Download as pdf or txt
Download as pdf or txt
You are on page 1of 70

Real Analysis

Yongmin Park

November 25, 2021

Contents

1 Measures 1
1.1 Measure spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Construction of measures by outer measures . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Lebesgue measure and Lebesgue-Stieltjes measures . . . . . . . . . . . . . . . . . . . . 8
1.4 Properties of Lebesgue measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Integration 13
2.1 Measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Limit theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Complex-valued functions and image measures . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Modes of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Lebesgue spaces 27
3.1 𝐿 𝑝 spaces: definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 𝐿 𝑝 spaces: properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Signed and complex measures 37


4.1 Hahn and Jordan decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Banach space of complex measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Radon-Nikodym theorem and Lebesgue decomposition . . . . . . . . . . . . . . . . . . 41

5 Product measures 47
5.1 Product measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Fubini’s theorem and applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6 Borel measures on locally compact Hausdorff spaces 53


6.1 Locally compact Hausdorff spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Riesz representation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3 Regularity properties of Radon measures . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.4 Lebesgue measure again . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.5 Lusin’s theorem and Vitali-Carathéodory theorem . . . . . . . . . . . . . . . . . . . . . 63

References 68

i
1 Measures

1.1 Measure spaces


We begin with discussing what are measurable sets.

Definition 1.1. Let 𝑋 be a set. A family A of subsets of 𝑋 is called an 𝜎-algebra on 𝑋 if the following
holds:

1. 𝑋 ∈ A.

2. A is closed under complementation.

3. A is closed under countable union.

4. A is closed under countable intersection.

If 𝐴 ∈ A, then 𝐴 is said to be A-measurable.

Remark 1.2.

1. If the word ’countable’ is replaced by ’finite’, then we get the definition of algebra on 𝑋.

2. By De Morgan’s law, if a family A is closed under complementation, then the following hold:

a) 𝑋 ∈ A if and only if ∅ ∈ A
b) A is closed under countable union if and only if A is closed under countable intersection.

3. For an arbitrary family of 𝜎-algebras on 𝑋, its intersection is also a 𝜎-algebra on 𝑋. Thus, for a
family F of subsets of 𝑋, we define the smallest 𝜎-algebra 𝜎(F) on 𝑋 containing F by
Ù
𝜎(F) B {A ⊆ P (𝑋) | A is a 𝜎-algebra on 𝑋 containing F }.

We call 𝜎(F) the 𝜎-algebra generated by F.

Example 1.3.

1. For a set 𝑋, the power set P (𝑋) is a 𝜎-algebra.

2. For a topological space 𝑋, we define the Borel 𝜎-algebra B(𝑋) on 𝑋 by the 𝜎-algebra generated
by the collection of all open subsets of 𝑋. In fact, B(𝑋) is the same with the 𝜎-algebra generated
by the collection of all closed subsets of 𝑋. Moreover, if 𝑋 is second-countable, then every base
of the topology of 𝑋 generates the Borel 𝜎-algebra on 𝑋.

We now discuss several ways to construct of a given family of subsets of 𝑋.

Definition 1.4. A collection C of subsets of 𝑋 is called a monotone class on 𝑋 if it is closed under


monotone limits.

Similarly to 𝜎-algebras, an arbitrary intersection of monotone classes on 𝑋 is also a monotone class


of 𝑋. Thus, for a family F of subsets of 𝑋, there is the smallest monotone class 𝑚(F) containing F.
Remark 1.5. Let A be a family of subsets of 𝑋 closed under complementation. Then, by De Morgan’s
law, A is closed under monotone increasing limits if and only if A is closed under monotone decreasing
limits.
Moreover, if A is an algebra on 𝑋, then A is a monotone class if and only if A is an 𝜎-algebra.

Theorem 1.6 (Monotone class theorem). Let A be an algebra on 𝑋. Then, 𝑚(A) = 𝜎(A).

1
Proof. It is clear that 𝑚(A) ⊆ A since 𝜎(A) is a monotone class containing A. To show the reverse
containment, it suffices to show that 𝑚(A) is an 𝜎-algebra. By the previous remark, we claim that 𝑚(A)
is an algebra.
We first show that 𝑚(A) is closed under complementation. Define

C B {𝐴 ∈ 𝑚(A) | 𝐴𝑐 ∈ 𝑚(A)}.

Then, since A is an algebra and A ⊆ 𝑚(A), A ⊆ C. Thus, if we show that C is a monotone class, then
𝑚(A) ⊆ C. Indeed, one can easily show that C is a monotone class by using De Morgan’s law.
Now we show that 𝑚(A) is closed under finite intersection. For each 𝐴 ∈ 𝑚(A), define

C 𝐴 B {𝐵 ∈ 𝑚(A) | 𝐴 ∩ 𝐵 ∈ 𝑚(A)}.

Then, by the distribution law for intersection and union, C 𝐴 is a monotone class. Observer that if 𝐴 ∈ A,
then since A is an algebra, A ⊆ C 𝐴. This shows that for each pair of 𝐴 ∈ A and 𝐵 ∈ 𝑚(A), 𝐴∩ 𝐵 ∈ 𝑚(A).
This means that for every 𝐵 ∈ 𝑚(A), A ⊆ C 𝐵 . This completes the proof. 

Definition 1.7. A collection D of subsets of 𝑋 is called a 𝜆-system (or a Dynkin class) on 𝑋 if

1. 𝑋 ∈ D

2. D is closed under complementation

3. D is closed under countable disjoint union.

Moreover, a collection of subsets of 𝑋 closed under finite intersection is called a 𝜋-system on 𝑋.

Remark 1.8.

1. If a 𝜆-system D is a 𝜋-system, then D is an algebra, and consequently, D is a 𝜎-algebra. In other


words, a 𝜆-system D is a 𝜎-algebra if and only if D is a 𝜋-system.

2. A 𝜆-system is a monotone class. To show this, let D be a 𝜆-system on 𝑋. Then, for 𝐴, 𝐵 ∈ D with
𝐴 ⊆ 𝐵, 𝐵 \ 𝐴 = (𝐵 𝑐 ∪ 𝐴) 𝑐 ∈ D. Thus, if ( 𝐴𝑛 ) 𝑛=1,2,... is an increasing sequence in D, then set
𝐵1 B 𝐴1 , and for each 𝑛 > 1,
𝐵𝑛 B 𝐴𝑛 \ 𝐴𝑛−1 ∈ D.
Note that (𝐵𝑛 ) is a disjoint collection, and

Ø ∞
Ø
𝐴𝑛 = 𝐵𝑛 ∈ D.
𝑛=1 𝑛=1

This shows that D is a monotone class. In fact, a collection D of subsets of 𝑋 is a 𝜆-system if and
only if the following hold.

a) 𝑋 ∈ D
b) For each pair of 𝐴, 𝐵 ∈ D with 𝐴 ⊆ 𝐵, 𝐵 \ 𝐴 ∈ D
c) D is closed under monotone increasing limits.
Similarly to 𝜎-algebras and monotone classes, an arbitrary intersection of 𝜆-systems is also a 𝜆-
system. Therefore, for a given family F of subsets of 𝑋, we can define the 𝜆-system D(F) generated by
F.

Theorem 1.9 (Dynkin’s 𝜋-𝜆 theorem). Let C be a 𝜋-system on 𝑋. Then, D(C) = 𝜎(C).

2
Proof. The proof is similar to the proof of the monotone class theorem. Note that 𝜎(C) is a 𝜆-system
containing C, so D(C) ⊆ 𝜎(C). To show the reverse containment, it suffices to show that D(C) is a
𝜋-system: if this is true, then D(C) is a 𝜎-algebra containing C, so it completes the proof.
For each 𝐴 ∈ D(C), define

F 𝐴 B {𝐵 ∈ D(C) | 𝐴 ∩ 𝐵 ∈ D(C)}.

We now show that F 𝐴 is a 𝜆-system. First, 𝑋 ∈ A since 𝐴 ∩ 𝑋 = 𝐴 ∈ D(C). Moreover, by the distribution
law, F 𝐴 is closed under monotone increasing limits. Now, if 𝐵 ∈ F 𝐴, then

𝐴 ∩ 𝐵 𝑐 = 𝐴 \ ( 𝐴 ∩ 𝐵) ∈ D(C)

since 𝐴 ∩ 𝐵 ⊆ 𝐴 and 𝐴, 𝐴 ∩ 𝐵 ∈ D(C). Thus, F 𝐴 is a 𝜆-system.


Now, claim that for every 𝐴 ∈ D(C), C ⊆ F 𝐴. If 𝐴 ∈ C, then C ⊆ F 𝐴 since C is a 𝜋-system. Therefore,
in this case, D(C) ⊆ F 𝐴. This means that for each pair of 𝐴 ∈ C and 𝐵 ∈ D(C), 𝐴 ∩ 𝐵 ∈ D(C). Hence,
for all 𝐵 ∈ D(C), C ⊆ F 𝐵 . This completes the proof. 

Example 1.10. Let C be the family of 𝑑-dimensional rectangles (including the empty set) in R𝑑 . Here,
a 𝑑-dimensional rectangle is the Cartesian product of 𝑑 intervals in R which have the form

(𝑎, 𝑏), (𝑎, 𝑏], [𝑎, 𝑏), or [𝑎, 𝑏].

Then, clearly, C is a 𝜋-system, and 𝜎(C) is the Borel 𝜎-algebra B(R𝑑 ). Thus, by generating the 𝜆-system
from C, we obtain the 𝜎-algebra B(R𝑑 ).
Moreover, let F be the family of all finite disjoint unions of 𝑑-dimensional rectangles in R𝑑 , then F
is an algebra. Clearly, 𝜎(F) is again the Borel 𝜎-algebra B(R𝑑 ). Hence, the monotone class generated
by F is exactly the Borel 𝜎-algebra B(R𝑑 ).
Remark 1.11. In many applications, the previous theorems are used in the following procedure. Let 𝑝 be
a property on subsets of 𝑋. For a given 𝜎-algebra Σ, we want to show that 𝑝 holds for every element in
Σ. To prove this, take a 𝜋-system (or an algebra) C which generates the 𝜎-algebra Σ. Define the family

F B { 𝐴 ∈ Σ | 𝑝( 𝐴) holds}.

Now, assume that C ⊆ F. Then, by showing F is a 𝜆-system (or a monotone class), we conclude that
Σ = F by Dynkin’s 𝜋-𝜆 theorem (or the monotone class theorem).

We now introduce the notion of measures.

Definition 1.12. Let A is a 𝜎-algebra on a set 𝑋. A map 𝜇 : A → [0, ∞] is said to be countably additive
if for every countable disjoint family ( 𝐴𝑛 ) 𝑛=1,2,... of sets that belong to A,


! ∞
Ø Õ
𝜇 𝐴𝑛 = 𝜇( 𝐴𝑛 ).
𝑛=1 𝑛=1

If a map 𝜇 : A → [0, ∞] is countably additive and 𝜇(∅) = 0, then 𝜇 is called a measure on A.


Similarly, we can define finite additive maps and finitely additive measures on an algebra

Definition 1.13. Let A be a 𝜎-algebra on a set 𝑋, and let 𝜇 be a measure on A. Then, the pair (𝑋, A) is
called a measurable space, and the triplet (𝑋, A, 𝜇) is called a measure space.

Example 1.14.

3
1. Let A be a 𝜎-algebra on a set 𝑋. Define 𝜇 : A → [0, ∞] as follows:
(
| 𝐴| if 𝐴 is a finite set
𝜇( 𝐴) B
∞ if 𝐴 is an infinite set,

where | 𝐴| is the cardinality of the set 𝐴. Then, 𝜇 is a measure on A, which is called the counting
measure on (𝑋, A).

2. Let A be a 𝜎-algebra on a nonempty set 𝑋. Fix 𝑥 ∈ 𝑋, and define 𝛿 𝑥 : A → [0, ∞] by


(
1 if 𝑥 ∈ 𝐴
𝛿 𝑥 ( 𝐴) B
0 otherwise.

Then, 𝛿 𝑥 is a measure on (𝑋, A), which is called the point mass concentrated at 𝑥 (or the Dirac
measure centered at 𝑥).

3. Later, we will construct a measure on (R, B(R)) which assigns the length to each interval in R,
which is called the Lebesgue measure on (R, B(R)).

We now study some basic properties of a measure.

Proposition 1.15 (Monotonicity). Let 𝜇 be a measure on (𝑋, A). Then, for each pair 𝐴, 𝐵 ∈ A with
𝐴 ⊆ 𝐵, 𝜇( 𝐴) 6 𝜇(𝐵). Moreover, if 𝜇( 𝐴) < ∞, then 𝜇(𝐵 \ 𝐴) = 𝜇(𝐵) − 𝜇( 𝐴).

Proof. By the additivity of 𝜇, observe that

𝜇(𝐵) = 𝜇( 𝐴) + 𝜇(𝐵 \ 𝐴).

Corollary 1.16 (Countable subadditivity). Let 𝜇 be a measure on (𝑋, A), and let ( 𝐴𝑛 ) be a countable
subfamily of A. Then, !
Ø∞ Õ∞
𝜇 𝐴𝑛 6 𝜇( 𝐴𝑛 ).
𝑛=1 𝑛=1

Proof. Define 𝐵1 B 𝐴1 and


𝑛−1
Ø
𝐵 𝑛 B 𝐴𝑛 \ 𝐴𝑖
𝑖=1

for each 𝑛 > 1. Then, (𝐵𝑛 ) is a disjoint subfamily of A, and its union is the same with the union of ( 𝐴𝑛 ).
Hence, by additivity and monotonicity,

! ∞
!
Ø Ø Õ Õ
𝜇 𝐴𝑛 = 𝜇 𝐵𝑛 = 𝜇(𝐵𝑛 ) 6 𝜇( 𝐴𝑛 ).
𝑛=1 𝑛=1 𝑛 𝑛

Proposition 1.17 (Continuity). Let (𝑋, A, 𝜇) be a measure space. Then, the following hold:
Ð
1. (continuity from below) If ( 𝐴𝑛 ) is an increasing sequence in A, then 𝜇( 𝑛 𝐴𝑛 ) = lim𝑛→∞ 𝜇( 𝐴𝑛 ).

2. (continuity from above) If ( 𝐴𝑛 ) is a decreasing sequence in A with 𝜇( 𝐴 𝑘 ) < ∞ for some 𝑘, then
Ñ
𝜇( 𝑛 𝐴𝑛 ) = lim𝑛→∞ 𝜇( 𝐴𝑛 ).

4
Proof. We first prove the first statement. Let ( 𝐴𝑛 ) be an increasing sequence in A. Then, define 𝐵1 B 𝐴1
and
𝐵𝑛 B 𝐴𝑛 \ 𝐴𝑛−1
𝐵𝑛 = 𝐴 𝑘 for each 𝑘 = 1, 2, . . . , and ∞
Ð𝑘 Ð
for each 𝑛 > 1. Then, (𝐵𝑛 ) is a disjoint family, 𝑛=1 𝑛=1 𝐵 𝑛 =
Ð∞
𝑛=1 𝐴𝑛 . Thus, by additivity,


! ∞
! ∞
Ø Ø Õ
𝜇 𝐴𝑛 = 𝜇 𝐵𝑛 = 𝜇(𝐵𝑛 ).
𝑛=1 𝑛=1 𝑛=1

Moreover, for each 𝑘 = 1, 2, . . . , by additivity,


𝑘 𝑘
!
Õ Ø
𝜇(𝐵𝑛 ) = 𝜇 𝐵𝑛 = 𝜇( 𝐴 𝑘 ).
𝑛=1 𝑛=1
Í∞
Therefore, 𝑛=1 𝜇(𝐵𝑛 ) = lim𝑛→∞ 𝜇( 𝐴𝑛 ).
For the second statement, let ( 𝐴𝑛 ) be a decreasing sequence in A with 𝜇( 𝐴 𝑘 ) < ∞ for some 𝑘. Note
that by monotonicity, 𝜇( 𝐴𝑛 ) 6 𝜇( 𝐴 𝑘 ) < ∞ if 𝑛 > 𝑘. Consider the family (𝐶𝑛 ) given by 𝐶𝑛 = 𝐴 𝑘 \ 𝐴𝑛 .
Then, for 𝑛 = 1, . . . , 𝑘, 𝜇(𝐶𝑛 ) = 𝜇(∅) = 0, and by monotonicity, 𝜇(𝐶𝑛 ) = 𝜇( 𝐴 𝑘 ) − 𝜇( 𝐴𝑛 ) for 𝑛 > 𝑘.
Note that (𝐶𝑛 ) is an increasing sequence. Thus, by the first statement,
!
Ø
𝜇 𝐶𝑛 = lim 𝜇(𝐶𝑛 ).
𝑛→∞
𝑛

Note that ! !𝑐
Ø Ø Ø Ù Ù
𝐶𝑛 = ( 𝐴𝑘 ∩ 𝐴𝑛𝑐 ) = 𝐴𝑘 ∩ 𝐴𝑛𝑐 = 𝐴𝑘 ∩ 𝐴𝑛 = 𝐴𝑘 \ 𝐴𝑛 ,
𝑛 𝑛 𝑛 𝑛 𝑛
Ð Ñ Ð
so 𝜇( 𝑛 𝐶𝑛 ) = 𝜇( 𝐴 𝑘 ) − 𝜇( 𝑛 𝐴𝑛 ) since 𝜇( 𝑛 𝐴𝑛 ) 6 𝜇( 𝐴 𝑘 ) < ∞. Moreover, lim𝑛→∞ 𝜇(𝐶𝑛 ) =
lim𝑛→∞ (𝜇( 𝐴 𝑘 ) − 𝜇( 𝐴𝑛 )) = 𝜇( 𝐴 𝑘 ) − lim𝑛→∞ 𝜇( 𝐴𝑛 ). Hence, we obtain the required identity. 

Remark 1.18. For the continuity from above, we need to assume that 𝜇( 𝐴 𝑘 ) < ∞ for some 𝑘. To find a
counterexample, let 𝜇 be the counting measure on (N, P (N)). Define

𝐴𝑛 B N \ {1, 2, . . . , 𝑛}.

Then, this is a decreasing sequence, and 𝜇( 𝐴𝑛 ) = ∞ for all 𝑛, so lim𝑛→∞ 𝜇( 𝐴𝑛 ) = ∞. However,


Ñ
𝑛 𝐴𝑛 = ∅. This gives a required counterexample.
In fact, the continuity from below characterizes measures among finitely additive measures.

Theorem 1.19. Let (𝑋, A) be a measurable space, and let 𝜇 be a finitely additive measure on A. Then,
𝜇 is a measure if and only if 𝜇 satisfies the continuity from below.
Moreover, 𝜇 is a measure if the following holds: if ( 𝐴𝑛 ) is an decreasing sequence in A and
Ñ
𝑛 𝐴𝑛 = ∅, then lim𝑛→∞ 𝜇( 𝐴𝑛 ) = 0.

Proof. Let ( 𝐴𝑛 ) be a disjoint countable subfamily of A. Define


𝑛
Ø
𝐵𝑛 B 𝐴𝑖 .
𝑖=1

Then, (𝐵𝑛 ) is an increasing sequence in A, and by the finite additivity,


𝑛
Õ
𝜇(𝐵𝑛 ) = 𝜇( 𝐴𝑖 ),
𝑖=1

5
so lim𝑛→∞ 𝜇(𝐵𝑛 ) = ∞
Í Ð Ð
𝑛=1 𝜇( 𝐴𝑛 ). As 𝑛 𝐵𝑛 = 𝑛 𝐴𝑛 , continuity from below implies 𝜇 is countably
additive.
Moreover, define 𝐶𝑛 = ∞
Ð Ñ
𝑖=𝑛 𝐴𝑛 . Then, (𝐶𝑛 ) is a decreasing sequence in A and 𝑛 𝐶𝑛 = ∅ since
( 𝐴𝑛 ) is a disjoint family. Now, observe that

! 𝑖−1
! 𝑖−1
Ø Ø Õ
𝜇 𝐴𝑛 = 𝜇 𝐴𝑛 + 𝜇(𝐶𝑖 ) = 𝜇( 𝐴𝑛 ) + 𝜇(𝐶𝑖 )
𝑛=1 𝑛=1 𝑛=1
Ð Í
by the finite additivity. Therefore, if lim𝑛→∞ 𝜇(𝐶𝑛 ) = 0, then 𝜇( 𝑛 𝐴𝑛 ) = 𝑛 𝜇( 𝐴𝑛 ). This completes
the proof. 

In the sequel, the following terms frequently appear.

Definition 1.20. Let (𝑋, A, 𝜇) be a measure space.

1. 𝜇 is said to be finite if 𝜇(𝑋) < ∞.

2. 𝜇 is said to be 𝜎-finite if 𝑋 is the union of a countable family ( 𝐴𝑛 ) of measurable subsets of 𝑋


satisfying 𝜇( 𝐴𝑛 ) < ∞ for all 𝑛.

3. 𝜇 is said to be continuous if 𝜇({𝑥}) = 0 for each 𝑥 ∈ 𝑋.

4. 𝜇 is said to be discrete if there is a countable measurable subset 𝐷 of 𝑋 such that 𝜇(𝐷 𝑐 ) = 0.

Remark 1.21. If 𝜇 is 𝜎-finite, as the proof of conuntable subadditivity, we can make 𝑋 the countable
disjoint union of measurable subsets 𝐴𝑛 of 𝑋 such that 𝜇( 𝐴𝑛 ) < ∞ for all 𝑛. Hear, considering the proof
of the previous theorem, we can take an increasing sequence of measurable subsets 𝐴𝑛 of 𝑋 such that
𝜇( 𝐴𝑛 ) < ∞.

Example 1.22.

1. Let 𝜇 be the counting measure on a measurable space (𝑋, A). Then, 𝜇 is finite if and only if 𝑋 is
a finite set. Moreover, 𝜇 is 𝜎-finite if and only if 𝑋 is a countable set.

2. Let 𝛿 𝑥 be a point mass on a measurable space (𝑋, A) centered at 𝑥 ∈ 𝑋. Suppose there is a


countable subset 𝐷 of 𝑋 such that 𝐷 ∈ A and 𝑥 ∈ 𝐷. Then, since 𝜇(𝑋 \ 𝐷) = 0, 𝛿 𝑥 is discrete.

3. We will see that the Lebesgue measure on (R, B(R)) is a continuous measure.

1.2 Construction of measures by outer measures


We now see the standard method to construct measures. The outline is as follows: first, construct an
outer measure on a set 𝑋, which is a function on the power set P (𝑋), and then define the measurability
relative to the outer measure; these measurable sets form a 𝜎-algebra, and the restriction of the outher
measure on this 𝜎-algebra becomes a measure.

Definition 1.23. An outer measure on a set 𝑋 is a map 𝜇∗ : P (𝑋) → [0, ∞] such that the following hold:

1. 𝜇∗ (∅) = 0.

2. 𝜇∗ is monotone.

3. 𝜇∗ is countably subadditive.

An important example of an outer measure is the Lebesgue outer measure on R𝑑 , but we will see the
definition in the next section. In this section, we just discuss the method to construct a measure from a
given outer measure.

6
Definition 1.24 (Carathéodory condition). Let 𝜇 ( be an outer measure on a set 𝑋. A subset 𝐵 of 𝑋 is
said to be 𝜇∗ -measurable if for every subset 𝐴 of 𝑋,

𝜇∗ ( 𝐴) = 𝜇∗ ( 𝐴 ∩ 𝐵) + 𝜇∗ ( 𝐴 ∩ 𝐵 𝑐 ).

Remark 1.25. By subadditivity, 𝐵 is 𝜇∗ -measurable if and only if for every subset 𝐴 of 𝑋 with 𝜇∗ ( 𝐴) < ∞,

𝜇∗ ( 𝐴) > 𝜇∗ ( 𝐴 ∩ 𝐵) + 𝜇∗ ( 𝐴 ∩ 𝐵 𝑐 ).

Remark 1.26. If either 𝜇∗ (𝐵) = 0 or 𝜇∗ (𝐵 𝑐 ) = 0, then 𝐵 is 𝜇∗ -measurable. This can be easily showed by
the monotonicity of 𝜇∗ . Consequently, ∅ and 𝑋 are 𝜇∗ -measurable.

Theorem 1.27 (Carathéodory’s theorem). Let 𝜇∗ be an outer measure on a set 𝑋, and let M 𝜇∗ be the
family of all 𝜇∗ -measurable subsets of 𝑋. Then, the following hold:

1. M 𝜇∗ is a 𝜎-algebra.

2. The restriction of 𝜇∗ on the 𝜎-algebra M 𝜇∗ is a measure.

Proof. The first part of the proof is not hard, but tedious. To prove this, first show that M 𝜇∗ is an algebra,
and then show that M 𝜇∗ is closed under countable disjoint union. For a detailed proof, see [Coh13, pp.
16-17]. Now, we only show that 𝜇∗ is countably additive on M 𝜇∗ .
We first show that 𝜇∗ is finitely additive on M 𝜇∗ by using induction. Suppose 𝜇∗ is additive for
𝑘 disjoint 𝜇∗ -measurable subsets of 𝑋. Let ( 𝐴𝑛 ) 𝑛=1,...,𝑘+1 be a finite disjoint family of 𝜇∗ -measurable
subsets of 𝑋. Then, since 𝐴 𝑘+1 is 𝜇∗ -measurable, by the induction hypothesis,

𝑘+1
! 𝑘+1
!! 𝑘+1
!!
Ø Ø Ø
𝜇∗ 𝐴𝑛 = 𝜇∗ 𝐴 𝑘+1 ∩ 𝐴𝑛 + 𝜇∗ 𝐴𝑐𝑘+1 ∩ 𝐴𝑛
𝑛=1 𝑛=1 𝑛=1
𝑘
Õ
= 𝜇∗ ( 𝐴 𝑘+1 ) + 𝜇∗ ( 𝐴𝑛 )
𝑛=1
𝑘+1
Õ
= 𝜇∗ ( 𝐴𝑛 ).
𝑛=1

Therefore, 𝜇∗ is finitely additive on M 𝜇∗ .


Now, let ( 𝐴𝑛 ) be a countable disjoint collection of 𝜇∗ -measurable subsets of 𝑋. By monotonicity,
countable subadditivity and finite additivity, we obtain that for each 𝑘,

𝑘 ∞
! ∞
Õ Ø Õ
∗ ∗
𝜇 ( 𝐴𝑛 ) 6 𝜇 𝐴𝑛 6 𝜇∗ ( 𝐴𝑛 ).
𝑛=1 𝑛=1 𝑛=1

Letting 𝑘 → ∞, we have

! ∞
Ø Õ

𝜇 𝐴𝑛 = 𝜇∗ ( 𝐴𝑛 ).
𝑛=1 𝑛=1

This shows the countable additivity. 

Definition 1.28. Let 𝜇 be a measure on a measurable space (𝑋, A). 𝜇 is said to be complete if the
following holds: if 𝐴 is measurable and 𝜇( 𝐴) = 0, then every subset of 𝐴 is measurable.

Remark 1.29. The measure constructed by the previous method is complete.

7
1.3 Lebesgue measure and Lebesgue-Stieltjes measures
Definition 1.30. Define a function 𝜆∗ : P (R𝑑 ) → [0, ∞] as follows. Let C𝑏 the collection of all
𝑑-dimensional bounded open rectangles in R𝑑 . Then, for each subset 𝐴 of R𝑑 , define
(∞ ∞
)
Õ Ø
𝜆∗ ( 𝐴) B inf vol(𝑅𝑖 ) 𝐴 ⊆ 𝑅𝑖 and 𝑅𝑖 ∈ C𝑏 for all 𝑖 .
𝑖=1 𝑖=1

We call 𝜆∗ the Lebesgue outer measure on R𝑑 .

Lemma 1.31. 𝜆∗ is indeed an outer measure, and for each rectangle 𝑅 in R𝑑 , 𝜆∗ (𝑅) = vol(𝑅).

Proof. See [Coh13, pp. 13-15] and [Tes21, Lemma 1.8]. 

Definition 1.32. A 𝜆∗ -measurable subset of R𝑑 is said to be Lebesgue measurable.

Proposition 1.33. Every Borel subset of R𝑑 is Lebesgue measurable.

Lemma 1.34. The Borel 𝜎-algebra on R𝑑 is generated by each of the following collection of sets:

1. the collection of all closed half-spaces in R𝑑 that have the form

{(𝑥 1 , . . . , 𝑥 𝑑 ) | 𝑥 𝑘 6 𝑏}

for some 𝑘 = 1, . . . , 𝑑 and 𝑏 ∈ R

2. the collection of all rectangles in R𝑑 that have the form

{(𝑥 1 , . . . , 𝑥 𝑑 ) | 𝑎 𝑖 < 𝑥𝑖 6 𝑏 𝑖 for all 𝑖 = 1, . . . , 𝑑}

with (𝑎 1 , . . . , 𝑎 𝑑 ), (𝑏 1 , . . . , 𝑏 𝑑 ) ∈ R𝑑 .

Proof. Let B1 and B2 be the 𝜎-algebras generated by the collections above, respectively. Define 𝐻 𝑘,𝑏 B
{(𝑥 1 , . . . , 𝑥 𝑑 ) | 𝑥 𝑘 6 𝑏}. Then,
𝑑
Ù
(𝑎 1 , 𝑏 1 ] × . . . (𝑎 𝑑 , 𝑏 𝑑 ] = 𝐻𝑖,𝑏𝑖 \ 𝐻𝑖,𝑎𝑖 .
𝑖=1

Thus, B2 ⊆ B1 . Moreover, it is clear that B1 ⊆ B(R𝑑 ).


We now want to show that B(R𝑑 ) ⊆ B2 . Note that the collection of all bounded open rectangles in
R𝑑 is a countable base for the topology of R𝑑 , and that it each bounded open rectangle in R𝑑 is the union
of a countable family of rectangles of the form (𝑎 1 , 𝑏 1 ] × · · · × (𝑎 𝑑 , 𝑏 𝑑 ]. Therefore, B(R𝑑 ) ⊆ B2 . 

Proof of Proposition. We begin with showing that every closed half-space in R𝑑 is Lebesgue measurable.
Let 𝐻 be a closed half-space in R𝑑 . Let 𝐴 be a subset of 𝑋 with 𝜆∗ ( 𝐴) < ∞. Then, for each 𝜖 > 0, there
Ð
exists a countable family of open rectangles 𝑅𝑖 such that 𝐴 ⊆ 𝑖 𝑅𝑖

Õ
vol(𝑅𝑖 ) < 𝜆∗ ( 𝐴) + 𝜖 .
𝑖=1

Now, observe that


Õ
𝜆∗ ( 𝐴 ∩ 𝐻) 6 𝜆∗ (𝑅𝑖 ∩ 𝐻)
𝑖
Õ

𝜆 (𝐴 ∩ 𝐻 ) 6 𝑐
𝜆∗ (𝑅𝑖 ∩ 𝐻 𝑐 ),
𝑖

8
so Õ
𝜆∗ ( 𝐴 ∩ 𝐻) + 𝜆∗ ( 𝐴 ∩ 𝐻 𝑐 ) 6 (𝜆∗ (𝑅𝑖 ∩ 𝐻) + 𝜆∗ (𝑅𝑖 ∩ 𝐻 𝑐 )) .
𝑖
Note that for 𝑅𝑖 ∩ 𝐻 and 𝑅𝑖 ∩ 𝐻𝑐 are disjoint rectangles, and by direct calculation, we can show that

vol(𝑅𝑖 ) = vol(𝑅𝑖 ∩ 𝐻) + vol(𝑅𝑖 ∩ 𝐻 𝑐 )

for each 𝑖. Hence,


𝜆∗ ( 𝐴 ∩ 𝐻) + 𝜆∗ ( 𝐴 ∩ 𝐻 𝑐 ) < 𝜆∗ ( 𝐴) + 𝜖 .
Since 𝜖 is arbitrary, 𝐻 is Lebesgue measurable.
Therefore, by the previous lemma, the 𝜎-algebra M𝜆∗ contains the Borel 𝜎-algebra B(R𝑑 ). 

Definition 1.35. The restriction of 𝜆∗ to M𝜆∗ (or B(R𝑑 )) is called the Lebesgue measure and is denoted
by 𝜆.

We now discuss the finite Lebesgue-Stieltjes measures on the real line R.1

Proposition 1.36. Let 𝜇 be a finite measure on (R, B(R)). Define 𝐹𝜇 : R → R by

𝐹𝜇 (𝑥) B 𝜇((−∞, 𝑥]), 𝑥 ∈ R,

calle the distribution function of 𝜇. Then, 𝐹𝜇 is bounded, nondecreasing and right-continuous, and
satisfies lim 𝑥→−∞ 𝐹𝜇 (𝑥) = 0.

Proof. By the monotinicity and finiteness of 𝜇, 𝐹𝜇 is bounded and nondecreasing. Moreover, the
right-continuity of 𝐹𝜇 follows from the continuity of 𝜇 from above. Similarly, we can show that
lim 𝑥→−∞ 𝐹𝜇 (𝑥) = 0. 

Remark 1.37. Note that for 𝑎 < 𝑏 in R,

𝜇((𝑎, 𝑏]) = 𝐹𝜇 (𝑏) − 𝐹𝜇 (𝑎).

Moreover, since 𝐹𝜇 is bounded and nondecreasing, there exists the limit 𝐹𝜇 (𝑥−) for all 𝑥 ∈ R. Thus, one
can show that
𝜇({𝑥}) = 𝐹𝜇 (𝑥) − 𝐹𝜇 (𝑥−).
Therefore, 𝐹𝜇 is countinuous at 𝑥 if and only if 𝜇({𝑥} = 0, and so 𝜇 is continuous if and only if 𝐹𝜇 is
continuous on R.
From this remark, we can reconstruct the measure 𝜇 from the distribution function 𝐹𝜇 .

Proposition 1.38. Let 𝐹 : R → R be a bounded, nondecreasing and right-continuous function, and


suppose that lim 𝑥→−∞ 𝐹 (𝑥) = 0. Then, there exists a unique finite measure on (R, B(R)) such that
𝐹 (𝑥) = 𝜇((−∞, 𝑥]) for all 𝑥 ∈ R.

To show the uniqueness, we use the following lemma.

Lemma 1.39. Let (𝑋, A) be a measurable space, and let C be a 𝜋-system on 𝑋 with 𝜎(C) = A. Suppose
that 𝜇 and 𝜈 are two finite measure on A such that 𝜇(𝑋) = 𝜈(𝑋) and 𝜇(𝐶) = 𝜈(𝐶) for every 𝐶 ∈ C.
Then, 𝜇 = 𝜈.

Proof. Define D = {𝐴 ∈ A | 𝜇( 𝐴) = 𝜈( 𝐴)}. Then, it is easy to show that D is a 𝜆-system since both
𝜇 and 𝜈 are finite measures, and 𝜇(𝑋) = 𝜈(𝑋). Moreover, C ⊆ D by the assumption. Therefore, by
Dynkin’s 𝜋-𝜆 theorem, 𝜎(C) = A ⊆ D. This completes the proof. 
1 Forinfinite Lebesgue-Stieltjes measures, see [Fol99, Theorem 1.16]. The discussion is quiet similar to the following, but
it needs to deal with some technical issues.

9
Proof of Proposition. To show the uniqueness, observe that the collection of all intervals in R of the
form (−∞, 𝑏] forms a 𝜋-system on R. Moreover, we already know that this 𝜋-system generates the Borel
𝜎-algebra B(R). Now, let 𝜇 and 𝜈 be two finite measure on (R, B(R)) such that 𝐹𝜇 = 𝐹𝜈 . Then, by
continuity from below, 𝜇(R) = 𝜈(R). Therefore, since 𝐹𝜇 = 𝐹𝜈 , 𝜇 = 𝜈 by the previous lemma. This
proves the uniqueness.
To show the existence, we proceed the standard construction. We first need to define an outher
measrue on R from the function 𝐹. Define
(∞ ∞
)
Õ Ø
𝜇∗ ( 𝐴) B (𝐹 (𝑏 𝑛 ) − 𝐹 (𝑎 𝑛 )) 𝐴 ⊆ (𝑎 𝑛 , 𝑏 𝑛 ] .
𝑛=1 𝑛=1

Then, as the proof of Lemma 1.31, one can show that 𝜇∗ is an outer measure on R. Moreover, it is not
hard to prove that 𝜇∗ ((−∞, 𝑥]) = 𝐹 (𝑥) for all 𝑥 ∈ R (see [Coh13, p. 20]). Moreover, as the proof of
Proposition 1.33, every Borel subset of R is 𝜇∗ -measurable.
Now, let 𝜇 be the restriction of 𝜇∗ on B(R). Then, 𝜇 is a measure and satisfies 𝐹𝜇 = 𝐹. Moreover,
since 𝐹 is bounded, 𝜇 is finite. 

1.4 Properties of Lebesgue measure


We begin this section with the regularity of the Lebesgue measure on R𝑑 .
Proposition 1.40. Let 𝐴 be a Lebesgue measurable subset of R𝑑 . Then,
1. (outer regularity) 𝜆( 𝐴) = inf{𝜆(𝑈) | 𝐴 ⊆ 𝑈 and 𝑈 is open}.
2. (inner regularity) 𝜆( 𝐴) = sup{𝜆(𝐾) | 𝐾 ⊆ 𝐴 and 𝐾 is compact}.
Proof. Define 𝛼 = inf{𝜆(𝑈) | 𝐴 ⊆ 𝑈 and 𝑈 is open} and 𝛽 = sup{𝜆(𝐾) | 𝐾 ⊆ 𝐴 and 𝐾 is compact}.
Then, by the monotonicity of 𝜆,
𝛽 6 𝜆( 𝐴) 6 𝛼.
We first show that 𝛼 6 𝜆( 𝐴). We may assume that 𝜆( 𝐴) < ∞. Then, for each 𝜖 > 0, there exists a
countable family of open rectangles 𝑅𝑖 in R𝑑 such that 𝐴 ⊆ ∞
Ð
𝑖=1 𝑅𝑖 and

Õ
vol(𝑅𝑖 ) < 𝜆( 𝐴) + 𝜖 .
𝑖=1

Moreover, by the definition of 𝛼,



! ∞
Ø Õ
𝛼6𝜆 𝑅𝑖 6 vol(𝑅𝑖 ).
𝑖=1 𝑖=1

Therefore,
𝛼 < 𝜆( 𝐴) + 𝜖,
but since 𝜖 is arbitrary, 𝛼 6 𝜆( 𝐴).
We now show that 𝛽 > 𝜆( 𝐴). Here, we may assume that 𝛽 < ∞. First, suppose that 𝐴 is bounded.
Then, there exists a compact rectangle 𝐶 with 𝐴 ⊆ 𝐶. Note that 𝜆( 𝐴) 6 𝜆(𝐶) < ∞. We now approximate
𝐶 \ 𝐴 by the outer regularity. Let 𝜖 > 0 be given. Then, there exists an open set 𝑈 such that 𝐶 \ 𝐴 ⊆ 𝑈
and
𝜆(𝑈) < 𝜆(𝐶 \ 𝐴) + 𝜖 = 𝜆(𝐶) − 𝜆( 𝐴) + 𝜖
Now, consider 𝐾 = 𝐶 \ 𝑈. Then, 𝐾 is closed and bounded, so it is compact by the Heine-Borel theorem.
Since 𝐾 ⊆ 𝐴, 𝜆(𝐾) 6 𝛽, and since 𝐶 ⊆ 𝐾 ∪ 𝑈, we have

𝜆(𝐶) 6 𝜆(𝐾) + 𝜆(𝑈) < 𝛽 + 𝜆(𝐶) − 𝜆( 𝐴) + 𝜖 .

10
Thus, 𝜆( 𝐴) < 𝛽 + 𝜖, and since 𝜖 is arbitrary, 𝜆( 𝐴) 6 𝛽.
Now, assume that 𝐴 is not bounded. For each 𝑛 ∈ N, define 𝑅𝑛 B [−𝑛, 𝑛] 𝑑 ∈ R𝑑 , and let
𝐴𝑛 B 𝐴 ∩ 𝑅𝑛 . Then, each 𝐴𝑛 is a bounded subset of R𝑑 , so we can apply the previous result to 𝐴𝑛 . Let
𝜖 > 0 be given. Then, for each 𝑛, there exists a compact set 𝐾𝑛 with 𝐾𝑛 ⊆ 𝐴𝑛 such that

𝜆( 𝐴𝑛 ) < 𝜆(𝐾𝑛 ) + 𝜖 .
Ð
Note that by the continuity from below, lim 𝜆( 𝐴𝑛 ) = 𝜆( 𝐴𝑛 ) = 𝜆( 𝐴), and moreover, 𝜆(𝐾𝑛 ) 6 𝛽 for
each 𝑛. Thus,
𝜆( 𝐴) 6 𝛽 + 𝜖 . (1.1)
Since 𝜖 is arbitrary, 𝜆( 𝐴) 6 𝛽. 

The following proposition is about the uniqueness of the Lebesgue measure.

Proposition 1.41. The Lebesgue meaures on R𝑑 is the only measure on (R𝑑 , B(R𝑑 )) which assigns the
volume to each 𝑑-dimensional rectangle.

Lemma 1.42. Let (𝑋, A) be a measurable space, and let C be a 𝜋-system on 𝑋 with 𝜎(C) = A. For two
measures 𝜇 and 𝜈 on A, suppose the following:

1. 𝜇(𝐶) = 𝜈(𝐶) for all 𝐶 ∈ C.


Ð
2. There is an increasing sequence (𝐶𝑛 ) in C such that 𝜇(𝐶𝑛 ) = 𝜈(𝐶𝑛 ) < ∞ and 𝑛 𝐶𝑛 = 𝑋.

Then, 𝜇 = 𝜈.

Proof. After taking some sequence (𝐶𝑛 ) as in the second statement, for each 𝑛, define measures 𝜇 𝑛 and
𝜈𝑛 by
𝜇 𝑛 ( 𝐴) B 𝜇( 𝐴 ∩ 𝐶𝑛 ), 𝜈𝑛 ( 𝐴) B 𝜈( 𝐴 ∩ 𝐶𝑛 ).
Then, since C is a 𝜋-system, 𝜇 𝑛 (𝐶) = 𝜈𝑛 (𝐶) for all 𝐶 ∈ C, so 𝜇 𝑛 = 𝜈𝑛 . Then, by continuity from below,
for all 𝐴 ∈ A,
𝜇( 𝐴) = lim 𝜇 𝑛 ( 𝐴) = lim 𝜈𝑛 ( 𝐴) = 𝜈( 𝐴).
This completes the proof. 

Proof of Proposition. Let C be the collection of 𝑑-dimensional rectangles in R𝑑 (including the empty
Ð
set). Then, C is a 𝜋-system with 𝜎(C) = B(R𝑑 ). Moreover, R𝑑 = 𝑛 (−𝑛, 𝑛) 𝑑 , and to each 𝑛,
vol((−𝑛, 𝑛) 𝑑 ) < ∞. Now, apply the lemma. 

We now show the translation invariance of the Lebesgue measure.

Proposition 1.43. Lebesgue outer measure is translation invariant. Moreover, a subset 𝐴 of R𝑑 is


Lebesgue measurable if and only if 𝐴 + 𝑥 is Lebesgue measurable for some (and hence for all) 𝑥 ∈ R𝑑 .

Proof. This follows from the fact that the volume of a rectangle is translation invariant and the definition
of the Lebesgue outer measure. 

The following proposition characterizes the Lebesgue measure.

Proposition 1.44. Let 𝜇 be a nonzero measure on (R𝑑 , B(R𝑑 )). Suppose that 𝜇(𝐵) < ∞ for every
bounded Borel subset 𝐵 of R𝑑 and that 𝜇 is translation invariant. Then, there exists a positive constant
𝑐 such that 𝜇 = 𝑐𝜆.

11
Proof. Let 𝑐 be the measure of the unit half-open rectangle in R𝑑 with respect to 𝜇. Then, 𝑐 < ∞.
Moreover, by translating the unit rectangle, the union of the copies is R𝑑 . Since this union is countable
and disjoint, 𝑐 > 0.
Define a measure 𝜈 on B(R𝑑 ) by
1
𝜈( 𝐴) = 𝜇( 𝐴).
𝑐
Then, 𝜈 is translation invariant, and the measure of the unit half-open rectangle with respect to 𝜈 is 1. We
now claim that 𝜈 = 𝜆.
It suffices to show that 𝜈(𝑅) = vol(𝑅) for every rectangle 𝑅. By the continuity of 𝜇, we may assume
that 𝑅 is a half-open rectangle. By the translation invariance of 𝜇, we may further assume that 𝑅 has the
form [0, 𝑎 1 ) × · · · × [0, 𝑎 𝑑 ). Moreover, by the density of rational numbers in R and the continuity of 𝜇,
we can also assume that 𝑎 1 , . . . , 𝑎 𝑑 are rational numbers, so write
𝑛𝑗
𝑎𝑗 = , 𝑛 𝑗 , 𝑚 𝑗 ∈ N.
𝑚𝑗

Let 𝐷 = [0, 1/𝑚 1 ) × · · · × [0, 1/𝑚 𝑑 ). Then, by the translation invariance,

𝑚 1 . . . 𝑚 𝑑 · 𝜈(𝐷) = 1 and 𝑛1 . . . 𝑛 𝑑 · 𝜈(𝐷) = 𝜈(𝑅).

Thus,
𝑛1 . . . 𝑛 𝑑
𝜈(𝑅) = = 𝜆(𝑅).
𝑚1 . . . 𝑚 𝑑
This completes the proof. 

For a discussion on the Cantor set and nonmeasurable sets, see [Coh13, pp. 26-29].

12
2 Integration

2.1 Measurable functions


We begin with the following simple observation.

Lemma 2.1. Let (𝑋, A) be a measurable space, and let 𝐴 be a measurable subset of 𝑋. Then, for a
function 𝑓 : 𝐴 → [−∞, ∞], the following are equivalent:

1. For each 𝑡 ∈ R, {𝑥 ∈ 𝐴 | 𝑓 (𝑥) 6 𝑡} ∈ A.

2. For each 𝑡 ∈ R, {𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑡} ∈ A.

3. For each 𝑡 ∈ R, {𝑥 ∈ 𝐴 | 𝑓 (𝑥) > 𝑡} ∈ A.

4. For each 𝑡 ∈ R, {𝑥 ∈ 𝐴 | 𝑓 (𝑥) > 𝑡} ∈ A.

Proof. Note that


Ø
{𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑡} = {𝑥 ∈ 𝐴 | 𝑓 (𝑥) 6 𝑡 − 1/𝑛}
𝑛
Ø
{𝑥 ∈ 𝐴 | 𝑓 (𝑥) > 𝑡} = {𝑥 ∈ 𝐴 | 𝑓 (𝑥) > 𝑡 + 1/𝑛}.
𝑛

Since A is a 𝜎-algebra, we obtain the desired equivalences. 

Definition 2.2. If one (and hence all) of the conditions in the previous lemma holds, then 𝑓 is said to be
measurable with respect to A or A-measurable. If 𝑋 is a topological space, 𝑓 is called Borel measurable
or a Borel function if 𝑓 is measurable with respect to B(𝑋). In the case that 𝑋 = R𝑑 , 𝑓 is called Lebesgue
measurable if 𝑓 is measurable with respect to M𝜆∗ .

Example 2.3.

1. Every continuous function 𝑓 : 𝑋 → R is a Borel function.

2. If 𝐼 is an interval in R, then every nondecreasing function 𝑓 : 𝐼 → R is Borel measurable.

3. Every Borel function 𝑓 : R𝑑 → [−∞, ∞] is Lebesgue measurable.

4. Let (𝑋, A) be a measurable space, and let 𝐵 be a subset of 𝑋. Define the characteristic function of
𝐵 by (
1 if 𝑥 ∈ 𝐵
𝜒𝐵 (𝑥) B
0 otherwise
for 𝑥 ∈ 𝑋. Then, 𝜒𝐵 is measurable if and only if 𝐵 is measurable.

5. A function is called simple if it has only finitely many values. Let (𝑋, A) be a measurable space,
and let 𝑓 : 𝑋 → [−∞, ∞] be a simple function. Suppose 𝛼1 , . . . , 𝛼𝑛 be the values of 𝑓 . Then, 𝑓
is measurable if and only if {𝑥 ∈ 𝑋 | 𝑓 (𝑥) = 𝛼𝑖 } is measurable for all 𝑖.

Lemma 2.4. Let (𝑋, A) be a measurable space, and let 𝐴 ⊆ 𝑋 be measurable. Let 𝑓 and 𝑔 be measurable
functions on 𝐴 with values in [−∞, ∞]. Then, the following sets are measurable:

{𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑔(𝑥)}, {𝑥 ∈ 𝐴 | 𝑓 (𝑥) 6 𝑔(𝑥)} and {𝑥 ∈ 𝐴 | 𝑓 (𝑥) = 𝑔(𝑥)}.

13
Proof. By the density of rational numbers in R,
Ø
{𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑔(𝑥)} = ({𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑟} ∩ {𝑥 ∈ 𝐴 | 𝑟 < 𝑔(𝑥)}) .
𝑟 ∈Q

Since A is a 𝜎-algebra, we obtain the required results. 

We now discuss operations preserving the measurability of functions. The proofs are easy but tedious,
so we just accept the results. For detailed proofs, see [Coh13, pp. 43-46]. We start with the following
notations: for extended-real-valued functions 𝑓 and 𝑔 on 𝐴, define

𝑓 ∨ 𝑔 B max( 𝑓 , 𝑔) and 𝑓 ∧ 𝑔 B min( 𝑓 , 𝑔).

Proposition 2.5. Let (𝑋, A) be a measurable space, and let 𝐴 ∈ A. Let 𝑓 and 𝑔 be extended-real-valued
measurable functions on 𝐴, and let ( 𝑓𝑛 ) be a sequence of extended-real-valued measurable functions on
𝐴. Then, the following functions are all measurable:

1. 𝛼 𝑓 and 𝛼 + 𝑓 , where 𝛼 ∈ R

2. 𝑓 ∨ 𝑔 and 𝑓 ∧ 𝑔

3. sup𝑛 𝑓𝑛 and inf 𝑛 𝑓𝑛

4. lim sup𝑛 𝑓𝑛 and lim inf 𝑛 𝑓𝑛

5. lim𝑛 𝑓𝑛 whose domain is {𝑥 ∈ 𝐴 | lim sup𝑛 𝑓 (𝑥) = lim inf 𝑛 𝑓 (𝑥)}.

Proposition 2.6. Let (𝑋, A) be a measurable space, and let 𝐴 ∈ A. Let 𝑓 and 𝑔 be [0, ∞]-valued
measurable functions on 𝐴. Then, 𝑓 + 𝑔 are measurable.

Proposition 2.7. Let (𝑋, A) be a measurable space, and let 𝐴 ∈ A. Let 𝑓 and 𝑔 be real-valued
measurable functions on 𝐴, and let 𝛼 be a real number. Then, 𝛼 𝑓 , 𝑓 + 𝑔, 𝑓 − 𝑔, 𝑓 𝑔 and 𝑓 /𝑔 (where the
domain of 𝑓 /𝑔 is {𝑥 ∈ 𝐴 | 𝑔(𝑥) ≠ 0}) are measurable.

From now on, we consider preparation for defining integrals. Let 𝑓 be an extended-real-valued
function on a set 𝐴. Define the positive part 𝑓 + and the negative part 𝑓 − of 𝑓 by

𝑓 + B 𝑓 ∨ 0 and 𝑓 − B (− 𝑓 ) ∨ 0.

Moreover, for convenience, if 𝑝 is a condition on elements of 𝐴, then we usually write {𝑝} for {𝑥 ∈ 𝐴 |
𝑝(𝑥)}. For example, for a function 𝑓 : 𝐴 → [−∞, ∞], { 𝑓 < 𝑡} denotes the set {𝑥 ∈ 𝐴 | 𝑓 (𝑥) < 𝑡}.

Lemma 2.8. Let (𝑋, A) be a measurable space, and let 𝐴 ∈ A. Then, an extended-real-valued function
𝑓 on 𝐴 is measurable if and only if both 𝑓 + and 𝑓 − are measurable.
Moreover, 𝑓 = 𝑓 + − 𝑓 − and | 𝑓 | = 𝑓 + + 𝑓 − . In particular, | 𝑓 | is measurable if 𝑓 is measurable.

Proof. By the definition, if 𝑓 is measurable, then both 𝑓 + and 𝑓 − are measurable. For the converse,
observe that 𝐴 = { 𝑓 + = 0} ∪ { 𝑓 − = 0}. Thus, for each 𝑡 ∈ R,

{ 𝑓 > 𝑡} = { 𝑓 + < 𝑡} ∩ { 𝑓 − = 0} ∪ { 𝑓 − > −𝑡} ∩ { 𝑓 + = 0} .


 

Thus, if both 𝑓 + and 𝑓 − are measurable, then 𝑓 is also measurable.


Moreover, 𝑓 + − 𝑓 − is well-defined. Indeed, if 𝑓 − (𝑥) = ∞, then 𝑓 (𝑥) = −∞, so at this point, 𝑓 + (𝑥) = 0.
Now, the identities 𝑓 = 𝑓 + − 𝑓 − and | 𝑓 | = 𝑓 + + 𝑓 − follow from the definitions. 

Lemma 2.9. Let (𝑋, A) be a measurable space, and let 𝐴 ∈ A. Let 𝑓 : 𝐴 → [−∞, ∞] be a function.

1. If 𝐵 ⊆ 𝐴 with 𝐵 ∈ A, then the restriction 𝑓 | 𝐵 of 𝑓 to 𝐵 is measurable.

14
Ð
2. If 𝐴 = 𝑛 𝐵𝑛 with each 𝐵𝑛 belongs to A and if each restriction 𝑓 | 𝐵𝑛 is measurable, then 𝑓 is
measurable.

Proof. Just observe that

{ 𝑓 | 𝐵 < 𝑡} = 𝐵 ∩ { 𝑓 < 𝑡}
Ø
{ 𝑓 < 𝑡} = { 𝑓 | 𝐵𝑛 < 𝑡}.
𝑛

The following proposition says that every nonnegative measurable function can be approximated by
an increasing sequence of nonnegative measurable simple functions.

Proposition 2.10 (Approximation by simple functions). Let (𝑋, A) be a measurable space, and let
𝐴 ∈ A. Let 𝑓 be a [0, ∞]-valued measurable function on 𝐴. Then, there exists a sequence ( 𝑓𝑛 ) of simple
[0, ∞)-valued measurable functions on A such that

𝑓1 6 𝑓2 6 . . . and 𝑓 = lim 𝑓𝑛 .
𝑛

Proof. For each 𝑛 ∈ N and for 𝑘 = 1, 2 . . . , 𝑛2𝑛 , define


 
𝑘 −1 𝑘
𝐴𝑛,𝑘 B 6 𝑓 < 𝑛 .
2𝑛 2

Then, each 𝐴𝑛,𝑘 are measurable. Now, define 𝑓𝑛 : 𝐴 → [0, ∞) for each 𝑛 by
(
𝑘−1
2𝑛 if 𝑥 ∈ 𝐴𝑛,𝑘 for some 𝑘
𝑓𝑛 (𝑥) B Ð
𝑛 if 𝑥 ∈ 𝐴 \ 𝑘 𝐴𝑛,𝑘 .

Then, the sequence ( 𝑓𝑛 ) has the desired properties. We first show the monotonicity.

1. If 𝑓 (𝑥) > 𝑛 + 1, then 𝑓𝑛+1 (𝑥) − 𝑓𝑛 (𝑥) = 𝑛 + 1 − 𝑛 = 1 > 0.

2. if 𝑛 6 𝑓 (𝑥) < 𝑛 + 1, then 𝑓𝑛 (𝑥) = 𝑛, but 𝑥 ∈ 𝐴 (𝑛+1) ,𝑘 for some 𝑘 = 𝑛2𝑛+1 + 1, . . . , (𝑛 + 1)2𝑛+1 .
Thus, 𝑓𝑛+1 (𝑥) > 𝑛 = 𝑓𝑛 (𝑥).

3. If 0 6 𝑓 (𝑥) < 𝑛, then there exists unique 𝑘 ∈ {1, . . . , 𝑛2𝑛 } so that 𝑥 ∈ 𝐴𝑛,𝑘 . Then,

2𝑘 − 2 𝑘 − 1 𝑘 2𝑘
𝑛+1
= 𝑛 6 𝑓 (𝑥) < 𝑛 = 𝑛+1 .
2 2 2 2
Thus, either 𝑥 ∈ 𝐴 (𝑛+1), (2𝑘−1) or 𝑥 ∈ 𝐴 (𝑛+1),2𝑘 . Observe that

𝑘 − 1 2𝑘 − 2
𝑓𝑛 (𝑥) = = 𝑛+1
2𝑛 2
and that
2𝑘 − 2
if 𝑥 ∈ 𝐴 (𝑛+1), (2𝑘−1)



 2𝑛+1

𝑓𝑛+1 (𝑥) =
2𝑘 − 1
if 𝑥 ∈ 𝐴 (𝑛+1),2𝑘 .


 𝑛+1
 2
Thus, 𝑓𝑛 (𝑥) 6 𝑓𝑛+1 (𝑥).

15
We now show the pointwise convergence. If 𝑓 (𝑥) = ∞, then 𝑓 (𝑥) > 𝑛 for every 𝑛, so 𝑓𝑛 (𝑥) = 𝑛 → ∞.
Now, assume 0 6 𝑓 (𝑥) < ∞. Then, choose a positive integer 𝑚 so that 0 6 𝑓 (𝑥) < 𝑚. Then, for all
integer 𝑛 with 𝑛 > 𝑚, 0 6 𝑓 (𝑥) < 𝑛. Moreover, if 𝑥 ∈ 𝐴𝑛,𝑘 , then

1
0 6 𝑓 (𝑥) − 𝑓𝑛 (𝑥) < .
2𝑛
Thus, by letting 𝑛 → ∞, we conclude that 𝑓𝑛 (𝑥) → 𝑓 (𝑥). This completes the proof. 

The following terminology will be frequently used.

Definition 2.11. Let (𝑋, A, 𝜇) be a measure space, and let 𝑝 be a property on elements of a set 𝐸. 𝑝 is
said to hold 𝜇-almost everywhere on 𝐸 if there is a measurable set 𝑁 with 𝜇(𝑁) = 0 such that

{𝑥 ∈ 𝐸 | 𝑝(𝑥) fails} ⊆ 𝑁,

that is, 𝑝(𝑥) holds for all 𝑥 ∈ 𝐸 \ 𝑁.

Remark 2.12. In general, the set {𝑥 ∈ 𝐸 | 𝑝(𝑥) fails} is not necessarily measurable, but if 𝜇 is a complete
measure, then this set is measurable.

Definition 2.13. Let (𝑋, A, 𝜇) be a measure space. Moreover, let ( 𝑓𝑛 ) be a sequence of functions on 𝑋,
and let 𝑓 be a function on 𝑋. ( 𝑓𝑛 ) is said to converge almost everywhere to 𝑓 if there is a measurable
subset 𝑁 of 𝑋 such that 𝜇(𝑁) = 0 and ( 𝑓𝑛 ) converges to 𝑓 pointwise on 𝑋 \ 𝑁. In this case, we write
𝑓𝑛 → 𝑓 (𝜇)-a.e..

In the case of complete measure spaces, the following hold.

Proposition 2.14. Let (𝑋, A, 𝜇) be a complete measure space, and let 𝑓 and 𝑔 be extended-real-valued
functions on 𝑋.

1. If 𝑓 = 𝑔 𝜇-a.e. and if 𝑓 is measurable, then 𝑔 is measurable.

2. If 𝑓𝑛 → 𝑓 𝜇-a.e. and if each 𝑓𝑛 is measurable, then 𝑓 is measurable.

Proof. See [Coh13, pp. 50-51]. 

2.2 Integrals
In this section, we define the integration. We consider the integration in the following sequence:
simple nonnegative measurable functions, nonnegative measurable functions and integrable measurable
functions.
Let (𝑋, A, 𝜇) be a measure space. Let S be the family of all simple real-valued measurable functions
on 𝑋, and let S+ be the family of nonnegative functions in S.

Definition 2.15 (Integrals of simple functions). Let 𝑓 ∈ S+ , and write


𝑚
Õ
𝑓 = 𝑎 𝑖 𝜒 𝐴𝑖 , (2.1)
𝑖=1

where 𝑎 𝑖 ’s are nonnegative real numbers (may not be distinct) and 𝐴𝑖 ’s are disjoint measurable subsets
of 𝑋. Define ∫ Õ𝑚
𝑓 𝑑𝜇 B 𝑎 𝑖 𝜇( 𝐴𝑖 )
𝑖=1

which is called the integral of 𝑓 with respect to 𝜇.

16
Remark 2.16. One can show that the integral of 𝑓 is independent of the choice of an expression (2.1) for
𝑓 (see [Coh13, p. 53]).

Proposition 2.17. Let 𝑓 , 𝑔 ∈ S+ , and let 𝛼 be a nonnegative real number. Then, the following hold.
∫ ∫
1. 𝛼 𝑓 𝑑𝜇 = 𝛼 𝑓 𝑑𝜇.
∫ ∫ ∫
2. ( 𝑓 + 𝑔)𝑑𝜇 = 𝑓 𝑑𝜇 + 𝑔𝑑𝜇.
∫ ∫
3. If 𝑓 6 𝑔, then 𝑓 𝑑𝜇 6 𝑔𝑑𝜇.

Proof. These are immediate results from the definition. For a detailed proof, see [Coh13, Proposition
2.3.1]. 

Definition 2.18. Let 𝑓 be a [0, ∞]-valued measurable function on 𝑋. Define


∫ ∫ 
𝑓 𝑑𝜇 B 𝑔𝑑𝜇 𝑔 ∈ S+ and 𝑔 6 𝑓 .

By the previous proposition, this definition agrees with the previous definition for 𝑓 ∈ S+ .

Recall that if 𝑓 is a [0, ∞]-valued measurable


∫ function,
∫ then there exists a nondecreasing sequence
( 𝑓𝑛 ) in S+ such that 𝑓𝑛 → 𝑓 . We expect that 𝑓𝑛 𝑑𝜇 → 𝑓 𝑑𝜇. We prove a more general statement.

Theorem 2.19 (Monotone convergence theorem, version 1). Let 𝑓 be a [0, ∞]-valued measurable
function on 𝑋, and let ( 𝑓𝑛 ) be a nondecreasing sequence of [0, ∞]-valued measurable functions on 𝑋
such that 𝑓𝑛 → 𝑓 . Then, ∫ ∫
𝑓𝑛 𝑑𝜇 → 𝑓 𝑑𝜇.
∫ ∫ ∫
Proof. First, the ∫sequence (∫ 𝑓𝑛 𝑑𝜇) is nondecreasing and bounded above by 𝑓 ∫𝑑𝜇. Thus, lim𝑛∫ 𝑓𝑛 𝑑𝜇
exists, and lim𝑛 𝑓𝑛 𝑑𝜇 6 𝑓 𝑑𝜇. Next, we claim that for every 0 < 𝜖 < 1, 𝜖 𝑓 𝑑𝜇 6 lim𝑛∫ 𝑓𝑛 𝑑𝜇.
∫ this, it suffices to prove that for every 𝑔 ∈ S+ with 𝑔 6 𝑓 and for all 0 < 𝜖 < 1, 𝜖 𝑔𝑑𝜇 6
To show
lim𝑛 𝑓𝑛 𝑑𝜇.
Let 𝜖 ∈ (0, 1) be given, and let 𝑔 ∈ S+ with 𝑔 6 𝑓 . Let 𝑎 1 , . . . , 𝑎 𝑘 be the values of 𝑔, and let
𝐴𝑖 = 𝑔 −1 (𝑎 𝑖 ) for each 𝑖. Then, 𝑔 = 𝑖=1
Í𝑘
𝑎 𝑖 𝜒 𝐴𝑖 . Note that for each 𝑥 ∈ 𝐴𝑖 , if 𝑎 𝑖 > 0, then

sup 𝑓𝑛 (𝑥) = 𝑓 (𝑥) > 𝑔(𝑥) = 𝑎 𝑖 > 𝜖 𝑎 𝑖 ;


𝑛

if 𝑎 𝑖 = 0, then
sup 𝑓𝑛 (𝑥) = 𝑓 (𝑥) > 𝑔(𝑥) = 𝑎 𝑖 = 𝜖 𝑎 𝑖 = 0.
𝑛

In both cases, there exists 𝑛 such that 𝑓𝑛 (𝑥) > 𝜖 𝑎 𝑖 . Hence, if we write

𝐴(𝑛, 𝑖) B {𝑥 ∈ 𝐴𝑖 | 𝑓𝑛 (𝑥) > 𝜖 𝑎 𝑖 },

then Ø
𝐴𝑖 = 𝐴(𝑛, 𝑖).
𝑛

Note that for each 𝑖, ( 𝐴(𝑛, 𝑖)) is a nondecreasing sequence of measurable subsets of 𝑋. Now, for each 𝑛,
define
Õ 𝑘
𝑔𝑛 B 𝜖 𝑎 𝑖 𝜒 𝐴(𝑛,𝑖) .
𝑖=1

17
∫ ∫
Then, 𝑔𝑛 ∈ S+ and 𝑔𝑛 6 𝑓𝑛 . Therefore, 𝑔𝑛 𝑑𝜇 6 𝑓𝑛 𝑑𝜇, and by the continuity of 𝜇,
∫ 𝑘
Õ 𝑘
Õ ∫
lim 𝑔𝑛 𝑑𝜇 = lim 𝜖 𝑎 𝑖 𝜇( 𝐴(𝑛, 𝑖)) = 𝜖 𝑎 𝑖 𝜇( 𝐴𝑖 ) = 𝜖 𝑔𝑑𝜇.
𝑛 𝑛
𝑖=1 𝑖=1
∫ ∫ ∫ ∫
This shows
∫ that 𝜖 𝑔𝑑𝜇 6 lim 𝑛 𝑓 𝑛 𝑑𝜇. Consequently, 𝜖 𝑓 𝑑𝜇 6 lim 𝑛 𝑓𝑛 𝑑𝜇. Since 𝜖 is arbitrary, we
have 𝑓 𝑑𝜇 6 lim𝑛 𝑓𝑛 𝑑𝜇. 

Corollary 2.20. The results of Proposition 2.17 still hold for [0, ∞]-valued measurable functions 𝑓 and
𝑔

Proof. Use the approximation by simple functions and the monotone convergence theorem. 

Definition
∫ 2.21. Let∫ 𝑓 be an extended-real-valued measurable
∫ function on 𝑋. 𝑓 is said to be integrable
+ −
if both 𝑓 𝑑𝜇 and 𝑓 𝑑𝜇 exist. In this case, we define the 𝑓 𝑑𝜇 of 𝑓 by
∫ ∫ ∫
+
𝑓 𝑑𝜇 = 𝑓 𝑑𝜇 − 𝑓 − 𝑑𝜇.
∫ ∫ ∫
Moreover, even if 𝑓 is not integrable, we can define 𝑓 𝑑𝜇 in the case that one of 𝑓 + 𝑑𝜇 and 𝑓 − 𝑑𝜇 is
finite.
In the case that 𝑋 = R𝑑 and 𝜇 = 𝜆, we define Lebesgue integrability and Lebesgue integrals.

Remark 2.22. Let 𝐴 be a measurable subset of 𝑋. Then, 𝑓 is said to be integrable over 𝐴 if 𝑓 𝜒 𝐴 is


integrable. In this case, we define ∫ ∫
𝑓 𝑑𝜇 B 𝑓 𝜒 𝐴 𝑑𝜇.
𝐴

Moreover, if 𝑓 is measurable function on 𝐴, then extend 𝑓 by setting 𝑓 = 0 on 𝐴𝑐 , and define 𝐴
𝑓 𝑑𝜇 by
this extension.

Proposition 2.23. Let 𝑓 and 𝑔 be real-valued integrable functions on 𝑋, and let 𝛼 be a real number.
Then, 𝛼 𝑓 and 𝑓 + 𝑔 are integrable, and the results of Proposition 2.17 are hold for 𝑓 , 𝑔 and 𝛼.

Proof. The proof follows from the definitions and the previous result. For a detailed proof, see [Coh13,
Proposition 2.3.6]. 

Proposition 2.24. Let 𝑓 be an extended-real-valued measurable function on a measure space (𝑋, A, 𝜇).
Then, 𝑓 is integrable if and only if | 𝑓 | is integrable. In this case,
∫ ∫
𝑓 𝑑𝜇 6 | 𝑓 |𝑑𝜇.

Proof. This follows from the identity | 𝑓 | = 𝑓 + + 𝑓 − . 

Example 2.25.

1. If 𝜇 is a finite measure, then every bounded measurable function on (𝑋, A, 𝜇) is integrable.

2. Every bounded Borel function on the compact interval [𝑎, 𝑏] is Lebesgue integrable. In particular,
every continuous function on [𝑎, 𝑏] is Lebesgue integrable, and in fact, its integral is same with
the Riemann integral.

3. One can prove that every Riemann integrable function on [𝑎, 𝑏] is Lebesgue integrable, and the
integral of such a function is the same to the Riemann integral. The key ingredient for proving
this fact is Lebesgue’s criterion: a function on [𝑎, 𝑏] is Riemann integrable if and only if 𝑓 is
continuous almost everywhere. See [Coh13, Section 2.5].

18

4. Let 𝜇 be the counting measure on (N, P (N)). Then, for a nonnegative function 𝑓 on N, 𝑓 𝑑𝜇 =
Í
𝑛 𝑓 (𝑛), and 𝑓 is integrable if and only if Í
the series is convergent. Moreover, for a function
𝑓∫ on N is integrable if and only if the sum 𝑛 𝑓 (𝑛) is absolutely convergent, and in this case,
Í
𝑓 𝑑𝜇 = 𝑛 𝑓 (𝑛).

We end this section with some basic properties of integrable functions.

Proposition 2.26. Let 𝑓 and 𝑔 be extended-real-valued measurable


∫ functions∫ on a measure space
(∫ 𝐴, A, 𝜇).∫ Suppose that 𝑓 = 𝑔 almost everywhere and that 𝑓 𝑑𝜇 exists. Then, 𝑔𝑑𝜇 also exists, and
𝑓 𝑑𝜇 = 𝑔𝑑𝜇.

Proof. Suppose that 𝑓 and 𝑔 are nonnegative. Define

𝐴 B { 𝑓 ≠ 𝑔}.

Then, 𝐴 is measurable since 𝑓 and 𝑔 are measurable. Now, consider a function ℎ defined by
(
∞ if 𝑥 ∈ 𝐴
ℎ(𝑥) B
0 otherwise.

Then, since 𝑛𝜒 𝐴 → ℎ nondecreasingly,
∫ ∫ by the monotone
∫ convergence
∫ ∫ ℎ𝑑𝜇 =∫ 0. Moreover,
theorem,
since
∫ 𝑓 6∫ 𝑔 + ℎ, we have 𝑓 𝑑𝜇 6 𝑔𝑑𝜇 + ℎ𝑑𝜇 = 𝑔𝑑𝜇. Similarly, 𝑔𝑑𝜇 6 𝑓 𝑑𝜇. Thus,
𝑓 𝑑𝜇 = 𝑔𝑑𝜇.
For general cases, just observe the following: since 𝑓 = 𝑔 almost everywhere, 𝑓 + = 𝑔 + and 𝑓 − = 𝑔 −
almost everywhere. 

The following ineqaulity is frequently useful.

Proposition 2.27 (Chebyshev’s inequality). Let 𝑓 be a [0, ∞]-valued measurable function on a measure
space (𝑋, A, 𝜇). Then, for each positive real number 𝑡, the following hold:
∫ ∫
1 1
𝜇({ 𝑓 > 𝑡}) 6 𝑓 𝑑𝜇 6 𝑓 𝑑𝜇.
𝑡 { 𝑓 >𝑡 } 𝑡

Proof. Let 𝐴𝑡 B { 𝑓 > 𝑡}. Then, 𝑓 > 𝑓 𝜒 𝐴𝑡 > 𝑡 𝜒 𝐴𝑡 > 0, so


∫ ∫ ∫
𝑡 𝜒 𝐴𝑡 𝑑𝜇 6 𝑓 𝜒 𝐴𝑡 𝑑𝜇 6 𝑓 𝑑𝜇.

This inequality implies some basic consequences on integrable functions.

Corollary 2.28. Let 𝑓 be an extended-real-valued measurable function on (𝑋, A, 𝜇).

1. If 𝑓 is integrable, then the set { 𝑓 ≠ 0} is 𝜎-finite.



2. If | 𝑓 |𝑑𝜇 = 0, then 𝑓 = 0 almost everywhere.

Proof. These follow from the inequality



𝜇({| 𝑓 | > 1/𝑛}) 6 𝑛 | 𝑓 |𝑑𝜇.



Remark 2.29. By the second item, if 𝑓 > 0 almost everywhere, then 𝑓 𝑑𝜇 > 0.

19

Corollary 2.30. Let 𝑓 be an extended-real-valued integrable function on (𝑋, A, 𝜇). If 𝐴 𝑓 𝑑𝜇 > 0 for
all measurable subsets 𝐴, then 𝑓 > 0 almost everywhere.
∫ ∫
Proof. Define 𝐴 B { 𝑓 < 0}. Then, 𝐴 𝑓 𝑑𝜇 6 0. On the other hand, by the assumption, 𝐴 𝑓 𝑑𝜇 > 0.

This implies that 𝑓 𝜒 𝐴 𝑑𝜇 = 0. Thus, 𝑓 𝜒 𝐴 = 0 almost everywhere, and so 𝑓 > 0 almost everywhere. 

Corollary 2.31. Let 𝑓 be an extended-real-valued integrable function on (𝑋, A, 𝜇). Then, | 𝑓 | < ∞
almost everywhere.

Proof. This follows from the inequalities



1
𝜇({| 𝑓 | = ∞}) 6 𝜇({| 𝑓 | > 𝑛}) 6 | 𝑓 |𝑑𝜇.
𝑛


Corollary 2.32. Let 𝑓 be an extended-real-valued measurable function on (𝑋, A, 𝜇). Then, 𝑓 is integrable
if and only if there is a real-valued integrable function on 𝑋 which is equal to 𝑓 almost everywhere.

Proof. By a previous result, it suffices to show that if 𝑓 is integrable, then there is such a real-valued
integrable function. Suppose 𝑓 is integrable. Define 𝐴 = {| 𝑓 | = ∞}, and let 𝑓0 = 𝑓 𝜒 𝐴𝑐 . Then, 𝑓0 is
integrable and 𝑓0 = 𝑓 almost everywhere. 

2.3 Limit theorems


In this section, we discuss the basic limit theorems of integration theory.

Theorem 2.33 (Monotone convergence theorem, version 2). Let (𝑋, A, 𝜇) be a measure space. Let 𝑓
be a [0, ∞]-valued measurable function on 𝑋, and let ( 𝑓𝑛 ) be a sequence of [0, ∞]-valued measurable
functions on 𝑋.
Suppose
𝑓1 6 𝑓2 6 . . .
and
𝑓𝑛 → 𝑓
hold almost everywhere. Then, ∫ ∫
𝑓𝑛 𝑑𝜇 → 𝑓 𝑑𝜇.

Proof. The only difference with the previous version is that the assumptions are now in the a.e. sense. By
the assumptions, we can take a measurable subset 𝑁 of measure zero such that on 𝑁 𝑐 , ( 𝑓𝑛 ) is pointwise
increasing and 𝑓𝑛 → 𝑓 pointwise. Therefore, for the sequence ( 𝑓𝑛 𝜒 𝑁 𝑐 ), it is pointwise increasing and
converges to 𝑓 𝜒 𝑁 𝑐 pointwise. By the version 1, we have
∫ ∫
𝑓 𝜒 𝑁 𝑐 𝑑𝜇 = lim 𝑓𝑛 𝜒 𝑁 𝑐 𝑑𝜇.
𝑛

Since 𝑓 = 𝑓 𝜒 𝑁 𝑐 a.e. and 𝑓𝑛 = 𝑓𝑛 𝜒 𝑁 𝑐 a.e. for each 𝑛, we have the desired identity. 

Corollary 2.34 (Beppo Levi’s theorem). Let ( 𝑓𝑛 ) be a sequence of [0, ∞]-valued measurable functions
on 𝑋. Then,
∫ Õ ∞ Õ∞ ∫
𝑓𝑛 𝑑𝜇 = 𝑓𝑛 𝑑𝜇.
𝑛=1 𝑛=1

20
Remark 2.35. Let 𝑓 be a nonnegative measurable function on 𝑋. Define the map 𝜈 : A → [0, ∞] by
∫ ∫
𝜈( 𝐴) B 𝑓 𝑑𝜇 = 𝑓 𝜒 𝐴 𝑑𝜇.
𝐴

We want to show that 𝜈 is a measure. Indeed, 𝜈(∅) = 0. Moreover, if ( 𝐴𝑛 ) be a disjoint family of


measurable subsets, then since Õ
𝜒Ð𝑛 𝐴𝑛 = 𝜒 𝐴𝑛 ,
𝑛
! ∫ Õ
Ø Õ∫ Õ
𝜈 𝐴𝑛 = 𝑓 𝜒 𝐴𝑛 𝑑𝜇 = 𝑓 𝜒 𝐴𝑛 𝑑𝜇 = 𝜈( 𝐴𝑛 ).
𝑛 𝑛 𝑛 𝑛

Thus, 𝜈 is countably additive.


Moreover, 𝜈 is finite if and only if 𝑓 is integrable.
The following lemma is often used to estimate the upper bound for an integral.

Lemma 2.36 (Fatou’s lemma). Let ( 𝑓𝑛 ) be a sequence of [0, ∞]-valued measurable functions on
(𝑋, A, 𝜇). Then, ∫ ∫
lim inf 𝑓𝑛 𝑑𝜇 6 lim inf 𝑓𝑛 𝑑𝜇.
𝑛 𝑛

Proof. Note that lim inf 𝑛 = lim𝑛 inf 𝑘 >𝑛 . Define 𝑔𝑛 B inf 𝑘 >𝑛 𝑓𝑛 . Then, each 𝑔𝑛 is measurable, and the
sequence (𝑔𝑛 ) is nondecreasing such that

lim inf 𝑓𝑛 = lim 𝑔𝑛 .


𝑛 𝑛
∫ ∫
Thus, by the monotone convergence theorem, lim inf 𝑛 𝑓𝑛 𝑑𝜇 = lim𝑛 𝑔𝑛 𝑑𝜇, but since 𝑔𝑛 6 𝑓𝑛 for each
𝑛, we have ∫ ∫
lim inf 𝑓𝑛 𝑑𝜇 6 lim inf 𝑓𝑛 𝑑𝜇.
𝑛 𝑛

The following result is fundamental in integration theory.

Theorem 2.37 (Lebesgue’s dominated convergence theorem). Let (𝑋, A, 𝜇) be a measure space. Let 𝑔
be a [0, ∞]-valued integrable function on 𝑋, let 𝑓 be a [−∞, ∞]-valued measurable function on 𝑋, and
let ( 𝑓𝑛 ) be a sequence of [−∞, ∞]-valued measurable functions on 𝑋.
Suppose
𝑓𝑛 → 𝑓
and
| 𝑓𝑛 | 6 𝑔
for all 𝑛 almost everywhere. Then, 𝑓 and 𝑓1 , 𝑓2 , . . . are integrable, and
∫ ∫
𝑓 𝑑𝜇 = lim 𝑓𝑛 𝑑𝜇.
𝑛

Proof. Since 𝑔 is integrable, each 𝑓𝑛 is integrable. Moreover, since 𝑓𝑛 → 𝑓 , | 𝑓 | 6 𝑔 almost everywhere,


so 𝑓 is also integrable. Note that since 𝑔 is integrable, 𝑔 < ∞ almost everywhere.
Now, we assume that 𝑓𝑛 → 𝑓 , | 𝑓𝑛 | 6 𝑔 and 𝑔 < ∞ everywhere on 𝑋. By Fatou’s lemma,
∫ ∫
(𝑔 + 𝑓 )𝑑𝜇 6 lim inf (𝑔 + 𝑓𝑛 )𝑑𝜇,
𝑛

21
so ∫ ∫
𝑓 𝑑𝜇 6 lim inf 𝑓𝑛 𝑑𝜇.
𝑛

Similarly, ∫ ∫
(𝑔 − 𝑓 )𝑑𝜇 6 lim inf (𝑔 − 𝑓𝑛 )𝑑𝜇,
𝑛

and so ∫ ∫
𝑓 𝑑𝜇 > lim sup 𝑓𝑛 𝑑𝜇.
𝑛
∫ ∫
Since lim inf 𝑛 𝑓𝑛 𝑑𝜇 6 lim sup𝑛 𝑓𝑛 𝑑𝜇, we have
∫ ∫
𝑓 𝑑𝜇 = lim 𝑓𝑛 𝑑𝜇.
𝑛

Now, consider our original assumptions. Let 𝑁 be a measurable subset of 𝑋 such that 𝜇(𝑁) = 0
and 𝑓𝑛 → 𝑓 , | 𝑓𝑛 | 6 𝑔 and 𝑔 < ∞ on 𝑁 𝑐 . Then, we can apply the previous result to 𝑔 𝜒 𝑁 𝑐 , 𝑓 𝜒 𝑁 𝑐 and
( 𝑓𝑛 𝜒 𝑁 𝑐 ). Then, ∫ ∫
𝑓 𝜒 𝑁 𝑐 𝑑𝜇 = lim 𝑓𝑛 𝜒 𝑁 𝑐 𝑑𝜇.
𝑛

Since 𝑓 = 𝑓 𝜒 𝑁 𝑐 almost everywhere and 𝑓𝑛 = 𝑓𝑛 𝜒 𝑁 𝑐 almost everywhere for each 𝑛, we obtain the
required identity. 

2.4 Complex-valued functions and image measures


Definition 2.38. Let (𝑋, A) and (𝑌 , B) be measurable spaces. A function 𝑓 : 𝑋 → 𝑌 is called measurable
with respect to A and B if for each 𝐵 ∈ B, 𝑓 −1 (𝐵) ∈ A.

Remark 2.39.

1. In the sense of the previous definition, the composition of two measurable functions are measurable.

2. If 𝜎(B0 ) = B, then 𝑓 is measurable with respect to A and B if and only if for all 𝐵 ∈ B0 ,
𝑓 −1 (𝐵) ∈ A.

3. A real-valued function 𝑓 on (𝑋, A) is measurable if and only if it is measurable with respect to A


and B(R).

4. Define the Borel 𝜎-algebra B( [−∞, ∞]) on [−∞, ∞] as the collection of all subsets of [−∞, ∞]
of the form 𝐵 ∪ 𝐶, where 𝐵 is a Borel subsets of R and 𝐶 ⊆ {−∞, ∞}. Then, an extended-real-
valued function 𝑓 on (𝑋, A) is measurable if and only if 𝑓 is measurable with respect to A and
B([−∞, ∞]).

5. An R𝑑 -valued function 𝑓 = ( 𝑓1 , . . . , 𝑓 𝑑 ) on (𝑋, A) is said to be measurable if it is measurable with


respect to A and B(R𝑑 ). Then, 𝑓 is measurable if and only if each 𝑓𝑖 is measurable.

6. We say that a complex-valued function 𝑓 is measurable if it is measurable as an R2 -valued function.


(Note that B(C) = B(R2 ).) Thus, 𝑓 is measurable if and only if Re 𝑓 and Im 𝑓 are real-valued
measurable functions.

Definition 2.40. Let 𝑓 be a complex-valued measurable function on a measure space (𝑋, A, 𝜇). We say
that 𝑓 is integrable if both Re 𝑓 and Im 𝑓 are integrable. In this case, we define
∫ ∫ ∫
𝑓 𝑑𝜇 = Re 𝑓 𝑑𝜇 + 𝑖 Im 𝑓 𝑑𝜇.

22
Remark 2.41.

1. Let 𝑓 and 𝑔 be integrable complex-valued ∫ 𝛼 ∈ C. Then,


∫ functions on∫ 𝑋, and let ∫ both 𝑓 + 𝑔∫ and 𝛼 𝑓
are integrable, and the following hold: ( 𝑓 + 𝑔)𝑑𝜇 = 𝑓 𝑑𝜇 + 𝑔𝑑𝜇 and (𝛼 𝑓 )𝑑𝜇 = 𝛼 𝑓 𝑑𝜇.

2. A∫ complex-valued
∫ measurable function 𝑓 is integrable∫if and only if | 𝑓 | is integrable. In this case,
| 𝑓 𝑑𝜇| 6 | 𝑓 |𝑑𝜇. For a proof, use the polar form of 𝑓 𝑑𝜇.

3. The dominated convergence theorem holds for complex-valued functions.


We now define image measures.

Definition 2.42. Let (𝑋, A, 𝜇) be a measure space, and let (𝑌 , B) be a measurable space. Let 𝑓 : 𝑋 → 𝑌
be a measurable function. Define the image measure of 𝜇 under 𝑓 by

𝜇 𝑓 −1 (𝐵) B 𝜇( 𝑓 −1 (𝐵)), 𝐵 ∈ B. (2.2)

One can show that 𝜇 𝑓 −1 is indeed a measure on (𝑌 , B).

Proposition 2.43. Let 𝑔 be an extended-real-valued B-measurable function on 𝑌 . Then, 𝑔 is integrable


with respect to 𝜇 𝑓 −1 if and only if 𝑔 ◦ 𝑓 is integrable with respect to 𝜇. If 𝑔 is nonnegative or integrable,
then ∫ ∫
−1
𝑔𝑑 (𝜇 𝑓 ) = (𝑔 ◦ 𝑓 )𝑑𝜇.
𝑌 𝑋

Proof. Note that 𝑔 ◦ 𝑓 is measurable by the definition. We first consider characteristic functions. Let 𝐵
be a B-measurable subset of 𝑋, and let 𝑔 = 𝜒𝐵 . Observe that
(
1 if 𝑥 ∈ 𝑓 −1 (𝐵)
𝑔( 𝑓 (𝑥)) = 𝜒𝐵 ( 𝑓 (𝑥)) =
0 otherwise.

In other words, 𝑔 ◦ 𝑓 = 𝜒 𝑓 −1 (𝐵) . Thus,


∫ ∫
𝑔𝑑 (𝜇 𝑓 −1 ) = 𝜇 𝑓 −1 (𝐵) = 𝜇( 𝑓 −1 (𝐵)) = (𝑔 ◦ 𝑓 )𝑑𝜇.
𝑌 𝑋

By the additivy and homogeneity, the identity (2.2) holds for every simple B-measurable functions on
𝑌 . Also, by the approximation by simple functions and the monotone convergence theorem, the identity
(2.2) also holds for all nonnegative B-measurable functions. Now, for an arbitrary B-measurable function
𝑔 on 𝑌 , by the identity 𝑔 = 𝑔 + − 𝑔 − , 𝑔 is integrable with respect to 𝜇 𝑓 −1 if and only if 𝑔 ◦ 𝑓 is integrable
with respect to 𝜇. In this case, clearly, the identity (2.2) holds. 

2.5 Modes of convergence


Define L1 (𝑋, A, 𝜇, R) as the set of all real-valued integrable unctions on 𝑋. Note that this space L1 is a
vector space over R, and the integral is linear on this space.

Definition 2.44. Let ( 𝑓𝑛 ) be a sequence in L1 . We say that ( 𝑓𝑛 ) converges to 𝑓 in mean if



lim | 𝑓𝑛 − 𝑓 |𝑑𝜇 = 0.
𝑛

Remark 2.45.

1. Let ( 𝑓𝑛 ) be a sequence in L1 , and let 𝑓𝑛 → 𝑓 a.e., where 𝑓 is a real-valued measurable function.


Suppose there exists an integrable function 𝑔 such that | 𝑓𝑛 | 6 𝑔 a.e. Then, by the dominated
convergence theorem, 𝑓 ∈ L1 and 𝑓𝑛 → 𝑓 in mean.

23
2. Let ( 𝑓𝑛 ) be a sequence∫in L1 consisting
∫ of nonnegative functions, and let 𝑓 ∈ L1 . Suppose that
𝑓𝑛 → 𝑓 a.e. and that 𝑓𝑛 𝑑𝜇 → 𝑓 𝑑𝜇. Then, 𝑓𝑛 → 𝑓 in mean. To show this, apply Fatou’s
lemma to the sequence ( 𝑓𝑛 + 𝑓 − | 𝑓𝑛 − 𝑓 |).
Suppose 𝑓𝑛 → 𝑓 in mean. Then, by Chevyshev’s inequality, for every 𝜖 > 0,

1
𝜇({| 𝑓𝑛 − 𝑓 | > 𝜖 }) 6 | 𝑓𝑛 − 𝑓 |𝑑𝜇 → 0.
𝜖
Motivated by this observation, we define another mode of convergence.

Definition 2.46. Let ( 𝑓𝑛 ) be a sequence of real-valued measurable functions on (𝑋, A, 𝜇). ( 𝑓𝑛 ) is said
to converge to 𝑓 in measure if for every 𝜖 > 0,

lim 𝜇({| 𝑓𝑛 − 𝑓 | > 𝜖 }) = 0.


𝑛

Remark 2.47.

1. By the previous observation, if 𝑓𝑛 → 𝑓 in mean, then 𝑓𝑛 → 𝑓 in measure.

2. Neither almost everywhere convergence nor convergence in measure implies convergence in mean:
consider ( 𝑓𝑛 ) for the measure space (R, B(R), 𝜆), where 𝑓𝑛 is defined so that it has value 𝑛 on
[0, 1/𝑛] and 0 elsewhere

3. Almost everywhere convergence does not imply convergence in measure: consider ( 𝜒 [𝑛,∞) ) for the
measure space (R, B(R), 𝜆).

4. Convergence in measure also does not imply almost everywhere convergence: consider ( 𝑓𝑛 ) with
𝑓𝑛 = 𝜒 [ 𝑗/2𝑘 , ( 𝑗+1)/2𝑘 ) , where 𝑛 = 2 𝑘 + 𝑗 with 0 6 𝑗 < 2 𝑘 .

However, there are some useful results.

Proposition 2.48. Let 𝑓𝑛 → 𝑓 in measure. Then, there is a subsequence ( 𝑓𝑛𝑘 ) of ( 𝑓𝑛 ) which converges
to 𝑓 a.e.

Proof. Since 𝑓𝑛 → 𝑓 in measure, we can construct a subsequence ( 𝑓𝑛𝑘 ) of ( 𝑓𝑛 ) satisfying the following:
for each 𝑘,  
1 1
𝜇 | 𝑓 𝑛𝑘 − 𝑓 | > 6 𝑘.
𝑘 2
Define 𝐴 𝑘 B {| 𝑓𝑛𝑘 − 𝑓 | > 1/𝑘 }. Then, for each 𝑗,

∞ ∞ ∞
©Ø ª Õ Õ 1 1
𝜇­ 𝐴𝑘 ® 6 𝜇( 𝐴 𝑘 ) 6 𝑘
= 𝑗−1 < ∞.
2 2
« 𝑘= 𝑗 ¬ 𝑘= 𝑗 𝑘= 𝑗

Thus,
∞ ∞
©Ù Ø ª 1
𝜇­ 𝐴 𝑘 ® 6 lim 𝑗−1 = 0.
𝑗→∞ 2
« 𝑗=1 𝑘= 𝑗 ¬
Ñ∞ Ð∞
Moreover, if 𝑥 ∉ 𝑗=1 𝑘= 𝑗 𝐴 𝑘 , then 𝑥 ∉ 𝐴 𝑘 for only finitely many 𝐴 𝑘 ’s. This means that if 𝑥 ∉
Ñ∞ Ð∞
𝑗=1 𝑘= 𝑗 𝐴 𝑘 , then there exists 𝑁 ∈ N such that if 𝑘 > 𝑁, then | 𝑓𝑛𝑘 (𝑥) − 𝑓 (𝑥)| 6 1/𝑘. Therefore,
𝑓𝑛𝑘 → 𝑓 pointwise on ( ∞
Ñ Ð∞ Ñ∞ Ð∞
𝑘= 𝑗 𝐴 𝑘 ) . Since 𝜇( 𝑗=1 𝑘= 𝑗 𝐴 𝑘 ) = 0, 𝑓 𝑛 → 𝑓 almost everywhere. 
𝑐
𝑗=1

Corollary 2.49. If 𝑓𝑛 → 𝑓 in mean, then there exists a subsequence of 𝑓 which converges to 𝑓 a.e.

Proposition 2.50. Suppose that 𝜇 is a finite measure and that 𝑓𝑛 → 𝑓 a.e. Then, 𝑓𝑛 → 𝑓 in measure.

24
Proof. Note that 𝑓𝑛 (𝑥) 9 𝑓 (𝑥) if and only if there exists 𝜖 > 0 such that for every 𝑗 = 1, 2, . . . , there
exists 𝑛 > 𝑗 such that | 𝑓𝑛 (𝑥) − 𝑓 (𝑥)| > 𝜖, i.e. 𝑥 ∈ ∞
Ñ Ð
𝐴 , where 𝐴𝑛, 𝜖 = {| 𝑓𝑛 − 𝑓 | > 𝜖 }.
Ñ∞ Ð 𝑗=1 𝑛> 𝑗 𝑛, 𝜖
Now, let 𝜖 > 0 be given. Then, if 𝑥 ∈ 𝑗=1 𝑛> 𝑗 𝐴𝑛, 𝜖 , then 𝑓𝑛 (𝑥) 9 𝑓 (𝑥). Since 𝑓𝑛 → 𝑓 a.e.,


©Ù Ø
𝐴𝑛, 𝜖 ® = 0.
ª
𝜇­
« 𝑗=1 𝑛> 𝑗 ¬
This means that !
Ø
lim 𝜇 𝐴𝑛, 𝜖 = 0
𝑗→∞
𝑛> 𝑗

since 𝜇 is a finite measure. In particular, for each 𝑗,

lim 𝜇( 𝐴 𝑗, 𝜖 ) = lim 𝜇({| 𝑓 𝑗 − 𝑓 | > 𝜖 }) = 0.


𝑗→∞ 𝑗→∞

We can improve this proposition as in the following theorem.

Theorem 2.51 (Egorov’s theorem). Suppose that 𝜇 is a finite measure and that 𝑓𝑛 → 𝑓 a.e. Then, for
each 𝜖 > 0, there exists a measurable subset 𝐵 of 𝑋 such that 𝜇(𝐵 𝑐 ) < 𝜖 and 𝑓𝑛 → 𝑓 uniformly on 𝐵.

Remark 2.52.

1. A sequence ( 𝑓𝑛 ) is said to converge to 𝑓 almost uniformly if for each 𝜖 > 0, there exists a measurable
subset 𝐵 of 𝑋 such that 𝜇(𝐵 𝑐 ) < 𝜖 and 𝑓𝑛 → 𝑓 uniformly on 𝐵. Thus, Egorov’s theorem says that
if 𝜇 is finite, then a.e. convergence implies almost uniform convergence.

2. Almost uniform convergence implies a.e. convergence. To show this, suppose 𝑓𝑛 → 𝑓 almost
uniformly. Then, for each 𝑘, there exists a measurable set 𝐴 𝑘 such that 𝑓𝑛 → 𝑓 uniformly on 𝐴 𝑘
Ñ
and 𝜇( 𝐴𝑐𝑘 ) < 1/𝑘. Define 𝐸 = { 𝑓𝑛 9 𝑓 }. Then, 𝐸 ⊆ 𝐴𝑐𝑘 for all 𝑘, and so 𝐸 ⊆ 𝑘 𝐴𝑐𝑘 . Moreover,
Ñ 𝑐
𝜇( 𝑘 𝐴 𝑘 ) 6 1/𝑘 → 0. This completes the proof. Thus, if 𝜇 is finite, almost uniform convergence
is equivalent to a.e. convergence by Egorov’s theorem.

3. Almost uniform convergence implies convergence in measure. To show this, suppose 𝑓𝑛 → 𝑓


almost uniformly, and let 𝜖 > 0 be given. Then, for each 𝜂 > 0, there exists a measurable set 𝐴
such that 𝜇( 𝐴𝑐 ) < 𝜂 and 𝑓𝑛 → 𝑓 uniformly on 𝐴. Thus, there exists 𝑁 such that if 𝑛 > 𝑁, then
| 𝑓𝑛 (𝑥) − 𝑓 (𝑥)| 6 𝜖 for all 𝑥 ∈ 𝐴. This means that if 𝑛 > 𝑁, then {| 𝑓𝑛 − 𝑓 | > 𝜖 } ⊆ 𝐴𝑐 , and so
𝜇({| 𝑓𝑛 − 𝑓 | > 𝜖 }) < 𝜂. Thus, we have shown that for every 𝜂 > 0, there exists 𝑁 such that if
𝑛 > 𝑁, then 𝜇({| 𝑓𝑛 − 𝑓 | > 𝜖 }) < 𝜂. This means that lim𝑛→∞ 𝜇({| 𝑓𝑛 − 𝑓 | > 𝜖 }) = 0.

Proof. Recall that 𝑓𝑛 → 𝑓 uniformly on 𝐴 if and only if for every 𝑘, there exists 𝑗 = 𝑗 (𝑘) such that if
𝑛 > 𝑗, then | 𝑓𝑛 (𝑥) − 𝑓 (𝑥)| 6 1/𝑘 for all 𝑥 ∈ 𝐴. This means that for every 𝑘, there exists 𝑗 (𝑘) such that
Ù
𝐴 ⊆ 𝐴 𝑗 (𝑘),𝑘 B {| 𝑓𝑛 − 𝑓 | 6 1/𝑘 }.
𝑛> 𝑗 (𝑘)

Moreover, for any function 𝑗 : N → N, 𝑓𝑛 → 𝑓 uniformly on 𝐴, where 𝐴 = ∞


Ñ
Ð 𝑐 𝑘=1 𝐴 𝑗 (𝑘),𝑘 . Thus, for a
given 𝜖 > 0, we want to find a function 𝑗 : N → N so that 𝜇( 𝑘 𝐴 𝑗 (𝑘),𝑘 ) > 0.
Í
Let 𝜖 > 0 be given. By subadditivity, it suffices to find a function 𝑗 such that 𝑘 𝜇( 𝐴𝑐𝑗 (𝑘),𝑘 ) < 𝜖. Let
Ø
𝐸 𝑗,𝑘 = 𝐴𝑐𝑗,𝑘 = {| 𝑓𝑛 − 𝑓 | > 1/𝑘 }.
𝑛> 𝑗

25
Ð
Then, since 𝑓𝑛 → 𝑓 a.e., 𝜇( 𝑗 𝐸 𝑗,𝑘 ) = 0 for each 𝑘. Thus, for a fixed 𝑘, since 𝜇 is finite, by continuity
from above, lim 𝑗 𝜇(𝐸 𝑗,𝑘 ) = 0. Therefore, there exists 𝑗 (𝑘) such that 𝜇(𝐸 𝑗 (𝑘),𝑘 ) < 𝜖/2 𝑘+1 . It follows that
∞ ∞ ∞
Õ Õ Õ 𝜖 𝜖
𝜇( 𝐴𝑐𝑗 (𝑘),𝑘 ) = 𝜇(𝐸 𝑗 (𝑘),𝑘 ) 6 𝑘+1
= < 𝜖.
𝑘=1 𝑘=1 𝑘=1
2 2

This completes the proof. 

We can prove the dominated convergence theorem for convergence in measure.

Proposition 2.53. Let 𝑓𝑛 → 𝑓 in measure. Suppose there exists a nonnegative integrable function 𝑔 such
that | 𝑓𝑛 | 6 𝑔 almost everywhere. Then, 𝑓𝑛 → 𝑓 in mean.

Proof. Since 𝑓𝑛 → 𝑓 in measure, every subsequence of ( 𝑓𝑛 ) has a subsequence which converges to 𝑓


a.e., and so this converges to 𝑓 in mean by the dominated convergence theorem. Suppose now that ( 𝑓𝑛 )
∫does not converge to 𝑓 in mean. Then, there exist 𝜖 > 0 and a subsequence ( 𝑓𝑛𝑘 ) of ( 𝑓𝑛 ) such that
| 𝑓𝑛𝑘 − 𝑓 |𝑑𝜇 > 𝜖 for all 𝑘. Then, 𝑓𝑛𝑘 have no subsequence which converges to 𝑓 in mean, but this is a
contradiction. 

The following figure is a summary for relations between different modes of convergence we have
discussed.

𝜇 is finite
a.e. convergence convergence in measure
passing to a subsequence

𝜇 is finite
DCT DCT
(Egorov’s theorem)

almost uniform convergence convergence in mean

26
3 Lebesgue spaces

3.1 𝐿 𝑝 spaces: definition


We begin with a review of the notion of a Banach space.

Definition 3.1. Let 𝑉 be a vector space over R (or C). A function k · k : 𝑉 → R is a norm if the following
hold: for all 𝑢, 𝑣 ∈ 𝑉 and 𝛼 ∈ R (or C),

1. (nonnegativity) k𝑣k > 0

2. (positive definiteness) k𝑣k = 0 if and only if 𝑣 = 0

3. (homogeneity) k𝛼𝑣k = |𝛼|k𝑣k

4. (triangle inequality) k𝑢 + 𝑣k 6 k𝑢k + k𝑣k.

A function 𝑝 : 𝑉 → R with nonnegativity, homogeneity and triangle inequality is called a seminorm. A


vector space with a norm is called a norm (vector) space.
Moreover, if 𝑉 is a normed space with norm k · k, then this is a metric space:

𝑑 (𝑢, 𝑣) B k𝑢 − 𝑣k.

If 𝑉 is complete with respect to this distance function, then 𝑉 is called a Banach space.

Example 3.2.

1. By the usual norms, R𝑑 and C𝑑 are Banach spaces.

2. Define a norm on 𝐶 [𝑎, 𝑏] by


∫ 𝑏
k 𝑓 k1 B | 𝑓 |𝑑𝜆.
𝑎
Then, 𝐶 [𝑎, 𝑏] with this norm is a normed space but is not a Banach space: for 𝑎 = −1 and 𝑏 = 1,
consider the sequence ( 𝑓𝑛 ) given by



 0 if −1 6 𝑥 6 0


𝑓𝑛 (𝑥) = 𝑛𝑥 if 0 < 𝑥 6 1/𝑛

if 1/𝑛 < 𝑥 6 1.

1

3. Define a norm on 𝐶 [𝑎, 𝑏] by


k 𝑓 k ∞ B sup | 𝑓 (𝑥)|.
𝑥 ∈ [𝑎,𝑏]

Then, 𝐶 [𝑎, 𝑏] with this norm is a Banach space: the convergence is same to uniform convergence,
and any uniformly Cauchy sequence in 𝐶 [𝑎, 𝑏] converges uniformly to a continuous function on
[𝑎, 𝑏].

Lemma 3.3. Let (𝑣 𝑛 ) be a Cauchy sequence in a normed space 𝑉. Suppose that (𝑣 𝑛 ) has a subsequence
(𝑣 𝑛𝑘 ) such that 𝑣 𝑛𝑘 → 𝑣 ∈ 𝑉. Then, 𝑣 𝑛 → 𝑣.

Proof. Observe that


k𝑣 𝑛 − 𝑣k 6 k𝑣 𝑛 − 𝑣 𝑛𝑘 k + k𝑣 𝑛𝑘 − 𝑣k.
Now, let 𝜖 > 0 be given. Then, since (𝑣 𝑛 ) is Cauchy and 𝑣 𝑛𝑘 → 𝑣, there exists 𝑁 such that the
following hold:

27
• if 𝑛, 𝑚 > 𝑁, then k𝑣 𝑛 − 𝑣 𝑚 k < 𝜖, and

• if 𝑘 > 𝑁, then k𝑣 𝑛𝑘 − 𝑣k < 𝜖.

Let 𝑛 > 𝑁, and choose some 𝑘 > 𝑁. Then, 𝑛 𝑘 > 𝑁, so k𝑣 𝑛 − 𝑣k < 2𝜖 by the previous observations. This
completes the proof. 

A series ∞
Í Í∞
𝑘=1 𝑣 𝑘 in a normed space 𝑉 is said to converge absolutely if 𝑘=1 k𝑣 𝑘 k < ∞.

Proposition 3.4. A normed space 𝑉 is complete if and only if every absolutely convergent series in 𝑉 is
convergent.
Í Í
Proof. Suppose 𝑉 is a Banach space, and suppose 𝑘 𝑣 𝑘 is absolutely convergent. Define 𝑠 𝑛 B 𝑛𝑘=1 𝑣 𝑘 .
Then, for 𝑚 < 𝑛,
Õ𝑛 Õ 𝑛
k𝑠 𝑛 − 𝑠 𝑚 k = 𝑣𝑘 6 k𝑣 𝑘 k.
𝑘=𝑚+1 𝑘=𝑚+1
Í Í
Since 𝑘 k𝑣 𝑘 k is convergent, (𝑠 𝑛 ) is Cauchy in 𝑉. Since 𝑉 is complete, (𝑠 𝑛 ) is convergent, that is, 𝑘 𝑣 𝑘
is convergent.
To show the converse, suppose every convergent series in 𝑉 is convergent. Let (𝑢 𝑛 ) be a Cauchy
sequence in 𝑉. By the lemma, it suffices to show that (𝑢 𝑛 ) has a convergent subsequence. Since (𝑢 𝑛 ) is
Cauchy, there exists a subsequence (𝑢 𝑛𝑘 ) of (𝑢 𝑛 ) such that k𝑢 𝑛𝑘+1 − 𝑢 𝑛𝑘 k < 1/2 𝑘+1 for each 𝑘. Define
𝑣 1 = 𝑢 𝑛1 and 𝑣 𝑘 = 𝑢 𝑛𝑘 − 𝑢 𝑛𝑘−1 for 𝑘 > 1. Then, since k𝑣 𝑘 k < 1/2 𝑘 for 𝑘 > 1, the series ∞
Í
Í∞ Í𝑛 𝑘=1 𝑣 𝑘
converges absolutely. By the assumption, 𝑘=1 𝑣 𝑘 converges. Now, observe that 𝑘=1 𝑣 𝑘 = 𝑢 𝑛𝑘 . Thus,
(𝑢 𝑛𝑘 ) converges. This completes the proof. 

We now define Lebesgue spaces.

Definition 3.5 (Lebesgue spaces). Let (𝑋, A, 𝜇) be a measure space, and let 𝑝 ∈ [1, ∞). Define
L 𝑝 (𝑋, A, 𝜇, R) (or simply L 𝑝 (𝑋, R)) as the set of all real-valued measurable functions 𝑓 on 𝑋 such that
| 𝑓 | 𝑝 is integrable. Similarly, we define L 𝑝 (𝑋, A, 𝜇, C) (or simply L 𝑝 (𝑋, C)).
Moreover, a real-valued (or complex-valued) measurable function 𝑓 is said to be essentially bounded
if there exists 𝑀 > 0 such that | 𝑓 | 6 𝑀 almost everywhere. For 𝑝 = ∞, L∞ (𝑋, A, 𝜇, R) is the set of all
real-valued essentially bounded function on 𝑋. Likewise, L∞ (𝑋, A, 𝜇, C) is the set of all complex-valued
essentially bounded functions on 𝑋.
We will use L 𝑝 (𝑋, A, 𝜇) (or simply L 𝑝 ) to denote either L 𝑝 (𝑋, A, 𝜇, R) or L 𝑝 (𝑋, A, 𝜇, C).

Remark 3.6.

1. L∞ is a vector space over R (or C).

2. For 𝑝 ∈ [1, ∞), L 𝑝 is a vector space over R (or C): just observe that

| 𝑓 + 𝑔| 𝑝 6 (| 𝑓 | + |𝑔|) 𝑝 6 (2(| 𝑓 | ∨ |𝑔|)) 𝑝 6 2 𝑝 | 𝑓 | 𝑝 + 2 𝑝 |𝑔| 𝑝 .

Definition 3.7. Define a nonnegative function k · k 𝑝 on L 𝑝 as follows:


 ∫  1/ 𝑝
 | 𝑓 | 𝑝 𝑑𝜇

 if 𝑝 ∈ [1, ∞)
k 𝑓 k𝑝 B
 ess sup| 𝑓 |
 if 𝑝 = ∞,

where ess sup 𝑔 is the essential supremum of 𝑔 over 𝑋 defined by

ess sup 𝑔 B inf{𝑀 > 0 | 𝑔 6 𝑀 almost everywhere}.

Now, we show that k · k 𝑝 is a seminorm on L 𝑝 .

28
Lemma 3.8. Let 𝜑 be a convex function on an open interval 𝐼 in R. Then, 𝜑 is continuous.
Proof. Recall that 𝜑 : 𝐼 → R is convex if and only if for all 𝑡 ∈ (0, 1) and 𝑥, 𝑦 ∈ 𝐼,

𝜑((1 − 𝑡)𝑥 + 𝑡𝑦) 6 (1 − 𝑡)𝜑(𝑥) + 𝑡𝜑(𝑦).

By elementary geometry, one can show that if 𝜑 : 𝐼 → R is convex, then for every triple 𝑥 < 𝑦 < 𝑧 of
points in 𝐼,
𝜑(𝑦) − 𝜑(𝑥) 𝜑(𝑧) − 𝜑(𝑥) 𝜑(𝑧) − 𝜑(𝑦)
6 6 .
𝑦−𝑥 𝑧−𝑥 𝑧−𝑦
This means that if for 𝑥 < 𝑦 ∈ 𝐼, 𝑠(𝑥, 𝑦) denotes the slope of the line segment joining (𝑥, 𝜑(𝑥)) and
(𝑦, 𝜑(𝑦)), then
𝑠(𝑥, 𝑦) 6 𝑠(𝑧, 𝑥) 6 𝑠(𝑧, 𝑦). (3.1)
Now, fix 𝑥 ∈ 𝐼, and take 𝑎, 𝑏 ∈ 𝐼 with 𝑎 < 𝑥 < 𝑏. This is possible since 𝐼 is an open interval. In addition,
let 𝑦 ∈ (𝑥, 𝑏). Then, by the inequalities (3.1), we have the following inequalities:

𝑠(𝑎, 𝑥) 6 𝑠(𝑎, 𝑦) 6 𝑠(𝑥, 𝑦) 6 𝑠(𝑥, 𝑏).

Therefore,
𝑠(𝑎, 𝑥) (𝑦 − 𝑥) 6 𝜑(𝑦) − 𝜑(𝑥) 6 𝑠(𝑥, 𝑏) (𝑦 − 𝑥).
Hence, letting 𝑦 → 𝑥, we have that 𝜑(𝑦) → 𝜑(𝑥).
Similarly, if 𝑦 ∈ (𝑎, 𝑥), then

𝑠(𝑎, 𝑥) 6 𝑠(𝑦, 𝑥) 6 𝑠(𝑦, 𝑏) 6 𝑠(𝑥, 𝑏),

so
𝑠(𝑎, 𝑥) (𝑦 − 𝑥) > 𝜑(𝑦) − 𝜑(𝑥) > 𝑠(𝑥, 𝑏) (𝑦 − 𝑥).
Thus, by letting 𝑦 → 𝑥, we obtain again that 𝜑(𝑦) → 𝜑(𝑥). 
Remark 3.9. Let 𝜑 : 𝐼 → R be a convex function. By induction, one can prove the following: if (𝑡𝑖 )𝑖=1 𝑁

be a finite set of positive numbers such that 𝑡1 + · · · + 𝑡 𝑁 = 1, then for any finite set (𝑥 𝑖 )𝑖=1 of points in 𝐼,
𝑁

𝑁
! 𝑁
Õ Õ
𝜑 𝑡𝑖 𝑥𝑖 6 𝑡𝑖 𝜑(𝑥 𝑖 ).
𝑖=1 𝑖=1

Note that for such a family (𝑡𝑖 ) of positive numbers, we can define a measure 𝜇 on the set 𝑋 = {1, . . . , 𝑁 }
by Õ
𝜇( 𝐴) B 𝑡𝑖 , 𝐴 ∈ P (𝑋).
𝑖∈𝐴
Then, 𝜇(𝑋) = 1, and a finite family (𝑥 𝑖 ) ⊆ 𝐼 corresponds to a measurable function 𝑓 : 𝑋 = {1, . . . , 𝑁 } →
𝐼. In this case,
Õ𝑁 ∫
𝑡𝑖 𝑥𝑖 = 𝑓 𝑑𝜇
𝑖=1 𝑋
𝑁
Õ ∫
𝑡𝑖 𝜑(𝑥 𝑖 ) = 𝜑 ◦ 𝑓 𝑑𝜇.
𝑖=1 𝑋

Thus, the previous inequality can be read as the following:


∫  ∫
𝜑 𝑓 𝑑𝜇 6 𝜑 ◦ 𝑓 𝑑𝜇. (3.2)
𝑋 𝑋

We can generalize this inequality to every finite measure space (𝑋, A, 𝜇) with 𝜇(𝑋) = 1.

29
Proposition 3.10 (Jensen’s inequality). Let 𝜇 be a finite measure on (𝑋, A) with 𝜇(𝑋) = 1, and let
𝜑 : (𝑎, 𝑏) → R be a convex function. If 𝑓 : 𝑋 → R is an integrable function such that the image of 𝑓 is
contained in (𝑎, 𝑏), then the inequality (3.2) holds.

Proof. Let 𝑡 = 𝑋 𝑓 𝑑𝜇 ∈ R. Then, since 𝑎 < 𝑓 < 𝑏, we have 𝑡 ∈ (𝑎, 𝑏). Recall now the following
property of convex functions.

• If 𝜑 : (𝑎, 𝑏) → R is convex, then for every 𝑐 ∈ (𝑎, 𝑏), there exists 𝑚 ∈ R such that

𝜑(𝑠) > 𝑚(𝑠 − 𝑐) + 𝜑(𝑐).

Here, the line represented by the function 𝑠 ↦→ 𝑚(𝑠 − 𝑐) + 𝜑(𝑐) is called a supporting line of 𝜑 at 𝑐.

Consider now a supporting line 𝑚(𝑠 − 𝑡) + 𝜑(𝑡) at 𝑡. Then, since 𝑎 < 𝑓 < 𝑏, for every 𝑥 ∈ 𝑋,

𝜑( 𝑓 (𝑥)) > 𝑚( 𝑓 (𝑥) − 𝑡) + 𝜑(𝑡).

The right-hand side is integrable in 𝑥 with the integral 𝜑(𝑡). Moreover, 𝜑 ◦ 𝑓 is measurable since 𝜑 is
continuous. Now, use the following fact.

• Let 𝑔 and
∫ ℎ be two extended-real-valued
∫ measurable on 𝑋 satisfying that 𝑔 > ℎ and ℎ is integrable.
Then, 𝑋 𝑔𝑑𝜇 exists, and 𝑋 𝑔𝑑𝜇 is either a real number or ∞.
− − − −
∫To show this fact, observe that 𝑔 > ℎ implies −𝑔 > −ℎ , that is. 𝑔 6 ℎ . Thus, since ℎ is integrable,
𝑋
𝑔 − 𝑑𝜇 < ∞. 

Example 3.11. Clearly, the exponential function exp : (−∞, ∞) → R is convex. Therefore, in this case,
Jensen’s inequality can be written as
∫  ∫
exp 𝑓 𝑑𝜇 6 𝑒 𝑓 𝑑𝜇.
𝑋 𝑋

In particular, if 𝑔 is a strictly positive function, then we can consider the measurable function log 𝑔, so
we have the inequalify of arithmetic means and geometric means:
∫  ∫
exp log 𝑔𝑑𝜇 6 𝑔𝑑𝜇.
𝑋 𝑋

Indeed, if 𝑋 = {1, . . . , 𝑁 }, and if 𝜇 is defined by

| 𝐴|
𝜇( 𝐴) B , 𝐴 ⊆ 𝑋,
𝑁
then for a positive function 𝑔 : 𝑋 → (0, ∞), the inequality of arithmetic means and geometric means
becomes !
𝑁 𝑁
p
𝑁 1 Õ 1 Õ
𝑔(1) . . . 𝑔(𝑁) = exp log 𝑔(𝑖) 6 𝑔(𝑖),
𝑁 𝑖=1 𝑁 𝑖=1

which is the classical version of the inequality of arithmetic means and geometric means.

Lemma 3.12. Let 1 < 𝑝, 𝑞 < ∞ with 1/𝑝 + 1/𝑞 = 1. Then, for 𝑥, 𝑦 > 0,

𝑥 𝑝 𝑦𝑞
𝑥𝑦 6 + .
𝑝 𝑞

30
Proof. We will use the general AM-GM inequality. Define a measure 𝜇 on 𝑋 = {1, 2} by

1 1
𝜇({1}) = and 𝜇({2}) = .
𝑝 𝑞

Then, for 𝑠, 𝑡 > 0, by the AM-GM inequality, we have


1 1 𝑠 𝑡
𝑠𝑝𝑡𝑞 6 + .
𝑝 𝑞

Now, for 𝑥, 𝑦 > 0, let 𝑠 = 𝑥 𝑝 and 𝑡 = 𝑦 𝑞 . Then, we obtain that

𝑥 𝑝 𝑦𝑞
𝑥𝑦 6 + .
𝑝 𝑞

Note that this inequality holds also when 𝑥 = 0 or 𝑦 = 0. 

Proposition 3.13 (Hölder’s inequality). Let (𝑋, A, 𝜇) be a measure space, and let 𝑝, 𝑞 ∈ [1, ∞] satisfy
1/𝑝 + 1/𝑞 = 1.2 If 𝑓 ∈ L 𝑝 and 𝑔 ∈ L𝑞 , then 𝑓 𝑔 ∈ L1 , and the following holds:

k 𝑓 𝑔k 1 6 k 𝑓 k 𝑝 k𝑔k 𝑞 .

Proof. Let 1 < 𝑝, 𝑞 < ∞. Then, by the previous lemma,

1 𝑝 1 𝑞
| 𝑓 𝑔| 6 | 𝑓 | + |𝑔| .
𝑝 𝑞

Thus, if k 𝑓 k 𝑝 = 1 and k𝑔k 𝑞 = 1, then

1 1
k 𝑓 𝑔k 1 6 + = 1.
𝑝 𝑞

In general, if both k 𝑓 k 𝑝 and k𝑔k 𝑞 are nonzero, then consider 𝑓 /k 𝑓 k 𝑝 and 𝑔/k𝑔k 𝑞 . In this case, we have

k 𝑓 𝑔k 1
6 1,
k 𝑓 k 𝑝 k𝑔k 𝑞

that is,
k 𝑓 𝑔k 1 6 k 𝑓 k 𝑝 k 𝑓 k 𝑞 .
Note that this inequality also holds when either k 𝑓 k 𝑝 = 0 or k𝑔k 𝑞 = 0.
Now, assume that 𝑝 = 1 and 𝑞 = ∞. Note that

| 𝑓 𝑔| 6 k𝑔k ∞ | 𝑓 |

almost everywhere. Thus, by integrating both sides, we obtain

k 𝑓 𝑔k 1 6 k 𝑓 k 1 k𝑔k ∞ .

This completes the proof. 

Proposition 3.14 (Minkowski’s inequality). Let 𝑝 ∈ [1, ∞]. If 𝑓 , 𝑔 ∈ L 𝑝 , then 𝑓 + 𝑔 ∈ L 𝑝 , and the
following holds:
k 𝑓 + 𝑔k 𝑝 6 k 𝑓 k 𝑝 + k𝑔k 𝑞 .
2 If 𝑝 = 1, then 𝑞 = ∞, and if 𝑝 = ∞, then 𝑞 = 1.

31
Proof. First, we assume that 𝑝 = ∞. Then, | 𝑓 + 𝑔| 6 | 𝑓 | + |𝑔| 6 k 𝑓 k ∞ + k𝑔k ∞ almost everywhere. Thus,
we obtain the inequality for the case 𝑝 = ∞. Moreover, suppose that 𝑝 = 1. Then, since | 𝑓 + 𝑔| 6 | 𝑓 | + |𝑔|,
by integraing both sides, we have the required inequality for the case 𝑝 = 1.
Now, let 1 < 𝑝 < ∞. Observe that if 1/𝑝 + 1/𝑞 = 1, then 𝑞 = 𝑝/( 𝑝 − 1), so
 𝑞
| 𝑓 + 𝑔| 𝑝−1 = | 𝑓 + 𝑔| 𝑝 .

Note that we already know that 𝑓 + 𝑔 ∈ L 𝑝 , so the previous identity shows that | 𝑓 + 𝑔| 𝑝−1 ∈ L𝑞 . Now,
by Hölder’s inequality,
∫ ∫
𝑝
| 𝑓 + 𝑔| 𝑑𝜇 6 (| 𝑓 | + |𝑔|)| 𝑓 + 𝑔| 𝑝−1 𝑑𝜇
∫  1/𝑞 ∫  1/𝑞
𝑝 𝑝
6 k 𝑓 k𝑝 | 𝑓 + 𝑔| 𝑑𝜇 + k𝑔k 𝑝 | 𝑓 + 𝑔| 𝑑𝜇
∫  1/𝑞
𝑝
= (k 𝑓 k 𝑝 + k𝑔k 𝑝 ) | 𝑓 + 𝑔| 𝑑𝜇 .

Here, we have used Hölder’s inequality in the second line. In the case that | 𝑓 + 𝑔| 𝑝 𝑑𝜇 > 0, then we
have ∫  1−1/𝑞
𝑝
k 𝑓 + 𝑔k 𝑝 = | 𝑓 + 𝑔| 𝑑𝜇 6 k 𝑓 k 𝑝 + k𝑔k 𝑝 .

Moreover, if | 𝑓 + 𝑔| 𝑝 𝑑𝜇 = 0, then the required inequality holds trivially. 
Corollary 3.15. For 𝑝 ∈ [1, ∞], k · k 𝑝 is a seminorm on L 𝑝 .
Remark 3.16 (Embeddings).
1. Consider (N, P (N), 𝜇), where 𝜇 is the counting measure. Suppose 1 6 𝑝 1 < 𝑝 2 6 ∞. Now, recall
that if ∞
Í
𝑛=1 𝑎 𝑛 converges, then lim 𝑎 𝑛 = 0. Thus, we can immediately obtain that L (N) ⊆ L (N)
𝑝1 𝑝2

when 𝑝 2 = ∞. Now assume 𝑝 2 < ∞, and let (𝑎 𝑛 ) ∈ L (N). Sincce |𝑎 𝑛 | → 0, there exists 𝑁 > 0
𝑝 1

such that |𝑎| 𝑛 6 1 for all 𝑛 > 𝑁. Thus, for 𝑛 > 𝑁,


|𝑎 𝑛 | 𝑝2 6 |𝑎 𝑛 | 𝑝1
Í Í
since 1 6 𝑝 1 < 𝑝 2 , so |𝑎 𝑛 | 𝑝2 < ∞ since |𝑎 𝑛 | 𝑝1 < ∞. Therefore, L 𝑝1 (N) ⊆ L 𝑝2 (N) when
𝑝 2 < ∞. For an explicit estimate between norms, see [Fol99, Proposition 6.11].
2. Let 𝜇 be a measure on (𝑋, A) with 𝜇(𝑋) = 1. If 𝑓 ∈ L∞ (𝑋), then since | 𝑓 | 𝑝 6 k 𝑓 k ∞ 𝑝
for
all 𝑝 ∈ [1, ∞), k 𝑓 k 𝑝 6 k 𝑓 k ∞ . Moreover, if 1 6 𝑝 1 < 𝑝 2 < ∞, then since 𝑝 2 /𝑝 1 > 1,
𝑝 1 /𝑝 2 = 1/( 𝑝 2 /𝑝 1 ) < 1. Thus, for 𝑞 ∈ (1, ∞) with 1/𝑞 + 𝑝 1 /𝑝 2 = 1, by Hölder’s inequality,
∫ ∫ ∫  𝑝1 / 𝑝2 ∫  𝑝1 / 𝑝2
𝑝1 𝑝1 𝑝2 1/𝑞 𝑝2
| 𝑓 | 𝑑𝜇 = | 𝑓 | · 1𝑑𝜇 6 | 𝑓 | 𝑑𝜇 𝜇(𝑋) = | 𝑓 | 𝑑𝜇 .

This shows that for 𝑝 1 , 𝑝 2 ∈ [1, ∞] with 𝑝 1 < 𝑝 2 , the following hold: L 𝑝2 (𝑋) ⊆ L 𝑝1 (𝑋) and
k 𝑓 k 𝑝1 6 k 𝑓 k 𝑝2 .
We now construct the normed space 𝐿 𝑝 from the seminormed space L 𝑝 .
Lemma 3.17. Let 𝑉 be a vector space over R (or C) with a seminorm 𝑝. Then,
𝑁 = ker 𝑝 B {𝑣 ∈ 𝑉 | 𝑝(𝑣) = 0}
is a linear subspace of 𝑉, and the quotient space 𝑉/𝑁 has the norm
k𝑣 + 𝑁 k B 𝑝(𝑣), 𝑣 ∈ 𝑉.

32
Proof. By the triangle inequality and the homogeneity of 𝑝, 𝑁 is a subspace of 𝑉. Thus, we can define
the quotient space 𝑉/𝑁. Now, we need to verify that k𝑣 + 𝑁 k is well-defined for 𝑣 ∈ 𝑉. Let 𝑣 − 𝑤 ∈ 𝑁,
then we can write 𝑣 = 𝑤 + 𝑛. Then, by the triangle inequality,

𝑝(𝑣) = 𝑝(𝑤 + 𝑛) 6 𝑝(𝑤) + 𝑝(𝑛) = 𝑝(𝑤),

and likewise,
𝑝(𝑤) = 𝑝(𝑣 − 𝑛) 6 𝑝(𝑣) + 𝑝(𝑛) = 𝑝(𝑣).
Thus, 𝑝(𝑤) = 𝑝(𝑣), so the function k · k on 𝑉/𝑁 is well-defined.
Moreover, the triangle inequality and the homogeneity is clear for k · k. Thus, it suffices to show that
k · k is positive definite. Indeed, if k𝑣 + 𝑁 k = 𝑝(𝑣) = 0, then 𝑣 ∈ 𝑁, so 𝑣 + 𝑁 = 𝑁. This completed the
proof. 

Let (𝑋, A, 𝜇) be a measure space, and let 1 6 𝑝 6 ∞. Then, N 𝑝 = kerk · k 𝑝 is the space of
measurable functions that vanishes almost everywhere. Thus, by forming the quotient space, which is
called the 𝐿 𝑝 space with respect to (𝑋, A, 𝜇),

𝐿 𝑝 (𝑋, A, 𝜇) B L 𝑝 /N 𝑝 ,

we identify every pair of functions which agree almost everywhere. For the induced norm on 𝐿 𝑝 (𝑋) is
now called the 𝐿 𝑝 -norm with respect to the measure space (𝑋, A, 𝜇). Note that we usually regard an
equivalence class 𝑓 + N 𝑝 as a function using some representative.
In the next chapter, we will discuss some basic properties of 𝐿 𝑝 spaces as metric spaces.

3.2 𝐿 𝑝 spaces: properties


We begin with this section by showing that 𝐿 𝑝 spaces are complete metric spaces.

Theorem 3.18 (Riesz–Fischer). Let (𝑋, A, 𝜇) be a measure space. Then, for every 𝑝 ∈ [1, ∞],
𝐿 𝑝 (𝑋, A, 𝜇) is complete with respect to the norm k · k 𝑝 .
Í
Proof. Assume 1 6 𝑝 < ∞, and let ( 𝑓 𝑘 ) be a sequence in 𝐿 𝑝 such that 𝑘 k 𝑓 𝑘 k 𝑝 < ∞. We want to show
Í Í
that the series 𝑘 𝑓 𝑘 is convergent. To obtain 𝑘 𝑓 𝑘 as a pointwise limit first, consider

Õ
𝑔(𝑥) = | 𝑓 𝑘 (𝑥)|, 𝑥 ∈ 𝑋.
𝑘=1

By Minkowski’s inequality, for each 𝑛,


𝑛
Õ 𝑛
Õ
| 𝑓𝑘 | 6 k 𝑓𝑘 k 𝑝 ,
𝑘=1 𝑝 𝑘=1

so we have !𝑝 !𝑝
∫ 𝑛
Õ 𝑛
Õ
| 𝑓𝑘 | 𝑑𝜇 6 k 𝑓𝑘 k 𝑝 .
𝑘=1 𝑘=1

By the monotone convergence theorem,


∫ ∞
!𝑝
Õ
𝑔 𝑝 𝑑𝜇 6 k 𝑓𝑘 k 𝑝 < ∞.
𝑘=1

33
Therefore, 𝑔 𝑝 is integrable, and so 𝑔 is finite almost everywhere. Thus, for almost all 𝑥 ∈ 𝑋, the series
Í
𝑘 𝑓 𝑘 (𝑥) converges absolutely in R, so converges in R. Therefore, we can define a real-valued measurable
function 𝑓 on 𝑋 by (Í

𝑘=1 𝑓 𝑘 (𝑥) if 𝑔(𝑥) < ∞
𝑓 (𝑥) =
0 otherwise.
Í
Note that | 𝑓 | 6 𝑔, and since 𝑔 ∈ 𝐿 𝑝 , 𝑓 ∈ 𝐿 𝑝 . We now want to show that 𝑘 𝑓 𝑘 = 𝑓 in 𝐿 𝑝 . Indeed, for
almost all 𝑥 ∈ 𝑋,
𝑛
Õ
𝑓 𝑘 (𝑥) − 𝑓 (𝑥) → 0 and
𝑘=1
𝑛 𝑝
Õ
𝑓 𝑘 (𝑥) − 𝑓 (𝑥) 6 𝑔(𝑥) 𝑝 .
𝑘=1

Since 𝑔 𝑝 is integrable, by the dominated convergence theorem,


𝑛
∫ Õ 𝑝

𝑓𝑘 − 𝑓 𝑑𝜇 → 0,
𝑘=1

that is,
𝑛
Õ
𝑓𝑘 − 𝑓 → 0.
𝑘=1 𝑝

Now, assume that 𝑝 = ∞, and let ( 𝑓 𝑘 ) be a sequence in 𝐿 ∞ such that


Í
𝑘 k 𝑓𝑘 k∞ < ∞. As the case of
1 6 𝑝 < ∞, we first show that
Õ∞
| 𝑓 𝑘 (𝑥)| < ∞
𝑘=1

for almost all 𝑥 ∈ 𝑋. However, this is easy to show: for each 𝑘, let 𝑁 𝑘 B {𝑥 ∈ 𝑋 | | 𝑓 𝑘 (𝑥)| > k 𝑓 𝑘 k ∞ },
and let 𝑁 B ∞
Ð
Í∞ 𝑘=1 𝑁 𝑘 , then 𝜇(𝑁) = 0. Note that for 𝑥 ∈ 𝑋 \ 𝑁 and for every 𝑘, | 𝑓 𝑘 (𝑥)| 6 k 𝑓 𝑘 k ∞ , so
Í∞
|
𝑘=1 𝑘 𝑓 (𝑥)| 6 𝑘=1 k 𝑓 𝑘 k ∞ < ∞. Thus, define a real-valued measurable function 𝑓 by

𝑘 𝑓 𝑘 (𝑥) if 𝑥 ∈ 𝑋 \ 𝑁
𝑓 (𝑥) B
0 if 𝑥 ∈ 𝑁.
Í∞
< ∞, so 𝑓 ∈ 𝐿 ∞ , and since | 𝑓 −
Í Í𝑛
Then, | 𝑓 | 6 𝑘 k 𝑓𝑘 k∞ 𝑘=1 𝑓 𝑘 | 6 𝑘=𝑛+1 k 𝑓 𝑘 k ∞ almost everywhere,
𝑛
Õ ∞
Õ
𝑓− 𝑓𝑘 6 k 𝑓 𝑘 k ∞ → 0.
𝑘=1 ∞ 𝑘=𝑛+1

𝑓 𝑘 = 𝑓 in 𝐿 ∞ .
Í
Therefore, 𝑘 

Remark 3.19. Let ( 𝑓𝑛 ) be a Cauchy sequence in 𝐿 𝑝 , where 1 6 𝑝 6 ∞. Then, by the previous theorem,
𝑓𝑛 → 𝑓 in 𝐿 𝑝 for some 𝑓 ∈ 𝐿 𝑝 . Thus, for 𝜖 > 0, if 1 6 𝑝 < ∞, then

1
𝜇{| 𝑓𝑛 − 𝑓 | > 𝜖 } 6 𝑝 | 𝑓𝑛 − 𝑓 |𝑑𝜇 → 0;
𝜖
if 𝑝 = ∞, then by taking 𝑁 and 𝐸 with 𝜇(𝐸) = 0 so that 𝑛 > 𝑁 implies | 𝑓𝑛 (𝑥) − 𝑓 (𝑥)| 6 𝜖 for all
𝑥 ∈ 𝑋 \ 𝐸, we obtain that for 𝑛 > 𝑁,

𝜇{| 𝑓𝑛 − 𝑓 | > 𝜖 } 6 𝜇(𝐸) = 0.

34
This means that 𝐿 𝑝 -convergence implies convergence in measure. In particular, every Cauchy sequence
in 𝐿 𝑝 has a subsequence which converges almost everywhere ("passing to a subsequence").
Í
On the other hand, the proof of the previous theorem shows that if 𝑘 𝑔 𝑘 converges absolutely in 𝐿 𝑝 ,
Í
then 𝑘 𝑔 𝑘 converges both in 𝐿 𝑝 and almost everywhere to the same function. This provides another
proof of the technique so-called "passing to a subsequence". Indeed, since ( 𝑓𝑛 ) is Cauchy in 𝐿 𝑝 , we can
take a subsequence ( 𝑓𝑛𝑘 ) so that
1
k 𝑓𝑛𝑘+1 − 𝑓𝑛𝑘 k < 𝑘 .
2
Í Í
Then, since the geometric series 𝑘 1/2 converges, the series 𝑓𝑛1 + 𝑘 ( 𝑓𝑛𝑘+1 − 𝑓𝑛𝑘 ) converges absolutely,
𝑘

and so this series converges in both 𝐿 𝑝 and almost everywhere to the same function 𝑓 . Note that the limit
function 𝑓 coincides with the a.e.-limit of the subsequence ( 𝑓𝑛𝑘 ), and 𝑓𝑛 → 𝑓 in 𝐿 𝑝 .
We now consider a dense subset of an 𝐿 𝑝 space.
Proposition 3.20. Let 1 6 𝑝 6 ∞. Then, the set of simple functions in 𝐿 𝑝 forms a dense subspace of 𝐿 𝑝 .
Proof. First, suppose 1 6 𝑝 < ∞. Let 𝑓 ∈ 𝐿 𝑝 with 𝑓 > 0. Then, there is a increasing sequence (𝑔 𝑘 ) of
nonnegative simple functions on 𝑋 such that 𝑔 𝑘 → 𝑓 pointwise. Then, since |𝑔 𝑘 | 6 𝑓 , so 𝑔 𝑘 ∈ 𝐿 𝑝 for
each 𝑘. Moreover, lim 𝑘 | 𝑓 𝑘 − 𝑓 | = 0, and by the construction of 𝑔 𝑘 , |𝑔 𝑘 − 𝑓 | 6 𝑓 . Thus, by the dominated
convergence theorem, 𝑔 𝑘 → 𝑓 in 𝐿 𝑝 . The general case follows from this observation.
Now, suppose 𝑝 = ∞, and let 𝑓 ∈ 𝐿 ∞ . Let 𝜖 > 0 be given, and choose a real numbers

𝑎0 < 𝑎1 < · · · < 𝑎 𝑛

such that the intervals (𝑎 𝑘−1 , 𝑎 𝑘 ] cover the interval [0, | 𝑓 | ∞ ] and 𝑎 𝑘 − 𝑎 𝑘−1 < 𝜖 for each 𝑘. Let
𝐴 𝑘 B 𝑓 −1 ((𝑎 𝑘−1 , 𝑎 𝑘 ]) for each 𝑘. Now define
𝑛
Õ
𝑔B 𝑎 𝑘 𝜒 𝐴𝑘 .
𝑘=1

Then, 𝑔 is a simple function with k 𝑓 − 𝑔k ∞ 6 𝜖. Similarly to the previous case, the general case also
follows. Thus, the proof is complete. 
Define 𝐿 𝑝 (R𝑑 ) as the Lebesgue space 𝐿 𝑝 (R𝑑 , B(R𝑑 ), 𝜆), where 𝜆 is the Lebesgue measure. In the
case of 𝐿 𝑝 (R𝑑 ), we can choose the simple functions in the approximation more nicely.
Í
First, observe that for a given simple function 𝑗 𝑎 𝑗 𝜒 𝐴 𝑗 with 𝜆( 𝐴 𝑗 ) < ∞ for every 𝑗, we can
approximate this function by a simple function with rational coefficients. Thus, if we approximate each
characteristic function 𝜒 𝐴 with 𝜆( 𝐴) < ∞ by 𝜒𝐵 , where 𝐵 is a member of a countable base for R𝑑 with
𝜆(𝐵) < ∞,3 then we obtain the following proposition.
Proposition 3.21. For each 𝑝 ∈ [1, ∞), 𝐿 𝑝 (R𝑑 ) is separable.
Proof. Let 𝐴 be a Borel set with 𝜆( 𝐴) < ∞. Then, by the outer regularity, there exists a decreasing
sequence (𝑂 𝑛 ) of open sets such that 𝐴 ⊆ 𝑂 𝑛 for all 𝑛 and 𝜆(𝑂 𝑛 ) → 𝜆( 𝐴). Since 𝜆( 𝐴) < ∞, we may
assume that 𝜆(𝑂 𝑛 ) < ∞ for all 𝑛. Then,

k 𝜒𝑂𝑛 − 𝜒 𝐴 k 𝑝𝑝 = 𝜆(𝑂 𝑛 \ 𝐴) = 𝜆(𝑂 𝑛 ) − 𝜆( 𝐴) → 0.

Now, let B be a countable base for R𝑑 . Then, it suffices to show that each open set 𝑂 with 𝜆(𝑂) < ∞
Ð Ð
can be approximated by members in B. Let 𝑂 = 𝑗 𝐵 𝑗 with 𝐵 𝑗 ∈ B. Now, consider 𝑂 𝑛 = 𝑛𝑗=1 𝐵 𝑗 .
Ð
Then, (𝑂 𝑛 ) is an increasing sequence of open sets with 𝑛 𝑂 𝑛 = 𝑂. Therefore,

k 𝜒𝑂𝑛 − 𝜒𝑂 k 𝑝𝑝 = 𝜆(𝑂 \ 𝑂 𝑛 ) = 𝜆(𝑂) − 𝜆(𝑂 𝑛 ) → 0.

Note that the set of all finite unions of members from B is a countable collection, 𝐿 𝑝 (R𝑑 ) is separable. 
3 In fact, this is not a precise explanation. See the proof of the next proposition.

35
In fact, we can generalize the idea of the proof to more general cases. For a detailed proof, see
[Coh13, pp. 102-104].

Proposition 3.22 ([Coh13, Proposition 3.4.5]). Let (𝑋, A, 𝜇) be a measure space, and let 𝑝 ∈ [1, ∞).
Suppose that 𝜇 is 𝜎-finite and that A is generated by a countable subfamily. Then, 𝐿 𝑝 (𝑋, A, 𝜇) is
separable.

Moreover, we can even obtain a nicer approximation result. For a topological space 𝑋, let 𝐶𝑐 (𝑋) be
the space of continuous functions with compact support.

Proposition 3.23. Let 𝑝 ∈ [1, ∞). Then, 𝐶𝑐 (R𝑑 ) is dense in 𝐿 𝑝 (R𝑑 ).

Proof. As in the proof of the separability, if we use the inner regularity, we can approximate each
characteristic function 𝜒 𝐴 of a Borel set 𝐴 with 𝜆( 𝐴) < ∞ by 𝜒𝐾 of a compact set 𝐾. Thus, it suffices to
prove that we can approximate each 𝜒𝐾 by continuous functions on 𝑋 with compact support.
Let 𝐾 be a compact set in R𝑑 , and let 𝜖 > 0 be given. By the outer regularity, there exists an open set
𝑂 such that 𝐾 ⊆ 𝑂 and 𝜆(𝑂 \ 𝐾) 6 𝜖. We can take a continuous function 𝑓 : R𝑑 → [0, 1] with compact
support such that 𝑓 = 1 on 𝐾 and 𝑓 = 0 outside 𝑂.4 Then,
∫ ∫
𝑝
| 𝜒𝐾 − 𝑓 | 𝑑𝜆 = | 𝑓 | 𝑝 𝑑𝜆 6 𝜆(𝑂 \ 𝐾) 6 𝜖 .
𝑂\𝐾

This completes the proof. 

Remark 3.24. For a closed interval [𝑎, 𝑏] in R, we can define the Lebesgue space 𝐿 𝑝 ( [𝑎, 𝑏]) for each
𝑝 ∈ [1, ∞]. Then, for 𝑝 ∈ [1, ∞), we can obtain better approximations, namely,

• the set of step functions on [𝑎, 𝑏] is dense in 𝐿 𝑝 ( [𝑎, 𝑏]), and

• the space 𝐶 ( [𝑎, 𝑏]) of continuous functions on [𝑎, 𝑏] is dense in 𝐿 𝑝 ( [𝑎, 𝑏]).

For a discussion about these facts, see [Coh13, pp. 101-102].

4 One can use a direct construction by using a bump function or use Urysohn’s lemma.

36
4 Signed and complex measures

4.1 Hahn and Jordan decompositions


Definition 4.1. Let (𝑋, A) be a measurable space. A map 𝜇 : A → [−∞, ∞] is said to be a singed
measure if it is countably additive and 𝜇(∅) = 0. Similarly, one can define a complex measure as a map
from A to C.
Moreover, a signed measure 𝜇 is finite if −∞ < 𝜇 < ∞.

Remark 4.2. Let 𝜇 be a signed measure. Notice that ∞ − ∞ is not defined, and consider for a measurable
subset 𝐴,
𝜇(𝑋) = 𝜇( 𝐴) + 𝜇( 𝐴𝑐 ).
Thus, if 𝜇( 𝐴) = ∞, then 𝜇(𝑋) = ∞; if 𝜇( 𝐴) = −∞, then 𝜇(𝑋) = −∞. Therefore, it is impossible that 𝜇
has the values ∞ and −∞ simultaneously. Indeed, if 𝜇( 𝐴) = ∞ and 𝜇(𝐵) = −∞ for measurable sets 𝐴
and 𝐵, then 𝜇(𝑋) must be both ∞ and −∞, but this is not possible. Hence, if 𝜇 is a signed measure, then
the range of 𝜇 is contained in either (−∞, ∞] or [−∞, ∞). Similarly, if 𝐵 is a measurable set with 𝜇(𝐵)
is finite, then for every measurable set 𝐴 contained in 𝐵, 𝜇( 𝐴) is finite.
Moreover, continuity also holds for signed measures, namely, the corresponding results to Proposi-
tion 1.17 and Theorem 1.19 hold.

Example 4.3. Suppose 𝑓 ∈ 𝐿 1 (𝑋, A, 𝜇, R). Then, define



𝜈( 𝐴) B 𝑓 𝑑𝜇, 𝐴 ∈ A.
𝐴

Then, 𝜈 is a signed measure.5 ∫Note that this signed measure


∫ is difference between two positive measures
𝜈1 and 𝜈2 defined by 𝜈1 ( 𝐴) = 𝐴 𝑓 + 𝑑𝜇 and 𝜈2 ( 𝐴) = 𝐴 𝑓 − 𝑑𝜇.
In fact, if 𝜈1 and 𝜈2 are positive measures and if at least one of them is finite, then 𝜈1 − 𝜈2 defines a
signed measure. In the discussion about the Jordan decomposition theorem, we will see that any signed
measure is given in this way.
In addition, one can provide an example of a complex measure by using some integrable complex-
valued function. Of course, such a measure also can be decomposed into positive measures.

We now want to study some standard decompositions of signed or complex measures. Let 𝜇 be a
signed measure on a measurable space (𝑋, A). A measurable subset 𝐴 of 𝑋 is called a positive set if for
every measurable set 𝐸 contained in 𝐴, 𝜇(𝐸) > 0. Likewise, we define negative sets.
Note that if 𝐴 ⊆ 𝐵 are positive sets, then since

𝜇(𝐵) = 𝜇( 𝐴) + 𝜇(𝐵 \ 𝐴),

𝜇( 𝐴) 6 𝜇(𝐵). Similarly, if 𝐴 ⊆ 𝐵 are negative sets, then 𝜇( 𝐴) > 𝜇(𝐵).

Theorem 4.4 (Hahn decomposition). Let 𝜇 be a signed measure on a measurable space (𝑋, A). Then,
𝑋 is the disjoint union of a positive set and a negative set.

Remark 4.5. Let 𝑋 = 𝑃1 ∪ 𝑁1 = 𝑃2 ∪ 𝑁2 be two Hahn decompositions of 𝜇. Then, by definition, 𝑃1 ∩ 𝑁2


and 𝑁1 ∩ 𝑃2 are both positive and negative. Consequently, they are of measure zero relative to 𝜇, so the
Hahn decomposition of 𝜇 is essentially unique.

Proof. We may assume that 𝜇 does not attain the value −∞. The idea of this proof is to construct a
negative set with the minimum signed measure. Let

𝐿 B inf{𝜇( 𝐴) | 𝐴 is a negative set}.


5 Moreover, this signed measure is finite.

37
Then, since ∅ is a negative set, 𝐿 6 0. Now, take a sequence ( 𝐴𝑛 ) of negative sets such that lim 𝜇( 𝐴𝑛 ) = 𝐿.
Define 𝑁 B ∞
Ð
𝑛=1 𝐴𝑛 . We first claim that
Ð
𝑁 is a negative set. To shows this, let 𝐸 be a measurable
Ð∞
set contained in 𝑁. Define 𝐵𝑛 B 𝐴𝑛 \ 𝑛−1 𝑗=1 𝐴 𝑗 . Then, 𝑁 = 𝑛=1 𝐵 𝑛 , and each 𝐵 𝑛 is a negative set.
Therefore, since (𝐵𝑛 ) is a disjoint family,

Õ
𝜇(𝐸) = 𝜇(𝐸 ∩ 𝐵𝑛 ) 6 0.
𝑛=1

Moreover, since 𝑁 is a negative set, for each 𝑛,

𝐿 6 𝜇(𝑁) 6 𝜇( 𝐴𝑛 ).

Letting 𝑛 → ∞, we have 𝐿 = 𝜇(𝑁).


Now, let 𝑃 = 𝑁 𝑐 , and suppose 𝑃 is not a positive set. To obtain a contradiction, we want to construct
a negative set 𝑁 0 such that 𝜇(𝑁 0) < 𝐿. Observe that if 𝐵 ⊆ 𝑃 is a negative set with 𝜇(𝐵) < 0, then

𝜇(𝑁 ∪ 𝐵) = 𝜇(𝑁) + 𝜇(𝐵) < 𝜇(𝑁) = 𝐿.

Thus, it suffices to show that there is a negative set 𝐵 ⊆ 𝑃 with 𝜇(𝐵) < 0. This can be done by the
following lemma. 

Lemma 4.6. Let 𝐴 be a measurable subset with −∞ < 𝜇( 𝐴) < 0. Then, there exists a negative set 𝐵
such that 𝐵 ⊆ 𝐴 and 𝜇(𝐵) 6 𝜇( 𝐴).

Proof. See [Coh13, Lemma 4.1.4]. 

Corollary 4.7 (Jordan decomposition). Let 𝜇 be a signed measure on a measurable space (𝑋, A). Then,
there are two positive measure 𝜇+ and 𝜇− such that

𝜇 = 𝜇+ − 𝜇− .

Moreover, at least one of these measures is finite.

Proof. Let 𝑋 = 𝑃 ∪ 𝑁 be a Hahn decomposition. For a measurable set 𝐴, define

𝜇+ ( 𝐴) B 𝜇( 𝐴 ∩ 𝑃)
𝜇− ( 𝐴) B −𝜇( 𝐴 ∩ 𝑁).

Then, 𝜇+ and 𝜇− are positive measures, and at least one of them is finite. Moreover, 𝜇 = 𝜇+ − 𝜇− . 

Remark 4.8. Let 𝐴 be a measurable set. Observe that if 𝐵 is a measurable set contained in 𝐴, then

𝜇(𝐵) = 𝜇+ (𝐵) − 𝜇− (𝐵) 6 𝜇+ (𝐵) 6 𝜇+ ( 𝐴).

Moreover, since 𝜇+ ( 𝐴) = 𝜇( 𝐴 ∩ 𝑃), we obtain that

𝜇( 𝐴) = sup{𝜇(𝐵) | 𝐵 is measurable and 𝐵 ⊆ 𝐴}.

Similarly,
𝜇− ( 𝐴) = − inf{𝜇(𝐵) | 𝐵 is measurable and 𝐵 ⊆ 𝐴}.
Therefore, 𝜇+ and 𝜇− does not depend on the choice of a Hahn decomposition. 𝜇+ and 𝜇− are called the
positive part and the negative part of 𝜇.

38
4.2 Banach space of complex measures
Definition 4.9. Let 𝑀 (𝑋, A, R) be the set of finite singed measures on the measurable space (𝑋, A).
Similarly, define the set 𝑀 (𝑋, A, C) of complex measures. Then, these sets are vector spaces over R and
C, respectively.

We want to provide appropriate norms on these spaces. First, for a signed or complex measure 𝜇, we
construct a smallest positive measure 𝜈 such that

|𝜇( 𝐴)| 6 𝜈( 𝐴) (4.1)

for all measurable subset 𝐴. One may expect that if we define

𝜈( 𝐴) B |𝜇( 𝐴)|,

then we have such a positive measure, but this is not true in general. Indeed, if 𝐴 and 𝐵 are disjoint
measurable sets, then
|𝜇( 𝐴 ∪ 𝐵)| = |𝜇( 𝐴) + 𝜇(𝐵)| 6 |𝜇( 𝐴)| + |𝜇(𝐵)|,
and one may construct a signed or complex measure 𝜇 and measurable sets 𝐴 and 𝐵 such that the
inequality is not equality.
Now, suppose the inequality (4.1) holds for all measurable sets 𝐴. Then, considering the additivity
of 𝜈, if ( 𝐴𝑛 ) is a countable disjoint family of measurable sets, then
!
Ø Õ Õ
𝜈 𝐴𝑛 = 𝜈( 𝐴𝑛 ) > |𝜈( 𝐴𝑛 )|.
𝑛 𝑛 𝑛

Thus, it is reasonable to consider the least upper bound of the right-hand side, so define
( )
Õ
|𝜇|( 𝐴) B sup |𝜇( 𝐴𝑛 )| ( 𝐴𝑛 ) is a partition of 𝐴 , (4.2)
𝑛

where 𝐴 is a measurable set, and a partition of 𝐴 means a countable disjoint family ( 𝐴𝑛 ) of measurable
Ð
sets such that 𝑛 𝐴𝑛 = 𝐴.

Definition 4.10. The set function |𝜇| defined in (4.2) is called the variation of 𝜇.

Remark 4.11. We can show that for any measurable set 𝐴,


Õ
 𝑛 


|𝜇|( 𝐴) = sup |𝜇( 𝐴 𝑗 )| ( 𝐴 𝑗 ) 𝑛𝑗=1 is a finite partition of 𝐴 .
 
 𝑗=1 
To show this, let 𝛼( 𝐴) be the right-hand side. Then, it is clear that 𝛼( 𝐴) 6 |𝜇|( 𝐴). We now show the
reverse inequality. If 𝛼( 𝐴) = ∞, then obviously |𝜇|( 𝐴) 6 𝛼( 𝐴). Now, suppose 𝛼( 𝐴) < ∞, and let
∞ be a partition of 𝐴. Then, for each 𝑁,
( 𝐴𝑛 ) 𝑛=1
𝑁
Õ Ð 

|𝜇( 𝐴 𝑗 )| + 𝜇 𝑗=𝑁 +1 𝐴 𝑗 6 𝛼( 𝐴) < ∞.
𝑗=1

By the continuity of 𝜇 from above, if we let 𝑁 → ∞, we have



Õ
|𝜇( 𝐴 𝑗 )| 6 𝛼( 𝐴).
𝑗=1

Since the partition ( 𝐴𝑛 ) is arbitrary, we have |𝜇|( 𝐴) 6 𝛼( 𝐴).

39
Moreover, suppose 𝜇 is a signed measure, and let 𝑋 = 𝑃 ∪ 𝑁 be a Hahn decomposition. Then, for a
given measurable set 𝐴 and a partition ( 𝐴𝑛 ) of 𝐴,
Õ Õ Õ Õ
|𝜇( 𝐴𝑛 )| = |𝜇( 𝐴𝑛 ∩ 𝑃) + 𝜇( 𝐴𝑛 ∩ 𝑁)| 6 𝜇+ ( 𝐴𝑛 ) + 𝜇− ( 𝐴𝑛 ) = 𝜇+ ( 𝐴) + 𝜇− ( 𝐴).
𝑛 𝑛 𝑛 𝑛

Therefore,
|𝜇|( 𝐴) 6 𝜇+ ( 𝐴) + 𝜇− ( 𝐴).
In addition, since 𝐴 = (𝑃 ∩ 𝐴) ∪ (𝑁 ∩ 𝐴) and this union is disjoint,

𝜇+ ( 𝐴) + 𝜇− ( 𝐴) = 𝜇(𝑃 ∩ 𝐴) + 𝜇(𝑁 ∩ 𝐴) 6 |𝜇|( 𝐴).

This shows that if 𝜇 is a signed measure, then

|𝜇| = 𝜇+ + 𝜇− . (4.3)

Proposition 4.12. The variation 𝜇 is the smallest positive measure 𝜈 such that |𝜇( 𝐴)| 6 𝜈( 𝐴) for all
measurable set 𝐴. Moreover, if 𝜇 is a real or complex measure, then |𝜇| is a finite positive measure.

Proof. First, suppose 𝜇 is a complex measure, and let

𝜇 = 𝜇1 − 𝜇2 + 𝑖𝜇3 − 𝑖𝜇4

be the Jordan decomposition of 𝜇. In other words, 𝜇1 − 𝜇2 is the Jordan decomposition of the real
part, and 𝜇3 − 𝜇4 is the Jordan decomposition of the imaginary part. Then, for any partition ( 𝐴𝑛 ) of a
measurable set 𝐴, Õ
|𝜇( 𝐴𝑛 )| 6 𝜇1 ( 𝐴) + 𝜇2 ( 𝐴) + 𝜇3 ( 𝐴) + 𝜇4 ( 𝐴) < ∞.
𝑛
Thus, |𝜇( 𝐴)| < 𝜇1 ( 𝐴) + 𝜇2 ( 𝐴) + 𝜇3 ( 𝐴) + 𝜇4 ( 𝐴) < ∞, so |𝜇| is finite.
Moreover, by the definition (4.2) of |𝜇|, it suffices to show that |𝜇| is a measure. In the case that 𝜇 is a
signed measure, this follows from (4.3). If 𝜇 is a complex measure, just show that |𝜇| is finitely additive
and satisfies continuity from above. (For a detailed proof, see [Coh13, pp. 118-119].) 

Definition 4.13. Let 𝜇 be a real or complex measure on a measurable space (𝑋, A). The total variation
of 𝜇 is defined by k𝜇k B |𝜇|(𝑋).

Remark 4.14. The total variation provides a norm on the space 𝑀 (𝑋, A, C). The homogeneity is clear
from the defintion of the total variation. To show the positive definiteness, suppose k𝜇k = |𝜇|(𝑋) = 0.
Then, for every measurable set 𝐴, |𝜇( 𝐴)| 6 |𝜇|( 𝐴) = 0, so 𝜇( 𝐴) = 0, that is, 𝜇 = 0. To show the triangle
inequality, just observe that for a measurable set 𝐴,

|𝜇( 𝐴) + 𝜈( 𝐴)| 6 |𝜇( 𝐴)| + |𝜈( 𝐴)| 6 |𝜇|( 𝐴) + |𝜈|( 𝐴).

This shows that |𝜇 + 𝜈| 6 |𝜇| + |𝜈|, and so k𝜇 + 𝜈k 6 k𝜇k + k𝜈k.


Similarly, 𝑀 (𝑋, A, R) is also a normed space with respect to the total variation.

Proposition 4.15. The spaces 𝑀 (𝑋, A, R) and 𝑀 (𝑋, A, C) are Banach spaces with respect to the total
variation norm.

Proof. Suppose (𝜇 𝑛 ) be a Cauchy sequence in 𝑀 (𝑋, A, R) or 𝑀 (𝑋, A, C). Then, for each measurable
set 𝐴,
|𝜇 𝑚 ( 𝐴) − 𝜇 𝑛 ( 𝐴)| 6 |𝜇 𝑚 − 𝜇 𝑛 |( 𝐴) 6 k𝜇 𝑚 − 𝜇 𝑛 k, (4.4)
so the sequence (𝜇 𝑛 ( 𝐴)) is Cauchy in R. Therefore, define

𝜇( 𝐴) B lim 𝜇 𝑛 ( 𝐴).
𝑛

40
We now need to verify that 𝜇 is a signed or complex measure. First, it is clear that 𝜇(∅) = 0, and
𝜇 is finitely additive. To show the countable additivity, let ( 𝐴 𝑘 ) be a sequence of measurable sets with
Ñ
𝑘 𝐴 𝑘 = ∅. We claim that lim 𝑘 𝜇( 𝐴 𝑘 ) = 0. First, observe the following estimate: for each 𝑁,

|𝜇( 𝐴 𝑘 )| 6 |𝜇( 𝐴 𝑘 ) − 𝜇 𝑁 ( 𝐴 𝑘 )| + |𝜇 𝑁 ( 𝐴 𝑘 )|.

Now, let 𝜖 > 0 be given, and take 𝑁 so that if 𝑚, 𝑛 > 𝑁, then k𝜇 𝑚 − 𝜇 𝑛 k 6 𝜖. By (4.4), if 𝑚, 𝑛 > 𝑁, then
for all measurable set 𝐴,
|𝜇 𝑚 ( 𝐴) − 𝜇 𝑛 ( 𝐴)| 6 𝜖 .
Thus, by letting 𝑚 → ∞ and taking 𝑛 = 𝑁, we obtain the following: for every measurable set 𝐴,

|𝜇( 𝐴) − 𝜇 𝑁 ( 𝐴)| 6 𝜖 .

Therefore, by the previous estimate,

|𝜇( 𝐴 𝑘 )| 6 𝜖 + |𝜇 𝑁 ( 𝐴 𝑘 )|.

Since 𝜇 𝑁 is a signed or complex measure, by letting 𝑘 → ∞, we have

lim sup |𝜇( 𝐴 𝑘 )| 6 𝜖 .


𝑘

Note that 𝜖 > 0 is arbitrary, we have lim sup 𝑘 |𝜇( 𝐴 𝑘 )| = 0, this means that lim 𝑘 |𝜇( 𝐴 𝑘 )| = 0. Hence, 𝜇 is
a signed or complex measure.
We now want to prove lim𝑛 k𝜇 𝑛 − 𝜇k = 0. From the first part of the inequality (4.4), if ( 𝐴 𝑗 ) 𝑁 𝑗=1 is a
finite partition of 𝑋, then
𝑁
Õ
|𝜇 𝑚 ( 𝐴 𝑗 ) − 𝜇 𝑛 ( 𝐴 𝑗 )| 6 |𝜇 𝑚 − 𝜇 𝑛 |(𝑋) = k𝜇 𝑚 − 𝜇 𝑛 k.
𝑗=1

Now, let 𝜖 > 0 be given, and let 𝑁 be a positive integer so that if 𝑚, 𝑛 > 𝑁, then k𝜇 𝑚 − 𝜇 𝑛 k 6 𝜖. Then,
for an arbitrary finite partition ( 𝐴 𝑗 ) 𝑁
𝑗=1 of 𝑋, by the previous inequality, if 𝑚, 𝑛 > 𝑁, then

𝑁
Õ
|𝜇 𝑚 ( 𝐴 𝑗 ) − 𝜇 𝑛 ( 𝐴 𝑗 )| 6 𝜖 .
𝑗=1

By letting 𝑚 → ∞, we obtain that if 𝑛 > 𝑁, then


𝑁
Õ
|𝜇 𝑛 ( 𝐴 𝑗 ) − 𝜇( 𝐴 𝑗 )| 6 𝜖 .
𝑗=1

Since the partition ( 𝐴 𝑗 ) 𝑁


𝑗=1 is arbitrary, and 𝑁 is independent of the choice of a partition, we conclude
that if 𝑛 > 𝑁, then
k𝜇 𝑛 − 𝜇k 6 𝜖 .
This completes the proof. 

4.3 Radon-Nikodym theorem and Lebesgue decomposition


Definition 4.16. Let 𝜇 and 𝜈 be positive measures on a measurable space (𝑋, A). 𝜈 is said to be absolutely
continuous with respect to 𝜇 if for each measurable set 𝐴 with 𝜇( 𝐴) = 0, 𝜈( 𝐴) = 0. In this case, we write
𝜈  𝜇.
Moreover, a measure on (R𝑑 , B(R𝑑 )) is said to be absolutely continuous if it is absolutely continuous
with respect to the Lebesgue measure.

41
Example 4.17. Let 𝑓 be an integrable nonnegative function on (𝑋, A, 𝜇). Then, the measure 𝜈 defined
by ∫
𝜈( 𝐴) B 𝑓 𝑑𝜇, 𝐴∈A
𝐴

is absolutely continuous with respect to 𝜇. We will show that if 𝜇 is 𝜎-finite, then every 𝜎-finite absolutely
continuous measure with respect to 𝜇 is given in this way.

In the case of finite measures, we can find an equivalent statement to absolute continuity in terms of
𝜖 and 𝛿.

Lemma 4.18. Suppose that 𝜈 is a finite measure. Then, 𝜈  𝜇 if and only if for every 𝜖 > 0, there exists
𝛿 > 0 such that for every measurable set 𝐴, 𝜇( 𝐴) < 𝛿 implies 𝜈( 𝐴) < 𝜖.

Proof. Suppose that for each 𝜖 > 0, there is a corresponding 𝛿 > 0. Suppose 𝜇( 𝐴) = 0 for a measurable
set 𝐴. Then, 𝜇( 𝐴) < 𝛿 for every 𝛿 > 0, so for every 𝜖 > 0, 𝜈( 𝐴) < 𝜖. This implies that 𝜈( 𝐴) = 0. Thus,
𝜈  𝜇.
Now, suppose that 𝜈  𝜇, and assume that there exists an 𝜖 > 0 such that there is no corresponding
𝛿. Then, for each 𝑘, there exists a measurable set 𝐴 𝑘 such that 𝜇( 𝐴 𝑘 ) < 1/2 𝑘 , but 𝜈( 𝐴 𝑘 ) > 𝜖. From this,
we now construct a measurable set 𝐴 such that 𝜇( 𝐴) = 0, but 𝜈( 𝐴) > 𝜖 > 0. By subadditivity, for each 𝑛,

! ∞
Ø Õ 1 1
𝜇 𝐴𝑘 6 𝑘
= 𝑛−1 < ∞.
𝑘=𝑛 𝑘=𝑛
2 2
Ð∞
Note that since ( 𝑘=𝑛 𝐴 𝑘 ) is a monotonically decreasing family,

∞ Ø

!
Ù
𝜇 𝐴 𝑘 = 0.
𝑛=1 𝑘=𝑛

Moreover, since 𝜈 is a finite measure,


∞ Ø

! ∞
!
Ù Ø
𝜈 𝐴 𝑘 = lim 𝜈 𝐴𝑘 > 𝜖 .
𝑛
𝑛=1 𝑘=𝑛 𝑘=𝑛
Ñ∞ Ð∞
Thus, let 𝐴 = 𝑛=1 𝑘=𝑛 𝐴 𝑘 , then this completes the proof. 

Theorem 4.19 (Radon-Nikodym, for positive measures). Let 𝜇 and 𝜈 be 𝜎-finite measures on a mea-
surable space (𝑋, A). Suppose 𝜈  𝜇. Then, there exists a measurable function 𝑔 : 𝑋 → [0, ∞) such
that ∫
𝜈( 𝐴) = 𝑔𝑑𝜇
𝐴

for every measurable set 𝐴. Moreover, the function 𝑔 is unique up to 𝜇-almost everywhere equality.

Proof. First, suppose that 𝜇 and 𝜈 are finite measures. We now consider the collection F of measurable
functions 𝑓 : 𝑋 → [0, ∞] such that ∫
𝑓 𝑑𝜇 6 𝜈( 𝐴)
𝐴

for every measurable set 𝐴, and observe that if 𝑓 ∈ F, then the map 𝐴 ∈ A ↦→ 𝜈( 𝐴) − 𝐴 𝑓 𝑑𝜇 defines a
positive measure. Moreover, if 𝑓 ∈ F is integrable, so by redefining the values of 𝑓 on a set of measure
zero, we∫ may assume that 𝑓 : 𝑋 → [0, ∞). Therefore, it suffices to show that there exists 𝑓 ∈ F with
𝜈(𝑋) = 𝑓 𝑑𝜇.

42
∫ To find such a member in F, we want to seek a function 𝑔 in the collection F such that the integral
𝑔𝑑𝜇 is the largest. This means that we want to find 𝑔 ∈ F such that
∫ ∫ 
𝑔𝑑𝜇 = sup 𝑓 𝑑𝜇 𝑓 ∈ F . (4.5)

Assume that such 𝑔 ∈ F exists. Then, define a positive measure 𝜈0 by



𝜈0 ( 𝐴) B 𝜈( 𝐴) − 𝑔𝑑𝜇, 𝐴 ∈ A.
𝐴

Note that 𝜈0 is finite and is absolutely continuous with respect to 𝜇. Now, suppose 𝜈0 (𝑋) > 0. Then,
𝜇(𝑋) > 0, and since 𝜇 is a finite measure, there exists an 𝜖 > 0 such that

𝜈0 (𝑋) > 𝜖 𝜇(𝑋).

Now, let 𝑋 = 𝑃 ∪ 𝑁 be a Hahn decomposition of the signed measure 𝜈0 − 𝜖 𝜇. Then,

0 < 𝜈0 (𝑃) + (𝜈0 (𝑁) − 𝜖 𝜇(𝑁)) − 𝜖 𝜇(𝑃) 6 𝜈0 (𝑃).

Moreover, for each measurable set 𝐴,


∫ ∫
𝜈( 𝐴) = 𝜈0 ( 𝐴) + 𝑔𝑑𝜇 > 𝜈0 ( 𝐴 ∩ 𝑃) + 𝑔𝑑𝜇
∫ 𝐴 ∫ 𝐴

> 𝜖 𝜇( 𝐴 ∩ 𝑃) + 𝑔𝑑𝜇 = (𝑔 + 𝜖 𝜒𝑃 )𝑑𝜇.


𝐴 𝐴
∫ ∫
Thus,
∫ 𝑔 + 𝜖 𝜒 𝑃 ∈ F. However, since 𝑔 ∈ F, 𝑔𝑑𝜇 6 𝜈(𝑋) < ∞, and since 𝜇(𝑃) > 0, (𝑔 + 𝜖 𝜒𝑃 )𝑑𝜇 >
𝑔𝑑𝜇. This is a contradiction, so 𝜈0 (𝑋) = 0.
We now construct 𝑔 ∈ F satisfying (4.5). First, note that F is nonempty since 0 ∈ F. Moreover,
note that if 𝑓1 , 𝑓2 ∈ F, then 𝑓1 ∨ 𝑓2 ∈ F. Indeed, for a measurable set 𝐴,
∫ ∫ ∫
( 𝑓1 ∨ 𝑓2 )𝑑𝜇 = 𝑓1 𝑑𝜇 + 𝑓2 𝑑𝜇
𝐴 𝐴∩{ 𝑓1 > 𝑓2 } 𝐴∩{ 𝑓1 < 𝑓2 }
6 𝜈( 𝐴 ∩ { 𝑓1 > 𝑓2 }) + 𝜈( 𝐴 ∩ { 𝑓1 < 𝑓2 })
= 𝜈( 𝐴).
∫ ∫
Now, let ( 𝑓𝑛 ) be a sequence in F such that 𝑓𝑛 𝑑𝜇 → sup{ 𝑓 𝑑𝜇 | 𝑓 ∈ F }. By redefining 𝑓𝑛 to be
𝑓1 ∨ 𝑓2 ∨ · · · ∨ 𝑓𝑛 , we may assume that ( 𝑓𝑛 ) is increasing. Let 𝑔 = lim 𝑓𝑛 .6 Then, by the monotone
convergence theorem, ∫ ∫ ∫ 
𝑔𝑑𝜇 = lim 𝑓𝑛 𝑑𝜇 = sup 𝑓 𝑑𝜇 𝑓 ∈ F

and since each 𝑓𝑛 ∈ F, by the monotone convergence theorem again, for every measurable set 𝐴,
∫ ∫
𝑔𝑑𝜇 = lim 𝑓𝑛 𝑑𝜇 6 𝜈( 𝐴).
𝐴 𝐴

This means that 𝑔 ∈ F.


Suppose now that 𝜇 and 𝜈 are 𝜎-finite. Then, 𝑋 is the disjoint union of countably many measurable
sets 𝐵𝑛 for which each 𝐵𝑛 has finite measure ∫under both 𝜇 and 𝜈. Then, for each 𝑛, let 𝑔𝑛 : 𝐵𝑛 → [0, ∞)
be a measurable function such that 𝜈(𝐵) = 𝐵 𝑔𝑛 𝑑𝜇 for every measurable subset 𝐵 of 𝐵𝑛 . Define the
function 𝑔 : 𝑋 → [0, ∞) whose restriction on 𝐵𝑛 is 𝑔𝑛 . Then, this function is the required function.
6 Notethat 𝑔 may attain the value ∞. This is why we define the family F as a subfamily of extended-real-valued functions,
not as a subfamily of real-valued functions.

43
To show the uniqueness, let 𝑔, ℎ : 𝑋 → [0, ∞) be measurable functions such that
∫ ∫
𝜈( 𝐴) = 𝑔𝑑𝜇 = ℎ𝑑𝜇
𝐴 𝐴

for every measurable set 𝐴. First, suppose 𝜈 is finite. Then, both 𝑔 and ℎ are integrable with respect to 𝜇,
so 𝑔 = ℎ 𝜇-almost everywhere. If 𝜈 is 𝜎-finite, by decomposing 𝑋 into measurable sets of finite measure,
we obtain the uniqueness. 

Remark 4.20. We need to assume that 𝜇 is 𝜎-finite. Let 𝑋 = [0, 1], and let the 𝜎-algebra be the Borel
𝜎-algebra of [0, 1]. Let 𝜇 be the counting measure, and ∫let 𝜈 be the Lebesgue measure. Then, 𝜈  𝜇,
but there is no measurable function 𝑓 such that 𝜈( 𝐴) = 𝐴 𝑓 𝑑𝜇 for every measurable set 𝐴. Indeed, if
such a function exists, then∫for every 𝑥 ∈ 𝑋, by taking 𝐴 = {𝑥}, we have 𝑓 (𝑥) = 0. Simultaneously, by
taking 𝐴 = 𝑋 = [0, 1], 1 = 𝑓 𝑑𝜇, but this is impossible.
We can extend the Radon-Nikodym theorem to signed or complex measures.

Definition 4.21. Let 𝜇 be a positive measure on a measurable space (𝑋, A). Let 𝜈 be a signed or complex
measure on (𝑋, A). Then, 𝜈 is said to be absolutely continuous with respect to 𝜇, denoted by 𝜈  𝜇, if
the variation |𝜈| is absolutely continuous with respect to 𝜇.

Remark 4.22. By the definition of the variation |𝜈|, 𝜈  𝜇 if and only if 𝜇( 𝐴) = 0 implies 𝜈( 𝐴) = 0 for
every measurable set 𝐴.
Moreover, for a signed measure 𝜈, since |𝜈| = 𝜈 + + 𝜈 − , 𝜈  𝜇 if and only if both 𝜈 + and 𝜈 − are
absolutely continuous with respect to 𝜇. Likewise, for a complex measure 𝜈 and its Jordan decomposition
𝜈 = 𝜈1 − 𝜈2 + 𝑖𝜈3 − 𝑖𝜈4 , 𝜈  𝜇 if and only if all the measures 𝜈1 , . . . , 𝜈4 are absolutely continuous with
respect to 𝜇. This follows from the inequalitis |𝜈1 − 𝜈2 |, |𝜈3 − 𝜈4 | 6 |𝜈| 6 𝜈1 + 𝜈2 + 𝜈3 + 𝜈4 .

Theorem 4.23 (Radon-Nikodym, for real or complex measures). Let 𝜇 be a 𝜎-finite measure on a
measurable space (𝑋, A). Let 𝜈 be a real or complex measure on (𝑋, A). Suppose 𝜈  𝜇. Then, there
exists a measurable function 𝑔 in L1 (𝑋, A, 𝜇) such that

𝜈( 𝐴) = 𝑔𝑑𝜇
𝐴

for every measurable set 𝐴. Moreover, the function 𝑔 is unique up to 𝜇-almost everywhere equality.

Definition 4.24. Let 𝜇 be a 𝜎-finite measure on a measurable space (𝑋, A), and let 𝜈 be a real, complex,
or 𝜎-finite positive measure on (𝑋, A). If 𝜈  𝜇, ∫then there exists a unique (up to 𝜇-almost everywhere
equality) measurable function 𝑔 such that 𝜈( 𝐴) = 𝐴 𝑔𝑑𝜇 for every measurable set 𝐴. This function 𝑔 is
called the Radon-Nikodym derivative of 𝜈 with respect to 𝜇, and this is denoted by
𝑑𝜈
.
𝑑𝜇
Remark 4.25. For the Radon-Nikodym derivative 𝑔 of 𝜈 with respect to 𝜇, one can show that

|𝜈|( 𝐴) = |𝑔|𝑑𝜇
𝐴

for every measurable set 𝐴 (see [Coh13, Proposition 4.2.5]).


Consequently, since 𝜈  |𝜈|, if 𝑓 = 𝑑𝜈/𝑑|𝜈|, then

|𝜈|( 𝐴) = | 𝑓 |𝑑|𝜈|
𝐴

for every measurable set 𝐴. By the uniqueness of the Radon-Nikodym derivative, | 𝑓 | = 1 |𝜈|-almost
everywhere.

44
Definition 4.26. Let (𝑋, A) be a measurable space. A measure 𝜇 is said to be concentrated on the
measurable set 𝐸 if 𝜇(𝐸 𝑐 ) = 0. Similarly, a signed or complex measure 𝜇 is said to be concentrated on
𝐸 if the variation |𝜇| is concentrated on 𝐸.
Two positive, signed, or complex measures 𝜇 and 𝜈 on (𝑋, A) are said to be mutually singular if there
exists a measurable set 𝐸 such that 𝜇 is concentrated on 𝐸 and 𝜈 is concentrated on 𝐸 𝑐 . In this case, we
also write 𝜇 ⊥ 𝜈 and say that 𝜈 is singular with respect to 𝜇, or that 𝜇 is singular with respect to 𝜈.
A positive, signed, or complex measure on (R𝑑 , B(R𝑑 )) is said to be singular if it is singular with
respect to the Lebesgue measure.

Example 4.27. Let 𝜇 be a signed measure, and let 𝜇 = 𝜇+ − 𝜇− be the Jordan decomposition. Then, by
taking a Hahn decomposition of 𝜇, we can show that 𝜇+ ⊥ 𝜇− .

Theorem 4.28 (Lebesgue decomposition). Let 𝜇 be a positive measure on a measurable space (𝑋, A),
and let 𝜈 be a real, complex, or 𝜎-finite positive measure on (𝑋, A). Then, there are unique real, complex,
or positive measures 𝜈 𝑎 and 𝜈𝑠 on (𝑋, A) such that the following hold.

1. 𝜈 𝑎 is absolutely continuous with respect to 𝜇.

2. 𝜈𝑠 is singular with respect to 𝜇.

3. 𝜈 = 𝜈 𝑎 + 𝜈𝑠 .

This decomposition 𝜈 = 𝜈 𝑎 + 𝜈𝑠 is called the Lebesgue decomposition of 𝜈 with respect to 𝜇.

Proof. Suppose first that 𝜈 is a finite positive measure. We consider a measurable set 𝑁 such that
𝜇(𝑁) = 0, and 𝜈(𝑁) has the maximum measure with respect to 𝜈. To construct such a measurable set, let
N 𝜇 be the collection of measurable sets 𝐵 with 𝜇(𝐵) = 0. Then, this family is nonempty since 𝜇(∅) = 0.
Now, consider a sequence (𝐵 𝑗 ) of sets in N 𝜇 such that

lim 𝜈(𝐵 𝑗 ) = sup{𝜈(𝐵) | 𝐵 ∈ N 𝜇 }.


𝑗
Ð
Then, let 𝑁 = 𝑗 𝐵 𝑗 . Then, by the subadditivity, 𝜇(𝑁) = 0, and since

𝜈(𝐵 𝑗 ) 6 𝜈(𝑁) 6 sup{𝜈(𝐵) | 𝐵 ∈ N 𝜇 },

for each 𝑗, we have


𝜈(𝑁) = sup{𝜈(𝐵) | 𝐵 ∈ N 𝜇 } < ∞.
Define for each measurable set 𝐴,

𝜈 𝑎 ( 𝐴) B 𝜈( 𝐴 ∩ 𝑁 𝑐 )
𝜈𝑠 ( 𝐴) B 𝜈( 𝐴 ∩ 𝑁).

Then, clearly, 𝜈 = 𝜈 𝑎 + 𝜈𝑠 , and since 𝜈𝑠 (𝑁 𝑐 ) = 0, 𝜇 and 𝜈𝑠 are mutually singular. Suppose now that
𝜇(𝐵) = 0. If 𝜈 𝑎 (𝐵) = 𝜈(𝐵 ∩ 𝑁 𝑐 ) > 0, then

𝜈(𝑁 ∪ 𝐵) = 𝜈(𝑁) + 𝜈(𝐵 ∩ 𝑁 𝑐 ) > 𝜈(𝑁),

but since 𝜇(𝑁 ∩ 𝐵) = 0, we have a contradiction. Thus, 𝜈 𝑎 is absolutely continuous with respect to 𝜇.
If 𝜈 is a real or complex measure, then by applying the previous argument to the finite measure |𝜈|,
we can obtain the measurable set 𝑁. Then, define 𝜈 𝑎 and 𝜈𝑠 in the same manner. Now, suppose 𝜈 is a
𝜎-finite positive measure. Then, by taking a partition (𝐷 𝑘 ) of 𝑋 such that 𝜇 is finite on each 𝐷 𝑘 , and
then apply the previous argument to the restriction of 𝜇 on 𝐷 𝑘 . Then, we obtain measurable subsets 𝑁 𝑘
Ð
of 𝐷 𝑘 . Now, take 𝑁 = 𝑘 𝑁 𝑘 , and likewise define 𝜈 𝑎 and 𝜈𝑠 .

45
We now show the uniqueness. Let 𝜈 = 𝜈 𝑎 + 𝜈𝑠 = 𝜈 𝑎0 + 𝜈𝑠0 be two Lebesgue decompositions. Suppose
𝜈 is a real or complex measure. Then, since

𝜆 = 𝜈 𝑎 − 𝜈 𝑎0 = 𝜈𝑠0 − 𝜈𝑠 ,

the signed or complex measure 𝜆 is both absolutely continuous and singular with respect to 𝜇. Take a
measurable set 𝐸 so that |𝜆|(𝐸) = 0 and 𝜇(𝐸 𝑐 ) = 0. Then, |𝜆|(𝐸 𝑐 ) = 0 since 𝜆  𝜇, so |𝜆|(𝑋) = 0.
Therefore, 𝜆 = 0. If 𝜈 is a 𝜎-finite positive measure, then by taking a partition of 𝑋, we can reduce this
case to the case of finite positive measures. 

46
5 Product measures

5.1 Product measures


Definition 5.1. Let (𝑋, A) and (𝑌 , B) be measurable spaces. A measurable rectangle is a set 𝐴 × 𝐵,
where 𝐴 ∈ A and 𝐵 ∈ B. The product 𝜎-algebra A × B is the 𝜎-algebra generated by measurable
rectangles in 𝑋 × 𝑌 .

Example 5.2. Let 𝑋 and 𝑌 be second countable topological spaces. Then, we can consider the 𝜎-algebras
B(𝑋 × 𝑌 ) and B(𝑋) × B(𝑌 ). We claim that these 𝜎-algebras are the same.
Let 𝜋 𝑋 and 𝜋𝑌 are the projections onto 𝑋 and 𝑌 , respectively. Then, these are continuous, so
measurable with respect to the Borel 𝜎-algebras. Therefore, for each pair of Borel subsets 𝐴 ⊆ 𝑋 and
𝐵 ⊆ 𝑌 , 𝜋 𝑋 −1 ( 𝐴) = 𝐴 × 𝑌 and 𝜋𝑌 −1 (𝐵) = 𝑋 × 𝐵 are Borel subsets, so

𝐴 × 𝐵 = ( 𝐴 × 𝑌 ) ∩ (𝑋 × 𝐵)

is a Borel subset of 𝑋 × 𝑌 . This shows that B(𝑋) × B(𝑌 ) ⊆ B(𝑋 × 𝑌 ). Moreover, since 𝑋 × 𝑌 is
second countable, any open subset 𝑈 of 𝑋 × 𝑌 is the countable union of open subsets of the form
𝐴 × 𝐵, where 𝐴 ⊆ 𝑋 and 𝐵 ⊆ 𝑌 are open, and hence 𝑈 belongs to B(𝑋) × B(𝑌 ). This shows that
B(𝑋 × 𝑌 ) ⊆ B(𝑋) × B(𝑌 ).
In particular, B(R2 ) = B(R) × B(R).

We now prove the section properties of a product space. Let 𝐸 be a subset of 𝑋 × 𝑌 . For each 𝑥 ∈ 𝑋,
define the section 𝐸 𝑥 by
𝐸 𝑥 B {𝑦 ∈ 𝑌 | (𝑥, 𝑦) ∈ 𝐸 }.
Similarly, for each 𝑦 ∈ 𝑌 , define the section 𝐸 𝑦 by

𝐸 𝑦 B {𝑥 ∈ 𝑋 | (𝑥, 𝑦) ∈ 𝐸 }.

Moreover, for a function 𝑓 on 𝑋 × 𝑌 , defines sections 𝑓 𝑥 and 𝑓 𝑦 , where 𝑥 ∈ 𝑋 and 𝑦 ∈ 𝑌 , by

𝑓 𝑥 (𝑦) B 𝑓 (𝑥, 𝑦),

and
𝑓 𝑦 (𝑥) B 𝑓 (𝑥, 𝑦).

Lemma 5.3. Let (𝑋, A) and (𝑌 , B) be measurable spaces.

1. For a measurable subset 𝐸 of 𝑋 × 𝑌 , then sections 𝐸 𝑥 and 𝐸 𝑦 are measurable for each 𝑥 ∈ 𝑋 and
each 𝑦 ∈ 𝑌 .

2. For a measurable extended-real-valued or complex-value function 𝑓 on 𝑋 × 𝑌 , the sections 𝑓 𝑥 and


𝑓 𝑦 are measurable for each 𝑥 ∈ 𝑋 and each 𝑦 ∈ 𝑌 .

Proof. Define the family F to be the collection of subsets 𝐸 of 𝑋 × 𝑌 such that 𝐸 𝑥 is measurable for
every 𝑥 ∈ 𝑋. We want to show that A × B ⊆ F. First, observe that for 𝐴 ∈ A, 𝐵 ∈ B and 𝑥 ∈ 𝑋,
(
𝐵 if 𝑥 ∈ 𝐴
( 𝐴 × 𝐵) 𝑥 =
∅ otherwise.

Therefore, F contains all measurable rectangles. Now, we show that F is a 𝜎-algebra. Clearly, 𝑋 ×𝑌 ∈ F.
Also, since 𝑦 ∈ (𝐸 𝑐 ) 𝑥 if and only if (𝑥, 𝑦) ∉ 𝐸, we have (𝐸 𝑐 ) 𝑥 = (𝐸 𝑥 ) 𝑐 . Thus, 𝐸 ∈ F implies 𝐸 𝑐 ∈ F.
Ð Ð
Now, let (𝐸 𝑛 ) be a countable subcollection of F. Then, 𝑦 ∈ 𝑛 (𝐸 𝑛 ) 𝑥 if and only if (𝑥, 𝑦) ∈ 𝑛 𝐸 𝑛 , and
Ð Ð Ð Ð
this is equivalent to the following: 𝑦 ∈ 𝑛 (𝐸 𝑛 ) 𝑥 . This shows that ( 𝑛 𝐸 𝑛 ) 𝑥 = 𝑛 (𝐸 𝑛 ) 𝑥 , so 𝑛 𝐸 𝑛 ∈ F.

47
Thus, by the definition of A × B, A × B ⊆ F. Hence, for each 𝐸 ∈ A × B and each 𝑥 ∈ 𝑋, 𝐸 𝑥 ∈ B.
Similarly, we can show the statement about 𝑦-sections.
Now, let 𝑓 be a measurable function on 𝑋 × 𝑌 . Then, for 𝑥 ∈ 𝑋 and a subset 𝐷 of 𝑌 , 𝑓 𝑥 −1 (𝐷) =
( 𝑓 −1 (𝐷)) 𝑥 . By this identity, 𝑓 𝑥 is measurable for each 𝑥 ∈ 𝑋. Similarly, 𝑓 𝑦 is measurable for each
𝑦 ∈ 𝑌. 

We now want to define the product measure 𝜇 × 𝜈 on (𝑋 × 𝑌 , A × B), where 𝜇 and 𝜈 are measures
on (𝑋, A) and (𝑌 , B), respectively. We want to require that for a measurable rectangle 𝐴 × 𝐵,

(𝜇 × 𝜈) ( 𝐴 × 𝐵) = 𝜒 𝐴×𝐵 𝑑 (𝜇 × 𝜈) = 𝜇( 𝐴)𝜈(𝐵).

Observe that 𝜒 𝐴×𝐵 (𝑥, 𝑦) = 𝜒 𝐴 (𝑥) 𝜒𝐵 (𝑦) for each pair of 𝑥 ∈ 𝑋 and 𝑦 ∈ 𝑌 , and that
∫ ∫  ∫ ∫ 
𝜇( 𝐴)𝜈(𝐵) = 𝜒 𝐴 (𝑥) 𝜒𝐵 (𝑦) 𝑑𝜈(𝑦) 𝑑𝜇(𝑥) = 𝜒 𝐴 (𝑥) 𝜒𝐵 (𝑦) 𝑑𝜇(𝑥) 𝑑𝜈(𝑦).
𝑋 𝑌 𝑌 𝑋

Now, let 𝑓 (𝑥, 𝑦) = 𝜒 𝐴×𝐵 (𝑥, 𝑦). Then, the previous identities are read as follows:
∫ ∫ ∫ ∫ ∫
𝑓 𝑑 (𝜇 × 𝜈) = 𝑓 𝑥 𝑑𝜈(𝑦)𝑑𝜇(𝑥) = 𝑓 𝑦 𝑑𝜇(𝑥)𝑑𝜈(𝑦).
𝑋 ×𝑌 𝑋 𝑌 𝑌 𝑋

We will generalize this identity for every measurable function 𝑓 whose integral over 𝑋 ×𝑌 can be defined.
In particular, if 𝑓 = 𝜒𝐸 , where 𝐸 is a measurable subset of 𝑋 × 𝑌 , then the following must hold:
∫ ∫
(𝜇 × 𝜈) (𝐸) = 𝜈(𝐸 𝑥 )𝑑𝜇(𝑥) = 𝜇(𝐸 𝑦 )𝑑𝜈(𝑦).
𝑋 𝑌

We will define the product measure 𝜇 × 𝜈 by these identities, but before that, we need to show that the
maps 𝑥 ↦→ 𝜈(𝐸 𝑥 ) and 𝑦 ↦→ 𝜇(𝐸 𝑦 ) are measurable.

Lemma 5.4. Let (𝑋, A, 𝜇) and (𝑌 , B, 𝜈) be 𝜎-finite measure spaces. Then, for every measurable subset
𝐸 of 𝑋 × 𝑌 , the maps 𝑥 ↦→ 𝜈(𝐸 𝑥 ) and 𝑦 ↦→ 𝜇(𝐸 𝑦 ) are measurable.

Proof. Assume first that 𝜈 is a finite measure. Let F be the collection of measurable subsets 𝐸 of 𝑋 × 𝑌
such that the map 𝑥 ↦→ 𝜈(𝐸 𝑥 ) is measurable. Clearly, F contains all the measurable rectangles. Note
that by the identity
( 𝐴1 × 𝐵1 ) ∩ ( 𝐴2 × 𝐵2 ) = ( 𝐴1 ∩ 𝐴2 ) × (𝐵1 ∩ 𝐵2 ),
the collection of all measurable rectangles is a 𝜋-system on 𝑋 × 𝑌 . Thus, by Dynkin’s 𝜋-𝜆 theorem, if F
is a 𝜆-system, then A × B ⊆ F, that is, A × B = F, and thus, for all measurable set 𝐸 of 𝑋 × 𝑌 , the map
𝑥 ↦→ 𝜈(𝐸 𝑥 ) is measurable.
We now show that F is a 𝜆-system. Clearly, 𝑋 ×𝑌 ∈ F. Let 𝐸, 𝐹 ∈ F with 𝐸 ⊆ 𝐹. Then, (𝐹 \ 𝐸) 𝑥 =
𝐹𝑥 \ 𝐸 𝑥 , and since 𝜈 is a finite measure, 𝜈((𝐹 \ 𝐸) 𝑥 ) = 𝜈(𝐹𝑥 ) − 𝜈(𝐸 𝑥 ). Thus, 𝐹 \ 𝐸 ∈ F. Moreover, if
Ð Ð
(𝐸 𝑛 ) is an increasing countable subfamily of F, then 𝜈(( 𝑛 𝐸 𝑛 ) 𝑥 ) = 𝜈( 𝑛 (𝐸 𝑛 ) 𝑥 ) = lim𝑛 𝜈((𝐸 𝑛 ) 𝑥 ), so
Ð
the map 𝑥 ↦→ 𝜈(( 𝑛 𝐸 𝑛 ) 𝑥 ) is measurable. Therefore, F is a 𝜆-system, and hence F = A × B.
Now, suppose that 𝜈 is 𝜎-finite. Let (𝐷 𝑛 ) be a disjoint countable subfamily of B such that each
Ð
𝜈(𝐷 𝑛 ) is finite and that 𝑛 𝐷 𝑛 = 𝑌 . For each 𝑛, define 𝜈𝑛 (𝐵) = 𝜈(𝐵 ∩ 𝐷 𝑛 ), 𝐵 ∈ B, then 𝜈𝑛 is a finite
measure on B. Therefore, for a measurable subset 𝐸 of 𝑋 × 𝑌 , each map 𝑥 ↦→ 𝜈𝑛 (𝐸 𝑥 ) is measurable, and
Í
so the map 𝑥 ↦→ 𝜈(𝐸 𝑥 ) is measurable since 𝜈(𝐸 𝑥 ) = 𝑛 𝜈𝑛 (𝐸 𝑥 ).
Under the assumption that 𝜇 is 𝜎-finite, similarly, we can prove that for a measurable set 𝐸 of 𝑋 × 𝑌 ,
the map 𝑦 ↦→ 𝜇(𝐸 𝑦 ) is measurable. 

Theorem 5.5 (Product measure). Let (𝑋, A, 𝜇) and (𝑌 , B, 𝜈) be 𝜎-finite measure spaces. Then, there
exists a unique measure 𝜇 × 𝜈 on (𝑋 × 𝑌 , A × B), called the product of 𝜇 and 𝜈, satisfying

(𝜇 × 𝜈) ( 𝐴 × 𝐵) = 𝜇( 𝐴)𝜈(𝐵) (5.1)

48
for all measurable subsets 𝐴 ∈ 𝑋 and 𝐵 ∈ 𝑌 . Moreover, for every measurable subset 𝐸 of 𝑋 × 𝑌 j,
∫ ∫
(𝜇 × 𝜈) (𝐸) = 𝜈(𝐸 𝑥 )𝑑𝜇(𝑥) = 𝜇(𝐸 𝑦 )𝑑𝜈(𝑦).
𝑋 𝑌

Proof. To show the uniqueness, observe that the collection of all measurable rectangles is a 𝜋-system,
and it generates the 𝜎-algebra A × B. Thus, since the measure 𝜇 and 𝜈 are 𝜎-finite, by Lemma 1.42, the
relation (5.1) determines a unique measure on A × B.
We now show the existence. Define the set function (𝜇 × 𝜈)1 on A × B by

(𝜇 × 𝜈)1 (𝐸) B 𝜈(𝐸 𝑥 )𝑑𝜇(𝑥).
𝑋

We need to prove that this is a measure and the relation (5.1) holds. First, note that for each 𝑥 ∈ 𝑋, since
∅ 𝑥 = ∅, (𝜇 × 𝜈)1 (∅) = 0. Moreover, if (𝐸 𝑛 ) is a disjoint countable subfamily of A × B, then since
((𝐸 𝑛 ) 𝑥 is a disjoint family,
∫ Õ∫ Õ
Ð Ð
(𝜇 × 𝜈)1 ( 𝑛 𝐸 𝑛 ) = 𝜈( 𝑛 (𝐸 𝑛 ) 𝑥 )𝑑𝜇(𝑥) = 𝜈((𝐸 𝑛 ) 𝑥 )𝑑𝜇(𝑥) = (𝜇 × 𝜈)1 (𝐸 𝑛 ).
𝑋 𝑛 𝑋 𝑛

Theerfore, (𝜇 × 𝜈)1 is a measure. Moreover, the relation (5.1) clearly holds. Similarly, if we define

(𝜇 × 𝜈)2 (𝐸) B 𝜇(𝐸 𝑦 )𝑑𝜈(𝑦), 𝐸 ∈ A × B,
𝑌

then we can prove that (𝜇 × 𝜈)2 is a measure and that the relation (5.1) holds. 

Example 5.6. Let 𝜆1 be the Lebesgue measure on (R, B(R)), and let 𝜆2 be the Lebesgue measure on
(R2 , B(R2 )). We already know that B(R2 ) = B(R) × B(R). Moreover, 𝜆2 = 𝜆1 × 𝜆1 since they agree for
all 2-dimensional rectangles (see Proposition 1.41).

5.2 Fubini’s theorem and applications


We now discuss the identities
∫ ∫ ∫ ∫ ∫
𝑓 𝑑 (𝜇 × 𝜈) = 𝑓 𝑥 𝑑𝜈(𝑦)𝑑𝜇(𝑥) = 𝑓 𝑦 𝑑𝜇(𝑥)𝑑𝜈(𝑦). (5.2)
𝑋 ×𝑌 𝑋 𝑌 𝑌 𝑋

Note that by the theorem in the previous section, the identities (5.2) hold for every characteristic function,
and hence they hold for every nonnegative measurable simple function. Thus, by using approximation
by simple functions and the monotone convergence theorem, we prove the following theorem.

Theorem 5.7 (Tonelli). Let (𝑋, A, 𝜇) and (𝑌 , B, 𝜈) be 𝜎-finite measure spaces, and let 𝑓 : 𝑋×𝑌 → [0, ∞]
be a measurable function. Then, the following hold.
∫ ∫
1. The maps 𝑥 ↦→ 𝑌 𝑓 𝑥 𝑑𝜈(𝑦) and 𝑦 ↦→ 𝑋 𝑓 𝑦 𝑑𝜇(𝑥) are measurable.

2. The identity (5.2) holds.

Proof. We already know that the assertions are true when 𝑓 is a characteristic function, and hence they
hold when 𝑓 is a nonnegative simple measurable function. Let 𝑓 : 𝑋 × 𝑌 → [0, ∞] be measurable,
and let ( 𝑓𝑛 ) be an increasing sequence of nonnegative measurable simple functions suth that 𝑓𝑛 → 𝑓
pointwise. Then, for each 𝑥 ∈ 𝑋, (( 𝑓𝑛 ) 𝑥 ) is also an increasing sequence of measurable functions such
that ( 𝑓𝑛 ) 𝑥 → 𝑓 𝑥 , and thus, by the monotone convergence theorem,
∫ ∫
lim ( 𝑓𝑛 ) 𝑥 𝑑𝜈(𝑦) = 𝑓 𝑥 𝑑𝜈(𝑦).
𝑛 𝑌 𝑌

49

Moreover, for each 𝑛, the map 𝑥 ↦→ 𝑌 ( 𝑓𝑛 ) 𝑥 𝑑𝜈(𝑦) is measurable since 𝑓𝑛 is a nonnegative measurable
∫ ∫
simple function, so the map 𝑥 ↦→ 𝑌 𝑓 𝑥 𝑑𝜈(𝑦) is also measurable. Moreover, the maps 𝑥 ↦→ 𝑌 ( 𝑓𝑛 ) 𝑥 𝑑𝜈(𝑦)
form an increasing sequence, so by the monotone convergence theorem again,
∫ ∫ ∫ ∫
𝑓 𝑥 𝑑𝜈(𝑦)𝑑𝜇(𝑥) = lim ( 𝑓𝑛 ) 𝑥 𝑑𝜈(𝑦)𝑑𝜇(𝑥).
𝑋 𝑌 𝑛 𝑋 𝑌

On the other hand, applying the monotone convergence theorem again,


∫ ∫ ∫ ∫
𝑓 𝑑 (𝜇 × 𝜈) = lim 𝑓𝑛 𝑑 (𝜇 × 𝜈) = lim ( 𝑓𝑛 ) 𝑥 𝑑𝜈(𝑦)𝑑𝜇(𝑥).
𝑋 ×𝑌 𝑛 𝑋 ×𝑌 𝑛 𝑋 𝑌

Hence, we obtain the first identity. The second identity can be proved similarly. 

We now prove the identity (5.2) for integrable functions.

Theorem 5.8 (Fubini). Let (𝑋, A, 𝜇) and (𝑌 , B, 𝜈) be 𝜎-finite measure spaces, and let 𝑓 : 𝑋 × 𝑌 →
[−∞, ∞] be an integrable function. Then, the following hold.

1. For almost all 𝑥 ∈ 𝑋, the section 𝑓 𝑥 is integrable.


∫ Similarly, for almost∫ all 𝑦 ∈ 𝑌 , the section 𝑓 is
𝑦

integrable. Thus, we can define the maps 𝑥 ↦→ 𝑌 𝑓 𝑥 𝑑𝜈(𝑦) and 𝑦 ↦→ 𝑋 𝑓 𝑦 𝑑𝜇(𝑥) for almost all 𝑥
and almost all 𝑦.

2. Moreover, the identity (5.2) holds.

Proof. By the previous theorem and the assumption,


∫ ∫ ∫
±
𝑓 𝑑 (𝜇 × 𝜈) = 𝑓 𝑥± 𝑑𝜈(𝑦)𝑑𝜇(𝑥) < ∞.
𝑋 ×𝑌 𝑋 𝑌

This means that the maps 𝑥 ↦→ 𝑌 𝑓 𝑥± 𝑑𝜈(𝑦) are integrable, and so they are finite almost everywhere. This
+ −
∫ all 𝑥 ∈ 𝑋, both 𝑓 𝑥∫ and 𝑓 𝑥 are integrable, and so 𝑓 𝑥 is integrable. Let 𝑁 be the set
means that for almost
of 𝑥 ∈ 𝑋 such that 𝑌 𝑓 𝑥+ 𝑑𝜈(𝑦) = ∞ or 𝑌 𝑓 𝑥− 𝑑𝜈(𝑦) = ∞. Then, 𝑁 is measurable and is of measure zero.
Moreover, define (∫ ∫ ∫
𝑓 𝑑𝜈 = 𝑓 + 𝑑𝜈 − 𝑓 − 𝑑𝜈 if 𝑥 ∉ 𝑁
𝑥 𝑥 𝑌 𝑥
𝐼 𝑓 (𝑥) = 𝑌 𝑌
0 if 𝑥 ∈ 𝑁.
Then, 𝐼 𝑓 is a real-valued integrable function. Moreover, since 𝑁 is of measure zero,
∫ ∫ ∫ ∫ ∫
+
𝐼 𝑓 𝑑𝜇(𝑥) = 𝑓 𝑥 𝑑𝜈 − 𝑓 𝑥− 𝑑𝜈
𝑋 𝑋 \𝑁 𝑌 𝑋 \𝑁 𝑌
∫ ∫ ∫ ∫
= 𝑓 𝑥+ 𝑑𝜈 − 𝑓 𝑥− 𝑑𝜈
∫𝑋 𝑌 𝑋 𝑌∫
+
= 𝑓 𝑑 (𝜇 × 𝜈) − 𝑓 − 𝑑 (𝜇 × 𝜈)
∫𝑋 ×𝑌 𝑋 ×𝑌

= 𝑓 𝑑 (𝜇 × 𝜈).
𝑋 ×𝑌

The remaining part can be proved similarly. 

Remark 5.9. We can extend the previous discussion to finite products of measure spaces. For 𝜎-finite
measure spaces (𝑋1 , A1 , 𝜇1 ), . . . , (𝑋𝑛 , A𝑛 , 𝜇 𝑛 ), first define the 𝜎-algebra

A1 × · · · × A 𝑛 .

50
Then, one can check that for each 𝑘 = 1, . . . , 𝑛,

(A1 × · · · × A 𝑘 ) × (A 𝑘+1 × · · · × A𝑛 ) = A1 × · · · × A𝑛 .

Moreover, the product measure 𝜇1 × · · · × 𝜇 𝑛 is defined by the unique measure on A1 × · · · × A𝑛 such that

(𝜇1 × · · · × 𝜇 𝑛 ) ( 𝐴1 × · · · × 𝐴𝑛 ) = 𝜇1 ( 𝐴1 ) . . . 𝜇 𝑛 ( 𝐴𝑛 )

for every 𝑛-tuple of measurable sets 𝐴1 ∈ 𝑋1 , . . . , 𝐴𝑛 ∈ 𝑋𝑛 . Integrals with respect to this product measure
can be evaluated by applying Tonelli’s or Fubini’s theorems repeatedly.
We now see some applications of the Tonelli’s or Fubini’s theorems.

Example 5.10. In this example, we will obtain some formulas for the measure of the region under the
graph of a function. Let (𝑋, A, 𝜇) be a 𝜎-finite measure space, and let 𝑓 : 𝑋 → [0, ∞] be a measurable
function. Define the region under the graph of 𝑓 by

𝐸 B {(𝑥, 𝑡) ∈ 𝑋 × R | 0 6 𝑡 < 𝑓 (𝑥)}.

Note that 𝐸 = (𝑋 × [0, ∞)) ∩ 𝑔 −1 ((0, ∞]), where 𝑔 : 𝑋 × R → [−∞, ∞] be the function given by
𝑔(𝑥, 𝑡) = 𝑓 (𝑥) − 𝑡, which is measurable with respect to the 𝜎-algebra A × B(R). Thus, 𝐸 is measurable
with respect to the 𝜎-algebra A × B(R).
Now, let 𝜆 be the Lebesgue measure on R. Then,
∫ ∫
(𝜇 × 𝜆) (𝐸) = 𝜆(𝐸 𝑥 )𝑑𝜇(𝑥) = 𝑓 (𝑥)𝑑𝜇(𝑥).
𝑋 𝑋

Moreover, ∫ ∫ ∞
(𝜇 × 𝜆) (𝐸) = 𝜇(𝐸 𝑡 )𝑑𝜆(𝑡) = 𝜇({ 𝑓 > 𝑡})𝑑𝜆(𝑦).
R 0

Thus, we also obtain the following identity:


∫ ∫ ∞
𝑓 (𝑥)𝑑𝜇(𝑥) = 𝜇({ 𝑓 > 𝑡})𝑑𝜆(𝑡).
𝑋 0

In fact, we can generalize this identity further. Let 𝑝 ∈ [1, ∞). Then,
∫ ∫ ∫ 𝑓 𝑝 ( 𝑥)

𝑓 𝑝 (𝑥)𝑑𝜇(𝑥) = 1𝑑𝜆(𝑡)𝑑𝜇(𝑥)
𝑋 𝑋 0
∫ ∫ 𝑓 ( 𝑥)
= 𝑝𝑠 𝑝−1 𝑑𝜆(𝑠)𝑑𝜇(𝑥)
∫𝑋 ∫0
= 𝑝𝑠 𝑝−1 𝜒 [0, 𝑓 ( 𝑥)) (𝑠)𝑑𝜆(𝑠)𝑑𝜇(𝑥)
∫𝑋 ∫R
= 𝑝𝑠 𝑝−1 𝜒 { 𝑓 >𝑠 } (𝑥) 𝜒 [0,∞) (𝑠)𝑑𝜆(𝑠)𝑑𝜇(𝑥)
∫𝑋 ∫ R
= 𝑝𝑠 𝑝−1 𝜒 { 𝑓 >𝑠 } (𝑥) 𝜒 [0,∞) (𝑠)𝑑𝜇(𝑥)𝑑𝜆(𝑠)
∫R ∞ 𝑋
= 𝑝𝑠 𝑝−1 𝜇({ 𝑓 > 𝑠})𝑑𝜆(𝑠).
0

In the second identity, we use the change of variable 𝑡 = 𝑠 𝑝 .7


7 Note that the inner integral in this identity can be regarded as a Riemann integral.

51
Example 5.11 (Convolution). Let 𝑓 and 𝑔 be two real-valued integrable functions on (R, B(R), 𝜆). Note
that the function (𝑥, 𝑡) ↦→ 𝑓 (𝑥 − 𝑡)𝑔(𝑡) is Borel measurable, and
∫ ∫ ∫
| 𝑓 (𝑥 − 𝑡)𝑔(𝑡)|𝑑 (𝜆 × 𝜆) (𝑥, 𝑡) = | 𝑓 (𝑥 − 𝑡)𝑔(𝑡)|𝑑𝜆(𝑥)𝑑𝜆(𝑡)

= k 𝑓 k 1 |𝑔(𝑡)|𝑑𝜆(𝑡)

= k 𝑓 k 1 k𝑔k 1 < ∞.

Thus, the map (𝑥, 𝑡) ↦→ 𝑓 (𝑥 − 𝑡)𝑔(𝑡) is integrable, and so for almost all 𝑥, the map 𝑡 ↦→ 𝑓 (𝑥 − 𝑡)𝑔(𝑡) is
integrable. Therefore, we can define the function

( 𝑓 ∗ 𝑔) (𝑥) = 𝑓 (𝑥 − 𝑡)𝑔(𝑡)𝑑𝜆(𝑡)

for almost all 𝑥. By redefining this function on a set of measure zero, we define the function 𝑓 ∗ 𝑔 on the
whole 𝑋, which is called the convolution of 𝑓 and 𝑔. Note that for almost all 𝑥,

| 𝑓 ∗ 𝑔(𝑥)| 6 | 𝑓 (𝑥 − 𝑡)𝑔(𝑡)|𝑑𝜆(𝑡),

and thus, by the previous calculation,

k 𝑓 ∗ 𝑔k 1 6 k 𝑓 k 1 k𝑔k 1 .

52
6 Borel measures on locally compact Hausdorff spaces
We have seen that the Lebesgue measure on a Euclidean space is constructed from the Lebesgue outer
measure. However, if we regard the Lebesgue measure as a Borel measure, a different method of
construction is possible. In this chapter, we discuss a characterization of Borel measures (with additional
properties) on a locally compact Hausdorff space, which is called the Riesz representation theorem.

6.1 Locally compact Hausdorff spaces


We now recall some basic topological notions. The term "neighborhood" always denotes an open
neighborhood. A Hausdorff space 𝑋 is said to be locally compact if for every point 𝑥 ∈ 𝑋, 𝑥 has a
neighborhood whose closure is compact. We briefly call locally compact Hausdorff spaces LCH spaces.
In general, an LCH space is not a normal space, but a weak version of Urysohn’s lemma holds, and we
will use this lemma in our discussion.
First, in a Hausdorff space, a pair of a compact set and a point can be separated by neighborhoods.

Proposition 6.1. Let 𝐾 be a compact subset in a Hausdorff space 𝑋, and let 𝑝 ∉ 𝐾. Then, there are
disjoint neighborhoods 𝑈 and 𝑉 of 𝑝 and 𝐾, respectively.

Proof. Since 𝑋 is a Hausdorff space, for each 𝑥 ∈ 𝐾, there exist disjoint neighborhoods 𝑈 𝑥 and 𝑉𝑥 of
𝑝 and 𝑥, respectively. Then, the collection (𝑉𝑥 ) 𝑥 ∈𝐾 forms an open cover of 𝑋, and since 𝐾 is compact,
there is a finite subcover (𝑉𝑥𝑖 )𝑖=1,..., 𝑁 of (𝑉𝑥 ) 𝑥 ∈𝐾 . Now, define
𝑁
Ø 𝑁
Ù
𝑉 B 𝑉 𝑥𝑖 and 𝑈 B 𝑈𝑥 𝑗 .
𝑖=1 𝑗=1

Then, 𝑈 and 𝑉 are neighborhoods of 𝑝 and 𝐾, respectively. Moreover, 𝑈 ∩ 𝑉 = ∅. 

The following corollary is immediate.

Corollary 6.2. Let 𝑋 be a Hausdorff space. Then, every compact subset of 𝑋 is closed.

Proof. Let 𝐾 ⊆ 𝑋 be compact. By the previous proposition, for each 𝑝 ∈ 𝑋 \ 𝐾, there exists a
neighborhood of 𝑈 of 𝑝 such that 𝑈 ⊆ 𝑋 \ 𝐾. This shows that 𝑋 \ 𝐾 is open, so 𝐾 is closed. 

It is known that every compact Hausdorff space is normal. However, there is a locally compact
Hausdorff space which is not normal. For example, see the counterexamples 65, 93 and 106 of [SS95].
However, we can show the following:

Proposition 6.3. An LCH space is completely regular.

Proof. Let 𝑋 be an LCH space. By the one-point compactification, there exists a compact space 𝑌 such
that 𝑋 is an open subset of 𝑌 . Then, since 𝑋 is locally compact, 𝑌 is Hausdorff. Thus, 𝑌 is a compact
Hausdorff space, so it is normal and Hausdorff. Thus, in particular, 𝑌 is completely regular (by Urysohn’s
lemma). Therefore, 𝑋 is also completely regular by the hereditary property for complete regularity. 

Corollary 6.4. An LCH space is regular.

Remark 6.5. We showed the regularity of an LCH space by the compactification, but we can show the
property directly. Let 𝑋 be an LCH space, and let 𝐶 ⊆ 𝑋 be a closed subset. Take 𝑝 ∈ 𝑋 \ 𝐶. From the
following lemma, there exists a neighborhood 𝑈 of 𝑝 such that 𝑈 is compact and that 𝑝 ∈ 𝑈 ⊆ 𝑈 ⊆ 𝑋 \𝐶.
Since 𝑋 is Hausdorff, 𝑈 is closed, so 𝑉 = 𝑋 \ 𝑈 is open. Now, observe that 𝑈 ∩ 𝑉 = ∅ and that 𝑈 and 𝑉
are neighborhoods of 𝑝 and 𝐶, respectively.

53
Lemma 6.6. Let 𝑋 be an LCH space, and let 𝑝 ∈ 𝑋. Suppose 𝑈 is a neighborhood of 𝑝. Then, there
exists a neighborhood 𝑉 of 𝑝 such that 𝑉 is compact and that 𝑝 ⊆ 𝑉 ⊆ 𝑉 ⊆ 𝑈.
Proof. Since 𝑋 is locally compact, there exists a neighborhood 𝑁 of 𝑝 such that 𝑁 is compact. Now,
consider the neighborhood 𝑈 ∩ 𝑁 of 𝑝. Then, 𝑈 ∩ 𝑁 ⊆ 𝑈 ∩ 𝑁 ⊆ 𝑁. Since 𝑁 is compact and closed,
𝑈 ∩ 𝑁 is also compact.
Now, write 𝑌 = 𝑈 ∩ 𝑁. Then, 𝑌 is a compact Hausdorff space. Note that 𝑌 is open in 𝑌 . Thus, 𝑌 \ 𝑌
is closed in 𝑌 , and so 𝑌 \ 𝑌 is compact in 𝑌 since 𝑌 is a compact space. By Proposition 6.1, there are
disjoint neighborhoods 𝑉 and 𝑊 of 𝑝 and 𝑌 \ 𝑌 = 𝜕𝑌 , respectively, in 𝑌 . Then, 𝑝 ∈ 𝑉 ⊆ 𝑌 , and 𝑉 is open
in 𝑌 . Since 𝑌 is open in 𝑋, 𝑉 is open in 𝑋. Moreover, since 𝑌 is closed in 𝑋, the closure of 𝑉 in 𝑌 is
exactly 𝑉, and in addition, 𝑉 is compact since 𝑌 is compact. Now, observe that 𝑉 ⊆ 𝑌 \ 𝑊 and that 𝑌 \ 𝑊
is closed in 𝑌 . Therefore, the closure of 𝑉 in 𝑌 is contained in 𝑌 \ 𝑊, and thus, 𝑉 ⊆ 𝑌 \ 𝑊. In particular,
𝑉 ∩ 𝜕𝑌 = ∅ and 𝑉 ⊆ 𝑌 . Therefore, 𝑉 ⊆ 𝑌 . This completes the proof. 
We can easily generalize this lemma as follows.
Proposition 6.7. Let 𝑋 be an LCH space, and let 𝐾 be a compact subset of 𝑋. Suppose 𝑈 is a
neighborhood of 𝐾. Then, there exists a neighborhood of 𝐾 such that 𝑉 is compact and that 𝐾 ⊆ 𝑉 ⊆
𝑉 ⊆ 𝑈.
Proof. For each 𝑥 ∈ 𝐾, take a neighborhood 𝑉𝑥 of 𝑥 as in the previous lemma. Then, (𝑉𝑥 ) 𝑥 ∈𝐾 is an
𝑁 . Now, define 𝑉 = Ð 𝑁 𝑉 . Then, 𝑉 is a
open cover of 𝐾, and thus, there exists a finite subcover (𝑉𝑥𝑖 )𝑖=1 𝑖=1 𝑥𝑖
Ð𝑁
neighborhood of 𝐾, and moreover, 𝑉 = 𝑖=1 𝑉𝑥𝑖 which is a compact set contained in 𝑈. 
We are now ready to prove Urysohn’s lemma for LCH spaces. Before proving that, we introduce
some notations. Let 𝑋 be a topological space. For a pair of a compact set 𝐾 and a continuous function 𝑓
on 𝑋 with compact support for which 0 6 𝑓 6 1, we write
𝐾≺ 𝑓
whenever 𝑓 = 1 on 𝐾. Moreover, for a pair of an open set 𝑈 and a continuous function 𝑓 on 𝑋 with
compact support for which 0 6 𝑓 6 1, we write
𝑓 ≺𝑈
whenever the support of 𝑓 is contained in 𝑈. Note that for a complex-valued function on 𝑋, the support
of 𝑓 is defined by
supp 𝑓 B {𝑥 ∈ 𝑋 | 𝑓 (𝑥) ≠ 0}.
Moreover, 𝐶 (𝑋) and 𝐶𝑐 (𝑋) denote the space of continuous complex-valued functions on 𝑋 and the space
of those functions with compact support, respectively.
Proposition 6.8 (Urysohn’s lemma for LCH spaces). Let 𝑋 be an LCH space. Suppose 𝐾 ⊆ 𝑈, where 𝐾
is a compact set and 𝑈 is an open set. Then, there exists 𝑓 ∈ 𝐶𝑐 (𝑋) with 0 6 𝑓 6 1 such that
𝐾 ≺ 𝑓 ≺ 𝑈.
Proof. By the previous proposition, there exists a neighborhood 𝑉 of 𝐾 such that 𝑉 is compact and that
𝐾 ⊆ 𝑉 ⊆ 𝑉 ⊆ 𝑈. Note that 𝑉 is a compact Hausdorff space, and in particular, it is normal. Moreover, 𝐾
is closed in 𝑉, and also, 𝑉 \ 𝑉 = 𝜕𝑉 is closed in 𝑉, which is disjoint to 𝐾. Thus, by Urysohn’s lemma for
normal spaces, there exists 𝑓 ∈ 𝐶 (𝑉) with 0 6 𝑓 6 1 such that 𝐾 ≺ 𝑓 and supp 𝑓 ⊆ 𝑉 ⊆ 𝑈.
We extend 𝑓 to 𝑋 by setting 𝑓 = 0 on 𝑋 \ 𝑉. We need to show that 𝑓 is continuous. Note that since
𝑓 ( 𝐴) = 𝑓 −1 ( 𝐴 ∩ [0, 1]) for a subset 𝐴 of R, 𝑓 is continuous if and only if for every closed subset 𝐶 of
−1

[0, 1], 𝑓 −1 (𝐶) is closed. Let 𝐶 ⊆ [0, 1] be closed. First, assume that 0 ∉ 𝐶. Then, 𝑓 −1 (𝐶) = ( 𝑓 | 𝑉 ) −1 (𝐶),
and this is a closed set in 𝑉, and thus, 𝑓 −1 (𝐶) is closed in 𝑋 since 𝑉 is closed. Now, assume that 0 ∈ 𝐶.
Then, 𝑓 −1 (𝐶) = ( 𝑓 | 𝑉 ) −1 (𝐶) ∪ (𝑋 \ 𝑉). Note that since 𝑓 | 𝜕𝑉 = 0, 𝜕𝑉 ⊆ ( 𝑓 | 𝑉 ) −1 (𝐶). Therefore,
𝑓 −1 (𝐶) = ( 𝑓 | 𝑉 ) −1 (𝐶) ∪ (𝑋 \ 𝑉), and so 𝑓 −1 (𝐶) is closed in 𝑋. Hence, 𝑓 is continuous. 

54
Remark 6.9. This version of Urysohn’s lemma provides another proof of the complete regularity of an
LCH space 𝑋. Indeed, if 𝐶 is closed and if 𝑝 ∉ 𝐶, then let 𝑈 = 𝑋 \ 𝐶, and 𝐾 = {𝑝}. Note that 𝐾 is
compact since it is a closed subset of a compact set. By applying Urysohn’s lemma to this setting, there
is 𝑓 ∈ 𝐶𝑐 (𝑋) with 0 6 𝑓 6 1 such that {𝑝} ≺ 𝑓 ≺ 𝑋 \ 𝐶. This means that 𝑓 ( 𝑝) = 1 and 𝑓 = 0 on 𝐶.
We can also prove the Tietze extension theorem for LCH spaces.

Proposition 6.10 (Tietze extension theorem for LCH spaces). Let 𝑋 be an LCH space, and let 𝐾 be a
compact set in 𝑋. Suppose 𝑓 ∈ 𝐶 (𝐾). Then, there exists 𝐹 ∈ 𝐶 (𝑋) such that 𝐹 | 𝐾 = 𝑓 . Moreover, we
can take 𝐹 so that it vanishes outside a compact set.

Proof. Take a neighborhood 𝑈 of 𝐾 for which 𝑈 is compact. Then, 𝑈 is a compact Hausdorff space, so
it is a normal space. Thus, by the Tietze extension theorem for normal spaces, there exists 𝑔 ∈ 𝐶 (𝑈)
such that 𝑔| 𝐾 = 𝑓 . Moreover, by applying Urysohn’s lemma to the closed sets 𝐾 and 𝑈 \ 𝑈 = 𝜕𝑈 in 𝑈,
there exists ℎ ∈ 𝐶 (𝑈) with 0 6 ℎ 6 1 such that ℎ = 1 on 𝐾 and ℎ = 0 on 𝜕𝑈. Now, define 𝐹 ∈ 𝐶 (𝑈)
by 𝐹 = ℎ𝑔. Then, still 𝐹 | 𝐾 = 𝑓 , and we also have 𝐹 = 0 on 𝜕𝑈. Now, extend 𝐹 to 𝑋 by letting 𝐹 = 0
on 𝑋 \ 𝑈. Then, as in the proof of Urysohn’s lemma for LCH spaces, we can show that 𝐹 is continuous.
Moreover, in this construction, 𝐹 = 0 outside the compact set 𝑈. 

Remark 6.11. The proof shows that if 𝑉 is a neighborhood of 𝐾, then we can take 𝐹 so that supp 𝐹 ⊆ 𝑉.
Indeed, for a neighborhood 𝑉 of 𝐾, take a neighborhood 𝑈 of 𝐾 such that 𝐾 ⊆ 𝑈 ⊆ 𝑈 ⊆ 𝑉, and apply
the proof. Then, 𝐹 = 0 outside 𝑈, and so supp 𝐹 ⊆ 𝑈 ⊆ 𝑉.
Using Urysohn’s lemma, we obtain a partition of unity subordinate to a finite open cover of a compact
set.

Definition 6.12. Let 𝑋 be a topological space, and let 𝐴 be a subset of 𝑋. A collection (ℎ 𝛼 ) of continuous
functions on 𝑋 with 0 6 ℎ 𝛼 6 1 is said to be a partition of unity on 𝐴 if the following hold:

1. (ℎ 𝛼 ) is locally finite, that is, for each 𝑥 ∈ 𝑋, there exists a neighborhood 𝑈 of 𝑥 such that only
finitely many ℎ 𝛼 ’s are nonzero on 𝑈.
Í
2. 𝛼 ℎ 𝛼 (𝑥) = 1 for every 𝑥 ∈ 𝐴.

Moreover, for an open cover (𝑈 𝛼 ) of 𝐴, we say that a partition of unity (ℎ 𝛼 ) is subordinate to the open
covr (𝑈 𝛼 ) if for each 𝛼, ℎ 𝛼 ≺ 𝑈 𝛼 .

Proposition 6.13. Let 𝑋 be an LCH space, and let 𝐾 be a compact subset of 𝑋. Suppose (𝑈𝑖 )𝑖=1 𝑁 is a

finite open cover of 𝐾. Then, there is a partition of unity (ℎ𝑖 )𝑖=1 on 𝐾 subordinate to the cover (𝑈𝑖 )𝑖=1
𝑁 𝑁

consisting of compactly supported functions.

Proof. For each 𝑥 ∈ 𝐾, there exists 𝑖(𝑥) ∈ {1, . . . , 𝑁 } such that 𝑥 ∈ 𝑈𝑖 ( 𝑥) . Then, there exists a
neighborhood 𝑉𝑥 of 𝑥 such that 𝑉𝑥 is compact and that 𝑉𝑥 ⊆ 𝑈𝑖 ( 𝑥) . Note that (𝑉𝑥 ) 𝑥 ∈𝐾 is an open cover
of 𝐾, and since 𝐾 is compact, there exists a finite subcover (𝑉𝑥 𝑗 ) 𝑀
𝑗=1 . For each 𝑖 = 1, . . . , 𝑁, define
Ø
𝑊𝑖 B 𝑉𝑥 𝑗 .
𝑖 ( 𝑥 𝑗 )=𝑖

Then, by the definition, for every 𝑖, 𝐹𝑖 B 𝑊𝑖 is compact, and 𝐹𝑖 ⊆ 𝑈𝑖 . Moreover, (𝑊𝑖 )𝑖=1
𝑁 forms an open

cover of 𝐾.
For each 𝑖 = 1, . . . , 𝑁, by applying Urysohn’s lemma to 𝐹𝑖 and 𝑈𝑖 , there exists 𝑔𝑖 ∈ 𝐶𝑐 (𝑋) with
Í𝑁
0 6 𝑔𝑖 6 1 such that 𝐹𝑖 ≺ 𝑔𝑖 ≺ 𝑈𝑖 . Then, 𝑖=1 𝑔𝑖 > 1 > 0 since 𝐹𝑖 ’s cover 𝐾. Now, define the open
Í𝑁
subset 𝑂 of 𝑋 given by 𝑂 = {𝑥 ∈ 𝑋 | 𝑖=1 𝑔𝑖 (𝑥) > 0}. Then, 𝐾 ⊆ 𝑂, and by Urysohn’s lemma again,
there exists 𝑓 ∈ 𝐶𝑐 (𝑋) with 0 6 𝑓 6 1 such that 𝐾 ≺ 𝑓 ≺ 𝑂. Observe the following: if 𝑥 ∈ 𝑂,
Í𝑁 Í𝑁
𝑖=1 𝑔𝑖 (𝑥) > 0 and 0 6 𝑓 (𝑥) 6 1, and if 𝑥 ∉ 𝑂, then 𝑖=1 𝑔𝑖 (𝑥) = 0 and 𝑓 (𝑥) = 0. Moreover, on 𝐾,
𝑓 = 1.

55
Í 𝑁 +1
Now, define 𝑔 𝑁 +1 = 1 − 𝑓 . Then, 𝑖=1 𝑔𝑖 > 0 everywhere. For each 𝑖 = 1, . . . , 𝑁, let
𝑔𝑖
ℎ𝑖 B Í 𝑁 +1 .
𝑖=1 𝑔𝑖
Í𝑁
Then, ℎ𝑖 ∈ 𝐶𝑐 (𝑋) and ℎ𝑖 ≺ 𝑈𝑖 for every 𝑖, and for 𝑥 ∈ 𝐾, since 𝑔 𝑁 +1 (𝑥) = 1 − 1 = 0, 𝑖=1 ℎ𝑖 (𝑥) = 1.
This completes the proof. 

6.2 Riesz representation theorem


Let 𝑋 be a topological space, and let 𝜇 be a Borel measure on 𝑋 for which 𝜇(𝐾) < ∞ for every compact
set 𝐾 in 𝑋. Then, every 𝑓 ∈ 𝐶𝑐 (𝑋) is integrable, and so we can define a linear functional on 𝐶𝑐 (𝑋) as
follows: ∫
𝑓 ↦→ 𝑓 𝑑𝜇.
𝑋

Note that this functional satisfies positivity, namely, if 𝑓 > 0, then 𝑋 𝑓 𝑑𝜇 > 0. In this section, we will
prove that every positive linear functional on 𝐶𝑐 (𝑋) arises in this manner if 𝑋 is locally compact and
Hausdorff.

Definition 6.14. Let 𝜇 be a Borel measure on an LCH space 𝑋, and let 𝐸 be a Borel set in 𝑋. 𝜇 is said
to be outer regular on 𝐸 if

𝜇(𝐸) = inf{𝜇(𝑈) | 𝐸 ⊆ 𝑈 and 𝑈 is open}.

In addition, 𝜇 is said to be inner regular on 𝐸 if

𝜇(𝐸) = sup{𝜇(𝐾) | 𝐾 ⊆ 𝐸 and 𝐾 is compact}.

If 𝜇 is both outer and inner regular on every Borel set in 𝑋, then 𝜇 is called regular.

However, for the uniqueness of a measure constructed by the Riesz representation theorem, a condition
weaker than regularity is required. To formulate that condition, we give the following definition.

Definition 6.15. Let 𝜇 be a Borel measure on an LCH space. 𝜇 is called a Radon measure if the following
hold.

1. 𝜇(𝐾) < ∞ for every compact set 𝐾.

2. 𝜇 is outer regular on each Borel set.

3. 𝜇 is inner regular on each open set.

Theorem 6.16 (Riesz representation theorem). Let 𝑋 be an LCH space, and let Λ be a positive
∫ linear
functional on 𝐶𝑐 (𝑋). Then, there exists a unique Radon measure 𝜇 on 𝑋 such that Λ 𝑓 = 𝑓 𝑑𝜇 for all
𝑓 ∈ 𝐶𝑐 (𝑋).

We will prove this theorem in the following propositions.



Proposition 6.17 (Uniqueness). A Radon measure 𝜇 on 𝑋 such that Λ 𝑓 = 𝑓 𝑑𝜇 for every 𝑓 ∈ 𝐶𝑐 (𝑋)
is at most one.

Proof. Let 𝜇 be such a Radon measure. Considering the outer regularity for Borel sets, we determine the
measure of an open set. Let 𝑈 be an open set in 𝑋. Observe that if 𝑓 ≺ 𝑈 for 𝑓 ∈ 𝐶𝑐 (𝑋), then 𝑓 6 𝜒𝑈 .
Then, by integration, we obtain that Λ 𝑓 6 𝜇(𝑈) for every 𝑓 ∈ 𝐶𝑐 (𝑋) with 𝑓 ≺ 𝑈. This means that

sup{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝑓 ≺ 𝑈} 6 𝜇(𝑈).

56
Moreover, we have inner regularity for open sets. For a compact set 𝐾 with 𝐾 ⊆ 𝑈, using Urysohn’s
lemma, take 𝑓 ∈ 𝐶𝑐 (𝑋) with 𝐾 ≺ 𝑓 ≺ 𝑈. Then, by integration again, we have

𝜇(𝐾) 6 Λ 𝑓 ,

and in particular,
𝜇(𝐾) 6 sup{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝑓 ≺ 𝑈}.
This holds for an arbitrary compact set 𝐾 with 𝐾 ⊆ 𝑈, so by the inner regularity, we have

𝜇(𝑈) 6 sup{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝑓 ≺ 𝑈}.

This shows that


𝜇(𝑈) = sup{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝑓 ≺ 𝑈}. (6.1)
By the outer regularity, we have the uniqueness. 

We use the identity (6.1) to construct a desired Radon measure on 𝑋. Define 𝜇 for open sets 𝑈 by (6.1).
Then, we now use the standard method to construct a measure. Define a set function 𝜇∗ : P (𝑋) → [0, ∞]
by
𝜇∗ (𝐸) B inf{𝜇(𝑈) | 𝐸 ⊆ 𝑈, where 𝑈 is open}.
Note that 𝜇∗ (𝑈) = 𝜇(𝑈) for every open set 𝑈. To show this, suppose 𝑈 ⊆ 𝑉, where 𝑈 and 𝑉 are
open. Then, for 𝑓 ∈ 𝐶𝑐 (𝑋) with 𝑓 ≺ 𝑈, we also have 𝑓 ≺ 𝑉. This shows that 𝜇(𝑈) 6 𝜇(𝑉), and so
𝜇(𝑈) 6 𝜇∗ (𝑈). Moreover, since 𝑈 ⊆ 𝑈, 𝜇∗ (𝑈), and hence, 𝜇∗ (𝑈) = 𝜇(𝑈).
We are asked to prove the following proposition.

Proposition 6.18. 𝜇∗ is an outer measure on 𝑋, and every Borel set is 𝜇∗ -measurable.

Proof. First, 𝜇∗ (∅) = 0 since the only continuous function with empty support is the zero function.
Moreover, the monotonicity of 𝜇∗ is clear by the nature of infima. We need to show countable subadditivity.
Ð
We now show countable additivity for open sets. Let (𝑈𝑛 ) be a sequence of open sets. Let 𝑈 = 𝑛 𝑈𝑛 ,
and let 𝑓 ≺ 𝑈. Then, 𝐾 = supp 𝑓 is compact with 𝐾 ⊆ 𝑈, so (𝑈𝑛 ) is an open cover of 𝐾. By compactness,
there exists a finite subcover (𝑈𝑛𝑘 ) 𝑘=1
𝑁 . Then, there exists a partition of unity (ℎ ) 𝑁 on 𝐾 with ℎ ≺ 𝑈 .
𝑘 𝑘=1 𝑘 𝑛𝑘
Í𝑁 Í𝑁
Since 𝑘=1 ℎ 𝑘 = 1 on 𝐾 = supp 𝑓 , we have 𝑓 = 𝑘=1 𝑓 ℎ 𝑘 . Moreover, 𝑓 ℎ 𝑘 ≺ 𝑈 𝑘 . Therefore,
𝑁
Õ 𝑁
Õ ∞
Õ
Λ𝑓 = Λ( 𝑓 ℎ 𝑘 ) 6 𝜇(𝑈𝑛𝑘 ) 6 𝜇(𝑈𝑛 ).
𝑘=1 𝑘=1 𝑛=1
Í
This hols for an arbitrary 𝑓 ∈ 𝐶𝑐 (𝑋) with 𝑓 ≺ 𝑈, we have 𝜇(𝑈) 6 𝑛 𝜇(𝑈𝑛 ).
Now, let (𝐸 𝑛 ) be an arbitrary sequence of subsets of 𝑋. We may assume that 𝑛 𝜇∗ (𝐸 𝑛 ) < ∞. In
Í
particular, 𝜇∗ (𝐸 𝑛 ) < ∞ for every 𝑛. Let 𝜖 > 0 be given. For each 𝑛, take an open set 𝑈𝑛 with 𝐸 𝑛 ⊆ 𝑈𝑛
such that
𝜖
𝜇(𝑈𝑛 ) < 𝜇∗ (𝐸 𝑛 ) + 𝑛 .
2
Then, by subadditivity for open sets,
Õ Õ Õ 𝜖 Õ
𝜇∗ (𝐸 𝑛 ) + 𝜇∗ (𝐸 𝑛 ) + 𝜖 .
Ð
𝜇 ( 𝑛 𝑈𝑛 ) 6 𝜇(𝑈𝑛 ) 6 𝑛
=
𝑛 𝑛 𝑛
2 𝑛
Ð Ð
Moreover, since 𝑛 𝐸𝑛 ⊆ 𝑛 𝑈𝑛 , by monotonicity,
Õ
𝜇∗ ( 𝜇∗ (𝐸 𝑛 ) + 𝜖 .
Ð
𝑛 𝐸𝑛) 6
𝑛

Since 𝜖 is arbitrary, we have subadditivity.

57
We now show that every Borel set is 𝜇∗ -measurable. It suffices to prove that every open set is
𝜇∗ -measurable.
Let 𝑈 be an open set. We want to show that for every subset 𝐴 of 𝑋 with 𝜇∗ ( 𝐴) < ∞,

𝜇∗ ( 𝐴) > 𝜇∗ ( 𝐴 ∩ 𝑈) + 𝜇∗ ( 𝐴 \ 𝑈).

First, suppose that such a subset 𝐴 is open. Then, 𝐴 ∩ 𝑈 is open. Let 𝜖 > 0 be given. Then, there exists
𝑓 ≺ 𝐴 ∩ 𝑈 such that
𝜇∗ ( 𝐴 ∩ 𝑈) − 𝜖 < Λ 𝑓
Moreover, since supp 𝑓 ⊆ 𝐴 ∩ 𝑈 ⊆ 𝑈, 𝜇∗ ( 𝐴 \ 𝑈) 6 𝜇∗ ( 𝐴 \ supp 𝑓 ). Now, also choose 𝑔 ≺ 𝐴 \ supp 𝑓
such that
𝜇∗ ( 𝐴 \ supp 𝑓 ) − 𝜖 < Λ𝑔.
Now, observe that

𝜇∗ ( 𝐴 ∩ 𝑈) + 𝜇∗ ( 𝐴 \ 𝑈) 6 𝜇 ∗ ( 𝐴 ∩ 𝑈) + 𝜇∗ ( 𝐴 \ supp 𝑓 ) < Λ( 𝑓 + 𝑔) + 2𝜖,

but since 𝑓 + 𝑔 ≺ 𝐴 (check!), we have Λ( 𝑓 + 𝑔) 6 𝜇∗ ( 𝐴). Thus, we obtain

𝜇∗ ( 𝐴 ∩ 𝑈) + 𝜇∗ ( 𝐴 \ 𝑈) < 𝜇∗ ( 𝐴) + 2𝜖 .

Since 𝜖 > 0 is arbitrary, we obtain the desired inequality.


Now, assume that 𝐴 is an arbitrary subset of 𝑋. Let 𝜖 > 0 be given. Then, choose an open subset 𝑉
with 𝐴 ⊆ 𝑉 such that
𝜇∗ ( 𝐴) + 𝜖 > 𝜇(𝑉).
Then,

𝜇∗ ( 𝐴) + 𝜖 > 𝜇(𝑉) > 𝜇∗ (𝑉 ∩ 𝑈) + 𝜇∗ (𝑉 \ 𝑈)


> 𝜇∗ ( 𝐴 ∩ 𝑈) + 𝜇∗ ( 𝐴 \ 𝑈).

Since 𝜖 > 0 is arbitrary, we obtain the required inequality again.


This completes the proof. 

By the previous propositions and Carathéodory theorem, the restriction 𝜇 of 𝜇∗ on the Borel 𝜎-algebra
on 𝑋 is a measure.8 Then, by the definition of 𝜇∗ , 𝜇 is outer regular. Therefore, to show that 𝜇 is a Radon
measure, it remains to prove that 𝜇 is inner regular for open sets and that 𝜇 is finite for compact sets. For
showing these, we prove the following proposition.

Proposition 6.19. For a compact set 𝐾,

𝜇(𝐾) = inf{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝐾 ≺ 𝑓 }.

Proof. Let 𝑈 be an open set with 𝐾 ⊆ 𝑈. Then, by Urysohn’s lemma, there exists 𝑓 ∈ 𝐶𝑐 (𝑋) with
𝐾 ≺ 𝑓 ≺ 𝑈. Then, by the definition of 𝜇(𝑈), Λ 𝑓 6 𝜇(𝑈). Thus,

inf{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝐾 ≺ 𝑓 } 6 𝜇(𝑈),

and by the outer regularity on 𝐾,

inf{Λ 𝑓 | 𝑓 ∈ 𝐶𝑐 (𝑋) and 𝐾 ≺ 𝑓 } 6 𝜇(𝐾).

To obtain the reverse inequality, let 𝐾 ≺ 𝑓 . Define for each 𝜖 ∈ (0, 1),

𝑈 𝜖 B {𝑥 ∈ 𝑋 | 𝑓 (𝑥) > 1 − 𝜖 }.
8 Note that we have already defined 𝜇 for open sets, but since this coincides with 𝜇∗ for open sets, there is no conflict of
notation.

58
Then, each 𝑈 𝜖 is open and satisfies 𝐾 ⊆ 𝑈 𝜖 . In particular, 𝜇(𝐾) 6 𝜇(𝑈 𝜖 ) for each 𝜖. We now estimate
each 𝜇(𝑈 𝜖 ). Let 𝑔 ≺ 𝑈 𝜖 . Then,
1
𝑓 > 𝑔,
1−𝜖
so we obtain
1
Λ 𝑓 > Λ𝑔
1−𝜖
by the positivity of Λ. In particular,
1
𝜇(𝑈 𝜖 ) 6 Λ𝑓
1−𝜖
for each 𝜖. This shows that
1
𝜇(𝐾) 6 Λ𝑓,
1−𝜖
and by letting 𝜖 → 0, we have 𝜇(𝐾) 6 Λ 𝑓 . Hence, we obtain the desired identity. 

Proposition 6.20. 𝜇 is a Radon measure on 𝑋.

Proof. For finiteness on compact sets, let 𝐾 be a compact set, and let 𝐾 ≺ 𝑓 . Then, since Λ 𝑓 ∈ R,
𝜇(𝐾) 6 Λ 𝑓 < ∞. Thus, it remains to show inner regularity on open sets. Let 𝑈 be an open set. It
suffices to show the following inequality:

𝜇(𝑈) 6 sup{𝜇(𝐾) | 𝐾 ⊆ 𝑈 and 𝐾 is compact}.

Let 𝛼 < 𝜇(𝑈). Then, there exists 𝑓 ∈ 𝐶𝑐 (𝑋) with 𝑓 ≺ 𝑈 such that

𝛼 < Λ𝑓

Note that supp 𝑓 is compact and supp 𝑓 ⊆ 𝑈. Now, let supp 𝑓 ≺ 𝑔, then since 𝑔 > 𝑓 , Λ𝑔 > Λ 𝑓 . This
shows that
Λ 𝑓 6 𝜇(supp 𝑓 ).
Moreover, since supp 𝑓 ∈ 𝑈,

𝜇(supp 𝑓 ) 6 sup{𝜇(𝐾) | 𝐾 ⊆ 𝑈 and 𝐾 is compact}.

Thus, we obtain the inequality

𝛼 < sup{𝜇(𝐾) | 𝐾 ⊆ 𝑈 and 𝐾 is compact}.

Since 𝛼 < 𝜇(𝑈) is arbitrary, we obtain the required inequality. 

The following proposition completes our proof of the theorem.



Proposition 6.21. For all 𝑓 ∈ 𝐶𝑐 (𝑋), Λ 𝑓 = 𝑓 𝑑𝜇.

Proof. We may assume that 𝑓 ∈ 𝐶𝑐 (𝑋) is nonnegative. Let 𝜖 > 0, and for each 𝑛, define a function
𝑓𝑛 : 𝑋 → R by


 0 if 𝑓 (𝑥) 6 (𝑛 − 1)𝜖


𝑓𝑛 (𝑥) B 𝑓 (𝑥) − (𝑛 − 1)𝜖 if (𝑛 − 1)𝜖 < 𝑓 (𝑥) 6 𝑛𝜖

if 𝑛𝜖 < 𝑓 (𝑥).

𝜖

Then, since 𝑓 is continuous, each 𝑓𝑛 is also continuous. In fact, each 𝑓𝑛 belongs to 𝐶𝑐 (𝑋). Indeed,
𝑓𝑛 (𝑥) > 0 implies 𝑓 (𝑥) > (𝑛−1)𝜖 > 0, so supp 𝑓𝑛 is a closed subset of the compact set supp 𝑓 . Moreover,
Í
since 𝑓 is bounded, there exists 𝑁 such that 𝑓𝑛 = 0 for all 𝑛 > 𝑁. Now, observe that 𝑓 = 𝑛 𝑓𝑛 , and for
each 𝑛, if we write 𝐾𝑛 B {𝑥 ∈ 𝑋 | 𝑓 (𝑥) > 𝑛𝜖 }, then

𝜖 𝜒𝐾𝑛 6 𝑓𝑛 6 𝜖 𝜒𝐾𝑛−1 .

59
Thus, by integration, for each 𝑛,

𝜖 𝜇(𝐾𝑛 ) 6 𝑓𝑛 𝑑𝜇 6 𝜖 𝜇(𝐾𝑛−1 ).

Moreover, if 𝑈 is an open set with 𝐾𝑛−1 ⊆ 𝑈, then 𝑓𝑛 /𝜖 ≺ 𝑈, and so Λ 𝑓𝑛 /𝜖 6 𝜇(𝑈). This implies that
Λ 𝑓𝑛 6 𝜖 𝜇(𝐾𝑛−1 ). Also, since 𝐾𝑛 ≺ 𝑓𝑛 /𝜖, 𝜖 𝜇(𝐾𝑛 ) 6 Λ 𝑓𝑛 , that is,

𝜖 𝜇(𝐾𝑛 ) 6 Λ 𝑓𝑛 6 𝜖 𝜇(𝐾𝑛−1 )
Í𝑁
for every 𝑛. Since 𝑓 = 𝑛=1 𝑓 𝑛 , we have
𝑁
Õ ∫ 𝑁
Õ
𝜖 𝜇(𝐾𝑛 ) 6 𝑓 𝑑𝜇 6 𝜖 𝜇(𝐾𝑛−1 )
𝑛=1 𝑛=1
𝑁
Õ 𝑁
Õ
𝜖 𝜇(𝐾𝑛 ) 6 Λ 𝑓 6 𝜖 𝜇(𝐾𝑛−1 ).
𝑛=1 𝑛=1

Note that
𝑁
Õ 𝑁
Õ
𝜇(𝐾𝑛−1 ) − 𝜇(𝐾𝑛 ) = 𝜇(𝐾0 ) − 𝜇(𝐾 𝑁 ) = 𝜇(supp 𝑓 ) − 𝜇(𝐾 𝑁 ).
𝑛=1 𝑛=1

Thus, ∫
𝑓 𝑑𝜇 − Λ 𝑓 6 𝜖 𝜇(supp 𝑓 ) − 𝜖 𝜇(𝐾𝑛 ) 6 𝜖 𝜇(supp 𝑓 ).

Since 𝜇(supp 𝑓 ) < ∞ and 𝜖 > 0 is arbitrary, we obtain that Λ 𝑓 = 𝑓 𝑑𝜇. 

6.3 Regularity properties of Radon measures


We first prove the following proposition.

Proposition 6.22. Let 𝜇 be a Radon measure on an LCH space 𝑋. Then, 𝜇 is inner regular on every
𝜎-finite set.

Proof. The proof is similar to the proof of the second part of Proposition 1.40. First, assume that 𝐸 is a
Borel set with 𝜇(𝐸) < ∞. Then, by outer regularity, given 𝜖 > 0, take an open set 𝑈 such that 𝐸 ⊆ 𝑈 and

𝜇(𝑈) < 𝜇(𝐸) + 𝜖 .

Then, by the inner regularity on 𝑈, we can choose a compact set 𝐹 with 𝐹 ⊆ 𝑈 and the following holds:

𝜇(𝑈) − 𝜖 < 𝜇(𝐹).

Note that 𝜇(𝑈 \ 𝐸) < 𝜖, and by outer regularity, there exists an open set 𝑉 with 𝑈 \ 𝐸 ⊆ 𝑉 such that
𝜇(𝑉) < 𝜖. Now, let 𝐾 = 𝐹 \ 𝑉. Then, 𝐾 is a closed subset of the compact set 𝐹, so it is compact.
Moreover, 𝐾 ⊆ 𝐸, and the following is satisfied:

𝜇(𝐾) = 𝜇(𝐹) − 𝜇(𝐹 ∩ 𝑉) > 𝜇(𝑈) − 𝜖 − 𝜇(𝑉) > 𝜇(𝑈) − 2𝜖 > 𝜇(𝐸) − 2𝜖 .

This shows that for every 𝜖 > 0,

𝜇(𝐸) − 2𝜖 6 sup{𝜇(𝐾) | 𝐾 ⊆ 𝐸 and 𝐾 is compact}.

Since 𝜖 > 0 is arbitrary, we have the inner regularity on 𝐸.

60
Ð
Now, suppose that 𝐸 is 𝜎-finite, and 𝜇(𝐸) = ∞. Then, write 𝐸 = 𝑛 𝐸 𝑛 with 𝜇(𝐸 𝑛 ) < ∞. Then, for
Ð𝑁 Ð𝑁
each positive number 𝛼, there exists 𝑁 such that 𝜇( 𝑛=1 𝐸 𝑛 ) > 𝛼. Moreover, since 𝜇( 𝑛=1 𝐸 𝑛 < ∞, by
Ð𝑁 Ð𝑁
taking 2𝜖 = 𝜇( 𝑛=1 𝐸 𝑛 ) − 𝛼 and using the previous argument, there exists a compact set 𝐾 ⊆ 𝑛=1 𝐸𝑛
such that
Ð𝑁
𝜇(𝐾) > 𝜇( 𝑛=1 𝐸 𝑛 ) − 2𝜖 = 𝛼.
Thus, we obtain that
𝛼 < sup{𝜇(𝐾) | 𝐾 ⊆ 𝐸 and 𝐾 is compact}.
Since 𝛼 > 0 is arbitrary, we obtain the inner regularity on 𝐸. 

Definition 6.23. A topological space is said to be 𝜎-compact if it is a countable union of compact sets.
We can also define 𝜎-compact subsets of 𝑋.

Corollary 6.24. Every 𝜎-finite Radon measure on an LCH space 𝑋 is regular. In particular, if 𝑋 is
𝜎-compact, then every Radon measure on 𝑋 is regular.

We now some approximation properties of Borel sets.

Definition 6.25. Let 𝑋 be a topological space. An 𝐹𝜎 set is a subset of 𝑋 which is a countable union of
closed sets. Also, a 𝐺 𝛿 set is a subset of 𝑋 that is a countable intersection of open sets.

Proposition 6.26. Let 𝜇 be a 𝜎-finite Radon measure on an LCH space 𝑋, and let 𝐸 be a Borel set.
Then, the following hold.

1. For each 𝜖 > 0, there exist an open set 𝑈 and a closed set 𝐹 such that 𝐹 ⊆ 𝐸 ⊆ 𝑈 and 𝜇(𝑈 \ 𝐹) < 𝜖.

2. There exist an 𝐹𝜎 set 𝐴 and a 𝐺 𝛿 set 𝐵 such that 𝐴 ⊆ 𝐸 ⊆ 𝐵 and 𝜇(𝐵 \ 𝐴) = 0.


Ð
Proof. Write 𝐸 = 𝑛 𝐸 𝑛 such that 𝜇(𝐸 𝑛 ) < ∞ for each 𝑛 and (𝐸 𝑛 ) is disjoint. Then, for each 𝑛, take an
Ð
open set 𝑈𝑛 for which 𝐸 𝑛 ⊆ 𝑈𝑛 and 𝜇(𝑈𝑛 ) < 𝜇(𝐸 𝑛 ) + 𝜖/2𝑛−1 . Let 𝑈 = 𝑛 𝑈𝑛 . Then, 𝑈 is open with
Ð
𝐸 ⊆ 𝑈, and since 𝑈𝑛 \ 𝐸 ⊆ 𝑈𝑛 \ 𝐸 𝑛 for each 𝑛, 𝑈 \ 𝐸 ⊆ 𝑛 (𝑈𝑛 \ 𝐸 𝑛 ). Therefore,

𝜇(𝑈 \ 𝐸) 6 𝜇(𝑈𝑛 \ 𝐸 𝑛 ) < 𝜖/2.

Apply the previous argument to 𝐸 𝑐 , then we have an open set 𝑉 such that 𝐸 𝑐 ⊆ 𝑉 and 𝜇(𝑉 \ 𝐸 𝑐 ) < 𝜖/2.
Now, let 𝑉 = 𝐹 𝑐 . Then, 𝐹 is closed with 𝐹 ⊆ 𝐸, and

𝜇(𝑈 \ 𝐹) = 𝜇(𝑈 \ 𝐸) + 𝜇(𝐸 \ 𝐹) = 𝜇(𝑈 \ 𝐸) + 𝜇(𝑉 \ 𝐸 𝑐 ) < 𝜖 .

Now, for each 𝑛, choose an open set 𝑈𝑛 and a closed set 𝐹𝑛 such that 𝐹𝑛 ⊆ 𝐸 ⊆ 𝑈𝑛 and 𝜇(𝑈𝑛 \ 𝐹𝑛 ) <
Ð Ñ
1/𝑛. Define 𝐴 = 𝑛 𝐹𝑛 and 𝐵 = 𝑛 𝑈𝑛 . Then, 𝐴 is 𝐹𝜎 , 𝐵 is 𝐺 𝛿 , and 𝐴 ⊆ 𝐸 ⊆ 𝐵. Moreover, for each
𝑛, 𝐵 \ 𝐴 ⊆ 𝑈𝑛 \ 𝐹𝑛 , so 𝜇(𝐵 \ 𝐴) < 1/𝑛. By letting 𝑛 → ∞, we obtain that 𝜇(𝐵 \ 𝐴) = 0. 

Remark 6.27. In this proof, we did not use the inner regularity of 𝜇.
The following theorem provides a sufficient condition that most of Borel measures on an LCH space
𝑋 are regular.

Theorem 6.28. Let 𝑋 be an LCH space such that every open set is 𝜎-compact. Then, every Borel
measure 𝜇 on 𝑋 which is finite on every compact set is regular.

Remark 6.29. Note that every second countable LCH space satisfies the assumption of the theorem. To
show this, suppose an LCH space 𝑋 is second countable. Then, we can take a countable base B of the
topology of 𝑋. Now, let 𝑈 be an open set. Then, for each 𝑥 ∈ 𝑈, take a neighborhood 𝑉𝑥 of 𝑥 such that
𝑉𝑥 is compact and 𝑥 ∈ 𝑉𝑥 ⊆ 𝑉𝑥 ⊆ 𝑈. Moreover, since B is a base, for each 𝑥 ∈ 𝑈, there exists 𝑊 𝑥 ∈ B
such that 𝑥 ∈ 𝑊 𝑥 ⊆ 𝑉𝑥 . Then, 𝑊 𝑥 ⊆ 𝑉𝑥 , so 𝑊 𝑥 is compact. Since 𝑥 ⊆ 𝑊 𝑥 ⊆ 𝑈 for every 𝑥 ∈ 𝑈, we

61
Ð
have that 𝑈 = 𝑥 ∈𝑈 𝑊 𝑥 . However, B is a countable set, so the union must be countable. This shows our
assertion.
Note that every Euclidean space R𝑑 is a second countable LCH space, and thus, by applying the
theorem, every Borel measure on R𝑑 that is finite on each compact set is a regular measure.

Proof. Define a positive linear functional Λ on 𝐶𝑐 (𝑋) by



Λ𝑓 B 𝑓 𝑑𝜇, 𝑓 ∈ 𝐶𝑐 (𝑋).

Then, by the Riesz representation theorem, there exists a Radon measure 𝜈 on 𝑋 such that Λ 𝑓 = 𝑓 𝑑𝜈 for
all 𝑓 ∈ 𝐶𝑐 (𝑋). We first show that 𝜇(𝑈) = 𝜈(𝑈) for every open set 𝑈. Let 𝑈 be an open set. Then, since 𝑈
Ð
is 𝜎-compact, we can write 𝑈 = 𝑛 𝐾𝑛 , where each 𝐾𝑛 is compact. Since a finite union of compact sets is
compact, we may assume that (𝐾𝑛 ) is an increasing sequence. Choose 𝑓1 ∈ 𝐶𝑐 (𝑋) so that 𝐾1 ≺ 𝑓1 ≺ 𝑈.
If 𝑓𝑛 ∈ 𝐶𝑐 (𝑋) was chosen, then choose 𝑓𝑛+1 ∈ 𝐶𝑐 (𝑋) so that (𝐾𝑛+1 ∪ (supp 𝑓𝑛 )) ≺ 𝑓𝑛+1 ≺ 𝑈. Then,
( 𝑓𝑛 ) is an increasing sequence, and moreover, for each 𝑥 ∈ 𝑈, there exists 𝑁 such that 𝑥 ∈ 𝐾 𝑁 , and so
for 𝑛 > 𝑁, since 𝐾 𝑁 ⊆ 𝐾𝑛 ≺ 𝑓𝑛 , we obtain that 𝑓𝑛 (𝑥) = 1. Moreover, for 𝑥 ∉ 𝑈, for every 𝑛, 𝑓𝑛 (𝑥) = 0
since 𝑓𝑛 ≺ 𝑈. Hence, 𝑓𝑛 → 𝜒𝑈 pointwise. Therefore, by the monotone convergence theorem,
∫ ∫
𝜇(𝑈) = lim 𝑓𝑛 𝑑𝜇 = lim 𝑓𝑛 𝑑𝜈 = 𝜈(𝑈).

We now show that 𝜇 is a regular measure. Let 𝐸 be a Borel set. Then, given 𝜖 > 0, there are an open
set 𝑈 and a closed set 𝐹 such that 𝐹 ⊆ 𝐸 ⊆ 𝑈 and 𝜈(𝑈 \ 𝐹) < 𝜖. However, since 𝑈 \ 𝐹 is open, we have
𝜇(𝑈 \ 𝐹) < 𝜖, and in particular,

𝜇(𝑈) = 𝜇(𝐸) + 𝜇(𝑈 \ 𝐸) 6 𝜇(𝐸) + 𝜇(𝑈 \ 𝐹) 6 𝜇(𝐸) + 𝜖 .

This shows that


inf{𝜇(𝑈) | 𝐸 ⊆ 𝑈 and 𝑈 is open} 6 𝜇(𝐸) + 𝜖
for every 𝜖 > 0. Since 𝜖 > 0 is arbitrary, we show that 𝜇 is outer regular on 𝐸. Moreover, since 𝑈 is
Ð
𝜎-compact, 𝐹 is also 𝜎-compact. Thus, we can write 𝐹 = 𝑛 𝐾𝑛 , where 𝐾𝑛 is compact, and (𝐾𝑛 ) is
increasing. Now, observe that

𝜇(𝐸) = 𝜇(𝐹) + 𝜇(𝐸 \ 𝐹) 6 𝜇(𝐹) + 𝜖,

and since 𝜇(𝐾𝑛 ) → 𝜇(𝐹),

𝜇(𝐹) 6 sup{𝜇(𝐾) | 𝐾 ⊆ 𝐸 and 𝐾 is compact}.

Therefore,
𝜇(𝐸) 6 sup{𝜇(𝐾) | 𝐾 ⊆ 𝐸 and 𝐾 is compact} + 𝜖
for every 𝜖 > 0. Since 𝜖 > 0 is arbitrary, we also prove that 𝜇 is inner regular on 𝐸. 

We now turn to the properties of 𝐿 𝑝 -spaces associated to a Radon measure. The proofs are essentially
the same with the previous ones, so we just mention the previous results. The following proposition can
be showed by using the argument in the proof of Proposition 3.21.

Proposition 6.30. Let 𝑝 ∈ [1, ∞), and let 𝑋 be a Radon measure. Then, 𝐿 𝑝 (𝑋, 𝜇) is separable.

Using the same argument in the proof of Proposition 3.23 and Proposition 6.22, we can prove the
following approximation property of 𝐿 𝑝 (𝑋, 𝜇) if 𝜇 is a Radon measure on an LCH space 𝑋.

Proposition 6.31. Let 𝑝 ∈ [1, ∞), and let 𝜇 be a Radon measure on an LCH space 𝑋. Then, 𝐶𝑐 (𝑋) is
dense in 𝐿 𝑝 (𝑋, 𝜇).

62
6.4 Lebesgue measure again
Note that every function∫ 𝑓 ∈ 𝐶𝑐 (R𝑑 ) is Riemann integrable over some compact rectangle containing
supp 𝑓 , so we can define R𝑑 𝑓 𝑑𝑥 using the Riemann integration. Define a linear functional Λ on 𝐶𝑐 (R𝑑 )
by ∫
Λ𝑓 B 𝑓 𝑑𝑥,
R𝑑
where the integration is the Riemann integration. This functional is clearly positive. One may guess that
this functional induces the Lebesgue measure 𝜆 on R𝑑 .

Theorem 6.32. The corresponding Radon measure to the positive linear functional Λ is the Lebesgue
measure on R𝑑 .

Remark 6.33. By this theorem, we prove the regularity of the Lebesgue measure again.

Proof. Let 𝜇 be the associated Radon measure with Λ. It suffices to show that 𝜇(𝑅) = 𝜆(𝑅) for every
𝑑-dimensional bounded rectangle 𝑅. By the continuity, we may assume that 𝑅 is an open rectangle.
Ð
Take an increasing sequence of compact rectangles (𝐾𝑛 ) so that 𝑛 𝐾𝑛 = 𝑅. Note that 𝜆(𝐾𝑛 ) → 𝜆(𝑅).
Now, choose 𝑓1 ∈ 𝐶𝑐 (R𝑑 ) so that 𝐾1 ≺ 𝑓1 ≺ 𝑅. If 𝑓𝑛 was chosen, take 𝑓𝑛+1 ∈ 𝐶𝑐 (R𝑑 ) so that
(𝐾𝑛 ∪ (supp 𝑓𝑛 )) ≺ 𝑓𝑛+1 ≺ 𝑅. Then, ( 𝑓𝑛 ) is an increasing sequence, and by the definition of Λ,
𝜆(𝐾𝑛 ) 6 Λ 𝑓𝑛 for each 𝑛. Moreover, since each 𝑓𝑛 satisfies 0 6 𝑓𝑛 6 1, and since supp 𝑓𝑛 ⊆ 𝑅,
∫ ∫
Λ 𝑓𝑛 = 𝑓𝑛 𝑑𝑥 6 𝑑𝑥 = 𝜆(𝑅).
𝑅 𝑅

Note that 𝑓𝑛 → 𝜒𝑅 pointwise, and by the monotone convergence theorem,



Λ 𝑓𝑛 = 𝑓𝑛 𝑑𝜇 → 𝜇(𝑅).

By taking the limit 𝑛 → ∞ in the following inequalities

𝜆(𝐾𝑛 ) 6 Λ 𝑓𝑛 6 𝜆(𝑅),

we obtain the identity 𝜆(𝑅) = 𝜇(𝑅). This completes the proof. 

6.5 Lusin’s theorem and Vitali-Carathéodory theorem


We already show that if 𝜇 is a Radon measure on an LCH space 𝑋, then 𝐶𝑐 (𝑋) is dense in 𝐿 𝑝 (𝑋, 𝜇)
whenever 1 6 𝑝 < ∞. We now show another approximation theorem.

Theorem 6.34 (Lusin). Let 𝜇 be a Radon measure on an LCH space 𝑋, and let 𝑓 be a complex-valued
measurable function on 𝑋 which vanishes outside a set of finite measure. Then, for every 𝜖 > 0, there
exists 𝑔 ∈ 𝐶𝑐 (𝑋) such that
𝜇({ 𝑓 ≠ 𝑔}) < 𝜖 .
Moreover, 𝑔 can be taken so that

sup |𝑔(𝑥)| 6 sup | 𝑓 (𝑥)|.


𝑥 ∈𝑋 𝑥 ∈𝑋

Proof. First, suppose that 𝑓 is bounded, and let 𝐸 B { 𝑓 ≠ 0}. Then, 𝜇(𝐸) < ∞ by the assumption. Note
that since 𝑓 is bounded, 𝑓 belongs to 𝐿 1 (𝑋, 𝜇), so there exists a sequence (𝑔𝑛 ) in 𝐶𝑐 (𝑋) such that 𝑔𝑛 → 𝑓
in 𝐿 1 . By passing to a subsequence, we may assume that 𝑔𝑛 → 𝑓 almost everywhere. Then, 𝑔𝑛 → 𝑓 on
𝐸, so by Egorov’s theorem, given 𝜖 > 0, there exists a Borel set 𝐴 ⊆ 𝐸 such that 𝑔𝑛 → 𝑓 uniformly on
𝐴 and 𝜇(𝐸 \ 𝐴) < 𝜖/3. Now, by inner regularity, take a compact set 𝐾 ⊆ 𝐴 so that 𝜇( 𝐴) − 𝜖/3 < 𝜇(𝐾),

63
that is, 𝜇( 𝐴 \ 𝐾) < 𝜖/3. Then, 𝑔𝑛 → 𝑓 uniformly on 𝐾, 𝑓 is continuous on 𝐾. Next, by outer regularity,
take an open set 𝑈 such that 𝐸 ⊆ 𝑈 and 𝜇(𝑈) < 𝜇(𝐸) + 𝜖/3, that is, 𝜇(𝑈 \ 𝐸) < 𝜖/3. Now, by the Tietze
extension theorem, we can find a continuous function ℎ on 𝑋 such that ℎ| 𝐾 = 𝑓 and supp ℎ ⊆ 𝑈. Note
that since ℎ = 𝑓 on 𝐾 ∪ 𝑈 𝑐 , {ℎ ≠ 𝐾 } ⊆ 𝑈 \ 𝐾. Now, observe that
𝜖
𝜇(𝑈 \ 𝐾) = 𝜇(𝑈 \ 𝐸) + 𝜇(𝐸 \ 𝐴) + 𝜇( 𝐴 \ 𝐾) < 3 · = 𝜖.
3
Thus, 𝜇({ℎ ≠ 𝑓 }) < 𝜖.
To show the second assertion, let 𝑅 = sup 𝑥 ∈𝑋 | 𝑓 (𝑥)|. Define 𝜑 : C → C by
(
𝑧 if |𝑧| 6 𝑅
𝜑(𝑧) B
𝑅 |𝑧𝑧 | if |𝑧| > 𝑅.

Then, 𝜑 is continuous and |𝜙| 6 𝑅. Therefore, if we define 𝑔 B 𝜑 ◦ ℎ, then sup 𝑥 ∈𝑋 |𝑔(𝑥)| 6 𝑅. Moreover,
since {ℎ = 𝑓 } ⊆ {𝑔 = 𝑓 }, we have 𝜇({𝑔 ≠ 𝑓 }) < 𝜖.
Now, suppose that 𝑓 is unbounded. For each 𝑛, define 𝐸 𝑛 B {0 < | 𝑓 | 6 𝑛}. Then, (𝐸 𝑛 ) is an
Ð
increasing sequence, and 𝑛 𝐸 𝑛 = 𝐸. Thus, given 𝜖 > 0, there exists 𝑁 such that 𝜇(𝐸 \ 𝐸 𝑁 ) < 𝜖/2.
By the previous part, let 𝑔 ∈ 𝐶𝑐 (𝑋) such that 𝜇({𝑔 ≠ 𝑓 𝜒𝐸𝑁 }) < 𝜖/2. Note that since 𝑓 = 𝑓 𝜒𝐸𝑁 on
𝐸 𝑁 ∪ 𝐸 𝑐 , we obtain that 𝑓 = 𝑔 on (𝐸 𝑁 ∪ 𝐸 𝑐 ) ∩ {𝑔 = 𝑓 𝜒𝐸𝑁 }. Therefore,

𝜇({ 𝑓 ≠ 𝑔)} 6 𝜇(𝐸 \ 𝐸 𝑁 ) + 𝜇({𝑔 ≠ 𝑓 𝜒𝐸𝑁 }) < 𝜖 .

We now discuss approximation by semicontinuous functions.

Definition 6.35. An extended-real-valued function 𝑓 on a topological space 𝑋 is said to be lower


semicontinuous (LSC) if the subset { 𝑓 > 𝑎} is open for every real number 𝑎. Similarly, 𝑓 is said to be
upper semicontinuous (USC) if the subset { 𝑓 < 𝑎} is open for every real number 𝑎.

Example 6.36. The characteristic function of an open set is LSC. Similarly, the characteristic function
of a closed set is USC.
To prove these statements, let 𝑋 be a topological space, and let 𝑈 be an open set. Note that for each
real number 𝑎,


 𝑋 if 𝑎 < 0


{ 𝜒𝑈 > 𝑎} = 𝑈 if 0 6 𝑎 < 1

 ∅ if 𝑎 > 1.


Thus, 𝜒𝑈 is LSC. Moreover, if 𝐹 is a closed set, then for a real number 𝑎,



 𝑋 if 𝑎 > 1


{ 𝜒𝐹 < 𝑎} = 𝑋 \ 𝐹 if 0 < 𝑎 > 1

∅

if 𝑎 6 0.

Thus, 𝜒𝐹 is USC.

Proposition 6.37. Let 𝑋 be a topological space.

1. Suppose 𝑓1 and 𝑓2 are LSC functions on 𝑋. Then, 𝑓1 + 𝑓2 is also an LSC function on 𝑋.

2. Suppose F is a family of LSC functions on 𝑋. Then, 𝑓 = sup F is also an LSC function on 𝑋.

64
Proof. To show the first statement, let 𝑎 ∈ R, and let 𝑥 0 ∈ 𝑋 with 𝑓1 (𝑥 0 ) + 𝑓2 (𝑥 0 ) > 𝑎. We choose
𝑏 1 , 𝑏 2 ∈ R so that
𝑥 0 ∈ { 𝑓1 > 𝑏 1 } ∩ { 𝑓2 > 𝑏 2 } ⊆ { 𝑓1 + 𝑓2 > 𝑎}.
To do this, we require that 𝑏 1 + 𝑏 2 > 𝑎, 𝑓1 (𝑥 0 ) > 𝑏 1 and 𝑓2 (𝑥 0 ) > 𝑏 2 . First, choose 𝜖 > 0 so that
𝑓1 (𝑥0 ) + 𝑓2 (𝑥 0 ) > 𝑎 + 𝜖 > 𝑎. Put 𝑏 1 = 𝑎 + 𝜖 − 𝑓2 (𝑥 0 ). Then, we need to take 𝑏 2 so that 𝜖 − 𝑓2 (𝑥 0 ) + 𝑏 2 > 0
and that 𝑓2 (𝑥 0 ) > 𝑏 2 . By choosing 𝑏 2 = 𝑓2 (𝑥 0 ) − 𝜖, we make the requirement satisfied. This means that
𝑥 0 has an open neighborhood contained in the set { 𝑓1 + 𝑓2 > 𝑎}. Therefore, { 𝑓1 + 𝑓2 > 𝑎} is open. Hence,
𝑓1 + 𝑓2 is LSC.
Now, we prove the second statement. Observe that for each real number 𝑎, sup𝑔 ∈F 𝑔(𝑥) > 𝑎 if and
only if 𝑔(𝑥) > 𝑎 for some 𝑔 ∈ F. This is clear by the definition of supremum. This means that
Ø
{ 𝑓 > 𝑎} = {𝑔 > 𝑎}.
𝑔 ∈F

Hence, the set { 𝑓 > 𝑎} is open. 

Note that if 𝑔 is a USC function, then −𝑔 is a LSC function. Thus, we also have the following
properties of USC functions.

Proposition 6.38. Let 𝑋 be a topological space.

1. Suppose 𝑓1 and 𝑓2 are USC functions on 𝑋. Then, 𝑓1 + 𝑓2 is also an USC function on 𝑋.

2. Suppose F is a family of USC functions on 𝑋. Then, 𝑓 = inf F is also an USC function on 𝑋.

We now prove the following approximation theorem.

Theorem 6.39 (Vitali-Carathéodory). Let 𝜇 be a Radon measure on an LCH space 𝑋, and let 𝑓 be a
𝜇-integrable real-valued function on 𝑋. Then, for each 𝜖 > 0, there exist functions 𝑔 and ℎ on 𝑋 such
that the following hold.

1. 𝑔 6 𝑓 6 ℎ.

2. 𝑔 is USC and bounded above.

3. ℎ is LSC and bounded below.

4. We have the following estimate: ∫


(ℎ − 𝑔)𝑑𝜇 < 𝜖 .

Proof. We first assume that 𝑓 is nonnegative. Then, take the standard simple approximation ( 𝑓𝑛 ) of 𝑓 ,
namely,
𝑛2𝑛
1 Õ
𝑠𝑛 = 𝑛 (𝑘 − 1) 𝜒𝐸𝑛,𝑘 + 𝑛𝜒 { 𝑓 >𝑛} ,
2 𝑘=1
where  
𝑘 −1 𝑘
𝐸 𝑛,𝑘 = 6 𝑓 < 𝑛 .
2𝑛 2
Then, (𝑠 𝑛 ) is an increasing sequence of simple measurable functions, and 𝑠 𝑛 → 𝑓 pointwise. Now,
define 𝑠0 = 0 and 𝑡 𝑛 = 𝑠 𝑛 − 𝑠 𝑛−1 for each 𝑛. Then, each 𝑡 𝑛 is a nonnegative simple function, and

Õ
𝑓 = 𝑡𝑛 .
𝑛=1

65
Note that we can write each 𝑡 𝑛 as a linear combination of characteristic functions with nonnegative
coefficients. (For an explicit computation, see the remark.) Moreover, we may assume that the coefficients
are strictly positive. Thus, since 𝑓 = ∞
Í
𝑛=1 𝑡 𝑛 , we can write


Õ
𝑓 = 𝑐 𝑛 𝜒 𝐸𝑛 ,
𝑛=1

where each 𝐸 𝑛 is a measurable set and each 𝑐 𝑛 is strictly positive.9 Now, by the monotone convergence
theorem, ∫ Õ
𝑓 𝑑𝜇 = 𝑐 𝑛 𝜇(𝐸 𝑛 ) < ∞.
𝑛

In particular, 𝜇(𝐸 𝑛 ) < ∞ for every 𝑛 since 𝑐 𝑛 > 0. Thus, 𝜇 is inner regular on 𝐸 𝑛 .
Now, given 𝜖 > 0 and for each 𝑛, choose a compact set 𝐾𝑛 and an open set 𝑈𝑛 such that 𝐾𝑛 ⊆ 𝐸 𝑛 ⊆ 𝑈𝑛
and that
𝜖
𝑐 𝑛 𝜇(𝑈𝑛 \ 𝐾𝑛 ) < 𝑛+1 .
2
Define
Õ∞ Õ𝑁
ℎ= 𝑐 𝑛 𝜒𝑈𝑛 and 𝑔 = 𝑐 𝑛 𝜒𝐾𝑛 ,
𝑛=1 𝑛=1

where 𝑁 is chosen later. Then, ℎ is LSC, 𝑔 is USC, and 𝑔 6 𝑓 6 ℎ. Moreover, 𝑔 is a bounded function,
and ℎ is nonnegative. Now, observe that
𝑁
Õ ∞
Õ
ℎ−𝑔 = 𝑐 𝑛 ( 𝜒𝑈𝑛 − 𝜒𝐾𝑛 ) + 𝑐 𝑛 𝜒𝑈𝑛
𝑛=1 𝑛=𝑁 +1
Õ𝑁 Õ∞
6 𝑐 𝑛 ( 𝜒𝑈𝑛 − 𝜒𝐾𝑛 ) + 𝑐 𝑛 ( 𝜒𝐾𝑛 + 𝜒𝑈𝑛 \𝐾𝑛 )
𝑛=1 𝑛=𝑁 +1

Õ ∞
Õ
= 𝑐 𝑛 𝜒𝑈𝑛 \𝐾𝑛 + 𝑐 𝑛 𝜒𝐾𝑛
𝑛=1 𝑛=𝑁 +1
Õ∞ Õ∞
6 𝑐 𝑛 𝜒𝑈𝑛 \𝐾𝑛 + 𝑐 𝑛 𝜒 𝐸𝑛 .
𝑛=1 𝑛=𝑁 +1
Í∞
Take 𝑁 large so that 𝑛=𝑁 +1 𝑐 𝑛 𝜇(𝐸 𝑛 ) < 𝜖/2. Then,
∫ ∞
Õ ∞
Õ
(ℎ − 𝑔)𝑑𝜇 6 𝑐 𝑛 𝜇(𝑈𝑛 \ 𝐾𝑛 ) + 𝑐 𝑛 𝜇(𝐸 𝑛 ) < 𝜖 .
𝑛=1 𝑛=𝑁 +1

Hence, we have proved our assertions for nonnegative 𝑓 ’s.


Suppose 𝑓 is real-valued. Then, decompose 𝑓 as 𝑓 = 𝑓 + − 𝑓 − . Given 𝜖 > 0, take 𝑔1 and ℎ1 for 𝑓 + ,
and 𝑔2 and ℎ2 for 𝑓 − so that they correspond to 𝜖/2. Then, since 𝑔1 6 𝑓 + 6 ℎ1 and 𝑔2 6 𝑓 − 6 ℎ2 , we
have
𝑔1 − ℎ 2 6 𝑓 = 𝑓 + − 𝑓 − 6 ℎ 1 − 𝑔2 .
Then, 𝑔 = 𝑔1 − ℎ2 and ℎ = ℎ1 − 𝑔2 satisfy our assertions. 

Remark 6.40. In this remark, we compute an explicit formula for 𝑡 𝑛 .

1. If 𝑓 (𝑥) > 𝑛, then 𝑡 𝑛 (𝑥) = 𝑛 − (𝑛 − 1) = 1.


9 (𝐸
𝑛) may not be a disjoint family.

66
2. Observe that
𝑛2𝑛
Ø
{𝑛 − 1 6 𝑓 < 𝑛} = 𝐸 𝑛,𝑘 .
𝑘=(𝑛−1)2𝑛 +1

Thus, if 𝑛 − 1 6 𝑓 (𝑥) < 𝑛, then by taking 𝑗 ∈ {1, . . . , 2𝑛 } so that 𝑥 ∈ 𝐸 𝑛, ( (𝑛−1)2𝑛 + 𝑗) B 𝐵𝑛, 𝑗 ,

(𝑛 − 1)2𝑛 + 𝑗 − 1 𝑗 −1
𝑡 𝑛 (𝑥) = 𝑛
− (𝑛 − 1) = 𝑛 .
2 2

3. Suppose 0 6 𝑓 (𝑥) < 𝑛 − 1. Then, let 𝑘 ∈ {1, . . . , (𝑛 − 1)2𝑛−1 } so that 𝑥 ∈ 𝐸 (𝑛−1),𝑘 . Then, either
𝑥 ∈ 𝐸 𝑛, (2𝑘−1) or 𝑥 ∈ 𝐸 𝑛,2𝑘 . Therefore,

0


 if 𝑥 ∈ 𝐸 𝑛, (2𝑘−1)
𝑡 𝑛 (𝑥) = 1

 𝑛 if 𝑥 ∈ 𝐸 𝑛,2𝑘 .
2
Ð (𝑛−1)2𝑛−1
Define 𝐴𝑛 B 𝑘=1 (𝐸 (𝑛−1),𝑘 ∩ 𝐸 𝑛,2𝑘 ).

Combining the observations above,


2𝑛
Õ
−𝑛 −𝑛
𝑡 𝑛 (𝑥) = 2 𝜒 𝐴𝑛 + 2 ( 𝑗 − 1) 𝜒𝐵𝑛, 𝑗 + 𝜒 { 𝑓 >𝑛} .
𝑗=2

67
References
[Bog07] V. Bogachev, Measure Theory, Springer, 2007.
[Coh13] D. L. Cohn, Measure Theory, 2nd ed., Birkhäuser Advanced Texts Basler Lehrbücher, Birkhäuser,
2013.
[Fol99] G. B. Folland, Real Analysis: Modern Techniques and Their Applications, 2nd ed., Wiley, 1999.
[Rud76] W. Rudin, Principles of Mathematical Analysis, 3rd ed., McGraw-Hill, 1976.
[SS95] L. A. Steen and J. A. Seebach Jr., Counterexample in Topology, Dover Books on Mathematics, Dover
Publications, 1995.
[Tes21] G. Teschl, Topics in Real Analysis, May 22, 2021, url:
https://www.mat.univie.ac.at/~gerald/ftp/book-ra/index.html.

68

You might also like