Math 515 Lecture Notes

Gabe Khan

January 21, 2022

2
Chapter 1

Introduction

The topic of this class is real analysis, with a particular focus on measure
theory and integration. Since the qualifying exam will primarily focus on
functions of a single real variable, this will also be our main focus. We will
develop the notion of the Lebesgue integral, which supplants the Riemann
integral by being robust enough to handle fairly pathological functions and
flexible enough to form the basis for modern functional analysis.
The lecture notes and home work are primarily taken from the textbook
by Royden and Fitzpatrick [RF88], which is the textbook for this course.
Furthermore, much of the material is also adapted from the 2015 lecture
notes by Timothy McNicholl.

1.1 Lecture 1: A review of Riemann integration

In this introductory lecture, we will review some notions of Riemann integra-
tion, which is the historical (and pedagogical) predecessor to the Lebesgue
integral that we will study in this course. We start with some basic defini-
tions.
Definition 1. A partition P of the interval ra, bs is a sequence of real num-
bers px0 , . . . , xn q with
a “ x0 ă x1 ă . . . xn “ b.
Intuitively, the partition cuts up the interval into n subintervals rxi , xi`1 s.
Definition 2. A real valued function f : R Ñ R is bounded on the interval
ra, bs if there is a positive number M so that
|f pxq| ď M

3
4 CHAPTER 1. INTRODUCTION

for all x P ra, bs.

Here, the domain of f can be any set which contains ra, bs, and not
necessarily the entire real line.

Definition 3. Suppose that f is a real-valued bounded function on ra, bs.

1. A lower Riemann sum for f on ra, bs is a sum of the form

n
ÿ
vj pxj ´ xj´1 q
j“1

where px0 , . . . , xn q is a partition of ra, bs and vj ď f pxq whenever x P

rxj´1 , xj s.

2. An upper Riemann sum for f is the same, except now we insist that
vj ě f pxq whenever x P rxj´1 , xj s.

Intuitively, the lower Riemann sum is an under-estimate for the area

underneath the function f pxq whereas the upper Riemann sum is an over-
estimate.

By taking finer and finer partitions of the interval ra, bs, we can refine our
estimates for the area under the curve and (hopefully), compute its Riemann
integral.
1.1. LECTURE 1: A REVIEW OF RIEMANN INTEGRATION 5

ş
Definition 4. The lower Riemann integral is the supremum of all lower
Riemann sums with respect to all possible partitions.
ş
Definition 5. The upper Riemann integral is the infimum of all upper
Riemann sums with respect to all possible partitions.

We say that a function f is Riemann integrable if the lower Riemann

integral is equal to the upper Riemann integral.

Example. If f is continuous on ra, bs, then f is integrable.

Proof. Since f is a continuous function on a closed and bounded interval, it

is uniformly continuous. Therefore, for any  ą 0, we can find a δ so that

|f pxq ´ f pyq| ă 

whenever |x ´ y| ă δ.
We then take a partition P whose mesh1 is smaller than δ and consider
the upper and lower Riemann sums
n
ÿ
LP “ f pxj qpxj ´ xj´1 q
j“1

and
n
ÿ
UP “ f pxj qpxj ´ xj´1 q,
j“1

where xj “ argminxPrxj´1 ,xj s f pxq and xj “ argmaxxPrxj´1 ,xj s f pxq.

We now compare LP and UP using the uniform continuity of f , we have
the following estimate

n
ÿ n
ÿ
UP ´ LP “ f pxj qpxj ´ xj´1 q ´ f pxj qpxj ´ xj´1 q
j“1 j“1
ÿn
` ˘
“ f pxj q ´ f pxj q pxj ´ xj´1 q
j“1
ÿn
ă pxj ´ xj´1 q
j“1
“  ¨ pb ´ aq
1
In other words, we consider a partition for which xj ´xj´1 ă δ for all j with 1 ď j ď n.
6 CHAPTER 1. INTRODUCTION

For any upper and lower sum, we have that

ż ż
LP ď ď ď UP .

By taking  arbitrarily small, this shows that L “ U.

In fact, we can Riemann integrate functions which have some disconti-

nuities (and we will prove a much stronger version of the following exercise
later in the course).

Exercise 1. Let f be a bounded real-valued function on ra, bs that is con-

tinuous except possibly at finitely many points. Show that f is Riemann
integrable.

On the other hand, by considering highly discontinuous functions, we

can create functions which are not Riemann integrable.

Non-Example. The function f : r0, 1s Ñ R with

#
1 whenever x P Q
1Q pxq “
0 whenever x R Q

is not Riemann integrable.

Proof.
řn In any interval pxj´1 , xj q, there are irrational number. Therefore, if
j“1 vj pxj ´ xj´1 is a lower Riemann sum, then vj ď 0 and thus we have
řn L ď 0. Similarly, in any interval there are rational numbers, so if
that
j“1 vj pxj ´ xj´1 is an upper Riemann sum, then vj ě 1 for all j. As such,
U ě 1.

With slightly more effort, it is possible to show that L “ 0 and U “ 1,

but the key thing is that they are not equal.
In this previous example, the rational numbers are “small” in that they
are a countable subset of an uncountable space, and thus we might expect
that
ż1
1Q dx “ 1.
0

We will later see that this is indeed the case if we use Lebesgue integral.
1.1. LECTURE 1: A REVIEW OF RIEMANN INTEGRATION 7

Figure 1.1: An upper and lower sum for a reasonably discontinuous function

1.1.1 Intuition

The idea of the Riemann integral is to slice a function vertically into narrow
strips which can be approximated by rectangles. However, what this exam-
ple shows is that for functions which are extremely discontinuous, it is not
possible to determine the correct “height” to make the slices.
The idea of the Lebesgue integral is to instead cut vertically, which
transforms the problem from one of determining how high to make the slices
to determining the size of sets.

Figure 1.2: A Riemann sum (left) and an approximation by simple functions

(right)
8 CHAPTER 1. INTRODUCTION

1.1.2 Why should we care?

At its core, measure theory studies the “size” of sets. In other words, given
a set X, we want a way to determine how large it is. This turns out to be
a fairly subtle question, and one with a long history.
There are many ways to discuss the size of a set. To give two basic exam-
ples, one could start by simply counting the elements of X (i.e., determining
its cardinality)2 . Another option for topological spaces is to determine the
dimension of the set. However, our focus is in measurements which corre-
spond to an integral of one form or another.
For some of you, the idea of extending the Riemann integral to integrate
more functions might be interesting and natural. For others, functions which
fail to be Riemann integrable are already quite pathological, so the idea of
extending the integral might appear to be an esoteric pursuit.
However, the significance of the Lebesgue integral is that it is much more
robust than the Riemann integral. Because of this, we will be able to prove
strong convergence theorems, which play a central role in modern analysis,
PDEs, probability, etc. To give an analogy, the mean and intermediate value
theorem play a central role in analysis, but these results do not hold in Q,
but require its metric completion R. In this context, the Riemann integral
is somewhat akin to the rational numbers, and the Lebesgue integral is R.

1.2 Resources
In this section, I have included some references which might be helpful for
your studying.

1. Sheldon Axler wrote a textbook which is freely available online [Axl20].

It’s a bit more elementary than our course, so might be a good place
to start.

2. I highly recommend the textbook by Gerald Folland [Fol99]. This text

considers measures in greater generality from the outset and covers the
material in a different order than the other books though, so might be
a good reference.

3. Terry Tao wrote a introduction to measure theory which is worth read-

ing [Tao11]. It is also available freely online.
2
For countable subsets, the cardinality can be phrased as an integral with respect to a
so called “counting measure.”
1.2. RESOURCES 9

4. Gerald Teschl wrote a book on functional analysis which discusses a

lot of topics that we cover in the course [Tes98]. It is freely available
online.
10 CHAPTER 1. INTRODUCTION
Chapter 2

Measurable Sets and their

Measures

In this chapter, we will build up measure theory with the goal of defining
the “length” of a set of real numbers. Our purpose is to motivate the notion
of the Lebesgue measure so we will not start with an abstract definition of a
measure. Instead, we will start with some natural properties of a “length”
and build from there.
There are a few properties of lengths which are “obvious.” For instance,
any notion of length should satisfy that the length of any finite interval ra, bs
is simply a ´ b. Similarly, the length of any unbounded interval should be
infinite.
Going further, it is natural to request that the length of a finite (or
countable) union of disjoint intervals is the sum of its lengths and that the
length of a subset should be less than the length of its parent.

2.1 Outer measures

We now give our first attempt to define the “length” of a set of real numbers,
which is the outer measure, denoted m˚ .
This measure satisfies some of the properties that we desire. For instance,
the outer measure of an interval is its length. The outer measure also has
the advantage that it is very intuitive and well-defined for any subset of the
real numbers. However, this flexibility comes at a cost and we will see that
this attempt fails to have some natural properties we might expect for a
notion of length.

11
12 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

For instance, given A, B Ă R with A X B “ H, it may be that

m˚ pA Y Bq ă m˚ pAq ` m˚ pBq.

In order to define the outer measure, we start by defining a “length”

function ` : PpRq Ñ R. Here, PpRq denotes the set of subsets of R

1. If I is a bounded interval, then `pIq denotes the usual length of I.

2. If I is an unbounded interval, then `pIq “ 8.

3. If tIj ujPF is a countable pairwise disjoint family of intervals1 , then

˜ ¸
ď ÿ
` Ij “ `pIj q. (2.1)
jPF jPF

Before moving on, we make some remarks.

1. In Equation 2.1, all the terms in the right hand side of the equation
are non-negative. Thus, if the series converges it does so absolutely
and the order of summation does not matter.

2. If U is any non-empty open set of reals, there is a countable

Ť pairwise
disjoint family of open intervals tIj ujPF so that U “ jPF Ij . Thus
`pU q is defined for every non-empty open sets of real numbers.

3. We shall refer to ` as the length of U

With this in mind, we now define the outer measure.

Definition 6. Suppose E is a set of real numbers. The outer measure of E,

denoted m˚ pEq, is defined to be the largest real number that is less than or
equal to the length of every open set including E. In other words,
# ˇ +
ÿ ˇ ď
˚
m pEq “ inf lpIj q ˇ E Ă Ij .
ˇ
ˇ
jPF jPF

Note that this definition is very slightly different from the definition
given in the book. However, this difference is not important.
1
We will often use the notation F to denote a countable set. Such a set may be finite,
which is why we are not using N as our index.
2.1. OUTER MEASURES 13

Example. There are a few outer measures that can readily be computed.

1.
m˚ pHq “ 0

2.
m˚ pR “ 8

3. Let a P R,
m˚ ptauq “ 0

Exercise 2. Show that the outer measure of a countable set is zero.

Proposition 1. The outer measure of an interval is its length.

Proof. Claim 1: `pIq ď m˚ pIq.

Suppose U is an open set with I Ă U . There is aŤcountable pairwise
disjoint family of open intervals tIj ujPF so the U “ jPF Ij . Since I is
connected, there is a j0 P F with I Ă Ij0 . Therefore,

`pIq ď `pIj0 q ď `pU q.

Taking infimums, we find that `pIq ď m˚ pIq.

Claim 2: m˚ pIq ď `pIq.
We break this part of the argument into two cases

1. I is unbounded. Then `pIq “ 8, and we are done.

2. If I is bounded, let a be the left end-point and b be the right end-point.

We let  ą 0 and take U “ pa ´ , b ` q. We see that I Ă U. Further-
more, we that that

m˚ I ď `pU q “ pb ´ aq ` 2 “ `pIq ` 2.

Since epsilon was arbitrary, we see that m˚ pIq ď `pIq.

Let us use what we’ve seen to solve a question from the qualifying exams.
14 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Exercise 3 (1997 Qualifying Exam, Problem 6). For every  ą 0 and every
subset E of the real numbers, define
8
ÿ 1
µε pEq “ inf r` pIn qs 2
n“1

where `p¨q denotes the length of an interval and the infimum is taken over all
countable collections of open intervals tI1 , I2 , . . .u which cover E for which
` pIn q ă  for all n. Prove the following:
a. Show that µ is an outer measure on the real line. b. Show that

µpEq :“ sup µε pEq “ lim µε pEq

εą0 εÑ0`

c. Show that µp0, 1q “ `8. (Hint: If one has intervals ˙ In of length ď ε

1 1
which cover p0, 1q, then r` pIn qs 2 “ `pIn q 1 ě ´ 2 ` pIn q ¨
r`pIn qs 2

2.1.1 Monotonicity and translation invariance

We now show that the outer measure has two natural and desired properties
of length, namely that subsets cannot be bigger than the original set (i.e.,
the outer measure is monotonic) and that if you translate a set, its outer
measure remains the same (i.e., the outer measure is translation invariant).

Proposition 2. If A Ă B Ă R, then m˚ paq ď m˚ pBq.

Proof. Suppose A Ă B Ă R. If U is an open set that contains B, then it

also contains A. Thus,

inf t`pU qB Ă U u ě rinf t`pU qA Ă U u ,

and so
m˚ pBq ě m˚ pAq

Definition 7. If A Ă R and if a P R, then

a ` A “ ta ` x | x P Au.

Proposition 3. Given A Ă R and a P R, m˚ pAq “ m˚ pa ` Aq

2.1. OUTER MEASURES 15

Proof. First note that if I is an interval then `pIq “ `pa ` Iq.

We then note that if tIj ujPF is a pairwise disjoint family of intervals,
ta ` Ij ujPF is as well and
ď ď
a` Ij “ pa ` Ij q.

If follows that if U is open `pU q “ `pa ` U q. If U is an open set that

includes A, then a ` U includes a ` A and has the same length as U . Thus,

inf t`pU q| A Ă U and U open u ě inf t`pV q| a ` A Ă U and V open u

and so,
m˚ pAq ě m˚ pa ` Aq.
Note that A “ ´a ` pa ` Aq, so by a symmetric argument we have that

m˚ pa ` Aq ě m˚ pAq.

Exercise 4. If A Ă R and if λ ą 0, then λA is defined to be the set

λA “ tλx | x P Au.

Show that m˚ pλAq “ λm˚ pAq.

Here is a more challenging version of the previous exercise (since we

haven’t defined outer measure in Rn ).

Exercise 5. If A Ă Rn and if M is an n ˆ n matrix, then M A is defined

to be the set
M A “ tM x | x P Au.
What is m˚ pM Aq?

Here is some food for thought, which we will revisit later in the course.

Exercise 6. If A, B Ă R, then

A ` B “ ta ` b | a P A, b P Bu.

What can we say about m˚ pA ` Bq in terms of m˚ pAq and m˚ pBq? Are

there sets with m˚ pAq “ m˚ pBq “ 0 but m˚ pA ` Bq ‰ 0?
16 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

2.1.2 Countable sub-additivity

We now establish one of the most important properties of length, which is
that it is countably sub-additive.

Definition 8. An measure is said to be countably sub-additive if

˜ ¸
8
ď ÿ8
˚
m Ek ď m˚ pEk q
i“1 i“1

for any countable collection of sets Ek (disjoint or otherwise).

If one of the Ek ’s has infinite outer measure, the right hand side is
infinite, and there is nothing to prove. We therefore suppose each of the Ek
’s has finite outer measure. Let  ą 0. For each natural number k, there is
a countable collection tIk,i u8
i“1 of open, bounded intervals for which

8
ď 8
ÿ
Ek Ď Ik,i and ` pIk,i q ă m˚ pEk q ` {2k .
i“1 i“1

Now, tIŤk,i u1ďi,kă8 is a countable collection of open, bounded invervals

that cover 8 k“1 Ek . Thus we have that

˜ ¸ « ff
8
ď ÿ 8
ÿ 8
ÿ
˚
m Ek ď ` pIk,i q “ ` pIk,i q
k“1 1ďk,iă8 k“1 i“1
8
ÿ” ı
ă m˚ pEk q ` {2k
k“1
« ff
8
ÿ
“ m˚ pEk q ` 
k“1

Note that all of the terms are positive, so we can rearrange terms in the
final equality. Since  was arbitrary, we obtain the desired inequality.
Note also that if tEk unk“1 is a finite collection of sets, we can prove finite
sub-additivity by taking Ek “ H for k ą n.

2.2 The need for σ-algebras

The outer measure gives our first definition of length for arbitrary sets, and
we have seen that it does a pretty good job. However, it has one major
2.2. THE NEED FOR σ-ALGEBRAS 17

drawback, which is that it fails to be countably additive, or even finitely

additive.
In other words, it is possible to find two sets A and B which are disjoint
and satisfy
m˚ pA Y Bq ă m˚ pAq ` m˚ pBq.
It is not really possible to construct these sets by hand. In particular, it
requires the Axiom of Choice. However, the fact that the counterexamples
are so pathological suggests that to fix this problem, all we need to do is
restrict our attention to sets which are somehow reasonable, which leads
directly into the notion of measurability and σ-algebras.

Definition 9. Let A be a set of reals. A is measurable if m˚ pXq “ m˚ pA X

Xq ` m˚ pAc X Xq for every X Ď R.

Here, we use the notation X c “ R ´ X whenever X Ď R..

Remark: A is measurable if and only if m˚ pXq ě m˚ pAXXq`m˚ pAc XXq
for every X Ď R.

Example. We can immediately find some examples of measurable sets.

1. If m˚ pAq “ 0, then A is measurable. To see this, suppose m˚ pAq “ 0.

Since m˚ is monotonic, m˚ pA X Xq “ m˚ pA X X c q “ 0 for every
X Ď R.
As a result, H, and every singleton are measurable.

2. Since pAc qc “ A, for any set A which is measurable, its complement

Ac must also be measurable. Thus, R (the complement of the empty
set) is measurable.

After looking at the definition of measurable sets, we see that there is

an algebraic structure to measurable sets, which we can make precise with
the notion of a σ-algebra.

Definition 10. Suppose S Ď PpRq. S is a σ-algebra over R if it meets the

following criteria.

1. R P S.

2. For every X P S, X c P S. (Closure under complements)

For every sequence of sets tAn u8

3. Ť n“0 so that An P S for every n P N,
8
A
n“0 n P S. (Closure under countable unions)
18 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Suppose S is a σ-algebra over R.

1. If X, Y P S, then X Y Y P S.ŤProof: Set X0 “ X, X1 “ Y , Xn “ H

when n ą 1. Then, X Y Y “ 8 n“0 Xn .

2. If Xn P S for each n P N, then 8

Ş
n“0 Xn P S.
Proof: DeMorgan’s laws.

Example. There are a few simple examples of σ-algebras that we can write
down immediately.

1. PpRq is a σ-algebra over R.

2. tH, Ru is a σ-algebra over R.

Our next goal now is to show that the set of measurable subsets of R is
a σ-algebra. To do so, we will need several lemmas.

Lemma 1. Suppose E1 , . . . , Em are measurable subsets of R. Then, m

Ť
j“1 Em
is measurable.

Proof. We will use induction (on m). The base case, where m “ 1 is trivial,
so the hard work
Ť will be the inductive step.
Suppose m´1
Ťm´1
j“1 Ej is a measurable set. Let F1 “ j“1 Ej . Let F2 “ Em .
We must show that F1 Y F2 is measurable. Let X Ď R. We need to show

m˚ pXq “ m˚ pX X pF1 Y F2 qq ` m˚ pX X pF1 Y F2 qc q.

What follows is a sequence of calculations by intersecting various different

sets to get the desired equality.
Claim 1: m˚ pXq “ m˚ pX XF1 q`m˚ pX XF1c XF2 q`m˚ pX XpF1 YF2 qc q.

Since F1 is measurable, we have that

m˚ pXq “ m˚ pX X F1 q ` m˚ pX X F1c q.

Since F2 is measurable,

m˚ pX X F1c q “ m˚ pX X F1c X F2 q ` m˚ pX X F1c X F2c q

“ m˚ pX X F1c X F2 q ` m˚ pX X pF1 Y F2 qc q.

So:

m˚ pXq “ m˚ pX X F1 q ` m˚ pX X F1c X F2 q ` m˚ pX X pF1 Y Fc qc qq.

2.2. THE NEED FOR σ-ALGEBRAS 19

Claim 2: m˚ pX X F1 q ` m˚ pX X F1c X F2 q “ m˚ pX X pF1 Y F2 qq.

Since F1 is measurable,

m˚ ppX X F1 q Y pX X pF1c X F2 qqq “ m˚ pppX X F1 q Y X X pF1c X F2 qq X F1 q

`m˚ pppX X F1 q Y pX X pF1c X F2 qqq X F1c q
“ m˚ pX X F1 q ` m˚ pX X pF1c X F2 qq.

But m˚ pX X F1 Y X X F1c X F2 q “ m˚ pX X pF1 Y F2 qq.

Therefore, m˚ pXq “ m˚ pX X pF1 Y F2 qq ` m˚ pX X pF1 Y F2 qc q.

This lemma has the following important corollary.

Corollary 1. If X, Y Ď R are measurable, then so are X X Y and X ´ Y .

The proof of this follows by applying De Morgan’s laws, so I’ll leave it

as a small exercise.
Using the same technique as the previous lemma, it is also possible to
prove the following lemma. Since the idea is exactly the same, we will not
write down the proof.

Lemma 2. Suppose A, E1 , . . . , Em Ď R. Suppose that E1 , . . . , Em are mea-

surable and that pE1 , . . . , Em q is pairwise disjoint. Then,
m
ď m
ÿ
m˚ pA X Ej q “ m˚ pA X Ej q.
j“1 j“1

We will now prove another helpful lemma towards the goal of showing
that measurable sets form a σ-algebra.

Lemma 3. Suppose S Ď PpRq. Then, S is a σ-algebra over R if and only

if it meets the following three criteria.

1. R P S.

2. X ´ Y P S whenever X, Y P S.
Ť8
3. n“0 An P S whenever tAn u8
n“0 is a pairwise disjoint sequence of sets
in S.
20 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Proof. If S is a σ-algebra, then by definition and the previous corollary, S

satisfies these three criteria.
Conversely, suppose S satisfies these three criteria. Then, R P S, and
(2) implies closure under complementation. It only remains to show that S
is closed under countable unions.

Claim 1: X Y Y P S whenever X, Y P S.
To show this, suppose X, Y P S. Set X0 “ X, X1 “ Y ´X, and Xn “ H
when n ą 1. It follows from (2) that H P S and Y ´ XŤP S. By construc-
8 8
tion, tXn uŤ
n“0 is a pairwise disjoint sequence. So, by 3, n“0 Xn P S. But,
8
X Y Y “ n“0 Xn .

Claim 2: S is closed under countable unions.

Let tAn u8
n“0 be a sequence of sets in S. For each n P N, let
ď
B n “ An ´ Aj .
jăn
Ť
By Claim 1, jăn Aj P S for each n. By (2), Bn P S for each n P N.
8
Ť8
By construction,
Ť8 tB n u
Ť8n“0 is pairwise disjoint. Thus, by (3), n“0 Bn P S.
However, n“0 An “ n“0 Bn .

Thus, S is a σ-algebra.

With this lemma in hand, we can finally show that the Lebesgue mea-
surable sets form a σ-algebra.

Theorem 4. The set of all measurable subsets of R is a σ-algebra over R.

Proof. We apply Lemma 3. We have already observed that R is measurable

and that the set of measurable sets is closed under set differences.
What remains to show is that the set of measurable sets is closed under
countable disjoint unions. Suppose tEŤn u8n“0 is a pairwise disjoint sequence
8
of measurable subsets of R. Set E “ n“0 En . We need to show that E is
measurable, i.e., for X Ď R, we have that

m˚ pXq “ m˚ pX X Eq ` m˚ pX X E c q.
ř8
Claim 1: m˚ pXqX ě ˚
n“0 m pX X En q ` m˚ pX X E c q.
2.2. THE NEED FOR σ-ALGEBRAS 21

Ťm
Let m P N. By Lemma 1, n“0 En
is measurable. So,
˜ ¸c
m
ď m
ď
˚ ˚ ˚
m pXq “ m pX X En q ` m pX X En q.
n“0 n“0
By Lemma 2,
m
ď m
ÿ
m˚ pX X En q “ m˚ pX X En q.
n“0 n“0
Ťm c
At the same time, p n“0 En q Ě Ec. So,
˜ ¸c
m
ď
m˚ pX X En q ě m˚ pX X E c q.
n“0

Thus,
m
ÿ
m˚ pXq ě m˚ pX X En q ` m˚ pX X E c q
n“0
for every m P N. Thus, Claim 1 holds.

Claim 2: m˚ pXq ě m˚ pX X Eq ` m˚ pX X E c q.

By Claim 1,
8
ÿ
m˚ pXq ě m˚ pX X En q ` m˚ pX X E c q.
n“0
Since outer measure is countably subadditive,
8
ÿ 8
ď
m˚ pX X En q ě m˚ p X X En q.
n“0 n“0
This establishes Claim 2.

Thus, m˚ pXq “ m˚ pX X Eq ` m˚ pX X E c q.
Example. We are able to use this idea to find many new examples of mea-
surable sets.
1. Every countable set of reals is measurable.
2. The set of irrational numbers is measurable.
Ş8
3. If An is a measurable set of reals for each n P N, then n“0 An is
measurable.
22 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

2.2.1 Borel sets

We now define a class of sets called Borel sets which are “always measurable”
(for any reasonable measure). It is worth noting that there are measurable
sets2 which are not Borel. However, Borel sets are those which must be
measurable whenever open intervals are measurable.

Definition 11. Suppose X Ď R. X is Borel if it belongs to every σ-algebra

over R that contains all the open subsets of R.

Remarks:

1. Every open set is Borel.

2. If X is a Borel set of reals, then X c is Borel.

Ť8
3. If
Ş8 Xn is a Borel set of reals for each each n P N, then n“0 Xn and
n“0 Xn are Borel.

Examples:
2
Here, we mean measurable with respect to the Lebesgue measure.
2.2. THE NEED FOR σ-ALGEBRAS 23

1. If x P R, then txu is Borel. Proof: txu “ pp´8, xq Y px, 8qqc .

2. Every countable set of reals is Borel.

3. The set of rational numbers is Borel, and the set of irrational numbers
is Borel.

4. pa, bs is Borel if a, b are real numbers such that a ă b.

Intuitively, the Borel sets are those which you can obtain by taking
countable unions, intersections and complements of open sets, all of which
preserve measurability. It is possible to explore the structure of these sets in
a much deeper way. In particular, it turns out there is a hierarchy of Borel
sets depending on how complicated they are. However, we will not need to
discuss this topic in this class.
We now show that Borel sets are always measurable. This fact should
not be surprising. However, there is still some work to do for this since we
haven’t yet shown that open intervals are measurable.

Theorem 5. Every Borel set of reals is measurable.

To prove this, we only must show that every open set is measurable. In
fact, we need only show that if a ă b, the open interval pa, bq is measurable
(since every open set is the countable union of disjoint intervals). Going
further, it suffices to show that if a P R, that the two sets

p´8, aq and pa, 8q

are measurable (Exercise: why?), so that’s what we are going to do.

Lemma 4. If a P R, the set p´8, aq is measurable.

Proof. Suppose a P R and X Ď R. We need to show that

m˚ pXq ě m˚ pX X p´8, aqq ` m˚ pX X ra, 8qq.

Note that

m˚ pX X ra, 8qq ď m˚ pX X pa, 8qq ` m˚ pX X tauq,

by sub-additivity but the latter term is zero, and thus

m˚ pX X ra, 8qq “ m˚ pX X pa, 8qq.

24 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Now we show that

m˚ pXq “ m ˚
pX X p´8, aqq ` m
loooooooooomoooooooooon
˚
pX X pa, 8qq
looooooooomooooooooon
X1 X2

Since outer measure is monotonic, we can assume that both X1 and X2

are non-empty (or else our job is already done). Now we take r ą m˚ pXq.
Then there is a non-empty set U Ď R such that X Ă U and r ą `pU q. Let

U1 “ U X p´8, aq and U2 “ U X pa, 8q.

Then U1 and U2 are both non-empty and `pU q “ `pU1 q ` `pU2 q. (For the
latter claim, see the solution to problem 6 on homework 1).
But `pU1 q ě m˚ pX Xp´8, aqq, by definition and `pU2 q ě m˚ pX Xpa, 8qq,
so r ě m˚ pX X p´8, aqq ` m˚ pX X pa, 8qq for all r ą m˚ pXq.
Thus,
m˚ pXq ě m˚ pX X p´8, aqq ` m˚ pX X pa, 8qq.

2.3 Approximating Measurable Sets

We now turn our attention to approximating measurable sets. These approx-
imations will be essential when we start using measure theory to compute
integrals. To do so, we define the notions of Gδ and Fσ sets, which are used
as outer and inner approximations of measurable sets, respectively.

Definition 12. Let X Ď R.

1. X is Gδ if it is the intersection of a countable family of open sets.

2. X is Fσ if it is the union of a countable family of closed sets.

Remarks:

1. If X Ď R, then X is Gδ if and only if X c is Fσ .

2. If X is a Gδ set of reals, then there

Ş8 is an increasing sequence of open
8
sets of reals tUn un“0 so that X “ n“0 Un . If X is an Fσ set of reals,
then thereŤis a decreasing sequence of closed sets of reals tCn u8
n“0 so
8
that X “ n“0 Cn .

Before discussing these sets in detail, let’s see a few examples

2.3. APPROXIMATING MEASURABLE SETS 25

Example. 1. The set of rational numbers is an Fσ set. Namely,

ď
Q“ tqu.
qPQ

2. Therefore the set of irrational numbers is a Gδ set.

3. If a, b are real numbers so that a ă b, then pa, bs is a Gδ set. Proof:

č
pa, bs “ pa, b ` 2´n q.
nPN

4. Every open interval is both Gδ and Fσ . Proof: every open set is Gδ .

If I is an open interval, then
ď
I“ ra, bs.
a,bPIXQ

However, not every set will be Gδ . For instance, Q is not Gδ .

Proof. Suppose that it were. In other words, suppose Q “ 8

Ş
n“0 Un where
Un is an open set of reals for each n P N. Let tqn u8n“0 be an enumeration
of the rational numbers. Thus, Un ´ tq u
Şn8 is open and dense for each n P N.
So, by the Baire Category Theorem n“0 pUn ´ tqn uq is dense. But, this
intersection is empty- a contradiction.

In order to use Gδ and Fσ sets to approximate measurable sets, we must

first establish the following excision property.

Lemma 5. If X is a measurable set of reals that has finite outer measure,

and if Y is a set of reals that includes X, then

m˚ pY ´ Xq “ m˚ pY q ´ m˚ pXq.

Proof. Suppose X is a measurable set of reals and m˚ pXq ă 8. Suppose

R Ě Y Ě X. Since X is measurable,

m˚ pY q “ m˚ pY X Xq ` m˚ pY X X c q “ m˚ pXq ` m˚ pY ´ Xq.

Thus, since m˚ pXq ă 8, m˚ pY q ´ m˚ pXq “ m˚ pY ´ Xq.

Using this lemma, we are finally able to prove the simple fact that the
outer measure of an open set is its length.
26 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Lemma 6. If U is a non-empty open set of reals, then m˚ pU q “ `pU q.

Proof. Suppose U is a non-empty open set of reals.

Let
S “ t`pV q : R Ě V Ě U | V ‰ H is openu.
Thus, m˚ pU q “ infpSq. Since `pU q P S, m˚ pU q ď `pU q. If V is a non-empty
open set of reals that includes U , then `pV q ě `pU q (See solution to problem
6 homework 1). Thus, m˚ pU q ě `pU q.

We can now describe the relationship between measurable sets and their
approximations.

Theorem 6. Let X Ď R. Then, the following are equivalent.

1. X is measurable.

2. For every  ą 0, there is an open U Ď R so that X Ď U and m˚ ˚ pU ´

Xq ă .

3. There is a Gδ set of reals V so that V Ě X and so that m˚ pV ´Xq “ 0.

4. For every  ą 0, there is a closed C Ď R so that X Ě C and so that

m˚ pX ´ Cq ă .

5. There is an Fσ set of reals V so that V Ď X and m˚ pX ´ V q “ 0.

Proving this is somewhat tedious, but it’s not too hard.

Proof. Proof:
(1) ñ (2): Suppose X is measurable. Let  ą 0.

Case 1: m˚ pXq ă 8.

By the definition of outer measure, there is a non-empty open set of

reals U so that `pU q ă m˚ pXq `  and so that U Ě X. Since m˚ pXq ă 8,
`pU q ´ m˚ pXq ă . By Lemma 6, `pU q “ m˚ pU q. Since X is measurable,
by Lemma 5, m˚ pU ´ Xq “ m˚ pU q ´ m˚ pXq. So, m˚ pU ´ Xq ă .

Case 2: m˚ pXq “ 8.
Ť
For each n P N, let Xn “ X X p´n, nq. Thus, X “ n Xn . For each
n P N, m˚ pXn q ď m˚ pp´n, nqq ă 8 since outer measure is monotonic.
By Case 1, for each n P N there is an open set of reals Un Ě Xn so that
2.3. APPROXIMATING MEASURABLE SETS 27

Ť Ť
m˚ pUn ´ Xn q ă 2´pn`1q . Set U “ n Un . Then, U ´ X Ď n pUn ´ Xn q.
So,
ď 8
ÿ
m˚ pU ´ Xq ď m˚ p Un ´ Xn q ă 2´pn`1q “ .
n n“0

(2) ñ (3): Suppose (2). Then, for each n P N, there is an openŞset of

reals Un so that Un Ě X and so that m˚ pUn ´ Xq ă 2´n . Set V “ n Un .
So, for each n P N,

m˚ pV ´ Xq ď m˚ pUn ´ Xq ă 2´n .

Thus, m˚ pV ´ Xq “ 0.

(3) ñ (1): Suppose there is a Gδ set V so that V Ě X and m˚ pV ´ Xq “

0. Thus, V is Borel and so V is measurable. Since m˚ pV ´ Xq “ 0, V ´ X
is measurable. Since pV ´ Xqc “ V c Y X, X “ V X pV ´ Xqc . Thus, X is
measurable.

The remaining implications are now proved by taking complements.

We can now show that measurable sets are open intervals, modulo a set
of small measure.
Theorem 7. Suppose X is a measurable set of reals whose outer measure
is finite.ŤThen, for every  ą 0, there are open intervals I0 , . . . , Ik so that
m˚ pX∆ kn“0 In q ă  and so that pI0 , . . . , Ik q is pairwise disjoint.
Proof. Let  ą 0. By Theorem 6, there is an open set of reals U so that
U Ě X and so that m˚ pU ´ Xq ă {2. There is a countable
Ť and pairwise
disjoint family of open intervals tIj ujPF so that U “ jPF Ij .

Claim 1: `pU q ă 8.

Proof Claim 1: Since m˚ pU q ď m˚ pU ´ Xq ` m˚ pXq ă {2 ` m˚ pXq and

m˚ pXq ă 8, m˚ pU q ă 8. By Lemma 6, m˚ pU q “ `pU q.
ř
Claim 2: There is a finite F1 Ď F so that jPF ´F1 m˚ pIj q ă {2.
ř
Proof of Claim 2: Since jPF `pIj q “ `pU q ă 8.
Ť
Claim 3: m˚ pX∆ jPF1 Ij q ă .
28 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Proof of Claim 3: By monotonicity of outer measure and Claim 2:

ď
m˚ p Ij ´ Xq ď m˚ pU ´ Xq ă {2
jPF1
ď ď
˚
m pX ´ Ij q ď m˚ pU ´ Ij q ă {2
jPF1 jPF1

By the countable subadditivity of m˚

ď
m˚ pX∆ Ij q ă .
jPF1

2.3.0.1 The Vitali covering lemma

Before defining the Lebesgue measure, let us also provide the Vitali covering
lemma, which will be quite useful later in the course.

Definition 13. An interval is degenerate if it consists of a single point.

Definition 14. Suppose E Ď R. Suppose F is a set of intervals that are

closed, bounded, and non-degenerate. F is a Vitali covering of E if for every
x P E and every  ą 0 there is an interval I P F so that x P I and `pIq ă .

Here are some examples and non-examples of Vitali coverings.

1. tr´1, 2su is a covering of r0, 1s but not a Vitali covering.

2. trx ´ 2´n , x ` 2´n s : x P r0, 1s and n P Nu is a Vitali covering of

r0, 1s.

Lemma 7 (Vitali Covering Lemma). Suppose E is a set of reals and m˚ pEq ă

8. Suppose F is a Vitali covering of E. Then, for every  ą 0, there exist
I0 , . . . , In P F so that pI0 , . . . , In q is pairwise disjoint and so that
n
ď
m˚ pE ´ Ij q ă .
k“0
Ťn
Proof. Case 1: There exist I0 , . . . , In P F so that E Ď j“0 Ij and so that
pI0 , . . . , In q is pairwise disjoint.

In this case, there is nothing to prove and we are done.

2.3. APPROXIMATING MEASURABLE SETS 29

Ťn
Case 2: There do not exist I0 , . . . , In P F so that E Ď j“0 Ij and so
that pI0 , . . . , In q is pairwise disjoint.

By the definition of outer measure, there is an open set U Ě E so that

`pU q ă 8. Let F 1 “ tI P F : I Ď U u.

Claim 1: F 1 is a Vitali covering of E.

Proof: Exercise.

Claim 2: If I0 , . . . , In P F 1 ,Ťand if pI0 , . . . , In q is pairwise disjoint, then

there exists I P F 1 so that I X nj“0 Ij “ H.

Proof: Exercise.

We inductively define a sequence of intervals in F 1 as follows. Let I0 be

any interval in F 1 . Assume I0 , . . . , In have been defined. Let
n
ď
Fn “ tI P F 1 : I X Ij “ Hu.
j“0

By Claim 2, Fn is nonempty. Let sn “ supt`pIq : I P Fn u. Since I Ď U

for each I P F 1 , sn ă 8. So, there exists I P Fn such that `pIq ą sn {2; let
In`1 denote such an interval.
Note that tIn u8 n“0 is pairwise disjoint.

ř8
Claim 3: n“0 `pIn q ă 8.

3: µp 8
Ť
ŤProof Claim n“0 In q ď µpU q ă 8. By countable additivity,
µp 8 8
ř
I
n“0 n q “ n“0 `pIn q.

Thus, limnÑ8 `pIn q “ 0.

For each n P N, let cn “ center of In and let rn “ radius In . i.e.

In “ rcn ´ rn , cn ` rn s. Let Jn “ rcn ´ 5rn , cn ` 5rn s. Thus, `pJn q “ 5`pIn q.
Ťn Ť8
Claim 4: For every n P N, E ´ k“0 Ik Ď k“n`1 Jk .
Ťn
Proof Claim
Ťn 4: Let x P E ´ k“0 Ik . 1 There is a δ ą 0 so that
px´δ, x`δqX k“0 Ik “ H. There is an I P F so that x P I Ď px´δ, x`δq.
30 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

Thus, I P Fn .

Subclaim 4.1: There exists n0 so that I X In0 ‰ H.

Proof of Subclaim 4.1: Suppose not. Then, I P Fn0 for all n0 . Thus,
`pIn`1 q ą `pIq{2 ą 0 for all n. Thus, limnÑ8 `pIn q ‰ 0- a contradiction.

Choose the least such n0 . Therefore, n0 ą n. Also, I P Fn0 ´1 , so

`pIn0 q ą `pIq{2.

Subclaim 4.2: x P Jn0 .

Proof of Subclaim 4.2: By Subclaim 4.1, there is a point x0 P I X In0 .

Then,
|x ´ x0 | ď `pIq ă 2`pIn0 q

and thus
x0 ´ xn0 ď `pIn0 q{2.

So, |x ´ xn0 | ď 4rn0 ` rn0 “ 5rn0 , which implies x P Jn0 .

This proves Claim 4.

Now, let  ą 0. There exists n so that

8
ÿ
`pIk q ă {5.
k“n`1

Since µpJk q “ 5µpIk q, we find that

8
ď
µp Jn q ă .
n“k`1

So, by Claim 4,
n
ď
µpE ´ Ik q ă .
k“0
2.4. THE LEBESGUE MEASURE 31

2.4 The Lebesgue Measure

With all these preliminaries taken care of, we are finally able to define the
notion of the Lebesgue measure.

Definition 15. If E is a measurable set of reals, then the outer measure of

E is also called the measure of E and is denoted µpEq.

From our previous examples, there are a few sets which we can immedi-
ately measure.

Example. 1. The Lebesgue measure of an interval is its length.

2. µpQq “ 0.

Before moving on to proving some properties of the Lebesgue measure,

let’s solve a problem from a past qualifying exam.

Exercise 7 (2000 Qual Problem 1). 1. Define the measure µ on the Borel-
σ-algebra of R by µpAq “ λpA X p0, 1qq for Borel sets A. (λ “ Lebesgue
measure) Let
č
K “ tA : A is a closed set such that µpAq “ 1u

and č
D“ tG : G is an open set such that µpGq “ 1u
Determine precisely which points belong to K and D and prove your claim.

2.4.1 Countable Additivity

The Lebesgue measure has the advantage of being countable additive, which
was the entire reason that we needed to discuss measurable sets.

Theorem 8 (Countable Additivity). : If Ť tEn u8

n“0 is ařpairwise disjoint
sequence of measurable sets of reals, then µp 8n“0 En q “
8
n“0 µpEn q.

Proof. Suppose tEn u8

n“0 is a pairwise disjoint sequence of measurable sets
of reals.

Claim 1: µp 8
Ť ř8
n“0 En q ď n“0 µpEn q.

This holds since outer measure is countably subadditive.

32 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

řm
Claim 2: For each m P N, µp 8
Ť
n“0 En q ě n“0 µpEn q.

ŤmTo establish this, we use the finite additivity. Let m P N. Set A “

n“0 En . By Lemma 2,

m
ď m
ÿ
µpA X En q “ µpA X En q.
n“0 n“0
Ťm Ťm
But, A X n“0 En “ n“0 En and A X En “ En . Since outer measure is
monotonic,
m
ď 8
ď
µp En q ď µp En q.
n“0 n“0

Ť8 ř8
Claim 3: µp n“0 En q ě n“0 µpEn q.

To see this, apply Claim 2 and take limits.

As a corollary, we immediately obtain finite additivity as well.

Corollary 2. If E0 , . . . , EmŤ
are measurable
řmsets of reals, and if pE0 , . . . , Em q
is pairwise disjoint, then µp m E
n“0 n q “ n“0 µpEn q.

We also obtain the measure of the difference of sets is the difference of

measures.

Corollary 3. Suppose X, Y are measurable sets of reals so that Y Ě X and

so that µpXq ă 8. Then, µpY ´ Xq “ µpY q ´ µpXq.

Proof. By Corollary 2, µpY q “ µpY ´ Xq ` µpXq. Since µpXq ă 8, µpY q ´

µpXq “ µpY ´ Xq.

2.4.2 Continuity of measure

The Lebesgue measure has another desirable property, which is that it is
continuous with respect to increasing and decreasing sequences of sets.

Theorem 9 (Continuity of measure). Suppose tEn u8 n“0 is a sequence of

measurable sets of reals.

1. If En Ď En`1 for all n P N, then µp 8

Ť
n“0 En q “ limmÑ8 Em .
2.4. THE LEBESGUE MEASURE 33

Ş8
2. If En Ě En`1 for all n P N, and if µpE0 q ă 8, then µp n“0 En q “
limmÑ8 µpEm q.
Proof. (1): Suppose En Ď En`1 for all n P N.

Case 1: µpEm q “ 8 for some m.

By monotonicity of outer measure, µp 8

Ť
n“0 En q “ 8. By monotonicity
of outer measure, µpEn q “ 8 for all n ą m. Thus, limnÑ8 µpEn q “ 8.

Case 2: µpEm q ă 8 for all m.

Ť8 Ť8
We have n“0 En “ E0 Y n“0 µpEn`1 ´ En q. By Theorem 8,
8
ď 8
ÿ
µpE0 Y µpEn`1 ´ En qq “ µpE0 q ` µpEn`1 ´ En q.
n“0 n“0

By Corollary 3, µpEn`1 ´ En q “ µpEn`1 q ´ µpEn q for all n. Thus,

8
ÿ 8
ÿ
µpEn`1 ´ En q “ µpEn`1 q ´ µpEn q “ lim µpEn q ´ µpE0 q.
nÑ8
n“0 n“0

Thus, µp 8
Ť
n“0 En q “ limnÑ8 µpEn q.

(2): Suppose EnŞĚ En`1 for all n P N and µpE0 q ă 8. Thus, µpEn q ă 8
for all n P N and µp 8
n“0 En q ă 8. Also, E0 ´ En Ď E0 ´ En`1 for all n P N.
Then, using the previous argument ( Part 1),
8
ď
µp E0 ´ En q “ lim µpE0 ´ En q.
nÑ8
n“0
Ť8 Ş8
However, n“0 E0 ´ En “ E0 ´ n“0 En . By Corollary 3,
8
č 8
č
µpE0 ´ En q “ µpE0 q ´ µp En q and µpE0 ´ En q “ µpE0 q ´ µpEn q.
n“0 n“0

As a result, we find that

8
č
µp En q “ lim µpEn q.
nÑ8
n“0
34 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

2.4.3 Everywhere and Almost Everywhere

Measure theory is a powerful tool for measuring the size of sets. However,
as we have seen it is possible for a non-empty set to have non-zero measure
and the Lebesgue measure is unable to distinguish these sets from the empty
set. As such, for the rest of the course, we will often talk about properties
that hold almost everywhere, as opposed to those that hold everywhere.
Definition 16. We say that a property holds almost everywhere in a set E
if the set of all x P E so that the property fails has measure 0. The set of
all x where the property fails is called an exceptional set.
Example. Almost all real numbers are irrational.
The following lemma is extremely important.
Lemma 8 (Borel-Cantelli).
ř8 Suppose tAn u8 n“0 is a sequence of measurable
sets of reals so that n“0 µpAn q ă 8. Then, almost every real number
belongs to only finitely many An ’s. i.e. the measure of
tx P R : x P An for infinitely many nu
is zero.
Proof. Let
E “ tx P R : x P An for infinitely many nu.
We need to show that µpEq “ 0. Note that x P E if and only if for every
n P N there exists m ě n so that x P Am . So:
č8 ď8
E“ Am
n“0 m“n
ř8
ă 8, µp 8
Ť Ť Ť
Note that m“n Am Ě m“n`1 Am . Since n“0 µpAn q n“0 An q ă
8. So by Theorem 2,
8
ď
µpEq “ lim µp Am q.
nÑ8
m“n
However,
8
ď 8
ÿ
µp Am q ď µpAm q.
m“n m“n
ř8
Since m“0 µpAm q ă 8,
8
ÿ
lim µpAm q “ 0.
nÑ8
m“n
Thus, µpEq “ 0.
2.5. NON-MEASURABLE SETS 35

2.5 Non-Measurable Sets

Most sets that we encounter “in the wild” are Borel sets, and thus measur-
able. However, there is certainly some sampling bias going on here, because
in some sense, “most” sets are non-measurable3 .
Constructing non-measurable sets requires some work, so we take a brief
moment to review equivalence relations. Let us first introduce some nota-
tion.

Definition 17. Suppose X is a set, and suppose „ is an equivalence relation

on X.

1. For each c P X, let rcs„ “ the „-equivalence class of c.

2. X{ „ denotes the set of all „-equivalence classes.

Definition 18. Suppose X is a set, and suppose „ is an equivalence relation

on X. A system of representatives for „ is a set R Ď X so that each „-
equivalence class contains exactly one element of R.

Example. Let n be a positive integer. Let „ denote the equivalence modulo

n relation on Z. Then t0, . . . , n ´ 1u is a system of representatives for „.
t1, . . . , nu is also a system of representatives for „.

The relationship between equivalence classes and system of representa-

tives quickly leads down the rabbit hole of foundational issues.

Question 1. Does every equivalence relation have a system of representa-

tives?

Definition 19 (Axiom of Choice). If X is a nonempty set, then there is a

function F : PpXq Ñ X so that F pY q P Y for each Y P PpXq.

With this axiom, we can find a system of representatives for any equiv-
alent relation.

Corollary 4. Assuming the Axiom of Choice, every equivalence relation has

a system of representatives.

Proof. Suppose X is a set and „ is an equivalence relation on X. If X “ H,

then H is a set of representatives for „. Let F : PpXq Ñ X be a function
so that F pY q P Y for each Y P PpXq. Let R “ F rX{ „s. We claim that
3
At least if we assume the axiom of choice.
36 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

R is a system of representatives for „. For, let a be a „-equivalence class.

Then, F paq P a. Suppose c P R and c ‰ F paq. There exists b P F { „ so
that F pbq “ c. Thus, b ‰ a. Therefore, b X a “ H. Since c P b, c R a. So,
a contains exactly one element of R.

The reason to consider equivalence relationships and systems of repre-

sentatives is that it allows us to construct non-measurable sets.
Definition 20. When x, y P R, write x „Q y if y ´ x P Q.
Note that „Q is an equivalence relation on R.
Proposition 10. Suppose R is a system of representatives for „Q . Then,
q ` R X q 1 ` R “ H whenever q, q 1 are rational numbers so that q ‰ q 1 .
Proof. Suppose q, q 1 P Q and that q ` R X q 1 ` R ‰ H. Then, there exist
c, c1 P R so that q ` c “ q 1 ` c1 . Thus, q ´ q 1 “ c1 ´ c. Thus, c „Q c1 . Since
R is a system of representatives for „Q , c “ c1 . Thus, q “ q 1 .

Theorem 11. Every system of representatives for „Q is non-measurable.

Proof. Let R be a set of representatives for „Q . By way of contradiction,
suppose R is measurable. Let Sn “ R X r´n, ns. Then, Sn is measurable for
each n P N. Also,
ď 8
ď ď 8
ď ď 8
ď
R“ λ`R“ λ ` Sn “ λ ` Sn .
λPQ λPQ n“0 m“0 λPQXr´m,ms n“0

We start by showing that if Sn was measurable, then µpSn q “ 0 for every

n P N.
Ť
Proof of Claim: Let n P N. Then, λPQXr´1,1s λ ` Sn Ď r´pn ` 1q, n ` 1s.
So, ď
µp λ ` Sn Ď r´pn ` 1q, n ` 1sq ă 8
λPQXr´1,1s

Thus, by Proposition 10,

ÿ
µpλ ` Sn q ă 8
λPQXr´1,1s

Since outer measure is translation-invariant:

ÿ
µpSn q “ 0
λPQXr´1,1s
2.5. NON-MEASURABLE SETS 37

Therefore, µpSn q “ 0.

From this, we see that µpRq “ 0, which is a contradiction.

As a result, if we assume the axiom of choice, we find that non-measurable

sets exist.

Corollary 5. There is a non-measurable set of reals.

Proof. By Corollary 4, there is a system of representatives for „Q . By

Theorem 11, such a system is non-measurable.

Using this, we can finally go back to show why the outer measure fails
finite additivity

Corollary 6. There exist A, B Ď R so that A X B “ H and m˚ pA Y Bq ă

m˚ pAq ` m˚ pBq.

Proof. Let E be a non-measurable set of reals. Then, there is a set of reals

X so that m˚ pXq ă m˚ pE X Xq ` m˚ pE c X Xq.

Proposition 12 (Vitali). Every set with positive outer measure includes a

non-measurable set.

Proof. Exercise.

2.5.0.1 Some remarks on non-measurable sets

The existence of measurable sets is an extremely thorny issue which is deeply
tied to foundational issues. In fact, Robert Solovay constructed a model (in
the sense of mathematical logic) of the real numbers where all subsets are
Lebesgue measurable. For this, he used the Zermelo Frenkel axioms with an
extra axiom on existence of an inaccessible cardinal. The technical details
are outside the scope of this course, but heuristically what this shows is that
the existence of non-measurable sets is very nearly equivalent to the axiom
of choice.
Historically, this phenomena, along with a higher-dimensional version
called the Banach-Tarski paradox4 lead many mathematicians to reject the
axiom of choice altogether.
4
The Banach-Tarski paradox decomposes a solid ball into 5 pieces and then rotates
and translates the pieces into two solid balls of the same radius. This construction heavily
relies on the axiom of choice and non-measurable sets.
38 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

For modern functional analysis, trying to avoid the axiom of choice or

the Banach-Tarski paradox is extremely restrictive, because you have to dis-
card the Hahn-Banach theorem [FW91], which is one of the most important
results in partial differential equations.
My advice is to not worry too much about the axiom of choice unless
you are interested in mathematical logic. In my experience, non-measurable
sets and other foundational matters don’t show up unless you go looking
for them, so it’s perfectly possible (and acceptable) for a mathematician to
consider such issues only in a superficial way.

2.6 The Cantor Set

We have seen that all countable sets have zero measure and that all Borel
sets are measurable. It is natural to ask whether the converse of these results
hold. In both cases, the answer is no, and we can see this by studying the
Cantor set.
There are several ways to define this set, so we will provide two.

Definition 21 (Cantor set - ternary expansion definition). The Cantor mid-

dle third set consists of all numbers in r0, 1s that have a ternary expansion
consisting of 0’s and 2’s. That is, x belongs to the Cantor ř8middle third
an
set if and only if there is a sequence tan u8n“0 so that x “ n“0 3n`1 and
an P t0, 2u for all n P N. We denote this set by C.

Definition 22 (Cantor set - intersection definition). Consider the closed

interval I “ r0, 1s and remove the open set p 31 , 23 q. Set
ˆ ˙ „  „ 
1 2 1 2
C1 “ I ´ , “ 0, Y ,1 .
3 3 3 3

Repeat the process of removing the middle third from each set to obtain
ˆ ˙ „  „  „  „ 
1 2 7 8 1 2 1 2 7 8
C2 “ C1 ´ p , q Y p , q “ 0, Y , Y , Y ,1 .
9 9 9 9 9 9 3 3 9 9

Cn is the disjoint union of 2n closed intervals of length 3´n . We then

define the Cantor set to be
č8
C“ Cn .
n“1

Exercise 8. Show that these two definitions of the Cantor set agree.
2.6. THE CANTOR SET 39

Figure 2.1: The Cantor Set

1
Exercise 9. Show that 4 is in the Cantor set.

Proposition 13. C is a closed set whose measure is zero.

Proof. C is closed since it is the intersection of closed sets. It has measure

zero since C Ă Cn for all n and thus

2n
µpCq ď µpCn q “ .
3n

Exercise 10. Show that C is uncountable.

Hint: Use the ternary definition and consider the map sending a num-
ber x written in ternary as .a1 a2 a3 . . . to the number written in binary as
. a21 a22 a23 . . .

Exercise 11. Use the Cantor set to construct a dense uncountable set of
real numbers with measure zero.
40 CHAPTER 2. MEASURABLE SETS AND THEIR MEASURES

2.6.1 Counting subsets

The Cantor set gives an example of an uncountable set with measure zero.
However, we can also use it to “construct” a Lebesgue measurable set which
is not Borel.

Proposition 14. Any subset of the Cantor set is measurable.

Proof. By the monotonicity of the Lebesgue measure, any subset of the

Cantor set has measure zero.

To see that this implies that there are Lebesgue measurable sets which
are not Borel, we can count the number of subsets of the Cantor sets, of
which there are
#pPpCqq “ #pPpRqq “ 2p#Rq .
On the other hand, there are #R open subsets of R (exercise: why?)
and every Borel set is constructed from countable intersections and unions
of open sets. Using transfinite induction and these two facts, it is possible
to show that there are #R Borel sets.
We will revisit the Cantor set in the next chapter to provide counterex-
amples for some intuitive ideas about continuous functions, their derivatives
and measures.
Chapter 3

Measurable Functions

We now turn our attention to measurable functions in order to lay the

foundations for the Lebesgue integral. The basic idea is to define it in a
similar way to continuity, but there are some subtleties.

Definition 23. A function f : X Ñ Y is continuous if the pre-image of

any open set f ´1 pU q is open.

Definition 24. A function f : E Ñ R (or r´8, 8s) is measurable if the

preimage of any open set f ´1 pU q is Lebesgue-measurable.

By considering f ´1 pRq, we can immediately see that the domain of the

function E must be measurable. We also see that every continuous real-
valued function on R is measurable.

Proposition 15. Suppose E is a measurable set of reals and f : E Ñ

r´8, 8s. Then, the following are equivalent..

1. f is measurable.

2. f ´1 rpa, 8ss is measurable for every a P R.

3. f ´1 rr´8, aqs is measurable for every a P R.

4. f ´1 rra, 8ss is measurable for every a P R.

5. f ´1 rr´8, ass is measurable for every a P R.

Proof. (Sketch) By definition, 1 implies all the other statements. If any of

2 - 5 hold, then by considering complements, countable unions, and finite
intersections, it follows that f is measurable.

41
42 CHAPTER 3. MEASURABLE FUNCTIONS

As an immediate consequence, we see the following corollary.

Corollary 7. Suppose f : E Ñ r´8, 8s is measurable. Then the preimage
f ´1 pU q of any Borel set is measurable.

3.0.1 A brief remark on σ-algebras

Much in the same way that we can use function(al)s to induce a topology (by
considering the coarsest topology in which a function (or family of functions)
is continuous, we can use functions to construct sigma algebras. This idea
will not be used heavily in this class, but if you are interested in stochastic
calculus or financial math, you will see it there.
Proposition 16. Suppose f : E Ñ R is a measurable function. Let S “
tX Ď R : f ´1 rXs is measurableu. Then, S is a σ-algebra over R.
Proof. Since f ´1 rRs “ E, and since E is measurable, R P S.
If f ´1 rXs is measurable, then f ´1 rX c s “ E ´ f ´1 rXs is measurable.
So, S is closed under
Ť8 complementation.
Ť8 If f ´1 rAn s is measurable for each
´1 ´1
n P N, then f r n“0 An s “ n“0 f rAn s is measurable. So, S is closed
under countable unions.

Now back to the regularly scheduled programming. . .

3.1 Sums, products, pointwise limits, etc.

We now show that some natural ways to combine measurable functions
produce more measurable functions.
Theorem 17. If f : E Ñ r´8, 8s is continuous, and if E is measurable,
then f is measurable.
Proof. Let U Ď r´8, 8s be an open set of reals. Since f is continuous,
there is an open set V so that f ´1 rU s “ V X E. Since V, E are measurable,
V X E is measurable.

Lemma 9. Suppose f : E Ñ r´8, 8s and g : E Ñ r´8, 8s are equal

almost everywhere (i.e. µptt P E : f ptq ‰ gptqu. If f is measurable, then g
is measurable.
Proof. Suppose f is measurable. Let E “ tx P E : f pxq ‰ gpxqu. Then, for
every U Ď r´8, 8s,

g ´1 rU s “ f ´1 rU s X pR ´ Eq Y g ´1 rU s X E.
3.1. SUMS, PRODUCTS, POINTWISE LIMITS, ETC. 43

Suppose U is open. Then, f ´1 rU s is measurable. Since E “ 0, g ´1 rU s X E

is measurable. Thus, g ´1 rU s is measurable.

Proposition 18. Suppose f : E1 Ñ r´8, 8s is measurable and g : E2 Ñ

r´8, 8s is continuous. Suppose E2 is open and E2 Ě ranpf q. Then, g ˝ f
is measurable.
Proof. Suppose U Ď r´8, 8s is open. Then, pg ˝ f q´1 rU s “ f ´1 rg ´1 rU ss.
Thus, g ´1 rU s is open. Thus, pg ˝ f q´1 rU s is measurable.

Theorem 19. Suppose f : E Ñ r´8, 8s and g : E Ñ r´8, 8s are mea-

surable. Suppose also that f and g are finite almost everywhere. Then, f g
is measurable and αf ` βg is measurable for all α, β P R.
Proof. By Lemma 9, we can assume that f and g are real-valued.

Claim 1: f ` g is measurable.

Proof Claim 1: Set h “ f ` g. Let a P R.

We have:
hpxq P pa, 8s ðñ f pxq ` gpxq ą a
ðñ f pxq ą a ´ gpxq
ðñ Dq P Q f pxq ą q ą a ´ gpxq
ðñ Dq P Q f pxq ą q and gpxq ą a ´ q
So, ď
h´1 rpa, 8ss “ f ´1 rpq, 8ss X g ´1 rpa ´ q, 8ss.
qPQ

Thus, h´1 rpa, 8ss

is measurable. Thus, by Proposition 15, h is measurable.
Claim 2: αf is measurable.

Proof of Claim 2: Set gpxq “ αx. Then, αf “ g ˝ f . Apply Proposition

18.

Claim 3: f g measurable.
Proof Claim 3: f g “ 21 rpf ` gq2 ´ pf ´ gq2 s. By Proposition 18, pf ` gq2
and pf ´ gq2 are measurable.

Definition 25. Suppose f : E Ñ r´8, 8s and g : E Ñ r´8, 8s. Then, for

all x P R, maxtf, gupxq “ maxtf pxq, gpxqu and mintf, gupxq “ mintf pxq, gpxqu.
44 CHAPTER 3. MEASURABLE FUNCTIONS

In other words, maxtf, gupxq is simply the larger of f pxq and gpxq and
conversely for mintf, gupxq.

Definition 26. Suppose f : E Ñ r´8, 8s. The positive part of f , denoted

f ` is defined as f ` “ maxtf, 0u. The negative part of f , denoted f ´ , is
defined f ´ “ ´ mintf, 0u.

Proposition 20. Suppose f : E Ñ r´8, 8s and g : E Ñ r´8, 8s. Sup-

pose f, g are measurable and finite almost everywhere. Then, maxtf, gu and
mintf, gu are measurable.

In fact, it is possible to prove this result without assuming that f and g

are finite almost everywhere, but this will not be necessary for our purposes.

Proof. Set h “ maxtf, gu. Suppose U is an open subset of r´8, 8s. Then,

h´1 rU s “ tx P E : f pxq P U and f pxq ą gpxqu

ď
tx P E : gpxq P U and gpxq ě f pxqu
ď
“ f ´1 rU s X pf ´ gq´1 rp0, 8ss g ´1 rU s X pg ´ f q´1 rr´8, 0ss.

Since f ´ g and g ´ f are measurable, it follows that h´1 rU s is measurable

if U is open. Thus, h is measurable.
mintf, gu “ ´ maxt´f, ´gu. Thus, mintf, gu is measurable.

Corollary 8. If f is measurable, then f ` and f ´ are measurable.

It is worth noting that these proofs would be easier were we able to

discuss measurability of higher dimensional sets (since max is naturally a
two-variable function). This is one of the downsides of sticking with one
variable, (the upside is that it is much easier to define the Lebesgue measure
on the real line compared to a higher dimensional space).
We now come to one of the most important facts about the measur-
able functions, which is that the pointwise limit of measurable functions is
measurable. Note that the pointwise limit of continuous functions may not
be continuous. It turns out that the pointwise limit of Riemann integrable
functions need not be Riemann integrable either, so this result will be one
of the major advantages of the Lebesgue integral.

Theorem 21. Suppose E is a measurable set of reals and that f, f0 , f1 , . . .

are extended real-valued functions on E so that f0 , f1 , . . . are measurable and
limnÑ8 fn pxq “ f pxq almost everywhere. Then, f is measurable.
3.2. THE CANTOR FUNCTION AND NON-MEASURABLE FUNCTIONS45

Proof. Let X “ tx P E : limnÑ8 fn pxq ‰ f pxqu. Suppose U Ď r´8, 8s is

open. Then,
f ´1 rU s “ X X f ´1 rU s Y X c X f ´1 rU s.
Suppose x R X. Then,

f pxq P U ô DN0 P N @n ě N0 fn pxq P U

ď8 8
č
ô xP fn´1 rU s.
N0 “0 n“N0

So,
ď 8
č
X c X f ´1 rU s “ X c X fn´1 rU s.
N0 PN n“N0

Thus, f ´1 rU s is measurable.

Definition 27. Notation: fn Ñ f a.e. is shorthand for tfn u8

n“0 converges to
8
f almost everywhere. fn Ñ f is shorthand for tfn un“0 converges pointwise
to f .
Finally, we note that the restriction of a measurable function to a mea-
surable sub-set is measurable.
Definition 28. f |E denotes the restriction of f to E when E Ď dompf q.
Namely, "
f pxq if x P E
f |E pxq “
undefined otherwise
Proposition 22. If f is measurable, and if E is a measurable subset of the
domain of f , then f |E is measurable.

3.2 The Cantor Function and non-measurable func-

tions
In the previous section, we saw multiple ways to combine measurable func-
tions to obtain new measurable functions. However, there was a conspicuous
absence: composition.
Observation 1. The composition of measurable functions is not necessarily
measurable!
To see this, we will revisit the Cantor set and use it define the Cantor-
Lebesgue function.
46 CHAPTER 3. MEASURABLE FUNCTIONS

an
Definition 29. Suppose x P C, and let x “ 8
ř
n“0 3n`1
where an P t0, 2u for
all n P N. Define
8
ÿ an {2
φpxq “ .
n“0
2n`1

Proposition 23. Suppose n0 P N. If x0 , x1 P C, and if |x0 ´ x1 | ă 3´n0 ,

then |φpx0 q ´ φpx1 q| ă 2´n0 .
To obtain the full Cantor-Lebesgue function, we extend φ to r0, 1s as
follows. For each x P r0, 1s ´ C, let

φpxq “ φpsuptx1 P C : x1 ă xuq.

φ is called the Cantor-Lebesgue function.

It is possible to calculate some values of this function explicitly.
1. If x P p 13 , 32 q, then φpxq “ 21 .

2. If x P p 79 , 98 q, then φpxq “ 43 .
We will now note some properties of this function
Proposition 24. If pa, bq Ď r0, 1s ´ C, then φ is constant on pa, bq and
φpaq “ φpbq.
Corollary 9. φ is non-decreasing.
Proposition 25. φ is continuous.
Proof. Let  ą 0 Choose n0 P N so that 2´n0 ă . Suppose x0 , x1 P r0, 1s
and |x0 ´ x1 | ă 3´n0 .
Claim: |φpx0 q ´ φpx1 q| ă 2´n0 .
Proof Claim: WLOG x0 ă x1 . Let:

x10 “ inftx P C : x0 ď xu
x11 “ suptx P C : x ď x1 u

Thus, x10 , x11 P C. By Proposition 24, φpx10 q “ φpx0 q and φpx11 q “ φpx1 q.

Case 1: x10 ą x1 or x11 ă x0 .

It follows that px0 , x1 q Ď r0, 1s ´ C. Thus, φpx0 q “ φpx1 q by Proposition

24.
3.2. THE CANTOR FUNCTION AND NON-MEASURABLE FUNCTIONS47

Figure 3.1: An approximation of the Cantor-Lebesgue function

48 CHAPTER 3. MEASURABLE FUNCTIONS

Case 1: x10 ď x1 and x0 ď x11 .

Thus, x10 , x11 P rx0 , x1 s. Therefore, |x10 ´ x11 | ă 3´n0 . So, by Proposition
23, |φpx10 q ´ φpx11 q| ă 2´n0 .

This proposition has the (perhaps surprising) consequence that φ is sur-

jective on r0, 1s

Corollary 10. ranpφq “ r0, 1s.

Proof. Since φp0q “ 0 and φp1q “ 1 and φ continuous.

The Cantor-Lebesgue function is continuous and increasing, but it will

be helpful to modify it so that it is strictly increasing. To do so, we define

ψpxq “ x ` φpxq.

Observation 2. The functionψ is an increasing and continuous map of

r0, 1s onto r0, 2s.

Observation 3. The inverse function ψ ´1 is continuous.

One somewhat surprising fact is that even though the Cantor set has
zero measure, its image under Ψ has positive measure.

Proposition 26. µpψrCsq “ 1.

Proof. Since ψ is one-to-one, r0, 2s “ ψrCs Y ψrr0, 1s ´ Cs. If pa, bq Ď r0, 1s ´

C, then µrψrpa, bqss “ µppa, bqq. Thus, µpψrr0, 1s ´ Csq “ 1. Therefore,
µpψrCsq “ 1.

Proposition 27. There is a measurable A Ď r0, 1s so that ψrAs is non-

measurable.

Proof. Since ψrCs has positive measure, it includes a non-measurable set B.

Set A “ ψ ´1 rBs. Then, A Ď C. So, A is measurable.

Corollary 11. There is a non-measurable function.

Proof. Suppose A is a measurable subset of r0, 1s so that ψrAs is non-

measurable. Let h “ χA ˝ ψ ´1 . Then, h´1 rp 21 , 23 qs “ ψrAs. Thus, h is
non-measurable.
3.2. THE CANTOR FUNCTION AND NON-MEASURABLE FUNCTIONS49

3.2.1 A remark for those interested in probability

The purpose of this subsection is to make some remarks on the Cantor-
Lebesgue function and its relation to probability. The reader is invited to
skip this section if they would like.
When one learns probability, one encounters many examples of proba-
bility measures (i.e., measures whose total mass is one). These generally fall
into two types.

1. The first type are those which are discrete and have probability mass
functions, such as the binomial distribution or negative binomial dis-
tribution. Such measures are called atomic, since they charge (i.e.,
give positive measure to) points.

2. The second type are measures such as normal distribution, which has
a probability density function
1 1 2
f pxq “ ? e´ 2 pxq
2π
and where the measure of a subset U can be computed as
ż
f pxqdx
U

(where this integral is done with respect to the Lebesgue measure).

It is tempting to intuitively think of measures as being some sort of lin-

ear combination of a continuous measure and an atomic measure. We will
not cover it in this class, but if you continue learning real analysis you will
encounter the Radon-Nikodym derivative (of which f pxq is an example) and
the Lebesgue decomposition theorem, which support this intuition. Intu-
itively, these results state that any reasonable1 measure can be decomposed
into a measure which has a probability density function (with respect to the
Lebesgue measure) and a measure which is singular (i.e., charges null sets).
However, it is important to realize that singular measures need not be
atomic. In particular, the Cantor-Lebesgue function can be understood
as a cumulative distribution function for a probability measure which is
supported on the Cantor set. Although this measure is Lebesgue singular
(since the Cantor set has measure zero), the fact that the Cantor-Lebesgue
1
There is a technical assumption here that the measures be σ-finite, but we will not
discuss this.
50 CHAPTER 3. MEASURABLE FUNCTIONS

function is continuous means that the measure is non-atomic. As such,

this is an important counter-example to keep in mind if you are studying
probability theory.
And now back to our regularly scheduled programming. . .

3.3 Approximating measurable functions

For Riemann integration, the general idea is to use step functions to ap-
proximate the desired function and then compute the area underneath the
step function. For the Lebesgue integral, we will instead approximate us-
ing simple functions. Intuitively, these functions cut horizontally instead of
vertically.
Definition 30. A simple function is a real-valued measurable function whose
range is finite.
Example. 1. The Heaviside function
#
1 xą0
f pxq “
0 xă0

2. The Dirichlet function

#
1 xPQ
f pxq “
0 xRQ

Definition 31. Suppose E Ď R. For each x P R let

"
1 xPE
χE pxq “
0 xRE

χE is called the characteristic function of E.

If E is measurable, then χE is a simple function.

Proof. Suppose U Ď r´8, 8s is open.

$
’
’ R 0, 1 P U
E c 0 P U and 1 R U
&
´1
χE rU s “
’
’ E 0 R U and 1 P U
%
H 0, 1 R U
3.3. APPROXIMATING MEASURABLE FUNCTIONS 51

It is not hard to show that every step function is also a simple function
Exercise 12. Show that every step function is a simple function.
There is an alternative characterization of simple function and the linear
combination of characteristic functions.
Proposition 28. A function φ : R Ñ R is simple if and only if there are
measurable sets E1 , . . . , En Ď R and c1 , . . . , cn P R so that
n
ÿ
φ“ cj χEj
j“1

Here is a quick sketch of the proof. For the forward direction, suppose
that φ simple. Let tc1 , . . . , cn u “ ranpφq. Take Ej “ φ´1 rtcj us. For the
reverse direction, note that the range of φ is finite. Since χEj is measurable,
by Theorem 19, we find that φ is measurable.
Theorem 29 (Simple Approximation Lemma). Suppose f : E Ñ R is a
bounded measurable function. Then, for every  ą 0, there exist φ, ψ so that
φ and ψ are simple functions on E, , φ ď f ď ψ and ψ ´ φ ă .
Proof. Since f is bounded, there is a positive number M so that |f pxq| ă M
for all x P E. Choose n0 P N so that 2M 2´n0 ă . Let cj “ ´M `
jp2M q2´n0 , and let dj “ ´M ` Ť pj ` 1qp2M q2´n0 . Set Vj “ rcj , dj q when
0 ď j ă 2n0 . Thus, r´M, M q “ jă2n0 Vj . Note ranpf q Ď r´M, M q. Set
Ej “ f ´1 rVj s. Therefore, Ej is measurable. Set:
ÿ
φ1 “ cj χEj
jă2n0
ÿ
ψ1 “ dj χEj
jă2n0

By Proposition 28, φ1 and ψ1 are simple. Let φ “ φ1 |E , and let ψ “ ψ1 |E .

Thus, φ and ψ are simple and φ ď f ď ψ by construction. Let x P E. Then,
x P Ej for exactly one j. By construction, cj “ φpxq and dj “ ψpxq. But,
dj ´ cj “ 2M 2´n0 ă .

In fact, we can prove a stronger approximation theorem as well.

Theorem 30. Suppose f : E Ñ r´8, 8s and suppose E is measurable.
Then, f is measurable. if and only if there is a sequence of simple functions
on E tσn u8
n“0 so that σn Ñ f a.e. and |σn | ď |f | for all n. If f ě 0, then
we can choose tσn u8
n“0 to be non-decreasing.
52 CHAPTER 3. MEASURABLE FUNCTIONS

Figure 3.2: The simple functions are converging almost everywhere to the
function f : E Ñ R (depicted in purple). As a disclaimer, these are not the
precise functions that are constructed in the proofs of the simple approxi-
mation lemmas.

Proof. The reverse direction follows directly from Theorem 21, so it remains
to show the forward direction.
Case 1: f ě 0.

Set En “ f ´1 rr0, n ` 1qs for every n P N. Set fn “ f ¨ χEn . Thus, fn is

measurable and bounded. By Theorem 29, there is a simple function φn so
that φn ď fn and fn ´ φn ă 2´n . Let
" `
φn pxq x P En
τn pxq “
n ` 1 x R En
Thus, τn is simple, and 0 ď τn ď f .

Claim: tτn u8
n“0 converges pointwise to f .

Let x P E. Let  ą 0. Choose N0 P N so that 2´N0 ă  and f pxq ă N0 .

Let n ě N0 . Therefore, x P En . Thus, τn pxq “ φ`
n pxq.

Subcase a: φn pxq ą 0.
3.4. LITTLEWOOD’S THREE PRINCIPLES 53

Then, τn pxq “ φn pxq and |f pxq ´ τn pxq| “ f pxq ´ φn pxq ă .

Subcase b φn pxq ď 0:

|f pxq ´ τn pxq| “ f pxq ď fn pxq ´ φn pxq ă .

Thus, limnÑ8 τn pxq “ f pxq.

Now, set σn “ maxjďn τj . Thus, σn is measurable. σn ď σn`1 by defini-

tion. τj ď f . Thus, 0 ď σn ď f . Since tτn u8
n“0 converges pointwise to f , it
8
follows that tσn un“0 converges pointwise to f .

Case 2: f ğ 0.

By Case 1, there exist a sequence of simple functions tσ1,n u8 n“0 that

` `
converges pointwise to f and so that 0 ď σ1,n ď f . By Case 1, there is a
sequence of simple functions tσ2,n u8
n“0 that converges pointwise to f
´ and

so that 0 ď σ2,n ď f ´ . Set σn “ σ1,n ´ σ2,n . Since f “ f ` ´ f ´ , tσn u8

n“0
converges to f pointwise.

3.4 Littlewood’s three principles

The extent of knowledge required is nothing like so great as is
sometimes supposed. There are three principles, roughly ex-
pressible in the following terms: Every [measurable] set is nearly
a finite union of intervals; every [measurable] function is nearly
continuous; every pointwise convergent sequence of [measurable]
functions is nearly uniformly convergent. Most of the results of
[the theory] are fairly intuitive applications of these ideas, and
the student armed with them should be equal to most occasions
when real variable theory is called for. If one of the principles
would be the obvious means to settle the problem if it were ‘quite’
true, it is natural to ask if the ‘nearly’ is near enough, and for a
problem that is actually solvable it generally is.

J.E. Littlewood

At first, measure theory might seem to be an imposing and very tech-

nical subject. However, Littlewood’s quote shows that measure theory only
54 CHAPTER 3. MEASURABLE FUNCTIONS

improves upon a naive approach to integration by  (albeit a very important

).
We have already seen Littlewood’s first principle precisely. The Vitali
covering lemma (Theorem 7) shows that given a measurable set E of finite
measure, then for each  ą 0, there is a finite collection of open intervals
whose union U is ”nearly equal to” E in the sense that
µpE ´ U q ` µpU ´ Eq ă .
In this section, we will work towards making the other principles precise.

3.4.1 The third principle: Egoroff ’s theorem

Here, we will show that pointwise convergence of measurable functions is
“nearly” uniform convergence. To do so, we start with the following lemma.
Lemma 10. Suppose µpEq ă 8, and suppose tfn u8 n“0 is a sequence of
measurable functions on E that converges pointwise to a real-valued function
f . Then, for all 1 , 2 ą 0 there exists N0 P N and a measurable A Ď E so
that µpE ´ Aq ă 2 and so that |f pxq ´ fk pxq| ă 1 whenever k ě N0 and
x P A.
Proof. Let 1 , 2 ą 0 and for each N0 P N, let
8
č
EN0 “ tx P E : |fn pxq ´ f pxq| ă 1 u
n“N0

Since fn Ñ f , ď
E“ EN0
N0 PN
Since EN0 Ď EN0 `1 ,
lim µpEN0 q “ µpEq.
N0 Ñ8

So, since µpEq ă 8, there is a non-negative integer N0 so that µpEq ´

µpEN0 q ă 2 . Set A “ EN0 .

With this lemma in hand, we can now prove a somewhat surprising theo-
rem, that pointwise convergence is “nearly” the same as uniform convergence
for measurable functions.
Theorem 31. (Egoroff ’s Theorem) Suppose µpEq ă 8 and tfn u8 n“0 is a
sequence of measurable functions on E that converges pointwise to a real-
valued function f . Then, for every  ą 0 there is a closed set F Ď E so that
tfn u8
n“0 converges uniformly to f on F and µpE ´ F q ă .
3.4. LITTLEWOOD’S THREE PRINCIPLES 55

Proof. Let  ą 0.

Claim 1: There is a measurable set A Ď E so that µpE ´ Aq ă {2 and

tfn u8
n“0 converges uniformly to f on A.

Proof of Claim 1: By Lemma 10, for each n P N, there is a non-negative

integer Mn and a measurable set An Ď E so that µpE ´ An q ăŞ2´pn`2q and
|f pxq ´ fk pxq| ă 2´n whenever x P An and k ě Mn . Set A “ 8 n“0 An .

Subclaim 1.a: µpE ´ Aq ă {2.

Proof:
8
ď 8
ÿ
µpE ´ Aq “ µp E ´ An q ă  2´pn`2q “ {2.
n“0 n“0

Subclaim 1.b: tfn u8

n“0 converges uniformly to f on E ´ A.

Proof: Let 1 ą 0. Choose n P N so that 2´n ă 1 . Let x P A, and let

k ě n. Since x P An , |f pxq ´ fk pxq| ă 2´n ă 1 .

Claim 2: There is a closed set F Ď E so that tfn u8

n“0 converges uni-
formly on F and µpE ´ F q ă .

Proof: By Theorem 6, there is a closed F Ď A so that µpA ´ F q ă {2.

So µpE ´ F q “ µpE ´ Aq Y µpA ´ F q ă .

3.4.2 The second principle: Lusin’s theorem

We can now make the second principle precise as well, to show that mea-
surable functions are “nearly” continuous functions. To establish this, we
start with a simpler version.

Lemma 11. Let f be a simple function defined on E. Then for each  ą 0,

there is a continuous function g : R Ñ R and a closed set F Ď E for which

f “ g on F and mpE ´ F q ă 

Proof. Let a1 , a2 , . . . , an be the distinct values taken by f , and let them

be taken on the sets E1 , E2 , . . . , En , respectively. The collection tEk unk“1 is
56 CHAPTER 3. MEASURABLE FUNCTIONS

disjoint since the ak ’s are distinct. According to Theorem 4, we may choose

closed sets F1 , F2 , . . . , Fn such that for each index k, 1 ď k ď n

Fk Ď Ek and m pEk ´ Fk q ă {n

Then we consider the set F “ nk“1 Fk , which must be closed (as the finite
Ť
union of closed sets). Since tEk unk“1 is disjoint,
˜ ¸
ďn n
ÿ
µpE ´ F q “ µ rEk ´ Fk s “ m pEk ´ Fk q ă 
k“1 k“1

Define g to be a function on F which takes the value ak on Fk for

1 ď k ď n. Since the collection tFk unk“1 is disjoint, g is properly defined.
Moreover, g is continuous on F since for a point x P FiŤ , there is an open
interval containing x which is disjoint from the closed set k‰i Fk and hence
on the intersection of this interval with F the function g is constant. But
g can be extended2 from a continuous function on the closed set F to a
continuous function on all of R. The continuous function g on R has the
required approximation properties.

Using this and Egoroff’s theorem, we can prove the full version of Lusin’s
Theorem.

Theorem 32 (Lusin’s Theorem). Suppose f : E Ñ R is measurable. Then,

for every  ą 0 there is a closed set F and a continuous function g : R Ñ R
so that F Ď E, µpE ´ F q ă , and gptq “ f ptq for all t P F .

Proof. We consider only provide details for the case that mpEq ă 8. Ac-
cording to the Simple Approximation Theorem, there is a sequence tfn u of
simple functions defined on E that converges to f pointwise on E. Let n be
a natural number. By the preceding proposition, with f replaced by fn and
 replaced by {2n`1 , we may choose a continuous function gn on R and a
closed set Fn Ď E for which

fn “ gn on Fn and µ pE ´ Fn q ă .
2n`1
According to Egoroff’s Theorem, there is a closed set F0 contained in E
such Ş
that tfn u converges to f uniformly on F0 and µ pE ´ F0 q ă {2. Define
F “ 8 n“0 Fn . Thus, we have that
2
This is known as the Tietze Extension Theorem. A proof can be found in your favorite
topology book (for instance, see [Mun14] Chapter 4 Section 35).
3.4. LITTLEWOOD’S THREE PRINCIPLES 57

˜ ¸
8 8
ď  ÿ 
µpE ´ F q “ m rE ´ F0 s Y rE ´ Fn s ď ` n`1
“ .
n“1
2 n“1 2

The set F is closed since it is the intersection of closed sets. Each fn is

continuous on F since F Ď Fn and fn “ gn on Fn . Finally, tfn u converges to
f uniformly on F since F Ď F0 . However, the uniform limit of continuous
functions is continuous, so the restriction of f to F is continuous on F .
Finally, there is a continuous function g defined on all of R whose restriction
to F equals f (again by the Tietze Extension Theorem). This function g
has the required approximation properties.

Exercise 13. Prove Lusin’s theorem in the case that µpEq “ 8.

58 CHAPTER 3. MEASURABLE FUNCTIONS
Chapter 4

Lebesgue Integration

After almost 60 pages of lecture notes (and over 2000 lines of LATEXcode),
we can finally start discussing how to integrate functions.
We start by defining the Lebesgue integral for non-negative simple func-
tions, which will be the fundamental building block for the Lebesgue integral.
Definition 32. If s : X Ñ r0, 8q is a simple function, and if E Ď X is
measurable, then the Lebesgue integral of s over E is defined to be
ż ÿ ` ˘
s dµ “ a µ s´1 ptauq X E .
E aPrangepsq

We denote this by ż
s dµ.
E

59
60 CHAPTER 4. LEBESGUE INTEGRATION

Note that in Definition 32, we use the convention 0 ¨ 8 “ 0 (since the

measure of the set where s vanishes will often be infinite).
We can also consider integrals over all of X, which is defined as
ż ÿ
sdµ “ aµps´1 rtausq
X aPranpsq

if s : X Ñ r0, 8q is simple.
Example. There are a few examples of integrals that we can compute im-
mediately.
ş
1. If E is a measurable set, then R χE dµ “ µpEq. Proof: let s “ χE .
Then, ranpsq “ t0, 1u. s´1 rt0us “ E c and s´1 rt1us “ E. So,
ż
s dµ “ 0 µpE c q ` 1 µpEq “ µpEq.
R
ş
2. If f is the Dirichlét function, then Rf dµ “ 0.
ş şb
3. If s is a non-negative step function on ra, bs, then ra,bs s dµ “ a s dx.
Proof: HW exercise.
Let us now establish some basic properties of the Lebesgue integral (for
simple functions).
Proposition 33. Suppose s : X Ñ r0, 8q is simple and that E Ď X is
measurable. Then, ż ż
s dµ “ sχE dµ.
E X
Proof. Let t “ sχE . Without loss of generality, we can take E ‰ X. Then,
ranptq “ srEs Y t0u. Thus, ranptq ´ ranpsq Ď t0u.
When a ‰ 0, tpxq “ a if and only if x P E and spxq “ a; i.e. x P
´1
s rtaus X E. Thus,
ż ÿ
t dµ “ aµpt´1 rtaus X Eq
X aPranptq
ÿ
“ aµpt´1 rtaus X Eq
aPranptq´t0u
ÿ
“ aµps´1 rtaus X Eq
aPsrEs
ÿ
“ aµps´1 rtaus X Eq
aPranpsq
 61

Lemma 12. Suppose s : X Ñ r0, 8q is a simple function. Suppose E1 , E2

are measurable sets so that E1 , E2 Ď X and so that E1 X E2 “ H. Then,
ż ż ż
sµ “ s dµ ` s dµ.
E1 YE2 E1 E2

Proof.
ż ÿ
s dµ “ aµps´1 rtaus X pE1 Y E2 qq
E1 YE2 aPranpsq
ÿ
“ apµps´1 rtaus X E1 q ` µps´1 rtaus X E2 qq
aPranpsq
ż ż
“ s dµ ` s dµ.
E1 E2

We now show that the Lebesgue integral is linear, which is one of its
most important properties.

Theorem 34. Suppose s1 , s2 : X Ñ r0, 8q are simple functions.

ş ş ş
1. X ps1 ` s2 q dµ “ X s1 dµ ` X s2 dµ.
ş ş
2. If c P R, then X cs1 dµ “ c X s1 dµ. 1

Proof. Part a:
pjq
For each a P ranpsj q, let Ea “ s´1
j rtaus. For each a P ranps1 q and b P
p1q p2q
ranps2 q, let Ea,b “ Ea X Eb . Note that:
p1q Ť
1. For each a P ranps1 q, Ea “ bPranps2 q Ea,b .

p2q Ť
2. For each b P ranps2 q, Eb “ aPranps1 q Ea,b .
Ť
3. X “ a,b Ea,b , and

4. Ea,b X Ea1 ,b1 “ H if pa, bq ‰ pa1 , b1 q.

1
If one wants to be pedantic here, they should require that c ą 0. For the purposes of
simplifying the proof of the next theorem, I will leave off this assumption.
62 CHAPTER 4. LEBESGUE INTEGRATION

By Lemma 12,
ż ÿż
ps ` tq dµ “ ps ` tq dµ.
X a,b Ea,b

On the other hand,

ż ÿ
ps1 ` s2 q dµ “ µpps1 ` s2 q´1 rtcus X Ea,b q
Ea,b cPranps`tq

“ pa ` bqµpps1 ` s2 q´1 rta ` bus X Ea,b q

“ pa ` bqµpEa,b q

So,
ż ÿ
ps1 ` s2 q dµ “ pa ` bqµpEa,b q
E a,b
¨ ˛ ¨ ˛
ÿ ÿ ÿ ÿ
“ ˝ aµpEa,b q‚` ˝ bµpEa,b q‚
aPranps1 q bPranps2 q bPranps2 q aPranps1 q
ÿ ÿ p2q
“ aµpEap1q q ` bµpEb q
aPranps1 q bPranps2 q
ż ż
“ s1 dµ ` s2 µ.
E E

Part b:
By definition,
ż ÿ
cs dµ “ caµpEa q
E caPranpcs1 q
ÿ
“ c aµpEa q
aPranps1 q
ż
“ c s dµ.
E

We now show that the integral is monotonic, which is another of its

essential features.
ş ş
Theorem 35. If s1 ď s2 , then X s1 dµ ď X s2 dµ.
4.1. THE LEBESGUE INTEGRAL FOR NON-NEGATIVE FUNCTIONS63

Proof. Consider the function t “ s2 ´ s1 . This is positive and simple func-

tion, so we have that
ż ż ż
s2 dµ ´ s2 dµ “ t dµ
X X X
ÿ
“ aµpEa q ě 0.
aPranptq

Exercise 14. Rearrange the proof of Theorem 35 to not require the differ-
ence of integrals (which we have technically not defined yet).

4.1 The Lebesgue integral for non-negative func-

tions
We can now define the Lebesgue integral for non-negative functions.

Definition 33. If f : X Ď r0, 8s is a measurable function, and if E is a

measurable subset of X, then the Lebesgue integral of f over E is
"ż *
sup s dµ : 0 ď s ď f and s is simple .
E

We denote this by ż
f dµ.
E

Note that if s : X Ñ r0, 8q is simple, and if E Ď X is measurable, then

by Theorem 34, Definitions 32 and 33 yield the same value.
This shows that the definition of the Lebesgue integral is consistent. We
can now state two principles for the Lebesgue integral, which are simple
consequences of the definition but are also quite useful.

Observation 4. Suppose f : X Ñ r0, 8s is measurable, and suppose E Ď X

is measurable.
ş
1. şIf s : X Ñ R is a simple function so that 0 ď s ď f , then E s dµ ď
E f dµ.
ş ş
2. Suppose α P r0, 8s. E f dµ ď α if and only if E sdµ ď α for every
simple s : X Ñ R so that 0 ď s ď f .
64 CHAPTER 4. LEBESGUE INTEGRATION

We now establish two propositions about the Lebesgue integral. First

we show that the Lebesgue integral of a function on a subset is the same as
the integral of the product of the function which the characteristic function
Proposition 36. Suppose
ş f :ş X Ñ r0, 8s is measurable and that E Ď X is
measurable. Then, E f dµ “ X f χE dµ.
ş ş
Proof. To do this, we şfirst show that E f dµ ď X f ¨ χE dµ.
For this, set α “ X f ¨ χE dµ and suppose s : X Ñ r0, 8s is a simple
function so that s ď f .
Then,
s ¨ χE ď f ¨ χE .
The function sχE is simple so we have that
ż ż
s ¨ χE dµ ď f ¨ χE dµ
X X
ş ş
by Definition 33. By Proposition 33, E s dµ “ X s ¨ χE dµ. So,
ż ż
s dµ ď f χE dµ
E X

and thus ż ż
f dµ ď f χE dµ.
E X

We now show the opposite inequality.

ż ż
f ¨ χE dµ ď f dµ.
X E

Suppose s : X Ñ r0, 8q is a simple function so that s ď f ¨ χE . Since f

is non-negative, s ď f . Thus,
ż ż
s dµ ď f dµ.
E E

But, ż ż
s dµ “ s dµ.
E X
So, ż ż
s dµ ď f dµ.
X E
4.1. THE LEBESGUE INTEGRAL FOR NON-NEGATIVE FUNCTIONS65

Proposition
ş ş 37. Suppose f, g : X Ñ r0, 8s are measurable. If f ď g then
X f µ ď X gdµ.

Proof. Suppose f ď g. If s is a simple function so that 0 ď s ď f , then

sďg .
So,
"ż *
s dµ : 0 ď s ď f and s is simple
X
"ż *
Ď sdµ : 0 ď s ď g and s is simple
X
ş ş
Thus, E f dµ ď E gdµ.

We now establish one of the most important results about the Lebesgue
integral, the monotone convergence theorem. This is a central and very
useful result which shows that we can interchange limits and integration for
increasing sequences of functions.

Theorem 38. (Monotone Convergence Theorem): Suppose f, f0 , f1 , . . . :

X Ñ r0, 8s are measurable and that fn ď fn`1 for all n P N. Suppose
f ptq “ limnÑ8 fn ptq for all t P X. Then,
ż ż
f dµ “ lim fn dµ.
X nÑ8 X

Proof. We
ş have that fn ď şf for all n P N since fn ď fn`1 . Let α “
limnÑ8 X fn dµ. Then, α ď X f dµ.
ş
Claim 1: X sdµ ď α whenever s : X Ñ R is a simple function so that
0 ď s ă f.

Proof Claim 1: Suppose s : X Ñ R is a simple function so that 0 ď s ă f .

For each n P N, let

Xn “ tx P X : spxq ď fn pxqu.
Ť
So, X “ n Xn (s ă f ). Also, Xn Ď Xn`1 . On the other hand,
ż ż ż
s dµ ď fn dµ ď fn dµ ď α.
Xn Xn X
66 CHAPTER 4. LEBESGUE INTEGRATION

and
ż ÿ
lim s dµ “ lim aµps´1 rtaus X Xn q
nÑ8 X nÑ8
n aPranpsq
ÿ ż
´1
“ aµps rtaus X Xq “ s dµ.
aPranpsq X

ş
SO, X s dµ ď α.
ş
Claim 2: X s dµ ď α whenever s : X Ñ R is a simple function so that
0 ď s ď f.

Proof Claim 2: Suppose s : X Ñ R is a simple function so thatş 0 ď s ď f.

Let sn “ p1 ´ 2´n qs for
ş each n P N. Thus, 0 ď s n ă f . Hence,
ş X sn dµ ď α.
´n
Therefore, p1 ´ 2 q X sn dµ ď α for each n P N. Thus, X s dµ ď α.
ş
Thus, X f dµ ď α.

Using the monotone convergence theorem, we can now show that the
Lebesgue integral is linear (we had previously only shown this for simple
functions).
Theorem 39. Suppose f, g : X Ñ r0, 8s are measurable.
ş ş ş
1. X pf ` gq dµ “ X f dµ ` X g dµ.
ş ş
2. If c ě 0, then X cf dµ “ c X f dµ.
Proof. By the proof of Theorem 30, there exist non-decreasing sequences
of non-negative simple functions tsn u8 8
n“0 and ttn un“0 so that sn Ñ f and
tn Ñ g. Apply MCT.

This has the following corollary, which is essentially a restatement of the

Monotone convergence theorem.
Corollary
ř 12. Suppose f0 , f1 , . . . : X Ñ r0, 8s are measurable and that
f pxq “ 8
n“0 fn pxq for all x P X. Then,
ż 8 ż
ÿ
f dµ “ fn dµ.
X n“0 X
řk
Proof. Set gk “ n“0 fn . Therefore, 0 ď gk ď gk`1 and gk Ñ f . Apply
MCT.
4.1. THE LEBESGUE INTEGRAL FOR NON-NEGATIVE FUNCTIONS67

In general, pointwise limits of functions need not preserve the Lebesgue

integral. However, there is an inequality which always holds, which is known
as Fatou’s lemma.
Lemma 13 (Fatou’s lemma). Suppose f0 , f1 , . . . : X Ñ r0, 8s are measur-
able and that f pxq “ lim inf n fn pxq for all x P X. Then,
ż ż
f dµ ď lim inf fn dµ
X n X

Proof. Set gk “ inf něk fn . Thus, gk ď f , f “ limk gk , and gk ď gk`1 and

gk ď f . Thus,
ż ż
f dµ “ lim gk dµ
X k Xż

“ lim inf gk dµ
k
żX
ď lim inf fk dµ.
k X

To see that Fatou’s inequality need not be an equality, consider the

following example.
Example 1. Consider the sequence of functions
1
fn “ χ .
n r0,ns
We have that fn Ñ f ” 0 uniformly. However, for all n
ż
1
fn dµ “ ¨ pn ´ 0q “ 1,
R n
whereas ż
f dµ “ 0.
R

Before moving on to the Lebesgue integral of functions which can be both

positive or negative, let us note one very important inequality in probability,
known as Chebychev’s inequality.
Theorem 40. Suppose f : X Ñ r0, 8s is measurable. Then, for all λ ą 0,
ż
1
µptx P X : f pxq ě λuq ď f dµ
λ X
68 CHAPTER 4. LEBESGUE INTEGRATION

ş
Proof. We can assume that X f dµ ă 8, or else the theorem is trivial.
Set Eλ “ tx P X : f pxq ě λu. ş ş
ş h “ λχEλ . Thus, 0 ď h ď f . Therefore, X hdµ ď X f dµ. But,
Set
X hdµ “ λµpEλ q.

4.2 The Lebesgue integral of real-valued functions

To define the Lebesgue integral of a real-valued function, we restrict our-
selves to functions whose absolute value is integrable. This might seem
strange initially, but this idea should be somewhat familiar because it is
exactly the distinction between convergent series and absolute convergent
series.

Definition 34 (Integrable functions). Suppose f : X Ñ r´8, 8s is mea-

surable. Suppose E is a measurable subset of X.
ş
1. f is integrable over E if E |f |dµ ă 8.

2. f is integrable if it is integrable over X.

For these functions, we can just define the Lebesgue integral of the dif-
ference between the integrals of the positive and negative parts.

Definition 35 (Lebesgue integral). Suppose f : X Ñ r´8, 8s is integrable

and E Ď X is measurable. The Lebesgue integral of f over E is defined to
be ż ż
f ` dµ ´ f ´ dµ.
E E
ş
This is denoted by E f dµ.

Before moving on, we make two brief remarks.

ş
Observation
ş ` 5. ş ´ X |f | dµ ă 8, and if E Ď X is measurable, then
1. If
E f dµ and E f dµ are finite.

2. If f is non-negative
ş and integrable, then Definitions 35 and 33 yield
same value for E f dµ.

Proposition 41. Suppose f : X Ñ r´8, 8s is integrable and E Ď X is

measurable. Then, ż ż
f dµ “ f ¨ χE dµ.
E X
4.2. THE LEBESGUE INTEGRAL OF REAL-VALUED FUNCTIONS69

Instead of giving a full proof, note that we already know this if f ě 0.

Otherwise, we set g “ f ¨ χE . Then we have that g ` “ f ` ¨ χE and
g ´ “ f ´ ¨ χE .
We can now show that the integral is monotonic and linear.

Theorem
ş ş 42. Suppose f, g : X Ñ r´8, 8s are integrable and f ď g. Then,
X f dµ ď X gdµ.

To see this, note that we have already established this if f ě 0. Other-

wise, note that f ` ď g ` and g ´ ď f ´ .

Theorem 43. Suppose f, g : X Ñ ş r´8, 8s are integrable

ş and that
ş α, β P R.
Then, αf ` βg is integrable and X pαf ` βgqdµ “ α X f dµ ` β X gdµ.

Proof. We start by noting that |αf ` βg| ď |α||f | ` |β||g|. Thus, αf ` βg is

integrable.
ş ş ş
Claim 1: X pf ` gq dµ “ X f dµ ` X g dµ.

Proof Claim 1: Set h “ f ` g. Then,

h` ´ h´ “ f ` ´ f ´ ` g ` ´ g ´

So,
h` ` f ´ ` g ´ “ f ` ` g ` ` h´ .
Thus,
ż ż ż ż ż ż
` ` ´ ` `
h dµ ` f dµ ` g dµ “ f dµ ` g dµ ` h´ dµ.
X X X X X X

Now rearrange terms.

ş ş
Claim 2: X αf dµ “ α X f dµ.

We will skip the proof of this claim.

Corollary 13. Suppose f : X Ñ r´8, 8s is integrable. Then,

ˇż ˇ ż
ˇ ˇ
ˇ f dµˇ ď |f |dµ.
ˇ ˇ
X X
70 CHAPTER 4. LEBESGUE INTEGRATION

Proof sketch: f, ´f ď |f |. Apply monotonicity and linearity. This result

is known in the book as the Integral Comparison test, where it is stated as
follows.
Corollary 14. Suppose f : X Ñ r´8, 8s is measurable and

|f | ă g

for some integrable function g. Then, f is integrable and

ˇż ˇ ż
ˇ ˇ
ˇ f dµˇ ď |f |dµ.
ˇ ˇ
X X

In any case, we can use this to show the following proposition

Proposition 44. Suppose f : X Ñ r´8, 8s is measurable
ş and either inte-
grable or non-negative. If E Ď X has measure 0, then E f dµ “ 0.

4.2.1 The dominated convergence theorem

We now prove one the most fundamental theorems in measure theory: the
dominated convergence theorem.
Theorem 45. Suppose f0 , f1 , . . . : X Ñ R are measurable and that f pxq “
limnÑ8 fn pxq for all x P X. Suppose there is an integrable g : X Ñ r´8, 8s
so that |fn pxq| ď gpxq for all x P X. Then, f is integrable,
ż
lim |fn ´ f |dµ “ 0, (4.1)
nÑ8 X

and ż ż
lim fn dµ “ f dµ (4.2)
nÑ8 X X

Proof. Since |f pxq| “ limnÑ8 |fn pxq| ď gpxq. Thus, f is integrable.

To establish Equation 4.1, we note that |fn ´ f | ď 2g and that

2g “ lim inf p2g ´ |fn ´ f |q.

n

By Fatou’s Lemma,
ż ż
2g dµ ď lim inf p2g ´ |fn ´ f |q dµ
X n X
ż ż
“ 2g dµ ´ lim sup |fn ´ f | dµ
X n X
4.2. THE LEBESGUE INTEGRAL OF REAL-VALUED FUNCTIONS71

So, ż
lim sup |fn ´ f | dµ ď 0.
n X
ş
Thus, limn X |fn ´ f | dµ “ 0.
Equation 4.2 now follows from Equation 4.1.

As a preliminary application, we can now give a disproportionately so-

phisticated proof that the harmonic series diverges.

Exercise 15. Using Example 1, show that the series

8
ÿ 1
n“1
n

diverges. Hint: Consider the function gpxq “ supn f pxq.

This argument was posted by Tao as an answer on MathOverflow, but

I actually find it helpful to remember the dominated convergence theorem
and Fatou’s inequality.

Theorem 46. Let f : E Ñ r´8, 8s be integrable. Suppose tEn u8 Ť is a

n“0
pairwise disjoint sequence of measurable subsets of E and set E “ n En .
Then,
ż 8 ż
ÿ
f dµ “ f dµ.
E n“0 En

Proof. Set
n
ď
Fn “ Ej
j“0
fn “ f ¨ χFn

Thus, limnÑ8 fn “ f . |fn | ď |f |. So, by DCT

ż ż
lim fn dµ “ f dµ
nÑ8 E E

But,
ż n ż
ÿ
fn dµ “ f dµ.
E k“0 Ek
72 CHAPTER 4. LEBESGUE INTEGRATION

This implies the following corollary.

Corollary 15. (Continuity of integration): Suppose f : X Ñ r´8, 8s is
integrable.
Ť
1. If E0 Ď E1 . . . Ď X are measurable, and if E “ n En , then
ż ż
f dµ “ lim f dµ.
E nÑ8 E
n

Ş
2. If X Ě E0 Ě E1 Ě . . ., and if E “ n En , then
ż ż
f dµ “ lim f dµ.
E nÑ8 E
n

4.2.2 Uniform integrability and the Vitali convergence the-

orem
We conclude our initial discussion of Lebesgue integration by finding a cri-
terion which allows us to interchange limits and integrals. Let us start by
making an observation which was actually an earlier homework exercise and
says that we can break up sets of finite measure into the disjoint union of
sets with small measure.
Lemma 14. Let E be a set of finite measure and δ ą 0. Then E is the
disjoint union of a finite collection of sets, each of which has measure less
than δ.
Proof. By the continuity of measure,

lim µpE „ r´n, nsq “ µpHq “ 0

nÑ8

Choose a natural number n0 for which µ pE „ r´n0 , n0 sq ă δ. By choosing

a fine enough partition of r´n0 , n0 s, express E X r´n0 , n0 s as the disjoint
union of a finite collection of sets, each of which has measure less than δ.

Proposition 47. Let f be a measurable function on E. If f is integrable

over E, then for each  ą 0, there is a δ ą 0 for which if A Ď E is measurable
and µpAq ă δ, then ż
|f | ă 
A
. Conversely, in the case µpEq ă 8, if for each  ą 0, there is aδ ą 0 for
which this inequality holds, then f is integrable over E.
4.2. THE LEBESGUE INTEGRAL OF REAL-VALUED FUNCTIONS73

Proof. The theorem follows by establishing it separately for the positive

and negative parts of f so WLOG we take f ě 0 on E. First, we assume
f is integrable over E and let  ą 0. By the definition of the integral of a
nonnegative integrable function, there is a measurable bounded function f
of finite support which satisfies
ż ż
0 ď f ď f on E and 0 ď f´ f ă {2
E E

Since f ´ f ě 0 on E, if A Ď E is measurable, then, by the linearity and

additivity over domains of the integral,
ż ż ż ż ż ż
f´ f “ rf ´ f s ď rf ´ f s “ f´ f ă {2.
A A A E E E

But f is bounded so we can find M ą 0 for which 0 ď f ă M on E0 .

Therefore, if A Ď E is measurable, then
ż ż
fă f ` {2 ď M ¨ µpAq ` {2 .
A A

We then take δ “ {2M . Then we have that if µpAq ă δ, then

ż
|f | ă .
A

Conversely, suppose şmpEq ă 8 and for each  ą 0, there is a δ ą 0 for

which µpAq ă δ implies A |f | ă .
ş
Let δ0 ą 0 be such that A |f | ă 1 whenever µpAq ď δ0 . Since µpEq ă 8,
the previous lemma shows that we can consider E as the disjoint union of a
finite collection of measurable subsets tEk uN
k“1 , each of which has measure
less than δ0 . Therefore
ÿN ż
f ăN
k“1 Ek

By the additivity over domains of integration it follows that if h is a non-

negative
ş measurable function of finite support and 0 ď h ď f on E, then
E h ă N. Therefore f is integrable.

With these results in mind, we give the following definition of integra-

bility for a sequence of functions.
74 CHAPTER 4. LEBESGUE INTEGRATION

Definition 36. Suppose f0 , f1 , . . . are measurable functions on X. tfn u8

n“0
is uniformly integrable if for every  ą 0 there is a δ ą 0 so that
ż
|fn |dµ ă 
A

whenever n P N and A is a subset of X whose measure is smaller than δ.

Theorem 48 (Vitali convergence theorem). Suppose µpXq ă 8 and tfn u8 n“0

is a uniformly integrable sequence of measurable functions on X that con-
verges pointwise almost everywhere to f . Then, f is integrable and
ż ż
lim fn dµ “ f dµ.
nÑ8 X X

We will not cover the proof in class. However, the proof is contained in
the book.

4.3 Convergence in measure

[This subsection is taken directly from Royden-Fitzpatrick with
minimal modifications.]
We have considered sequences of functions that converge uniformly, that
converge pointwise, and that converge pointwise almost everywhere. There
is one more mode of convergence that we will study in this class, known as
convergence in measure.

Definition 37. Let tfn u be a sequence of measurable functions and f a

measurable function (all defined on E). Suppose that f and fn are finite
a.e. The sequence tfn u is said to converge in measure on E to f provided
for each η ą 0,

lim µ tx P E | |fn pxq ´ f pxq| ą ηu “ 0

nÑ8

Note that if a sequence of (finite a.e.) measurable functions converges

uniformly, then it also converges in measure (Exercise: why?). However, we
can also say something stronger.

Proposition 49. Assume E has finite measure. Let tfn u be a sequence

of measurable functions on E that converges pointwise a.e. on E to f and
suppose that f is finite a.e. Then tfn u Ñ f in measure on E.
4.3. CONVERGENCE IN MEASURE 75

Proof. The function f is measurable since it is the pointwise limit almost

everywhere of a sequence of measurable functions. Let η ą 0. To prove
convergence in measure we let  ą 0 and seek an index N such that

µ tx P E|fn pxq ´ f pxq| ą ηu ă  for all n ě N.

Egoroff’s Theorem tells us that there is a measurable subset F of E with

µpE „ F q ă  where tfn u converges to f . As such, we can find N large
enough so that
|fn ´ f | ă η on F for all n ě N .
Thus, for n ě N, tx P E | | fn pxq ´ f pxq |ą ηu Ď E „ F , which implies
that the measure is less than .

This result fails if E has infinite measure, and the converse to this propo-
sition also is false. In other words, there are sequences which converge in
measure which do not converge pointwise. However, given a sequence of
functions which converge in measure, there is a subsequence which converges
pointwise almost everywhere.

Theorem 50. If tfn u Ñ f in measure on E, then there is a subsequence

tfnk u that converges pointwise a.e. on E to f .

Proof. By the definition of convergence in measure, there is a strictly in-

creasing sequence of natural numbers tnk u for which
1
µ tx P E | | fj pxq ´ f pxq |ą 1{ku ă for all j ě nk .
2k
For each index k, define

Ek “ tx P E | | fnk ´ f pxq |ą 1{ku

Since µ pEk q ă 21k , we have that 8

ř
k“1 m pEk q ă 8. The Borel-Cantelli
Lemma then implies that for almost all x P E, there is an index Kpxq such
that x R Ek if k ě Kpxq, that is,

|fnk pxq ´ f pxq| ď 1{k for all k ě Kpxq

Therefore
lim fnk pxq “ f pxq.
kÑ8
76 CHAPTER 4. LEBESGUE INTEGRATION

The following proposition shows that many of the theorems in this course
can be weakened to require convergence in measure rather than pointwise
everywhere convergence.

Proposition 51. Fatou’s Lemma, the Monotone Convergence Theorem,

the Lebesgue Dominated Convergence Theorem, and the Vitali Convergence
Theorem remain vahid if ”pointwise convergence a.e.” is replaced by ”con-
vergence in measure.”

The proof is left as an exercise.

4.4 The Riemann and the Lebesgue integral

At this point, we have defined the Lebesgue integral and discussed some of
its fundamental properties. However, it is natural to ask how this integral
relates to the Riemann integral, which is the focus of this section. To do so,
we start by proving some small lemmas

Lemma 15. Suppose f : ra, bs Ñ R is Riemann integrable. Then, there is

a non-decreasing sequence of step functions ts1,n u8
n“0 and a non-increasing
sequence of step functions ts2,n u8
n“0 so that for all n P N s1,n ď f ď s2,n
and

żb
ps2,n ´ s1,n q ă 2´n .
a

The proof of this is left as a home work exercise.

4.4. THE RIEMANN AND THE LEBESGUE INTEGRAL 77

We now prove a “sandwich” lemma for these integrals.

Lemma 16. Suppose tφn u8 n“0 is a non-decreasing sequence of integrable
function on X, and suppose tψn u8n“0 is a non-increasing sequence of inte-
grable functions on X. Suppose φn ď f ď ψn for all n P N and that
ż
lim pψn ´ φn q dµ “ 0.
nÑ8 X

Then:
1. φn Ñ f a.e. and ψn Ñ f a.e.
2. f is integrable
3. ż ż ż
lim φn dµ “ lim ψn dµ “ f dµ.
nÑ8 X nÑ8 X X

Proof. (a): Let φş “ limnÑ8 φn and ψ “ limnÑ8 ψn . Thus, 0 ď ψ ´ φ ď

ψn ´ φn . Thus, X pψ ´ φqdµ “ 0. Therefore, ψ “ φ a.e. Thus, φ “ f “ ψ
a.e. Thus, f is measurable.

(b): We have 0 ď f ´ φ0 ď ψ0 ´ φ0 . Thus, f ´ φ0 is integrable. Thus, f

is integrable.

(c): This follows from monotonicity.

78 CHAPTER 4. LEBESGUE INTEGRATION

With these two results in hand, we can now show that the Riemann in-
tegral is equal to the Lebesgue integral, whenever the former is well-defined.

Theorem 52. Suppose f : ra, bs Ñ R is Riemann integrable. Then, f is

Lebesgue integrable and the two integrals are equal. i.e.
żb ż
f“ f dµ.
a ra,bs

Proof. By Lemma 15, there is a non-increasing sequence

şb of step functions
ts2,n u8
n“0 so that for all n P N s1,n ď f ď s2,n and a psn,2 ´ sn,1 q ă 2´n . It
follows from Lemma 16 that f is integrable. Since
ż żb żb żb żb
sn,j dµ “ sn,j , and since sn,1 ď fď sn,2 ,
ra,bs a a a a

it follows that żb żb ż
f “ lim s1,n “ f dµ.
a nÑ8 a ra,bs

As such, all that remains is to determine necessary and sufficient condi-

tions of a function to be Riemann integrable. For this, we have the following
result.

Theorem 53. Suppose f : ra, bs Ñ R is bounded. Then, f is Riemann

integrable if and only if f is continuous a.e..

Proof. pñq: Suppose f is Riemann integrable. By Lemma 15, there is a

non-increasing sequence of step functions ts2,n u8
n“0 so that for all n P N
şb ´n
s1,n ď f ď s2,n and a psn,2 ´ sn,1 q ă 2 .
Let

Ej “ tx : Dn P N sj,n is discontinuous at xu j “ 1, 2
Ej`2 “ tx : lim sj,n pxq ‰ f pxqu j “ 1, 2
nÑ8
E “ E1 Y E2 Y E3 Y E4

Thus E1 Y E2 is countable. So µpEq “ 0.

Claim: If x0 P ra, bs ´ E, then f is continuous at x0 .

4.4. THE RIEMANN AND THE LEBESGUE INTEGRAL 79

Proof Claim: Suppose x0 P ra, bs ´ E. Let  ą 0. Choose N0 so that

s2,N0 pxq´s1,N0 pxq ă . There exist δ ą 0 so that s1,N0 and s2,N0 are constant
on px0 ´ δ, x0 ` δq. Therefore, if |x ´ x0 | ă δ, then

|f pxq ´ f px0 q| ď |s2,N0 pxq ´ s1,N0 px0 q| ă .

pðq: Suppose f is continuous a.e.. For each n, let Pn denote the uniform
pnq
partition of ra, bs of width pb ´ aq{2´n . Let aj denote the j-th point in Pn
pnq pnq pnq
where j “ 0, . . . , 2n . Set Ij “ paj , aj`1 s.
Let
pnq pnq
mj “ inftf px1 q : x1 P Ij u
pnq pnq
Mj “ suptf px1 q : x1 P Ij u

pnq
When x P Ij , set

pn0 q
φn pxq “ mj
pnq
ψn pxq “ Mj .

We then define the set

ď ! pnq )
E “ tx P ra, bs : f is discontinuous at xu aj : n P N and 0 ď j ď 2n .

Thus, µpEq “ 0.

Claim 1: If x0 P ra, bs ´ E, then limnÑ8 φn px0 q “ limnÑ8 ψn px0 q “

f px0 q.

Proof Claim 1: Suppose x0 P ra, bs ´ E. Let  ą 0. Choose δ ą 0 so that

|x ´ x0 | ă δ ñ |f pxq ´ f px0 q| ă .

Choose n0 P N so that pb ´ aq2´n0 ă δ. Let n ě n0 . There is a unique

pnq pnq pnq
j so that x P Ij . Therefore, Mj ´ mj ď pb ´ aq2´n ď 2. So,
|s2,n pxq ´ f pxq|, |s1,n pxq ´ f pxq| ď 2.
ş
Claim 2: limnÑ8 ra,bs pψn ´ φn qdµ “ 0.
80 CHAPTER 4. LEBESGUE INTEGRATION

Proof Claim 2: Since f is bounded, there is a positive number M so that

0 ď ψn ´ φn ă M for all n P N. Thus, by DCT
ż
lim pψn ´ φn q dµ “ 0.
nÑ8 ra,bs

Claim 3: f is Riemann integrable.

Proof Claim 3:
ż żb
ψn dµ “ ψn
ra,bs a
żb
ě f
a
ż b
ě f
a
ż b ż
ě φn “ φn dµ
a ra,bs

Therefore,
żb żb
f“ f.
a a

Although it is genuinely a stronger condition for a function to be Rie-

mann integrable versus Lebesgue integrable, it turns out that the building
blocks for each of these integrals are not so different. In fact, we have the
following proposition (proof taken from Mathstackexchange [wh])

Proposition 54. Step functions are dense inside the space of simple func-
tions.

Proof. It suffices to show that the characteristic of a measurable set of finite

measure can be approximated by step functions. Take a measurable set A
with µpAq ă `8; by regularity, we can find an open set U Ą A such that
µpU zAq ă .
U can be written as a countable union of disjoint intervals:
8
ď
U“ In
n“0
4.4. THE RIEMANN AND THE LEBESGUE INTEGRAL 81

so, we can find N such that

˜ ¸
ď
µ In ă.
nąN

Hence, we define
N
ÿ
hpxq “ χIn pxq
n“0

and we have ż
|hpxq ´ χA pxq} ď 2 .

So, the step functions are dense among the simple functions.
82 CHAPTER 4. LEBESGUE INTEGRATION
Chapter 5

Differentiation and
Integration

The fundamental theorems of calculus provide a foundational relationship

between the Riemann integral and the derivative. We now want to establish
versions of these theorems for the Lebesgue integral.

5.1 Monotone functions

A function is monotone if it is either increasing or decreasing. Such func-
tions play an essential role in the Lebesgue versions of the fundamental
theorems of calculus for two reasons.

1. A theorem of Lebesgue asserts that a monotone function on an open

interval is differentiable almost everywhere.

2. A theorem of Jordan shows that a very general family of functions

(i.e., those of bounded variation) may be expressed as the difference of
monotone functions and therefore they also are differentiable almost
everywhere on the interior of their domain.

Theorem 55. Every function that is monotone on an open interval is con-

tinuous except at a countable set of points.
Proof. Suppose I is an open interval, and suppose f is a monotone function
on I. Without loss of generality, f is non-decreasing. Set
" *
E “ a P I : lim f pxq ‰ lim f pxq .
xÑa´ xÑa`

83
84 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

For each a P E, let

`a “ lim f pxq
xÑa´
ra “ lim f pxq
xÑa`

Since f is non-decreasing, `a ă ra for all a P E.

Claim 1: if a, b P E, and if a ‰ b, then p`a , ra q X p`b , rb q “ H.

Proof Claim 1:

Suppose a, b P E and a ‰ b. WLOG a ă b. Since f is non-decreasing,

for all x0 , x1 P I,

x0 ă a and x1 ą b ñ f px0 q ď f px1 q.

Thus, ra ď `b .

Claim 2: E is countable.

For each a P E, choose a rational number qa P pra , `a q. By Claim 1,

a ÞÑ qa is one-to-one. Thus, E is countable.

We want to show that monotone functions are differentiable almost ev-

erywhere. For this, we will need to define some concepts related to deriva-
tives.

Definition 38. Suppose f : ra, bs Ñ R is integrable. When h ą 0 and

x P ra, bs, let
f px ` hq ´ f pxq
Diff h pf qpxq “ .
h
We call Diff h pf q a divided difference function.

Definition 39. Suppose f : A Ñ R and x0 belongs to the interior of A.

1. Dpf qpx0 q :“ limhÑ0` sup0ă|t|ďh Diff t pf qpx0 q. We call Dpf q the upper
derivative of f .

2. Dpf qpx0 q :“ limhÑ0` inf 0ă|t|ďh Diff t pf qpx0 q. We call Dpf q the lower
derivative of f .
5.1. MONOTONE FUNCTIONS 85

If Dpf qpx0 q “ Dpf qpx0 q, then f is differentiable at x0 and we set f 1 px0 q “

Dpf qpx0 q.

We can immediately see that Dpf qpx0 q ď Dpf qpx0 q.

Lemma 17. Suppose f : ra, bs Ñ R is non-decreasing. Then:

1. For all α ą 0,

1
µptx P pa, bq : Dpf qpxq ě αuq ă pf pbq ´ f paqq.
α

2. µptx P pa, bq : Dpf qpxq “ 8uq “ 0.

Proof. (1): Suppose α ą 0. Set

Eα “ tx P pa, bq : Dpf qpxq ě αu.

Set
F “ trc, ds : rc, ds Ď pa, bq and f pdq ´ f pcq ě αpd ´ cqu.
Claim 1: F is a Vitali covering (recall Definition 14) of Eα .

Proof Claim 1: Let x0 P Eα , and let  ą 0. Dpf qpx0 q ě α. So, there is a

δ ą 0 so that for all h P p0, δq, there is a t so that |t| ď h and

f px0 ` tq ´ f px0 q
ě α. (5.1)
t
Choose h ą 0 so that h ă δ, {2. and rx0 ´ h, x0 ` hs Ď pa, bq. Choose t so
that |t| ď h and (5.1). Set

c “ mintx0 , x0 ` tu
d “ maxtx0 , x0 ` tu

Since f non-decreasing, f pdq ´ dpcq ě αpd ´ cq. Thus, rc, ds P F, x0 P rc, ds,
and d ´ c ă .

Claim 2: µ˚ pEα q ď α1 pf pbq ´ f paqq.

86 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

Proof Claim 2: Let  ą 0. By the Vitali Covering Lemma, there exist

rc0 , d0 s, . . . , rcj , dj s P F so that prc0 , d0 s, . . . , rcn , dn sq is pairwise disjoint and
so that
ď n
˚
µ pEα ´ rcj , dj sq ă .
j“0

Therefore,
n
ÿ
µ˚ pEα q ď  ` pdj ´ cj q
j“0
n
ÿ
ď  ` α´1 pf pdj q ´ f pcj qq
j“0

ď  ` α´1 pf pbq ´ f paqq since f non-decreasing

Therefore, µ˚ pEα q ď α´1 pf pbq ´ f paqq.

(2): The second claim follows directly from the first one.

We can not show that monotone functions are differentiable almost ev-
erywhere.
Theorem 56. Suppose I is an open interval and f : I Ñ R is monotone.
Then, f is differentiable a.e..
Proof. Suppose I “ pa, bq where ´8 ă a ă b ă 8 and set

E “ tx P pa, bq : Dpf qpxq ą Dpf qpxqu.

For all α, β P Q with α ą β, let

Eα,β “ tx P pa, bq : Dpf qpxq ą α ą β ą Dpf qpxqu.

Ť
Therefore, E “ α,β Eα,β .

Claim 1: µ˚ pEα,β q “ 0.

Proof Claim 1: Let α, β P Q with β ą α and take  ą 0.

Let O be an open set so that µpOq ď µ˚ pEα,β q ` . We then se F be
the collection of closed, bounded intervals rc, ds contained in O for which
f pdq ´ f pcq ă βpd ´ cq.

F “ trc, ds : rc, ds Ď O and f pdq ´ f pcq ă βpd ´ cqu.

5.1. MONOTONE FUNCTIONS 87

As in the proof of the previous lemma (Lemma 17), F is a Vitali covering

of Eα,β (since x P Eα,β implies that Dpf qpxq ă β).
By the Vitali Covering Lemma, there exist rc0 , d0 s, . . . , rcn , dn s P F so
that prc0 , d0 s, . . . , rcn , dn sq is pairwise disjoint and
n
ď
µ˚ pEα,β ´ rcj , dj sq ă .
j“0

Using Lemma 17, we also have that

µ˚ pEα,β X rcj , dj sq ď α´1 pf pdj q ´ f pcj qq.

Combining both of these, we find that

n
ÿ
˚ ´1
µ pEα,β q ă  ` α pf pdj q ´ f pcj qq
j“0
n
ÿ
ă  ` α´1 β pdj ´ cj q
j“0

ď  ` α´1 βµpOq
ď  ` α´1 βpµ˚ pEα,β q ` q.

Since  was arbitrary, this shows that

µ˚ pEα,β q ď α´1 βµ˚ pEα,β q.

However, since α´1 β ă 1 and µ˚ pEα,β q ă 8,

µ˚ pEα,β q “ 0.

Our goal now is is prove the fundamental theorem of calculus for mono-
tonic functions. However, before we do so we need to provide two more
definitions and prove a small proposition.
Definition 40. Suppose f : ra, bs Ñ R is integrable.
1. When h ą 0 and x P ra, bs, let
f px ` hq ´ f pxq
Diff h pf qpxq “ .
h
We call Diff h pf q a divided difference function.
88 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

2. When h ą 0 and x P ra, bs, let

ż
1
Avh pf qpxq “ f dµ
h rx,x`hs

We call Avh pf q an average value function.

: In (a) and (b) we take f pb ` hq to be f pbq.

Proposition 57. Suppose f : ra, bs Ñ R is integrable. Then,

żv
Diff h pf qdµ “ Avh pvq ´ Avh pf q.
u

whenever a ď u ă v ď b and h ą 0.

Proof. żv „ż v żv 
1
Diff h pf qdµ “ f px ` hqdµpxq ´ f dµ .
u h u u

To see this, note that

żv ż v`h
f px ` hq dµpxq “ f dµ.
u u`h

Then note that

ż v`h żv ż v`h ż u`h
f dµ ´ f dµ “ f dµ ´ f dµ.
u`h u v u

To see this, note that there are two separate cases, where u ` h ă v and
where v ď u ` h. Combining these factos together, we find that
ż
Diff h pf qdµ “ Avh pf qpvq ´ Avh pf qpuq.

Finally, over 80 pages (and over 3000 lines of LATEX) in, we can (almost)
prove the fundamental theorem of calculus (for monotonic functions).

Proposition 58. Suppose f : ra, bs Ñ R is non-decreasing. Then, f 1 is

integrable and
żb
f 1 dµ ď f pbq ´ f paq.
a
5.1. MONOTONE FUNCTIONS 89

Proof. From our previous work, we immediately see the following.

1. f is measurable (applying Theorem 55 since f is monotonic).

2. f is differentiable a.e. (applying Theorem 56).

3. f 1 ě 0. Since f is non-decreasing.
şb 1
af dµ ď lim inf n rAv2´n pf qpbq ´ Av2´n pf qpaqs.

Proof Claim 1: We first note that

f 1 “ lim Diff 2´n pf q.

nÑ8

By Fatou’s Lemma and Proposition 57,

żb żb
1
f dµ ď lim inf Diff 2´n pf q dµ
a n a
“ lim inf rAv2´n pf qpbq ´ Av2´n pf qpaqs
n

Claim 2: Av2´n pf qpbq ´ Av2´n pf qpaq ď f pbq ´ f paq.

By definition
ż b`2´n
n
Av2´n pf qpbq “ 2 f dµ “ 2n f pbqpb ` 2´n ´ bq “ f pbq.
b

And, since f is non-decreasing,

ż a`2´n
n
Av2´n pf qpaq “ 2 f dµ ě 2n f paq2´n “ f paq.
a

Note that the inequality in this proposition can be strict. For instance,
if φ is the Cantor-Lebesgue function, then
ż1
φ1 dµ “ 0 ă φp1q ´ φp0q.
0
90 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

5.2 Functions of bounded variation

In order to take what we have done for monotonic functions and apply it
more generally, we first need to discuss the notion of variation for a function.
Definition 41. Suppose f : ra, bs Ñ R.
1. If P “ px0 , . . . , xn q is a partition of ra, bs, then we define
ÿ
V pf, P q “ |f pxj`1 q ´ f pxj q|.
jăn

V pf, P q is called the variation of f with respect to P .

2. The total variation of f is

T V pf q :“ sup V pf, P q.
P

3. f has bounded variation if T V pf q ă 8.

We can give some examples and non-examples.

Example. 1. If f is non-decreasing, then T V pf q “ f pbq ´ f paq.

2. Lipschitz functions have bounded variation. More precisely, suppose

there is a positive real M so that |f pxq ´ f pyq| ď M |x ´ y| whenever
x, y P ra, bs. Then, T V pf q ă 8.
Non-Example. When 0 ď x ď 1, let

x cosp πx q x ‰ 0
"
f pxq “
0 x“0
5.2. FUNCTIONS OF BOUNDED VARIATION 91

For each n P N, let

1 1 1 1
Pn “ p0, , , . . . , , , 1q.
n`1 n 3 2
Then,
n`1
ÿ 1
V pf, Pn q “ 1 ` 2 .
j“2
j

So, T V pf q “ 8.
The main result about functions with bounded variation that we will use
is that they can be written as the difference of two monotonic functions.

Theorem 59. f : ra, bs Ñ R has bounded variation if and only if it can be

expressed as a difference of non-decreasing functions.

We will skip the proof in class, but for completeness I will include the
proof in the notes.

Proof.
` To˘prove this, we start by defining the total variation function x ÞÑ
T V fra,xs , which is defined to be the total variation of f restricted to the
interval ra, xs. We can immediately see several properties of this function.

1. If a ă x ă b, then

T V pf q “ T V pf |ra,xs q ` T V pf |rx,bs q.

(You should convince yourself why this is the case. As a hint, try
refining all the partitions with x and see what happens.)

2. Rearranging this equation, we find that

` ˘ ` ˘ ` ˘
T V fra,vs ´ T V fra,us “ T V fru,vs ě 0 for all a ď u ă v ď b,

so the total variation function is increasing.

3. Furthermore, for a ď u ă v ď b, we have

` ˘ ` ˘ ` ˘ ` ˘
f puq´f pvq ď |f pvq´f puq| “ V fru,vs , P ď T V fru,vs “ T V fra,vs ´T V fra,us

which implies that ` ˘

f pxq ` T V fra,xs
is also an increasing function.
92 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

Thus, we can write

` ˘
f pxq “ f pxq ` T V pf |ra,xs q ´ T V pf |ra,xs q

as the difference of two increasing functions.

To prove the converse direction, it suffices to note that if f “ g ´ h for
non-decreasing functions g and h, then for any partition P of ra, bs

V pf, P q ď pgpbq ´ gpaqq ` phpbq ´ hpaqq,

which immediately implies that f has bounded total variation.

Corollary 16. Every function of bounded variation is differentiable almost

everywhere.

5.3 Absolutely continuous functions

At this point, we have proven the fundamental inequality of calculus for a
wide class of functions, which are those of bounded variation. However, we
have also seen that this inequality need not be an inequality and so it is
natural to ask the following question:
“When does the fundamental theorem of calculus hold?”
It turns out that in order to answer this question, we need to define the
notion of absolute continuity

Definition 42. Suppose f : ra, bs Ñ R. f is absolutely continuous if for

every  ą 0 there is a δ ą 0 so that
ÿ
|f pyj q ´ f pxj q| ă 
jďn

whenever px0 , y0 q, . . ., pxn , yn q are disjoint subintervals of ra, bs so that

ÿ
pyj ´ xj q ă δ.
jďn

In essence, absolute continuity is a stronger form of uniform continuity

where you also require the total change in a function to be small when you
take the union of many small intervals (whose total size is less than δ). The
main result of this section is the following.

Theorem 60. Suppose f : ra, bs Ñ R is absolutely continuous. Then:

5.3. ABSOLUTELY CONTINUOUS FUNCTIONS 93

1. f is differentiable a.e.

2. f 1 is integrable.
ş
3. ra,bs f 1 dµ “ f pbq ´ f paq.

However, before we prove this, let us first discuss absolute continuity in

a bit more detail.

Example. Every Lipschitz continuous function is absolutely continuous.

Non-Example. The Cantor-Lebesgue function is not absolutely continuous.

Proof. Consider all intervals of the form

p0.a0 a1 . . . an , 0.a0 . . . an 1q (base 3)

where a0 , . . . , an P t0, 2u.

The length of each such interval is 1{3n`2 and there are 2n`1 such in-
tervals, so the sum of their lengths is 2n`1 {3n`2 which Ñ 0 as n Ñ 8.
Furthermore, they are non-overlapping.
On the other hand, the change in φ across each such interval is φp0.a0 . . . an 1q´
φp0.a0 . . . an q “ 1{2n`2 . So the sum of all these changes in φ is 1{2! There-
fore, φ is not absolutely continuous.

Not all functions with bounded variation are absolutely continuous, but
all absolutely continuous functions have bounded variation.

Theorem 61. 1. Every absolutely continuous function has bounded vari-

ation.

2. In particular, every absolutely continuous function can be expressed as

the difference of two non-decreasing absolutely continuous functions.

Proof. For the first statement, suppose f : ra, bs Ñ R is absolutely continu-

ous. Then, there is a δ ą 0 so that if px0 , y0 q, . . ., pxn , yn q are subintervals
of ra, bs such that

1. ÿ
pyj ´ xj q ă δ, and
jďn

2. such that pxj , yj q X pxk , yk q “ H whenever j ‰ k,

94 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

then we have that ÿ

|f pyj q ´ f pxj q| ă 1.
jďn

Choose m P N so that pb ´ aq2´m ă δ.

We now claim that if P is any partition of ra, bs, then V pf, P q ă 2m . To

show this, let P be a partition of ra, bs. Let Q denote the uniform partition
of ra, bs of width pb ´ aq2´m . Let I0 , . . . , I2m ´1 denote the subintervals of
Q. Let P 1 “ P Y Q. Then, V pf, P q ď V pf, P 1 q. By the choice of m, δ,
V pf, P 1 X Ij q ă 1 for each j. Thus,
ÿ
V pf, P 1 q “ V pf, P 1 X Ij q ă 2m .
jă2m

Thus, f has bounded variation.

To show the second claim, note that f has bounded variation so we set

g1 pxq “ f pxq ` T V pf |ra,xs q

g2 pxq “ T V pf |ra,xs q

By our previous work, g1 and g2 are non-decreasing.

We now want to show that g1 and g2 are absolutely continuous. It suffices

to show that g2 is absolutely continuous. Let  ą 0. Choose δ ą 0 so that
ÿ
|f pyj q ´ f pxj q| ă {2
jăn

whenever px0 , y0 q, . . ., pxn , yn q are non-overlapping subintervals of ra, bs so

that the sum of their lengths is smaller than δ.
Now, suppose px0 , y0 q, . . ., pxn , yn q are non-overlapping subintervals of
ra, bs so that the sum of their lengths is smaller than δ. We want to show
that ÿ
|g2 pyj q ´ g2 pxj q| ă .
jďn

To do so, we use the fact that

g2 pyj q ´ g2 pxj q “ T V pf |ra,yj s q ´ T V pf |ra,xj s q “ T V pf |rxj ,yj s q.

If Pj is a partition of rxj , yj s for each j ď n, then by the choice of δ,

ÿ
V pf, Pj q ă {2
jďn
5.3. ABSOLUTELY CONTINUOUS FUNCTIONS 95

So, ÿ
T V pf |rxj ,yj s q ď {2 ă .
jďn

With this fact, we can now establish a lemma that gets us most of the
way to Theorem 60.

Lemma 18. Suppose f : ra, bs Ñ R is absolutely continuous. Then, tDiff 2´n pf qu8
n“0
is uniformly integrable.

Proof. Suppose f is absolutely continuous. Without loss of generality, we

can suppose f is non-decreasing. Let  ą 0 and choose δ ą 0 as in the
definition of absolute continuity for {2.
Suppose n P N, and set h “ 2´n . Now let E is a measurable subset of
ra, bs so that µpEq ă δ{2. Then, there is an open set U so that U Ě E
and µpU q ă δ. There is a Ť pairwise disjoint family of bounded open intervals
tpcj , dj qujPF so that U “ jPF pcj , dj q. Then:
ż ż
Diff h pf qdµ “ Diff h pf qdµ
E U ż
ÿ
“ Diff h pf qdµ
jPF pcj ,dj q
ÿ
“ rAvh pdj q ´ Avh pcj qs
jPF

1 ÿ h
ż
“ rf pdj ` tq ´ f pcj ` tqsdµptq
h jPF 0

If F 1 Ď F is finite, then

1 ÿ h
ż żh ÿ
1
rf pdj ` tq ´ f pcj ` tqs dµptq “ |f pdj ` tq ´ f pcj ` tq| dµptq
h jPF 1 0 h 0 jPF 1

1 
ă h “ {2.
h 2
Thus,
1 ÿ h
ż
rf pdj ` tq ´ f pcj ` tqsdµptq ď {2 ă .
h jPF 0
96 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

Using this lemma, we can now prove the main result from this section.

Proof. The first claim follows immediately from the fact that f has bounded
variation.

(2) and (3): By the Lemma, tDiff 2´n pf qu8

n“0 is uniformly integrable.
Furthermore,

lim Diff 2´n pf q “ f 1 a.e. by the first claim.

nÑ8

By the Vitali Convergence Theorem, f 1 is integrable and

ż ż
1
f dµ “ lim Diff 2´n pf qdµ.
ra,bs nÑ8 ra,bs

However,
ż
lim Diff 2´n pf qdµ “ lim Av2´n pf qpbq ´ Av2´n pf qpaq “ f pbq ´ f paq.
nÑ8 ra,bs nÑ8

5.4 Indefinite Integrals and the Fundamental The-

orem of Calculus
We can now use our work to define the notion of the indefinite integral.
Definition 43. Suppose f, g : ra, bs Ñ R. We say that f is the indefinite
integral of g if g is integrable and
ż
f pxq “ f paq ` gdµ
ra,xs

for all x P ra, bs.

Theorem 62. Suppose f : ra, bs Ñ R. Then, f is absolutely continuous if
and only if f is the indefinite integral of a function on ra, bs.
Proof. pñq: By Theorem 60.

pðq: Suppose g : ra, bs Ñ R is integrable and

ż
f pxq “ f paq ` g dµ
ra,xs
5.4. INDEFINITE INTEGRALS AND THE FUNDAMENTAL THEOREM OF CALCULUS97

for all x P ra, bs. Let  ą 0.

By a homework exercise, there is a δ ą 0 so that
ż
|g| dµ ă 
E

whenever E is a measurable subset of ra, bs with µpEq ă δ. Suppose

px0 , y0 q, . . . , pxn , yn q
are non-overlapping subintervals of ra, bs so that the sum of their lengths is
smaller than δ. Set ď
E “ pxj , yj q.
j

Then:
ˇ ˇ
ÿ ÿ ˇˇż ˇ
|f pyj q ´ f pxj q| “ g dµˇ
ˇ
ˇ
jďn rxj ,yj s
jďn
ˇ ˇ
ÿ ż
ď |g| dµ
jďn rxj ,yj s
ż
“ |g| dµ ă .
E

For monotone functions, we can state this result a bit differently.

Theorem 63. Suppose f : ra, bs Ñ R is monotone. Then, f is absolutely
continuous if and only if
ż
f 1 dµ “ f pbq ´ f paq.
ra,bs

Proof. Before proving this, it is worth noting that f being monotonic implies
that f 1 is integrable, which implies that the left hand side is well defined.
The forward direction follows immediately from Theorem 60.

For the converse direction, we suppose, without loss of generality, f is

non-decreasing.
We need to show f is absolutely continuous. To do so, it suffices to show
that f is the indefinite integral of f 1 . Let x P ra, bs. On the one hand:
ż żx żb
1 1
f pbq ´ f paq “ f dµ “ f dµ ` f 1 dµ
ra,bs a x
98 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

But, f pbq ´ f paq “ f pbq ´ f pxq ` f pxq ´ f paq. So, we conclude

żx żb
1
0“ f dµ ´ rf pxq ´ f paqs ` f 1 dµ ´ rf pbq ´ f pxqs
a x

However, since f is non-decreasing, by Proposition 58,

żx
f 1 dµ ď f pxq ´ f paq
a
żb
f 1 dµ ď f pbq ´ f pxq
x
Thus, both differences in the above equation are non-positive. So,
żx
f 1 dµ “ f pxq ´ f paq
a

We are almost ready to prove the second form of the fundamental theo-
rem of calculus. Before doing so, we need a small lemma.
Lemma 19. Suppose f : ra, bs Ñ R is integrable Then, f “ 0 a.e. if and
only if ż x2
f dµ “ 0
x1

whenever px1 , x2 q Ď ra, bs

Instead of giving a full proof of this, note that the forward şdirection is
immediate. For the reverse direction, note that it implies that U f dµ “ 0
for every open set U . Then, we can approximate measurable sets with open
sets to show that ż
f dµ “ 0
E
for any measurable set. Then we can take the positive and negative parts
of f and apply Chebychev’s inequality (Theorem 40).
Theorem 64. Suppose f : ra, bs Ñ R is integrable. Then,
d x
ż
f dµ “ f pxq a.e.
dx a
1

1
What is the complexity of this exceptional set?
5.5. A BRIEF REMARK ON RADON-NIKODYM DERIVATIVES 99

Proof. Set żx
F pxq “ f dµ
0
So, F is absolutely continuous. Thus, F is differentiable a.e. When px0 , x1 q Ď
ra, bs, by definition F and Theorem 63
ż x1
pF 1 ´ f qdµ “ 0
x0

So by the Lemma, F 1 “ f a.e.

5.4.1 The Lebesgue Decomposition

Before concluding this chapter, let me make one further remark about func-
tions of bounded variation. This will probably not be meaningful until you
study measure theory in more depth, but it is worth noting now.
Definition 44. Suppose f : ra, bs Ñ R. f is singular if f 1 “ 0 a.e.
Given any function of bounded variation, it is possible to decompose it
as a sum of an absolutely continuous function with a singular function
Theorem 65. Suppose f : ra, bs Ñ R has bounded variation. Then, f can
be written as the sum of an absolutely continuous function and a singular
function. This decomposition is known as the Lebesgue Decomposition
Proof. Since f has bounded variation, f is differentiable a.e.. Set
żx
gpxq “ f 1 dµ
a
hpxq “ f pxq ´ gpxq

Thus, g is absolutely continuous. h1 “ 0 a.e. and f “ g ` h.

5.5 A brief remark on Radon-Nikodym derivatives

In this course, we will unfortunately not be able to discuss measures other
than the Lebesgue measure. However, the real power of absolute continuity
appears when one considers more general measures, so I felt it was necessary
to include a brief remark about this.
To begin, we start with a definition of absolute continuity, not for func-
tions, but rather for measures.
100 CHAPTER 5. DIFFERENTIATION AND INTEGRATION

Definition 45. A measure µ on Borel subsets of the real line is absolutely

continuous with respect to the Lebesgue measure if for every measurable set
E, µpEq “ 0 implies νpEq “ 0. This is written as ν ! µ.
There is a straightforward way to create measures which are Lebesgue
absolutely continuous. To do so, one takes a non-negative measurable func-
tion f : R Ñ R` , and defines the measure to be
ż
νpEq “ f dµ.
E

Example. 1. The uniform measure on r0, 1s is generated by the function

f “ χr0,1s

2. The Gaussian measure is generated from the function

1 1 x´µ 2
f pxq “ ? e´ 2 p σ q ,
σ 2π
where µ and σ are parameters.
It turns out that all measures which are absolutely continuous can be
expressed in this way.
Theorem 66 (Radon-Nikodym Theorem). if ν ! µ, then there exists a
µ-measurable function2 f : R Ñ r0, 8q, such that for any measurable set
EĎR ż
νpEq “ f dµ.
E

dν
In this case, the function f (also written as is said to be the Radon-
dµ
Nikodym derivative of ν with respect to µ.
It is natural to ask what the relationship between absolutely continuous
functions and absolutely continuous measures are. In fact, this comes down
to the following simple fact.
Proposition 67. A finite measure ν on Borel subsets of the real line is
absolutely continuous with respect to Lebesgue measure if and only if the
point function
F pxq “ νp´8, xsq
is an absolutely continuous real function.
2
More precisely, a function which is measurable with respect to the associated σ-
algebra.
5.5. A BRIEF REMARK ON RADON-NIKODYM DERIVATIVES 101

Note that if ν has total mass one, then F is simply the cumulative
distribution function. Also note that if you assume Proposition 67, then
proving Theorem 66 is essentially just a matter of appealing to Theorem
62, so you can intuitively think about the latter result as being a Radon-
Nikodym theorem in disguise.

5.5.1 An even briefer remark on the Lebesgue Decomposi-

tion theorem
The motivation for Theorem 65 might seem somewhat mysterious. In fact,
this theorem is a special case of a better known theorem for measures, which
is the following.

Theorem 68. Let µ and ν be σ-finite measures on a measurable space

pΩ, Fq. Then ν can be uniquely decomposed into ν “ νc ` νs where νc ! µ
and νs K µ.

Here, the notation νs K µ means that νs is singular with respect to µ.

Definition 46. A measure νs is singular with respect to µ if it is possible to

decompose F into two disjoint subsets Eνs and Eµ so that for any set A P Eνs ,
µpAq “ 0 and for any B P Eµ , νs pBq “ 0.

Out of context (and with some of the terms as yet undefined), this result
might just be word salad. However, this result, when combined with the
Radon-Nikodym theorem, give a good way to understand the relationship
between arbitrary measures on reasonable measure spaces.
102 CHAPTER 5. DIFFERENTIATION AND INTEGRATION
Chapter 6

Convex functions

We now turn our attention to the notion of convexity.

Definition 47. Suppose f : pa, bq Ñ R where ´8 ď a ă b ď 8.

1. f is convex if

f pλx1 ` p1 ´ λqx2 q ď λf px1 q ` p1 ´ λqf px2 q

whenever a ă x1 , x2 ă b and 0 ď λ ď 1.

2. f is strictly convex if

f pλx1 ` p1 ´ λqx2 q ă λf px1 q ` p1 ´ λqf px2 q

whenever a ă x1 , x2 ă b and 0 ă λ ă 1.

3. A function is strongly convex if

1
f ptx ` p1 ´ tqyq ď tf pxq ` p1 ´ tqf pyq ´ mtp1 ´ tqpx ´ yq2 .
2

In other words, convex functions are those which lie below their secant
lines, strictly convex functions are those which lie strictly below their secant
lines and strongly convex functions lie below a quadratic function

103
104 CHAPTER 6. CONVEX FUNCTIONS

Definition 48. A function f is concave (respectively, strictly, strongly) if

´f is convex (strictly convex, strongly convex).

There are several examples of convex functions which should be familiar.

1. When p ą 0, x ÞÑ xp strictly convex if p ą 1, strictly concave if p ă 1.

2. The exponential function is strictly convex.

We can now show an alternative characterization of convexity in terms

of increasing difference quotients.

Proposition 69. f : pa, bq Ñ R is convex if and only if

f pxq ´ f px1 q f px2 q ´ f pxq

ď
x ´ x1 x2 ´ x
whenever a ă x1 ă x ă x2 ă b.

Proof. pñq: Suppose f is convex. Suppose a ă x1 ă x ă x2 ă b. Then,

there exists 0 ď λ ď 1 so that x “ λx1 ` p1 ´ λqx2 . Thus,

f pxq ´ f px1 q λf px1 q ` p1 ´ λqf px2 q ´ f px1 q

ď
x ´ x1 λx1 ` p1 ´ λqx2 ´ x1
f px2 q ´ f px1 q
“
x2 ´ x1

Similarly:
f px2 q ´ f px1 q f px2 q ´ f pxq
ď .
x2 ´ x1 x2 ´ x
6.1. THE DERIVATIVES AND SUB-DERIVATIVES OF CONVEX FUNCTIONS105

pðq: Suppose
f pxq ´ f px1 q f px2 q ´ f pxq
ď
x ´ x1 x2 ´ x
whenever a ă x1 ă x ă x2 ă b. Suppose 0 ď λ ď 1. We can assume
0 ă λ ă 1. Set x “ λx1 ` p1 ´ λqx2 . It follows that
x2 ´ x
λ“
x2 ´ x1
and that
x ´ x1
1´λ“ .
x2 ´ x1
Since 0 ă λ ă 1, x1 ă x ă x2 . Thus,

f pxq ´ f px1 q f px2 q ´ f pxq

ď .
x ´ x1 x2 ´ x
Therefore
x2 ´ x x ´ x1
pf pxq ´ f px1 qq ď pf px2 q ´ f pxqq
x2 ´ x1 x2 ´ x1
i.e.
λpf pxq ´ f px1 qq ď p1 ´ λqpf px2 q ´ f pxqq.
Which rearranges to f pxq ď λf px1 q ` p1 ´ λqf px2 q.

6.1 The derivatives and sub-derivatives of convex

functions
One nice property of convex functions is that their derivatives satisfy several
nice properties. In order to discuss this, we first define the notion of the left
and right derivative.

Definition 49. Suppose f : pa, bq Ñ R. Suppose a ă x0 ă b.

f px0 `hq´f px0 q
1. B´ pf qpx0 q “ limhÑ0´ h .
B´ pf q is called the left hand derivative of f .
f px0 `hq´f px0 q
2. B` pf qpx0 q “ limhÑ0` h .
B` pf q is called the right hand derivative of f .
106 CHAPTER 6. CONVEX FUNCTIONS

The reason to define these two derivatives is that they are always well
defined and increasing for convex functions.

Proposition 70. Suppose f is convex on pa, bq. Then,

1. f has left and right hand derivatives at each point in pa, bq.

2. If a ă u ă v ă b, then

f pvq ´ f puq
B´ pf qpuq ď B` pf qpuq ď ď B´ pf qpvq ď B` pf qpvq.
v´u

Proof. (1) : Suppose x1 ą u. By Proposition 69,

f px1 q ´ f puq
x1 ´ u
decreases as x1 Ñ u. On the other hand, if a ă c ă u, by Proposition 69
again,
f puq ´ f pcq f px1 q ´ f puq
ď
u´c x1 ´ u
whenever x1 ą u. Thus, B` pf qpuq exists. Similarly, B´ pf qpuq exists.

(2): Proof of (1) shows that B´ pf q ď B` pf q. Suppose u ă x1 ă v. By

Proposition 69,
f puq ´ f px1 q f pvq ´ f puq
ď .
u ´ x1 v´u
Now, let x1 Ñ u` . We obtain

f pvq ´ f puq
B` pf qpuq ď .
v´u
We similarly show that

f pvq ´ f puq
ď B´ pf qpvq.
v´u

This result has several important corollaries. For instance, it shows that
convex functions are Lipschitz.

Corollary 17. If f is convex on I, and if ra, bs Ď I, then f is Lipschitz on

ra, bs.
6.1. THE DERIVATIVES AND SUB-DERIVATIVES OF CONVEX FUNCTIONS107

Proof. By the Proposition, if a ď x ă y ď b,

f pyq ´ f pxq
B` pf qpaq ď ď B´ pbq.
y´x

Set M “ maxt|B` pf qpaq|, |B´ pf qpbq|u. Then, |f pxq ´ f pyq| ď M |x ´ y|.

Theorem 71. Suppose f is convex on I. Then

1. f is differentiable except at countably many points of I, and

2. f 1 is non-decreasing.

Proof. By Proposition 70, B´ pf q and B` pf q are non-decreasing. and B´ pf q ď

B` pf q, where the inequality is strict only at the points at which f 1 does not
exist. As such, these points are in one-to-one correspondence with a pairwise
disjoint family of open intervals. When f 1 px0 q exists, f 1 px0 q “ B` pf qpx0 q.
So, f is non-decreasing.

This has the following corollary.

Corollary 18 (Alexandrov’s theorem). Suppose f is convex on I. Then f

is twice differentiable almost everywhere.

In the proof of Theorem ??, the subdifferential played an important role,

so we will investigate it further now.

Definition 50. The set Bf pxq “ rB´ f pxq, B` f pxqs is known as the sub-
differential of f at x.

The sub-differential plays a crucial role in the analysis of convex func-

tions. The standard definition of the sub-differential is the following.
č
Bf pxq “ tv | f pzq ě f pxq ` v ¨ pz ´ xqu
zPdom f

Note that by considering v as a vector in Euclidean space, the sub-

differential is well defined for functions f : Rn Ñ R. It turns out that the
sub-differential is always a closed and convex set. This is obvious for convex
functions in R, but holds true more generally as well.
108 CHAPTER 6. CONVEX FUNCTIONS

6.2 Jensen’s inequality

In this course, and on the qualifying exams, one of the best tools that you
have involving convex functions is Jensen’s inequality. As such, we will
spend some time discussing it now.
To begin, we note that convex functions lie above their tangent lines
(where tangent is defined to be any point in the sub-differential).

Lemma 20. Suppose f is convex on pa, bq and that a ă x0 ă b. If

B´ pf qpx0 q ď m ď B` pf qpx0 q, then f pxq ě mpx ´ x0 q ` f px0 q.

Using this, we can now state and prove Jensen’s inequality.

Theorem 72. (Jensen’s Theorem): If φ is a convex function on R, and if

f , φ ˝ f integrable on E where µpEq “ 1, then
ˆż ˙ ż
φ f dµ ď φ ˝ f dµ.
E E
ş
Proof. Set α “ E f dµ. Suppose B´ pφqpαq ď m ď B` pf qpαq. By Lemma 20,

φpf pxqq ě mpf pxq ´ αq ` φpαq.

Now, take the integral of both sides.

Example: If f : r0, 1s Ñ R is integrable, then

ˆż 1 ˙2 ż 1
f dµ ď f 2 dµ.
0 0

Question 2 (2011 Analysis Qualifying Exam, Problem 3). Let f : r0, 1s Ñ

R be continuous with f pxq ą 0 for x P r0, 1s. Show that
ˆż 1 ˙ ż1
exp log f ď f.
0 0

6.3 A brief remark on convex duality

Any discussion of convex functions would be incomplete without some dis-
cussion of convex duality. One of the fundamental properties of convex
functions is that they come in pairs.
6.3. A BRIEF REMARK ON CONVEX DUALITY 109

Definition 51 (Legendre transform). Let I Ă R be an interval, and f : I Ñ

R a function. The Legendre transform of f (denoted f ˚ ) is the function
f ˚ : I ˚ Ñ R defined by

f ˚ px˚ q “ sup px˚ x ´ f pxqq , x˚ P I ˚

xPI

where the domain I ˚ is defined to be

" *
I ˚ “ x˚ P R : sup px˚ x ´ f pxqq ă 8
xPI

Proposition 73. The Legendre transform of any function is a convex func-

tion.

Proof. Let p and q be two points in the domain I ˚ where f ˚ is defined and
let t be within r0, 1s. Then,

f ˚ ptp ` p1 ´ tqqq “ sup tx rtp ` p1 ´ tqqs ´ f pxqu

x
“ sup tt rxp ´ f pxqs ` p1 ´ tq rxq ´ f pxqsu
x
ď t sup txp1 ´ f pxqu ` p1 ´ tq sup txq ´ f pxqu
x x
“ tf ˚ ppq ` p1 ´ tqf ˚ pqq .

For convex functions, it turns out that this transformation is involutive.

We will not prove this here because I was not able to find a good proof using
only what we’ve covered thus far.

Proposition 74. A function is convex iff f ˚˚ “ f.

There is a lot more that can be said about the Legendre transform.
For instance, there is a deep relationship between the sub-gradients of the
Legendre pair.
Furthermore, this duality appears throughout mathematics. To give a
short and very incomplete list, the it shows up in

1. Mathematical Physics (Classical mechanics, thermodynamics . . . )

2. Optimal transport

3. Mirror symmetry (in particular, T-duality).

110 CHAPTER 6. CONVEX FUNCTIONS
Chapter 7

Basics of functional analysis

In analysis, we will often want to establish quantitative control over our

functions. For instance, if we are studying a surface of the ocean, we want
to understand how large the waves are and how smooth or rough the chop
is. In order to do that, it is necessary to develop a more complete theory of
how to bound functions. In this chapter, we cover some preliminary ideas
in this direction, which lay the groundwork for modern functional analysis.

7.1 Normed linear spaces and Banach spaces

Definition 52. Suppose V is a vector space over R. A norm on V is a
function } ¨ } : V Ñ r0, 8q so that
• }αv} “ |α|}v} for all α P R and all v P V . (Homogeneity)
• For all v P V , if }v} “ 0, then v “ 0. (Positivity)
• For all u, v P V , }u ` v} ď }u} ` }v}. (The triangle inequality)
Definition 53. A normed linear space consists of a vector space together
with a norm.
Example. 1. Rn with its usual norm is a normed linear space.
2. Let Cr0, 1s denote the space of all continuous functions
f : r0, 1s Ñ R.

Cr0, 1s is a vector space under the usual pointwise addition and scalar
multiplication. For each f P Cr0, 1s, we let
}f }sup “ maxt|f ptq| : t P r0, 1su.

111
112 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

It follows that }f }sup is a norm on Cr0, 1s.

3. Integrable functions with the norm

ż
}f }1 “ |f | dµ.

The reason for the “1” in the subscript of the previous norm will soon
become clear. The examples that we will be interested in are mostly function
spaces such as the latter two spaces, as much of what will say is trivial for
finite dimensional spaces.
The main reason to study normed linear spaces is that they induce a
notion of convergence.

Definition 54. Suppose V is a normed linear space, and suppose tvn u8

n“0
is a sequence of vectors.

1. We say that tvn u8

n“0 converges in V (or converges in norm) if there is
a vector v P V so that limnÑ8 }vn ´ v} “ 0.

2. We say that tvn u8n“0 is Cauchy if for every  ą 0 there is an N P N

so that }vn ´ vm } ă  whenever m, n ě N .

In this class we will not cover notions of convergence other than conver-
gence in norm, but you may well do so in future classes so we will use the
correct terminology from the start.
There is a relationship between convergent sequences and Cauchy se-
quences, which should hopefully be familiar from a previous course in anal-
ysis.

Proposition 75. Suppose V is a normed linear space, and suppose tvn u8

n“0
is a sequence of vectors in V

1. If tvn u8
n“0 converges in V , then there is exactly one vector v in V so
that limnÑ8 }vn ´ v} “ 0; denote this vector by limnÑ8 vn .

2. If tvn u8
n“0 converges in V , then it is Cauchy.

3. Suppose tvn u8 8
n“0 is Cauchy. If a subsequence of tvn un“0 converges in
8
V , then tvn un“0 converges in V .

Definition 55. A normed linear space is complete if all of its Cauchy se-
quences converge; in this case we say that it is a Banach space.

Example. Cr0, 1s with the supremum norm is a Banach space.

7.2. LP SPACES 113

Exercise 16. Show this fact. Hint: Use uniform convergence.

Non-Example. Cr0, 1s with the norm

dż
}f }L2 “ f 2 dµ

is an incomplete normed linear space.

Exercise 17. Find a Cauchy sequence in this space which does not converge.

Before moving on, let us mention two lemmas which will be useful some
future considerations.

Lemma 21. Suppose V is a normed linear space, and suppose tvn u8 n“0
is a Cauchy sequence of vectors in V . Suppose taj u8 j“0 is a sequence of
positive reals that converges to 0. Then, there exist n0 ă n1 ă . . . so that
}vn2j ´ vn2j`1 } ă aj for all j P N.

The proof is left as a small exercise.

Lemma 22. Suppose tvn u8 n“0 is a convergent sequence of vectors in a

normed linear space V and that v denotes its limit. Then, limnÑ8 }vn } “
}v}.

Proof sketch: Observe that

|}vn } ´ }v}| ď }vn ´ v}.

7.2 Lp spaces
Definition 56. Suppose E is a measurable set of reals. Suppose 1 ď p ă 8.

1. Lp pEq consists of all measurable f : E Ñ r´8, 8s so that

ż
|f |p dµ ă 8
E

2. When f P Lp pEq, let

ˆż ˙1{p
p
}f }p “ |f | dµ
E
114 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Before moving on, we make a few remarks. First, note that L1 pEq con-
sists of all integrable functions on E.
Furthermore, when f, g P Lp pEq and f “ g a.e., we identify f and g.
As such, Lp spaces are really equivalence classes of functions rather than
functions.
There is also an L8 -norm. However, in order to define it, we need to
define some more terminology

Definition 57. Suppose S Ď R.

1. M P p´8, 8s is an essential upper bound on S if M ą a for almost

all a P S.

2. The essential supremum of S is the greatest lower bound of the set of

all essential upper bounds of S.

Exercise 18. To make sense of this defintion, consider the following space

r0, 1q Y t3, 4, 5u.

What is its essential supremum?

With this definition, we can define the L8 norm.

Definition 58. Suppose E is a measurable set of reals.

1. L8 pEq consists of all measurable f : E Ñ r´8, 8s so that the essential

supremum of ranp|f |q is finite.

2. When f P L8 pEq, let }f }8 denote the essential supremum of f .

The relationship between Lp functions and L8 ones can be a bit subtle.

Here is an exercise to help think about it.

Exercise 19. Suppose f is measurable and satisfies

ż1
expp|f pxq|qdx ă 8
0

Prove that f P Lp pr0, 1sq for all 1 ď p ă 8. Is it true that such an f must
necessarily belong to L8 pr0, 1sq?
7.3. CONJUGATE EXPONENTS AND SOME FUNDAMENTAL INEQUALITIES115

7.3 Conjugate exponents and some fundamental

inequalities
One of the most important features of the Lp -spaces is that they come in
conjugate pairs and satisfy several fundamental inequalities.
Definition 59. Suppose 1 ď p, q ď 8. We say that p, q are conjugate if
1 1
p ` q “ 1.

Example. 1. 1 and 8 are conjugate.

2. 2 and 2 are conjugate.

3
3. 3 and 2 are conjugate.
Using some algebra, we can observe two basic, but very important ob-
servations
Observation 6. 1. If q, p P R are conjugate then q “ p{pp ´ 1q.

2. If p, q are conjugate, and if p ą 2, then 1 ď q ă 2.

7.3.1 Young’s inequality and the Peter-Paul version

Theorem 76 (Young’s Inequality). If p, q are conjugate then
ap bq
ab ď `
p q
for all reals a, b ě 0.
Proof. Without loss of generality, we can consider a, b ‰ 0. Set: s1 “ p lnpaq
and s2 “ q lnpbq. Since exp is convex,
1 s1 1 s2
es1 {p`s2 {q ď e ` e
p q
1 p 1 q
“ a ` b
p q

But, ab “ es1 {p`s2 {q .

Oftentimes in analysis, it will be useful to use a slightly different version

of this inequality, which is known as the Peter-Paul inequality. The basic
idea is that you “rob Peter in order to pay Paul”. In other words, you gain
tighter control of the second term at the expense at the cost of losing some
116 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

control of the first term. For the case p “ q “ 2, this inequality is the
following.

a2 εb2
ab ď ` . (7.1)
2ε 2

7.3.2 Holder’s Inequality

Theorem 77. (Hölder’s inequality): Suppose E is a measurable set of reals.
Suppose p, q are conjugate, f P Lp pEq, and g P Lq pEq. Then, f g P L1 pEq
and
}f g}1 ď }f }p }g}q

Proof. We assume that 0 ă }f }p , }f }q , or else the proof is trivial.

Case 1: If one of p, q is infinite, then the inequality holds directly from
the monotonicity of the integral.‘
Case 2: p, q ă 8.

Claim 1:
|f | |g| 1 |f |p 1 |g|q
ď p ` .
}f }p }g}q p }f }p q }g}qq
Proof Claim 1: By Young’s Inequality.

Claim 2: ż
|f ||g|
ď 1.
E }f }p }g}q
Proof Claim 2: By Claim 1,

|f |p 1 |g|q
ż ż ż
|f | |g| 1
dµ ď p dµ ` q dµ
E }f }p }g}q p E }f }p E q }g}q
1 1
“ ` “ 1.
p q
˝Claim 2 .

The theorem now follows directly from Claim 2.

Holder’s inequality is saturated in by the conjugate function, which is

the following.
7.3. CONJUGATE EXPONENTS AND SOME FUNDAMENTAL INEQUALITIES117

Proposition 78. If f is not identically 0, the function 2 f ˚ “ }f }p1´p ¨sgnpf q¨

|f |p´1 belongs to Lq pX, µq and satisfies
ż
f ¨ f ˚ “ }f }p and }f ˚ }q “ 1.
E

Here, sgnpf q is the function which is 1 whenever f is non-negative and ´1

whenever f is negative so that sgnpf q ¨ f “ |f | a.e.

This is a favorite topic of the people who write the qualifying exam.

Exercise 20 (Fall 2021 Qual: Problem 4). Let f P Lp pRq, }f }p ‰ 0. Prove

that there exists a g P Lq pRq such that }g}q “ 1 and
ż
1
f pxqgpxqdx “ }f }p .
R 2
The case of Holder’s inequality with p “ q “ 2 is of particular impor-
tance, and is better known as the Cauchy-Schwarz inequality.

Corollary 19 (Cauchy-Schwarz inequality). If E is a measurable set of

reals, and if f, g P L2 pEq, then }f g}1 ď }f }2 }g}2 .

7.3.3 Minkowski’s inequality

Theorem 79. Suppose E is a measurable set of reals, 1 ď p ď 8, and
f, g P Lp pEq. Then,
}f ` g}p ď }f }p ` }g}p

Proof. WLOG p ă 8. Let q be conjugate of p.

Claim 1: pf ` gqp “ f pf ` gqp{q ` gpf ` gqp{q .

To see this, note that since p, q are conjugate, p ´ 1 “ pq . So

pf ` gqp “ f pf ` gqp´1 ` gpf ` gqp´1

“ f pf ` gqp{q ` gpf ` gqp{q .

Claim 2:
}f ` g}pp ď }|f ` g|p{q }q }f }p ` }g}p q.
Proof Claim 2: By Claim 1 and the triangle inequality,

|f ` g|p ď |f ||f ` g|p{q ` |g||f ` g|p{q .

118 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Thus, ż ż ż
p p{q
|f ` g| dµ ď |f ||f ` g| ` dµ |g||f ` g|p{q dµ
E E E
Note that |f `g|p{q P Lq pEq. So, by Hölder’s Inequality, |f ||f `g|p{q P L1 pEq
and |g||f ` g|p{q P L1 pEq and
ż ż
p{q
|f ||f ` g| dµ ` |g||f ` g|p{q dµ “ }|f ||f ` g|p{q }1 ` }|g||f ` g|p{q }1
E E
ď }f }p }|f ` q|p{q }q ` }g}p }|f ` g|p{q }q .

Finally, we can show the desired inequality.

}f ` g}p ď }f }p ` }g}p .
To see this, note that by Claim 2,
}f ` g}pp }|f ` g|p{q }´1
q ď }f }p ` }g}p

But,
ˆż ˙1´1{q
}f ` g}pp }|f ` g|p{q }´1
q “ pf ` gq dµ p
“ }f ` g}p .
E

This has the following important corollary, which is that Lp spaces form
a normed linear space.
Corollary 20. If E is a measurable set of reals, and if 1 ď p ď 8, then
Lp pEq is a normed linear space.
Before moving on, it is worthwhile to write out some functions which are
in various Lp -spaces.
Example. 1. Set E “ p0, 1s. Suppose 1 ď p1 ă p2 ă 8. Choose α so
that ´1{p1 ă α ă ´1{p2 . Set f pxq “ xα when 0 ă x ď 1. It follows
that f P Lp2 pEq ´ Lp1 pEq.
2. Set E “ p0, 8q. For all x ą 0, let
x´1{2
f pxq “
1 ` lnpxq
It follows that f P Lp pEq iff p “ 2.ş Proof sketch: use change of
8
variables u “ 1 ` lnpxq. When p ą 2, 1 |f pxq|p dx “ 8. When p ă 2,
ş1 p
0 |f pxq| dx “ 8.
7.3. CONJUGATE EXPONENTS AND SOME FUNDAMENTAL INEQUALITIES119

7.3.4 Embeddings of Lp -spaces

Theorem 80. Suppose µpEq ă 8 and 1 ď p1 ă p2 ă 8. Then, Lp2 pEq Ď

Lp1 pEq.

Proof. Set p “ p2 {p1 ą 1. Let f P Lp2 pEq. Thus, |f |p1 P Lp pEq. Set q “ p´1
p
so that p, q are conjugate. Set g “ χE . Since µpEq ă 8, g P Lq pEq. So,

ż ż
|f |p1 dµ “ |f |p1 gdµ
E E
ď }|f |p1 }p }g}q
“ }} fpp21 µpEqq ă 8.

Thus, f P Lp1 pEq.

7.3.4.1 Interpolation of Lp spaces

On R, none of the Lp spaces are subspaces of others. However, there is a

foundational result which states that if f P Lp0 X Lp1 for p0 ă p1 , it is also
in Lq for all q with p0 ă q ă p1 and we can estimate the Lq norm in terms
of the Lp0 - and Lp1 -norms.
To explain this result, we first introduce some notation.

Definition 60. Let p0 , p1 be two numbers such that 0 ă p0 ă p1 ď 8. Then

for 0 ă θ ă 1 define pθ by: p1θ “ 1´θ θ
p0 ` p1 .

Theorem 81. Each f P Lp0 X Lp1 satisfies:

}f }pθ ď }f }p1´θ
0
}f }θp1 .
120 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Proof. Consider |f | “ |f |θ |f |1´θ and apply Holder’s inequality:

ż ż ´ ¯pθ
|f |pθ “ |f |θ |f |1´θ (7.2)
ż
“ |f |θpθ |f |p1´θqpθ (7.3)
θp ˙ p1´θqpθ
ˆż ´ ¯ p1 ˙ p1θ ˆż ´ ¯ p0
p0
θpθ θpθ p1´θqpθ p1´θqpθ
ď |f | |f | (7.4)
ˆż ˙ θpθ ˆż ˙ p1´θqpθ
p1 p0
p1 p0
“ |f | |f | (7.5)
ˆż ˙ θpθ ˆż ˙ p1´θqpθ
p1 p0
p1 p0
“ |f | |f | (7.6)

“ }f }θp θ p1´θqpθ
p1 }f }p0 (7.7)
Then take the pθ -th roots of both sides.

7.4 Lp and pointwise convergence

We can now discuss the relationship between convergence in Lp and usual
pointwise convergence of functions.
Theorem 82. Suppose 1 ď p ď 8 and E is a measurable set of reals. Then,
every Cauchy sequence of vectors in Lp pEq has a subsequence that converges
pointwise almost everywhere.
Proof. Suppose tfn u8 p
n“0 is a Cauchy sequence of vectors in L pEq. By
Lemma 21, there is a sequence n0 ă n1 ă . . . so that }fnj ´ fnj`1 }pp ă 2´2j
for all j.
For each j P N, let
Ej “ tx P E : |fnj pxq ´ fnj`1 pxq|p ą 2´j u.
Claim 1: limj fnj pxq exists if x P Ej for only finitely many j.

Proof Claim 1: Suppose x P Ej for only finitely many j. Let  ą 0.

Choose k so that x R Ej for all j ě k and 2´k`1 ă . Suppose j0 , j1 ě k
and j1 ą j0 . Then, by the triangle inequality
8
ÿ ´j
|fnj pxq ´ fnj`1 pxq| ď 2 p

j“j0

“ 2´j0 `1 ď 2´k`1 ă .
7.4. LP AND POINTWISE CONVERGENCE 121

Claim 2: For almost every x P E, x P Ej for only finitely many j.

Proof Claim 2: By Chebychev,

ż
j
µpEj q ď 2 |fnj ´ fnj`1 |p dµ ă 2´j .
E
ř
So j µpEj q ă 8. By Borel-Cantelli, Claim 2 follows.

Note that the passing to a subsequence might be necessary. It is possible

to find sequences which converge to zero in Lp but do not converge pointwise.

Similarly, it is possible to construct families of functions which converge

pointwise but do not converge in Lp .

Example 2. 1. fn “ χrn,n`1s

2. fn “ nχp0, 1 s
n

Lemma ř 23. Suppose E is a measurable set of reals. Suppose 1 ď p ă 8.

Suppose 8n“0 }fn }p ă 8. Then:
122 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

ř8
1. n“0 fn converges almost everywhere.
řm
2. If f “ ř8 p
ř
n“0 fn , then f P L pEq and limmÑ8 }f ´ n“0 fn }p “ 0 and
}f }p ď 8n“0 }f n } p .
ř8
Proof. (2): Suppose f “ n“0 fn . Set:

m
ÿ
gm “ |fn |
n“0
g “ lim gm
mÑ8

Claim 1: g P Lp pEq and limmÑ8 }gm ´ g}p “ 0.

By definition, |gm |p ď |gm`1 |p . So, by MCT

ż ż
p
lim |gm | dµ “ |g|p dµ.
mÑ8 E E

i.e.
ż
lim }gm }pp “ }g}pp .
mÑ8 E

But,
m
ÿ
}gm }p ď }fn }p .
n“0
ř8
Thus, }g}p ă 8 since n“0 }fn }p .
If m P N, then
ˇ ˇp
ż ˇm`k ˇ
ˇÿ
gm }pp lim ˇ |fn |ˇ dµ
ˇ
}g ´ “
E kÑ8 ˇ ˇ
n“m
ˇp
ˇ
ż ˇm`k ˇ
ˇÿ
“ lim |fn |ˇ dµ (MCT)
ˇ
ˇ
kÑ8 E ˇ ˇ
n“m
m`k
ÿ
“ lim } fn }pp
kÑ8
n“m
7.4. LP AND POINTWISE CONVERGENCE 123

So,
m`k
ÿ
}g ´ gm }p “ lim } |fn |}p
kÑ8
n“m
m`k
ÿ
ď lim }fn }p
kÑ8
n“m
8
ÿ
“ }fn }p .
n“m
ř8
Since n“0 }fn }p ă 8, it follows that limmÑ8 }g ´ gm }p “ 0. ˝Claim 1
ř8
Claim 2: n“0 fn converges a.e..

Proof Claim 2: Since g P Lp pEq, gpxq ă 8 a.e.. Let x P E so that

gpxq ă 8. When k, m P N and m ą 0,
ˇ ˇ
ˇm`k ˇ
ˇÿ
fn pxqˇ ď gpxq ´ gm´1 pxq
ˇ
ˇ
ˇn“m ˇ

ř8
It follows
ř8 that the partial sums of n“0 fn pxq form a Cauchy sequence.
Thus, n“0 fn converges almost everywhere. ˝Claim 2 .
ř8
Set f “ n“0 .

Claim 3: f P Lp pEq.

Proof Claim 3: |f |p ď |g|p by definition. Apply Claim 2. ˝Claim 3 .

Claim 4: limmÑ8 }f ´ m
ř
n“0 fn }p “ 0.

Proof Claim 4: Again, by MCT we have that

m
ÿ
}f ´ fn }p ď }g ´ gm }p .
n“0

Apply Claim 2. ˝Claim 4

Claim 5: }f }p ď 8
ř
n“0 }fn }p .

Proof Claim 5: By Claim 4 and Lemma 22.

124 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Theorem 83. Lp pEq a Banach space

Proof. Suppose tfn u8 p

n“0 is a Cauchy sequence in L pEq. Choose
ř8 a sequence
´j
n0 ă n1 ă . . . so that }fnj`1 ´ fnj }p ď 2 . Set fř“ fn0 ` j“0 fnj`1 ´ fnj .
By Lemmař23, f P Lp pEq and limmÑ8 }f ´pfn0 ` m j“0 fnj q}p “ 0. However,
f ´pfn0 ` m f
j“0 nj q “ f ´f nj`1 . Thus lim jÑ8 }f ´f 8
nj }p “ 0. Since tfn un“0
is a Cauchy sequence, limnÑ8 }fn ´ f }p “ 0 (see Proposition 75).

7.5 Approximations and Separability

One major strategy in analysis is to try to approximate functions of one
class by those with better properties. This is a version of Littlewood’s
second principle, but it has many applications (e.g., the continuity method,
using test functions to define distributions, etc.). Here, we are able to make
this idea precise, using the notion of dense subsets.

Definition 61. Let X be a normed linear space with norm } ¨ }. Given two
subsets F and G of X with F Ď G, we say that F is dense in G, provided
for each function g in G and  ą 0, there is a function f in F for which
}f ´ g} ă .

Example 3. 1. The rational numbers Q are dense in the real numbers

R.

2. Weierstrauss’ theorem shows that the space of polynomials is dense

inside Cra, bs for any bounded interval.

There is an important principle, which is often useful:

Observation 7. If F is dense in G and G is dense in H, then F is dense

in H.

Proposition 84. Let E be a measurable set and 1 ď p ď 8. Then the

subspace of simple functions in Lp pEq is dense in Lp pEq.

Proof. Suppose that g P Lp pEq. First consider the case p “ 8. There is a

subset E0 of E of measure zero for which g is bounded on E „ E0 . From the
Simple Approximation Lemma, we find that there is a sequence of simple
functions on E „ E0 that converge uniformly on E „ E0 to g. Therefore,
these functions converge with respect to the L8 pEq norm. Thus the simple
functions are dense in L8 pEq.
7.5. APPROXIMATIONS AND SEPARABILITY 125

Now suppose 1 ď p ă 8. The function g is measurable and therefore,

by the Simple Approximation Theorem, there is a sequence tϕn u of simple
functions on E such that tϕn u Ñ g pointwise on E and

|ϕn | ď |g| on E for all n.

By comparison, we see that ϕn P Lp pEq for all n. Claim 1: tϕn u Ñ g in

Lp pEq. Proof of claim 1: For all n

|ϕn ´ g|p ď 2p t|ϕn |p ` |g|p u ď 2p`1 |g|p on E.

Since |g|p is integrable over E, we infer from the Lebesgue Dominated Con-
vergence Theorem that tϕn u Ñ g in Lp pEq.

Proposition 85. Let ra, bs be a closed, bounded interval and 1 ď p ă 8.

Then the subspace of step functions on ra, bs is dense in Lp ra, bs.

The proof of this is essentially the same as the proof of Proposition 54,
so we will omit the proof.

Definition 62. A topological space is separable if it admits a dense count-

able subset.

Example 4. The real numbers are separable (which is something you are
probably tired of hearing by this point in the course).

Proposition 86. Let E be a measurable subset. The space Lp pEq is sepa-

rable for 1 ď p ă 8

Proof. (Sketch) On the interval r´n, ns, consider the space Fn of step func-
tions which take rational values and whose jumps occur at rational numbers
and which are defined to be identically zero outside this interval. We con-
sider the union ď
F“ Fn .
n

By our previous work, F is a countable collection of functions that is

dense in Lp pRq. Finally, let E be a general measurable set. Then the
restrictions of functions in F to E is a countable dense subset of Lp pEq.

Corollary 21. Cc pEq is dense in Lp pEq

On the other hand L8 pRq is not separable (see the book for details.)
126 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

7.6 The duality of Lp

We have already seen that the conjugate exponents appear in pairs. In this
section, we will discuss this duality in a little more depth. We start with
some definitions.

7.6.1 Bounded and continuous functionals

Definition 63. Suppose V is a vector space. A linear functional on V is a
map T : V Ñ R such that for all g, h P V , α, β P R,

T pαg ` βhq “ αT pgq ` βT phq.

Note that the collection of linear functionals on V is itself a vector space.

We can give a few examples.
Example 5. 1. If V “ Rn , the space of linear functionals is simply the
space of linear functions.

2. Suppose E is measurable, 1 ď p ă 8 and p1 ` 1

q “ 1. For g P Lq pEq,
we define the functional T : Lp pEq Ñ R by
ż
T pf q “ f gdµ
E

for f P Lp pEq. Using Holder’s inequality, we know that |T f | ď

}f }p }g}q , so T pf q is well defined.
For normed spaces, we can also discuss the notion of a bounded linear
functional.
Definition 64. For a normed linear space X, a linear functional T is said
to be bounded provided there is an M ě 0 for which

|T pf q| ď M }f } for all f P X.

The infimum of the upper-bounds is known as the norm of T and is

denoted by }T }˚ .
Exercise 21. Show that }T }˚ “ sup tT f | f P X, }f } ď 1u “ sup tT f | f P X, }f } “ 1u
From this, we have the following.
Proposition 87. The function }¨}˚ is a norm, and the collection of bounded
linear functionals on V is a normed linear space, which is denoted V ˚ .
7.6. THE DUALITY OF LP 127

Proof. It is immediate that }λT }˚ “ λ}T }˚ and that }T }˚ “ 0 if and only if

T ” 0. What remains to show is that

}S ` T }˚ ď }S}˚ ` }T }˚ .

To see this, we use the following observation.

}S ` T }˚ “ sup tpS ` T qf | f P X, }f } “ 1u
ď sup tSf | f P X, }f } “ 1u ` sup tT f | f P X, }f } “ 1u
“ r}S}˚ ` r}T }˚ .

One immediate consequence of a linear functional being bounded is the

following inequality, which holds for any f, h P X :

|T pf q ´ T phq| ď }T }˚ }f ´ h}.

From this, it immediately follows that bounded linear functionals are

continuous. In other words, whenever a linear functional is bounded, it is
also continuous. At first, this might seem like a strange statement, because
if you are used to finite dimensional linear algebra, all linear functionals are
both continuous and bounded. As such, it’s important to keep in mind that
things are more complicated for function spaces, which are generally infinite
dimensional. Actually, a linear functional being bounded is equivalent to its
continuity.

Proposition 88. For a linear functional T in a normed linear space V , the

following conditions are equivalent.

1. T is bounded.

2. T is continuous.

3. T is continuous at one point of V.

Let us now focus on the case where V is Lp pEq, and try to understand
the dual space V ˚ .

Proposition 89. Let E be measurable and 1 ď p ă 8. Suppose that q is

conjugate to p and that g P Lq pEq. ş
Define Tg : Lp pEq Ñ R by Tg pf q “ E f gd mu.
Then Tg P pLp pEqq˚ and }T }˚ “ }g}q
128 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Proof. Hölder’s inequality shows that Tg P pLp pEqq˚ and that }T }˚ ď }g}q .
Therefore, the only thing to prove is the equality. In the case where
1 ă p ă 8, we take
|g|q´1 |
f “ psgnpgqq q´1
}g}q
to be the conjugate function to g. Doing so, we have that f P Lp and that
}f }p “ 1. Furthermore, we find that

|g|q |
ż
|Tg pf q| “ “ }g}q ,
}g}q´1
q

which implies that

}Tg }˚ ě }g}q .
We then turn to the case where p “ 1. For the sake of contradiction, we
assume that }Tg } ă }g}8 .
Then there is a set A with µpAq ą 0 for which |g| ą }Tg } on A. Then
χA
we take f “ sgnpgq µpAq . Then f P L1 and }f }1 “ 1. However,
ż ż
χA
Tg pf q “ f g “ |g| ą }Tg }˚ ,
µpAq
which is a contradiction.

Before we come to the main result of this section, let us state a small
lemma which will be helpful.

Lemma 24. Let E be a measurable set, 1 ď p ă 8 and we have a function

g which is integrable on E and satisfies the following property.
There exists and M ě 0 so that
ˇż ˇ
ˇ ˇ
ˇ gf ˇ ď M }f }p
ˇ ˇ
E

for every simple function f P Lp pEq.

Then g P Lq pEq and }g}q ď M.

Proof. (Sketch)

1. g is finite almost everywhere since it is integrable.

2. As such, we can use the simple approximation lemma to find a sequence

of simple functions φn so that 0 ď φn ď |g| and φn Ñ |g| as n Ñ 8.
7.6. THE DUALITY OF LP 129

3. Then we consider the function

fn “ sgnpgqφnq´1

and compute its integral against g.

4. Finally, we apply Fatou’s lemma.

We can now state and prove the main result of this section, which char-
acterizes the dual space of Lp (except when p is infinite). We start with the
case where the domain is an interval.

Theorem 90 (Riesz representation theorem). Let ra, bs be a closed, bounded

interval and 1 ď p ă 8. Suppose T is a bounded linear functional on
Lp ra, bs. Then there is a function g P Lq ra, bs, for which
ż
T pf q “ g ¨ f for all f in Lp ra, bs
I

[The following proof is taken verbatim from Royden-Fitzpatrick.]

Proof. We will only consider the case p ą 1 (the proof of the case p “ 1 is
similar).For x in ra, bs, we define

Φpxq “ T pχra, xq

We first show t this real-valued function Φ is absolutely continuous on ra, bs.

Indeed, by the linearity of T , for each rc, ds Ď ra, bs, since χrc,dq “ χra,dq ´
χra,cq , ` ˘ ` ˘ ` ˘
Φpdq ´ Φpcq “ T χra,dq ´ T χra,cq “ T χrc,dq
Thus if tpak , bk qunk“1 is a finite disjoint collection of intervals in pa, bq, by the
linearity of T , we have that
˜ ¸
ÿn ÿn ÿn
` ˘
|Φ pbk q ´ Φ pak q| “ sgn rΦ pbk q ´ Φ pak qs¨T χrak ,bk q “ T sgn rΦ pbk q ´ Φ pak qs ¨ χrak ,bk q
k“1 k“1 k“1
řn
Moreover, for the simple function f “ k“1 sgn rΦ pbk q ´ Φ pak qs¨χrak ,bk q ,
« ff1{p
n
ÿ
|T pf q| ď }T }˚ ¨ }f }p and }f }p “ pbk ´ ak q
k“1
130 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

Thus « ff1{p
n
ÿ n
ÿ
|Φ pbk q ´ Φ pak q| ď }T }˚ ¨ pbk ´ ak q
k“1 k“1

Therefore, for  ą 0, we take δ “ p{}T }˚ qp and see that Φ is absolutely

continuous on ra, bs.
The function g “ Φ1 is integrable over ra, bs and
żx
Φpxq “ g for all x P ra, bs
0

Therefore, for each rc, ds Ď pa, bq

żb
` ˘
T χrc,dq “ Φpdq ´ Φpcq “ g ¨ χrc,dq
a
şb
Since the functional T and the functional f ÞÑ a g ¨f are linear on the linear
space of step functions, it follows that
żb
T pf q “ g ¨ f for all step functions f on ra, bs
a

If f is a simple function on ra, bs, there is a sequence of step functions

tϕn u that converges to f in Lp ra, bs and also is uniformly pointwise bounded
on ra, bs. Since the linear functional T is bounded on Lp ra, bs, it follows from
the continuity property that

lim T pϕn q “ T pf q
nÑ8

On the other hand, the Lebesgue Dominated Convergence Theorem implies

that żb żb
lim g ¨ ϕn “ g¨f
nÑ8 a a
Therefore
żb
T pf q “ g ¨ f for all simple functions f on ra, bs
a

Since T is bounded,
ˇż b ˇ
ˇ ˇ
ˇ g ¨ f ˇ “ |T pf q| ď }T }˚ ¨ }f }p for all simple functions f on ra, bs.
ˇ ˇ
a
7.7. PROBLEMS 131

According to the previous lemma, g belongs to Lq ra, bs so the linear func-

şb
tional f ÞÑ a g ¨ f is bounded on Lp ra, bs. This functional agrees with the
bounded functional T on the simple functions, which is a dense subspace of
Lp ra, bs, so these two functionals agree on all of Lp ra, bs.

In fact, this result holds for general measurable sets, not just intervals.

Theorem 91 (The Riesz Representation Theorem for the Dual of L p pEq).

Let E be a measurable set, 1 ď p ă 8, and q the conjugate of p. For each
g P Lq pEq, define the bounded linear functional Rg on Lp pEq by
ż
Rg pf q “ g ¨ f for all f in Lp pEq
E

Then for each bounded linear functional T on Lp pEq, there is a unique func-
tion g P Lq pEq for which

Rg “ T, and }T }˚ “ }g}q

Observation 8. Remark Let ra, bs be a nondegenerate closed, bounded in-

terval. Hölder’s Inequality shows that if f belongs to L1 ra, bs, then the func-
şb
tional g ÞÑ a f ¨ g is a bounded linear functional on L8 ra, bs. However, there
are bounded linear functionals on L8 ra, bs that are not of this form.

In Chapter 19 Section 3 of the book, Royden-Fitzpatrick explain a result

of Kantorovich which details the dual of L8 . Feel free to look at this section
for details.

7.7 Problems
Lp spaces appear on nearly every qualifying exam, so we will spend some
time working on problems involving them.

1 1
1. August 2020 # 2 Let 1 ă p, q ă 8 be such that p ` q “ 1. For any
f P Lp pr0, 1sq, set
żx
gpxq “ f ptqdt.
0

Prove that g P Lq pr0, 1sq and }g}q ď 2´1{q }f }p .

132 CHAPTER 7. BASICS OF FUNCTIONAL ANALYSIS

2. January 2020 # 4 Let f, g P L1 pRq. Given n P N , define pTn f q “

f px ´ nq. Prove that

lim }Tn f ` g}1 “ }f }1 ` }g}1

nÑ8

3. Fall 2011 # 2 Suppose f is a measurable function on r0, 1s such that

for every 1 ď p ă 8, f P Lp r0, 1s, and suppose there exists a B such
that }f }p ď B. Prove that f P L8 r0, 1s.

4. Fall 2014 # 3 Let p ě 1 and pfn q be a sequence of measurable functions

in Lp r0, 1s such that limnÑ8 fn pxq “ f pxq almost everywhere. Show
that limnÑ8 }fn ´ f }p “ 0 if and only if limnÑ8 }fn }p “ }f }p

5. Fall 2016
ş1 # 1 Let f : r0, 1s Ñ R be a Lebesgue integrable function and
In “ 0 |f pxq|n dx.

(a) if }f }8 ą 1, show that limnÑ8 In is equal to 8.

(b) if }f }8 ď 1, show that limnÑ8 In exists and evaluate the limit.
Chapter 8

Fubini and Tonelli Theorems

in Rn

In this chapter, we discuss multivariate integration and determine when it

is possible to compute a multiple integral as an iterated integral, as you saw
in multivariate calculus. This material is covered in Chapter 20 of Royden-
Fitzpatrick. Unfortunately, it uses quite a few terms that we have not yet
defined, so it might be worthwhile to read [Bal08] for more details.

8.1 A crash course in more general measures and

integration
In order to discuss this, we need to introduce measurable spaces in a bit
more generality.

Definition 65. A measure space pX, M, µq is a set X with a σ-algebra M

of subsets of X and a measure µ, which is a function µ : M Ñ r0, 8s which
satisfies

1. µpHq “ 0

2. µ is countably additive.

Definition 66. A measure space is

1. finite if µpXq ă 8

2. σ-finite if X is the union of countably many sets, each of which has

finite measure.

133
134 CHAPTER 8. FUBINI AND TONELLI THEOREMS IN RN

3. complete if µpEq “ 0 implies that every subset of E is measurable.

Example 6. 1. The Lebesgue measure on r0, 1s is finite, σ-finite and
complete.

2. The Lebesgue measure on R is not finite, but is σ-finite and complete.

3. The Borel measure on R is σ-finite, but neither finite nor complete.

4. The counting measure on R is neither finite nor σ-finite, but is com-

plete.

8.1.1 Product measures

The goal now is define the Lebesgue measure on R2 (or Rn . To this end,
we will define the product measure from the Lebesgue measure on R. More
generally, given two measure spaces pX, A, µq and pY, B, νq, we can build a
product measure on the Cartesian product X ˆ Y .
Definition 67 (Measurable rectangles). If A Ď X and B Ď Y , we call AˆB
a rectangle. If A P A and B P B, we call A ˆ B a measurable rectangle.
We can define the measure of a measurable rectangle to be the product
of the measures of A and B.
Definition 68. If R “ A ˆ B is a measurable rectangle, we define the
measure of R to be λpRq “ µpAq ˆ νpBq.
Example 7. Let ra, bs Ă R and rc, ds Ă R, both induced with the Lebesgue
measure. What is λpra, bs ˆ rc, dsq?
To make sure that this definition is consistent, we’ll verify the following
lemma.
Lemma 25. Let tAk ˆ Bk u8 k“1 be a countable disjoint collection of measur-
able rectangles whose union also is a measurable rectangle A ˆ B. Then
8
ÿ
µpAq ˆ νpBq “ µ pAk q ˆ ν pBk q
k“1

Proof. Fix a point x P A. For each y P B, the point px, yq belongs to exactly
one Ak ˆ Bk . Therefore we have the following disjoint union:
ď
B“ Bk
tk|xPAk u
8.1. A CRASH COURSE IN MORE GENERAL MEASURES AND INTEGRATION135

Figure 8.1: A measurable rectangle as the union of countable measurable

rectangles

By the countable additivity of the measure ν,

ÿ
νpBq “ ν pBk q
tk|xPAk u

Rewriting this equality in terms of characteristic functions, we find the

following
8
ÿ
νpBq ¨ χA pxq “ ν pBk q ¨ χAk pxq for all x P A
k“1
Since each Ak is contained in A, this equality also clearly holds for x P XzA.
Therefore
ÿ8
νpBq ¨ χA “ ν pBk q ¨ χAk on X
k“1
By the Monotone Convergence Theorem,
ż 8 ż
ÿ 8
ÿ
µpAq ˆ νpBq “ νpBq ¨ χA dµ “ ν pBk q ¨ χAk dµ “ µ pAk q ˆ ν pBk q
X k“1 X k“1
136 CHAPTER 8. FUBINI AND TONELLI THEOREMS IN RN

Technically speaking, λ is actually a pre-measure, and we need to con-

sider its Caratheodory extension to define the measure. Getting into the
details of this would take several weeks, so we will not do so. However, the
Lebesgue measure is a complete outer measure, so there is a short-cut.

Definition 69. Given two measure spaces pX, A, µq and pY, B, νq and a set
E Ă X ˆ Y , we define the outer measure
8
ÿ
pµ ˆ νq˚ pEq “ inf λpRk q,
k“1

where the infimum is taken over all countable collections of measurable rect-
angles Rk which cover E.

Definition 70. A set E Ă X ˆ Y is measurable if

pµ ˆ νq˚ pSq “ pµ ˆ νq˚ pE X Sq ` pµ ˆ νq˚ pE c X Sq

for all sets S.

Combining these, we can define the product measure.

Definition 71. The space pX ˆY, measurable sets, µˆνq is a measure space,
and the measure is known as the product measure.

8.1.2 Some measurable sets

It is helpful to know a good collection of measurable sets in Rn , since our
definition of measurable makes it hard to construct them by hand. For this,
we have the following proposition.

Proposition 92. Any open subset U of R2 is measurable with respect to the

product measure.

In fact, open sets are Borel (i.e., the intersection/union/complements of

countably many measurable rectangles. However, it will not be necessary to
prove this.
As with the case in R, there are many Lebesgue measurable sets which
are not Borel. As a note of caution, measurable sets in higher dimensions
can be very complicated, so it is worthwhile to verify that you are sets are
measurable.
8.2. INTEGRATION WITH RESPECT TO A GENERAL MEASURE137

8.2 Integration with respect to a general measure

Before we can discuss the Fubini and Tonelli theorems, we need to say what
it means to integrate with respect to a more general measure. The way that
this is done is essentially the same as in the one-dimensional case, by using
simple functions.
Definition 72 (Simple functions). Given a measure space pX, A, µq, a sim-
ple function s : X Ñ R is a function whose range is finite and for which the
inverse image
s´1 paq “ tx P X | f pxq “ au
is a measurable set (i.e., in A) for all a P R.
Definition 73. For non-negative simple functions, we define the integral
ż ÿ
s dµ “ aµps´1 paqq.
X aP rangepsq

Then, given a function f : X Ñ R` which is non-negative and measur-

able (i.e., the pre-image of every open set is in A), we define the integral of
f as ż "ż *
f dµ “ sup s dµ | s simple and s ď f .
X X
Finally, we say that a function is integrable if both its positive and
negative parts are integrable.

8.3 Product measures and iterated integrals

We are now able to discuss the Fubini and Tonelli theorems. Before stating
these theorems, let us give one further definition.
If f : x ˆ y Ñ R is a measurable function, for x P X, we define the
function f px, q : Y Ñ R to be the x-slice of f , For y P Y , we say that the
function f p¨, yq : X Ñ R is the y-slice of f .
Theorem 93 (Fubini’s Theorem). Let pX, A, µq and pY, B, νq be two mea-
sure spaces and ν be complete. Let f be integrable1 over X ˆ Y with respect
to the product measure µ ˆ ν. Then for almost all x P X, the x-slice of
f, f px, ¨q, is integrable over Y with respect to ν and
ż ż „ż 
f dpµ ˆ νq “ f px, yqdνpyq dµpxq.
1
ş
This means that f is measurable and that XˆY
|f | dpµ ˆ νq ă 8.
138 CHAPTER 8. FUBINI AND TONELLI THEOREMS IN RN

Theorem 94 (Tonelli’s theorem). let pX, A, µq and pY, B, νq be σ-finite mea-

sure spaces. Let f be a measurable function on X ˆY with respect to pµˆνq.
If 0 ď f ď 8, then

1. for almost all x P X, the x-slice is ν-measurable,

2. for almost all y P X, the y-slice is µ-measurable,

ş
3. the function Y f p¨, yq dνpyq is µ-measurable,
ş
4. the function X f px, ¨q dµpxq is ν-measurable.

5.
ż ż „ż  ż „ż 
f dpµˆνq “ f px, yqdνpyq dµpxq “ f px, yqdµpxq dνpyq
XˆY X Y Y X

Before proving these results, let us provide two examples which show
why the assumptions in these theorems are necessary.
ş1 ”ş1 2 ´y2 ı
Non-Example. 1. Prove 0 0 pxx2 `y 2 q2
dy dx “ ` π4

ş1 ”ş1 x2 ´y 2
ı
2. Prove 0 0 px2 `y 2 q2 dx dy “ ´ π4

3. Explain why the answer to the above parts do not violate Fubini’s the-
orem.
2 2
Proof. The function f “ pxx2 `y
´y
2 q2
is continuous on r0, 1s ˆ r0, 1s except at
the origin, hence is measurable.
Doing the first integral, we find that
ż1 1
x2 ´ y 2
„
y 1
2 2 2 dy “ x2 ` y 2 “
1 ` x2
0 px ` y q y“0

and ż 1 ˆż 1 ż1
x2 ´ y 2
˙
1 π
dy dx “ dx “
0 0 px2 ` y 2 q2 0 1`x2 4
Proceeding to the integral computation in the opposite order we get
ż 1 ˆż 1
x2 ´ y 2
˙
π
dx dy “ ´ .
0 0 px2 ` y 2 q2 4
8.3. PRODUCT MEASURES AND ITERATED INTEGRALS 139

ş To understand why this doesn’t contradict Fubini-Tonelli, we estimate

IˆI |f px, yq|dpx, yq. We have

ˇ 2 ż 1 ˜ż π{2 2 2 pθq ´ r 2 sin2 pθq|

¸
ˇ x ´ y2 ˇ
ż ˇ
|r cos
ˇ px2 ` y 2 q2 ˇ dpx, yq ě rdθ dr
ˇ ˇ
IˆI 0 0 r4
ż 1 ż π{2
| cosp2θq|
“ rdθdr
0 0 r2
ż 1 ż π{4
cosp2θq
“ 2 dθdr
0 0 r
1 11
ż
“2 dθdr
2 0 r
“ lim r Ñ 0 ´ lnprq “ 8.

As such, f R L1 pI ˆ Iq, so this does not contradict Tonelli’s theorem.

It’s also good to have an example in mind of why it is necessary to

assume that the measures are σ-finite. For this, we consider the following
example, which is taken from Axler [Axl20] 5.30.

Example 8. Suppose B is the σ-algebra of Borel subsets of r0, 1s, λ is

Lebesgue measure on pr0, 1s, Bq, and µ is counting measure on pr0, 1s, Bq.
Let D denote the diagonal of r0, 1s ˆ r0, 1s; in other words,

D “ tpx, xq : x P r0, 1su.

Then ż ż ż
χD px, yqdµpyqdλpxq “ 1dλ “ 1,
r0,1s r0,1s r0,1s

but ż ż ż
χD px, yqdλpxqdµpyq “ 0dµ “ 0.
r0,1s r0,1s r0,1s

The proof of Fubini and Tonelli’s theorem in Royden requires a fair

amount of background, so I recommend reading the proof in Axler [Axl20],
where it is Theorem 5.28.
140 CHAPTER 8. FUBINI AND TONELLI THEOREMS IN RN
Bibliography

[Axl20] Sheldon Axler. Measure, integration & real analysis. Springer

Nature, 2020.

[Bal08] Prakash Balachandran. 241 lecture: Product measures - Fubini

and Tonelli, 2008.

[Fol99] Gerald B Folland. Real analysis: modern techniques and their

applications, volume 40. John Wiley & Sons, 1999.

[FW91] Matthew Foreman and Friedrich Wehrung. The hahn-banach the-

orem implies the existence of a non-lebesgue measurable set. Fun-
damenta Mathematicae, 138:13–19, 1991.

[Mun14] James Munkres. Topology. Pearson Education, 2014.

[RF88] Halsey Lawrence Royden and Patrick Fitzpatrick. Real analysis,

volume 32. Macmillan New York, 1988.

[Tao11] Terence Tao. An introduction to measure theory, volume 126.

American Mathematical Society Providence, RI, 2011.

[Tes98] Gerald Teschl. Topics in real analysis. Amer. Math. Soc., Provi-
dence, to appear, 1998.

[wh] wisefool (https://math.stackexchange.com/users/51807/wisefool).

Simple function approximation of a function in lp . Mathematics
Stack Exchange. URL:https://math.stackexchange.com/q/256589
(version: 2012-12-11).

141