
and, since 21 (gk + gn ) 2 M; we get

4d2 + k gn gk k22 2(k f gk k22 + k f gn k22 ):

Here the right hand converges to 4d2 as k and n go to in…nity and we conclude
that (gn )1 2
n=1 is a Cauchy sequence. Since L ( ) is complete and M closed
there exists a g 2 M such that gn ! g as n ! 1: Moreover,

d =k f g k2 :

We claim that f g 2 M ? : To prove this choose h 2 M and > 0

arbitrarily and use the inequality

k (f g) + h k22 k f g k22

to obtain
kf g k22 +2 hf g; hi + 2
k h k22 k f g k22
and
2hf g; hi + k h k22 0:
By letting ! 0; hf g; hi 0 and replacing h by h; hf g; hi 0: Thus
f g 2 h? and it follows that f g 2 M ? :
The uniqueness in Theorem 4.2.1 follows from the remark just before the
formulation of Theorem 4.2.1. The theorem is proved.

A linear mapping T : L2 ( ) ! R is called a linear functional on L2 ( ):

If h 2 L2 ( ); the map h ! hf; hi of L2 ( ) into R is a continuous linear
 141

functional on L2 ( ): It is a very important fact that every continuous linear

functional on L2 ( ) is of this type.

Theorem 4.2.2. Suppose T is a continuous linear functional on L2 ( ):

Then there exists a unique w 2 L2 ( ) such that

T f = hf; wi all f 2 L2 ( ):

PROOF. Uniqueness: If w; w0 2 L2 ( ) and hf; wi = hf; w0 i for all f 2 L2 ( );

then hf; w w0 i = 0 for all f 2 L2 ( ): By choosing f = w w0 we get f ? f
that is w = w0 :

Existence: The set M =def T 1 (f0g) is closed since T is continuous and

M is a linear subspace of L2 ( ) since T is linear. If M = L2 ( ) we choose
w = 0: Otherwise, pick a g 2 L2 ( ) n M: Without loss of generality it can be
assumed that T g = 1 by eventually multiplying g by a scalar. The previous
theorem gives us a vector h 2 M such that u =def g h 2 M ? : Note that
0 <k u k22 = hu; g hi = hu; gi:
To conclude the proof, let …xed f 2 L2 ( ) be …xed; and use that (T f )g
f 2 M to obtain
h(T f )g f; ui = 0
or
(T f )hg; ui = hf; ui:
By setting
1
w= u
k u k22
we are done.

###
4.3. The Haar Basis and Wiener Measure
142

In this section we will show the existence of Brownian motion with continu-
ous paths as a consequence of the existence of linear measure in the unit
interval. The so called Wiener measure is the probability law on C [0; 1] of
real-valued Brownian motion in the time interval [0; 1] : The Brownian mo-
tion process is named after the British botanist Robert Brown (1773-1858).
It was suggested by Lous Bachelier in 1900 as a model of stock price ‡uctua-
tions and later by Albert Einstein in 1905 as a model of the physical phenom-
enon Brownian motion. The existence of the mathematical Brownian motion
process was …rst established by Norbert Wiener in the twenties. Wiener also
proved that the model can be chosen such that the path t ! W (t); 0 t 1;
is continuous a.s. Today Brownian motion is a very important concept in
probability, …nancial mathematics, partial di¤erential equations and in many
other …elds in pure and applied mathematics.
Suppose n is a non-negative integer and set In = f0; :::; ng : A sequence
(ei )i2In in L2 ( ) is said to be orthonormal if ei ? ej for all i 6= j; i; j 2 In
and k ei k= 1 for each i 2 In : If (ei )i2In is orthonormal and f 2 L2 ( );

f i2In hf; ei iei ? ej all j 2 I

and Theorem 4.2.1 shows that

kf i2In hf; ei iei k2 k f i2In i ei k2 all real 1 ; :::; n:

Moreover

k f k22 =k f i2In hf; ei iei k22 + k i2In hf; ei iei k22

and we get
2
i2In hf; ei i k f k22 :
We say that (en )n2In is an orthonormal basis in L2 ( ) if it is orthonormal
and
f = i2In hf; ei iei all f 2 L2 ( ):
A sequence (ei )1 2 n
i=0 in L ( ) is said to be orthonormal if (ei )i=0 is ortho-
normal for each non-negative integer n: In this case, for each f 2 L2 ( );
1 2
i=0 hf; ei i k f k22

and the series

1
i=0 hf; ei iei
 143

converges since the sequence

n 1
( i=0 hf; ei iei )n=0

of partial sums is a Cauchy sequence in L2 ( ): We say that (ei )1

i=0 is an
2
orthonormal basis in L ( ) if it is orthonormal and
1
f= i=0 hf; ei iei for all f 2 L2 ( ):

Theorem 4.3.1. An orthonormal sequence (ei )1 2

i=0 in L ( ) is a basis of
L2 ( ) if
(hf; ei i = 0 all i 2 N) ) f = 0

Proof. Let f 2 L2 ( ) and set

1
g=f i=0 hf; ei iei :

Then, for any j 2 N;

1
hg; ej i = hf i=0 hf; ei iei ; ej i

1
= hf; ej i i=0 hf; ei ihei ; ej i = hf; ej i hf; ej i = 0:
Thus g = 0 or
1
f= i=0 hf; ei iei :

The theorem is proved.

As an example of an application of Theorem 4.3.1 we next construct an

orthonormal basis of L2 ( ), where is linear measure in the unit interval.
Set
H(t) = [0; 1 [ (t)
2 [ 1 ;1] (t); t 2 R
2

Moreover, de…ne h00 (t) = 1; 0 t 1; and for each n 1 and j = 1; :::; 2n 1 ,

n 1
hjn (t) = 2 2 H(2n 1 t j + 1); 0 t 1:
144

Stated otherwise, we have for each n 1 and j = 1; :::; 2n 1

8 n 1
j 12
>
> 2 2 ;
j 1
t < ;
>
> 2 n 1 2n 1
>
<
hjn (t) = n 1 j 12
> 2 2 ;
2n 1
t 2nj 1 ;
>
>
>
>
:
0; elsewhere in [0; 1] :

It is simple to show that the sequence h00; hjn ; j = 1; :::; 2n 1 ; n 1; is

2
orthonormal in L ( ). We will prove that the same sequence constitute an
orthonormal basis of L2 ( ): Therefore, suppose f 2 L2 ( ) is orthogonal to
each of the functions h00; hjn ; j = 1; :::; 2n 1 ; n 1: Then for each n 1 and
j = 1; :::; 2n 1
Z jn 211 Z nj 1
2 2
fd = j 1 fd
j 1 2
2n 1 2n 1

and, hence,
Z j Z 1
2n 1 1
fd = fd = 0
j 1 2n 1
0
2n 1

since Z Z
1 1
fd = f h00 d = 0:
0 0

Thus Z k
2n 1
f d = 0; 1 j k 2n 1
j
2n 1

and we conclude that

Z 1 Z b
1[a;b] f d = f d = 0; 0 a b 1:
0 a

Accordingly from this, f = 0 and we are done.

The above basis (hk )1 2
k=0 = (h00; h11 ; h12 ; h22 ; h13 ; h23 ; h33 ; h43 ; :::) of L ( )
is called the Haar basis.
Let 0 t 1 and de…ne for …xed k 2 N
Z 1 Z t
ak (t) = [0;t] (x)hk (x)dx = hk d
0 0
 145

so that
1
[0;t] = k=0 ak (t)hk in L2 ( ):
Then, if 0 s; t 1;
Z 1
1
min(s; t) = [0;s] (x) [0;t] (x)dx =h k=1 ak (s)hk ; [0;t] i
0

1 1
= k=0 ak (s)hhk ; [0;t] i = k=0 ak (s)ak (t):

Note that
1 2
t= k=0 ak (t):

If (Gk )1
k=0 is a sequence of N (0; 1) distributed random variables based on
a probability space ( ; F; P ) the series
1
k=0 ak (t)Gk

converges in L2 (P ) and de…nes a Gaussian random variable which we denote

by W (t): From the above it follows that (W (t))0 t 1 is a real-valued centred
Gaussian stochastic process with the covariance

E [W (s)W (t)] = min(s; t):

Such a process is called a real-valued Brownian motion in the time interval

[0; 1] :
Recall that

(h00; h11 ; h12 ; h22 ; h13 ; h23 ; h33 ; h43 ; :::) = (hk )1
k=0 :

We de…ne
(a00; a11 ; a12 ; a22 ; a13 ; a23 ; a33 ; a43 ; :::) = (ak )1
k=0

and
(G00; G11 ; G12 ; G22 ; G13 ; G23 ; G33 ; G43 ; :::) = (Gk )1
k=0 :

It is important to note that for …xed n;

Z t
ajn (t) = [0;t] (x)hjn (x)dx 6= 0 for at most one j:
0

Set
U0 (t) = a00 (t)G00
146

and
2n 1
Un (t) = j=1 ajn (t)Gjn ; n 2 N+ :
We know that
1
W (t) = n=0 Un (t) in L2 (P )
for …xed t:
The space C [0; 1] will from now on be equipped with the metric

d(x; y) =k x y k1

where k x k1 = max0 t 1 j x(t) j : Recall that every x 2 C [0; 1] is uniformly

continuous. From this, remembering that R is separable, it follows that the
space C [0; 1] is separable. Since R is complete it is also simple to show that
the metric space C [0; 1] is complete. Finally, if xn 2 C [0; 1] ; n 2N; and
1
n=0 k x n k1 < 1

the series
1
n=0 xn

converges since the partial sums

n
sn = k=0 xk ; k2N

forms a Cauchy sequence.

We now de…ne
1
= f! 2 ; n=0 k Un k1 < 1g :

Here 2 F since
k Un k1 = sup j Un (t) j
0 t 1
t2Q

for each n: Next we prove that n is a null set.

To this end let n 1 and note that
n n
P k Un k1 > 2 4 P maxn 1 (k ajn k1 j Gjn j) > 2 4 :
1 j 2

But
1
k ajn k1 = n+1
2 2
 147

and, hence,
n
h n
i
n 1 + 12
P k Un k1 > 2 4 2 P j G00 j> 2 4 :

Since Z 1
y 2 =2 dy x2 =2
x 1 ) P [j G00 j x] 2 ye p e
x x 2
we get
n
2n=2
P k Un k1 > 2 4 2n e
and conclude that
"1 #
X X
1
n
E 1[kUn k1 >2 n
4 ] = P k Un k1 > 2 4 < 1:
n=0 n=0

From this and the Beppo Levi Theorem (or the …rst Borel-Cantelli Lemma)
P [ ] = 1:
The trajectory t ! W (t; !); 0 t 1; is continuous for every ! 2 :
Without loss of generality, from now on we can therefore assume that all
trajectories of Brownian motion are continuous (by eventually replacing
by ):
Suppose
0 t1 < ::: < tn 1
and let I1 ; :::; In be open subintervals of the real line. The set

S(t1 ; :::; tn ; I1 ; :::; In ) = fx 2 C [0; 1] ; x(tk ) 2 Ik ; k = 1; :::; ng

is called an open n-cell in C [0; 1] : A set in C [0; T ] is called an open cell if

there exists an n 2 N+ such that it is an open n-cell. The -algebra generated
by all open cells in C [0; 1] is denoted by C: The construction above shows
that the map
W : ! C [0; 1]
which maps ! to the trajectory

t ! W (t; !); 0 t 1

is (F; C)-measurable. The image measure PW is called Wiener measure in

C [0; 1] :
148

The Wiener measure is a Borel measure on the metric space C [0; 1] : We

leave it as an excersice to prove that

C = B(C [0; 1]):

"""
 149

CHAPTER 5
DECOMPOSITION OF MEASURES

Introduction

In this section a version of the fundamental theorem of calculus for Lebesgue

integrals will be proved. Moreover, the concept of di¤erentiating a measure
with respect to another measure will be developped. A very important result
in this chapter is the so called Radon-Nikodym Theorem.

5:1: Complex Measures

Let (X; M) be a measurable space. Recall that if An X; n 2 N+ , and

Ai \ Aj = if i 6= j, the sequence (An )n2N+ is called a disjoint denumerable
collection. The collection is called a measurable partition of A if A = [1
n=1 An
and An 2 M for every n 2 N+ :
A complex function on M is called a complex measure if
1
(A) = n=1 (An )

for every A 2 M and measurable partition (An )1 n=1 of A: Note that ( ) = 0

if is a complex measure. A complex measure is said to be a real measure
if it is a real function. The reader should note that a positive measure need
not be a real measure since in…nity is not a real number. If is a complex
measure = Re + i Im , where Re =Re and Im =Im are real measures.
If (X; M; ) is a positive measure and f 2 L1 ( ) it follows that
Z
(A) = fd ; A 2 M
A

is a real measure and we write d = f d .

150

A function : M ! [ 1; 1] is called a signed measure measure if

(a) : M ! ] 1; 1] or : M ! [ 1; 1[
(b) ( ) = 0
and
(c) for every A 2 M and measurable partition (An )1
n=1 of A;

1
(A) = n=1 (An )
where the latter sum converges absolutely if (A) 2 R:

Here 1 1 = 1 and 1 + x = 1 if x 2 R: The sum of a

positive measure and a real measure and the di¤erence of a real measure and
a positive measure are examples of signed measures and it can be proved that
there are no other signed measures (see Folland [F ]). Below we concentrate
on positive, real, and complex measures and will not say more about signed
measures here.
Suppose is a complex measure on M and de…ne for every A 2 M
1
j j (A) = sup n=1 j (An ) j;

where the supremum is taken over all measurable partitions (An )1

n=1 of A:
Note that j j ( ) = 0 and

j j (A) j (B) j if A; B 2 M and A B:

The set function j j is called the total variation of or the total variation
measure of : It turns out that j j is a positive measure. In fact, as will
shortly be seen, j j is a …nite positive measure.

Theorem 5.1.1. The total variation j j of a complex measure is a positive

measure.

PROOF. Let (An )1

n=1 be a measurable partition of A:
 151

For each n; suppose an <j j (An ) and let (Ekn )1

k=1 be a measurable
partition of An such that
1
an < k=1 j (Ekn ) j :

Since (Ekn )1
k;n=1 is a partition of A it follows that

1 1
n=1 an < k;n=1 j (Ekn ) j j j (A):

Thus
1
n=1 j j (An ) j j (A):
To prove the opposite inequality, let (Ek )1
k=1 be a measurable partition of
A: Then, since (An \ Ek )n=1 is a measurable partition of Ek and (An \ Ek )1
1
k=1
a measurable partition of An ;
1 1 1
k=1 j (Ek ) j= k=1 j n=1 (An \ Ek ) j
1 1
k;n=1 j (An \ Ek ) j n=1 j j (An )
and we get
1
j j (A) n=1 j j (An ):
Thus
1
j j (A) = n=1 j j (An ):
Since j j ( ) = 0, the theorem is proved.

Theorem 5.1.2. The total variation j j of a complex measure is a …nite

positive measure.

PROOF. Since
j j j Re j+j Im j
there is no loss of generality to assume that is a real measure.
Suppose j j (E) = 1 for some E 2 M: We …rst prove that there exist
disjoint sets A; B 2 M such that

A[B =E
152

and
j (A) j> 1 and j j (B) = 1:
To this end let c = 2(1+ j (E) j) and let (Ek )1
k=1 be a measurable partition
of E such that
n
k=1 j (Ek ) j> c

for some su¢ ciently large n: There exists a subset N of f1; :::; ng such that
c
j k2N (Ek ) j> :
2
c
Set A = [k2N Ek and B = E n A: Then j (A) j> 2
1 and

j (B) j=j (E) (A) j

c
j (A) j j (E) j>
j (E) j= 1:
2
Since 1 =j j (E) =j j (A)+ j j (B) we have j j (A) = 1 or
j j (B) = 1: If j j (B) < 1 we interchange A and B and have
j (A) j> 1 and j j (B) = 1:
Suppose j j (X) = 1: Set E0 = X and choose disjoint sets A0 ; B0 2 M
such that
A0 [ B0 = E0
and
j (A0 ) j> 1 and j j (B0 ) = 1:
Set E1 = B0 and choose disjoint sets A1 ; B1 2 M such that

A1 [ B1 = E1

and
j (A1 ) j> 1 and j j (B1 ) = 1:
By induction, we …nd a measurable partition (An )1
n=0 of the set A =def
1
[n=0 An such that j (An ) j> 1 for every n: Now, since is a complex
measure,
(A) = 1n=0 (An ):

But this series cannot converge, since the general term does not tend to zero
as n ! 1: This contradiction shows that j j is a …nite positive measure.
 153

If is a real measure we de…ne

+ 1
= (j j+ )
2
and
1
= (j j ):
2
The measures + and are …nite positive measures and are called the
positive and negative variations of ; respectively . The representation
+
=

is called the Jordan decomposition of :

Exercises

1. Suppose (X; M; ) is a positive measure space and d = f d ; where

f 2 L1 ( ): Prove that d j j=j f j d :

2. Suppose ; ; and are real measures de…ned on the same -algebra and
and : Prove that

min( ; )

where
1
min( ; ) = ( + j j):
2

3. Suppose : M ! C is a complex measure and f; g : X ! R measurable

functions. Show that

j (f 2 A) (g 2 A) j j j (f 6= g)

for every A 2 R:
154

5.2. The Lebesque Decomposition and the Radon-Nikodym Theo-

rem

Let be a positive measure on M and a positive or complex measure

on M: The measure is said to be absolutely continuous with respect to
(abbreviated << ) if (A) = 0 for every A 2 M for which (A) = 0: If
we de…ne
Z = fA 2 M; (A) = 0g
it follows that << if and only if

Z Z :

For example, n << vn and vn << n :

The measure is said to be concentrated on E 2 M if = E , where
E
(A) =def (E \ A) for every A 2 M: This is equivalent to the hypoth-
esis that A 2 Z if A 2 M and A \ E = : Thus if E1 ; E2 2 M, where
E1 E2 ; and is concentrated on E1 ; then is concentrated on E2 : More-
over, if E1 ; E2 2 M and is concentrated on both E1 and E2 ; then is
concentrated on E1 \ E2 : Two measures 1 and 2 are said to be mutually
singular (abbreviated 1 ? 2 ) if there exist disjoint measurable sets E1 and
E2 such that 1 is concentrated on E1 and 2 is concentrated on E2 :

Theorem 5.2.1. Let be a positive measure and ; 1; and 2 complex

measures.
(i) If 1 << and 2 << ; then ( 1 1 + 2 2) << for all complex
numbers 1 and 2:
(ii) If 1 ? and 2 ? ; then ( 1 1 + 2 2) ? for all complex
numbers 1 and 2 :
(iii) If << and ? ; then = 0:
(iv) If << ; then j j<< :

PROOF. The properties (i) and (ii) are simple to prove and are left as exer-
cises.
 155

To prove (iii) suppose E 2 M is a -null set and = E : If A 2 M, then

(A) = (A \ E) and A \ E is a -null set. Since << it follows that
A \ E 2 Z and, hence, (A) = (A \ E) = 0: This proves (iii)
To prove (iv) suppose A 2 M and (A) = 0: If (An )1 n=1 is measurable
partition of A; then (An ) = 0 for every n: Since << ; (An ) = 0 for
every n and we conclude that j j (A) = 0: This proves (vi).

Theorem 5.2.2. Let be a positive measure on M and a complex measure

on M: Then the following conditions are equivalent:
(a) << :
(b) To every " > 0 there corresponds a > 0 such that j (E) j< " for
all E 2 M with (E) < :

If is a positive measure, the implication (a) ) (b) in Theorem 5.2.2 is,

in general, wrong. To see this take = 1 and = v1 : Then << and if
we choose An = [n; 1[ ; n 2 N+ ; then (An ) ! 0 as n ! 1 but (An ) = 1
for each n:

PROOF. (a))(b). If (b) is wrong there exist an " > 0 and sets En 2 M,
n 2 N+ ; such that j (En ) j " and (En ) < 2 n : Set

An = [1 1
k=n Ek and A = \n=1 An :

Since An An+1 A and (An ) < 2 n+1 , it follows that (A) = 0 and
using that j j (An ) j (En ) j; Theorem 1.1.2 (f) implies that

j j (A) = lim j j (An ) ":

n!1

This contradicts that j j<< :

(b))(a). If E 2 M and (E) = 0 then to each " > 0; j (E) j< "; and we
conclude that (E) = 0: The theorem is proved:
156

Theorem 5.2.3. Let be a -…nite positive measure and a real measure

on M.
(a) (The Lebesgue Decomposition of ) There exists a unique pair
of real measures a and s onM such that
= a + s; a << ; and s ? :
If is a …nite positive measure, a and s are …nite positive measures.
(b) (The Radon-Nikodym Theorem) There exits a unique g 2 L1 ( )
such that
d a = gd :
If is a …nite positive measure, g 0 a.e. [ ] :

The proof of Theorem 5.2.3 is based on the following

Lemma 5.2.1. Let (X; M; ) be a …nite positive measure space and suppose
f 2 L1 ( ):
(a) If a 2 R and
Z
fd a (E); all E 2 M
E

then f a a.e. [ ].
(b) If b 2 R and
Z
fd b (E); all E 2 M
E

then f b a.e. [ ].

PROOF. (a) Set g = f a so that

Z
gd 0; all E 2 M:
E

Now choose E = fg > 0g to obtain

Z Z
0 gd = E gd 0
E X
 157

as E g 0 a.e. [ ] : But then Example 2.1.2 yields Eg = 0 a.e. [ ] and we

get E 2 Z : Thus g 0 a.e. [ ] or f a a.e. [ ] :

Part (b) follows in a similar way as Part (a) and the proof is omitted
here.

(k) (k)
PROOF. Uniqueness: (a) Suppose a and s are real measures on M such
that
= (k) (k)
a + s ;
(k)
a << ; and (k)
s ?
for k = 1; 2: Then
(1) (2) (2) (1)
a a = s s

and
(1) (2) (1) (2)
a a << and a a ? :
(1) (2) (1) (2)
Thus by applying Theorem 5.2.1, a a = 0 and a = a : From this
we conclude that (1)
s = s .
(2)

(b) Suppose gk 2 L1 ( ); k = 1; 2; and

d a = g1 d = g2 d :
Then hd = 0 where h = g1 g2 : But then
Z
hd = 0
fh>0g

and it follows that h 0 a.e. [ ] : In a similar way we prove that h 0 a.e.

[ ]. Thus h = 0 in L1 ( ); that is g1 = g2 in L1 ( ):

Existence: The beautiful proof that follows is due to von Neumann.

First suppose that and are …nite positive measures and set = + :
Clearly, L1 ( ) L1 ( ) L2 ( ): Moreover, if f : X ! R is measurable
Z Z sZ
p
jf jd jf jd f 2d (X)
X X X

and from this we conclude that the map

Z
f! fd
X
158

is a continuous linear functional on L2 ( ): Therefore, in view of Theorem

4.2.2, there exists a g 2 L2 ( ) such that
Z Z
fd = f gd all f 2 L2 ( ):
X X

Suppose E 2 M and put f = E to obtain

Z
0 (E) = gd
E

and, since ; Z
0 gd (E):
E

But then Lemma 5.2.1 implies that 0 g 1 a.e. [ ] : Therefore, without

loss of generality we can assume that 0 g(x) 1 for all x 2 X and, in
addition, as above
Z Z
fd = f gd all f 2 L2 ( )
X X

that is Z Z
f (1 g)d = f gd all f 2 L2 ( ):
X X

Put A = f0 g < 1g, S = fg = 1g ; a = A ; and s = S : Note that

= A + S : The choice f = S gives (S) = 0 and hence s ? : Moreover,
the choice
f = (1 + ::: + g n ) E
where E 2 M; gives
Z Z
n+1
(1 g )d = (1 + ::: + g n )gd :
E E

By letting n ! 1 and using monotone convergence

Z
(E \ A) = hd :
E

where
h = lim (1 + ::: + g n )g:
n!1
 159

Since h is non-negative and

Z
(A) = hd
X

it follows that h 2 L1 ( ): Moreover, the construction above shows that =

a + s:
In the next step we assume that is a -…nite positive measure and
a …nite positive measure. Let (Xn )1n=1 be a measurable partition of X such
that (Xn ) < 1 for every n: Let n be …xed and apply Part (a) to the pair
Xn
and Xn to obtain …nite positive measures ( Xn )a and ( Xn )s such that
Xn Xn Xn Xn Xn Xn Xn
=( )a + ( )s ; ( )a << ; and ( )s ?

and
Xn Xn Xn Xn
d( )a = hn d (or ( )a = hn )
where 0 hn 2 L1 ( Xn ): Without loss of generality we can assume that
hn = 0 o¤ Xn and that ( Xn )s is concentrated on An Xn where An 2 Z :
In particular, ( Xn )a = hn : Now
1 Xn
=h + n=1 ( )s

where
1
h= n=1 hn

and Z
hd (X) < 1:
X
Xn
Thus h 2 L1 ( ): Moreover, s =def 1 n=1 ( )s is concentrated on [1
n=1 An 2
Z : Hence s ? :
Finally if is a real measure we apply what we have already proved to
the positive and negative variations of and we are done.

Example 5.2.1. Let be Lebesgue measure in the unit interval and the
counting measure in the unit interval restricted to the class of all Lebesgue
measurable subsets of the unit interval. Clearly, << : Suppose there is an
160

h 2 L1 ( ) such that d = hd . We can assume that h 0 and the Markov

inequality implies that the set fh "g is …nite for every " > 0: But then
n
(h 2 ]0; 1]) = lim (h 2 )=0
n!1
R
and it follows that 1 = (h = 0) = fh=0g
hd = 0; which is a contradiction.

Corollary 5.2.1. Suppose is a real measure. Then there exists

h 2 L1 (j j)

such that j h(x) j= 1 for all x 2 X and

d = hd j j:

PROOF. Since j (A) j j j (A) for every A 2 M, the Radon-Nikodym

Theorem implies that d = hd j j for an appropriate h 2 L1 (j j): But
then d j j=j h j d j j (see Exercise 1 in Chapter 5.1): Thus
Z
j j (E) = j h j d j j; all E 2 M
E

and Lemma 5.2.1 yields h = 1 a.e. [j j] : From this the theorem follows at
once.

Theorem 5.2.4. (Hahn’s Decomposition Theorem) Suppose is a

real measure. There exists an A 2 M such that
+ A Ac
= and = :

PROOF. Let d = hd j j where j h j= 1: Note that hd = d j j : Set

A = fh = 1g : Then

+ 1 1
d = (d j j +d ) = (h + 1)d = Ad
2 2
 161

and
+
d =d d =( A 1)d = Ac d :
The theorem is proved.

If a real measure is absolutely continuous with respect to a -…nite

positive measure , the Radon-Nikodym Theorem says that d = f d for an
approprite f 2 L1 ( ): We sometimes write

d
f=
d
and call f the Radon-Nikodym derivate of with respect to :

Exercises

1. Suppose and n ; n 2N, are positive measures de…ned on the same

-algebra and set = 1 n=0 n . Prove that
a) ? if n ? ; all n 2 N:
b) << if n << ; all n 2 N:

2. Suppose is a real measure and = 1 2; where 1 and 2 are …nite

+
positive measures. Prove that 1 and 2 :

3. Let 1 and 2 be mutually singular complex measures on the same -

algebra: Show that j 1 j?j 2 j :

4. Let (X; M; ) be a -…nite positive measure space and suppose and

are two probability measures de…ned on the -algebra M such that <<
and << : Prove that
Z
1 d d
sup j (A) (A) j= j jd :
A2M 2 X d d
162

5.3. The Wiener Maximal Theorem and the Lebesgue Di¤erentia-

tion Theorem

We say that a Lebesgue measurable function f in Rn is locally Lebesgue in-

tegrable and belongs to the class L1loc (mn ) if f K 2 L1 (mn ) for each compact
subset K of Rn : In a similar way f 2 L1loc (vn ) if f is a Borel function such
that f K 2 L1 (vn ) for each compact subset K of Rn : If f 2 L1loc (mn ); we
de…ne the average Ar f (x) of f on the open ball B(x; r) as
Z
1
Ar f (x) = f (y)dy:
mn (B(x; r)) B(x;r)
It follows from dominated convergence that the map (x; r) ! Ar f (x) of
Rn ]0; 1[ into R is continuous. The Hardy-Littlewood maximal function
f is, by de…nition, f = supr>0 Ar j f j or, stated more explicitly,
Z
1
f (x) = sup j f (y) j dy; x 2 Rn :
r>0 m n (B(x; r)) B(x;r)

The function f : (Rn ; B(Rn )) ! ([0; 1] ; R0;1 ) is measurable since

f = sup Ar j f j :
r>0
r2Q

Theorem 5.3.1. (Wiener’s Maximal Theorem) There exists a positive

constant C = Cn < 1 such that for all f 2 L1 (mn );
C
mn (f > ) k f k1 if > 0:

The proof of the Wiener Maximal Theorem is based on the following

remarkable result.

Lemma 5.3.1. Let C be a collection of open balls in Rn and set V = [B2C B:

If c < mn (V ) there exist pairwise disjoint B1 ; :::; Bk 2 C such that
k n
i=1 mn (Bi ) >3 c:
 163

PROOF. Let K V be compact with mn (K) > c; and suppose A1 ; :::; Ap 2 C

cover K: Let B1 be the largest of the A0i s (that is, B1 has maximal radius),
let B2 be the largest of the A0i s which are disjoint from B1 ; let B3 be the
largest of the A0i s which are disjoint from B1 [ B2 ; and so on until the process
stops after k steps. If Bi = B(xi ; ri ) put Bi = B(xi ; 3ri ): Then [ki=1 Bi K
and
c < ki=1 mn (Bi ) = 3n ki=1 mn (Bi ):
The lemma is proved.

PROOF OF THEOREM 5.3.1. Set

E = ff > g :

For each x 2 E choose an rx > 0 such that Arx j f j (x) > : If c < mn (E );
by Lemma 5.3.1 there exist x1 ; :::; xk 2 E such that the balls Bi = B(xi ; rxi );
i = 1; :::; k; are mutually disjoint and
k n
i=1 mn (Bi ) >3 c:

But then
Z Z
n k 3n k 3n
c<3 i=1 mn (Bi ) < i=1 j f (y) j dy j f (y) j dy:
Bi Rn

The theorem is proved.

Theorem 5.3.2. If f 2 L1loc (mn );

Z
1
lim f (y)dy = f (x) a.e. [mn ] :
r!0 mn (B(x; r)) B(x;r)

PROOF. Clearly, there is no loss of generality to assume that f 2 L1 (mn ):

Suppose g 2 Cc (Rn ) =def ff 2 C(Rn ); f (x) = 0 if j x j large enoughg. Then

lim Ar g(x) = g(x) all x 2 Rn :

r!0
164

Since Ar f f = Ar (f g) (f g) + Ar g g;

lim j Ar f fj (f g) + j f gj:
r!0

Now, for …xed > 0;

mn (lim j Ar f f j> )
r!0

mn ((f ) + mn (j f g j> )
g) >
2 2
and the Wiener Maximal Theorem and the Markov Inequality give

mn (lim j Ar f f j> )
r!0

2C 2
( + )kf g k1 :

Remembering that Cc (Rn ) is dense in L1 (mn ); the theorem follows at once.

If f 2 L1loc (mn ) we de…ne the so called Lebesgue set Lf to be

Z
1
Lf = x; lim j f (y) f (x) j dy = 0 :
r!0 mn (B(x; r)) B(x;r)

Note that if q is real and

Z
1
Eq = x; lim j f (y) q j dy =j f (x) qj
r!0 mn (B(x; r)) B(x;r)

then mn ([q2Q Eqc ) = 0: If x 2 \q2Q Eq ;

Z
1
lim j f (y) f (x) j dy 2 j f (x) qj
r!0 mn (B(x; r)) B(x;r)

for all rational numbers q and it follows that mn (Lcf ) = 0:

A family Ex; = (Ex;r )r>0 of Borel sets in Rn is said to shrink nicely to a
point x in Rn if Ex;r B(x; r) for each r and there is a positive constant ;
independent of r; such that mn (Ex;r ) mn (B(x; r)):
 165

Theorem 5.3.3. (The Lebesgue Di¤erentiation Theorem) Suppose

f 2 L1loc (mn ) and x 2 Lf : Then
Z
1
lim j f (y) f (x) j dy = 0
r!0 mn (Ex;r ) E
x;r

and Z
1
lim f (y)dy = f (x):
r!0 mn (Ex;r ) Ex;r

PROOF. The result follows from the inequality

Z Z
1 1
j f (y) f (x) j dy j f (y) f (x) j dy:
mn (Ex;r ) Ex;r mn (B(x; r)) B(x;r)

Theorem 5.3.4. Suppose is a real or positive measure on Rn and suppose

? vn : If is a positive measure it is assumed that (K) < 1 for every
compact subset of Rn . Then

(Ex;r )
lim = 0 a.e. [vn ]
r!0 vn (Ex;r )

If Ex;r = B(x; r) and is the counting measure cQn restricted to Rn then

? vn but the limit in Theorem 5.3.4 equals plus in…nity for all x 2 Rn : The
hypothesis " (K) < 1 for every compact subset of Rn " in Theorem 5.3.4 is
not super‡ous.

PROOF. Since j (E) j j j (E) if E 2 Rn ; there is no restriction to assume

that is a positive measure (cf. Theorem 3.1.4). Moreover, since

(Ex;r ) (B(x; r))

vn (Ex;r ) vn (B(x; r))

it can be assumed that Ex;r = B(x; r): Note that the function (B( ; r))
is Borel measurable for …xed r > 0 and (B(x; )) left continuous for …xed
x 2 Rn :
166

Suppose A 2 Z and vn = (vn )A : Given > 0; it is enough to prove that

F 2 Zvn where
(B(x; r))
F = x 2 A; lim >
r!0 mn (B(x; r))

To this end let " > 0 and use Theorem 3.1.3 to get an open U A such that
(U ) < ": For each x 2 F there is an open ball Bx U such that

(Bx ) > vn (Bx ):

If V = [x2F Bx and c < vn (V ) we use Lemma 5.3.1 to obtain x1 ; :::; xk such

that Bx1 ; :::; Bxk are pairwise disjoint and
1
c < 3n k
i=1 vn (Bxi ) < 3n k
i=1 (Bxi )
1 1
3n (U ) < 3n ":
Thus vn (V ) 3n 1 ": Since V F 2 Rn and " > 0 is arbitrary, vn (F ) = 0
and the theorem is proved.

Corollary 5.3.1. Suppose F : R !R is an increasing function. Then F 0 (x)

exists for almost all x with respect to linear measure.

PROOF. Let D be the set of all points of discontinuity of F: Suppose 1 <

a < b < 1 and " > 0: If a < x1 < ::: < xn < b; where x1 ; :::; xn 2 D and

F (xk +) F (xk ) "; k = 1; :::; n

then
n
n" k=1 (F (xk +) F (xk )) F (b) F (a):
Thus D \ [a; b] is at most denumerable and it follows that D is at most
denumerable. Set H(x) = F (x+) F (x); x 2 R; and let (xj )N j=0 be an
enumeration of the members of the set fH > 0g : Moreover, for any a > 0;
X X
H(xj ) (F (xj +) F (xj ))
jxj j<a jxj j<a

F (a) F ( a) < 1:
 167

Now, if we introduce
N
(A) = 1 H(xj ) xj (A); A2R

then is a positive measure such that (K) < 1 for each compact subset
K of R. Furthermore, if h is a non-zero real number;
1 1 1
j (H(x + h) H(x) j (H(x + h) + H(x)) 4 (B(x; 2 j h j)
h h 4jhj
and Theorem 5.3.4 implies that H 0 (x) = 0 a.e. [v1 ]. Therefore, without loss
of generality it may be assumed that F is right continuous and, in addition,
there is no restriction to assume that F (+1) F ( 1) < 1:
By Section 1.6 F induces a …nite positive Borel measure such that

(]x; y]) = F (y) F (x) if x < y:

Now consider the Lebesgue decomposition

d = f dv1 + d

where f 2 L1 (v1 ) and ? v1 : If x < y;

Z y
F (y) F (x) = f (t)dt + (]x; y])
x

and the previous two theorems imply that

F (y) F (x)
lim = f (x) a.e. [v1 ]
y#x y x
If y < x; Z x
F (x) F (y) = f (t)dt + (]y; x])
y

and we get
F (y) F (x)
lim = f (x) a.e. [v1 ] :
y"x y x
The theorem is proved.

Exercises
168

1. Suppose F : R ! R is increasing and let f 2 L1loc (v1 ) be such that

F 0 (x) = f (x) a.e. [v1 ] : Prove that
Z y
f (t)dt F (y) F (x) if 1 < x y < 1:
x

5.4. Absolutely Continuous Functions and Functions of Bounded

Variation

Throughout this section a and b are reals with a < b and to simplify notation
we set ma;b = mj[a;b] : If f 2 L1 (ma;b ) we know from the previous section that
the function Z x
(If )(x) =def f (t)dt; a x b
a

has the derivative f (x) a.e. [ma;b ] ; that is

Z x
d
f (t)dt = f (x) a.e. [ma;b ] :
dx a

Our next main task will be to describe the range of the linear map I:
A function F : [a; b] ! R is said to be absolutely continuous if to every
" > 0 there exists a > 0 such that
n n
i=1 j bi ai j< implies i=1 j F (bi ) F (ai ) j< "

whenever ]a1 ; b1 [ ; :::; ]an ; bn [ are disjoint open subintervals of [a; b]. It is ob-
vious that an absolutely continuous function is continuous. It can be proved
that the Cantor function is not absolutely continuous.

Theorem 5.4.1. If f 2 L1 (ma;b ); then If is absolutely continuous.

PROOF. There is no restriction to assume f 0: Set

d = f dma;b :
 169

By Theorem 5.2.2, to every " > 0 there exists a > 0 such that (A) < "
for each Lebesgue set A in [a; b] such that ma;b (A) < : Now restricting A to
be a …nite disjoint union of open intervals, the theorem follows.

Suppose 1 < 1 and F : ] ; [ ! R: For every x 2 ] ; [ we

de…ne
n
TF (x) = sup i=1 j F (xi ) F (xi 1 ) j
where the supremum is taken over all positive integers n and all choices
(xi )ni=0 such that
< x0 < x1 < ::: < xn = x:
The function TF : ] ; [ ! [0; 1] is called the total variation of F: Note that
TF is increasing. If TF is a bounded function, F is said to be of bounded varia-
tion. A bounded increasing function on R is of bounded variation. Therefore
the di¤erence of two bounded increasing functions on R is of bounded vari-
ation. Interestingly enough, the converse is true. In the special case ] ; [ =
R we write F 2 BV if F is of bounded variation.

Theorem 5.4.2. Suppose F 2 BV:

(a) The functions TF + F and TF F are increasing and
1 1
F = (TF + F ) (TF F ):
2 2
In particular, F is di¤erentiable almost everywhere with respect to linear
measure.
(b) If F is right continuous, then so is TF :

PROOF. (a) Let x < y and " > 0: Choose x0 < x1 < ::: < xn = x such that
n
i=1 j F (xi ) F (xi 1 ) j Tf (x) ":

Then
TF (y) + F (y)
170

n
i=1 j F (xi ) F (xi 1 ) j + j F (y) F (x) j +(F (y) F (x)) + F (x)
TF (x) " + F (x)
and, since " > 0 is arbitrary, TF (y) + F (y) TF (x) + F (x): Hence TF + F is
increasing. Finally, replacing F by F it follows that the function TF F
is increasing.

(b) If c 2 R and x > c;

n
Tf (x) = TF (c) + sup i=1 j F (xi ) F (xi 1 ) j

where the supremum is taken over all positive integers n and all choices
(xi )ni=0 such that
c = x0 < x1 < ::: < xn = x:
Suppose TF (c+) > TF (c) where c 2 R: Then there is an " > 0 such that

TF (x) TF (c) > "

for all x > c: Now, since F is right continuous at the point c, for …xed x > c
there exists a partition

c < x11 < ::: < x1n1 = x

such that
n1
i=2 j F (x1i ) F (x1i 1 ) j> ":
But
TF (x11 ) TF (c) > "
and we get a partition

c < x21 < ::: < x2n2 = x11

such that
n2
i=2 j F (x2i ) F (x2i 1 ) j> ":
Summing up we have got a partition of the interval [x21 ; x] with
n2 n1
i=2 j F (x2i ) F (x2i 1 ) j + i=2 j F (x1i ) F (x1i 1 ) j> 2":
 171

By repeating the process the total variation of F becomes in…nite, which is

a contradiction. The theorem is proved.

Theorem 5.4.3. Suppose F : [a; b] ! R is absolutely continuous. Then

there exists a unique f 2 L1 (ma;b ) such that
Z x
F (x) = F (a) + f (t)dt; a x b:
a

In particular, the range of the map I equals the set of all real-valued absolutely
continuous maps on [a; b] :

PROOF. Set F (x) = F (a) if x a and F (x) = F (b) if x b: There exists a

> 0 such that
n n
i=1 j bi ai j< implies i=1 j F (bi ) F (ai ) j< 1

whenever ]a1 ; b1 [ ; :::; ]an ; bn [ are disjoint subintervals of [a; b] : Let p be the
least positive integer such that a + p b: Then TF p and F 2 BV: Let
F = G H; where G = 21 (TF + F ) and H = 12 (TF F ): There exist …nite
positive Borel measures G and H such that

G (]x; y]) = G(y) G(x); x y

and
H (]x; y]) = H(y) H(x); x y:
If we de…ne = G H;

(]x; y]) = F (y) F (x); x y:

Clearly,
(]x; y[) = F (y) F (x); x y
since F is continuous.
Our next task will be to prove that << v1 . To this end, suppose A 2 R
and v1 (A) = 0: Now choose " > 0 and let > 0 be as in the de…nition of the
absolute continuity of F on [a; b] : For each k 2 N+ there exists an open set
172

Vk A such that v1 (Vk ) < and limk!1 (Vk ) = (A): But each …xed Vk is
a disjoint union of open intervals (]ai ; bi [)1
i=1 and hence

n
i=1 j bi ai j<

for every n and, accordingly from this,

1
i=1 j F (bi ) F (ai ) j "

and
1
j (Vk ) j i=1 j (]ai ; bi [) j ":
Thus j (A) j " and since " > 0 is arbitrary, (A) = 0: From this << v1
and the theorem follows at once.

Suppose (X; M; ) is a positive measure space. From now on we write

f 2 L1 ( ) if there exist a g 2 L1 ( ) and an A 2 M such that Ac 2 Z and
f (x) = g(x) for all x 2 A: Furthermore, we de…ne
Z Z
fd = gd
X X

(cf the discussion in Section 2). Note that f (x) need not be de…ned for every
x 2 X:

Corollary 5.4.1. A function f : [a; b] ! R is absolutely continuous if and

only if the following conditions are true:
(i) f 0 (x) exists for ma;b -almost all x 2 [a; b]
(ii) f 0 2 L1 (ma;b ) R
x
(iii) f (x) = f (a) + a f 0 (t)dt; all x 2 [a; b] :

Exercises

1. Suppose f : [0; 1] ! R satis…es f (0) = 0 and

1
f (x) = x2 sin if 0 < x 1:
x2
 173

Prove that f is di¤erentiable everywhere but f is not absolutely continuous.

2. Suppose is a positive real number and f a function on [0; 1] such that

f (0) = 0 and f (x) = x sin x1 ; 0 < x 1. Prove that f is absolutely
continuous if and only if > 1:

3. Suppose f (x) = x cos( =x) if 0 < x < 2 and f (x) = 0 if x 2 Rn ]0; 2[ :

Prove that f is not of bounded variation on R.

4 A function f : [a; b] ! R is a Lipschitz function, that is there exists a

positive real number C such that

j f (x) f (y) j Cjx yj

for all x; y 2 [a; b] : Show that f is absolutely continuous and j f 0 (x) j C

a.e. [ma;b ] :

5. Suppose f : [a; b] ! R is absolutely continuous. Prove that

Z x
Tg (x) = j f 0 (t) j dt; a < x < b
a

if g is the restriction of f to the open interval ]a; b[ :

6. Suppose f and g are real-valued absolutely continuous functions on the

compact interval [a; b]. Show that the function h = max(f; g) is absolutely
continuous and h0 max(f 0 ; g 0 ) a.e. [ma;b ].

###

5.5. Conditional Expectation

174

Let ( ; F; P ) be a probability space and suppose 2 L1 (P ): Moreover,

suppose G F is a -algebra and set

(A) = P [A] ; A 2 G

and Z
(A) = dP; A 2 G:
A
It is trivial that Z = ZP \ G Z and the Radon-Nikodym Theorem shows
there exists a unique 2 L1 ( ) such that
Z
(A) = d all A 2 G
A

or, what amounts to the same thing,

Z Z
dP = dP all A 2 G:
A A

Note that is (G; R)-measurable. The random variable is called the con-
ditional expectation of given G and it is standard to write = E [ j G] :
A sequence of -algebras (Fn )1
n=1 is called a …ltration if

Fn Fn+1 F:

If (Fn )1 1
n=1 is a …ltration and ( n )n=1 is a sequence of real valued random
variables such that for each n;

(a) n 2 L1 (P )
(b) n is (Fn ; R)-measurable
(c) E n+1 j Fn = n

then ( n ; Fn )1
n=1 is called a martingale. There are very nice connections
between martingales and the theory of di¤erentiation (see e.g Billingsley [B]
and Malliavin [M ]):
"""
 175

CHAPTER 6
COMPLEX INTEGRATION

Introduction

In this section, in order to illustrate the power of Lebesgue integration, we

collect a few results, which often appear with uncomplete proofs at the un-
dergraduate level.

6.1. Complex Integrand

So far we have only treated integration of functions with their values in R or

[0; 1] and it is the purpose of this section to discuss integration of complex
valued functions.
Suppose (X; M; ) is a positive measure. Let f; g 2 L1 ( ): We de…ne
Z Z Z
(f + ig)d = fd + i gd :
X X X

If and are real numbers,

Z Z
( + i )(f + ig)d = (( f g) + i( g + f ))d
X X
Z Z
= ( f g)d + i ( g + f )d
X X
Z Z Z Z
= fd gd + i gd + i fd
X X X X
Z Z
= ( + i )( f d + i gd )
X X
Z
=( +i ) (f + ig)d :
X
176

We write f 2 L1 ( ; C) if Re f; Im f 2 L1 ( ) and have, for every f 2 L1 ( ; C)

and complex ; Z Z
fd = fd :
X X

Clearly, if f; g 2 L1 ( ; C); then

Z Z Z
(f + g)d = fd + gd :
X X X

Now suppose is a complex measure on M: If

f 2 L1 ( ; C) =def L1 ( Re ; C) \ L1 ( Im ; C)

we de…ne Z Z Z
fd = fd Re +i fd Im :
X X X

It follows for every f; g 2 L1 ( ; C) and 2 C that

Z Z
fd = fd :
X X

and Z Z Z
(f + g)d = fd + gd :
X X X

###
6.2. The Fourier Transform

Below, if x = (x1 ; :::; xn ) and y = (y1 ; :::; yn ) 2 Rn ; we let

n
hx; yi = k=1 xk yk :

and p
j x j= hx; yi:
If is a complex measure on Rn (or Rn ) the Fourier transform ^ of is
de…ned by Z
^ (y) = e ihx;yi d (x); y 2 Rn :
Rn
 177

Note that
^ (0) = (Rn ):
The Fourier transform of a function f 2 L1 (mn ; C) is de…ned by

f^(y) = ^ (y) where d = f dmn :

Theorem 6.2.1. The canonical Gaussian measure n in Rn has the Fourier

transform
jyj2
^ n (y) = e 2 :

PROOF. Since
n = 1 ::: 1 (n factors)
it is enough to consider the special case n = 1: Set
Z
1 x2
g(y) = ^ 1 (y) = p e 2 cos xydx:
2 R
Note that g(0) = 1: Since
cos x(y + h) cos xy
j j jxj
h
the Lebesgue Dominated Convergence Theorem yields
Z
0 1 x2
g (y) = p xe 2 sin xydx
2 R
(Exercise: Prove this by using Example 2.2.1). Now, by partial integration,
Z
0 1 h x2 ix=1 y x2
g (y) = p e 2 sin xy p e 2 cos xydx
2 x= 1 2 R
that is
g 0 (y) + yg(y) = 0
and we get
y2
g(y) = e 2 :
178

If = ( 1 ; :::; n ) is an Rn -valued random variable with k 2 L1 (P );

k = 1; :::; n; the characteristic function c of is de…ned by

c (y) = E eih ;yi

= P^ ( y); y 2 Rn :

For example, if 2 N (0; ); then = G; where G 2 N (0; 1); and we get

c (y) = E eihG; yi
= ^1( y)
2 y2
=e 2 :
Choosing y = 1 results in
1
E ei =e 2
E [ 2 ] if 2 N (0; ):

Thus if ( k )nk=1 is a centred real-valued Gaussian process

n 1
E ei k=1 yk k = exp( E ( n
k=1 yk k )
2
2
1 n 2 2
= exp( k=1 yk E k 1 j<k n yj yk E j k ):
2
In particular, if
E j k = 0; j 6= k
we see that 2
yk
n E[ 2
]
E ei k=1 yk k = n
k=1 e
2 k

or
n
E ei k=1 yk k = n
k=1 E eiyk k :
Stated otherwise, the Fourier tranforms of the measures P( 1 ;:::; n ) and nk=1 P k
agree. Below we will show that complex measures in Rn with the same
Fourier transforms are equal and we get the following

Theorem 6.2.2. Let ( k )nk=1 be a centred real-valued Gaussian process with

uncorrelated components, that is

E j k = 0; j 6= k:
 179

Then the random variables 1 ; :::; n are independent.

6.3 Fourier Inversion

Theorem 6.3.1. Suppose f 2 L1 (mn ): If f^ 2 L1 (mn ) and f is bounded

and continuous Z
dy
f (x) = eihy;xi f^(y) n
; x 2 Rn :
Rd (2 )

PROOF. Choose " > 0: We have

Z Z Z
ihy;xi "2
jyj 2
^ dy "2
jyj2 dy
e e 2 f (y) n
= f (u) eihy;x ui
e 2 du
R n (2 ) R n R n (2 )n
where the right side equals
Z Z Z
ihv; x " u i 1
jvj2 dv du 1 2 du
f (u) e e 2 p n p n = f (u)e 2"2 ju xj p n
2 2 " n 2 "n
Rn Rn Rn
Z
1 2 dz
= f (x + "z)e 2 jzj p n :
Rn 2
Thus Z Z
"2 2 dy 1
jzj2 dz
e ihy;xi
e 2 f^(y)jyj
= f (x + "z)e 2 p n:
Rn (2 )n Rn 2
By letting " ! 0 and using the Lebesgue Dominated Convergence Theorem,
Theorem 6.3.1 follows at once.

Recall that Cc1 (Rn ) denotes the class of all functions f : Rn ! R

with compact support which are in…nitely many times di¤erentiable. If f 2
Cc1 (Rn ) then f^ 2 L1 (mn ). To see this, suppose yk 6= 0 and use partial
integration to obtain
Z Z
^ ihx;yi 1
f (y) = e f (x)dx = e ihx;yi fx0 k (x)dx
Rd iyk Rd
and Z
1
f^(y) = e ihx;yi (l)
fxk (x)dx; l 2 N:
(iyk )l Rd
180

Thus Z
j yk j j f^(y) j
l
j fx(l)k (x) j dx; l 2 N
Rn

and we conclude that

sup (1+ j y j)n+1 j f^(y) j< 1:

y2Rn

and, hence, f^ 2 L1 (mn ):

Corollary 6.3.1. If f 2 Cc1 (Rn ); then f^ 2 L1 (mn ) and

Z
dy
f (x) = eihy;xi f^(y) n
; x 2 Rn :
R n (2 )

Corollary 6.3.2 If is a complex Borel measure in Rn and ^ = 0; then

= 0:

f^(y)
PROOF. Choose f 2 Cc1 (Rn ). We multiply the equation ^ ( y) = 0 by (2 )n
and integrate over Rn with respect to Lebesgue measure to obtain
Z
f (x)d (x) = 0:
Rn

Since f 2 Cc1 (Rn ) is arbitrary it follows that = 0: The theorem is proved.

6.4. Non-Di¤erentiability of Brownian Paths

Let N D denote the set of all real-valued continuous function de…ned on the
unit interval which are not di¤erentiable at any point. It is well known that
N D is non-empty. In fact, if is Wiener measure on C [0; 1], x 2 N D
a.e. [ ] : The purpose of this section is to prove this important property of
Brownian motion.
 181

Let W = (W (t))0 t 1 be a real-valued Brownian motion in the time

interval [0; 1] such that every path t ! W (t); 0 t 1 is continuous. Recall
that
E [W (t)] = 0
and
E [W (s)W (t)] = min(s; t):
If
0 t0 ::: tn 1
and 1 j<k n

E [(W (tk ) W (tk 1 ))(W (tj ) W (tj 1 )]

= E [(W (tk )W (tj )] E [W (tk )W (tj 1 )] E [W (tk 1 )W (tj )]+E [W (tk 1 )W (tj 1 )]
= tj tj 1 tj + tj 1 = 0:
From the previous section we now infer that the random variables

W (t1 ) W (t0 ); :::; W (tn ) W (tn 1 )

are independent.

Theorem 7. The function t ! W (t); 0 t 1 is not di¤erentiable at

any point t 2 [0; 1] a.s. [P ] :

PROOF. Without loss of generality we assume the underlying probability

space is complete. Let c; " > 0 and denote by B(c; ") the set of all ! 2
such that

j W (t) W (s) j< c j t s j if t 2 [s "; s + "] \ [0; 1]

for some s 2 [0; 1] : It is enough to prove that the set

[
1 [
1
1
B(j; ):
j=1 k=1
k

is of probability zero. From now on let c; " > 0 be …xed. It is enough to

prove P [B(c; ")] = 0 :
182

Set
j+1 j
Xn;k = max j W ( ) W( ) j
k j<k+3 n n
for each integer n > 3 and k 2 f0; :::; n 3g :
Let n > 3 be so large that
3
":
n
We claim that
6c
B(c; ") min Xn;k :
0 k n 3 n
If ! 2 B(c; ") there exists an s 2 [0; 1] such that

j W (t) W (s) j cjt s j if t 2 [s "; s + "] \ [0; 1] :

Now choose k 2 f0; :::; n 3g such that

k k 3
s2 ; + :
n n n

If k j < k + 3;
j+1 j j+1 j
j W( ) W( ) j j W( ) W (s) j + j W (s) W( ) j
n n n n
6c
n
6c
and, hence, Xn;k n
: Now

6c
B(c; ") min Xn;k
0 k n 3 n

and it is enough to prove that

6c
lim P min Xn;k = 0:
n!1 0 k n 3 n

But
6c X
n 3
6c
P min Xn;k P Xn;k
0 k n 3 n k=0
n
 183

6c 6c
= (n 2)P Xn;0 nP Xn;0
n n
1 6c 3 6c
= n(P j W ( ) j ) = n(P (j W (1) j p )3
n n n
12c 3
n( p ):
2 n
where the right side converges to zero as n ! 1. The theorem is proved.

Recall that a function of bounded variation possesses a derivative a.e.

with respect to Lebesgue measure. Therefore, with probability one, a Brown-
ian path is not of bounded variation. In view of this an integral of the type
Z 1
f (t)dW (t)
0

cannot be interpreted as an ordinary Stieltjes integral. Nevertheless, such

an integral can be de…ned by completely di¤erent means and is basic in, for
example, …nancial mathematics.

"""
184

CHAPTER 6
COMPLEX INTEGRATION

Introduction

In this section, in order to illustrate the power of Lebesgue integration, we

collect a few results, which often appear with uncomplete proofs at the un-
dergraduate level.

6.1. Complex Integrand

So far we have only treated integration of functions with their values in R or

[0; 1] and it is the purpose of this section to discuss integration of complex
valued functions.
Suppose (X; M; ) is a positive measure. Let f; g 2 L1 ( ): We de…ne
Z Z Z
(f + ig)d = fd + i gd :
X X X

If and are real numbers,

Z Z
( + i )(f + ig)d = (( f g) + i( g + f ))d
X X
Z Z
= ( f g)d + i ( g + f )d
X X
Z Z Z Z
= fd gd + i gd + i fd
X X X X
Z Z
= ( + i )( f d + i gd )
X X
Z
=( +i ) (f + ig)d :
X
 185

We write f 2 L1 ( ; C) if Re f; Im f 2 L1 ( ) and have, for every f 2 L1 ( ; C)

and complex ; Z Z
fd = fd :
X X

Clearly, if f; g 2 L1 ( ; C); then

Z Z Z
(f + g)d = fd + gd :
X X X

Now suppose is a complex measure on M: If

f 2 L1 ( ; C) =def L1 ( Re ; C) \ L1 ( Im ; C)

we de…ne Z Z Z
fd = fd Re +i fd Im :
X X X

It follows for every f; g 2 L1 ( ; C) and 2 C that

Z Z
fd = fd :
X X

and Z Z Z
(f + g)d = fd + gd :
X X X

###
6.2. The Fourier Transform

Below, if x = (x1 ; :::; xn ) and y = (y1 ; :::; yn ) 2 Rn ; we let

n
hx; yi = k=1 xk yk :

and p
j x j= hx; yi:
If is a complex measure on Rn (or Rn ) the Fourier transform ^ of is
de…ned by Z
^ (y) = e ihx;yi d (x); y 2 Rn :
Rn
186

Note that
^ (0) = (Rn ):
The Fourier transform of a function f 2 L1 (mn ; C) is de…ned by

f^(y) = ^ (y) where d = f dmn :

Theorem 6.2.1. The canonical Gaussian measure n in Rn has the Fourier

transform
jyj2
^ n (y) = e 2 :

PROOF. Since
n = 1 ::: 1 (n factors)
it is enough to consider the special case n = 1: Set
Z
1 x2
g(y) = ^ 1 (y) = p e 2 cos xydx:
2 R
Note that g(0) = 1: Since
cos x(y + h) cos xy
j j jxj
h
the Lebesgue Dominated Convergence Theorem yields
Z
0 1 x2
g (y) = p xe 2 sin xydx
2 R
(Exercise: Prove this by using Example 2.2.1). Now, by partial integration,
Z
0 1 h x2 ix=1 y x2
g (y) = p e 2 sin xy p e 2 cos xydx
2 x= 1 2 R
that is
g 0 (y) + yg(y) = 0
and we get
y2
g(y) = e 2 :
 187

If = ( 1 ; :::; n ) is an Rn -valued random variable with k 2 L1 (P );

k = 1; :::; n; the characteristic function c of is de…ned by

c (y) = E eih ;yi

= P^ ( y); y 2 Rn :

For example, if 2 N (0; ); then = G; where G 2 N (0; 1); and we get

c (y) = E eihG; yi
= ^1( y)
2 y2
=e 2 :
Choosing y = 1 results in
1
E ei =e 2
E [ 2 ] if 2 N (0; ):

Thus if ( k )nk=1 is a centred real-valued Gaussian process

n 1
E ei k=1 yk k = exp( E ( n
k=1 yk k )
2
2
1 n 2 2
= exp( k=1 yk E k 1 j<k n yj yk E j k ):
2
In particular, if
E j k = 0; j 6= k
we see that 2
yk
n E[ 2
]
E ei k=1 yk k = n
k=1 e
2 k

or
n
E ei k=1 yk k = n
k=1 E eiyk k :
Stated otherwise, the Fourier tranforms of the measures P( 1 ;:::; n ) and nk=1 P k
agree. Below we will show that complex measures in Rn with the same
Fourier transforms are equal and we get the following

Theorem 6.2.2. Let ( k )nk=1 be a centred real-valued Gaussian process with

uncorrelated components, that is

E j k = 0; j 6= k:
188

Then the random variables 1 ; :::; n are independent.

6.3 Fourier Inversion

Theorem 6.3.1. Suppose f 2 L1 (mn ): If f^ 2 L1 (mn ) and f is bounded

and continuous Z
dy
f (x) = eihy;xi f^(y) n
; x 2 Rn :
Rd (2 )

PROOF. Choose " > 0: We have

Z Z Z
ihy;xi "2
jyj 2
^ dy "2
jyj2 dy
e e 2 f (y) n
= f (u) eihy;x ui
e 2 du
R n (2 ) R n R n (2 )n
where the right side equals
Z Z Z
ihv; x " u i 1
jvj2 dv du 1 2 du
f (u) e e 2 p n p n = f (u)e 2"2 ju xj p n
2 2 " n 2 "n
Rn Rn Rn
Z
1 2 dz
= f (x + "z)e 2 jzj p n :
Rn 2
Thus Z Z
"2 2 dy 1
jzj2 dz
e ihy;xi
e 2 f^(y)jyj
= f (x + "z)e 2 p n:
Rn (2 )n Rn 2
By letting " ! 0 and using the Lebesgue Dominated Convergence Theorem,
Theorem 6.3.1 follows at once.

Recall that Cc1 (Rn ) denotes the class of all functions f : Rn ! R

with compact support which are in…nitely many times di¤erentiable. If f 2
Cc1 (Rn ) then f^ 2 L1 (mn ). To see this, suppose yk 6= 0 and use partial
integration to obtain
Z Z
^ ihx;yi 1
f (y) = e f (x)dx = e ihx;yi fx0 k (x)dx
Rd iyk Rd
and Z
1
f^(y) = e ihx;yi (l)
fxk (x)dx; l 2 N:
(iyk )l Rd
 189

Thus Z
j yk j j f^(y) j
l
j fx(l)k (x) j dx; l 2 N
Rn

and we conclude that

sup (1+ j y j)n+1 j f^(y) j< 1:

y2Rn

and, hence, f^ 2 L1 (mn ):

Corollary 6.3.1. If f 2 Cc1 (Rn ); then f^ 2 L1 (mn ) and

Z
dy
f (x) = eihy;xi f^(y) n
; x 2 Rn :
R n (2 )

Corollary 6.3.2 If is a complex Borel measure in Rn and ^ = 0; then

= 0:

f^(y)
PROOF. Choose f 2 Cc1 (Rn ). We multiply the equation ^ ( y) = 0 by (2 )n
and integrate over Rn with respect to Lebesgue measure to obtain
Z
f (x)d (x) = 0:
Rn

Since f 2 Cc1 (Rn ) is arbitrary it follows that = 0: The theorem is proved.

6.4. Non-Di¤erentiability of Brownian Paths

Let N D denote the set of all real-valued continuous function de…ned on the
unit interval which are not di¤erentiable at any point. It is well known that
N D is non-empty. In fact, if is Wiener measure on C [0; 1], x 2 N D
a.e. [ ] : The purpose of this section is to prove this important property of
Brownian motion.
190

Let W = (W (t))0 t 1 be a real-valued Brownian motion in the time

interval [0; 1] such that every path t ! W (t); 0 t 1 is continuous. Recall
that
E [W (t)] = 0
and
E [W (s)W (t)] = min(s; t):
If
0 t0 ::: tn 1
and 1 j<k n

E [(W (tk ) W (tk 1 ))(W (tj ) W (tj 1 )]

= E [(W (tk )W (tj )] E [W (tk )W (tj 1 )] E [W (tk 1 )W (tj )]+E [W (tk 1 )W (tj 1 )]
= tj tj 1 tj + tj 1 = 0:
From the previous section we now infer that the random variables

W (t1 ) W (t0 ); :::; W (tn ) W (tn 1 )

are independent.

Theorem 7. The function t ! W (t); 0 t 1 is not di¤erentiable at

any point t 2 [0; 1] a.s. [P ] :

PROOF. Without loss of generality we assume the underlying probability

space is complete. Let c; " > 0 and denote by B(c; ") the set of all ! 2
such that

j W (t) W (s) j< c j t s j if t 2 [s "; s + "] \ [0; 1]

for some s 2 [0; 1] : It is enough to prove that the set

[
1 [
1
1
B(j; ):
j=1 k=1
k

is of probability zero. From now on let c; " > 0 be …xed. It is enough to

prove P [B(c; ")] = 0 :
 191

Set
j+1 j
Xn;k = max j W ( ) W( ) j
k j<k+3 n n
for each integer n > 3 and k 2 f0; :::; n 3g :
Let n > 3 be so large that
3
":
n
We claim that
6c
B(c; ") min Xn;k :
0 k n 3 n
If ! 2 B(c; ") there exists an s 2 [0; 1] such that

j W (t) W (s) j cjt s j if t 2 [s "; s + "] \ [0; 1] :

Now choose k 2 f0; :::; n 3g such that

k k 3
s2 ; + :
n n n

If k j < k + 3;
j+1 j j+1 j
j W( ) W( ) j j W( ) W (s) j + j W (s) W( ) j
n n n n
6c
n
6c
and, hence, Xn;k n
: Now

6c
B(c; ") min Xn;k
0 k n 3 n

and it is enough to prove that

6c
lim P min Xn;k = 0:
n!1 0 k n 3 n

But
6c X
n 3
6c
P min Xn;k P Xn;k
0 k n 3 n k=0
n
192

6c 6c
= (n 2)P Xn;0 nP Xn;0
n n
1 6c 3 6c
= n(P j W ( ) j ) = n(P (j W (1) j p )3
n n n
12c 3
n( p ):
2 n
where the right side converges to zero as n ! 1. The theorem is proved.

Recall that a function of bounded variation possesses a derivative a.e.

with respect to Lebesgue measure. Therefore, with probability one, a Brown-
ian path is not of bounded variation. In view of this an integral of the type
Z 1
f (t)dW (t)
0

cannot be interpreted as an ordinary Stieltjes integral. Nevertheless, such

an integral can be de…ned by completely di¤erent means and is basic in, for
example, …nancial mathematics.

"""
 193

REFERENCES
[B] Billingsley, P. (1995) Probability and Measure. Wiley&Sons.

[D] Dudley, R. M. (1989) Real Analysis and Probability. Wadsworth&Brooks.

[DI] Diedonné, J. (1960) Foundations of Modern Analysis. Academic Press.

[F ] Folland, G. B. (1999) Real Analysis; Modern Techniques and Their Ap-

plications. Second Edition. Wiley&Sons.

[M ] Malliavin, P. (1995) Integration and Probability. Springer.

[R] Rudin, W. (1966) Real and Complex Analysis. McGraw-Hill.

[S] Solovay R. M. (1970) A model of set theory in which every set of reals
is Lebesgue measurable. Annals of Mathematics 92, 1-56.