Lecture Notes in Real Analysis

Lewis Bowen∗
University of Texas at Austin
December 8, 2014

Contents
1 Outer measure and measurable sets 3

2 Measures and measurable sets 4

3 Cantor sets and the Cantor-Lebesgue function 5

4 Measurable functions 5

5 Borel functions (tangential and optional) 7

6 Semi-continuity (tangential) 8

7 Littlewood’s 3 principles 8
7.1 An aside . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

8 Convergence in measure 11

9 Integration for bounded functions 12

10 Integration for nonnegative functions 15

11 Integrable functions 16

12 Convergence Theorems 17

13 Riemannian integration 20
∗
supported in part by NSF grant DMS-0968762, NSF CAREER Award DMS-0954606 and BSF grant
2008274

1
14 Differentiation 21

15 Differentiation of monotone functions 25

15.1 Functions of Bounded Variation . . . . . . . . . . . . . . . . . . . . . . . . . 28

16 Absolutely continuous functions 29

17 Convex functions 31

18 Lp spaces 33
18.1 Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

19 Hilbert spaces 37

20 Signed measures 39

21 Radon-Nikodym Theorem 41

22 The Dual of Lp 43

23 Outer measure 45

24 Carathéodory’s Extension Theorem 47

25 Product measures 50

26 Fubini’s Theorem 52

27 Convolution 54
27.1 Norm inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
27.2 Smoothness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
27.3 Approximate identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

28 The space of measures 59

28.1 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
28.2 Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

29 Fourier series 63
29.1 Definition and convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
29.2 L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
29.3 Derivatives and absolute convergence . . . . . . . . . . . . . . . . . . . . . . 68
29.4 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

2
1 Outer measure and measurable sets
Definition 1. A subset O ⊂ R is open if for every x ∈ O there is an  > 0 such that the
interval (x − , x + ) ⊂ O. A set F ⊂ R is closed if its complement is open.

Exercise 1. For every open subset O ⊂ R there is a finite or countable collection {Ii } of
pairwise disjoint open intervals such that O = ∪i Ii .

Definition 2. The outer measure of a subset E ⊂ R is defined by m∗ (E) = inf C (x,y)∈C |x−
P

y| where the infimum is over all subsets C ⊂ R × R satisfying

• (x, y) ∈ C ⇒ x < y

• E ⊂ ∪(x,y)∈C (x, y).

We are using (x, y) to mean two different things: (x, y) is either an element of R2 or is an
open interval of R. The context should make clear which meaning is meant.

Exercise 2. The outer measure of an interval is its length.

Exercise 3. For any subsets E1 , E2 , . . . , ⊂ R, m∗ (∪i Ei ) ≤ i m∗ (Ei ). This means that outer
P
measure is countably sub-additive.
Observation 1. If E1 ⊂ E2 then m∗ (E1 ) ≤ m∗ (E2 ).
Exercise 4. For any E ⊂ R and any  > 0 there exists an open set O ⊃ E such that
m∗ (O) < m∗ (E) + .

Definition 3. We say that a subset E ⊂ R is measurable if for every  > 0 there exists an
open set O ⊃ E with m∗ (O \ E) < .

Observation 2. Open sets are measurable. If m∗ (E) = 0 then E is measurable.

Exercise 5. If E1 , E2 , . . . are measurable then ∪i Ei is also measurable.
Exercise 6. Compact sets are measurable.
Exercise 7. Closed sets are measurable.
Exercise 8. E is measurable if and only if R \ E is measurable.
Exercise 9. If E1 , E2 , . . . are measurable then ∩i Ei is also measurable.

Definition 4. Let X be a set and C a collection of subsets of X. We say C is a σ-algebra if

it is nonempty and for every E1 , E2 , . . . ∈ C, X \ Ei ∈ C (for all i, ∪i Ei ∈ C and ∩i Ei ∈ C.

Observation 3. The measurable subsets of R form a σ-algebra.

Definition 5. The Borel σ-algebra is the smallest sigma-algebra containing all of the open
sets. A set is Borel if it is in the Borel sigma-algebra. Note that all Borel sets are measurable.

3
2 Measures and measurable sets
Exercise 10. If E ⊂ R is measurable then for every  > 0 there exists a closed set F ⊂ E
with m∗ (E \ F ) < .
Proof. Because E is meas., its complement E c is also meas. So if  > 0 then there exists
an open set O ⊃ E c with m∗ (O \ E c ) < . Now Oc is a closed set, Oc ⊂ E and E \ Oc =
E ∩ O = O \ E c . So m∗ (E \ Oc ) < .
Exercise 11. If E1 , E2 , . . . ⊂ R are measurable and pairwise disjoint then
X
m∗ (∪i Ei ) = m∗ (Ei ).
i

Definition 6. Let X be a set and C a σ-algebra on X. We say that (X, C) is a measurable

space (or Borel space). Also let µ : C → [0, ∞] be a function satisfying: if E1 , E2 , . . . ∈ C are
pairwise disjoint then X
µ(∪i Ei ) = µ(Ei ).
i

Then µ is a measure on (X, C) and (X, C, µ) is a measure space. Often we omit C from the
notation and just say “µ is a measure on X”.

Observation 4. m∗ is a measure on (R, M) where M denotes the collection of measurable

subsets. From now on, we let m denote the restriction of m∗ to M. This is called Lebesgue
measure on R.
Exercise 12. Let (X, C, µ) be a measure space. Suppose E1 ⊂ E2 ⊂ · · · ∈ C and F1 ⊃ F2 ⊃
· · · ∈ C. Then
lim µ(Ei ) = µ(∪i Ei ).
i

If µ(F1 ) < ∞ then limi µ(Fi ) = µ(∩i Fi ).

Definition 7. A countable intersection of open subsets is called a set of type Gδ . A countable

union of closed subsets is called a set of type Fσ .

Example 1. The irrational numbers are a dense Gδ subset of the real line.
Exercise 13. Let E ⊂ R. Prove that the following are equivalent.

1. E is measurable.

2. for every  > 0 there exists an open set O ⊃ E with m∗ (O \ E) < ;

3. there exists a set G ⊃ E of type Gδ such that m∗ (G \ E) = 0;

4. for every  > 0 there exists a closed set F ⊃ E with m∗ (E \ F ) < ;

5. there exists a set F ⊃ E of type Fσ such that m∗ (E \ F ) = 0;

4
 6. there exists a Borel set B such that m(E M B) = 0 where E M B = (E \ B) ∪ (B \ E)
is the symmetric difference of B and E;

7. for every set A ⊂ R,

m∗ (A) = m∗ (A ∩ E) + m∗ (A \ E).

3 Cantor sets and the Cantor-Lebesgue function

...

4 Measurable functions
Definition 8. Let X, Y be topological spaces. We will assume X is endowed with a sigma-
algebra so that we can meaningfully discuss measurable subsets of X. A function f : X → Y
is

• continuous if for every open O ⊂ Y , f −1 (O) is open;

• measurable if for every open O ⊂ Y , f −1 (O) is measurable.

Observation 5. Every continuous function is measurable.

We will concern ourselves with measurable functions into the extended reals R∪{−∞, +∞}.
Exercise 14. Let f : X → R ∪ {±∞} be a function. TFAE

1. f is measurable;

2. for every a ∈ R, f −1 (a, +∞] is measurable;

3. for every a ∈ R, f −1 [a, +∞] is measurable;

4. for every Borel subset B ⊂ R ∪ {±∞}, f −1 (B) is measurable.

Proof. Clearly (1) ⇒ (2). So assume (2). By taking complements, we see f −1 [−∞, a] is
measurable ∀a. Since
f −1 ([−∞, a)) = ∪r∈Q,r<a f −1 [−∞, r]
it follows that f −1 ([−∞, a)) is measurable. By taking complements again we see that
f −1 [a, +∞] is measurable. So (2) ⇒ (3).
Now assume (3). We will prove (1). Since every open subset is a countable union of open
intervals, it suffices to show that f −1 (I) is measurable whenever I is an open interval. By
taking complements we see that this is true if I = [−∞, a) for some a. Because f −1 (a, +∞] =
∪r∈Q,r>a f −1 [r, +∞], it is also true if I = f −1 (a, +∞] for some a. So it’s true whenever
I is an infinite interval. If I is finite then I = (a, b) for some a, b ∈ R in which case
f −1 (I) = f −1 (a, +∞] ∩ f −1 [−∞, b). So it’s true in this case too.

5
 We have now shown that 1,2,3 are equivalent. Since open sets are Borel, (4) implies
(1). To see that (1) implies (4), it suffices to observe that the collection C of all subsets
A ⊂ R ∪ {±∞} such that f −1 (A) is measurable forms a sigma-algebra. This is because for
any sets E1 , E2 , . . . ,
f −1 (∪i EI ) = ∪i f −1 (Ei ), f −1 (∩i EI ) = ∩i f −1 (Ei ), f −1 (E c ) = f −1 (E)c .

Exercise 15. Let f, g, b : X → R be functions. Suppose f, g are measurable and b is contin-

uous. Then
• f + g is measurable
• f g is measurable
• f /g is measurable if g is never 0.
• b ◦ f is measurable
Proof. Let a ∈ R. It suffices to show {x ∈ R : f (x) + g(x) > a} is measurable. This follows
from
[
{x ∈ X : f (x) + g(x) > a} = {x ∈ X : f (x) > q} ∩ {x ∈ X : g(x) > a − q}.
q∈Q

Similarly, if a ≥ 0 then
[
{x ∈ X : f (x)g(x) > a} = {x ∈ X : f (x) > q} ∩ {x ∈ X : g(x) > a/q}
q∈Q,q>0
[
∪ {x ∈ X : f (x) < q} ∩ {x ∈ X : g(x) < a/q}.
q∈Q,q<0

The case a < 0 is similar. This shows the first two items. The third one is similar.
To show b ◦ f is measurable, suppose O ⊂ R is open. Then b−1 (O) is open. So f −1 (O)
is measurable.
Exercise 16. Suppose fi : X → R for i = 1, 2, . . . are measurable functions. Then supi fi , inf i fi , lim supi fi
and lim inf i fi are measurable.
Proof. For any a ∈ R,
{x ∈ X : sup fi (x) > a} = ∩n ∪i {x ∈ X : fi (x) > a − 1/n}.
i

This shows supi fi is measurable. inf i fi is similar.

{x ∈ X : lim sup fi (x) ≥ a} = ∩i {x ∈ X : sup fj (x) ≥ a}.
i j>i

This shows lim supi fi is measurable. lim inf i fi similar.

6
Definition 9. Let f, g : X → R be functions. We say that f = g almost everywhere (a.e.)
if {x ∈ X : f (x) 6= g(x)} has measure zero.
Exercise 17. Suppose f = g a.e. If f is measurable then so is g.
Proof. Let Z = {x ∈ X : f (x) 6= g(x)} and let O ⊂ R ∪ {±∞} be open. Then g −1 (O) \ Z =
f −1 (O) \ Z. So there exists a subset Z 0 ⊂ Z such that g −1 (O) = Z 0 ∪ (f −1 (O) \ Z). Because
Z, Z 0 have measure zero, they are both measurable. Thus g −1 (O) is measurable. Since O is
arbitrary, this proves g is measurable.
Definition 10. The characteristic function or indicator function of a set E ⊂ X is the
function χE : X → R given by χE (x) = 1 if x ∈ E and χE (x) = 0 otherwise. A simple
function is a finite linear combination of characteristic functions of measurable subsets.
Exercise 18. Let f : X → R be a function. TFAE
1. f is simple

2. f is measurable and the range of f is finite

3. there exist disjoint measurable sets E1 , . . . , Ek and real numbers r1 , . . . , rk such that
X
f (x) = ri χEi .
i

Later on, we will use simple functions to develop integration theory. It will be useful to
have the following approximation result:
Exercise 19. Let f be a measurable function on X. Then there exist simple functions
{fi }i such that f = limi fi pointwise. Moreover, if f ≥ 0 then we can choose fi so that
0 ≤ f1 ≤ f2 ≤ · · · ≤ f .
Proof. Define fn by

 −n f (x) < −n
−n
f (x) = k2 k2−n ≤ f (x) < (k + 1)2−n
n + 1 f (x) ≥ n + 1


5 Borel functions (tangential and optional)

Definition 11. Let X, Y be topological spaces. A function f : X → Y is Borel if for every
open O ⊂ Y , f −1 (O) is Borel.
Observation 6. Every continuous function is Borel and every Borel function is measurable
(as long as Borel sets are measurable which is usually the case).

7
Exercise 20. A function f : X → Y is Borel if and only if for every Borel set B ⊂ Y , f −1 (Y )
is Borel. Hence compositions of Borel functions are Borel.
Exercise 21. There is a Borel function f : [0, 1] → [0, 1] and a measurable set X ⊂ [0, 1] such
that f −1 (X) is not measurable.
Exercise 22. There are measurable functions f, g : R → R whose composition is not mea-
surable.

6 Semi-continuity (tangential)
We won’t use this much but it is good to have in your vocabulary:

Definition 12. f : X → R ∪ {±∞} is upper semi-continuous if f −1 [−∞, a) is open for every

a ∈ R (where X is a topological space). f is lower semi-continuous if f −1 (a, +∞] is open
for every a ∈ R.

Exercise 23. Let X be a compact metric space. Then f : X → R ∪ {±∞} is upper semi-
continuous if and only if there exist continuous functions f1 , f2 , . . . such that f = inf i fi .
Similarly, f is lower semi-continuous if and only if there exist continuous functions f1 , f2 , . . .
such that f = supi fi .

7 Littlewood’s 3 principles
Littlewood’s three principles are:

1. Every subset of the real line of finite measure is nearly a finite union of intervals.

2. Every measurable function is nearly continuous.

3. Every convergent sequence of functions is nearly uniformly convergent.

Let us make this rigorous:

Exercise 24 (First principle). Suppose E ⊂ R is measurable and has finite measure. Prove:
for every  > 0 there exists a finite union of open intervals O such that m(O M E) < .
Proof. There exists an open set O ⊃ E with m(O \ E) < . Since O is a countable union of
intervals, this means there is a finite union of intervals O0 ⊂ O with m(O \O0 ) < . Therefore
m(O0 M E) < 2.
We will prove:

Theorem 7.1 (Second principle: Lusin’s Theorem). Suppose f : R → R measurable func-

tion. For every  > 0 there exists a continuous function g such that

m({x ∈ R : f (x) 6= g(x)}) < .

8
Theorem 7.2 (Third principle: Egorov’s Theorem). Suppose {fn } is a sequence of measur-
able functions defined on a set E ⊂ R of finite measure. Suppose also that fn → f pointwise
a.e. For every  > 0 there exists a subset B ⊂ E such that m(E \ B) <  and fn converges
to f uniformly on E \ B.

We will prove Egorov’s Theorem first. So suppose {fn }, f, E are as in Egorov’s Theorem.
Exercise 25. For every δ > 0 there exists a measurable set B ⊂ E and an integer N > 0
such that for every n > N ,

m({x ∈ E : |fn (x) − f (x)| ≥ δ}) ≤ δ.

Hint: let
Gn = {x ∈ E : |fn (x) − f (x)| ≥ δ}
and set EN = ∪∞
n=N Gn . Prove limN →∞ m(EN ) = 0 and derive the result from this.

Proof. Note that EN ⊃ EN +1 ⊃ · · · is a decreasing sequence. So

m(∩N EN ) = lim m(EN ).

N →∞

Suppose x ∈ ∩N EN . Then lim supN →∞ |fn (x) − f (x)| ≥ δ. However, the latter is satisfied
only on a measure zero set. So

0 = m(∩N EN ) = lim m(EN ).

N →∞

Choose N large enough so that m(EN ) ≤ δ. Because

{x ∈ E : |fn (x) − f (x)| ≥ δ} ⊂ EN

(for n ≥ N ) we’re done.

Exercise 26. Prove Egorov’s Theorem. Hint: apply the previous exercise repeatedly with
δn = 2−n .
Proof. By the previous exercise, there exist measurable sets Bn and integers Nn such that

• m(Bn ) < 2−n 

• for every k > Nn and x ∈ E \ Bn ,

|fk (x) − f (x)| < 2−n .

Let B∞ = ∪n Bn . So m(B∞ ) <  and if x ∈ E \ B∞ and k > Nn then

|fk (x) − f (x)| < 2−n .

This proves {fk } converges uniformly to f on E \ B∞ .

9
 We now start with the proof of Lusin’s theorem.
Exercise 27. If f : E → R is simple and  > 0 then there exists a closed set F ⊂ E such
that m(E \ F ) <  and f  F is continuous.
Proof. Because f is simple f = ni=1 ci χEi for some coefficients ci and some pairwise disjoint
P
measurable sets Ei . Let Fi ⊂ Ei be closed sets with m(Ei \ Fi ) < /n. Let F = ∪Fi . Note
m(E \ F ) <  and f restricted to F is continuous.
Exercise 28. If f : R → R is measurable and  > 0 then there exists a closed set F ⊂ R such
that m(R \ F ) <  and f  F is continuous.
Proof. Let us first assume that m(E) < ∞.
Let {fk } be a sequence of simple functions that converge pointwise to f . For each k, let
Fk ⊂ E be a closed set with m(R \ Fk ) < /2k such that fk  Fk is continuous. By Egorov’s
Theorem there exists B ⊂ R such that m(B) <  and {fk } converges uniformly to f on
R \ B. Wlog, we may require B to be open so that F0 := R \ B is closed. (if B is not open
we may replace it with an open set O ⊃ B such that m(O) < ).
Because
T a uniformly convergent sequence of continuous functions is continuous, f  F∞ :=
F0 ∩ k Fk is continuous. Note
∞
X
m(E \ F∞ ) ≤ m(E \ Fk ) < 2.
k=0

So this handles the case m(E) < ∞. To obtain the general case, we find finite measure
sets E0 ⊂ E1 ⊂ E2 ⊂ · · · ⊂ R such that R = ∪Ei . Then for each i there exists a closed set
Fi ⊂ Ei with m(Ei \ Fi ) < /2i such that f  Fi is continuous. Note that m(R \ ∪i Fi ) < . So
there exists finite I such that m(R\∪Ii=0 Fi ) < . Observe that f is continuous on ∪Ii=0 Fi .
Exercise 29. Prove Tietze’s Extension Theorem: if F ⊂ R is any closed set and f : F → R
any continuous function then there exists a continuous function g : R → R whose restriction
to F equals f .
Proof. There exist pairwise disjoint open intervals O1 , O2 , . . . such that the complement of
F is the union ∪i Oi . Write Oi = (ai , bi ). Now define g by g(x) = f (x) if x ∈ F and
   
bi − x x − ai
g(x) = f (ai ) + f (bi )
b i − ai b i − ai

if x ∈ (ai , bi ).
Remark 1. Tietze’s Extension Theorem holds in much greater generality: you can allow F to
be a closed subset of any normal topological space. (Normal means that every two disjoint
closed subsets have disjoint open neighborhoods. For example, metric spaces are normal).
Lusin’s Theorem is an immediate consequence of the previous two exercises.
Lusin’s Theorem can be generalized a great deal. To explain we need:

10
Definition 13. A measure m on a topological space X is regular if for every measurable set
E ⊂ X, every  > 0 there exists an open set O ⊃ E and a compact set F ⊂ E such that

m(O \ E) < , m(E \ F ) < .

For example, Lebesgue measure is regular.

Regular measures are common:

Theorem 7.3. Every Borel measure on a locally compact secound countable space is regular.

(We won’t prove this here).

Exercise 30 (Lusin’s Theorem). Suppose X is a normal Hausdorff topological space and m
is a σ-finite regular measure on X. Then for any measurable function f : X → R and  > 0
there exists a continuous g : X → R such that

m({x ∈ X : f (x) 6= g(x)}) < .

7.1 An aside
It can be shown that a measurable function f : [a, b] → R is Riemann-integrable if and only
if the set of all x ∈ [a, b] such that f is discontinuous at x has measure zero.
Example: Let f = χQ . Claim: f is discontinuous everywhere.
Proof: For any x ∈ R and any δ > 0 we have that (x − δ, x + δ) ∩ Q 6= ∅ and (x −
δ, x + δ) ∩ Qc 6= ∅. So there y, z with |x − y| < δ such that y ∈ Q and z ∈ / Q. Then either
|f (x) − f (y)| = 1 or |f (x) − f (z)| = 1. In any case, f is discontinuous at x.

8 Convergence in measure
Definition 14. Let f1 , f2 , . . . , f∞ be measurable functions all defined on some measurable
subset E ⊂ R. We say

• {fn } converges pointwise to f∞ if f∞ (x) = limn fn (x) for every x ∈ E;

• {fn } converges pointwise a.e. to f∞ if f∞ (x) = limn fn (x) for a.e. x ∈ E;

• {fn } converges in measure to f∞ if for every  > 0,

lim m({x ∈ X : |fn (x) − f∞ (x)| > }) = 0.

n

Exercise 31. Suppose E has finite measure and fn → f∞ on E pointwise a.e. Show that
fn → f∞ in measure.
Proof. This follows from Egorov’s Theorem. Note: it is necessary that E have finite measure.

11
Exercise 32. Construct measurable functions f1 , f2 , . . . , f∞ : [0, 1] → R such that fn → f∞
in measure but fn does not converge pointwise a.e.
Proof. Define fi = χEi where Ei ⊂ [0, 1] is measurable, m(Ei ) → 0 and lim supi Ei =
[0, 1].
Exercise 33. Suppose fn → f∞ in measure. Show there exists a subsequence {fni } such that
fni → f∞ pointwise a.e.
Proof. Let ni be large enough so that if Ei = {x ∈ X : |fni (x) − f (x)| > 1/i} then

m(Ei ) ≤ 2−i .

Let Z = {x ∈ X : lim supi |fni (x) − f (x)| > 0} = lim supi Ei = ∩∞

j=1 ∪i>j Ei . By continuity
m(Z) = 0. So fni → f∞ pointwise a.e.

9 Integration for bounded functions

We will assume throughout that (X, C, m) is a measure space.

Definition 15. Let f = ni=1 ci χEi be a simple function with finite measure support. As-
P
sume that the sets E1 , . . . , En are pairwise disjoint and ci 6= cj if i 6= j. (This property
determines the Ei ’s and ci ’s from f uniquely up to pemuting the indices). Then we define
Z n
X
f dm = ci m(Ei ).
i=1

In other words Z X
f dm = cm(f −1 (c)).
c∈R

Exercise 34. Suppose f, g are simple functions (each with finite measure support) and a, b ∈
R. Then af + bg is simple and
Z Z Z
af + bg dm = a f dm + b g dm.
R R R R R
Proof. It’s clear that af dm = a f dm so it suffies to prove f +g dm = f dm+ g dm.

12
We observe
Z X
f + g dm = cm({x ∈ X : f (x) + g(x) = c})
c∈R
X X
= c m({x ∈ X : f (x) = b, g(x) = c − b})
c∈R b∈R
XX
= (b + c)m({x ∈ X : f (x) = b, g(x) = c})
b∈R c∈R
XX XX
= bm({x ∈ X : f (x) = b, g(x) = c}) + cm({x ∈ X : f (x) = b, g(x) = c})
b∈R c∈R b∈R c∈R
X X
= bm({x ∈ X : f (x) = b}) + cm({x ∈ X : g(x) = c})
b∈R c∈R
Z Z
= f dm + g dm.

Exercise 35. Suppose E ⊂ X has finite measure and f : E → R is a bounded measurable

function. Z Z
sup φ dm = inf ψ dm
φ≤f f ≤ψ

where the sup and inf are over simple functions from E to R.
Proof. It’s easy to see that ≤ must occur. Indeed, if φ ≤ f ≤ ψ are as above and
n
X m
X
φ= ci χEi , ψ= di χFi
i=1 i=1

then after refining the partition {E1 , . . . , En }, {F1 , . . . , Fm } we may assume that they are
equal. That is, we may assume that
k
X k
X
φ= ci χGi , ψ= di χGi
i=1 i=1

for some measuable sets G1 , . . . , Gk . Since ci ≤ di , the inequality ≤ follows.

Because f is bounded there is some M > 0 such that |f (x)| ≤ M for all x. For n > 0
and k ∈ Z, |k| ≤ 2n let
En,k = {x ∈ X : k2−n ≤ f (x)/M < (k + 1)2−n }.
Let φn (x) = M k2−n and ψn (x) = M (k + 1)2−n on En,k . These are both simple functions
and by design |φn − ψn | ≤ M 2−n . So
Z Z X
φn dm − ψn dm ≤ M 2−n m(En,k ) = M 2−n m(E).
k

Since this tends to 0 as n → ∞, it proves the exercise.

13
Definition 16. With f, X as above we define
Z Z Z
f dm = sup φ dm = inf ψ dm.
E φ≤f f ≤ψ

If Y ⊂ X is measurable then Z Z
f dm = f χY dm.
Y

If [a, b] ⊂ X ⊂ R is an interval and m is Lebesgue measure then

Z Z b
f dm = f (x) dx.
[a,b] a

If X = E or Rif E is understood from the context then we usually drop the subscript and
simply write f dm.

Exercise 36. Suppose E ⊂ X has finite measure and f, g : E → R are bounded measurable
functions and a, b ∈ R. Then
R R R
1. (linearity) E af + bg dm = a f dm + b g dm,
R R R
2. R(monotonicity) if f ≤ g a.e. then f dm ≤ g dm. So f = g a.e. then f dm =
g dm.
R
3. If A ≤ f ≤ B then Am(E) ≤ f dm ≤ Bm(E),

P R if A1 , A2 , . . . , An ⊂ X are disjoint, measurable and ∪i Ai = E then

4. R(finitely additive)
f dm = i Ai f dm.

Proof. Suppose that a, b > 0. If φ1 ≤ f ≤ ψ1 and φ2 ≤ g ≤ ψ2 are simple functions then

aφ1 + bφ2 ≤ af + bg ≤ aψ1 + bψ2

are simple functions and

Z Z Z
aφ1 + bφ2 dm = a φ1 dm + b φ2 dm
E

and a similar formula holds with ψi in place of φi . Using the definition of the integral, this
implies (1).
R To prove R (2) we simply notice that if φ ≤ f is a simple function since φ ≤ g we also have
φ dm ≤ g dm by definition. So taking the sup over all φ ≤ f we obtain (2).
Item (3) follows from item (4) by letting
P f (or g) be a constant.
To see item (4), observe that f = i f χAi . So (4) follows from (1).

14
10 Integration for nonnegative functions
Definition 17. Let f : X → [0, ∞] be measurable (where (X, C, m) is a measure space).
We define Z Z
f dm = sup g dm
g≤f

where the sup is over all bounded measurable functions g ≤ f with finite measure support.
Exercise 37. Suppose f, g : X → [0, ∞] are measurable and c > 0 is a constant. Then
R R
1. cf dm = c f dm
R R R
2. f + g dm = f dm + g dm.
R R
3. f ≤ g ⇒ f dm ≤ g dm.
4. if X = A ∪ B where A, B are disjoint measurable sets, then
Z Z Z
f dm = f dm + f dm.
A B

Proof. The last two inequalities are obvious. To prove the others, let f 0 ≤ f, g 0 ≤ g be
bounded measurable functions with finite measure supports. Because cf 0 ≤ cf and f 0 + g 0 ≤
f + g are bounded measurable functions with finite supports we have
Z Z Z
0 0
c f dm = cf dm ≤ cf dm
Z Z Z Z
0 0 0 0
f dm + g dm = f + g dm ≤ f + g dm.

Taking the sup over all such f 0 , g 0 yields

Z Z
cf dm ≥ c f dm
Z Z Z
f + g dm ≥ f dm + g dm.

It follows that
Z Z Z Z Z
f dm = (1/c)cf dm ≥ (1/c) cf dm ≥ (1/c)c f dm = f dm.
R R
Since equality holds throughout, we must have that cf dm = c f dm as claimed.
Let k ≤ f + g be a bounded measurable function with finite measure support. Let
k = min(f, k) and k 00 = k − k 0 . Then k 0 ≤ f and k 00 ≤ g (since k ≤ f + g implies k − f ≤ g
0

implies k − k 0 ≤ g). So
Z Z Z Z
0 00
k dm + k dm = k dm ≤ f + g dm.

Taking the sup over all k proves the exercise.

15
Exercise 38. Let f : R → R be a nonnegative function and R = {(x, y) : 0 ≤ y ≤ f (x)} be
the region under the graph of f . Then
Z
f dm = m(R).

Proof. By linearity, it suffices to prove this in the special case in which f has finite support.
This is clear if f is the characteristic function of an interval. Therefore, it is also true
if f is a finite linear combination of such characteristic functions. Since any measurable set
with finite measure is nearly a finite union of intervals (Littlewood’s first principle), it is also
true if f = χE where E is measurable. Therefore it is true if f is simple.
Using Egorov’s Theorem and approximation of f by simple functions, we obtain the
result for arbitrary f .

11 Integrable functions
Definition 18. A measurable function f : X → R is integrable if
Z
|f | dm < ∞.
R
Because |f | is a nonnegative measurable function, |f | dm is well-defined. In this case, let

f + = max(f, 0)

f − = max(−f, 0).
Note: f + , f − are measurable functions, f = f + − f − and |f | = f + + f − . Define
Z Z Z
f dm = f dm − f − dm.
+

Exercise 39. Suppose f, g : X → R are integrable and c ∈ R is a constant. Then

R R
1. cf is integrable and cf dm = c f dm,
R R R
2. f + g is integrable and f + g dm = f dm + g dm,
R R R R
3. if f ≤ g a.e. then f dm ≤ g dm. In particular if f = g a.e. then f dm = g dm.

4. if X = A ∪ B where A, B are disjoint measurable sets, then

Z Z Z
f dm = f dm + f dm.
A B

16
Proof. Suppose that c > 0. Then (cf )+ = cf + and (cf )− = cf − . This implies (1) when
c > 0. The case c ≤ 0 is similar.
To prove (2), suppose f = f1 − f2 where f1 , f2 are nonnegative measurable functions.
Then
f = f + − f − = f1 − f2
implies f + + f2 = f − + f1 . By the previous section on integrals of nonnegative functions,
Z Z Z Z Z Z
+ + − −
f dm + f2 dm = f f2 dm = f + f1 dm = f dm + f1 dm.

Therefore, Z Z Z Z Z
+ −
f dm = f dm − f dm = f1 dm − f2 dm.

Now let us apply this to the situation at hand:

Z Z
f + g dm = f + − f − + g + − g − dm
Z
= (f + + g + ) − (f − + g − ) dm
Z Z
= (f + g ) dm − (f − + g − ) dm
+ +

Z Z Z Z 
+ + − −
= f dm + g dm − f dm + g dm
Z Z
= f dm + g dm.

This proves (2). (3) follows from (2) since

Z Z Z
0 ≤ g − f dm = g dm − f dm.

(4) follows from (2) since f = χA f + χB f .

12 Convergence Theorems
There are four convergence
R Rtheorems which state that if fn converges in some sense to a
function f then fn dm → f dm (or at least we have an inequality if not an equality).
Exercise 40 (Bounded Convergence Theorem). Suppose m(X) < ∞. Let {fn } be a uni-
formly bounded sequence of measurable functions on X that converge pointwise a.e. to f .
(uniformly bounded means there is a number M > 0 such that |fn (x)| ≤ M for a.e. x and
every n). Then Z Z
fn dm → f dm
as n → ∞.

17
Proof. This follows from Egorov’s Theorem. Indeed, if  > 0 then, by Egorov’s Theorem
there is a set Y ⊂ X such that m(X \ Y ) < /(4M ) and an N such that n > N implies
|fn (x) − f (x)| ≤ /(2µ(X))
for all x ∈ Y . Therefore
Z Z Z
| fn (x) dm − f (x) dm| ≤ |fn (x) − f (x)| dm
Z Z
= |fn (x) − f (x)| dm + |fn (x) − f (x)| dm
Y X\Y
≤ (/(2µ(X)))µ(X) + (/(4M ))(2M ) < .

Exercise 41 (Fatou’s Lemma). Let {fn } be an sequence of nonnegative functions. Then

Z Z
lim inf fn dm ≥ lim inf fn dm
n n
R
as n → ∞. Hint: use the definition of lim inf n fn dm directly.
Proof. Let h ≤ lim inf n fn be a bounded measurable function with finite measure support.
Let
hn = min(h, fn ).
Note hn → h pointwise. So the bounded convergence theorem implies
Z Z Z
h dm = lim hn dm ≤ lim inf fn dm
n n

Take the sup over all h to obtain the exercise.

Exercise 42. Find an example of a sequence of functions as in Fatou’s Lemma such that the
inequality is strict.
Proof. For example, we could have fn = (+∞)χ[0,1/n] or fn = χ[n,+∞) .
Exercise 43 (Monotone Convergence Theorem). Let {fn } be an increasing sequence of non-
negative functions. (This means 0 ≤ f1 ≤ f2 ≤ · · · ). Let f (x) = limn fn (x) ∈ [0, +∞].
Then Z Z
fn dm → f dm
as n → ∞.
Proof. By Fatou’s Lemma, Z Z
lim inffn dm ≥ f dm.
n
R R
But for each n we have fn ≤ f which implies fn dm ≤ f dm. Taking limits implies the
opposite inequality.

18
Exercise 44. If X1 , X2 , . . . are pairwise disjoint measurable sets whose union is X and f ≥ 0
is measurable then Z ∞ Z
X
f dm = f dm.
i=1 Xi
R
(therefore if µ(E) = E f dm Pthen µ is a measure on X). More generally, if {fn } are
nonnegative functions and f = n fn then
Z XZ
f dm = fn dm.
n

Exercise 45 (Lebesgue’s Dominated Convergence Theorem). Let {fn } be a sequence of mea-

surable functions on X that converge pointwise a.e. to f . Suppose there is an integrable
function g such that 0 ≤ fn ≤ g for all n. Then
Z Z
fn dm → f dm

Proof. Proof #1:One way to do this is to realize that g dm is a measure. That is: we define
a new measure µ on X by Z
µ(E) = g dm.
E
We’ve already proven that this really defines a measure. Moreover, if k is any bounded
measurable function then Z Z
k dµ = kg dm.

(To see this, note that it’s true for characteristic functions and therefore true for simple
functions and therefore true for arbitrary bounded measurable functions).
After replacing X with the support of g if necessary we may assume that g > 0 a.e. By
the Bounded Convergence Theorem,
Z Z
fn /g dµ → f /g dµ.
R R R
(This is because µ(X) = g dm < ∞ and 0 ≤ fn /g ≤ 1 is bounded). Since k dµ = kg dm
for any bounded measurable function k, this implies the exercise.
Proof #2: By Fatou’s Lemma,
Z Z
lim inf fn dm ≥ f dm.
n

On the other hand, g − fn ≥ 0 converges pointwise a.e. to g − f . So Fatou’s Lemma implies

Z Z
lim inf g − fn dm ≥ g − f dm
n

In other words, Z Z
lim sup fn ≤ f dm.
n

19
Exercise 46. The hypotheses of Lebesgue’s Dominated Convergence Theorem can be weak-
ened by replacing 0 ≤ fn ≤ g with h ≤ fn ≤ g where h, g are integrable functions.
R R
Proof. Observe
R that 0 ≤ f n −h ≤ g −h. By the previous exercise, f n −h dm → f −h dm.
Cancelling −h dm from both sides proves the result.
Definition 19. If (X, m) is a measure space and p > 0, we let Lp (X, m) be the set of all
equivalence classes classes of measurable functions f : X → C such that
Z
|f |p dm < ∞.

Here: two functions f, g are equivalent if they agree a.e. By abuse of notation, we may write
f ∈ L1 (X, m) to mean that f is integrable (for example). Later on, we will show that if
p ≥ 1 and f, g ∈ Lp (X, m) then
Z 1/p
p
kf − gkp = |f − g| dm

is a metric on Lp (X, m). We will study the topological aspects of these spaces later.
Also, we let L∞ (X, m) denote the set of equivalence classes of bounded measurable func-
tions. We write
kf k∞ = sup{a ≥ 0 : m(f −1 [a, ∞]) > 0}.
(This is called the essential supremum of f ). This also gives a metric on L∞ (X, m): the
distance between f and g is kf − gk∞ .

13 Riemannian integration
Let f be a function on an interval [a, b] ⊂ R. Let Σ = {x0 , x1 , . . . , xn } ⊂ [a, b] with
x0 = a, xn = b. Define
n
X
IΣ+ (f ) = (sup f  [xi−1 , xi ])|xi − xi−1 |
i=1

and n
X
IΣ− (f ) = (inf f  [xi−1 , xi ])|xi − xi−1 |.
i=1

Let diam(Σ) = maxi |xi − xi−1 |. We define the Riemannian integral of f (if it exists) by
Z
R f dx = lim IΣ+ (f ) = lim IΣ− (f ).
diam(Σ)→0 diam(Σ)→0

We say f is Riemann-integrable if its Riemann integral exists and is finite.

We will prove:

20
Theorem 13.1. Suppose f : [a, b] → R is Riemann-integrable. Then f is integrable and
Z Z
R f dx = f dm.
R
Exercise 47. If f is measurable, f ≥ 0 a.e. and f dm = 0 then f = 0 a.e.
Exercise 48. Suppose f : X → R is a function satisfying
Z Z
sup φ dm = inf ψ dm
φ≤f ψ≥f

where the sup is over all simple functions φ ≤ f and the inf is over all simple functions
ψ ≤ f . Then f is measurable. R
Hint: there exist simple functions φn , ψn such that φn ≤ f ≤ ψn and ψn −φn dm ≤ 1/n.
Let φ∗ = supn φn and ψ ∗ = inf n ψn .
Exercise 49. Finish the proof of Theorem 13.1.
Proof. If R is Riemann-integrable then
Z Z
sup φ dm = inf ψ dm
φ≤f ψ≥f

where the sup is over all step functions φ ≤ f and the inf is over all step functions ψ ≤ f .
So the previous exercise implies f is measurable.
Now
Z Z Z Z Z
sup φ dm ≤ sup φ dm ≤ f dm ≤ inf ψ dm ≤ inf ψ dm
φ≤f φ≤f ψ≥f ψ≥f

where the first sup and the last inf are over step functions and the secondR sup and first inf
are over simple functions. Since the first and last quantity are equal to R f dm this proves
Z Z
R f dx = f dm.

Exercise 50. A bounded function f : [a, b] → R is Riemann-integrable if and only if the set
of discontinuities of f has measure zero.

14 Differentiation
Motivation:

1. when does the derivative f 0 exist?

21
 Rb
2. when does f 0 dx = f (b) − f (a)?
a
d
Rx
In this subsection we will show that dx ( a f dm) exists a.e. under very general conditions.
First,
Exercise 51. Let f : X → R be an integrable function where (X, C, m) is a measure
R space.
For every  > 0 there exists δ > 0 such that if E ⊂ X satisfies m(E) < δ then E |f | dm ≤ .
(InR Wheedon-Zygmund this is referred to as “absolute continuity” of the “set function”
E 7→ E f dm.)
Our next goal is to prove:
Theorem 14.1 (Lebesgue’s Differentiation Theorem). Let f : Rn → R be integrable. Then
for a.e. x ∈ Rn , Z
1
f (x) = lim f dm
Q&x m(Q) Q

where the limit is over cubes Q centered at x.

We will need the following approximation result:
Exercise 52. Let f ∈ L1 (Rn ). Then for every  > 0 there exists a continuous function g with
compact support such that kf − gk1 ≤ . In other words, Cc (Rn ) is dense in L1 (Rn ) where
Cc (Rn ) denotes the subspace of compactly-supported continuous functions.
Definition 20. The Hardy-Littlewood maximal function of f is
Z
∗ 1
f (x) = sup |f | dm
Q m(Q) Q

where the sup is over all cubes centered at x.

Exercise 53. Let E ⊂ Rn be bounded with positive measure. Let f (x) = χE (x). Prove that
f ∗ is not integrable.
Definition 21. We say a measurable function f is in weak L1 if there is a constant C > 0
such that
kf k1
m({x ∈ Rn : |f (x)| ≥ t}) ≤ C
t
for all t > 0.
Exercise 54. Prove that all integrable functions are in weak L1 (with constant C = 1). This
is known as Chebyshev’s inequality.
We will prove:
Theorem 14.2 (Hardy-Littlewood maximal inequality). For any f ∈ L1 (Rn ) and any t > 0,
kf k1
m({x ∈ Rn : f ∗ (x) ≥ t}) ≤ Cn
t
where Cn > 0 is a contant depending only on the dimension n. In particular, f ∗ is in weak
L1 .

22
 First we need:
Exercise 55 (The Simple Vitali Covering Lemma). Let E ⊂ Rn satisfy m∗ (E) < ∞. Let C
be a collection of cubes covering E. Then there exists a constant Cn > 0 (depending only
on n) and a subcollection C0 ⊂ C such that

• the cubes in C0 are pairwise disjoint

• m(∪{Q : Q ∈ C0 }) ≥ Cn m∗ (E).

Hint: use a greedy algorithm.

Exercise 56. Prove the Hardy-Littlewood maximal inequality in the special case. Hint: for
∗ 1
R
each x with f (x) > t choose a cube Qx such that m(Qx ) Qx |f | dm ≥ t. These cubes cover
{x ∈ Rn : f ∗ (x) ≥ t}. Use Vitali’s Covering Lemma. What can you say about E |f | dm
R

where E = ∪{Q : Q ∈ C0 }?
Proof. Following the hint, we observe that,
Z Z
|f | dm ≤ |f | dm = kf k1 .
E

On the other hand, because the cubes in C0 are pairwise disjoint,

Z XZ X
|f | dm = |f | dm ≥ tm(Q) = tm(E).
E Q∈C0 Q Q∈C

kf k1
Thus m(E) ≤ t
. By the way we chose E, m(E) ≤ (1/Cn )m({x : f ∗ (x) ≥ t}). So

kf k1
m({x : f ∗ (x) ≥ t}) ≤ Cn .
t

Definition 22. A point x ∈ Rn is a Lebesgue point of a locally integrable function f if

Z
1
lim |f (y) − f (x)| dm(y) = 0.
Q&0 m(Q) Q

Exercise 57. For any Lebesgue measurable function f , almost every point is a Lebesgue point.
Note: this is formally stronger than Lebesgue’s Differentation Theorem. Hint: First consider
the case in which f is continuous with compact support. Then approximation arbitrary f
by compact supported continuous functions. Use the maximal inequality to control the error
term.
Proof. The special case in which f is continuous with compact support is easy. So let us
assume f is arbitrary.

23
 Let  > 0 and let g be a continuous function with compact support. Let
Z
1
E = {x ∈ X : lim sup |f (x) − f (y)| dm(y) > }.
Q&0 m(Q) Q

By the triangle inequality,

|f (x) − f (y)| ≤ |f (x) − g(x)| + |g(x) − g(y)| + |g(y) − f (y)|.

Because Z
1
lim sup |g(x) − g(y)| dm = 0,
Q&0 m(Q) Q

E ⊂ E0 ∪ E00 where

E0 = {x ∈ X : |f (x) − g(x)| ≥ /2}
Z
00 1
E = {x ∈ X : lim sup | |g − f | dm| ≥ /2}.
Q&x m(Q) Q
Note Z
2 2
m(E0 ) ≤ |f (x) − g(x)| dm ≤ kf − gk1 .
 E 
Note that
E00 ⊂ {x ∈ X : (g − f )∗ (x) ≥ /2}.
So the maximal inequality implies m(E00 ) ≤ 2Cn

kf − gk1 for some constant Cn > 0. Thus
2Cn + 2
m(E ) ≤ m(E0 ) + m(E00 ) ≤ kf − gk1 .

By a previous exercise, we can make kf − gk1 as small as we wish. Therefore m(E ) = 0.
Since  is arbitrary, this implies the theorem.
Remark 2. The proof above is typical of a general strategy involving maximal inequalities:
suppose there is a subset P ⊂ L1 (X). We wish to prove that P = L1 (X). (In the Differ-
entiation Theorem, P is the set of all functions f satisfying the Differentiation Theorem).
First we prove that P contains a dense subset (in the example, the dense subset was the
set of compactly-supported continuous functions). Second we use a maximal inequality to
prove that P is closed. This same strategy is employed to prove Birkhoff’s pointwise ergodic
theorem.
R
Exercise 58. A measurable function f on Rn is locally integrable if K |f | dm < ∞ for every
compact K ⊂ Rn . The conclusion to Lebesgue’s Differentiation Theorem holds for locally
integrable functions.
Exercise 59. For any measurable subset E ⊂ Rn ,
m(Q ∩ E)
χE (x) = lim
Q&x m(Q)
for a.e. x ∈ E where the limit is over all cubes Q centered at x.

24
Definition 23. A family {S} of measurable sets shrinks regularly to x if
• the diameters of the sets S tend to 0

• if Q is the smallest cube with center x containing S then there is a constant k inde-
pendent of S such that m(Q) ≤ km(S).
Exercise 60. Suppose x is a Lebesgue point of f . Then
Z
1
|f (y) − f (x)| dm(y) → 0
m(S) S

for any family of sets {S} that shrink regularly to x.

15 Differentiation of monotone functions

Motivation:
1. when does f 0 exist?
Rb
2. when does a f 0 dx = f (b) − f (a)?
Here we will study the second question. We begin with monotone functions.
Definition 24. f : [a, b] → R is monotone increasing if f (x) ≤ f (y) whenever x ≤ y. f is
monotone decreasing if f (x) ≥ f (y) whenever x ≤ y.
We will prove:
Theorem 15.1. If f : (a, b) → R is monotone increasing and measurable then f 0 exists a.e.,
f 0 is measurable and Z b
f 0 dx ≤ f (b− ) − f (a+ ).
a

Exercise 61. Use the Cantor-Lebesgue function to prove that the inequality above can be
strict.
Definition 25. We say that a collection I of intervals covers a set E ⊂ R in the sense of
Vitali if for every  > 0 and x ∈ E there exists I ∈ I with x ∈ I and m(I) < . The intervals
may be closed, open or half-open but we do not allow an interval to degenerate to a point.
Exercise 62 (The not-as-simple Vitali’s Covering Lemma). Let E ⊂ R have finite outer
measure. Suppose I covers E in the sense of Vitali. Then for any  > 0 there exists a
countable subset I0 ⊂ I such that
• the intervals in I0 are pairwise disjoint

• m∗ (E \ ∪I∈I0 I) = 0.

25
 • m∗ (∪I∈I0 ) ≤ (1 + )m∗ (E).

Proof. Let G ⊃ E be an open set with m(G) ≤ (1 + )m∗ (E). Without loss of generality,
we may assume that every cube of I lies in G. So the (3) is automatic.
By the Simple Vitali Covering Lemma, there exist a finite collection I1 ⊂ I such that (a)
I1 is pairwise disjoint and (b)
m(∪I∈I1 I) ≥ C1 m∗ (E)
for some C1 > 0. Thus

m∗ (E \ ∪I∈I1 I) ≤ m(G \ ∪I∈I1 I) = m(G) − m(∪I∈I1 I)

≤ (1 +  − C1 )m∗ (E).

By choosing  < C1 /2 we obtain

m∗ (E \ ∪I∈I1 I) ≤ (1 − /2)m∗ (E).

Repeating the above argument we find a finite pairwise disjoint collection I2 ⊂ I \ I1 such
that
m∗ (E \ ∪I∈I1 ∪I2 I) ≤ (1 − /2)m∗ (E \ ∪I∈I1 I) ≤ (1 − /2)2 m∗ (E).
Repeating in this way we obtain the result with I0 = ∪n In .

Exercise 63. Show that in Vitali’s not-so-simple Covering Lemma, we can choose I0 to be
finite if we relax the second condition to

m∗ (E \ ∪I∈I0 I) < .

Definition 26. Let f : [a, b] → R. For x ∈ [a, b] define the Dini derivatives of f by:

f (x + h) − f (x)
Dne f (x) := lim sup
h&0 h
f (x + h) − f (x)
Dse f (x) := lim inf
h&0 h
f (x) − f (x − h)
Dnw f (x) := lim sup
h&0 h
f (x) − f (x − h)
Dsw f (x) := lim inf .
h&0 h
(nw stands for northwest, etc). If all of these are equal at a point x then we let f 0 (x) denote
the common value. This is the derivative of f .

Exercise 64. Prove Theorem 15.1. Hint: For rational number u < v, let

Eu,v = {x : Dsw f (x) < u < v < Dne f (x)}.

26
In order to prove that m∗ (Eu,v ) = 0, let I be the collection of all intervals of the form [x−h, x]
such that
f (x) − f (x − h) < uh.

P the previous exercise to obtain a nice finite subcollection I ⊂ I. Get an upper bound
0
Apply
on [c,d]∈I0 f (d) − f (c).
Then let Fu,v = Eu,v ∩(∪I∈I0 I). Let J be the collection of all intervals of the form [x, x+h]
with [x, x + h] ⊂ I ∈ I0 for some I and such that such that f (x + h) − f (x) > vh. Apply
P exercise again to get a nice finite subcollection J ⊂ J. This
0
the previous time get a lower
bound on [c,d]∈J0 f (d) − f (c). Then use the fact that each interval in J is a subinterval of
0

an interval in I0 .
Proof. Let  > 0. Following the hint, we obtain a finite collection I0 ⊂ I of intervals such
that

• the intervals in I0 are pairwise disjoint

• m∗ (Eu,v \ ∪I∈I0 I) < .

• m(∪I∈I0 ) ≤ (1 + )m∗ (Eu,v ).

If I = [c, d] ∈ I0 then f (d) − f (c) < u(d − c) = um(I). Therefore,

X
f (d) − f (c) < um(∪I∈I0 ) ≤ u(1 + )m∗ (Eu,v ).
[c,d]∈I0

Let Fu,v = Eu,v ∩ (∪I∈I0 I). Note that m∗ (Fu,v ) ≥ m∗ (Eu,v ) − .

By the previous exercise, we obtain a finite collection J0 ⊂ J of intervals such that

• the intervals in J0 are pairwise disjoint

• m∗ (Fu,v \ ∪J∈J0 J) < .

• m(∪J∈J0 ) ≤ (1 + )m∗ (Fu,v ).

If J = [c, d] ∈ J0 then f (d) − f (c) > v(d − c) = vm(J). Therefore,

X
f (d) − f (c) > vm(∪J∈J0 J) ≥ v(m∗ (Fu,v − ) ≥ v(m∗ (Eu,v − 2).
[c,d]∈J0

Because each interval in J0 is a subinterval of an interval in I0 and f is monotone increasing,

X X
f (d) − f (c) ≤ f (d) − f (c).
[c,d]∈J0 [c,d]∈I0

So
u(1 + )m∗ (Eu,v ) ≥ v(m∗ (Eu,v − 2).

27
Because  > 0 is arbitrary and u < v, this implies m∗ (Eu,v ) = 0. Since this holds for all
rational u, v it follows that Dsw f (x) = Dne f (x) for a.e. x. In a similar way, it can be shown
that all of the derivatives are equal. Therefore, f 0 exists a.e.
Now let
f (x + 1/k) − f (x)
fk (x) =
1/k
where we define f (x) = f (b− ) for x ≥ b. Observe that fk → f 0 a.e. Moreover fk ≥ 0 since f
is monotone increasing. Fatou’s Lemma implies
Z b Z b
0
f dm ≤ lim inf fk dm.
a k a

However, (setting h = 1/k),

Z b Z b+h Z b Z a+h
−
fk dm = (1/h) f dm − f dm = f (b ) − (1/h) f dm.
a a+h a a
R a+h Rb
Since f (a+ ) ≤ (1/h) a
f dm ≤ f (a+ + 1/h), we obtain a
fk dm → f (a+ ). This implies
the theorem.

15.1 Functions of Bounded Variation

Let f : [a, b] → R be a function. Let Γ = {x1 , x2 , . . . , xn } ⊂ [a, b] be a finite set with
x1 < x2 < · · · < xn . Define
n−1
X
sΓ = SΓ [f ; a, b] = |f (xi+1 ) − f (xi )|.
i=1

Let
V = V [f ; a, b] = sup sΓ
Γ

be the variation of f over [a, b]. We say that f has bounded variation on [a, b] if V [f ; a, b] <
∞.
Exercise 65. 1. If f is monotone on [a, b] then it has bounded variation.

2. If f, g have bounded variation and c, d ∈ R are scalars then cf + dg has bounded

variation.

3. If [a0 , b0 ] ⊂ [a, b] then V [f ; a0 , b0 ] ≤ V [f ; a, b].

4. If a < c < b then V [f ; a, b] = V [f ; a, c] + V [f ; c, b].

Exercise 66. Construct a continuous function on [0, 1] that is of unbounded variation on
every subinterval. Hint: modify the Cantor-Lebesgue construction.

28
Exercise 67. Suppose f is continuously differentiable on [a, b]. Prove that V [f ; a, b] =
Rb 0
a
|f | dx.
For Γ as above, let
n−1
X
PΓ = [f (xi+1 ) − f (xi )]+
i=1
n−1
X
NΓ = [f (xi+1 ) − f (xi )]− .
i=1

Observe that SΓ = PΓ + NΓ and f (b) − f (a) = PΓ − NΓ (assuming a, b ∈ Γ). Let P =

P [f ; a, b] = supΓ PΓ , N [f ; a, b] = supΓ NΓ .
Exercise 68. P + N = V , P − N = f (b) − f (a). Thus
1 1
P = [V + f (b) − f (a)], N = [V − f (b) + f (a)].
2 2
Proof. Because V ≥ SΓ = PΓ + NΓ , V ≥ P + N . On the other hand, SΓ ≤ P + N . So taking
the sup over Γ yields V ≤ P + N .
Since NΓ + f (b) − f (a) = PΓ , we have N + f (b) − f (a) = P .
Exercise 69. Prove that f as bounded variation on [a, b] if and only if f is the difference of
two monotone functions.
Proof. By a previous exercise we only need to show that if f has bounded variation then it
is the difference of two monotone functions. Let P (x) = P [f ; a, x] and N (x) = N [f ; a, x].
These are monotone functions and the previous exercise shows f (x) = P (x)−N (x)+f (a).
Exercise 70. If f has bounded variation then the set of discontinuities of f is countable.
Moreover, every discontinuity is a jump discontinuity.
Exercise 71. If f is of bounded variation on [a, b] and V (x) = V [f ; a, x] for a ≤ x ≤ b then
V 0 (x) = |f 0 (x)|
for a.e. x ∈ [a, b].
hints: ??

16 Absolutely continuous functions

Rb
Next, we investigate: when does a
f 0 dx = f (b) − f (a)?
Definition 27. We say that f is absolutely continuous on an interval [a, b] if for every  > 0
P exists a δ > 0 such that if I is any collection of pairwise disjoint intervals in [a, b] with
there
I∈I m(I) < δ then X
|f (d) − f (c)| < .
[c,d]∈I

29
Exercise 72. If f is absolutely continuous then f is continuous.
Exercise 73. If f is Lipschitz then f is absolutely continuous. For example, if f is continu-
ously differentiable on [a, b] then f is absolutely continuous.
Exercise 74. If f is absolutely continuous on [a, b] then f has bounded variation on [a, b]. So
f 0 exists a.e.
Proof. Let  > 0 and let δ > 0 be as in the definition of absolute continuity. Let Γ = {a =
x0 , x1 , . . . , xn , b = xn+1 } be any partition of the interval into subintervals of length ≤ δ.
Then by partitioning Γ into collections of subintervals whose total less is ≤ δ we see that

sΓ ≤ (1 + d1/δe).

(The partition of Γ can be chosen to have at most (1 + d1/δe) parts). Since this is true for
every Γ, we are done.

Definition 28. A function f for which f 0 = 0 a.e. is said to be singular. For example, the
Cantor-Lebesgue function is singular.

Exercise 75. If f is absolutely continuous and singular on [a, b] then f is constant. Hint: use
Vitali’s Covering Lemma.
Proof. Let E be the set of all x ∈ [a, b] such that f 0 (x) = 0. So m(E) = b − a. Let  > 0.
Because f is absolutely continuous there P exists δ > 0 such that if J is any pairwise disjoint
collection of subintervals of [a, b] then [c,d]∈J |f (d) − f (c)| < .
Let I be the collection of all intervals of the form [x, x + h] with x ∈ E and

|f (x + h) − f (x)| < h.

By the exercise after Vitali’s Covering Lemma, there exists a pairwise disjoint finite subcol-
lection I0 ⊂ I such that X
m(I ∩ [a, b]) ≥ b − a − δ.
I∈I0

Let J be the collection of intervals in the complement of ∪I∈I0 I in [a, b]. We now have
X
|f (d) − f (c)| <  + (b − a).
[c,d]∈I0 ∪J

Since X
f (b) − f (a) = f (d) − f (c)
[c,d]∈I0 ∪J

we have X
|f (b) − f (a)| ≤ |f (d) − f (c)| <  + (b − a).
[c,d]∈I0 ∪J

Since  is arbitrary, this implies f (b) = f (a).

30
 So we have shown that if f 0 = 0 a.e. and f is absolutely continuous on [a, b] then
f (b) = f (a). We can apply this again to any subinterval of [a, b] to see that f must be
constant throughout.

Exercise 76. f is absolutely continuous on [a, b] if and only if f 0 exists a.e. on [a, b], f 0 is
integrable on [a, b] and Z x
f (x) − f (a) = f 0 dm.
a
Hint: use the previous exercise.
Proof. We have already shown that integrals are absolutely continuous. So one direction is
clear.
Now assume f is absolutely continuous
Rx 0 on [a, b]. We have already shown that f 0 exists
a.e. and is integrable. Let F (x) = a f dm. By Lebesgue’s Differentiation Theorem, F 0 = f 0
a.e. and F is absolutely continuous. So F − f is absolutely continuous and singular on
[a, b]. By the previous Rexercise, F − f is constant. Since F (a) − f (a) = −f (a) this implies
x
F (x) = f (x) − f (a) = a f 0 dx.
Exercise 77. If f has bounded variation on [a, b] then f = g+h for some absolutely continuous
function g and some singular function h on [a, b]. Moreover, g and h are unique up to additive
constants.
Rx
Proof. For the first part, let g(x) = a f 0 dm and h = f − g. For the second part, note that
if f = g1 + h1 = g2 + h2 then g1 − g2 = h2 − h1 is both absolutely continuous and singular.
Exercise 78. If f, g are absolutely continuous on [a, b] then
Z b Z b
0
gf d = g(b)f (b) − g(a)f (a) − g 0 f dx.
a a

Proof. it is easy to check that gf is absolutely continuous on [a, b] and the product rule

(gf )0 = g 0 f + gf 0

holds. The formula above follows by integrating the product rule.

17 Convex functions
Definition 29. A function φ : (a, b) → R is convex if for every pair a0 , b0 with a < a0 < b0 < b
the graph of φ  [a0 , b0 ] lies under (or on) the graph of the line from (a0 , φ(a0 )) to (b0 , φ(b0 )).
We allow a = −∞, b = +∞.
Exercise 79. φ is convex on (a, b) if and only if for 0 < t < 1 and x < y with x, y ∈ (a, b) we
have
φ(tx + (1 − t)y) ≤ tφ(x) + (1 − t)φ(y).

31
Exercise 80. If φ is twice continuously differentiable and φ00 ≥ 0 then φ is convex.
Exercise 81. If φ is convex then it is continuous.
Exercise 82. If φ is convex then φ0 (x) exists for all but countably many x’s. Moreover, φ0 is
monotone increasing.
Exercise 83. If φ is convex then φ is locally Lipschitz. In other words, for every pair a0 , b0
with a < a0 < b0 < b there is a constant C > 0 such that if x, y ∈ [a0 , b0 ] then
|f (y) − f (x)| ≤ C|x − y|.
The last exercise implies that φ is absolutely continuous.
Exercise 84 (Jensen’s inequality). Let (X, µ) be a probability space and f ∈ L1 (X, µ).
Suppose that φ is convex on the essential range of f . Then
Z Z
φ( f dµ) ≤ φ(f ) dµ.

Hint: to get an intuition for why this result is true, consider the special case in which µ is
the probability measure on R satisfying µ({a}) = t, µ({b}) = 1 − t for some 0 < t < 1.
R
Proof. Let γ = f dµ and let m0 be the slope of a supporting line L at γ. In other words
m0 = φ0 (γ) if this exists. If it does not exist, then we could, for example, set m0 equal to
any of the Dini derivatives of φ at γ.
Because φ is convex, for a.e. x ∈ X,
φ(f (x)) ≥ m0 (f (x) − γ) + φ(γ).
(This is because m0 (f (x) − γ) + φ(γ) is “y-coordinate” of the point on L that has “x-
coordinate” equal to f (x)). Integrating over x, we obtain the exercise.
For example, take φ(x) = ex . Then Jensen’s inequality becomes,
Z Z
exp( f dµ) ≤ ef dµ.

If X = {p1 , . . . , pn }, µ({pi }) = 1/n and f (pi ) = xi then we obtain

exp((x1 + · · · + xn )/n) ≤ (1/n)(ex1 + · · · + exn ).
If we put y = exi then we obtain
(y1 · · · yn )1/n ≤ (1/n)(y1 + · · · + yn )
in other words: the arithmetic mean dominates the geometric mean (for a sequence of positive
numbers).
More generally, if µ({pi }) = wi then we have
X
y1w1 · · · ynwn ≤ wi yi .
i

32
 ap bq
Exercise 85. Suppose a, b ≥ 0, p, q > 1 and 1/p + 1/q = 1. Then ab ≤ p
+ q
.
Proof. Wlog a, b > 0.
Let φ(x) = ex and let µ be the probability measure on {log(a), log(b)} defined by
µ({log(a)}) = 1/p, µ({log(b)}) = 1/q. Define f on {log(a), log(b)} by f (log(a)) = p log(a)
and f (log(b)) = q log(b). By Jensen’s inequality,

ep log(a)/p+q log(b)/q ≤ (1/p)ep log(a) + (1/q)eq log(b) .

18 Lp spaces
Recall that if (X, µ) is a measure space and 0 < p < ∞ then Lp (X, µ) is the set of all mod
0 equivalence classes of measurable functions f on X satisfying
Z 1/p
p
kf kp L = |f | dµ < ∞.

We also define
kf k∞ = sup{t ≥ 0 : µ({x ∈ X : |f (x)| ≥ t}) > 0}.
By the way, we allow f to be complex-valued by defining
Z Z Z
f dµ = Re(f ) dµ + i Im(f ) dµ.

Exercise 86 (Young’s Inequality). Let φ : [0, ∞) → [0, ∞) be continuous, strictly increasing

with φ(0) = 0. Let φ−1 denote its composition inverse (so φ ◦ φ−1 (x) = x). Then for a, b > 0,
Z a Z b
ab ≤ φ(x) dx + φ−1 (x) dx.
0 0

Equality holds if and only if b = φ(a).

Hint: there is an almost immediate geometric proof. Simply draw the graph of φ and
realize that it is also the graph of φ−1 (if we switch the axes) and interpret the integrals as
areas under the curves.
1 1
Definition 30. We say numbers p, q ∈ [1, ∞] are conjugate if p
+ q
= 1. For examples, 2 is
conjugate with itself and 1 is conjugate with +∞.

Exercise 87 (Hölder’s inequality). If p, q are conjugate then kf gk1 ≤ kf kp kgkq . Hints: note
that if the result holds for f, g and a, b are scalars then the result holds for af and bg. So we
may assume without loss of generality, that kf kp = kgkq = 1. Now use Young’s inequality
with φ(x) = xp−1 .

33
Proof. The case when p or q = ∞ is obvious. So assume 1 < p, q < ∞. Then Young’s
inequality implies
Z a Z b
p−1 ap b1/(p−1)+1
ab ≤ x dx + x1/(p−1) dx = +
0 0 p 1 + 1/(p − 1)
p 1
Notice that 1/(p − 1) + 1 = p−1
= 1−1/p
= q. So

ab ≤ ap /p + bq /q.

Applying this to a = |f (x)|, b = |g(x)| and integrating over x, we obtain

kf gk1 ≤ kf kpp /p + kgkqq /q.

In the special case in which kf kp = kgkq = 1, this implies the result.

Schwarz’s inequality is the special case p = q = 2.
Exercise 88 (Minkowski’s inequality). If 1 ≤ p ≤ ∞ and f, g ∈ Lp (X) then kf + gkp ≤
kf kp + kgkp . Hint: the cases p ∈ {1, ∞} are straightforward so assume 1 < p < ∞. Note
that
Z Z Z Z
|f + g| dµ = |f + g| |f + g| dµ ≤ |f + g| |f | dµ + |f + g|p−1 |g| dµ.
p p−1 p−1

Now think about how to apply Hölder’s inequality.

Proof. Hölder’s inequality implies
Z
|f + g|p−1 |f | dµ ≤ kf kp k|f + g|p−1 kq .

Since (1/p) + (1/q) = 1, multiplying out the denominators gives p + q = pq. So (p − 1)q = p.
So
k|f + g|p−1 kq = kf + gkp/q
p .

Similarly, Z
|f + g|p−1 |f | dµ ≤ kgkp kf + gkp/q
p .

So we obtain Z
kf + gkpp |f + g|p dµ ≤ (kf kp + kgkp )kf + gkp/q
p .

Observe that p/q = p − 1. So dividing both sides by kf + gkp−1

p finishes it.
Minkowski’s inequality implies that Lp (X) is a metric space with metric d(f, g) = kf −gkp
if 1 ≤ p ≤ ∞. By contrast,
Exercise 89. Show using explicit examples that if 0 < p < 1 then there exist f, g ∈ Lp (X)
such that kf − gkp > kf kp + kgkp .

34
Exercise 90. Show that, if p ≥ 1, then Lp (X) is a complete metric space. Hint: Suppose
p < ∞. If {fn } is Cauchy in Lp then Chebyshev’s inequality implies {fn } is Cauchy in
measure. This implies (by a previous exercise?) the existence of f such that fn → f is
measure. Now apply Fatou’s Lemma.
Proof. If p = ∞ then for a.e. x, {fn (x)} is Cauchy in R. So there exists a limit function
f ∈ L∞ and fn converges uniformly a.e. to f .
We may assume p < ∞. Assuming the hint, we have fn → f in measure. So Fatou’s
Lemma implies
Z Z
p
lim inf lim inf |fnj − fm | dµ ≤ lim inf |fnj − fm |p dµ.
m nj nj ,m

The RHS is 0. So the LHS is also 0. We can choose the subsequence {nj } so that fnj → f
pointwise a.e.
So kfm − f kp → 0 as m → ∞. This implies, in particular, that f ∈ Lp since kf kp ≤
kf − fn kp + kfn kp .
Exercise 91. If 1 ≤ p < ∞ then Lp (Rn ) is separable. However, L∞ (Rn ) is not.
Proof. Consider the set S of simple functions of the form ni=1 qi χEi where qi ∈ Q and Ei
P
is a rectangle with vertices at rational coordinates. This set is countable and with a bit of
effort, you can show that it is dense in Lp (Rn ).

18.1 Banach spaces

Definition 31. A vector space V with a function k · k : V → [0, ∞) (called the norm) is a
Banach space if

• k · k really is a norm. This means that kcvk = |c|kvk for every scalar c and vector v
and kv + wk ≤ kvk + kwk and kvk = 0 if and only if v = 0.

• V is complete with respect to the metric d(v, w) = kv − wk.

We have already proven that each Lp (X, µ) is a Banach space if 1 ≤ p ≤ ∞.

Of course, Rn , Cn , `p , C(X), C0 (X) are also Banach spaces (where e.g. X is a topological
space and C(X) is the space of continuous functions with norm kf k = supx |f (x)|).
Exercise 92. If X is a topological space, the C(X) is a Banach space.
Proof. If {fn } ⊂ C(X) is Cauchy, then for each x ∈ X, {fn (x)} is Cauchy in R (or C). So
f (x) = limn fn (x) exists for every x. Standard results imply f is continuous.

Definition 32. If V is a Banach space, then V ∗ is the set of all linear maps f : V → C (or
R whichever is the scalar field), such that

sup |f (v)| < ∞.

kvk=1

35
V ∗ is called the dual space. Its elements are bounded linear functionals of V . For f ∈ V ∗ , let

kf k := sup |f (v)|.
kvk=1

Definition 33. A linear map f : V → C is in V ∗ if and only if it is continuous.

Exercise 93. If V is a Banach space then V ∗ is also a Banach space.

Proof. It is elementary to check that k · k is a norm. Suppose {fn } is a Cauchy sequence. By
the previous exercise, we may regard fn as a continuous function on V1 , the unit norm ball of
V . That is, V ∗ ⊂ C(V1 ). Note that the norm on V ∗ coincides with the norm it inherits from
C(V1 ). So {fn } is Cauchy as a sequence in C(V1 ). Since C(V1 ) is complete, {fn } converges
to f ∈ C(V1 ). It is easy to check that f must be linear.
Exercise 94. Suppose 1 ≤ p ≤ q ≤ ∞ are conjugate. Show that Lq (X) naturally sits inside
the dual Lp (X)∗ . Moreover, the map from Lq (X) into Lp (X)∗ is an isometric embedding.
In case q = ∞, assume the measure is semi-finite (this means every subset Y ⊂ X with
µ(Y ) > 0 contains a subset Y 0 ⊂ Y with 0 < µ(Y 0 ) < ∞).
Proof. Case 1. Suppose 1 < p, q < R∞.
Given φ ∈ Lq (X), the map f 7→ f φ dµ is a bounded linear function on Lp (by Hölder’s
inequality). This shows Lq (X) embeds into Lp (X)∗ . However, you may wonder: why is it
an isometric embedding? In other words, we have to prove:
R
| f φ dµ|
kφkq = sup .
f ∈Lp ,f 6=0 kf kp

First assume that φ ≥ 0 and kφkq = 1. Define f (x) = φq−1 . Observe that f ∈ Lp ,
Z Z
p pq−p
kf kp = φ dµ = φq dµ = 1
R R
and φf dµ = φq dµ = 1. This proves it.
Now suppose kφkq = 1 but don’t assume φ ≥ 0. Set f (x) = |φq−1 (x)|u(x) where u is
chosen so that u(x) = 0 if φ(x) = 0 and u(x) = φ(x)
φ(x)
otherwise. The same argument as above
works.
For the general case, we observe that the statement is scale-invariant. So we are done
with the case 1 < p, q < ∞.
Case 2. Suppose q = 1 and p = ∞.
φ(x)
Define f ∈ L∞ by f (x) = 0 of φ(x) = 0 and f (x) = |φ(x)| otherwise. Note that f (x)φ(x) =
R
|φ(x)|. Therefore f (x)φ(x) dµ(x) = kφk1 which proves this case.
Case 3. Suppose q = ∞ and p = 1.
We assume wlog that kφk∞ = 1. For n > 0 let En = {x ∈ X : |φ(x)| > 1 − 1/n}. Note
µ(En ) > 0. Because µ is semifinite, there exists a subset En0 ⊂ En with 0 < µ(En0 ) < ∞.

36
 φ(x)
/ En0 and fn (x) = |φ(x)|µ(E
Define fn (x) = 0 if x ∈ 0 if x ∈ En0 . Then
n)
Z Z
fn (x)φ(x) dµ(x) = |φ(x)|/µ(En0 ) dµ(x) ≥ (1 − 1/n).
0
En

Note that Z
1
kfn k1 = dµ = 1.
0
En µ(En0 )
Therefore, R
| f φ dµ|
1 = sup
f ∈L1 ,f 6=0 kf k1
as required.

Later we will show that Lq = (Lp )∗ (unless p = ∞).

19 Hilbert spaces
For f, g ∈ L2 (X), define the inner product:
Z
hf, gi = f ḡ dµ.

By Schwarz’ inequality,
|hf, gi| ≤ kf k2 kgk2 .
In particular, it is finite.
Exercise 95. • the inner product is sesquilinear: for any f, g, h ∈ L2 (X),

hf, g + hi = hf, gi + hf, hi,

hf + g, hi = hf, hi + hg, hi.

Also if c is a constant then
hcf, gi = chf, gi,
hf, cgi = c̄hf, gi.

• hf, f i = kf k22 .

• hv, wi = hw, vi.

More generally,
Definition 34. A Hilbert space is a Banach space H with an inner product satisfying the
conditions of the previous exercise. So L2 (X) is a Hilbert space. Also `2 is also a Hilbert
space.

37
 Most of the results we will consider are true for an arbitrary Hilbert space. So from now
on, fix a Hilbert space H.
Definition 35. We say v, w ∈ H are orthogonal if hv, wi = 0. A subset B ⊂ H is orthonor-
mal if kvk = 1 for every v ∈ B and every pair of distinct vectors in B are orthogonal. If
B ⊂ H is any set then its span is the set of all finite linear combinations of the form b∈B cb b
P
(with cb ∈ C). We say B is a basis if its span is dense in H.
Exercise 96. If E ⊂ Rn has positive measure then L2 (E) has a countable orthonormal basis.
Moreover, every orthonormal basis of L2 (E) is countable.
Proof. By a previous exercise, L2 (E) is separable. Let S = {s1 , s2 , . . .} ⊂ L2 (E) be a
countable dense subset. Let S 0 = {s01 , s02 , . . .} be defined inductively by: s01 = s1 and s0i = sj
where j is the smallest number so that sj is not contained in the span of {s01 , . . . , s0j−1 }.
Define vectors vi inductively by the Gram-Schmidt process:
s1
v1 = ,
ks1 k
P
si − j<i hsi , vj ivj
vi = P .
ksi − j<i hsi , vj ivj k
Then the vi ’s are orthonormal. They form a basis because they have the same span as S.
Exercise 97. Let {vi }ni=1 be a finite ON set and x ∈ H. Let y = ni=1 hx, vi ivi . Then
P

• (x − y) ⊥ vi for all i.
• (x − y) ⊥ z for every z ∈ Span(v1 , . . . , vn ).
• y is the unique closest point to x with y ∈ Span(v1 , . . . , vn )
Proof. It’s easy to check that (x − y) ⊥ vi for all i which implies (x − y) ⊥ z for every
z ∈ Span(v1 , . . . , vn ). Note that, in general, if z1 ⊥ z2 then kz1 − z2 k2 = kz1 k2 + kz2 k2 . So if
w ∈ Span(v1 , . . . , vn ) then

kx − wk2 = hx − y + y − w, x − y + y − wi
= kx − yk22 + ky − wk22 .
Thus kx − wk2 ≥ kx − yk2 with equality iff y = w.

Exercise 98. {vi } is an ON basis for H if and only if every x ∈ H can be uniquely expressed
in the form X
x= hx, vi ivi
i
for some coefficients ci := hx, vi i ∈ C, called the Fourier coefficients. Moreover,
X
kxk2 = |ci |2 .
i

38
Proof. Let ci = hx, vi i. Let yn = ni=1 ci vi . The previous exercise implies that yn is the
P
closestPpoint to x on Span(v1 , . . . , vn ). Because {vi } isP a basis, we must have lim Pn yn = x.
Thus ∞ c
i=1 i iv = x as required. Observe that ky n k2
= n
|c
i=1 i |2
implies kxk 2
= |c
i i |2
.
P Why is
P this representation unique? Well, if there were two such representations v =
i ci v i = i di vi then we could subtract them to obtain a representation of 0 of the form
X
0= (ci − di )vi .
i

However this implies

X X
0 = k0k2 = k (ci − di )vi k2 = |ci − di |2 .
i i

So ci = di for all i.
Exercise 99. If E ⊂ R has positive measure then L2 (E) is isomorphic as a Hilbert space to
`2 .
Exercise 100. Show that {(2π)−1/2 einx }n∈Z is an orthonormal set in L2 ([0, 2π), m).
Theorem 19.1 (Stone-Weierstrauss). Let X be a locally compact Hausdorff space and A ⊂
C0 (X) an algebra. Suppose A separates points and vanishes nowhere (i.e. for any x, y ∈ X
with x 6= y there exists f, g ∈ A with f (x) 6= f (y) and g(x) 6= 0). Then A is dense in C0 (X).
Exercise 101. Using the above theorem prove that {(2π)−1/2 einx }n∈Z is a basis in L2 ([0, 2π), m).

20 Signed measures
Definition 36. A signed measure on a measurable space (X, B) is a function ν : B →
R ∪ {±∞} such that
• ν assumes at most one of the values −∞, +∞
P∞
• ν(∪∞i=1 Ei ) = i=1 ν(Ei ) for any sequence {Ei } of pairwise disjoint measurable sets.

Definition 37. If ν is a signed measure then E ∈ B is

• positive if ν(E 0 ) ≥ 0 for every measurable subset E 0 ⊂ E
• negative if ν(E 0 ) ≤ 0 for every measurable subset E 0 ⊂ E
• null if ν(E 0 ) = 0 for every measurable subset E 0 ⊂ E.
Exercise 102. Let E be a measurable subset with 0 < ν(E) < ∞. Then there is a positive
set A ⊂ E with ν(A) > 0.
Hint: If E is not positive, then let n1 be the smallest positive integer such that there is
a set E1 ⊂ E with ν(E1 ) < −1/n1 . Inductively define nk to be the smallest positive integer
such that there is a set Ek ⊂ E \ ∪i<k Ei with ν(Ek ) < −1/nk (assuming E \ ∪i<k Ei is not
already positive). Show that A = E \ ∪i Ei satisfies the exercise.

39
Proof. Following the hint, we observe that ν(∪i Ei ) < i −1
P
ni
. Since ν(E) = ν(∪i Ei ) + ν(A)
is finite we must have that ν(∪i Ei ) is also finite, ν(A) > 0 and ni → +∞ as i → ∞.
To show A is positive, let B ⊂ A be measurable. So B is in the complement of ∪i Ei .
By definition of Ei this means that ν(B ∪ Ei ) = ν(B) + ν(Ei ) ≥ −1/(ni − 1). So ν(B) ≥
−1/(ni − 1) − ν(Ei ) ≥ −1/(ni − 1). Since ni → +∞, this implies ν(B) ≥ 0.
Since B is arbitrary, A is positive.
Exercise 103 (Hahn Decomposition Theorem). Let ν be a signed measure on (X, B). Then
there are a positive set P and a negative set N such that X = P ∪ N , P ∩ N = ∅.
Hint: let λ be the sup of ν(A) over all positive sets A. Observe that a countable union
of positive sets is positive.
Proof. We assume wlog that +∞ = 6 ν(A) for any A. Following the hint, we obtain positive
set Ai with ν(Ai ) ≥ λ − 1/i and set P = ∪i Ai and N = X \ P .
Definition 38 (Jordan decomposition). Let ν be a signed measure and P, N be sets as
above. Let ν + , ν − be the measures defined by
ν + (E) = ν(E ∩ P ), ν − (E) = −ν(E ∩ N ).
Observe that these are measures on (X, B) in the usual sense (non-signed). Moreover,
ν = ν + − ν −.
This is called the Jordan decomposition of ν. The total variation of ν is
|ν| = ν + + ν − .
Definition 39. Let ν, µ be two measures on X. We say
• ν is absolutely continuous to µ, denoted ν << µ if for every measurable E, µ(E) =
0 ⇒ ν(E) = 0;
• ν is singular to µ, denoted ν ⊥ µ, if there exists a measurable set A such that ν(A) =
0, µ(Ac ) = 0.
Exercise 104. Prove that the Jordan decomposition of ν is unique in the sense that if ν =
µ1 − µ2 for two positive measures µ1 , µ2 that are mutually singular then ν + = µ1 , ν − = µ2 .
Proof. Suppose ν = µ1 − µ2 . Because µ1 , µ2 are singular, there exists a decomposition
X = S1 ∪ S2 such that µi (Si+1 ) = 0. If µ1 (N ) > 0 then µ1 (N ∩ S1 ) > 0 which implies
ν(N ∩ S1 ) = µ1 (N ∩ S1 ) > 0
a contradiction. So µ1 (N ) = 0. Similarly µ2 (P ) = 0. Similarly, ν + (S2 ) = 0 = ν − (S1 ). If
E ⊂ X is measurable then
ν(E ∩ P ) = ν + (E ∩ P ) = ν + (E) = µ1 (E ∩ P ) = µ1 (E)
ν(E ∩ N ) = ν − (E ∩ N ) = ν − (E) = µ2 (E ∩ N ) = µ2 (E).

40
21 Radon-Nikodym Theorem
Before proving the Radon-Nikodym Theorem (below) we will need two exercises.
Exercise 105. For motivation, suppose f : X → R is a measurable real-valued function. For
each rational number r we may consider the sets f −1 ((−∞, r]). Do these sets determine f ?
In other words, if g : X → R is another measurable function such that µ(f −1 ((−∞, r]) M
g −1 ((−∞, r])) = 0 for every rational r, then does g = f a.e. ?
Proof. Yes. To obtain a contradiction suppose there is an  > 0 such that the set E =
{x ∈ X : f (x) − g(x) > } has positive measure. There must exist rational r0 such that
f −1 ((r0 , r0 + ]) ∩ E has positive measure. However g −1 ((r0 − , r0 ]) ∩ E ⊃ f −1 ((r0 , r0 + ]) ∩ E
which implies that g −1 (−∞, r0 ] is not the same as f −1 ((−∞, r0 ]) (even up to measure zero).

Exercise 106. Can we reverse the previous exercise? Suppose we are given a collection of
measurable sets {Xr }r∈Q satisfying r < s ⇒ Xr ⊂ Xs and X = ∪r Xr . Is there a measurable
function f such that f ≤ r on Xr and f ≥ r on Xrc ? If there is one, is it unique?
Proof. Define f (x) to be the infimum of r over all r such that x ∈ Xr . To see that f is
measurable, note that for any r > 0

f −1 (−∞, r] = (∪s≤r Xs ) ∪ (∩t>r Xt ).

Clearly, f satisfies the exercise. Suppose g is another such function. If g 6= f a.e. then
there is a positive measure set E ⊂ X and disjoint closed intervals If , Ig such that f (x) ∈ If
for x ∈ E and g(x) ∈ Ig for x ∈ Ig . Let r be strictly between If and Ig . Wlog If < r < Ig .
Then E ⊂ Xr (because If < r) but E is not a subset of Xr (because r < Ig ). This
contradiction proves it.
Exercise 107 (The Radon-Nikodym Theorem). Suppose ν << µ and µ is σ-finite. Then
there exists a measurable function f on X with f ≥ 0 and
Z
ν(E) = f dµ
E

for all measurable E. This function is unique up to measure zero.

Proof. Wlog we may assume µ(X) < ∞.
For each rational r, consider the signed measure ν − rµ. By the Hahn Decomposition
Theorem, we may write X = Pr ∪ Nr where Pr ∩ Nr = ∅ are measurable sets such that Pr
is positive for ν − rµ and Nr is negative.
We don’t yet know that these sets are nested so we can’t yet use the previous exercise.
However, note that if s < r then

0 ≥ ν − sµ(Ns \ Nr ) ≥ ν − rµ(Ns \ Nr ) ≥ 0.

41
So after ignoring a measure zero set, we can in fact assume that these are nested. So
s < r ⇒ Ns ⊂ Nr .
So the previous exercise implies there is a function f such that f ≤ r on Nr and f ≥ r
on Nrc = Pr .
Without loss of generality, we may assume f ≥ 0 since ν ≥ 0 implies we can choose
P0 = X.
Let k > 0 be an integerSand let En = Pk/n ∩ N(k+1)/n . Let E∞ be the complement of
∪n En in X. So X = E∞ ∪ n En .
Let S ⊂ X be measurable with µ(S) < ∞. Because En ⊂ Pk/n we have

ν(S ∩ En ) − (k/n)µ(S ∩ En ) ≥ 0

which implies ν(S ∩ En ) ≥ (k/n)µ(S ∩ En ). Because En ⊂ N(k+1)/n we have

(k/n)µ(S ∩ En ) ≤ ν(S ∩ En ) ≤ ((k + 1)/n)µ(S ∩ En ).

Because
k/n ≤ f (x) ≤ (k + 1)/n
on Pk/n ∩ N(k+1)/n we have
Z
k/nµ(S ∩ En ) ≤ f dµ ≤ (k + 1)/nµ(S ∩ En ).
S∩En

Thus Z
|ν(S ∩ En ) − f dµ| ≤ (1/n)µ(S ∩ En ).
S∩En

Adding up over n we have

Z
|ν(S ∩ (X \ E∞ )) − f dµ| ≤ (1/n)µ(S).
S∩(X\E∞ )

Since n is arbitrary and µ(S) < ∞ this shows

Z
|ν(S ∩ (X \ E∞ )) − f dµ.
S∩(X\E∞ )

Now observe that f = +∞ on E∞ . If µ(E∞ ) = 0 then we’re done of course since ν << µ.
If µ(E∞ ∩ S) > 0 then ν(E∞ ∩ S) = +∞ since ν(E∞ ∩ S) − rµ(E∞ ∩ S) ≥ 0 for every r.
Also Z
f dµ = +∞ = ν(E∞ ∩ S).
E∞ ∩S

So we’re done for µ(E) < ∞. The general case follows from this case by countable
R additivity.
Suppose g is another function satisfying the exercise. Then 0 = E f − g dµ for all
measurable sets E. By a previous exercise, this implies f = g a.e.

42
Definition 40. The function f appearing in the previous exercise is called the Radon-
Nikodym derivative and sometimes it is denoted by
dν
f (x) = (x).
dµ

Exercise 108 (Lebesgue Decomposition Theorem). Let µ, ν be σ-finite measures on (X, B).
Then there exist unique measures νac , νsing such that

ν = νac + νsing

νac << µ and νsing ⊥ µ.

Proof. Let λ = µ+ν. Observe that µ << λ and ν << λ. So there exist measurable functions
f, g such that Z Z
µ(E) = f dλ, ν(E) = g dλ
E E

for all measurable E. Let A = {x ∈ X : f (x) > 0} and B = {x ∈ X : f (x) = 0}. Then
X = A ∪ B, A ∩ B = ∅. Define

νac (E) = ν(E ∩ A), νsing (E) = ν(E ∩ B).

The rest is clear.

22 The Dual of Lp
Exercise 109. Let 1 ≤ p ≤ q ≤ ∞ be conjugate exponents. Recall from a previous exercise
that Lq (X) isometrically embeds into Lp (X)∗ . Prove that this is onto. That is Lq (X) is
canonically isomorphic to Lp (X)∗ .
Hint: suppose Φ ∈ Lp (X)∗ . We may assume wlog that the Lp spaces are real (as opposed
to complex). Handle the case µ(X) < ∞ first. Define a set function ν by

ν(E) = Φ(χE ).

Prove that ν is a signed measure. Apply the Radon-Nikodym Theorem to the measures ν +
and ν − from its Jordan Decomposition.
Proof. First we show ν is a measure. Let {Ei } be pairwise disjoint measurable sets. Define
f on X by X
f (x) = sign(Φ(χEi ))χEi (x)
i

and fn by
n
X
fn (x) = sign(Φ(χEi ))χEi (x).
i=1

43
Because µ(X) < ∞, f ∈ Lp (X) and fn → f in Lp (X) (by the bounded convergence theorem).
Because Φ is continuous, Φ(fn ) → Φ(f ). Thus
n
X X
Φ(f ) = lim Φ(fn ) = lim |ν(Ei )| = |ν(Ei )| < ∞.
n n
i=1 i

Thus X
ν(E) = ν(Ei )
i

(again using continuity of Φ and the absolute summability of ν(Ei )). This shows ν is a
signed measure.
By the Radon-Nikodym Theorem (and the Jordan Decomposition), there is a measurable
function g such that Z
Φ(χE ) = ν(E) = g dµ.
E
if f is any simple function then we must have
Z
Φ(f ) = f g dµ.

If g is bounded then g ∈ Lq . In this case, we have that Φ = Φg onRthe dense set of simple
functions (where Φg ∈ (Lp )∗ is the linear functional given by f 7→ f g dµ). Because both
Φ and Φg are continuous, this implies Φ = Φg and kΦk = kgkq (by a previous exercise).
Now suppose g is not necessarily bounded. Let En = {x ∈ X : |g(x)| ≤ n}. We observe

kΦk ≥ kΦ  Lp (En )k = kg  En kq .

Here  means “restricted to”. For example, kΦ  Lp (En )k is the operator norm of the
restriction of Φ to Lp (En ) which we may regard as a subspace of Lp (X). So kΦ  Lp (En )k =
sup{|Φ(f )| : f ∈ Lp (En ), kf kp ≤ 1}. The reason that kΦ  Lp (En )k = kg  En kq is that g is
bounded on En . So the results of the previous paragraph apply.
As n → ∞, kg  En kq → kgkq (since ∪n En = X). Therefore, kΦk ≥ kgkq , and in
particular g ∈ Lq . Once again, Φ = Φg on the dense set of simple functions implies Φ = Φg
by continuity.
It is clear that g is unique (up to measure zero).
Now extend to the σ-finite case: suppose X = ∪i Xi where µ(Xi ) R< ∞. Then for each
i there is a unique gi ∈ Lq (Ei ) such that for any f ∈ Lp , Φ(χXi f ) = Ei f gi dµ. Note that
gi+1 = gi a.e. on Ei . So we can define g on X by g = gi on Ei . Note that kgi kq ≤ kΦk. So
taking the limit as i → ∞ we obtain kgkq ≤ kΦk < ∞ (this uses the monotone convergence
theorem). In particular,
R g ∈ Lq . Also using Lebesgue’s Dominated Convergence Theorem,
we obtain Φ(f ) = f g dµ for any f ∈ Lp , finishing it.

Exercise 110. Show that L1 (R) is not the dual of L∞ (R).

44
23 Outer measure
Definition 41. An outer measure on a set X is a function µ∗ defined on all subsets of X
satisfying
• µ∗ (∅) = 0

• A ⊂ B ⇒ µ∗ (A) ≤ µ∗ (B)

• E ⊂ ∪i Ei ⇒ µ ∗ (A) ≤ i µ∗ (Ei ).
P

Definition 42. A subset E ⊂ X is measurable with respect to an outer measure µ∗ if for

every A ⊂ X,
µ∗ (A) = µ∗ (A ∩ E) + µ∗ (A \ E).
Exercise 111. The class B is µ∗ -measurable sets is a sigma-algebra. Hints:
1. First show that if E1 , E2 ∈ B then E1 ∪ E2 ∈ B.

2. Next, show that if E1 , . P

. . , En are pairwise disjoint and measurable and A is arbitrary
then µ∗ (A ∩ ∪ni=1 Ei ) = i µ∗ (A ∩ Ei ).

3. Next show that if {Ei } are pairwise disjoint measurable sets then ∪i Ei is measurable.
Proof. Clearly, ∅ ∈ B and E ∈ B ⇒ E c ∈ B. So it suffices to show B is closed under
countable unions. First we will show it is closed under finite unions.
Let E1 , E2 ∈ B. By subadditivity, we have

µ∗ (A) ≤ µ∗ (A ∩ E) + µ∗ (A \ E)

for any A, E. So it suffices to prove the opposite inequality when E = E1 ∪ E2 and A is

arbitrary.
Because E2 is measurable,

µ∗ (A) = µ∗ (A ∩ E2 ) + µ∗ (A ∩ E2c ).

Because E1 is measurable

µ∗ (A ∩ E2c ) = µ∗ (A ∩ E2c ∩ E1 ) + µ∗ (A ∩ E2c ∩ E1c ).

So we have
µ∗ (A) = µ∗ (A ∩ E2 ) + µ∗ (A ∩ E2c ∩ E1 ) + µ∗ (A ∩ E2c ∩ E1c ).
Note
A ∩ (E1 ∪ E2 ) = (A ∩ E2 ) ∪ (A ∩ E1 ∩ E2c ).
By subadditivity,

µ∗ (A ∩ (E1 ∪ E2 )) ≤ µ∗ (A ∩ E2 ) + µ∗ (A ∩ E1 ∩ E2c ).

45
So
µ∗ (A) ≥ µ∗ (A ∩ (E1 ∪ E2 )) + µ∗ (A ∩ E2c ∩ E1c ).
Since (E1 ∪ E2 )c = E1c ∩ E2c this proves E1 ∪ E2 is measurable.
Note that E1 \ E2 = E1 ∩ E2c = (E1c ∪ E2 )c . So relative complements of measurable sets
are measurable. For this reason, it suffices now to show that if {Ei } be a sequence of disjoint
measurable sets then the union ∪i Ei is measurable.
Let Gn = ∪ni=1 Ei . Then Gn is measurable and

µ∗ (A) = µ∗ (A ∩ Gn ) + µ∗ (A ∩ Gcn ) ≥ µ∗ (A ∩ Gn ) + µ∗ (A ∩ Gc∞ ).

Unfortunately, we don’t know that µ∗ is continuous, so we can’t just take a limit as n → ∞.

However, because En is measurable,

µ∗ (A ∩ Gn ) = µ∗ (A ∩ Gn ∩ En ) + µ∗ (A ∩ Gn ∩ Enc )
= µ∗ (A ∩ En ) + µ∗ (A ∩ Gn−1 ).

By induction,
n
X
∗
µ (A ∩ Gn ) = µ∗ (A ∩ Ei ).
i=1

So n
X
µ∗ (A) ≥ µ∗ (A ∩ Gc∞ ) + µ∗ (A ∩ Ei ).
i=1

Now we take a limit as n → ∞ to obtain

∞
X
µ∗ (A) ≥ µ∗ (A ∩ Gc∞ ) + µ∗ (A ∩ Ei ) ≥ µ∗ (A ∩ Gc∞ ) + µ∗ (A ∩ G∞ )
i=1

where the last inequality uses subadditivity.

Exercise 112. Show B is complete in the sense that if E ∈ B has µ∗ (E) = 0 then every
subset of E is measurable.
Proof. Let F ⊂ E and A ⊂ X be arbitrary. Then

µ∗ (A) ≥ µ∗ (A ∩ F c ) = µ∗ (A ∩ F ) + µ∗ (A ∩ F c ).

The first inequality is from monotonicity and the second is also from monotonicity (since
A ∩ F ⊂ F ⊂ E). Subadditivity gives the oppositive inequality and thereby proves F is
measurable.
Exercise 113. If µ = µ∗  B then µ is a measure. Hint: first show finite additivity.

46
Proof. If E1 , E2 ∈ B are disjoint then, because E2 is measurable,
µ(E1 ∪ E2 ) = µ((E1 ∪ E2 ) ∩ E2 ) + µ((E1 ∪ E2 ) \ E2 ) = µ(E2 ) + µ(E1 ).
So we have finite addiitivity.
Suppose {Ei } ⊂ B are pairwise disjoint. Then
n
X
µ(∪∞
i=1 Ei ) ≥ µ(∪ni=1 Ei ) = µ(Ei ).
i=1

By taking limits we have

∞
X
µ(∪∞
i=1 Ei ) ≥ µ(Ei ).
i=1
The oppositive inequality follows from subadditivity.

24 Carathéodory’s Extension Theorem

Suppose A is algebra on a set X and µ0 is like a measure on A in the sense that µ0 : A →
[0, ∞) is countably
P additive: if E1 , E2 , . . . ∈ A are pairwise disjoint and if ∪Ei ∈ A then
µ0 (∪i Ei ) = i µ0 (Ei ). Can we promote µ0 to a measure on the sigma-algebra generated
by A? The extension theorem says ‘yes’. We will use the extension theorem to construct
product measures and (if there’s time in the semester) to prove the Riesz-Markov Theorem
characterizing positive linear functionals on Cc (X) for locally compact spaces X in terms
of measures. It can also be used to construct Hausdorff measure and Haar measure but we
won’t do that (not enough time in the semester).
To begin, we use µ0 to construct an outer measure. Define µ∗ by:
X
µ∗ (A) = inf µ0 (Ei )
i

where the infimum is taken over all collections {Ei } of sets Ei ∈ A with A ⊂ ∪i Ei . We need
to show that µ∗ is an outer measure. But first:
Exercise 114. If A ∈ A then µ∗ (A) = µ0 (A).
Proof. if {Ei } ⊂ A satisfies A ⊂ ∪i Ei then we’d like to say µ0 (A) ≤ i µ0 (Ei ). Unfortu-
P
nately, we don’t know ∪i Ei ∈ A so we can’t do this directly. So let Bn = (A ∩ En ) \ (∪i<n Ei ).
Then Bn ∈ A and A = ∪n Bn is a disjoint union. So
X X
µ0 (A) = µ0 (Bi ) ≤ µ0 (Ei )
i i

(the last inequality occurs because Bi ⊂ Ei . Since {Ei } is arbitrary, this implies µ0 (A) ≤
µ∗ (A). The opposite inequality is also true since we can take E1 = A and Ei = ∅ for all
i > 1.

47
Exercise 115. µ∗ is an outer measure.
Proof. It’s clear that µ∗ (∅) = 0 and µ∗ is monotone. To show countable subadditivity, let
∗
P ∗
E ⊂ X and let {Ei } be subsets with E ⊂ ∪i Ei . We must show µ (E) ≤ i µ (Ei ). If
µ(Ei ) = +∞ for some i then we are done. So we can asume µ∗ (Ei ) < ∞ for each i. Let
 > 0. By definition of µ∗ there exist sets Aij ∈ A such that Ei ⊂ ∪j Aij and
X
µ∗ (Ei ) ≥ −/2i + µ0 (Aij ).
j

Since the sets {Aij } cover E, we have

X X X
µ(E) ≤ µ0 (Aij ) ≤ µ∗ (Ei ) + /2i ≤  + µ∗ (Ei ).
ij i i

Since  is arbitrary, this implies the exercise.

Exercise 116. Every set in A is measurable.
Proof. Let E ⊂ X be arbitrary and A ∈ A. Because of subadditivity, it suffices to show

µ∗ (E) ≥ µ∗ (E ∩ A) + µ∗ (E \ A).

We can assume wlog that µ∗ (E) < ∞. Let  >P0. By definition of µ∗ there exist sets
{Ai } ⊂ A such that E ⊂ ∪i Ai and µ∗ (E) > − + i µ(Ai ).
Because µ restricted to A is additive, we have

µ(Ai ) = µ(Ai ∩ A) + µ(Ai \ A)

for each i. Adding up over i gives

X
µ∗ (E) > − + µ(Ai ∩ A) + µ(Ai ∩ Ac ) ≥ − + µ∗ (E ∩ A) + µ∗ (E ∩ Ac )
i

where in the last line we use subadditivity and monotonicity. Since this is true for every ,
we are done.
We now know that we can extend µ0 originally defined on A to a measure µ defined on
a sigma-algebra containing A. Is this extension unique?
Exercise 117. Let Aσ be the collection of all countable unions of sets in A and Aσδ the
collection of all countable intersections of sets in Aσ . Suppose µ is σ-finite. Then for every
µ∗ -meaurable set E there exists A ∈ Aσδ with E ⊂ A and µ(A \ E) = 0.

48
Proof. Since µ is sigma-finite, X = ∪i Xi where µ(Xi ) < ∞.
By definition of µ, for every n there exist sets An,i,j ∈ A such that
∞
X
−i
−2 /n + µ(An,i,j ) ≤ µ(E ∩ Xi )
j=1

and E ∩ Xi ⊂ ∪j An,i,j .
So if An,i = ∪j An,i,j then An,i ∈ Aσ and −2−i /n + µ(Ani ) ≤ µ(E ∩ Xi ). Now let
An = ∪i An,i . Then An ∈ Aσ , E ⊂ An and
∞
X
µ(An \ E) ≤ µ(An,i \ E ∩ Xi ) ≤ 1/n.
i=1

Finally let A = ∩n An . Then A ∈ Aσδ , E ⊂ A and we have µ(A \ E) ≤ 1/n for all n (by
monotonicity) which implies µ(A \ E) = 0.
Exercise 118. If µ0 is σ-finite and B is the smallest sigma-algebra containing A then µ  B
is the unique measure on B that agrees with µ0 on A.
Proof. Let ν be a measure on B that agrees with µ0 on A. By the previous exercise it is
enough to show that for every E ∈ Aσδ , ν(E) = µ(E).
i=1 ⊂ A then ∪i Ai can be expressed as a disjoint union of sets in A. Therefore
If {Ai }∞
ν(∪i Ai ) = µ(∪i Ai ). So ν = µ on Aσ .
Consider sets Ai , Bj ∈ A (i, j ∈ N). Note that (∪i Ai ) ∩ (∪j Bj ) = ∪i,j Ai ∩ Bj ∈ Aσ . So
finite intersections of sets in Aσ are in Aσ .
Now if A1 , A2 , . . . ∈ Aσ and Bn = ∩ni=1 Ai then Bn ∈ Aσ and ∩i Ai = ∩n Bn . Moreover
∩n Bn is a decreasing intersection. That is B1 ⊃ B2 ⊃ · · · . Assuming A1 = B1 has finite
measure, then

µ(∩i Ai ) = µ(∩n Bn ) = lim µ(Bn ) = lim ν(Bn ) = ν(∩n Bn ) = ν(∩i Ai ).

n n

In the general case, write X = ∪i Xi where µ(Xi ) < ∞ for all i. Then we have shown
µ(Xi ∩ ∩j Aj ) = ν(Xi ∩ ∩j Aj ) for all i. As we can assume the Xi ’s are pairwise disjoint, this
proves that µ(∩j Aj ) = ν(∩j Aj ) and therefore µ = ν on Aσδ and so µ = ν on B.

If we only assume that µ0 is finitely additive (instead of countably additive) then it

might not extend to a measure on the sigma-algebra generated by A. Indeed there are
finitely additive probability measures on Z (defined on all subsets) that are shift-invariant.
No such measure extends to a countably-additive measure.

49
25 Product measures
Given measure spaces (X, X, µ) and (Y, Y, ν) we’d like to construct a canonical measure
space on the product X × Y . If A ∈ X, B ∈ Y then A × B is a rectangle. Let R denote
the collection of all rectangles. Define λ : R → [0, ∞] by λ(A × B) = µ(A)ν(B) (with the
convention that 0∞ = 0).
We would like to use Caratheodory’s Theorem to extend λ to a measure on X × Y . First
we need to extend it to the algebra generated by R. For that we need:
Exercise 119. If A × B = ti Ai × Bi then
X
λ(A × B) = λ(Ai × Bi ).
i

Proof. Observe that X

ν(Bi )χAi (x) = ν(B)χA (x).
i
By the monotone convergence theorem,
XZ Z X Z
ν(Bi )χAi (x) dµ(x) = ν(Bi )χAi (x) dµ(x) = ν(B)χA (x) dµ(x) = ν(B)µ(A).
P
By direct computation the LHS is i λ(Ai × Bi ).
Let A be the algebra generated by R.
Exercise 120. Every set R ∈ A is a finite union of rectangles.
Proof. It suffices to show that if A0 the set of all finite unions of rectangles, then A0 is an
algebra. Note that
(A1 × B1 ) ∩ (A2 × B2 ) = (A1 ∩ A2 ) × (B1 × B2 )
so R (and thus A0 ) is closed under finite intersections. Also
(A × B)c = Ac × B ∪ A × B c ∪ Ac × B c
So A0 is closed under complementation. Obviously, A0 is closed under finite unions; so this
finishes it.
Exercise 121. Any finite union of rectangles can be expressed as a finite disjoint union of
rectangles.
Proof.
(A1 × B1 ) ∪ (A2 × B2 )
= (A1 ∩ Ac2 ) × (B1 ∩ B2c ) ∪ (A1 ∩ Ac2 ) × (B1 ∩ B2 )
∪(Ac1 ∩ A2 ) × (B1c ∩ B2 ) ∪ (Ac1 ∩ A2 ) × (B1 ∩ B2 )
∪(A1 ∩ A2 ) × (B1 ∩ B2c ) ∪ (A1 ∩ A2 ) × (B1c ∩ B2 ) ∪ (A1 ∩ A2 ) × (B1 ∩ B2 ).

50
Exercise 122. If A × B ∈ R and Ai ×
PBi ∈ R are countably many pairwise disjoint with
A × B = ∪i Ai × Bi then λ(A × B) = i λ(Ai × Bi ).
Proof. For any x ∈ X, X
ν(Bi )χAi (x) = ν(B)χA (x).
i

This is because {x} × B is the disjoint union of {x} × Bi over those i with x ∈ Ai . Now
integrate over x,
Z X Z
ν(Bi )χAi (x) = ν(B)χA (x) dµ(x) = λ(A × B).
i

The Monotone Convergence Theorem allows us to switch the sum and integral in the LHS.
So we obtain XZ X
λ(A × B) = ν(Bi )χAi (x) dµ(x) = λ(Ai × Bi ).
i i

Exercise 123. If R1 , . . . , Rn are pairwise disjoint rectangles and S1 , . . . , Sm are also pairwise
disjoint rectangles and ∪i Ri = ∪j Sj then
X X
λ(Ri ) = λ(Sj ).
i j
P
Proof. Observe that Ri = ∪j Ri ∩ Sj . So λ(Ri ) = j λ(Ri ∩ Sj ) and
X X
λ(Ri ) = λ(Ri ∩ Sj ).
i i,j
P P
Similarly, j λ(Sj ) = i,j λ(Ri ∩ Sj ).
P
If R1 , . . . , Rn are disjoint rectangles then we define λ(∪i Ri ) = i λ(Ri ). The previous
exercise show that this depends only on ∪i Ri . So we have defined λ on the algebra generated
by R. The previous exercise also implies λ is countably additive. So Caratheodory’s extension
Theorem implies that λ extends to a measure on the sigma-algebra containing R. If µ and
ν are sigma-finite then λ is also sigma-finite and in this case λ is uniquely defined. We call
it the product measure and denote it by µ × ν.
Example: note that the direct product of Lebesgue measure on the real line with itself
is Lebesgue measure on R2 .

51
26 Fubini’s Theorem
Theorem 26.1 (Fubini’s Theorem). Let (X, A, µ), (Y, B, ν) be two complete measure spaces
and f be an integrable function on X × Y . Then

1. for a.e. x ∈ X, fx : Y → C, fx (y) = f (x, y) is integrable on Y

2. for a.e. y ∈ Y , f y : X → C, f y (x) = f (x, y) is integrable on X

R
3. Y f (x, y) dν(y) is an integrable function of x
R
4. X f (x, y) dµ(x) is an integrable function of y
R R R R R
5. ( Y f (x, y) dν(y))dµ(x) = X×Y f dµ × ν = ( X f (x, y) dµ(x))dν(y).

We will prove the theorem first for characteristic functions of sets. So if E ⊂ X × Y ,

then we set Ex = {y ∈ Y : (x, y) ∈ E}, E y = {(x, y) : (x, y) ∈ E}. These are called cross
sections.
Recall that the collection of measurable sets of X ×Y is the smallest σ-algebra containing
all of the rectangles which is complete wrt µ × ν.
Exercise 124. If E ⊂ X × Y is in the sigma-algebra generated by Ω then Ex and E y are
measurable for a.e. x. Hint: first show this is true for rectangles then the collection of sets
for which it is true is a sigma-algebra.
Proof. Let Ω be the collection of all sets E ⊂ X × Y such that Ex is measurable for a.e. x.
Observe that Ω contains all measurable rectangles Ω ⊃ R. Observe that Ω is a sigma-algebra
because

• X ×Y ∈Ω

• E ∈ Ω ⇒ (E c )x = (Ex )c ⇒ E c ∈ Ω

• if Ei ∈ Ω and E = ∪i Ei then Ex = ∪i (Ei )x hence E ∈ Ω

Exercise 125. Suppose µ, ν are sigma-finite. Then for any measurable E ⊂ X × Y , Ex and
E y are measurable for a.e. x. Moreover,
Z Z
y
µ × ν(E) = µ(E ) dν(y) = ν(Ex ) dµ(x).

Hint: Let Ω be the collection of all measurable sets E satisfying the above. Show Ω
contains the rectangles and finite unions of rectangles. Conclude that Ω contains the algebra
A generated by rectangles. Then show Ω contains Aσ and Aσδ . Use a previous approximation
result to conclude the exercise. It may be easier to handle the case that µ(X) and ν(Y ) are
finite first.

52
Proof. We follow the hint. Clearly Ω contains all rectangles.
If E1 , E2 , . . . ∈ Ω are pairwise disjoint and E = ∪i Ei then
X XZ
µ × ν(E) = µ × ν(Ei ) = µ(Eiy ) dν(y)
Zi X i
Z
= µ(Eiy ) dν(y) = µ(E y ) dν(y)
i

where the third equality follows from the Monotone Convergence Theorem. So E ∈ Ω. Since
every set in the algebra generated by the rectangles is a finite disjoint union of rectangles,
this proves that the algebra A generated by the rectangles is contained in Ω.
If E1 ⊂ E2 ⊂ · · · is an increasing sequence of subsets of Ω then by the Monotone
Convergence Theorem, E := ∪i Ei is also in Ω. So Aσ ⊂ Ω.
Now assume that µ(X) and ν(Y ) are finite.
If E1 ⊃ E2 ⊃ · · · ∈ Ω then ∩i Ei ∈ Ω by Lebesgue’s Dominated Convergence Theorem.
So Aσδ ⊂ Ω.
If µ × ν(E) = 0 then by a previous exercise there exists F ∈ Aσδ with E ⊂ F and
µ × ν(F ) = 0. Since E y ⊂ F y for every y, this implies E ∈ Ω.
Now let E be an arbitrary measurable set. By a previous exercise, there exist F ∈ Aσδ
such that E ⊂ F and µ × ν(F \ E) = 0. So
Z Z
µ × ν(E) = µ × ν(F ) = µ(F ) dν(y) = µ(F y ) − µ((F \ E)y ) dν(y)
y

Z
= µ(E y ) dν(y).

So E ∈ Ω. This proves the exercise when µ × ν(X × Y ) < ∞.

The general case follows from the finite measure case because Ω is closed under countable
disjoint unions.
Exercise 126 (Tonelli’s Theorem). If f ≥ 0 is measurable on X × Y then
ZZ ZZ Z
f (x, y) dµ(x)dν(y) = f (x, y) dν(y)dµ(x) = f dµ × ν.

In particular x 7→ f (x, y) is measurable for a.e. y.

Proof. The previous exercise states that this result is true when f is a characteristic function.
By linearity, it holds for simple functions. By the Monotone Convergence Theorem, it holds
for nonnegative functions.
Exercise 127. Prove Fubini’s Theorem.
Proof. This follows from Tonelli’s Theorem by splitting f into real and imaginary parts and
then splitting each of these parts into positive and negative parts.

53
27 Convolution
Given measurable functions f, g on Rn we define the function f ∗ g by
Z
f ∗ g(x) = f (t)g(x − t) dt

whenever this integral exists. This is called the convolution. Before we go further, there
is one minor detail to contend with: how do we know f ∗ g is measurable? Well suppose we
knew
R that the two variables function F (x, t) = f (t)g(x − t) is measurable. Then f ∗ g(x) =
F (x, t) dt is measurable by Fubini’s Theorem (assuming as we do, that F is integrable).
But how do we know that F is measurable? First of all, the function G(x, t) = f (t)g(x)
is measurable. In fact, since f and g are pointwise limits of simple functions, G is also a
pointwise limit of simple functions. Now F = G◦H where H(x, t) = (t, x−t). H is continuous
but we know that pre-compositions of continuous functions with measurable functions do
not have to be measurable. (Post-compositions are OK). What to do? Well, it is an exercise
to show that any measurable function agrees with a Borel function almost everywhere. (To
prove this, use the lemma before the proof of the Radon-Nikodym Theorem). Compositions
of Borel functions are Borel. So we can find a Borel function G0 that agrees with G almost
everywhere and G0 ◦ H = G ◦ H almost everywhere because H −1 takes measure zero sets to
measure zero sets.
Observe that if s = x − t then
Z Z
f (t)g(x − t) dt = f (x − s)g(s) ds.

So f ∗ g = g ∗ f .
1
Example: suppose f = χ
2 [−,]
. Then
Z 
1
f ∗ g(x) = g(x − t) dt
2 −

is the average value of g in the interval [x − , x + ]. Observe:

1. if g ∈ L1 then f ∗ g ∈ L1 and kf ∗ gk1 ≤ kgk1 .
2. if g ∈ L1 then f ∗ g is continuous!
3. lim&0 f ∗ g(x) = g(x) a.e. Why? (there is a one line answer).
The main technical goal of this section is to generalize this example.
Some motivation:
1. Probability theory: suppose f and g are probability distributions, X, Y ∈ Rn are
chosen independently at random with laws f, g respectively. What’s the distribution
of X +Y ? Of course, it is f ∗g. In particular, if X1 , X2 , . . . , Xn are iid random variables
with distribution f then X1 + · · · + Xn has distribution equal to the n-fold convolution
of f with itself. (These facts are especially easy to verify if we replace Rn with Zn ).

54
 2. Image processing: it is common to ‘clean up’ photographs by convolving with a Gaus-
sian. For example, suppose a satellite image gives the height h(x, y) above sea level of
a point (x, y) on the Earth. This function can be very ‘noisy’. So we replace it with
2 2
h ∗ g where g(x, y) = Ce−x −y is aRGaussian. This has the effect of replacing the value
h(x, y) with the weighted average h(x − t, y − w)g(t, w) dtdw allowing features of the
landscape to be more easily identifiable.

3. Fourier analysis: the Fourier transform of f ∈ L1 (R) is the function

Z
fˆ(t) = f (x)e−ixt dx.

∗ g = fˆĝ. So convolution is transformed into ordinary pointwise

It can be shown that f[
multiplication.

4. PDE’s: for example, solutions to the heat equation can be expressed in terms of
convolution. For example, let
1
Φt (x) = √ exp(−x2 /4t).
4πt
Let f (x) denote an initial temperature distribution on R. Let u(x, t) denote the tem-
perature distribution at time t ≥ 0 (so u(x, 0) = f (x)). Then
Z
u(x, t) = f ∗ Φt (x) = f (y)Φt (x − y) dy.

On physical considerations you might expect that u(x, t) is smooth in the spatial
variable for t > 0. This is correct if, for example, f ∈ L1 . (BTW Φ is the solution to
the heat equation ut = uxx with initial distribution δ0 ).

27.1 Norm inequalities

Exercise 128.
kf ∗ gk1 ≤ kf k1 kgk1
with equality if f ≥ 0 and g ≥ 0.
Proof. Since |f ∗ g| ≤ |f | ∗ |g|, we may assume that f ≥ 0 and g ≥ 0.
ZZ ZZ Z Z 
kf ∗ gk1 = f (t)g(x − t) dtdx = f (t)g(x − t) dxdt = f (t) g(x − t) dx dt
Z  Z 
= f dt g dx = kf k1 kgk1 .

55
 One nice consequence of this formula is that convolution defines a product structure on
L . Indeed, L1 (Rn ) is a Banach algebra: it is a Banach space with a product satisfying
1

associativity, distributivity and kxyk ≤ kxkkyk. Associativity is nontrivial: you should check
this:
Exercise 129. Prove that (f ∗ g) ∗ h = f ∗ (g ∗ h) assuming f, g, h ∈ L1 (Rn ).
Exercise 130. If f ∈ Lp and g ∈ L1 then

kf ∗ gkp ≤ kf kp kgk1 .
1 1
More generally, if r
= p
+ 1q − 1, f ∈ Lp , g ∈ Lq then

kf ∗ gkr ≤ kf kp kgkq .

Proof. Let us just prove the first statement. The second is a homework exercise. We assume
wlog that f, g ≥ 0. The cases p = 1, p = ∞ are easy and left to the reader. Let q be a
conjugate exponent to p. Then
Z Z
f ∗ g(x) = f (t)g(x − t) dt = f (t)g(x − t)1/p g(x − t)1/q dt
Z 1/p Z 1/q
p
≤ f (t)g(x − t) dt g(x − t) dt
1/q
= (f p ∗ g(x))1/p kgk1 .

Raising to the p-power and integrating over x and using the previous exercise we obtain
p/q
kf ∗ gkpp ≤ kf kpp kgk1 kgk1 .

Taking p-th roots finishes it.

27.2 Smoothness
Let K ∈ L1 (Rn ). Then convolution with K defines an operator L1 (Rn ) → L1 (Rn ), f 7→ f ∗K.
This operator has norm = kKk1 by the previous theorems.
One very practical purpose of convolution is as a tool for increasing the regularity of a
function. That is: the convolution of an arbitrary L1 function with a smooth function will
be smooth. So if K is smooth enough then convolution with K defines an operator from
L1 (Rn ) to smooth functions on Rn .
To make this precise, recall that C m (Rn ) is the space of functions all of whose derivatives
of order ≤ m exist and are continuous.
Exercise 131. Let 1 ≤ p < ∞. For any f ∈ Lp (Rn ) and t ∈ Rn let τt (f ) ∈ Lp (Rn ) be the
function τt (f )(x) = f (x − t). Show that

lim kτt (f ) − f kp = 0.
t→0

56
Proof. This is obvious if f is continuous with compact support. Such functions are dense
in Lp (Rn ). Alternatively, it is easy to check that the result is true if f is a characteristic
function (using absolute continuity of the integral). By taking linear combinations, it is true
for simple functions and because simple functions are dense, it is true.
Exercise 132. Let f ∈ Lp (Rn ), K ∈ C m (Rn ) and suppose K has compact support. Then
f ∗ K ∈ C m and
Dα (f ∗ K) = f ∗ Dα K
P
if α = (α1 , . . . , αn ) with |α| = i αi ≤ m.
Hint: work out the case m = 0 first. This means that K is continuous and you have to
show that f ∗ K is continuous.
Proof. Let’s first handle the case m = 0. By changing variables,
Z Z
|f ∗ K(x) − f ∗ K(y)| = | f (t)[K(x − t) − K(y − t)] dt| = | f (x − t)[K(t) − K(t − x + y)] dt|
≤ kf kp kK − τx−y (K)kq .
Because K is continuous and compactly supported, kK − τx−y (K)kq → 0 as x − y → 0. This
proves continuity (in fact, uniform continuity).
Now suppose m > 0. Let ei ∈ Rn denote the i-th basis vector. Then
f ∗ K(x + hei ) − f ∗ K(x) K(x − t + hei ) − K(x − t)
Z
= f (t) dt
h h
Z
∂K
= f (t) (x − t + h0 ei ) dt
∂xi
where h0 ∈ [0, h] is determined by the Mean Value Theorem (h0 depends on t and x). As
∂K
h → 0, ∂x i
(x − t + h0 ei ) → ∂x
∂K
i
(x − t) uniformly in t (because K has compact support). By
the Uniform Convergence Theorem, ∂(f ∗ K)/∂xi = f ∗ (∂K/∂xi ). The general case follows
by induction.

27.3 Approximate identities

As mentioned above, convolution defines a product structure on L1 . Is there a multiplicative
identity element? This would have to be a function g ∈ L1 such that for any f ∈ L1 ,
f ∗ g = f.
Such a thing exists only if you allow g to be a measure instead of a function. To be
precise, if f ∈ L1 and µ is a finite measure on Rn then their convolution µ ∗ f is the function
defined by Z
µ ∗ f (x) = f (x − t) dµ(t).
Let δ0 be the Dirac measure concentrated on 0. Then δ0 ∗ f = f . So δ0 is the multiplicative
identity. Of course, its not a function. However, we will see that there exist sequences {Kj }
of functions such that f ∗ Kj → f as j → ∞. Convergence will either mean in Lp (for
appropriate p) or pointwise at continuity points.

57
Remark 3. BTW, we can also speak of convolution of measures:
ZZ
µ ∗ ν(E) = χE (x + y) dµ(x)dν(y)

on Rn (or on any group with a measurableR structure). If f ∈ L1 (Rn ) then we can identify f
with the measure given by f (E) = E f dm where m is Lebesgue measure. You can check
that these notions are compatible. To summarize the paragraph above: the convolution
identity is δ0 , a measure. However, we can find functions Kj ∈ L1 such that if we interpret
these functions as measures then they converge (in a weak* sense to be explained in later
lectures) to δ0 . Since convolution is continuous in L1 (and in the weak* topology), it shouldn’t
surprise us that f ∗ Kj → f .
Exercise 133. Suppose {Kj } ⊂ L1 (Rn ) is a sequence satisfying
R
1. Kj dm = 1

2. supj kKj k1 < ∞

3. for every δ > 0, Z

lim |Kj (x)| dm(x) = 0.
j |x|>δ

Then for every f ∈ Lp , 1 ≤ p < ∞

lim f ∗ Kj = f
&0

where the limit is in Lp (Rn ). In other words, kf ∗ Kj − f kp → 0 as  & 0.

Proof.
Z
|f ∗ Kj (x) − f (x)| = [f (x − t) − f (x)]Kj (t) dm(t)
Z
≤ |(f (x − t) − f (x))Kj (t)| dm(t)
Z
1 1
= |f (x − t) − f (x)||Kj (t)| p |Kj (t)| q dm(t)
Z 1/p Z 1/q
p
≤ |f (x − t) − f (x)| |Kj (t)| dm(t) |Kj (t)| dm(t)
Z 1/p
p 1/q
= |f (x − t) − f (x)| |Kj (t)| dm(t) kKj k1 .

58
So for any δ > 0,
Z
kf ∗ Kj − f kpp = |f ∗ Kj (x) − f (x)|p dm(x)
Z Z 
p/q p
≤ kKj k1 |f (x − t) − f (x)| |Kj (t)| dm(t) dm(x)
ZZ
p/q
= kKj k1 |f (x − t) − f (x)|p |Kj (t)| dm(x)dm(t)
Z
p/q
= kKj k1 kτt (f ) − f kpp |Kj (t)| dm(t)
Z
p/q
≤ kKj k1 kτt (f ) − f kpp |Kj (t)| dm(t)
|t|<δ
Z
p/q
+kKj k1 2p kf kpp |Kj (t)| dm(t).
|t|≥δ

Let η > 0. Then there exists δ > 0 such that kτt (f ) − f kpp < η for all |t| < δ. So for this
particular δ we have
Z
p p/q+1 p/q
lim sup kf ∗ Kj − f kp ≤ lim sup kKj k1 η + lim sup kKj k1 2p kf kpp |Kj (t)| dm(t)
j→∞ j→∞ j→∞ |t|≥δ
p/q
≤ lim sup kKj k1 η.
j→∞

Since η is arbitrary, this does it.

Exercise 134. If 1 ≤ p < ∞ then Cc∞ (Rn ) is dense in Lp (Rn ).
Proof. The functions with compact support are dense in Lp ; so it suffices to prove that if
f ∈ Lp has compact support then there exists a sequence of compact supported smooth
functions converging to f (in Lp norm). R
Let K be a compactlyRsupport smooth function with K = 1. For  > 0, let K (x) =
−n K(x/). Observe that K = 1.
Then for any f ∈ Lp with compact support f ∗K → f as  & 0 and f ∗K are compactly
supported and smooth.

28 The space of measures

Let X be a topological space and let C0 (X) denote the space of continuous functions on X
that “vanish at infinity”. This means that for every f ∈ C0 (X) and  > 0 there exists a
compact set K ⊂ X such that for all x ∈/ K, |f (x)| < . The space C0 (X) is a Banach space
with the norm kf k = supx∈X |f (x)|.
Let Cc (X) ⊂ C0 (X) denote the subspace of continuous functions that are compactly
supported. This space is dense in C0 (X).

59
 Suppose µ is a Borel measure on X and f is a continuous function on X. Then we define
Z
µ(f ) := f dµ

whenever this integral exists. This defines a linear functional on Cc (X) (assuming that µ is
finite on compact sets). Moreover if f ≥ 0 then µ(f ) ≥ 0. The Riesz-Markov Theorem is a
converse to this:
Theorem 28.1 (Riesz-Markov Theorem). Let X be a locally compact Hausdorff space. Let
Λ be a positive linear functional on Cc (X). Being positive means that Λ(f ) ≥ 0 for every
nonnegative f ∈ Cc (X). Then there is a unique Radon measure µ on X such that Λ(f ) =
µ(f ) for every f ∈ Cc (X). A Radon measure is a Borel measure which is finite on compact
sets and is inner regular as explained below.
From this result we can obtain a complete picture of C0 (X)∗ :
Theorem 28.2 (Riesz Representation Theorem). For every ρ ∈ C0 (X)∗ there is a finite
signed Borel measure µ such that
ρ(f ) = µ(f ).
One of the applications of these results (that we will not prove) is the existence and
uniqueness of Haar measure. To be precise, let G be a locally compact group (e.g. GL(n, R),
O(n), the isometry group of hyperbolic n-space, the absolute Galois group of a number field,
etc). Then there exists a Radon measure µ on G that is invariant under left multiplication
(µ(gE) = µ(E) for Borel E ⊂ G and g ∈ G). It is unique up to scalar multiplication. For
example, Lebesgue measure is the Haar measure on Rn (as an additive group).
The first issue related to the Riesz-Markov Theorem we will deal with is the uniqueness.
Why should it be true that if µ(f ) = ν(f ) for every f ∈ Cc (X) that µ = ν? This question
is connected with the issue of regularity of a measure explained next.

28.1 Regularity
Definition 43. A measure µ on a topological space X is inner regular if for every Borel set
E,
µ(E) = sup{µ(K) : K ⊂ E}
where the sup is over all compact sets in E. The measure is outer regular if for every Borel
E,
µ(E) = inf{µ(O) : O ⊃ E}
where the inf is over all open sets containing E. The measure is regular if it is both inner
and outer regular. For example, Lebesgue measure on Rn is regular.
A Polish space is a separable complete metric space. A locally compact space is a topo-
logical space in which every point has an open neighborhood whose closure is compact. In
this section we will study Borel measures on a locally compact Polish space (X, d). While
many of the results we obtain generalize beyond this setting we will restrict our attention to
these spaces in order to avoid technicalities.

60
Exercise 135. Suppose X is a compact metric space. Let µ denote a finite Borel measure on
X. Then µ is regular.
Proof. Let Ω denote the collection of all Borel subsets E of X such that

µ(E) = sup{µ(K) : K ⊂ E}

and
µ(E) = inf{µ(O) : O ⊃ E}.
Let E ⊂ X be closed (and therefore compact) and let On = {x ∈ X : d(x, E) < 1/n}. Then
On is open and E = ∩n On . This shows that E ∈ Ω.
Observe that E ∈ Ω if and only if E c ∈ Ω.
Now suppose E1 , E2 , . . . ∈ Ω. Let  > 0 and Ki ⊂ Ei be compact such that µ(Ei \
Ki ) < /2i , µ(Oi \ Ei ) < /2i . Then µ(∪i Ei \ ∪i Ki ) < . So there is some n such that
µ(∪∞ n n
i=1 Ei \ ∪i=1 Ki ) < . Since ∪i=1 Ki is compact and  > 0 is arbitrary, this shows µ(E) =
sup{µ(K) : K ⊂ E}.
Similarly, Ei ⊂ Oi be open such that µ(Ei \ Oi ) < /2i . Then µ(∪i Oi \ ∪i Ei ) < . Since
∪i Oi is open and  > 0 is arbitrary, this shows µ(E) = inf{µ(O) : O ⊃ E}.
So we have shown Ω contains all closed sets and is closed under complementation and
countable unions. So it contains all Borel sets. This proves µ is regular.
Exercise 136. If K ⊂ X is compact and X is locally compact and Polish then there exist
precompact open sets O1 ⊃ O2 ⊃ · · · such that ∩i Oi = K.
Proof. Because K is compact and X is locally compact, K can be covered by a finite number
of open sets whose closure is compact. Taking the union of the sets, we see K ⊂ U for some
open set U whose closure is compact. Let Wn = {x ∈ U : d(x, K) < 1/n}. This is an open
set with compact closure and ∩n Wn = K.
Exercise 137. Suppose X is a locally compact Polish space. Let µ denote a Borel measure
on X such that µ(K) < ∞ for every compact K ⊂ X. Then µ is regular.
Proof. Let E ⊂ X be Borel. Let K1 ⊂ K2 ⊂ .. be increasing compact sets such that
X = ∪i Ki . For each i there exists a compact set Li ⊂ Ei ∩Ki with µ(Li ) > −/2i +µ(E ∩Ki ).
Without loss of generality we may assume L1 ⊂ L2 ⊂ ... So

µ(∪i Li ) = lim µ(Li ) ≥ lim −/2i + µ(Ei ∩ Ki ) = µ(E).

i i

So µ(E) = sup{µ(K) : K ⊂ E}.

Let Oi ⊃ E ∩ Ki be an open subset of Ki with µ(Oi ) ≤ /2i + µ(E ∩ Ki ). Now Oi
need not be open in X. However, because it is open there exists an open set Oi0 ⊂ X with
Oi0 ∩ Ki = Oi .
By the previous exercise there exists a precompact open set Vi ⊂ X such that Ki ⊂ Vi
and µ(Vi \ Ki ) < /2i .

61
 After replacing Oi0 with Oi0 ∩ Vi if necessary, we may assume that µ(Oi0 ) < µ(E ∩ Ki ) +
/2i−1 . So ∪i Oi0 is an open set containing E with µ(∪i Oi0 \ E) < 2. This proves that

µ(E) = inf{µ(O) : O ⊃ E}.

Since E is arbitrary, µ is regular.

Definition 44. A Radon measure on a space X is an inner regular Borel measure µ such
that µ(K) < ∞ for every compact K. We have just shown that every Borel measure on a
locally compact Polish space which is finite on compact sets is a Radon measure.

We will need:

Theorem 28.3 (Tietze’s Extension Theorem). Suppose X is a normal space (this means
that X is Hausdorff and if C1 , C2 are disjoint closed subsets of X then there exist disjoint
open sets O1 , O2 with Ci ⊂ Oi for i = 1, 2). Then if K ⊂ O ⊂ X where K is closed and O
is open then there exists a continuous function f : X → [0, 1] such that f (x) = 1 for x ∈ K
and f (x) = 0 for x ∈/ O.

Exercise 138. Suppose µ, ν are two Radon measures on a locally compact Polish space X
and µ(f ) = ν(f ) for every f ∈ Cc (X). Then µ = ν. Hint: first prove that µ(K) = ν(K) for
compact sets K. Use Tietze’s extension theorem.
Proof. Let K ⊂ X be compact and O1 ⊃ O2 ⊃ · · · precompact open sets with ∩n On = K.
By Tietze’s extension theorem there exist continuous functions fn : X → [0, 1] with fn (x) = 1
for x ∈ K and fn (x) = 0 for x ∈
/ On .

µ(K) = lim µ(fn ) = lim ν(fn ) = ν(K)

n n

by the bounded convergence theorem. Since µ and ν are regular, this implies µ = ν.

28.2 Some examples

Now that we know that Radon measures are determined by their values of compactly sup-
ported continuous functions, we may consider some topological considerations. Let M (X)
denote the space of all Radon measures on X with the following topology. We say that
{µn }n∈N converges to µ∞ if
µn (f ) → µ∞ (f )
for every f ∈ Cc (X).
Exercise 139. Suppose X = Rn . Construct a sequence of probability measures µn that are
absolutely continuous to Lebesgue measure and converge to the Dirac measure δ0 (this is the
measure on Rn defined by δ0 (E) = 1 if 0 ∈ E and δ0 (E) = 0 otherwise.

62
Exercise 140. Construct a Borel probability measure on the middle thirds Cantor set as a
limit of measures on R each of which is absolutely continuous to Lebesgue measure. This
example (and the previous one) show that a limit of absolutely continuous measures can be
singular.
Proof. Let K1 = [0, 1], K2 = [0, 1/3] ∪ [2/3, 1], K3 = [0, 1/9] ∪ [2/9, 1/3] ∪ [2/3, 7/9] ∪ [8/9, 1]
etc be the sets in the constructions of the Cantor set. So if C denotes the Cantor set then
C = ∩n Kn is a decreasing intersection. Observe that m(Kn ) = (2/3)n−1 . Define measures
µn on R by
µn (E) = (3/2)n−1 m(E ∩ Kn ).
Each µn is a probability measure. We claim that the limit µ∞ = limn µn exists and is
supported on C. Observe that if f is a characteristic function of an interval then µn (f )
eventually stabilizes. In particular, limn µn (f ) exists. By linearity, the same holds for simple
functions. Also limn |µn (f )| ≤ kf k∞ since this holds for each n individually. We can ap-
proximate any f ∈ Cc (R) by step functions in the k · k∞ norm. So we obtain that limn µn (f )
exists for every f ∈ Cc (R). This explains that µ∞ = limn µn exists. The fact that it is
supported on C follows because: if O ⊂ C is any nonempty open set then it contains one of
the intervals I in the construction. If I is an interval of Kn then µn (O) ≥ µn (I) = (1/2)n−1 .
Moreover, if k > n then µk (I) = µn (I). So we must have µ∞ (O) ≥ (1/2)n−1 > 0. This shows
that the support of µ∞ contains C. It is obvious that the support is contained in C.
If X is compact then it can be proven that the space of probability measures P (X) ⊂
M (X) is also compact (note: P (X) embeds into [−1, 1]C(X) which is compact).
Exercise 141. Suppose T : X → X is a homeomorphism and µ ∈ P (X). Let T∗ µ ∈ P (X)
be the measure T∗ µ(E) = µ(T −1 E). Show that T∗ is a homeomorphism of P (X). Also show
that if µ ∈ P (X) then any limit point of
n
1X i
T µ
n i=1 ∗

is a fixed point for T . Therefore, there exists a T -invariant measure on X.

The latter exercise is a fundamental result in dynamics...
(Stuff I might add in later: weak* convergence, the Cantor set example, the space of
invariant measures, the existence of invariant measures, horseshoe example?, Gauss’ trans-
formation?, circle rotations)

29 Fourier series
29.1 Definition and convolution
Let T ⊂ C denote the unit circle. We can consider T as a group under multiplication.
Exercise 142. Classify the homomorphisms from T to itself.

63
 We consider T with Lebesgue measure normalized to have mass 1. In other words,
consider the map π : R → T given by π(x) = eix . This maps [0, 2π) onto T and we consider
T with the push-forward measure divided by 2π.
Exercise 143. If h : T → T is a homomorphism and f ∈ L1 (T). Let
Z
fˆ(h) := f (x)h(x−1 ) dx.
T

Show that for any f, g ∈ L1 (T),

∗ g(h) = fˆ(h)ĝ(h).
f[

Proof.
Z Z Z
−1
∗ g(h) =
f[ f[∗ g(x)h(x ) dx = f (xy −1 )g(y)h(x−1 ) dydx
ZT Z T T

= f (xy −1 )g(y)h(x−1 y)h(y −1 ) dydx

ZT ZT
= f (xy −1 )g(y)h(x−1 y)h(y −1 ) dxdy
T T
= fˆ(h)ĝ(h).

To put this another way, we identify T with [0, 2π). Then the exercise states that if
f ∈ L1 (0, 2π), n ∈ Z and Z 2π
ˆ 1
f (n) = f (x)e−inx dx
2π 0
∗ g(n) = fˆ(n)ĝ(n) where the convolution is taken mod 2π.
then f[
Remark 4. We can consider Fourier analysis on an arbitrary locally compact abelian group
G. Let Ĝ = Hom(G, T). This is also an abelian group under pointwise addition. (We will
write the group law of both G and Ĝ additively). It is a theorem that Ĝ is also locally
compact. Because these groups are locally compact, they each admit a Haar measure. Now
if h ∈ Hom(G, T) and f ∈ L1 (G) then we write
Z
ˆ
f (h) = f (x)h(−x) dx.

This defines a homomorphism from the Banach algebra L1 (G) to C in the sense that it is
linear and f[∗ g(h) = fˆ(h)ĝ(h). In other words, we have a natural map form Hom(G, T) to
Hom(L (G), C) given by h 7→ (f 7→ fˆ(h)). It can be shown that this map is bijective.
1

64
29.2 L2
Let en denote the function on T given by en (x) = einx where we have identified T with
[0, 2π). (We could instead write en (z) = z n where we consider T ⊂ C). A trigonometric
polynomial is a finite linear combination of the en ’s. That is to say, a trigonometric
polynomial is any function in the span of the en ’s.
Exercise 144. The functions en (n ∈ Z) are orthonormal in L2 (T).
Note that fˆ(n) = hf, en i.
We’d like to prove that if f ∈ L2 (T) then
X X
f= hf, en ien = fˆ(n)en
n∈Z n∈Z

in the sense that this sum converges in L2 . However, we would need to know that the en ’s
form a basis. This is not obvious.
To put things another way, we would like to show that
N
X
f = lim hf, en ien
N →∞
n=−N

(in L2 ). This implies that the trigonometric polynomials are dense in L2 .

Consider
XN
DN (x) = einx .
n=−N

These are called Dirichlet kernels.

Exercise 145. If f ∈ L2 (T) then
N
X
hf, en ieinx = DN ∗ f (x).
n=−N

Proof. f ∗ en (x) = T f (t)ein(x−t) dt = einx T f (t)e−int dt = hf, en ieinx . So the result follows
R R

by summing n from −N to N .
Unfortunately it’s not true that the DN ’s form an approximation to the identity. So we
let
N
1 X
FN (x) = Dk .
N + 1 k=0
These are called Fejér kernels. (They are Cesaró sums of Dirichlet kernels).
Exercise 146. By the way, why should the Fejer kernels behave better than P the Dirichlet ker-
nels? Well, if {xi } is sequence we could consider the partial sums sn = ni=1 xi . Sometimes
the partial sums do not converge. For example, this occurs if xi = (−1)i . The Cesaro sums

65
of are defined by cn := n1 ni=1 si . It can happen that the Cesaro sums converge even when
P
the partial sums do not. For example, this is true when xi = (−1)n . (Check!). However if
the partial sums converge then the Cesaro sums also converge. And they converge to the
same limit.
Exercise 147. If f ∈ L2 (T) then
N k
1 X X
hf, en ieinx = FN ∗ f (x).
N + 1 k=0 n=−k

We will show that FN does form an approximation to the identity. From this it follows
that FN ∗ f → f as N → ∞ in L2 (T). This FN ∗ f is a trigonometric polynomial (that is, it
is in the span of the en ’s), it follows that the en ’s form an ON basis.
First we study the Dirichlet kernels in a bit more detail. Observe that
2N
−iN x
X 1 − ei(2N +1)x
DN (x) = e einx = e−iN x
n=0
1 − eix
−iN x
e − ei(N +1)x exp(−i(N + 1/2)x) − exp(i(N + 1/2)x)
= ix
=
1−e exp(−ix/2) − exp(ix/2)
−2 sin((N + 1/2)x) sin((N + 1/2)x)
= = .
−2 sin(x/2) sin(x/2)

Exercise 148. Show that DN is an even periodic function that oscillates rapidly when N is
large, DN (0) = 2N + 1 and DN (π) = (−1)N . Also
Z
DN dx = 1.
T

Now
sin((k + 1/2)x) exp(i(k + 1/2)x) − exp(−i(k + 1/2)x)
Dk = = .
sin(x/2) (2i) sin(x/2)
So
(2i) sin(x/2)Dk = exp(i(k + 1/2)x) − exp(−i(k + 1/2)x).
So

(2i sin(x/2))2 Dk (x) = [exp(ix/2) − exp(−ix/2)][exp(i(k + 1/2)x) − exp(−i(k + 1/2)x)]

= exp(i(k + 1)x) − exp(ikx/2) − exp(−ikx/2) + exp(−i(k + 1)x).

66
So
N
2 1 2
X
(2i sin(x/2)) FN (x) = (2i sin(x/2)) Dk (x)
N +1 k=0
N
1 X
= exp(i(k + 1)x) − exp(ikx/2) − exp(−ikx/2) + exp(−i(k + 1)x)
N + 1 k=0
1
= (exp(i(N + 1)x) + exp(−i(N + 1)x) − 2)
N +1
1
= (exp(i(N + 1)x/2) − exp(−i(N + 1)x/2))
N +1
1
= (2i sin((N + 1)x/2))2 .
N +1
Thus
sin((N + 1)x/2)2
FN (x) = .
(N + 1) sin(x/2)2
R
Exercise 149. Show FN (x) ≥ 0, FN (0) = N + 1, T FN (x) dx = 1 and for any 0 < δ < 2π,
Z
FN (x) dx → 0
|x|>δ,x∈T

as N → ∞.
Exercise 150. For any f ∈ Lp (T) (1 ≤ p < ∞), f ∗ FN → f as N → ∞ where convergence
is in Lp -norm. Similarly, if f ∈ C(T) then f ∗ FN → f uniformly.
Exercise 151. Trigonometric polynomials are dense in Lp (T) (1 ≤ p < ∞). Of course, they
are not dense in L∞ (T).
Exercise 152 (Weierstrauss Approximation Theorem). Trigonometric polynomials are dense
in C(T).
Exercise 153 (Weierstrauss Approximation Theorem). Polynomials are dense in C([0, 1]).
Exercise 154 (Weyl’s Equidistribution Theorem). Let α be an irrational number and f ∈
C(T). Then for every x ∈ T,
n Z
1 X
lim f (x + kα) = f (t) dt.
n→∞ n + 1 T
k=0

Moreover the same holds true if f is the characteristic function of an interval.

Exercise 155 (Parseval’s Identity). For any f, g ∈ L2 (T),
Z +∞
X
hf, gi = f ḡ dx = hfˆ, ĝi = fˆ(n)ĝ(n).
T n=−∞

67
Exercise 156. Show that ∞ 1 2 π−x
P
n=1 n2 = π /6. Hint: consider the Rfunction f (x) = 2 on [0, 2π).
2π
1
f (x)2 dx = n∈Z |fˆ(n)|2 .
P
Compute the Fourier coefficients of this function and use 2π 0

Exercise 157. The map f 7→ fˆ gives an isomorphism from L2 (T) to `2 (Z).

Proof. Parseval’s identity shows that f 7→ fˆ is an isometric embedding of L2 (T) into `2 (Z).
But
P how do we know it’s surjective? Let c = (ck ) ∈ `2 (Z). Define fn ∈ L2 (T) by fn =
ˆ ˆ
|k|≤n ck ek . Note fn (k) = ck if |k| ≤ n and fn (k) = 0 otherwise. Note that {fn } is Cauchy
with limit f = k∈Z ck ek . Since f 7→ fˆ is continuous, we have fˆ = limn fˆn . Of course, this
P
could be computed directly now.
Exercise 158 (Fourier Inversion P Theorem). Let f ∈ L1 (T) and assume that the Fourier
series of f converges absolutely: n∈Z |fˆ(n)| < ∞. Then there exists a continuous function
g ∈ C(T) such that f = g a.e. Moreover, the Fourier series converges uniformly to g.
Proof. Let g = n∈Z fˆ(n)en . This converges absolutely, so g ∈ C(T). Of course, ĝ = fˆ, so
P

g = f a.e. (this is because g[− f = 0 so g − f = 0). We have shown that the Cesaro means
of the Fourier series of g converge to g uniformly. This implies the same statement about
the Fourier series.
By contrast, there exist continuous functions f ∈ C(T) such that the Fourier series of f
does not converge to f pointwise on an uncountable set. A deep Theorem of Carleson and
Hunt shows that if f ∈ Lp (T), 1 < p < ∞ then the Fourier series of f converges almost
everywhere. Kolmogorov provided a counterexample with f ∈ L1 (T).

29.3 Derivatives and absolute convergence

Under what conditions on f can we guarantee that its Fourier transform fˆ converges abso-
lutely? We will see that if f is continuously differentiable then this is true. Moreover, we
can express fˆ0 in terms of fˆ. Actually, f need not be continuously differentiable for some
of these results to hold; it need only be absolutely continuous. We will also see that the
regularity of f is reflected in the rate of decay of fˆ.
As usual, we will identify T with the interval [0, 2π). We will say that a function f is
periodic if it is defined on [0, 2π] and f (0) = f (2π). Of course, we could extend f to all of R
by the formula f (x) = f (x + 2πk) for k ∈ Z.
Exercise 159. Suppose f is an absolutely continuous function on [0, 2π] and f (0) = f (2π).
Then
fb0 (n) = fˆ(n)(in).
(In other words, if f ∼ cn einx then f 0 ∼ cn (in)einx ). Hint: integration by parts.
P P

68
Proof. Following the hint,
Z Z 2π
1
f 0 e−inx dx = f 0 (x)e−inx dx
T 2π 0
Z 2π
1 −inx 2π 1
= [f (x)e ]0 − (−in)f (x)e−inx dx
2π 2π 0
= fˆ(n)(in).

Recalling that the indefinite integral of any integrable function is absolutely continuous,
we obtain the first part of:
Rx R 2π
Exercise 160. If F (x) = 0 f (t) dt where f ∈ L2 ([0, 2π)) satisfies 0 f dt = 0 then

fb(n) = F̂ (n)(in).

Moreover, if f ∈ L2 ([0, 2π)) then the Fourier series n∈Z F̂ (n)einx converges absolutely and
P
uniformly to F . In particular, this is true whenever F is continuously differentiable.
P ˆ 2
P fˆ(n)
Proof. We have f = n f (n)en in the sense of L . Also F ∼ n in en . So Cauchy’s
inequality implies
!1/2 !1/2
X fˆ(n) X X
| en | ≤ |fˆ(n)|2 |1/n|2 < ∞.
n
in n n

So the Fourier series of F converges absolutely and uniformly.

Exercise 161 (Riemann-Lebesgue). If f ∈ L1 (T) then fˆ(n) → 0 as n → ±∞. In fact,

|fˆ(n)| ≤ (1/2)kf − τπ/n f k1

where τπ/n f (x) = f (x + π/n). Hint: compute fˆ(n) and use the change of variables x 7→
x + π/n.
Proof.
Z 2π
1
fˆ(n) = f (x)e−inx dx
2π 0
−1 2π
Z
= f (x + π/n)e−inx dx.
2π 0
So
Z 2π
1 1
fˆ(n) = (f (x) − f (x + π/n)) e−inx dx.
2π 0 2

69
Definition 45. f is Hölder of exponent α > 0 if
|f (x + h) − f (x)| ≤ C|h|α
for some constant C > 0 and for all x, h ∈ T.
Exercise 162. If f is Hölder of exponent 0 < α < 1 then fˆ(n) = O(|n|−α ). If f is Lipschitz
then fˆ(n) = o(|n|−1 ). If f ∈ C 2 (T) then fˆ(n) = O(|n|−2 ).
Proof. The first statement follows immediately from the previous exercise. If f is Lipschitz
then it is absolutely continuous. So this statement follows from previous exercises.

29.4 Localization
−
Theorem 29.1. Suppose f ∈ L1 (T), x0 ∈ T, f (x+
0 ) and f (x0 ) both exist and there is a
constant C > 0 such that
|f (y) − f (x+
0 )| ≤ C(y − x0 ), y > x0
|f (x−
0 ) − f (y)| ≤ C(x0 − y), y < x0
for all y in some neighborhood of x0 . (E.g. f could be Lipschitz in a neighborhood of x0 ).
Then
N
X f (x− +
0 ) + f (x0 )
lim fˆ(n) exp(inx0 ) = .
N →∞
n=−N
2
Proof. Recall that
N
X
f ∗ DN (x0 ) = fˆ(n) exp(inx0 ) = DN ∗ f (x0 )
n=−N
sin((N +1/2)x0 )
where DN (x0 ) = sin(x0 /2)
. So
Z 2π Z 2π
1 1 sin((N + 1/2)t)
DN ∗ f (x0 ) = DN (t)f (x0 − t) dt = f (x0 − t) dt.
2π 0 2π 0 sin(t/2)
f (x− +
0 )+f (x0 )
R
Let L = 2
. Since T Dn (t) dt = 1, we have
Z 2π
1 sin((N + 1/2)t)
DN ∗ f (x0 ) − L = [f (x0 − t) − L] dt
2π 0 sin(t/2)
Z π Z 0
1 sin((N + 1/2)t) 1 sin((N + 1/2)t)
= [f (x0 − t) − L] dt + [f (x0 − t) − L] dt
2π 0 sin(t/2) 2π −π sin(t/2)
Z π Z π
1 sin((N + 1/2)t) 1 sin((N + 1/2)t)
= [f (x0 − t) − L] dt + [f (x0 + t) − L] dt
2π 0 sin(t/2) 2π 0 sin(x0 /2)t
Z π
1 sin((N + 1/2)t)
= [f (x0 − t) + f (x0 + t) − 2L] dt
2π 0 sin(t/2)
Z π
1
= φ(t) sin((N + 1/2)t) dt
2π 0

70
where
[f (x0 − t) + f (x0 + t) − 2L] [f (x0 − t) − f (x− +
0 )] + [f (x0 + t) − f (x0 )]
φ(t) = = .
sin(t/2) sin(t/2)
Note φ(t) is integrable for t ∈ [, π] and bounded for t ∈ [0, ] (for some  > 0). Therefore
φ ∈ L1 ([0, π]). So the Riemann-Lebesgue Lemma implies
Z π
1
lim φ(t) sin((N + 1/2)t) dt = 0.
N →∞ 2π 0

This part might not seem so clear because we proved the Riemann-Lebesgue Lemma in a
slightly different form. Let’s go through the steps again (it’s good practice anyway). We
consider φ to be a π-periodic function (by extending its domain of definition). We make the
change of variables s = t + (π/(N + 1/2)). Then
Z π Z π+(π/(N +1/2))
φ(t) sin((N + 1/2)t) dt = − φ(s − (π/(N + 1/2))) sin((N + 1/2)s) ds.
0 (π/(N +1/2))

So
Z π
φ(t) sin((N + 1/2)t) dt
0
"Z #
π Z π+(π/(N +1/2))
1
= φ(t) sin((N + 1/2)t) dt − φ(s − (π/(N + 1/2))) sin((N + 1/2)s) ds .
2 0 (π/(N +1/2))

We bound the above in absolute value by

1 π
Z Z (π/(N +1/2)) Z π
≤ |φ(t)−φ(t−(π/(N +1/2)))| dt+ |φ(t)| dt+ |φ(t)| dt.
2 (π/(N +1/2)) 0 π−(π/(N +1/2))

Because φ ∈ L1 , the later tends to zero as N → ∞.

For example, we let f denote the sign function on [−π, π) which we identify with the
circle. Then Z π
1 1 − (−1)n
fˆ(n) = sign(x)e−ixn dx = .
2π −π inπ
So fˆ(n) = 0 if n is even and fˆ(n) = −2i
nπ
otherwise. Taking advantage of the fact that sin is
odd and cos is even and eix = cos(x) + i sin(x) we have
N N
X
ˆ inx 4 X sin(nx)
f (n)e = .
n=−N
π n=1,odd
n

So if 0 < x < π, we obtain

N
4 X sin(nx)
lim = 1.
N →∞ π n
n=1,odd

71
With x = π/2, we obtain
∞
X (−1)n
= π/4.
n=0
(2n + 1)
Theorem 29.2. If f ∈ L1 (T) and there is an open interval I ⊂ T such that f = 0 in I
then
PN sN (f ) converges uniformly to f on every compact subset of I where sN (f ) = f ∗ DN =
ˆ
n=−N f (n)en .
Proof. The proof of this theorem is similar to the proof of the previous result. To be precise,
we assume x0 ∈ I so that L = 0. Then
Z 2π
1 sin((N + 1/2)t)
DN ∗ f (x0 ) = f (x0 − t) dt
2π 0 sin(t/2)
Z 2π
1
= φx0 (t) sin((N + 1/2)t) dt
2π 0
where
f (x0 − t)
φx0 (t) = .
sin(t/2)
Let  > 0 be small enough so that (− + x0 ,  + x0 ) ∈ I. For |t| <  we have φx0 (t) = 0. So we
can write φx0 (t) = f (x0 − t)χ(t) where χ is a continuous function satisfying χ(t) = sin(t/2)
whenever |t| ≥ . By the change of variables s = t + (π/(N + 1/2)), it follows that
Z 2π
2πDN ∗ f (x0 ) = f (x0 − t)χ(t) sin((N + 1/2)t) dt
0
Z 2π
=− f (x0 − s + 2π/(N + 1/2))χ(s − 2π/(N + 1/2)) sin((N + 1/2)s) ds.
0
So
Z 2π
|4πDN f ∗(x0 )| ≤ |f (x0 −t)χ(t)−f (x0 −t+2π/(N +1/2))χ(t−2π/(N +1/2))|| sin((N +1/2)t)| dt
0
Z 2π
≤ |f (x0 − t)χ(t) − f (x0 − t + 2π/(N + 1/2))χ(t − 2π/(N + 1/2))| dt.
0
Z 2π
≤ |f (x0 − t) − f (x0 − t + 2π/(N + 1/2))||χ(t)| dt
0
Z 2π
+ |f (x0 − t + 2π/(N + 1/2))||χ(t) − χ(t − 2π/(N + 1/2))| dt.
0
The last integral tends to zero as N → ∞ uniformly in x0 because χ is continuous. We let
M = kχk. We bound the first integral by
Z 2π Z 2π
M |f (x0 − t) − f (x0 − t + 2π/(N + 1/2))| dt = M |f (t) − f (t + 2π/(N + 1/2))| dt.
0 0
Since this is independent of x0 and tends to zero as N → ∞ we are done.

72