Introduction to partial differential equations

802635S

Lecture Notes
3rd Edition

Valeriy Serov
University of Oulu
2011

Edited by Markus Harju

Contents
0 Preliminaries 1

1 Local Existence Theory 10

2 Fourier Series 23

3 One-dimensional Heat Equation 32

4 One-dimensional Wave Equation 44

5 Laplace Equation in Rectangle and in Disk 51

6 The Laplace Operator 57

7 The Dirichlet and Neumann Problems 70

8 Layer Potentials 82

9 The Heat Operator 100

10 The Wave Operator 108

Index 119

i
0 Preliminaries
We consider Euclidean space Rn , n ≥ 1 with elements x = (x1 , . . . , xn ). The Euclidean
length of x is defined by q
|x| = x21 + · · · + x2n
and the standard inner product by

(x, y) = x1 y1 + · · · + xn yn .

We use the Cauchy-Schwarz-Bunjakovskii inequality in Rn

|(x, y)| ≤ |x| · |y|.

By BR (x) we denote the ball of radius R > 0 with center x

BR (x) := {y ∈ Rn : |x − y| < R}.

We say that Ω ⊂ Rn , n ≥ 2 is an open set if for any x ∈ Ω there is R > 0 such that

BR (x) ⊂ Ω.

If n = 1 by open set we mean the open interval (a, b), a < b.

An n-tuple α = (α1 , . . . , αn ) of non-negative integers will be called a multi-index .
We define
P
(i) |α| = nj=1 αj

(ii) α + β = (α1 + β1 , . . . , αn + βn ) with |α + β| = |α| + |β|

(iii) α! = α1 ! · · · αn ! with 0! = 1

(iv) α ≥ β if and only if αj ≥ βj for each j = 1, 2, . . . , n. Moreover, α > β if and

only if α ≥ β and there exists j0 such that αj0 > βj0 .

(v) if α ≥ β then α − β = (α1 − β1 , . . . , αn − βn ) and |α − β| = |α| − |β|.

(vi) for x ∈ Rn we define

xα = xα1 1 · · · xαnn
with 00 = 1.
We will use the shorthand notation
∂ ∂ |α|
∂j = , ∂ α = ∂1α1 · · · ∂nαn ≡ .
∂xj ∂xα1 1 · · · ∂xαnn
This text assumes that the reader is familiar also with the following concepts:
1) Lebesgue integral in a bounded domain Ω ⊂ Rn and in Rn .

1
2) Banach spaces (Lp , 1 ≤ p ≤ ∞, C k ) and Hilbert spaces (L2 ): If 1 ≤ p < ∞ then
we set
Z 1/p
p p
L (Ω) := {f : Ω → C measurable : kf kLp (Ω) := |f (x)| dx < ∞}
Ω

while
L∞ (Ω) := {f : Ω → C measurable : kf kL∞ (Ω) := ess sup |f (x)| < ∞}.
x∈Ω

Moreover
X
C k (Ω) := {f : Ω → C : kf kC k (Ω) := max |∂ α f (x)| < ∞},
x∈Ω
|α|≤k

where Ω is the closure of Ω. We say that f ∈ C ∞ (Ω) if f ∈ C k (Ω1 ) for all k ∈ N

and for all bounded subsets Ω1 ⊂ Ω. The space C ∞ (Ω) is not a normed space.
The inner product in L2 (Ω) is denoted by
Z
(f, g)L2 (Ω) = f (x)g(x)dx.
Ω
2
Also in L (Ω), the duality pairing is given by
Z
hf, giL2 (Ω) = f (x)g(x)dx.
Ω

′
3) Hölder’s inequality: Let 1 ≤ p ≤ ∞, u ∈ Lp and v ∈ Lp with
1 1
+ ′ = 1.
p p
Then uv ∈ L1 and
Z Z  p1 Z  1′
p
p p′
|u(x)v(x)|dx ≤ |u(x)| dx |v(x)| dx ,
Ω Ω Ω

where the Hölder conjugate exponent p′ of p is obtained via

p
p′ =
p−1
with the understanding that p′ = ∞ if p = 1 and p′ = 1 if p = ∞.
4) Lebesgue’s theorem about dominated convergence:
Let A ⊂ Rn be measurable and let {fk }∞ k=1 be a sequence of measurable functions
converging to f (x) point-wise in A. If there exists function g ∈ L1 (A) such that
|fk (x)| ≤ g(x) in A, then f ∈ L1 (A) and
Z Z
lim fk (x)dx = f (x)dx.
k→∞ A A

2
 5) Fubini’s theorem about the interchange of the order of integration:
Z Z Z  Z Z 
|f (x, y)|dxdy = dx |f (x, y)|dy = dy |f (x, y)|dx ,
X×Y X Y Y X
if one of the three integrals exists.
Exercise 1. Prove the generalized Leibnitz formula
X
∂ α (f g) = Cαβ ∂ β f ∂ α−β g,
β≤α

where the generalized binomial coefficients are defined as

α!
Cαβ = = Cαα−β .
β!(α − β)!

Hypersurface
A set S ⊂ Rn is called hypersurface of class C k , k = 1, 2, . . . , ∞, if for any x0 ∈ S there
is an open set V ⊂ Rn containing x0 and a real-valued function ϕ ∈ C k (V ) such that
∇ϕ ≡ (∂1 ϕ, . . . ∂n ϕ) 6= 0 on S ∩ V
S ∩ V = {x ∈ V : ϕ(x) = 0} .
By implicit function theorem we can solve the equation ϕ(x) = 0 near x0 to obtain
xn = ψ(x1 , . . . , xn−1 )
k
for some C function ψ. A neighborhood of x0 in S can then be mapped to a piece of
fn = 0 by
the hyperplane x
x 7→ (x′ , xn − ψ(x′ )),
where x′ = (x1 , . . . , xn−1 ). The vector ∇ϕ is perpendicular to S at x ∈ S ∩ V . The
vector ν(x) which is defined as
∇ϕ
ν(x) := ±
|∇ϕ|
is called the normal to S at x. It can be proved that
(∇ψ, −1)
ν(x) = ± p .
|∇ψ|2 + 1
If S is the boundary of a domain Ω ⊂ Rn , n ≥ 2 we always choose the orientation so
that ν(x) points out of Ω and define the normal derivative of u on S by
∂u ∂u
∂ν u := ν · ∇u ≡ ν1 + · · · + νn .
∂x1 ∂xn
Thus ν and ∂ν u are C k−1 functions.
Example 0.1. Let Sr (y) = {x ∈ Rn : |x − y| = r}. Then
n
x−y 1X ∂
ν(x) = and ∂ν = (xj − yj ) .
r r j=1 ∂xj

3
The divergence theorem
Let Ω ⊂ Rn be a bounded domain with C 1 boundary S = ∂Ω and let F be a C 1 vector
field on Ω. Then Z Z
∇ · F dx = F · νdσ(x).
Ω S

Corollary (Integration by parts). Let f and g be C 1 functions on Ω. Then

Z Z Z
∂j f · gdx = − f · ∂j gdx + f · gνj dσ(x).
Ω Ω S

Let f and g be locally integrable functions on Rn , i.e. integrable on any bounded

set from Rn . The convolution f ∗ g of f and g is defined by
Z
(f ∗ g)(x) = f (x − y)g(y)dy = (g ∗ f )(x),
Rn

provided that the integral in question exists. The basic theorem on the existence of
convolutions is the following (Young’s inequality for convolution):
Proposition 1 (Young’s inequality). Let f ∈ L1 (Rn ) and g ∈ Lp (Rn ), 1 ≤ p ≤ ∞.
Then f ∗ g ∈ Lp (Rn ) and
kf ∗ gkLp ≤ kf kL1 kgkLp .
Proof. Let p = ∞. Then
Z Z
|(f ∗ g)(x)| ≤ |f (x − y)||g(y)|dy ≤ kgkL∞ |f (x − y)|dy = kgkL∞ kf kL1 .
Rn Rn

Let 1 ≤ p < ∞ now. Then it follows from Hölder’s inequality and Fubini’s theorem
that
Z Z Z p
p
|(f ∗ g)(x)| dx ≤ |f (x − y)||g(y)|dy dx
Rn Rn Rn
Z Z p/p′ Z
≤ |f (x − y)|dy |f (x − y)||g(y)|p dydx
Rn Rn R n
Z Z
p/p ′
≤ kf kL1 |f (x − y)||g(y)|p dydx
ZR R
n n
Z
p/p′ p
≤ kf kL1 |g(y)| dy |f (x − y)|dx
Rn Rn
p/p′ p/p′ +1
= kf kL1 kgkpLp kf kL1 = kf kL1 kgkpLp .

Thus, we have finally

1/p′ +1/p
kf ∗ gkLp ≤ kf kL1 kgkLp = kf kL1 kgkLp .

4
 1 1 1
Exercise 2. Suppose 1 ≤ p, q, r ≤ ∞ and p
+ q
= r
+ 1. Prove that if f ∈ Lp (Rn )
and g ∈ Lq (Rn ) then f ∗ g ∈ Lr (Rn ) and
kf ∗ gkr ≤ kf kp kgkq .
In particular,
kf ∗ gkL∞ ≤ kf kLp kgkLp′ .
Definition. Let u ∈ L1 (Rn ) with
Z
u(x)dx = 1.
Rn

Then uε (x) := ε−n u(x/ε), ε > 0 is called an approximation to the identity.

Proposition 2. Let uε (x) be an approximation to the identity. Then for any function
ϕ ∈ L∞ (Rn ) which is continuous at {0} we have
Z
lim uε (x)ϕ(x)dx = ϕ(0).
ε→0+ Rn

Proof. Since uε (x) is an approximation to the identity we have

Z Z
uε (x)ϕ(x)dx − ϕ(0) = uε (x)(ϕ(x) − ϕ(0))dx
Rn Rn

and thus
Z  Z
 
 u (x)ϕ(x)dx − ϕ(0) ≤ |uε (x)||ϕ(x) − ϕ(0)|dx
 n ε  √
R |x|≤ ε
Z
+ √
|uε (x)||ϕ(x) − ϕ(0)|dx
|x|> ε
Z Z
≤ sup√
|ϕ(x) − ϕ(0)| |uε (x)|dx + 2 kϕkL∞ √
|uε (x)|dx
|x|≤ ε Rn |x|> ε
Z
≤ sup
√
|ϕ(x) − ϕ(0)| · kukL1 + 2 kϕkL∞ √
|u(y)|dy → 0
|x|≤ ε |y|>1/ ε

as ε → 0.
Example 0.2. Let u(x) be defined as
(
sin x1
2
· · · sin2xn , x ∈ [0, π]n
u(x) =
0, x∈/ [0, π]n .

Then uε (x) is an approximation to the identity and

Z επ Z επ Yn
−n xj
lim(2ε) ··· sin ϕ(x)dx = ϕ(0).
ε→0 0 0 j=1
ε

5
Fourier transform
If f ∈ L1 (Rn ) its Fourier transform fb or F(f ) is the bounded function on Rn defined
by Z
b
f (ξ) = (2π) −n/2
e−ix·ξ f (x)dx.
Rn


Clearly f (ξ) is well-defined for all ξ and fb
b ≤ (2π)−n/2 kf k1 .
∞

The Riemann-Lebesgue lemma

If f ∈ L1 (Rn ) then fb is continuous and tends to zero at infinity.
Proof. Let us first prove that Ff (ξ) is continuous (even uniformly continuous) in Rn .
Indeed,
Z
−n/2
|Ff (ξ + h) − Ff (ξ)| ≤ (2π) |f (x)| · |e−i(x,h) − 1|dx
R n
Z Z
≤ √ |f (x)||x||h|dx + 2 √ |f (x)|dx
|x||h|≤ |h| |x||h|> |h|
p Z
≤ |h| kf kL1 + 2 √ |f (x)|dx → 0
|x|>1/ |h|

as |h| → 0 since f ∈ L1 (Rn ).

To prove that Ff (ξ) → 0 as |ξ| → 0 we proceed as follows. Since eiπ = −1 then
Z Z
−n/2 −i(x,ξ) −n/2 2
2Ff (ξ) = (2π) f (x)e dx − (2π) f (x)e−i(x−πξ/|ξ| ,ξ) dx
ZR
n
ZR
n

= (2π)−n/2 f (x)e−i(x,ξ) dx − (2π)−n/2 f (y + πξ/|ξ|2 )e−i(y,ξ) dy

R n R n
Z
= −(2π)−n/2 (f (x + πξ/|ξ|2 ) − f (x))e−i(x,ξ) dx.
Rn

Hence
Z
−n/2
2|Ff (ξ)| ≤ (2π) |f (x + πξ/|ξ|2 ) − f (x)|dx
Rn

= (2π)−n/2 f (· + πξ/|ξ|2 ) − f (·) L1 → 0

as |ξ| → ∞ since f ∈ L1 (Rn ).

Exercise 3. Prove that if f, g ∈ L1 (Rn ) then f[

∗ g = (2π)n/2 fbgb.

Exercise 4. Suppose f ∈ L1 (Rn ). Prove that

1. If fh (x) = f (x + h) then fbh = eih·ξ fb.

6
 2. If T : Rn → Rn is linear and invertible then f[
◦ T = |det T |−1 fb((T −1 )′ ξ), where
′
T is the adjoint matrix.

3. If T is rotation, that is T ′ = T −1 (and |det T | = 1) then f[

◦ T = fb ◦ T .
Exercise 5. Prove that
\α f ,
∂ α fb = (−ix) ∂d
α f = (iξ)α fb.

Exercise 6. Prove that if f, g ∈ L1 (Rn ) then

Z Z
f (ξ)b
g (ξ)dξ = fb(ξ)g(ξ)dξ.
Rn Rn

For f ∈ L1 (Rn ) define the inverse Fourier transform of f by

Z
−1 −n/2
F f (x) = (2π) eix·ξ f (ξ)dξ.
Rn

It is clear that
F −1 f (x) = Ff (−x), F −1 f = F(f )
and for f, g ∈ L1 (Rn )
(Ff, g)L2 = (f, F −1 g)L2 .
The Schwartz space S(Rn ) is defined as
 
n ∞ n α β
S(R ) = f ∈ C (R ) : sup |x ∂ f (x)| < ∞, for any multi-indices α and β .
x∈Rn

The Fourier inversion formula

If f ∈ S(Rn ) then (F −1 F)f = f .
Exercise 7. Prove the Fourier inversion formula for f ∈ S(Rn ).

The Plancherel theorem

The Fourier transform on S extends uniquely to a unitary isomorphism of L2 (Rn ) onto
itself, i.e.
b
f = kf k2 .
2
This formula is called the Parseval equality.
The support of a function f : Rn → C, denoted by supp f , is the set

supp f = {x ∈ Rn : f (x) 6= 0}.

Exercise 8. Prove that if f ∈ L1 (Rn ) has compact support then fb extends to an entire
holomorphic function on Cn .

7
Exercise 9. Prove that if f ∈ C0∞ (Rn ) i.e. f ∈ C ∞ (Rn ) with compact support, is
supported in {x ∈ Rn : |x| ≤ R} then for any multi-index α we have
|(iξ)α fˆ(ξ)| ≤ (2π)−n/2 eR|Im ξ| k∂ α f k1 ,

that is, fb(ξ) is rapidly decaying as |Re ξ| → ∞ when |Im ξ| remains bounded.

Distributions
We say that ϕj → ϕ in C0∞ (Ω), Ω ⊂ Rn open, if ϕj are all supported in a common
compact set K ⊂ Ω and
sup |∂ α ϕj (x) − ∂ α ϕ(x)| → 0, j→∞
x∈K

for all α. A distribution on Ω is a linear functional u on C0∞ (Ω) that is continuous, i.e.,
1. u : C0∞ (Ω) → C. The action of u to ϕ ∈ C0∞ (Ω) is denoted by hu, ϕi. The set of
all distributions is denoted by D′ (Ω).
2. hu, c1 ϕ1 + c2 ϕ2 i = c1 hu, ϕ1 i + c2 hu, ϕ2 i
3. If ϕj → ϕ in C0∞ (Ω) then hu, ϕj i → hu, ϕi in C as j → ∞. It is equivalent to the
following condition: for any K ⊂ Ω there is a constant CK and an integer NK
such that for all ϕ ∈ C0∞ (K),
X
|hu, ϕi| ≤ CK k∂ α ϕk∞ .
|α|≤NK

Remark. If u ∈ L1loc (Ω), Ω ⊂ Rn open, then u can be regarded as a distribution (in

that case a regular distribution) as follows:
Z
hu, ϕi := u(x)ϕ(x)dx, ϕ ∈ C0∞ (Ω).
Ω

The Dirac δ-function

The δ-function is defined as
hδ, ϕi = ϕ(0), ϕ ∈ C0∞ (Ω).
It is not a regular distribution.
Example 0.3. Let uε (x) be an approximation to the identity. Then
Z Z
−n/2 −n −i(x,ξ) −n/2
ubε (ξ) = (2π) ε u(x/ε)e dx = (2π) u(y)e−i(y,eξ) dy = u
b(εξ).
Rn Rn

In particular,
b(εξ) = (2π)−n/2 .
lim ubε (ξ) = lim u
ε→0+ ε→0+

Applying Proposition 2 we may conclude that

8
 1) limε→0+ huε , ϕi = ϕ(0) i.e. limε→0+ uε = δ in the sense of distributions, and

2) δb = (2π)−n/2 · 1.

We can extend the operations from functions to distributions as follows:

h∂ α u, ϕi = hu, (−1)|α| ∂ α ϕi,

hf u, ϕi = hu, f ϕi, f ∈ C ∞ (Ω),

e
hu ∗ ψ, ϕi = hu, ϕ ∗ ψi, ψ ∈ C0∞ (Ω),
e
where ψ(x) = ψ(−x). It is possible to show that u ∗ ψ is actually a C ∞ function and

∂ α (u ∗ ψ) = u ∗ ∂ α ψ.

A tempered distribution is a continuous linear functional on S(Rn ). In addition to the

preceding operations for the tempered distributions we can define the Fourier transform
by
hb
u, ϕi = hu, ϕi,
b ϕ ∈ S.

Exercise 10. Prove that if u is a tempered distribution and ψ ∈ S then

u[ bu.
∗ ψ = (2π)n/2 ψb

Exercise 11. Prove that

1. δb = (2π)−n/2 · 1, b
1 = (2π)n/2 δ

2. ∂d
α δ = (iξ)α (2π)−n/2

cα = i|α| ∂ α (b
3. x 1) = i|α| (2π)n/2 ∂ α δ.

9
1 Local Existence Theory
A partial differential equation of order k ∈ N is an equation of the form


F x, (∂ α u)|α|≤k = 0, (1.1)

where F is a function of the variables x ∈ Ω ⊂ Rn , n ≥ 2, Ω an open set, and (uα )|α|≤k .

A complex-valued function u(x) on Ω is a classical solution of (1.1) if the derivatives
α
∂ u occurring in F exist on Ω and


F x, (∂ α u(x))|α|≤k = 0

pointwise for all x ∈ Ω. The equation (1.1) is called linear if it can be written as
X
aα (x)∂ α u(x) = f (x) (1.2)
|α|≤k

for some known functions aα and f . In this case we speak about the (linear) differential
operator X
L(x, ∂) ≡ aα (x)∂ α
|α|≤k

and write (1.2) simply as Lu = f. If the coefficients aα (x) belong to C ∞ (Ω) we can apply
the operator L to any distribution u ∈ D′ (Ω) and u is called a distributional solution
(or weak solution) of (1.2) if the equation (1.2) holds in the sense of distributions, i.e.
X
(−1)|α| hu, ∂ α (aα ϕ)i = hf, ϕi,
|α|≤k

where ϕ ∈ C0∞ (Ω). Let us list some examples. Here and throughout we denote
2
ut = ∂u
∂t
, utt = ∂∂t2u and so forth.

1. The eikonal equation

|∇u|2 = c2 ,
where ∇u = (∂1 u, . . . , ∂n u) is the gradient of u.

2. a) Heat (or evolution) equation

ut = k∆u

b) Wave equation
utt = c2 ∆u
c) Poisson equation
∆u = f,
where ∆ ≡ ∇ · ∇ = ∂12 + · · · + ∂n2 is the Laplacian (or the Laplace operator ).

10
 3. The telegrapher’s equation

utt = c2 ∆u − αut − m2 u

4. Sine-Gordon equation
utt = c2 ∆u − sin u

5. The biharmonic equation

∆2 u ≡ ∆(∆u) = 0

6. The Korteweg-de Vries equation

ut + cu · ux + uxxx = 0.

In the linear case, a simple measure of the ”strength” P of a differential operator is

α
provided by the notion of characteristics. If L(x, ∂) = |α|≤k aα (x)∂ then its charac-
teristic form (or principal symbol ) at x ∈ Ω is the homogeneous polynomial of degree
k defined by X
χL (x, ξ) = aα (x)ξ α , ξ ∈ Rn .
|α|=k

A nonzero ξ is called characteristic for L at x if χL (x, ξ) = 0 and the set of all such ξ
is called the characteristic variety of L at x, denoted by charx (L). In other words,

charx (L) = {ξ 6= 0 : χL (x, ξ) = 0} .

In particular, L is said to be elliptic at x if charx (L) = ∅ and elliptic in Ω if it is elliptic

at every x ∈ Ω.
Example 1.1. 1. L = ∂1 ∂2 , charx (L) = {ξ ∈ R2 : ξ1 = 0 or ξ2 = 0, ξ12 + ξ22 > 0} .

2. L = 12 (∂1 + i∂2 ) is the Cauchy-Riemann operator on R2 . It is elliptic in R2 .

3. L = ∆ is elliptic in Rn .
P
4. L = ∂1 − nj=2 ∂j2 , charx (L) = {ξ ∈ Rn \ {0} : ξj = 0, j = 2, 3, . . . , n} .
P n P o
5. L = ∂12 − nj=2 ∂j2 , charx (L) = ξ ∈ Rn \ {0} : ξ12 = nj=2 ξj2 .

Let ν(x) be the normal to S at x. A hypersurface S is called characteristic for L

at x ∈ S if ν(x) ∈ charx (L), i.e.

χL (x, ν(x)) = 0

and S is called non-characteristic if it is not characteristic at any point, that is, for
any x ∈ S
χL (x, ν(x)) 6= 0.

11
Let us consider the linear equation of the first order
n
X
Lu ≡ aj (x)∂j u + b(x)u = f (x), (1.3)
j=1

where aj , b and f are assumed to be C 1 functions of x. We assume also that aj , b and f

are real-valued. Suppose we wish to find a solution u of (1.3) with given initial values
u = g on the hypersurface S (g is also real-valued). It is clear that
n o
~·ξ =0 ,
charx (L) = ξ 6= 0 : A

~ = (a1 , . . . , an ). It implies that charx (L) ∪ {0} is the hyperplane orthogonal to

where A
~ and therefore, S is characteristic at x if and only if A
A ~ is tangent to S at x (A
~ · ν = 0).
Then n n
X X
aj (x)∂j u(x) = aj (x)∂j g(x), x ∈ S,
j=1 j=1

is completely determined as certain directional derivatives of ϕ (see the definition of

S) along S at x, and it may be impossible to make it equal to f (x) − b(x)u(x) (in
order to satisfy (1.3)). Indeed, let us assume that u1 and u2 have the same value g on
S. This means that u1 − u2 = 0 on S or (more or less equivalently)

u1 − u2 = ϕ · γ,

where ϕ = 0 on S (ϕ defines this surface) and γ 6= 0 on S. Next,

~ · ∇)u1 − (A
(A ~ · ∇)u2 = (A
~ · ∇)(ϕγ) = γ(A
~ · ∇)ϕ + ϕ(A
~ · ∇)γ = 0,

~ · ∇)ϕ = 0 ⇔ (A
since S is characteristic for L ((A ~ · ∇ )ϕ = 0 ⇔ A ~ · ν = 0). That’s
|∇|
why to make the initial value problem well-defined we must assume that S is non-
characteristic for this problem.
Let us assume that S is non-characteristic for L and u = g on S. We define the
integral curves for (1.3) as the parametrized curves x(t) that satisfy the system

~
ẋ = A(x), x = x(t) = (x1 (t), . . . , xn (t)) (1.4)

of ordinary differential equations, where

ẋ = (x′1 (t), . . . , x′n (t)).

Along one of those curves a solution u of (1.3) must satisfy

X ∂u n
du d ~ · ∇)u = f − bu ≡ f (x(t)) − bu(x(t))
= (u(x(t))) = ẋj = (A
dt dt j=1
∂x j

12
or
du
= f − bu. (1.5)
dt
By the existence and uniqueness theorem for ordinary differential equations there is a
unique solution (unique curve) of (1.4) with x(0) = x0 . Along this curve the solution
u(x) of (1.3) must be the solution of (1.5) with u(0) = u(x(0)) = u(x0 ) = g(x0 ).
Moreover, since S is non-characteristic, x(t) ∈
/ S for t 6= 0, at least for small t, and the
curves x(t) fill out a neighborhood of S. Thus we have proved the following theorem.

Theorem 1. Assume that S is a surface of class C 1 which is non-characteristic for

(1.3), and that aj , b, f and g are C 1 and real-valued functions. Then for any sufficiently
small neighborhood U of S in Rn there is a unique solution u ∈ C 1 of (1.3) on U that
satisfies u = g on S.

Remark. The method which was presented above is called the method of characteristics.
Let us consider some examples where we apply the method of characteristics.

Example 1.2. In R3 , solve x1 ∂1 u + 2x2 ∂2 u + ∂3 u = 3u with u = g(x1 , x2 ) on the plane

x3 = 0.
Since S = {x ∈ R3 : x3 = 0} then ν(x) = (0, 0, 1) and since χL (x, ξ) = x1 ξ1 +
2x2 ξ2 + ξ3 we have

χL (x, ν(x)) = x1 · 0 + 2x2 · 0 + 1 · 1 = 1 6= 0

so that S is non-characteristic. The system (1.4)-(1.5) to be solved is

ẋ1 = x1 , ẋ2 = 2x2 , ẋ3 = 1, u̇ = 3u

with initial conditions

(x1 , x2 , x3 )|t=0 = (x01 , x02 , 0), u(0) = g(x01 , x02 )

on S. We obtain

x1 = x01 et , x2 = x02 e2t , x3 = t, u = g(x01 , x02 )e3t .

These equations imply

x01 = x1 e−t = x1 e−x3 , x02 = x2 e−2t = x2 e−2x3 .

Therefore
u(x) = u(x1 , x2 , x3 ) = g(x1 e−x3 , x2 e−2x3 )e3x3 .

Example 1.3. In R3 , solve ∂1 u + x1 ∂2 u − ∂3 u = u with u(x1 , x2 , 1) = x1 + x2 .

Since S = {x ∈ R3 : x3 = 1} then ν(x) = (0, 0, 1). That’s why

χL (x, ν(x)) = 1 · 0 + x1 · 0 − 1 · 1 = −1 6= 0

13
and S is non-characteristic. The system (1.4)-(1.5) for this problem becomes

ẋ1 = 1, ẋ2 = x1 , ẋ3 = −1, u̇ = u

with
(x1 , x2 , x3 )|t=0 = (x01 , x02 , 1), u(0) = x01 + x02 .
We obtain
t2
x1 = t + x01 , x2 = + tx01 + x02 , x3 = −t + 1, u = (x01 + x02 )et .
2
Then,
t = 1 − x3 , x01 = x1 − t = x1 + x3 − 1,
(1 − x3 )2 1 x2
x02 = x2 − − (1 − x3 )(x1 + x3 − 1) = − x1 + x2 − x3 + x1 x3 + 3
2 2 2
and, finally,  
x23 1
u= + x1 x3 + x2 − e1−x3 .
2 2
Now let us generalize this technique to quasi-linear equations or to the equations
of the form n
X
aj (x, u)∂j u = b(x, u), (1.6)
j=1

where aj , b and u are real-valued. If u is a function of x, the normal to the graph of u

in Rn+1 is proportional to (∇u, −1), so (1.6) just says that the vector field

~ y) := (a1 , . . . , an , b) ∈ Rn+1
A(x,

is tangent to the graph y = u(x) at any point. This suggests that we look at the
~ in Rn+1 given by solving the ordinary differential equations
integral curves of A

ẋj = aj (x, y), j = 1, 2, . . . , n, ẏ = b(x, y).

Suppose u is a solution of (1.6). If we solve

ẋj = aj (x, u(x)), j = 1, 2, . . . , n,

with xj (0) = x0j then setting y(t) = u(x(t)) we obtain that

n
X n
X
ẏ = ∂j u · ẋj = aj (x, u)∂j u = b(x, u) = b(x, y).
j=1 j=1

Suppose we are given initial data u = g on S. If we form the submanifold

S ∗ := {(x, g(x)) : x ∈ S}

14
in Rn+1 then the graph of the solution should be the hypersurface generated by the
integral curves of A~ passing through S ∗ . Again, we need to assume that S is non-
characteristic in a sense that the vector

(a1 (x, g(x)), . . . , an (x, g(x))) , x ∈ S,

should not be tangent to S at x. If S is represented parametrically by a mapping

~ : Rn−1 → Rn (for example ϕ
ϕ ~ (x1 , . . . , xn−1 ) = (x1 , . . . , xn−1 , ψ(x1 , . . . , xn−1 ))) and we
have the coordinates x′ = (x1 , . . . , xn−1 ) ∈ Rn−1 this condition is just
 ∂ϕ1 ∂ϕ1 ′ ′

∂x1
. . . ∂xn−1
a 1 (~
ϕ (x ), g(~
ϕ (x )))
 
det  ... ... ..
.
..
.  6= 0.
∂ϕn ∂ϕn
∂x1
... ∂xn−1
ϕ(x′ ), g(~
an (~ ϕ(x′ )))

Remark. If S is parametrized as

xn = ψ(x1 , . . . , xn−1 ), x′ = (x1 , . . . , xn−1 ) ∈ S ′ ⊂ Rn−1

then S can be represented also by

φ(x1 , . . . , xn ) = 0,

where φ(x1 , . . . , xn ) ≡ ψ(x′ ) − xn and ν(x) is proportional to

 
∂ψ ∂ψ
∇φ = ,..., , −1 .
∂x1 ∂xn−1

Then S is non-characteristic if and only if

∂ψ ∂ψ
a1 + · · · + an−1 − an 6= 0
∂x1 ∂xn−1
or  
1 0 ··· 0 a1 (x, g(x))
 0 1 ··· 0 a2 (x, g(x)) 
 
 .. .. ... .. .. 
det  . . . .  6= 0,
 
 0 0 ··· 1 an−1 (x, g(x)) 
∂ψ ∂ψ
∂x1
··· ··· ∂xn−1
an (x, g(x))
where x ∈ S.

Example 1.4. In R2 , solve u∂1 u + ∂2 u = 1 with u = s/2 on the segment x1 = x2 = s,

where s > 0, s 6= 2 is a parameter.
~ (s) = (s, s) then (x′ = x1 = s)
Since ϕ
 ∂x1   
a1 (s, s, s/2) 1 s/2
det ∂x2 ∂s = det = 1 − s/2 6= 0,
∂s
a2 (s, s, s/2) 1 1

15
for s > 0, s 6= 2. The system (1.4)-(1.5) for this problem is

ẋ1 = u, ẋ2 = 1, u̇ = 1

with
x01
(x1 , x2 , u)|t=0 = (x01 , x02 , ) = (s, s, s/2).
2
Then
u = t + s/2, x2 = t + s, ẋ1 = t + s/2
t2 st
so that x1 = 2
+ 2
+ s. This implies

x1 − x2 = t2 /2 + t(s/2 − 1).

For s and t in terms of x1 and x2 we obtain

 
s 1 t2 2(x1 − x2 )
=1+ x1 − x2 − , t= .
2 t 2 x2 − 2

Hence
2(x1 − x2 ) x1 − x2 t
u = +1+ −
x2 − 2 t 2
2(x1 − x2 ) x2 − 2 x1 − x2
= +1+ −
x2 − 2 2 x2 − 2
x1 − x2 x2 − 2 x1 − x2 x2
= +1+ = +
x2 − 2 2 x2 − 2 2
2x1 − 4x2 + x22
= .
2(x2 − 2)

Exercise 12. In R2 , solve x21 ∂1 u + x22 ∂2 u = u2 with u ≡ 1 when x2 = 2x1 .

Exercise 13. In R2 , solve u∂1 u + x2 ∂2 u = x1 with u(x1 , 1) = 2x1 .

Example 1.5. Consider the Burgers equation

u∂1 u + ∂2 u = 0

in R2 with u(x1 , 0) = h(x1 ), where h is a known C 1 function. It is clear that S :=

{x ∈ R2 : x2 = 0} is non-characteristic for this quasi-linear equation, since
 
1 h(x1 )
det = 1 6= 0,
0 1

and ν(x) = (0, 1). Now we have to solve the ordinary differential equations

ẋ1 = u, ẋ2 = 1, u̇ = 0

16
with

(x1 , x2 , u)|t=0 = x01 , 0, h(x01 ) .
We obtain
x2 = t, u ≡ h(x01 ), x1 = h(x01 )t + x01
so that
x1 − x2 h(x01 ) − x01 = 0.
Let us assume that
−x2 h′1 (x01 ) − 1 6= 0.
By this condition last equation defines an implicit function x01 = g(x1 , x2 ). That’s why
the solution u of the Burgers equation has the form

u(x1 , x2 ) = h(g(x1 , x2 )).

Let us consider two particular cases:

1. If h(x01 ) = ax01 + b, a 6= 0, then

ax1 + b 1
u(x1 , x2 ) = , x2 6= − .
ax2 + 1 a

2. If h(x01 ) = a(x01 )2 + bx01 + c, a 6= 0, then

p !
−x2 b − 1 + (x2 b + 1)2 − 4ax2 (cx2 − x1 )
u(x1 , x2 ) = a
2ax2
p !
2
−x2 b − 1 + (x2 b + 1) − 4ax2 (cx2 − x1 )
+ b + c,
2ax2

with D = (x2 b + 1)2 − 4ax2 (cx2 − x1 ) > 0.

Let us consider again the linear equation (1.2) of order k i.e.

X
aα (x)∂ α u(x) = f (x).
|α|≤k

Let S be a hypersurface of class C k . If u is a C k function defined near S, the quantities

u, ∂ν u, . . . , ∂νk−1 u (1.7)

on S are called the Cauchy data of u on S. And the Cauchy problem is to solve (1.2)
with the Cauchy data (1.7). We shall consider Rn , n ≥ 2, as Rn−1 × R and denote the
coordinates by (x, t), where x = (x1 , . . . , xn−1 ). We can make a change of coordinates
from Rn to Rn−1 × R so that x0 ∈ S is mapped to (0, 0) and a neighborhood of x0 in

17
 ∂
S is mapped into the hyperplane t = 0. In that case ∂ν = ∂t on S = {(x, t) : t = 0}
and equation (1.2) can be written in the new coordinates as
X
aα,j (x, t)∂xα ∂tj u = f (x, t) (1.8)
|α|+j≤k

with the Cauchy data

∂tj u(x, 0) = ϕj (x), j = 0, 1, . . . , k − 1. (1.9)
Since the normal ν = (0, 0, . . . , 0, 1) then the assumption ”S is non-characteristic”
means that
χL (x, 0, ν(x, 0)) ≡ a0,k (x, 0) 6= 0.
Hence by continuity a0,k (x, t) 6= 0 for small t, and we can solve (1.8) for ∂tk u:
 
X
∂tk u(x, t) = (a0,k (x, t))−1 f − aα,j ∂xα ∂tj u (1.10)
|α|+j≤k,j<k

with the Cauchy data (1.9).

Example 1.6. The line t = 0 is characteristic for ∂x ∂t u = 0 in R2 . That’s why
we will have some problems with the solutions. Indeed, if u is a solution of this
equation with Cauchy data u(x, 0) = g0 (x) and ∂t u(x, 0) = g1 (x) then ∂x g1 = 0,
that is, g1 ≡ constant. Thus the Cauchy problem is not solvable in general. On
the other hand, if g1 is constant, then there is no uniqueness, because we can take
u(x, t) = g0 (x) + f (t) with any f (t) such that f (0) = 0 and f ′ (0) = g1 .
Example 1.7. The line t = 0 is characteristic for ∂x2 u − ∂t u = 0 in R2 . Here if we are
given u(x, 0) = g0 (x) then ∂t u(x, 0) is already completely determined by ∂t u(x, 0) =
g0′′ (x). So, again the Cauchy problem has ”bad” behaviour.
Let us now formulate and give ”a sketch” of the proof of the famous Cauchy-
Kowalevski theorem for linear case.
Theorem 2. If aα,j (x, t), ϕ0 (x), . . . , ϕk−1 (x) are analytic near the origin in Rn , then
there is a neighborhood of the origin on which the Cauchy problem (1.10)-(1.9) has a
unique analytic solution.
Proof. The uniqueness of analytic solution follows from the fact that an analytic func-
tion is completely determined by the values of its derivatives at one point (see the
Taylor formula or the Taylor series). Indeed, for all α and j = 0, 1, . . . , k − 1
∂xα ∂tj u(x, 0) = ∂xα ϕj (x).
That’s why
 
X
∂tk u|t=0 = (a0,k )−1 f (x, 0) − aα,j (x, 0)∂xα ϕj (x)
|α|+j≤k,j<k

18
and moreover
 
X
∂tk u(x, t) = (a0,k )−1 f (x, t) − aα,j (x, t)∂xα ∂tj u .
|α|+j≤k,j<k

Then all derivatives of u can be defined from this equation by

∂tk+1 u = ∂t ∂tk u .

Next, let us denote by yα,j = ∂xα ∂tj u and by Y = (yα,j ) this vector. Then equation
(1.10) can be rewritten as
 
X
y0,k = (a0,k )−1 f − aα,j yα,j 
|α|+j≤k,j<k

or  
X
∂t (y0,k−1 ) = (a0,k )−1 f − aα,j ∂xj y(α−~j),j 
|α|+j≤k,j<k

and therefore the Cauchy problem (1.10)-(1.9) becomes

( Pn−1
∂t Y = j=1 Aj ∂xj Y + B
(1.11)
Y (x, 0) = Φ(x), x ∈ Rn−1 ,

where Y, B and Φ are analytic vector-valued functions and Aj ’s are analytic matrix-
valued functions. Without loss of generality we can assume that Φ ≡ 0. Let Y =
(j)
(y1 , . . . , yN ), B = (b1 , . . . , bN ), Aj = (aml )N
m,l=1 . We seek a solution Y = (y1 , . . . , yN ) in
the form X (m)
ym = Cα,j xα tj , m = 1, 2, . . . , N.
(m)
The Cauchy data tell us that Cα,0 = 0 for all α and m, since we assumed Φ ≡ 0. To
(m)
determine Cα,j for j > 0, we substitute ym into (1.11) and get for m = 1, 2, . . . , N
X (j)
∂ t ym = aml ∂xj yl + bm (x, y)

or X X X  (j)  X (m) X
(m) ~
Cα,j jxα tj−1 = aml xβ t r Cα,j αj xα−j tj + b(m) α j
αj x t .
βr
j,l β,r

(m)
It can be proved that this equation determines uniquely the coefficients Cα,j and
therefore the solution Y = (y1 , . . . , yN ).

19
Remark. Consider the following example in R2 , due to Hadamard, which sheds light
on the Cauchy problem:
√
∆u = 0, u(x1 , 0) = 0, ∂2 u(x1 , 0) = ke− k
sin(x1 k), k ∈ N.

This problem is non-characteristic on R2 since ∆ is elliptic in R2 . We look for

u(x1 , x2 ) = u1 (x1 )u2 (x2 ). Then

u′′1 u2 + u′′2 u1 = 0

which implies that

u′′1 u′′
= − 2 = −λ = constant.
u1 u2
Next, the general solutions of
u′′1 = −λu1
and
u′′2 = λu2
are √ √
u1 = A sin( λx1 ) + B cos( λx1 )
and √ √
u2 = C sinh( λx2 ) + D cosh( λx2 ),
√
respectively.
√ But u2√(0) = 0, u′2 (0) = 1 and u1 (x1 ) = ke− k sin(kx1 ). Thus D = 0, B =
0, k = λ, A = ke− k and C = k1 = √1λ . So we finally have
√ 1 √
u(x1 , x2 ) = ke− k
sin(kx1 ) sinh(kx2 ) = e− k sin(kx1 ) sinh(kx2 ).
k
As k → +∞, the Cauchy √data and their derivatives (for x2 = 0) of all orders tend
uniformly to zero since e− k decays faster than polynomially. But if x2 6= 0 (more
precisely, x2 > 0) then
√
lim e− k
sin(kx1 ) sinh(kx2 ) = ∞,
k→+∞

at least for some x1 and some subsequence of k. Hence u(x1 , x2 ) is not bounded. But
the solution of the original problem which corresponds to the limiting case k = ∞ is
of course u ≡ 0, since u(x1 , 0) = 0 and ∂2 u(x1 , 0) = 0 in the limiting case. Hence the
solution of the Cauchy problem may not depend continuously on the Cauchy data. It
means by Hadamard that the Cauchy problem for elliptic operators is ”ill-posed”, even
in the case when this problem is non-characteristic.
Remark. This example of Hadamard shows that the solution of the Cauchy problem
may not depend continuously on the Cauchy data. By the terminology of Hadamard
”the Cauchy problem for the Laplacian is not well-posed or it is ill-posed”. Due to
Hadamard and Tikhonov any problem is called well-posed if the following are satisfied:

20
 1. existence

2. uniqueness

3. stability or continuous dependence on data

Otherwise it is called ill-posed .

Let us consider one more important example due to H. Lewy. Let L be the differ-
ential operator of the first order in R3 ((x, y, t) ∈ R3 ) given by

∂ ∂ ∂
L≡ +i − 2i(x + iy) . (1.12)
∂x ∂y ∂t

Theorem 3 (The Hans Lewy example). Let f be a continuous real-valued function

depending only on t. If there is a C 1 function u satisfying Lu = f , with the operator
L from (1.12), in some neighborhood of the origin, then f (t) necessarily is analytic at
t = 0.

Remark. This example shows that the assumption of analyticity of f in Theorem 2

in the linear equation can not be omitted (it is very essential). It appears necessarily
since Lu = f with L from (1.12) has no C 1 solution unless f is analytic.
Proof. Suppose x2 + y 2 < R2 , |t| < R and set z = x + iy = reiθ . Denote by V (t) the
function Z Z 2π
V (t) := u(x, y, t)dσ(z) = ir u(r, θ, t)eiθ dθ,
|z|=r 0
1
where u(x, y, t) is the C solution of the equation Lu = f with L from (1.12). We keep
denoting u in polar coordinates also by u. By the divergence theorem for F := (u, iu)
we get
Z Z   Z
∂u ∂u
i ∇ · F dxdy ≡ i +i dxdy = i (u, iu) · νdσ(z)
|z|<r |z|<r ∂x ∂y |z|=r
Z  x Z
y
= i u + iu dσ(z) = i ueiθ dσ(z)
|z|=r r r |z|=r
Z 2π
= ir ueiθ dθ ≡ V (t).
0

But on the other hand, in polar coordinates,

Z   Z r Z 2π  
∂u ∂u ∂u ∂u
V (t) ≡ i +i dxdy = i +i (ρ, θ, t)ρdρdθ.
|z|<r ∂x ∂y 0 0 ∂x ∂y

21
This implies that
Z 2π   Z  
∂V ∂u ∂u ∂u ∂u dσ(z)
= ir +i (r, θ, t)dθ = +i (x, y, t)2r
∂r 0 ∂x ∂y |z|=r ∂x ∂y 2z
Z    Z 
∂u f (t) ∂V dσ(z)
= 2r i + dσ(z) = 2r i + f (t)
|z|=r ∂t 2z ∂t |z|=r 2z
 
∂V
= 2r i + iπf (t) .
∂t

That’s why we have the following equation for V :

 
1 ∂V ∂V
=i + πf (t) . (1.13)
2r ∂r ∂t

Let us introduce now a new function U (s, t) = V (s) + πF (t), where s = r2 and F ′ = f .
The function F exists because f is continuous. It follows from (1.13) that

1 ∂V ∂V ∂U ∂V ∂U ∂U
≡ , = , =i .
2r ∂r ∂s ∂s ∂s ∂s ∂t
Hence
∂U ∂U
+i = 0. (1.14)
∂t ∂s
Since (1.14) is the Cauchy-Riemann equation then U is a holomorphic (analytic) func-
tion of the variable w = t + is, in the region 0 < s < R2 , |t| < R and U is continuous
up to s = 0. Next, since U (0, t) = πF (t) (V = 0 when s = 0 ⇔ r = 0) and f (t) is real-
valued then U (0, t) is also real-valued. Therefore, by the Schwarz reflection principle
(see complex analysis), the formula

U (−s, t) := U (s, t)

gives a holomorphic continuation of U to a full neighborhood of the origin. In partic-

ular, U (0, t) = πF (t) is analytic in t, hence so is f (t) ≡ F ′ (t).

22
2 Fourier Series
Definition. A function f is said to be periodic with period T > 0 if the domain D(f )
of f contains x + T whenever it contains x, and if

f (x + T ) = f (x), x ∈ D(f ). (2.1)

It follows that if T is a period of f then mT is also a period for any integer m > 0.
The smallest value of T > 0 for which (2.1) holds is called the fundamental period of
f.
For example, the functions sin mπx L
and cos mπx
L
, m = 1, 2, . . . are periodic with
2L
fundamental period T = m . Note also that they are periodic with the common period
2L.
Definition. Let us assume that the domain of f is symmetric with respect to {0}, i.e.
if x ∈ D(f ) then −x ∈ D(f ). A function f is called even if

f (−x) = f (x), x ∈ D(f )

and odd if
f (−x) = −f (x), x ∈ D(f ).
Definition. The notations f (c ± 0) are used to denote the limits

f (c ± 0) = lim f (x).
x→c±0

Definition. A function f is said to be piecewise continuous on an interval a ≤ x ≤ b if

the interval can be partitioned by a finite number of points a = x0 < x1 < · · · < xn = b
such that
1. f is continuous on each subinterval xj−1 < x < xj .
2. f (xj ± 0) exists for each j = 1, 2, . . . , n − 1 and f (x0 + 0) and f (xn − 0) exist.
The following properties hold: if a piecewise continuous function f is even then
Z a Z a
f (x)dx = 2 f (x)dx (2.2)
−a 0

and if it is odd then Z a

f (x)dx = 0. (2.3)
−a

Definition. Two real-valued functions u and v are said to be orthogonal on a ≤ x ≤ b

if Z b
u(x)v(x)dx = 0.
a
A set of functions is said to be mutually orthogonal if each distinct pair in the set is
orthogonal on a ≤ x ≤ b.

23
Proposition. The functions 1, sin mπx
L
and cos mπx
L
, m = 1, 2, . . . form a mutually or-
thogonal set on the interval −L ≤ x ≤ L. In fact,
Z L (
mπx nπx 0, m = 6 n
cos cos dx = (2.4)
−L L L L, m = n
Z L
mπx nπx
cos sin dx = 0 (2.5)
−L L L
Z L (
mπx nπx 0, m 6= n
sin sin dx = (2.6)
−L L L L, m = n
Z L Z L
mπx mπx
sin dx = cos dx = 0. (2.7)
−L L −L L
Proof. Let us derive (for example) (2.5). Since
1
cos α sin β = (sin(α + β) − sin(α − β))
2
we have for m 6= n
Z L Z Z
mπx nπx 1 L (m + n)πx 1 L (m − n)πx
cos sin dx = sin dx − sin dx
−L L L 2 −L L 2 −L L
( )L ( )L
1 − cos
(m+n)πx  1 − cos (m−n)πx 
L  L 
=  − 
2 (m+n)π  2 (m−n)π 
L −L L −L
( )
1 − cos(m + n)π cos(m + n)π
= (m+n)π
+ (m+n)π
2
L L
( )
1 − cos(m − n)π cos(m − n)π
− (m−n)π
+ (m−n)π
= 0.
2
L L

If m = n we have
Z L Z L
mπx nπx 1 2mπx
cos sin dx = sin dx = 0
−L L L 2 −L L

since sine is odd. Other identities can be proved in a similar manner and are left to
the reader.
Let us consider the infinite trigonometric series

a0 X  mπx 
∞
mπx
+ am cos + bm sin . (2.8)
2 m=1
L L

24
This series consists of 2L−periodic functions. Thus, if the series (2.8) converges for all
x, then the function to which it converges will be periodic of period 2L. Let us denote
the limiting function by f (x), i.e.

a0 X  mπx 
∞
mπx
f (x) = + am cos + bm sin . (2.9)
2 m=1
L L

To determine am and bm we proceed as follows: assuming that the integration can be

legitimately carried out term by term, we obtain
Z L Z L X∞ Z L
nπx a0 nπx mπx nπx
f (x) cos dx = cos dx + am cos cos dx
−L L 2 −L L m=1 −L L L
X∞ Z L
mπx nπx
+ bm sin cos dx
m=1 −L L L

for each fixed n. It follows from the orthogonality relations (2.4),(2.5) and (2.7) that
the only nonzero term on the right hand side is the one for which m = n in the first
summation. Hence, Z L
nπx
f (x) cos dx = Lan
−L L
or Z
1 L nπx
an = f (x) cos dx. (2.10)
L −L L
A similar expression for bn may be obtained by multiplying (2.9) by sin nπx
L
and inte-
grating termwise from −L to L. Thus,
Z
1 L nπx
bn = f (x) sin dx. (2.11)
L −L L

To determine a0 we use (2.7) to obtain

Z L Z L ∞
X Z L X∞ Z L
a0 mπx mπx
f (x)dx = dx + am cos dx + bm sin dx = a0 L.
−L 2 −L m=1 −L L m=1 −L L

Hence Z L
1
a0 = f (x)dx. (2.12)
L −L

Definition. Let f be a piecewise continuous function on the intervel [−L, L]. The
Fourier series of f is the trigonometric series (2.9), where the coefficients a0 , am and
bm are given by (2.10), (2.11) and (2.12).

25
 It follows from this definition and (2.2)-(2.3) that if f is even on [−L, L] then the
Fourier series of f has the form
∞
a0 X mπx
f (x) = + am cos (2.13)
2 m=1
L
and if f is odd then
∞
X mπx
f (x) = bm sin . (2.14)
m=1
L
The series (2.13) is called the Fourier cosine series and (2.14) is called the Fourier
sine series.
Example 2.1. Find the Fourier series of


−1, −π ≤ x < 0
sgn(x) = 0, x=0


1, 0<x≤π
on the interval [−π, π].
Since L = π and sgn(x) is odd function we have a Fourier sine series with
Z Z  π
1 π 2 π 2 cos(mx) 
bm = sgn(x) sin(mx)dx = sin(mx)dx = − 
π −π π 0 π m 0
    (
m
2 cos(mπ) 1 2 1 − (−1) 0, m = 2k, k = 1, 2, . . .
= − + = = 4
π m m π m πm
, m = 2k − 1, k = 1, 2, . . . .
That’s why
∞
X 4
sgn(x) = sin((2k − 1)x).
k=1
π(2k − 1)
In particular,
∞ ∞
π X sin((k − 1/2)π) X (−1)k+1
= = .
2 k=1
k − 1/2 k=1
k − 1/2
Example 2.2. Let us assume that f (x) = |x|, −1 ≤ x ≤ 1. In this case L = 1 and
f (x) is even. Hence we will have a Fourier cosine series (2.13), where
Z 1 Z 1
a0 = |x|dx = 2 xdx = 1
−1 0
and
Z  1 Z 1
1
sin(mπx)  sin(mπx)
am = 2 x cos(mπx)dx = 2 x  −2 dx
0 mπ 0 0 mπ
 1  
cos(mπx)  cos(mπ) 1
= 2
(mπ)2  = 2 (mπ)2 − (mπ)2
0
(
m 0, m = 2k, k = 1, 2, . . .
2((−1) − 1)
= = 4
(mπ)2 − (mπ)2 , m = 2k − 1, k = 1, 2, . . . .

26
So we have ∞
1 4 X cos((2k − 1)πx)
|x| = − 2 .
2 π k=1 (2k − 1)2
In particular,
∞
π2 X 1
= 2
.
8 k=1
(2k − 1)

Exercise 14. Find the Fourier series of f (x) = x, −1 ≤ x ≤ 1.

Let us consider the partial sums of the Fourier series defined by

a0 X  mπx 
N
mπx
SN (x) = + am cos + bm sin .
2 m=1
L L

We investigate the speed with which the series converges. It is equivalent to the
question: how large value of N must be chosen if we want SN (x) to approximate
f (x) with some accuracy ε > 0? So we need to choose N such that the residual
RN (x) := f (x) − SN (x) satisfies
|RN (x)| < ε
for all x, say, on the interval [−L, L]. Consider the function f (x) from Example 2.2.
Then ∞
4 X cos((2k − 1)πx)
RN (x) = 2
π k=N +1 (2k − 1)2
and
∞  
4 X 1 4 1 1
|RN (x)| ≤ < 2 + + ···
π 2 k=N +1 (2k − 1)2 π (2N )(2N + 1) (2N + 1)(2N + 2)
 
4 1 1 1 1 4 2
= − + − + · · · = = <ε
π 2 2N 2N + 1 2N + 1 2N + 2 2N π 2 N π2

if and only if N > επ22 . Since π 2 ≈ 10 then if ε = 0.04 it is enough to take N = 6, for
ε = 0.01 we have to take N = 21.
The function f (x) = |x| is ”good” enough with respect to ”smoothness” and the
smoothness of |x| guarantees a good approximation by the partial sums. We would
like to formulate a general result.

Theorem 1. Suppose that f and f ′ are piecewise continuous on the interval −L ≤

x ≤ L. Suppose also that f is defined outside the interval −L ≤ x ≤ L so that it is
periodic with period 2L. Then f has a Fourier series (2.8) whose coefficients are given
by (2.10)-(2.12). Moreover, the Fourier series converges to f (x) at all points where f
is continuous, and to 21 (f (x + 0) + f (x − 0)) at all points x where f is discontinuous
(at jump points).

27
Corollary. When f is a 2L−periodic function that is continuous on (−∞, ∞) and has
a piecewise continuous derivative, its Fourier series not only converges at each point
but it converges uniformly on (−∞, ∞), i.e. for every ε > 0 there exists N0 (ε) such
that
|f (x) − SN (x)| < ε, N ≥ N0 (ε), x ∈ (−∞, ∞).
Example 2.3. For sgn(x) on [−π, π) we had the Fourier series
∞
4 X sin((2k − 1)x)
sgn(x) = .
π k=1 2k − 1

Let us extend sgn(x) outside the interval −π ≤ x < π so that it is 2π-periodic. Hence,
this function has jumps at xn = πn, n = 0, ±1, ±2, . . . and
∞
4 X sin((2k − 1)πn) 1
= (sgn(πn + 0) + sgn(πn − 0)) = 0.
π k=1 2k − 1 2

Example 2.4. Let (

0, −L < x < 0
f (x) =
L, 0 < x < L
and let f be defined outside this interval so that f (x + 2L) = f (x) for all x, except at
the points x = 0, ±L, ±2L, . . .. We will temporarily leave open the definition of f at
these points. The Fourier coefficients are
Z Z
1 L 1 L
a0 = f (x)dx = Ldx = L,
L −L L 0
Z Z L L
1 L mπx mπx sin mπx 
am = f (x) cos dx = cos dx = mπL  = 0
L −L L L0 L 0

and
Z Z L
1 L
mπx mπx L
− cos mπx
L 

bm = f (x) sin dx =
sin dx = mπ 
L −L L 0 L L 0
(
L L 0, m = 2k, k = 1, 2, . . .
= (1 − cos(mπ)) = (1 − (−1)m ) = 2L
mπ mπ mπ
, m = 2k − 1, k = 1, 2, . . . .

Hence
∞
L 2L X sin
(2k−1)πx
L
f (x) = + .
2 π k=1 2k − 1
It follows that for any x 6= nL, n = 0, ±1, ±2, . . .,
N
L 2L X sin
(2k−1)πx
L
SN (x) = + → f (x), N → ∞,
2 π k=1 2k − 1

28
where f (x) = 0 or L. At any x = nL,

L L
SN (x) ≡ → , N → ∞.
2 2
But nevertheless, the difference

RN (x) = f (x) − SN (x)

cannot be made uniformly small for all x simultaneously. In the neighborhood of

points of discontinuity (x = nL), the partial sums do not converge smoothly to the
mean value L2 . This behavior is known as the Gibbs phenomenon. However, if we
consider the pointwise convergence of the partial sums then Theorem 1 still applies.

Complex form of the Fourier series

Since
eiα + e−iα eiα − e−iα
cos α = and sin α =
2 2i
then the series (2.8) becomes
∞
X mπx
cm e i L ,
m=−∞

where 
am −ibm

 2 , m = 1, 2, . . .
cm = a20 , m=0

 a−m +ib−m
2
, m = −1, −2, . . . .
If f is real-valued then cm = c−m and
Z L
1 mπx
cm = f (x)e−i L dx, m = 0, ±1, ±2, . . . .
2L −L

In solving problems in differential equations it is often useful to expand in a Fourier

series of period 2L a function f originally defined only on the interval [0, L] (instead
of [−L, L]). Several alternatives are available.

1. Define a function g of period 2L so that

(
f (x), 0≤x≤L
g(x) =
f (−x), −L < x < 0

(f (−L) = f (L) by periodicity). Thus, g(x) is even and its Fourier (cosine) series
represents f on [0, L].

29
 2. Define a function h of period 2L so that


f (x), 0<x<L
h(x) = 0, x = 0, L


−f (−x), −L < x < 0.

Thus, h is the odd periodic extension of f and its Fourier (sine) series represents
f on (0, L).
3. Define a function K of period 2L so that
K(x) = f (x), 0≤x≤L
and let K(x) be defined on (−L, 0) in any way consistent with Theorem 1. Then
its Fourier series involves both sine and cosine terms, and represents f on [0, L].
Example 2.5. Suppose that
(
1 − x, 0 < x ≤ 1
f (x) =
0, 1 < x ≤ 2.

As indicated above, we can represent f either by a cosine series or sine series. For
cosine series we define an even extension of f as follows:


 1 − x, 0 ≤ x ≤ 1

0, 1<x≤2
g(x) =

 1 + x, −1 ≤ x < 0


0, −2 ≤ x < −1,
see Figure 1.

Figure 1: The extension of f .

This is an even 4-periodic function. The Fourier coefficients are

Z Z 2 Z 1
1 2 1
a0 = g(x)dx = g(x)dx = (1 − x)dx =
2 −2 0 0 2

30
and
Z 2 Z 1
1 mπx mπx
am = g(x) cos dx = (1 − x) cos dx
2 −2 2 0 2
1 Z 1
sin mπx 2 
 2 mπx
= (1 − x) mπ  + sin dx
2 0
mπ 0 2
1
2 cos mπx 2 
 4  mπ 
= − mπ  = 1 − cos
mπ 2 0
m2 π 2 2
(
4
m2 π 2
, m = 2k − 1, k = 1, 2, . . .
= 4 k


m2 π 2
1 − (−1) , m = 2k, k = 1, 2, . . . .

Hence the Fourier cosine series has the form

∞ ∞  
4 X cos 4 X 1 − (−1)k
(2k−1)πx
1 2
+ + 2 cos(kπx)
4 π 2 k=1 (2k − 1)2 π k=1 (2k)2

or
∞ ∞
4 X cos 2 X cos((2k − 1)πx)
(2k−1)πx
1 2
+ + 2 .
4 π 2 k=1 (2k − 1)2 π k=1 (2k − 1)2
This representation holds for all x ∈ R. In particular, for all x ∈ [1, 3] we have
∞ ∞
4 X cos 2 X cos((2k − 1)πx)
(2k−1)πx
1 2
+ + 2 = 0.
4 π 2 k=1 (2k − 1)2 π k=1 (2k − 1)2

Exercise 15. Find the corresponding Fourier sine series of f .

31
3 One-dimensional Heat Equation
Let us consider a heat conduction problem for a straight bar of uniform cross section
and homogeneous material. Let x = 0 and x = L denote the ends of the bar (x-axis
is chosen to lie along the axis of the bar). Suppose that no heat passes through the
sides of the bar. We also assume that the cross-sectional dimensions are so small that
temperature u can be considered the same on any given cross section.

x=0 u(x, t) x=L

Then u is a function only of the coordinate x and the time t. The variation of temper-
ature in the bar is governed by a partial differential equation

α2 uxx (x, t) = ut (x, t), 0 < x < L, t > 0, (3.1)

where α2 is a constant known as the thermal diffusivity. This equation is called the
heat conduction equation or heat equation.
In addition, we assume that the initial temperature distribution in the bar is given
by
u(x, 0) = f (x), 0 ≤ x ≤ L, (3.2)
where f is a given function. Finally, we assume that the temperature at each end of
the bar is given by

u(0, t) = g0 (t), u(L, t) = g1 (t), t > 0, (3.3)

where g0 and g1 are given functions. The problem (3.1), (3.2), (3.3) is an initial value
problem in time variable t. With respect to the space variable x it is a boundary value
problem and (3.3) are called the boundary conditions. Alternatively, this problem can
be considered as a boundary value problem in the xt-plane:
t

u(0, t) = g0 (t) α2 uxx = ut u(L, t) = g1 (t)

x
x=0 u(x, 0) = f (x) x=L

We start by considering the homogeneous boundary conditions when the functions g0 (t)

32
and g1 (t) in (3.3) are identically zero:

 2
α uxx = ut , 0 < x < L, t > 0
u(0, t) = u(L, t) = 0, t > 0 (3.4)


u(x, 0) = f (x), 0 ≤ x ≤ L.

We look for a solution to the problem (3.4) in the form

u(x, t) = X(x)T (t). (3.5)

Such method is called a separation of variables. Substituting (3.5) into (3.1) yields

α2 X ′′ (x)T (t) = X(x)T ′ (t)

or
X ′′ (x) 1 T ′ (t)
= 2
X(x) α T (t)
in which the variables are separated, that is, the left hand side depends only on x and
the right hand side only on t. This is possible only when both sides are equal to the
same constant:
X ′′ 1 T′
= 2 = −λ.
X α T
Hence, we obtain two ordinary differential equations for X(x) and T (t)

X ′′ + λX = 0,

T ′ + α2 λT = 0. (3.6)
The boundary condition for u(x, t) at x = 0 leads to

u(0, t) = X(0)T (t) = 0.

It follows that
X(0) = 0
(since otherwise T ≡ 0 and so u ≡ 0 which we do not want). Similarly, the boundary
condition at x = L requires that
X(L) = 0.
So, for the function X(x) we obtain the homogeneous boundary value problem
(
X ′′ + λX = 0, 0<x<L
(3.7)
X(0) = X(L) = 0.

The values of λ for which nontrivial solutions of (3.7) exist are called eigenvalues and
the corresponding nontrivial solutions are called eigenfunctions. The problem (3.7) is
called an eigenvalue problem.

33
Lemma 1. The problem (3.7) has an infinite sequence of positive eigenvalues

n2 π 2
λn = , n = 1, 2, . . .
L2
with the corresponding eigenfunctions
nπx
Xn (x) = c sin ,
L
where c is an arbitrary nonzero constant.
Proof. Suppose first that λ > 0, i.e. λ = µ2 . The characteristic equation for (3.7) is
r2 + µ2 = 0 with roots r = ±iµ, so the general solution is

X(x) = c1 cos µx + c2 sin µx.

Note that µ is nonzero and there is no loss of generality if we assume that µ > 0. The
first boundary condition in (3.7) implies

X(0) = c1 = 0,

and the second reduces to

c2 sin µL = 0
or
sin µL = 0
as we do not allow c2 = 0 too. It follows that

µL = nπ, n = 1, 2, . . .

or
n2 π 2
λn = , n = 1, 2, . . . .
L2
Hence the corresponding eigenfunctions are
nπx
Xn (x) = c sin .
L
If λ = −µ2 < 0, µ > 0, then the characteristic equation for (3.7) is r2 − µ2 = 0 with
roots r = ±µ. Hence the general solution is

X(x) = c1 cosh µx + c2 sinh µx.

Since
eµx + e−µx eµx − e−µx
cosh µx = and sinh µx =
2 2
this is equivalent to
X(x) = c′1 eµx + c′2 e−µx .

34
The first boundary condition requires again that c1 = 0 while the second gives

c2 sinh µL = 0.

Since µ 6= 0 (µ > 0), it follows that sinh µL 6= 0 and therefore we must have c2 = 0.
Consequently, X ≡ 0, i.e. there are no nontrivial solutions for λ < 0.
If λ = 0 the general solution is

X(x) = c1 x + c2 .

The boundary conditions can be satisfied only if c1 = c2 = 0 so there is only the trivial
solution in this case as well.
n2 π 2
Turning now to (3.6) for T (t) and substituting L2
for λ we have
nπα 2
T (t) = ce−( L ) .
t

Hence the functions

un (x, t) = e−( ) sin nπx
nπα 2
L
t
(3.8)
L
satisfy (3.1) and the homogeneous boundary conditions from (3.4) for each n = 1, 2, . . ..
The linear superposition principle gives that any linear combination
N
X
u(x, t) = c n e −( L ) sin nπx
nπα 2
t

n=1
L

is also a solution of the same problem. In order to take into account infinitely many
functions (3.8) we assume that
∞
X
u(x, t) = c n e −( L ) sin nπx ,
nπα 2
t
(3.9)
n=1
L

where the coefficients cn are yet undetermined, and the series converges in some sense.
To satisfy the initial condition from (3.4) we must have
∞
X nπx
u(x, 0) = cn sin = f (x), 0 ≤ x ≤ L. (3.10)
n=1
L

In other words, we need to choose the coefficients cn so that the series (3.10) converges
to the initial temperature distribution f (x).
It is not difficult to prove that for t > 0, 0 < x < L, the series (3.9) converges (with
any derivative with respect to x and t) and solves (3.1) with boundary conditions (3.4).
Only one question remains: can any function f (x) be represented by a Fourier sine
series (3.10)? Some sufficient conditions for such representation are given in Theorem
1 of Chapter 2.

35
Remark. We can consider the boundary value problem for any linear differential equa-
tion
y ′′ + p(x)y ′ + q(x)y = g(x) (3.11)
of order two on the interval (a, b) with the boundary conditions
y(a) = y0 , y(b) = y1 , (3.12)
where y0 and y1 are given constants. Let us assume that we have found a fundamental
set of solutions y1 (x) and y2 (x) to the corresponding homogeneous equation
y ′′ + p(x)y ′ + q(x)y = 0.
Then the general solution to (3.11) is
y(x) = c1 y1 (x) + c2 y2 (x) + yp (x),
where yp (x) is a particular solution to (3.11) and c1 and c2 are arbitrary constants.
To satisfy the boundary conditions (3.12) we have the linear nonhomogeneous al-
gebraic system (
c1 y1 (a) + c2 y2 (a) = y0 − yp (a)
(3.13)
c1 y1 (b) + c2 y2 (b) = y1 − yp (b).
If the determinant  
 y1 (a) y2 (a) 
 
 y1 (b) y2 (b) 
is nonzero, then the constants c1 and c2 can be determined uniquely and therefore the
boundary value problem (3.11)-(3.12) has a unique solution. If
 
 y1 (a) y2 (a) 
 
 y1 (b) y2 (b)  = 0

then (3.11)-(3.12) either has no solutions or has infinitely many solutions.

Example 3.1. Let us consider the boundary value problem
(
y ′′ + µ2 y = 1, 0<x<1
y(0) = y0 , y(1) = y1 ,
1
where µ > 0 is fixed. This differential equation has a particular solution yp (x) = µ2
.
Hence, the system (3.13) becomes
(
c1 sin 0 + c2 cos 0 = y0 − µ12
c1 sin µ + c2 cos µ = y1 − µ12
or ( 1
c 2 = y0 − µ2  
1 1
c1 sin µ = y1 − µ2
− y0 − µ2
cos µ.

36
If  
 0 1 
  6= 0
 sin µ cos µ 
i.e. sin µ 6= 0 then c1 is uniquely determined and the boundary value problem in
question has a unique solution. If sin µ = 0 then the problem has solutions (actually,
infinitely many) if and only if
 
1 1
y1 − 2 = y0 − 2 cos µ.
µ µ

If µ = 2πk then sin µ = 0 and cos µ = 1 and the following equation must hold
1 1
y1 − 2
= y0 − 2
µ µ
i.e. y1 = y0 . If µ = π + 2πk then sin µ = 0 and cos µ = −1 and we must have
2
y1 + y0 = .
µ2
Suppose now that one end of the bar is held at a constant temperature T1 and the
other is maintained at a constant temperature T2 . The corresponding boundary value
problem is then

 2
α uxx = ut , 0 < x < L, t > 0
u(0, t) = T1 , u(L, t) = T2 , t > 0 (3.14)


u(x, 0) = f (x).

After a long time (t → ∞) we anticipate that a steady temperature distribution v(x)

will be reached, which is independent of time and the initial condition. Since the
solution of (3.14) with T1 = T2 = 0 tends to zero as t → ∞, see (3.9), then we look for
the solution to (3.14) in the form

u(x, t) = v(x) + w(x, t). (3.15)

Substituting (3.15) into (3.14) leads to


 2
α (vxx + wxx ) = wt
v(0) + w(0, t) = T1 , v(L) + w(L, t) = T2


v(x) + w(x, 0) = f (x).

Let us assume that v(x) satisfies the steady-state problem

(
v ′′ (x) = 0, 0 < x < L
(3.16)
v(0) = T1 , v(L) = T2 .

37
Then w(x, t) satisfies the homogeneous boundary value problem for the heat equation:

 2
α wxx = wt , 0 < x < L, t > 0
w(0, t) = w(L, t) = 0 (3.17)

 ˜
w(x, 0) = f (x),

where f˜(x) = f (x) − v(x). Since the solution of (3.16) is

T2 − T1
v(x) = x + T1 (3.18)
L
the solution of (3.17) is
∞
X
w(x, t) = c n e −( L ) sin nπx ,
nπα 2
t
(3.19)
n=1
L

where the coefficients cn are given by

Z  
2 L T2 − T1 nπx
cn = f (x) − x − T1 sin dx.
L 0 L L

Combining (3.18) and (3.19) we obtain

X∞
T2 − T1 nπα 2 nπx
u(x, t) = x + T1 + cn e−( L ) t sin .
L n=1
L

Let us slightly complicate the problem (3.14), namely assume that


 2
α uxx = ut + p(x), 0 < x < L, t > 0
u(0, t) = T1 , u(L, t) = T2 , t > 0 (3.20)


u(x, 0) = f (x).

We begin by assuming that the solution to (3.20) consists of a steady-state solution

v(x) and a transient solution w(x, t) which tends to zero as t → ∞:

u(x, t) = v(x) + w(x, t).

Then for v(x) we will have the problem

(
v ′′ (x) = α12 p(x), 0 < x < L
(3.21)
v(0) = T1 , v(L) = T2 .

To solve this, integrate twice to get

Z x Z y
1
v(x) = 2 dy p(s)ds + c1 x + c2 .
α 0 0

38
The boundary conditions yield c2 = T1 and
 Z L Z y 
1 1
c1 = T2 − T1 − 2 dy p(s)ds .
L α 0 0

Therefore, the solution of (3.21) has the form

Z L Z y Z x Z y
T2 − T1 x 1
v(x) = x− dy p(s)ds + 2 dy p(s)ds + T1 .
L Lα2 0 0 α 0 0

For w(x, t) we will have the homogeneous problem


 2
α wxx = wt , 0 < x < L, t > 0
w(0, t) = w(L, t) = 0, t > 0


w(x, 0) = f˜(x) := f (x) − v(x).

A different problem occurs if the ends of the bar are insulated so that there is no
passage of heat through them. Thus, in the case of no heat flow, the boundary value
problem is 
α2 uxx = ut , 0 < x < L, t > 0

ux (0, t) = ux (L, t) = 0, t > 0 (3.22)


u(x, 0) = f (x).
This problem can also be solved by the method of separation of variables. If we let
u(x, t) = X(x)T (t) it follows that

X ′′ + λX = 0, T ′ + α2 λT = 0. (3.23)

The boundary conditions yield now

X ′ (0) = X ′ (L) = 0. (3.24)

If λ = −µ2 < 0, µ > 0, then (3.23) for X(x) becomes X ′′ − µ2 X = 0 with general
solution
X(x) = c1 sinh µx + c2 cosh µx.
Therefore, the conditions (3.24) give c1 = 0 and c2 = 0 which is unacceptable. Hence
λ cannot be negative.
If λ = 0 then
X(x) = c1 x + c2 .
Thus X ′ (0) = c1 = 0 and X ′ (L) = 0 for any c2 leaving c2 undetermined. Therefore
λ = 0 is an eigenvalue, corresponding to the eigenfunction X0 (x) = 1. It follows from
(3.23) that T (t) is also a constant. Hence, for λ = 0 we obtain the constant solution
u0 (x, t) = c2 .

39
 If λ = µ2 > 0 then X ′′ + µ2 X = 0 and consequently

X(x) = c1 sin µx + c2 cos µx.

The boundary conditions imply c1 = 0 and µ = nπ L

, n = 1, 2, . . . leaving c2 arbitrary.
2 2
Thus we have an infinite sequence of positive eigenvalues λn = nLπ2 with the corre-
sponding eigenfunctions
nπx
Xn (x) = cos , n = 1, 2, . . . .
L
If we combine these eigenvalues and eigenfunctions with zero eigenvalue and X0 (x) = 1
we may conclude that we have the infinite sequences
n2 π 2 nπx
λn = , Xn (x) = cos , n = 0, 1, 2, . . . .
L2 L
and
nπx −( nπα 2
un (x, t) = cos e L ) t , n = 0, 1, 2, . . . .
L
Each of these functions satisfies the equation and boundary conditions from (3.22). It
remains to satisfy the initial condition. In order to do it we assume that u(x, t) has
the form ∞
c0 X nπx −( nπα 2
u(x, t) = + cn cos e L ) t, (3.25)
2 n=1
L
where the coefficients cn are determined by the requirement that
∞
c0 X nπx
u(x, 0) = + cn cos = f (x), 0 ≤ x ≤ L.
2 n=1
L

Thus the unknown coefficients in (3.25) must be the Fourier coefficients in the Fourier
cosine series of period 2L for even extension of f . Hence
Z
2 L nπx
cn = f (x) cos dx, n = 0, 1, 2, . . .
L 0 L
and the series (3.25) provides the solution to the heat conduction problem (3.22) for a
rod with insulated ends. The physical interpretation of the term
Z
c0 1 L
= f (x)dx
2 L 0
is that it is the mean value of the original temperature distribution.
Exercise 16. Let v(x) be a solution of the problem
(
v ′′ (x) = 0, 0 < x < L
v ′ (0) = T1 , v ′ (L) = T2 .

40
Show that the problem

 2
α uxx = ut , 0 < x < L, t > 0
ux (0, t) = T1 , ux (L, t) = T2 , t > 0


u(x, 0) = f (x)

has a solution of the form u(x, t) = v(x) + w(x, t) if and only if T1 = T2 .

Example 3.2. 

uxx = ut , 0 < x < 1, t > 0
u(0, t) = u(1, t) = 0

 P
u(x, 0) = ∞ 1
n=1 n2 sin(nπx) := f (x).

As we know the solution of this problem is given by

∞
X 2
u(x, t) = cn sin(nπx)e−(nπ) t .
n=1

Since ∞ ∞
X X 1
u(x, 0) = cn sin(nπx) = 2
sin(nπx)
n=1 n=1
n
1
then we may conclude that cn = n2
necessarily (since the Fourier series is unique).
Hence the solution is ∞
X 1 2
u(x, t) = 2
sin(nπx)e−(nπ) t .
n=1
n

Exercise 17. Find a solution of the problem



uxx = ut , 0 < x < π, t > 0
ux (0, t) = ux (π, t) = 0, t > 0


u(x, 0) = 1 − sin x

using the method of separation of variables.

Let us consider a bar with mixed boundary conditions at the ends. Assume that
the temperature at x = 0 is zero, while the end x = L is insulated so that no heat
passes through it: 
 2
α uxx = ut , 0 < x < L, t > 0
u(0, t) = ux (L, t) = 0, t > 0


u(x, 0) = f (x).
Separation of variables leads to
(
X ′′ + λX = 0, 0 < x < L
(3.26)
X(0) = X ′ (L) = 0

41
and
T ′ + λT = 0, t > 0.
As above, one can show that (3.26) has nontrivial solutions only for λ > 0, namely
(2m − 1)2 π 2 (2m − 1)πx
λm = 2
, Xm (x) = sin , m = 1, 2, 3, . . . .
4L 2L
The solution to the mixed boundary value problem is
X∞
(2m − 1)πx −( (2m−1)πα 2
)t
u(x, t) = cm sin e 2L

m=1
2L
with arbitrary constants cm . To satisfy the initial condition we have
X∞
(2m − 1)πx
f (x) = cm sin , 0 ≤ x ≤ L.
m=1
2L
This is a Fourier sine series but in some specific form. We show that the coefficients
cm can be calculated as
Z
2 L (2m − 1)πx
cm = f (x) sin dx
L 0 2L
and such representation is possible.
In order to prove it, let us first extend f (x) to the interval 0 ≤ x ≤ 2L so that
it is symmetric about x = L, i.e. f (2L − x) = f (x) for 0 ≤ x ≤ L. Then extend
the resulting function to the interval (−2L, 0) as an odd function and elsewhere as a
periodic function fe of period 4L. In this procedure we need to define
fe(0) = fe(2L) = fe(−2L) = 0.
Then the Fourier series contains only sines:
X∞
e nπx
f (x) = cn sin
n=1
2L
with the Fourier coefficients
Z 2L
2 nπx
cn = fe(x) sin dx.
2L 0 2L
Let us show that cn = 0 for even n = 2m. Indeed,
Z
1 2L e mπx
c2m = f (x) sin dx
L 0 L
Z Z
1 L mπx 1 2L mπx
= f (x) sin dx + f (2L − x) sin dx
L 0 L L L L
Z Z
1 L mπx 1 0 mπ(2L − y)
= f (x) sin dx − f (y) sin dy
L 0 L L L L
Z Z
1 L mπx 1 0 mπy
= f (x) sin dx + f (y) sin dy = 0.
L 0 L L L L

42
That’s why
∞
X (2m − 1)πx
fe(x) = c2m−1 sin ,
m=1
2L
where
Z
1 2L e (2m − 1)πx
c2m−1 = f (x) sin dx
L 0 2L
Z Z
1 L (2m − 1)πx 1 2L (2m − 1)πx
= f (x) sin dx + f (2L − x) sin dx
L 0 2L L L 2L
Z
2 L (2m − 1)πx
= f (x) sin dx
L 0 2L
as claimed. Let us remark that the series
∞
X (2m − 1)πx
cm sin
m=1
2L

represents f (x) on (0, L].

Remark. For the boundary conditions

ux (0, t) = u(L, t) = 0

the function f (x) must be extended to the interval 0 ≤ x ≤ 2L as f (x) = −f (2L − x)

with f (L) = 0. Furthermore, fe is an even extension to the interval (−2L, 0). Then
the corresponding Fourier series represents f (x) on the interval [0, L).

43
4 One-dimensional Wave Equation
Another situation in which the separation of variables applies occurs in the study of a
vibrating string. Suppose that an elastic string of length L is tightly stretched between
two supports, so that the x-axis lies along the string. Let u(x, t) denote the vertical
displacement experienced by the string at the point x at time t. It turns out that if
damping effects are neglected, and if the amplitude of the motion is not too large, then
u(x, t) satisfies the partial differential equation
a2 uxx = utt , 0 < x < L, t > 0. (4.1)
Equation (4.1) is known as the one-dimensional wave equation. The constant a2 = T /ρ,
where T is the force in the string and ρ is the mass per unit length of the string material.

x
x=0 x=L
u(x, t)

To describe the motion completely it is necessary also to specify suitable initial and
boundary conditions for the displacement u(x, t). The ends are assumed to remain
fixed:
u(0, t) = u(L, t) = 0, t ≥ 0. (4.2)
The initial conditions are (since (4.1) is of second order with respect to t):
u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L, (4.3)
where f and g are given functions. In order for (4.2) and (4.3) to be consistent it is
also necessary to require that
f (0) = f (L) = g(0) = g(L) = 0. (4.4)
Equations (4.1)-(4.4) can be interpreted as the following boundary value problem for
the wave equation:
t

u(0, t) = 0 a2 uxx = utt u(L, t) = 0

x
x=0 u(x, 0) = f (x) x=L
ut (x, 0) = g(x)

44
 Let us apply the method of separation of variables to this homogeneous boundary
value problem. Assuming that u(x, t) = X(x)T (t) we obtain

X ′′ + λX = 0, T ′′ + a2 λT = 0.

The boundary conditions (4.2) imply that

(
X ′′ + λX = 0, 0 < x < L
X(0) = X(L) = 0.

This is the same boundary value problem that we have considered before. Hence,

n2 π 2 nπx
λn = , Xn (x) = sin , n = 1, 2, . . . .
L2 L
Taking λ = λn in the equation for T (t) we have
 nπa 2
T ′′ (t) + T (t) = 0.
L
The general solution to this equation is
nπat nπat
T (t) = k1 cos + k2 sin ,
L L
where k1 and k2 are arbitrary constants. Using the linear superposition principle we
consider the infinite sum
X∞  
nπx nπat nπat
u(x, t) = sin an cos + bn sin , (4.5)
n=1
L L L

where the coefficients an and bn are to be determined. It is clear that u(x, t) from (4.5)
satisfies (4.1) and (4.2) (at least formally). The initial conditions (4.3) imply
∞
X nπx
f (x) = an sin , 0 ≤ x ≤ L,
n=1
L
∞ (4.6)
X nπa nπx
g(x) = bn sin , 0 ≤ x ≤ L.
n=1
L L

Since (4.4) are fulfilled then (4.6) are the Fourier sine series for f and g, respectively.
Therefore,
Z
2 L nπx
an = f (x) sin dx,
L 0 L
Z L (4.7)
2 nπx
bn = g(x) sin dx.
nπa 0 L

45
Finally, we may conclude that the series (4.5) with the coefficients (4.7) solves (at least
formally) the boundary value problem (4.1)-(4.4).
Each displacement pattern
 
nπx nπat nπat
un (x, t) = sin an cos + bn sin
L L L
is called a natural mode of vibration and is periodic in both space variable x and time
variable t. The spatial period 2L
n
in x is called the wavelength, while the numbers nπa
L
are called the natural frequencies.
Exercise 18. Find a solution of the problem


uxx = utt , 0 < x < 1, t > 0
u(0, t) = u(1, t) = 0, t ≥ 0


u(x, 0) = x(1 − x), ut (x, 0) = sin(7πx)

using the method of separation of variables.

If we compare the two series
∞
X  
nπx nπat nπat
u(x, t) = sin an cos + bn sin
n=1
L L L
∞
X nπx −( nπα 2
u(x, t) = cn sin e L )t
n=1
L

for the wave and heat equations we can see that the second series has the exponential
factor that decays fast with n for any t > 0. This guarantees convergence of the series
as well as the smoothness of the sum. This is not true anymore for the first series
because it contains only oscillatory terms that do not decay with increasing n.

The boundary value problem for the wave equation with free ends of the string can
be formulated as follows:

 2
a uxx = utt , 0 < x < L, t > 0
ux (0, t) = ux (L, t) = 0, t ≥ 0


u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L.

Let us first note that the boundary conditions imply that f (x) and g(x) must satisfy

f ′ (0) = f ′ (L) = g ′ (0) = g ′ (L) = 0.

The method of separation of variables gives that the eigenvalues are

 nπ 2
λn = , n = 0, 1, 2, . . .
L

46
and the formal solution u(x, t) is
∞  
b 0 t + a0 X nπx nπat nπat
u(x, t) = + cos an cos + bn sin .
2 n=1
L L L
The initial conditions are satisfied when
∞
a0 X nπx
f (x) = + an cos
2 n=1
L
and ∞
b0 X nπa nπx
g(x) = + bn cos ,
2 n=1
L L
where
Z
2 L nπx
an = f (x) cos dx, n = 0, 1, 2, . . .
L 0 L
Z
2 L
b0 = g(x)dx
L 0
and
Z L
2 nπx
bn = g(x) cos dx, n = 1, 2, . . . .
nπa 0 L
Let us consider the wave equation on the whole line. It corresponds, so to say, to
the infinite string. In that case we no more have the boundary conditions but we have
the initial conditions:
(
a2 uxx = utt , −∞ < x < ∞, t > 0
(4.8)
u(x, 0) = f (x), ut (x, 0) = g(x).
Proposition. The solution u(x, t) of the wave equation is of the form
u(x, t) = ϕ(x − at) + ψ(x + at),
where ϕ and ψ are two arbitrary C 2 functions of one variable.
Proof. By the chain rule
∂tt u − a2 ∂xx u = 0
if and only if
∂ξ ∂η u = 0,
where ξ = x + at and η = x − at (and so ∂x = ∂ξ + ∂η , a1 ∂t = ∂ξ − ∂η ). It follows that
∂ξ u = Ψ(ξ)
or
u = ψ(ξ) + ϕ(η),
where ψ ′ = Ψ.

47
 To satisfy the initial conditions we have

f (x) = ϕ(x) + ψ(x), g(x) = −aϕ′ (x) + aψ ′ (x).

It follows that
1 1 1 1
ϕ′ (x) = f ′ (x) − g(x), ψ ′ (x) = f ′ (x) + g(x).
2 2a 2 2a
Integrating we obtain
Z x Z x
1 1 1 1
ϕ(x) = f (x) − g(s)ds + c1 , ψ(x) = f (x) + g(s)ds + c2 ,
2 2a 0 2 2a 0

where c1 and c2 are arbitrary constants. But ϕ(x) + ψ(x) = f (x) implies c1 + c2 = 0.
Therefore the solution of the initial value problem is
Z x+at
1 1
u(x, t) = (f (x − at) + f (x + at)) + g(s)ds. (4.9)
2 2a x−at

This formula is called the d’Alembert formula.

Exercise 19. Prove that if f is a C 2 function and g is a C 1 function, then u from

(4.9) is a C 2 function and satisfies (4.8) in the classical sense.

Exercise 20. Prove that if f and g are merely locally integrable, then u from (4.9) is
a distributional solution of (4.8) and the initial conditions are satisfied pointwise.

Example 4.1. The solution of

(
uxx = utt , −∞ < x < ∞, t > 0
u(x, 0) = f (x), ut (x, 0) = 0,

where (
1, |x| ≤ 1
f (x) =
0, |x| > 1
is given by the d’Alembert formula
1
u(x, t) = (f (x − t) + f (x + t)) .
2
Some solutions are graphed below.

48
 u(x, t)

1 t=0

x
−1 1
u(x, t)

1 t= 1
2

x
−1 1
u(x, t)

t=2

1
2

x
−1 1

We can apply the d’Alembert formula for the finite string also. Consider again the
boundary value problem with homogeneous boundary conditions with fixed ends of the
string.  2

 a uxx = utt , 0 < x < L, t > 0

u(0, t) = u(L, t) = 0, t ≥ 0

 u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L


f (0) = f (L) = g(0) = g(L) = 0.
Let h(x) be the function defined for all x ∈ R such that
(
f (x), 0≤x≤L
h(x) =
−f (−x), −L ≤ x ≤ 0
and 2L-periodic and let k(x) be the function defined for all x ∈ R such that
(
g(x), 0≤x≤L
k(x) =
−g(−x), −L ≤ x ≤ 0

49
and 2L-periodic. Let us also assume that f and g are C 2 functions on the interval
[0, L]. Then the solution to the boundary value problem is given by the d’Alembert
formula Z x+at
1 1
u(x, t) = (h(x − at) + h(x + at)) + k(s)ds.
2 2a x−at
Remark. It can be checked that this solution is equivalent to the solution which is given
by the Fourier series.

50
5 Laplace Equation in Rectangle and in Disk
One of the most important of all partial differential equations in applied mathematics
is the Laplace equation:
uxx + uyy = 0 2D-equation
(5.1)
uxx + uyy + uzz = 0 3D-equation
The Laplace equation appears quite naturally in many applications. For example, a
steady state solution of the heat equation in two space dimensions
α2 (uxx + uyy ) = ut
satisfies the 2D-Laplace equation (5.1). When considering electrostatic fields, the elec-
tric potential function must satisfy either 2D or 3D equation (5.1).
A typical boundary value problem for the Laplace equation is (in dimension two):
(
uxx + uyy = 0, (x, y) ∈ Ω ⊂ R2
(5.2)
u(x, y) = f (x, y), (x, y) ∈ ∂Ω,
where f is a given function on the boundary ∂Ω of the domain Ω. The problem (5.2)
is called the Dirichlet problem (Dirichlet boundary conditions). The problem
(
uxx + uyy = 0, (x, y) ∈ Ω
∂u
∂ν
(x, y) = g(x, y), (x, y) ∈ ∂Ω,

where g is given and ∂u

∂ν
is the outward normal derivative is called the Neumann problem
(Neumann boundary conditions).
y
ν |ν| = 1

∂Ω

Dirichlet problem for a rectangle

Consider the boundary value problem in most general form:


wxx + wyy = 0, 0 < x < a, 0 < y < b
w(x, 0) = g1 (x), w(x, b) = f1 (x), 0 < x < a


w(0, y) = g2 (y), w(a, y) = f2 (y), 0 ≤ y ≤ b,

51
for fixed a > 0 and b > 0. The solution of this problem can be reduced to the solutions
of 

uxx + uyy = 0, 0 < x < a, 0 < y < b
u(x, 0) = u(x, b) = 0, 0<x<a (5.3)


u(0, y) = g(y), u(a, y) = f (y), 0 ≤ y ≤ b,
and 

uxx + uyy = 0, 0 < x < a, 0 < y < b
u(x, 0) = g1 (x), u(x, b) = f1 (x), 0 < x < a


u(0, y) = 0, u(a, y) = 0, 0 ≤ y ≤ b.
Due to symmetry in x and y we consider (5.3) only.
y
u(x, b) = 0
b

u(0, y) = g(y) Ω u(a, y) = f (y)

x
u(x, 0) = 0 a

The method of separation of variables gives for u(x, y) = X(x)Y (y),

(
Y ′′ + λY = 0, 0 < y < b,
(5.4)
Y (0) = Y (b) = 0,

and
X ′′ − λX = 0, 0 < x < a. (5.5)
From (5.4) one obtains the eigenvalues and eigenfunctions
 nπ 2 nπy
λn = , Yn (y) = sin , n = 1, 2, . . . .
b b
Substitute λn into (5.5) to get the general solution
nπx nπx
X(x) = c1 cosh + c2 sinh .
b b
As above, represent the solution to (5.3) in the form

nπy  nπx 
∞
X nπx
u(x, y) = sin an cosh + bn sinh . (5.6)
n=1
b b b

52
The boundary condition at x = 0 gives
∞
X nπy
g(y) = an sin ,
n=1
b

with Z b
2 nπy
an = g(y) sin dy.
b 0 b
At x = a we obtain

nπy  nπa 
∞
X nπa
f (y) = sin an cosh + bn sinh .
n=1
b b b

It is a Fourier sine series for f (y). Hence,

Z b
nπa nπa 2 nπy
an cosh + bn sinh = f (y) sin dy := ebn .
b b b 0 b
It implies
ebn − an cosh nπa
b
bn = . (5.7)
sinh nπa
b

Substituting (5.7) into (5.6) gives

!
nπx ebn − an cosh nπa
∞
X nπy b nπx
u(x, y) = sin an cosh + nπa sinh
n=1
b b sinh b b
∞
X nπy e sinh nπx
b
= sin bn nπa
n=1
b sinh b
X∞  
nπy cosh nπx
b
sinh nπa
b
− cosh nπa
b
sinh nπx
b
+ sin an
n=1
b sinh nπa
b
∞
X X ∞
nπ(a−x)
nπy e sinh nπx
b nπy sinh b
= sin bn + sin an ,
n=1
b sinh nπa
b n=1
b sinh nπa
b

because cosh α sinh β − sinh α cosh β = sinh(β − α).

Exercise 21. Find a solution of the problem



uxx + uyy = 0, 0 < x < 2, 0 < y < 1
u(x, 0) = u(x, 1) = 0, 0<x<2


u(0, y) = 0, u(2, y) = y(1 − y), 0 ≤ y ≤ 1

using the method of separation of variables.

53
Dirichlet problem for a disk
Consider the problem of solving the Laplace equation in a disk {x ∈ R2 : |x| < a}
subject to boundary condition
u(a, θ) = f (θ), (5.8)
where f is a given function on 0 ≤ θ ≤ 2π. In polar coordinates x = r cos θ, y = r sin θ,
the Laplace equation takes the form
1 1
urr + ur + 2 uθθ = 0. (5.9)
r r
We apply again the method of separation of variables and assume that

u(r, θ) = R(r)T (θ). (5.10)

Substitution for u in (5.9) yields

1 1
R′′ T + R′ T + 2 RT ′′ = 0
r r
or (
r2 R′′ + rR′ − λR = 0
T ′′ + λT = 0.
There are no homogeneous boundary conditions, however we need T (θ) to be 2π-
periodic and also bounded. This fact, in particular, leads to

T (0) = T (2π), T ′ (0) = T ′ (2π). (5.11)

It is possible to show that (5.11) require λ to be real. In what follows we will consider
the three possible cases.
If λ = −µ2 < 0, µ > 0, then the equation for T becomes T ′′ − µ2 T = 0 and
consequently
T (θ) = c1 eµθ + c2 e−µθ .
It follows from (5.11) that
(
c1 + c2 = c1 e2πµ + c2 e−2πµ
c1 − c2 = c1 e2πµ − c2 e−2πµ

so that c1 = c2 = 0.
If λ = 0 then T ′′ = 0 and T (θ) = c1 + c2 θ. The first condition in (5.11) implies
then that c2 = 0 and therefore T (θ) ≡ constant.
If λ = µ2 > 0, µ > 0, then

T (θ) = c1 cos(µθ) + c2 sin(µθ).

54
Now the conditions (5.11) imply that
(
c1 = c1 cos(2πµ) + c2 sin(2πµ)
c2 = −c1 sin(2πµ) + c2 cos(2πµ)
or (
c1 sin2 (µπ) = c2 sin(µπ) cos(µπ)
c2 sin2 (µπ) = −c1 sin(µπ) cos(µπ).
If sin(µπ) 6= 0 then (
c1 = c2 cot(µπ)
c2 = −c1 cot(µπ).
Hence c21 + c22 = 0 i.e. c1 = c2 = 0. Thus we must have sin(µπ) = 0 and so
λ n = n2 , Tn (θ) = c1 cos(nθ) + c2 sin(nθ), n = 0, 1, 2, . . . . (5.12)
Turning now to R, for λ = 0 we have r2 R′′ + rR′ = 0 i.e. R(r) = k1 + k2 log r. Since
log r → −∞ as r → 0 we must choose k2 = 0 in order for R (and u) to be bounded.
That’s why
R0 (r) ≡ constant. (5.13)
For λ = µ2 = n2 the equation for R becomes
r2 R′′ + rR′ − n2 R = 0.
Hence
R(r) = k1 rn + k2 r−n .
Again, we must choose k2 = 0 and therefore
Rn (r) = k1 rn , n = 1, 2, . . . . (5.14)
Combining (5.10),(5.12), (5.13) and (5.14) we obtain
∞
a0 X n
u(r, θ) = + r (an cos(nθ) + bn sin(nθ)). (5.15)
2 n=1

The boundary condition (5.8) then requires

∞
a0 X n
u(a, θ) = + a (an cos(nθ) + bn sin(nθ)) = f (θ).
2 n=1

Hence the coefficients are given by

Z
1 2π
a0 = f (θ)dθ,
π 0
Z 2π
1
an = f (θ) cos(nθ)dθ
πan 0

55
and
Z 2π
1
bn = f (θ) sin(nθ)dθ.
πan 0

This procedure can be used also to study the Neumann problem, i.e. the problem
in the disk with the boundary condition
∂u
(a, θ) = f (θ). (5.16)
∂r
Also in this case the solution u(r, θ) has the form (5.15). The boundary condition
(5.16) implies that
 X∞
∂u 

(r, θ) = nan−1 (an cos(nθ) + bn sin(nθ)) = f (θ).
∂r r=a n=1

Hence Z 2π
1
an = f (θ) cos(nθ)dθ
πnan−1 0
and Z 2π
1
bn = f (θ) sin(nθ)dθ.
πnan−1 0
Remark. For the Neumann problem a solution is defined up to an arbitrary constant
a0
2
. Moreover, f must satisfy the consistency condition
Z 2π
f (θ)dθ = 0
0

since integrating
∞
X
f (θ) = nan−1 (an cos(nθ) + bn sin(nθ))
n=1

termwise gives us zero.

56
6 The Laplace Operator
We consider what is perhaps the most important of all partial differential operators,
the Laplace operator (Laplacian) on Rn , defined by
n
X
∆= ∂j2 ≡ ∇ · ∇.
j=1

We will start with a quite general fact about partial differential operators.

Definition. 1. A linear transformation T on Rn is called a rotation if T ′ = T −1 .

2. Let h be a fixed vector in Rn . The transformation Th f (x) := f (x + h) is called

a translation.

Theorem 1. Suppose that L is a linear partial differential operator on Rn . Then L

commutes
P with translations and rotations if and only if L is a polynomial in ∆, that
is, L ≡ m a
j=0 j ∆ j
.

Proof. Let X
L(x, ∂) ≡ aα (x)∂ α
|α|≤k

commute with a translation Th . Then

X X
aα (x)∂ α f (x + h) = aα (x + h)∂ α f (x + h).
|α|≤k |α|≤k

This implies that aα (x) must be constants (because aα (x) ≡ aα (x + h) for all h), say
aα . Next, since L now has constant coefficients we have (see Exercise 5)

c
Lu(ξ) = P (ξ)b
u(ξ),

where the polynomial P (ξ) is defined by

X
P (ξ) = aα (iξ)α .
|α|≤k

Recall from Exercise 4 that if T is a rotation then

u[
◦ T (ξ) = (b
u ◦ T ) (ξ).

Therefore
\x)(ξ) = Lu(T
(Lu)(T c ξ)

or
\
P (ξ)u(T x)(ξ) = P (T ξ)b
u(T ξ).

57
This forces
P (ξ) = P (T ξ).
Write ξ = |ξ|θ, where θ ∈ Sn−1 = {x ∈ Rn : |x| = 1} is the direction of ξ. Then
T ξ = |ξ|θ′ with some θ′ ∈ Sn−1 . But
0 = P (ξ) − P (T ξ) = P (|ξ|θ) − P (|ξ|θ′ )
shows that P (ξ) does not depend on the angle θ of ξ. Therefore P (ξ) is radial, that is,
X
P (ξ) = P1 (|ξ|) = a′α |ξ||α| .
|α|≤k

But since we know that P (ξ) is a polynomial then |α| must be even:
X
P (ξ) = aj |ξ|2j .
j

By Exercise 5 we have that

c
∆u(ξ) = −|ξ|2 u
b(ξ).
It follows by induction that
d
∆ j u(ξ) = (−1)j |ξ|2j u
b(ξ), j = 0, 1, . . . .
Taking the inverse Fourier transform we obtain
X X X
Lu = F −1 (P (ξ)b
u(ξ)) = F −1 aj |ξ|2j u
b(ξ) = F −1 d
a′j ∆ j u(ξ) = a′j ∆j u.
j j j

Conversely, let X
Lu = aj ∆j u.
j

It is clear by the chain rule that Laplacian commutes with a translation Th and a
rotation T . By induction the same is true for any power of ∆ and so for L as well.
n−1 ′
Lemma 1. If f (x) = ϕ(r), r = |x|, that is, f is radial, then ∆f = ϕ′′ (r) + r
ϕ (r).
∂r xj
Proof. Since ∂xj
= r
then
n
X n
X x 
j ′
∆f = ∂j (∂j ϕ(r)) = ∂j ϕ (r)
j=1 j=1
r
Xn x  n
X x2j
′ j
= ϕ (r)∂j + ϕ′′ (r)
j=1
r j=1
r2
Xn  2  n
X
1 xj ′
x2j
= − 3 ϕ (r) + ϕ′′ (r)
j=1
r r j=1
r2
n
X
n ′ 1 n−1 ′
= ϕ (r) − 3 x2j ϕ′ (r) + ϕ′′ (r) = ϕ′′ (r) + ϕ (r).
r r j=1
r

58
Corollary. If f (x) = ϕ(r) then ∆f = 0 on Rn \ {0} if and only if
(
a + br2−n , n 6= 2
ϕ(r) =
a + b log r, n = 2,

where a and b are arbitrary constants.

Proof. If ∆f = 0 then by Lemma 1 we have

n−1 ′
ϕ′′ (r) + ϕ (r) = 0.
r
Denote ψ(r) := ϕ′ (r). Since ψ solves the first order differential equation
n−1
ψ ′ (r) + ψ(r) = 0
r
it can be found by the use of integrating factor. Indeed, multiply by e(n−1) log r = rn−1
to get
rn−1 ψ ′ (r) + (n − 1)rn−2 ψ(r) = 0
or
′
rn−1 ψ(r) = 0.
It follows that
ϕ′ (r) = ψ(r) = cr1−n .
Integrate once more to arrive at

( ar + b, n=1
cr 2−n
2−n
+ c1 , n 6= 2 
ϕ(r) = = a log r + b, n = 2
c log r + c1 , n = 2 
 2−n
ar + b, n ≥ 3.

In the opposite direction the result follows from elementary differentiation.

Definition. A C 2 function u on an open set Ω ⊂ Rn is said to be harmonic on Ω if

∆u = 0 on Ω.

Exercise 22. For u, v ∈ C 2 (Ω) ∩ C 1 (Ω) and for S = ∂Ω, which is a surface of class
C 1 , prove the following Green’s identities:

a) Z Z
(v∆u − u∆v) dx = (v∂ν u − u∂ν v) dσ
Ω S

b) Z Z
(v∆u + ∇v · ∇u) dx = v∂ν udσ.
Ω S

59
Exercise 23. Prove that if u is harmonic on Ω and u ∈ C 1 (Ω) then
Z
∂ν udσ = 0.
S

Corollary (from Green’s identities). If u ∈ C 1 (Ω) is harmonic on Ω and

1. u = 0 on S, then u ≡ 0

2. ∂ν u = 0 on S, then u ≡ constant.
Proof. By resorting to real and imaginary parts it suffices to consider real-valued func-
tions. If we let u = v in part b) of Exercise 22 we obtain
Z Z
2
|∇u| dx = u∂ν udσ(x).
Ω S

In the case 1) we get ∇u ≡ 0 or u ≡ constant. But u ≡ 0 on S implies that u ≡ 0. In

the case 2) we can conclude only that u ≡ constant.
Theorem 2 (The mean value theorem). Suppose u is harmonic on an open set Ω ⊂ Rn .
If x ∈ Ω and r > 0 is small enough so that Br (x) ⊂ Ω, then
Z Z
1 1
u(x) = n−1 u(y)dσ(y) ≡ u(x + ry)dσ(y),
r ωn |x−y|=r ωn |y|=1
2π n/2
where ωn = Γ(n/2)
is the area of the unit sphere in Rn .

Proof. Let us apply Green’s identity a) with u and v = |y|2−n , if n 6= 2 and v = log |y|
if n = 2 in the domain

Br (x)\Bε (x) = {y ∈ Rn : ε < |x − y| < r} .

Then for v(y − x) we obtain (n 6= 2)

Z
0 = (v∆u − u∆v)dy
Br (x)\Bε (x)
Z Z
= (v∂ν u − u∂ν v)dσ(y) − (v∂ν u − u∂ν v)dσ(y)
|x−y|=r |x−y|=ε
Z Z
2−n 1−n
= r ∂ν udσ(y) − (2 − n)r udσ(y)
|x−y|=r |x−y|=r
Z Z
2−n 1−n
− ε ∂ν udσ(y) + (2 − n)ε udσ(y). (6.1)
|x−y|=ε |x−y|=ε

In order to get (6.1) we took into account that

x−yx−y d d
∂ν = ν · ∇ = =
r r dr dr

60
for the sphere. Since u is harmonic then due to Exercise 23 we can get from (6.1) that
for any ε > 0, ε < r,
Z Z
1−n 1−n
ε udσ(y) = r udσ(y).
|x−y|=ε |x−y|=r

That’s why
Z Z
1−n
lim ε u(y)dσ(y) = lim u(x + εθ)dθ
ε→0 |x−y|=ε ε→0 |θ|=1
Z
1−n
= ωn u(x) = r u(y)dσ(y).
|x−y|=r

This proves the theorem because the latter steps hold for n = 2 also.
Corollary. If u and r are as in Theorem 2 then
Z Z
n n
u(x) = n u(y)dy ≡ u(x + ry)dy, x ∈ Ω. (6.2)
r ωn |x−y|≤r ωn |y|≤1

Proof. Perform integration in polar coordinates and apply Theorem 2.

Remark. It follows from the latter formula that
ωn
vol {y : |y| ≤ 1} = .
n
Exercise 24. AssumeR that u is harmonic in Ω. Let χ(x) ∈ C0∞ (B1 (0)) be such that
χ(x) = χ1 (|x|) and Rn χ(x)dx = 1. Define an approximation to the identity by
χε (·) = ε−n χ(ε−1 ·). Prove that
Z
u(x) = χε (x − y)u(y)dy
Bε (x)

for x ∈ Ωε := {x ∈ Ω : Bε (x) ⊂ Ω}.

Corollary 1. If u is harmonic on Ω then u ∈ C ∞ (Ω).
Proof. The statement follows from Excercise 24 since the function χε is compactly
supported and we may thus differentiate under the integral sign as often as we please.

Corollary 2. If {uk }∞
k=1 is a sequence of harmonic functions on an open set Ω ⊂ R
n

which converges uniformly on compact subsets of Ω to a limit u, then u is harmonic

on Ω.
Theorem 3 (The maximum principle). Suppose Ω ⊂ Rn is open and connected. If u
is real-valued and harmonic on Ω with supx∈Ω u(x) = A < ∞, then either u < A for
all x ∈ Ω or u(x) ≡ A in Ω.

61
Proof. Since u is continuous on Ω then the set {x ∈ Ω : u(x) = A} is closed in Ω. On
the other hand due to Theorem 2 (see (6.2)) we may conclude that if u(x) = A in
some point x ∈ Ω then u(y) = A for all y in a ball about x. Indeed, if y0 ∈ Bσ′ (x) and
u(y0 ) < A then u(y) < A for all y from small neighborhood of y0 . Hence, by Corollary
of Theorem 2,
Z
n
u(x) = n u(y)dy
r ωn |x−y|≤r
Z Z
n n
= n u(y)dy + n u(y)dy
r ωn |x−y|≤r,|y0 −y|>ε r ωn |y−y0 |≤ε
 Z Z 
n n
< A n dy + n dy
r ωn |x−y|≤r,|y0 −y|>ε r ωn |y−y0 |≤ε
Z
n
= A n dy = A,
r ωn |x−y|≤r
that is, A < A. This contradiction proves our statement. This fact also means that
the set {x ∈ Ω : u(x) = A} is also open. Hence it is either Ω (in this case u ≡ A in Ω)
or the empty set (in this case u(x) < A in Ω).
Corollary 1. Suppose Ω ⊂ Rn is open and bounded. If u is real-valued and harmonic
on Ω and continuous on Ω, then the maximum and minimum of u on Ω are achieved
only on ∂Ω.
Corollary 2 (The uniqueness theorem). Suppose Ω is as in Corollary 1. If u1 and
u2 are harmonic on Ω and continuous in Ω (might be complex-valued) and u1 = u2 on
∂Ω, then u1 = u2 on Ω.
Proof. The real and imaginary parts of u1 − u2 and u2 − u1 are harmonic on Ω. Hence
they must achieve their maximum on ∂Ω. These maximum are, therefore zero, so
u1 ≡ u 2 .
Theorem 4 (Liouville’s theorem). If u is bounded and harmonic on Rn then u ≡
constant.
Proof. For any x ∈ Rn and |x| ≤ R by Corollary of Theorem 2 we have
Z Z  Z
n  
 n
|u(x) − u(0)| = n  u(y)dy − u(y)dy  ≤ n |u(y)|dy,
R ωn BR (x) BR (0) R ωn D
where
D = (BR (x)\BR (0)) ∪ (BR (0)\BR (x))
is the symmetric difference of the balls BR (x) and BR (0). That’s why we obtain
Z Z Z
n kuk∞ n kuk∞ R+|x| n−1
|u(x) − u(0)| ≤ dy ≤ r dr dθ
Rn ωn R−|x|≤|y|≤R+|x| Rn ωn R−|x| |θ|=1
 
(R + |x|)n − (R − |x|)n 1
= n
kuk∞ = O kuk∞ .
R R
Hence the difference |u(x) − u(0)| vanishes as R → ∞, that is, u(x) = u(0).

62
Definition. A fundamental solution for a partial differential operator L is a distribu-
tion K ∈ D′ such that
LK = δ.
Remark. Note that a fundamental solution is not unique. Any two fundamental solu-
tions differ by a solution of the homogeneous equation Lu = 0.
Exercise 25. Show that the characteristic function of the set

(x1 , x2 ) ∈ R2 : x1 > 0, x2 > 0

is a fundamental solution for L = ∂1 ∂2 .

1 i
Exercise 26. Prove that the Fourier transform of x1 +ix2
in R2 is equal to − ξ1 +iξ 2
.
Exercise 27. Show that the fundamental solution for the Cauchy-Riemann operator
L = 12 (∂1 + i∂2 ) on R2 is equal to
1 1
.
π x1 + ix2
Since the Laplacian commutes with rotations (Theorem 1) it should have a radial
fundamental solution which must be a function of |x| that is harmonic on Rn \ {0}.
Theorem 5. Let (
|x|2−n
(2−n)ωn
, n 6= 2
K(x) = 1
(6.3)
2π
log |x|, n = 2.
Then K is a fundamental solution for ∆.
Proof. For ε > 0 we consider a smoothed out version Kε of K as
( 2 2 2−n
(|x| +ε ) 2
, n=6 2
Kε (x) = (2−n)ωn (6.4)
1 2 2
4π
log(|x| + ε ), n = 2.

Then Kε → K pointwise (x 6= 0) as ε → +0 and Kε and K are dominated by a fixed

locally integrable function for ε ≤ 1 (namely, by |K| for n > 2, | log |x|| + 1 for n = 2
and (|x|2 + 1)1/2 for n = 1). So by the Lebesgue’s dominated convergence theorem
Kε → K in L1loc (or in the topology of distributions) when ε → +0. Hence we need to
show only that ∆Kε → δ as ε → 0 in the sense of distributions, that is,

h∆Kε , ϕi → ϕ(0), ε→0

for any ϕ ∈ C0∞ (Rn ).

Exercise 28. Prove that

∆Kε (x) = nωn−1 ε2 (|x|2 + ε2 )−( 2 +1) ≡ ε−n ψ(ε−1 x)

n

for ψ(y) = nωn−1 (|y|2 + 1)−( 2 +1) .

n

63
 Exercise 28 allows us to write
Z Z Z
−n −1
h∆Kε , ϕi = ϕ(x)ε ψ(ε x)dx = ϕ(εz)ψ(z)dz → ϕ(0) ψ(z)dz
Rn Rn Rn

as ε → +0. So it remains to show that

Z
ψ(z)dz = 1.
Rn

Using Exercise 28 we have

Z Z Z ∞ Z
n 2 −( n +1 ) n n−1 2 − ( n
+1)
ψ(x)dx = (|x| + 1) 2 dx = r (r + 1) 2 dr dθ
Rn ωn Rn ωn 0 |θ|=1
Z ∞ Z
n−1 2 −( n +1) n ∞ (n−1)/2 n 1
=n r (r + 1) 2 dr = t (1 + t)− 2 −1 √ dt
0 2 0 t
Z ∞ Z 1 n/2−1
n n n 1 n ds
= tn/2−1 (1 + t)− 2 −1 dt = −1 s 2 +1 2
2 0 2 0 s s
Z 1 Z 1
n n
= (1 − s)n/2−1 ds = τ n/2−1 dτ = 1.
2 0 2 0

It means that ε−1 ψ(ε−1 x) is an approximation to the identity and

∆Kε → δ.

But Kε → K and so ∆K = δ also.

Theorem 6. Suppose that

1. f ∈ L1 (Rn ) if n ≥ 3
R
2. R2 |f (y)| (| log |y|| + 1) dy < ∞ if n = 2
R
3. R |f (y)| (1 + |y|) dy < ∞ if n = 1.

Let K be given by (6.3). Then f ∗ K is well-defined as a locally integrable function and

∆(f ∗ K) = f in the sense of distributions.

Proof. Let n ≥ 3 and set (

1, x ∈ B1 (0)
χ1 (x) =
0, x ∈
/ B1 (0).
Then χ1 K ∈ L1 (Rn ) and (1 − χ1 )K ∈ L∞ (Rn ). So, for f ∈ L1 (Rn ) we have that
f ∗ (χ1 K) ∈ L1 (Rn ) and f ∗ (1 − χ1 )K ∈ L∞ (Rn ) (see Proposition 1 of Chapter 0).

64
Hence f ∗ K ∈ L1loc (Rn ) by addition and we may calculate
h∆(f ∗ K), ϕi = hf ∗ K, ∆ϕi, ϕ ∈ C0∞ (Rn )
Z Z Z
= (f ∗ K)(x)∆ϕ(x)dx = f (y)K(x − y)dy∆ϕ(x)dx
ZR
n Rn Rn
Z Z
= f (y) K(x − y)∆ϕ(x)dxdy = f (y)hK(x − y), ∆ϕ(x)idy
Rn Rn Rn
Z Z
= f (y)h∆K(x − y), ϕ(x)idy = f (y)hδ(x − y), ϕ(x)idy
Rn Rn
Z
= f (y)ϕ(y)dy = hf, ϕi.
Rn

Hence ∆(f ∗ K) = f .

Exercise 29. Prove Theorem 6 for n = 2.

Exercise 30. Prove Theorem 6 for n = 1.
Theorem 7. Let Ω be a bounded domain in Rn (for n = 1 assume that Ω = (a, b))
with C 1 boundary ∂Ω = S. If u ∈ C 1 (Ω) is harmonic in Ω, then
Z


u(x) = u(y)∂νy K(x − y) − K(x − y)∂ν u(y) dσ(y), x ∈ Ω, (6.5)
S

where K(x) is the fundamental solution (6.3).

Proof. Let us consider Kε from (6.4). Then since ∆u = 0 in Ω, by Green’s identity a)
(see Exercise 22) we have
Z Z


u(y)∆y Kε (x − y)dy = u(y)∂νy Kε (x − y) − Kε (x − y)∂ν u(y) dσ(y).
Ω S

As ε → 0 the right hand side of this equation tends to the right hand side of (6.5)
for each x ∈ Ω. This is because for x ∈ Ω and y ∈ S there are no singularities in K.
On the other hand, the left hand side is just (u ∗ ∆Kε ) (x) if we set u ≡ 0 outside Ω.
According to the proof of Theorem 5
(u ∗ ∆Kε ) (x) → u(x), ε → 0,
completing the proof.
Remark. If we know that u = f and ∂ν u = g on S then
Z


u(x) = f (y)∂νy K(x − y) − K(x − y)g(y) dσ(y)
S

is the solution of ∆u = 0 with Cauchy data on S. But this problem is overdetermined

because we know from Corollary 2 of Theorem 3 that the solution of ∆u = 0 is uniquely
determined by f alone.

65
 The following theorem concerns spaces C α (Ω) and C k,α (Ω) which are defined by

C α (Ω) ≡ C 0,α (Ω) = {u ∈ L∞ (Ω) : |u(x) − u(y)| ≤ C|x − y|α , x, y ∈ Ω}


C k,α (Ω) ≡ C k+α (Ω) = u : ∂ β u ∈ C α (Ω), |β| ≤ k

for 0 < α < 1 and k ∈ N.

Theorem 8. Suppose k ≥ 0 is an integer, 0 < α < 1 and Ω ⊂ Rn open. If f ∈ C k+α (Ω)

k+2+α
and u is a distributional solution of ∆u = f in Ω, then u ∈ Cloc (Ω).

Proof. Since ∆(∂ β u) = ∂ β ∆u = ∂ β f we can assume without loss of generality that

k = 0. Given Ω1 ⊂ Ω such that Ω1 ⊂ Ω pick ϕ ∈ C0∞ (Ω) such that ϕ ≡ 1 on Ω1 and
let g = ϕf .
Since ∆(g ∗ K) = g (see Theorem 6) and therefore ∆(g ∗ K) = f in Ω1 , then
u − (g ∗ K) is harmonic in Ω1 and hence C ∞ there. It is therefore enough to prove
that if g is a C α function with compact support, then g ∗ K ∈ C 2+α . To this end
we consider Kε (x) and its derivatives. Straightforward calculations lead to following
formulae (n ≥ 1):

∂
Kε (x) = ωn−1 xj (|x|2 + ε2 )−n/2
∂xj
( (6.6)
∂2 −nxi xj (|x|2 + ε2 )−n/2−1 , i 6= j
Kε (x) = ωn−1
∂xi ∂xj (|x|2 + ε2 − nx2j )(|x|2 + ε2 )−n/2−1 , i = j.

Exercise 31. Prove formulae (6.6).

Since Kε ∈ C ∞ then g ∗ Kε ∈ C ∞ also. Moreover, ∂j (g ∗ Kε ) = g ∗ ∂j Kε and
∂i ∂j (g ∗ Kε ) = g ∗ ∂i ∂j Kε . The pointwise limits in (6.6) as ε → 0 imply

∂
K(x) = ωn−1 xj |x|−n
∂xj
( (6.7)
∂2 −nωn−1 xi xj |x|−n−2 , i 6= j
K(x) =
∂xi ∂xj ωn−1 (|x|2 − nx2j )|x|−n−2 , i = j,

for x 6= 0. The formulae (6.7) show that ∂j K(x) is a locally integrable function
and since g is bounded with compact support then g ∗ ∂j K is continuous. Next,
g ∗ ∂j Kε → g ∗ ∂j K uniformly as ε → +0. It is equivalent to ∂j Kε → ∂j K in the
topology of distributions (see the definition). Hence ∂j (g ∗ K) = g ∗ ∂j K.
This argument does not work for the second derivatives because ∂i ∂j K(x) is not
integrable. But there is a different procedure for these terms.
Let i 6= j. Then ∂i ∂j Kε (x) and ∂i ∂j K(x) are odd functions of xi (and xj ), see (6.6)
and (6.7). Due to this fact their integrals over any annulus 0 < a < |x| < b vanish.
For Kε we can even take a = 0.
Exercise 32. Prove this fact.

66
 That’s why for any b > 0 we have
Z Z
g ∗ ∂i ∂j Kε (x) = g(x − y)∂i ∂j Kε (y)dy − g(x) ∂i ∂j Kε (y)dy
Rn |y|<b
Z Z
= (g(x − y) − g(x))∂i ∂j Kε (y)dy + g(x − y)∂i ∂j Kε (y)dy.
|y|<b |y|≥b

If we let ε → 0 we obtain
Z Z
lim g ∗ ∂i ∂j Kε (x) = (g(x − y) − g(x))∂i ∂j K(y)dy + g(x − y)∂i ∂j K(y)dy.
ε→0 |y|<b |y|≥b

This limit exists because

|g(x − y) − g(x)||∂i ∂j K(y)| ≤ c|y|α |y|−n

(g is C α ) and because g is compactly supported. Then, since b is arbitrary, we can let

b → +∞ to obtain
Z
∂i ∂j (g ∗ K)(x) = lim (g(x − y) − g(x))∂i ∂j K(y)dy
b→∞ |y|<b
Z
+ lim g(x − y)∂i ∂j K(y)dy
b→∞ |y|≥b
Z
= lim (g(x − y) − g(x))∂i ∂j K(y)dy. (6.8)
b→∞ |y|<b

A similar result holds for i = j. Indeed,

1 −n
∂j2 Kε (x) = ε ψ(ε−1 x) + Kjε (x),
n
where ψ(x) = nωn−1 (|x|2 + 1)−n/2−1 and Kjε = ωn−1 (|x|2 − nx2j )(|x|2 + ε2 )−n/2−1 (see
(6.6)). The integral Ij of Kjε over an annulus a < |y| < b vanishes. Why is it so? First
of all, Ij is independent of
Pjn by symmetryPin the coordinates, that is, Ij = Ii for i 6= j.
n
ε
So nIj is the integral of j=1 Kj . But j=1 Kjε = 0. Hence Ij = 0 also. That’s why
we can apply the same procedure. Since

g ∗ (ε−n ψ(ε−1 x)) → g, ε → 0,

(because ε−n ψ(ε−1 x) is an approximation to the identity) then

Z
2 g(x)
∂j (g ∗ K)(x) = + lim (g(x − y) − g(x))∂j2 K(y)dy. (6.9)
n b→∞ |y|<b

Since the convergence in (6.8) and (6.9) is uniform then at this point we have shown
that g ∗ K ∈ C 2 . But we need to prove more.

67
Lemma 2 (Calderon-Zigmund). Let N be a C 1 function on Rn \ {0} that is homoge-
neous of degree −n and satisfies
Z
N (y)dy = 0
a<|y|<b

for any 0 < a < b < ∞. Then if g is a C α function with compact support, 0 < α < 1,
then Z
h(x) = lim (g(x − z) − g(x))N (z)dz
b→∞ |z|<b

belongs to C α .
Proof. Let us write h = h1 + h2 , where
Z
h1 (x) = (g(x − z) − g(x))N (z)dz,
|z|≤3|y|
Z
h2 (x) = lim (g(x − z) − g(x))N (z)dz.
b→∞ 3|y|<|z|<b

We wish to estimate h(x + y) − h(x). Since α > 0 we have

Z
|h1 (x)| ≤ c |z|α |z|−n dz = c′ |y|α .
|z|≤3|y|

and hence
|h1 (x + y) − h1 (x)| ≤ |h1 (x + y)| + |h1 (x)| ≤ 2c′ |y|α .
On the other hand
Z
h2 (x + y) − h2 (x) = lim (g(x − z) − g(x))N (z + y)dz
b→∞ 3|y|<|z+y|<b
Z
− lim (g(x − z) − g(x))N (z)dz
b→∞ 3|y|<|z|<b
Z
= lim (g(x − z) − g(x))(N (z + y) − N (z))dz
b→∞ 3|y|<|z|<b
Z
+ lim (g(x − z) − g(x))N (z + y)dz
b→∞ {3|y|<|z+y|<b}\{3|y|<|z|<b}
= I1 + I2 .

It is clear that

{3|y| < |z + y|} \ {3|y| < |z|} ⊂ {2|y| < |z|} \ {3|y| < |z|}
= {2|y| < |z| ≤ 3|y|} .

68
That’s why
Z
|I2 | ≤ |g(x − z) − g(x)||N (z + y)|dz
2|y|<|z|≤3|y|
Z
≤ c |z|α |z + y|−n dz
Z2|y|<|z|≤3|y|
≤ c′ |z|α−n dz = c′′ |y|α .
2|y|<|z|≤3|y|

Now we observe that for |z| > 3|y|

|N (z + y) − N (z)| ≤ |y| sup |∇N (z + ty)|
0≤t≤1

≤ c|y| sup |z + ty|−n−1 ≤ c′ |y||z|−n−1 ,

0≤t≤1

because ∇N is homogeneous of degree −n − 1, since N is homogeneous of degree −n.

Hence
Z Z ∞
α −n−1 ′
|I1 | ≤ c |z| |y||z| dz = c |y| ρα−2 dρ = c′′ |y|α .
|z|>3|y| 3|y|

Note that the condition α < 1 is needed here. Collecting the estimates for I1 and I2
we can see that the lemma is proved.
In order to end the proof of Theorem it remains to note that ∂i ∂j K(x) satisfies all
conditions of Lemma 2. Thus the Theorem is also proved.
Exercise 33. Show that a function K1 is a fundamental solution for ∆2 ≡ ∆(∆) on
Rn if and only if K1 satisfies the equation
∆K1 = K,
where K is the fundamental solution for the Laplacian.
Exercise 34. Show that the following functions are the fundamental solutions for ∆2
on Rn :
1. n = 4:
log |x|
−
4ω4
2. n = 2:
|x|2 log |x|
8π
3. n 6= 2, 4:
|x|4−n
.
2(4 − n)(2 − n)ωn
Exercise 35. Show that (4π|x|)−1 e−c|x| is the fundamental solution for −∆ + c2 on R3
for any constant c ∈ C.

69
7 The Dirichlet and Neumann Problems
The Dirichlet problem
Given functions f in Ω and g on S = ∂Ω, find a function u in Ω = Ω ∪ ∂Ω satisfying
(
∆u = f, in Ω
(D)
u = g, on S.

The Neumann problem

Given functions f in Ω and g on S, find a function u in Ω satisfying
(
∆u = f, in Ω
(N)
∂ν u = g, on S.

We assume that Ω is bounded with C 1 boundary. But we shall not, however, assume
that Ω is connected. The uniqueness theorem (see Corollary of Theorem 3 of Chapter
6) shows that the solution of (D) will be unique (if it exists), at least if we require
u ∈ C(Ω). For (N) uniqueness does not hold: we can add to u(x) any function that is
constant on each connected component of Ω. Moreover, there is an obvious necessary
condition for solvability of (N). If Ω′ is a connected component of Ω then
Z Z Z Z
∆udx = ∂ν udσ(x) = g(x)dσ(x) = f dx,
Ω′ ∂Ω′ ∂Ω′ Ω′

that is, Z Z
f (x)dx = g(x)dσ(x).
Ω′ ∂Ω′
It is also clear (by linearity) that (D) can be reduced to the following homogeneous
problems: (
∆v = f, in Ω
(DA )
v = 0, on S
(
∆w = 0, in Ω
(DB )
w = g, on S
and u := v + w solves (D). Similar remarks apply to (N), that is
(
∆v = f, in Ω
∂ν v = 0, on S
(
∆w = 0, in Ω
∂ν w = g, on S
and u = v + w.

70
Definition. The Green’s function for (D) in Ω is the solution G(x, y) of the boundary
value problem (
∆x G(x, y) = δ(x − y), x, y ∈ Ω
(7.1)
G(x, y) = 0, x ∈ S, y ∈ Ω.
Analogously, the Green’s function for (N) in Ω is the solution G(x, y) of the boundary
value problem (
∆x G(x, y) = δ(x − y), x, y ∈ Ω
(7.2)
∂νx G(x, y) = 0, x ∈ S, y ∈ Ω.

This definition allows us to write

G(x, y) = K(x − y) + vy (x), (7.3)

where K is the fundamental solution of ∆ in Rn and, for any y ∈ Ω, the function vy (x)
satisfies (
∆vy (x) = 0, in Ω
(7.4)
vy (x) = −K(x − y), on S
in the case of (7.1) and
(
∆vy (x) = 0, in Ω
∂νx vy (x) = −∂νx K(x − y), on S

in the case of (7.2). Since (7.4) guarantees that vy is real then so is G corresponding
to (7.1).

Lemma 1. The Green’s function (7.1) exists and is unique.

Proof. The uniqueness of G follows again from Corollary 2 of Theorem 3 of Chapter

6, since K(x − y) in (7.4) is continuous for all x ∈ S and y ∈ Ω (x 6= y). The existence
will be proved later.

Lemma 2. For both (7.1) and (7.2) it is true that G(x, y) = G(y, x) for all x, y ∈ Ω.

Proof. Let G(x, y) and G(x, z) be the Green’s functions for Ω corresponding to sources
located at fixed y and z, y 6= z, respectively. Let us consider the domain

Ωε = (Ω\ {x : |x − y| < ε}) \ {x : |x − z| < ε} ,

see Figure 2.

71
 b
z

b
y

Figure 2: The domain Ωε .

If x ∈ Ωε then x 6= z and x 6= y and, therefore, ∆x G(x, z) = 0 and ∆x G(x, y) = 0.

These facts imply
Z
0 = (G(x, y)∆x G(x, z) − G(x, z)∆x G(x, y)) dx
Ωε
Z
= (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ(x)
Z S

− (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ(x)

|x−y|=ε
Z
− (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ(x).
|x−z|=ε

Hence, by (7.1) or (7.2), for arbitrary ε > 0 (small enough)

Z
(G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ(x)
|x−y|=ε
Z
= (G(x, z)∂νx G(x, y) − G(x, y)∂νx G(x, z)) dσ(x).
|x−z|=ε

Let n ≥ 3. Due to (7.3) for ε → 0 we have

Z
(G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ(x)
|x−y|=ε
Z  
2−n (x − y, x − z)
≈ cn ε (2 − n) + ∂νx vz (x) dσ(x)
|x−y|=ε |x − y||x − z|n
Z
− G(x, z)∂νx G(x, y)dσ(x)
|x−y|=ε
Z
2−n n−1 1 (εθ, εθ + y − z)
≈ cn (2 − n)ε ε dθ − I1 ≈ −I1 ,
ε θ |εθ + y − z|n
where we have denoted
Z
I1 = G(x, z)∂νx G(x, y)dσ(x).
|x−y|=ε

72
The same is true for the integral over |x − z| = ε, that is,
Z
(G(x, z)∂νx G(x, y) − G(x, y)∂νx G(x, z)) dσ(x) ≈ −I2 , ε → 0,
|x−z|=ε

where Z
I2 = G(x, y)∂νx G(x, z)dσ(x).
|x−z|=ε

But using the previous techniques we can obtain that

Z
1−n n−1
I1 ≈ (2 − n)cn ε ε G(εθ + y, z)dθ → (2 − n)cn ωn G(y, z), ε→0
|θ|=1

and
Z
1−n n−1
I2 ≈ (2 − n)cn ε ε G(εθ + z, y)dθ → (2 − n)cn ωn G(z, y), ε → 0.
|θ|=1

It means that G(y, z) = G(z, y) for all z 6= y. This proof holds for n = 2 (and even for
n = 1) with some simple changes.

Lemma 3. In three or more dimensions

K(x − y) < G(x, y) < 0, x, y ∈ Ω, x 6= y

where G(x, y) is the Green’s function for (D).

Proof. For each fixed y, the function vy (x) := G(x, y) − K(x − y) is harmonic in Ω, see
(7.4). Moreover, on S = ∂Ω, vy (x) takes on the positive value

|x − y|2−n
−K(x − y) ≡ − .
ωn (2 − n)

By the minimum principle, it follows that vy (x) is strictly positive in Ω. This proves
the first inequality.
Exercise 36. Prove the second inequality in Lemma 3.

Exercise 37. Show that for n = 2 Lemma 3 has the following form:

1 |x − y|
log < G(x, y) < 0, x, y ∈ Ω,
2π h
where h ≡ maxx,y∈Ω |x − y|.

73
Exercise 38. Obtain the analogue of Lemma 3 for n = 1. Hint: show that the Green’s
d2
function for the operator dx 2 on Ω = (0, 1) is

(
x(y − 1), x < y
G(x, y) =
y(x − 1), x > y.

Remark. G(x, y) may be extended naturally (because of the symmetry) to Ω × Ω by

setting G(x, y) = 0 for y ∈ S.
Now we can solve both problems (DA ) and (DB ). Indeed, let us set f = 0 in (DA )
outside Ω and define
Z Z
v(x) := G(x, y)f (y)dy ≡ (f ∗ K)(x) + (G(x, y) − K(x − y)) f (y)dy.
Ω Ω

Then the Laplacian of the first term is f (see Theorem 6 of Chapter 6), and the second
term is harmonic in x (since vy (x) is harmonic). Also v(x) = 0 on S because the same
is true for G. Thus, this v(x) solves (DA ).
Consider now (DB ). We assume that g is continuous on S and we wish to find
w which is continuous on Ω. Applying Green’s identity a) (together with the same
limiting process as in the proof of Lemma 2) we obtain
Z
w(x) = (w(y)∆y G(x, y) − G(x, y)∆w(y)) dy
ZΩ Z
= w(y)∂νy G(x, y)dσ(y) = g(y)∂νy G(x, y)dσ(y).
S S

Let us denote the last integral by (P). Since ∂νy G(x, y) is harmonic in x and continuous
in y for x ∈ Ω and y ∈ S then w(x) is harmonic in Ω. In order to prove that this w(x)
solves (DB ) it remains to prove that w(x) is continuous in Ω and w(x) on S is g(x).
We will prove this general fact later.
Definition. The function ∂νy G(x, y) on Ω × S is called the Poisson kernel for Ω and
(P) is called the Poisson integral .
Now we are in the position to solve the Dirichlet problem in a half-space. Let

Ω = Rn+1
+ = (x′ , xn+1 ) ∈ Rn+1 : x′ ∈ Rn , xn+1 > 0 ,

where n ≥ 1 now, and let xn+1 = t. Then

∆n+1 = ∆n + ∂t2 , n = 1, 2, . . .

Denote by K(x, t) a fundamental solution for ∆n+1 in Rn+1 , that is,

( 2 2 1−n
(|x| +t ) 2
, n>1
K(x, t) = (1−n)ωn+1
1
4π
log(|x|2 + t2 ), n = 1.

74
Let us prove then that the Green’s function for Rn+1
+ is
G(x, y; t, s) = K(x − y, t − s) − K(x − y, −t − s). (7.5)
It is clear (see (7.5)) that G(x, y; t, 0) = G(x, y; 0, s) = 0 and
∆n+1 G = δ(x − y, t − s) − δ(x − y, −t − s) = δ(x − y)δ(t − s)
because for t, s > 0, −t − s < 0 and, therefore, δ(−t − s) = 0. Thus G is the Dirichlet
Green’s function for Rn+1+ . From this we immediately have the solution of (DA ) in
n+1
R+ as Z Z ∞
u(x, t) = G(x, y; t, s)f (y, s)dsdy.
Rn 0
To solve (DB ) we compute the Poisson kernel for this case. Since the outward normal
derivative on ∂Rn+1
+
∂
is − ∂t then the Poisson kernel becomes
∂ ∂
− G(x, y; t, s)|s=0 = − (K(x − y, t − s) − K(x − y, −t − s)) |s=0
∂s ∂s
2t
= n+1 . (7.6)
ωn+1 (|x − y|2 + t2 ) 2
Exercise 39. Prove (7.6).
Note that (7.6) holds for any n ≥ 1. According to the formula for (P), the candidate
for a solution to (DB ) is:
Z
2 tg(y)
u(x, t) = dy. (7.7)
ωn+1 Rn (|x − y|2 + t2 ) n+1
2

In other words, if we set

2t
Pt (x) := n+1 , (7.8)
ωn+1 (|x|2 + t2 ) 2

which is what is usually called the Poisson kernel for Rn+1

+ , the proposed solution (7.7)
is simply equal to
u(x, t) = (g ∗ Pt )(x). (7.9)
Exercise 40. Prove that Pt (x) = t−n P1 (t−1 x) and
Z
Pt (y)dy = 1.
Rn

Theorem 1. Suppose g ∈ Lp (Rn ), 1 ≤ p ≤ ∞. Then u(x, t) from (7.9) is well-defined

on Rn+1
+ and is harmonic there. If g is bounded and uniformly continuous, then u(x, t)
is continuous on Rn+1
+ and u(x, 0) = g(x), and
ku(·, t) − g(·)k∞ → 0
as t → +0.

75
Proof. It is clear that for any t > 0, Pt (x) ∈ L1 (Rn ) ∩ L∞ (Rn ), see (7.8). Hence
Pt (x) ∈ Lq (Rn ) for all q ∈ [1, ∞] with respect to x and for any fixed t > 0. Therefore,
the integral in (7.9) is absolutely convergent and the same is true if Pt is replaced by
its derivatives ∆x Pt or ∂t2 Pt (due to Young’s inequality for convolution).
Since G(x, y; t, s) is harmonic for (x, t) 6= (y, s) then Pt (x) is also harmonic and
∆x u + ∂t2 u = g ∗ (∆x + ∂t2 )Pt = 0.
It remains to prove that if g is bounded and continuous then
ku(·, t) − g(·)k∞ → 0

as t → +0 and, therefore, u(x, 0) = g(x) and u is continuous on Rn+1 + .

We have (see Exercise 40)
Z Z 
 
kg ∗ Pt − gk∞ = sup   g(x − y)Pt (y)dy − g(x)Pt (y)dy 
x∈Rn
ZR
n Rn

≤ sup |g(x − y) − g(x)||Pt (y)|dy

x∈Rn Rn
Z
= sup |g(x − tz) − g(x)||P1 (z)|dz
x∈Rn Rn
Z
= sup |g(x − tz) − g(x)||P1 (z)|dz
x∈Rn |z|<R
Z 
+ |g(x − tz) − g(x)||P1 (z)|dz
|z|≥R
Z
≤ sup |g(x − tz) − g(x)| + 2 kgk∞ |P1 (z)|dz < ε
x∈Rn ,|z|<R |z|≥R

for t small enough.

The first term in the latter sum can be made less than ε/2 since g is uniformly
continuous on Rn . The second term can be made less than ε/2 for R large enough
since P1 ∈ L1 (Rn ). Thus, the theorem is proved.
Remark. The solution of the considered problem is not unique: if u(x, t) is a solution
then so is u(x, t) + ct for any c ∈ C. However, we have the following theorem.
Theorem 2. If g ∈ C(Rn ) and limx→∞ g(x) = 0 then u(x, t) := (g ∗ Pt )(x) → 0 as
(x, t) → ∞ in Rn+1
+ and it is the unique solution with this property.
Proof. Assume for the moment that g has compact support, say g = 0 for |x| > R.
Then g ∈ L1 (Rn ) and
kg ∗ Pt k∞ ≤ kgk1 kPt k∞ ≤ ct−n ,
so u(x, t) → 0 as t → ∞ uniformly in x. On the other hand, if 0 < t ≤ T , then
2t
|u(x, t)| ≤ kgk1 sup |Pt (x − y)| = kgk1 sup n+1 ≤ cT |x|−n−1 ,
|y|<R |y|<R ωn+1 (|x − y|2 + t2 ) 2

76
for |x| > 2R. Hence u(x, t) → 0 as x → ∞ uniformly for t ∈ [0, T ]. This proves
that u(x, t) vanishes at infinity if g(x) has compact support. For general g, choose a
sequence {gk } of compactly supported functions that converges uniformly (in L∞ (Rn ))
to g and let
uk (x, t) = (gk ∗ Pt )(x).
Then
Z 
 
kuk − ukL∞ (Rn+1 ) = sup  (gk − g)(y)Pt (x − y)dy 
t,x
R
n
Z 
≤ sup kgk − gkL∞ (Rn ) sup |Pt (x − y)|dy
t x Rn
Z
= kgk − gkL∞ (Rn ) sup |Pt (y)|dy = kgk − gkL∞ (Rn ) → 0
t>0 Rn

as k → ∞.
Hence u(x, t) vanishes at infinity. Now suppose v is another solution and let w :=
v − u. Then w vanishes at infinity and also at t = 0 (see Theorem 1). Thus |w| < ε on
the boundary of the region {(x, t) : |x| < R, 0 < t < R} for R large enough.

t=R

b
t=0
R

But since w is harmonic then by the maximum principle it follows that |w| < ε on this
region. Letting ε → 0 and R → ∞ we conclude that w ≡ 0.
Let us consider now the Dirichlet problem in a ball. We use here the following
notation:
B = B1 (0) = {x ∈ Rn : |x| < 1} , ∂B = S.

Exercise 41. Prove that  

x 

|x − y| =  − y|x|
|x|
for x, y ∈ Rn , x 6= 0, |y| = 1.

77
 Now, assuming first that n > 2, define
 
x2−n
G(x, y) := K(x − y) − |x| K −y
|x|2
 2−n !
1  x 
= |x − y|2−n −  − y|x| , x 6= 0. (7.10)
(2 − n)ωn |x|

Exercise 41 shows that G(x, y) from (7.10) satisfies G(x, y) = 0, x ∈ B, y ∈ S. It is

also clear that G(x, y) = G(y, x). This is true because
 2  2
x   
 − y|x|  =  x  − 2(x, y) + |y|2 |x|2 = 1 − 2(x, y) + |y|2 |x|2
 |x|   |x| 
 2  2
y  y 
=   − 2(y, x) + |x| |y| =  − x|y| .
  2 2 
|y| |y|
Next, for x, y ∈ B we have that
 
 x  |x| 1
 
 |x|2  = |x|2 = |x| > 1
x
and y 6= |x|2
. Hence,
 
2−n x
G(x, y) − K(x − y) ≡ −|x| K −y
|x|2
is harmonic in y. But symmetry of G and K shows also that G(x, y) − K(x − y) is
harmonic in x. Thus, G(x, y) is the Green’s function for B. It also makes clear how
to define G at x = 0 (and at y = 0):
1
G(0, y) = (|y|2−n − 1)
(2 − n)ωn
since  
x 
 − y|x| →1
 |x| 
as x → 0.
For n = 2 the analogous formulae are:
  
1 x  1
G(x, y) = 
log |x − y| − log  − y|x| , G(0, y) = log |y|.
2π |x| 2π
Now we can compute the Poisson kernel P (x, y) := ∂νy G(x, y), x ∈ B, y ∈ S. Since
∂νy = y · ∇y on S, then
  
x
1  (y, x − y) |x|
− y|x|, y|x| 1 − |x|2
P (x, y) = − −    ≡ , n ≥ 2. (7.11)
ωn |x − y|n x n ωn |x − y|n
 − y|x||x|

78
Exercise 42. Prove (7.11).
Theorem 3. If f ∈ L1 (S) then
Z
u(x) = P (x, y)f (y)dσ(y), x ∈ B,
S

is harmonic. If f ∈ C(S) then u extends continuously to B and u = f on S.

Proof. For each x ∈ B (see (7.11)), P (x, y) is a bounded function of y ∈ S, so u(x) is
well-defined for f ∈ L1 (S). It is also harmonic in B, because P (x, y) is harmonic for
x 6= y. Next, we claim that
Z
P (x, y)dσ(y) = 1. (7.12)
S

Since P is harmonic in x then the mean value theorem implies (y ∈ S)

Z
1 = ωn P (0, y) = P (ry ′ , y)dσ(y ′ )
S

for any 0 < r < 1. But

P (ry ′ , y) = P (y, ry ′ ) = P (ry, y ′ )

if y, y ′ ∈ S. The last formula follows from

|ry ′ − y|2 = r2 − 2r(y ′ , y) + 1 = |ry − y ′ |2 .

That’s why we may conclude that

Z Z
′ ′
1= P (ry , y)dσ(y ) = P (x, y ′ )dσ(y ′ )
S S

with x = ry. This proves (7.12). We claim also that for any y0 ∈ S and for the
neighborhood Bσ (y0 ) ⊂ S,
Z
lim P (ry0 , y)dσ(y) = 0. (7.13)
r→1−0 S\Bσ (y0 )

Indeed, for y0 , y ∈ S and 0 < r < 1,

|ry0 − y| > r|y0 − y|

and therefore
|ry0 − y|−n < (r|y0 − y|)−n ≤ (rσ)−n
if y ∈ S\Bσ (y0 ), i.e., |y − y0 | ≥ σ. Hence |ry0 − y|−n is bounded uniformly for r → 1− 0
and y ∈ S\Bσ (y0 ). In addition, 1 − |ry0 |2 ≡ 1 − r2 → 0 as r → 1 − 0. This proves
(7.13).

79
 Now, suppose f ∈ C(S). Hence f is uniformly continuous since S is compact.
That’s why for any ε > 0 there exists δ > 0 such that

|f (x) − f (y)| < ε, x, y ∈ S, |x − y| < δ.

For any x ∈ S and 0 < r < 1, by (7.12),

Z 
 

|u(rx) − f (x)| =  (f (y) − f (x))P (rx, y)dσ(y)
ZS
≤ |f (y) − f (x)||P (rx, y)|dσ(y)
|x−y|<δ
Z
+ |f (y) − f (x)||P (rx, y)|dσ(y)
S\Bδ (x)
Z Z
≤ ε |P (rx, y)|dσ(y) + 2 kf k∞ |P (rx, y)|dσ(y)
S S\Bδ (x)
Z
≤ ε + 2 kf k∞ P (rx, y)dσ(y) → 0,
S\Bδ (x)

as ε → 0 and r → 1 − 0 by (7.13). Hence u(rx) → f uniformly as r → 1 − 0.

Corollary (without proof). If f ∈ Lp (S), 1 ≤ p ≤ ∞, then

ku(r·) − f (·)kp → 0

as r → 1 − 0.

Exercise 43. Show that the Poisson kernel for the ball BR (x0 ) is

R2 − |x − x0 |2
P (x, y) = , n ≥ 2.
ωn R|x − y|n

Exercise 44 (Harnack’s inequality). Suppose u ∈ C(B) is harmonic on B and u ≥ 0.

Then show that for |x| = r < 1,
1−r 1+r
n−1
u(0) ≤ u(x) ≤ u(0).
(1 + r) (1 − r)n−1

Theorem 4 (The Reflection Principle). Let Ω ⊂ Rn+1 , n ≥ 1, be open and carry

the property that (x, −t) ∈ Ω if (x, t) ∈ Ω. Let Ω+ = {(x, t) ∈ Ω : t > 0} and Ω0 =
{(x, t) ∈ Ω : t = 0}. If u(x, t) is continuous on Ω+ ∪ Ω0 , harmonic in Ω+ and u(x, 0) =
0, then we can extend u to be harmonic on Ω by setting u(x, −t) := −u(x, t).

Definition. If u is harmonic on Ω\ {x0 }, Ω ⊂ Rn open, then u is said to have a

removable singularity x0 if u can be defined at x0 so as to be harmonic in Ω.

80
Theorem 5. Suppose u is harmonic on Ω\ {x0 } and u(x) = o (|x − x0 |2−n ) for n > 2
and u(x) = o (log |x − x0 |) for n = 2 as x → x0 . Then u has a removable singularity
at x0 .

Proof. Without loss of generality we assume that Ω = B := B1 (0) and x0 = 0. Since

u is continuous on ∂B then by Theorem 3 there exists v ∈ C(B) satisfying
(
∆v = 0, in B
v = u, on S.

We claim that u = v in B\ {0}, so that we can remove the singularity at {0} by setting
u(0) := v(0). Indeed, given ε > 0 and 0 < δ < 1 consider the function
(
u(x) − v(x) − ε(|x|2−n − 1), n > 2
gε (x) =
u(x) − v(x) + ε log |x|, n=2

in B\Bδ (0). These functions are real (we can assume without loss of generality),
harmonic and continuous for δ ≤ |x| ≤ 1. Moreover gε (x) = 0 on ∂B and gε (x) < 0
on ∂Bδ (0) for all δ small enough. By the maximum principle, it is negative in B\ {0}.
Letting ε → 0 we see that u − v ≤ 0 in B\ {0}. By the same arguments we may
conclude that also v − u ≤ 0 in B\ {0}. Hence u = v in B\ {0} and we can extend u
to the whole ball by setting u(0) = v(0). This proves the theorem.

81
8 Layer Potentials
In this chapter we assume that Ω ⊂ Rn is bounded and open, and that S = ∂Ω is a
surface of class C 2 . We assume also that both Ω and Ω′ := Rn \ Ω are connected.

Definition. Let ν(x) be a normal vector to S at x. Then

∂ν− u(x) := lim ν(x) · ∇u(x + tν(x))

t→−0

∂ν+ u(x) := lim ν(x) · ∇u(x + tν(x))

t→+0

are called the interior and exterior normal derivatives, respectively, of u.

The interior Dirichlet problem (ID)

Given f ∈ C(S), find u ∈ C 2 (Ω) ∩ C(Ω) such that ∆u = 0 in Ω and u = f on S.

The exterior Dirichlet problem (ED)

Given f ∈ C(S), find u ∈ C 2 (Ω′ ) ∩ C(Ω′ ) such that ∆u = 0 in Ω′ and at infinity and
u = f on S.

Definition. We say that u is harmonic at infinity if

  (
x o(|x|2−n ), n 6= 2
|x|2−n u =
|x|2 o (log |x|) , n = 2

as x → 0.

The interior Neumann problem (IN)

Given f ∈ C(S), find u ∈ C 2 (Ω) ∩ C(Ω) such that ∆u = 0 in Ω and ∂ν− u = f exists
on S.

The exterior Neumann problem (EN)

Given f ∈ C(S), find u ∈ C 2 (Ω′ ) ∩ C(Ω′ ) such that ∆u = 0 in Ω′ and at infinity and
∂ν+ u = f exists on S.

Theorem 1 (Uniqueness). 1. The solutions of (ID) and (ED) are unique.

2. The solutions of (IN) and (EN) are unique up to a constant on Ω and Ω′ , re-
spectively. When n > 2 this constant is zero on the unbounded component of
Ω′ .

82
Proof. If u solves (ID) with f = 0 then u ≡ 0 because this is just the uniqueness
theorem for harmonic functions (see Corollary 2 of Theorem 3 of Chapter  6).  If u
solves (ED) with f = 0 we may assume that {0} ∈ e = |x|2−n u |x|x2 solves
/ Ω′ . Then u
n o
(ID) with f = 0 for bounded domain Ωe = x : x2 ∈ Ω′ . Hence u e ≡ 0 so that u ≡ 0
|x|
and part 1) is proved.
 
Exercise 45. Prove that if u is harmonic then u e = |x| u |x|x2 , x 6= 0, is also
2−n

harmonic.
Concerning part 2), by Green’s identity we have
Z Z Z
2
|∇u| dx = − u∆udx + u∂ν− udσ(x).
Ω Ω S

Thus ∇u = 0 in Ω so that u is constant in Ω.

For (EN) let r > 0 be large enough so that Ω ⊂ Br (0). Again by Green’s identity
we have
Z Z Z Z
2
|∇u| dx = − u∆udx + u∂r udσ(x) − u∂ν+ udσ(x)
Br (0)\Ω Br (0)\Ω ∂Br (0) S
Z
= u∂r udσ(x),
∂Br (0)

d
where ∂r u ≡ dr
u. Since for n > 2 and for large |x| we have



u(x) = O |x|2−n , ∂r u(x) = O |x|1−n
then
Z  Z
 
  2−n 1−n
u∂r udσ(x) ≤ cr r dσ(x) = cr3−2n rn−1 = cr2−n → 0

∂Br (0) ∂Br (0)

as r → ∞. Hence Z
|∇u|2 dx = 0.
Ω′
It implies that u is constant on Ω′ and u = 0 on the unbounded component of Ω′
because for large |x|,

u(x) = O |x|2−n , n > 2.
If n = 2 then ∂r u(x) = O (r−2 ) for function u(x) which is harmonic at infinity.
Exercise 46. Prove that if u is harmonic at infinity then u is bounded and ∂r u(x) =
O (r−2 ) as r → ∞ if n = 2 and ∂r u(x) = O(|x|1−n ), r → ∞, if n > 2.
Due to Exercise 46 we obtain
Z 
 
 u∂ udσ(x)  ≤ cr−2 r = cr−1 → 0, r → ∞.
 r 
∂Br (0)

Hence ∇u = 0 in Ω′ and u is constant in (each component of) Ω′ .

83
 We now turn to the problem of finding the solutions (existence problems). Let us
try to solve (ID) by setting
Z
e(x) :=
u f (y)∂νy K(x − y)dσ(y), (8.1)
S

where K is the (known) fundamental solution for ∆.

Remark. Note that (8.1) involves only the known fundamental solution and not the
Green’s function (which is difficult to find in general) as in the Poisson integral
Z
w(x) = f (y)∂νy G(x, y)dσ(y). (P)
S

We know that u e(x) is harmonic in Ω, because K(x−y) is harmonic for x ∈ Ω, y ∈ S.

e will not have the right boundary
It remains to verify the boundary conditions. Clearly u
values but in a sense it is not far from right. We shall prove (very soon) that on S,
f
e=
u + T f,
2
where T is a compact operator on L2 (S). Thus, what we really want is to take
Z
u(x) = ϕ(y)∂νy K(x − y)dσ(y), x ∈ / S, (8.2)
S

where ϕ is the solution of

1
ϕ + T ϕ = f.
2
Similarly, we shall try to solve (IN) (and (EN)) in the form
Z
u(x) = ϕ(y)K(x − y)dσ(y), x ∈ / S. (8.3)
S

Definition. The functions u(x) from (8.2) and (8.3) are called the double and single
layer potentials with moment (density) ϕ, respectively.

Definition. Let I(x, y) be continuous on S × S, x 6= y. We call I a continuous kernel

of order α, 0 ≤ α < n − 1, n ≥ 2, if

|I(x, y)| ≤ c|x − y|−α , 0 < α < n − 1,

and
|I(x, y)| ≤ c1 + c2 |log |x − y|| , α = 0,
where c > 0 and c1 , c2 ≥ 0.

Remark. Note that a continuous kernel of order 0 is also a continuous kernel of order
α, 0 < α < n − 1.

84
 We denote by Ib the integral operator
Z
b
If (x) = I(x, y)f (y)dσ(y), x∈S
S
with kernel I.
Lemma 1. If I is a continuous kernel of order α, 0 ≤ α < n − 1, then
1. Ib is bounded on Lp (S), 1 ≤ p ≤ ∞.

2. Ib is compact on L2 (S).
Proof. It is enough to consider 0 < α < n − 1. Let us assume that f ∈ L1 (S). Then
Z Z
b
If 1 ≤ |I(x, y)||f (y)|dσ(y)dσ(x)
L (S) S S
Z Z
≤ c |f (y)|dσ(y) |x − y|−α dσ(x)
S S
Z d
≤ c kf kL1 (S) rn−2−α dr = c′ kf kL1 (S) ,
0

where d = diam S = supx,y∈S |x − y|.

If f ∈ L∞ (S) then
Z d
b
If ∞ ≤ c kf kL∞ (S) rn−2−α dr = c′ kf kL∞ (S) .
L (S) 0

For 1 < p < ∞ part 1) follows now by interpolation.

For part 2), let ε > 0 and set
(
I(x, y), |x − y| > ε
Iε (x, y) =
0, |x − y| ≤ ε.

Since Iε is bounded on S × S then Ibε is a Hilbert-Schmidt operator in L2 (S) so that

Ibε is compact for each ε > 0.
Exercise 47. Prove that a Hilbert-Schmidt operator i.e. an integral operator whose
kernel I(x, y) satisfies Z Z
|I(x, y)|2 dxdy < ∞
S S
2
is compact in L (S).
On the other hand, due to estimates for convolution,
Z Z 2 !1/2

b
If − Ibε f 2 ≤ c |f (y)||x − y|−α dσ(y) dσ(x)
L (S) |x−y|<ε
Z ε
≤ c kf kL2 (S) rn−2−α dr → 0, ε → 0.
0

Thus, Ib as limit of Ibε , is also compact in L2 (S).

85
Lemma 2. 1. If I is a continuous kernel of order α, 0 ≤ α < n − 1, then Ib trans-
forms bounded functions into continuous functions.
2. If Ib is as in part 1 then u + Iu
b ∈ C(S) for u ∈ L2 (S) implies u ∈ C(S).

Proof. Let |x − y| < δ. Then

Z
b b
|If (x) − If (y)| ≤ |I(x, z) − I(y, z)||f (z)|dσ(z)
S
Z
≤ (|I(x, z)| + |I(y, z)|) |f (z)|dσ(z)
|x−z|<2δ
Z
+ |I(x, z) − I(y, z)||f (z)|dσ(z)
S\{|x−z|<2δ}
Z


≤ c kf k∞ |x − z|−α + |y − z|−α dσ(z)
|x−z|<2δ
Z
+ |I(x, z) − I(y, z)||f (z)|dσ(z) := I1 + I2 .
S\{|x−z|<2δ}

Since |z − y| ≤ |x − z| + |x − y| then
Z 3δ
I1 ≤ c kf k∞ rn−2−α dr → 0, δ → 0.
0

On the other hand for |x − y| < δ and |x − z| ≥ 2δ we have that

|y − z| ≥ |x − z| − |x − y| > 2δ − δ = δ.

So the continuity of I outside of the diagonal implies that

I(x, z) − I(y, z) → 0, x → y,
uniformly in z ∈ S\ {|x − z| < 2δ}. Hence, I1 and I2 will be small if y is sufficiently
close to x. This proves part 1.
For part 2, let ε > 0 and let ϕ ∈ C(S × S) be such that 0 ≤ ϕ ≤ 1 and
(
1, |x − y| < ε/2
ϕ(x, y) =
0, |x − y| ≥ ε.

b = ϕIu
Write Iu c + (1\ − ϕ)Iu := Ib0 u + Ib1 u. By the Cauchy-Schwarz-Buniakowsky
inequality we have
Z 1/2
|Ib1 u(x) − Ib1 u(y)| ≤ kuk2 2
|I1 (x, z) − I1 (y, z)| dσ(z) → 0, y → x,
S

since I1 is continuous (see the definition of ϕ). Now if we set

b − Ib1 u ≡ u + Ib0 u
g := u + Iu

86
then g is continuous for u ∈ L2 (S) by the conditions of this lemma. Since the operator
norm of Ib0 can be made on L2 (S) and L∞ (S) less than 1 (we can do it due to the
choice of ε > 0 small enough), then
 −1
u = I + Ib0 g,

where I is the identity operator. Since g is continuous and the operator norm is less
than 1, then
X∞  j
u= −Ib0 g.
j=0

This series converges uniformly and therefore u is continuous.

Let us consider now the double layer potential (8.2) with moment ϕ,
Z
u(x) = ϕ(y)∂νy K(x − y)dσ(y), x ∈ Rn \S.
S

First of all
(x − y, ν(y))
∂νy K(x − y) = − . (8.4)
ωn |x − y|n
Exercise 48. Prove that (8.4) holds for any n ≥ 1.
It is clear also that (8.4) defines a harmonic function in x ∈ Rn \S, y ∈ S. Moreover,
it is O (|x|1−n ) as x → ∞ (y ∈ S) so that u is also harmonic at infinity.
Exercise 49. Prove that (8.4) defines a harmonic function at infinity.
Lemma 3. There exists c > 0 such that

|(x − y, ν(y))| ≤ c|x − y|2 , x, y ∈ S.

Proof. It is quite trivial to obtain

|(x − y, ν(y))| ≤ |x − y||ν(y)| = |x − y|.

But the latter inequality allows us to assume that |x − y| ≤ 1. Given y ∈ S, by

a translation and rotation of coordinates we may assume that y = 0 and ν(y) =
(0, 0, . . . , 0, 1). Hence (x − y, ν(y)) transforms to xn , and near y, S is the graph of the
equation xn = ψ(x1 , . . . , xn−1 ), where ψ ∈ C 2 (Rn−1 ), ψ(0) = 0 and ∇ψ(0) = 0. Then
by Taylor’s expansion

|(x − y, ν(y))| = |xn | ≤ c|(x1 , . . . , xn−1 )|2 ≤ c|x|2 = c|x − y|2 .

We denote ∂νy K(x − y) by I(x, y).

87
Lemma 4. I is a continuous kernel of order n − 2, n ≥ 2.

Proof. If x, y ∈ S then I(x, y) is continuous, see (8.4), for x 6= y. It follows from

Lemma 3 that
c|x − y|2
|I(x, y)| ≤ = c′ |x − y|2−n .
ωn |x − y|n

Lemma 5. 
Z 
1, x∈Ω
I(x, y)dσ(y) = 0, x ∈ Ω′ (8.5)
S 
1
2
, x ∈ S.

Proof. If x ∈ Ω′ then K(x − y) is harmonic in x ∈

/ S, y ∈ S and it is also harmonic in
′
y ∈ Ω, x ∈ Ω . Hence (see Exercise 23)
Z
∂νy K(x − y)dσ(y) = 0
S

or Z
I(x, y)dσ(y) = 0, x ∈ Ω′ .
S

If x ∈ Ω, let δ > 0 be such that Bδ (x) ⊂ Ω. Then K(x−y) is harmonic in y in Ω\Bδ (x)
and therefore by Green’s identity
Z
0 = (1 · ∆y K(x − y) − K(x − y)∆1) dy
Ω\Bδ (x)
Z Z
= ∂νy K(x − y)dσ(y) − ∂νy K(x − y)dσ(y)
S |x−y|=δ
Z 1−n Z
δ
= I(x, y)dσ(y) − dσ(y)
S ωn |x−y|=δ
Z
= I(x, y)dσ(y) − 1
S

or Z
I(x, y)dσ(y) = 1.
S

Now suppose x ∈ S and Sδ = S\ (S ∩ Bδ (x)). In this case

Z Z
I(x, y)dσ(y) = lim I(x, y)dσ(y). (8.6)
S δ→0 Sδ

88
 ∂Bδ+
S
δ
Bδ+
Ω′ b Ω
x
Bδ−

Sδ Ωδ ∂Bδ−

If y ∈ Ωδ := Ω\Bδ (x) then for x ∈ S we have that x 6= y. It implies that

Z
0 = ∆y K(x − y)dy
Ωδ
Z Z
= ∂νy K(x − y)dσ(y) − ∂νy K(x − y)dσ(y).
Sδ ∂Bδ−

That’s why, see (8.4),

Z Z
lim ∂νy K(x − y)dσ(y) = lim ∂νy K(x − y)dσ(y)
δ→0 Sδ δ→0 ∂Bδ−
Z
δ 1−n
= lim dσ(y)
δ→0 ωn ∂Bδ−

δ 1−n  n−1 ωn  1
= lim δ + o(δ n−1 ) = .
δ→0 ωn 2 2
It means that the limit in (8.6) exists and (8.5) is satisfied.
Lemma 6. There exists c > 0 such that
Z
|∂νy K(x − y)|dσ(y) ≤ c, x ∈ Rn .
S

Proof. It follows from Lemma 4 that

Z Z
c
|∂νy K(x − y)|dσ(y) ≤ |x − y|2−n dσ(y) ≤ c1 , x ∈ S.
S ωn S
Next, for x ∈
/ S define dist(x, S) = inf y∈S |x − y|.
There are two possibilities now: if dist(x, S) ≥ δ/2 then |x − y| ≥ δ/2 for all y ∈ S
and therefore Z Z
1−n
|∂νy K(x − y)|dσ(y) ≤ cδ dσ(y) = c′ , (8.7)
S S
′
where c does not depend on δ > 0 (because δ is fixed).
Suppose now that dist(x, S) < δ/2. If we choose δ > 0 small enough then there is
unique x0 ∈ S such that

x = x0 + tν(x0 ), t ∈ (−δ/2, δ/2).

89
Denote Bδ = {y ∈ S : |x0 − y| < δ}. We estimate the integral of |I(x, y)| over S\Bδ
and Bδ separately. If y ∈ S\Bδ then
|x − y| ≥ |x0 − y| − |x − x0 | > δ − δ/2 = δ/2
and
|I(x, y)| ≤ cδ 1−n
so that the integral over S\Bδ satisfies (8.7), where again c′ does not depend on δ.
To estimate the integral over Bδ we note that (see (8.4)),
|(x − y, ν(y))| |(x − x0 , ν(y)) + (x0 − y, ν(y))|
|I(x, y)| = n
=
ωn |x − y| ωn |x − y|n
|x − x0 | + c|x0 − y|2
≤ . (8.8)
ωn |x − y|n
The latter inequality follows from Lemma 3 since x0 , y ∈ S. Moreover, we have (due
to Lemma 3)
|x − y|2 = |x − x0 |2 + |x0 − y|2 + 2(x − x0 , x0 − y)
 
2 2 x − x0
= |x − x0 | + |x0 − y| + 2|x − x0 | x0 − y,
|x − x0 |
2 2
≥ |x − x0 | + |x0 − y| − 2|x − x0 ||(x0 − y, ν(x0 ))|
≥ |x − x0 |2 + |x0 − y|2 − 2c|x − x0 ||x0 − y|2
≥ |x − x0 |2 + |x0 − y|2 − |x − x0 ||x0 − y|,
1
if we choose δ > 0 such that |x0 − y| ≤ 2c , where constant c > 0 is from Lemma 3.
1
Since |x − x0 ||x0 − y| ≤ 2 (|x − x0 | + |x0 − y|2 ) then finally we obtain
2

1

|x − y|2 ≥ |x − x0 |2 + |x0 − y|2
2
and (see (8.4) and (8.8))
|x − x0 | + |x0 − y|2 |x − x0 | c
|I(x, y)| ≤ c ≤c + .
|2
(|x − x0 + |x0 − y|2 )n/2 |2
(|x − x0 + |x0 − y|2 )n/2 |x0 − y|n−2
This implies
Z Z δ Z δ
′ |x − x0 | n−2 ′ rn−2
|I(x, y)|dσ(y) ≤ c n/2
r dr + c dr
Bδ (|x − x0 |2 + r2 )
0 0 rn−2
Z ∞
′ ′ arn−2
≤ cδ+c dr,
0 (a2 + r2 )n/2
where a := |x − x0 |. For the latter integral we have (t = r/a)
Z ∞ Z ∞
arn−2 tn−2
dr = dt < ∞.
0 (a2 + r2 )n/2 0 (1 + t2 )n/2

90
If we combine all estimates then we may conclude that there is c0 > 0 such that
Z
|∂νy K(x − y)|dσ(y) ≤ c0 , x ∈ Rn ,
S

and this constant does not depend on x.

Theorem 2. Suppose ϕ ∈ C(S) and u is defined by the double layer potential (8.2)
with moment ϕ. Then for any x ∈ S,
Z
ϕ(x)
lim u(x + tν(x)) = + I(x, y)ϕ(y)dσ(y)
t→−0 2
ZS
ϕ(x)
lim u(x + tν(x)) = − + I(x, y)ϕ(y)dσ(y)
t→+0 2 S

uniformly on S with respect to x.

Proof. If x ∈ S and t < 0, with |t| small enough, then xt := x + tν(x) ∈ Ω and
u(x + tν(x)) is well-defined by
Z Z
u(x + tν(x)) = ϕ(y)I(xt , y)dσ(y) = (ϕ(y) − ϕ(x))I(xt , y)dσ(y) + ϕ(x)
S S
Z Z
→ ϕ(x) + ϕ(y)I(x, y)dσ(y) − ϕ(x) I(x, y)dσ(y)
ZS S

= ϕ(x) + ϕ(y)I(x, y)dσ(y) − ϕ(x)/2, t → −0.

S

If t > 0, the arguments are the same except that

Z
I(xt , y)dσ(y) = 0.
S

The uniformity of convergence follows from the fact that S is compact and ϕ ∈ C(S).

Corollary. For x ∈ S,
ϕ(x) = u− (x) − u+ (x),
where u± = limt→±0 u(xt ).
We state without a proof that the normal derivative of the double layer potential
is continuous across the boundary in the sense of the following theorem.
Theorem 3. Suppose ϕ ∈ C(S) and u is defined by the double layer potential (8.2)
with moment ϕ. Then for any x ∈ S,

lim (ν(x) · ∇u(x + tν(x)) − ν(x) · ∇u(x − tν(x))) = 0

t→+0

uniformly on S with respect to x.

91
 Let us now consider the single layer potential
Z
u(x) = ϕ(y)K(x − y)dσ(y)
S

with moment ϕ ∈ C(S).

Lemma 7. The single layer potential u is continuous on Rn .

Proof. Since u is harmonic in x ∈/ S we need only to show continuity for x ∈ S. Given

x0 ∈ S and δ > 0, let Bδ = {y ∈ S : |x0 − y| < δ}. Then
Z
|u(x) − u(x0 )| ≤ (|K(x − y)| + |K(x0 − y)|) |ϕ(y)|dσ(y)
Bδ
Z
+ |K(x − y) − K(x0 − y)||ϕ(y)|dσ(y)
S\Bδ
1
≤ cδ(or δ log for n = 2)
Z δ
+ kϕk∞ |K(x − y) − K(x0 − y)|dσ(y) → 0
S\Bδ

as x → x0 and δ → 0.

Exercise 50. Prove that

Z (
δ, n>2
(|K(x − y)| + |K(x0 − y)|) |ϕ(y)|dσ(y) ≤ c kϕk∞ 1
Bδ δ log δ , n = 2.

Definition. Let us set

(x − y, ν(x))
I ∗ (x, y) := ∂νx K(x − y) ≡ .
ωn |x − y|n

Theorem 4. Suppose ϕ ∈ C(S) and u is defined on Rn by the single layer potential

(8.3) with moment ϕ. Then for x ∈ S,
Z
ϕ(x)
lim ∂ν u(xt ) = − + I ∗ (x, y)ϕ(y)dσ(y),
t→−0 2
ZS
ϕ(x)
lim ∂ν u(xt ) = + I ∗ (x, y)ϕ(y)dσ(y).
t→+0 2 S

Proof. Consider the double layer potential on Rn \S with moment ϕ

Z
v(x) = ϕ(y)∂νy K(x − y)dσ(y)
S

92
and define the function f on the tubular neighborhood V of S by
(
v(x) + ∂ν u(x), x ∈ V \S
f (x) = (8.9)
b
Iϕ(x) + Ib∗ ϕ(x), x ∈ S,

where u is defined by (8.3).

Here the tubular neighborhood of S is defined as

V = {x + tν(x) : x ∈ S, |t| < δ} .

We claim that f is continuous on V . It clearly is (see (8.9)) continuous on V \S and

S, so it suffices to show that if x0 ∈ S and x = x0 + tν(x0 ) then f (x) − f (x0 ) → 0 as
t → ±0. We have
b 0 ) − Ib∗ ϕ(x0 )
f (x) − f (x0 ) = v(x) + ∂ν u(x) − Iϕ(x
Z Z
= I(x, y)ϕ(y)dσ(y) + ϕ(y)∂νx K(x − y)dσ(y)
ZS ZS
− I(x0 , y)ϕ(y)dσ(y) − I ∗ (x0 , y)ϕ(y)dσ(y)
ZS S

= (I(x, y) + I ∗ (x, y) − I(x0 , y) − I ∗ (x0 , y)) ϕ(y)dσ(y).

S

Write this expression as an integral over Bδ = {y ∈ S : |x0 − y| < δ} plus an integral

over S\Bδ . The integral over S\Bδ tends uniformly to 0 as x → x0 , because |y − x| ≥ δ
and |y − x0 | ≥ δ so that the functions I and I ∗ have no singularities in this case.
On the other hand, the integral over Bδ can be bounded by
Z
kϕk∞ (|I(x, y) + I ∗ (x, y)| + |I(x0 , y) + I ∗ (x0 , y)|) dσ(y).
Bδ

Since
(x − y, ν(y))
I(x, y) = −
ωn |x − y|n
and ν(x) = ν(x0 ) for x = x0 + tν(x0 ) ∈ V we have
(x − y, ν(x)) (x − y, ν(x0 ))
I ∗ (x, y) = I(y, x) = n
≡ . (8.10)
ωn |x − y| ωn |x − y|n
Hence
 
 (x − y, ν(x0 ) − ν(y))  |x − y||ν(x0 ) − ν(y)|
∗ 
|I(x, y) + I (x, y)| =  ≤
ωn |x − y|n  ωn |x − y|n
|x − y||x0 − y| |x0 − y|
≤ c n
≤ c′ = c′ |x0 − y|2−n ,
ωn |x − y| |x0 − y|n−1
because |x0 − y| ≤ |x0 − x| + |x − y| ≤ 2|x − y|. Here we have also used the fact that
|ν(x0 ) − ν(y)| ≤ c|x0 − y| since ν is C 1 .

93
 x
S

x0
y

This estimate allows us to obtain that the corresponding integral over Bδ can be
dominated by
Z Z δ
2−n ′
c |x0 − y| dσ(y) = c r2−n rn−2 dr = c′ δ.
|y−x0 |≤δ 0

Thus f := v + ∂ν u extends continuously across S. That’s why for x ∈ S,

b 1
Iϕ(x) b
+ Ib∗ ϕ(x) = v− (x) + ∂ν− u(x) = ϕ(x) + Iϕ(x) + ∂ν− u(x).
2
It follows that
ϕ(x) b∗
∂ν− u(x) = − + I ϕ(x).
2
By similar arguments we obtain

b 1
Iϕ(x) b
+ Ib∗ ϕ(x) = v+ (x) + ∂ν+ u(x) = − ϕ(x) + Iϕ(x) + ∂ν+ u(x)
2
and therefore
ϕ(x) b∗
∂ν+ u(x) = + I ϕ(x).
2
This finishes the proof.
Corollary.
ϕ(x) = ∂ν+ u(x) − ∂ν− u(x),
where u is defined by (8.3).
Lemma 8. If f ∈ C(S) and
ϕ b∗
+I ϕ=f
2
then Z Z
ϕdσ = f dσ.
S S

Proof. It follows from (8.10) and Lemma 5 that

Z Z Z Z
1
f (x)dσ(x) = ϕ(x)dσ(x) + ϕ(y)dσ(y) I ∗ (x, y)dσ(x)
S 2
ZS S
Z S
Z
1 1
= ϕ(x)dσ(x) + ϕ(y)dσ(y) = ϕ(y)dσ(y).
2 S 2 S S

94
Lemma 9. Let n = 2.
1. If ϕ ∈ C(S) then the single layer potential u with moment ϕ is harmonic at
infinity if and only if Z
ϕ(x)dσ(x) = 0.
S

2. Let ϕ ∈ C(S) with Z

ϕ(x)dσ(x) = 0
S

and u as in part 1. If u is constant on Ω then ϕ ≡ 0, and hence u ≡ 0.

Proof. Since n = 2 then
Z
1
u(x) = log |x − y|ϕ(y)dσ(y)
2π S
Z Z
1 1
= (log |x − y| − log |x|) ϕ(y)dσ(y) + log |x| ϕ(y)dσ(y).
2π S 2π S

But log |x − y| − log |x| → 0 as x → ∞ uniformly for y ∈ S and therefore, this term is
harmonic at infinity
R (we have a removable singularity). Hence u is harmonic at infinity
if and only if S ϕ(x)dσ(x) = 0 and in this case u(x) vanishes at infinity. This proves
part 1.
In part 2, u is harmonic at infinity. If u is constant on Ω then it solves (ED) with
f ≡ constant on S. But a solution of such problem must be constant and vanish at
infinity. Therefore this constant is zero. Thus ϕ ≡ 0 and, hence, u ≡ 0.
We assume (for simplicity and without loss of generality) that Ω and Ω′ are simply-
connected, that is, ∂Ω has only one C 2 component. For f ∈ C(S) consider the integral
equations
1 b = f,
± ϕ + Iϕ (1± )
2
1
± ϕ + Ib∗ ϕ = f, (1∗± )
2
where I(x, y) = ∂νy K(x − y) and I ∗ (x, y) = I(y, x). Theorems 2 and 4 show that the
double layer potential u with moment ϕ solves (ID) (respectively (ED)) if ϕ satisfies
(1+ ) (respectively (1− )) and the single layer potential u with moment ϕ solves (IN)
(respectively (EN)) if ϕ satisfies (1∗− ) (respectively (1∗+ )). For n = 2 we need the extra
necessary condition for (EN) given by Lemma 9.
We proceed to study the solvability of (1± ) and (1∗± ). Let us introduce
 
b 1
V± = ϕ : Iϕ = ± ϕ
2
  (8.11)
1
W± = ϕ : Ib∗ ϕ = ± ϕ ,
2
where ϕ is allowed to range over either L2 (S) or C(S).

95
Fredholm’s Theorem
Let A be a compact operator on a Hilbert space H. For each λ ∈ C, let
Vλ = {x ∈ H : Ax = λx}
and
Wλ = {x ∈ H : A∗ x = λx} .
Then
1. The set {λ ∈ C : Vλ =6 {0}} is finite or countable with only one possible accumu-
lation point at {0}. Moreover, dimVλ < ∞ for λ 6= 0.
2. dimVλ = dimWλ if λ 6= 0.
3. R(A − λI) and R(A∗ − λI) are closed if λ 6= 0.
Here and throughout the use of symbol I for the identity operator is to be distinguished
from a function I = I(x, y) by the context in which it appears.
Corollary 1. Suppose λ 6= 0. Then
1. (A − λI)x = y has a solution if and only if y⊥Wλ .
2. A − λI is surjective (onto) if and only if it is injective (invertible).
In other words, either (A−λI)ϕ = 0 and (A∗ −λI)ψ = 0 have only the trivial solutions
ϕ = ψ = 0 for λ 6= 0 or they have the same number of linearly independent solutions
ϕ1 , . . . , ϕm , ψ1 , . . . , ψm , respectively. In the first case (A−λI)ϕ = g and (A∗ −λI)ψ = f
have unique solutions (A − λI and A∗ − λI are invertible) for every g, f ∈ H. In the
second case (A − λI)ϕ = g and (A∗ − λI)ψ = f have the solutions if and only if ϕj ⊥g
and ψj ⊥f for every j = 1, 2, . . . , m.
Proof. It is known and not so difficult to show that

⊥
R(A − λI) = Ker A∗ − λI , (8.12)
where M ⊥ denotes the orthogonal complement of M ⊂ H defined by
M ⊥ = {y ∈ H : (y, x)H = 0, x ∈ M }.
Exercise 51. Prove (8.12)
But by part 3 of Fredholm’s theorem (λ 6= 0) we know that R(A − λI) = R(A−λI)
and, therefore R(A − λI) = Ker (A∗ − λI)⊥ . It is equivalent to the fact that
y ∈ R(A − λI) ⇔ y⊥Ker (A∗ − λI)
or
(A − λI)x = y, x ∈ H ⇔ y⊥Wλ .
For the second part, A − λI is surjective if and only if R(A − λI) = H, that is,
Ker (A∗ − λI) = 0. But this is equivalent to A∗ − λI being invertible or A − λI being
invertible (injective).

96
Corollary 2.

L2 (S) = V+⊥ ⊕ W+ = V−⊥ ⊕ W−

= V+ ⊕ W+⊥ = V− ⊕ W−⊥

where V± and W± are defined by (8.11) and the direct sums are not necessarily orthog-
onal.
Proof. By Lemma 5 we know that
Z
1
I(x, y)dσ(y) = , x ∈ S.
S 2
It can be interpreted as follows: ϕ(x) ≡ 1 belongs to V+ . Hence dimV+ ≥ 1. But
by part 2 of Fredholm’s theorem dimV+ = dimW+ . Since the single layer potential
uniquely solves the (EN) and (IN) then dimW+ ≤ 1. Hence dimV+ = dimW+ = 1.
Therefore, in order to prove the equality

L2 (S) = V+⊥ ⊕ W+

it is enough to show that V+⊥ ∩ W+ = {0} (because V+⊥ is a closed subspace of codi-
mension 1).
Suppose ϕ ∈ V+⊥ ∩ W+ . Then Ib∗ ϕ = 12 ϕ (ϕ ∈ W+ ) and there is ψ ∈ L2 (S) such
that ϕ = − ψ2 + Ib∗ ψ (ϕ ∈ V+⊥ ), see Corollary 1 for λ = 1/2 and A = Ib∗ . Next, since
Ib∗ ϕ − 21 ϕ ≡ 0 and ϕ ∈ L2 (S) then part 2 of Lemma 2 gives that ϕ is continuous and
hence ψ is continuous too.
Let u and v be the single layer potentials with moments ϕ and ψ, respectively.
Then by Theorem 4
ϕ b∗
∂ ν− u = − +I ϕ=0
2
ψ ϕ
∂ν− v = − + Ib∗ ψ = ϕ = + Ib∗ ϕ = ∂ν+ u.
2 2
It follows that
Z Z Z
0 = (u∆v − v∆u)dx = (u∂ν− v − v∂ν− u)dσ(x) = u∂ν+ udσ(x).
Ω S S

But on the other hand

Z Z
u∂ν+ udσ(x) = − (u∆u + |∇u|2 )dx.
S Ω′

Hence Z
|∇u|2 dx = 0
Ω′
′
and so u is constant in Ω . This gives finally ϕ = ∂ν+ u = 0. The other equalities can
be proved in a similar manner.

97
Remark. Since we know that
 ⊥  
b 1 b 1
W∓⊥ ∗
= Ker I ± I =R I± I
2 2

(see (8.12) and part 3 of Fredholm’s theorem) we can rewrite Corollary 2 as

   
1 1
L (S) = V+ ⊕ R Ib − I = V− ⊕ R Ib + I .
2
2 2

Theorem 5. [Main theorem]

1. (ID) has a unique solution for any f ∈ C(S)

2. (ED) has a unique solution for any f ∈ C(S)

R
3. (IN) has a solution for any f ∈ C(S) if and only if S
f dσ = 0. The solution is
unique up to a constant.
R
4. (EN) has a solution for any f ∈ C(S) if and only if S
f dσ = 0. The solution is
unique up to a constant.

Proof. We have already proved uniqueness (see Theorem 1) and the necessity of the
conditions on f (see Exercise 23). So all that remains is to establish existence. It turns
out that in each case this question is reduced to the question of the solvability of an
integral equation.
Note first that Z
f dσ = 0
S
if and only if
(f, 1)L2 (S) = 0
or f ∈ V+⊥ since 1 ∈ V+ and dimV+ = 1. But f ∈ V+⊥ is necessary and sufficient
condition (see Corollary 1) to solve the integral equation
ϕ b∗
− + I ϕ = f.
2
If ϕ is a solution of this equation, ϕ is continuous (see part 2 of Lemma 2). Hence, by
Theorem 4 the single layer potential Rwith moment ϕ solves (IN)
Similarly, for (EN), we have that S f dσ = 0 if and only if f ∈ V−⊥ . In this case we
can solve the equation
ϕ b∗
+I ϕ=f
2
and then solve (EN) by the single layer potential with moment ϕ, see again Theorem
4.

98
 Consider now (ID). By Corollary 2 of Fredholm’s Theorem and Remark after it we
can write for f ∈ C(S) ⊂ L2 (S),
ϕ 
f= b
+ Iϕ + ψ, (8.13)
2
where ψ ∈ V− ⊂ C(S) and ϕ is continuous since f − ψ is continuous (part 2 of Lemma
2).
b = 0. Let us prove that this condition implies that
Since ψ ∈ V− then 12 ψ + Iψ
ψ = 0. Consider the double layer potential
Z
v(x) = ψ(y)I(x, y)dσ(y), x ∈/ S.
S

b = 0 (see Theorem 2). Hence v ∈ C(Ω)

It is harmonic outside of S and v− = 12 ψ + Iψ
and the uniqueness result for the interior Dirichlet problem ensures that v = 0 in Ω.
Therefore ∂ν− v = 0 and hence ∂ν+ v = 0 follows from the jump relation ∂ν+ v − ∂ν− v = 0
(Theorem 3). It means that v is constant in Rn \Ω. If n > 2 the uniqueness theorem
for the exterior Neumann problem implies that v = 0 in Rn \Ω. If n = 2 the argument
is slightly different. We know that ∆v = 0 in Rn \Ω and ∂ν+ v = 0. By part 4 the
unique solution of this problem is
Z
v(x) = ψ1 (y)K(x − y)dσ(y),
S
R
with S ψ1 (y)dσ(y) = 0 since otherwise v is not harmonic at infinity and we do not
even have uniqueness. Thus Lemma 9 implies that v = 0.
So v ≡ 0 which means that ψ = v− − v+ = 0. We have also proved above that
the operator 21 I + Ib is injective. Hence it is surjective too and the integral equation
(8.13) is solvable for any f ∈ C(S). By Theorem 2, the double layer potential v with
moment ϕ now solves (ID).
Exercise 52. Prove part 2 of Theorem 5.

99
9 The Heat Operator
We turn our attention now to the heat operator

L = ∂t − ∆x , (x, t) ∈ Rn × R.

The heat operator is a prototype of parabolic operators. These are operators of the
form X
∂t + aα (x, t)∂xα ,
|α|≤2m

where the sum satisfies the strong ellipticity condition

X
(−1)m aα (x, t)ξ α ≥ ν|ξ|2m ,
|α|=2m

for all (x, t) ∈ Rn × R and ξ ∈ Rn \ {0} with ν > 0 constant.

We begin by considering the initial value problem
(
∂t u − ∆u = 0, in Rn × (0, ∞)
u(x, 0) = f (x).

This problem is a reasonable problem both physically and mathematically.

Assuming for the moment that f ∈ S, the Schwartz space, and taking the Fourier
transform with respect to x only, we obtain
(
b(ξ, t) + |ξ|2 u
∂t u b(ξ, t) = 0
(9.1)
b
b(ξ, 0) = f (ξ).
u

If we solve the ordinary differential equation (9.1) we obtain

b(ξ, t) = e−|ξ| t fb(ξ).

2
u

Thus (at least formally)

   
u(x, t) = F −1 e−|ξ| t fb(ξ) = (2π)−n/2 f ∗ F −1 e−|ξ| t (x, t) = f ∗ Kt (x),
2 2

where  
2 |x|2
Kt (x) = (2π)−n/2 F −1 e−|ξ| t ≡ (4πt)−n/2 e− 4t , t > 0 (9.2)

is called the Gaussian kernel . We define Kt (x) ≡ 0 for t ≤ 0.

Exercise 53. Prove (9.2).

100
 Let us first prove that Z
Kt (x)dx = 1.
Rn
Indeed, using polar coordinates,
Z Z Z ∞ Z
|x|2 2
−n/2 − −n/2 n−1 − r4t
Kt (x)dx = (4πt) e 4t dx = (4πt) r e dr dθ
Rn Rn 0 |θ|=1
Z ∞
r2
−n/2
= ωn (4πt) rn−1 e− 4t dr
Z0 ∞
n−1 1 √ ds
= ωn (4πt)−n/2 (4st) 2 e−s 4t √
2 s
Z ∞0
ωn −n/2
= π sn/2−1 e−s ds
2 0
ωn −n/2 1 2π n/2 −n/2
= π Γ(n/2) = π Γ(n/2) = 1.
2 2 Γ(n/2)
Theorem 1. Suppose that f ∈ L∞ (Rn ) is uniformly continuous. Then u(x, t) :=
(f ∗ Kt )(x) satisfies ∂t u − ∆u = 0 and
ku(·, t) − f (·)kL∞ (Rn ) → 0
as t → +0.
Proof. For fixed t > 0
 
−n/2 −
|x−y|2 |x − y|2 n
∆x Kt (x − y) = (4πt) e 4t
2
−
4t 2t
and for fixed |x − y| 6= 0
 
−n/2 −
|x−y|2 |x − y|2 n
∂t Kt (x − y) = (4πt) e 4t − .
4t2 2t
Therefore (∂t − ∆x )Kt (x − y) = 0.
But we can differentiate (with respect to x and t) under the integral sign since this
integral will be absolutely convergent for any t > 0. That’s why we may conclude that
∂t u(x, t) − ∆x u(x, t) = 0.
It remains only to verify the initial condition. We have
Z
u(x, t) − f (x) = (f ∗ Kt )(x) − f (x) = f (y)Kt (x − y)dy − f (x)
Rn
Z Z
= f (x − z)Kt (z)dz − f (x)Kt (z)dz
Rn Rn
Z
= (f (x − z) − f (x))Kt (z)dz
ZR
n

√
= (f (x − η t) − f (x))K1 (η)dη.
Rn

101
The assumptions on f imply that
√ Z
|u(x, t) − f (x)| ≤ sup |f (x − η t) − f (x)| K1 (η)dη
x∈Rn ,|η|<R Rn
Z
+ 2 kf kL∞ (Rn ) K1 (η)dη < ε/2 + ε/2
|η|≥R

for small t and for R large enough. So we can see that u(x, t) is continuous (even
uniformly continuous and bounded) for (x, t) ∈ Rn × [0, ∞) and u(x, 0) = f (x).

Corollary 1. u(x, t) ∈ C ∞ (Rn × R+ ).

Proof. We can differentiate under the integral defining u as often as we please, because
the exponential function decreases at infinity faster than any polynomial. Thus, the
heat equation takes arbitrary initial data (bounded and uniformly continuous) and
smooths them out.

Corollary 2. Suppose f ∈ Lp (Rn ), 1 ≤ p < ∞. Then u(x, t) := (f ∗ Kt )(x) satisfies

∂t u − ∆u = 0 and
ku(·, t) − f (·)kLp (Rn ) → 0
as t → +0. And again u(x, t) ∈ C ∞ (Rn × R+ ).

Theorem 2 (Uniqueness). Suppose u(x, t) ∈ C 2 (Rn × R+ ) ∩ C(Rn × R+ ) and satisfies

∂t u − ∆u = 0 for t > 0 and u(x, 0) = 0. If for every ε > 0 there exists cε > 0 such that
2 2
|u(x, t)| ≤ cε eε|x| , |∇x u(x, t)| ≤ cε eε|x| (9.3)

then u ≡ 0.

Proof. For two smooth functions ϕ and ψ it is true that

n
X
ϕ(∂t ψ − ∆ψ) + ψ(∂t ϕ + ∆ϕ) = ∂j (ψ∂j ϕ − ϕ∂j ψ) + ∂t (ϕψ) = ∇x,t · F~ ,
j=1

where F~ = (ψ∂1 ϕ − ϕ∂1 ψ, . . . , ψ∂n ϕ − ϕ∂n ψ, ϕψ). Given x0 ∈ Rn and t0 > 0 let us
take
ψ(x, t) = u(x, t), ϕ(x, t) = Kt0 −t (x − x0 ).
Then
∂t ψ − ∆ψ = 0, t > 0
∂t ϕ + ∆ϕ = 0, t < t0 .

If we apply the divergence theorem in the region

Ω = {(x, t) ∈ Rn × R+ : |x| < r, 0 < a < t < b < t0 }

102
we obtain
Z Z Z
0 = F~ · νdσ = u(x, b)Kt0 −b (x − x0 )dx − u(x, a)Kt0 −a (x − x0 )dx
∂Ω |x|≤r |x|≤r
Z b Z n
X xj
+ dt (u(x, t)∂j Kt0 −t (x − x0 ) − Kt0 −t (x − x0 )∂j u(x, t)) dσ(x).
a |x|=r j=1 r

Letting r → ∞ the last sum vanishes by assumptions (9.3). That’s why we have
Z Z
0= u(x, a)Kt0 −a (x − x0 )dx − u(x, b)Kt0 −b (x − x0 )dx.
Rn Rn

As we know from the proof of Theorem 1 for b → t0 − 0 the second term tends to
u(x0 , t0 ) and for a → +0 the first term tends to
Z
u(x, 0)Kt0 (x − x0 )dx = 0
Rn

because u(x, 0) = 0. Hence we have finally that u(x0 , t0 ) = 0 for all x0 ∈ Rn , t0 > 0.

Theorem 3. The kernel Kt (x) is a fundamental solution for the heat operator.

Proof. Given ε > 0, set (

Kt (x), t ≥ ε
Kε (x, t) =
0, t < ε.
Clearly Kε (x, t) → Kt (x) as ε → 0 in the sense of distributions. Even more is true,
namely, Kε (x, t) → Kt (x) pointwise as ε → 0 and
Z Z Z
|Kε (x, t)| dx = Kε (x, t)dx ≤ Kt (x)dx = 1.
Rn Rn Rn

That’s why we can apply the dominated convergence theorem and obtain
Z Z
lim Kε (x, t)dx = Kt (x)dx.
e→+0 Rn Rn

So it remains to show that, as ε → 0,

∂t Kε (x, t) − ∆x Kε (x, t) → δ(x, t),

or
h∂t Kε − ∆x Kε , ϕi → ϕ(0), ϕ ∈ C0∞ (Rn+1 ).

103
Using integration by parts we obtain
Z ∞ Z
h∂t Kε − ∆x Kε , ϕi = hKε , −∂t ϕ − ∆ϕi = dt Kt (x)(−∂t − ∆)ϕ(x, t)dx
ε Rn
Z Z ∞
= − dx Kt (x)∂t ϕ(x, t)dt
Rn ε
Z ∞ Z
− dt Kt (x)∆x ϕ(x, t)dx
ε Rn
Z Z ∞ Z
= Kε (x)ϕ(x, ε)dx + dt ∂t Kt (x)ϕ(x, t)dx
Rn ε Rn
Z ∞ Z
− dt ∆x Kt (x)ϕ(x, t)dx
ε Rn
Z Z ∞ Z
= Kε (x)ϕ(x, ε)dx + dt (∂t − ∆)Kt (x)ϕ(x, t)dx
Rn ε Rn
Z
= Kε (x)ϕ(x, ε)dx → ϕ(0, 0), ε → 0
Rn

as we know from the proof of Theorem 1.

Theorem 4. If f ∈ L1 (Rn+1 ), then

Z t Z
u(x, t) := (f ∗ Kt )(x) ≡ ds Kt−s (x − y)f (y, s)dy
−∞ Rn

is well-defined almost everywhere and is a distributional solution of ∂t u − ∆u = f .

Exercise 54. Prove Theorem 4.

Let us now consider the heat operator in a bounded domain Ω ⊂ Rn over a time
interval t ∈ [0, T ], 0 < T ≤ ∞. In this case it is necessary to specify the initial
temperature u(x, 0), x ∈ Ω, and also to prescribe a boundary condition on ∂Ω × [0, T ].
∂Ω
t=T

Ω t=0

The first basic result concerning such problems is the maximum principle.

Theorem 5. Let Ω be a bounded domain in Rn and 0 < T < ∞. Suppose u is a

real-valued continuous function on Ω × [0, T ], that satisfies ∂t u − ∆u = 0 in Ω × (0, T ).
Then u assumes its maximum and minimum either on Ω × {0} or on ∂Ω × [0, T ].

104
Proof. Given ε > 0, set v(x, t) := u(x, t) + ε|x|2 . Then ∂t v − ∆v = −2nε. Suppose 0 <
T ′ < T . If maximum of v in Ω×[0, T ′ ] occurs at an interior point of Ω×(0, T ′ ) then the
first derivatives ∇x,t v vanish there and the second derivative ∂j2 v for any j = 1, 2, . . . , n
is nonpositive (consider v(x, t) as a function of one variable xj , j = 1, 2, . . . , n). In
particular, ∂t v = 0 and ∆v ≤ 0, which contradicts with ∂t v − ∆v = −2nε < 0 and
∆v = 2nε > 0.
Likewise, if the maximum occurs in Ω×{T ′ }, then ∂t v(x, T ′ ) ≥ 0 and ∆v(x, T ′ ) ≤ 0
which contradicts with ∂t v − ∆v < 0. Therefore,

max u ≤ max v ≤ max u + ε max |x|2 .

Ω×[0,T ′ ] Ω×[0,T ′ ] (Ω×{0})∪(∂Ω×[0,T ′ ]) Ω

It follows that for ε → 0 and T ′ → T ,

max u ≤ max u.
Ω×[0,T ] (Ω×{0})∪(∂Ω×[0,T ])

Replacing u by −u we can obtain the same result for the minimum.

Corollary (Uniqueness). There is at most one continuous function u(x, t) in Ω ×

[0, T ], 0 < T < ∞, which agrees with a given continuous function f (x) in Ω × {0}, with
g(x, t) on ∂Ω × [0, T ] and satisfies ∂t u − ∆u = 0.

Let us look now more closely at the following problem:



∂t u − ∆u = 0, in Ω × (0, ∞)
u(x, 0) = f (x), in Ω (9.4)


u(x, t) = 0, on ∂Ω × (0, ∞).

This problem can be solved by the method of separation of variables. We begin by

looking for solution of the form

u(x, t) = F (x)G(t).

Then
∂t u − ∆u = F G′ − G∆x F = 0
if and only if
G′ ∆F
= := −λ2
G F
or
G′ + λ2 G = 0, ∆F + λ2 F = 0,
for some constant λ. The first equation has the general solution
2
G(t) = ce−λ t ,

105
where c is an arbitrary constant. Without loss of generality we assume that c = 1. It
follows from (9.4) that (
∆F = −λ2 F, in Ω
(9.5)
F = 0, on ∂Ω,
because u(x, t) = F (x)G(t) and G(0) = 1.
It remains to solve (9.5) which is an eigenvalue (spectral) problem for the Laplacian
with Dirichlet boundary condition. It is known that the problem (9.5) has infinitely
∞ 2 ∞ 2
many solutions {Fj (x)}j=1 with corresponding λj j=1 . The numbers −λj are called
eigenvalues and Fj (x) are called eigenfunctions of the Laplacian. It is also known that
λj > 0, j = 1, 2, . . ., λ2j → ∞ and {Fj (x)}∞
j=1 can be chosen as complete orthonormal
∞
set in L (Ω) (or {Fj (x)}j=1 forms an orthonormal basis of L2 (Ω)). This fact allows us
2

to represent f (x) in terms of Fourier series

∞
X
f (x) = fj Fj (x), (9.6)
j=1

where fj = (f, Fj )L2 (Ω) are called the Fourier coefficients of f with respect to {Fj }∞
j=1 .
If we take now ∞
X 2
u(x, t) = fj Fj (x)e−λj t , (9.7)
j=1

then we may conclude (at least formally) that

∞
X ∞
X
2 2
∂t u = − fj λ2j Fj (x)e−λj t = fj ∆Fj (x)e−λj t = ∆u,
j=1 j=1

that is, u(x, t) from (9.7) satisfies the heat equation and u(x, t) = 0 on ∂Ω × (0, ∞). It
remains to prove that u(x, t) satisfies the initial condition and to determine for which
functions f (x) the series (9.6) converges and in what sense. This is the main question
in the Fourier method.
It is clear that the series (9.6) and (9.7) (for t ≥ 0) converge in the sense of
L (Ω). It is also clear that if f ∈ C 1 (Ω) vanishes at the boundary then u will vanish
2

on ∂Ω × (0, ∞) and one easily verifies that u is a distributional solution of the heat
equation (t > 0). Hence it is a classical solution since u(x, t) ∈ C ∞ (Ω × (0, ∞)) (see
Corollary 2 of Theorem 1).
Similar considerations apply to the problem


∂t u − ∆u = 0, in Ω × (0, ∞)
u(x, 0) = f (x), in Ω


∂ν u(x, t) = 0, on ∂Ω × (0, ∞).
This problem boils down to finding orthonormal basis of eigenfunctions for Laplacian
with the Neumann boundary condition. Let us remark that for this problem, {0} is
always an eigenvalue and 1 is an eigenfunction.

106
Exercise 55. Prove that u(x, t) of the form (9.7) is a distributional solution of the
heat equation in Ω × (0, ∞).
Rπ
Exercise 56. Show that 0 |u(x, t)|2 dx is a decreasing function of t > 0, where u(x, t)
is the solution of (
ut − uxx = 0, 0 < x < π, t > 0
u(0, t) = u(π, t) = 0, t > 0.

107
10 The Wave Operator
The wave equation is defined as

∂t2 u(x, t) − ∆x u(x, t) = 0, (x, t) ∈ Rn × R. (10.1)

The wave equation is satisfied exactly by the components of the classical electromag-
netic field in vacuum.
The characteristic variety of (10.1) is

charx (L) = (ξ, τ ) ∈ Rn+1 : (ξ, τ ) 6= 0, τ 2 = |ξ|2

and it is called the light cone. Accordingly, we call

{(ξ, τ ) ∈ charx (L) : τ > 0}

and
{(ξ, τ ) ∈ charx (L) : τ < 0}
the forward and backward light cone, respectively.
The wave operator is a prototype of hyperbolic operators. It means that the main
symbol X
aα (x, t)ξ α τ j
|α|+j=k

has k distinct real roots with respect to τ .

Theorem 1. Suppose u(x, t) is C 2 function and that ∂t2 u − ∆u = 0. Suppose also that
u = 0 and ∂ν u = 0 on the ball B = {(x, 0) : |x − x0 | ≤ t0 } in the hyperplane t = 0.
Then u = 0 in the region Ω = {(x, t) : 0 ≤ t ≤ t0 , |x − x0 | ≤ t0 − t}.

Proof. By considering real and imaginary parts we may assume that u is real. Denote
by Bt = {x : |x − x0 | ≤ t0 − t}. Let us consider the following integral
Z
1

E(t) = (ut )2 + |∇x u|2 dx
2 Bt

which represents the energy of the wave in Bt at time t. Next,

Z n
!
X
E ′ (t) = ut utt + ∂j u(∂j u)t dx
Bt j=1
Z
1

− (ut )2 + |∇x u|2 dσ(x) := I1 + I2 .
2 ∂Bt

108
Straightforward calculations using the divergence theorem show us that
Z n n
!
X X
I1 = ∂j [(∂j u)ut ] − (∂j2 u)ut + ut utt dx
Bt j=1 j=1
Z Z n
X
= ut (utt − ∆x u)dx + (∂j u)νj ut dσ(x)
Bt ∂Bt j=1
Z Z
1

≤ |ut | |∇x u|dσ(x) ≤ |ut |2 + |∇x u|2 dσ(x) ≡ −I2 .
∂Bt 2 ∂Bt

Hence
dE
≤ −I2 + I2 = 0.
dt
But E(t) ≥ 0 and E(0) = 0 due to Cauchy data. Therefore E(t) ≡ 0 if 0 ≤ t ≤ t0 and
thus ∇x,t u = 0 in Ω. Since u(x, 0) = 0 then u(x, t) = 0 also in Ω.
(x0 , t0 )
b

x0 b
t0

Remark. This theorem shows that the value of u at (x0 , t0 ) depends only on the Cauchy
data of u on the ball {(x, 0) : |x − x0 | ≤ t0 }.
Conversely, the Cauchy data on a region R in the initial (t = 0) hyperplane influ-
ence only those points inside the forward light cones issuing from points of R. Sim-
ilar result holds when the hyperplane t = 0 is replaced by a space-like hypersurface
S = {(x, t) : t = ϕ(x)}. A surface S is called space-like if its normal vector ν = (ν ′ , ν0 )
satisfies |ν0 | > |ν ′ | at every point of S, i.e., if ν lies inside the light cone. It means that
|∇ϕ| < 1.

Let us consider the Cauchy problem for the wave equation:

(
∂t2 u − ∆u = 0, x ∈ Rn , t > 0
(10.2)
u(x, 0) = f (x), ∂t u(x, 0) = g(x).

Definition. If ϕ is a continuous function on Rn and r > 0, we define the spherical

mean Mϕ (x, r) as follows:
Z Z
1 1
Mϕ (x, r) := n−1 ϕ(z)dσ(z) = ϕ(x + ry)dσ(y).
r ωn |x−z|=r ωn |y|=1

109
Lemma 1. If ϕ is a C 2 function on Rn , then Mϕ (x, 0) = ϕ(x) and
 
2 n−1
∆x Mϕ (x, r) = ∂r + ∂r Mϕ (x, r).
r

Proof. It is clear that

Z
1
Mϕ (x, 0) = ϕ(x)dσ(y) = ϕ(x).
ωn |y|=1

For the second part we have, by the divergence theorem, that

Z n
X Z
1 1
∂r Mϕ (x, r) = yj ∂j ϕ(x + ry)dσ(y) = r∆ϕ(x + ry)dy
ωn |y|=1 j=1
ωn |y|≤1
Z
1
= ∆ϕ(x + z)dz
rn−1 ωn |z|≤r
Z r Z
1 n−1
= n−1 ρ dρ ∆ϕ(x + ρy)dσ(y).
r ωn 0 |y|=1

That’s why we have

Z
n−1


rn−1
∂r r ∂r Mϕ (x, r) = ∆ϕ(x + ry)dσ(y) ≡ rn−1 ∆x Mϕ (x, r).
ωn |y|=1

It implies that

(n − 1)rn−2 ∂r Mϕ (x, r) + rn−1 ∂r2 Mϕ (x, r) = rn−1 ∆x Mϕ (x, r)

and proves the claim.

Corollary. Suppose u(x, t) is a C 2 function on Rn+1 and let

Z Z
1 1
Mu (x, r, t) = n−1 u(z, t)dσ(z) = u(x + ry, t)dσ(y).
r ωn |x−z|=r ωn |y|=1

Then u(x, t) satisfies the wave equation if and only if

 
2 n−1
∂r + ∂r Mu (x, r, t) = ∂t2 Mu (x, r, t). (10.3)
r

Lemma 2. If ϕ ∈ C k+1 (R), k ≥ 1, then

 k−1  k
1
∂r
∂r2 ∂r r 2k−1
ϕ(r) = (r2k ϕ′ ).
r r

110
Proof. We employ induction with respect to k. If k = 1 then
 k−1
2 1

∂r ∂r r2k−1 ϕ(r) = ∂r2 (rϕ) = ∂r (ϕ + rϕ′ ) = 2ϕ′ + rϕ′′
r
and  k  
∂r 2k ′ ∂r
(r ϕ ) = (r2 ϕ′ ) = 2ϕ′ + rϕ′′ .
r r
Assume that  k−1  k
1
∂r
∂r2 ∂r r 2k−1
ϕ(r) = (r2k ϕ′ ).
r r
Then
 k  k−1  
2 1 2k+1

2 1 ∂r 2k+1

∂r ∂r r ϕ(r) = ∂r ∂r r ϕ
r r r
 k−1
2 1

= ∂r ∂r (2k + 1)r2k−1 ϕ + r2k ϕ′
r
 k−1  k−1
2 1 2k−1

2 1

= (2k + 1)∂r ∂r r ϕ + ∂r ∂r r2k ϕ′
r r
 k  k
∂r
∂r
= (2k + 1) r2k ϕ′ + (r2k (rϕ′ )′ )
r r
 k
∂r

= (2k + 1)r2k ϕ′ + r2k (rϕ′ )′
r
 k
∂r

= (2k + 1)r2k ϕ′ + r2k ϕ′ + r2k+1 ϕ′′
r
 k
∂r

= (2k + 2)r2k ϕ′ + r2k+1 ϕ′′
r
 k+1
∂r

= r2k+2 ϕ′ .
r

Corollary of Lemma 1 gives that if u(x, t) is a solution of the wave equation (10.1)
in Rn × R then Mu (x, r, t) satisfies (10.3), i.e.,
 
2 n−1
∂r + ∂r Mu = ∂t2 Mu ,
r
with initial conditions:

Mu (x, r, 0) = Mf (x, r), ∂t Mu (x, r, 0) = Mg (x, r), (10.4)

since u(x, 0) = f (x) and ∂t u(x, 0) = g(x).

111
 Let us set
  n−3
∂r 2

e(x, r, t) :=
u rn−2 Mu ≡ T Mu ,
r (10.5)
fe(x, r) := T Mf , ge(x, r) := T Mg

for n = 2k + 1, k = 1, 2, . . ..

Lemma 3. The following is true:

(
e = ∂t2 u
∂r2 u e
(10.6)
u e
e|t=0 = f , ∂t u
e|t=0 = ge,

e, fe and ge are defined in (10.5).

where u

Proof. Since n = 2k + 1 then n−3 2

= k − 1 and n − 2 = 2k − 1. Hence we obtain from
Lemmata 1 and 2 that
 k−1  k
2 2 2 ∂r 2k−1

∂r

∂r ue = ∂ r T Mu = ∂ r r Mu = r2k ∂r Mu
r r
 k−1
∂r

= 2kr2k−2 ∂r Mu + r2k−1 ∂r2 Mu
r
 k−1     k−1
∂r 2k−1 2 n−1 ∂r

= r ∂ r Mu + ∂ r Mu = r2k−1 ∂t2 Mu
r r r
 k−1
∂r

= ∂t2 r2k−1 Mu = ∂t2 u e.
r

Moreover, the initial conditions are satisfied due to (10.4) and (10.5).
But now, since (10.6) is a one-dimensional problem, we may conclude that u e(x, r, t)
from Lemma 3 is equal to
 Z r+t 
1 e e
e(x, r, t) =
u f (x, r + t) + f (x, r − t) + ge(x, s)ds . (10.7)
2 r−t

Lemma 4. If n = 2k + 1, k = 1, 2, . . ., then

e(x, r, t)
u
Mu (x, 0, t) = lim ,
r→0 (n − 2)!!r

where (n − 2)!! = 1 · 3 · 5 · · · (n − 2), is the solution of (10.2). We have even more,

namely,  
1 e
u(x, t) = ∂r f |r=t + ge(x, t) . (10.8)
(n − 2)!!

112
Proof. By (10.5) we have
 k−1  k−2
∂r 2k−1

∂r

e(x, r, t) =
u r Mu = (2k − 1)r2k−3 Mu + r2k−2 ∂r Mu
r r
= (2k − 1)(2k − 3) · · · 1 · Mu r + O(r2 )

or
e(x, r, t)
u
= Mu + O(r).
(n − 2)!!r
Hence
e(x, r, t)
u
Mu (x, 0, t) = lim .
r→0 (n − 2)!!r

But by definition of Mu we have that Mu (x, 0, t) = u(x, t), where u(x, t) is the solution
of (10.2). The initial conditions in (10.2) are satisfied due to (10.5). Next, since
e(x, r, t) satisfies (10.7) then
u
Z !
e(x, r, t)
u 1 fe(x, r + t) + fe(x, r − t) 1 r+t
lim = lim + lim ge(x, s)ds
r→0 (n − 2)!!r 2(n − 2)!! r→0 r r→0 r r−t

1  
= ∂r fe|r=t + ∂r fe|r=−t + ge(x, t) − ge(x, −t) ,
2(n − 2)!!

because fe(x, t) and ge(x, t) are odd functions of t. That’s why we finally obtain
e(x, r, t)
u 1  
lim = ∂r fe|r=t + ge(x, t) .
r→0 (n − 2)!!r (n − 2)!!

Now we are in the position to prove the main theorem for odd n ≥ 3.
n+3 n+1
Theorem 2. Suppose that n = 2k+1, k = 1, 2, . . .. If f ∈ C 2 (Rn ) and g ∈ C 2 (Rn )
then
(   n−3  Z 
1 ∂t 2 n−2
u(x, t) = ∂t t f (x + ty)dσ(y)
(n − 2)!!ωn t |y|=1
  n−3  Z ) (10.9)
∂t 2
+ tn−2 g(x + ty)dσ(y)
t |y|=1

solves (10.2).
Proof. Due to Lemmata 3 and 4 u(x, t) given by (10.8) is the solution of the wave
equation. It remains only to check that this u satisfies the initial conditions. But
(10.9) gives us for small t that

u(x, t) = Mf (x, t) + tMg (x, t) + O(t2 ).

113
It implies that

u(x, 0) = Mf (x, 0) = f (x), ∂t u(x, 0) = ∂t Mf (x, 0) + Mg (x, 0) = g(x).

The last equality follows from the fact that Mf (x, t) is even in t and so its derivative
vanishes at t = 0.
Remark. If n = 3 then (10.9) becomes
  Z  Z 
1
u(x, t) = ∂t t f (x + ty)dσ(y) + t g(x + ty)dσ(y)
4π |y|=1 |y|=1
Z Z
1
≡ f (x + ty)dσ(y) + t ∇f (x + ty) · ydσ(y)
4π |y|=1 |y|=1
Z 
+ t g(x + ty)dσ(y) .
|y|=1

The solution of (10.2) for even n is readily derived from the solution for odd n by
”the method of descent”. This is just the trivial observation: if u is a solution of the
wave equation in Rn+1 × R that does not depend on xn+1 then u satisfies the wave
equation in Rn × R. Thus to solve (10.2) in Rn × R with even n, we think of f and g
as functions on Rn+1 which are independent of xn+1 .
n+4 n+2
Theorem 3. Suppose that n is even. If f ∈ C 2 (Rn ) and g ∈ C 2 (Rn ) then the
function
(   n−2 Z !
2 ∂t 2 f (x + ty)
u(x, t) = ∂t tn−1 p dy
(n − 1)!!ωn+1 t |y|≤1 1 − y2
  n−2 Z !) (10.10)
∂t 2 g(x + ty)
+ tn−1 p dy
t |y|≤1 1 − y2

solves the Cauchy problem (10.2).

Proof. If n is even then n + 1 is odd and n + 1 ≥ 3. That’s why we can apply (10.9)
in Rn+1 × R to get that
(   n−2 Z !
1 ∂t 2
u(x, t) = ∂t tn−1 f (x + ty + tyn+1 )dσ(ey)
(n − 1)!!ωn+1 t y12 +···+yn
2 +y 2
n+1 =1
  n−2 Z !)
∂t 2
+ tn−1 g(x + ty + tyn+1 )dσ(e y) ,
t y12 +···+yn 2 +y 2
n+1 =1

(10.11)

where ye = (y, yn+1 ), solves (10.2) in Rn+1 × R (formally). But if we assume now
that f and g do not depend on xn+1 then u(x, t) does not depend on xn+1 either and

114
solves (10.2) in Rn × R. It remains only to calculate the integrals in (10.11) under this
assumption. We have
Z Z
f (x + ty + tyn+1 )dσ(e
y) = f (x + ty)dσ(e
y)
2
|y|2 +yn+1 =1 2
|y|2 +yn+1 =1
Z
dy
= 2 f (x + ty) p ,
|y|≤1 1 − |y|2

because we have the upper and lower hemispheres of the sphere |y|2 + yn+1
2
= 1.
Similarly for the second integral in (10.11). This proves the theorem.
Remark. If n = 2 then (10.10) becomes
( Z ! Z )
1 f (x + ty) g(x + ty)
u(x, t) = ∂t t p dy + t p dy .
2π |y|≤1 1 − y2 |y|≤1 1 − y2

Now we consider the Cauchy problem for the inhomogeneous wave equation
(
∂t2 u − ∆x u = w(x, t)
(10.12)
u(x, 0) = f (x), ∂t u(x, 0) = g(x).

We look for the solution u(x, t) of (10.12) as u = u1 + u2 , where

(
∂t2 u1 − ∆u1 = 0
(A)
u1 (x, 0) = f (x), ∂t u1 (x, 0) = g(x),

and (
∂t2 u2 − ∆u2 = w
(B)
u2 (x, 0) = ∂t u2 (x, 0) = 0.
For the problem (B) we will use a method known as Duhamel’s principle.

Theorem 4. Suppose w ∈ C [ 2 ]+1 (Rn × R). For s ∈ R let v(x, t; s) be the solution of
n

(
∂t2 v(x, t; s) − ∆x v(x, t; s) = 0
v(x, 0; s) = 0, ∂t v(x, 0; s) = w(x, s).

Then Z t
u(x, t) := v(x, t − s; s)ds
0

solves (B).

115
Proof. By definition of u(x, t) it is clear that u(x, 0) = 0. We also have
Z t
∂t u(x, t) = v(x, 0; t) + ∂t v(x, t − s; s)ds.
0

It implies that ∂t u(x, 0) = v(x, 0; 0) = 0. Differentiating once more in t we can see that
Z t
2
∂t u(x, t) = ∂t (v(x, 0; t)) + ∂t v(x, 0; t) + ∂t2 v(x, t − s; s)ds
0
Z t
= w(x, t) + ∆x v(x, t − s; s)ds
0
Z t
= w(x, t) + ∆x v(x, t − s; s)ds = w(x, t) + ∆x u.
0

Thus u solves (B) and the theorem is proved.

Let us consider again the homogeneous Cauchy problem (10.2). Applying the
Fourier transform with respect to x gives
(
∂t2 u
b(ξ, t) + |ξ|2 u
b(ξ, t) = 0
u b
b(ξ, 0) = f (ξ), ∂t u b(ξ, 0) = gb(ξ).

But this ordinary differential equation with initial conditions can be easily solved to
obtain
 
b sin(|ξ|t) b sin(|ξ|t) sin(|ξ|t)
ub(ξ, t) = f (ξ) cos(|ξ|t) + gb(ξ) ≡ f (ξ)∂t + gb(ξ) .
|ξ| |ξ| |ξ|

It implies that
   
sin(|ξ|t) sin(|ξ|t)
u(x, t) = F −1
fb(ξ)∂t +F −1
gb(ξ)
|ξ| |ξ|
     
−n/2 −1 sin(|ξ|t) −n/2 −1 sin(|ξ|t)
= f ∗ ∂t (2π) F + g ∗ (2π) F
|ξ| |ξ|
= f ∗ ∂t Φ(x, t) + g ∗ Φ(x, t), (10.13)
 
where Φ(x, t) = (2π)−n/2 F −1 sin(|ξ|t)
|ξ|
.
The next step is to try to solve the equation

∂t2 F (x, t) − ∆x F (x, t) = δ(x)δ(t).

By Fourier transform in x we obtain

∂t2 Fb(ξ, t) + |ξ|2 Fb(ξ, t) = (2π)−n/2 δ(t).

116
That’s why Fb must be a solution of ∂t2 u + |ξ|2 u = 0 for t 6= 0. Therefore
(
a(ξ) cos(|ξ|t) + b(ξ) sin(|ξ|t), t < 0
Fb(ξ, t) =
c(ξ) cos(|ξ|t) + d(ξ) sin(|ξ|t), t > 0.

To obtain the delta function at t = 0 we require that Fb is continuous at t = 0 but ∂t Fb

has a jump of size (2π)−n/2 at t = 0. So we have

a(ξ) = c(ξ), |ξ|(d(ξ) − b(ξ)) = (2π)−n/2 .

This gives two equations for the four unknown coefficients a, b, c and d. But it is
1
reasonable to require F (x, t) ≡ 0 for t < 0. Hence, a = b = c = 0 and d = (2π)−n/2 |ξ| .
That’s why (
b (2π)−n/2 sin(|ξ|t)
|ξ|
, t>0
F (ξ, t) = (10.14)
0, t < 0.
If we compare (10.13) and (10.14) we may conclude that
 
−n/2 −1 sin(|ξ|t)
F (x, t) = (2π) Fξ , t>0
|ξ|

and (
Φ(x, t), t > 0
Φ+ (x, t) =
0, t<0
is the fundamental solution of the wave equation, i.e., F (x, t) with t > 0.
There is one more observation. If we compare (10.9) and (10.10) with (10.13) then
we may conclude that these three formulae are the same. Hence, we may calculate the
inverse Fourier transform of
sin(|ξ|t)
(2π)−n/2
|ξ|
in odd and even dimensions respectively with (10.9) and (10.10). Actually, the result
is presented in these two formulae.
When solving the wave equation in the region Ω × (0, ∞), where Ω is a bounded
domain in Rn , it is necessary to specify not only Cauchy data on Ω × {0} but also some
conditions on ∂Ω × (0, ∞) to tell the wave what to do when it hits the boundary. If the
boundary conditions on ∂Ω × (0, ∞) are independent of t, the method of separation of
variables can be used.
Let us (for example) consider the following problem:

 2
∂t u − ∆x u = 0, in Ω × (0, ∞)
u(x, 0) = f (x), ∂t u(x, 0) = g(x), in Ω (10.15)


u(x, t) = 0, on ∂Ω × (0, ∞).

117
We can look for solution u in the form u(x, t) = F (x)G(t) and get
(
∆F (x) + λ2 F (x) = 0, in Ω
(10.16)
F (x) = 0, on ∂Ω,

and
G′′ (t) + λ2 G(t) = 0, 0 < t < ∞. (10.17)
The general solution of (10.17) is

G(t) = a cos(λt) + b sin(λt).

 2 ∞
Since (10.16) has infinitely many solutions {Fj }∞ 2
j=1 with corresponding λj j=1 , λj →
+∞ (λj > 0) and {Fj }∞ 2
j=1 can be chosen as an orthonormal basis in L (Ω), the solution
u(x, t) of (10.15) is of the form
∞
X
u(x, t) = Fj (x) (aj cos(λj t) + bj sin(λj t)) . (10.18)
j=1

At the same time f (x) and g(x) have the L2 (Ω) representations
∞
X ∞
X
f (x) = fj Fj (x), g(x) = gj Fj (x), (10.19)
j=1 j=1

where fj = (f, Fj )L2 and gj = (g, Fj )L2 . It follows from (10.15) and (10.18) that
∞
X ∞
X
u(x, 0) = aj Fj (x), ut (x, 0) = λj bj Fj (x). (10.20)
j=1 j=1

Since (10.19) must be satisfied also we obtain

1
aj = f j , bj = gj .
λj

Therefore, the solution u(x, t) of (10.15) has the form

X∞  
1
u(x, t) = Fj (x) fj cos(λj t) + gj sin(λj t) .
j=1
λj

It is clear that all series (10.18),(10.19) and (10.20) converge in L2 (Ω), because {Fj }∞
j=1
is an orthonormal basis in L2 (Ω). It remains only to investigate the convergence of
these series in stronger norms (which depends on f and g, or more precisely, on their
smoothness).
The Neumann problem with ∂ν u(x, t), x ∈ ∂Ω, can be considered in a similar
manner.

118
Index
δ-function, 8 Gaussian kernel, 100
Gibbs phenomenon, 29
a translation, 57 gradient, 10
approximation to the identity, 5 Green’s function, 71
biharmonic equation, 11 Green’s identities, 59
Burgers equation, 16 Hans Lewy example, 21
Cauchy data, 17 harmonic function, 59
Cauchy problem, 17 Harnack’s inequality, 80
Cauchy-Kowalevski theorem, 18 heat equation, 10, 32
Cauchy-Riemann operator, 11, 63 heat operator, 100
characteristic, 11 hyperplane, 3
characteristic form, 11 hypersurface, 3
characteristic variety, 11 ill-posed problem, 21
continuous kernel, 84 integral curves, 12
convolution, 4 interior Dirichlet problem, 82
d’Alembert formula, 48 interior Neumann problem, 82
differential operator, 10 Korteweg-de Vries equation, 11
Dirichlet problem, 51
distribution, 8 Laplace equation, 51
distributional solution, 10 Laplace operator, 10, 57
divergence theorem, 4 Laplacian, 10, 57
double layer potential, 84 linear superposition principle, 35
Duhamel’s principle, 115 Liouville’s theorem, 62

eigenvalue problem, 33 maximum principle, 61, 104

eikonal equation, 10 mean value theorem, 60
elliptic differential operator, 11 method of characteristics, 13
even function, 23 multi-index, 1
evolution equation, 10 mutually orthogonal functions, 23
exterior Dirichlet problem, 82
exterior Neumann problem, 82 Neumann problem, 51
non-characteristic, 11
Fourier cosine series, 26 normal, 3
Fourier inversion formula, 7
Fourier series, 25 odd function, 23
Fourier sine series, 26 orthogonal complement, 96
Fourier transform, 6 orthogonal functions, 23
fundamental period, 23 periodic function, 23
fundamental solution, 63 piecewise continuous function, 23

119
Plancherel theorem, 7
Poisson equation, 10
Poisson integral, 74
Poisson kernel, 74
principal symbol, 11

quasi-linear equation, 14

Reflection Principle, 80
regular distribution, 8
removable singularity, 80
Riemann-Lebesgue lemma, 6
rotation, 57

Schwartz space, 7
separation of variables, 33
Sine-Gordon equation, 11
single layer potential, 84
spherical mean, 109
support, 7

telegrapher’s equation, 11
tempered distribution, 9
tubular neighborhood, 93

wave equation, 10, 44, 108

wave operator, 108
weak solution, 10
well-posed problem, 20

Young’s inequality for convolution, 4

120