Lecture notes

on
Distributions

Hasse Carlsson

2011
2
Preface

Two important methods in analysis is differentiation and Fourier trans-

formation. Unfortunally not all functions are differentiable or has a Fourier
transform. The theory of distribution tries to remedy this by imbedding
classical functions in a larger class of objects, the so called distributions (or
general functions). The basic idea is not to think of functions as pointwise
defined but rather as a ”mean value”. A locally integrable function f is
identified with the map Z
ϕ 7→ f ϕ,

where ϕ belongs to a space of ”nice” test functions, for instance C0∞ .

As an extension of this we let a distribution be a linear functional on the
space of test functions. When extending operations such as differentiation
and Fourier transformation, we do this by transfering the operations to the
test functions, where they are well defined.
Let us for instance see how to define the derivative of a locally integrable
function f on R. If f is continuously differentiable, an integration by parts
implies that Z Z
f ϕ = − f ϕ0 .

Now we use this formula to define the differential of f , when f is not classi-
cally differentiable. f 0 is the map
Z
ϕ 7→ − f ϕ0 .

In these lectures we will study how differential calculus and Fourier anal-
ysis can be extended to distributions and study some applications mainly in
the theory of partial differential equations.
The presentation is rather short and for a deeper study I recommend the
following books:
Laurent Schwartz. Théorie des Distributions I, II. Hermann, Paris, 1950–
51.
Lars Hörmander. The Analysis of Linear Partial Differential Operators
I, 2nd ed. Springer, Berlin, 1990.
Contents

1 A primer on C0∞ -functions 6

2 Definition of distributions 11

3 Operations on distributions 17

4 Finite parts 21

5 Fundamental solutions of the Laplace and heat equations 28

6 Distributions with compact support 31

7 Convergence of distributions 32

8 Convolution of distributions 36

9 Fundamental solutions 43

10 The Fourier transform 47

11 The Fourier transform on L2 55

12 The Fourier transform and convolutions 57

13 The Paley-Wiener theorem 63

14 Existence of fundamental solutions 66

15 Fundamental solutions of elliptic differential operators 68

16 Fourier series 70

4
17 Some applications 74
17.1 The central limit theorem . . . . . . . . . . . . . . . . . . . . 74
17.2 The mean value property for harmonic functions . . . . . . . . 75
17.3 The Heisenberg uncertainty principle . . . . . . . . . . . . . . 76
17.4 A primer on Sobolev inequalities . . . . . . . . . . . . . . . . 78
17.5 Minkowski’s theorem . . . . . . . . . . . . . . . . . . . . . . . 81

5
Chapter 1

A primer on C0∞-functions

When we shall extend differential calculus to distributions, it is suitable to

use infintely differentiable functions with compact support as test functions.
In this chapter we will show that there is ”a lot of” C0∞ -functions.

Notation
Let Ω be a domain in Rn . C k (Ω) denotes the k times comtinuously differ-
entiable functions on Ω. (k may be +∞.) C0k (Ω) are those functions in
C k (Ω) with compact support. We denote points in Rn with x = (x1 , . . . , xn )
and dx = dx1 . . . dxn denotes the Lebesgue measure. For a vector α =
(α1 , . . . αn ) ∈ Nn we let

|α| = α1 + . . . + αn , α! = α1 ! . . . αn !, xα = xα1 1 . . . xαnn

and
∂ αf ∂ α1 ∂ αn
∂ αf = = α . . . f.
∂xα ∂x1 1 ∂xαnn
Example 1.1. With these notations the Taylorpolynomial of f of degree N
can be written as
X ∂ α f (a)
xα .
α!
|α|≤N

As described in the preface, to a function f ∈ L1loc , we will associate the

map Λf , given by Z
ϕ 7→ f ϕ dx, ϕ ∈ C0∞ .
Rn

6
Problem. Does the map Λf determine f ?
More precisely, if f, g ∈ L1loc and
Z Z
f ϕ dx = gϕ dx, ϕ ∈ C0∞ ,
Rn Rn

does this imply that f = g a.e.? 2

To be able to solve this problem we need to construct functions ϕ ∈ C0∞ .

We start with

Example 1.2. There are functions f ∈ C ∞ (R) with f (x) = 0 when x ≤ 0

and f (x) > 0 when x > 0.

Remark 1.3. There is no such real analytic function. 2

Proof. Such a function must satisfy f (n) (0) = 0 for all n. Thus f (x) =
0(xn ), x → 0, for all n. Guided by this, we put

 e−1/x , x > 0
f (x) =
 0, x≤0.

We have to prove that f ∈ C ∞ . By induction we have


1
1
e− x , x > 0
 P

(n) n x
f (x) =
 0, x≤0

for some polynomials Pn . This is clear when x 6= 0. But at the origin we

have if h > 0,

f (n) (h) − f (n) (0)

 
1 1 −1
= Pn e h → 0, h → 0.
h h h

Example 1.4. There are non-trivial functions in C0∞ (Rn ).

Proof. Let f be the function in Example 2 and put ϕ(x) = f (1 − |x|2 ).

7
Approximate identities
Pick a function ϕ ∈ C0∞ (Rn ) with ϕ = 1 and
R
−n ∞ n
R ϕ ≥ 0. For δ > 0 we let
ϕδ (x) = δ ϕ(x/δ). Then ϕδ ∈ C0 (R ) and ϕδ = 1. {ϕδ ; δ > 0} is called
an approximate identity.

”ϕδ → δ”

Regularization by convolution
The convolution of two functions f and ϕ is defined by
Z
f ∗ ϕ(x) = f (x − y)ϕ(y)dy.
Rn

The convolution is defined for instance if f ∈ L1loc and ϕ ∈ C0∞ . Then

f ∗ ϕ = ϕ ∗ f, f ∗ ϕ ∈ C ∞ and ∂ α (f ∗ ϕ) = f ∗ ∂ α ϕ.
Exercise 1.1. Verify this.

Theorem 1.5.

a) If f ∈ C0 , then f ∗ ϕδ → f, δ → 0, uniformly.

b) If f is continuous in x, then f ∗ ϕδ (x) → f (x), δ → 0.

c) If f ∈ Lp , 1 ≤ p < +∞, f ∗ ϕδ → f i Lp (and a.e.).

Remark 1.6. a) implies that C0∞ (Ω) is dense in C0 (Ω) (in the supremum
norm).
Exercise 1.2. Verify this.

8
Proof.

a) Take R so that supp ϕ ⊂ {x; |x| ≤ R}. We have

Z
|f ∗ ϕδ (x) − f (x)| ≤ |f (x − y) − f (x)|ϕδ (y)dy
|y|≤δR
R
≤ uniform continuity ≤  Rn ϕδ (y)dy = , if δ is small enough.

b) Exercise 1.3.

c) Jensen’s inequality implies

Z p
p
|f ∗ ϕδ (x) − f (x)| ≤ |f (x − y) − f (x)|ϕδ (y)dy
Z Rn Z
p
≤ |f (x − y) − f (x)| ϕδ (y)dy = |f (x − δt) − f (x)|p ϕ(t)dt.
Rn Rn

Using Fubini’s theorem and the notation f δt (x) = f (x − δt), we get

Z Z
p
kf ∗ ϕδ − f kp ≤ ϕ(t)dt |f (x − δt) − f (x)|p dx
R n R n
Z
= kf δt − f kpp ϕ(t)dt → 0,
Rn

That the limit is zero follows by dominated convergence and that translation
is continuous on Lp . This in turn follows since C0 is dense in Lp , 1 ≤ p < +∞:
If g ∈ C0 , then
Z
δ p
kg − gkp = |g(x − δ) − g(x)|p dx → 0, δ → 0,
K

by dominated convergence. Now approximate f ∈ Lp with g ∈ C0 , kf −gkp <

. Minkowski’s inequality (the triangle inequality) implies

kf δ − f kp ≤ kf δ − g δ kp + kg δ − gkp + kg − f kp ≤ 2 + kg δ − gkp ≤ 3,

if δ is small enough.

Exercise 1.4. a) Let Br = {x; |x| < r}. Construct a function ψδ ∈ C0∞ (Rn ) such that
0 ≤ ψδ ≤ 1, ψδ = 1 on Br and supp ψδ ⊂ Br+δ . How big must kδ α ψδ k∞ be?
b) Let K ⊂ Ω where K is compact and Ω is open in Rn . Construct ψ ∈ C0∞ (Ω) with
ψ = 1 on a neighborhood of K and 0 ≤ ψ ≤ 1. How big must k∂ α ψk∞ be?

9
 Now we are able to answer yes to the problem on page 7.

Theorem 1.7. A locally integrable function that is zero as a distribution is

zero a.e.

f ϕ = 0 for all ϕ ∈ C0∞ . According to Theorem 1 a),

R
Proof. We R assume that
we have f Φ = 0 for all Φ ∈ C0 , and thus f = 0 a.e. (for instance by the
Riesz representation theorem.)
Alternatively we can argue as follows: Take ψn ∈ C0∞ with ψn (x) = 1
when |x| ≤ n. Then f ψn ∈ L1 and
Z
f ψn ∗ ϕδ (x) = f (y)ψn (y)ϕδ (x − y)dy = 0,
Rn

since y 7→ ψn (y)ϕδ (x − y) is C0∞ . But f ψn ∗ ϕδ → f ψn in L1 according to

Theorem 1 c). Hence f ψn = 0 a.e., and thus f = 0 a.e.

10
Chapter 2

Definition of distributions

Definition 2.1. Let Ω be an open domain in Rn . A distribution u in Ω is

a linear functional on C0∞ (Ω), such that for every compact set K ⊂ Ω there
are constants C and k such that
X
|u(ϕ)| ≤ C k∂ α ϕk∞ , (2.1)
|α|≤k

for all ϕ ∈ C0∞ with supp ϕ ⊂ K. 2

We denote the distributions on Ω by D 0 (Ω). If the same k can be used
for all K, we say that u has order ≤ k. These distributions are denoted
Dk0 (Ω). The smallest k that can be used is called the order of the distribution.
DF0 = ∪k Dk0 are the distributions of finite order.
Example 2.2.
(a) A function f ∈ L1loc is a distribution of order 0.
(b) A measure is a distribution of order 0.
(c) u(ϕ) = ∂ α ϕ(x0 ) defines a distribution of order |α|.
(d) Let xj be a sequence without limit point in Ω and let
X
u(ϕ) = ∂ αj ϕ(xj ).

Then u is a distribution. u has finite order if and only if

sup |αj | < ∞ and then the order is sup |αj |. 2
We will use the notation D(Ω) to denote the set C0∞ (Ω), in particular
when we consider D(Ω) with a topology that corresponds to the the following
convergence of test functions.

11
Definition 2.3. ϕj → 0 in D(Ω) if, for all j, supp ϕj are contained in a fix
compact set and k∂ α ϕj k∞ → 0, j → ∞, for all α. 2

Theorem 2.4. A linear functional u on D(Ω) is a distribution if and only

if u(ϕj ) → 0 when ϕj → 0 in D(Ω).

Proof. ⇒): Trivial.

⇐): Assume that (1) doesn’t hold. We have to prove that u(ϕj ) 6→ 0,
although ϕj → 0 in D(Ω). That (1) doesn’t hold implies that there is a
compact set K and a function ϕj ∈ D(Ω), with ϕj ⊂ K, u(ϕj ) = 1 and
X
|u(ϕj )| > j k∂ α ϕj k∞ .
|α|≤j

This implies k∂ α ϕj k∞ ≤ 1
j
if j ≥ |α|. Thus ϕj → 0 in D(Ω).

Theorem 2.5. A distribution u ∈ Dk0 (Ω) can uniquely be extended to a linear

functional on C0k (Ω). For every compact set K ⊂ Ω there is a constant
C = CK such that X
|u(ϕ)| ≤ C k∂ α ϕk∞ , (2.2)
|α|≤k

for all ϕ ∈ C0k (Ω) with support in K.

Corollary 2.6. Measures and distributions of order 0 coincides.

Proof of Theorem 5. Let ϕ be a fix function in C0k (Ω). Let Φδ ∈ C0∞ be

an approximate identity and put ϕn = ϕ ∗ Φ 1 , n ≥ N . Then all ϕn are
n
supported in a fix compact set K in Ω and if |α| ≤ k then

k∂ α (ϕ − ϕn )k∞ = k∂ α ϕ − (∂ α ϕ) ∗ Φ 1 k∞ → 0, n → ∞ . (2.3)
n

Hence, if u has an extension satisfying (2), then u(ϕ) = limn→∞ u(ϕn ). This
proves the uniqueness of the extension and makes it natural to define

u(ϕ) = lim u(ϕn ) .

n→∞

The limit exists since u(ϕn ) is a Cauchy sequence:

X
|u(ϕn ) − u(ϕm )| = |u(ϕn − ϕm )| ≤ C k∂ α (ϕn − ϕm )k → 0,
|α|≤k

as n, m → ∞.
It is easy to see, by taking limits in (1), that u satisfies (2).

12
Exercise 2.1. Verify this.

Theorem 2.7. A positive distribution is a positive measure.

Definition 2.8. A distribution u is positive if ϕ ≥ 0 implies u(ϕ) ≥ 0 2

Proof. By Corollary 6 it is enough to show that u ∈ D00 .

Assume first that ϕ is real valued. Let K ⊂⊂ Ω and take χ ∈ C0∞ (Ω), 0 ≤
χ ≤ 1 with χ = 1 on K. If supp ϕ ⊂ K, then χkϕk∞ ± ϕ ≥ 0. Hence
u(χkϕk∞ ± ϕ) ≥ 0, or

|u(ϕ)| ≤ u(χkϕk∞ ) = u(χ)kϕk∞ .

So (1) holds with k = 0, C = u(χ).

If ϕ = f + ig is complex valued,we get

|u(ϕ)| ≤ |u(f )| + |u(g)| ≤ u(χ)(kf k∞ + kgk∞ ) ≤ 2u(χ)kϕk∞ .

Theorem 2.9. A distribution is determined by its local behavior.

More precisely: Assume that Ω = ∪Ωi and that ui ∈ D 0 (Ωi ). Furthermore
we assume that ui = uj on Ωi ∩Ωj , i.e. if ϕ ∈ C0∞ (Ωi ∩Ωj ) then ui (ϕ) = uj (ϕ).
Then there is a unique distribution u on Ω with u = ui on Ωi .

To prove this we need a C0∞ partition of unity.

∞
Proposition 2.10. Let K ⊂ ∪N
1 Ωi . Then there are ϕi ∈ C0 (Ωi ), 0 ≤ ϕi ≤ 1
and Σϕi = 1 on K.

Proof of Theorem 9. Assume that u = ui on Ωi . Let supp P ϕ = K and ϕi be

a partition av unity as above. By linearity, since ϕ = i ϕϕi ,
X X
u(ϕ) = u(ϕϕi ) = ui (ϕϕi ) (2.4)
i i

This shows the uniqueness.

To prove the existence, we need to show that (4) gives a well
P defined distri-
butionPu. But if ϕ̃k isP
another
P partition of P unity,
P then ϕ̃k = i ϕ Pi ϕ̃k on K and
thus k uk (ϕϕ̃k ) = k i uk (ϕϕ̃k ϕi ) = i k ui (ϕϕ̃k ϕi ) = i ui (ϕϕi ), so
(4) defines u uniquely.
It is easy to show that u satisfies (1), and the theorem is proved.

Exercise 2.2. Do it!

13
Proof of Proposition 10. We shall show the following
Claim. There are open sets Vi with V i ⊂ Ωi and K ⊂ ∪N
1 Vi .
∞
Assuming this take ϕ̃i ∈ C0 (Ωi ), 0 ≤ ϕ̃i ≤ 1 with ϕ̃i = 1 on V i .
Then Σϕ̃i > 0 on a neighborhood U of K. Take χ with χ = 1 on K and
supp χ ⊂ U . Put
ϕ̃i
ϕi = χ .
Σϕ̃i
It is clear that ϕi satisfy the conditions in the proposition.
To prove the claim, take to S x ∈ K a neighborhood Vx with x S ∈ Vx ⊂
N
V x ⊂ Ωj for [some j. Then K ⊂ V x . By compactness we get K ⊂ 1 Vxk .
Let Vi = V xk .
Vxk ⊂Ωi

The support of a distribution

R
If f ∈ C then supp f = {x; f (x) 6= 0}. This implies that f ϕ = 0 for all
ϕ ∈ C0∞ whos support doesn’t intersect the support of f .

Definition 2.11. If u ∈ D 0 (Ω) then supp u = {x ∈ Ω; There is no neighbor-

hood of x with u = 0 in this neighborhood.}

Exercise 2.3. Show that supp u is closed.

Theorem 2.12. If supp u ∩ supp ϕ = ∅, then u(ϕ) = 0.

Proof. This follows directly from Theorem 9, since u = 0 locally on

Ω \ supp u.

An important extension of Theorem 12 is the following theorem and its

corollary.

Theorem 2.13. Assume that u ∈ Dk0 (Ω) and ϕ ∈ C0k (Ω) with ∂ α ϕ(x) = 0 if
|α| ≤ k and x ∈ supp u. Then u(ϕ) = 0.

Corollary 2.14. If u ∈ D 0 (Ω) and supp u = {x0 } ⊂ Ω, then u is of the

form X
u(ϕ) = aα ∂ α ϕ(x0 ).
|α|≤k

Proof of Theorem 13. Let K = supp u ∩ supp ϕ. If K = ∅, the result follow

from Theorem 5. But K can be non empty. Then, let K = {x; d(x, K) < }

14
and take χ ∈ C0∞ (K ) with χ = 1 in a neighborhood of K. Then, by
Theorem 5,
u(ϕ) = u(χ ϕ + (1 − χ )ϕ) = u(χ ϕ).
If k = 0 this implies
|u(ϕ)| ≤ Ckχ ϕk∞ → 0,  → 0.
If k > 0 we get
X X
|u(ϕ)| ≤ C k∂ α (χ ϕ)k∞ ≤ C k∂ α χ ∂ β ϕk∞ .
|α|≤k |α|+|β|≤k

We can choose χ such that k∂ α χ k∞ ≤ C−|α| . To estimate k∂ β ϕk∞ we

consider the Taylor expansion of ϕ at a point x ∈ K. Let y ∈ K and take
x ∈ K with |x − y| ≤ . Put
g(t) = ∂ β ϕ(x + t(y − x)).
By the Taylorexpansion of g at t = 0 of order k − |β| − 1, we get
 X g (i) (0) 
|∂ β ϕ(y)| = |g(1)| =  + R(y).
 
i!
i≤k−|β|−1

Now g (i) (0) = 0 and

X
|R(y)| ≤ C sup |∂ k−|β| g(s)| ≤ Ck−|β| k∂ β ϕkK .
0≤s≤1
|β|=k

This implies
X X
|u(ϕ)| ≤ C k−|α|−|β| k∂ β ϕkK → 0,  → 0.
|α|+|β|≤k |β|=k

Proof of the corollary. u is of finite order k for some k. Fix χ ∈ C0∞ (Ω) with
χ = 1 near x0 and put
X ∂ α ϕ(x0 )
ψ(x) = ϕ(x) − χ(x) (x − x0 )α .
α!
|α|≤k

Then ∂ α ψ(x0 ) = 0 if |α| ≤ k. By Theorem 13, u(ψ) = 0 or

X  (x − x )α  X
0
u(ϕ) = ∂ α ϕ(x0 )u χ(x) = aα ∂ α ϕ(x0 ).
α!
|α|≤k |α|≤k

15
Exercise 2.4. H 2.2
Exercise 2.5. H 3.1.7.
P∞
Exercise 2.6. Show that u(ϕ) = 1 nα (ϕ( n1 ) − ϕ(− n1 )) is a distribution of order ≤ 1 if
α < 0. Also show that supp u = {0, ±1, ± 21 , ± 13 , . . .}, but if K is a closed set with

k
X
|u(ϕ)| ≤ C sup |∂ i ϕ|, ϕ ∈ C0∞ (R) ,
i=0 K

then either α < −1 or else K contains a neighborhood of the origin. (In particular
we can not choose K = supp u.)
Exercise 2.7. Assume that u ∈ Dk0 (R) and supp u ⊂ I where I is a compact interval.
Show that X
|u(ϕ)| ≤ C sup |∂ α ϕ|, ϕ ∈ C0∞ (R).
I
|α|≤k

(Hint. Theorem 13.)

Exercise 2.8. Is there a linear functional u on C0∞ that isn’t a distribution?

16
Chapter 3

Operations on distributions

The derivative of distributions

If u is a continuously differentiable function in Rn , an integration by parts
gives Z Z
∂k u · ϕ dx = − u · ∂k ϕ dx, ϕ ∈ D,
Rn Rn

as ϕ has compact support. This motivates the following definition.

Definition 3.1. If u ∈ D 0 (Ω), we define ∂k u ∈ D 0 (Ω) by

∂k u(ϕ) = −u(∂k ϕ).

That ∂k u defines a distribution follows since

X X
|∂k u(ϕ)| = |u(∂k ϕ)| ≤ C k∂ α (∂k ϕ)k∞ ≤ C k∂ α ϕk∞ .
|α|≤k |α|≤k+1

If u ∈ C 1 the distribution derivative coincides with the classsical derivative.

Example 3.2. Let the Heavisidefunktionen H be defined by


 1, x ≥ 0
H(x) =
 0, x < 0.

Then Z ∞
0 0
H (ϕ) = −H(ϕ ) = − ϕ0 (x)dx = ϕ(0).
0

17
 The Dirac measure at x0 ∈ Rn is given by δx0 (ϕ) = ϕ(x0 ). So we have
showed that H 0 = δ0 . The derivatives of the Dirac measure are given by
∂ α δx0 (ϕ) = (−1)|α| δx0 (∂ α ϕ) = (−1)|α| ∂ α ϕ(x0 ). With this notation, by Corol-
lary 2.14, a distribution supported at x0 can be written as
X
u= Cα δx(α)
0
.
|α|≤k

A generalization of Exempel 2 is given by

Proposition 3.3. Let u be a function in Ω ⊂ R, which is continuously
differentiable for x 6= x0 . Assume that the derivative v is integrable near x0 .
Then
u0 = v + (u(x0 + 0) − u(x0 − 0))δx0 .
Proof. We start by showing that the limits exist. Let x0 < x < y. Then
Z y
u(x) = u(y) − v(t)dt.
x

Since v is integrable we obtain as x ↓ x0

Z y
u(x0 + 0) = u(y) − v(t)dt.
x0

By the same argument also u(x0 − 0) exists. We get

Z Z
0 0 0
u (ϕ) = −u(ϕ ) = − uϕ dx = lim − u(x)ϕ0 (x)dx
R →0 |x−x0 |>
n h i∞ h ix0 − Z o
= lim − u(x)ϕ(x) − u(x)ϕ(x) + v(x)ϕ(x)dx
→0 x0 + −∞ |x−x0 |>
Z
= (u(x0 + 0) − u(x0 − 0))ϕ(x0 ) + v(x)ϕ(x)dx.
R

Theorem 3.4. Let u be a distribution on an interval I ⊂ R. If u0 = 0, then

u is constant.
Proof. That u0 = 0 as a distribution means that u0 (ϕ) = 0 or u(ϕ0 ) = 0 for
all ϕ ∈ D. To compute u(φ), we 0
R want to decide if φ = ψR for some ψ ∈ D.
x
This is the case exactly when φ = 0 and then ψ(x) = −∞ φ(t)dt. Thus,
R
if φ = 0, then u(φ) = 0. We shall reduce the general case to this special
case. Fix ψ0 ∈ D with ψ0 = 1. Put φ̃ = φ − ψ0 φ. Then φ̃ = 0 so
R R R
R R
0 = u(φ̃) = u(φ) − u(ψ0 ) φ or u(φ) = u(ψ0 ) φ. Thus u is the constant
u(ψ0 ).

18
Multiplication by functions
D(Ω) is a linear space, since we can add distributions and multiply a distri-
bution with a scalar in a natural way. We also want to multiply a distribution
with a function f . If u is a locally integrable function, then
Z Z
f u(ϕ) = (f u)ϕ dx = u(f ϕ) dx = u(f ϕ) .
Rn Rn

To be able to use this to define f u when u is a distribution we need that

f ϕ ∈ C ∞.
Definition 3.5. If f ∈ C ∞ we define f u by

f u(ϕ) = u(f ϕ).

Exercise 3.1. Show that f u ∈ D 0 (Ω).

Remark 3.6. If u is of order k, it is enough to demand that f ∈ C k . 2

Proposition 3.7.
(a) ∂j ∂k u = ∂k ∂j u
(b) ∂k (f u) = (∂k f )u + f ∂k u.
Exercise 3.2. Prove Proposition 7.

Remark 3.8. By (a), the distributional derivatives commutes and we can

use the notation ∂ α u, ∂ α u(ϕ) = (−1)|α| u(∂ α ϕ) where α is a multiindex. 2
Theorem 3.9. If u ∈ D 0 (Ω), Ω ⊂ R, and u0 + au = f where f ∈ C and
a ∈ C ∞ , Then u ∈ C 1 and the equation is holds classically.
Proof. Assume first that a ≡ 0. Let F be a (classsical) primitive function
of f . Then F ∈ C 1 and (u − F )0 = u0 − F 0 = f − f = 0 as a distribution.
Theorem 1 implies that u = F + C, and thus u ∈ C 1 and u0 = F 0 = f
classically.
If a 6≡ 0, we multiply the equation with its integrating factor. Let A be a
primitive function of a. Then A and eA are C ∞ functions. Furthermore, we
have
(eA u)0 = eA u0 + eA au = eA (u0 + au)
in the distributional sense. Therefore, the equation is equivalent to

(eA u)0 = eA f,

and we can use the case a ≡ 0.

19
Exercise 3.3. H 3.1.1
Exercise 3.4. H 3.1.5
Exercise 3.5. H 3.1.14
Exercise 3.6. H 3.1.21
Exercise 3.7. H 3.1.22.
Exercise 3.8. Assume that u ∈ D 0 (Ω), Ω ⊂ R, satisfies u(m) + am−1 u(m−1) + . . . + a0 u =
f , where f ∈ C and aj ∈ C ∞ . Show that u ∈ C m , and that the equation holds
classically.

20
Chapter 4

Finite parts

In this chapter we will extend Proposition 3.3 to the case where the derivative
is not locally integrable.
 
d 1
Example 4.1. What is √ ?
dx x+
We have
1 1
h( √ )0 , ϕi = −h √ , ϕ0 i =
x+ x+
Z ∞ i∞ 1 Z ∞ ϕ(x) 
1 0  h 1
= lim − √ ϕ (x)dx = lim − √ ϕ(x) − dx
→0  x →0 x  2  x3/2
1 Z ∞ 1 2ϕ(0) 
= − lim ϕ(x)dx − √
2 →0  x3/2 
2
Definition 4.2.
Z ∞ 
1 ϕ(x) 2ϕ(0)
hfp 3/2
, ϕi = lim 3/2
dx − √ .
x+ →0  x 
2
Thus we have shown that
d 1 1 1
( √ ) = − fp 3/2 .
dx x+ 2 x+

A version of the definition that is easier to remember is

Z ∞
1 ϕ(x) − ϕ(0)
hfp 3/2 , ϕi = dx.
x+ 0 x3/2

21
 1
Example 4.3. We define fp by
|x|5/2
ϕ(x) − ϕ(0) − xϕ0 (0)
Z
1
hfp 5/2 , ϕi = dx .
|x| R |x|5/2
1
The order of fp is 2. To show this we split the integral into two pieces,
|x|5/2
ϕ(x) − ϕ(0) − xϕ0 (0)
Z Z
1
hfp 5/2 , ϕi = + dx = I+II.
|x| |x|≤1 |x|>1 |x|5/2
To estimate the first integral we use that
1
|ϕ(x) − ϕ(0) − xϕ0 (0)| ≤ x2 kϕ00 k∞ .
2
This implies Z
1 1
|I| ≤ kϕ00 k∞ 1/2
dx ≤ Ckϕ00 k∞ .
2 |x|≤1 |x|
For the second integral we have
2kϕk∞ + |x|kϕ0 k∞
Z
|II| ≤ 5/2
dx ≤ C(kϕk∞ + kϕ0 k∞ ) .
|x|>1 |x|
Thus the order is at most two.
To show that the order can not be smaller, we let ϕ ∈ C0∞ , 0 ≤ ϕ ≤ 1,
supp ϕ ⊂ (0, 3) and ϕ = 1 on [1, 2] and put ϕ (x) = ϕ(x/). Then
Z Z 2
1 ϕ (x) 1 1
|hfp 5/2 , ϕ i| = 5/2
dx ≥ 5/2
dx ≥ c 3/2 .
|x| R x  x 
Furthermore kϕ k∞ + kϕ0 k∞ ≤ C/. Thus if the order were less than 2, we
1
would have c/3/2 ≤ |hfp 5/2 , ϕi| ≤ C/, a contradiction.
|x|
1 1
Since the order of fp 5/2 is 2 and |x|5/2 ∈ C 2 , |x|5/2 fp 5/2 is well
|x| |x|
defined and
1 1
h|x|5/2 fp 5/2 , ϕi = hfp 5/2 , |x|5/2 ϕi
|x| |x|
5/2
|x| ϕ(x)
Z Z
= dx = ϕ(x)dx = h1, ϕi.
R |x|5/2 R

Here we have used that |x|5/2 ϕ(x) and its derivative vanishes at x = 0.
Thus fp |x|15/2 solves the division problem |x|5/2 u = 1. 2

22
 1 3 1
Exercise 4.1. Show that (fp 3/2
)0 = − fp 5/2 .
x+ 2 x+

The above examples can be generalized to to define fp x−a+ , fp |x|

−a
and
−a
(for certain a) fp x etc. when a is not an integer, for instance
Z ∞
1 ϕ(x) − P (x)
hfp a , ϕi = dx,
|x| −∞ |x|a

where P is the Taylorpolynomial of ϕ at the origin of order [a] - 1. Then

 
1 0 sgn x 1 1
(fp a ) = −a fp a+1 = a fp a+1 − fp a+1
|x| |x| x+ x−
och
1
|x|a fp = 1.
|x|a
1
Another important property of fp a , a 6= −1, −2, . . . , is that it is ho-
|x|
mogeneous of degree −a. As we shall see later this fact simplifies the com-
putation of its Fourier transform.
1
To show that fp a is homogeneous we first must define what this means.
|x|
If u(x) is a function on Rn , u is homogeneous of degree α if u(tx) = tα u(x), t >
0. This can be reformulated in a way that is meaningful for distributions.
For a function u, we have
Z Z
1 y
hu(tx), ϕi = u(tx)ϕ(x)dx = [y = tx] = u(y) n ϕ( )dy = hu, ϕt i.
R R t t
But if u is homogeneous of degree α, we also have
Z Z
hu(tx), ϕi = u(tx)ϕ(x)dx = tα u(x)ϕ(x)dx = tα hu, ϕi.
R R

Therefore we make the following definition.

Definition 4.4. u ∈ D 0 (Rn ) is homogeneous of degree α if

hu, ϕt i = tα hu, ϕi, t > 0.

2
1 1
Proposition 4.5. fp a
och fp a are homogeneous of degree −a if a 6=
|x| x+
1, 2, 3, . . .

23
 5 1 1
Proof when a = . We have ϕt (0) = ϕ(0) and ϕ0t (0) = 2 ϕ0 (0). Thus
2 t t
Z
1 1  1 x 1 x 
hfp 5/2 , ϕt i = 5/2 t
ϕ( ) − ϕ(0) − ϕ(0) dx
|x| R |x| t t t2
Z
h xi 1 0 1 1
= y= = 5/2 |x|5/2
(ϕ(x) − ϕ(0) − xϕ (0))dx = hfp , ϕi.
t R t t5/2 |x|5/2

Example 4.6. Compute (log |x|)0 .

We have
Z
h(log |x|) , ϕ)i = −hlog |x|, ϕ i = − ϕ0 (x) log |x|dx
0 0

Z R
n
= − lim ϕ (x) log |x|dx = lim − [ϕ(x) log |x|]−
0
−∞
→0 |x|> →0
Z
∞ dx o
+ [ϕ(x) log |x|] − ϕ(x)
|x|> x
nZ dx o
= lim ϕ(x) + (ϕ() − ϕ(−)) log 
→0 |x|> x
Z Z
dx ϕ(x)
= lim ϕ(x) = pv dx ,
→0 |x|> x R x
where the last equality is a definition. 2

Z
1 ϕ(x)
Definition 4.7. hpv , ϕi = lim dx.
x →0 |x|> x
If we instead differentiate log x+ , we get
Z ∞
hlog x+ , ϕi = lim − ϕ0 (x) log xdx =
→0 
 Z ∞ 
∞ ϕ(x)
= lim −[ϕ(x) log x] + dx =
→0  x
Z ∞  Z ∞ Z 1 
ϕ(x) ϕ(x) ϕ(0)
= lim dx + ϕ(0) log  = lim dx − dx =
→0  x →0 0 x  x
Z ∞
ϕ(x) − χ(x)ϕ(0)
= dx,
0 x
where χ = χ[−1,1] , as in the rest of this chapter.
Thus with the following

24
 ∞
ϕ(x) − χ(x)ϕ(0)
Z
1
Definition 4.8. hfp , ϕi = dx 2
x+ 0 x
1
we have proved that (log x+ )0 = fp .
x+
1
Exercise 4.2. Show that fp solves the division problem xu = H.
x+

The above examples can be generalized to the following

Definition 4.9.
xn−1
1
Z ∞ ϕ(x) − P (x) − (n−1)!
ϕ(n−1) (0)χ(x)
hfp n , ϕi = dx
|x| −∞ |x|n

if n = 1, 2, 3, . . . and P is the Taylorpolynomial of ϕ of degree n − 2. 2

1 0 1 1
Example 4.10. (fp 2
) = −2fp 3 + δ (2) . 2
x+ x+ 2
Proof.
Z ∞ 0
1 0 1 0 ϕ (x) − ϕ0 (0) − xϕ00 (0)χ(x)
h(fp 2 ) , ϕi = −hfp 2 , ϕ i = − dx =
x+ x+ 0 x2
n Z ∞ ϕ0 (x) − ϕ0 (0) Z 1
xϕ00 (0) o
− lim dx − dx .
→0  x2  x2
As ϕ(x)−ϕ(0)−xϕ0 (0) is a primitive function of ϕ0 (x)−ϕ0 (0), an integration
by parts in the first integral implies that
1 nh ϕ(x) − ϕ(0) − xϕ0 (0) i∞
h(fp 2 )0 , ϕi = − lim +
x+ →0 x2 
Z ∞ Z 1
ϕ(x) − ϕ(0) − xϕ0 (0) xϕ00 (0) o
2 dx − dx .
 x3  x2
Now h ϕ(x) − ϕ(0) − xϕ0 (0) i∞ 1
lim = − ϕ00 (0)
→0 x2  2
and
n Z ∞ ϕ(x) − ϕ(0) − xϕ0 (0) Z 1
xϕ00 (0) o
lim 2 dx − dx =
→0  x3  x2
Z ∞
ϕ(x) − ϕ(0) − xϕ0 (0) − 21 x2 ϕ00 (0)χ(x)
2 dx .
0 x3

25
 Hence
∞
ϕ(x) − ϕ(0) − xϕ0 (0) − 12 x2 ϕ00 (0)χ(x)
Z
1 1 00
h(fp 2 )0 , ϕi = ϕ (0) − 2 dx
x+ 2 0 x3
1 1 (2)
= h−2fp 3 + δ , ϕi.
x+ 2

1
Example 4.11. fp is not homogeneous of degree −3 since
|x|3
2
ϕ( t ) − 1t ϕ(0) − tx2 ϕ0 (0) − 21 xt3 ϕ00 (0)χ(x)
Z 1 x
1 t
hfp 3 , ϕt i = dx
|x| R |x|3
ϕ(x) − ϕ(0) − xϕ0 (0) − 12 x2 ϕ00 (0)χ(xt)
Z
h xi 1
= y= = 3 dx
t t R |x|3
Z
1 1 1 00 dx
= (If we assume that t > 1) = 3 hfp 3 , ϕi + 3 ϕ (0)
t |x| 2t 1
t
<|x|<1 |x|
1 1 00 log t
= hfp , ϕi + ϕ (0) .
t3 |x|3 t3
2
Exercise 4.3. What happens if t < 1?
Exercise 4.4. Is fp x13 homogeneous of degree −3?
N
X −1
Exercise 4.5. Show that the equation xN u = 0 has the solution u = cn δ (n) .
n=0

Since xN fp x1N = 1, Exercise 4.5 implies that the equation

N −1
N 1 X
x u = 1 has the general solution u = fp N + cn δ (n) .
x n=0

In the same way the equation

N −1
N 1 X
(x − a) u = 1 has the solution u = fp + cn δa(n) .
(x − a)N n=0

1 1
where fp N
is defined in the same way as fp N .
(x − a) x

26
 Now we can solve the division problem P u = 1, where P is a polynomial
of one variable. In a neighborhood where P 6= 0, u = 1/P is a nice function.
So the the only problem is to understand 1/P near a real zero a of P . But
there we have P (x) = (x − a)n Q(x) where Q(a) 6= 0. Hence, near x = a,
1
we have (x − a)n Q(x)u = 1. This is satisfied if Qu = fp . Hence
(x − a)n
1 1
u= fp solves P u = 1 near x = a. By Theorem 2.9, u is a well
Q(x) (x − a)n
defined distribution on R that solves P u = 1.
Exercise 4.6. H 3.1.14
Exercise 4.7. H 3.1.20
Exercise 4.8. Let u be a continuous function on Rn \ {0} that is homogeneous of degree
−n. Show that we can define a distribution pv u by
Z
hpv u, ϕi = lim u(x)ϕ(x)dx,
→0 |x|>
R
if and only if |x|=1
u(x)dσ(x) = 0.
+ is by analytic continuation. If ϕ ∈ D
Exercise 4.9. An alternative method to define fp xα
and Re α > −1 the map Z ∞
Fϕ (α) = xα ϕ(x)dx
0
is analytic. Show that this map can be continued to a meromophic function in C,
whose only singularities are simple poles in −1, −2, −3, . . .. Compute the residues
R−k of Fϕ and show that if we for k = 1, 2, 3, . . . extend the definition of Fϕ by
 
R−k
Fϕ (−k) = lim Fϕ (α) − ,
α→−k α+k

we have
Fϕ (α) = hfp xα
+ , ϕi.

This approach gives an alternative proof that

−α
xα
+ fp x+ = H, α∈C

and
0 α−1
(fp xα
+ ) = αfp x+ , α 6= 0, −1, −2, −3, . . .

27
Chapter 5

Fundamental solutions of the

Laplace and heat equations

Definition 5.1. Let P (D) be a differential operator. A distribution E with

P (D)E = δ, is called a fundamental solution of P .
2

In Example 3.2, we saw that the Heaviside function H is a fundamental

solution of d/dx. A little more general, H(x1 ) . . . H(xn ) is a fundamental
solution of ∂1 . . . ∂n .
In this chapter we will treat the Laplace operator
n
X ∂ 2u
∆u = ,
i=1
∂x2i

and the heat operator

n
∂u X ∂ 2 u
 
∂
− ∆x u = − 2
.
∂t ∂t i=1
∂x i

To accomplish this we need to be able to integrate by parts in Rn , and we

remind the reader about

Green’s identity
Z Z  
∂v ∂u
(u∆v − v∆u) dx = u −v dσ
Ω ∂Ω ∂n ∂n
where ∂/∂n is the exterior normal derivative.

28
Theorem 5.2.
1


 log |x|, n = 2,
E(x) = 2π
1
 −
 , n ≥ 3,
ωn (n − 2)|x|n−2
is a fundamental solution of the Laplace operator in Rn .
(ωn is the surface measure of the unit sphere in Rn .)
Exercise 5.1. Compute ωn in terms of the Γ function,
Z ∞
Γ(s) = ts−1 e−t dt, Res > −1.
0

Exercise 5.2. Visa att ∆E(x) = 0 om x 6= 0.

Proof.
Z
h∆E, ϕi = hE, ∆ϕi = lim E∆ϕdx
→0 |x|>
Z
= Exercise 2 = lim (E∆ϕ − ϕ∆E)dx = Green’s identity =
→0 |x|>
Z  
∂ϕ ∂E
= lim E −ϕ dσ = lim(I + II ).
→0 |x|= ∂n ∂n →0

We only consider the case n ≥ 3, and leave the case n = 2 as

Exercise 5.3.
We have
∂ϕ 1 n−1
|I | ≤ C ∂n n−2 ωn 
−→ 0,  → 0 ,
∞
and as ∂/∂n = −∂/∂r,
−(n − 2)
Z Z
∂E 1
II = ϕ dσ = − ϕ(x) dσ(x)
|x|= ∂r (n − 2)ωn |x|= |x|n−1
Z Z
ϕ(0) dσ(x) 1 dσ(x)
= n−1
+ (ϕ(x) − ϕ(0)) n−1
ωn |x|=  ωn |x|= 
−→ ϕ(0),  → 0.

Theorem 5.3.
|x|2

1
exp(− ), t > 0,


E(x, t) = (4πt)n/2 4t

 0, t < 0,
is a fundamental solution of the heat equation in Rn+1 .

29
 ∂
Exercise 5.4. Show that ( ∂t − ∆x )E(t, x) = 0 if t 6= 0.
Proof. Let φ(x) = E(x, 12 ). When n = 1, this is the density of a N (0, 1) dis-
tributed stochastic variable and, when n >R 1 the product of n such densities.
Furthermore, E(x, t) = φ√2t (x) and thus Rn E(x, t)dx = 1 for all t > 0 and
E ∈ L1loc (Rn+1 ). Now
    Z Z
∂E ∂ϕ ∂ϕ
, ϕ = − E, = lim − dx E dt
∂t ∂t →0 n t> ∂t
Z Z RZ 
∂E
= lim E(x, )ϕ(x, )dx + ϕ dxdt ,
→0 Rn Rn t> ∂t
and
Z Z Z Z
h∆x E, ϕi = hE, ∆x ϕi = lim E∆x ϕdxdt = lim ∆x Eϕdxdt
→0 Rn t> →0 Rn t>
Thus
  
∂
− ∆x E, ϕ =
∂t
nZ Z Z  
∂E o
lim E(x, )ϕ(x, )dx + ϕ − ∆x E dxdt =
→0 Rn n ∂t
ZR t>
= Exercise 4 = lim E(x, )ϕ(x, )dx.
→0 Rn
Since E(x, t) = φ√2t (x) is an approximate identity, we ought to have
Z
I = E(x, )ϕ(x, )dx → ϕ(0),  → 0.
Rn
This does not follow directly from Theorem 1.4 since the support√of ϕ(x, )
is not compact and depend on . But the change of variables x = 2 y gives
Z √
I = φ(x)ϕ( 2 x, )dx .
Rn
√
Since φ ∈ L1 and |ϕ( 2 x, )| ≤ kϕk∞ , we get by dominated convergence
Z √ Z
lim I = φ(x) lim ϕ( 2 x, )dx = φ(x)ϕ(0, 0)dx = ϕ(0) .
→0 Rn →0 Rn

1 ∂
Exercise 5.5. Show that is a fundamental solution to in C.
πz ∂z
∂
Exercise 5.6. Compute log |z| and ∆ log |z| in C.
∂z
Exercise 5.7. H 3.3.9
Exercise 5.8. H 3.3.11
Exercise 5.9. H 3.3.12

30
Chapter 6

Distributions with compact

support

Theorem 6.1. Assume that u ∈ D 0 (Ω) has compact support Then u has a
unique extension to C ∞ (Ω) that satisfies u(ϕ) = 0 if supp u and supp ϕ are
disjoint.
If K is a compact set that contains a neighborhood of supp u, then
X
|u(ϕ)| ≤ C k∂ α ϕkK , ϕ ∈ C ∞ (Ω). (6.1)
|α|≤k

Proof. Take χ ∈ C0∞ (K) with χ = 1 in a neighborhood of supp u. If ϕ ∈ C0∞ ,

then according to Theorem 2.12
u(ϕ) = u(χϕ + (1 − χ)ϕ) = u(χϕ) + u((1 − χ)ϕ) = u(χϕ).
Thus
u(ϕ) = u(χϕ), ϕ ∈ C ∞
defines an extension of u. (1) follows Leibnitz’ rule.
Assume on the other hand that u1 is an extension to C ∞ . The condition
on the support implies that u1 ((1 − χ)ϕ) = 0, and consequently u1 (ϕ) =
u1 (χϕ) = u(χϕ) and thus the extension is unique.

Remark 6.2. Exercise 2.6 shows that it is not always possible to take K =
supp u in (1). 2
Exercise 6.10. State and prove a converse of Theorem 1.

Thus we can identify distributions with compact support with the linear
functionals on C ∞ (Ω) that satisfies (1). These distributions are denoted
E 0 (Ω).

31
Chapter 7

Convergence of distributions

Definition 7.1. A sequence uj ∈ D 0 (Ω) converges to u ∈ D 0 (Ω) if

uj (ϕ) → u(ϕ), j → ∞,
for every test function ϕ ∈ D(Ω). We denote this by uj → u in D 0 . 2
If uj → u inPD 0 , we also have ∂ α uj → ∂ α u in D 0 for every multiindex α.
We write u = uj in D 0 if the partial sums PconvergesPin αD . If the series
0
α
converges, it is differentiable and we have ∂ ( uj ) = ∂ uj .
Remark 7.2. Convergence in D 0 is a ”weak” condition, if for instance fj → f
in Lp then fj → f i D 0 . 2
Exercise 7.1. Prove that.

Definition 7.3. uj ∈ D 0 (Ω) is a Cauchy sequence in D 0 (Ω) if uj (ϕ) is a

Cauchy sequence in C for every ϕ ∈ D(Ω).
Theorem 7.4. D 0 (Ω) is complete.
Since uj (ϕ) is a Cauchy sequence in C, the following limit exist
u(ϕ) = lim uj (ϕ),
j→∞

and defines a linear functional on D(Ω). The difficulty is to show that u is a

distribution, i.e. that u satisfies the norm inequality (2.1), or the equivalent
formulation in Theorem 2.4. This is a consequence of the Banach-Steinhaus
theorem.
Let K be a compact set in Ω. We shall study the space X = XK =
{ϕ ∈ C ∞ (Ω); supp ϕ ⊂ K}. We introduce a metric on X by
X kϕ1 − ϕ2 kk
d(ϕ1 , ϕ2 ) = 2−k ,
k
1 + kϕ1 − ϕ2 kk

32
where kϕkk = |α|≤k supK |∂ α ϕ| and put kϕk = d(ϕ, 0).
P

Observe that if  > 0, and we take N = N so that ∞ −k

< 2 , then
P
N +1 2

N N
X  X −k  
kϕk ≤ 2−k kϕkk + ≤ 2 kϕkN + ≤ kϕkN + <  ,
k=1
2 k=1
2 2

if kϕkN < 2 .

Exercise 7.2. Show that d is a metric on X.

Exercise 7.3. Show that X is complete.
Exercise 7.4. Show that ϕj → 0 in D(K) if and only if kϕj k → 0.
Exercise 7.5. Show that if kϕj k → 0 in D(K), there are positive numbers cj with cj → ∞
but kcj ϕj k → 0.

The Banach-Steinhaus theorem

Let Λα be a family of linear functionals on X with |Λα ϕ| ≤ Cα kϕk. Then,
either
1) there are r > 0 and C < ∞ with

sup |Λα ϕ| ≤ C
α

for all ϕ ∈ X with kϕk ≤ r ,

or

2) supα |Λα ϕ| = ∞ for some ϕ ∈ X.

Now we can complete the proof of Theorem 4. Take ϕ with support in

K. Since uj (ϕ) converges, 2) can not hold. Thus 1) holds, ie.

|u(ϕ)| ≤ sup |uj (ϕ)| ≤ C if kϕk ≤ r .

j

Hence if ϕk → 0 in D(K), k → ∞, Exercise 7.5 implies that |u(ck ϕk )| ≤ C

if k is large enough. Thus |u(ϕk )| ≤ cCk → 0, k → ∞, ie. u ∈ D 0 .
2

33
 The Banach-Steinhaus theorem is a consequence of

Baire’s theorem
Assume that X is a complete metric space. Let V1 , V2 , . . . be open dense sets
in X. Then ∩i Vi is non-empty.

Proof. Let Br (φ) = {ϕ ∈ X; d(ϕ, φ) < r}. Since Vi are open and dense vi
can successivly choose φi and ri with ri < 1i such that Br1 (φ1 ) ⊂ V1 and
Bri (φi ) ⊂ Vi ∩ Bri−1 (φi−1 ), i = 1, 2, 3, . . ..
If i, j ≥ n, then φi , φj ∈ Brn (φn ), and therefore d(φ1 , φj ) < n2 . Thus φn
is a Cauchy sequence, and φn → φ0 for some φ0 ∈ X. But φi ∈ Brn (φn ) if
i ≥ n. Hence φ0 ∈ Brn (φn ) ⊂ Vn for all n and φ0 ∈ ∩Vi .

Proof of the Banach-Steinhaus theorem. Let φ(ϕ) = sup |Λα ϕ|. φ is lower
semi-continuous, and hence Vn = {ϕ; φ(ϕ) > n} is open. If some VN isn’t
dense, then there are ϕ0 , r with Br (ϕ0 ) ⊂ VNc ie.

{ϕ; kϕ − ϕ0 k < r} ⊂ VNc .

Thus if kϕk < r, then |Λα (ϕ0 + ϕ)| ≤ N . This implies

|Λα ϕ| ≤ |Λα (ϕ + ϕ0 )| + |Λα ϕ0 | ≤ 2N = C if kϕk < r.

On the other hand if all Vn are dense, then there are ϕ ∈ ∩Vn , ie. φ(ϕ) =
∞ or supα |Λα ϕ| = ∞.

Theorem 7.5. Assume that uj → u0 in D 0 (Ω) and that uj ≥ 0. Then uj

converges weakly to a positive measure u0 .

Proof. Since u0 is the limit of positive distributions, u0 is a positve distribu-

tion. By Theorem 2.7, u0 is a positive measure. If χ ∈ C0∞ is equal to 1 on
K, the proof of Theorem 2.7 gave the estimate

|uj (ϕ)| ≤ 2uj (χ)kϕk∞ ,

when ϕ ∈ C0∞ is supported in K.

Since uj (χ) → u0 (χ), we have supj |uj (χ)| ≤ C, and we obtan

|uj (ϕ)| ≤ Ckϕk∞ , ϕ ∈ C0∞ , j = 0, 1, 2, . . .

By taking limits, compare Theorem 2.5, this also holds when ϕ ∈ C0 .

34
 Now let ϕ ∈ C0 . We have to prove that uj (ϕ) → u0 (ϕ), j → ∞. Take
ϕn ∈ C0∞ whit ϕn → ϕ uniformly. Then

|uj (ϕ) − u0 (ϕ)| ≤ |uj (ϕ) − uj (ϕn )| + |uj (ϕn ) − u0 (ϕn )| + |u0 (ϕn ) − u0 (ϕ)|
= |uj (ϕ − ϕn )| + |uj (ϕn ) − u0 (ϕn )| + |u0 (ϕn − ϕ)|
≤ 2Ckϕ − ϕn k + |uj (ϕn ) − u0 (ϕn )| .

Hence
lim |uj (ϕ) − u0 (ϕ)| ≤ 2Ckϕ − ϕn k∞ < 
j→∞

if n is large enough.

Exercise 7.6. Assume that f is analytic in Ω = I × (0, δ) ⊂ C, where I is an open

interval. Show that if |f (z)| ≤ C|Imz|−N , then f (x + i0) = limy→0 f (x + iy) exists
in the distribution sense and f (x + i0) ∈ DN0
+1 (I).
Exercise 7.7. Compute
1 1
a) x+i0 + x−i0
and
1 1
b) x+i0 − x+i0 .
Exercise 7.8. H 2.5
Exercise 7.9. H 2.6
Exercise 7.10. H 2.7
Exercise 7.11. H 2.9
Exercise 7.12. H 2.16

35
Chapter 8

Convolution of distributions

If u ∈ L1loc and ϕ ∈ C0∞ , then u ∗ ϕ(x) =

R
u(y)ϕ(x − y)dy. This motivates
the following

Definition 8.1. If u ∈ D 0 (Rn ) and ϕ ∈ D(Rn ), then

u ∗ ϕ(x) = huy , ϕ(x − y)i.

The notation huy , ϕ(x − y)i means that the distribution u acts on the test
function y 7→ ϕ(x − y). Sometimes we also write hu, ϕ(x − ·)i.

Remark 8.2. This definition can also be used in the case where u ∈ E 0 (Rn ),
ϕ ∈ C ∞ (Rn ). 2

Theorem 8.1. If u ∈ D 0 (Rn ) and ϕ ∈ D(Rn ), then

a) u ∗ ϕ ∈ C ∞ (Rn )
b) supp(u ∗ ϕ) ⊂ supp u + supp ϕ
c) ∂ α (u ∗ ϕ) = u ∗ ∂ α ϕ = (∂ α u) ∗ ϕ

Proof. We first show that u ∗ ϕ is continuous. Let x → x0 . If |x − x0 | ≤ 1,

then y 7→ ϕ(x − y) has support in a fixed compact set. Furthermore
∂yα (ϕ(x−y)−ϕ(x0 −y)) → 0, x → x0 , uniformly. Hence ϕ(x−y) → ϕ(x0 −y),
in D when x → x0 , and we get u ∗ ϕ(x) = huy , ϕ(x − y)i → huy , ϕ(x0 − y)i =
u ∗ ϕ(x0 ), x → x0 .
Since u ∗ ϕ is continuous, to prove b) it is enough to show that if x ∈/
supp u + supp ϕ, then u ∗ ϕ(x) = 0. But if x ∈ / supp u + supp ϕ, there are
no y ∈ supp u with x − y ∈ supp ϕ. So there is no y with y ∈ supp u and
y ∈ supp ϕ(x − ·). Hence supp u ∩ supp ϕ(x − ·) = ∅ och u ∗ ϕ(x) = 0.

36
 The proof of the second equality in c) is simple.

∂ α u ∗ ϕ(x) = h∂ α uy , ϕ(x − y)i = (−1)|α| huy , ∂yα ϕ(x − y)i =

= huy , ϕ(α) (x − y)i = u ∗ (∂ α ϕ)(x).

The first equality follows by induction if we can prove it in the special case
α = (1, 0, . . . , 0). Thus it is enough to show that

1
lim (u ∗ ϕ(x + he1 ) − u ∗ ϕ(x)) = u ∗ ∂1 ϕ(x).
h→0 h

Let φx,h (y) = h1 (ϕ(x+he1 −y)−ϕ(x−y)). Then h1 (u∗ϕ(x+he1 )−u∗ϕ(x)) =

u(φx,h ). But φx,h (y) → ∂x∂ϕ
1
(x−y) i D. Hence ∂ α (u∗ϕ)(x) = limh→0 u(φx,h ) =
∂ϕ
uy ( ∂x1
(x − y)) = u ∗ ∂1 ϕ(x).
Since a) follows from c) the theorem is proved.

Exercise 8.1. Show that φx,h (y) → ∂ϕ

∂x1 (x − y) i D.
Exercise 8.2. Show that the convolution of functions is associative.

Theorem 8.2. If u ∈ D 0 (Rn ) and ϕ, ψ ∈ D(Rn ), then (u∗ϕ)∗ψ = u∗(ϕ∗ψ).

Remark 8.3. If u ∈ E 0 (Rn ), it is enough that one of ϕ, ψ has compact

support. 2

Proof. We have
Z
?
u ∗ (ϕ ∗ ψ)(x) = huy , ϕ ∗ ψ(x − y)i = huy , ϕ(x − y − t)ψ(t)dti =
Rn
Z Z
?
= huy , ϕ(x − y − t)iψ(t)dt = u ∗ ϕ(x − t)ψ(t)dt
Rn Rn
= (u ∗ ϕ) ∗ ψ(x).

?
To prove that = holds, we approximate the integral with its Riemann sum.
By Lemma 4 below the Riemann sum converges to the convolution in D and
?
thus = holds.

Lemma 8.4. If ϕ ∈ C0j (Rn ) and ψ ∈ C0 (Rn ), then

X
ϕ(x − kh)ψ(kh)hn −→ ϕ ∗ ψ(x) i C0j ,
k∈Zn

as h → 0.

37
Proof of Lemma 4. The sum is supported in supp ϕ + supp ψ. The function
(x, y) 7→ ϕ(x − y)ψ(y) is uniformly continuous. Hence the Riemann sum
converges uniformly to ϕ ∗ ψ(x). Since ∂ α (ϕ ∗ ψ) = ∂ α ϕ ∗ ψ om |α| ≤ j, this
also holds for the derivatives.

Theorem 8.5 (Regularisation of distributions.). Let u ∈ D 0 (Rn ) and ϕδ be

an approximate identity. Then u ∗ ϕδ → u in D 0 (Rn ), δ → 0 .
∨ ∨ ∨
Proof. Define ψ by ψ (x) = ψ(−x). Then u(ψ) = u∗ ψ (0). By Theorem 2,
this implies
∨
uδ (ψ) = u ∗ ϕδ (ψ) = (u ∗ ϕδ )∗ ψ (0) =
∨
= u ∗ (ϕδ ∗ ψ)(0) .
∨ ∨
But, since ϕδ is an approximate identity, ϕδ ∗ ψ→ψ in D(Rn ), δ → 0. Hence
∨ ∨
lim uδ (ψ) = lim u ∗ (ϕδ ∗ ψ)(0) = u∗ ψ (0) = u(ψ).
δ→0 δ→0

Exercise 8.3. Let u ∈ D 0 (Ω). Show that there are C0∞ functions un with un → u in
D 0 (Ω), n → ∞.

Example 8.6. An alternative proof that u is constant if u0 = 0.

Let uδ = u ∗ ϕδ ∈ C ∞ . Then u0δ = u0 ∗ ϕδ = 0 ∗ ϕδ = 0. Hence uδ = Cδ .
But uδ → u in D 0 , and Cδ → C for some constant C and u = C. 2

Exercise 8.4. Let u ∈ D 0 (R). Show that

a) If u0 ≥ 0, then u is an increasing function.
b) If u00 ≥ 0, then u is a convex function.

Example 8.7. Harmonic functions.

If u ∈ C 2 (Rn ) satisfies ∆u = 0, we say that u is a harmonic function.
Harmonic functions satisfies the mean value property.
Z
1
u(x) = u(y)dσ(y).
|Sr (x)| Sr (x)

In R2 , this follows from the Cauchy integral formula since a harmonic function
locally is the real part of a holomorphic function. The general case follows
from

38
Exercise 8.5. Prove the mean value property.
Hint. We may assume that x = 0. First apply Green’s identity to the functions u and
1 on Br = {|x| ≤ r}, and then on u and E (E is the fundamental solution of ∆) on
Ω = { ≤ |x| ≤ 1}. Let  → 0.

A different proof is given in Section 17.2.

Theorem 8.8 (Weyl’s lemma). If u ∈ D 0 (Rn ) and ∆u = 0, then u ∈ C ∞
and ∆u = 0 classically.
R
Proof. Let ϕδ be an approximate identity, ϕ(x) = ϕ(|x|), ϕ ≥ 0 and ϕ = 1.
Put uδ = u ∗ ϕδ . Then uδ ∈ C ∞ and ∆uδ = (∆u) ∗ ϕδ = 0 ∗ ϕδ = 0. So uδ
satisfies the mean value property. Hence
Z ∞ Z
n−1
uδ ∗ ϕ(x) = r ϕ(r)dr uδ (x − rω)dσ(ω)
0 S n−1
Z ∞ Z
n−1
= ωn uδ (x) r ϕ(r)dr = uδ (x) ϕ(y)dy = uδ (x).
0 Rn

Thus uδ = uδ ∗ ϕ. Now let δ → 0. We get u = u ∗ ϕ ∈ C ∞ and ∆u =

∆u ∗ ϕ = 0 ∗ ϕ = 0.
Next we will define the convolution of two distributions. We want to do
it in such a way that the associativity is preserved. To be able to that we
assume that at least one of the distributions has compact support.
Definition 8.9. Assume that u, v ∈ D 0 (Rn ), and at least one of them has
compact support. Then u ∗ v is the (uniquely determined) distribution that
satisfies
(u ∗ v) ∗ ϕ = u ∗ (v ∗ ϕ), ϕ ∈ D(Rn ).
2
Is this a definition?
We first observe that u ∗ (v ∗ ϕ) is well-defined. If v has compact support,
then v ∗ ϕ ∈ D and u ∗ (v ∗ ϕ) is well-defined by Definition 1.1. On the
other hand, if u has compact support, then v ∗ ϕ ∈ C ∞ and u ∗ (v ∗ ϕ) is
well-defined by Remark 1.2.
That there is at most one U = u ∗ v is also clear. Namely, if there
were two such distribution U and U e , then U ∗ ϕ = u ∗ (v ∗ ϕ) = Ue ∗ ϕ and
∨ ∨
U (ϕ) = U ∗ ϕ (0) = U e ∗ ϕ (0) = U
e (ϕ).
To show the existence we will study the map ϕ 7→ u ∗ ϕ.
Proposition 8.10. Let T ϕ = u ∗ ϕ. Then we have

a) If u ∈ D 0 (Rn ), then T is a continuous linear map D(Rn ) → C ∞ (Rn ).

39
 b) If u ∈ E 0 (Rn ), then T is a continuous linear map D(Rn ) → D(Rn ) and
C ∞ (Rn ) → C ∞ (Rn ).

Proof. We prove a) and leave b) as an exercise. Thus we assume that ϕj → 0

i D(Rn ) and shall prove that ∂ α (u∗ϕj ) → 0 uniformly on compact sets. Since
∂ α ϕj → 0 in D(Rn ) if ϕj → 0 in D(Rn ), and ∂ α (u ∗ ϕj ) = u ∗ ∂ α ϕj , we may
assume that α = 0. If x is contained in a compact set and if all ϕj are
supported in another compact set, then also y 7→ ϕj (x − y) is supported in
a fix compact set. Thus
X
|u ∗ ϕj (x)| = |u(ϕj (x − ·)| ≤ C k∂ α ϕj (x − ·)k∞ −→ 0, j → ∞.
|α|≤k

Exercise 8.6. Prove Proposition 10 b).

Let τh be the translation operator, τh ϕ(x) = ϕ(x − h). Then we have

Proposition 8.11. Convolution and translation commutes.
Proof.
u ∗ τh ϕ(x) = huy , τh ϕ(x − y)i = huy , ϕ(x − h − y)i
= u ∗ ϕ(x − h) = τh (u ∗ ϕ)(x).

An important converse of this is

Theorem 8.12. Assume that T is a continuous linear map from D(Rn )
into C ∞ (Rn ) that commutes with translations. Then there is a distribution
u ∈ D 0 (Rn ) with
T ϕ = u ∗ ϕ, ϕ ∈ D(Rn ).
∨ ∨
Proof. If T ϕ = u ∗ ϕ, then in particular u(ϕ) = u∗ ϕ (0) = T ϕ (0). We
therefore define u by
∨
u(ϕ) = T ϕ (0).
The continuity assumption implies that u is a distribution. Furthermore we
have
∨ ∨
u ∗ ϕ(h) = hu, ϕ(h − x)i = hu, τh ϕi = T ((τh ϕ)∨ )(0)
= T (τ−h ϕ)(0) = τ−h T (ϕ)(0) = T ϕ(h) .

40
 The above results implies that Definition 9 is a definition. Proposition 10
shows that ϕ 7→ u ∗ (v ∗ ϕ) satisfies the conditions in Theorem 12 and, u ∗ v
is this distribution. 2

Remark 8.13.

a) If v ∈ D(Rn ), Definitions 1 and 2 coincides.

b) If both u and v have compact support, then (u ∗ v) ∗ ϕ = u ∗ (v ∗ ϕ) for

all ϕ ∈ C ∞ (Rn ).
2

Example 8.14. u ∗ δ = u since (u ∗ δ) ∗ ϕ = u ∗ (δ ∗ ϕ) = u ∗ ϕ. 2

Theorem 8.15.

a) u ∗ v = v ∗ u

b) supp (u ∗ v) ⊂ supp u + supp v

c) u ∗ (v ∗ w) = (u ∗ v) ∗ w
if at least two of the distributions have compact support.

Proof. a) To show that two distributions U and V coincides, it is enough to

show that U ∗ (ϕ ∗ ψ) = V ∗ (ϕ ∗ ψ), if ϕ, ψ ∈ D(Rn ). Namely, in that case,
(U ∗ ϕ) ∗ ψ = U ∗ (ϕ ∗ ψ) = V ∗ (ϕ ∗ ψ) = (V ∗ ϕ) ∗ ψ, according to Theorem 2.
This implies U ∗ ϕ = V ∗ ϕ, and U = V .
Now
(u ∗ v) ∗ (ϕ ∗ ψ) = u ∗ (v ∗ (ϕ ∗ ψ)) = u ∗ ((v ∗ ϕ) ∗ ψ)
= u ∗ (ψ ∗ (v ∗ ϕ)) = (u ∗ ψ) ∗ (v ∗ ϕ).

If v has compact support, the last equality follows by Theorem 8.2. If v does
not have compact support it follows from the next exercise.
We also have

(v ∗ u) ∗ (ϕ ∗ ψ) = (v ∗ u) ∗ (ψ ∗ ϕ) = (v ∗ ϕ) ∗ (u ∗ ψ) = (u ∗ ψ) ∗ (v ∗ ϕ),

and a) is proved.
b) By the commutativity we may assume that v has compact support.
∨ ∨ ∨
Definie v by hv, ϕi = hv, ϕi. If x ∈ supp (u ∗ v), there is to every  > 0 a
∨
ϕ ∈ D(Rn ), supp ϕ ⊂ {y; |x − y| < } = O , with 0 6= u ∗ v(ϕ) = u ∗ v∗ ϕ (0)
∨ ∨ ∨
= u((v∗ ϕ)∨ ) = u(v ∗ϕ). So E = supp u ∩ supp (v ∗ϕ) 6= ∅. Let y ∈ E. Then

41
 ∨
y ∈ supp u och y ∈ supp v ∗ϕ, or y = −z + x + δ, where z ∈ supp v and
|δ| < . Thus x = y + z − δ ∈ supp u + supp v + O . Now let  → 0.
c) Assume first that w has compact support. Then w ∗ ϕ ∈ D, and we
get
((u ∗ v) ∗ w) ∗ ϕ = (u ∗ v) ∗ (w ∗ ϕ) = u ∗ (v ∗ (w ∗ ϕ)).
But also,
(u ∗ (v ∗ w)) ∗ ϕ = u ∗ ((v ∗ w) ∗ ϕ) = u ∗ (v ∗ (w ∗ ϕ))
and hence u ∗ (v ∗ w) = (u ∗ v) ∗ w.
If w does not have compact support, both u and v have, and a) implies
u ∗ (v ∗ w) = (v ∗ w) ∗ u = v ∗ (w ∗ u) = (w ∗ u) ∗ v
= w ∗ (u ∗ v) = (u ∗ v) ∗ w.

Exercise 8.7. Show that u ∗ (ψ ∗ ϕ) = (u ∗ ψ) ∗ ϕ if u ∈ E 0 , ψ ∈ D and ϕ ∈ C ∞ .

Theorem 8.16. ∂ α (u ∗ v) = ∂ α u ∗ v = u ∗ ∂ α v if at least one of the distri-

butions have compact support.
Proof. If u ∈ D 0 (Rn ), we have ∂ α u = ∂ α δ ∗ u, since
∂ α u ∗ ϕ = u ∗ ∂ α ϕ = u ∗ (δ ∗ ∂ α ϕ) = u ∗ (∂ α δ ∗ ϕ) = (u ∗ ∂ α δ) ∗ ϕ.
Using this we get
∂ α (u ∗ v) = ∂ α δ ∗ (u ∗ v) = (∂ α δ ∗ u) ∗ v = ∂ α u ∗ v.
The second equality follows from Theorem 15 a).
Theorem 8.17. Assume that u ∈ Dk0 and v ∈ C0k (or u ∈ Ek0 , v ∈ C k ). Then
u ∗ v is the continuous function x 7→ huy , v(x − y)i.
Proof. If x → x0 , then v(x − ·) → v(x0 − ·) i C0k . But u is continuous on C0k ,
and we get huy , v(x − y)i → huy , v(x0 − y)i. Thus h(x) = huy , v(x − y)i is a
continuous function.
According to Definition 9, (u ∗ v) ∗ ψ = u ∗ (v ∗ ψ). As in the proof of
Theorem 2 one can show that h ∗ ψ = u ∗ (v ∗ ψ). Hence h = u ∗ v.

Exercise 8.8. Let u, v ∈ D 0 (R) with support in {x ≥ 0}. Define u ∗ v.

Exercise 8.9. H 4.2.1
Exercise 8.10. H 4.2.2
Exercise 8.11. H 4.2.3
Exercise 8.12. H 4.2.4

42
Chapter 9

Fundamental solutions

Let X
P = aα ∂ α
|α≤N |

be a differential operator with constant coefficients and E a fundamental

solution to P , ie. E ∈ D 0 (Rn ) and P E = δ. Then

P (E ∗ f ) = f, f ∈ E 0 (Rn ), (9.1)

and
E ∗ P u = u, u ∈ E 0 (Rn ). (9.2)
Thus E is both a left and a right invers to P on E 0 . (But on different domains,
so it does not imply that P is bijective.) So (1) gives a solution u = E ∗ f
of the equation P u = f if f has compact support. (2) can be used to study
regularity of solutions of P u = f .
Remark 9.1. In Chapter 14 we will show that every differential operator
with constant coefficients has a fundamental solution. 2
In Chapter 5, we obtained fundamental solutions to the Laplace and heat
equations. Another example is that

 (x . . . x )k /(k!)n , all x > 0
1 n i
Ek (x) =
 0 otherwise,

is a fundamental solution to Pk+1 = ∂1k+1 . . . ∂nk+1 . Using this we can prove

Theorem 9.2. If u ∈ Em0 (Rn ), there is a continuous function f with

∂1m+2 . . . ∂nm+2 f = u.

43
Proof. Em+1 is a fundamental solution to Pm+2 . Thus f = Em+1 ∗ u satisfies
Pm+2 f = u. By Theorem 8.17, f is continuous.
A corollary of Theorem 1 is the following representation theorem for
distributions.
Theorem 9.3. If u ∈ D 0 (Ω), there are functions fα ∈ C(Ω) with
X
u= ∂ α fα

in D 0 . The sum is locally finite, and if u has finite order the sum is finite.
Proof. Choose a partition of unity ψi ∈ C0∞ and χi ∈ C0∞ with χi = 1 on
supp ψi . This can be done in such a way that Σχi is locally finite. We get
X X
u(ϕ) = ψi u(ϕ) = χi u(ψi ϕ).
i i

The distribution χi u has compact support, and hence finite order. Theorem 1
implies that χi u = ∂ αi fi , fi ∈ C. Hence
X X Z
αi |αi |
u(ϕ) = ∂ fi (ψi ϕ) = (−1) fi ∂ αi (ψi ϕ) dx .
i i Rn

If we compute ∂ αi (ψi ϕ), we get

XX Z XX
|α|
u(ϕ) = (−1) fi,α ∂ α ϕ dx = ∂ α fi,α (ϕ) .
i α Rn α i
P
Finally, we let fα = i fi,α .
To study the regularity of solutions of the equation P u = f , we want to
study the set where u is not C ∞ .
Definition 9.4. The singular support of a distribution u ∈ D 0 (Ω) is denoted
sing supp u, and consists of those points in Ω that have no neighborhood
where u is C ∞ . 2
sing supp u is the smallest closed set such that u is C ∞ in its complement.
It is clear that sing supp u ⊂ supp u.
Theorem 9.5. If u, v ∈ D 0 (Rn ), and at least one of them have compact
support, then

sing supp (u ∗ v) ⊂ sing supp u + sing supp v. (9.3)

44
Proof. Put u1 = u and u2 = v. Let us first assume that both distributions
have compact support. Let Ki = sing supp ui , and Ωi a neighborhood of Ki
and take ψi ∈ C0∞ (Ωi ) with ψi = 1 on Ki . Then
u1 ∗ u2 = (ψ1 u1 + (1 − ψ1 )u1 ) ∗ (ψ2 u2 + (1 − ψ2 )u2 )
=ψ1 u1 ∗ ψ2 u2 + ψ1 u1 ∗ (1 − ψ2 )u2
+(1 − ψ1 )u1 ∗ ψ2 u2 + (1 − ψ1 )u1 ∗ (1 − ψ2 )u2 .

Since (1 − ψi )ui ∈ C0∞ , Theorem 8.1 implies that the last three terms are
C ∞ . Thus
sing supp (u1 ∗ u2 ) = sing supp (ψ1 u1 ∗ ψ2 u2 )
⊂ supp (ψ1 u1 ∗ ψ2 u2 ) ⊂ supp ψ1 + supp ψ2 ⊂ Ω1 + Ω2 .

If we let Ωi ↓ Ki , we obtain (9.3).

If only one of the distributions have compact support, we can by Theo-

rem 8.15 asssume that v ∈ E 0 .
To show that sing supp u ∗ v ⊂ sing supp u+sing supp v, it is enough to
show that sing supp u ∗ v ∩ B1 (x) ⊂ (sing supp u+sing supp v)∩B1 (x) for each
x ∈ Rn .
Take R ≥ 1 so large that supp v ⊂ BR (0) and |x| ≤ R, and choose
χ ∈ C0∞ (B6R (0)) with χ = 1 on B5R (0). Put u1 = χu and u2 = (1 − χ)u.
C
Thus u = u1 + u2 where u1 has compact support and supp u2 ⊂ B5R (0).
C C
Then supp u2 ∗ v ⊂ B5R (0) + BR (0) ⊂ B4R (0) and B1 (x) ⊂ B2R (0). Hence
u2 ∗ v = 0 on B1 (x). Since both u1 and v have compact support, we get
sing supp (u ∗ v) ∩ B1 (x) = sing supp (u1 ∗ v) ∩ B1 (x)
⊂ (sing supp u1 + sing supp v) ∩ B1 (x) = (sing supp u + sing supp v) ∩ B1 (x),
and the theorem is proved.
Theorem 9.6. If P has a fundamental solution with sing supp E = {0}, then
sing supp u = sing supp P u, u ∈ D 0 .
Remark 9.7. The converse is also true. Thus if there is a fundamental solu-
tion with singular support at the origin, then this is true for all fundamental
solutions. 2
Proof. sing supp P u ⊂ sing supp u allways holds since P u is C ∞ if u is.
For the other inclusion, we first observe that if u has compact support,
then u = E ∗ P u and by Theorem 5

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

45
If u is not compactly supported, take ψ ∈ C0∞ with ψ = 1 on an open set Ω.
Then
sing supp ψu ⊂ sing supp P (ψu).
But on Ω we have P (ψu) = P u and ψu = u, and the result follows.
A differential operator P is called hypoelliptisk if every solution u of P u =
f is C ∞ if f is. Theorem 6 thus implies that P is hypoelliptic if P has a
fundamental solution E with sing supp E = {0}.
The Laplace and the heat operators are hypoelliptic. In Chapter 12, we
will show that all elliptic operators are hypoelliptic. The Laplace operator is
elliptic but not the heat operator.
d
Exercise 9.1. Let P be a polynomial of one variable. Show that P ( dx ) has a fundamental
solution.
Exercise 9.2. H 4.4.3
Exercise 9.3. H 4.4.4
Exercise 9.4. H 4.4.5
Exercise 9.5. H 4.4.6
Exercise 9.6. H 4.4.9

46
Chapter 10

The Fourier transform

If u is a ”nice” periodic function with period T , u can be written as

X
u(x) = cm e2πimx/T . (10.1)
m

Then X
u(x)e−2πiνx/T = cν e2πi(m−ν)x/T ,
ν

and integration over [− T2 , T2 ] gives formally that

Z T
2
T cν = u(x)−2πiνx/T dx
− T2

or
Z T
1 2
cν = u(x)e−2πiνx/T dx
T − T2

cν are the Fourier coefficients of u. (1) is the inversion theorem. We also

have Parseval’s identity
X 1
|cν |2 = kuk22 .
2π
How can this be generalised to Rn ? Let us first consider the case n = 1.
Let u ∈ C0∞ (R) and choose T so that supp u ⊂ (− T2 , T2 ). Let uT be the
periodic extension of u,
X
uT (x) = u(x − kT ).
k∈Z

47
Then we have Z T
1 2 2πν
cT (ν) = u(x)e−i T
x
dx.
T − T2

Thus for |x| < T2 , (1) implies

ν
X
u(x) = uT (x) = cT (ν)e2πi T x .
ν

Define Z
u
b(ξ) = u(x)e−iξx dx, ξ ∈ R.
R
1 2πν
We observe that cT (ν) = T
u
b( T ), and we can write
X 1 2πν ν 1 X 2π 2πν i 2πν x
u(x) = u
b( )e2πi T x = u
b( )e T .
ν
T T 2π ν
T T

This is a Riemann sum of the integral

Z
1
ub(ξ)eiξx dξ.
2π R
So, if we let T → ∞, we obtain
Z
1
u(x) = b(ξ)eiξx dξ.
u (10.2)
2π R

With some care, the above argument can be used to prove (2) when u is a nice
function. We will not do this, but instead prove (2) (and its generalization
to Rn ) directly. The theory for Fourier series will then be a corollary of the
theory of the Fourier transform.
Definition 10.1. Assume that f ∈ L1 (Rn ). The Fourier transform of f is
defined by Z
f (ξ) =
b f (x)e−ixξ dx,
Rn
Pn
where xξ = 1 xi ξi . We sometimes write Ff instead of fb.
2
We will prove the following important properties of the Fourier transform.

I. The inversion formula. If f and fb ∈ L1 , then

Z
1
f (x) = fb(ξ)eixξ dξ.
(2π)n Rn

48
 1
II. Parseval’s identity. If f ∈ L1 ∩ L2 , then fb ∈ L2 and kf k2 = (2π)n
kfbk2 .

III. If f, g ∈ L1 , then (f ∗ g)∧ = fb gb.

IV. F(P (D)f )(ξ) = P (ξ)fb(ξ) where Dj = −i∂j .

Exercise 10.1. Prove the Riemann-Lebesgue lemma: If f ∈ L1 , then fb is continuous

and fb(ξ) → 0 when |ξ| → ∞.
Exercise 10.2. Prove III and IV.

To solve the constant coefficient differential equation P (D)u = f , we

can use I and IV. By Fourier transformation, we get P (ξ)b u(ξ) = fb(ξ). Thus
−1
b(ξ) = fb(ξ)/P (ξ) and u = F (fb/P ). To be able to use this method ”often”,
u
we want to extend the Fourier transform to distributions. As a motivation
for the definition, we observe that, by Fubini’s theorem we have
Z Z
1
Proposition 10.2. If f, g ∈ L , then f gb dx = fbg dx.
Rn Rn

Exercise 10.3. Prove Proposition 2.

Hence for a L1 function we have

Z Z
hf , ϕi =
b f ϕ dx =
b b dx = hf, ϕi.
fϕ b
Rn Rn

b when u ∈ D 0 by
It is therefore natural to define u

hb
u, ϕi = hu, ϕi.
b (10.3)

But if ϕ ∈ D, ϕ 6≡ 0, then ϕ
b can not have compact support and hence ϕ b∈/ D.
So we can not define hu, ϕi
b by (3).
So what to do? Well, we will consider a different class of test functions,
that is preserved by the Fourier transform. This is the Schwartz space S ,
that consists of C ∞ rapidly decreasing functions.
Definition 10.3.
(a) ϕ ∈ S (Rn ) if ϕ ∈ C ∞ and supx∈Rn (1 + |x|2 )k |∂ α ϕ(x)| < ∞ for alla k
and α.
(b) ϕj → 0 in S (Rn ) if

sup (1 + |x|2 )k |∂ α ϕj (x)| → 0

x∈Rn

for all k and α. 2

49
Definition 10.4.

(a) A tempered distribution on Rn is a linear functional on S , such that

u(ϕj ) → 0 when ϕj → 0 in S . We write u ∈ S 0 .

(b) A sequence uj ∈ S 0 converges to u ∈ S 0 if

uj (ϕ) → u(ϕ),

for every testfunction ϕ ∈ S .

To show that (3) works as a definition of the Fourier transform if u ∈ S 0

we need to study the Fourier transform on S . We start with the following

Proposition 10.5. If f, g ∈ S , then

(a) F(xα f (x)) = iα ∂ α fb

(b) (∂ α f )∧ (ξ) = (iξ)α fb(ξ)

(c) (τh f )∧ (ξ) = e−iξh fb(ξ)

(d) F(eixh f (x)) = τh fb

(e) fba (ξ) = fb(aξ)

(f) (f (ax))∧ = (fb)a

Z Z
(g) f gb = fbg
Rn Rn

(h) (f ∗ g)∧ = fb gb
and

(i) fb ∈ S .

Exercise 10.4. Prove Proposition 5.

Theorem 10.6 (The inversion theorem). If f ∈ S , then

Z
1
f (x) = eixξ fb(ξ)dξ. (10.4)
(2π)n Rn

50
 To prove this we need to find one function that satisfies (4). Then (4)
2
follows in general by Proposition 5. We make the choice G(x) = e−|x| /2 .
b = (2π)n/2 G.
Lemma 10.7. G
Proof. By Fubinin’s theorem it is enough to consider the case n = 1. G
satisfies the differential equation

G0 (x) + xG(x) = 0.

If we take the Fourier transform of this equation, Proposition 5 (a) and (b)
implies
iξ G(ξ)
b + iG b0 (ξ) = 0,
or
b0 (ξ) + ξ G(ξ)
G b = 0.
Hence G(ξ)
b = CG(ξ). If we let ξ = 0, we get
Z
2 √
C = G(0)
b = e−x /2 dx = 2π.
R

Exercise 10.5.
a) Prove Lemma 7 using the Cauchy theorem.
b) Prove Lemma 7, letting ξ = ζ ∈ C, and compute G(iη).
b

1
Proof of Theorem 6. (2π) n (G)δ is an approximate identity. Proposition 5 f)
b
and g) implies that
Z Z
1 1
f (x)(G)δ (x)dx =
b fb(ξ)G(δξ)dξ.
(2π)n Rn (2π)n Rn
Letting δ → 0, we get
Z Z
1 1
f (0) = G(0) fb(ξ)dξ = fb(ξ)dξ.
(2π)n Rn (2π)n Rn

Exercise 10.6. Prove this.

If we apply this to τ−x f , we get

Z Z
1 ∧ 1
f (x) = τ−x f (0) = (τ−x f ) (ξ)dξ = fb(ξ)eixξ dξ.
(2π)n Rn (2π)n Rn

51
Remark 10.8. If we only assume that f ∈ L1 , then
Z Z
1 1 1 2 2

n
f ∗(G)δ (x) =
b
n
f (x+y)Gδ (y)dy =
b
n
fb(ξ)eixξ e−δ |ξ| /2 dξ.
(2π) (2π) Rn (2π) Rn
This implies Z
1 2 2
f (x) = lim fb(ξ)eixξ e−δ |ξ| /2 dξ,
δ→0 (2π)n Rn
1
with convergence in L .
In particular, if fb ∈ L1 , then
Z
1
f (x) = fb(ξ)eixξ dξ a.e.
(2π)n Rn

2
Theorem 10.9 (Plancherel). If φ, ψ ∈ S , then
Z Z
1
φψ̄ dx = φb ψb dξ .
Rn (2π)n Rn
Corollary 10.10 (Parseval). If φ ∈ S , then
1
kφk2 = kφk
b 2
(2π)n/2
.
Proof. Proposition 2 g) implies
Z Z
φψ0 dx =
b φψ
b 0 dx .
Rn Rn

Let ψb0 = ψ̄. By the inversion theorem,

Z Z
1 ixξ b 1
ψ0 (x) = e ψ0 (ξ)dξ = eixξ ψ̄(ξ)dξ
(2π)n Rn (2π)n Rn
Z
1 1 b
= n
e−ixξ ψ(ξ)dξ = ψ(x).
(2π) Rn (2π)n
Z Z
1
Thus φψ̄ = n
φb ψ.
b The corollary follows by taking ψ = φ.
Rn (2π) Rn

Remark 10.11. The Parseval formula also holds if φ, ψ ∈ L2 . We will prove

this in the next chapter. 2
To prove that ub, defined by u
b(ϕ) = u(ϕ),
b is a tempered distribution we
need the following

52
Lemma 10.12. F : S → S continuously, i.e., if ϕj → 0 i S , then ϕ bj →
0 i S.


Proof. Proposition 5 a) and b) implies ξ β ∂ α ϕ bj (ξ) = cF ∂ β (xα ϕj (x)) (ξ) .
Hence
Z 
−ixξ β
β α
 α

sup |ξ ∂ ϕ bj (ξ)| ≤ c sup  e ∂ (x ϕj (x))dx
ξ ξ Rn
Z
≤c |∂ β (xα ϕj (x))|dx → 0,
Rn

as ϕj → 0 in S .

We are now ready to make the following definition.

Definition 10.13. If u ∈ S 0 , then u

b is the tempered distribution given by

u
b(ϕ) = u(ϕ).
b

Remark 10.14. We observe that the two definitions of fb when f ∈ L1

coincides. 2

Theorem 10.15. The Fourier transform is a continuous linear bijection

∨
from S 0 to S 0 with u
b = (2π)n u.
b

∨ ∨ ∨
Proof. We remind the reader that u is defined by u (ϕ) = u(ϕ), and that
uj → u in S 0 means that uj (ϕ) → u(ϕ) for all ϕ ∈ S . The theorem is an
easy consequence of the corresponding properties on S :
∨ ∨
u
b(ϕ) = u
b b(ϕ) bb = (2π)n u(ϕ
b = u(ϕ) ) = (2π)n u(ϕ)

and
u b → u(ϕ)
bj (ϕ) = uj (ϕ) b =u
b(ϕ)
if uj → u i S 0 .

Example 10.16. a) A measure µ with Rn (1 + |x|2 )−k dµ(x) < ∞ for some
R

k is a tempered distribution.
b) If f ∈ Lp , 1 ≤ p ≤ ∞, then f ∈ S 0 . (Proof. Hölder’s inequality.)
c) δb = 1 and b1 = (2π)n δ.
x
d) e is not a tempered distribution. 2

53
Proposition 10.17. If u ∈ S 0 , then
a) (xj u)∧ = −Dj ub
∧
b) (Dj u) = ξj ub
∧
c) (τh u) (ξ) = exp(−ihξ)b
u(ξ)
and
d) F(exp(ixh)u) = τh ub.

Proof. It is easy to see that Dj u, xj u, . . . are tempered distributions. Then

the formulas follows from Proposition 5. (Remember that Dj = −i∂j .)
Exercise 10.7. Show that ex cos(ex ) ∈ S 0 .
Exercise 10.8. Show that u ∈ S 0 if and only if
X
|u(ϕ)| ≤ C sup(1 + |x|2 )k |∂ α ϕ(x)|
k+|α|≤N

for some N .
Exercise 10.9. H 7.1.10
Exercise 10.10. H 7.1.19
Exercise 10.11. H 7.1.20
Exercise 10.12. H 7.1.21
Exercise 10.13. H 7.1.22
Exercise 10.14. H 7.6.1

54
Chapter 11

The Fourier transform on L2

According to Exempel 10.16b), f ∈ L2 has a Fourier transform defined as a

tempered distribution. In fact we have the following result.
Theorem 11.1. If f ∈ L2 (R2 ), then fˆ ∈ L2 (Rn ) and
1
kf k2 = kfˆk2 .
(2π)n/2

Furthermore fˆ is given by
Z
fˆ(ξ) = lim e−iξx f (x)dx ,
N →∞ |x|≤N

with convergence in L2 .
Proof. Take fn ∈ C0∞ , fn → f in L2 . Then fn is a Cauchy sequence in L2 .
By the Plancherel theorem, we have
kfˆn − fˆm k2 = ckfn − fm k2 → 0, n, m → ∞.
Hence fˆn is a Cauchy sequence in L2 . By the completeness of L2 , we have
fˆn → g i L2 for some g ∈ L2 . This implies that fˆn → g in S 0 . Furthermore
fn → f in S 0 , and since the Fourier transform is continuous on S 0 , we have
fˆn → fˆ in S 0 . Hence fˆ = g ∈ L2 and we get
1 1
kfˆk2 = lim kfˆn k2 = n/2
lim kfn k2 = kf k2 ,
n→∞ (2π) n→∞ (2π)n/2
and the first part is proved.
Put fN = f χ{|x|≤N } . Then fN → f in L2 and fN ∈ L1 . Hence
Z
fˆN (ξ) = e−ixξ f (x)dx,
|x|≤N

55
and by the Plancherel theorem, we obtain

kfˆ − fˆN k2 = ckf − fN k2 → 0, N → ∞,

and the proof is complete.

56
Chapter 12

The Fourier transform and

convolutions

We shall show that under suitable conditions (u ∗ v)∧ = u

b vb.
First we observe that D ⊂ S ⊂ C . The inclusions are continuous, ie. if
∞

ϕj → 0 i D, then ϕj → 0 in S , and this in turn implies that ϕj → 0 in C ∞ .

Furthermore, D is dense in S , and S is dense in C ∞ . (Show that!) Hence

E 0 (Rn ) ⊂ S 0 (Rn ) ⊂ D 0 (Rn ).

Definition 12.1. If u ∈ S 0 and φ ∈ S , we define the convolution u ∗ φ by

u ∗ φ(x) = huy , φ(x − y)i. 2

Theorem 12.2. If u ∈ S 0 and φ ∈ S , then

(a) u ∗ φ ∈ C ∞ och ∂ α (u ∗ φ) = ∂ α u ∗ φ = u ∗ ∂ α φ

(b) u∗φ is bounded by a polynomial (and hence u∗φ ∈ S 0 ), and (u∗φ)∧ =

φb u
b.

(c) u ∗ (φ ∗ ψ) = (u ∗ φ) ∗ ψ (ψ ∈ S )
and

b ∗ φb = (2π)n (φu)∧ .
(d) u

Sketch of proof. (a) We assume that n = 1. The second equality is proved

in the same way as in Theorem 8.1. As in the proof of Theorem 8.1, the first
equality folows if we can prove that

φ(x + h) − φ(x)
−→ φ0 (x) in S .
h
57
To do this is elementary but tedious. The simplest way is (probably) to use
the Fourier transform.
(b) By Exercise 10.8, we have
|u ∗ φ(x)| = |huy , φ(x − y)i|
X
≤ C sup (1 + |y|2 )k |∂ α φ(x − y)|
y
k+|α|≤N
X
≤ C sup (1 + |x|2 )k (1 + |x − y|2 )k |∂ α φ(x − y)|
y
k+|α|≤N

≤ C(1 + |x|2 )N .
If ψ ∈ D, we also have
∨
Z
∧ n n
b = (u ∗ φ)(ψb ) = (2π) (u ∗ φ)(ψ) = (2π)
(u ∗ φ) (ψ) u ∗ φ(x)ψ(−x)dx
b
Rn
Z
n
= (2π) huy , ψ(−x)φ(x − y)idx = Approximate with a Riemann sum =
−K
Z Z
n n
= (2π) huy , ψ(−x)φ(x − y)dxi = (2π) huy , ψ(x)φ(−y − x)dxi
Rn Rn
∧
= (2π)n huy , (φ ∗ ψ)∨ )i = huy , (φ ∗ ψ)∧ i = u
b((φ ∗ ψ)∧ ) = u
b(φbψ)
b = φb u
b(ψ).
b

But Db is dense in S , and (b) follows.

(c) From the proof of (b), we get
∨
u ∗ φ(ψ) = u((φ ∗ ψ)∨ ),
first for ψ ∈ D, and by contiuity also for ψ ∈ S . This can be written
(u ∗ φ) ∗ ψ(0) = u ∗ (φ ∗ ψ)(0).
The general case follows if we replace ψ with τ−x ψ.
∨∨
(d) By (b), we have (b b ∧ = φb
u∗φ)
bb
b = (2π)2n φu= (2π)2n (φu)∨ = (2π)n (φu)b.
u
b

Thus, by the inversion theorem we get u b ∗ φb = (2π)n (φu)b.

Theorem 12.3. If u ∈ E 0 (Rn ), then u

b ∈ C ∞ and u
b(ξ) = ux (e−ixξ ).
Proof. Let ψ ∈ C0∞ be 1 on a neighborhood of supp u. Then u = ψu and by
Theorem 2(d) we get ub = (ψu)b = (2π)−n u
b ∗ ψb ∈ C ∞ . Thus
∨

b(ξ) = (2π)−n u
u b = (2π)−n hb
b ∗ ψ(ξ) b − x)i = (2π)−n hb
ux , ψ(ξ ux , ψb (x − ξ)i
−2n 3 −2n 3
= (2π) hb ux , F ψ(x − ξ)i = (2π) hb ux , τξ F ψ(x)i
−2n −ixξ 4 −iξx
= (2π) hux , e F ψ(x)i = ux (e ψ(x)) = ux (e−iξx ).

58
Remark 12.4. In the next chapter we will prove the Paley-Wiener theorem
b when u ∈ E 0 .
that gives much more precise information of u 2
1
Example 12.5. Determine the Fourier transform of pv . 2
x
1
Method 1. Let u = pv . Then u is the sum of a distribution in E 0 and
x
an L2 function. Thus
Z
dx
u
b(ξ) = lim e−ixξ .
→0
N →∞ <|x|<N
x

If ξ > 0, the change of variables y = xξ implies,

Z ∞
e−iy
Z
sin y
u
b(ξ) = lim dy = −i dy = −iπ.
→0
N →∞ <|x|<N
y −∞ y

When ξ < 0, we instead get u

b(ξ) = iπ. Hence

b(ξ) = −π sgn ξ.
u
R∞ sin y
Exercise 12.1. Prove that −∞ y dy = π.

1
Method 2. We have xpv u0 = 2πδ, u
= 1. This implies ib b0 = −2πiδ and
x
b. Hence c = − 12 och
b = −2πi(H + c). Since u is odd, so is u
u

b(ξ) = −iπ sgn ξ.

u

In the last argument we used that if a distribution u is odd, then u

b is also
1
odd. This is clear if u ∈ L (a simple change of variables).
∨ ∨
Definition 12.6. A distribution is even if u= u, and odd if u= −u. 2

Proposition 12.7. If u is an odd tempered distribution, then its Fourier

transform is also odd.

Proof. By Theorem 10.15

∨  ∧ b∨
−n −n b
u
b= (2π) u
b = (2π) u b = u = (u is odd) = −b
u.
b
b

In the same way we see that the Fourier transform of an even distribution
is even.

59
 1
Remark 12.8. The map Hϕ = pv ∗ ϕ, is called the the Hilbert transform.
x
The Hilbert transform is an important example of a so called singular integral
operator. The Hilbert transform is bounded on Lp , 1 < p < ∞, and of weak-
type (1, 1).
When p = 2, this follows from Exampel 5 and the Plancherel theorem. 2

Next we will study invariance properties of the Fourier transform. Let

F : Rn → Rn be a diffeomorphism (i.e. a C ∞ bijection). If u is a function,
we have Z Z
ϕ
u ◦ F (x)ϕ(x)dx = u(y) 0 ◦ F −1 (y)dy.
Rn Rn |F |
Therefore, if u ∈ D 0 , we define u ◦ F by
ϕ
hu ◦ F, ϕi = hu, ◦ F −1 i
|F 0 |

In particular, if F = Λ is linear, then

hu ◦ Λ, ϕi = | det Λ|−1 hu, ϕ ◦ Λ−1 i

Definition 12.9. A distribution u is radial if u ◦ O = u for all orthogonal

maps O.

Theorem 12.10. If u is a radial tempered distribution, then u

b is radial.

Proof. First, we observe that if ϕ ∈ S , then

Z
b ◦ O(ξ) = ϕ(Oξ)
ϕ b = e−ixOξ ϕ(x)dx
Rn
 
∗
y=O x
Z
∗
= e−iO xξ ϕ(x)dx =  
Rn x = Oy
Z
= e−iyξ ϕ(Oy)dy = (ϕ ◦ O)∧ (ξ).
Rn

This implies

hb
u ◦ O, ϕi = hbu, ϕ ◦ O−1 i = hu, (ϕ ◦ O∗ )∧ i
b ◦ O∗ i = hu, ϕ
= hu, ϕ b ◦ O−1 i = hu ◦ O, ϕb i = hu, ϕ
b i = hb
u, ϕi.

60
Theorem 12.11. If u is a tempered distribution that is homogeneous of
b is homogeneous of degree −n − α.
degree α, then u
Proof. By Definition 4.4, u ∈ S 0 is homogeneous of degree α if hu, ϕt i =
tα hu, ϕi. Therefore,

hb
u, ϕt i = hu, ϕ
bt i = huξ , ϕ(tξ)i
b = t−n hu, (ϕ)
b 1/t i
= t−(n+α) hu, ϕi
b = t−(n+α) hb
u, ϕi.

Example 12.12. A fundamental solution of the Laplace operator when n ≥

3.
By Fourier transformation of

∆u = δ,

we get
−|ξ|2 u
b(ξ) = 1.
One solution is
1
b(ξ) = −
u .
|ξ|2
Observe that |ξ|12 ∈ L1loc (Rn ) if n ≥ 3, and that |ξ|12 is radial and homogeneous
of degree−2. Hence u is radial and homogeneous of degree 2−n. This implies
cn
u(x) = .
|x|n−2

Argument. If n = 3, then |ξ|12 ∈ L1 + L2 , and thus u is a function. If

1
n > 4, then |x|n−2 ∈ L1 + L2 , and we can argue as above using the inversion
1
theorem. When n = 4, we have u = ∈ L1 + L2 , and thus its Fourier
|x|2+
1
transform is a constant times and the statement follows by letting
|ξ|2−
 → 0.
An alternative way is to use Exercise 4 below.
Exercise 12.2. What is cn
Exercise 12.3. What happens if n = 2?
Exercise 12.4. Determine all radial distributions in Rn that are homogeneous of degree
α.
Hint. Consider first n = 1. Compute the derivate of hu, ϕt i = tα hu, ϕi with respect
to t.
Warning. Be careful when −α = n, n + 2, n + 4, . . ..

61
Exercise 12.5. What is the Fourier transform of fp|x|α in R?
Exercise 12.6. What is a reasonably definition of fp|x|α in Rn ? What is its Fourier
transform?
Exercise 12.7. Determine a fundamental solution to the heat equation.
Hint. Determine FE(x, t), where F is the Fourier transform with respect to x ∈ Rn .

Theorem 12.13. If u ∈ S 0 and v ∈ E 0 , then u ∗ v ∈ S 0 and

(u ∗ v)∧ = vb u
b.

Proof. If ϕ ∈ C0∞ , then

∨ ∨ ∨ ∨
u ∗ v(ϕ) = (u ∗ v)∗ ϕ (0) = u ∗ (v∗ ϕ)(0) = u((v∗ ϕ)∨ ) = u(v ∗ϕ).
∨
To see that u ∗ v ∈ S 0 , we need to show that v ∗ϕj → 0 in S when
ϕj → 0 in S . Let K be a compact neigborhood of av supp v and k the order
of v. Then
∨
X
|∂ β (v ∗ϕj )(x)| = |hvy , ∂ β ϕj (y − x)i| ≤ C sup |∂ β+γ ϕj (x − y)|.
y∈K
|γ|≤k

Thus
∨
X
(1+|x|2 )` |∂ β (v ∗ϕj )(x)| ≤ C (1+|x|2 )` sup |∂ β+γ ϕj (x−y)| → 0, j → ∞.
y∈K
|γ|≤k

To compute the Fourier transform, we observe that

∨
(u ∗ v)∧ (ϕ) = u ∗ v(ϕ)
b = u(v ∗ϕ) b
= (2π)−n u(bvb ∗ ϕ) v ϕ)∧ ) = u
b = u((b b(b
v ϕ) = vb u
b(ϕ).

2
∗n
Exercise 12.8. 1
Compute ( 1+x 2) and (e−x )∗n .
Exercise 12.9. H 7.1.6
Exercise 12.10. H 7.1.7
Exercise 12.11. H 7.1.9
Exercise 12.12. H 7.1.11
Exercise 12.13. H 7.1.18
Exercise 12.14. H 7.1.28

62
Chapter 13

The Paley-Wiener theorem

If u ∈ E 0 (Rn ), we know that û ∈ C ∞ and that

û(ξ) = u(e−ixξ ).

We shall show that û can be extended to an analytic function in Cn , ie.

û(ζ1 , . . . , ζn ) is an analytic function in each variable ζ1 , . . . , ζn . We start with
a version of the theorem for test functions.
Proposition 13.1.

(a) If φ ∈ C0∞ and supp φ ⊂ {x; |x| ≤ R}, then

Z
φ̂(ζ) = e−ixζ φ(x)dx
Rn

is an entire function with

|φ̂(ζ)| ≤ CN (1 + |ζ|)−N eR|Im ζ| (13.1)

for all N .

(b) Conversely, if φ̂ is entire and satifies (1), then φ ∈ C0∞ and

supp φ ⊂ {x; |x| ≤ R}.

Proof. (a) By differentiation under the integral sign, we see that φ̂ is analytic.
Furthermore, if ζ = ξ + iη,
Z
|φ̂(ζ)| ≤ exη |φ(x)|dx ≤ CeR|η| . (13.2)
|x|≤R

If we apply (2) to Dα φ, (1) follows.

63
 (b) Since φ̂ is rapidly decreasing we can differentiate under the integral
sign in the Fourier inversion formula. Hence φ ∈ C ∞ .
Again, by the rapid decrease of φ̂ we can use the Cauchy theorem to
change the contour of integration and integrate along {ζ; Imζi = ηi }. We get
 Z 
|φ(x)| = (2π)−n eix(ξ+iη) φ̂(ξ + iη)dξ  ≤ Ce−xη eR|η| .
 
Rn

If we let η = tx, we obtain

|φ(x)| ≤ Ce−t|x|(|x|−R) .
Thus if |x| > R, and we let t → ∞, we obtain φ(x) = 0. Hence
supp φ ⊂ {x; |x| ≤ R}.

For distributions we have

Theorem 13.2 (The Paley-Wiener thorem).

(a) If u is a distribution of order N with support in {x; |x| ≤ R}, then û is

an entire function and
|û(ζ)| ≤ C(1 + |ζ|)N eR |Im ζ| . (13.3)

(b) Conversely, if û is an entire function that satisfies (3) for some N , then
u is a distribution that is supported in {x; |x| ≤ R}.
Proof. (a) That û is entire follows since
∂ ∂  ∂ 
û(ζ) = u(e−ixζ ) = u (e−ixζ ) .
∂ζi ∂ζi ∂ζi
The last equality holds as
e−ix(ζ+ωi ) − e−ixζ ∂ −ixζ
→ (e ) i C ∞, ωi → 0.
ωi ∂ζi
To prove (3), we fix χδ ∈ C0∞ with χδ = 1 in a neighborhood of {x; |x| ≤
R} and supp χδ ⊂ {x; |x| < R + δ}. We can choose χδ such that kDα χδ k∞ ≤
Cδ −|α| . We obtain
|û(ζ)| = |u(e−ixζ )| = |u(χδ (x)e−ixζ )|
X
≤ CN sup |Dxα (χδ (x)e−ixζ )|
|α|≤N

≤ Ce(R+δ)|Imζ|
X
δ −|β| (1 + |ζ|)N −β .
|β|≤N

64
 1
If we let δ = 1+|ζ| , (3) follows.
(b) The polynomial growth of û implies that û, and hence also u, is in
S 0 . Let ϕδ ∈ S be an approximative identity and let uδ = u ∗ ϕδ . Then
uδ ∈ C ∞ , uδ → u as δ → 0, and

|ûδ (ζ)| = |û(ζ)ϕ̂(δζ)| ≤ CM,δ (1 + |ζ|)−M exp((R + cδ)|Imζ|).

Here we have used (3) and (1) in Proposition 1. If we apply Proposition 1

to uδ , we get supp uδ ⊂ {x; |x| ≤ (R + cδ)}. If we let δ → 0, we obtain supp
u ⊂ {x; |x| ≤ R}.
Exercise 13.1. Assume that u, v ∈ E 0 (Rn ) and that u ∗ v = 0. Show that then u = 0 or
v = 0. What happens if only one of u and v have compact support?
Exercise 13.2. H 7.1.40.

65
Chapter 14

Existence of fundamental
solutions

Let P (D) be a differential operator with constant coefficients in Rn . We shall

show that P (D) has a fundamental solution E.
Let us first make a formal computation. By Fourier transformation of
P (D)E = δ, we get P (ξ)E(ξ)
b = 1 and E(ξ)
b = P (ξ)−1 . Now
∨
∨ ∨
b ∨
hE, ϕi = hE, ϕi = (2π)−n hE, ϕi
b = (2π)−n hE,
b ϕi
b .
Hence it is natural to define E by
Z
−n
hE, ϕi = (2π) P (ξ)−1 ϕ(−ξ)dξ.
b
Rn

Then (formally)
Z
−n
hP (D)E, ϕi = hE, P (−D)ϕi = (2π) P (ξ)−1 (P (−D)ϕ)b(−ξ)dξ
Z Rn

= (2π)−n P (ξ)−1 P (ξ)ϕ(−ξ)dξ

b = ϕ(0) = hδ, ϕi.
Rn

However, this does not always work since P (ξ) may vanish. Therefore, we
will change the contour of integration and define hE, ϕi by an integral along
a set in Cn that contains no zero of P .
Theorem 14.1. Every linear differential operator with constant coefficients
has a fundamental solution E ∈ D 0 .
Proof. Let m = grad P . After a linear change of variables P is of the form
P (ξ) = Pξ0 (ξn ) = ξnm + Pm−1 (ξ 0 )ξnm−1 + . . . + P0 (ξ 0 )
= (ξn − α1 (ξ 0 )) . . . (ξn − αm (ξ 0 )).

66
Here ξ = (ξ1 , . . . , ξn ) = (ξ 0 , ξn ) and αi (ξ 0 ) are the zeros of Pξ0 (ξn ). We
can choose φ(ξ 0 ) ∈ R such that |φ(ξ 0 )| ≤ m + 1 and |φ(ξ 0 ) − αi (ξ 0 )| ≥
|φ(ξ 0 ) − Im αi (ξ 0 )| ≥ 1 for i = 1, 2, . . . , m. Define hE, ϕi, when ϕ ∈ D, by
Z Z
−n 0
hE, ϕi = (2π) dξ P (ζ)−1 ϕ(−ζ)dζ
b n.
Rn−1 Im ζn =φ(ξ )
0

By the Paley-Wiener theorem, ϕ(ζ)

b is an entire function and

C X
|ϕ(ζ)| ≤ N
kDα ϕk∞ .
(1 + |ζ|)
b
|α|≤N

Furthermore |P (ζ)−1 | ≤ 1, and hence, if N is large enough, we get

X
|hE, ϕi| ≤ C kDα ϕk∞ .
|α|≤N

Thus E ∈ D 0 . Finally, we see that

hP (D)E, ϕi = hE, P (−D)ϕi

Z Z
−n 0
= (2π) dξ P (ζ)−1 (P (−D)ϕ)∧ (−ζ)dζn
R n−1 Im ζn =φ(ξ )
0
Z Z
= (2π)−n dξ 0 ϕ(−ζ)dζ
b n
Rn−1 Im ζn =φ(ξ0 )
Z Z
−n 0
= the Cauchy theorem = (2π) dξ ϕ(−ξ)dξ
b n
Z Rn−1 R

= (2π)−n ϕ(ξ)dξ
b = ϕ(0) = hδ, ϕi.
Rn

Exercise 14.1. Determine a fundamental solution to the Schrödinger equation

n
X
(Dt − Dx2i )E = δ.
1

(D = −i∂)
Hint. See Exercise 10.14 and the hint to Excercise 12.7

67
Chapter 15

Fundamental solutions of
elliptic differential operators

Let P (D) be a differential operator with constant coefficients. We write the

polynomial P as
P = Pm + Pm−1 + . . . + P0 ,
where Pk is a homogeneous polynomial of degree k. The operator P (D) is
called elliptic if Pm (ξ) 6= 0 for ξ 6= 0, ξ ∈ Rn .

Example 15.1. ∆ and ∂¯ are elliptic. The heat and wave operators are not
elliptic. 2

Theorem 15.1. Let P (D) be an elliptic differential operator Then there is

a distribution E ∈ S 0 (Rn ) such that sing supp E = {0} and P (D)E = δ − ω,
for some ω ∈ S (Rn ).

Corollary 15.2. If P is elliptic, then P is hypoelliptic.

Proof of the corollary. We shall show that u is C ∞ if P (D)u is. If u has

compact support, we have u = δ ∗u = (P (D)E +ω)∗u = E ∗P (D)u+ω ∗u ∈
C ∞ . The general case follows by considering ψn u where ψn ∈ C0∞ with ψn = 1
on {|x| ≤ n} (compare Theorem 9.6)

Proof of the theorem. Since P is elliptic, |Pm (ξ)| ≥ δ > 0 when |ξ| = 1. By
homogenity this implies
|Pm (ξ)| ≥ δ|ξ|m .
Hence, if |ξ| > R where R is large enough,

|P (ξ)| ≥ c|ξ|m .

68
Take χ ∈ C0∞ (Rn ) with χ(ξ) = 1 if |ξ| ≤ R. Then (1 − χ)P −1 is bounded
and hence a tempered distribution. Thus we can define E ∈ S 0 (Rn ) by

b = 1 − χ.
E
P
Then,
1−χ
(P (D)E)∧ = P E
b=P = 1 − χ = δb − χ.
P
b = χ, then ω ∈ D
If we define ω by ω b ⊂ S and P (D)E = δ − ω. It remains
to show that E ∈ C ∞ (Rn \ {0}). Observe that

1 − χ(ξ)
(xβ Dα E)∧ (ξ) = cDβ (ξ α ) = O(|ξ|−|β|−m+|α| ), |ξ| → ∞.
P (ξ)

If we choose |β| large enough, we get (xβ Dα E)∧ ∈ L1 . Thus xβ Dα E ∈ C

and hence Dα E ∈ C(Rn \ {0}), and the proof is complete.

69
Chapter 16

Fourier series

Let u be a distribution that is periodic with period 2π in each variable, i.e.

hu, τ2πk ϕi = hu, ϕi,

if k ∈ Zn . Intuitively u is determined by its ”values” on

T = {x; 0 ≤ xi < 2π}.

Lemma 16.1. If u is periodic, then u ∈ S 0 .

Proof. Let ψ ∈ C0∞ with 0 ≤ ψ ≤ 1 and ψ = 1 on T . Put
X
ψ(x)
e = ψ(x − 2πk).
k∈Zn

Then ψe is a periodic C ∞ -function with ψe ≥ 1. Thus φ = ψ/ψe ∈ C0∞ and

X
φ(x − 2πk) = 1.
k

If ϕ ∈ D, then
X
hu, ϕi = hux , φ(x − 2πk)ϕ(x)i = a finite sum =
k
X
= hux , φ(x − 2πk)ϕ(x)i = periodicity =
k
X X
= hux , φ(x)ϕ(x + 2πk)i = hux , φ(x) ϕ(x + 2πk)i.
k k

But if ϕj → 0 in S , then φ(x) k ϕj (x + 2πk) → 0 i D. (Prove that!) Hence

P
the right hand side defines an extension of u to S 0 .

70
 To compute u
b, we first show the following result.
Theorem 16.2 (The Poisson summation formula). If ϕ ∈ S , then
X X
ϕ(2πk)
b = ϕ(k).
k∈Zn k∈Zn
P
Proof. Let u = k∈Zn δ2πk . Then δ2πl ∗ u = u, since δ2πl ∗ δ2πk = δ2π(k+l) .
(Prove that!)
Hence
(e−2πilξ − 1)b
u = 0.
But e−2πilξ − 1 6= 0 if ξ ∈ / Zn , and consequently u b is supported on Zn .
By choosing different l, we see that close to the origin we have ξi u b = 0,
b = cδ0 there. Furthermore e−ikx u = u, and hence u
i = 1, 2, . . . , n. Thus u b is
invariant under translation by integers. From this, we obtain
X
u
b=c δk .
k∈ZN

This means that X X

ϕ(2πk)
b =c ϕ(k).
k∈Zn k∈Zn
If we replace ϕ with a translation of ϕ, we get
X X
2πikx
ϕ(2πk)e
b =c ϕ(k + x).
k∈Zn k∈Zn

Integration over {x; 0 ≤ xi < 1} gives

Z
ϕ(0)
b =c ϕ(x)dx = cϕ(0)
b .
Rn

Thus c = 1 and the proof is complete.

Let us return to the computation of u
b when u is periodic. Using the
Poisson summation formula on ϕ(y) = ψ(y)e−ixy , we get, as ϕ(ξ)
b = ψ(x b + ξ),
X X X X
ψ(x
b + 2πk) = ϕ(2πk)
b = ϕ(k) = e−ixk ψ(k).
k k k k

From the proof of Lemma 1, we have

X
hb
u, ψi = hu, ψi
b = hu, φ(x) ψ(x
b + k)i
k
X
= hu, φ(x) e−ixk ψ(k)i
k
X
= ψ(k)hu, φ(x)e−ixk i.
k

71
 P
Hence u
b= k ck δk ,where

ck = hu, φ(x)e−ixk i.

In particular, if u is an integrable function on T , we have

Z
−ixk
ck = hu, φ(x)e i= u(x)φ(x)e−ixk dx
Rn
XZ
= u(x − 2πj)φ(x − 2πj)e−i(x−2πj)k dx
j T
Z X Z
−ixk
= u(x)e φ(x − 2πj)dx = u(x)e−ixk dx.
T j T

Hence ck are ”our old” Fourier coefficients. The inversion theorem implies
that
1 X ikx
u(x) = ck e in S 0 .
(2π)n k
If u ∈ C l , then ck = O(|k|−l ), |k| → ∞, and the sum is uniformly convergent
if l > n. Thus we have proved
Theorem 16.3. If u ∈ C l (Rn ), l > n, and u is periodic with period 2π in
each variable, then
1 X ixk
u(x) = ck e ,
(2π)n k
where the series is uniformly convergent.
We finish this chapter by proving
Theorem 16.4 (The Plancherel theorem). If u ∈ L2 (T ) with Fourier coef-
ficients ck , then
1 X ixk
u(x) = ck e in L2 ,
(2π)n
and Z
1 X
|u|2 dx =
n
|ck |2 .
T (2π)
2 1 ixk
P P
Conversely, if |ck | < ∞, then u(x) = (2π) n k ck e is a function in
2
L (T ) with Fourier coefficients ck .
Proof. If u ∈ C n+1 , the series is uniformly convergent, and we get
Z Z
2 1 X ix(k−l) 1 X
|u| dx = 2n
c k c̄ l e dx = n
|ck |2 .
T (2π) k,l T (2π)

72
As C n+1 is dense in L2 , we can extend this to u ∈ L2 : Take un ∈ C n+1 , un → u
in L2 . Then, also un → u in S 0 and u bn → ub in S 0 . But also, by the isometry,
bn is a Cauchy sequence in l2 . This implies u
u bn → u b in l2 . Hence
Z Z
2 1 X 1 X
|u| dx = lim |un |2 dx = lim |c k (un )|2
= |ck |2 .
T n→∞ T n→∞ (2π)n (2π) n
k

|ck |2 < ∞, let

P
Conversely, if
1 X ixk 1 X
u(x) = c k e and uN (x) = ck eixk .
(2π)n k (2π)n
|k|≤N

Then, uN → u in L2 and S 0 , and we get

X
u
b = lim u
bN = c k δk .
N →∞

Remark 16.5. If u is a function with period t, then ut (x) = u( 2πx

t
) has
period 2π. Using this, we can generalise Fourier series to functions with
arbitrary period. 2
Exercise 16.1. H 7.2.1
Exercise 16.2. H 7.2.5
Exercise 16.3. H 7.2.8
P∞ 1
P+∞ 1
P∞ (−1)n
Exercise 16.4. Compute a) −∞ 1+n2 b) −∞ (n+a)2 and c) n=0 (2n+1)3 .

73
Chapter 17

Some applications

17.1 The central limit theorem

Let X, X1 , X2 , . . . be independent identically distributed stochastic variables
with E[X] = m and Var[X] = σ 2 . Then
 X + X + . . . + X − nm Z x
1 2 n
 1 1 2
lim P √ ≤x = √ e− 2 y dy. (17.1)
n→∞ σ n 2π −∞

Some background: To a stochastic variable X we associate a probability

measure µ on R (we write X ∼ µ) by
Z x
P (X ≤ x) = dµ(y).
−∞

If µ1 are µ2 probability measures, we define a new probability measure µ1 ∗µ2

by ZZ
hµ1 ∗ µ2 , ϕi = ϕ(x + y)dµ1 (x)dµ2 (y).
R2

Then (µ1 ∗µ2 )∧ = µ b2 . (Show that!) If X ∼ µ1 and Y ∼ µ2 are independent,

b1 µ
then X + Y ∼ µ1 ∗ µ2 .
Proof. We may assume that m = 0 and σ = 1. Let
X1 + . . . + Xn
Sn = √
n

and
µn∗ = µ ∗ . . . ∗ µ .
| {z }
n times

74
Then Sn ∼ µn , where
Z  
x
hµn , ϕi = ϕ √ dµn∗ (x),
R n
and   n
ξ
µ b √
cn (ξ) = µ .
n
Since Var[X] < ∞, Z
µ
b(ξ) = e−ixξ dµ(x)
R
is a C 2 -function with
b 0 (0) = −im = 0 and µ
µ b 00 (0) = −σ 2 = −1.
Thus
1
b(ξ) = 1 − ξ 2 + o(ξ 2 ),
µ ξ → 0,
2
and
  n   2
ξ 1 2 ξ n 1 2
µ b √
cn (ξ) = µ = 1− ξ +o → e− 2 ξ , n → ∞,
n 2n n
for each fixed ξ. But, since |b
µ(ξ)| ≤ 1, we get, by dominated convergence,
that
1 2
cn (ξ) → e− 2 ξ in S 0 .
µ
Hence Fourier inversion implies that
1 1 2
µn → √ e − 2 x
2π
in S 0 , and hence also in D 0 . But µn are positive measures, and by Theo-
rem 7.4
1 1 2
µn → √ e − 2 x
2π
weakly, and hence we obtain (1).

17.2 The mean value property for harmonic

functions
If u ∈ C ∞ is harmonic in a neighborhood of {|x| ≤ 1}, then
Z
1
u(0) = u(y)dσy).
ωn S n−1

75
Remark 17.1. By Weyl’s lemma, the assumption that u ∈ C ∞ is unneces-
sary. 2
Proof. Define a distribution Λ by
Z
hΛ, ϕi = ϕ(y)dσ(y) − ωn ϕ(0).
S n−1

Then Λ ∈ E 0 , and hence Λ b is an entire function. Furthermore, Λ, and

therefore also Λ,
b is radial. Hence Λ(ζ)
b = G(|ζ|), where G(t) = Λ(t,
b 0, . . . , 0)
2
is holomorphic. Also, G is even, so G(z) = F (z ) for some entire function F .
Since F (0) = Λ(0)
b = Λ(1) = 0,

Λ(ξ)
b F (|ξ|2 ) − F (0)
=
|ξ|2 |ξ|2
is the restriction of an entire function. By the Paley-Wiener theorem, there
is a distribution µ ∈ E 0 with µb(ξ) = −F (|ξ|2 )/|ξ|2 , and so

(∆µ)b(ξ) = −|ξ|2 µ
b(ξ) = Λ(ξ).
b

Hence ∆µ = Λ, which gives

hΛ, ui = h∆µ, ui = hµ, ∆ui = hµ, 0i = 0.

17.3 The Heisenberg uncertainty principle

If f ∈ L2 (R), then
r
π
kxf (x)k2 kξ fb(ξ)k2 ≥ kf k22 , (17.2)
2
with equality only if f (x) = exp(−kx2 ), k > 0.
Quantum mechanical background: The state of a particle is described by
a function ψ ∈ L2 (R) with kψk2 = 1. We interprete
Z
|ψ|2
E

as the probability that the particle is in the set E. An observable quantity

A is a symmetric operator on a suitable subspace of L2 . The mean value of
A in the state ψ is Z
E[A] = Aψ · ψ̄ = hAψ, ψi.

76
That A is symmetric means that A = A∗ and hence we have

hAψ, ψi = hψ, A∗ ψi = hψ, Aψi = hAψ, ψi.

Thus the mean value is real.

Example 17.2.
a) Position. Aψ(x) = xψ(x)
b) Momentum. Bψ = 2πiψ 0 . 2

We have
Z Z Z Z
0 b 2 dξ.
E[B] = Bψ · ψ̄ = 2πi ψ ψ̄ = Plancherel = ξ ψ(ξ)ψ(ξ)
b = ξ|ψ(ξ)|

b 2 as the density of the momentum.

Hence we can interprete |ψ(ξ)|
The general form of the Heisenberg uncertainty principle is

1 2
E[(A − E[A])2 ]E[(B − E[B])2 ] ≥ E[AB − BA] (17.3)
4

for arbitrary A and B.

Exercise 17.1. Show that if A and B are position and momentum, then AB − BA =
−2πi.

Exercise 17.2. Prove that (2) implies (3), when A and B are position and momentum.

Proof. If f ∈ S , then

kxf (x)k2 kξ fb(ξ)k2 = kxf (x)k2 kfb0 (ξ)k2 = Parseval =

√ √ Z
= 2πkxf (x)k2 kf (x)k2 ≥ Schwartz ≥ 2π |xf (x)f 0 (x)|dx
0

√ Z 1 0

≥ (|xz̄w| ≥ x Re z̄w) ≥ 2π x f (x)f (x) + f (x)f (x) dx 0
2
r Z r Z r
π 2 0 π 2 π
= x(|f (x)| ) dx = Integration by parts = |f | dx = kf k22 .
2 2 2

The proof that the theorem holds for functions in L2 , and the statement of
equality is left to the reader.

77
17.4 A primer on Sobolev inequalities
A benefit of the theory of distributions is that we can find solutions to prob-
lems that has no classical solutions. But we often want our solutions to be
nice functions. Therefore it is natural to ask the question
When is a distributional solution a function?
The theory of Sobolev spaces gives us a method to answer that question.
We start with the simplest result in Sobolev theory,

The Sobolev L1 -inequality Let f be an integrable function on

R . Assume that the distributional derivatives ∂ α f also are integrable for all
n

|α| ≤ n. Then f is a bounded continuous function and

X
kf k∞ ≤ k∂ α f k1 . (17.4)
|α|≤n

If furthermore ∂ α f are integrable for all |α| ≤ n+k, then f is a C k -function.

Proof. We start with the case n = 1 where we shall show that

kf k∞ ≤ kf k1 + kf 0 k1 . (17.5)

If ϕ ∈ C0∞ , then
x x ∞
Z  Z Z
0 0
|ϕ0 (t)|dt .
 
|ϕ(x)| =  ϕ (t)dt ≤ |ϕ (t)|dt ≤
−∞ −∞ −∞

This implies that

kϕk∞ ≤ kϕ0 k1 . (17.6)
This inequality is sharper than (5), but we have obtained it asssuming two
strong extra conditions, C ∞ and compact support. (The function 1 shows
that (6) can not be true in general.)
If f ∈ C ∞ is not compactly supported, we choose a sequence of cut off
functions χn ∈ C ∞ , with χn = 1 as |x| ≤ n and kχ0n k∞ ≤ 1. If we apply (6)
to ϕ = χn f ,we get

kχn f k∞ ≤ k(χn f )0 k1 ≤ kχ0n f k1 + kχn f 0 k1 ≤ kf k1 + kf 0 k1 .

Since n is arbitrary, (5) is proved for C ∞ -functions.

If f is not C ∞ , we let φδ be an approximative identity. Then fδ =
φδ ∗ f ∈ C ∞ and we can apply (5) to fδ . We get, as kφδ ∗ f k1 ≤ kf k1 och
k(φδ ∗ f )0 k1 = kφδ ∗ f 0 k1 ≤ kf 0 k1 , that

kφδ ∗ f k∞ ≤ kf k1 + kf 0 k1 .

78
But φδ ∗ f → f a.e. and we have proved (5) in the general case.
To finish the proof, we apply (5) to f − fδ , to obtain

kf − fδ k∞ ≤ kf − φδ ∗ f k1 + kf 0 − φδ ∗ f 0 k1 → 0, δ → 0 .

Thus fδ → f uniformly, and f is a continuous function.

The last claim follows by applying this argument to the functions ∂ i f ,
i = 1, 2, . . . , k.
The argument when n ≥ 2 is similar. The case n = 2 shows how but
without too cumbersome notation. If ϕ ∈ C0∞ , we now get
Z x Z y  Z ∞Z ∞
 (1,1)

|ϕ(x, y)| =   ∂ ϕ(s, t)dsdt ≤ |∂ (1,1) ϕ(s, t)|dsdt
∞ ∞ ∞ ∞

and hence
kϕk∞ ≤ k∂ (1,1) ϕk1 .
When we apply this to χn f , f ∈ C ∞ , we get, as ∂ (1,1) (χn f ) = ∂ (1,1) χn f +
∂ (1,0) χn ∂ 0,1 f + ∂ (0,1) χn ∂ 1,0 f + χn ∂ (1,1) f , that

kf k∞ ≤ kf k1 + k∂ 1,0 f k1 + k∂ 0,1 f k1 + k∂ (1,1) f k1 , f ∈ C ∞ .

The rest of the argument works exactly the same as in the case n = 1.

Remark 17.3. The proof shows that it is enough to consider α = (α1 , . . . , αn ),

where each αi is either 0 or 1, in the sum (4). 2

The Sobolev L2 -inequality Let f ∈ D 0 (Ω), where Ω is an open set

in R , and let r and k ≥ 0 be integers. If ∂ α f ∈ L2loc for all α, 0 ≤ |α| ≤ r
n

where r > k + n2 , then f ∈ C k (Ω).

Proof when n = 1 and k = 0.

The assumption means that f ∈ L2loc and f 0 ∈ L2loc . Let ω be an open
set, ω ⊂⊂ Ω, and take χ ∈ C0∞ (Ω) with kχk∞ ≤ 1, kχ0 k∞ ≤ 1 and χ = 1 in
a neighborhood of ω. Define F (x) = Fω (x) = χ(x)f (x). (F = 0 outside the
support of χ.) Since F ∈ L2 (R) and F 0 = χ0 f + χf 0 ∈ L2 (R), the Parseval
identity implies that
Z Z
2
|F | dξ < ∞ and
b ξ 2 |Fb|2 dξ < ∞ .
R R

79
Hence Z
(1 + |ξ|)2 |Fb|2 dξ < ∞ .
R

The Cauchy-Schwartz inequality implies

Z 2 Z 2
dξ
|Fb|dξ= (1 + |ξ|)|F |
b
R R 1 + |ξ|
Z Z
dξ
≤ (1 + |ξ|)2 |Fb|2 dξ 2
<∞.
R R (1 + |ξ|)

Thus Fb is integrable and hence F is continuous. As ω is an arbitrary open

subset of Ω, it follows that f ∈ C(Ω).

The general case.

00 2 2b 2
Z k = 1, we also know that F ∈ L . Thus ξ F (ξ) ∈ L
If n = 1 and
and we have (1 + |ξ|)4 |Fb|2 dξ < ∞. By the Cauchy inequality, this gives
Z R

(1+|ξ|)|F |dξ < ∞. In particular, ξ Fb(ξ) is integrable and F 0 is continuous.

b
R
The case for arbitrary k follows in the same way.
If n > 1, the condition on ∂ α f implies that ξil Fb(ξ) ∈ L2 , l ≤ r. Using the
inequality (1 + |ξ|)2l ≤ Cl (1 + ξ12l + . . . + ξn2l ), and the Cauchy inequality, we
obtain (1 + |ξ|)k Fb ∈ L1 and hence F ∈ C k .

Sobolev spaces
Let us abstract the ideas in the proof of the L2 -inequality. We saw that
if f and its derivatives up to order r are in L2 , then (1 + |ξ|)r fb ∈ L2 or
equivalently (1 + |ξ|2 )r/2 fb ∈ L2 . In this condition, r may be an arbitary real
number, and we can make the following definition.

Definition 17.4. A distribution f ∈ D 0 (Rn ), is in the Sobolev space H s (Rn ),

s ∈ R, if
kf kH s = k(1 + |ξ|2 )s/2 fb(ξ)kL2 < +∞

Proposition 17.5. If f ∈ H r (Rn ), where r > s + n2 , then

(1 + |ξ|)s fb ∈ L1 (Rn ).

Remark 17.6. If s = k is a non-negative integer, this implies that f ∈ C k .

80
Proof. We have (1 + |ξ|2 )r/2 fb(ξ) ∈ L2 . Hence, by the Cauchy inequality, we
get
Z
(1 + |ξ|2 )s/2 fb(ξ)dξ
Rn
Z
dξ
= (1 + |ξ|2 )r/2 fb(ξ) ≤ Ckf kH r ,
Rn (1 + |ξ|2 )(r−s)/2

as (1 + |ξ|2 )−(s−r) is integrable when r − s > n2 .

17.5 Minkowski’s theorem

Let B be a convex set in Rn that is symmetric at the origin. If |B| ≥ 2n ,
then B contains more than one lattice point.
Proof. We assume that 0 is the only lattice point in B, and show that this
implies that |B| < 2n . Let f = χB ∗ χB . Since B is symmetric, χ
bB is real.
2 2
Hence f = (b
b χB ) = |b
χB | .
We observe that if f (2k) 6= 0, ie.
Z
f (2k) = χB (2k − x)χB (x)dx 6= 0 ,

then there is x ∈ B with 2k−x ∈ B. But then we have k = 21 (2k−x)+ 12 x ∈ B,

as B is convex. Hence if f (2k) 6= 0 we have k = 0. Furthermore
Z Z
f (0) = χB (−x)χB (x)dx = |χB |2 dx = |B|.

The Poisson summation formula, applied to the lattices (2Z)n and (πZ)n ,
gives X X
f (2j) = 2−n fb(πj) .
j∈Zn j∈Zn

Hence
X X
|B| = f (0) = f (2j) = 2−n fb(πj)
j j
!
X 2
X 2
= 2−n χB (πj)| = 2−n |B|2 +
|b |b
χB (πj)| .
j j6=0

If we can show that X

χB (πj)|2 > 0,
|b
j6=0

81
we obtain |B| > 2−n |B|2 , or |B| < 2n , and we are done.
bB (πj) = 0 when j 6= 0, then
But if χ
X
χ(x) = χB (x + 2j)
j

is constant. This follows from the Poisson summation formula since

X X
χ(x) = τ−x χB (2j) = 2−n bB (πj) = 2−n χ
eiπxj χ bB (0).
j j

But this is a contradiction as

χ(0) = 1 6= 0 = χ(1, 0, . . . , 0).

Exercise 17.3. The proof is ”wrong”. Why? Correct it!

82
Index

C0∞ , 6 Heisenberg uncertainty principle, 76

L2 , 55 Hilbert transform, 60
D 0 (Ω), 11 homogeneous, 23, 61
D(Ω), 11 hypoelliptic, 46, 68
E 0 (Ω), 31
S , 49 inversion theorem, 50
S 0 , 50 Laplace operator, 29, 61
approximate identities, 8 mean value property, 75
measure, 11
Cauchy sequence, 32 Minkowski’s theorem, 81
Central limit theorem, 74
compact support, 31 odd, 59
complete, 32
Paley-Wiener theorem, 63, 64
convergence, 32, 50
Parseval formula, 52, 55
convolution, 8, 36
partition of unity, 13
derivative, 17 periodic, 70
Dirac measure, 18 Plancherel theorem, 52, 72
distribution, 11 Poisson summations formula, 71
division problem, 25 positive distribution, 13
principal value, 24
elliptic, 68
rapidly decreasing functions, 49
even, 59
regularisation, 8, 38
finite part, 21 Schwartz space, 49
Fourier coefficient, 47, 72 singular support, 44
Fourier series, 48, 70 Sobolev inequalities, 78
Fourier transform, 47, 48 Sobolev spaces, 80
fundamental solution, 28, 43, 66 support of a distribution, 14
harmonic function, 38 tempered distribution, 50
harmonic functions, 75 translation, 40
heat eqution, 29
Heaviside function, 17 Weyl’s lemma, 39

83