Partial Differential Equations 1

Lecture Notes

Martin Olbrich and Guendalina Palmirotta

typed by Léa Mia Micard (student)

Master in Mathematics – Winter Semester 2022

Department of Mathematics, University of Luxembourg

June 27, 2024

1
Contents
1 Introduction and motivation 3
1.1 Reminder and Motivation : Banach spaces, Hilbert spaces, Fourier series . . . . . . . . 5
1.1.1 Hilbert spaces and orthonormal systems . . . . . . . . . . . . . . . . . . . . . . 8
1.1.2 Classical Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Definition of the Fourier transform and Schwartz functions . . . . . . . . . . . . . . . . 16
1.3 Convolution, density theorems and Fourier transform on L2 (Rn ) . . . . . . . . . . . . . 23
1.4 Continuous Linear maps and continuous extensions . . . . . . . . . . . . . . . . . . . . 27
1.5 An application to constant coefficient differential operators . . . . . . . . . . . . . . . . 30

2 The classical differential equations 32

2.1 The Laplace/Poisson equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 Typical problems connected with the Laplacian . . . . . . . . . . . . . . . . . . . . . . 34
2.3 Gradient, divergence, volume forms and integral formulas . . . . . . . . . . . . . . . . 38
2.4 Volume of spheres: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5 The solution of the Dirichlet problem for the ball U = B(0, R) ⇢ Rn . . . . . . . . . . 43
2.6 Applications of Theorem 1 to harmonic functions . . . . . . . . . . . . . . . . . . . . . 46

2
1 Introduction and motivation
Partial differential equations (PDE) (functions of several variable) are in contrast to ordinary differen-
tial equations (ODE) (functions of one variable).

Ordinary differential equations (ODE) - Recall

Let I, J ⇢ R be open intervals and ⌦ ⇢ R an open subset. We look for sufficiently often differentiable
functions f : I ! J ⇢ R satisfying a relation of the form

F (x, f (x), f 0 (x), f 00 (x), · · · , f (n) (x)) = 0 (1.1)

t
I t | {z }
I 2⌦⇢Rn

where F is a certain function (at least continuous)

F : I ⇥ J ⇥ ⌦ ! R.

Eq. (1.1) is the (general) ordinary differential equations of order k 2 N = {1, 2, 3, · · · } in the implicit
form. If possible, one tries to bring (1.1) into explicit form, i.e. one isolates the highest derivative on
one side:
f (k) (x) = G(x, f (x), f 0 (x), · · · , f (k 1) (x)). (1.2)
In order to have a chance to get unique solutions, one usually fixes initial conditions

f (x0 ) = y0 , f 0 (x0 ) = y1 , · · · , f (k 1)
(x0 ) = yk 1; (x0 , y0 , y1 , · · · , yk 1) given. (1.3)

A similar thing works for systems of equations:

f = (f1 , · · · , fm ), f : I ! U ⇢ Rm , ⌦ ⇢ Rk m
and F : I ⇥ U ⇥ ⌦ ! Rm .

By reduction of order any k-th order ODE is equivalent to a system of first order ODE. This gives a
general solution theory (of systems) of ODE in the form (1.2). For instant, if G satisfies a certain Lip-
schitz condition (e.g. if G is continuously differentiable), then (1.2) together with the initial conditions
has a unique local (defined at least in a smart neighborhood of x0 ) solution.
The situation become even easier if F (and hence G) depends linearly on f (x), f1 (x), fk 1 (x) (and
fk (x)). This means that (1.1) has the form
k
X
ae (x)f (l) (x) + b(x) = 0, 8x 2 I,
l=0

where al are the “coefficients” of the equation. Then there is a global solution (defined outside the
zeroes of ak (x)). If the coefficients are constants, solutions can be determined explicitly.

Partial differential equations (PDE)

Consider similar relations between the partial derivatives up to a certain order for functions of several
orders. Let U ⇢ Rn be open interval and K = R, C. We look for twice continuously differentiable
functions f : U ! K, i.e. f 2 C 2 (U ):
Xn
@if
= 0. (1.4)
@xi
i=0
P @2
Solution of (1.4) are called harmonic functions. One sets = ni=1 @x 2 and writes (1.4) as
i

f = 0. (1.5)

Here “delta” is called the Laplace operator. Eq. (1.4) (or (1.5)) is called a linear PDE with constant
coefficients. Instead of (1.5) we can also consider for given g : U ! C the inhomogeneous equation

f = g. (1.6)

3
The general PDE order k could be written of the form
⇣ @f @f @2f @kf ⌘
F x, f (x), (x), · · · , (x), (x), · · · , k (x) = 0. (1.7)
@x1 @xk @x1 @x2 @xn

In other words, using multi-indices ↵ = (↵1 , · · · , ↵n ), with ↵i 2 N0 = {0, 1, 2, · · · }, we denote the

partial derivatives by
@ (k)
D ↵ = ↵1 ↵2 ,
@x1 @x2 · · · @x↵kk
P
where |↵| = ni=1 ↵i . Note that F depends on x and D↵ f (x) for all ↵ with |↵|  k (by convention for
↵ · (0, · · · , 0) we take D↵ f = f ). Similar as for ODE, Eq. (1.7) is called linear if f depends linearly
from all the D↵ f , |↵|  k. Observe that (1.5) and (1.6) are linear.

Examples (Examples of non-linear equations).

@f @f
(a) @x1 · @x2 = 0.
2
(b) det(( @x@i @x
f
j
)i,j=1,...,n ) = g.

(c) f (x) = f (x)k is non-linear for k 6= 1, where is a fixed number.

Instead of initial conditions, one often considers a kind of “boundary conditions”: N ⇢ U “submanifold
of a dimension 1” (e. g. an affine hyperplane). We require conditions like f |N = f0 , where f0 is given
a function on N .

Differences to ODE
• There is no general solution theory for the general equation (1.7). One exception: if F is an-
alytic (given by convergent power series in n variables) then there is a general theorem, the
Cauchy–Kovalevskaya theorem.

• Non-linear equations are very difficult (will play no role in this course).

• There is no easy way to reduce higher order equations to systems of first order equations. Many
of the interesting equations have order 2 and should be studied directly.

In this course, we are mainly concerned with linear PDE with constant coefficients
X
a↵ f (x) = g(x),
|↵|k

where f is unknown, and a↵ 2 C and g are both given. In particular, we will study

f = g (Laplace (or Poisson) equation)

@f
(t, x) = x f (t, x) (heat equation) (1.8)
@t
@2f
(t, x) = x f (t, x) (wave equation), (1.9)
@t
where for the two last, we consider the function f : I ⇥ Rn ! C.

Tools:

• For the classical equations, a lot of “computational tricks” are known.

• Separation of variables and reduction of order.

4
 • For linear differential operators D, we can try to use methods of functional analysis (Fourier
transform, spectral theory), i.e. consider D as a linear operator between (infinite - dimensional)
vector space of functions, e.g.
D : C k (U ) ! C 0 (U ).

Plan of the course. There will be two parts:

1. Banach and Hilbert spaces, Fourier transform with applications to PDE.
2. The classical equations (1.5), (1.8) and (1.9).

1.1 Reminder and Motivation : Banach spaces, Hilbert spaces, Fourier series
Let V be a vector space over K = R or C, not necessarily finite - dimensional. In fact, we are mainly
interested in vector spaces of real or complex-valued functions on some set X (there are vector spaces
by pointwise addition and multiplication by scalars). A norm on V is a map

k·k:V ! [0, 1)

satisfying
1. kvk = 0 , v = 0, v 2V,
2. k vk = | | kvk, v 2 V, 2 K,
3. (triangle inequality) kv + wk  kvk + kwk, v, w 2 V .
p
Here, for 2 C, | | = Re ( )2 + Im ( )2 = . A pair (V, k · k) of a vector-space with a norm is
called a normed vector-space. A norm k · k on a vector-space defines a distance function on V by

d(v, w) = kv wk.

With this function, V be comes a metric space. In particular we can speak of

• convergent sequence in V : vi ! v : ! kv wk ! 0 in [0, 1).
i!1 i!1

• Cauchy sequences: 8" > 0 9N 8i, j N : kvi vj k  ". Every convergent sequence is Cauchy.
The converse is not always true, even in the context of normed vector spaces. But it is always
true for finite-dimensional normed vector spaces.
Recall: A metric space is called complete, if every Cauchy sequence in it converges.
Definition 1 (Banach space). A Banach space is a normed vector space which is complete as a metric
space, i.e. every Cauchy sequence in (v, k · k) converges.
Let V be a vector space over K = R or C. For 2 C, we denote by 2 C the complex conjugate. A
scalar product on V is a function
h·, ·i : V ⇥ V ! K
with
• hv1 + v2 , wi = hv1 , wi + hv2 , wi, v1 , v2 , w 2 V
• h v, wi = hv, wi, v, w 2 V, 2R
• hw, vi = hv, wi.
These 3 conditions mean: h·, ·i is a symmetry bilinear form (K = R) or a hemiliton form (K = C). In
addition, we require the following is positive definite

hv, vi 0, 8v 2 V,

if hv, vi = 0 =) v = 0. A pair (V, h·, ·i) of a vector space and a scalar product on it is called a vector
space with scalar product.

5
Lemma 1 (Cauchy-Schwarz inequality). Let (V, h·, ·i) be a vector space with scalar product. Then,
|hv, wi|2  hv, vi · hw, wi.
p
Given (V, h·, ·i) a vector scalar product. We can define norm on it by kvk = hv, vi. In terms of k · k,
the Cauchy-Schwarz inequality can be written:
|hv, wi|  kvk · kwk.
Lemma 2. Let (V, h·, ·i) be a vector space with scalar product. Then k · k defined as above is a norm.
Hence (V, k · k) is a normed vector space.
p
Proof. We have to check the axioms (1), (2), (3) for kvk = hv, vi.
(1) and (2) are checked immediately. Only (3) (triangle inequality) is more involved. For it, we use
Cauchy-Schwarz (Lem. 1):
kx + yk2  hx + y, x + yi = hx, xi + hy, yi + hx, yi + hy, xi
= kxk2 + kyk2 + 2 · Re (hx, yi)
 kxk2 + kyk2 + 2 · |hx, yi|
Cauchy-Schwarz
 kxk2 + kyk2 + 2kxk kyk
= (kxk + kyk)2 .
Hence kx + yk  kxk + kyk.

Definition 2 (Hilbert space). A Hilbert space is a vector space with scalar product (V, h·, ·i) such that
the corresponding normed vector space (V, k · k) is complete (i.e. every Cauchy sequence converges).
Examples (Banach and Hilbert spaces). m
• Finite dimensional examples: Let V = Cn .
P 1
1. Consider V with norms (p 2 [1, 1)) kXkp := ( ni=1 |xi |p ) p , X = (x1 , · · · , xn ), xi 2 C.
For p = 2, we get the usual Euclidean norm. In this case, we have a Hilbert space with
n
X
scalar product hx, yi := x i yi .
i=1

Note that for p 6= 1, 2 the triangle inequality is not very easily checked. One needs the
Minkowski inequality (which follows from Hölder’s inequality). We will come back to this.
Completeness follows easily from the completeness of R(C ' R2 ).
2. Consider V with norm
kXk1 := max {|xi |, i = 1, · · · , n} .
It is easily checked that (V, k · k1 ) is a complete normed vector space.
• Infinite dimensional examples:

3. Let X be a metric space (e.g. X = U or U (closure) with U ⇢ Rn open).

We consider the vector space of continuous and bounded functions in C
C(X) = {f : X ! C | f continuous and bounded} ,
with norm
kf k1 = sup |f (x)| < 1.
x2X
The norm axioms are easy to check and the completeness follows from the next.
Limits of uniformly convergent continuous functions are continuous for X = {1, 2, · · · , n}.
Then
(C(X), k · k1 ) ⇠
= (Cn , k · k1 )
f 7 ! (f (1), f (2), · · · , f (n)) 2 Cn .

6
 4. Let (X, µ) be a measure space such that X 6= ; set with distinguished “ -algebra”. Consider
A as a family of subsets of X closed under taking countable unions and under taking
complements. Take ; 2 A. We define
– µ : A ! [0, 1] (1 is included) with µ(;) = 0.
1
[ 1
X
– (Ei )i is a sequence in A, pairwise disjoint, this means that µ( Ei ) = µ(Ei ).
i=1 i=1
A main example is X = U or X = U with U ⇢ Rn open and µ is the restriction to U (or
U ) of the usual Lebesgue measure on Rn .
In this case, A = B(U ) “Borel algebra”. This is the algebra “generated” by all open
(or closed) set.

Given (X, A) as above, one has measurable functions {f : X ! C 8U 2 C open f 1 (U ) 2 A}, and
| {z }
preimage
R
integrable functions {f : X ! C measurable X |f (x)| dµ(x) < 1}.
Z
For integrable functions, we have that f (x)d µ(x) 2 C is well defined.
X

• For p 2 [1, 1), we define the Lp -space

n Z o.
Lp (X) = Lp (X, µ) := f : X ! C measurable |f (x)|p dµ(x) < 1 ⇠
X
Z
1
(f1 ⇠ f2 ! µ({x 2 X | f1 (x) + f2 (x)})) with norm kf kp := ( |f (x)|p dµ(x)) p .
X
One can show that those spaces are Banach spaces.
R
• For p = 2, it is a Hilbert space with scalar product hf, gi = X f (x)g(x) dµ(x) for f, g 2 L2 (X).

• Let us show that

⇢ Z Z
1
p
L (X) = f : X ! C measurable |f (x)|p d µ(x) < 1 , kf kp = ( |f (x)|p dµ(x)) p .
X X

is a normed vector space. To do this, we need that f, g 2 Lp =) f + g 2 Lp (X) and the

Minkowskis inequality kf + gkp  kf kp + kgkp .
For p = 1, this is trivial. Let p > 1. First we show f + g 2 Lp (X) :

|f (x) + g(x)|p  (2 · max(|f (x)|, |g(x)|))p = 2p (max(|f (x)|p , |g(x)|p ))

 2p (|f (x)|p + |g(x)|p ).

Now by taking the integrand, this implies that kf + gkpp  2p (kf (x)kpp + kg(x)kpp ) < 1 and hence
f + g 2 Lp (X).
In order to get the right inequality, we write

|f (x) + g(x)|p = |f (x) + g(x)| · |f (x) + g(x)|p 1

 |f (x)| |f (x) + g(x)|p 1

+ |g(x)| |f (x) + g(x)|p 1
.
1 1 1 1 p 1
Now we can introduce q with + = 1, i.e. = 1 = . Hence q = p p 1 .
p q q p p
This means that (|f (x) + g(x)|p 1 )q = |f (x) + g(x)|p is integrable. This implies that |f (x) +
g(x)|p 1 2 Lq (X).
Now we can apply Hölder’s inequality (see Exercise Sheet 1) for f ·h 2 L1 (X) with f 2 Lp (X), h 2
Lq (X):
kf · hk1  kf kp khkq .

7
 Taking h = |f + g|p 1, we obtain

kf + gkpp = k |f | · |f + g|p 1
k1 + k |g| · |f + g|p 1
k1
Hölder
 kf kp · k |f + g|p 1
kq + kgkp · k |f + g|p 1
kq (1.10)
R 1
p
and k |f + g|p 1 kq = ( X (|f + g|p 1 )q dµ) q = kf + gkpq = kf + gkpp 1 . If kf + gk = 0, then the
equality is trivial.
In the other case, we divide (1.10) by kf + gkpp 1 and obtain the Minkowski’s inequality

kf + gkp  kf kp + kgkp .

1.1.1 Hilbert spaces and orthonormal systems

Let (H, h·, ·i) be a space with scalar product (e.g. L2 (X)). For a subset A ⇢ H, we define

A? := {w 2 H | hw, ai = 0, 8a 2 A} .

We also set
spank A := {finite linear combination of element of A}
and also consider its closure spank (A). Then spank (A) is a linear subspace of A.
Lemma 3. (a) Let w 2 H. Then the map H 3 x 7 ! hx, wi 2 K is continuous.
?
(b) A? = spank (A)? = spank (A) .
Proof. (a) Let xn ! x 2 H, i.e. kx xn k ! 0.
This implies that |hx, wi hxn , wi| = |hx xn , wi|  kx xn k kwk ! 0.
c.s.
Hence hxn , wi ! hx, wi and our map is continuous.
?
(b) We have A ⇢ spank (A) ⇢ spank (A). This implies A? spank (A)? spank (A) .
If w 2 A? , i.e., hw, ai = 0, 8a 2 A this implies that hw, a+µbi = hw, ai+µhw, bi = 0 8a, b 2 A.
Then, hw, ci = 0 8c 2 spank (A).
Hence, A? ⇢ spank (A)? and thus A? = spank (A)? .
?
Let w 2 spank (A) , x 2 spank (A). We have found xn 2 spank (A) with xn ! x in H. This
?
means that hw, xi = lim hw, xn i = 0 and thus w 2 spank (A) .
n!1 | {z }
=0
?
This implies that spank (A) ⇢ spank (A)? , hence equality.

Remark. • A? is a closed linear subspace of H for every subset A ⇢ H. The closedness is a

consequence of Lem. 3(a).
• If V ⇢ H is a linear subspace of a normal vector space, then V ⇢ H is also a linear subspace (of
course closed).
• If A ⇢ H is a closed subspace of a Hilbert space, then there is a unique map P : H ! A which
is characterized by X PX 2 A? . P is also called the orthogonal projection to A, H = A ⌦ A? .
See exercises in second sheet.

From now on, let H be a Hilbert space.

Definition 3. Let I be an index set. A collection of elements {ei | i 2 I} ⇢ H is called an orthogonal
system (ONS) if (
1 i = j,
hei , ej i =
0 i 6= j.

An ONS is called complete, if {ei | i 2 I}? = {0} .

8
Remark. In many sources, a complete ONS is called an orthonormal basis for H. This is justified
by Thm. 1 below. However, if H is infinite dimensional, then a complete ONS is not a basis in the
algebraic sense.
Proposition 1. Let {ei | i 2 I} ⇢ H be a ONS. Then the following are equivalent
(a) {ei | i 2 I} is complete.

(b) spank {ei | i 2 I} = H.

(c) if ⇢ : H ! spank {ei | i 2 I} is the orthogonal projection, then ⇢ = id.

Proof. It is easy to show that:
? Lem. 3(b)
(a) =) (b) : spank {ei | i 2 I} = H =) {ei | i 2 I}? = spank {ei | i 2 I} = H ? = {0}.
Hence {ei | i 2 I} is complete.

(c) =) (b) : If ⇢ = id is the orthogonal projection to space V ⇢ H, in particular, id : H ! V then

V = H.
Lem. 3 ?
It remains to show (a) =) (c) : {ei | i 2 I} is complete = spank {ei | i 2 I} = {0}.
?
Hence, the orthogonal projection to the latter space satisfies X ⇢X 2 spank {ei | i 2 I} = {0}, i.e.
X ⇢X , which implies that ⇢ = id.

Example. H = L2 ([ ⇡, ⇡]). I = Z. We consider for n 2 Z

1
en (x) = p einx .
2⇡
By the exercises (Sheet 2) one can see that {en | n 2 Z} is an ONS. We will see later that it is also
complete.
Theorem 1 (Abstract Fourier series). Let I be at most countable, we may order it and assume I =
{1, · · · , N } or I = N = {1, 2, · · · }.
X
(a) For all x 2 H, the “Fourier series” hx, ei iei , converges in H to PX , where
i2I

P : H ! spank {ei | i 2 H}

is the orthogonal projection.

X
(b) {ei | i 2 I} is complete () x = hx, ei iei . (The above series converges to x for all x 2 H and
i2I
hx, ei i is the “Fourier coefficients”).
X X
(c) For x, y 2 H we have h⇢x , ⇢y i = hx, ei ihy, ei i. In particular k⇢x k2 = h|x, ei i|2 .
i2I i2I
X
(d) Let (ai )i2I be a “sequence” in K with |ai |2 < 1.
X i2I
Then, ai ei converges in H to some x with hx, ei i = ai .
i2I

Proof. (i) First we prove (d).

n
X
We may assume I = N. We consider sn = ai ei . Then, for n m + 1, we have
i=1

n
X n
X n
X 1
X
ksn sm k2 = k ai ei k2 = h·, ·i = ai aj hei , ej i = |ai |2  |ai |2 ! 0.
i=n+1 i,j=n+1 i=n+1 i=n+1

9
 1
X
Hence, |ai |2 converges when n ! 1.
i=n+1
n
X
Then, (sn ) is a Cauchy sequence and it converges, i.e. ai ei converges in H to some x.
i=1
Lem. 3(a)
We have hx, ej i = lim hsn , ej i = aj . We may assume I = {1, 2, 3, · · · } = N. For finite I,
n!1
Thm. 1 is easy to check.
1
X
(ii) Second, we show that for x 2 H, |hx, ei i|2 < 1.
i=1
For n 2 N, let ⇢n : H ! spank {e1 , e2 , · · · , en } be the orthogonal projection. Then, x ⇢n x 2
spank {e1 , e2 , · · · , en }? and thus hx ⇢n x, ei i = 0, i = 1, · · · , n, i.e. hx, ei i = h⇢n x, ei i.
X n n
X
We have that hx, ei iei = h⇢n x, ei iei = ⇢n x (holds for finite dimensional vector spaces with
i=1 i=1
scalar product).
Pn
This implies that i=1 |hx, ei i|
2 = k⇢n xk2  kxk2 because kxk2 = k⇢n xk2 + kx ⇢n xk2 .
This last inequality holds for all n. We obtain for n ! 1:
1
X
|hx, ei i|2  kxk2 < 1.
i=1
P1
Now, by (i) and (ii), we obtain that i=1 hx, ei iei converges. Thus the map x 7 ! ⇢x has the following
properties ⇢ : H ! spank {ei | i 2 I}. Moreover,
n
X n
X
Lem. 3(a)
hej , x ⇢x i = hej , xi hej , lim hx, ei iei i = hej , xi lim hej , hx, ei iei i = 0,
n!1 n!1
i=1 i=1

where (
n
X 0 if n < j,
hej , hx, ei iei i =
i=1 hx, ej i if n j.
Lem. 3(b) ?
Hence, x ⇢x 2 {ei | i 2 I}? = spank {ei | i 2 I} and thus rho is the orthogonal projection to
spank {ei | i 2 I}. This proves (a).

For (b), we have that for all x 2 H:

1
X P rop. 1
x= hx, ei iei () ⇢ = id.
i=1

By Prop. 1, this equivalent that {ei | i 2 I} is complete..

Concerning (c), we have

n
X n
X n
X
Lem. 3(a)
h⇢x , ⇢y i = lim h hx, ei iei , ⇢y i = lim hx, ei i · hei , ⇢y i = hx, ei ihy, ei i.
n!1
i=1
n!1
i=1
| {z } i=1
=hei ,yi=hy,ei i

1.1.2 Classical Fourier series

Let us recall some important tools from integration theory:

10
Lebesgue’s theorem on dominated convergence
Let (fn ) 2 L1 (X) be a series such that f (x) = lim fn (x) exists for (almost) all x 2 X. We assume in
n!1
addition that there exists g 2 L1 (X), g Z
: X ! [0, 1) such
Z that |fn (x)|  g(x) for (almost) all x 2 X.
Then f 2 L (X), fn ! f 2 L (X) and
1 1 f dµ = lim fn dµ.
X n!1
Z Z
Shortly : (lim fn )dµ = lim fn dµ.
X n!1 n!1 X

Product measures and Fubini’s Theorem

We consider two finite measure spaces (X, µ) and (Y, ⌫) (where finite and X is a countable union
of sets with finite measure). Then one can construct a distinguished measure µ ⇥ ⌫ on X ⇥ Y (the
product measure).
If X 2 Rp , with µ the p dim Lebesgue measure and if Y 2 Rq , with ⌫ the q dim Lebesgue measure,
then µ ⇥ ⌫ is the p + q dim Lebesgue measure of X ⇥ Y .

Fubini’s Theorem
(a) Let f : X ⇥ Y ! [0, 1] and µ ⇥ ⌫ the measure. Then
Z Z Z Z Z
f d(µ ⇥ ⌫) = ( f (x, y)d⌫(X)) dµ(X) = ( f (x, y)dµ(X)) d⌫(Y ). (1.11)
X⇥Y X
| Y {z } Y
| X {z }
measurable function of X measurable function of Y

(b) Let f 2 L1 (X). Then the functions

Z Z
x7 ! f (x, y)d⌫(Y ) and y 7 ! f (x, y)dµ(X)
Y X

are defined for almost all x 2 X and y 2 Y , respevtively, and they define elements of L1 (X) and
L1 (Y ) respectively. In addition, (1.11) holds.

We consider the space L1 ([ ⇡, ⇡]) L2 ([ ⇡, ⇡]).

| {z } | {z }
Banach space Hilbert space
In fact, we have for all p > 1 and (X, µ) with µ(X) < 1 that L1 (X) Lp (X).
Indeed: Let f 2 Lp (X). Then
Hölder
kf k1 = kf 1k1  kf kp kf kq  1,
1 1
where 1 2 Lq (X) is a constant function and + = 1.
p q
Let f 2 L1 ([ ⇡, ⇡]) (it can be considered as a 2⇡ periodic function on R).
We define a new function (taking “Fourier coefficients”, sometimes called “Fourier transform”).

fb : Z ! C by
Z ⇡
b 1 ik⇡
f (k) = p f (x)e dx 2 C.
2⇡ ⇡
Since |e ik⇡ | = 1, the integral exists.
If f 2 L2 ([ ⇡, ⇡]), we have
1
fb(k) = hf, ek i, where ek (x) = p eik⇡ .
2⇡

Recall that {ek | k 2 Z} is an ONS in L2 ([ ⇡, ⇡]).

11
Theorem 2. The “Fourrier transform” fb of f 2 L1 ([ ⇡, ⇡]) is injective, i.e. fb = 0 implies f = 0. In
particular, the ONS {ek | k 2 Z} ⇢ L2 ([ ⇡, ⇡]) is complete.

Proof. The second assertion follows directly from the first one.
Let f 2 {ek | k 2 Z}? ⇢ L2 ([ ⇡, ⇡]). Then fb = 0. By the first assertion, f = 0. Hence {ek | k 2 Z}? =
{0}, which means that the ONS is complete.

(i) It suffices to consider a real valued function f .

We observe that
fb(k) = fb( k).

\ f\
+f \ f\ f
Let fb = 0, then R e (f ) = = 0 and Im(f ) = = 0.
2 2
So we can consider Re (f ) and Im(f ) separately.

(ii) Let g 2 C([ ⇡, ⇡]), real-valued, with g( ⇡) = g(⇡) with gb = 0. Then g = 0.

We have already seen (Exercise 4. Sheet 3) that for a real-valued function g 2 C([ ⇡, ⇡]) with
gb = 0 we have g(0) = 0.
If g(⇡) = g( ⇡), we can extend g to a 2⇡-periodic function on R. Then, we can translate such a
function by y 2 R :
Z ⇡ Z ⇡
1 ikx 1
⌧y g(x) := g(x y)⌧cy g(k) = p g(x y)e dx = p g(z)e ik(x+y) dz = e iky gb(k).
2⇡ ⇡ 2⇡ ⇡

Hence, gb = 0 implies that ⌧c

y g = 0, thus ⌧y g(0) = g( y) = 0 for all y 2 R. Therefore, we have
that gb = 0 =) ⌧c
y g = 0 implies g = 0.

(iii) Let f 2 L1 ([ ⇡, ⇡]) with fb = 0. Then for all x 2 [ ⇡, ⇡] we have

Z ⇡
g(x) := f (t)dt = 0. (1.12)
⇡

Proof. Define g as in (1.12). Then

p g is continuous, this follows for instance from dominated
b
convergence. Moreover, g(⇡) = 2⇡ f (0) = 0, g( ⇡) = 0.
For k 6= 0, we compute
p Z ⇡ Z ⇡ Z ⇡ Z ⇡
Fubini
2⇡bg (k) = ( f (t)dt)e ikx dx = ( e ikx dx)f (t)dt
⇡ ⇡
Z ⇡⇡ ⇡
1
= (( 1)k e ikt )f (t)dt.
⇡ ik

Hence, gb(k) = 1
1)k fb(0) fb(k)) = 0.
ik ((
1
Now consider h := g p gb(0). Then, b h(k) = gb(k) = 0 for k 6= 0 and
2⇡
Z ⇡
b 1
h(0) = gb(0) p gb(0)dx = 0.
2⇡ ⇡
| {z }
gb(0)

The function h satisfies the assertion of (ii). We conclude that h = 0. Hence g is constant. But
g( ⇡) = 0 implies g = 0.

(iv) It remains to conclude from g = 0, in (1.12), that f = 0. R

Let f be as in (iii). Then for all Borel subset A ⇢ [ ⇡, ⇡] , A f (x)dx = 0.

12
(v) will imply that f = 0, more precisely, f (x) = 0 for almost all x 2 [ ⇡, ⇡].
For instance, since f is measurable, consider the Borel subset A := {x 2 [ ⇡, ⇡] | f (x) 0} .
Then, we have that
Z
f (x)dx = 0 =) f (x) = 0, for almost all x in A.
A integration theory

Consider B := Ac = {x 2 [ ⇡, ⇡] | f (x) < 0}.

Z Z
f (x)dx = 0 =) f (x)dx = 0 =) f (x) = 0, for almost x 2 B,
B B as above

but f (x) < 0, 8x 2 B, this implies that B has measure zero. Hence [ ⇡, ⇡] = A [ B implies
f (x) = 0 for almost all x 2 [ ⇡, ⇡].

Proof of (v). (a bit technical)⇢ Z

We want to show that M := A ⇢ [ ⇡, ⇡] Borel f (x)dx = 0 is equal to B, the family of
A
all Borel subset of [ ⇡, ⇡].
Note that M has the following property. Assume
1
[
A1 ⇢ A2 ⇢ · · · ⇢ Ai ⇢ Ai+1 =) Ai 2 M
i=1
x
yin M (1.13)
1
\
A1 A2 ··· Ai Ai+1 =) Ai 2 M
i=1

This follows from the dominated convergence theorem by writing

Z Z
f (x)dx = XAi (x)f (x)dx.
Ai X

A system satisfying (1.13) is called a “monotone class”.

By (iii), [ ⇡, ⇡] 2 M. The same is true for the corresponding open and hall-open in intervals.
This implies that all intervals as well as their disjoint unions (we denote this family by (1.12)
belong to M : A ⇢ M.

Claim: Let M0 be the smallest monotone class containing A. Then M0 is a algebra (closed
under complement and countable unions).

Assuming the claim: M0 is a algebra containing all intervals. On the other hand, every open
set can be written as the union of all open intervals with rational end points contained in it.
Here, M0 is a algebra containing all open sets. By definition B is the smallest algebra
containing all open sets. This implies that B ⇢ M0 ⇢ M = B, hence M = B.

Proof of the claim. Since M0 is a monotone class, it suffices to show that M0 is stable under
taking finite unions and complements. Since [ ⇡, ⇡] 2 A ⇢ M0 , it suffices to show that M0 is
closed under finite unions and differences.
For A 2 M0 , we introduce the system

N(A) = {B 2 B | A [ B, A \ B, B \ A 2 M0 } .

N(A) is again a monotone class.

If A 2 A, then A [ B, A \ B and B \ A are disjoint unions of intervals.
Hence, A 2 N(A) and thus M0 ⇢ N(A).

13
 The condition B 2 (A) () A 2 (B) is symmetric in A and B.
From the above B 2 M0 implies that B 2 N (A) and thus A 2 N(B). We have
A2A

A ⇢ N(B) =) M0 ⇢ N(B), 8B 2 M0 .

The last assertion just means that A, B 2 M0 =) A [ B, A \ B, B \ A 2 M0 .

Therefore, M0 is a algebra.

Remark. In many books, the completeness of the ONS {ek | k 2 Z} is proven in a different way. One
shows that spanC {ek | k 2 Z} = L2 ([ ⇡, ⇡]), i.e. spanC {ek | k 2 Z} 2 L2 ([ ⇡, ⇡]) is dense.
As a first step, by the approximation theorem of Stone-Weierstrass spanC {ek | k 2 Z} is dense in
Cper (R) (continuous periodic functions) with respect to the supremum norm k · k1 . As a second step,
one shows that Cper (R) ⇢ L2 ([ ⇡, ⇡]) is dense (see e.g. “Real and computer analysis”).
Definition 4. A linear map A : H1 ! H2 between Hilbert spaces is called unitary, if

hAv; AwiH2 = hv, wiH1 .

Theorem 3 (Classical Fourier series). The map L2 ([ ⇡, ⇡]) 3 f 7 ! fb 2 L2 (Z) is a unitary bijection1 .
It’s inverse is given by the Fourier series

L2 (Z) 3 g 7 ! g ⌫ 2 L2 ([ ⇡, ⇡]),
1 X P
where g ⌫ := p g(k)eikx ( g(k)eikx is convergent in L2 ([ ⇡, ⇡]).
2⇡ k2Z k2Z
X
In particular: kf k22 = |fb(k)|2 .
k2Z

Proof. Direct consequence of the completeness of {ek | k 2 Z} and Thm. 1 (abstract Fourier series).

The theorem says in particular that every f 2 L2 ([ ⇡, ⇡]) has a unique representation

1
X Z⇡
ikx fb(k) 1 ikx
f= ek e , where ck = p = f (x)e dx,
2⇡ 2⇡
k=1 ⇡

where the convergences is w.r.t the L2 norm.

In general, one cannot expect pointwise convergence of the Fourier series, even not for general f 2
Cper (R). However, if f 2 Cper
1 (R) we even have uniform convergence (see exercises Sheet 5).

Higher dimensional Fourier series

We consider [ ⇡, ⇡]n ⇢ Rn and L1 ([ ⇡, ⇡]n ) L2 ([ ⇡, ⇡]n ). For f 2 L1 ([ ⇡, ⇡]n ), we define fb :
Zn ! C by
Z
b 1 ihk,xi if f 2 L2 ([ ⇡, ⇡]n )
f (k) := n f (x)e dx = hf, ek i,
(2⇡) 2 n
[ ⇡,⇡]

where ek (X) = 1
n eihk,ki , and hk, xi = k1 x1 + · · · + kn xn , k = (k1 , · · · , kn ).
(2⇡) 2

Theorem 4. (a) f 2 L1 ([ ⇡, ⇡]n ), fb = 0 =) f = 0.

(b) ek | k 2 Zn is a complete ONS.

Proof. This was derived from Thm. 2 in Exercises 3 and 4, Series 4 (using Fubini).
1
Z is a counting measure.

14
Combining Thm. 4 with Thm. 1, we obtain Thm. 5.

Theorem 5 (Higher dimensional Fourier series). The map

L2 ([ ⇡, ⇡]n ) 3 f ! fb 2 L2 (Zn ).

is a unitary bijection with inverse

1 X
L2 (Zn ) 3 g 7 ! gb := n g(k)eihk,xi
(2⇡) k2Zn
2

In particular, X
kf k22 = |fb(k)|2 .
k2Zn

Among other things, Fourier series can be used to study constant coefficient PDE on periodic functions
(2⇡ periodic in every variable).

Let X
D= c↵ D ↵ (1.14)
|↵|k

@ |↵|
where D↵ = , ↵ = (↵1 , · · · , ↵n ), ↵i 2 N0 , |↵| = ↵1 + · · · + ↵n . Given D of order k,
@x↵11
· · · @x↵nn
g 2 Cper (Rn ) (2⇡ periodic in every variable). Find f 2 Cper k (Rn ) such that

Df = g.

c
@f
1
Proposition 2. Let f 2 Cper (Rn ). Then (k) = i · ke fb(k), k = (k1 , · · · , kl ).
@xe
Proof. Partial integration as in the one-dimensional case: (compare Exercise 1 Series 4).
(↵)
Corollary 1. (a) If f 2 Cper (Rn ) then

d
D ↵ f (k) = (ik)↵ fb(k),

where for z 2 Cn , we set z ↵ := z ↵1 z ↵2 · · · z ↵ .

(b) Let D be a differential operator of order k as in (1.14). We define the polynomial

X
PD : Rn ! C by PD (y) = i|↵| c↵ y ↵ .
|↵|k

c (k) = PD (k) · fb(k).

Then Df

Proof. (a) Is just iteration of Prop. 2.

(b) Take linear combinations of (a).

This means that Df = g is just equivalent to PD (k)· fb(k) = g(k), k 2 Zn (polynomial equation on Zn ).

15
1.2 Definition of the Fourier transform and Schwartz functions
The Fourier coefficient of 2⇡ periodic functions (in several variables) were just given by integration of
f against exponentials eihk,xi , k 2 Zn , which are exactly the 2⇡ periodic eigenfunction for the partial
@ @
derivation ,··· , .
@x1 @xn
@ @
Dropping the periodicity condition, all eigenfunctions of ,··· , are given by eih⇠,xi , ⇠ 2 Cn . If
@x1 @xn
we require in addition boundedness, then ⇠ 2 Rn .

Definition 5. For f 2 L1 (Rn ) and ⇠ 2 Rn we define

Z
1
fb(⇠) = n f (x)e ih⇠,xi
dx
(2⇡) 2 Rn

The function fb : Rn ! C is called the Fourier transform of f .

Lemma 4. If f 2 L1 (Rn ) then fb is a well-defined, bounded, continuous function on Rn (fb 2 Cb (Rn ))

satisfying
1
kf k1  n kf k1 .
(2⇡) 2

Proof. For ⇠ 2 Rn we have |e ih⇠,xi | = 1 and here the integrand in Def. 5 belongs to L1 (R). Hence the
integral is defined. We estimate
Z
b 1 1
|f (⇠)|  n |f (x)e ih⇠,xi | dx = n kf k1 .
(2⇡) 2 Rn | {z } (2⇡) 2
=|f (x)|

1
) kf k1  n kf k1 , in part f is bounded.
(2⇡) 2
For continuity we employ the dominated convergence theorem :
Let ⇠n ! ⇠ in Rn . We consider f (x) := f (x)e ih⇠n ,xi . fn converges pointwise to f (x)e ih⇠,xi . We
have |fn | = |f |, that implies
Z Z
1 1
lim n f (x)dx = n f (x)e ih⇠,xi dx.
n!1 (2⇡) 2 Rn (2⇡) 2 R
| {z } | {z }
=fb(⇠n ) =fb(⇠)

So fb(⇠n ) ! fb(⇠) for all ⇠n ! ⇠. Thus f is continuous.

Problem. L2 (Rn ) * L1 (Rn ), but we want to define and study the Fourier transform on L2 (Rn ).
Approach : Study the Fourier transform on nice large subspaces on L1 (Rn ) \ L2 (Rn ) and try to extend
by continuity to L2 (Rn ).
One could consider L1 (Rn ) \ L2 (Rn ) for this space, but one the “Schwartz space”

S(Rn ) ⇢ L1 (Rn ) \ L2 (Rn )

has better properties.

In the following, we will write |x| instead of kxk (Euclidean norm), for x 2 Rn .

Definition 6. A function f : Rn ! C is called a Schwartz function if

(a) f 2 C 1 (Rn ).

(b) f and all its derivatives are “rapidly decreasing” (for |x| ! 1), i.e. for all N 2 N0 , and all
multi-indices ↵, there exists a constant CN,↵ such that

(1 + |x|2 )N |D↵ f (x)|  CN,↵ . (1.15)

16
By S(Rn ), we denote the vector space of all Schwartz functions on Rn .
Remark. (a) The condition (1.15) can be expressed in a more symmetric way : for all multi-indices
↵, there is
C↵, = |X ↵ D f (x)|  C↵, . (1.16)

Indeed (1.15) implies that a similar estimate holds for all polynomials. (1+|x|2 )N is a polynomial.
Hence (1.15) =) (1.16).
Vice versa, one uses that |x↵ |  C(1 + |x|2 )N , whenever 2N |↵|. Hence (?) =) (??).
(b) There are not so many “classical functions” that belong to S(Rn ). For instance,
• e x / S(R) since e
2 x ! 0 for x ! 1.
• e |x| 2
/ S(Rn ), since it is not differentiable at x = 0.
|x2 |
• the “Gaussian” g(x) := e 2 belongs to S(Rn ).
1
• 2 S(Rn ).
cosh(x)
A lot of Schwartz functions are obtained by taking smooth “compactly supported” functions, i.e. smooth
functions on Rn vanishing outside a sufficiently large ball. We will see that there are really many of
them.
Recall (Notion of support). The support of a continuous function f : X ! C is defined by

supp f = {x 2 X | f (x) 6= 0}.

Alternative expression for suppf useful later

[
suppf = ( U )C ,
{U ✓X open | f |U =0}

where C is the complement on X. We set Cc (X) := {f : X ! C continuous | supp f compact}.

For X = U ⇢ Rn , we set Here, Cc1 (U ) = Cc (U ) \ C 1 (U ).
Cc (U ) is the vector space of completely supported smooth functions on U . In particular, we have
Cc1 (Rn ).
not⇢
Note: Even if U ⇢ Rn is bounded, then Cc1 (U ) 6= C 1 (U ). The support of f 2 Cc1 (U ) looks like in
Fig. 1.

Figure 1: Support of f 2 Cc1 (U ).

Lemma 5. (a) Cc1 (Rn ) ⇢ S(Rn ).

(b) S(Rn ) ✓ Lp (Rn ) for all p 1.
Proof. (a) We can take as CN,↵ in (?) just the following

sup (1 + |x|2 )N |D↵ f (x)| < 1,

x2suppf| {z }
continuous

where suppf is compact.

17
 (b) For (b), we observe using polar coordinates that the function hN (x) := 1
(1+|x|2 )N
is integrable on
Rn . If N > n2 , then hN 2 Lp (Rn ). On the other hand by (1.16)

1
|f (x)|  C0,N = C0,N hN (x), for all N ,
(1 + |x|2 )N

where f 2 Lp (Rn ) and hN (x) 2 Lp (Rn ) for N sufficiently large.

In particular, S(Rn ) ⇢ L1 (Rn ). So for f 2 S(Rn ) the Fourier transform fb is defined.

We first want to study the Fourier transform on S(Rn ).
The defining integral for fb is a “parameter dependent” integral with parameter ⇠ 2 Rn . We need
differentiation of parameter dependent integrands.

Differentiation of parameter dependent integrands

Let (X, µ) be a measure space, U ⇢ Rn an open (parameter) subspace, and k 2 N0 the order of
differentiation.
Let f : U ⇥ X ! C be a measurable function such that for all multi-indices ↵ with |↵|  k the
function
y 7 ! Dy↵ f (y, x)
exists and is continuous for almost all x 2 X. Moreover, for all x0 2 U there is a neighborhood U0 and
a integrable function
g↵ : X ! [0; 1[
with |Dy↵ f (y, x)|  g(x) for all y 2 U0 , and almost all x 2 X. Then, the function
Z
U 3y7 ! f (y, x)dµ(x)
X

belongs to C k (U ) and for each ↵ such that |↵|  k, we have

Z Z
Dy↵ f (y, x)dµ(x) = Dy↵ f (y, x)dµ.
X X

This is a consequence of the dominated convergence.

|x2 |
Lemma 6 (Fourier transform of the Gaussian). Let g(x) := e 2 on R. Then gb = g.

Proof. We first do it for n = 1 Z(crucial step).

+1
1 x2
We first compute gb(0) = p e 2 dx.
2⇡ 1
The usual way to evaluate is by using Fubini’s theorem as follows:
Z +1 Z Z 2⇡ Z 1
x2 Fubini x2 polar coordinates r2
2
( e 2 dx) = e 2 dx = e 2 r dr = 2⇡,
1 R2 0 0

r2 r2
where e 2 = (e 2 )0 . Hence, gb(0) = 1 = g(0).
There are at least 2 variants to proceed with the proof.

1 variant: (Differentiation of parameter dependant integrals)

For this we observe that g satisfies the first order ODE, g 0 = xg, which becomes unique by requiring

18
g(0) = 1. We compute
Z +1 Z +1
0 1 x2 d i⇠x 1 x2
i⇠x
g (⇠) = p e 2 e dx = p ixe 2 e dx
2⇡ 1 d⇠ 2⇡ 1
Z +1
1 d x2
i⇠x
= p (e 2 )e dx
2⇡ 1 dx
Z +1
IPP i x2 d i⇠
= p e 2 e dx
2⇡ 1 dx
Z +1
i x2
i⇠
= p e 2 e dx
2⇡ 1

= ⇠b
g (⇠).

Hence, gb0 = g , this is the ODE above. We conclude that g = gb.

⇠b

2 variant: (use of complex) Let us write the exponetial in the following way
x2 2 2 2 ⇠2 z2
i⇠x ( x2 +i⇠x) ( z2 + ⇠2 )
e 2 e =e =e =e 2 e 2 ,

where z := x + i⇠ and z 2 = x2 + 2i⇠ ⇠ 2 . Then,

Z +1 Z i⇠+1
1 x2
i⇠x z=x+i⇠ ⇠2 1 x2
gb(⇠) = p e 2 e dx = e 2 p e 2 dz.
2⇡ 1 2⇡ i⇠ 1

x2
By Cauchy’s theorem, the integral over the closed path indicated of e 2 is 0 (see Fig. 2), hence

x2
Figure 2: Visualisation of Cauchy’s theorem with the integral of e 2 along a closed path.

Z i⇠+r Z r Z Z
··· = ··· + +
i⇠ r r ↵2 (r) ↵3 (r)

Z i⇠+1 Z +1 p p
r!1 x2 x2
=) e 2 dz = e 2 dz = 2⇡b
g (0) = 2⇡,
i⇠ 1 1
Z Z
⇠2
where both integrals and converge to 0 when n ! 1. This implies that gb(⇠) = e 2 =
↵2 (r) ↵3 (r)
g(⇠).
|x|2 Qn x2
j
For n 1, we use Fubini multiplicativity of exponentials e 2 e ih⇠,xi = j=1 e
2 e i⇠j xj and we

19
obtain
Z Z Z Z Y
n x2
1 |x|2
ih⇠,xi Fubini 1 j
i⇠j xj
gb(⇠) = n e 2 e dx = n ··· (e 2 e )dx1 dx2 · · · dxn
(2⇡) 2 Rn (2⇡) 2 R R R j=1
| {z }
n
n
Y Z x2
1 j
i⇠j xj
= n e 2 e dxj
j=1
(2⇡) 2 R
n
Y ⇠j2
(⇤)
= e 2

j=1
= g(⇠),

where in (⇤) we applied the result for n = 1. This finishes the proof of the second variant.

Proposition 3. Let f 2 S(Rn ), then

(i) fb 2 S(Rn ).
d
(ii) D ↵b
x f (⇠) = (i⇠) f (⇠) for every multi-index ↵ = (↵1 , · · · , ↵n ).
↵

(iii) D⇠↵ fb(⇠) = ( \

ix)↵ f (⇠) for every multi-index ↵ = (↵1 , · · · , ↵n ).

Proof. We first show (ii). As in the case of Fourier series, we do partial integration. In fact, it suffices
@
to show (ii) for |↵| = 1, say D↵ = . We write x = (x0 , xn ), x0 2 Rn 1 . Then,
@xn
c Z Z
@f 1 @f
(⇠) = n ( (x)e ih⇠,xi dxn )dx0
@xn (2⇡) 2 Rn 1 R @xn
Z
@f
= (x)e ih⇠,xi dxn
R @xn
Z r
@f
= lim (x)e ih⇠,xi dxn
n!1 r @x n
h ir Z r
IPP 0 ih⇠,(x0 ,xn )i
= f (x0 , xn )e ih⇠,(x ,xn )i f (x0 , xn )(i⇠)e dxn
r r
Z
r!1 h⇠,xi
! i⇠ f (x)e dxn
R

c
Inserting in the above, we obtain @f
@xn (⇠) = i⇠n fb(⇠).

Now let us show that fb 2 C 1 (Rn ) and that (iii) holds.

We want to apply differentiation in parameter dependent integrals. First of all, D⇠↵ (f (x)e h⇠,xi ) exists
and is continuous for all ↵
h⇠,xi hx,⇠i
|D⇠↵ f (x)e | = |( ix)↵ f (x)e | = |x↵ f (x)|.

Note that x 7 ! x↵ f (x) is again in S(Rn ), where x↵ f (x) := g(x) 2 S(Rn ) ⇢ L1 (Rn ) independent of ⇠.
By differentiation of parameter independent integrals we have fb 2 C k (Rn ) for all k, i.e. 2 C 1 (Rn ) and
Z Z
1 1
↵b
D⇠ f (⇠) = n
↵
D⇠ (f (x)e hi⇠,xi
)dx = n ( ix)↵ f (x)e ih⇠,xi dx = ( \
ix)↵ f (⇠).
(2⇡) 2 Rn (2⇡) 2 Rn

20
Finally, we show (i). According to the remark after Def. 6 (1.16) it suffices to show that for all
multi-indices ↵, the function g↵, = ⇠ ↵ D⇠ fb(⇠) is bounded. In fact, we have

g↵, = ⇠ ↵ ( \
ix) f (⇠) = ( i)|↵| (i⇠)↵ ( \
ix) f (⇠) = ( i)|↵| D⇠↵\
(ii)
( ix) f (⇠).

Hence, g↵, is the Fourier transform of a Schwartz function, which is an element of L1 (Rn ) and thus
by Lem. 1 g↵, is bounded.

Remark. • The some proof applies to any function which is of sufficient decay to produce a certain
degree of differentiability of fb. The differentiability of f is not importantrfor this point. E.g.,
1 2 |⇠|
the function 2
has a Fourier transform which is even not C 1 , namely e (Exercise).
1+x ⇡
• Let D be a constant coefficient differential operation on Rn . Let PD : Rn ! C be the associated
polynomial as in the end of Subsect. 1.1. Then, Prop. 3 (ii) implies as in case of Fourier series
and we get
c (⇠) = PD (⇠)fb(⇠).
Df
Theorem 6 (Fourier inversion on S(Rn )). The Fourier restricted to S(Rn ) gives a bijection

^ : S(Rn ) ! S(Rn ).

Moreover, for f, g 2 S(Rn ), we have

Z
1
(a) f (x) = n fb(⇠)eih⇠,xi dx (inversion formula).
(2⇡) 2 Rn

(b) hf, giL2 (Rn ) = hfb, gbiL2 (Rn ) .

Remark. • Observe the analog of (a) (and also (b)) to the theory of Fourier series. Every Schwartz
function can be represented as the integral (“continuous sum”) of exponentials with the “Fourier
coefficients” fb(⇠) as coefficients.

• (a) holds whenever it makes sense, in particular for f 2 L1 (R) \ C(R) (critical condition).
Proof. Idea: The most important thing is to establish (a). For this we have to evaluate the double
integral Z Z Z
1 b ih⇠,xi 1
n f (⇠)e d⇠ = n f (y)e ih⇠,yi dy eih⇠,xi d⇠
(2⇡) 2 Rn (2⇡) 2 Rn Rn
Z Z
1
= n f (y)e ih⇠,x yi dy d⇠.
(2⇡) 2 R n R n

We would like to interchange the order of integration. This is impossible since the function

(y, ⇠) 7 ! f (y)eih⇠,x yi

is not integrable in Rn ⇥ Rn . We will introduce an additional factor '(⇠), ' 2 S(Rn ) making the
integrand integrable and then study what happens if ' ! 1 (constant function).
Let ' 2 S(Rn ) be arbitrary. We consider
Z Z
1
'(⇠)fb(⇠)e ih⇠,xi d⇠ = n ( f (y)eih⇠,x yi dy)d⇠
R n (2⇡) 2 R n
Z Z
Fubini 1
= n '(⇠)e ih⇠,x yi d⇠ f (y)dy
(2⇡) 2 Rn Rn
Z
= b
'(y x)f (y)dy
Z R n

z=y x
= b
'(z)f (z + x)dz.
Rn

21
Hence,
Z Z
'(⇠)fb(⇠)e ih⇠,xi
d⇠ = b
'(z)f (z + x)dz (1.17)
Rn Rn

Consider another g 2 S(Rn ) and introduce for " > 0, '" (⇠) := g("⇠). Then, '" 2 S(Rn ) and we have
pointwise convergence '" ! g(0) (constant function). We want to put '" into (1.17) and let then
"!0
" ! 0 so that we can use dominated convergence.
We start with the left hand side
|'" (⇠)|  sup |g(x)|  C.
x2Rn

This implies that |'" (⇠)fb(⇠)e ih⇠,xi |  C|fb(⇠)|, where lim '" (⇠)fb(⇠)e ih⇠,xi = g(0)fb(⇠)e ih⇠,xi . Hence,
"!0
fb is Schwartz, and thus in L1 . So C|fb(⇠)| 2 L1 (R) (independent of ✏).
By the dominated convergence
Z Z Z
lim b
'" (⇠)f (⇠)e ih⇠,xi
dx = lim '" (⇠)fb(⇠)eih⇠,xi dx = g(0) fb(⇠)eih⇠,xi dx. (1.18)
"!0 Rn Rn "!0 Rn

For the right hand side, we have

" =h z ,yi
z }| {
Z Z y
1 1 ihz, i 1 z
b" (z) =
' n g("x)e ihz,xi dx = n g(y)e " dy = gb( ).
(2⇡) 2 Rn y="x (2⇡) 2 Rn "n "

Hence
Z Z
1 z
b" (z)f (z + x)dz
' = gb( )f (z + x)dz
Rn "n Rn "
Z
y= z"
= gb(y)f (x + "y) dy
Rn | {z }
| |  C 2 |b
g (y)| L1 and independent of "
Z Z
"!0
! gb(y)f (x)dy = gb(y)dy f (x)
dominated convergence Rn Rn
R R
Now by (1.17) and (1.18), we obtain ( Rn gb(y)dy)f (x) = g(0) Rn fb(⇠)e ih⇠,xi dx for all f, g 2 S(R)n .
|x|
Next, we evaluate the constants be taking g(x) := e 2 . Then g(0) = 1 and, by Lem. 6,
Z
n n n
g(x)dy = (2⇡) 2 = g(0) = (2⇡) 2 g(0) = (2⇡) 2 .
Rn
R
We conclude that f (x) = 1
n
Rn f (⇠)ehi⇠,xi dx, which is (a).
(2⇡) 2

We now show that ^ : S(R) ! S(R) is bijective.

We define the “inverse Fourier transform” _ : S(R) ! S(R) by
_
Z
1
f (x) := n f (⇠)eih⇠,xi dx.
(2⇡) 2 Rn
_
Then, (a) says (fb)_ = f . We show that also (f )n = f . Then _ is the inverse of ^ and therefore ^ is
injective. We observe that fe(x) = fe( x). We define f 1 (x) := f ( x). We obtain (as above for '" )
d1 = (fb) 1 (change of variables).
'
_ above _
(f )^ (x) = ((fb) 1)
^
(x) = (fb)_ ( x) = (fb)_ (x) = f (x), i.e.(f )^ = f.

22
This implies bijectivity.

As for (b): hf, giL2 (Rn ) = hfb, gbiL2 (Rn ) . We return to (1.17) and set x = 0. We then obtain that
R R
b
Rn f (z)'(z)dz = Rn fb(⇠)'(⇠)d⇠. By taking '(⇠) := gb(⇠), we have that ' 2 S(Rn ) and the right hand
side of both equations coincide. By computing
Z Z
1 ihz,⇠i 1
b
'(z) = n gb(⇠)e d⇠ = n gb(⇠)e ihz,⇠i d⇠ = g(z).
(2⇡) 2 Rn (2⇡) 2 Rn (a)

we have that the left hand side of the two equations also coincides. Hence, the scalar product formula
follows.

1.3 Convolution, density theorems and Fourier transform on L2 (Rn )

def
Recall. A subset A ⇢ X of a metric space is called dense ! A = X.
Equivalently: 8x 2 A, 9(an ) ⇢ A : an ! x or 8x 2 X, 8" > 0, 9a 2 A : d(x, a) < ".
n!1
E.g. Q ⇢ R is dense.
We are interested here in X = Lp (Rn ), in particular p = 2.

Goal among other things:

• Show that Cc1 (Rn ) ⇢ Lp (Rn ) is dense. In particular S(Rn ) ⇢ Lp (Rn ) is dense.
• Extend the Fourier transform from the dense space S(Rn ) ⇢ L2 (Rn ) to L2 (Rn ) by continuity.

First we want to “produce” many elements in Cc1 (Rn ). We start by stating that there is at least one
such non-zero function.
Lemma 7. There exists 2 Cc1 (Rn ) with
(a) supp 2 B(0, 1) ⇢ Rn .
(b) (x) 0, 8x 2 Rn .
Z
(c) (x)dx = 1.
Rn

Proof. We consider the following function f : R ! [0, 1] ⇢ R given by

( 1
e t if t > 0,
f (t) =
0 if t  0,

see Fig. 3.

Figure 3: Function f (t).

Then limf (k) (t) = 0. This implies f 2 C 1 (R) (see basic analysis courses or books). We define
t!0
Z [0, 1] by '(x) := f (1 |x| ). Then f 2 Cc (R) and satisfies condition (a) and (b).
' : Rn ! 2 1
1
Let ↵ := '(x)dx > 0. Then set (x) := '(x), we have that satisfies all requirements.
Rn ↵

23
We want to produce a lot more of compact smooth functions by using “convolution integrals”. First
we need some definitions.
Definition 7. Let U ⇢ Rn be open (e.g. U = Rn ). We define the space L1loc (U ) of “locally integrable
functions” by
⇢ Z
1 ,
Lloc (U ) = f : X ! C measurable 8K ⇢ U compact : |f (x)|dx < 1
K
⇠
where ⇠ is the usual equivalence relation.
Remark. We have
L1 (U ) ⇢ L1loc (U ), Lp (U ) ⇢ L1loc (U )
since Lp (K) ⇢ L1 (K) for K compact (this implies that K has finite measure) and that C(U ) ⇢ L1loc (U )
(continuous functions are bounded on each compact K).
Definition 8 (support of a measurable function). Let U ⇢ Rn be open, f : U ! C be measurable.
We define [
supp f = ( V )c ⇢ U.
V ⇢U
f |V =0 a.e.

Remark. • If f1 ⇠ f2 , then supp f1 = supp f2 . So supp f is well defined for f 2 L1loc (U ), Lp (U ).

• supp f ⇢ U is closed.
• If f 2 C(U ) (continuous), then this definition gives the same supp f as the original one.
• We define L1c (U ) = f 2 L1loc (U ) | supp f is compact . Then L1c (U ) ⇢ L1 (U ).
Definition 9 (Convolution of two functions). Let ' 2 Cc (Rn ), f 2 L1loc (Rn ). Then we define a new
function Z
' ⇤ f : Rn ! C by ' ⇤ f (x) := '(y)f (x y)dy (1.19)
Rn
Remark. Under
Z the assumption that the integrand is real integrable we have by a charge of variables
' ⇤ f (x) = '(x y)f (y)dy = “f ⇤ '00 .
Rn
Lemma 8. (a) ' ⇤ f is well-defined and thus also (1.19) holds.
(b) ' ⇤ f is continuous. Moreover, if ' 2 Cck (Rn ), then also ' ⇤ f 2 C k (Rn ) (here k = 0, k = 1 is
allowed) and D↵ (' ⇤ f ) = (D↵ ') ⇤ f .
Remark. This is a way to produce a lot of smooth functions starting from one (say ').

In the next lemma we will see that this construction gives a lot of elements of Cc1 (Rn ).

Proof of Lem. 8. (a) We have to check that for all x 2 Rn , the function
Rn 3 y 7 ! '(y)f (x y)
is integrable. In fact, it is measurable and
Z Z Z
|'(y)f (x y)|dy = |'(y)||f (x y)|dy  C |f (x y)|dy
Rn supp' supp'
Z
= C |f (z)|dz
{x} supp'
< 1,
since f 2 L1loc (Rn ). Here C := supy2supp' |'(y)| < 1 and {x} supp' =: {x y | y 2 supp'} is
compact. This proves (a).

24
 (b) We use the formula (1.19) and differentiation of parameter dependent integrals. We have to
consider Dx↵ '(x y). For x varying in a bounded set B, we have |Dx↵ '(x y)|  C↵ for all
y 2 Rn (because of the compact support of ').
Thus, for some compact B 0 , |Dx↵ '(x y)f (y)|  C↵ XB 0 (y)|f (y)| independently for x 2 B, where
|f (y)| 2 L1 (Rn ), since f 2 L1loc(Rn ) . Hence, ' ⇤ f 2 C k (Rn ) and
Z
k
D (' ⇤ f )(x) = Dxk '(x y)f (y)dy = (Dk ') ⇤ f (x).
Rn

Notation. Let A, B ⇢ Rn be subsets. We define A + B := {x + y | x 2 A, y 2 B} .

Properties (Exercises). .

• A closed, B compact =) A + B closed.

• A, B compact =) A + B is compact.

Warning: A, B closed not implies that A + B closed. For instance, B(0, R) + B(0, ") = B(0, R + ").

Figure 4: Nonclosure under addition of closed sets.

Lemma 9. Let ' 2 Cc (Rn ), f 2 L1loc (Rn ) then supp ' ⇤ f ⇢ supp ' + supp f .
In particular, if f 2 L1c (Rn ) (compact support) then ' ⇤ f 2 Cc (Rn ).
If in addition ' 2 Cc1 (Rn ), then ' ⇤ f 2 Cc1 (Rn ).
Z
Proof. Let x 2 R be such that ' ⇤ f (x) =
n '(x)f (x y)dy 6= 0.
supp'
Then, µ {y 2 supp' | f (x y) 6= 0} > 0, hence 9y ⇢ supp', x y 2 suppf. This means

x = y + (x y) 2 supp' + suppf,

which is closed since supp' is compact and suppf is closed. Taking the closure of the set of all such
x, we obtain supp(' ⇤ f ) ⇢ supp' + suppf .
R
Lemma 10. Let ' 2 Cc1 (Rn ) be as in Lem. 7, ' 0, supp' = B(0, 1), R '(x)dx = 1.
1 ⇣x⌘
For " 0, we define '" (x) := n · ' .2 Let f 2 Cc (Rn ).
" "
Then

• supp(' ⇤ f ) ⇢ supp(f ) + B(0, "),

• '" ⇤ f 2 Cc1 .
"!0 "!0
• k'" ⇤ f f k1 ! 0, i.e. '" ⇤ f ! f uniformly.
2
R
supp'" = B(0, "), Rn
'" (x)dx = 1.

25
Proof. The first assertions follow immediately from Lem. 8 and 9. We prove the convergence.
First we observe Z
'" ⇤ f (x) f (x) = f" (y)(f (x y) f (x))dx
B(0,")
Since f is compactly supported and continuous, it is uniformly continuous. This means that 8 >
0, 9"0 > 0 : 8y 2 B(0, "0 ) : |f (x y) f (x)|  . Hence,
Z Z
'" 0 if "+"0
|'" ⇤ f (x) f (x)|  '" (y)|f (x y) f (x)|dx  '" (y)dy ·
B(0,") B(0,")
| {z }
=1
= .
Taking supremum over x 2 Rn the assertion follows.
Theorem 7. Let U ⇢ Rn be open. Then Cc1 (U ) ⇢ Lp (U ), p 1, is dense.
In particular, Cc1 (Rn ) ⇢ Lp (Rn ), S(Rn ) ⇢ Lp (Rn ) are dense!

For the proof of Thm. 7, we need the following proposition.

Proposition 4. Let U ⇢ Rn be open. Then Cc (U ) ⇢ Lp (U ) is dense.
Proof. For the proof see e.g. Rudin “Real and complex analysis”.
Proof of Thm. 7. By Prop. 4 we have Cc (U ) = Lp (X).
If we could show that Cc (U ) ⇢ Cc1 (U ), then would have Lp (X) = Cc (U ) ⇢ Cc1 (U ) ⇢ Lp (X), which
would imply that Cc1 (U ) = Lp (X).
Thus it suffices to show: 8g 2 Cc (U ), 9 sequence (hn ) in Cc1 (U ) such that khn gkp ! 0.
n!1
We will instead produce h" for " ! 0. We identify
Cc (U ) ⇠
= {g 2 Cc (Rn ) | supp g ⇢ U }
Cc1 (U ) ⇠
= {f 2 Cc1 (Rn ) | supp f ⇢ U } (extension by zero).

Figure 5: Visualising support sets with Cc (U ) and smooth functions in Cc1 (U ).

Let g 2 Cc (U ), then there is "0 > 0 such that supp g + B(0, "0 ) ⇢ U (since g is compact). Let "  "0 .
We set h" := '" ⇤ g, '" as in Lem. 10. Then h" 2 Cc1 (Rn ), supp h" ⇢ suppg + B(0, ") ⇢ U , this implies
that h" 2 Cc1 (U ). Hence
Z
p
kh" gkp = |'" ⇤ g g|p dx
supp h" [ supp g

 k'" ⇤ g gkp1 · vol(supp h" [ supp g) ! 0

| {z } "!0
=:C<1

This proves convergence in Lp (Rn ).

26
1.4 Continuous Linear maps and continuous extensions
Let (V1 , k · k), (V2 , k · k) be normed vector spaces and A : V1 ! V2 be linear map.

A is continuous () A is continuous at 0 2 V1 () 9C : kAvk2  Ckvk1 , 8v 2 V1 . (1.20)

Then smallest C in (1.20) is called the operator norm kAk of A. We have

kAk = sup kAvk2 .

v2V1 , kvk1 =1

Then, kAvk2  kAk kvk1 .

Proposition 5. Let V1 , V2 be as above. In addition we assume that V2 is a Banach space. Consider the
normed vector space (V0 , k · k), where V0 ⇢ V1 is a dense subspace. Let A : V0 ! V1 be a continuous
linear map.
Then there is a unique continuous linear map A e : V1 ! V2 with A|e V = A (continuous extension).
0
e = kAk.
We have kAk

Proof. Uniqueness: Holds for general continuous maps f1 , f2 : X1 ! X2 between metric spaces that
coincide on a dense subset X0 ⇢ X1 . Indeed, the set A := {x 2 X1 | f1 (x) = f2 (x)} is closed.
On the other hand X0 = X1 ⇢ A. But X 0 ⇢ A ⇢ X1 , so A = X1 , i.e. f1 = f2 .

The existence uses much more specific properties:

Let v 2 V1 . We choose a sequence (vn ) ⇢ V0 such that vn ! v in V1 .
n!1
bv := lim Avn (if it exists). Let vn be a Cauchy sequence. Then
We define A
n!1

kAvn AVm k2 = kA(vn vm )k2  kAk kvn v m k2

b is
It follows that (Avn ) is a Cauchy sequence in V2 . Since V2 is complete (Banach), lim Avn exists. Av
n!1
independent of the choice of the sequence. Let vbn ! v be a sequence of this kind. Then (e
vn vn ) ! 0
in V0 .
It follows by continuity of A: A(evn vn ) ! 0. In other words: lim Avn = lim Ae vn .
n!1 n!1

e
Now we can check basic properties of A:
e V = A (take a constant sequence).
• A| 0

e is linear.
• A
e is continuous with kAk
• A e = kAk, hence

e 2 = k lim Avn k2 = lim kAvn k2  lim kAk kvn k1 = kAk lim kvn k1 = kAk kvk1 .
kAvk
n!1 n!1 n!1 n!1

e  kAk. On the other hand, every extension satisfies kAk

A is continuous with kAk e kAk. Hence
e = kAk.
kAk

Theorem 8 (Fourier transform on L2 (R)). The Fourier transform ^ : S(Rn ) ! S(Rn ) extends
uniquely to a continuous linear map F : L2 (Rn ) ! L2 (Rn ).
F is unitary and bijective.

Proof. We apply Prop. 5 to V1 = V2 = L2 (Rn ), which is Hilbert-, in particular a Banach space.

We have V0 = S(Rn ) which is dense in V1 = L2 (Rn ) by Thm. 7. We consider A : S(Rn ) ! S(Rn )
given by A(f ) := fb 2 S(Rn ) ⇢ L2 (Rn ). By Thm. 6 (b) we have

hA(f ), A(g)i = hfb, gbi = hf, gi.

27
In particular, kA(f )k = kf k, and therefore is continuous.
e of A to L2 (Rn ),
By Prop. 5, there exists a unique continuous extension F := A

F : L2 (Rn ) ! L2 (Rn ), F(f ) = fe,

if f 2 S(Rn ).

F is bijective: We have the inverse Fourier transform B : S(Rn ) ! L2 (Rn ), B(f ) := f _ (see the proof
of Thm. 6). Applying the same procedure to B, we obtain an extension G := B e : L2 (Rn ) ! L2 (Rn ).
By (the proof of) Thm. 6 we have:

A B = B A = i : S(Rn ) ! L2 (Rn )

where i is the natural injection. F G is a continuous extension of A B = i. But the unique extension
of i is idL2 (Rn ) . We obtain that F G = id. Similarly, G F is the continuous extension of B A = i.
Hence G F = id implies G = F 1 , and therefore F is bijective.

Unitary: We use formula (a) above (from Thm. 6). Let f, g 2 L2 (Rn ). We choose sequences (fn ), (gn ) in
S(Rn ) such that fn ! f , gn ! g in L2 (Rn ). Then, by construction F(f ) = lim fbn , F(g) = lim gbn .
n!1 n!1
Hence
(a)
hF(f ), F(g)i = lim hfbn , F(g)i = lim lim hfbn , gbn i = lim lim hfn , gn i = hf, gi.
n!1 n!1n!1 n!1n!1

We now have two versions of the Fourier transform:

^ : L1 (Rn ) ! C0 (Rn )

and
F : L2 (Rn ) ! L2 (Rn )
given by F(f ) = lim fbn , where fn ! f , fn 2 S(Rn ) (approximating sequence for f ) is arbitrary.
n!1

Problems:

• We know that ^ |S(Rn ) = F|S(Rn ) .

But we would like to have ^ |L1 (Rn ) \ L2 (Rn ) ! S(Rn ) = F|L1 (Rn ) \ L2 (Rn ) ! S(Rn ) .

• For f 2 L2 (Rn ) we would like to have formula for F(f ) without using the approximation of f
by Schwartz functions.

These problems are solved in the next proposition.

Proposition 6. (a) Let f 2 L1 (Rn ) \ L2 (Rn ). Then F(f ) = fb 2 L2 (Rn ) (\C0 (Rn )).

(b) Let f 2 L2 (Rn ). Then

Z
L2 1 ih⇠,xi
F(f ) = lim n f (x)e dx
R!1 (2⇡) 2
B(0,R)

(a kind of improper integral).

Proof. (a) Claim 1: Let f 2 L1 (Rn ) \ L2 (Rn ). Then there exists a sequence (fn ) in Cc1 (Rn ) such
that kfn f k1 ! 0 and kfn f k2 ! 0 (The point is that we can use the same sequence for
n!1 n!1
k · k1 and k · k2 ).

28
Proof of the Claim 1. We first observe that

XB(0,n) f 2 L1 (Rn ) \ L2 (Rn ) and kXB(0,n) f f k1 ! 0, kXB(0,n) f f k2 ! 0 . (1.21)

n!1 n!1

We apply Thm. 7 to p = 2 and U = B(0, n). We find fn 2 Cc1 (B(0, n)) ⇢ Cc1 (Rn ) such that
1
kfn XB(0,n) f k2  p . Now we estimate
n vol(B(0, n))
Z
kfn XB(0,n) f k1 = |fn (x) f (x)|dx = h |fn XB(0,n) f |, XB(0,n) iL2 (Rn )
B(0,n)
Cauchy-Schwartz
 kfn XB(0,n) f k2 · kXB(0,n) k2
above 1
 ,
n
p
where kXB(0,n) k2 = vol(B(0, n)). We obtain kfn XB(0,n) f k1 ! 0, kfn XB(0,n) f k2 ! 0.
n!1 n!1
Combined with (1.21) (and triangle inequality) we obtain the claim.

We equip the vector space L1 (Rn ) \ L2 (Rn ) with the norm kf k := k f k1 + kf k2 . Claim 1 tells
us that Cc1 (Rn ) is dense in L1 (Rn ) \ L2 (Rn ) with respect to k · k. For fixed R > 0 we define 2
linear maps:
^R : L1 (Rn ) \ L2 (Rn ) ! L2 (B(0, R))
FR : L1 (Rn ) \ L2 (Rn ) ! L2 (B(0, R))
by
fbR := fb|B(0,R) , FR (f ) := F(f )|B(0,R) ,

where fb|B(0,R) is bounded continuous on B(0, R) if and only if is in L2 (B(0, R)). Both maps are
well-defined, linear.

Claim 2: The two above maps are continuous with respect to k · k.

Proof. We have
p p
kfbR k2  kfbR k1 · vol(B(0, R))  kfbk1 · vol(B(0, R))
p
Lem. 4 vol(B(0, R))
 kf k1 · n
(2⇡) 2
p
vol(B(0, R))
 kf k · n
(2⇡) 2
| {z }
=:C

By unity of F (Thm. 8), we have kFR f k2  kFf k2 = kf k2  kf k.

Conclusion: Now we have two continuous linear maps on (L1 (Rn ) \ L2 (Rn ), k · k) to L2 (B(0, R))
that coincide on S(Rn ) in part on Cc1 (Rn ).
By Claim 2, we have that Cc1 (Rn ) is dense in (L1 (Rn ) \ L2 (Rn ), k · k). By the easy uniqueness
part of Prop. 5 the two maps coincide on the whole space. This means: for almost all ⇠ 2 B(0, R)
[1
we have fb(⇠) = Ff (⇠). Now Rn = B(0, N ) countable union. We conclude that fb(⇠) = Ff (⇠)
N =1
for almost all ⇠ 2 Rn , i.e., F(f ) = fb as elements of L2 (Rn ).

29
 (b) Let f 2 L2 (Rn ). Then kXB(0,R) f f k2 ! 0
R!1
Note that XB(0,R) f 2 L1 (Rn ). The continuity of F now implies
Z
L2 (a) L2 \ f = lim L2 1 ih⇠,xi
F(f ) = lim F(XB(0,R)f ) = lim XB(0,R) n f (x)e dx.
R!1 R!1 R!1 (2⇡) 2
B(0,R)

Remark (Concerning the evaluation of limL2 ). Using the fact (F) below, the limL2 can be simply
R!1 R!1
computed as follows. Consider a sequence Rn ! 1 (e.g. Rn = n). Then, possibly after going to a
n!1
subsequence, the integrals Z
ih⇠,xi
f (x)e dx
B(0,Rn )

converge for almost all ⇠ 2 Rn . The resulting class of measurable functions then defines F(f ) 2 L2 (Rn ).

(F) Let (fn ) be a sequence in L2 (X) converging to f 2 L2 (X). Then there is a subsequence fnk such
that fnk converges to f (for k ! 1) pointwise almost everywhere.

1.5 An application to constant coefficient differential operators

X
Let D = c↵ D↵ , c↵ 2 C be a constant coefficient differential operator of order k. As in Sect. 1.1 we
|↵|k
P
consider the corresponding polynomial PD (⇠) := i|↵| c↵ ⇠ ↵ . By Prop. 3, we obtain for f 2 S(Rn )
|↵|k
that Df 2 S(Rn ) and
c (⇠) = PD (⇠) · fb(⇠) .
Df (1.22)

Proposition 7. Let D =
6 0 be as above. The linear map

D : S(Rn ) ! S(Rn )

is injective. This means in terms of PDE, for any given PDE

Df = g (1.23)

has at most one solution f belonging to S(Rn ) (it has much more solutions if we chop the condition
f 2 S(Rn )).

Proof. We use the fact that if a polynomial on Rn vanishes on a non-empty open set, then it is zero.
If D =
6 0, then PD 6= 0, and hence

A := {⇠ 2 Rn | PD (⇠) 6= 0} ⇢ Rn

is dense. Now assume Df = 0. Then by (F) 0 = Df c (⇠) = PD (⇠) · fb(⇠) and hence fb(⇠) = 0 for all
b b
⇠ 2 A. Since f constant and A dense, we get f = 0, by Thm. 6 this implies that f = 0.

6 0 as above. Then D : Cc1 (Rn ) ! Cc1 (Rn ) is injective.

Corollary 2. D =

For g 2 Cc1 (Rn ) the equation Df = g has a most one solution f in Cc1 (Rn ).
There is a more quantitative version of Cor. 2 saying that the support of f cannot be much bigger
than the support of Df = g (if f 2 Cc1 (Rn )). More precisely, if A ⇢ Rn we consider its convex hull
conv(A) = smallest convex subset of Rn containing A.

30
 Figure 6: Convex hull of A ⇢ Rn .

Note that for general f 2 C 1 (Rn ), supp(Df ) ⇢ supp(f ). For f 2 Cc1 (Rn ), we get not the inclusion in
the other direction, but we get it for the convex hulls.

Proposition 8. Let D =
6 0 be a constant coefficient and consider the same differential operator as
above. Then,
conv(supp(Df )) = conv(supp(f )).

Proof. Omitted but not difficult if one uses the Fourier transform.

Remark. Prop. 8 implies Cor. 2.

Indeed, we have that f 2 Cc1 (Rn ) with Df = 0 implies conv(supp(Df )) = 0 and by Prop. 8, we get
conv(supp(f )) = ;. Hence, supp(f ) = ; implies f = 0.

d
Sample example. Consider D = on R1 , and f 2 Cc1 (Rn ) such that
dx
f |[ 1,1] = 1, supp(f ) = [ 2, 2] .

Figure 7: Smooth function f with compact support, f |[ 1,1] = 1 and supp(f ) = [ 2, 2].

Then supp(Df ) ⇢ [ 2, 2] [ [1, 2] implies conv(supp(Df )) ⇢ [ 2, 2]. More precisely, we must have
equality.

Remark. Prop. 8 plays a key role for the proof of the following existence theorem.

Theorem 9. Let D 6= 0 be a coefficient differential operator, U ⇢ Rn be open and convex, and

g 2 C 1 (U ). Then there exists f 2 C 1 (U ) such that Df = g.

31
2 The classical differential equations
2.1 The Laplace/Poisson equations
Definition 10. Let U ⇢ Rn be open, f : U ! C. Then f is called harmonic, if f 2 C 2 (U ) and

Xn
@2f
f (= ) = 0 on U.
i=1
@x2i

So harmonic functions are thus just solutions of the homogeneous Laplace (or Poisson) equation

f = 0,

where is the “Laplace operator”. The corresponding inhomogeneous equation is f = g, g 2 C(U )

given.

Examples of harmonic functions:

Pn
• Consider polynomials of (at most) first order: f (x) = i=1 ai x
i +b. For n = 1 these are harmonic
functions.
For n 2 there are much more harmonic functions.

• Let n = 2, then is the product of two first order differential operator. This gives a relation
between harmonic and holomorphic functions on U ⇢ R2 ⇠ = C, (x, y) 7 ! z = x + iy.

• Claim 1: Let U ⇢ C ⇠
= R2 be open, and let f : U ! C be holomorphic. Then,

f, Re (f ), Im (f ) : U ! R/C

are harmonic.

Proof of the Claim 1. If f is harmonic, then automatically Re (f ) and Im (f ) are harmonic, since

Re ( f ) = (Re (f )), Im ( f ) = (Im (f )) ( has real coefficients).

We want to show that f is harmonic if f is holomorphic. For this we introduce the complex
partial derivatives:
@ 1 @ @
:= ( i )
@z 2 @x @y
@ 1 @ @
:= ( +i )
@z 2 @z @y

By the Cauchy-Riemann equations in complex analysis, we have f is holomorphic implies that

@f
f is differentiable as a function of two real variables and = 0. Moreover, every holomorphic
@z
function belongs to C 1 (U ), in particular C 2 (U ), since it is locally given as a converging power
series. We compute

=0 on C 2 (U )
z }| {
@ @ 1 @ @ @ @ 1 @2 @2 @2 @2 1
= ( i )( + i ) = ( 2 + 2 + i( )) = on C 2 (U ).
@z @z 4 @x @y @x @y 4 @x @y @x@y @y@x 4
@f
Now let f : U ! C be holomorphic. Then f 2 C 2 (U ) and = 0.
@z
1
f
z 4}| {
@ @f
Hence, f 2 C 2 (U ) and ( ) and thus f is harmonic.
@z @z

32
Example. U = C \ ( 1, 0], f : U ! C principal branch of the logarithm, i.e., the unique extension
of the usual logarithm log : (0, +1) ! R to holomorphic function on C \ ] 1, 0]. In polar
coordinates we have f (rei' ) = log(r) + i', ' 2 ] ⇡, ⇡[. Then Re (f (rei' )) = log(r). In other words
Re (f (z)) = log |r|. This is a harmonic function on U by (a)). In fact log |z|. This is a harmonic
function on U by the Claim 1. In fact, log |z| is harmonic on C \ {0} ⇠
= R2 \ {0}.

The Claim 1 above has a certain reverse.

Let U ⇢ C be open and simply connected (e.g. U = C, B(z0 , R) but not C \ {0}).
Let f : U ! C be harmonic. Then there exist holomorphic functions f1 , f2 : U ! C such
that f = f1 + f 2 (f 2 is “anti holomorphic”).
If f : U ! R is harmonic, then there exists g : U ! C holomorphic such that f = Re (g).

Proof. Concerning the last assertion: f real valued implies that f = Re (f ). By the Claim above, we
=:g
z }| {
have that Re (f1 + f 2 ) = Re (f1 + f2 ).

Concerning the first assertion: Let f : U ! C be harmonic. Using the computation in the Claim
@ @f @f
above, we find ( ) = 0, i.e. is holomorphic by the Cauchy-Riemann equations. By complex
@z @z @z
analysis holomorphic functions on simply connected domains are holomorphic primitives, i.e. there
@f
exists a holomorphic function f1 : U ! C such that f10 (z) = (z).
@z
@f1 @
Moreover, as for every holomorphic function we have f10 = . This implies (f f1 ) = 0.
@z @z
@ @ @
We define f2 := f f1 (=) f = f1 + f 2 ). Then f2 = (f f1 ) = (f f1 ) = 0, i.e. f2 is also
@z @z @z
holomorphic.

Remark. • The reversed version of the Claim above shows that every harmonic function is at
least locally expressible as f1 + f 2 , f1 , f2 holomorphic.

• This implies that for n = 2 every holomorphic function is of class C 1 (later we will see that this
holds for arbitrary n).

• Now let n 2 be arbitrary. We extend the scalar product h·, ·i from Rn to Cn , C bilinearly:
n
X
hw, zi = wi zi . Then there are a lot of isotropic c vectors 0 6= ⇠ 2 Cn :
i=1

n
X
h⇠, ⇠i = ⇠i2 = 0.
i=1

For instance, in C2 the isotropic vectors are given by ⇠ = (z, ±iz), z 2 C.

We consider the following linear function

g⇠ : Rn ! C, g⇠ (x) := h⇠, xi

Claim 2: Let ⇠ 2 Cn be isotropic and let h : R2 ⇠

= C ! C be harmonic (e.g h holomorphic),
then
f := h g⇠ : Rn ! C
is harmonic. For instance take (for example).

• h = z k =) x 7 ! h⇠, xik is harmonic (harmonic polynomials).

• h = ez =) x 7 ! eh⇠, xi is harmonic.

33
Proof. Consider h : C ⇠
= R2 ! C differentiable as a function on R2 and v ⇢ C ⇠
= R2 . We have that
@h @h
dhz (v) = (z)Re (v) + (z)Im (v),
@x @x
v+v v v @h @h
with Re (v) = , Im (v) = = (z)v + (z)v. (Check this!).
z zi @z @z
Using this, we compute
@g
= @x⇠ (x)
i
@f @h z}|{ @h
(x) = (g⇠ (x)) ⇠i + (g⇠ (x))⇠ i
@xi chain rule @z @z

@2f @2h @2h @2h

and we obtain @x2i
(x) = (g (x))⇠i2
@z 2 ⇠
+ (g (x))⇠ i
@z 2 ⇠
+2 |⇠i |2 .
@z@z
| {z }
1
4
h=0
@2h @2h
By summing everything up, we get f= g⇠ h⇠, ⇠i + 2 g⇠ h⇠, ⇠i = 0 (⇠ isotropic).
@z 2 | {z } @z | {z }
=0 =0

Remark. • The last example implies that for all n 2 and all non-empty open U ⇢ Rn

dim {f : U ! C | f harmonic} = 1.

• f harmonic implies that D↵ f is harmonic (at least if f 2 C |↵|+2 (U )) ( (D↵ f ) = D↵ ( f ) = 0)

The some is true for Df , D const. coeff. D.O. on Rn (if f 2 C ord(D)+2 (U )).

• If U = Rn and f : Rn ! C harmonic, ' 2 Cc (Rn ) then ' ⇤ f : Rn ! C is harmonic

( (' ? f ) = ' ⇤ f = 0).
|{z}
=0

2.2 Typical problems connected with the Laplacian

(a) The inhomogeneous Poisson equation
Let U ⇢ Rn be open, g 2 C(U ) given. Find f 2 C 2 (U ) such that

f = g. (2.1)

Remark. If f solves (2.1), and f0 is harmonic on U then f + f0 is also a solution of (??).

Since there are a lot of harmonic functions solutions of (2.1), they are for from being unique.
In order to get something unique, one usually considers additional “boundary conditions” for f .

(b) Boundary value problems

We consider the following type of problems:
U ⇢ Rn open, bounded, connected with smooth boundary @U = U \ U , i.e. @U ⇢ Rn is a smooth
submanifold of dimension n 1 of Rn .

Figure 8: Domain U and smooth boundary @U .

34
 This means: @U can (locally) be parametrised by a C 1 -map
open
: Rn 1
V ! @U ⇢ Rn

such that d (g) has full rank n 1.

We have even more:
8x 2 @U 9 open neighborhood W ⇢ Rn of x and a diffeomorphism of class C 1
e : Rn Ve !W

such that n o
e( y 2 Ve | yn > 0 ) = W \ U
n o
e( y 2 Ve | yn = 0 ) = W \ @U.

Figure 9: Diffeomorphism e: Smooth mapping from Rn to open neighborhoods around @U points.

Figure 10: Examples of smooth maps: allowed and not allowed.

We can speak of functions f 2 C k (U ) by requiring that f e 2 C k ({y 2 V | yn 0}), where e

@
as above (all partial derivatives up to order k exist and are continuous, is only taken from
@yn
the above).

(b.1) Dirichlet problem

Given ' 2 C(@U ). Look for f 2 C(U ) such that f |U is harmonic and f |@U = '.
Shortly: Look for f 2 C 2 (U ) \ C(U ) with

(i) f = 0.

(ii) f |@U = '.

A physical interpretation of the Dirichlet problem

' can be interpreted as the distribution of temperature on @U (if n = 3: on the surface of the
solid body U ) which is kept constant in time. On expects that after a while the temperature f in
the interior will also be constant in time. Thus process is usually described by the heat equation
@f
= xf =) Xf = 0,
@t

35
 @f
where = 0 for t 0. By continuity of f on U we should have f |@U = '. We will see in the
@t
lecture that solutions of ?? are unique and will give an explicit solution for U = B(0, R) ball of
radius R.

(b.2) Poisson equation with boundary conditions

Given g 2 C(U ) (or g 2 C 1 (U )). Find f 2 C 2 (U ) \ C(U ) such that

(i) f = 0 on U .

(ii) f |@U = 0.

Remark. (b.1) and (b.2) are just special cases of the more general problem, say (b.0).

(b.0) More general problem

Given ' 2 C(@U ), g 2 C(U ) (or C 1 (U )). Find f 2 C 2 (U ) \ C(U ) such that

(i) f = g on U .

(ii) f |@U = '.

Assume (b.1) and (b.2) have (unique) solution. Then so has (b.0).
In fact, let f1 be a solution of (b.1) (corresponding to ') and f2 a solution of (b.2) (corresponding
to g). Then f := f1 + f2 solves (b.0):

f= f + f = g,
|{z}1 |{z}1
=0 =g

f |@U = f1 |@U + f2 |@U = '.

| {z } | {z }
=' =0

Uniqueness: Let f3 , f4 be two solutions of (b.0). Then F = f3 f4 is a solution of (b.1) (with

' = 0). By uniqueness of solution of (b.1), we would get F = 0, i.e. f3 = f4 .
Conclusion: It suffices to consider (b.1) and (b.2) in order to solve (b.0). Of different nature.

(b.3) Dirichlet eigenvalues (and eigenfunctions)

We are looking for number 2 C and functions 0 6= f 2 C 2 (U ) \ C(U ) such that

(i) f + f = 0 on U .
(2.2)
(ii) f |@U = 0,

i.e. is an eigenvalue of , f corresponding eigenfunction vanishing on @U .

Remarks. • (b.3) is usually treated as a problem of special theory of a certain self-adjoint

unbounded operator acting on the Hilbert space L2 (U ) (see e.g. Chapters 6 8 in the book
of Reed-Simon3 ). We refer some results we get for problem (b.3).
• If (2.2) has a non-zero solution, then 2 (0, 1).
• The space of solutions of (2.2) is a vector space over C (easy) which is finite-dimensional
(for fixed , not so easy). Multiplicity of := dimension of that space.(One expects that
this dimension is  1 if U has no symmetries).
3
Methods of modern mathematical physics, Volume 1: Functional analysis, http://www.astrosen.unam.mx/~aceves/
Metodos/ebooks/reed_simon1.pdf

36
 • 6= µ =) hf , fµ iL2 (U ) = 0.
• One can order the appearing eigenvalues by size, (there are only countably many). We get
a sequence 1  2  3  · · ·  n ! 1 (no finite accumulation point).
Here we repeat eigenvalues according to their “multiplicity”. We choose corresponding eigen-
functions fn := f n with kfn k2 = 1 and in case of higher multiplicity fn , fn+1 , · · · , fn+k is
an orthonormal basis of the solution space of (2.2) with = n , if
n 1 < n = ··· = n+k < n+k+1 .

Then: {fn | n 2 N} is a complete ONS of L2 (U ).

• For n = 2: It can be imagine as a drum. The sequence of eigenvalues describes the “sound”
of the drum. More precisely, 1 is base tone, 2  3  · · · are the overtones.

Figure 11: Domain representation for n = 2 with drum boundary.

Remark. Knowing (theoretically) solutions of (b.3) one can find solutions of (b.2):
{fn } complete ONS in L2 (U )
1
X
From Sect. 1, g 2 C(U ) ⇢ L2 (U ) such that g = hg, fn ifn . Then at least formally the solution
n=1
f of (b.2) is given by
1
X 1
f= hg, fn ifn .
n=1 n

(b.4) Neumann problem

is of order 2. In analog to ODE one expects that there are interesting boundary conditions
involving also first derivatives of the function at the boundary. For that we introduce the outer
unit normal vector field N along @U ,
N : @U 3 x 7 ! N (x) 2 Rn
It is characterised by the following conditions:

(i) N (x) ? all tangent vectors to @U at X. In terms of parametrization : Rn 1 V ! @U

this means ⌧
@
N ( (y)), (y) =0
@yk Rn
for k = 1, · · · , n 1 and x = (y).
(ii) N (x) points inside U .
(iii) |N (x)| = 1

Let f 2 C 1 (U ). We define the outer normal derivative N f 2 C(@U ) by

n
X @f
N f (x) := dfx (N (x)) = hgrad fx , N (x)i = Ni (x) (x), x 2 @U.
@xi
i=1

Neumann problem:
Given ' 2 C(@U ). Find f 2 C 1 (U ) \ C 2 (U ) such that

37
 (i) f = 0 on U .

(ii) N f = ' on @U.

Remark. Uniqueness of solutions can only be up to additions of constant.

2.3 Gradient, divergence, volume forms and integral formulas

Let U ⇢ Rn be open, and let X : U ! Rn (or Cn ) be a vector field on U of class C 1 .

Figure 12: Vector field on U of class C 1 .

Then we define continuous function, the divergence of X, by

div X : U ! R/C
n
X @Xi
div X(x) = (x),
@xi
i=1

where X(x) = (X1 (x), · · · , Xn (x)).

If f 2 C 2 (U ) then grad f : U ! Cn is a vector field on U of class C 1 . Then
div (grad f ) = f.
@f @ grad(f )i @2f
Indeed, (grad f )i = , hence = . The Euclidean geometry in Rn induces a “volume
@xi @xi @x2i
form” on @U (a measure on @U ) as follows.
Let : Rn 1 V ! @U ⇢ Rn be a local parametrisation. Then we set
Z Z s ⌧
@ (y) @ (y)
f d(@U ) = f ( (y)) det( , )i,j=1,··· ,n ,
@xi @xj
(V )⇢@U V | {z }
>0

where f : @U ! C is continuous function, and dy is a (n 1) dimensional Lebesgue measure.

S
r
Integral over @U is obtained as follows: @U = k (vk ), here r is the parametrisation of the pieces.
k=1
We decompose f = f1 + f2 + · · · + fr , with suppfk ⇢ k (vk ). We set
Z r
X Z
f d(@U ) = fk d(@U ).
@U k=1
k (vk )

This definition is independent of the choice of the parametrisations and the decomposition of f . (This
follows from the transformation formula for integrals in Rn 1 ).
Example. Let n = 2. We can parameterise @U a closed curve : [a, b] ! R2 , (a) = (b). Then
Z Z b
f d(@U ) = f ( (t))| (t)| dt,
U a

where (t) is the length of the tangent vector.

38
 R R
In the following, we will simply write @U f (y)dy for @U f d(@U ). We now formulate the Classical
Gauß(-Ostrogradski) formula (a kind of higher dimensional partial integration)
Z Z
div X(x)dx = hX(y), N (y)i dy. (2.3)
U @U

Here, U ⇢ Rn open, bounded, connected, @U smooth, X 2 C 1 (U , Rn ) vector field and N the outer
unit normal as in (b.4).
Proof. Omitted.
Now let f : U ! C be an additional function of class C 1 . Then f · X : U ! Cn is a new vector field
of class C 1 . We compute
Leibniz rule
div(f · X) = hgradf, XiRn + f · divX.
Then by (2.3)
Z Z
(f · div(X) + hgradf, Xi)dx = f hX, N idy. (2.4)
U @U

From this we obtain important partial integration formulas for .

Proposition 9 (Green’s formula). Let f, g 2 C 2 (U ). Then we have
R R R
(G1) U f · g dx = U hgrad f, grad gidx + @U f · N g dy.
R R
(G2) U (f · g f · g)dx = @U (f · N g N f · g)dy.
Proof. We insert X =: grad g into (2.4) and use hgrad g, N i = N g, we obtain (G1).
(G2) arises by taking the difference between (G1) and (G2) with the roles of f and g interchanged
(hgrad f, grad gi is symmetric and cancels).

2.4 Volume of spheres:

We set Z
n 1
!n := voln 1 (S (0, 1)) = 1 d(S n 1
(0, 1)),
Sn 1 (0,1)

where S n 1 is a sphere of radius 1 in Rn . It can be computed explicitly (Exercise):

n
(2⇡) 2
!n = .
( n2 )
Z 1
Here (S) := e t ts 1
dt is the Gamma function, s > 04 . In particular,
0
2⇡
!2 = = 2⇡ (length of unit circle).
(1)
3 3
(2⇡) 2 (2⇡) 2
!3 = = = 4⇡ (area of surface of the 2-sphere).
3
(2) 1 1
2 ( )
| {z2}
p
= ⇡

The radial harmonic functions on Rn \ {0} are given (up to addition of constants and multiples) by
f (x) = log |x| if n=2 and
1
f (x) = if n 3.
|x|n 2

We have observed that f 2 L1loc (Rn ).

4 p
( (1) = 1, s (s) = (s + 1) =) (n) = (n 1)! (n 2 N), ( 12 ) = ⇡).

39
Definition 11 (Fundamental solutions of ). Let n 2. We define n : Rn \ {0} ! R by
1 1
• 2 (x) := log |x| (= log |x|).
!2 2⇡
1 1
• n (x) := · , n 3.
(n 2)!n xn 2

Proposition 10. Let U ⇢ Rn be open, connected, bounded with smooth boundary. Let f 2 C 2 (U ),
x 2 U . Then
Z Z
f (x) = n (x y) f (y) dy + n (x y)N f (y) f (y)Ny (x y) dy. (?) (2.5)
U @U

Remarks. • Since n 2 L1loc (U ) there is no problem with the convergence of the first integral.

• On @U , (x .) is smooth.

• The first integral is a kind of convolution (difference : we integrate only over U instead of Rn ).

Proof. Fix x 2 U . Let " > 0 such that B(x, ") ⇢ U and consider the new region U" := U \ B(x, ").
Then @U" = @U [ S n 1 (x, ").

Figure 13: New region U" with boundary components @U and S n 1

(x, ").

We apply (G2) to f and n (x .). On U" we have y n (x .) = 0 ( n harmonic on Rn \ {0}). Hence,

by (G2), we obtain
Z Z
n (x y) f (y) dy = (f (y)Ny n (x y) n (x y)N f (y)) dy
U @U
Z Z
(G2)
=) n (x y) f (y) dy = + (f (y)Ny n (x y) n (x y)N f (y)) dy
U R Sn 1 (x,")

and the right hand side of (2.5) is equal to lim S n 1 (x,") (f (y)Ny n (x y) n (x y)N f (y)) dy.
"!0

It remains to show that this limit is equal to f (x). We have that n (x y) is constant for y 2 S n 1 (x, ")
(namely n (")). Since B(x, "0 ) is compact and f 2 C 2 ⇢ C 1 , this implies that |N f (y)|  C on B(x, "0 )
"n 1· !
R z }| n {
n 1
and thus Sn 1 (x,") n (x y)N f (y) dy  | n (")| · C · vol (S (x, ")).
Consider 8
< 1 n 3
0
| n (")| = Cn · "n 2
:| log "| n = 2

40
 R
then Sn 1 (x,") ··· ! 0. For the first summand we observe that in polar coordinates centered at x
"!0
we have for any
@h
h : Nh = .
@0r
Now take h = n (x .). We do it for n 3 (n = 2 is similar):
|x y| = r 1 1 @ 2 n
n (x y) = r2 n =) Ny n (x y) = r
(n 2)!n (n 2)!n @r r="
1
= ( (2 n))r1 "
(n 2)!n "
1
=
! n "n 1
1
= n 1 (x, ")
R R vol S
Hence, S n 1 (x,y) f (y)Ny n (x y) dy = vol S n 1 (x,✏) S n 1 (x,") f (y) dy =: If (").
1

Since f continuous, we have that for all > 0, 9"0 > 0 such that |f (y) f (x)|  wherever y 2 B(x, ").
By integrating over the sphere for " < "0 , we get |If (") f (x)|  ↵, and thus lim If (") = f (x). This
"!0
implies the assertion.

Corollary 3. Let f 2 Cc2 (Rn ). Then f = n ⇤ f.

In particular, f can be recovered from f and the map

: Cc2 (Rn ) ! Cc0 (Rn )

is injective.

Proof. Exercise.

Proposition 11 (Special solution of the Poisson equations f = g, when g has compact support).
Let g 2 Cc2 (Rn ). Then f := n ⇤ g 2 C (R ) with
2 n f = g.

Proof. Exercise.

Remarks. • The solution f is not compactly supported in general.

• The proposition also holds for g 2 Cc1 (Rn ) (more complicated proof) but not in general for
g 2 Cc (Rn ).

• If we consider additional boundary equations, e.g. Problem (b.2), one often finds similar integra-
tion formulas. But then n (x y) has to be replaced by another function of 2 variables Gn (x, y),
the Green’s function for U .

We now exploit Prop. 10 in order to derive interesting properties of harmonic functions.

Proposition 12 (Mean value property of harmonic functions). Let U ⇢ Rn be open, f : U ! C be

harmonic. Let x 2 U , R > 0 such that B(x, R) ⇢ U . Then
Z
1
f (x) = f (y)dy,
!n R n 1
Sn 1 (x,R)

where dy is the volume form.

Proof. We Rapply Prop. 10 to B(x, R) instead of U . We have f (x) = 0 8x 2 B(x, R). Hence,
f (x) = ( n (x y)N f (y) f (y)Ny n (x y))dy.
Sn 1 (x,R)

41
We have that y 7 ! n (x y) is constant on S n 1 (x, R), namely equal to CR := (R), hence
Z Z
CR N f dy = (CR N f f · N CR )dy
| {z }
Sn 1 Sn 1 =0
Z
= (CR f f · N CR )dy = 0.
(G2) |{z} | {z }
B(x,R) =0 =0

We have seen in the proof of Prop. 10 that Ny n (x y) = @r @

n (r) r=R , now N y = @r pointing outside
@
R
the ball, hence this is equal to !n ·Rn 1 . We obtain that f (x) = !n ·Rn 1 S n 1 (x,") f (y) dy.
1 1

Remark. One can show that, vice versa, every C 2 -function having the mean value property is harmonic.

Proposition 13 (Maximum principle for harmonic functions). Let U ⇢ Rn be open, bounded and
connected5 . Let f 2 C(U ) \ C 2 (U ) be harmonic on U .
If f is real valued and assumes its maximum or minimum in U , then f is constant.
In general, if |f | takes its maximum in U , then f is constant (not true if |f | takes its minimum in U .)

Shortly: Non-constant harmonic function take its maximum only on the boundary @U .

Proof. We first assume f real valued and assume that f takes a maximum at x0 2 U (for a minimum
consider just the harmonic function f ).
Let M = supf (x) = maxf (x) = N < 1 (U is compact). We consider A := {x 2 U | f (x) = M }. By
x2U x2U
assumption: A 6= ;. By continuity of f : A ⇢ U is closed, therefore Ac ⇢ U is open. We want to show
that A is also open.
Now let x 2 A. Choose r > 0 with B(x, r) ⇢ U . Then by the mean property of f (Prop. 12) we have
for all R < r : Z
1
f (x) = f (y) dy.
|{z} !n · Rn 1 S n 1 (x,R)
=M | {z }
M
R
Hence, 1
!n ·Rn 1 Sn 1 (x,R) (f (y) M ) dy = 0, and thus f (y) M = 0 8y 2 S n 1 (x, R). This means
| {z }
0, continuous
that S n 1 (x, R) ⇢ A 8R < r and hence B(x, r) ⇢ A. This holds for all x 2 A =) A is open.
We have U = A [ Ac . Since U is connected, one of the two sets should be empty. Since A 6= ;, we
open open
conclude that A = U . Then f (x) = U for all x 2 U , i.e. f is constant.
Now let f be complex valued. We consider the standard scalar product on R2 ⇠
= C:

hz, wi = Re (z) · Re (w) + Im (z) · Im (w),

where z, w 2 C. Then for any z 2 C

x 7 ! hz, f (x)i = Re (z) · Re (f (x)) + Im (z) · Im (f (x))

| {z } | {z }
harmonic harmonic

is harmonic and real valued.

Now, let x0 2 U such that |f | take its maximum at x0 . We consider the real valued harmonic function

fe(x) := hf (x0 ), f (x)i.

Then, by Cauchy-Schwartz we obtain

|fe(x)|  |f (x0 )| · |f (x)|  M · M = M 2

5
No condition on @U .

42
and |fe(x0 )| = M
f =) fe assume its maximum or minimum at x0 . By the above, we get fe is constant
and hence f (x) = f (x0 ) + g(x) for some g with hf (x0 ), g(x)i = 0. By Pythagoras, we obtain

|f (x)|2 = |f (x0 )|2 + |g(x)|2 .

| {z } | {z } | {z }
M 2 =M 2 0

This implies that g(x) = 0 8x 2 U and thus f (x) = f (x0 ) 8x 2 U, i.e. f is constant.

2.5 The solution of the Dirichlet problem for the ball U = B(0, R) ⇢ Rn
⇣ |y| R ⌘
For R > 0, consider R (x, y) := n x y , n 2 and the “Green’s function”
R |y|
GR (x, y) := n (x, y) R (x, y)

for B(0, R). We will consider GR mainly as function for x 2 B(0, R), y 2 @B(0, R) = S n 1 (0, R).

Lemma 11. For y 2 @B(0, R) = S n 1 (0, R) we have for normal derivative with respect to y

|x|2 R2 1
Ny GR (x, y) = · < 0, for x 2 B(0, R).
R · !n |x y|n
Proof. We do it for n 3 (n = 2 runs similarly).
In Ex. 1 Sheet 11, we have seen that R (x, y) = R (y, x). Hence, we also have GR (x, y) = GR (y, x):

1 |x| R 2 n
GR (x, y) = GR (y, x) = (|y x|2 n
y x )
n 3 (n 2) · !n R |x|
1 |x| R2
= (|y x|2 n
( )2 n (y x))2 n
.
(n 2) · !n R |x|2

Figure 14: Representation of N (y).

y 1 Pn @f
We have N (y) = = (y1 , · · · , yn ). Then, N f (y) = 1
R i=1 yi @yi (y).By computation,
R R
@ @ n n n
|y z|2 n
= hy z, y zi1 2 = (1 )hy z, y zi 2 2(yi zi )
@yi @yi 2
= (2 n)|y z| n (yi zi )
R2
!
@ 1 yi xi ⇣ |x| ⌘2 n yi x
|x|2 i
and thus GR (x, y) = R2
.
@yi !n |y x|n R |y |x| 2 x|
n

43
|x| R2 ? |y| R2 |y|=R
y 2
x = y y = |x y|,
R |x| R |y|2
?: Symmetric in x, y proved as the symmetry of R (Ex. 1 Sheet 11).
Therefore, we obtain
!
@ 1 1 ⇣ |x| ⌘2 ⇣ R2 ⌘
GR (x, y) = yi xi yi
@yi !n |y x|n R |x|2
!
1 1 |x|2
= 1 yi .
!n |y x|n R2
P
Now since 1
R yi @y@ i , we get

=|y|2 =R2
! z n}| {
1 1 |x|2 X 2 1 R2 |x2 |
Ny GR (x, y) = 1 yi = .
R!n |y x|n R2 R!n |x y|n
i=1

Theorem 10. Let ' 2 C(S n We define for x 2 B(0, R)

1 (0, R)).

Z
R2 |x|2 '(y)
f (x) := dy.
R !n S n 1 (0,R) |x y|n

Then

(a) f is harmonic on B(0, R).

(b) f extends continuously to B(0, R) with lim f (x) = '(y0 ).

x!y0
|x|<R

Shortly: (The continuous extension of ). f solves the Dirichlet problem (b.1) for U = B(0, R).

Proof. By Lem. 11, we have

Z
f (x) = Ny GR (x, y)'(y)dy.
Sn 1 (0,R)

Hence, x Ny GR (x, y) = Ny x GR (x, y) and x GR (x, y) = x n (x y) x R (x y) = 0.

| {z } | {z }
=0 =0 (by Ex. 1 Sheet 11)

differentiation of parameter dependent integrals6 . This implies that

Now, part (a) follows easily by Z
f 2 C 2 (B(0, R)) and f (x) = x Ny GR (x, y)'(y)dy = 0.
S n 1 (0,R) | {z }
=0
The main part is to prove (b).
We first observe
Z Z Z
Ny GR (x, y)dy = Ny R (x, y) dy Ny n (x y) dy
Sn 1 (0,R) Sn 1 (0,R) Sn 1 (0,R)
| {z }
=0
Z Z
Prop.2
= 1+ n (x y) 1(y) dy n (x y)(N1 )(y) dy
for f = 1 B(0,R) Sn 1 (0,R)
| {z } | {z }
=0 =0
= 1 (independently of x).
6
Use that S n 1
(0, R) is compact.

44
Let y0 2 S n 1 (0, R), x 2 B(0, R) (x close to y0 ). Then
Z
f (x) '(y0 ) = Ny GR (x, y)('(y) '(y0 )) dy.
Sn 1 (0,R)

We want to estimate |f (x) '(y0 )|. For this we split the domain of integration for > 0 into 2 parts.

1 = y 2 Sn 1
(0, R) | |y y0 | 
n 1
2 = y2S (0, R) | |y y0 | > .

Figure 15: Splitting the integration domain into 2 parts: 1 and 2 for > 0.

Then,
Z Z
Ny GR (x, y)('(y) '(y0 )) dy  max |'(y) '(y0 )| |Ny GR (x, y)| dy
|y y0 | 1
| {z }
1 <0

Z 0 on S n 1
z }| {
 max |'(y) '(y0 )| Ny GR (x, y) dy
|y y0 | 1
Z
 max |'(y) '(y0 )| Ny GR (x, y) dy.
|y y0 | Sn 1 0,R)

By the above, we obtain

Z
Ny Gr (x, y)('(y) '(y0 ))dy  max |'(y) '(y0 )| · 1 (independent of x). (2.6)
1
|y y0 |

For integration of 2 we have the explicit formula for Ny GR (x, y). We assume that |x y0 | < . Then
2
Z Z
R |x|2 1
Ny GR (x, y)('(y) '(y0 )) dy  max 2 · |'(y)| · dy.
2
R !n y2S n 1 (0,R)
2
|x y|n
| {z }
2k'k1

on 2 ⇣
Here, we did |x y| |x y0 | + |y0 y| > ⇣ 2 = ⇣2 . Hence,

Z =Rn 1 !
z }| n {
R2 |x|2 1 n 1
Ny GR (x, y)('(y) '(y0 )) dy  · 2 · ||'||1 · · vol(S (0, R))
2
R !n (⇣/2)n
 C(', ⇣) · (R2 |x|2 ), (2.7)

45
where C is independent of x. Now we fix " > 0. We want to find a neighborhood V of y0 such that
|f (y) '(y0 )|  " for all x 2 V \ B(0, R).
Choose ⇣ such that
"
max |'(y) '(y0 )|  (2.8)
|y y0 | 2
⇢
"
this is possible since ' continuous. We set V = B(y0 , ) \ x R2 |x|2 < , where C is
2 2C(', ⇣)
the constant in (2.7).

Figure 16: Region where V is defined.

Then
Z Z  "
|f (x) '(y0 )|  ··· + · · · (2.6), (2.7), (2.8) + C(', ⇣)(R2 |x|2 ) < ".
1 2
2 | {z }
"
< on V
2

2.6 Applications of Theorem 1 to harmonic functions

Corollary 4. Let U ⇢ Rn open, f : U ! C harmonic, x0 2 U , R > 0 such that B(x0 , R) ⇢ U . Then
Z
R2 |x x0 |2 f (y)
f (x) = dy for x 2 B(x0 , R).
R !n S n 1 (x0 ,R) |x y|n

Proof. f is the unique solution of the Dirichlet problem for B(x0 , R) with ' = f |S n 1 (x ,R)
0
. We shift
x0 to 0 and apply Thm. 1.

Proposition 14. Let U ⇢ Rn be open, f : U ! C harmonic. Then f 2 C 1 (U ).

Shortly: Every harmonic function is smooth, not only C 2 .

Proof. Apply differentiation of parameter dependent integrals to Cor. 4. Then, f |B(x0 ,R) is of class C 1 .
Since for every x0 such that an R > 0 with B(x0 , R) ⇢ U exists, we obtain f 2 C 1 (U ).

Proposition 15 (Lionville). Let f : Rn ! C be harmonic. If f is bounded, i.e. |f (x)|  C, then f

is constant.
If f is real valued and bounded from above or below, then f is constant.

Proof. If f is real valued and bounded below, then fe = f + C 0 and harmonic. If f is real valued
and bounded from above, then fe = f C 0 and harmonic.
If f is complex valued and bounded, then Re (f ) and Im (f ) are harmonic, real valued, bounded from
above and below.
By this discussion it suffices to show if f : Rn ! R is a real valued, harmonic and non-negative
(f (x) 0), then f is constant.
Let x 2 Rn , let R > |x|. Then by Cor. 4 (with x0 = 0):
Z
R2 |x|2 f (y)
f (x) = dy, for |x y| |y| |x| = R |x|.
R !n S n 1 (0,R) |x y|n

46
Hence,
Z Pro. 4
R2 |x|2 1 mean value property R2 |x|2
f (x)  · f (y) dy = · Rn 1
!n f (0)
R !n (R |x|)n S n 1 (0,R) R !n (R |x|)n
R2 |x|2
= · Rn 2 · f (0),
(R |x|)n
| {z }
! 1
R!1

where x is fixed. Thus f (x)  f (0) and hence f has a maximum at 0. By the maximum principle for
B(0, R), we get that f |B(0,R) is constant. Since R was arbitrary, we obtain that f constant on Rn .

47