Theory of function and distribution spaces

Willem van Zuijlen

27th April 2022

Contents
Introduction 3
1 Spaces of differentiable functions and testfunctions 5
2 Distributions 10
3 Topological vector spaces 19
4 Topologies on the spaces of testfunctions and distributions 25
5 Compactly supported distributions 36
6  Structure theorems 42
7 Intermezzo: Convolutions of functions 45
8 Convolution of distributions with testfunctions 51
9 Distributions of finite order 60
10 Convolutions of distributions 65
11 Fundamental solutions of partial differential operators 68
12 Sobolev spaces 74
13 Solutions to elliptic PDEs in Sobolev spaces 80
14 The Schwartz space 85
15 Tempered distributions 92
16 The Fourier transformation 97
17 Convolution of tempered distributions 108
18  The Fourier transformation on E 0 116
19 Fourier multipliers 118
20 Fractional Sobolev spaces 123
21 Besov spaces defined by Littlewood–Paley decompositions 127
21.1 Comments ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
22  Diversity of Besov spaces and inclusions 148
23 Embeddings of Besov spaces and Sobolev spaces 151
24  Besov spaces related to other spaces 157
25 The Hörmander-Mikhlin inequalities for Fourier multipliers 160
25.1 Comments ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

1
26 Products of tempered distributions 171
26.1 Comments... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
27 Paraproducts and resonance products in Besov spaces 179
27.1 Comments ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
28 Solutions to elliptic PDEs in Besov spaces 196
29 Overview 200
29.1 Spaces of functions and distributions . . . . . . . . . . . . . . . . . . . . . 200
29.2 Characterisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
29.3 Convergences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
29.4 Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
29.5 Continuity of operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
29.6 Metrizability and completeness statements . . . . . . . . . . . . . . . . . . 203
29.7 Denseness and separability statements . . . . . . . . . . . . . . . . . . . . 204
29.8 Continuous embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
29.9 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
29.10Other statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
A Preliminaries on Lp spaces 208
B Taylor’s formula 212
B.1 For one dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
B.2 Taylor expansion in higher dimensions . . . . . . . . . . . . . . . . . . . . 213
C Integration by parts 214
Bibliography 215

2
Introduction

The aim of these notes is to give a self-contained precise introduction into the spaces of
distributions, Sobolev and in particular Besov spaces. For the latter we rely on the Fourier
transform on tempered distributions. Those spaces are commonly used in the theory of
partial differential equations, hence also some sections are devoted to applications to
partial differential equations.
The first part of these notes concentrate on introducing the essentials of distribution
spaces; their topologies, convolution and mollifiers and the Fourier transformation on
tempered distributions. There is a lot of good literature on the theory of distributions,
for example [Don69] (quite an old fashioned way of writing and no inner references,
though quite complete), [DK10] [Fri98] (both good introduction to distribution spaces,
comparable to the first part of these notes, approach the theory without requiring know-
ledge on topology), [Str03] (similar the previous two, though different from the taste
of the author, very sparse on topological issues), [Hor66] (is more an introduction into
topological vector spaces, with distributions a final section), [Leo17] (a brief introduction
to distributions in order to be able to introduce Sobolev spaces and also Besov spaces,
but without using the Fourier transformation).
The second part of these notes concentrates on Besov spaces. The notes differ from
the literature in that the proofs contain more details and rigour. Other references in
which Besov spaces are introduced are [BCD11] (these notes have a lot of overlap with
Section 2 of that book, which focusses less on the details but contains more content
on the applications), [Gra14] (contains a brief introduction into Besov spaces, but also
contains many more function spaces), [Saw18] (contains most of the contents of these
notes, though written in a different style), [ST87] , [Tri83] , [Tri92] , [Tri78] , [Trè06] (all
very extensive books on function spaces, containing also Besov spaces, also the style here
differs from the one in these notes).
Acknowledgements The first version of these notes was written for the course
“Theory of Function Spaces and Applications” in spring 2020 at the FU Berlin (online).
Five strong students followed the course and gave valuable feedback on the lecture notes.
Special thanks go to A.C.M. van Rooij, who carefully read the lecture notes and gave a
lot of feedback and suggestions.

3
Conventions and notation
• N = {1, 2, 3, . . . }, N0 = N ∪ {0} and N−1 = {−1, 0} ∪ N.
• d is an element of N.
• Ω is a nonempty open subset of Rd .
• F is either R or C. qP
d
• For x ∈ Rd or x ∈ Cd
we write |x| for its Euclidean norm 2
i=1 |xi | , |x|1 =
Pd
i=1 |xi | and |x|∞ = maxi∈{1,...,d} |xi |.
• For a, b ∈ R we write a ∨ b = max{a, b} and a ∧ b = min{a, b}.
• For x ∈ Rd , r > 0 we write B(x, r) for the (Euclidean) ball in Rd with centre x and
radius r:

B(x, r) = {y ∈ Rd : |x − y| < r}.

(See also 1.7.)

• For a set A ⊂ Ω we write A◦ for its interior and A for its closure, see also 3.1.
• We write ∇ for the gradient of a function, which is the vector consisting of first
derivatives

∇f = (∂1 f, · · · , ∂d f ).

• For a set A ⊂ Ω we write 1A for the indicator function of A, see Definition 2.4.
• We write k · kLp for the norm on the Lp spaces. See Section A. As is common, we
do not distinguish between a function in Lp and its corresponding equivalence class
in Lp .
• For the inner product on L2 we write h·, ·iL2 (to avoid confusion with the notation
h·, ·i for the pairing between distributions and test functions). So
Z
hf, giL2 = f g.

• (Notation of limits of partial sums) Suppose (vn )n∈N is a sequence in a topological

N →∞
vector space X. We say that ∞
PN
n=1 vn −−−−→ v for some v ∈ X
P
P∞ n=1 vn exists if
and write n=1 vn for v.
If I is a countable set and vi ∈ X for i ∈ I, then we say that
X
vi
i∈I
P∞
exists, if there exists a v ∈ X such that for each bijection q : N → I, n=1 vq(n) =v
P
in X, and write i∈I vi for v.

4
1 Spaces of differentiable functions and testfunctions

In the next section, Section 2, we introduce the objects called distributions, which play
the central role of this text. A distribution can be viewed as a sort of generalised function.
As we will see, many functions like for example all continuous functions “are” or “can be
viewed” as distributions. Moreover, the distributions, like the differentiable functions,
form a vector space on which operations like translation, multiplication with differentiable
functions and differentiation are defined and follow the usual formal rules of calculus.
With the important difference that all distributions are differentiable, in the sense that
a derivative of a distribution always exists as a distribution.
This is a huge advantage which makes the theory of distributions very suitable as a
tool for (partial) differential equations, of which we will see a little bit for example in
Section 11.
In this section we consider the space D(Ω) of testfunctions on Ω, which are smooth
functions with compact support. A distribution, as defined later in Definition 2.1, is a
linear function D(Ω) → F with certain continuity properties. In this section we show
that the space D(Ω) is not empty, and on the contrary, is large enough in the sense that
a compact set K and a closed set F with K ∩ F = ∅ can be separated by a testfunction
in the sense that there exists a testfunction which equals 1 on K and equals 0 on F
(Lemma 1.13). Moreover, the notion of a partition of unity will be introduced which will
show its use multiple times (Definition 1.10).
Remember that Ω is a nonempty open subset of Rd and that the underlying field F
is either R or C.
Before we define the notion of a testfunction in Definition 1.5, we introduce some
definitions and recall a fact about the space of k-times continuously differentiable func-
tions.

Definition 1.1. Let f : Ω → F be a continuous function. We define the support of f ,

supp f , to be the closure in Ω of the set

{x ∈ Ω : f (x) 6= 0}. (1.1)

This means that it is the set of all x such that for all neighbourhoods V of x (open set
that contains x) there exists an element y in that neighbourhood such that f (y) 6= 0.
Let A ⊂ Ω. We say that f vanishes on A if f = 0 on A, i.e., f (x) = 0 for all x ∈ A.
Let U be the collection of all open subsets of Ω on which f vanishes. Then U := U is
S

the largest open subset of Ω on which f vanishes, and supp f = Ω \ U .

If F is a set of functions Ω → F, then we write Fc for the subset of compactly
supported functions in F , i.e., Fc = {f ∈ F : supp f is compact}.

Definition 1.2. • We write ei for the basis vector in Rd in the i-th direction, for
i ∈ {1, . . . , d}.

5
 Pd
• For α ∈ Nd0 , we write |α| = i=1 αi .

• For α ∈ Nd0 and any |α|-times continuously differentiable function f : Ω → F, we

write
∂ α f = ∂1α1 . . . ∂dαd f,
where ∂i is the partial derivation with respect to the i-th coordinate, i.e., for a
differentiable function f : Ω → F,
f (x + hei ) − f (x)
∂i f (x) = lim (x ∈ Ω).
h→0 h
• For a one-dimensional differentiable function f : R → F we will also write ∂f or
f 0 to denote its derivative, and ∂ k f for its k-th derivative (the letter D will not be
used for a derivative, but will be reserved for Fourier multipliers, see Section 19).
Definition 1.3. • We write C(Ω, F) or C(Ω) for the set of continuous functions
Ω → F. We will also write C 0 (Ω) = C(Ω) and
kϕkC 0 (Ω) = sup |ϕ(x)| (ϕ ∈ C(Ω)),
x∈Ω

observe that kϕkC 0 (Ω) = kϕkL∞ (Ω) for ϕ ∈ C(Ω).

In general, we will write k · kC 0 and k · kLp instead of k · kC 0 (Ω) and k · kLp (Ω) .

• For k ∈ N we write C k (Ω, F) or C k (Ω) for the k-times continuously differentiable

functions Ω → F, and k · kC k (Ω) : C k (Ω) → [0, ∞] for

kf kC k (Ω) = max k∂ β f kL∞ (f ∈ C k (Ω, F)). (1.2)

β∈Nd0 :|β|≤k

In general, we will write k · kC k instead of k · kC k (Ω) .

• C ∞ (Ω, F) or C ∞ (Ω) is the set of ∞-times continuously differentiable functions
Ω → F, i.e.,
C ∞ (Ω) =
\
C k (Ω).
k∈N

A function f : Ω → F is called smooth if it is in C ∞ (Ω).

• For a subset A ⊂ F and for k ∈ N0 ∪{∞} we write C k (Ω, A) for the set of functions in
C k (Ω, F) which take their values in A, i.e., those f ∈ C k (Ω, F) for which f (Ω) ⊂ A.
Pn
1.4. As on Rn for some n ∈ N the norm | · |1 and | · |∞ defined by |x|1 = i=1 |xi | and
|x|∞ = maxni=1 |xi | for x ∈ Rn are equivalent, or more specifically:
|x|∞ ≤ |x|1 ≤ n|x|∞ (x ∈ Rn ),
it is easy to see that there exists a C > 0 such that
X
kf kC k ≤ k∂ β f kL∞ ≤ Ckf kC k (f ∈ C k (Ω, F)).
β∈Nd0 :|β|≤k

6
Definition 1.5 (Testfunctions). D(Ω) is defined to be the vector space Cc∞ (Ω). An
element of D(Ω) is called a testfunction.
Later, in Definition 4.1, we equip D(Ω) with a topology.

The next lemma shows there exist many testfunctions, namely one can separate points
from closed sets that do not contain that point.
Lemma 1.6. Let x ∈ Rd and U be an open subset of Rd such that x ∈ U . There exists
a testfunction ϕ : Rd → [0, 1] such that ϕ(x) = 1 and supp ϕ ⊂ U .

Proof. We take x = 0 and show that for every ε > 0 there exists a function ψε such that
ψε (0) = 1 and supp ψε ⊂ B(0, ε). Consider the function ψε : Rd → [0, ∞) defined by
 1

e |y|2 −ε2 if |y| < ε,
ψε (y) =
0 if |y| ≥ ε.
One can prove that this function is C ∞ by using that limt→∞ p(t)e−t = 0 for any poly-
nomial p. Then ψε is strictly positive at 0 with support in B(0, ε).

Exercise 1.A. Prove that the function f : R → [0, 1] defined by

(
0 t ≤ 0,
f (t) = 1
e− t t > 0.
is smooth.
Exercise 1.B. Let d = 1 and ϕ be a nonzero testfunction. How that ϕ0 is nonzero as
well and conclude that ∂ k ϕ is nonzero for all k ∈ N0 .
Exercise 1.C. Prove the following statement. For any sequence (xn )n∈N in Ω of dis-
tinct elements such that for each compact K ⊂ Ω there are only finitely many elements
of the sequence in K (that is, no subsequence converges), and for any sequence (λn )n∈N
there exists a smooth function ψ on Ω with ψ(xn ) = λn for all n ∈ N.

We will now prepare ourselves to show that there exist so called partitions of unity
(Definition 1.10). They will be used often in the following. First we recall a definition to
show that Ω can be written as a union of compact sets.
1.7 (Notation). For x ∈ Rd , r > 0 we write B(x, r) for the (Euclidean) ball in Rd with
centre x and radius r:
B(x, r) = {y ∈ Rd : |x − y| < r}.
Theorem 1.8. There exists an increasing sequence of compact sets (Kn )n∈N such that
◦
Kn ⊂ Kn+1 for all n ∈ N and
[
Ω= Kn .
n∈N

Consequently, for each compact set K ⊂ Ω there exists an n ∈ N such that K ⊂ Kn .

7
Proof. Observe that if Ω = Rd , then we can take Kn to be the closure of the ball around
0 with radius n: B(0, n).
Let us now assume that Ω 6= Rd . We first prove that Ω is the union of countably
many closed sets in Rd . Let f : Ω → [0, ∞) be such that f (x) is the distance from x to
Rd \ Ω, i.e.,

f (x) = inf{|x − y| : y ∈ Rd \ Ω} (x ∈ Ω).

Then f is a continuous function and therefore An = f −1 [ n1 , ∞) is a closed subset of Ω,

An ⊂ An+1 for all n ∈ N and Ω = n∈N An .
S

Now it is straightforward to check that Kn = An ∩ B(0, n) satisfies the conditions.

S ◦
The consequence follows by using the fact that Ω = n∈N Kn .

Definition 1.9. Let E be an open subset of Ω. A collection of subsets of E, U, is called

a covering of E if U = E. It is called an open covering if each element in U is an open
S

set. If U and V are covers of E, then V is called a refinement of U or finer than U if for
each V ∈ V there exists a U ∈ U with V ⊂ U . A covering U is called locally finite if for
all x ∈ E there exists a neighbourhood V of x such that V intersects only finitely many
elements of U. If V and U are coverings of E and V ⊂ U, then V is called a subcovering
of U.

With the help of Theorem 1.8 one can show that there exists an open locally finite
covering of Ω. We will see this in the proof of Theorem 1.11.

Definition 1.10. Let U be a covering of Ω. A partition of unity on Ω subordinated to

U, is a sequence (χn )n∈N in D(Ω) with
X
0 ≤ χn (x) ≤ 1, χn (x) = 1 (x ∈ Ω),
n∈N
for each n ∈ N there exists a U ∈ U with supp χn ⊂ U.

Let us show that partitions of unity exist.

Theorem 1.11. Let U be an open covering of Ω. Then there exists a partition of unity
on Ω subordinated to U, (χn )n∈N , such that the sets {x ∈ Ω : χn (x) > 0} form a locally
finite covering of Ω. Consequently, for each ϕ ∈ D(Ω) there exists an N ∈ N such that
ϕ= N
P
n=1 χn ϕ.

Proof. In this proof, let us call an element ϕ of D “small” if supp ϕ is contained in an

element of U.
Step 1 Let A be a compact subset of an open set W ⊂ Ω. Let x ∈ A and let U ∈ U
be such that x ∈ U . By applying Lemma 1.6 to the compact set {x} and the open set
U ∩ W , we find a small ϕ ∈ D with ϕ(x) > 0 and supp ϕ ⊂ U ∩ W .

8
 It follows from the compactness of A that there exist N ∈ N and small ϕ1 , . . . , ϕN ∈ D
with ϕi ≥ 0 and supp ϕi ⊂ W for each i and A ⊂ i∈N {x ∈ Ω : ϕi (x) > 0} ⊂ W .
S

Step 2 Let (Kn )n∈N be as in Theorem 1.8 and put K0 = ∅. For every n ∈ N,
An := Kn \ Kn−1◦ ◦
is a compact set, contained in the open set Wn := Kn+1 \ Kn−1 . By
applying, for each n ∈ N, Step 1, to An and Wn , one obtains a sequence (ϕi )i∈N of small
elements of D and a sequence 1 = N1 < N2 < · · · in N such that for all n ∈ N

0 ≤ ϕi , supp ϕi ⊂ Wn if Nn ≤ i < Nn+1 ,

Nn+1
[
An ⊂ Ui , where Ui := {x ∈ Ω : ϕi (x) > 0}.
i=Nn

The collection of open sets {Ui : i ∈ N} forms a covering of Ω. We prove it to be locally

finite. To that end, let n ∈ N; it suffices to show that Kn◦ intersects only finitely many
of the sets Ui , which will be the case if Ui ⊂ Ω \ Kn for all i > Nn+1 . Take i ∈ N,
i > Nn+1 . There is an m ∈ N with Nm ≤ i < Nm+1 . Then m ≥ n + 1, whence
Ui ⊂ Wm = Km+1 ◦ \ Km−1 ⊂ Ω \ Km−1 ⊂ Ω \ Kn .
Step 3 It follows that we can define a function ϕ : Ω → (0, ∞) by ϕ(x) = n∈N ϕn (x)
P

for x ∈ Ω, and that ϕ ∈ C ∞ (Ω). By setting χn := ϕϕn for n ∈ N one obtains the desired
partition of unity on Ω.

Remark 1.12. Observe that for (χn )n∈N as in Theorem 1.11: For any sequence (λn )n∈N
in F, the formula ψ(x) = ∞
n=1 λn χn (x) for x ∈ Ω defines a C
∞ function ψ on Ω.
P

With the help of the partition of unity of Theorem 1.11 we can extend the statement
of Lemma 1.6 in such a way that we can find a testfunction that equals 1 on a compact
set:

Lemma 1.13. Let K ⊂ Ω be a compact set and U be an open subset of Rd such that
K ⊂ U . There exists a testfunction ϕ : Rd → [0, 1] such that ϕ = 1 on K and supp ϕ ⊂ U .

Exercise 1.D. Prove Lemma 1.13.

Let us recall the Leibniz’ differentiation rule.

1.14 (Leibniz’ rule). If k ∈ N0 , f, g ∈ C k (Ω) and α ∈ Nd0 with |α| ≤ k, then

!
α
X α
∂ (f g) = (∂ β f )(∂ α−β g), (1.3)
β
β∈Nd0
β≤α

Qd
where β ≤ α means βi ≤ αi for all i ∈ {1, . . . , d} and with α! = i=1 αi !,

d
! !
α α! Y αi
= = .
β (α − β)!β! i=1 βi

9
Exercise 1.E. Let α ∈ Nd0 .

(a) Show there exists a smooth function ψ ∈ C ∞ (Rd ) such that ∂ α ψ(0) = 1.
(b) Let ψ ∈ C ∞ . Let ϕ ∈ D(Ω) be such that ϕ = 1 on a neighbourhood of 0 (i.e., on
B(0, r) for some r > 0). Show that

∂ α (ψϕ)(0) = ∂ α ψ(0).

(c) Prove that for all α ∈ Nd0 there exists a testfunction ϕ : Rd → [0, 1] with ∂ α ϕ(0) = 1.

−1 34 −1 12 −1 14 −1 − 34 − 12 − 41 0 1 1 3 1 1 41 3
1 43
4 2 4 2

Figure 1: An example of a function χ as in Exercise 1.F.

Exercise 1.F. See also Figure 1.

(a) Show that there exists a smooth function χ : Rd → [0, 1] such that χ = 1 on [− 14 , 14 ]d
and χ = 0 outside (− 34 , 43 )d and such that
X
χ(x − k) = 1 (x ∈ Rd ).
k∈Zd

(b) Show that one can find such a function χ such that it is of the form χ(x) =
Qd d
i=1 η(xi ) for x = (x1 , . . . , xd ) ∈ R for a testfunction η : R → [0, 1] with η = 1 on
1 1 d 3 3 d
[− 4 , 4 ] and η = 0 outside (− 4 , 4 ) .
Pn
(c) Let n ∈ N and ψ : R → [0, 1] be the function given by ψ(x) = k=−n η(x − k),
with η as in (b). Show that kψkC m = kηkC m for all m ∈ Nd0 .
(d) Let n ∈ N and ϕ : Rd → [0, 1] be the function given by ϕ(x) = k∈[−n,n]d ∩Zd χ(x −
P

k), which χ as in (b). Show that kψkC m = kχkC m for all m ∈ Nd0 .

2 Distributions

In this section we introduce the notion of a distribution and show that all locally integ-
rable functions can be viewed as distributions. Motivated by formulae that hold for those
functions, we define operations on distributions like derivation and multiplication with
smooth functions. Then we consider the order of a distribution and Radon measures,
which also can be viewed as distributions.

10
Definition 2.1 (Distributions). A linear function u : D(Ω) → F, is called a distribution
if for all compact sets K ⊂ Ω, there exist C > 0 and k ∈ N0 such that

|u(ϕ)| ≤ CkϕkC k (ϕ ∈ D(Ω), supp ϕ ⊂ K). (2.1)

2.2. Observe that if u and v are distributions (on Ω) and λ, µ ∈ F, then w : D(Ω) → F
defined by w(ϕ) = λu(ϕ) + µv(ϕ) is a distribution.

Definition 2.3. We define D0 (Ω) to be the vector space of distributions.

Let us first consider some examples of distributions. A large class of distributions is

given by locally integrable functions:

Definition 2.4. For a set A ⊂ Ω we define its indicator function, 1A , by

(
1 x ∈ A,
1A (x) =
0 x∈/ A.

Definition 2.5. We say that a function f : Ω → F is locally integrable if f 1K is an

integrable function for all compact sets K ⊂ Ω. We write L1loc (Ω) for the space of
all locally integrable functions and L1loc (Ω) for the space of their equivalence classes
similarly as for Lp and Lp (see Section A). Similarly, Lploc (Ω) is written for those functions
f : Ω → Ω for which f 1K ∈ Lp for all compact sets K ⊂ Ω and Lploc (Ω) for the space of
equivalence classes in Lploc (Ω).

Of course all continuous functions are locally integrable, but also all elements of
Lp (Ω):

Exercise 2.A. Prove that every function in Lp (Ω) is locally integrable, where p is an
element of [1, ∞]. Conclude that Lploc (Ω) ⊂ L1loc (Ω).

2.6 (Locally integrable functions as distributions). Let f be a locally integrable function

on Ω. Define uf : D(Ω) → F by
Z Z
uf (ϕ) = fϕ = f (x)ϕ(x) dx (ϕ ∈ D(Ω)). (2.2)
Ω Ω

It is straightforward to verify that uf is a distribution.

Similarly, if f ∈ L1loc (Ω) then we define uf also by (2.2) (this is well-defined as if g
and h are locally integrable functions which are equal a.e. (almost everywhere), then
uf = ug ).

Definition 2.7. Let V and W be vector spaces. We say that V is embedded in W if

there exists a linear injection V → W , which will also be called an embedding.

For Lemma 2.9 we recall Lebesgue’s differentiation theorem:

11
Theorem 2.8 (Lebesgue’s differentiation theorem). [HvNVW16, Theorem 2.3.4] For
all f ∈ L1loc (Rd ) almost every point in Rd is a Lebesgue point, i.e., for almost all points
x,
Z
ε↓0
ε−d |f (y) − f (x)| dy −−→ 0. (2.3)
B(x,ε)

Lemma 2.9. Let f ∈ L1loc (Ω), uf = 0. Then f = 0 almost everywhere. In other words,
the function L1loc (Ω) → D0 (Ω) given by f 7→ uf is an embedding.

Proof. Let B ⊂ Ω be a ball. Then by Theorem 1.11 there exists a sequence (ϕn )n∈N
of positive functions in D(Ω) with ϕn ↑ 1B , indeed take ϕn = ni=1 χi where (χn )n∈N
P

is
R
a partition ofR unity on B. Then, by Lebesgue’s Dominated Convergence Theorem,
B f = limn→∞ f ϕn = limn→∞ uf (ϕn ) = 0. As, by Theorem 2.8, almost every point of
Ω is a density point of f , f = 0 almost everywhere on Ω.

2.10 (Convention/Notation). It is customary to identify locally integrable functions

with their equivalence classes in L1loc and regard such an equivalence class as a function.
Similarly, it is customary to identify the distribution uf with the function f for any
f ∈ L1loc or f ∈ L1loc , and to say that uf “is” a function. We mostly follow this habit, but
tend to be careful.

A distribution is determined by its “local behaviour”, in the sense that if it is equal

to zero around each point in Ω, then it is equal to zero, in the sense of the following
theorem.
Theorem 2.11. If u, v ∈ D0 (Ω) are such that for all x ∈ Ω there exists an open neigh-
bourhood U of x such that u(ϕ) = v(ϕ) for all ϕ ∈ D(Ω) with supp ϕ ⊂ U , then u = v.

Exercise 2.B. Prove Theorem 2.11.

Before we define operations on distributions, we motivate these by identities for locally

integrable functions and their corresponding distributions in 2.13.
2.12 (Notation). Let f : Ω → F and y ∈ Rd . We define

−Ω = {−x : x ∈ Ω},
Ω + y = {x + y : x ∈ Ω},

and define the functions fˇ : −Ω → F and Ty f : Ω + y → F by

fˇ(x) = f (−x), Ty f (x) = f (x − y) (x ∈ Rd ). (2.4)

We also write Rf = fˇ.

2.13. Let f ∈ L1loc (Ω). The following statements follow by applying the change of
variables formulae and integration by parts. (For the notation uf see (2.2).)

12
 (a) ufˇ is a distribution on −Ω and for ϕ ∈ D(−Ω)
Z Z
ufˇ(ϕ) = f (−x)ϕ(x) dx = f (x)ϕ(−x) dx = uf (ϕ̌) (2.5)
−Ω Ω

(b) uTy f is a distribution on Ω + y and for ϕ ∈ D(Ω + y)

Z Z
uTy f (ϕ) = f (x − y)ϕ(x) dx = f (x)ϕ(x + y) dx = uf (T−y ϕ). (2.6)
Ω+y Ω

(c) Suppose f ∈ C k (Ω) for some k ∈ N. Let α ∈ Nd0 with |α| ≤ k. Then u∂ α f is a
distribution and for ϕ ∈ D(Ω)
Z Z
u∂ α f (ϕ) = ∂ α f (x)ϕ(x) dx = (−1)|α| f (x)∂ α ϕ(x) dx = (−1)|α| uf (∂ α ϕ).
Ω Ω
(2.7)

(d) Let ψ ∈ C ∞ (Ω). Then uψf is a distribution and for ϕ ∈ D(Ω)

Z
uψf (ϕ) = ψ(x)f (x)ϕ(x) dx = uf (ψϕ). (2.8)
Ω

(e) Let l : Rd → Rd be linear and bijective. Then f ◦ l is locally integrable and uf ◦l is

a distribution on l−1 (Ω) and for ϕ ∈ D(l−1 (Ω))
1
Z Z
uf ◦l (ϕ) = f ◦ l(x)ϕ(x) dx = f (x)ϕ ◦ l−1 (x) dx
l−1 (Ω) | det l| Ω
1
= uf (ϕ ◦ l−1 ). (2.9)
| det l|

Exercise 2.C. Let u ∈ D0 (Ω). Define w : D(−Ω) → F as follows. For ϕ ∈ D(−Ω)

define w(ϕ) to be equal to the right-hand side of (2.5) with “u” instead of “uf ”, i.e.,
w(ϕ) = u(ϕ̌). Check that w is a distribution (on −Ω). Do the same for (2.6), (2.7), (2.8)
and (2.9).

The analogous operations for distributions generalise the previous relations.

Definition 2.14. Let y ∈ Rd , α ∈ Nd0 , ψ ∈ C ∞ (Ω) and l : Rd → Rd linear and bijective.
For a distribution u ∈ D0 (Ω) we define the following distributions (it is easy to check
that these are indeed distributions, see Exercise 2.C)

(a) ǔ ∈ D0 (−Ω) by

ǔ(ϕ) = u(ϕ̌) (ϕ ∈ D(−Ω)),

(b) Ty u ∈ D0 (Ω + y) by

Ty u(ϕ) = u(T−y ϕ) (ϕ ∈ D(Ω + y)),

13
 (c) ∂ α u ∈ D0 (Ω) by

∂ α u(ϕ) = (−1)|α| u(∂ α ϕ) (ϕ ∈ D(Ω)),

(d) ψu ∈ D0 (Ω) by

ψu(ϕ) = u(ψϕ) (ϕ ∈ D(Ω)),

(e) u ◦ l ∈ D0 (l(Ω)) by
1
u ◦ l(ϕ) = u(ϕ ◦ l−1 ) (ϕ ∈ D(l(Ω))).
| det l|

Observe that all the above operations are “linear in u”. Moreover, observe that

Ty ∂ α u = ∂ α Ty u (y ∈ Rd , α ∈ Nd0 ).

Example 2.15. Let d = 1 and let f : R → R be the absolute value function: f (x) = |x|
for x ∈ R. Then one can compute (see Exercise 2.D)

∂uf = ug , for g = 1(0,∞) − 1(−∞,0) ,

2
∂ uf (ϕ) = ∂ug (ϕ) = 2ϕ(0).

Of course for x 6= 0 we have g(x) = f 0 (x). On the other hand, the second derivative
∂ 2 uf , which equals ∂ug is not equal to uh for any function h for which h(x) = g 0 (x) for
all x 6= 0. Actually, ∂ 2 uf is given by a distribution corresponding to a specific Radon
measure, namely the so-called Dirac δ-measure. We will now turn to such measures and
their corresponding distributions.

Exercise 2.D. For f as in Example 2.15 check that ∂uf (ϕ) = ug (ϕ) and ∂ 2 uf (ϕ) =
2ϕ(0) for ϕ ∈ D(Ω).

Exercise 2.E. In this exercise we consider dimension one and want to consider the
function x 7→ x1 as a distribution. However, there is a problem of defining the integral
by testing it against a testfunction and integrating around zero. Therefore we define the
distribution differently.
(1) First prove that for all ϕ ∈ D(R) the limit

ϕ(x)
Z
lim dx
ε↓0 R\[−ε,ε] x
0 (x)f (x)
R
exists, and equals − Rϕ dx, where f : R → R is given by
(
log |x| x 6= 0,
f (x) =
0 x = 0.

14
For this check that f is integrable around zero and conclude that it is locally integrable.
(2) Prove that u : D(R) → R defined by

ϕ(x)
Z
u(ϕ) := lim dx (ϕ ∈ D(R)),
ε↓0 R\[−ε,ε] x

defines a distribution and u = ∂uf , where f is as in (1).

Exercise 2.F. (a) Prove that for any u ∈ D0 (R), the following are equivalent

(1) ∂u = 0,
(2) there exists a c ∈ F such that u = cu1 .
R R
(Hint: Let ψ ∈ D with ψ = 1. Prove that for all ϕ ∈ D, ϕ−( ϕ)ψ has a primitive
in D.)
(b) Prove that for all u ∈ D0 (R) there exists a v ∈ D0 (R) such that u = ∂v.

Besides locally integrable functions, we can also interpret Radon measures as distri-
butions. We introduce the notion of a Radon measure first, for some basic definitions
from measure theory we refer the reader to Appendix A.

Definition 2.16. We write Borel(Ω) for the smallest σ-algebra that contains all open
subsets of Ω. Borel(Ω) is called the Borel-σ-algebra on Ω. A Radon measure on Ω is
a measure on Borel(Ω) (i.e., a countably additive function µ : Borel(Ω) → [0, ∞] with
µ(∅) = 0) such that µ(K) < ∞ for all compact sets K ⊂ Ω.

2.17 (Radon measures as distributions). Let µ be a Radon measure on Ω. Then uµ :

D(Ω) → F defined by
Z
uµ (ϕ) = ϕ dµ (ϕ ∈ D(Ω)) (2.10)
Ω

is a distribution.
Let us write λ for the Lebesgue measure. If f ∈ L1loc (Ω) and f ≥ 0, then f λ is a Radon
measure, where f λ(A) = 1A f . Then uf defined as in 2.6 equals uf λ . Not all Radon
R

measures are of the form f λ with f ∈ L1loc (Ω), see for example the Dirac δ-measure:

Definition 2.18. For x ∈ Rd define the measure δx as follows; for a Borel measurable
set A set
(
1 x ∈ A,
δx (A) =
0 x∈/ A.

Then δx is a Radon measure, which is called the Dirac δ-measure centered at x. We call
δ0 the Dirac δ-measure. A point measure is a Dirac δ-measure centered at some point.
Observe that δx (ϕ) = ϕ(x) for any Borel measurable function ϕ.

15
Theorem 2.19. If µ is a Radon measure on Ω, then for all open sets U ⊂ Ω
µ(U ) = sup{µ(K) : K ⊂ U, K is compact} (2.11)
Z
= sup{ ϕ dµ : ϕ ∈ Cc∞ (Ω, [0, 1]), supp ϕ ⊂ U }. (2.12)
R
Consequently, if uµ (ϕ) = ϕ dµ = 0 for all ϕ ∈ D(Ω), then µ = 0. In other words,
the map from the space of Radon measures into the space of distributions, µ 7→ uµ , is
injective.

Proof. (2.11) follows from Theorem 1.8. (2.12) follows similarly as in the proof of
Lemma 2.9; by RTheorem 1.11 there exists a sequence (ϕn )n∈N in D(Ω) with ϕn ↑ 1U .
Then ϕn dµ ↑ 1U µ = µ(U ).
R

2.20. Now the space of Radon measures is not a vector space, but it is closed under
additions and multiplication with positive scalars. As the map µ 7→ uµ preserves addition
and multiplication with positive scalars, one could also say that the space of Radon
measures is “embedded” into the space of distributions. As Radon measures can attain
the value ∞ on a set, there is not a straightforward way of making sense of the difference
of two Radon measures as a function on the σ-algebra (take for example λ the Lebesgue
P
measure on R and let µ = z∈Z δz ).

We introduce the order of a distribution and then discuss how Radon measures cor-
respond to distributions with order 0.
Definition 2.21. If u is a distribution and there exist a C > 0 and k ∈ N0 such that
|u(ϕ)| ≤ CkϕkC k (ϕ ∈ D(Ω)), (2.13)
then u is said to be of order at most k. In other words, for this C and k (2.1) holds for
all compact sets K. If u is of order at most 0, we also say that u is of order 0. If k ∈ N
and u is of order at most k but not of order at most k − 1, then u is said to be of order
k.
2.22. Every distribution of the form uf for an integrable function f and every distribution
of the form uµ for a Radon measure µ with µ(Ω) < ∞ is of order 0.
On the other hand, the distribution u1 corresponding to the function 1 which is
equal to 1 everywhere is not of any finite order: Let χ be as in Exercise 1.F and define
ϕn (x) = k∈[−n,n]∩Zd χ(x − k). Then kϕn kC m = kϕ1 kC m for all n ≥ 2 and m ∈ Nd0 . On
P

the other hand u1 (ϕn ) ≥ 1[−(n−1),n−1] ≥ (2(n − 1))d as ϕn ≥ 1[−(n−1),n−1] for all n ∈ N.
R

Exercise 2.G.

(a) Let k ∈ N0 , d = 1. Show that ∂ k δ is a distribution of order at most k.

(b) Show that ∂ k δ is not of order at most k−1, i.e., show that it is of order k. (Hint: Test
it against ϕε (x) = xk ϕ( xε ) for ε > 0 with ϕ ∈ D(Ω) with ϕ = 1 on a neighbourhood
of 0.)

16
 (c) What is the order of ∂ α δ for α ∈ Nd0 ?
(d) Construct a distribution which is not represented by a locally integrable function
and not of any finite order.

Actually, all distributions of order 0 correspond to linear combinations of finite Radon

measures (a Radon measure µ is called finite if µ(Ω) is finite), see Theorem 2.28. Such
linear combinations are called signed Radon measures if F = R and complex Radon
measures if F = C, see Definition 2.23 and Theorem 2.24. For the proof of Theorem 2.28,
besides the Riesz representation theorem (Theorem 2.30) we will use convolutions. As
convolutions are treated in Section 7, we postpone the proof of Theorem 2.28 to Section 9
(see below Corollary 9.8).

Definition 2.23. A Radon measure µ is called finite if µ(Ω) < ∞. We write M(Ω, F) or
M(Ω) for the set of countably additive functions µ : Borel(Ω) → F. Elements of M(Ω, R)
are called signed measures and elements of M(Ω, C) are called complex measures. We say
that a signed or complex Radon measure µ is positive, if µ(A) ≥ 0 for all A ∈ Borel(Ω).
We define k · kM : M(Ω, F) → [0, ∞] by
( n )
X
kµkM = sup |µ(Ai )| : A1 , . . . , An is a partition of Ω in Borel(Ω) ,
i=1

where a partition of Ω in Borel(Ω) is a finite number of pairwise disjoint sets A1 , . . . , An

in Borel(Ω) such that ni=1 Ai = Ω.
S

We state the Hahn–Jordan Decomposition Theorem without proof; a proof can be

found in [Con90, Theorem C.1] .

Theorem 2.24 (Hahn–Jordan Decomposition). For each µ ∈ M(Ω, R) there exist ex-
actly one pair of disjunct positive finite Radon measures µ+ , µ− such that there exist meas-
urable sets E+ , E− ∈ Borel(Ω) such that E+ ∩ E− = ∅ and E+ ∪ E− = X, µ+ (E− ) = 0,
µ− (E+ ) = 0 and

µ = µ+ − µ− .

Moreover,

kµkM = µ+ (Ω) + µ− (Ω).

For a µ in M(Ω, C) we can define the functions <µ, =µ : Borel(Ω) → R defined by

(<µ)(A) = <(µ(A)) (=µ)(A) = =(µ(A)) (A ∈ Borel(Ω)),

are signed measures, or in other words, elements of M(Ω, R). By applying the previous
theorem to <µ and =µ we obtain the following.

17
Corollary 2.25. For each µ ∈ M(Ω, C) there exist four positive finite Radon measures
µ1 , µ2 , µ3 , µ4 such that

µ = µ1 − µ2 + i[µ3 − µ4 ],

and

kµkM = µ1 (Ω) + µ2 (Ω) + µ3 (Ω) + µ4 (Ω).

The Hahn–Jordan decomposition in particular implies:

Lemma 2.26. k · kM is a norm on M(Ω). L1 (Ω, F) is embedded in M(Ω, F) by the
function L1 (Ω, F) → M(Ω, F), f 7→ f λ, where λ is the Lebesgue measure on Ω. Moreover,

kf kL1 = kf λkM (f ∈ L1 (Ω, F)). (2.14)

Proof. That k · kM is a norm follows by the Hahn–Jordan decompositions. In case F = R

(2.14) follows as kf kL1 = f + f , where f + (x) = f (x) ∨ 0 and f − (x) = −f (x) ∨ 0
R + R −

for x ∈ Ω. The case F = C is similar.

2.27. For a signed or complex Radon measure µ ∈ M(Ω, F) we define uµ : D(Ω) → F by

(2.10). As a consequence of Theorem 2.24 and Theorem 2.19 M(Ω, F) is embedded in
D0 (Ω) by the embedding µ 7→ uµ .

Theorem 2.28. A distribution u is of order 0 if and only if u = uµ for a µ ∈ M(Ω, F).

The following theorem, the Riesz representation theorem, will be used for the proof of
Theorem 2.28 in Section 9 (it is given below Corollary 9.8). For a proof of Theorem 2.30
see [Con90, Theorem C.18] .
Definition 2.29. We write C0 (Ω, F) or C0 (Ω) for the continuous functions Ω → F such
that for all ε > 0 there exists a compact set K ⊂ Ω such that |f | < ε on Ω \ K. C0 (X, F)
is equipped with the norm k · kC 0 .
Theorem 2.30 (Riesz representation theorem). For µ ∈ M(Ω, F) define Ψµ : C0 (Ω, F) →
F by
Z
Ψµ (f ) = f dµ (f ∈ C0 (Ω, F)).

Then ψµ ∈ C0 (Ω, F)0 and the map M(Ω, F) → C0 (Ω, F)0 , µ 7→ Ψµ is an isometric
isomorphism.
Corollary 2.31. M(Ω, F) is a Banach space under the norm k · kM .

Proof. This follows by Theorem 2.30 as the dual of a normed space, in this case C0 (Ω, F)0 ,
is a Banach space. See for example [Con90, Proposition III.5.4] or [Rud91, Theorem
4.1].

18
3 Topological vector spaces

In this section we introduce some topological notions like topological vector spaces, loc-
ally convex spaces and weak topologies on dual pairs. We discuss a few properties like
metrizability and separation properties. These notions will be used in Section 4 to define
the topologies on D(Ω) and D0 (Ω). We recall some definitions of topology in 3.1, but will
not require the reader to be familiar with nets; see 3.2 for some comments.

3.1 (Topological vocabulary). Let X be a set. A topology τ on X is a set of subsets of

X, called open sets, such that ∅, X ∈ τ , if A, B ∈ τ then A ∩ B ∈ τ and if U ⊂ τ , then
U ∈ τ . Then the pair (X, τ ) is called a topological space; but it is also common to say
S

that X is a topological space itself.

If X is a topological space with topology τ , then we use the following vocabulary.
A set F ⊂ X is called closed if its complement X \ F is open. The closure A of a set
A ⊂ X is the smallest closed set that contains A; it is the intersection of all closed sets
that contain A. The interior A◦ of a set A ⊂ X is the largest open set contained in A;
it is the union of all open subsets of A. A set K ⊂ X is compact if each open covering of
K has a finite subcovering. A neighbourhood of a point x ∈ X is a set that contains an
open set that contains x. For a set A ⊂ X we call an set that contains an open set that
contains A also a neighbourhood of A. If Y ⊂ X and σ = {U ∩ Y : U ∈ τ }, then σ is
called the relative topology on Y . If σ is a collection of subsets of X, then τ is said to be
generated by σ, or we also say, σ generates the topology τ if τ is the smallest topology
that contains σ. One can show that τ consists of those sets which are unions of finite
intersections of elements in σ. If x ∈ X and U is a collection of subsets of X with x ∈ U
for all U ∈ U, then U is called a local base for x if for each neighbourhood V of x there
exists a U ∈ U with U ⊂ V . X is called connected if it is not the union of two disjoint
non-empty open subsets of X. X is called metrizable if there exists a metric d on X such
that the topology is generated by the balls {x ∈ X : d(x, y) < r} with y ∈ X and r > 0;
in that case one says that d is compatible with the topology on X. A metric d on a vector
space X is called translation invariant if d(x + z, y + z) = d(x, y) for all x, y, z ∈ X.
Let X and Y be topological spaces. A function f : X → Y is called continuous if
f −1 (U ) is open in X for each open set in Y . The product topology on X × Y is the
topology generated by the collection of sets U × V , with U being an open set in X and Y
an open set in Y . Equivalently, the product topology on X × Y is the smallest topology
such that the projections X × Y → X and X × Y → Y are continuous. If not mentioned
otherwise, for two topological spaces X and Y the space X × Y is equipped with the
product topology.

3.2 (Sequences and nets). For metric spaces X and Y we have that f : X → Y is
continuous if and only if xn → x implies f (xn ) → f (x) for each sequence (xn )n∈N and x
in X. Moreover, a set F in X is closed if and only if each sequence in F that converges
in X has its limit in F . This does not hold in general for any topological space. But
one can replace the usage of sequences by nets. In order to keep the text readable for

19
those who are not so familiar with that notion, we avoid it. Moreover, for the theory of
distributions, as we will see, it is often enough to consider convergence of sequences.
Definition 3.3. A vector space X equipped with a topology τ is called a topological vector
space if both the operations addition and scalar multiplication, that is the functions

X × X → X, (x, y) 7→ x + y,
F × X → X, (λ, X) 7→ λx,

are continuous.
It is easy to see that each normed vector space is a topological vector space.
Observe that if U is a local base for 0, then the topology of X is generated by the
sets x + U with x ∈ X and U ∈ U.

We recall the definition of a seminorm.

Definition 3.4. Let X be a vector space. A seminorm on X is a function p : X → [0, ∞)
such that

p(x + y) ≤ p(x) + p(y) (x, y ∈ X),

p(λx) = |λ|p(x) (λ ∈ F, x ∈ X).

A seminorm p is a norm if p(x) = 0 implies x = 0.

Definition 3.5. Let P be a collection of seminorms on a vector space X. Let τ be the
topology generated by the sets

{x ∈ X : p(x − y) < r} (y ∈ X, r > 0).

Then τ is called the topology generated by P.

A set U is open in X, i.e., U ∈ τ if and only if for each y ∈ U there exist n ∈ N,
p1 , . . . , pn ∈ P, r, . . . , rn > 0 such that
n
\
{x ∈ X : pi (x − y) < ri } ⊂ U.
i=1

Moreover, X equipped with the topology generated by P is a topological vector space

(the proof of this is rather straightforward and left to the reader).
Definition 3.6. Let P be a collection of seminorms. A vector space X equipped with
the topology generated by P is called a locally convex space if
\
{x ∈ X : p(x) = 0} = {0}. (3.1)
p∈P

If P is such that (3.1) holds, then P is called a separating family of seminorms and the
topology generated by P is called a locally convex topology.

20
 A locally convex space is Hausdorff, that is, if x, y ∈ X and x 6= y, then there exist
open sets U, V such that U ∩ V = ∅, x ∈ U and y ∈ V : By (3.1) there exists a p ∈ P such
that p(x − y) 6= 0. Let ε > 0 be such that p(x − y) > 2ε. Take U = {z ∈ X : p(z − x) < ε}
and V = {z ∈ X : p(z − y) < ε}.
Of course, every normed space is a locally convex space.
3.7 (Why is it called “locally convex”?). From this definition it might not be clear why
this is called “locally convex”. This is due to the fact that the topology is such that for
all x and open neighbourhood U of x there exists a convex open neighbourhood V with
x ∈ V ⊂ U , see for example [Rud91, Theorem 1.37].

As we will see in 3.11 the most obvious choice of equipping the testfunctions with
the k · kC k norms with k ∈ N, leads to a metrizable locally convex topology with is not
complete. That it is metrizable follows from the following theorem.
Theorem 3.8. A locally convex vector space X is metrizable if and only if it is generated
by countably many seminorms. Moreover, if the topology of X is generated by a countable
number of seminorms p1 , p2 , . . . , then the topology is compatible with the translation
invariant metric d on X defined by
2−n ∧ pn (x − y)
X
d(x, y) = (x, y ∈ X). (3.2)
n∈N

Proof. (Optional) If X is generated by countably many seminorms p1 , p2 , . . . , then it is

easy to check that the metric defined in (3.2) is compatible with the topology.
Suppose X is metrizable and d be a metric on X that is compatible with the topology.
Then it is easy to check that d(x, y) := d(0, y − x) for x, y ∈ X defines a translation
invariant metric on X. Moreover, as the neighbourhoods of 0 are the same for d and d,
also d is compatible with the topology on X (as a translation by x of a local base for 0
is a local base for x).
X has a countable local base for each y ∈ X (such X are called “first countable”),
namely Uy = {Vn : n ∈ N}, where
1
Vn := {x ∈ X : d(x, y) < } (n ∈ N).
n
1
As d is translation invariant we have Uy = x + U0 , indeed {x ∈ X : d(x, y) < n} =
y + {x ∈ X : d(0, y − x) < n1 } for all n ∈ N.
Let P be a collection of seminorms that generates the topology on X. Define
1
Up,m := {x ∈ X : p(x) < } (p ∈ P, m ∈ N).
m
Then {Up,m : p ∈ P, m ∈ N} forms a local base for 0. Therefore, for each n ∈ N, we can
choose a pn ∈ P and mn ∈ N such that
Upn ,mn ⊂ Vn .

21
This implies that {Upn ,mn : n ∈ N} is a local base for 0. Therefore the topology is
generated by the countable set of seminorms {pn : n ∈ N}.

Let us introduce the notion of a Cauchy sequence and completeness for locally convex
spaces. In 3.10 we comment on how this notion agrees with the notions of Cauchy
sequences and completeness for metric spaces.
Definition 3.9. Let X be a locally convex topological space and P be a collection of
seminorms that generate the topology of X. A sequence (xn )n∈N in X is called P-Cauchy
or just Cauchy, if for each p ∈ P it is Cauchy with respect to p:

for all ε > 0 there exists an N ∈ N such that p(xn − xm ) < ε for all n, m ≥ N.

If each P-Cauchy sequence in X converges, X is called P-complete or just complete.

3.10 (P-Cauchy and Cauchy with respect to a metric). Let X and P be as in Defin-
ition 3.9. If X is a normed space, which means that its topology is generated by one
norm, then a sequence is P-Cauchy if and only if it is Cauchy with respect to the norm
(and thus distance). Consequently, the space is P-complete if and only if it is complete
with respect to the norm.
If X is metrizable, which means that one can take P to be countable, then a sequence
is P-Cauchy if and only if it is d-Cauchy with d the translation invariant metric as in
(3.2).
It is important to have such a translation invariant metric, as the following example
illustrates: Take X = R and let d and d be metrics on R given by d(x, y) = |x − y|
and d(x, y) = |e−x − e−y | for x, y ∈ R. Then both metrics generate the same topology,
however the sequence 1, 2, 3, . . . is d-Cauchy but not d-Cauchy, and R is complete under
d but not under d.
3.11 (A metrizable topology on D(Ω) which is not complete).
On D(Ω) the function k · kC k is a (semi)norm for each k ∈ N0 and therefore these
seminorms generate a locally convex topology on D(Ω). This topology is metrizable, see
Theorem 3.8, however the topology is not complete: Let χ be a testfunction which equals
1 at 0 with support in (− 21 , 12 ). Let (εn )n∈N be in (0, ∞) and εn ↓ 0 and let χn = Tn χ.
Take ϕn = ni=1 εi χi . Then (ϕn )n∈N is a Cauchy sequence with respect to the metric d
P

of Theorem 3.8 with pk = k · kC k but does not have a limit in D(Rd ). One can adapt this
argument to show that also D(Ω) is not complete for any open set Ω ⊂ Rd (for example
in the spirit of Exercise 1.C or with the use of a partition of unity).

Exercise 3.A. Show that (for a general open set Ω ⊂ Rd ) the space D(Ω) is not
complete in the metrizable topology generated by the seminorms k · kC k for k ∈ N.

The topology mentioned in 3.11 is not the topology which D(Ω) is equipped with. In
order to introduce that topology, we make some notions and introduce some definitions.
First, we observe that distributions define seminorms on the space of testfunctions:

22
3.12. Each distribution u defines a seminorm on D(Ω) by

ϕ 7→ |u(ϕ)|. (3.3)

By Lemma 2.9 applied to f ∈ D(Ω) it follows that if uf (ϕ) = uϕ (f ) = 0 for all ϕ ∈ D(Ω),
then f = 0. Therefore the collection of seminorms generated by the distributions, that
is the seminorms (3.3) for u ∈ D0 (Ω) form a separating family.

In the notation D0 (Ω) we have used the symbol “ 0 ”, which is commonly used to denote
the topological dual of a space, as in Definition 3.13. The definition of the topologies
of D(Ω) and D0 (Ω) are given in Definition 4.1 in terms of the topologies defined in
Definition 3.16. The topology on D(Ω) is exactly the one generated by the seminorms
defined by the distributions as in 3.12. In the following definition we introduce the
notion of a topological dual of a topological vector space. It turns out that D0 (Ω) is
indeed the topological dual of D(Ω) with the topology that we consider. This relies on
Theorem 3.19, as we mention in Definition 4.1. First we introduce the notion of a dual
and use the Hahn–Banach theorem to show Lemma 3.15, which shows that for a locally
convex space X with its dual X 0 forms a pair as defined in Definition 3.16. As those
theorems are not essential for the further theory of distributions, the reader may skip
those theorems which are indicated by (  ).

Definition 3.13. Let X be a topological vector space over F. The space of linear
continuous maps X → F is called the dual or topological dual of X. We write X 0 for the
dual of X. Each element of x ∈ X determines a seminorm on X 0 by f 7→ |f (x)|. We
equip X 0 with the locally convex topology generated by these seminorms, which is also
called the weak∗ topology.

The elements of X 0 for a locally convex space X separate the points in X, see
Lemma 3.15. To prove this lemma we use the Hahn–Banach theorem, which we state
without a proof (for a proof see for example [Con90, Corollary III.6.4] or [Rud91, Theorem
3.3]).

Theorem 3.14 (Hahn–Banach). () Let X be a vector space, M be a linear subspace of

X and p be a seminorm on X. If f : M → F is a linear function such that |f (x)| ≤ p(x)
for all x ∈ M , then there is a linear extension of f to X, F : X → F (F |M = f ) with
|F (x)| ≤ p(x) for all x ∈ X.

Lemma 3.15. (  ) Let X be a locally convex vector space. Let M be a linear subspace
of X. Let f : M → F be linear and continuous (with respect to the relative topology),
then f has a continuous linear extension to X.
Consequently, for each x, y ∈ X with x 6= y and x 6= 0 there exists an F ∈ X 0 such
that F (x) = 1 and F (y) = 0.

Proof. Suppose that P is a collection of seminorms that generates the topology of X. As

23
f is continuous on M there exist seminorms p1 , . . . , pn ∈ P and r1 , . . . , rn > 0 such that
n
{x ∈ X : pi (x) < ri } ⊂ f −1 (B(0, 1)).
\

i=1

Let p = maxni=1 prii . It is easy to check that p is a continuous seminorm on X. As p(x) < 1
implies f (x) < 1, by linearity we have |f (x)| ≤ p(x). By the Hahn–Banach theorem there
exists an exists a linear extension F : X → F of f , i.e., F (x) = f (x) for x ∈ M , and
|F (x)| ≤ p(x). The latter implies that F is continuous.

Definition 3.16. Let X and Y be vector spaces over F and h·, ·i : X × Y → F be a

bilinear form that satisfies the separation axioms:

hx, yi = 0 for all y ∈ Y implies x = 0,

hx, yi = 0 for all x ∈ X implies y = 0.

Such a pair (X, Y ) is called a dual pair.

The weak topology σ(X, Y ) on X is the coarsest topology on X such that all maps
h·, yi with y ∈ Y are continuous. In other words, this topology is generated by the
seminorms x 7→ |hx, yi| for y ∈ Y and therefore a locally convex topology. Similarly one
defines the weak topology σ(Y, X) on Y ; it is the coarsest topology on Y such that the
maps hx, ·i with x ∈ X are continuous.

Example 3.17.
(a) Let V be a vector space and V # be its algebraic dual. The pair (V, V # ) forms a
dual pair under the bilinear form V × V # → F given by (x, f ) 7→ f (x).
(b) Let X be a locally convex vector space. By Lemma 3.15 the pair (X, X 0 ) forms a
dual pair under the bilinear form X × X 0 → F given by (x, f ) 7→ f (x).
(c) As the distributions separate the testfunctions as we have observed in 3.12, the pair
(D(Ω), D0 (Ω)) forms a dual pair under the bilinear form D(Ω) × D0 (Ω) → F given
by (ϕ, u) 7→ u(ϕ).

We will show that for X and Y as in Definition 3.16 Y is the dual of X equipped
with σ(X, Y ) in Theorem 3.19, so that the name “dual pair” now actually makes sense.
The next lemma is a preparation for the proof.

Lemma 3.18. (  ) Let X and Y be a dual pair. Let n ∈ N and y1 , . . . , yn ∈ Y be

linearly independent. Then there exist linearly independent x1 , . . . , xn ∈ X such that
hxi , yj i = δi,j for all i, j ∈ {1, . . . , n}, where δi,j is the Kronecker symbol, which equals 1
if i = j and 0 otherwise.

Proof. For n = 1 this follows immediately by the separation axioms. We continue by

induction. Let n ∈ N and y1 , . . . , yn+1 ∈ Y be linearly independent. Let x̃1 , . . . , x̃n be
such that hx̃j , yi i = δi,j for i, j ∈ {1, . . . , n}. Let Mn be the linear span of x̃1 , . . . , x̃n and

24
let Fn = {x ∈ X : hx, yi i = 0, i ∈ {1, . . . , n}}. Then X is the direct sum of Fn and Mn ,
i.e., X = Fn + Mn . The map h·, yn+1 i cannot vanish on Fn , because if it would, then
yn+1 would be a linear combination of y1 , . . . , yn . Therefore there exists an xn+1 ∈ Fn
with hxn+1 , yn+1 i = 1. By defining xi = x̃i − hx̃i , yn+1 ixn+1 we have hxi , yj i = δi,j for all
i, j ∈ {1, . . . , n + 1}, because of this the xi have to be independent.

Theorem 3.19. (  ) Let X and Y be a dual pair. The dual of the topological vector
space (Y, σ(Y, X)) is X. This means that if f : Y → F is continuous and linear, then
there exists a unique x ∈ X such that f (y) = hx, yi for y ∈ Y .

Proof. As f −1 (B(0, 1)) is an open set, there exist x1 , . . . , xn ∈ X such that

n
{y ∈ Y : |hxi , yi| < 1} ⊂ f −1 (B(0, 1)).
\

i=1

So maxni=1 |hxi , yi| < 1 implies f (y) < 1 for all y ∈ Y . As f is linear, we obtain
n
|f (y)| ≤ max |hxi , yi| (y ∈ Y ). (3.4)
i=1

We may assume that x1 , . . . , xn are linearly independent. Let fi ∈ Y # be defined by

fi (y) = hxi , yi for y ∈ Y . Then f1 , . . . , fn are linearly independent. By Lemma 3.18
f cannot be linearly independent from f1 , . . . , fn because of (3.4). Hence f is a linear
combination of f1 , . . . , fn , from which it follows that there exists an x (which is a linear
combination of x1 , . . . , xn ) such that f (y) = hx, yi for all y ∈ Y .

4 Topologies on the spaces of testfunctions and distribu-

tions

In this section we equip D(Ω) and D0 (Ω) with topologies.

Definition 4.1 (Pairing and topologies on D(Ω) and D0 (Ω)).

We define h·, ·i : D0 (Ω) × D(Ω) → F by

hu, ϕi = u(ϕ) ((u, ϕ) ∈ D0 (Ω) × D(Ω)). (4.1)

We equip the space of testfunctions D(Ω) with the weak topology σ(D(Ω), D0 (Ω)) and
the space of distributions D0 (Ω) with the weak∗ topology, that is, with the weak topology
σ(D(Ω)0 , D(Ω)).
By Theorem 3.19 it follows that D0 (Ω) is the dual space of D(Ω) (equipped with the
σ(D(Ω), D0 (Ω)) topology).

4.2 (Convention/Notation). As we mentioned already in 2.10, it is customary to identify

locally integrable functions with their corresponding distribution. That is, one may write

25
“f ” instead of “uf ” for a locally integrable function (or equivalence class) f . However,
we still prefer not to write “f (ϕ)” instead of “uf (ϕ)” so we will write “hf, ϕi” instead.
The notation “h·, ·i” is also commonly used for inner products and this might cause
confusion. Indeed, say weRtake f, g ∈ D and as mentioned above, view f as the distri-
bution uf . Then
R
hf, gi is f g which is not the same (at least not for general C-valued
functions) as f g, the latter is the inner product of f and g, for which we write hf, giL2 .
Remark 4.3 (Another way to introduce the topology on D(Ω)). In the literature there
are basically two approaches to the topologies on testfunctions and the distributions.
The one presented here, where first the distributions are defined and then the topology
as a weak topology. Or one where one first defines a topology on the testfunctions via
an inductive limit approach, also called .... Then the space of distributions is defined to
be the dual of this space, i.e., the space of linear functionals on the testfunctions that
are continuous with respect to the topology (this approach is followed by the books on
functional analysis [Con90] and [Rud91]). The topologies differ slightly (we comment on
this in ...), but the convergence of sequences is the same, and, as we will see, the space
of distributions is determined by that (and therefore is the same for both approaches).
4.4. Let y ∈ Rd , α ∈ Nd0 , ψ ∈ C ∞ (Ω) and l : Rd → Rd be a linear bijection. Observe
that the operations ˇ , Ty , ∂ α , multiplication by ψ and composition with l, i.e.,
D(Ω) → D(−Ω), ϕ 7→ ϕ̌, D0 (Ω) → D0 (−Ω), u 7→ ǔ,
0 0
D(Ω) → D(Ω + y), ϕ 7→ Ty ϕ, D (Ω) → D (Ω + y), u 7→ Ty u,
α 0 0
D(Ω) → D(Ω), ϕ 7→ ∂ ϕ, D (Ω) → D (Ω), u 7→ ∂ α u,
D(Ω) → D(Ω), ϕ 7→ ψϕ, D0 (Ω) → D0 (Ω), u 7→ ψu,
0 0
D(Ω) → D(l(Ω)), ϕ 7→ ϕ ◦ l, D (Ω) → D (l(Ω)), u 7→ u ◦ l,
are continuous. As an example we consider multiplication with ψ. Let us write Πψ :
D(Ω) → D(Ω) for ϕ 7→ ψϕ:
As the operations is linear, it is sufficient to show continuity at zero. Instead of showing
that the preimage of any open neighbourhood of 0 under the multiplication is an open set,
it is sufficient to consider the preimage of neighbourhoods generated by the distributions.
The following sets form a local base for 0 in the topology on D(Ω) (Exercise 4.A):
n
(n ∈ N, u1 , . . . , un ∈ D0 (Ω)).
\
{ϕ ∈ D(Ω) : |hui , ϕi| < 1} (4.2)
i=1
Tn
Fix u1 , . . . , un in D0 (Ω) and let U = i=1 {ϕ ∈ D(Ω) : |hui , ϕi| < 1}. Then
n
Π−1
\
ψ (U ) = {ϕ ∈ D(Ω) : ψϕ ∈ U } = {ϕ ∈ D(Ω) : |hui , ψϕi| < 1}
i=1
\n
= {ϕ ∈ D(Ω) : |hψui , ϕi| < 1},
i=1

which is again a set of the form (4.2), hence open. With similar arguments the other
operations can be shown to be continuous; this is left to the reader.

26
Exercise 4.A. Show that indeed the open sets
n
\
{ϕ ∈ D : |hui , ϕi| < 1}
i=1

for u1 , . . . , un ∈ D0 (Ω) form a local base at 0 in D(Ω).

The characterisation of convergence of sequences is given in Theorem 4.11. Before

some auxiliary lemmas are given. First we show that we can embed D(U ) in D(Ω) and
D0 (Ω) in D0 (U ) whenever U is an open subset of Ω.
Definition 4.5. Let X and Y be topological vector spaces. We say that X is continuously
embedded in Y , and write X ,→ Y , if there exists a continuous embedding X → Y .
If Z is another topological vector space, then we will write “X ,→ Y ,→ Z” instead
of “X ,→ Y and Y ,→ Z”, etc.
4.6 (Restriction of a distribution to a smaller set). Suppose U is an open subset of Ω.
As every compact set in U is compact in Ω, there exists a linear injection
ι : D(U ) → D(Ω),
with ι(ϕ)(x) = ϕ(x) for x ∈ U and ι(ϕ)(x) = 0 for x ∈ Ω \ U , and ϕ ∈ D(U ). On the
other hand, for u ∈ D0 (Ω) we define ρ(u) : D(U ) → F by
hρ(u), ϕi = hu, ι(ϕ)i (ϕ ∈ D(U )).
Then ρ(u) is a distribution, and so ρ forms a map D0 (Ω) → D0 (U ). It follows that both
ρ and ι are linear and continuous. So D(U ) can be continuously embedded in D(Ω) and
D0 (Ω) can be continuously embedded in D0 (U ), i.e.,
D(U ) ,→ D(Ω), D0 (Ω) ,→ D0 (U ).

For this reason, we will view ρ(u) as the restriction of u to D(U ). Therefore, when
v ∈ D0 (U ) we will say “u = v on U ” instead of “ρ(u) = v”. Moreover, if u ∈ D0 (U ) and
ϕ ∈ D(Ω), then we will write “u(ϕ)” instead of “u(ϕ|U )”.
Observe, moreover, that ι is also continuous with respect to the seminorms k · kC k
with k ∈ N0 , and
kι(ϕ)kC k (Ω) = kϕkC k (U ) (ϕ ∈ D(U )). (4.3)

The following lemma is an application of the mean-value theorem to higher dimen-

sions.
Lemma 4.7. Let V, U ⊂ Rd , V be convex, U be open and V ⊂ U . If ψ ∈ C 1 (U ) and
d
M = max sup |∂i ψ(x)| < ∞,
i=1 x∈V

then ψ is Lipschitz continuous on V with Lipschitz constant M .

27
 d
Proof. Let x, y ∈ U . For t ∈ [0, 1] we have that dt ψ(tx + (1 − t)y) equals the directional
derivative of ψ in the direction x − y, and thus
d
ψ(tx + (1 − t)y) = ∇ψ(tx + (1 − t)y) · (x − y).
dt
Therefore by the mean-value theorem
Z 1
d d
|ψ(x) − ψ(y)| = ψ(tx + (1 − t)y) dt ≤ max k∂i ψkL∞ |x − y|.
0 dt i=1

Observe that if Ω is a convex open subset of Rd and ψ ∈ C 1 (Ω), kψkC 1 (Ω) < ∞, then
ψ is Lipschitz continuous. However, if Ω is not convex, this need not be true.
Exercise 4.B. Construct a function ψ which is C 1 on (0, 1) ∪ (1, 2) and kψ 0 kL∞ < ∞,
but which is not Lipschitz continuous.

We will use the previous lemma in combination with the following lemma.
Lemma 4.8. Let X ⊂ Rd be compact. Suppose (fn )n∈N is a sequence of uniformly
Lipschitz continuous functions, i.e., there exists an M > 0 such that

|fn (x) − fn (y)| ≤ M |x − y| (n ∈ N, x, y ∈ X).

If (fn )n∈N converges pointwise to zero, i.e., fn (x) → 0 for all x ∈ X, then (fn )n∈N
converges uniformly to zero, i.e., kfn kL∞ → 0.

Proof. Let ε > 0. As X is compact, there exist x1 , . . . , xk ∈ X such that X ⊂

Sk ε ε
i=1 B(xi , 2M ). Let N ∈ N be such that |fn (xi )| < 2 for all i ∈ {1, . . . , k} and all
n ≥ N . By the Lipschitz continuity we have that for all y ∈ X that there exists a i such
ε
that y ∈ B(xi , 2M ) and thus for all n ≥ N
ε ε
|fn (y)| ≤ |fn (y) − fn (xi )| + |fn (xi )| < M + = ε.
2M 2

Definition 4.9. Let K ⊂ Ω be compact and m ∈ N0 . We define the seminorm k · kC m ,K

on C m (Ω) by

kf kC m ,K = kf |K kC m (K) = max sup |∂ β f (x)| (f ∈ C ∞ (Ω)). (4.4)

β∈Nd0 :|β|≤m x∈K

Lemma 4.10. Let (ϕn )n∈N be a sequence in D(Ω). If for all α ∈ Nd0 , (∂ α ϕn )n∈N is
uniformly bounded and if ∂ α ϕn converges pointwise to zero, that is, if

sup k∂ α ϕn kL∞ < ∞ (α ∈ Nd0 ),

n∈N
α n→∞
∂ ϕn (x) −−−→ 0 (x ∈ Ω),

28
then for all m ∈ N0 and compact sets K ⊂ Ω
n→∞
kϕn kC m ,K −−−→ 0.

Proof. Let ι be the continuous embedding D(Ω) → D(Rd ), see 4.6. Then ∂ α ι(ϕk ) =
ι(∂ α ϕk ) and so (∂ α ι(ϕk ))k∈N is uniformly bounded for any α ∈ Nd0 . By an application
of Lemma 4.7 it follows that (∂ α ι(ϕk ))k∈N and thus (∂ α ϕk )k∈N are uniformly Lipschitz.
Hence by Lemma 4.8 it follows that k∂ α ϕk kL∞ → 0 for all α ∈ Nd0 .

Theorem 4.11. A sequence (ϕn )n∈N converges to a ϕ in D(Ω) if and only if (a) and
(b):

(a) There exists a compact set K ⊂ Ω such that the supports of ϕn and ϕ lies within
K for all n ∈ N.
(b) kϕn − ϕkC m → 0 for all m ∈ N.
Exercise 4.C. Prove the “if” part of Theorem 4.11: that (a) and (b) imply ϕn → ϕ
in D(Ω).

Proof of the “only if” part of Theorem 4.11. Suppose that ϕn → 0 in D(Ω). We deduce
(a) and (b), arguing by contradiction.
Suppose (a) is not satisfied. Then no compact subset of Ω contains the supports
of all functions ϕ1 , ϕ2 , . . . . Let (Kn )n∈N be as in Theorem 1.8. Inductively, choose
n1 < n2 < · · · in N such that for all i ∈ N
i−1
[
supp ϕnj ⊂ Ki , supp ϕni 6⊂ Ki .
j=1

For i ∈ N choose xi in Ω \ Ki with ϕni (xi ) 6= 0. If i, j ∈ N and j < i, then ϕnj (xi ) = 0
since xi ∈
/ Kj and supp ϕni ⊂ Kj .
Now let us define a measure with support being equal to the set of xi ’s as follows.
We let µ = i∈N ai δxi , where the ai ’s are chosen such that ki=1 ai ϕnk (xi ) = 1; this can
P P

always be done inductively. By assumption on the sequence (xk )k∈N , this measure is a
Radon measure, as any compact R
set K ⊂ Ω contains only finitely many xk ’s. Therefore
it defines a distribution. But ϕn dµ = 1 for all n ∈ N, which contradicts the hypothesis
that ϕn → 0 in D(Ω).
In order to show (b) we use Lemma 4.10. As ϕn → 0 in D(Ω) and ∂ α δx ∈ D0 (Ω) for
n→∞
all x ∈ Ω and α ∈ Nd0 , we have ∂ α ϕn (x) −−−→ 0 for all x ∈ Ω and α ∈ Nd0 . Hence by
Lemma 4.10 it is sufficient to show

for all α ∈ Nd0 (∂ α ϕn )n∈N is uniformly bounded.

As ∂ α is a continuous function D(Ω) → D(Ω) for all α ∈ Nd0 , it is sufficient to show

ψn → 0 in D(Ω) =⇒ (ψn )n∈N is uniformly bounded. (4.5)

29
 To prove the statement (4.5) let us assume that ψn → 0 in D(Ω) and that ψn is not
uniformly bounded. Therefore, by possibly passing to a subsequence, we may assume
that kψn kL∞ > 3n for all n ∈ N. Then we can find a sequence (xn )n∈N in Ω such that
|ψn (xn )| = kψn kL∞ .
As ψn converges pointwise to zero, we may and do assume –by possibly passing to a
subsequence– that n−1
i=1 |ψn (xi )| < 1 for all n ∈ N. As we did before let us construct a
P

Radon measure. We let

ψi (xi )
ai = 3−i
X
µ= ai δxi , (i ∈ N).
i∈N
|ψi (xi )|

Then
Z n−1
X ∞
X
ψn dµ = ai ψn (xi ) + an ψn (xn ) + ai ψn (xi ).
i=1 i=n+1
P∞ −i
As |ai | ≤ 1
3 for all i and i=n+1 3 = 12 3−n , by the assumptions
∞
1 n−1
Z
|ψn (xi )| + 3−n kψn kL∞ − kψn kL∞ 3−i
X X
| ψn dµ| ≥ −
3 i=1 i=n+1
1 1 1
≥ − + = > 0.
3 2 6
R
Therefore ψn dµ does not converge to zero, which contradicts our hypothesis.
ε↓0
Corollary 4.12. Let ϕε ∈ D(Ω) for all ε > 0 and ϕ ∈ D(Ω). Then ϕε −−→ ϕ in D(Ω) if
and only if (a) and (b):

(a) There exists a compact set K ⊂ Ω such that the supports of ϕε and ϕ lies within
K for all ε > 0.
ε↓0
(b) kϕε − ϕkC m −−→ 0 for all m ∈ N.

Proof. Again the “if” part is left as an exercise. If (a) does not hold, then there exists
a sequence (εn )n∈N in (0, ∞) with εn ↓ 0 such that (a) of Theorem 4.11 does not hold,
and thus ϕεn 6→ ϕ and so ϕε 6→ ϕ in D(Ω). Similarly, if (b) does not hold, then one can
conclude that ϕε 6→ ϕ in D(Ω).

Definition 4.13. Let X and Y be topological vector spaces. A function f : X → Y is

called sequentially continuous if for any sequence (xn )n∈N and x in X:
xn → x =⇒ f (xn ) → f (x).
We say that X is sequentially continuously embedded in Y , and write X ,→seq Y if there
exists a sequentially continuous embedding X → Y .
As a continuous map is sequentially continuous, if X is continuously embedded in Y ,
then X is sequentially continuously embedded in Y .

30
 In general (topological spaces), sequential continuity does not imply continuity. (For
those who are familiar to the notion of nets, continuity of a function f : X → Y , with X
and Y topological vector spaces, means f (xι ) → u(x) for any net (xι )ι∈I with xι → x in
X.) However, as the next theorem implies, a linear function on testfunctions with values
in F is continuous if and only if it is sequentially continuity.

Theorem 4.14. A linear function u : D(Ω) → F is a distribution if and only if it is

sequentially continuous, i.e., ϕn → ϕ implies u(ϕn ) → u(ϕ) for all sequences (ϕn )n∈N
and ϕ in D(Ω).

Exercise 4.D. Prove Theorem 4.14. (Prove that a linear function D(Ω) → F which is
not a distribution is not sequentially continuous.)

4.15 (D is not metrizable). Let us show that D(Ω) is not metrizable. We show that if
there is a metric on D(Ω), then it generates a different topology. Suppose d is a metric
on D(Ω), such that under the topology of d, D(Ω) is a topological vector space. We can
find a increasing sequence of compact sets (Kn )n∈N who’s union equals Ω. For n ∈ N,
let χn be a test function that equals 1 on Kn . We can and do choose λn ∈ R such that
d(λn χn , 0) ≤ 2−n . Then λn χn converges to 0 but (a) of Theorem 4.11 is not satisfied,
which means that λn χn converges in the topology generated by d but not in the weak
topology σ(D, D0 ).

Remember the seminorms k · kC m ,K defined in (4.4).

Definition 4.16. We define E(Ω) to be the set C ∞ (Ω) equipped with the topology
generated by the seminorms k · kC m ,K with K ⊂ Ω compact and m ∈ N0 .
By Theorem 4.11 the space D(Ω) is sequentially continuously embedded in E(Ω);

D(Ω) ,→seq E(Ω).

Exercise 4.E. Justify the statement D(Ω) ,→seq E(Ω).

Definition 4.17. Let K ⊂ Ω be compact. We define DK (Ω) to be the space

{ϕ ∈ D(Ω) : supp ϕ ⊂ K},

equipped with the topology generated by the seminorms k · kC m with m ∈ N0 .

Definition 4.18. A locally convex space is called a Fréchet space if it is complete and
metrizable with a translation invariant metric.

Theorem 4.19.

(a) For each m ∈ N0 the space C m (Ω), equipped with the seminorms k · kC m ,K with
K ⊂ Ω compact, is a Fréchet space.
(b) E(Ω) is a Fréchet space.

31
 (c) DK (Ω) is a Fréchet space for each compact set K ⊂ Ω.
S
Proof. (a) Let (Kn )n∈N be the increasing sequence of compact sets such that n∈N Kn =
Ω as in Theorem 1.8. Then the topology of C m (Ω) is generated by the seminorms
k · kC m ,Kn with n ∈ N, so that by Theorem 3.8 we see that C m (Ω) is metrizable with a
translation invariant metric. Let us show that C m (Ω) is complete.
For m = 0 this follows from the fact that for any compact set K ⊂ Ω the space C(K)
is complete under the supremum norm. Indeed, if (fn )n∈N is a Cauchy sequence in C 0 (Ω),
then for each compact K ⊂ Ω there exists a continuous function fK such that fn → fK
uniformly on K. As fK equals fK̃ for a compact K̃ ⊂ Ω with K ⊂ K̃. Then there exists
a continuous function f ∈ C(Ω) such that kfn − f kC 0 ,K → 0 for each compact K ⊂ Ω
◦
(take f (x) = fKn (x) for x ∈ Kn \ Kn−1 with K0 = ∅).
Suppose m ∈ N and (fn )n∈N is a Cauchy sequence in C m (Ω). Then for all β ∈ Nd0
with |β| ≤ m, the sequence (∂ β fn )n∈N is a Cauchy sequence in C(Ω); hence there exists
a gβ in C(Ω) such that ∂ β fn → gβ in C 0 (Ω). Let us write f for g0 . It is sufficient to
show that ∂ β f exists and equals gβ . By performing an induction argument, we may as
well assume that |β| = 1, i.e., β = ei for some i ∈ {1, . . . , d} (where ei is the i-th unit
vector in Rd ). For all x ∈ Rd and h ∈ R we have

f (x + hei ) − f (x) = lim fn (x + hei ) − fn (x)

n→∞
Z h Z h
= lim ∂i fn (x + tei ) dt = gei (x + tei ) dt,
n→∞ 0 0

and thus gei = ∂i f .

(b) As in (a), there are countably many seminorms that generate the topology, hence
E(Ω) is metrizable with a translation invariant metric. The completeness basically follows
from the completeness of C m (Ω): Suppose (fn )n∈N is a Cauchy sequence in E(Ω). Then,
for m ∈ N, it is a Cauchy sequence in C m (Ω) and therefore there exists a gm such that
limn→∞ fn = gm in C m (Ω). As convergence in C m+1 (Ω) implies convergence in C m (Ω)
for each m ∈ N, we have gm+1 = gm and thus gm = g0 for all m ∈ N. Therefore
f := g0 ∈ C m (Ω) for each m ∈ N and thus f ∈ E(Ω) and fn → f in E(Ω).
(c) By (b) it is sufficient to show that DK (Ω) is closed in E(Ω). Suppose ϕ ∈ E and
(ϕn )n∈N is a sequence in DK (Ω) that converges to ϕ in E. Then for x ∈ Ω \ K

|ϕ(x)| ≤ |ϕn (x)| + |ϕ(x) − ϕn (x)| ≤ kϕ − ϕn kC 0 (n ∈ N).

By taking n → ∞ on the right-hand side, we see that ϕ = 0 outside K, i.e., ϕ ∈

DK (Ω).

For the proof that D0 (Ω) is weak∗ complete, see Theorem 4.26, we use Baire’s Category
Theorem. A proof can be found for example in [Rud91, Theorem 2.2] or [dPvR13,
Theorem 11.1] .

32
Theorem 4.20 (Baire’s Category Theorem). Let X be a complete nonempty metric
space and let U1 , U2 , . . . be dense open subsets of X. Then the intersection of those sets,
T
n∈N Un is dense in X.

4.21. Equivalent to the above statement of Baire’s Category Theorem, one obtains the
following statement by taking complements: If X is a complete nonemtpy metric space
S
and A1 , A2 , . . . are closed subsets of X such that the interior of n∈N An is nonempty,
then there exists an n such that the interior of An is nonempty.
Lemma 4.22. Let X be a topological vector space whose topology is compatible with a
translation invariant metric and assume that X is complete with respect to this metric.
Let q be a seminorm on X such that the set {x ∈ Ω : q(x) ≤ 1} is closed. Then q is
continuous.

Proof. For t ∈ [0, ∞) put [q ≤ t] = {x ∈ Ω : q(x) ≤ t}. As X = m∈N [q ≤ m] and every

S

[q ≤ m] is closed, the Baire Category Theorem tells us that for some m the set [q ≤ m] has
an interior point, a ∈ [q ≤ m]◦ , say. Then 0 = a − a ∈ [q ≤ m]◦ + [q ≤ m]◦ = [q ≤ 2m]◦ .
Then every set [q ≤ t] for t > 0 is a neighbourhood of 0, so that q is continuous at 0. As
|q(x) − q(y)| ≤ q(x − y) for all x, y ∈ X, it follows that q is continuous (Exercise 4.F).

Exercise 4.F. Let q be a seminorm on a topological vector space X. Prove that the
following statements are equivalent:

(a) q is continuous.
(b) {x ∈ X : q(x) < 1} is open.
(c) 0 ∈ {x ∈ X : q(x) < 1}◦ .
(d) 0 ∈ {x ∈ X : q(x) ≤ 1}◦ .
(e) q is continuous at 0.
(f) There exists a continuous seminorm p on X such that q ≤ p (i.e., q(x) ≤ p(x) for
all x ∈ X).

Exercise 4.G. (a) Suppose that P is a collection of seminorms on a vector space X

such that

q(x) = sup p(x) < ∞ (x ∈ X).

p∈P

Show that q is a seminorm.

(b) Suppose that F is a collection of lower semicontinuous functions on a topological
space X such that

g(x) = sup f (x) < ∞ (x ∈ X).

f ∈F

Show that g is lower semicontinuous (which means that [g ≤ c] = {x ∈ X : g(x) ≤

c} is closed for all c > 0).

33
 Exercise 4.H. Prove: Suppose X is a locally convex vector space whose topology is
generated by countably many seminorms p1 , p2 , . . . with p1 ≤ p2 ≤ · · · . A linear function
f : X → F is continuous if and only if there exist C > 0 and n ∈ N such that

|f (x)| ≤ Cpn (x) (x ∈ X).

Definition 4.23. We write E 0 (Ω) for the space of continuous linear functions u : E(Ω) →
F. That is, u ∈ E 0 (Ω) if and only if there exist a compact set K, a m ∈ N0 and a C > 0
such that

|u(ϕ)| ≤ CkϕkC m ,K (ϕ ∈ E(Ω)). (4.6)

We equip E 0 (Ω) with the weak* topology σ(E 0 (Ω), E(Ω)).

The following theorem is a principle of uniform boundedness. For a more general

statement, see [Rud91, Theorem 2.6], which is a consequence of the Banach–Steinhaus
Theorem.
Theorem 4.24. (a) Let U ⊂ D0 (Ω) and assume

sup |u(ϕ)| < ∞ (ϕ ∈ D(Ω)).

u∈U

Then, for each compact K ⊂ Ω there exists a C > 0 and m ∈ N0 such that

|u(ϕ)| ≤ CkϕkC m (ϕ ∈ DK (Ω), u ∈ U). (4.7)

(b) Let U ⊂ E 0 (Ω) and assume

sup |u(ϕ)| < ∞ (ϕ ∈ E(Ω)).

u∈U

Then, there exists a compact K ⊂ Ω, C > 0 and m ∈ N0 such that

|u(ϕ)| ≤ CkϕkC m ,K (ϕ ∈ E(Ω), u ∈ U). (4.8)

Proof. (a) Let K ⊂ Ω be compact. Define the function q : DK (Ω) → [0, ∞) by

q(ϕ) := sup |u(ϕ)|.

u∈U

This defines a seminorm as it is the supremum of a family of seminorms and it is lower

semicontinuous as it is the supremum of continuous functions (see Exercise 4.G). There-
fore {ϕ ∈ DK (Ω) : q(ϕ) ≤ 1} is closed in DK (Ω). Therefore, by Lemma 4.22 q is
continuous on DK (Ω). As the topology of DK (Ω) is generated by the seminorms k · kC m
for m ∈ N and k · kC m ≤ k · kC k for k ≥ m, there exists a C > 0 and m ∈ N such that
(4.7) (see Exercise 4.H).
(b) follows by the above argument but with “D(Ω)” and “DK (Ω)” both replaced by
“E(Ω)” and using that the topology on E(Ω) is generated by the seminorms k · kC m ,K for
m ∈ N and K ⊂ Ω compact.

34
Proposition 4.25. The pairing maps D0 (Ω) × D(Ω) → F, (u, ϕ) 7→ u(ϕ) = hu, ϕi and
E 0 (Ω) × E(Ω) → F, (v, ψ) 7→ v(ψ) = hv, ψi are sequentially continuous.

Proof. Let (un , ϕn )n∈N and (u, ϕ) be in D0 (Ω) × D(Ω) such that (un , ϕn ) → (u, ϕ) in
D0 (Ω) × D(Ω), i.e., un → u in D0 (Ω) and ϕn → ϕ in D(Ω). By Theorem 4.11 (a)
there exists a compact set K ⊂ Ω such that supp ϕn , supp ϕ ⊂ K for all n ∈ N. By
Theorem 4.24 there exists a C > 0 and m ∈ N0 such that

|un (ϕ)| ≤ CkϕkC m (ϕ ∈ DK (Ω), n ∈ N).

Therefore

|un (ϕn ) − u(ϕ)| ≤ |un (ϕn − ϕ)| + |un (ϕ) − u(ϕ)| ≤ Ckϕn − ϕkC m + |un (ϕ) − u(ϕ)|.

The latter converges to zero by Theorem 4.11 (b).

The sequential continuity of the map E 0 (Ω) × E(Ω) → F, (v, ψ) 7→ v(ψ) = hv, ψi
follows similarly and the proof is left to the reader.

Theorem 4.26. D0 (Ω) and E 0 (Ω) are weak* sequentially complete.

Proof. First we prove that D0 (Ω) is weak∗ complete. The completeness of E 0 (Ω) can be
proved similarly; we comment on this at the end of the proof.
Suppose that (un )n∈N is a sequence in D0 (Ω) such that (hun , ϕi)n∈N is a Cauchy
sequence for all ϕ ∈ D(Ω). It will be clear what the limit should be: We define u :
D(Ω) → F such that hu, ϕi = limn→∞ hun , ϕi for any ϕ ∈ D(Ω). Clearly u is linear, so
let us show that it is a distribution. By Theorem 4.24, for each compact K ⊂ Ω, there
exist a C > 0 and m ∈ N0 such that

|u(ϕ)| ≤ sup |un (ϕ)| ≤ CkϕkC m (ϕ ∈ DK (Ω)).

n∈N

Therefore u is a distribution.
If (un )n∈N is a Cauchy sequence in E 0 (Ω), then one can follow the above prove with
“D(Ω)” and “DK (Ω)” both replaced by “E(Ω)”.

4.27. We equip the space of locally integrable functions on Ω with the topology defined
by the seminorms k · kL1 ,K with K ⊂ Ω being compact, where
Z
kϕkL1 ,K := kϕ1K kL1 = |ϕ| (ϕ ∈ D(Ω)).
K

Similarly, for p ∈ [1, ∞], Lploc (Ω) is equipped with the seminorms k · kLp ,K with K ⊂ Ω
being compact, defined by kϕkLp ,K := kϕ1K kLp .
It is rather straightforward to check for p ∈ [1, ∞], Lp (Ω) ,→ Lploc (Ω) and

E(Ω) ,→ Lploc (Ω) ,→ L1loc (Ω) (p ∈ [1, ∞]). (4.9)

35
Exercise 4.I. Show (4.9). (Hint: Exercise 2.A.)
Theorem 4.28. Let p ∈ [1, ∞]. The embedding D(Ω) → Lp (Ω) is sequentially continu-
ous. Moreover, the function Lploc (Ω) → D0 (Ω), f 7→ uf is a continuous embedding:

D(Ω) ,→ Lp (Ω) ,→ Lploc (Ω) ,→ D0 (Ω) (p ∈ [1, ∞]).

Proof. With Theorem 4.11 it follows that D(Ω) → Lp (Ω) is sequentially continuous. By
(4.9) it is sufficient to show that the map Lploc (Ω) → D0 (Ω), f 7→ uf is a continuous
embedding for p = 1. The injectivity follows from Lemma 2.9. The continuity is left as
an exercise (see Exercise 4.J).

Exercise 4.J. Prove the continuity of the functions Lploc (Ω) → D0 (Ω), f 7→ uf and
M(Ω) → D0 (Ω), µ 7→ uµ (see Definition 2.23 for M(ω)).

The following theorem is a kind of counterpart to Theorem 2.11.

Theorem 4.29. Let U be an open covering of Ω. For each U ∈ U let uU be a distribution
on U . Suppose that, if U, V ∈ U, then uU = uV on U ∩V , in the sense that uU (ϕ) = uV (ϕ)
for all ϕ ∈ D(Rd ) with supp ϕ ⊂ U ∩ V . Then there exists a unique distribution u on Ω
such that u = uU on U for all U ∈ U.

Proof. Choose a partition (χn )n∈N of unity subordinated to U as in Theorem 1.11. As

for every ϕ ∈ D(Ω) there is an N ∈ N with ϕ = N
P
n=1 χn ϕ, there is a unique linear
u : D(Ω) → F described by
∞
X
u(ϕ) = uUn (χn ϕ) (ϕ ∈ D(Ω)),
n=1

where Un = {x ∈ Ω : χn (x) > 0}. We are done if this u is a distribution.

That u is a linear function on D(Ω) is straightforward to check. For the continuity
we use Theorem 4.14 to restrict to sequential continuity. By Theorem 4.11 we know that
if ϕn → ϕ in D(Ω), then there exists a compact set K that contains the supports of all
ϕn ’s. Therefore, there are only finitely many k such that uUk (χk ϕn ) is nonzero for some
n. That is, there exists a L ∈ N such that u(ϕn ) = L k=1 uUk (χk ϕn ) for all n ∈ N. As
P

for all k we have χk ϕn → χk ϕ, we have uUk (χk ϕn ) → uUk (χk ϕ). From this we conclude
the continuity of u.

5 Compactly supported distributions

In Definition 4.16 we have introduced the topological space E(Ω) consisting of all smooth
functions on Ω and in Definition 4.23 we defined its topological dual E 0 (Ω). In this section
we will further study E 0 (Ω), which is in one to one correspondence to the set of compactly
supported distributions, see 5.2.

36
Definition 5.1. Let u be a distribution on Ω. We call a (relatively) closed subset A of
Ω a carrier of u if u(ϕ) = 0 for every ϕ ∈ D(Ω) with supp ϕ ∩ A = ∅. The intersection
of all carriers of u is defined to be the support of u,

supp u.

We prove supp u to be a carrier of u (hence, the smallest carrier of u). Let U be the
collection of all complements (in Ω) of the carriers of u; take ϕ ∈ D(Ω) with supp ϕ ⊂ U.
S

We wish to prove u(ϕ) = 0. By Theorem 1.11 there exist χ1 , . . . , χN ∈ D(Ω) with

ϕ = N n=1 χn ϕ, where for each n the support of χn is contained in an element of U.
P

Then supp χn ϕ ⊂ supp χn , whence u(χn ϕ) = 0 for each n, and u(ϕ) = 0.

We say that u vanishes on an open set U ⊂ Ω if u(ϕ) = 0 for all ϕ ∈ D(Ω) with
supp ϕ ⊂ U . Then Ω \ U is a carrier of u. Moreover, let U be the collection of all open
sets U ⊂ Ω on which u vanishes. Then u vanishes on U and supp u is the complement
S

of U.
S

From this the following statements are immediate

ϕ ∈ D(Ω), supp ϕ ∩ supp u = ∅ =⇒ u(ϕ) = 0,

χ ∈ C ∞ (Ω), χ = 1 on a neighbourhood of supp u =⇒ χu = u.

Note that supp ϕ ∩ supp u = ∅ means that ϕ equals 0 on an open set that contains supp u.
If ϕ equals 0 on supp u, then the evaluation u(ϕ) might not equal zero as the following
example illustrates: Take u = ∂δ0 and let ϕ be a testfunction such that ϕ(x) = x around
0 (see Exercise 1.E).
Observe moreover that for α ∈ Nd0 and ψ ∈ C ∞ (Ω)

supp ∂ α u ⊂ supp u, supp ψu ⊂ supp ψ ∩ supp u.

Exercise 5.A. Show that each compactly supported distribution is of finite order.

5.2 (Each element of E 0 defines a compactly supported distribution). Let u ∈ E 0 (Ω) and
let K ⊂ Ω be compact, m ∈ N0 and C > 0 be such that

|u(ϕ)| ≤ CkϕkC m ,K (ϕ ∈ E(Ω)).

If ϕ ∈ E(Ω) and supp ϕ∩K = ∅ then kϕkC m ,K = 0 and thus u(ϕ) = 0. Hence supp u ⊂ K
and so an element of E 0 (Ω) defines a distribution with compact support. In 5.6 we will
prove that a distribution with compact support can be extended to an element of E 0 (Ω).

Proposition 5.3. For each m ∈ N there exists a C > 0 such that

kf gkC m ≤ Ckf kC m kgkC m (f, g ∈ C m (Ω)) (5.1)

m
kf gkC m ,K ≤ Ckf kC m ,K kgkC m ,K (f, g ∈ C (Ω), K ⊂ Ω compact). (5.2)

37
Consequently, the functions
E(Ω) × D(Ω) → D(Ω), (ψ, ϕ) 7→ ψϕ,
0 0
D (Ω) × E(Ω) → D (Ω), (u, ϕ) 7→ ϕu,
0 0
E (Ω) × E(Ω) → E (Ω), (v, ψ) 7→ ψv,
are sequentially continuous.

Proof. By Leibniz’ differentiation rule (see 1.14) we have for x ∈ Ω

!
α
X α
max |∂ (f g)(x)| ≤ max |∂ β f (x)||∂ γ g(x)|
α∈Nd0 α∈Nd0 β
β∈Nd0
|α|≤k |α|≤k
β≤α
!!
X α
≤ max max max |∂ β f (x)||∂ γ g(x)|. (5.3)
α∈Nd0 β β∈Nd0 γ∈Nd0
β∈Nd0
|α|≤k |β|≤k |γ|≤k
β≤α
P α

Hence with C = maxα∈Nd :|α|≤k β∈Nd0 :β≤α β one has (5.1) and (5.2).
0

The continuity of the product maps D0 (Ω)×E(Ω) → D0 (Ω) and E 0 (Ω)×E(Ω) → E 0 (Ω)
follow by Proposition 4.25.

5.4. Observe that Leibniz’ rule (1.3) extends to the product of a distribution with a
smooth function. That is, if u ∈ D(Ω) and ψ ∈ C ∞ (Ω), then
!
α
X α
∂ (ψu) = (∂ β ψ)(∂ α−β u),
β
β∈Nd0
β≤α

5.5. Observe that for each compact set K ⊂ Ω there exists a χ ∈ Cc∞ (Ω, [0, 1]) such that
χ = 1 on a neighbourhood of K: Let K ⊂ Ω be compact and such that K ⊂ K◦ (which
exists by Theorem 1.8). By Lemma 1.13 there exists such a χ which equals 1 on K and
thus on K◦ which is an open set that contains K, i.e., K is a neighbourhood of K.
5.6 (Each compactly supported distribution extends to an element of E 0 ).
Let u be a distribution on Ω with compact support K. We will show that there exists
exactly one v ∈ E 0 such that u(ϕ) = v(ϕ) for all ϕ ∈ D(Ω).
We have already seen in Definition 5.1 that if ϕ ∈ D(Ω) and supp ϕ ⊂ Ω \ K, then
u(ϕ) = 0. Let χ be a testfunction that is equal to 1 on a neighbourhood of K. Then
supp(ϕ − χϕ) ⊂ Ω \ K and thus
u(ϕ) = u(χϕ) (ϕ ∈ D(Ω)). (5.4)
Let K0 = supp χ. As u is a distribution, there exist C1 > 0 and m ∈ N0 such that
|u(ϕ)| ≤ C1 kϕkC m for all ϕ ∈ D(Ω) with supp ϕ ⊂ K0 . This implies
|u(ϕ)| = |u(χϕ)| ≤ C1 kχϕkC m (ϕ ∈ D(Ω)).

38
Let C > 0 be such that (5.1) holds, then with C 0 = C1 CkχkC m ,

|u(χϕ)| ≤ C 0 kϕkC m ,K0 (ϕ ∈ E(Ω)). (5.5)

Therefore v : E(Ω) → F defined by v(ϕ) = u(χϕ) for ϕ ∈ E is an element of E 0 .

Let us show that this v is the only element of E 0 such that v(ϕ) = u(ϕ) for ϕ ∈ D(Ω).
Suppose w ∈ E 0 is such that w(ϕ) = u(ϕ) for ϕ ∈ D(Ω). Let L ⊂ Ω be compact, m ∈ N
and M > 0 such that

|w(ϕ)| ≤ M kϕkC m ,L (ϕ ∈ E(Ω)).

Let η ∈ D(Ω) be equal to 1 on a neighbourhood of L. Then w(ηψ) = w(ψ) for all

ψ ∈ E(Ω). Therefore

w(ψ) = w(ηψ) = u(ηψ) = u(χηψ) = u(χψ) = v(ψ) (ψ ∈ E(Ω)).

With a partition of unity we have the following approximations.

Lemma 5.7. Let (χn )n∈N be a partition of unity as in Theorem 1.11. Then
N
X N →∞
χn ϕ −−−−→ ϕ in D(Ω) (ϕ ∈ D(Ω)), (5.6)
n=1
N
N →∞
χn u −−−−→ u in D0 (Ω) (u ∈ D0 (Ω)),
X
(5.7)
n=1
N
X N →∞
χn ϕ −−−−→ ϕ in E(Ω) (ϕ ∈ E(Ω)), (5.8)
n=1
N
N →∞
χn u −−−−→ u in E 0 (Ω) (u ∈ E 0 (Ω)).
X
(5.9)
n=1

Proof. As for each ϕ ∈ D(Ω) there exists an N ∈ N such that N

P
n=1 χn ϕ = ϕ, (5.6) and
(5.7) follow immediately. Let K ⊂ Ω be compact. Let K ⊂ Ω be a compact set such that
K ⊂ K◦ and let N ∈ N be such that N n=1 χn (x) = 1 for all x ∈ K. Then
P

N
X
ψ− χn ψ =0 (ψ ∈ E(Ω), m ∈ N0 ),
n=1 C m ,K

so that (5.8) and (5.9) follow.

Exercise 5.B. Show that if u ∈ E 0 (Ω) and u 6= 0, then there exists a ϕ ∈ D(Ω) such that
u(ϕ) 6= 0. Moreover, show that u(ψ) = 0 for every ψ ∈ E(Ω) with supp ψ ∩supp(u|D ) = ∅.

39
Definition 5.8. Let X be a topological space. We call a set A ⊂ X sequentially dense
in X if for each x ∈ X there exists a sequence (an )n∈N in A that converges to x.
If there exists a countable sequentially dense subset of X, we call X sequentially
separable.
Any sequentially dense set is also dense. If X is a metric space, then any dense set is
also sequentially dense. Not every dense set is sequentially dense:

Example 5.9. Let X = R (or a uncountable set) and call a set U ⊂ X open if its
complement is either countable or equal to X. With this topology the set X \ {0} is
dense in X but not sequentially dense as if a sequence (xn )n∈N converges to some x in
X, then there exists an N ∈ N such that xn = x for all n ≥ N .

Exercise 5.C. Verify the statement in Example 5.9.

Theorem 5.10. The embedding D(Ω) → E(Ω) is a sequential continuous embedding; the
map ι : E 0 (Ω) → D0 (Ω) defined by ι(u) = u|D(Ω) is a continuous embedding and its image
is the set of compactly supported distributions and the map ζ : E 0 (Ω) → E 0 (Rd ) defined by
(ζu)(ψ) = u(ψ|Ω ) is a continuous embedding:

D(Ω) ,→seq E(Ω), E 0 (Ω) ,→ D0 (Ω), E 0 (Ω) ,→ E 0 (Rd ).

Moreover, D(Ω) is sequentially dense in E(Ω) and E 0 (Ω) is sequentially dense in D0 (Ω).

Proof. That D(Ω) → E(Ω) is sequentially continuous follows from Theorem 4.11. That
ι is continuous follows from the fact that D(Ω) ⊂ E(Ω). That ι forms a bijection on to
the set of compactly supported distributions follows from 5.2 and 5.6. That the image
of E 0 (Ω) is sequentially dense in D0 (Ω) follows from Lemma 5.7.
That the map ζ is continuous follows as ψ|Ω is in E(Ω) for each ψ ∈ E(Rd ).
That D(Ω) is sequentially dense in E(Ω) and that E 0 (Ω) is sequentially dense in D0 (Ω)
follows from Lemma 5.7.

Definition 5.11. Let y ∈ Rd , α ∈ Nd0 , ψ ∈ C ∞ (Ω) and l : Rd → Rd linear and bijective.

For a u ∈ E 0 (Ω) we define ǔ, Ty u, ∂ α u, ψu and u ◦ l by the formulas as in Definition 2.14
but replacing “D” everywhere by “E”.
Again, it is straightforward to check that ǔ, Ty u, ∂ α u, ψu and u ◦ l are all in E 0 (Ω) and
moreover that with ι : E 0 (Ω) → D0 (Ω) as in Theorem 5.10,

ι(ǔ) = ι(u)ˇ,
ι(Ty u) = Ty ι(u),
ι(∂ α u) = ∂ α ι(u),
ι(ψu) = ψι(u),
ι(u ◦ l) = ι(u) ◦ l.

40
Definition 5.12. For v ∈ E 0 (Ω) we define

supp v = supp(v|D ).

By Exercise 5.B, supp v is the smallest closed set A such that

v(ψ) = 0 for all ψ ∈ E(Ω) with supp ψ ∩ A = ∅.

In 5.6 the inequality (5.5) holds for K0 which is larger than K. The next exercise
illustrates that (5.5) may not hold for K0 = supp u.

Exercise 5.D. [DK10, Exercise 8.3] Let d = 1. Let (xn )n∈N be a sequence of distinct
elements and x be in R such that xn → x and such that xn 6= x for all n ∈ N.

(a) Show that there exists a sequence (an )n∈N in (0, ∞) such that
X X
an = ∞, an |xn − x| < ∞.
n∈N n∈N

(b) Prove that the formula

X
u(ϕ) = an (ϕ(xn ) − ϕ(x)) (ϕ ∈ D(R))
n∈N

defines a distribution u of order ≤ 1. Prove that the support of u is the compact

set {xn : n ∈ N} ∪ {x}.
(c) Show that for all n ∈ N there exists a ϕn ∈ D such that ϕn = 1 on a neighbourhood
of xi for all i ∈ {1, . . . , n} and ϕn = 0 on a neighbourhood of xj for all j > n and
ϕn = 0 on a neighbourhood of x. Prove that for all m ∈ N
n
X
kϕn kC m ,supp u = 1, u(ϕn ) = ai .
i=1

(d) Conclude that for K = supp u, (4.6) does not hold for any k ∈ N.

The following example illustrates that the embeddings D(Ω) → E(Ω) and E 0 (Ω) →
D0 (Ω) from Theorem 5.10 are not homeomorphisms on their images.

5.13. We show that (a) the relative topology of D(Ω) as a subspace of E(Ω) is different
from the topology on D(Ω), namely σ(D(Ω), D0 (Ω)); (b) E 0 (Ω) does not have the same
topology as ι(E 0 (Ω)) (However, (E 0 (Ω), σ(E 0 (Ω), D(Ω))) is homeomorphic to ι(E 0 (Ω)).)
and (c) E 0 (Ω) is not metrizable.
Choose (χn )n∈N as in Theorem 1.11, letting U consist of all open subsets of Ω (we
assume χn 6= 0 for all n ∈ N). For each n ∈ N choose xn such that χn (xn ) 6= 0.

41
 R
(a) Let (λn )n∈N in (0, ∞) be such that λn χn = 1 for all n ∈ N. Let ϕn = λn χn for
all n ∈ N. For all compact sets K there exists an N ∈ N such that supp ϕn ∩ K =
n→∞
supp χn ∩ K = ∅ for all n ≥ N . Therefore kϕn kC m ,K −−−→ 0 for all m ∈ N
and compact K ⊂ Ω, i.e., ϕn → 0 in E 0 (Ω). However, for u the distribution
corresponding to the Lebesgue measure, or equivalently to the constant function 1,
we have u(ϕn ) = 1 for all n, whence (ϕn )n∈N does not converge in D(Ω).
(b) δxn is an element of D0 (Ω) and of E 0 (Ω) for all n ∈ N. We have δxn → 0 in D0 (Ω)
but not in E 0 (Ω), as we have δn (1) = 1 for all n ∈ N.
(c) E 0 (Ω) is not metrizable, as we proceed to show. Suppose its topology is given by a
metric, d. For every n ∈ N we have limλ↓0 λδxn = 0 in E 0 (Ω), so there is a λn > 0
with d(λn δxn , 0) < n1 . By Remark 1.12 there exists a ψ ∈ C ∞ (Ω) = E(Ω) with
ψ(x) = ∞ 1 0
n=1 λn χn (xn ) χn (x); then λn δxn (ψ) ≥ 1 for all n. But λn δxn → 0 in E (Ω)
P

since d(λn δxn , 0) → 0. Contradiction.

6  Structure theorems

In this section we show that every distribution is a linear combination of derivatives of

continuous functions, that is, we describe the global structure of distributions. These
theorems are often called “structure theorems”. In this section we write DK instead of
DK (Ω) or DK (Rd ), where the identification for a compact K ⊂ Ω of an element DK (Rd )
with an element DK (Ω) is the obvious one.
We start by describing a distribution on DK .

Theorem 6.1. Let u ∈ D0 (Ω) and K ⊂ Ω be compact. Then there exists an f ∈ C(Ω)
and an α ∈ Nd0 ,

u(ϕ) = ∂ α uf (ϕ) (ϕ ∈ DK ). (6.1)

α can be chosen to be (N + 2, . . . , N + 2), where N ∈ N is such that there exists a C > 0

such that |u(ϕ)| ≤ CkϕkC N for all ϕ ∈ DK .

Proof. By performing a rescaling and a translation, we may as well assume that the
support of u lies within the unit cube Q = [0, 1]d (which itself does not need to be
included in Ω). By the mean value theorem we have

kψkL∞ ≤ max |∂i ψ(x)| (ψ ∈ DQ , i ∈ {1, . . . , n}). (6.2)

x∈Q

Qd
Let T = ∂1 ∂2 · · · ∂d , i.e., T = ∂ (1,...,1) . For y ∈ Q let Q(y) = i=1 [0, yi ]. Then
Z
ψ(y) = (T ψ)(x) dx (ψ ∈ DQ ). (6.3)
Q(y)

42
For N ∈ N we have by (6.2) and (6.3)
Z
kψkC N ≤ max |T N ψ(x)| ≤ |T N +1 ψ(x)| dx (ψ ∈ DQ ).
x∈Q Q

Let C > 0 and N ∈ N be such that

|u(ϕ)| ≤ CkϕkC N (ϕ ∈ DK ),

so that
Z
|u(ϕ)| ≤ C max |T N ϕ(x)| ≤ C |T N +1 ϕ(x)| dx (ϕ ∈ DK ). (6.4)
x∈Q Q

By (6.3) it follows that T is injective on DK , hence T N +1 is injective on DK . Let

Y = {T N +1 ϕ : ϕ ∈ DK }. Define u1 : Y → F by u1 = u ◦ (T N +1 )−1 , i.e.,

u1 (T N +1 ϕ) = u(ϕ) (ϕ ∈ DK ).

By (6.4) we have
Z
|u1 (ψ)| ≤ C |ψ(x)| dx (ψ ∈ Y ).
K

By the Hahn–Banach Theorem (Theorem 3.14) u1 extends to a bounded linear functional

on L1 (K), i.e., an element of L1 (K)0 which can be represented by a bounded Borel–
measurable function g (in L∞ (K)):
Z
u(ϕ) = u1 (T N +1 ϕ) = g(x)(T N +1 ϕ)(x) dx (ϕ ∈ DK ). (6.5)
Q

Extend g as a function on Rd by setting g = 0 on Rd \ K and put

Z
f (y) = (−1)(N +1)d Qd g (y ∈ Rd ).
i=1
(−∞,yi ]

Then f is continuous, and by applying integration by parts d times, (6.5) gives

Z
u(ϕ) = (−1)(N +2)d f (x)T N +2 ϕ(x) dx (ϕ ∈ DK ).
Ω

This implies (6.1).

We use the previous theorem to represent compactly supported distributions:

Theorem 6.2. Let u ∈ D0 (Ω) be compactly supported with support K. Let U ⊂ Ω be open
such that K ⊂ U . Suppose u has order N . Then there exist finitely many continuous
functions fβ ∈ C(Ω) with supp fβ ⊂ U for β ∈ Nd0 with β ≤ (N + 2, . . . , N + 2), such that
X
u= ∂ β ufβ .
β∈Nd0
β≤(N +2,...,N +2)

43
Proof. Let V, W ⊂ Ω be open sets such that V and W are compact and K ⊂ V ⊂ W ⊂ U .
(use Theorem 1.8 with “U ” (or “W ”) instead of “Ω”). Let α = (N + 2, · · · , N + 2). By
Theorem 6.1 applied with K = W there exists a f ∈ C(Ω) such that

u(ϕ) = u∂ α f (ϕ) (ϕ ∈ D(W )).

Let χ ∈ D(Ω) be supported in W and equal to 1 on V . Then by Leibniz’ formula (see

1.14)
Z
u(ϕ) = u(χϕ) = (−1)|α| f · ∂ α (χϕ)
!
α
Z
|α|
X
= (−1) f· (∂ α−β χ)(∂ β ϕ)
β
β∈Nd0
β≤α
Z
(−1)|β|
X
= fβ (∂ β ϕ),
β∈Nd0
β≤α

for
!
α
fβ = (−1)|α−β| f · (∂ α−β χ) (β ∈ Nd0 : β ≤ α).
β

By using the partition of unity to represent a distribution by a sum of compactly

supported distributions, we obtain the following representation of general distributions.
Theorem 6.3. Let u ∈ D0 (Ω). There exist (gα )α∈Nd in C(Ω) such that for each compact
0
set K ⊂ Ω, supp gα ∩ K 6= ∅ for only finitely many α and
X
u= ∂ α ugα .
α∈Nd0

If u has finite order N , one can choose the gα such that gα = 0 for α 6≤ (N +2, . . . , N +2).

Proof. Let (χn )n∈N be a partition of unity as in Theorem 1.11. By Theorem 6.2 for
each n ∈ N there exists an Nn ∈ N and continuous functions fn,β for β ∈ Nd0 with
β ≤ (Nn , . . . , Nn ) with supports in Un = {x ∈ Ω : χn (x) > 0}, such that
X
χn u = ∂ β ufn,β .
β∈Nd0
β≤(Nn ,...,Nn )

For β ∈ Nd0 with β 6≤ (Nn , . . . , Nn ) set fn,β = 0. Define

∞
X
gα = fn,α
n=1

44
By the fact that the sets Un form a locally finite cover of Ω it follows that for all compact
K ⊂ Ω the intersection supp fn,β ∩ K being a subset of Vn ∩ K is nonempty only for
finitely many β and n. Therefore supp gα ∩ K 6= ∅ for only finitely many α. Furthermore
X X X X
u= χn u = ∂ β ufn,β = ∂ α ugα .
n∈N n∈N β∈Nd α∈Nd0
0

If u has order N , then one can choose Nn = N + 2 for all n ∈ N by Theorem 6.2.

7 Intermezzo: Convolutions of functions

We still consider Ω to be an open subset of Rd , though most statements are about

functions on Rd .
Before we define the convolution of two functions, we recall some measure theoretic
statements. With “measurable” in this section we mean “Borel measurable”.

7.1. Because the operations of addition and multiplication are measurable, the following
statement holds: If f, g : Rd → F are measurable and z ∈ Rd , then the following functions
are measurable

Rd → F, x 7→ f (x − z),
d d
R × R → F, (x, y) 7→ f (x − y),
Rd × Rd → F, (x, y) 7→ f (x)g(y),
d
R → F, x 7→ f (x)g(x).

We write f g for the function x 7→ f (x)g(x).

We recall Fubini’s theorem (for the product space Rd × Rd only). For a proof see for
example [Bog07, Theorem 3.4.4] or [Hal74, Theorem 36.C].

Theorem 7.2 (Fubini’s Theorem). Let f : Rd × Rd → F be integrable. Then for almost

all x ∈ Rd the functions y 7→ f (x, y) and y 7→ f (y, x) are integrable. If gR: Rd → RF is such
that g(x) = Rd f (x, y) dy for almost all x ∈ Rd , then g is integrable and Rd g = Rd ×Rd f .
R

Definition 7.3. Let f and g be measurable functions on Rd . The function (x, y) 7→

f (x)g(y − x) is measurable, and for almost every y ∈ Rd the function x 7→ f (x)g(y − x)
is measurable (see 7.1). Then we define a function f ∗ g on Rd by
(R
f (x)g(y − x) dx if x 7→ f (x)g(y − x) is integrable,
f ∗ g(y) =
0 otherwise

f ∗ g is called the convolution of f and g.

45
7.4. (1) Putting R
it differently: If f Ty ǧ is integrable for almost every y, then f ∗ g exists,
and f ∗ g(y) = f Ty ǧ for almost every y.
(2) If f ∗ g exists and if f˜, g̃ are functions with f˜ = f a.e. and g̃ = g a.e., then f˜∗ g̃ exists
and equals f ∗ g. Thus, we can see ∗ as an operation on equivalence classes of functions.
Also, “f ∗ g” is often viewed as an equivalence class.
(3) If f1 ∗ g and f2 ∗ g exist, then (f1 + f2 ) ∗ g exists and is equal (a.e.) to f1 ∗ g + f2 ∗ g.
(4) If f ∗ g exists, then g ∗ f exists an equals f ∗ g. Indeed, for almost every y the function
f Ty ǧ is integrable; then so are (T−y g)fˇ and gTy fˇ. The rest is easy.

Let us also recall the following theorem, which is sometimes also referred to as the
Fubini theorem (the statement of Theorem 7.2 is most commonly known under the name
Fubini’s theorem).

Theorem 7.5. If f : Rd × Rd → [0, ∞) is measurable, then

Z Z Z Z Z
f= f (x, y) dx dy = f (x, y) dy dx.
Rd ×Rd Rd Rd Rd Rd

Theorem 7.6. Let f and g be integrable functions on Rd . Then R

for almostRevery Ry the
function x 7→ f (x)g(y − x) is integrable, f ∗ g is integrable, and Rd (f ∗ g) = Rd f · Rd g,
kf ∗ gkL1 ≤ kf kL1 kgkL1 .

Proof. The function F : (x, y) 7→ f (x)g(y − x) on Rd × Rd , being the product of two

measurable functions, is measurable R
(see 7.1). For every x the function y 7→ F (x, y) is
integrable and its integral is f (x) Rd g. That F is integrable follows from Theorem
R R
7.5.
Therefore, by Fubini’s theorem f ∗g is defined and integrable and its integral is Rd f Rd g.

The following theorem will be used often later on. It generalises the inequality kf ∗
gkL1 ≤ kf kL1 kgkL1 .

Theorem 7.7 (Young’s inequality). Let p, q, r ∈ [1, ∞] be such that

1 1
p + q = 1 + 1r .

For f ∈ Lp (Rd ), g ∈ Lq (Rd ) we have f ∗ g ∈ Lr (Rd ) and

kf ∗ gkLr ≤ kf kLp kgkLq .

Proof. We assume f, g ≥ 0. Observing that p and q play symmetrical roles, we consider

four cases.
Case 1: r = 1 and thus p = q = 1. This case is covered by Theorem 7.6.
Case 2: r = ∞ and thus p1 + 1q = 1. Then for every y ∈ Rd , Rd f (x)g(y − x) dx exists
R

and is at most kf kLp kgkLq by Hölder’s inequality (see Theorem A.4). Consequently, f ∗ g
exists and kf ∗ gkL∞ ≤ kf kLp kgkLq .

46
 Case 3: 1 < r < ∞. As f p and g q are integrable, by Theorem 7.5 the function
x 7→ f p (x)g q (y − x) is integrable for almost all y ∈ Rd .
p q r−p r−q
Fix such a y. Put h(x) = g(y − x) for x ∈ Rd . As f h = f r h r · f r · h r , an
application of the Generalized Hölder inequality (see Theorem A.5) with n = 3,
pr qr
p1 = r, p2 = , p3 = ,
r−p r−q
1 1 1 1
(with the convention that 0 = ∞) so that p1 + p2 + p3 = 1, shows that f h is integrable
and
Z Z  1 Z  r−p Z  r−q Z 1
r pr qr r
p q p q p q
fh ≤ f h f h = f (x) g(y − x) dx A,

r−p r−q
with A = ( f p ) ( hq )
R R
pr qr , which is independent of y.
The above implies that the function x 7→ f (x)g(y − x) is integrable for almost every
y, so that f ∗g exists, and (f ∗g)r ≤ (f p ∗g q )Ar . By Theorem 7.6 it follows that f ∗g ∈ Lr
and
Z  Z Z
(kf ∗ gkLr )r ≤ f p ∗ g q Ar ≤ fp · g q · Ar
Z Z Z  r−p Z  r−q
p q
p q p q
= f · g · f g = kf krLp kgkrLq .

The following is a consequence of Young’s inequality.

Corollary 7.8. Let p, q, r ∈ [1, ∞] be such that

1 1 1
+ + = 2.
p q r

For f ∈ Lp (Rd ), g ∈ Lq (Rd ) and h ∈ Lr (Rd ) we have that (f ∗ g)h is integrable and
Z Z Z
(f ∗ g)h = f (g ∗ ȟ) = f (ǧ ∗ h) (7.1)
k(f ∗ g)hkL1 ≤ kf kLp kgkLq khkLr . (7.2)

Exercise 7.A. Prove Corollary 7.8.

In the following definition we define the essential support of a measurable function.

This agrees with the support of a distribution when it is given by a locally integrable
function. We consider the essential support of a convolution in Theorem 7.10.

47
Definition 7.9. Let f : Ω → F. We say that an open set U ⊂ Ω is f -null if f = 0 almost
everywhere on U . We define the essential support of f ,

ess supp f,

to be the complement of the union of all f -null open sets.

Actually, this union itself is f -null, hence is the largest f -null open set. To see
that, observe that there exist a countable number of f -null open sets (Un )n∈N such that
S
U = n∈N Un by Theorem 1.11.
As the essential support of f is equal to the one of g if f and g are equal almost
everywhere, one can make sense of the essential support for equivalence classes of locally
integrable functions in the usual way by identifying an equivalence class with an element
in it: If f ∈ L1loc (Ω), then ess supp f is defined by the essential support of any function
representing f .
Of course, for a continuous function f we have

supp f = ess sup f.

If f ∈ L1loc , then by Lemma 2.9 it follows for an open set U ⊂ Ω that Ω \ U is a carrier
for uf if and only if U is f -null, hence

ess sup f = supp uf .

Theorem 7.10. For any two measurable functions f, g on Rd such that f ∗ g exists, we
have

ess supp f ∗ g ⊂ ess supp f + ess supp g.

Proof. Let A = ess supp f , B = ess supp g, and take c ∈ Rd such that f ∗ g(c) 6= 0.
It suffices to prove c ∈ AR+ B. Now f = f 1A a.e. and g = g 1B a.e., so 0 6= f ∗
g(c) = (f 1A ) ∗ (g 1B )(c) = (f 1A )(x)(g 1B )(c − x) dx. Therefore there is an x such that
(f 1A )(x)(g 1B )(c − x) 6= 0: Then 1A (x) 6= 0 and 1B (c − x) 6= 0, whence c = c − x + x ∈
A + B.

Theorem 7.10 states that the support is included in the closure of the sum of two
closed sets. It is necessary to take the closure as Example 7.13 illustrates. We first
recall the following facts about closedness of sums of closed sets in Lemma 7.11 and
Example 7.12.

Lemma 7.11. Let A, B ⊂ Rd and A be compact and B closed. Then A + B is closed.

If, moreover, B is compact then A + B is compact.

Proof. Let (dn )n∈N be a sequence in A + B that converges to an element d in Rd . We

prove that d ∈ A + B. By definition, for each n there exist an ∈ A and bn ∈ B such that

48
dn = an + bn . As A is compact, (an )n∈N has a convergent subsequence. Let us assume
(an )n∈N itself converges in A to an element a. Then dn − an → d − a and as dn − an ∈ B
for all n and B is closed, d − a ∈ B, which implies d = a + d − a ∈ A + B.
If B is compact, then A + B is the image of the compact set A × B of the addition
function + : Rd × Rd → Rd , (x, y) 7→ x + y, which is continuous, hence the image is
compact.

The assumption that A is not only closed, but also bounded (which together is the
same as compact for subsets of Rd ) is essential as the following example illustrates.
1
Example 7.12. Let A = N and B = {−m + m : m ∈ N, m ≥ 2}. Then A + B is not
1
closed as m is an element of A + B for all m ∈ 1 + N but 0 is not.
Example 7.13. [ess supp f + ess supp g ( supp f ∗ g = ess supp f + ess supp g]
We adapt Example 7.12 to obtain two measurable functions f and g which are not almost
everywhere equal to zero. We define the sets A, B ⊂ R by
∞ h ∞ h
[ 1i [ 1 2i
A= n, n + , B= −m+ , −m + .
n=2
n m=2
m m

We define f, g : R → R by

f (x) = |x|−2 1A (x), g(x) = |x|−2 1B (x) (x ∈ R).

Then f and g are integrable functions and so f ∗ g exists (and is integrable). Moreover,
supp f = A, supp g = B,
∞ h
[ 1 2 1i
A+B = n−m+ ,n − m + + .
n,m=2
m m n

1
As in Example 7.12, the set A + B is not closed as 0 is not in A + B but m is for all
1 2 1
m ∈ 1 + N. For each n, m ∈ 1 + N and z ∈ (n − m + m , n − m + m + n ) we can show
that f ∗ g(z) 6= 0, so that as the support of a function is closed, A + B ⊂ supp f ∗ g. And
thus in this case A + B ( supp f ∗ g = A + B
Exercise 7.B. (a) Choose f, g ∈ L1 (R) such that f and g are not continuous but
f ∗ g is.
(b) Let α > −1. Let f (x) = g(x) = xα for x ∈ (0, 1) and f (x) = g(x) = 0 for x ∈
/ (0, 1).
Show that there exists a c > 0 such that f ∗ g(y) = cy 2α+1 . Conclude that f ∗ g is
not continuous for α ∈ (−1, − 21 ).
7.14 (Notation). For any closed set A ⊂ Rd we write [A]ε for the set of those points in
Rd that are at distance at most ε from A, so that

[A]ε = A + B(0, ε) = {y ∈ Rd : inf |x − y| ≤ ε}.

x∈A

By Lemma 7.11 [A]ε is closed.

49
Theorem 7.15. Let f ∈ L1loc (Rd ) and ψ ∈ Cc (Rd ). For ε > 0 we write ψε for the
function defined by ψε (x) = ε−d ψ(ε−1 x). Then the following statements hold.
ε↓0
(a) f ∗ ψε (x) −−→ ( ψ)f (x) for all Lebesgue points x ∈ Rd of f .
R

(b) If f is continuous on an open set U ⊂ Rd , then f ∗ ψε → ( ψ)f uniformly on all

R

compact subsets of U .
(c) If p ∈ [1, ∞) and f ∈ Lploc (Rd ), then f ∗ ψε → ( ψ)f in Lploc (Rd ).
R

(d) If p ∈ [1, ∞) and f ∈ Lp (Rd ), then f ∗ ψε → ( ψ)f in Lp (Rd ).

R

R R
Proof. As ψε = ψ for all ε > 0, we have
Z Z  
f ∗ ψε (x) − ( ψ)f (x) = ψε (x − y) f (y) − f (x) dy.

As we can find an ε > 0 such that supp ψε ⊂ B(0, 1), we may without loss of generality
assume that supp ψ ⊂ B(0, 1). Then
Z Z
|f ∗ ψε (x) − ( ψ)f (x)| ≤ kψkL∞ ε−d |f (y) − f (x)| dy. (7.3)
B(x,ε)

From this (a) follows. Suppose f is continuous on an open set U and K ⊂ U is compact.
Let δ > 0 be such that [K]δ ⊂ U . As f is uniformly continuous on [K]δ , the convergence
in (2.3) is valid uniformly for x ∈ K. Hence (b) also follows from (7.3).
Let us turn to the proof of (c). Let K ⊂ Rd be compact. We will show
kf ∗ ψε − (∫ ψ)f kLp ,K → 0.

But first we observe that for all h ∈ Lploc we have

Z Z
|h ∗ ψε (x)| ≤ |h(y)ψε (x − y)| dy ≤ |(h1[K]ε )(y)ψε (x − y)| dy (x ∈ K),

so that with Young’s inequality and as kψε kL1 = kψkL1

k(h ∗ ψε )1K kLp ≤ kψkL1 kh1[K]ε kLp = kψkL1 khkLp ,[K]ε . (7.4)
Let δ > 0. There exists a function g (by Lemma A.14) that is continuous on [K]1 and
equals 0 outside [K]1 such that
kf − gkLp ([K]1 ) < δ.
R R R
Then, as |f ∗ ψε − (∫ ψ)f | ≤ |f ∗ ψε − g ∗ ψε | + |g ∗ ψε − ( ψ)g| + |( ψ)g − ( ψ)f |, we
obtain for ε ∈ (0, 1) by using (7.4)
kf ∗ ψε − (∫ ψ)f kLp ,K ≤ k(f − g) ∗ ψε kLp ,K + kg ∗ ψε − (∫ ψ)gkLp ,K
+ kψkL1 kg − f kLp ,K
Z
1
≤ 2δkψkL1 + ( 1K ) p kg ∗ ψε − (∫ ψ)gkL∞ ,K .

50
 ε↓0
As g is continuous on the set U = [K]◦1 , (b) implies kg ∗ ψε − (∫ ψ)gkL∞ ,K −−→ 0, which
in turn implies (c).
(d) follows similarly as (c), because Cc (Rd ) is dense in Lp (Rd ) (Lemma A.14), so that
for all δ > 0 there exists a g ∈ Cc (Rd ) such that kf − gkLp < δ.

8 Convolution of distributions with testfunctions

In this section we consider the convolution of a distribution with a testfunction. As we

have seen in 7.4: For ϕ ∈ D(Rd ) and f ∈ L1loc (Rd ) we have
Z
f ∗ ϕ(y) = f Ty ϕ̌ = uf (Tx ϕ̌) (y ∈ Rd ),

see 7.4. This equality motivates the following generalisation of the notion of convolution
between functions to convolution between distributions and testfunctions:

Definition 8.1. Let (u, ϕ) be in D0 (Rd ) × D(Rd ) or in E 0 (Rd ) × E(Rd ). We define the
convolution of u with ϕ to be the function Rd → F defined by

u ∗ ϕ(x) = u(Tx ϕ̌) (x ∈ Rd ).

It is easy (see Exercise 8.A) to check the following properties.

Lemma 8.2. Let (u, ϕ) be in D0 (Rd ) × D(Rd ) or in E 0 (Rd ) × E(Rd ). Then

δ0 ∗ ϕ = ϕ,
δy ∗ ϕ = Ty ϕ (y ∈ Rd ),
R(u ∗ ϕ) = Ru ∗ Rϕ,
Ty (u ∗ ϕ) = (Ty u) ∗ ϕ = u ∗ (Ty ϕ),
u(ϕ) = u ∗ ϕ̌(0).

Exercise 8.A. Prove Lemma 8.2.

Before we turn to the differentiability of the convolution u ∗ ϕ, in the following lemma

we will show that the convergence of difference quotients of testfunctions hold in the
topology of D.

Lemma 8.3. Let ϕ ∈ D(Rd ), ψ ∈ E(Rd ) and i ∈ {1, . . . , d}. Then

T0 − Thei
 
h→0
ϕ −−−→ ∂i ϕ in D(Rd ), (8.1)
h
T0 − Thei
 
h→0
ψ −−−→ ∂i ψ in E(Rd ). (8.2)
h

51
Consequently, for u ∈ D0 (Rd ) and v ∈ E 0 (Rd ),
T0 − Thei
 
h→0
u −−−→ ∂i u in D0 (Rd ),
h
T0 − Thei
 
h→0
v −−−→ ∂i v in E 0 (Rd ).
h

Proof. Let us first show (8.1). A simple argument by contradiction shows that it is
sufficient to show that
T0 − Thn ei
 
lim ϕ = ∂i ϕ
n→∞ hn
for every sequence (hn )n∈N in R\{0} that converges to 0. Let (hn )n∈N be such a sequence
and, for n ∈ N, put
T0 − Thn ei
 
ψn (x) := ϕ(x) − ∂i ϕ(x) (x ∈ Rd ). (8.3)
hn

We apply Theorem 4.11 to prove ψn → 0 in D(Rd ). That (a) of Theorem 4.11 is

satisfied can be easily seen from the compactness of supp ϕ (see also Lemma 7.11). It
follows from Taylor’s Theorem (see Theorem B.4) that
f (0) − f (−h)
− f 0 (0) ≤ |h|kf 00 kL∞ (h ∈ R \ {0}),
h
for every smooth function f on R for which f 00 is bounded. By choosing f (h) = ∂ α ϕ(x +
hei ), where α ∈ Nd0 and x ∈ Rd , we obtain
∂ α ϕ(x) − ∂ α ϕ(x − hei )
− ∂i ∂ α ϕ(x) ≤ |h|k∂i2 ∂ α ϕkL∞ . (8.4)
h
Therefore, for each α ∈ Nd0 ,
∂ α ϕ(x) − ∂ α ϕ(x − hn ei )
k∂ α ψn kL∞ = sup − ∂i ∂ α ϕ(x) → 0,
x∈Rd hn
which implies (b) of Theorem 4.11.
For (8.2): Let K ⊂ Rd be compact. As the set K1 := K + B(0, 1) is compact,
k∂i2 ∂ α ϕkL∞ (K1 )
< ∞. Also by using Taylor’s Theorem, similarly as how we obtained
(8.4), one obtains
∂ α ψ(x) − ∂ α ψ(x − hei )
− ∂i ∂ α ψ(x) ≤ |h|k∂i2 ∂ α ψkL∞ (K1 ) (h ∈ B(0, 1), x ∈ K).
h
Therefore, for each α ∈ Nd0 ,
T0 − Thei ∂ α ϕ(x) − ∂ α ϕ(x − hei )
  
∂α ϕ − ∂i ϕ = sup − ∂i ∂ α ϕ(x) → 0,
h L∞ (K) x∈K h
 T −T 
0 hei
and thus h ϕ − ∂i ϕ → 0 for all m ∈ N.
C m ,K

52
Theorem 8.4. Let (u, ϕ) be in D0 (Rd ) × D(Rd ) or in E 0 (Rd ) × E(Rd ). Then u ∗ ϕ ∈
C ∞ (Rd ) and
∂ α (u ∗ ϕ) = u ∗ (∂ α ϕ) = (∂ α u) ∗ ϕ (α ∈ Nd0 ). (8.5)

Proof. Let ∈ {1, . . . , d}. It is enough to prove that ∂i (u ∗ ϕ) = u ∗ (∂i ϕ). Let us write
R here for the reflector operator E(Rd ) → E(Rd ), ϕ 7→ ϕ̌. Then ∂i R = −R∂i and
u ∗ ϕ(x) = u(Ty Rϕ) for x ∈ Rd . Then for every x ∈ Rd we obtain by Lemma 8.3
u ∗ ϕ(x + hei ) − u ∗ ϕ(x)
∂i (u ∗ ϕ)(x) = lim
h→0 h
u(Tx+hei Rϕ) − u(Tx Rϕ)
= lim
h→0 h
T−hei ϕ − ϕ
  
= lim u Tx R
h→0 h
T−hei ϕ − ϕ
  
= u Tx R lim
h→0 h
= u(Tx R∂i ϕ) = u ∗ (∂i ϕ)(x).
Moreover, u(Tx R∂i ϕ) = −u(∂i Tx Rϕ) = ∂i u(Tx Rϕ) = (∂i u) ∗ ϕ.

The statement of Theorem 7.10, which states that the support of the convolution of
two functions is included in the closure of the sum of the supports, extends to distributions
(see Theorem 8.6).
8.5 (Convention for the notation “supp”). As we have observed in Definition 7.9, we
have
supp uf = supp f (f ∈ C(Ω)),
supp ug = ess sup g (f ∈ L1loc (Ω).
As we in general do not distinguish between a locally integrable function f and its
corresponding distribution uf , we will also write “supp f ” instead of “supp uf ”, which
then corresponds to “ess sup f ” or if f is continuous with “supp f ” (of course supp f as
in Definition 1.1 also appears in the literature for functions that are not continuous, but
as our focus is on distributions, we assume that no confusion arises).
Theorem 8.6. Let (u, ϕ) be in D0 (Rd ) × D(Rd ) or in E 0 (Rd ) × E(Rd ). Then
supp u ∗ ϕ ⊂ supp u + supp ϕ.
Exercise 8.B. Prove Theorem 8.6.
Theorem 8.7. Convolution is a sequentially continuous operation: The maps
D0 × D → E, (u, ϕ) 7→ u ∗ ϕ, (8.6)
0
E × D → D, (v, ϕ) 7→ v ∗ ϕ, (8.7)
0
E × E → E, (v, ψ) 7→ v ∗ ψ, (8.8)
are sequentially continuous.

53
Proof. We prove sequential continuity of (8.6) and (8.8); the sequential continuity of
(8.7) follows from (8.6) by Theorem 4.11.
Proof of the sequential continuity of (8.6).
Suppose (un , ϕn ) → (u, ϕ) in D0 × D, i.e., un → u in D0 and ϕn → ϕ in D. Let
K ⊂ Rd be compact and m ∈ N0 . We show that

kun ∗ ϕn − u ∗ ϕkC m ,K → 0,

by showing

kun ∗ (ϕn − ϕ)kC m ,K → 0, (8.9)

k(un − u) ∗ ϕkC m ,K → 0. (8.10)

By Theorem 4.11 (a) there exists a compact set L such that the supports of ϕ̌n and ϕ̌
are contained in L for all n ∈ N. Write u∞ = u. As the set K + L is compact (see
Lemma 7.11), by Theorem 4.24 (a), there exist C > 0 and k ∈ N such that

|∂ α un (η)| ≤ CkηkC k (η ∈ DK+L , n ∈ N ∪ {∞}, α ∈ Nd0 , |α| ≤ m + 1). (8.11)

Then

supp Tx ϕ̌n , supp Tx ϕ̌ ⊂ K + L (n ∈ N, x ∈ K),

hence, using Theorem 8.4

kun ∗ (ϕn − ϕ)kC m ,K ≤ sup max |∂ α un (Tx R(ϕn − ϕ))|

x∈K α∈Nd0
|α|≤m

≤ Ckϕn − ϕkC k → 0,

which means (8.9) holds.

To prove (8.10) we use Lemma 4.8. Without loss of generality we may assume K to
be convex (as we can always choose a larger compact convex set). First observe, that
as un → u in D0 , we have ∂ α (un − u) ∗ ϕ(x) → 0 for all x ∈ K and α ∈ Nd0 , |α| ≤ m.
Therefore, by Lemma 4.8, (8.10) follows when (∂ α (un − u) ∗ ϕ)n∈N is a sequence of
uniformly Lipschitz continuous functions on K for all α ∈ Nd0 with |α| ≤ m. For this, by
Lemma 4.7, it is sufficient to show that supn∈N∪{∞} kun ∗ ϕkC m+1 ,K < ∞. But by (8.11)
for all n ∈ N ∪ {∞} we have

kun ∗ ϕkC m+1 ,K ≤ max sup |∂ α un ∗ ϕ(x)|

α∈Nd0 x∈K
|α|≤m

≤ C sup kTx ϕ̌kC k = CkϕkC k .

x∈K

This finishes the proof of the sequential continuity of (8.6).

54
 Proof of the sequential continuity of (8.8). Suppose (un , ϕn ) → (u, ϕ) in E 0 × E. Let
K ⊂ Rd be compact and m ∈ N0 . We show both (8.9) and (8.10). Write u∞ = u. By
Theorem 4.24 (b) there exist a compact set L ⊂ Rd , C > 0 and k ∈ N0 such that

|∂ α un (η)| ≤ CkηkC k ,L (η ∈ E, n ∈ N ∪ {∞}, α ∈ Nd0 , |α| ≤ m). (8.12)

Then

kun ∗ (ϕn − ϕ)kC m ,K ≤ sup max |∂ α un (Tx R(ϕn − ϕ))|

x∈L α∈Nd0
|α|≤m

≤ Ckϕn − ϕkC k ,K+L → 0,

so that (8.9) holds.

To prove (8.10) by the same reasoning (using Lemma 4.8 and Lemma 4.7) it is suffi-
cient to show supn∈N∪{∞} kun ∗ ϕkC m+1 ,K < ∞. By (8.12) we have for all n ∈ N ∪ {∞}
we have

kun ∗ ϕkC m+1 ,K ≤ max sup |∂ α un ∗ ϕ(x)|

α∈Nd0 x∈K
|α|≤m

≤ C sup kTx ϕ̌kC k ,L = CkϕkC k ,K+L .

x∈K

Let ϕ, ψ ∈ D(Rd ) and u ∈ D0 (Rd ). The convolution η = ϕ ∗ ψ is a smooth function

with compact support by Theorem 8.6 (and Lemma 7.11). Hence it is a testfunction and
u ∗ η = u ∗ (ϕ ∗ ψ) is a smooth function. On the other hand, v = u ∗ ϕ is a smooth
function and therefore can be identified with its corresponding distribution, moreover,
v∗ψ = (u∗ϕ)∗ψ is a smooth function. In Theorem 8.9 we will see that these functions are
equal, that is, the convolution obeys an associativity rule. Before, we prove an auxiliary
lemma that considers an approximation of ϕ ∗ ψ. This lemma will be used later on for
example to prove that D and D0 are sequentially separable (in Theorem 8.15). We give
two proofs of Theorem 8.9, one which relies on the Structure Theorem 6.2 and one which
relies on Lemma 8.8 (hence one may postpone reading Lemma 8.8 if one is happy to use
Theorem 6.2).
Lemma 8.8. Let ϕ ∈ D(Rd ). Then there is a sequence (wj )j∈N in D0 consisting of finite
linear combinations of point masses, supported in supp ϕ, such that wj ∗ ψ → ϕ ∗ ψ in
D(Rd ) for all ψ ∈ D(Rd ).

Proof. For j ∈ N, take a locally finite partition of unity (χn )n∈N as in Theorem 1.11 with
diam supp χn < 1j for every n ∈ N, and define
X Z
wj = ( ϕχn )δan ,
n∈N

55
where, for n ∈ N, an ∈ supp χn and if ϕχn 6= 0 then also an ∈ supp ϕ.
By definition we have supp wj ⊂ supp ϕ, so that supp wj ∗ ψ ⊂ supp ϕ + supp ψ.
Therefore, by Theorem 4.11, for ψ ∈ D(Rd ), wj ∗ ψ → ϕ ∗ ψ in D(Rd ) when ∂ α (wj ∗ ψ) →
∂ α (ϕ ∗ ψ) uniformly for all α ∈ Nd0 . As ∂ α (wj ∗ ψ) = wj ∗ (∂ α ψ) and ∂ α (ϕ ∗ ψ) = ϕ ∗ (∂ α ψ)
and ∂ α ψ is a testfunction, it is sufficient to show:
Claim: If ψ ∈ D(Rd ), then wj ∗ ψ → ϕ ∗ ψ uniformly.
1
Let ε > 0. Take j ∈ N such that |ψ(x) − ψ(y)| < ε whenever |x − y| < j. It is
sufficient to show that
Z
|(wj ∗ ψ)(a) − (ϕ ∗ ψ)(a)| ≤ ε |ϕ| (a ∈ Rd ). (8.13)

Fix a ∈ Rd and let η = Ta ψ̌. As wj ∗ ψ(a) = wj (η) and ϕ ∗ ψ(a) =

R
ϕη, instead of (8.13)
we may as well show
Z Z
|wj (η) − ϕη| ≤ ε |ϕ|. (8.14)

Observe that η satisfies, like ψ, |η(x) − η(y)| < ε whenever |x − y| < 1j . We have
Z X Z  Z XZ
wj (η) − ϕη = ϕχn η(an ) − ϕη = ϕχn (η(an ) − η).
n∈N n∈N
R R R
As an ∈ supp χn it follows that |wj (η) − ϕη| ≤ |ϕ|χn ε = ε |ϕ|, which implies
P
n∈N
(8.13).

Theorem 8.9. (a) Let u ∈ D0 and ϕ, ψ ∈ D. Then

(u ∗ ψ) ∗ ϕ = u ∗ (ψ ∗ ϕ). (8.15)

(b) Let v ∈ E 0 , ϕ ∈ D and η ∈ E. Then

v ∗ (ϕ ∗ η) = (v ∗ ϕ) ∗ η = (v ∗ η) ∗ ϕ (8.16)

Proof. (a) It suffices to prove (u ∗ ψ) ∗ ϕ(0) = u ∗ (ψ ∗ ϕ)(0) for any u ∈ D0 , ϕ, ψ ∈ D

(take translations and use the rules in Lemma 8.2). Choose χ ∈ Cc∞ (Rd , [0, 1]) such that
χ = 1 on an open set that contains − supp ψ − supp ϕ. Both ((1 − χ)u ∗ ψ) ∗ ϕ and
((1 − χ)u) ∗ (ψ ∗ ϕ) have supports in supp(1 − χ) + supp ψ + supp ϕ, so both vanish at 0.
Continuation of the proof using Theorem 6.2. χu has compact support, therefore it is
equal to a finite linear combination of derivatives of continuous functions by Theorem 6.2.
As for any continuous function f ∈ C(Ω) and α ∈ Nd0

(∂ α f ∗ ψ) ∗ ϕ = ∂ α ((f ∗ ψ) ∗ ϕ) = ∂ α (f ∗ (ψ ∗ ϕ)) = ∂ α f ∗ (ψ ∗ ϕ),

it follows that (χu ∗ ψ) ∗ ϕ = χu ∗ (ψ ∗ ϕ).

56
 Continuation of the proof using Lemma 8.8 Let (wj )j∈N be as in Lemma 8.8, so that
wj ∗ η → ϕ ∗ η in D(Rd ) for all η ∈ D(Rd ). Write v = χu. As (v ∗ ψ) and ϕ are
testfunctions, we have (v ∗ ψ) ∗ ϕ = ϕ ∗ (v ∗ ψ) (see 7.4(4)). Therefore, by the properties
of (wj )i∈N and the continuity as in Theorem 8.7, we are done if we can show that

wj ∗ (v ∗ ψ) = v ∗ (wj ∗ ψ) (j ∈ N).

But this follows as wj is a linear combination of point measures for each i ∈ N, and

δy ∗ (v ∗ ψ)(x) = Ty (v ∗ ψ)(x) = v(Tx−y ψ̌) = v(Tx (Ty ψ)ˇ) = v ∗ (Ty ψ)(x)

= v ∗ (δy ∗ ψ)(x).

(b) If η ∈ D, then this follows by (a) (and the fact that ϕ ∗ η = η ∗ ϕ). As D is
sequentially dense in E (Theorem 5.10), (8.16) follows by the sequential continuity of the
convolution map E → E, ψ 7→ w ∗ ψ with w being the element in E 0 given by either v,
v ∗ ϕ or ϕ.

8.10. As a direct consequence of Theorem 8.9 we have for u ∈ D0 (Rd ) and ϕ, ψ ∈ D(Rd )

hu ∗ ψ, ϕi = (u ∗ ψ) ∗ ϕ̌(0) = u ∗ (ψ ∗ ϕ̌)(0) = hu, ψ̌ ∗ ϕi,

and for v ∈ E 0 (Rd ) and η ∈ E(Rd )

hv ∗ ψ, ηi = (v ∗ ψ) ∗ η̌(0) = v ∗ (ψ ∗ η̌)(0) = hv, ψ̌ ∗ ηi.

Observe that if u = uf for an f ∈ L1 (Rd ), then the above relation agrees with (f ∗ψ)ϕ =
R
R
f (ψ̌ ∗ ϕ) as in Corollary 7.8.
R
Definition 8.11. Let ψ be a testfunction such that supp ψ ⊂ B(0, 1) and ψ = 1 (the
existence is guaranteed by Lemma 1.6). Such a function is called a mollifier. For a
mollifier ψ and for ε > 0 we define ψε to be the function on Rd defined by

ψε (x) = ε−d ψ( xε ) (x ∈ Rd ).
R
Then supp ψε ⊂ B(0, ε) and ψε = 1 for all ε > 0. For a distribution u we call

u ∗ ψε (8.17)

a mollification of u (with respect to ψ of order ε).

Let ψ be a mollifier. By Theorem 8.4 we know that uε = u ∗ ψε is a smooth function

for all ε > 0. For a function f in Lploc we also know that fε := f ∗ ψε → f in Lploc , by
Theorem 7.15. So in particular,
Z Z
fε ϕ → fϕ (ϕ ∈ D(Rd )),

which implies that fε → f in D0 (Rd ). This “extends” to any distribution, see the following
theorem. This theorem follows by Theorem 7.15.

57
Theorem 8.12. Let ψ be a mollifier, ϕ ∈ D(Rd ), u ∈ D0 (Rd ), η ∈ E(Rd ) and v ∈ E 0 (Rd ).
Then uε = u ∗ ψε ∈ E(Rd ), vε = v ∗ ψε ∈ D(Rd ),

supp uε ⊂ [supp u]ε , supp vε ⊂ [supp v]ε ,

ε↓0 ε↓0
ϕ ∗ ψε −−→ ϕ in D(Rd ), η ∗ ψε −−→ η in E(Rd ),
ε↓0 ε↓0
uε −−→ u in D0 (Rd ) vε −−→ v in E 0 (Rd ).

Proof. supp uε ⊂ [supp u]ε follows from Theorem 8.6.

ε↓0
For the convergence ϕ ∗ ψε −−→ ϕ in D(Rd ) we use Corollary 4.12. As for ε ≤ 1 the
supports supp ϕ ∗ ψε are contained in the compact set [supp ϕ]1 it is sufficient to show
that kϕ ∗ ψε − ϕkC m → 0 for all m ∈ N0 . As ∂ α η ∈ D(Rd ) and ∂ α (η ∗ ψε ) = (∂ α η) ∗ ψ for
all η ∈ D(Rd ), it is sufficient to show that for all ϕ ∈ D(Rd ) ϕ ∗ ψε → ϕ uniformly. But
this follows from Theorem 7.15 (b)
ε↓0
The convergence uε −−→ u in D0 (Rd ) follows from the identity in 8.10 and the con-
vergence ψ̌ε ∗ ϕ → ϕ in D(Rd ) for ϕ ∈ D(Rd ).
ε↓0 ε↓0
The convergences η ∗ ψε −−→ η and vε −−→ v follow in a similar fashion, their proof is
left to the reader.

The next example illustrates that the inclusion supp uε ⊂ [supp u]ε can be a strict
inclusion.

Example 8.13 (supp uε 6= [supp u]ε ). Consider d = 1. By choosing a mollifier ψ (for

example the one of Lemma 1.6) which is supported in (−1, 1) we can define another
mollifier ψ̃ that is supported in (− 34 , − 41 ) ∪ ( 14 , 34 ) as follows:

ψ̃ = 12 (T− 1 ψ 1 + T 1 ψ 1 ).
2 4 2 4

Then δ ∗ ψ̃ equals ψ̃, which is zero around zero. Hence 0 ∈

/ supp δ ∗ ψ̃ but 0 ∈ (supp δ)1 ,
i.e., the inclusion in Theorem 8.12 may be strict.

By using the previous theorem it is immediate that E(Rd ) is sequentially dense in

D0 (Rd ). As D(Rd ) is sequentially dense in E(Rd ), also D(Rd ) is sequentially dense in
D0 (Rd ) as E(Rd ) is continuously embedded in D0 (Rd ) (see Theorem 5.10, that E(Rd ) is
continuously embedded in D0 (Rd ) follows for example by Theorem 4.28). Moreover, we
have the following:

Lemma 8.14. Let Ω ⊂ Rd be open. D(Ω) is sequentially dense in D0 (Ω) and in E 0 (Ω).

Proof. We will show that there exist (ϕε )ε>0 in D(Ω) that converge to u in D0 (Ω). As the
compactly supported distributions are sequentially dense in D0 (Ω) (see Theorem 5.10),
we may assume u has compact support. Let χ ∈ D(Ω) be equal to 1 on a neighbourhood
of supp u, so that χu = u. Let u ∈ D0 (Rd ) be the distribution given by u(ϕ) = u(χϕ|Ω ).

58
Observe supp u = supp u. Let ψ be a mollifier. Then u ∗ ψε → u in D0 (Rd ) by The-
orem 8.12. Moreover u ∗ ψε ∈ D(Rd ) and there exists a δ > 0 such that [supp u]δ ⊂ Ω,
and thus supp(u ∗ ψε ) ⊂ Ω for all ε ∈ (0, δ). Therefore ϕε := u ∗ ψε |Ω is an element of
D(Ω) and by continuity of the embedding ρ : D0 (Rd ) → D0 (Ω) as in 4.6 is follows that
ϕε → u in D0 (Ω).
That D(Ω) is sequentially dense in E 0 (Ω) follows in a similar fashion, this proof is left
to the reader.

Theorem 8.15. Let Ω ⊂ Rd be open. Then D(Ω), D0 (Ω) and E 0 (Ω) are sequentially
separable.

Proof. By Lemma 8.14 it is sufficient to show that D(Ω) is sequentially separable.

We invoke Lemma 8.8. Let ψ ∈ D(Rd ) be a mollifier. By Theorem 8.12 we know that
the set of functions

{ϕ ∗ ψε : ϕ ∈ D, ε > 0, [supp ϕ]ε ⊂ Ω} (8.18)

is dense in D(Ω). Let A ⊂ D(Ω) be the collection of functions of the form

m
X
λk δxk ∗ ψε
k=1

for some m ∈ N, λk ∈ Q + iQ, xk ∈ Qd ∩ Ω for all k ∈ {1, . . . , m} and with ε ∈ (0, ∞) ∩ Q

such that

[{x1 , . . . , xm }]ε ⊂ Ω.

Then A is a countable set. By Lemma 8.8 it follows that A is dense in (8.18) and thus
in D(Ω).

Similarly to the proof that D(Ω) is dense in D0 (Ω) we can (and will) prove that D(Ω)
is dense in Lp (Ω) for finite p. For this we use the following lemma to show that the
compactly supported functions in Lp (Ω) are dense in Lp (Ω). For this reason we do not
need to include the “∂ α ” in the lemma, but we do so as this will be used later on.
Lemma 8.16. Let p ∈ [1, ∞). Let χ ∈ D(Rd ) be equal to 1 a neighbourhood of 0. For
R > 0 write χR = l 1 χ = χ( R1 ·). Then
R

ku∂ α (χR − 1)kLp −−−−→ 0

R→∞
(α ∈ Nd0 , u ∈ Lp (Rd )).

Proof. Observe that for R ≥ 1,

k∂ α (χR − 1)kL∞ ≤ 1 + R−|α| k∂ α χkL∞ ≤ 1 + k∂ α χkL∞ .

Therefore the convergence follows by Lebesgue’s dominated convergence theorem, be-

cause ∂ α (χR − 1) converges pointwise to 0 as R → ∞.

59
Theorem 8.17. Let p ∈ [1, ∞). D(Ω) is dense in Lp (Ω). Consequently, Lp (Ω) is
separable.

Proof. By Lemma 8.16 it is sufficient to show that for all f ∈ Lp (Ω) with compact support
ε↓0
there exist fε ∈ D(Ω) for ε > 0 with fε −−→ f in Lp (Ω). Let f ∈ Lp (Ω) be compactly
supported and let ψ be a mollifier. Let f be the function in Lp (Rd ) that equals f on Ω
and equals 0 elsewhere. Then f ∗ ψε ∈ D(Ω) for all ε > 0. As in the proof of Lemma 8.14,
let δ > 0 be such that [supp f ]δ ⊂ Ω and thus supp f ∗ ψε ⊂ Ω for all ε ∈ (0, δ). By
Theorem 8.12 f ∗ ψε ∈ D(Rd ) and thus fε := (f ∗ ψε )|Ω ∈ D(Ω). By Theorem 7.15 (c)
we have f ∗ ψε → f in Lp (Rd ) and therefore conclude fε → f in Lp (Ω).
As D(Ω) is separable (Theorem 8.15) and is continuously embedded in Lp (Ω) (The-
orem 4.28), it follows that Lp (Ω) is separable.

Exercise 8.C. Let ψ ∈ D(Rd ) and F ⊂ Rd be a closed set. Show that ψ ∗ 1F is a

smooth function and that all of its derivatives are bounded.
Exercise 8.D. For each of the following cases, find u ∈ D0 (R) and ϕ ∈ D(R) such that:

(a) u ∗ ϕ(x) = 0 for all x ∈ R, but u 6= 0 and ϕ 6= 0,

(b) u ∗ ϕ(x) = 1 for all x ∈ R,
(c) u ∗ ϕ(x) = x for all x ∈ R,
(d) u ∗ ϕ(x) = sin x for all x ∈ R.

Exercise 8.E. Consider the distribution on R given by h = 1[0,∞) , also called the
Heaviside function. For ϕ ∈ D(R) calculate h ∗ ϕ0 , where ϕ0 denotes the derivative of
ϕ. Calculate the derivative of the distribution corresponding to h, i.e., calculate ∂uh .
Validate by these calculations that (h ∗ ϕ)0 = h ∗ ϕ0 = ∂uh ∗ ϕ.

9 Distributions of finite order

In this section we prove additional properties of distributions of finite order. First, we

have seen that for a distribution u and a testfunction ϕ one has u(ϕ) = 0 as soon as
ϕ = 0 on a neighbourhood of the support of u. And that in general, the condition ϕ = 0
on supp u does not imply u(ϕ) = 0, see Definition 5.1. In Theorem 9.4 we will see that
if a distribution u is of order k, then

∂ α ψ = 0 on supp u for all α ∈ N0 with |α| ≤ k,

implies u(ψ) = 0. With this theorem we prove that a distribution supported in a point
is given by a linear combination of derivatives of the point measure at the supporting
point, in Theorem 9.5. Moreover, we show that distributions of order k can be extended
to continuous linear functions C0k (Ω) → F and with that prove Theorem 2.28.
First we consider some auxiliary lemmas to prove Theorem 9.4.

60
Lemma 9.1. Let ϕ ∈ D(Ω), a ∈ Ω and k ∈ N0 . Suppose that that ∂ α ϕ(a) = 0 for all
α ∈ Nd0 with |α| ≤ k. Then, for all ε > 0 with B(a, ε) ⊂ Ω,
X
|∂ α ϕ(x)| ≤ k∂ β ϕkL∞ εk+1−|α| (x ∈ B(a, ε), α ∈ Nd0 , |α| ≤ k). (9.1)
β∈Nd0 :|β|≤k+1

Proof. By Taylor’s formula (see Theorem B.7) we know that ϕ equals its remainder of
order k at a as its Taylor polynomial of order k at a equals zero. Hence by (B.1) with
l = k + 1 we see that for M = β∈Nd :|β|≤k+1 k∂ β ϕkL∞
P
0

|ϕ(x)| ≤ M |x − a|k+1 ≤ M εk+1 (x ∈ B(a, ε)).

By a repetition of the above argument for the derivatives of ϕ, we obtain (9.1).

Definition 9.2. For k ∈ N0 we define Cbk (Ω) to be the space

Cbk (Ω) = {f ∈ C k (Ω) : kf kC k < ∞}, (9.2)

equipped with the norm k · kC k . Thus Cbk (Ω) is the space of smooth functions whose
derivatives up to order k are bounded.
Cb∞ (Ω) is the space

Cb∞ (Ω) =
\
Cbk (Ω), (9.3)
k∈N

equipped with the seminorms k · kC k for k ∈ N0 .

For k ∈ N0 , Cbk (Ω) is a Banach space and Cb∞ (Ω) is a Fréchet space; this follows from
Theorem 4.19 (see Exercise 9.A). Moreover, as for any compact set K ⊂ Ω and k ∈ N0

kf kC k ,K ≤ kf kC k (f ∈ Cb∞ (Ω)),

Cb∞ (Ω) is continuously embedded in E(Ω);

Cb∞ (Ω) ,→ E(Ω).

Exercise 9.A. Prove that Cbk (Ω) is a Banach space for k ∈ N0 and Cb∞ (Ω) a Fréchet
space.

In 5.6 we have used that for any compact set we can find a testfunction that equals
1 on that compact set. Now with the tools of convolution, we can also construct such
functions by a convolution of a mollifier. Moreover, for any closed set F we can also find
smooth functions which are equal to 1 on a neighbourhood of F with a control on the
growth of the derivatives:

61
Lemma 9.3. Let F ⊂ Rd be a closed set. For each ε > 0 there exists a ηε ∈ Cb∞ (Rd )
such that

ηε = 1 on [F ]ε , supp ηε ⊂ [F ]3ε , (9.4)

and for each k ∈ N there exists a C > 0 such that

k∂ α ηε kL∞ ≤ Cε−|α| (α ∈ Nd0 , |α| ≤ k, ε > 0). (9.5)

Consequently, if F is a compact set in Ω, then for each ε > 0 such that [F ]3ε ⊂ Ω, there
exists a ηε ∈ D(Ω) with (9.4) and (9.5).

Proof. Let ψ be a positive mollifier. For ε > 0 define ηε = 1[F ]2ε ∗ ψε . Then supp ηε ⊂
[F ]3ε . As 1[F ]2ε (x − y) = 1 for all y ∈ B(0, ε) and x ∈ [F ]ε , and as ψε integrates to 1 on
B(0, ε) and has support in B(0, ε):
Z Z
ηε (x) = 1[F ]2ε (x − y)ψε (y) dy = ψε (y) dy = 1 (x ∈ [F ]ε ),
B(0,ε)

i.e., ηε = 1 on [F ]ε . Let α ∈ Nd0 . We have ∂ α ηε = 1[F ]2ε ∗ ∂ α ψε and ψε (x) = ε−d ψ( xε ) for
x ∈ Rd . Therefore, by Young’s inequality, for C = α∈Nd :|α|≤k k∂ α ψkL1
P
0

k∂ α ηε kL∞ ≤ k1[F ]2ε kL∞ k∂ α ψε kL1 ≤ Cε−|α| .

Theorem 9.4. Let k ∈ N0 and u ∈ D0 (Ω) be a distribution of order k. Suppose that

ψ ∈ D(Ω) and

∂ α ψ = 0 on supp u for all α ∈ N0 with |α| ≤ k. (9.6)

Then u(ψ) = 0.

Proof. Let us first show that we may assume that u is compactly supported. In order to
do that we assume that (χn )n∈N is a partition of unity as in Theorem 1.11. Then, for any
ϕ ∈ D(Ω), u(ϕ) equals the sum over χn u(ϕ), for which only finitely many are possibly
nonzero. As supp χn u ⊂ supp u, it suffices to show the statement for “χn u” instead of
“u”.
Instead, we will assume that u is compactly supported. Let F = supp u. Let C > 0
be such that

|u(ϕ)| ≤ CkϕkC k (ϕ ∈ D(Ω)).

Let ε > 0 be such that [F ]3ε ⊂ Ω. By Lemma 9.3 there exists a ηε ∈ D(Ω) with (9.4)
and (9.5). Then u = ηε u and thus

|u(ϕ)| ≤ Ckηε ϕkC k (ϕ ∈ D(Ω)).

62
Let ψ ∈ D(Ω) be such that (9.6). By the above inequality it suffices to show kηε ψkC k → 0
as ε ↓ 0. By Leibniz’ rule 1.14, for x ∈ [F ]3ε and α ∈ Nd0 , |α| ≤ k,
Let M = β∈Nd :|β|≤k+1 k∂ β ψk∞ and C > 0 be as in (9.5). Then by Leibniz’ rule 1.14
P
0
we obtain for all α ∈ Nd0 with |α| ≤ k and x ∈ [F ]3ε
!
α
X α
|∂ (ηε ψ)(x)| ≤ |∂ β ηε (x)||∂ α−β ψ(x)|
β
β∈Nd0
β≤α
! !
α  X α 
M ε−|β| Cεk+1−|α−β| ≤ M C
X
≤ ε.
β β
β∈Nd0 d β∈N0
β≤α β≤α

ε↓0
Therefore kηε ψkC k −−→ 0 and thus u(ψ) = 0.

Theorem 9.5. If u is a distribution supported by {x}, then there exist a k ∈ N0 and

cα ∈ F for α ∈ Nd0 , |α| ≤ k such that
X
u= cα ∂ α δx .
α∈Nd0 :|α|≤k

Moreover, cα = hι−1 (u), xα i, where x : x 7→ x (and thus xα : x 7→ xα ) and with

ι : E 0 (Ω) → D(Ω) as in Theorem 5.10.

Proof. By taking a translation of the distribution, we may as well assume that x = 0.

Let ε > 0 be such that B(0, ε) ⊂ Ω. Let k ∈ N0 be the order of u. By Taylor’s formula
(see Theorem B.7) ϕ = P + ψ on B(0, ε), for a polynomial P of order k given by
X 1
P (x) = ∂ α ϕ(0)xα (x ∈ Rd ),
α!
α∈Nd0
|α|≤k

and ψ satisfying ∂ α ψ(0) = 0 for all α ∈ Nd0 with |α| ≤ k. Let χ be a testfunction that
equals 1 on B(0, 2ε ) and has support within B(0, ε). Then u(ϕ) = u(χϕ) = u(P χ) by the
previous theorem. And thus,
X 1
u(ϕ) = u(P χ) = ∂ α ϕ(0)u(xα χ).
α!
α∈Nd0
|α|≤k

Theorem 9.6. Let k ∈ N0 . Let u ∈ D0 (Ω) be of order k. Then there exists exactly one
linear extension of u, v : Cck (Ω) → F which is continuous with respect to k · kC k , i.e.,
v = u on D(Ω) and there exists a C > 0 such that

|v(f )| ≤ Ckf kC k (f ∈ Cck (Ω)). (9.7)

63
Proof. Let C > 0 be such that

|u(ϕ)| ≤ CkϕkC k (ϕ ∈ D(Ω)).

Let ψ be a mollifier. By Theorem 7.15 (b) follows that ∂ α f ∗ ψε → ∂ α f uniformly for all
f ∈ Cck (Ω) and α ∈ Nd0 , |α| ≤ k. Therefore
ε↓0
kf ∗ ψε − f kC k −−→ 0 (f ∈ Cck (Ω)).

Therefore, (u(f ∗ ψn−1 ))n∈N is a Cauchy sequence in F for all ψ ∈ Cck (Ω). We define
v : Cck (Ω) → F by

v(f ) = lim u(f ∗ ψn−1 ) (f ∈ Cck (Ω)).

n→∞

Then v is linear, v = u on D(Ω) and for f ∈ Cck (Ω) we have for all n ∈ N

|v(f )| ≤ |v(f − f ∗ ψn−1 )| + |u(f ∗ ψn−1 )| ≤ |v(f ) − u(f ∗ ψn−1 )| + Ckf ∗ ψn−1 kC k ,

so by taking the limit n → ∞, we obtain (9.7). The uniqueness follows by the continuity.

Definition 9.7. For k ∈ N0 ∪{∞} we define C0k (Ω) to be the space of functions f ∈ C k (Ω)
with ∂ α f ∈ C0 (Ω) for all α ∈ Nd0 , |α| ≤ k (C0 (Ω) is defined in Definition 2.29). For
k ∈ N0 the space C0k (Ω) is equipped with the norm k · kC k and C0∞ (Ω) is equipped with
the seminorms k · kC k for k ∈ N0 .
For k ∈ N0 , C0k (Ω) is a Banach space and C0∞ (Ω) is a Fréchet space, and D(Ω) is
sequentially dense in C0k (Ω) for each k ∈ N0 ∪ {∞} (see Exercise 9.B).

Exercise 9.B. Let k ∈ N0 . Prove:

(a) C0k (Ω) is a Banach space,

(b) C0∞ (Ω) a Fréchet space,
(c) D(Ω) is dense in C0k (Ω) and in C0∞ (Ω),
(d) For each linear v : Cck (Ω) → F that is continuous with respect to k · kC k there exists
exactly one continuous extension of v, w : C0k (Ω) → F.

Corollary 9.8. Let k ∈ N0 . Let u ∈ D0 (Ω) be of order k. Then there exists exactly one
continuous linear extension of u, v : C0k (Ω) → F.

Proof. This follows by Theorem 9.6 and the fact that Cck (Ω) is dense in C0k (Ω) (see also
Exercise 9.B).

We are ready to give the proof of Theorem 2.28:

Theorem 2.28. A distribution u is of order 0 if and only if u = uµ for a µ ∈ M(Ω, F).

64
Proof. The “if” statement is trivial. Let u ∈ D0 (Ω) be of order 0. By Corollary 9.8
u extends to a continuous linear C0 (Ω, F), which by the Riesz representation theorem
(Theorem 2.30) is given by a µ ∈ M(Ω, F). It follows that u = uµ .

9.9 (The distributional structure of M(Ω, F)). Observe that if µ ∈ M(Ω, F) and µ has
compact support, in the sense that supp uµ is compact, then by Theorem 6.1 there exists
a continuous function f ∈ C(Ω) such that for α = (2, . . . , 2), uµ = ∂uf .
Moreover, by Theorem 6.3, for any µ ∈ M(Ω, F) there exist continuous functions
gα ∈ C(Ω) for α ∈ Nd0 with α ≤ (2, . . . , 2) such that uµ = α∈Nd :α≤(2,...,2) ∂ α ugα .
P
0

For d = 1 this means that any Radon measure equals (as a distribution) the sum of
ug0 + ∂ug1 + ∂ 2 ug2 , for some continuous functions g0 , g1 and g2 .

Question 9.1. Like for the structure theorems, can any distribution u of order k be
written as α∈Nd ,|α|≤k ∂ α uµα for some µα ∈ M(Ω, F)?
P
0

10 Convolutions of distributions

In this section we consider convolution as an operation between distributions. Like

for locally integrable functions, one cannot expect to be able to define the convolution
between any two distributions; consider for example the function 1, of which convolution
with itself does not exist (in the sense of Definition 7.3). But when one considers two
distributions of which one has compact support, one can define a convolution between
them as we will see. Observe that the convolution between two locally integrable functions
of which at least one has compact support exists in the sense of Definition 7.3.
In this section we consider only Ω = Rd and write ‘E’ and ‘D’ instead of ‘E(Rd )’ and
‘D(Rd )’. First, we start by characterising the operation of convoluting with a distribution
as a sequentially continuous map D → E that commutes with translation.

Definition 10.1. We say that a linear function A : D → E or A : E → E commutes

with translations if it commutes with the translation operators, i.e., if Tx A = ATx for all
x ∈ Rd .

Theorem 10.2.

(a) Let A : D → E be linear. Then A is sequentially continuous and commutes with

translations if and only if there exists a u ∈ D0 such that Aϕ = u ∗ ϕ for all ϕ ∈ D.
(b) Let A : E → E be linear. Then A is sequentially continuous and commutes with
translations if and only if there exists a u ∈ E 0 such that Aϕ = u ∗ ϕ for all ϕ ∈ D.

For both (a) and (b), if A is sequentially continuous and commutes with translations,
then there exists exactly one such u such that Aϕ = u ∗ ϕ for all ϕ ∈ D.

65
Proof. For both (a) and (b) the “if” statement, follows from the fact that convolutions
commutes with translations as we have seen in Lemma 8.2 and from the fact that convo-
lution as an operation is sequentially continuous Theorem 8.7. Therefore we assume A
to be sequentially continuous and to commute with translations.
(a) Define u : D → F by u(ϕ) = Aϕ̌(0) for ϕ ∈ D. u is linear and sequentially
continuous, therefore it is a distribution (Theorem 4.14). Then for every x ∈ Rd

Aϕ(x) = T−x Aϕ(0) = A[T−x ϕ](0) = u([T−x ϕ]ˇ) = u(Tx ϕ̌) = u ∗ ϕ(x). (10.1)

(b) As in (a), define u : E → F by u(ϕ) = Aϕ̌(0) for ϕ ∈ E. Then u linear is

sequentially continuous, hence continuous as E is a metric space, which implies u ∈ E 0 .
That Aϕ = u ∗ ϕ follows by (10.1).

10.3. Let v ∈ E 0 and u ∈ D0 . As u∗ϕ ∈ E for all ϕ ∈ D, we can compose the maps D → E,
ϕ 7→ u∗ϕ and E → E, ψ 7→ v ∗ψ. By Lemma 8.2 and Theorem 8.7 this composition forms
a sequentially continuous linear map D → E that commutes with translations. Therefore,
by Theorem 10.2 there exists a w ∈ D0 such that

w ∗ ϕ = u ∗ (v ∗ ϕ) (ϕ ∈ D).

We could take this w as our definition of u ∗ v; instead we define u ∗ v via another formula
and show in Theorem 10.6 that it equals w.

10.4. Remember (7.1) in Corollary 7.8, which tells us that for integrable Rf , g and a
testfunction ϕ we have (by viewing f ∗ g as a distribution and thus hf, ϕi = f ϕ)

hg ∗ f, ϕi = hf ∗ g, ϕi = hg, fˇ ∗ ϕi.

Moreover, if (u, ψ) ∈ D0 × D or (u, ψ) ∈ E 0 × E and ϕ ∈ D, then by the associativity

property (Theorem 8.9)

hu ∗ ψ, ϕi = (u ∗ ψ) ∗ ϕ̌(0) = u ∗ (ψ ∗ ϕ̌)(0) = hu, ψ̌ ∗ ϕi.

This identity motivates the definition of the convolution between a distribution and a
compactly supported distribution. First we make the following observation. For v ∈ E 0
we know by Theorem 8.7 that D → D, ϕ 7→ v̌∗ϕ is a sequentially continuous map. Hence,
if v ∈ E 0 and u ∈ D0 , then D → F, ϕ 7→ u(v̌ ∗ ϕ) is a distribution as it is sequentially
continuous and linear (see Theorem 4.14).

Definition 10.5. For u ∈ D0 and v ∈ E 0 we define u ∗ v to be the distribution given by

u ∗ v(ϕ) = u(v̌ ∗ ϕ) (ϕ ∈ D).

Moreover, we define v ∗ u to be the distribution

v ∗ u(ϕ) = v(ǔ ∗ ϕ) (ϕ ∈ D).

66
Theorem 10.6. Let u ∈ D0 , v ∈ E 0 and ϕ ∈ D. Then
(u ∗ v) ∗ ϕ = u ∗ (v ∗ ϕ) = v ∗ (u ∗ ϕ) = (v ∗ u) ∗ ϕ. (10.2)
Consequently, u ∗ v = v ∗ u.

Proof. The first equality in (10.2), and similarly the last one by interchanging the roles
of u and v, follows from the observation that for all x ∈ Rd
(u ∗ v) ∗ ϕ(x) = u ∗ v(Tx ϕ̌) = u(v̌ ∗ Tx ϕ̌) = u(Tx (v ∗ ϕ)ˇ) = u ∗ (v ∗ ϕ)(x).
We are therefore left to prove
u ∗ (v ∗ ϕ) = v ∗ (u ∗ ϕ). (10.3)
By Theorem 8.9 (a) and by the commutativity of convolution on functions (see 7.4), for
ψ ∈ D we have (10.2) with “ψ” instead of “v”:
u ∗ (ψ ∗ ϕ) = ψ ∗ (u ∗ ϕ).
By the fact that D is sequentially dense in E 0 (Theorem 5.10) and by using the sequential
continuity of Theorem 8.7 we obtain (10.3) and thus (10.2).

If u, v ∈ E 0 , then the map E → F, ψ 7→ u(v̌ ∗ ψ) is an element of E 0 as it is se-

quentially continuous and linear (that it is continuous follows similarly to the proof of
Theorem 4.14).
Definition 10.7. For u, v ∈ E 0 we define u ∗ v ∈ E 0 by the formula
u ∗ v(ψ) = u(v̌ ∗ ψ) (ψ ∈ E).
Lemma 10.8. Let ι be the embedding E 0 (Ω) → D0 (Ω) as in Theorem 5.10. Then
ι(u ∗ v) = ι(u) ∗ v = u ∗ ι(v) (u, v ∈ E 0 ).

Proof. The proof is rather straightforward and left to the reader (see Exercise 10.A).

Exercise 10.A. Verify the statement in 10.7.

10.9 (Convention). As we have the one-to-one correspondence of elements of E 0 and
compactly supported distributions, the identity Lemma 10.8 and that supp v = supp(v|D )
for all v ∈ E 0 (Definition 5.12), for v ∈ E 0 we will often write “v” instead of “ι(v)”.
Similarly, for u, v ∈ D0 with v compactly supported we will write u ∗ v for u ∗ ι−1 (v), etc.
Lemma 10.10. Let (u, v) ∈ (E 0 × D0 ) ∪ (D0 × E 0 ). Then
δ0 ∗ u = u,
δy ∗ u = Ty u (y ∈ Rd ),
R(u ∗ v) = R(u) ∗ R(v),
Ty (u ∗ v) = (Ty u) ∗ v = u ∗ (Ty v) (y ∈ Rd ),
∂ α (u ∗ v) = (∂ α u) ∗ v = u ∗ (∂ α v) (α ∈ Nd0 ).

67
Proof. The proof is left for the reader.

Theorem 10.11. For u ∈ D0 and v ∈ E 0

supp u ∗ v ⊂ supp u + supp v.

Proof. Let x ∈ supp u ∗ v. For all ε > 0 there exists a ϕ ∈ D supported in B(x, ε) such
that u ∗ v(ϕ) 6= 0, i.e., u(v̌ ∗ ϕ) 6= 0. Therefore supp u ∩ (supp v̌ ∗ ϕ) 6= ∅. Let y be in this
intersection. By Theorem 8.6 we know that there exists a z ∈ supp v and w ∈ supp ϕ
such that y = −z + w. Then w = y + z ∈ supp u + supp v and |x − w| < ε. As we can find
such w for each ε and supp u + supp v is closed, we conclude that x ∈ supp u + supp v.

Remark 10.12. One can also define the convolution of two distributions, where instead
of assuming that one of the two has compact support the map Σ : Rd × Rd → Rd ,
Σ(x, y) = x + y is proper on supp u × supp v, meaning that Σ−1 (K) ∩ supp u × supp v is
a compact subset of Rd × Rd for all compact sets K ⊂ Rd . The details can be found for
example in [DK10, Section 11] .

11 Fundamental solutions of partial differential operators

In this section we consider partial differential operators and corresponding fundamental

solutions.

Definition 11.1. We call a map P : D0 (Ω) → D0 (Ω) a linear partial differential operator
with constant coefficients if there exist an m ∈ N and cα ∈ F for α ∈ Nd0 with |α| ≤ m
such that
X
P = cα ∂ α .
α∈Nd0
|α|≤m

Often, the following notation is also used. When we take p : Rd → F the polynomial
X
p(x) = cα xα ,
α∈Nd0
|α|≤m

then it is common to write “p(∂)” for “P ”, so that one interpret p(∂) as the formal
polynomial (i.e., finite formal powerseries) evaluated at ∂. One also uses “D” instead of
“∂” in literature, so that one writes “p(D)” for “P ”. We will not use “D” in this context
because we will use this in the context of Fourier multipliers in Section 19.
A distribution E is called a fundamental solution to P if P E = δ, where δ is the
Dirac measure at zero.

68
Theorem 11.2. Let P be a linear partial differential operator with constant coefficients
and E a fundamental solution to P . For all v ∈ E 0 (Rd ) we have

P (E ∗ v) = v = E ∗ (P v).

Proof. This follows by the fact that ∂ α (E ∗ v) = (∂ α E) ∗ v = E ∗ (∂ α v).

11.3. Observe that if E is a fundamental solution to a linear partial differential operator

with constant coefficients P , and if u ∈ D0 (Rd ) satisfies P u = 0, then E + u is also a
fundamental solution to P .
11.4. Let P , E be as in Theorem 11.2 and v ∈ E 0 (Rd ). One says that u = E ∗ v is a
solution to the partial differential equation

P u = v. (11.1)

Hence, by Theorem 11.2 one can derive solutions of partial differential equations of the
form (11.1) when one knowns a fundamental solution to P .
Definition 11.5 (Laplacian). We write ∆ for the linear partial differential operator
d
X
∆= ∂i2 ,
i=1

and call it the Laplacian.

Example 11.6 (Fundamental solution to ∆). Let E be the function on Rd (for d ≥ 2)
defined by E(0) = 0 and

1 2−d
(2−d)Vd |x| d 6= 2,

E(x) = (11.2)
 1 log |x| d = 2,
2π

where Vd is the d − 1 dimensional volume of the sphere {x ∈ Rd : |x| = 1} (observe that

2π = V2 ). Then E is a fundamental solution to ∆ (see Exercise 11.A).
Exercise 11.A. (a) For i ∈ {1, . . . , d} let vi be the function on Rd defined by vi (0) =
0 and
xi
vi (x) = (x ∈ Rd \ {0}).
|x|d

Prove that vi is locally integrable on Rd and that in D0

d
X
∂i vi = Vd δ,
i=1

where Vd is the d − 1 dimensional volume

R
of the sphere {x ∈ Rd : |x| = 1}. (Hint:
Observe that h∂i vi , ϕi = − limε↓0 Rd \B(0,ε) vi ∂i ϕ and apply integration by parts
(see Theorem C.1).)

69
 (b) Prove that E as in (11.2) is locally integrable on Rd and that E is a fundamental
solution to ∆, i.e., ∆E = δ (first you might want to prove that ∂i E = cvi for some
c ∈ R).
11.7. With E being a fundamental solution to ∆ as defined in (11.2), we conclude that
for v ∈ E 0 (Rd ) we have a solution to the Poisson equation

∆u = v,

given by u = E ∗ v ∈ D0 (Rd ).
Definition 11.8. A function f ∈ C 2 (Ω) is called harmonic, or an harmonic function if
∆f = 0. A distribution u ∈ D0 (Ω) is called harmonic if ∆u = 0.
Exercise 11.B. For F = C and d ≥ 2, check that for all k ∈ N0 the polynomial
x 7→ (x1 + ix2 )k is harmonic.
Observe that if Ω is connected and f ∈ C 2 (R) (i.e., d = 1), then ∆f = f 00 = 0 if and
only if f (x) = a + bx for some a, b ∈ F.

As is mentioned in 11.3, if u is a harmonic distribution, then E + u is a fundamental

solution to ∆. Wehl’s theorem, see Theorem 11.14, states that each harmonic distribution
is actually (represented by) a harmonic function in C ∞ (Rd ). We prove this theorem by
proving a more general result, Theorem 11.12, which is about singular supports; this
support indicates “where a distribution is smooth”.
Definition 11.9 (Singular support). Let u ∈ D0 (Ω). If UR ⊂ Ω is open we say that u
is smooth on U if there exists an f ∈ C ∞ (U ) with u(ϕ) = U f ϕ for all ϕ ∈ D(U ). Let
U be the collection of all open subsets of Ω on which u is smooth. By Theorem 1.11
u is smooth on U, as u = n∈N χn u for some partition of unity (χn )n∈N subordinate
S P

to U, so that χn u is (represented by) a smooth function with compact support. The

complement of U is defined to be the singular support of u,
S

sing supp u.

Observe that if U ⊂ Ω is open and u vanishes on U , then u is smooth on U . Con-

sequently

sing supp u ⊂ supp u (u ∈ D0 ),

If U ⊂ Ω is open and u, v ∈ D0 are both smooth on U , then u + v is smooth on U .

Consequently

sing supp(u + v) ⊂ sing supp u ∪ sing supp v (u, v ∈ D0 ),

Let χ ∈ Cc∞ (Ω, [0, 1]) and U ⊂ Ω be open. Then u is smooth on U if and only if χu and
(1 − χ)u are smooth on U . Consequently

sing supp u = sing supp(χu) ∪ sing supp((1 − χ)u).

70
The above extends in the following sense to a partition of unity (χn )n∈N as in The-
orem 1.11, due to the fact that the sets {x ∈ Ω : χn (x) > 0} form a locally finite cover;
[
sing supp u = sing supp(χn u). (11.3)
n∈N

The singular support satisfies the same rule as the support does for convolutions:

Lemma 11.10. Let u ∈ D0 (Rd ) and v ∈ E 0 (Rd ). Then

sing supp u ∗ v ⊂ sing supp u + sing supp v. (11.4)

Proof. Let us write A for sing supp u and B for sing supp v. Let δ > 0. By Lemma 9.3
there exists a χ ∈ C ∞ such that χ is equal to 1 on [A] δ and 0 outside [A]δ . Then
2
u2 := (1 − χ)u is (represented by) a smooth function and so u = u1 + u2 for u1 = χu,
and supp u1 ⊂ [A]δ . Similarly, we can write v = v1 + v2 , where supp v1 ⊂ [B]δ and v2 is
(represented by) a smooth function. Then

u ∗ v = u1 ∗ v1 + u1 ∗ v2 + u2 ∗ v1 + u2 ∗ v2 .

The last three terms are smooth (by Theorem 8.4) and the support of u1 ∗ v1 is included
in [A]δ + [B]δ (Theorem 10.11), which in turn is included in [A + B]2δ . Therefore

sing supp u ∗ v ⊂ [A + B]2δ .

Now observe that B is compact as it is a subset of the support of v. As δ is chosen

T
arbitrarily and the set A + B is closed (see Lemma 7.11), we have δ>0 [A + B]2δ = A + B
and conclude (11.4).

Definition 11.11. Let P be a linear partial differential operator with constant coeffi-
cients. A distribution E is called a parametrix of P if there exists a ψ ∈ E(Rd ) such that
P E = δ + ψ.
Observe that any fundamental solution to P is a parametrix of P .

Theorem 11.12. Let P be a linear partial differential operator with constant coefficients.
Suppose E is a parametrix of P with sing supp E = {0}. Then for all open Ω ⊂ Rd

sing supp u = sing supp P u (u ∈ D0 (Ω)). (11.5)

Proof. Similarly to 8.5 we have sing supp P u ⊂ sing supp u, which basically means that
‘P u is smooth where u is’.
By (11.3) we may assume that u has compact support, so that we may as well assume
that u ∈ D0 (Ω). Let ψ ∈ E(Rd ) be such that P E = δ + ψ. Then

E ∗ (P u) = (P E) ∗ u = (δ + ψ) ∗ u = u + ψ ∗ u.

71
Therefore sing supp u = sing supp E ∗ (P u) as ψ ∗ u ∈ E(Rd ) (Theorem 8.4). Therefore,
by Lemma 11.10

sing supp u ⊂ sing supp E + sing supp P u = sing supp P u,

as sing supp E = {0}.

11.13. Let P and E are as in Theorem 11.12. This theorem tells us that a solution u
to P u = v for a v ∈ E 0 (Rd ) is smooth where v is, in the sense that if U is open and v is
smooth on U , then u is smooth on U . Therefore, in particular we obtain Weyl’s theorem
as a consequence.

Theorem 11.14 (Weyl’s Theorem). Every harmonic distribution is (represented by) a

smooth harmonic function. Moreover, if u ∈ D0 is such that ∆u = ψ for a ψ ∈ E, then
u is a smooth function.

Example 11.15. For t > 0 we define the function ht : Rd → R by

d 1 2
ht (x) = (4πt)− 2 e− 4t |x| (x ∈ Rd ). (11.6)

Then (see Exercise 11.C)

∂
ht (x) = ∆ht (x) ((t, x) ∈ (0, ∞) × Rd ). (11.7)
∂t

Exercise 11.C. Show that (11.7) is satisfied for ht as in (11.6).

11.16. Observe that

Z Z d
1 1 2
ht (x) dx = (4πt)− 2 e− 4t s ds = 1,
Rd R

which follows by the fact that

Z
2 √
e−x dx = π.
R

The latter identity can be proved using polar coordinates:

Z 2 Z Z Z ∞
2 2 +y 2 ) 2
e−x dx = e−(x dx dy = 2π re−r dr
R R R 0
Z ∞
1 −s
= 2π e ds = π.
0 2
From this we can show that
t↓0
hht , ϕi −−→ ϕ(0) (ϕ ∈ Cb (Rd )). (11.8)

72
Indeed,
Z
hht , ϕi − ϕ(0) = ht (x)(ϕ(x) − ϕ(0)) dx.
Rd
x
By a substitution y = √
t
we have
Z Z √
ht (x)(ϕ(x) − ϕ(0)) dx = h1 (y)(ϕ( ty) − ϕ(0)) dy.
Rd Rd

So that by the Lebesgue dominated convergence theorem we indeed obtain (11.8).

Exercise 11.D. Calculate the limit in D0 (Rd ) of ∂
∂t ht as t ↓ 0.

In the theory of partial differential equations one is often looking for a function on
(0, ∞) × Rd where the first variable represents the “time variable”. As one distinguishes
the “time variable” from the “space variables”, it makes sense to introduce the following
notation.
Definition 11.17. Let R† = R. Let Ω ⊂ R† × Rd be open. For α ∈ Nd0 , we write ∂ α for
the operation D(Ω) → D(Ω) given by
∂ α1 ∂ αd
∂ α ϕ(t, x) = α1 · · · ϕ(t, x) (ϕ ∈ D(Ω), (t, x) ∈ R† × Rd ),
∂x1 ∂xαd 1
in other words, ∂ α is written for the operation ∂ (0,α) when we view Ω as a subset of R1+d .
Moreover, we write ∂† for the operation D(Ω) → D(Ω) given by
∂
∂† ϕ(t, x) = ϕ(t, x) (ϕ ∈ D(Ω), (t, x) ∈ R† × Rd ),
∂t
i.e., ∂† is the operation ∂ (1,0) when we view Ω as a subset of R1+d .
Remark 11.18. In the literature it is rather common to write “∂t ” instead of “∂† ”. We
avoid this as, on the one hand, we have already defined ∂i for i ∈ {1, . . . , d}, on the other
hand we prefer not to attach a meaning to “t” other than a variable which can be equal
to 1, 2, etc.
Definition 11.19 (Heat operator). The heat operator is the linear partial differential
operator ∂† − ∆.
Example 11.20. Let ht for t > 0 be as in Example 11.15. The function f : (0, ∞)×Rd →
R defined by f (t, x) = ht (x) for (t, x) ∈ (0, ∞)×Rd solves the heat equation on (0, ∞)×Rd :
∂† f = ∆f.
Example 11.21. Define E : R† × Rd → R by
(
ht (x) (t, x) ∈ (0, ∞) × Rd ,
E(t, x) =
0 (t, x) ∈ (−∞, 0] × Rd .
Then (see Exercise 11.E) E is a fundamental solution to the heat operator ∂† − ∆ (one
also says, E is a fundamental solution to the heat equation).

73
Definition 11.22. The gamma function is the function Γ : (0, ∞) → (0, ∞) given by
Z ∞
Γ(s) = ts−1 e−t dt (s ∈ (0, ∞)).
0

It is sometimes also defined on the complex plane for those numbers for which the real
part is strictly positive. By partial integration it follows that Γ(s + 1) = sΓ(s). As
√
Γ(1) = 1, it follows that Γ(n) = (n − 1)! for n ∈ N. Moreover, Γ( 12 ) = π.

Exercise 11.E. Let E be as in Example 11.21.

R∞
(a) Calculate 0 ht (x) dt for x 6= 0 (in terms of the gamma function).
ht (x)ϕ(t, x) dx = ϕ(0) for any ϕ ∈ D(Rd+1 ).
R
(b) Show that limt↓0 Rd
(c) Show that E is locally integrable and conclude that the order of (∂† − ∆)E is at
most 2.
(d) Calculate sing supp E.
(e) Show that supp(∂† − ∆)E ⊂ {0}.
(f) Show that E is a fundamental solution to ∂† − ∆ (Hint: Observe that
h(∂† − ∆)E, ϕi = lims↓0 limT ↑∞ − sT Rd ht (x)(∂† + ∆)ϕ(t, x) dx dt and apply in-
R R

tegration by parts.)
(g) Conclude that if v ∈ E 0 (Rd+1 ) is smooth on an open set U , then so is a solution u
of (∂† − ∆)u = v.
Remark 11.23. In [DK10, Section 12] one finds references for the proof of the statement
that every linear partial differential operator with constant coefficients, of which at least
one coefficient is nonzero, has a fundamental solution.

12 Sobolev spaces

In this section we consider Sobolev spaces as subspaces of D0 . These spaces are subsets
of Lp for which not only the function itself, but also its derivatives (in the distributional
sense) up to a certain order are all included in Lp .
Definition 12.1. Let f ∈ L1loc (Ω) and α ∈ Nd0 . A g ∈ L1loc (Ω) is called the α-th weak
partial derivative of f if ug = ∂ α uf , i.e., if
Z Z
gϕ = f · (−1)|α| ∂ α ϕ (ϕ ∈ D(Ω)).

12.2. Let f, g, h ∈ L1loc (Ω) and α, β ∈ Nd0 .

(1) By Lemma 2.9 f has at most one α-th weak partial derivative.
(2) If g is an α-th weak partial derivative of f , then we write ∂ α f = g.
(3) If g = ∂ α f and h = ∂ β g, then h = ∂ α+β f .

74
 (4) Following our convention 4.2 that we may write “f ” instead of “uf ”, for a distri-
bution u ∈ D0 (Ω) we write “∂ α u ∈ L1loc ” if there exists a g ∈ L1loc such that ∂ α u = ug .

Definition 12.3. Let p ∈ [1, ∞] and k ∈ N0 . We define the Sobolev space of order k and
integrability p, denoted W k,p (Ω), by

W k,p (Ω) = {u ∈ D0 (Ω) : ∂ β u ∈ Lp (Ω) for all β ∈ Nd0 with |β| ≤ k}.

Observe:
(1) ∂ α u ∈ W k−|α|,p (Ω) for all u ∈ W k,p (Ω) and α ∈ Nd0 with |α| ≤ k.
(2) If U is an open subset of Ω and u ∈ W k,p (Ω), then u|U ∈ W k,p (U ).

Exercise 12.A. Consider Ω = (0, 2), f, g ∈ L1loc (Ω) given by

( (
x x ∈ (0, 1], x x ∈ (0, 1],
f (x) = g(x) =
1 x ∈ (1, 2), 2 x ∈ (1, 2).

(a) Show that f has a weak derivative that is in Lp , so that f ∈ W 1,p (Ω) for all
p ∈ [1, ∞].
(b) Show that g has no weak derivative, but calculate ∂ug .
(c) Give an example of an element h ∈ W 1,p (0, 2) such that the function g defined on
R by g(x) = h(x) for x ∈ (0, 2) and g(x) = 0 for other x, is not in W 1,p (R).

Definition 12.4. We equip the Sobolev space W k,p (Ω) for p ∈ [1, ∞] with the norm

kukW k,p = max k∂ β ukLp .

β∈Nd0 ,|β|≤k

Exercise 12.B. Verify that k · kW k,p indeed defines a norm on W k,p (Ω).

Let us recall the definition of equivalent norms:

Definition 12.5. Let X be a normed space and k · k1 , k · k2 : X → [0, ∞) be norms on X.

They are said to be equivalent if they define the same topology.
Two norms k·k1 , k·k2 are equivalent if and only if (see for example [Con90, Proposition
III.1.5]) there exist c, C > 0 such that

ckf k1 ≤ kf k2 ≤ Ckf k1 (f ∈ X).

1
12.6. For p ∈ [1, ∞), n ∈ N and x ∈ Rn let |x|p = ( ni=1 |xi |p ) p . In 1.4, we mentioned
P

that | · |1 is equivalent to | · |∞ (where |x|∞ = maxdi=1 |xi | for x ∈ Rn ). Similarly, | · |p and

| · |∞ are equivalent, as
1
|x|∞ ≤ |x|p ≤ n p |x|∞ (x ∈ Rd ).

75
Therefore, for each q ∈ [1, ∞) there exists a C > 0 such that
 1
k∂ β ukqLp
q
X
kukW k,p ≤ ≤ CkukW k,p ,
β∈Nd0 ,|β|≤k

1
i.e., the norm u 7→ ( β∈Nd ,|β|≤k k∂ β ukqLp ) q is equivalent to k · kW k,p (in other books one
P
0
might find the norm on W k,p to be defined by this norm with either q = 1 or q = p).

Observe that W 0,p (Ω) = Lp (Ω), so that the Sobolev space of 0-th order is a Banach
space. This extends to any order:

Theorem 12.7. For all p ∈ [1, ∞] and k ∈ N0 , W k,p (Ω) is a Banach space.

Proof. Suppose that (un )n∈N is a Cauchy sequence in W k,p (Ω). Then (∂ α un )n∈N is a
Cauchy sequence in Lp (Ω) for all α ∈ Nd0 with |α| ≤ k. As Lp (Ω) is a Banach space,
there exist u(α) ∈ Lp (Ω) such that ∂ α un → u(α) in Lp for all such α.
Let us write u for u(0) . We are finished by showing that ∂ α u = u(α) in D0 (Ω) for
all such α ∈ Nd0 , as this implies un → u in WRk,p (Ω). This follows by testing against a
p
R
testfunction ϕ, using; if fn → f in L , then fn ϕ → f ϕ (which follows by Hölder’s
inequality):
Z Z Z
h∂ α u, ϕi = u · (−1)|α| ∂ α ϕ = lim un · (−1)|α| ∂ α ϕ = lim ∂ α un · ϕ = hu(α) , ϕi.
n→∞ n→∞

As this holds for all ϕ ∈ D(Ω), we have ∂ α u = u(α) (by Lemma 2.9).

Let us consider the continuity of the operator ∂ α on the C m space and embeddings
of such spaces and then show the analogue statements when we instead of C m spaces
consider Sobolev spaces.

Lemma 12.8. Let k, m ∈ N0 and α ∈ Nd0 . If k ≤ m, then for all compact sets K ⊂ Rd ,

k∂ α f kC k−|α| ,K ≤ kf kC m ,K (f ∈ C m ),

and

k∂ α f kC k−|α| ≤ kf kC m (f ∈ Cbm ).

In particular, ∂ α : C m → C m−|α| is continuous for all m ∈ N0 and α ∈ Nd0 , |α| ≤ m, and

C m ,→ C k , Cbm ,→ Cbk (k ≤ m).

Proof. This easily follows by the definitions of the norms k · kC m ,K and k · kC m , we leave
it to the reader to check this.

76
Lemma 12.9. Let p ∈ [1, ∞]. Let k, m ∈ N0 and α ∈ Nd0 . If k ≤ m, then

k∂ α ukW k−|α|,p ≤ kukW m,p (u ∈ W m,p ).

In particular, ∂ α : W m,p → W m−|α|,p is continuous for all m ∈ N0 and α ∈ Nd0 , |α| ≤ m

and

W m,p ,→ W k,p (k ≤ m).

Proof. The proof is again rather straightforward and left to the reader.

Similar to Proposition 5.3 one has:

Lemma 12.10. For all k ∈ N0 there exists a C > 0 such that for all p ∈ [1, ∞] and
r, q ∈ [1, ∞] with 1r + 1q = p1 ,

kuvkW k,p ≤ CkukW k,r kvkW k,q (u ∈ W k,r (Ω), v ∈ W k,q (Ω)).

Consequently, the function W k,r (Ω) × W k,q (Ω) → W k,p (Ω) given by (u, v) 7→ uv is con-
tinuous.

Proof. See the proof of Proposition 5.3 and use additionally that kf gkLp ≤ kf kLr kgkLq
for f ∈ Lp (Ω) and g ∈ L∞ (Ω), which follows by Hölder’s inequality.

Theorem 12.11. Let p ∈ [1, ∞). Then D(Ω) is dense in W k,p (Ω).

Proof. Let χ and χR for R > 0 be as in Lemma 8.16. By Lemma 12.10 χR |Ω u ∈ W k,p (Ω)
for all R > 0. By Lemma 8.16 limR→∞ χR |Ω u = u in W k,p (Ω). Therefore it is sufficient
to show that for all compactly supported u ∈ W k,p (Ω) there exist uε ∈ D(Ω) for ε > 0
ε↓0
such that uε −−→ u in W k,p (Ω). This follows similarly as in Theorem 8.17 by using (8.5).
We leave the details to the reader.

Definition 12.12. Let k ∈ N0 and p ∈ [1, ∞].

(a) We write W0k,p (Ω) for the closure of Cc∞ (Ω) in W k,p (Ω).
(b) We write H k (Ω) = W k,2 (Ω) and H0k (Ω) = W0k,2 (Ω).
(c) We define h·, ·iH k : H k (Ω) × H k (Ω) → F by
X
hu, viH k = h∂ α u, ∂ α viL2 (u, v ∈ H k (Ω)),
α∈Nd0 :|α|≤k

and k · kH k : H k (Ω) → [0, ∞) by

q
kukH k = hu, uiH k (u ∈ H k (Ω)).

77
Remark 12.13. One interprets W0k,p (Ω) as the subspace of W k,p (Ω) of elements that
vanish at the boundary of Ω, in symbols, u = 0 on ∂Ω.

Similarly to Theorem 12.7, in which we showed that W k,p is a Banach space by using
that Lp is a Banach space, one can show that H k is a Hilbert space because L2 is:
Theorem 12.14. Let k ∈ N0 . h·, ·iH k is an inner product on H k (Ω), so that H k (Ω)
(and H0k (Ω)) equipped with this inner product is a Hilbert space.

Proof. We leave it for the reader to check that h·, ·iH k defines an inner product. The
rest follows from Theorem 12.7 and the fact that k · kH k is equivalent to k · kW 2,k (see
12.6).

There is a lot of theory on Sobolev spaces, which we will not treat here. Sobolev
spaces play a central role in the theory of partial differential equations. In the following
section we consider an application to elliptic partial differential operators. One classical
reference for PDE theory, which contains a whole section on Sobolev spaces is [Eva98]
(see Section 5). There are various estimates that are useful, of which we present one
important example, the Poincaré inequality.
Definition 12.15 (Lp (Ω, Rd ) and Lp (Ω, Rd )). Lp (Ω, Rd ) is the space of (Lebesgue)
measurable f : Ω → Rd such that
Z 1
p
kf kLp (Ω,Rd ) := |f (x)|p dx < ∞.

Lp (Ω, Rd ) is the space of all equivalence classes in Lp (Ω, Rd ). We mostly write “k · kLp ”
instead of “k·kLp (Ω,Rd ) ”. Lploc (Ω, Rd ) is the space of functions that are locally in Lp (Ω, Rd ),
i.e., f ∈ Lploc (Ω, Rd ) if and only if f 1K ∈ Lp (Ω, Rd ) for all compact sets K ⊂ Ω.
Let f : Ω → Rd be measurable and f1 , . . . , fd : Ω → Rd be its coordinates, i.e., f (x) =
(f1 (x), . . . , fd (x)). Then fi is measurable for each i ∈ {1, . . . , d} and f ∈ Lp (Ω, Rd ) if
and only if fi ∈ Lp (Ω) for all i ∈ {1, . . . , d}. Moreover,
Z  p 1
p
2 2 2
kf kLp = |f1 (x)| + · · · + |fd (x)| dx .

On the other hand, k · kLp is equivalent to (see 12.6)

Z 1
p
f 7→ |f1 (x)|p + · · · + |fd (x)|p dx ,

and to f 7→ kf1 kLp + · · · + kfd kLp .

Definition 12.16. An open set Ω ⊂ Rd is said to be of finite width if there exist a, b ∈ R,
a < b a y ∈ Rd with |y| = 1 and

{x · y : x ∈ Ω} ⊂ [a, b], (12.1)

78
where · is the inner product on Rd . The Ω is said to be of width w if
n o
w = inf b − a : a, b ∈ R, a < b, y ∈ Rd , |y| = 1, {x · y : x ∈ Ω} ⊂ [a, b] .

Theorem 12.17 (The Poincaré inequality). Let w > 0. Let Ω be a nonempty open
subset of Rd of finite width w. For all p ∈ [1, ∞)
w
kukLp ≤ 1 k∇ukLp (u ∈ W01,p (Ω)). (12.2)
p p

Proof. Let u ∈ W01,p (Ω). We extend u to Rd by defining it to be equal to 0 outside Ω.

Let y ∈ Rd , |y| = 1 and a, b ∈ R, a < b with b = a + w be such that (12.1). In the
first step we show the inequality in case y = e1 . In the second step we show how we can
reduce the general case to the case where y = e1 .
Step 1 Let x ∈ Rd , x0 = (x2 , . . . , xd ). Then by Hölder’s inequality,

|u(x1 , x0 )| = |u(x1 , x0 ) − u(a, x0 )|

Z x1
≤ |∂1 u(s, x0 )| ds
a
p−1
≤ (x1 − a) p k∂1 u(·, x0 )kLp (R) .

Therefore (using Fubini’s theorem)

Z Z Z b
p
p
|u| = u(x1 , x0 ) dx1 dx0
Rd−1 a
Z Z b
≤ (x1 − a)p−1 dx1 k∂1 u(·, x0 )kpLp (R) dx0
Rd−1 a
(b − a)p wp
= k∂1 ukLp (Rd ) ≤ k∇ukLp (Rd ) .
p p

Step 2 Let v1 , . . . , vd be an orthonormal basis of Rd with v1 = y. Then the matrix

which columns equal v1 , . . . , vd ; R = [v1 v2 · · · vd ] is such that Re1 = v1 = y and, by
Pythagoras’ theorem,

|Rx| = |x| (x ∈ Rd ).

Then, writing also R for the linear bijection v 7→ Rv, R−1 Ω ⊂ [a, b]e1 + (Re1 )⊥ (where
(Re1 )⊥ is the orthogonal complement of Re1 ) and kukLp = ku ◦ RkLp , ∇(u ◦ R) =
R[(∇u) ◦ R] and

k∇ukLp = kR[(∇u) ◦ R]kLp .

Therefore the Poincaré inequality follows from Step 1.

79
Corollary 12.18. Let Ω be a nonempty open subset of Rd of finite width. Then the
norms W01,p → [0, ∞),

u 7→ kukW 1,p , u 7→ k∇ukLp ,

are equivalent.
Exercise 12.C. Prove the following statement. Let Ω be a bounded open subset of
Rd . Let p ∈ [1, ∞). For each r ∈ [1, p] there exists a C > 0 such that

kukLr ≤ Ck∇ukLp (u ∈ W01,p (Ω)).

Exercise 12.D. In this exercise we show that for open sets that do not have finite
width, an inequality as the Poincaré inequality (12.2) does not hold (for any w > 0):
Suppose Ω ⊂ Rd is an open set that contains each ball B(xn , n) for n ∈ N and some
sequence (xn )n∈N in Rd . Show that for each m ∈ N there exists a um ∈ W01,p (Ω) with
k∇um kLp = 1 and kum kLp ≥ m.

13 Solutions to elliptic PDEs in Sobolev spaces

In this section we consider the existence of solutions to elliptic partial differential equa-
tions equations. The notion of solution will be defined in the language of Sobolev spaces.
In this section the scalar field is the real numbers, i.e., F = R and Ω is an open subset
of Rd .
Definition 13.1. We call a map P : D(Ω) → D0 (Ω) a linear partial differential operator
with variable coefficients if there exist an m ∈ N and uα ∈ D0 (Ω) for α ∈ Nd0 with |α| ≤ m
such that
X
P = uα ∂ α ,
α∈Nd0
|α|≤m

i.e.,
X
Pϕ = (∂ α ϕ)uα (ϕ ∈ D(Ω)).
α∈Nd0
|α|≤m

m is called the order of P .

Definition 13.2. A linear partial differential operator P of order 2 is called a second
order linear partial differential operator. Let P be such an operator. Then there exist
aij , bi , c ∈ D0 (Ω) for i, j ∈ {1, . . . , d} such that
d
X d
X
P =− aij ∂i ∂j + bi ∂i + c.
i,j=1 i=1

80
If aij ∈ L1loc (Ω) for all i, j ∈ {1, . . . , d}, then P is called elliptic if there exists a θ > 0
such that
d
X
aij (x)yi yj ≥ θ|y|2 (almost all x ∈ Ω, y ∈ Rd ). (13.1)
i,j=1

Let a ∈ L1loc (Ω, Rd×d ) be the matrix valued function such that (a)ij = aij . Then P is
elliptic if and only if a − θI is positive definite for some θ > 0, where I is the identity
matrix.

Observe that −∆ is an elliptic operator.

13.3. As we allow the coefficients uα to be distributions, the domain of the operator P
is D(Ω) and not D0 (Ω) as in Definition 11.1. But observe the following.

• If uα ∈ L1loc (Ω) for each α ∈ Nd0 with |α| ≤ m, then P ϕ ∈ L1loc (Ω) for all ϕ ∈ D(Ω)
and
X
P ϕ(x) = uα (x)∂ α ϕ(x) (ϕ ∈ D(Ω), x ∈ Ω).
α∈Nd0
|α|≤m

• If uα ∈ E(Ω) for each α ∈ Nd0 with |α| ≤ m, then P ϕ ∈ D(Ω) for all ϕ ∈ D(Ω) and
P extends to an operator D0 (Ω) → D0 (Ω).
13.4 (Assumptions). In this section we consider the following setting. Let

aij , bi , c ∈ L∞ (Ω), (i, j ∈ {1, . . . , d}).

Let L : D(Ω) → D0 (Ω) be defined by

d
X d
X
Lϕ = − ∂i (aij ∂j ϕ) + bi ∂i ϕ + cϕ (ϕ ∈ D(Ω)). (13.2)
i,j=1 i=1
Pd
Observe that with b̃i = bi − j=1 ∂j aij we have
d
X d
X
Lϕ = − aij ∂ij ϕ + b̃i ∂i ϕ + cϕ (ϕ ∈ D(Ω)), (13.3)
i,j=1 i=1

from which we see that L is a second order linear partial differential operator with variable
coefficients. Let a ∈ L∞ (Ω, Rd×d ) be the matrix valued function given by (a)ij = aij and
b ∈ L∞ (Ω, Rd ) the vector valued function given by (b)i = bi . Then a∇ϕ is an element of
L∞ (Ω, Rd ). By writing ∇ · f = di=1 ∂i f for any f ∈ L∞ (Ω, Rd ), L can be written as
P

Lϕ = −∇ · (a∇ϕ) + b · ∇ϕ + cϕ (ϕ ∈ D(Ω)).

∇ · f is also called the divergence of f.

81
 We consider the following problem. Fix a function f : Ω → R. We want to find a
distribution u such that Lu = f and u “equals zero on ∂Ω”. For a distribution (or an
element of L1loc ) “u = 0 on ∂Ω” does not make sense. Instead, we consider the problem
of finding a u in a Sobolev space W0k,p such that Lu = f (remember Remark 12.13). We
formalise this in Definition 13.5 with the help of the bilinear form associated with L.
Observe that
d
X d
X
hLϕ, ψi = − h∂i (aij ∂j ϕ), ψi + hbi ∂i ϕ, ψi + hcϕ, ψi
i,j=1 i=1
 
Z d
X d
X
=  aij (∂i ϕ)(∂j ψ) + bi (∂i ϕ)ψ + cϕψ  (ϕ, ψ ∈ D(Ω)).
Ω i,j=1 i=1

As the aij , bi and c are bounded, we have

 
d
X d
X
|hLϕ, ψi| ≤  kaij kL∞ + kbi kL∞ + kckL∞  kϕkH 1 kψkH 1 (ϕ, ψ ∈ D(Ω)).
i,j=1 i=1
(13.4)

Therefore, as D(Ω) is dense in H01 (Ω) by definition, the bilinear form D(Ω) × D(Ω) → R,
(ϕ, ψ) 7→ hLϕ, ψi extends to a bilinear form H01 (Ω) × H01 (Ω) → R.
Definition 13.5. Let L be as in 13.4.

(a) The bilinear form associated with L is the function B : H01 (Ω)×H01 (Ω) → R defined
by
Z d
X d
X
B(u, v) = aij (∂i u)(∂i v) + bi (∂i u)v + cuv.
Ω i,j=1 i=1

(b) If f ∈ L2 (Ω), u ∈ H01 (Ω) and

B(u, v) = hf, viL2 (v ∈ H01 (Ω)),

then we call u a weak solution to the Dirichlet boundary problem

(
Lu = f on Ω,
(13.5)
u=0 on ∂Ω.

Exercise 13.A. Let L be as in 13.4 and suppose that aij = aji and bi = 0 for all
i, j ∈ {1, . . . , d}. Show that B is symmetric; B(u, v) = B(v, u) for all u, v ∈ H01 (Ω).

We will use tools from functional analysis to prove that under certain conditions there
exists a weak solution to the Dirichlet boundary problem (13.5). Let us first recall the
Riesz-Fréchet theorem, for a proof see for example [Rud91, Theorem 12.5].

82
Theorem 13.6 (Riesz–Fréchet). Let H be a Hilbert space over R with inner product
h·, ·i. If A : H → R is a bounded linear functional, then there exists exactly one a ∈ H
such that

Ax = ha, xi (x ∈ H).

Theorem 13.7 (Lax–Milgram). Let H be a Hilbert space over R, with inner product
h·, ·i and norm k · k. Let B : H × H → R be a bilinear map. Suppose there exist c, C > 0
such that

|B(u, v)| ≤ Ckukkvk (u, v ∈ H), (13.6)

2
ckuk ≤ B(u, u) (u ∈ H). (13.7)

(a) There exists a linear homeomorphism A : H → H such that

B(u, v) = hAu, vi (v ∈ H).

(b) Let g : H → R be a bounded linear functional. Then there exists exactly one u ∈ H
such that

B(u, v) = g(v) (v ∈ H).

Proof. Observe that (b) follows from (a) and the Riesz-Fréchet theorem, Theorem 13.6.
Let us prove (a).
As for u ∈ H the map v 7→ B(u, v) is a bounded linear functional, the Riesz-Fréchet
theorem implies that there exists an element in H, for which we write Au, such that

B(u, v) = hAu, vi (v ∈ H).

Then A defines a linear map H → H. By the Inverse Mapping Theorem, see for example
[Con90, Theorem III.12.5], it is sufficient to show that A is bounded and bijective.
We have

kAuk2 = hAu, Aui = B(u, Au) ≤ CkukkAuk (u ∈ H).

Therefore kAuk ≤ Ckuk for u ∈ H, so that A is bounded.

By (13.7) we have

ckuk2 ≤ B(u, u) = hAu, ui ≤ kAukkuk (u ∈ H),

and thus

ckuk ≤ kAuk (u ∈ H),

from which follows that A is injective and its range, AH, is closed in H: If (un )n∈N is
a sequence in H such that Aun converges, then it follows that (un )n∈N is Cauchy and
therefore has a limit u in H. By the boundedness of A follows that Aun → Au.

83
 Now, let us prove that AH, the range of A, equals H. As A(H) is closed we have
AH +(AH)⊥ = H (where (AH)⊥ is the orthogonal complement of AH), so it is sufficient
to show that (AH)⊥ = {0}. Let w ∈ (AH)⊥ . Then 0 = hAw, wi = B(w, w) ≥ ckwk2 . So
w = 0.

Exercise 13.B. Let H and B be as in Theorem 13.7. Suppose furthermore that B is

symmetric, in the sense that B(u, v) = B(v, u) for all u, v ∈ H. Why does statement (b)
directly follow from the Riesz–Fréchet theorem?

Let us verify the assumptions of the Lax-Milgram theorem for the bilinear form
associated with L.

Theorem 13.8. Let L be as in 13.4 and B be the bilinear form associated with L.
Suppose L is elliptic and Ω is of finite width. There exist a γ ≥ 0 and c, C > 0 such that

|B(u, v)| ≤ CkukH 1 kvkH 1 (u, v ∈ H01 (Ω)), (13.8)

ckuk2H 1 ≤ B(u, u) + γkuk2L2 (u ∈ H01 (Ω)). (13.9)

Proof. (13.8) follows from the following estimate, see also (13.4),
 
d
X d
X
|B(u, v)| ≤  kaij kL∞ + kbi kL∞ + kckL∞  kukH 1 kvkH 1 (u, v ∈ H01 (Ω)).
i,j=1 i=1

On the other hand, for θ > 0 as in (13.1) we have for u ∈ H01 (Ω)
d Z
X Z d
X
θ |∂i u|2 ≤ aij (∂i u)(∂j u)
i=1 Ω Ω i,j=1

d
Z X Z
= B(u, u) − bi u∂i u − cu2
Ω i=1 Ω
d
X Z Z
≤ B(u, u) + kbi k L∞ |∂i u||u| + kck L∞ u2 .
i=1 Ω Ω

1 2
As ab ≤ εa2 + 4ε b for any a, b ∈ R and ε > 0 we have

1
Z Z Z
2
|∂i u||u| ≤ ε |∂i u| + |u|2 (u ∈ H01 (Ω)).
Ω Ω 4ε Ω
Pd
Let M = i=1 kbi kL∞ and take ε small enough such that εM ≤ 2θ . Then

d d
θX M
X Z Z Z
kbi kL∞ |∂i u||u| ≤ |∂i u|2 + u2 .
i=1 Ω 2 i=1 4ε Ω

84
By the Poincaré inequality (see Theorem 12.17) there exists a β > 0 such that
d Z
X
βkuk2H 1 ≤ |∂i u|2 (u ∈ H01 (Ω)).
i=1 Ω

Thus
 d Z
θ θ X M
  Z
2 2
β kukH 1 ≤ θ − |∂i u| ≤ B(u, u) + kckL∞ + u2 .
2 2 i=1 Ω 4ε Ω

Now we can prove that under certain conditions (13.5) has a weak solution.

Theorem 13.9. Let L be as in 13.4. Suppose L is elliptic and Ω is of finite width. There
exists a γ ≥ 0 such that for all β ≥ γ and f ∈ L2 (Ω) there exists a unique weak solution
u ∈ H01 (Ω) of the Dirichlet boundary problem
(
Lu + βu = f on Ω,
(13.10)
u=0 on ∂Ω.

Proof. Let γ ≥ 0 be as in Theorem 13.8. Let β ≥ γ. We apply the Lax-Milgram

theorem to Bβ , the bilinear operator corresponding to the elliptic operator Lβ given by
Lβ u = Lu + βu:

Bβ (u, v) = B(u, v) + βhu, viL2 (u, v ∈ H01 (Ω)).

Observe that for f ∈ L2 (Ω) the map g : H01 (Ω) → R given by g(v) = hf, viL2 is bounded
and linear, because kvk2L2 ≤ kvk2H 1 . So by the Lax-Milgram theorem there exists exactly
one u ∈ H01 (Ω) such that Bβ (u, v) = hf, viL2 for all v ∈ H01 (Ω), which means that u is a
weak solution to (13.10).

For more theory on weak solutions of elliptic Dirichlet boundary problems, we refer
the reader to [Eva98, Section 6.2] . Moreover, one can show that the solutions have a
certain regularity that depends on the regularity of the coefficients ai,j , bi , c, see [Eva98,
Section 6.3] .

Exercise 13.C. Show that one can choose γ = 0 in Theorem 13.8 and Theorem 13.9
in case bi = 0 for all i and c = 0.

14 The Schwartz space

In this section we introduce the Schwartz space, which is the space of smooth functions
that are of rapid decay. This space is suitable for the Fourier transformation, as the

85
Fourier transformation maps the Schwartz space onto itself. We will turn to that later
and first discuss here the topological properties of the Schwartz space. In Section 15
we consider its dual, the space of tempered distributions. As our underlying space we
consider Rd (only). For this reason we can leave out the part “(Rd )” in the notation of
function spaces or spaces of distributions.
Definition 14.1. We say that a function f : Rd → F is of rapid decay if

lim P (x)f (x) = 0,

|x|→∞

for all polynomials P , where lim|x|→∞ g(x) = a means that for all ε > 0 there exists an
R > 0 such that for all x ∈ Rd with |x| > R, |g(x) − a| < ε.
Observe that f is of rapid decay if and only if lim|x|→∞ xα f (x) = 0 for all α ∈ Nd0 .
As xj ≤ (1 + |x|2 ) for all j ∈ {1, . . . , d} and x ∈ Rd it follows that for each polynomial P
there exists a C > 0 and k ∈ N such that

|P (x)| ≤ C(1 + |x|)k (x ∈ Rd ).

Therefore, f is of rapid decay if and only if lim|x|→∞ (1 + |x|)k f (x) = 0 for all k ∈ N0 .
See Exercise 14.A for equivalent descriptions of rapid decay for continuous functions.
Exercise 14.A. Prove the following statement (Hint: First prove: xj ≤ (1 + |x|2 ) for
all j ∈ {1, . . . , d} and x ∈ Rd ):
Let f : Rd → F be continuous. The following statements are equivalent:

(a) f is of rapid decay.

(b) x 7→ P (x)f (x) is bounded for all polynomials P .
(c) For all k ∈ N0 there exists an M < ∞ such that

|f (x)| ≤ M (1 + |x|)−k (x ∈ Rd ).

(d)

k(1 + | · |)k f k∞ = sup (1 + |x|)k |f (x)| < ∞ (k ∈ N0 ).

x∈Rd

Definition 14.2. A smooth function ϕ is called a Schwartz function if the function and
all its derivatives are of rapid decay: if ∂ α ϕ is of rapid decay for all α ∈ Nd0 .
We write S (or S(Rd )) for the space of Schwartz functions and call it the Schwartz
space. For k ∈ N0 we define k · kk,S : S → [0, ∞) by

kϕkk,S := max k(1 + | · |)k ∂ α ϕkL∞ (ϕ ∈ S). (14.1)

α∈Nd0
|α|≤k

k · kk,S is a norm for all k ∈ N0 . The space S is equipped with the topology generated
by the seminorms k · kk,S .

86
 Each testfunction is a Schwartz function. Gaussian functions are examples of Schwartz
functions that are not compactly supported:
Definition 14.3. A function f : Rd → R is called a Gaussian function if there exist
a, b ∈ R, a > 0, y ∈ Rd such that
2
f (x) = be−a|x−y| .

Each such function f is smooth and is a Schwartz function: Let α ∈ Nd0 and k = |α|.
Then
2
|∂ α f (x)| ≤ |b|(2a)k |x − y|k e−a|x−y| (x ∈ Rd ),

and thus, using that (1 + |x + y|)k ≤ (1 + |y|)k (1 + |x|)k ,

2
kf kk,S ≤ |b|(2a(1 + |y|))k sup (1 + |x|)2k e−a|x| < ∞. (14.2)
x∈Rd

14.4 (Equivalent norms). In the literature one finds different definitions of norms or
seminorms on S, which all generate the same topology. For example, for k ∈ N0 the
function ||| · |||k,S : S → [0, ∞) defined by
k
|||ϕ|||k,S = max k(1 + | · |2 ) 2 ∂ α ϕ(x)kL∞ ,
α∈Nd0
|α|≤k

is a norm that is equivalent to k · kk,S , which can be seen by the estimate

1 + |x|2 ≤ (1 + |x|)2 ≤ 2(1 + |x|2 ) (x ∈ Rd ). (14.3)

On the other hand, the topology on S generated by the seminorms k · kk,S with k ∈ N0
is equal to the topology generated by the seminorms |||·|||k,α with k ∈ N0 and α ∈ Nd0 ,
which are defined by

||| · |||k,α = k(1 + | · |2 )k ∂ α ϕkL∞ (ϕ ∈ S).

Exercise 14.B. Verify that the seminorms |||·|||k,α with k ∈ N0 and α ∈ Nd0 generate
the same topology on S as the norms k · kk,S with k ∈ N0 .

Exercise 14.C. Give an example of function ϕ in C ∞ (R) for which ϕ is of rapid decay
but its derivative ϕ0 is not.

The operations of reflection, translation and derivation are operations S → S. On the

other hand, a Schwartz function multiplied by a smooth function may not be a Schwartz
2 2
function; take ϕ ∈ S, ψ ∈ S given by ϕ(x) = e−|x| and ψ(x) = e2|x| for x ∈ Rd , then
2
ϕψ(x) = e|x| which is clearly not of rapid decay and therefore not a Schwartz function.
Multiplication with a Cb∞ function is an operation S → S, but we may allow for more
smooth functions:

87
Definition 14.5. Let Ω ⊂ Rd be open. A function f : Ω → F is said to be of at most
polynomial growth if either Ω is bounded or there exists a polynomial P such that

f (x)
lim = 0,
|x|→∞ P (x)

or equivalently (as in Definition 14.1), there exists a k ∈ N0 such that

f (x)
lim = 0.
|x|→∞ (1 + |x|)k

Exercise 14.D. Prove the following statement:

Let f : Rd → F be continuous. The following statements are equivalent:

(a) f is of at most polynomial growth.

(b) There exists a polynomial p : R → R such that |f (x)| ≤ p(|x|) for all x ∈ Rd .
(c) There exist a C > 0 and a k ∈ N0 such that

|f (x)| ≤ C(1 + |x|)k (x ∈ Ω).

(d) There exists a k ∈ N0 such that

k(1 + | · |)−k f kL∞ < ∞.

Definition 14.6. We write Cp∞ (Ω) for the set of smooth functions σ such that for all
α ∈ Nd0 , the function ∂ α σ is of at most polynomial growth.
For Cp∞ = Cp∞ (Rd ) and σ ∈ C ∞ we have σ ∈ Cp∞ if and only if for all m ∈ N0 there
exists an k ∈ N0 such that

qm,k (σ) := max k(1 + | · |2 )−k ∂ α σkL∞ < ∞. (14.4)

α∈Nd0
|α|≤m

Observe that qm,0 (σ) = kσkC m for σ ∈ Cb∞ (Ω).

m (Ω) be the space of functions f ∈ C m (Ω) such
Exercise 14.E. Let m, k ∈ N0 . Let Cp,k
m (Ω) under which it is a Banach
that qm,k (f ) < ∞. Show that qm,k is a norm on Cp,k
space.
∞ (Ω) be the space of functions f ∈ C ∞ (Ω) such that q
Let Cp,k m,k (f ) < ∞ for all
∞ (Ω) equipped with the seminorms (q
m ∈ N0 . Show that Cp,k m,k m∈N is a Fréchet space
)
and
∞
Cp,k (Ω) ,→ E.

88
Lemma 14.7. For all σ ∈ Cp∞ and ϕ ∈ S we have σϕ ∈ S. Moreover, for all m ∈ N0
there exists a C > 0 such that

kσϕkm,S ≤ Cqm,k (σ)kϕkm+k,S (σ ∈ Cp∞ , k ∈ N0 , ϕ ∈ S), (14.5)

in particular,

kσϕkm,S ≤ CkσkC m kϕkm,S (σ ∈ Cb∞ , ϕ ∈ S), (14.6)

kϕψkm,S ≤ Ckϕkm,S kψkm,S (ϕ, ψ ∈ S). (14.7)

Proof. Let k, m ∈ N0 . By (5.3), which relies on Leibniz’ rule, there exists a C > 0 such
that for all σ ∈ Cp∞ and ϕ ∈ S

kσϕkm,S = max sup (1 + |x|)m |∂ α (σϕ)(x)|

α∈Nd0 x∈Rd
|α|≤m
  
≤ C max sup (1 + |x|)−k |∂ α σ(x)| max sup (1 + |x|)m+k |∂ β ϕ)(x)| .
α∈Nd0 x∈Rd β∈Nd0 x∈Rd
|α|≤m |β|≤m

∞ be as in Exercise 14.E. Show that C ∞ × S → S, (σ, ϕ) 7→ σϕ

Exercise 14.F. Let Cp,k p,k
is continuous.
14.8. The reflection operator R (see 2.12), the translation operator Ty , the derivation
operator ∂ α , multiplication with a Cp∞ function and the operation of composing with a
linear bijection form continuous maps S → S.
14.9 (Notation). For λ ∈ R \ {0} and a function f : Ω → F we write lλ f for the function
1
λ Ω → F given by

lλ f (x) = f (λx) (x ∈ λ1 Ω).

The following rather elementary estimates and convergences will be used multiple
times.
Lemma 14.10.

(a) Let K ⊂ Rd be compact and m ∈ N0 . Then

kϕkC m ,K ≤ kϕkC m ≤ kϕkm,S (ϕ ∈ S),

m
kϕkm,S ≤ (1 + sup |x|) kϕkC m (ϕ ∈ D).
x∈supp ϕ

(b) Let r > 0. If ψ ∈ S and ψ = 0 on B(0, r), then for k ∈ N0

kψkk+1,S
kψkk,S ≤ .
1+r

89
 (c) Let χ ∈ Cc∞ (Rd , [0, 1]) be equal to 1 on a neighbourhood of 0. Then
λ↓0
(lλ χ)ϕ −−→ ϕ in S (ϕ ∈ S).
PN
(d) Let (χn )n∈N be a partition of unity with supN ∈N k n=1 χn kC k < ∞ for all k ∈ N0 .
Then
N
X N →∞
χn ϕ −−−−→ ϕ in S (ϕ ∈ S).
n=1

Proof. (a) can be easily checked. (b) follows by

(1 + |x|)k+1 α
kψkk,S = max sup (1 + |x|)k |∂ α ψ(x)| ≤ max sup |∂ ψ(x)|
α∈Nd0 x∈Rd α∈Nd0 x∈Rd 1+r
|α|≤k |x|≥r |α|≤k
kψkk+1,S
≤ .
1+r
(c) follows by (b) and Lemma 14.7 and using that klλ χ − 1kC k ≤ kχkC k + 1 for all
λ ∈ (0, 1). We leave the details and the proof of (d) to the reader, see Exercise 14.G.

Exercise 14.G. Prove Lemma 14.10(c) and (d).

Theorem 14.11. D is sequentially continuously embedded in S and S is continuously
embedded in Cb∞ , so that
D ,→seq S ,→ Cb∞ ,→ E.

Proof. D ,→seq S ,→ Cb∞ follows by Lemma 14.10 (a), Cb∞ ,→ E is already observed in
Definition 9.2.

Theorem 14.12. S is a separable Fréchet space and D is dense in S.

Proof. As S is equipped with a countable number of seminorms, it is metrizable with a

translation invariant metric Theorem 3.8. If (ϕn )n∈N is a Cauchy sequence in S, then
there exists a ϕ in Cb∞ such that
kϕn − ϕkC k → 0.
as S is continuously embedded in Cb∞ . As (ϕn )n∈N is Cauchy, it is bounded; for each
k ∈ N0 there exists an M > 0 such that kϕn kk,S < M for all n ∈ N, and thus
|∂ α ϕn (x)| ≤ M (1 + |x|)−k (x ∈ Rd ),

so that |∂ α ϕ(x)| ≤ M (1 + |x|)−k for all x ∈ Rd as ∂ α ϕn converges uniformly to ∂ α ϕ.

Therefore ϕ ∈ S (see also Exercise 14.A).
That D is dense in S follows from Lemma 14.10(c). The separability then follows by
the fact that D is separable, see Theorem 8.15.

90
 Let us recall the integrability of the function (1 + | · |)α .
Lemma 14.13. Let α ∈ R.
α
(a) The functions Rd → R, x 7→ (1 + |x|)−α and x 7→ (1 + |x|2 )− 2 are integrable if and
only if α > d.
α
(b) The functions Zd → R, k 7→ (1 + |k|)−α and k 7→ (1 + |k|2 )− 2 are summable if and
only if α > d.

Proof. (a) It sufficient to show that x 7→ (1 + |x|)−α is integrable if and only if α > d
(see (14.3)). Integrating this function on B(0, 1) gives a finite integral for each α ∈ R. It
will be clear that α > 0 is required. By changing to spherical coordinates and observing
that (2r)−α ≤R (1 + r)−α ≤ r−α for α > 0 and r ≥ 1, we see that (1 + |x|)−α is integrable
if and only if 1∞ rd−1−α dr is finite. The latter is of course the case if and only if α > d.
α
(b) It sufficient to show that k 7→ (1 + |k|2 )− 2 is summable if and only if α > d (see
(14.3)). We write bxc = (bx1 c, . . . , bxd c) for x ∈ Rd where bxR1 c is the largest integer that
2 −α 2 −α
√ d (1 + |k| ) 2 = Rd (1 + |bxc| ) 2 dx. Note
P
is smaller than or√ equal to x1 . Then k∈Z
that |x − bxc| ≤ d. Therefore, if |x| ≥ 2 d we have
√ √
1
2 |x| ≤ |x| − d ≤ |bxc| ≤ |x| + d ≤ 32 |x|.

Hence, 14 (1 + |x|2 ) ≤ (1 + |bxc|2 ) ≤ 49 (1 + |x|2 ) for those x and so the statement follows
by (a).

By Lemma 14.13 it follows that all measurable functions of rapid decay are integrable,
and in Lp for any p ∈ [1, ∞].
Lemma 14.14. Let p ∈ [1, ∞). For all m ∈ N such that pm > d there exists a C > 0
such that

k · kLp ≤ Ck · km,S .

As k · kL∞ = k · k0,S , we therefore have for all p ∈ [1, ∞] that S is continuously embedded
in Lp ,

S ,→ Lp (p ∈ [1, ∞]).

Moreover, S is dense in Lp .

Proof. Let m ∈ N be such that pm > d. By Lemma 14.13 C := k(1 + | · |)−m kLp is finite.
Let f ∈ S. By definition of k · km,S we have

|f (x)| ≤ kf km,S (1 + |x|)−m (x ∈ Rd ),

and thus kf kLp ≤ Ckf km,S .

The fact S is dense in Lp follows from the fact that D is dense in Lp , see Theorem 8.17.

91
 By the continuity of the partial derivation one then derives that the Schwartz space
is continuously embedded in the Sobolev spaces. The Schwartz space is also dense in the
Sobolev space W k,p when p is not infinite. This statement and its proof are postponed
to Theorem 23.8.

Lemma 14.15. Let k ∈ N0 and p ∈ [1, ∞]. Then

S ,→ W k,p .

Exercise 14.H. Prove Lemma 14.15.

15 Tempered distributions

In this section we consider tempered distributions, which form a subspace of the space of
distributions. Even though not every locally integrable function is a tempered distribu-
tion, the space of tempered distributions has the benefit that we can define the Fourier
transform on it, which we heavily use in the sequel.

Definition 15.1. A linear continuous map S → F is called a tempered distribution. We

write S 0 (or S 0 (Rd )) for the space of tempered distributions. In other words, u ∈ S 0 if
and only if u is linear and there exist a k ∈ N0 and a C > 0 such that

|u(ϕ)| ≤ Ckϕkk,S (ϕ ∈ S).

If u is a tempered distribution, then u|D is a distribution by Lemma 14.10 (a). Therefore

each tempered distribution corresponds to a u ∈ D0 (Rd ) that continuously extends to a
function S(Rd ) → F.
S 0 is equipped with the σ(S 0 , S) topology, where h·, ·i : S 0 × S → F is given by
hu, ϕi = u(ϕ) for u ∈ S 0 , ϕ ∈ S (we expect that no confusion will arise with h·, ·i as
defined in Definition 4.1).

By Lemma 14.10 (a) it follows that u|S is a tempered distribution for all u ∈ E 0 .

Theorem 15.2. S 0 → D0 , u 7→ u|D is a sequentially continuous embedding and E 0 → S 0 ,

u 7→ u|S is a continuous embedding,

E 0 ,→ S 0 ,→ D0 .

Proof. Both maps are injective as D is dense in S, and, D and thus S are dense in E
(Theorem 5.10). The sequential continuity of the embeddings S 0 → D0 and E 0 → S 0 is
straightforward.

As in 5.13 we show in 15.11 that the embeddings in Theorem 14.11 and Theorem 15.2
are not homeomorphisms onto their image.

92
 We define the support and the operations we defined for distributions in a similar way.
For the multiplication with a smooth function we restrict to those of at most polynomial
growth. If ψ ∈ Cp∞ and u ∈ S 0 , then ϕ 7→ u(ψϕ) is again in S 0 due to Lemma 14.7.
Definition 15.3. For u ∈ S 0 we define the support of u, supp u, to be the support of the
corresponding distribution,

supp u = supp u|D .

Let y ∈ Rd , α ∈ Nd0 and l : Rd → Rd linear and bijective. For a u ∈ S 0 we define

ǔ, Ty u, ∂ α u and u ◦ l by the formulas as in Definition 2.14 but replacing “D” everywhere
by “S”.
Let σ ∈ Cp∞ . By Lemma 14.7 we have σϕ ∈ S for all ϕ ∈ S and the function S → S,
ϕ 7→ σϕ is continuous. Thus for each u ∈ S 0 the function ϕ 7→ u(σϕ) is a tempered
distribution.
For σ ∈ Cp∞ and u ∈ S 0 we define σu ∈ S 0 by

σu(ϕ) = u(σϕ) (ϕ ∈ S).

Again, it is straightforward to check that ǔ, Ty u, ∂ α u and u ◦ l are all in S 0 and

moreover that with ι : S 0 → D0 being the embedding function of S 0 into D0 ,

ι(ǔ) = ι(u)ˇ,
ι(Ty u) = Ty ι(u),
ι(∂ α u) = ∂ α ι(u),
ι(σu) = σι(u),
ι(u ◦ l) = ι(u) ◦ l.

15.4. Let y ∈ Rd , α ∈ Nd0 , σ ∈ Cp∞ and l : Rd → Rd be a linear bijection. Observe that

as in 4.4 the operations ˇ , Ty , ∂ α , multiplication by σ and composition with l, i.e.,

S → S, ϕ 7→ ϕ̌, S 0 → S 0, u 7→ ǔ,
0 0
S → S, ϕ 7→ Ty ϕ, S →S, u 7→ Ty u,
α 0 0
S → S, ϕ 7→ ∂ ϕ, S →S, u 7→ ∂ α u,
S → S, ϕ 7→ σϕ, S 0 → S 0, u 7→ σu,
0 0
S → S, ϕ 7→ ϕ ◦ l, S →S, u 7→ u ◦ l,

are continuous. We leave it to the reader to check this (as S is metrizable, it is sufficient
to show sequential continuity for the operations S → S).
15.5 (Convention). Following 10.9, we will identify elements of S 0 with their correspond-
ing distributions and elements of E 0 with their corresponding tempered distribution, e.g.,
for u ∈ S 0 we will also write “u” for “u|D ”, and, for a v ∈ E 0 we will also write “v” instead
of “v|S ”.

93
 Let us consider more examples of tempered distributions, than only the compactly
supported distributions. Each locally integrable function defines a distribution, but it
2
might not be tempered. Consider for example the function g given by g(x) = e|x| for
x ∈ Rd . In Theorem 15.6 (b) we see that certain “tempered” functions define tempered
distributions.
2
Exercise 15.A. Verify that x 7→ e|x| is not in S 0 .
Theorem 15.6. (a) Let p ∈ [1, ∞]. Lp is continuously embedded in S 0 ,

Lp ,→ S 0 .

(b) Let h : Rd → F be Borel measurable. If there exist k ∈ N0 and p ∈ [1, ∞] such that
(1 + | · |)−k h ∈ Lp , then h ∈ S 0 .

Proof. (a) Let g ∈ Lp . Let q ∈ [1, ∞] and m ∈ N be such that

1 1
+ = 1, qm > d.
p q
By Hölder’s inequality and Lemma 14.14 there exists a C > 0 such that
Z
|hg, ϕi| ≤ |gϕ| ≤ kgkLp kϕkLq ≤ CkgkLp kϕkm,S (ϕ ∈ S).

The injectivity follows from the injectivity of Lp → D0 and as D is dense in S.

(b) follows by (a) as (1 + | · |)k ∈ Cp∞ and multiplication with such a function is an
operation S 0 → S 0 .

Corollary 15.7. Let k ∈ N0 and p ∈ [1, ∞]. Then

W k,p ,→ S 0 .

Proof. This follows by Theorem 15.6 because W k,p ,→ Lp .

Observe that Theorem 15.6 implies that every Borel measurable h : Rd → F of at

most polynomial growth defines a tempered distribution.
Exercise 15.B. Let f : R → F be given by
(
log |x| x 6= 0,
f (x) =
0 x = 0.

Show that f and u defined by

ϕ(x)
Z
u(ϕ) := lim dx (ϕ ∈ S(R)),
ε↓0 R\[−ε,ε] x
define tempered distributions. (Hint: Exercise 2.E.)

94
 Exercise 15.C. Show that if f ∈ Lp for some p ∈ [1, ∞], then there exists an m ∈ N0
such that (1 + | · |)−m f ∈ L1 .
Conclude the following. If h : Rd → F is Borel measurable and there there exist
k ∈ N0 and p ∈ [1, ∞] such that (1 + | · |)−k h ∈ Lp , then there exists an m ∈ N0 such
that (1 + | · |)−m h ∈ L1 .

Exercise 15.D. Let Cp,k m be the Banach space equipped with the norm q
m,k as in
Exercise 14.E. Prove that for all k ∈ N0 and all m ∈ N0 ∪ {∞}
m
Cp,k ,→ S 0 .

Exercise 15.E. Let p ∈ [1, ∞] and k ∈ N0 . Define Lpp,k to be the space of (equivalence
classes of) Borel measurable functions g : Rd → F such that np,k (g) := k(1 + | · |)−k gkLp <
∞. Show that np,k is a norm such that Lpp,k equipped with this norm is a Banach space
and

Lpp,k ,→ S 0 .

Conclude by Exercise 15.C that there exists an m ∈ N0 such that

Lpp,k ,→ L1p,m .

For Radon measures there exist analogous statements to Theorem 15.6:

Theorem 15.8. (a) M is continuously embedded in S 0 ,

M ,→ S 0 .

(b) µ : Borel(Rd ) → [0, ∞] is a measure and there exists a k ∈ N0 such that

Z
(1 + | · |)−k dµ < ∞,

then µ defines a tempered distribution.

Proof. See Exercise 15.F.

Exercise 15.F. Prove Theorem 15.8.

Exercise 15.G. (a) Prove that for any u ∈ S 0 (R), the following are equivalent

(1) ∂u = 0,
(2) there exists a c ∈ F such that u = c1.
R R
(Hint: Let ψ ∈ S with ψ = 1. Prove that for all ϕ ∈ S, ϕ − ( ϕ)ψ has a primitive
in S.)
(b)  Prove that for all u ∈ S 0 (R) there exists a v ∈ S 0 (R) such that u = ∂v.

95
 P∞
Exercise 15.H. Show that n=1 nδn ∈ S 0.

Example 15.9. There exist Borel measurable functions g : Rd → F which define

tempered distributions but are not as in Theorem 15.6 (b), that is, for which for all
k ∈ N0 and p ∈ [1, ∞],

k(1 + | · |)−k hkLp = ∞.

Consider for example d = 1 and h = g 0 where g defined by

2
g(x) = sin(ex ).
2 2
Then g ∈ L∞ and thus g and its derivative h are in S 0 , but h(x) = 2xex cos(ex ) for all
x ∈ R so that h is not of the form as in Theorem 15.6 (b).

Theorem 15.10. (a) Let U ⊂ S 0 and assume

sup |u(ϕ)| < ∞ (ϕ ∈ S).

u∈U

Then, there exist C > 0 and m ∈ N0 such that

|u(ϕ)| ≤ Ckϕkm,S (ϕ ∈ S, u ∈ U).

(b) The space S 0 is weak* sequentially complete.

(c) The pairing map S 0 × S → F, (u, ϕ) 7→ u(ϕ) = hu, ϕi is sequentially continuous.
Consequently (by Lemma 14.7), the product map S 0 × S → S 0 , (u, ϕ) 7→ ϕu is
sequentially continuous.

Proof. Both statements (a) and (b) follow by the arguments as in the proofs of The-
orem 4.24 and Theorem 4.26 with “D(Ω)” and “DK (Ω)” both replaced by “S”. (c)
follows from (a) similarly as the proof of Proposition 4.25.

15.11. We show that (a) the relative topology of D as a subspace of S is different from
the topology on D; (b) the relative topology of S as a subspace of E is different from
the topology on E; (c) the relative topology of S 0 as a subset of D0 is not equal to the
topology of S 0 and (d) the relative topology of E 0 as a subset of S 0 is not equal to the
topology of E 0 .
We consider d = 1 for convenience.
2
(a) Let fn be the Gaussian function given by fn (x) = e−n(1+x ) for x ∈ Rd and n ∈ N.
By (14.2) limn→∞ kfn kk,S = 0 for all k ∈ N0 . Let φ ∈ D be nonzero and define
ψn = fn Tn φ. Observe that ψn ∈ D. By Lemma 14.7 kψn kk,S ≤ kφkC k kfn kk,S → 0
as n → ∞ for all k ∈ N0 . Hence ψn → 0 in S. However, by Theorem 4.11 (ψn )n∈N
does not converge in D as there is no compact set which contains the support of
ψn for all n ∈ N.

96
 (b) Let φn = n1 Tn φ for n ∈ N, where φ ∈ Cc∞ (Rd , [0, 1]) and φ(0) = 1. Then φn → 0 in
E, but (φn )n∈N does not converge in S as

sup (1 + |x|)|φn (x)| ≥ 1 (n ∈ N0 ).

x∈Rd

2
(c) en δn → 0 in D0 but not in S 0 (and not in E 0 ), indeed, for ϕ ∈ S the Gaussian
2 2
function ϕ(x) = e−x for x ∈ R, we have en δn (ϕ) = 1 for all n ∈ N.
(d) δn → 0 in S 0 but not in E 0 .

15.12. Observe that by Lemma 14.7 the following holds. If σn ∈ Cp∞ for all n ∈ N and
n→∞
σ ∈ Cp∞ and for all m ∈ N there exists a k ∈ N such that qm,k (σn − σ) −−−→ 0, then
n→∞
σn ϕ −−−→ σϕ in S (ϕ ∈ S),
n→∞ 0
σn u −−−→ σu in S (u ∈ S 0 ).

16 The Fourier transformation

In this section we consider Rd to be the space on which our functions and distributions
are defined, we therefore leave out the notation “(Rd )” in the considered function spaces
or spaces of distributions.

Definition 16.1 (Fourier transform of a function). Let f : Rd → F be an integrable

function. The Fourier transform of f , fb : Rd → C is given by
Z
fb(ξ) = e−2πihx,ξi f (x) dx, (16.1)
Rd

where hx, ξi is the inner product on Rd (the notation h·, ·i is of course also used as the
pairing between distributions, but we trust that there will be no confusing arising).
In case g is another integrable function that equals f almost everywhere, then fb =
gb. This enables us to define the Fourier transform of an element of L1 as the Fourier
transform of one of its representatives and we will use the formula (16.1) also for f ∈ L1 .

Example 16.2. Let a, b ∈ R, a < b. The Fourier transform of the indicator function
1[a,b] is given by
e−2πiaξ −e−2πibξ
(
ξ ∈ R \ {0},
1[
[a,b] (ξ) =
2πiξ
b−a ξ = 0.

Observe that 1
[[a,b] is continuous.

97
Example 16.3. The Fourier transform of the function f : R → R,

f (x) = max(1 − |x|, 0) (x ∈ R),

is given by
( (1−cos 2πξ)
2π 2 ξ 2
ξ 6= 0,
fb(ξ) =
1 ξ = 0.

Exercise 16.A. (a) Verify that 1

[[a,b] and f are given by the formulas in Example 16.2
b
and Example 16.3.
(b) Check that they are both continuous (at 0).
(c) The function sinc (also found under the name “cardinal sinus”) is defined by
(
sin x
x x 6= 0,
sinc(x) =
1 x = 0.

Prove that

1\
[− 12 , 21 ] (ξ) = sinc πξ, fb(ξ) = sinc2 πξ.

The Fourier transform of an integrable function is an element of C0 (Rd , F) (Defini-

tion 2.29), see Theorem 16.6. In order to prove that statement, we first introduce some
auxiliary lemmas.
Lemma 16.4. Let f ∈ L1 . Then for all a ∈ Rd

lim kTa f − Tx f kL1 = 0.

x→a

Proof. For ϕ ∈ Cc it holds that limx→a kTa ϕ − Tx ϕkL∞ = 0 by uniform continuity, and
therefore limx→a kTa ϕ − Tx ϕkL1 = 0. As Cc is dense in L1 , by a 3ε argument one can
finish the proof.

Lemma 16.5 (Lemma of Riemann–Lebesgue). Let g ∈ L1 (R). Then

1
|gb(a)| ≤ kg − T 1 gkL1 (a ∈ R, a 6= 0).
2 2a

Proof. Let a ∈ R, a 6= 0. As eπi = −1 we have

1 −2πia(x− 1 )
Z Z Z
g(x)e−2πiax dx = g(x − )e 2a dx = − T 1 g(x)e−2πiax dx.
R R 2a R 2a

Therefore
1
Z Z
−2πiax
g(x)e dx = [g(x) − T 1 g(x)]e−2πiax dx,
R 2 R 2a

so that the desired inequality follows.

98
Theorem 16.6. If f ∈ L1 , then fb ∈ C0 (Rd , C) and

kfbkL∞ ≤ kf kL1 .

Proof. The norm estimate is straightforward. The continuity follows by Lebesgue’s dom-
inated convergence theorem, so that f ∈ Cb (Rd , C) for all f ∈ L1 . That lim|ξ|→∞ fb(ξ) = 0
follows as by Lemma 16.5

|fb(ξ)| ≤ 21 kf − T 1
e f kL1 (i ∈ {1, . . . , d}, ξ ∈ Rd ),
2ξi i

and kf −T 1
e f kL1 converges to zero as |ξi | → ∞, for each i ∈ {1, . . . , d}, by Lemma 16.4.
2ξi i

Definition 16.7 (Fourier transformation). We write F for the linear function L1 →

C0 (Rd , C), f 7→ fb and call this map the Fourier transformation.

The Fourier transformation turns out very useful as it turns certain operations into
other operations, see for example Theorem 16.9 and Theorem 16.13.

16.8 (Notation). The symbol X is used to denote the identity map Rd → Rd .

By substitution rules for integration we obtain the following.

Theorem 16.9. Let f ∈ L1 .

(a) For y ∈ Rd

F(Ty f ) = e−2πihX,yi fb, Ty fb = F(e2πihX,yi f ). (16.2)

(b) Let l : Rd → Rd be linear and bijective. Then

1 b
F(f ◦ l) = f ◦ l∗ ,
| det l|

where l∗ is the transpose of l−1 , which means that hl−1 y, ξi = hy, l∗ (ξ)i for all
x, ξ ∈ Rd .
In particular, for λ ∈ R \ {0} (for the notation see 14.9)

F(lλ f ) = F(f (λX)) = |λ|−d l 1 fb.

λ

Exercise 16.B. Verify the statements of Theorem 16.9.

Theorem 16.10. Let f, g ∈ L1 . Then f gb, fbg ∈ L1 and

Z Z
f gb = fbg. (16.3)

99
Proof. The integrability follows by Theorem 16.6. The identity follows by Fubini’s the-
orem (Exercise 16.C).

Exercise 16.C. Check that (16.3) holds.

Definition 16.11. For f ∈ L1loc (R) we say that a function g : R → R is an indefinite

integral of f if
Z b
g(b) − g(a) = f (a, b ∈ R, a < b).
a

Observe that any continuously differentiable function g is the indefinite integral of its
derivative g 0 .

Theorem 16.12. Let g ∈ L1 (R).

(a) If Xg ∈ L1 (R), then gb is continuously differentiable and

gb0 = F(−2πiXg). (16.4)

(b) If g is an indefinite integral of a function h ∈ L1 (R), then h

b = 2πiXgb.

In particular, if g is continuously differentiable and g 0 ∈ L1 (R), then F(g 0 ) = 2πiXgb.

Proof. (a) Let a, b ∈ R, a < b. Then, by Theorem 16.10 (see also Example 16.2)
Z b Z
F(−2πiXg) = F(−2πiXg)1[a,b]
a ZR
= −2πixg(x)F(1[a,b] )(x) dx
R
e−2πibx − e−2πiax
Z
= −2πixg(x) dx
ZR
−2πix
= g(x)(e−2πibx − e−2πiax ) dx = gb(b) − gb(a).
R

As the Fourier transform of an integrable function is continuous, we conclude that gb is

continuously differentiable with derivative given by (16.4).
(b) For ξ ∈ R we have (integration by parts)
Z N
h(ξ)
b − 2πiξ gb(ξ) = lim h(x)e−2πiξx + g(x)(−2πiξ)e−2πiξx dx
N →∞ −N

= lim (g(N )e−2πiN ξ − g(−N )e2πiN ξ ).

N →∞

Therefore it suffices to show that lim|x|→∞ g(x) = R0. As g is the indefinite integral
of h, which means for example that g(y) = g(0) + 0y h(x) dx, both limy→∞ g(y) and
limy→−∞ g(y) exist. By the integrability of g, these limits need to be equal to zero.

100
 We can so to say ‘apply’ Theorem 16.12 to any of the directions in Rd , to obtain the
following.

Theorem 16.13. Let k ∈ N0 and f ∈ L1 .

(a) If Xβ f ∈ L1 for all β ∈ Nd0 with |β| ≤ k, then fb ∈ C k and

∂ β fb = F((−2πiX)β f ) (β ∈ Nd0 , |β| ≤ k).

(b) If f ∈ C k and ∂ β f ∈ L1 for all β ∈ Nd0 with |β| ≤ k, then Xβ fb ∈ C0 and

F(∂ β f ) = (2πiX)β fb (β ∈ Nd0 , |β| ≤ k).

Proof. By an induction argument it suffices to consider k = 1 and β ∈ Nd0 with |β| = 1.

Let j ∈ {1, . . . , d} be such that βj = 1, i.e., β = ej . For j ∈ {1, . . . , d} let Fj denote the
one-dimensional Fourier transformation acting on the j-th coordinate. That is
Z
Fj f (x) = e−2πixj yj f (x1 , . . . , yj , . . . , xd ) dyj .
R

Then it follows for i 6= j that Fj commutes with ∂i and with multiplication by Xj , where
Xj= hX, ej i;

∂i Fj (f ) = Fj (∂i f ), Xi Fj (f ) = Fj (Xi f ) (i, j ∈ {1, . . . , d}, i 6= j). (16.5)

(a) then follows as by Theorem 16.13 (a) the function ξj 7→ Fj f (x1 , . . . , ξj , . . . , xd ) is

continuously differentiable and
d
Fj f (x1 , . . . , ξj , . . . , xd ) = Fj ((−2πiX)β f ).
dξj

(b) follows also by the above commutation rules (16.5) and as by Theorem 16.13 (b)

Fj (∂j f ) = 2πiXj fb.

As functions of rapid decay are integrable (see for example Lemma 14.13), we obtain
the following corollary of Theorem 16.13.

Corollary 16.14. Let f : Rd → F be measurable.

(a) If f is of rapid decay, then fb is smooth.

(b) If f is smooth and ∂ β f ∈ L1 for all β ∈ Nd0 , then fb is of rapid decay.
(c) If f is a Schwartz function, then fb is a Schwartz function.

101
 The Fourier transformation actually forms a bijection S(Rd , C) → S(Rd , C), see The-
orem 16.16. In the proof of that theorem we use that the Fourier transform of a Gaussian
function is another Gaussian function:

Theorem 16.15. Let a > 0, y ∈ Rd and f : Rd → R be the Gaussian function f (x) =

2
e−a|x−y| , then f ∈ L1 and
 d
π 2 π 2 |ξ|2
fb(ξ) = e− a e−2πihy,ξi . (16.6)
a

Proof. We consider the specific case with d = 1, a = 1 and y = 0 and leave it to the
reader to prove the general case (Exercise 16.D).
2
Let g : R → R be the Gaussian function given by g(x) = e−x for x ∈ Rd . By
Theorem 16.12
d 2
 2

gb(ξ) = F(−2πiXe−X )(ξ) = πiF ∂e−X (ξ) = −2π 2 ξ gb(ξ) (ξ ∈ Rd ).
dξ
R −x2 √
By 11.16 we have gb(0) = Re dx = π. Therefore
√ 2 ξ2
gb(ξ) = πe−π (ξ ∈ Rd ),

as h = gb is the unique solution to the ordinary differential equation

(
h0 (ξ) = −2π 2 ξh(ξ),
√
h(0) = π.

Exercise 16.D. Prove Theorem 16.15.

Theorem 16.16. The Fourier transformation F forms a linear homeomorphism

S(Rd , C) → S(Rd , C) with
Z
f (x) = F(fb)(−x) = fb(ξ)e2πihx,ξi dξ (x ∈ Rd ). (16.7)
Rd

Proof. We already know that F maps S(Rd , C) into S(Rd , C), see Corollary 16.14. Let
us write “S” for “S(Rd , C)” here (or differently said, assume F = C for this proof). Let
us first prove that F forms a bijection S → S by proving (16.7). Let f ∈ S and x ∈ Rd .
Let g = T−x f . Then gb = fbe2πihx,Xi by Theorem 16.9. Therefore, it is sufficient to show
(16.7) for x = 0, which means it is sufficient to show
Z
f (0) = fb(ξ) dξ.
Rd

102
Let ht be as in Example 11.15, i.e.,
d 1 2
ht (x) = (4πt)− 2 e− 4t |x| (x ∈ Rd ).

Let gt = hbt . By Theorem 16.15 we have (take a = tπ 2 t)

2 t|x|2
gt (x) = e−4π (x ∈ Rd ),

and gbt = ht so that hbt = ht and (16.7) holds with f = ht for any t > 0. By Theorem 16.10
b

Z Z
f (x)ht (x) dx = fb(ξ)gt (ξ) dξ.
Rd Rd

As f is continuous and bounded, the left-hand side converges to f (0) as t ↓ 0 by (11.8).

As fb is an element of S it is integrable, therefore by Lebesgue’s
R b
dominated convergence
theorem we have that the right-hand side converges to Rd f (ξ) dξ as t ↓ 0, because
gt (ξ) ↑ 1 as t ↓ 0 for all ξ ∈ Rd . This proves (16.7).
Now let us prove that F defines a homeomorphism F : S → S. As the inverse is given
by the composition of F with the reflection operator R defined in 2.12, i.e., F −1 = RF ,
it is sufficient to show continuity of F . Let k ∈ N0 and α ∈ Nd0 . It is sufficient (see for
example 14.4) to show that there exists an n ∈ N0 and a C > 0 such that

k(1 + | · |2 )k ∂ α fbkL∞ ≤ Ckf kn,S (f ∈ S). (16.8)

Let us first observe the following. As L1 is continuously embedded in S (Theorem 15.6 (a)),
there exists an m ∈ N0 and a C > 0 such that

kfbkL∞ ≤ kf kL1 ≤ Ckf km,S (f ∈ S). (16.9)

By Theorem 16.13

∆ k 
  
2 k α
(1 + |ξ| ) ∂ fb(ξ) = F 1− 2 (2πiX)α f (ξ) (ξ ∈ Rd ).
4π

Therefore by (16.9)
 ∆ k  
k(1 + | · |2 )k |∂ α fbkL∞ ≤ C 1− (2πi X) α
f
4π 2 m,S

 k
∆
As multiplication with (2πiX)α and the operation 1 − 4π 2 are continuous as functions
S → S, see 15.4, there exists a C > 0 and n ∈ N0 such that (16.8).

Actually, the previous theorem extends in the following way, in the sense that the
Fourier transformation is a bijection on a larger space.

103
Theorem 16.17. Suppose that f ∈ L1 (Rd , C) is such that also fb ∈ L1 (Rd , C). Then
Z
f (x) = F(fb)(−x) = fb(ξ)e2πihx,ξi dξ for almost all x ∈ Rd . (16.10)
Rd

Consequently, F also forms a bijection

{f ∈ L1 (Rd , C) : fb ∈ L1 (Rd , C)} → {f ∈ L1 (Rd , C) : fb ∈ L1 (Rd , C)}.

Proof. For all ϕ ∈ S we have by Theorem 16.10 and Theorem 16.16.

Z Z Z Z Z
F(fb)ϕ = fbϕb = f F(ϕ)
b = f ϕ̌ = fˇϕ.
Rd Rd Rd Rd Rd

Therefore, by Lemma 2.9, we have F(fb) = fˇ almost everywhere.

16.18. Observe that by Theorem 16.6 the set {f ∈ L1 : fb ∈ L1 } is included in Cb (where

Cb is viewed as subset of L∞ ).

Example 16.19. Let f be as in Example 16.3. Then |fb(x)| ≤ ξ −2 , so that fb ∈ L1 (R).

By Theorem 16.17 fb(a) = f (−a) for a ∈ R. By taking a = 0 we obtain
b

(1 − cos 2πξ)
Z
dξ = 1,
R 2π 2 ξ 2
and thus
(1 − cos x)
Z
dx = π,
R x2

We mentioned fb in Example 16.19 but not 1\ 1

[−1,1] as it is not in L . In Example 16.28
we come back to this.

Exercise 16.E. Prove

cos(xs)
Z
dx = πe−|s| (s ∈ R).
R 1 + x2

Exercise 16.F. Calculate the Fourier transform of the function f : R → R given by

f (x) = e−|x| for x ∈ R.

16.20. For f ∈ L1 (Rd , C) and ϕ ∈ S(Rd , C) we have by Theorem 16.10

Z Z
fbϕ = f ϕ.
b

So that with the notation of 2.6, we have

ufb(ϕ) = uf (ϕ).
b

104
As the Fourier transformation is a continuous function S(Rd , C) → S(Rd , C) it is natural
to define the Fourier transform of u to be the tempered distribution ϕ 7→ u(ϕ). b We give
the definition in Definition 16.22, but first discuss the situation for F = R. As in that
case, if we have a u ∈ S 0 (Rd , R), that is a continuous linear u : S(Rd , R) → R, then we
a priori are not able to pair u with ϕb for a ϕ ∈ S(Rd , R) as the Fourier transform of ϕ
might attain non-real values (take for example a Gaussian function as in Theorem 16.15
with y 6= 0).
Let us show how we overcome this situation. First of all, let us assume that u is
represented by a locally integrable function f : Rd → R. Observe that for any Schwartz
function ϕ ∈ S(Rd , C) the functions <ϕ, =ϕ are Schwartz functions, where
(<ϕ)(x) = <(ϕ(x)), (=ϕ)(x) = =(ϕ(x)) (x ∈ Rd ),
and <a and =a are the real and imaginary part of a, for a ∈ C. Let us for the moment be
extra careful and write g for the locally integrable function Rd → C given by g(x) = f (x)
for Rd . We can pair g with Schwartz functions, and have
Z Z Z
hg, ϕi = gϕ = f <ϕ + i f =ϕ = hf, <ϕi + ihf, =ϕi (ϕ ∈ S(Rd , C)).

By this identity it is clear that g and therefore f corresponds to a tempered distribution

in S 0 (Rd , C). We can use the above identity to generalise this to general tempered
distributions:
Definition 16.21. For any u ∈ S 0 (Rd , R) we define its complex extension uC ∈ S 0 (Rd , C)
by
uC (ϕ) = u(<ϕ) + iu(=ϕ) (ϕ ∈ S(Rd , C)).

As it is common to identify the function f : Rd → R with the function g : Rd → C

with g(x) = f (x) for all x ∈ Rd , we identify a u ∈ S 0 (Rd , R) with uC in the sense if
“u(ϕ)” is written for some ϕ ∈ S(Rd , C), then it is interpreted as “uC (ϕ)”.
Definition 16.22. Let u ∈ S 0 . We define the Fourier transform of u, u
b by

u
b(ϕ) = u(ϕ)
b (ϕ ∈ S).
(As mentioned in Definition 16.21 for F = R we interpret “u(ϕ)”
b to be “uC (ϕ)”.)
b
From here on we write F for the map S 0 → S 0 , u 7→ u
b.
Example 16.23. The function 1 represents a tempered distribution, and so does δ0 . We
calculate their Fourier transforms. For ϕ ∈ S we have
Z
hδ0 , ϕi = δ0 (ϕ) = ϕ(0) = ϕ = h1, ϕi,
b b b
Z
h1
b , ϕi = ϕ b = ϕ(0) = hδ0 , ϕi,

where we used the inversion formula in the second line. Hence

δb0 = 1, 1b = δ0 .

105
 Exercise 16.G. Let µ ∈ M(Rd , C). It is customary to define the Fourier transform of
µ to be the function u
b given by
Z
u
b(ξ) = e−2πihx,ξi dµ(x) (ξ ∈ Rd ).
Rd

Show that this definition is consistent with the definition of the Fourier transform of uµ ,
that is, of µ as a tempered distribution. Moreover, prove that ub is bounded and uniformly
continuous.

The following theorem is a consequence of Theorem 16.9, Theorem 16.13 and The-
orem 16.16.

Theorem 16.24. The Fourier transformation F : S 0 (Rd , C) → S 0 (Rd , C), u 7→ u

b is a
linear homeomorphism. Moreover, with R the reflection operator defined as in 2.12

F −1 = FR = RF,

and for u ∈ S 0 , β ∈ Nd0 , y ∈ Rd , l : Rd → Rd a linear bijection and λ ∈ R,

F(∂ β u) = (2πiX)β u
b, ∂β u
b = F((−2πiX)β u), (16.11)
−2πihX,yi 2πihX,yi
F(Ty u) = e u
b, Ty u
b = F(e u), (16.12)
1 1
F(u ◦ l) = b ◦ l∗ ,
u F(lλ u) = l1 u
b, (16.13)
| det l| |λ|d λ

where l∗ is the transpose of l−1 as in Theorem 16.9 and as in 14.9 “lλ u” is written for
“u ◦ lλ ”.

Exercise 16.H. Prove the following.

(a) If v ∈ S 0 (R) and Xv = 0, then there exists a c ∈ F such that v = cδ0 . (Hint:
Exercise 15.G.)
(b) If v ∈ S 0 (R) and Xv = 1, then there exists a c ∈ F such that v = cδ0 + u, where u
is as in Exercise 15.B, i.e.,
ϕ(x)
Z
u(ϕ) = lim dx (ϕ ∈ S(R)).
ε↓0 R\[−ε,ε] x

Exercise 16.I. Determine the Fourier transform in S 0 (R, C) of the following distribu-
tions:

(a) X2 .

(b) 1[0,∞) . (Hint: Exercise 8.E and Exercise 16.H, and 1(−∞,0] = R1[0,∞) .)
(c) 1(−∞,a) , for a ∈ R.

Exercise 16.J. (a) Give an example of a u ∈ S(Rd , C) such that u 6= 0 and u = u

b.

106
 (b) Give an example of a u ∈ S 0 (Rd , C) \ S(Rd , C) such that u 6= 0 and u = u
b. (Hint:
Example 16.23.)
(c) Suppose u ∈ S 0 and u
b = cu for some c ∈ C, what can be said about c?

Exercise 16.K. Prove that in S 0 (Rd , C)

(1 − ∆)e−|X| = 2δ0 .

Observe that this proves that 21 e−|X| is a fundamental solution to (1 − ∆).

Definition 16.25. The inverse of the Fourier transformation on S 0 (Rd , C), F −1 , is called
the inverse Fourier transformation. For an u ∈ S 0 (Rd , C) we call F −1 (u) the Fourier
inverse of u.

16.26. Let us write f for the complex conjugate of a function f : Rd → C, i.e., f (x) =
<(f (x)) − i=(f (x)) for x ∈ Rd . Then

ˇ
fb = fˇ (f ∈ L1 (Rd )).
b
fb = fb

Therefore, as a consequence of Theorem 16.10 we have

Z Z Z
hf, ϕi
b L2 = f ϕb = f ϕ̌
b
= fbϕ̌ = hfb, ϕ̌iL2 (f ∈ L1 , ϕ ∈ S). (16.14)

By the above observation and the Fourier inversion formula, we obtain the following
identity, which is due to Parseval and Plancherel.

Theorem 16.27 (Parseval, Plancherel). F forms an isometric isomorphism

L2 (Rd , C) → L2 (Rd , C),

so that in particular

kfbkL2 = kf kL2 (f ∈ L2 (Rd , C)), (16.15)

and that the Fourier inversion formula (16.10) holds for any f ∈ L2 (Rd , C).

Proof. We refrain from writing “(Rd , C)” in this proof. As S is dense in L2 , see Lemma 14.14,
it turns out, as we argue later, that it is sufficient to show

kϕk
b L2 = kϕkL2 (ϕ ∈ S). (16.16)

Let ϕ ∈ S. By applying (16.14) with f = ϕb and using that F ϕb = ϕ̌,

b 2L2 = hϕ,
kϕk b ϕ̌iL2 = hϕ, ϕiL2 = kϕk2L2 .
b L2 = hF(ϕ),
b ϕi

107
As the Fourier transform is bijective on S 0 , it is sufficient to show that the Fourier
transform of any L2 function is given by an L2 function. Now take any f ∈ L2 and let
(ϕn )n∈N be a sequence in S that converges to f in L2 :

kϕn − f kL2 → 0.

By (16.16) it follows that (ϕbn )n∈N is a Cauchy sequence in L2 and thus converges in L2
to some g. As L2 is continuously embedded in S 0 (Theorem 15.6 (a)), it follows that
g = fb in L2 .

Example 16.28. As 1 b 1 1 = sinc π X is in L2 , see Example 16.2 and Exercise 16.A, by

[− 2 , 2 ]
Theorem 16.27 we have

F(sinc) = π 1[− 1 , 1 ] (π X).

2 2

17 Convolution of tempered distributions

We have already seen that the Fourier transform turns certain operations into other
operations, like differentiation become multiplication with polynomials. In this section
the Fourier transform is used to consider convolutions, namely it turns the operation of
convolution into the operation of multiplication.

Theorem 17.1. Let f, g ∈ L1 . Then f ∗ g ∈ L1 and

F(f ∗ g) = fbgb.

Proof. By Young’s inequality, Theorem 7.7 we have f ∗ g ∈ L1 . Therefore, by Fubini’s

theorem, we have for ξ ∈ Rd ,
Z
F(f ∗ g)(ξ) = f ∗ g(x)e−2πihx,ξi dx
d
ZR Z
= f (y)g(x − y) dy e−2πihx,ξi dx
Z Rd Rd Z
= f (y) g(x − y)e−2πihx,ξi dx dy
d d
ZR ZR
= f (y) g(x)e−2πihx+y,ξi dx dy
d Rd
ZR
= f (y)gb(ξ)e−2πihy,ξi dy = fb(ξ)gb(ξ),
Rd

where we used Theorem 16.9 (a).

Exercise 17.A. Show that 1[− 1 , 1 ] ∗1[− 1 , 1 ] = f , where f is as in Example 16.3. Observe
2 2 2 2

that indeed fb = (1\ 2

[− 1 , 1 ] ) , see also Exercise 16.A.
2 2

108
 As a direct consequence, by Theorem 16.16 and as multiplication is a continuous
operation S × S → S (see Lemma 14.7):
Theorem 17.2. Let ϕ, ψ ∈ S. Then

F(ϕ ∗ ψ) = ϕbψ,
b F(ϕψ) = ϕb ∗ ψ.
b (17.1)

Consequently, ϕ ∗ ψ ∈ S and the function S × S → S, (f, g) 7→ f ∗ g is continuous.

Analogously to Definition 8.1 we define the convolution between elements of S 0 and

S as follows.
Definition 17.3. Let u ∈ S 0 and ϕ ∈ S. We define the convolution of u with ϕ to be
the function Rd → F defined by

u ∗ ϕ(x) = u(Tx ϕ̌) (x ∈ Rd ).

The following is the analogues statement of Lemma 8.2.

Lemma 17.4. Let (u, ϕ) be in S 0 (Rd ) × S(Rd ). Then

δ0 ∗ ϕ = ϕ,
δy ∗ ϕ = Ty ϕ (y ∈ Rd ),
R(u ∗ ϕ) = Ru ∗ Rϕ,
Ty (u ∗ ϕ) = (Ty u) ∗ ϕ = u ∗ (Ty ϕ),
u(ϕ) = u ∗ ϕ̌(0).

Proof. The proof is left to the reader.

Similarly to Theorem 8.4 we have that the convolution between a Schwartz function
and a tempered distribution is smooth, as we will see in Theorem 17.6. However, it need
not be a Schwartz function as will be clear from the following exercise.
Exercise 17.B. Compute the convolution of the tempered distribution 1 with the
2
Schwartz function e−|X| .

Let us consider the convergence of difference quotients as we did in Lemma 8.3.

Lemma 17.5. Let ϕ ∈ S(Rd ) and j ∈ {1, . . . , d}. Then
T0 − Thej
 
h→0
ϕ −−−→ ∂j ϕ in S(Rd ). (17.2)
h

Consequently, for u ∈ S 0 (Rd ),

T0 − Thej
 
h→0
u −−−→ ∂j u in S 0 (Rd ).
h

109
Proof. Let us write Xj = hX, ej i for j ∈ {1, . . . , d}. Observe that for ϕ ∈ S, j ∈ {1, . . . , d}
and h ∈ R \ {0}
  T−he − T0  
j e2πihhej ,Xi − 1
F ϕ = ϕ,
b F(∂j ϕ) = 2πiXj ϕ.
b
h h
Therefore (17.2) holds if and only if the following convergence holds in S, where
!
e2πihXj − 1 h→0
− 2πiXj ϕb −−−→ 0. (17.3)
h

By Lemma 14.7 (see (14.5)), the latter is the case if there exists a k ∈ N0 such that
!
e2πihXj − 1 h→0
qm,k − 2πiXj −−−→ 0 (m ∈ N0 ). (17.4)
h

Let us consider the function R → C given by

eiht − 1
gh (t) = − it (t ∈ R),
h
2πihξj
so that gh (2πξj ) = e h −1 − 2πiξj for all ξ ∈ Rd . Now (17.4) follows for k = 2, by the
following –which we will show–

2
|h||t| n = 0,

dn

gh (t) ≤ |h||t| n = 1, (h ∈ R \ {0}, t ∈ R).
dtn 
|h|n−1 n≥2


For n ≥ 2 the above inequality follows as

dn
gh (t) = in hn−1 eiht (n ≥ 2).
dtn
By rewriting gh (t) as
R h irt R h R r iut
eiht − 1 0 e − 1 dr 2 0 0 e du dr
gh (t) = − it = it = (it) ,
h h h
we see that the above inequality for n = 0 holds and also that
Z h
gh0 (t) = i(eiht − 1) = it eirt dr,
0

from which the inequality for n = 1 follows.

Theorem 17.6. Let u ∈ S 0 and ϕ ∈ S. Then u ∗ ϕ is smooth and of at most polynomial

growth, that is u ∗ ϕ ∈ Cp∞ . Strictly speaking, u ∗ ϕ is represented by a function in Cp∞ .
For all α ∈ Nd0

∂ α (u ∗ ϕ) = u ∗ (∂ α ϕ) = (∂ α u) ∗ ϕ. (17.5)

110
Moreover, if C > 0, k ∈ N0 and u ∈ S 0 are such that

|u(ψ)| ≤ Ckψkk,S (ψ ∈ S), (17.6)

then (with qm,k as in (14.4))

qm,k (u ∗ ϕ) ≤ Ckϕkm+k,S (ϕ ∈ S). (17.7)

Proof. That u ∗ ϕ is smooth and that (17.5) holds follow by Lemma 17.5 as in the proof
of Theorem 8.4. As (see Exercise 17.C)

kTx ϕ̌km,S ≤ (1 + |x|)m kϕkm,S (x ∈ Rd , m ∈ N0 ), (17.8)

we have for u ∈ S 0 such that (17.6),

|u ∗ ϕ(x)| = |u(Tx ϕ̌)| ≤ C(1 + |x|)k kϕkk,S (x ∈ Rd ).

By (17.5) we obtain (17.7).

Exercise 17.C. Prove (17.8).

Theorem 17.7. Let u ∈ S 0 and ϕ ∈ S. Then in S 0

F(u ∗ ϕ) = ϕbu
b, F(ϕu) = u
b ∗ ϕ.
b

Strictly speaking, F(ϕu) is the tempered distribution represented by the function u

b ∗ ϕ.
b

Proof. As ϕb is a Schwartz function, ϕbu

b is a tempered distribution. Recall that in 8.10
we have observed that

hu ∗ η, ψi = hu, η̌ ∗ ψi (η, ψ ∈ D).

Therefore, if ϕ ∈ D and ψ ∈ S is such that ψb ∈ D, then as ϕ̌ = ϕ

bb

hF(u ∗ ϕ), ψi = hu ∗ ϕ, ψi
b = hu, ϕ̌ ∗ ψi
b = hu, ϕ
bb ∗ ψi
b
= hu, F(ϕψ)i
b = hu b = hϕ
b, ϕψi bub, ψi.

As D is dense in S, the Fourier transformation is a homeomorphism, also {ψ ∈ S : ψb ∈ D}

is dense in S. Therefore, using that ϕ 7→ u ∗ ϕ is continuous as a function S → S 0 , we
obtain hF(u ∗ ϕ), ψi = hϕbub, ψi for all ϕ, ψ ∈ S. The identity F(ϕu) = u
b∗ϕ
b then follows
by the first (Exercise 17.D).

Exercise 17.D. Prove that the identity F(ϕu) = ϕb ∗ u b holds for all u ∈ S 0 and ϕ ∈ S
b holds for all u ∈ S 0 and ϕ ∈ S.
by using that F(u ∗ ϕ) = ϕbu

As we mentioned in 15.5, we identify elements of v ∈ E 0 with their corresponding

tempered distribution with compact support, v|S . For the Fourier transform of v|S we
therefore also write vb.

111
Lemma 17.8. If v ∈ E 0 , then vb ∈ Cp∞ . Consequently, if u ∈ S 0 and supp u
b is compact,
∞ ∞
then u ∈ Cp (in the sense that u is represented by a function in Cp ).
Moreover, if v ∈ E 0 and ϕ ∈ S, then v ∗ ϕ ∈ S.

Proof. Let χ ∈ D be such that v = χv (see 5.6). Then

vb = F(χv) = χ
b ∗ vb.

b Therefore vb ∈ Cp∞ by Theorem 17.6.

As χ is a Schwartz function, so is χ.
For the “Moreover” part: for v ∈ E 0 and ϕ ∈ S, then vb ∈ Cp∞ and therefore vbϕb and
v ∗ ϕ = F −1 (vbϕ)
b are in S.

17.9. Let k ∈ N0 and K ⊂ Rd be compact. Let

C = sup (1 + |x|)k .
x∈K

Then

kψkC m ,K ≤ Cqm,k (ψ) (ψ ∈ E, m ∈ N0 ). (17.9)

Therefore, we can already conclude by Theorem 17.6 that for u ∈ S 0 , the function
S → E, ϕ 7→ u∗ϕ is continuous. Moreover, see for example Exercise 15.D also the function
S → S 0 , ϕ 7→ u ∗ ϕ is continuous. In Theorem 17.10 we prove that the convolution is
continuous as a function of both the tempered distribution and the Schwartz function:

Theorem 17.10. (a) Let u ∈ S 0 , ϕ ∈ S, (un )n∈N be a sequence in S 0 , (ϕn )n∈N be a

sequence in S and suppose

(un , ϕn ) → (u, ϕ) in S 0 × S.

Then there exists a k ∈ N0 such that

n→∞
qm,k (un ∗ ϕn − u ∗ ϕ) −−−→ 0 (m ∈ N0 ). (17.10)

(b) The functions

S 0 × S → S 0, (u, ϕ) 7→ u ∗ ϕ, (17.11)
0
S × S → E, (u, ϕ) 7→ u ∗ ϕ, (17.12)
0
E × S → S, (v, ϕ) 7→ v ∗ ϕ, (17.13)

are sequentially continuous.

(c) If (vn )n∈N and v are in E 0 and vn → v in E 0 , then there exists a k ∈ N0 such that
n→∞
qm,k (vbn − vb) −−−→ 0 (m ∈ N0 ).

112
Proof. (a) Write u∞ = u. As un → u in S 0 , by Theorem 15.10 (a) there exist C > 0 and
k ∈ N0 such that

|un (ψ)| ≤ Ckψkk,S (ψ ∈ S, n ∈ N ∪ {∞}). (17.14)

We show (17.10) by showing by showing

qm,k (un ∗ (ϕn − ϕ)) → 0, (17.15)

qm,k ((un − u) ∗ ϕ) → 0. (17.16)

(17.15) follows by (17.14) and Theorem 17.6. To prove (17.16) we use Lemma 4.8:
Without loss of generality we may assume K to be convex (as we can always choose
a larger compact convex set). First observe, that as un → u in S 0 , we have ∂ α (un − u) ∗
ϕ(x) → 0 for all x ∈ K and α ∈ Nd0 , |α| ≤ m. Therefore, by Lemma 4.8, (17.16) follows
when ((1 + | · |2 )−k ∂ α (un − u) ∗ ϕ)n∈N is a sequence of uniformly Lipschitz continuous
functions on K for all α ∈ Nd0 with |α| ≤ m. For this, by Lemma 4.7, it is sufficient to
show
d
M := sup max max k∂i [(1 + | · |2 )−k (∂ α (un − u) ∗ ϕ]kL∞ < ∞.
n∈N α∈Nd0 i=1
|α|≤m

As

∂i [(1 + | · |2 )−k ≤ 2k(1 + | · |2 )−k ,

L∞

we have, using Theorem 17.6 and (17.14),

M
≤ max [(1 + | · |2 )−k (∂ α (un − u) ∗ ϕ]
2k + 1 α∈Nd0 L∞
|α|≤m+1

= sup qm+1,k ((un − u) ∗ ϕ) ≤ 2Ckϕkm+1+k,S .

n∈N

(b) That the functions (17.11) and (17.12) are sequentially continuous follows by
Exercise 15.D and the estimate (17.9) in 17.9. For the sequential continuity of the function
(17.13) it is sufficient to prove (c) due to the identity v ∗ ϕ = F −1 (vbϕ)
b (Theorem 17.7),
Lemma 14.7 and the continuity of the Fourier transform on S(Rd , C) (Theorem 16.16).
(c) Let χ ∈ E be such that v = χv and thus vb = χ b ∗ vb. By continuity of the product
map E 0 × E → E (Proposition 5.3) it follows that χvn → χv = v and (1 − χ)vn → 0 in E 0 .
The latter implies (1 − χ)vn → 0 in S 0 and thus F((1 − χ)vn ) → 0. Therefore, we may
as well assume that vn = χvn for all n ∈ N. As vbn → vb in S 0 and χ b ∈ S, by (a) there
n→∞
exists a k ∈ N0 such that qm,k (vbn ∗ χ
b − vb ∗ χ) −−−→ 0 for all m ∈ N0 .

As in Theorem 8.9 we have the following associativity rule.

113
Theorem 17.11. If u ∈ S 0 and ϕ, ψ ∈ S then

u ∗ (ϕ ∗ ψ) = (u ∗ ϕ) ∗ ψ. (17.17)

Proof. By Theorem 17.7 and Theorem 17.2 we have

F(u ∗ (ϕ ∗ ψ)) = F(ϕ ∗ ψ)u

b=ϕ
bψbu
b,
F((u ∗ ϕ) ∗ ψ) = ψF(u
b ∗ ϕ) = ψbϕbu
b.

As the Fourier transformation is injective on S 0 (Theorem 16.24), we have (17.17).

In Definition 10.5 we defined the convolution between distributions u and v, of which

at least one has compact support, by

u ∗ v(ϕ) = u(v̌ ∗ ϕ) (ϕ ∈ D).

Suppose that u ∈ S 0 and v ∈ E 0 . Then the functions S → F given by ϕ 7→ u(v̌ ∗ ϕ) and

ϕ 7→ v(ǔ ∗ ϕ) are tempered distributions as they are linear and sequentially continuous
by Theorem 17.10 (b) (because S → S, ϕ 7→ v̌ ∗ ϕ and S → E, ϕ 7→ ǔ ∗ ϕ are sequentially
continuous). On other words, the distributions u ∗ v and v ∗ u as in Definition 10.5 extend
to tempered distributions, for which we use the same notation:
Definition 17.12. For u ∈ S 0 and v ∈ E 0 we define u ∗ v to be the tempered distribution
given by

u ∗ v(ϕ) = u(v̌ ∗ ϕ) (ϕ ∈ S).

Moreover, we define v ∗ u to be the tempered distribution

v ∗ u(ϕ) = v(ǔ ∗ ϕ) (ϕ ∈ S).

Theorem 17.13. Let u ∈ S 0 , v ∈ E 0 and ϕ ∈ S.

(u ∗ v) ∗ ϕ = u ∗ (v ∗ ϕ) = v ∗ (u ∗ ϕ) = (v ∗ u) ∗ ϕ, (17.18)
u ∗ v = v ∗ u, (17.19)
F(u ∗ v) = vbu
b. (17.20)

Proof. (17.18) holds in case ϕ ∈ D, by Theorem 10.6. As D is dense in S (Theorem 14.12),

by the continuity of the function S → E, ϕ 7→ w∗ϕ for w being either one of the tempered
distributions u ∗ v, u and v ∗ u, and the continuity of the function S → S, ϕ 7→ v ∗ ϕ one
then obtains (17.18) and thus (17.19).
For ϕ ∈ S we have, using FF = R,

hF(u ∗ v), ϕi = hu ∗ v, ϕi
b = u ∗ v ∗ ϕ(0)
b̌ = hu, Rv ∗ F(ϕ)i
= hu, F(vb) ∗ F(ϕ)i = hu, F(vbϕ)i = hu
b, vbϕi = hvbu
b, ϕi,

so that we conclude (17.20).

114
 The relations in Lemma 8.2 between the reflection and translation extend naturally
to the convolution operation between elements of E 0 and S 0 , we summarize:

Lemma 17.14. Let (u, v) ∈ (E 0 × S 0 ) ∪ (S 0 × E 0 ). Then

δ0 ∗ u = u,
δy ∗ u = Ty u (y ∈ Rd ),
R(u ∗ v) = R(u) ∗ R(v),
Ty (u ∗ v) = (Ty u) ∗ v = u ∗ (Ty v) (y ∈ Rd ),
∂ α (u ∗ v) = (∂ α u) ∗ v = u ∗ (∂ α v) (α ∈ Nd0 ).

Proof. The proof is left for the reader.

The operation of convolution is like in Theorem 10.2 characterised by maps S → E

that is sequentially continuous and commute with translations. The proof is similar to
the proof of Theorem 10.2, and left for the reader.

Theorem 17.15. Let A : S → E be linear. Then A is sequentially continuous and

commutes with translations if and only if there exists a u ∈ S 0 such that Aϕ = u ∗ ϕ for
all ϕ ∈ S.
If A is sequentially continuous and commutes with translations, then there exists
exactly one such u such that Aϕ = u ∗ ϕ for all ϕ ∈ S.

Lemma 17.16. The function S 0 × E 0 → S 0 , (u, v) 7→ u ∗ v is sequentially continuous.

Proof. The proof is left for the reader (Exercise 17.E).

Exercise 17.E. Prove Lemma 17.16.

Exercise 17.F. Suppose u : [0, ∞) × Rd → R satisfies

(
∂† u = ∆u on (0, ∞) × Rd ,
u(0, ·) = u0 ,

for some u0 ∈ S. Take the Fourier transform of the space variable, that is, let U (t, ξ) =
F(u(t, X))(ξ) for (t, ξ) ∈ [0, ∞) × Rd . Derive
2 t|ξ|2
U (t, ξ) = e−4π U (0, ξ) ((t, ξ) ∈ (0, ∞) × Rd ).

Conclude that

u(t, x) = ht ∗ u0 (x) ((t, x) ∈ (0, ∞) × Rd ),

d 1 2
where ht (x) = (4πt)− 2 e− 4t |x| for (t, x) ∈ (0, ∞) × Rd , as in Example 11.15.

115
18  The Fourier transformation on E 0

In the previous section we have seen that the Fourier transformation forms a bijection
S(Rd , C) → S(Rd , C) and S 0 (Rd , C) → S 0 (Rd , C). Moreover, Example 16.2 and Ex-
ample 16.3 illustrate that the Fourier transform of a compactly supported distribution
may not be compactly supported. In this section we show that the only compactly
supported distribution whose Fourier transform is compact is the zero distribution (The-
orem 18.7). This implies D ∩ F(D) = {0} and E 0 ∩ F(E 0 ) = {0}. Moreover, as we will see,
one can say much more about Fourier transforms of compactly supported distributions
than we have seen in Lemma 17.8.

Definition 18.1. An entire function is a function g : Cd → C, given by

∞ X
X
g(z) = cα z α (z ∈ Cd ),
k=0 α∈Nd
0
|α|=k

where (cα )α∈Nd is a family of complex numbers satisfying

0

∞ X
|cα |R|α| < ∞ for all R > 0.
X

k=0 α∈Nd
0
|α|=k

Lemma 18.2. With ha, bi denoting the inner product of a and b in Cd , for a, b ∈ Cd ,
∞ X α
aα b
eha,bi =
X
.
k=0 α∈Nd
α!
0
|α|=k

Proof. Write eha,bi = ea1 b1 · · · ead bd and use the power series representation of the expo-
nential function:
∞ ∞
(a1 b1 )α1 · · · (ad bd )αd
eha,bi =
X X
··· .
α1 =0 αd =0
α1 ! · · · αd !

Lemma 18.3. Let v ∈ E 0 . Define g : Cd → C by

g(z) = hv, ehz,Xi i (z ∈ Cd ).

Then g is an entire function.

116
Proof. Let m ∈ N0 be such that v is at most of order m. First observe
∞ X
1
kXα kC m ,supp v R|α| < ∞
X
(R > 0). (18.1)
k=0 α∈Nd
α!
0
|α|=k

Therefore in particular

K X ∞ X
X z α Xα X z α Xα K→∞
− −−−−→ 0 (z ∈ Cd ).
k=0 α∈Nd
α! k=0 α∈Nd
α!
0 0
|α|=k |α|=k C m ,supp v

Therefore, for z ∈ Cd we have by Lemma 18.2,

∞ X ∞ X
* +
X z α Xα X hv, Xα i α
g(z) = v, = z ,
k=0 α∈Nd
α! k=0 α∈Nd
α!
0 0
|α|=k |α|=k

whereas, for all R ∈ [0, ∞),

∞ X
X |hv, Xα i| |α|
R < ∞,
k=0 α∈Nd
α!
0
|α|=k

by (18.1).

Theorem 18.4 (Paley-Wiener). If v ∈ E 0 and

g(z) = hv, e−2πihz,Xi i (z ∈ Cd ), (18.2)
then g is entire and g|Rd equals vb in S 0 .

Proof. By Lemma 18.3 g is entire. In order to prove that g|Rd equals vb in S 0 , let χ ∈ D
be equal to 1 on a neighbourhood of supp v, so that v = χv. It suffices to prove that g
b̌ = F −1 (χ), for all x ∈ Rd :
b on Rd . As χ
equals vb ∗ χ
vb ∗ χ(x)
b = hvb, Tx F −1 (χ)i = hv, FTx F −1 (χ)i = hv, e−2πihx,Xi χi
= hχv, e−2πihx,Xi i = hv, e−2πihx,Xi i = g(x).

Definition 18.5. Let Ω ⊂ Rd be open. A function f : Ω → F is called analytic if for

each y ∈ Rd there exist r > 0 and (cα )α∈Nd in F such that ∞ α∈Nd0 ,|α|=k cα (x − y)
α
P P
0
k=0
exists for all x ∈ B(y, r) and
∞ X
X
f (x) = cα (x − y)α (x ∈ B(y, r)). (18.3)
k=0 α∈Nd0
|α|=k

117
 Every analytic function is smooth and if (18.3) holds, then ∂ α f (y) = cα α!.
If g : Cd → C is an entire function, then f : Rd → C given by f = g|Rd , i.e.,
f (x) = g(x) for x ∈ Rd , is an analytic function.

Lemma 18.6. Let Ω be a connected and open subset of Rd . Let f : Ω → F be an analytic

function. Then either f = 0 or supp f = Ω.

Proof. Let U = Ω \ supp f . For y ∈ Ω, y ∈ U if and only if there exists an r > 0 such
that f = 0 on B(y, r). As Ω is connected, it is sufficient to show that supp f is open.
Let y ∈ supp f and let r > 0 and (cα )α∈Nd in F be such that (18.3). Then there exists
0
an α ∈ Nd0 such that cα 6= 0 and thus ∂ α f (y) = cα 6= 0. As ∂ α f is continuous, there
exists an s > 0 such that ∂ α f 6= 0 on B(y, s). Then f cannot be equal to 0 on B(y, s).
Therefore B(y, s) ⊂ supp f .

Theorem 18.7. If v ∈ E 0 and vb has compact support, then v = 0.

Proof. We are done if vb = 0, which in the language of Theorem 18.4 is the case if g = 0
on Rd . As g is entire g|Rd is analytic and therefore this follows by Lemma 18.6.

Actually, Theorem 18.4 is only a small part of the statements of the Paley-Wiener
theorem, which characterizes the Fourier transforms of compactly supported distributions
explicitly. We pose the statements here and refer to [Rud91, Theorem 7.23] for a proof.

Theorem 18.8 (Paley-Wiener).

(a) If v ∈ E 0 , R > 0, supp v ⊂ B(0, R), v has order k and

g(z) = hv, e−2πihz,Xi i (z ∈ Cd ), (18.4)

then g is entire, g|Rd = vb and there exists a C > 0 such that

|g(z)| ≤ C(1 + |z|)k eR|=z| (z ∈ Cd ). (18.5)

(b) Conversely, if g is an entire function on Cd which satisfies (18.5) for some k ∈ N0

and C > 0, then there exists a v ∈ E 0 with support in B(0, R) such that (18.2)
holds.

19 Fourier multipliers

Let us motivate the definition of Fourier multipliers by recalling some facts from the
previous sections. By Theorem 16.24 we have

∂ β u = F −1 ((2πiX)β u
b) (β ∈ Nd0 , u ∈ S 0 ).

118
Therefore, by writing
1
Dβ = ∂β (β ∈ Nd0 ),
(2πi)β

we have

Dβ u = F −1 (Xβ u
b) (β ∈ Nd0 , u ∈ S 0 ).

Moreover, for any polynomial p, with p(D) the polynomial p “evaluated in D” (as in
Definition 11.1), so that Xα (D) = Dα ,

p(D)u = F −1 (pu
b) pu = F −1 (p(D)u
b) (u ∈ S 0 ).

Also, we have

Ty u = F −1 (e2πihX,yi u
b) (y ∈ Rd , u ∈ S 0 ).

So both operations Dβ and Ty can be described by the composition of the Fourier inverse
with the multiplication of the Fourier transform with a certain function, namely with
Xβ and e2πihX,yi , respectively. In this section we consider those operators of the form
u 7→ F −1 (σ u
b) for a certain class of functions σ such that σ u
b is again tempered and hence
in the domain of F −1 . Those operators are called Fourier multipliers.

Definition 19.1 (Fourier multiplier). For σ ∈ Cp∞ we define σ(D) : S 0 → S 0 by

σ(D)u = F −1 (σ u
b) (u ∈ S 0 ),

and call σ(D) a Fourier multiplier.

Lemma 19.2. Let σ ∈ Cp∞ . Then σ(D) is continuous as function S 0 → S 0 and it forms
a continuous function S → S. Moreover,

hσ(D)u, ϕi = hu, σ̌(D)ϕi (u ∈ S 0 , ϕ ∈ S).

Proof. As multiplication with σ is a continuous operation S 0 → S 0 and S → S (see

15.4) and because F is continuous as a function S 0 → S 0 and as a function S → S
(Theorem 16.16 and Theorem 16.24), it follows that σ(D) is continuous as a function
S 0 → S 0 and as a function S → S.
For σ ∈ Cp∞ , u ∈ S 0 and ϕ ∈ S

b), Fϕi = hσ̌F −1 (u), ϕi

hσ(D)u, ϕi = hR(σ u b = hu, F −1 (σ̌ ϕii
b
= hu, σ̌(D)ϕi.

119
Example 19.3. By Theorem 16.24 we have ∂ β = σ(D) for σ = (2πiX)β and Ty = σ(D)
for σ = e−2πihX,yi , i.e.,

∂ β u = (2πiX)β (D)u, Ty u = e−2πihX,yi (D)u (β ∈ Nd0 , y ∈ Rd , u ∈ S 0 ). (19.1)

Let v ∈ E 0 . In Lemma 17.8 we have seen that vb ∈ Cp∞ . Therefore

v ∗ u = vb(D)u (u ∈ S 0 ).

By the commutativity of multiplication, we obtain that Fourier multipliers commute

as well. Moreover, if F −1 (σ) is compactly supported, then the Fourier multiplier σ(D)
equals convolution with F −1 (σ):

Lemma 19.4. Let σ, τ ∈ Cp∞ . Then

τ (D)σ(D)u = (στ )(D)u = σ(D)τ (D)u (u ∈ S 0 ).

Consequently, Fourier multipliers commute with partial differential operators with con-
stant coefficients, with translations and with convolutions. Moreover,

σ(D)(lλ u) = lλ [(lλ σ)(D)u] (u ∈ S 0 , λ > 0), (19.2)

σ(D)ϕ = F −1 (σ) ∗ ϕ (ϕ ∈ S),

b ∈ E 0 or if σ ∈ S, then
and if σ

σ(D)u = F −1 (σ u
b) = F −1 (σ) ∗ u (u ∈ S 0 ). (19.3)

Proof. Most identities are trivial. The proof of (19.2) is left as an exercise, see Exer-
cise 19.A.

Exercise 19.A. Prove (19.2).

19.5 (Bessel potentials). Let g ∈ S 0 . We consider the problem of finding a u ∈ S 0 such

that

(1 − ∆)u = g.

We can write (1 − ∆) as a Fourier multiplier (by, for example (19.1)), namely (1 − ∆) =

σ(D), for

σ(ξ) = (1 + 4π 2 |ξ|2 ) (ξ ∈ Rd ).

As this function is strictly positive, we can divide by it: We define τ : Rd → R by

τ (ξ) = (1 + 4π 2 |ξ|2 )−1 (ξ ∈ Rd ).

120
It is not too difficult to show that τ ∈ Cp∞ . As τ σ = 1, with I : S 0 → S 0 being the
identity map ϕ 7→ ϕ, we have

I = (τ σ)(D) = τ (D)σ(D) = τ (D)(1 − ∆).

As also (1 − ∆)τ (D) = I, we have τ (D) = (1 − ∆)−1 and thus u = τ (D)g.

Let σ s for s ∈ R be the function given by σ s (ξ) = (σ(ξ))s for ξ ∈ Rd . Then

(1 − ∆)k = σ k (D) (k ∈ N0 ).

For s ∈ R \ N0 one writes “(1 − ∆)s ” instead of “σ s (D)”, i.e.,

(1 − ∆)s u := σ s (D)u (u ∈ S 0 ). (19.4)

For s < −d the function σ s is integrable and thus F −1 (σ s ) is a bounded continuous func-
tion. It turns out that for any s > 0 the tempered distribution F −1 (σ −s ) is represented
by a function that is smooth on Rd \ {0} (see for example [Gra14, Theorem 6.1.5]). Such
functions (on Rd \ {0}) are called Bessel potentials. For more on Bessel potentials we
refer to [Eva98, Section 4.3] and [Gra14, Section 6.1.2] . In the last reference, not the
function F −1 (σ −s ) but the operator (1 − ∆)−s is called a Bessel potential. We come back
to Bessel potentials in 20.9 and 20.10.

Exercise 19.B. Show that (1 − ∆)s forms a homeomorphism S → S and S 0 → S 0 for

each s ∈ R.

Similarly to the definition of (1−∆)s we will define ∆s for s ∈ R\N0 . As the function
ξ 7→ |ξ|2s is not smooth at 0, in order to define ∆s via Fourier multipliers, we extend
the notion of a Fourier multiplier σ(D) to the situation where the domain of σ is not
necessarily all of Rd .

19.6. Let Ω ⊂ Rd be open. Let σ ∈ C ∞ (Ω). Suppose u ∈ S 0 and [supp ub]3δ ⊂ Ω for some
∞
δ > 0. Let χ ∈ Cb be equal to 1 on [supp u]δ and 0 outside [supp u]2δ (see Lemma 9.3).
b b
By abusing notation by writing χσ for the function
(
χ(x)σ(x) x ∈ Ω,
x 7→
0 x∈/ Ω,

we have χσ ∈ Cp∞ if supp u b is compact or if σ ∈ Cp∞ (Ω). Therefore, in both cases,

(χσ)(D)u is defined and would be a good candidate for a definition of “σ(D)u” as, σ and
χσ are equal on [supp u
b ]δ .
We show that we can take this as our definition by showing that it does not depend
on the choice of χ. Indeed, if η ∈ Cb∞ equals 1 on [supp u
b]δ and 0 outside [supp u
b]2δ , then
χu
b = ηub and thus χσ ub = ησ ub, i.e., (χσ)(D)u = (ησ)(D)u.

121
Definition 19.7. Let Ω ⊂ Rd be open, σ ∈ C ∞ (Ω). Suppose u ∈ S 0 , δ > 0 and
b]3δ ⊂ Ω. If σ ∈ Cp∞ (Ω) or supp u
[supp u b is compact, we define

σ(D)u := F −1 (σχu
b) = (σχ)(D)u,

where χ ∈ Cb∞ (Rd , [0, 1]) equals 1 on [supp u

b]δ with supp χ ⊂ [supp u
b]2δ . In this sense we
obtain an operator σ(D) on the set

{u ∈ S 0 : there exists a δ > 0 such that [supp u

b]δ ⊂ Ω}. (19.5)

Observe that this extends Definition 16.22, as for Ω = Rd , one may choose χ = 1
(remember also 19.6).

Definition 19.8. We define S◦0 to be the space

n o
S◦0 = u ∈ S 0 : supp u
b ⊂ Rd \ {0} .

Observe that S◦0 is a subset of (19.5) for Ω = Rd \ {0}.

Exercise 19.C. Prove

n o
ϕ ∈ S : ∂ α ϕ(0)
b = 0 for all α ∈ Nd0
= {ϕ ∈ S : hP, ϕi = 0 for all polynomials P } .

19.9 (Fractional Laplacian/Riesz potential). For s ∈ R the function σ s : Rd \ {0} → R

given by

σ s (ξ) = |2πξ|2s , (ξ ∈ Rd \ {0})

is in Cp∞ (Rd \{0}). Let u ∈ S◦0 . Then there exists an ε > 0 such that B(0, ε) ⊂ Rd \supp u
b.
Therefore there exists a δ > 0 such that [supp u d s
b]3δ ⊂ R \{0} and hence σ (D)u is defined
for all s ∈ R. For k ∈ N0 σ k extends to a smooth function on Rd and

(−∆)k = σ k (D) on S 0 .

For s ∈ R \ N0 one writes “(−∆)s ” instead of “σ s (D)”, i.e.,

(−∆)s u = σ s (D)u (u ∈ S◦0 ).

The operator (−∆)s is called a fractional Laplacian but is sometimes also called a Riesz
α
potential. In probability theory for α ∈ (0, 2) the operator (−∆) 2 plays the role of the
generator of an α-stable Lévy process.
1
Exercise 19.D. Let d = 1. Show that (−∆) 2 6= ∂.

122
19.10 (Pseudo differential operators). Each linear partial differential operator with
constant coefficients P can be written as p(D) for some polynomial p: Suppose P =
α cα α
α∈Nd :|α|≤k cα ∂ for some scalars cα ∈ F. Then P = p(D) for p =
P P
0 α∈Nd :|α|≤k (2πi)α X . 0

Let P : S → S 0 be a linear partial differential operator with variable coefficients. Say

fα ∈ L∞ for α ∈ Nd0 with |α| ≤ k are such that
X
P = fα ∂ α .
α∈Nd0 ,|α|≤k

Then

P ϕ(x) = p(x, D)ϕ(x) (x ∈ Rd , ϕ ∈ S), (19.6)

where
X
p(x, ξ) = fα (x)ξ α (x, ξ ∈ Rd ).
α∈Nd0 ,|α|≤k

Pseudo differential operators are operators of the form (19.6) for certain “nice” functions
p. Fourier multipliers are examples of pseudo differential operators. For example (−∆)s
is a pseudo differential operator.

20 Fractional Sobolev spaces

In this section we return to Sobolev spaces and describe them in terms of Fourier trans-
forms and Fourier multipliers.

20.1 (Sobolev spaces described by Fourier transforms). Let k ∈ N0 . In Theorem 12.14

we have seen that H k (reminder: H k = H k (Rd )), being the Sobolev space given by

H k = W k,2 = {u ∈ D0 : ∂ β u ∈ L2 (Ω) for all β ∈ Nd0 with |β| ≤ k},

is a Hilbert space with inner product

X
hu, viH k = h∂ α u, ∂ α viL2 (u, v ∈ H k ),
α∈Nd0 :|α|≤k

so that it is equipped with the norm

 1
2
X
kukH k = k∂ α uk2L2 (u ∈ H k ).
α∈Nd0 :|α|≤k

This space can also be described using the Fourier transformation, as we will see in
Lemma 20.3.

123
Lemma 20.2. Let θ ∈ [0, 1]. The function [0, ∞) → [0, ∞), x 7→ xθ is subadditive, that
is,

(a + b)θ ≤ aθ + bθ (a, b ∈ [0, ∞). (20.1)

Proof. It suffices to show (20.1) for a = 1 (otherwise one can divide by aθ ). This in turn
follows by
d
(1 + t)θ − tθ = θ((1 + t)θ−1 − tθ−1 ) ≤ 0 (t ∈ [0, ∞)).
dt

Lemma 20.3. Let k ∈ N0 .

H k = {u ∈ S 0 : (1 + |X|)k u
b ∈ L2 }, (20.2)

the norm k · kH k is equivalent to u 7→ k(1 + |X|)k u

bkL2 , which means there exists a C > 1
such that
1
kukH k ≤ k(1 + |X|)k u
bkL2 ≤ CkukH k (u ∈ H k ), (20.3)
C
and

H k × H k → F, (u, v) 7→ h(1 + |X|)k u

b, (1 + |X|)k vbiL2 (20.4)

is an inner product on H k that generates the same topology as the inner product h·, ·iH k .

Proof. That (20.4) defines an inner product can be easily checked. It suffices to show
(20.3). We show that for k = 1 there exists a C > 0 such that

k(1 + |X|)k u
bkL2 ≤ CkukH k (u ∈ H k ),

and leave the rest for the reader, see Exercise 20.A.
1
By Lemma 20.2 we have (1 + |x|) 2 ≤ 1 + |x1 | + · · · + |xd | for all x ∈ Rd . Therefore
d
1 X
bkL2 ≤ ku
k(1 + |X|) 2 u bkL2 + kxj u
bkL2
j=1
d
X 1
≤ ku
bkL2 + kF(∂j u)kL2
j=1
2π
d
X 1
= kukL2 + k∂j ukL2 ≤ kukW 1,2 .
j=1
2π

The latter is equivalent to k · kH k , see 12.6.

124
 To prove (20.3) and thus (20.2) the Multinomial Theorem can be beneficial.

Theorem 20.4 (Multinomial Theorem). For x = (x1 , . . . , xd ) ∈ Fd and k ∈ N

!
k
X k α
(x1 + · · · + xd ) = x , (20.5)
α
α∈Nd0
|α|=k

where with α! = α1 !α2 ! · · · αd !,

!
k k! k!
= = .
α α! α1 !α2 ! · · · αd !

Proof. This follows by induction. For d = 1 the formula is trivial for all k ∈ N. For
d = 2 it is the usual binomial formula. Suppose (20.5) holds for a fixed d ∈ N and for
any k ∈ N. Then for y = x1 + · · · + xd and z = (x1 , . . . , xd ) we have
!
k
X k m k−m
(y + xd+1 ) = y xd+1
m∈N0
m
m≤k
! !
X k X m α k−m
= z xd+1 .
m∈N0
m α
α∈Nd0
m≤k |α|=k

As for β = (α1 , . . . , αd , k − m) we have |β| = k and

! ! !
k m k! m! k
= = ,
m α (k − m)!m! α1 ! · · · αd ! β

it follows that (20.5) is valid also for for d + 1.

The description of H k as in Lemma 20.3 extends naturally to non-integer values of

k, as follows.

Definition 20.5. For s ∈ R \ N0 we define the fractional Sobolev space H s by

H s = {u ∈ S 0 : (1 + | · |)s u
b ∈ L2 }, (20.6)

h·, ·iH s : H s (Ω) × H s (Ω) → F by

hu, viH s = h(1 + |X|)s u

b, (1 + |X|)s vbiL2 (u, v ∈ H s (Ω)),

and k · kH s : H s (Ω) → [0, ∞) by

q
kukH s = hu, uiH s (u ∈ H s (Ω)).

125
Example 20.6. We have already seen that δb0 = 1. As (1 + | · |)s is in L2 if and only if
2s < −d by Lemma 14.13, it follows that δ0 ∈ H s if and only if s < − d2 .

Exercise 20.A. Prove the following statements.

(a) For each k ∈ N0 there exists a C > 0 such that for all α ∈ Nd0 with |α| ≤ k,

k∂ α ukL2 ≤ Ck(1 + |X|)k u

bkL2 (u ∈ H k ).

Consequently, there exists a C > 0 such that

kukH k ≤ Ck(1 + |X|)k u

bkL2 (u ∈ H k ).

s
(b) For all s ∈ [0, ∞), the functions u 7→ k(1 + |X|)s u
bkL2 and u 7→ k(1 + |X|2 ) 2 u
bkL2 are
norms and they are equivalent.
s
(c) For s ∈ R the norm k · kH s is equivalent to u 7→ k(1 − ∆) 2 ukL2 .
(d) For each k ∈ N0 there exists a C > 0 such that

k(1 + |X|2 )k u
bkL2 ≤ CkukH 2k

(e) For each k ∈ N0 there exists a C > 0 such that (20.3) holds.

Theorem 20.7. Let s ∈ R. h·, ·iH s is an inner product on H s , so that H s equipped with
this inner product is a Hilbert space.

Exercise 20.B. Prove Theorem 20.7.

Like for the spaces of continuously differentiable functions and for Sobolev spaces,
see Lemma 12.8 and Lemma 12.9, we have the following.

Lemma 20.8. Let s, r ∈ R and α ∈ Nd0 . If r ≤ s, then

k∂ α ukH r−|α| ≤ kukH s (u ∈ H s ).

In particular, ∂ α : H s → H s−|α| is continuous for all s ∈ R and α ∈ Nd0 , and

H s ,→ H r (r ≤ s).

The Fourier multiplier Dα is thus a continuous function H s → H s−|α| . We consider

conditions on the functions σ such that the Fourier multiplier σ(D) is a continuous func-
tion between Besov spaces (in Section 25). Those spaces are generalisations of fractional
Sobolev spaces and are introduced in the following section.

Exercise 20.C. Prove Lemma 20.8.

126
20.9. By Exercise 20.A(b) and (c) we see that the spaces H s can be described in terms
of the Bessel potentials:

H s = {u ∈ S 0 : (1 − ∆)s u ∈ L2 }.

Actually, H s is a particular case of a Bessel potential space, see 20.10 for a definition.
20.10 (Bessel potential spaces). We have only considered a generalisation of the Sobolev
space W k,p for p = 2 but for any p one can actually define fractional Sobolev spaces, also
called Bessel potential spaces. In [Tri83, p.88] , for example, it is shown that
k
W k,p = {f ∈ S 0 : (1 − ∆) 2 f ∈ Lp }, (20.7)

and that k · kW k,p is equivalent to

k
f 7→ k(1 − ∆) 2 f kLp . (20.8)

Similarly to Definition 20.5 one defines the fractional Sobolev space Hps for s ∈ R and
p ∈ [1, ∞] by replacing “k” in (20.7) by “s”:
s
Hps = {u ∈ S 0 : (1 − ∆) 2 u ∈ Lp }, (20.9)

and defines a norm on Hps by

s
kukHps = k(1 − ∆) 2 ukLp (u ∈ S 0 ).

Then Hpk = W k,p and the norms k · kHpk and k · kW k,p are equivalent for k ∈ N0 and by
Plancherel’s identity it follows that k · kH2s = k · kH s and thus H2s = H s for s ∈ R. One
can also show that Hps is a Banach space for all s ∈ R and p ∈ [1, ∞].

21 Besov spaces defined by Littlewood–Paley decomposi-

tions

In this section we introduce Besov spaces. The definition is more cumbersome than those
of the normed spaces like W k,p , H k , H s and Hps . First, we take a specific partition of
unity of Rd , (χn )n∈N , such that in S 0

(u ∈ S 0 ),
X
u= χn (D)u
n∈N

and then describe Besov spaces in terms of the Fourier multipliers χn (D).
21.1 (Motivation of Besov spaces by the space H s ). We will see later that H s is a
special case of a Besov space. Let us motivate the definition of a general Besov space by
describing H s in a different way. For n ∈ N0 we define jn : Rd → F by

jn (x) = 1[n,n+1) (|x|) (x ∈ Rd ).

127
We have u ∈ H s if and only if (1 + |X|)s u
b ∈ L2 . As gjn is orthogonal to gjk in L2 when
k 6= n and g ∈ L , it follows that: u ∈ H s if and only if
2

(1 + |X|)s u
bIn ∈ L2 (n ∈ N0 ),
 
k(1 + |X|)s u
bjn kL2 ∈ `2
n∈N

and, moreover,
 
kukH s = k(1 + |X|)s u
bjn kL2 (u ∈ H s ),
n∈N `2

where we wrote “k · k`2 ” for “k · k`2 (N0 ) ”. From this it follows that u ∈ H s if and only if

(1 + n)s u
bjn ∈ L2 (n ∈ N0 ),
 
(1 + n)s ku
bjn kL2 ∈ `2
n∈N0

and, moreover, that k · kH s is equivalent to

 
u 7→ (1 + n)s ku
bjn kL2 (u ∈ H s ),
n∈N0 `2

which follows from the fact that for s ≥ 0 (and something similar for s < 0)

(1 + n)s jn ≤ (1 + |X|)s jn ≤ (2 + n)s jn ≤ 2s (1 + n)s jn (n ∈ N0 ).

Instead of taking intervals of length 1, we can also split the function into intervals of
dyadic length: For j ∈ N0 define hj : Rd → F by

hj (x) = 1[2j ,2j+1 ) (|x|) (x ∈ Rd ),

and let h−1 = j0 , i.e., h−1 = 1[0,1) (|X|). Then u ∈ H s if and only if

2js u
bhj ∈ L2 (j ∈ N0 ∪ {−1}),
 
2js ku
bhn kL2 ∈ `2 ,
j∈N0 ∪{−1}

and, k · kH s is equivalent to
 
u 7→ 2js ku
bhj kL2 (u ∈ H s ). (21.1)
j∈N0 ∪{−1} `2

As (16.3) of Theorem 16.10 extends to f, g ∈ L2 (see Exercise 21.A), by Theorem 16.27

it follows that

bhj kL2 = kF −1 (hj ) ∗ ukL2 .

ku (21.2)

Now Besov spaces are basically spaces with a norm like (21.1) where instead of “L2 ”
and “`2 ” one has “Lp ” and “`q ” for some p, q ∈ [1, ∞] and with smooth functions “ϕj ”

128
with compact support which are similar to “hj ” in the sense that like for hj , one has the
scaling relation ϕj (x) = ϕ0 (2−j x) for all x ∈ Rd . The latter is being done so that the
convolution with F −1 (ϕj ) is defined for any tempered distribution and defines a smooth
function.
We will first introduce such functions ϕj .
Exercise 21.A. Show that
Z Z
f gb = g fb (f, g ∈ L2 ).

21.2 (Notation N−1 ). We write “N−1 ” for the set {−1, 0, 1, 2, . . . }.

Definition 21.3 (Annulus). Let a, b ∈ (0, ∞), a < b. We write
A(a, b) = {x ∈ Rd : a < |x| < b}.
Such a set A(a, b) is called an annulus.
Definition 21.4. A function f : Rd → F is called radial if f (x) = f (y) for all x, y ∈ Rd
with |x| = |y|.
For each radial function f there exists a function g : [0, ∞) → F such that f (x) =
g(|x|) for x ∈ Rd .

We introduce the notion of a dyadic partition of unity, which consists of one function
that is supported in a ball and equals 1 on a smaller ball centered at the origin and of
functions that are supported in annuli which are scaled versions of each other.

Figure 2: An example of a dyadic partition of unity.

Definition 21.5. A sequence of radial functions (ϕj )j∈N−1 in Cc∞ (Rd , [0, 1]) is called a
dyadic partition of unity if
supp ϕ−1 is equal to the closure of a ball centered at the origin,
supp ϕ0 is equal to the closure of an annulus,
ϕj (ξ) = ϕ0 (2−j ξ) (ξ ∈ Rd , j ∈ N0 ),
X 1 X
ϕj (ξ) = 1, ≤ ϕj (ξ)2 ≤ 1 (ξ ∈ Rd ), (21.3)
j∈N−1
2 j∈N
−1

|i − j| ≥ 2 =⇒ supp ϕi ∩ supp ϕj = ∅ (i, j ∈ N−1 ). (21.4)

We say that a radial function ϕ in Cc∞ (Rd , [0, 1]) supported in an annulus generates a
partition of unity if there exists a partition of unity (ϕj )j∈N−1 with ϕ0 = ϕ
Observe

129
 (a) ϕj = l2−j ϕ = ϕ(2−j · ) for j ∈ N0 .
(b) ϕ−1 = 1 −
P
j∈N0 ϕj , and

J−1
l2−J ϕ−1 = ϕ−1 (2−J · ) =
X
ϕj (J ∈ N0 ). (21.5)
j=−1

(c) ϕ0 = (l2−1 − l1 )ϕ−1 = ϕ−1 ( 21 · ) − ϕ−1 .

Theorem 21.6. There exists a radial function ϕ in Cc∞ (Rd , [0, 1]) that generates a dyadic
partition of unity.

Proof. As the functions ϕj need to be radial, we first construct functions ηj on [0, ∞)

such that the functions ϕj defined by ϕj (ξ) = ηj (|ξ|) form a partition of unity.
Due to the above observations, we may as well start by showing the existence of
a function that is the function ϕ−1 of a partition of unity (ϕj )j∈N−1 , as then ϕ0 =
(l2−1 − l1 )ϕ−1 .
Let a ∈ ( 34 , 1). Let θ : [0, ∞) → [0, 1] be smooth on (0, ∞) and equal to 1 on [0, 34 ],
nonzero on [ 34 , a) and equal to 0 on [a, ∞). Define η = l 1 θ − θ = θ( 12 · ) − θ. Then η = 1
2
on [a, 64 ] and supp η ⊂ [0, 2a] \ [0, 34 ] ⊂ [ 34 , 2a]. Define η−1 = θ, ηj = l2−j η = η(2−j · ) for
j ∈ N0 . Then by definition
X
ηj = 1.
j∈N−1

3
As 2 < 3 = 4 · 4 we have [ 34 , 2] ∩ 4[ 34 , 2] = ∅ and thus, as supp η−1 ⊂ [0, 1] and supp ηj ⊂
2j [ 34 , 2],

|i − j| ≥ 2 =⇒ supp ηi ∩ supp ηj = ∅ (i, j ∈ N−1 ). (21.6)

We are left to show that for t ≥ 0,

1 X
≤ ηj (t)2 . (21.7)
2 j∈N
−1

P P
Let us write Σodd = j∈2N0 −1 ηj and Σeven = j∈2N0 ηj . As the functions ηj with
j ∈ 2N0 − 1 have disjoint supports by (21.6), we have Σ2odd = j∈2N0 −1 ηj2 . Similarly,
P

Σ2even = j∈2N0 ηj2 . Therefore,

P

X
1 = (Σodd (t) + Σeven (t))2 ≤ 2(Σ2odd (t) + Σ2even (t)) = 2 ηj (t)2 .
j∈N−1

By defining ϕ(ξ) = η0 (|ξ|) for ξ ∈ Rd , we see that ϕ is radial function in Cc∞ (Rd , [0, 1])
that generates a dyadic partition of unity.

130
Exercise 21.B. (a) Let a > 0. Show that there exists a radial function ϕ with
ϕ(ξ) = 1 for all ξ ∈ A(a − δ, a + δ) for some δ > 0, that generates a partition of
unity. (Hint: Observe that for ϕ in the proof of Theorem 21.6 one has ϕ(ξ) = 1 for
ξ ∈ Rd , |ξ| = 1.)
(b) Show that there exists a radial function ϕ with ϕ = 1 on A( 34 , 54 ) that generates a
partition of unity. (Hint: In the proof, replace “ 34 ” by another suitable number.)

21.7 (Notation of ordered and unordered infinite sums). In the following we use the
P
notation “ j∈N−1 ”, which denotes the unordered sum over a function on N−1 (with
values for example in S 0 ). We recall the notation from the section Conventions and
notation.
If I is a countable set and vi is an element of a topological vector space X for all i ∈ I,
then we say that
X
vi
i∈I
P∞
exists, if there exists a v ∈ X such that for each bijection q : N → I, n=1 vq(n) = v in
P
X, and write i∈I vi for v.

Lemma 21.8. Let (ϕj )j∈N−1 be a dyadic partition of unity. Then

X
ϕj ψ = ψ in S (ψ ∈ S).
j∈N−1

Moreover, for all finite subsets F ⊂ N−1

X
ϕj ≤ 4kϕ−1 kC k .
Ck
j∈F

Proof. We invoke Lemma 14.10 (d). Observe that kf + gkC k = kf kC k ∨ kgkC k for f and
g in C k with disjoint supports.
Therefore, by (21.4), it follows that for all finite subsets F ⊂ N−1

X X X
ϕj ≤ ϕj + ϕj
j∈F j∈2N0 −1∩F Ck j∈2N0 ∩F Ck
Ck
≤ sup kϕj kC k + sup kϕj kC k ≤ 2 sup kϕj kC k
j∈2N0 −1 j∈2N0 j∈N−1

≤ 2(kϕ−1 kC k ∨ kϕ0 kC k ) ≤ 4kϕ−1 kC k ,

where we used that kϕj kC k = kϕ0 (2−j · )kC k ≤ kϕ0 kC k and that ϕ0 = l 1 ϕ−1 − ϕ−1.
2
From this by Lemma 14.10 (d) it follows that ∞ n=1 ϕq(n) ψ = ψ in S for any bijection
P

q : N → N−1 .

131
 ϕj ψb = ψb in S for all ψ ∈ S. Therefore, by
P
By the above lemma we also have j∈N−1
the continuity of F,
 X 
b = F −1
F −1 (ϕj ψ)
X X
ϕj (D)ψ = ϕj ψb = ψ in S (ψ ∈ S). (21.8)
j∈N−1 j∈N−1 j∈N−1

The partition of unity (ϕj )j∈N−1 is used to obtain the operators ϕj (D). After showing
some properties of Besov spaces, one can basically only work with those operators. It is
therefore customary to use a shorter notation for these operators: ∆j = ϕj (D). Those
operators ∆j for j ∈ N−1 are also called the Littlewood–Paley operators.

Lemma 21.9. Let (ϕj )j∈N−1 be a dyadic partition of unity. Write ∆j = ϕj (D). Then

ψ ∈ S =⇒ ∆j ψ ∈ S, u ∈ S 0 =⇒ ∆j u ∈ Cp∞ , f ∈ Lp =⇒ ∆j f ∈ Lp (j ∈ N−1 ),

and

h∆j u, ψi = hu, ∆j ψi (j ∈ N−1 , u ∈ S 0 , ψ ∈ S) (21.9)

X
∆j ψ = ψ in S (ψ ∈ S), (21.10)
j∈N−1

in S 0 (u ∈ S 0 ),
X
∆j u = u (21.11)
j∈N−1

c0 kL1 kf kLp ≤ 2kϕ

k∆j f kLp ≤ kϕ d−1 kL1 kf kLp (j ∈ N0 , f ∈ Lp ), (21.12)
J
X
∆j f −1 kL1 kf kLp
≤ kϕd (J ∈ N−1 , f ∈ Lp ). (21.13)
j=−1 Lp

Proof. (21.10) follows from Lemma 21.8, see (21.8). The rest we leave as an exercise
(Exercise 21.C).

Exercise 21.C. Finish the proof of Lemma 21.9.

Definition 21.10. Let X be a vector space (over the scalar field F). We call a function
n : X → [0, ∞] a norm-like function if

n(x + y) ≤ n(x) + n(y), n(λx) = |λ|n(x), n(x) = 0 ⇐⇒ x = 0 (x, y ∈ X, λ ∈ F).

Like for norms, two norm-like functions n and m on X are called equivalent if there exists
a C > 1 such that
1
m(x) ≤ n(x) ≤ Cm(x) (x ∈ X).
C

Observe that Y = {x ∈ X : n(x) < ∞} is a vector space and n forms a norm on Y .

132
21.11 (Convention). Let M denote the space of measurable functions Rd → F. For Lp
spaces we have that a measurable function (or better said, an equivalence class) f ∈ M
is in Lp if and only if NLp (f ) < ∞, where NLp : M → [0, ∞] is the norm-like function
given by
Z 1
p
p
N (f ) =
Lp |f | (f ∈ M ).

Of course, on Lp , NLp is equal to the norm k · kLp . It may be convenient to work with
such norm-like functions, which are allowed to take the value ∞. In general, we will not
distinguish anymore between the norm-like function NLp and the norm k · kLp . Moreover,
for q ∈ [1, ∞] and a countable set I and a ∈ FI we will also write kak`q (I) for lq (a), where
lq : FI → [0, ∞] is the norm-like function given by
 1q
q
i∈I |a(i)| ) q < ∞,
(P
lq (a) = (a ∈ FI ).
i∈I |a(i)| q = ∞,
sup

Mostly, we write k · k`q instead of k · k`q (I) .

Definition 21.12 (Besov Space). Let s ∈ R and p, q ∈ [1, ∞]. Let ϕ be a function that
generates a dyadic partition of unity (ϕj )j∈N−1 . We define the function k · kBp,q 0
s [ϕ] : S →

[0, ∞] by
 
kukBp,q
s [ϕ] := 2js kϕj (D)ukLp (u ∈ S 0 ). (21.14)
j∈N−1 `q

Here we wrote “k · k`q ” as an abbreviation for “k · k`q (N−1 ) ”. Then k · kBp,q

s [ϕ] is a norm-like

function (observe that if ∆j u = 0 for all j ∈ N−1 , then u = 0 by Lemma 21.9). We

define the nonhomogeneous Besov space generated by ϕ, Bp,q s [ϕ], to be the space of all

tempered distributions u such that kukBp,qs [ϕ] < ∞. We write ∆j = ϕj (D). ∆j is called a

Littlewood–Paley operator and for u ∈ S 0 one also calls ∆j u the Littlewood–Paley block.
With this notation,
 
kukBp,q
s [ϕ] = 2js k∆j ukLp (u ∈ S 0 ).
j∈N−1 `q

We will drop the notation “[ϕ]” later, as the space does not depend on ϕ; this follows
from Theorem 21.18, see Corollary 21.20. First we consider some properties of tempered
distributions u ∈ S 0 such that supp u
b is a subset of an annulus, or of a ball, which will
be used to prove Theorem 21.18.
t [ϕ] ⊂ B s [ϕ] for s < t and
Exercise 21.D. Let p, q ∈ [1, ∞], s, t ∈ R. Show that Bp,q p,q
s [ϕ] ⊂ B s−ε [ϕ] for ε > 0.
Bp,∞ p,q

Exercise 21.E. Let p ∈ [1, ∞]. Show that for f ∈ Lp and λ > 0, with the convention
d
that ∞ = 0,
− dp
klλ f kLp = λ kf kLp , (21.15)
where lλ f (x) = f (λx) (as in 14.9).

133
Definition 21.13. For λ ∈ R \ {0} we define lλ∗ to be the operation λ−d l 1 .
λ

Observe the following facts.

(a) lλ∗ is the adjoint of lλ as an operator on L2 , i.e.,

hlλ f, giL2 = hf, lλ∗ giL2 (f, g ∈ L2 ),

and (Definition 2.14 (e) and Definition 15.3)

hlλ u, ψi = hu, lλ∗ ψi, hlλ∗ u, ψi = hu, lλ ψi (u ∈ S 0 , ψ ∈ S).

In particular, if d = 1, δλ = lλ∗ δ1 .
(b) By Theorem 16.24 we know that for a distribution u ∈ S 0 ,

F(lλ u) = lλ∗ u
b F(lλ∗ u) = lλ u
b. (21.16)

(c) Furthermore, by (21.15)

klλ∗ f kL1 = kf kL1 (λ > 0, f ∈ L1 ). (21.17)

The following lemma will be used both for the proof of Theorem 21.18 but also gives
us Bernstein inequalities, see Theorem 21.15.
Lemma 21.14. Let A be an annulus and B be a ball centered at the origin in Rd . Let
χ ∈ Cc∞ be equal to 1 on a neighbourhood of B. Let φ ∈ Cc∞ be supported in an annulus
and be equal to 1 on a neighbourhood of A. Let λ > 0.

b ⊂ λB, then for all α ∈ Nd0

(a) If u is a tempered distribution with supp u

∂ α u = λ|α|+d (lλ hα ) ∗ u,

where hα ∈ S is given by hα = ∂ α F −1 (χ).

b ⊂ λA and k ∈ N0 , then
(b) If u is a tempered distribution with supp u

u = λ−k
X
λd (lλ gα ) ∗ ∂ α u,
α∈Nd0 :|α|=k

where gα ∈ S for α ∈ Nd0 with |α| = k is given by

!
k  
gα = F −1 (−2πiX)α |2π X|−2k φ .
α

Proof. First observe that l 1 χ equals 1 on a neighbourhood of λB and l 1 φ equals 1 on a

λ λ
neighbourhood of λA. As

∂ α F −1 (l 1 χ) = ∂ α l∗1 F −1 (χ) = λ|α| l∗1 ∂ α F −1 (χ),

λ λ λ
   
−1 −2k |α|−2k ∗ −1
F α
(−2πiX) |2π X| l1 φ = λ l1 F (−2πiX)α |2π X|−2k φ ,
λ λ

134
it is sufficient to prove the statements for λ = 1.
(a) follows from the fact that u
b = χu
b (see Definition 5.1).
b is supported on an annulus, we can divide (and multiply) by |2π X|2k .
For (b), as u
By the multinormial theorem (see Theorem 20.4, take xj = |2πξj |2 = (−2πiξj )(2πiξj )):
!
X k
|2πξ|2k = (−2πiξ)α (2πiξ)α (ξ ∈ Rd ). (21.18)
α
α∈Nd0 :|α|=k

Therefore by Lemma 19.4 and because (2πiX)α (D) = ∂ α ),

P 
k
α α
α∈Nd0 :|α|=k α (−2πiX) (2πiX)
u = F −1 (φu
b) =  φ (D)u
|2π X|2k
!
k (−2πiX)α
X 
= φ (D) ∂ α u,
α |2π X|2k
α∈Nd0 :|α|=k

which proves (b).

The next theorem follows by the previous lemma and Young’s inequality.
Theorem 21.15 (Bernstein’s inequalities). Let A be an annulus and B be a ball around
the origin in Rd . For all k ∈ N there exists a C > 0 such that such that for all p, q ∈ [1, ∞]
with q ≥ p and any u ∈ S 0 we have for all λ > 0
k+d( p1 − 1q )
b ⊂ λB =⇒ max k∂ α ukLq ≤ Cλ
supp u kukLp , (21.19)
α∈Nd0
|α|=k
1 k
b ⊂ λA =⇒
supp u λ kukLp ≤ max k∂ α ukLp ≤ Cλk kukLp . (21.20)
C α∈Nd0
|α|=k

Proof. The lower bound in (21.20) follows by Lemma 21.14 and Young’s inequality,

kukLp ≤ λ−k
X
kλd (lλ gα )kL1 k∂ α ukLp
α∈Nd0 :|α|=k
  
−k
X
α
≤λ kgα kL1 max k∂ uk Lp ,
α∈Nd0 :|α|=k
α∈Nd0 :|α|=k

because kλd (lλ gα )kL1 = kl∗1 gα kL1 = kgα kL1 and gα ∈ S ⊂ L1 for all α ∈ Nd0 .
λ

The upper bound in (21.20) follows from (21.19), so we are left to prove the latter.
Let r ∈ [1, ∞] be such that 1q + 1 = 1r + p1 . By Young’s inequality, Lemma 21.14 and an
easy calculation (see Exercise 21.E)
d
k∂ α ukLq ≤ λk+d klλ hα kLr kukLp ≤ λk+d− r khα kLr kukLp

135
so that (21.19) follows because d − dr = d(1 + 1
p − 1
q − 1) = d( p1 − 1q ) and because
khα kLr ≤ khα kL1 + khα kL∞ , see Corollary A.12.

Exercise 21.F. Let s ∈ R. Show that there exists a C > 0 such that for all p, q ∈ [1, ∞],
q ≥ p, any u ∈ S 0 and λ > 0
s s+d( p1 − 1q )
b ⊂ λA =⇒ k(−∆) 2 ukLq ≤ Cλ
supp u kukLp (u ∈ S 0 , λ > 0).
(Hint: Adapt Lemma 21.14 (a) and follow the proof of Theorem 21.15.)
21.16 (Towards Theorem 21.18). After one more preparation, we turn to a technical
theorem, Theorem 21.18, so we want to prepare the reader a little bit. This theorem
implies that a Besov space does not depend on the dyadic partition of unity, but it tells
us more than that. Indeed, in order to calculate kukBp,q
s [ϕ] one needs to first calculate
p
the L norm of all Littlewood–Paley blocks ∆j u. Theorem 21.18 considers sequences of
tempered distributions (uj )j∈N−1 with
b−1 ⊂ B,
supp u bj ⊂ 2j A for j ∈ N0 ,
supp u (21.21)
 
2js kuj kLp < ∞. (21.22)
j∈N−1 `q

First, observe that if u ∈ S 0 , then (21.21) is satisfied for uj = ∆j u. And if, moreover,
s , then also (21.22) is satisfied with u = ∆ u and k(2js ku k p )
u ∈ Bp,q j j j L j∈N−1 k`q = kukBp,q
s .

Also observe that uj ∈ Cp∞ for all j ∈ N−1 due to (21.21), see Lemma 17.8.
Theorem 21.18 tells us on the one hand that if (21.21) and (21.22) are satisfied, then
0
j∈N−1 uj exists in S , and, on the other hand, that for u =
P P
j∈N−1 uj the Besov norm
kukBp,q js
s [ϕ] is bounded from above by k(2 kuj kLp )j∈N−1 k`q .

This theorem is useful as it can be used to estimate kukBp,q

s [ϕ] in case calculating

∆j u might be more effort. A very simple example is the following. Suppose u ∈ Bp,q s

and A ( N−1 . Then Theorem 21.18 tells us that v := j∈N−1 ∩A ∆j exists in S , and
P 0

moreover, it gives an estimate on the Besov norm k∆j vkBp,q

s [ϕ] without calculating ∆j v

for all j ∈ N−1 .

The Young’s inequality will be helpful multiple times, not only for convolutions of
functions on Rd (as in Theorem 7.7), but also on Z. Both Rd and Z are groups, and one
can formulate the Young’s inequality more generally for locally compact groups that are
equipped with a so-called Haar measure. More on such spaces can be found in books on
“abstract harmonic analysis”. We formulate the Young’s inequality for functions on Z
separately:
Theorem 21.17 (Young’s inequality for functions on Z).
Let p, q, r ∈ [1, ∞] be such that
1 1
p + q ≥ 1 + 1r .
For f ∈ `p (Z), g ∈ `q (Z) we have f ∗ g ∈ `r (Z) and
kf ∗ gk`r ≤ kf k`p kgk`q .

136
Proof. Let s ∈ [1, ∞], s ≥ r, and p1 + 1q = 1 + 1s . Then k · k`r ≤ k · k`s , see Lemma A.8.
Therefore, we may as well assume p1 + 1q = 1 + 1r .
Now the proof follows in a similar fashion as the proof of Theorem 7.7, but with
applying Hölder’s inequality (Theorem A.4) to the sequence spaces `p (Z) (observe `p (Z)
equals Lp (µ) where µ is the counting measure on Z).

Now we are ready to prove the rather technical theorem with a long statement. Let
us mention beforehand that (b) and (c) are similar to (a).

Theorem 21.18. Let s ∈ R and p, q ∈ [1, ∞]. Let B be a ball centered at the origin and
A be an annulus. Let ϕ be a function that generates a partition of unity (ϕj )j∈N−1 .

(a) There exist C > 0 and m ∈ N0 such that for all sequences of tempered distributions
(uj )j∈N−1 with
 
b−1 ⊂ B,
supp u bj ⊂ 2j A for j ∈ N0 ,
supp u 2js kuj kLp < ∞,
j∈N−1 `q

one has that


uj exists in S 0 ,
P

 u := j∈N−1
(21.23)
2js kuj kLp


|hu, ψi| ≤ C kψkm,S (ψ ∈ S),


j∈N−1 ` q

and
 
kukBp,q
s [ϕ] ≤ C 2js kuj kLp . (21.24)
j∈N−1 `q

−1 kL1 for (21.24).)

(Here one may take C = 2kϕd
(b) If s > 0, then there exist C > 0 and m ∈ N0 such that for all sequences of tempered
distributions (uj )j∈N−1 with
 
bj ⊂ 2j B for all j ∈ N−1 ,
supp u 2js kuj kLp < ∞, (21.25)
j∈N−1 `q

(21.23) and (21.24) hold. (Here there exists an N ∈ N such that we may take
2N s
C = 2kϕd−1 kL1 2s −1 for (21.24).)

(c) If s = 0 and q = 1, then there exist C > 0 and m ∈ N0 such that for all sequences
of tempered distributions (uj )j∈N−1 with (21.25), (21.23) holds and

kukBp,∞
0 [ϕ] = sup kϕj (D)ukLp ≤ (kuj kLp )j∈N−1 . (21.26)
j∈N−1 `1

Proof. First observe that if (uj )j∈N−1 is a sequence of tempered distributions, then
0 0
j∈N−1 uj exists in S if there exists a u ∈ S such that for all bijections q : N → N−1
P

137
the sum Jj=1 uq(j) converges to u in S 0 as J → ∞. As S 0 is weak∗ sequentially complete
P

(Theorem 15.10 (b)), it is sufficient to show ∞ j=−1 |huj , ψi| < ∞ for all ψ ∈ S. In order
P

to also show that (21.23) holds, it is sufficient to prove

∞
X  
|huj , ψi| ≤ C 2js kuj kLp kψkm,S (ψ ∈ S). (21.27)
j∈N−1 `q
j=−1

Without loss of generality, as we may take a larger ball and annulus, we may assume
supp ϕ−1 ⊂ B and supp ϕ0 ⊂ A. We write ∆j = ϕj (D) for j ∈ N−1 .
In the rest of the proof we let r ∈ [1, ∞] be such that
1 1
+ = 1.
p r

Let n ∈ N and a C1 > 0 be such that (see Lemma 14.14)

kψkLr ≤ C1 kψkn,S (ψ ∈ S). (21.28)

−1 kL1 see (21.12) and (21.13) with

Let C2 > 0 be such that (for example C2 = 2kϕd
J = −1)

k∆j f kLp ≤ C2 kf kLp (f ∈ Lp ). (21.29)

Finally, let N ∈ N be such that

2j A ∩ A = 2j A ∩ B = ∅ (j ≥ N ). (21.30)

We start with (b) and (c) as these are the easier cases.
Proof of (c) Suppose (uj )j∈N−1 is as in (c). By Hölder’s inequality
∞
X ∞
X
|huj , ψi| ≤ kuj kLp kψkLr ≤ C1 (kuj kLp )j∈N−1 kψkn,S ,
`1
j=−1 j=−1

so that we obtain (21.27) with q = 1. We obtain (21.26) by using

X
k∆j ukLp ≤ k∆j ui kLp .
i∈N−1

Proof of (b) Suppose that s > 0 and (uj )j∈N−1 is as in (b). By Hölder’s inequality we
obtain
 
|huj , ψi| ≤ kuj kLp kψkLr ≤ C1 2−js 2js kuj kLp kψkn,S .
j∈N−1 `q

2−js .
P
As s > 0, we obtain (21.27) with C = C1 j∈N−1

138
 By (21.30) we have ∆j ui = 0 for all j ≥ i + N and so by (21.29)
X
2js k∆j ukLp ≤ 2(j−i)s 2is k∆j ui kLp
i∈N−1 :i>j−N
X
≤ C2 2(j−i)s 2is kui kLp = C2 (a ∗ b)(j),
i∈N−1 :i>j−N

where a(k) = 2ks 1(−∞,N ) (k) for k ∈ Z and b(k) = 2ks kuk kLp for k ∈ N−1 and b(k) = 0
for k ∈ Z \ N−1 . By Young’s inequality (Theorem 21.17), we obtain the desired bound as
2(N −1)s 2N s
2(N −1−k)s =
X X
kak`1 = 2ks = = .
k∈Z:k<N k∈N0
1 − 2−s 2s − 1

Proof of (a) Suppose (uj )j∈N−1 is as in (a).

• Let us first prove (21.27). Let k ∈ N0 be such that k > −s. Let ψ ∈ S. For all
j ∈ N0 we have by Lemma 21.14 (b)
uj = 2−jk
X
2jd (l2j gα ) ∗ ∂ α uj .
α∈Nd0 :|α|=k

Therefore,
huj , ψi = 2−jk (−1)k
X
huj , 2jd (l2j ǧα ) ∗ ∂ α ψi.
α∈Nd0 :|α|=k

By Hölder’s and by Young’s inequality, and then by (21.28) and (21.17),

|huj , ψi| ≤ 2−jk kuj kLp
X
k2jd (l2j ǧα )kL1 k∂ α ψkLr
α∈Nd0 :|α|=k

≤ C1 2−jk kuj kLp

X
kgα kL1 kψkn+k,S .
α∈Nd0 :|α|=k

Therefore with
X
C3 = kgα kL1 ,
α∈Nd0 :|α|=k

we have for all j ∈ N0

|huj , ψi| ≤ C1 C3 2−jk kuj kLp kψkn+k,S
 
≤ C1 C3 2−j(k+s) 2js kuj kLp kψkn+k,S .
j∈N−1 `q

We may assume that the above also holds for j = −1, as by a direct application of
Hölder’s inequality we have
|hu−1 , ψi| ≤ ku−1 kLp kψkLr ≤ C1 ku−1 kLp kψkn,S
 
≤ C1 2 s 2js kuj kLp kψkn+k,S .
j∈N−1 `q

139
As k + s > 0, there exists a C > 0 such that (21.27) holds with m = n + k.
• We prove (21.24). By (21.30) ∆j ui = 0 for all i, j ∈ N−1 with |i − j| ≥ N . As
2(j−i)s ≤ 2|s|N for i, j ∈ N−1 with |i − j| ≤ N , with (21.29), for all j ∈ N−1
j+N
X j+N
X
2js k∆j ukLp ≤ 2js k∆j ui kLp ≤ C2 2is kui kLp = C2 (a ∗ b)j ,
i=(j−N )∨−1 i=(j−N )∨−1

where a(k) = 1[−N,N ) (k) for k ∈ Z and b(k) = 2ks kuk kLp for k ∈ N−1 and b(k) = 0 for
k ∈ Z \ N−1 . Therefore by Young’s inequality (Theorem 21.17),

(2js k∆j ukLp )j∈N−1 ≤ C2 ka ∗ bk`q (Z) ≤ C2 kak`1 (Z) kbk`q (Z) .
`q (N−1 )

2js kuj kLp

As kbk`q (Z) = j∈N−1 `q (N ) and kak`1 (Z) = 2N + 1, this finished the proof
−1
for (a).

Remark 21.19. So again, Theorem 21.18 gives us an upper bound of the Besov norm
of a tempered distribution u if u can be separated into smooth pieces uj for j ∈ N−1
with certain properties which are similar to those that the ∆j u satisfy, although less
restrictive. Indeed, for u = 0 one may take u−1 and u0 to be nonzero, in Lp , such that
u0 = −u−1 and supp ud −1 = supp uc0 ⊂ B ∩ A. In this case we see that the inequality
(21.24) cannot be reversed as the left-hand side of (21.24) is zero and the right-hand side
is not.
Corollary 21.20. Let s ∈ R and p, q ∈ [1, ∞]. Let ϕ and ψ generate dyadic partitions
s [ϕ] = B s [ψ] and there exists a C > 1 such that
of unity. Then Bp,q p,q

1
kukBp,q
s [ψ] ≤ kukB s [ϕ] ≤ CkukB s [ψ] (u ∈ S 0 ). (21.31)
C p,q p,q

In particular, k · kBp,q s
s [ϕ] and k · kB s [ψ] are equivalent norms on Bp,q [ϕ].
p,q

Proof. Let (ϕj )j∈N−1 and (ψj )j∈N−1 be the dyadic partitions of unity generated by ϕ and
ψ, respectively. There exist a ball B and an annulus A such that supp ϕ−1 , supp ψ−1 ⊂ B
and supp ϕ0 , supp ψ0 ⊂ A. By Theorem 21.18 (a) there exists a C > 0 such that
kukBp,q
s [ϕ] ≤ CkukB s [ψ]
p,q
(u ∈ S 0 ).

By interchanging the roles of ϕ and ψ we obtain (21.31).

Definition 21.21. Let s ∈ R and p, q ∈ [1, ∞]. We define the nonhomogeneous Besov
s to be equal to B s [ϕ], where ϕ is a function that generates a dyadic partition
space Bp,q p,q
of unity. Often we write k · kBp,q
s for k · kBp,q
s [ϕ] and without mentioning the partition of

unity, write ∆j for the Littlewood–Paley operators corresponding to a partition of unity.

Basically all statements that follow about the function k · kBp,q
s are about estimations
with respect to a scalar times another function. For such statements, the choice of
partition is of course irrelevant.

140
Exercise 21.G. Let p ∈ [1, ∞].

(a) Let u ∈ Lp and suppose that supp u

b is compact. Show that for all s ∈ R and
s .
q ∈ [1, ∞], u ∈ Bp,q
s for some q ∈ [1, ∞] and s ∈ R. Suppose that supp u
(b) Let u ∈ Bp,q b is compact. Show
that u ∈ Lp .
Example 21.22. Remember that in Example 20.6 we found that δ0 ∈ H s if and only if
s < − d2 .
Now let us consider in which Besov space the Dirac delta, δ0 , lies. Using that δb0 = 1,
one has ∆j δ0 = F −1 (ϕj ) for all j ∈ N−1 . Let p, q ∈ [1, ∞]. Then, by (21.16) and
d
Exercise 21.E, with the convention that ∞ = 0,
−j( pd −d)
k∆j δ0 kLp = kF −1 (ϕj )kLp = 2 kF −1 (ϕ0 )kLp (j ∈ N0 ),
k∆−1 δ0 kLp = kϕd
−1 kLp .

Therefore,
s
δ0 ∈ Bp,q ⇐⇒ s < −d(1 − p1 ) (q < ∞).
−d(1− p1 ) −d(1− p1 )−ε
So that in particular, δ0 ∈ Bp,∞ and δ0 ∈ Bp,q for all ε > 0.
s if and only if s < − d . As the
Observe that for any q ∈ [1, ∞] we have δ0 u ∈ B2,q 2
s = H s . The proof
reader may already expect from the motivation in 21.1, one has B2,2
of this will be given in Theorem 23.2.
−ε
Exercise 21.H. Show that for ε > 0 the function z 7→ δz is continuous on B1,∞ but
0 1 d
that it is not continuous on B1,∞ . Hint: First prove: for any f ∈ L and z ∈ R \ {0},

lim kTaz f − f kL1 = 2kf kL1 .

a→∞

This statement can be proved by approximating f by f 1K for a ‘large’ compact set K.

Similarly to Lemma 12.8, Lemma 12.9 and Lemma 20.8, we have the following.
Theorem 21.23. For all α ∈ Nd0 there exists a C > 0 such that for all s, t ∈ R,
p1 , p2 , q1 , q2 ∈ [1, ∞], with
p2 ≥ p1 , q 2 ≥ q1 , t ≤ s − d( p11 − 1
p2 ), (21.32)
one has
k∂ α ukB t−|α| ≤ CkukBps (u ∈ S 0 ). (21.33)
p2 ,q2 1 ,q1

s−|α|
In particular, ∂ α : Bp,q
s → B
p,q is continuous for all s ∈ R, p, q ∈ [1, ∞] and α ∈ Nd0 ,
and for all s, t ∈ R, p1 , p2 , q1 , q2 ∈ [1, ∞], p2 ≥ p1 , q2 ≥ q1 ,
Bps1 ,q1 ,→ Bpt 2 ,q2 t ≤ s − d( p11 − 1


p2 ) .

141
Proof. This follows by Bernstein’s inequality, Theorem 21.15: By taking λ = 2j and
using that ∆j ∂ α = ∂ α ∆j it implies that there exists a C > 0 such that
j(|α|+d( p1 − p1 ))
k∆j ∂ α ukLp2 ≤ C2 1 2 k∆j ukLp1 (u ∈ S 0 ).
Therefore
k∂ α ukBpt ≤ Ckuk t+|α|+d( p1 − p1 ) (u ∈ S 0 ).
2 ,q2 1 2
Bp1 ,q2

By monotonicity of the norm k · k`q in q (see Lemma A.8) and by monotonicity of the
norm k · kBp,q
s in s (see Exercise 21.D) we obtain (21.33).

Exercise 21.I. Let t ∈ R, p, q ∈ [1, ∞]. Show that for u ∈ S◦0 (see Definition 19.8)
s
k(−∆) 2 ukBp,q
t−s ≤ kukB t .
p,q

(Hint: Exercise 21.F.)

Exercise 21.J. 0
(a) Let ψ ∈ S. Show that ψ ∈ Bp,∞ for all p ∈ [1, ∞].
s for all
(b) (Optional, will be proven also in the next section.) Show that ψ ∈ Bp,q
s ∈ R, p, q ∈ [1, ∞]. (Hint: First consider q = ∞ and use (a) and Bernstein’s
inequalities.)
Exercise 21.K. Show that the distribution |X|2 is a tempered distribution which is
s for any s ∈ R, p, q ∈ [1, ∞].
not in Bp,q
21.24 (Some formal language as explanation).
Let us formally describe what Theorem 21.23 means, but first give names to the para-
s ” by calling “s” the regularity
meters of Besov spaces. We refer to the parameters in “Bp,q
parameter and calling “p” and “q” the first and second integration parameter. In the
literature one also says that Besov spaces are examples of function spaces with mixed
regularity.
First, let us describe what it means for an element to be in a Besov space in this
formal language. A tempered distribution u is in Bp,q s for some s, p and q, if ∆ u is
j
p
in L for all j and k∆j ukLp cannot grow too rapidly. Take for example q = ∞. By
writing N for the Besov norm of u, N = kukBp,q js
s , we see that 2 k∆j ukLp < N and thus
−js
k∆j ukLp ≤ 2 N for all j. Now we see that if s is strictly positive, then k∆j ukLp needs
to decrease fast enough as j increases and if s is negative, then one only allows a not too
fast growth of k∆j ukLp .
Now Theorem 21.23 tells us on the one hand by taking a derivative, the regularity
parameter decreases by the number of (directional) derivatives taken. On the other
hand, it tells us that for an element in a Besov space we can increase the first integration
parameter at the cost of decreasing the regularity parameter. Also we can increase the
second integration parameter without changing the first and the regularity parameter.
The following lemma states that one can also decrease the second parameter by paying
the littlest amount of regularity.

142
Lemma 21.25. For all q1 , q2 ∈ [1, ∞] and ε > 0 there exists a C > 0 such that for all
s ∈ R and p ∈ [1, ∞]
kukBp,q
s−ε ≤ CkukB s
p,q
(u ∈ S 0 ). (21.34)
2 1

That is,
s s−ε
Bp,q 1
,→ Bp,q2
.

Proof. If q1 ≤ q2 , then this follows directly from Theorem 21.23 (even in case ε = 0).
Therefore we assume q1 > q2 . The case q1 = ∞ has already been treated in Exercise 21.D.
s−ε and a := k∆ uk p . Observe that
Let u ∈ Bp,q 2 j j L

q2 q1 − q2
+ = 1.
q1 q1
q1 q2
Therefore, by Hölder’s inequality, letting r = q1 −q2 ,
 X  q1
2
j(s−ε) −jεq2 js q2
s−ε = k(2
kukBp,q aj )j∈N−1 k`q2 = 2 (2 aj )
2
j∈N−1
 X q1 q2
 qq1 −q 2  q1
−jε q 1 q2 X 1
js q1
≤ 2 1 −q2 (2 aj )
j∈N−1 j∈N−1
−jε js
= k(2 )j∈N−1 kLr k(2 aj )j∈N−1 k`q1 .
So that with C = k(2−jε )j∈N−1 kLr we obtain (21.34).

Theorem 21.23 and Lemma 21.25 imply for all s ∈ R, q1 , q2 ∈ [1, ∞] with q2 ≥ q1 and
ε>0
s s s s−ε
B∞,q 1
⊂ B∞,q 2
, B∞,q 2
⊂ B∞,q 1
.
The following example illustrates (at least for d = 1) that those inclusions are strict, i.e.,
s s s s−ε
B∞,q 1
( B∞,q 2
, for q2 > q1 B∞,q 2
( B∞,q 1
.
We extend the idea behind the following example in the proof of Theorem 22.5 to show
that all inclusions that one obtains from Theorem 21.23 and Lemma 21.25 are strict.
Example 21.26 will be considered again in the next section.
Example 21.26. Let d = 1. Let (ϕj )j∈N−1 be a dyadic partition of unity such that
ϕ0 = 1 on a neighbourhood V of 1 (so that ϕi = 1 on ±2j if i = j and 0 otherwise). For
n ∈ N, define vn , un ∈ S 0 either by
vn = 12 (δ2n + δ−2n ) un = cos(2π2n X),
or by
vn = 1
2i (δ2
n − δ−2n ) un = sin(2π2n X).
Observe that for n ∈ N,

143
 • vcn = un and ucn = vn ,

• un ∈ L∞ and kun kL∞ = 1,

• ϕj vn = vn and ∆j un = un if j = n and 0 otherwise for all j ∈ N−1 .

1
Let q ∈ [1, ∞], s ∈ R. Define a, b : N → R by ( ∞ = 0)

− 1q
a(n) = n , b(n) = 2−ns a(n) (n ∈ N),

so that

a ∈ `r ⇐⇒ r ≥ q, k(2ns b(n))n∈N k`r = kak`r (r ∈ [1, ∞]).

Because k(2ns kb(n)un kL∞ )n∈N k`q = kak`q < ∞, by Theorem 21.18 (a),

b(n)un exists in S 0 ,
X
u :=
n∈N

Moreover, for r ∈ [1, ∞] and t ∈ R

kukB∞,r
t = k(2n(t−s) a(n))n∈N k`r

which is finite if and only if t = s and r ≥ q or t < s. Therefore

(
t t = s, r ≥ q,
u∈ B∞,r ⇐⇒
t < s.

The previous example can be used to explain the usage of the word “frequencies”, a
term that is used regularly in the literature:

21.27 (Formal language in terms of frequencies).

Let us consider the formal language in 21.24 with regard to Example 21.26. We can
say that the frequency of un is equal (or proportional to) 2n . Observe that ∆n u is a
multiple of un , which is either of the functions cos(2π2n X) and sin(2π2n X) for n ∈ N. It
is rather natural to call “2n ” the frequency of such a un . Formally, for general tempered
distributions u, one often says: “∆j u captures the behaviour of u at those frequencies of
order 2j ”. Moreover, for example in [Saw18], the support of ub is also called the “frequency
support of u”.
With this language, the decomposition of u in terms of its Littlewood–Paley blocks
∆j u, also called the Littlewood–Paley decomposition, is a decomposition of u into fre-
quency levels (or blocks of frequencies) of different (dyadic) order.
Now with this “frequency” language, one could say that u ∈ Bp,q s if the behaviour

of u at the frequencies of order 2j does not increase too fast (s < 0) or decreases fast
enough (s > 0).

144
 In Theorem 21.32 we will show that Besov spaces are Banach spaces. Moreover,
we prove another property, namely: Every bounded sequence in a Besov space has a
subsequence that converges in S 0 to an element of that Besov space whose Besov norm
is bounded by the lim inf of the norms of the subsequence. Probably due to this limiting
inequality, this property is in the literature sometimes called the “Fatou property”.
In order to prove this property for Besov spaces, we first prove it for the spaces Lp and
M. For that, let us recall two facts from Functional Analysis. For a proof of the following
theorem, see for example [Con90, Theorems III.5.5 and III.5.7]. (For the definitions of
M and C0 see Definition 2.23 and Definition 2.29.)

Theorem 21.28. (a) Let p ∈ (1, ∞] and q ∈ [1, ∞) be such that p1 + 1q = 1. Then Lp is
isometrically isomorphic to (Lq )0 , the dual of Lq . In particular, with hv, f i = vf
R

for v ∈ Lp , f ∈ Lq ,

kvkLp = sup{|hv, f i| : f ∈ Lq , kf kLq ≤ 1} (v ∈ Lp ).

(b) M
R
is isometrically isomorphic to C00 , the dual of C0 . In particular, with hµ, f i =
f dµ for µ ∈ M, f ∈ C0 ,

kµkM = sup{|hµ, f i| : f ∈ C0 , kf kC0 ≤ 1} (µ ∈ M).

The other fact we recall from Functional Analysis is the separable version of Alaoglu’s
theorem. For a proof combine [Con90, Theorem III.3.1 and III.5.1] or [Rud91, Theorem
3.15 and 3.16].

Theorem 21.29 (Alaoglu’s theorem, separable version).

Let X be a separable normed space. Then the closed unit ball in X0 , {f ∈ X0 : kf kX0 ≤ 1}
is weak*-sequentially compact.

Lemma 21.30. Let p ∈ (1, ∞]. Let X be either the Banach space Lp or M. If (un )n∈N is
a sequence in X that is bounded with respect to the norm on X, then it has a subsequence
(unm )m∈N that converges in S 0 to an element u of X with

kukX ≤ lim inf kunm kX . (21.35)

m→∞

Proof. Let Y be either Y = Lq or Y = C0 (see Theorem 21.28), so that X is isometrically

isomorphic to Y0 , the dual of Y. Then there is a pairing h·, ·i between X and Y with

kukX = sup{|hu, f i : f ∈ Y, kf kY ≤ 1} (u ∈ X).

Alaoglu’s theorem ensures the existence of a subsequence (unm )m∈N and an element
u ∈ Y0 such that

lim hunm , f i = hu, f i (f ∈ Y).

m→∞

145
For f ∈ Y we have

|hu, f i| = lim inf |hunm , f i| ≤ lim inf kunm kX kf kY ,

m→∞ m→∞

so that kukX ≤ lim inf m→∞ kunm kX .

The continuous embedding S ,→ Y entails that u may be viewed as an element of S 0
and that limm→∞ unm = u in S 0 .

Exercise 21.L. Show that the statement in Lemma 21.30 for p = 1 does not hold.

With an additional assumption on the supports of the Fourier transforms, we do

derive the same conclusion as in Lemma 21.30 even for p = 1.

Lemma 21.31. Let p ∈ [1, ∞]. Let K ⊂ Rd be compact. If (un )n∈N is a sequence in Lp
that is bounded in the Lp norm and such that supp ubn ⊂ K for all n ∈ N, then it has a
0
subsequence (unm )m∈N that converges in S to an element u of Lp , supp u
b ⊂ K, and

kukLp ≤ lim inf kunm kLp . (21.36)

m→∞

Proof. If p > 1, then this immediately follows from Lemma 21.30. If p = 1 and (un )n∈N
is bounded in L1 , then it is bounded in M as (see Lemma 2.26)

kf kL1 = kf kM (f ∈ L1 ), (21.37)

(where we identified f with the distribution f which corresponds to the Radon measure
f λ with λ the Lebesgue measure). By Lemma 21.30 it has a subsequence (unm )m∈N that
converges in S 0 to an element u of M such that (21.35) holds. As u bn → ub, supp ub ⊂ K.
This implies u ∈ Cp∞ , and therefore, by (21.37) we have u ∈ L1 and (21.36) for p = 1.
s is a Banach space that is continuously
Theorem 21.32. Let s ∈ R and p, q ∈ [1, ∞]. Bp,q
0 s that is bounded in the B s
embedded in S . Moreover, if (un )n∈N is a sequence in Bp,q p,q
norm, then it has a subsequence (unm )m∈N that converges in S 0 to an element u, which
s and
is also in Bp,q

kukBp,q
s ≤ lim inf kunm kBp,q
s .
m→∞

s is continuously embedded in S 0 follows from (21.23) in Theorem 21.18.

Proof. That Bp,q
s is complete after proving the “Moreover” statement.
We will prove that Bp,q
s . Without loss of generality we
Let (un )n∈N be a sequence that is bounded in Bp,q
may assume that kun kBp,q
s ≤ 1 for all n ∈ N. Then

k∆j un kLp ≤ 2−sj (n ∈ N, j ∈ N−1 ).

146
By applying Lemma 21.31 to (∆j un )n∈N for each j, and applying Cantor’s diagonal
argument, we find a subsequence (unm )m∈N of (un )n∈N and vj ∈ S 0 for all j ∈ N−1 such
that
m→∞
∆j unm −−−−→ vj in S 0 , kvj kLp ≤ lim inf k∆j unm kLp ≤ 2−sj (j ∈ N−1 ).
m→∞

As the support of the Fourier transform of ∆j un is in the annulus 2j A (or ball B), so is
the support of vbj for j ∈ N0 (for j = −1).
As k(2sj kvj kLp )j∈N−1 k`∞ ≤ 1, Theorem 21.18 (a) implies that v := j∈N−1 vj exists
P

in S 0 . Observe that vj = ∆j v for all j ∈ N−1 : As ∆j is continuous as a map S 0 → S 0 for

j ∈ N−1 , we have by letting ∆−2 u = 0 for all u ∈ S 0 ,
j+1
X j+1
X
∆j unm = ∆j ∆i unm → ∆j vj = ∆j v (j ∈ N−1 ).
i=j−1 i=j−1

Therefore
 
kvkBp,q
s = 2js kvj kLp
j∈N−1 `q
 
≤ 2js lim inf k∆j unm kLp
m→∞ j∈N−1 `q
 
≤ lim inf 2js k∆j unm kLp = lim inf kunm kBp,q
s .
m→∞ j∈N−1 `q m→∞

s is complete, we assume that the sequence (u )

To prove that Bp,q n n∈N as above is also
Cauchy. It suffices to show that with u = v as above, un → u in Bp,q s . Let ε > 0 and

N ∈ N be such that m, k ≥ N implies kuk − um kBp,q s < ε. Let k ≥ N . Apply the above
limiting argument to the sequence (un − uk )n∈N , so that for some sequence (nm )m∈N in
N

ku − uk kBp,q
s ≤ lim inf kunm − uk kBp,q
s < ε.
m→∞

s .
Then un → u in Bp,q

21.1 Comments ...

For our purposes we did not need this, but one can be a bit more specific about the
dependence of the C in the Bernstein inequalities on k as in Theorem 21.15 (or differently
said, in the statement we can interchange the “for all k ∈ N0 ” and “there exists a C > 0”),
as follows. For a proof see Exercise 25.A.

Theorem 21.33 (Bernstein inequalities). Let A be an annulus and B be a ball around

the origin in Rd . There exists a C > 0 such that for all k ∈ N and p, q ∈ [1, ∞] with

147
q ≥ p and any u ∈ S 0 we have for all λ > 0
k+d( p1 − 1q )
b ⊂ λB =⇒ max k∂ α ukLq ≤ C k+1 λ
supp u kukLp , (21.38)
α∈Nd0
|α|=k

b ⊂ λA =⇒ C −k−1 λk kukLp ≤ max k∂ α ukLp ≤ C k+1 λk kukLp .

supp u (21.39)
α∈Nd0
|α|=k

22  Diversity of Besov spaces and inclusions

In Example 21.26 we have seen that the inclusions that one obtains from the embeddings
in Theorem 21.23 and Lemma 21.25 are strict in case p1 = p2 = ∞. In this section
we show this for general p1 and p2 in [1, ∞] and consider under which conditions Besov
spaces are equal or included in each other.
We recall the Inverse Mapping Theorem, which is a consequence of the Open Mapping
Theorem (see for example [Con90, Theorem III.12.1 and III.12.5] and [Rud91, 2.11 and
2.12]).

Theorem 22.1 (Inverse Mapping Theorem). Let X and Y be Banach spaces. If A :

X → Y is a continuous linear bijection, then A is a homeomorphism.

Theorem 22.2. Let Z be a topological vector space. Let k · k1 and k · k2 be norm-like

functions on Z (Definition 21.10). Let

Xi = {x ∈ Z : kxki < ∞} (i ∈ {1, 2}).

Suppose that (X1 , k · k1 ) and (X2 , k · k2 ) are Banach spaces, that are both continuously
embedded in Z,

(X1 , k · k1 ) ,→ Z, (X2 , k · k2 ) ,→ Z. (22.1)

If X1 ⊂ X2 , then there exists a C > 0 such that k · k2 ≤ Ck · k1 . Consequently, if the sets

X1 and X2 are equal, then k · k1 and k · k2 are equivalent; (X1 , k · k1 ) and (X2 , k · k2 ) are
homeomorphic.

Proof. Suppose X1 ⊂ X2 . Let us write X = X1 . Define k · k3 : Z → [0, ∞] by k · k3 =

k · k1 + k · k2 . Then k · k3 is a norm-like function and X = {x ∈ Z : kxk3 < ∞}. By
definition, the identity map (X, k·k3 ) → (X, k·k1 ) is continuous (and linear and bijective).
We are done if we show that (X, k · k3 ) is a Banach space, as then, by Theorem 22.1, it
follows that also the identity map (X, k · k1 ) → (X, k · k3 ) is continuous, i.e., there exists
a C > 0 such that k · k3 ≤ Ck · k1 and thus k · k2 ≤ (C − 1)k · k1 . Let (xn )n∈N be a
Cauchy sequence in (X, k · k3 ). Then it converges in (X, k · k1 ) to some limit x and in
(X2 , k · k2 ) to some limit y. By (22.1) it follows that x = y. Therefore kxn − xk3 → 0
and so (X, k · k3 ) is indeed a Banach space.

148
 1
Lemma 22.3. Let q ∈ [1, ∞]. Let a : N → R be given by ( ∞ = 0)
1
1

q
a(n) = (n ∈ N).
(n + 1)(log(n + 1))2

Then a ∈ `r if and only if r ≥ q.R Let f : [2, ∞) → R be given by f (x) = (x(log x)2 )−1 (x ∈
[2, ∞)). Then, for α ∈ (0, ∞), 2∞ f α < ∞ if and only if α ≥ 1.

Proof. It is sufficient to show the statement for the function f as a(n)q ≤ f on [n, n + 1]
and f ≤ a(n)q on [n + 1, n + 2] for all n ∈ N with n ≥ 2. That f α is integrable for α > 1
will be clear. That f itself is integrable, follows as it is the derivative of (log X)−1 . Let
α ∈ (0, 1) and choose ε > 0 such that α(1 + 2ε) < 1. As x−ε log x → 0 as x → ∞, there
exists a C > 0 such that log x ≤ Cxε on [2, ∞), therefore
 1 α  1 α 1
2
≥ 1+2ε
≥ α α(1+2ε (x ∈ [2, ∞)).
x(log x) Cx C x )

So that f α is not integrable.

Remark 22.4. Observe that one needs the square on the logarithm, in the sense that
(x log x)−1 is not integrable.

Theorem 22.5. Let s, s1 , s2 ∈ R, p, p1 , p2 , q, q1 , q2 ∈ [1, ∞] and ε > 0.

s such that for all m, r ∈ [1, ∞] and t ∈ R,
(a) There exists a u ∈ Bp,q

t ≤ s − d( 1 − 1 ), r ≥ q,
t p m
u ∈ Bm,r ⇐⇒ (22.2)
t < s − d( 1 − 1 ), r < q.
p m

(b) Bps11,q1 = Bps22,q2 ⇐⇒ p1 = p2 , q1 = q2 , s1 = s2 .

(c)
s−d( p1 − p1 )
p1 < p2 =⇒ Bps1 ,q1 ( Bp2 ,q2 1 2
, s
q2 > q1 =⇒ Bp,q2
s−ε
( Bp,q 1
.

(d)

s ≤ s − d( 1 − 1 ), q 1 ≤ q2 ,
2 1
Bps11,q1 ⊂ Bps22,q2 ⇐⇒ p1 ≤ p2 and p1 p2
s2 < s1 − d( 1 − 1 ), q1 > q2 .
p1 p2

Proof. We consider a similar setting as in Example 21.26 (although, without restricting

d to be equal to 1. Let (ϕj )j∈N−1 be a dyadic partition of unity such that ϕ := ϕ0 = 1 on
the annulus A(1 − 2ε, 1 + 2ε) and ϕ−1 = 1 on the ball B(0, 12 + 2ε) for some ε > 0 (such
a dyadic partition of unity exists, see the proof of Theorem 21.6). Let ψ ∈ Cc∞ be such

149
that supp ψ ⊂ A(1 − ε, 1 + ε). Observe that ψb ∈ Lp for all p ∈ [1, ∞] as ψb is a Schwartz
function. For n ∈ Z define

vn = l2∗n ψ = 2−n l2−n ψ, un = vbn = l2n ψ.

b

Then,
nd
• un ∈ Lm and kun kLm = 2− m kψk
b Lm for all n ∈ Z, m ∈ [1, ∞] (Exercise 21.E).

• For n ∈ N and j ∈ N−1 , ϕj vn = vn and ∆j un is un if j = n and 0 if j 6= n.

b−n ⊂ 2−n A(1 − ε, 1 + ε) ⊂ B(0, 12 + ε) for all n ∈ N.

• supp u

• For n ∈ N and j ∈ N−1 , ∆j u−n = u−n if j = −1 and 0 if j ∈ N0 .

(a) Let us construct the u such that (22.2). Let p, q, m, r ∈ [1, ∞], s, t ∈ R. Let a be as
in Lemma 22.3, so that a ∈ `r if and only if r ≥ q. Define b : N → R by
−ns+ nd
b(n) = 2 p a(n) (n ∈ N).

exists in S 0 (Theorem 21.18 (a)) and

P
Then u := n∈N b(n)un

nt
1
n(t−s+d( p1 − m ))
kukBm,r
t [ϕ] = k(2 kb(n)un kLm )n∈N k`r = k(2 a(n))n∈N k`r kψk
b Lm .

Therefore, (22.2) holds.

(b) Then, for all m, r ∈ [1, ∞] and t ∈ R
nd
ku−n kBm,r
t = ku−n kLm = 2 m kψk
b Lm .

By Theorem 22.2 k · kBps1 ,q and k · kBps2 ,q cannot be equivalent if p1 6= p2 . Moreover, if

1 1 2 2
p1 < p2 , then by the above we see that for each C > 0 there exists an n ∈ N such that
nd nd
ku−n kBps1 ,q = 2 p1 > C2 p2 = Cku−n kBps2 ,q , so that there does not exist a C > 0 such
1 1 2 2
that k · kBps1 ,q ≤ Ck · kBps2 ,q , hence Bps22,q2 6⊂ Bps11,q1 :
1 1 2 2

p1 < p2 =⇒ Bps22,q2 6⊂ Bps11,q1 . (22.3)

s be as in (a). If t > s, then u ∈
Let u ∈ Bp,q t for any r ∈ [1, ∞]:
/ Bp,r
s1 s2
s1 6= s2 =⇒ Bp,q 1
6= Bp,q 2
.

And if r < q, then u ∈ s :

/ Bp,r
s s
q1 6= q2 =⇒ Bp,q 1
6= Bp,q 2
.

By the above three implications we obtain (b).

(c) follows by (b) and Theorem 21.23 and Lemma 21.25.

150
 (d) The ⇐= we have already seen in Theorem 21.23. Suppose

Bps11,q1 ⊂ Bps22,q2 . (22.4)

By (a) with p = p1 , q = q1 and s = s1 this implies either one of the following cases

s ≤ s − d( 1 − 1 ), q 1 ≤ q2 ,
2 1 p1 p2
(22.5)
s2 < s1 − d( 1 − 1 ), q1 > q2 .
p1 p2

(22.3) implies p1 ≤ p2 .

23 Embeddings of Besov spaces and Sobolev spaces

In this section we consider embeddings between Besov and Sobolev spaces. First we
s is equal to the fractional Sobolev spaces H s (Definition 20.5), as the
prove that B2,2
motivation of Besov spaces 21.1 already suggested. Then we compare Besov spaces with
Sobolev spaces and consider in which of these the spaces the testfunctions are dense.
s = H s we use the following lemma.
For the proof of B2,2

Lemma 23.1. Let (ϕj )j∈N−1 be a dyadic partition of unity. For all s ∈ R there exist
c, C > 0 such that
 s X  s
c 1 + |ξ|2 ≤ 22sj ϕj (ξ)2 ≤ C 1 + |ξ|2 (ξ ∈ Rd ). (23.1)
j∈N−1

Proof. Let a, b > 0, a < b be such that supp ϕ0 ⊂ A(a, b). Because supp ϕ0 ⊂ A(a, b), we
have supp ϕ−1 ⊂ B(0, 2b ) = 2−1 B(0, b) (as for example ϕ−1 + ϕ0 = l 1 ϕ−1 and supp ϕ−1 +
2
ϕ0 ⊂ B(0, b)). Without loss of generality, we may assume a < 1 and b > 1. Using that
1 + 22j b2 ≤ b2 (1 + 22j ) and that 1 ≤ 4 · 22j for all j ∈ N−1 , we obtain

ξ ∈ supp ϕj =⇒ ξ ∈ 2j B(0, b) =⇒ 1 + |ξ|2 ≤ 5b2 22j (j ∈ N−1 ),

j 2 2j 2
ξ ∈ supp ϕj =⇒ ξ ∈ 2 A(a, b) =⇒ 1 + |ξ| ≥ 2 a (j ∈ N0 ),
ξ ∈ supp ϕ−1 =⇒ 1 + |ξ|2 ≥ 1 ≥ 2−2 a2 .

Therefore
1 + |ξ|2 1 + |ξ|2
ξ ∈ supp ϕj =⇒ ≤ 1 ≤ (j ∈ N0 ).
5b2 22j a2 22j
Let Aj = {ξ ∈ Rd : ϕ2j (ξ) ≥ 41 } for j ∈ N−1 . Observe that by definition of the dyadic
partition of unity, see (21.3) and (21.4), it follows that
[
Aj = R d .
j∈N−1

151
Let t ≥ 0. Then
!t !−t
1 1 + |ξ|2 1 1 + |ξ|2
ξ ∈ Aj =⇒ ≤ ϕ2j (ξ) and ≤ ϕ2j (ξ),
4 5b2 22j 4 a2 22j
!t !−t
1 + |ξ|2 1 + |ξ|2
ξ ∈ supp ϕj =⇒ ϕ2j (ξ) ≤ and ϕ2j (ξ) ≤ .
a2 22j 5b2 22j
Then, because
|i − j| ≥ 2 =⇒ Ai ∩ Aj = supp ϕi ∩ supp ϕj = ∅ (i, j ∈ N−1 ),
we have
1  t X
−2t
 t
1 + |ξ|2
≤ 2 2tj
ϕj (ξ)2
≤ a 1 + |ξ|2
(ξ ∈ Rd ),
4(5b2 )t j∈N −1

a2t  −t  −t

2−2tj ϕj (ξ)2 ≤ (5b2 )t 1 + |ξ|2
X
1 + |ξ|2 ≤ (ξ ∈ Rd ).
4 j∈N−1

Theorem 23.2. For all s ∈ R we have

s
B2,2 = H s,
with equivalent norms.

Proof. By the Plancherel formula (Theorem 16.27),

X X
kuk2B2,2
s = 22sj kϕj (D)uk2L2 = 22sj kϕj u
bk2L2
j∈N−1 j∈N−1
Z X
= 22sj |ϕj (ξ)|2 |u
b(ξ)|2 dξ.
Rd j∈N
−1

The rest follows from Lemma 23.1.

23.3. In particular, Theorem 23.2 implies L2 = B2,20 . However, there do not exist s ∈ R,
1 s
p, q ∈ [1, ∞] such that L = Bp,q , see Exercise 23.A.

Exercise 23.A. Show that there do not exist s ∈ R, p, q ∈ [1, ∞] such that L1 =
s . Hint: Use the property of Theorem 21.32 and Exercise 21.L and observe that by
Bp,q
s are equal if and only if the norm-like functions k · k
Theorem 22.2 the sets L1 and Bp,q L1
and k · kBp,q
s are equivalent.
Theorem 23.4. Let k ∈ N0 and p, q ∈ [1, ∞]. Then
k
Bp,1 ,→ W k,p ,→ Bp,∞
k
,
t
Bp,q ,→ W k,p ,→ Bp,q
s
(s, t ∈ R, s < k < t).
More specifically:

152
 (a) For all k ∈ N0 and s ∈ R with s < k there exists a C > 0 such that
kukBp,q
s ≤ CkukW k,p (u ∈ S 0 , p, q ∈ [1, ∞]) (23.2)
kukBp,∞
k ≤ CkukW k,p (u ∈ S 0 , p ∈ [1, ∞]). (23.3)

(b) For all k ∈ N0 and t ∈ R with t > k there exists a C > 0 such that
kukW k,p ≤ CkukBp,q
t (u ∈ S 0 , p, q ∈ [1, ∞]) (23.4)
kukW k,p ≤ CkukB k (u ∈ S 0 , p ∈ [1, ∞]). (23.5)
p,1

Proof. Proof of (a). Let k ∈ N0 . By Bernstein’s inequality (Theorem 21.15) there exists
a C1 > 0 such that
k∆j ukLp ≤ C1 2−kj max k∂ β ∆j ukLp (u ∈ S 0 , j ∈ N0 , p ∈ [1, ∞]).
β∈Nd0 :|β|=k

By Lemma 21.9 (in particular (21.12) and (21.13)) there exists a C2 > 0 such that
k∆j ukLp ≤ C2 kukLp (u ∈ S 0 , j ∈ N−1 , p ∈ [1, ∞]).
Let s ∈ R and q ∈ [1, ∞] be such that either s < k or s = k and q = ∞. Then
M = k2(s−k)j )j∈N−1 k`q < ∞.
and thus, with C 0 = C1 C2 M ,
 
kukBp,q
s ≤ C 0 kukLp ∨ max k∂ β ukLp ≤ C 0 kukW k,p (u ∈ S 0 , p ∈ [1, ∞]).
β∈Nd0 :|β|=k

This implies (23.3). If s < k, then M ≤ k2(s−k)j )j∈N−1 k`1 (Lemma A.8) so that (23.2)
holds for all q ∈ [1, ∞] with C = C1 C2 k2(s−k)j )j∈N−1 k`1 .
1
Proof of (b). First, observe the following. Let r ∈ [1, ∞] being such that 1 = r + 1q .
Then by Hölder’s inequality (Corollary A.9),
∞
2−aj 2aj k∆j ukLp ≤ k(2−aj )j∈N−1 k`r kukBp,q (u ∈ S 0 , a ∈ R, p ∈ [1, ∞]).
X
kukLp ≤ a

j=−1

By Theorem 21.23 there exists a C3 > 0 such that

k∂ α ukBp,q
t−k ≤ C3 kukB t
p,q
(u ∈ S 0 , α ∈ Nd0 , |α| ≤ k, p, q ∈ [1, ∞]).

Let t ∈ R and q ∈ [1, ∞] be such that either t > k or t = k and q = 1 (and thus r = ∞).
Then
N = k(2−(t−k)j )j∈N−1 k`r < ∞,
and thus for all α ∈ Nd0 with |α| ≤ k,
k∂ α ukLp ≤ N k∂ α ukBp,q
t−k ≤ C3 N kukB t
p,q
(u ∈ S 0 , p ∈ [1, ∞]).

This implies (23.5). If t > k, then N ≤ k(2−(t−k)j )j∈N−1 k`1 so that (23.4) holds for all
p, q ∈ [1, ∞] with C = C3 k(2−(t−k)j )j∈N−1 k`1 .

153
 0
For p = ∞ we do not only have that B∞,1 is embedded into the bounded functions, but
also into the continuous bounded functions. Moreover, we have the following statement.
Theorem 23.5. For all k ∈ N0
k
B∞,1 ,→ Cbk ,→ B∞,∞
k
.

Proof. That Cbk ,→ B∞,∞

k follows by the fact that W k,∞ ,→ B∞,∞
k and Cbk ,→ W k,∞ . By
k k
Theorem 23.4 it is sufficient to show the inclusions B∞,1 ⊂ Cb . As for u ∈ B∞,1 k one
α 0 d
has ∂ u ∈ B∞,1 for all α ∈ N0 with |α| ≤ k (by Theorem 21.23), it is sufficient to show
0
B∞,1 ⊂ Cb0 . We leave the proof of this for the reader (see Exercise 23.B).
0
Exercise 23.B. Prove that any element of B∞,1 is (represented by) a continuous func-
tion.
Corollary 23.6. Let s ∈ R and p, q ∈ [1, ∞]. Then
s
D ,→seq S ,→ Bp,q .

Proof. By Theorem 14.11 D ,→seq S. Let k ∈ N0 be such that k > s. By Lemma 14.15
and Theorem 23.4

S ,→ W k,p ,→ Bp,q
s
.

s in case p and q are both

In Theorem 23.8 we will show that D is also dense in Bp,q
finite. For this we will use the following lemma.
PJ
Lemma 23.7. Let s ∈ R and p, q ∈ [1, ∞]. Suppose q < ∞. Then j=−1 ∆j u → u in
s as J → ∞ for all u ∈ B s .
Bp,q p,q

Exercise 23.C. Prove Lemma 23.7.

Theorem 23.8. Let s ∈ R and p, q ∈ [1, ∞].

(a) If q < ∞ then Cb∞ ∩ Bp,q

s is dense in B s .
p,q
s .
(b) If p < ∞ and q < ∞, then D is dense in Bp,q

Proof. By (b) it is sufficient to prove (a) for p = ∞. But this follows from Lemma 23.7
as ∆j u ∈ L∞ for all j ∈ N−1 and u ∈ Bp,q s .

s such that
Let p, q < ∞. By Lemma 23.7 it is sufficient to show that for all u ∈ Bp,q
R→∞ s .
b is compact, there exist uR ∈ D for R > 0 such that uR −
supp u −−−→ u in Bp,q
s be such that supp u
Let u ∈ Bp,q b is compact. Then u ∈ Cp∞ (Lemma 17.8) and ∂ α u ∈
Lp for all α ∈ N0 by Bernstein’s inequality (Theorem 21.15) and because u = Jj=−1 ∆j u
d P

for some J ∈ N−1 and ∆j u ∈ Lp for all j ∈ N−1 . Consequently, u ∈ W k,p for all k ∈ N0 .

154
 Let χ and χR for R > 0 be as in Lemma 8.16. Then χR u ∈ D and

k∂ α (χR − 1) · ∂ β ukLp −−−−→ 0

R→∞
(α, β ∈ Nd0 ). (23.6)

Let k ∈ N0 , k > s. As W k,p ,→ Bp,q

s (by Theorem 23.4) it is sufficient to show that

k(χR − 1)ukW k,p −−−−→ 0. But this follows by (23.6) because by Leibniz’ rule (see 1.14)
R→∞

there exists a C > 0 such that

k(χR − 1)ukW k,p ≤ C max max k∂ α (χR − 1) · ∂ β ukLp .

β∈Nd0 α∈Nd0
|β|≤k |α|≤k

Exercise 23.D. Show that 1 ∈ B∞,q s if and only if either s = 0 and q = ∞ or s < 0.
∞ s
Conclude: Cb ⊂ B∞,q if and only if either s = 0 and q = ∞ or s < 0.
s ). As ∆ 1 = 1, we have 1 ∈ B s
23.9 (S is not dense in B∞,q −1 ∞,q for all s ∈ R and
q ∈ [1, ∞]. For each ϕ ∈ S we have ∆−1 ϕ ∈ S and thus k∆−1 ϕ − ∆−1 1kL∞ = 1.
s
Therefore S and thus D are not dense in B∞,q for any s ∈ R and q ∈ [1, ∞].

The following lemma implies that if S is dense in Bp,q s , that for all u ∈ B s one has
p,q
limj→∞ 2 k∆j ukLp = 0. Observe however, that the converse is not the case: If u ∈ S 0 is
js
s (even though
such that (23.8), this need not to imply that u is in the closure of D in Bp,q
this is claimed to be obvious in [BCD11, Remark 2.75]); indeed, for p = ∞ this is the
case for 1, see 23.9.

Lemma 23.10. For ψ ∈ S

lim 2js k∆j ψkLp = 0 (s ∈ R, p ∈ [1, ∞]). (23.7)

j→∞

s and u is the limit of testfunctions in B s ,

Consequently, if s ∈ R, p, q ∈ [1, ∞], u ∈ Bp,q p,q
s , then
in other words, u is in the closure of S in Bp,q

lim 2js k∆j ukLp = 0. (23.8)

j→∞

Proof. (23.7) basically follows because S ⊂ Bp,qr for all r ∈ R and p, q ∈ [1, ∞]. The

details and the proof of the rest of the statement are left for the reader (Exercise 23.E).

Exercise 23.E. Complete the proof of Lemma 23.10.

In the following example we will show that the inclusion W k,∞ ⊂ B∞,∞
k which one
∞ 0
obtains by Theorem 23.4 is strict. Therefore, in particular L ( B∞,∞ .

155
Example 23.11 (W k,∞ ( B∞,∞
k ). Consider the setting of Example 21.26 with q = ∞.
Let

un = cos(2π2n X), wn = sin(2π2n X) (n ∈ N).

Then Uk :=
P −nk u and Wk :=
P −nk w are tempered distributions for k ∈ N0
n∈N 2 n n∈N 2 n
and
(
t t = k, r = ∞
Uk , Wk ∈ B∞,r ⇐⇒ .
t<k

And, as ∂2−n wn = 2πun , ∂2−n un = −2πwn for all n ∈ N,

∂Wk+1 = 2πUk and ∂Uk+1 = −2πWk in S 0 (k ∈ N0 ).

Let k ∈ N. We will show that L∞ ( B∞,∞

k / L∞ (remember
by showing that ∂ k Uk ∈
L∞ = W 0,∞ ).
It is sufficient to show that ∂ k Uk ∈
/ L1loc .
∂ k Uk is a multiple of W0 or U0 . Therefore, we may as well assume k = 0. Write
u = U0 and w = 2π 1
W1 so that ∂w = u in S 0 .
w is a Weierstrass function, as Hardy showed, see [Har16] (he showed that functions
of the form n∈N0 an cos(bn π X) or n∈N0 an sin(bn π X) for a ∈ (0, 1) and b ∈ (1, ∞) with
P P

ab ≥ 1 are nowhere differentiable). This means that w is a continuous function that is

nowhere differentiable.
By showing the following statement we conclude that u is not given by a locally
integrable function and therefore not in L∞ :
Lemma 23.12. Let u, w ∈ L1loc (R) and suppose ∂w = u in D0 (R). Then w is almost
everywhere differentiable with derivative u.

Proof. First observe that for a, b ∈ R, a < b and ϕ ∈ D(R), ϕ ∗ 1[a,b] (x) =
R x−a
x−b ϕ for
x ∈ R and thus

∂(ϕ ∗ 1[a,b] ) = Ta ϕ − Tb ϕ in D(R).

Choose a mollifier ψ with 0 ≤ ψ(x) ≤ 1 for all x ∈ R and let ψε for ε > 0 be as in
Definition 8.11.
Let a, b ∈ R, a < b. By Theorem 7.15 ψε ∗ 1[a,b] −−→ 1[a,b] pointwise except possibly
ε↓0

at a and at b, whereas 0 ≤ ψε 1[a,b] (x) ≤ 1 for all x ∈ R and ε > 0. By Lebesgue’s

dominated convergence theorem it follows that
Z b Z Z
u= u1[a,b] = lim u · ψε ∗ 1[a,b] = limhu, ψε ∗ 1[a,b] i = limhw0 , ψε ∗ 1[a,b] i
a ε↓0 ε↓0 ε↓0

= − limhw, ∂(ψε ∗ 1[a,b] )i = limhw, Tb ψε − Ta ψε i = w(b) − w(a).

ε↓0 ε↓0

156
By Lebesgue’s differentiation theorem it then follows that w is almost everywhere differ-
entiable with ∂w = u almost everywhere, because for example
Z x+|h|
w(x + h) − w(x) 1
− u(x) ≤ |u(s) − u(x)| ds (h 6= 0).
h h x−|h|

Exercise 23.F. Show that Uk as in Example 23.11 does not lie in the closure of D in
k
B∞,∞ .

24  Besov spaces related to other spaces

In this section we give an overview of Besov spaces and some other spaces, and of em-
beddings between them.
Definition 24.1 (Hölder spaces). Let Ω be an open subset of Rd and k ∈ N0 . We write
also C k,0 (Ω) for C k (Ω). Let α ∈ (0, 1].

• A function f : Ω → F is α-Hölder continuous if there exists a C > 0 such that

|f (x) − f (y)| ≤ C|x − y|α (x, y ∈ Ω). (24.1)

• A function is called Lipschitz continuous if it is 1-Hölder continuous.

• C 0,α (Ω) is defined to be the space of α-Hölder continuous functions Ω → F. The
α-Hölder coefficient of a function f is given by
|f (x) − f (y)|
|f |C 0,α (Ω) = sup .
x,y∈Ω:x6=y |x − y|α

• C k,α (Ω) is defined to be the space of functions Ω → F that are k-times continuously
differentiable for which their derivatives of order k are α-Hölder continuous.

We defined Cbk (Ω) to be the subspace of C k (Ω) that consists of functions f for which
kf kC k is finite. Similarly we define
X
kf kC k,α (Ω) = kf kC k (Ω) + |∂ β f |C 0,α (Ω) (f ∈ C k,α (Ω)), (24.2)
β∈Nd0 :|β|=k

Cbk,α (Ω) = {f ∈ C k,α (Ω) : kf kC k,α (Ω) < ∞}. (24.3)

24.2. For the rest of this section we consider Ω = Rd and write “C k,α ” instead of
“C k,α (Rd )”.
Observe that C 0,1 consists of all the Lipschitz functions and that for k ∈ N, Cbk+1 (
Cbk,1 .
For s ∈ (0, ∞) \ N it is also common in literature to write C s for C k,α , where k = bsc
and α = s − bsc.

157
Exercise 24.A. Can you classify the space of α-Hölder functions with α > 1, that is,
which functions f satisfy (24.1) for α > 1?

In Definition 12.3 we introduced the Sobolev spaces W k,p for k ∈ N0 and p ∈ [1, ∞].
In Definition 20.5 and 20.10 we introduced the fractional Sobolev or Bessel–potential
spaces Hps for s ∈ R \ N0 and p ∈ [1, ∞]. We will now consider Slobodeckij spaces, W s,p
with s ∈ (0, ∞) \ N as subspaces of W k,p with k = bsc in a similar way as C s or C k,α for
α ∈ (0, 1] is defined to be a subspace of C k .

Definition 24.3 (Slobodeckij spaces). Let p ∈ [1, ∞) and s ∈ (0, ∞) \ N. Let k ∈ Z and
α ∈ (0, 1) be given by

k = bsc, α = s − k.

We define the norm-like function k · kW s,p : W k,p → [0, ∞] by

!1
|∂ β f (x) − ∂ β f (y)|p p
X Z Z
kf kW s,p := kf kW k,p + dx dy (f ∈ W k,p ).
Rd Rd |x − y|d+αp
β∈Nd0 :|β|=k

The Slobodeckij space W s,p is then defined by

W s,p = {f ∈ W k,p : kf kW s,p < ∞}.

Definition 24.4 (Zygmund spaces). Let s ∈ (0, ∞). Let k ∈ N0 and α ∈ (0, 1] be given
by

k = ds − 1e, α = s − k, (24.4)

in other words, k is such that s − k ∈ (0, 1]. We define the norm-like function k · kC s :
C k → [0, ∞], by

X k(Th − 1)2 ∂ β f kC 0
kf kC s = kf kC k + sup (f ∈ C k ).
h∈Rd \{0} |h|α
β∈Nd0 :|β|=k

The Zygmund space C s is then defined by

C s = {f ∈ C k : kf kC s < ∞}.

Observe that

(Th − 1)2 g(x) = (Th − 1)(Th − 1)g(x) = (Th − 1)g(x − h) − (Th − 1)g(x)
= g(x − 2h) − 2g(x − h) + g(x).

158
Definition 24.5 (Besov–Lipschitz spaces). Let s ∈ (0, ∞). Let k ∈ Z and α ∈ (0, 1] be
as in (24.4). For p, q ∈ [1, ∞] we define the norm-like function k · kΛsp,q : W k,p → [0, ∞]
by
  1
 k(Th −1)2 ∂ β f kqLp
R q
P q < ∞,

β∈Nd0 :|β|=k Rd dh
(f ∈ W k,p ).
|h| d+αq
kf kΛsp,q := kf kW k,p +
 k(Th −1)2 ∂ α f kqLp
q = ∞,
P
β∈Nd0 :|β|=k suph∈Rd \{0}


|h|α

For p, q ∈ [1, ∞] we define the Besov-Lipschitz space Λsp,q to be the set of functions

Λsp,q = {f ∈ W k,p : kf kΛsp,q < ∞}.

The Triebel–Lizorkin spaces are defined as the Besov spaces, but with the “Lp ” and
“`q ”norm interchanged:

Definition 24.6 (Triebel–Lizorkin spaces). Let (ϕj )j∈N−1 be a dyadic partition of unity.
Let s ∈ R. For p ∈ [1, ∞) and q ∈ [1, ∞] we define the norm-like function k · kFp,q
s : S0 →
[0, ∞] by

kukFp,q
s := k(2js |∆i u|)j∈N−1 k`q Lp
(u ∈ S 0 ),

for example, for q < ∞ this means

 1
p
Z  X p
q
kukFp,q
s = 2qjs |∆i u(x)|q dx .
Rd j∈N−1

s to be the set
We define the Triebel–Lizorkin space Fp,q
s
Fp,q = {u ∈ S 0 : kukFp,q
s < ∞}.

Remark 24.7. As for Besov spaces, the norm of Fp,q s depends on the choice of dyadic

partition, but the space itself does not. This is shown in [Tri83, Section 2.3.2] .

24.8. Let us summarize for which parameters we have either continuous embeddings or
equality between spaces with equivalent norms. Here, “A =∼ B” means that A and B are
the same space with equivalent norms, i.e., A ,→ B ,→ A.
∼ C s for s ∈ (0, ∞) \ N (C s is as in 24.2).
(a) [Tri83, p.90, (9)] Cbs =
(b) [Tri83, p.90, (9)] W s,p ∼
= Λsp,p for s ∈ (0, ∞) \ N and p ∈ (1, ∞).
(c) [Tri83, p.88] H s ∼
p= F s for s ∈ R and p ∈ (1, ∞).
p,2

(d) [Tri83, p.88] Hpk ∼

= W k,p for k ∈ N and p ∈ (1, ∞).
0 ,→ Lp ,→ B 0
(e) [Tri83, p.89] Bp,1 p,∞ for p ∈ [1, ∞). (See also Theorem 23.4.)
(f) [Tri83, p.89] 0
B∞,1 ,→ Cb0 ,→ B∞,∞
0 . (See also Theorem ??.)

159
 (g) [Tri83, p.90, p.113] Λsp,q ∼ s for s > 0, p ∈ [1, ∞) and q ∈ [1, ∞].
= Bp,q
(h) [Tri83, p.90, p.113] C s ∼ s
= B∞,∞ for s > 0.
s s ,→ B s
(i) [Tri83, p.47] Bp,min{p,q} ,→ Fp,q p,max{p,q} for s ∈ R, p ∈ [1, ∞) and q ∈ [1, ∞].
(j) [Tri83, p.60] For s1 , s2 ∈ R, p1 , p2 , q1 , q2 ∈ [1, ∞]:
Bps11,q1 (Rd ) ∼
= Bps22,q2 (Rd ) if and only if s1 = s2 and p1 = p2 , q1 = q2 (see The-
orem 22.5).
For p1 , p2 < ∞:
Fps11,q1 (Rd ) ∼
= Fps22,q2 (Rd ) if and only if s1 = s2 and p1 = p2 , q1 = q2 ,
Fp11,q1 (R ) ∼
s d
= Bps22,q2 (Rd ) if and only if s1 = s2 and p1 = p2 = q1 = q2 .
Observe that we can combine some of the above to obtain:

Cbs = C s = B∞,∞
s
(s ∈ (0, ∞) \ N),
s,p
W = Λsp,p s
= Bp,p s
= Fp,p (s ∈ (0, ∞) \ N, p ∈ (1, ∞)).
Hpk = W k,p k
= Fp,2 (k ∈ N, p ∈ (1, ∞)).

24.9. In 23.3 we mentioned that no Besov space is equal to L1 . We can generalise this
as follows: For r ∈ [1, 2) ∪ (2, ∞) there are no s ∈ R, p, q ∈ [1, ∞] such that Lr = Bp,q
s .

s = Lr if and only if p = q = r = 2 and

Exercise 24.B. Let r ∈ (1, ∞). Show that Bp,q
0 r
s = 0. Hint: Hr = L (see 20.10).

Remark 24.10. The proof of C α = B∞,∞

α for α ∈ (0, 1) can also be found in [MS13,
Lemma 8.6] .

25 The Hörmander-Mikhlin inequalities for Fourier multi-

pliers

The derivation operator ∂ α maps C k into C k−|α| . It behaves similar on Besov spaces as
we have seen in Theorem 21.23, namely, ∂ α maps Bp,q s continuously into B s−|α| .
p,q

In this section we consider the action of Fourier multipliers on Besov spaces, and ba-
sically show that those who behave similarly as the Fourier multiplier ∂ α , i.e., (2πiX)α (D),
also map Bp,qs continuously into B s−|α| . Moreover, we can also use this to treat “inverse
p,q
derivation operators”: for example (1 − ∆)−1 forms a continuous map from Bp,q s into
s+1 . This turns out to be very useful in order to solve (elliptic) partial differential
Bp,q
equations, as we will see in Section 28 (it allows us to find a solution by finding a fixed
point of a map that involves an inverse of the form (β − ∆)−1 ).
Like Theorem 21.23 it proven by using the Bernstein inequalities Theorem 21.15,
which describe the action of ∂ α on Lp functions whose Fourier transforms are supported
within annuli and balls, we start here by considering similar inequalities for Fourier
multipliers on Lp functions whose Fourier transforms are supported within annuli or

160
balls. Then we use this to obtain the action of Fourier multipliers on Besov spaces and
apply this to the fractional Laplacian (−∆)s and Bessel potentials (1 − ∆)s (see 19.9 and
19.5). A summary of the main results in this section is given at the end, without the use
of some of the introduced notation in this section, see Theorem 25.18.
The first theorem is a direct consequence of Young’s inequality and describes the
action of a Fourier multiplier of a function whose Fourier transform is integrable on both
Lp and Besov spaces.

Theorem 25.1. If σ ∈ Cp∞ and σ

b ∈ L1 , then there exists a C > 0 such that for all
s ∈ R, p, q ∈ [1, ∞]

kσ(D)ukLp ≤ CkukLp (u ∈ Lp ),
s
kσ(D)ukBp,q
s ≤ CkukBp,q
s (u ∈ Bp,q ).

So σ(D) forms a continuous map Lp → Lp and Bp,q

s → B s . One may take C = kσ
p,q b kL1 .

Proof. The first inequality follows as σ(D)u = F −1 (σ) ∗ u for all u ∈ S 0 , so that
kσ(D)ukLp ≤ kσ b kL1 kukLp . The inequality for the Besov norms follows by applying the
inequality for the Lp norm to ∆j u for all j ∈ N−1 .

Now we continue to consider Fourier multipliers for smooth σ whose Fourier transform
may not be integrable, but which satisfy some other conditions. We first turn to the action
of Fourier multipliers σ(D) on Lp functions u whose Fourier transforms u b are supported
in annuli. For these Fourier multipliers the function σ does not need to be defined on
the whole of Rd (Definition 19.7).
For example if Ω ⊂ Rd is open, σ ∈ C ∞ (Ω) and u ∈ Lp , supp u b is compact. Then
σ(D)u = (σχ)(D)u for some χ ∈ Cc . So φ = σχ is in Cc . As φ(D)u = F −1 (φ) ∗ u,
∞ ∞

by Young’s inequality we have kφ(D)ukLp ≤ kF −1 φkL1 kukLp . Furthermore we have

kF −1 φkL1 = kRFφkL1 = kφkb 1 , so that this motivates us to consider estimates of kφk
L
b 1
L
first:

Lemma 25.2. Let k = 2b1 + d2 c and r ∈ [1, ∞). There exists an M > 0 such that
k
kgbkLr ≤ M k(1 − ∆) 2 gkL1 (g ∈ S 0 ). (25.1)

Moreover, for all compact K ⊂ Rd there exists a C > 0 such that

kφk
b Lr ≤ Ckφk k
C (φ ∈ Cc∞ , supp φ ⊂ K).

Proof. We have already used the following inequality a couple of times, but let us recall
k
it. Observe that (1 + 4π 2 |X|2 )− 2 and thus the r-th power of this function are integrable
k
(Lemma 14.13). Let M = k(1 + 4π 2 |X|2 )− 2 kLr . Then
k k k
kf kLr = k(1 + 4π 2 |X|2 )− 2 (1 + 4π 2 |X|2 ) 2 f kLr ≤ M k(1 + 4π 2 |X|2 ) 2 f kL∞ (f ∈ S 0 ),

161
which implies (25.1).
By the Multinomial Theorem (Theorem 20.4) there exist cα ∈ R for α ∈ Nd0 with
|α| ≤ k such that
k X
(1 + 4π 2 |x|2 ) 2 = cα (2πix)α (x ∈ Rd ),
α∈Nd0
|α|≤k

and thus
k X
(1 − ∆) 2 = cα ∂ α .
α∈Nd0
|α|≤k

{|cα | : α ∈ Nd0 , |α| ≤ k}. Then, with L ∈ [0, ∞) being the Lebesgue measure
P
Let C =
of K,
k
(φ ∈ Cc∞ , supp φ ⊂ K).
X
k(1 − ∆) 2 φkL1 ≤ |cα | · k∂ α φkL1 ≤ CLkφkC k
α∈Nd0
|α|≤k

Exercise 25.A. Let A, B, χ and φ be as in Lemma 21.14: Let A be an annulus and B

be a ball centered at the origin in Rd , χ ∈ Cc∞ be equal to 1 on a neighbourhood of B
and φ ∈ Cc∞ be supported in an annulus equal to 1 on a neighbourhood of A.
Show that there exists a a C > 0 such that for all r ∈ [1, ∞]

khα kLr , kgα kLr ≤ C |α|+1 (α ∈ Nd0 ).

With this, prove Theorem 21.33.

Exercise 25.B. Prove the following statement: Let p ∈ [1, ∞] and k = 2b1 + d2 c. Let
u ∈ Lp . Then there exists a C > 0 such that

|hu
b, ψi| ≤ Ckψkk,S (ψ ∈ S).

As a consequence of the previous lemma we have the following, which is already

mentioned in the text preceding Lemma 25.2.

Lemma 25.3. Let k = 2b1 + d2 c. Then there exists a C > 0 such that for all open
Ω ⊂ Rd , σ ∈ Cp∞ (Ω), φ ∈ Cc∞ such that supp φ ⊂ Ω and p ∈ [1, ∞]

k(σφ)(D)ukLp ≤ CkφkC k kσkC k ,supp φ kukLp (u ∈ Lp ). (25.2)

Proof. This follows by Young’s inequality, Lemma 25.2 and Proposition 5.3.

162
 As we will consider the complement of {0} and of closed balls numerous times in this
section, it makes sense to introduce a short notation. Remember Definition 21.3.

Definition 25.4. For a ∈ [0, ∞) we write

A(a, ∞) = {x ∈ Rd : |x| > a}.

Observe that A(0, ∞) = Rd \ {0} and that for λ > 0,

λA(a, ∞) = A(λa, ∞), λA(b, c) = A(λb, λc) (a ∈ [0, ∞), b, c ∈ (0, ∞), b < c).

We use the following functions, called Mikhlin norms, to describe the action of Fourier
multipliers on Lp .

Definition 25.5 (Mikhlin norm). Let m ∈ R, k = 2b1 + d2 c and θ ≥ 0. We define

Mm,θ : C k (A(θ, ∞)) → [0, ∞] by

Mm,θ (σ) = max sup |x||α|+m |∂ α σ(x)| (σ ∈ C k (A(θ, ∞))).

α∈Nd0 :|α|≤k x∈A(θ,∞)

Even though Mm,θ is only a norm on the space {σ ∈ C k (A(θ, ∞)) : Mm,θ (σ) < ∞}, we
call Mm,θ a Mikhlin norm.
We will also write “Mm ” instead of “Mm,0 ”.
Let σ ∈ C k (A(θ, ∞)). Then we observe the following facts.

(a) Mm,θ (σ) < ∞ if and only if there exists a C > 0 such that

|∂ α σ(x)| ≤ C|x|−m−|α| (x ∈ A(θ, ∞), α ∈ Nd0 , |α| ≤ k). (25.3)

Moreover, if Mm,θ (σ) < ∞, then (25.3) is valid for C = Mm,θ (σ).
(b)

Mm,a (σ) ≤ Mm,θ (σ) (a ≥ θ). (25.4)

(c)

Mm,a (lλ σ) = λ−m Mm,λa (σ) (λ > 0, a ≥ λθ ). (25.5)

(d) If θ > 0 and Mm,a (σ) is finite for some a ≥ θ, then Mm,b (σ) is finite for all b ≥ θ
(we view C k (A) as a subset of C k (B) if B ⊂ A).

Example 25.6. (a) For all β ∈ Nd0 we have Mm (Xβ ) < ∞ if and only if m =
−|β|, and thus M−n (P ) < ∞ for n ∈ N0 and P being a polynomial of the form
α
α∈Nd ,|α|=n cα X for some cα ∈ F.
P
0

(b) Mm,1 (ψ) < ∞ for m ∈ R and ψ ∈ S, as every derivative of ψ is of rapid decay.

163
 Now by the previous lemma we obtain the following inequalities for Fourier multipliers
acting on Lp functions whose Fourier transforms are supported in annuli.
Lemma 25.7. Let m ∈ R. Let A be an annulus in Rd . There exists a C > 0 such that
for all p ∈ [1, ∞], λ > 0 and all σ ∈ C ∞ (Rd \ {0})

b ⊂ λA =⇒ kσ(D)ukLp ≤ CMm (σ)λ−m kukLp

supp u (u ∈ Lp ). (25.6)

Moreover, if a > 0 is such that A ⊂ A(a, ∞), then there exists a C > 0 such that for all
p ∈ [1, ∞], λ > 0 and all σ ∈ C ∞ (Rd \ {0})

b ⊂ λA =⇒ kσ(D)ukLp ≤ CMm,λa (σ)λ−m kukLp

supp u (u ∈ Lp ). (25.7)

Proof. It is sufficient to prove the “Moreover” statement because Mm,λa (σ) ≤ Mm (σ)
for all λ > 0 and σ ∈ C ∞ (Rd \ {0}).
Step 1 Let us first argue that it is sufficient to consider λ = 1 only. Let λ > 0 and
u ∈ Lp be such that supp u b ⊂ λA. Then supp lλ u b ⊂ A and thus supp F(l 1 u) ⊂ A. Let
λ
us write v = l 1 u so that u = lλ v. By Lemma 19.4 (19.2)
λ

− dp
kσ(D)vkLp = klλ [(lλ σ)(D)u]kLp = λ k(lλ σ)(D)ukLp ,
d
kvkLp = kl 1 ukLp ≤ λ p kukLp .
λ

Therefore, if (25.6) holds for λ = 1, we can apply it to v and then obtain (25.6) for any
λ > 0 because Mm (lλ σ) = λ−m Mm (σ), see (25.5).
Step 2 Let b, c, d > 0 be such that a < b < c < d and A = A(b, c). As A ⊂ A(a, ∞),
we have a < b. Let φ ∈ Cc∞ be such that φ = 1 on a neighbourhood of A and supp φ ⊂
A(a, d). Then σ(D)u = (σφ)(D)u for all u ∈ S 0 with supp u
b ⊂ A and all σ ∈ Cp∞ (Rd \{0}).
By Lemma 25.3 we can conclude the existence of a C > 0 such that (25.6) by observing
that for all β ∈ Nd0 with |β| ≤ k we have

sup |∂ β σ(ξ)| ≤ Mm,a (σ) sup |ξ|−m−|β| (σ ∈ Cp∞ (Rd \ {0})),

ξ∈supp φ ξ∈A(a,d)

and that maxβ∈Nd ,|β|≤k supξ∈A(a,d) |ξ|−m−|β| < ∞.

0

Exercise 25.C. Let A be an annulus and k ∈ N0 . Show that there exists a C > 1 such
that

sup |ξ|−m−|β| ≤ C |m| (β ∈ Nd0 , |β| ≤ k).

ξ∈A

From this conclude the slightly more general statement than the one of Lemma 25.7
(similar to Theorem 21.33, see also Exercise 25.A): Let A be an annulus in Rd . There
exists a C > 0 such that for all m ∈ R, p ∈ [1, ∞], λ > 0 and all σ ∈ C ∞ (Rd \ {0})

b ⊂ λA =⇒ kσ(D)ukLp ≤ C 1+|m| Mm (σ)λm kukLp

supp u (u ∈ Lp ).

164
 By these inequalities we obtain the following action of Fourier multipliers acting on
elements of Besov spaces whose Fourier transforms are supported away from the origin.
For this we introduce the following notation:

Definition 25.8. Let A ⊂ Rd . We write

0
SA = {u ∈ S 0 : supp u
b ⊂ A}.

Theorem 25.9. Let m ∈ R and ε > 0. Then there exists a C > 0 such that for all
s ∈ R, p, q ∈ [1, ∞], and σ ∈ Cp∞ (Rd \ {0})
s 0
kσ(D)ukBp,q
s+m ≤ CMm (σ)kukB s
p,q
(u ∈ Bp,q ∩ SA(ε,∞) ). (25.8)

In other words, σ(D) forms a continuous operator

s 0 s−m 0
Bp,q ∩ SA(ε,∞) → Bp,q ∩ SA(ε,∞) .

Proof. We can consider a dyadic partition of unity (ϕj )j∈N−1 such that for the correspond-
ing Littlewood–Paley operators (∆j )j∈N−1 one has j∈N0 ∆j u = u for all u ∈ SA(ε,∞)
P 0 .
By Lemma 25.7 (applied with A an annulus that contains the support of ϕ0 ) there
exists a C > 0 such that
0
2jm kσ(D)∆j ukLp ≤ CMm (σ)k∆j ukLp s
(j ∈ N0 , u ∈ Bp,q ∩ SA(ε,∞) ).

As σ(D)∆j u = ∆j σ(D)u, (25.8) follows.

Remark 25.10 (Homogeneous Besov spaces). The statement of Theorem 25.9 is rather
0
ugly as one has to take the intersection with SA(ε,∞) . This requirement is done so that
one only has to deal with the Littlewood–Paley blocks ∆j u for j ∈ N0 , that is, those of
˙ j = ϕ̇j (D) for j ∈ Z,
the form ϕj (D)u, where ϕj = l2−j ϕ. Let us define ϕ̇j = l2−j ϕ and ∆
so that ϕ̇j = ϕj for j ∈ N0 . Analogously to the definition of nonhomogeneous Besov
spaces, one defines homogeneous Besov spaces Ḃp,q s for those tempered distributions u
P∞
for which limJ→∞ j=−J ∆ ˙ j u = u in S 0 . We will not introduce these spaces but want
s .
mention that the inequality in (25.8) holds for all u ∈ Ḃp,q

A typical example of such a Fourier multiplier is the fractional Laplacian. For that,
let us first show that the corresponding Mikhlin norm is finite.

Lemma 25.11. Let l ∈ R and β ∈ Nd0 . Then

Mm (|X|l ) < ∞ ⇐⇒ m = −l, Mm (Xβ |X|l ) < ∞ ⇐⇒ m = −(l + |β|).

Proof. First we observe that for α = 0 we have

|X||α|+m · |∂ α (Xβ |X|l )| = |X|m |X|l+|β| .

165
The supremum over Rd \ {0} over this function is finite if and only if m = −(l + |β|).
Let i ∈ {1, . . . , d}. Then

d d l
|x|l = (x21 + · · · + x2d ) 2 = l|x|l−2 xi (x ∈ Rd ), (25.9)
dxi dxi

and thus (on Rd \ {0})

∂i (Xβ |X|l ) = βi Xβ−ei |X|l + lXβ+ei |X|l−2 . (25.10)

By induction it follows that ∂ α (Xβ |X|l ) is a linear combination of functions of the form
Xγ |X|a with γ ∈ Nd0 , a ∈ R such that a+|γ| = l+|β|−|α|. Therefore |X||α|+m ·|∂ α (Xβ |X|l )|
is a bounded function for all α ∈ Nd0 if m = −(l + |β|).

Corollary 25.12. Let s, t ∈ R, p, q ∈ [1, ∞]. For all ε > 0 the fractional Laplacian
(−∆)s forms a homeomorphism
t 0 t−s 0
Bp,q ∩ SA(ε,∞) → Bp,q ∩ SA(ε,∞) .

Proof. The continuity follows by Theorem 25.9 and Lemma 25.11. That it is a homeo-
morphism follows from the fact that (−∆)−s is the inverse of (−∆)s on SA(ε,∞)
0 .

One of the other main examples that we will consider is the Bessel potential (1−∆)s =
(1 + |X|2 )s (D). We have seen that Mm (|2π X|s ) is finite if and only if m = −s, but
Mm (1 + |X|2 )s is infinite for all s > 0 and m ∈ R (because (1 + |X|2 )s equals 1 at 0).
However, as we will see, M−2s,θ ((1 + |X|2 )s ) is finite for all θ > 0.

Theorem 25.13. Let m ∈ R and σ ∈ Cp∞ , Mm,1 (σ) < ∞. Then there exists a C > 0
such that for all s ∈ R, p, q ∈ [1, ∞],
s
s+m ≤ CkukB s
kσ(D)ukBp,q p,q
(u ∈ Bp,q ). (25.11)

s → B s+m .
In other words, σ(D) forms a continuous operator Bp,r p,r
b ∈ L1 and m > 0, then there exists a C > 0 such that for all s ∈ R,
If additionally, σ
p, q ∈ [1, ∞],
−m s
s+m ≤ C(µ
k(lµ σ)(D)ukBp,q ∨ 1)kukBp,q
s (u ∈ Bp,q , µ > 0). (25.12)

Proof. Besov norms corresponding to different dyadic partitions of unity are equivalent,
we may as well consider a dyadic partition of unity (ϕj )j∈N−1 such that supp ϕ0 ⊂
A(2, ∞). By Lemma 25.21 (applied with A ⊂ A(2, ∞) such that supp ϕ0 ⊂ A) there
exists a C1 > 0 such that

2jm k(lµ σ)(D)∆j ukLp ≤ C1 Mm,2j (lµ σ)k∆j ukLp (µ > 0, σ ∈ Cp∞ , j ∈ N0 , u ∈ S 0 ).
(25.13)

166
Let χ ∈ Cc∞ be such that χϕ−1 = ϕ−1 and thus χ(D)∆−1 = ∆−1 and σ(D)∆−1 =
(σχ)(D)∆−1 . Then by Young’s inequality

kσ(D)∆−1 ukLp ≤ kF −1 (σχ)kL1 k∆−1 ukLp (σ ∈ Cp∞ , u ∈ S 0 ). (25.14)

By (25.14) and (25.13) for µ = 1 we obtain (25.11), because Mm,2j (σ) ≤ Mm,1 (σ).
Let σ ∈ Cp∞ be such that σ
b ∈ L1 and assume m ≥ 0. Then, (see also Theorem 25.1)

k(lµ σ)(D)∆j ukLp ≤ klµ∗ σ

b kL1 k∆j ukLp ≤ kσ
b kL1 k∆j ukLp (µ > 0, j ∈ N−1 , u ∈ S 0 ).
(25.15)

By this we may as well assume m > 0. Then it is sufficient to show that there exists a
C > 0 such that

2jm k(lµ σ)(D)∆j ukLp ≤ Cµ−m k∆j ukLp (u ∈ S 0 , j ∈ N0 , µ > 0). (25.16)

By the observations in Definition 25.5

Mm,2j (lµ σ) = µ−m Mm,µ2j (σ) ≤ µ−m Mm,1 (σ) (µ ≥ 2−j , j ∈ N0 ).

As for µ < 2−j one has 2jm < µ−m , by (25.13) and (25.15), for all µ > 0,
(
µ−m Mm,1 (σ)k∆j ukLp if µ ≥ 2−j ,
2jm k(lµ σ)(D)∆j ukLp ≤ (j ∈ N0 ).
µ−m kσ
b kL1 k∆j ukLp if µ < 2−j

So that (25.16) follows (with C = Mm,1 (σ) ∨ kσ

b kL1 ).

One can formulate the condition Mm,1 (σ) < ∞ for smooth σ differently:
Lemma 25.14. Let m ∈ R, k = 2b1 + d2 c and σ ∈ C ∞ . Then Mm,1 (σ) < ∞ if and only
if there exists a C > 0 such that

|∂ α σ(x)| ≤ C(1 + |x|)−m−|α| (x ∈ Rd , α ∈ Nd0 , |α| ≤ k). (25.17)

Proof. As
1
(1 + |x|) ≤ |x| ≤ 1 + |x| (x ∈ Rd , |x| ≥ 1),
2
by observation (a) of Definition 25.5 we have Mm,1 (σ) < ∞ if and only if there exists a
C > 0 such that

|∂ α σ(x)| ≤ C(1 + |x|)−m−|α| (x ∈ Rd , |x| ≥ 1, α ∈ Nd0 , |α| ≤ k). (25.18)

Let us show that (25.18) is equivalent to (25.17).

As 1 ≤ 1 + |x| ≤ 2 for all x ∈ B(0, 1), there exists an C1 > 0 such that

1 ≤ C1 (1 + |x|)−m−|α| (x ∈ B(0, 1), α ∈ Nd0 , |α| ≤ k).

167
As σ is smooth, its restriction to B(0, 1) is bounded in C k -norm, i.e., C2 = kσkC k (B(0,1)) <
∞. Then

|∂ α σ(x)| ≤ C2 ≤ C1 C2 (1 + |x|)−m−|α| (x ∈ B(0, 1), α ∈ Nd0 , |α| ≤ k).

This shows that (25.18) is equivalent to (25.17).

25.15. The previous lemma can be used to prove the following (Exercise 25.D): For all
σ ∈ Cp∞ there exists a m0 ∈ R such that Mm,1 (σ) < ∞ for all m ∈ R with m ≤ m0 .

Exercise 25.D. Prove the statement in 25.15.

Lemma 25.16. Let a, b, l ∈ R, a > 0, β ∈ Nd0 . Then

Mm,1 (Xβ |X|l (1 + |X|a )b ) < ∞ ⇐⇒ m ≤ −(|β| + l + ab).

In particular,

M−ab,1 ((1 + |x|a )b ) < ∞.

Proof. For x ∈ Rd , |x| ≥ 1 we have |x|a ≤ 1 + |x|a ≤ 2|x|a and thus

sup |x|m · xβ |x|l (1 + |x|a )b < ∞ ⇐⇒ m ≤ −(|β| + l + ab).

x∈Rd
|x|≥1

Let i ∈ {1, . . . , d}. Then, see (25.9),

d
(1 + |x|a )b ) = b(1 + |x|a )b−1 · a|x|a−2 xi (x ∈ Rd ).
dxi

Therefore, by (25.10) (on Rd \ {0})

∂i Xβ |X|l (1 + |X|a )b = βi Xβ−ei |X|l (1 + |X|a )b + lXβ+ei |X|l−2 (1 + |X|a )b

+ aXβ+ei |X|l+a−2 b(1 + |x|a )b−1 .

By induction it follows that ∂ α (Xβ |X|l (1 + |X|a )b ) is a linear combination of functions of

the form (Xγ |X|c (1+|X|a )d ) with γ ∈ Nd0 , c, d ∈ R such that |γ|+c+ad = |β|+l +ab−|α|.
Therefore |X||α|−m |∂ α (Xβ |X|l (1 + |X|a )b )| is bounded for all α ∈ Nd0 if m ≤ −(|β| + l +
ab).

Corollary 25.17. Let s, t ∈ R, p, q ∈ [1, ∞]. The Bessel potential (1 − ∆)s forms a
homeomorphism
t t−2s
Bp,q → Bp,q .

Proof. This is a consequence of Theorem 25.13 and Lemma 25.16.

168
 We summarize the different inequalities that we have obtained, but without the intro-
duced notations Mm,θ and A(a, ∞) (however, incorporating the scaling properties that
one obtains via for example (25.5)).
Theorem 25.18 (Hörmander–Mikhlin inequalities). Let m ∈ R.

(a) (Lemma 25.7) Let A be an annulus. Let θ ≥ 0. Let σ ∈ C ∞ (Rd \ {0}) and suppose
that there exists a M > 0 such that

|∂ α σ(x)| ≤ M |x|−m−|α| (x ∈ Rd , |x| > θ, α ∈ Nd0 , |α| ≤ k). (25.19)

For all λ0 > 0 there exists a C > 0 such that for all p ∈ [1, ∞] and λ > λ0

b ⊂ λA =⇒ k(lµ σ)(D)ukLp ≤ Cµ−m λ−m kukLp

supp u (u ∈ Lp ).

(b) (Theorem 25.9) Let m ∈ R and ε > 0. Then there exists a C > 0 such that for all
s ∈ R, p, q ∈ [1, ∞], and σ ∈ Cp∞ (Rd \ {0}) such that (25.19) holds for θ = 0,
−m s
k(lµ σ)(D)ukBp,q
s+m ≤ Cµ kukBp,q
s (µ > 0, u ∈ Bp,q b ∩ B(0, ε) = ∅).
, supp u

(c) (Theorem 25.13) Let σ ∈ Cp∞ be such that (25.19) holds for some θ > 0 or equival-
ently, such that there exists an M > 0 such that

|∂ α σ(x)| ≤ C(1 + |x|)−m−|α| (x ∈ Rd , α ∈ Nd0 , |α| ≤ k).

Then there exists a C > 0 such that for all s ∈ R, p, q ∈ [1, ∞],
s
s+m ≤ CkukB s
kσ(D)ukBp,q p,q
(u ∈ Bp,q ).

(d) (Theorem 25.1) Let σ ∈ Cp∞ be such that σ

b ∈ L1 . There exists a C > 0 such that
for all s ∈ R, p, q ∈ [1, ∞],
s
kσ(D)ukBp,q
s ≤ CkukBp,q
s (u ∈ Bp,q ).

(e) (Theorem 25.13) Let m ≥ 0 and σ ∈ Cp∞ be such that σ

b ∈ L1 . Then there exists a
C > 0 such that for all s ∈ R, p, q ∈ [1, ∞],
−m s
s+m ≤ C(µ
k(lµ σ)(D)ukBp,q ∨ 1)kukBp,q
s (u ∈ Bp,q , µ > 0).

Let us apply Theorem 25.18 (e) to the heat semigroup. We consider this semigroup
later on again to find solutions to the heat equation.
Definition 25.19. For t > 0 let ht : Rd → R be the Schwartz function (as in Ex-
ample 11.15) given by
d 1 2
ht (x) = (4πt)− 2 e− 4t |x| (x ∈ Rd ).

169
We define Ht : S 0 → S 0 by

Ht u = ht ∗ u (u ∈ S 0 ),

and H0 : S 0 → S 0 to be the identity map, H0 u = u. The collection (Ht )t∈[0,∞) is called

the Heat semigroup, in the sense that
Exercise 25.E. Show that (Ht )t∈[0,∞) is a semigroup, i.e.,

Ht Hs u = Ht ◦ Hs u = Ht+s u (s, t ∈ [0, ∞), u ∈ S 0 ).

As the convolution of a tempered distribution with a Schwartz function is a Cp∞

function, Ht u is smooth for all t > 0. By the following theorem it follows that if u is in
Lp , then Ht u is in the Sobolev space W k,p for all k ∈ N0 . Moreover, if u is in a Besov
s for some s ∈ R, then u is in B t for all t ∈ R.
space Bp,q p,q

Theorem 25.20. Let s ∈ R, m ≥ 0, k ∈ N0 , p, q ∈ [1, ∞].

(a) There exists a C > 0 such that

k
kHt f kW k,p ≤ C(t− 2 ∨ 1)kf kLp (f ∈ Lp , t > 0).

(b) There exists a C > 0 such that

m
−2 s
kHt ukBp,q
s+m ≤ C(t ∨ 1)kukBp,q
s (u ∈ Bp,q , t > 0).

Exercise 25.F. Prove Theorem 25.20. (Hint for (a): Show that ∂ α ht ∈ L1 for all
2
α ∈ N0 by showing that y 7→ |y|n e−|y| is integrable for all n ∈ N0 ; either by showing
2
that (1 + |y|2 )n e−|y| is a Schwartz function or one can use that the gamma function Γ is
finite everywhere (Definition 11.22). Hint for (b)): Example 25.6.)

25.1 Comments ...

In this section we have considered Fourier multipliers of smooth functions on Rd \{0}. The
Mikhlin norm Mm,θ is however defined for C k functions on A(θ, ∞). One can actually
also consider Fourier multipliers corresponding to such C k (A(θ, ∞)) functions. Let us
comment on “taking k ∈ N0 instead of k = ∞” and on “θ ≥ 0 instead of θ = 0” separately.
First of all, instead of taking smooth functions one may take C k functions basically
because of Exercise 25.B: For v ∈ S 0 for which there exists a C > 0 such that |hv, ϕi| ≤
Ckϕkk,S One can show that if σ ∈ Cpk = m∈N0 Cp,m k
S
(see Exercise 14.E), that is, there
exists an m ∈ N0 such that qk,m (σ) < ∞, then

ϕ 7→ hv, σϕi

defines a tempered distribution, which we call σv. Then, by Exercise 25.B for any u ∈ Lp
we have σ ub ∈ S 0 and so we may define σ(D)u to be the tempered distribution F −1 (σ u
b).

170
Similar as in Definition 19.7 one may extend this definition in case σ is only Cpk on a
neighbourhood of the support of u. The arguments in this section do not depend on the
smoothness of σ, well, only in terms of the Mikhlin norm, which only requires σ to be
Ck.
Consider the situation as in Lemma 25.7 but with σ ∈ C ∞ (A(θ, ∞)) for some θ > 0.
Then σ(D)u may not be defined for all λ > 0 (and u ∈ Lp with supp u b ⊂ λA), as λA
needs to be a subset of A(θ, ∞). We have λA ⊂ A(θ, ∞) for λ > aθ if a > 0 is such that
A ⊂ A(a, ∞). The following lemma is an extension of Lemma 25.7 which entails this in
its premise.
Lemma 25.21. Let m ∈ R. Let A be an annulus in Rd . Let θ ≥ 0 and a > 0 be such
that A ⊂ A(a, ∞). There exists a C > 0 such that for all p ∈ [1, ∞], λ > aθ and all
σ ∈ C ∞ (A(θ, ∞))

b ⊂ λA =⇒ kσ(D)ukLp ≤ CMm,λa (σ)λ−m kukLp

supp u (u ∈ Lp ). (25.20)

Proof. For θ = 0 this is the “Moverover” statement of Lemma 25.7. Let θ > 0. Let
r > 1 be such that A ⊂ A(ra, ∞). Let χ ∈ Cb∞ (Rd ) be equal to 1 on A(rθ, ∞) and
supp χ ⊂ A(θ, ∞). Then σ(D)u = (σχ)(D)u for σ ∈ C ∞ (A(θ, ∞)) and u ∈ S 0 with
b ⊂ λA for λ > aθ . As there exists a C > 0
b ⊂ A(rθ, ∞), which is the case if supp u
supp u
such that Mm,λa (χσ) = Mm,λa (σ) for λ > aθ we obtain (25.20) by (25.7).

Remark 25.22. We called the inequalities in Theorem 25.18 Hörmander–Mikhlin in-

equalities as they are closely related to what in literature are called the Hörmander–
Mikhlin multiplier theorems, see [Sha66] for example, or [Mic57] (in Russian) or [H6̈0]
for the work of Mikhlin and Hörmander. As unfortunately happens with names from
languages with different alphabets, we also found instead of Mikhlin the names Michlin
or Mihlin.
The theorems of Hörmander and Mikhlin deal with the case m = 0. See for example
also [HvNVW16, Theorem 5.5.10] (which looks again a bit different). We decided to call
the norm the Mikhlin norm as that seems to align with the literature and it seems that
the Hörmander and Mikhlin statements are slightly different.

26 Products of tempered distributions

For functions f, g : Rd → F their product f g (or f · g) is the function Rd → F defined

by f g(x) = f (x) · g(x) for x ∈ Rd . We defined the product of a smooth function with
a distribution in Definition 2.14 and the product of a Cp∞ function with a tempered
distribution in Definition 15.3 in such a way that they extend the product of functions:
if f ∈ L1loc and ψ ∈ E, then the product of the testfunction with the corresponding
distribution ψuf equals the distribution uψf that corresponds to the product of the
functions f and ψ (similarly this equality holds in S 0 if f ∈ L1loc is such that uf ∈ S 0 and
ψ ∈ Cp∞ ).

171
 In this section we investigate for which distributions one can make sense of their
“product”. It would make sense to call an operation × : S 0 × S 0 → S 0 that extends the
product Cp∞ × S 0 → S 0 , (σ, u) 7→ σu a “product” if it is continuous in both variables,
commutative and associative.
However, such an operation does not exist and so we may only make sense of a kind
of “product” for certain pairs of (tempered) distributions.
First of all, let us observe that there does not exist an operation × : S 0 × S 0 → S 0
that extends the product map

Cp∞ × S 0 → S 0 , (σ, u) 7→ σu, (26.1)

and is continuous in both variables, as then, δ0 × δ0 by some approximation argument

would be the tempered distribution that is equal to the Fourier transform of 1 ∗ 1 (by
the continuity of the product and the Fourier transform), which is not defined (or, if one
wants, is infinity everywhere). On the other hand, let u be defined as in Exercise 2.E,
see also Exercise 15.B:
ϕ(x)
Z
u(ϕ) := lim dx (ϕ ∈ S(R)).
ε↓0 R\[−ε,ε] x
Then Xu = 1, 1δ0 = δ0 and Xδ0 = 0, and 0u = 0. By means of these identities let us
show that there does not exist an operation × : S 0 × S 0 → S 0 that extends the product
map (26.1) and is commutative and associative, that is,

ψ × u = ψu (ψ ∈ Cp∞ , u ∈ S 0 ),
u × v = v × u, (u × v) × w = u × (u × w) (u, v, w ∈ S 0 ).

If × were such an operation, then we would have

X × (u × δ0 ) = (X × u) × δ0 = (Xu) × δ0 = 1 × δ0 = 1δ0 = δ0 ,
X × (u × δ0 ) = u × (X × δ0 ) = u × (Xδ0 ) = u × 0 = 0u = 0.

We use the decomposition of tempered distribution in terms of Littlewood–Paley

blocks as a starting point to define a product on a class of pairs of distributions. Let ∆j
for j ∈ N−1 be the Littlewood–Paley operators as in Definition 21.12 (for a given dyadic
partition of unity). For u, v ∈ S 0 , by Lemma 21.9 we have
X X
u= ∆i u, v= ∆i v.
i∈N−1 i∈N−1

Let us write “ · ” also for the product between functions, so that ∆i u · ∆j v is (∆i u)∆j v
(we do not want to write ∆i u∆j v as this can be read as ∆i (u∆j v)). For those tempered
u and v such that
X
∆i u · ∆j v (26.2)
i,j∈N−1

172
exists in S 0 (for the summation notation see 21.7), one could call the tempered distribu-
tion (26.2) a “product” of u and v. A priori it is not clear whether this product agrees
with f g, or better said, uf g using the notation as in 2.6, if u and v are represented by
locally integrable functions f and g, respectively; i.e., u = uf and v = ug .

In this section we introduce a certain product on tempered distributions, and call

it the Bony product. In Section 27 we consider this Bony product between elements of
Besov spaces, here we it between Cp∞ functions and Schwartz functions (Theorem 26.5)
and between Cp∞ functions and tempered distributions (Theorem 26.7) and show that
in these cases the product agrees with the pointwise product and with the product as
defined in Definition 15.3. Moreover, we show that the Bony product between a function
in Lp and a function in Lq for p and q such that p1 + 1q = 1 is equal to the product of the
two functions, which is an L1 function by Hölders inequality. Finally, in Example 26.11
we given examples of locally integrable functions which represent tempered distributions
for which their Bony product exists but is not equal to their product as functions (the
pointwise product).
Definition 26.1. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 and ∆j = ϕj (D)
for j ∈ N−1 . Let u, v ∈ S 0 . If
J
X
∆i u · ∆j v
i,j=−1

converges in S 0 as J → ∞, then we say that the ϕ-Bony product of u and v exists and
write u •ϕ v or v •ϕ u for the limit and call it the ϕ-Bony product of u and v, i.e.,
J
X
u •ϕ v = lim ∆i u · ∆j v.
J→∞
i,j=−1

If for all ψ that generate dyadic partitions of unity the ψ-Bony product of u and v exists
and u •ϕ v = u •ψ v, then we say that the Bony product of u and v exists and call it the
Bony product of u and v and write u • v instead of u •ϕ v.

The Bony product can be viewed as a bilinear operation in the following sense:
Lemma 26.2. Let ϕ generate a dyadic partition of unity. Let u, v, w ∈ S 0 and λ ∈ F.

(a) If u •ϕ v and w •ϕ v exist, then (u + w) •ϕ v exists and

(u + w) •ϕ v = u •ϕ v + w •ϕ v.

(b) If u •ϕ v and u •ϕ w exist, then u •ϕ (v + w) exists and

u •ϕ (v + w) = u •ϕ v + u •ϕ w.

(c) If u •ϕ v exist, then (λu) •ϕ v and u •ϕ (λv) exist and

λ(u •ϕ v) = (λu) •ϕ v + u •ϕ (λv).

Proof. The proof is straightforward and left to the reader.

173
(                          )

We have already seen that ∞ j=−1 ∆j ϕ = ϕ in S for ϕ ∈ S and that

∞
P P
j=−1 ∆j u = u in
S for u ∈ S , see Lemma 21.9. We consider some similar convergence for Cp∞ , that we
0 0

will use for the Bony products of such functions with Schwartz functions and tempered
distributions. We state this convergence in Lemma 26.4 as a consequence of the following
convergence of mollifiers of Cp∞ functions.
Lemma 26.3. Let ψ be a Schwartz function with ψ = 1. Let ψε = lε∗ ψ for ε > 0. For
R

all η ∈ Cp∞ and m ∈ N0 there exists a k ∈ N0 such that

ε↓0
qm,k (ψε ∗ η − η) −−→ 0.

In particular, if n ∈ N0 is such that qm,n (η) < ∞, then it suffices to choose k = n + 1.

Proof. As ∂ α (ψε ∗ η) = ψε ∗ (∂ α η) and ∂ α η ∈ Cp∞ for all α ∈ Nd0 and η ∈ Cp∞ , it is

sufficient to prove that for all η ∈ Cp∞ there exists a k ∈ N0 such that
ε↓0
k(ψε ∗ η − η)(1 + |X|2 )−k kL∞ −−→ 0. (26.3)

Let n ∈ Nd0 be such that η(1 + |X|2 )−n is bounded. Let M = kη(1 + |X|2 )−n kL∞ and
N = kψ(1 + |X|2 )n kL1 ∨ 1. As

(1 + |x|2 )n ≤ 2n (1 + |x − y|2 )n (1 + |y|2 )n (x, y ∈ Rd ),

for ε ∈ (0, 1) we have

Z
2 −n
k(ψε ∗ η)(1 + |X| ) kL∞ = sup ψε (x − y)η(y)(1 + |x|2 )−n dy
x∈Rd Rd
Z
≤ sup ψε (x − y)(1 + |x − y|2 )n η(y)(1 + |y|2 )−n dy
x∈Rd Rd

= k(ψε (1 + |X|2 )n ) ∗ (η(1 + |X|2 )−n )kL∞

≤ kψε (1 + |X|2 )n kL1 kη(1 + |X|2 )−n kL∞
≤ M klε∗ (ψ(1 + |X|2 )n )kL1 ≤ M N,

where we used that (1 + |X|2 )n ≤ (1 + | Xε |2 )n for all ε ∈ (0, 1).

Let δ, ε ∈ (0, 1). Let R > 0 be such that 2M N (1 + R2 )−1 < ε. Let k = n + 1. Then

k(ψε ∗ η − η)(1 + |X|2 )−k kL∞

≤ sup |(ψε ∗ η(x) − η(x))(1 + |x|2 )−k |
x∈B(0,R)
 
+ (1 + R2 )−1 k(ψε ∗ η)(1 + |X|2 )−n kL∞ + kη(1 + |X|2 )−n kL∞
≤ sup |ψε ∗ η(x) − η(x)| + δ.
x∈B(0,R)

174
By Theorem 7.15 (b)
ε↓0
sup |ψε ∗ η(x) − η(x)| −−→ 0,
x∈B(0,R)

so that we conclude (26.3).

Lemma 26.4. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 . Let ∆j = ϕj (D) for
j ∈ N−1 . Let η ∈ Cp∞ . For all m ∈ N0 there exists a k ∈ N0 such that

J
 X 
J→∞
qm,k ∆j η − η −−−→ 0.
j=−1

Proof. For ψ R= F −1R(ϕ−1 ) we have Jj=−1 ∆j η = F −1 ( Jj=−1 ϕj ) ∗ η = l2∗−J+1 ψ ∗ η (see

P P

(21.10)) and ψ = ϕb−1 = ϕ−1 (0) = 1. Therefore this follows by Lemma 26.3.

Theorem 26.5. For η ∈ Cp∞ and ψ ∈ S, η • ψ exists in S and

ηψ = η • ψ (η ∈ Cp∞ , ψ ∈ S).

Proof. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 . Let ∆j = ϕj (D) for j ∈ N−1 .
It is sufficient to show Ji,j=−1 ∆i η · ∆j ψ → ηψ in S as J → ∞. This follows as
P

J
X J
 X J
 X J
 X 
∆i η · ∆j ψ − ηψ = ∆i η − η ∆i ψ + η · ∆j ψ − ψ . (26.4)
i,j=−1 i=−1 i=−1 j=−1

The second term converges to zero in S by Lemma 21.9. For the first term we use
Lemma 14.7. Let m ∈ N0 and let C > 0 be as in (14.5). Then for all k ∈ N0
 
J
X J
X I
X J
 X 
 ∆i η − η  ∆i ψ ≤ C sup ∆i ψ qm,k ∆j η − η .
i=−1 i=−1 I∈N−1 i=−1 j=−1
m,S m+k,S

PJ PI
As j=−1 ∆j ψ → ψ in S, the supremum supI∈N−1 i=−1 ∆i ψ m+k,S is finite for all
k ∈ N0 . By Lemma 26.4 there exists a k ∈ N0 such that
J
 X 
J→∞
qm,k ∆j η − η −−−→ 0,
j=−1

from which we conclude that the first term on the right-hand side of (26.4) also converges
to zero in S.

The following auxiliary lemma will be used for Theorem 26.7, in which we consider
the product of Cp∞ functions with tempered distributions.

175
Lemma 26.6. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 . Let ∆j = ϕj (D) for
j ∈ N−1 . For all m ∈ N0 there exists a k ∈ N0 and a C > 0 such that
X
∆j ψ ≤ Ckψkk,S (F ⊂ N−1 , F is a finite set, ψ ∈ S).
m,S
j∈F

Proof. By Lemma 14.7, by Lemma 21.8 and the continuity of the Fourier transformation
in S there exist C1 , C2 , C3 > 0 and k, l ∈ N0 such that for all finite subsets F ⊂ N−1 and
ψ∈S
X X X
∆j ψ ≤ C1 ϕj ψb ≤ C1 C2 ϕj kψk
b
k,S
m,S k,S Ck
j∈F j∈F j∈F

≤ C1 C2 C3 kψkl,S .

Theorem 26.7. For η ∈ Cp∞ and u ∈ S 0 , η • u exists in S 0 and

ηu = η • u in S 0 .

Proof. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 . Let ∆j = ϕj (D) for j ∈ N−1 .
We have to show that
J
* +
X J→∞
∆i η · ∆j u, ψ −−−→ hηu, ψi (ψ ∈ S). (26.5)
i,j=−1

Observe that hηu, ψi = hu, ηψi and

J J J
* + * +
X X  X 
∆i η · ∆j u, ψ = u, ∆j ∆i η · ψ (J ∈ N−1 ).
i,j=−1 j=−1 i=−1
PJ PJ
Hence it suffices to show j=−1 ∆j ( i=−1 ∆i η · ψ) → ηψ in S. Observe that
J
X J
 X  J
X J
X J
X
∆j ∆i η · ψ − ηψ = ∆j ( ∆i η · ψ) − ∆i η · ψ
j=−1 i=−1 j=−1 i=−1 i=−1
J
X
+ ∆i η · ψ − ηψ. (26.6)
i=−1
PJ
By Lemma 26.4 and Lemma 14.7 it follows that i=−1 ∆i η · ψ → η · ψ in S, and by those
lemmas and Lemma 26.6 it follows that
J
X J
 X  
∆j ∆i η − η · ψ → 0.
j=−1 i=−1

176
 Now we turn to the product of Lp functions with Lq functions, in case p1 + 1q = 1. As
before, we first consider the convergence of the partial sums Jj=−1 ∆j f as J → ∞ in Lp
P

for f ∈ Lp . For this we adapt Theorem 7.15 (d):

Lemma 26.8. Let ψ ∈ C(Rd ) ∩ L1 (Rd ). Let p ∈ [1, ∞). Write ψε = lε∗ ψ for ε > 0.
Then, for any f ∈ Lp (Rd )
Z
ψε ∗ f → ( ψ)f. (26.7)

Proof. Let χ ∈ Cc (Rd ) be equal to 1 on a neighbourhood of zero. Write χR = l 1 χ for

R
R > 0. By Theorem 7.15 (d) we have
Z
ε↓0
lε∗ (ψχR ) ∗ f − ( ψχR )f −−→ 0.
Lp

By Young’s inequality we have

klε∗ (ψ(1 − χR )) ∗ f kLp ≤ kψ(1 − χR )kL1 kf kLp (ε > 0).

As both kψ(1 − χR )kL1 −−−−→ 0 and

R→∞ R R→∞ R
ψχR −−−−→ ψ, and
Z
ψε ∗ f − ( ψ)f ≤ klε∗ (ψ(1 − χR )) ∗ f kLp
Lp
Z Z Z
+ lε∗ (ψχR ) ∗ f − ( ψχR )f + ( ψχR )f − ( ψ)f
Lp Lp
≤ kψ(1 − χR )kL1 kf kLp
Z Z Z
+ lε∗ (ψχR ) ∗ f − ( ψχR )f + ( ψχR ) − ( ψ) kf kLp ,
Lp

we conclude (26.7).

Lemma 26.9. Let p ∈ [1, ∞). Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 . Let
∆j = ϕj (D) for j ∈ N−1 . Then ∞ p
P
j=−1 ∆j f = f in L .

Proof. This follows from Lemma 26.8 similarly as in the proof of Lemma 26.4.

Theorem 26.10. Let p, q ∈ [1, ∞] be such that

1 1
+ = 1.
p q

Let f ∈ Lp and g ∈ Lq . Then f g ∈ L1 and f • g exists in L1 (and thus in S 0 ) and

f g = f • g in L1 .

Or, more precisely, uf • ug exists in S 0 and uf g = uf • ug .

177
Proof. Without loss of generality we may assume q < ∞ (otherwise one has p < ∞ and
we can interchange roles). We have
J
X J
X
∆i f ∆j g − f g
i=−1 j=−1 L1
J
X J
X J
X
≤ ∆i f ∆j g − g + ∆i f − f kgkLq ,
i=−1 Lp j=−1 Lq i=−1 Lp

which converges to 0 as J → ∞ by Lemma 26.8 and by Lemma 21.9 (in particular

(21.13)).

Example 26.11. Consider d = 1. Let f, g : R → R be given by

 
−1
 √1 x > 0, −1
0 x ≥ 0,
f (x) = x+ 2 = x g(x) = x− 2 = (x ∈ R).
0 x ≤ 0,  √1 x < 0,
−x

Then both f ang g are locally integrable. f g = 0 everywhere but, as we will show, f • g
exists in S 0 and
π
f • g = δ0 . (26.8)
2
Indeed, let ψ ∈ S, ψ = ψ(−X) and ψ = 1 and write ψε = lε∗ ψ. It suffices to show (see
R

for example the proof of Lemma 26.4)

ε↓0 π
(ψε ∗ f ) · (ψε ∗ g) −−→ δ0 in S 0
2
For h ∈ L1loc one has
y x
Z Z
ψε ∗ h = ε−d ψ( )h(x − y) dy = ψ(z)h(x − εz) dz = ψ ∗ lε h( ) (x ∈ Rd ).
R ε R ε
Therefore, for ϕ ∈ S (the convergence follows by Lebesgue’s dominated convergence
theorem, as ϕ is bounded)
Z
h(ψε ∗ f ) · (ψε ∗ g), ϕi = (ψε ∗ f )(x)(ψε ∗ g)(x)ϕ(x) dx
Z Z R
 Z 
= ψ(y)f (x − εy) dy ψ(y)g(x − εz) dz ϕ(x) dx
R
Z R R
Z  Z 
=ε ψ(y)f (εw − εy) dy ψ(y)g(εw − εz) dz ϕ(εw) dw
Z R Z wR  ∞
R
1 1
Z

= ψ(y) √ dy ψ(z) √ dz ϕ(εw) dw
R −∞ w−y w z−w
Z Z w Z ∞
ε↓0 1  1 
−−→ ϕ(0) ψ(y) √ dy ψ(z) √ dz dw
R −∞ w−y w z−w
Z ∞ Z z
1 1
Z
= ϕ(0) ψ(y) ψ(z) √ √ dw dz dy.
R y y w − y z −w

178
Now, for all y, z ∈ R, z > y,
Z z Z z−y
1 1 1 1
√ √ dw = √ √ dw
y w−y z−w 0 w z−y−w
Z 1
1 1
= √ √ dw,
0 w 1−w
which equals π; the latter equals the Beta function evaluated in ( 12 , 12 ), which equals
Γ( 12 )2
Γ(1) = π (see for example [AAR99, Section 1.1], and see Definition 11.22 for Γ). By
using that ψ is symmetric, i.e., ψ = Rψ, we have
Z Z ∞ Z Z y
ψ(y) ψ(z) dz dy = ψ(y) ψ(z) dz dy
R y R −∞
R R 1
and as Rψ Rψ = 1, the above integrals equal 2 and thus
ε↓0 π
h(ψε ∗ f ) · (ψε ∗ g), ϕi −−→ ϕ(0) (ϕ ∈ S),
2
i.e.,
π
(ψε ∗ f ) · (ψε ∗ g) → δ0 .
2
We conclude (26.8).

26.1 Comments...

The statements in this section are similar to those of Johnsen [Joh95], though different
tools are used for the proofs. Moreover, the statements differ slightly in the sense that in
this section we have considered the Bony product in S 0 , whereas [Joh95] only considers
it in D0 . [Joh95, Theorem 3.8] considers a more general statement than Theorem 26.10:
Instead, f ∈ Lploc , g ∈ Lqloc and p1 + 1q ≤ 1. It is shown that f g = f • g in D0 , and moreover
that the product is in Lrloc , with r ∈ [1, ∞] being such that 1r = p1 + 1q .
For a more comprehensive reference on products of distributions we refer to the book
by Oberguggenberger [Obe92].
Example 26.11 is [Obe92, Example 2.3] (and is also mentioned in [Joh95, Example
3.2].

27 Paraproducts and resonance products in Besov spaces

In the previous section we defined the Bony product of two tempered distributions u and
v to be the limit (if it exists) of Ji,j=−1 ∆i u · ∆j v as J → ∞. In this section we consider
P

the Bony product between elements of Besov spaces by considering limits of three parts
of Ji,j=−1 ∆i u · ∆j v, namely one that considers the sum over a left-upper triangle of
P

{−1, 0, 1, . . . , J}2 , one over a right-lower triangle and one over a thickened diagonal:

179
Definition 27.1. Let ϕ generate a dyadic partition of unity (ϕj )j∈N−1 and ∆j = ϕj (D)
for j ∈ N−1 . For j ∈ N−1 we define
j
(j ∈ N−1 , u ∈ S 0 ).
X
∆−3 u := 0, ∆−2 u := 0, ∆≤j u := ∆i u
i=−1

Let u, v ∈ S 0 .

(a) If
J
X J j−2
X X
∆≤j−2 u · ∆j v = ∆i u · ∆j v
j=−1 j=1 i=−1

converges in S 0 as J → ∞, then we say that the ϕ-paraproduct of u with v exists

and write u 4ϕ v or v 5ϕ u for the limit and call it the ϕ-paraproduct of u with v,
i.e.,
∞
X ∞ j−2
X X
u 4ϕ v = v 5ϕ u = ∆≤j−2 u · ∆j v = ∆i u · ∆j v.
j=−1 j=1 i=−1

If for all ψ that generate dyadic partitions of unity the ψ-paraproduct of u and v
exists and u 4ϕ v = u 4ψ v, then we say that the paraproduct of u with v exists and
write u 4 v instead of u 4ϕ v and call u 4 v the paraproduct product of u with v.
(b) If
J
X
∆j−1 u · ∆j v + ∆j u · ∆j v + ∆j u · ∆j−1 v
j=−1

converges in S 0 as J → ∞, then we say that the ϕ-resonance product of u and v

exists and write u ϕ v for the limit and call it the ϕ-resonance product of u and
v, i.e.,
∞
X
u ϕ v = ∆j−1 u · ∆j v + ∆j u · ∆j v + ∆j u · ∆j−1 v.
j=−1

If for all ψ that generate dyadic partitions of unity the ψ-resonance product of u
and v exists and u ϕ v = u ψ v, then we say that the resonance product of u and
v exists and write u  v instead of u ϕ v and call u  v the resonance product of
u with v.

Observe that u ϕ v = v ϕ u.

• If both u 4ϕ v and u ϕ v exist, then we write u < ϕ v = u 4ϕ v + u ϕ v.

• If both u 5ϕ v and u ϕ v exist, then we write u = ϕ v = u 5ϕ v + u ϕ v.

180
 • If both u 4ϕ v and u 5ϕ v exist, then we write u <
= ϕ v = u 4ϕ v + u 5ϕ v.

We refer the reader to Remark 27.18 about the different notations for paraproducts and
resonance products in the literature.

Let us make the following observation, which one could interpret as bilinearity of the
different products:

Lemma 27.2. Let ϕ generate a dyadic partition of unity. Let u, v, w ∈ S 0 and λ ∈ F.

(a) If u 4ϕ v and w 4ϕ v exist, then (u + w) 4ϕ v exists and

(u + w) 4ϕ v = u 4ϕ v + w 4ϕ v.

(b) If u 4ϕ v and u 4ϕ w exist, then u 4ϕ (v + w) exists and

u 4ϕ (v + w) = u 4ϕ v + u 4ϕ w.

(c) If u 4ϕ v exist, then (λu) 4ϕ v and u 4ϕ (λv) exist and

λ(u 4ϕ v) = (λu) 4ϕ v + u 4ϕ (λv).

The statements (a), (b) and (c) are also valid if we replace each occurrence of “4ϕ ” by
“ϕ ”, or each occurrence of “4ϕ ” by “5ϕ ”.

Proof. The proof is straightforward and left to the reader.

The existence of each of the paraproducts and the resonance product imply the ex-
istence of the Bony product:

Lemma 27.3. Let ϕ generate a dyadic partition of unity. Let u, v ∈ S 0 . If u 4ϕ v, u ϕ v

and u 5ϕ v exist, then u •ϕ v exists and

u •ϕ v = u 4ϕ v + u ϕ v + u 5ϕ v.

Proof. See Exercise 27.A.

Exercise 27.A. Prove Lemma 27.3.

In Theorem 27.5 we consider estimates on paraproducts in Besov spaces. These

rely on the Hölder inequalities and the following theorem, which essentially says: For a
negative regularity index s the Besov norm is equivalent to the norm-like function which
looks like the Besov norm but with “∆≤j ” instead of “∆j ”.

181
Theorem 27.4. Let s < 0 and p, q ∈ [1, ∞]. Then we have for u ∈ S 0
s
u ∈ Bp,q ⇐⇒ k(2js k∆≤j ukLp )j∈N−1 k`q < ∞.

Moreover,

(1 + 2s )−1 kukBp,q
s ≤ k(2js k∆≤j ukLp )j∈N−1 k`q ≤ (1 − 2s )−1 kukBp,q
s (u ∈ S 0 ). (27.1)

Proof. It is sufficient to prove (27.1). For the inequality on the left–hand side of (27.1):

2js k∆j ukLp ≤ 2js k∆≤j ukLp + 2s 2(j−1)s k∆≤j−1 ukLp .

Therefore

kukBp,q
s ≤ (1 + 2s )k(2js k∆≤j ukLp )j∈N−1 k`q .

For the inequality on the right-hand side of (27.1):

j
X j
X
js js
2 k∆≤j ukLp ≤ 2 k∆i ukLp = 2(j−i)s 2is k∆i ukLp = (a ∗ b)(j),
i=−1 i=−1

where a, b : Z → R are given for j ∈ Z by

( (
2js j ∈ N0 , 2js k∆j ukLp j ∈ N−1 ,
a(j) = b(j) =
0 j ≤ −1, 0 j ≤ −2.

Hence, by Young’s inequality Theorem 21.17

k(2js k∆≤j ukLp )j∈N−1 k`q = ka ∗ bk`q ≤ kak`1 kbk`q = kak`1 kukBp,q
s .

As s < 0 we have kak`1 = 2js = (1 − 2s )−1 .

P
j∈N0

Theorem 27.5. Let ϕ generate a dyadic partition of unity. Let p, p1 , p2 , q, q1 , q2 , r ∈

[1, ∞] be such that p11 + p12 ≤ 1 and

1 1 1 1
p = p1 + p2 , q = min{1, q11 + 1
q2 }, (27.2)

(a) For all s ∈ R, u ∈ Lp1 , and v ∈ Bps2 ,r the paraproduct of u with v exists in Bp,r
s .

Moreover, there exists a C > 0 such that

ku 4ϕ vkBp,r
s ≤ CkukLp1 kvkBps (u ∈ Lp1 , v ∈ Bps2 ,r ). (27.3)
2 ,r

(b) For all s < 0, t ∈ R, u ∈ Bps1 ,q1 and v ∈ Bpt 2 ,q2 the paraproduct of u with v exists
s+t and
in Bp,q

ku 4ϕ vkBp,q
s+t ≤ CkukB s
p
kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ). (27.4)
1 ,q1 2 ,q2

182
(If the Besov norms in (27.3) and (27.4) are with respect to ϕ, as in Definition 21.12,
2s s −1 in
(21.14), then we may take C = 3kϕd −1 kL1 in (27.3) and C = 2kϕ d−1 kL1 2 (1 − 2 )
(27.4).)

Proof. Let u, v ∈ S 0 . For j ∈ N−1 let wj = ∆≤j−2 u · ∆j v = j−2

i=−1 ∆i u · ∆j v. For both
P

(a) and (b) we apply Theorem 21.18 (a). Let us first show that there exists a ball B and
an annulus A such that
b−1 ⊂ B,
supp w bj ⊂ 2j A
supp w (j ∈ N0 ).
Observe that ∆≤−3 u · ∆−1 v = 0 and ∆≤−2 u · ∆0 v = 0, and so their Fourier transform
is trivially supported in any ball and annulus. Let us check that for j ∈ N the Fourier
transform of ∆≤j−2 u · ∆j v is supported in 2j A for some annulus A.
Let a, b > 0, a < b be such that supp ϕ ⊂ A(a, b). By the disjointness property of
dyadic partitions of unity, (21.4), we may assume that a and b are such that 4A(a, b) ∩
A(a, b) = ∅, i.e., 4a > b. Moreover, supp ϕ−1 ⊂ B(0, 2b ) = 2−1 B(0, b) (as for example
ϕ−1 + ϕ0 = l 1 ϕ−1 and supp ϕ−1 + ϕ0 ⊂ B(0, b)). Let C, B and A be given by
2

C = A(a, b), B = B(0, b), A = A(a − 14 b, 54 b) = 14 B + C.

As F(∆i u · ∆j v) = (ϕi u
b) ∗ (ϕj vb), we have (see Theorem 7.10 and Lemma 7.11):

supp F(∆i u · ∆j v) ⊂ 2j (2i−j B + C) ⊂ 2j ( 41 B + C) (i ∈ N−1 , j ∈ N0 , i ≤ j − 2),

and thus
cj ⊂ 2j A
supp w (j ∈ N0 ).

For (a), by Theorem 21.18 (a) it is then sufficient to show

 
2js kwj kLp −1 kL1 kukLp1 kvkBps
≤ kψd .
j∈N−1 `r 2 ,r

This follows by Hölder’s inequality (Theorem A.4) and by (21.13):

kwj kLp = k∆≤j−2 u · ∆j vkLp ≤ k∆≤j−2 ukLp1 k∆j vkLp2
≤ kψd
−1 kL1 kukLp1 k∆j vkLp2 (j ∈ N−1 ).

And for (b) it is sufficient to show

  22s
2j(s+t) kwj kLp ≤ kukBps ,q kvkBpt ,q . (27.5)
j∈N−1 `q 1 − 2s 1 1 2 2

By Hölder’s inequality (Corollary A.9), we get

   
2j(s+t) kwj kLp ≤ 2js k∆≤j−2 ukLp1 2jt k∆j vkLp2
j∈N−1 `q j∈N−1 `q
   
≤ 2js k∆≤j−2 ukLp1 2jt k∆j vkLp2 .
j∈N−1 `q1 j∈N−1 `q2

183
By Theorem 27.4
    22s
2js k∆≤j−2 ukLp1 = 22s 2js k∆≤j ukLp1 ≤ kukBps ,q ,
j∈N−1 `q1 j∈N−1 `q1 1 − 2s 1 1

so that we conclude (27.5).

27.6. Here we consider some consequences of Theorem 27.5 in combination with The-
orem 21.23 and Theorem 23.4. The statements that we obtain for the “ <
= ” product are
summarized in Corollary 27.7.
Let C1 ≥ 1 as in Theorem 21.23 be such that

kukBp,q
s ≤ C1 kukBp,q
t (u ∈ S 0 , p, q ∈ [1, ∞], t, s ∈ R, t ≥ s).

Let C2 ≥ 1 as in Theorem 23.4 (b) be such that

kukLp ≤ C2 kukBp,q
t (u ∈ S 0 , t > 0)
kukLp ≤ C2 kukB 0 (u ∈ S 0 ).
p,1

Let ϕ generate a dyadic partition of unity. We write C3 = kϕd −1 kL1 (so that (27.3) holds
with C = C3 for all p, p1 , p2 , q, q1 , q2 ∈ [1, ∞] such that (27.2) is satisfied). For s < 0 we
define Cs = (1 − 2s )−1 (so that (27.4) holds with C = Cs for all p, p1 , p2 , q, q1 , q2 ∈ [1, ∞]
such that (27.2) is satisfied). Let s, t ∈ R and p, p1 , p2 , q, q1 , q2 , r ∈ [1, ∞] be such that
(27.2) is satisfied. Let u ∈ Bps1 ,q1 and v ∈ Bpt 2 ,q2 .
s+t .
(a) If s < 0 and t ∈ R, then u 4ϕ v exists in Bp,q
Observe that if t < 0, then the regularity of u 4ϕ v is worse than the regularity of
each of the terms u and v and that if t = 0, the regularity of u 4ϕ v equals the
regularity of u and if t > 0 the regularity of u 4ϕ v is larger than the one of u.
(b) If s = 0, t ∈ R and r ∈ [1, ∞] then u 4ϕ v exists in Bp,r t if u ∈ Lp1 and v ∈
Bp2 ,r . Observe that if u ∈ Bp1 ,1 (or differently said; if q1 = 1), then u ∈ Lp1 (by
s 0

Theorem 23.4). Therefore,

ku 4ϕ vkBp,r
t ≤ C3 kukLp1 kvkBpt ≤ C2 C3 kukB 0 kvkBpt .
2 ,r p1 ,1 2 ,r

t for all r ≥ q and

(c) If s > 0 and t ∈ R, then u 4ϕ v exists in Bp,r 2

ku 4ϕ vkBp,r
t ≤ C3 kukLp1 kvkBpt ≤ C1 C2 C3 kukBps kvkBpt .
2 ,r 1 ,q1 2 ,q2

By the above we obtain the following.

• If both s and t are in (−∞, 0), then u 4ϕ v and u 5 v and thus u < s+t ,
= ϕ v exist in Bp,q
and

ku < s+t ≤ (Cs + Ct )kukB s

= ϕ vkBp,q p
kvkBpt .
1 ,q1 2 ,q2

184
 s+t by (a) and u 5 v = v 4 u is in B s for all
• If s < 0 and t > 0, then u 4 v is in Bp,q p,r
r ≥ q1 by (c). Moreover, as r ≥ q1 implies r ≥ q,

ku <
= ϕ vkBp,r
s ≤ ku 4 vkBp,q
s+t + kv 4 ukB s
p,r

≤ (Cs + C1 C2 C3 )kukBps kvkBpt (r ≥ q1 ).

1 ,q1 2 ,q2

t
• If s > 0 and t > 0, then u 4 v ∈ Bp,q s
and v 4 u ∈ Bp,q s∧t
by (c) so that u 4 v ∈ Bp,r
2 1
for all r ≥ q1 ∨ q2 . Moreover,

ku < s∧t ≤ 2C1 C2 C3 kukB s

= ϕ vkBp,r p
kvkBpt (r ≥ q1 ∨ q2 ).
1 ,q1 2 ,q2

s by (a)
• (  ) If s < 0 and t = 0 and q2 = 1 (so that q = 1), then u 4 v exists in Bp,1
s
and u 5 v = v 4 u exists in Bp,q1 by (b) and thus u < s
= ϕ v exists in Bp,q1 . Moreover,

ku <
= ϕ vkBp,q
s ≤ (C1 Cs + C2 C3 )kukBps kvkBpt
1 1 ,q1 2 ,1

t and
• (  ) If s = 0 and t = 0 and q1 = q2 = 1 (so that q = 1), then u 4 v ∈ Bp,1
s s∧t
v 4 u ∈ Bp,1 by (b) so that u 4 v ∈ Bp,1 . Moreover,

ku < s∧t ≤ 2C2 C3 kukB s

= ϕ vkBp,1 p
kvkBpt .
1 ,1 2 ,1

t
• (  ) If s > 0 and t = 0 and q1 = 1 (so that q = 1), then u 4 v exists in Bp,q by (c)
2
s
and u 5 v = v 4 u exists in Bp,q by (b) and thus u < 0
= ϕ v exists in Bp,q2 . Moreover,

ku <
= ϕ vkBp,q
0 ≤ (C1 Cs + C2 C3 )kukBps kvkBp0
2 1 ,1 2 ,q2

We summarize the above observations in the following corollary.

Corollary 27.7. Let ϕ generate a dyadic partition of unity.

(a) Let s, t ∈ R \ {0}. There exists a C > 0 such that for all p, p1 , p2 , q, q1 , q2 ∈ [1, ∞]
with p11 + p12 ≤ 1 and
1 1 1 1
p = p1 + p2 , q = min{1, q11 + 1
q2 },

for all u ∈ Bps1 ,q1 and v ∈ Bpt 2 ,q2 : u < s∧t for all r ∈ [1, ∞] such that
= ϕ v exists in Bp,r



 q if s, t < 0,

q

if s > 0, t < 0,
1
r≥


 q2 if s < 0, t > 0,

q1 ∨ q2 if s > 0, t > 0,



and for such r,

ku <
= ϕ vkB s∧t∧(s+t) ≤ CkukBps kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ).
p,r 1 ,q1 2 ,q2

185
 (b) (  ) Let s, t ∈ R, s ≤ t and either s = 0 and t 6= 0 or s 6= 0 and t = 0. Then there
exists a C > 0 such that for all p, p1 , p2 ∈ [1, ∞] with p11 + p12 ≤ 1 and p1 = p11 + p12 ,
for all q ∈ [1, ∞], for all u ∈ Bps1 ,q and v ∈ Bpt 2 ,1 , u < s and
= ϕ v exists in Bp,q

ku <
= ϕ vkBp,q
s ≤ CkukBps kvkBpt (u ∈ Bps1 ,q , v ∈ Bpt 2 ,1 ).
1 ,q 2 ,1

(c) (  ) Then there exists a C > 0 such that for all p, p1 , p2 ∈ [1, ∞] with p11 + p12 ≤ 1
and p1 = p11 + p12 , for all q ∈ [1, ∞]: for all u ∈ Bp01 ,1 and v ∈ Bpt 2 ,1 , u <
= ϕ v exists
0
in Bp,1 and

ku <
= ϕ vkB 0 ≤ CkukB 0 kvkB 0 (u ∈ Bp01 ,1 , v ∈ Bp02 ,1 ).
p,1 p1 ,1 p2 ,1

Example 27.8. We consider the situation as in Example 21.26: Let d = 1. Let (ϕj )j∈N−1
be a dyadic partition of unity such that ϕ0 = 1 on A( 43 , 54 ) (see Exercise 21.B). For n ∈ N,
consider vn , un ∈ S 0 given by

vn = 21 (δ2n + δ−2n ) un = cos(2π2n X).

As in Example 21.26 for s ∈ R one has

2−2ns u2n ∈ B∞,∞

X
s
us := .
n∈N

We considered here the summation over even numbers, as this simplifies the calculations
that we make later.
We consider the existence of the paraproduct, resonance product and Bony product
us 4ϕ ut , us ϕ ut and us •ϕ ut for s and t in R.
• Paraproducts. Observe that ∆j us = 2−2ns u2n if j = 2n for some n ∈ N and
∆j us = 0 otherwise. Therefore
J−2
(
X wn,s,t J = 2n for some n ∈ N,
∆≤J−2 us · ∆J vt = ∆i us · ∆J ut =
i=−1 0 otherwise,

where
n−1 n−1
∆2i us · ∆2n ut = 2−2nt 2−2is u2i u2n
X X
wn,s,t := (n ∈ N).
i=1 i=1

It follows that us 4ϕ ut exists in S 0 if n∈N wn,s,t exists in S 0 . To show the latter we

P

invoke Theorem 21.18 (a). By considering the trigonometric identity for the product of
two cosines, or equivalently, the convolution of v2i ∗ v2n one finds

u2i u2n = 1
2 cos(2π(22n + 22i )X) + 21 cos(2π(22n − 22i )X) (i, n ∈ N). (27.6)

186
Observe that for all n ∈ N and i ∈ {1, . . . , n − 1} we have
3 2n
42 = (1 − 41 )22n ≤ 22n − 22i ≤ 22n + 22i ≤ (1 + 14 )22n = 54 22n .

Hence
[
2n 3 5
supp w
[ n,s,t = 2i u2n ⊂ 2 A( 4 , 4 ),
supp u\
i∈{1,...,n−1}

and thus, by our choice of dyadic partition of unity,

(
  wn,s,t j = 2n,
∆j wn,s,t = (27.7)
0 otherwise.

It is not difficult to see that (evaluate at the origin)

n−1
(
(n − 1)2−2nt s = 0,
= 2−2nt 2−2is =
X
wn,s,t −2ns −2s
L∞
i=1 2−nt 2 2−2s−2
−1
s 6= 0.

Now, as for example

2−2ns − 2−2s 1 − 22(n−1)s 1

sup 22ns −2s
= sup −2s
= −2s (s < 0),
n∈N 2 −1 n∈N 2 −1 2 −1

it follows by Theorem 21.18 (a) that us 4ϕ ut =

P s+t for s < 0
n∈N wn,s,t exists in B∞,∞
and t ∈ R and by (27.7)

1 22s
kus 4ϕ ut kB∞,∞
s+t
[ϕ] = = kus kB∞,∞
s [ϕ] kut kB∞,∞
t [ϕ] (s < 0, t ∈ R).
2−2s − 1 1 − 22s
On the other hand, if s > 0, then

2−2ns − 2−2s 1
sup −2s − 1
= 2s ,
n∈N 2 2 −1
t
and thus us 4ϕ ut exists in B∞,∞ for s > 0 and t ∈ R and

1 1
kus 4ϕ ut kB∞,∞
t [ϕ] = = 2s kus kB∞,∞
s [ϕ] kut kB∞,∞
t [ϕ] (s > 0, t ∈ R).
22s −1 2 −1
If s = 0, then

sup 2nr (n − 1) < ∞ ⇐⇒ r < 0,

n∈N

r+t for any r < 0, but u 4 u is not an element of B t

so that u0 4ϕ ut exists in B∞,∞ 0 ϕ t ∞,∞
for any t ∈ R.

187
 • Resonance products. By our choice of our dyadic partition of unity we have

∆j−1 us · ∆j ut + ∆j us · ∆j ut + ∆j us · ∆j−1 ut
(
∆2n us · ∆2n ut if j = 2n for some n ∈ N,
=
0 otherwise.

Similarly as above, see (27.6), we have

1
cos(2π22(n+1) X) + 1 ,

∆2n us · ∆2n ut = 2−n(t+s)
2
and thus
n n n+1
1 X 1
2−i(t+s) 1 +

2−(i−1)(t+s) u2i .
X X
∆2i us · ∆2i ut =
i=1
2 i=1 i=2
2

Therefore
−i(t+s) 1
 P 
1 n
n
X 


 2 i=1 2 j = −1,
∆j ∆2i us · ∆2i ut = 2 −(n−1)(t+s) 1 u j = 2n for some n ∈ N, n ≥ 2,
 2 2n
i=1 
0 otherwise.


From this we see (Exercise 27.B) that if

n
X
∆2i us · ∆2i ut
i=1

converges in S 0 as n → ∞ then
∞
1 X 
2−i(t+s) < ∞,
2 i=1

n∈N ∆2n us · ∆2n ut

P
which is only the case for t + s > 0. And if t + s > 0, then us ϕ ut =
t+s and
exists in B∞,∞

−(s+t) 1 2−(s+t) 1
kus ϕ ut kB∞,∞
s+t
[ϕ] = 2 −(s+t)
∨ 2−(t+s)
21−2 2
Pn
Exercise 27.B. Show that in the situation of Example i=1 ∆2i us · ∆2i ut converges
in S 0 as n → ∞ only if
∞
1 X 
2−i(t+s) .
2 i=1

(Hint: Show there exists a symmetric ψ ∈ S such that ∆j ψb = 0 for j ∈ N0 and hψ, 1i =
6 0
and test against ϕ.)

188
27.9 (Formal explanation of paraproducts and resonance products). In 21.27 we discussed
the language of frequencies corresponding to the Example 21.26 which of course closely
corresponds to Example 27.8. In this language, one may say that the paraproduct u 4 v
is the distribution that considers the product between the “low frequencies” of u times
the “high frequencies” of v.
In Example 27.8 we have seen that the frequencies of ∆≤2n−2 us · ∆2n ut are of order
22n , which is the same order as the frequences of ∆2n ut . See for example also Figure 3
and Figure 4 for an illustration that the frequency of the product of two functions of
different frequencies is close to the larger one of the two.

Figure 3: A function with high and one Figure 4: The product of the functions
with low frequency. with high and low frequencies.

Observe again, as we have seen in 27.6, that the regularity of us 4 ut is at most the
regularity of t. One therefore could say, in the above language: “One cannot improve the
regularity of ut by multiplying the high frequencies of ut by the low frequencies of us ”.
If, however, we turn around the roles of us and ut , we see that if us is of low regularity,
say s < 0, then: “One can improve the regularity of us by multiplying the low frequences
of us by the high frequencies of ut ”. Possibly this formal language helps the reader with
having some intuition about the estimates we have for paraproducts.
The term “resonance product” can be explained as follows. Contrary to the effect of
the paraproduct, the order of the frequencies of ∆j−1 us · ∆j ut + ∆j us · ∆j ut + ∆j us ·
∆j−1 ut may range between frequencies of order 1 to frequencies of order 2n+1 . Indeed, in
Example 27.8, we have seen that taking the product of two functions of equal frequencies
may give a function of larger and a function of lower frequency: cos2 X = 12 + 21 cos 2X
(see Figure 5). This effect relates to the word “resonance”, as it relates to two ‘systems’
that interact on the same frequency, which may ‘strengthen’ the outcome.

Theorem 27.10. Let ϕ generate a dyadic partition of unity. Let p, p1 , p2 , q, q1 , q2 ∈

[1, ∞] be such that p11 + p12 ≤ 1 and

1 1 1 1
p = p1 + p2 , q = min{1, q11 + 1
q2 },

(a) For all s, t ∈ R with s + t > 0, for u ∈ Bps1 ,q1 and v ∈ Bpt 2 ,q2 the resonance product

189
 cos x cos2 x 1 1
cos2 x = 2 + 2 cos 2x
Figure 5: The sine function and its square and the decomposition of the square of the
sine function in a low and high frequency function.

s+t . Moreover, there exists a C > 0 such that

of u and v exists in Bp,q

ku ϕ vkBp,q
s+t ≤ CkukB s
p
kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ), (27.8)
1 ,q1 2 ,q2

(b) (  ) For all s, t ∈ R with s + t = 0, for u ∈ Bps1 ,1 and v ∈ Bpt 2 ,∞ the resonance
0 . Moreover, there exists a C > 0 such that
product of u and v exists in Bp,∞

ku ϕ vkBp,∞
0 ≤ CkukBps kvkBpt (u ∈ Bps1 ,1 , v ∈ Bpt 2 ,∞ ). (27.9)
1 ,1 2 ,∞

(If the Besov norms in (27.3) and (27.4) are with respect to ϕ, as in Definition 21.12,
2N s s t
(21.14), then there exists an N ∈ N such that we may take C = 2kϕd
−1 kL1 2s −1 (2 +1+2 )
in (27.8) and in (27.9).)

Proof. Let u, v ∈ S 0 . For j ∈ N−1 let wj = ∆j−1 u · ∆j v + ∆j u · ∆j v + ∆j u · ∆j−1 v. For

(a) we apply Theorem 21.18 (b) and for (b) we apply Theorem 21.18 (c). Let us first
show that there exists a ball B such that

b j ⊂ 2j B
supp w (j ∈ N−1 ). (27.10)

Let C be a ball such that supp ϕj ⊂ 2j C for all j ∈ N−1 . Similarly as in the proof of
Theorem 27.5, by using that F(∆i u · ∆j v) = (ϕi u
b) ∗ (ϕj vb), we have
[
cj ⊂
supp w 2j−i C + 2j C ⊂ 2j (2C) (j ∈ N−1 ),
i∈{0,1}

from which we have (27.10) for B = 2C. For (a) and (b) (the latter case we have q1 = 1
and q2 = ∞ and thus q = 1), by Theorem 21.18 (b) and (c) it is sufficient to show

(2j(s+t) kwj kLp )j∈N−1 ≤ (2s + 1 + 2t )kukBps kvkBpt . (27.11)

`q 1 ,q1 2 ,q2

We use kwj kLp ≤ k∆j−1 u · ∆j vkLp + k∆j u · ∆j vkLp + k∆j u · ∆j−1 vkLp . By Hölder’s

190
inequality (both Theorem A.4 and Corollary A.9):
 
j(s+t) j(s+t)
(2 k∆j−1 u · ∆j vkLp )j∈N−1 ≤ 2 ∆j−1 u k∆j vkLp2
`q Lp1 j∈N−1 `q
   
≤ 2s 2(j−1)s ∆j−1 u Lp1
2jt k∆j vkLp2
j∈N−1 `q2
j∈N−1 `q1

= 2s kukBps kvkBpt .
1 ,q1 2 ,q2

One obtains similar estimates with “∆j−1 u·∆j v” replaced by “∆j u·∆j v” or “∆j u·∆j−1 v”
by which one can conclude (27.11).

Exercise 27.C. Consider the situation as in Theorem 27.10: Let p, p1 , p2 , q, q1 , q2 ∈

[1, ∞] be such that p11 + p12 ≤ 1 and
1 1 1 1
p = p1 + p2 , q = min{1, q11 + 1
q2 }.

Let s, t ∈ R, u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 . Show that if either s + t > 0 or s + t = 0 and
{q1 , q2 } = {1, ∞}, then ∆J+i u · ∆J+j v → 0 in S 0 as J → ∞ for all i, j ∈ N0 .

In the following theorem we consider the product between elements of Besov spaces.
Theorem 27.11. (a) Let s, t ∈ R \ {0}, s + t > 0. Let δ > 0. There exists a C > 0
such that for all p, p1 , p2 , q, q1 , q2 ∈ [1, ∞] such that p11 + p12 ≤ 1,
1 1 1
p = p1 + p2 ,

we have for all r ∈ [1, ∞] such that


 q1 if s > 0, t < 0,


r≥ 2 q if s < 0, t > 0,

q ∨ q

if s > 0, t > 0,
1 2

s∧t for all u ∈ B s

u • v exists in Bp,r t
p1 ,q1 and v ∈ Bp2 ,q2 , and for such r the operation
s t
• forms a bilinear continuous map Bp1 ,q1 × Bp2 ,q2 → Bp,r s∧t ;

ku • vkBp,r
s∧t ≤ CkukB s
p
kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ) (27.12)
1 ,q1 2 ,q2

s∧t−δ for all u ∈ B s

and, consequently, u • v exists in Bp,q t
p1 ,q1 and v ∈ Bp2 ,q2 , and the
operation • forms a bilinear continuous map Bps1 ,q1 × Bpt 2 ,q2 → Bp,q
s∧t−δ ;

ku • vkBp,q
s∧t−δ ≤ CkukB s
p
kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ). (27.13)
1 ,q1 2 ,q2

(b) (  ) Let s ∈ (0, ∞). There exists a C > 0 such that for all p, p1 , p2 , q ∈ [1, ∞]
with p11 + p12 ≤ 1 and p1 = p11 + p12 , for all u ∈ Bps1 ,1 and v ∈ Bp02 ,q , u • v exists in
0 and
Bp,1

ku • vkBp,q
0 ≤ CkukBps kvkBp0 (u ∈ Bps1 ,1 , v ∈ Bp02 ,q ). (27.14)
1 ,1 2 ,q

191
 (c) (  ) There exists a C > 0 such that for all p, p1 , p2 , ∈ [1, ∞] with p11 + p12 ≤ 1 and
1 1 1 0 0 0
p = p1 + p2 , for all u ∈ Bp1 ,∞ and v ∈ Bp2 ,1 : u • v exists in Bp,∞ and

ku • vkBp,∞
0 ≤ CkukB 0 kvkB 0 (u ∈ Bp01 ,1 , v ∈ Bp02 ,1 ). (27.15)
p1 ,1 p2 ,1

In all cases, also v • u exists and is equal to u • v.

Proof. Let ϕ generate a dyadic partition of unity. For each of the cases, we first show the
statements for “u •ϕ v” instead of “u • v” and then show that the product is independent
of the choice of ϕ. The bilinearity follows by Lemma 26.2. The continuity follows by the
norm estimates, see Exercise 27.D.
(a): (27.13) follows from (27.12) by Lemma 21.25: Let r be as in the statement and
q any element of [1, ∞]. Then for all δ > 0 there exists a C > 0 such that

kukBp,q
s−δ ≤ kukB s
p,r
(u ∈ S 0 ).

The existence of u •ϕ v and the estimate (27.12) for u •ϕ v, i.e.,

ku •ϕ vkBp,r
s∧t ≤ CkukB s
p
kvkBpt (u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ),
1 ,q1 2 ,q2

follows from Corollary 27.7 (a) and Theorem 27.10 (a).

(b): (27.14) follows from Corollary 27.7 (c) and Theorem 27.10 (a) as k·kBp,q
0 ≤ k·kB 0 .
p,1

(c): (27.15) follows from Corollary 27.7 (c) and Theorem 27.10 (b) as k · kBp,∞
0 ≤
k · kB 0 .
p,1

We are left to prove that u •ϕ v, so to say, is independent of the choice of ϕ (that

is, if η generates a dyadic partition of unity, then u •η v = u •ϕ v). If either u or v is in
Cp∞ , then this follows by Theorem 26.7. We continue by using that we can approximate
elements of Besov spaces whose second integration parameter is finite by Cp∞ functions,
see Theorem 23.8 (a) or Lemma 23.7. We give the proof in the situation of (a); the proof
in case of (b) and (c) is similar. Let us first assume that either q1 < ∞ or q2 < ∞. We may
assume q1 < ∞. Then there exists a sequence (ηn )n∈N in Cp∞ ∩ Bps1 ,q1 (take for example
ηn = ∆≤n u) such that ηn → u in Bps1 ,q1 . so that limn→∞ ηn v = limn→∞ ηn •ϕ v = u •ϕ v,
so that the limit is independent of ϕ. Now suppose q1 = q2 = ∞. We may assume s > 0.
Then we choose κ > 0 such that s − κ > 0 and s + t − κ > 0 and thus u ∈ Bps−κ 1 ,1
and
s∧t
u •ϕ v ∈ Bp,r for all r ∈ [1, ∞] and

ku •ϕ vkBp,r
s∧t ≤ Ckuk s−κ kvkB t
B p
(u ∈ Bps1 ,q1 , v ∈ Bpt 2 ,q2 ).
p1 ,1 2 ,q2

We may repeat the above limiting argument and obtain that in u •ϕ v is independent of
the choice of ϕ.

192
 Exercise 27.D. Let X, Y and Z be normed vector spaces with norms k · kX , k · kY and
k · kZ , respectively. Suppose F : X × Y → Z is bilinear and

kF (x, y)kZ ≤ kxkX kykY (x ∈ X, y ∈ Y ).

Show that F is continuous, i.e., if xn → x in X and yn → y in Y , then

F (xn , yn ) → F (x, y) in Z.

Lemma 27.12. Let s1 , s2 , s3 ∈ R \ {0}, s1 + s2 > 0 and s2 + s3 > 0. Let p1 , p2 , p3 , q1 , q2 ,

q3 ∈ [1, ∞] be such that p11 + p12 + p13 ≤ 1. For u1 ∈ Bps11,q1 , u2 ∈ Bps22,q2 and u3 ∈ Bps33,q3 ,
u1 • u2 , u2 • u3 , (u1 • u2 ) • u3 and u1 • (u2 • u3 ) exist in S 0 and

(u1 • u2 ) • u3 = u1 • (u2 • u3 ). (27.16)

Proof. By choosing κ1 , κ2 , κ3 > 0 such that s1 −κ1 , s2 −κ2 , s3 −κ3 6= 0, s1 −κ1 +s2 −κ2 > 0
and s2 − κ2 + s3 − κ3 > 0, we may as well assume that q1 = q2 = q3 = 1 (as u1 ∈
−κ1 , u ∈ B s2 −κ2 and u ∈ B s3 −κ3 ).
Bps11,q 1 2 p2 ,q2 3 p3 ,q3
Let p, p1,2 , p2,3 ∈ [1, ∞] be such that

1 1 1 1 1 1 1 1 1 1
= + + , = + , = + .
p p1 p2 p3 p1,2 p1 p2 p2,3 p2 p3
1 ∧s2 s2 ∧s3
By Theorem 27.11 (a) it follows that u1 • u2 exists in Bps12 ,1 , u2 • u3 exists in Bp23 ,1 and
s1 ∧s2 ∧s3
(u1 • u2 ) • u3 and u1 • (u2 • u3 ) exist in Bp,1 . Let ηn = ∆≤n u1 and ζn = ∆≤n u2 for
all n ∈ N. Then ηn , ζn ∈ Cp and ηn → u1 in Bps11,1 and ζn → u2 in Bps22,1 . Moreover, by
∞

Theorem 26.7

(ηn • ζn ) • u3 = ηn ζn u3 = ηn • (ζn • u3 ) (n ∈ N).

By the continuity of the product • as a map

1 ∧s2
Bps11,1 × Bps22,1 → Bps12 ,1 ,
2 ∧s3
Bps22,1 × Bps33,1 → Bps23 ,1 ,
1 ∧s2 s1 ∧s2 ∧s3
Bps12 s3
,1 × Bp3 ,1 → Bp,1 ,
2 ∧s3 s1 ∧s2 ∧s3
Bps11,1 × Bps23 ,1 → Bp,1 ,

(27.16) follows.

Definition 27.13. For s ∈ R we define C s = B∞,∞

s and write “k · kC s ” instead of
“k · kB∞,∞ ”.
s

For those who know the notation C t for t > 0 as in Definition 24.4 one may take the
above definition only for s ≤ 0. For t > 0 it does not make much of a difference as the
norm-like functions k · kC t and k · kB∞,∞
t are equivalent, see 24.8.

193
 By Theorem 23.2 the following statement is a consequence of Theorem 27.11.

Corollary 27.14. Let s > 0, t ∈ R and s + t > 0. Let δ > 0.

(a) If t < 0, u ∈ H s and v ∈ C t , then u • v exists and is an element of H t . Moreover,

the map

H s × Ct → H t, (u, v) 7→ u • v,

is a bilinear continuous map and there exists a C > 0 such that

ku • vkH t ≤ CkukH s kvkC t (u, v ∈ S 0 ). (27.17)

(b) If u ∈ H s and v ∈ C t , then u • v exists and is an element of H s∧t−δ . Moreover, the

map

H s × C t → H s∧t−δ , (u, v) 7→ u • v,

is a bilinear continuous map and there exists a C > 0 such that

ku • vkH s∧t−δ ≤ CkukH s kvkC t (u, v ∈ S 0 ). (27.18)

Proof. As H s = B2,2 s and C t = B t

∞,∞ with equivalent norms, (a) follows directly by
Theorem 27.11, (27.12) (with r = 2 and q1 = 2, q2 = ∞) and (b) follows similarly by
(27.13) if t 6= 0. If t = 0, then (27.13) implies

ku • vkH s∧t−δ . kukH s kvk δ . kukH s kvkC t .

C t− 2

We turn now to a specific case of products between Besov spaces with p = q = ∞.

By Proposition 5.3 it follows that if k, m ∈ N0 , then the product map Cbk × Cbm →
Cbk∧m , (f, g) 7→ f g is a continuous bilinear map. The following statement is of a similar
nature.
It is widely used in the theory of SPDEs. See for example [Hai14, Proposition 4.14]
and [GIP15, Lemma 2.1 and text below] .

Corollary 27.15. Let s, t ∈ R \ {0} and s + t > 0. If u ∈ C s and v ∈ C t , then u • v exists

and is an element of C s∧t . Moreover, the map

C s × C t → C s∧t , (u, v) 7→ u • v,

is a bilinear continuous associative symmetric map and there exists a C > 0 such that

ku • vkC s∧t ≤ CkukC s kvkC t (u, v ∈ S 0 ). (27.19)

194
 The following corollary is a consequence of Theorem 27.5 and Theorem 27.10 and is
left as an exercise:

Corollary 27.16. Let s ∈ (0, ∞). There exists a C > 1 such that for all p, q ∈ [1, ∞]
 
ku • vkBp,q
s s kvkL∞ + kukL∞ kvkB s
≤ C kukBp,q p,q
(u, v ∈ S 0 ).

Consequently, L∞ ∩ Bp,q
s is a Banach algebra under the norm C(k · k ∞ + k · k s ).
L Bp,q

Exercise 27.E. Prove Corollary 27.16.

Another consequence is the following:

Theorem 27.17. Let s, t ∈ R and s+t > 0. Let δ > 0 be such that s∧t−δ ∈ (0, ∞)\N0 .
For u ∈ H s and v ∈ H t we have u • v exists in W s∧t−δ,1 . Moreover, the product map
H s × H t → W s∧t−δ,1 , (u, v) 7→ u • v is continuous, i.e., there exists a C > 0 such that

ku • vkW s∧t−δ,1 ≤ CkukH s kvkH t (u ∈ H s , v ∈ H t ).

Exercise 27.F. Prove Theorem 27.17. (Hint: Observe that it is allowed for either s or
t to be equal to zero. Furthermore: Theorem 23.4.)

27.1 Comments ...

Remark 27.18 (About notation and latex). In many textbooks one writes “Tu v” for the
paraproduct instead of “u 4 v” (for example in [BCD11] ). In this sense one views Tu as
an multiplying operator. Also “Π(u, v)” or “R(u, v)” is written for the resonance product.
In the application to SPDEs in the authors of the paper [GIP15] wrote “u ≺ v” and
“u ◦ v” for the para- and resonance product, respectively. The latter notation changed in
the SPDE literature, with some authors creating new symbols, for example “<” and “=”
with circles around them. In the latter case, “≤” with a circle around it is then used for
the sum of the paraproduct and the resonance product, for which the authors of [GIP15]
used “”.
For the sum of the paraproduct and the resonance product we write 4. The following
table presents the latex commands for the symbols used in these notes.
\varolessthan 4
\varogreaterthan 5
\varodot 
\mathrlap{\odot}{\olessthan} <
\mathrlap{\odot}{\ogreaterthan} =
\mathrlap{\olessthan}{\ogreaterthan} =
<
\mathrlap{\mathrlap{\odot}{\olessthan}}{\olessthan} =
<

195
27.19 (The notation .). We have seen multiple arguments in which different estimates
were combined, which were for example of the form
kΦ(x)kY ≤ CkxkX (x ∈ X),
with C > 0, (X, k · kX ) and (Y, k · kY ) normed vector spaces and Φ : X → Y . Let us call
such a “C” a “toleration constant” for the moment. For example in 27.6 we had toleration
constants C1 , C2 , C3 and Cs . Keeping track of those toleration constants becomes more
cumbersome and is a rather uninteresting job as the number of those constants grows.
Indeed, often one is only interested in the existence of such a toleration constant. For
this reason the notation “.” has been invented. Its usage is the following. If f and g are
functions on a set X with values in [0, ∞], we write
f (x) . g(x) (x ∈ X),
to indicated the existence of a number C > 0 such that f (x) ≤ Cg(x) for all x ∈ X.
For sets X, Y and functions f, g : X × Y → [0, ∞] observe the difference between
f (x, y) . g(x, y) ((x, y) ∈ X × Y ), (27.20)
and
if y ∈ Y, then f (x, y) . g(x, y) (x ∈ X). (27.21)
In the literature (27.20) is also rendered
f (x, y) . g(x, y) (x ∈ X), uniformly in y ∈ Y,
also, sometimes (27.21) is rendered
f (x, y) .y g(x, y) (x ∈ X).

28 Solutions to elliptic PDEs in Besov spaces

In Section 13 we have considered elliptic operators with bounded coefficients (see 13.4),
i.e., second order linear partial differential operators of the form
d
X d
X
P =− aij ∂i ∂j + bi ∂i + c,
i,j=1 i=1

aij , bi , c ∈ L∞ (Ω) for i, j ∈ {1, . . . , d} such that there exists a θ > 0 such that
d
X
aij (x)yi yj ≥ θ|y|2 (almost all x ∈ Ω, y ∈ Rd ).
i,j=1

In this section we on the one hand restrict ourselves to Ω = Rd and aii = 1 and aij = 0
if i 6= j for i, j ∈ {1, . . . , d} but on the other hand, instead of considering bi and c in
L∞ , we allow bi and c to be in the Besov space B∞,∞ s for some s < 0 (in which L∞ is
embedded, see Theorem 23.4).

196
Lemma 28.1. Let s ∈ R and
s+1 s
bi ∈ B∞,∞ , c ∈ B∞,∞ (i ∈ {1, . . . , d}).
Let t > 0 be such that s + t > 0 and s + 1 ≤ t. Then bi • ∂i u and c • u exist in H s−δ for
all i ∈ {1, . . . , d} and there exists a C > 0 such that
d  d 
(u ∈ S 0 ).
X X
bi • ∂i u + c • u ≤ C kckB∞,∞
s + kbi kB∞,∞
s+1 kukH t
i=1 H s−δ i=1

Proof. This follows by Theorem 23.2, Theorem 21.23 and Theorem 27.11. Observe in
particular that as s ≤ t − 1, (for the . notation see 27.19)
kbi • ∂i ukH s−δ . kbi • ∂i ukH (s+1)∧(t−1)−δ
. kbi kB∞,∞
s+1 k∂i ukH t−1 . kbi k s+1 kukH t
B∞,∞ (i ∈ {1, . . . , d}, u ∈ H t ).

By this lemma we can extend the elliptic operator from D to H t for large enough t:
28.2 (Assumptions). Let s ∈ R, t > 0, s + t > 0, s ≤ t and δ > 0. Let
s+1 s
bi ∈ B∞,∞ , c ∈ B∞,∞ (i ∈ {1, . . . , d}).
Let L : H t → H (s−δ)∧(t−2) be defined by
d
X
Lu = −∆u − bi ∂i u − cu (u ∈ H t ). (28.1)
i=1

Let us describe the goal of this section for L, s and t as in 28.2. Similarly to The-
orem 13.9 we want to consider conditions on L and f (in some Besov space) such that for
some class of real numbers β there exists a u in H t that satisfies
Lu + βu = f.
This equation can be rewritten by the formula
d
X
(β − ∆)u = bi • ∂i u + c • u + f,
i=1

which, with (β − ∆)−1 = (β + 4π 2 |X|2 )−1 (D), can be rewritten as

d
X 
u = (β − ∆)−1 bi • ∂i u + c • u + f . (28.2)
i=1

The following strategy is often used to find a u that satisfies this formula. First one
defines a function Φ by the formula Φ(v) = (β − ∆)−1 ( di=1 bi • ∂i v + c • v + f) (basically
P

by “the right-hand side” of (28.2)). Then u satisfies (28.2) if and only if Φ(u) = u. One
also says that u is a fixed point of Φ. In order to show the existence of fixed points, we
recall Banach’s fixed point theorem, and leave its proof for the reader.

197
Definition 28.3. Let (X, d) be a non-empty complete metric space. A function Φ : X →
X is called a contraction if there exists a θ ∈ (0, 1) such that

d(Φ(x), Φ(y)) ≤ θd(x, y) (x, y ∈ X).

Theorem 28.4 (Banach’s fixed point theorem). Let (X, d) be a non-empty complete
metric space. Suppose that Φ : X → X is a contraction. Then there exists a precisely one
x∗ in X such that Φ(x∗ ) = x∗ . Moreover, by defining Φ1 = Φ and Φk = Φ ◦ Φk−1 for
k ∈ N with k ≥ 2, we have for each x ∈ X that

lim Φk (x) = x∗ .
k→∞

Before we prove the existence of a fixed point, let us have a closer look at the operator
(β − ∆)−1 on Besov spaces.
Lemma 28.5. Let m ∈ (0, 2]. There exists a C > 0 such that for all s ∈ R, p, q ∈ [1, ∞]
m
k(β − ∆)−1 ukBp,q
s+m ≤ Cβ
−1
(β 2 ∨ 1)kukBp,q
s (u ∈ S 0 , β > 0). (28.3)

Proof. Let σ : Rd → (0, ∞) be defined by

σ = (1 + 4π 2 |X|2 )−1 .

By Lemma 25.16 we have Mm (σ) < ∞ for m ∈ (−∞, 2]. Observe that

(β + 4π 2 |X|2 )−1 = β −1 (1 + 4π 2 | β X|2 )−1 = β −1 l

p
1 σ.
β− 2

Therefore

k(β − ∆)−1 ukBp,q

s+m = β
−1
(l 1 σ)(D)u s+m (u ∈ S 0 , β > 0).
β− 2 Bp,q

Hence by Theorem 25.18 (e) we obtain (28.3).

Theorem 28.6. Let s ∈ (−2, 2). Let

s+1 s
bi ∈ B∞,∞ , c ∈ B∞,∞ , f ∈ Hs (i ∈ {1, . . . , d}).

For all t > 0 with t + s > 0, 1 + s < t < 2 + s there exists a γ > 0 such that for all β > γ,
the formula
d
X 
−1
Φf (v) = (β − ∆) bi • ∂i v + c • v + f (v ∈ H t )
i=1

defines a contraction H t → H t . Consequently, for such t and γ, for each β > γ there
exists exactly one u ∈ H t such that

Lu + βu = f,

where L is as in (28.1) of 28.2.

198
Proof. First observe that as t > |s|, we have t > s, t > 0 and t + s > 0. Let δ > 0 be
such that t < 2 − δ + s. Then we can and do choose m ∈ (0, 2) such that t < s − δ + m.
By Lemma 28.5 there exists a C1 > 0 and by Lemma 28.1 there exists a C2 > 0 such
that for all v ∈ H t
d
m
kΦf (v)kH s−δ+m ≤ C1 β −1 (β 2 ∨ 1)
X
bi • ∂i v + c • v + f
i=1 H s−δ
 d 
m
β −1 (β 2 ∨ 1)kvkH t + kf kH s .
X
≤ C1 C2 kckB∞,∞
s + kbi kB∞,∞
s+1

i=1

As
kΦf (v)kH t ≤ kΦf (v)kH s−δ+m (v ∈ H t ),
we see that Φ defines a function H t → H t . As Φf (v − w) = Φ0 (v − w), we see that Φf is
a contraction if
 d 
m
β −1 (β 2 ∨ 1) < 1.
X
C1 C2 kckB∞,∞
s + kbi kB∞,∞
s+1

i=1

From this one can obtain a γ such that the statement holds, for example
 d  2
2−m
X
γ = 1 ∨ C1 C2 (kckB∞,∞
s + kbi kB∞,∞
s+1 ) .
i=1

The methods used above do not restrict to second order linear partial differential
operators, as the following theorem shows. The proof follows the same strategy as above
and is left to the reader.
Theorem 28.7. Let k ∈ N. Let s ∈ (−2k, 2k). Let
s+|α| s
bα ∈ B∞,∞ , f ∈ B∞,∞ (α ∈ Nd0 , |α| ≤ 2k − 1).
For all t > 0 with t + s > 0, 2k − 1 + s < t < 2k + s there exists a γ > 0 such that for
all β > γ, the formula
 
Φf (v) = (β − ∆k )−1
X
bα • ∂ α v + f (v ∈ H t )
α∈Nd0
|α|≤2k−1

defines a contraction H t → H t . Consequently, for such t and γ, for each β > γ there
exists exactly one u ∈ H t such that
X
βu − ∆k u − bα • ∂ α u = f.
α∈Nd0

Exercise 28.A. Prove Theorem 28.7.

199
29 Overview

In this section we group statements according to their type and give references to where
the full statement can be found. Some formula’s are given without any description of the
different symbols.

29.1 Spaces of functions and distributions

C(Ω), C k (Ω), C ∞ (Ω) : 1.3. D(Ω) : 1.5 and 4.1. D0 (Ω) : 2.1 and 2.3. L1loc : 2.5. M(Ω :
2.23. Cbk , Cb∞ : 9.2. W k,p : 12.3. S : 14.2. Cp∞ : 14.6. S 0 : 15.1. H s : 20.5. Hps : 20.10.
s [ϕ] : 21.12. B s : 21.21.
Bp,q p,q

29.2 Characterisations

Distributions of order 0 : 2.28.

Distributions in terms of sequentially continuous maps : Theorem 4.14.
Compactly supported distributions and E 0 : 5.2 and 5.6.
E 0 (Ω) in terms of linear combinations of derivatives of continuous functions : 6.2.
D0 (Ω) in terms of linear combinations of derivatives of continuous functions : 6.3.
Distributions of order k : 9.6.
Convolutions : 10.2 and 17.15.
H k = {u ∈ S 0 : (1 + |X|)k u
b ∈ L2 } : 20.3.
s s
B2,2 = H : 23.2.

29.3 Convergences

• Convergences of mollifications.
Pointwise for elements in L1loc . Theorem 7.15 (a).
Uniformly on compacts for elements in C(U ). Theorem 7.15 (b).
In Lploc . Theorem 7.15 (c).
In Lploc . Theorem 7.15 (d).
In D, D0 , E, E 0 . Theorem 8.12.
ε↓0
“In Cp∞ ”: qm,k (ψε ∗ η − η) −−→ 0 Lemma 26.3.
• Other convergences.

200
 λ↓0
(lλ χ)ϕ −−→ ϕ Lemma 14.10 (c).
PN N →∞
χ ϕ −−−−→ ϕ Lemma 14.10 (d).
Pn=1 n
ϕ ψ=ψ Lemma 21.8.
Pj∈N−1 j
∆ ψ=ψ Lemma 21.9.
Pj∈N−1 j
j∈N ∆j u = u Lemma 21.9.
−1 
J J→∞
− η −−−→ 0
P
qm,k j=−1 ∆j η Lemma 26.4.
P
i∈N−1 ∆i f = f Lemma 26.9.

29.4 Identities

• Leibniz’ Rule and the Multinomial Theorem.

∂ α (f g) = β∈Nd αβ (∂ β f )(∂ α−β g)
P

1.14.
0
β≤α
k
α
(x1 + · · · + xd )k =
P
α∈Nd0 α x Theorem 20.4.
|α|=k
• Identities involving convolutions (and no Fourier transforms).
δ0 ∗ ϕ = ϕ Lemma 8.2 and Lemma 17.4
δy ∗ ϕ = Ty ϕ Lemma 8.2 and Lemma 17.4
(u ∗ ϕ)ˇ= ǔ ∗ ϕ̌ Lemma 8.2 and Lemma 17.4
Ty (u ∗ ϕ) = (Ty u) ∗ ϕ = u ∗ (Ty ϕ) Lemma 8.2 and Lemma 17.4
u(ϕ) = u ∗ ϕ̌(0) Lemma 8.2 and Lemma 17.4
∂ α (u ∗ ϕ) = u ∗ (∂ α ϕ) = (∂ α u) ∗ ϕ Theorem 8.4, Theorem 17.6.
(u ∗ ϕ) ∗ ψ = u ∗ (ϕ ∗ ψ) Theorem 8.9, Theorem 17.11.
u∗v =v∗u Theorem 10.6, Theorem 17.13.
(u ∗ v) ∗ ψ = u ∗ (v ∗ ψ) Theorem 10.6, Theorem 17.13.
δ0 ∗ u = u Lemma 10.10, Lemma 17.14.
δy ∗ u = Ty u Lemma 10.10, Lemma 17.14.
R(u ∗ v) = R(u) ∗ R(v) Lemma 10.10, Lemma 17.14.
Ty (u ∗ v) = (Ty u) ∗ v = u ∗ (Ty v) Lemma 10.10, Lemma 17.14.
∂ α (u ∗ v) = (∂ α u) ∗ v = u ∗ (∂ α v) Lemma 10.10, Lemma 17.14.
• Identities involving Fourier transforms.

201
 R R
f gb = fbg Theorem 16.10.
F −1 = FR = RF Theorem 16.24.
F(∂ β u) = (2πiX)β u b Theorem 16.24.
β
∂ u = F((−2πiX) u)
b β Theorem 16.24.
F(Ty u) = e−2πihX,yi u b Theorem 16.24.
2πih X,yi
Ty u = F(e
b u) Theorem 16.24.
1 b
F(u ◦ l) = | det l| ◦ l∗
u Theorem 16.24.
1
F(lλ u) = |λ|d l 1 ub Theorem 16.24.
λ
F(f ∗ g) = fbgb Theorem 17.1.
F(ϕ ∗ ψ) = ϕbψb Theorem 17.2.
F(u ∗ ϕ) = ϕbub Theorem 17.7.
F(u ∗ v) = vbu
b Theorem 17.13.
• Identities involving norms.
kfbkL2 = kf kL2 Theorem 16.27.
kvkLp = sup{|hv, f i| : f ∈ Lq , kf kLq ≤ 1} Theorem 21.28.
kµkM = sup{|hµ, f i| : f ∈ C0 , kf kC0 ≤ 1} Theorem 21.28.
• Identities involving Fourier multipliers.
τ (D)σ(D)u = (στ )(D)u = σ(D)τ (D)u Lemma 19.4.
σ(D)(lλ u) = lλ [(lλ σ)(D)u] Lemma 19.4.
σ(D)ϕ = F −1 (σ) ∗ ϕ Lemma 19.4.
σ(D)u = F −1 (σ u b) = F −1 (σ) ∗ u Lemma 19.4.
h∆j u, ψi = hu, ∆j ψi Lemma 21.9.
∂ α u = λ|α|+d (lλ hα ) ∗ u Lemma 21.14.
u = λ−k α∈Nd :|α|=k λd (lλ gα ) ∗ ∂ α u
P
Lemma 21.14.
0

• Identities involving “products”.

(u + w) •ϕ v = u •ϕ v + w •ϕ v Lemma 26.2.
u •ϕ (v + w) = u •ϕ v + u •ϕ w Lemma 26.2.
λ(u •ϕ v) = (λu) •ϕ v + u •ϕ (λv) Lemma 26.2.
ηψ = η • ψ Theorem 26.5.
ηu = η • u Theorem 26.7.
fg = f • g Lemma 26.10.
(u + w) 4ϕ v = u 4ϕ v + w 4ϕ v Lemma 27.2.
u 4ϕ (v + w) = u 4ϕ v + u 4ϕ w Lemma 27.2.
λ(u 4ϕ v) = (λu) 4ϕ v + u 4ϕ (λv) Lemma 27.2.
u •ϕ v = u 4ϕ v + u ϕ v + u 5ϕ v Lemma 27.3.
(u1 • u2 ) • u3 = u1 • (u2 • u3 ) Lemma 27.12.

29.5 Continuity of operations

Continuity of the reflection R, translations Ty , derivations ∂ α , multiplication with a

smooth function and composition with a linear bijection: 4.4 and 15.4

202
• Continuity of the pairing map
D0 (Ω) × D(Ω) → F, E 0 (Ω) × E(Ω) → F 4.25.
S 0 (Ω) × S(Ω) → F. Theorem 15.10 (c).
• Continuity of the product maps
E(Ω) × D(Ω) → D(Ω) 5.3.
D0 (Ω) × E(Ω) → D0 (Ω) 5.3.
E 0 (Ω) × E(Ω) → E 0 (Ω) 5.3.
W k,p (Ω) × W k,r (Ω) → W k,q (Ω) 12.10.
Cb∞ × S → S 14.7.
S ×S →S 14.7.
• Continuity of convolution maps
Lp × Lq → Lr . Theorem 7.7.
D0 × D → E. Theorem 8.7.
E 0 × D → D. Theorem 8.7.
E 0 × E → E. Theorem 8.7.
S 0 × S → S 0. Theorem 17.10.
S 0 × S → E. Theorem 17.10.
E 0 × S → S. Theorem 17.10.
E 0 × S 0 → S 0. Lemma 17.16.
• Continuity of the Fourier transformation
L1 → C0 (Rd , ¸) Theorem 16.6.
S(Rd , C) → S(Rd , C) Theorem 16.16.
S 0 (Rd , C) → S 0 (Rd , C) Theorem 16.24.
L2 (Rd , C) → L2 (Rd , C) Theorem 16.27.
• Continuity of Fourier multipliers
σ(D) : S 0 → S 0 Lemma 19.2.

29.6 Metrizability and completeness statements

203
 D(Ω) is NOT metrizable. 4.15
Countably many seminorms =⇒ metrizability. Theorem 3.8
C m (Ω) is a Fréchet space. Theorem 4.19
E(Ω) is a Fréchet space. Theorem 4.19
DK (Ω) is a Fréchet space. Theorem 4.19
D0 (Ω) is weak∗ complete. Theorem 4.26
E 0 (Ω) is weak∗ complete. Theorem 4.26
E 0 (Ω) is NOT metrizable. 5.13.
W k,p (Ω) is a Banach space. Theorem 12.7.
H k (Ω) is a Hilbert space. Theorem 12.14.
S is a Fréchet space. Theorem 14.12.
S 0 is weak∗ complete. Theorem 15.10 (b).
H s is a Hilbert space. Theorem 20.7.
s is a Banach space.
Bp,q Theorem 21.32.

29.7 Denseness and separability statements

• Denseness:
D(Ω) is sequentially dense in E(Ω). Theorem 5.10
E 0 (Ω) is sequentially dense in D0 (Ω). Theorem 5.10
D(Ω) is sequentially dense in D0 (Ω) and E 0 (Ω). Lemma 8.14.
D(Ω) is dense in Lp (Ω). Theorem 8.17.
D(Ω) is dense in W k,p (Ω). Theorem 12.11.
D is dense in S. Theorem 14.12
Cb∞ ∩ Bp,qs is dense in B s
p,q Theorem 23.8.
D is dense in Bp,qs . Theorem 23.8.
S is NOT dense in B∞,q s 23.9.
• Separability:
D(Ω), D0 (Ω) and E 0 (Ω) are separable. Theorem 8.15.
S is separable. Theorem 14.12.

29.8 Continuous embeddings

• Continuous embeddings

204
 D(Ω) ,→ Lp (Ω) ,→ Lploc (Ω) ,→ D0 (Ω). Theorem 4.28
D(Ω) ,→seq E(Ω). Theorem 5.10
E 0 (Ω) ,→ D0 (Ω). Theorem 5.10
E 0 (Ω) ,→ E 0 (Rd ). Theorem 5.10
C m ,→ C k , Cbm ,→ Cbk . Lemma 12.8
W m,p ,→ W .k,p Lemma 12.9
D ,→seq S ,→ Cb∞ ,→ E. Theorem 14.11.
S ,→ Lp Lemma 14.14.
S ,→ W k,p Lemma 14.15.
E 0 ,→ S 0 ,→ D0 Theorem 15.2.
Lp ,→ S 0 Theorem 15.6.
W k,p ,→ S 0 Lemma 14.15.
H s ,→ H r Lemma 20.8.
Bps1 ,q1 ,→ Bpt 2 ,q2 t ≤ s − d( p11 − p12 )


Theorem 21.23.
s
Bp,q s−ε
,→ Bp,q Lemma 21.25.
1 2
s s
B2,2 ,→ H ,→ B2,2 s Theorem 23.2.
k ,→ W k,p ,→ B k
Bp,1 Theorem 23.4.
p,∞
t ,→ W k,p ,→ B s
Bp,q (s, t ∈ R, s < k < t) Theorem 23.4.
p,q
k k
B∞,1 ,→ Cb ,→ B∞,∞ k Theorem 23.5.
D ,→seq S ,→ Bp,q s Corollary 23.6.
• Embeddings which are not bijective.
W k,∞ ,→ B∞,∞
k Example 23.11.
• Embeddings which are no homeomorphisms on their image.
D(Ω) ,→seq E(Ω) 5.13
E 0 (Ω) ,→ D0 (Ω) 5.13
D ,→ S 15.11
S ,→ E 15.11
S 0 ,→ D0 15.11
E 0 ,→ S 0 15.11

29.9 Inequalities

• Inequalities involving ‘products’ and convolutions.

kf ∗ gkLr ≤ kf kLp kgkLq Theorem 7.7.
kuvkW k,p ≤ CkukW k,r kvkW k,q Lemma 12.10.
kσϕkm,S ≤ Cqm,k (σ)kϕkm+k,S Lemma 14.7.
kf ∗ gk`r ≤ kf k`p kgk`q Theorem 21.17.

205
 kσϕkm,S ≤ CkσkC m kϕkm,S . Lemma 14.7.
kϕψkm,S ≤ Ckϕkm,S kψkm,S . Lemma 14.7.
ku 4ϕ vkBp,r
s ≤ CkukLp1 kvkBps ,r Theorem 27.5.
2
ku 4ϕ vkBp,q
s+t ≤ CkukB s
p1 ,q1
kvk Bpt 2 ,q2 Theorem 27.5.
ku <
= ϕ vkB s∧t∧(s+t) ≤ CkukBp ,q kvkBpt
s Corollary 27.7 (a).
p,r 1 1 2 ,q2
ku <
= ϕ vkBp,q
s ≤ CkukBps ,q
kvkBpt Corollary 27.7 (b).
1 2 ,1
ku <
= ϕ vkB 0 ≤ CkukB 0 kvkB 0 Corollary 27.7 (c).
p,1 p1 ,1 p2 ,1
ku ϕ vkBp,q
s+t ≤ Ckuk kvkBpt ,q
Bps1 ,q1 Theorem 27.10.
2 2
ku ϕ vkBp,∞
0 ≤ CkukBps ,q kvkBpt ,q Theorem 27.10.
1 1 2 2
s∧t ≤ CkukB s
ku • vkBp,r p1 ,q1
kvkBpt ,q Theorem 27.11 (a).
2 2
ku • vkBp,q
s∧t−δ ≤ CkukB s
p1 ,q1
kvkBpt ,q Theorem 27.11 (a).
2 2
ku • vkBp,q
0 ≤ CkukBps ,1 kvkBp0 ,q Theorem 27.11 (b).
1 2
ku • vkBp,∞
0 ≤ CkukB 0 kvkB 0 Theorem 27.11 (c).
p1 ,1 p2 ,1
ku • vkH t ≤ CkukH s kvkC t Corollary 27.14.
ku • vkH s∧t−δ ≤ CkukH s kvkC t Corollary 27.14.
ku • vkC s∧t ≤ CkukC s kvkC t Corollary 27.15.
ku • vkBp,q
s ≤ Corollary 27.16
 
C kukBp,q
s kvkL∞ + kukL∞ kvkB s
p,q

• Inequalities of operations.
k∂ α f kC k−|α| ,K ≤ kf kC m ,K Lemma 12.8.
k∂ α f kC k−|α| ≤ kf kC m Lemma 12.8.
k∂ α ukW k−|α|,p ≤ kukW m,p Lemma 12.9.
k∂ α ukH r−|α| ≤ kukH s Lemma 20.8.
k∆j f kLp ≤ kϕ c0 kL1 kf kLp Lemma 21.9.
PJ
j=−1 ∆j f p
≤ kϕd −1 kL1 kf kLp Lemma 21.9.
L
k+d( p1 − 1q )
maxα∈Nd k∂ α ukLq ≤ Cλ kukLp Theorem 21.15.
0
|α|=k
1 k
C λ kukL
p ≤ maxα∈Nd k∂ α ukLp ≤ Cλk kukLp Theorem 21.15.
0
|α|=k
k∂ α ukB t−|α| ≤ CkukBps Theorem 21.23.
p2 ,q2 1 ,q1

kσ(D)ukLp ≤ CkukLp Theorem 25.1.

kσ(D)ukBp,q
s ≤ CkukBp,q
s Theorem 25.1 and
Theorem 25.18 (d).
k(lµ σ)(D)ukLp ≤ Cµ−m λ−m kukLp Theorem 25.18 (a).
k(lµ σ)(D)ukBp,q −m kuk s
s+m ≤ Cµ Bp,q Theorem 25.18 (b).
kσ(D)ukBp,q
s+m ≤ CkukB s
p,q
Theorem 25.18 (c).
k(lµ σ)(D)ukBp,q −m ∨ 1)kuk s
s+m ≤ C(µ Bp,q Theorem 25.18 (e).
≤ Ckψkk,S
P
j∈F ∆j ψ Lemma 26.6.
m,S

206
• Inequalities related to embeddings.
kϕkC m ,K ≤ kϕkC m ≤ kϕkm,S Lemma 14.10.
kϕkm,S ≤ (1 + supx∈supp ϕ |x|)m kϕkC m Lemma 14.10.
kψkk+1,S
kψkk,S ≤ 1+r Lemma 14.10.
k · kLp ≤ Ck · km,S . Lemma 14.14.
2js kuj kLp j∈N−1 q


s [ϕ] ≤ C
kukBp,q Theorem 21.18.
`
1
C kuk B s [ψ] ≤ kukB s [ϕ] ≤ CkukB s [ψ]
p,q p,q p,q
Corollary 21.20.
k∂ α ukB t−|α| ≤ CkukBps ,q Theorem 21.23.
p2 ,q2 1 1

kukBp,q
s−ε ≤ CkukB s
p,q1
Lemma 21.25.
2
kukX ≤ lim inf m→∞ kunm kX Lemma 21.30.
kukLp ≤ lim inf m→∞ kunm kLp Lemma 21.31.
kukBp,q
s ≤ lim inf m→∞ kunm kBp,q
s Theorem 21.32.

s
c 1 + |ξ|2 ≤ j∈N−1 22sj ϕj (ξ)2 ≤
P
Lemma 23.1.

s
C 1 + |ξ|2
kukBp,q
s ≤ CkukW k,p Theorem 23.4.
kukBp,∞
k ≤ CkukW k,p Theorem 23.4.
kukW k,p ≤ CkukBp,q t Theorem 23.4.
kukW k,p ≤ CkukB k Theorem 23.4.
p,1
(1 + 2s )−1 kukBp,q
s ≤ k(2js k∆≤j ukLp )j∈N−1 k`q Theorem 27.4.
k(2 k∆≤j ukLp )j∈N−1 k`q ≤ (1 − 2s )−1 kukBp,q
js s Theorem 27.4.
• Other inequalities
Poincaré inequality
kukLp ≤ Ck∇ukLp (for u ∈ W01,p (Ω)) Theorem 12.17.
Lemma of Riemann–Lebesgue
kgb(a)k ≤ 21 kg − T 1 gkL1 . Lemma 16.5.
2a
(a + b)θ ≤ aθ + bθ Lemma 20.2.

29.10 Other statements

supp u ∗ v ⊂ supp u ∗ supp v. Theorem 7.10, Theorem 8.6,

Theorem 10.11.
v ∈ E 0 implies vb ∈ Cp∞ Lemma 17.8.
Principle of uniform boundedness for D0 , E 0 . Theorem 4.24
supp u = {x} =⇒ u = α∈Nd :|α|≤k cα ∂ α δx
P
Theorem 9.5
0
Principle of uniform boundedness for S 0 . Theorem 15.10 (a)

207
A Preliminaries on Lp spaces

In this section we let (X, A, µ) be a measure space (see Definition A.1. For the purpose
of this course, X will mostly be either be Rd , A the set of Lebesgue measurable sets and
µ the Lebesgue measure or X with be N or Zd or any countable space, A the power set
of X and µ the counting measure. As usual, F is either R or C.
Before we turn to Lp spaces, let us recall some standard definitions from measure
theory.
Definition A.1. Let X be a set. A collection A of subsets of X is called a σ-algebra on
X if X ∈ A and

A, B ∈ A =⇒ A \ B ∈ A,
[
A1 , A2 , · · · ∈ A =⇒ An ∈ A.
n∈N

If A is a σ-algebra on X, then the pair (X, A) is called a measurable space. A function f

on a measurable space (X, A) with values in R, i.e., f : X → F, is called measurable (or
A-measurable) if f −1 (U ) ∈ A for all open sets U ⊂ F.
Let A be a σ-algebra on X. A function µ : A → [0, ∞] with µ(∅) = 0 is called a
measure if it is countably additive, i.e., if for any sequence (An )n∈N of pairwise disjoint
elements of A,
[ X
µ( An ) = µ(An ).
n∈N n∈N

A measure is called finite if µ(X) < ∞. A signed measure is a countably additive

function µ : A → R such that µ(∅) = 0. A complex measure is a countably additive
function µ : A → C such that µ(∅) = 0.
Observe that not every measure is a signed measure, for example the Lebesgue meas-
ure is not a signed measure. But every finite measure is a signed measure.
Definition A.2. We say that a subset A of X is an (µ-)null set, if there exists a B ∈ A
with A ⊂ B and µ(B) = 0. We write Ac for the complement of A in X, so that Ac = X\A.
Definition A.3. Let p ∈ [1, ∞). Lp (µ) is the space of measurable functions f : X → F
for which
Z
|f (x)|p dµ(x) < ∞.

We say that two measurable functions f and g are equivalent, written f ∼ g if there
exists a null set A ∈ A such that f = g on Ac . We write Lp (µ) for the space that
consists of all equivalence classes in Lp (µ), in formula Lp (µ) = Lp (µ)/ ∼ or when we
define [f ]∼ = {g ∈ Lp : g ∼ f } for f ∈ Lp , then

Lp (µ) = {[f ]∼ : f ∈ Lp (µ)}.

208
We define
Z 1
p
p
kf kLp := |f (x)| dµ .

Similarly, we define L∞ to be the space of measurable functions f : X → F for which

there exists a null set A such that f is bounded on Ac . We define

kf kL∞ = inf{M > 0 : |f | ≤ M a.e. },

where we abbreviated “almost everywhere” by “a.e.”. Similarly as for p ∈ [1, ∞), we

define

L∞ (µ) = {[f ]∼ : f ∈ L∞ (µ)},

and write for f ∈ L∞ (µ) and g ∈ f (the following is independent of the choice of g)

kf kL∞ = kgkL∞ .

But from now on we ‘identify’ functions f with their equivalence class [f ]∼ , and so
use also consider elements of Lp as functions.
Theorem A.4 (Hölder’s inequality). [BCD11, Theorem 1.1] Let p, q, r ∈ [1, ∞] satisfy
1 1 1
+ = .
p q r
If f ∈ Lp (µ) and g ∈ Lq (µ), then f g ∈ Lr (µ) and

kf gkLr ≤ kf kLp kgkLq .

Theorem A.5 (Generalized Hölder inequality). Let n ∈ N and p1 , . . . , pn , r ∈ [1, ∞].

Suppose
1 1 1
+ ··· + = .
p1 pn r
For i ∈ {1, . . . , n} let fi ∈ Lpi (µ). Then f1 · · · fn ∈ Lr and

kf1 · · · fn kLr ≤ kf1 kLp1 · · · kfn kLpn .

Proof. Let q ∈ [1, ∞] be such that

1 1 1
= + ··· + .
q p1 pn−1
1 1 1
Let g = f1 · · · fn−1 . If g ∈ Lq , then by the Hölder inequality, as q + pn = r

kgfn kLr ≤ kgkLq kfn kLpn .

From this one can finish the proof by an induction argument.

209
Lemma A.6. We have Lp (µ) ⊂ L1 (µ) + L∞ (µ) for all p ∈ [1, ∞].

Proof. Let f ∈ Lp (µ). Then [|f | > 1] has finite measure. Define f1 := f 1[|f |≤1] and
f2 := f 1[|f |>1] . Then f1 ∈ L∞ (µ) and with Hölder’s inequality we have
kf2 kL1 ≤ kf kLp k1[|f |>1] kLq < ∞,
1 1
for q ≥ 1 such that p + q = 1.
Definition A.7. Let p ∈ [1, ∞] and I be a countable set. We write k · k`p (I) or k · k`p for
the function FI → [0, ∞] given by

 1
 p p
i∈I |x(i)| p < ∞,
 P
kxk`p (I) =
i∈I |x(i)| p = ∞, (x ∈ FI ).
sup


We write `p (I) for the set of x ∈ FI such that kxk`p (I) < ∞. In other words, `p (I) is the
space Lp (µ) in case µ is the counting measure on I.
Lemma A.8. Let p, r ∈ [1, ∞] and p ≤ r. Then
kxk`r ≤ kxk`p (x ∈ FI ).
In particular, `p ⊂ `r .

Proof. If r = ∞, this is immediate. For convenience, we may assume I = N. Let θ ≤ 1.

It suffices (take θ = pr ) to show that
∞
X θ ∞
X
|x(i)| ≤ |x(i)|θ (x ∈ FI ).
i=1 i=1

But this follows from the subadditivity of the function (0, ∞) → (0, ∞), t 7→ ta , see
Lemma 20.2.
Corollary A.9 (Hölder’s inequality for `p spaces). Let p, q ∈ [1, ∞] and r ∈ [1, ∞] be
such that
1 1 1
min{1, + } = .
p q r
If f ∈ `p and g ∈ `q , then f g ∈ `r with
kf gk`r ≤ kf k`p kgk`q .

Proof. Suppose that p1 + 1q > 1, in the other case we can apply Hölder’s inequality
immediately. Then both p and q are finite, and we can find p̃, q̃ with p ≤ p̃ < ∞,
q ≤ q̃ < ∞ such that
1 1
+ = 1,
p̃ q̃
Let f ∈ `p and g ∈ `q . Then f ∈ `p̃ and g ∈ `q̃ and kf gk`1 ≤ kf k`p̃ kgk`q̃ ≤ kf k`p kgk`q .

210
Theorem A.10 (Log-convexity of Lp norms). Let p, r be such that 1 ≤ p < r ≤ ∞.
Then Lp (µ) ∩ Lr (µ) ⊂ Lq (µ) for all q with p ≤ q ≤ r and with θ ∈ [0, 1] such that

1 θ 1−θ
= + .
q p r
we have
1−θ
kf kLq ≤ kf kθLp kf kLr (f ∈ Lp ∩ Lr ).

θq (1−θ)q
Proof. As 1 = p + r , we obtain by Hölder’s inequality,
Z
(1−θ)q
kf kqLq = |f |θq |f |(1−θ)q ≤ k|f |θq k p k|f |(1−θ)q k r = kf kθq
Lp kf kLr .
L θq L (1−θ)q

1 1
Lemma A.11 (Young’s inequality for products). For p, q > 0 with p + q = 1,

1 1
ab ≤ ap + bq (a, b ≥ 0).
p q

In an other formulation; if θ ∈ [0, 1] then aθ b1−θ ≤ θa + (1 − θ)b for all a, b ≥ 0.

Proof. As the exponential function is convex, we have for p, q as above and a, b ≥ 0,

1 1 1 1 1 1
 
ab = exp log ap + log bq ≤ exp (log ap ) + exp (log bq ) = ap + bq .
p q p q p q

Corollary A.12. Let p, q ∈ [1, ∞] be such that 1

p + 1q = 1. Then L1 (µ) ∩ L∞ (µ) ⊂ Lp (µ)
and
1 1
kf kLp ≤ kf kL1 + kf kL∞ (f ∈ L1 ∩ L∞ ).
p q

1 1 θ (1−θ)
Proof. Note that θ = p is such that p = 1 + ∞ . Apply Theorem A.10, to obtain
1 1
p q
kf kLp ≤ kf kL1 kf kL∞ . Then apply Lemma A.11.

A.13 (Notation). Let d ∈ N and Ω ⊂ Rd be open. We write B(Ω) for the Borel-σ-algebra
on Ω. Let λ be the Lebesgue measure on the measurable space (Ω, B λ (Ω)), where B λ (Ω)
is the completion of the Borel-σ-algebra on Ω, which consists of all Lebesgue measurable
sets. For p ∈ [1, ∞] we write Lp (Ω) instead of Lp (λ).

Lemma A.14. Let p ∈ [1, ∞). Then Cc (Rd ) is dense in Lp (Rd ).

211
B Taylor’s formula

B.1 For one dimension

Let us first recall the fundamental theorem of calculus.

Theorem B.1. [RS82, §15] Let g : [a, b] → R be continuous. Then
Z x
d
g(y) dy = g(x).
dx a

The following is a direct consequence.

Corollary B.2. If f ∈ C 1 [a, b], then
Z x
f (x) = f (a) + f 0 (y) dy.
a

B.3. If f ∈ C 2 , then we have

Z y
f 0 (y) = f 0 (a) + f 00 (z) dz,
a
and thus
Z x
f (x) = f (a) + f 0 (y) dy
Zax  Z y 
= f (a) + f 0 (a) + f 00 (z) dz dy
a a
Z xZ y
= f (a) + (x − a)f 0 (a) + f 00 (z) dz dy.
a a

This can be iterated:

For f ∈ C k [a, b], we have
k−1
X (x − a)i i k
f (x) = ∂ f (a) + Rf,a (x),
i=0
i!
where by Fubini
Z x Z y1 Z yk−1
k
Rf,a (x) = ··· ∂ k f (yk ) dyk dyk−1 · · · dy1
Za a a

= 1{y:a≤yk ≤yk−1 ≤···≤y1 ≤x} (y)∂ k f (yk ) dy.

[a,b]k
Z xZ xZ x Z x
= ··· dy1 dy2 · · · dyk−1 ∂ k f (yk ) dyk
a yk yk−1 y2

By induction one can easily see that

(x − yk )k−1
Z xZ x Z x
··· dy1 dy2 · · · dyk−1 = .
yk yk−1 y2 (k − 1)!
So we have obtained the following.

212
Theorem B.4. Let f ∈ C k [a, b], then
k−1
(x − a)i i (x − y)k−1 k
X Z x
f (x) = ∂ f (a) + ∂ f (y) dy
i=0
i! a (k − 1)!
k
(x − a)i i (x − y)k−1
X Z x
= ∂ f (a) + [∂ k f (y) − ∂ k f (a)] dy.
i=0
i! a (k − 1)!

Let
L = max |∂ k f (y)|
y∈[a,b]

M = max |∂ k f (y) − ∂ k f (a)|.

y∈[a,b]

Then
k−1
X (x − a)i i L
f (x) − ∂ f (a) ≤ (x − a)k ,
i=0
i! k!
k
X (x − a)i i M
f (x) − ∂ f (a) ≤ (x − a)k .
i=0
i! k!

B.2 Taylor expansion in higher dimensions

Definition B.5. Let f ∈ C k (U, Rp ) for U ⊂ Rd open. Let a ∈ U . The Taylor polynomial
k , is given by
of order k at the point a, written Tf,a
k
X 1 α
Tf,a (x) = ∂ f (a)(x − a)α .
α!
α∈Nd0 :|α|≤k

k (x) = f (x) − T k (x).

The remainder of order k at the point a is given by Rf,a f,a

Lemma B.6. [DK10, Lemma 6.1] Let f ∈ C k (U, Rd ). Then for l ∈ {0, 1, . . . , k} and
a, h ∈ Rd and t ∈ R such that a + th ∈ U we have
1 dj X hα α
f (a + th) = ∂ f (a + th)
j! dtj α!
α∈Nd0 :|α|=j

Theorem B.7 (Taylor’s Formula). [DK10, Theorem 6.2] Let f ∈ C k (U, Rp ) for U ⊂ Rd
being an open ball. Let a ∈ U . For all l ∈ {1, . . . , k} and x ∈ U
(x − a)α (1 − s)l−1 α
Z 1
l−1
X
f (x) = Tf,a (x) + ∂ f (a + s(x − a)) ds (B.1)
α! 0 (l − 1)!
α∈Nd0 :|α|=l

(x − a)α (1 − s)l−1
X Z 1
l
= Tf,a (x) + [∂ α f (a + s(x − a)) − ∂ α f (a)] ds.
α! 0 (l − 1)!
α∈Nd0 :|α|=l
(B.2)

213
For a, x ∈ U let us define

(x − a)α (1 − s)l−1
X Z 1
l
Rf,a (x) = (∂ α f (a + s(x − a)) − ∂ α f (a)) ds. (B.3)
α! 0 (l − 1)!
α∈Nd0 :|α|=l

The map U × U → R given by (a, x) 7→ Rf,al (x) is in C k−l , and for every compact K ⊂ U

and every ε > 0 there exists a δ > 0 such that

l
|Rf,a (x)| ≤ ε|x − a|l for x, a ∈ K and |x − a| < δ.
l
Moreover, for all a ∈ U the map Rf,a : U → R is in C k and ∂ α Rf,a
l (a) = 0 for all α ∈ Nd
0
with |α| ≤ l.

Proof. Let g be the one-dimensional function given by g(t) = f (a + t(x − a)). Then by
Theorem B.4
l−1 i i
(t − s)l−1 dl
Z t
X t d
g(t) = i
g(0) + g(s) ds.
i=0
i! dt 0 (l − 1)! dsl

So that with Lemma B.6 one obtains (B.1) and (B.2).

C Integration by parts

Theorem C.1. [Eva98, Appendix C.2, Theorem 2] Let Ω be a bounded open set with
C 1 boundary ∂Ω. We write σ for the d − 1 dimensional “surface” measure on ∂Ω. For
f, g ∈ C(Ω) which are differentiable on Ω, and i ∈ {1, . . . , d} we have
Z Z Z
f ∂i g = − g∂i f + f gni dσ,
U U ∂U

where n(x) for x ∈ ∂U is the outward pointing normal vector and ni its i-th component.

214
References
[AAR99] George E. Andrews, Richard Askey, and Ranjan Roy. Special functions,
volume 71 of Encyclopedia of Mathematics and its Applications. Cambridge Uni-
versity Press, Cambridge, 1999.

[AC] Romain Allez and Khalil Chouk. The continuous anderson hamiltonian in di-
mension two. Preprint available at https://arxiv.org/abs/1511.02718.

[BCD11] Hajer Bahouri, Jean-Yves Chemin, and Raphaël Danchin. Fourier analysis and
nonlinear partial differential equations, volume 343 of Grundlehren der Math-
ematischen Wissenschaften [Fundamental Principles of Mathematical Sciences].
Springer, Heidelberg, 2011.

[Bog07] V.I. Bogachev. Measure theory. Vol. I, II. Springer-Verlag, Berlin, 2007.

[Con90] John B. Conway. A Course in Functional Analysis, volume 96 of Graduate Texts

in Mathematics. Springer-Verlag, New York, second edition, 1990.

[Die76] J. Dieudonné. Treatise on analysis. Vol. II. Academic Press [Harcourt Brace
Jovanovich, Publishers], New York-London, 1976. Enlarged and corrected print-
ing, Translated by I. G. Macdonald, With a loose erratum, Pure and Applied
Mathematics, 10-II.

[DK10] J. J. Duistermaat and J. A. C. Kolk. Distributions. Cornerstones. Birkhäuser

Boston, Inc., Boston, MA, 2010. Theory and applications, Translated from the
Dutch by J. P. van Braam Houckgeest.

[Don69] William F. Donoghue, Jr. Distributions and Fourier transforms, volume 32 of

Pure and Applied Mathematics. Academic Press, New York, 1969.

[dPvR13] B. de Pagter and A.C.M. van Rooij. An Invitation to Functional Analysis.

Epsilon Uitgaven. Epsilon Uitgaven, 2013.

[Eva98] Lawrence C. Evans. Partial differential equations, volume 19 of Graduate Studies

in Mathematics. American Mathematical Society, Providence, RI, 1998.

[Fri75] Avner Friedman. Stochastic differential equations and applications. Vol. 1. Aca-
demic Press [Harcourt Brace Jovanovich, Publishers], New York-London, 1975.
Probability and Mathematical Statistics, Vol. 28.

[Fri98] F. G. Friedlander. Introduction to the theory of distributions. Cambridge Uni-

versity Press, Cambridge, second edition, 1998. With additional material by M.
Joshi.

[GIP15] Massimiliano Gubinelli, Peter Imkeller, and Nicolas Perkowski. Paracontrolled

distributions and singular PDEs. Forum Math. Pi, 3:e6, 75, 2015.

215
[Gra14] Loukas Grafakos. Modern Fourier analysis, volume 250 of Graduate Texts in
Mathematics. Springer, New York, third edition, 2014.

[H6̈0] Lars Hörmander. Estimates for translation invariant operators in Lp spaces. Acta
Math., 104:93–140, 1960.

[Hai14] M. Hairer. A theory of regularity structures. Invent. Math., 198(2):269–504,

2014.

[Hal74] Paul R. Halmos. Measure Theory. Springer, 1974.

[Har16] G. H. Hardy. Weierstrass’s non-differentiable function. Trans. Amer. Math. Soc.,

17(3):301–325, 1916. Pdf available at ams.org.

[Hor66] John Horváth. Topological vector spaces and distributions. Vol. I. Addison-
Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1966.

[HvNVW16] Tuomas Hytönen, Jan van Neerven, Mark Veraar, and Lutz Weis. Analysis
in Banach spaces. Vol. I. Martingales and Littlewood-Paley theory, volume 63 of
Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern
Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series.
A Series of Modern Surveys in Mathematics]. Springer, Cham, 2016.

[Joh95] Jon Johnsen. Pointwise multiplication of Besov and Triebel-Lizorkin spaces.

Math. Nachr., 175:85–133, 1995.

[Leo17] Giovanni Leoni. A first course in Sobolev spaces, volume 181 of Graduate Studies
in Mathematics. American Mathematical Society, Providence, RI, second edition,
2017.

[Mar18] Jörg Martin. Refinements of the Solution Theory for Singular SPDEs. PhD
thesis, Humboldt-Universität zu Berlin, 2018.

[Mic57] S. G. Michlin. Fourier integrals and multiple singular integrals. Vestnik Lenin-
grad. Univ. Ser. Mat. Meh. Astr., 12(7):143–155, 1957.

[MS13] Camil Muscalu and Wilhelm Schlag. Classical and multilinear harmonic analysis.
Vol. I, volume 137 of Cambridge Studies in Advanced Mathematics. Cambridge
University Press, Cambridge, 2013.

[Obe92] M. Oberguggenberger. Multiplication of distributions and applications to par-

tial differential equations, volume 259 of Pitman Research Notes in Mathematics
Series. Longman Scientific & Technical, Harlow; copublished in the United States
with John Wiley & Sons, Inc., New York, 1992.

[PT16] David J. Prömel and Mathias Trabs. Rough differential equations driven by
signals in Besov spaces. J. Differential Equations, 260(6):5202–5249, 2016.

[R88] A. C. M. van Rooij. Fouriertheorie. Epsilon Uitgaven, Utrecht, 1988.

216
[RS82] A. C. M. van Rooij and W. H. Schikhof. A Second Course on Real Functions.
Cambridge University Press, Cambridge-New York, 1982.

[RS96] Thomas Runst and Winfried Sickel. Sobolev spaces of fractional order, Nemytskij
operators, and nonlinear partial differential equations, volume 3 of De Gruyter
Series in Nonlinear Analysis and Applications. Walter de Gruyter & Co., Berlin,
1996.

[Rud91] Walter Rudin. Functional analysis. International Series in Pure and Applied
Mathematics. McGraw-Hill, Inc., New York, second edition, 1991.

[Saw18] Yoshihiro Sawano. Theory of Besov spaces, volume 56 of Developments in Math-

ematics. Springer, Singapore, 2018.

[Sch71] Helmut H. Schaefer. Topological Vector Spaces. Springer-Verlag, New York-

Berlin, 1971. Third printing corrected, Graduate Texts in Mathematics, Vol.
3.

[Sha66] Eliahu Shamir. A remark on Mikhlin-Hörmander multipliers theorem. J. Math.

Anal. Appl., 16:104–107, 1966.

[SS03] Elias M. Stein and Rami Shakarchi. Fourier analysis, volume 1 of Princeton
Lectures in Analysis. Princeton University Press, Princeton, NJ, 2003. An intro-
duction.

[ST87] Hans-Jürgen Schmeisser and Hans Triebel. Topics in Fourier analysis and
function spaces. A Wiley-Interscience Publication. John Wiley & Sons, Ltd.,
Chichester, 1987.

[Str03] Robert S. Strichartz. A guide to distribution theory and Fourier transforms.

World Scientific Publishing Co., Inc., River Edge, NJ, 2003. Reprint of the 1994
original [CRC, Boca Raton; MR1276724 (95f:42001)].

[Trè06] Francois Trèves. Topological vector spaces, distributions and kernels. Dover Pub-
lications, Inc., Mineola, NY, 2006. Unabridged republication of the 1967 original.

[Tri78] Hans Triebel. Interpolation theory, function spaces, differential operators,

volume 18 of North-Holland Mathematical Library. North-Holland Publishing
Co., Amsterdam-New York, 1978.

[Tri83] Hans Triebel. Theory of function spaces, volume 78 of Monographs in Mathem-

atics. Birkhäuser Verlag, Basel, 1983.

[Tri92] Hans Triebel. Theory of function spaces. II, volume 84 of Monographs in Math-
ematics. Birkhäuser Verlag, Basel, 1992.

217