THE DIRICHLET PROBLEM

FOR THE LAPLACE OPERATOR

Stefano Meda

Università di Milano-Bicocca

c Stefano Meda 2013

ii

A Francesco
Contents

I The Dirichlet problem

via Perron’s method 1

1 The classical Dirichlet problem 3

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Background and preliminary results . . . . . . . . . . . . . . . 7
1.3 Subharmonic functions . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Maximum principle and uniqueness . . . . . . . . . . . . . . . 17
1.5 Green’s function . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 The Poisson kernel for the half space . . . . . . . . . . . . . . 24
1.7 The Poisson kernel for the ball . . . . . . . . . . . . . . . . . . 27
1.7.1 The two dimensional case . . . . . . . . . . . . . . . . 27
1.7.2 The higher dimensional case . . . . . . . . . . . . . . . 30
1.8 Perron’s method . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.9 The Lebesgue spine . . . . . . . . . . . . . . . . . . . . . . . . 45

II The Dirichlet problem

via integral equations 49

2 Linear operators on Banach spaces 53

2.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.2 Bounded operators . . . . . . . . . . . . . . . . . . . . . . . . 54
2.3 The spectrum of a linear operator . . . . . . . . . . . . . . . 57
2.4 The adjoint of a bounded operator . . . . . . . . . . . . . . . 58
2.5 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . 65

iii
iv CONTENTS

2.6 The spectra of compact operators . . . . . . . . . . . . . . . . 68

2.7 The spectral theorem . . . . . . . . . . . . . . . . . . . . . . . 75

3 Dirichlet via integral equations 79

3.1 Kernels of type α . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2 The double layer potential . . . . . . . . . . . . . . . . . . . . 88
3.3 The single layer potential . . . . . . . . . . . . . . . . . . . . . 98
3.4 Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

III The Dirichlet problem

via L2 methods 107
4.0.1 Introduction to Dirichlet’s principle . . . . . . . . . . . 109
4.0.2 Why weak solutions? . . . . . . . . . . . . . . . . . . . 113
4.0.3 Plan of Part III . . . . . . . . . . . . . . . . . . . . . . 114
4.1 Details left behind . . . . . . . . . . . . . . . . . . . . . . . . 115

5 Distributions and Sobolev spaces 119

5.1 Continuous functions and measures . . . . . . . . . . . . . . . 119
5.2 Smooth functions and distributions . . . . . . . . . . . . . . . 129
5.3 Derivatives of distributions . . . . . . . . . . . . . . . . . . . . 136
5.4 Fundamental solutions . . . . . . . . . . . . . . . . . . . . . . 141
5.5 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 145

6 Dirichlet L2 155
6.1 Poincare’s inequality . . . . . . . . . . . . . . . . . . . . . . . 155
6.2 Solution to the minimization problem . . . . . . . . . . . . . . 157
6.3 The dual of H01 (Ω) . . . . . . . . . . . . . . . . . . . . . . . . 162
6.4 Weak solutions and regularity . . . . . . . . . . . . . . . . . . 166
 Part I

The Dirichlet problem

via Perron’s method

1
Chapter 1

The classical Dirichlet problem

1.1 Introduction
Suppose that Ω is a bounded open set in Rn . We denote by ∆ the Laplace
operator, which acts on a function f in C 2 (Ω) by
n
X
∆f (x) = ∂j2 f (x) ∀x ∈ Ω.
j=1

The classical Dirichlet problem is the following: given a continuous func-

tion g on ∂Ω, find a function u in C 2 (Ω) ∩ C(Ω) such that
(
∆u = 0 in Ω
u|∂Ω = g.

This is a very challenging problem, which has been considered by many

outstanding mathematicians in the past two centuries. The reader is referred
to the interesting article [G] of L. Garding for an historical account of the
research on the Dirichlet problem until the first half of the twentieth century.
As we shall see, it is comparatively easy to prove that if a solution to
the Dirichlet problem exists, then it is unique. By contrast, it is entirely
nontrivial to prove that, under suitable assumptions on the domain Ω, the
Dirichlet problem is solvable.
In some textbooks, mainly bounded domains with C 2 boundary are con-
sidered. From the point of view of applications, this assumption is far inade-
quate. Indeed, Dirichlet problems in a square, or in domains of the plane with
polygonal boundary are quite common, if not paradigmatic. Thus, it seems

3
4 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

reasonable to focus on theories which at least cover domains with Lipschitz

boundary.
In this chapter, we take up the classical theory of harmonic functions on
a very general class of domains, which include Lipschitz domains. Then we
will illustrate Perron’s method of solving the Dirichlet problem and give a
characterisation of the domains in Rn where the Dirichlet problem is solvable.
We mainly follow [GT, Chapter 2].
There are many problems in Mathematics and in Physics that lead to
consider the Dirichlet problem. Here we recall two of them.

The first is concerned with the heat diffusion in a body Ω ⊂ Rn (n = 2, 3

are the most important cases). Assume that a fixed temperature distribution
at the boundary is maintained by a heating and refrigeration system. By
suitably normalising the physical constants involved, we may assume that
the temperature u(x, t) of the point x ∈ Ω at time t satisfies the following
equation (known as the heat equation)

∂t u(x, t) − ∆u(x, t) = 0 ∀x ∈ Ω ∀t > 0, (1.1.1)

where the Laplacian ∆ acts on the x variable. Denote by g(x) the tempera-
ture at which the heating and refrigeration system keeps the point x ∈ ∂Ω.
Then we must have

u(x, t) = g(x) ∀x ∈ ∂Ω ∀t > 0. (1.1.2)

Of course, the body has an initial temperature u(x, 0) at each point x ∈ ∂Ω.
Given that the boundary temperature is kept at a steady state, it is plau-
sible that the system will evolve towards an equilibrium state u(x), that is

u(x) = lim u(x, t).

t→∞

Since for all T > 0 the function uT (x, t) := u(x, T + t) satisfies the boundary
value problem (1.1.1)–(1.1.2), it is reasonable to expect that the same will
happen to u. Since u does not depend on t, u will satisfy
(
∆u = 0 in Ω
u|∂Ω = g.

The second problem leading to the Dirichlet boundary value problem

is internal to Mathematics. Riemann tackled the problem of constructing
a conformal mapping ϕ between a simply connected domain Ω in the
1.1. INTRODUCTION 5

complex plane and the unit disc D := {z ∈ C : |z| < 1}. By this we
mean that ϕ is a holomorphic bijection between Ω and D. Suppose that Ω
has smooth boundary. We look for a homeomorphism ϕ : Ω → D, which,
restricted to Ω, is a biolomorphism between Ω and D. We may assume
that 0 ∈ Ω. By possibly composing ϕ with a Möbius transformation, we
may assume that ϕ(0) = 0. If such a conformal mapping ϕ exists, then
z 7→ ϕ(z)/z is nonvanishing (recall that ϕ0 (z) 6= 0 for all z ∈ Ω, for ϕ is
conformal), whence there exists a holomorphic function f on Ω such that

ϕ(z)
= ef (z) ∀z ∈ Ω.
z
Therefore
log ϕ(z) = log |z| + Re f (z) ∀z ∈ Ω.
Clearly Re f is harmonic in Ω, for it is the real part of a holomorphic function,
and it is a solution to the boundary value problem
(
∆u = 0 in Ω
u|∂Ω = − log |z|.

Before closing this section, we illustrate an important method, due to

Daniel Bernoulli and known as the method of separation of variables,
which allows us to construct a solution to the Dirichlet problem in some cases
where Ω has suitable geometric features.
Denote by R the strip

{(x, y) ∈ R2 : 0 ≤ x ≤ π, y ≥ 0}.

Assume that the stationary temperature u(x, y) is continuous and bounded

and satisfies the following boundary conditions

u(0, y) = 0 = u(π, y) ∀y > 0, u(x, 0) = f (x) ∀x ∈ [0, π], (1.1.3)

where f is an assigned continuous function. We seek functions of the form

u(x, y) = X(x) Y (y) that satisfy the Laplace equation and the boundary
conditions on the vertical edges of the strip. We substitute u in the Laplace
equation, and obtain
X 00 Y 00
=− .
X Y
Since the left hand side depends only on x and the right hand side depends
only on y, neither can depend on either. Hence both are equal to a constant,
6 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

which we write as −c, where c is, for the time being, a complex number.
Then, we are led to solve the ordinary differential equations

X 00 + c X = 0 and Y 00 − c Y = 0.

Furthermore, X must be a solution to the following boundary value problem

X 00 + c X = 0 and X(0) = 0 = X(π). (1.1.4)

It is straightforward to check that this problem has a nontrivial solution only

if c is of the form k 2 , where k is a positive integer (prove this!). The solutions
of (1.1.4) are then all the multiples of sin(kx). The function Y must be a
solution to the following problem

Y 00 − k 2 Y = 0 and Y bounded on [0, ∞). (1.1.5)

All the solutions to this problem are multiples of e−ky . Thus, we are led to
consider the functions

uk (x, y) := e−ky sin(kx).

They satisfy

∆uk = 0, uk (0, y) = 0 = uk (π, y) ∀y > 0.

Unless the datum f is one of the functions sin(kx), none of the functions uk
will match the boundary condition uk (x, 0) = f (x). The idea to circumvent
this difficulty is to consider superpositions of the functions uk , i.e., to see
whether functions of the form
∞
X
ck e−ky sin(kx) (1.1.6)
k=1

satisfy the given Dirichlet problem. Of course, this idea is suggested by the
fact that the Laplace operator v 7→ ∆v is linear. At least formally, if the
function above satisfies the Dirichlet problem, then its value at (x, 0) must
be equal to f (x). In other words, f should have the following Fourier sine
series expansion
∞
X
f (x) = ck sin(kx).
k=1

It is a nontrivial fact, proved long after the appearance of Fourier’s original

paper (1822), that the sequence (2/π)1/2 sin(kx) is a complete orthonor-
mal system in L2 ((0, π)). We shall come back to this fact later. This will force
ck to be the k th Fourier coefficient of the sine expansion of the function f .
1.2. BACKGROUND AND PRELIMINARY RESULTS 7

Note that if f is the restriction to [0, π] of a π periodic function in C 2 (R)

such that f (0) = 0, then
ck = O(k −2 ),
and the series in (1.1.6) converges uniformly on [0, π] × [0, ∞). Therefore
its sum is a continuous function on R that matches the boundary values.
Furthermore, the sum of the series is infinitely many times differentiable at
all the points in R with y > 0, and satisfies the Laplace equation therein.

Exercise 1.1.1 Prove all the assertions

 above on Fourier sine series, except
for the completeness of the system (2/π)1/2 sin(kx) .

Exercise 1.1.2 Show that the method of separation of variables for the
Dirichlet problem discussed above leads, in the case where f (x) = 1, to the
solution
4 h −y 1 1 i
u(x, y) := e sin x + e−3y sin(3x) + e−5y sin(5x) + · · ·
π 3 5
2 sin x
= arctan .
π sinh y
Show that u satisfies the boundary conditions except at the corners. Draw
the isothermals for small x and y. Hint. To find the sum of the series within
square brackets, write z = x + iy and observe that the series is the imaginary
part of a power series in the variable eiz , whose sum is (1/2) log[(1+eiz )/(1−
eiz )].

1.2 Background and preliminary results

We need the following definitions and results. A subset of Rn is a domain
if it is an open connected set.
Now we recall some basic facts from advanced calculus. A subset S of
Rn is a hypersurface of class C k if for every x0 ∈ S there exists an open
subset V of Rn containing x0 and a real valued function φ ∈ C k (V ) such that
∇φ does not vanish on S ∩ V and

S ∩ V = x ∈ V : φ(x) = 0 .

For the sake of definiteness, suppose that ∂n φ does not vanish on S ∩ V .

Then, by the implicit function theorem, there exists a C k function ψ such
that

xn = ψ x1 , . . . , xn−1
8 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

for all x1 , . . . , xn−1 , xn in S ∩ V . For convenience, denote by x0 the point

(x1 , . . . , xn−1 ) in Rn−1 . Note that the map

(x0 , xn ) 7→ x0 , xn − ψ(x0 )

maps V ∩ S onto a neighbourhood of the point x00 of the hyperplane xn = 0.

The inverse of the map

x0 7→ x0 , ψ(x0 ) ,

which is defined in a suitable neighbourhood of x00 , gives a local chart of S

around x0 .
Now, for every x ∈ S, the vector ∇φ(x) is orthogonal to S (i.e., it is
orthogonal to the tangent plane to S at x0 ). We shall assume that S is
oriented, i.e. there is a choice of a unit normal vector ν(x) at each point x
of S that varies continuously with x. In particular

∇φ(x)
ν(x) = ± .
∇φ(x)

This formula shows that the normal field x 7→ ν(x) is of class C k−1 on S.
We say that a domain Ω has C k boundary if ∂Ω is a hypersurface of class
k
C . We recall the classical divergence theorem.

Theorem 1.2.1 Suppose that Ω is a bounded domain with C 1 boundary, and

denote by ν the unit outward normal to Ω. Suppose that w ∈ C 1 (Ω). Then
Z Z
div w dV = w · ν dσ,
Ω ∂Ω

where σ denotes the surface measure of ∂Ω.

In the applications, the boundary does not always satisfy the assumptions
of Theorem 1.2.1. There is a generalisation of the divergence theorem which
covers domains with a very general boundary. It is beyond the scope of these
notes to deal with such generalisations, for which the reader is referred to the
book of Evans and Gariepy [EG]. We just make a few comments concerning
domains with Lipschitz boundary.
Recall that a map ϕ : Rn−1 → R is Lipschitz if there exists a constant L
such that
ϕ(x) − ϕ(y) ≤ L |x − y| ∀x, y ∈ Rn−1 .
1.2. BACKGROUND AND PRELIMINARY RESULTS 9

The infimum of all the constants L such that the above inequality holds is
called the Lipschitz constant of ϕ.
Lipschitz functions are possibly not differentiable at some points (for in-
stance, x 7→ |x| is Lipschitz with constant 1, but it is not differentiable at
the origin), but the set where it is not differentiable is small in the measure
theoretic sense. In fact, the following nontrivial result holds. We refer the
reader to [EG] for the proof of this result.

Theorem 1.2.2 (Rademacher) Suppose that ϕ is Lipschitz. Then the set

of points where ϕ is not differentiable is of null measure.

Definition 1.2.3 A Lipschitz domain is a domain Ω with the following

property:
for each y ∈ ∂Ω there exists a system of coordinates (x0 , xn ) (x0 is an (n − 1)-
dimensional vector) centred at y, a ball Br (y), a neighbourhood U of x0 = 0,
and a Lipschitz map ϕ : U → R, with ϕ(0) = 0, such that

(i) ∂Ω ∩ Br (y) = {(x0 , xn ) ∈ Br (y) : xn = ϕ(x0 ) for all x0 in U }

(ii) Ω ∩ Br (y) = {(x0 , xn ) ∈ Br (y) : xn > ϕ(x0 ) for all x0 in U }.

A consequence of the definition and of Rademacher’s theorem is that a Lips-

chitz domain Ω admits a well defined outward unit normal ν at almost every
point of ∂Ω (with respect to the surface measure). The surface measure of
∂Ω has the following expression in terms of the local coordinates (x0 , ϕ(x0 ))
in a neighbourhood of the point y (see Definition 1.2.3 for the notation)
q
2
dσ(x , ϕ(x )) = 1 + ∇ϕ(x0 ) dx0 .
0 0

The following generalised divergence theorem holds.

Theorem 1.2.4 Suppose that Ω is a bounded domain with Lipschitz bound-

ary, and denote by ν the unit outward normal to Ω, which is defined at almost
every point of the boundary. Suppose that w ∈ C 1 (Ω). Then
Z Z
div w dV = w · ν dσ,
Ω ∂Ω

where σ denotes the surface element of ∂Ω.

Corollary 1.2.5 Suppose that Ω is a bounded Lipschitz domain, and that u

and v are in C 2 (Ω). The following hold:
10 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

(i) Z Z
∆u dV = ∂ν u dσ;
Ω ∂Ω

(ii) ( first Green’s identity)

Z Z Z
v ∆u dV + ∇v · ∇u dV = v ∂ν u dσ;
Ω Ω ∂Ω

(iii) ( second Green’s identity)

Z Z


(v ∆u − u∆v) dV = v ∂ν u − u ∂ν v dσ.
Ω ∂Ω

(iv) ( integration by parts)

Z Z Z
v ∂j u dV + u ∂j v dV = u v νj dσ.
Ω Ω ∂Ω

Proof. To prove (i), just take w = ∇u in the divergence theorem.

To prove (ii), just take w = v ∇u in the divergence theorem.
By interchanging the role of u and v in (ii), and subtracting the resulting
equalities, we get (iii).
Finally, (iv) follows from the divergence theorem by taking w to be a
vector field, all of whose components vanish but the j th , which is equal to
uv. 2

Denote by ωn the surface measure of the sphere ∂B1 (0) in Rn .

Exercise 1.2.6 Prove that the measure of the unit ball in Rn is ωn /n.
Hint: use the divergence theorem for an appropriate vector field.

Exercise 1.2.7 By following the steps below, prove that the measure of the
unit ball in Rn is π n/2 /Γ(n/2 + 1) (here Γ denotes Euler’s Gamma function):

(i) show that if a > 0, then

Z

n/2
exp − a|x|2 dx = π/a


;
Rn
1.2. BACKGROUND AND PRELIMINARY RESULTS 11

(ii) compute
Z


I := p θ1 , . . . , θn−1 dθ1 · · · dθn−1 ,
R

by computing Z
exp − a|x|2 dx


Rn

with (i), and using polar coordinates;

(iii) compute V B1 (0) , by integrating the characteristic function of B1 (0)
in polar coordinates.

Exercise 1.2.8 Suppose that Ω is a domain in Rn . Prove that the diver-

gence operator is the “adjoint” of the gradient, in the sense that for every
smooth vector field w and every smooth function ϕ with compact support
contained in Ω Z Z
ϕ div w dV = − ∇ϕ · w dV.
Ω Ω

Exercise 1.2.9 Suppose that φ is a C 2 diffeomorphism between the do-

mains Ω and φ(Ω) and that f is a function in C 2 (φ(Ω)). We denote by x the
variable in Ω, by y that in φ(Ω), and by J(x) the determinant of the differ-
ential map φ0 (x). Prove the following change of variables formulae (in these
formulae y = φ(x), and the gradient of a scalar function is a row vector):

(i) ∇y f (y) = ∇x f ◦ φ (x) φ0 (x)−1 ;

1
divx J (F ◦ φ) [(φ0 )t ]−1 (x);
 
(ii) divy F(y) =
J(x)

1
−1 
divx J ∇x (f ◦ φ) (φ0 )t φ0

(iii) ∆y f (y) = (x);
J(x)

(iv) if φ is an orthogonal transformation, then the columns of φ0 (x) are

of the form aj uj , where u1 , . . . , un are orthonormal vectors. Then
φ0 (x)t φ0 (x) = diag(a21 , . . . , a2n ) and J(x) = a1 · · · an . Compute ∆y f (y);

(v) prove that polar coordinates are associated to an orthogonal transfor-

mation, and compute the Laplacian in polar coordinates by using (iii)
above, or by specialising the formula found in (iv);
12 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

(vi) suppose that φ is a conformal mapping, i.e., that φ0 (x) = %(x) U (x),
where %(x) is a scalar factor and U (x) is an orthogonal matrix. Prove
that
1  n−2 
∆y f (y) = div x % ∇x (f ◦ φ) (x);
%(x)n

(vii) prove that the inversion map y = x/|x|2 is a conformal mapping

between appropriate domains, with conformal factor |x|−2 . Compute
the Laplace operator with respect to the new coordinates;

(viii) suppose that Ω is a domain in C and that φ : Ω → f (Ω) is a one-to-one

holomorphic function. Prove that

∆f (w) = |φ0 (z)|2 (∆(f ◦ φ−1 ) φ−1 (w) ,

where w = φ(z).

Hint: Exercise 1.2.8 may be helpful to prove (ii).

Exercise 1.2.10 Suppose that u is a smooth function on R2 \ {0}. Prove

that
1 1
∆u(r, θ) = ∂r2 u + ∂r u + 2 ∂θ2 u.
r r
Then find all radial solutions to the equation ∆u = 0 in R2 \ {0}.

Exercise 1.2.11 Suppose that u is a smooth function on R3 \ {0}. Prove

that
2 1 h 1 i
∆u(r, θ, ψ) = ∂r2 u + ∂r u + 2 ∂ 2
u + ∂ 2
u + cot ψ ∂ ψ u .
r r (sin ψ)2 θ ψ

Then find all radial solutions to the equation ∆u = 0 in R3 \ {0}. Prove also
that the differential operator within square brackets in the formula above is
the Laplace–Beltrami operator on the sphere S2 with respect to the Rieman-
nian metric induced by the standard Euclidean metric of R3 .

1.3 Subharmonic, superharmonic and harmonic

functions
Definition 1.3.1 Suppose that Ω is a domain in Rn . A function u ∈ C 2 (Ω)
is subharmonic (superharmonic, harmonic) if ∆u ≥ 0 (≤ 0, = 0) in Ω.
1.3. SUBHARMONIC FUNCTIONS 13

Exercise 1.3.2 Prove that the real and the imaginary parts of a holomor-
phic function in a domain Ω are harmonic in Ω. Prove that the functions
x2 − y 2 and xy are harmonic in R2 , and so are the functions rj cos(jθ) and
rj sin(jθ) for each nonnegative integer j.

Exercise 1.3.3 Suppose that u is harmonic in Rn \ {0}. Prove that the

function
u∗ (x) = |x|2−n u x/|x|2

is harmonic in Rn \ {0}. Hint: Exercise 1.2.9 (vii) may be useful.

Exercise 1.3.4 Suppose that Ω is a domain in the complex plane, that

φ : Ω → φ(Ω) is a one-to-one holomorphic mapping, and that f is har-
monic in φ(Ω). Prove that the function f ◦ φ is is harmonic in φ(Ω). Hint:
Exercise 1.2.9 (viii) may be useful.

Exercise 1.3.5 Prove that if u is harmonic in Ω, and Ω0 ⊂⊂ Ω, then

Z
∂ν u dσ = 0.
∂Ω

Conversely, show that if u ∈ C 2 (Ω) satisfies

Z
∂ν u dσ = 0
∂B

for every ball B ⊂⊂ Ω, then u is harmonic in Ω.

Exercise 1.3.6 For which values of α is the function x 7→ |x|α subhar-

monic? Answer: α + n − 2 ≥ 0. This shows that if n ≥ 2, then a convex
function need not be subharmonic.

The next result relates subharmonic, superharmonic and harmonic functions

with their spherical and solid means over balls. In fact, these kind of functions
may be characterised by the behaviour of their means, as we shall see later.
For a ball B, we denote by σ(∂B) the surface measure of ∂B and by V (B)
the Lebesgue measure of B.

Theorem 1.3.7 (Mean value inequalities) Suppose that Ω is a domain.

The following hold:
14 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

(i) for every subharmonic function u and every ball B ⊂⊂ Ω

Z Z
1 1
u(y) ≤ u dσ and u(y) ≤ u dV ;
σ(∂B) ∂B V (B) B

(ii) for every superharmonic function u and every ball B ⊂⊂ Ω

Z Z
1 1
u(y) ≥ u dσ and u(y) ≥ u dV ;
σ(∂B) ∂B V (B) B

(iii) for every harmonic function u and every ball B ⊂⊂ Ω

Z Z
1 1
u(y) = u dσ and u(y) = u dV.
σ(∂B) ∂B V (B) B

Proof. We shall prove (i). The proof of (ii) is similar and is omitted. Part
(iii) is a straightforward consequence of (i) and (ii).
To prove (i), suppose that B = BR (y), that 0 < ρ < R, and observe that,
by Corollary 1.2.5 (i),
Z
0≤ ∂ν u dσ
∂Bρ (y)
ω0
Z
= ∇u(y + ω 0 ) · 0 dσρ (ω 0 )
∂Bρ (0) |ω |
Z
= ρn−1 ∇u(y + ρω) · ω dσ1 (ω)
∂B1 (0)
Z
n−1 d 
=ρ u(y + ρω) dσ1 (ω)
∂B1 (0) dρ
Z
n−1 d
=ρ u(y + ρω) dσ1 (ω)
dρ ∂B1 (0)
h 1 Z
n−1 d
i
=ρ u dσ .
dρ ρn−1 ∂Bρ (y)

Now we integrate both sides with respect to ρ between ε and R and obtain
Z Z
1 1
0 ≤ n−1 u dσ − n−1 u dσ.
R ∂BR (y) ε ∂Bε (y)

Since u is continuous in y,
Z
1
lim u dσ = ωn u(y).
ε→0 εn−1 ∂Bε (y)
1.3. SUBHARMONIC FUNCTIONS 15

By combining the last two formulae, we obtain that

Z
1
u(y) ≤ u dσ,
ωn Rn−1 ∂BR (y)

which is equivalent to the first formula in (i).

To prove the second formula in (i), we multiply by Rn−1 both sides of
the inequality above and integrate with respect to R between 0 and r. We
obtain Z r Z
1 1
u(y) ≤ dR u dσ,
n ωn 0 ∂BR (y)

which is equivalent to the required inequality. 2

Exercise 1.3.8 Find all harmonic functions on an interval I of R. Give a

characterisation of subharmonic functions on I. Conclude that a continuous
function u on I is subharmonic if and only if for every interval [a, b] ⊂⊂ I

u(x) ≤ Ha,b (x) ∀x ∈ [a, b]

where Ha,b is the harmonic function on [a, b] such that Ha,b (a) = u(a) and
Ha,b (b) = u(b).

It is an important fact that a converse of Theorem 1.3.7 holds.

Theorem 1.3.9 Suppose that u is a continuous function in the domain Ω

and that for every ball B ⊂⊂ Ω the following holds:
Z
1
u(cB ) = u dσ,
σ(∂B) ∂B

where cB denotes the centre of B. Then u is harmonic in Ω.

Proof. Denote by φ aRsmooth radial function with support contained in the

ball B1 (0) such that φ dV = 1, and denote by ψ its profile, i.e., ψ(|x|) =
φ(x). For every ε > 0, set φε (x) = ε−n φ(x/ε). Clearly the support of φε is
contained in Bε (0). If x belongs to

Ωε := {x : B ε (x) ⊂ Ω},
16 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

then the support of x 7→ φε (x − y) is contained in Ω. Observe that

Z Z
u(y) φε (x − y) dy = u(x − εy) φ(y) dy
Rn |y|<1
Z 1Z
= u(x − εry 0 ) ψ(r) rn−1 dσ(y 0 ) dr
0 |y 0 |=1
Z 1
= ωn u(x) ψ(r) rn−1 dr
Z 0
= u(x) φ dV
Rn
= u(x).

Now, the left hand side can be differentiated infinitely many times (see Propo-
sition ?? (iii), whence u is in C ∞ (Ωε ). Since ε is arbitrary, u is in C ∞ (Ω).
It remains to show that u is harmonic. Suppose that B r (x) ⊂ Ω. By
Corollary 1.2.5 (i),
Z Z
∆u dV = ∂ν u dσ
Br (x) ∂Br (x)
Z
ω
= grad u(x + ω) · dσ(ω)
|ω|=r |ω|
Z
=r n−1
grad u(x + rω 0 ) · ω 0 dσ(ω 0 )
|ω|=1
Z
n−1 d
=r u(x + rω 0 ) dσ(ω 0 ).
dr |ω0 |=1

The last integral is equal to

Z
1
ωn u dσ,
σ(∂Br (x)) ∂Br (x)

which, in turn is equal to ωn u(x), for u possesses the mean value property
by assumption. Thus, we have proved that
Z
d
∆u dV = ωn rn−1 u(x) = 0
Br (x) dr

for every ball Br (x) such that B r (x) ⊂ Ω. Since ∆u is continuous, ∆u = 0,

as required. 2
1.4. MAXIMUM PRINCIPLE AND UNIQUENESS 17

Exercise 1.3.10 Show that if u is a function in C 2 (Ω) and x0 ∈ Ω, then

Z
2n h 1 i
−∆u(x0 ) = lim 2 u(x0 ) − u(x 0 + r ω) dσ(ω) .
r↓0 r σ(Sn−1 ) Sn−1

Deduce that if u satisfies the mean value inequality of Theorem 1.3.7 (i),
then u is subharmonic.

1.4 Maximum principle and uniqueness for

the Dirichlet problem
In this section we derive some consequences of the mean value inequalities
proved in the previous section. In particular, we shall prove the maximum
and the minimum principles for harmonic functions. As a corollary we shall
obtain that the classical Dirichlet problem on a domain Ω has, at most, one
solution. The problem of the existence of solutions to the Dirichlet problem
will be addressed later.

Theorem 1.4.1 Suppose that Ω is a domain in Rn , and that u ∈ C 2 (Ω).

The following hold:

(i) (strong maximum principle) if u is subharmonic and there exists

a point y ∈ Ω such that u(y) = supΩ u, then u is constant;

(ii) (strong minimum principle) if u is superharmonic and there exists

a point y ∈ Ω such that u(y) = inf Ω u, then u is constant;

(iii) if u is harmonic, then u cannot assume an interior maximum or min-

imum, unless it is constant.

Proof. We prove (i). The proof of (ii) is similar and is omitted. Part (iii) is
a direct consequence of (i) and (ii).
Set M := supΩ u, and

ΩM := {x ∈ Ω : u(x) = M }.

Clearly ΩM contains y, and it is closed, for u is continuous. We shall prove

that ΩM is open. It will follow that ΩM = Ω, for Ω is connected by assump-
tion, i.e. u = M in Ω, as required.
18 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

To prove that ΩM is open, choose z in ΩM . Observe that u − M is

subharmonic (∆(u − M ) = ∆u), so that, by the mean value inequality,
Z
1
0 = u(z) − M ≤ (u − M ) dV ≤ 0
V (B) B

for every ball B with radius small enough. Therefore

Z
(u − M ) dV = 0.
B

If u were strictly less than M on an open subset of B, then the integral

above would be strictly negative. Therefore u = M on B, and ΩM is open,
as required. 2

Corollary 1.4.2 Suppose that Ω is a bounded domain in Rn , and that u ∈

C 2 (Ω) ∩ C(Ω). The following hold:

(i) (weak maximum principle) if u is subharmonic, then

sup u = sup u;
Ω ∂Ω

(ii) (weak minimum principle) if u is superharmonic, then

inf u = inf u;
Ω ∂Ω

(iii) if u is harmonic, then

inf u ≤ u ≤ sup u.
∂Ω ∂Ω

Proof. To prove (i), observe that if supΩ u > sup∂Ω u, then there exists a
point y in Ω for which u(y) = supΩ u. Consequently u is constant in Ω by
Theorem 1.4.1 (i). Since u is continuous on Ω, supΩ u = sup∂Ω u, thereby
contradicting the assumption.
The proof of (ii) is almost verbatim the same as the proof of (i), and is
omitted.
Part (iii) follows directly from (i) and (ii). 2
1.5. GREEN’S FUNCTION 19

Exercise 1.4.3 Suppose that u and v are in C 2 (Ω) ∩ C(Ω) and that
u|∂Ω = v|∂Ω .
Prove that if u is harmonic and v is subharmonic, then v ≤ u on Ω. This
result justifies the term subharmonic. State and prove a corresponding result
for superharmonic functions.

A noteworthy consequence of Corollary 1.4.2 is the following uniqueness re-

sult for the Dirichlet problem.

Theorem 1.4.4 Suppose that u and v are functions in C 2 (Ω) ∩ C(Ω) that
solve the Dirichlet problem
(
∆u = f in Ω
.
u|∂Ω = g
Then u = v.

Proof. Set w = u − v. Clearly w solves

(
∆w = 0 in Ω
w|∂Ω = 0.

In particular, w is a harmonic function in C 2 (Ω) ∩ C(Ω) which vanishes on

∂Ω. By Corollary 1.4.2 (iii), w = 0 on Ω, i.e. u = v, as required. 2
We wish to emphasize the fact that no claims are made concerning the exis-
tence of a solution to the Dirichlet problem.

1.5 Green’s function

In this section we establish an important representation formula for solutions
of the Dirichlet problem. We emphasize that we do not prove the existence
of a solution, rather we give a formula for the solution under the assumption
that a solution exists.

Definition 1.5.1 The Newtonian potential in Rn is the function N :

Rn \ {0} → R, defined by
1


 log |x| if n = 2
N (x) = 2π 1 (1.5.1)
 |x|2−n if n ≥ 3.
(2 − n) ωn

20 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Given a point x in Rn , the Newtonian potential with pole x is the function

N (x − ·).

Note that if n ≥ 3, then N is a negative function. By contrast, if n =

2, then N has not a definite sign. This simple fact will have far reaching
consequences. The reason for the normalisation of N will be clarified below
(see Definition 5.4.2). It will be proved later that N is a fundamental solution
of the Laplace operator: it will play an important role in finding distributional
solutions to the Poisson equation

∆u = f,

where f is a given datum. All this will be discussed in Section 5.4.

Exercise 1.5.2 Prove, by direct calculation, that N is harmonic in Rn \

{0}.

Exercise 1.5.3 Prove that

Z
∂ν N (x − y) dσ(y) = 1,
∂Br (x)

Exercise 1.5.4 Recall that the Laplacian in Rn may be written in polar

coordinates as follows
n−1 1
∆ = ∂r2 + ∂r + 2 ∆Sn−1 ,
r r
where ∆Sn−1 is the Laplace–Beltrami operator on the sphere Sn−1 (it is a
second order differential operator in the angular variables only). Find all
radial harmonic functions in Rn .

We now establish Green’s representation formula.

Theorem 1.5.5 (Green’s representation formula) Suppose that Ω is a

bounded Lipschitz domain and that u ∈ C 2 (Ω). The following hold:

(i) for every x in Ω

Z h i
u(x) = u(y) ∂ν N (x − y) − ∂ν u(y) N (x − y) dσ(y)
∂Ω Z
(1.5.2)
+ N (x − y) ∆u(y) dV (y);
Ω
1.5. GREEN’S FUNCTION 21

(ii) if the support of u is a compact set in Ω, then

Z
u(x) = N (x − y) ∆u(y) dV (y) ∀x ∈ Ω;
Ω

(iii) if u is harmonic in Ω, then

Z h i
u(x) = u(y) ∂ν N (x − y) − ∂ν u(y) N (x − y) dσ(y) ∀x ∈ Ω;
∂Ω

(iv) suppose that x ∈ Ω and that there exists a function hx ∈ C 1 (Ω) which is
harmonic on Ω and satisfies (hx )|∂Ω = −N (x − ·)|∂Ω . Denote by G(x, ·)
the function hx + N (x − ·). Then
Z Z
u(x) = u(y) ∂ν G(x, y) dσ(y) + G(x, y) ∆u(y) dV (y).
∂Ω Ω

The normal derivative of G in the integral above is taken with respect

to the variable y.

Proof. We prove the theorem in the case where n ≥ 3. The case n = 2 is left
to the reader.
First we prove (i). Consider the domain Ω \ B r (x), for r small. Apply
second Green’s identity (see Corollary 1.2.5 (iii)) with N (x − ·) in place of v
and Ω \ B r (x) in place of Ω. We obtain
Z
N (x − y) ∆u(y) dV (y)
Ω\B r (x)
Z
 
= N (x − y)∂ν u(y) − u(y) ∂ν N (x − y) dσ(y)
∂Ω Z
 
+ N (x − y)∂ν u(y) − u(y) ∂ν N (x − y) dσ(y).
∂Br (x)

The required formula will follow from this and the dominated convergence
theorem once we prove that
Z
lim N (x − y) ∂ν u(y) dσ(y) = 0 (1.5.3)
r↓0 ∂Br (x)

and that Z
lim u(y) ∂ν N (x − y) dσ(y) = u(x). (1.5.4)
r↓0 ∂Br (x)
22 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

To prove (1.5.3), note that

N (x − y) ≤ C r2−n ∀y ∈ ∂Br (x)
and that
∂ν u(y) ≤ max ∇u(y) ,
y∈Ω

which is finite, because u ∈ C 2 (Ω). Thus,

Z
N (x − y) ∂ν u(y) dσ(y) ≤ C max ∇u(y) r2−n σ ∂Br (x) ,


∂Br (x) y∈Ω


which tends to 0 as r tends to 0, for σ ∂Br (x)  rn−1 , as required.
To prove (1.5.4), write the integral in (1.5.4) as
Z Z
 
u(y) − u(x) ∂ν N (x − y) dσ(y) + u(x) ∂ν N (x − y) dσ(y).
∂Br (x) ∂Br (x)

Since u is smooth,
|u(y) − u(x)| ≤ r max |∇u| ∀y ∈ ∂Br (x).
Ω

Recall that N is homogeneous of degree 2 − n, hence its partial derivatives

are homogeneous of degre 1 − n. Thus,
∂ν N (x − y) ≤ C r1−n ∀y ∈ ∂Br (x).
Therefore
Z
u(y) − u(x) ∂ν N (x − y) dσ(y) ≤ C r r1−n σ ∂Br (y) ,
 

∂Br (x)

which tends to 0 as r tends to 0. Finally, by Exercise 1.5.3,

Z
∂ν N (x − y) dσ(y) = 1,
∂Br (x)

which completes the proof of (i).

Note that (ii) and (iii) are direct consequences of the representation for-
mula established in (i).
It remains to prove (iv). Write N (x − y) = G(x, y) − hx (y) in (1.5.2). We
obtain
Z h i Z
u(x) = u ∂ν G(x, ·) − ∂ν u G(x, ·) dσ + G(x, y) ∆u(y) dV (y)
∂Ω Z Z ZΩ
+ ∂ν u hx dσ − u ∂ν hx dσ − hx ∆u dV.
∂Ω ∂Ω Ω
1.5. GREEN’S FUNCTION 23

Note that, by second Green’s identity and the fact that hx is harmonic,
Z Z Z
∂ν u hx dσ = u ∂ν hx dσ + hx ∆u dV,
∂Ω ∂Ω Ω

and the required formula follows. 2

The last part of the theorem above suggest that the function G defined
therein may play an important role in the theory. This justifies the following
definition.

Definition 1.5.6 Suppose that Ω is a bounded domain in Rn . Assume

that for every x ∈ Ω the Dirichlet problem
(
∆u = 0 in Ω
u|∂Ω = −N (x − ·)

is solvable, and that the solution hx is in C 1 (Ω). The function G : Ω × Ω \

{(z, z) : z ∈ Ω}, defined by

G(x, y) := hx (y) + N (x − y),

is called the Green function for the domain Ω.

Note that we do not assert the existence of a Green’s function for a generic
domain Ω. In fact, there are domains which do not admit a Green’s function.

Exercise 1.5.7 Prove that the Green’s function for Ω, if it exists, is unique.

We explicitly state a straightforward but important consequence of Theo-

rem 1.5.5 (iv): a representation formula for the solution to the Dirichlet
problem (
∆u = 0 in Ω
u|∂Ω = g,
under the assumption that a solution u exists and it is of class C 1 (Ω).

Corollary 1.5.8 Suppose that the domain Ω admits a Green’s function G.

If u ∈ C 1 (Ω) is a solution to the Dirichlet problem above, then
Z
u(x) = g(y) ∂ν G(x − y) dσ(y) ∀x ∈ Ω.
∂Ω
24 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Definition 1.5.9 Suppose that G is the Green function for the domain Ω.
The function (x, y) 7→ ∂ν G(x, y) (here the normal derivative is taken with
respect to the y variable) is called a Poisson kernel for the domain Ω.

Corollary 1.5.8 indicates that it is reasonable to produce efforts to determine

the Poisson kernel of a given domain Ω. This is virtually impossible in most
cases. However, the Poisson kernel may be explicitly found in some important
cases, where the domain Ω has a suitable shape. The next section is devoted
to the case of the upper half space in Rn .

Exercise 1.5.10 Suppose that G is the Green’s function of the bounded

domain Ω. Prove the following:

(i) G(x, y) = G(y, x) for every x, y in Ω, x 6= y;

(ii) G(x, y) < 0 for every x, y in Ω, x 6= y;

(iii) prove that G(x, y) ≤ 2 N (x − y) for every x and y in Ω;

R
(iv) if f is bounded in Ω, then Ω G(x, y) f (y) dV (y) → 0 as x → ∂Ω.

Hints: (i) for every z in Ω and every (small) ε > 0, denote by Ωε (z) the set Ω\
B ε (z). Then write the harmonic function G(x, ·) by using the representation
formula in Theorem 1.5.5 (iii) with Ωε (x) in place of Ω and similarly write
the harmonic function G(y, ·) by using the representation formula with Ωε (y)
in place of Ω. Then use second Green’s identity and the fact that G(x, ·) =
N (x − ·) + hx (·);
(ii) note that G(x, y) tends to −∞ if y tends to x. Therefore, G < 0
near x. Apply the maximum principle to the domain Ω \ Bε (x);
(iii) observe that, by the minimum principle, hx > 0 in Ω and use (ii)
and the definition of G;
(iv) use (iii) and the Lebesgue dominated convergence theorem.

1.6 The Poisson kernel for the half space

Our aim is to compute the Poisson kernel for the upper half space Rn+ in Rn ,
defined as follows

Rn+ := {(x0 , xn ) ∈ Rn−1 × R : xn > 0}.

1.6. THE POISSON KERNEL FOR THE HALF SPACE 25

First we need to compute the Green’s function for Rn+ . Fix a point x =
(x0 , xn ) in the upper half space, and consider the Newtonian potential N (x−·)
with pole x. We must find a function hx in C 1 (Rn+ ) that is harmonic on Rn+
and such that N (x − ·) + hx vanishes identically when xn = 0. Observe
that the Newtonian potential is, up to a constant, the potential generated
by a unit negative charge placed at the point x∗ := (x0 , −xn ). By symmetry,
its values on the hyperplane xn = 0 are the same of those of a Newtonian
potential generated by a unit negative charge placed at the point x∗ . Thus,
the Green’s function G of Rn+ is given by

 1 log |x − y|

∗ − y|
if n = 2
G(x, y) = 2π |x (1.6.1)
 1 ∗
 2−n 2−n

(2−n) ωn
|x − y| − |x − y| if n ≥ 3.

Exercise 1.6.1 Check that the function G defined above is the Green’s
function for the upper half space.

First we consider the case where n ≥ 3. By definition, for every x ∈ Rn+ and
y 0 ∈ Rn−1 the Poisson kernel P (x, y 0 ) is then given by
P (x, y 0 ) = ∂ν G(x, y 0 )
= −∂yn G(x, y 0 )
1 h y n − xn y n + xn i
=− − .
ωn |x − y|n |x∗ − y|n |yn =0
Observe that |x − y| = |x∗ − y| when yn = 0. Hence
2 xn
P (x, y 0 ) =
ωn |x − y|n
2 xn (1.6.2)
= ∀xn > 0, x0 , y 0 ∈ Rn−1 .
ωn |x0 − y 0 |2 + x2 n/2


n

A similar computation shows that the formula above holds also in the case
where n = 2.

Exercise 1.6.2 Compute the Poisson kernel for a quadrant of the plane.
Hint: refine the method of images illustrated above for the half plane.

Note that, given a function g on Rn−1 the Dirichlet problem

(
∆u = 0 in Rn+
u|∂Rn+ = g
26 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

may have more than one solution. In particular, if g vanishes identically, then
the null function and the function u(x0 , xn ) = xn both solve the Dirichlet
problem above. Thus, if we want to recover uniqueness, we must impose
further restriction on the solution. A typical statement which holds in this
case is the following.

Theorem 1.6.3 Suppose that g is continuous and bounded on Rn−1 . Then

there is a unique bounded continuous function u on Rn+ which solves the
Dirichlet problem (
∆u = 0 in Rn+
u|∂Rn+ = g.
Furthermore
Z
u(x) = g(y 0 ) P (x, y 0 ) dy 0
Rn−1
2
Z
xn (1.6.3)
= g(y 0 )
n/2 dy
0
∀x ∈ Rn+ ,
ωn Rn−1 |x0 − y 0 |2 + x2n
and u(x0 , 0) = g(x0 ) for every x0 ∈ Rn−1 .

Remark 1.6.4 Set

2 1
p(x0 ) := ∀x0 ∈ Rn−1 .
ωn |x0 |2 + 1 n/2

Note that p is in L1 (Rn−1 ) and that

P (x, y 0 ) = pxn (x0 − y 0 ),
where, as customary, we write pxn to be the dilated and normalised ver-
sion of p in L1 (Rn−1 ). Explicitly,
pxn (w0 ) = x1−n p w0 /xn ∀w0 ∈ Rn−1 ∀xn > 0.


n

Observe that kpxn k1 = kpk1 , thereby justifying the terminology. With this
notation, formula (1.6.3) may be rewritten as follows
u(x) = g ∗ pxn (x0 ) ∀x ∈ Rn+ , (1.6.4)
where ∗ stands for convolution on Rn−1 .

Exercise 1.6.5 Prove that (1.6.3) gives, indeed, a solution to the Dirichlet
problem on the upper half space with boundary datum g.

By using a celebrated theorem of Liouville and the Schwartz reflection princi-

ple, we may prove that if g is continuous and bounded, then (1.6.4) gives the
unique bounded solution to the Dirichlet problem in the upper half space.
1.7. THE POISSON KERNEL FOR THE BALL 27

1.7 The Poisson kernel for the ball

The aim of this section is to provide the solution to the classical Dirichlet
problem in the ball BR (0). The theory developed here is important at least
for two reasons: (i) it leads to an explicit formula for the solution; (ii) the
fact that the classical Dirichlet problem in every ball is solvable is a key
step towards the solution of the classical Dirichlet problem in more general
domains.

1.7.1 The two dimensional case

We first look at the two dimensional case, where the theory of functions of
a complex variable is of valuable help. Suppose that ϕ is a conformal map
between the domains Ω ⊂ C and ϕ(Ω). In particular, ϕ is a bijection between
Ω and ϕ(Ω) and ϕ0 (z) 6= 0 for every z in Ω.

Proposition 1.7.1 Suppose that Ω and ϕ are as above, that ϕ extends to

a biholomorphic map between Ω and ϕ(Ω) and assume that ϕ(Ω) admits a
Green’s function Gϕ(Ω) . Then Ω admits a Green’s function GΩ and
GΩ (z, ζ) = Gϕ(Ω) (ϕ(z), ϕ(ζ)


∀(z, ζ) ∈ Ω × Ω \ {(ω, ω) : ω ∈ Ω}.

Proof. For each ζ ∈ ∂Ω the point ϕ(ζ) is in ∂(ϕ(Ω)). Therefore for each
z∈Ω
Gϕ(Ω) (ϕ(z), ϕ(ζ) = 0,

because Gϕ(Ω) is the Green’s function for ϕ(Ω). For the same reason
∆w Gϕ(Ω) (w, ϕ(ζ)) = 0.
Since ϕ is conformal,
∆w Gϕ(Ω) (w, ϕ(ζ)) = |ϕ0 (z)|2 ∆z Gϕ(Ω) (ϕ(z), ϕ(ζ)),
by (1.2.9) (viii), whence GΩ (·, ζ) is harmonic for every ζ in Ω.
Finally,
1
GΩ (z, ζ) − log |z − ζ|
2π
1
= Gϕ(Ω) (ϕ(z), ϕ(ζ) −


log |z − ζ|
2π
1
1
= log |ϕ(z) − ϕ(ζ)| + hϕ(z) ϕ(ζ) − log |z − ζ|
2π 2π
1 ϕ(z) − ϕ(ζ)

= log + hϕ(z) ϕ(ζ)
2π z−ζ
28 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Observe that ϕ(z) − ϕ(ζ)/z − ζ tends to ϕ0 (z) as ζ tends to z. Therefore the

function
1 ϕ(z) − ϕ(ζ)

ζ 7→ log + hϕ(z) ϕ(ζ)
2π (z − ζ)
is harmonic in Ω \ {z} and bounded in a neighbourhood of z, hence it is
harmonic in Ω. We have proved that GΩ (z, ·) may be written as the sum of
the Newtonian potential with pole z and a function harmonic in Ω.
This completes the proof of the proposition. 2

We apply this result to the case where the domain Ω is just the unit disc
D := {z ∈ C : |z| < 1} and

1+z
ϕ(z) = i ∀z ∈ D.
1−z
It is not hard to check that ϕ maps D conformally onto the upper half
plane Π, and ∂D onto R ∪ ∞. In the last section we have proved that the
Green’s function for the upper half plane is given by

1 |w − ω|
GΠ (w, ω) := log .
2π |w − ω|

Denote by z and ζ the points in D such that

ϕ(z) = w and ϕ(ζ) = ω.

It is straightforward to check that w = ϕ(1/z), i.e., w is the image under ϕ

of the point obtained from z by the mapping

1
z 7→ ,
z
which is the inversion with respect to the circle {|z| = 1}. Observe that

GD (z, ζ) = GΠ ϕ(z), ϕ(ζ)

1 |ϕ(z) − ϕ(ζ)|
= log
2π |ϕ(1/z) − ϕ(ζ)|
1 | 1+z
1−z
− 1+ζ
1−ζ
|
= log 1+(1/z) 1+ζ
2π | 1−(1/z) − 1−ζ |
1 |z − ζ|
= log .
2π |1 − z ζ|
1.7. THE POISSON KERNEL FOR THE BALL 29

We write z = r eiφ and ζ = s eiθ , and obtain

1 |r − s ei(θ−φ) |
GD (r eiφ , s eiθ ) =log
2π |1 − rs ei(θ−φ) |
1 r2 − 2rs cos(θ − φ) + s2
= log
4π 1 − 2rs cos(θ − φ) + r2 s2
The Poisson kernel is then
1 h −2r cos(θ − φ) + 2 −2r cos(θ − φ) + 2r2 i
∂s GD (r eiφ , s eiθ )|s=1 = −
4π r2 − 2r cos(θ − φ) + 1 1 − 2r cos(θ − φ) + r2
1 1 − r2
= .
2π 1 − 2r cos(θ − φ) + r2
Therefore, if a solution u ∈ C 1 (D) to the Dirichlet problem
(
∆u = 0 in D
u|∂D = g,
exists, then
2π
1 − r2
Z
iφ 1
u(r e ) = g(eiθ ) dθ. (1.7.1)
2π 0 1 − 2r cos(θ − φ) + r2

The aim of the following series of exercises is to determine the Green’s func-
tion, hence the Poisson kernel, of certain domains in the complex plane that
are conformally equivalent to the unit disc.

Exercise 1.7.2 Compute the Green’s function and the Poisson kernel of
the half disc
F := {z ∈ C : |z| < 1, Im(z) > 0}.

Exercise 1.7.3 Prove that the map z 7→ ez is a conformal map of the strip
S := {z ∈ C : 0 < Im(z) < π}
onto the upper half plane. Then compute the Green’s function and the
Poisson kernel of the strip S.

Exercise 1.7.4 Prove that the map z 7→ z α/π is a conformal map of the
upper half plane onto the sector
Γ := {z ∈ C : 0 < arg(z) < α}.
Then compute the Green’s function and the Poisson kernel of the sector Γ.
30 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

1.7.2 The higher dimensional case

We now consider the case where n ≥ 3. First, we deal with the case where
R = 1. The idea is that the potential on the unit sphere ∂B1 (0) induced by a
unit charge at x 6= 0 is equal to the potential on ∂B1 (0) induced by a charge
of magnitude |x|2−n at the point x/|x|2 , obtained by inversion with respect
to ∂B1 (0) of x.

Lemma 1.7.5 Suppose that Y ∈ ∂B1 (0) and that x ∈ B1 (0) \ {0}. Then
x
|Y − x| = |x| Y − .
|x|

Proof. This lemma is geometrically obvious (just draw a picture). For an

analytic proof, just square both sides of the formula above and make the
required computation. 2

Proposition 1.7.6 Suppose that R > 0. The following hold:

(i) the function, defined for every x ∈ B1 (0) and for every y ∈ B1 (0) by

G1 (x, y) = N (x − y) − |x|2−n N x/|x|2 − y

1 2−n 
|x − y|2−n − |x|−1 x − |x| y

= ,
(2 − n) ωn

is the Green’s function of B1 (0);

(ii) the Poisson kernel for the ball BR (0) is given by

R2 − |x|2
PR (x, Y ) = ∀x ∈ BR (0) ∀Y ∈ ∂BR (0). (1.7.2)
ωn R |x − Y |n

Proof. To prove (i) observe that the first equation shows that G1 (x, ·) is
harmonic in B1 (0), and, by Lemma 1.7.5, the second proves that G1 (x, y) =
N (x − y) + hx (y), where hx is harmonic in B1 (0) and hx (Y ) = −N (x − Y )
when Y ∈ ∂B1 (0). Since the Green’s function, if it exists, is unique, G1 is
the Green’s function of B1 .
Next we prove (ii). It is straightforward to check that the function
1
GR (x, y) = G1 (x/R, y/R)
Rn−2
1.7. THE POISSON KERNEL FOR THE BALL 31

is the Green’s function for BR (0). Indeed, by Lemma 1.7.5, if Y ∈ ∂BR (0)
and x ∈ BR (0), then

|x| x
|Y − x| = Y −R . (1.7.3)
R |x|

Clearly GR (x, ·) is harmonic in BR (0). Furthermore, if Y ∈ ∂BR (0), then

1 2−n
h x Y 2−n R x |x| Y 2−n i
GR (x, Y ) = R − − −
(2 − n) ωn R R |x| R R R
1 h
2−n R |x| 2−n i
= |x − Y | − x− Y
(2 − n) ωn |x| R
= 0,

where the last equality follows from (1.7.3). Denote by ν the outward normal
of ∂BR (0) at the point Y . By the definition of Poisson kernel,

PR (x, Y ) = ∂ν GR (x, Y )
|x| R
1 h Y −x Y |x| R
Y − |x|
x Y i
= n
· − n ·
ωn |x − Y | |Y | R R
x− |x|
Y |Y |
|x| R

R2 − |x|2
= ,
ωn R |x − Y |n

as required. 2

A direct consequence of Corollary 1.5.8 and Definition 1.5.9 is the following

representation formula for solutions to the Dirichlet problem on BR (0).

Corollary 1.7.7 Suppose that u ∈ C 1 (B R (0)) solves the Dirichlet problem

(
∆u = 0 in Ω
u|∂Ω = g.

Then
R2 − |y|2
Z
g(X)
u(y) = dσ(X) ∀y ∈ Ω.
ωn R ∂BR (0) |y − X|n

In addition, we can prove that a solution of the Dirichlet problem above with
continuous boundary data exists. This is the main result of this section.
32 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Theorem 1.7.8 Suppose that g ∈ C(∂BR (0)). Then the Dirichlet problem
(
∆u = 0 in Ω
u|∂Ω = g

has a unique solution u ∈ C(B R (0)), given by

 2 2 Z
 R − |y| g(X)
dσ(X) ∀y ∈ Ω
n
u(y) = ωn R ∂BR (0) |y − X|
∀y ∈ ∂BR (0).

g(y)

Proof. By differentiating under the integral sign, it is straightforward to check

that the function u defined above is harmonic in BR (0). Thus, it remains to
prove that for every X0 ∈ ∂BR (0)

lim u(y) = g(X0 ), (1.7.4)

y→X0

as y tends to X0 within BR (0). For y ∈ BR (0) we write

R2 − |y|2 g(X) − g(X0 )

Z
u(y) = dσ(X)
ωn R ∂BR (0) |y − X|n
R2 − |y|2
Z
dσ(X)
+ g(X0 ) n
.
ωn R ∂BR (0) |y − X|

We shall prove that

R2 − |y|2 g(X) − g(X0 )

Z
lim dσ(X) = 0 (1.7.5)
y→X0 ωn R ∂BR (0) |y − X|n

and that
R2 − |y|2
Z
dσ(X)
=1 ∀y ∈ BR (0). (1.7.6)
ωn R ∂BR (0) |y − X|n
These formulae will imply the required conclusion (1.7.4). The second for-
mula is a direct consequence of the representation formula of Corollary 1.7.7
(with the constant function 1 in place of g and u).
To prove (1.7.5), we proceed as follows. Since g is continuous in X0 , for
ε > 0 there exists δ > 0 such that

g(X) − g(X0 ) < ε ∀X ∈ B2δ (X0 ) ∩ ∂BR (0).

1.7. THE POISSON KERNEL FOR THE BALL 33

We split the integral over ∂BR (0) as the sum of the integral over ∂BR (0) ∩
B2δ (X0 ) and the integral over ∂BR (0) \ B2δ (X0 ), and estimate them sepa-
rately.
To estimate the first, observe that
g(X) − g(X0 )
Z Z
n
dσ(X) < ε |y − X|−n dσ(X)
∂BR (0)∩B2δ (X0 ) |y − X|
Z∂BR (0)∩B2δ (X0 )
≤ε |y − X|−n dσ(X)
∂BR (0)
ωn R
=ε ;
R2 − |y|2
we have used (1.7.6) in the equality above. Hence
R2 − |y|2 g(X) − g(X0 )
Z
dσ(X) < ε ∀y ∈ BR (0).
ωn R ∂BR (0)∩B2δ (X0 ) |y − X|n
Moreover
g(X) − g(X0 )
Z Z
n
dσ(X) ≤ 2kgk∞ |y − X|−n dσ(X)
∂BR (0)\B2δ (X0 ) |y − X| ∂BR (0)\B2δ (X0 )

We observe that |y − X| ≥ δ for all y in BR (0) ∩ Bδ (X0 ). Therefore the

integral on the right hand side is estimated from above by
δ −n σ ∂BR (0) ∩ B2δ (X0 )),
and
R2 − |y|2 g(X) − g(X0 )
Z
dσ(X)
ωn R ∂BR (0)∩B2δ (X0 ) |y − X|n
R2 − |y|2
2kgk∞ δ −n σ ∂BR (0)


≤ ∀y ∈ Bδ (X0 ).
ωn R
By combining the estimates above we see that given ε > 0, there exists δ > 0
such that for every y in BR (0) ∩ Bδ (X0 )
R2 − |y|2 g(X) − g(X0 )
Z
dσ(X) ≤ ε + C δ −n (R − |y|) .
 
ωn R |y − X| n
∂BR (0)∩B2δ (X0 )

Thus, for every ε > 0

R2 − |y|2 g(X) − g(X0 )
Z
lim dσ(X) ≤ ε,
y→X0 ωn R ∂BR (0) |y − X|n
which clearly implies (1.7.5).
The proof of the theorem is complete. 2
34 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

1.8 Perron’s method

In this section we characterise all domains in Rn , n ≥ 2, for which the
Dirichlet problem with continuous boundary data is solvable. The method
we employ is due to Perron. One of its merits lies in the fact that the theory
separates in a natural way the existence of a solution in Ω from the boundary
behaviour of the solution.
We need to enlarge the definition of subharmonic and superharmonic
functions to functions in C(Ω). Recall that so far we have given the definition
of C 2 (Ω) subharmonic functions.

Definition 1.8.1 Suppose that Ω is a domain in Rn . A function u ∈ C(Ω)

is subharmonic in Ω if for every ball B ⊂⊂ Ω and for every function h,
which is continuous on B, harmonic in B, and satisfies u ≤ h on ∂B, we have
that u ≤ h in B.

The definition of C(Ω) superharmonic function is similar (just revert all ≤

signs into ≥). Henceforth by subharmonic function in Ω we mean a C(Ω)
subharmonic function in the sense of the above definition.

Exercise 1.8.2 Prove that a function u ∈ C 2 (Ω) such that ∆u ≥ 0 in Ω

is subharmonic in the sense of the definition above.

Perron’s method hinges of the notion of harmonic lifting, which we now

define.

Definition 1.8.3 Suppose that Ω is a domain in Rn , that u is subharmonic

in Ω, and that B ⊂⊂ Ω. We denote by HB (u) the harmonic lifting of u
in B, defined as the function on Ω that agrees with u on Ω \ B, and with the
solution to the Dirichlet problem
(
∆h = 0 in B
h|∂B = u

on B. Clearly HB (u) is given by the Poisson integral in B of the restriction

of u to ∂B.

The following lemma summarizes some useful properties of subharmonic

functions.
1.8. PERRON’S METHOD 35

Lemma 1.8.4 Suppose that Ω is a domain in Rn . The following hold:

(i) if u is subharmonic in Ω, then it satisfies the strong maximum principle

in Ω;

(ii) if Ω is bounded, u, v are in C(Ω), u is subharmonic in Ω, v is super-

harmonic in Ω and u ≤ v on ∂Ω, then either u < v in Ω, or u = v;

(iii) if u is subharmonic in Ω and B ⊂⊂ Ω, then the harmonic lifting HB (u)

is subharmonic in Ω;

(iv) if u1 , . . . , uN are subharmonic functions in Ω, then

max u1 , . . . , uN

is subharmonic in Ω.

We leave the reader the task of stating and proving the corresponding prop-
erties for superharmonic functions.
Proof. We first prove (i). Suppose that y is a point in Ω such that

u(y) = sup u.
Ω

We need to prove that u is constant. Clearly the set Ω0 := {x ∈ Ω : u(x) =

u(y)} is closed, because u is continuous in Ω. We prove that Ω0 is open.
Suppose that x ∈ Ω0 and that BR (x) ⊂⊂ Ω. Denote by h the solution to
the Dirichlet problem on BR (x) with boundary values u|∂BR (x) . Since u is
subharmonic, u ≤ h on BR (x). Notice that

sup h = sup u ≤ u(x) ≤ h(x).

∂BR (x) ∂BR (x)

By the strong maximum principle, h is constant on B R (y) and the constant

must be h(y). Thus,
u|∂BR (x) = u(x).
By repeating the same reasoning for all r < R, we find that u = u(x) on
BR (x). Thus Ω0 is open. Since Ω is connnected, and Ω0 is nonempty, Ω0 = Ω,
i.e., u is constant on Ω, as required.
To prove (ii), we observe preliminarily that −v is a subharmonic function,
and that the sum of two subharmonic functions is subharmonic. Indeed, if
B ⊂⊂ Ω and h is continuous on B, harmonic in Ω, and satisfies −v|∂B ≤
h|∂B , then v|∂B ≥ −h|∂B . Since v is superharmonic and −h is harmonic, we
36 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

may conclude that v ≥ −h in B, equivalently that −v ≤ h. Hence −v is

subharmonic. Suppose that u1 and u2 are subharmonic functions in Ω and
fix B ⊂⊂ Ω. Denote by Uj the unique solution to the Dirichlet problem
(
∆w = 0 in B
w|∂B = (uj )|∂B .

Clearly uj ≤ Uj in B, because uj is subharmonic. Assume that h is continu-

ous on B, harmonic in B, and satisfies (u1 +u2 )|∂B ≤ h|∂B . Then U1 +U2 ≤ h
in B by the maximum principle. Therefore u1 + u2 ≤ h on B, as required.
As a consequence, u − v is subharmonic in Ω. Clearly u − v belongs
to C(Ω), and, by assumption, u − v ≤ 0 on ∂Ω. Set

M := sup (u − v).
Ω

We first prove that M ≤ 0. Indeed, if M were strictly positive, then, by

compactness, there would exists a point x0 ∈ Ω such that

M = (u − v)(x0 ).

Then, by (i) (i.e. by the strong maximum principle), u − v would be constant

(equal to M > 0) in Ω. This, together with the continuity of u − v on C(Ω),
would then contradict the fact that u − v ≤ 0 on ∂Ω.
Thus, M ≤ 0. Clearly, if M < 0, then u < v on Ω. To conclude the proof
of (ii), assume that M = 0. Again, either u < v in Ω, or the inequality u < v
fails at some point of Ω. In the latter case there exists x0 ∈ Ω such that

(u − v)(x0 ) = 0.

Then u = v by the strong maximum principle (i).

Next we prove (iii). Suppose that B 0 is a ball, which is relatively compact
in Ω. We need to show that for every function h, which is continuous on B 0 ,
harmonic in B 0 , and satisfies HB (u)|∂B 0 ≤ h|∂B 0 , we have HB (u) ≤ h in B 0 .
This is clear if B 0 ⊂ B or B 0 ⊂ (Ω \ B). It remains to consider the case where
B 0 ∩ (∂B) 6= ∅.
Since h ≥ HB (u) on ∂B 0 and HB (u) ≥ u on Ω, h ≥ u on ∂B 0 . Then
h ≥ u on B 0 , because u is subharmonic. Recall that HB (u) = u on Ω \ B,
whence h ≥ HB (u) on ∂(B ∩ B 0 ). Since both h and HB (u) are harmonic on
B ∩ B 0 , h ≥ HB (u) on (B ∩ B 0 ) by the maximum principle. This concludes
the proof of (iii).
1.8. PERRON’S METHOD 37

Finally we prove (iv). Suppose that B is a relatively compact ball in Ω,

and that h is continuous on B, harmonic in B, and satisfies max(u1 , . . . , uN )|∂B ≤
h|∂B . Then (uj )|∂B ≤ h|∂B for each j. By assumption uj is subharmonic, so
that uj ≤ h in B for every j, whence max(u1 , . . . , uN ) ≤ h on B. Therefore
max(u1 , . . . , uN ) is subharmonic, as required. 2

To proceed further, we need a compactness criterion for harmonic functions.

First we recall the classical compactness result of Ascoli and Arzelà.
Recall that a sequence {fj } of continuous functions on a set E ⊂ Rn is
uniformly bounded if there exists a constant C such that

sup |fj | ≤ C ∀j.

E

Similarly, {fj } is equicontinuous if for every ε > 0 there exists a constant δ

such that

x, y ∈ E, |x − y| < δ =⇒ fj (x) − fj (y) < ε ∀j.

Suppose that E is an interval of the real line, that fj is differentiable, and

that there exists a constant C such that

sup |fj0 | ≤ C ∀j.

E

Then {fj } is equicontinuous.

Exercise 1.8.5 Prove the assertion above.

Theorem 1.8.6 Suppose that {fj } is a sequence of uniformly bounded and

equicontinous functions on the compact set K ⊂ Rn . Then there exists a sub-
sequence {j` } such that {fj` } is uniformly convergent on K to a continuous
function f .

We refer the reader to [] for a proof of this result.

We shall apply Ascoli-Arzelà’s theorem to sequences of harmonic func-
tions on domains of Rn . The main point is that the equicontinuity of such a
sequence is a consequence of its uniform boundedness. A key step towards a
compactness criterion for harmonic functions is the following result, of inde-
pendent interest, which gives estimates for derivatives of harmonic functions
in terms of bounds of the functions themselves.
38 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Proposition 1.8.7 Suppose that u is harmonic in the domain Ω, and that K

is a compact subset of Ω. Then, for every multiindex α

α
 n |α| |α|
sup D u ≤ sup |u|.
K dist(K, ∂Ω) Ω

Proof. We prove the required estimate in the case where |α| = 1. Since ∂j u
is harmonic in Ω (because ∂j and ∆ commute), we may use the mean value
property for harmonic functions and write
Z
1
∂j u(x) =
∂j u dV
V BR (x) BR (x)
Z
1
=
u νj dσ.
V BR (x) ∂BR (x)

We have used the formula of integration by parts in the last equality. Here
R < dist(K, ∂Ω), so that BR (x) ⊂ Ω. Therefore
1

∂j u(x) ≤
σR ∂BR (x) sup |u|
V BR (x) Ω
n
= sup |u|
R Ω
The required estimate follows by taking the supremum of both sides with
respect to R < dist(K, ∂Ω). 2

Exercise 1.8.8 Prove that every bounded harmonic function on Rn is

constant.

Corollary 1.8.9 Every bounded sequence of harmonic functions on a do-

main Ω contains a subsequence converging uniformly on compact subdomains
of Ω to a harmonic function.

Proof. Suppose that {uj } is a bounded sequence of harmonic functions on Ω,

and that K is a compact subdomain of Ω. By Proposition 1.8.7,
n
sup |∇uj | ≤ sup |uj |
K dist(K, ∂Ω) Ω
≤ Cn,K .

Therefore {uj } is an equicontinuous sequence of functions on K. By Ascoli–

Arzelà’s theorem, there exists a subsequence {j` } such that {uj` } converges
1.8. PERRON’S METHOD 39

uniformly on K to a continuous function u. It remains to prove that u is

harmonic in the interior of K. It suffices to show that u satisfies the mean
value theorem therein. Suppose that x is in the interior of K. Then BR (x)
is contained in the interior of K for small R. Since uj` is harmonic in Ω,
Z
1
uj` (x) = uj dσ.
σ(∂BR (x)) ∂BR (x) `
The uniform convergence of {uj` } to u implies that
Z
1
u(x) = u dσ.
σ(∂BR (x)) ∂BR (x)
Hence u is a continuous function on the interior of K that satisfies the mean
value property. The required conclusion then follows from Theorem 1.3.9. 2

Definition 1.8.10 Suppose that Ω is a bounded domain and that ϕ is

a bounded function defined on ∂Ω. We say that a function v ∈ C(Ω) is a
subfunction relative to ϕ if v is subharmonic in Ω and v ≤ ϕ on ∂Ω. The
set of all subfunctions relative to ϕ is denote by Sϕ .

Theorem 1.8.11 Suppose that Ω is a bounded domain and that ϕ is a

bounded function defined on ∂Ω. The function u, defined by
u(x) := sup v(x) ∀x ∈ Ω,
v∈Sϕ

is harmonic in Ω.

Proof. Note that every subfunction relative to ϕ is bounded above by sup ϕ,

so that u is well defined and bounded above by sup ϕ. Note that inf ϕ is a
subfunction relative to ϕ. Hence u ≥ inf ϕ in Ω.
Fix y ∈ Ω. We shall show that u is harmonic in a neighbourhood of y.
By letting y vary in Ω we may then conclude that u is harmonic in Ω, as
required.
By the definition of u, there exists a sequence {v` } of subfunctions relative
to ϕ such that
u(y) = lim v` (y).
`→∞

It may very well happen that {v` } is unbounded from below. However, the
sequence {max(vn , inf ϕ)} is bounded from below, and still
u(y) = lim max(v` (y), inf ϕ),
`→∞
40 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

so that we may assume that {v` } is bounded. Now, we choose R such that
BR (y) ⊂⊂ Ω. We consider the harmonic lifting HBR (y) (v` ) of v` . Recall that
HBR (y) (v` ) is subharmonic in Ω, and harmonic in BR (y).
Clearly HBR (y) (v` ) is a subfunction relative to ϕ and
u(y) = lim HBR (y) (v` )(y).
`→∞

Observe that {HBR (y) (v` )} is a sequence of uniformly bounded harmonic func-
tions on BR (y). By Corollary 1.8.9, there exists a subsequence {`j } such that
{HBR (y) (v`j )} is uniformly convergent on BR/2 (y) to a function v, which is
harmonic in BR/2 (y).
Obviously v ≤ u in BR/2 (y). We claim that v = u in BR/2 (y).
We argue by contradiction. Suppose that there exists z ∈ BR/2 (y) such
that
v(z) < u(z).
By the definition of the function u, there exists a subfunction ve relative to ϕ
in Ω such that
v(z) < ve(z) ≤ u(z). (1.8.1)
Consider the sequence


wj := max ve, HBR (y) (v`j ) .
Clearly wj is a subfunction relative to ϕ. Its harmonic lifting HBR (y) (wj ) is
harmonic in BR (y),
u(y) = lim HBR (y) (wj )(y),
j→∞

and, possibly passing to a subsequence, {HBR (y) (wj )} is uniformly convergent

on BR/2 (y) to a function w, which is harmonic in BR/2 (y). Clearly w satisfies

v ≤ w ≤ u on BR/2 (y), v(z) < ve(z) ≤ w(z) and v(y) = w(y) = u(y).
Observe that v − w is harmonic in BR/2 (y) and that (v − w)|∂BR/2 (y) ≤ 0.
Since (v − w)(y) = 0, v − w = 0 in BR (y) by the maximum principle. This
contradicts (1.8.1). Therefore v = u on BR/2 (y). This proves the claim.
The proof of the theorem is complete. 2

Definition 1.8.12 The function

u(x) := sup v(x) ∀x ∈ Ω,
v∈Sϕ

is called the Perron solution with boundary datum ϕ.

1.8. PERRON’S METHOD 41

Corollary 1.8.13 Suppose that Ω is a bounded domain and that ϕ is a con-

tinuous function on ∂Ω. If the classical Dirichlet problem
(
∆w = 0 in Ω
w|∂Ω = ϕ

has a solution w ∈ C 2 (Ω)∩C(Ω), then w agrees on Ω with the Perron solution

with boundary datum ϕ.

Proof. On the one hand, observe that w is in Sϕ . Hence u = supv∈Sϕ v ≥ w

in Ω.
On the other hand, if v ∈ Sϕ , then v ≤ w on ∂Ω, whence v ≤ w on Ω,
because v is subharmonic. Then

u = sup v ≤ w in Ω.
v∈Sϕ

Thus, w = u in Ω, as required. 2

Now, we examine the boundary behaviour of the Perron solution to the

Dirichlet boundary value problem. We need the following definition.

Definition 1.8.14 Suppose that Ω is a domain and that X ∈ ∂Ω. A

function w, which is continuous on Ω is a barrier function at X relative to
Ω if the following hold

(i) w is superharmonic in Ω;

(ii) w > 0 in Ω \ {X} and w(X) = 0.

Suppose that Ω is the upper half space {x ∈ Rn : xn > 0} in Rn , n ≥ 3.

Choose y ∈ Ω and R > 0. Set ye = (y 0 , −R). Then the function

w(x) := R2−n − |x − ye|2−n

is a barrier at the point (y 0 , 0).

Exercise 1.8.15 Prove that the function above is a barrier at (y 0 , 0). Do

the same for the function
|x − ye|
w(x) := log
R
in the case where n = 2.
42 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Exercise 1.8.16 Suppose that Ω is a domain that satisfies the following

uniform exterior sphere condition: there exists R > 0 such that for
every X ∈ ∂Ω there exists a ball BR (y) such that BR (y) ∩ Ω = {X}. Prove
that for each X ∈ ∂Ω there exists a barrier at X relative to Ω.

Observe that the concept of barrier function is a local property of Ω. Indeed,

suppose that w is a local barrier at X relative to Ω, i.e., there exists R > 0
such that w is continuous on Ω ∩ BR (X) and

(i) w is superharmonic in Ω ∩ BR (X);

(ii) w > 0 in Ω ∩ BR (X) \ {X} and w(X) = 0.

Choose a ball B ⊂⊂ BR (X) with centre X, and such that

m := inf w > 0.
BR (X)\B

Then the function

(

min m, w(x) ∀x ∈ Ω ∩ B
w(x)
e =
m ∀x ∈ Ω \ B,

is a barrier at X relative to Ω.
Indeed, w
e is clearly continuous on Ω, and is superharmonic in Ω by
e > 0 on Ω \ {X} and w(X) = 0.
Lemma 1.8.4 (iv). Furthermore, w

Definition 1.8.17 A point X ∈ ∂Ω is regular if there exists a barrier at

that point.

Lemma 1.8.18 Denote by u the Perron solution in Ω relative to the da-

tum ϕ. If X ∈ ∂Ω is a regular point and ϕ is a function on ∂Ω that is
continuous at X, then
lim u(x) = ϕ(X);
x→X

here x tends to the boundary point X within Ω.

Proof. Set M := sup |ϕ| and choose ε > 0. Since X is a regular point, there
exists a barrier w at X relative to Ω. Since ϕ is continuous at X, there exists
δ > 0 such that
|ϕ(Y ) − ϕ(X)| < ε
1.8. PERRON’S METHOD 43

for all Y ∈ ∂Ω such that |X − Y | < δ. Furthermore, there exists a constant

k such that

k w(x) ≥ 2M ∀x such that |x − X| ≥ δ. (1.8.2)

It is straightforward to check that the funtions ϕ(X) + ε + kw and ϕ(X) −

ε − kw are superfunction and subfunction relative to X, respectively. By
the definition of u and the fact that every superfunction dominates every
subfunction, we have that

ϕ(X) − ε − kw(x) ≤ u(x) ≤ ϕ(X) − ε − kw(x) ∀x ∈ Ω.

This is clearly equivalent to

|u(x) − ϕ(X)| ≤ ε + kw(x) ∀x ∈ Ω.

Since w is a barrier at X, w(x) → 0 as x tends to X, so that for every ε > 0

lim |u(x) − ϕ(X)| ≤ ε,

x→X

which implies the required conclusion. 2

Corollary 1.8.19 Suppose that Ω is a bounded domain. Then the classical

Dirichlet problem is uniquely solvable in Ω if and only if every point of ∂Ω
is a regular point.

Proof. If every point of ∂Ω is regular, then, by Lemma 1.8.18, the classical

Dirichlet problem is solvable. Uniqueness then follows from the maximum
principle.
Conversely, if the classical Dirichlet problem is solvable, and X is a point
of ∂Ω, then the solution to the problem
(
∆u = 0 in Ω
u|∂Ω = g,

where g(Y ) := |X − Y | is a barrier function at X, hence X is a regular point.

Consequently, every point of ∂Ω is regular, as required. 2

Now, the problem becomes to give criteria which imply that every point of
the boundary of a given domain are regular. A simple criterion is discussed
in the following exercise.
44 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Exercise 1.8.20 Suppose that Ω is a bounded C 2 domain. Prove that the

classical Dirichlet problem is uniquely solvable in Ω.

We conclude this section with a criterion which ensures the solvability of

the classical Dirichlet problem in Lipschitz domains. We need the following
definition.

Definition 1.8.21 A bounded domain Ω satisfies the uniform exterior

cone condition if there exist constants h and α such that for each point
X ∈ ∂Ω there exists a cone Γ(X, α, h) with vertex X, aperture α and height
h such that Γ(X, α, h) ∩ Ω = {X}.

Exercise 1.8.22 Prove that every bounded Lipschitz domain satisfies the
exterior cone condition.

Proposition 1.8.23 Suppose that Ω is a bounded domain, which satisfies

the uniform exterior cone condition. Then the classical Dirichlet problem is
uniquely solvable in Ω.

Proof. Fix a point X ∈ ∂Ω. Since Ω satisfies the uniform exterior cone
condition, there exist constants α and h such that the cone Γ(X, α, h) is
such that Γ(X, α, h) ∩
Ω = {X}. We normalise our setting by assuming that
Γ(X, α, h) ∩ X + Sn−1 6= ∅. To construct a barrier at X we argue as follows.
Consider all functions v of the form
w(x) = rλ ϕ0 (ω),
where x = r ω (polar coordinates centred at X), λ > 0, and ϕ0 is the
eigenfunction associated to the lowest eigenvalue µ of the eigenvalue problem
∆Sn−1 U = −µ U
on Sn−1 \ Γ(0, α, h) with Dirichlet boundary conditions. A well known result
in spectral theory, which we do not prove in these notes, asserts that µ > 0
and that ϕ0 is a smooth function, which is strictly positive in Sn−1 \Γ(0, α, h).
By using the formula for the Laplacian ∆ in polar coordinates, we find that
∆w(x) = rλ−2 ϕ0 α2 + (n − 2) α − µ .
 
p 
It is straightforward to check that if α = (1/2) (n − 2)2 + 4µ − (n − 2) ,
then w is harmonic in Ω. Furthermore, w(X) = 0, and w > 0 in Ω \ {X}.
Thus, w is a barrier at X. Hence X is a regular point relative to Ω.
Since this holds for every point in ∂Ω, the classical Dirichlet problem
is solvable by Lemma 1.8.18. The uniqueness follows from the maximum
principle. 2
1.9. THE LEBESGUE SPINE 45

1.9 The Lebesgue spine

The aim of this section is to provide a counterexample to the solvability of the
classical Dirichlet problem on bounded domains. Specifically, we shall prove
that there exists a bounded domain in R3 , rotationally invariant with respect
to the z axis, for which the classical Dirichlet problem has no solutions. This
domain has a sharp inward cusp, which is responsible for the nonsolvability
of the Dirichlet problem. The counterexample is due to Lebesgue and dates
back to 1912.
We shall denote by (x, y, z) the coordinates in R3 , and set r2 = x2 + y 2 .
We consider the domain Ω, defined by
Ω := {(x, y, z) ∈ R3 : r2 + z 2 < 1, r > e−1/(2z) if z > 0}.
We call the inward cusp at (0, 0, 0) a Lebesgue spine. We shall prove that
(0, 0, 0) is not a regular point relative to Ω.
Indeed, consider the positive Radon measure µ, defined as the continuous
linear functional on Cc (R3 ), defined by
Z 1
µ(ϕ) = ϕ(0, 0, τ ) τ dτ
0

(see Section 5.1 and, in particular, Theorem 5.1.2, for more on the connection
between measures and continuous linear functionals on Cc (Rn )). Note that
the support of µ is contained in the segment
Σ := {(0, 0, z) ∈ R3 : z ∈ [0, 1]}.
We consider the potential associated to the measure µ, i.e., the function u
given by
u = N ∗ µ.
Since µ has compact support, the convolution on the right hand side of the
formula above makes perfect sense (see Definition 5.4.9 below). We shall
compute u explicitly. For every ϕ ∈ Cc∞ (R3 )
hϕ, N ∗ µi = ϕ ∗ Ň , µ
Z 1
= ϕ ∗ Ň (0, 0, τ ) τ dτ
0
Z 1 Z
= τ dτ ϕ(x, y, z) N (x, y, z − τ ) dx dy dz
0 R3
Z Z 1
1 τ
= ϕ(x, y, z) dx dy dz p dτ.
4π R3 0 x + y + (z − τ )2
2 2
46 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM

Therefore
Z 1
1 τ
N ∗ µ(x, y, z) = p dτ.
4π 0 x2 + y 2 + (z − τ )2
The right hand side can be computed explicitly. We find that, up to a
constant factor,
p √
N ∗ µ(x, y, z) = r2 + (z − 1)2 − r2 + z 2
√ p 
+ z log r2 + z 2 + z r2 + (z − 1)2 + 1 − z
− 2z log r ∀(x, y, z) ∈ Ω such that z > 0.

Exercise 1.9.1 Compute N ∗ µ(x, y, z) for z < 0.

It is straighforward to check that the restriction of u to ∂Ω is a continuous

function. Clearly u is harmonic in Ω. However, u does not admit a limit as
(x, y, z) tends to (0, 0, 0) within Ω.
Indeed, suppose that α ∈ (0, 1), and consider the surfaces

rα (z) := e−α/(2z) , z > 0,

which, for z ∈ (0, 1) is contained in Ω. It is straighforward to check that

lim u(rα (z) cos θ, rα (z) sin θ, z) = α,

z→0+

as required. We have used cylindrical coordinates in the formula above.

We are now ready to prove that (0, 0, 0) is not a regular point relative
to Ω, i.e., the Dirichlet problem with datum u|∂Ω does not admit a classical
solution.
We argue by contradiction. If v were a classical solution to the Dirichlet
problem with datum u|∂Ω , then its restriction to the subdomain Ωε , defined by

Ωε := {(x, y, z) ∈ Ω : r2 > ε2 },

solves the Dirichlet problem

(
∆w = 0 in Ωε
w|∂Ωε = u|∂Ωε .

Since every point of ∂Ωε is regular relative to Ωε and u is clearly a solution of

the problem above, u = v on Ωε for every ε > 0. Then v = u on Ω \ (0, 0, 0),
a fact which contradicts the continuity of v at the point (0, 0, 0).
1.9. THE LEBESGUE SPINE 47

It may be worth noticing that there are domains with inward cusps for which
the classical Dirichlet problem is solvable. For instance, it may be shown that
this is the case for the domain

Ω := {(x, y, z) ∈ R3 : r2 + z 2 < 1, r > z 2k if z > 0}.

for every positive integer k. The proof of this fact uses the theory of capacity,
which we do not tackle in this notes. The interested reader is referred to [K].
Similar results hold in Rn for n ≥ 3. However, the case of bounded domains
in the plane is quite different. Indeed, it is straightforward to show that if
Ω is a planar domain and z0 is a boundary point for which the following is
true: there exists a straight line segment [z0 , z1 ] entirely contained in R2 \ Ω,
then z0 is a regular point relative to Ω.

Exercise 1.9.2 Prove the assertion above. Hint: consider a branch of

 − z0 ) adapted to the situation and study the function − Re 1/ log(z −
log(z
z0 ) .

This criterion implies that the two dimensional analogue of the Lebesgue
spine is a regular point relative to Ω.
48 CHAPTER 1. THE CLASSICAL DIRICHLET PROBLEM
 Part II

The Dirichlet problem

via integral equations

49
 51

In this part we discuss a method for solving the Dirichlet problem on a

bounded domain Ω that hinges on certain integral equations on the bound-
ary ∂Ω. The idea behind the method is very simple. We recall that if
u ∈ C(Ω) ∩ C 2 (Ω) is the classical solution of the Dirichlet problem
(
∆u = 0 in Ω
u|∂Ω = g,

then Z
u(x) = g(Y ) ∂ν(Y ) G(x, Y ) dσ(Y ) ∀x ∈ Ω
∂Ω

(see Corollary 1.5.8), where G is the Green’s function of Ω, and ν(Y ) denotes
the outward unit normal to ∂Ω at the point Y . Furthermore,

lim u(x) = g(X) ∀X ∈ ∂Ω.

x→X
x∈Ω

The idea that leads to the method of integral equations is to replace the
Green’s function of Ω with the Newtonian potential N in the integral above,
and to consider the operator that associates to each reasonable function g on
∂Ω the function Dg, defined by
Z
Dg(x) = g(Y ) ∂ν(Y ) N (x − Y ) dσ(Y ) ∀x ∈ Ω.
∂Ω

Clearly Dg will not be the solution of the Dirichlet problem with datum g
for the obvious reason that N (x − Y ) does not agree with G(x, Y ). However,
the operator D is related to the solution of the Dirichlet problem, as we shall
explain below. The operator g 7→ Dg is called the double layer potential
and the function Dg is called the double layer potential with density g.
Its properties will play a key role in the theory.
We shall prove that for ϕ ∈ C(∂Ω), the integral
Z
ϕ(Y ) ∂ν(Y ) N (X − Y ) dσ(Y )
∂Ω

makes sense also when X ∈ ∂Ω. Thus, we may define the operator K on
C(∂Ω) by
Z
Kϕ(X) := ϕ(Y ) ∂ν(Y ) N (X − Y ) dσ(Y ) ∀X ∈ ∂Ω.
∂Ω
52

Furthermore, the following formula holds for every ϕ ∈ C(∂Ω):

1
lim Dϕ(x) = ϕ(X) + Kϕ(X) ∀X ∈ ∂Ω.
x→X 2
x∈Ω

Thus, the boundary value of the double layer potential Dϕ with density ϕ is
the datum g provided that ϕ satisfies the following integral equation
1
g(X) = ϕ(X) + Kϕ(X) ∀X ∈ ∂Ω.
2
It is a nontrivial fact, which will be proved below, that for every contin-
uous function g on ∂Ω the integral equation above has a unique solution
ϕ ∈ C(∂Ω). In the language of spectral theory, this statement may be refor-
mulated by saying that −1/2 is in the resolvent set of the operator K.
The operator K is a compact operator. This is the reason why in the
following we shall study in some detail the class of compact operators and
their spectra.

One of the major drawbacks of the method of integral equation is that

it is hard to make it work unless the domain Ω has a reasonable boundary,
say C 2 . The method can be pushed to cover Lipschitz domains, and Lp
boundary data, but this requires a lot of very refined analysis, which is well
beyond the scope of these notes. The interested reader is referred to the
paper of B. Dahlberg and C. Kenig [DK] for a detailed exposition of this
comparatively recent piece of research.
Chapter 2

Linear operators on Banach

spaces

Throughout this chapter B will denote a Banach space over C.

2.1 Basic definitions

Definition 2.1.1 A linear operator on B is a pair (X, A), where X is a
(possibly not closed) linear subspace of B and A : X → B is a linear operator:
X is called the domain of A and will be denoted by Dom(A).

Note that we do not assume that Dom(A) is dense in B, unless explicitly

stated.

Definition 2.1.2 Given two linear operators A and B on B, we say that

B is an extension of A, and we write B ⊃ A, if Dom(B) ⊃ Dom(A) and
the restriction of B to Dom(A) agrees with A.

Consider the operators A and B on L2 (R) defined by

Dom(A) = Cc∞ (R) Af = i f 0 ∀f ∈ Dom(A)

Dom(B) = Cc1 (R) Bf = i f 0 ∀f ∈ Dom(B).

Clearly B ⊃ A.

We consider two more examples of linear operators, which are one dimen-
sional models of the Dirichlet and Neumann Laplacian. Suppose that

53
54 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

−∞ < a < b < ∞. Let HD and HN be the operators on L2 ((a, b)) defined by

Dom(HD ) = {f ∈ C 2 ([a, b]) : f (a) = 0 = f (b)} HD f = −f 00

Dom(HN ) = {f ∈ C 2 ([a, b]) : f 0 (a) = 0 = f 0 (b)} HN f = −f 00 .

It is not hard to check that the domains of HD and HN are dense in L2 ((a, b)).

Exercise 2.1.3 Suppose that A is a densely defined linear operator in a

Hilbert space H such that (Af, f ) = 0 for all f in Dom(A). Prove that
A = 0. Note that this fails on real vector spaces (take, e.g., a rotation of
π/2 on R2 ).

2.2 Bounded operators

Definition 2.2.1 A linear operator on B is bounded if Dom(A) = B and

kAf k
sup < ∞.
f 6=0 kf k

The supremum above is called the operator norm of A and it is denoted

by |||A|||.

Exercise 2.2.2 Suppose that B is a finite dimensional (Banach) space.

Prove that every linear operator with domain B is bounded. (Hint: It suffices
to “control” the operator on a basis of B.)

Exercise 2.2.3 Prove that for a linear operator A on B the following are
equivalent:

(i) A is a bounded operator on B;

(ii) A is continuous at every point of B;

(iii) A is continuous at 0.

Exercise 2.2.4 Suppose that A is a linear operator on B with dense do-

main. Assume that
kAf k
sup < ∞.
f ∈Dom(A)\{0} kf k

Prove that there exists a unique bounded linear operator A

e that extends A.
2.2. BOUNDED OPERATORS 55

An interesting application of the exercise above is the following special case

of Poincaré’s inequality, whose general form will be considered in Section ??
below.

Definition 2.2.5 Suppose that I is an open interval of the real line. We

consider the following norm on Cc∞ (I):

kϕkH01 (I) := kϕ0 kL2 (I) .

Denote by H01 (I) the completion of Cc∞ (I) with respect to the norm above.

The following proposition illustrates the role of Exercise 2.2.4 in proving the
boundedness of certain operators.

Proposition 2.2.6 Suppose that I is a bounded interval of the real line. The
identity map j : ϕ 7→ ϕ, initially defined on Cc∞ (I), extends to a bounded
linear operator from H01 (I) to L2 (I).

Proof. By Exercise 2.2.4 it suffices to show that

kϕkL2 (I) ≤ |I| kϕ0 kL2 (I) ∀ϕ ∈ Cc∞ (I).

This is elementary. Indeed, denote by p the left endpoint of I and write

Z x
ϕ(x) = ϕ0 (s) ds
Z px
≤ ϕ0 (s) ds
p
≤ kϕ0 k1
(by Hölder’s inequality) ≤ |I|1/2 kϕ0 k2 ∀x ∈ I,

and the required estimate follows by integrating the squares of both sides
on I. 2

It is worth observing that the space H01 (I) may be realised as a (nonclosed)
subspace of L2 (I).
Indeed, to prove that the map j in Proposition 2.2.6 is an immersion, we
have to show that if f ∈ H01 (I) satisfies j(f ) = 0, then f = 0. Note that
this is trivial if f ∈ Cc∞ (I), because j(f ) = f as functions. However, we do
not know a priori how the map j acts on elements of H01 (I), so we need to
prove that j(f ) = 0 implies f = 0. Since H01 (I) is the completion of Cc∞ (I),
56 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

given f ∈ H01 (I), there exists a sequence {fn } of functions in Cc∞ (I) such
that kfn − f kH01 (I) → 0 as n tends to ∞. Note that kfn k2 → 0, because j is
continuous by Proposition 2.2.6, and fn = j(fn ) → j(f ) = 0.
Furthermore, {fn0 } is a Cauchy sequence in L2 (I) by the definition of the
norm in H01 (I). Since L2 (I) is complete, there exists a function g ∈ L2 (I) such
that kfn0 − gk2 → 0 as n tends to ∞. By possibly passing to a subsequence,
we deduce that {fn } is convergent to 0 a.e. By the fundamental theorem of
calculus Z x
fn (x) = fn0 (s) ds,
0
so that, by taking the limit of both sides,
Z x
0= g(s) ds
0

for all x ∈ I. Therefore g = 0 a.e. Thus,

kfn kH01 (I) = kfn0 k2 = kfn0 − gk2 → 0

as n tends to ∞. By the uniqueness of the limit in H01 (I), we may conclude

that f = 0, as required.
Thus, elements of H01 (I) may be interpreted as equivalence classes of
square integrable functions.

Exercise 2.2.7 Prove that the operators A and B defined just below Def-
inition 2.1.2 do not admit bounded extensions to L2 (R).

It is straightforward to check that the set of bounded operators on B is a

vector space, which we denote by L(B), and that the operator norm is a
norm on L(B).

Proposition 2.2.8 The space L(B) is a Banach space with respect to the
operator norm.

Proof. Suppose that {An } is a Cauchy sequence of operators in L(B). Then

for every ε > 0 there exists an integer ν such that

|||Am − An ||| ≤ ε ∀m, n ≥ ν.

Consequently,

kAm f − An f k ≤ ε kf k ∀m, n ≥ ν ∀f ∈ B. (2.2.1)

2.3. THE SPECTRUM OF A LINEAR OPERATOR 57

Thus, for every f in B, {An f } is a Cauchy sequence in B. Since B is complete,

there exists an element in B that we denote Af such that

lim An f = Af, (2.2.2)

n→∞

where the convergence is in the norm of B. It is straightforward to check

that the operator A thus defined is linear. By letting m tend to ∞ in (2.2.1),
we obtain
kAf − An f k ≤ ε kf k ∀n ≥ ν ∀f ∈ B.
This implies that A − An is a bounded operator on B, and that

|||A − An ||| ≤ ε ∀n ≥ ν.

Thus, A is bounded (because it is the sum of the bounded operators An and

A − An ), and limn→∞ |||An − A||| = 0, as required. 2

2.3 The spectrum of a linear operator

Definition 2.3.1 Suppose that A is a linear operator on B. We say that
the complex number ζ is in the resolvent set of A if ζI − A maps Dom(A)
in a one-to-one fashion onto B and the resolvent operator

R(ζ; A) := (ζI − A)−1

is a bounded operator on B. The resolvent set of A will be denoted by %(A).

The spectrum σ(A) of A is defined to be C \ %(A).

If B is finite dimensional, then the spectrum of a linear operator A is just

the set of its eigenvalues.
One of the reasons for which spectral theory on infinite dimensional Ba-
nach spaces is far more difficult than in the finite dimensional case is that the
following result, whose proof we leave as an exercise, is false in the infinite
dimensional case.

Proposition 2.3.2 Suppose that B is finite dimensional, and that A is a

linear operator defined on B. The following hold:

(i) A is bounded;

(ii) A is invertible if and only if A is injective if and only if A is surjective;

58 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

(iii) Ran(A) is dense if and only if Ran(A) = B;

(iv) if A is invertible, then A−1 is bounded.

Proof. Exercise. 2

Consider the operator A defined just below Definition 2.1.2. There exists a
(purely algebraic) extension of A to a linear operator on B 1 . Note that this
extension cannot be bounded for we have proved in Exercise 2.2.7 that A
does not admit a bounded extension to all of B.
This extension provides an example of an unbounded linear operator
which is defined everywhere.

Exercise 2.3.3 Find counterexamples to statements (ii)-(iv) in the case

of operators on infinite dimensional Banach spaces, by looking at the opera-
tors L and R, defined on `2 by

L(x1 , x2 , x3 , . . .) = (x2 , x3 , . . .) R(x1 , x2 , x3 , . . .) = (0, x1 , x2 , x3 , . . .),

and C, defined on `1 by
 x x 
2 3
C(x1 , x2 , x3 , . . .) = x1 , , , . . . .
2 3

Definition 2.3.4 Suppose that A is a closed operator on B. We say that

a point ζ in σ(A) is in the point spectrum of A if ζ is an eigenvalue of A.
We say that a point ζ in σ(A) is in the discrete spectrum of A if it is an
eigenvalue and it is an isolated point in σ(A).

The main reason for introducing the residual spectrum of an operator is that
self adjoint operators have empty residual spectrum.

2.4 The adjoint of a bounded operator

In this section we assume that A is a bounded linear operator on the Hilbert
space H.
In the case where H is a finite dimensional Hilbert space with inner
product (·, ·), once we fix a orthonormal basis {v1 , . . . , vn } of H, each linear
1
See, for instance, p. 38 of A.E. Taylor, D.C. Lay, Introduction to functional analysis,
John Wiley & Sons, 1980.
2.4. THE ADJOINT OF A BOUNDED OPERATOR 59

operator A may be represented by a unique matrix MA = {mj,k }, where

mj,k := (Avj , vk ), and



A(x1 v1 + . . . + xn vn ), y1 v1 + . . . + yn vn = MA x, y ,
where x is the column vector (x1 , . . . , xn )t , and similarly for y, and the inner
product on the right hand side is just the Hermitian inner product on Cn ,
defined by
Xn
(X, Y) = Xj Y j .
j=1
t
Denote by MA∗ the matrix M A (conjugate transpose of MA ). Then
MA x, y = x, MA∗ y ∀x, y ∈ Cn .

Denote by A∗ the operator on H associated to the matrix MA∗ . Then, clearly

A(x1 v1 +. . .+xn vn ), y1 v1 +. . .+yn vn = x1 v1 +. . .+xn vn , A∗ (y1 v1 +. . .+yn vn ) .

The operator A∗ is called the adjoint of A.

This notion may be easily generalised to all bounded operators on infinite
dimensional Hilbert spaces. Suppose that A is a bounded operator on the
Hilbert space H. For each g in H, the functional
f 7→ (Af, g)
is bounded on H, because, by Schwarz’s inequality,
(Af, g) ≤ kAf k kgk
(2.4.1)
(since A is bounded) ≤ |||A||| kf k kgk.
Therefore, by the Riesz representation theorem, there exists a unique element
γg ∈ H such that
(Af, g) = (f, γg )
It is straightforward to check that the assignment g 7→ γg is linear, whence
it defines a linear operator, which we call A∗ . Thus,
(Af, g) = (f, A∗ g) ∀f, g ∈ H.
Observe that A∗ is bounded on H, for
kA∗ gk = sup (f, A∗ g)
kf k≤1

= sup (Af, g)
kf k≤1

(by (2.4.1)) ≤ sup |||A||| kf k kgk

kf k≤1

= |||A||| kgk,
60 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

and that |||A∗ ||| ≤ |||A|||. By reversing the role of A and A∗ , we find that
|||A||| ≤ |||A∗ |||, whence |||A||| = |||A∗ |||.

Definition 2.4.1 Suppose that A is a bounded operator on the Hilbert

space H. The operator A∗ defined above is called the (Hilbert space)
adjoint of A.

Before looking at examples of adjoints, we define an important class of op-

erators, acting on L2 (Rn ).

Definition 2.4.2 An operator A, acting on L2 (Rn ), is a Hilbert–Schmidt

operator if it is of the following form
Z
Af (x) := K(x, y) f (y) dy ∀f ∈ L2 (Rn ),
Rn

where K ∈ L2 (Rn × Rn ). The function K is called the integral kernel (or

simply the kernel) of A.

Exercise 2.4.3 Prove that a Hilbert–Schmidt operator A with kernel K

is bounded on L2 (Rn ) with norm ≤ kKkL2 (Rn ×Rn ) , and that the adjoint op-
erator A∗ is the Hilbert–Schmidt operator with kernel (x, y) 7→ K(y, x).

Exercise 2.4.4 Consider the convolution operator T on L2 (R), defined by

Tf = f ∗ ϕ ∀f ∈ L2 (R),

where ϕ is in L1 (R). Prove that T is bounded on L2 (R) and find T ∗ .

Exercise 2.4.5 Consider the left shift operator L on `2 , defined by

L(x1 , x2 , . . .) = (x2 , x3 , . . .)

for every sequence {xn } in `2 . Compute L∗ . Do the same for the right shift
operator R, defined by

R(x1 , x2 , . . .) = (0, x1 , x2 , . . .).

The following proposition summarises the main properties of the ∗ operator.

Proposition 2.4.6 The following hold:

2.4. THE ADJOINT OF A BOUNDED OPERATOR 61

(i) ∗ is a conjugate linear isometric isomorphism of L(H);

(ii) (A∗ )∗ = A, i.e., ∗ is involutive;
(iii) (ST )∗ = T ∗ S ∗ for all S and T in L(H);
(iv) if A and A−1 are in L(H), then A∗ is invertible in L(H) and (A∗ )−1 =
(A−1 )∗ ;
(v) |||A∗ A||| = |||A|||2 for every A in L(H).

Proof. The proof of (i)-(iv) is left to the reader.

To prove (v) observe that, on the one hand,
|||A∗ A||| ≤ |||A∗ ||| |||A||| = |||A|||2 ,
where we have used the formula |||A∗ ||| = |||A|||. On the other hand
|||A|||2 = sup kAf k 2
kf k=1

= sup |(Af, Af )|
kf k=1

= sup |(f, A∗ Af )|
kf k=1

≤ |||A∗ A|||,
and the required equality follows. 2
The properties of the involution ∗ contained in the proposition above and
the fact that L(H) is complete are usually summarised by saying that L(H)
is a C ∗ -algebra (with involution ∗).

Exercise 2.4.7 Prove statements (i)-(iv) in Proposition 2.4.6.

We now introduce a very important class of operators in Hilbert spaces.

Definition 2.4.8 A bounded operator A on H is self adjoint if A∗ = A.

Bounded self adjoint operators on Hilbert spaces generalise Hermitian ma-

trices on Cn , and play a distinguished role in Analysis. In particular, every
bounded self adjoint operator may be “diagonalised”. In other words, for
bounded self adjoint operator a generalisation of the Spectral Theorem for
Hermitian matrices holds. This result will be proved in the next section for
a noteworthy subclass of bounded operators, namely the compact operators.
One of the step that will be used in its proof is the following proposition,
which is of independent interest.
62 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

Proposition 2.4.9 Suppose that A is a bounded self adjoint operator on the

Hilbert space H. Then
|||A||| = sup (Af, f ) .
kf k=1

Proof. Denote by ηA the right hand side of the equality above. The inequality
ηA ≤ |||A||| is trivial and its proof is left to the reader.
It remains to prove that |||A||| ≤ ηA . We use the following polarisation
identity: for every f , g in H

4 (Af, g) = (A(f + g), f + g) − (A(f − g), f − g)

(2.4.2)
+ i(A(f + ig), f + ig) − i(A(f − ig), f − ig),

which is easily verified. Observe that for each ϕ in H the number (Aϕ, ϕ) is
real, for
(Aϕ, ϕ) = (ϕ, A∗ ϕ) = (ϕ, Aϕ) = (Aϕ, ϕ).
Here we have used the assumption that A is self adjoint. From (2.4.2) we
then deduce that

4 Re(Af, g) = (A(f + g), f + g) − (A(f − g), f − g),

whence
4 Re(Af, g) ≤ ηA kf + gk 2 + kf − gk 2
 
(2.4.3)
= 2ηA kf k 2 + kgk 2 .
 

Now, we may choose a real number θ so that

(Af, g) = Re eiθ (Af, g) = Re(Af, e−iθ g) .

 

We now use (2.4.3) with e−iθ g in place of g, and obtain

ηA  2
Re(Af, g) ≤ kf k 2 + ke−iθ gk . (2.4.4)
2
By taking the supremum over all f and g with kf k = kgk = 1, we see that
|||A||| ≤ ηA , as required to conclude the proof of the proposition. 2

Next, we illustrate a connection between bounded bilinear forms on H × H

and bounded operators on H. Unbounded operators and unbounded bilinear
forms may also be considered.

Definition 2.4.10 A bounded bilinear form B on H is a complex val-

ued map on H × H so that
2.4. THE ADJOINT OF A BOUNDED OPERATOR 63

(i) f 7→ B(f, g) is a linear functional on H for all g in H;

(ii) g 7→ B(f, g) is a conjugate linear functional on H for all f in H;
(iii) there exists a constant M such that

B(f, g) ≤ M kf k kgk ∀f, g ∈ H.

The norm of B is the infimum of all M for which the above inequality holds.

Suppose that A is a bounded operator on H. We define the bilinear form B

associated to A as follows

B(f, g) := (Af, g) ∀f, g ∈ H.

It is straighforward to check that B is a bounded bilinear form, with norm |||A|||.

Conversely, suppose that B is a bounded bilinear form on H and that
f ∈ H. Then, by (iii) above, the functional λf : g 7→ B(f, g) is bounded on
H. By the Riesz representation theorem there exists a unique element of H
that represents λf . We call it Af . Then B(f, g) = (g, Af ), whence

B(f, g) = (Af, g) ∀f, g ∈ H.

It is straightforward to check that A is a linear operator on H. Finally,

kAf k = sup (Af, g)
kgk=1

= sup B(f, g)
kgk=1

≤ M kf k,
where we have used the boundedness of the form B in the last inequality.
Therefore the operator A is bounded: A will be called the bounded linear
operator associated to the form B.
Thus, bounded bilinear forms and bounded operators on H are somewhat
“interchangeable”. Furthermore, if A is self adjoint, then the associated
bilinear form is symmetric, i.e.

B(f, g) = (Af, g) = (f, Ag) = (Ag, f ) = B(g, f ) ∀f, g ∈ H.

Conversely, given a bounded bilinear form satisfying

B(f, g) = B(g, f ) ∀f, g ∈ H,

the bounded linear operator A associated to B is self adjoint.

64 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

Definition 2.4.11 A linear operator on H is said to be unitary if

U ∗ U = U U ∗ = I.

Exercise 2.4.12 Prove that the following are equivalent:

(i) U is unitary;

(ii) Ran(U ) = H and (U f, U g) = (f, g) for every f and g in H.

(iii) Ran(U ) = H and kU f k = kf k for every f in H;

Show that if H is finite dimensional, we may omit the requirement Ran(U ) =

H in (ii) and (iii) above. An operator U such that kU f k = kf k for every f
in H is called an isometry. Construct an isometry on `2 which is not onto.

The next proposition relates the kernel and the range of the bounded oper-
ators A and A∗ .

Proposition 2.4.13 Suppose that A is a bounded linear operator on H.

Then
Ker(A∗ ) = Ran(A)⊥ and Ker(A) = Ran(A∗ )⊥ .

Proof. Note that A∗ y = 0 if and only if

(x, A∗ y) = 0 ∀x ∈ H,

which is equivalent to

(Ax, y) = 0 ∀x ∈ H,

which, of course, says that y is in the orthogonal complement of Ran(A)⊥ .

This proves the first assertion.
The second follows from the first applied to A∗ and the elementary fact
that A∗∗ = A (see Proposition 2.4.6 (ii)). 2

Note that Ker(A) and Ker(A∗ ) are always closed subspaces of H, but this is,
in general, no longer true for Ran(A) and Ran(A∗ ). Therefore the statements

Ker(A∗ )⊥ = Ran(A) and Ker(A)⊥ = Ran(A∗ )

are, in general, false.

2.5. COMPACT OPERATORS 65

Exercise 2.4.14 Construct examples of bounded operators for which the

statements above fail.

Definition 2.4.15 A bounded linear operator E on H is a projection if

E 2 = E. A projection E is said to be orthogonal when it is self adjoint.

Exercise 2.4.16 Suppose that E is a projection on H. The following are

equivalent:

(i) E is self adjoint;

(ii) Ran(E) = Ker(E)⊥ ;

(iii) (Ev, v) = kEvk 2 for every v in H.

2.5 Compact operators

In this section we begin the study of compact operators, which will play a key
role in the solution to the Dirichlet problem via integral equations. In a sense
that we shall make precise later, compact operators on infinite dimensional
Banach spaces have spectral properties not dissimilar from those of linear
operators on finite dimensional Banach spaces,

Definition 2.5.1 A bounded operator T on a Banach space B is said to be

compact if T B is relatively compact (with respect to the norm topology):
here B denotes the unit ball in B. The vector space of compact operators
on B will be denoted by K(B).

It is a well known fact (that we shall not prove here) that the closed unit ball
in a Banach space B is compact if and only if B is finite dimensional. As a
consequence, the identity operator I on a Banach space B is compact if and
only if B is finite dimensional.

Exercise 2.5.2 Suppose that T is in L(B). Prove that the following are
equivalent:

(i) T is a compact operator;

(ii) T transforms bounded sets into relatively compact sets;

66 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

(iii) for every bounded sequence {fk } in B, there exists a subsequence {fnk }
such that T fnk is convergent in B.

Definition 2.5.3 A bounded linear operator T on B is said to have finite

rank provided Ran(T ) is finite-dimensional.

Clearly, finite rank operators are compact, because the image of the unit
ball B in B is a bounded set in the finite-dimensional Banach space Ran(T ),
and therefore it is relatively compact.

Compact operators need not have finite rank. Indeed, consider the inter-
val I = [0, 1] and the operator T : C(I) → C(I), defined by
Z x
T f (x) = f (s) ds ∀f ∈ C(I) ∀x ∈ I.¯
0

Observe that T is compact. Indeed, suppose that {fk } is a bounded sequence

in C(I). Then {T fk } is a bounded and equicontinuous sequence of contin-
uous functions in I. By the Ascoli–Arzelà theorem, {T fk } is a relatively
compact sequence in C(I). Therefore there exists a subsequence {nk } such
that {T fnk } is convergent in the uniform norm on I, and hence in C(I).
However, Ran(T ) = C 1 (I), which is infinite dimensional, whence T does
not have finite rank.

By using a celebrated compactness criterion of M. Riesz, Fréchet and

Kolmogorov, it is not hard to show that the operator T above is compact
from Lp (I) to Lp (I) for all p ∈ [1, ∞).

It is a celebrated result of P. Enflo that there are Banach spaces B such

that the vector space of all finite rank operators is not dense in K(B) (with
respect to the operator norm). Fortunately, finite rank operators in Hilbert
spaces are dense in the space of compact operators, as the following result
shows.

Exercise 2.5.4 Prove that the vector space K(B) of all compact linear
operators on B is a two sided closed ideal in L(B).

Exercise 2.5.5 Suppose that K is a continuous function on C([0, 1]2 ).

Prove that the operator
Z 1
Af (s) := K(s, t) f (t) dt ∀f ∈ C([0, 1])
0
2.5. COMPACT OPERATORS 67

is compact. Prove that if K is of the following form

J
X
K(s, t) := φj (s) ψj (t) ∀s, t ∈ [0, 1],
j=1

then A has finite rank.

Exercise 2.5.6 Suppose that A is as in the previous exercise, except that

K is a function on [0, 1]2 such that its points of discontinuity are contained
in the graphs of a finite number ϕ1 , . . . , ϕN of continuous functions. Prove
that A is a compact operator.

Proposition 2.5.7 Suppose that H is a Hilbert space. Then the space of

finite rank operators is dense in K(H) with respect to the operator norm.

Proof. Pick C in K(H). Denote by {Pn } a sequence of orthogonal projections

such that Pn ↑ I (in the strong operator topology, i.e., limn→∞ kPn f − f k = 0
for every f ∈ H). Then Pn C is a finite rank operator. We prove that
lim |||C − Pn C||| = 0.
n→∞

Indeed,
|||C − Pn C||| = sup k(I − Pn )Cxk ≤ sup k(I − Pn )yk,
x∈B y∈K

where K = CB. Since C is compact, so is K. Now, set

fn (y) = k(I − Pn )yk ∀y ∈ K.
Then fn ↓ 0. Since K is compact and fn continuous, fn ↓ 0 uniformly, so
that
|||C − Pn C||| ≤ sup k(I − Pn )yk → 0,
y∈K

as required. 2

Exercise 2.5.8 By using Proposition 2.5.7, prove that a bounded operator

A on H is compact if and only if A∗ is compact.

Exercise 2.5.9 By using Proposition 2.5.7, prove that Hilbert–Schmidt

operators are compact. It may be useful to show preliminarily that if {ϕj }
is an orthonormal basis of L2 (Rn ), then the set of functions ψj,k , deifined by
ψj,k (x, y) := ϕj (x) ϕk (y)
is an orthonormal basis of L2 (Rn × Rn ).
68 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

Exercise 2.5.10 Suppose that H is a Hilbert space and that J is a nonnull

norm closed ideal in L(H). Prove that K(H) is contained in J , by showing
that any finite rank operator is in J . Hint: follow the following steps: First
step: Given v, w in H, define
Rv,w f := (f, v) w.
Show that Rv,w is a rank one operator with Ran(Rv,m ) = hwi, and that the
following formulae hold:
Rv,w Rv0 ,w0 = (w0 , v) Rv0 ,w αRv,w = Rv,αw , Rv,v = kvk 2 Phvi ,
where Phvi denotes the orthogonal projection onto the subspace generated
by v. Furthermore, if T is in L(H), then
Rv,w T = RT ∗ v,w and T Rv,w = Rv,T w .
Finally, if T 6= 0 and T h 6= 0, then
1
RT h,w T Rw,h = Rw,w .
kT hk 2
Second step: Show that J contains all finite rank operators. Third step: use
Proposition 2.5.7

2.6 The spectra of compact operators

In this section we discuss some properties of the spectrum of a compact op-
erator acting on a Hilbert space H. The spectrum of a bounded operator
on H may be any closed bounded subset of the complex plane. By contrast,
the spectrum of a compact operator is a bounded discrete set of the com-
plex plane. If the spectrum is not a finite set, then it has 0 as the unique
accumulation point.

Exercise 2.6.1 Suppose that A is a compact operator on an infinite di-

mensional Hilbert space. Prove that 0 belongs to the spectrum of A.

As a preliminary step towards a better understanding of the spectral

properties of compact operators, we consider operators of a very special form.
Suppose that {ϕk } is an orthonormal basis of H, and that A acts diagonably
on {ϕk }, i.e., there exist constants λk such that
Aϕk = λk ϕk k = 1, 2, 3, . . . (2.6.1)
2.6. THE SPECTRA OF COMPACT OPERATORS 69

We claim that A is compact if and only if λk → 0 as k tends to infinity.

Indeed, denote by PN the orthogonal projection onto the subspace of H
generated by ϕ1 , . . . , ϕN . Clearly PN is compact, for it has finite rank.
Now, if λk → 0, then
|||A − PN A||| = sup{kAf k : kf k = 1, f ∈ Ran(PN )⊥ }.
Every f ∈ Ran(PN )⊥ with norm equal to one is of the form
P
j>N cj ϕj ,
P 2
where the constants cj satisfy j>N cj = 1. Then
X
Af = cj λj ϕj .
j>N

Hence X
kAf k 2 = |cj |2 |λj |2 .
j>N

For every ε > 0 there exists ν such that |λj | < ε for every N ≥ ν. Therefore
X
|||A − PN A||| ≤ sup ε2 |cj |2
kf k=1,f ∈Ran(PN )⊥ j>N

= sup ε kf k 2 ,
2
kf k=1,f ∈Ran(PN )⊥

so that PN A tends to A in the uniform operator topology. Since PN A is

compact, A is compact by Porposition 2.5.7, as required.
Conversely, assume that A is compact, and consider the bounded se-
quence {ϕk }. Then there must exist a subsequence {kj } such that {Aϕkj }
is convergent in H as j tends to ∞. Since Aϕkj = λkj ϕkj and {Aϕkj } is a
Cauchy sequence, the difference
kAϕkj − Aϕkj0 k 2 = kλkj ϕkj − λkj0 ϕkj0 k 2 = |λkj |2 + |λkj0 |2
must converge to 0 as j and j 0 tend to infinity. Therefore λkj must converge
to 0. By applying a similar argument to every subsequence of the sequence
{ϕk }, we see that λk is convergent to 0.
As a consequence, we see that a compact operator, which acts diagonably
on an orthonormal basis, has the following property:
for each µ > 0, the set Eµ of all the eigenvalues λ of A with |λ| > µ is finite,
and for each λ ∈ Eµ the nullspace Ker(λ I − A) is finite dimensional.
The next theorem shows that this property holds for all compact opera-
tors.
70 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

Exercise 2.6.2 Prove that the spectrum of the operator A, defined in

(2.6.1), is the set
{λk : k = 1, 2, 3, . . .} ∪ {0},
and that 0 may or may not be an eigenvalue.

Theorem 2.6.3 Suppose that A is a compact operator on H. Then for each

µ > 0, the set Eµ of all the eigenvalues λ of A with |λ| > µ is finite, and for
each λ ∈ Eµ the nullspace Ker(λ I − A) is finite dimensional.

Proof. We argue by contradiction. Suppose that there exists a sequence {ϕj }

of linearly independent eigenvectors, associated to the eigenvalues {λj } with
kϕj k = 1 and
µ ≤ |λj | ≤ |||A|||.
By (possibly) passing to a subsequence, we may assume that {λj } is conver-
gent to λ, say. Obviously
µ ≤ |λ| ≤ |||A|||.
Set
Hm := span{ϕ1 , . . . , ϕm },
⊥
and denote by gm and element of Hm ∩ Hm−1 such that kgm k = 1.
Since gm ∈ Hm , there exist constants {cm
j : j = 1, . . . , m} such that

m
X
gm = cm
j ϕj .
j=1

Thus,
m
X
cm
m ϕm = gm − cm
j ϕj . (2.6.2)
j=1

Since ϕj is an eigenvector associated to the eigenvalue λj

m
X
Agm = cm
j λj ϕj
j=1
m−1
X
= cm
m λm ϕm + cm
j λj ϕj
j=1
m−1
X
(by (2.6.2)) = λm gm + cm
j (λj − λm ) ϕj .
j=1
2.6. THE SPECTRA OF COMPACT OPERATORS 71

Therefore
Agm − λm gm ∈ Hm−1 .
Clearly {gm } is a bounded sequence in H. We shall prove that {Agm } does
not admit a convergent subsequence, thereby contradicting the compactness
of A.
Indeed, suppose that m < n. Then
kAgn − Agm k2 = kλn gn − vk2 ,
⊥
where, by (2.6.2), v ∈ Hn−1 . Since gn ∈ Hn−1 ,
kAgn − Agm k2 ≥ kλn gn k2
= |λn |2 kgn k2
≥ µ2 .
Hence {Agm } has no Cauchy subsequences, as required.
2
Operators of the form λ I − A, where A is compact, arise also in the theory of
second order elliptic equations, as we shall see. Fredholm studied some of the
properties of these operators, and the resulting theory is now now commonly
referred to as Fredholm theory. The following theorem is the main result
of Fredholm theory. We point out that Fredholm theory was extended to
Banach spaces, mainly by Riesz and Schauder. See [Br] or [Y] for excellent
expositions of this generalisation.

Theorem 2.6.4 Suppose that A is a compact operator on H and that λ 6= 0.

Then the following hold:

(i) the range of λ I − A is closed;

(ii) the range of λ I − A is H if and only if the kernel of λ I − A∗ is trivial;
(iii) λ I − A is injective if and only if λ I − A∗ is injective;
(iv) λ I − A is injective if and only if λ I − A is surjective;
(iv) Ker(λ I − A) and Ker(λ I − A∗ ) have the same dimension.

Proof. To prove (i), suppose that gj → g, where gj = (λ I − A)fj , for some

fj in H. We must show that there exists f ∈ H such that g = (λ I − A)f .
We have a unique decomposition
fj = uj + vj ,
72 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

where uj ∈ Ker(λ I − A), and vj ∈ Ker(λ I − A)⊥ . Recall that Ker(λ I − A)

is closed, because it is finite dimensional. Thus,

gj = λvj − Avj j = 1, 2, . . . (2.6.3)

We claim that {vj } is bounded. We argue by contradiction. Suppose that

{vj } is not bounded. We may assume that kvj k → ∞ (otherwise we consider
a suitable subsequence of {vj }). Set
vj
wj := j = 1, 2, . . .
kvj k
Since A is compact, there exists a subsequence, which we still denote by
{wj }, such that {Awj } is convergent to z, say. From (2.6.3), and the fact
that {gj } is bounded (for it is a convergent sequence) we deduce that

0 = λ lim wj − z.
j

In particular, kzk = |λ|. Observe that z is in Ker(λI − A), for

λz − Az = λ lim(λwj − Awj ) = 0.
j

But vj ∈ Ker(λ I − A)⊥ by assumption, hence so is wj . Since vj ∈ Ker(λ I −

A)⊥ is closed and λ 6= 0, z is in Ker(λ I − A)⊥ . Therefore z = 0, which
contradicts the fact that kzk = |λ| and λ 6= 0.
The claim is proved.
Since {vj } is bounded, there exists a subsequence, which we denote still
by {vj }, and a point ζ ∈ H such that Avj → ζ. Thus, by (2.6.3),

λ lim vj = g + ζ,
j

and
(λ I − A)((g + ζ)/λ) = lim (λ I − A)vj = lim gj = g.
j→∞ j→∞

This concludes the proof of (i).

Part (ii) follows directly from (i) and Proposition 2.4.13.
Next we prove (iii). We begin by showing that if Aλ is injective, then
A∗λ is injective. We argue by reductio ad absurdum. Suppose that A∗λ is not
injective. Observe that, by (i) and Proposition 2.4.13, there is an orthogonal
decomposition

H = Ran(Aλ ) + Ker(A∗λ ) (orthogonal sum).

2.6. THE SPECTRA OF COMPACT OPERATORS 73

Since Aλ∗ is not injective, Ker(A∗λ ) is not trivial, whence Ran(Aλ ) is a proper
(closed) subspace of H. Set Rjλ := Ran(Ajλ ).
We claim that

Rjλ ⊃ Rj+1
λ j = 1, 2, 3, . . . (proper inclusion). (2.6.4)

For otherwise there is an index j such that Rjλ = Rj+1λ . Then there are
j−1 j j
v ∈ Rλ \Rλ and w ∈ Rλ such that Aλ v = Aλ w. But this and the injectivity
of Aλ imply that v = w, which contradicts the fact that v ∈ Rj−1
λ \ Rjλ .
By (i), each Rjλ is a closed subspace of H, hence a proper closed subspace
of Rj−1 j j+1
λ . Therefore there exists a unit vector fj ∈ Rλ , orthogonal to Rλ .

We claim that {Afj } has no Cauchy subsequences. Indeed, suppose that

j < k. Then
Afk − Afj = λ fk − Aλ fk − λ fj + Aλ fj .
Observe that Aλ fk + λ fj − Aλ fj is in Rk+1
λ , and that λ fk is orthogonal to
k+1
Rλ . Therefore
kAfk − Afj k2 ≥ |λ|2 kfk k2 = |λ|2 ,
as claimed.
This contradicts the compactness of A, and concludes the proof of one
implication of (iii).
To prove the reverse implication, we may repeat the same reasoning with
Aλ∗ playing the role of Aλ , and using the fact that (A∗λ )∗ = Aλ .
The proof of (iii) is complete.
To prove (iv), observe that, by Proposition 2.4.13 and (i), λ I − A is
injective if and only if Ran(λ I − A∗ ) = H.
Finally we prove (v). For the sake of simplicity, set

Nλ := Ker(λ I − A) and Nλ∗ := Ker(λ I − A∗ ).

Denote by δ and δ ∗ the dimensions of Nλ and Nλ∗ , respectively.

We prove that δ ∗ ≤ δ. Suppose that δ < δ ∗ . Then there exists a linear
operator Λ : Nλ → Nλ∗ , which is injective but not surjective. Denote by P
the orthogonal projection onto Nλ , and consider the operator

S := A + Λ ◦ P.

Note that Λ ◦ P is of finite rank, hence compact. Thus, S is compact, for it

is the sum of two compact operators.
74 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

We claim that Ker(λ I − S) = {0}. Indeed, suppose that u is in Ker(λ I −

S). Then
λ u − Au − Λ(P u) = 0.
Since Λ(P u) is in Nλ∗ , λ u−Au is in Ran(λ I −A), and Ran(λ I −A) = (Nλ∗ )⊥
by Proposition 2.4.13 and (i), we may concude that

λu − Au = 0 and Λ(P u) = 0.

These equations imply that Λu = 0, whence u = 0 because Λ is injective.

The proof of the claim is complete.
By the claim and (iv), λ I − S is surjective. But this contradicts the fact
that Λ is not surjective. Indeed, suppose that g belongs to Nλ∗ \ Ran(Λ), and
consider the equation
λu − Su = g,
which is equivalent to
λu − Au − Λ(P u) = g.
Since g ∈ Nλ∗ and λu − Au is orthogonal to Nλ∗ , we must have g = Λ(P u),
which contradicts the assumption that g be off the range of Λ.
This concludes the proof that δ ∗ ≤ δ. The same reasoning, with λ − A∗
in place of λI − A, and the fact that

λI − A = (λ − A∗ )∗ ,

prove that δ ≤ δ ∗ .
This concludes the proof of (v), and of the theorem. 2

We conclude this section by deriving some consequences of the result above.

Recall the following consequence of the closed graph theorem.

Proposition 2.6.5 Suppose that B is a Banach space, that A is in L(B) is

a bijective operator (onto B). Then A−1 (which clearly exists) is in L(B).

The following corollary describes the spectrum of a compact operator.

Corollary 2.6.6 Suppose that A is a compact operator on the Hilbert space

H. If λ is in σ(A) \ {0}, then λ is an isolated eigenvalue of A of finite
multiplicity.
2.7. THE SPECTRAL THEOREM 75

Proof. Suppose that λ ∈ σ(A). If λ is not an eigenvalue, then λ I − A is

injective, hence it is surjective by Theorem 2.6.4 (iv). Then (λ I − A)−1 ∈
L(H) by Proposition 2.6.5, so that λ is in the resolvent set of A.
Thus, λ is an eigenvalue. The associated eigenspace is finite dimensional
by Theorem 2.6.3, as required. 2

Thus, the specrtum of a compact operator on an infinite dimensional sepa-

rable Hilbert space is one of the following

(i) 0 (which may or may not be an eigenvalue) and a finite number (pos-
sibly 0) of nonzero eigenvalues with finite multiplicities;

(ii) 0 (which may or may not be an eigenvalue) and a sequence {λk } of

nonzero eigenvalues with finite multiplicities that tends to 0 as k tends
to ∞.

An interesting fact, which we shall use in the sequel without any further
reference, is that all the results in this section hold, with different proofs, for
compact operators on Banach spaces (see the books [Br, Y] for the corre-
sponding proofs).

2.7 The spectral theorem for compact self ad-

joint operators
In this section we prove the spectral theorem for compact self adjoint opera-
tors on a Hilbert space H that generalises the spectral theorem for Hermitian
matrices on finite dimensional Hilbert spaces. This result may be further gen-
eralised to bounded and even to unbounded linear operators on H. The result
for bounded operators is due to Hilbert and that for unbounded operators
generalisation is a deep result of von Neumann, which we do not discuss here.

Theorem 2.7.1 Suppose that A is a compact self adjoint operator on the

Hilbert space H. Then there exists an orthonormal basis {ϕk }, which consists
of eigenvectors of A. Furthermore, the sequence {λk } of the corresponding
eigenvalues is real and is convergent to 0 as k tends to ∞.

We have already proved that if {ϕk } is an orthonormal basis of H, and A

acts diagonally on {ϕk } by

Aϕk = λk ϕk ,
76 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

then λk → 0 as k tends to infinity if and only if A is a compact operator. If

all the eigenvalues λk are real, then A is self adjoint.
Proof. The proof is quite long. We split it up in several steps.

Step I: if λ is an eigenvalue of A, then λ is real.

Suppose that λ is a nonzero eigenvalue and that ϕ is a corresponding
eigenvector. Then

λ (ϕ, ϕ) = (Aϕ, ϕ) = (ϕ, Aϕ) = λ (ϕ, ϕ),

which implies the required conclusion.

Step II: eigenvectors associated to distinct eigenvalues are or-

thogonal.
Suppose that ϕ1 and ϕ2 are eigenvectors associated to the eigenvalues λ1
and λ2 , with λ1 6= λ2 . Then

λ1 (ϕ1 , ϕ2 ) = (Aϕ1 , ϕ2 ) = (ϕ1 , Aϕ2 ) = λ2 (ϕ1 , ϕ2 ),

so that (ϕ1 , ϕ2 ) = 0, as required.

Step III: for each µ > 0, the set Eµ of all the eigenvalues λ of A
with |λ| > µ is finite, and for each λ ∈ Eµ the nullspace Ker(λ I − A)
is finite dimensional.
This is just Theorem 2.6.3 above, which holds for all compact operators.

Step IV: either |||A||| or −|||A||| is an eigenvalue of A.

By Proposition 2.4.9, either

|||A||| = sup (Af, f ) or − |||A||| = inf (Af, f ).

kf k=1 kf k=1

For the sake of definiteness, assume that the first of the two equalities above
holds. Clearly, there exists a sequence {fk } ∈ H with kfk k = 1, such that

|||A||| = lim (Afk , fk ). (2.7.1)

k→∞

Since A is compact, there exists a subsequence of {fk }, which we still denote

by {fk }, such that {Afk } is convergent in H to a function, g say. Write
Afk = Afk − g + g on the right hand side of (2.7.1) above. Observe that

(Afk − g, fk ) ≤ kAfk − gk kfk k → 0

2.7. THE SPECTRAL THEOREM 77

as k tends to infinity, so that

|||A||| = lim (g, fk ).

k→∞

Note also that for every k

|||A||| ≥ sup kAf k ≥ kAfk k.

kf k=1

By taking the limit of both sides as k tends to infinity, we get that

|||A||| ≥ kgk. (2.7.2)

Now, notice that

kAfk − |||A||| fk k 2 = kAfk k 2 − 2|||A||| (Afk , fk ) + |||A|||2 kfk k 2 .

Then we take the limit of both sides as k tends to infinity, and obtain

lim kAfk − |||A||| fk k 2 = kgk 2 − |||A|||2 .

k→∞

From (2.7.2) and the fact that the left hand side is nonnegative, we deduce
that |||A||| = kgk and, consequently, that

lim kAfk − |||A||| fk k 2 = 0.

k→∞

Therefore fk tends to g/|||A|||, and Ag = |||A||| g, so that g is an eigenvector

associated to the eigenvalue |||A|||, as required to conclude the proof of Step IV.

Step V: conclusion.
We argue by reductio ad absurdum. Suppose that the closure S of the
span of all the eigenvectors of A is not dense in H. Then S ⊥ is a nontrivial
closed subspace of H, hence an Hilbert space itself. It is straightforward
to check that AS ⊆ S and that AS ⊥ ⊆ S ⊥ . Furthermore, A, restricted to
S ⊥ , is clearly a bounded self adjoint operator on S ⊥ . Therefore, by Step IV
(applied to A|S ⊥ ), it has at least one nonzero eigenvalue. This clearly would
be also an eigenvalue of the original operator A, which contradicts the fact
that S is the closure of the span of all the eigenvalues of A.
This concludes the proof of Step V, and of the theorem. 2

Note that 0 may or may not be an eigenvalue of a bounded self adjoint oper-
ator. Furthermore, if 0 is an eigenvalue, then the corresponding eigenspace
may be finite or infinite dimensional. The following examples should help
clarifying the situation.
78 CHAPTER 2. LINEAR OPERATORS ON BANACH SPACES

Example 2.7.2 Suppose that S is a finite dimensional subspace of the

infinite dimensional Hilbert space H and denote by E the orthogonal projec-
tion of H onto S. Clearly E s a compact operator which has two eigenvalues,
0 and 1: the eigenspace associated to the eigenvalue 0 is just S ⊥ , which is
infinite dimensional.

Example 2.7.3 Suppose that {ϕk : k = 1, 2, . . .} is an orthonormal basis

of H. Define the operator A on this basis by
1
Aϕk = ϕk ,
k
and then extend it by linearity on H. Then A is a compact self adjoint
operator with eigenvalues {1/k}: thus, 0 is not an eigenvalue.

Exercise 2.7.4 Give all the details of the example above.

Example 2.7.5 Denote by I the interval [0, 1] and by K : I × I → R the

function defined by (
1 if t ≤ x
K(t, x) =
0 if t > x.
Let A : L2 (I) → L2 (I) be the Hilbert–Schmidt operator with kernel K.
By Exercise 2.5.9, the operator A is compact. Since the kernel K is not
Hermitian, A is not self adjoint. It may be proved that A has no eigenvalues
at all.

Exercise 2.7.6 Give all the details of the example above.

Example 2.7.7 Denote I the interval [0, 1], and by A : L2 (I) → L2 (I) the
operator defined by

Af (x) := x2 f (x) ∀x ∈ I.

It is straightforward to check that A is a bounded self adjoint operator, with

|||A||| = 1. It may be proved that A is not compact and has no eigenvalues.

Exercise 2.7.8 Give all the details of the example above.

Chapter 3

Dirichlet problem via integral

equations

3.1 Kernels of type α

In this section we analyse the properties of a class of integral operators on
smooth compact hypersurfaces of Rn . In particular, we investigate their
compactness. The results we shall obtain will be applied to the double layer
potential, and to a suitable derivative of the single layer potential.
Definition 3.1.1 Suppose that Ω is a bounded domain in Rn and that α
is a number in (0, n − 1). A measurable function K : ∂Ω × ∂Ω → C is a
kernel of type α if
K(X, Y ) = A(X, Y ) |X − Y |−α ∀X, Y ∈ ∂Ω, X 6= Y,
where A is a bounded function on ∂Ω × ∂Ω. If K is of the form
K(X, Y ) = A(X, Y ) log |X − Y | + B(X, Y ) ∀X, Y ∈ ∂Ω, X 6= Y,
where A and B are bounded functions on ∂Ω × ∂Ω, then we say that K
is a kernel of type 0. If K is a kernel of type α ∈ (0, n − 1) and A is a
continuous function on ∂Ω×∂Ω, then we say that K is a continuous kernel
of type α. Similarly, if K is a kernel of type 0, and the functions A and B
are continuous on ∂Ω × ∂Ω, then we say that K is a continuous kernel of
type 0.
If K is a kernel of order α ∈ [0, n − 1), then we define the operator TK
associated to K formally by
Z
TK f (X) := K(X, Y ) f (Y ) dσ(Y ) ∀X ∈ ∂Ω. (3.1.1)
∂Ω

79
80 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

It is straightforward to check that TK f makes sense for every bounded func-

tion f on ∂Ω. We now investigate possible extensions of TK to Lp (∂Ω).
We shall need the following boundedness criterion for operators acting on
Lp (∂Ω). We state it in the more general setting of a σ-finite measure space.

Proposition 3.1.2 Suppose that M is a measure space (with measure µ) and

that K is a complex valued measurable function on M × M , which satisfies
the following estimates
Z
C0 := sup |K(x, y)| dµ(y) < ∞
x∈M M

and Z
C1 := sup |K(x, y)| dµ(x) < ∞.
y∈M M

Then the operator TK , formally defined by

Z
TK f (x) = K(x, y) f (y) dµ(y),
M

extends to a bounded operator on Lp (M ) for all p ∈ [1, ∞]. Furthermore,

1/p0 1/p
|||TK ||| ≤ C0 C1 ,

where p0 denotes the index conjugate to p.

Proof. If p = ∞, the hypothesis C1 < ∞ is superfluous, and the conclusion

is obvious. Thus, we may assume that p < ∞. By Hölder’s inequality
Z 1/p0  Z 1/p
|TK f (x)| ≤ |K(x, y)| dµ(y) |K(x, y)| |f (y)|p dµ(y)
M M
0
Z 1/p
1/p
≤ C0 |K(x, y)| |f (y)|p dµ(y) .
M

We now compute the p norm of both sides (with respect to the variable x),
and obtain
Z Z 1/p
1/p0
kTK f kp ≤ C0 dµ(x) |K(x, y)| |f (y)|p dµ(y)
M M
1/p0 1/p
≤ C0 C1 kf kp ,

as required. 2
3.1. KERNELS OF TYPE α 81

Some important properties of kernels of type α are established in the next

proposition. We shall need two lemmata. Loosely speaking, Lemma 3.1.3
below says that, at a local scale, the distance between two points in ∂Ω is
(uniformly) comparable to that of the corresponding images via a (suitable)
coordinate function.

Lemma 3.1.3 Suppose that Ω is a bounded domain with C 2 boundary. Then

there exist local charts (Uj , Φj ), j ∈ {1, . . . , N }, such that the following hold:

(i) Φj is an homeomorphism between Uj and Φj (Uj ) =: Vj , and its inverse

Ψj := Φ−1 2
j : Vj → Uj is of class C (Vj );

(ii) Ψj has the following form

Ψj (ξ) = ξ, fj (ξ) ∀ξ ∈ Vj ,

i.e., that ∂Ω ∩ Uj is the graph of the function fj : Vj → R (clearly

fj ∈ C 1 (Vj ), because ∂Ω is assumed to be of class C 2 );
S
(iii) j Uj = ∂Ω;

(iv) there exist positive constants C0 and δ0 such that for each X ∈ ∂Ω there
exists j ∈ {1, . . . , N } with B4δ0 (X) ∩ ∂Ω ⊂ Uj , and




Φj Bδ (X) ∩ ∂Ω ⊆ Bδ Φj (X) ⊆ Φj BC0 δ (X) ∩ ∂Ω ∀δ ∈ (0, δ0 ],

equivalently


Bδ (X) ∩ ∂Ω ⊆ Ψj Bδ (Φj (X)) ⊆ BC0 δ (X) ∩ ∂Ω ∀δ ∈ (0, δ0 ].

Proof. For each point Z ∈ ∂Ω there is a local chart (UZ , ΦZ ) such that Z
is in UZ , the map ΦZ is an homeomorphism between UZ and Φ(UZ ) =: VZ
and its inverse ΨZ := Φ−1 2
Z is of class C (VZ ). By possibly shrinking the open
set UZ , and remaining the variables, we may assume that ΨZ is of the form
ξ 7→ (ξ, fZ (ξ)), where fZ is a function of class C 2 (VZ ), and that VZ is a ball
n−1
S
with centre 0 ∈ R . Clearly Z∈∂Ω UZ = ∂Ω. Since ∂Ω is compact, we can
choose a finite subcovering UZ1 , . . . , UZN , which clearly satisfies (i)-(iii). In
the sequel we shall write Uj instead of UZj .
Now we prove (iv). For each X ∈ ∂Ω and j ∈ {1, . . . , N }, set
(
0 if X ∈ / Uj
τj (X) :=
sup{r > 0 : Br (X) ∩ ∂Ω ⊆ Uj } if X ∈ Uj .
82 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Then set
τ (X) := max{τ1 (X), . . . , τN (X)}
If τ (X) = τk (X), then we say that X is related to Uk , and write X ∼ Uk . .
Since {U1 , . . . , UN } is a covering of ∂Ω, each X belongs to at least to one of
the sets U1 , . . . , UN , so that τ (X) > 0 for every X ∈ ∂Ω. Define
τ := inf τ (X). (3.1.2)
X∈∂Ω

We claim that τ > 0. We argue by reductio ad absurdum. Suppose that

τ = 0. Then there exists a sequence {Xk } of points in ∂Ω such that
lim τ (Xk ) = 0. (3.1.3)
k→∞

By compactness, there exists a subsequence {Xk` } of {Xk } which is conver-

gent to some point, X∗ say, in ∂Ω. Suppose that ε = (1/2) τ (X∗ ), and that
X∗ is related to Uh . Then there exists `0 such that
|Xk` − X∗ | < ε ∀` ≥ `0 .
Then, by the triangle inequality, Bε/2 (Xk` ) ⊂ Uh for every ` ≥ `0 , whence
τ (Xk` ) > ε/2 for all such `, thereby contradicting (3.1.3). Thus, the claim is
proved.
Now, we define
[

Kj := Bτ /2 (X) ∩ ∂Ω . (3.1.4)
X∈Uj

Each point in Bτ /2 (X) ∩ ∂Ω has distance at least τ /2 from ∂Uj , for X ∼ Uj

and τ < τ (X) = τj (X). Thus, Kj is a compact subset of Uj . Define


κj := co Φj (Kj ) , (3.1.5)
the convex hull of Kj . Clearly Φj (Kj ) is compact, hence κj is compact (see
Exercise 3.1.4 below). Furthermore, Φj (Kj ) ⊂ VZj , whence κj is a compact
subset of VZj (recall that VZj is an open ball in Rn−1 ). Set
q
Cj := sup ||| dΨj (ξ)||| and C0 := max 1 + Cj2 . (3.1.6)
ξ∈κj j=1,...,N

Suppose that X and Y are in Uj and set Φj (X) = ξ and Φj (Y ) = η. Observe

that
|ξ − η|2 ≤ |Ψj (ξ) − Ψj (η)|2 = |X − Y |2
= |ξ − η|2 + |fj (ξ) − fj (η)|2
≤ |ξ − η|2 + |∇fj (ω) · (ξ − η)|2 (3.1.7)
≤ |ξ − η|2 1 + sup |∇fj (ω)|2 .
 
ω∈[ξ,η]
3.1. KERNELS OF TYPE α 83

Choose any δ0 > 0 such that

τ
δ0 < , (3.1.8)
4C0
and suppose that δ ≤ δ0 . Fix a point X ∈ ∂Ω. For the sake of definiteness,
suppose that X is related to Uj . Assume that Y ∈ Bδ (X) ∩ ∂Ω. Then X
and Y belong to Kj , and estimate (3.1.7) applies. Notice that the segment
[ξ, η] is contained in κj , so that
|ξ − η|2 ≤ |X − Y |2 ≤ C02 |ξ − η|2 . (3.1.9)


The left hand inequality in (3.1.9) implies that Φj (Bδ (X)) ⊂ Bδ Φj (X)
for
every δ ≤ δ0 .
The right hand inequality in (3.1.9) implies that Bδ Φj (X) ⊂
Φj BC0 δ (X) for every δ ≤ δ0 .
This concludes the proof of (iv), and of the lemma. 2

Exercise 3.1.4 Prove that the convex hull of a compact set in Rn is com-
pact.

The next lemma is a sort of substitute for integration of “radial” functions

in polar coordinates on ∂Ω.

Lemma 3.1.5 Suppose that Ω is a bounded domain with C 2 boundary and

that α ∈ [0, n − 1). Then there exists a constant C1 (see (3.1.13) below) such
that Z
sup |X − Y |−α dσ(Y ) ≤ C1 δ n−1−α ∀δ > 0.
X∈∂Ω Bδ (X)∩∂Ω

Proof. Let δ0 > 0, U1 , . . . , UN and Φj : Uj → Rn−1 , j = 1, . . . , N be as in

Lemma 3.1.3. The left hand inequality in (3.1.9) implies that
|X − Y |−α ≤ |ξ − η|−α ∀ξ, η ∈ Φ(Kj ). (3.1.10)

the left hand inequality in (3.1.9) implies that Φj (Bδ (X)) ⊂

Furthermore,
Bδ Φj (X) for every δ ≤ δ0 . Thus, if X is related to Uj ,
Z Z r
−α
|X − Y | dσ(Y ) ≤ |ξ − η|−α 1 + sup |∇fj |2 dη
Bδ (X)∩∂Ω Bδ (Φj (X)) Φj (Kj )
Z
≤ C0 |ξ − η|−α dη
Bδ (Φj (X))
Z δ
= C0 ωn−1 rn−2−α dr
0
ωn−1
= C0 δ n−1−α ∀δ ∈ (0, δ0 ].
n−1−α
(3.1.11)
84 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

This proves the required result for δ ≤ δ0 . Suppose now that δ > δ0 . Then
Z
|X − Y |−α dσ(Y )
Bδ (X)∩∂Ω
Z
≤ |X − Y |−α dσ(Y ) + δ0−α σ(∂Ω)
Bδ0 (X)∩∂Ω
ωn−1
≤ C0 δ0n−1−α + δ0−α σ(∂Ω) (3.1.12)
n − 1 −h α
ωn−1 i
≤ δ n−1−α C0 + δ 1+α−n δ0−α σ(∂Ω)
hn − ω
1−α i
n−1
≤ δ n−1−α C0 + δ01+α−n δ0−α σ(∂Ω)
n−1−α
n−1−α
= C1 δ ,
where
ωn−1
C1 := C0 + δ01−n σ(∂Ω). (3.1.13)
n−1−α
This concludes the proof of the lemma. 2

Proposition 3.1.6 Suppose that K is a kernel of type α ∈ [0, n − 1) and

that TK denotes the corresponding integral operator (see (3.1.1)). The fol-
lowing hold:

(i) for every p ∈ [1, ∞] the operator TK is bounded on Lp (∂Ω);

(ii) if ε > 0, and K(X, Y ) = 0 for all X, Y ∈ ∂Ω such that |X − Y | ≥ ε,
then
|||TK |||p ≤ C εn−1−α kAk∞
when α > 0, and

|||TK |||p ≤ C εn−1 kAk∞ log ε−1 + kBk∞

 

when α = 0;
(iii) if α ∈ [0, n − 1), then TK is a compact operator on L2 (∂Ω).

Proof. To prove (i), we apply Proposition 3.1.2. We shall prove the required
result in the case where α > 0. The case where α = 0 is left to the reader.
Observe that for every δ > 0
Z Z
|K(X, Y )| dσ(Y ) ≤ kAk∞ |X − Y |−α dσ(Y )
∂Ω ∂Ω
≤ kAk∞ C1 δ n−1−α ,
3.1. KERNELS OF TYPE α 85

the last inequality being a consequence of Lemma 3.1.5. Thus,

Z
sup |K(X, Y )| dσ(Y ) < ∞. (3.1.14)
X∈∂Ω ∂Ω

A similar argument, with the roles of X and Y interchanged, proves that

Z
sup |K(X, Y )| dσ(X) < ∞, (3.1.15)
Y ∈∂Ω ∂Ω

and (i) follows from this and Proposition 3.1.2.

Now we prove (ii). We give full details in the case where α > 0. The case
α = 0 is left to the interested reader. A close examination of the proof of (i)
and of Lemma 3.1.5 shows that if ε is small enough, then

|||TK |||p ≤ C εn−1−α ,

as required.
Finally we prove (iii). For each ε > 0 we set

Dε := {(X, Y ) ∈ ∂Ω × ∂Ω : |X − Y | ≤ ε}.

Correspondingly, we define

Kε (X, Y ) := K(X, Y ) 1Dε (X, Y ).

Clearly K − Kε is bounded on ∂Ω × ∂Ω, so that K − Kε ∈ L2 (∂Ω × ∂Ω) for

∂Ω is compact. Therefore the corresponding operator TK−Kε is a Hilbert–
Schmidt operator, hence it is compact on L2 (∂Ω).
We shall prove that
lim |||TKε |||2 = 0. (3.1.16)
ε↓0

The required conclusion then follows from the fact that the ideal of compact
operators is closed in L(L2 (∂Ω)) (see Exercise 2.5.4).
To prove (3.1.16) observe that TKε is the integral operator whose kernel
vanishes in Dεc and is equal to K(X, Y ) in Dε . By (ii),

|||TKε |||2 ≤ C εn−1−α ,

which tends to 0 as ε tends to 0, as required. 2

So far, we have studied properties of operators associated to generic kernels

of type α, where α ∈ [0, n − 1). Now we concentrate on operators associated
to continuous kernels of order α.
86 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Proposition 3.1.7 Suppose that K is a continuous kernel of order α, where

α ∈ [0, n − 1). The following hold:

(i) the operator TK maps bounded functions on ∂Ω into continuous func-

tions;

(ii) if λ 6= 0, u ∈ L2 (∂Ω) and λu + TK u ∈ C(∂Ω), then u ∈ C(∂Ω).

Proof. First we prove (i). Suppose that f is a bounded function on ∂Ω. We

shall prove that TK f is continuous on ∂Ω. Fix X ∈ ∂Ω, and suppose that
Y ∈ Bδ (X) ∩ ∂Ω. Then
Z
 
TK f (Y ) − TK f (X) = K(Y, Z) − K(X, Z) f (Z) dσ(Y )
∂Ω Z h i
≤ kAk∞ kf k∞ |Y − Z|−α + |X − Z|−α dσ(Y )
B (X)
Z 2δ
+ kf k∞ K(Y, Z) − K(X, Z) dσ(Z).
∂Ω\B2δ (X)

Notice that for each δ > 0

Z
lim K(Y, Z) − K(X, Z) dσ(Z) = 0 (3.1.17)
Y →X ∂Ω\B2δ (X)

by the dominated convergence theorem. Indeed,

K(Y, Z) − K(X, Z) ≤ K(Y, Z) + K(X, Z)

≤ 2 kAk∞ δ −α ∀Y ∈ Bδ (X) ∀Z ∈ ∂Ω \ B2δ (X),

and
lim K(Y, Z) − K(X, Z) = 0 ∀Z ∈ ∂Ω \ B2δ (X).
Y →X

Now, (3.1.17) and Lemma 3.1.5 imply that

TK f (Y ) − TK f (X)
Z
n−1−α
≤ kAk∞ kf k∞ C1 δ + kf k∞ K(Y, Z) − K(X, Z) dσ(Z).
∂Ω\B2δ (X)

By taking the limit of both sides as Y tends to X, we see that

lim TK f (Y ) − TK f (X) ≤ kAk∞ kf k∞ C1 δ n−1−α ∀δ > 0.

Y →X
3.1. KERNELS OF TYPE α 87

By taking the infimum of both sides over all δ > 0, we obtain that
lim TK f (Y ) − TK f (X) = 0,
Y →X

as required to conclude the proof of (i).

Next we prove (ii). For each ε > 0 denote by φε a continuous function on
∂Ω × ∂Ω that vanishes on
{(X, Y ) ∈ ∂Ω × ∂Ω : |X − Y | ≥ ε}
and is equal to 1 in
{(X, Y ) ∈ ∂Ω × ∂Ω : |X − Y | ≤ ε/2}.
Set
Kε0 := φε K and Kε1 := (1 − φε ) K.
Note that Kε0 is supported near the diagonal of ∂Ω × ∂Ω, whereas Kε1 is
supported off the diagonal.
We claim that TKε1 u is continuous on ∂Ω. Indeed,
Z
 1
Kε (Y, Z) − Kε1 (X, Z) |u(Z)| dσ(Y )

TKε1 u(Y ) − TKε1 u(X) ≤
∂Ω
hZ  2 1/2
(by Schwarz) ≤ kuk2 Kε1 (Y, Z) − Kε1 (X, Z) |u(Z)| dσ(Y )
∂Ω
(Kε1 is continuous) →0 as Y tends to X,
(3.1.18)
as claimed.
Now, set

g := λu + TK u − TKε1 u.
Notice that g is continuous, for λu + TK u is continuous by assumption,
and we have just proved that TKε1 u is continuous. Recall that, by Propo-
sition 3.1.6 (ii),
TKε1 p ≤ C εn−1−α .
for all p ∈ [1, ∞]. Thus, we may choose ε so small that
TKε1 p
< |λ| (3.1.19)

(recall that λ 6= 0 by assumption). Then the operator λ I + TKε1 is invertible

both in L2 (∂Ω) and in L∞ (∂Ω). By a classical result (see Exercise 3.1.8)
∞ 
−1 −1 −1 −1 −1
X TK 1 j
ε
(λ I + TKε1 ) =λ (I + λ TKε1 ) ) = λ ,
j=0
λ
88 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

so that
∞ 
−1
X TK 1 jε
u=λ g.
j=0
λ

Note that, by (i), each summand of the series above is a continuous function
on ∂Ω. Since
X∞
kTKj ε1 gk∞ X∞
|||TKε1 |||j
≤ kgk ∞ < ∞,
j=0
λj j=0
λj

(the last inequality follows from (3.1.19)) the sum of the series above is
continuous on ∂Ω, whence u is continuous on ∂Ω, as required to conclude
the proof of (ii), and of the proposition. 2

Exercise 3.1.8 Suppose that A is a bounded linear operator on the Banach

space B, and that |||A||| < 1. Then I − A is invertible in the Banach algebra
L(B), and
X ∞
−1
(I − A) = Aj ,
j=0

the series being convergent in L(B). Hint: prove that the series above is
convergent in L(B) by using the analogue
P∞ of the classical Weierstrass test for
series of elements in B. Then set B := j=0 Aj , and show that (I − A) B =
I = B (I − A).

3.2 The double layer potential

Suppose that Ω is a bounded domain with C 2 boundary. In this section we
study some properties of the double layer potential. First note that

1 (Y − x) · ν(Y )
∂ν(Y ) N (x − Y ) = ∀Y ∈ ∂Ω ∀x 6= Y. (3.2.1)
ωn |x − Y |n

We already know that the Newtonian potential x 7→ N (x − Y ) is harmonic

(with respect to the x variable) in Rn \ {Y }. Since
     
∆x ∂ν(Y ) N (x−Y ) = ∆x ∇Y N (x−Y )·ν(Y ) = ∇Y ∆x N (x−Y ) ·ν(Y ) = 0,

we see that x 7→ ∂ν(Y ) N (x − Y ) is harmonic in Rn \ {Y }.

3.2. THE DOUBLE LAYER POTENTIAL 89

Definition 3.2.1 The double layer potential with density g is the func-
tion Z
Dg(x) = g(Y ) ∂ν N (x − Y ) dσ(Y ) ∀x ∈ Rn \ ∂Ω.
∂Ω

The function (x, Y ) 7→ ∂ν N (x − Y ) is called the kernel of the double layer

potential.

Lemma 3.2.2 If g ∈ C(∂Ω), then Dg is harmonic in Rn \ ∂Ω.

Proof. The required property follows from the remark above by differentiat-
ing under the integral sign. The reader should check that differentiation is
permitted. 2

Now, we study Dg(X) when X ∈ ∂Ω. The difficulty here is that the ker-
nel of the double layer potential has a singularity on ∂Ω, so that even the
convergence of the integral
Z
g(Y ) ∂ν(Y ) N (X − Y ) dσ(Y )
∂Ω

must be proved for every X ∈ ∂Ω. As a first attempt, we examine the order
of magnitude of the kernel. By (3.2.1),

∂ν(Y ) N (X − Y ) ≤ C |X − Y |1−n ∀X 6= Y.

Unfortunately, the function Y 7→ |X − Y |1−n is nonintegrable in a neigh-

bourhood of X on every smooth hypersurface containing X.

Exercise 3.2.3 Show that if S is a smooth hypersurface in Rn and X ∈ S,

then the restriction of function Y 7→ |X − Y |1−n to S is nonintegrable with
respect to the surface measure of S.

However, for Y close to X, the vector X − Y is almost orthogonal to ν(Y ),

because ∂Ω is assumed to be of class C 2 . Thus, Y 7→ ∂ν(Y ) N (X − Y ) is, in
fact, much smaller than Y 7→ |X − Y |1−n , as we shall show below. This will
ensure the integrability of the former.

Lemma 3.2.4 There exists a constant C such that

|(X − Y ) · ν(Y )| ≤ C |X − Y |2 ∀X, Y ∈ ∂Ω.

90 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Proof. This is clear if |X − Y | ≥ δ for some δ > 0. Indeed, by Schwarz’s

inequality
|(X − Y )|2
|(X − Y ) · ν(Y )| ≤ |(X − Y )| ≤ .
δ
Thus, we may assume that |X − Y | ≤ δ. Suppose that X is related to Uj
(see the proof of Lemma 3.1.3 for the terminology), and that δ ≤ δ0 (δ0 is as
in (3.1.8)). For the sake of simplicity, in the rest of the proof we shall write Φ
and Ψ instead of Φj and Ψj , and κ instead of κj (see (3.1.5) for the definition
of κj ). Set η := Φ(Y ) and ξ := Φ(X). Moreover, we write Ψ = (ψ1 , . . . , ψn ),
where ψj : U → R are functions of class C 2 . By Taylor’s formula for ψj
1
ψj (ξ) − ψj (η) = ∇ψj (η) · (ξ − η) + hHψj (ζj )(ξ − η), ξ − ηi
2
n−1 n−1
X 1 X 2
= ∂` ψj (η) (ξ` − η` ) + ∂ ψj (ζj ) (ξk − ηk ) (ξ` − η` ),
`=1
2 k,`=1 k,`

where ζj are suitable points on the segment (η, ξ), so that

Ψ(ξ) − Ψ(η)
n−1 n−1
X 1 X 2 2
= ∂` Ψ(η) (ξ` − η` ) + [∂k,` ψ1 (ζ1 ), . . . , ∂k,` ψn (ζn )]T (ξk − ηk ) (ξ` − η` ).
`=1
2 k,`=1

Observe that the vectors ∂1 Ψ(η), . . . , ∂n−1 Ψ(η) are tangent to ∂Ω at the
point Y , whence they are orthogonal to ν(Y ). Therefore
n−1
hX i
∂` Ψ(η) (ξ` − η` ) · ν(Y ) = 0, (3.2.2)
`=1

and
|(X − Y ) · ν(Y )|


= | Ψ(η) − Ψ(0) · ν(Y )|
n−1
1 X 2 2
≤ [∂ ψ1 (ζ1 ), . . . , ∂k,` ψn (ζn )]T · ν(Y ) (ξk − ηk ) (ξ` − η` )
2 k,`=1 k,`
2
≤ C max max sup |∂k,` ψh (ω)| |ξ − η|2
h=1,...,n k,`=1,...,n−1 ω∈κ
2
≤ C max max sup |∂k,` ψh (ω)| C02 |X − Y |2 .
h=1,...,n k,`=1,...,n−1 ω∈κ

We have used (3.1.9) in the last inequality. Since ∂Ω is covered by a finite

number of neighbourhoods Uj , the required estimate follows. 2
3.2. THE DOUBLE LAYER POTENTIAL 91

Notation 3.2.5 The kernel of the double layer potential is denoted

by K. Thus,
1 Y −x
K(x, Y ) := · ν(Y ) ∀Y ∈ ∂Ω ∀x ∈ Rn \ {Y }.
ωn |Y − x|n
Proposition 3.2.6 The kernel K of the double layer potential is a continu-
ous kernel of order n − 2.

Proof. Indeed, by Lemma 3.2.4,

|X − Y |2
|K(X, Y )| ≤ C = C |X − Y |2−n ∀X, Y ∈ ∂Ω, X 6= Y,
|X − Y |n
as required. 2
Next, we establish further properties of the double layer potential. We need
a lemma which describes a special system of coordinates in a neighbourhood
of the boundary of a C 2 domain.

Lemma 3.2.7 Suppose that S is a compact oriented hypersurface of class C 2 .

Then there exists a neighbourhood V of S in Rn and a number ε > 0 such
that the map F : S × (−ε, ε) → V , defined by
F (X, t) = X + tν(X),
is a C 1 diffeomorphism onto V .

Proof. We give only a sketch of the proof. The details are left to the interested
reader.
Clearly F is of class C 1 . Furthermore dF (X, 0) is nonsingular for every
X ∈ S. By the inverse function theorem, F can be inverted in a neighbour-
hood WX of the point (X, 0), yielding a map
FX−1 : WX → (S ∩ WX ) × − ε(X), ε(X)

for some ε(X) S > 0. Since S is compact, we can choose points X1 , . . . , XN

in S such that j WXj covers S, and the maps FX−1j patch together to yield
a C 1 inverse of F from a neighbourhood V of S to S × (−ε, ε), where ε :=
minj εj (X). 2
The neighbourhood V of the lemma above is called a tubular neighbour-
hood of S. It is convenient to extend the definition of normal derivative
from S to the tubular neighbourhood V as follows



∂ν u X + t ν(X) := ∇u X + t ν(X) · ν(X) ∀t ∈ (−ε, ε). (3.2.3)
92 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Proposition 3.2.8 Suppose that Ω is a bounded domain with C 2 boundary.

The kernel K of the double layer potential possesses the following properties:

(i) 
Z 1
 ∀x ∈ Ω
K(x, Y ) dσ(Y ) = 1/2 ∀x ∈ ∂Ω
∂Ω 
0 ∀x ∈ Ωc ;


(ii) Z
sup K(x, Y ) dσ(Y ) < ∞;
x∈Rn \∂Ω ∂Ω

(iii) for every ϕ ∈ C(∂Ω), and every X ∈ ∂Ω

1
lim Dϕ(x) = ϕ(X) + TK ϕ(X),
x→X,x∈Ω 2
and
1
lim Dϕ(x) = − ϕ(X) + TK ϕ(X).
x→X,x∈Ωc 2
2
Here TK denotes the operator on L (∂Ω) whose kernel is K.

Proof. First, we prove (i). Recall that the Newtonian potential with pole x
is harmonic in Rn \ {x}. Suppose that x ∈ Ωc . Then N (x − ·) is harmonic
in a neighbourhood of Ω, and
Z Z
∂ν(Y ) N (x − Y ) dσ(Y ) = ∆Y N (x − Y ) dV (Y ) = 0,
∂Ω Ω

as required.
Now suppose that x ∈ Ω. For ε small, consider the domain
Ωε := Ω \ Bε (x).
Then N (x − ·) is harmonic in a neighbourhood of Ω, and, by a standard
consequence of the divergence theorem
Z Z
∂ν(Y ) N (x − Y ) dσ(Y ) = ∆Y N (x − Y ) dV (Y ) = 0.
∂Ωε Ωε

Observe that
Z
∂ν(Y ) N (x − Y ) dσ(Y )
∂Ωε
Z Z
= ∂ν(Y ) N (x − Y ) dσ(Y ) − ∂ν(Y ) N (x − Y ) dσ(Y ).
∂Ω ∂Bε (x)
3.2. THE DOUBLE LAYER POTENTIAL 93

The last integral is equal to 1 (see Exercise 1.5.3). Hence

Z
∂ν(Y ) N (x − Y ) dσ(Y ) = 1,
∂Ω

as required.
Finally suppose that x ∈ ∂Ω. For ε > 0, we define

∂Bε (x)0 := ∂Bε (x) ∩ Ω

Sε := ∂Ω \ ∂Ω ∩ Bε (x) ,

and
∂Bε (x)00 := {y ∈ ∂Bε (x) : ν(x) · y < 0}.
Since Y 7→ ∂ν(Y ) N (x − Y ) is integrable (see Lemma 3.1.5),
Z Z
∂ν(Y ) N (x − Y ) dσ(Y ) = lim ∂ν(Y ) N (x − Y ) dσ(Y ).
∂Ω ε↓0 Sε

Observe that Y 7→ ∂ν(Y ) N (x−Y ) is harmonic in a neighbourhood of Ω \ Bε (x).

Therefore
Z
0= ∂ν(Y ) N (x − Y ) dσ(Y )
∂(Ω\Bε (x))
Z Z
= ∂ν(Y ) N (x − Y ) dσ(Y ) + ∂ν(Y ) N (x − Y ) dσ(Y ).
Sε ∂Bε (x)0

We then deduce that

Z Z
∂ν(Y ) N (x − Y ) dσ(Y ) = − ∂ν(Y ) N (x − Y ) dσ(Y )
Sε ∂Bε (x)0
1−n
ε
= σ(∂Bε (x)0 ).
ωn
It is straightforward to check that the assumption ∂Ω ∈ C 2 implies that

σ(∂Bε (x)0 ) − σ(∂Bε (x)00 ) = O(εn )

as ε tends to 0. Thus, we see that

ε1−n
Z
∂ν(Y ) N (x − Y ) dσ(Y ) = lim σ(∂Bε (x)0 )
∂Ω ε↓0 ωn
ε1−n
= lim σ(∂Bε (x)00 )
ε↓0 ωn
1
= ,
2
94 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

as required to conclude the proof of (i).

Next we prove (ii). We choose δ ≤ δ0 (δ0 is as in (3.1.8)) so small that

δ < ε, where ε is as in the statement of Lemma 3.2.7. We distinguish between
two cases.
(I) If d(x, ∂Ω) ≥ δ/2, then

1 2n−1 1−n
∂ν(Y ) N (x − Y ) ≤ |x − Y |1−n ≤ δ ,
ωn−1 ωn−1

whence
2n−1 1−n
Z
∂ν(Y ) N (x − Y ) dσ(Y ) ≤ δ σ(∂Ω).
∂Ω ωn−1
(II) Suppose that d(x, ∂Ω) < δ/2. By Lemma 3.2.7, there exists a unique
point X ∈ ∂Ω such that

x = X + |x − X| ν(X). (3.2.4)

We write
Z
∂ν(Y ) N (x − Y ) dσ(Y )
∂ΩZ Z
= ∂ν(Y ) N (x − Y ) dσ(Y ) + ∂ν(Y ) N (x − Y ) dσ(Y ).
∂Ω\Bδ (X) Bδ (X)
(3.2.5)
We denote the integrals on the right hand side by J1 and J2 , respectively,
and estimate them separately.
To estimate the first, observe that if Y ∈
/ (Bδ (X) ∩ ∂Ω), then

|x − Y | ≥ |X − Y | − |X − x| ≥ δ/2,

so that
2n−1 1−n
∂ν(Y ) N (x − Y ) ≤ δ ,
ωn−1
and
2n−1 1−n
Z
∂ν(Y ) N (x − Y ) dσ(Y ) ≤ δ σ(∂Ω).
∂Ω\Bδ (X) ωn−1
Thus, to conclude the proof of (ii), it remains to estimate
Z
∂ν(Y ) N (x − Y ) dσ(Y ). (3.2.6)
Bδ (X)
3.2. THE DOUBLE LAYER POTENTIAL 95

Notice that, by triangle’s inequality, Schwarz’s inequality, and Lemma 3.2.4,

|(x − X) · ν(Y )| |(Y − X) · ν(Y )|

ωn ∂ν(Y ) N (x − Y ) ≤ +
|x − Y |n |x − Y |n
(3.2.7)
|x − X| |Y − X|2
≤ + C ,
|x − Y |n |x − Y |n

where C does not depend on Y and X in ∂Ω and of x ∈ Rn \ ∂Ω.

We claim that if δ is small enough (see (3.2.9) below) then

1
|x − Y |2 ≥ |Y − X|2 + |x − X|2 .

(3.2.8)
2
Given the claim, we show how to conclude the proof of (ii). The claim and
(3.2.7) imply that there exists a constant C such that

ωn ∂ν(Y ) N (x − Y )
|x − X| |Y − X|2
≤C n/2 + C  n/2
|Y − X|2 + |x − X|2 |Y − X|2 + |x − X|2
|x − X| 2−n
≤C  n/2 + C |Y − X| .
|Y − X|2 + |x − X|2

Therefore the integral in (3.2.6) may be estimated by

|x − X|
Z Z
C  n/2 dσ(Y )+C |Y − X|2−n dσ(Y ).
Bδ (X)∩∂Ω |Y − X|2 + |x − X|2 Bδ (X)∂Ω

The second of the integrals above may be estimated from above by C1 δ, by

Lemma 3.1.5. To estimate the first, suppose that X is related to Uj , where
(Uj , Φj ) is one of the local charts described in Lemma 3.1.3 (see its proof
for the terminology). Write ξ = Φj (X) and η = Φj (Y ), and a := |x − X|.
Observe that
|x − X|
Z
 n/2 dσ(Y )
Bδ (X)∩∂Ω |Y − X|2 + |x − X|2
Z
a
q
≤  n/2 1 + |∇fj (η)|2 dη
Bδ (Φj (X)) |η − ξ|2 + a2
Z δ
a rn−2
≤ C0 ωn−1  n/2 dη,
0 r 2 + a2
96 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

where C0 is defined in (3.1.6). The change of variables v = r/a transforms

the last integral into
Z δ/a
v n−2
dv,
0 (1 + v 2 )n/2
which is dominated by the convergent integral
Z ∞
v n−2
dv,
0 (1 + v 2 )n/2
which does not depend on a.
Therefore the integral in (3.2.6) is finite, as required.
It remains to prove the claim (3.2.8). Observe that the vector (∇fj (η), −1)
is orthogonal to ∂Ω at the point Ψj (η). Write

x − Y = X − Y + |x − X| ν(X)

(∇fj (η), −1)
= ξ − η, fj (ξ) − fj (η) + p |x − X|
1 + |∇fj (ξ)|2

Then

2
|x − Y |2 = (ξ − η, fj (ξ) − fj (η)
2

+p |x − X| ∇fj (η)(ξ − η) − fj (ξ) + fj (η)
1 + |∇fj (ξ)|2
+ |x − X|2

Now,
2
≥ |ξ − η|2 ,


(ξ − η, fj (ξ) − fj (η)
and, by Taylor’s formula,

≤ C |ξ − η|2 .


∇fj (η)(ξ − η) − fj (ξ) + fj (η)

Therefore
2

p |x − X| ∇fj (η)(ξ − η) − fj (ξ) + fj (η)
1 + |∇fj (ξ)|2
≤ |x − X| sup |||Hfj (ω)||| |ξ − η|2 .
ω∈κj

We may choose δ so that

δ 1
max sup |||Hfj (ω)||| ≤ . (3.2.9)
2 j=1,...,N ω∈κj 2
3.2. THE DOUBLE LAYER POTENTIAL 97

Recall that we are assuming that |x − X| < δ/2. Therefore

1 1
|x − Y |2 ≥ |ξ − η|2 − |ξ − η|2 + |x − X|2 ≥ |ξ − η|2 + |x − X|2 ,

2 2
and the claim is proved. This concludes the proof of (ii).
Finally, we prove (iii). We shall prove the first formula. The proof of the
second is similar and is omitted. Fix X ∈ ∂Ω. Since
Z
K(x, Y ) dσ(Y ) = 1 ∀x ∈ Ω,
∂Ω

we have, for each x ∈ Ω,

Z
 
Dϕ(x) = K(x, Y ) ϕ(Y ) − ϕ(X) dσ(Y ) + ϕ(X)
Z∂Ω
  
= K(x, Y ) − K(X, Y ) ϕ(Y ) − ϕ(X) dσ(Y )
∂Ω Z
 
+ K(X, Y ) ϕ(Y ) − ϕ(X) dσ(Y ) + ϕ(X)
Z ∂Ω
  
= K(x, Y ) − K(X, Y ) ϕ(Y ) − ϕ(X) dσ(Y )
∂Ω
1
+ ϕ(X) + TK ϕ(X).
2
Thus, to conclude the proof of (iii), it remains to show that
Z
  
lim K(x, Y ) − K(X, Y ) ϕ(Y ) − ϕ(X) dσ(Y ) = 0. (3.2.10)
x→X,x∈Ω ∂Ω

We denote by J the last integral. Since, by assumption, ϕ is continuous at

X, for every ε > 0 there exists δ > 0 such that
ϕ(Y ) − ϕ(X) < ε ∀Y ∈ B2δ (X) ∩ ∂Ω.
Then
Z Z
|J | ≤ ε |K(x, Y )| dσ(Y ) + ε |K(X, Y )| dσ(Y )
B2δ (X)∩∂Ω B2δ (X)∩∂Ω
Z
+ 2 kϕk∞ K(x, Y ) − K(X, Y ) dσ(Y ).
∂Ω\B2δ (X)

By the Lebesgue dominated convergence theorem

Z
lim K(x, Y ) − K(X, Y ) dσ(Y ) = 0.
x→X,x∈Ω ∂Ω\B2δ (X)
98 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Therefore
Z
lim |J | ≤ 2ε sup |K(x, Y )| dσ(Y ) = C ε.
x→X,x∈Ω x∈Rn ∂Ω

We have used (ii) in the last equality. By taking the infimum of both sides
with respect to ε > 0, we obtain that

lim |J | = 0,
x→X,x∈Ω

as required to conclude the proof of (iii), and of the proposition. 2

3.3 The single layer potential

Suppose that Ω is a bounded domain with C 2 boundary. In this section we
study some properties of the single layer potential.

Definition 3.3.1 The single layer potential with density g is the func-
tion Z
Sg(x) := g(Y ) N (x − Y ) dσ(Y ) ∀x ∈ Rn \ ∂Ω.
∂Ω

The function (x, Y ) 7→ N (x − Y ) is called the kernel of the single layer

potential.

Lemma 3.3.2 If g ∈ C(∂Ω), then Sg is harmonic in Rn \ ∂Ω.

Proof. The required properties follow from the remark above by differentiat-
ing under the integral sign. The reader should check that differentiation is
permitted. 2

Recall that ∂Ω is assumed to be of class C 2 . Hence there is a tubular neigh-

bourhood V of ∂Ω, where the operator ∂ν is well defined (see (3.2.3)). In
particular, given ϕ ∈ C(∂Ω), the single layer potential Sϕ, generated by ϕ,
is harmonic in V \ ∂Ω by Lemma 3.3.2. Thus, we may compute ∂ν (Sϕ)(x)
for every x ∈ V \ ∂Ω (see (3.2.3) for the definition of ∂ν ), and obtain

x−Y
Z
1
∂ν (Sϕ)(x) = · ν(X) ϕ(Y ) dσ(Y ) ∀x ∈ V \ ∂Ω. (3.3.1)
ωn ∂Ω |x − Y |n
3.3. THE SINGLE LAYER POTENTIAL 99

Definition 3.3.3 Define K ∗ : ∂Ω × ∂Ω \ {(X, X) : X ∈ ∂Ω} → R by

1 X −Y
K ∗ (X, Y ) := · ν(X).
ωn |X − Y |n

Notice that if X and Y are in ∂Ω with X 6= Y , then

K ∗ (X, Y ) = K(Y, X). (3.3.2)

Remark 3.3.4 Clearly, K ∗ is a continuous kernel of type n−2, because K is.

Observe that the operator TK ∗ is the adjoint of TK , when acting on L2 (∂Ω).
Indeed, suppose that ϕ and ψ are in L2 (∂Ω). Then
Z


TK ϕ, ψ = TK ϕ(X) ψ(X) dσ(X)
Z∂Ω Z
= dσ(X) ψ(X) K(X, Y ) ϕ(Y ) dσ(Y ).
∂Ω ∂Ω

Since K is a kernel of type n − 2, the operator TK is bounded on L2 (∂Ω) by

Proposition 3.1.6 (i). Hence
Z Z
|ϕ(Y )| |K(X, Y )| |ψ(X)| dσ(Y ) dσ(X) ≤ |||TK |||L2 (∂Ω) kϕk2 kψk2 .
∂Ω ∂Ω

Thus, we may apply Fubini’s theorem, and conclude that

Z Z


TK ϕ, ψ = dσ(Y ) ϕ(Y ) K(X, Y ) ψ(X) dσ(X)
∂Ω ∂Ω
Z Z
(K is real) = dσ(Y ) ϕ(Y ) K ∗ (Y, X) ψ(X) dσ(X)
∂Ω ∂Ω


= ϕ, TK ∗ ψ ,
as required.

For each X ∈ ∂Ω we define

∂ν − (Sϕ)(X) := lim− ∂ν (Sϕ)(X + tν(X)) (3.3.3)
t→0

and
∂ν + (Sϕ)(X) := lim+ ∂ν (Sϕ)(X + tν(X)), (3.3.4)
t→0
whenever the limits above exist.
Some of the main properties of the single layer potential are summarised in
the next proposition.
100 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Proposition 3.3.5 Suppose that ∂Ω is of class C 2 , and that ϕ ∈ C(∂Ω).

The following hold:

(i) Sϕ is continuous in Rn ;

(ii) ∂ν − (Sϕ)(X) and ∂ν + (Sϕ)(X) exist for each X ∈ ∂Ω, and

1
∂ν − (Sϕ)(X) = − ϕ(X) + TK ∗ ϕ(X) (3.3.5)
2
and
1
∂ν + (Sϕ)(X) = ϕ(X) + TK ∗ ϕ(X). (3.3.6)
2
Furthermore, the convergence of ∂ν (Sϕ)(X + tν(X)) to ∂ν − (Sϕ)(X)
as t tends to 0− and of ∂ν (Sϕ)(X + tν(X)) to ∂ν + (Sϕ)(X) as t tends
to 0+ is uniform on ∂Ω.

Proof. First we prove (i). Clearly Sϕ is continuous on Rn \ ∂Ω. It remains

to prove that Sϕ is continuous at each X ∈ ∂Ω. Suppose that δ > 0 is small,
and that d(x, X) ≤ δ/10. Then d(x, ∂Ω) ≤ δ/10. Write

u(x) − u(X)
Z
1
≤ |x − Y |2−n |ϕ(Y )| dσ(Y )
(n − 2) ωn Bδ (X)∩∂Ω
Z
1
+ |X − Y |2−n |ϕ(Y )| dσ(Y )
(n − 2) ωn Bδ (X)∩∂Ω
Z
1
+ |X − Y |2−n − |X − Y |2−n |ϕ(Y )| dσ(Y ).
(n − 2) ωn ∂Ω\Bδ (X)

Denote by J1 , J2 and J3 the integrals above.

By Lemma 3.1.5,
|J2 | ≤ C1 kf k∞ δ, (3.3.7)
with C independent of x and X. Furthermore

lim J3 = 0, (3.3.8)
x→X,x∈Ω

by the Lebesgue dominated convergence theorem.

To estimate J1 we argue as follows. Since δ is small, there exists a unique
point ζ ∈ ∂Ω such that
x = ζ − |x − ζ| ν(ζ).
3.3. THE SINGLE LAYER POTENTIAL 101

Then
x − Y = ζ − Y − |x − ζ| ν(ζ).
We distinguish between two cases.


(I) If Y ∈ Bδ (X) ∩ ∂Ω \ Bδ/5 (ζ), then

|Y − x| ≥ |Y − ζ| − |ζ − x| ≥ δ/5 − δ/10 = δ/10.

Thus,
Z
|x − Y |2−n |ϕ(Y )| dσ(Y ) ≤ C kϕk∞ δ 2−n σ Bδ (X) ∩ ∂Ω


Bδ (X)∩∂Ω

≤ C kϕk∞ δ 2−n δ n−1

≤ C kϕk∞ δ.

(II) If Y ∈ Bδ/5 (ζ) ∩ ∂Ω and δ is small enough, then, by (3.2.8) with ζ in

place of X, there exists a constant c > 0 such that

|x − Y | ≥ c |ζ − Y |.

Therefore
Z Z
2−n
|x − Y | |ϕ(Y )| dσ(Y ) ≤ C kϕk∞ |ζ − Y |2−n dσ(Y )
Bδ/5 (ζ)∩∂Ω Bδ/5 (ζ)∩∂Ω

≤ C kϕk∞ δ.

Thus,
|J3 | ≤ C kf k∞ δ. (3.3.9)
By combining (3.3.9), (3.3.7), and (3.3.8), we see that for all δ > 0 small
enough
lim |u(x) − u(X)| ≤ C kf k∞ δ.
x→X,x∈Ω

Thus, limx→X,x∈Ω |u(x) − u(X)| = 0, as required to conclude the proof of (i).

Next we prove the first formula in (ii). The proof of the second formula
is similar and is omitted. Set
1
Et (X) := ∂ν (Sϕ)(X + tν(X)) + ϕ(X) − TK ∗ ϕ(X).
2
We must show that
lim sup Et (X) = 0. (3.3.10)
t→0− X∈∂Ω
102 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

By Proposition 3.2.8 (i) and (ii),

1
ϕ(X)
2
X −Y X + t ν(X) − Y i
Z h
1
= − · ν(Y ) ϕ(X) dσ(Y )
ωn ∂Ω |X − Y |n |X + t ν(X) − Y |n
X −Y X + t ν(X) − Y i
Z h
1
= − · ν(Y ) [ϕ(X) − ϕ(Y )] dσ(Y )
ωn ∂Ω |X − Y |n |X + t ν(X) − Y |n
X −Y X + t ν(X) − Y i
Z h
1
+ − · ν(Y ) ϕ(Y ) dσ(Y )
ωn ∂Ω |X − Y |n |X + t ν(X) − Y |n
(3.3.11)
Furthermore
X −Y
Z
1
TK ∗ ϕ(X) = · ν(X) ϕ(Y ) dσ(Y ) (3.3.12)
ωn ∂Ω |X − Y |n

and
X + t ν(X) − Y
Z
1
∂ν (Sϕ)(X + tν(X)) = · ν(X) ϕ(Y ) dσ(Y )
ωn ∂Ω |X + t ν(X) − Y |n
(3.3.13)
Therefore

Et (X)
X −Y X + t ν(X) − Y i
Z h
1
= − · ν(Y ) [ϕ(X) − ϕ(Y )] dσ(Y )
ωn ∂Ω |X − Y |n |X + t ν(X) − Y |n
X −Y X + t ν(X) − Y i
Z h
1
− − · [ν(X) − ν(Y )] ϕ(Y ) dσ(Y )
ωn ∂Ω |X − Y |n |X + t ν(X) − Y |n
(3.3.14)
t 2
We denote the two integrals above by It (X) and It (X), respectively.
Fix ε > 0, and choose δ > 0 such that

|ν(X) − ν(Y )| < ε and |ϕ(X) − ϕ(Y )| < ε (3.3.15)

for all X and Y in ∂Ω such that |X − Y | < 2δ. This is possible for both ν
and ϕ are continuous, hence uniformly continuous, functions on the compact
set ∂Ω.
We estimate It1 (X). We write
Z Z
1
It (X) = + .
B2δ (X)∩∂Ω ∂Ω\B2δ (X)
3.4. SOLVABILITY 103

Observe that
Z Z
 
≤ ε ωn K(X, Y ) + K(X + tν(X), Y ) dσ(Y )
B2δ (X)∩∂Ω ∂Ω
Z (3.3.16)
≤ 2 ε ωn sup K(x, Y ) dσ(Y ),
x∈Rn ∂Ω

where K(x, Y ) is the kernel of the double layer potential. Recall that, by
Proposition 3.2.8 (ii), the supremum above is finite.
Next, we apply the mean value theorem, and obtain that
X −Y X + t ν(X) − Y
sup n
− n
≤ C |t| δ −n ∀t ∈ (−δ, 0),
Y ∈∂Ω\B2δ (X) |X − Y | |X + t ν(X) − Y |

where C is a constant, which depends only on the dimension n. Thus,

Z
≤ 2 C |t| δ −n kϕk∞ σ(∂Ω). (3.3.17)
∂Ω\B2δ (X)

By combining (3.3.16) and (3.3.17), we obtain that for every ε > 0 there
exists δ > 0 such that

sup It1 (X) ≤ C ε + |t| δ −n ,

X∈∂Ω

whence
lim sup It1 (X) ≤ C ε.
t→0− X∈∂Ω

Therefore
lim sup It1 (X) = 0.
t→0− X∈∂Ω

A similar reasoning shows that

lim sup It2 (X) = 0.

t→0− X∈∂Ω

The last two formulae imply (3.3.10), as required to conclude the proof of
(ii), and of the proposition. 2

3.4 Solvability of the Dirichlet problem

The purpose of this section is to prove Theorem 3.4.2 below. We shall need
the following lemma.
104 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

Lemma 3.4.1 The operator (1/2) I + TK ∗ is injective.

Proof. Observe preliminarily that, since ∂Ω is a compact oriented hypersur-

face of class C 2 , we may apply Lemma 3.2.7, which ensures the existence of
a neighbourhood V of ∂Ω and a C 1 diffeomorphism F of ∂Ω × (−ε, ε) onto
V , defined by
F (X, t) = X + tν(X) ∀X ∈ ∂Ω ∀t ∈ (−ε, ε),
where ν(X) denotes the exterior unit normal to Ω. It is straightforward to
check that for each t ∈ (−ε, ε) the set of all points of the form
X + tν(X) ∀X ∈ ∂Ω
is a compact hypersurface, denoted by St , which is contained in Ωc for all
t ∈ (0, ε). Denote by Ωt the domain
Ω ∪ {ζ ∈ Rn : ζ = X + sν(X), with X ∈ ∂Ω and s ∈ [0, t)}.
The boundary of Ωt is St .
Since Ωt is bounded, Ωt is contained in BR (0) for R large enough. For
such R, consider the bounded domain
Ωt,R := BR (0) \ Ωt .


Suppose that f belongs to Ker (1/2) I + TK ∗ . We must show that f = 0.
Observe that f is continuous on ∂Ω, by Proposition 3.1.7 (ii). We form the
single layer potential Sf of f . Since Sf is harmonic on Rn \ ∂Ω,
Z
2
I:= ∇(Sf )(x) dV (x)
n
R \Ω
Z
2
= lim ∇(Sf )(x) dV (x)
t↓0,R↑∞ Ωt,R

by the Lebesgue monotone convergence theorem.

Now, Sf is harmonic in a neighbourhood of Ωt,R , hence we may apply
the first Green’s identity, and obtain that
Z
I = lim Sf (X) ∂ν(X) (Sf )(X) dσ(X) (3.4.1)
t↓0,R↑∞ ∂Ωt,R

The last integral is the sum of

Z
Jt := − Sf (X) ∂ν(X) (Sf )(X) dσ(X)
∂St
3.4. SOLVABILITY 105

and Z
HR := Sf (X) ∂ν(X) (Sf )(X) dσ(X).
∂BR (0)

We claim that limR↑∞ HR = 0. Indeed, for all x such that d(x, ∂Ω) ≥ |x|/2
kf k∞
Z
|Sf (x)| ≤ |x − Y |2−n dσ(Y )
(n − 2) ωn ∂Ω
kf k∞
≤ d(x, ∂Ω)2−n σ(∂Ω)
(n − 2) ωn
kf k∞
≤ 2n−2 |x|2−n σ(∂Ω).
(n − 2) ωn
Similarly, for all x such that d(x, ∂Ω) ≥ |x|/2
kf k∞
Z
|∂ν(X) Sf (x)| ≤ |x − Y |1−n dσ(Y )
ωn ∂Ω
kf k ∞
≤ 2n−2 |x|1−n σ(∂Ω).
ωn
Thus,
kf k2∞
|HR | ≤ 22(n−2) R3−2n σ(∂Ω)2 ωn Rn−1
(n − 2) ωn2
kf k2∞
= 22(n−2) R2−n σ(∂Ω)2 ,
(n − 2) ωn
which tends to 0 as R tends to infinity (we are assuming n ≥ 3), as claimed.
We claim that limt↓0 Jt = 0. This follows from Proposition 3.3.5 (i)-(ii).
We leave the verification of this fact to the reader.
The two claims above and (3.4.1) imply that I = 0. Therefore |∇(Sf )| =
0 a.e. on Rn \ Ω, hence |∇(Sf )| = 0 everywhere on Rn \ Ω, for |∇(Sf )| is
continuous on Rn \ Ω. Since Rn \ Ω is connected, Sf is constant on Rn \ Ω.
We have already proved that lim|x|→∞ Sf (x) = 0, whence Sf = 0 on Rn \ Ω.
By (i), Sf is continuous on Rn , so that Sf = 0 on ∂Ω. By the maximum
principle for Sf on Ω, it follows that Sf = 0 on Rn . Thus,
∂ν − (Sf ) = 0 on ∂Ω.
By (ii),
1
∂ν − (Sf )(X) = − f (X) + TK ∗ f (X) ∀X ∈ ∂Ω.
2
Therefore
1  1
f (X) = f (X) + TK ∗ f (X) − − f (X) + TK ∗ f (X)] = 0,
2 2
106 CHAPTER 3. DIRICHLET VIA INTEGRAL EQUATIONS

1
thereby proving the injectivity of 2
I + TK ∗ , and concluding the proof of the
lemma. 2

Recall that K denotes the kernel of the double layer potential.

Theorem 3.4.2 Suppose that Ω is a bounded domain with ∂Ω of class C 2 .

For every g ∈ C(∂Ω) there exists a unique function u ∈ C(Ω) which solves
the Dirichlet problem (
∆u = 0 in Ω
(3.4.2)
u|∂Ω = g.
Furthermore, the solution u is given by
Z
u(x) = K(x, Y ) ϕ(Y ) dσ(Y ) ∀x ∈ Ω,
∂Ω

where ϕ is the unique solution to the integral equation

Z
1
ϕ(X) + K(X, Y ) ϕ(Y ) dσ(y) = g(X) ∀X ∈ ∂Ω.
2 ∂Ω

Proof. Recall that the operator TK is compact, because K is a continuous

kernel of order n − 2 (see Proposition 3.2.6 and Proposition 3.1.6).
By Lemma 3.4.1 the operator (1/2) I + TK is injective (hence bijective by
Theorem 2.6.4 (iv)). Therefore the equation (1/2)ϕ + TK ϕ = g has exactly
one solution for every g ∈ L2 (∂Ω). Since g is continuous by assumption, the
solution ϕ is also continuous, by Proposition 3.1.7 (ii). By Lemma 3.2.2,
the double layer potential Dϕ generated by ϕ is harmonic in Ω and it is
continuous on Ω by Proposition 3.2.8 (iii). 2
 Part III

The Dirichlet problem

via L2 methods

107
 109

4.0.1 Introduction to Dirichlet’s principle

We have already seen how to prove the existence of a classical solution to the
Dirichlet problem (
∆u = 0 in Ω
(4.0.1)
u|∂Ω = g
via Perron’s method and via integral equations on the boundary ∂Ω. Here we
study a different approach to the problem, based on the so-called Dirichlet
principle. The latter is based on an idea of wide utility in mathematics and
physics: to find the equilibrium state of a system one seeks to minimize an
appropriate energy or action.
In the present case the role of this energy is played by the Dirichlet
integral. Given a continuous function g on the boundary ∂Ω of Ω, denote
by Yg the affine space of all functions ϕ ∈ C 2 (Ω) with ϕ|∂Ω = g. Clearly
Y0 is a vector space. The Dirichlet principle states that the solution
to the Dirichlet problem above agrees with the solution to the following
minimization problem
min D(u), (4.0.2)
u∈Yg

where D(u) denotes the Dirichlet integral, defined by

Z
2
D(u) := ∇u dV ; (4.0.3)
Ω

as usual, dV denotes the volume element in Rn .

There are two basic questions concerning Dirichlet’s principle:

(i) does a solution to the minimization problem (4.0.2) solve the Dirichlet
problem (4.0.1)?

(ii) does the solution to the the Dirichlet problem (4.0.1) solve the mini-
mization problem (4.0.2)?

First we discuss the first question. The application of Dirichlet’s principle

was thought to have been justified by the following proposition, whose proof
we postpone to the next section.

Proposition 4.0.1 (Dirichlet’s principle I) Suppose that Ω is a bounded

domain with Lipschitz boundary and that g ∈ C(∂Ω), that Yg is nonempty
and that u ∈ Yg . The following are equivalent:
110

(i) ∆u = 0 in Ω (in the classical sense);

(ii) u is a critical point of the functional D in the sense that

d
D(u + εϕ)|ε=0 = 0 ∀ϕ ∈ Y0 ;
dε

(iii) u minimizes D in the sense that

D(u) ≤ D(w) ∀w ∈ Yg .

In fact, the proposition above does not establish the existence of a solution to
the Dirichlet problem (4.0.1). Rather, it converts (4.0.1) into the problem of
minimizing the Dirichlet integral, under the assumption that g is the “trace”
on the boundary of some functions in C 2 (Ω).
In order to avoid interruptions in the flow of information, we postpone
all the proofs of the results stated in this introduction to sections below (see
Subsection 4.0.3 for details).

Weierstrass pointed out that it was not clear (and had not been proved)
that a minimizing function for the Dirichlet integral exists, so there might
simply be no winner to the implied competition in Proposition 4.0.1. He ar-
gued by analogy with a simpler one-dimensional problem: that of minimizing
the integral Z 1
2
I(ϕ) := xϕ0 (x) dx
−1
1
amongst all C functions ϕ on [−1, 1] that satisfy ϕ(−1) = −1 and ϕ(1) = 1.

Exercise 4.0.2 Prove that the infimum of I is equal to 0. Prove that the
infimum is not a minimum.

This suggests the possibility that the functional D may reach its infimum in
a set of competitors larger than Yg , but possibly not in Yg , and demands a
proof that a minimizer, in fact, exists. After all, the problem

min {x2 : x ∈ Q, x2 ≥ 2}

has no solutions, but it has indeed solutions in R, the completion of Q with

respect to the Euclidean distance.
To develop this line of reasoning further, let us provisionally assume
that Ω is a bounded domain with Lipschitz boundary. This is a tech-
nical assumption, which guarantees that the Sobolev space H 1 (Ω) may be
 111

defined as the completion of C 2 (Ω) with respect to a suitable norm. This is

no longer true for more general domains. Observe that
Z Z 1/2
2 2
kϕkH 1 := |ϕ| dV + ∇ϕ dV (4.0.4)
Ω Ω

is a norm on C 2 (Ω), where the latter is the space of all functions ϕ in C 2 (Ω)
such that Dα ϕ admits a continuous extension to Ω for each multiindex α
with |α| ≤ 2.

Definition 4.0.3 Assume that Ω is a bounded domain and that ∂Ω is

Lipschitz. Define H 1 (Ω) to be the completion of C 2 (Ω) with respect to the
norm (4.0.4) above.

It is tempting to enlarge the set of competitors in the minimizing problem

(4.0.2) to the affine space of all u ∈ H 1 (Ω) such that u|∂Ω = g, but there
is a problem here: what is the precise meaning of the relation u|∂Ω = g?
In particular, is it true that C(Ω) is included in H 1 (Ω), so that the relation
above may be intended in the obvious sense? The answer the second question
is positive if n = 1, and negative if n ≥ 2 (see Exercise 5.5.15 for the case
n = 2). The following result gives a partial answer to the first question.

Proposition 4.0.4 Suppose that Ω is a bounded domain with Lipschitz bound-

ary. The restriction map u 7→ u|∂Ω from C 2 (Ω) to C(∂Ω) extends to a bounded
operator γ from H 1 (Ω) to L2 (∂Ω).

The operator γ is called the trace operator. It is important to observe

that γ is not surjective, as a celebrated counterexample of Weierstrass shows
(see Proposition 4.0.9 below). In fact, there are many continuous functions
on ∂Ω that do not belong to γ(H 1 (Ω)). Loosely speaking, a function on ∂Ω
is in γ(H 1 (Ω)) if and only if it possesses “half derivative”, i.e., it belongs to
a certain Sobolev space, often denoted by H 1/2 (Ω). It is beyond the scope
of these notes to develop the theory of such space. We refer the interested
reader to [AF].
Now we may reformulate the minimization problem (4.0.2) as follows. Given
a function g in γ(H 1 (Ω)), find w ∈ H 1 (Ω) such that
D(w) = min {D(u) : u ∈ H 1 (Ω), γ(u) = g}. (4.0.5)
For technical reasons, it may be more convenient to give a slightly different
reformulation of the Dirichlet principle. To proceed further, we need to
introduce another function space, which will play a key role in the sequel.
112

Definition 4.0.5 We define H01 (Ω) to be the completion of Cc∞ (Ω) with
respect to the Sobolev norm k·kH 1 (Ω) .

Since Cc∞ (Ω) is a subspace of C 2 (Ω), the space H01 (Ω) is a (closed) subspace
of H 1 (Ω). Since the restriction of the trace operator γ to Cc∞ (Ω) vanishes, the
restriction of γ to H01 (Ω) vanishes as well. Thus, we may say that elements
of H01 (Ω) have “vanishing boundary values”. The converse statement is also
true. For the proof, see [AF, Thm 5.37].

Proposition 4.0.6 Suppose that Ω is a bounded domain with Lipschitz bound-

ary. If u ∈ H 1 (Ω) and γ(u) = 0, then u ∈ H01 (Ω).

Now we give another formulation of the minimization problem (4.0.5).

Given a function g in γ(H 1 (Ω)), denote by G any function in H 1 (Ω) such
that γ(G) = g. The problem consists in finding w ∈ G + H01 (Ω) such that

D(w) = min {D(u) : u ∈ H 1 (Ω), u − G ∈ H01 (Ω)}

= min1 D(u). (4.0.6)
u∈G+H0 (Ω)

It is not hard to prove the following result, whose proof we postpone to

Section 6.2.

Theorem 4.0.7 Suppose that Ω is a bounded domain with Lipschitz bound-

ary. Then the minimization problem (4.0.6) has a unique solution, which is
harmonic in Ω.

The following result states that, under some mild conditions, the classical
solution and the solution to the minimization problem (4.0.6) agree.

Theorem 4.0.8 Suppose that Ω is a bounded domain with Lipschitz bound-

ary. Suppose that g ∈ C(∂Ω), and that there exists G ∈ H 1 (Ω) such that
γ(G) = g. Then the solution to the minimization problem (4.0.6) is a classi-
cal solution.

We shall not prove this result. Its proof is a consequence of some deep
estimates of the oscillation of solutions on balls close to the boundary ∂Ω.
The interested reader is referred to [GT, Ch. 8], especially Corollary 8.28 and
the comments at the beginning of p. 206.
Now we discuss briefly the second question posed at the beginning of
this subsection. We show that the assumption that the boundary datum g
 113

is the trace of a function in H 1 (Ω) is necessary in Theorem 4.0.8. Indeed,

Weierstrass proved that solutions to the classical Dirichlet problem need not
have finite Dirichlet integral.

Proposition 4.0.9 There exists a continuous function g on the boundary

of the unit disc B1 (0) in R2 such that the classical solution to the Dirichlet
problem (
∆u = 0 in B1 (0)
u|∂B1 (0) = g
has infinite Dirichlet integral.

Thus, we cannot expect to solve all classical Dirichlet problems by

L2 methods. However, L2 methods are quite flexible and may be adapted to
a number of different problems, and they are most efficient in a great variety
of situations.

4.0.2 Why weak solutions?

The discussion in Subsection 4.0.1 emphasises the role that the Sobolev
spaces H 1 (Ω) and H01 (Ω) play in the approach to the Dirichlet problem
via Dirichlet’s principle. It is straightforward to check that the following
continuous inclusions hold
H01 (Ω) ,→ H 1 (Ω) ,→ L2 (Ω). (4.0.7)
Thus, elements of H 1 (Ω) and H01 (Ω) may be regarded as functions in L2 (Ω).
It would be desirable to have simple and direct characterisations of those
functions in L2 (Ω) that belong to each of these spaces. This will lead to the
notion of weak derivatives of a function in L2 (Ω) (see Definition 5.5.5).
Weak derivatives, in turn, are special instances of distributional deriva-
tives (see Definition 5.3.1), and are best understood in the framework of the
theory of distributions.
There is a further important reason to introduce weak (hence distribu-
tional) derivatives. We shall prove that solving the minimization problem
(4.0.6) is equivalent to finding a weak solution of the Laplace equation (see
Definition 6.4.1), which, in turn, is, via Weyl’s lemma (see Theorem ??),
equivalent to finding a distributional solution (see Definition 5.4.8) of the
Laplace equation. The definition of weak solution suggests that functions
in L2 (Ω) with first order derivatives (in the sense of distributions) in L2 (Ω)
may be an interesting space.
114

In conclusion, there are good reasons to introduce the so-called weak deriva-
tives of L2 functions.

4.0.3 Plan of Part III

Here is an outline of Part III.

In Section 4.1 we prove some of the statements made in Subsection 4.0.1,
specifically Proposition 4.0.1, the estimate for traces (Proposition 4.0.6), and
Weierstrass counterexample (Proposition 4.0.9).
Distributions, weak derivatives and Sobolev spaces are treated in Chap-
ter 5. In particular, in Section 5.1 we shall look at complex measures
on Rn and regard them as linear functionals on Cc (Rn ) that are continuous
with respect to a certain topology τSI . Then we do the same for complex
measures on a domain Ω in Rn .
In Section 5.2 we modify the arguments of Section 5.1 to study the dual
∞
of the space Cc∞ (Rn ) of test functions, endowed with the topology τSI , which
n
generalises the topology τSI on Cc (R ) studied before. The dual of test
functions is the space of distributions.
It turns out that a distribution has derivatives of all orders: derivatives
of a distribution are defined in Section 5.3, and applications to the solution
of Poisson equation ∆u = f , where f is a given distribution, are discussed
in Section 5.4.
In Section 5.5, we take up the study of Sobolev spaces. First we illus-
trate the subtleties that may arse when we try to substantiate the intent of
defining derivatives in the L2 sense. Then we define the Sobolev spaces H 1 (Ω)
and H01 (Ω) and state their properties. The proofs thereof are quite lengthy,
and would occupy a large portion of the course. Since Sobolev spaces are
the subject of many books or chapters of treatises, we have made the choice
of stating their properties without proofs, for which the reader is referred to
[AF, Br, EG, GT, Fo2, Tr].
In Chapter 6 we develop the L2 theory for the Dirichlet problem. In
particular, we study weak solutions of the Dirichlet problem, and prove The-
orems 4.0.7 and 4.0.8.
4.1. DETAILS LEFT BEHIND 115

4.1 Some details left behind

In this section we prove some of the results stated without proof in the
introduction above.
Proof of Proposition 4.0.1. For every ϕ in Y0
Z
D(u + εϕ) = ∇(u + εϕ) · ∇(u + εϕ) dV
Ω Z
∇ϕ · ∇u + ∇u · ∇ϕ dV + ε2 D(ϕ).


= D(u) + ε
Ω

Hence
d
D(u + εϕ)|ε=0
dε Z


= ∇ϕ · ∇u + ∇u · ∇ϕ dV
ΩZ Z



=− ϕ div(∇u) + div(∇u) ϕ dV + ϕ ∂n u + (∂n u) ϕ dσ
ZΩ ∂Ω


=− ϕ div(∇u) + div(∇u) ϕ dV :
Ω

the penultimate equality is obtained by integrating by parts, and the last

equality follows from the fact that ϕ vanishes at the boundary ∂Ω. We now
rewrite this formula in the cases where ϕ is real valued and where ϕ is purely
imaginary, ϕ = iψ with ψ real valued say. We obtain the two formulae
Z
d
D(u + εϕ)|ε=0 = −2 ϕ ∆(Re u) dV, ∀ϕ ∈ Re Y0
dε Ω

and Z
d
D(u + iεψ)|ε=0 = 2i ψ ∆(Im u) dV ∀ψ ∈ Re Y0 .
dε Ω

Now, if (i) holds, then ∆(Re u) = 0 = ∆(Im u), and the two formulae above
imply that u is a critical point of D, i.e. (ii) holds.
Conversely, if (ii) holds, then

∆(Re u) = 0 and ∆(Im u) = 0

a.e., hence in the classical sense, for u is of class C 2 (Ω). Therefore (i) holds.
116

Next, trivially (iii) implies (ii). To conclude the proof of the proposition,
we need to show that (ii) implies (iii). Observe that

D(w) = D(u + w − u)
Z


= D(u) + ∇u · ∇(w − u) + ∇(w − u) · ∇u dV + D(w − u).
Ω
(4.1.1)
The middle term of the right hand side vanishes. Indeed, and by integrating
by parts Z


∇u · ∇(w − u) + ∇(w − u) · ∇u dV
Ω Z


=− ∆u (w − u) + (w − u) ∆u dV
Ω
= 0;
note that there are no boundary terms arising from the integration by parts,
because w − u is in Y0 , hence it vanishes on the boundary of Ω. Furthermore,
we already know that (ii) is equivalent to (i), whence ∆u = 0, thereby
justifying the last equality above.
Therefore, by (4.1.1),

D(w) = D(u) + D(w − u) ≥ D(u),

as required to complete the proof of the implication (ii) ⇒ (iii). 2

Proof of Proposition 4.0.4. We give a simple proof in the case where ∂Ω

is smooth. For a proof in the case where ∂Ω is Lipschitz, see [EG, Thm 1,
p. 133].
Extend the unit normal field ν on ∂Ω to a vector field on a neighbour-
hood of Ω (for instance, use Lemma 3.2.7 to extend ν to a vector field in a
neighbourhood of ∂Ω, and then multiply it by a smooth cutoff function). By
the divergence theorem
Z Z
2
|u| dσ = (|u|2 ν) · ν dσ
∂Ω ∂Ω
n Z
X
= ∂j (|u|2 ν) dV
j=1 Ω
Xn Z
|u (∂j u) νj | + |(∂j u) u νj | + |u|2 |∂j νj | dV ∀u ∈ C 2 (Ω).
 
≤
j=1 Ω
4.1. DETAILS LEFT BEHIND 117

Now observe that ν and div ν are uniformly bounded in Rn . Therefore there
exists a constant C such that
Z Xn Z
2
|u (∂j u)| + |(∂j u) u| + |u|2 dV
 
|u| dσ ≤ C
∂Ω j=1 Ω

h Xn i
2
(by Schwartz’s inequality) ≤C 2 kukL2 (Ω) k∂j ukL2 (Ω) + n kukL2 (Ω)
j=1

≤ C kuk2H 1 (Ω) ∀u ∈ C 2 (Ω).

Denote by γ : C 2 (Ω) → L2 (∂Ω) the linear map, defined by

γ(u) := u|∂Ω .

We have proved that there exists a constant C such that

kγ(u)kL2 (∂Ω) ≤ C kukH 1 (Ω) ∀u ∈ C 2 (Ω).

By Exercise 2.2.4, the map γ extends uniquely to a bounded linear map from
H 1 (Ω) to L2 (∂Ω), for C 2 (Ω) is dense in H 1 (Ω) (see Theorem 5.5.12 below).
This concludes the proof of the proposition. 2

Exercise 4.1.1 Prove that if u is a smooth function defined in the unit

ball B1 (0) of R2 , then
Z Z
2
h
2 1 2
i
∇u dV = ∂r u + 2 ∂θ u r dr dθ.
Ω (0,1)×(0,2π) r

Proof of Proposition 4.0.9. We shall be a bit sketchy, leaving to the reader

the task of providing full details. For α ∈ (0, 1], we consider the function gα ,
defined by
∞
k
X
gα (θ) := 2−αk ei2 θ ∀θ ∈ [0, 2π]. (4.1.2)
k=0

It is straighforward to check that the series is uniformly convergent in [0, 2π],

whence gα is continuous therein. Furthermore, the function
∞
k kθ
X
uα (r, θ) := r2 2−αk ei2 ∀θ ∈ [0, 2π] ∀r ∈ [0, 1]
k=0

agrees with gα on the boundary of the unit disc in the plane, is harmonic in
the (open) unit disc (the series can be differentiated term by term as long
118

as r < 1) and it is continuous on the closure of the unit disc. Therefore u

solves the classical Dirichlet problem on the unit disc with boundary datum
gα . By using polar coordinates, we see that
Z
2
D(uα ) = ∇uα dV
ZΩ h
2 1 2
i
= ∂r uα + 2 ∂θ uα r dr dθ
(0,1)×(0,2π) r
Z h
2 1 2
i
= lim ∂r uα + 2 ∂θ uα r dr dθ.
ρ↑1 (0,ρ)×(0,2π) r

We may compute ∂r u and ∂θ u in Ωρ by differentiating the series (4.1.2) term

by term. We find
∞ ∞
k −1 kθ k k
X X
∂r uα = r2 2(1−α)k ei2 and ∂θ uα = r2 2(1−α)k iei2 θ .
k=0 k=0

By using Plancherel’s identity, we see that

Z Z ρX
h
2 1 2
i k+1
lim ∂r uα + 2 ∂θ uα r dr dθ ∼ lim 22k(1−α) r2 −1 dr.
ρ↑1 (0,ρ)×(0,2π) r ρ↑1 0
k

An elementary computation shows that the right hand side is infinite for all
α ≤ 1/2. Thus, for these values of α the classical solution to the Dirichlet
problem with boundary datum gα has infinite Dirichlet integral. 2
Chapter 5

Distributions and Sobolev

spaces

5.1 Continuous functions with compact sup-

port and measures
We aim at generalising the concept of complex valued function of several real
variables. As a first step in this direction, we shall reinterpret such a function
as a (continuous) linear functional on a suitable class of test functions. It
will be convenient to treat from the beginning measures instead of functions.
Recall that a complex measure on Rn is a complex valued function µ
on bounded Borel sets, with the property that if E is a bounded Borel set
and {Ej } is a sequence of pairwise disjoint Borel sets such that
∞
[
E= Ej ,
j=1

then ∞
X
µ(E) = µ(Ej ),
j=1

the series being absolutely convergent. Note that µ is not defined on un-
bounded Borel sets. It may be shown that, given a complex measure µ,
there exist four nonnegative measures µj , j = 1, . . . , 4, such that
4
X
µ(E) = µj (E).
j=1

119
120 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Clearly, we may integrate continuous functions ϕ with compact support with

respect to each µj , and then define
Z 4 Z
X
ϕ dµ = ϕ dµj ∀ϕ ∈ Cc (Rn ).
Rn j=1 Rn

Thus, for every complex measure µ we have defined a complex linear func-
tional T µ on Cc (Rn ), defined by
Z
(T µ)(ϕ) = ϕ dµ ∀ϕ ∈ Cc (Rn ).
Rn

It is straightforward to check that this functional is linear. Thus, we have

defined a map µ 7→ T µ from complex measures to linear functionals on
Cc (Rn ). Clearly, this map is linear. It is not hard to check that T is injective;
in other words, if T µ is the null functional on Cc (Rn ), then µ = 0 on bounded
Borel sets.

Exercise 5.1.1 The map T is injective.

An important example of complex measure is the delta measure δx at

a point x, defined by
(
1 if x ∈ E
δx (E) =
0 if x ∈
/ E.

The linear functional T δx is then given by

Z
T δx (ϕ) = ϕ dδx
Rn
= ϕ(0) ∀ϕ ∈ Cc (Rn ).

We now briefly discuss continuity of the functionals T µ, and the continu-

ity of the linear map T . To do this, we need to define a topology on the space
Cc (Rn ) and a topology on the space of complex measures. Remember that
such measures are defined only on bounded Borel sets. As a first attempt, we
may consider the sup norm on Cc (Rn ). But clearly this choice is not optimal.
Consider, for instance, the measure
∞
X
µ= δj
j=1
5.1. CONTINUOUS FUNCTIONS AND MEASURES 121

on R, where δj denotes the delta function at the point j. Clearly µ is a

complex measure in the above sense. Note, however, that the linear func-
tional T µ is not continuous with respect to the sup norm. Indeed, let g a
continuous function, which vanishes at infinity, i.e. such that

lim g(x) = 0.
x→∞

It is straightforward to show that g is the limit in the uniform norm of a

sequence of functions {ϕk } in Cc (R), with support contained in the interval
[−k, k]. If T µ were a continuous linear functional on Cc (R) with respect to
the sup norm, we should have

lim (T µ)(ϕk ) = (T µ)(g),

k→∞

and (T µ)(g) must be a complex number. But, observe that

X
lim (T µ)(ϕk ) = lim ϕk (j).
k→∞ k→∞
j≤k

P
If g is chosen so that j g(j) = ∞, then the right hand side in the preceding
formula is equal to +∞, thereby contradicting the fact that (T µ)(g) is a
complex number.
Thus, a different topology on Cc (Rn ) is needed. A basic result we shall
use is the following classical result of Riesz. The reader is referred to [Ru,
Ch. 6] for the proof of a slight generalisation thereof. Recall that if X is
a locally compact Hausdorff space, we denote by C0 (X) the completion of
Cc (X) with respect to the uniform norm.

Theorem 5.1.2 Let X be a locally compact Hausdorff space. Then to each

bounded linear functional F on C0 (X) there corresponds a unique complex
Borel measure ν on X such that
Z
F (f ) = f dν
X

for each f in C0 (X). Moreover, ||F || = |ν|(X), where |ν|(X) denotes the
total variation of the measure ν.

The total variation |ν| of ν is a positive measure such that

|ν(E)| ≤ |ν|(E)
122 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

for every measurable set E. In fact, |ν| is the minimal solution to the problem
of finding a positive measure µ that dominates ν, in the sense that

|ν(E)| ≤ µ(E)

for every measurable E. The measure |ν| may be defined as follows:

∞
X
|ν|(E) := sup |ν(Ej )|
j=1

S
where E = j Ej , the measurable sets Ej are pairwise disjoint, and the
supremum is taken with respect to all partitions of E as countable union of
pairwise disjoint measurable subsets of E. For these facts, and many more,
the reader is referred to [Ru, Ch. 6].

Going one step further, for each K ⊂⊂ Rn denote by %K the seminorm

on Cc (Rn ), defined by
%K (f ) = max |f (x)|.
x∈K

The family {%K }K⊂⊂Rn induces a locally convex topology τU C on Cc (Rn ),

which is called the topology of uniform convergence on compact sets.
We recall the definition τU C . For each N -tuple K1 , . . . , KN of compact sets
in Rn and for each positive ε, consider the set

U %K1 , . . . , %KN ; ε := g ∈ Cc (Rn ) : %Kj (g) < ε, j = 1, . . . , N }. (5.1.1)



It is straightforward to check that the collection of all such sets (obtained

by let ε, N , and the compact subsets K1 , . . . , KN of Rn chosen, vary) is
a local base of convex neighbourhoods of the null function in Cc (Rn ). A
base of neighbourhoods at f is obtained by simply translating the above
neighbourhoods by f , i.e., it consists of all the sets of the form


f + U %K1 , . . . , %KN ; ε .

It is not hard to see that τU C is induced by a distance on Cc (Rn ). Indeed,

first choose an exhaustion of Rn by compact subsets. It is convenient to write
∞
[
Rn = Bj ,
j=1

where Bj := B(0, j), and then consider the countable family %B j of the
associated seminorms, which we shall denote by %j for the sake of simplicity.
5.1. CONTINUOUS FUNCTIONS AND MEASURES 123

Note that a base of neighbourhoods at 0 is given by the collection of sets of

the form
U %j ; ε := g ∈ Cc (Rn ) : %j (g) < ε}.


(5.1.2)
Consider the metric on Cc (Rn ) induced by the distance
∞
X %j (f − g)
d(f, g) = 2−j .
j=1
1 + %j (f − g)

Exercise 5.1.3 Prove that d is indeed a metric on Cc (Rn ), and that τU C

is the topology induced by d.

It is straightforward to check that a sequence {fj } of functions in Cc (Rn ) is

convergent to f in this topology if and only if for every compact subset K
of Rn the sequence {fj } is uniformly convergent to f in K.
However, note that Cc (Rn ) is not complete with respect to the distance d.
Indeed, observe preliminarily that {fj } is a Cauchy sequence in Cc (Rn )
if and only if {fj } is a Cauchy sequence with respect to each seminorm %k .


Then set xj = j + (1/2) e1 where e1 denotes the first vector of the
canonical basis of Rn . Clearly xj belongs to Bj+1 \ B j and the distance
c
between xj and B j ∪ B j+1 is clearly equal to 1/2. Pick a function ψ in
Cc (Rn ), whose support is contained in B(0, 1/4) and define

ψj (x) = ψ(x − xj ) ∀x ∈ Rn .

Clearly, the support of ψj is contained in B(xj , 1/4), whence in B j+1 \ B j .

Now, define
XJ
ϕJ := ψj .
j=1

Obviously %N (ϕj − ϕk ) = 0 for every j, k > N . Hence {ϕj } is a Cauchy

P∞d. Furthermore, the sequence {ϕj } is
sequence with respect to the metric
convergent to the function ϕ := j=1 ψj , which, however, does not belong
to Cc (Rn ).

We now define another topology on Cc (Rn ), strictly finer than τU C .

Definition 5.1.4 For each ϕ ∈ Cc (Rn ), set

%00 (ϕ) := max

n
|ϕ|, %0j (ϕ) := max
n
|ϕ| j = 1, 2, 3, . . .
R R \Bj
124 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Note that the sequence {%0j (ϕ)} is decreasing. For each decreasing sequence
ε := {εj } of positive real numbers we consider the set

Vε := {f ∈ Cc (Rn ) : %0j (f ) < εj , j = 1, 2, 3, . . .}. (5.1.3)

It is straightforward to check that the (uncountable) family of the sets

V(0) := {Vε : ε as above}

is a fundamental system of convex neighbourhoods of 0 of a topology τSI on

Cc (Rn ), called the strict inductive limit topology of Cc (Rn ). A base of
neighbourhoods of a function f in Cc (Rn ) is then

V(f ) := f + V(0).

We claim that τSI is strictly finer than τU C
. On the one hand, it is
clear that for each j the neighbourhood U %j ; ε of 0 in the τU C topology
(see (5.1.2)) contains the neighbourhood Vε of 0 in the τSI topology, where
ε = {2−j ε : j = 0, 1, 2, . . .}. Thus, τSI is finer than τU C .
On the other hand, the neighbourhood

V := {g ∈ Cc (Rn ) : %0j (g) < 2−j ε, j = 0, 1, 2, . . .}

of 0 does not contain any set of the form U %j ; η , whence τSI is strictly finer
than τU C , as claimed.
It will be important to understand τSI better. In particular, it is inter-
esting to characterise convergent sequences. We have the following result.

Proposition 5.1.5 Let {fN } be a sequence of functions in Cc (Rn ). The

following are equivalent:

(i) {fN } is convergent to a function f ∈ Cc (Rn ) in the strict inductive

limit topology τSI ;

(ii) there exists ` such that the support of {fN } is contained in B ` and {fN }
is uniformly convergent to f .

Proof. By replacing fN by fN − f we may assume that f is the function

identically 0 in Rn .
Since τSI is finer than τU C , (ii) implies (i).
To prove that (i) implies (ii), we argue by reductio ad absurdum. Suppose
that {fN } is not contained in Cc (B ` ) for any `. Then there would exist two
5.1. CONTINUOUS FUNCTIONS AND MEASURES 125

increasing sequences of indices Nj and nj such that fNj belongs to Cc (B nj+1 )\

Cc (B nj ). Therefore there exist points xnj in Bnj+1 \ B nj such that


fNj xnj 6= 0, j = 1, 2, 3, . . . .


We may also assume that the sequence fNj xnj is decreasing.
Then
choose a decreasing sequence ε := {εk } such that εNj = fNj xnj . Con-
sider any tail {fh , fh+1 , fh+2 , . . .} of the original sequence. We claim that
this tail is not contained in Vε , where ε is the sequence chosen above. In-
deed, the sequence Nj tends to infinity as j tends to infinity, so that at least
one (in fact, infinitely many) function fNj belongs to the tail, and

%nj (fNj ) = sup fNj (x)

x∈B
/ nj

≥ fNj (xnj )
= εnj+1 ,

so that fNj is not in Vε , as claimed.

This contradicts the fact that {fj } tends to 0 as j tends to infinity. There-
fore, the supports of the functions {fj } must be contained in B ` for some `,
as required. 2

Corollary 5.1.6 Every Cauchy sequence for the topology τSI on Cc (Rn ) is
convergent to an element of Cc (Rn ).

Proof. Let {fj } be a Cauchy sequence with respect to τSI . Since τSI is
finer than τU C , {fj } is a Cauchy sequence with respect to the topology τU C .
Therefore {fj } is a Cauchy sequence with respect to each of the seminorms
%k . In particular, the restrictions of the functions fj to B k is a Cauchy
sequence in C(B k ) for every k. Since C(B k ) is complete with respect to the
uniform norm, there exists a function gk in C(B k ) such that

sup fj (x) − gk (x) → 0 as j tends to ∞.

x∈B k

It is clear that the restriction of gk+1 to B k agrees with gk . Therefore, we

may define a continuous function g on Rn by

g(x) = gk (x) ∀x ∈ B k .

This function is the uniform limit of {fj } on each ball B k .

126 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

It remains to show that the support of g is compact and that {fj } is

convergent to g with respect to τSI .
We argue by contradiction. Suppose that the support of g is not compact.
Denote by nj an increasing sequence of indices such that the support of fNj
is contained in Cc (B nj ). Denote by xnj a sequence of points in Bnj+1 \ B nj
for which g(xnj ) does not vanish, j = 1, 2, 3, . . .. Then, choose positive real
numbers ηnj > 1 so that εnj := |g(xnj )|/ηnj form a decreasing sequence.
Then complete {εnj } to a decreasing sequence {εk }.
We claim that for each positive integer h the tail {fh − g, fh+1 − g, . . .}
is not contained in Vε , where ε is the sequence {εk }, defined above. Indeed,
the tail contains an element of the form fNj − g, and

%nj (fNj − g) ≥ |(fNj − g)(xnj )| > |(fNj − g)(xnj )|/ηnj = εnj ,

as claimed.
This contradicts the fact that {fN } is convergent to g. Hence we have
proved that the support of g is compact.
The proof that {fN } is convergent to g in the τSI topology is straightfor-
ward and is left to the reader. 2

Now we show that, given a complex measure µ on Rn , the linear func-

tional T µ, defined by
Z
(T µ)(ϕ) = ϕ dµ ∀ϕ ∈ Cc (Rn ),
Rn

is continuous on Cc (Rn ), τSI .
We need to show that for every ε > 0, the set
Z
−1

n o
(T µ) Bε (0) = ϕ ∈ Cc (Rn ) : ϕ dµ < ε
Rn

(here Bε (0) denotes the open disc in the complex plane with centre 0 and
radius ε) contains a neighbourhood of the origin in the strict inductive limit
topology, i.e., there exists a decreasing sequence η := {ηj } such that

(T µ)(ϕ) < ε ∀ϕ ∈ Vη .
5.1. CONTINUOUS FUNCTIONS AND MEASURES 127

Set ηj := 2−j−1 ε/|µ|(B j+1 ). Then, given ϕ ∈ Vη , we write

Z ∞ Z
X
(T µ)(ϕ) ≤ ϕ dµ + ϕ dµ
B1 j=1 B j+1 \Bj
∞


X


≤ sup |ϕ| |µ| B 1 + sup |ϕ| |µ| B j+1 \ Bj
B1 j=1 (Bj )c
∞
X



≤ η0 |µ| B 1 + ηj |µ| B j+1
j=1

< ε,

as required.

It is an interesting fact that the converse of this result holds.

Theorem 5.1.7 To each continuous linear functional F on Cc (Rn ), τSI
there corresponds a unique complex Borel measure ν on X such that
Z
F (ϕ) = ϕ dν ∀ϕ ∈ Cc (Rn ).
Rn

Proof. We give a sketch of the proof,

leaving the details to the reader. Since
n
F is continuous on Cc (R ), τSI , for every ε > 0 there exists a sequence
η := {ηj } such that
F (ϕ) < ε ∀ϕ ∈ Vη . (5.1.4)
Recall that ϕ ∈ Vη if and only if

sup |ϕ| < η0 and sup |ϕ| < ηj j = 1, 2, 3, . . .

Rn Bjc

We prove that for every positive integer j the restriction of F to C0 (Bj ) is a

continuous linear functional on C0 (Bj ) (with respect to the uniform norm).
Indeed, suppose that ϕ is a function in C0 (Bj ) such that kϕk∞ ≤ ηj . Then
clearly ϕ belongs to Vη . Therefore |F (ϕ)| < ε, as required. By the Riesz
representation theorem 5.1.2 there exists a unique finite complex measure νj
such that Z
F (ϕ) = ϕ dνj ∀ϕ ∈ C0 (Bj ).
Bj

Since C0 (Bj ) is continuously included in C0 (Bj+1 ), the restriction of νj+1 to

C0 (Bj ) agrees with νj . Thus, there is a uniquely defined Radon measure ν
128 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

on Cc (Rn ) such that its restriction to C0 (Bj ) agrees with νj , and

Z
F (ϕ) = ϕ dνj ∀ϕ ∈ Cc (Rn ),
Bj

as required. 2

Finally, we describe a mild generalisation of the theory developed so far, in

which a fixed open set Ω plays the role of Rn . Recall the following classical
result.

Theorem 5.1.8 Suppose that Ω is an open subset of Rn . The following hold:

(i) there exists a sequence {VN } of precompact sets in Ω such that V N ⊂

VN +1 for every positive integer N and ∞
S
N =1 V N = Ω;

(ii) there exists a sequence {fN } of functions in Cc∞ (Ω) such that supp (fN ) ⊂
VN +1 and fN = 1 on V N .

Proof. [Fo1, Prop. 4.31, Prop. 4.32, Lemma 8.18]. 2

The sequence of sets {VN } is called an exhaustion of Ω. There are many

exhaustions of the same set Ω. The topologies we shall introduce below will
be independent of the chosen exhaustion.

The role played above by the space Cc (Rn ) will now be played by the
space Cc (Ω). The relevant topology on Cc (Ω) will be the strict inductive
limit topology τSI (Ω), which is defined similarly to the strict inductive
limit topology τSI on Cc (Rn ), with the role of the seminorms %0j played now
by the seminorms ρ0j , defined by

ρ00 (f ) := sup |f (x)| and ρ0j (f ) := sup |f (x)| j = 1, 2, 3, . . .

x∈Ω x∈Ω\Vj

By arguing as in Proposition 5.1.5, it may be shown that a sequence {fN } of

functions in Cc (Ω) is convergent to f in the strict inductive limit topology if
and only if there exists ` such that the supports of all the functions fN are
contained in V ` , and {fN } tends uniformly to f on V` .
Furthermore, every Cauchy sequence with respect to τSI (Ω) is convergent
to an element of Cc (Ω). We express this fact by saying that Cc (Ω), endowed
with the strict inductive limit topology, is complete.
5.2. SMOOTH FUNCTIONS AND DISTRIBUTIONS 129

Now, we attribute to the Lebesgue measure λ a distinguished role. For

every locally integrable function f (with respect to the Lebesgue measure),
we consider the measure
dµf = f dλ
absolutely continuous with respect to the Lebesgue measure and with den-
sity f . Thus, to every locally integrable function with respect to λ, we may
associate the linear functional T µf on Cc (Rn ), defined by
Z
(T µf )(ϕ) = ϕ f dλ ∀ϕ ∈ Cc (Rn ).
Rn

5.2 Smooth functions and distributions

There are physical situations which lead naturally to notions more compli-
cated than that of masses, and require mathematical tools more refined and
general than measures.
For instance, consider a positive charge 1/ε placed at the point ε on the
real line, and a negative charge −1/ε at the point −ε. Does the “limit” of
such a configuration make sense?
If we proceed naively, by just taking the

δε (E) − δ−ε (E)

lim ,
ε→0 ε
i.e., the limit of the measures, as additive set functions, we get into trouble.
Indeed, the limit vanishes over all intervals, except over those with an end
at the origin, for which it is undetermined.
If, instead, we adopt the point of view that measures correspond to con-
tinuous linear functionals on Cc (Rn ), and define Tε to be the linear functional

δε (ϕ) − δ−ε (ϕ)

Tε (ϕ) := lim ∀ϕ ∈ Cc (Rn ),
ε→0 ε
then we see that
ϕ(ε) − ϕ(ε)
lim Tε (ϕ) = lim
ε→0 ε→0 ε
0
= ϕ (0),
at least for every function ϕ in Cc1 (Rn ). Thus, the linear form

T (ϕ) := lim Tε (ϕ)

ε→0
130 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

is defined only on a proper linear subspace of Cc (Rn ), namely C 1 (Rn ). More-

over, T is discontinuous (with respect to the topology of uniform convergence
on compact subsets of Rn ). Indeed, given a sequence {ϕj } of functions in
C 1 (Rn ), uniformly convergent to 0 on compact subsets of Rn , the sequence
of their derivatives {ϕ0j } need not converge uniformly to 0, as simple exam-
ples show. Thus, in order to consider multipoles of any order, we are led
to consider linear functionals on spaces of infinitely differentiable functions
endowed with a finer topology than the topology τSI considered in the last
section.

Definition 5.2.1 Let K be a compact subset of Ω. Denote by DK (Ω) the

space of all functions in C ∞ (Ω) with support contained in K, endowed with
the family of seminorms

%m,K (f ) := max max |Dα f (x)|,

|α|≤m x∈K

where m is a nonnegative integer.

Note that DK (Ω) and DK (Rn ) are just the same space.

Exercise 5.2.2 Produce an explicit example of a function ϕ in DB(0,1) (Rn ).

Then produce an example of a smooth function on Rn with support contained
in B(0, 1) that is equal to 1 on B(0, 1/2).

Exercise 5.2.3 Suppose that K is a compact set in Rn , and g is a con-

tinuous function on K. Prove that there exists a sequence {Gj } of smooth
functions on Rn such that Gj → g uniformly on K, by completing the steps
below:

(i) if K0 and K1 are two disjoint compact sets in Rn , show that the con-
tinuous function

d(x, K0 )
η(x) := ∀x ∈ Rn
d(x, K0 ) + d(x, K1 )

vanishes on K0 , it is equal to 1 on K1 , and satisfies 0 ≤ η ≤ 1;

(ii) show that it suffices to prove the result for functions g satisfying 0 ≤
g ≤ 1;
5.2. SMOOTH FUNCTIONS AND DISTRIBUTIONS 131

(iii) define K0 := {x ∈ K : 2/3 ≤ f (x) ≤ 1} and K1 := {x ∈ K : 0 ≤

g(x) ≤ 1/3}, and denote by F1 a continuous function on Rn which is
equal to 1/3 on K0 , to 0 on K1 , and satisfies 0 ≤ F1 ≤ 1/3 on Rn .
Show that
0 ≤ g(x) − F1 (x) ≤ 2/3 ∀x ∈ K;

(iv) by iterating the procedure in (iii), find continuous functions F2 , . . . , Fj

on Rn such that

0 ≤ g(x) − F1 (x) − · · · − Fj ≤ (2/3)j ∀x ∈ K,

and 0 ≤ Fj ≤ (1/3)(2/3)j−1 on Rn ;
(v) set F = j Fj . Prove that F is continuous on Rn and F|K = g;
P

(vi) set Hε := F ∗ ϕε , where ϕ ∈ Cc∞ (B(0, 1)), ϕ dV = 1, and ϕε (x) =

R

ε−n ϕ(x/ε). Prove that

Hε (x) − F (x) ≤ sup F (x) − F (y)

|x−y|≤ε

for every x ∈ K. Show that {H1/j } is the required sequence.

Observe that DK (Ω) is metrizable. Indeed, its topology is induced by the

distance
∞
X %m,K (f − g)
dK (f, g) := 2−m ∀f, g ∈ DK (Ω).
m=0
1 + %m,K (f − g)

A sequence {fN } is convergent to f in DK (Ω) if and only if for every multi-

index α, the sequence {Dα fN } tends to Dα f uniformly in K.

Definition 5.2.4 Suppose that Ω is an open subset of Rn , and let {VN } be

an exhaustion of Ω as in Theorem 5.1.8. We denote by D(Ω) the vector space
Cc∞ (Ω), endowed with the strict inductive limit topology τSI
∞
(Ω). Explicitly,
∞
a fundamental system of neighbourhoods of 0 in τSI (Ω) is given by the sets

Vε,m := {f ∈ Cc∞ (Ω) : %0m0 ,Ω (f ) < ε0 , %0mj ,Ω\Vj (f ) < εj , j = 1, 2, . . .},

as ε := {εj } and m := {mj } vary among all possible decreasing sequences ε

that converge to 0 and all possible incresing sequences of integers that tend
to infinity. Here

%0m0 ,Ω (f ) := max max Dα f , %0mj ,Ω\Vj (f ) := max max Dα f . (5.2.1)

|α|≤m0 Ω |α|≤mj Ω\Vj
132 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

∞
Observe that τSI (Ω) is not the metrizable topology induced by the countable
family of seminorms %m,KN (here KN = V N ). It is a finer topology, which
induces on DKN (Ω) the topology introduced in Definition 5.2.1.
∞
We might argue that τSI (Ω) may depend on the chosen exhaustion {KN }
of Ω, but, in fact, it is straightforward to prove that two different exhaustion
give rise to the same topology. Henceforth, for every domain Ω we fix once
and for all an exhaustion, and all seminorms are associated to that particular
exhaustion.

Proposition 5.2.5 Let {fN } be a sequence of functions in D(Ω). The fol-

lowing are equivalent:

(i) {fN } is convergent to the function f ∈ D(Ω) in the strict inductive

∞
limit topology τSI (Ω);

(ii) there exists ` such that the support of fN is contained in V ` for ev-
ery N and for every mullti-index α the sequence {Dα fN } is uniformly
convergent to Dα f on V ` .

Proof. The proof is similar to the proof of Proposition 5.1.5, and it is left as
an exercise. 2

∞
Corollary 5.2.6 Every Cauchy sequence for the topology τSI (Ω) on D(Ω) is
convergent to an element of D(Ω).

Proof. The proof is similar to the proof of Corollary 5.1.6, and it is left as
an exercise. 2

Definition 5.2.7 A distribution on Ω is a continuous linear functional

on D(Ω). The linear space of all distributions on Ω is denoted by D0 (Ω). We
write D and D0 instead of D(Rn ) and D0 (Rn ), respecitvely.

It is a natural question to ask whether there are simple criteria which imply
∞
that a linear functional on D(Ω) is also continuous (with respect to the τSI
topology). The following theorem gives a handy characterisation of distribu-
tions.

Theorem 5.2.8 Suppose that T is a linear functional on D(Ω). The follow-

ing are equivalent:
5.2. SMOOTH FUNCTIONS AND DISTRIBUTIONS 133

∞
(i) T is a distribution (i.e., T is continuous with respect to the τSI topol-
ogy);

(ii) for every compact set K ⊂ Ω, there exist a constant CK and a number
mK such that

hϕ, T i ≤ CK %mK ,K (ϕ) ∀ϕ ∈ DK (Ω)

(see Definition 5.2.1 for the meaning of %j,K ).

Proof. Suppose first that T ∈ D0 (Ω). Then for every ε > 0 there exist
sequences η and m such that

hϕ, T i < ε ∀ϕ ∈ Vη,m .

In particular, if K is a compact subset of Ω, ϕ ∈ DK (Ω), and K ⊂ V` , then

%0h,Ω\Vj (ϕ) = 0 ∀j ≥ ` ∀h ∈ N,

because ϕ vanishes identically in Ω \ Vj . Since η is decreasing and m is

increasing, the set 
ϕ ∈ DK (Ω) : %m` ,K (ϕ) < η`
is contained in Vη,m (check the details). Now, suppose that ψ is any function
in DK (Ω) not identically equal to 0. We claim that the function
ψ
ψe := η`
%mK ,K (ψ)

belongs to Vη,m . Indeed, %0h,Ω\Vj (ψ)

e = 0 for every j ≥ `, and, if j < `, then

%0mj ,Ω\Vj (ψ)

e = max max Dα ψe
|α|≤mj Ω\Vj

≤ max max Dα ψe
|α|≤mj K

Dα ψ
= η` max max
|α|≤mj K %m` ,K (ψ)
< η`
< ηj ,
the last inequality being a consequence of the facts that η is decreasing and
j < `. Therefore
ε > ψ,e T = η` ψ, T ,
%m` ,K (ψ)
134 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

whence
ε
ψ,
eT < %m` ,K (ψ),
η`

as required, with CK := ε/η` .

Conversely, suppose that for every compact set K ⊂ Ω there exist a

constant CK and an integer jK such that

ψ, T ≤ CK %jK ,K (ψ) ∀ψ ∈ DK (Ω). (5.2.2)

We must show that T is a distribution on Ω, i.e., that for every ε > 0 there
exist two sequences η and m such that

hϕ, T i < ε ∀ϕ ∈ Vη,m .

Choose as m any increasing sequence {m` } such that m` ≥ jV ` . There exist

test functions {ω` } such that

supp (ω0 ) ⊂ V2 and supp (ω` ) ⊂ V`+2 \ V` ` = 1, 2, . . .

and
∞
X
1= ω` (x) ∀x ∈ Ω.
`=0

Given a test function ϕ on Ω, write ϕ = ∞

P
`=0 ω` ϕ. Observe that only finitely
many summands in the series above do not vanish identically on Ω, for ϕ has
compact support.
For the sake of simplicity we write %`+2 instead of %jV ,V `+2 . It is straight-
`+2
forward to show that there exists a constant C`+2 such that

%`+2 (ω` ϕ) ≤ C`+2 %0jV ,Ω\V` (ϕ)

`+2

(expand the derivatives of ω` ϕ with Leibnitz’s formula). Denote by η any

decreasing sequence of positive numbers such that

2−`−1 ε
η` < .
CV `+2 C`+2
5.2. SMOOTH FUNCTIONS AND DISTRIBUTIONS 135

Note that for every ϕ ∈ Vη,m

∞
X
|hϕ, T i| ≤ |hω` ϕ, T i|
`=0
X∞
≤ CV `+2 %`+2 (ω` ϕ)
`=0
∞
X (5.2.3)
≤ CV `+2 C`+2 %0jV ,Ω\V` (ϕ)
`+2
`=0
X∞
< ε 2−`−1
`=0
≤ ε,
as required.
The proof of the theorem is complete. 2
A consequence of the characterisation above is that each locally integrable
function and, more generally, every complex measure (in the sense of Sec-
tion 5.1) on Ω, is a distribution on Ω. This partially justifies the term gen-
eralised function often used to denote a distribution.

Proposition 5.2.9 Suppose that Ω is an open subset of Rn . The following

hold:

(i) Lp (Ω) is continuously included in D0 (Ω) for each p ∈ [1, ∞];

(ii) every locally integrable function f on Ω “is” a distribution on Ω.

Proof. First we prove (i). Suppose that f is in Lp (Ω), and consider the linear
functional Tf on D(Ω), given by
Z
hϕ, Tf i = ϕ f dλ ∀ϕ ∈ D(Ω),
Ω

where λ denotes the Lebesgue measure on Rn . We prove that Tf is continuous

∞
with respect to the τSI topology. Suppose that K is a compact subset of Ω.
Observe that for every ϕ ∈ DK (Ω)
Z 
hϕ, Tf i ≤ sup |ϕ| |f | dλ
K K
0
≤ sup |ϕ| kf kLp (K) λ(K)1/p
K
136 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

and the continuity of Tf follows from Theorem 5.2.8.

The proof of (ii) is left as an exercise. 2
Note that two locally integrable functions which are equal a.e. represent the
same distribution.

Exercise 5.2.10 Prove that every complex measure is a distribution.

Exercise 5.2.11 Prove that the linear functional T on D(R), defined by

hϕ, T i := ϕ0 (0),
is a distribution. More generally, prove that for every multiindex α, and for
every x ∈ Ω, the the linear functional Tx,α on D(Ω), defined by
hϕ, Tx,α i := Dα ϕ(x),
is a distribution.

5.3 Derivatives of distributions

One of the motivations for the introduction of distributions is that every
distribution may be differentiated infinitely many times. More precisely, for
every multi-index α, we may define a continuous linear map Dα : D0 (Ω) →
D0 (Ω) with the property that if T is a distribution which may be represented
by a function fT in C |α| (Ω), then Dα T is represented by the usual derivative
Dα fT . Thus, the notion of derivative of distributions generalises that of
derivative of smooth functions.
If f is in C 1 (Ω), then for every test function ϕ in D(Ω) we may write
Z
hϕ, ∂k f i = ϕ(x) ∂k f (x) dx
ΩZ

(by integrating by parts) = − ∂k ϕ(x) f (x) dx

Ω
= − h∂k ϕ, f i ;
note that there are no boundary terms in the integration by parts above,
for ϕ vanishes in a neighbourhood of the boundary of Ω. Thus, for any
continuously differentiable function f in Ω, we have the formula
hϕ, ∂k f i = − h∂k ϕ, f i ∀ϕ ∈ D(Ω).
The key observation is that the right hand side makes sense for every distri-
bution f . This leads to the following definition.
5.3. DERIVATIVES OF DISTRIBUTIONS 137

Definition 5.3.1 The j th partial derivative of the distribution T in

D0 (Ω) is the distribution ∂j T , defined by

hϕ, ∂j T i = − h∂j ϕ, T i ∀ϕ ∈ D(Ω).

Of course we need to check that the right hand side of the formula above
defines a distribution, i.e., the linear functional ϕ 7→ − h∂j ϕ, T i is continuous
for the strict inductive limit topology of D(Ω). By Theorem 5.2.8, ∂j T is con-
tinuous if and only if for every compact set K ⊂ Ω, there exist a constant CK
and a nonnegative integer mK such that

|hψ, ∂j T i| ≤ CK %mK ,K (ψ) ∀ψ ∈ DK (Ω). (5.3.1)

Since T is a distribution, for every compact set K ⊂ Ω, there exists a constant

CK and an integer m0K such that

hϕ, T i ≤ CK %m0K ,KN (ϕ) ∀ϕ ∈ DK (Ω).

Therefore
hϕ, ∂j T i = h∂j ϕ, T i
≤ CK %m0K ,K (∂j ϕ)
≤ CK %m0K +1,K (ϕ) ∀ϕ ∈ DK (Ω),
so that (5.3.1) is satisfied with mK = m0K + 1, as required.
It is clear that we can iterate the reasoning above, and define Dα T for
every distribution T . This will be the distribution defined by

hϕ, Dα T i = (−1)|α| hDα ϕ, T i ∀ϕ ∈ D(Ω). (5.3.2)

2 2
Note, in particular, that ∂j,k T = ∂k,j T for every pair of indices.

Example 5.3.2 Consider the absolute value function T on the real line

T (x) = |x| ∀x ∈ R.

Clearly T is differentiable in the classical sense off the origin with derivative

T 0 (x) = sgn(x),

where sgn denotes the signum function, which vanishes at 0, is equal to −1

on the negative real line and to +1 on the positive real line. We claim that
138 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

T 0 = sgn in the sense of distributions. Indeed, by the definition of derivative

of a distribution

hϕ, T 0 i = − hϕ0 , T i
Z ∞
(T is a loc. int. function) =− ϕ0 (x) |x| dx
−∞
Z 0 Z −∞
0
= ϕ (x) x dx − ϕ0 (x) x dx
−∞ 0
Z 0 Z −∞
(by parts) =− ϕ(x) dx + ϕ(x) dx
−∞ 0
Z ∞
= ϕ(x) sgn(x) dx ∀ϕ ∈ D,
−∞

as claimed.

Example 5.3.3 We show that

H 0 = δ0

in the sense of distributions. Here H is the Heaviside function, which

vanishes in (−∞, 0] and is equal to 1 in (0, ∞), and δ0 denotes the Dirac
mass at the origin. Indeed,

hϕ, H 0 i = − hϕ0 , Hi
Z ∞
(H is a loc. int. function) =− ϕ0 (x) dx
0
= ϕ(0)
= hϕ, δ0 i ∀ϕ ∈ D,

as required. Similarly, for every positive integer k ≥ 2 one can prove that

(k)
ϕ, δ0 = (−1)k ϕ(k) (0) ∀ϕ ∈ D.

Example 5.3.4 Consider the distribution T , which agrees with the locally
integrable function x 7→ x−1/2 1(0,∞) (x). We want to compute T 0 (in the sense
of distributions). Clearly, T 0 is not the function x 7→ −(1/2) x−3/2 1(0,∞) (x),
5.3. DERIVATIVES OF DISTRIBUTIONS 139

for this function is not locally integrable. Observe that

hϕ, T 0 i = − hϕ0 , T i
Z ∞
(T is a loc. int. function) =− ϕ0 (x) x−1/2 dx
0
Z ∞
= − lim ϕ0 (x) x−1/2 dx
ε↓0 ε
1 ∞
h Z i
−1/2 −3/2
(by parts) = − lim − ϕ(ε) ε − ϕ(x) x dx
ε↓0 2 ε
1 ∞ ϕ(x) − ϕ(ε)
Z
= − lim dx
ε↓0 2 ε x3/2
1 ∞ ϕ(x) − ϕ(0)
Z
=− dx ∀ϕ ∈ D.
2 0 x3/2

The last inequality is justified by an application of the Lebesgue dominated

convergence theorem. We leave the details to the reader. We already know
that the linear functional
1 ∞ ϕ(x) − ϕ(0)
Z
ϕ 7→ − dx
2 0 x3/2

is continuous on D, because it is the distributional derivative of a distribu-

tion. Prove directly the continuity of the above functional. Note also that
“the restriction” of T 0 to (0, ∞) “agrees” with the function − 12 x−3/2 in the
following sense: for every ϕ in D with support contained in (0, ∞), then
Z ∞
0 1 ϕ(x)
hϕ, T i = − dx.
2 0 x3/2

Example 5.3.5 (Cauchy’s principal value) Denote by T the linear

functional on D(R), defined by
Z
ϕ(x)
hϕ, T i = lim+ dx ∀ϕ ∈ D(R),
ε→0 Iεc x

where Iε = [−ε, ε]. We show that T is a distribution. It suffices to prove

that the restriction of T to D[−R,R] (R) is a continuous linear functional (with
respect to the topology of D[−R,R] (R). Observe that if 0 < ε < 1, then we
may write Z Z
ϕ(x) ϕ(x)
hϕ, T i = lim+ dx + dx
ε→0 I1 \Iε x I1c x
140 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

and that Z Z
ϕ(x) 1
dx ≤ kx ϕk∞ dx
I1c x I1c x2
≤ 2R kϕk∞ .
Furthermore
Z Z 1
ϕ(x) ϕ(x) − ϕ(−x)
lim+ dx = lim+ dx
ε→0 I1 \Iε x ε→0 ε x
Z 1 Z x
1
= lim+ ϕ0 (s) ds dx
ε→0 ε x −x
0
≤ 2 kϕ k∞ ∀ϕ ∈ D[−R,R] (R).

Hence
|hϕ, T i| ≤ CR (kϕ0 k∞ + kϕk∞ ) ∀ϕ ∈ D[−R,R] (R).
Therefore the restriction of T to D[−R,R] (R) is continuous, as required.

Exercise 5.3.6 Prove that δ00 is not a measure.

Exercise 5.3.7 Compute the derivative (in the sense of distributions) of

the distribution T , which agrees with the locally integrable function x 7→
|x|−1/2 .

Exercise 5.3.8 Compute the second derivative (in the sense of distribu-
tions) of the distribution T , which agrees with the locally integrable function
x 7→ x−1/2 1(0,∞) (x).

Exercise 5.3.9 Compute the derivative of the distribution T , which agrees

with the locally integrable function x 7→ log |x|. (Answer: T 0 = p.v.(1/x).)

Exercise 5.3.10 Compute the derivative of the distribution T , which

agrees with the locally integrable function x 7→ 1R+ (x) log |x|.

Exercise 5.3.11 Compute the derivative of the distribution “Cauchy prin-

cipal value”.

Exercise 5.3.12 Compute T 00 , where T is the distribution on the real line

that agrees with the locally integrable function x 7→ max(1 − |x|, 0).
5.4. FUNDAMENTAL SOLUTIONS 141

Exercise 5.3.13 Denote by T the linear functional on D(R2 ), defined by

Z
hϕ, T i = ϕ(x, x) dx.
R

Prove that T is a distribution, and compute ∂x1 T + ∂x2 T .

Exercise 5.3.14 Denote by σ the normalized surface measure of the unit

sphere in R2 , and let f and g be the functions on R2 , defined by
p
g(x, y) = (x2 + y 2 )−1/2 1B(0,1) (x, y).


f (x, y) = max 1 − x2 + y 2 , 0 e

Prove that ∆f = σ − g in the sense of distributions.

Exercise 5.3.15 For each r in R+ denote by σr the surface measure of the

sphere with centre 0 and radius r in Rn . Prove that
2n
h 1 i
lim+ 2 σ − δ0 = ∆δ0
n−1 r
r→0 r sn r

in the sense of distributions. Here sn denotes the surface measure of the unit
sphere.

2
Exercise 5.3.16 Denote by p f the distribution which agrees on R \{(0, 0)}
2 2
with the function (x, y) 7→ x + y . Compute the partial derivatives of f
up to the second order, and, then the Laplacian of f .

Exercise 5.3.17 Let Ω be an open subset of Rn and {fN } a sequence in

L1loc (Ω) such that Z
lim fN dλ = 0
N →∞ K

for every K ⊂⊂ Ω. Show that the sequence {Dα fN } is convergent to 0 in

D0 (Ω) for every multi-index α.

5.4 Fundamental solutions

In this section we introduce the concept of fundamental solution of a differ-
ential operator, and illustrate its application to the solution of the Poisson
equation.
142 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Definition 5.4.1 Suppose that J is a nonnegative integer and that

X
P (x) := aα x α .
|α|≤J

We denote by P (D) the differential operator, defined by

X
P (D)f = aα Dα f.
|α|≤J

Definition 5.4.2 A distribution E is a fundamental solution of the

differential operator P (D) if

P (D)E = δ0 .

The following remarkable result is due to Malgrange and Ehrenpreis.

Theorem 5.4.3 Every differential operator P (D) admits a fundamental so-

lution.

We do not prove this result. The interested reader is referred to [Fo2,

Thm 1.56, p. 62] or [Me, Thm 3.3.20].

Proposition 5.4.4 The Newtonian potential is a fundamental solution of

the Laplace operator.

Proof. We already know that N is harmonic in Rn \ {0}. Thus,

hϕ, ∆N i = 0 = hϕ, δ0 i (5.4.1)

for all ϕ ∈ Cc∞ (Rn ) such that 0 does not belong to the support of ϕ.
Now suppose that 0 belongs to the support of ϕ. Then

hϕ, ∆N i = h∆ϕ, N i
Z
= ∆ϕ N dV
Rn Z

= lim ∆ϕ N dV.
ε↓0 Rn \Bε (0)

The last equality follows from the local integrability of N near the origin and
the Lebesgue dominated convergence theorem. Suppose that the support of ϕ
5.4. FUNDAMENTAL SOLUTIONS 143

is contained in BR (0). Then, by the second Green’s identity the harmonicity

of N in Rn \ Bε (0) and the fact that ϕ vanishes on ∂BR (0),
Z Z

∆ϕ N dV = ϕ ∂ν N − ∂ν ϕ N ] dσ,
Rn \Bε (0) ∂Bε (0)

where ν denotes the outward unit normal to ∂Bε (0). By arguing as in the
proof of Theorem 1.5.5, we see that
Z Z
ϕ ∂ν N dσ → ϕ(0) and that ∂ν ϕ N dσ → 0
∂Bε (0) ∂Bε (0)

as ε ↓ 0. Thus, we have proved that

hϕ, ∆N i = ϕ(0) = hϕ, δ0 i

for all ϕ, whose support contains 0. This, together with (5.4.1), proves the
required result. 2

Exercise 5.4.5 Prove that the function E(x) := (1/2) |x| is a fundamental
solution of the operator d2 / dx2 (with pole 0).

Exercise 5.4.6 Prove that the function

(
1 for x1 > ξ1 , x2 > ξ2
u(x1 , x2 ) :=
0 elsewhere

is a fundamental solution of the operator ∂ 2 /∂x1 ∂x2 with pole (ξ1 , ξ2 ).

Exercise 5.4.7 Prove that for every complex number c the function

e−c|x|
uc (x) := ∀x ∈ R3 \ {0}
4π|x|

is a fundamental solution of the operator −∆ + c2 with pole 0.

We shall apply Proposition 5.4.4 to the solution of the Poisson equation.

Definition 5.4.8 Suppose that P (D) is a differential operator with con-

stant coefficients and that f is a distribution on Ω. We say that a distribu-
tion T on Ω is a distributional solution of the equation P (D)u = f if

hϕ, P (D)T i = hϕ, f i ∀ϕ ∈ Cc∞ (Ω).

144 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Note that no a priori regularity assumptions are made on the solution u.

Observe that if T is a classical solution to the Poisson equation (then f
must be continuous), then a repeated integration by parts shows that T is a
distributional solution.
Now we show how the fundamental solution of the Laplacian can be used to
solve Poisson’s equation. We need the definition of convolution between a
distribution with compact support and a test function.

Definition 5.4.9 Suppose that T is a distribution and that ϕ is in D.

Then the convolution T ∗ ϕ is the distribution, defined by
hψ, T ∗ ϕi := hψ ∗ ϕ̌, T i ∀ψ ∈ D. (5.4.2)

Exercise 5.4.10 Prove that T ∗ ϕ defines indeed a distribution on Rn , i.e.,

show that the linear functional on D, defined by (5.4.10), is continuous with
∞
respect to the τSI topology.

Exercise 5.4.11 Suppose that T is a distribution with compact sup-

port. This means that there exists a compact set K ⊂ Rn such that
hϕ, T i = 0 ∀ϕ : supp (ϕ) ∩ K = ∅.
Prove that the linear functional T ∗ N on D, defined by
hϕ, T ∗ N i = ϕ ∗ Ň , T ∀ϕ ∈ D,
is a distribution on Rn . Hint: Observe that ϕ ∗ Ň is a smooth function, and
that its behaviour outside K is “irrelevant”, for the support of T is contained
in K.

Proposition 5.4.12 Suppose that f is a distribution on Rn with compact

support. The following hold:

(i) the distribution N ∗f is a distributional solution of the equation ∆u = f ;

(ii) if f ∈ Cc2 (Rn ), then N ∗f is a classical solution (i.e., a C 2 (Rn ) solution)
of the equation ∆u = f .

Proof. To prove (i), observe that the convolution N ∗ f is well defined,

because f has compact support. Then, by the definition of distributional
Laplacian and of convolution of distributions,
hϕ, ∆(N ∗ f )i = h∆ϕ, N ∗ f i
(5.4.3)
= (∆ϕ) ∗ Ň , f .
5.5. SOBOLEV SPACES 145

Recall that Ň = N , and observe that

Z
(∆ϕ) ∗ N (x) = ∆ϕ(y) N (x − y) dV (y)
Rn Z

= lim ∆ϕ(y) N (x − y) dV (y),

ε↓0 Rn \Bε (x)

the last equality being a consequence of the dominated convergence theorem.

Now we apply the second Green’s identity to the domain BR (0) \ Bε (x),
where R is chosen so large that BR (0) contains the support of ϕ. We see
that
Z
(∆ϕ) ∗ N (x) = lim [ϕ(Y )∂ν N (x − Y ) − ∂ν ϕ(Y ) N (x − Y )] dσ(Y ).
ε↓0 ∂Bε (x)

By arguing much as in the proof of Theorem 1.5.5 (i), we see that

Z
lim ϕ(Y )∂ν N (x − Y ) dσ(Y ) = ϕ(x)
ε↓0 ∂Bε (x)

and Z
lim ∂ν ϕ(Y ) N (x − Y ) dσ(Y ) = 0.
ε↓0 ∂Bε (x)
Therefore
h(∆ϕ) ∗ N, f i = hϕ, f i .
The formula above, together with (5.4.3), shows that ∆(N ∗ f ) = f in the
sense of distributions, as required.
Next we prove (ii). We already know that N ∗ f is a smooth function
on Rn . It remains to prove that
∆(N ∗ f )(x) = f (x) ∀x ∈ Rn .
By (i), ∆(N ∗ f ) = f in the sense of distributions. Since both ∆(N ∗ f )
and f are locally integrable functions (for they are continuous), they must
be equal a.e. Since both ∆(N ∗ f ) and f are continuous, they must agree at
every point, as required. 2

5.5 The Sobolev spaces H k (Ω)

Suppose that I is an open interval of the real line. Various notions of “reg-
ularity” of a distribution u ∈ D0 (I) may be expressed by requiring that its
distributional derivative u0 belongs to a certain function space. Here are some
examples:
146 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

(i) u is in C 1 (I) if and only if u0 is in C(I);

(ii) u is in Lip (I) if and only if u0 is in L∞ (I);
(iii) u is absolutely continuous in I if and only if u0 is in L1 (I);
(iv) u is of bounded variation in I if and only if u0 is in M(I), the class of
all finite (complex) Borel measures on I.

Note that if a distribution u “is” a function, then any function v which is

equal to u a.e. represents the same distribution u. Thus, for instance the
statement (i) above should be interpreted in the following way: u is equal a.e.
to a C 1 (I) function if and only if u0 (in the distributional sense) is equal a.e.
to a continuous functions. Similar comments apply to the other statements
above.
We prove statement (i) above in full details to illustrate the sort of subtleties
which arise when distributional derivatives are considered. The proofs of
statements (ii)-(iv) are similar, and are omitted.
Suppose first that u is a distribution in C 1 (I). This means that there
exists a function v in C 1 (I) which agrees pointwise a.e. with u. We want to
show that the distributional derivative u0 of u agrees pointwise a.e. with a
continuous function C(I). For every ϕ in D(I), we have

hϕ, u0 i = − hϕ0 , ui
Z
= − ϕ0 (x) u(x) dx
ZI
(because u = v a.e.) = − ϕ0 (x) v(x) dx
Z I
(by parts, for ϕ and v are functions in C 1 (I)) = ϕ(x) v 0 (x) dx
I
= hϕ, v 0 i ,

so that u0 = v 0 in the sense of distributions. Hence u0 is a function, which

agrees a.e. with the continuous function v 0 .
Conversely, assume that u0 agrees (as a distribution) with a function w
in C(I). Fix a point x0 in I and define
Z x
v(x) = w(t) dt ∀x ∈ I.
x0

Clearly, v is a distribvtion in C 1 (I) and v 0 = w. Any other distribution

of the form v + c, where c is a constant, satisfies (v + c)0 = w. Let U be
5.5. SOBOLEV SPACES 147

a distribution such that U 0 = w (in the senso of distributions). May we

assert that U agrees with v + c for some constant c? In other words, if the
distributional derivative T 0 of a distribution T vanishes, may we conclude
that T is constant (as a distribution). The answer is “yes”, as the following
result shows.

Proposition 5.5.1 Suppose that I is an open interval of R and that u is

in D0 (I). If u0 = 0 in the sense of distributions, then u is constant (as a
distribution).

Proof. The assumption u0 = 0 is equivalent to hϕ0 , ui = 0 for every ϕ in D(I).

R prove that hψ, ui = 0 for every ψ in D(I) with vanishing integral.
First we
Indeed, if I ψ(x) dx = 0, then its primitive function
Z x
ϕ(x) := ψ(s) ds ∀x ∈ R
−∞

is a function in D(I), and ϕ0 = ψ. Hence hψ, ui = 0 by assumption.

Now, suppose that ψ is R a generic function in D(I). Denote R by ψ
0 a
function in D(I) such that I ψ0 (x) dx = 1. The function ψ − I ψ(x) dx ψ0
is in D(I) and its integral vanishes. By the first part of the proof, it follows
that  Z  
ψ− ψ(x) dx ψ0 , u = 0,
I
i.e. Z 
hψ, ui = ψ(x) dx hψ0 , ui .
I
Therefore u = hψ0 , ui 1I , as required. 2
Coming back to the point, we may now assert that u = v + c (in the sense
of distributions) for some constant c, hence u ∈ C 1 (I), as required.

Exercise 5.5.2 Prove that the function f , defined by

(
x2 sin(1/x2 ) if x 6= 0
f (x) :=
0 if x = 0,
is differentiable (in the classical sense) at every point x ∈ R, and compute its
derivative. Show that the distributional derivative f 0 of f does not agree
with its classical derivative. Prove that
ϕ(x) − ϕ(0)
Z Z
0 2
hϕ, f i = ϕ(x) 2x sin(1/x ) dx − 2 cos(1/x2 ) dx.
R R x
148 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

An important notion of “regularity” of a distribution u is obtained by re-

quiring that one or more derivatives of u belong to some Lp space. We
shall mainly concerned with the case where p = 2, but the other values of p
are important as well. The spaces thus obtained are referred to as Sobolev
spaces. Before going into the details, it may be helpful to discuss the main
motivation behind the introduction of such spaces.
There are many ways of defining the notion of function with derivative
in L2 (I). Most of them are equivalent and useful. There is one which is
definitely not good, i.e., that there exists a function g ∈ L2 (I) such that

f (x + h) − f (x)
→ g(x)
h
pointwise a.e.. The following example illustrate this statement.

Example 5.5.3 Denote by F : [0, 1] → [0, 1] the Cantor function. Extend

F to a continuous function on R, still denoted by F , which is equal to 0 when
x < 0 and equal to 1 when x > 1. We compute the distributional derivative
of F .
It is well known that F is differentiable in the classical sense in R \ C1/3 ,
where C1/3 denotes Cantor’s ternary set, and that the derivative is equal to
0 therein. Denote by µF the Lebesgue–Stieltjes measure on R associated to
F . In particular,

µF (a, b] = F (b) − F (a)

∀a, b ∈ R : a ≤ b.

Now, for every ϕ in D

hϕ, F 0 i = − hϕ0 , F i
Z
= − ϕ0 (x) F (x) dx
ZR Z x
0
=− dx ϕ (x) dµF (y)
R −∞

By Fubini’s theorem,
Z Z ∞
0
hϕ, F i = − dµ (y) F
ϕ0 (x) dx
R y
Z
= ϕ(y) dµF (y)
R
= ϕ, µF .
5.5. SOBOLEV SPACES 149

Therefore F 0 = µF in the sense of distributions. Note that is ϕ ∈ D with

support contained in R \ C1/3 , then hϕ, F 0 i = 0. Therefore, the support of
µF is contained in the Cantor ternary set.

The following result shows that there are at least five equivalent definitions
of L2 -derivative of a function in L2 (Rn ).

Proposition 5.5.4 Suppose that f ∈ L2 (Rn ). The following are equivalent:

(i) the distributional derivative ∂1 f is in L2 (Rn );

(ii) ξ1 fb(ξ) is in L2 (Rn );

(iii) h−1 f (x1 + h, x2 , . . . , xn ) − f (x1 , x2 , . . . , xn ) is convergent in L2 (Rn )
 

as h tends to 0;
(iv) there exists a sequence {fn } of test functions such that
kf − fn kL2 (Rn ) → 0
and {∂1 fn } is a Cauchy sequence in L2 (Rn );
(v) there exists g in L2 (Rn ) such that
Z Z
− ∂1 ϕ f dV = ϕ g dV ∀ϕ ∈ Cc1 (Ω).
Rn Rn

Proof. We prove only that (i) and (v) are equivalent. The rest of the proof
is left to the interested reader.
Suppose that (i) holds. Then
Z Z
ϕ ∂1 f dV = − ∂1 ϕ f dV ∀ϕ ∈ D. (5.5.1)
Rn Rn

We must show that (5.5.1) holds for all ϕ ∈ C 1 (R n 1

R ). For ϕ ∈ C (R ),
n
n
consider ϕ ∗ ψε , where ψ ∈ DB1 (0) (R ) is such that Rn ψ dV = 1, and
ψε (x) := ε−n ψ(x/ε) ∀x ∈ Rn .
Since the support of both ϕ and ψ is compact, so is the support of ϕ ∗ ψε .
Furthermore, ϕ ∗ ψε is smooth. Thus, we may rewrite (5.5.1) with ϕ ∗ ψε in
plave of ϕ, and get
Z Z
(ϕ ∗ ψε ) ∂1 f dV = − ∂1 (ϕ ∗ ψε ) f dV
Rn ZR
n

  (5.5.2)
=− (∂1 ϕ) ∗ ψε f dV.
Rn
150 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Note that

ϕ ∗ ψε − ϕ 2
→0 and (∂1 ϕ) ∗ ψε − ∂1 ϕ 2
→0

as ε tends to 0. Therefore, by taking the limit of both sides of (5.5.2), we

obtain Z Z
ϕ ∂1 f dV = − ∂1 ϕ f dV.
Rn Rn

Since this holds for every ϕ ∈ Cc1 (Rn ), (v) is proved, with g = ∂1 f . Con-
versely, if (v) holds, then, in particular,

− h∂1 ϕ, f i = hϕ, gi ∀ϕ ∈ D.

Thus, g is the distributional derivative ∂1 f of f , whence ∂1 f ∈ L2 (Rn ), as

required. 2

Definition 5.5.5 Suppose that Ω is an open subset of Rn and that α is a

multiindex. We say that a function u ∈ L2 (Ω) admits a weak α-derivative,
if there exists a function g ∈ L2 (Ω) such that
Z Z
|α|
(−1) α
(D ϕ) u dV = ϕ g dV ∀ϕ ∈ Cc|α| (Ω).
Rn Rn

The function g is then called the weak α-derivative of u.

A slight generalisation to α-derivatives on Ω of the equivalence between (i)

and (v) in Proposition 5.5.4 shows that for a function u ∈ L2 (Ω) the following
are equivalent:

(i) u admits weak α-derivative;

(ii) the distributional derivative Dα u of u belongs to L2 (Ω).

We shall use this equivalence without any further comment in the sequel.

If u ∈ C k (Ω), then an integration by parts shows that for every multiin-

dex α with |α| ≤ k the weak α-derivative of u agrees with the classical
derivative Dα u.

Exercise 5.5.6 Prove directly that the function sgn on the real line does
not admit a weak first derivative.
5.5. SOBOLEV SPACES 151

Definition 5.5.7 Suppose that Ω is a bounded open set in Rn . Denote

by C k (Ω) the vector space of all functions u in C k (Ω) such that for every
multi-index α such that |α| ≤ k, Dα u extends to a continuous function on Ω.

Definition 5.5.8 Suppose that Ω is a possibly unbounded open subset

of Rn . We denote by H k (Ω) the space of all complex valued functions u in
L2 (Ω) whose distributional derivatives up to the order k belong to L2 (Ω),
endowed with the norm
X Z 1/2
α 2
kukH k (Ω) = |D u| dλ . (5.5.3)
|α|≤k Ω

Note that this norm is associated to the following natural inner product in
H k (Ω)
XZ
(u, v) := Dα u Dα v dλ.
|α|≤k Ω

More generally, for each p in [1, ∞] we denote by W k,p (Ω) the space of all
complex valued functions u in Lp (Ω) whose distributional derivatives up to
the order k belong to Lp (Ω), endowed with the norm
X Z 1/p
kukW k,p (Ω) = |Dα u|p dλ
|α|≤k Ω

in the case where p < ∞, and with the norm

X
kukW k,∞ (Ω) = kDα uk∞
|α|≤k

in the case where p = ∞.

Clearly W k,2 (Ω) = H k (Ω). In the rest of these notes we focus on the case
where p = 2.

Exercise 5.5.9 Suppose that Ω is a domain in Rn containing the origin.

Prove that the function
rα (x) := |x|−α
is k times weakly differentiable if and only if k + α < n. Does rα belong to
any Sobolev space?
152 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES

Exercise 5.5.10 Consider the function

−1
uα (x) := (1 + x2 )−α/2 log(2 + x2 ) ∀x ∈ R.

Prove that uα ∈ W 1,p (R) for all p ∈ [1/α, ∞] and that uα ∈

/ Lq (R) for all
q ∈ [1, 1/α).

The following results summarises some relevant properties of Sobolev spaces.

For the proof, the reader is referred to [AF, Br, EG, GT, Fo2, Tr].

Proposition 5.5.11 Suppose that Ω is an open subset of Rn . The following

hold:

(i) H k (Ω) is an Hilbert space;

(ii) H k (Ω) ∩ C ∞ (Ω) is dense in H k (Ω).

To state the second result we need the following definition. Suppose that α
is a number in (0, 1]. We say that a function u ∈ C k (Ω) is in C k,α (Ω) if for
every multiindex β of order k, the derivatives Dβ u are Hölder continuous of
order α on Ω.

Theorem 5.5.12 Suppose that Ω is a bounded domain with Lipschitz bound-

ary. The following hold:

(i) H k (Ω) ∩ C ∞ (Ω) is dense in H k (Ω);

(ii) there exists a continuous linear map (called extension operator) E :

H 1 (Ω) → H 1 (Rn ) such that Eu has compact support and Eu = u on Ω;

(iii) the imbedding H 1 (Ω) ,→ L2 (Ω) is a compact operator;

(iv) if m > n/2, then H m (Ω) ,→ C k,α (Ω); here k := m − [n/2] − 1, and, if
n is odd, then α := m − n/2 − k, whereas, if n is even, then α is any
number in (0, 1).

Remark 5.5.13 By Theorem 5.5.12 (i), the Sobolev space H m (Ω) may be
equivalently defined as the completion of C ∞ (Ω) with respect to the norm
(5.5.3).
5.5. SOBOLEV SPACES 153

Remark 5.5.14 For a generic domain Ω, the space C ∞ (Ω) is not dense
in H k (Ω), as the following example in one dimension shows. Denote by
Ω the set (−1/2, 0) ∪ (0, 1/2) in R. We argue by contradiction. Suppose
that C ∞ (Ω) is dense in H 1 (Ω). Denote by f the function, which is equal
to −1 on (−1/2, 0) and equal to 1 on (0, 1/2). Clearly f belongs to H 1 (Ω)
and kf kH 1 (Ω) = kf kL2 (Ω) , because the distributional derivative of f vanishes.
Suppose that {ϕn } is a sequence of functions in C ∞ ([−1/2, 1/2]) that is
convergent to f in H 1 (Ω). Then {ϕn } is convergent to f in L2 (Ω), hence
lim kϕn kL2 (Ω) = kf kL2 (Ω) = 1.
n→∞

By possibly taking a subsequence, which we still denote by {ϕn }, ϕn is con-

vergent to f a.e. Therefore, there exists at least a point xn in (−1/2, 1/2)
such that ϕn (xn ) = 0. By the fundamental theorem of calculus,
ϕn (x) = ϕn (x) − ϕn (xn )
Z x
= ϕ0n dλ,
xn

whence Z 1/2
|ϕn (x)| ≤ |ϕ0n | dλ
−1/2

≤ kϕ0n kL2 (Ω) ∀x ∈ [−1/2, 1/2]

(the last inequality follows from Schwarz’s inequality). Thus, by taking the
supremum over all the points x in [−1/2, 1/2], we obtain
kϕn k∞ ≤ kϕ0n k2 ∀n ∈ N.
But kϕn k2 ≤ kϕn k∞ , so that
kϕn k2 ≤ kϕ0n k2 ∀n ∈ N.
Therefore
kf kH 1 (Ω) = lim kϕn kH 1 (Ω) ≥ 2 lim kϕn k2 = 2 = 2 kf kH 1 (Ω) .
n→∞ n→∞

As a consequence, kf kH 1 (Ω) = 0, which is clearly impossible.

Exercise 5.5.15 Prove that the function u, defined by

1
u(x) = log log ∀x ∈ R \ {0},
|x|
is in H 1 (B1/2 (0)). Therefore H 1 (B1/2 (0)) 6⊂ {bounded functions}.

Exercise 5.5.16 Prove directly that H 1 ((−1, 1)) ⊂ C([−1, 1]). Hint: use
the fundamental theorem of calculus (for the Lebesgue integral).
154 CHAPTER 5. DISTRIBUTIONS AND SOBOLEV SPACES
Chapter 6

The Dirichlet problem with L2

methods

6.1 Poincare’s inequality

One of the cornerstones of the L2 methods for the Dirichlet problem is the
so-called Poincaré inequality, which relates the Dirichlet integral and the
L2 norm of a function in H01 (Ω). We need to introduce the following notation.
Suppose that Ω is a (possibly unbounded) open set in Rn . Given a unit
vector ω in Rn , we denote by δΩ (ω) “the diameter of Ω in the direc-
tion ω”, namely the length of the projection of Ω in the direction of ω.
Then define
δΩ := inf δΩ (ω). (6.1.1)
|ω|=1

Note that δΩ may be finite even if Ω is unbounded. For instance, a strip S

of width a in the plane is obviously unbounded, but δS = a. Notice that in
the following result no assumptions are made on the boundary of the open
set Ω.

Lemma 6.1.1 Suppose that Ω is an open subset of Rn . Then

kϕkL2 (Ω) ≤ δΩ k∇ϕkL2 (Ω) ∀ϕ ∈ H01 (Ω).

Proof. Since H01 (Ω) is the completion of Cc∞ (Ω), it suffices to show that the
required estimate holds for all ϕ ∈ Cc∞ (Ω).
The one dimensional case has already been proved in Proposition 2.2.6.

155
156 CHAPTER 6. DIRICHLET L2

Suppose that n ≥ 2. Choose a unit vector ω in Rn so that δΩ (ω) is finite,

and then choose an orthonormal basis of Rn whose last vector is ω. For x
in Rn write x = (x0 , xn ), where xn is the component of x in the direction ω
and x0 is the component of x in the subspace orthogonal to ω. For each x0
fixed, ϕ(x0 , ·) vanishes outside an interval I(x0 ) of length at most δΩ (ω). By
the one dimensional result
Z Z
0 2 0 2 2
ϕ(x , t) dt ≤ I(x ) ∂xn ϕ(x0 , t) dt.
I(x0 ) I(x0 )

Recall that I(x0 ) ≤ δΩ (ω). By integrating both sides with respect to x0 , we

obtain Z Z
2 2 2
ϕ dV ≤ δΩ (ω) ∂xn ϕ dV
Ω ZΩ
2
≤ δΩ (ω)2 ∇ϕ dV.
Ω
By taking the infimum of both sides with respect to ω, we obtain the required
estimate. 2

Note that we cannot hope to control the L2 norm of a generic smooth function
with the L2 norm of its gradient (just take a nonzero constant function on an
interval of the real line). Thus, the assumption that ϕ vanish at some point
in Ω is really needed.
√
Observe that D is a norm on H01 (Ω). Indeed, observe that if u ∈ H01 (Ω),
then
D(u) = 0 ⇒ u is constant a.e. on each connected component of Ω
⇒ u = 0 a.e. on Ω,

the last implication being justified by the

√fact that γ(u) must vanish, for u
is in H01 (Ω). The next result shows that D is a norm on H01 (Ω) equivalent
to the H 1 (Ω) norm.

Corollary 6.1.2 Suppose that Ω is an open subset of Rn . Then

p p
D(u) ≤ kukH01 (Ω) ≤ (1 + δΩ ) D(u) ∀u ∈ H01 (Ω).

Proof. The left hand inequality follows directly from the definition of kukH 1 (Ω) ;
the right hand inequality is a consequence of Poincaré’s inequality in the pre-
ceding lemma. 2
6.2. SOLUTION TO THE MINIMIZATION PROBLEM 157

Remark 6.1.3 A consequence of Corollary 6.1.2 is that the inner product

induced on H01 (Ω) by restricting that of H 1 (Ω) is equivalent to the following
Z
(u, v)H01 := ∇u · ∇v dV ∀u, v ∈ H01 (Ω). (6.1.2)
Ω

Henceforth, we shall endow H01 (Ω) p

with the inner product (6.1.2), and with
the corresponding Hilbertian norm D(u), which, by abuse of notation, will
still be denoted by kukH01 (Ω) .

6.2 Solution to the minimization problem

In this section we prove (a slight generalisation of) Theorem 4.0.7. In or-
der to formulate the new minimization problem, we need a few preliminary
consideratons.
If X is a complex Banach space of functions on an open set Ω (possibly
n
R ), we denote by Re X the real Banach space of all real valued functions in
X.
Suppose that f and g are real valued functions. We aim at convincing
the reader that any solution of the Dirichlet problem problem
(
∆u = f in Ω
u|∂Ω = g

is real valued. Indeed, if u = u1 + iu2 is a solution, with u1 and u2 real

valued, then u1 and u2 are solutions to the Dirichlet problems
(
∆u1 = f in Ω
u1 |∂Ω = g

and (
∆u2 = 0 in Ω
u2 |∂Ω = 0.
Clearly, u2 = 0, thereby showing that u is real valued.
Thus, when trying to solve the Dirichlet problem with real data, it suffices
to look for solutions in a real Banach space. Note that a fundamental role in
the reasoning below is played by the fact that ∆ preserves the class of real
valued functions, This may fail for more general operators.
158 CHAPTER 6. DIRICHLET L2

We consider the following minimization problem. Given f ∈ Re L2 (Ω)

and G ∈ Re H 1 (Ω), find u in Re H 1 (Ω) that minimizes the functional
Z Z
1 2
J(w) := ∇w dV − f w dV (6.2.1)
2 Ω Ω

subject to the condition

w − G ∈ Re H01 (Ω). (6.2.2)

We follow a recent expository article of N.H. Friedel [Fr]. Note that there
are no restrictions on ∂Ω in the next statement. An estimate of the unique
solution to the minimization problem (4.0.6) is obtained by just setting f = 0
in the next statement.

Theorem 6.2.1 Suppose that Ω is a bounded domain in Rn . The minimiza-

tion problem above has a unique solution u, which satisfies the estimate
q p

kukH 1 ≤ δΩ2 + 1 δΩ kf kL2 + D(G) + kGkH 1 ,

where δΩ is the constant in Poincaré’s inequality for the domain Ω.

Proof. We make the following change of variables:

v := w − G.

Observe that
Z
1 1
J(v + G) = D(v) + Re ∇v · ∇G dV + D(G) − (v, f )L2 − (G, f )L2
2 Ω 2
= J(v)
e + const

where Z
e := 1 D(v) +
J(v) ∇v · ∇G dV − (v, f )L2 , (6.2.3)
2 Ω

and const = 21 D(G) − (f, G)L2 . Thus, w solves the above minimization
problem if and only if v minimizes the functional J,
e subject to the condition

v ∈ Re H01 (Ω).

Denote by Λf,G the linear functional on H01 (Ω) defined by

Z
Λf,G (v) := ∇v · ∇G dV − (v, f )L2 , (6.2.4)
Ω
6.2. SOLUTION TO THE MINIMIZATION PROBLEM 159

Observe that Λf,G is continuous on H01 (Ω). Indeed, by Schwarz’s inequality,

Λf,G (v) ≤ k|∇G|kL2 k|∇v|kL2 + kf kL2 kvkL2

(6.2.5)
≤ kGkH01 + δΩ kf kL2 kvkH01 :

we have used Poincaré’s inequality for the domain Ω in the second inequal-
ity above. Then, by Riesz’s representation theorem, there exists a unique
element vΛf,G ∈ H01 (Ω), with |||Λf,G |||H01 = kvΛf,G kH01 , that represents the
functional Λf,G , i.e.

Λf,G (v) = (v, vΛf,G )H01 ∀v ∈ H01 (Ω).

Since Λf,G is real on Re H01 (Ω), the element vΛf,G must be real valued, i.e.,
vΛf,G is in Re H01 (Ω).
Now, the functional Je on Re H01 (Ω) may be rewritten as follows

e = 1 (v, v)H 1 + (v, vΛ )H 1 .

J(v) f,G
2 0 0

The advantage of this rewriting is that the existence and uniqueness of a

minimizer for Je is now apparent. Indeed, the minimizer is just the same as
the minimizer for the functional
1 1
v 7→ (v, v)H01 + (v, vΛf,G )H01 + (vΛf,G , vΛf,G )H01 = kv + vΛf,G k2H 1 ,
2 2 0

which is obviously equal to −vΛf,G . Thus, the minimizer of the original func-
tional J is u := G − vΛf,G . To conclude the proof of the theorem, it remains
to estimate the norm of u in terms of the data. The triangle inequality and
(6.2.5) imply that
kukH01 ≤ kGkH01 + kvΛf,G kH01
≤ 2kGkH01 + δΩ kf kL2 ,
as required. 2

In view of Theorem 6.2.1, we have a solution map

S : Re L2 (Ω) × Re H 1 (Ω) → Re H 1 (Ω)

which associates to the pair (f, G) the unique solution S(f, G) of the mini-
mization problem with data (f, G). The proof of Theorem 6.2.1 above shows
that
S(f, G) = G − vΛf,G , (6.2.6)
160 CHAPTER 6. DIRICHLET L2

where vΛf,G is the unique function in Re H01 (Ω) that represents the continuous
linear functional
Z
Λf,G (v) = ∇v · ∇G dV − (v, f )L2 .
Ω

Note that if H is a function in Re H 1 (Ω) such that G − H is in Re H01 (Ω),

then
S(f, G) = S(f, H). (6.2.7)
Indeed,
S(f, G) − S(f, H) = G − vΛf,G + vΛf,H − H
= G − H − vΛ0,G−H
= S(0, G − H).
Thus, (6.2.7) is equivalent to the statement that S(0, G − H) = 0. Since
Λ0,G−H is clearly represented as an inner product in Re H01 (Ω) by the function
G − H (which belongs to Re H01 (Ω), we have

G − H − vΛ0,G−H = G − H − (G − H) = 0,

as required.
Thus, if fact, the solution map (f, G) 7→ S(f, G) does not depend on the
element G in Re H 1 (Ω), but on the coset G + Re H01 (Ω) in the coset space
Re H 1 (Ω)/ Re H01 (Ω). Therefore there exists a unique map Se : Re L2 (Ω) ×
Re H 1 (Ω)/ Re H01 (Ω) → Re H 1 (Ω) such that the following diagram

Re L2 (Ω) × Re H 1 (Ω)
Q
Q
Q
Q
I ×π Q
Q
S
Q
Q
? Q
s
Q
2 1
Re L (Ω) × Re H (Ω)/ Re H01 (Ω) - Re H 1 (Ω)
Se

is commutative. Here π denotes the canonical projection of Re H 1 (Ω) onto

Re H 1 (Ω)/ Re H01 (Ω). Thus, Re H 1 (Ω)/ Re H01 (Ω) may be interpreted as the
space of all possible boundary data of the minimization problem (6.2.1).

Now, we know that the minimization problem (6.2.1) subject to the con-
dition (6.2.2) has a unique solution. We would like to relate the minimizimg
6.2. SOLUTION TO THE MINIMIZATION PROBLEM 161

function of the functional J to some sort of solution to the original Dirichlet

problem.
In the proof of Theorem 6.2.1 it is shown that proving the existence of a
minimizing function for the problem (6.2.1) subject to the condition (6.2.2) is
equivalent to proving the existence of a minimizing function for the problem
(6.2.3) subject to the condition v ∈ Re H01 (Ω).
Furthermore, a close examination of the proof of Theorem 6.2.1 shows
that the assumption that f belong to Re L2 (Ω) is used only to ensure that
the linear functional v 7→ (v, f )L2 extends to a bounded linear functional on
H01 (Ω). It is straightforward to check that slight modifications of the proof
of Theorem 6.2.1 also prove the following.

Theorem 6.2.2 Suppose that Ω is a bounded domain in Rn , and that F is

in the dual space Re H01 (Ω)∗ . The minimization problem
h1 i
min1 J(v)
e := min1 D(v) − F (v) (6.2.8)
v∈Re H0 (Ω) v∈Re H0 (Ω) 2

has a unique solution u, which satisfies the estimate

kukH 1 ≤ kF kRe H01 (Ω)∗ .

This result suggest to study the Banach dual of H01 (Ω) in some detail. This
will be done in Section 6.3.
We conclude this section will the following result, which goes in the di-
rection of interpreting the minimizing function of the problem (6.2.8) as a
solution of a differential problem associated to the Laplace operator.

Theorem 6.2.3 Suppose that Ω is a bounded domain in Rn , that F is in

the dual space Re H01 (Ω)∗ and that v is in Re H01 (Ω). The following are
equivalent:

(i) v is a solution to the minimization problem (6.2.8)

(ii) for each ϕ ∈ Re H01 (Ω), the function v satisfies the equation
Z
∇ϕ · ∇v dV = F (ϕ). (6.2.9)
Ω

This result may be regarded as a modified form of the Dirichlet prin-

ciple (compare with Proposition 4.0.1). Note that here we know that the
162 CHAPTER 6. DIRICHLET L2

minimization problem has always a solution. The equation in (ii) is called

the Euler equation associated to the functional J.
e

Proof. Suppose that ε > 0, and observe that

hZ i ε2
J(v + εϕ) − J(v) = ε
e e ∇ϕ · ∇v dV − F (ϕ) + D(ϕ).
Ω 2

It is clear that v is a minimizing function of Je if and only if the left hand

side is nonnegative for every ε > 0 and every ϕ ∈ Re H01 (Ω), and this holds
if and only if the coefficient of ε on the right hand side vanishes for every
ϕ ∈ Re H01 (Ω), i.e., if and only if v satisfies the equation in (ii). 2

6.3 The dual of H01(Ω)

In this section we characterise the Banach dual of H01 (Ω). Note that, by the
Riesz representation theorem, every continuous linear functional on H01 (Ω)
is of the form
u 7→ (u, w)H01
for a suitable w ∈ H01 (Ω). For reasons which will become clear later, it is
important to give another representation of elements in the dual of H01 (Ω).
First, observe that the dual of H01 (Ω) may be viewed as a subspace of the
space of distributions D0 (Ω). Indeed, if Λ is a continuous linear functional
on H01 (Ω), then its restriction to D(Ω) satisfies the following

|Λ(ϕ)| ≤ |||Λ||| kϕkH01 ∀ϕ ∈ D(Ω).

Thus, given a compact subset K of Ω,

|Λ(ϕ)| ≤ |||Λ||| λ(K)1/2 %1,K (ϕ) ∀ϕ ∈ DK (Ω).

Then Λ is a distribution on Ω, by Theorem 5.2.8.

Furthermore, the restriction of Λ to D(Ω) identifies Λ. Indeed, suppose
that Λ is a continuous linear functionals on H01 (Ω) such that Λ(ϕ) = 0 for
all ϕ in D(Ω). Then Λ = 0 on H01 (Ω). Indeed, for every u ∈ H01 (Ω) there
exists a sequence {ϕk } of test functions in Ω such that ku − ϕk kH01 → 0 as k
tends to ∞. Hence
Λ(u) = lim Λ(ϕk ) = 0,
k→∞

as required.
6.3. THE DUAL OF H01 (Ω) 163

Theorem 6.3.1 Suppose that Ω is a bounded open subset of Rn . The fol-

lowing hold:

(i) every distribution in Ω of the form f + nj=0 ∂j gj , where f and gj are

P

in L2 (Ω), j ∈ {1, . . . , n}, extends to a bounded linear functional on

H01 (Ω);

(ii) for every bounded linear functional Λ on H01 (Ω) there exist f and gj in
L2 (Ω), j ∈ {1, . . . , n}, such that
Z Z
Λ(u) = u f dV + u div G dV ∀u ∈ H01 (Ω).
Ω Ω

Proof. First we prove (i). Define

D n
X E
Λ(ϕ) = ϕ, f + ∂j gj ∀ϕ ∈ D(Ω).
j=0

Observe that
D n
X E n
X
ϕ, f + ∂j gj ≤ |hϕ, f i| + h∂j ϕ, gj i
j=0 j=0
n
X
≤ kϕkL2 kf kL2 + k∂j ϕkL2 kgj kL2
j=0
h n
X i
≤ C kϕkH01 kf kL2 + kgj kL2 ∀ϕ ∈ D(Ω).
j=0

We have used Poincaré’s inequality in the last inequality above. We have

proved that
h n
X i
|Λ(ϕ)| ≤ C kϕkH01 kf kL2 + kgj kL2 ∀ϕ ∈ D(Ω).
j=0

Thus, Λ extends to a bounded linear functional on H01 (Ω) and

h n
X i
|||Λ||| ≤ C kf kL2 + kgj kL2 ,
j=0

as required.
164 CHAPTER 6. DIRICHLET L2

Next we prove (ii). Suppose that Λ is a continuous linear functional on

H01 (Ω). By the observation just above the statement of the theorem, Λ may
be interpreted as a distribution. By the Riesz representation theorem there
exists a unique function vΛ in H01 (Ω) such that

hϕ, Λi = (ϕ, vΛ )H01 ∀ϕ ∈ D(Ω).

Observe that
(ϕ, vΛ )H01 = (∇ϕ, ∇vΛ )L2
= − hϕ, div ∇vΛ i
= − hϕ, ∆vΛ i ∀ϕ ∈ D(Ω).
We have proved that

Λ(ϕ) = − hϕ, ∆vΛ i ∀ϕ ∈ D(Ω). (6.3.1)

Thus, Λ, viewed as a distribution, is just −∆vΛ . Hence, Λ has the desired

representation (take f = 0 and G = ∇vΛ ). 2

Definition 6.3.2 The dual of H01 (Ω), endowed with the operator norm,
will be denoted by H −1 (Ω).

Note that the Dirac delta at 0 is in H −1 ((−1, 1)), for it is the distributional
derivative of the Heaviside function.

Exercise 6.3.3 Prove that δ0 is not in H −1 (B1 (0)) for n ≥ 2.

Exercise 6.3.4 Prove that ∇1B1 (0) is in H −1 (B1 (0)) for all n. Which is
the support of ∇1B1 (0) ?

We conclude this section with the following result, which we shall use later.

Lemma 6.3.5 The operator ∆ is an isometric isomorphism between H01 (Ω)

and H −1 (Ω).

Proof. First we show that ∆ is a continuous map from H01 (Ω) to H −1 (Ω).
This is implicit in the proof of Theorem 6.3.1 (ii). However, for the sake of
clarity, we repeat the argument here. Since H01 (Ω) is included in D0 (Ω), ∆u
is a distribution on Ω. For every ϕ in D(Ω)

hϕ, ∆ui = hϕ, div ∇ui

= − h∇ϕ, ∇ui .
6.3. THE DUAL OF H01 (Ω) 165

Since u is in H01 (Ω), the components of ∇u are in L2 (Ω). Therefore, by

Schwarz’s inequality,

hϕ, ∆ui ≤ k∇ϕkL2 (Ω) k∇ukL2 (Ω)

= kϕkH01 (Ω) kukH01 (Ω) .

Here the space H01 (Ω) is endowed with the norm kukH01 (Ω) := k|∇u|kL2 (Ω) ,
which is equivalent to the H 1 (Ω)-norm, as already mentioned. We now take
the supremum of both sides with respect to all test functions ϕ such that
kϕkH01 (Ω) ≤ 1, and obtain

k∆ukH −1 (Ω) = sup hϕ, ∆ui

kϕkH 1 (Ω) ≤1
0 (6.3.2)
≤ kukH01 (Ω) .

Hence ∆ : H01 (Ω) → H0−1 (Ω) with operator norm at most one.

Next we prove that ∆ is injective. Note that

kuk2H 1 (Ω) = k∇uk2L2 (Ω)

0

= hu, −∆ui (6.3.3)

≤ kukH01 (Ω) k∆ukH −1 (Ω) ∀u ∈ H01 (Ω),

which obviously implies

kukH01 (Ω) ≤ k∆ukH −1 (Ω) ∀u ∈ H01 (Ω). (6.3.4)

The injectivity of ∆ follows directly from this inequality, which, together

with (6.3.2), shows that ∆ is an isometry.

We claim that ∆ has closed range. Indeed, suppose that {gk } is a Cauchy
sequence in Ran(∆). Denote by uk the unique (for ∆ is injective) function
in H01 (Ω) such that ∆uk = gk . By (6.3.4) we have that

kuj − uk kH01 (Ω) ≤ C kgj − gk kH −1 (Ω)

for all j and k. Hence {uk } is a Cauchy sequence in H01 (Ω). Since H01 (Ω)
is complete, there exists u in H01 (Ω) such that kuk − ukH01 (Ω) tends to 0 as
j and k tend to infinity. Since ∆ is continuous from H01 (Ω) to H −1 (Ω), it
follows that
∆u = lim gk ,
k→∞

whence the range of ∆ is closed.

166 CHAPTER 6. DIRICHLET L2

Finally, observe that ∆ is onto. Indeed, we have already proved (see

(6.3.1)), that, given Λ ∈ H −1 (Ω), there exists vΛ ∈ H01 (Ω) such that

Λ(ϕ) = − hϕ, ∆vΛ i ∀ϕ ∈ D(Ω).

Since D(Ω) is dense in H01 (Ω), the restriction of Λ to D(Ω) identifies Λ

completely. Thus, Λ, viewed as a distribution, is just ∆(−vΛ ), as required.
By a standard consequence of the open mapping theorem, ∆ is a Banach
space isomorphism between H01 (Ω) and H −1 (Ω), as required to conclude the
proof of the lemma. 2
Here is the main result of this section.

Theorem 6.3.6 Suppose that Ω is a bounded C 2 domain. For each f in

H −1 (Ω), there is a unique function u in H01 (Ω) such that

∆u = f.

Furthermore,
kukH01 (Ω) = kf kH −1 (Ω) .

Proof. Denote by ∆−1 : H −1 (Ω) → H01 (Ω) its inverse (which, by Lemma 6.3.5,
is a continuous linear map). Then, given f in H −1 (Ω), then u := ∆−1 f in
H01 (Ω) and ∆u = f . Thus, u is the unique solution in H01 (Ω) of the Dirichlet
problem associated to the Poisson equation.
The required norm estimate is a consequence of the fact, established
in Lemma 6.3.5, that ∆ is an isometric isomorphism between H01 (Ω) and
H −1 (Ω). 2

6.4 Weak solutions and regularity

In Section 6.2, we have proved that the minimization problem (6.2.1)-(6.2.2)
has a unique solution u in H 1 (Ω). Suppose that f = 0 in (6.2.1). It is natural
to speculate whether u solves the Dirichlet equation ∆u = 0 is the classical
sense. In other words, we ask whether a solution of the minimization problem
(6.2.1)-(6.2.2) is an harmonic function in Ω. In this section, we shall prove
that this is indeed the case.
Recall the definition of distributional solution to the equation

∆u = f (6.4.1)
6.4. WEAK SOLUTIONS AND REGULARITY 167

(see Definition 5.4.8). The following argument leads to an a priori slightly

different definition of solution. Suppose for a moment that f is a continuous
function and that u is a (classical) solution of (6.4.1). Then multiply both
sides of (6.4.1) by a function ϕ in Cc1 (Ω) and integrate on Ω. We obtain
Z Z
ϕ(x) ∆u(x) dx = ϕ(x) f (x) dx.
Ω Ω

Notice that both integrals above are absolutely convergent, because ϕ van-
ishes near the boundary of Ω. We can integrate by parts in the left hand side
and get Z Z
− ∇ϕ(x) · ∇u(x) dx = ϕ(x) f (x) dx. (6.4.2)
Ω Ω

The key observation here is that we do not need to assume that u is in C 2 (Ω)
to give meaning to the left hand side of (6.4.2). For instance, it suffices to
assume that u and its distributional derivatives ∂j u are in L1loc (Ω). Obviously
this is a much weaker requirement than u belongs to C 2 (Ω).

Definition 6.4.1 Suppose that f ∈ L2 (Ω). A locally integrable function

u with locally integrable first derivatives is a weak solution of the equation
∆u = f if (6.4.2) holds for every ϕ ∈ Cc1 (Ω).

Exercise 6.4.2 Suppose that f ∈ L2 (Ω). Prove that a weak solution of

the equation ∆u = f is a distributional solution.

In principle, there might be distributional solutions of the Poisson equation

(6.4.1) with datum f ∈ L2 (Ω) that are not weak solutions. In fact, this is
not the case. Indeed, if u1 and u2 are distributional solutions of the Poisson
equation with datum f ∈ D0 (Ω), then u1 − u2 is a distributional solution to
the Laplace equation. Weyl’s lemma, Lemma 6.4.3 below, shows that u1 = u2
in the sense of distributions.

Lemma 6.4.3 (Weyl’s lemma) Suppose that u is a distributional solution

of the Laplace equation ∆u = 0. Then u is a classical solution (i.e., u is
harmonic in Ω).

Proof. We prove this result under the additional assumption that the distri-
butional solution u be locally integrable. For the proof of the general case,
see [Me, Thm 3.3.14 and Corollary 3.3.15].
168 CHAPTER 6. DIRICHLET L2

Choose a function ϕ ∈ Cc∞ (B1 (0)) such that

R
Rn
ϕ dV = 1, and for every
ε > 0 denote by ϕε the function, defined by
ϕε (x) := ε−n ϕ(x/ε) ∀x ∈ Rn .
Set
Ωε := {x ∈ Rn : Bε (x) ⊂ Ω}.
Clearly Ωε ⊂ Ω. Observe that ϕε ∗ u(x) makes sense for every x ∈ Ωε , and
that ϕε ∗ u is smooth in Ωε . Furthermore, if ψ ∈ D(Ωε ), then ψ ∗ ϕε ∈ D(Ω).
We compute
hψ, ∆(u ∗ ϕε )i = h∆ψ, u ∗ ϕε i
= h(∆ψ) ∗ ϕ̌ε , ui
= h∆(ψ ∗ ϕ̌ε ), ui
= hψ ∗ ϕ̌ε , ∆ui
= 0.
Thus, ∆(u ∗ ϕε ) = 0 on Ωε , i.e., u ∗ ϕε is harmonic in Ωε . Fix Ω0 ⊂ Ω with
compact closure in Ω. By the mean value theorem, for every r small enough
Z
n
u ∗ ϕε (x) = n (u ∗ ϕε )(x + y) dV (y) ∀x ∈ Ω0 . (6.4.3)
r ωn Br (0)
Now, by standard properties of convolution, for every compact subset K of Ω
lim ku ∗ ϕε − ukL1 (K) = 0. (6.4.4)
ε→0

Therefore
sup u ∗ ϕε (x) − u ∗ ϕε0 (x)
x∈Ω0
Z
n
≤ n sup (u ∗ ϕε )(x + y) − (u ∗ ϕε0 )(x + y) dV (y)
r ωn x∈Ω0 Br (0)
n
≤ n sup (u ∗ ϕε )(x + ·) − (u ∗ ϕε0 )(x + ·) L1 (Br (0))
r ωn x∈Ω0
n
≤ n u ∗ ϕε − u ∗ ϕε0 L1 (Ω0 +Br (0))
r ωn
→0
as ε and ε0 tend to 0. Thus, {u ∗ ϕε } is convergent uniformly in Ω0 as ε tends
to 0. By (6.4.4), its uniform limit must be u. Hence u is continuous. We
then deduce from (6.4.3) that for r small enough
Z
n
u(x) = n u(x + y) dV (y) ∀x ∈ Ω0 .
r ωn Br (0)
Therefore u is harmonic in Ω0 by Theorem 1.3.9. 2
Bibliography

[AF] R.A. Adams and J.J.F. Fournier, Sobolev spaces, Pure and Ap-
plied Mathematics Series, vol. 140, Academic Press, IInd edition.

[Br] H. Brezis, Functional analysis, Sobolev spaces and partial differ-

ential equations, Springer-Verlag, 2010.

[DK] B. Dahlberg and C. Kenig, Harmonic Analysis and Partial Differ-

etial Equations Department of Mathematics, Chalmers University
of Technology and the University of Göteborg, 1985/1996.

[DS] N. Dunford and J.T. Schwartz, Linear Operators. Part I. General

Theory, Wiley Classic Library Edition, 1988.

[EG] L. C. Evans, R. F. Gariepy, Measure Theory and Fine Proper-

ties of Functions, Studies in Advanced Mathematics, CRC Press,
Boca Raton, FL, 1992.

[Fo1] G.B. Folland, Real analysis. Modern techniques and their appli-
cations, Second edition, J. Wiley & Sons, 1999.

[Fo2] G.B. Folland, Introduction to Partial Differential Equations,

Princeton University Press, 1976.

[Fr] N.H. Friedel, Short proof of Dirichlet’s principle, arXiv:

1012.5198v1 [math.FA], 2010.

[G] L. Garding, The Dirichlet problem, The Mathematical Intelli-

gencer, 2 (1979), 43–53.

[GT] D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equa-

tions of Second Order Second edition, Springer-Verlag, 1983.

[K] O.D. Kellogg, Foundations of Potential Theory, Dover Publica-

tions, 1953.

169
170 BIBLIOGRAPHY

[Me] S. Meda, Appunti di analisi funzionale, disponibili sul sito del

docente.

[R] J. Rauch, Partial differential equations, Springer Verlag, Berlin,

Heidelberg, New York, 1991.

[RS1] M. Reed and B. Simon, Methods of modern Mathematical Physics.

Vol. I, Academic Press.

[RS2] M. Reed and B. Simon, Methods of modern Mathematical Physics.

Vol. II, Academic Press.

[Ru] W. Rudin, Real and complex analysis, Third Edition, McGraw-

Hill, 1987.

[S] S. Salsa, Equazioni a derivate parziali, Springer-Verlag Italia, IIˆ

edizione, 2010.

[Tr] F. Treves, Basic partial differential equations, Dover Publications,

1975.

[Y] K. Yosida, Functional Analysis, Grundlehren der mathematis-

chen Wissenschaften 123, Springer Verlag, Berlin Heidelberg New
York, 1985.