Lectures on Real Analysis

Mikhail Tyaglov

14.06.2018
2
Contents

1 Introduction to the set theory 7

1.1 Open, closed and compact sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Comparison of sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Countable sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Sets of cardinality continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Sets of higher cardinalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Introduction to the measure theory 21

2.1 Rings and algebras of sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 General measure theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 The Lebesgue measure on Rn 31

3.1 The measure on bricks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 The measure on E n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 The outer measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 The Lebesgue measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Extension of the notion of measurability. The class of measurable sets. . . . . . . . . . . . . 40
3.6 Borel σ-algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.7 Borel regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.8 Invariance properties of Lebesgue measure and non-Lebesgue measurable sets . . . . . . . . 48
3.9 Other examples of measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.9.1 Lebesgue–Stieltjes measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.9.2 Discrete measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.10 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4 Measurable functions 55
4.1 Definition of measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Properties of measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 Sequences of measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4 Approximation of measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5 Integration theory 71
5.1 Integral of simple functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2 Integral of nonnegative measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 Integral of measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3.1 Invariance Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.4 Difference between Riemann and Lebesgue definite integrals . . . . . . . . . . . . . . . . . . 90
5.4.1 Discontinuities of Riemann integrable functions . . . . . . . . . . . . . . . . . . . . . 92

3
4 CONTENTS

5.4.2 Approximation of Riemann integrable functions . . . . . . . . . . . . . . . . . . . . . 96

5.4.3 Approximation of Lebesgue integrable functions . . . . . . . . . . . . . . . . . . . . 97
5.5 Fubini’s theorem and its applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.5.1 Statement and proof of Fubini’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.5.2 Applications of Fubini’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.1.1 Basic properties of Lebesgue integral . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.1.2 Difference between Lebesgue and Riemann definite integrals . . . . . . . . . . . . . . 110
5.1.3 Fubini’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

6 Differentiation of integrable functions 117

6.1 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.1.1 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.1.2 Functions of bounded variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.1.3 Absolutely continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7 Linear spaces 123

7.1 Hölder and Minkowski inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2 Linear normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.3 Spaces Lp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.4 Relations between different types of convergence . . . . . . . . . . . . . . . . . . . . . . . . 134
7.4.1 Approximation of integrable functions . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.5 Basic theory of inner product spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.5.1 Gram-Schmidt orthogonalization method . . . . . . . . . . . . . . . . . . . . . . . . 144
7.5.2 Orthonormal systems in infinitely dimensional spaces . . . . . . . . . . . . . . . . . . 145
7.6 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.6.1 Isomorphism of separable Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.6.2 Orthogonal projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.6.3 Linear functionals and the Riesz representation theorem . . . . . . . . . . . . . . . . 153
7.6.4 Weak convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.6.5 The space L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.7.1 Linear normed spaces. Spaces Lp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.7.2 Linear inner product spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

8 Divergent series. Cesàro and Abel methods of summation. 171

8.1 Divergent series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.1.1 General methods of summing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.1.2 Cesàro summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.1.3 Abel summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.1.4 Product of series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.2 Tauberian theorems. Higher methods of summation of divergent series. . . . . . . . . . . . . 180
8.2.1 Relation between Cesàro and Abel summability methods . . . . . . . . . . . . . . . 180
8.2.2 Tauberian theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.3 Some other summability methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8.3.1 Hölder summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8.3.2 Voronoy summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
8.3.3 Higher Cesàro summability methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
8.3.4 Borel summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
8.3.5 Euler summability method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
8.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
CONTENTS 5

9 Trigonometric Fourier series 191

9.1 Trigonometric Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.2 Basic theory of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.2.1 Series by sines or cosines with monotonically decreasing coefficients. . . . . . . . . . 194
9.3 Applications of Fourier series to ODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.3.1 Differentiation of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.3.2 Applications of Fourier series to finding periodic solutions of ODE . . . . . . . . . . 196
9.3.3 Boundary problems of the theory of ODE . . . . . . . . . . . . . . . . . . . . . . . . 198
9.3.4 Regular Sturm-Liouville problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
9.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
9.4.1 Basic theory of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
9.4.2 Applications of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6 CONTENTS
Chapter 1

Introduction to the set theory

This chapter is devoted to the set theory in Cantor’s sense.

Notations
Let A and B be two sets.
S
• the set A B called the union of the sets A and B, is the set of all elements of A that belong to A
or to B or to both of them.
T
• the set A B called the intersection of the sets A and B, is the set of all elements of A that belong
to A and to B.

• the set A \ B called the difference of the sets A and B, is the set of all elements of A that do not
belong to B.
S T
• the set A M B = (A B) \ (A B) is called the symmetric difference of the sets A and B.

1.1 Open, closed and compact sets

The open ball in Rn centered at x and of radius r is defined by

Br (x) = {y ∈ Rn : |y − x| < r}.

A subset C ⊂ Rn is open if for every x ∈ C there exists r > 0 with Br (x) ⊂ C. By definition, a set is
closed if its complement is open.
We note that any (not necessarily countable) union of open sets is open, while in general the intersection
of only finitely many open sets is open. A similar statement holds for the class of closed sets if one
interchanges the roles of unions and intersections.
A set C is bounded if it is contained in some ball of finite radius. A bounded set is compact if it is
also closed. Compact sets enjoy the Heine-Borel covering property:
S
• Assume C is compact, C ⊂ Ak , and each Ak is open. Then there are finitely many of the open
k
m
S
sets, Ak1 , Ak2 , . . . , Akm , such that C ⊂ Ak j .
j=1

In words, any covering of a compact set by a collection of open sets contains a finite subcovering.
A point x ∈ Rn is a limit point of the set C if for every r > 0, the ball Br (x) contains points of C.
This means that there are points in C which are arbitrarily close to x. An isolated point of C is a point
x ∈ C such that there exists an r > 0 where Br (x) ∩ C is equal to {x}.

7
8 CHAPTER 1. INTRODUCTION TO THE SET THEORY

A point x ∈ C is an interior point of C if there exists r > 0 such that Br (x) ⊂ C. The set of all
interior points of C is called the interior of C. Also, the closure C of the C consists of the union of C
and all its limit points. The boundary of a set C, denoted by ∂C, is the set of points which are in the
closure of C but not in the interior of C.
Note that the closure of a set is a closed set; every point in C is a limit point of C; and a set is closed
if and only if it contains all its limit points. Finally, a closed set C is perfect if C does not have any
isolated points.
The symbol ∅ denotes the empty set, the set with no elements at all. It is supposed to be open and
closed simultaneously.

1.2 Comparison of sets

We start this section the following question. Given two sets, it is asked which one of them contains more
elements than the other one. It is easy to answer this question if the sets contain only finitely (and rather
few) many elements, or if one of the sets is finite, while the other one is infinite. However, it is not easy
to answer the question if both sets are infinite, or contain quite a large number of elements that it is
impossible to count them. For example, can we say what sets contains more elements: the set of natural
numbers or even numbers; the set of rational numbers or integers; the set of irrational numbers or rational?
As we show below, intuitive answers to these seemingly evident questions might be wrong. To find the
correct answers to these questions we have to introduce the following definition.
Definition 1.2.1. Two sets A and B are called equivalent, and we write A ∼ B, if there exists a one-to-one
correspondence between their elements, that is, there exists a bijection f : A ↔ B of the set A on the
set B.
It is easy to see that this definition defines an equivalens relation, because A ∼ A, A ∼ B implies
B ∼ A, and A ∼ B, B ∼ C imply A ∼ C.
Definition 1.2.2. If the sets sets A and B are equivalent, then they are said to be of equal cardinalities
or cardinal numbers: CardA = CardB.
For finite sets, the equality of their cardinality numbers means the equality of the number of their
elements. Analogously, we will say that if two infinite sets are equivalent, then they contain equal number
of elements.
Example 1.2.3. Consider the set of all natural numbers N and the set of all even natural numbers
2N = {2n : n ∈ N}. The mapping f : N ↔ 2N that corresponds the number 2n to each natural number n
is obviously a bijection. Therefore, the sets N and 2N are equivalent.
Definition 1.2.4. The cardinal number of a set A is said to be less than or equal to the cardinal number
of a set B:
CardA 6 CardB,
if the set A is equivalent to some subset B0 of the set B (B0 can coincide with B).
Definition 1.2.5. The cardinal number of a set A is said to be less than the cardinal number of a set B:

CardA < CardB,

if the set A is equivalent to some subset B0 of the set B0 but is not equivalent to B.

1.3 Countable sets

Definition 1.3.1. The cardinal number of the set of all natural numbers N is called the countable cardi-
nality and will be denoted by the letter a, Card N = a.
1.3. COUNTABLE SETS 9

Definition 1.3.2. The sets equivalent to N are called countable.

Thus, if A is a countable set, Card A = a, then each element of A can be one-to-one corresponded
to a natural number n. In other word, the elements of a countable set can be enumerated, that is, be
represented as a sequence. So, if A is a countable set, then it can be written as follows

A = {a1 , a2 , . . . , an , . . .}.

The countable cardinality is the least cardinality among the ones of all infinite sets. This follows from
the following theorem.
Theorem 1.3.3. Any infinite set contains a countable subset.
Proof. Let us choose two distinct elements of the set A and denote them a1 and b1 . The set A \ {a1 , b1 } is
obviously infinite, so we can choose two distinct elements a2 and b2 of this set. We continue in this manner
to get, at an n-th step the set
[
A \ {{a1 , a2 , . . . , an } {b1 , b2 , . . . , bn }}.

This set is infinite, so we can choose two distinct elements an+1 and bn+1 of this set and so on. So we
continue this procedure infinitely many times, and obtain two sequences

a1 , a2 , . . . , an , . . . ; b1 , b2 , . . . , bn , . . .

of distinct elements of the set A. Let us introduce the notations

A1 = {a1 , a2 , . . . , an , . . .}, B1 = {b1 , b2 , . . . , bn , . . .}.

The set A1 ⊂ A is countable, as required.

Remark 1.3.4. We proved even more than it was claimed in the assertion of the theorem. Namely, we
established that we can always split an infinite set into two part, one of which is countable and the other
one is infinite. S S
Indeed, the set A \ A1 = B1 (A \ (A1 B1 )) contains the countable set B1 , so it is infinite.
We will use this Remark in what follows.
Theorem 1.3.5. A union of at most countably many of at most countable sets is at most countable.
Proof. The assertion of the theorem contains four statements which we proof separately.
1) A union of finitely many finite sets is finite.
This statement is obvious.
2) A union of countably many finite sets is at most countable set.
Let us have the sets
A1 = {a11 , a12 , . . . , a1n1 },

A2 = {a21 , a22 , . . . , a2n2 },

......................................

Ak = {ak1 , ak2 , . . . , aknk },

......................................
∞
S
Consider the set A = Ak , and show that it is at most countable. To do this it suffices to show
k=1
that we can enumerate the elements of A or, equivalently, order them into a sequence.
10 CHAPTER 1. INTRODUCTION TO THE SET THEORY

Let us order the elements of A as follows: First, we write elements of the set A1 , then we write those
elements of A2 that do not belong to A1 , and so on. At the kth step, we write the elements of the
set Ak omitting those ones of them that do belong to the previous sets, and so on. Thus, we finally
obtain
A = {a11 , a12 , . . . , a1n1 ; a21 , a22 , . . . , a2n2 ; . . . ; ak1 , ak2 , . . . , aknk ; . . .}.
It is clear that in this manner, each element of each set Ak will be chosen, so we obtain the set A,
indeed. It suffices now to enumerate the elements of the set A. If we use only finitely many numbers
during enumeration, then the set A is finite (it might happen if, from some point, all elements of the
rest sets are already written). If we use the whole set N during the enumeration, then A is countable.
3) A union of finitely many countable sets is a countable set.
Let us have the sets
A1 = {a11 , a12 , . . . , a1n1 , . . .},

A2 = {a21 , a22 , . . . , a2n2 , . . .},

......................................

Ak = {ak1 , ak2 , . . . , aknk , . . .}.

k
S
Consider the set A = Aj , and show that it is countable. To do this, let us order the elements
j=1
of A as follows: First, we write all the first elements of all the sets Ak . Then we write the second
elements of Ak omitting those ones that already appeared in the list, and so on. Finally, we get

A = {a11 , a21 , . . . , ak1 ; a12 , a22 , . . . , ak2 ; . . . ; a1n , a2n , . . . , akn ; . . .}.

It is clear that in such a manner, we mention all the elements of all sets. Moreover, the elements of
A can be enumerated by natural numbers, so A is countable.
4) A union of countably many countable sets is a countable set.
Let us have the sets
A1 = {a11 , a12 , a13 , . . . , a1n1 , . . .},

A2 = {a21 , a22 , a23 , . . . , a2n2 , . . .},

A3 = {a31 , a32 , a33 , . . . , a3n2 , . . .},

......................................

Ak = {ak1 , ak2 , ak3 , . . . , aknk , . . .}.

......................................
∞
S
Consider the set A = Ak , and show that it is countable. In this case, we order the elements of Ak
k=1
“along diagonals” omitting those ones that already appeared in the list:

A = {a11 , a12 , a21 , a13 , a22 , a31 ; . . . ; a1n , a2,n−1 , a3,n−2 , . . . , an1 ; . . .}.

Again, in such a manner every element of every sets will finally appear in the list. During enumeration
of the elements of the set A, we will obviously use all the natural numbers, so A is countable.
Note that in the cases 3) and 4) it is not necessary to have all the sets Ak countable. It is enough if
some of them are countable.
1.3. COUNTABLE SETS 11

If we unite finitely many disjoint finite sets, their cardinalities are summed. Analogously, for infinite
sets we can put, by Theorem 1.3.5,
n1 + n2 + · · · + nk + · · · = a,

a + a + · · · + a = a,

a + a + · · · + a + · · · = a.
Example 1.3.6. Card Z = a, Card Q = a, where Z is the set of all integers, and Q is the set of rational
numbers.
Indeed, the set Z can be represented as
[ [
Z = N {0} (−N),

where −N = {−1, −2, −3, . . . , −n, . . .} is a countable set. By Theorem 1.3.5, Z is a countable set.
Furthermore, the set of rational numbers Q = {m/n : m ∈ Z, n ∈ N} can be represented as
∞
[
Q= Qn , where Qn = {m/n : m ∈ Z},
n=1

so Q is countable as a countable union of countable sets.

Theorem 1.3.7. It is possible to subtract a countable set from any infinite set such that the rest set is
equivalent to the initial set.
Proof. Let A be an infinite set. Select from A two countable sets A1 and B1 as it was done in the proof
of Theorem 1.3.3 (see Remark 1.3.4). Then, we have

A = (A1 ∪ B1 ) ∪ (A \ (A1 ∪ B1 )) ,

A \ A1 = B1 ∪ (A \ (A1 ∪ B1 )) .

The sets A1 ∪ B1 and B1 are countable, so equivalent. Moreover, the set A \ (A1 ∪ B1 ) and A1 ∪ B1 do
not have common elements. Therefore, A and A \ A1 are equivalent.
Theorem 1.3.8. The Cartesian product of finitely many countable sets is a countable set.
Proof. Clearly, it is sufficient to prove the theorem for the Cartesian product of two countable sets.
Given two countable sets

A = {a1 , a2 , . . . , an , . . .}, B = {b1 , b2 , . . . , bn , . . .},

let us consider their Cartesian product

A × B = {(an , bn ) : an ∈ A, bn ∈ B}.

The sets Cm = {(an , bm ) : n ∈ N} are countable because there exists a bijection f : (an , bm ) ↔ an . Since
∞
[
A×B = Cm ,
m=1

the set A × B is countable by Theorem 1.3.5.

Example 1.3.9. Card Qm = a. That is, the set of all m-dimensional vectors with rational coordinates is
countable, since
Qm = Q × Q × · · · × Q .
| {z }
m
12 CHAPTER 1. INTRODUCTION TO THE SET THEORY

Example 1.3.10. The set of all polynomials with rational coefficients is countable.
Let
Q[x] = {pm (x) = r0 xm + r1 xm−1 + · · · + rm : r0 , r1 , . . . , rm ∈ Q, m ∈ N}
be the set of all polynomials with rational coefficients. Each polynomial pm (x) ∈ Q[x] is uniquely defined
by its coefficients (r0 , r1 , . . . , rm ), so if Qm [x] is the set of all polynomials of degree m with rational
∞
coefficients, then Qm [x] ∼ Qm+1 , Qm [x] is countable. Therefore, Q[x] =
S
Qm [x] is countable as a
m=1
countable union of countable sets.
Definition 1.3.11. A real number r is called algebraic if it is a root of a polynomial with rational
coefficients.
Example 1.3.12. The set of all algebraic numbers is countable.
Indeed, any polynomial has finitely many roots. The set Q[x] is countable. Therefore, the set of all
algebraic numbers is countable as a countable union of finite sets, according to Theorem 1.3.5.
Definition 1.3.13. A real number r which is not algebraic is called transcendental.
Do such numbers exist? To answer this question, we should answer the question about existence of
infinite sets that are not countable, and define the cardinality of R. We do this in the next Section.

1.4 Sets of cardinality continuum

To continue study of cardinalities of sets, let us remind the following classical fact from mathematical
analysis.
Theorem 1.4.1 (Cantor). Any decreasing nested sequence of intervals has a unique common point.
Recall that a sequence of closed intervals of the real line {[an , bn ]}, n ∈ N, is called decreasing nested if

[a1 , b1 ] ⊃ [a2 , b2 ] ⊃ [a3 , b3 ] ⊃ . . . ⊃ [an , bn ] ⊃ . . . ,

and
lim (bn − an ) = 0.
n→∞

We asked at the end Section 1.3 whether exist infinite sets that are not countable. The next theorem
gives an affirmative answer to this question.
Theorem 1.4.2 (Cantor). The set of all points of the interval [0, 1] is not countable.
Proof. It is clear that Card [0, 1] > a, since the interval [0, 1] contains the countable subset {1/n : n ∈ N}
(in fact, we can also use Theorem 1.3.3). Suppose, by contradiction, that Card [0, 1] = a. Then we
can enumerate all the points of the interval [0, 1], that is, we can represent the interval [0, 1] as follows
[0, 1] = {x1 , x2 , . . . , xn , . . .}. Let us now apply Cantor’s procedure. To do this, let us denote the interval
[0, 1] as ∆0 and split it into three equal (by length) subintervals: [0, 1/3], [1/3, 2/3], [2/3, 1]. Denote by
∆1 one of them that does not contain the point x1 , that is, x1 6∈ ∆1 (If the point x1 does not belong to
two subintervals, we can choose any of them). Then, we split ∆1 into three equal subintervals and denote
by ∆2 one of them that does not contain the point x2 , and so on. At the n-th step, we split the interval
∆n−1 into three equal subintervals and denote by ∆n one of them that does not contain the point xn , and
so on.
Finally, we get a decreasing nested sequence of intervals

∆0 ⊃ ∆1 ⊃ ∆2 . . . ∆ n ⊃ . . . ,

whose lengths tend to zero. By Cantor’s theorem 1.4.1, there exists a unique point x0 which belong to all
the intervals ∆n , n = 1, 2, . . .. Since x0 ∈ [0, 1], there exists a number m ∈ N such that x0 = xm . Now, on
1.4. SETS OF CARDINALITY CONTINUUM 13

the one hand, xm = x0 ∈ ∆m , while on the other hand, xm 6∈ ∆m by construction, a contradiction. So,
the points of interval [0, 1] cannot be enumerated by natural numbers, that is,

Card [0, 1] > a,

as required.
Definition 1.4.3. The cardinal number of the set of the points of the interval [0, 1] is called the continuum
cardinality and will be denoted by the letter c, Card [0, 1] = c.
S
Theorem 1.4.4. If Card A > a, Card B 6 a, then Card (A B) = Card A.
T
S assume that A B = ∅ (otherwise, we change B to B \ A, and
Proof. Without loss of generality, we can
this changing will not affect the set A B).
Let us select from A a countable sets A0 as it was done in the proof of Theorem 1.3.3. Then we have
S
A = A0 (A \ A0 ),
S S S
A B = (A0 B) (A \ A0 ).
S
The sets A0 and A0 B are both countable, and so equivalent. SinceS the first and the seconds summands
S
in the unions above do not have common elements, the sets A and A B are equivalent, so Card (A B) =
Card A, as required.
Theorem 1.4.5. If Card A > a, Card B 6 a, then Card (A \ B) = Card A.
S
Proof. The set A \ B cannot be finite or countable, since, otherwise, the set A ⊂ (A \ B) B was finite or
countable.
Obviously, Card(A \ B) 6 CardA, so by Theorem 1.4.4 we have
[
Card A 6 Card (A \ B) B = Card (A \ B) 6 Card A,

therefore, Card (A \ B) = Card A, as required.

Example 1.4.6. Card (0, 1) = Card (0, 1] = Card [0, 1) = c.
Example 1.4.7. Card [a, b] = Card (a, b) = Card (a, b] = Card [a, b) = c (b > a).
The mapping y = a + x(b − a) is a bijection between the intervals [0, 1] and [a, b], so Card [a, b] = c.
Thus, any closed, half-closed, or open interval have cardinality continuum.
Example 1.4.8. Card R = c, since the mapping y = arctan x set a bijection between R and the interval
(−π/2, π/2).
Corollary 1.4.9. Transcendental numbers exist. The cardinality of the set of transcendental numbers is
continuum.
Proof. Let A be the set of all algebraic numbers, CardA = a, and T be the set of all transcendental
numbers. Since T = R \ A, we have Card T = c by Theorem 1.4.5.
It is clear that a finite union of sets Xk of cardinality continuum is of cardinality continuum whenever
T
Xk Xj = ∅ for k 6= j. Indeed, each Xk is equivalent to an interval of the form [a, b) and the interval
[0, 1) can be represented as follows:
n
[
[0, 1) = [ck−1 , ck ),
k=1
n
S
where 0 = c0 < c1 < c2 < . . . < cn−1 < cn = 1. Thus, Xk ∼ [0, 1).
k=1
14 CHAPTER 1. INTRODUCTION TO THE SET THEORY

n
S
Analogously, a countable union Xk of sets of cardinality continuum is of cardinality continuum if
T k=1
Xk Xj = ∅ for k =6 j. In this case we represent the interval [0, 1) as a countable union
∞
[
[0, 1) = [ck−1 , ck ),
k=1

1 ∞
S
where ck = 1 − , k = 0, 1, 2, . . .. Since Xk ∼ [ck−1 , ck ) for k ∈ N, we obtain that Xk ∼ [0, 1).
k+1 k=1
Thus,
c + c + · · · + c = nc = c,
and
c + c + · · · + c + · · · = ac = c.

Theorem 1.4.10. The cardinality of the set of all sequences whose elements are 0s and 1s, is continuum.

Proof. Let A be the considered set:

A = {(a1 , a2 , . . . , an , . . .) : an ∈ {0; 1}, n ∈ N}.

Together with the set A, let us consider the set of binary fractions:

F = {0. a1 a2 . . . an . . . : an ∈ {0; 1}, n ∈ N}.

Each such a fraction defines a real number

+∞
X ak
x = 0. a1 a2 . . . an . . . = ,
2k
k=1

from the interval [0, 1], since the minimal binary fraction from the set F is

0. 000 . . . 0 . . . = 0,

while the largest one is

+∞
X 1 1 1
0. 111 . . . 1 . . . = = · 1 = 1.
2k 2 1− 2
k=1

Conversely, every number x ∈ [0, 1] can be expressed as

x = 0. a1 a2 . . . an . . . .

However, some numbers, namely, the numbers of the form 2mk , and only they, have two representations by
binary fractions. One of these representations have 1 in period, while the other one has 0 in period. For
example,
3
= 0. 011000 . . . 00 . . . = 0. 010111 . . . 11 . . .
8
Taking this fact into account, let us split the set F into two sets F1 and F2 , so that F1 consists only
of binary fractions with 0 in period but the fraction 0. 000 . . . 0 . . . = 0, and F2 contains all other binary
fractions. It is clear that F2 ∼ [0, 1], so Card F2 = c. The set F1 is equivalent to the set of fractions of
the form 2mk where 0 < m < 2n , n ∈ N. This set is an infiniteSsubset of all rational numbers, therefore
Card F1 = a. By Theorem 1.4.5, one has Card F = Card (F1 F2 ) = c.
Since, obviously, A ∼ F , we obtain Card A = c.
1.4. SETS OF CARDINALITY CONTINUUM 15

Example 1.4.11. The cardinality of the of all (finite and infinite) sequences of natural numbers is con-
tinuum.
Indeed, let L be the set of all binary fractions of the form

0. a1 a2 a3 . . . an . . . , ak ∈ {0, 1}, k = 1, 2, 3, . . . .

Since L is in one-to-one correspondence with the set of all sequences whose elements are 0s and 1s, we have
Card L = c by Theorem 1.4.10. Let us assign to every sequence (n1 , n2 , . . . , nm , . . .) of natural numbers
a fraction 0. a1 a2 a3 . . . an . . . according the following rule:

an1 = an1 +n2 = · · · = an1 +···+nm = · · · = 1,

while all other ak equal zero. Clearly, we construct a one-to-one correspondence between the set of all
sequences (finite or infinite) of natural numbers and the set of the mentioned binary fractions.
Proposition 1.4.12. The Cartesian product of finitely many sets of cardinality continuum is a set of
cardinality continuum.
Proof. Without loss of generality, it is enough to consider the Cartesian product of two sets. Let CardA =
CardB = c. Consider M = A × B = {(a, b) : a ∈ A, b ∈ B}. For every m ∈ M we have m = (a, b), and
by Example 1.4.11, we have that there is a sequence (p1 , p2 , . . .) of natural numbers corresponding to a,
and there is a sequence (q1 , q2 , . . .) of natural numbers corresponding to b. Thus, we can associate with
the element m the following sequences of natural numbers:

(p1 , q1 , p2 , q2 , . . .)

Thus, we obtain a one-to-one correspondence between the set M and the set of all sequences of natural
numbers which is of cardinality continuum by Example 1.4.11.
Corollary 1.4.13. For any n ∈ N, Card Rn = c.
Corollary 1.4.14. The continuum union of pair-wise non-intersecting sets of cardinality continuum is of
cardinality continuum.
Proof. Indeed, there exists a one-to-one corresponding between the considered union and the set of all
straight lines in R2 parallel to the Ox axis.
Theorem 1.4.15. The Cartesian product of countably many sets of cardinality continuum is a set of
cardinality continuum.
∞
N
Proof. Let M = Ak , CardAk = c, k = 1, 2, . . .. Every element m of the union M has the form
n=1
m = (a1 , a2 , a3 , . . .) where ak ∈ Ak , k ∈ N. By Example 1.4.11, with every ak we can associate a sequence
with natural elements. Thus we have
a1 ∼ (n1 , n2 , n3 , . . .),

a2 ∼ (m1 , m2 , m3 , . . .),

a3 ∼ (l1 , l2 , l3 , . . .),

......................................
so with the element m we associate the sequence

(n1 , n2 , m1 , n3 , m2 , l1 , . . .).

This is a necessary one-to-one correspondence between M and the set of all sequences with natural elements.
16 CHAPTER 1. INTRODUCTION TO THE SET THEORY

The following example plays an important role for our further study of measures.
Example 1.4.16. Cantor’s set.
Let C0 = [0, 1]. Split the interval
[0, 1] into three equal (by length) parts by the points 31 and 23 , and
delete the middle open interval 13 , 23 . The resulting set we denote as C1 , so
   
1 [ 2
C1 = 0, ,1 .
3 3
   
1 2 1 2
Let us now split each of the intervals 0, and , 1 split into three part by the numbers 2 , 2 , and
3 3 3 3
7 8
, , respectively, and delete the middle intervals. The resulting set consisting of four intervals we
32 32
denote as C2 :        
1 [ 2 1 [ 2 7 [ 8
C2 = 0, 2 , , , 1 .
3 32 3 3 32 32
Continuing the procedure, at the nth step we have the set Cn−1 consisting of 2n−1 intervals. We split each
of these intervals into three equal parts and delete the middle intervals. The resulting set consisting of 2n
intervals we denote as Cn , and so on. Finally, we will have the sequence of sets Cn : {Cn }∞
n=0 . Let us denote
their intersection as C:
∞
\
C= Cn .
n=0
The set C is called Cantor’s set. Let us study some properties of this set.
1) Card C = c.
Indeed, let us represent the numbers of the interval [0, 1] as ternary fractions
x = 0. a1 a2 a3 . . . an . . . ,
where ak can assume only values 0, 1, or 2. Clearly, when we delete the middle interval at the first
step of constructing Cantor’s set, we exclude from C0 all the numbers whose first digit in the ternary
fraction representation is 1. At the second step, we delete from C1 all the numbers whose second
digit in the ternary fraction representation is 1, and so on. Therefore, Cantor’s set consists of all
the numbers of the interval [0, 1] whose ternary fraction representations do not have 1, that is, all
the digits in their representations are 0s and 2s. Thus, Cantor’s set is equivalent to the set of all
sequences whose elements are 0s and 1s, and Card C = c by Theorem 1.4.10.
2) Cantor’s set C is closed, so it is compact, since C ⊂ [0, 1].
Each set Cn is closed as a union finitely many closed intervals. But the intersection of any number
of closed sets is a closed set, so C is closed.
3) The “length” of Cantor’s set C equals 0.
A strict definition of the “length” (measure) of a set is given below in Chapter 2. Now let us calculate
the sum of the lengths of all intervals that was deleted during the process of constructing of Cantor’s
set.
At the first step we deleted an interval of length 31 . At the second step we deleted two intervals of
1
length 2 . At the third step, there were deleted four intervals of length 313 , and so on. Thus, the
3
sum of the lengths of all deleted open intervals equals
+∞
X 2k−1
1 1 1 1 1
+ 2 · 2 + 4 · 3 + ··· = k
= · 2 = 1.
3 3 3 3 3 1− 3
k=1

Since we deleted from [0, 1] a union of intervals whose length equals 1, then the “length” of the rest
set (Cantor’s set) equals zero.
1.5. SETS OF HIGHER CARDINALITIES 17

The “paradoxical” properties 1) and 3) of Cantor’s set were ones of the reasons to develop the so-called
measure theory, which we study below in Chapter 2.
We finish this section with the following theorem.
Theorem 1.4.17. Every open set in R can represented as a countable union of disjoint open intervals.
Proof. Let A be an open set in R. For each x ∈ A, let Ix denote the largest open interval containing x
and contained in A. More precisely, since A is open, x is contained in some small (non-trivial) interval,
and therefore if

ax = inf{a < x : (a, x) ⊂ A} and bx = sup{b > x : (x, b) ⊂ A}

we must have ax < x < bx (with possibly infinite values for ax and bx ). If we now let Ix = (ax , bx ), then
by construction we have x ∈ Ix as well as Ix ⊂ A. Hence
[
A= Ix .
x∈A

Now suppose that two intervals Ix and Iy intersect. Then their union (which is also an open interval)
is contained in A and contains x. Since Ix is maximal, we must have (Ix ∪ Iy ) ⊂ Ix , and similarly
(Ix ∪ Iy ) ⊂ Iy . This can happen only if Ix = Iy , therefore, any two distinct intervals in the collection
I = {Ix }x∈A must be disjoint. Moreover, the collection I of open intervals Ix is countable, since every open
interval Ix contains a rational number. Since different intervals are disjoint, they must contain distinct
rationals, and therefore I is countable, as required.

1.5 Sets of higher cardinalities

While the countable cardinality is the minimal among all the cardinal numbers, there is no maximal
cardinal number, as the following theorem shows.
Theorem 1.5.1. The cardinal numbers of the set of all subsets of a given set B is larger than the cardinal
number of the set B.
S S
Proof. Let B = b be some set, and let M = m be the set of all subsets of B. Here we denote by b the
elements of the set B, and by m the elements of the set M . We notice that the empty set ∅ and the set
B itself are elements of the set M . There obviously exists a bijection between the elements b of the set B
and the one-element sets m = {b} of the set B, therefore, Card M > Card B. On the contrary, suppose
that Card M = Card B. Then, there exists a bijection f : M ↔ B. This means that each element b of
the set B can be represented as b = f (m), where the element m ∈ M is unique.
Let us split the set B into two classes. The first class contains all the elements b of the set B such that
b = f (m) ∈ m. The second class contains all the elements b of the set B such that b = f (m) 6∈ m. Since f
is a bijection, each element of B cannot belong to both these classes simultaneously, and must belong to
one of them.
Consider the set m0 of the elements of the second class. This class is a subset of B, so m0 ∈ M , and
the bijection f defines an element b0 ∈ B such that b0 = f (m0 ). So, does the element b0 belong to the
first class or to the second one? If b0 belongs to the first class, then b0 = f (m0 ) ∈ m0 , that is impossible,
since m0 is the set of elements of the second class, a contradiction. If b0 belongs to the second class, then
b0 = f (m0 ) 6∈ m0 . But m0 contains all the elements of the second class, a contradiction again.
Thus, b0 cannot belong neither to the first nor the second class. Therefore, no bijections f : M ↔ B
exist.
set (1 = n0 ), n = n1 of



If the set B is finite, Card B = n, then the set M contains one empty
one-element sets, n2 of two-element sets,. . . , and the set B itself (1 = nn ). Since

     
n n n
1+ + + ··· + + 1 = (1 + 1)n = 2n ,
1 2 n−1
18 CHAPTER 1. INTRODUCTION TO THE SET THEORY

then Card M = 2n . By analogy, if B is an infinite set and Card B = α, then it is assumed that

Card M = 2α .

Example 1.5.2. The following formula is valid:

2a = c

Let T be the set of all subsets of the set N, and let L be the set of all the binary fractions of the form

0. a1 a2 a3 . . . an . . . , ak ∈ {0, 1}, k = 1, 2, 3, . . . .

It is clear that Card T = 2a , since Card N = a, while Card L = c by Theorem 1.4.10.

To each element t of T (t is a finite or countable sequence of growing natural numbers) we can assign
an element x = 0. a1 a2 a3 . . . an . . . of L by the the following rule. If a natural number k belongs to the
sequence t, then ak = 1. If k 6∈ t, then ak = 0. Clearly, in such a manner we obtain a one-to-one
correspondence between T and L, so their cardinalities are equal.
It can be shown that any two cardinal numbers can be compared, that is, if A and B are two sets, then
the following relations are exclusive (there exists only one of them):

Card A < Card B, Card A = Card B, Card A > Card B. (1.5.1)

It can be proved by establishing first that if A ⊃ A∗ ⊃ A∗∗ and Card A = Card A∗∗ , then Card A∗ =
Card A. Then, the fact that the relations (1.5.1) are exclusive follows from the fact that if a set A is
equivalent to a subset of a set B, while the set B is equivalent to a subset to the set A, then A and B are
equivalent.
Remark 1.5.3. Despite of the well-developed theory of sets comparisons, there is no answer to the question
known as continuum hypothesis. The hypothesis (due to G. Cantor) claims that there is no sets whose
cardinality is between a and c. However, it was proved in mid-XX century that it is impossible to prove or
disprove the hypothesis in the system of axioms known as ZFC (Zermelo–Fraenkel set theory). In fact, we
implicitly use this system of axioms in our life and, in particular, in these lecture notes. So we can only
postulate here the affirm answer to the question as an additional axiom. For further details, see [2].
1.6. PROBLEMS 19

1.6 Problems
Problem 1.1. Prove explicitly (by finding a bijection) that
1) the sets [0, 1] and (0, 1) are equivalent;
2) the sets [0, 1] and (a, b] are equivalent. Here a, b ∈ R;
3) the sets (0, 1] and [0, 1) are equivalent;
4) the sets [0, 1] and (a, b) are equivalent, where a, b ∈ R, a < b.
Hint: Find a bijection that transfer (one-to-one) a countable subset of the interval [0, 1] containing the
point 1 to a countable subset of the the interval [0, 1) that does not contain the point 1. Use the composition
of bijections.
Problem 1.2. Prove explicitly (by finding a bijection) that
1) the real axis R is equivalent to the closed interval [−1, 1];
2) the real axis R is equivalent to the open interval (−a, a), a > 0;
3) the real axis R is equivalent to the interval [−a, a), a > 0.
Hint: Find a bijection that transfer (one-to-one) a countable subset of the interval [0, 1] containing the
point 1 to a countable subset of the the interval [0, 1) that does not contain the point 1. Use the composition
of bijections.
Problem 1.3. Prove that the border of the square with vortices A(−1, 1), B(1, 1), C(−1, 1), and D(−1, −1)
is equivalent to the circumference {(x, y) : x2 + y 2 = 1}.
Problem 1.4. Prove that the circumference {(x, y) : x2 + y 2 = 1} is equivalent to the ellipse {(x, y) :
2x2 + 4y 2 = 16}.
Problem 1.5. Prove that the disc {(x, y) : x2 + y 2 6 9} is equivalent to the set {(x, y) : |x| 6 1, |y| 6 1}.
Problem 1.6. Prove that the disc {(x, y) : x2 + y 2 6 25} is equivalent to the set {(x, y) : 4x2 + 8y 2 6 16}.
Problem 1.7. Prove that the set of points of discontinuity of a monotone function on an interval [a, b],
is at most countable.
Problem 1.8. Prove that the set of all intervals of R, whose ends are rational, is countable.
Problem 1.9. Prove that the set of all triangles on the plane, whose vertices have rational coordinates,
is countable.
Problem 1.10. Prove that the cardinal number of the set {(x, y) : x2 + y 2 = 1} equals c (continuum).
Problem 1.11. Prove that any set of open disjoint discs on the plane is at most countable.
S
Problem 1.12. Let a set A be finite, and let a set B be countable. Prove that A B is equivalent to the
set of all integers Z.
S
Problem 1.13. Let a set A be finite, and let a set B be countable. Prove that A B is equivalent to the
set of all numbers of the form 2k , k ∈ Z.
S
Problem 1.14. Let a set A = B C and CardA = c. Prove that at least one of the sets B and C is of
cardinality c.
Problem 1.15. Prove that the cardinality the set of all (at most countable) sequences {uk }+∞
k=1 with real
elements (uk ∈ R) is c.
Hint: Use Theorem 1.4.15.
20 CHAPTER 1. INTRODUCTION TO THE SET THEORY

Problem 1.16. Prove that the cardinality of the set Φ of all continuous functions defined on [0, 1] is equal
to c.
Hint: Prove that CardΦ > c by finding a subset of Φ of cardinality c. To prove CardΦ 6 c, use the
continuity of the functions in Φ, the fact that the set of rational numbers is countable, and the result of
Problem 1.15.
Problem 1.17. Prove that the cardinality of the set F of all real functions defined on [0, 1] is greater
than c. There is the notation Card F = f .
Hint: Prove that [0, 1] 6∼ F and ∃ F ∗ ⊂ F such that [0, 1] ∼ F ∗ .
Problem 1.18. Prove that 2c = f .
Hint: First prove the fact that the cardinality of the strip [0, 1] × R is continuum.
Problem 1.19. Prove that the Cantor set C is totally disconnected and perfect. In other words, given
two distinct points x, y ∈ C, there is a point z ∈
/ C that lies between x and y, and yet C has no isolated
points.
Hint: If x, y ∈ C and |x − y| > 31k , then x and y belong to two different intervals in Ck . Also, given any
x ∈ C there is an end-point yk of some interval in Ck that satisfies x 6= yk and |x − yk | 6 31k .
Problem 1.20. Prove that the interval [0, 1] cannot be represented as a countable union of disjoint closed
intervals (where at least two of intervals are non-empty).
Hint: Use the Cantor’s theorem about the sequence of nested intervals to get a contradiction.
Do not use Baire categories!!!
Problem 1.21. Prove that the interval [0, 1] cannot be represented as a countable union of disjoint closed
sets (where at least two of sets are non-empty).
Hint: Use the fact that any open set on [0, 1] can be represented as a countable union of open sets.
Do not use Baire categories!!!
Chapter 2

Introduction to the measure theory

In this chapter, we give a short review of rings and algebras of sets and introduce the general measure
theory.

2.1 Rings and algebras of sets

In what follows we consider collections of sets whose elements are also sets. We mainly assume that all
the considered sets are subsets of some basic set (a universe).

Definition 2.1.1. A non-empty collection of sets K is called a ring if it possesses the following properties:
S
1) A, B ∈ K =⇒ A B ∈ K;

2) A, B ∈ K =⇒ A \ B ∈ K.

Rings also possess the properties:

3) ∅ ∈ K;
T
4) A, B ∈ K =⇒ A B ∈ K;

5) A, B ∈ K =⇒ A M B ∈ K;

These properties follow from the properties 1) and 2) and from the equalities:

∅=A\A (A ∈ K),
T
A B = A \ (A \ B),
S
A M B = (A \ B) (B \ A),

that can be easily checked.

Thus, a ring is a non-empty collection of sets containing the empty set and closed w.r.t. operations of
union, intersection, substraction, and symmetric substraction.
T
Definition 2.1.2. A set E ∈ K is called the unit of the ring K if A E = A for any A ∈ K.

Definition 2.1.3. A ring with unit is called algebra.

We denote algebras by the letter A. Any algebra possesses one more property (additionally to the
properties 1) − 5)).

6) A ∈ A =⇒ Ac := E \ A ∈ A.

21
22 CHAPTER 2. INTRODUCTION TO THE MEASURE THEORY

The set Ac is called the complement of the set A in the algebra A.

Example 2.1.4. Let X be a non-empty set. Then the collection A = {∅, X} is an algebra with the
unit X.

Example 2.1.5. Let X be a non-empty set. Then the collection of all its subsets P(X) is an algebra with
the unit X.

Example 2.1.6. Let X be a non-empty set, and let Pf (X) be the collection of all finite subsets of X.
Then Pf (X) is a ring. The ring Pf (X) is an algebra with the unit X if, and only if, X is a finite set.
However, in this case, Pf (X) coincides with P(X).

Example 2.1.7. Let Pb (R) be the collection of all bounded subsets of R. The set Pb (R) is a ring (without
units).

Our next example is not as trivial as the previous ones, and requires some preparations. Consider the
n-dimensional Euclidean space Rn .

Definition 2.1.8. The set1

K = {x = (xi )ni=1 : ai < xi < bi ∨ ai 6 xi < bi ∨ ai < xi 6 bi ∨ ai 6 xi 6 bi , i = 1, 2, . . . , n}. (2.1.1)

is called a brick or n-dimensional rectangle (parallelepiped).

In (2.1.1) we assume that ai 6 bi for any i = 1, 2, . . . , n, and that for different i there might hold
different inequalities (one of the four mentioned inequalities). Thus, ∅ is a brick (ai = bi for at least one
index i, and ai < xi < bi for all i); the points are bricks (ai = bi and ai 6 xi 6 bi for all i); finite intervals
are bricks (for one fixed index i, ai < bi , and for all the rest indices i, ai = bi ); parallelepipeds of any
dimension k, 2 6 k 6 n, are bricks, etc.

Definition 2.1.9. A set B ∈ Rn is called elementary if it can be represented as a union of finitely many
bricks:
[l
B= Kj . (2.1.2)
j=1

The collection of all elementary sets of the space R is denoted E n .

n

Note that the representation of an elementary set in the form (2.1.2) is not unique, since the bricks Kj
might have non-empty intersections. However, it is possible to find the representation of an elementary
sets in the form of union of finitely many disjoint bricks. To do this, we can take any representation (2.1.2)
of an elementary set and cut each brick in the representation by hyperplanes along all the sides of all the
bricks in the representation. We finally obtain the representation
m
B = ⊍ Ki0 , (2.1.3)
i=1
\
where the symbol ⊍ stands for the union of pairwise disjoint sets (here Ki0 Kj0 = ∅, i 6= j).

Theorem 2.1.10. E n is a ring.

Proof. It is clear that the intersection of two bricks is a brick. So, if

l
[ m
[
B= Kj , C= Ki0
j=1 i=1

1 The symbol ∨ denotes the exclusive “or”.

2.1. RINGS AND ALGEBRAS OF SETS 23

are two elementary sets, then the set

  !
\ l
[ \ m
[ l [
[ m  \ 
B C= Kj  Ki0 = Ki Kj0
j=1 i=1 i=1 j=1

is elementary, as well.
The difference of two bricks K \ K 0 is clearly an elementary set. Consequently, if K is a brick, and
[l
B= Ki is an elementary set, then
i=1

l
! l
[ \
K \B =K \ Ki = (K \ Ki )
i=1 i=1

is an elementary set too.

Let now B and C be two elementary sets. They are bounded as finite unions of bounded sets (bricks),
so there exists a brick K containing both sets. Therefore, the sets
[   [   \ 
B C=K\ K\ B C = K \ (K \ B) (K \ C)
\
B\C =B (K \ C).

are elementary.
The ring E n is not an algebra, since it contains no units. Nevertheless, the collection E n (K) of all
bricks containing in some brick K is an algebra whose unit is K.
Definition 2.1.11. An algebra A is called σ-algebra (sigma-algebra) if it is closed under countable union
of its elements, that is, if it possesses the property
∞
[
7) (Ak )∞
k=1 ⊂ A =⇒ A = Ak ⊂ A.
k=1

If A is a σ-algebra, then it also possesses the following property

∞
\
8) (Ak )∞
k=1 ⊂ A =⇒ A = Ak ⊂ A.
k=1

Indeed, if (Ak )∞
k=1 ⊂ A, then by the laws of duality

∞ ∞
!c
\ [
A= Ak = Ack ∈ A.
k=1 k=1

Here Ac = E \ A, where E is the unit of the algebra A.

Thus, σ-algebra is an algebra closed under countable unions and intersections of its elements. Among
the examples (2.1.4)–(2.1.7), the set P(X) is a σ-algebra only.
Example 2.1.12. Let X = R, and let

A = {A ⊂ A : A is countable or Ac is countable}.

It is easy to check that A is a σ-algebra.

Example 2.1.13. Let X = [0, 1], and let A = {∅, X, [0, 21 ], ( 12 , 1]}. Then A is a σ-algebra.
24 CHAPTER 2. INTRODUCTION TO THE MEASURE THEORY

Remark 2.1.14. For rings, there is a somewhat different situation. One has to differ σ-rings that are
closed under countable unions of their elements, and δ-rings that are closed under countable intersections
of their elements.

Let now X be a set, and let (Ak )∞

k=1 be a sequence of subsets of the set X.

Definition 2.1.15. The set A consisting of all the elements of X (and only them) that belong to infinitely
many sets Ak of the sequence (Ak )∞ ∞
k=1 , is called the limit superior of the sequence (Ak )k=1 . We will denote
it as A = lim sup Ak = lim Ak

Thus, a set A is the limit superior of the sequence (Ak )∞ k=1 if, and only if, for any x ∈ A, there exists
a sequence kj , j ∈ N, of indices such that x ∈ Akj , j = 1, 2, . . ..

Definition 2.1.16. The set A consisting of all the elements of X (and only them) that belong to all the
sets Ak starting from a fixed index, is called the limit inferior of the sequence (Ak )∞
k=1 . We will denote it
as A = lim inf Ak = lim Ak

Thus, a set A is the limit inferior of the sequence (Ak )∞k=1 if, and only if, for any x ∈ A, there exists
an index k0 = k0 (x) such that x ∈ Ak for all k > k0 .
It obviously follows from the definitions 2.1.15–2.1.16 that any sequence has the limit superior and the
limit inferior (probably empty sets), and that A ⊂ A. The embedding here can be strict as the following
example shows.

Example 2.1.17. Let Ak = [0, 1 + 1/k] for odd k, and Ak = [1 − 1/k, 2] for even k. Then we have
lim Ak = [0, 2], lim Ak = {1}.

Definition 2.1.18. A sequence of sets (Ak )∞

k=1 is called convergent if

lim Ak = lim Ak = A.

In this case, A is called the limit of the sequence (Ak )∞

k=1 .

Theorem 2.1.19. The following relations hold:

∞ ∞ ∞ ∞
! !
\ [ [ \
A= Al , A= Al . (2.1.4)
k=1 l=k k=1 l=k

Proof. Let us prove the first of the relations (2.1.4). Suppose that x ∈ A. Then, by definition, there
exists a sequence of indices kj such that x ∈ Akj , j = 1, 2, . . .. So, for any!k ∈ N, there exists a number
∞ \∞ [∞
S
l = kj > k, so that x ∈ Al for any k ∈ N. Consequently, x ∈ Al .
l=k k=1 l=k
∞ ∞ ∞
!
\ [ [
Conversely, let x ∈ Al . Then x ∈ Al for any k ∈ N. This means that for any k ∈ N, there
k=1 l=k l=k
exists a number l > k such that x ∈ Al , that is, x belongs to infinitely many sets of the sequence (Ak )∞
k=1 .
Prove now the second relation (2.1.4). Let x ∈ A. Then, by definition, there exists a index k = k(x)
∞ ∞ ∞
!
\ [ \
such that x ∈ Al for all l > k. So x ∈ Al , therefore, x ∈ Al .
l=k k=1 l=k
∞ ∞ ∞
!
[ \ \
Conversely, let x ∈ Al . This means that there exists a natural number k such that x ∈ Al ,
k=1 l=k l=k
that is, x ∈ Al for all l > k.

From this theorem we can obtain one more property of σ-algebras. Namely, if A is a σ-algebra, then
it possesses the following property
2.2. GENERAL MEASURE THEORY 25

9) (Ak )∞ ∞
k=1 ⊂ A =⇒ lim Ak , lim Ak ∈ A. Moreover, if the sequence (Ak )k=1 converges, then its limit
also belongs to A, lim Ak ∈ A.

This property is a consequence of the properties 7) and 8), and of the formulæ (2.1.4).

Definition 2.1.20. A sequence of sets (Ak )∞

k=1 is called monotone if it satisfies one the following relations

A1 ⊂ A2 ⊂ · · · ⊂ Ak ⊂ · · · (increasing sequence),

or
A1 ⊃ A2 ⊃ · · · ⊃ Ak ⊃ · · · (decreasing sequence).

Lemma 2.1.21. A monotone sequence of sets (Ak )∞ k=1 converges. Moreover, if the sequence is increasing,
∞
S ∞
T
then lim Ak = Ak , and if the sequence decreases, then lim Ak = Ak .
k=1 k=1

∞ ∞ ∞ ∞ ∞
! !
∞ ∞ \ [ \ [ [
Proof. Let (Ak )∞
S S
k=1 increases. Then Al = Al , therefore, Al = Al = Al .
l=k l=1 k=1 l=k k=1 l=1 l=1
∞
S
Consequently, A = Ak (see (2.1.4)).
k=1
∞ ∞ ∞ ∞
!
\ [ \ [
On the other hand, Al = Ak , so A = Al = Ak .
l=k k=1 k=l k=1
∞
[
Thus, A = A, which means that an increasing sequence of sets always converges, and lim Ak = Ak .
k=1
The corresponding fact for decreasing sequences can be proved analogously.

2.2 General measure theory

The concept of the measure of a set is a natural extension of the such concepts as the length of an interval,
the square of a flat body, the volume of a body, the increment ϕ(b)−ϕ(a) of a non-decreasing function ϕ(x)
on the interval [a, b], the integral of a nonnegative function over a certain domain, and many other notions.

Let K be a ring.

Definition 2.2.1. A function µ : K 7→ R+ := [0, ∞) is called a measure if it possesses the following

property

1) µ (A1 ⊍ A2 ) = µA1 + µA2 , for any A1 , A2 ∈ K.

The property 1) is called the additivity property of measures. Let us study some properties of measures.

2) µ∅ = 0.
Indeed, µ∅ = µ(∅ ⊍ ∅) = µ∅ + µ∅ = 2µ∅, so µ∅ = 0.

3) A ⊂ B =⇒ µA 6 µB (monotonicity).
Since B = A ⊍(B \ A) and since B \ A ∈ K, we have

µB = µA + µ(B \ A) > µA, (2.2.1)

due to the additivity and nonnegativity of measures.

4) A ⊂ B =⇒ µ(B \ A) = µB − µA.
This property follows from (9.1.1).
26 CHAPTER 2. INTRODUCTION TO THE MEASURE THEORY

S T
5) µ(A B) = µA + µB − µ(A B).
In fact, A B = A ⊍(B \ (A B)), so the property 5) follows now from the properties 1) and 4).
S T

 l  X l
6) µ ⊍ Ak = µAk (finite additivity).
k=1 k=1
This property follows from 1) by induction.
l
! l
[ X
7) µ Ak 6 µAk (finite semi-additivity).
k=1 k=1
 
k−1
[
Let us introduce the sets A01 = A1 , A0k = Ak \  Aj , k = 2, 3, . . . , l. Since from each set we
j=1
subtract all the elements that belong to the previous sets, we get A0k ⊂ Ak for all k, and A0k A0j = ∅
T

[l [l l
whenever k 6= j. Obviously, Ak = A0j = ⊍ A0j . Now from the properties 4) and 3) it follows
k=1 j=1 j=1

l
! ! l l
[ l X X
µ Ak =µ ⊍ A0j = µA0j 6 µAk .
k=1 j=1 j=1 k=1

∞ ∞
8) If A ⊃ ⊍ Ak , where A, Ak ∈ K, k ∈ N, then the series
P
µAk converges, and
k=1 k=1

∞
X
µA > µAk (2.2.2)
k=1

l
Since A ⊃ ⊍ Ak for any l ∈ N, so by the properties 3) and 6)
k=1

 l
 l
X
µA > µ ⊍ Ak = µAk , ∀l ∈ N.
k=1 k=1

∞
P
This estimate shows that the series µAk converges, and by tending l to infinity we get (9.2.2).
k=1

The properties 1)–8) exhaust all the properties of additive measures on a ring K. However, these
properties turn to be not enough in a lot of cases, so it is useful to introduce a specific class of measures.
Definition 2.2.2. A measure µ defined on a ring K is called countably additive, or σ-additive, if it possesses
the following property
∞
9) If A = ⊍ Ak , where A, Ak ∈ K, k ∈ N, then
k=1

∞
X
µA = µAk . (2.2.3)
k=1

Let us formulate two more definitions (properties).

Definition 2.2.3. A measure µ defined on a ring K is called countably semi-additive, or σ-semi-additive,
if it possesses the following property
2.2. GENERAL MEASURE THEORY 27

∞
S
10) If A ⊂ Ak , where A, Ak ∈ K, k ∈ N, then
k=1

∞
X
µA 6 µAk . (2.2.4)
k=1

Definition 2.2.4. A measure µ defined on a ring K is called continuous if it possesses the property
11) If a sequences (Ak )∞
k=1 (⊂ K) is monotone, and lim Ak ∈ K, then

µ(lim Ak ) = lim µAk . (2.2.5)

Theorem 2.2.5. The properties of σ-additivity, σ-semi-additivity, and the continuity of the measures are
equivalent.
Proof. Theorem asserts that if a measure possesses one of the properties 9)–11), then it possesses the rest
ones.
First, we prove that the property 9) is equivalent to the property 10).
Let the measure µ is σ-additive. Consider an arbitrary sequence of sets (Ak )∞
k=1 ⊂ K and a set A ∈ K
∞
S
such that A ⊂ Ak . Introduce the following new sets
k=1
 
\ 
\  k−1
[
A01 = A1 A, A0k = Ak A \ A0j  .
j=1

∞
It is easy to check that A0k ∈ K (k ∈ N), A0k A0j = ∅ (k 6= j), A0k ⊂ Ak (k ∈ N), and A = ⊍ A0k . Thus,
T
k=1
by the property 3), one has µA0k 6 µAk (k ∈ N), and by the property 9)
∞
X ∞
X
µA = µA0k 6 µAk ,
k=1 k=1

so 9) =⇒ 10).
Conversely, let the measure µ is σ-semi-additive, and let (Ak )∞
k=1 ∈ K be a sequence of disjoint sets,
∞ ∞
and A = ⊍ Ak ∈ K. Then the inclusion A ⊂
S
Ak holds, and due to the σ-semi-additivity of the
k=1 k=1
measure µ we have
∞
X
µA 6 µAk . (2.2.6)
k=1
∞
On the other hand, the opposite inclusion A ⊃ ⊍ Ak holds, as well, so by the property 8)
k=1

∞
X
µA > µAk . (2.2.7)
k=1

From (9.4.3)–(2.2.7), it follows that µA is σ-additive whenever it is σ-semi-additive.

Let us prove now equivalence of the properties 9) and 11). Suppose that µ is σ-additive. Let a sequence
(Ak )∞
k=1 (⊂ K) be monotone (see Definition 2.1.20) and A = lim Ak ∈ K.
If the sequence (Ak )∞
k=1 increases, then according to Lemma 2.1.21,

∞
[ ∞
A = lim Ak = Ak = ⊍ (Ak \ Ak−1 ) (A0 = ∅),
k=1 k=1
28 CHAPTER 2. INTRODUCTION TO THE MEASURE THEORY

so
∞
X m
X m
X
µA = µ(Ak \ Ak−1 ) = lim µ(Ak \ Ak−1 ) = lim (µAk − µAk−1 ) = lim µAm .
m→∞ m→∞ m→∞
k=1 k=1 k=1
by the properties 3), 4), and 9). Thus, 9) =⇒ 11) whenever the sequence (Ak )∞ k=1 is increasing.
Suppose now that (Ak )∞ k=1 is decreasing. Then the sequence (A 1 \ A ) ∞
k k=1 is increasing, and
∞ ∞
!
[ \
lim(A1 \ Ak ) = (A1 \ Ak ) = A1 \ Ak = A1 \ lim Ak .
k=1 k=1

Since the measure µ is proved to be continuous for increasing sequences, we have

µ(A1 \ lim Ak ) = lim µ(A1 \ Ak ),
or, equivalently,
µA1 − µ(lim Ak ) = lim(µA1 − µAk ) = µA1 − lim µAk ,
so
µ(lim Ak ) = lim(µAk ).
Thus, we proved the implication 9) =⇒ 11).
Conversely, let µ be a continuous measure, and let (Ak )∞
k=1 (⊂ K) be a sequence of disjoint sets such
∞ k
that A = ⊍ Ak ∈ K. Introduce the sets A0k = ⊍ Aj , k ∈ N. It is clear that the sequence (A0k )∞
k=1 is
k=1 j=1
increasing, and according to Lemma 2.1.21,
∞ ∞
[
A = ⊍ Ak = A0k = lim A0k .
k=1 k=1

Therefore, by the continuity and finite additivity of the measure µ, one obtains
∞
! k
k X X
0
µA = lim µAk = lim µ ⊍ Aj = lim µAj = µAj .
j=1 j=1 j=1

Thus, 11) implies 9), as required.

Theorem 2.2.6. Let µ be a σ-additive measure defined on a σ-algebra A, and let (Ak )∞k=1 be an arbitrary
sequence of elements of the algebra A. Then the measure µ possesses the following properties:
12) µ(lim Ak ) > lim µAk ,
13) µ(lim Ak ) 6 lim µAk ,
14) if the sequence (Ak )∞
k=1 converges, then

µ(lim Ak ) = lim µAk .

Proof. We, first, prove the property 12). From (2.1.4) it follows that
∞ ∞
!
\ [
A = lim Ak = Al .
k=1 l=k
∞
S
Consider the sequence of the sets Bk = Al , k ∈ N that belong to A, since it is a σ-algebra by
l=k
∞
Al decreases, so the sequence (Bk )∞
S
assumption. As k grows the union k=1 is decreasing, and lim Ak =
l=k
∞
T
Bk = lim Bk . Since the measure µ is continuous by Theorem 2.2.5, we have
k=1

µ(lim Ak ) = lim µBk . (2.2.8)

2.2. GENERAL MEASURE THEORY 29

Consider now the sequence (µAk )∞ k=1 . It is bounded from above by the number µE where E is the
unit of the algebra A, therefore, there exists finite lim µAk . Consequently, in the sequence (Ak )∞ k=1 one
can find a subsequence (Akj )∞
j=1 converging to lim µA k . Since obviously B kj ⊃ A kj , from (2.2.8) and from
the monotonicity of the measure µ (the property 3)) we obtain

µ(lim Ak ) = lim µBk = lim µBkj > lim µAkj = lim µAk ,
k j j

so the property 12) is true. The property 13) can be proved analogously.
Suppose now that the sequence (Ak )∞k=1 converges. Then

lim Ak = lim Ak = lim Ak ,

so by the properties 12) and 13)

lim µAk 6 µ(limAk ) = µ(lim Ak ) = µ(limAk ) 6 lim µAk 6 lim µAk ,

therefore,
µ(lim Ak ) = lim µAk ,
as required.
The difference between the property 11) and the property 14) of measures on σ-rings and on σ-algebras
is that in the property 11) we have to consider only monotone sequences, while in the property 14) we
avoid this restriction.
Definition 2.2.7. If there is defined a measure on a ring K, then the sets of the ring K are called
measurable w.r.t. the measure µ, or µ-measurable.
30 CHAPTER 2. INTRODUCTION TO THE MEASURE THEORY

2.3 Problems
Problem 2.1. Prove that the definition of the ring given in the class is equivalent to the following one: a
non-empty systemTof sets K is called a ring if it possesses the following two properties
1) A, B ∈ K ⇒ A B ∈ K;
2) A, B ∈ K ⇒ A M B ∈ K.
Problem 2.2. Let X = {a, b, c} (a set consisting of three elements), and let P(X) is the set of all subsets
of X.
a) Describe all rings that can be constructed with the elements of P(X).
b) Describe all algebras that can be constructed with the elements of P(X).
Problem 2.3. Prove that
a) intersection of rings is a ring,
b) intersection of algebras with common unit is an algebra,

c) intersection of σ-algebras with common unit is a σ-algebra.

Problem 2.4. Let A(X) be σ-algebra with the unit X, and let X0 ∈ A(X). Show that the collection of
sets \
A(X0 ) = {A X0 : A ∈ A(X)}
is a σ-algebra.
Problem 2.5. Let X be a set, and let (En )n∈N be a sequence of sets such that En ∈ X for any n ∈ N.
Prove that
X\(limEn ) = lim(X\En ); X\(limEn ) = lim(X\En ).
Problem 2.6. Let X be a set, and M(X) be a σ-algebra of subsets of X. Suppose a σ-additive measure
µ is defined on M(X). Prove that the collection of all subsets of X of measure 0 is a σ-ring.
Problem 2.7. Let ϕ be an additive real function defined on a ring of sets K. Prove that

1) if ϕ is finite on at least one element of K, then ϕ(∅) = 0,

2) for any A, B ∈ K,
ϕ(A ∪ B) + ϕ(A ∩ B) = ϕ(A) + ϕ(B),

3) if G is the set of all elements of K such that ϕ(A) is finite for any A ∈ G, then G is a ring.
Problem 2.8. Let a measure µ (a nonnegative additive function) be defined on a ring K. Prove that if

µ(A M B) = 0,

then µ(A) = µ(B).

Problem 2.9. On the algebra of all subsets of rational numbers in [0, 1], define a measure such that the
measure of each rational number is positive and the measure of the whole set of all rational numbers on
[0, 1] equals 1.
Chapter 3

The Lebesgue measure on Rn

The Lebesgue measure is a natural extension of the concepts of the length of an interval, of the square
of flat bodies, of the volume of bodies, and it is an extension of the Jordan measure (see homework for
definition) to a more wide class of sets with the additional property of σ-additivity. Here we construct the
Lebesgue measure in three steps. First, we define the measure on bricks, then we extend it to the ring E n
of all elementary sets, and then we, finally, extend the obtained measure to a more wide class of sets which
we call the set of Lebesgue measurable sets.

3.1 The measure on bricks

Let K be a brick (see Definition 2.1.8). The following number
n
Y
m0 K = (bi − ai ) (3.1.1)
i=1

is called the measure of the brick K.

From (3.1.1) it is easy to see that m0 K is the length of an interval for n = 1, the square of a rectangle
for n = 3, and the volume of a parallelepiped for n = 3. It is also clear that if a brick is degenerate, that
is, bi = ai for at least one index i, then m0 K = 0.
The measure of bricks introduced by (3.1.1) possesses the following properties:

1) m0 K > 0,

l l
X
2) K = ⊍ Kj =⇒ m0 K = m0 Kj .
j=1 j=1

The property 1) is obvious. For brevity, we prove the property 2) for the case R2 .
Let, first, the brick K is split into the bricks Kj by vertical lines

x = xs , 0 6 s 6 p, x0 = a1 , xp = b1 ,

and by horizontal lines

y = yt , 0 6 t 6 q, y0 = a2 , yq = b2 , pq = l.
Then we have
l
X p X
X q
m0 Kj = (xs − xs−1 )(yt − yt−1 ) = (b1 − a1 )(b2 − a2 ) = m0 K.
j=1 s=1 t=1

31
32 CHAPTER 3. THE LEBESGUE MEASURE ON RN

l
If now K = ⊍ Kj is an arbitrary representation of K as a union of disjoint bricks Kj , then we split
j=1
the bricks Kj by vertical and horizontal lines drawn along the all sides of all the bricks Kj . The obtained
new bricks we denote as Ki0 , 1 6 i 6 r. Thus,
l r
K = ⊍ Kj = ⊍ Ki0 ,
j=1 i=1

so  
r
X l
X X l
X
m0 K = m0 Ki0 =  m0 Ki0  = m0 Kj .
i=1 j=1 i:Ki0 ⊂Kj j=1

as we showed above.
Remark 3.1.1. Generally speaking, the measure m0 on bricks is not a measure by Definition 2.2.1, since
the collection of all bricks in Rn is not a ring. However, one can introduce the notion of semi-rings of sets
(the collection of bricks is a semi-ring) and then to define measures on semi-rings. For details, see e.g. [6].

3.2 The measure on E n

Now we define a measure on the ring E n of elementary sets in Rn . Let B be an elementary set (see
Definition 2.1.9). It can be represented (see (2.1.3)) as a union of disjoint bricks Kj0
r
B = ⊍ Ki0 .
i=1

We set
r
X
mB := m0 Ki0 . (3.2.1)
i=1

First of all, we must show that the measure (3.2.1) is defined correctly, that is, that mK = m0 K for
any brick, and that the value mB does not depend on the representation of B as a union of bricks.
Let
l r
B = ⊍ Kj = ⊍ Ki0 .
j=1 i=1

We want to show that

l
X r
X
m0 Kj = m0 Ki0 . (3.2.2)
j=1 i=1

To do this, let us introduce the following sets

\
Ke j,i = Kj Ki0 , j = 1, 2, . . . , l, i = 1, 2, . . . , r.

The sets K e j,i are bricks as intersections of two bricks. Moreover, since the bricks Kj and K 0 are pairwise
i
1
disjoint , we have \
K
e j ,i
1 1
Ke j ,i = ∅,
2 2
(j1 , i1 ) 6= (j2 , i2 ).
r l
Obviously, Kj = ⊍ K
e j,i for any j = 1, . . . , l, and K 0 = ⊍ K
i
e j,i for any i = 1, . . . , r. Since the “measure”
i=1 j=1
m0 is additive (see the property 2) of m0 ) on bricks, one gets
r
X l
X
m0 Kj = m0 K
e j,i , j = 1, . . . , l, m0 Kj0 = m0 K
e j,i , i = 1, . . . , r.
i=1 j=1

1 That Ki01 Ki02

T T
is, Kj1 Kj2 = ∅ if j1 6= j2 , and, respectively, = ∅ if i1 6= i2 .
3.2. THE MEASURE ON E N 33

Therefore, !
l l r l r r l r
 
⊍ Kj = ⊍ ⊍K
e j,i e j,i = ⊍
= ⊍ ⊍K ⊍K
e j,i = ⊍ Ki0 .
j=1 j=1 i=1 j=1 i=1 i=1 j=1 i=1

0
Now from additivity of m on bricks, we obtain
!  
l
X l
X r
X l X
X r r
X l
X r
X
m0 Kj = m0 K
e j,i = m0 K
e j,i =  m0 K
e j,i  = m0 Ki0 ,
j=1 j=1 i=1 j=1 i=1 i=1 j=1 i=1

so the identity (3.2.2) is proved. In particular, from (3.2.2), it follows that

mK = m0 K

for any brick, since the identity K = K is one of the representations of the elementary set K as a finite
union of bricks. Thus, the measure m is an extension of the “measure” m0 to the ring E n of all elementary
sets in Rn .
As well as m0 , the measure m possesses the following properties:

1) mB > 0, ∀B ∈ E n ,
l l
2) B = ⊍ Bj =⇒ mB =
P
mBj .
j=1 j=1

The first property is evident. Moreover, it is sufficient to prove the property 2) for the case l = 2. So,
r1 r2
let B1 = ⊍ K1,j and B2 = ⊍ K2,j . Then we have
j=1 j=1
! !
r1 r2 2 ri
B = B1 ⊍ B2 = ⊍ K1,j ⊍ ⊍ K2,j = ⊍ ⊍ Ki,j ,
j=1 j=1 i=1 j=1

so  
X ri
2 X 2
X Xri
mB = m0 Ki,j =  m0 Ki,j  = mB1 + mB2 .
i=1 j=1 i=1 j=1

Consequently, the function m defined on the ring E n is a measure according to the Definition 2.2.1,
therefore, it possesses the properties 2) − 8) of measures.

Theorem 3.2.1. The measure m on E n is σ-semi-additive.

∞
S
Proof. We need to prove that if B and Bk , k ∈ N, are elementary sets, and if B ⊂ Bk , then
k=1

∞
X
mB 6 mBk . (3.2.3)
k=1

If the series in the right hand side of (3.2.3) diverges, then we are done. Suppose now that this series
converges, and fix some number ε > 0. For the set B, one can find a closed elementary set B 0 ⊂ B such
that
ε
mB 0 > mB − . (3.2.4)
2
r
It can be done as follows. Let B = ⊍ Ki . In each brick Ki we choose a closed brick K 0,i so that
i=1

ε
m0 K 0,i > m0 Ki − ,
2r
34 CHAPTER 3. THE LEBESGUE MEASURE ON RN

and set
r
B 0 = ⊍ K 0,i .
i=1

So, we obtain
r r  r
X X ε X 0 ε ε
mB 0 = m0 K 0,i > m0 Ki − = m Ki − = mB − .
i=1 i=1
2r i=1
2 2

Furthermore, for each number k ∈ N, let us choose an open elementary set B

ek ⊃ Bk such that

ε
mB
ek < mBk + . (3.2.5)
2k+1

(It can be done in the same manner as we chose B 0 above.)

Then we have
∞
[ [∞
B0 ⊂ B ⊂ Bk ⊂ B
ek .
k=1 k=1

By the Heine-Borel theorem, from every cover of the closed set B 0 by open sets, one can choose a finite
subcover, that is, to choose sets B
ek , j = 1, . . . , l, such that
j

l
[
B0 ⊂ B
ek .
j
j=1

As we mentioned above, the measure m is finite semi-additive (satisfies the property 7) of measures).
Therefore,
X l
mB 0 6 mBek .
j
j=1

Now from (3.2.4)–(3.2.5) we obtain

l ∞ ∞ ∞
ε X e ε X e ε X ε  ε X
mB < mB 0 + 6 mBkj + 6 mBk + < mBk + k+1 + = mBk + ε.
2 j=1 2 2 2 2
k=1 k=1 k=1

Since ε is arbitrary, the inequality (3.2.3) holds.

A simple consequence of Theorems 3.2.1 and 2.2.5 is the following fact.

Theorem 3.2.2. The measure m on the ring E n is σ-additive.

Thus, by the formula (3.2.1), we define the σ-additive measure m on E n . Our next goal is to extend
this measure onto a wider class of sets keeping its σ-additivity.

3.3 The outer measure

Let A(⊂ Rn ) be an arbitrary set.

Definition 3.3.1. The number

 
X [ 
µ∗ A = inf m0 Kj : A ⊂ Kj (3.3.1)
 
j j

is called the outer measure of the set A.

3.3. THE OUTER MEASURE 35

We emphasize that the infimum in (3.3.1) is taken over all covers of the set A by finite or infinite
collections of bricks. Also notice that µ∗ A can be infinite (for example, in the case A = Rn ).
Let us study properties of the measure µ∗ .

1) µ∗ A > 0, A ⊂ Rn .
This property is evident.

2) A1 ⊂ A2 =⇒ µ∗ A1 6 µ∗ A2 (The outer measure is monotone).

Since every cover of the set A2 is also a cover of the set A1 , we have
   
X [  X [ 
µ∗ A1 = inf m0 Kj : A1 ⊂ Kj 6 inf m0 Kj : A2 ⊂ Kj = µ∗ A2 .
   
j j j j

[ X
3) A ⊂ Ak =⇒ µ∗ A 6 µ∗ Ak (The outer measure is σ-semi-additive).
k k
Here (Ak )k>0 is a finite or countable collection of sets. As we mentioned above, the outer measure
can be infinite. If one of the measures µ∗ Ak is infinite or all of them are finite but the series
P ∗
µ Ak
k
is infinite, then the inequality µ∗ A 6 µ∗ Ak holds. Suppose now that all the measures µ∗ Ak are
P
k
finite and X
µ∗ Ak < +∞.
k

∞
Let us fix a number ε > 0. By Definition
S 3.3.1, for every k there exists a collection of bricks (Kk,j )j=1
(finite or countable) such that Ak ⊂ Kk,j and
j

X ε
µ∗ Ak > m0 Kk,j − .
j
2k

[ [[
Moreover, A ⊂ Ak ⊂ Kk,j , so by Definition 3.3.1,
k k j

 
XX X X X ε X ∗
µ∗ A 6 m0 Kk,j =  m0 Kk,j  < m∗ Ak + k 6 µ Ak + ε.
j j
2
k k k k

Since ε is arbitrary, the inequality µ∗ A 6 µ∗ Ak holds.

P
k

4) A ∈ E n =⇒ µ∗ A = mA (on E n the outer measure coincides with m).

r r
Let A = ⊍ Ki be an elementary set. Since ⊍ Ki is a cover of the set A and since µ∗ A is the
i=1 i=1
infimum over all covers, we have
r
X
µ∗ A 6 m0 Ki = mA. (3.3.2)
i=1

Kj0 is an arbitrary cover of the set A, then due to σ-additivity of the

S
On the other hand, if
j
measure m (by Theorem 3.2.1), one obtains
X X
mA 6 mKj0 = m0 Kj0 ,
j j
36 CHAPTER 3. THE LEBESGUE MEASURE ON RN

therefore,  
X [ 
mA 6 inf m0 Kj0 : A ⊂ Kj = µ∗ A. (3.3.3)
 
j j

Now the property 4) follows from (3.3.2)–(3.3.3).

Example 3.3.2. Let A = {ai : ai ∈ Rn , i ∈ N} be a countable set. Then µ∗ A = 0.

ε
Indeed, let ε > 0 be arbitrary, and for every i ∈ N, we take a brick Ki such that ai ∈ Ki and m0 Ki < i
p S 2
(for example, Ki is the n-dimensional cube with edges of length n ε/2i ). Thus, we have A ⊂ Ki , and
i

∞ ∞
X X 1
µ∗ A 6 m0 Ki < ε = ε,
i=1 i=1
2i

so µ∗ A = 0, since ε is arbitrary.

Let us introduce the class of sets in Rn with finite outer measure and denote it M∗ (Rn ), that is,

M∗ (Rn ) = {A ⊂ Rn : µ∗ A < +∞}.

From the properties 2) and 3) of outer measure, it follows that M∗ (RS

n
) is a ring, since if A1 , A2 ∈
M (R ), then A1 A2 and A1 \ A2 (⊂ A1 ) also have finite measures, so A1 A2 ∈ M∗ (Rn ) and A1 \ A2 ∈
∗
n
S
M∗ (Rn ).

3.4 The Lebesgue measure

Now we are in a position to introduce the Lebesgue measure and the Lebesgue measurable sets on Rn .

Definition 3.4.1. A set A ∈ M∗ (Rn ) is called Lebesgue measurable if for any ε > 0 there exists an
elementary set B such that
µ∗ (A M B) < ε. (3.4.1)
The class of all Lebesgue measurable sets is denoted as M(Rn ).

Definition 3.4.2. The function µ∗ restricted to M(Rn ) is called the Lebesgue measure, and is denoted
µ.

Our next goal is to show that M(Rn ) is a ring, and that the function µ defined on this ring is a measure
according to Definition 2.2.1. But before that we define the geometrical meaning of Lebesgue measurable
sets. In fact, as it follows from (3.4.1), any Lebesgue measurable set can be approximated by elementary
sets arbitrary accurately. That is, for any Lebesgue measurable set A, one can find an elementary set B
such that A and B protrude from each other “not too far”.
Before we start to solve the intended problem, let us notice the following properties of the function µ.

1) µA > 0, ∀A ∈ M(Rn ).
This property is obvious.

2) E n ⊂ M(Rn ), and if A ∈ E n , then µA = mA.

Indeed, if A ∈ E n , then according to the property 4) of the outer measure, A ∈ M∗ (Rn ), since
µ∗ A = mA < +∞. Let us set B := A, then for any ε > 0, we have

µ∗ (A M B) = µ∗ ∅ = 0 < ε,

therefore, A ∈ M(Rn ), and by the property 4) of the outer measure µA = µ∗ A = mA.

3.4. THE LEBESGUE MEASURE 37

Lemma 3.4.3. The following relations hold:

S S S
a) (A1 A2 ) M (B1 B2 ) ⊂ (A1 M B1 ) (A2 M B2 ),
T T S
b) (A1 A2 ) M (B1 B2 ) ⊂ (A1 M B1 ) (A2 M B2 ),
S
c) (A1 \ A2 ) M (B1 \ B2 ) ⊂ (A1 M B1 ) (A2 M B2 ),
S
d) (A1 M A2 ) M (B1 M B2 ) ⊂ (A1 M B1 ) (A2 M B2 ).

Proof. We prove here S the relations

S a) and c). The relations b) S and d) can be proved
S analogously.S
a) Let
S x ∈ (A 1 A )
2S M (B 1 B 2 ). Then
S either x ∈ A 1 A 2 but x ∈
6 B 1 B 2 or x ∈ B1
, B2 but
x 6∈ A1 A2 . If x ∈ A1 A2 but x 6∈ B1 BS 2 , then x ∈ A 1 or x
S ∈ A 2 but x ∈
6 B 1 and x ∈
6 B 2 . Therefore,
x ∈ A1 M B1 or x ∈ A2 M B2 . If x ∈ B1 B2 but x 6∈ A1 A2 , then the inclusion a) can be proved
analogously.
c) Let x ∈ (A1 \ A2 ) M (B1 \ B2 ). Then either x ∈ A1 \ A2 but x 6∈ B1 \ B2 , or x 6∈ A1 \ A2 but
x ∈ B1 \ B2 . If x ∈ A1 \ A2 but x 6∈ B1 \ B2 , then x ∈ A1 but x 6∈ A2 , and either x ∈ B1 and x ∈ B2 , or
x 6∈ B1 . This means that either x ∈ A1 M B1 , or x ∈ A2 M B2 . If x 6∈ A1 \ A2 but x ∈ B1 \ B2 , then the
inclusion c) can be prove analogously.

Lemma 3.4.4. For any A1 , A2 ∈ M∗ (Rn ), the following inequality holds

|µ∗ A1 − µ∗ A2 | 6 µ∗ (A1 M A2 ) (3.4.2)

(A1 M A2 ). By additivity of the outer measure µ∗ , one has

S
Proof. It is easy to see that A1 ⊂ A2

µ∗ A1 6 µ∗ A2 + µ∗ (A1 M A2 ),

that is,
µ∗ A1 − µ∗ A2 6 µ∗ (A1 M A2 ).
By interchanging A1 and A2 , we get

µ∗ A2 − µ∗ A1 6 µ∗ (A1 M A2 ).

The last two inequalities give us (3.4.2), as required.

Theorem 3.4.5. The set M(Rn ) is a ring.

Proof. Due to definition of rings, it is sufficient to prove that if A1 , A2 ∈ M(Rn ), then A1 A2 ∈ M(Rn )
S
n
and A1 \ A2 ∈ M(R ).
So, let A1 , A2 ∈ M(Rn ). Then for an arbitrary ε > 0, one can find elementary sets B1 and B2 such
that
ε ε
µ∗ (A1 M B1 ) < and µ∗ (A2 M B2 ) < .
2 2
Let us denote A := A1 A2 and B := B1 B2 . it is clear that A1 , A2 ∈ M∗ (Rn ), so A ∈ M∗ (Rn ). The
S S
set B is an elementary set as the union of two elementary sets. By Lemma 3.4.4,
[
A M B ⊂ (A1 M B1 ) (A2 M B2 ).

Then by the property of semi-additivity of the outer measure, we obtain

ε ε
µ∗ (A M B) 6 µ∗ (A1 M B1 ) + µ∗ (A2 M B2 ) < + = ε,
2 2
so A = A1 A2 ∈ M(Rn ).
S
The proof of the fact that A1 \ A2 ∈ M(Rn ), can be proved in the same manner.
38 CHAPTER 3. THE LEBESGUE MEASURE ON RN

Remark 3.4.6. In fact, using Lemma 3.4.4, it is possible to prove that if A1 , A2 ∈ M(Rn ), then A1 A2 ∈
T
M(Rn ) and A1 M A2 ∈ M(Rn ). But this also follows from Theorem 3.4.5 and the properties 4) and 5) of
rings.
Corollary 3.4.7. The set M([0, 1]) of all Lebesgue measurable subsets of the interval [0, 1] is an algebra.
Proof. Since M([0, 1]) ⊂ M(R) and [0, 1] ∈ M([0, 1]), the set M([0, 1]) is a ring with the unit [0, 1].
Theorem 3.4.8. The function µ is additive on M(Rn ).
Proof. We have to prove that if A = A1 ⊍ A2 , then

µA = µA1 + µA2 . (3.4.3)

So, let A1 , A2 ∈ M(Rn ) and A = A1 ⊍ A2 . By Theorem 3.4.5 A ∈ M(Rn ). Since the outer measure is
semi-additive by the property 3) of the outer measure, and since µ∗ coincides with µ on M(Rn ), we have

µA 6 µA1 + µA2 (3.4.4)

To prove the opposite inequality, let us fix a number ε > 0 and take elementary sets B1 and B2 such that
ε ε
µ∗ (A1 M B1 ) < and µ∗ (A2 M B2 ) < . (3.4.5)
6 6
Such sets B1 and B2 exist,
S since A1 and A2 are Lebesgue
T measurable by assumption.
Let us set B := B1 B2 . By assumption A1 A2 = ∅, but B1 and B2 can have common elements.
However, from Lemma 3.4.3 it follows that
\ \ \ [
B1 B2 = (A1 A2 ) M (B1 B2 ) ⊂ (A1 M B1 ) (A2 M B2 ),

so \ \ ε ε ε
m(B1 B2 ) = µ∗ (B1 B2 ) 6 µ∗ (A1 M B1 ) + µ∗ (A2 M B2 ) < + = .
6 6 3
This inequality together with the property 5) of measures on rings imply
 \  ε
mB = mB1 + mB2 − m B1 B2 > mB1 + mB2 − . (3.4.6)
3
Now, according to Lemma 3.4.4, we obtain
ε
mB1 > µA1 − µ∗ (A1 M B1 ) > µA1 − , (3.4.7)
6
and, analogously,
ε
mB2 > µA2 − . (3.4.8)
6
S
Recall the relation A M B ⊂ (A1 M B2 ) (A1 M B2 ) (Lemma 3.4.3). By semi-additivity of the outer
measure, one has
ε
µ∗ (A M B) 6 µ∗ (A1 M B2 ) + µ∗ (A1 M B2 ) < ,
3
so from Lemma 3.4.4 it follows that
ε
µA > mB − µ∗ (A M B) > mB − . (3.4.9)
3
Now the estimates (3.4.6)–(3.4.9) imply
ε 2ε
µA > mB − > mB1 + mB2 − > µA1 + µA2 − ε.
3 3
Since ε > 0 is arbitrary, we have
µA > µA1 + µA2 .
This inequality together with (3.4.4) implies (3.4.3).
3.4. THE LEBESGUE MEASURE 39

Theorem 3.4.9. The measure µ defined on the ring M(Rn ) is σ-additive.

Proof. The measure µ is the restriction of the outer measure µ∗ to the class M(Rn ). Since the function
µ∗ is σ-semi-additive, so is the measure µ on the ring M(Rn ). Now by Theorem 2.2.5, the measure µ is
σ-additive.
∞
Theorem 3.4.10. Let A ∈ M∗ (Rn ) and A = Ak , where Ak ∈ M(Rn ), k ∈ N. Then A ∈ M(Rn ).
S
k=1

In other word, if the outer measure of a set is finite, and the set can be represented as a countable
union of Lebesgue measurable sets, then it is Lebesgue measurable.

k−1
Proof. Let us set A0 = ∅, and A0k = Ak \ Aj , k ∈ N. It is easy to see that by construction A0k A0l = ∅
S T
j=1
whenever k 6= l. At the same time, every point of A belongs to (at least) one of the sets A0k . Therefore,
∞
A = ⊍ A0k , and A0k ∈ M(Rn ), since M(Rn ) is a ring by Theorem 3.4.5.
k=1
l
Now, since ⊍ A0k ⊂ A for any l ∈ N, by the monotonicity of the outer measure (the property 2), and
k=1
by additivity of the Lebesgue measure, we obtain
l
X  l
  l

µA0k =µ ⊍ A0k =µ ∗
⊍ A0k 6 µ∗ A < +∞.
k=1 k=1 k=1

∞
µA0k with nonnegative terms are bounded from above, so the series
P
Thus, the partial sums of the series
k=1
∞
µA0k converges.
P
k=1
Let ε > 0 be an arbitrary positive number. Then there exists a number m0 ∈ N such that
∞
X ε
µA0k < .
2
k=m0 +1

m0
The set ⊍ A0k is Lebesgue measurable as a finite union of Lebesgue measurable sets. Therefore, there
k=1
exists an elementary set B such that
 m  
0 ε
µ∗ ⊍ k A 0
M B <
k=1 2

It is easy to check that !

∞
 m  [
0
AMB⊂ ⊍ A0k M B ⊍ A0k .
k=1 k=m0 +1

Thus, from the property of σ-additivity of the outer measure, we obtain

 m   ∞
∗ ∗
0
0
X ε ε
µ (A M B) 6 µ ⊍ Ak M B + µA0k < + < ε,
k=1 2 2
k=m0 +1

as required.

Corollary 3.4.11. The union of countably many null sets (sets whose Lebesgue measure is 0) has Lebesgue
measure 0, that is, it is a null set.
40 CHAPTER 3. THE LEBESGUE MEASURE ON RN

∞
Ak , where Ak ∈ M(Rn ) and µ(Ak ) = 0, k ∈ N. Then due to σ-semi-additivity of the
S
Proof. Let A =
k=1
outer measure, we have
∞
X
µ∗ (A) 6 µ(Ak ) = 0.
k=1

Now by Theorem 3.4.10, the set A is Lebesgue measurable, so its Lebesgue measure coincides with its
outer measure and equals 0.
Definition 3.4.12. A measure µ defined on a ring K is called complete if every subset of a null set is
measurable w.r.t. the measure µ.
Thus, if a measure µ defined on a ring K is complete, then by the monotonicity of measures, every
subset of a null set is a null set.
Theorem 3.4.13. The Lebesgue measure is complete.
Proof. Let µA = 0, A ∈ M(R), and let A0 ⊂ A. Then for any ε > 0 and for B = ∅ ∈ E n , we have

µ∗ (A0 M B) = µ∗ (A0 M ∅) = µ∗ A0 6 µ∗ A = µA = 0 < ε.

Thus, by Definition 3.4.1, the set A0 is Lebesgue measurable.

3.5 Extension of the notion of measurability. The class of mea-

surable sets.
The Lebesgue measure defined above possesses a substantial defect. Namely, only sets with finite (outer)
measure can be Lebesgue measurable. So, many “good” infinite sets left unmeasurable w.r.t. our Lebesgue
measure (e.g. Rn , quadrants, strips, interiors of parabolas, etc). Let us eliminate this defect.
Let Q
b l be an n-dimensional cube with the centre at the origin and with edge of length 2l, that is,

b l = {x = (xi )ni=1 : |xi | 6 l, i = 1, 2, . . . , n}, l ∈ R.

Q (3.5.1)

Definition 3.5.1. AT set A ⊂ Rn is called Lebesgue measurable in extended sense, or σ-measurable, if for
any l ∈ N the set A Q b l is Lebesgue measurable.

If a set A is σ-measurable, then we put

 \ 
µA := lim µ A Q
bl . (3.5.2)
l→∞

The class of all σ-measurable sets is denoted as Mσ (Rn ).  T 

Tb
Since the sequences A Q l increases as l grows, then the sequences µ A Q
b l also increases. There-
fore, the limit in (3.5.2) exists (but can be infinite).
Theorem 3.5.2. M(Rn ) ⊂ Mσ (Rn ).

Proof. Let A ∈ M(Rn ). Then for any l ∈ N, the set A Q

Tb
l is a Lebesgue measurable set as an intersection
of two Lebesgue measurable sets. Then by Definition 3.5.1, the set A is σ-measurable.
Theorem 3.5.3. If for a set A ∈ Mσ (Rn ) the limit (3.5.2) is finite, then A ∈ M(Rn ).
Tb
Proof. Let Al := A Q l , l ∈ N. The sequence Al is increasing, and

∞
[ ∞
A= Al = ⊍ (Al \ Al−1 ), (A0 = ∅).
l=1 l=1
3.5. EXTENSION OF THE NOTION OF MEASURABILITY. THE CLASS OF MEASURABLE SETS.41

Since A ∈ Mσ (Rn ), the sets Al are Lebesgue measurable, and since M(Rn ) is a ring, Al \ Al−1 ∈ M(Rn )
for any l ∈ N.
The existence of finite limit (3.5.2), that is, of lim µAl , means that the following series converges
l→∞

∞
X
µ(Al \ Al−1 ),
l=1

since
∞
X l
X l
X
µ(Al \ Al−1 ) = lim µ(Ak \ Ak−1 ) = lim (µAk − µAk−1 ) = lim µAl .
l→∞ l→∞ l→∞
l=1 k=1 k=1

Now from σ-additivity of the outer measure, we have

∞
X ∞
X
µ∗ A 6 µ∗ (Al \ Al−1 ) = µ(Al \ Al−1 ) < +∞.
l=1 l=1

Therefore, A ∈ M∗ (Rn ), so A ∈ M(Rn ) by Theorem 3.4.10.

Theorem 3.5.3 shows that extension of M(Rn ) to Mσ (Rn ) is made by adding the sets of infinite
measure to M(Rn ).

Theorem 3.5.4. The set Mσ (Rn ) is a σ-algebra with the unit Rn .

Proof. First, we prove that Mσ (Rn ) is a ring. Let A, B ∈ Mσ (Rn ), then for any l ∈ N, the sets A
Tb
Ql
Tb
and B Q l are Lebesgue measurable. But
[ \ \ [ \
(A B) Q b l = (A Q b l ) (B Q b l ) ∈ M(Rn ),

and \ \ \
(A \ B) Q
b l = (A b l ) \ (B
Q b l ) ∈ M(Rn ).
Q

So A B, A \ B ∈ Mσ (Rn ). Obviously, Rn is the unit of Mσ (Rn ), therefore, Mσ (Rn ) is an algebra. It is

S
left to show that it is a σ-algebra.
∞
Consider Ak ∈ Mσ (Rn ), k ∈ N, and put A := n
S Tb
Ak . Then for any l ∈ N, one has Ak Q l ∈ M(R ),
k=1
and
∞ ∞
!
\ [ \ [ \
A Q
bl = Ak Q
bl = (Ak bl ) ⊂ Q
Q bl .
k=1 k=1

∗b l ) < +∞, by monotonicity of the outer measure we obtain A T Q b l ∈ M∗ (Rn ). Now from
Since µ (Q
Theorem 3.4.10 it follows that A Ql ∈ M(Rn ), so Mσ (Rn ) is a σ-algebra.
Tb

Corollary 3.5.5. The set M([0, 1]) of all Lebesgue measurable subsets of the interval [0, 1] is a σ-algebra.

Proof. By Corollary 3.4.7, M([0, 1]) is an algebra. If Ak ∈ M([0, 1]), k ∈ N, then

∞
[
A= Ak ⊂ [0, 1],
k=1

so, as in the proof of Theorem 3.5.4, µ∗ (A) 6 µ∗ ([0, 1]) = µ([0, 1]) = 1. Thus, A ∈ M∗ ([0, 1]), and
according to Theorem 3.4.10, one has A ∈ M([0, 1]), as required.

Remark 3.5.6. The corollary is true for M(X) where X is a compact set in Rn .
42 CHAPTER 3. THE LEBESGUE MEASURE ON RN

Definition 3.5.7. The triple (X, M, µ e) is called a measure space. If µe(X) < +∞, then the measure µe is
called finite. If additionally µ(X) = 1, it is called a probability measure. If µ(X) = +∞, but there exist
∞
S
sets Xi such that µ e(Xi ) < +∞, i ∈ N, and X = Xi , then the measure µ
e is called σ-finite.
i=1

n
The Lebesgue measure on Mσ (R ) is σ-finite.

Now let us discuss how large the class Mσ (Rn ) is.

Definition 3.5.8. Let K1 and K2 be two bricks, and letTK1◦ and K2◦ denote their interiors, respectively.
We say that bricks K1 and K2 are almost disjoint if K1◦ K2◦ = ∅, meaning that they intersect at most
along their boundaries.

Theorem 3.5.9. Every open set in Rn , n > 1, can represented as a countable union of almost disjoint
closed bricks.

Proof. Let A ⊂ Rn be open. We construct a family of closed cubes (bricks of equal sides) as follows.
First, we bisect Rn into almost disjoint closed cubes {Qi : i ∈ N} of side one with integer coordinates. If
Qi ⊂ A, we include Qi in the family, and if Qi is disjoint from A, we exclude it. Otherwise, we bisect the
sides of Qi to obtain 2n almost disjoint closed cubes of side one-half and repeat the procedure. Iterating
this process arbitrarily many times, we obtain a countable family of almost disjoint closed cubes.
The union of the cubes in this family is contained in A, since we only include cubes that are contained
in A. Conversely, if x ∈ A, then since A is open some sufficiently small cube in the bisection procedure
that contains x is entirely contained in A, and the largest such cube is included in the family. Hence the
union of the family contains A, and is therefore equal to A.

From this theorem and from Theorem 3.5.4, we now obtain the following fact.

Corollary 3.5.10. Any open set in Rn is σ-measurable.

Moreover, it is clear now that closed sets (as complements of open sets), countable and finite unions
and intersections of open and closed sets in Rn are σ-measurable.
Recall that for R Theorem 3.5.9 has a more improved version, see Theorem 1.4.17.

3.6 Borel σ-algebra

Definition 3.6.1. The Borel σ-algebra B(Rn ) on Rn is the σ-algebra generated by the open sets, that is,
\
B(Rn ) := σ(T (Rn )) = {A ⊂ P(Rn ) : A ⊃ T (Rn ) and A is a σ-algebra} ,

where T (Rn ) is the set of all open sets in Rn .

A set that belongs to the Borel σ-algebra is called a Borel set.

Thus, the Borel σ-algebra is the minimal σ-algebra on Rn that contains all opens sets.
Since σ-algebras are closed under complementation, the Borel σ-algebra is also generated by the closed
sets in Rn . Moreover, since Rn is σ-compact (i.e. it is a countable union of compact sets) its Borel
σ-algebra is generated by the compact sets.
So, all the finite and countable union and intersections of opens and closed sets on Rn (taken in arbitrary
order) are Borel sets.

Proposition 3.6.2. The Borel algebra B(Rn ) is generated by the collection of closed bricks K(Rn ):
\
B(Rn ) = σ(K(Rn )) := {A ⊂ P(Rn ) : A ⊃ K(Rn ) and A is a σ-algebra} .

Every Borel set is Lebesgue measurable.

3.7. BOREL REGULARITY 43

Proof. Since K(Rn ) is a subset of the set of closed sets, we have σ(K(Rn )) ⊂ B(Rn ). Conversely, by
Theorem 3.5.9, σ(K(Rn )) ⊃ T (Rn ) , so σ(K(Rn )) ⊃ σ(T (Rn )) = B(Rn ), and therefore B(Rn ) = σ(K(Rn )).
From the property 2) of Lebesgue measures and from Theorem 3.5.2, we obtain K(Rn ) ⊂ Mσ (Rn ). Since
Mσ (Rn ) is a σ-algebra, it follows that σ(K(Rn )) ⊂ Mσ (Rn ), so B(Rn ) ⊂ Mσ (Rn ).

Note that if
∞
[
A= Kj
j=1

is a decomposition of an open set A into a union of almost disjoint closed bricks, then
∞
A ⊃ ⊍ Kj◦
j=1

is a disjoint union, and therefore

∞
X ∞
X
µ(Kj◦ ) 6 µ(A) 6 µ(Kj )
j=1 j=1

Since µ(Kk◦ ) = µ(Kj ), it follows that

∞
X
µ(A) = µ(Kj ) (3.6.1)
j=1

for any such decomposition and that the sum is independent of the way in which A is decomposed into
almost disjoint rectangles.
The Borel σ-algebra B(Rn ) is not complete and is strictly smaller than the Lebesgue σ-algebra Mσ (Rn ).
In fact, one can show that the cardinality of B(Rn ) is equal to the cardinality c of the real numbers, whereas
the cardinality of Mσ (Rn ) is equal to 2c . For example, the Cantor set is a set of measure zero with the
same cardinality as R and every subset of the Cantor set is Lebesgue measurable (see Homework). We can
obtain examples of sets that are Lebesgue measurable but not Borel measurable by considering subsets of
sets of measure zero.
Examples of Lebesgue measurable sets that are not Borel sets may also arise from the theory of product
measures in Rn for n > 2. For example, let N = E × {0} ⊂ R2 where E ⊂ R is a non-Lebesgue measurable
set in R. Then N is a subset of the x-axis, which has two-dimensional Lebesgue measure zero, so N belongs
to Mσ (R2 ) since Lebesgue measure is complete. One can show, however, that if a set belongs to B(R2 )
then every section with fixed x or y coordinate belongs to B(R); thus, N cannot belong to B(R2 ) since the
y = 0 section E is not Borel (because it is not Lebesgue measurable).

3.7 Borel regularity

Regularity properties of measures refer to the possibility of approximating in measure one class of sets (for
example, non-measurable sets) by another class of sets (for example, measurable sets). Lebesgue measure is
Borel regular in the sense that Lebesgue measurable sets can be approximated in measure from the outside
by open sets and from the inside by closed sets, and they can be approximated by Borel sets up to sets of
measure zero. Moreover, there is a simple criterion for Lebesgue measurability in terms of open and closed
sets. The following theorem expresses a fundamental approximation property of Lebesgue measurable sets
by open and compact sets. Equations (3.7.1) and (3.7.2) are called outer and inner regularity, respectively.

Theorem 3.7.1. If A ⊂ Rn , then

µ∗ (A) = inf{µ(G) : A ⊂ G, G open} (3.7.1)

and if A ∈ Mσ (Rn ), then

µ(A) = sup{µ(F ) : F ⊂ A, F compact} (3.7.2)
44 CHAPTER 3. THE LEBESGUE MEASURE ON RN

Proof. First, we prove (3.7.1). If µ∗ (A) = +∞, we are done. Suppose now that µ∗ (A) is finite. If A ⊂ G,
then µ∗ (A) 6 µ∗ (G) by the property 2) of the outer measure, so

µ∗ (A) 6 inf{µ(G) : A ⊂ G, G open}

It is left to prove the reverse inequality

µ∗ (A) > inf{µ(G) : A ⊂ G, G open} (3.7.3)

Let ε > 0. By Definition 3.3.1 of the outer measure, there exists a cover of A by bricks Ki , i ∈ N, such
that
∞
X ε
µ(Ki ) 6 µ∗ (A) +
i=1
2

e i be an open brick such that Ki ⊂ K

Let K e i and
ε
µ(K
fi ) 6 µ(Ki ) + . (3.7.4)
2i+1
e i , i ∈ N, covers A and
Then the collection of open bricks K
∞
[
G := K
ei
i=1

is an open set that contains A. Moreover, by (3.7.4) and by σ-semi-additivity of the Lebesgue measure,
one has
∞ ∞
X X ε
µ(G) 6 µ(K
ei) 6 µ(Ki ) + ,
i=1 i=1
2
and therefore
µ(G) 6 µ∗ (A) + ε, (3.7.5)
which proves (3.7.3) since ε > 0 is arbitrary.
Next, we prove (3.7.2). If F ⊂ A, then by property 2) of the outer measure µ(F ) 6 µ(A), so

sup{µ(F ) : F ⊂ A, F compact} 6 µ(A).

Therefore, we just need to prove the reverse inequality

µ(A) 6 sup{µ(F ) : F ⊂ A, F compact}. (3.7.6)

To do this we apply the previous result to the complement Ac and use the measurability of A.
First, suppose that A is a bounded measurable set, so µ(A) < +∞ since there exists a brick such that
A ⊂ K, and µ(A) 6 µ(K) < +∞. Let H ⊂ Rn be a compact set that contains A. By the preceding result,
for any ε > 0, there is an open set G ⊃ H \ A such that

µ(G) 6 µ(H \ A) + ε (3.7.7)

G and H = A ⊍(H \ A), so

S
Then F = H \ G is a compact set such that F ⊂ A. Moreover, H ⊂ F

µ(H) 6 µ(F ) + µ(G), µ(H) = µ(A) + µ(H \ A). (3.7.8)

It follows from (3.7.7)–(3.7.8) that

µ(A) = µ(H) − µ(H \ A) 6 µ(H) − µ(G) + ε 6 µ(F ) + ε,

which implies (3.7.6) and proves the result for bounded measurable sets.
3.7. BOREL REGULARITY 45

Now suppose that A ∈ Mσ (Rn ) is unbounded, and consider Al := A Ql , l ∈ Rn , where Q

Tb b l is defined
in (3.5.1). By Definition 3.5.1 (see (3.5.2)), we have
µ(A) = lim µ(Al ). (3.7.9)
l→+∞

If µ(A) = +∞, then µ(Al ) → +∞ as l → +∞. Since Al is bounded and measurable, by previous result,
we can find a compact set Fl ⊂ Al ⊂ A such that
µ(Fl ) + 1 > µ(Al ),
so that µ(Fk ) → +∞. Therefore,
sup{µ(F ) : F ⊂ A, F compact} = +∞,
which proves the result in this case.
Finally, suppose that A is unbounded and A ∈ M(Rn ), so µ(A) < +∞. From (3.7.9), for any ε > 0
one can choose l ∈ N such that (Al ⊂ Al+1 ⊂ A for any l)
ε
µ(A) 6 µ(Al ) + .
2
Moreover, since Al is bounded, there is a compact set F ⊂ Al such that
ε
µ(Al ) 6 µ(F ) + .
2
Therefore, for any ε > 0, there exists a compact set F ⊂ A such that
µ(A) 6 µ(F ) + ε,
which gives (3.7.6) and completes the proof.
It follows that we may determine the Lebesgue measure of a measurable set in terms of the Lebesgue
measure of open or compact sets by approximating the set from the outside by open sets or from the inside
by compact sets. The outer approximation in (3.7.1) does not require that A is measurable. Thus, for any
set A ⊂ Rn , given ε > 0, we can find an open set G ⊃ A such that µ(G) − µ∗ (A) < ε. If A is measurable,
we can strengthen this condition to get that µ∗ (G \ A) < ε; in fact, this gives a necessary and sufficient
condition for measurability.
Theorem 3.7.2. A ∈ Mσ (Rn ) if and only if for every ε > 0 there is an open set G ⊃ A such that
µ∗ (G \ A) < ε (3.7.10)
Proof. First we assume that A is σ-measurable and show that it satisfies the condition given in the theorem.
Suppose that µ(A) < +∞ and let ε > 0. From (3.7.5) there is an open set G ⊃ A such that
µ(G) 6 µ∗ (A) + ε. Then, since A is measurable, G \ A is measurable and
 \ 
µ∗ (G \ A) = µ(G \ A) = µ(G) − µ G A = µ(G) − µ (A) = µ(G) − µ∗ (A) < ε,

which proves the result when A has

T bfinite measure.
If µ(A) = +∞, define Al = A Q l ⊂ A with Ql given by (3.5.1), and let ε > 0. Since Al is measurable
b
with finite measure, the argument above shows that for each l ∈ N, there is an open set Gl ⊃ Al such that
ε
µ(Gl \ Al ) < l .
2
∞
S
Then G = Gl is an open set that contains A, and
l=1

∞ ∞ ∞
!
[ X X
∗ ∗
µ (G \ A) = µ (Gl \ A) 6 µ∗ (Gl \ A) 6 µ∗ (Gl \ Al ) < ε.
l=1 l=1 l=1
46 CHAPTER 3. THE LEBESGUE MEASURE ON RN

Conversely, suppose that A ⊂ Rn satisfies the condition of the theorem. Let ε > 0, and choose an open
ε
set G such that µ∗ (G \ A) < .
2
If µ∗ (A) < +∞, then due to semi-additivity of the outer measure
ε
µ(G) = µ∗ (G) = µ∗ (A ⊍(G \ A)) 6 µ∗ (A) + µ∗ (G \ A) < µ∗ (A) + ,
2
∞
S
So µ(G) < +∞. By Theorem 3.5.9, one can represent G = Ki , where Ki are almost disjoint closed
i=1
bricks, and
∞
X
µ(G) = m(Ki ) < +∞.
i=1

Consequently, there exists a number i0 ∈ N such that

∞
X ε
m(Ki ) <
i=i0 +1
2

i0
Ki ∈ E n . Then
S
Let B :=
i=1
 [ 
µ∗ (A M B) = µ∗ (A \ B) (B \ A) 6 µ∗ (A \ B) + µ∗ (B \ A). (3.7.11)

On the other hand, since A ⊂ G, we have

∞
i0
!
[ [
A\B =A\ Ki ⊂ Ki ,
i=1 i=i0 +1

and !
i0
[
B\A= Ki \ A ⊂ G \ A.
i=1

Now from (3.7.11) we obtain

∞
!
∗ ∗ ∗ ∗
[ ε ε
µ (A M B) 6 µ (A \ B) + µ (B \ A) 6 µ Ki + µ∗ (G \ A) < + = ε.
i=i0 +1
2 2

Thus, A ∈ Mσ (Rn ).
Suppose now that µ∗ (A) = +∞. We have to prove that Al = A Q n
Tb
l ∈ M(R ), l ∈ N, where Ql is
b
◦
defined in (3.5.1). Any cube Ql is closed by definition. Let us denote by Ql its interior. Then the set
b b
T b◦
Gl := G Q l is open, and we have
 \   \   \   \  \
Gl \ Al = G Q b ◦l \ A Q bl ⊂ G Q bl \ A Q b l = (G \ A) Q b l ⊂ G \ A.

From semi-additivity of the outer measure we obtain

ε
µ∗ (Gl \ Al ) 6 µ∗ (G \ A) < , l ∈ N.
2
Therefore, Al is a bounded set with finite outer measure for which there exists an open set Gl such that
ε
µ∗ (Gl \ Al ) < . By the previous result, Al ∈ M(Rn ), so A ∈ Mσ (Rn ).
2
The following theorem gives another characterization of Lebesgue measurable sets, as ones that can be
“squeezed” between open and closed sets.
3.7. BOREL REGULARITY 47

Theorem 3.7.3. A ∈ Mσ (Rn ) if, and only if, for every ε > 0 there is an open set G and a closed set F
such that F ⊂ A ⊂ G, and
µ(G \ F ) < ε (3.7.12)
If µ(A) < +∞, then F may be chosen to be compact.
Proof. If for a given A ⊂ Rn , for any ε > 0 there exist an open set G ⊃ A and a closed set F ⊂ A such
that (3.7.12) satisfied, then by monotonicity of the outer measure we have µ∗ (G \ A) 6 µ∗ (G \ F ) < ε, so
A ∈ Mσ (Rn ) according to Theorem 3.7.2.
Conversely, let A ∈ Mσ (Rn ). Then Ac ∈ Mσ (Rn ), and by Theorem 3.7.2 given ε > 0, there exist open
sets G ⊃ A and H ⊃ Ac such that
ε ε
µ∗ (G \ A) < , µ∗ (H \ Ac ) < .
2 2
Then, defining the closed set F := H c , we have G ⊃ A ⊃ F and

µ(G \ F ) 6 µ∗ (G \ A) + µ∗ (A \ F ) = µ∗ (G \ A) + µ∗ (H \ Ac ) < ε.

Finally, suppose that µ(A) < +∞ and let ε > 0. According to Theorem 3.7.1, there exists a compact
ε
set F ⊂ A such that µ(A) < µ(F ) + , and
2
ε
µ(A \ F ) = µ(A) − µ(F ) < .
2
As before, from Theorem 3.7.2, there is an open set G ⊃ A such that
ε
µ(G) < µ(A) + .
2
It follows that G ⊃ A ⊃ F , and

µ(G \ F ) = µ(G \ A) + µ(A \ F ) < ε,

which shows that we may take F compact when A ∈ M(Rn ).

From the previous results, we can approximate measurable sets by open or closed sets, up to sets
of arbitrarily small but, in general, nonzero measure. By taking countable intersections of open sets or
countable unions of closed sets, we can approximate measurable sets by Borel sets, up to sets of measure
zero.
Definition 3.7.4. The collection of sets in Rn that are countable intersections of open sets is denoted by
Gδ (Rn ), and the collection of sets in Rn that are countable unions of closed sets is denoted by Fσ (Rn ).
Gδ (Rn ) and Fσ (Rn ) sets are Borel. Thus, it follows from the next result that every Lebesgue measurable
set can be approximated up to a set of measure zero by a Borel set. This is the Borel regularity of Lebesgue
measure.
Theorem 3.7.5. Suppose that A ∈ Mσ (Rn ). Then there exist sets G ∈ Gδ (Rn ) and F ∈ Fσ (Rn ) such
that
G ⊃ A ⊃ F, µ(G \ A) = µ(A \ F ) = 0.
Proof. For each k ∈ N, choose an open set Gk and a closed set Fk such that Gk ⊃ A ⊃ Fk and
1
µ(Gk \ Fk ) < .
k
Then
∞
\ ∞
[
G := Gk , F := Fk
k=1 k=1
are Gδ and Fσ sets with the required properties.
48 CHAPTER 3. THE LEBESGUE MEASURE ON RN

As a corollary of this result, we get that the Lebesgue σ-algebra is the completion of the Borel σ-algebra
w.r.t. Lebesgue measure in the sense that Mσ (Rn ) is the space of σ-Lebesgue measurable sets that differs
from B(Rn ) only by subsets of Borel null sets, and Mσ (Rn ) is complete w.r.t. Lebesgue measure.

Theorem 3.7.6. The Lebesgue σ-algebra Mσ (Rn ) is the completion of the Borel σ-algebra Bσ (Rn ).

Proof. Lebesgue measure is complete from Theorem 3.4.13. By the previous theorem, if A ∈ Mσ (Rn ),
then there is a Fσ set F ⊂ A such that M = A \ F has Lebesgue measure zero. S It follows by the same
theorem that there is a Borel set N ∈ Gδ with µ(N ) = 0 and M ⊂ N . Thus, A = F M where F ∈ B(Rn )
and M ⊂ N ∈ B(Rn ) with µ(N ) = 0, which proves that Mσ (Rn ) is the completion of B(Rn ).

3.8 Invariance properties of Lebesgue measure and non-Lebesgue

measurable sets
A crucial property of Lebesgue measure in Rn is its translation-invariance, which can be stated as follows:
if A is a measurable set and h ∈ Rn , then the set A + h = {x + h : x ∈ A} is also measurable, and
µ(A + h) = µ(A). The invariance of outer measure µ∗ is an immediate consequence of Definition 3.3.1,
since {Ki + h : i ∈ N} is a cover of A + h if and only if {Ki : i ∈ N} is a cover of A, and µ(K + h) = µ(K)
for every brick K. Clearly, µ(B + h) = µB for any elementary set B. To prove the measurability of
A + h under the assumption that A is measurable, we note that for any ε > 0, one can find an elementary
set B such that µ∗ (A M B) < ε. And since µ∗ is translation-invariant (as we proved above), one has
µ∗ ((A + h) M (B + h)) < ε.
In the same way one can prove the relative dilation-invariance of Lebesgue measure. Suppose δ > 0,
and denote by δA the set {δx : x ∈ A}. We can then assert that δA is measurable whenever A is, and
µ(δA) = δ n µ(A). One can also easily see that Lebesgue measure is reflection-invariant. That is, whenever
A is measurable, so is −A = {−x : x ∈ A} and µ(−A) = µ(A).
It also can be shown that the Lebesgue measure is rotation-invariant, see [4, Section 2.8].

Above we mentioned non-Lebesgue measurable sets. But do such sets exists? The following example
answer this question affirmative.

Example 3.8.1 (Non-Lebesgue measurable set). Let us consider the interval [0, 2π) as the unit circum-
ference Γ = {z ∈ C : |z| = 1} on the complex plane (that is, there is one-to-one correspondece between Γ
and [0, 2π)). Then every point ϕ of the interval [0, 2π) becomes the complex number z = eiϕ .
Let α be an irrational number. Define an equivalence relation ∼ on Γ by z2 ∼ z1 if z2 = z1 eπkαi , k ∈ Z.
It is easy to see that the relation ∼ is reflexive (z ∼ z), symmetric (z2 ∼ z1 =⇒ z1 ∼ z2 ) and transitive
(z2 ∼ z1 , z3 ∼ z2 =⇒ z3 ∼ z1 ), so circumference Γ is split by this equivalence relation into equivalence
classes (each class is evidently a countable set). Let the set Φ0 contain exactly one element from each
equivalence class, and let
Φm = {zeπmαi : z ∈ Φ0 }, m ∈ Z.

We now show that

Γ = ⊍ Φm . (3.8.1)
m∈Z

Indeed, Φm Φl = ∅, m 6= l, since if, on the contrary, z ∈ Φm and z ∈ Φl , then z = z0 eπαmi and

T
z = z00 eπαli , where z0 , z00 ∈ Φ0 . So we obtain z0 eπαmi = z00 eπαli , or z00 = zeπα(m−l)i , that is, z00 ∼ z0 . But
Φ0 contains exactly one element from each equivalence class, so there are no equivalent elements in Φ0 .
Therefore, z00 = z0 , so eπαli = eπαmi , or eπα(m−l)i = 1, or πα(m T − l)i = 2πki, k ∈ Z, or α = m−l ∈ Q, a
2k

contradiction, since α is irrational by assumption. Thus, Φm Φl = ∅ whenever m 6= l.

On the other hands, every point z ∈ Γ belongs to one of the equivalence class by construction. Therefore,
z ∼ z0 ∈ Φ0 , so z = z0 eπαmi , that is, z ∈ Φm . Thus, the identity (3.8.1) is proved.
3.9. OTHER EXAMPLES OF MEASURES 49

Suppose now that the set Φ0 is Lebesgue measurable. Then each set Φm is also measurable, and
µΦm = µΦ0 , m ∈ Z, since Φm is obtained from Φ0 by a rotation on Γ (a shift on R), but the Lebesgue
measure is invariant with respect to rotations and shifts. The Lebesgue measure is σ-additive, so we have
+∞
X
µΓ = 2π = µΦm .
m=−∞

However, this equality is impossible, since the sum of the series here is either equals zero (if µΦ0 = 0), or
equals infinity (if µΦ0 = a > 0).
Thus, the set Φ0 is non-Lebesgue measurable.

3.9 Other examples of measures

3.9.1 Lebesgue–Stieltjes measure
Here we shortly describe a generalization of one-dimensional Lebesgue measure, called Lebesgue-Stieltjes
measure on R. This measure is obtained from a non-decreasing and left continuous function F : R 7→ R
by assigning to intervals the following measure

m0F (a, b) = F (b) − F (a + 0),

m0F [a, b] = F (b + 0) − F (a),

m0F (a, b] = F (b + 0) − F (a + 0),

m0F [a, b) = F (b) − F (a).

It is easy to see that the defined measure m0 is nonnegative and additive (in fact, m0 is not a measure
by Definition 2.2.1, since the system of intervals on R is not a ring, see Remark 3.1.1). Note that the
system of bricks on R is the system of intervals. As well, the measure m0 defined on bricks (see (3.1.1)) is
a particular case of m0F as n = 1 (in this case, F (t) = t, t ∈ R).
Applying the extension process described in Sections 3.2–3.5 to the function m0F , we define on R the
ring MF of sets measurable w.r.t. σ-additive measure µF (extension of m0F ) which is called the Lebesgue–
Stieltjes measure. The ring MF depends on F but it always contains all Borel sets (see [4, Section 2.9], [3,
Section 1.5] for more details). If the function F has a finite variation on R, that is, if F (+∞) − F (−∞) <
+∞, then, obviously, MF is a σ-algebra with the unit R, and µF (R) = F (+∞) − F (−∞). Moreover,
if F (+∞) − F (−∞) = 1, then the measure µF is called normed but if additionally F (−∞) = 0 and
F (+∞) = 1, then µF is called a probability measure.
We emphasize once again that if F (t) = t, then µF = µ is the ordinary Lebesgue measure.

3.9.2 Discrete measure

Let X = {x1 , x2 , . . . , xn , . . .} be an arbitrary countable set, and let (pn )∞
n=1 be a sequence of positive
numbers satisfying the condition
X∞
pn < +∞.
n=1

For any subset A of the set X we define

X
mA = pk . (3.9.1)
k: xk ∈A

The formula (3.9.1) defines a σ-additive measure m on the σ-algebra P(X) of all subsets of the set X.
50 CHAPTER 3. THE LEBESGUE MEASURE ON RN

If X = {x1 , x2 , . . . , xn , . . .} ⊂ R, then we connect this set with the function F : R 7→ R+ which defined
as follows X
F (x) = pk (3.9.2)
k: xk <x

This function is called a saltus function, because

F (xk + 0) − F (xk ) = pk ,

and on the intervals free of points xk , the function F is constant.

By the function (3.9.2), one can defined a Lebesgue-Stieltjes measure µF . Such measure is called
discrete measure.
3.10. PROBLEMS 51

3.10 Problems
Problem 3.1. Let {Ak }+∞ n
k=1 be decreasing sequence of sets from Mσ (R ), and µAk = +∞ (k ∈ N). Can
∞
\
the set A = Ak have
k=1

a) infinite measure;
b) finite positive measure;
c) measure 0?
Problem 3.2. Given T two Lebesgue measurable sets A1 and A2 on the interval [0, 1] such that µA1 +µA2 >
1. Prove that µ (A1 A2 ) > 0.
Pn
Problem 3.3. Given n Lebesgue measurable sets A1 , A2 , . . . , An on the interval [0, 1] such that µAk >
 n  k=1
T
n − 1. Prove that µ Ak > 0.
k=1

Problem 3.4. Given n, construct Lebesgue measurable sets A1 , A2 , . . . , An on the interval [0, 1] such
Pn
that µAk = n − 1 and
k=1 !
\n
µ Ak = 0.
k=1
n
Definition. A set A ∈ R is called measurable in Jordan sense (or simply Jordan measurable) if for any
ε > 0 there exist elementary sets B1 , B2 ∈ E n such that B1 ⊂ A ⊂ B2 , and

m(B2 \ B1 ) < ε.

In this case, the (Jordan) measure µJ of the set A is defined as follows:

µJ = inf{mB : A ⊂ B, B ∈ E n }.

Problem 3.5. Prove that if A ∈ Rn is Jordan measurable, then it is Lebesgue measurable, and its
Lebesgue measure equals its Jordan measure.
Hint: Prove first the semi-additivity of the Jordan measure.
T
Problem 3.6. Prove that the set A := Q [0, 1] is Lebesgue measurable but non-Jordan measurable.
Find the Lebesgue measure of A.
Problem 3.7. Find the Lebesgue measure of the Cantor set and prove that any subset of the Cantor set
is Lebesgue measurable.
Problem 3.8. Find the Lebesgue measure of the subset of the interval [0, 1] consisting of the numbers
(of [0, 1]) whose decimal form does not contain a digit n, n = 0, 1, 2 . . . , 9.
Problem 3.9. Find the Lebesgue measure of a subset of the interval [0, 1] consisting of the numbers
(of [0, 1]) in whose decimal form the digit 2 always stays earlier than the digit 3.
Problem 3.10 (The Borel-Cantelli lemma). Let {Ak }+∞
k=1 be a sequence of Lebesgue measurable sets such
that
+∞
X
µAk < +∞.
k=1

Prove that A = limAk is Lebesgue measurable, and µA = 0.

52 CHAPTER 3. THE LEBESGUE MEASURE ON RN

Problem 3.11. Find the Lebesgue measure of the subset of the plane square {(x, y) : 0 6 x 6 1, 0 6
1
x 6 1} consisting of points (x, y) such that 0 6 sin x 6 and cos(x + y) is irrational.
2
Problem 3.12. Find the Lebesgue measure of the subset of the plane square {(x, y) : 0 6 x 6 1, 0 6
x 6 1} consisting of points (x, y) whose Descartes and polar coordinates are irrational.
Problem 3.13. Prove that any set of R with positive Lebesgue measure is of cardinality continuum.
Problem 3.14. Prove that the cardinality of Mσ (Rn ) is 2c .
T
Problem 3.15. Let a set A ⊂ [0, 1] is Lebesgue measurable. Prove that the function f (x) = µ (A [0, x])
is continuous on [0, 1].
∞
 
1 1 α
Problem 3.16. Prove that the set B = ⊍ , + ⊂ [0, 1] is Lebesgue measurable and
n=2 n n n(n − 1)
T
µ (B [0, x])
lim = α,
x→+0 x
where α ∈ (0, 1).
Problem 3.17. Let A ⊂ [0, 1] be a measurable set w.r.t. the Lebesgue measure µ on [0, 1], and let
µ(A ∩ (a, b)) 6 α(b − a), 0 < α < 1, for any interval (a, b) ⊂ [0, 1]. Prove that µA = 0.
Problem 3.18. Let A ⊂ [0, 1] and µ(A) > 0. Prove that there exists a pair of points x, y ∈ A such that
the distance |x − y| is irrational. Here µ is the Lebesgue measure.
Problem 3.19. Let A ⊂ [0, 1] and µ(A) > 0. Prove that there exist a pair of points x, y ∈ A such that
the distance |x − y| is rational. Here µ is the Lebesgue measure.
Problem 3.20. Let a set A(∈ T R) be non-Lebesgue measurable, and a set A0 (∈ R) be of Lebesgue
measure 0. Prove that the set A (Ac0 ) is non-Lebesgue measurable. Is the statement true for the case
when A, A0 ∈ Rn ?
Problem 3.21. Construct a Lebesgue measurable set A ⊂ [0, 1] × [0, 1] such that both its projections on
the coordinate axes Ox and Oy are non-Lebesgue measurable.
Problem 3.22. Let µ be the Lebesgue measure on Rn , A ∈ M, and µ(A) > 0. Prove that there exists a
Lebesgue non-measurable set B ⊂ A.
Problem 3.23. Construct a σ-additive measure on Rn such that any subset of Rn is measurable.
Problem 3.24. Construct a set A ⊂ [0, 1] × [0, 1] measurable w.r.t. the Lebesgue measure on [0, 1] × [0, 1]
whose projections on the axes OX and OY are not measurable w.r.t. the Lebesgue measure on [0, 1].
Problem 3.25. Construct a set A ⊂ [0, 1] × [0, 1] which is not measurable w.r.t. the Lebesgue measure on
[0, 1] × [0, 1] and whose projections on the axes OX and OY are measurable w.r.t. the Lebesgue measure
on [0, 1].
Problem 3.26. Find the (two-dimensional) Lebesgue measure of the set
 
1
A = (x, y) ∈ [0, 1] × [0, 1] : x ∈ [0, 1] \ Q[0,1] and sin y < ,
2

where Q[0,1] is the set of all rational numbers of the interval [0, 1].
Problem 3.27. Prove that the outer measure is not additive.
Hint: Use Lebesgue non-measurable sets.
3.10. PROBLEMS 53

Problem 3.28. Prove that Q ∈ Fσ (R) and Q 6∈ Gδ (R).

Hint: Using the fact that a set A is dense in R if it has a common point with any interval on R, prove that
a countable intersection of sets in Gδ (R) dense in R is a set in Gδ (R) which is also dense in R. Supposing
that Q belongs to Gδ (R) find a sequence of Gδ -sets dense in R whose intersection is not dense in R that
will give a contradiction.

Problem 3.29. Let

A = {(x, y) : a 6 x 6 b and 0 6 y 6 f (x)},
where f (x) is a positive continuous function on [a, b]. Prove that A is measurable w.r.t. the Lebesgue
measure on R2 and its Lebesgue measure is

Zb
µA = f (x)dx,
a

where the integral is the Riemann integral.

54 CHAPTER 3. THE LEBESGUE MEASURE ON RN
Chapter 4

Measurable functions

In this chapter, we introduce measurable functions and study their main properties as a first step of the
construction of the Lebesgue integral.

4.1 Definition of measurable functions

Let X be a set, M = M(X) be a σ-algebra of subsets of the set X, which is the unit of this σ-algebra,
and let µ be a complete σ-additive measure defined on M. In what follows, we call the three (X, M, µ)
a measure space (or a µ-measure space).

Definition 4.1.1. A function f : X 7→ R is called measurable on the set X w.r.t. the measure µ (or
µ-measurable) if for any real number c the set X(f > c) := {x ∈ X : f (x) > c} is µ-measurable.

We denote the set of all µ-measurable on X functions as S(X, M, µ). If it is clear from the context
what measure is used, then we omit the mention of the measure and denote the set of all measurable
functions as S(X).

Example 4.1.2. Let µX = 0. Then any function f : X 7→ R is measurable on X, since from the
completeness of the considered measure, we obtain for any c ∈ R:

X(f > c) ∈ M(X).

Lemma 4.1.3. The function f ∈ S(X) if, and only if, one of the following conditions holds:

X(f > c) ∈ M(X), (4.1.1)

X(f < c) ∈ M(X), (4.1.2)

X(f 6 c) ∈ M(X). (4.1.3)

Proof. Let S(x), and let c ∈ R. First we show that

∞  
\ 1
X(f > c) = X f >c− . (4.1.4)
k
k=1

Indeed,
1
x ∈ X(f > c) ⇐⇒ f (x) > c ⇐⇒ ∀k ∈ N f (x) > c − ⇐⇒
k
∞
   
1 T 1
⇐⇒ ∀k ∈ N x ∈ X f (x) > c − ⇐⇒ x ∈ X f (x) > c − .
k k=1 k

55
56 CHAPTER 4. MEASURABLE FUNCTIONS

If f ∈ S(X), then any set in the right-hand side of (4.1.4) is measurable. Therefore, X(f > c) is also
measurable as a countable intersection of measurable sets, since M(x) is a σ-algebra by assumption. Thus,
we proved that if f ∈ S(X), then X(f > c) ∈ M(X).
Conversely, let the set X(f > c) be measurable. Then from the identity (which can be easily checked)
∞  
[ 1
X(f > c) = X f >c+ .
k
k=1

it follows that f ∈ S(X).

The equivalence of the condition (4.1.2) to the inclusion f ∈ S(X) follows from the following obvious
identities:
X(f 6 c) = X \ X(f > c), X(f > c) = X \ X(f 6 c).
The equivalence of the condition (4.1.3) to the inclusion f ∈ S(X) follows from the obvious identities:
X(f < c) = X \ X(f > c), X(f > c) = X \ X(f < c).

Lemma 4.1.3 claims that anyone of the conditions (4.1.1)–(4.1.3) can be used in the definition of
measurable functions.
Corollary 4.1.4. If f ∈ S(X), then for any a, b ∈ R, a 6 b, the following sets are measurable:
X(a 6 f 6 b), X(a < f 6 b), X(a 6 f < b), X(a < f < b), X(f = a).
Proof. Indeed, \
X(a 6 f 6 b) = X(f 6 b) X(f > a) ∈ M(X).
The rest assertions of the theorem can be proved analogously.
m
Example 4.1.5. Let X = ⊍ Xk . Define the function h : X 7→ R putting h(x) = ck whenever x ∈ Xk ,
k=1
k = 1, . . . , m, where ck ∈ R are arbitrary real numbers (that may coincide for different indices). Such
function is called a simple function.
If all the sets Xk are measurable, then for any c ∈ R, the set
X(h > c) = ⊍ Xk
k: ck >c

is measurable as a union of finitely many measurable sets (if c > ck for all k = 1, . . . , m, then the set
X(h > c) = ∅ is measurable), therefore, the function h is measurable.
If all the numbers ck are distinct, then from h ∈ S(x) it follows that Xk ∈ M(X) for all k = 1, . . . , m.
Theorem 4.1.6. Let X be R or an interval in R. Any continuous function f : X 7→ R is σ-Lebesgue
measurable (that is, measurable w.r.t. the Lebesgue measure).
Proof. Let X0 be the set of all interior points of the set X (if X is an interval, then X0 is the same interval
but without the end points). For an arbitrary c ∈ R, consider the set X0 (f > c). If this set is empty, then
f (x0 )−c
it is measurable. Suppose now that X0 (f > c) 6= ∅. Let T x0 ∈ X0 (f > c), and let ε := 2 . Since f is
continuous, there exists δ > 0 such that for all x ∈ X0 (x0 − δ, x0 + δ), we have |f (x) − f (x0 )| < ε. So,
f (x0 ) − c c + f (x0 )
f (x) > f (x0 ) − ε = f (x0 ) − = > c.
2 2
Clearly, we can choose δ > 0 so small that (x0 − δ, x0 + δ) ⊂ X0 , and (x0 − δ, x0 + δ) ⊂ X0 (f > c)
as we proved above. Therefore, the set X0 (f > c) is open, so X0 (f > c) is σ-Lebesgue measurable by
Corollary 3.5.10. The set X(f > c) differs from X0 (f > c) by one or two points (the ends of the interval).
But one-point sets (degenerated bricks) are σ-Lebesgue measurable, so the set X(f > c) is σ-Lebesgue
measurable, thus f ∈ S(X).
4.2. PROPERTIES OF MEASURABLE FUNCTIONS 57

4.2 Properties of measurable functions

Let us now study some properties of measurable functions.

1) f ∈ S(X), l ∈ R =⇒ f + l ∈ S(X).
Obviously, X(f + l > c) = X(f > c − l) ∈ M(X) for any real c, so f + l ∈ S(X).

2) f ∈ S(X), k ∈ R =⇒ kf ∈ S(X).
This property follows from the identity


 X(f > c/k), k > 0,

 X(f < c/k), k < 0,
X(kf > c) =


 X, k = 0, c < 0,
k = 0, c > 0,

∅,

since all the sets in the right-hand side of this identity are measurable.

To establish the next property, we need the following lemma.

Lemma 4.2.1. If f, g ∈ S(X), then X(f > g) ∈ M(X).

Proof. Let us enumerate all the rational numbers Q = {rk : k ∈ N}. This is possible, since Card Q = a.
Let x ∈ X(f > g), that is, fT(x) > g(x). Then there exists a rational number rk such that f (x) > rk >
g(x), therefore, x ∈ X(f > rk ) X(g < rk ), so
∞ 
[ \ 
X(f > g) ⊂ X(f > rk ) X(g < rk ) .
k=1

The converse inclusion is obvious, so we have

∞ 
[ \ 
X(f > g) = X(f > rk ) X(g < rk ) .
k=1

Since M(X) is a σ-algebra by assumption, the right-hand side of the last identity is measurable, so is the
set X(f > g).

Now we are in a position to establish the next properties of measurable functions.

3) f, g ∈ S(X) =⇒ f + g ∈ S(X).
Clearly,
X(f + g > c) = X(f > −g + c),
so f + g is measurable by Lemma 4.2.1 and by the properties 1) and 2) of measurable functions.

4) f ∈ S(X) =⇒ f 2 ∈ S(X).
Indeed, the set ( √ S √
2 X(f > c) X(f < − c), c > 0,
X(f > c) =
X, c < 0.
is measurable for any real c.

Note that converse is not true. That is, if f 2 is measurable, then this does not mean that the function f
is measurable, generally speaking.
58 CHAPTER 4. MEASURABLE FUNCTIONS

Example 4.2.2. Let X = [0, 1], and X0 be a non-Lebesgue measurable subset of X. Suppose that
(
1, x ∈ X0 ,
f (x) =
−1, x ∈ X \ X0 .

The function f is non-Lebesgue measurable, since the set X(f = 1) (which is exactly X0 ) is non-Lebesgue
measurable. At the same time, the function f 2 ≡ 1 is Lebesgue measurable on X.
5) f, g ∈ S(X) =⇒ f · g ∈ S(X).
This property follows from the identity
(f + g)2 − f 2 − g 2
f ·g =
2
and from the properties 2), 3), and 4).
f
6) f, g ∈ S(X), g(x) 6= 0 for x ∈ X =⇒ ∈ S(X).
g
1
It suffices to prove that the function is measurable and then to apply the property 5). So, from
g
the identity 
   X(0 < g < 1/c), c > 0,
1 
X >c = X(g > 0), c = 0,
g  S
X(g > 0) X(g < 1/c), c < 0,


1
we have ∈ S(X).
g
Definition 4.2.3. The function (
+ f (x), f (x) > 0,
f (x) =
0, f (x) < 0,
is called the positive part of f , and the function
(
0, f (x) > 0,
f − (x) =
−f (x), f (x) 6 0,

is called the negative part of f .

7) f ∈ S(X) =⇒ |f |, f + , f − ∈ S(X).
This property follows from the identities
( S
X(f < −c) X(f > c), c > 0,
X(|f | > c) =
X, c < 0,

and
|f | + f |f | − f
f+ = , f− = .
2 2
Example 4.2.2 shows that |f | ∈ S(X) =
6 ⇒ f ∈ S(X).
8) f ∈ S(X), X0 ∈ M(X) =⇒ f ∈ S(X0 ).
The proof of this property Tis based on the fact that if A is a σ-algebra with the unit X, and if
X0 ∈ A, then A(X0 ) := {A X0 : A ∈ A(X)} is a σ-algebra with the unit X0 (see Homework no.4).
Now for any c ∈ R, one has
\
X0 (f > c) = X0 X(f > c) ∈ M(X0 ).
4.3. SEQUENCES OF MEASURABLE FUNCTIONS 59

ω
Xk , where Xk ∈ M(X), and1 1 6 ω 6 +∞, and if f ∈ S(Xk ), 1 6 k 6 ω,
S
9) If f : X 7→ R, X =
k=1
then f ∈ S(X).
This property follows from the identity
[
X(f > c) = Xk (f > c),
k

which can be easily checked.

10) Let f ∈ S(X) and µX(f = ±∞) = 0, then for any ε > 0 there exists g ∈ S(X) such that g is bounded
on X and µX(f 6= g) < ε.
Consider the sets

Ak = X(|f | > k), A = X(f = ±∞) = X(|f | = +∞).

∞
T
By assumption, µA = 0. Moreover, A = Ak , and
k=1

A1 ⊃ A2 ⊃ A3 ⊃ . . .

According to Lemma 2.1.21, lim Ak = A, and due to σ-additivity of the measure µ, we obtain
k→∞
 
µA = µ lim Ak = lim µAk = 0
k→∞ k→∞

by Theorem 2.2.5. Therefore, for any ε > 0 there exists a number N ∈ N such that µAN < ε for
any k > N . Introduce the following function on X
(
f (x), x ∈ X \ AN ,
g(x) =
0, x ∈ AN .

It is clear that g ∈ S(x). Moreover, it is bounded, since |g(x)| 6 N on X. At the same time,
X(f 6= g) = AN , so µX(f 6= g) < ε.

Definition 4.2.4. Some property is said to hold on the set X “almost everywhere” (a.e.) if the set of the
points of X for which this property does not hold, is a null set (a set of measure zero).
For example, the statement “the function f equals zero almost everywhere on the set X” means that
µX(f 6= 0) = 0.
Thus, the property 10) shows that any measurable function which is almost everywhere finite on X
becomes bounded if we neglect a subset of X of arbitrary small measure.

4.3 Sequences of measurable functions

In what follows, we consider operations on measurable functions related to passage to the limit. Since
during this process we can obtain infinite values, we extend the set of functions under consideration.
Namely, let us consider functions f : X 7→ R = [−∞, +∞]. Since we need to operate with infinite values,
let us define the following:
1) a + (±∞) = ±∞, 2) a − (±∞) = ∓∞, a ∈ R,
3) (+∞) + (+∞) = +∞, 4) −∞ + (−∞) = −∞,
5) a · (±∞) = ±∞, a > 0, 6) a · (±∞) = ∓∞, a < 0, 7) 0 · (±∞) = 0,
8) (±∞) · (±∞) = +∞, 9) (±∞) · (∓∞) = −∞,
1 That is, the union can be finite or countable.
60 CHAPTER 4. MEASURABLE FUNCTIONS

a
10) | ± ∞| = +∞, 11)= 0, a ∈ R.
±∞
±∞ ±∞ a
The symbols (±∞) − (±∞), , , are considered to be meaningless.
±∞ ∓∞ 0
The definition of measurability for the functions f : X 7→ R is left the same as Definition 4.1.1. There-
fore, all the properties proved for measurable functions are valid, provided the corresponding operations
are acceptable.

Theorem 4.3.1. Let (fk )∞ k=1 be a sequence of functions measurable on X. Then the functions sup{fk },
inf{fk }, lim{fk }, lim{fk }, lim fk (if any) are measurable on X.

All the operations mentioned in the theorem are pointwise. For instance, the function f ∗ = sup{fk } is
defined as follows
f ∗ (x) = sup{fk (x) : k ∈ N}, x ∈ X. (4.3.1)

Proof of Theorem 4.3.1. First we prove that f ∗ = sup{fk } is measurable. To do this, let us establish the
following identity
∞
\
X(f ∗ 6 c) = X(fk 6 c), c ∈ R. (4.3.2)
k=1

∗ ∗
Assume that x ∈ X(f 6 c). Then f (x) 6 c, and from (4.3.1) it follows that fk (x) 6 c for any k ∈ N.
∞ ∞
X(fk 6 c), so X(f ∗ 6 c) ⊂
T T
Therefore, x ∈ X(fk 6 c).
k=1 k=1
∞
T
Conversely, if x ∈ X(fk 6 c), then fk (x) 6 c for any k ∈ N. Consequently,
k=1

sup{fk (x) : k ∈ N} = f ∗ (x) 6 c,

∞
that is, x ∈ X(f ∗ 6 c), so X(f ∗ 6 c) ⊃
T
X(fk 6 c). Thus, the identity (4.3.2) is true, and the
k=1
function f ∗ is measurable by Lemma 4.1.3.
The function inf{fk } is measurable, since

inf{fk } = − sup{−fk }.

The function f = limfk is measurable, since

 
f (x) = limfk (x) = inf sup{fi (x)} .
k∈N i>k

From the identity limfk = −lim(−fk ), it follows that the function f = limfk is measurable.
If lim fk exists, it is measurable, since in this case

lim fk = limfk = limfk .

a.e.
Definition 4.3.2. Two functions f and g defined on the set X are called equivalent, or equal a.e., f = g,
if their values coincide a.e. on X, that is, if

µX(f 6= g) = 0.
a.e.
Lemma 4.3.3. If f, g : X 7→ R, f ∈ S(X), and f = g, then g ∈ S(X).
4.3. SEQUENCES OF MEASURABLE FUNCTIONS 61

Proof. Let X0 := X(f 6= g). By assumption, µX0 = 0. Consider the set X1 = X \ X0 . The set X1 is
measurable as a difference of two measurable sets. We have X = X0 ⊍ X1 , and

X(g > c) = X0 (g > c) ⊍ X1 (g > c) = X0 (g > c) ⊍ X1 (f > c).

Since µ is assumed to be a complete measure, the set X0 (g > c) is measurable as a subset of the null
set X0 . At the same time, the set X1 (f > c) is measurable by the property 8) of measurable functions.
Therefore, the set X(g > c) is measurable for any c ∈ R.

Definition 4.3.4. A sequences of functions (fk )∞

k=1 is said to be convergent a.e. on the set X to a
a.e.
function f (we write fk −→ f ) if
µX(fk 6→ f ) = 0.

Example 4.3.5. Let X = [0, 1], µ be the Lebesgue measure, fk (x) = xk , and f (x) ≡ 0.
Since lim xk = 0 for 0 6 x < 1, and since lim 1k = 1, we have X(fk 6→ f ) = {1}. Therefore,
k→+∞ k→+∞
a.e.
fk −→ f , because µ({1}) = 0.
a.e.
Theorem 4.3.6. If (fk )∞
k=1 is a sequence of measurable functions, and if fk −→ f on X, then f ∈ S(X).

Proof. Define X1 := X(fk → f ) and X0 = X(f 6→ f ). By assumption, µX0 = 0, therefore, the set
X1 = X \ X0 is measurable. Let us consider the following functions:
( (
fk (x), x ∈ X1 , f (x), x ∈ X1 ,
gk (x) = g(x) =
0, x ∈ X0 , 0, x ∈ X0 .

a.e. a.e.
Since µX0 = 0, we get fk = gk , k ∈ N, and f = g. At the same time, gk → g on X pointwise.
By Lemma 4.3.3, the functions gk are measurable on X, so g ∈ S(X) as a pointwise limit of measurable
a.e.
functions, see Theorem 4.3.1. Now f ∈ S(X) by Lemma 4.3.3, since f = g.

Theorem 4.3.7 (Egoroff). Let µX < +∞, and let (fk )∞ k=1 be a sequence of measurable functions that are
2 a.e.
a.e. finite on X. If fk −→ f , where f is a.e. finite on X, then for any δ > 0 there exists a measurable
set Xδ ⊂ X such that µXδ < δ, and the sequences (fk )∞ k=1 converges to f uniformly on X \ Xδ .

Proof. By Theorem 4.3.6, the function f is measurable on X. Let us introduce the following sets
∞
!
[ [
X0 = X(fk = ±∞) X(fk 6→ f ), X1 = X \ X0 .
k=0

The set X0 is measurable as a countable union of measurable sets. Moreover, µX0 = 0 due to σ-semi-
additivity of the measure µ, since X0 is a countable union of null sets. The set X1 is measurable as
difference of two measurable sets. Therefore, X1 is measurable, and additionally, f , fk , k ∈ N, are finite
on X1 , and fk → f pointwise on X1 .
Consider the sequence
gk (x) = |fk (x) − f (x)|, k = 1, 2, 3, . . .
The functions gk are nonnegative and measurable on X1 , and gk (x) → 0 for any x ∈ X1 .
Let (εj )∞
j=1 be a sequence of positive numbers such that εj → 0 as j → ∞. Consider the following sets

∞
\
Xk,j = X1 (gl < εj ). (4.3.3)
l=k

2 Not 1
necessary bounded! Consider e.g. the functions fk (x) = that are finite a.e. on X := [0, 1] but not bounded on X.
xk
62 CHAPTER 4. MEASURABLE FUNCTIONS

The sequences (Xk,j ) increases as k grows, since

∞
\ ∞
\
Xk,j = X1 (gl < εj ) ⊂ X1 (gl < εj ) = Xk+1,j .
l=k l=k+1

Let us show that

lim Xk,j = X1 j = 1, 2, 3, . . . (4.3.4)
k→∞

Let x ∈ X1 . Since lim gk = 0, then for εj > 0 there exists a number k0 such that gk (x) < εj for any
k→+∞
k > k0 . This means that
∞
\
x∈ X1 (gl < εj ) = Xk0 ,j ,
l=k0

therefore,
∞
[
x∈ Xk,j = lim Xk,j ,
k→∞
k=1

thus, X1 ⊂ lim Xk,j . The converse inclusion is obvious, so the identities (4.3.4) are true. Since µ is
k→∞
assumed to be σ-additive, it is continuous. Therefore, from (4.3.4), we obtain

lim µXk,j = µX1 , j = 1, 2, 3, . . . ,

k→∞

So, for any j ∈ N there exists a number k = k(j) such that

δ
µ(X1 \ Xk(j),j ) < ,
2j
where δ is an arbitrary positive number (here we use the fact that µX1 = µX < +∞).
The set  
[ [ ∞
Xδ = X0  (X1 \ Xk(j),j )
j=1

is the set defined in the assertion. Indeed,

∞ ∞
X X δ
µXδ 6 µX0 + µ(X1 \ Xk(j),j ) < 0 + j
= δ.
j=1 j=1
2

If x ∈ X \ Xδ , then x 6∈ X0 , and x 6∈ (X1 \ Xk(j),j ) for any j ∈ N. But since x 6∈ X0 , we have x ∈ X1 .

And since x 6∈ (X1 \ Xk(j),j ) for any j ∈ N, we obtain that x ∈ Xk(j),j for all j ∈ N. Thus, from (4.3.3) we
obtain
|fl (x) − f (x)| = gl (x) < εj , l > k(j).
The number k(j) does not depend on x, so fk converges to f uniformly on X \ Xδ .
Example 4.3.8. Consider the sequence fk (x) = sink πx, k ∈ N on the set X = [0, 1] with the Lebesgue
measure µ.
1
It is easy to see that for x ∈ [0, 1], x 6= , sink πx → 0, so
2
a.e.
fk −→ 0.
     
1 1 1 1 S 1
Choose δ such that 0 < δ < , and define the set Xδ = − δ, + δ , so X\Xδ = 0, − δ + δ, 1 .
2 2 2  2 2
π    π 
Obviously, for any x ∈ X\Xδ , the inequality 0 6 sink πx 6 sink π − δ holds, and sink π −δ →
∞
2 2
0 as k → ∞. Thus, the sequence (fk )k=1 converges uniformly to zero on the set X \ Xδ .
4.3. SEQUENCES OF MEASURABLE FUNCTIONS 63

For sequences of measurable functions one can define one more kind of convergence, the convergence
in measure.
Definition 4.3.9. A sequences (fk )∞
k=1 of functions measurable on X is said to be convergent to a mea-
µ
surable on X function f in measure (we write fk −→ f ) if for any σ > 0,
µX (|fk − f | > σ) −→ 0.
k→∞

Let us find interrelation between the convergence almost everywhere and the convergence in measure.
Theorem 4.3.10. Let µX < +∞. If a sequences (fk )∞ k=1 of functions measurable on X converges a.e. to
a function f , then (fk )∞
k=1 converges to f on X in measure:
a.e. µ
fk −→ f =⇒ fk −→ f.
Proof. By Theorem 4.3.6, the function f is measurable. On the contrary, suppose that (fk )∞
k=1 does not
converge to f in measure on the set X. Then for a certain σ0 > 0 there exist δ0 > 0 and a sequence of
indices (kj )∞
j=1 such that
µX(|fkj − f | > σ0 ) > δ0 , j = 1, 2, 3, . . .
Let
X 0 = lim X(|fkj − f | > σ0 ).
j

Then by Theorem 2.2.6, µX > δ0 . So, if x ∈ X 0 , then according to the definition of the limit superior
0

of a sequence of sets, among indices kj , there exist infinitely many indices such that x ∈ X(|fkj − f | > σ0 ).
This, in particular, means that for infinitely many indices we have |fkj (x) − f (x)| > σ0 , that is, fk (x) 6−→
f (x).
a.e.
Thus, fk 6−→ f on the set X 0 of positive measure, a contradiction, since fk −→ f on X by assumption.

x
Note that we cannot avoid the condition µX < +∞ here.Consider, for example, the sequence fn (x) =
n
on R.
However, the converse statement of Theorem 4.3.10 is not true.
Example 4.3.11. Let X = [0, 1], and µ be the Lebesgue measure. Consider the functions
 
j−1 j


 1, x ∈ , ,
l l



ϕl,j (x) =
 
j−1 j


 0, x 6∈

 , ,
l l
where l = 1, 2, . . ., j = 1, 2, . . . , l.
Let us represent the functions ϕl,j as one-index sequence as follows:
f1 (x) = ϕ11 (x), f2 (x) = ϕ21 (x), f3 (x) = ϕ22 (x), f4 (x) = ϕ31 (x), f5 (x) = ϕ32 (x), . . . ,
so
l−1
X
fn (x) = ϕl,j (x), n=j+ k, l = 1, 2, . . . , j = 1, 2, . . . , l.
k=1

The sequence fn converges in measure on [0, 1] to the function f0 (x) ≡ 0. Indeed, every function
fn belongs to a group of functions ϕl,j with fixed index l each of which is nonzero only on the interval
j−1 j 1
, of length . Thus, if we take σ 6 1, we get
l l l
1
µX(|fn − f0 | > σ) = −→ 0.
l n→∞
64 CHAPTER 4. MEASURABLE FUNCTIONS

At the same time, for l > 2, and for any x ∈ [0, 1] in each group of functions ϕl,j with fixed l, there
exist functions that equal 1 at x and functions that equal 0 at x. Consequently, for any x ∈ [0, 1] the
sequence fn (x) consists of infinitely many 1s and infinitely many 0s, so it is not convergent at all.
Thus, the sequence fn converges in measure on [0, 1] to the function f0 (x) ≡ 0, but it does not converge
at any point of the interval [0, 1].
However, the following theorem is true.
Theorem 4.3.12 (F. Riesz). Any sequence (fn )∞ n=1 of functions measurable on X convergent in measure
on X to a measurable function f0 contains a subsequence convergent a.e. on X to f0 .
Proof. Let (εj )∞ ∞
j=1 and (ηj )j=1 be two sequences of positive numbers such that

∞
X
εj −→ 0, ηj < +∞.
j→∞
j=1

Consider the sequence of indices

n1 < n2 < . . . < nj < . . .
constructed as follows: The index n1 is chosen such that
µX(|fn1 − f0 | > ε1 ) < η1 .
Next we find an index n2 > n1 such that
µX(|fn2 − f0 | > ε2 ) < η2 .
If we already have the indices n1 < n2 < . . . < nj−1 , then we take an index nj such that
µX(|fnj − f0 | > εj ) < ηj .
µ
For any j we can always find an index nj described above, because fn −→ f0 by assumption. Since the
process described can be continued indefinitely, as a result we obtain a subsequence (fnj )∞
j=1 . Let us show
a.e.
that fnj −→ f0
For brevity sake, we denote Xj = X(|fnj − f0 | > εj ), j ∈ N, and define the following set
 
\∞ [∞
X0 = lim Xj =  Xj  .
l=1 j=l

∞
S
Since for any l ∈ N, X0 ⊂ Xj , we obtain
j=l
 
∞
[ ∞
X ∞
X
µX0 6 µ  Xj  6 µXj < ηj . (4.3.5)
j=l j=l j=l

∞
P
The series ηj converges by assumption, so its reminder vanishes, therefore from (4.3.5) we get µX0 = 0.
j=1 S
Let now x 6∈ X0 . Then there exists an index l such that x 6∈ Xj , that is, x 6∈ Xj for all j > l.
j=l
Consequently, for any j > l, one has
|fnj (x) − f0 (x)| < εj .
Since εj −→ 0 by assumption, it follows that for any x ∈ X \ X0 , fnj (x) → f0 (x).
j→∞

In Example 4.3.11, we can take, for example, the subsequence of the sequence (fn )∞ n=1 consisting only
of functions ϕl,1 . This subsequence tends to zero at any point of the interval [0, 1] but the point x = 0.
4.4. APPROXIMATION OF MEASURABLE FUNCTIONS 65

4.4 Approximation of measurable functions

In Example 4.1.5 we introduced simple functions that play a very important role in the theory of Lebesgue
integral. We recall the definition of these function but using different notations.
The characteristic function (or indicator function) of a subset A ⊂ X is the function χA : X 7→ R
defined by (
1 if x ∈ A
χA =
0 if x 6∈ A
The function χA is measurable if and only if A is a measurable set.
m
Definition 4.4.1. Let X = ⊍ Ek . Define a function h : X 7→ R as follows: h(x) = ck whenever x ∈ Ek ,
k=1
where ck ∈ R. The function h is called a simple function. We will represent h in the form
m
X
h(x) = ck χEk (x). (4.4.1)
k=1

Here we suppose that ck = 0 if µEk = +∞.

In (4.4.1) some numbers ck can coincide. However, for a given simple function h we can always find
another representation of the form (4.4.1) with distinct numbers ck . ItScan be done as follows. Let c01 , c02 ,
. . . , c0l be all distinct values of the function h. Define the sets Ej0 = Ek , j = 1, 2 . . . , l. The sets Ej0
k: ck =c0j
are disjoint, since the sets Ek are disjoint. So, we obtain a new (evidently, unique) representation of the
function h:
Xl
h(x) = c0j χEj0 (x), (4.4.2)
j=1

where all c0j

are distinct.
As it was noticed in Example 4.1.5 if all the sets Ek are measurable, then the function h is measurable.
Moreover, the sets Ej0 are measurable if, and only if, the function h is measurable.

Theorem 4.4.2. (on approximation) Any nonnegative function measurable on X can be represented as
the limit of a non-decreasing sequence of nonnegative measurable simple functions.

Proof. Let f ∈ S + (X), where S + (X) = S + (X, M, µ) is the set of all nonnegative functions measurable
on X. The measure µ is σ-finite on X, so
+∞
[
X= Xl ,
l=1

where X1 ⊆ X2 ⊆ X3 ⊆ · · · .
For l ∈ N, consider the truncations

 f (x)
 if x ∈ Xl and f (x) 6 l,
Fl (x) = l if x ∈ Xl and f (x) > l,

0 otherwise.


Then,
Fl (x) % f (x), as l→∞ ∀x ∈ X. (4.4.3)
For a fixed l ∈ N, we partition the range of Fl (x), namely [0, l], as follows. We divide the interval [0, l]
into subintervals  
k−1 k
Il,k = , , k = 1, 2, . . . , l · 2l ,
2l 2l
66 CHAPTER 4. MEASURABLE FUNCTIONS

and consider the sets

Xl,0 = Xl (Fl = l) = Xl (f > l)
and    
k−1 k k−1 k
Xl,k = Xl 6 Fl < l = Xl 6f < l , k = 1, 2, . . . , l · 2l .
2l 2 2l 2
Obviously, !
l·2l
Xl = Xl,0 ⊍ ⊍ Xl,k .
k=1

We consider the simple functions


 l, x ∈ Xl,0 ,
hl (x) = k−1
 , x ∈ Xl,k ,
2l
and prove that the sequence (hl )∞ l=1 is non-decreasing and convergent to f .
a) The sets Xl,k , k = 1, 2, . . . , l · 2l , are measurable by Corollary 4.1.4. Therefore, for each l ∈ N, the
function hl (x) is a nonnegative measurable simple function.
b) Compare now the function hl and hl+1 . If the point x ∈ Xl is such that f (x) < l, then there exists an
k−1
index k, 1 6 k 6 l·2l such that x ∈ Xl,k and hl (x) = . When we consider the function hl+1 , we divide
2l  
2k − 2 2k − 1 2k − 2 k−1
each interval Il,k into two parts. Therefore, either f (x) ∈ l+1
, l+1 (so hl+1 (x) = l+1 = ),
  2 2 2 2l
2k − 1 2k 2k − 1
or f (x) ∈ l+1
, l+1 (so hl+1 (x) = l+1 ). In both cases we have hl+1 (x) > hl (x).
2 2 2
If Fl (x) = l for a given x ∈ Xl (that is, f (x) > l), then hl (x) = l, and hl+1 (x) > l = hl (x). So, for any
x ∈ Xl , hl (x) 6 hl+1 (x), and since hl (x) = 0 for any x ∈ X \ Xl , one has hl (x) 6 hl+1 (x) for any x ∈ X.
Thus, the sequence (hl )∞ l=1 is non-decreasing on X.
c) Let us show now that hl (x) % f (x) as l → ∞ for any x ∈ X.
Let x ∈ X. For any ε > 0 there exists l0 ∈ N such that for any l > l0 ,
ε
0 6 f (x) − Fl (x) < (4.4.4)
2
k−1
by (4.4.3). On the other hand, for Fl (x) 6 l, so there exists a number k = k(l) such that 6 Fl (x) 6
2l
k k−1
l
. At the same time, hl (x) = , so
2 2l
1
0 6 Fl (x) − hl (x) < . (4.4.5)
2l
Thus, there exists l1 ∈ N such that for any l > l1 we have
ε
0 6 Fl (x) − hl (x) < . (4.4.6)
2
Now from (4.4.4) and (4.4.6) we obtain that

0 6 f (x) − hl (x) < ε

for any l > max{l0 , l1 }, so hl (x) % f (x), as required.

Corollary 4.4.3. If f is a nonnegative bounded measurable function, then there exists a non-decreasing
sequence of simple functions convergent to f uniformly on X if µX < +∞.
4.4. APPROXIMATION OF MEASURABLE FUNCTIONS 67

Proof. If µX < +∞, then we can construct the sequence hl (x) without introducing the truncations Fl (x).
Since f is bounded, there exists a number l0 such that f (x) 6 l0 on X. Then for the sequence (hl )∞ l=1
constructed in the proof of Theorem 4.4.2 the inequality (4.4.5) holds for any l > l0 and for any x ∈ X.

Let us establish some additional approximation theorems for measurable functions.

Theorem 4.4.4. Let f ∈ S(X) and let f be a.e. finite on X. Then for any ε > 0 there exists a bounded
function g ∈ S(X) s.t. µX(f 6= g) < ε.
Proof. Let Xn = X(|f | > n), and X∞ = X(|f | = ∞). By assumption µX∞ = 0. Moreover,
∞
\
X1 ⊃ X2 ⊃ X3 ⊃ · · · , X∞ = Xn = lim Xn .
n→∞
n=1

Since µ is σ-additive, by Theorem 2.2.5 we have

µX∞ = lim µXn .

n→∞

So, given ε > 0, there exists N > 0 s.t. µXn < ε for any n > N . Now define the following function
(
f (x), x ∈ X \ XN ,
g(x) =
0, x ∈ XN .

This is a desirable function, since g ∈ S(X), |g(x)| 6 N on X, and the X(f 6= g) = XN is of measure not
exceeding ε.
Remark 4.4.5. Thus, every measurable a.e. finite function becomes bounded outside of a set of arbitrary
small measure.
Measurable functions can also be approximated by a class of piece-wise constant functions called step
functions.
Definition 4.4.6. A function f of the form
m
X
f (x) = cj χKj (x),
j=1

where Kj , j = 1, . . . , m, are bricks in Rn , and cj are some real numbers, is called a step function.
Step functions are particular cases of simple functions. But unlike simple functions, every step function
is measurable.
Theorem 4.4.7. Let f ∈ S(X). Then there exists a sequence of step functions (ϕk )∞
k=1 that converges
to f (x) a.e. on X.
Proof. By Theorem 4.4.2, every nonnegative measurable function can be represented as a limit of a non-
decreasing sequence of simple nonnegative functions. It is clear that an arbitrary measurable function can
also be approximated by simple functions. Indeed, if f ∈ S(X), then f = f + − f − , where f + and f −
are nonnegative measurable functions. Then by Theorem 4.4.2 there exist two non-decreasing sequences
of simple nonnegative functions (hn )∞ ∞
n=1 and (gn )n=1 such that

f + (x) = lim hn (x) f − (x) = lim gn (x), x ∈ X.

n→∞ n→∞

Now for the functions ψn (x) = hn (x) − gn (x) we obviously have

|ψn (x)| = |hn (x)| + |gn (x)| 6 |ψn+1 (x)| and lim ψn (x) = f (x) ∀x ∈ X.
n→∞
68 CHAPTER 4. MEASURABLE FUNCTIONS

Thus, it is enough to establish that if A is a measurable set with finite measure, then f (x) = χA (x)
can be approximated by step functions. By Definition 3.4.1 for any ε > 0 there exists an elementary set
B s.t. µ(A M B) < ε. As we mentioned in Lecture ??, the elementary set B can be represented as follows
m
B = ⊍ Ki ,
i=1

where Ki are some bricks. Consider the step function

m
X
ψ(x) = χKi (x).
i=1

Then we have that µX(f 6= ψ) < ε. Consequently, for every k ∈ N, there exists a step function ψk (x) s.t.
1
if Ek = X(f 6= ψk ) then µXk 6 k . Consider now the limit superior of the sequence (Ek )∞
k=1 :
2
∞ [
\ ∞
E= El .
k=1 l=k
∞  ∞ 
−k
S S
It is easy to see that µ El 62 , so µE = lim µ El = 0. We thus obtain that
l=k k→∞ l=k

χA (x) = lim ψk (x) ∀x ∈ X \ E,

k→∞

and µE = 0, as required.
Now we are in a position to show that measurable functions can be approximated by continuous
functions. First we prove the Luzin theorem stating that any a.e. finite measurable function on X is
”nearly” continuous.
Theorem 4.4.8 (Luzin). Suppose f is measurable and a.e. finite on X, and µX < +∞. Then for any
ε > 0 there exists a closed set Fε ⊂ X such that
µ(X \ Fε ) < ε
and f is continuous on Fε (that is, the restriction f |Fε of the function f to the set Fε is continuous on Fε ).
By f F we mean the restriction of f to the set Fε . The conclusion of the theorem states that if f is
ε
viewed as a function defined only on Fε , then f is continuous. However, the theorem does not make the
stronger assertion that the function f defined on X is continuous at the points of Fε .
a.e.
Proof. By Theorem 4.4.7, there exists a sequence (fn )∞
n=1 of step functions so that fn −−→ f on X. Then
1
for every n ∈ N we may find sets Xn so that µXn 6 n and fn is continuous outside Xn . By Egorov’s
2
ε
theorem 4.3.7, we may find a set A ε on which fn → f uniformly and µA ε < . Then we consider the set
3 3 3
∞
!
[
F0 = A \ Xn
n=N
∞
P 1 ∞
P ε
for N so large that µXn = n
< . Now for every n > N the function fn is continuous on
n=N n=N 2 3
F0 thus f (being the uniform limit of (fn )∞
n=1 ) is also continuous on F0 . To finish the proof, we merely
ε
need to approximate the set F0 by a closed set Fε ⊂ F0 such that µ(F \ Fε ) < , that is possible by
3
Theorem 3.7.1.
Thus, any measurable function is continuous outside of a set of arbitrary small measure. The next
theorem gives us a tool for approximation of measurable functions by continuous functions.
4.5. PROBLEMS 69

4.5 Problems
√
Problem 4.1. Let f ∈ S + (X). Prove that f ∈ S + (X).
√
Problem 4.2. Let f ∈ S(X). Prove that 3 f ∈ S(X).
Problem 4.3. Let f ∈ S(X). Is sgn f measurable?
(
f (x) if |f (x)| 6 n
Problem 4.4. Let f ∈ S(X), and let [f (x)]n = . Is [f (x)]n measurable?
0 if |f (x)| > n

Problem 4.5. Prove that if the set X(f > r) is measurable for any r ∈ Q, then f is measurable. Is
converse statement true? Explain the answer.
Problem 4.6. Prove that if for any a, b ∈ R either the sets X(a 6 f 6 b), or the sets X(a < f 6 b), or
the sets X(a 6 f < b), or the sets X(a < f < b) are measurable, then f ∈ S(X).
Problem 4.7. Let a function f : Rn 7→ R be continuous, and the functions gk : R 7→ R, k = 1, . . . , n, be
measurable on R. Prove that the function h : R 7→ R defined as follows

h(x) = f (g1 (x), . . . , gn (x)),

is measurable.
Problem 4.8. Prove that if {fk (x)}+∞
k=1 ∈ S(X), then the set of points x ∈ X such that lim fk (x) exists
k→+∞
is measurable.
Problem 4.9. Let the function f : [0, 1] 7→ R be differentiable on [0, 1]. Prove that its derivative (the
function f 0 : [0, 1] 7→ R) is Lebesgue measurable.
Problem 4.10. A complex-valued function f : X 7→ C (f (x) = u(x) + iv(x), u, v : X 7→ R) is called
measurable if its real and imaginary parts u and v are measurable. Prove that the absolute value and the
argument of a complex-valued measurable function are measurable functions.
a.e. a.e. a.e.
Problem 4.11. Let {fn }+∞
n=1 ∈ S(X), fn −→ f and fn −→ g on X. Prove that f = g on X.
µ µ a.e.
Problem 4.12. Let {fn }+∞
n=1 ∈ S(X), fn −→ f and fn −→ g on X. Prove that f = g on X.

Problem 4.13. Let X = [0, 1], and µ be the Lebesgue measure on X. Let fn = e−nx , n ∈ N, and f0 ≡ 0.
µ
Prove that fn −→ f0 on X.
Problem 4.14. Let X = [0, π], and µ be the Lebesgue measure on X. Let fn = (sin nx)n , n ∈ N, and
µ
f0 ≡ 0. Prove that fn −→ f0 on X. Does {fn }+∞
n=1 converge to f0 ≡ 0 almost everywhere on X?

Hint: Use the followig Chebyshev’s theorem:

Theorem 4.14.9. For any irrational number α and for any real number β the inequality
3
|αq − p − β| <
p

has infinitely many solutions (p, q), where p ∈ N, q ∈ Z.

α
Problem 4.15. Prove that the sequence fn (x) = | sin nx|n does not converge to f0 ≡ 0 a.e. on X = [0, π]
if 0 6 α 6 2. Does it have an a.e. limit if α > 2? For what values of α does it converge in measure?
Problem 4.16. For any δ > 0 find the “Egoroff set” Xδ (µXδ < δ) such that the sequence fn (x) =
n sin x
, x ∈ [0, π], n ∈ N, converges uniformly on [0, π] \ Xδ .
1 + n2 sin2 x
70 CHAPTER 4. MEASURABLE FUNCTIONS

Problem 4.17. For any δ > 0 find the “Egoroff set” Xδ (µXδ < δ) such that the sequence fn (x), x ∈ [0, 1],
n ∈ N, converges uniformly to f ≡ 0 on [0, 1] \ Xδ . Here

2nx xn
a) fn (x) = e−n(1−x) ; b) fn (x) = ; b) fn (x) = ;
1 + n2 x2 1 + xn
 1
 n2 x, 06x6 ,
n


  
 2 1 2
d) fn (x) = n2 −x , <x< ,
 n n n
2



 0, 6 x 6 1.
n
Problem 4.18. For
! any δ > 0 find the “Egoroff set” Xδ (µXδ < δ) such that the sequence fn =
r
1 √
n x + − x , x ∈ [0, 1], n ∈ N, converges uniformly on [0, 1] \ Xδ .
n

Problem 4.19 (Lusin’s theorem). Suppose f is measurable and finite-valued on X, and µX < +∞. Then
for any ε > 0 there exists a closed set Fε ⊂ X such that

µ(X \ Fε ) < ε

and f is continuous on Fε (that is, the restriction f |Fε of the function f to the set Fε is continuous on Fε ).
Hint: Use approximation of measurable functions by simple functions and Egorov’s theorem.
P.S. This problem actually says that every measurable and a.e. finite function is nearly continuous.
Problem 4.20. Let X be a compact set in Rn , and let F ⊂ X be closed. Suppose that ψ is defined and
continuous on F . Prove that there exists a function ϕ(x) defined on X satisfying the following properties
a) ϕ(x) is continuous on X;

b) ϕ(x) = ψ(x) for any x ∈ F ;

c) max |ϕ(x)| = max |ψ(x)|.
Using this fact and Problem 4.19 prove that for any f ∈ S(X) with µX(f = ±∞) = 0 and for any ε > 0,
there exists a function g(x) continuous on X such that µX(f 6= g) < ε, and if |f (x)| 6 M , then |g(x)| 6 M .

Problem 4.21. Construct a non-decreasing sequence of Lebesgue measurable simple functions hn (x)
converging to a function f (x) on X where
a) f (x) = e−(3x+1) , X = [0, +∞);
b) f (x) = e2x+1 , X = [1, +∞);
1
c) f (x) = , X = (2, 20];
x−2
1
d) f (x) = , X = [0, +∞);
(2x + 1)(2x + 3)
Problem 4.22. Prove that a function f monotone on an interval [a, b] ⊂ R is measurable (w.r.t. the
Lebesgue measure).
Problem 4.23. Let (X, M, µ) be a measure space, and A ∈ M with µ(A) < +∞. Suppose that f
is a finite measurable function defined on A. Prove that the function g(t) = µA(f > t) is non-strictly
decreasing (non-increasing) and right-continuous on R.
Chapter 5

Integration theory

In definition of the Lebesgue integral of a measurable function, we approximate the function by simple
functions. By contrast, in definition of the Riemann integral of a function f : [a, b] 7→ R, we partition the
domain [a, b] into subintervals and approximate f by step functions that are constant on these subintervals.
This difference is sometime expressed by saying that in the Lebesgue integral we partition the range, and in
the Riemann integral we partition the domain. This trick allows to extend the set of integrable functions.
Moreover, the Lebesgue integral is defined in Rn identically for any n, while the Riemann integral must be
defined first on R, and then extended to Rn , n > 2. For functions defined on an abstract measure space,
the Riemann integral is not defined at all.
Let (X, M, µ) be a measure space, where the measure µ is supposed to be σ-additive and complete.
And let S(X, M, µ) = S(X) be the set of all functions µ-measurable on X. In what follows we assume
all the considered sets to belong to the σ-algebra M and all the considered functions to be measurable
(unless otherwise mentioned).
We construct the Lebesgue integral into three steps: First, we define the integral for simple functions,
then we extend it to all nonnegative measurable functions, and then to all measurable functions.

5.1 Integral of simple functions

Let h be a measurable simple function such that all the sets in its representation (4.4.1) are measurable
and of finite measure.

Definition 5.1.1. The expression

Z m
X
h(x)dµ = ck µEk (5.1.1)
X k=1

is called the integral of the simple function h over the set X w.r.t. the measure µ. Here we use the
convention that 0 · ∞ = 0.

One can verify that the value of the integral in (5.1.1) is independent on the representation (4.4.1).
Indeed, among all the representations of the function h there exists a unique representation (4.4.2) in
which all the numbers c0j are distinct. Following the way of construction the representation (4.4.2), one
has    
Xm Xl [ X l [ X∞
ck µEk = c0j  µEk  = c0j µ  Ek  = c0j µEj0 .
k=1 j=1 k: ck =c0j j=1 k: ck =c0j j=1

Thus, the integral of a simple function does not depend on its representation.

Let us study some properties of the integral of simple functions.

71
72 CHAPTER 5. INTEGRATION THEORY
Z
1) If µX = 0, then h(x)dµ = 0.
X
Z Z
2) αh(x)dµ = α h(x)dµ ∀α ∈ R.
X X
These two properties immediately follow from the definition of the integral.
Z Z Z
3) (h1 (x) + h2 (x))dµ = h1 (x)dµ + h2 (x)dµ.
X X X
Let
m
X l
X
h1 (x) = ck χEk (x), h2 (x) = c0j χEj0 (x).
k=1 k=1
Consider the sets \
Ek,j = Ek Ej0 , k = 1, 2, . . . , m j = 1, 2, . . . , l.
Clearly,
l m l m
Ek = ⊍ Ek,j , Ek0 = ⊍ Ek,j , X = ⊍ ⊍ Ek,j .
j=1 k=1 j=1 k=1

It is also easy to see that h1 (x)+h2 (x) = ck +c0j if x ∈ Ek,j , k = 1, 2, . . . , m, j = 1, 2, . . . , l. Therefore,

the function h1 (x) + h2 (x) is simple, so
Z m X
X l m
X l
X l
X m
X
(h1 (x) + h2 (x)) = (ck + c0j )µEk,j = ck µEk,j + c0j µEk,j =
X k=1 j=1 k=1 j=1 j=1 k=1

m
X m
X Z Z
= ck µEk + c0j µEj0 = h1 (x)dµ + h2 (x)dµ.
k=1 j=1 X X

4) If X = X 0 ⊍ X 00 , then Z Z Z
h(x)dµ = h(x)dµ + h(x)dµ.
X X0 X 00

m
P
Let h(x) = ck χEk (x). Consider the sets
k=1
\ \
Ek0 = X 0 Ek , Ek00 = X 00 Ek , k = 1, 2, . . . , m.

Then we have Ek = Ek0 ⊍ Ek00 , k = 1, 2, . . . , m, and

Z Xm X m X m m
X Z Z
h(x)dµ = ck µEk = ck (µEk0 + µEk00 ) = ck µEk0 + ck µEk00 = h(x)dµ + h(x)dµ.
X k=1 k=1 k=1 k=1 X0 X 00

5) If h1 (x) 6 h2 (x) for all x ∈ X, then

Z Z
h1 (x)dµ 6 h2 (x)dµ.
X X
Z Z
If h(x) > 0 on X, then, clearly, h(x)dµ > 0. Therefore, (h1 (x) − h2 (x))dµ > 0, and the
X X
property 5) follows now from the properties 2) and 3).
5.1. INTEGRAL OF SIMPLE FUNCTIONS 73

6) If a non-increasing (i.e. non-strictly decreasing) sequence of nonnegative simple functions (hn )∞

n=1
converges to 0 almost everywhere on X, then
Z
lim hn (x)dµ = 0. (5.1.2)
n→∞
X

If µX = 0, then we are done, since all the integrals equal zero.

Suppose now that 0 < µX < +∞. The simple function h1 (x) has only finitely many values (in
its range), therefore, it is bounded, so there exists a number M > 0 such that h1 (x) 6 M on X.
Since the sequence (hn )∞ n=1 is non-increasing and nonnegative, we have

0 6 hn (x) 6 M, n ∈ N, x ∈ X. (5.1.3)

ε
Fix a number ε > 0 and define δ := . By Egoroff’s theorem, Theorem 4.3.7, there exists a
2M
measurable set Xδ ⊂ X such that µXδ < δ and the sequence (hn )∞ n=1 converges to 0 uniformly on
X \ Xδ . So, there exists a number n0 ∈ N such that for any n > n0
ε
0 6 hn (x) < , x ∈ X \ Xδ . (5.1.4)
2 µX
Consider the following simple function


 M, x ∈ Xδ ,

hε (x) =
ε
x ∈ X \ Xδ .


 ,
2 µX
From (5.1.3)–(5.1.4) it follows that hn (x) 6 hε (x) for any x ∈ X whenever n > n0 . So by the
property 5), one has for n > n0
Z Z
ε
0 6 hn (x)dµ 6 hε (x)dµ = M · µXδ + · µ(X \ Xδ ) <
2 µX
X X

ε ε ε
<M ·δ+ · µX = M · + = ε.
2 µX 2M 2

Finally, if µX = +∞, then there exists X1 ⊂ X such that µX1 < +∞ and h1 (x) = 0 for any
x ∈ X \ X1 . Since the sequence of non-negative simple functions hn (x) is non-increasing, we have
that hn (x) = 0, n ∈ N, for any x ∈ X \ X1 . Now by the previous result, we have that (5.1.2) is true.
7) If a sequence (hn )∞
n=1 of nonnegative simple functions is non-increasing and
Z
lim hn (x)dµ = 0, (5.1.5)
n→+∞
X

a.e.
then hn −→ 0 on X.
As before, suppose µX > 0 (otherwise, the assertion is trivial). Since the sequence (hn )∞
n=1 is
non-increasing and is bounded from below, it converges at any point x of the set X. Let

g(x) = lim hn (x), x ∈ X.

n→+∞

The functions hn are nonnegative on X, so is the function g(x), that is, g(x) > 0 for any x ∈ X. It
a.e.
suffices to show that g(x) = 0 on X.
74 CHAPTER 5. INTEGRATION THEORY

Indeed, it is clear that

∞   ∞
[ 1 [
X0 = X(g 6= 0) = X g> = Xk .
k
k=1 k=1

On the contrary, suppose that there exists an index k0 such that µXk0 = δ0 > 0, and consider a
simple function h0 defined as follows.
1

 k ,

 x ∈ Xk0 ,
0
h0 (x) =


0, x ∈ X \ Xk0 .


For the function h0 , one has

hn (x) > g(x) > h0 (x), n ∈ N, x ∈ X.

The property 5) now implies

Z Z
1 1
hn (x)dµ > h0 (x)dµ = · µX0 + 0 · µ(X \ X0 ) = · δ0 > 0, n ∈ N,
k0 k0
X X

where we use our convention that 0 · ∞ = 0. This contradicts with the condition (5.1.5), so µXk = 0
for all k ∈ N. From σ-additivity of the measure µ we have
∞
X
µX0 6 µXk = 0.
k=0

5.2 Integral of nonnegative measurable functions

If f ∈ S + (X) (that is, f is nonnegative and measurable on X), then by Approximation Theorem 4.4.2
there exists a non-decreasing sequence of nonnegative simple functions (hn )∞
n=1 convergent to f on the set
X. By the property 5) of integral of simple functions the sequence of integrals
 
Z
 hn (x)dµ
X n∈N

is non-decreasing and has a limit (finite or infinite).

Definition 5.2.1. The expression
Z Z
f (x)dµ = lim hn (x)dµ (5.2.1)
n→+∞
X X

is called
Z the integral of the nonnegative function f over the set X w.r.t. measure µ.
If f (x)dµ is finite, then the function f is called integrable (or summable) on X. The set of all
X
nonnegative integrable functions on X is denoted L+ (X, M, X), or, briefly, L+ (X) if the measure µ is
defined in advance.
Let us emphasize once again that the integral exists for any nonnegative measurable function but it
can be infinite.
Before we start to study properties of the integral of nonnegative functions, we have to prove that
the integral (5.2.1) is defined properly. Namely, we have to prove that it is independent on the sequence
(hn )∞
n=1 which we use to approximate f . We prove even a more general fact.
5.2. INTEGRAL OF NONNEGATIVE MEASURABLE FUNCTIONS 75

Theorem 5.2.2. Let f, g ∈ S + (X) and f (x) 6 g(x) for any x ∈ X. Suppose that (hn )∞ ∞
n=1 and (vn )n=1 are
non-decreasing sequences of nonnegative simple functions approximating the functions f and g, respectively.
Then Z Z
lim hn (x)dµ 6 lim vn (x)dµ. (5.2.2)
n→+∞ n→+∞
X X

Proof. Consider the difference hk − vn for a fixed index k, n ∈ N. Since the sequences (vn )∞
n=1 is non-
decreasing, the sequence (hk − vn )∞
n=1 is non-increasing, so it has limit as n → +∞

lim (hk (x) − vn (x)) 6 lim (f (x) − vn (x)) = f (x) − g(x) 6 0

n→+∞ n→+∞

for any x ∈ X.
Additionally, the sequence (hk − vn )+
n∈N of positive parts of the sequence (hk − vn )n∈N is non-increasing
as well. Therefore,
lim (hk (x) − vn (x))+ = 0
n→+∞

for all x ∈ X.
By the property 6) of integrals of simple functions, we have
Z
lim (hk (x) − vn (x))+ dµ = 0.
n→+∞
X

From the inequality (hk (x) − vn (x)) 6 (hk (x) − vn (x))+ , it follows that
Z
lim (hk (x) − vn (x))dµ 6 0. (5.2.3)
n→+∞
X

Here the integral (finite or equal −∞) exists, since the sequence of integrals is non-increasing by the
property 5). From (5.2.3) we obtain
Z Z
hk (x)dµ 6 lim vn (x)dµ.
n→+∞
X X

From this inequality, we get (5.2.2) as k → +∞.

If we put g(x) := f (x) and take two different non-decreasing sequences of nonnegative simple functions
approximating f , we get by Theorem (5.2.2) that, on one hand,
Z Z
lim hn (x)dµ 6 lim vn (x)dµ,
n→+∞ n→+∞
X X

and, on the other hand, interchanging hn and vn (we can do this because of symmetry)
Z Z
lim vn (x)dµ 6 lim hn (x)dµ.
n→+∞ n→+∞
X X

Consequently, Z Z
lim vn (x)dµ = lim hn (x)dµ.
n→+∞ n→+∞
X X
Thus, the definition of integral of nonnegative measurable functions does not depend on the non-
decreasing sequences of simple nonnegative functions approximating these functions. Moreover, this fact
implies that for any nonnegative measurable simple function h(x), Definition 5.1.1 of the integral coincides
with Definition 5.2.1, since we can put in Definition 5.2.1 hn (x) := h(x).

Now we are in a position to study properties of integral of nonnegative measurable functions.

76 CHAPTER 5. INTEGRATION THEORY
Z
1) If µX = 0, then f (x)dµ = 0 for any nonnegative function f .
X

On a null set any function is measurable, since the measure µ is assumed to be complete (see
Example 4.1.2), so f ∈ S + (X). The property is obvious for simple functions, and can be proved for
any function f in S + (X) by passage to the limit (when we approximate f by simple functions).

2) If α > 0, then
Z Z
αf (x)dµ = α f (x)dµ.
X X
+ +
Moreover, if f ∈ L (X), then αf ∈ L (X).
If f (x) > 0, then αf (x) > 0 for any x ∈ X. If hn % f , then αhn % αf . Therefore,
Z Z Z Z
αf dµ = lim αhn dµ = α lim hn dµ = α f dµ.
n→+∞ n→+∞
X X X X

The second assertion is obvious.

3) If f, g ∈ S + (X), then
Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ,
X X X

and if f, g ∈ L+ (X), then f + g ∈ L+ (X).

Let hn % f and vn % g. Then hn + vn % f + g, and
Z Z
(f (x) + g(x))dµ = lim (hn (x) + vn (x))dµ =
n→+∞
X X
Z Z Z Z
= lim hn (x)dµ + lim vn (x)dµ = f (x)dµ + g(x)dµ
n→+∞ n→+∞
X X X X

The second assertion follows from the proved identity.

4) If X = X 0 ⊍ X 00 , where X 0 , X 00 ∈ M(X), then

Z Z Z
f (x)dµ = f (x)dµ + f (x)dµ,
X X0 X 00

and f ∈ L+ (X) if, and only if, f ∈ L+ (X 0 ) and f ∈ L+ (X 00 ).

If hn % f on X, then hn % f on both X 0 and X 00 . Conversely, if h0n % f on X 0 and h00n % f X 00 ,
then we set hn (x) := h0n (x) for x ∈ X 0 , and hn (x) := h00n (x) for x ∈ X 00 to get hn (x) % f for x ∈ X.
It is left to use the property 4) of integral of simple functions and pass to the limit as n → +∞.

5) If f (x) 6 g(x), x ∈ X, then

Z Z
f (x)dµ 6 g(x)dµ.
X X

In particular, if g ∈ L+ (X), then f ∈ L+ (X).

This property is, in fact, Theorem 5.2.2.
5.2. INTEGRAL OF NONNEGATIVE MEASURABLE FUNCTIONS 77

6) If f ∈ L+ (X), then µX(f = +∞) = 0. In other word, a nonnegative integrable function on X is

finite almost everywhere on X.
Suppose that it is not true. Then µX0 = a > 0, where X0 = X(f = +∞). Consider the following
sequence of simple functions: (
n, x ∈ X0 ,
hn (x) =
0, x ∈ X \ X0 .
Since f (x) > hn (x) for x ∈ X, n ∈ N, then by the property 5), one has
Z Z Z
f (x)dµ > hn (x)dµ = ndµ = n · a ∀n ∈ N.
X X X0
R
The number a is positive, therefore, the integral f (x)dµ cannot be finite, a contradiction, since
X
+
f ∈ L (X) by assumption.
7) (Chebyshev’s inequality) If f ∈ L+ (X), then
Z
1
µX(f > c) 6 f (x)dµ, ∀c > 0.
c
X

Let c > 0. Denote Xc = X(f > c). Then we have

Z Z Z Z
f (x)dµ = f (x)dµ + f (x)dµ > f (x)dµ > c · µXc ,
X Xc X\Xc Xc

as required.
Z
+
8) If f ∈ S (X) and f (x)dµ = 0, then f (x) = 0 almost everywhere on X.
X
∞

S 1
It is easy to check that X(f > 0) = X f> . By Chebyshev’s inequality we obtain
n=1 n
  Z
1
µX f > 6 n · f (x)dµ = 0.
n
X

Due to σ-additivity of the measure µ, we have

∞  
X1
µX(f > 0) 6 µX f > = 0.
n=1
n

a.e.
9) If f = g on X, then Z Z
f (x)dµ = g(x)dµ.
X X
+ +
So f ∈ L (X) if, and only, if g ∈ L (X).
Let X 0 = X(f 6= g) and X 00 = X \ X 00 . By assumption µX 0 = 0, and by construction f (x) = g(x)
for any x ∈ X 00 . From the properties 4) and 1) we have
Z Z Z Z Z Z
f (x)dµ = f (x)dµ + f (x)dµ = g(x)dµ + g(x)dµ = g(x)dµ.
X X0 X 00 X0 X 00 X
78 CHAPTER 5. INTEGRATION THEORY

∞
fk (x), where fk ∈ S + (X), k ∈ N. Then f ∈ S + (X), and
P
Theorem 5.2.3 (B. Levi). Let f (x) =
k=1
Z ∞ Z
X
f (x)dµ = fk (x)dµ. (5.2.4)
X k=1 X

n
fk (x). By the property 3) of measurable functions, we have sn (x) ∈ S + (X).
P
Proof. Let sn (x) =
k=1
Therefore, f ∈ S + (X) by Theorem 4.3.1 as the limit of measurable functions. It is left to prove the
identity (5.2.4).
Since f (x) > sn (x) for all n ∈ N and x ∈ X, by the properties 3) and 5) of integral of nonnegative
functions, we have
Z Z Xn Z
f (x)dµ > sn (x)dµ = fk (x)dµ.
X X k=1 X

If here n → +∞, one obtains

Z ∞ Z
X
f (x)dµ > fk (x)dµ. (5.2.5)
X k=1 X

Let us now prove the opposite inequality. Let (hk,j )∞

j=1 be a sequence of nonnegative simple functions
such that hk,j % fk as j → ∞, k ∈ N, for any x ∈ X. Consider the functions
n
X
gn (x) = hk,n (x).
k=1

The functions gn (x) are nonnegative measurable simple functions as sums of nonnegative measurable simple
functions. Moreover, the sequence (gn )∞
n=1 is non-decreasing, since

n+1
X n
X
gn+1 (x) = hk,n+1 (x) > hk,n (x) = gn (x),
k=1 k=1

therefore, the sequence (gn )∞

n=1 converges on X to a function g which is nonnegative and measurable by
Theorem 4.3.1.
Let us fix n ∈ N and take an arbitrary p ∈ N. Then
n
X n+p
X n+p
X
hk,n+p (x) 6 hk,n+p (x) = gn+p (x) 6 fk (x) 6 f (x).
k=1 k=1 k=1

As p → +∞, we obtain
n
X
sn (x) = fk (x) 6 g(x) 6 f (x). (5.2.6)
k=1

Since lim sn (x) = f (x), then from (5.2.6) it follows that g(x) = f (x) for x ∈ X. Consequently,
n→+∞
gn (x) % f (x) for all x ∈ X. So we get by definition of the integral
Z Z Z X n
f (x)dµ = lim gn (x)dµ = lim hk,n (x)dµ 6
n→+∞ n→+∞
X X X k=1

n
Z X n Z
X ∞ Z
X
lim fk (x)dµ = lim fk (x)dµ = fk (x)dµ.
n→+∞ n→+∞
X k=1 k=1 X k=1 X

This inequality together with (5.2.5) imply the identity (5.2.4).

5.2. INTEGRAL OF NONNEGATIVE MEASURABLE FUNCTIONS 79

Remark 5.2.4. Both parts of the identity (5.2.4) can be infinite.

Corollary 5.2.5. If (fn )∞
n=1 is a monotone non-decreasing sequence of nonnegative measurable functions,
and f (x) = lim fn (x), x ∈ X, then
n→+∞
Z Z
f (x)dµ = lim fn (x)dµ. (5.2.7)
n→+∞
X X

Proof. Consider the functions ϕ1 (x) = f1 (x), ϕk (x) = fk (x) − fk−1 (x), k > 2. The functions ϕk (x) are
∞
P
measurable and nonnegative on X, and f (x) = ϕk (x). So by Levi’s theorem, Theorem 5.2.3, we have
k=1
Z ∞ Z
X n Z
X n
Z X Z
f (x)dµ = ϕk (x)dµ = lim ϕk (x)dµ = lim ϕk (x)dµ = lim fk (x)dµ.
n→+∞ n→+∞ n→+∞
X k=1 X k=1 X X k=1 X

Remark 5.2.6. Both parts of the identity (5.2.7) can be infinite.

Theorem 5.2.7 (Fatou). Let (fn )∞
n=1 be a sequence of nonnegative measurable functions on X. Then
Z Z
lim fn (x)dµ 6 lim fn (x)dµ. (5.2.8)
X X

Proof. For any x ∈ X, we have  

lim fn (x) = sup inf {fk (x)} .
n∈N k>n

Introduce the functions gn (x) = inf {fk (x)}. Then it is clear that
k>n

gn (x) 6 fn (x), ∀x ∈ X, ∀n ∈ N. (5.2.9)

The sequence gn (x) is non-decreasing, since the number of terms in the sequence {fn (x) : k > n}
decreases as n grows. Consequently, the limit lim gn (x) exists, and
n→+∞
 
lim gn (x) = sup{gn (x)} = sup inf {fk (x)} = lim fn (x).
n→+∞ n∈N n∈N k>n

Now from Corollary 5.2.5 it follows that

Z Z Z
lim fn (x)dµ = lim gn (x)dµ = lim gn (x)dµ. (5.2.10)
n→+∞ n→+∞
X X X
 ∞
Z
The sequence  fn (x)dµ has the limit inferior (finite or equal to +∞), since its terms are
X n=1
nonnegative. So there exists a subsequences of indices (nj )∞
j=1 such that
Z Z
lim fn (x)dµ = lim fnj (x)dµ (5.2.11)
j→+∞
X X

Now from (5.2.9)–(5.2.11) we obtain

Z Z Z Z Z
lim fn (x)dµ = lim gn (x)dµ = lim gnj (x)dµ 6 lim fnj (x)dµ = lim fn (x)dµ.
n→+∞ j→+∞ j→+∞
X X X X X
80 CHAPTER 5. INTEGRATION THEORY

Corollary 5.2.8. Let a sequence (fn )∞

n=1 of nonnegative measurable functions on X converge to a non-
negative function f a.e. on X, and
Z
fn (x)dµ 6 C, n ∈ N.
X

Then Z
f (x)dµ 6 C
X

Proof. Indeed, the function lim fn is nonnegative by assumption and is measurable by Theorem 4.3.1.
a.e.
Since f = lim fn (= lim fn ) on X by assumption, the function f is measurable as well (by Lemma 4.3.3).
From the property 9) of integral of nonnegative functions and from Theorem 5.2.7, we obtain
Z Z Z
f (x)dµ = lim fn (x)dµ 6 lim fn (x)dµ 6 C, n ∈ N,
X X X

as required.

Example 5.2.9. Note that in (5.2.8) strict inequality is possible. To show this, let us consider the
Lebesgue measure µ on the interval [0, 1], and the sequence

 n, 0 6 x 6 1 ,

fn (x) = n
1
 0,
 < x 6 1.
n
Then we have Z  
1 1
fn (x)dµ = n · + 0 · 1 − = 1, ∀n ∈ N,
n n
[0,1]
Z
therefore, lim fn (x)dµ = 1. At the same time,
[0,1]

(
+∞, x = 0,
f (x) = lim fn (x) =
0, 0 < x 6 1.
Z Z
a.e.
so f (x) = g(x) ≡ 0 on [0, 1]. Thus, f (x)dµ = g(x)dµ = 0, and we get
[0,1] [0,1]
Z Z
0= lim fn (x)dµ < lim fn (x)dµ = 1.
X X

5.3 Integral of measurable functions

Let f be an arbitrary measurable function on a set X. Then it can be represented as follows

f (x) = f + (x) − f − (x),

where f + and f − are defined in Definition 4.2.3.

The functions f + and f − are nonnegative and measurable on X by the property 7) of measurable
functions.
5.3. INTEGRAL OF MEASURABLE FUNCTIONS 81

Definition 5.3.1. The expression

Z Z Z
f (x)dµ = f + (x)dµ − f − (x)dµ (5.3.1)
X X X

is called the integral of the function f over the set X w.r.t. measure µ.

From the definition it follows that the integral (5.3.1) does not always exist. In fact, the following four
cases are possible:
Z Z Z
+ −
1) The integrals f (x)dµ and f (x)dµ are finite, so f (x)dµ is also finite.
X X X
Z Z
2) The integral f + (x)dµ is infinite but the integral f − (x)dµ is finite. In this case, the integral
Z X X

f (x)dµ = +∞.
X
Z Z
3) The integral f + (x)dµ is finite but the integral f − (x)dµ is infinite. In this case, the integral
Z X X

f (x)dµ = −∞.
X
Z Z Z
+ −
4) The integrals f (x)dµ and f (x)dµ are infinite. In this case, the integral f (x)dµ does not
X X X
exist.

Definition 5.3.2.
Z A measurable function f is called (Lebesgue) integrable (or summable) on X w.r.t.
measure µ if f (x)dµ exists and finite.
X

The set of all integrable functions on X is denoted L(X, M, X), or, briefly, L(X) if the measure µ is
defined in advance. It is clear that L+ (X) ⊂ L(X), and that f ∈ L(X) if, and only if, f + , f − ∈ L+ (X).
Let us study properties of the integral of measurable functions.
Z
1) If µX = 0, then f (x)dµ = 0 for any function f .
X
This property follows from the property 1) of integral of nonnegative functions.

2) The function f is integrable on X if, and only if, the function |f | is integrable on X. Moreover, if
f ∈ L(X), then
Z Z
f (x)dµ 6 |f (x)|dµ (5.3.2)
X X

Indeed, let f ∈ L(X), then f + , f − ∈ L+ (X). Since |f | = f + + f − , we obtain |f | ∈ L+ (X) by the

property 2) of integral of nonnegative functions.
Conversely, let |f | ∈ L+ (X). Since f + (x) 6 |f (x)| and f − (x) 6 |f (x)| for any x ∈ X, by the
property 5) of integral of nonnegative functions we get f + , f − ∈ L+ (X), so f ∈ L(X). The inequal-
ity (5.3.2) follows from the fact that |f | = f + + f − .
82 CHAPTER 5. INTEGRATION THEORY

Z Z Z Z Z
f (x)dµ = f + (x)dµ − f − (x)dµ 6 f + (x)dµ + f − (x)dµ =
X X X X X
Z Z Z
= f + (x)dµ + f − (x)dµ = |f (x)|dµ.
X X X
Z Z
a.e.
3) If f = g and the integral f (x)dµ exists (but can be infinite), then the integral g(x)dµ exists,
X X
and Z Z
f (x)dµ = g(x)dµ.
X X

a.e. a.e. a.e.

Since f = g, we have f + = g + and f − = g − , so g + , g − ∈ L+ (X) by the property 9) of integral
of nonnegative functions.
a.e.
Corollary 5.3.3. If f ∈ L(X), f = g on X, then g ∈ L(X).

4) If f ∈ L(X), then µX(f = ±∞) = 0.

If f ∈ L(X), then f + , f − ∈ L+ (X), and by the property 8) of integral of nonnegative functions we
have µX(f + = +∞) = 0 and µX(f − = +∞) = 0, so
[
µX(f = ±∞) = µX(f + = +∞) µX(f − = +∞) = 0

due to additivity of the measure µ.

Z
5) If X = X 0 ⊍ X 00 , where X 0 , X 00 ∈ M(X), and the integral f (x)dµ exists, then
X
Z Z Z
f (x)dµ = f (x)dµ + f (x)dµ. (5.3.3)
X X0 X 00

Z Z Z
If the integral f (x)dµ exists, then one of the integrals +
f (x)dµ or f − (x)dµ is finite. With-
X ZX X
+
out loss of generality we can suppose that the integral f (x)dµ is finite. Then both integrals
Z Z X
+ +
f (x)dµ and f (x)dµ are finite, so both integrals in the right-hand side of (5.3.3) exist. The
X0 X 00
identity (5.3.3) now follows from Definition 5.3.1 and from the property 4) of integral of nonnegative
functions.

Remark 5.3.4. Z Existence of both integrals in the right-hand side of (5.3.3) does not imply the existence
of the integral f (x)dµ.
X

Example 5.3.5. Let (

+∞, x ∈ X 0,
f (x) =
−∞, x ∈ X 00 ,
5.3. INTEGRAL OF MEASURABLE FUNCTIONS 83

and µX 0 > 0, µX 00 > 0. Then

Z Z
f (x)dµ = +∞, f (x)dµ = −∞,
X0 X 00
Z
but the integral f (x)dµ does not exists.
X

Corollary 5.3.6. If f ∈ L(X) and X = X 0 ⊍ X 00 , where X 0 , X 00 ∈ M(X), then f ∈ L(X 0 ) and f ∈ L(X 00 ).
Conversely, if f ∈ L(X 0 ) and f ∈ L(X 00 ), then f ∈ L(X). In both cases the identity (5.3.3) holds.
Z Z
6) If the integral f (x)dµ exists and α ∈ R, then the integral αf (x)dµ exists, and
X X
Z Z
αf (x)dµ = α f (x)dµ.
X X

Z
If the integral f (x)dµ exists, then in the right-hand side of the identity
X
Z Z Z
f (x)dµ = f + (x)dµ − f − (x)dµ.
X X X
Z
one of integrals is finite. Without loss of generality, suppose that f + (x)dµ is finite. Furthermore,
X
we have (
αf + (x) − αf − (x), α > 0,
αf (x) =
|α|f − (x) − |α|f + (x), α < 0.
Therefore, by Definition 5.3.1,
 Z Z



+
αf (x)dµ − αf − (x)dµ, α > 0,


X X
Z 

αf (x)dµ = Z Z (5.3.4)

−

X |α|f (x)dµ − |α|f + (x)dµ, α < 0.





X X

By the property 2) of integral of nonnegative functions, in (5.3.4), the integrals with f + are finite,
so the integral in the left-hand side of (5.3.4) exists. Moreover, by the same property,
 Z Z


 α f +
(x)dµ − α f − (x)dµ, α > 0,


X X
Z 

αf (x)dµ = Z Z =

|α| f − (x)dµ − |α| f + (x)dµ, α < 0.

X 




X X
 
Z Z Z
= α  f + (x)dµ − f − (x)dµ = α f (x)dµ.
X X X
84 CHAPTER 5. INTEGRATION THEORY

Corollary 5.3.7. If f ∈ L(X), then αf ∈ L(X), α ∈ R.

Z Z
7) If the integrals f (x)dµ and g(x)dµ exist, and if at least one of them is finite, then the integral
Z X X

(f (x) + g(x))dµ exists, and

X
Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ. (5.3.5)
X X X

Z Z
Without loss of generality, suppose that the integral f (x)dµ is finite. Then the integrals f + (x)dµ
Z Z X Z X
− + −
and f (x)dµ, and one of the integrals g (x)dµ and g (x)dµ are also finite. Without loss
X Z X X

of generality, suppose that the integral g + (x)dµ. By the property 4) of integral of nonnegative
X
functions, integrals of f − and +
Z f over any measurable
Z subset of X are finite, as well. We must prove
that one of the integrals (f (x) + g(x))+ dµ or (f (x) + g(x))− dµ is finite, and then prove the
X X
identity (5.3.5).
We split the set X into the following six subsets

X1 = X(f > 0, g > 0), X4 = X(f < 0, g < 0),

X2 = X(f > 0, g < 0, f + g > 0), X5 = X(f > 0, g < 0, f + g < 0),

X3 = X(f < 0, g > 0, f + g > 0), X5 = X(f < 0, g > 0, f + g < 0).
6
It is clear that X = ⊍ Xj and
j=1


3
 f (x) + g(x),

 X = ⊍ Xj ,
j=1
(f + g)+ = 6


 0, X = ⊍ Xj ,
j=4


3


 0, X = ⊍ Xj ,
j=1
(f + g)− = 6
 −(f (x) + g(x)),

 X = ⊍ Xj .
j=4

On the set X1 the functions f and g are nonnegative. Therefore, by the property 3) of integral of
nonnegative functions, one has
Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ. (5.3.6)
X1 X1 X1

Moreover, all the integrals here are finite by assumption.

5.3. INTEGRAL OF MEASURABLE FUNCTIONS 85
Z
On the set X2 , we have −g(x) 6 f (x), and the integral f (x)dµ is finite by assumption. Therefore,
Z Z X2

the integral (−g(x))dµ is also finite, so g(x)dµ is finite by the property 6). Furthermore, in the
X2 X2
identity
f (x) = (f (x) + g(x)) + (−g(x))
both functions in the right-hand side are nonnegative on X2 , so by the property 3) of integral of
nonnegative functions and by the property 6) we get
Z Z Z
f (x)dµ = (f (x) + g(x))dµ − g(x)dµ,
X2 X2 X2

or Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ. (5.3.7)
X2 X2 X2

All the integrals here are finite by assumption.

On the set X3 , we have g(x) = (f (x) + g(x)) + (−f (x)) where both functions in the right-hand side
are nonnegative, so as above
Z Z Z
g(x)dµ = (f (x) + g(x))dµ − f (x)dµ,
X3 X3 X3

or Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ. (5.3.8)
X3 X3 X3

Moreover, all the integrals in (5.3.8) are finite by assumption.

According to the property 4) of integral of nonnegative functions, we now obtain
Z Z Z 3 Z
X
+ +
(f (x) + g(x)) dµ = (f (x) + g(x)) dµ + 0 · dµ = (f (x) + g(x))dµ. (5.3.9)
X 3 6
j=1 X
j
⊍ Xj ⊍ Xj
j=1 j=4

On the set X4 , we have −(f (x) + g(x)) = (−f (x)) + (−g(x)) where both functions in the right-hand
side are nonnegative, so by the property 3) of integral of nonnegative functions and by the property 6)
we get Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ. (5.3.10)
X4 X4 X4
Z Z
By assumption the integral f (x)dµ is finite, but g(x)dµ can be equal to −∞, so the integral in
X4 X4
the left-hand side of the identity (5.3.10) is either finite, or equal to −∞.
Analogously, for the sets X5 and X6 we obtain
Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ (5.3.11)
X5 X5 X5
Z Z Z
(f (x) + g(x))dµ = f (x)dµ + g(x)dµ (5.3.12)
X6 X6 X6
86 CHAPTER 5. INTEGRATION THEORY

Moreover, the integral in the left-hand side of (5.3.12) is finite by assumption, while the integral in
the left-hand side of (5.3.11) can equal −∞.
Now by the property 4) of integral of nonnegative functions and by the property 6) we have
Z Z Z 6 Z
X
(f (x) + g(x))− dµ = 0 · dµ + [−(f (x) + g(x))] dµ = − (f (x) + g(x))dµ, (5.3.13)
X 3 6
j=4 X
j
⊍ Xj ⊍ Xj
j=1 j=4

where the integral on the left-hand side can be either finite or equal to +∞.
Z
Thus, we proved that the integral (f (x) + g(x))dµ exists. Now from the property 5) and from the
X
identities (5.3.6)–(5.3.13) one obtains
Z Z Z 6 Z
X
(f (x) + g(x))dµ = (f (x) + g(x))+ dµ − (f (x) + g(x))− dµ = (f (x) + g(x))dµ =
X X X j=1X
j

 
6
X Z Z Z Z
= f (x)dµ + g(x)dµ = f (x)dµ + g(x)dµ.
 
 
j=1 Xj Xj X X

Other situations (when integral of f is infinite but the integral of g is finite and so on) can be proved
analogously.

Remark 5.3.8. In the conditions of the property 7) there can appear such a situation when the sum
f (x) + g(x) is not defined at some points of the set (for instance, when f (x) = +∞ but g(x) = −∞).
However, by the property 4) the set of such points is a null set (for example, in the proof of the property 7)
we had µX(f = ±∞) = 0), therefore, by the property 3), we can change the function whose integral is
finite to another function which is equal to the function we change a.e. on X and is everywhere finite on X
(leaving the same notation for the Znew function). After that the sum f (x) + g(x) is defined everywhere
on X but the value of the integral (f (x) + g(x))dµ remains unchanged.
X

Corollary 5.3.9. If f, g ∈ L(X), then f + g ∈ L(X).

8) If f (x) 6 g(x) a.e. on X, then Z Z

f (x)dµ 6 g(x)dµ, (5.3.14)
X X

provided both integrals exist.

Without loss of generality we can assume that f (x) 6 g(x) everywhere on X. Otherwise, we redefine
either one or both functions on the null-set where the inequality f (x) 6 g(x) fails. According to the
property 3), this operation does not affect existence and values of the corresponding integrals. It is
clear that
f + (x) 6 g + (x), f − (x) > g − (x), x ∈ X.

By the property 5) of integral of nonnegative functions, we have

Z Z Z Z
+ +
f (x)dµ 6 g (x)dµ, f (x)dµ > g − (x)dµ.
−

X X X X

Now we subtract the second inequality from the first one to obtain (5.3.14).
5.3. INTEGRAL OF MEASURABLE FUNCTIONS 87

9) If |f (x)| 6 g(x) a.e. on X, and g ∈ L(X), Then f ∈ L(X), and

Z Z
f (x)dµ 6 g(x)dµ (5.3.15)
X X

Without loss of generality we can assume that |f (x)| 6 g(x) everywhere on X. Otherwise, we redefine
either one or both functions on the null-set where the inequality |f (x)| 6 g(x) fails. According to the
property 3), this operation does not affect existence and values of the corresponding integrals. The
inequality (5.3.15) follows from the property 5) of nonnegative functions and from the property 2).
Corollary 5.3.10. If a function f is measurable and a.e. bounded on X, and µX < +∞, then f ∈ L(X).
Moreover, if m 6 f (x) 6 M a.e. on X, then
Z
m · µX 6 f (x)dµ 6 M · µX. (5.3.16)
X

Proof. Since |f (x)| 6 max{|m|, M } =: C a.e. on X, and the simple function h(x) ≡ C is integrable on X,
the function f is integrable on X, as well. To prove (5.3.16), consider the simple functions h1 (x) ≡ m and
h2 (x) ≡ M on X. We have h1 (x) 6 f (x) 6 h2 (x) a.e. on X. Now (5.3.16) follows from (5.3.14).
Remark 5.3.11. By Theorem 4.4.4 every a.e. bounded measurable functions is Lebesgue integrable. So
every a.e. finite measurable function is integrable out of a set of arbitrary small measure.
10) (Absolute continuity of integral) If f ∈ L(X), then for any ε > 0, there exists δ > 0 such that
Z
f (x)dµ < ε
Xδ

for any subset Xδ ⊂ X with µXδ < δ.

Due to the property 2), it is enough to prove the absolute continuity of integral of nonnegative
functions. So, suppose that f ∈ L+ (X), and fix ε > 0. By Definition 5.2.1 of integral of nonnegative
functions, there exists a nonnegative simple function h such that h(x) 6 f (x) on X and
Z Z Z
ε
f (x)dµ − h(x)dµ = (f (x) − h(x))dµ <
2
X X X

m
P ε
Let h(x) = ck χEk (x), and M = max{ck : k = 1, 2, . . . , m}. Take the number δ :=
, and let
k=1 2M
Xδ ⊂ X be such that µXδ < δ. Put hδ ≡ M on Xδ , then h(x) 6 hδ (x) for all x ∈ Xδ , so by the
property 4) of integral of simple functions, we have
Z Z
ε ε
h(x)dµ 6 hδ (x)dµ = M · µXδ = M · = .
2M 2
Xδ Xδ

Thus, we obtain
Z Z Z Z Z Z
f (x)dµ = f (x)dµ − h(x)dµ + h(x)dµ = (f (x)dµ − h(x))dµ + h(x)dµ 6
Xδ Xδ Xδ Xδ Xδ Xδ

Z Z
ε ε
6 (f (x) − h(x))dµ + h(x)dµ < + = ε,
2 2
X Xδ

as required.
88 CHAPTER 5. INTEGRATION THEORY

∞
11) (Complete additivity of integral) Let f ∈ L(X) and the set X can be represented as X = ⊍ Xk ,
k=1
where all the sets Xk are measurable. Then
Z ∞ Z
X
f (x)dµ = f (x)dµ.
X k=1X
k

Suppose first that f ∈ L+ (X) and set

(
f (x), x ∈ Xk ,
fk (x) =
0, x ∈ X \ Xk .
All the nonnegative functions fk are measurable on X by the property 8) of measurable functions,
and
∞
X
f (x) = fk (x), x ∈ X.
k=1

Now by Levi’s Theorem 5.2.3, we have

Z ∞ Z
X ∞ Z
X
f (x)dµ = fk (x)dµ = f (x)dµ.
X k=1 X k=1X
k

If f is an arbitrary integrable function on X, then f = f + − f − , and the functions f + and f − are

integrable and nonnegative on X. Therefore,
Z X∞ Z Z X∞ Z
f + (x)dµ = f + (x)dµ, f − (x)dµ = f − (x)dµ. (5.3.17)
X k=1X X k=1X
k k

Since both series in the right-hand side of the identities (5.3.17) converge, we obtain
 
Z Z Z X∞ Z Z X∞ Z
+ − + −
f (x)dµ = f (x)dµ − f (x)dµ =  f (x)dµ − f (x)dµ =  f (x)dµ.
X X X k=1 Xk Xk k=1X
k

∞
Proposition 5.3.12. Let X = ⊍ Ek , Ek ∈ M(X), k ∈ N, and
k=1
∞
X
f (x) = an χEn (x).
n=1

The function f ∈ L(X) if, and only if

∞
X
|an |µEn < +∞
n=1
Proof. It is sufficient to prove the proposition for the case when f > 0 on X. So ak > 0, k ∈ N. Consider
n
ak χEk (x). Clearly, the sequence (hn )∞
P
the simple functions hn (x) = n=1 is non-decreasing and
k=1

lim hn (x) = f (x).

n→+∞

Now by Definition 5.2.1, we have

Z Z n Z
X n
X ∞
X
f (x)dµ = lim hn (x)dµ = lim ak χEk (x)dµ = lim ak µEk = ak µEk .
n→+∞ n→+∞ n→+∞
X X k=1 X k=1 k=1

∞
Thus, f ∈ L+ (X) if, and only if, the series
P
ak µEk converges.
k=1
5.3. INTEGRAL OF MEASURABLE FUNCTIONS 89

Theorem 5.3.13 (Lebesgue’s Dominated Convergence Theorem). Let a sequence (fn )∞ n=1 of integrable
functions on X converge to a function f almost everywhere on X. If the inequality |fn (x)| 6 g(x), n ∈ N,
holds almost everywhere on X, and g(x) ∈ L(X), then f ∈ L(X), and
Z Z
f (x)dµ = lim fn (x)dµ. (5.3.18)
n→+∞
X X

Proof. From the conditions of the theorem it follows that |f (x)| 6 g(x) a.e. on X, so f ∈ L(X) by the
property 9). If necessary, we redefine the functions fn and f on sets of measure zero, so that fn converges
to f everywhere on X (this operation does not affect the value of the integrals of these functions), and the
inequality |fn (x)| 6 g(x) holds everywhere on X.
Applying Fatou’s Theorem 5.2.7 to the sequence (g + fn )∞ n=1 of nonnegative functions on X, we get
Z Z
lim (g(x) + fn (x))dµ 6 lim (g(x) + fn (x))dµ. (5.3.19)
X X

Now since fn (x) −→ f (x) on X, we have

n→+∞

lim(g(x) + fn (x)) = lim(g(x) + fn (x)) = g(x) + lim fn (x) = g(x) + f (x),

so the left-hand side of (5.3.19) has the form

Z Z
g(x)dµ + f (x)dµ. (5.3.20)
X X

At the same time, the right-hand side of (5.3.19) have the form
Z Z
g(x)dµ + lim fn (x)dµ. (5.3.21)
X X

From (5.3.19)–(5.3.21) we obtain Z Z

f (x)dµ 6 lim fn (x)dµ. (5.3.22)
X X

Now let us consider the sequence (g − fn )∞

n=1 of nonnegative functions on X. In the same way as above,
one can prove that  
Z Z
− f (x)dµ 6 lim − fn (x)dµ ,
X X
or Z Z
f (x)dµ > lim fn (x)dµ. (5.3.23)
X X
R
From (5.3.22) and (5.3.23) it follows that lim fn (x)dµ exists and the identity (5.3.18) holds.
n→+∞
X

5.3.1 Invariance Properties

Definition 5.3.14. Let f be a function defined on Rn . The function fc , defined as fc (x) = f (x − c) is
called the translation of f by a vector c ∈ Rn .

The following theorem is true.

90 CHAPTER 5. INTEGRATION THEORY

Theorem 5.3.15. Let f ∈ L(Rn ). Then fc ∈ L(Rn ), and

Z Z
f (x − c)dµ = f (x)dµ. (5.3.24)
Rn Rn

Proof. Suppose first that f (x) = χA (x), where A is a measurable set. Then obviously fc (x) = χAc , where
Ac = {x + c : x ∈ A}. In this case, the assertion of the theorem holds, since µA = µAc as we proved in
Section 3.8, and Z Z
f (x − c)dµ = µAc = µA = f (x)dµ.
Rn Rn

As a result of linearity, the identity (5.3.24) holds for all simple functions.
Now if f (x) is non-negative and (hn (x))∞ n=1 is a sequence of simple functions that increase pointwise
a.e to f (such a sequence exists by Theorem 4.4.2), then (hn (x − c))∞ n=1 is a sequence of simple functions
that increase to fc (x) pointwise a.e, and Corollary 5.2.5 implies (5.3.24) in this special case. In general
case, we represent the function as f = f + − f − where the nonnegative functions f + and f − are defined in
Definition 4.2.3. Since the assertion of the theorem holds for f + and f − , it holds for the function f .

In the same way one can prove the following theorem.

Theorem 5.3.16. Let f ∈ L(Rn ). Then f (ax) ∈ L(Rn ) for a ∈ R \ {0}, and
Z Z
f (ax)dµ = |a|−n f (x)dµ. (5.3.25)
Rn Rn

5.4 Difference between Riemann and Lebesgue definite integrals

One can compare the Riemann and Lebesgue integrals only if both are defined, that is, if X is rectifiable
set in Rn and µ is the Lebesgue measure. For the sake of simplicity, let us consider the case X = [a, b]. In
Zb Z
what follow, we denote Lebesgue integral as (L) f (x)dx instead of f (x)dµ, omitting the symbol (L)
a [a,b]
Zb
whenever it is clear that we use Lebesgue integration. As well, we denote Riemann integral as (R) f (x)dx.
a
To define Riemann integral, one needs to divide the integration domain into l parts Ik . Then, in each
Pl
part we choose a point and consider the integral sum σ = f (ξk )|Ik |. The Riemann integral exists, that
k=1
is, the integral sums converge if the arbitrariness of choosing points ξk does not affect the value of the
integral sum, that is, if the integrand f is not “too discontinuous”.
Lebesgue’s approach to integration is different in essence. Let us state it briefly. Suppose that f be a
bounded Lebesgue measurable function on [a, b], and set

m := inf{f (x) : x ∈ [a, b]}, M := sup{f (x) : x ∈ [a, b]}.

Split the interval [m, M ] into several parts by points m = y0 < y1 < y2 < . . . < yl = M , and put
Xk := {x ∈ [a, b] : yk−1 < f (x) < yk }, k = 1, . . . , l. Then we choose points ξk ∈ Xk and construct the
l
P
following integral sum σ = f (ξk )µXk .
k=1
l
P l
P
Furthermore, one introduces the lower and upper integral sums, s = yk−1 ·µXk and S = yk ·µXk
k=1 k=1
whose properties are similar to the ones of the lower and upper Darboux sums. It can be proved that for
5.4. DIFFERENCE BETWEEN RIEMANN AND LEBESGUE DEFINITE INTEGRALS 91

any bounded Lebesgue measurable function on [a, b], the following identity holds

Zb
lim s = lim S = (L) f (x)dx.
l→+∞ l→+∞
a

Thus, any bounded Lebesgue measurable function is Lebesgue integrable.

In fact, Definition 4.2 of integral of nonnegative functions defines Lebesgue integral as the limit of lower
integral sums.
Thus, in Lebesgue’s approach we divide the integration domain into parts where the values of the
integrand are close to each other, while in Riemann integration the integration domain is split into small
parts regularly. So Lebesgue’s approach allows us to substantially extend the class of integrable function.
For example, Dirichlet’s function ( T
1, x ∈ Q [a, b],
D(x) =
0, x ∈ [a, b] \ Q,
is non-Riemann integrable (see Example 5.4.8 below) but it is a simple function from Lebesgue’s point of
view, so
Zb  \ 
(L) D(x)dx = 1 · µ Q [a, b] + 0 · µ ([a, b] \ Q) = 1 · 0 + 0 · 1 = 0.
a

The following theorem shows that the class of Lebesgue integrable functions contains the class of
Riemann integrable functions.

Theorem 5.4.1. If a function is Riemann integrable on [a, b], then it is Lebesgue integrable on [a, b], and

Zb Zb
(L) f (x)dx = (R) f (x)dx.
a a

(n) (n) (n) (n)

Proof. Let (Pn )n∈N be a partition of the interval [a, b] such that a = x0 < x1 < x2 < . . . < xkn = b.
Moreover, let the partition Pn+1 be a refinement of Pn for any n ∈ N, and ∆n → 0 as n → +∞, where
(n) (n) (n) (n) (n)
∆n = max{|xj − xj−1 | : j = 1, . . . , kn } is the norm of the partition Pn . Let us set Ii := [xi−1 , xi ],
(n) (n) (n)
|Ii | = xi − xi−1 , and

(n) (n) (n) (n)

mi = inf{f (x) : x ∈ Ii }, Mi = sup{f (x) : x ∈ Ii },

where i = 1, . . . , kn , n ∈ N. Introduce the lower and the upper Darboux sums

kn kn
(n) (n) (n) (n)
X X
sn = mi |Ii |, Sn = Mi |Ii | (5.4.1)
i=1 i=1

The Riemann integrability of the function f means that

Z b
lim sn = lim Sn = (R) f (x)dx. (5.4.2)
n→+∞ n→+∞ a

Consider two sequences of simple functions

 
(n) (n) (n)
hn (x) = mi , x ∈ xi−1 , xi , i = 1, . . . , kn , n ∈ N,
 
(n) (n) (n)
κn (x) = Mi , x ∈ xi−1 , xi , i = 1, . . . , kn , n ∈ N.
92 CHAPTER 5. INTEGRATION THEORY

(n)
At the points xi , these functions can be defined arbitrarily. Indeed, for each n we have finitely many
such points, and for all n ∈ N we get countably many points where we define the functions hn and κn
arbitrarily. But any countable set is a null-set, so it is not important for constructing the Lebesgue integral
(n)
of the function f how we define the functions hn and κn at the points xi .
The functions hn and κn are measurable on [a, b]. Moreover, since each partition Pn is a refinement
of the previous partition Pn−1 , the functions hn cannot decrease (a.e. on [a, b]), while the functions κn
cannot increase (a.e. on [a, b]). Consequently, the following limits exist a.e. on [a, b]

g1 (x) = lim hn (x), g2 (x) = lim κn (x),

n→+∞ n→+∞

where the functions g1 (x) and g2 (x) are measurable on [a, b] by Theorem 4.3.6.
Since the following holds

hn (x) 6 f (x) 6 κn (x), a.e. on [a, b],

in the limit case we have

g1 (x) 6 f (x) 6 g2 (x), a.e. on [a, b]. (5.4.3)
Consider now the sequence (κn − hn )∞
This sequence converges to g2 (x) − g1 (x) a.e. on [a, b],
n=1 .
and is majorized by the integrable function κ1 − h1 . Consequently, we can apply Lebesgue’s Dominated
Convergence Theorem 5.3.13 to the sequence (κn − hn )∞ n=1 :

Zb Zb
(g2 (x) − g1 (x))dx = lim (κn (x) − hn (x))dx =
n→+∞
a a

Zb Zb
= lim κn (x)dx − lim hn (x)dx = lim Sn − lim sn = 0.
n→+∞ n→+∞ n→+∞ n→+∞
a a

The function g2 (x)−g1 (x) > 0 a.e. on [a, b], so from the property 8) of integral of nonnegative functions
a.e.
it follows that g2 (x) − g1 (x) = 0 a.e. on [a, b]. Now the inequalities (5.4.3) imply f = g1 on [a, b]. Thus,
f is measurable on [a, b] by Lemma 4.3.3.
Since f is bounded and measurable on [a, b], it is Lebesgue integrable on [a, b] by Corollary 5.3.10, so
according to Corollary 5.2.5 we obtain

Zb Zb Zb Z b
(L) f (x)dx = (L) g1 (x)dx = lim hn (x)dx = lim sn = (R) f (x)ds,
n→+∞ n→+∞ a
a a a

as required.

5.4.1 Discontinuities of Riemann integrable functions

It is very well known that continuous, piecewise continuous, and monotone functions are Riemann inte-
grable (see e.g. [9, Chapter 6]). In this section we completely describe the class of Riemann integrable
functions in terms of their discontinuities. To do this, let us introduce the following objects.
Let J := [a, b] and let I(c, r) := (c − r, c + r) be the open interval centered at c of radius r > 0. Suppose
that f : [a, b] 7→ R is a bounded function.
Definition 5.4.2. The oscillation of f on I(c, r) is the following value

osc(f, c, r) = sup |f (x) − f (y)|

T
where the supremum is taken over all x, y ∈ J I(c, r). This quantity exists since f is bounded.
5.4. DIFFERENCE BETWEEN RIEMANN AND LEBESGUE DEFINITE INTEGRALS 93

Definition 5.4.3. The oscillation of f at c is the value

osc(f, c) = lim osc(f, c, r).

r→0

This limit exists because osc(f, c, r) > 0 is a non-strictly decreasing function of r.

It is clear from Definitions 5.4.2 and 5.4.3 that f is continuous at c if, and only if, osc(f, c) = 0 (This
is a Baire theorem, see [8, Ch. V, §4]). For each ε > 0 we define a set Aε by

Aε = {c ∈ J : osc(f, c) > ε}.

S
It is easy to see that the set of points in J where f is discontinuous is Aε .
ε>0

Lemma 5.4.4. If ε > 0, then the set Aε is closed (and therefore compact).
Proof. Suppose that cn ∈ Aε converges to c and assume that c 6∈ Aε , say, osc(f, c) = ε − δ where δ > 0.
δ r T  r
Select r so that osc(f, c, r) < ε − , and choose n with |cn − c| < . So if x, y ∈ J I cn , , then
T  2 r 2 2
x, y ∈ J I (c, r), therefore, osc f, cn , < ε which implies osc(f, cn ) < ε, a contradiction.
2
Now we are in s position to describe the class of Riemann integrable functions in terms of their
discontinuities.
Theorem 5.4.5 (Lebesgue). A bounded function f on [a, b] is Riemann integrable if, and only if, it is
continuous almost everywhere on [a, b].
Proof. By assumption, there exists a number M > 0 such that |f (x)| 6 M on [a, b].
Suppose that the set D of discontinuities of f has Lebesgue measure 0, and let ε > 0. Since Aε ⊂ D,
we have µ(Aε ) = 0, since Lebesgue measure µ is complete by Theorem 3.4.13. The set Aε is measurable,
and its Lebesgue measure coincides with its outer
S measure. So by Definition 3.3.1, given ε > 0 there exists
a cover of Aε by intervals (bricks in R), Aε ⊂ Ij , such that
j
X ε ε
|Ij | 6 µ∗ (Aε ) + = .
j
2 2

Moreover, we can find open intervals Iej ⊃ Ij such that

ε
|Iej | < |Ij | + ,
2j+1
so we have X X ε ε ε
|Iej | < |Ij | + 6 + = ε.
j j
2 2 2

Se N
S
The cover Ij ⊃ Aε of the closed set Aε by open sets Iej contains a finite subcover Aε ⊂ Iej =: I such
j j=1
that
N
X
|I| = |Iej | < ε.
j=1

The complement CI = [a, b] \ I of I is compact, and around each point z in this complement we can S find
an interval Fz with sup |f (x) − f (y)| < ε, since z 6∈ Aε . We may now choose a finite subcover of Fz ,
x,y∈Fz z∈CI
which we denote by IN +1 , ..., IN 0 . Now, taking all the end points of the intervals I1 , I2 , ..., IN 0 we obtain
a partition P of [a, b] with
XN
SN 0 − sN 0 6 2M |Ij | + ε(b − a) 6 Cε,
j=1
94 CHAPTER 5. INTEGRATION THEORY

where the Darboux sums sN 0 and SN 0 are defined in (5.4.1). Hence f is integrable on [a, b] by (5.4.2), as
required.
Conversely, suppose that f is integrable on [a, b], and let D be its set of discontinuities. Since D equals
∞
S
A n1 , it suffices to prove that each A n1 has measure 0 according to Corollary 3.4.11. Let ε > 0 and
n=1
ε
choose a partition P = {x0 , x1 , ..., xN } so that SN − sN < . Then, if A n1 intersects Ij = (xj−1 , xj ) we
n
1
must have sup f (x) − inf f (x) > , and this shows that
x∈Ij x∈Ij n

1 X X ε
|Ij| 6 [ sup f (x) − inf f (x)]|Ij| 6 SN − sN < .
n T T x∈Ij x∈Ij n
{j: Ij A 1 6=∅} {j: Ij A 1 6=∅}
n n

So by taking intervals intersecting A n1 and making them slightly larger, we can cover A n1 with open
intervals of total length less than 2ε. Therefore, A n1 has measure 0, and we are done.

Corollary 5.4.6. Continuous, piece-wise continuous, and monotone bounded functions on [a, b] are Rie-
mann integrable on [a, b].
Proof. Indeed, the set of discontinuity is empty for continuous functions, finite for piece-wise continuous
functions, and at most countable for monotone functions (see Homework), so its Lebesgue measure is 0
for all these classes of functions.
Example 5.4.7. Let us consider the Dirichlet function
( T
1, x ∈ Q [0, 1],
D(x) =
0, x ∈ [0, 1] \ Q.

The function D(x) is discontinuous

T at any point of [0, 1], since for any 0 < ε < 1, δ > 0, and for any
x0 ∈ [0, 1] there exist x ∈ [0, 1] (x0 − δ, x0 + δ) such that |D(x) − D(x0 )| = 1 > ε. Thus, D(x) 6∈ R[0, 1]
by Theorem 5.4.5, since its set of discontinuities is the whole interval [0, 1] whose Lebesgue measure is
positive (equals 1). T
However, the set X0 := Q [0, 1] is a subset of the countable set Q. Since µQ = 0 and the Lebesgue
a.e.
measure is complete, we have µX0 = 0. Therefore, D = 0 on [0, 1], and the function g(x) ≡ 0 is measurable
and integrable on [0, 1]. Thus, D ∈ L+ [0, 1] and by property 3) of integral of measurable functions
Z Z
D(x)dµ = 0 · dµ = 0.
[0,1] [0,1]

Example 5.4.8. Consider the Riemann function



 0, x ∈ [0, 1] \ Q,

f (x) =
 1 , x = m where m and n are coprime, m < n.


n n
T
The set of discontinuities of the function f is X0 := Q [0, 1]. Indeed, if c ∈ X0 , then there exists a
1 1
number n such that f (c) = . Let us take an arbitrary number 0 < ε0 < . Then ∀δ > 0 there exists
T n n
x ∈ [0, 1] (c − δ, c + δ) such that f (x) = 0, so we have
1
|f (x) − f (c)| = > ε0 .
n
Consequently, X0 is a subset of the set of discontinuities of the function f .
5.4. DIFFERENCE BETWEEN RIEMANN AND LEBESGUE DEFINITE INTEGRALS 95

Let us show that f is continuous on [0, 1] \ X0 . In fact, suppose that a ∈ [0, 1] \ X0 . For any ε > 0,
1
there exist only finitely many positive integers not exceeding . So there are only finitely many rational
ε
numbers in X0 such that f (m/n) > ε. One can choose δ > 0 so small that the interval (a − δ, a + δ) does
not contain these rational numbers, we have

|f (x) − f (a)| < ε

T
for any x ∈ (a − δ, a + δ) [0, 1]. Thus, f is continuous on [0, 1] \ X0 .
a.e.
The set X0 is dense in [0, 1] but µX0 = 0, so f ∈ R[0, 1] by Theorem 5.4.5. Moreover, f = 0 on [0, 1],
so Z Z
f (x)dµ = 0 · dµ = 0.
[0,1] [0,1]

Example 5.4.9. Let X = [0, 1]. For n ∈ N, consider the sets

        
1 1 1 1 1 1 1 1 1 1
An = , , Bn = , + , Cn = + , ,
n+1 n n+1 2 n+1 n 2 n+1 n n

and construct the function

∞
X
f (x) = (n + 1) [χCn (x) − χBn (x)] .
n=1

Note that An = Bn ⊍ Cn , n ∈ N, f ∈ L(An ), and

Z
f (x)dµ = (n + 1)[µCn − µBn ] = 0,
An

1
since µBn = µCn = . However, f 6∈ L[0, 1] by Proposition 5.3.12, since
2n(n + 1)
∞ ∞ ∞
X X n+1 X 1
(n + 1)[µBn + µCn ] = = = +∞.
n=1 n=1
n(n + 1) n=1
n

1 1
Consider now a real number a ∈ (0, 1). There exists a natural number n ∈ N such that 6a< .
n+1 n
Consequently, on the interval [a, 1] the function f is bounded and has finitely many points of discontinuities.
Thus, by Theorem 5.4.5, the function f is Riemann integrable on [a, 1], and by Theorem 5.4.1

Z1 Z n−1
X Z Z
(R) f (x)dx = (L) f (x)dµ = f (x)dµ + f (x)dµ =
a k=1 1 1  1
(a,1)
 
k+1 , k a, n

Z  
1 1 1
= f (x)dµ 6 (n + 1) − = .
n n+1 n
1
 
a, n

Clearly, a → +0 if, and only if, n → +∞, so

Z1
1
lim (R) f (x)dx 6 lim = 0,
a→+0 n→+∞ n
a
96 CHAPTER 5. INTEGRATION THEORY

thus, f ∈ R(+0, 1], and

Z1
(R) f (x)dx = 0.
+0

This example show that even if a function is non-Lebesgue integrable, there can exist an improper
Riemann integral.

5.4.2 Approximation of Riemann integrable functions

Theorem 5.4.10. Suppose f is Riemann integrable on [a, b], and is bounded by M . Then there exists a
sequence (fk )∞
k=1 of continuous functions on [a, b] so that

sup |fk (x)| 6 M, k = 1, 2, . . . ,

x∈[a,b]

and Z b
|f (x) − fk (x)|dx → 0 as k → ∞.
a

Proof. Given ε > 0, we may choose a partition a = x0 < x1 < · · · < xN = b of the interval [a, b] so that
the upper and lower Darboux sums of f differ by at most ε. Denote by f ∗ the step function defined by

f ∗ (x) = sup f (y) if x ∈ [xj−1 , xj ) for 1 6 j 6 N.

xj−1 6y6xj

By construction we have |f ∗ | 6 M , and moreover

Zb Zb
∗
|f (x) − f (x)|dx = (f ∗ (x) − f (x))dx < ε. (5.4.4)
a a

Now we can modify f ∗ to make it continuous and still approximate f in the sense of the lemma. For small
δ > 0, let fe(x) = f ∗ (x) when the distance of x from any of the division points x0 , . . . , xN exceeds δ. In
the δ-neighborhood of xj for j = 1, . . . , N − 1, define fe(x) to be the linear function for which fe(xj ± δ) =
f ∗ (xj ± δ). Near x0 = a, fe is linear with fe(a) = 0 and fe(a + δ) = f ∗ (a + δ). Similarly, near xN = b the
function fe is linear with fe(b) = 0 and fe(b − δ) = f ∗ (b − δ). The absolute value of this extension is also
bounded by M . Moreover, fe differs from f ∗ only in the N intervals of length 2δ surrounding the division
points. Thus
Zb
|f ∗ (x) − fe(x)|dx 6 2M N · 2δ.
a
If we choose δ sufficiently small, we get
Zb
|f ∗ (x) − fe(x)|dx < ε. (5.4.5)
a

As a result, equations (5.4.4)–(5.4.5), and the triangle inequality yield

Zb
|fe(x) − f (x)|dx < 2ε.
a

1
Denoting by fk the fe so constructed, when 2ε = , we see that the sequence (fk )∞
k=1 has the properties
k
required by the lemma.
5.5. FUBINI’S THEOREM AND ITS APPLICATIONS 97

Remark 5.4.11. We construct fe(x) such that fe(a) = fe(b), so we may extend fe to a continuous and
(b − a)- periodic function on R. If the function f satisfies the condition f (a) = f (b), we also can consider it
to be (b − a)-periodic on the real line. Thus, our approximating function fe(x) can be used to approximate
periodic functions.

5.4.3 Approximation of Lebesgue integrable functions

To be written soon...

5.5 Fubini’s theorem and its applications

In elementary calculus integrals of continuous functions of several variables are often calculated by iterating
one-dimensional integrals. We shall now examine this important analytic device from the general point of
view of Lebesgue integration in Rn , and we shall see that a number of interesting issues arise. In general,
we can represent Rn as a product Rn = Rn1 × Rn2 where n = n1 + n2 , and n1 , n2 > 1. A point in Rn then
takes the form (x, y), where x ∈ Rn1 and y ∈ Rn2 . With such a decomposition of Rn in mind, the general
notion of a slice, formed by fixing one variable, becomes natural.
Definition 5.5.1. Let f be a function defined on Rn1 × Rn2 . The slice of f corresponding to y ∈ Rn2 is
the function f y of the x ∈ Rn1 variable, given by

f y (x) = f (x, y).

Similarly, the slice of f for a fixed x ∈ Rn1 is fx (y) = f (x, y).

In the case of a set A ⊂ Rn1 × Rn2 we define its slices by

Ay = {x ∈ Rn1 : (x, y) ∈ A} and Ax = {y ∈ Rn2 : (x, y) ∈ A}.

Recall that even if f is measurable on Rn , it is not necessarily true that the slice f y is measurable on
n2
R for each y; nor does the corresponding assertion necessarily hold for a measurable set: the slice E y may
not be measurable for each y. An easy example arises in R2 by placing a one-dimensional non-measurable
set on the x-axis; the set A in R2 has measure zero, but E y is not measurable for y = 0. What saves us is
that, nevertheless, measurability holds for almost all slices.

5.5.1 Statement and proof of Fubini’s theorem

Theorem 5.5.2. Suppose f (x, y) is integrable on Rn1 × Rn2 . Then for almost every y ∈ Rn2 :
(i) The slice f y (x) is integrable on Rn1 .
Z
(ii) The function f y (x)µ(dx) is integrable on Rn2 , and
Rn1

(iii) The following identity holds:

Z Z  Z
f (x, y)µ(dx) µ(dy) = f (x, y)dµ.
Rn1
Rn2 Rn

Clearly, the theorem is symmetric inZ x and y so that we also may conclude that the slice fx (y) is
integrable on Rn2 for a.e. x. Moreover, fx (y)µ(dy) is integrable, and
Rn2
Z Z  Z
f (x, y)µ(dy) µ(dx) = f (x, y)dµ.
Rn2
Rn1 Rn
98 CHAPTER 5. INTEGRATION THEORY

In particular, Fubini’s theorem states that the integral of f on Rn can be computed by iterating lower-
dimensional integrals, and that the iterations can be taken in any order
Z Z  Z Z  Z
f (x, y)µ(dx) µ(dy) = f (x, y)µ(dy) µ(dx) = f (x, y)dµ.
Rn1 Rn2
Rn2 Rn1 Rn

The proof of Fubini’s theorem which we give next consists of a sequence of six steps. We begin by
letting F denote the set of integrable functions on Rn which satisfy all three conclusions in the theorem,
and set out to prove that L(Rn ) ⊂ F.
We proceed by first showing that F is closed under operations such as linear combinations (Step 1)
and limits (Step 2). Then we begin to construct families of functions in F. Since any integrable function
is the “limit” of simple functions, and simple functions are themselves linear combinations of sets of
finite measure, the goal quickly becomes to prove that the function χA (x) belongs to F whenever A is a
measurable subset of Rn with finite measure. To achieve this goal, we begin with bricks and work our way
up to sets of type Gδ (Rn ) (Step 3), and sets of measure zero (Step 4). Finally, a limiting argument shows
that all integrable functions are in F. This will complete the proof of Fubini’s theorem.

Proof of Theorem 5.5.2.

1) Any finite linear combination of functions in F also belongs to F.
Indeed, let (fm )N
m=1 ⊂ F. For each m there exists a set Am ⊂ R
n2 y
of measure 0 so that fm is
N
integrable on Rn1 whenever y 6∈ Am . Then, if A =
S
Am , the set A has measure 0, and in the
m=1
complement of A, the y-slice corresponding to any finite linear combination of the fm is measurable,
and also integrable. By linearity of the integral, we then conclude that any linear combination of the
fm ’s belongs to F.
2) Suppose (fm )∞ m=1 is a sequence of measurable functions in F so that fm % f or fm & f , where f
is integrable (on Rn = Rn1 × Rn2 ). Then f ∈ F.
By taking −fm instead of fm if necessary, we note that it suffices to consider the case of an increasing
sequence. Also, we may replace fm by fm − f1 and assume that the fm ’s are non-negative. Now, we
observe that an application of Corollary 5.2.5 yields
Z Z
lim fm (x, y)µ(dx)µ(dy) = f (x, y)µ(dx)µ(dy). (5.5.1)
m→∞
Rn Rn

By assumption, for each m there exists a set Am ⊂ Rn2 , so that fm y

is integrable on Rn1 whenever
∞
Am , then µA = 0 in Rn2 , and if y 6∈ A, then fm
y
is integrable on Rn1 for all m,
S
y 6∈ Am . If A =
m=1
and, by Corollary 5.2.5, we find that
Z Z
y
gm (y) = fm (x)µ(dx) % g(y) = f y (x)µ(dx)
Rn1 Rn1

as m → ∞. By assumption, each gm (y) is integrable, so that another application of Corollary 5.2.5

yields Z Z
gm (y)µ(dy) −→ g(y)µ(dy) as m → ∞. (5.5.2)
Rn2 Rn2

By the assumption that fm ∈ F we have

Z Z
gm (y)µ(dy) = fm (x, y)µ(dx)µ(dy),
Rn2 Rn
5.5. FUBINI’S THEOREM AND ITS APPLICATIONS 99

and combining this fact with (5.5.1) and (5.5.2), we conclude that
Z Z
g(y)µ(dy) = f (x, y)µ(dx)µ(dy).
Rn2 Rn

Since f is integrable, the right-hand integral is finite, and this proves that g is integrable. Con-
sequently g(y) is finite-valued a.e. on Rn2 in variable y, hence f y is integrable for a.e. y ∈ Rn2 ,
and Z Z  Z
f (x, y)µ(dx) µ(dy) = f (x, y)dµ.
Rn1
Rn2 Rn

This proves that f ∈ F as desired.

3) Any characteristic function of a set E that is a Gδ (Rn ) and of finite measure belongs to F.
We proceed in stages of increasing order of generality.
(a) First suppose E is a bounded open cube in Rn , such that X = Q1 × Q2 , where Q1 and Q2 are
open cubes in Rn1 and Rn2 , respectively. Then, for each y the function χE (x, y) is measurable
in x, and integrable with
(
y ∈ Q2 ,
Z
µ(Q1 ),
g(y) = χE (x, y)µ(dx) =
0, y 6∈ Q2 .
Rn1

Consequently, g(y) = µ(Q1 )χQ2 (y) is also measurable and integrable, with
Z
g(y)µ(dy) = µ(Q1 )µ(Q2 ).
Rn2
Z
Since we initially have χE (x, y)µ(dy) = µ(E) = µ(Q1 )µ(Q2 ), we deduce that χE ∈ F.
Rn
(b) Now suppose E is a subset of the boundary
Z of some closed cube. Then, since the boundary of
n
a cube has measure 0 in R , we have χE (x, y)µ(dx)µ(dy) = 0.
Rn
Next, we note, after an investigation of the various possibilities,
Z that for almost every y, the
y n1
slice E has measure 0 in R , and therefore if g(y) = χE (x, y)µ(dx) we have g(y) = 0 for
Z Rn1

a.e. y ∈ Rn2 . As a consequence, g(y)µ(dy) = 0, and therefore χE ∈ F.

Rn2
N
S
(c) Suppose now E is a finite union of closed cubes whose interiors are disjoint, E = Qk . Then,
k=1
if Q◦k denotes the interior of Qk , we may represent χE (x) as a linear combination of the the
functions χQ◦k (x) and χAk (x) where Ak is a subset of the boundary of Qk for k = 1, . . . , N .
By our previous analysis, we know that χQ◦k (x) and χAk (x) belong to F for all k, and since
Step 1 guarantees that F is closed under finite linear combinations, we conclude that χE ∈ F,
as desired.
(d) Next, we prove that if E is open and of finite measure, then χE ∈ F. This follows from taking a
limit in the previous case. Indeed, by Theorem 3.5.9, we may represent E as a countable union
of almost disjoint closed cubes
∞
[
E= Qj .
j=1
100 CHAPTER 5. INTEGRATION THEORY

m
S
Consequently, if we let fm (x) = χQj (x), then we note that the functions fm increase to
j=1
f (x) = χE (x), which is integrable since µE is finite. Therefore, we may conclude by Step 2 that
f ∈ F.
(e) Finally, if E ∈ Gδ (Rn ) of finite measure, then χE ∈ F. Indeed, by definition, there exist open
sets G
e1 , G
e 2 , . . ., such that
∞
\
E= Gek .
k=1

e 0 of finite measure with E ⊂ G

Since E has finite measure, there exists an open set G e 0 . If we
set  
\ \ k
Gk = G e0  Gej 
j=1

then (Gk )∞
k=1 is a decreasing sequence of open sets of finite measure such that

G1 ⊃ G2 ⊃ · · ·

and
∞
\
E= Gk .
k=1

Therefore, the sequence of functions fk (x) = χGk (x) decreases to f (x) = χE (x), and since
χGk ∈ F for all k by (d) above, we conclude by Step 2 that χE (x) belongs to F.

4) If E has measure 0, then χE (x) belongs to F.

Indeed, since E is measurable, we may choose a set G ∈ Gδ (Rn ) with E ⊂ G and µG = 0 (see
Theorem 3.7.5). Since χG ∈ F (by the previous step) we find that
Z Z  Z
χG (x, y)µ(dx) µ(dy) = χG (x, y)dµ = 0.
Rn1
Rn2 Rn

Therefore Z
χG (x, y)µ(dx) = 0 for a.e. y ∈ Rn2 .
Rn1

Consequently, the slice G has measure 0 for a.e. y ∈ Rn2 .Z The simple observation that E y ⊂ Gy
y

then shows that E y has measure 0 for a.e. y ∈ Rn2 , and χE (x, y)µ(dx) = 0 for a.e. y ∈ Rn2 .
Rn1
Therefore, Z Z  Z
χE (x, y)µ(dx) µ(dy) = 0 = χE (x, y)dµ
Rn1
Rn2 Rn

and thus χE ∈ F, as was to be shown.

5) If E is any measurable subset of Rn with finite measure, then χE (x) belongs to F.

To prove this, recall first that by Theorem 3.7.5 there exists a set of finite measure G ∈ Gδ (Rn ),
with E ⊂ G and µ(G \ E) = 0. Since

χE (x) = χG (x) − χG\E (x),

and F is closed under linear combinations, we find that χE ∈ F, as desired.

5.5. FUBINI’S THEOREM AND ITS APPLICATIONS 101

6) If f ∈ L(Rn ), then f ∈ F.
We note first that f has the decomposition f = f + −f − , where non-negative and integrable functions
f + and f − are defined in Definition 4.2.3. So by Step 1 we may assume that f is itself non-negative.
By Theorem 4.4.2, there exists a sequence (hm )∞ m=1 of simple functions that increase to f . Since
each hm (x) is a finite linear combination of characteristic functions of sets with finite measure, we
have hm ∈ F by Steps 5 and 1, hence f ∈ F by Step 2.

5.5.2 Applications of Fubini’s theorem

Theorem 5.5.3 (Tonelly). Let f (x, y) ∈ S + (Rn1 × Rn2 ). Then for almost every y ∈ Rn2 :
(i) The slice f y (x) is measurable on Rn1 .
Z
(ii) The function defined by f y (x)µ(dx) is measurable on Rn2 .
Rn1
Moreover:
(iii) Z Z  Z
f (x, y)µ(dx) µ(dy) = f (x, y)dµ,
Rn1
Rn2 Rn

where the integrals exist (but can be infinite).

In practice, this theorem is often used in conjunction with Fubini’s
Z theorem. Indeed, suppose we
n
are given a measurable function f on R and asked to compute f dµ. To justify the use of iterated
Rn
integration, we first apply the present theorem to |f |. Using it, we may freely compute (or estimate) the
iterated integrals ofZthe non-negative function |f |. If these are finite, Theorem 5.5.3 guarantees that f is
integrable, that is, |f |dµ < +∞. Then the hypothesis in Fubini’s theorem is verified, and we may use
Rn
that theorem in the calculation of the integral of f .
Proof of Theorem 5.5.3. Consider the truncations
(
f (x, y), if |(x, y)| < l and f (x, y) < l,
fl (x, y) =
0, otherwise,

where |(x, y)| < l means the cube Q b l in Rn (see (3.5.1))

Each fl is integrable, and by part (i) in Fubini’s Theorem 5.5.2 there exists aSset El ⊂ Rn2 of measure 0
such that the slice fly (x) is measurable for all y ∈ Elc . Then, if we set E = El , we find that fly (x) is
l
measurable for all y ∈ E c and all l. Moreover, µ(E) = 0. Since fly % f y , Corollary 5.2.5 implies that if
y 6∈ E, then Z Z
fl (x, y)µ(dx) % f (x, y)µ(dx) as l → ∞.
Rn1 Rn1
Z Z
Again by Fubini’s Theorem 5.5.2, fl (x, y)µ(dx) is measurable for all y ∈ E c , hence so is f (x, y)µ(dx).
Rn1 Rn1
Another application of Corollary 5.2.5 then gives
Z Z  Z Z 
fl (x, y)µ(dx) µ(dy) → f (x, y)µ(dx) µ(dy), (5.5.3)
Rn1 Rn1
Rn2 Rn2
102 CHAPTER 5. INTEGRATION THEORY

By part (iii) in Fubini’s Theorem 5.5.2 we know that

Z Z  Z
fl (x, y)µ(dx) µ(dy) = fl (x, y)dµ, (5.5.4)
Rn1
Rn2 Rn

A final application of Corollary 5.2.5 directly to fl also gives

Z Z
fl (x, y)dµ → f (x, y)dµ. (5.5.5)
Rn Rn

Combining (5.5.3), (5.5.4), and (5.5.5) completes the proof of the theorem.
Corollary 5.5.4. If E is a measurable set in Rn1 × Rn2 , then for almost every y ∈ Rn2 the slice

E y = {x ∈ Rn1 : (x, y) ∈ E}

is a measurable subset of Rn1 . Moreover µ(E y ) is a measurable function of y and

Z
µ(E) = µ(E y )µ(dy).
Rn2

This is an immediate consequence of the first part of Theorem 5.5.3 applied to the function χE (x).
Clearly, a symmetric result holds for the x-slices in Rn2 .
We have thus established the basic fact that if E is measurable on Rn1 × Rn2 , then for almost every
y ∈ Rn2 the slice E y is measurable in Rn1 (and also the symmetric statement with the roles of x and y
interchanged). One might be tempted to think that the converse assertion holds. To see that this is not
the case let us recall an example considered for non-Borel measurable sets.
Let N denote a non-measurable subset of R, and then define

E = [0, 1] × N ⊂ R × R,

we see that (
[0, 1], y ∈ N,
Ey =
0, y 6∈ N .
Thus E y is measurable for every y. However, if E were measurable, then the corollary would imply that
Ex = {y ∈ R : (x, y) ∈ E} is measurable for almost every x ∈ R, which is not true since Ex is equal to N
for all x ∈ [0, 1].
Let us now study measurability and measures of product sets of the form E1 × E2 where Ek ∈ Rnk ,
k = 1, 2.
Theorem 5.5.5. If E = E1 × E2 is a measurable subset of Rn , and µ∗ (E2 ) > 0, then E1 is measurable.
Proof. By Corollary 5.5.4, we have that for a.e. y ∈ Rn2 , the slice function

(χE1 ×E2 )y (x) = χE1 (x)χE2 (y)

is measurable as a function of x. In fact, we claim that there is some y ∈ E2 such that the above slice
function is measurable in x; for such a y we would have χE1 ×E2 (x, y) = χE1 (x), and this would imply that
E1 is measurable.
To prove the existence of such a y, we use the assumption that µ∗ (E2 ) > 0. Indeed, let F denote the
of y ∈ Rn2 such that the slice E yTis measurable. Then µ(F c ) = 0 according T
set T Corollary
S 5.5.4. However,
E2 F is not empty because µ∗ (E2 F ) > 0. To see this, note that E2 = (E2 F ) (E2 F c ), hence
T
 \   \   \ 
0 < µ∗ (E2 ) 6 µ∗ E2 F + µ∗ E2 F c = µ∗ E2 F ,

F c is a subset of a set of measure zero.

T
because E2
5.5. FUBINI’S THEOREM AND ITS APPLICATIONS 103

To deal with a converse of the above result, we need the following lemma.
Lemma 5.5.6. If E1 ∈ Rn1 and E2 ∈ Rn2 , then
µ∗ (E1 × E2 ) 6 µ∗ (E1 )µ∗ (E2 ),
and if one of the sets Ek has outer measure zero, then µ∗ (E1 × E2 ) = 0.
(1) (2)
Proof. Let ε > 0. By Definition 3.3.1, we there exist bricks {Kj }∞
j=1 in R
n1
and {Kj }∞
j=1 in R
n2
such
that
∞ ∞
(1) (2)
[ [
E1 ⊂ Kj and E2 ⊂ Kj
j=1 j=1

and
∞ ∞
(1) (2)
X X
mKj 6 µ∗ (E1 ) + ε and mKj 6 µ∗ (E2 ) + ε.
j=1 j=1
∞
S (1) (2)
Since E1 × E2 ⊂ Kj × Kl , the semi-additivity of the outer measure yields
j,l=1
 
∞ ∞ ∞
!
(1) (2) (1) (2)
X X X
∗
µ (E1 × E2 ) 6 m(Kj × Kl ) = mKj  mKl 6 (µ∗ (E1 ) + ε)(µ∗ (E2 ) + ε).
j,l=1 j=1 l=1

If neither E1 nor E2 has outer measure 0, then from the above we find
µ∗ (E1 × E2 ) 6 µ∗ (E1 )µ∗ (E2 ) + o(ε),
and since ε is arbitrary, we have µ∗ (E1 × E2 ) 6 µ∗ (E1 )µ∗ (E2 ).
If for instance µ∗ (E1 ) = 0, consider for each positive integer m the set E2m = E2 {y ∈ Rn2 : |y| 6 m}.
T
Then, by the above argument, we find that µ∗ (E1 × E2m ) = 0. Since (E1 × E2m ) % (E1 × E2 ) as m → ∞,
we conclude that µ∗ (E1 × E2 ) = 0.
Theorem 5.5.7. Suppose E1 and E2 are measurable subsets of Rn1 and Rn2 , respectively. Then E =
E1 × E2 is a measurable subset of Rn . Moreover,
µ(E) = µ(E1 )µ(E2 ),
and if one of the sets Ek , k = 1, 2, has measure zero, then µ(E) = 0.
Proof. It suffices to prove that E is measurable, because then the assertion about µ(E) follows from
Corollary 5.5.4. Since each set Ek , k = 1, 2, is measurable, there exist sets Gk ⊂ Rnk of type Gδ (Rnk ) such
that Ek ⊂ Gk and µ∗ (Gk \ Ek ) = 0 for each k = 1, 2 according to Theorem 3.7.5. Clearly, G = G1 × G2 is
measurable in Rn1 × Rn2 and
[
(G1 × G2 ) \ (E1 × E2 ) ⊂ ((G1 \ E1 ) × G2 ) (G1 × (G2 \ E2 )).

By Lemma 5.5.6 we conclude that µ∗ (G \ E) = 0, hence E is measurable.

As a consequence of this proposition we have the following.
Theorem 5.5.8. Suppose f is a measurable function on Rn1 . Then the function fb defined by fb(x, y) =
f (x) is measurable on Rn1 × Rn2 .
Proof. Recall first that if c ∈ R and E1 = {x ∈ Rn1 : f (x) < c}, then E1 is measurable by definition.
Since
{(x, y) ∈ Rn1 × Rn2 : fb(x, y) < c} = E1 × Rn2 ,
Theorem 5.5.7 shows that {fb(x, y) < c} is measurable for each c ∈ R. Thus fb(x, y) is a measurable function
on Rn1 × Rn2 , as desired.
104 CHAPTER 5. INTEGRATION THEORY

This theorem will be of use when we will study the Fourier transform.
Theorem 5.5.9. If f is a measurable function on Rn , then the function fb(x, y) = f (x − y) is measurable
on Rn × Rn .
Proof. By picking E = {z ∈ Rn : f (z) < c}, we see that it suffices to prove that whenever E is a
measurable subset of Rn , then Ee = {(x, y) : x − y ∈ E} is a measurable subset of Rn × Rn .
Note first that if G is an open set, then Ge is also open. Taking countable intersections shows that if
n eTQ
E ∈ Gδ (R ), then so is E. Assume now that µ(E
e ej ) = 0 for each j, where E
 T 
ej = E b j = {|y| < j}.
b j and Q
Again, take G to be open in Rn , and let us calculate µ G e Q b j . We have that

χGe T Qbj (x, y) = χG (x − y)χQbj (y).

Hence  \  Z Z 
µ G
e Q
bj = χG (x − y)χQbj (y)µ(dy) µ(dx) =
Rn
Rn
Z Z 
= χG (x − y)µ(dx) χQbj (y)µ(dy) = µ(G)µ(Q
b j ),
Rn
Rn
by the translation-invariance of the measure. Now if µ(E) = 0, there is a sequence ofopen sets Gm such
that E ⊂ Gm and µ(Gm ) → 0. It follows from the above that E ej ⊂ Gem T Q b j and µ Gem T Qb j → 0 as
m → ∞ for each fixed j. This shows µ(E ej ) = 0, and hence µ(E)
e = 0. The proof of the proposition is
concluded once we recall that any measurable set E can be written as the difference of a Gδ (Rn ) and a
set of measure zero.
Finally, let us establish one of the fundamental facts of the integration theory, which is similar to a
correspondent property of the Riemann integral.
Theorem 5.5.10. Suppose f (x) is a non-negative function on Rn , and let
A = {(x, y) ∈ Rn × R : 0 6 y 6 f (x)}.
Then:
1) f is measurable on Rn if and only if A is measurable in Rn+1 .
2) If the conditions in 1) hold, then Z
f (x)dx = µ(A).
Rn

Proof. If f is measurable on Rn , then by Theorem 5.5.8 the function

F (x, y) = y − f (x)
is measurable on Rn+1 , so we immediately see that A = {y > 0} {F 6 0} is measurable.
T
Conversely, suppose that A is measurable. We note that for each x ∈ Rn the slice Ax = {y ∈ R :
(x, y) ∈ A} is a closed segment, namely Ax = [0, f (x)]. Consequently Corollary 5.5.4 (with the roles of x
and y interchanged) yields the measurability of µ(Ax ) = f (x). Moreover
Z Z Z
µ(A) = χA (x, y)dxdy = µ(Ax )dx = f (x)dx,
Rn+1 Rn Rn

as was to be shown.
Later we will prove a somewhat converse statement by introducing the so-called Radon-Nikodym deriva-
tive.
5.6. PROBLEMS 105

5.6 Problems
5.1.1 Basic properties of Lebesgue integral
Problem 5.1. Prove that the function
(
x2 ,
T
x ∈ Q [0, 1],
f (x) = √
x, x ∈ [0, 1] \ Q,

is Lebesgue integrable on [0, 1], find its Lebesgue integral over [0, 1]. Is f (x) Riemann integrable on [0, 1]?
Z
Problem 5.2. Calculate the integral f (x)dµ, where µ is the Lebesgue measure and
[0,1]


 sin x, x ∈ Q,
f (x) =

cos x, x 6∈ Q.


Z
Problem 5.3. Calculate the integral f (x)dµ, where µ is the Lebesgue measure and
[0,1]

 sin x,
 x ∈ Q,
f (x) =
sin2 x,

x 6∈ Q.


Problem 5.4. Calculate the integrals

Z Z Z
a) e−[x] dµ, b) e−[3x+1] dµ, c) e−[2x+2] dµ.
(0,+∞) (3,+∞) (4,+∞)

where [x] is the integer part of x (not exceeding x).

Problem 5.5. Calculate
Z Z Z
dµ dµ dµ
a) , b) , c) ,
[x + 1] · [x + 2] [2x + 1] · [2x + 3] [3x + 1]
(0,+∞) (0,+∞) (3,+∞)

(−1)[x] (−1)[x]
Z Z
d) dµ, e) dµ,
[x + 1] · [x + 2] [2x + 1] · [2x + 2]
(3,+∞) (0,+∞)

where [x] is the integer part of x (not exceeding x).

Hint: In the problem e) use the fact that
∞
X (−1)n π
= .
n=0
2n + 1 4

Problem 5.6. Does the integral

(−1)[x]
Z
dµ,
[x + 1]
(0,+∞)

exist? Here µ is the Lebesgue measure, and [x] is the integer part of x (not exceeding x).
106 CHAPTER 5. INTEGRATION THEORY

Problem 5.7. Does the integral

(−1)[2x+1]
Z
dµ,
[x + 1]
(0,+∞)

exist? Here µ is the Lebesgue measure, and [x] is the integer part of x (not exceeding x).

Problem 5.8. Calculate the integral Z

f (x, y)dµ,
D

where D = {(x, y) : 0 6 x 6 1, 0 6 y 6 1}, µ is the Lebesgue measure, and

(
1, where xy 6∈ Q
f (x, y) =
0, where xy ∈ Q.

Problem 5.9. Let fn ∈ S + (X), and Z

fn (x)dµ −→ 0.
n→∞
X
µ
Prove that fn −→ 0. Does this sequence converge to 0 a.e. on X?
n→∞

Problem 5.10. Prove that the condition

|fn (x)|
Z
dµ −→ 0
1 + |fn (x)| n→∞
X

µ
is equivalent to the fact fn −→ 0 whenever µX < +∞. Construct a counterexample in the case when
n→∞
X = R.

Problem 5.11. Let f ∈ L(Rn ), and let Xy = {x ∈ Rn : |f (x)| > y}. Prove that

Z +∞
Z
|f (x)|dµ = µ(Xy )dy.
Rn 0

Problem 5.12. Let f, g ∈ S + (X). Prove that

Z +∞
Z
f (x)g(x)dµ = Φ(y)dy,
X 0
Z
where Φ(y) = f (x)dx and Xy = X(g > y).
Xy

Hint: Use approximation of nonnegative measurable functions and Levi’s theorem.

Problem 5.13. Let f ∈ L(X) and for any set E ⊂ X, E ∈ M(X), the identity
Z
f (x)dµ = 0
E

a.e.
holds. Prove that f (x) = 0 on X.
5.6. PROBLEMS 107

Problem 5.14. Let f ∈ S(X), µX < +∞, and Xn = X(n − 1 6 f < n). Prove that f ∈ L(X) if, and
+∞
X
only if, the series |n| · µXn converges.
n=−∞

Problem 5.15. Construct a finite-valued function f ∈ S + (R) such that

+∞
X
n · µXn < +∞,
n=1

where Xn = X(n 6 f < n + 1), but such that f 6∈ L+ (R).

Problem 5.16. Let µX < +∞, f ∈ S(X), and there exist constants A > 0 and α > 1 such that for any
ε > 0 the following inequality holds

A
µ{x ∈ X : |f (x)| > ε} <
εα
Prove that f ∈ L(X) w.r.t. the measure µ.

Problem 5.17. Prove that if f ∈ L(X) and Xn = X(|f | > n), then lim n · µXn = 0.
n→+∞

Problem 5.18. Let µX < +∞ and f ∈ S(X). Define the sets Xn = {x ∈ X : |f (x)| > n}, n = 0, 1, 2, . . ..
Prove that f ∈ L(X) if, and only if,
X∞
µXn < ∞.
n=1

Problem 5.19. Construct an example of a function such that f ∈ S + (R), and

∞
X
µXn < ∞.
n=1

where Xn are defined in Problem 5.18, and f 6∈ L+ (R).

Problem 5.20. Let f ∈ S + (x), µX < +∞, and Xn = X(n − 1 6 f < n), n = 1, 2, . . .. Prove that
+∞
X
f m ∈ L+ (X) if, and only if, the series nm · µXn converges, m ∈ N.
n=1

Problem 5.21. Let f ∈ S + (x), µX < +∞, and Bn = X(f > n−1), n = 1, 2, . . .. Prove that f m ∈ L+ (X)
+∞
X
if, and only if, the series nm−1 · µBn converges, m ∈ N.
n=1

1 1
Problem 5.22. Prove that the function f (x) = cos is non-Lebesgue integrable on [0, 1].
x x
Problem 5.23. Let f be unbounded integrable function on X. Put
( (
f (x), |f (x)| 6 n, 0 f (x), |f (x)| 6 n,
[f (x)]n = [f (x)]n =
n , |f (x)| > n, 0 , |f (x)| > n.

Prove that Z Z Z
f (x)dµ = lim [f (x)]n dµ = lim [f (x)]0n dµ.
X n→+∞ X n→+∞ X
108 CHAPTER 5. INTEGRATION THEORY

µ
Problem 5.24. Let f, fn ∈ L+ (X), n ∈ N. Suppose that fn −→ f on X, and
Z Z
lim fn (x)dµ = f (x)dµ.
n→∞
X X

Prove that for any A ⊂ X, A ∈ M(X),

Z Z
lim fn (x)dµ = f (x)dµ.
n→∞
A A

Hint: Use Riesz’s Theorem on subsequences of sequences of measurable functions convergent in measure
and the Fatou theorem.

Problem 5.25. Construct an example showing that in Problem 5.24 the condition fn (x) > 0 on X is
substantial.

Problem 5.26. Let f ∈ L+ (X), and let g ∈ S(X), |g(x)| 6 M on X. Prove that there exists a number K,
inf{g(x) : x ∈ X} 6 K 6 sup{g(x) : x ∈ X}, such that
Z Z
f (x)g(x)dµ = K f (x)dµ.
X X

Construct an example showing that the nonnegativity of the function f on X is crucial.

µ
Problem 5.27. Let µX < +∞ and f, fn ∈ L(X), n ∈ N. Suppose that fn −→ f on X, the functions
(fn )+∞
n=1 have uniformly absolute continuous integrals over X, that is, for any ε > 0 there exists δ > 0 such
that for any A ⊂ X, A ∈ M(X), µA < δ, one has

Z
fn (x)dµ < ε ∀n ∈ N.
A

Prove that Z Z
lim fn (x)dµ = f (x)dµ.
n→∞
X X

a.e.
Problem 5.28. Let {fn }+∞
∈ L(X), and fn −→ f on X as n → +∞. Suppose that |fn | 6 g(x), n ∈ N
n=1
almost everywhere on X, where g(x) ∈ L(X) and |g(x)| 6 M almost everywhere on X. Prove that
Z Z
lim fn (x)g(x)dµ = f (x)g(x)dµ.
n→+∞ X
X

Problem 5.29. Let f ∈ L(X), and let (Xn )+∞

n=1 ⊂ X be a monotone decreasing sequence of measurable
sets, and X0 := lim Xn . Is the identity
n→∞
Z Z
f (x)dµ = lim f (x)dµ.
n→+∞ Xn
X0

true?
Z
Problem 5.30. Let µX < +∞. Prove that if the integral f (x)g(x)dµ exists and finite for any f ∈ L(X),
X
then g(x) is bounded almost everywhere on X.
5.6. PROBLEMS 109

Problem 5.31. Prove that if Z

f (x)g(x)dµ > 0
X

for any f such that f (x) > 0 on X, then g(x) > 0 almost everywhere on X.
1
Problem 5.32. Let f (x) = and µ be the Lebesgue measure on (0, 1). Prove that
x2
Z
f (x)dµ = +∞
(0,1)

using only the definition of Lebesgue integral for nonnegative functions.

1
Problem 5.33. Let f (x) = and µ be the Lebesgue measure on (0, 1). Prove that
x
Z
f (x)dµ = +∞
(0,1)

using only the definition of Lebesgue integral for nonnegative functions.

Problem 5.34. Construct a sequence of functions (fn )+∞ +
n=1 , fn ∈ L ([0, 1]) such that fn (x) → 0 for any
x ∈ [0, 1] but Z
fn (x)dµ 6→ 0 as n → ∞.
[0,1]

Problem 5.35. Show that there exists a sequence of functions (fn )+∞ +
n=1 , fn ∈ L ([0, 1]) such that fn (x) → 0
for any x ∈ [0, 1] and Z
fn (x)dµ → 0 as n → ∞,
[0,1]

but F (x) = sup fn (x) 6∈ L+ ([0, 1]).

n

1 n
Hint: Consider the sequence fn (x) = 2 χAn (x), where An = (2−n , 2−n+1 ), n ∈ N.
n
Problem 5.36. Construct an example of a sequence (fn )∞
n=1 ∈ L(R) such that fn −→ 0 uniformly on R
n→∞
but Z
fn (x)dµ −→ ∞.
n→∞
R

Problem 5.37. Let (X, M, µ) be a measure set, and let f ∈ L+ (X). Define
Z
ν(A) = f (x)dµ.
A

Prove that
1) ν is a σ-additive measure;
2) if g is integrable w.r.t. ν, then f g is integrable w.r.t. µ, and
Z Z
g(x)dν(x) = f (x)g(x)dµ(x).
X X
110 CHAPTER 5. INTEGRATION THEORY

5.1.2 Difference between Lebesgue and Riemann definite integrals

In this section, µ is the Lebesgue measure.

Problem 5.38. Consider the functions


 0, x = 0,
f (x) = 1
 ln , x ∈ (0, 1],
x
and 
x ∈ 0, n1 ,


0,
fn (x) = 1
x ∈ n1 , 1 .
 
 ln ,
x
a.e.
Prove that fn −→ f on [0, 1]. Is f (x) Riemann integrable on the interval [0, 1]? Is it Lebesgue integrable
on [0, 1]? If so, find the Lebesgue integral of this function on [0, 1].

Problem 5.39. Is the function ( T

1, x ∈ Q [0, 1],
f (x) =
x3 , x ∈ [0, 1] \ Q,
Riemann integrable on the interval [0, 1]? Is it Lebesgue integrable on [0, 1]? If so, find the Lebesgue
integral of this function on [0, 1].

Problem 5.40. Let K ∈ Rn be a brick, and f ∈ R(K). Prove that f ∈ L(K), and
Z Z
(L) f (x)dµ = (R) f (x)dxn .
K K

Problem 5.41. Let f ∈ R(a + 0, b], that is, let the improper integral

Zb Zb
(R) f (x)dx := lim (R) f (x)dx
c→a+0
a+0 c

be finite. Suppose that f (x) > 0 on (a, b]. Prove that f ∈ L+ (a, b) and

Z Zb
(L) f (x)dµ = (R) f (x)dx.
(a,b) a+0

Problem 5.42. Let f ∈ R[a0 , b] for any a0 ∈ (a, b). Prove that |f | ∈ R(a + 0, b] if, and only if, f ∈ L(a, b).
And if f is integrable, then
Z Zb
(L) f (x)dµ = (R) f (x)dx.
(a,b) a+0

Hint: Putting a0 := b and considering some sequence an & a, prove that

Zb Zb
(R) f (x)dx = lim (R) f (x)dx,
n→∞
a+0 an

then, use the complete additivity of the Lebesgue integral.

5.6. PROBLEMS 111

Problem 5.43. Let α ∈ R, and (

xα , x ∈ (0, 1],
f (x) =
0, x = 0.
Find all α such that

a) f ∈ R[0, 1];

b) f ∈ R(+0, 1];

c) f ∈ L[0, 1].

Problem 5.44. Let α ∈ R, and f (x) = xα on [1, +∞). Find all α such that

a) f ∈ R[1, +∞);

b) f ∈ L[1, +∞).

Problem 5.45. Let α ∈ R, β > 0, and

(
xα sin xβ , x ∈ (0, 1],
f (x) =
0, x = 0.

Find all pairs (α, β) such that

a) f ∈ R[0, 1];

b) f ∈ R(+0, 1];

c) f ∈ L[0, 1].

Problem 5.46. Let α ∈ R, β < 0, and

(
xα sin xβ , x ∈ (0, 1],
f (x) =
0, x = 0.

Find all pairs (α, β) such that

a) f ∈ R[0, 1];

b) f ∈ R(+0, 1];

c) f ∈ L[0, 1].

Problem 5.47. Let α, β ∈ R, and f (x) = xα sin xβ on [1, +∞). Find all pairs (α, β) such that

a) f ∈ R[1, +∞);

b) f ∈ L[1, +∞).

Problem 5.48. Find the limit Z

dµ(x)
lim (L) 1 .
n→∞ x n


(0,+∞)
xn 1+ n

1
Hint: To prove that integrands 1
n are Lebesgue integrable on (0, +∞), estimate them by
xn 1 + nx
Lebesgue integrable functions on (0, 1) and [1, +∞) separately. Then use Lebesgue’s Dominated Conver-
gence Theorem.
112 CHAPTER 5. INTEGRATION THEORY

[0, 1] = {rn }∞
T
Problem 5.49. Let Q n=1 and ε > 0. Prove that the series
∞
X 1
p .
n=1 n1+ε |x − rn |
converges a.e. on [0, 1].

Hint: Using Problem 5.42 and Levi’s theorem, prove that the sum of the series is Lebesgue integrable on (0, 1).
Problem 5.50. Let X = (0, 1). Construct nonnegative functions (fn )∞
n=1 and f measurable w.r.t. the
Lebesgue measure µ such that f (x) = limfn (x) on X,
Z
fn (x)dµ 6 C ∀n ∈ N
X

for some constant C > 0, but f 6∈ L(X).

Problem 5.51. Let C be the Cantor set on [0, 1], and


 1 if x ∈ C,




  
1


x ∈ 0, \ C,

f (x) x if
 2



  
1


x2 if x ∈ , 1 \ C.



2
Find Z
(L) f (x)dµ.
(0,1)

Problem 5.52. Let C be the Cantor set on [0, 1] and let

∞
[0, 1] \ C = ⊍ (an , bn ).
n=1

Define
 the functions
 fn as follows. For any n ∈ N, we have fn (x) = 0 for anyx ∈ [0, 1] \ (an , bn ),
an + bn an + bn an + bn
fn = 1, and fn (x) is continuous and linear on the intervals an , , , bn . Let
2 2 2
∞
X
f (x) = fn (x).
n=1

Find Z
(L) f (x)dµ.
(0,1)

Problem 5.53. Let C be the Cantor set on [0, 1] and let

∞
[0, 1] \ C = ⊍ (an , bn ).
n=1

Define
 the functions
 fn as follows. For any n ∈ N, we have fn (x) = 0 for any x ∈ [0,  1] \ (an , bn),
an + bn bn − an an + bn an + bn
fn = , and fn (x) is continuous and linear on the intervals an , , , bn .
2 2 2 2
Let
∞
X
f (x) = fn (x).
n=1
5.6. PROBLEMS 113

Find Z
(L) f (x)dµ.
(0,1)

Problem 5.54. Let

∞
X xi
x=
i=1
2i

be the binary representation of a number x ∈ [0, 1), where xi ∈ {0, 1}, and lim xi 6= 1.
i→∞
Let fk (x) = 2xk − 1 for k ∈ N and x ∈ (0, 1). Prove that the sequence {fk (x)}∞
k=1 is orthonormal on
(0, 1), that is, prove that Z
(L) fk (x)fj (x)dµ = 0, for j 6= k,
(0,1)

and Z
(L) fk2 (x)dµ = 1, k ∈ N.
(0,1)

Problem 5.55. Find functions f, g ∈ R((+0, 1]) such that

h(x) = max{f (x), g(x)} 6∈ R((+0, 1]).

5.1.3 Fubini’s theorem

Let X1 ⊂ Rn1 and X2 ⊂ Rn2 . Let µ1 and µ2 be two Lebesgue-Stieltjes σ-finite (or finite) measures defined
on M1 = M(X1 ) and M2 = M(X2 ), respectively. The Descartes product of these two measures spaces
is the following space (X, M, µ), where X = X1 × X2 , M = M1 × M2 , that is, any A ∈ M can be
represented as A = A1 × A2 with A1 ∈ M1 and A2 ∈ M2 . The measure µ of A ∈ M is defined as
µ(A) = µ1 (A1 ) · µ2 (A2 ). it is easy to see that µ is a Lebesgue-Stieltjes measure defined on M = M(X).
In this section, we consider such three measure spaces.

Problem 5.56. Let A ∈ M, µA < ∞. Prove that for µ1 -a.e. x in X1 , the slice Ax is µ2 -measurable, that
is, Ax ∈ M2 . Also prove that the function µ2 (Ax ) is µ1 -measurable on X1 , and
Z
µ(A) = µ2 (Ax )dµ2 (x).
X1

Problem 5.57. Let A ∈ M, and f (x, y) is integrable on A w.r.t the measure µ. Prove that

1) fx is µ2 -integrable on Ax for µ1 -a.e. x in X1 , where Ax is the x slice of A;

2) the function Z
fx (y)dµ2 (y)
Ax

is µ1 -integrable on X1 ;

3) and Z Z Z
f (x, y)dµ = fx (y)dµ2 (y)dµ1 (x).
A X1 Ax

Problem 5.58. Let A ∈ M, and f (x, y) ∈ S + (A). Prove that

114 CHAPTER 5. INTEGRATION THEORY

1) fx is µ2 -measurable for µ1 -a.e. x in X1 ;

2) the function Z
fx (y)dµ2 (y)
Ax

is µ1 -measurable, where Ax is the x slice of A;

3) and Z Z Z
f (x, y)dµ = fx (y)dµ2 (y)dµ1 (x).
A X1 A x

Problem 5.59. Construct a function f (x, y) 6∈ L((0, 1) × (0, 1)) such that f (x0 , y) is integrable on (0, 1)
for any fixed x0 ∈ (0, 1), and f (x, y0 ) is integrable on (0, 1) for any fixed y0 ∈ (0, 1), and such that
Z Z
f (x0 , y)dµ(y) = f (x, y0 )dµ(x) = 0.
(0,1) (0,1)

Problem 5.60. Construct a finite function f (x, y) measurable on (0, 1) × (0, 1) such that
 
Z Z
f (x, y)dµ(x) dµ(y) = 0,
 

(0,1) (0,1)

and  
Z Z
f (x, y)dµ(y) dµ(x) = 1.
 

(0,1) (0,1)

Hint: Consider the function

∞
X
4n χAn ×An (x, y) − 2χAn+1 ×An (x, y) ,
 
f (x, y) =
n=0

where An = (2−n , 2−n+1 ), n ∈ N.

Problem 5.61. Let a, b ∈ R, a 6= b. Construct a finite function f (x, y) measurable on (0, 1) × (0, 1) such
that  
Z Z
f (x, y)dµ(x) dµ(y) = a,
 

(0,1) (0,1)

and  
Z Z
f (x, y)dµ(y) dµ(x) = b.
 

(0,1) (0,1)

Problem 5.62. Prove that the function f (x, y) = e−xy sin x sin y is integrable on (0, +∞)×(0, +∞) (w.r.t.
Lebesgue measure).
Problem 5.63. Let f (x) ∈ Lµ1 (X1 ), g(x) ∈ Lµ2 (X2 ). Prove that f (x) · g(y) ∈ Lµ (X). Here Lν means
integrability w.r.t. the measure ν.
Problem 5.64. Prove that if A = A1 × A2 (A1 ⊂ X1 and A2 ⊂ X2 ) is measurable (w.r.t. µ) and has
positive measure, then the sets A1 and A2 are measurable w.r.t measures µ1 and µ2 , respectively.
5.6. PROBLEMS 115

Problem 5.65. Let a function f be measurable on X = X1 × X2 (w.r.t. µ), and let the following integral
exist  
Z Z
I =  |f (x, y)|dµ2 (y) dµ1 (x) < +∞.
X1 X2

Prove that f ∈ L(X, M, µ).

Problem 5.66. Let f, g ∈ L(R). Prove that the function h(x, t) = f (t)g(x − t) ∈ L(R) for a.e. x ∈ R and
that the convolution Z
(f ∗ g)(x) = f (t)g(x − t)dµ(t)
R

belongs to L(R).
116 CHAPTER 5. INTEGRATION THEORY
Chapter 6

Differentiation of integrable functions

I cover pages 98–108, 115–136 and 285–292 from the book [11].

117
118 CHAPTER 6. DIFFERENTIATION OF INTEGRABLE FUNCTIONS

6.1 Problems
6.1.1 Differentiation
Problem 6.1. Prove that if f is integrable on Rn , and f is not identically zero, then
c
f ∗ (x) > , for some c > 0 and all |x| > 1,
|x|n
p
where |x| = x21 + · · · + x2n . Conclude that f ∗ is not integrable on Rn . Then, show that the weak type
estimate
c
µ({x : F ∗ > α}) 6
α
R
for all α >R 0 whenever |f |dµ = 1, is the best possible in the following sense: if f is supported in the unit
ball with |f |dµ = 1, then
c0
µ({x : F ∗ > α}) >
α
0
for some c > 0 and all sufficiently small α.
R
Hint: For the first part, use the fact that |f |dµ > 0 for some ball B.
B

Problem 6.2. Consider the function on R defined by

1 1

 if |x| 6 ,
 |x|(ln 1 )2


|x|
2
f (x) =



 0 otherwise.

(a) Verify that f is integrable.

(b) Establish the inequality

c 1
f ∗ (x) > 1 for some c > 0 and all |x| 6 ,
|x|(ln |x| ) 2

to conclude that the maximal function f ∗ is not locally integrable.

6.1.2 Functions of bounded variation

By V [a, b] we denote the set of all functions of bounded variation defined on the interval [a, b]. By Vab (f )
we denote the (total) variation of the function f on [a, b]:
N
X
Vab (f ) = sup |F (tk ) − F (tk−1 )|,
T
k=1

where the supremum is taken over all partitions of the interval [a, b]:

T := {tk }N
k=0 , a = t0 < t1 < . . . < tN −1 < tN = b.

Problem 6.3. Let c ∈ (a, b). Prove that f ∈ V [a, b] if, and only if, f ∈ V [a, c] ∩ V [c, b], and

Vab (f ) = Vac (f ) + Vcb (f ).

Problem 6.4. Let f ∈ V [a, b]. Prove that f can be represented as a difference of strictly increasing
functions.
6.1. PROBLEMS 119

Problem 6.5. Let 

 1
 for x = a ∈ [0, 1],
f (x) =

0 for x ∈ [0, 1] \ {a}.


Represent f as a difference of two non-strictly increasing functions.

Problem 6.6. Let f (x) = sin x defined on [0, 2π]. Represent f as a difference of two non-strictly increasing
functions.
Problem 6.7. Let f ∈ V [a, b]. Prove that f is non-strictly increasing on [a, b] if, and only if, Vab (f ) =
f (b) − f (a).
Problem 6.8. Let f be finite on [a, b], and let ϕ be a strictly increasing continuous function on [a, b] with
ϕ(a) = a and ϕ(b) = b. Prove that f ∈ V [a, b] if, and only if, g(x) = f (ϕ(x)) ∈ V [a, b]. Show by an
example that continuity of ϕ is crucial in this result. Find f 6∈ V [0, 1] and strictly increasing function ϕ
with ϕ(0) = 0, ϕ(1) = 1 such that g(x) = f (ϕ(x)) ∈ V [0, 1].
Problem 6.9. Let f, g ∈ V [a, b]. Prove that f and g are bounded, f · g ∈ V [a, b], and

Vab (f · g) 6 sup |g(x)| · Vab (f ) + sup |f (x)| · Vab (g).

x∈[a,b] x∈[a,b]

g
Problem 6.10. Let f, g ∈ V [a, b], and |f (x)| > C > 0 for x ∈ [a, b]. Prove that ∈ V [a, b].
f
Problem 6.11. Let f, g ∈ V [a, b]. Prove that max{f, g} ∈ V [a, b].
Problem 6.12. Let f ∈ V [a, b]. Prove that |f | ∈ V [a, b], and Vab (|f |) 6 Vab (f ).
Problem 6.13. Find all real α and β such that the function
(
xα sin xβ for x ∈ (0, 1],
f (x) =
0 for x = 0,

is of bounded variation on [0, 1].

Problem 6.14. Let f ∈ V [a, b]. Prove that f ∈ C[a, b] if, and only if, Vax (f ) ∈ C[a, b].
Problem 6.15. Let a set A ⊂ [a, b] be of Lebesgue measure zero, µA = 0. Find a continuous (non-strictly)
increasing function on [a, b] s.t. f 0 (x) = ∞ for any x ∈ A.
Hint: To construct a necessary function, use the following auxiliary function
∞
X
g(x) = n2 χGn (x),
n=1

where G1 ⊃ G2 ⊃ . . . ⊃ A, and µGn < 2−n .

Problem 6.16 (Fubini’s Little Theorem). Let {fn (x)}∞
n=1 be a sequence of non-strictly increasing func-
tions on [a, b]. Let also the series
X∞
fn (x)
n=1

converge everywhere on [a, b], and let f (x) be its sum. Prove that
∞
X
f 0 (x) = fn0 (x), a.e. on [a, b].
n=1
120 CHAPTER 6. DIFFERENTIATION OF INTEGRABLE FUNCTIONS

Problem 6.17. Construct a sequence {fn (x)}∞

n=1 of non-strictly increasing continuous functions on [−1, 1]
such that the series
∞
X
fn (x)
n=1

converges uniformly on [−1, 1] to a function f (x), and

∞
X
fn (0) = 0,
n=1

but f 0 (0) = ∞.
Hint: To construct a necessary function, use the following auxiliary functions

 0, 0 6 x 6 2−n−1 ,
 √
gn (x) = 2n+1 x(x − 2−n−1 ), 2−n−1 < x < 2−n ,
√
2−n 6 x 6 1.

x,


Problem 6.18. Given a sequence {fn (x)}∞

n=1 ⊂ V [a, b] such that

∞
X ∞
X
|fn (a)| < ∞ and Vab (fn ) < ∞,
n=1 n=1

prove that the series

∞
X
fn (x)
n=1

converges to some function f (x) ∈ V [a, b] uniformly on [a, b], and that
∞
X
Vab (f ) 6 Vab (fn ).
n=1

Problem 6.19. Consider the following jump function

∞
X
h(x) = ak gk (x),
k=1

where
∞
X
α= |ak | < ∞
k=1

and gk (x) = χ(ck ,b] with ck ∈ [a, b], ck 6= cj if k 6= j. Prove that Vab (h) = α < ∞.

Problem 6.20. Construct a strictly increasing function f (x) ∈ C[a, b] ∩ V [a, b] such that f 0 (x) = 0 a.e.
on [a, b].

Problem 6.21. Let {fn (x)}∞ n=1 be a sequence of non-strictly increasing functions on [a, b]. Suppose that
there exists a constant C > 0 such that |fn (x)| 6 C for any x ∈ [a, b] and for any n ∈ N. Prove that
there exists a subsequence {fnk }∞
k=1 converging everywhere on [a, b] to a non-strictly increasing function
on [a, b].

Problem 6.22. Let {fn (x)}∞ b

n=1 ⊂ V [a, b]. Suppose that there exists a constant C > 0 such that Va (fn )
and |fn (x)| 6 C for any x ∈ [a, b] and for any n ∈ N. Prove that there exists a subsequence {fnk }∞ k=1
converging everywhere on [a, b] to a function f ∈ V [a, b].
6.1. PROBLEMS 121

Problem 6.23. Let f ∈ C[a, b]. For a given partition T = {a = t0 < t1 < · · · < tn−1 < tn = b} of the
interval [a, b] we define the intervals ∆k = [tk−1 , tk ] and the magnitudes Mk = max f (t) and mk = min f (t)
t∈∆k t∈∆k
for k = 1, . . . , n. Let
n
X
Ω(T ) = (Mk − mk ).
k=1
Prove that
Vab (f ) = lim VT (f ) = lim Ω(T ),
λ(T )→0 λ(T )→0

where λ(T ) = max |∆k |. Here the value Vab (f ) can be finite or infinite.
16k6n
Construct a function f 6∈ V [0, 1] and a sequence {Ti }∞
i=1 of partitions of the interval [0, 1] such that
λ(Ti ) → 0 as i → ∞ and VTi (f ) = 0.
Hint: Show that lim inf VT (f ) > α for any α < Vab (f ).
λ(T )→0

Problem 6.24. Let f ∈ C[a, b]. The function g(y) equal to the number of solutions of the equation
f (x) = y is called Banach indicatrix (g(y) = ∞ is possible). Prove that g is measurable and Lebesgue
integrable, and Z
Vab (f ) = g(y)dµ(y).
R

Hint: Approximate g(y) by an increasing sequence of some measurable functions, and use Levi’s theorem
and Problem 6.23.
Problem 6.25. Let f ∈ V [a, b] and f (x) = ψ(x) + j(x), where j(x) is the jump function related to f (x).
Prove that Vab (f ) = Vab (ψ) + Vab (j).
Problem 6.26. Let f ∈ C[a, b] ∩ V [a, b]. Prove that there exist nonnegative non-strictly increasing
continuous function f1 (x) and f2 (x) such that f (x) = f1 (x) − f2 (x) and Vab (f ) = Vab (f1 ) + Vab (f2 ).
Problem 6.27. Let f ∈ V [a, b]. Prove that there exist nonnegative non-strictly increasing function f1 (x)
and f2 (x) such that f (x) = f1 (x) − f2 (x) and Vab (f ) = Vab (f1 ) + Vab (f2 ).

6.1.3 Absolutely continuous functions

If a function f is absolutely continuous on [a, b], then we write f ∈ AC[a, b].
f
Problem 6.28. Let f, g ∈ AC[a, b]. Prove that f ·g ∈ AC[a, b], and if g(x) 6= 0 on [a, b], then ∈ AC[a, b].
g
Problem 6.29. Prove that if f ∈ AC[a, b], then |f | ∈ AC[a, b].
And conversely, prove that if |f | ∈ AC[a, b] and f ∈ C[a, b], then f ∈ AC[a, b].
Problem 6.30. Prove that if f ∈ AC[a, b], then f ∈ V [a, b].
Problem 6.31. Construct a function f ∈ C[0, 1] ∩ V [0, 1] such that f 6∈ AC[0, 1].
Problem 6.32. Prove that if f ∈ AC[a, b], then Vax (f ) ∈ AC[a, b].
Problem 6.33. Prove that f ∈ AC[a, b] if, and only if, for any ε > 0 there exists δ > 0 such that for any
countable system S = {(ak , bk )}∞
k=1 of disjoint open intervals in [a, b] with
∞
X
|bk − ak | < δ
k=1

one has
∞
X
|f (bk ) − f (ak )| < ε
k=1
122 CHAPTER 6. DIFFERENTIATION OF INTEGRABLE FUNCTIONS

Problem 6.34. Let f ∈ AC[a, b], and suppose that g(x) is a non-strictly increasing on [c, d], g(x) ∈
AC[c, d], g(c) = a, g(d) = b. Prove that f (g(x)) ∈ AC[c, d].

Problem 6.35. A function f defined on [a, b] is said to possess Luzin N -property if for any Lebesgue
measurable set E, E ⊂ [a, b], with µE = 0, the set f (E) is Lebesgue measurable and µ (f (E)) = 0.
Prove that if f ∈ AC[a, b], then f possesses Luzin N -property.
Hint: Use the result of Problem 6.33.
Problem 6.36. Let f ∈ C[a, b]. Prove that f (E) is Lebesgue measurable for any Lebesgue measurable
set E ⊂ [a, b] if, and only if, f (x) possesses Luzin N -property.
Hint: Use the fact that continuous functions transfers compact sets to compact sets and the result of
Problem 3.22.
Problem 6.37 (Banach–Zarecki). Let f ∈ C[a, b]∩V [a, b], and suppose that f possesses Luzin N -property.
Prove that f ∈ AC[a, b].
Problem 6.38. Let f ∈ AC[a, b] be non-strictly increasing on [a, b], and let a set A ⊆ [a, b] be Lebesgue
measurable. Prove that Z
µ(f (A)) = f 0 (t)dt,
A

where µ is the Lebesgue measure.

Hint: Prove first that the Lebesgue-Stieltjes measure µf constructed with the function f is absolutely
continuous w.r.t. the Lebesgue measure µ.

Problem 6.39. Let F, G ∈ AC[a, b]. Prove that

Z Z
F (x)G0 (x)dµ = F (b)G(b) − F (a)G(a) − F 0 (x)G(x)dµ.
[a,b] [a,b]

Problem 6.40. Let f ∈ L[a, b], and suppose that the function G ∈ AC[c, d] is strictly increasing on
[c, d]. Moreover, let G(c) = a, G(d) = b, and the inverse function G−1 (y) belong to AC[a, b]. Prove that
f (G(y))G0 (y) ∈ L[c, d], and Z Z
f (x)dµ(x) = f (G(y))G0 (y)dµ(y).
[a,b] [c,d]

Problem 6.41. Let f (x) be Lipschitz continuous on [a, b]. Prove that f ∈ AC[a, b].
1
Prove also that the function f (x) = on (0, 1], f (0) = 0, is absolutely continuous on [0, 1], but not
ln x2
Lipschitz continuous.
Problem 6.42. Find all real α and β such that the function
(
xα sin xβ for x ∈ (0, 1],
f (x) =
0 for x = 0,

is absolutely continuous on [0, 1].

Hint: Use Problems 6.41, 6.35, 6.13, and 6.37.
Problem 6.43. Let f ∈ C[a, b]∩V [a, b]. Prove that there exists a unique representation f (x) = g(x)+h(x),
where g ∈ AC[a, b], g(a) = f (a), and h(x) ≡ 0 on [a, b] or h(x) ∈ C[a, b] ∩ V [a, b] is such that h(x) 6≡ const
and h0 (x) = 0 a.e. on [a, b].
Chapter 7

Linear spaces

7.1 Hölder and Minkowski inequalities

On the plane (ξ, η), consider the line η = ξ p−1 for ξ > 0, p > 1. Let the number q be such that

1 1
+ = 1. (7.1.1)
p q
1
From this identity it follows that q − 1 = , so q > 1, since p > 1 by assumption. Moreover, the
p−1
p−1
equation η = ξ is equivalent to the equation ξ = η q−1 .
For a, b > 0, let S1 be the square of the set bounded by the axis Oξ, by the graph of the function
η = ξ p−1 , and by the line ξ = a, and let S2 be the square of the set bounded by the axis Oη, by the graph
of the function η = ξ p−1 , and by the line η = b. Obviously,

S1 + S2 > ab, (7.1.2)

where equality holds if, and only if, b = ap−1 . The squares S1 and S2 can be calculated as follows

Za Zb
p−1 ap bq
S1 = ξ dξ = , S2 = η q−1 dη = . (7.1.3)
p q
0 0

Substituting these values of S1 and S2 into (7.1.2), we obtain the following inequality

ap bq
ab 6 + , (7.1.4)
p q

called the Young inequality. Note that the equality in (7.1.4) is possible if, and only if, b = ap−1 , that is,
if, and only if, bq = ap .

Hölder inequality
Let (X, M(X), µ) be a measure space. Suppose that the measure µ is complete and σ-additive. Suppose
also that f, g ∈ S(X) such that |f |p , |g|q ∈ L(X), where p and q are related as in (7.1.1). Put now in the
Young inequality (7.1.4)

|f (x)| |g(x)|
a := , b := ,
 1  1
R p R q
|f (x)|p dµ |g(x)|q dµ
X X

123
124 CHAPTER 7. LINEAR SPACES

then we have
|f (x)g(x)| |f (x)|p |g(x)|q
1 1 6  +  .
    R R
R p R q p p
|f (x)| dµ q q
|g(x)| dµ
|f (x)|p dµ |g(x)|q dµ X X
X X

The right-hand side of this inequality is Lebesgue integrable on X by assumption. Therefore, the left-
hand side is also Lebesgue integrable on X by property 8) of integral of measurable functions. Thus, we
have Z Z Z
|f (x)g(x)|dµ |f (x)|p dµ |g(x)|q dµ
X X X 1 1
6  +   = + = 1,
 1  1 R R p q
R p R q p |f (x)|p dµ q |g(x)|q dµ
|f (x)|p dµ |g(x)|q dµ X X
X X
so
 1  1
Z Z p Z q

|f (x)g(x)|dµ 6  |f (x)|p dµ  |g(x)|q dµ . (7.1.5)

X X X
This inequality is called the Hölder inequality. For the case p = q = 2 we get
 1  1
Z Z 2 Z 2
2 2
|f (x)g(x)|dµ 6  |f (x)| dµ   |g(x)| dµ  .
X X X

This inequality is called Cauchy–Schwarz–Bunyakovsky inequality.

Z
Remark 7.1.1. We prove the Hölder inequality under implicit assumption that the integrals |f (x)|p dµ,
Z Z X
q p a.e.
|g(x)| dµ are non-zero. However, if, for instance, |f (x)| dµ = 0, then f (x) = 0 on X by the property
X X
8) of integral of nonnegative functions. But in this case, the left-hand side of (7.1.5) is zero together with
its right-hand side, so the inequality (7.1.5) holds.
Remark 7.1.2. In the proof of the Hölder inequality we assumed that both integrals in the right-hand
side of (7.1.5) are finite. In fact, the inequality (7.1.5) holds if one or both integrals are infinite. But in
this case it becomes useless. The main meaning of the Hölder inequality is that from the integrability of
the functions |f |p and |g|q we get the integrability of the product |f g|.
Remark 7.1.3. Since the equality in the Young inequality appears if, and only if, ap = bq , we have that
a.e.
in the Hölder inequality the equality appears if, and only if, |f (x)|p = C|g(x)|q on X for some constant
C > 0.
Remark 7.1.4. In the same way, from the Young inequality, one can prove the following finite sum form
of the Hölder inequality for real xk , yk , k = 1, . . . , n,

n n
!1 n
!1
X X p X q
xk yk 6 |xk |p |yk |q , (7.1.6)
k=1 k=1 k=1

where equality holds if, and only if,

|xj |p |yj |q
n
P = P
n , sgn xj yj = const, j = 1, . . . , n.
|xk | p |yk |q
k=1 k=1
7.1. HÖLDER AND MINKOWSKI INEQUALITIES 125

Remark 7.1.5. From (7.1.6), one can easily deduce the following form of the Hölder inequality

∞ ∞
!1 ∞
!1
X X p X q
xk yk 6 |xk |p |yk |q ,
k=1 k=1 k=1

provided the series in the right hand-side of this inequality converge.

Minkowski inequality
Let now f, g ∈ S(X) and |f |p , |g|p ∈ L(X), p > 1. Consider two sets X1 = X(|f | > |g|) and X2 = X(|f | <
|g|). Then we have
Z Z Z Z Z
p p p p p p
|f (x) + g(x)| dµ = |f (x) + g(x)| dµ + |f (x) + g(x)| dµ 6 2 |f (x)| dµ + 2 |g(x)|p dµ,
X X1 X2 X1 X2

so |f + g|p ∈ L(X). Moreover,

Z Z
|f (x) + g(x)| dµ = |f (x) + g(x)|p−1 · |f (x) + g(x)|dµ 6
p

X X
Z Z
6 |f (x) + g(x)|p−1 · |f (x)|dµ + |f (x) + g(x)|p−1 · |g(x)|dµ
X X

Applying Hölder inequality to both summands in the right-hand side of the inequality above and recalling
that (p − 1)q = p, we get

1  1  1
 

Z Z p Z p Z q
p p p p
 
|f (x) + g(x)| dµ 6 
 |f (x)| dµ +  |g(x)| dµ 

 |f (x) + g(x)| dµ
X X X X

 1
q
p
R
Now we divide both parts of the inequality by |f (x) + g(x)| dµ to obtain the Minkowski inequality
X

 1  1  1
Z p Z p Z p
 |f (x) + g(x)|p dµ 6  |f (x)|p dµ +  |g(x)|p dµ . (7.1.7)
X X X

1 1
Here we used the identity 1 − = .
q p
Z
Remark 7.1.6. We prove the Minkowski under implicit assumption that the integral |f (x) + g(x)|p dµ
X
is non-zero. However, if it is zero then the inequality (7.1.7) is obvious.

Remark 7.1.7. We deduce the Minkowski inequality from the Hölder inequality which holds for p > 1.
However, the Minkowski holds for p > 1, since for p = 1 it is obvious.
a.e.
Remark 7.1.8. Since the equality in the Hölder inequality appears if, and only if, |f (x)|p = C|g(x)|q on
X for some constant C > 0, we have that in the Minkowski inequality the equality appears if, and only if,
a.e. a.e.
either f (x) = Cg(x) on X for some constant C > 0, or g(x) = 0 on X.
126 CHAPTER 7. LINEAR SPACES

Remark 7.1.9. In the same way, from the Hölder inequality, one can prove the following finite sum form
of the Minkowski inequality for real xk , yk , k = 1, . . . , n,

n
!1 n
!1 n
!1
X p X p X p
|xk + yk |p 6 |xk |p + |yk |p . (7.1.8)
k=1 k=1 k=1

where equality holds if, and only if,

xj = λyj , λ > 0, or yj = 0, j = 1, . . . , n.

Remark 7.1.10. As above, from (7.1.8), one can easily deduce the series form of the Minkowski inequality

∞
!1 ∞
!1 ∞
!1
X p X p X p
|xk + yk | 6 |xk |p + |yk |p , (7.1.9)
k=1 k=1 k=1

provided the series in the right hand-side of this inequality converge.

7.2 Linear normed spaces

Definition 7.2.1. A non-empty set V is called a linear or vector space over a field F if it satisfies the
following conditions.

I. For any x, y ∈ V , there is a uniquely determined element z ∈ V called the sum of x and y, and
denoted as x + y. Moreover,

1) x + y = y + x (commutativity),
2) x + (y + z) = (x + y) + z (associativity),
3) There exists an element 0 ∈ V such that x + 0 = x for all x ∈ V (existence of zero),
4) For any x ∈ V there exists in element −x ∈ V such that x + (−x) = 0 (existence of the opposite
element),

II. For any number α ∈ F and for any element x ∈ V , the element αx ∈ V is defined (the product of
the element x and the number α). Moreover,

1) α(βx) = (αβ)x,
2) There exists a number 1 in the field F such that 1 · x = x for any x ∈ V ,
3) (α + β)x = αx + βx,
4) α(x + y) = αx + αy.

Example 7.2.2.

1) The real line R with ordinary operations of addition and product by a real number is a linear space
over the field F = R.

2) The collection of all possible n-tuple of real numbers x = (x1 , . . . , xn ) with the following product
and addition
(x1 , . . . , xn ) + (y1 , . . . , yn ) = (x1 + y1 , . . . , xn + yn ),
α(x1 , . . . , xn ) = (αx1 , . . . , αxn ),
is a linear space over the field F = R. This space is denoted Rn . Analogously, one can define the
space Cn over the field F = C.
7.2. LINEAR NORMED SPACES 127

3) The set of all (real or complex) continuous functions on [a, b] with ordinary addition and multiplica-
tion by a number (real or complex) is a linear space over the field R or C, denoted C[a, b].

4) The space lp , p > 1, whose elements are sequences of numbers (real or complex)

x = (x1 , . . . , xn , . . .)

satisfying the condition

∞
X
|x|p < +∞,
k=1

with operations

(x1 , . . . , xn , . . .) + (y1 , . . . , yn , . . .) = (x1 + y1 , . . . , xn + yn , . . .),

α(x1 , . . . , xn , . . .) = (αx1 , . . . , αxn , . . .),

is a linear space over the field R or C. The fact that the sum of two element in lp belongs lp , follows
from the Minkowski inequality (7.1.9).

5) The set of all convergent (real or complex) sequences with coordinate-wise summation and multipli-
cation by a (real or complex) number is a linear space.

6) The set of all (real or complex) sequences convergent to 0 with coordinate-wise summation and
multiplication by a (real or complex) number is a linear space.

7) The set of all bounded (real or complex) sequences with coordinate-wise summation and multiplica-
tion by a (real or complex) number is a linear space.

8) The set of all (real or complex) sequences (R∞ or C∞ ) with coordinate-wise summation and multi-
plication by a (real or complex) number is a linear space.

Definition 7.2.3. Two linear spaces V and V 0 over a field F are called isomorphic if there exists a one-
to-ne correspondence between their elements which agrees with the linear operations in V and V 0 . This
means that from x ↔ x0 and y ↔ y 0 (x, y ∈ V , x0 , y 0 ∈ V 0 ) it follows that

x + y ↔ x0 + y 0

and
αx ↔ αx0 ∀α ∈ F.

Isomorphic spaces can be considered as different realizations of the same linear space. For example, the
space Rn (respectively, Cn ) is isomorphic to the space Rn−1 [x] (respectively, Cn−1 [x]) of all polynomials
of degree n − 1 with real (complex) coefficients.

Definition 7.2.4. Elements x1 , x2 , . . . , xm of a linear space V are called linearly dependent if there exist
m
P
numbers αk ∈ F, k = 1, . . . , m, such that |αk | =
6 0 and
k=1

α1 x1 + α2 x2 + · · · + αm xm = 0. (7.2.1)

Otherwise, the elements x1 , x2 , . . . , xm are called linearly independent. In other word, the elements
x1 , x2 , . . . , xm are linearly independent if the identity (7.2.1) implies α1 = α2 = . . . = αm = 0.

Definition 7.2.5. An infinite system of elements x1 , x2 , . . . , of a linear space V is called linearly inde-
pendent if any its finite subsystem is linearly independent.
128 CHAPTER 7. LINEAR SPACES

If in the linear space V one can find n linearly independent elements, and any n + 1 its elements
are linearly dependent, then the space V is said to have dimension n. If in the space V one can find
arbitrary many linearly independent elements, then the space V is called infinitely dimensional. Any
linearly independent system with n elements of n-dimensional space V is called a basis of the space V .
It is easy to show that the spaces Rn and Cn are n-dimensional.
Definition 7.2.6. A linear space V over a field F is called normed if for any x ∈ V there exists the number
kxk called the norm of the element x, satisfying the following conditions:
1) kxk > 0 (= 0 if, and only if x = 0),
2) kx + yk 6 kxk + kyk (triangle inequality),
3) kαxk = |α| · kxk,
for all x, y ∈ V , and for any α ∈ F.
Note that the function ρ(x, y) = kx − yk, x, y ∈ V , is a metric, since it satisfies the following properties
1) ρ(x, y) > 0 (= 0 if, and only if, x = y),
2) ρ(x, y) = ρ(y, x),
3) ρ(x, z) 6 ρ(x, y) + ρ(y, z).
Thus, any linear normed space is a metric space!
Theorem 7.2.7. Let V be a linear normed space, and let a sequence {xn }∞
n=1 ⊂ V is convergent to x ∈ V
in the norm of the space V . Then kxn k → kxk as n → ∞.
Proof. From the triangle inequality we obtain the inequalities
kxn k 6 kxn − xk + kxk, kxk 6 kxn − xk + kxn k,
which imply
|kxn k − kxk| 6 kxn − xk,
as required, since kxn − xk → 0 as n → ∞ by assumption.
Converse, of course, is not true. For example, the sequence of real numbers xn = (−1)n−1 , n ∈ N, has
no limit but the sequence |xn | = 1, n ∈ N, converges to 1.
Definition 7.2.8. A linear normed space V is called complete w.r.t. its norm if any Cauchy sequence of
its elements converges, that is, from kxn − xm k −→ 0 it follows that the sequence (xn )∞
n=1 converges.
n,m→∞

Definition 7.2.9. Linear normed spaces complete w.r.t. their norm are called Banach spaces.
Example 7.2.10.
1) The real line R is a complete linear normed space with the norm kxk = |x|.
2) The spaces Rn and Cn become complete linear normed spaces of elements x = (x1 , . . . , xn ) if we set

n
!1
X p
kxkp = |xk |p , 1 6 p < +∞.
k=1

We can also introduce the norm

kxk∞ = max |xk |.
16k6n

It is easy to show that all the introduced norms are equivalent, that is for any p1 , p2 ∈ [1, +∞] there
exist numbers cp1 ,p2 and Cp1 ,p2 such that
cp1 ,p2 kxkp1 6 kxkp2 6 Cp1 ,p2 kxkp1 .
7.2. LINEAR NORMED SPACES 129

3) The set C[a, b] of (real or complex) continuous functions on [a, b] is a complete normed linear space
with the norm
kf k = max |f (x)|.
a6x6b

Indeed, kf k > 0 for any f ∈ C[a, b]. Moreover, if kf k = 0, it is clear that f (x) ≡ 0 on [a, b].
Furthermore, for f, g ∈ C[a, b] and for any x ∈ [a, b], one has

|f (x) + g(x)| 6 |f (x)| + |g(x)| 6 kf k + kgk.

Since it is true for any x ∈ [a, b], this is true for kf + gk. Thus, kf k satisfies the property 2) of norms.
The property 3) is obvious in this case.
Furthermore, if (fn )∞
n=1 ⊂ C[a, b] is a Cauchy sequence: kfn − fm k −→ 0, then we have that
n,m→∞
for any c ∈ [a, b] the sequence (fn (c))∞
n=1is a Cauchy sequence of real (complex) numbers, so it is
convergent as R and C are complete spaces. Thus, the sequence (fn (x))∞ n=1 converges pointwise to
a function f (x) on [a, b]. Since (fn (x))∞
n=1 is a Cauchy sequence, we have that

∀ε > 0 ∃n0 ∈ N : ∀n > n0 ∀p ∈ N max |fn+p (x) − fn (x)| < ε.

x∈[a,b]

If p → ∞, we obtain

∀ε > 0 ∃n0 ∈ N : ∀n > n0 max |fn (x) − f (x)| 6 ε.

x∈[a,b]

Thus, the sequence (fn (x))∞

n=1 uniformly converges to f (x) on [a, b], so f ∈ C[a, b] as a uniform limit
of a sequence of continuous functions. Consequently, C[a, b] is complete.
4) The space lp , p > 1, is a complete linear normed space with the norm

∞
!1
X p
p
kxkp = |x| .
k=1

Indeed, it is clear that the function kxkp : lp 7→ [0, +∞) satisfies the properties 1) and 3) of norms.
Moreover, by Minkowsky’s inequality (7.1.9) it satisfies the property 2), so it is a norm by Defini-
tion 7.2.6.
Suppose (x(n) )∞
n=1 is a Cauchy sequence of elements lp , that is,
∞
(n) (m)
X
|xk − xk |p → 0 as n, m → +∞. (7.2.2)
k=1

This, in particular, means that for any k ∈ N

(n) (m)
|xk − xk | → 0 as n, m → +∞,
(n)
so for any k ∈ N the sequence (xk )∞ n=1 converges to a number xk . Thus, we have an infinite sequence
x = (xk )∞
k=1 . It is left to prove that x ∈ lp . In its turn it follows from Minkowski inequality. In fact,
from (7.2.2) we obtain that for any M ∈ N and for any ε > 0, there exists a number n0 ∈ N such
that
M
(n) (m)
X
|xk − xk |p < ε n, m > n0 .
k=1

As m → ∞, we obtain for any M ∈ N,

M
(n)
X
|xk − xk |p 6 ε n > n0 .
k=1
130 CHAPTER 7. LINEAR SPACES

So when M → ∞, one has

∞
(n)
X
|xk − xk |p 6 ε n > n0 .
k=1
Now by Minkowski’s inequality (7.1.9), we get

∞
!1 ∞
!1 ∞
!1
p p p
(n) (n)
X X X
p p
|xk | 6 |xk − xk | + |xk |p .
k=1 k=1 k=1

Since both series in the right-hand side of this inequality are convergent, the series in its left-hand
side is convergent as well, so x ∈ lp .
However, not every linear normed space is complete.
Example 7.2.11. Consider the set of all functions continuous on [−1, 1], and introduce the norm
Z Z1
kf k = (L) |f (x)|dµ = (R) |f (x)|dx. (7.2.3)
[−1,1] −1

It is clear that the number kf k satisfies all the properties of norms. Thus, we get a linear normed space
that we denote as R[−1, 1].
Let (fn (x))∞
n=1 ⊂ R[−1, 1] be the following sequence

1
 −1 if − 1 6 x 6 − n ,

1 1
fn (x) = nx if − n 6 x 6 n ,
1

1 if n 6 x 6 1.


It is easy to see that

Z1
1 1
kfn (x) − fm (x)k = |fn (x) − fm (x)|dx = − →0 as n, m → ∞,
m n
−1

and
Z1
kfn (x) − sgn(x)k = |fn (x) − sgn(x)|dx → 0 as n → ∞.
−1

At the same time, for any function f continuous on [−1, 1] one has
0 < kf (x) − sgn(x)k 6 kf (x) − fn (x)k + kfn (x) − sgn(x)k,
so the sequence (fn )∞
n=1 does not have continuous limit on [−1, 1]. Thus, the space R[−1, 1] is not complete.

As we will se below, the completion of the space R[−1, 1] (that is, roughly speaking, adding all the
limits of sequences of elements of the space R[−1, 1]) is exactly the space L[−1, 1] with the norm (7.2.3).

7.3 Spaces Lp
Let (X, M(X), µ) be a measure space, and the measure µ is supposed to be σ-additive and complete.
a.e.
Recall that two functions f, g ∈ S(X)) are called equivalent on X and denoted f ∼ g if f = g on X.
Let us split the set S(X) into equivalence classes. It is possible since the equivalence relation f ∼ g is
reflexive (f ∼ f ), symmetric (f ∼ g =⇒ g ∼ f ), and transitive (f ∼ g, g ∼ h =⇒ f ∼ h). In what
follows, we denote the class of functions equivalent to a function f by the same letter f , and do not
distinguish functions that belong to the same equivalence class. So dealing with an equivalence class we
pick a representative function of this class (it may be any function in the class) and deal with this function.
7.3. SPACES LP 131

Definition 7.3.1. The collection of all equivalence classes of measurable on X functions satisfying the
inequality Z
|f (x)|p dµ < +∞, p > 1, (7.3.1)
X

is defined to be a normed linear space Lp (X, M, µ) with the norm

 1
Z p

kf kp :=  |f (x)|p dµ (7.3.2)

X

Remark 7.3.2. The norm defined in (7.3.2) is finite and does not depend on the representative functions
of equivalence classes, since all function from one equivalence class have the same integral over X.
Remark 7.3.3. In what follows we denote the class Lp (X, M, µ) as Lp (X) if it does not lead to any
disambiguations.
It is clear that Lp (X) is a linear space. Indeed, for any number λ (real or complex) and for any
f ∈ Lp (X), the class of functions λf belongs to the space Lp (X) by the property 6) of integral of
measurable functions. Moreover, from the Minkowski inequality (7.1.7). it follows that f + g ∈ Lp
whenever f, g ∈ Lp (X).
It is also easy to check that the number kf kp satisfies all the properties of norms. Indeed, kf kp > 0 for
a.e.
any f ∈ Lp (X). Moreover, if kf kp = 0, then f = 0 on X, so it is belong to the class of functions which
has the function f (x) ≡ 0 as a representative function. Thus, the norm satisfies the property 1) of norms.
The property 3) is obvious, and the property 2) follows from the Minkowski inequality (7.1.7).
Definition 7.3.4. The convergence in the norm of the space Lp (X) is called the convergence in the pth
mean. For p = 1, it is usually called the mean convergence, and for p = 2 it is called the mean-square
Lp
convergence. If a sequence (fn )∞
n=1 of functions in Lp (X) converges to f0 ∈ Lp (X), we denote fn −→ f0 .

Theorem 7.3.5 (Uniqueness of limit). Any convergent sequences in Lp (X), p > 1, has a unique limit.
Lp Lp
Proof. Let (fn )∞
n=1 be such that fn −→ f and fn −→ g. Then by Minkowski’s inequality (7.1.7) we have
n→∞ n→∞

kf − gkp = k(f − fn ) + (fn − g)kp 6 kf − fn kp + kg − fn kp → 0 as n → ∞,

a.e.
so kf − gkp = 0. Thus, f = g on X, therefore, they belong to one class and are considered as one element
in the space Lp (X) by Definition 7.3.1.
By Theorem 7.2.7, we have
Lp
Theorem 7.3.6 (Continuity of norm). If fn −→ f0 , p > 1, then kfn kp −→ kf kp .
n→∞ n→∞

Lp
Theorem 7.3.7. If fn −→ f0 , p > 1, then (fn )∞
n=1 is a Cauchy sequence w.r.t. the norm k · kp .
n→∞

Proof. By assumption, for any ε > 0, there exists a number n0 ∈ N such that for any n > n0 one has
ε
kfn − f kp < .
2
Let n, m > n0 . Then from Minkowski’s inequality (7.1.7), it follows that

kfn − fm kp = k(fn − f ) + (f − fm )kp 6 kfn − f kp + kf − fm kp < ε,

as required.
132 CHAPTER 7. LINEAR SPACES

The converse statement is not so easy to prove and means that the space Lp (X) is Banach, provided
µX < +∞. To prove this statement we need the following lemma which can also be helpful in some other
cases.

Lemma 7.3.8. If f ∈ S(X) and f ∈ Lp (X) for some number p > 1, then f ∈ Lp1 (X) for any p1 ∈ [1, p).
Moreover, the following inequality holds
 1  1
p1 p
Z p−p1 Z
 |f (x)|p1 dµ 6 (µX) p1 p  |f (x)|p dµ (7.3.3)
X X

Proof. Let us introduce the numbers p0 and q 0 as follows

p p
p0 = , q0 = .
p1 p − p1
1 1
It is easy to see that + 0 = 1.
p0 q Z
Applying the Hölder inequality with powers p0 and q 0 to the integral |f (x)|p1 dµ, we obtain
X

  p1   p−p1   p1
p p p
Z Z Z p Z p p−p1 Z
p1 · p
|f (x)|p1 dµ = |f (x)|p1 · 1dµ 6  |f (x)| 1  ·  1 p−p 1  = (µX) p  |f (x)| p
,
X X X X X

1
that gives us (7.3.3) after taking powering in .
p1
Theorem 7.3.9 (Riesz-Fischer). The space Lp (X), p > 1, µX < +∞, is a complete linear normed space,
that is, it is Banach w.r.t. the norm k · kp .

Proof. Let (fn )+∞

n=1 be a Cauchy sequence of functions in the space Lp . For any ε > 0, there exists a
number n0 ∈ N such that Z
|fm (x) − fn (x)|p dµ < εp , m, n > n0 . (7.3.4)
X

If p > 1,then from (7.3.3) with p1 = 1 we obtain

Z p−1
|fm (x) − fn (x)|dµ < (µX) p · ε, m, n > n0 .
X

Consequently, if (fn )∞
n=1 is a Cauchy sequence in Lp (X), p > 1, it is a Cauchy sequence in L1 (X).
1
Furthermore, for the number ε = k there exists a number n0 (k) such that
2
Z
1
|fm (x) − fn (x)|dµ < k , m, n > n0 (k).
2
X

For k = 1, we take n1 > n0 (1), then for k = 2, we take n2 > max{n1 , n0 (2)}, etc. When we have the
numbers n1 < n2 < · · · < nk−1 , we take nk > max{nk−1 , n0 (k)}, etc. Thus, we obtain a subsequence
(fnk )∞
k=1 such that Z
1
|fnk+1 (x) − fnk (x)|dµ < k , k = 1, 2, . . . .
2
X
7.3. SPACES LP 133

Introduce the following function

∞
X
g(x) = |fn1 (x)| + |fnk (x) − fnk−1 (x)|. (7.3.5)
k=2

By Lemma 7.3.8, all the terms of this series are integrable functions. According to Levy’s Theorem 5.2.3,
one has
Z Z ∞ Z Z ∞
X X 1
g(x)dµ = |fn1 (x)|dµ + |fnk (x) − fnk−1 (x)|dµ < |fn1 (x)|dµ + < +∞
2k−1
X X k=2 X X k=2

By the property 6) of integral of nonnegative functions, the function g(x) is finite a.e. on X. In other
word, the series (7.3.5) converges a.e. on X. Therefore, the series
∞
X
fn1 (x) + (fnk (x) − fnk−1 (x))
k=2

also converges a.e. on X, and its sum f0 (x) is a measurable and a.e. finite on X. Moreover, we have
 
X k
f0 (x) = lim fn1 (x) + (fnj (x) − fnj−1 (x)) = lim (fnk (x)).
k→+∞ k→+∞
j=2

Thus, the subsequence (fnk )∞

k=1 converges to f0 (x) a.e. on X.
Put now m = nk in the inequality (7.3.4). Since

|fnk (x) − fn (x)|p −→ |f0 (x) − fn (x)|p a.e. on X,

k→∞

by Corollary 5.2.8 Z
|f0 (x) − fn (x)|p dµ < εp , n > n0 . (7.3.6)
X
Lp
Now we prove that f0 ∈ Lp (X) and that fn −→ f0 . Indeed, for any number n > n0 , by the Minkowski
inequality (7.1.7) and by (7.3.6), we have
 1  1
Z p Z p
 |f0 (x)|p dµ =  |(f0 (x) − fn (x)) + fn (x)|p dµ 6
X X

 1  1
Z p Z p
p p
6  |f0 (x) − fn (x)| dµ  +  |fn (x)| dµ 6 ε + Mn < +∞,
X X

 1
Z p
p
where Mn =  |fn (x)| dµ . Thus, f0 ∈ Lp (X).
X
Now from (7.3.6) it follows that

kfn − f0 kp < ε, n > n0 ,

which means that the sequence (fn )∞

n=1 converges to f0 in p
th
mean.
So, we get that a Cauchy sequence in Lp converges, as required.
134 CHAPTER 7. LINEAR SPACES

Note that the condition µX < +∞ was used only to establish that any Cauchy sequence in Lp (X),
p > 1, is a Cauchy sequence in L1 (X). This means that we can omit this condition for the space L1 (X).

Corollary 7.3.10. Let X be the space Rn or any subset of Rn with finite or infinite measure. Then the
space L1 (X) is complete w.r.t. the norm k · k1 .

From the proof of Theorem 7.3.9 we obtain the following fact.

Corollary 7.3.11. Let X be the space Rn or any subset of Rn with finite or infinite measure. If (fn )∞
n=1
converges to f in L1 (X), then there exists a subsequence (fnk )∞
k=1 such that

a.e.
fnk −−→ f on X.

7.4 Relations between different types of convergence

Let (X, M, µ) be a measure space with µX < +∞, and let (fn )∞
n=1 be a sequence of measurable functions
on X, and f0 ∈ S(X). We know the following kinds of convergence.

1) Uniform convergence on X,

2) Convergence a.e. on X (pointwise convergence),

3) Convergence in the space Lp (convergence in pth mean),

4) Convergence in measure.

Let us consider interrelations between these kinds of convergence.

I. Uniform convergence implies all other kinds of convergence.

a) If (fn (x))∞
n=1 converges to f0 (x) uniformly on X, then it converges to f0 (x) for any x ∈ X, so it
converges pointwisely (almost everywhere) on X.

b) By definition of the uniform convergence, for any ε > 0 there exists a number n0 ∈ N such that for
any n > n0 and for any x ∈ X the following inequality holds

|fn (x) − f0 (x)| < ε.

This implies
 1
p
1
Z
kfn − f0 kp =  |fn (x) − f0 (x)|p  < ε(µX) p ,
X

Lp
so fn −→ f0 .

c) From the definition of the uniform convergence we get

X(|fn − f0 | > ε) = ∅
µ
for any n > n0 . Therefore, fn −→ f0 .

However, convergence in measure, converges in pth mean, and point-wise convergence do not imply the
uniform convergence, generally speaking.
7.4. RELATIONS BETWEEN DIFFERENT TYPES OF CONVERGENCE 135

Example 7.4.1. Let X = [0, 1], and µ be the Lebesgue measure. Consider the sequence fn (x) = xn .

This sequence converges to f0 (x) ≡ 0 a.e. on X (everywhere but the point x = 1). Moreover, for
any p > 1,
1
1

Z p 
n p 1 p
kfn − f0 kp =  |x − 0| = −→ 0
np + 1 n→∞
X
Lp
so fn −→ f0 . Additionally, for any 0 < σ < 1, one has
√
µX(|fn − f0 | > σ) = 1 − n
σ −→ 0,
n→∞
µ
therefore, fn −→ f0 . However, fn (x) does not converge to f0 (x) uniformly on X, since
sup |fn (x) − f0 (x)| = sup |xn | = 1 6−→ 0.
x∈[0,1] x∈[0,1] n→∞

II. Pointwise convergence and convergence in pth mean are incomparable.

Example 7.4.2. Let X = [0, 1], and µ be the Lebesgue measure, and let
 1
 np , 0 6 x 6 1 ,

fn (x) = n
 0, 1
 < x 6 1.
n
a.e.
Clearly, fn −→ f0 where f0 (x) ≡ 0. However,
 1  p1
 1
Z p Zn
kfn − f0 kp =  |fn (x) − f0 (x)|p dµ = 
 
6−→ 0,
 ndx = 1 n→∞

X 0

so fn 6→ f0 in Lp (X).
Example 7.4.3. Let X = [0, 1], and µ be the Lebesgue measure, and let
  
 1, x ∈ m − 1 , m ,

 k k

ϕkm (x) =
m − 1 m
 0, x 6∈ , ,


k k
where k = 1, 2, . . ., and m = 1, 2, . . . , k. Let us represent the functions φl,j as one sequence as follows:
f1 (x) = ϕ11 (x), f2 (x) = ϕ21 (x), f3 (x) = ϕ22 (x), f4 (x) = ϕ31 (x), f5 (x) = ϕ32 (x), . . . ,
so
l−1
X
fn (x) = ϕl,j (x), n=j+ k, l = 1, 2, . . . , j = 1, 2, . . . , l.
k=1

The sequence fn converges in pth mean on [0, 1] to the function f0 (x) ≡ 0, since
1  m
1

p p
Z Zk  1
1 p
|fn (x) − f0 (x)|p dµ = 
 
kfn − f0 kp =  1 · dx
 = −→ 0.
 
 k n→∞
[0,1] m−1
k

Obviously, k → ∞ ⇐⇒ n → ∞. Thus, the sequence fn converges in pth mean on [0, 1] to the function
f0 (x) ≡ 0, but it does not converge at any point of the interval [0, 1] as it was established in Example 4.3.11.
136 CHAPTER 7. LINEAR SPACES

III. Pointwise convergence implies convergence in measure. Converse is not true, generally speaking.
This was established in Theorem (4.3.10) and Example 4.3.11.
IV. Convergence in pth mean with exponent p2 implies convergence in pth mean with exponent p1 for any
1 6 p1 < p2 .
This follows from the estimate (7.3.3) proved in Lemma 7.3.8.
The converse statement is not true, generally speaking, as the following example shows.
Example 7.4.4. Let X = [0, 1], and µ be the Lebesgue measure, and let
 n 1

 p 1
 , 06x6 ,
fn (x) = ln n n
1
0, < x 6 1,


n
for some p > 1. Put f0 (x) ≡ 0.
Since we have

 1  1
 p1
p 1
Z Zn 
p
 n  1 p
kfn − f0 kp =  |fn (x)| dµ =  dx = −→ 0,
 
 ln n  ln n n→∞
[0,1] 0

Lp
so fn −→ f0 , but fn (x) 6→ f0 in pth
1 mean for any p1 > p, because

 1  p1
1
Zn 1
  n  pp1  p1 −p

1 p
kfn − f0 kp1 = dx = n p1 p · −→ +∞.
 ln n  ln n n→∞
0

V. Convergence in pth mean for any p > 1 implies convergence in measure.

Since convergence in pth mean for any p > 1 implies convergence in mean (that is, with the exponent
1 L µ
p = 1), it is sufficient to show that fn −→ f0 implies fn −→ f0 .
L1
So let fn −→ f0 . For any σ > 0, δ > 0, and ε = σδ, there exists a natural number n0 = n0 (ε) such that
Z
kfn − f0 k1 = |fn (x) − f0 (x)|dµ < ε.
X

By Chebyshev’s inequality (see the property 7) of integral of nonnegative functions), one has for any n > n0
Z
1 ε
µX(|fn − f0 | > σ) 6 |fn (x) − f0 (x)|dµ < = δ,
σ σ
X

µ
so fn −→ f0 .
The following example shows that converse is not true.
Example 7.4.5. Let X = [0, 1] and µ be the Lebesgue measure. Suppose that

 n, 0 6 x 6 1 ,

fn (x) = n
1
 0,
 < x 6 1.
n
7.4. RELATIONS BETWEEN DIFFERENT TYPES OF CONVERGENCE 137

The sequence fn converges to the function f0 (x) ≡ 0 in measure on X, since for any σ > 0
1
µX(|fn − f0 | > σ) 6 .
n
However, for any n ∈ N Z
kfn − f0 k1 = |fn (x) − f0 (x)|dµ = 1,
X
th
so fn 6−→ 0 in p mean for any p > 1.

7.4.1 Approximation of integrable functions

We say that a family G of integrable functions is dense in Lp (X), p > 1, if for any f ∈ Lp (X) and ε > 0,
there exists g ∈ G so that kf − gkp < ε. Fortunately, we are familiar with many families that are dense
in Lp (X), and we describe some in the theorem that follows. These are useful when one is faced with
the problem of proving some fact or identity involving integrable functions. In this situation a general
principle applies: the result is often easier to prove for a more restrictive class of functions (like the ones
in the theorem below), and then a density (or limiting) argument yields the result in general.

Theorem 7.4.6. Let X be a compact set in Rn , and let F ⊂ X be closed. If ψ is defined and continu-
ous on F , then there exists a function ϕ(x) defined on X satisfying the following properties

a) ϕ(x) is continuous on X;

b) ϕ(x) = ψ(x) for any x ∈ F ;

c) max |ϕ(x)| = max |ψ(x)|.

Proof. For the case X = [a, b], a proof can be found in the book [8, Chapter 4, § 4, Lemma 2]. A similar
proof for arbitrary compact set X ⊂ Rn is more difficult.
Another way (more topological) is to take the function
 
ρ(x, y)


 inf ψ(y) + − 1 , x ∈ X \ F,
 y∈F ρ(x, F )
ϕ(x) = (7.4.1)


x ∈ F,

ψ(x),

which satisfies all the conditions of the theorem. Here ρ(x, y) = kx − yk2 and ρ(x, F ) = inf ρ(x, y).
y∈F

Problem. Prove that the function (7.4.1) is continuous on X.

From Theorems 4.4.8 and 7.4.6 it immediately follows the following fact.

Theorem 7.4.7 (Luzin). Let X ⊂ Rn be a compact set. Then for any f ∈ S(X) and for any ε > 0 there
exists a function g continuous on X and such that µX(f 6= g) < ε.
If |f (x)| 6 M , then |g(x)| 6 M , as well.

Corollary 7.4.8 (Frechet). Let X ⊂ Rn be a compact set. Then for any f ∈ S(X) there exists sequence
(fn )∞
n=1 of continuous functions converging to f a.e. on X.

Corollary 7.4.9 (Borel). Let X ⊂ Rn be a compact set. Then for any f ∈ S(X) there exists sequence
(fn )∞
n=1 of continuous functions converging to f in measure on X.

Now we can study function families dense in the space of all integrable functions.

Theorem 7.4.10. Let X ⊂ Rn be a compact set. The following families of functions are dense in L1 (X):
138 CHAPTER 7. LINEAR SPACES

(i) The bounded measurable functions.

(ii) The simple functions.
(iii) The step functions.
(iv) The continuous functions of compact support.
Here X is supposed to be the space Rn or a (bounded or unbounded) subset of the space Rn .
Recall that the set suppf = {x ∈ X : f (x) 6= 0} is called the support of the function f (note that
support of a function is always a closed set by definition).
Proof. Suppose that f ∈ L1 (X). By the absolute continuity of Lebesgue integral (see the property 10) of
integral of measurable functions), for any ε > 0 there exists δ > 0 such that
Z
|f |dµ < ε
Xδ

for any Xδ with µXδ < δ.

Now by the property 4) of integral of measurable functions, we have µX(f = ±∞) = 0. By Theo-
rem 4.4.4, for any δ > 0 there exists a bounded function g ∈ S(X) such that µX(f 6= g) < δ, and we set
g(x) = 0 for any x ∈ X(f 6= g). So finally we get
Z Z
|f − g|dµ = |f |dµ < ε.
X X(f 6=g)

Thus, kf − gk1 < ε, as required.

Let again f be an integrable function on X. We can always represent f as follows f = f + − f − , where
f and f − are nonnegative functions defined in Definition 4.2.3, and it now suffices to prove the theorem
+

when f > 0.
For (ii), Theorem 4.4.2 guarantees the existence of a sequence (hk )∞
k=1 of non-negative simple functions
that increase to f pointwise. By the Lebesgue Dominated Convergence Theorem 5.3.13, we then have that
Z Z
f (x)dµ = lim hn (x)dµ,
n→∞
X X
so Z
lim (f (x) − hn (x))dµ = 0,
n→∞
X
that is equivalent to
kf − hn k1 → 0 as n → ∞,
since f − hn > 0 for any n ∈ N. Thus, there are simple functions that are arbitrarily close to f in the L1
norm.
For (iii), we first note that by (ii) it suffices to approximate simple functions by step functions. Then,
we recall that a simple function is a finite linear combination of characteristic functions of sets of finite
measure, so it suffices to show that if A is such a measurable set, then there is a step function ϕ so that
kχA − ϕk1 is small. However, since A is measurable, then by Definition 3.4.1 for any ε > 0 there exists an
elementary set B such that
µ∗ (A∆B) = µ(A∆B) < ε.
m
P
The elementary set B = Kl is a finite sum of bricks that we can always consider as almost disjoint (see
l=1
m
P
Definition 3.5.8). Thus χA (x) and ϕ(x) = χKl (x) differ at most on a set of measure ε, and as a result
k=1
we find that kχA − ϕk1 < ε.
7.4. RELATIONS BETWEEN DIFFERENT TYPES OF CONVERGENCE 139

By (iii), it suffices to establish (iv) when f is the characteristic function of a brick. In the one-
dimensional case, where f is the characteristic function of an interval [a, b], we may choose a continuous
piecewise linear function g defined by
(
1, x ∈ [a, b],
g(x) =
0, x 6∈ [a − ε, b + ε],

and with g linear on the intervals [a − ε, a] and [b, b + ε]. Then kf − gk1 < ε. In n dimensions, it suffices to
note that the characteristic function of a rectangle is the product of characteristic functions of intervals.
Then, the desired continuous function of compact support is simply the product of functions like g defined
above.

Remark 7.4.11. Note that there exist some other ways to approximate the characteristic function of
measurable set by continuous functions with compact support. For further details, see [5, 4].

Remark 7.4.12. In fact, bounded measurable functions are dense in the space Lp (X). To prove this it
is sufficient to note that if f ∈ Lp (X), then f p ∈ L1 (X), and use absolute continuity of Lebesgue integral
for the function f p , at the same time, approximating the function f by a bounded measurable function.

As an application of Theorem 7.4.10, we now examine how continuity properties of f are related to the
way the translations fc vary with c (see Definition 5.3.14). Note that for any given x ∈ Rn , the statement
that fc (x) → f (x) as c → 0 is the same as the continuity of f at the point x.
However, a general f ∈ L1 (Rn ) may be discontinuous at every x, even when corrected on a set of
measure zero. Nevertheless, there is an overall continuity that an arbitrary f ∈ L1 (Rn ) enjoys, one that
holds in the norm.

Proposition 7.4.13. Let f ∈ L1 (Rn ). Then

kf − fc k1 → 0 as c → 0.

The proof is a simple consequence of the approximation of integrable functions by continuous functions
of compact support as given in Theorem 7.4.10.

Proof. By Theorem 7.4.10 for any ε > 0, one can find a continuous function g of compact support such
that kf − gk1 < ε. Now
fc − f = (gc − g) + (fc − gc ) + (g − f ).
Since kfc − gc k1 = kf − gk1 < ε and since g is continuous of compact support we have that clearly
Z
kgc − gk1 = |g(x − c) − g(x)|dx → 0 as c → 0.
Rn

So if |c| < δ, where δ is sufficiently small, then kgc − gk1 < ε, and as a result kfc − f k1 < 3ε, whenever |c| <
δ.

The results above for L1 (Rn ) lead immediately to an extension in which Rn can be replaced by any
fixed subset X of positive measure. In fact, if X is such a subset, we can define L1 (X) and carry out the
arguments that are analogous to L1 (Rn ). Better yet, we can proceed by extending any function f on X
by setting fe = f on X and fe = 0 on Rn \ X, and defining kf kL1 (X) = kfekL1 (Rn ) . Moreover, for measure
spaces (X, M, µ) with compact X, the statement of Theorem 7.4.10 can be partially established for the
spaces Lp (X) with p > 1.

Theorem 7.4.14. The following families of functions are dense in Lp (X), p > 1, where X ⊂ Rn is a
compact set:

(i) The bounded measurable functions.

140 CHAPTER 7. LINEAR SPACES

(ii) The continuous functions.

Proof. The proof of the statement (i) is similar to the one in Theorem 7.4.10. Indeed, if f ∈ Lp (X), then
by the absolute continuity of Lebesgue integral (see the property 10) of integral of measurable functions),
for any ε > 0 there exists δ > 0 such that
Z
|f |p dµ < εp
Xδ

for any Xδ with µXδ < δ.

Now by the property 4) of integral of measurable functions, we have µX(f = ±∞) = 0. By Theo-
rem 4.4.4, for any δ > 0 there exists a bounded function g ∈ S(X) such that µX(f 6= g) < δ, and we set
g(x) = 0 for any x ∈ X(f 6= g). So finally we get
Z Z
p
|f − g| dµ = |f |p dµ < εp .
X X(f 6=g)

Thus, kf − gk1 < ε, as required.

Now it is enough to approximate a bounded measurable function in the norm Lp . Let f ∈ S(X) be
bounded, so |f (x)| 6 M . By Theorem 7.4.7, for any ε > 0 there exists a continuous function g s.t.

εp
µX0 = µX(f 6= g) < .
(2M )p

Then we have Z Z
kf − gkpp = |f − g|p dµ = |f − g|p dµ 6 (2M )p · µX0 < εp .
X X0

From this theorem and from Weierstrass’ theorems it easy to obtain the following facts.

Corollary 7.4.15. The families of polynomials and trigonometric polynomials are dense in Lp [−π, π],
p > 1.

Recall that trigonometric polynomials are functions of the form

n
a0 X
Tn (x) = + (am cos mx + bm sin mx).
2 m=1

7.5 Basic theory of inner product spaces

In this section we introduce a more narrow class of linear normed spaces, namely, the inner product spaces.
Let X be a linear space over the field1 F = C.

Definition 7.5.1. The function (·, ·) : X × X 7→ C is called inner (scalar) product of the elements of the
space X if it possesses the following properties:

1) for any x ∈ X, (x, x) > 0 (= 0 ⇐⇒ x = 0);

2) (x, y) = (y, x) for any2 x, y ∈ X;

3) (αx + βy, z) = α(x, z) + β(y, z) for any x, y, z ∈ X and α, β ∈ C.

1 In fact, the theory discussed here can be easily transferred to the case F = R.
2 In the case F = R, this property looks as (x, y) = (y, x)
7.5. BASIC THEORY OF INNER PRODUCT SPACES 141

From the properties of the inner product it follows that3

(x, αy) = (αy, x) = α(y, x) = α · (y, x) = α(x, y).
Definition 7.5.2. A linear space X over a field F with an inner product is called Euclidean if F = R, and
is called unitary if F = C.
Let X be a Euclidean (unitary) space. For any x ∈ X, introduce the following number
p
kxk = (x, x). (7.5.1)
This number is denoted as norm. And we have to prove that this is a norm, indeed. To do this, we need
the following inequality:
|(x, y)| 6 kxk · kyk, ∀x, y ∈ X. (7.5.2)
This inequality is called Cauchy–(Bunyakovsky)–Schwarz inequality. For finite sums it was proved by
Cauchy in 1821. For integrals, it was proved by Bunyakovsky in 1859 and by Schwarz in 1888.
To prove the inequality (7.5.2), let us represent the inner product as follows (x, y) = |(x, y)|eiθ . Then
for any t ∈ R we have
(te−iθ x + y, te−iθ x + y) = (te−iθ x, te−iθ x) + (te−iθ x, y) + (y, te−iθ x) + (y, y) =

= teiθ · te−iθ · (x, x) + te−iθ (x, y) + tiθ (y, x) + kyk2 = t2 kxk2 + kyk2 + t · |(x, y)| + t · |(y, x)| =

= t2 kxk2 + 2t · |(x, y)| + kyk2 > 0.

The quadratic polynomial is nonnegative if, and only if, its discriminant is non-positive. In our case, the
discriminant has the form |(x, y)|2 − kxk2 · kyk2 , and its non-positivity implies (7.5.2).
Prove now that the number (7.5.1) is a norm.
1) kxk > 0 (= 0 ⇐⇒ x = 0).
This property follows from the first property of inner products.
2) The triangle inequality
By the Cauchy–Bunyakovsky-Schwarz inequality, we have
kx + yk2 = (x + y, x + y) = kxk2 + (x, y) + (x, y) + kyk2 = kxk2 + 2 Re(x, y) + kyk2 6

6 kxk2 + 2|(x, y)| + kyk2 6 kxk2 + 2kxk · kyk + kyk2 = (kxk + kyk)2 .
Consequently,
kx + yk 6 kxk + kyk, (7.5.3)
as required.
p p p
3) kαxk = (αx, αx) = α · α(x, x) = |α| · kxk2 = |α| · kxk.
Thus, we have that any Euclidean (unitary) space is a linear normed space.
Proposition 7.5.3. The inner product of a Euclidean (unitary) space X is a continuous function of its
variables.
Proof. Indeed, let xn → x and yn → y by norm in X. Then by Theorem 7.2.7, kyn k → kyk as n → ∞, so
there exists a constant C > 0 such that kyn k 6 C for all n ∈ N. Thus, by Cauchy-Bunyakovsky-Schwarz
inequality (7.5.2) we have
|(xn , yn ) − (x, y)| 6 |(xn , yn ) − (x, yn )| + |(x, yn ) − (x, y)| = |(xn − x, yn )| + |(x, yn − y)| 6

6 kxn − xk · kyn k + kyn − yk · kxk 6 Ckxn − xk + kyn − yk · kxk −→ 0.

n→∞

Thus, lim (xn , yn ) = (x, y), as required.

n→∞
3 If F = R, then we have (x, αy) = α(x, y).
142 CHAPTER 7. LINEAR SPACES

Definition 7.5.4. For any two non-zero elements x, y of a Euclidean space X, cosine of the angle α
between x and y is defined by the following formula
(x, y)
cos α = .
kxk · kyk
π
It is clear that if α = , then (x, y) = 0. Thus, we can introduce for both Euclidean and unitary spaces
2
the following notion.
Definition 7.5.5. Two non-zero elements x and y of a Euclidean (unitary) space are called orthogonal if
(x, y) = 0. We denote this as x ⊥ y
For orthogonal elements of Euclidean and inner product spaces there exists the Pythagorean theorem.
Theorem 7.5.6 (Pythagorean theorem). If x ⊥ y, then
kx + yk2 = kxk2 + kyk2 .
Proof. Indeed, we have
kx + yk2 = (x + y, x + y) = kxk2 + (x, y) + (y, x) + kyk2 = kxk2 + (x, y) + kyk2 = kxk2 + kyk2 .

Definition 7.5.7. Let X be a Euclidean or unitary space. A system (en )ω n=1 of elements of the space X
is called orthogonal if (en , em ) = 0 whenever n 6= m. Here ω can be finite or infinite (if the space X is
infinite-dimensional).
The system (en )ω
n=1 is called orthonormal if
(
1 if m = n,
(en , em ) =
0 if m 6= n.

Definition 7.5.8. Let (en )ω

n=1 , 1 6 ω 6 ∞, be an orthonormal system of elements of a Euclidean (unitary)
space X. For any x ∈ X, then numbers
cn (x) = (x, en ), n = 1, 2, . . . , ω,
are called Fourier coefficients of the element x with respect to the system (en )ω
n=1 .

Theorem 7.5.9 (On projection). Let (ek )nk=1 be an orthonormal system of elements of a Euclidean (uni-
tary) space X. Suppose that x ∈ X and
Xn
y= ck (x)ek .
k=1
Then
x − y ⊥ y.
Proof. Note first that for any k = 1, . . . , n,
ck (x) = ck (y),
so that for 1 6 k 6 n we have
(x − y, ek ) = (x, ek ) − (y, ek ) = ck (x) − ck (y) = 0.
Therefore,
n
X n
X
(x − y, y) = (x − y, ck (x)ek ) = ck (x)(x − y, ek ) = 0,
k=1 k=1
as required.
7.5. BASIC THEORY OF INNER PRODUCT SPACES 143

Corollary 7.5.10. If (ek )nk=1 is an orthonormal system of elements of a Euclidean (unitary) space X,
then
n 2 n
X X
ak ek = |ak |2 .
k=1 k=1

Proof.

n 2 n 2 n 2 n 2
X X 2
X X
2 2
a k ek = a 1 e1 + ak ek = ka1 e1 k + ak ek = |a1 | · ke1 k + a k ek =
k=1 k=2 k=2 k=2

n
X
= |a1 |2 + |a2 |2 + · · · + |an |2 = |ak |2 .
k=1

Theorem 7.5.11 (Bessel’s inequality). Let X be an infinitely dimensional Euclidean or unitary space. If
(en )∞
n=1 is an orthonormal system of elements of the space X, then

∞
X
|cn (x)|2 6 kxk2 ∀x ∈ X. (7.5.4)
n=1

m
P
Proof. Let y = cn (x)en . Then by Theorems 7.5.6 and 7.5.9, we have
n=1

kxk2 = kx − yk2 + kyk2 . (7.5.5)

Since kx − yk2 > 0, from (7.5.5) we obtain

m
X
|cn (x)|2 = kyk2 6 kxk2 ∀m ∈ N.
n=1

Tending now m to +∞ we get (7.5.4), as required.

∞
|cn (x)|2 converges, thus we have the
P
By the Bessel inequality (7.5.4) for any x ∈ X the series
n=1
following corollary.

Corollary 7.5.12 (Riemann–Lebesgue’s Lemma). Let X be an infinitely dimensional Euclidean or unitary

space, and let (en )∞
n=1 be an orthonormal system of elements of the space X. Then for any x ∈ X its Fourier
coefficients cn (x) satisfy the following
cn (x) −→ 0.
n→∞

Theorem 7.5.13. Let (ek )nk=1

is an orthonormal system of elements of a unitary space X. Suppose that
x ∈ X, and ck (x), k = 1, . . . , n, are its Fourier coefficients with respect to the system (ek )nk=1 . Then the
functional
Xn
∆(λ1 , λ2 , . . . , λn ) = x − λ k ek , λ1 , λ2 , . . . , λn ∈ C
k=1

achieves its minimal value for

λk = ck (x), k = 1, . . . , n.

This property is called the minimization property of Fourier coefficients.

144 CHAPTER 7. LINEAR SPACES

Proof. Let
n
X
y= ck (x)ek .
k=1

Since (x − y) By Corollary 7.5.10 and by Theorem 7.5.6, we have

n 2 n 2 n 2
X X X
2
x− λk ek = x−y+ (ck (x) − λk )ek = kx − yk + (ck (x) − λk )ek =
k=1 k=1 k=1

n
X
= kx − yk2 + |λk − ck |2 .
k=1

It is clear now that

n 2 n
X X
min x− λk ek = kx − yk2 = kxk2 − kyk2 = kxk2 − |ck (x)|2 .
λ1 ,...,λn
k=1 k=1

The geometric meaning of this theorem is that the shortest distance from a point to its orthogonal
projection to a subspace is the perpendicular from the point to the projection.

7.5.1 Gram-Schmidt orthogonalization method

Let X be an infinitely dimensional unitary (Euclidean) space, and let (xn )∞ n=1 ⊂ X be a countable system
of linearly independent elements of X. We show that in X there exists a countable orthonormal system
(en )∞n=1 of elements of X such that for any n ∈ N the element en is a linear combination of elements x1 ,
x2 , . . . , xn . Our proof of this fact is constructive.
x1
Since the system (xn )∞ n=1 is linearly independent, kx1 k 6= 0. Set now e1 := . Then we set
kx1 k
(1) (1)
y2 := α2 e1 − x2 , and find α2 from the condition (e1 , y2 ) = 0. Then we have
(1)
α2 = (x2 , e1 ).

Since the system (xn )∞

n=1 is linearly independent, y2 6= 0, so we can set

y2
e2 := ,
ky2 k
(1) (2) (1) (2)
and (e1 , e2 ) = 0. Furthermore, let y3 = α3 e1 + α3 e2 − x3 . We find α3 and α3 from the conditions
(y3 , e1 ) = (y3 , e2 ) = 0. Obviously,
(1) (2)
α3 = (x3 , e1 ), α3 = (x3 , e2 ),

and y3 6= 0, so we set
y3
e3 := .
ky3 k
Let we already have an orthonormal system (en )m n=1 such that each en is a linear combination of
(1) (2) (n)
elements x1 , x2 , . . . , xn . We set yn+1 = αn+1 e1 + αn+1 e2 + · · · + αn+1 en − xn , and find the coefficients
(k)
αn+1 , k = 1, . . . , n, from the conditions

(yn+1 , ek ) = 0, k = 1, . . . , n.
7.5. BASIC THEORY OF INNER PRODUCT SPACES 145

Then
(k)
αn+1 = (xn+1 , ek ), k = 1, . . . , n,
and yn+1 6= 0, so
yn+1
en+1 := .
kyn+1 k
Continuing in the same manner, we will get a countable orthonormal system (en )∞ n=1 of elements of X
such that for any n ∈ N the element en is a linear combination of elements x1 , x2 , . . . , xn . This process
is called the Gram–Schmidt orthogonalization process.

7.5.2 Orthonormal systems in infinitely dimensional spaces

Let again X be an infinitely dimensional unitary (Euclidean) space, and let (en )∞
n=1 be an orthonormal
system of elements of X (that always exists as we established above).
Definition 7.5.14. For any x ∈ X the series
∞
X
cn (x)en ,
n=1

where cn (x) = (x, en ), n ∈ N, is called the Fourier series of the element x.

From the identity (7.5.5) it follows that the Fourier series of any element x ∈ X converges to x (in the
norm of the space X induced by the inner product) if and only if the following identity holds:
∞
X
kxk2 = |cn (x)|2 . (7.5.6)
n=1

This identity is called Parseval’s identity.

Definition 7.5.15. An orthonormal system (en )∞ n=1 of elements of X is called closed (in Steklov’s sense)
if for any x ∈ X Parseval’s identity (7.5.6) holds.
Thus, we obtain the following fact.
Theorem 7.5.16. Any element of an infinitely dimensional unitary (Euclidean) space X can be expanded
into its Fourier series w.r.t. an orthonormal system (en )∞ ∞
n=1 ⊂ X if, and only if, the system (en )n=1 is
closed (in Steklov’s sense).
Let now
xα , α ∈ A,
be a system of linearly independent elements of a linear normed space X over a field F. Here A is the set
of indices (finite, countable or of higher cardinality).
P
Definition 7.5.17. The collection of all elements of X of the form λα xα , λk ∈ F, is called the linear
α∈A
span of this system, and denoted span (xα ).
α∈A

4
Definition 7.5.18. The system (xα )α∈A of elements of a linear normed space X is called complete in X
if its linear span is dense in X, that is, if for any x ∈ X the following conditions hold:
for any ε > 0 there exist elements xα1 , xα2 , . . . , xαn of the system (xα )α∈A and numbers λ1 , λ2 ,
. . . , λn ∈ F such that
Xn
x− λk xαk < ε.
k=1
4 It is supposed to be linearly independent.
146 CHAPTER 7. LINEAR SPACES

Example 7.5.19. By Theorem 7.4.10 the following systems of elements of Lp [a, b], p > 1, are complete in
Lp [a, b]: bounded measurable functions, step functions, simple functions, continuous functions of compact
support.

Theorem 7.5.20. Let X be in infinitely dimensional unitary (Euclidean) space, and let (en )∞ n=1 be an
orthonormal system of elements of X. If the system (en )∞ n=1 is complete, then every element x ∈ X can
be expanded into its Fourier series w.r.t. the system (en )∞
n=1 .

Proof. Since the system (en )∞

n=1 is complete, for any x ∈ X and for any ε > 0 there exist numbers α1 , α2 ,
. . . , αNε ∈ F such that
Nε 2
X
x− αn en < ε.
n=1

From Theorem (7.5.13) and from (7.5.5), we have

Nε Nε 2 Nε 2
X X X
2 2
kxk − |(x, en )| = x − (x, en )en 6 x− αn en < ε.
n=1 n=1 n=1

This means that

∞
X
kxk2 = |(x, en )|2
n=1

So the system (en )∞

n=1 is closed (in Steklov’s sense), therefore, every element x ∈ X can be expanded
into its Fourier series w.r.t. to the system (en )∞
n=1 .

7.6 Hilbert spaces

This section is devoted to complete inner product spaces.

Definition 7.6.1. A Euclidean (unitary) space is called a real (complex) Hilbert space if it is complete
with respect to the norm induced by the inner product (that is, defined by the formula (7.5.1)).

Example 7.6.2. The (complex) space l2 is a (complex) Hilbert space with inner product
∞
X
(x, y) = xn y n , ∀x, y ∈ l2 ,
n=1

since l2 is complete with respect to the norm

∞
!1
X 2 p
kxk = |x|2 = (x, x),
n=1

as we proved in Example 7.2.10.

Example 7.6.3. Another example is the space L2 [a, b] with inner product

Zb
(f, g) = f (x)g(x)dx, ∀f, g ∈ L2 [a, b],
a

which is complete w.r.t. the inner product induced by this norm. We will study this space in detail in
Section 7.6.5.
7.6. HILBERT SPACES 147

Theorem 7.6.4. If a = (an )∞ ∞

n=1 ∈ l2 , then for any orthonormal system (en )n=1 of elements of a Hilbert
∞
X
space H, the series an en converges. Moreover, there exists an element x ∈ H such that this series is
n=1
the Fourier series of x.
Proof. From the Pythagorean theorem 7.5.6 it follows that for any natural p < m,
p 2 p
X X
an en = |an |2 .
n=m n=m
∞
X
Since the series |an |2 converges (because a ∈ l2 ), we obtain that the sequence of partial sums Sm =
n=1
m
X ∞
X
an en of the series an en is a Cauchy sequence:
n=1 n=1
p
X
kSp − Sm k = an en → 0 as p, m → ∞.
n=m

Since the space H is complete, the sequence Sn has a limit x ∈ H which is the sum of the considered series
∞
X
x= an en
n=1

Furthermore, by Proposition 7.5.3, the inner product is a continuous functional of its variables, therefore,
m
! m
X X
(x, en ) = lim ak ek , en = lim (ak ek , en ) = an .
m→∞ m→∞
k=1 k=1
∞
X
Thus, the series an en is the Fourier series of the element x, as required.
n=1

Theorem 7.6.5. Let H be a Hilbert space (real or complex). An orthonormal system (en )∞ n=1 ⊂ H is
complete if, and only if, the only element of H orthogonal to any en , n ∈ N, is the zero element.
Proof. If the system (en )∞n=1 is complete, then by Theorem 7.5.20 Parseval’s identity (7.5.6) holds. So if
(x, en ) = cn (x) = 0, n ∈ N, then kxk = 0, so x = 0 by the first property of the inner product.
Conversely, suppose that the system (en )∞ n=1 ⊂ H is such that the only element of H orthogonal to all
elements en is zero. Consider an element x ∈ H and show that if a series
∞
X
an en
n=1

is the Fourier series of x, then it converges to x in the norm of the space H induced by its inner product.
X∞
Indeed, from Bessel’s inequality (7.5.4) it follows that the series |an |2 converges. So by Theo-
n=1
rem 7.6.4, there exists an element y ∈ H such that
∞
X
y= an en ,
n=1
∞
X
and an en is the Fourier series for y. Thus, we have
n=1

(x − y, en ) = 0, ∀n ∈ N,
so x − y = 0, as required.
148 CHAPTER 7. LINEAR SPACES

Remark 7.6.6. Some authors call an orthonormal system in a Hilbert space H complete if only if the
zero element of H is orthogonal to all the elements of the system. Theorem 7.6.5 says that this definition
of complete systems is equivalent to our Definition 7.6.5.

Proposition 7.6.7. An orthonormal system (en )∞

n=1 of elements of a Hilbert space X is closed (in Steklov’s
sense) if, and only if, it is complete.

Proof. Let (en )∞ n=1 be closed. From Parseval’s identity (7.5.6) it immediately follows that if for some
x ∈ X we have cn (x) = 0, n ∈ N, then kxk = 0, so x = 0. Therefore, the system (en )∞ n=1 is complete by
Theorem 7.6.5.
If (en )∞
n=1 is complete, then by Theorem 7.5.20 any element x ∈ X can be expanded into its Fourier
series. Now Theorem 7.5.16 implies that the system (en )∞n=1 is closed.

The following theorem will allow us to study closed systems of functions in the space of integrable
functions.

Theorem 7.6.8. Let a set L ⊂ H is dense in a Hilbert space H, and let an orthonormal system (en )∞
n=1 ⊂
L is closed in L. Then it is closed in H.

Proof. Let x ∈ H. Consider the partial sums

n
X
Sn (x) = ck (x)ek .
k=1

It is clear that Sn (A1 x1 + A2 x2 ) = A1 Sn (x1 ) + A2 Sn (x2 ) for any x1 , x2 ∈ H and for any A1 , A2 ∈ F.
Moreover, by Bessel’s inequality (7.5.4) we have kSn (x)k 6 kxk.
ε
By assumption, for any ε > 0 there exists an element y ∈ L such that kx − yk < . Thus, we obtain
3

kx − Sn (x)k 6 kx − yk + kSn (y) − Sn (x)k + ky − Sn (y)k. (7.6.1)

Note that by Bessel’s inequality (7.5.4) and by assumption

ε
kSn (y) − Sn (x)k = kSn (y − x)k 6 ky − xk < . (7.6.2)
3

Moreover, by assumption the system (en )∞

n=1 is closed in L, so for the element y Parseval’s identity (7.5.6)
holds:
X∞
kyk2 = |ck (y)|2 ,
k=1

and by (7.5.5) we have

n
2
X ε
ky − Sn (y)k = kyk2 − |ck (y)|2 < (7.6.3)
3
k=1

for sufficiently large n. Now from (7.6.1)–(7.6.3) it follows that

kx − Sn (x)k < ε

for sufficiently large n. This means that the system (en )∞

n=1 is complete, so by Theorems 7.5.16 and 7.5.20
it is closed, as required.
7.6. HILBERT SPACES 149

7.6.1 Isomorphism of separable Hilbert space

Definition 7.6.9. Two Euclidean (unitary) spaces X1 and X2 are called isomorphic if there exists a
bijection (one-to-one correspondence) F : X1 7→ X2 such that

(F (x), F (y))2 = (x, y), x, y ∈ X1 ,

where (·, ·)1 and (·, ·)2 are the inner products of the spaces X1 and X2 , respectively. F is called an
isomorphism.

Definition 7.6.10. A subset L0 of a linear space L is called a subspace of L if L0 is a linear space. A

subset is called closed if for any sequence (xn )∞ 0
n=1 ⊂ L convergent to an element x ∈ L, we have that
x ∈ L0 .

Definition 7.6.11. A real (complex) Hilbert space H is a completion of a Euclidean (unitary) space X if
there is a subspace X 0 in H isomorphic to X and dense in H.

Theorem 7.6.12. Any Euclidean (unitary) space has a completion.

This is a particular case of the corresponding theorem for metric spaces and we leave a proof of this
theorem for the class of ”Functional Analysis”.

Definition 7.6.13. A linear normed space X is called separable if there exists a countable complete system
of elements of the space X.

Definition 7.6.14. A countable system (en )∞ n=1 of elements of a linear normed space X is called a basis
of the space X if every element x of the space X has a unique expansion w.r.t. this system, that is, there
exists a unique series sequence (λn )∞
n=1 such that

∞
X
x= λn en
n=1

Here the series converges to x w.r.t. the norm of the space X, that is,
n
X
∀ε > 0 ∃N : ∀n > N, x− λk ek < ε.
k=1

Example 7.6.15. By the 1st Weierstrass theorem, the system (xn )∞

n=1 complete in C[a, b], but it is not
a basis in C[a, b]. Later we will show that the system

1, sin x, cos x, sin 2x, cos 2x, . . . , sin nx, cos nx, . . .

is a basis in L2 [−π, π]. We will also show that this system is not a basis in C[−π, π].

Theorem 7.6.16. In any separable Euclidean (unitary) space there exists an orthonormal basis.

Proof. Let X be a separable Euclidean (unitary) space. Then there exists a countable complete linearly
independent system (xn )∞n=1 of elements of the space X. Using the Gram–Schmidt orthogonalization
process, from (xn )∞
n=1 we construct an orthonormal system (en )∞
n=1 which is complete in X. By Theo-
rem 7.5.20 any element of X can be expanded into its Fourier series w.r.t. this system, so (en )∞
n=1 is an
orthonormal basis in X.

Theorem 7.6.17. Any separable real (or complex) Hilbert space H is isomorphic to the space l2 over R
(or C).
150 CHAPTER 7. LINEAR SPACES

Proof. By Theorem 7.6.16, in H there exists an orthonormal basis (en )∞ n=1 . Then to each element x ∈ H
we can correspond the sequence cn (x) = (x, en ), n ∈ N, of its Fourier coefficients. By Theorem 7.5.20,
the system (en )∞ n=1 is closed in Steklov’s sense, so the Fourier series of the element x satisfies Parseval’s
identity 7.5.6, so (cn (x))∞
n=1 ∈ l2 .
Conversely, let a = (an )∞ n=1 ∈ l2 . Then by Theorem 7.6.4, the numbers an , n ∈ N, are the Fourier
coefficients of an element x of the space H.
Thus, between the spaces H and l2 there exists a one-to-one correspondence. Furthermore, it is obvious
that if x ∼ (an ) and y ∼ (bn ), then (αx+βy) ∼ α(an )+β(bn ), where α, β ∈ F. Moreover, by Theorem 7.5.20
we have
∞
X X∞
kxk2 = |an |2 , kyk2 = |bn |2 ,
n=1 n=1

and from the continuity of the inner product (see Proposition 7.5.3) it follows that
∞
X
(x, y)H = an bn = (a, b)l2 . (7.6.4)
n=1

It is clear that the space l2 is separable, thus, from Theorem 7.6.17 we obtain the following fact.
Theorem 7.6.18. All separable infinitely dimensional Hilbert spaces are isomorphic to each other.
Proof. Let H and H 0 be separable infinitely dimensional Hilbert spaces. According to Theorem 7.6.16
∞
X
there exist orthonormal bases (en )∞
n=1 ⊂ H and (e0 ∞
)
n n=1 ⊂ H 0
. Then to each element x = an en ∈ H
n=1
∞
X
we correspond the element x0 = an e0n ∈ H 0 . Clearly, this is the required isomorphism.
n=1

From the facts established above it follows that if (en )∞

n=1 is a basis in a separable Euclidean (unitary)
space X, then the completion of the space X is a Hilbert space consisting of all the series of the form
∞
X
an en
n=1

where (an ) ∈ l2 , with the inner product of the form

∞
X
(x, y) = an bn ,
n=1

whenever
∞
X ∞
X
x= an en , y= bn e n .
n=1 n=1

Note that not every Hilbert space is separable.

Example 7.6.19. Let HR be the set of all functions defined on R and having at most countable number
of non-zero (real) values satisfying
X
x2 (t) < ∞, ∀x(t) ∈ HR .
t∈R

This is a Euclidean space with the inner product

X
(x(t), y(t)) = x(t)y(t). (7.6.5)
t∈R
7.6. HILBERT SPACES 151

For every function f (t) defined on R the set {t ∈ R : f (t) 6= 0} is called the support of f and is denoted
as suppf . Thus, the space HR is the space of functions with countable support.
This space is complete. Indeed, let {xn (t)}∞ 1 is a Cauchy sequence in the norm induced by the inner
product (7.6.5): for any ε > 0 there exists a number Nε > 0 such that ∀n, m > Nε
ω
X
|xn (tk ) − xm (tk )|2 < ε, 1 6 ω 6 ∞,
k=1
∞
S
where tk ∈ T , and T = supp[xn (t)]. It is clear that T is at most countable as at most countable union
n=1
of at most countable sets. Thus, we have that for any tk ∈ T ,

|xn (tk ) − xm (tk )|2 < ε,

so the sequence xn (tk ) is a Cauchy sequence for every tk ∈ T , thus xn (tk ) converges to x(tk ). The function
x(t) has at most countable support. Moreover, for any finite number M 6 ω, we obtain
M
X
|xn (tk ) − xm (tk )|2 < ε,
k=1

and running m to infinity

M
X
|xn (tk ) − x(tk )|2 6 ε,
k=1
so
ω
X
kxn − xk2 = |xn (tk ) − x(tk )|2 6 ε.
k=1

Now for sufficiently large n we get

kxk2 6 kxn − xk2 + kxn k2 < ∞.

Therefore, x ∈ HR , so HR is a real Hilbert space.

However, HR is not separable. Indeed, if
(
1, t = τ,
xτ (t) =
0, t 6= τ,
√
then kxτ − xτ 0 k = 2 whenever τ 6= τ 0 . There exists a bijection between the system {xτ (t)}τ ∈R and the
real line R.
∞
If HR were separable, there would exist a countable dense system (en )∞
S
n=1 . However, supp[en (t)] is
n=1
countable, so for any linear combination of any elements of this system there exists a point τ ∈ R that do
∞
S
not belong the supp[en (t)], so there are always exist functions xτ that cannot be approximated by any
n=1
countable system of elements of HR , so this space is not separable.

7.6.2 Orthogonal projections

Let L be a subspace of a Euclidean (unitary) subspace X.

Definition 7.6.20. An element y0 ∈ L is called the orthogonal projection of an element x0 ∈ X onto the
subspace L if
(x0 − y0 , y) = 0 ∀y ∈ L. (7.6.6)
152 CHAPTER 7. LINEAR SPACES

Clearly, every element x ∈ X can have only one orthogonal projection onto L. Indeed, if there exist
y1 , y2 ∈ L such that
(x − y1 , y) = 0, (x − y2 , y) = 0 ∀y ∈ L,
then
ky1 − y2 k2 = (y1 − y2 , y1 − x + x − y2 ) = (y1 − y2 , y1 − x) + (y1 − y2 , x − y2 ) = 0,
since y1 − y2 ∈ L. Thus, y1 = y2 .

Theorem 7.6.21. An element y0 ∈ L is the orthogonal projection of an element x0 ∈ X if, and only if,

kx0 − y0 k = inf kx0 − yk. (7.6.7)

y∈L

Proof. If y0 ∈ L satisfies the condition (7.6.6), then for any y ∈ L

kx0 − yk2 = ((x0 − y0 ) + (y0 − y), (x0 − y0 ) + (y0 − y)) = kx0 − y0 k2 + ky0 − yk2 ,

so the condition (7.6.7) holds.

Conversely, let the condition (7.6.7) holds for an element y0 ∈ L. Let us consider the function

f (t) = kx0 − y0 + tyk2 , t ∈ R,

where y is an arbitrary element of L.

If X is Euclidean, then we have

f (t) = kx0 − y0 k2 + 2t(x0 − y0 , y) + t2 kyk2 .

Since this function has the minimal value at the point t = 0 by assumption, we obtain f 0 (0) = 0 that
implies the condition (7.6.6).
If X is a unitary space, then

f (t) = kx0 − y0 k2 + 2t Re(x0 − y0 , y) + t2 kyk2 ,

so analogously,
Re(x0 − y0 , y) = 0 ∀y ∈ L.
And considering the function
g(t) = kx0 − y0 + ityk2 , t ∈ R,
in the same way we obtain
Im(x0 − y0 , y) = 0 ∀y ∈ L,
therefore, the condition (7.6.6) holds in the case of unitary spaces too.

Theorem 7.6.22. If a subspace L of a Euclidean (unitary) space X is complete, then for any x ∈ X,
there exists an orthogonal projection onto the subspace L.

Proof. According to Theorem 7.6.21 it suffices to prove that for any x0 ∈ X there exists y0 ∈ L such that

kx0 − y0 k = inf kx0 − yk.

y∈L

Given x0 ∈ X, define the number

d := inf kx0 − yk2 .
y∈L

Then there exists a sequence (yn )∞

n=1 ⊂ L such that

lim kx0 − yn k2 = d. (7.6.8)

n→∞
7.6. HILBERT SPACES 153

It is easy to see that this system is Cauchy. Indeed, from the Apollonius identity
2
ym + yn
kym − yn k2 = 2kyn − x0 k2 + 2kym − x0 k2 − 4 x0 −
2

which is true for any n, m ∈ N, it follows that

2
ym + yn
x0 − > d,
2
ym + yn
since ∈ L. So we have
2
kym − yn k2 6 2kyn − x0 k2 + 2kym − x0 k2 − 4d, (7.6.9)

and (7.6.8) implies that for any ε > 0 there exists a number N ∈ N such that for any n > N ,

ε2
kx0 − yn k < d + .
4
Thus from (7.6.9) we get
kyn − ym k < ε ∀n, m > N.
Since the subspace L is a complete linear normed space, there exists an element y0 ∈ L such that kyn −y0 k →
0 as n → ∞. From the continuity of norm we obtain

lim kx0 − yn k = kx0 − y0 k,

n→∞

as required.

7.6.3 Linear functionals and the Riesz representation theorem

Let again X be a Euclidean or unitary space.

Definition 7.6.23. Any number-valued mapping f is called a functional. That is, if f : X 7→ C or R,

then f is a functional.

Definition 7.6.24. A functional f : X 7→ F defined on X is called linear if

f (c1 x1 + c2 x2 ) = c1 f (x1 ) + c2 f (x2 ), ∀c1 , c2 ∈ F, x1 , x2 ∈ X,

where F = C or R.

Definition 7.6.25. The set K = {x ∈ X : f (x) = 0} is called the kernel of the functional f .

Definition 7.6.26. A functional f : X 7→ C or R is called bounded if there exists a constant M > 0 such
that
|f (x)| 6 M kxk ∀x ∈ X.

Lemma 7.6.27. A linear functional defined on a linear normed space X is continuous on X if it is

continuous at 0.

Proof. Let f (xn ) → 0 as n → ∞ whenever xn → 0 in X as n → ∞, that is kxn k → 0.

Suppose that kxn − xk −→ 0, that is, xn → x in X. Then f (xn ) − f (x) = f (xn − x) → 0 as n → ∞,
n→∞
so f (xn ) → f (x) as n → ∞ for any x ∈ X.

Theorem 7.6.28. In a normed space a linear functional is continuous if and only if it is bounded.
154 CHAPTER 7. LINEAR SPACES

Proof. Let f is bounded, and let kxn k → 0 as n → ∞. Then

|f (xn )| 6 C · kxn k → 0 as n → ∞.

So f is continuous at zero, therefore, is continuous by Lemma 7.6.27.

Conversely, suppose that f is continuous but unbounded. Then there exists a sequence (xn )∞
n=1 ⊂ X
such that |f (xn )| > nkxn k. Introduce a new sequence
xn
ξn = .
nkxn k
1
Since f is linear, we have |f (ξn )| > 1. At the same time, kξn k =
→ 0 as n → ∞. Thus, we have ξn → 0
n
as n → ∞ in the norm of X, so by continuity f (ξn ) → 0, a contradiction.
Definition 7.6.29. Let f be a linear continuous functional defined on a normed space X. The number
kf k := inf M where inferior is taken over all numbers M > 0 satisfying the inequality

|f (x)| 6 M kxk ∀x ∈ X,

is called the norm of the functional f .

Lemma 7.6.30. The norm of a linear continuous functional defined on a normed space X can be defined
by the formula
|f (x)|
kf k = sup = sup |f (x)| = sup |f (x)|. (7.6.10)
x∈X, kxk x∈X, x∈X,
kxk6=0 kxk61 kxk=1

Proof. It is easy to see that the norm kf k is the inferior of the numbers M > 0 satisfying the inequality
|f (x)|
6 M, ∀x ∈ X, x 6= 0.
kxk
b ∈ X, kb
This means that for any ε > 0 there exists an element x xk =
6 0, such that
|f (b
x)|
> kf k − ε.
kbxk
By definition of supremum, we have
|f (x)|
kf k = sup .
x∈X, kxk
kxk6=0
 
|f (x)| x
Furthermore, since =f , so
kxk kxk
kf k = sup |f (x)| 6 sup |f (x)|.
x∈X, x∈X,
kxk=1 kxk61

Now if kxk 6 1, then

|f (x)| 6 kf k · kxk 6 kf k,
so
sup |f (x)| 6 kf k.
kxk61

Consequently,
This implies that
sup |f (x)| = kf k.
kxk61
7.6. HILBERT SPACES 155

It is clear that the norm of lineal continuous functionals possesses all three properties of norms. So
we can introduce the linear normed space of linear bounded functional defined on X. This space is called
dual to the space X and is usual denoted X ∗ .
Example 7.6.31. Let tk ∈ [a, b], and ck ∈ R, k = 1, . . . , n, a < t1 and tn < b. In C[a, b] consider the
following functional
n
def X
f (x) = ck x(tk ), x ∈ C[a, b].
k=1

This functional is obviously linear.Moreover,

n n
!
X X
|f (x)| 6 |ck | · |x(tk )| 6 |ck | · kxk, ∀x ∈ C[a, b],
k=1 k=1

so
n
|f (x)| X
kf k = sup 6 |ck |.
x∈C[a,b], kxk k=1
kxk6=0

Thus, f is a bounded linear functional.

Consider now the function y(t) ∈ C[a, b] which satisfies y(tk ) = sgn ck , k = 1, . . . , n, and is linear
between the points a, t1 , . . . , tn , b, and y(a) = y(b) = 0. It is clear that kyk = 1. Then we have
n
X n
X
f (y) = ck · sgn ck = |ck |.
k=1 k=1

So we found a function in C[a, b] on which the functional f achieves the value equal to a bound of its
norm, so the norm is equal to this bound:
n
X
kf k = |ck |.
k=1

In Hilbert spaces, it is possible to find a general form of any linear bounded functional.
Lemma 7.6.32. For any element a ∈ X, where X is a Euclidean or unitary space, the functional

f (x) = (x, a) x∈X (7.6.11)

is linear and bounded. Moreover,

kf k = kak.
Proof. The linearity of f is obvious, and the boundness follows from the Cauchy-Bunyakovsky-Schwarz
inequality:
|f (x)| 6 kak · kxk ∀x ∈ X.
Since f (a) = kak2 , we have kf k = kak.
Lemma 7.6.33. If a functional defined on X is given by the formula (7.6.11), then the element a is
uniquely defined.
Proof. Indeed, let
f (x) = (x, a) and f (x) = (x, b),
then (x, b − a) = 0 for any x ∈ X. In particular, (b − a, b − a) = kb − ak = 0, so a = b.
Theorem 7.6.34 (Riesz). Any linear continuous functional defined on a Hilbert space H can be uniquely
represented as in (7.6.11) for a certain element a ∈ H.
156 CHAPTER 7. LINEAR SPACES

Proof. Let K be the kernel of the functional f . From the continuity of the functional f it follows that K
is a closed subspace of the space H: if f (yn ) = 0, yn ∈ Y , n ∈ N, and yn −→ y, then f (y) = 0. Moreover,
n→∞
since the space H is complete, the subspace K us complete as well (any Cauchy sequence in K converges,
and the limit must belong to K).
If K = H, then f (x) ≡ 0 on H, so clearly

f (x) = (0, x) x ∈ H.

Suppose now that K 6= H. Then there exists an element x0 ∈ H such that f (x0 ) 6= 0. Let y0 ∈ K be
the orthogonal projection of the element x0 onto the subspace K which exists according to Theorem 7.6.22.
We set
z0 = x0 − y0 .
Then we have f (z0 ) = f (x0 ) 6= 0, and (z0 , y) = 0 for any y ∈ K.
Since  
f (x)
f x− z0 = 0 x ∈ H,
f (z0 )
it follows that
f (x)
x− z0 ∈ K ∀x ∈ H.
f (z0 )
Therefore,  
f (x)
x− z0 , z0 =0 ∀x ∈ H.
f (z0 )
Thus, we obtain
kz0 k2
(x, z0 ) = · f (x).
f (z0 )
f (z0 )
Note that f (z0 ) 6= 0 implies kz0 k =
6 0. Consequently, the element a = z0 exists, and
kz0 k2

f (x) = (x, a), a ∈ H,

as required.

This theorem and Lemma 7.6.32 show that the space H ∗ dual to a Hilbert space H is isomorphic to
H. This is principal difference between Hilbert and Banach spaces that isomorphic to a subspace of their
dual space.

Example 7.6.35. In l2 let us consider the following functional

∞
X xn + xn+1
f (x) = ,
n=1
2n

1
where x = (x1 , x2 , . . . , xn , . . .) ∈ l2 . This functional is bounded. Indeed, if y ∈ l2 is such that yn = n ,
2
then by Cauchy-Bunyakovski-Schwarz inequality we obtain
∞ ∞
X 1 X 1
|f (x)| 6 n
· |x n | + 2 n+1
· |xn+1 | 6 3kykl2 · kxkl2 .
n=1
2 n=1
2

To find the norm we will use Theorem 7.6.34. The functional f can be rewritten in the form
∞
x1 X xn
f (x) = +3 = (a, x),
2 n=2
2n
7.6. HILBERT SPACES 157

where a = (a1 , a2 , . . . , an , . . .) with

1 3
a1 = , an = , n = 2, 3, . . . .
2 2n
Thus, by Lemma 7.6.32, we have
∞
1 9 X 1 1 9
kf k = kakl2 = − 9 − + 9 n
= − 9 − + 12 = 1.
4 4 n=0
4 4 4

7.6.4 Weak convergence

Definition 7.6.36. A sequence (xn )∞
n=1 of elements of a linear normed space X is called weakly convergent
to an element x ∈ X if
lim f (xn ) = f (x) ∀f ∈ X ∗ .
n→∞

Due to the Riesz Theorem 7.6.34, we can define the weak convergence in Hilbert spaces in the inner
product form.

Definition 7.6.37. A sequence (xn )∞

n=1 of elements of a Hilbert space H is called weakly convergent to
an element x ∈ H if
lim (xn , y) = (x, y) ∀y ∈ H.
n→∞

This definition will allow us later to understand the notion of weak solutions of differential equations.

Theorem 7.6.38. If the sequences (xn )∞ n=1 of elements of a linear normed space X converges to x ∈ X
in the norm of X, then it is weakly convergent to x.

Proof. By linearity of the inner product and by the boundness of functionals in the space X ∗ we have for
any y ∈ X
|f (xn ) − f (x)| = |f (xn − x)| 6 kf k · kxn − xk → 0 as n → ∞.

Since any Hilbert space is a normed space, we have the corresponding theorem for Hilbert spaces which
we state separately for the sake of convenience in the future.

Theorem 7.6.39. If the sequences (xn )∞ n=1 of elements of a Hilbert space H converges to x ∈ H in the
norm of H, then it is weakly convergent to x.

Generally speaking, converse is not true.

Example 7.6.40. Consider the sequence of coordinate vectors (en )∞

n=1 in l2 , where en = (0, . . . , 0, 1, 0, . . .).
| {z }
n−1
This sequence weakly convergent in l2 . Indeed, for any a ∈ l2 we have

(en , a) = an → 0 as n → ∞,
∞
|an |2 converges. However, the sequence (en )∞
P
since the series n=1 is not convergent in l2 , because for
n=1
any a ∈ l2
ken − ak2 = ken k2 − 2 Re(en , a) + kak2 −→ 1 + kak2 > 1.
n→∞

However, in finite-dimensional case the weak convergence is equivalent to the convergent in norm.
158 CHAPTER 7. LINEAR SPACES

Example 7.6.41. Consider the space Cn with standard inner product. Let (ek )nk=1 be an arbitrary
orthonormal basis in Cn , and suppose that a sequence (xm )∞ n
m=1 weakly converges to an element x ∈ C .
Then we have
Xn Xn
xm = x(k)
m ek and x= x(k) ek ,
k=1 k=1

and
x(k)
m = (xm , ek ) −→ (x, ek ) = x
(k)
, k = 1, . . . , n.
m→∞

This means that the sequence (xm )∞

m=1 converges to x entrywise. This implies

n
X
2
kxm − xk = |x(k)
m −x
(k) 2
| →0 as m → ∞.
k=1

This example gives us an idea that in l2 the coefficient-wise convergence does not imply the convergence
(n) 1
in norm. Indeed, if x(n) ∈ l2 is such that xk = 1 1 , then
k2+n
∞
!2 ∞
(n) 2
X 1 X 1
kx kl2 = 1 1 = 2 < +∞ ∀n ∈ N.
k=1 k2+n k=1 k
1+ n

1
The coefficient-wise limit of the sequence x(n) is the sequence x whose coefficients are xk = 1 , so
k2
∞
!2 ∞
X 1 X 1
kxk2l2 = 1 = = +∞,
k
k=1 k2 k=1

thus, x 6∈ l2 .

With an additional property the weak convergence might imply the convergence in norm. In fact, by
Theorem 7.2.7 it follows that if kxn − xk → 0 as n → ∞, then kxn k → kxk. And converse is not true.

Theorem 7.6.42. Let a sequence (xn )∞n=1 of elements of Hilbert space H weakly converges to x ∈ H, and
let kxn k → kxk as n → ∞. Then (xn )∞
n=1 converges to x in norm of H.

Proof. Since (xn )∞ 2 2

n=1 weakly converges to x, we have (xn , x) → kxk and (x, xn ) → kxk . Thus, we have

kxn − xk2 = kxn k2 − (xn , x) − (x, xn ) + kxk2 =

     
= kxn k2 − kxk2 + kxk2 − (xn , x) + kxk2 − (x, xn ) → 0 as n → ∞.

7.6.5 The space L2

Consider now the (complex) space L2 [−π, π]. According to Theorem 7.3.9, the space is complete w.r.t.
the norm5
 π 1
Z 2
2
kf k2 =  |f (x)| dx , f ∈ L2 [−π, π].
−π

Zπ Z
5 In spite of simplicity, in what follows we use the standard notation g(x)dx for the Lebesgue integral g(x)dµ.
−π [−π,π]
7.6. HILBERT SPACES 159

Moreover, we can introduce the following function of two variables

 1
Zπ 2

(f, g) =  f (x)g(x)dx , f, g ∈ L2 [−π, π]. (7.6.12)

−π

It is easy to see that this function defines an inner product on the space L2 [−π, π] since it satisfies all
three properties of inner products. Indeed, for f ∈ L2 [−π, π], (f, f ) = kf k22 > 0 (= 0 ⇐⇒ f = 0). This
is the property the norm k · k2 which we established earlier. The linearity of the function (7.6.12) follows
from the linearity of Lebesgue integral (see Properties 6 and 7 of Lebesgue integral). Finally, the property
(f, g) = (g, f ) is obvious.
Thus, the space L2 [−π, π] is a Hilbert space with inner product defined by (7.6.12).
Theorem 7.6.43. The space L2 [−π, π] is separable.
Proof. Let f ∈ L2 [−π, π]. By Theorem 7.4.14, for any ε > 0 there exists a function g ∈ C[−π, π] such that
ε
kf − gk < . In it its turn, the function g can be approximated by a polynomial p by the first Weierstrass
3
ε
theorem: max |g(x) − p(x)| < . This implies that
x∈[−π,π] 6π
 1
Zπ 2 Zπ
2 ε ε
kg − pk2 =  |g(x) − p(x)| dx 6 max |g(x) − p(x)| dx < · 2π = .
x∈[−π,π] 6π 3
−π −π

Furthermore, every polynomial p with complex coefficients can be approximated by a polynomial q with
complex rational coefficients (that is, the real and imaginary part of every coefficient is rational). So if
n n
ak xk , ak ∈ C, and q(x) = bk xk , Re bk , Im bk ∈ Q, then
P P
p(x) =
k=0 k=0

n
X ε
kp − qk2 6 2π · max |p(x) − q(x)| < 2π |ak − bk |π k < ,
x∈[−π,π] 3
k=0

ε
since we can choose a polynomial q such that |ak − bk | < . Combining all the previous we
6(n + 1)π k+1
get
kf − qk2 6 kf − gk2 + kg − pk2 + kp − qk2 < ε.
Thus, the system of all polynomials with complex rational coefficients is complete in the space L2 [−π, π].
By Example 1.3.10 the set of polynomials with rational coefficients is countable. But every polynomial Q
with complex rational coefficients has the form Q(x) = Re Q(x) + i Im Q(x). So the set of all polynomials
with complex rational coefficients is the union of two countable sets, so it is countable by Theorem 1.3.5.
Since all the polynomials are Lebesgue integrable on [−π, π], we get that the space L2 [−π, π] contains a
countable complete system, so it is separable.
Note that this theorem can be proved much simpler. Indeed, according to Corollary 7.4.15, the count-
able system (xn )∞n=1 is complete in Lp [−π, π], p > 1, so even Lp [−π, π] is separable. But this system is not
orthonormal in L2 [−π, π] (for p 6= 2, Lp [a, b] has no inner product). So, the proof of Theorem 7.6.43 gives
an idea that the orthonormal polynomial basis in L2 [−π, π] can be a system of polynomials with rational
coefficients. In fact, such a system exists, and the polynomials are called Legendre polynomials. These
polynomials have the form
1 dn
Ln (x) = n · n (x2 − 1)n .
2 n! dx
Corollary 7.6.44. If Parseval’s identity w.r.t. a system (en )∞ n
n=1 holds for all the functions x , n =
0, 1, 2, . . ., then the system (en )∞
n=1 is closed.
160 CHAPTER 7. LINEAR SPACES

Proof. Indeed, consider the polynomial

m
X
P (x) = ak xk .
k=0

Then
m
X
Sn (P ) = ak Sk (xk ),
k=0
so
m
X
kP − Sn (P )k2 6 |ak | · kxk − Sn (xk )k2 → 0 as n → ∞.
k=0

Thus, Parseval’s identity holds for any polynomials, and the set of all polynomials is dense in L2 [−π, π]
by Theorem 7.4.14.
But do orthonormal bases exist in L2 [−π, π]? Yes, as we mentioned above. Legendre polynomials form
such a basis. The existence of such a basis is provided by Theorems 7.6.43 and 7.6.16.
Let us consider the following countable system of continuous functions

1, sin x, cos x, . . . , sin nx, cos nx, . . . (7.6.13)

it is easy to see that this system is orthonormal in L2 [−π, π].

Lemma 7.6.45. If n ∈ N ∪ {0}, then
Zπ
sin nxdx = 0.
−π

Proof. If n = 0, then the statement is obvious. Let n > 0.. Then we have
Zπ π
− cos nx
sin nxdx = = 0.
n −π
−π

Analogously we have the following

Lemma 7.6.46. If n ∈ N, then
Zπ
cos nxdx = 0.
−π

Lemma 7.6.47. For any n ∈ N,

Zπ Zπ
2
cos nxdx = sin2 nxdx = π.
−π −π

Proof. The formulæ follow from the well-known identities

1 + cos 2x 1 − cos 2x
cos2 x = , sin2 x =
2 2
and from Lemmata 7.6.45–7.6.46.
So we get
Lemma 7.6.48. The system (7.6.13) is orthogonal in L2 [−π, π].
7.6. HILBERT SPACES 161

Proof. This fact follows from the Lemmata 7.6.45–7.6.47 and the standard formulae of sines and cosines
sums of sums of angles:
cos(n − m)x − cos(n + m)x
sin nx sin mx = ,
2
cos(n − m)x + cos(n + m)x
cos nx cos mx = ,
2
sin(n − m)x + sin(n + m)x
sin nx cos mx = ,
2

Corollary 7.6.49. The trigonometric system (7.6.13) is an orthogonal basis in L2 [−π, π].

Proof. Indeed, the trigonometric polynomials

n
a0 X
Tn (x) = + (am cos mx + bm sin mx)
2 m=1

are dense in L2 [−π, π] by Corollary 7.4.15. But Parseval’s identity holds for any trigonometric polynomial:

Zπ n
π|a0 |2 X
|Tn (x)|2 dx = +π (|am |2 + |bm |2 ).
2 m=1
−π

Consequently, the system (7.6.13) is closed in L2 [−π, π] by Theorem 7.6.8. By Proposition 7.6.7 the
system (7.6.13) is complete, so it is a basis in L2 [−π, π] by Theorem 7.5.20.

Thus, any function f from the space L2 [−π, π] can be expanded in to the series
∞
a0 X
f (x) = + (am cos mx + bm sin mx)
2 m=1

convergent in the norm of L2 [−π, π], where the coefficients am and bm have the form

Zπ Zπ
1 1
am = f (x) cos mxdx, bm = f (x) sin mxdx, m = 0, 1, 2, . . . (7.6.14)
π π
−π −π

are the Fourier coefficients of the function f w.r.t. the system (7.6.13).
From the properties of separable Hilbert spaces, we have following fact.

Proposition 7.6.50. A function f belongs to the space L2 [−π, π] if, and only if, the sequences of its
Fourier coefficients defined by (7.6.14) belong to the space l2 . Moreover, the Parseval’s identity holds:

Zπ ∞
π|a0 |2 X
kf k22 = |f (x)|2 dx = +π (|am |2 + |bm |2 ). (7.6.15)
2 m=1
−π

From the proof of Theorem 7.6.17 (see formula (7.6.4)) it follows that for any two functions f, g ∈
L2 [−π, π] with Fourier coefficients

Zπ Zπ
1 1
am = f (x) cos mxdx, bm = f (x) sin mxdx, m = 0, 1, 2, . . .
π π
−π −π
162 CHAPTER 7. LINEAR SPACES

and
Zπ Zπ
1 1
cm = g(x) cos mxdx, dm = g(x) sin mxdx, m = 0, 1, 2, . . .
π π
−π −π

one has
Zπ ∞
a0 c0 X
(f, g) = f (x)g(x)dx = + (am cm + bm dm ). (7.6.16)
4 m=1
−π

This formula implies the following interesting fact.

Proposition 7.6.51. The Fourier series of a function f ∈ L2 [−π, π] w.r.t. the system (7.6.13) can be
integrated term by term over any measurable set A ⊂ [−π, π], i.e.
 
Z ∞
X Z Z
f (x)dµ = a0 µA + an cos nxdµ + bn sin nxdµ . (7.6.17)
A n=1 A A

Proof. In fact, let g(x) = χA (x), the characteristic function of the set A. It is bounded and measurable,
since A is measurable. Substituting g(x) into the formula (7.6.16) we get (7.6.17), as required.

Remark 7.6.52. Note that if (en )∞ n=1 is an arbitrary orthonormal basis in L2 [−π, π], then Proposi-
tions 7.6.50–7.6.51 are also true for the corresponding Fourier series w.r.t. this basis.
7.7. PROBLEMS 163

7.7 Problems
7.7.1 Linear normed spaces. Spaces Lp .
Problem 7.1. Prove that any normed linear space is a metric (linear) space with the metric ρ(f, g) =
kf − gk.
Problem 7.2. Prove that the space C 1 [a, b] of continuously differentiable functions on [a, b] with norm

kf k = max |f (x)| + max |f 0 (x)|

x∈[a,b] x∈[a,b]

is a complete linear normed space. (Prove first that kf k is a norm.)

Problem 7.3. Prove that the space of all bounded (real or complex) sequences is a complete linear normed
space with norm
kxk∞ = sup |xk |.
k∈N

Problem 7.4. Let fn → f in L1 (X). Prove that |fn | → |f | in L1 (X).

a.e.
Problem 7.5. Let fn → f in L1 (X), and fn → g almost everywhere on X. Prove that f = g on X.
µ
Problem 7.6. Let {fn (x)}+∞ n=1 ∈ S(X) and f0 (x) ∈ S(X), where µX < +∞. Prove that if fn −→ f0 on
X, and the functions fn (x), n > 1, are bounded almost everywhere on X as a whole (that is, there is one
1 L
constant for all functions for almost all points x), then fn −→ f0 on X.
Problem 7.7. Prove that the sequence fn = n2 xe−nx , n ∈ N, converges everywhere on [0, 1] to the
function f0 (x) ≡ 0, but fn does not converge in L2 [0, 1]. Does it converge in L1 [0, 1] and C[0, 1]?
Problem 7.8. Does the sequence fn (x) = (sin nx)n converge to f0 ≡ 0 in Lp norm, p > 1, on [0, π]?
Hint: Use Stirling’s formula:
√  n n   
1
n! = 2πn 1+O as n → ∞.
e n
Problem 7.9. Do the following expressions define norms?
1) x 7→ | arctan x| for x ∈ R;
 n
1
p
p
for x = (x1 , . . . , xn ) ∈ Rn if 0 < p < 1 and n > 2;
P
2) x 7→ |xk |
k=1

3) x 7→ max |x(t)| for x ∈ C[a, b];

a+b
a6t6 2

4) x 7→ |x(a)| + max |x0 (t)| for x ∈ C 1 [a, b];

a6t6b

5) x 7→ |x(b) − x(a)| + max |x0 (t)| for x ∈ C 1 [a, b].

a6t6b

Problem 7.10. Prove that in the space C[0, π] (with maximum norm) the functions 1, cos t, cos2 t are
linearly independent but the functions 1, cos 2t, cos2 t are linearly dependent.
Problem 7.11. Consider the space B of all n times continuously differentiable functions on [a, b] with
the norm
kxks = max { sup pj (t)|x(j) (t)|},
06j6n t∈[a,b]

where pj (t) ∈ C[a, b], j = 0, 1, . . . , n, are some positive functions.

Prove that k · ks defines a norm and that B is a complete space.
164 CHAPTER 7. LINEAR SPACES

Problem 7.12 (Extended Hölder inequality). Let f ∈ L1 (X) and g ∈ L∞ (X). Prove that f g ∈ L1 (X)
and Z
|f (x)g(x)|dµ 6 kf k1 · kgk∞ .
X

Problem 7.13. Let 1 6 s < p < ∞. Show that there exists a function f ∈ Lr (0, +∞) for any r ∈ [s, p]
such that f 6∈ Lr (0, +∞) for any r 6∈ [s, p].
Problem 7.14. Let f ∈ Lp (X) for all p > p0 > 1. Prove that there exists a finite or infinite limit of the
magnitudes kf kp as p → ∞. Prove also that if the limit is finite, then f ∈ L∞ (X) and

lim kpk = kf k∞ .
p→∞

If the limit is infinite, prove that f 6∈ L∞ (X).

µ
Problem 7.15. Let (fn )∞
n=1 ⊂ Lp (X) and f0 ∈ Lp (X) for some p > 1. Suppose that fn −
→ f0 , and
Lp
|fn (x)| 6 |g(x)| on X, ∀n ∈ N ∪ {0}, where g ∈ Lp (X). Prove that fn −−→ f .
Hint: Prove first that Lebesgue’s Dominated Convergence Theorem is true if we change the convergence a.e.
to convergence in measure.
Problem 7.16. Let µX < +∞ and 1 6 r < p < ∞. Given a sequence (fn )∞ n=1 ⊂ Lp (X) such that
µ
fn −→ f , suppose that there exists a number C > 0 so that kfn kp 6 C for any n ∈ N. Prove that
n→∞
Lr (X)
fn −→ f .
n→∞

Problem 7.17. Let 1 6 p < ∞. Given a sequence {fn (x)}∞

n=1 ⊂ Lp (X), suppose that

∞
X
kfn kp < +∞.
n=1

Prove that the series

∞
X
fn (x)
n=1

converges absolutely a.e. in Lp (X), and

∞
X ∞
X
fn (x) 6 kfn kp .
n=1 p n=1

Problem 7.18. Does the sequence {xn }∞ n=1 converge in the space E? Here
 
1 1 1
1) E = l1 , xn = 0, . . . , 0, σ , , , . . ., σ > 1;
| {z } n (n + 1)σ (n + 2)σ
n−1
 
1
2) E = l2 , xn =  , 0, . . . , 0, 1, 0, 0, . . . ;
n | {z }
n−2

tn+1 tn+2
3) E = C 1 [0, 1], xn (t) = − ;
n+1 n+2
( t
e− n , t ∈ R \ Q,
4) E = L1 [0, 1], xn (t) = ;
0, t ∈ Q,
7.7. PROBLEMS 165

√ √
  
1

 n − n nt, t ∈ 0, ,
n



5) E = L2 [0, 1], xn (t) = .
 
1


t∈


 0, ,1 ,
n
Problem 7.19. Let [a, b] ⊂ R, 1 6 p < ∞, and let f ∈ Lp [a, b]. Define its modulus of continuity as follows

 1
p
Z
ω(f ; δ)p = sup  |f (x + h) − f (x)|p dµ .
 
06h6δ
[a,b−h]

Prove that ω(f ; δ)p → 0 as δ → +0. Here µ is the Lebesgue measure.

Problem 7.20. Let f, g ∈ L1 (R) and f (x) is bounded. Prove that the convolution
Z
(f ∗ g)(x) = f (t)g(x − t)dµ(t)
R

is continuous on R.

Problem 7.21. Let f ∈ Lp (R), 1 6 p < +∞ and h > 0. Define

Z
1
fh (x) = f (t)dµ(t).
2h
(x−h,x+h)

Prove that kfh kLp 6 kf kLp , and kf − fh kLp → 0 as h → +0.

Problem 7.22. Prove that a subspace L0 of a Banach space L is a Banach space if and only if L0 is
closed.
∞
 
Problem 7.23. Is the set L = x = (xn )∞
P
n=1 ∈ l p : x k = 0, x k ∈ R a linear closed subspace of the
k=1
∞
P
spaces lp , p > 1? Remind that in l1 the norm is kxk1 = |xn |.
n=1
 
 1 1 1 
Hint: Consider the cases p = 1 and p > 1 separately. Consider the sequence x(n) = 
1, − n , − n , . . . , − n , 0, 0, . . ..

| {z }
n

7.7.2 Linear inner product spaces

Problem 7.24. Let X be a complex normed space. Prove that X is a unitary space if and only if the
parallelogram identity holds:
kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ). (7.24.1)
Hint: As an inner product, consider the construction
1 i
(x, y) = {kx + yk2 − kx − yk2 } + {kx + iyk2 − kx − iyk2 }, (7.24.2)
4 4
√
where i = −1. Prove first that (x, y) is continuous w.r.t. its both variables. Then use that fact that any
real number can be approximated by a sequence of rational numbers.
Prove the same for real normed linear spaces.
166 CHAPTER 7. LINEAR SPACES

Problem 7.25. Prove that in lp , p 6= 2, it is impossible to introduce an inner product agreed with the
norm
+∞
!1
X p
kxkp = |xk |p .
k=1

Hint: Use the parallelogram identity.

Problem 7.26. Prove that in Lp [a, b], p 6= 2, it is impossible to introduce an inner product agreed with
the norm
 b  p1
Z
kf kp =  |f (x)|p dµ .
a

Problem 7.27. Prove that the space C[a, b] of continuous functions on [a, b] with norm

kf k = max |f (x)|
x∈[a,b]

cannot be a Hilbert space.

Problem 7.28. Prove that in a Euclidean (unitary) space X the following identity holds
2
1 x+y
kz − xk2 + kz − yk2 = kx − yk2 + 2 z −
2 2
for any x, y, z ∈ X. This identity is called the Apollonius identity.
Problem 7.29. Let X be the linear space of functions x(t) defined on (−∞, ∞) and satisfying the condition
+∞
Z
2
|x(t)|2 e−t dt.
−∞

Prove that the space X with the inner product

+∞
Z
2
x(t)y(t)e−t dt, x, y ∈ X
−∞

is a Hilbert space.
Problem 7.30. Prove that in an infinite dimensional Hilbert space H any closed unit ball contains
1
infinitely many non-intersecting closed balls with radius .
4
A closed ball B(a, r) in H of radius r with the centre at a is the set B(a, r) = {x ∈ H : kx − ak 6 r}.
Hint: Use the fact that any infinite dimensional Hilbert space contains an infinite orthonormal system of
elements (due to Gram-Schmidt process).
Problem 7.31. Let L be a subspace of a Hilbert space H. The subspace L⊥ = {x ∈ H : (x, y) = 0 ∀y ∈
L} of H is called the orthogonal complement of L. Prove that L⊥ is closed.
Problem 7.32. Let L be a closed subspace of a Hilbert space H. Prove that H = L ⊕ L⊥ , that is, any
x ∈ H can be uniquely represented in the form x = u + v where u ∈ L and v ∈ L⊥ and kx − uk = kvk. In
this case, u is the projection of x onto L.
Problem 7.33. Let L be a subspace of a Hilbert space H. Prove that L is dense in H (i.e. the closure
of L coincides with H: L = H) if, and only if, L⊥ = {0}.
Remind that the closure of L is the set consisting of all points of L and all limit points of L.
7.7. PROBLEMS 167

Problem 7.34. In the space L2 [−1, 1], construct projections of any function x ∈ L2 [−1, 1] onto the
subspaces of even and odd functions.
Problem 7.35. Prove that for a fixed n ∈ N the set
( n
)
X
Ln = x ∈ l2 , x = (ξ1 , ξ2 , . . .) : ξk = 0
k=1

is a closed subspace of l2 . Describe L⊥

n and find the distance between e1 = (1, 0, 0, . . .) and L:

ρ(e1 , Ln ) = inf ky − e1 kl2 .

y∈Ln

Find also lim ρ(e1 , Ln ).

n→∞

Problem 7.36. For the function et , find the polynomial p(t) of degree 2 such that the norm ket − p(t)k
is minimal in L2 [−1, 1].
Problem 7.37. Prove that the system (tn )∞
n=1 is not a basis in L2 [0, 1].

Hint: Prove that if (tn )∞

n=1 were a basis in L2 [0, 1], thenZany function f (t) ∈ L2 [0, 1] would be infinitely
s
differentiable. To do this, consider the integral g(s) = f (t)dt as a function on [0, 1], where f is an
0
arbitrary function on L1 [0, 1], and prove that g(s) is an infinitely differentiable function whenever (tn )∞
n=1
is a basis.
√
2
Problem 7.38. Prove that the system en (t) = sin µn t, where µn are positive solutions of the
sin µn
equation tan µ = µ, is an orthogonal system in L2 [0, 1].
Problem 7.39. Let α = (α1 , α2 , . . .) be a sequence of positive numbers: αn > 0, n ∈ N, and let l2,α be a
space of complex sequences x = (x1 , x2 , . . .) satisfying the condition
∞
X
αk |xk |2 < +∞.
k=1

Prove that the space l2,α with the inner product

∞
X
αk xk yk < +∞
k=1

is a complex separable Hilbert space. Construct an orthonormal basis in this space if

1) αn = e−n , n ∈ N;
2) αn = n, n ∈ N;
3) αn = n2 , n ∈ N.
Problem 7.40. Does the mapping f acting on the space X define a linear and/or continuous functional:
1)
Z2
f (x) = x(2) + tx(t)dt, X = C[−2, 2];
−2

2)
f (x) = x1 − 4x3 , X = l3 ;
168 CHAPTER 7. LINEAR SPACES

3)
Z1
f (x) = x2 (t)dt, X = C[0, 1];
0

4)
Z1
f (x) = x2 (t)dt, X = L2 [0, 1];
0

5)
Z1
f (x) = x(t) sin2 tdt, X = L2 [0, 1];
0

Problem 7.41. Does the mapping f (x) = x0 (0) acting on the set L of differentiable functions in C[−1, 1]
define a linear continuous functional?
∞
X ∞
X
Problem 7.42. Does the mapping f (x) = xk acting on the subspace L = {x ∈ l2 : xk < ∞}
k=1 k=1
define a linear continuous functional?
 
1 1 1 
 n , n , . . . , n , 0, 0, . . ..
Hint: Consider the sequence xn =  
| {z }
n

Problem 7.43. Let Rnp be the linear normed space of n-dimensional vectors x = (x1 , . . . , xn ) with the
norm
1

 n
 p
p
 P
 |xk | if 1 6 p < +∞;
kxkp = k=1
max |xk |


 if p = +∞.
16k6n

Find the general representation of linear functionals in Rnp .

Problem 7.44. Prove that any linear functional defined on a finite-dimensional space is continuous.
Problem 7.45. Prove that a linear functional in a normed space is continuous if and only if its kernel is
closed.
Problem 7.46. Let f be a linear functional defined on a linear normed space L. Prove that f is continuous
if and only if the sets {x ∈ L : f (x) < c} and {x ∈ L : f (x) > c} are open in L.
Problem 7.47. Find the norm of the following functionals
1)
x(ε) + x(−ε) − 2x(0)
f (x) = on C[−1, 1], ε ∈ (0, 1) is fixed;
ε2
2)
Z1 n  
1 X k
f (x) = x(t)dt − x on C[−1, 1], n∈N is fixed;
2n + 1 n
−1 k=−n

3)
∞
X k−1
f (x) = [1 − (−1)k ] xk where x = (xk )∞
k=1 ∈ l1 ;
k
k=1
7.7. PROBLEMS 169

4)
∞
X xk
f (x) = p where x = (xk )∞
k=1 ∈ l2 .
k=1
k(k + 1)

5)
1
Z2 √
f (x) = tx(t2 )dt on L2 [0, 1];
0

6)
1
Z2  
1
f (x) = x(t) sgn t − dt on L2 [0, 1];
2
0

Hint: To prove that a number M satisfying |f (x)| 6 M kxk is the norm of the functional f , try to find an
f (xn )
element x (kxk = 1) such that f (x) = M or a sequence (xn )∞
n=1 such that > M − εn , and εn → 0
kxn k
as n → ∞.
170 CHAPTER 7. LINEAR SPACES
Chapter 8

Divergent series. Cesàro and Abel

methods of summation.

8.1 Divergent series.

By definition, a number series
∞
X
S= am
m=0

converges to the number S if its partial sums

n−1
X
Sn = am
m=0

have finite limit as n → ∞ which is equal S.

According to this definition the series

1 − 1 + 1 − 1 + 1 − 1 + ··· (8.1.1)

diverges. Indeed, its partial sums Sn

S1 = 1, S2 = 0, S3 = 1, S4 = 0, . . . ,

that is,
S2k+1 = 1, S2k = 0, k ∈ N.
However, if we follow Euler and set (formally)

a = 1 − 1 + 1 − 1 + 1 − 1 + ··· ,

then we obtain
a = 1 − (1 − 1 + 1 − 1 + 1 − 1 + · · · ) = 1 − a,
1
so 2a = 1, and a = .
2
We can get the same result in another way. To do this consider the function
∞
X 1
f (z) = zm = . (8.1.2)
m=0
1−z

171
172 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

This series converges for |z| < 1 but the function f (z) exists for z = −1, so we get
∞
X 1 1
a = f (−1) = (−1)m = = .
m=0
1 − (−1) 2

However, the situation is not so good as it seems from the first view, because it can be shown that the
1
“sum” a of the series (8.1.1) can be a number different from . Indeed, let us represent the series (8.1.1)
2
as follows
a = 1 − 0 − 1 + 1 − 0 − 1 + 1 − 0 − 1 + ··· (8.1.3)
and consider the function
∞ ∞
X X 1 z2 1 − z2 1+z
g(z) = z 3m − z 3m+2 = 3
− 3
= = . (8.1.4)
m=1 m=1
1−z 1−z 1 − z3 1 + z + z2

From (8.1.3)–(8.1.4) it follows that

2
a = g(1) = .
3
Moreover, it is clear that for any m, n ∈ N, m < n,

1 + x + · · · + xm 1 − xm
gmn (x) = = = 1 − xm + xn − xm+n + x2n − · · · ,
1 + x + · · · + xn 1 − xn
so
m
a = gmn (1) = .
n
Additionally, if we substitute z = −2 into (8.1.2), then we get
+∞
X 1 1
f (−2) = 1 − 2 + 22 − 23 + · · · = (−2)n = = . (8.1.5)
n=0
1+2 3

1
Here the “sum” of a series with integer terms is fractional. There is no paradox here, since is not the
3
sum in usual sense. This is just a number functionally dependent on the series (8.1.5). Moreover, if we
take z = 2 in (8.1.2), then we obtain a really surprising thing
+∞
X 1
f (2) = 1 + 2 + 22 + 23 + · · · = 2n = = −1. (8.1.6)
n=0
1−2

Thus, without specific rules regarding “summing” of divergent series, we can obtain a lot of paradoxical
results.

8.1.1 General methods of summing

Let D be the set of all number series, Dc (⊂ D) be the set of all convergent series, and Da (⊂ Dc ) be the
set of all absolute convergent series. Let S : Dc 7→ C denote the ordinary summation operator defined on
the set Dc . To every convergent series S corresponds its sum: ∀σ ∈ Dc ∃S(σ) ∈ C, where S(σ) is the sum
of the series σ.
Consider now some other summation operator S ∗ defined on B ∗ ⊂ D. It is natural to assume that Dc ⊂
B . Then the operator S ∗ must possess the following properties:
∗

1) Regularity:
∀σ ∈ Dc S ∗ (σ) = S(σ).
8.1. DIVERGENT SERIES. 173

2) Shift invariance:
∞ ∞
! !
X X
S∗ a0 + an = a0 + S ∗ an .
m=1 m=1

3) Linearity: ∀α, β ∈ C, and ∀σ1 σ2 ∈ B

S ∗ (ασ1 + βσ2 ) = αS ∗ (σ1 ) + βS ∗ (σ2 ).

The second condition is less important, and some significant methods, such as Borel summation (see
below), do not possess it.
Sometimes S ∗ possess an additional property
4) S ∗ : B ∗ 7→ C is a homomorphism.
This means that S ∗ preserves the product of series. For example, the operator S of standard summation is
a homomorphism on the set Da , since the sum of the product of two absolutely convergent series is equal
to the product of their sums.
Two summation methods (operators) S1 and S2 defined on sets B1 and B2 are called consistent and
S2 is stronger than S1 if B1 ⊂ B2 ⊂ D, and
S2 (σ) = S1 (σ) ∀σ ∈ B1 .
In fact, it is possible to construct a hierarchy of some regular summation methods. That is, we can
find a sequence of summation operators Sk defined on sets Bk such that
Da ⊂ Dc ⊂ B1 ⊂ B2 ⊂ · · · ⊂ D
and
Sk+1 (σ) = Sk (σ) ∀σ ∈ Bk
for k = 0, 1, 2, . . .. Here S0 = S is the standard summation method.
Of course, it is possible to construct summation operators that are incomparable. In what follows, we
will consider a hierarchy of a sequence of summation operators.
Let us study two the most important for us summability methods.

8.1.2 Cesàro summability method

Let (sn )∞
n=1 be a sequence of numbers. Then the numbers

s1 + s2 + · · · + sN
σN = (8.1.7)
N
are called Cesàro means. The sequence (sn )∞ ∞
n=1 is called Cesàro convergent if the sequence (σN )N =1 of
Cesàto means has a finite limit.
Respectively, the series
X∞
an
m=0
are called Cesàro summable if the sequence of its partial sums is Cesàro convergent.
For instance, for the series (8.1.1), Cesàro means of the sequence of its partial sums have the form
1 k
σ2k = , σ2k−1 = , k = 1, 2, 3, . . . .
2 2k − 1
1 1
It is clear that lim σN = , so the sum a = of the series (8.1.1) obtained by Euler is the Cesàro sum
N →∞ 2 2
of this series.
Note that the Cesàro summability method possesses all three properties of generalized summation
operators.
174 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

1) Cesàro summation method is regular.

∞
P
Indeed, let a series am converges to a number A. Suppose first that A = 0, then for any δ > 0
m=0
n−1
P
there exists a number M1 > 0 such that for any n > M1 one has |Sn | < δ, where Sn = am .
m=0
Moreover, since the sequence Sn is convergent, it is bounded, so there exists a number B > 0 such
that |Sn | < B for all n ∈ N. Choose ε > δ > 0. Then there exists a number M2 > 0 such that
∀n > M2
M1 (B − δ)
< ε − δ.
n
Let N > max{M1 , M2 }, then we have

M
P1 N
P
|Sn | |Sn |
S1 + S1 + · · · + SN n=1 n=M1 +1 M1 (B − δ)
|σN | = 6 + 6 + δ < ε.
N N N N
∞
P
If A 6= 0, we can consider a new series a0 = a0 − A, and e
am such that e
e am = am , m > 1. Then
m=0
we have that the sum A eN → 0 as N → ∞. But clearly Sen = Sn − A
e of this series equals zero, so σ
eN = σN − A. Therefore, σN → A as N → ∞.
for all n, so σ

2) Cesàro summation method is shift invariant.

This property was, in fact, established in the previous item.

3) Cesàro summation method is linear.

This property is obvious because of linearity of the Cesàro means (8.1.7).

However, Cesàro summation operator is not a homomorphism as the following example shows.
1
Example 8.1.1. As we found above the series (8.1.1) is Cesàro summable to . Consider the square of
2
this series:
∞
!2 ∞ ∞
X X X
2 m
b=a = (−1) = ((−1)m + (−1)m + · · · + (−1)m )= (−1)m (m + 1).
m=0 m=0
| {z } m=0
m+1

The partial sum of this series are

2k−1
X 2k−2
X
S2k = (−1)m (m + 1) = −k, S2k−1 = (−1)m (m + 1) = k, k = 1, 2, . . . .
m=0 m=0

The series b is not summable, since S2k → +∞ and S2k−1 → −∞ as k → +∞. However, this series is also
not Cesàro summable. Indeed, its Cesàro means have the form

S1 + S1 + · · · + SN  0, N = 2k,
σN = = k+1
N  , N = 2k + 1,
2k + 1
so the sequence σN has no limit, and b is not Cesàro summable.
∞
P
Theorem 8.1.2. If a series an is Cesàro summable, then an = o(n) as n → ∞.
n=0
8.1. DIVERGENT SERIES. 175

N +1
Proof. Indeed, if σN → σ as N → ∞, then σN → σ. Therefore, we have
N
(N + 1)σN +1 − N σN SN +1
= −→ σ − σ = 0,
N N N →∞
−1
NP
where SN = am . Thus, we obtain
m=0

an Sn+1 − Sn Sn+1 n−1 Sn

= = − · −→ 0 − 1 · 0 = 0.
n n n n n − 1 n→∞

∞
(−1)m (m + 1) is non-Cesàro convergent according to this theorem.
P
It is clear that the series
m=0
The next summability method possesses all four properties of summability methods we mentioned
above.

8.1.3 Abel summability method

For a given series
∞
X
am (8.1.8)
m=0

consider the function

∞
X
A(r) = am r m .
m=0

where the series A(r) is supposed to be convergent for 0 6 r < 1. The series (8.1.8) is called Abel summable
if there exists the final limit
SA = lim A(r).
r→1−0

In this case, the number SA is called the Abel sum of the series (8.1.8).
First, note that the Abel summability method possesses all three properties of generalized summation
operators.

1) Abel summation method is regular.

This property can be proved directly. But it also follows from Theorem 8.2.2 below.

2) Abel summation method is shift invariant.

Indeed,
∞
X ∞
X ∞
X
lim am rm = a0 + lim r am rm−1 = a0 + lim am+1 rm
r→1−0 r→1−0 r→1−0
m=0 m=1 m=0

3) Abel summation method is linear.

This property is obvious.

Let us check whether the series b from Example 8.1.1 is Abel summable. For 0 6 r < 1, consider (formally)
the series
∞ ∞ ∞
X X d X m d 1 1
A(r) = (−1)m (m + 1)rm = (m + 1)(−r)m = z = = .
m=0 m=0
dz m=0 dz 1 − z z=−r (1 + r)2
z=−r
176 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

Since we supposed that r ∈ [0, 1), the differentiation of the series is possible. Moreover, the series A(r)
converges for r ∈ [0, 1), and we have
1 1
SA = lim 2
= .
r→1−0 (1 + r) 4
The Abel summation is related to the analytic continuation of functions.
Example 8.1.3. Consider the Riemann zeta-function
+∞
X 1
ζ(s) = .
n=1
ns

This series converges for Re s > 1, but the function ζ(s) can be analytically continued to the domain C\{1}
where it is meromorphic. For example, by analytic continuation it is known that
∞
X 1
ζ(−1) = m=− .
m=1
12

We can get this result using the previous example. Indeed,

∞
X ∞
X
−3ζ(−1) = ζ(−1) − 4ζ(−1) = m− 4m = 1 + (2 − 4) + 3 + (4 − 8) + 5 + (6 − 12) + · · ·
m=1 m=1

Therefore,
∞
X A 1
−3ζ(−1) = (−1)m (m + 1) = ,
m=0
4
thus
A 1
ζ(−1) = − .
12
The following example will be useful later.
Example 8.1.4. Consider the series
+∞
1 X
+ cos nθ, θ 6= 0. (8.1.9)
2 n=1

The limit lim cos nθ does not exist, so the series (8.1.9) diverges. Let us find out whether this series
n→∞
Abel summable. To do this, consider the series
+∞
1 X n
+ r cos nθ, (8.1.10)
2 n=1

where r ∈ [0, 1). The series of absolute values of its terms

+∞
1 X n
+ r | cos nθ|,
2 n=1
p p
converges, for instance, by Cauchy test, since n rn | cos nθ| = r n | cos nθ| 6 r < 1. Thus, the series (8.1.10)
converges absolutely, and therefore converges. Let us find its sum. Note that
+∞ +∞
!
reiθ
 
1 X n 1 X n inθ 1
+ r cos nθ = Re + r e = Re + .
2 n=1 2 n=1 2 1 − reiθ
8.1. DIVERGENT SERIES. 177

It is easy to see that

+∞
1 X n inθ 1 reiθ (1 − re−iθ ) 1 reiθ − r2
+ r e = + iθ −iθ
= + . (8.1.11)
2 n=1 2 (1 − re )(1 − re ) 2 1 − 2r cos θ + r2

Taking the real part of this expression we have

+∞
1 X n 1 r cos θ − r2 1 1 − r2
+ r cos nθ = + 2
= · .
2 n=1 2 1 − 2r cos θ + r 2 1 − 2r cos θ + r2

Thus, the Abel sum of the series (8.1.9) is the following.

+∞
!
1 X n inθ 1 1 − r2
lim Re + r e = lim · = 0, θ 6= 0.
r→1−0 2 n=1 r→1−0 2 1 − 2r cos θ + r 2

Furthermore, the series

∞
X
sin nθ, θ 6= 0, (8.1.12)
n=1
diverges. However, the function
∞
X
rn sin nθ, 0 6 r < 1,
n=1
is the imaginary part of the series (8.1.11), so
∞ ∞
!
X
n
X
n inθ r sin θ
r sin nθ = Im r e = ,
n=1 n=1
1 − 2r cos θ + r2

and the Abel sum of the series (8.1.12) has the form
∞
X r sin θ 1 sin θ 1 2 sin θ2 sin θ2 1 θ
lim rn sin nθ = lim = · = · = cot
r→1−0
n=1
r→1−0 1 − 2r cos θ + r 2 2 1 − cos θ 2 2 sin2 θ2 2 2

Note that the series

∞
X
qn , q > 1,
n=0
is not Abel summable, since
∞
X
lim rn q n = +∞.
r→1−0
n=0
Thus, the series (8.1.5) and (8.1.6) are not Abel summable. Later we will introduce the Euler summability
method that allow to “sum” the series (8.1.5).

8.1.4 Product of series

Historically, one of the first methods of “summing” divergent series was the Abel (Abel–Poisson) summation
method which appeared to be useful for the product of series. Implicitly, Leibnitz and Euler used this
method.
P∞ P∞
So, it is known that if two series an and bn converge absolutely to some numbers A and B,
n=0 n=0
respectively, then their (Cauchy) product
∞
X ∞
X ∞
X
an · bn = (a0 bn + a1 bn−1 + · · · + an−1 b1 + an b0 )
n=0 n=0 n=0

converges to the number C = A · B.

178 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

Example 8.1.5. Consider the series

∞
1 X (2m − 1)!! 1 1·3
√ = (−1)m =1− + − ··· , (8.1.13)
2 m=0 (2m)!! 2 2·4

that can be obtained from the functional series

∞
1 X (2m − 1)!! m
√ = x , (8.1.14)
1 − x m=0 (2m)!!

whose convergence radius is 1. Indeed, from the Stirling formula

√  m m
m! ≈ 2πm · as m → +∞
e
we have
√
1 2 πm · 4m · m2m e2m
 
(2m − 1)!! 1 (2m)! 1 2m 1
am = = m· = m ≈ m· 2m
· 2m
=√ ,
(2m)!! 4 m! m! 4 m 4 e 2πm · m πm
so s
1 √ 1
= lim m
am = lim m
√ = 1.
R m→∞ m→∞ πm
Thus, the series (8.1.14) converges absolutely for |x| < 1. It obviously diverges for x = 1, but it
converges for x = −1. Indeed, since
1 (2m)! 1 (2m + 2)! 1
am − am+1 = m
· − m+1 · = am · > 0,
4 m! m! 4 (m + 1)!(m + 1)! 2(m + 1)
and
1
am ≈ √ −→ 0,
πm m→∞
the series (8.1.13) converges by Leibnitz’s test. But this series is not absolutely convergent! We prove now
that its square is a divergent series. To do this let us recall the Chu-Vendermond identity
n     
X α β α+β
= .
m=0
k n−k n

1
If α = β = − , then we have
2
(−1)n · n!
   
α+β −1 (−1)(−2) · · · (−n)
= = = = (−1)n ,
n n n! n!
and
    
1 1 1
1 − − − 1 ··· − − k + 1
(−1)k (2k − 1)!! (−1)k (2k)! (−1)k 2k
  
−2 2 2 2
= = = k = ,
k k! 2k · k! 2 · k! 2k · k! 4k k
 1 
(−1)n−k 2(n − k)
 
−2
= .
n−k 4n−k n−k
So from the Chu-Vandermond identity it follows that
n  1 
− 12
  
X −2 −1
= ,
m=0
k n−k n
8.1. DIVERGENT SERIES. 179

or
n   
X 2k 2n − 2k
= 4n .
m=0
k n − k
Consider now the square of the series (8.1.13):
∞
!2 ∞ n ∞ n     X ∞
X
m
X X X X 1 2k 1 2n − 2k
(−1) am = (−1)n ak an−k = (−1)n k n−k
= (−1)n .
m=0 n=0 n=0
4 k 4 n − k n=0
k=0 k=0

1
However, the square of the left-hand side of (8.1.13) equals . An we know that this is the Abel (and
2
∞
(−1)n .
P
Cesàro) sum of the series
n=0

The following theorem explain this phenomenon.

∞
P ∞
P
Theorem 8.1.6. If series an and bn are Abel summable with sums A and B, respectively, then
n=0 n=0
their product is Abel summable with sum A · B.
Thus, the Abel summation operator is a homomorphism.
Proof. By assumption the series
∞
X ∞
X
an rn and bn rn
n=0 n=0

converge for any r ∈ [0, 1). Therefore, the series

∞
X ∞
X
A(z) := an z n and B(z) := bn z n
n=0 n=0

converge absolutely for |z| < 1 by Abel’s theorem (from the power series theory). So we have
∞
! ∞ !
X X
lim an rn bn rn = lim A(r) · B(r) = A · B.
r→1−0 r→1−0
n=0 n=0

1
This theorem confirms that the square of the series (8.1.13) is Abel summable to .
2
Example 8.1.7. Consider again the series
∞
X
(−1)m .
m=0

It is easy to see that

!2  
∞
X ∞
X X∞
(−1)m = (−1)m + (−1)m + · · · (−1)m  = (−1)m−1 m.
m=0 m=0
| {z } m=0
mtimes
 2
1 1
The initial series is Cesàro summable to but its square is not Cesàro summable. It is Abel summable to .
2 2
∞
P ∞
P
Corollary 8.1.8. If series an and bn are summable with sums A and B, respectively, then their
n=0 n=0
product is Abel summable with sum A · B.
180 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

8.2 Tauberian theorems. Higher methods of summation of di-

vergent series.
8.2.1 Relation between Cesàro and Abel summability methods
To find out how Cesàro and Abel summability methods are related, we need the so-called Abel summation
formula.

Theorem 8.2.1 (Abel summation formula). Let (an )N N

n=0 and (bn )n=0 be finite number sequences. Then

N
X N
X −1
an bn = aN BN +1 − aM BM − (an+1 − an )Bn+1 . (8.2.1)
n=M n=M

where
m−1
X
B0 = 0, Bm = bn
n=0

Proof. Since bm = Bm+1 − Bm , we have

N
P N
P N
P N
P
an bn = an (Bn+1 − Bn ) = (an Bn+1 − an−1 Bn ) − (an Bn − an−1 Bn ) =
n=M n=M n=M n=M

−1
NP −1
NP
= aN BN +1 − aM −1 BM − (an+1 − an )Bn+1 = aN BN +1 − aM BM − (an+1 − an )Bn+1 ,
n=M −1 n=M

as required.

Theorem 8.2.2 (Frobenius). The Cesàro and Abel summation methods are consistent, and the Abel
summation method is stronger.

Proof. We have to prove that if the series (8.1.8) is Cesàro summable to a number σ, then it is Abel
summable to the same number. Due to shift invariance of both methods, it is sufficient to prove the
theorem in the case σ = 0.
So, by assumption the Cesàro means σN converge to zero as N → ∞. This, in particular, means that
the sequence (σN )∞
N =1 is bounded. Thus, there exists a positive number B such that |σN | 6 B for any
N ∈ N, and for any ε > 0, there exists a number N1 ∈ N such that for any N > N1 ,
ε
|σN | < . (8.2.2)
3
n−1
P
Let Sn = am , n ∈ N, and S0 := 0. By Abel summation formula (8.2.1) we have
m=0

N
X N
X −1 N
X −1
m N m+1 m N
am r = SN +1 r − S0 − (r − r )Sm+1 = SN +1 r + (1 − r) Sm+1 rm . (8.2.3)
m=0 m=0 m=0

From the same formula and from (8.1.7) we get

N
X −1 N
X −1 N
X −2
Sm+1 rm = [(m + 1)σm+1 − mσm ]rm = N σN rN −1 − (m + 1)σm+1 (rm+1 − rm ) =
m=0 m=0 m=0
(8.2.4)
N
X −2
= N σN rN −1 + (1 − r) (m + 1)σm+1 rm ,
m=0
8.2. TAUBERIAN THEOREMS. HIGHER METHODS OF SUMMATION OF DIVERGENT SERIES.181

where we supposed that σ0 = 0. So, from (8.2.3)–(8.2.4) we obtain

N
X N
X −2
am rm = SN +1 rN + (1 − r)N σN rN −1 + (1 − r)2 (m + 1)σm+1 rm . (8.2.5)
m=0 m=0

Next, it is clear that

Sn n−1
= σn − σn−1 −→ 0,
n n n→∞
and that
|N σN | 6 BN ∀N ∈ N,
so the sequence N σN grows (if any) slower than N , that is σN = O(N ) and SN = o(N ). Thus, for
any ε > 0 there exists a number N2 ∈ N such that for all N > N2 ,
ε ε
SN +1 rN < . and |N σN rN −1 | < . (8.2.6)
3 3
Therefore, (8.2.2), (8.2.5) and (8.2.6) imply that for any N > max{N1 + 2, N2 },
N NX1 −1 N −2
X ε ε ε X
am rm < + (1 − r) + (1 − r)2 (m + 1)σm+1 rm + (1 − r)2 (m + 1)rm (8.2.7)
m=0
3 3 m=0
3
m=N1

Let now N → ∞. Then from (8.2.7) we get

∞ NX1 −1
X ε ε ε (1 + N1 − N1 r)rN1
|A(r)| = am rm < + (1 − r) + (1 − r)2 B (m + 1)rm + (1 − r)2 =
m=0
3 3 m=0
3 (1 − r)2

N1
ε 2
X ε
= (2 − r) + (1 − r) B (m + 1)rm + (1 + N1 − N1 r)rN1 .
3 m=0
3

It is easy to see now that |A(r)| < ε as r → 1 − 0. Since ε is arbitrary, we have

lim A(r) = 0,
r→1−0

as required.
Example 8.2.3. Consider again the series (8.1.9). Its partial sums have the form
m−1 m−1
1 X 1 1 X θ
Sm (θ) = + cos kθ = + cos kθ sin =
2 2 θ 2
k=1 sin k=1
2
 
1
m−1
X      sin m − θ
1 1 1 1 2
= − sin k − θ − sin k + θ = ,
2 θ 2 2 θ
2 sin k=1 2 sin
2 2
so the Cesàro means are the following
N N  
1 X 1 X 1 θ
σN (θ) = Sm (θ) = sin m − θ sin =
N m=1 θ 2 2
2N sin2 m=1
2

Nθ 2
 
1
N
1 − cos N θ 1  sin
2 
X
= [cos (m − 1) θ − cos mθ] = = .
2 θ 2 θ 2N
 θ 
4N sin m=1 4N sin sin
2 2 2
182 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

It is easy to see that σN (θ) → 0 as N → ∞ for any θ 6= 0 that agrees with Example 8.1.4 and Theorem 8.2.2.
Consider again the series (8.1.12). Its partial sums have the form

m m
X 1 X θ
Sm (θ) = sin kθ = sin kθ sin =
θ 2
k=1 sin k=1
2
 
θ 1
m       cos − cos m + θ
1 X 1 1 2 2
= cos k − θ − cos k + θ = ,
θ 2 2 θ
2 sin k=1 2 sin
2 2

so the Cesàro means are the following

N N    
1 X 1 X θ 1 θ
σN (θ) = Sm (θ) = cos − cos m + θ sin =
N m=1 θ 2 2 2
2N sin2 m=1
2
N
1 θ 1 X 1 θ sin θ − sin(N + 1)θ
= cot + [sin mθ − sin (m + 1) θ] = cot + .
2 2 θ 2 2 θ
4N sin2 m=1 4N sin2
2 2

1 θ
It is easy to see that σN (θ) → cot as N → ∞ for any θ 6= 0 that agrees with Example 8.1.4 and
2 2
Theorem 8.2.2.

Remark 8.2.4. As we mentioned above, from this theorem it follows that the Abel summation method
is regular. Indeed, if a series is summable, then it is Cesàro summable, since Cesàro summability method
is regular. Now Theorem 8.2.2 guarantees that the series is Abel summable. This fact is usually known as
Abel’s theorem.

∞
P ∞
P
Corollary 8.2.5. If series an and bn are Cesàro summable with sums A and B, respectively, then
n=0 n=0
their product is Abel summable with sum A · B.

Finally, let us prove a fact that is of use later.

n−1 ∞
am rm of a series
P P
Proposition 8.2.6. Let the Abel partial sums am be bounded as r → 1 − 0. If,
  m=0 m=0
1 n−1
P
additionally, am = O , then the partial sums am of the given series are bounded.
m m=0

1
Proof. Let rN = 1 − . Then N (1 − rN ) = 1, and rN → 1 − 0 whenever N → +∞. Moreover, since
  N
1
am = O , there exists M > 0 such that m|am | 6 M for any m ∈ N (for large m it follows from
m
the assumption of the proposition, and for other, finitely many, m’s we can always find such a bound).
8.2. TAUBERIAN THEOREMS. HIGHER METHODS OF SUMMATION OF DIVERGENT SERIES.183

Consider the difference

N
X −1 ∞
X N
X −1 ∞
X
m m m m
|SN − ArN | = (am − am rN )− am rN 6 |am | · (1 − rN )+ |am | rN 6
m=0 m=N m=0 m=N

N −1 ∞ m
X
m−1
X rN
6 (1 − rN ) |am | (1 + rN + · · · + rN )+ m · |am | · 6
m=0
m
m=N

N −1 ∞ m N −1 ∞
X X rN X M X m
6 (1 − rN ) m · |am | + M 6 (1 − rN )M 1+ rN 6
m=0
m m=0
N
m=N m=N

∞
M X m M 1
6 (1 − rN )M N + r =M+ · = 2M as N → +∞.
N m=0 N N 1 − rN

8.2.2 Tauberian theorems

From the properties of Abel and Cesàro summation methods we get that the Abel summation method is
stronger than the Cesàro summation method, which, in its turn, is stronger than the standard summation
method. This means that any convergent series is Abel and Cesàro convergent, and any Cesàro convergent
series is Abel convergent. But are there some conditions on the given series such that the converse
statements becomes true? In fact, such conditions exist, and the corresponding results on this topic are
usually called Tauberian theorems after A. Tauber who proved the following theorem in 1897.
∞
 
P 1
Theorem 8.2.7 (Tauber). Let the series am be Abel summable to a number S, and let am = o .
m=0 m
P∞
Then am is summable to S.
m=0

−1
NP
Proof. Let SN = am , and
m=0
∞
X
lim am rm = lim A(r) = S. (8.2.8)
r→1−0 r→1−0
m=0
 
1
Since am = o , the sequence (mam )∞ m=1 is bounded, that is, there exists B > 0 such that
m
|mam | 6 B for any m ∈ N. Moreover, for any ε > 0 there exists a number N1 ∈ N such that for
any m > N1
ε2
|mam | 6 .
9B
Now from (8.2.8) it follows that for any ε > 0 there exists δ > 0 such that
ε
|A(r) − S| < whenever 0 6 (1 − r) < δ.
3
Choose a number N2 ∈ N such that
ε
<δ ∀N > N2 .
3N B
ε
Let r := 1 − . Then we have
3N B
ε ε
1−r = <δ and (1 − r)N = .
3N B 3B
184 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

Thus, for any N > max{N1 , N2 }, one obtains

−1 −1 ∞ ∞
N N
!
X X X X
m m m
|SN − S| = am − S = am (1 − r ) − am r + am r −S 6
m=0 m=0 m=N m=0

N −1 ∞ ∞
X X rm X ε2 1 ε
6 (1 − r) m|am | + m|am | + am rm − S 6 (1 − r)N B + · + < ε.
m=0
m m=0
9B N (1 − r) 3
m=N

From this theorem and from Theorem 8.2.2 we immediately obtain the following fact.
∞
 
P 1
Corollary 8.2.8. Let the series am be Cesàro summable with the sum σ, and let am = o . Then
m=0 m
∞
P
am is summable with the sum σ.
m=0

8.3 Some other summability methods

Here we briefly consider some additional regular summability methods.

8.3.1 Hölder summability method

This method is the iteration of Cesàro method. Namely, if we have a series
∞
X
am ,
m=0

n−1
(0)
am are called initial or 0th Hölder sums. Respectively, Cesàro means
P
then its partial sums σn =
m=0
1 Pn
(1) (0)
σn = σ are called 1st Hölder sums. And in general, the lth Hölder sums are
n k=1 k
n
1 X (l−1)
σn(l) = σk .
n
k=1

(0) (1)
So if the sequence (σn )∞ ∞
n=1 diverges, we use the sequence (σn )n=1 . If this diverges too, we use the next
(2)
Hölder sums (σn )∞ n=1 , and so on.
(l)
The series is called lth Hölder summable with the sum S if lim σn = S.
n→∞

Example 8.3.9. Consider the series

∞
X
(−1)m (m + 1).
m=0

As we established earlier, for N ∈ N,

( 
−k, N = 2k,  0, N = 2k,
(0) (1)
σN = σN = k
k, N = 2k − 1,  , N = 2k − 1.
2k − 1
8.3. SOME OTHER SUMMABILITY METHODS 185

Then we have1
 
N +1 N +1 N +1
h i h i h i
2 2   2
(2) 1 X k 1 X 2k − 1 + 1  N +1 +
1  X 1 
σN = = = 6
N 2k − 1 2N 2k − 1 2N  2 2k − 1 
k=1 k=1 k=1

  N
1 N +1 1 X1 1
6 + −→ ,
2N 2 2N k N →+∞ 4
k=1

1 1 PN 1
since the sequence → 0 as k → ∞, so its Cesàro sums → 0 as N → ∞. On the other hand, from
k N k=1 k
1 1 ∞
(2) (2)
(−1)m (m + 1)
P
the formula above it follows that σN > , thus σN → as N → ∞. Thus the series
4 4 m=0
is second Hölder summable, and its Hölder sum coincides with the Abel sum of this series.

8.3.2 Voronoy summability method

Let (pn )∞
n=1 be a sequence of some positive numbers. Construct the sequences

Pn = p1 + p2 + · · · + pn

and
pn S1 + pn−1 S2 + · · · + p1 Sn
ωn = ,
Pn

where (Sn )∞
n=1 is the sequence of partial sums of the given series.
The series is called Voronoy summable with the sum S if lim ωn = S.
n→∞

8.3.3 Higher Cesàro summability methods

These methods are particular cases of the Voronoy summability method and are denoted (C, l), l = 1, 2, . . ..
In Voronoy’s method we set
 
n+l−2
pn := , n = 1, 2, . . . .
l−1

Then we have by induction

n    
X m+l−2 n+l−1
Pn = = .
m=1
l−1 l

Consider the numbers

n+l−2 n+l−3 l−1




l−1 S1 + l−1 S2 + ··· + l−1 Sn
ωn = n+l−1

,
l

where (Sn )∞
n=1 is the sequence of partial sums of the given series.
The series is called lth Cesàro summable with the sum S if lim ωn = S.
n→∞
Note that the method (C, 1) is the Cesàro summability method defined above.

1 Here [α] denote the largest integer not exceeding α whenever α > 0.
186 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

8.3.4 Borel summability method

∞
P
For a given series am , consider the function
m=0

∞
P xm
Sm+1 ∞
m=0 m! X xm
F (x) = ∞ m = e−x Sm+1 ,
P x m!
m=0
m=0 m!

n−1
P
where Sn = am .
m=0
∞
P
The series am is called Borel summable with the sum S if lim F (x) = S.
m=0 x→+∞

8.3.5 Euler summability method

∞
P
For a given series am , consider the following new series
m=0

∞
X bn
n+1
,
n=0
2

where      
n n n
bn = a0 + a1 + · · · + an = ∆n a0 . (8.3.9)
0 1 n
Here ∆ak = ak+1 − ak is a finite difference.
∞
P ∞
P
The series am is called Euler summable with the sum S if the series bn is summable (in the
m=0 n=0
standard sense) with the sum S.
Note that if we formally set
∞ ∞
X X bn y n
fb(x) = am , gb(y) = ,
m=0 n=0
2n+1

where bn are defined in (8.3.9), then

 y 
2
gb(y) = fb y .
1− 2

Example 8.3.10. Consider the series

∞
X
qn . (8.3.10)
n=0

Here an = q n . The numbers bn in the Euler summation method have the form
     
n n n
bn = a0 + a1 + · · · + an = (1 + q)n ,
0 1 n

so the Euler sum of the series (8.3.10) is the following

∞ ∞
X bn X (1 + q)n 1
S= n+1
= n+1
= .
n=0
2 n=0
2 1−q
8.3. SOME OTHER SUMMABILITY METHODS 187

1+q
Here the Euler series converges if < 1, that is, if −3 < q < 1. Thus, the series (8.3.10) is Euler
2
summable if −3 < q < 1, while it is ordinary summable only if −1 < q < 1. If q = −2 we have
∞
X E 1
(−2)n = .
n=0
3

∞ 1
q n is Borel summable with the (Borel) sum
P
It is also easy to show that the series whenever q < 1.
n=0 1−q
188 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.

8.4 Problems
Problem 8.1. Using Abel’s summation formula, prove the Dirichlet test for convergence of a series: if
−1
NP
the partial sums of the series BN = bm are bounded, and (an )∞
n=1 is a sequence of real numbers that
m=0
∞
P
decreases monotonically to 0, then the series an bn converges.
n=0
∞
(−1)m am converges if (an )∞
P
From Dirichlet’s test deduce Leibnitz theorem claiming the the series n=1
m=0
is a sequence of real numbers that decreases monotonically to 0.
Problem 8.2. Prove that any lth Hölder summable series is (l + 1)th Hölder summable to the same sum.
Are the Hölder summability methods regular?
Problem 8.3. Prove that the Voronoy summability method is regular if, and only if, the following con-
dition holds
pn
lim = 0.
n→∞ Pn

Problem 8.4. Using the result of Problem 8.3 prove that the Cesàro summability methods (C, l), l ∈ N,
are regular.
Problem 8.5. Prove that the Euler summability method is regular.
Problem 8.6. Prove that the Borel summability method is regular.
Problem 8.7. Prove that if a series is k th Cesàro summable, l ∈ N, then it is (k + 1)th Cesàro summable
with the same sum. Show by example that converse is not true.
∞
(−1)n−1 nk , k ∈ N.
P
Hint: Consider the series
n=1

Problem 8.8. Prove that if a series is lth Cesàro summable, then it is Abel summable to the same sum.
+∞
X
Problem 8.9. A series an is summable if, and only if, it is Cesàro summable with the same sum, and
n=0

un = a1 + 2a2 + · · · + nan = o (n) as n → +∞.

+∞
X
Problem 8.10. A series an is summable if, and only if, it is Abel summable with the same sum, and
n=0

un = a1 + 2a2 + · · · + nan = o (n) as n → +∞.

+∞
un X un rn+1
Hint: For necessity, prove the formula SN − σN = . For sufficiency, prove that lim =
n r→1−0
n=1
n(n + 1)
+∞
X
S − a0 , where S = lim an rn , and use Theorem 8.1.2.
r→1−0
n=0

Problem 8.11. Prove that if a series is Cesàro summable (first Cesàro summable) and mam > −C,
m ∈ N, for some C > 0, then it is summable with the same sum.
Problem 8.12. Find Cesàro and Abel sums of the series
∞
X
(−1)n−1 n2k sin nθ, k = 1, 2, . . . ,
n=1

and
∞
X
(−1)n−1 n2k cos nθ, k = 0, 1, 2, . . .
n=1
8.4. PROBLEMS 189

Problem 8.13. Find Cesàro and Abel sums of the series

∞
X
(−1)n−1 n2k+1 sin nθ,
n=1

and
∞
X
(−1)n−1 n2k+1 cos nθ,
n=1

for k = 0, 1, 2, . . . .
Problem 8.14. Prove the identities
12k − 22k + 32k − · · · = 0, k = 1, 2, . . . ,

12k+1 − 32k+1 + 52k+1 − · · · = 0, k = 0, 1, 2, . . . ,

22k+2 − 1
12k+1 − 22k+1 + 32k+1 − · · · = B2k+2 , k = 0, 1, 2, . . . ,
2k + 2
where Bk are Bernoulli numbers.
Here the “sums” of the given series are considered in Abel sense.
Hint: Consider the Taylor series of the function tan θ.
Problem 8.15. Find Cesàro and Abel sums of the series
∞
X
(−1)n−1 (2n − 1)2k cos nθ, k = 1, 2, . . . ,
n=1

and prove the identity

1
12k − 32k + 52k − · · · =
E2k , k = 1, 2, . . . ,
2
where Ek are Euler numbers, and the value in the right-hand side of the identity is supposed to be the
Abel sum.
Hint: Consider the Taylor series of the function sec θ.
190 CHAPTER 8. DIVERGENT SERIES. CESÀRO AND ABEL METHODS OF SUMMATION.
Chapter 9

Trigonometric Fourier series

I cover pages 2–15 from the book [10].

9.1 Trigonometric Fourier series

In Section 7.6.5 we studied properties of Fourier series w.r.t. an orthonormal basis (en )∞n=1 in L2 [a, b]. In
this section we will study mostly trigonometric Fourier series in L1 [a, b], C[a, b], and C (k) [a, b] as well as
on the set of all piece-wise continuous or Lipschitz functions.

Definition 9.1.1. A function f is called periodic on R if there exists a number A > 0 such that

f (x) = f (x + A) ∀x ∈ R. (9.1.1)

The minimal number T satisfying (9.1.1) is called the period of the function f . In this case, the function
f is called T -periodic.

Suppose now f ∈ L1 [−π, π] and 2π-periodic. This, in fact, means that f is Lebesgue integrable over
any interval of the real line (see Homework N13).

Definition 9.1.2. The series

∞
a0 X
+ (am cos mx + bm sin mx) (9.1.2)
2 m=1

is called the trigonometric Fourier series of the function f if its coefficients are related to the function f
as follows
Zπ Zπ
1 1
am = f (x) cos mxdx, bm = f (x) sin mxdx, m = 0, 1, 2, . . .
π π
−π −π

The coefficients am and bm exist, since f is integrable and cos mx and sin mx are bounded. We will
also consider a complex form of the trigonometric Fourier series of the function f :
+∞
X
cn einx , (9.1.3)
n=−∞

where the coefficients cn have that form

Zπ
1
cn = f (x)e−inx dx, n ∈ Z.
2π
−π

191
192 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

In what follows we will consider all functional spaces over the field of real numbers, F = R. In this
case, the coefficients of the series (9.1.2) and (9.1.3) are related as follows

a0 an − ibn an + ibn
c0 = , cn = , c−n = , n = 1, 2, . . .
2 2 2

One of the principal topic in the theory of Fourier series is the convergence which can be understanding
in difference senses. We already studied convergence of Fourier series in L2 [a, b]. However, the theory of
Fourier series in L1 is more complicated. Also, if we try to study the pointwise (a.e.) convergence of Fourier
series, we will see that from this point of view Fourier series might have rather weird behaviour even for
continuous functions. Nevertheless, the uniform convergence1 (the convergence in the space C[−π, π]) do
not bring any troubles to researchers, as the following theorem shows.

Theorem 9.1.3. Let a function f (x) is defined on the interval [−π, π] and is expanded into a trigonometric
series of the form (9.1.2) which is uniformly convergent on [−π, π]. Then the coefficients am and bm of
the series are uniquely determined by the function f , and the series is the Fourier series of the function f .

Proof. By assumption, we have ∀x ∈ [−π, π],

∞
a0 X
f (x) = + (am cos mx + bm sin mx), (9.1.4)
2 m=1

where the series converges uniformly on [−π, π]. So we can integrate this series over the interval [π, π]:
 
Zπ Zπ ∞ Zπ Zπ
a0 X
f (x)dx = dx + am cos mxdx + bm sin mxdx = a0 π,
2 m=1
−π −π −π −π

so
Zπ
1
a0 = f (x)dx. (9.1.5)
π
−π

Now we multiply the series (9.1.5) by cos nx. The resulting series is also uniformly convergent on
[−π, π], since cos nx is bounded. So we can integrate this series:
 
Zπ Zπ ∞ Zπ Zπ
a0 X
f (x) cos nxdx = cos nxdx + am cos mx cos nxdx + bm sin mx cos nxdx = an π,
2 m=1
−π −π −π −π

so
Zπ
1
an = f (x) cos nxdx. (9.1.6)
π
−π

Analogously we obtain
Zπ
1
bn = f (x) sin nxdx. (9.1.7)
π
−π

Thus, if the function is expanded into a uniformly convergent series on [−π, π], then this series is its
Fourier series which is uniquely determined by f .
1 As we established in Section 7.4, the uniform convergence is the strongest convergence among all other types of convergence

we considered.
9.2. BASIC THEORY OF FOURIER SERIES 193

For a given function f ∈ L1 [−π, π], we will write

∞
a0 X
f∼ + (am cos mx + bm sin mx),
2 m=1
or
+∞
X
f∼ cn einx
n=−∞

to emphasize that the corresponding trigonometric series (in real or complex form) is the Fourier series of
the function f . We will use symbol “=” instead of “∼” if it is known that the Fourier series converges to
f pointwise.

9.2 Basic theory of Fourier series

We start to study properties of trigonometric Fourier series in L1 [−ππ] from proving the Riemann-
Lebesgue. For the space L2 [−π, π], this lemma is a simple consequence of the convergence of the series of
Fourier coefficients (see Corollaries 7.5.12 and 7.6.49).
First we prove a more general fact.
Theorem 9.2.4 (Lebesgue). Let (ϕn )∞ n=1 be a system of measurable functions defined on [a, b]. Suppose
that there exists a constant K > 0 such that for any n ∈ N and t ∈ [a, b],
|ϕn (t)| 6 K. (9.2.8)
If for any c ∈ [a, b],
Zc
lim ϕn (t)dt = 0, (9.2.9)
n→∞
a

then for any f ∈ L1 [a, b], the identity

Zb
lim f (t)ϕn (t)dt = 0 (9.2.10)
n→∞
a

holds.
Proof. Let [α, β] ⊂ [a, b]. Then (9.2.9) implies that
Zβ
lim ϕn (t)dt = 0. (9.2.11)
n→∞
α

Consider now a continuous function f defined on [a, b], and for a fixed ε > 0 split the interval [a, b] into
subintervals [tk−1 , tk ], k = 1, . . . , m, (t0 = a, tm = b), so that the oscillation2 osc(f, tk ) < ε, k = 0, . . . , m,
that is possible because of continuity of f (t). Then we have
Zb tZk+1 tZk+1
m−1
X m−1
X
f (t)ϕn (t)dt = [f (t) − f (tk )]ϕn (t)dt + f (tk ) ϕn (t)dt. (9.2.12)
a k=1 t k=1 tk
k

But we have
tZk+1

[f (t) − f (tk )]ϕn (t)dt 6 Kε(tk+1 − tk ), (9.2.13)

tk
2 See Definition 5.4.3.
194 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

so the first sum in (9.2.12) does not exceed Kε(b − a), while the second sum in (9.2.12) tends to zero as
n → ∞ according to (9.2.11). Thus, there exists N ∈ N such that for n > N ,

Zb
f (t)ϕn (t)dt 6 ε[K(b − a) + 1],
a

and (9.2.10) is established for continuous functions.

Now by Theorem 7.4.14, for any f ∈ L1 [a, b] and for any ε > 0 there exists a continuous function g
ε
such that kf − gkL1 < , so we have
2K(b − a)

Zb Zb Zb
ε ε
f (t)ϕn (t)dt 6 [f (t) − g(t)]ϕn (t)dt + g(t)ϕn (t)dt < + =ε
2 2
a a a

for sufficiently large n.

Example 9.2.5. Let ϕn (t) = cos nt. Then

Zc
sin nc − sin na
ϕn (t)dt = −→ 0,
n n→∞
a

so the system (ϕn (t)∞

n=0 satisfy all the conditions of Theorem 9.2.4.

This implies the so-called Riemann–Lebesgue lemma.

Theorem 9.2.6 (Riemann-Lebesgue). For any function f ∈ L1 [a, b], the following holds

Zb Zb
lim f (t) cos ntdt = lim f (t) sin ntdt = 0.
n→∞ n→∞
a a

In particular, we have that for any function f ∈ L1 [−π, π], its Fourier coefficients w.r.t. the sys-
tem (7.6.13) vanish as n → ∞:
Zπ
1
an = f (t) cos ntdt −→ 0,
π n→∞
−π

and
Zπ
1
bn = f (t) sin ntdt −→ 0.
π n→∞
−π

Remark 9.2.7. If, for a given system (ϕn )∞ n=1 , the identity (9.2.10) holds for any f ∈ L1 [a, b], then the
system (ϕn )∞
n=1 is called weakly convergent to 0.

9.2.1 Series by sines or cosines with monotonically decreasing coefficients.

It is clear that if a function f is odd (resp. even), then an = 0 (resp. bn = 0) for any n.

Theorem 9.2.8. If an & 0, then the series

+∞
a0 X
+ an cos nx
2 n=1
9.2. BASIC THEORY OF FOURIER SERIES 195

converges at any point x ∈ R except possibly the points 2πm, m ∈ Z. Moreover, for any δ ∈ (0, π) this
series converges uniformly on [δ, 2π − δ].
If bn & 0, then the series
+∞
X
bn sin nx
n=1

converges at any point x ∈ R except possibly the points 2πm, m ∈ Z. Moreover, for any δ ∈ (0, π) this
series converges uniformly on [δ, 2π − δ].

Theorem 9.2.9. Then the series

∞
X
bn sin nx
n=1

converges uniformly on [0, 2π−] if and only if nbn → 0 as n → +∞.

Remark 9.2.10. The previous theorem shows that there exist trigonometric series that converge uniformly
but no absolutely. Indeed, by Theorem 9.2.9 one has that the series
∞
X sin nx
n=1
n ln n

1
converges uniformly since nbn = → 0 as n → +∞. At the same time, it does not converge absolutely,
ln n
since the series
∞
X 1
n=1
n ln n

diverges.

Theorem 9.2.11. If bn & 0, and there exists C > 0 such that |nbn | 6 C for any n ∈ N. Then there exists
a a constant M > 0 such that
Xm
bn sin nx 6 M ∀m ∈ N.
n=1

Corollary 9.2.12. There exists a constant C > 0 such that for any m ∈ N and any x ∈ R,
m
X sin nx
6 C.
n=1
n

Theorem 9.2.13. If an & 0 and

an − 2an+1 + an+2 > 0 ∀n ∈ N ∪ {0},

then the series

∞
a0 X
+ an cos nx
2 n=1

converges at any point x ∈ R, except possibly the points 2πm, m ∈ Z, to a nonnegative 2π-periodic even
function f ∈ L1 [−π, π]. Moreover, this series is the Fourier series of f .

Corollary 9.2.14. The series

∞
X cos nx
.
n=2
ln n

is the Fourier series of a function f ∈ L1 [−π, π].

196 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

9.3 Applications of Fourier series to ODE

I covered pages 17-22 of the book [10].

9.3.1 Differentiation of Fourier series

Theorem 9.3.1. Let a function f be 2π-periodic and integrable on [−π, π] (so on R). Suppose that its
Fourier coefficients fb(n) have the following estimate
+∞
σn X
fb(n) = k , where |σn |2 < ∞,
n n=−∞

then f ∈ C (k−1) [−π, π].

Proof. If we formally differentiate the Fourier series of the function f k − 1 times, then the Fourier
coefficients of its formal derivative have form
σn
(in)k−1 fb(n) = ik−1 ,
n
it converges, since
σn |σn |2 1
|(in)k−1 fb(n)| 6 6 + .
n 2 2|n|2
This justifies the differentiation, so f is k − 1 times continuously differentiable.

9.3.2 Applications of Fourier series to finding periodic solutions of ODE

Let us consider a linear ordinary differential equation

p0 y (n) (x) + p1 y (n−1) (x) + · · · + pn y(x) = q(x) (9.3.1)

with constant coefficients and a 2π-periodic function q. The problem is to find out whether there exists a
periodic solution of the equation (9.3.1).
We will search the solution in the Fourier series form
+∞
X
y(x) = yk eikx .
k=−∞

Let us expand the right hand side of (9.3.1) in to the Fourier series:
+∞
X
q(x) = qk eikx .
k=−∞

then for all k ∈ Z one has

p0 (ik)n yk + p1 (ik)n−1 yk + · · · + pn yk = qk
(this is the equality of k th Fourier coefficients of the left hand side and the right hand side of (9.3.1)).
Putting
P (z) = p0 z n + p1 z n−1 + · · · + pn ,
we obtain
P (ik)yk = qk (k ∈ Z).
If
P (ik) 6= 0 (∀k ∈ Z), (9.3.2)
9.3. APPLICATIONS OF FOURIER SERIES TO ODE 197

then
qk
yk = ,
P (ik)
that is, we found the Fourier coefficients of the expected periodic solution y(x) of the equation (9.3.1),
and therefore, the function itself is also found. But in order to the function y(x) to be a periodic solution
of (9.3.1) we must justify the possibility of n-times differentiating. To do this we require additionally that
q(x) is continuously differentiable.
From the asymptotics P (ik) ∼ p0 (ik)n for k → ∞ we obtain

|P (ik)| > c|k|n

for some c > 0, so

|σk |
|yk | 6 ,
c|k|n+1
where σk are the Fourier coefficients of the function q 0 (x). Since q 0 is continuous by assumption, the series

+∞
X
|σk |2
k=−∞

converges, so we can apply Theorem 9.3.1 to infer that the sum of the series
+∞
X
yk eikx
k=−∞

converges uniformly to a n-times continuously differentiable function y(x). So y(x) is a 2π-periodic solution
of (9.3.1).
Moreover, this solution is unique. In fact, if there exists on more 2π-periodic solution of (9.3.1), say,
ye(x), then the difference f (x) = y(x) − ye(x) is also 2π-periodic and satisfies the equation

p0 f (n) (x) + p1 f (n−1) (x) + · · · + pn f (x) = 0. (9.3.3)

The general solution of (9.3.3) can be expressed via the exponentials eµx , where µ are the roots of the
characteristic equation
P (z) = 0,

so the equation (9.3.3) has a periodic solution (except trivial f (x) ≡ 0) only if µ = ik for some k ∈ Z, that
is, P (ik) = 0 for some k. This contradicts with the assumption (9.3.2).
Thus we get that for 2π-periodic q 0 ∈ C 1 [−π, π] the condition (9.3.2) is the condition of the existence
and uniqueness of 2π-periodic solution of the equation (9.3.1).
Note that if for some integer m
P (im) = 0,

and qm = 0, then the equation (9.3.3) has a 2π-periodic solution

f (x) = Aeimx ,

where A is an arbitrary constant. So in this case, the equation (9.3.1) has infinitely many periodic solutions.
Finally, if
P (im) = 0,

but qm 6= 0, then the equation (9.3.3) has no periodic solutions.

198 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

9.3.3 Boundary problems of the theory of ODE

The main problem of the theory of ODE is the Cauchy problem (the initial value problem). But in
applications to physics, more important problems are the so called boundary value problems. We consider
now the boundary problems for ODE of the second order. In a sufficiently general form such a problem
can be posed as follows.
On an interval [a, b] we want find solutions y = y(x) of the differential equation

p2 (x)y 00 + p1 (x)y 0 + p0 (x)y = f (x), (9.3.4)

satisfying the boundary conditions (

α0 y(a) + α1 y 0 (a) = c1 ,
(9.3.5)
β0 y(b) + β1 y 0 (b) = c2 .
We assume that the functions p0 , p1 , p2 , and f are continuous on [a, b]. If c1 = c2 = 0, then the
boundary conditions are called uniform. Here we consider this case. This restriction (c1 = c2 = 0) allows
us to treat the set of all continuously differentiable functions on [a, b] satisfying the uniform boundary
conditions as a linear space. We denote this space as V1 . That is,

α0 y(a) + α1 y 0 (a) = 0,
 
1
V1 = y(x) : y ∈ C [a, b] and
β0 y(b) + β1 y 0 (b) = 0.

We also denote by V2 the subspace of V1 consisting of two-times continuously differentiable functions.

Then the boundary value problem tales the following form: to find y ∈ V2 such that

L(y) = f, (9.3.6)

where L is a linear differential operator defined on V2 as follows

L(y) = p2 (x)y 00 + p1 (x)y 0 + p0 (x)y.

The natural question is: for what kind of f the problem (9.3.6) has a solution? Is this solution unique?
To answer this question, we need to study the spectral properties of the operator L. This means that we
need to study the solvability of the eigenvalue problem of L, that is, to find all the pairs (L, y), where
λ ∈ C and y ∈ V2 , y 6= 0 such that
L(y) = λy.
We will see that the eigenvalue problem for the operators of boundary value problems is a source of
many various orthonormal (in a certain sense) systems, the systems of eigenfunctions. If the eigenvalue
problem for L is solved, and the complete orthonormal (in a certain sense) system of eigenfunctions ϕ1 ,
ϕ1 , . . . is found, moreover, if there are no zero eigenvalues λn , then the solution of the problem (9.3.6)
can be solved elementary. Indeed, let us expand the function f into the Fourier series with respect to the
orthonormal system {ϕn }∞ n=0 :
+∞
X
f= fb(n)ϕn ,
n=1

where fb(n) are the Fourier coefficients of the function f with respect to the
We will search the solution y of the boundary value problem in the form of Fourier series
+∞
X
y= yb(n)ϕn .
n=1

Suppose that the rate of convergence of this series allows us to differentiate it sufficiently many times:
+∞
! +∞
X X
L yb(n)ϕn = yb(n)L(ϕn ),
n=1 n=1
9.3. APPLICATIONS OF FOURIER SERIES TO ODE 199

so the equation (9.3.6) has the form

λn yb(n) = fb(n), n = 1, 2, . . .

since L(ϕn ) = λn ϕn . Thus, we have

+∞ b
X f (n)
y= ϕn .
n=1
λn

The question of the convergence of this series should be considered separately.

To study the boundary value problem, it is more convenient to rewrite the differential equation for
eigenvalues in a symmetric form. Namely, we multiply the equation

p2 y 00 + p1 y 0 + p0 y = λy

by a function ρ such that this equation turns into the following

−(py 0 )0 + qy = λρy.

Obviously, the function ρ can be found from the equation

p1 ρ = (p2 ρ)0 ,

where p = −p2 ρ. Thus, if p2 6≡ 0, then

R p1
p2 dx ,
p
p = Ce ρ=− .
p2

The operator L defined formally by the equality

−(py 0 )0 + qy
L(y) = (9.3.7)
ρ

is called the Sturm-Liouville operator. The boundary value problem for eigenvalues and eigenfunctions
(λ, y) of the operator L
− (py 0 )0 + qy = λρy (9.3.8)
(
α0 y(a) + α1 y 0 (a) = 0,
(9.3.9)
β0 y(b) + β1 y 0 (b) = 0.
is called the Sturm-liouville problem. Here we assume that the functions p, q, and ρ are real continuous
functions. Moreover, p is continuously differentiable, while q and ρ are nonnegative. The coefficients α1 ,
α2 , β1 , β2 are assumed to be real and such that

α12 + α22 6= 0, (9.3.10)

β12 + β22 6= 0.
Note that the boundary conditions of the form

y(a) = 0, y(b) = 0

are called the Dirichlet conditions, while the boundary conditions of the form

y 0 (a) = 0, y 0 (b) = 0

are called the Neumann conditions.

200 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

The differential equation of the Sturm-Liouville problem can be further transformed. Indeed, if we
introduce the new variable t by the equality
Z
dx
t= ,
p(x)

where we assume that p 6= 0, then from the identity

d dx d d
= · =p ,
dt dt dx dx
we obtain ta new form of our equation:

d2
− y + pqy = λpρy.
dx2
Let now Z
dt
y = k(t)u(s), s= ,
k2
where the function k is not defined yet, while s and u are the new variable and new sought-for function
(instead of y), respectively. At the same time, we have

dy dk du ds dk 1 du
= u+k · = u+ · ,
dt dt ds dt dt k ds
dy 2 d2 k dk du ds 1 dk du 1 d2 u ds d2 k 1 d2 u
2
= 2u+ · · − 2· · + · 2 · = 2u+ 3 · 2,
dt dt dt ds dt k dt ds k ds dt dt k ds
that implies
1 d2 u d2 k
 
− 3 · 2 + pqk − 2 = λpρku.
k ds dt
Now we choose the function k such that
dk 2
pρk 4 = 1, r = pqk 4 − k 3
dt2
to obtain the following new form of the considered equation:

d2 u
− + ru = λu.
ds2
The transformations described above preserve the form of the uniform boundary conditions (with some
new coefficients).

9.3.4 Regular Sturm-Liouville problem

The operator L defined in (9.3.7) is called regular Sturm-Liouville operator if p, ρ > 0 on [a, b]. the
eigenvalue boundary problem (9.3.8)–(9.3.9) is called regular if L is regular.
Let now specify some properties of the regular Sturm-Liouville problem.

Proposition 9.3.2. All the zeros of the eigenfunctions are simple (of multiplicity one).

Proof. Indeed, if y(x0 ) = 0 and y 0 (x0 ) = 0 for some x0 ∈ [a, b], then y(x) ≡ 0 on [a, b], so y cannot be an
eigenfunction of L by definition.

Proposition 9.3.3. To every eigenvalue of L there corresponds a unique (up to a constant factor) eigen-
function, that is, all the eigenvalues of the regular Sturm-Liouville operator are simple.
9.3. APPLICATIONS OF FOURIER SERIES TO ODE 201

Proof. Let y1 and y2 be two eigenfunctions corresponding to an eigenvalue µ of the operator L. Since the
uniform system (with respect to variables α0 and α1 )
(
α0 y1 (a) + α1 y10 (a) = 0,
(9.3.11)
α0 y2 (a) + α1 y20 (a) = 0

must have a nontrivial solution by (9.3.10). Consequently, the determinant of this system must vanish.
Note that the determinant of this system is the Wronskian W [y1 , y2 ] of the functions y1 and y2 calculated
at the point a. The Wronskian of the functions y1 and y2 has the form

y1 (x) y10 (x)

W [y1 , y2 ](x) = = y1 (x)y20 (x) − y2 (x)y10 (x).
y2 (x) y20 (x)

So the system (9.3.11) has a nontrivial solution if, and only if, W [y1 , y2 ](a) = 0.
On the other hand, by (9.3.8) we have that

d
{p(x)W [y1 , y2 ](x)} = y1 (x)(p(x)y20 (x))0 − y2 (x)(p(x)y10 (x))0 =
dx

= y2 (x)[λρ(x) − q(x)]y1 (x) − y1 (x)[λρ(x) − q(x)]y2 (x) = 0,

so the function ρ(x)W [y1 , y2 ](x) is constant, so W [y1 , y2 ](x) ≡ 0, since W [y1 , y2 ](a) = 0. Consequently,
y1 (x) = cy2 (x), where c is a constant, as required.

Proposition 9.3.4. All the eigenvalues of the Sturm-Liouville problem are real. The corresponding eigen-
functions can be chosen real.

Proof. Note that by integration by part one can obtain

Z b b
Z b
[qf − (pf 0 )0 ]gdx = pW [f, g] + f [qg − (pg 0 )0 ]dx. (9.3.12)
a a a

If the functions f and g satisfy the boundary conditions (9.3.9) (that is, a, b ∈ V2 ), then W [f, g](a) = 0
and W [f, g](b) = 03 . Let us introduce the following weighted inner product
Z b p
hf, gi = f gρdx, kf k = hf, f i. (9.3.13)
a

Then the identity (9.3.12) takes the form

hL[f ], gi = hf, L[g]i.

Such a property is called the symmetry of the operator L. If now y is an eigenfunction of L corresponding
to an eigenvalue λ, then
λkyk2 = hL[y], yi = hy, L[y]i = λkyk2 .
Since kyk =
6 0 (by definition of eigenfunctions), we obtain λ = λ, so λ ∈ R, as required.

Note that since the functions p, q and ρ are real, and since the equation L[y] = λy is real and linear,
the eigenfunctions of the operator L can be chosen real (up to a constant complex factor). So in what
follows we assume that the eigenfunctions of the operator L are real.

Proposition 9.3.5. If λ1 = 6 λ2 , then the corresponding eigenfunctions y1 and y2 are orthogonal with
respect to the inner product (9.3.13).
3 Here we use the fact that the boundary conditions have real coefficients.
202 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

Proof. Indeed,
(λ1 − λ2 )hy1 , y2 i = hL[y1 ], y2 i − hy1 , L[y2 ]i = 0.
So,
hy1 , y2 i = 0,
since λ1 − λ2 6= 0 by assumption.
Proposition 9.3.6. The sequence of eigenvalues of the operator L is infinite and monotone increasing

λ1 < λ2 · · · λn < · · · , λn −−−−→ ∞.

n→∞

We will not prove this proposition.

Consider the space V1 with the inner product (9.3.13). Suppose that the eigenfunctions yn of the
Sturm-Liouville problem are normed:
Z b
2
kyn k = |yn (x)|2 ρ(x)dx = 1.
a

Then they generate an orthonormal system. The Fourier coefficients of a function f (from the space V1
with inner product (9.3.13)) with respect to this orthonormal system are defined by the following formula4
Z b
fb(n) = hf, yn i = f (x)yn (x)ρ(x)dx.
a

The expansion of the function f into the Fourier series with the eigenfunctions of the Sturm-Liouville
problem has the form
+∞
X
f∼ fb(n)yn (x). (9.3.14)
n=1

Let us establish that the system of the eigenfunctions of the Sturm-Liouville problem is closed (or
complete), so the series (9.3.14) converges to the function f in the norm of the space V1 with inner
product (9.3.13):
XN
kf − fb(n)yn (x)k −−−−→ 0. (9.3.15)
N →∞
n=1

We will give a scheme of the proof of (9.3.15) under additional conditions:

1. ρ = 1;
2. V1 consists of real continuously differentiable functions satisfying Dirichlet’s boundary conditions:

y(a) = y(b) = 0.

In this case, the inner product (9.3.13) has the form

Z b
hf, gi = f gdx. (9.3.16)
a

Let Z b Z b
I(y) ≡ hL(y), yi = [−(py 0 )0 + qy]ydx = [p(y 0 )2 + qy 2 ]dx. (9.3.17)
a a

It can be proved (we skip this proof) that the minimal value of the functional I(y) under conditions

y(a) = y(b) = 0, kyk = 1,

4 Recall that the eigenfunctions yn are supposed to be real.
9.3. APPLICATIONS OF FOURIER SERIES TO ODE 203

equals the following

min I(y) = λ1 , λ1 = I(y1 ),
where λ1 is the minimal eigenvalue of the considered Sturm-Liouville problem, and y1 is the corresponding
eigenfunction, kyk = 1.
Moreover, the following variational principle holds.
Theorem 9.3.7. Let y1 , y2 , . . . , yn−1 be the orthonormal system of eigenfunctions of the Sturm-Liouville
problem satisfying the eigenvalues λ1 , λ2 , . . . , λn−1 such that

λ1 , λ2 , . . . , λn−1 .

Then the minimal value of the functional I(y) defined in (9.3.17) under the conditions

y(a) = y(b) = 0, kyk = 1, y⊥yk , ∀k = 1, 2, . . . , n − 1,

can be achieved. Furthermore,

min I(y) = λn , λn = I(yn ),
where λn is the n-th eigenvalue of the Sturm-Liouville problem, and yn is the corresponding eigenfunction.
We will use this theorem to prove the completeness of the system of eigenfunctions of the Sturm-
Liouville operator L in the space V1 .
Let
n−1
X
rn = f − ck y k ,
k=1

where ck = hf, yk i are the Fourier coefficients of the function f . So yk ⊥rn for k < n. Therefore,
n−1
X n−1
X
hL(rn ), rn i = hL(f ), rn i − ck hL(yk ), rn i = hL(f ), rn i − ck λk hyk , rn i,
k=1 k=1

and since hyk , rn i = 0, we have

hL(rn ), rn i = hL(f ), rn i.
Suppose that krn k =
6 0, so
rn = krn ken ,
where
ken k = 1, en ⊥y1 , y2 , . . . , yn−1 .
By Theorem 9.3.7, we have

hL(rn ), rn i = krn k2 hL(en ), en i > krn k2 min hL(y), yi = λn krn k2 .

kyk=1
y⊥y1 ,...,yn−1

This implies, due to the Cauchy-Bunyakovsky-Schwarz inequality, that

λn krn k2 6 hL(f ), rn i 6 kL(f )k · krn k,

so since λn → +∞ as n → ∞, we obtain
kL(f )k
λn krn k 6 −−−−→ 0.
λn n→∞

But this is equivalent to the closedness (completeness) of the system of the eigenvalues yk of the Sturm-
Liouville operator of the Dirichlet boundary problem with ρ = 1.
Let us return to the equation
L(y) = f,
204 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

and note that since L is symmetric, we have

hL(y), yn i = hy, L(yn )i = hy, λn yn i = λn hy, yn i,

so
λn hy, yn i = hf, yn i.
The question whether the function
+∞
X hf, yn i
y= yn
n=1
λn
is, in fact, the solution of the Sturm-Liouville problem depends on the rate of convergence of this series.
9.4. PROBLEMS 205

9.4 Problems
9.4.1 Basic theory of Fourier series
Problem 9.1. Prove the formula
 
1
n sin n + x
X 1 1 2
cos mx = − + · x (9.1.1)
m=1
2 2 sin
2

eiy + e−iy eiy − e−iy

Hint: Use the formulæ cos y = , sin y = . An alternative proof is to exploit the
2 2
trigonometric formula of the product of a sine and cosine.
Problem 9.2. Applying integration by parts, deduce from (9.1.1) that
   
1 1 t
n cos n + x π cos n + t · cos
π−x
Z
X sin mx 2 1 2 2
= − x + 2n + 1 dt (9.2.2)
m 2 t
m=1 (2n + 1) sin x sin2
2 2
Problem 9.3. Prove that
∞
X sin nx π−x
=
n=1
n 2
for any x ∈ (0, π).
π+x
Analogously, prove that for any x ∈ (−π, 0), the series converges to − .
2
Problem 9.4. Prove the formula
n
X 1 sin 2nx
cos(2m − 1)x = · .
m=1
2 sin x

Using this formula prove that for any x ∈ (0, π)

∞
X sin(2n − 1)x π
= . (9.4.3)
m=1
2n − 1 4

Problem 9.5. Using the identity (9.4.3), prove Steklov’s formula

+∞ Z
4 X x+π
f (x + π) − f (x) = − f (u) cos(2m + 1)(u − x)du,
π m=0 x

where f (u) is differentiable on the interval (x, x + π).

Problem 9.6. Expand into Fourier series the following periodic functions
1) {x}, the fractional part of x, {x} = x − [x], where [x] is the integer part of x;
2) (x), the distance from x to the closest integer;
Problem 9.7. Find the Fourier series of the a 3-periodic function defined on [0, 3] by the following formulæ


 x if x ∈ [0, 1],
1 if x ∈ [1, 2],

3 − x if x ∈ [2, 3],


and study its convergence.

206 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

Problem 9.8. Verify that

+∞
1 X einx
2i n=−∞ n
n6=0

is the Fourier series of the 2π-periodic sawtooth function defined by f (0) = 0 and
 π x
− 2 − 2


if − π < x < 0,
f (x) =
 π−x

if 0 < x < π.

2 2
Note that this function is not continuous. Show that nevertheless, the series converges for every x (by
which we mean, as usual, that the symmetric partial sums of the series converge). In particular, the value
of the series at the origin, namely 0, is the average of the values of f (x) as x approaches the origin from
the left and the right.
P
Hint: Use Dirichlet’s test for convergence of a series an bn .
n

Problem 9.9. Expand f (x) = x3 into the cos-series in the interval (0, π).

Problem 9.10. Expand f (x) = x3 into the sin-series in the interval (0, π).

Problem 9.11. Expand f (x) = ex into Fourier series in the interval (−h, h).

Problem 9.12. Expand f (x) = cosh(ax) into Fourier series in the interval (−π, π).

Problem 9.13. Expand f (x) = sinh(ax) into Fourier series in the interval (−π, π).

Problem 9.14. Using the change of variables t = eix prove that

1)
+∞
a sin x X
= an sin(nx), |a| < 1,
1 − 2a cos x + a2 n=1

2)
+∞ n
X a
ln(1 − 2a cos x + a2 ) = −2 cos(nx), |a| < 1,
n=1
n

3)
+∞
x X (−1)n−1 cos(nx)
ln 2 cos = ,
2 n=1 n

where x ∈ (−π, π).

Hint: For 2) use differentiation/integration, for 3) use the series 2).

Problem 9.15. Is it possible to find Fourier series of the 2π-periodic function

x
tan
2
on the interval (−π, π)? Explain the answer.

Problem 9.16. Prove that

9.4. PROBLEMS 207

1)
cos 2x cos 4x cos 6x 1 π
+ + + · · · = − sin x, 0 < x < π,
1·3 3·5 5·7 2 4
2)
+∞
X sin(2n + 1)x πx
3
= (π − x), 0 < x < π.
n=1
(2n + 1) 8

3)
cos 3x cos 5x cos 7x π 1 π π
− + − · · · = cos2 x − cos x, − <x< .
1·3·5 3·5·7 5·7·9 8 3 2 2
Problem 9.17. Prove that
+∞
X sin(2n + 2)x
= sin 2x − (π − 2x) sin2 x − sin x cos x ln(4 sin2 x), 0 < x < π.
n=1
n(n + 1)

Problem 9.18. Prove that

+∞  2 
X cos(2n + 1)x π 2
= 2(π − 2x) sin x + − 2πx + 2x − 3 cos x, 0 < x < π.
n=1
n2 (n + 1)2 3
Z 1
1
Problem 9.19. Noting that = tn−1 dt find the sum of the series
n 0

+∞
X sin nx
, 0 < x < π.
n=1
n

Problem 9.20. Find the sum of the following series

1)
+∞
X cos(nx)
, 0 < x < 2π,
n=1
n2

2)
+∞
X sin(nx)
, 0 < x < 2π,
n=1
n3

3)
+∞
X cos nx
, 0 < x < 2π,
n=1
n

4)
+∞
X cos(2n + 1)x
, 0 < x < π,
n=0
2n + 1

5)
+∞
X cos(nx)
, 0 < x < 2π,
n=2
n2 − 1

6)
+∞
X sin(nx)
, 0 < x < 2π,
n=2
n2 − 1
208 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

7)
+∞
X sin nx
, 0 < x < π,
n=1
n(n2 + a2 )

8)
+∞
X (−1)n sin nx
, 0 < x < π,
n=1
n(n2 + a2 )

9)
+∞
X sin(2n + 1)x
, 0 < x < π,
n=0
2n + 1

10)
+∞
X (−1)n cos nx
, −π < x < π,
n=0
n2 − 4

Hint: For 1) and 2) use the result of Problem 9.19 and differentiation.

Problem 9.21. Show that for α not an integer, the Fourier series of
π
ei(π−x) α
sin πα
on [0, 2π] is given by
+∞
X einx
n=−∞
n+α

Apply Parseval’s identity to show that

+∞
X 1 π2
= .
n=−∞
(n + α)2 (sin πα)2

Problem 9.22. Suppose f is a periodic function of period 2π which belongs to the class C k . Show that
cn einx is the Fourier series of f (x), then
P
if
n
 
1
cn = o as |n| → ∞.
|n|k

Hint: Use Riemann-Lebesgue lemma.

Problem 9.23. Let f ∈ C[−π, π]. Prove that

Zπ π  Zπ π 
f (x) cos nxdx 6 Cω ,f , f (x) sin nxdx 6 Cω ,f
n n
−π −π

where C > 0 is a constant and

ω (δ, f ) = sup |f (x) − f (y)|, x, y ∈ [−π, π].

|x−y|<δ
9.4. PROBLEMS 209

Problem 9.24. Let f is continuous and 2π-periodic. Prove that

Zπ
1 4n (n!)2  u 2n
fn (x) = · f (x + u) cos du
2π (2n)! 2
−π

is a singular integral, that is, fn (x) tends to f (x) uniformly on R. Show also that fn (x) is a trigonometric
polynomial.

Problem 9.25. Improve the rate of convergence of the following series

1)
+∞ 2
X n sin nx
n=1
n3 + 1

from order n−1 to order n−4 using the result of Problem 6,

2)
+∞ 2
X n +1
4+1
cos nx
n=1
n

from order n−2 to order n−4 using Problem 6.

Problem 9.26. Find the sum of the series

+∞
X cos nx
, 0 < x < 2π.
n=2
n2 − 1

and use it to improve the rate of convergence of the series

+∞ 2
X n +1
cos nx
n=1
n4 + 1

from order n−2 to order n−6 .

Problem 9.27. Let f ∈ L1 [a, b]. Prove that

Zb
lim f (x)eiRx dx = 0.
R→∞
a

9.4.2 Applications of Fourier series

Problem 9.28. Use the method of eigenfunctions expansion to solve the following boundary value prob-
lem: (
y 00 + 4y = x3
y(0) = 0, y 0 (1) = 0.

Problem 9.29. Find the eigenvalues and eigenfunctions of the following boundary value problem:

 (xy 0 )0 + 2y = −λ y
x x
 y 0 (1) = 0, y 0 (2) = 0.
210 CHAPTER 9. TRIGONOMETRIC FOURIER SERIES

Problem 9.30. Use the method of eigenfunctions expansion to solve the following boundary value prob-
lem: 
 (xy 0 )0 + y = 1
x x
 y(1) = 0, y(e) = 0.

Problem 9.31. The eigenvalue problem x2 y 00 −λxy 0 +λy = 0 with y(1) = y(2) = 0 is not a Sturm-Liouville
eigenvalue problem. Show that none of the eigenvalues are real by solving this eigenvalue problem.
Problem 9.32. Find 2π-periodic solutions of the following ordinary differential equation:

y 000 + y = sin 2x

Problem 9.33. Find the solution (tending to zero as |x| → ∞) of the following difference equations:
1
f (x + h) − 2f (x) = ,
1 + x2

Hint: Represent the solution as a functional (not power or Fourier!) series depending on parameter h.
1
When h = 0, the series must degenerate to the function f (x) = − .
1 + x2
Problem 9.34. Find 2π-periodic solutions of the following difference equation:

f (x + h) + 3f (x) + f (x − h) = g(x), h 6= 2π,

where g(x) is the 2π-periodic function defined by the equality

g(x) = x, for x ∈ (0, 2π).

Hint: Compare the Fourier coefficients of the left and right hand sides of the equation.
Problem 9.35. Find the eigenvalues and eigenfunctions of the following problem
(
y 00 + y = λy,
y(1) = y 0 (l) = 0.

and solve the problem (

y 00 + y = x2 ,
y(1) = y 0 (l) = 0.

Problem 9.36. Find the eigenvalues and eigenfunctions of the following problem

 x2 y 00 + 1 y = λy,
4
 y(1) = y 0 (e) = 0.

and solve the problem 

1
 x2 y 00 + y = x,
4
 y(1) = y 0 (e) = 0.
Bibliography

[1] R. Bass, Real analysis for graduate students: measure and integration theory, 2011.
[2] P. Cohen, Set theory and the continuum hypothesis, Dover Publications, Mineola, New York, 1966.
[3] G.B. Folland, Real Analysis, 2nd ed., Wiley, New York, 1999.

[4] J.K. Hunter, Measure theory, Lecture notes: https://www.math.ucdavis.edu/~hunter/measure_

theory/measure_theory.html.
[5] A. Kolmogorov and S. Fomin, Introduction to Real Analysis, Dover Publication Inc., New York, 1970.
[6] A. Kolmogorov and S. Fomin, Measure, Lebesgue integrals, and Hilbert space, Academic Press, New
York and London, 1960.
[7] T. Körner, Excersices for Fourier Analysis, Cambridge University Press, 1993.
[8] I. Natanson, Theory of Functions of a Real Variable, Vol. I and Vol. II, Frederick Ungar Publishing
Co, New York, 1955, 1961, 1964.

[9] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill Inc., 1976.

[10] E. Stein and R. Shakarchi, Fourier Analysis, Princeton University Press, Princeton and London, 2003.
[11] E. Stein and R. Shakarchi, Real Analysis, Princeton University Press, Princeton and London, 2005.

211