Math 3306

Measure theory and integration II

Lecture notes of Prof. Hicham Gebran

[email protected]

Lebanese University, Fanar, Spring 2019

www.hichamgebran.com
2

We continue our study of measure theory and integration started in Math 3301. Topics
covered are the Lebesgue measure on IRN , product of measure spaces, the change of variables
theorem, Fubini theorems, Lp spaces, convolutions and Fourier transforms.
All questions, comments, remarks and suggestions are welcome.

References

Paul. R. Halmos, Measure theory (Princeton, Van Nostrand, 1950).

Roger Jean, Mesure et intégration (Presses de l’Université du Québec, 1982).
Walter Rudin, Real and complex analysis (McGraw-Hill, 1977).
Marc Troyanov, Mesure et intégration (Lecture notes, EPFL, 2005).
Thierry Gallay, Théorie de la mesure et de l’intégration (Lecture notes, Université Joseph
Fourier, 2009).
Contents

1 The Lebesgue measure on IRN 5

1.1 The Lebesgue outer measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 The Lebesgue σ−algebra LN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 The Lebesgue measure λN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Change of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Product of measure spaces 9

2.1 Product measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Tonelli and Fubini’s theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Lp spaces 15
3.1 Convex functions and inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4 Convolutions and Fourier transforms 27

4.1 Convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.1 Convolution of nonnegative functions . . . . . . . . . . . . . . . . . . . . . 27
4.1.2 Convolution of integrable functions . . . . . . . . . . . . . . . . . . . . . . 29
4.1.3 Convolution with smooth functions and approximation . . . . . . . . . . . 30
4.2 Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Fourier transforms in L1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2.2 Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.3 Fourier transforms in L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3
4 CONTENTS
Chapter 1

The Lebesgue measure on IRN

Our target in this chapter is to give a mathematical meaning to the intuitive notions of area
and volume. We shall assign a measure (a volume) to many subsets of IRN . This measure is
called the Lebesgue measure and the class of subsets having a measure is called the Lebesgue
σ−algebra.
Here is our plan for this chapter.
1. Construct an outer measure λ∗N on IRN and establish its special properties.
2. Use the first part of Caratheodory’s theorem to construct a σ−algebra LN on IRN and
establish its properties.
3. Use the second part of Caratheodory’s theorem to construct a measure λN on (IRN , LN )
called the Lebesgue measure. Establish some important properties λN .

1.1 The Lebesgue outer measure

A subset P ⊂ IRN is called an N −cell if it is of the form I 1 × · · · × I N where each I k is an
interval of IR (of any type). Thus, a 1−cell is an interval, a 2−cell is a rectangle, a 3−cell is
a parallelepiped. The volume of an N −cell P = I 1 × · · · × I N is `(I 1 ) × `(I 2 ) · · · × `(I N ). We
will denote it by volN (P ). An open N −cell is a product of open intervals. If N = 2, we will
write area(P ) instead of vol2 (P ). Note that according to our convention ∞ × 0 = 0, we have
area(IR × {0}) = 0, vol3 (IR2 × {0}) = 0 and more generally volN (IRN −1 × {0}) = 0. This means
that lines have zero area and planes have zero volume.
The Lebesgue outer measure on IRN is the function λ∗N : P(IRN ) → [0, ∞] defined by
(∞ ∞
)
X [
λ∗N (A) = inf volN (Pn ) Pn is an N -cell and A ⊂ Pn . (1.1)
n=1 n=1

If
P∞ A ⊂ ∪∞ n=1 Pn , then the sequence (Pn ) is called a countable covering of A by N −cells and
vol(P ∗
n=1 n ) is called the total volume of the covering. In words, λN (A) is the the smallest
total volume of countable coverings of A by N -cells. Thus, measuring a subset of IRN involves
approximating this set by N −cells from outside; and this is why we call λ∗N an outer measure.
Some comments are in order. First, λ∗ is well defined. Indeed, let
(∞ ∞
)
X [
XA = volN (Pn )) Pn is an N -cell and A ⊂ Pn .
n=1 n=1
S∞ P∞
Then XA ⊂ [0, ∞]. Since A ⊂ n=1 ] − n, n[N and n=1 (2n)N = +∞, we have +∞ ∈ XA .
Therefore XA is not empty and so it has an infinimum. It is possible that XA = {+∞}. For
example, XIRN = {+∞}.

5
6 CHAPTER 1. THE LEBESGUE MEASURE ON IRN

Second, we can write

(∞ ∞
)
X [
λ∗N (A) = inf volN (Pn ) Pn is an open N -cell and A ⊂ Pn .
i=1 n=1

Proposition 1.1 λ∗N is an outer measure on IRN .

Definition 1.1 Let E ⊂ IRN and a ∈ IRN . We set E + a = {x + a|x ∈ E}. We say that
E + a is a translate of E. If a ∈ IR, we set aE = {ax|x ∈ E}. aE is the image of E under the
homothety x 7→ ax.
Exercise. Prove the following
1. If E is an N −cell then E + a is an N −cell and volN (E + a) = volN (E).
2. If a ∈ IR and P is an N −cell, then volN (aP ) = |a|N volN (P ).

1.2 The Lebesgue σ−algebra LN

Since λ∗N is an outer measure on IRN , it follows from the first part of Caratheodory’s theorem
that the set of λ∗N −measurable sets is a σ−algebra on IRN . We call it the Lebesgue σ−algebra
and we denote it by LN . Therefore E ∈ LN if and only if
λ∗N (A) = λ∗N (A ∩ E) + λ∗N (A\E)
for every subset A ⊂ IRN . An element in LN is called a Lebesgue-measurable set. It turns out
that LN is a big set. It contains the Borel σ−algebra of IRN , but it is much bigger. Recall that
the Borel σ−algebra B(IRN ) is the smallest σ−algebra containing the open subsets of IRN . It
is also the smallest σ−algebra containing the closed subsets of IRN . An element of B(IRN ) is
called Borel-measurable. We also denote the Borel σ− algebra on IRN by BN .

Lemma 1.1 An open subset of IRN is a countable union of open and bounded N −cells.

Corollary 1.1 The Borel σ−algebra of IRN is generated by the open and bounded N −cells.

Lemma 1.2 The Borel σ−algebra on IRN is generated by the family of open half spaces, i,e.,
by the sets of the form

] − ∞, a[×IR × · · · × IR; IR×] − ∞, a[× · · · IR; · · · ; IR × IR × · · · ] − ∞, a[.

Proposition 1.2 BN ⊂ LN , that is, every Borel subset of IRN is Lebesgue-measurable.

Proposition 1.3 LN is invariant under translations and homotheties.

Recall that f : (X, A) → (Y, B) is called (A, B)−measurable if f −1 (B) ⊂ A. Since on IRN ,
there are two main σ−algebras: the Lebesgue σ−algebra LN and the Borel σ−algebra BN , we
distinguish between two types of measurable functions. Recall that a function f : IRN → IR is
called Borel-measurable if it is (BN , B(IR))−measurable, that is, if the inverse image under f
of every Borel set is a Borel set. On the other hand we have the following definition.

Definition 1.2 A function f : IRN → IR is called Lebesgue-measurable (sometimes just mea-

surable) if it is (LN , B(IR))−measurable, that is, if the inverse image under f of every Borel set
is Lebesgue measurable.

Remark 1.1 Every Borel-measurable function is Lebesgue-measurable. The converse is not

true. Why?
1.3. THE LEBESGUE MEASURE λN 7

1.3 The Lebesgue measure λN

The second part of Caratheodory’s theorem ensures that λ∗N restricted to LN is a measure.
We call it the Lebesgue measure on IRN and we denote by λN . Therefore (IRN , LN , λN ) is a
measure space. In addition, to the properties satisfied by any measure, the Lebesgue measure
satisfies some special properties.

Proposition 1.4 The Lebesgue measure λN satisfies the following properties.

1. The measure of a countable set is zero.

2. The measure of an N −cell is equal to its volume.

3. If E ⊂ IRN −1 × {0}, then E ⊂ LN and λN (E) = 0.

4. λN is invariant under translations and symmetries.

5. The measure space (IRN , LN , λN ) is complete.

6. λN is σ−finite.

Here is another important relation between the Lebesgue σ−algebra and the Borel σ−algebra
on IRN .

Proposition 1.5 The Lebesgue σ−algebra is the completion of the Borel σ−algebra. This
means that a subset A ⊂ IRN is Lebesgue measurable if and only if there exists a Borel set B
and a negligible set M , such that A = B ∪ M .

Corollary 1.2 Let f : IRN → IR be Lebesgue measurable. Then there exists a Borel measur-
able function g : IRN → IR that coincides with f almost everywhere.

1.4 Change of variables

Let ϕ : U → V be a differentiable function between two open subsets of IRN . Recall that the
Jacobian J(ϕ) of the transformation ϕ is the determinant of the Fréchet derivative of ϕ:
 ∂ϕ1 ∂ϕ1 
∂x1 ··· ∂xN
J(ϕ) = det  ... .. 

. 
∂ϕN ∂ϕn
∂x1 ··· ∂xN .

Theorem 1.1 (Invariance of the Lebesgue σ−algebra under differentiable homeo-

morphisms) Let U and V be two open sets of IRN and let ϕ : U → V be a differentiable
homeomorphism. Let E ⊂ U be Lebesgue measurable. Then ϕ(E) is also Lebesgue measurable
and Z
λN (ϕ(E)) = |J(ϕ)| dλN .
E

The proof is difficult and can be found in the book of Rudin (Theorem 8.26).
It follows in particular that the Lebesgue measure is invariant under rotations and symmetries
and more generally under any differentiable homeomorphism with Jacobian ±1.
8 CHAPTER 1. THE LEBESGUE MEASURE ON IRN

Theorem 1.2 (Change of variables) Let U and V be two open sets of IRN and let ϕ :
U → V be a differentiable homeomorphism. Let A ⊂ U be a Lebesgue measurable set and
f : ϕ(A) → IR be Lebesgue measurable. If either f has constant sign or is summable, then
Z Z
f (x) dx = f (ϕ(y))|J(ϕ)|dy.
ϕ(A) A

Remark 1.2 If f : A ⊂ V and f : A → IR is Lebesgue measurable which is either of constant

sign or summable, then the change of variables theorem also reads
Z Z
f (x) dx = f (ϕ(y))|J(ϕ)|dy.
A ϕ−1 (A)

∂ old variables
In this form, the Jacobian should be thought of as .
∂ new variables
Chapter 2

Product of measure spaces

2.1 Product measures

Let (X, A) and (Y, B) be two measurable spaces. The product σ−algebra A ⊗ B on X × Y

A ⊗ B = σ ({A × B ⊂ X × Y such that A ∈ A and B ∈ B}) .

It is convenient to denote by A × B the set {A × B ⊂ X × Y such that A ∈ A and B ∈ B},

which is the set of rectangles with measurable sides. Then A ⊗ B = σ(A × B). Observe that
A × B is a π−system on X × Y . Indeed, (A1 × B1 ) ∩ (A2 × B2 ) = (A1 ∩ A2 ) × (B1 ∩ B2 ).

Proposition 2.1 Let p1 : X × Y → X and p2 : X × Y → Y denote the projections onto X

and Y . Then the product σ−algebra is the smallest σ−algebra on X × Y that makes p1 and
p2 measurable.

Proof. a) Recall that p1 : X × Y → X is defined by p1 (x, y) = x and p2 : X × Y → Y is

defined by p2 (x, y) = y. Observe that p−1 −1
1 (A) = A × Y . Therefore if A ∈ A, then p1 (A) ∈
A × B ⊂ A ⊗ B. This means that p1 is measurable when X × Y is equipped with the product
σ−algebra. Similarly, we show that p2 is measurable.
b) Let now D be a sigma algebra on X × Y such that p1 and p2 are measurable. Let A ∈ A and
B ∈ B. Then A×Y = p−1 −1
1 (A) ∈ D and X ×B = p2 (B) ∈ D. Then A×B = (A×Y )∩(X ×B) ∈
D. This means that A × B ⊂ D. It follows that A ⊗ B ⊂ D. 

Proposition 2.2 Let f : (Z, C) → (X × Y, A ⊗ B) where (Z, C) is a measurable space. We

denote by f1 and f2 the components of f . Then f is measurable if and only if f1 and f2 are
measurable.

Proof. Suppose first that f is measurable. Then f1 = p1 ◦ f and f2 = p2 ◦ f . By the previous

proposition, p1 and p2 are measurable. Therefore f1 and f2 are measurable as compositions of
measurable functions. Now suppose conversely, that f1 and f2 are measurable. It is enough to
prove that f −1 (A × B) is measurable and this follows from the fact that

f −1 (A × B) = f1−1 (A) ∩ f2−1 (B).

Indeed x ∈ f −1 (A × B) ⇔ f (x) ∈ A × B ⇔ (f1 (x), f2 (x)) ∈ A × B ⇔ f1 (x) ∈ A and f2 (x) ∈

B ⇔ x ∈ f1−1 (A) and x ∈ f2−1 (B) ⇔ x ∈ f1−1 (A) ∩ f2−1 (A). 

Proposition 2.3 Let (X, A), (Y, B) and (Z, C) be measurable spaces. Let f : X × Y → Z be
measurable. Then the partial maps f (x, ·) : Y → Z and f (·, y) : X → Z are measurable.

9
10 CHAPTER 2. PRODUCT OF MEASURE SPACES

Proof. Let for every x ∈ X, gx : Y → X × Y be defined by gx (y) = (x, y). Then first,
each component of gx is measurable(the first component is constant and the second component
is the identity). By the previous proposition, gx is measurable . Second f (x, ·) = f ◦ gx .
Therefore f (x, ·) is measurable as a composition of measurable functions. The proof that f (·, y)
is measurable is similar. 
Now more generally, let (Xi , Ai ), i = 1, . . . , d, be a sequence of measurable spaces. Then we
denote by A1 ⊗ · · · ⊗ Ad the sigma algebra on X1 × · · · × Xd generated by the sets of the form
A1 × · · · × Ad , where Ai ∈ Ai , that is by the rectangles with measurable sides. This the smallest
sigma algebra on X1 × · · · × Xd that makes the projections pi : (x1 , . . . , xd ) 7→ xi measurable.

Proposition 2.4 [Associativity of ⊗] Let (X1 , A1 ), (X2 , A2 ) and (X2 , A3 ) be measurable spaces.
Then
(A1 ⊗ A2 ) ⊗ A3 = A1 ⊗ (A2 ⊗ A3 ) = A1 ⊗ A2 ⊗ A3 .
Proof. The σ−algebra (A1 ⊗ A2 ) ⊗ A3 is the smallest sigma algebra that makes the maps
((x1 , x2 ), x3 ) 7→ (x1 , x2 ) and ((x1 , x2 ), x3 ) 7→ x3 measurable. These maps are measurable if
and only if the maps (x1 , x2 , x3 ) 7→ x1 , (x1 , x2 , x3 ) 7→ x2 and (x1 , x2 , x3 ) 7→ x3 measurable.
But the smallest sigma algebra that makes these maps measurable is A1 ⊗ A2 ⊗ A3 . Thus,
(A1 ⊗ A2 ) ⊗ A3 = A1 ⊗ A2 ⊗ A3 .
A similar reasoning show that A1 ⊗ (A2 ⊗ A3 ) = A1 ⊗ A2 ⊗ A3 . 
Let (X, A) and (Y, B) be two measurable spaces and let E ∈ A ⊗ B. For every x ∈ X, the
section of E by x is defined by

Ex = {y ∈ Y |(x, y) ∈ E}.

Note that Ex is a measurable set, i.e., Ex ∈ B. Indeed, the characteristic function of Ex satisfies
1Ex (y) = 1E (x, y). The result follows from Proposition 2.3 and the fact that a set is measurable
if and only if its characteristic function is measurable.
Similarly, for y ∈ Y , the section of E by y is defined by

E y = {x ∈ X|(x, y) ∈ E}.

E y is a measurable set, i.e., E y ∈ A.

Lemma 2.1 Let (X, A, µ) and (Y, B, ν) be measure spaces with ν σ−finite. Then for every
E ∈ A ⊗ B the map α : X → [0, ∞] defined by α(x) = ν(Ex ) is A−measurable.

Theorem 2.1 Let (X, A, µ) and (Y, B, ν) be σ−finite measure spaces. Then there exists a
unique measure denoted by µ ⊗ ν on the product space (X × Y, A ⊗ B) satisfying

(µ ⊗ ν)(A × B) = µ(A)ν(B)

for all A ∈ A and B ∈ B. This measure is σ−finite and is called the product measure of µ and
ν. Moreover, for every E ∈ A ⊗ B, we have
Z Z
(µ ⊗ ν)(E) = ν(Ex ) dµ = µ(E y ) dν.
X Y

Consider now the set IRN for N ≥ 2. Let r and s be two positive integers such that
r + s = N . Then IRN can be identified with IRr × IRs (they have the same algebraic and
topological structures). Four natural σ−algebras can therefore be defined on IRN : the Borel
σ−algebra BN , the product σ−algebras Br ⊗ Bs and Lr ⊗ Ls and the Lebesgue σ−algebra
LN . What are then the relation between these σ−algebras? The answer is given in the next
proposition.
2.2. TONELLI AND FUBINI’S THEOREMS 11

Proposition 2.5 Let r and s be two positive integers such that r + s = N . Then the following
hold.

(i) BN = Br ⊗ Bs Lr ⊗ Ls LN .

(ii) The measure space (IRN , LN , λN ) is the completion of the space (IRN , Lr ⊗ Ls , λr ⊗ λs ).

2.2 Tonelli and Fubini’s theorems

Tonelli and Fubini’s theorems are concerned with the interchange of the order of integration,i.e.,
under which conditions do we have
Z Z Z Z
f (x, y) dx dy = f (x, y) dy dx?

Fubini’s theorems give sufficient conditions under which the interchange is possible. These
conditions cannot be easily removed as we shall see in the exercises.

Theorem 2.2 (Tonelli) Let (X, A, µ) and (Y, B, ν) be σ−finite measure R spaces and let f :
X
R × Y → [0, ∞] be A ⊗ Bmeasurable. Then, the functions x →
7 Y f (x, y) dν(y) and y 7→
X f (x, y) dµ(x) are measurable and
Z Z Z  Z Z 
f (x, y)d(µ ⊗ ν) = f (x, y) dν(y) dµ(x) = f (x, y) dµ(x) dν(y). (E)
X×Y X Y Y X

Tonelli theorem states that under certain assumptions, a double integral can be computed as a
nested integral and the order of integration is immaterial. The assumptions are

1. The measures involved are sigma finite and

2. the integrated function is nonnegative.

We shall see in the exercises that the conclusion of the theorem may fail if one of these assump-
tions is not satisfied.
Proof. We prove the theorem in three steps. The first step is not necessary because it is a
particular case of step 2, but we do it because it makes the proof more clear.
Step 1.R Suppose first that
R f is a characteristic
R function. Then f = 1E for some E ∈ A ⊗ B.
Then, Y f (x, y) dν(y) = 1
Y E (x, y) dν = 1
Y Ex (y) dν = ν(Ex ) by definition of the Lebesgue
y
R
integral. Similarly, X f (x, y) dµ(x) = µ(E ). Lemma 2.1 ensures these inner integrals are
measurable. Since they are nonnegative, we can integrate them with respect to the other
measure. Now
Z Z
• f d(µ ⊗ ν) = 1E d(µ ⊗ ν) = (µ ⊗ ν)(E) by definition of the Lebesgue integral,
X×Y X×Y
Z Z  Z
• f (x, y) dν(y) dµ(x) = ν(Ex ) dµ(x);
X Y X
Z Z  Z
• f (x, y) dµ(x) dµ(y) = µ(E y ) dν(y).
Y X Y

By Theorem 2.1, Z Z
(µ ⊗ ν)(E) = ν(Ex ) dµ = µ(E y ) dν.
X Y
Hence equation (E) when f is a characteristic function.
Pm
Step 2. Suppose now that f is a simple function. Then f = i=1 αi 1Ei . Therefore
12 CHAPTER 2. PRODUCT OF MEASURE SPACES

Z Z m
X m
X
• f d(µ ⊗ ν) = αi 1Ei d(µ ⊗ ν) = αi (µ ⊗ ν)(Ei ) by definition of the Lebesgue
X×Y X×Y i=1 i=1
integral;
Z Z  m
X Z
• f (x, y) dν(y) dµ(x) = αi ν((Ei )x ) dµ(x) by linearity of the integral;
X Y i=1 X

Z Z  m Z
µ(Eiy ) dν(y) by linearity of the integral.
X
• f (x, y) dµ(x) dµ(y) = αi
Y X i=1 Y

By Theorem 2.1, the three integrals are equal.

Step 3. Let now f be an arbitrary measurable nonnegative function. Then there exists a
increasing sequence of nonnegative simple functions (fn ) that converges pointwise to f . By step
2

Z Z Z  Z Z 
fn (x, y)d(µ ⊗ ν) = fn (x, y) dν(y) dµ(x) = fn (x, y) dµ(x) dν(y).
X×Y X Y Y X

Letting n → ∞ and using the monotone convergence theorem (twice for the nested integrals) we
get the result (the inner integrals are measurable as pointwise limits of measurable functions) .

Z Z  Z Z
Remark 2.1 Instead of f (x, y) dν(y) dµ(x) we can write dµ(x) f (x, y) dν(y)
Z Z X Y X Y

or dµ f dν.
X Y

Corollary 2.1 Let (X, A, µ) and (Y, B, ν) be σ−finite measure spaces and let f : X × Y → IR
be measurable. Then the following conditions are equivalent

(i) f is µ ⊗ ν summable.
R R

(ii) X Y |f (x, y)| dν dµ < ∞.
R R

(iii) Y X |f (x, y)| dµ dν < ∞.
2
Example. Let f (x, y) = e−x(1+y ) . Then
Z ∞ Z ∞ 
π
f (x, y) dx dy = ,
0 0 2

whereas 2
Z ∞ Z ∞  Z ∞
−t2
f (x, y) dy dx = 2 e dt .
0 0 0
Z ∞ √
2 π
By Tonelli, e−t dt = (write the details).
0 2

Theorem 2.3 (Fubini) Let (X, A, µ) and (Y, B, ν) be σ−finite measure spaces and let f :
X × Y → IR be µ ⊗ ν summable. Then

1. The function y 7→ f (x, y) is ν−summable for µ−almost every x ∈ X and the function
x 7→ f (x, y) is µ−summable for ν−almost every y ∈ Y .
2.2. TONELLI AND FUBINI’S THEOREMS 13

R R
2. The function x 7→ Y f (x, y) dν is µ−summable and the function y 7→ X f (x, y) dµ is
ν−summable.

3. We have Z Z Z  Z Z 
f d(µ ⊗ ν) = f dν dµ = f dµ dν.
X×Y X Y Y X

The conclusion of Fubini’s theorem is the same as that of Tonelli. The assumptions however
are

1. The measures involved are sigma finite and

2. the integrated function is summable.

Proof. We apply Fubini-Tonelli’s theorem to f + and f − which are nonnegative measurable

functions to get
Z Z Z  Z Z 
± ± ±
f d(µ ⊗ ν) = f dν dµ = f dµ dν < ∞.
X×Y X Y Y X

The inner integrals are therefore finite almost everywhere. It follows that
Z Z Z
|f | dν = f + dν + f − dν < ∞
Y Y Y
Z Z Z
for almost every x ∈ X. Similarly, |f | dµ = +
f dµ + f − dµ < ∞ for almost every
X X X
y ∈ Y . This proves 1.
Z Z Z
2. Note that f dν = +
f dν − f − dν. By Tonelli’s theorem, the partial integrals
Z Y Z Y Y
f + dν and f − dν are measurable. Therefore x 7→ Y f dν is measurable on its
R
x 7→
Y Y
domain (that is µ − a.e). Next
Z Z Z Z Z Z
f dν dµ ≤ |f | dν dµ = (f + + f − ) dν dµ
X Y
ZX ZY ZX ZY
= +
f dν dµ + f − dν dµ < ∞.
X Y X Y

For the other partial integral, the proof is similar.

3. Finally, we have
Z Z Z
f d(µ ⊗ ν) = f + d(µ ⊗ ν) − f − d(µ ⊗ ν) by definition of the integral
X×Y X×Y X×Y
Z Z  Z Z 
+ −
= f dν dµ − f dν dµ by Tonelli’s theorem
X Y X Y
Z Z 
+ −
= (f − f ) dν dµ by linearity of the integral
ZX ZY 
= f dν dµ.
X Y

Since X and Y play the same role, we get the conclusion. 

14 CHAPTER 2. PRODUCT OF MEASURE SPACES

Corollary 2.2 Consider a domain ∆ ⊂ IR2 of the form

∆ = {(x, y) ∈ IR2 | a ≤ x ≤ b; ϕ1 (x) ≤ y ≤ ϕ2 (x)}

where ϕ1 , ϕ2 : [a, b] → IR are given Borel measurable functions. Let f : ∆ → IR be Borel

measurable. If either f has constant sign or is summable, then
!
Z Z Z b ϕ2 (x)
f (x, y) dλ2 = f (x, y) dy dx
∆ a ϕ1 (x)

Proof. Observe first that our assumptions imply that ∆ is Borel measurable (write the
details). Now extend f to IR2 by setting
(
f (x, y) if (x, y) ∈ ∆
f¯(x, y) =
0 if (x, y) ∈
/ ∆.

We know from Math 3301 that f¯ is still Borel measurable. Note that
Z Z
¯
f (x, y) dλ2 = f (x, y) dλ2 .
IR2 ∆

By Tonelli’s theorem (if f has constant sign) or Fubini’s theorem (if f is summable), we have
Z Z Z 
f¯(x, y) dλ2 = f¯(x, y) dy dx.
IR2 IR IR

But because f¯ vanishes outside ∆ and coincide with f on ∆, the iterated integral is

 :0
Z Z 
 Z Z  Z Z
f¯(x, y) dy f¯(x, y) dy dx + f¯(x,

dx =   y) dy dx

IR IR [a,b] IR 
IR\[a,b]
  IR

:0
 


Z Z Z
f¯(x, y) dy + ¯ 
= 
 f (x, y) dy
  dx
[a,b] [ϕ1 (x),ϕ2 (x)]

IR\[ϕ
 1 (x),ϕ2 (x)]

Z Z !
= f (x, y) dy dx.
[a,b] [ϕ1 (x),ϕ2 (x)]

Remark 2.2 Indeed, we have a similar result for domains of the form

∆ = {(x, y) ∈ IR2 | a ≤ y ≤ b; ϕ1 (y) ≤ x ≤ ϕ2 (y)}.

More generally, we have the following.

Corollary 2.3 Consider a domain ∆ ⊂ IRN +1 of the form

∆ = {(x1 , . . . , xN , xN +1 ) ∈ IRN +1 | (x1 , . . . , xN ) ∈ D; ϕ1 (x1 , . . . , xN ) ≤ xN +1 ≤ ϕ2 (x1 , . . . , xN )}

where D is a Borel measurable subset of IRN and ϕ1 , ϕ2 : D → IR are given Borel measurable
functions. Let f : ∆ → IR be Borel measurable. If either f has constant sign or is summable,
then
Z Z Z ϕ2 (x1 ,...,xN ) !
f (x1 , . . . , xN , xN +1 ) dλN +1 = f (x1 , . . . , xN , xN +1 ) dxN +1 dλN .
∆ D ϕ1 (x1 ,...,xN )
Chapter 3

Lp spaces

Lp spaces are normed spaces of functions that are very important in Analysis and especially in
the theory of partial differential equations. Before we introduce them, we prove some important
inequalities related to the very important notion of convexity.

3.1 Convex functions and inequalities

Definition 3.1 Let I be an interval of IR. A function ϕ : I → IR is said to be convex if the
following inequality holds.

ϕ((1 − t)x + ty) ≤ (1 − t)ϕ(x) + tϕ(y)

for every t ∈]0, 1[ and every x, y ∈ I. Geometrically, a function is convex if the part of its graph
between any two points is below the straight line segment joining these points.
A function ϕ : I → IR is said to be concave if −ϕ is convex, or equivalently

ϕ((1 − t)x + ty) ≥ (1 − t)ϕ(x) + tϕ(y)

for every t ∈]0, 1[ and every x, y ∈ I. Geometrically, a function is concave if the part of its
graph between any two points is above the straight line segment joining these points.

Remark 3.1 The epigraph of a function ϕ : I → IR is the set of points in the plane that lie
above the graph of ϕ. It is denoted by epi(ϕ). So

epi(ϕ) = {(x, y) ∈ I × IR | y ≥ ϕ(x)}.

Then ϕ is convex if and only if its epigraph is a convex subset of IR2 .

Proposition 3.1 Let I be an interval of IR and let ϕ : I → IR be a function. Then the following
conditions are equivalent

(i) ϕ is convex.

(ii) For every x, y, z ∈ I such that y < x < z, we have

ϕ(x) − ϕ(y) ϕ(z) − ϕ(y)

≤ .
x−y z−y

(iii) For every x, y, z ∈ I such that y < x < z, we have

ϕ(z) − ϕ(y) ϕ(z) − ϕ(x)

≤ .
z−y z−x

15
16 CHAPTER 3. LP SPACES

(iv) For every x, y, z ∈ I such that y < x < z, we have

ϕ(x) − ϕ(y) ϕ(z) − ϕ(x)

≤ .
x−y z−x

Inequality (ii) means that the slope of the line joining (x, ϕ(x)) and (y, ϕ(y)) is smaller than the
slope of the line joining (y, ϕ(y)) to (z, ϕ(z)). Inequality (iii) means that the slope of the line
joining (y, ϕ(y)) and (z, ϕ(z)) is smaller than the slope of the line joining (x, ϕ(x)) to (z, ϕ(z)).
Inequality (iv) means that the slope of the line joining (y, ϕ(y)) and (x, ϕ(x)) is smaller than
the slope of the line joining (x, ϕ(x)) to (z, ϕ(z)).
x−y
Proof. (i)⇒(ii). Observe that x = (1 − t)y + tz where t = z−y . Since 0 < t < 1, convexity of
ϕ implies that ϕ(x) ≤ (1 − t)ϕ(y) + tϕ(z). Therefore
x−y
ϕ(x) − ϕ(y) ≤ t(ϕ(z) − ϕ(y)) = (ϕ(z) − ϕ(y)).
z−y
Dividing by x − y > 0, we get the result. Going back through the steps, we can prove that
(ii)⇒(i).
(i)⇒(iii). As above, we have ϕ(x) ≤ (1 − t)ϕ(y) + tϕ(z). Therefore

ϕ(x) − ϕ(z) ≤ (1 − t) (ϕ(y) − ϕ(z)) .

x−y z−x x−z
But 1 − t = 1 − z−y = z−y = y−z . Therefore dividing by x − z < 0, we get

ϕ(z) − ϕ(x) ϕ(x) − ϕ(z) ϕ(y) − ϕ(z) ϕ(z) − ϕ(y)

= ≥ = .
z−x x−z y−z z−y

Going back through the steps, you can prove that (iii)⇒(i).
(i)⇒(iv). (iv) follows from (ii) and (iii).
(iv)⇒(i). Let y, z ∈ I such that y < z and let t ∈]0, 1[. Let x = (1 − t)y + tz. Then y < x < z
and so
ϕ(x) − ϕ(y) ϕ(z) − ϕ(x)
≤ .
x−y z−x
It follows that
x−y t
ϕ(x) − ϕ(y) ≤ (ϕ(z) − ϕ(x)) = (ϕ(z) − ϕ(x)) .
z−x 1−t
Therefore
ϕ(x) ≤ (1 − t)ϕ(y) + tϕ(z)
that is
ϕ((1 − t)y + tz) ≤ (1 − t)ϕ(y) + tϕ(z).

Observe now the following. Inequality (ii) implies that for each y ∈ I the function s 7→
ϕ(s)−ϕ(y)
s−y is increasing for s > y andinequality (iii) implies that for each z ∈ I, the function
ϕ(s)−ϕ(z) ϕ(s)−ϕ(a)
s 7→ s−z It follows that for any a ∈ I the function s 7→
is increasing for s < z. s−a is
increasing on I\{a}. Actually this condition is equivalent to the convexity of ϕ.

A convex function need not be differentiable. For example x 7→ |x| is not differentiable at 0.
However, it has a left and right derivative at 0. This is not a coincidence. A convex function
has a left and right derivative at every point inside its domain.
3.1. CONVEX FUNCTIONS AND INEQUALITIES 17

Proposition 3.2 Let I be an open interval of IR and let ϕ : I → IR be convex. Then the
following hold.

(i) At each point x ∈ I, ϕ has a left derivative ϕ0` (x) and right derivative ϕ0r (x) with ϕ0` (x) ≤
ϕ0r (x).

(ii) ϕ is continuous.

(iii) The left and right derivatives are increasing functions.

(iv) For every x ∈ I, for every β ∈ [ϕ0` (x), ϕ0r (x)], and for every y ∈ I, we have

ϕ(y) ≥ ϕ(x) + β(y − x).

Proof. (i). Let x, y, z ∈ I satisfy y < x < z. By the previous proposition

ϕ(y) − ϕ(x) ϕ(z) − ϕ(x)

≤ . (3.1)
y−x z−x

Now fix x and z and let y vary. Then the function y 7→ ϕ(y)−ϕ(x)
y−x is increasing and bounded
from above. Therefore it has a limit as y % x. This means that ϕ has a left derivative at x.
If now we fix x and y and let z vary, we see that g(z) := ϕ(z)−ϕ(x)
z−x has a limit as z & x.1
This means that ϕ has a right derivative at x.
Now, if we fix x and let y % x and z & x in (3.1), we get ϕ0` (x) ≤ ϕ0r (x).
(ii). Let x ∈ I. Tthe existence of a left derivative at x implies the ϕ is continuous from the
left at x, that is, lim ϕ(y) = ϕ(x). Also the existence of a right derivative at x implies ϕ is
y%x
continuous from the right at x, that is, lim ϕ(y) = ϕ(x). It follows that ϕ is continuous at x.
y&x

(iii). Let x, z ∈ I satisfy x < z. We have to show that ϕ0` (x) ≤ ϕ0` (z). So let y and w satisfy
y < x < w < z. In this reasining, x and z are fixed and y and w vary. By the previous
proposition
ϕ(y) − ϕ(x) ϕ(x) − ϕ(w) ϕ(w) − ϕ(z)
≤ ≤ .
y−x x−w w−z
Letting y % x and w % z, we get ϕ0` (x) ≤ ϕ0` (z). Acompletely similar reasoning shows that
ϕ0r (x) ≤ ϕ0r (z).
(iv). Let x, y, z ∈ I satisfy y < x < z. The proof of (i) shows that

ϕ(y) − ϕ(x) ϕ(z) − ϕ(x)

≤ ϕ0` (x) ≤ ϕ0r (x) ≤ .
y−x z−x
Therefore if ϕ0` (x) ≤ β ≤ ϕ0r (x), we have ϕ(y) ≥ ϕ(x) + β(y − x) and ϕ(z) ≥ ϕ(x) + β(z − x).
Since y and z were arbitrary, this proves the result. 

Remark 3.2 A convex function ϕ : I → IR need not be continuous at the end point of its
domain. For example the function ϕ : [0, 1] → IR defined by ϕ(0) = 1 and ϕ(x) = 0 for
0 < x ≤ 1 is convex but discontinuous at x = 0. This is why we assumed that I is open in the
previous proposition.

Proposition 3.3 Let I be an open interval of IR and let ϕ : I → IR be differentiable. Then

the following conditions are equivalent.

(i) ϕ is convex.
1
As z decreases to x, g(z) decreases and is bounded from below (by g(y)). Therefore it has a limit.
18 CHAPTER 3. LP SPACES

(ii) ϕ(y) ≥ ϕ(x) + ϕ0 (x)(y − x) for every y ∈ I (this means that the graph of ϕ is above any
of its tangent lines).

(iii) ϕ0 is increasing.

If moreover ϕ is twice differentiable, then (i), (ii) and (iii) are equivalent to the condition
ϕ00 ≥ 0.

Proof. (i)⇒(ii) follows from point (iv) of the the previous proposition.
ϕ(y)−ϕ(x)
(ii)⇒(iii). Let x, y ∈ I satisfy y < x. Then ϕ(y) ≥ ϕ(x) + ϕ0 (x)(y − x) and so y−x ≤ ϕ0 (x).
ϕ(x)−ϕ(y)
Similarly we have ϕ(x) ≥ ϕ(y) + ϕ0 (y)(x − y) and so x−y ≥ ϕ0 (y). Thus,

ϕ(x) − ϕ(y) ϕ(y) − ϕ(x)

ϕ0 (y) ≤ = ≤ ϕ0 (x).
x−y y−x

(iii)⇒(i). Suppose that ϕ is not convex. By Proposition 3.1 (iv), there exists x, y, z ∈ I such
that y < x < z and
ϕ(x) − ϕ(y) ϕ(z) − ϕ(x)
> .
x−y z−x
ϕ(x)−ϕ(y)
Now by the value theorem, there exists ξ1 between y and x such that x−y = ϕ0 (ξ1 ) and
there exists ξ2 between x and z such that ϕ(z)−ϕ(x)
z−x = ϕ0 (ξ2 ). But then we have ϕ0 (ξ1 ) > ϕ0 (ξ2 )
although ξ1 < ξ2 and this contradicts the monotonicity of ϕ0 . 

Remark 3.3 If I is not necessarily open but ϕ is continuously differentiable, then the previous
corollary still hold.

Examples. (1) The function x 7→ ex is convex on IR.

(2) Let p ≥ 1. The function x 7→ xp is convex on [0, ∞[. It follows for example that if a, b ≥ 0
then
a+b p 1 p
 
≤ (a + bp )
2 2
and so (a + b)p ≤ 2p−1 (ap + bp ).
(3) The function x 7→ ln x is concave on ]0, ∞[.
(4) The function x 7→ sin x is concave on [0, π] and convex on [π, 2π]. It follows that sin x ≥ π2 x
for every x ∈ [0, π2 ] (draw a figure).
(5) The function x 7→ arctan x is convex on ] − ∞, 0] and concave on [0, ∞[. It follows that
arctan x ≥ π4 x for every x ∈ [0, 1].

Definition 3.2 We say that two numbers p, q > 1 are conjugate if

1 1
+ = 1,
p q
q
or equivalently if pq = p + q, or p = q−1 , or pq = q − 1. We extend this definition to include the
case p = 1 and q = ∞ so that 1 and ∞ are called conjugate.

Proposition 3.4 (Young’s inequality) Let p and q be two conjugate numbers > 1 and let
a, b be nonnegative. Then
ap bq
ab ≤ + .
p q
3.1. CONVEX FUNCTIONS AND INEQUALITIES 19

Proof. The inequality is trivial if a = 0 or b = 0. So we assume that ab 6= 0. By concavity

of the logarithm we have  
1 p 1 q 1 p 1 q
ln a + ln b ≤ ln a + b .
p q p q
1
But p ln ap + 1q ln bq = ln a + ln b = ln(ab). Since ln is increasing, we get the result. 

Proposition 3.5 (Hölder’s inequality) Let (X, A, µ) be a measure space and let f, g : X →
IR be two measurable functions. Let p and q be two conjugate numbers > 1. Then
Z Z 1/p Z 1/q
p q
|f g| dµ ≤ |f | dµ |g| dµ .
X X X

1/p
Proof. Let ||u||p = X |u|p dµ
R
. We shall see shortly that this is a semi-norm. Then the
inequality can be written
||f g||1 ≤ ||f ||p ||g||q .

Note that ||αu|| = |α||u||p for any constant α.

R
Step 1. If ||f ||p = 0, then f = 0 a.e and so f g = 0. So the inequality is satisfied. Similarly
if ||g||q = 0.
Step 2. Assume that ||f ||p = ||g||q = 1. Now by Young’s inequality

1 1
|f (x)g(x)| ≤ |f (x)|p + |g(x)|q .
p q

Integrating over X, we get Z Z Z

1 1
p
|f g| ≤ |f | + |g|q ,
X p X q X

that is
1 1 1 1
||f g||1 ≤ ||f ||pp + ||g||qq = + = 1,
p q p q
since ||f ||p = ||g||q = 1.
Step 3. Assume that ||f ||p 6= 0 and ||g||q 6= 0. Set

f g
f1 = and g1 =
||f ||p ||g||q

so that ||f1 ||p = ||g1 ||q = 1. It follows from Step 2 that ||f1 g1 ||1 ≤ 1, that is

||f g||1
≤ 1.
||f ||p ||g||q

Proposition 3.6 (Minkowski’s inequality) Let (X, A, µ) be a measure space and let f, g :
X → IR be measurable functions. Let p ∈ [1, ∞[. Then

||f + g||p ≤ ||f ||p + ||g||p ,

1/p
|f |p dµ
R
where as before ||f ||p = X
20 CHAPTER 3. LP SPACES

Proof. Note that |f + g|p = |f + g|p−1 |f + g| ≤ |f + g|p−1 |f | + |f + g|p−1 |g|. Integrating we

get Z Z
p p−1
||f + g||p ≤ |f + g| |f | + |f + g|p−1 |g|.
X X
Since (p − 1)q = p, it follows that
Z 1/q Z (p−1)/p
(p−1)q p
|f + g| = |f + g| = ||f + g||pp−1 .
X X

Therefore by Hölder’s inequality,

||f + g||pp ≤ ||f + g||p−1 p−1 p−1

p ||f ||p + ||f + g||p ||g||p = ||f + g||p (||f ||p + ||g||p ) .

Dividing by ||f + g||p−1 , we get the result.

3.2 Definitions
Definition 3.3 Let 1 ≤ p < ∞ and let (X, A, µ) be a measure space. We denote by Lp (X) the
set of measurable functions f : X → IR, such that |f |p is summable on X:
Z
p
L (X) = {f : X → IR | f is measurable and |f |p dµ < ∞}.
X

Other notations: Lp (X, A, µ), Lp (µ), Lp .

Examples. 1) Equipp ]0, 1[ with the Lebesgue measure. Let f (x) = √1 . Then f ∈ L1 (]0, 1[)
x
/ L2 (]0, 1[).
but f ∈
2) Equipp ]1, ∞[ with the Lebesgue measure. and let f (x) = x1 . Then f ∈ Lp (]1, ∞[) for every
/ L1 (]1, ∞[).
p ∈]1, ∞[ but f ∈
3) Equipp IN∗ with the counting measure let f (n) = 1
n. Then f ∈ Lp (IN∗ ) for every p ∈]1, ∞[
/ L1 (IN∗ ).
but f ∈

Proposition 3.7 Let 1 ≤ p < ∞ and let (X, A, µ) be a measure space. Then

(i) Lp (X) is a vector space.

(ii) If we set for f ∈ Lp (X)

Z 1/p
p
||f ||p = |f | dµ ,
X

then || · ||p is a semi-norm on Lp (X)

Proof. (i) Lp is a subspace of the space of all functions from X → IR equipped with the usual
addition and multiplication by a scalar. Indeed, let f, g ∈ Lp (X) and let α be a constant. Then
first, f + g and αf are measurable. Next, |f + g|p ≤ (|f | + |g|)p ≤ 2p−1 (|f |p + |g|p ). Integrating
over X, we get Z Z Z 
|f + g|p ≤ 2p−1 |f |p + |g|p < ∞.
X X X
Z Z
Thus f + g ∈ Lp (X). Next |αf |p = |α|p |f |p < ∞ and so αf ∈ Lp (X).
X X
(ii) We clearly have

1. ||f ||p ≥ 0.
3.2. DEFINITIONS 21

2. ||0||p = 0.
3. ||αf ||p = |α|||f ||p (Homogeneity)
4. ||f + g||p ≤ ||f ||p + ||g||p (Triangle inequality)
The triangle inequality is just Minkowski’s inequality. However, ||f ||p = 0 ⇒ f = 0 a.e. So
|| · ||p is not strictly speaking a norm.

Definition 3.4 Let (X, A, µ) be a measure space and let f : X → IR be measurable. We say
that f is essentially bounded if there exists a constant C ≥ 0 such that |f (x)| ≤ C a.e. We
say that C is an essential bound of |f |. The set of measurable essentially bounded functions is
denoted by L∞ (X), that is,

L∞ (X) = {f : X → IR | f is measurable and ∃ C such that|f (x)| ≤ C a.e. on X}.

Other notations: L∞ (X, A, µ), L∞ (µ), L∞ .

Examples. 1) A bounded function f is of course essentially bounded. The converse is however
( For example, equip IR with the Lebesgue measure and let f : IR → IR be defined by
not true.
x if x ∈ IN
f (x) = . Then f is essentially bounded but not bounded. So f ∈ L∞ (IR).
1 otherwise
/ L∞ (]0, 1[)
2) Equip ]0, 1[ with the Lebesgue measure and Let f (x) = x1 for x ∈]0, 1[. Then f ∈
(reason by contradiction). More generally, a function f :]a, b[→ IR such that limx→a |f (x)| =
+∞ is not essentially bounded (for the Lebesgue measure).

Proposition 3.8 Let (X, A, µ) be a measure space. Then

(i) L∞ (X) is a vector space.
(ii) If we set for f ∈ L∞ (X)

||f ||∞ = inf{C | |f (x)| ≤ C a.e.},

then || · ||∞ is a semi-norm on L∞ (X)

Proof. (i) Let f, g ∈ L∞ (X) and let α be a constant. Then first, f +g and αf are measurable.
Next, |f (x)| ≤ C1 and |g(x)| ≤ C2 a.e and so

|f (x) + g(x)| ≤ |f (x)| + |g(x)| ≤ C1 + C2 a.e.

Thus f + g ∈ L∞ (X). Also |αf (x)| ≤ |α|C1 a.e. and so αf ∈ L∞ (X).

(ii) We claim first that ||f ||∞ is an essential bound of |f | when f ∈ L∞ (the inf is a min).
Indeed, by a property of the inf, for all n ∈ IN∗ , there exists an essential bound C such that
C < ||f ||∞ + n1 . Therefore |f (x)| < ||f ||∞ + n1 a.e. Then, for each n, there exists a set An
on which the last inequality holds and µ(Acn ) = 0. Let A = ∩An . Then for x ∈ A, we have
|f (x)| ≤ ||f ||∞ and µ((∩An )c ) = µ(∪Acn ) = 0. This proves the claim.
Now, the first two properties of a semi norm are indeed satisfied. So we prove homogenity.
Let f ∈ L∞ . From the inequality |f (x)| ≤ ||f ||∞ a.e, it follows that |αf (x)| ≤ |α|||f ||∞ . This
means that |α|||f ||∞ is an essential bound for |αf | and so ||αf ||∞ ≤ |α|||f ||∞ . To prove the
reverse inequality we assume that α 6= 0 because otherwise the inequality is satisfied. From
1
|αf (x)| ≤ ||αf ||∞ a.e. it follows after dividing by 1/|α| that ||f ||∞ ≤ |α| ||αf ||∞ . Thus,
|α|||f ||∞ ≤ ||αf ||∞ . Hence equality.
We prove now the triangle inequality. For two elements in f, g ∈ L∞ (X), we have |f (x)| ≤
||f ||∞ and |g(x)| ≤ ||g||∞ a.e. Therefore |f (x)+g(x)| ≤ |f (x)|+|g(x)| ≤ ||f ||∞ +||g||∞ a.e. This
implies that ||f ||∞ + ||g||∞ is an essential bound for |f + g| and so ||f + g||∞ ≤ ||f ||∞ + ||g||∞ .
22 CHAPTER 3. LP SPACES

Remark 3.4 Let (X, A, µ) = (IRN , LN , λN ) and let f : IRN → IR be continuous and bounded
then sup |f (x)| = ||f ||∞ . In fact we have |f (x)| ≤ ||f ||∞ everywhere because the set where
f (x) > ||f ||∞ is open (by continuity of f ) and negligible and so it is empty. Therefore sup |f | ≤
||f ||∞ . Conversely, sup |f | is essential bound for |f | so ||f ||∞ ≤ sup |f |.

On Lp (X), we define the relation f ∼ g if f = g a.e. It should be clear that this is an equivalence
relation which is moreover compatible with the addition an multiplication by scalars. The
quotient set of this relation is denoted by Lp (X). So an element of Lp (X) is a class of functions
that coincide a.e. between each other. But there is a natural structure of vector space on
Lp (X). Indeed, let us denote by [f ] the equivalence class containing the function f . Define and
addition and multiplication by a scalar by

[f ] + [g] = [f + g] and α[f ] = [αf ]

This makes sense because if f1 ∈ [f ] and g1 ∈ [g], then f1 + g1 = f + g a.e. so [f1 + g1 ] = [f + g].
Also αf1 = αf a.e and so [αf1 ] = [αf ]. Now one can easily check that Lp is a vector space on
IR (the neutral element of the addition being [0], the class of all functions that are equal a.e.
to zero). Finally we set ||[f ]||p = ||f ||p . This also makes sense since if f ∼ g then ||f ||p = ||g||p .
And now ||f ||p is a norm on Lp since ||[f ]|| = 0 ⇒ f = 0 a.e., and so [f ] = [0].

Remark 3.5 In practise we write ||f ||p or ||f ||Lp instead of ||[f ]||p and of we think of elements
of Lp as functions that are defined a.e. and (not as classes). With this interpretation it makes
no sense to speak about the value of a function on a set of measure zero.

3.3 Properties
Theorem 3.1 (Riesz-Fisher) (Lp (X), || · ||p ) is a Banach space for every 1 ≤ p ≤ +∞.
Proof.
1) Suppose first that p = +∞. Let (fn ) be a Cauchy sequence in L∞ . For any k ∈ IN, there
exists Nk such that
1
||fm − fn ||∞ ≤ for all m, n ≥ Nk .
k
So there exists Ak of measure zero such that
1
|fm (x) − fn (x)| ≤ for all x ∈ X\Ak , m, n ≥ Nk (3.2)
k
S
Let A = k Ak . Then A is of measure zero and for all x ∈ X\A, the sequence (fn (x)) is a
Cauchy sequence in IR. It converges therefore to a function f (x). Letting m → ∞ in (3.2), we
get
1
|f (x) − fn (x)| ≤ for all x ∈ X\A, n ≥ Nk .
k
Setting f (x) = 0 for x ∈ A, we see that f ∈ L∞ (because |f (x)| ≤ |fN1 (x)| + 1 for all x ∈ X)
and ||f − fn ||∞ ≤ k1 for all n ≥ Nk . Consequently, ||f − fn ||∞ → 0.
2) Suppose now that 1 ≤ p < ∞. Let (fn ) be a Cauchy sequence in Lp . It is enough to show
that (fn ) contains a convergent subsequence. Since (fn ) is a Cauchy sequence, there exists n1
such that ||fm − fn ||p ≤ 21 for m, n ≥ n1 . Next, there exists n2 > n1 such that ||fm − fn ||p ≤ 212
for m, n ≥ n2 . Next, there exists n3 > n2 such that ||fm − fn ||p ≤ 213 for m, n ≥ n3 and so on.
This defines a subsequence (fnk ) such that

1
||fnk+1 − fnk ||p ≤ ∀k ≥ 1. (3.3)
2k
3.3. PROPERTIES 23

We show that (fnk ) converges in Lp . To simplify the notation we write fk instead of fnk , so
that
1
||fk+1 − fk ||p ≤ ∀k ≥ 1. (3.4)
2k
Let
n−1
X
gn (x) = |fk+1 (x) − fk (x)|.
k=1
Then by Minkowski’s inequality and (3.4)
n−1 n−1
X X 1
||gn ||p ≤ ||fk+1 − fk ||p ≤ < 1. (3.5)
2k
k=1 k=1

It is clear that ((gn )p ) is an increasing sequence of nonegative functions. Therefore it converges

p
pointwise Rto some nonegative function h. R By the monotone convergence theorem ||gn ||p con-
verges to X h dµ. By inequality (3.5), X h dµ ≤ 1. Therefore h is summable. Note that h
could take the value +∞. However, summability of h dictates that h(x) < ∞ almost every-
where. Modifying h on a set of measure zero, we may assume that h is finite everywhere and
so g := h1/p ∈ Lp . But now we have that ((gn )p ) converges almost everywhere to h. Continuity
of t 7→ t1/p implies that (gn ) converges to g almost everywhere.
On the other hand, we have for m > n and for every x ∈ X,

|fm (x) − fn (x)| ≤ |fm (x) − fm−1 (x)| + · · · + |fn+1 (x) − fn (x)| ≤ gm (x) − gn−1 (x) (3.6)
≤ gm (x). (3.7)

It follows that for almost every x ∈ X, (fn (x)) is a Cauchy sequence in IR. Since IR is complete
(fn (x)) converges almost everywhere to a limit f (x). It follows by letting m → ∞ in inequality
(3.7), that
|f (x) − fn (x)| ≤ g(x) a.e. (3.8)
Then first, f ∈ Lp since |f (x)| ≤ |fn (x)| + g(x). Second, |fn (x) − f (x)|p → 0 a.e. with
|fn (x) − f (x)|p ≤ g(x)p . Thus, by the dominated convergence theorem, ||fn − f ||p → 0. 

Corollary 3.1 L2 (X) is a Hilbert space for the scalar product

Z
hf, gi = f g dµ.
X

Proposition 3.9 Let (fn ) be a sequence of Lp that converges in Lp to some function f . Then
there exists a subsequence (fnk ) such that
(a) fnk → f a.e.

(b) |fnk | ≤ h a.e for all k with h ∈ Lp .

Proof. If p = ∞, the result follows from part 1) of the proof of the previous theorem (in
fact, we can take fnk = fn ). Suppose now that 1 ≤ p < ∞. From the proof of part 2), there is
a subsequence fnk satisfying (3.3). Continuing as in the proof, we see that fnk converges a.e to
a limit f ∗ (we need to distinguish first between f and f ∗ ). Moreover, thanks to (3.7),

|f ∗ (x) − fnk (x)| ≤ g(x) a.e ∀k, with g ∈ Lp . (3.9)

It follows first that f ∗ ∈ Lp . Second, by the dominated convergence theorem fnk → f ∗ in Lp .

Thus, f = f ∗ a.e. Inequality (3.9) yields |fnk (x)| ≤ |f ∗ (x)| + g(x) a.e. with |f ∗ | + g ∈ Lp . This
proves (b). 
24 CHAPTER 3. LP SPACES

Remark 3.6 We can formulate the following Lp version of the dominated convergence theorem.
Let p ∈ [1, ∞[ and let fn be a sequence of measurable functions such that 1. fn converges to
f a.e. and 2. |fn | ≤ g a.e. where g ∈ Lp . Then fn converges to f in Lp . Indeed, |fn − f |p
converges to zero a.e. and |fn − f |p ≤ 2p g p (because |fn − f | ≤ |fn | + |f | ≤ 2g). The classical
dominated convergence theorem yields the result. However the result is not true in L∞ (consider
for instance the sequence 1[n,∞[ ).

Proposition 3.10 Let (X, A, µ) be a measure space with µ(X) < ∞. If 1 ≤ p ≤ q ≤ ∞ then
Lq (X) ⊂ Lp (X) and the natural injection is continuous.
Suppose first that q = ∞ and p < ∞. Then X |f |p dµ ≤ ||f ||p∞ µ(X) and so
R
Proof.
||f ||p ≤ ||f ||∞ µ(X)1/p . This estimate proves the result. Next we assume that p < q < ∞. Let
q
f ∈ Lq . Applying Hölder’s inequality to |f |p and 1 with exponents α = q/p and β = q−p , we
obtain q−p
||f ||p ≤ ||f ||q µ(X) pq .


Remark 3.7 If µ(X) = ∞ then the conclusion need not be true. For instance the function f
defined by f (x) = √x12 +1 belongs to L2 (IR) but not to L1 so L2 " L1 .

Let Ω ⊂ IRN be Lebesgue measurable. We consider Ω as a measure subspace of (IRN , LN , λN ).

Let p ∈ [1, +∞[. Then LpR(Ω) denotes the space of functions f : Ω → IR which are Lebesgue
measurable and such that Ω |f |p dλN < +∞. L∞ (Ω) denotes the space of Lebesgue measurable
functions f : Ω → IR which are bounded almost everywhere.
Functions of Lp (Ω) may be discontinuous, but can we approximate them with continuous
functions? The answer is yes. Actually we can approximate functions of Lp by continuous
functions that vanish outside a compact set. These functions are said to have compact support.
Let f : Ω → IR. The support of f is the smallest closed set on which f does not vanish. It is
denoted by supp f . Thus, supp f = {x ∈ Ω|f (x) 6= 0}. The space of continuous functions with
compact support is denoted by Cc (Ω).
Remarks. 1) Cc (Ω) is a vector space.
2) A function in Cc (Ω) is uniformly continuous (prove this).

Theorem 3.2 (Density theorem) Let Ω ⊂ IRN be an open set and let 1 ≤ p < ∞. Then
Cc (Ω) is dense in Lp (Ω).
To prove this theorem we need a lemma which hold in a more general context.

Lemma 3.1 Let (X, A, µ) be a measurable space. Then any function in Lp (X) is the limit in
Lp of a sequence of simple functions in Lp .
Proof. Let (gn ) be an increasing sequence of simple nonnegative functions that converges
(pointwise) to f + and let (hn ) be an increasing sequence of of simple nonnegative functions
that converges to f − . Let ϕn = gn − hn . Then ϕn → f pointwise and |f − ϕn | ≤ |f | + |ϕn | ≤
|f | + hn + gn ≤ 2|f |. Then |f − ϕn |p ≤ 2p |f |p . It follows first that ϕn ∈ Lp and second that
||f − ϕn ||p → 0 by the Lebesgue dominated convergence theorem. 

Proof of Theorem 3.2. Thanks to Lemma 3.1, it is enough to prove that any simple
function in Lp (Ω) can be approximated in the Lp norm by a function in Cc (Ω). We prove this
in two steps.
Step 1. For simplicity, we will assume that Ω = IRN (we will prove a more general result in
the next chapter). Let f = 1A ∈ Lp where A is a Lebesgue measurable subset of IRN . Then
3.3. PROPERTIES 25

necessarily λN (A) = ||1A ||pp < +∞. Let  > 0 be given. We know from an exercise of chapter 1
that
λN (A) = inf{λN (O)|O open and A ⊂ O}.
From a property of the inf, there exists an open set O such that A ⊂ O and λN (O) < λN (A)+p
Then
||1O − 1A ||p = ||1O\A ||p = λ(O\A)1/p < .
Set now, for each n ∈ IN∗ , On = O ∩ B(0, n). Then
1. On is open.
2. On is compact (closed and bounded).
S∞
3. The sequence (On ) is increasing and O = n=1 On .

This implies that 1O = lim 1On . Using the Lebesgue dominated convergence theorem we
see that ||1O − 1On ||p → 0. Therefore we can choose some n such that ||1O − 1On ||p < . Let
V = On . Then V is an open set whose closure is compact and

||1O − 1V ||p < .

Now for each k ∈ IN∗ , set ϕk (x) = min(1, kd(x, V c )). Then first, ϕk is Lipshitz continuous with
constant k (check this). Second, supp ϕk ⊂ V . Therefore ϕk has a compact support; and so
ϕk ∈ Cc (IRN ). Third, ϕk → 1V pointwise. Again using the Lebesgue dominated convergence
theorem we see that ||ϕk − 1V ||p → 0. Thus we can choose k such that

||ϕk − 1V ||p < .

By the triangle inequality,

||1A − ϕk ||p < 3.
Since  was arbitrary, this means that f = 1A can be approximated in the Lp norm by a function
in Cc (IRN ).
Step 2. Let f be a simple function in Lp (Ω). Then we can write f = m
P
j=1 αj 1ARj where the

R(Aj ) form a partition of Ω. Moreover, for each j = 1, . . . , m, λN (Aj ) < ∞ because Ω |f |p dλ ≥

p p
Aj |f | dλN = |aj | λN (Aj ).
According to step 1, for each j =P 1, . . . , m there exists a function ϕj ∈ Cc (Ω) such that
k1Aj − ϕj ||p < m(|αj |+1) . Letting ϕ = mj=1 αj ϕj , we get ϕ ∈ Cc (Ω) and

||f − ϕ||p < .

Remark 3.8 The above theorem does not hold for p = +∞. We can prove for example that
the closure of Cc (IR) in L∞ (IR) is the space of continuous functions that vanish at infinity,
which is a proper subspace of L∞ (IR). Prove this if you want.
We end this chapter by a useful result that we shall use later.

Proposition 3.11 (Continuity in the mean) Let 1 ≤ p < +∞ and let f ∈ Lp (IRN ). Then
Z
lim |f (x + h) − f (x)|p dx = 0.
h→0 IRN

Otherwise sated, if τh f is the function defined by τh f (x) = f (x + h), then ||τh f − f ||p → 0 as
h → 0.
Proof. See the exercises. 
26 CHAPTER 3. LP SPACES
Chapter 4

Convolutions and Fourier transforms

4.1 Convolutions
The convolution of two functions f, g : IRN → IR is formally defined by
Z
(f ∗ g)(x) = f (x − y)g(y) dy
IRN

There are conditions under which this integral is well defined. This integral exists if (1) f and g
are both Lebesgue measurable and nonnegative or (2) f ∈ L1 (IRN ) and g ∈ Lp (IRN ). We shall
deal with these cases, but first an example.
2 2 2
Example. e−x ∗ e−x = π2 e−x /2 .
p

4.1.1 Convolution of nonnegative functions

Proposition 4.1 Let f, g : IRN → [0, +∞] be Lebesgue measurable. Then

1. For every x ∈ IRN , the function y 7→ f (x − y)g(y) is Lebesgue measurable.

Z
2. The function x 7→ f (x − y)g(y) dy is Borel measurable.
IRN

Proof. 1) First, the function y 7→ g(y) is Lebesgue measurable by assumption. Second, we

claim that for every x ∈ IRN , the function h defined by h(y) = f (x − y) is Lebesgue measurable.
Indeed, let α ∈ IR be given and let A = {t ∈ IRN |f (t) < α}. Then A is Lebesgue measurable
because f is Lebesgue measurable. If B = {t ∈ IRN |h(t) < α}, then B = x − A and so B
is also Lebesgue measurable (the Lebesgue sigma algebra is invariant under translations and
symmetries). Finally y 7→ f (x − y)g(y) is Lebesgue measurable as a product of two Lebesgue
measurable functions.
2) There exists a Borel measurable function f1 : IRN → IR and a Borel mesurable function
g1 : IRN → IR such that f1 = f a.e and g1 = g a.e. Now, the function (x, y) 7→ x−y is continuous
and therefore Borel measurable. It folows that (x, y) 7→ f1 (x − y) is Borel measurable as a
composition of two Borel measurable functions. Since (x, y) 7→ g1 (y) is also Borel measurable,
Z (x, y) 7→ f1 (x − y)g1 (y) is Borel measurable. By Tonelli theorem,
it follows that the function
the partial integral x 7→ f1 (x − y)g1 (y) dy is Borel measurable.
IRN
Now, for Zevery x, and for almostZevery y, we have f (x − y)g(y) = f1 (x − y)g1 (y). Therefore
for every x, f (x − y)g(y) dy = f1 (x − y)g1 (y) dy. Hence the conclusion. 
IRN IRN

27
28 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

R
Definition 4.1 The function x 7→ IRN f (x − y)g(y) dy is denoted by f ∗ g, so that
Z
(f ∗ g)(x) = f (x − y)g(y) dy
IRN

and it is called the convolution of f and g. The map M+ × M+ → M+ , (f, g) 7→ f ∗ g is called

the product of convolution.

Proposition 4.2 The product of convolution satisfies the following properties

1. f ∗ g = g ∗ f (commutativity).

2. f ∗ (g ∗ h) = (f ∗ g) ∗ h (associativity).

3. f ∗ (g + h) = f ∗ g + f ∗ h (distributivity with respect to addition)

Z Z  Z 
4. f ∗g = f g .
IRN IRN IRN

5. {f ∗ g 6= 0} ⊂ {f 6= 0} + {g 6= 0}.

Proof. 1) Consider the change of variables u = x − y, or equivalently, y = x − u. In absolute

value, the Jacobian is 1 and the image of the new domain is IRN . Therefore
Z Z Z
(f ∗ g)(x) = f (x − y)g(y) dy = f (u)g(x − u)du = g(x − u)f (u) du = (g ∗ f )(x).
IRN IRN IRN

2) On the first hand, using Tonelli theorem,

Z Z Z 
(f ∗ g) ∗ h(x) = (f ∗ g)(y)h(x − y) dy = (f (z)g(y − z) dz h(x − y) dy
N IRN IRN
ZIR
= f (z)g(y − z)h(x − y) dy dz.
IRN ×IRN

On the other hand

Z Z Z 
f ∗ (g ∗ h)(x) = f (z)(g ∗ h)(x − z) dz = f (z) g(y)h(x − z − y) dy dz
N IRN IRN
ZIR
= f (z)g(y)h(x − z − y) dy dz.
IRN ×IRN

The above two integrals are equal by the change of variable u = y + z; v = z whose Jacobian
is 1 (write the details).

3) Follows from the linearilty of the integral.

4) Follows from Tonelli theorem and a change of variables (write the details).

5) Let x ∈ IRN be such that x ∈ / {f 6= 0} + {g 6= 0}. We claim that for any y ∈ IRN , either
g(y) = 0 or f (x−y) = 0. Indeed, otherwise, there exists y0 such that g(y0 ) 6= 0 and f (x−y0 ) 6= 0.
This means that y0 ∈ {g 6= 0} and x−y0 ∈ {f 6= 0}. Then x = (x−y0 )+y0 ∈ {f 6= 0}+{g 6= 0},
contrary to the assumption. It follows from the claim that g(y)f (x − y) = 0 for every y ∈ IRN .
This implies that (f ∗ g)(x) = 0. 
4.1. CONVOLUTIONS 29

4.1.2 Convolution of integrable functions

For simplicity, we write Lp for Lp (IRN ).

Proposition 4.3 Let f ∈ L1 and g ∈ Lp for some p ∈ [1, +∞]. Then

1. For almost every x ∈ IRN , the function y 7→ f (x − y)g(y) is Lebesgue-integrable.

Z
2. The function f ∗ g defined almost everywhere by (f ∗ g)(x) = f (x − y)g(y) dy belongs
IRN
to Lp (IRN ) and
||f ∗ g||p ≤ ||f ||1 ||g||p (Y oung 0 s inequality).

3. The class of f ∗ g in Lp (IRN ) depends only on the classes of f and g, that is, if f1 = f2
a.e. and g1 = g2 a.e, then f1 ∗ g1 = f2 ∗ g2 a.e.

Proof. We distinguish between three cases.

Z Z  Z 
Case 1. p = 1. We have |f | ∗ |g| = |f | |g| < ∞. This means that |f | ∗ |g|
IRN IRN IRN
is integrable. Next, let F (x, y) = f (x − y)g(y). Then, by Tonelli’s theorem
Z Z Z  Z
|F (x, y)| dx dy = |f (x − y)||g(y)|dy dx = |f | ∗ |g| dλN < ∞.
IRN ×IRN IRN IRN IRN

By Fubini’s theorem, the partial function y 7→ f (x − y)g(y) is integrable for almost every x.
Therefore the convolution is well defined almost
R everywhere.
R
The inequality in point 2. follows from | h| ≤ |h| and Tonelli theorem (interchange of
the order of integration). Indeed,
Z Z Z Z Z
||f ∗ g||1 = |f ∗ g| dx = f (x − y)g(y) dy dx ≤ |f (x − y)||g(y)| dy dx
Z Z Z  Z 
= |f (x − y)||g(y)| dx dy = |g| |f | = ||f ||1 ||g||1 .

Case 2. 1 < p < +∞. Let p0 be the conjugate of p. Writing

0
|f (x − y)g(y)| = |f (x − y)|1/p |g(y)| · |f (x − y)|1/p

and applying Hölder’s inequality, we get for every x

Z Z 1/p Z 1/p0
p
|f (x − y)||g(y)| dy ≤ |f (x − y)||g(y)| dy |f (x − y)| dy

1/p 1/p0
= (|f | ∗ |g|p )(x) ||f ||1 < ∞.

By case 1, the term on the right is finite for almost every x. This proves the integrability of
y 7→ f (x − y)g(y) for almost every x.
Raising to the power p both sides of the above inequality, we get
Z p
p p/p0
|(f ∗ g)(x)| ≤ |f (x − y)||g(y)| dy ≤ (|f | ∗ |g|p )(x)||f ||1 .

Integrating with respect to x and applying the estimate ||u ∗ v||1 ≤ ||u||1 ||v||1 established in
case 1, we see that f ∗ g ∈ Lp and
p/p0
||f ∗ g||pp ≤ ||f ||1 ||g p ||1 ||f ||1 ,
30 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

that is,
p/p0
||f ∗ g||pp ≤ ||f ||1 ||g||p ||f ||1 .
Raising to 1/p, we finally get ||f ∗ g||p ≤ ||f ||1 ||g||p .
Case 3. p = +∞. This case is straightforward (write the details). But note that in this case
f ∗ g is defined everywhere. Indeed, we have for every x and almost every y, |f (x − y)g(y)| ≤
|f (x − y)|||g||∞ . Therefore
Z Z
|f (x − y)g(y)| dy ≤ ||g||∞ |f (x − y)| dy = ||g||∞ ||f ||1 .

Therefore the function x 7→ f (x − y)g(y) is summable for every x and not only for almost every
x.
3. Let f1 = f2 a.e. and g1 = g2 a.e, and let

Ax = {y ∈ IRN |f1 (x − y)g1 (y) 6= f2 (x − y)g2 (y)}.

Then

Ax ⊂ {y ∈ IRN |f1 (x − y) 6= f2 (x − y)} ∪ {y ∈ IRN |g1 (y) 6= g2 (y)} = (x − {f1 6= f2 }) ∪ {g1 6= g2 }

Therefore λN (Ax ) ≤ λ(x − {f1 6= f2 }) + λ{g1 6= g2 } = λ({f1 6= f2 }) + λ{g1 6= g2 } = 0. This

N
(x − y)g1 (y) = f2 (x − y)g
means that for every x, f1Z Z 2 (y) for almost every y ∈ IR . Integrating
with respect to y, we get f1 (x − y)g1 (y) dy = f2 (x − y)g2 (y) dy when both integrals are
defined, that is for almost every x. This means that f1 ∗ g1 = f2 ∗ g2 a.e. . 

Proposition 4.4 Let p ∈ [1, ∞]. The convolution satisfies the following properties.

(i) If f ∈ L1 and g ∈ Lp then f ∗ g = g ∗ f (commutativity).

(ii) If f ∈ L1 , g ∈ L1 and h ∈ Lp then f ∗ (g ∗ h) = (f ∗ g) ∗ h (associativity).

(iii) If f ∈ L1 and g, h ∈ Lp then f ∗ (g + h) = f ∗ g + f ∗ h (distributivity with respect to

addition)

(iv) supp (f ∗ g) ⊂ supp (f ) + supp (g). In particular, if f and g have compact support, then
f ∗ g has also compact support.

Proof. The proof of the (i), (ii) and (iii) is similar to the proof of Proposition 4.2. As in
that proposition, we have {f ∗ g 6= 0} ⊂ {f 6= 0} + {g 6= 0} ⊂ supp (f ) + supp (g). Taking the
closure, we get the result. Finally, the sum of two compact sets of IRN is a compact set. 

Corollary 4.1 For every p ∈ [1, ∞], the convolution defines a bilinear continuous map from
L1 × Lp to Lp and from Lp × L1 to Lp .

4.1.3 Convolution with smooth functions and approximation

Definitions and notations. The space of k times continuously differentiable functions
with compact support is denoted by Cck (IRN ), so that Cck (IRN ) = C k (IRN ) ∩ Cc (IRN ). Let
α = (α1 , · · · , αN ) ∈ INN . It is called a multi-index. We set |α| = α1 + · · · + αN and call this
quantity the length of α. We write

∂ |α|
Dα f = .
∂xα1 1 · · · ∂xαNN
4.1. CONVOLUTIONS 31

∂7f
For example, D(1,4,2) f = . The space of infinitely many times differentiable functions
∂x∂y 4 ∂z 2
with compact support is denoted by Cc∞ (IRN ), that is, Cc∞ (IRN ) = C ∞ (IRN ) ∩ Cc (IRN ). This
space is not trivial for it contains the function ρ defined by
( 1
1
e ||x||2 −1 if ||x|| < 1
2
ρ(x) = e ||x|| −1 1B(0,1) =
0 if ||x|| ≥ 1

where ||x|| denotes the Euclidean norm of x ∈ IRN . We will prove in fact that for any open
subset Ω ⊂ IRN , the space Cc∞ (Ω) is dense in Lp (Ω).

Proposition 4.5 Let f ∈ L1 (IRN ) and g ∈ Cck (IRN ). Then f ∗ g ∈ C k (IRN ) and for every
multi-index α ∈ INN such that |α| ≤ k, we have

Dα (f ∗ g) = f ∗ Dα g.

Proof. Note first that our assumptions imply that f ∗g is defined everywhere because g ∈ L∞ .
We prove first the result for k = 0. We need to show the continuity of f ∗ g. Since g is
continuous, the function x 7→ f (y)g(x − y) is continuous and |f (y)g(x − y)| ≤ M |f (y)| where
M =Rsup |g|. By a theorem on the continuity of integrals depending on a parameter, the function
x 7→ IRN f (y)g(x − y) is continuous, that is f ∗ g is continuous.
∂F ∂g
We prove next the result for k = 1. Let F (x, y) = f (y)g(x − y). Then ∂xi
= f (y) ∂x i
(x − y)
∂F ∂g
and so | ∂x i
| ≤ M |f (y)| where M = sup | ∂x i
| and f ∈ L1 . The result follows from the theorem
on differentiation under the integral sign. The relation
∂(f ∗ g)
Z
∂g
(x) = f (y) (x − y) dy
∂xi IRN ∂xi
∗g)
shows that ∂(f
∂xi is continuous (by the theorem on the continuity of integrals depending on a
parameter). The general case follows by induction on k. 

Remark 4.1 The result of the previous proposition remains true if f ∈ Lp for an arbitrary
p ∈ [1, +∞]. Indeed, let x0 ∈ IRN be given. We will prove that f ∗ g is differentiable on a
neighborhood of x0 . Since g has compact support, the set B(x0 , 1) − supp g is bounded and so
it is contained in a compact set K. Now we can write
Z
(f ∗ g)(x) = f (y)g(x − y) dy ∀x ∈ B(x0 , 1)
K

because if y ∈
/ K, then x − y ∈ / supp g and so g(x − y) = 0. Keeping the same notation as
in the above proof, we have | ∂xi | ≤ M |f (y)| and now f ∈ Lp (IRN ) ⊂ Lp (K) ⊂ L1 (K). The
∂F

continuation is the same as above.

Corollary 4.2 Let f ∈ L1 (IRN ) and g ∈ Cc∞ (IRN ). Then f ∗ g ∈ C ∞ (IRN ) ∩ Lp (IRN ) for any
p ∈ [1, ∞].

Definition 4.2 A sequence of mollifiers1 is a sequence of functions (ρn ) ⊂ Cc∞ (IRN ) such that

1. ρn ≥ 0.

2. supp ρn ⊂ B 0 (0, n1 ).
R
3. IRN ρn = 1.
1
Suite régularisante.
32 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

1
A sequence of mollifiers always exist. Indeed, let for example ρ be defined by e ||x||2 −1 1B(0,1) and
let Z −1
ρn (x) = ρ nN ρ(nx).

Then ρn is a sequence of mollifiers. An exact formula for ρn is irrelevant to what follows.

If (ρn ) is sequence of mollifiers and f ∈ Lp , then by the previous proposition, f ∗ ρn is C ∞ ;
thus convoluting with a sequence of mollifiers makes the function very smooth and this is why
the elements of such sequence are called mollifiers.

Theorem 4.1 Let (ρn ) be a sequence of mollifiers and let f ∈ Lp (IRN ) for some p ∈ [1, +∞[.
Then
lim ||f ∗ ρn − f ||p = 0,
n→∞
that is, the sequence (f ∗ ρn ) converges to f in Lp .
We prove this theorem in two steps. Let us record the first step in a lemma.

Lemma 4.1 Let (ρn ) be a sequence of mollifiers and let f ∈ Cc (IRN ). Then (f ∗ ρn ) converges
to f uniformly on compact subsets of IRN . In particular for any compact subset K ⊂ IRN ,
||ρn ∗ f − f ||Lp (K) → 0 as n → ∞.

Proof. Let K ⊂ IRN be a fixed compact set and let ε > 0 be given. Uniform continuity of f
on the compact set K + B 0 (0, 1) implies that there exists δ > 0 (depending on K and ε) such
that
|f (x − y) − f (x)| < ε ∀x ∈ K, ∀y ∈ B(0, δ).
Z
Since ρn (y) dy = 1, we can write
IRN
Z Z
(f ∗ ρn )(x) − f (x) = f (x − y)ρn (y) dy − f (x) ρn (y) dy
N IRN
ZIR
= [f (x − y) − f (x)] ρn (y) dy
N
ZIR
= [f (x − y) − f (x)] ρn (y) dy
1
B(0, n )

Let therefore n > 1δ and x ∈ K. Then n1 < δ and so B(0, n1 ) ⊂ B(0, δ). If follows that
|f (x − y) − f (x)| < ε when y ∈ B(0, 1/n) and so
Z
|(f ∗ ρn )(x) − f (x)| ≤ ε ρn (y) dy = ε.
1
B(0, n )

This means that f ∗ ρn converges to f uniformly on K. It follows that

Z
|(f ∗ ρn )(x) − f (x)|p ≤ λN (K)εp
K

and so
||f ∗ ρn − f ||Lp (K) ≤ λN (K)1/p ε
This means that ||ρn ∗ f − f ||Lp (K) → 0. 

Proof of Theorem 4.1. Let ε > 0 be given. By Theorem 3.2, there exists f1 ∈ Cc (IRN )
such that ||f − f1 ||p < ε. Now we know from a previous proposition that

supp (f1 ∗ ρn ) ⊂ supp (f1 ) + B 0 (0, 1/n) ⊂= supp (f1 ) + B 0 (0, 1/n) ⊂ supp (f1 ) + B 0 (0, 1).
4.2. FOURIER TRANSFORMS 33

If we set K := supp (f1 ) + B 0 (0, 1), we see that K is a fixed compact set (independent of n)
that contains both supp (f1 ∗ ρn ) and supp (f1 ).
By the previous lemma
Z
|(f1 ∗ ρn )(x) − f1 (x)|p dx → 0.
K

But Z Z
|(f1 ∗ ρn )(x) − f1 (x)|p = |(f1 ∗ ρn )(x) − f1 (x)|p dx.
IRN K
Therefore ||f1 ∗ ρn − f1 ||p → 0 and so ||f1 ∗ ρn − f1 ||p < ε for all n large enough.
Now by the triangle inequality

||f ∗ ρn − f ||p ≤ ||(f − f1 ) ∗ ρn ||p ≤ ||f1 ∗ ρn − f1 ||p + ||f1 − f ||p .

By Young’s inequality, ||(f − f1 ) ∗ ρn ||p ≤ ||f − f1 ||p ||ρn ||1 = ||f − f1 ||p < ε. Therefore

||f ∗ ρn − f ||p < 3ε.

for all n large enough. 

Corollary 4.3 For 1 ≤ p < +∞, Lp (IRN ) ∩ C ∞ (IRN ) is dense in Lp (IRN ).

Corollary 4.4 Let Ω ⊂ IRN be open and let 1 ≤ p < ∞. Then Cc∞ (Ω) is dense in Lp (Ω).
Proof. Let f ∈ Lp (Ω). We will find a sequence (fn ) in Cc∞ (Ω) such that ||fn − f ||Lp (Ω) → 0.
Let (
f (x) if x ∈ Ω
f¯(x) =
0 otherwise.

Then f¯ ∈ Lp (IRN ). Let

Kn = {x ∈ IRN | ||x|| ≤ n and dist (x, Ωc ) ≥ 2/n}

S
Then (Kn ) is an increasing sequence of compact subsets of Ω such that Ω = Kn .
Set gn = 1Kn f¯ and fn = gn ∗ ρn where (ρn ) is a smoothing sequence. Then

supp (fn ) ⊂ B 0 (0, 1/n) + Kn ⊂ Ω.

It follows that fn ∈ Cc∞ (Ω). Now

||fn − f ||Lp (Ω) = ||fn − f¯||Lp (IRN ) = ||gn ∗ ρn − f¯k|Lp (IRN )

≤ ||gn ∗ ρn − f¯ ∗ ρn k|p + ||f¯ ∗ ρn − f¯||p by the triangle inequality
≤ ||ρn ||1 ||gn − f¯||p + ||f¯ ∗ ρn − f¯||p by Young’s inequality
= ||gn − f¯||p + ||f¯ ∗ ρn − f¯||p .

It follows from the dominated convergence theorem that ||gn − f¯||p → 0 (check this). By the
previous theorem ||f¯ ∗ ρn − f¯||p → 0. Hence ||fn − f ||Lp (Ω) → 0. 

4.2 Fourier transforms

In this section, we denote by L1 the space of complex valued integrable functions on IRN . The
usual dot product in IRN is denoted by x · y, that is, for x = (x1 , . . . , xN ) and y = (y1 , . . . , yN ),
x · y = x 1 y1 + · · · + x N yN .
34 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

4.2.1 Fourier transforms in L1

Lemma 4.2 Let f ∈ L1 (IRN ) and ω ∈ IRN , then the function x 7→ e−2iπω·x f (x) is integrable
on IRN .
Proof. The function x 7→ e−2iπω·x is continuous and therefore Lebesgue measurable. Thus
the function x 7→ Re−2iπω·x f (x) is Lebesgue
R measurable as a product of two Lebesgue measurable
functions. Next, |e −2iπω·x f (x)| dx = |f (x)|dx < ∞. 

Definition 4.3 Let f ∈ L1 (IRN ). The Fourier transform of f is the complex valued function
fˆ defined by Z
fˆ(ω) = e−2πiω·x f (x) dx.
IRN

Example 4.1 Let f = 1[−1,1] . Then, fˆ(0) = 2 and for ω 6= 0

x=1
e−2πiω·x e2πiω − e−2πiω
Z Z 1
−2πiωx sin(2πω)
ˆ
f (ω) = e f (x) dx = e−2πiωx dx = = = .
IR −1 −2πiω x=−1 2πiω πω

Observe that fˆ is continuous although f is not. This is not a coincidence.

Notation. We denote by C0 = C0 (IRN , C) the space of all continuous functions ϕ : IRN → C

such that lim||x||→∞ ϕ(x) = 0. A function in C0 is bounded and uniformly continuous. If we set
||ϕ||∞ = supx∈IRN |ϕ(x)|, then || · ||∞ is a norm on C0 .

Proposition 4.6 The following hold.

1. For all f ∈ L1 , fˆ ∈ C0 .
2. The map F : (L1 , || · ||1 ) → (C0 , || · ||∞ ), f 7→ fˆ is linear and bounded.
Proof. The continuity of fˆ follows from a theorem on the continuity of integrals depending
on a parameter. Next, we need to show that lim|ω|→∞ fˆ(ω) = 0. For this, it is enough to
understand the case N = 1. Let ωn be a sequence of real numbers that tend to ∞. Since
eiπ = −1, we have
Z
f (ωn ) = e−2iπωn x f (x) dx
ˆ (4.1)
Z
1
= − e−2iπωn (x− 2ωn ) f (x) dx.

The change of variables u = x − 2ω1n gives

Z   Z  
ˆ −2iπωn u 1 −2iπωn x 1
f (ωn ) = − e f u+ du = − e f x+ dx. (4.2)
2ωn 2ωn
Adding together (4.1) and (4.2), we get
Z  
ˆ −2iπωn x 1
2f (ωn ) = e f (x) − f (x + ) dx.
2ωn
It follows that
1
|fˆ(ωn )| ≤ τ 1 f −f .
2 2ωn 1
But the last quantity tends to zero as n → ∞ by the continuity in the mean proposition. This
proves the first statement.
For the second statement, the linearity of F follows from linearity of the integral and the
boundedness follows from |fˆ(ω)| ≤ ||f ||1 for all ω ∈ IRN and so ||fˆ||∞ ≤ ||f ||1 , i.e ||F(f )||∞ ≤
||f ||1 . 
4.2. FOURIER TRANSFORMS 35

Corollary 4.5 If (fn ) converges to f in L1 , then (fˆn ) converges uniformly to fˆ.

∗ g = fˆĝ.
Proposition 4.7 Let f, g ∈ L1 . Then f[

Proof. Recall that f ∗ g ∈ L1 and so f[ ∗ g exists. Recall also that, by Tonelli’s theorem, the
function (x, y) 7→ f (x − y)g(y) is integrable on IRN × IRN . Therefore, by Fubini’s theorem,
Z
f ∗ g(ω) = (f ∗ g)(x)e−2πiω·x dx
[
Z Z 
= f (x − y)g(y) dy e−2πiω·x dx
Z Z
= f (x − y)e−2πiω·(x−y) g(y)e−2πiω·y dx dy
Z Z 
= g(y)e−2πiω·y f (x − y)e−2πiω·(x−y) dx dy

= fˆ(ω)ĝ(ω).

Proposition 4.8 (Exchange theorem) Let f, g ∈ L1 . Then f ĝ and fˆg are integrable and
Z Z
f ĝ dλN = fˆg dλN .

Proof. f ĝ is integrable because f is integrable and ĝ is bounded. Similarly for fˆg. By Tonelli’s
theorem the function (x, ω) 7→ f (ω)g(x)e−2iπω·x is integrable and so by Fubini’s theorem,
Z Z Z 
−2πiω·x
f (ω)ĝ(ω)dω = f (ω) g(x)e dx dω
Z Z 
−2πiω·x
= g(x) f (ω)e dω dx
Z
= g(x)fˆ(x) dx.

Theorem 4.2 (Inversion theorem) Let f ∈ L1 be such that fˆ ∈ L1 . Then

Z
f (x) = fˆ(ω)e2πiω·x dω
IRN

for almost every x ∈ IRN .

Proof. See the exercises.

Corollary 4.6 The operator F : L1 → C0 , f 7→ fˆ is injective.

Remark 4.2 Let F be the operator defined on L1 by

Z Z
+2πiω·x
F(f )(x) = f (ω)e dω = f (y)e+2πiy·x dy.
IRN IRN

Then F has exactly the same properties as the operator F. Observe also the following: given
a function f , let f˜ be defined by f˜(x) = f (−x). Then F(f ) = F(f˜).
36 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

4.2.2 Differentiability
Lemma 4.3 Let f : IR → C be continuously differentiable, integrable and such that f 0 is
integrable. Then limx→±∞ f (x) = 0.
Rx
Proof. RBy the fundamental theorem of calculus, f (x) = f (0) + 0 f 0 (t) dt. Since f 0 is
x
integrable, 0 f 0 (t) dt has a limit as x → ±∞. Therefore f has a limit as x → ±∞. Since f is
integrable, the limit is necessarily zero. 

Remark 4.3 If g : IR → C is integrable then g need not have a limit at ±∞ (give an example).
However, if g is integrable and has a limit at +∞ or −∞, then necessarily this limit is zero.
Indeed, let L = limx→∞ g(x). Then, for every ε > 0 there exists a ∈ IR such that |g(x) − L| < ε
for Rall x ≥ a. If L 6= 0, take ε = |L|/2 > 0. Then, by the triangle inequality, |g(x)| > |L|/2 and
∞ R∞
so a |g(x)| dx ≥ a (|L|/2) dx = ∞. But this contradicts the integrability of g. Therefore
L = 0. A similar reasoning yields limx→−∞ g(x) = 0.

Proposition 4.9 Let f : IRN → C be of class C k for some k ∈ N ∗ . Suppose that Dα f ∈ L1

for all α such that |α| ≤ k. Then

F(Dα f )(ω) = (2πi)|α| ω α F(f )(ω).

Proof. By induction on the length of α, it is enough to treat the case k = 1, N = 1. Therefore,

we assume that f ∈ C 1 (IR, C) and that f, f 0 ∈ L1 (IR). We need to show that fb0 (ω) = 2πiω fˆ(ω).
For this, observe that
Z Z
fb0 (ω) − 2πiω fˆ(ω) = f 0 (x)e−2πiωx dx − 2πiω f (x)e−2πiωx dx
ZIR IR

f 0 (x)e−2πiωx − 2πiωf (x)e−2πiωx dx

=
ZIR
∂
f (x)e−2πiωx dx


=
IR ∂x
Z n
∂
f (x)e−2πiωx dx


= lim
n→∞ −n ∂x

= lim f (n)e−2πinω − f (−n)e2πinω

n→∞
=0

by the previous lemma. 

Notation. Let α = (α1 , . . . , αN ) ∈ INN . For x = (x1 , . . . , xN ) ∈ IRN , we set

M α (x) = xα1 1 · xαNN .

Proposition 4.10 Let k ∈ IN∗ and let f : IRN → C be a function such that M α f ∈ L1 for
every multi-index α such that |α| ≤ k. Then fˆ ∈ C k and

Dα fˆ = (−2πi)|α| M
[ αf .

Proof. This is a consequence of the theorem on differentiation under the integral. 

Remark 4.4 In the literature, there are other definitions of the Fourier transform. For example
some authors define the Fourier transform by
Z
1
fˆ(ω) = f (x)e−iω·x dx.
(2π)N/2 IRN
With this definition we have
4.2. FOURIER TRANSFORMS 37

∗ g = (2π)N/2 fˆĝ.
1. f[

2. fˆ0 (ω) = iω fˆ(ω).

Z
1
3. f (x) = (2π)N/2 fˆ(ω)e+iω·x dω.
IRN

4.2.3 Fourier transforms in L2

We cannot define the Fourier transform on L2 directly as we did for functions in L1 . So we use
an indirect way: we first restrict the definition of Fourier transform to a space S common to
L1 and L2 and which is dense in L2 and then we extend the definition to L2 by density and
continuity.

Definition 4.4 We set

S = {f ∈ C ∞ (IRN , C)|∀α, β ∈ INN , sup |xα Dβ f (x)| < ∞}.

x∈IRN

S is called the space of rapidly decreasing functions on IRN .

Comments. 1) A function f in S vanishes at infinity together with its derivatives. In

particular f and all its derivatives are bounded.
2
2) S contains the space Cc∞ (IRN , C) but it contains more, for example the function x 7→ e−||x|| .
3) The condition supx∈IRN |xα Dβ f (x)| < ∞ can be replaced by the condition

lim |xα Dβ f (x)| = 0.

||x||→∞

4) S is vector space over C.

Proposition 4.11 The following hold.

1. S is stable under multiplication.

2. ∀f ∈ S, ∀α, β ∈ INN , M α Dβ f ∈ S where M α (f )(x) = xα f (x). In particular, all the

derivatives of f are in S (S is stable under differentiation).

3. S ⊂ Lp (IRN , C) for all 1 ≤ p ≤ ∞.

4. S is dense in Lp (IRN , C) for all 1 ≤ p < ∞.

Proof. We prove the theorem for N = 1 because it is easier to write and understand.
1. By the Leibniz rule (f g)(n) = nk=0 Cnk f (k )g (n−k) . Therefore
P

n
X
xm (f g)n = Cnk xm f (k) g (n−k) .
k=0

Since all the derivatives are bounded the result follows.

2. We have to prove that xγ (xα f (β) )(n) is bounded. By the Leibniz rule, this is a sum of powers
of x multiplied by derivatives of f and each term is bounded by assumption.
3.
R We claim first that L1 ∩ L∞ ⊂ Lp for every p ∈ [1, ∞]. Indeed, let f ∈ L1 ∩ L∞ , then
|f | ≤ ||f ||p−1
p ∞
∞ ||f ||1 . Now we already observed that a function in S is bounded so S ⊂ L . It
38 CHAPTER 4. CONVOLUTIONS AND FOURIER TRANSFORMS

is therefore enough (and necessary) to prove that S ∈ L1 . So let f ∈ S Then |f (x)| ≤ A and
A+B
|x2 f (x)| ≤ B. Therefore |f (x)| ≤ 1+x 1
2 and the last function is in L .

4. This is because that S contains Cc∞ (IRN , C) which is dense in Lp .



Since S ⊂ L1 , the Fourier transform is defined on S by the same formula, i.e.

Z
fˆ(ω) = f (x)e−2πix·ω dx.
IRN

Theorem 4.3 The Fourier transform is a bijection from S to itself.

Proof. Let f ∈ S. We show first that fˆ ∈ S. We know that M β f and Dα (M β f ) are in

S ⊂ L1 . Using two previous propositions, we have
1
ω α Dβ fˆ(ω) = (−2πi)|β| ω α F(M β f )(ω) = (−2πi)|β| F(Dα (M β f )) ∈ C0
(−2πi)|α|

This means that fˆ ∈ S.

Next we prove that F is bijective. Let
Z
F(f )(x) = f (y)e+2πiy·x dy.
IRN

Then the inversion theorem implies that f = FF(f ) if both f and F(f ) are in L1 . Now if
f ∈ S then by the above F(f ) ∈ S ⊂ L1 and so we have f = FF(f ) . This identity proves
that F is injective. Now let f ∈ S. Then, f and fˆ belong to S ⊂ L1 . Let h(ω) = fˆ(−ω). Then
h ∈ S. By the inversion theorem,
Z Z
f (x) = ˆ
f (ω)e+2πix·ω
dω = fˆ(−ω)e−2πix·ω dω by the change of variables ω → −ω
IRN IRN
Z
= h(ω)e−2πix·ω dω = F(h)(x)
IRN

Thus, f = F(h). This proves that F is surjective. The proof shows that in fact F −1 = F. 

Proposition 4.12 For all f, g ∈ S, we have F(f ∗ g) = F(f )F(g) and F(f g) = F(f ) ∗ F(g).

Proof. The first identity holds for functions in L1 and so also for functions in S. Exactly in
the same way F −1 takes convolutions into products. Let now f, g ∈ S, and let ϕ = F(f ) and
ψ = F(g). Then

F(f g) = F(F −1 ϕF −1 ψ) = F(F −1 (ϕ ∗ ψ) = ϕ ∗ ψ = F(f ) ∗ F(g).

Remark 4.5 It follows from the computations above that S is stable under convolutions.

In the following, L2 denotes the space L2 (IRN , C) equipped with the complex dot product
Z
hf, gi = f ḡ dλ.
IRN

Proposition 4.13 For all f, g ∈ S, we have hfˆ, ĝi = hf, gi. In particular ||fˆ||2 = ||f ||2 .
4.2. FOURIER TRANSFORMS 39

Proof. Let f, g ∈ S. Using the exchange theorem and the inversion theorem we get
Z Z
¯
ˆ ˆ¯
hf , ĝi = f (t)g((t) dt = fˆ(t)g((−t)
ˆ ˆ dt
Z Z Z
ˆ
ˆ
= f (t)ḡ(−t) dt = f (−t)ḡ(−t) dt = f (t)ḡ(t) dt = hf, gi.


It follows that F is an isometry between (S, || · ||2 ) into (L2 , || · ||2 ). Since S is dense in L2
and L2 is complete, there is a unique continuous extension of F to L2 . Indeed, let f ∈ L2 ,
then there exists a sequence (fn ) in S that converges to f in L2 . Then F(fn ) is a Cauchy
sequence L2 because ||F(fn ) − F(fm )|| = ||fn − fm ||. Therefore F(fn ) is convergent in L2 . We
set G(f ) = lim F(fn ). This makes sense because the result is independent of the sequence fn .
Indeed, if gn is another sequence of S that converges to f , then ||F(fn ) − F(gn )|| = ||fn − gn ||.
To prove uniqueness of this extension, let G 0 be another continuous extension to L2 of F.
Then G and G 0 coincide on S. Let f ∈ S and let (fn ) be a sequence in S that converges to f in
L2 . Then G(fn ) = G 0 (fn ). Letting n → ∞, we get G(f ) = G 0 (f ).
It is not difficult to see that G is also linear and bounded.
In the sequel we still denote this extension by F.

Theorem 4.4 (Fourier transform in L2 ) The operator F : L2 → L2 is an isometric isomor-

phism that conserves the dot product.
Proof. We know that F is linear and bounded. Let f, g ∈ L2 and let (fn ) and (gn ) be
two sequences in S that converge in L2 to f and g respectively. By the previous proposition
hF(fn ), F(gn )i = hfn , gn i. Letting n → ∞ and using the continuity of F and the dot product
we get
hF(f ), F(g)i = hf, gi (Plancherel identity).
In particular
||F(f )||2 = ||f ||2 (Parseval identity).
Now we extend the operator F to L2 by the formula

F(f ) = F(f˜) ∀f ∈ L2 where f˜(x) = f (−x).

This is extension of the previous formula for F since whenever f ∈ L1 , we have

Z Z
˜ −2πit·x
F(f )(x) = f (−t)e dt = f (y)e+2πiy·x dy.
IRN IRN

It follows that ||F(f )||2 = ||F(f˜)||2 = ||f˜||2 = ||f ||2 . This means that F is also bounded. Next,
let fn be a sequence of S that converges to f in L2 . We know that FFfn = fn . By continuity
of F and F we get FFf = f . Similarly we get FF(f ) = f . 
sin(2πω)
Example of use of Parseval’s identity. Let f = 1[−1,1] . We know that fˆ(ω) = πω .
Parseval’s identity implies that

sin(2πω) 2
Z ∞ Z ∞ 
(1[−1,1] )2 dx = dω.
−∞ −∞ πω

sin(2πω) 2
Z ∞ 
Therefore dω = 2. By the change of variables t = 2πω, we get
−∞ πω

sin t 2 sin t 2
Z ∞  Z ∞ 
π
dt = π and so dt = .
−∞ t 0 t 2