Ralph Chill

Partial differential equations

– Sobolev spaces and variational methods –

April 29, 2016

c by R. Chill
 v

Vorwort:

Ich habe dieses Skript zur Vorlesung Partielle Differentialgleichungen an

der TU Dresden nach meinem besten Wissen und Gewissen geschrieben.
Mit Sicherheit schlichen sich jedoch Druckfehler oder gar mathematische
Ungenauigkeiten ein, die man beim ersten Schreiben eines Skripts nicht
vermeiden kann. Möge man mir diese Fehler verzeihen.
Obwohl die Vorlesung auf Deutsch gehalten wird, habe ich mich
entschieden, dieses Skript auf Englisch zu verfassen. Auf diese Weise wird
eine Brücke zwischen der Vorlesung und der (meist englischsprachigen) Lit-
eratur geschlagen. Mathematik sollte jedenfalls unabhängig von der Sprache
sein in der sie präsentiert wird.
Ich danke Susanne Stimpert fr die Hilfe beim Erstellen dieses Skripts.
Für weitere Kommentare, die zur Verbesserungen beitragen, bin ich sehr
dankbar.
Contents

1 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The space of test functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Weak derivatives and Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Sobolev spaces in one dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 The extension property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 The Sobolev embedding theorems . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1,p
1.7 The space W0 (Ω) and the Poincaré inequality . . . . . . . . . . . . . . 31

2 Elliptic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1 The Lax-Milgram lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 The Laplace operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 General elliptic operators in divergence form . . . . . . . . . . . . . . . 42
2.4 The comparison and maximum principles . . . . . . . . . . . . . . . . . . 45
2.5 Regularity of weak solutions of elliptic equations . . . . . . . . . . . . 49

3 Evolution equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Wellposedness results for abstract diffusion equations, wave
equations and Schrödinger equations . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 The comparison principle for diffusion equations . . . . . . . . . . . . 52

4 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1 The topology in D(Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.3 The product and the convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Tempered distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.5 The Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.5.1 The Fourier transform on L1 (RN ) . . . . . . . . . . . . . . . . . . . . 63
4.5.2 The Fourier transform on S(RN ) . . . . . . . . . . . . . . . . . . . . . 68
4.5.3 The Fourier transform on L2 . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.5.4 The Fourier transform on S(RN )0 . . . . . . . . . . . . . . . . . . . . 70

vii
viii Contents

4.6 The theorem of Malgrange-Ehrenpreis . . . . . . . . . . . . . . . . . . . . . . 70

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Chapter 1
Sobolev spaces

1.1 The space of test functions

For subsets Ω of RN , we denote by

C(Ω) := {u : Ω → C : u is continuous}

the space of continuous complex-valued functions. Sometimes, we only con-

sider real-valued functions, especially when order properties such as posi-
tivity or comparison are involved, but this will be clear from the context. If
Ω is in addition measurable, then we denote by
Z
Lp (Ω) := {u : Ω → C : u is measurable, |u|p < ∞} (p ∈ [1, ∞)) and
Ω
L∞ (Ω) := {u : Ω → C : u is measurable, and {u > c} is a null set for some C ≥ 0}

the usual Lebesgue spaces which are equipped with the norms
Z ! p1
p
kukLp := |u| (p ∈ [1, ∞)), and
Ω
kukL∞ := inf{C ≥ 0 : {u > C} is a null set}.

Strictly speaking, Lp (Ω) is a space of equivalence classes of measurable func-

tions but it is often convenient to work with representatives instead of equiv-
alence classes. For measurable Ω ⊆ RN , we set
Z
p
Lloc (Ω) := {u : Ω → C : u is measurable and for every K compact |u|p < ∞},
K

the space of locally p-integrable functions. For every p ∈ [1, ∞] one has

1
2 1 Sobolev spaces
p
Lp (Ω) ⊆ Lloc (Ω) ⊆ L1loc (Ω) and
C(Ω) ⊆ L∞ 1
loc (Ω) ⊆ Lloc (Ω).

Hence, among all spaces defined above, L1loc (Ω) is the largest space.
For a function u ∈ L1loc (Ω) we define the support by
[
supp u := Ω \ U.
U⊆Ω rel. open
u=0 a.e. on U

By definition, the support is relatively closed in Ω (but it is in general not

closed in RN ). The definition of the support depends on the underlying set
Ω, and although the spaces Lp ([0, 1]) and Lp ((0, 1)) coincide, the support of
their elements do not coincide in general; if nothing is said explicitly, the
underlying set Ω is usually supposed to be open in RN .
We say that a function u ∈ L1loc (Ω) has compact support if supp u is compact,
and we set
p
Lc (Ω) := {u ∈ Lp (Ω) : supp u is compact},
Cc (Ω) := {u ∈ C(Ω) : supp u is compact},
Ckc (Ω) := Cc (Ω) ∩ Ck (Ω),
C∞ ∞
c (Ω) := Cc (Ω) ∩ C (Ω).

For the definition of the latter two spaces we suppose that Ω is open. The
space C∞ ∞
c (Ω) is sometimes also denoted by D(Ω). Elements of Cc (Ω) are
∞
called test functions; Cc (Ω) is called the space of test functions.

Example 1.1. The function

− 1 2

PN
e 1−kxk if kxk2 = 2
i=1 xi < 1,



u(x) := 

0

else

is a test function on RN and supp u is the closed unit ball.

Exercise 1.2 Show that for a function u ∈ C(Ω)

Ω
supp u = {x ∈ Ω : u(x) , 0} ,

where the closure is understood in the relative Euclidean topology induced on Ω.

Exercise 1.3 Show that every u ∈ L1loc (Ω) is almost everywhere equal to 0 on Ω \
supp u.
1.2 The convolution 3

1.2 The convolution

Theorem 1.4 (Young). Fix p ∈ [1, ∞], f ∈ L1 (RN ) and g ∈ Lp (RN ). Then, for almost
every x ∈ RN , the function
y 7→ f (x − y)g(y)
is integrable on RN , and the function f ∗ g : RN → C defined by
Z
( f ∗ g)(x) := f (x − y)g(y) dy
RN

belongs to Lp (RN ). Moreover,

k f ∗ gkLp ≤ k f kL1 kgkLp

Proof. The case p = 1. Let us first assume that p = 1. Then Tonelli’s theorem
yields
Z Z Z Z
| f (x − y) g(y)| dy dx = | f (x − y)| dx|g(y)| dy
RN RN N N
ZR ZR
= | f (x)| dx |g(y)| dy
RN RN
= k f kL1 kgkL1 < ∞.

In particular, the left-hand Rside is finite which is only possible if for almost
every x ∈ RN , the integral RN | f (x − y) g(y)| dy is finite, that is, the function
y 7→ f (x − y)g(y) is integrable. Since the left-hand side equals k f ∗ gkL1 , the
above inequality also yields k f ∗ gkL1 ≤ k f kL1 kgkL1 .
The case 1 < p < ∞. The assumption g ∈ Lp (RN ) implies |g|p ∈ L1 (RN ), and
then the first step implies that for almost every x ∈ RN , the function y 7→ | f (x−
y)| |g(y)|p is integrable. This, in turn, is equivalent to saying that for almost
1
every x ∈ RN , the function y 7→ | f (x − y)| p |g(y)| is p-integrable. On the other
1 0 p
hand, | f | p0 ∈ Lp (RN ), where p0 = p−1 is the conjugate exponent. Hence, by
Hölder’s inequality, for almost every x ∈ RN , the function y 7→ | f (x − y)| |g(y)|
is integrable, and
Z
| f ∗ g(x)| ≤ | f (x − y)| |g(y)| dy
RN
Z ! 10 Z ! 1p
p
p
≤ | f (x − y)| dy | f (x − y)| |g(y)| dy
RN RN
p−1 Z ! 1p
p
= kfk | f (x − y)| |g(y)| dy .
p
L1
RN
4 1 Sobolev spaces

Raising both sides of this inequality to the p-th power and integrating over
RN yields, by another application of Tonelli’s theorem,
Z
p
k f ∗ gkLp = | f ∗ g(x)|p dx
RN
Z Z
p−1
≤ kfk 1 | f (x − y)| |g(y)|p dy dx
L
RN RN
Z Z
p−1
= kfk 1 | f (x − y)| dx |g(y)|p dy
L
RN RN
p p
= k f k 1 kgkLp .
L

From here follows the claim for p ∈ (1, ∞).

The statement for p = ∞ is evident.
p p
Corollary 1.5. Fix p ∈ [1, ∞], f ∈ L1loc (RN ) and g ∈ Lc (RN ) (or f ∈ Lloc (RN ) and
g ∈ L1c (RN )). Then, for almost every x ∈ RN , the function

y 7→ f (x − y)g(y)

is integrable on RN , and the function f ∗ g : RN → C defined by

Z
( f ∗ g)(x) := f (x − y)g(y) dy
RN
p
belongs to Lloc (RN ).
We call the function f ∗ g from Theorem 1.4 and Corollary 1.5 the convo-
lution of f and g. Let us discuss some properties of the convolution and its
use in connection with the approximation of Lp -functions by test functions.
For subsets A, B ⊆ RN we set

A + B := {x + y : x ∈ A, y ∈ B}.

Lemma 1.6 (Support of a convolution). For appropriate functions f and g one

has
supp f ∗ g ⊆ supp f + supp g.
Proof. For x < supp f + supp g one has

(x − supp f ) ∩ supp g = ∅,

and therefore
Z
( f ∗ g)(x) = f (x − y) g(y) dy = 0.
(x−supp f )∩supp g

In other words, f ∗ g = 0 everywhere on the complement of supp f + supp g.

1.2 The convolution 5

For every h ∈ RN and every function f : RN → C we define the shift

τh f : RN → C,
τh f (x) := f (x + h) (x ∈ RN ).
Lemma 1.7 (Strong continuity of the shifts). For every p ∈ [1, ∞) and every
f ∈ Lp (RN ) one has
lim kτh f − f kLp = 0.
x→0
Proof. For characteristic functions f = 1Q with Q = (a1 , b1 ) × · · · × (aN , bN ), the
claim follows from Lebesgue’s dominated convergence theorem. In fact, this
function is continuous at almost every point, except on the null set ∂Q. By the
triangle inequality, the claim thus holds for every f ∈ U, where U is the linear
span of characteristic functions of the form 1Q . By properties of the Lebesgue
measure and by properties of p-integrable functions, the space U is dense in
Lp (RN ). Let f ∈ Lp (RN ) and ε > 0. Choose g ∈ U such that k f − gkLp < ε. Then

lim sup kτh f − f kLp

h→0
≤ lim sup[kτh f − τh gkLp + kτh g − gkLp + kg − f kLp ]
h→0
< 2ε.

Since ε > 0 was arbitrary, we obtain the claim.

TheoremR1.8 (Approximation of the identity). Let ϕ ∈ L1 (RN ) be positive and
such that RN ϕ = 1. Set

ϕn (x) := nN ϕ(nx) (x ∈ RN , n ∈ N),

R
so that ϕn ∈ L1 (RN ) is positive and RN ϕn = 1. Then, for every p ∈ [1, ∞) and every
f ∈ Lp (RN ),
lim k f ∗ ϕn − f kLp = 0
n→∞

Proof. Assume first that p = 1, that is, f ∈ L1 (RN ). Then, by an application

of Tonnelli’s theorem, Lemma 1.7, and Lebesgue’s dominated convergence
theorem,
Z Z
k f ∗ ϕn − f kL1 = f (x − y)ϕn (y) dy − f (x) dx
N N
ZR ZR
= [ f (x − y) − f (x)] ϕn (y) dy dx
N N
ZR Z R
y
≤ | f (x − ) − f (x)| dx ϕ(y) dy
N N n
ZR R
= kτ y f − f kL1 ϕ(y) dy
n
RN
→ 0 as n → ∞.
6 1 Sobolev spaces

Assume next that p ∈ (1, ∞). Let f ∈ L1 ∩ L∞ (RN ). Recall the interpolation
inequality
1 1
p p0
k f kLp ≤ k f k 1 k f kL∞ ,
L
which is a straightforward consequence of Hölder’s inequality. Applying
this inequality to f ∗ ϕn − f ∈ L1 ∩ L∞ (RN ) (by Young’s inequality), we obtain,
by applying also the first step,
1 1
p0
k f ∗ ϕn − f kLp ≤ k f ∗ ϕn − f k 1 k f ∗ ϕn − f kL∞
p
L
1 1
p
≤ k f ∗ ϕn − f k 1 (2k f kL∞ ) p0
L
→ 0 as n → ∞.

Now let f ∈ Lp (RN ) be arbitrary, and let ε > 0. Since L1 ∩ L∞ (RN ) is dense
in Lp (Ω), there exists g ∈ L1 ∩ L∞ (RN ) such that k f − gkLp < ε. Hence, by an
application of Young’s inequality and the preceding inequality,

lim sup k f ∗ ϕn − f kLp

n→∞
≤ lim sup[k f ∗ ϕn − g ∗ ϕn kLp + kg ∗ ϕn − gkLp + kg − f kLp ]
n→∞
≤ lim sup[k f − gkLp kϕn kL1 + kg − f kLp ]
n→∞
< 2ε.

Since ε > 0 was arbitrary, the claim follows.

Theorem 1.9 (Smoothing effect of the convolution). For every f ∈ L1loc (RN )
and every g ∈ Ckc (RN ) (with k ∈ N0 ∪ {∞}) one has f ∗ g ∈ Ck (RN ) and for every
multi-index α ∈ NN 0
of length |α| ≤ k one has

∂α ( f ∗ g) = f ∗ (∂α g).

Proof. Assume first k = 0. For every x0 ∈ RN one has, by continuity of g and

by an application of Lebesgue’s dominated convergence theorem,
Z
lim sup |( f ∗ g)(x) − ( f ∗ g)(x0 )| ≤ lim sup | f (y)| |g(x − y) − g(x0 − y)| dy = 0,
x→x0 x→x0 RN

and hence f ∗ g is continuous at x0 . Since x0 was arbitrary, we have proved

that f ∗ g is continuous.
Assume next that k = 1. Fix x ∈ RN , j ∈ {1, . . . , N}, and let e j ∈ RN be the
j-th canonical unit vector. Then, by an application of Lebesgue’s dominated
convergence theorem again,
1.2 The convolution 7

( f ∗ g)(x + te j ) − ( f ∗ g)(x) g(x + te j − y) − g(x − y)

Z
= f (y) dy
t RN t
∂g
Z
→ f (y) (x − y) dy
RN ∂x j
∂g
= (f ∗ )(x) as t → 0.
∂x j

Since x ∈ RN was arbitrary, we have shown that the function f ∗ g is partially

differentiable with respect to x j , and the partial derivative ∂x∂ ( f ∗ g) coincides
j
∂g
with f ∗ ∂x j
. Since the latter function is continuous by the first part of the
proof (k = 0), we obtain that f ∗ g is continuously partially differentiable,
which is equivalent to saying that f ∗ g is continuously differentiable, that is,
f ∗ g ∈ C1 (RN ).
The case k > 1 follows by induction.

For every subset A ⊆ RN and every ε > 0 we define

Aε := {x ∈ RN : dist (x, A) < ε},

where
dist (x, A) := inf{kx − yk : y ∈ A}.

Lemma 1.10 (Further examples of test functions). For every compact subset
K ⊆ RN and every ε > 0 there exists a test function ψ ∈ C∞ N
c (R ) such that

0 ≤ ψ(x) ≤ 1 for every x ∈ RN ,

ψ(x) = 1 for every x ∈ K, and
ψ(x) = 0 for every x ∈ RN \ Kε .

Proof. Let K ⊆ RN be compact and ε > 0. Let ϕRbe a scalar multiple of the test
function from Example 1.1, so that ϕ ≥ 0 and RN ϕ = 1. Put

1 x
ϕε (x) := ϕ( ) (x ∈ RN ).
ε N ε

Then ϕε ∈ C∞ c (R ) is a positiveR test, too, supp ϕε = B̄(0, ε) (the closed ball

N

with center 0 and radius ε) and RN ϕε = 1. Let

ψ := 1Kε ∗ ϕε .

By the smoothing effect of the convolution (Theorem 1.9), ψ ∈ C∞ (RN ). By

Lemma 1.6 on the support of the convolution,

supp ψ ⊆ Kε + B̄(0, ε) = K2ε .

8 1 Sobolev spaces

In particular, the support of ψ is compact, that is, ψ ∈ C∞ N

c (R ), but this
inclusion also shows that ψ = 0 on R \ K . Since 1Kε and ϕε are positive,
N 2ε

their convolution is positive, too, that is, ψ ≥ 0. Moreover, for every x ∈ RN ,

Z Z
ψ(x) = 1Kε (x − y)ϕε (y) dy ≤ ϕε (y) dy = 1,
RN RN

so that 0 ≤ ψ ≤ 1 on RN . Finally, for every x ∈ K, the preceding inequality

turns into an equality, that is, ψ = 1 on K. Since ε > 0 was arbitrary, we have
proved the claim.
Theorem 1.11 (The test functions are dense in Lp (Ω)). For every open subset
Ω ⊆ RN and every p ∈ [1, ∞) the space of test functions C∞ p
c (Ω) is dense in L (Ω).

Proof (by truncation and regularization). Let Ω ⊆ RN be open, and let f ∈ Lp (Ω).
Truncation. First,S
we choose an increasing sequence (Kn ) of compact sub-
sets of Ω such that n Kn = Ω. For example, the sequence given by

1
Kn := {x ∈ Ω : dist (x, ∂Ω) ≥ and kxk ≤ n}
n
will do. Next, we choose a sequence (εn ) of positive reals such that

4 εn < dist (Kn , ∂Ω).

Finally, for each n we choose a test function (cut-off function) ψn ∈ C∞ c (Ω)

such that 0 ≤ ψn ≤ 1 on Ω, ψn = 1 on Kn and ψn = 0 on Ω \ Knεn (see Lemma
1.10). By choice of the Kn and the ψn , the sequence (ψn ) is uniformly bounded
by 1 and converges pointwise to the constant function 1. In particular, by
Lebesgue’s dominated convergence theorem,
Z
p
lim sup k f ψn − f kLp = lim sup | f (x)ψn (x) − f (x)|p dx = 0.
n→∞ n→∞ RN

Regularization. LetR ϕ ∈ C∞ N
c (R ) be a positive test function such that
supp ϕ ⊆ B̄(0, 1) and RN ϕ = 1. For example, one may take a scalar multi-
ple of the test function from Example 1.1. For every m ∈ N we set
x
ϕm (x) := ε−N
m ϕ( ) (x ∈ RN ).
εm
R
Then ϕm ∈ C∞ c (R ) is positive, supp ϕm ⊆ B̄(0, m ) and
N 1
RN
ϕm = 1. By Theorem
1.8, for every n ∈ N,

lim k f ψn − ( f ψn ) ∗ ϕm kLp = 0,
m→∞

where we have extended the function f ψn by 0 outside Ω in order to take the

convolution in RN . Since the sequence ( f ψn ) is convergent by the first step,
1.3 Weak derivatives and Sobolev spaces 9

and since it is thus relatively compact, the above limit is uniform in n ∈ N.

In particular, when we put

fn := ( f ψn ) ∗ ϕn ,

then
lim k fn − f kLp = 0.
n→∞

Moreover, by Theorem 1.9 on the smoothing effect of the convolution, fn ∈

C∞ (RN ). By Lemma 1.6 on the support of the convolution,

supp fn ⊆ supp ( f ψn ) + supp ϕn

⊆ supp ψn + B̄(0, εn )
⊆ Knεn + B̄(0, εn )
⊆ Kn3εn .

By the choice of εn , the support of fn is thus contained in Ω, so that fn ∈ C∞

c (Ω).
Summing up, ( fn ) is a desired sequence of test functions in C∞ c (Ω) which
converges in Lp (Ω) to f . Since f was arbitrary, this implies the claim.
In the truncation part of the preceding proof, it was actually not really im-
portant that the truncation function (cut-off function) ψn was a test function:
the characteristic function 1Kn would also be fine. However, as a corollary
of the above proof, which uses the slightly more complicated truncation
functions, we obtain the following result.
Theorem 1.12 (Uniqueness by testing). Let Ω ⊆ RN be open and f ∈ L1loc (Ω). If
Z
f ϕ = 0 for every ϕ ∈ C∞
c (Ω),
Ω

then f = 0.
Proof. Choose (ψn ) and (ϕn ), and define ( fn ) as in the proof of Theorem 1.11
above. If f ∈ L1 (Ω), then, by Theorem 1.11, fn → f in L1 (Ω). On the other
hand, the assumption implies fn = 0 on Ω, and hence f = 0. If f is only
locally integrable, then one may apply this reasoning to f 1ω , where ω ⊆ Ω is
an open subset such that ω̄ is compact and contained in Ω.

1.3 Weak derivatives and Sobolev spaces

We use the following notations for the partial derivative with respect to the
j-th variable in RN :
∂
or ∂x j or ∂ j .
∂x j
10 1 Sobolev spaces

Given a multi-index α ∈ NN
0
, we use also the abbreviation
α α
∂α := ∂x11 · · · · · ∂xNN

for the partial derivatives of order |α|.

Let Ω ⊆ RN be an open set, f ∈ L1loc (Ω) and α ∈ NN
0
. We say that ∂α f ∈ L1loc (Ω)
if there exists g ∈ L1loc (Ω) such that
Z Z
f ∂α ϕ = (−1)|α| gϕ for every ϕ ∈ C∞
c (Ω). (1.1)
Ω Ω

It follows from Theorem 1.12 that the function g is uniquely determined by

this identity. We write g =: ∂α f and call ∂α f the weak α-th partial derivative of
f . We say that f is k times weakly differentiable if ∂α f ∈ L1loc (Ω) for all multi-
indices α ∈ NN 0
with |α| ≤ k. By Gauß’ Theorem, every k times continuously
differentiable function f : Ω → C is k times weakly differentiable, and the
classical partial derivatives and the weak partial derivatives coincide.

Example 1.13. Consider on Ω := (−1, 1) the absolute value function f (x) = |x|
(x ∈ (−1, 1)). Then, for every ϕ ∈ C∞
c ((−1, 1)),
Z 1 Z 0 Z 1
f ϕ0 = (−x)ϕ0 (x) dx + xϕ0 (x) dx
−1 −1 0
Z 0 Z 1
0 1
= (−x)ϕ(x) −1
− (−1)ϕ(x) dx + xϕ(x) 0
− ϕ(x) dx
−1 0
Z 1
=− gϕ,
−1

where 
−1 if x ≤ 0,


g(x) := 
if x > 0.

1

We have thus shown that the absolute value function f is weakly differen-
tiable and f 0 = g. The interval (−1, 1) in this example may be replaced by any
other open set in R.

For every k ∈ N and every p ∈ [1, ∞] we define the Sobolev space

W k,p (Ω) := { f ∈ Lp (Ω) : f is k times weakly differentiable

and ∂α f ∈ Lp (Ω) for every α ∈ NN
0 , |α| ≤ k}.

One easily verifies that W k,p (Ω) is a vector space. We equip this space with
the norm
1.3 Weak derivatives and Sobolev spaces 11
  1p
 
 X 
α
 
k f kWk,p :=  k∂ f kLp  if p ∈ [1, ∞)
 
 α∈NN 
0
|α|≤k

and
k f kWk,∞ := sup k∂α f kL∞ if p = ∞,
α∈NN
0
|α|≤k

so that W k,p (Ω) is a normed space. We also define the Sobolev spaces

k,p W k,p
W0 (Ω) := C∞
c (Ω) ,

that is, the closure of the space of test functions in W k,p (Ω) with respect to
the norm k · kWk,p . Finally, we set

Hk (Ω) := W k,2 (Ω) and

H0k (Ω) := W0k,2 (Ω).

Both spaces are equipped with the inner product

X
h f, giHk := h∂α f, ∂α giL2 ,
α∈NN
0
|α|≤k

which turns them into inner product spaces. Note that the norm induced by
h·, ·iHk coincides with the norm k · kWk,2 =: k · kHk .
k,p
Theorem 1.14. The Sobolev spaces W k,p (Ω) and W0 (Ω) are Banach spaces. They
are separable if p ∈ [1, ∞), and reflexive if p ∈ (1, ∞). The spaces Hk (Ω) and H0k (Ω)
are separable Hilbert spaces.

Proof. Let A := {α ∈ NN 0
: |α| ≤ k} be the set of all multi-indices of length less
than or equal to k. Consider the cartesian product Lp (Ω)A equipped with the
norm
  p1
 
 X 
k( f α )α∈A k(Lp )A :=  k f α kLp  if p ∈ [1, ∞)

 
 
 α∈NN 
0
|α|≤k

and
k( f α )α∈A k(L∞ )A := sup k f α kL∞ if p = ∞,
α∈NN
0
|α|≤k

Then the mapping

12 1 Sobolev spaces

j : W k,p (Ω) → Lp (Ω)A , f 7→ (∂α f )α∈A

is an isometry. Since Lp (Ω) is a Banach space, it thus suffices to show that the
range of j is closed in Lp (Ω)A .
Completeness. Let ( fn ) be a sequence in W k,p (Ω) such that ((∂α fn )α∈A ) con-
verges in Lp (Ω)A to some element ( f α )α∈A . Put f := f (0,...,0) . By definition of
W k,p (Ω) and the weak partial derivatives, for every α ∈ A, every n ∈ N and
every ϕ ∈ C∞ c (Ω), Z Z
fn ∂α ϕ = (−1)|α| ∂α fn ϕ.
Ω Ω
The convergences fn → f and ∂α fn → f α in Lp (Ω) thus
imply that
Z Z
f ∂α ϕ = (−1)|α| f α ϕ for every α ∈ A, ϕ ∈ C∞
c (Ω).
Ω Ω

Hence, by definition of the weak partial derivatives, f ∈ W k,p (Ω) and ∂α f = f α .

In other words, ( f α )α∈A = j( f ), and we have shown that the range of j is closed.
In other words, W k,p (Ω) is complete.
Separability. The space Lp (Ω) being separable if p ∈ [1, ∞), the cartesian
product Lp (Ω)A is separable, too. Since every subset of a separable metric
space is separable (Exercise!), j(W k,p (Ω)) = W k,p (Ω) is separable if p ∈ [1, ∞).
Reflexivity. The space Lp (Ω) being reflexive if p ∈ (1, ∞), the cartesian prod-
uct Lp (Ω)A is reflexive, too. Since every closed subspace of a reflexive Banach
space is reflexive, j(W k,p (Ω)) = W k,p (Ω) is reflexive if p ∈ (1, ∞).

Theorem 1.15. For every p ∈ [1, ∞), the space C∞ N 1,p N

c (R ) is dense in W (R ). In
1,p
other words, W0 (RN ) = W 1,p (RN ).

Proof. Truncation and regularization. In fact, one may take the same sequence
( fn ) as constructed in the proof of Theorem 1.11, ensuring, however, that the
cut-off functions ψn have uniformly bounded gradient. This can easily be
achieved by looking into the proof of Lemma 1.10.
1,p
Remark 1.16. For general open sets Ω ⊆ RN , the spaces W0 (Ω) and W 1,p (Ω)
need not coincide. For example, they do not coincide for any interval Ω =
(a, b) , R (see Theorem ?? below).

Theorem 1.17 (Friedrichs). Let p ∈ [1, ∞) and let Ω ⊆ RN be open. Then for every
f ∈ W 1,p (Ω) there exists a sequence ( fn ) in C∞
c (Ω) such that

fn → f in Lp (Ω) and
∇ fn → ∇ f in Lp (ω) for every ω ⊂⊂ Ω.

Proof. Truncation and regularization. In fact, one may take the same sequence
of approximations as constructed in the proof of Theorem 1.11.
1.3 Weak derivatives and Sobolev spaces 13

Let Ω ∈ RN be open. For every f ∈ L1loc (Ω), ω ⊂⊂ Ω and h ∈ RN with

|h| < dist (ω, ∂Ω) one defines τh f by

τh f (x) := f (x + h) (x ∈ ω).

Theorem 1.18. Let Ω ⊆ RN be open, p ∈ (1, ∞], and f ∈ Lp (Ω). Then the following
assertions are equivalent:
(i) f ∈ W 1,p (Ω).
(ii) There exists a constant C ≥ 0 such that, for every ϕ ∈ C∞
c (Ω) and every
j ∈ {1, . . . , N}, Z
f ∂ j ϕ ≤ C kϕkLp0 .
Ω

(iii) There exists a constant C ≥ 0 such that, for every ω ⊂⊂ Ω and every h ∈ RN
with |h| < dist (ω, ∂Ω)
k f − τh f kLp ≤ C |h|.
Moreover, if the assertions (i)–(iii) are true, then one may take C = k∇ f kLp in (ii)
and (iii).

Proof. (i)⇒(ii) follows from the definition of weak derivative and an appli-
cation of Hölder’s inequality.
R Fix j ∈ {1, . . . , N}. By the assumption in (ii), the linear functional
(ii)⇒(i)
L j : ϕ → Ω f ∂ j ϕ is continuous on C∞ c (Ω) equipped with the norm induced
0 0
from L (Ω). Since the test functions are dense in Lp (Ω) by Theorem 1.11, we
p
0
may extend L j to a bounded linear functional on Lp (Ω). Recall that the dual
0
space of Lp (Ω) can be identified with Lp (Ω) (here we use p , 1), that is, there
exists a function g j ∈ Lp (Ω) such that, for every ϕ ∈ C∞ c (Ω)
Z Z
L j (ϕ) = f ∂ jϕ = g j ϕ.
Ω Ω

By definition of the weak derivative, this means that ∂ j f ∈ Lp (Ω) (and ∂ j f =

−g j ). Since j was arbitrary, we have thus proved f ∈ W 1,p (Ω).
(i)⇒(iii) Assume first that f ∈ C∞c (Ω). Fix ω ⊂⊂ Ω and h ∈ R such that
N

|h| < dist (ω, ∂Ω). Then

Z 1
d
f (x + h) − f (x) = f (x + th) dt
0 dt
Z 1
= ∇ f (x + th) h dt,
0

and therefore
14 1 Sobolev spaces
Z Z Z 1
p
|τh f (x) − f (x)| dx ≤ |h| p
|∇ f (x + th)|p dt dx
ω ω 0
Z 1Z
= |h|p |∇ f (x + th)|p dx dt
0 ω
Z 1
p p
≤ |h| k∇ f kLp (ω0 ) dt
0
p
= |h|p k∇ f kLp (ω0 ) ,

where we have set ω0 := ω|h| = {x ∈ RN : dist (x, ω) < |h|}. Note that ω0 ⊂⊂ Ω.
If f ∈ W 1,p (Ω), then, by Friedrichs’ theorem (Theorem 1.17), there exists a
sequence ( fn ) in C∞ p p 0
c (Ω) such that fn → f in L (Ω) and ∇ fn → ∇ f in L (ω ).
From the above inequality we thus obtain

kτh f − f kLp (ω) ≤ |h| k∇ f kLp (ω0 ) ≤ |h| k∇ f kLp (Ω) .

(iii)⇒(ii) Let ϕ ∈ Cc (Ω) and j ∈ {1, . . . , N}. Choose ω ⊂⊂ Ω such that supp ϕ ⊆
ω. Then
Z Z τ ϕ−ϕ
te j
f ∂ j ϕ = lim f
Ω t→0 Ω t
Z τ
−te j − f
f
= lim ϕ.
t→0 Ω t

By assumption, for every t ∈ R \ {0} small enough,

kτ−te j f − f kLp (ω) ≤ C |t|,

and hence, by Hölder’s inequality,

Z
f ∂ j ϕ ≤ C kϕkLp0 .
Ω

Remark 1.19. Theorem 1.18 is not true for p = 1. Only the implications

(i) ⇒ (ii) ⇔ (iii)

remain true. The space

Z
c (Ω)∀ j ∈ {1, . . . , N}
{ f ∈ L1 (Ω) : ∃C ≥ 0∀ϕ ∈ C∞ f ∂ j ϕ ≤ C kϕkL∞ }
Ω

is usually denoted by BV(Ω). Elements of this space are called functions of

bounded variation.
Theorem 1.20 (Composition rule). Let Ω ⊆ RN be open and p ∈ [1, ∞]. Let
g ∈ C1 (R) satisfy g(0) = 0 and kg0 kL∞ =: M < ∞, and let f ∈ W 1,p (Ω). Then the
1.3 Weak derivatives and Sobolev spaces 15

composition g ◦ f belongs to W 1,p (Ω) and

∂ j (g ◦ f ) = (g0 ◦ f ) · ∂ j f.

Proof. The assertion is true for f ∈ C∞ 1,p

c (Ω) and follows for general f ∈ W (Ω)
from an approximation argument using Friedrichs’ theorem (Theorem 1.17).
Note that for every x ∈ Ω and every f ∈ W 1,p (Ω),

|g( f (x))| = |g( f (x)) − g(0)| ≤ M | f (x) − 0| = M | f (x)|,

and therefore g ◦ f ∈ Lp (Ω). Moreover, |(g0 ◦ f ) · ∂ j f | ≤ M |∂ j f |, and hence (g0 ◦

f ) · ∂ j f ∈ Lp (Ω). If f ∈ W 1,p (Ω), then there exists a sequence ( fn ) in C∞
c (Ω) such
that fn → f ∈ Lp (Ω) and ∇ fn → ∇ f in Lp (ω) for every ω ⊂⊂ Ω. Fix ϕ ∈ C∞ c (Ω)
and j ∈ {1, . . . , N}. Then, for every n ∈ N,
Z Z
g ◦ fn ∂ j ϕ = − (g0 ◦ fn ) ∂ j fn ϕ.
Ω Ω

Letting n → ∞, we thus obtain

Z Z
g ◦ f ∂ j ϕ = − (g0 ◦ f ) ∂ j f ϕ,
Ω Ω

and from here and the above estimates follows the claim.

For real functions f , g : Ω → R we set

f ∨ g := sup{ f, g} (pointwise),
f ∧ g := inf{ f, g} (pointwise),
f + := f ∨ 0,
f − := (− f ) ∨ 0, and
| f | := f + + f − .

Corollary 1.21. For every open subset Ω ⊆ RN and every p ∈ [1, ∞], the real space
W 1,p (Ω) is a vector lattice, that is, for every f , g ∈ W 1,p (Ω) one has f + , f − , | f |,
f ∨ g, f ∧ g ∈ W 1,p (Ω). Moreover,

∂ j f + = 1{ f >0} ∂ j f,
∂ j f − = −1{ f <0} ∂ j f,
∂ j | f | = sgn( f ) ∂ j f,
∂ j ( f ∨ g) = 1{ f >g} ∂ j f + 1{ f ≤g} ∂ j g, and
∂ j ( f ∧ g) = 1{ f ≤g} ∂ j f + 1{ f >g} ∂ j g.

Proof. For f + one would like to apply Theorem 1.20 with g given by
16 1 Sobolev spaces

s if s ≥ 0,


g(s) := 
if s < 0.

0

However, since this function is not continuously differentiable, one applies

Theorem 1.20 with gε given by
 ε


s− 2 if s ≥ ε,


if 0 ≤ s < ε,

1 2
gε (s) := 
 2ε s


if s < 0,

0


which is continuously differentiable, and then one passes to the limit ε → 0+.
The other cases follow from this case by noting succesively that f − = (− f )+ ,
| f | = f + + f − , f ∨ g = g + ( f − g)+ and f ∧ g = f − ( f − g)+ .

1.4 Sobolev spaces in one dimension

In this section, let I = (a, b) be an interval in R with −∞ ≤ a < b ≤ ∞.

Lemma 1.22. Let f ∈ L1loc (I) be weakly differentiable with weak derivative f˙ = 0.
Then f is constant.
R R
Proof. Choose ψ ∈ C∞ c (I) such that I ψ = 1, and put c := I f ψ. Then, for every
R
ϕ ∈ C∞c (I) the function ϕ − ( I ϕ) ψ has integral equal to 0 and is therefore the
derivative of an other test function. By definition of the weak derivative and
the assumption, we therefore obtain
Z Z
0 = f (ϕ − ( ϕ)ψ)
ZI ZI Z
= f ϕ − ( ϕ) ( f ψ)
ZI ZI I

= fϕ−c ϕ
ZI I

= ( f − c)ϕ.
I

By Theorem 1.12, this implies f = c, and therefore f is constant.

Theorem 1.23. For every p ∈ [1, ∞] the following assertions are true:
a) Every function in f ∈ W 1,p (I) admits a continuous representative which ex-
tends continuously to the closed interval Ī. If this continuous representative is
denoted by f , too, then for every s, t ∈ I, s < t,
1.4 Sobolev spaces in one dimension 17
Z t
f (t) − f (s) = f˙(τ) dτ.
s

The continuous representative is bounded, and vanishes at +∞ (if b = +∞)

and −∞ (if a = −∞) if p < ∞.
b) The embedding W 1,p (I) → C0 (Ī) is continuous. More precisely,

− 1p
k f kL∞ ≤ (L ∧ 1) k f kW1,p for every f ∈ W 1,p (I),

where L := b − a ∈ (0, ∞] is the length of the interval I.

Rt
Proof. Fix s ∈ I and define g : I → C by s f˙(τ) dτ. It follows from Lebesgue’s
dominated convergence theorem that g is continuous on I and also that
it admits a continuous extension to the closure of I. Moreover, for every
ϕ ∈ C∞c (I),
Z Z bZ t
gϕ̇ = f˙(τ) dτϕ̇(t) dt
I a s
Z sZ s Z bZ t
=− f˙(τ) dτϕ̇(t) dt + f˙(τ) dτϕ̇(t) dt
a t s s
Z sZ τ Z bZ b
=− ϕ̇(t) f˙(τ)dτ + ϕ̇(t) dt f˙(τ) dτ
a a s τ
Z s Z b
=− ϕ(τ) f˙(τ) dτ − ϕ(τ) f˙(τ) dτ
a s
Z
=− f˙ϕ.
I

By definition, g is thus weakly differentiable and ġ = f˙. Hence, by Lemma

1.22, g − f is a constant function. In particular, f = g + c admits a continuous
representative which extends continuously to the closure of I. If we denote
this continuous representative by f , too, then we see from g(s) = 0 that c = f (s),
and therefore, by the definition of g,
Z t
f (t) − f (s) = f˙(τ) dτ.
s

Let us show boundedness of functions in W 1,p (I). By definition, every func-

tion f ∈ W 1,∞ (I) belongs also to L∞ (I) and k f kL∞ ≤ k f kW1,∞ . So consider the
case when f ∈ W 1,p (I) and p < ∞. Let J ⊆ I be an interval of length equal to
L ∧ 1. Then, by the above formula, for every s, t ∈ J, s < t,
18 1 Sobolev spaces
Z t
| f (t)| ≤ | f (s)| + | f˙|
s
1
≤ | f (s)| + |t − s| p0 k f˙kLp (J) .

Integrating this inequality over s ∈ J, we obtain

Z
1+ 1
(L ∧ 1) | f (t)| ≤ | f (s)| + (L ∧ 1) p0 k f˙kLp (J)
J
1 1+ p10
≤ (L ∧ 1) p0 k f kLp (J) + (L ∧ 1) k f˙kLp (J) ,

or
1
− 1p
k f kL∞ (J) ≤ (L ∧ 1) k f kLp (J) + (L ∧ 1) p0 k f˙kLp (J) .

From this inequality one deduces the claim in (b). It remains to show that f
vanishes at +∞ (if b = +∞) and −∞ (if a = −∞) if p < ∞. Assume that b = +∞.
Then, by the preceding inequality (applied with L = 1), for every s ∈ I,
p p p
kfk = k f kLp + k f˙kLp
W 1,p
∞
X
p p
≥ [k f kLp (s+n,s+n+1) + k f˙kLp (s+n,s+n+1) ]
n=0
∞
X
p
≥ k f kL∞ (s+n,s+n+1) ,
n=0

and since the left-hand side is finite, this yields

lim | f (t)| = lim k f kL∞ (s+n,s+n+1) = 0.

t→∞ n→∞

The case a = −∞ is discussed similarly.

Corollary 1.24. Let I ⊆ R be an open interval. A continuous function f : I → R
is Lipschitz continuous if and only if f is weakly differentiable and f˙ ∈ L∞ (I), and
then the Lipschitz constant of f equals k f˙kL∞ .
Proof. Assume first that f˙ ∈ L∞ (I). Then, by Theorem 1.23, for every s, t ∈ I,
s < t, Z t
| f (t) − f (s)| ≤ | f˙(τ)| dτ ≤ k f˙kL∞ |t − s|,
s

and thus f is Lipschitz continuous with Lipschitz constant L = k f˙kL∞ .

Conversely, assume now that f is Lipschitz continuous with Lipschitz
constant L. Let I0 ⊂⊂ I be an open, bounded subinterval. Since the closure
of I0 is compact and contained in I, f is bounded on I0 . Let J ⊂⊂ I0 be a
subinterval, and let h ∈ R be such that |h| < dist (J, ∂I0 ). Then
1.5 The extension property 19

kτh f − f kL∞ (J) = sup | f (t + h) − f (t)| ≤ L |h|.

t∈J

By Theorem 1.18, this implies f ∈ W 1,∞ (I0 ) and k f˙kL∞ (I0 ) ≤ L. Since I0 ⊂⊂ I was
arbitrary, we find k f˙kL∞ (I) ≤ L.

1.5 The extension property

We say that an open subset Ω ⊆ RN has the W k,p extension property if there
exists a bounded linear operator E : W k,p (Ω) → W k,p (RN ) such that, for every
f ∈ W k,p (Ω), E f |Ω = f .
We say that the boundary of an open set Ω ⊆ RN is C1 -regular (and we
write ∂Ω ∈ C1 ) if it is a C1 -manifold. The aim of this section is to prove the
following result.
Theorem 1.25. Let Ω ⊆ RN be an open set with compact, C1 -regular boundary.
Then Ω has the W 1,p extension property for every p ∈ [1, ∞].
The proof of this theorem is based on two lemmas.
Lemma 1.26. The half-space

R+N := {x = (x1 , . . . , xN ) ∈ RN : xN > 0}

has the W 1,p extension property for every p ∈ [1, ∞].

Proof. For every f ∈ W 1,p (R+N ) we define f˜ : RN → C by

 f (x , xN ) if xN > 0,
 0

f˜(x) := 

 f (x0 , −x ) if x < 0,

N N

where x0 = (x1 , . . . , xN−1 ) ∈ RN−1 for x = (x1 , . . . , xN ). Moreover, for every j ∈

{1, . . . , N} we define

(∂ j f )(x , xN ) if xN > 0,

0



g j (x) := 
(−1)δ jN (∂ f )(x0 , −x ) if x < 0,

j N N

where δ jN is the Kronecker delta. Next, we choose a cut-off function η : R → R

such that

η(s) = 1 if |s| ≥ 1,
1
η(s) = 0 if |s| ≤ ,
2
η(s) = η(−s) for every s ∈ R,
20 1 Sobolev spaces

and we put ηε (s) = η(s/ε) for ε > 0. Finally, we define η̃ε (x) := ηε (xN ) (x =
(x0 , xN ) ∈ RN ).
Let ϕ ∈ C∞ c (R ). Then the support of the test function ϕ η̃ε does not in-
N

tersect the hyperplane {xN = 0}. Therefore, by Lebesgue’s dominated con-

vergence theorem, the substitution (x1 , . . . , xN−1 , xN ) 7→ (x1 , . . . , xN−1 , −xN ), the
product rule and the definition of weak derivative, for every j ∈ {1, . . . , N},
Z
f˜∂ j ϕ =
RN
Z
= lim f˜η̃ε ∂ j ϕ
ε→0+ RN
Z Z 
= lim  f (x)η̃ε (x)(∂ j ϕ)(x) dx + f (x)η̃ε (x)(∂ j ϕ)(x , −xN ) dx
0
 
ε→0+ RN R+N
Z + Z 
δ
= lim  f (x)η̃ε (x)∂ j ϕ(x) dx + (−1) jN f (x)η̃ε (x)∂ j ϕ(x , −xN ) dx
0
 
ε→0+ RN R+N
Z + Z
= lim  f (x)∂ j (η̃ε (xN )ϕ(x)) dx + (−1)δ jN f (x)∂ j (η̃ε (x)ϕ(x0 , −xN )) dx

ε→0+ R+N R+N
Z Z 
δ jN
f (x)(∂ j η̃ε )(x)ϕ(x) dx − (−1) f (x)(∂ j η̃ε )(x)ϕ(x , −xN ) dx
0

−
R+N R+ N
 Z Z
δ jN
= lim − (∂ j f )(x)η̃ε (x)ϕ(x) dx − (−1) (∂ j f )(x)η̃ε (x)ϕ(x0 , −xN ) dx


ε→0+ R+ N R+N
Z Z 
δ jN
f (x)(∂ j η̃ε )(x)ϕ(x) dx − (−1) f (x)(∂ j η̃ε )(x)ϕ(x , −xN ) dx
0

−
R N R+ N
Z+ Z
=− ∂ j f (x)ϕ(x) dx − (−1)δ jN ∂ j f (x)ϕ(x0 , −xN ) dx
N
R+ N
R+
Z Z 
δ jN
f (x)(∂ j η̃ε )(x)ϕ(x) dx + (−1) f (x)(∂ j η̃ε )(x)ϕ(x , −xN ) dx .
0
 
− lim 
ε→0+ N
R+ N
R+

If j , N, then ∂ j η̃ε = 0, and thus the function under the lim-sign on the
right-hand side is equal to 0. If j = N, then, for R ≥ 0 large enough (so that
supp ϕ ⊆ B(0, R)),
1.5 The extension property 21
Z Z
lim sup f (x)(∂N η̃ε )(x)ϕ(x) dx + (−1)δ jN f (x)(∂N η̃ε )(x)ϕ(x0 , −xN ) dx
ε→0+ N
R+ N
R+
Z
≤ lim sup { ε <xN <ε,
| f (x)| |η0ε (xN )| |ϕ(x0 , xN ) − ϕ(x0 , −xN )| dx
ε→0+ 2
|x0 |≤R}

kη0 kL∞
Z
≤ lim sup { ε <xN
| f (x)| k∇ϕk∞ 2ε dx
ε→0+ 2
<ε, ε
|x0 |≤R}

= 0.

Hence, the limit on the right-hand side of the display above equals 0 for
every j ∈ {1, . . . , N}. We have thus proved
Z Z
f˜∂ j ϕ = − g jϕ
RN RN

for every j ∈ {1, . . . , N} and every ϕ ∈ C∞ N ˜ 1,p N

c (R ). In other words, f ∈ W (R )
˜ ˜
and ∂ j f = g j for every j ∈ {1, . . . , N}. By definition, f is an extension of f , and
k f˜kW1,p (RN ) = 2 k f kW1,p (RN ) . Thus E : W 1,p (R+N ) → W 1,p (RN ), f 7→ f˜ is a desired
+
extension operator.

Lemma 1.27 (Partition of unity). Let K ⊆ RN be a compact set, and let (Ui )1≤i≤n
be a finite open covering of K. Then there exists a finite family (ϕi )0≤i≤1 in C∞ (RN )
such that

0 ≤ ϕi ≤ 1 for every 0 ≤ i ≤ n,
Xn
ϕi = 1 on RN ,
i=0
supp ϕi ⊆ Ui for every 1 ≤ i ≤ n, and
supp ϕ0 ⊆ RN \ K.

We call the family (ϕi )0≤i≤n a partition of unity subordinate to the covering
(Ui )1≤i≤n .

Proof. Choose δ > 0, and define, for every i ∈ {1, . . . , n},

Vi := {x ∈ Ui : dist (x, ∂Ui ) > δ}.

Then Vi is open and a subset of Ui . A simple compactness argument shows

that for δ > 0 small enough, (Vi )1≤i≤n is a finite covering of K, too. Then we
define
22 1 Sobolev spaces

A1 := V̄1 ,
i−1
[
Ai := V̄i \ A j for every 2 ≤ i ≤ n, and
j=1
n
[
A0 := RN \ A j.
j=1

δ
Then Ai2 ⊆ Ui for every i ∈ {1, . . . , n}, the Ai are mutually disjoint, and
Sn R
i=0 Ai = R . Let ϕ ∈ Cc (R ) be a positive test function such that RN ϕ = 1
N ∞ N

and supp ϕ ⊆ B(0, 4δ ), and put ϕi := 1Ai ∗ ϕ (0 ≤ i ≤ n). Then ϕi ∈ C∞ (RN ) by the
smoothin effect of the convolution (Theorem 1.9), supp ϕi ⊆ Ui by construc-
tion, and
X n Xn
ϕi = ( 1Ai ) ∗ ϕ = 1Sn Ai ∗ ϕ = 1 ∗ ϕ = 1.
i=0
i=0 i=0

Proof (of Theorem 1.25). Let ((Ui , ψi ))1≤i≤n be a finite atlas of ∂Ω, where Ui are
open sets and ψi : Ui → RN are C1 -diffeomorphisms from Ui onto Vi := ψi (Ui )
such that ψi (Ui ∩ Ω) ⊆ R+N , ψi (Ui ∩ ∂Ω) ⊆ {x ∈ RN : xN = 0} and ψi (Ui ∩ Ω̄c ) ⊆
{x ∈ RN : xN < 0}. We can always find such a finite atlas by definition of a
C1 -manifold and since ∂Ω is compact. Moreover, without loss of generality,
we may assume that the sets Vi are symmetric with respect to the hyperplane
{xN = 0}.
Let (ϕi )0≤i≤n be a partition of unity subordinate to the covering (Ui )1≤i≤n .
Given f ∈ W 1,p (Ω), we put fi := ( f ϕi )|Ω (0 ≤ i ≤ 1). Then fi ∈ W 1,p (Ω) and

supp fi ⊆ supp ϕi ∩ Ω ⊆ Ui ∩ Ω.

The function f0 has compact support in Ω and may thus be extended by

0 to RN . We denote this extension by f˜0 and note that f˜0 ∈ W 1,p (RN ) and
k f˜0 kW1,p (RN ) = k f0 kW1,p (Ω) .
For 1 ≤ i ≤ N, the functions fi ◦ ψ−1
i
belong to W 1,p (R+N ) and have support
in Vi ∩ R+ . They may thus be extended to functions in W 1,p (RN ) by reflection
N

(see Lemma 1.26 and its proof). We denote these extensions by f ] ◦ ψ−1 and
i i
note that these extensions have support in Vi (here we use that the Vi are
symmetric with respect to the hyperplane {xN = 0}). Now put f˜i := fi]
◦ ψ−1
i
◦
˜
ψi . Then fi ∈ W (R ) is an extension of fi . Now, it suffices to define E :
1,p N

W 1,p (Ω) → W 1,p (RN ) by

n
X
E f := f˜i
i=0

and to verify that E is a desired, bounded extension operator.

1.6 The Sobolev embedding theorems 23

1.6 The Sobolev embedding theorems

In the following, given two Banach spaces X and Y, we write X ,→ Y if X ⊆ Y

and if the identity map i : X → Y, x 7→ x is continuous.

Theorem 1.28 (Sobolev-Gagliardo-Nirenberg). Fix p ∈ [1, N) and define p∗ :=

Np 1,p N p∗ N
N−p . Then every f ∈ W (R ) belongs to L (R ) and

k f kLp∗ ≤ C k∇ f kLp
∗
In particular, W 1,p (RN ) ,→ Lp (RN ).

For the proof of this theorem, we need the following technical lemma.

Lemma 1.29. Let N ≥ 2 and f1 , . . . , fN ∈ LN−1 (RN−1 ). For every 1 ≤ i ≤ N and

every x ∈ RN we put x̃i := (x1 , . . . , xi−1 , xi+1 , . . . , xN ). Then the function
N
Y
f (x) := fi (x̃i )
i=1

belongs to L1 (RN ) and

N
Y
k f kL1 (RN ) ≤ k fi kLN−1 (RN−1 ) .
i=1

Proof (incomplete). The assertion of this lemma is straightfoward to verify for

N = 2. Now suppose that the assertion is true for some N, and let us prove
that it is true for N + 1, too. In the case N + 1, fix xN+1 . We have, by Hölder’s
inequality,
Z N
 N−1
Z  Y N  N
| f (x1 , . . . , xN+1 )| dx1 . . . dxN ≤ k fN+1 kLN (RN ) 
 | fi (x̃i )| N−1 dx1 . . . dxN  .
RN RN i=1

By assumption, fi ∈ LN (RN ), that is,

Z
| f (x1 , . . . , xi−1 , xi+1 , . . . , xN+1 )|N dx1 . . . dxi−1 dxi+1 . . . dxN+1 < ∞.
RN

In particular, for almost every xN+1 ∈ R,

Z
| f (x1 , . . . , xi−1 , xi+1 , . . . , xN+1 )|N dx1 . . . dxi−1 dxi+1 . . . dxN < ∞ for every 1 ≤ i ≤ N.
RN−1

Hence, by induction hypothesis, for almost every xN+1 ∈ R,

24 1 Sobolev spaces
Z N N 1
N
Y Y
| fi (x̃i )| N−1 dx1 . . . dxN ≤ N .
k fi k N−1
L
RN i=1 i=1

N
or, equivalently, | fi | N−1 ∈ LN−1 (RN ).

Proof (of Theorem 1.28). Assume first that p = 1 and f ∈ C∞ N

c (R ). Then
Z x1 Z ∞
| f (x)| = ∂1 f (ξ, x2 , . . . , xN ) dξ ≤ |∂1 f (ξ, x2 , . . . , xN )| dξ.
−∞ −∞

Similarly,
Z ∞
| f (x)| ≤ |∂i f (x1 , . . . , xi−1 , ξ, xi+1 , . . . , xN )| dξ =: fi (x̃i ),
−∞

where x̃i := (x1 , . . . , xi−1 , xi+1 , . . . , xN ). Hence,

N
Y
| f (x)|N ≤ fi (x̃i ).
i=1

1
Note that fi N−1 ∈ LN−1 (RN−1 ) and

Z ! 1
1 N−1
k fi N−1
kLN−1 (RN−1 ) = fi (x̃i ) dx̃i
RN−1
Z Z ! 1
N−1
= |∂i f (x1 , . . . , xi−1 , ξ, xi+1 , . . . , xN )| dξ dx̃i
RN−1 R
1
= k∂i f k N−1
1 N
.
L (R )

As a consequence, by Lemma 1.29,

Z N 1
N
Y
| f (x)| N−1 dx ≤= k∂i f k N−1
1 N
,
L (R )
RN i=1

or
N
Y 1
kfk N ≤ k∂i f k N1 .
L N−1 L (RN )
i=1

Applying this formula to | f |t−1 f with t ≥ 1, we obtain

1.6 The Sobolev embedding theorems 25

N
Y 1
k f kt tN ≤ kt | f |t−1 ∂i f k N1
L N−1 L (RN )
i=1
N
Y 1 1
≤t k| f |t−1 k N p k∂i f k Np (1.2)
L (RN )
i=1 L p−1
N
Y 1
= t k f kt−1
(t−1)p k∂i f k Np
L (RN )
L p−1 i=1

(t−1)p
Choose t such that tN
N−1 = p−1 , that is, t = N−1 ∗
N p . Then

N
Y 1
k f kLp∗ ≤ C k∂i f k Np
L (RN )
i=1

for some constant C which depends only on t, and from here follows imme-
diately
k f kLp∗ ≤ C k∇ f kLp (RN ) .

The claim for general f ∈ W 1,p (RN ) follows from this inequality and an ap-
proximation argument using Theorem 1.15.
Corollary 1.30. For every p ∈ [1, N) and every q ∈ [p, p∗ ] one has

W 1,p (RN ) ,→ Lq (RN ).

Proof. By definition of the Sobolev spaces, W 1,p (RN ) ,→ Lp (RN ), and by The-
∗
orem 1.28, W 1,p (RN ) ,→ Lp (RN ). The interpolation inequality (a straightfor-
ward application of Hölder’s inequality) yields

Lp ∩ Lp∗ (RN ) ,→ Lq (RN ) for every q ∈ [p, p∗ ],

and putting the embeddings together yields the claim.

Corollary 1.31 (The limit case p = N). For every q ∈ [N, ∞) one has

W 1,N (RN ) ,→ Lq (RN ).

c (R ). We apply (1.2) with p = N and obtain

Proof. Let f ∈ C∞ N

k f kt tN (t−1)N k∇ f kLN (RN ) .

≤ t k f kt−1
L N−1 L N−1

By Young’s inequality,

kfk tN ≤ C [k f k (t−1)N + k∇ f kLN (RN ) ]. (1.3)

L N−1 L N−1

Choosing t = N in this inequality, we obtain

26 1 Sobolev spaces

kfk N2 ≤ C k f kW1,N ,
L N−1

and thus, by interpolation,

N2
k f kLq ≤ C k f kW1,N for every q ∈ [N, ]
N−1
Iterating this argument by choosing now t = N + 1, t = N + 2, . . . in (1.3), we
obtain
k f kLq ≤ C k f kW1,N for every q ∈ [N, ∞).
This inequality holds, by an approximation argument, finally for all f ∈
W 1,N (RN ).

Theorem 1.32 (Morrey). For every p ∈ (N, ∞),

W 1,p (RN ) ,→ L∞ (RN ).

More precisely, for α := 1 − Np there exists a constant C = C(p, N) ≥ 0 such that for
every f ∈ W 1,p (RN ), x, y ∈ RN ,

| f (x) − f (y)| ≤ C |x − y|α k∇ f kLp .

Proof. Let Q be cube with sides parallel to the coordinate axes and of length
r > 0, such that 0 ∈ Q. For every x ∈ Q one has
Z 1
d
| f (x) − f (0)| = f (tx) dt
0 dt
Z 1
≤ |∇ f (tx) · x| dt
0
Z 1
≤r |∇ f (tx)| dt.
0
R
Put f¯ := |Q|
1
Q
f . By integrating the above inequality over x ∈ Q, and by using
Hölder’s inequality, we obtain
Z Z 1
r
| f¯− f (0)| ≤ |∇ f (tx)| dt dx
|Q| Q 0
Z 1Z
1
= |∇ f (tx)| dx dt
rN−1 0 Q
Z 1Z
1 dt
= |∇ f (x)| dx
rN−1 0 tQ tN
1.6 The Sobolev embedding theorems 27

Z 1 Z ! 1p
1 p−1 dt
≤ |∇ f |p |tQ| p
rN−1 0 tQ tN
Z 1
1− N −N
≤r p k∇ f kLp (Q) t p dt
0
1− N
r p
= k∇ f kLp (Q) .
1 − Np

Clearly, this inequality holds also with x ∈ RN instead of 0 (and with a cube
Q containing x). Hence, for every cube Q with sides parallel to the coordinate
axes and of length r > 0, and for every x, y ∈ Q,

1− N
2r p
| f (x) − f (y)| ≤ k∇ f kLp (Q) .
1 − Np

Since we may choose Q minimal (so that, for example, r = 2 |x − y|), and by an
approximation argument (Theorem 1.15), we obtain the Hölder continuity of
f ∈ W 1,p (RN ). Moreover, applying again the above inequality with 0 replaced
by x ∈ RN and with a cube of side length r = 1, we obtain

| f (x)| ≤ | f¯| + | f¯− f (x)|

1
≤ k f kLp (Q) + k∇ f kLp (Q)
1 − Np
≤ C k f kW1,p (RN ) ,

and from here and an approximation argument, we obtain the boundedness

of f ∈ W 1,p (RN ).

Summarizing the results of this section, we are in the position to state the
first version of the Sobolev embedding theorem. For an open set Ω ⊆ RN ,
m ∈ N0 and θ ∈ [0, 1] we set

| f (x) − f (y)|
C0,θ (Ω) := { f ∈ C(Ω) : sup < ∞,
x,y∈Ω |x − y|θ
x,y

and

Cm,α (Ω) := { f ∈ Cm (Ω) : ∂α f ∈ C0,θ (Ω) for all α ∈ NN

0 , |α| = m}.

Theorem 1.33 (Sobolev embedding theorem). Let k ∈ N and p ∈ [1, ∞). Then
the following assertions are true:
a) If kp < N, then
28 1 Sobolev spaces

Np
W k,p (RN ) ,→ Lq (RN ) for every q ∈ [p, ].
N − kp

b) If kp = N, then

W k,p (RN ) ,→ Lq (RN ) for every q ∈ [p, ∞).

c) If kp > N, then
W k,p (RN ) ,→ L∞ (RN ) ∩ Cm,α (RN ),
where 
N
[k − p ]


 if k − Np < N,
m := 
k − Np − 1 if k − Np < N,



and θ = k − Np − m.

Proof. For k = 1, the assertions (a)-(c) follow respectively from Corollary

1.30, Corollary 1.31 and Theorem 1.32. For k ≥ 2, the assertions follow by an
iteration (or induction) argument.
Corollary 1.34 (Sobolev embedding theorem). Let k ∈ N, p ∈ [1, ∞) and let
Ω ⊆ RN be an open set which has the W 1,q -extension property for every q ∈ [1, ∞) (for
example, Ω has compact, C1 regular boundary ∂Ω). Then the following assertions
are true:
a) If kp < N, then

Np
W k,p (Ω) ,→ Lq (Ω) for every q ∈ [p, ].
N − kp

b) If kp = N, then

W k,p (Ω) ,→ Lq (Ω) for every q ∈ [p, ∞).

c) If kp > N, then
W k,p (Ω) ,→ L∞ (Ω) ∩ Cm,α (Ω),
where 
N
[k − p ]


 if k − Np < N,
m := 
k − Np − 1 if k − Np < N,



and θ = k − Np − m.

These assertions (a)-(c) remain true for arbitrary open Ω ⊆ RN if the spaces W k,p (Ω)
k,p
are replaced by W0 (Ω).
Proof. For k = 1, the assertions follow from the assumption that Ω has the
W 1,q -extension property for every q ∈ [1, ∞) and from the Sobolev embedding
1.6 The Sobolev embedding theorems 29

theorem for the W k,p (RN )-spaces (Theorem 1.33). For k ≥ 2, the assertions
follow by an iteration (or induction) argument.
If Ω ⊆ RN is an arbitrary open set, one may argue similarly by noting
k,p
that functions in W0 (Ω) may be extended by 0 outside Ω to functions in
W k,p (RN ). In fact, for test functions in C∞
c (Ω) this is obvious, and for general
k,p
functions in W0 (Ω) this follows by an approximation argument using the
k,p
definition of W0 (RN ).

Lemma 1.35. For every open, bounded subset Ω ⊆ RN , every p ∈ [1, ∞] and every
ϕ ∈ L1 (RN ) the operator

Tϕ : Lp (RN ) → Lp (Ω),
f 7→ ( f ∗ ϕ)|Ω ,

is compact.

Proof. For ϕ ∈ Cc (RN ) the assertions follows from the theorem of Arzelà-
Ascoli. For general ϕ ∈ L1 (RN ) the assertion follows by approximation of ϕ,
Young’s inequality (Theorem 1.4) and the fact that the compact operators
form a closed subspace of the space of bounded linear operators. In fact, if
ϕ ∈ L1 (RN ), then we find a sequence (ϕn ) in C∞ N
c (R ) such that limn→∞ kϕ −
ϕn kL1 = 0 (Theorem 1.11). As a consequence

kTϕ − Tϕn k = sup kTϕ f − Tϕn f kLp (Ω)

k f kLp (Ω) ≤1

≤ sup k f ∗ ϕ − f ∗ ϕn kLp (RN )

k f kLp (Ω) ≤1

≤ sup k f kLp (RN ) kϕ − ϕn kL1 (RN )

k f kLp (Ω) ≤1

→ 0 as n → ∞.

Since the Tϕn are compact and converge in operator norm to Tϕ , Tϕ is com-
pact, too.

Lemma 1.36. For every f ∈ C∞c (Ω) one has

Z
−1 y
f (0) = ∇ f (y) N dy.
σN−1 RN |y|

Proof.

Theorem 1.37 (Rellich-Kondrachev). Let k ∈ N, p ∈ [1, ∞) and let Ω ⊆ RN be

a bounded, open set. Then the following assertions are true:
a) If kp < N, then
30 1 Sobolev spaces

k,p c Np
W0 (Ω) ,→ Lq (Ω) for every q ∈ [p, ).
N − kp

b) If kp = N, then
k,p c
W0 (Ω) ,→ Lq (Ω) for every q ∈ [p, ∞).

c) If kp > N, then
k,p c 0
W0 (Ω) ,→ L∞ (Ω) ∩ Cm,θ (Ω̄),
where 
N
[k − p ]


 if k − Np < N,
m := 
k − Np − 1 if k − Np < N.



and θ0 ∈ (0, k − Np − m).

These assertions (a)-(c) remain true for bounded, open sets Ω ⊆ RN which have the
k,p
W 1,q -extension property for every q ∈ [1, ∞), and then the spaces W0 (Ω) may be
replaced by W k,p (Ω).

Proof. First assume that k = 1. Choose R > 0 such that Ω ⊆ B(0, R) and define
ϕ ∈ L1 (RN ) by
1 x
ϕ(x) := 1B(0,2R) (x) (x ∈ RN ).
σN−1 |x|N
Let f ∈ C∞ ∞ N
c (Ω) ⊆ Cc (R ). By Lemma 1.36 (applied to the function f (x − ·)),
for every x ∈ Ω,
Z
1 y
f (x) = ∇ f (x − y) dy
σN−1 RN |y|
Z
1 y
= ∇ f (x − y) dy
σN−1 B(0,2R) |y|
= ∇ f ∗ ϕ(x).

Hence, the continuous operator

1,p
W0 (Ω) → Lp (Ω; RN ) → Lp (Ω),
f 7→ ∇f 7→ (∇ f ∗ ϕ)|Ω ,

coincides with the identity on the space C∞

c (Ω). This space being dense in
1,p
W0 (Ω) (by definition), the above operator thus is the canonical embedding
1,p
of W0 (Ω) into Lp (Ω). However, the second operator is compact by Lemma
1.35, and thus the embedding itself is compact.
 1,p
1.7 The space W0 (Ω) and the Poincaré inequality 31

Np θ
Next, assume p < N, and let q ∈ [p, p∗ ), where p∗ = N−p . Then 1
q = p +
1−θ
p∗ for some θ ∈ (0, 1]. Let ( fn ) be a bounded sequence in Then, W 1,p (Ω).
by the preceding step, there exists a subsequence (denoted again by ( fn ))
which converges in Lp (Ω) to some element in f . Moreover, by the Sobolev
embedding theorem (Corollary 1.34 (a)), the sequence ( fn ) is bounded in
Lp∗ (Ω). Hence, by the interpolation inequality, and since θ > 0,

k fn − f kLq ≤ k fn − f kθLp k fn − f k1−θ

p∗ → 0,
L

that is, ( fn ) converges also in Lq (Ω).

Since the bounded sequence ( fn ) was
1,p
arbitrary, we have proved that the embedding W0 (Ω) ,→ Lq (Ω) is compact.
The case p = N is treated similarly, and in the case p > N one uses the case (c)
of the Sobolev embedding theorem (Corollary 1.34) and Arzelà-Ascoli.
The case k ≥ 2 follows by induction.
If Ω ⊆ RN is bounded, open and has the W 1,p (Ω)-extension property, and
k,p
if we replace the spaces W0 (Ω) by W k,p (Ω), then one may argue as follows.
Again, we consider first the case k = 1. We choose a test function ψ ∈ C∞ c (R )
N

such that ψ = 1 on Ω. Moreover, we choose R > 0 such that supp ψ ⊆ B(0, R).
Let E : W 1,p (Ω) → W 1,p (RN ) be any extension operator, and consider then
1,p
the extension operator Ẽ : W 1,p (Ω) → W0 (B(0, R)) given by Ẽ f := (ψ · E f )|Ω .
1,p 1,p
Using now the first step (with W0 (Ω) replaced by W0 (B(0, R))), we obtain
the claim for k = 1. The case k ≥ 2 follows again by induction.

Remark 1.38. Note that the range of possible exponents q in Theorem 1.37
(a) differs from the range of possible exponents in Corollary 1.34 (a): the limit
case q = p∗ is excluded. In fact, the embedding
1,p ∗
W0 (Ω) ,→ Lp (Ω)

is continuous, but in general not compact.

1,p
1.7 The space W0 (Ω) and the Poincaré inequality

Theorem 1.39 (Poincaré inequality). Let Ω ⊆ RN be contained in a strip of the

form
S := RN−1 × (a, b),
and let p ∈ [1, ∞]. Then

1 p p 1,p
kukLp ≤ k∇ukLp for every u ∈ W0 (Ω).
(b − a)p
32 1 Sobolev spaces

Proof. Assume first that u ∈ C∞ N

c (Ω). We extend u by 0 to R and denote this
extension by u, too. Then, for every x = (x̄, xN ) ∈ Ω (x̄ ∈ R N−1 , xN ∈ R),

u(x) = u(x̄, xN ) − u(x̄, a)

Z xN
= ∂N u(x̄, ξ) dξ.
a

As a consequence, if p < ∞, then

Z Z Z b
|u(x)| dx =
p
|u(x̄, xN )|p dxN dx̄
Ω RN−1 a
Z Z b Z xN !p
≤ |∂N u(x̄, ξ)| dξ dxN dx̄
RN−1 a a
Z Z b Z b
≤ (b − a) p−1
|∂N u(x̄, ξ)|p dξ dxN dx̄
RN−1 a a
Z Z b
= (b − a)p |∂N u(x̄, ξ)|p dξ dx̄
R N−1 a
Z
= (b − a) p
|∂N u(x)|p dx
ZΩ
p
≤ (b − a) |∇u|p dx.
Ω

This inequality and an approximation argument yield the claim.

Chapter 2
Elliptic equations

2.1 The Lax-Milgram lemma

Let V be a Hilbert space. A sesquilinear form a : V × V → C is coercive if there

exists η > 0 such that

Re a(u, u) ≥ η kuk2V for every u ∈ V.

If H is a second Hilbert space such that V ,→ H with dense, continuous

embedding, then the form a is called H-elliptic (or shortly elliptic if the
Hilbert space H is clear from the context) if there exist η > 0 and ω ∈ R such
that
Re a(u, u) + ω kuk2H ≥ η kuk2V for every u ∈ V,
Clearly, a coercive sesquilinear form is H-elliptic for every Hilbert space H
into which V is continuously and densely embedded.
Lemma 2.1 (Lax-Milgram). Let a : V ×V → C be a continuous, coercive, sesquilin-
ear form on a Hilbert space V. Then, for every continuous, antilinear f : V → C
there exists a unique element u ∈ V such that

a(u, v) = h f, viV0 ,V for all v ∈ V.

If V is a real Hilbert space, then “sesquilinear” and “antilinear” should be replaced

by “bilinear” and “linear”, respectively. Moreover, if V is real and if a is in addition
symmetric, then the unique element u ∈ V above is characterised by the equality

1 1
a(u, u) − h f, uiV0 ,V = min a(v, v) − h f, viV0 ,V (2.1)
2 v∈V 2

Proof. For every u ∈ V we denote by Au the continuous, antilinear form

V → C, v 7→ a(u, v). We thus obtain a continuous, linear operator A : V → V 0
(where V 0 is in this proof the space of all continuous, antilinear forms on V).
We have to show that A is bijective.

33
34 2 Elliptic equations

The operator A is injective and has closed range. By coercivity of a, for every
v ∈ V,

η kvk2V ≤ Re a(v, v)
≤ |a(v, v)|
= |hAv, viV0 ,V |
≤ kAvkV0 kvkV ,

which implies
η kvkV ≤ kAvkV0 for every v ∈ V.
As a consequence of this inequality, A is injective and has closed range.
The operator A is surjective. Since A has closed range, it suffices to show
that A has dense range in V 0 . If the range of A is not dense in V 0 , then, by the
Hahn-Banach theorem and by reflexivity of V, there exists u ∈ V \ {0} such
that
hAv, uiV0 ,V = a(v, u) = 0 for every v ∈ V.
Choosing v = u in this equality, we obtain a contradiction to coercivity of a.
Hence, the range of A is dense in V 0 which, together with the preceding step,
yields that A is surjective.
Now let us assume that V is a real Hilbert space and that a is symmetric.
Let u ∈ V be such that a(u, v) = h f, viV0 ,V for every v ∈ V. Then, for every v ∈ V,

1 1 1
a(u + v, u + v) − h f, u + viV0 ,V = a(u, u) + a(u, v) + a(v, v) − h f, uiV0 ,V − h f, viV0 ,V
2 2 2
1 1
= a(u, u) + a(v, v) − h f, uiV0 ,V
2 2
1
≥ a(u, u) − h f, uiV0 ,V ,
2
which yields (2.1).
Conversely, assume that (2.1) holds. Then, for every v ∈ V,

1 1
a(u, u) − h f, uiV0 ,V ≤ a(u + v, u + v) − h f, u + viV0 ,V ,
2 2
which implies

1
0 ≤ a(u, v) + a(v, v) − h f, viV0 ,V for every v ∈ V.
2
Replacing v by tv with t > 0, diving the resulting inequality by t, and letting
t → 0+, we deduce

0 ≤ a(u, v) − h f, viV0 ,V for every v ∈ V.

By noting that this inequality holds both for v and −v, we conclude that
2.1 The Lax-Milgram lemma 35

0 = a(u, v) − h f, viV0 ,V for every v ∈ V.

Proof (Second proof of Lemma 2.1 under the additional assumption that V is sep-
arable). Existence. Let (Vn ) S
be an increasing sequence of finite dimensional
subspaces of V such that n Vn is dense in V. For the existence of such a
sequence, we use separability of V; for example, if (wk ) is a dense sequence
in V, then one may choose Vn := span {wk : 1 ≤ k ≤ n}.
For each n we consider the finite dimensional, linear problem of finding
un ∈ Vn such that
a(un , v) = f (v) for every v ∈ Vn . (2.2)
In order to translate this into a linear problem in Cm (m = dim Vn ), one
may choose a basis (bi )1≤i≤m of Vn , define the coefficients βi := f (bi ) ∈ C and
αi j := a(bi , b j ) ∈ C (1 ≤ i, j ≤ m), and write un = i ξi bi with ξi ∈ C. Then the
P
problem (2.2) is equivalent to the problem
m
X
αi j ξi = β j for every 1 ≤ j ≤ m.
i=1

If un ∈ Vn is a solution of (??), then the coercivity of a yields

k f kV0 kun kV ≥ Re f (un )

= Re a(un , un )
= η kun k2V ,

or, equivalently,
1
kun kV ≤ k f kV 0 . (2.3)
η
This inequality shows first injectivity of the underlying linear operator, which
together with the fact that Vn is finite dimensional implies bijectivity of
the underlying linear operator. Hence, the problem (2.2) admits a unique
solution un ∈ Vn . However, inequality 2.3 also implies that the resulting
sequence (un ) is bounded in V. Since V is reflexive, there exists a subsequence
of (un ) (which we denote again by (un )) which converges weakly to some
element u ∈ V. In particular, for the continuous, linear functionals a(·, v) : V →
C (v ∈ V) one has a(un , v) → a(u, v) (n → ∞). However, for all v ∈ Vk and all
n ≥ k one has a(un , v) = f (v). Hence

a(u, v) = f (v) for all v ∈ Vn and alln.

S
Since n Vn is dense in V and since a and f are continuous, we obtain

a(u, v) = f (v) for all v ∈ V.

36 2 Elliptic equations

Uniqueness. Uniqueness of a solution u ∈ V of the above equation follows

again from coercivity of a, similarly as in the finite dimensional case.

2.2 The Laplace operator

In this section we consider the problem

λu − ∆u = f in Ω,
(2.4)
u = 0 on ∂Ω,

where Ω ⊆ RN is an open set, λ ∈ R, f : Ω → C is a given function, u : Ω → C

is the unknown function, and
N
X ∂2
∆ :=
i=1
∂x2i

is the Laplace operator. While the first line in (2.4) is a partial differential
equation in which the unknown function u and its partial derivatives (here,
the second, not mixed partial derivatives) appear, the second line in (2.4) is
a boundary condition. It is called (homogeneous) Dirichlet boundary condi-
tion.
Note that if u ∈ C2 (Ω) ∩ C(Ω̄) is a classical solution of (2.4), that is, u
satisfies (2.4) in the usual sense (using classical partial derivatives), then we
may multiply the first line in (2.4) by the complex conjugate of a test function
ϕ ∈ C∞c (Ω) and integrate over Ω. An integration by parts then yields
Z Z Z
f ϕ̄ = λ uϕ̄ − ∆uϕ̄
Ω Ω Ω
N
∂2 u
Z Z X
=λ uϕ̄ − ϕ̄
Ω i=1 ∂xi
2
Ω
N
∂u ∂ϕ
Z Z X
=λ uϕ̄ −
Ω Ω i=1 ∂x i ∂xi
Z Z
=λ uϕ̄ − ∇u∇ϕ.
Ω Ω

Given f ∈ L2 (Ω), we now call a function u ∈ H01 (Ω) a weak solution of (2.4) if
Z Z Z
λ uϕ̄ − ∇u∇ϕ = f ϕ̄ for every ϕ ∈ H01 (Ω). (2.5)
Ω Ω Ω
2.2 The Laplace operator 37

Theorem 2.2 (The Laplace operator with Dirichlet boundary conditions).

Let λ1 := λ1 (Ω) be the Poincaré constant of the set Ω, that is, the optimal (largest)
constant λ ≥ 0 such that
Z Z
λ |u|2 ≤ |∇u|2 for every u ∈ H01 (Ω).
Ω Ω

Then, for every λ > −λ1 and every f ∈ L2 (Ω) the problem (2.4) admits a unique
weak solution u ∈ H01 (Ω). For this weak solution we have the estimate

1
kukL2 ≤ k f kL2 .
λ + λ1

Proof. Consider on the Hilbert space H01 (Ω) the sesquilinear form a : H01 (Ω) ×
H01 (Ω) → C given by
Z Z
a(u, v) := λ uv̄ − ∇u∇v (u, v ∈ H01 (Ω)).
Ω Ω

Then a is continuous since for every u, v ∈ H01 (Ω), by the Cauchy-Schwarz

inequality,

|a(u, v)||λ| kukL2 kvkL2 + k∇ukL2 kvkL2

(1 + |λ|) kukH1 kvkH1 .
0 0

We show that a is also coercive. Choose first ε > 0 such that λ + λ1 > ε, and
choose next µ ∈ (0, 1] such that ε − µλ1 > 0. Then, for every u ∈ H01 (Ω), by the
definition of λ1 ,
Z Z
a(u, u) = λ |u| +
2
|∇u|2
Ω Ω
Z Z
= (λ + λ1 − ε) |u| − (λ1 − ε)
2
|u|2
Ω Ω
Z Z
+ (1 − µ) |∇u| + µ
2
|∇u|2
Ω Ω
Z Z Z
≥ (λ + λ1 − ε) |u| + (ε − µλ1 )
2
|u| + µ
2
|∇u|2
Ω Ω Ω
≥ ηkuk2 1 ,
H0

where η = min{λ + λ1 − ε, µ} > 0. Hence, a is coercive.

R
Consider next the mapping ` : H01 (Ω) → C, v 7→ Ω f v̄, which is well defined
and continuous by the Cauchy-Schwarz inequality, and antilinear. Existence
and uniqueness of a weak solution of (2.4) thus follows from the Lax-Milgram
lemma (Lemma 2.1) applied to a and `. For the estimate, note that we have,
38 2 Elliptic equations

by the Cauchy-Schwarz inequality,

Z
k f kL2 kukL2 ≥ fu
Ω
Z Z
=λ |u| + 2
|∇u|2
Ω Ω
Z
≥ (λ + λ1 ) |u|2
Ω
= (λ + λ1 ) kuk2L2 .

Theorem 2.3 (The limit case λ = −λ1 ). Assume that Ω ⊆ RN is open and
bounded. Then the problem

−λ1 u − ∆u = 0 in Ω,
(2.6)
u = 0 on ∂Ω,

admits a weak solution u ∈ H01 (Ω) which is not = 0.

Proof. By definition of λ1 ,
R
|∇u|2
Z
Ω
λ1 = inf R = inf |∇u|2 .
u∈H1 (Ω)
0 Ω
|u|2 u∈H1 (Ω)
0 Ω
u,0 kuk 2 =1
L

By definition of the infimum, there exists a sequence (un ) in H01 (Ω) such that
Z
|∇un |2 → λ1 as n → ∞, and
Ω
kun kL2 = 1 for every n.

This sequence is thus bounded in H01 (Ω). By the theorem of Rellich-

Kondrachev (Theorem 1.37), there exists a subsequence of (un ) (which we
denote again by (un )) and u ∈ L2 (Ω) such that

un → u in L2 (Ω).

Moreover, since H01 (Ω) is reflexive, there exists a further subsequence

(again denoted by (un )) which converges weakly in H01 (Ω) to some element
v ∈ H01 (Ω). Since weak convergence in H01 (Ω) implies weak convergence in
L2 (Ω), since strong convergence implies weak convergence, and since weak
limits are unique, we obtain u = v ∈ H01 (Ω). By the geometric
R version of the
2
Hahn-Banach theorem, and since the function v 7→ Ω |∇v| is convex and
continuous,
2.2 The Laplace operator 39
Z Z
|∇u|2 ≤ lim inf |∇un |2 .
Ω n→∞ Ω

Moreover, by the strong convergence in L2 (Ω),

kukL2 = lim kun kL2 = 1.

n→∞

In particular, u , 0. The preceding two (in-)equalities, the definition of λ1 ,

and the choice of the sequence (un ) imply
Z
λ1 = λ1 |u|2
Ω
Z
≤ |∇u|2
Ω
Z
≤ lim inf |∇un |2
n→∞ Ω
= λ1 .

In particular, the inequality signs in this chain of inequalities can be replaced

by equality signs. This means that
Z Z
|∇u| − λ1
2
|u|2 = 0.
Ω Ω
R
In other words, u is a global minimizer of the (positive) function v 7→ Ω |∇v|2 −
R
λ1 Ω |v|2 . Proceeding now like in the second step of the proof of the lemma
of Lax-Milgram (Lemma 2.1), we deduce from this
Z Z
∇u∇v − λ1 uv̄ = 0 for every v ∈ H01 (Ω).
Ω Ω

In other words, u is a weak solution of (2.6).

Theorem 2.4 (The case N = 1). Let Ω = (a, b) be an interval in R (−∞ ≤ a < b ≤
∞). Then every weak solution u ∈ H01 (a, b) of (2.4) (λ ∈ R, f ∈ L2 (a, b)) belongs to
H2 (a, b) ∩ H01 (a, b) and λu − u00 = f . Moreover, u belongs to C0 (a, b) which means
in particular that u admits a continuous extension to a and b (if they are finite) and
u(a) = u(b) = 0.

Proof. Since u is a weak solution,

Z b Z b Z b
f v̄ = λ uv̄ + u0 v̄0 for every H01 (a, b).
a a a

In particular, since C∞ 1
c ((a, b)) ⊆ H0 (a, b),
40 2 Elliptic equations
Z b Z b
u ϕ =−
0 0
(λu − f )ϕ for every ϕ ∈ C∞
c ((a, b)).
a a

By definition of the weak derivative, and since u − f ∈ L2 (a, b), this equality
implies
u0 ∈ H1 (a, b) and (u0 )0 = λu − f.
In other words,
u ∈ H2 (a, b) and λu − u00 = f.
The fact that u admits a continuous extension to the closure of the interval
(a, b) follows from properties of the Sobolev spaces on intervals (see Theorem
). The fact that u vanishes in a and b (if they are finite) follows from Theorem
.

Now let us consider the problem

λu − ∆u = f in Ω,
(2.7)
∂ν u = 0 on ∂Ω,

where Ω ⊆ RN is an open set with C1 -regular boundary ∂Ω, ν is the outer

normal vector, ∂ν u = ∇u · ν is the outer normal derivative, λ ∈ R, f : Ω → C is a
given function, u : Ω → C is the unknown function. The boundary condition
in (2.7) is called (homogeneous) Neumann boundary condition.
We call a function u ∈ H1 (Ω) a weak solution of problem (2.7) if
Z Z Z
λ uϕ̄ − ∇u∇ϕ = f ϕ̄ for every ϕ ∈ H1 (Ω). (2.8)
Ω Ω Ω

Theorem 2.5 (The Laplace operator with Neumann boundary conditions).

For every λ > 0 and every f ∈ L2 (Ω) the problem (2.7) admits a unique weak solution
u ∈ H1 (Ω). For this weak solution we have the estimate

1
kukL2 ≤ k f kL2 .
λ
Proof. Consider on the Hilbert space H1 (Ω) the sesquilinear form a : H1 (Ω) ×
H1 (Ω) → C given by
Z Z
a(u, v) := λ uv̄ + ∇u∇v (u, v ∈ H1 (Ω)).
Ω Ω

Then a is continuous since for every u, v ∈ H1 (Ω), by the Cauchy-Schwarz

inequality,

|a(u, v)| ≤ λ kukL2 kvkL2 + k∇ukL2 kvkL2

≤ (1 + λ) kukH1 kvkH1 .
2.2 The Laplace operator 41

We show that a is also coercive. In fact, for every u ∈ H01 (Ω),

Z Z
a(u, u) = λ |u|2 + |∇u|2
Ω Ω
≥ ηkuk2 1 ,
H0

where η = min{λ, 1} > 0. Hence, a is coercive. R

Consider next the mapping ` : H1 (Ω) → C, v 7→ Ω f v̄, which is well defined
and continuous by the Cauchy-Schwarz inequality, and antilinear. Existence
and uniqueness of a weak solution of (2.7) thus follows from the Lax-Milgram
lemma (Lemma 2.1) applied to a and `. For the estimate, note that we have,
by the Cauchy-Schwarz inequality,
Z
k f kL2 kukL2 ≥ fu
Ω
Z Z
=λ |u| +
2
|∇u|2
Ω Ω
Z
≥λ |u|2
Ω
= λ kuk2L2 .

Theorem 2.6 (Neumann boundary conditions in the case N = 1). Let Ω =

(a, b) be an interval in R (−∞ ≤ a < b ≤ ∞). Then every weak solution u ∈ H1 (a, b) of
(2.7) (λ ∈ R, f ∈ L2 (a, b)) belongs to H2 (a, b) and λu − u00 = f . Moreover, u0 admits
a continuous extension to a and b (if they are finite) and u0 (a) = u0 (b) = 0.

Proof. Since u is a weak solution,

Z b Z b Z b
f v̄ = λ uv̄ + u0 v̄0 for every H1 (a, b).
a a a

In particular, since C∞ 1
c ((a, b)) ⊆ H (a, b),
Z b Z b
u0 ϕ0 = − (λu − f )ϕ for every ϕ ∈ C∞
c ((a, b)).
a a

By definition of the weak derivative, and since u − f ∈ L2 (a, b), this equality
implies
u0 ∈ H1 (a, b) and (u0 )0 = λu − f.
In other words,
u ∈ H2 (a, b) and λu − u00 = f.
The fact that u0 admits a continuous extension to the closure of the interval
(a, b) follows from properties of the Sobolev spaces on intervals (see Theorem
42 2 Elliptic equations

). Assume, for simplicity, that both a and b are finite. For every ϕ ∈ H1 (a, b),
we have, by an integration by parts,
Z b Z b Z b
f ϕ̄ = λ uϕ̄ + u0 ϕ̄0
a a a
Z b h ib Z b
=λ uϕ̄ + u0 ϕ̄ − u00 ϕ̄
a
a a
Z b
= (λu − u00 )ϕ̄ + u0 (b)ϕ̄(b) − u0 (a)ϕ̄(b)
a
Z b
= f ϕ̄ + u0 (b)ϕ̄(b) − u0 (a)ϕ̄(b).
a

In other words,

u0 (b)ϕ̄(b) − u0 (a)ϕ̄(b) = 0 for every ϕ ∈ H1 (a, b).

Choosing now successively ϕ(x) := x−a

b−a and ϕ(x) = b−x
b−a , we obtain

u0 (b) = u0 (a) = 0.

2.3 General elliptic operators in divergence form and

inhomogeneous Dirichlet boundary conditions

Consider the elliptic operator L which is formally given by

N
X N
X
Lu = ∂i (ai j (x)∂ j u) + [∂i (bi (x)u) + ci (x)∂i u] + d(x)u,
i,j=1 i=1

where
ai j , bi , ci , d ∈ L∞ (Ω) (1 ≤ i, j ≤ N),
and Ω ⊆ RN is an open set. We assume that the coefficients ai j are uniformly
elliptic in the sense that there exists η > 0 such that
N
X
ai j (x)ξ j ξi ≥ η |ξ|2 for all ξ ∈ CN , x ∈ Ω. (2.9)
i,j=1

There are no further conditions on the lower order conditions bi , ci and d,

and in fact, the boundedness condition on these coefficients may be relaxed
a little bit. All coefficients may be complex valued. We then consider the
problem
2.3 General elliptic operators in divergence form 43

λu − Lu = f + divg in Ω,
(2.10)
u = h on ∂Ω,

where

f, g1 , . . . , gN ∈ L2 (Ω) and
h ∈ H1 (Ω).

We say that a function u ∈ H1 (Ω) is a weak solution of the problem

λu − Lu = f + divg in Ω, (2.11)

if, for all ϕ ∈ C∞

c (Ω),

Z N Z
X
λ uϕ̄ + ai j (x)∂ j u∂i ϕ+
Ω i,j=1 Ω

N Z
X Z
+ [−bi (x)u∂i ϕ + ci (x)∂i uϕ̄ + d(x)uϕ̄ (2.12)
i=1 Ω Ω
Z N Z
X
= f ϕ̄ + gi ∂i ϕ.
Ω i=1 Ω

Note that by an approximation argument, if the above equality holds for

all test functions ϕ ∈ C∞
c (Ω), then it holds for all ϕ ∈ H0 (Ω), and vice versa.
1

Next, we say that the inhomogeneous Dirichlet boundary condition

u = h on ∂Ω (2.13)

is satisfied, if
u − h ∈ H01 (Ω). (2.14)
Accordingly, u ∈ H1 (Ω) is a weak solution of (2.10) if it satisfies both (2.12)
and (2.14).

Theorem 2.7. Assume that Ω, ai j , bi , ci , d, f , gi and h are as above. Then there

exists a real number λ̂ such that for all λ ∈ C with Re λ > λ̂ the problem (2.10)
admits a unique weak solution u ∈ H1 (Ω).

Proof. Assume first that h = 0. Then we define the sesquilinear form a :

H01 (Ω) × H01 (Ω) → C by
44 2 Elliptic equations
Z N Z
X
a(u, v) = λ uv̄ + ai j (x)∂ j u∂i v+
Ω i, j=1 Ω

N Z
X Z
+ [−bi (x)u∂i v + ci (x)∂i uv̄ + d(x)uv̄.
i=1 Ω Ω

Then a is continuous since for every u, v ∈ H01 (Ω), by the Cauchy-Schwarz

inequality,
N
X
|a(u, v)| ≤ |λ| kukL2 kvkL2 + kai j kL∞ k∂ j ukL2 k∂i vkL2 +
i,j=1
N
X
+ [kbi kL∞ kukL2 k∂i vkL2 + kci kL∞ k∂i ukL2 kvkL2 ] + kdkL∞ kukL2 kvkL2
i=1
≤ C kukH1 kvkH1 ,

where C ≥ 0 is for example the sum of |λ| and the L∞ -norms of the coefficients
ai j , bi , ci , d.
We show that a is also coercive whenever
N
1 X
Re λ > λ̂ := kdkL∞ + (kbi k2L∞ + kci k2L∞ ).
2η
i=1

In fact, for every such λ ∈ C and every u ∈ H01 (Ω), by the uniform ellipticity
condition, by the Cauchy-Schwarz inequality and by Young’s inequality,
Z N Z
X
Re a(u, u) = Re λ |u|2 + Re ai j ∂ j u∂i u+
Ω i,j=1 Ω

N Z
X Z
+ Re [−bi (x)u∂i u + ci (x)∂i uū + Re d(x)|u|2
i=1 Ω Ω

≥ (Re λ − kdkL∞ ) kuk2L2 + η k∇uk2L2 −

N  12
X 
−  (kbi kL∞ + kci kL∞ ) kukL2 k∇ukL2
 2 2

i=1
η
≥ (Re λ − λ̂)kuk2L2 + k∇uk2L2
2
≥ η̃kuk 1 ,
2
H0

η
where η̃ = min{Re λ − λ̂, 2 } > 0. Hence, a is coercive.
Consider next the mapping ` : H1 (Ω) → C given by
2.4 The comparison and maximum principles 45
Z N Z
X
`(v) = f v̄ − gi ∂i v,
Ω i=1 Ω

which is well defined and continuous by the Cauchy-Schwarz inequality,

and antilinear. Existence and uniqueness of a weak solution of (2.10) thus
follows from the Lax-Milgram lemma (Lemma 2.1) applied to a and `.
Let now h ∈ H1 (Ω) be arbitrary, and define
N
X
fˆ := f + λh + ci ∂i h + dh ∈ L2 (Ω) and
i=1
N
X
ĝi := gi + ai j ∂ j h + bi h ∈ L2 (Ω) (1 ≤ i ≤ N).
j=1

Then one easily verifies that u ∈ H1 (Ω) is a weak solution of (2.10) if and only
if w = u − h ∈ H01 (Ω) is a weak solution of

λw − Lw = fˆ+ div ĝ in Ω,
w = 0 on ∂Ω,

and from this equivalence and the first step one obtains existence and unique-
ness of a weak solution of (2.10).

2.4 The comparison and maximum principles

Theorem 2.8 (Comparison principle I). Let Ω ⊆ RN be open, f ∈ L2 (Ω) and

λ > −λ1 (Ω). Let u ∈ H01 (Ω) be the unique weak solution of (2.4). If f ≥ 0, then
u ≥ 0, and similarly if f ≤ 0, then u ≤ 0.

Proof. Let f , λ and u be as in the assumption and assume that f ≤ 0. Taking

ϕ = u+ ∈ H01 (Ω) as a test function in the definition of a weak solution, we
obtain
Z
0≥ f u+
Ω
Z Z
=λ u u+ + ∇u∇u+
ZΩ ZΩ
+ 2
= λ (u ) + |∇u+ |2
Ω Ω
Z
+ 2
≥ (λ + λ1 ) (u ) .
Ω
46 2 Elliptic equations

Since λ + λ1 > 0, this inequality implies u+ = 0 which means that u ≤ 0. The

case f ≥ 0 is proved similarly, or follows from this case by multiplying the
equation (2.4) by −1.

We consider now again the problem (2.10) with a general elliptic oper-
ator in divergence form and inhomogeneous boundary conditions, but we
assume that the coefficients ai j , bi , ci and d, as well as the functions f , gi and
h are real valued. Repeating the existence and uniqueness proof (proof of
Theorem 2.7) in the real Hilbert space H1 (Ω), we see that the unique weak
solution u ∈ H1 (Ω) is real valued, too.
We say that u ∈ H1 (Ω) is a subsolution (resp. supersolution) of (2.11) and
we write
λu − Lu ≤ f + divg in Ω (resp. ≥) (2.15)
if, for every positive test function ϕ ∈ C∞
c (Ω),

Z N Z
X
λ uϕ + ai j (x)∂ j u∂i ϕ+
Ω i, j=1 Ω

N Z
X Z
+ [−bi (x)u∂i ϕ + ci (x)∂i uϕ + d(x)uϕ (2.16)
i=1 Ω Ω
Z N Z
X
≤ fϕ+ gi ∂i ϕ (resp. ≥).
Ω i=1 Ω

Moreover, given a constant k ∈ R, we write

u ≤ k on ∂Ω

if
(u − k)+ ∈ H01 (Ω),
and we define

sup u := {k ∈ R : u ≤ k on ∂Ω}
∂Ω
= {k ∈ R : (u − k)+ ∈ H01 (Ω)}, and
inf u := − sup(−u)
∂Ω ∂Ω
= sup{k ∈ R : (u + k)− ∈ H01 (Ω)}.

Theorem 2.9 (Comparison principle II). Let Ω ⊆ RN be open, f ∈ L2 (Ω) and

λ > −λ1 (Ω). Let u ∈ H1 (Ω) be such that

λu − ∆u ≤ 0 in Ω,
u ≤ 0 on ∂Ω.
2.4 The comparison and maximum principles 47

Then u ≥ 0.

Proof. Let λ and u be as in the assumption. Taking ϕ = u+ ∈ H01 (Ω) as a test

function in the definition of a weak solution, we obtain
Z Z
+
0≥λ uu + ∇u∇u+
Ω Ω
Z Z
+ 2
= λ (u ) + |∇u+ |2
Ω Ω
Z
≥ (λ + λ1 ) (u+ )2 .
Ω

Since λ + λ1 > 0, this inequality implies u+ = 0 which means that u ≤ 0.

Theorem 2.10. Let Ω ⊆ RN be open, λ > 0 and f ∈ L2 ∩ L∞ (Ω). Let u ∈ H01 (Ω) be
the unique weak solution of (2.4). Then u ∈ L∞ (Ω) and

k f kL∞
kukL∞ ≤ .
λ
k f kL∞
Proof. Let k := Since u ∈ H01 (Ω) is a weak solution of (2.4),
λ .
Z Z Z
λ uϕ + ∇u∇ϕ = f ϕ for every ϕ ∈ H01 (Ω).
Ω Ω Ω

Taking $ = (u − k)+ ∈ H01 (Ω) as a test function (here we use λ > 0!), we obtain
the equality
Z Z Z
+ +
λ (u − k)(u − k) + ∇u∇(u − k) = ( f − λk)(u − k)+ ,
Ω Ω Ω

and hence
Z Z Z
+ 2 +2
λ ((u − k) ) + |∇(u − k) | = ( f − k f kL∞ )(u − k)+ ≤ 0.
Ω Ω Ω

k f kL∞
This inequality implies (u − k)+ = 0 which is only possible if u ≤ k = λ .
k f kL∞
Similarly, one proves u ≥ − λ . Taking both inequalities together yields the
claim.

For measurable functions f : Ω → R we denote by supΩ f the essential

supremum, that is,

sup f := inf{k ∈ R : f ≤ k almost everywhere},

Ω

and accordingly for infΩ f .

48 2 Elliptic equations

Theorem 2.11. Let Ω ⊆ RN be open and bounded, λ > −λ1 and f ∈ L2 (Ω). Let
u ∈ H1 (Ω) be such that
λu − ∆u ≤ f in Ω.
Then
u ≤ sup u + k(λ)(sup f − λ sup u)+ ,
∂Ω Ω ∂Ω

where k(λ) := kwkL∞ and w ∈ H01 (Ω) is the unique weak solution of

λw − ∆w = 1 in Ω,
(2.17)
w = 0 on ∂Ω.

Similarly, if
λu − ∆u ≥ f in Ω,
then
u ≥ inf u − k(λ)(λ inf u − inf f )+ .
∂Ω ∂Ω Ω

Proof. Set m := sup∂Ω u and M := supΩ f − λ sup∂Ω u. Then the assumption

and the definition of m and M imply

λ(u − m) + ∆(u − m) ≤ f − λm ≤ M in Ω,
u − m ≤ 0 on ∂Ω.

If M ≤ 0, then Theorem 2.9 implies u − m ≤ 0, so that u ≤ sup∂Ω u. If M ≥ 0,

then
λ(u − m − Mw) + ∆(u − m − Mw) ≤ 0 in Ω,
u − m − Mw ≤ 0 on ∂Ω,

where w is the unique solution of (2.17). Now Theorem 2.9 implies u − m −

Mw ≤ 0, so that u ≤ sup∂Ω u + kwk∞ M. Hence, in both cases we obtain the
required estimate. The case with λu − ∆u ≥ f follows from the first case by
multiplying this inequality with −1 and by replacing u by −u.

A function u ∈ H1 (Ω) is harmonic if −∆u = 0 in Ω.

Corollary 2.12 (Weak maximum principle). Let Ω ⊆ RN be open and bounded.

Then, for every harmonic function u ∈ H1 (Ω),

inf u ≤ u ≤ sup u.
∂Ω ∂Ω

Proof. Apply Theorem 2.11 with λ = 0 and f = 0.

2.5 Regularity of weak solutions of elliptic equations 49

2.5 Regularity of weak solutions of elliptic equations

Theorem 2.13 (Elliptic equations in RN ). Let the coefficients ai j ∈ W 1,∞ (RN ) be

uniformly elliptic, f ∈ L2 (RN ), and let u ∈ H1 (RN ) be a weak solution of
N
X
− ∂ j (ai j (x)∂i u) = f in RN .
i, j=1

Then u ∈ H2 (RN ) and

k∂i ∂ j ukL2 ≤ k f kL2 for every 1 ≤ i, j ≤ N.

Proof. We prove the statement in the case when ai j = δi j , which corresponds

to the Laplace operator. The argument for coefficients in W 1,∞ (RN ) is very
similar. For every h ∈ RN , h , 0 and every function u : RN → C we define
Dh u : RN → C by

u(x + h) − u(x)
Dh u(x) = (x ∈ RN ).
|h|

Recall from Theorem 1.18 that a function u ∈ L2 (RN ) belongs to H1 (RN ) if

and only if there exists a constant C ≥ 0 such that kDh ukL2 ≤ C for every
h ∈ RN , h , 0, and then one may choose C = k∇ukL2 . We shall apply this
characterisation to the partial derivatives / the gradient of u. Since u ∈ H1 (RN )
is a weak solution of −∆u = f in RN , for every ϕ ∈ H1 (RN )
Z Z
∇u∇ϕ = f ϕ.
RN RN

Inserting ϕ = D−h (Dh u) into this equality yields

Z Z
f D−h (Dh u) = ∇u ∇D−h (Dh u)
RN N
ZR
= |∇Dh u|2
N
ZR
= |Dh (∇u)|2 .
RN

As a consequence, by the Cauchy-Schwarz inequality and the characterisa-

tion from Theorem 1.18,

kDh (∇u)k2L2 ≤ k f kL2 kD−h (Dh u)kL2 ≤ k f kL2 kDh (∇u)kL2 ,

or
kDh (∇u)kL2 ≤ k f kL2 .
50 2 Elliptic equations

From here and Theorem 1.18 follows the claim.

Corollary 2.14 (Elliptic equations on domains - inner regularity). Let Ω ⊆

RN be open. Let the coefficients ai j ∈ W 1,∞ (Ω) be uniformly elliptic, f ∈ L2 (Ω), and
let u ∈ H1 (Ω) be a weak solution of
N
X
− ∂ j (ai j (x)∂i u) = f in Ω.
i,j=1

2 (Ω), and for every compact K ⊆ Ω there exists a constant C ≥ 0 such

Then u ∈ Hloc K
that
k∂i ∂ j ukL2 (K) ≤ CK k f kL2 (Ω) for every 1 ≤ i, j ≤ N.

Proof.

Corollary 2.15 (Elliptic equations in the half-space; Dirichlet boundary

conditions). Let the coefficients ai j ∈ W 1,∞ (R+N ) be uniformly elliptic, f ∈ L2 (R+N ),
and let u ∈ H01 (R+N ) be a weak solution of

N
X
− ∂ j (ai j (x)∂i u) = f in RN .
i,j=1

Then u ∈ H2 (RN ) and

k∂i ∂ j ukL2 ≤ k f kL2 for every 1 ≤ i, j ≤ N.

Theorem 2.16 (Elliptic equations in domains - regularity up to the bound-

ary; Dirichlet boundary conditions). Let Ω ⊆ RN be open with bounded, C2 -
regular boundary ∂Ω. Let the coefficients ai j ∈ W 1,∞ (Ω) be uniformly elliptic,
f ∈ L2 (Ω), and let u ∈ H01 (Ω) be a weak solution of

N
X
− ∂ j (ai j (x)∂i u) = f in Ω.
i,j=1

Then u ∈ H2 (Ω), and there exists a constant C ≥ 0 depending only on the boundary
∂Ω and the coefficients ai j such that

kukH2 (Ω) ≤ C k f kL2 (Ω) .

Chapter 3
Evolution equations

3.1 Wellposedness results for abstract diffusion equations,

wave equations and Schrödinger equations

Let V and H be complex Hilbert spaces such that V ⊆ H with dense and
continuous embedding. Let a : V ×V → C be an elliptic, bounded, sesquilinear
form. Associated with this form is an operator A : H ⊇ dom A → H given by

dom A := {u ∈ V : ∃ f ∈ H∀v ∈ V : a(u, v) = h f, viH },

Au := f.

If the form a is in addition symmetric, the following theorem asserts that

A is “essentially” a multiplication operator on an abstract L2 -space. It is a
very general form of the theorem which states that every hermitian matrix
is diagonalisable over an orthonormal basis of eigenvectors.

Theorem 3.1 (Spectral theorem for symmetric, elliptic forms). Let V, H, a

and A be as above. Assume in addition that a is symmetric. Then there exists a mea-
sure space (B, B, µ), a (real) measurable function m : B → R which is bounded from
below, and a unitary operator U : H → L2 (B, dµ) such that, given the multiplication
operator M : L2 (B, dµ) ⊇ dom M → L2 (B, dµ) with

dom M := { f ∈ L2 (B, dµ) : m f ∈ L2 (B, dµ)}

= L2 (B, (1 + m2 )dµ),
M f := m f,
1
one has U(dom A) = dom M, U(V) = L2 (B, (1 + m2 ) 2 dµ) and the diagram

51
52 3 Evolution equations

A
dom A −−−−−→ H

 x

U∗ =U−1

yU
 

M
dom M −−−−−→ L2 (B, dµ)

commutes.

We shall not prove this result here. For a proof, see for instance
[Reed and Simon (1980)].

Theorem 3.2 (Wellposedness of abstract diffusion equations). Let V, H, a

and A be as above. Assume in addition that a is symmetric. Then for every u0 ∈ H
there exists a unique function u ∈ C(R+ ; H) ∩ C∞ ((0, ∞); H) such that u(t) ∈ dom A
for every t > 0 and

u̇(t) + Au(t) = 0 in (0, ∞), u(0) = u0 . (3.1)

Proof.

Theorem 3.3 (Wellposedness of abstract wave equations). Let V, H, a and A

be as above. Assume in addition that a is symmetric. Then for every u0 ∈ dom A,
u1 ∈ V there exists a unique function u ∈ C2 (R+ ; H) ∩ C1 (R+ ; V) such that u(t) ∈
dom A for every t ≥ 0 and

ü(t) + Au(t) = 0 in (0, ∞), u(0) = u0 , u̇(0) = u1 . (3.2)

Moreover, one has energy conservation in the sense that for every t ≥ 0

ku̇(t)k2H + a(u(t), u(t)) = ku1 k2H + a(u0 , u0 ).

Theorem 3.4 (Wellposedness of abstract Schrödinger equations). Let V, H,

a and A be as above. Assume in addition that a is symmetric. Then for every
u0 ∈ dom A there exists a unique function u ∈ C1 (R+ ; H) such that u(t) ∈ dom A
for every t ≥ 0 and

u̇(t) + iAu(t) = 0 in (0, ∞), u(0) = u0 . (3.3)

Moreover, one has energy conservation in the sense that for every t ≥ 0

ku(t)k2H = ku0 k2H .

3.2 The comparison principle for diffusion equations

Let Ω ⊆ RN be open, f : R → R be Lipschitz continuous and u0 ∈ L2 (Ω)

real-valued. Let a solution u of
3.2 The comparison principle for diffusion equations 53

∂t u − ∆u − f (u) = 0 in (0, T) × Ω,
u(0, ·) = u0 in Ω,

be a function u ∈ C([0, T]; L2 (Ω)) ∩ C1 (]0, T[; L2 (Ω)) ∩ C(]0, T[; H1 (Ω)), such that
for all φ ∈ C∞c (Ω) and all t ∈]0, T]
Z Z Z
∂t u(t, x)φ(x)dx + ∇u(t, x)∇φ(x)dx + f (u(t, x))φ(x)dx = 0
Ω Ω Ω

and u(0) = u0 . Via approxiation one gets

Z τZ Z τZ Z τZ
∂t u(t, x)φ(t, x)dxdt + ∇u(t, x)∇φ(t, x)dxdt + f (u(t, x))φ(t, x)dxdt = 0
0 Ω 0 Ω 0 Ω

for all φ ∈ C([0, T], H01 (Ω)) with φ(0) = 0 and all τ ∈]0, T].
Notation: QT B (0, T) × Ω.
The parabolic boundary of QT is the set

ΓT = ({0} × Ω̄) ∪ ([0, T] × ∂Ω).

We denote

edge∂t u − ∆u + f (u) ≤ ∂t v − ∆v + f (v)

in QT if u, v ∈ C([0, T]; L2 (Ω)) ∩ C1 (]0, T[; L2 (Ω)) ∩ C(]0, T[; H1 (Ω)) and
Z τZ Z τZ Z τZ Z τZ Z τZ Z τZ
∂t uφ + ∇u∇φ + f (u)φ ≤ ∂t vφ + ∇v∇φ + f (v)φ
0 Ω 0 Ω 0 Ω 0 Ω 0 Ω 0 Ω

for all φ ∈ C([0, T]; H01 (Ω)), φ ≥ 0, φ(0) = 0 and all τ ∈]0, T].
We denote u ≤ v in ΓT if

(u − v)+ ∈ C([0, T]; L2 (Ω)) ∩ C(]0, T[; H01 (Ω))

(u − v)+ (0) = 0.

Theorem 3.5 (Comparison principle). Let

∂t u − ∆u + f (u) ≤ ∂t v − ∆v + f (v) in QT
u ≤ v on ΓT .

Then u ≤ v in QT .

Corollary 3.6. Let f (0) = 0 and u a solution of

54 3 Evolution equations

∂t u − ∆u + f (u) = 0 in QT
u = 0 on (0, T) × ∂Ω
u(0, ·) = u0 in Ω.

If u0 ≥ 0 in Ω, then u ≥ 0 in QT .

Proof. Choose v = 0 in QT as comparison.

Proof (Comparison principle). As required for all 0 < τ0 ≤ τ ≤ T holds

Z τZ Z τZ Z τZ
+ +
(∂t u − ∂t v)(u − v) + (∇u − ∇v)∇(u − v) ≤ − ( f (u) − f (v))(u − v)+
τ0 Ω τ0 Ω τ0 Ω
Z τZ
≤L |u − v|(u − v)+
τ0 Ω
Z τZ
≤L ((u − v)+ )2
τ0 Ω

with Lipschitz constant L. Set

Z
a(t) = ((u − v)+ (t, x))2 dx,
Ω

then a ∈ C([0, T]) and with the inequality we get

Z τ
0
a(τ) − a(τ ) ≤ 2L a(t)dt
τ0

due to
τZ
1 τ d
Z Z Z
+ 1
∂t (u − v)(u − v) = ((u − v)+ )2 = (a(τ) − a(τ0 )),
2 τ0 dt Ω 2
Zτ τ ZΩ
0
Z τZ
∇(u − v)∇(u − v)+ = |∇(u − v)+ |2 = 0
τ0 Ω τ0 Ω

For t0 → 0 and with a(0) = 0 we get

Z T
a(τ) ≤ 2L a(t)dt ∀τ ∈ [0, T].
0

Then (with Gronwall’s Lemma)

a≡0 =⇒ (u − v)+ = 0 in QT ,

i.e. u ≤ v in QT .
Chapter 4
Distributions

4.1 The topology in D(Ω)

Let Ω ⊆ RN be open. In this chapter we write D(Ω) B C∞ c (Ω) to denote the

space of test functions in Ω.
In order to equip the space of test functions with a topology,Schoose a
sequence of bounded, open sets (Ωk )k∈N such that Ω̄k ⊆ Ωk+1 and k∈N Ωk =
Ω, for example

1
 
Ωk B x ∈ Ω : dist(x, Ωc ) ≥ , |x| ≤ k , k ∈ N.
k
n o
Let Ek B φ ∈ D(Ω) : supp φ ⊆ Ω̄k . Every space Ek is equipped with the
countable family (pα )α∈NN of seminorms given by
0

pα (φ) B |∂α φ|∞ ,

and from here one can construct a metric dk as following:

X  
dk (φ, ψ) B cα pα (φ − ψ) ∧ 1 , (φ, ψ ∈ Ek ),
α∈NN
0

where the cα > 0 are such that < ∞.

P
α∈NN cα
0

Remark 4.1. For every k ∈ N the space (Ek , dk ) is a Fréchet space, that is, a
complete, metric vector space (proof: exercise).

Let the topology τ on D(Ω) be the finest topology on D(Ω), such that
every mapping

gk : (Ek , dk ) → (D(Ω), τ)
φ 7→ φ

55
56 4 Distributions

is continuous (final/inductive topology). Together with this topology D(Ω)

becomes a topological vector space, which means that D(Ω) is not only a
vector space and a topological space, but also that the addition and the
multiplication,

+ : D(Ω) × D(Ω) → D(Ω) and

· : K × D(Ω) → D(Ω),

are continuous.
Exercise 4.2 A sequence (φn )n∈N in D(Ω) converges to φ ∈ D(Ω) with respect to τ
if and only if there exists a k ∈ N such that φn , φ ∈ Ek , that is, supp φ, supp φn ⊆ Ω̄k ,
and dk (φn , φ) → 0 for k → ∞.
Exercise 4.3 The topology τ does not depend on the choice of the sequence (Ωk )k∈N .

4.2 Distributions

We denote by

D(Ω)0 B {T : D(Ω) → K : T is linear and continuous}

the dual space of (D(Ω), τ). The elements of D(Ω)0 are called distributions
on Ω, D(Ω)0 is called space of distributions on Ω. We equip D(Ω)0 with the
weak-* topology τ0 , which is the coarsest topology such that all mappings

(D(Ω)0 , τ0 ) → K
T 7→ T(φ),

with φ ∈ D(Ω) are continuous. In the following, given φ ∈ D(Ω) and T ∈

D(Ω)0 , we shall also use the notation

hT, φi B T(φ).

A sequence (Tn )n∈N in D(Ω)0 converges to T ∈ D(Ω)0 with respect to the

topology τ0 if and only if Tn (φ) → T(φ) for all φ ∈ D(Ω).

Examples 4.4 (Distributions).

a) Let f ∈ L1loc (Ω) (⊇ Lp (Ω), C(Ω)). Then

T f : D(Ω) → K,
Z
φ 7→ hT f , φi B f · φ,
Ω
4.2 Distributions 57

is continuous and linear, that is, a distribution. By uniqueness by testing

(Theorem 1.12),

Tf = 0 =⇒ f = 0.

Thus all locally integrable functions are distributions and the mapping

L1loc (Ω) → D(Ω)0

f 7→ T f

is linear and injective, and so L1loc (Ω) ⊆ D(Ω)0 . This embedding is even
continuous, if one equips L1loc (Ω) with the usual Fréchet topology as
follows: let (Ωk )k∈N be a sequence of open, bounded subsets of Ω, such
that
[
Ω̄k ⊆ Ωk+1 , Ωk = Ω.
k∈N

Define for every k ∈ N the seminorm pk by

Z
pk ( f ) B |f| ( f ∈ L1loc (Ω)),
Ωk

and then the metric

X
d( f, g) = 2−k (pk ( f − g) ∧ 1).
k∈N

Then d is a complete metric on L1loc (Ω).

b) Let

Mbloc (Ω) B µ : B(Ω) → R : µ is a signed Radon measure



be the space of all signed Radon measures µ, i.e. µ = µ1 − µ2 , with µ1 ,

µ2 positive Radon measures, i.e. positive regular Borel measures with
the property that

µi (K) < ∞ ∀i ∈ { 1, 2 } ∀K ⊆ Ω compact.

For all µ ∈ Mbloc (Ω)

Tµ : D(Ω) → R
Z
φ 7→ hTµ , φi B φdµ
Ω

is linear and continuous, that is, a distribution.

58 4 Distributions

For example, point evaluation

Tδx0 : D(Ω) → R
Z
φ 7→ hTδx0 , φi B φ(x0 ) = φdδx0
Ω

is a distribution for every x0 ∈ Ω. It is called the Dirac distribution in

the point x0 .
c) For all α ∈ NN
0
, Ω ⊆ RN , x0 ∈ Ω

T : D(Ω) → R
φ 7→ hT, φi B Dα φ(x0 )

is linear and continuous, that is, a distribution.

We call m ∈ N0 the order of a distribution T ∈ D(Ω)0 , if for all compact

subsets K ⊆ Ω there exists a constant cK ≥ 0, such that for all φ ∈ D(Ω) with
supp φ ⊆ K the estimate

|hT, φi| ≤ cK · kφkCm with

X
kφkCm B kDα φkL∞
α∈NN
0
,|α|≤m

holds.

Remark 4.5. a) Distributions of the form T f or Tµ with f ∈ L1loc (Ω) or µ ∈

Mbloc (Ω) are of order 0. Indeed, for all f ∈ L1loc (Ω), all compact K ⊆ Ω and
all φ ∈ D(Ω) with supp(φ) ⊆ K one has
Z Z Z !
|hT f , φi| = | f φ| ≤ | f · g| ≤ | f | kφkC∞
Ω K K

b) We now equip Cm c (Ω) with a topology in a similar way as we did with

C∞c (Ω). Let (Ωk )k∈N Sbe a sequence of open, bounded subsets of Ω such
that Ω̄k ⊆ Ωk+1 and k Ωk . Then we set

c (Ω) : supp φ ⊆ Ω̄k }

Ek B {φ ∈ Cm

and equip this space with the norm

X
kφkCm B kDα φkL∞ .
α∈NN
0
,|α|≤m

Then let τm be the finest topology on Cm

c (Ω), such that every mapping
4.2 Distributions 59

(Ek , k · kCm ) → (Cm

c (Ω), τm )
φ 7→ φ

is continuous. A distribution T ∈ D(Ω)0 is of order m ∈ N0 , if and only

if T extends to a linear, continuous mapping Cmc (Ω) → K.

Theorem 4.6 (Riesz-Markov).

{T ∈ D(Ω)0 : T is of order 0} = Mbloc (Ω) = Cc (Ω)0

(without proof)

Theorem 4.7. Let T ∈ D(Ω)0 be positive, that is, hT, φi ≥ 0 for all φ ∈ D(Ω) with
φ ≥ 0. Then T is of order 0, that is, T is a positive Radon measure.

Let T ∈ D(Ω)0 . Define

[
OT B {U ⊆ Ω : U is open and ∀φ ∈ D(Ω) with supp φ ⊆ U one has hT, φi = 0}.

We call supp T B Ω \ OT the support of T.

Example 4.8. If

hT, φi = Dα φ(x0 )

then
supp T = {x0 }.

For all T ∈ D(Ω)0 , α ∈ NN

0
we define Dα T ∈ D(Ω)0 by

hDα T, φi B (−1)|α| hT, Dα φi (φ ∈ D(Ω)).

The distribution Dα T is called α-th partial derivative of T.

Theorem 4.9 (Consistency of the distributional partial derivatives). For all

1,1
f ∈ Wloc (Ω) = {g ∈ L1loc (Ω) : ∂i g ∈ L1loc (Ω) ∀1 ≤ i ≤ N} and all 1 ≤ i ≤ N one has

∂i T f = T∂i f .

Thus, the partial derivative ∂i in the distributional sense coincides with the weak
1,1
partial derivative ∂i of functions in the Sobolev space Wloc (Ω). In particular, it
1
coincides with the classical partial derivate on C (Ω).

Proof. By definition of ∂i on D(Ω)0 , using the definitions of T f and the weak

derivatives we get
60 4 Distributions

h∂i T f , φi = −hT f , ∂i φi
Z
=− f ∂i φ
Z Ω
= ∂i f φ
Ω
= hT∂i f , φi ∀φ ∈ D(Ω).

4.3 The product and the convolution

We define the product T · ψ ∈ D(Ω)0 of a distribution T ∈ D(Ω)0 and a test

function ψ ∈ D(Ω) by

hT · ψ, φi B hT, ψ · φi (φ ∈ D(Ω)).

Lemma 4.10 (Consistency of the product). If f ∈ L1loc (Ω) and ψ ∈ D(Ω) then

T f · ψ = T f ψ.

Proof. For every f ∈ L1loc (Ω) and ψ, φ ∈ D(Ω) one has

Z
hT f · ψ, φi = hT f , ψ · φi = f · ψ · φ = hT f ·ψ , φi.
Ω

Lemma 4.11. For every T ∈ D(Ω)0 and every ψ ∈ D(Ω) one has

supp T · ψ ⊆ supp T ∩ supp ψ (⊆ supp ψ)

For a distribution T ∈ D(Ω)0 and a test function φ ∈ D(RN ) we define the

convolution T ∗ φ by setting

T ∗ φ(x) B hT, φ(x − ·)i

for all x ∈ Ω, such that φ(x − ·) ∈ D(Ω), so x − supp φ ⊆ Ω.

One observes that the set {x ∈ Ω : φ(x − ·) ∈ D(Ω)} = Uφ is an open subset
with respect to Ω.

Lemma 4.12. If T ∈ D(Ω)0 and φ ∈ D(Ω) then

T ∗ φ ∈ C∞ (Uφ ).

If supp T + supp φ ⊆ Ω and supp T are compact, and if we extend T ∗ φ by 0, then

T ∗ φ ∈ D(Ω). Moreover,

Dα (T ∗ φ) = T ∗ Dα φ = Dα T ∗ φ.
4.4 Tempered distributions 61

Theorem 4.13. The space D(Ω) is sequentially dense in D(Ω)0 with respect to the
weak-* topology τ0 , that is, for all τ ∈ D(Ω) there exists a sequence (φn )n∈N ∈ D(Ω),
such that

Tφn → T in D(Ω)0 .

Proof. Truncation and regularization.

R Let (φn )n∈N be an approximation of the identity, that is, φ1 ∈ D(R ), φ1 ≥ 0,

N

φ = 1 and
RN 1

φn x) B nN φ1 (n · x) (n ∈ N, x ∈ RN ).
R
Then φn ≥ 0 and RN
φn ≥ 1 for every n. Moreover, for all ψ ∈ D(RN )
Z
hTφn , ψi = φn (x)ψ(x)dx → ψ(0) = hTδ0 , ψi, (n → ∞),
RN

that is, the approximation of the identity (φn ) converges to the Dirac distri-
bution Tδ0 in D(RN )0 .

Remark 4.14. For all φ ∈ D(RN ) one has

Tδ0 ∗ φ(x) = hTδ0 , φ(x − ·)i = φ(x),

or

Tδ0 ∗ φ = φ.

4.4 Tempered distributions

We define the space

Z
S(R ) := { f ∈ C (R ) : ∀α, β ∈
N ∞ N
NN
0 : |xβ ∂α f (x)|2 dx < ∞}.
RN

Elements of S(RN ) are called the rapidly decreasing functions or Schwartz

c (R ) =
(test) functions. Clearly, the space of (classical) test functions C∞ N
N N
D(R ) is a subspace of S(R ), but it is a proper subspace since the function
2
f (x) = e−x is an example of a Schwartz test function which does not have
compact support.
It is an exercise to show that
62 4 Distributions
Z
S(RN ) = { f ∈ C∞ (RN ) : ∀α, β ∈ NN
0 : |xβ ∂α f (x)| dx < ∞}
RN
β α
0 : sup |x ∂ f (x)| < ∞}.
= { f ∈ C∞ (RN ) : ∀α, β ∈ NN
x∈RN

The space S(RN ) is equipped with the topology induced by the countable
family of seminorms (k · kα,β )α,β∈NN , where
0

Z !1
2
β α
k f kα,β := |x ∂ f (x)| dx 2
.
RN

This countable family of seminorms induces in a natural way a metric d

given by
X k f − gkα,β
d( f, g) := cα,β ,
1 + k f − gkα,β
N α,β∈N0

where the coefficients cα,β > 0 are fixed such that < ∞. We have
P
α,β∈NN cα,β
0

fn → f in S(RN ) ⇔ ∀α, β ∈ NN
0 : k fn − f kα,β → 0
⇔ d( fn , f ) → 0,

and the space S(RN ) is complete. In other words, the countable family of
seminorms turns S(RN ) into a Fréchet space.
From the definition of the space S(RN ) we immediately obtain the follow-
ing lemma which is, however, worth of being stated separately.
Lemma 4.15. For every f ∈ S(RN ) and every polynomial p : CN → C the product
p f and the (sum of) partial derivatives p(∂) f belong again to S(RN ). In other words,
the mappings

f 7→ p f and
f 7→ p(∂) f

leave the space S(RN ) invariant.

Elements of the dual space

S(RN )0 := {T : S(RN ) → C : T is linear and continuous}

are called tempered distributions. Since D(RN ) ⊆ S(RN ) with dense and
continuous embedding, we obtain the inclusion

D(RN ) ⊆ S(RN ) ⊆ S(RN )0 ⊆ D(RN )0 .

In particular, every tempered distribution is a distribution. For every multi-

index α ∈ NN 0
and every tempered distribution T ∈ S(RN )0 we define the
4.5 The Fourier transform 63

partial derivative ∂α T ∈ S(RN )0 and the product xα T ∈ S(RN )0 by

h∂α T, ϕi := (−1)|α| hT, ∂α ϕi and

α α
hx T, ϕi := hT, x ϕi.

These linear operations are consistent with the classical partial derivatives
and the product, respectively, on the space S(RN ).

4.5 The Fourier transform

4.5.1 The Fourier transform on L1 (RN )

For every f ∈ L1 (RN ) we define the Fourier transform F f and the adjoint
Fourier transform F̄ f by
Z
F f (x) := e−ixy f (y) dy and
N
ZR
F̄ f (x) := eixy f (y) dy (x ∈ RN ).
RN

The integrals are absolutely convergent, and we have the trivial estimates

|F f (x)|, |F̄ f (x)| ≤ k f kL1 for every x ∈ RN .

In particular, the functions F f and F̄ f are bounded.

Theorem 4.16 (Riemann-Lebesgue). For every f ∈ L1 (RN ) one has F f , F̄ f ∈
C0 (RN ).
Proof. The fact that the Fourier transform F f is continuous follows easily
from Lebesgue’s dominated convergence theorem. Next, for every x ∈ RN ,
x , 0,
Z
1 iπ x·x
F f (x) = (e−ixy − e−ixy e |x|2 ) f (y) dy
2 RN
πx
Z
1
= eixy ( f (y) − f (y + 2 )) dy.
2 RN |x|

Since the shift group on L1 (RN ) is strongly continuous by Lemma 1.7, we

thus obtain
πx
Z
1
|F f (x)| ≤ | f (y) − f (y + 2 )| dy → 0 as |x| → ∞.
2 RN |x|

The arguments for the adjoint Fourier transform are similar.

64 4 Distributions

Corollary 4.17. The Fourier transform F and the adjoint Fourier transform are
bounded, linear operators from L1 (RN ) into C0 (RN ).

We need the following basic lemma in order to prove the inversion formula
for the Fourier transform.

Lemma 4.18 (Féjer kernel). One has, for a > 0,

sin2 ax
Z
dx = aπ.
R x2

Proof. We define
∞
sin2 ax
Z
f (λ) := e−λx dx (λ ∈ (0, ∞)).
0 x2

Then f ∈ C∞ ((0, ∞)) and

∞
sin2 ax sin2 ax
Z Z
1
lim f (λ) = dx = dx, and
λ→0+ 0 x2 2 R x2
lim f (λ) = 0.
λ→∞

A simple computation shows

∞
sin2 ax
Z
f (λ) = −
0
e−λx dx, and
0 x
Z ∞
f (λ) =
00
e−λx sin2 ax dx
0
∞ !2
eiax − e−iax
Z
= e−λx dx
0 2i
1 1 2 1
 
=− − +
4 λ − 2ia λ λ + 2ia
1 2 2λ
 
= − 2 .
4 λ λ + 4a2
As a consequence,
1 λ2
f 0 (λ) =
log 2 .
4 λ + 4a2
In order to integrate this function, we make the ansatz

λ2
!
1
f (λ) = λ log 2 + g(λ) ,
4 λ + 4a2

which leads to the equation

4.5 The Fourier transform 65

8a2
g0 (λ) = − ,
λ2 + 4a2
that is,
λ
g(λ) = −4a arctan +C
2a
Together with the condition limλ→∞ f (λ) = 0 we thus find

λ2 π λ
!
1
f (λ) = λ log 2 + 4a( − arctan ) .
4 λ + 4a2 2 2a

This yields
π
lim f (λ) = a ,
λ→0 2
which implies the claim.
Before stating the following theorem we define for every r ∈ RN with rk ≥ 0
the set
N
Qr := [−rk , rk ].
k=1

Theorem 4.19 (Inversion formula for the Fourier transform I). Let f ∈
L1 (RN ). For every R > 0 we put
Z Z
1
gR (x) := eixy F f (y) dy dr (x ∈ RN ).
(2πR)N [0,R]N Qr

Then gR ∈ L1 (RN ) and

lim kgR − f kL1 = 0.
R→∞

Proof. For every R > 0 and every x ∈ RN we compute, using Fubini’s theorem,
Z Z
1
eixy F f (y) dy dr
(2πR)N [0,R]N Qr
Z Z Z
1
= eiy(x−z) dy f (z) dz dr
(2πR)N [0,R]N RN Qr
Z Z N
1 Y sin(rk (xk − zk ))
= dr f (z) dz
(πR)N RN [0,R]N xk − zk
k=1
N
sin2 ( R (xk − zk ))
Z Y
2
= f (z) dz
2 π(xk − zk )
R 2
RN k=1
Z
= kR (x − z) f (z) dz
R
= kR ∗ f (x)
66 4 Distributions

where
N
Y sin2 ( R xk )
2
kR (x) := (x ∈ RN )
k=1
R
2 πx 2
k

is the Féjer kernel. Note that

kR ∈ L1 (RN ),
kR ≥ 0,
kR (x) = RN k2 (Rx) for every x ∈ RN and
Z
kR (x) dx = 1 (Lemma 4.18)
RN

for every R > 0. Hence, (kR )R%∞ is an approximate identity, and the claim
follows from Young’s inequality and Theorem 1.8.

Corollary 4.20 (Inversion formula for the Fourier transform II). Let f ∈
L1 (RN ) be such that F f ∈ L1 (RN ). Then F̄ f ∈ L1 (RN ) and

1
f= F̄ (F f ) and
(2π)N
1
f= F (F̄ f ).
(2π)N

Proof. Since
Z
F̄ f (x) = eixy f (y) dy = F f (−x) for every x ∈ RN ,
RN

we immediately obtain F̄ f ∈ L1 (RN ).

Now let gR be defined as in the preceding theorem. For every R > 0 and
every x ∈ RN we then have

1
gR (x) − F̄ (F f )(x) =
(2π)N
" Z Z Z #
1 1
= ixy
e F f (y) dy dr − ixy
e F f (y) udy
(2π)N RN [0,R]N Qr RN
Z Z
1 1
= eixy F f (y) dy dr,
(2π) RN [0,R]N Qcr
N

and hence, for every L > 0

4.5 The Fourier transform 67

1
lim sup gR (x) − F̄ (F f )(x) ≤
R→∞ (2π)N
Z Z
1 h 1
≤ lim sup eixy F f (y) dy dr
(2π)N R→∞ RN [L,R]N Qcr
Z Z
1 i
+ lim sup N eixy F f (y) dy dr
R→∞ R [0,R]N \[L,R]N Qcr
(R − L)N NLRN−1
Z
1 h i
≤ N
lim sup N
|F f (y)| dy + lim sup N
kF f kL1
(2π) R→∞ R ([−L,L]N )c R→∞ R
Z
1
≤ |F f (y)| dy
(2π)N ([−L,L]N )c

Since L > 0 was arbitrary, and since

Z
lim |F f (y)| dy = 0,
L→∞ ([−L,L]N )c

we thus obtain
1
lim gR (x) = F̄ (F f )(x) for every x ∈ RN .
R→∞ (2π)N

Combining this with the first inversion formula, we obtain the first identity.
The second identity is proved similarly.

Corollary 4.21. The Fourier transforms F , F̄ : L1 (RN ) → C0 (RN ) are injective.

Remark 4.22. The Fourier transform F on L1 is not surjective onto C0 .

Lemma 4.23 (Fourier transform and convolution). For every f , g ∈ L1 (RN )

one has

F ( f ∗ g) = F f F g and
F̄ ( f ∗ g) = F̄ f F̄ g.

Proof. For every x ∈ RN we compute, using Fubini’s theorem,

68 4 Distributions
Z
F ( f ∗ g)(x) = e−ixy f ∗ g(y) dy
RN
Z Z
= e−ixy f (y − z)g(z) dy dz
RN RN
Z Z
= e−ix(y+z) f (y) dy g(z) dz
RN RN
Z Z
= e−ixy
f (y) dy e−ixz g(z) dz
RN RN
= F f (x) F g(x).

The second identity is proved similarly.

4.5.2 The Fourier transform on S(RN )

Lemma 4.24. For every f ∈ S(RN ) and every polynomial p : CN → C one has F f ,
F̄ f ∈ C∞ (RN ), and

F (p(∂) f ) = p(i·)F f,
F (p(−i·) f ) = p(∂)F f,
F̄ (p(∂) f ) = p(−i·)F̄ f, and
F̄ (p(i·) f ) = p(∂)F̄ f.

Proof. Let f ∈ S(RN ) and k ∈ {1, . . . , N}. Then

Z
F (−i ·k f )(x) = − e−ixy iyk f (y) dy
RN
∂ −ixy
Z
= (e ) f (y) dy
R N ∂xk
Z
= ∂k e−ixy f (y) dy
RN
= ∂k F f (x).

Moreover, by an integration by parts,

4.5 The Fourier transform 69

∂
Z
F (∂k f )(x) = e−ixy f (y) dy
RN ∂yk
∂ −ixy
Z
=− (e ) f (y) dy
RN ∂yk
Z
= −ixk e−ixy f (y) dy
RN
= −ixk F f (x).

The first two equalities follow from these two identities and by induction.
The proofs for the adjoint Fourier transform F̄ are similar.
Theorem 4.25. For every f ∈ S(RN ) one has F f , F̄ f ∈ S(RN ) and the Fourier
transforms F , F̄ : S(RN ) → S(RN ; X) are linear, continuous isomorphisms.
Proof. The statement essentially follows from the preceding Lemma 4.24.

4.5.3 The Fourier transform on L2

Theorem 4.26 (Parseval’s identity).

a) For every T ∈ L1 (RN ; L(X, Y)) and every f ∈ L1 (RN ; X) one has
Z Z
F T(x) f (x) dx = T(x)F f (x) dx.
RN RN

b) For every f , g ∈ L1 (RN ) one has

Z Z
¯
F f (x) g(x) dx = f (x)F̄ g(x) dx.
RN RN

c) For every f , g ∈ L1 (RN ) such that F f , F g ∈ L1 (RN ) one has

Z Z
¯ dx = 1
f (x) g(x) F f (x)F g(x) dx.
RN (2π)N RN

Similar identities hold if we replace everywhere F by F̄ and vice versa.

Proof. (a) We calculate, using Fubini’s theorem,
Z Z Z
F T(x) f (x) dx = e−ixy T(y) dy f (x) dx
RN RN RN
Z Z
= T(y) e−ixy f (x) dx dy
RN RN
Z
= T(y)F f (y) dy.
RN
70 4 Distributions

(b) is proved in a similar way and (c) follows from (b) by using the Inversion
Formula II (Corollary 4.20).
Theorem 4.27 (Plancherel). The Fourier transforms F , F̄ : C∞ N 2
c (R ) → L (R )
N
2 N 1
extend uniquely to bounded, linear operators on L (R ). The operators √
N
F,
2π
√ 1 F̄ : L2 (RN ) → L2 (RN ) are unitary and
2πN

1 1
(√ F )∗ = √ F̄ .
2πN 2πN
Proof. From Parseval’s identity (Theorem 4.26 (c)) we obtain, that for every
f , g ∈ C∞ N
c (R )
hF f, F giL2 = (2π)N h f, giL2 ,
and in particular,
kF f k2L2 = (2π)N k f k2L2 .

As a consequence, since C∞ N 2 N
c (R ) is dense in L (R ), F extends in a unique
2 N
way to a bounded, linear operator on L (R ). Moreover, we see from the
above equality that √ 1 N F is isometric. As a consequence, this operator is
2π
injective and has closed range. However, from the inversion formula we see
1
that C∞ N
c (R ) is contained in the range. Hence, √2πN F is surjective, and thus
unitary.
The arguments for F̄ are similar.

4.5.4 The Fourier transform on S(RN )0

4.6 The theorem of Malgrange-Ehrenpreis

In this section we state and prove the theorem of Malgrange-Ehrenpreis.

The proof given here follows the lines of the proof of H. König and may be
found, for example, in [Walter (1994)].

Let p : CN → C be a complex polynomial of degree m ∈ N,

X
p(z) = aα zα (z ∈ CN )
α∈NN
0
|α|≤m

with fixed coefficients aα ∈ C. Denote by ṗ the main part of this polynomial,

that is, X
ṗ(z) := aα zα (z ∈ CN )
α∈NN
0
|α|=m
4.6 The theorem of Malgrange-Ehrenpreis 71

with ṗ , 0.
We call a tempered distribution T ∈ S(RN )0 a fundamental solution for
the partial differential operator with constant coefficients
X
p(∂) = aα ∂α
α∈NN
0
|α|≤m

(which in principle acts on D(Ω)0 for every open Ω ⊆ RN ) if

p(∂)T = δ0 ,

where δ0 is the Dirac distribution in 0. Given a fundamental solution for the

differential operator p(∂), one may solve the partial differential equation

p(∂)u = f in RN

for given right-hand side f ∈ D(RN ) by putting u := T ∗ f ∈ C∞ (RN ). In fact,

for this function u one has

p(∂)u = p(∂)(T ∗ f ) = (p(∂)T) ∗ f = δ0 ∗ f = f.

Theorem 4.28 (Malgrange-Ehrenpreis). For every polynomial p : CN → C the

differential operator p(∂) admits a fundamental solution.

Let us note some observations which are useful for the proof of this
theorem.

Observation 4.29. For every a = (a1 , . . . , aN ) ∈ CN and every n ∈ N one has

 n
X N  X n!
 a = aα

j
α!
 

j=1 |α|=n

and therefore  n
N
 X  X n!
1 + a j  = aα .
 
  (n − |α|)!α!
j=1 |α|≤n

If z = (z1 , . . . , zn ) ∈ CN and a j = z j z̄ j = |z j |2 , then one has

X n!
(1 + |z|2 )n = zα z̄α .
(n − |α|)!α!
|α|≤n

Observation 4.30. Let Γ be the unit circle in C, ΓN ⊆ CN the N-fold cartesian

product. We write
72 4 Distributions
Z Z 2π
1
f (z)dτ(z) := f (eiθ ) dθ.
Γ 2π 0

Note that for every c ∈ Γ,

Z Z
f (cz)dτ(z) = f (z)dτ(z).
Γ Γ
R
Accordingly, we define for f : ΓN → C the iterated integral ΓN
f (z) dτN (z).
For every pair of multi-indices α, β ∈ NN
0
one has

if α , β,

0
Z 
α β
z z̄ dτN (z) = δαβ = 

ΓN if α = β.

1

Let us define the trace of the main part of the polynomial p by

X
s(p) = |aα |2 > 0.
|α|=m

One has Z
s(p) = |ṗ(z)|2 dτN (z),
ΓN
because

|ṗ(z)|2 = ṗ(z) ṗ(z)

X X
= aα aβ zα z̄β .
|α|=m |β|=m

P Let kg : C → C be holomorphic and let q : C → C be a poly-

Observation 4.31.
nomial, q(z) = m
k=0 ck z with cm , 0. Then Cauchy’s theorem implies
Z
g(0) cm = g(z) q(z) zm dτ(z).
Γ

Indeed, note that for z ∈ Γ one has z̄ = z−1 and therefore

m
X
q(z) z =
m
ck zm−k .
k=0

Observation 4.32. Let f : CN → C be holomorphic. Then, for every z ∈ CN ,

Z
f (z) s(p) = f (z + s) p(z + s) ṗ(s) dτN (s).
ΓN
4.6 The theorem of Malgrange-Ehrenpreis 73

In fact, if one fixes z, s ∈ CN , and if one sets

g(t) := f (z + ts) (t ∈ C), and

m−1
X
q(t) := p(z + ts) = ṗ(s)tm + A j (z, s) t j ,
j=0

where the A j are independent of t, then g(0) = f (z), cm = ṗ(s), and the preced-
ing observation yields

f (z) |ṗ(s)|2 = f (z)ṗ(s)ṗ(s)

Z
= f (z + ts) p(z + ts) tm ṗ(s) dτ(t)
Γ
Z
= f (z + ts) p(z + ts) ṗ(ts) dτ(t).
Γ

Integrating both sides over s ∈ ΓN yields

Z Z
f (z) s(p) = f (z + ts)p(z + ts)ṗ(ts) dτN (s) dτ(t)
ZΓ Γ
N

= f (z + s)p(z + s)ṗ(s) dτN (s).

ΓN

Proof (of Theorem 4.28). Let us define the function h·i : C → C by

 t̄

t
 if t , 0,
hti := 
if t = 0.

0

Let u ∈ D(RN ) be a test function. Then its Fourier transform F u : RN → C

extends to a holomorphic function CN → C which we also denote by F u.
The inversion formula for the Fourier transform yields
Z
1
u(0) = F u(x) dx,
(2π)N RN

and therefore
74 4 Distributions
Z
(2π)N s(p)u(0) = s(p)F u(x) dx
N
ZR Z
= F u(x + s) p(x + s) ṗ(s) dτN (s) dx
ZR ZΓ
N N

= p(x + s) F u(x + s) hp(x + s)i ṗ(s) dτN (s) dx

ZR ZΓ
N N

= F (p(−i∂)u)(x + s) hp(x + s)i ṗ(s) dτN (s) dx.

RN ΓN

Setting v = p(−i∂)u, we have

X
(1 + |z|2 )n F v(z) = cα hzα iz2α F v(z)
|α|≤n
X
= cα hzα iF (∂2α v)(z)(−1)|α| ,
|α|≤n

n!
where cα := (n−|α|)!α! . Hence,

h(x + s)α p(x + s)i

Z Z X
(2π)N s(p)u(0) = (−1)|α| cα ṗ(s) F (∂2α v)(x+s) dτN (s) dx.
RN ΓN |α|≤n (1 + |x + s|2 )n

We take n large enough, so that N

2 < n. Then the function Iα given by

h(x + s)α p(x + s)i

Iα (x, s) := (−1)|α| cα ṗ(s)
(1 + |x + s|2 )n

is integrable over RN × ΓN . By interchanging a sum and some integrals, we

obtain
Z Z X
(2π) s(p)u(0) =
N
Iα (x, s) F (∂2α v)(x + s) dτN (s) dx
RN ΓN |α|≤n
Z Z X Z
= Iα (x, s) e−i(x+s)y ∂2α v(y) dy dτN (s) dx
RN ΓN |α|≤n RN
Z  
 X Z Z 
= Iα (x, s) e−i(x+s)y dτN (s) dx ∂2α v(y) dy.
 

R N  R N Γ N
|α|≤n


Setting
XZ Z
Kp (y) := Iα (x, s) e−i(x+s)y dτN (s) dx (y ∈ RN ),
|α|≤n RN ΓN

we have Kp ∈ L1loc (RN ). Now it suffices to put

4.6 The theorem of Malgrange-Ehrenpreis 75

1 X
T := ∂2α Kp ,
s(p) (2π)N
|α|≤m

and one obtains

hp(i∂)T, ui = hT, p(−i∂)ui = hT, vi = u(0),

or, in different notation,

p(i∂)T = δ0 .
References

[Brézis (1992)] Brézis, H. : Analyse fonctionnelle. Masson, Paris, 1992.

[Dautray and Lions (1988)] Dautray, R., Lions, J.-L. : Mathematical analysis and numerical
methods for science and technology. Vol. 2. Springer-Verlag, Berlin, 1988.
[Dautray and Lions (1990)] Dautray, R., Lions, J.-L. : Mathematical analysis and numerical
methods for science and technology. Vol. 1. Springer-Verlag, Berlin, 1990.
[Dautray and Lions (1992)] Dautray, R., Lions, J.-L. : Mathematical analysis and numerical
methods for science and technology. Vol. 5. Springer-Verlag, Berlin, 1992.
[Dautray and Lions (1993)] Dautray, R., Lions, J.-L. : Mathematical analysis and numerical
methods for science and technology. Vol. 6. Springer-Verlag, Berlin, 1993.
[Evans (1998)] Evans, L. C. : Partial Differential Equations. Vol. 19 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 1998.
[Gilbarg and Trudinger (2001)] Gilbarg, D., Trudinger, N. S. : Elliptic Partial Differential
Equations of Second Order. Springer Verlag, Berlin, Heidelberg, New York, 2001.
[Reed and Simon (1980)] Reed, M., Simon, B. : Methods of modern mathematical physics. I,
2nd Edition. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York,
1980.
[Walter (1994)] Walter, W. : Einführung in die Theorie der Distributionen, 3rd Edition. Bibli-
ographisches Institut, Mannheim, 1994.

77
Index

Fejer kernel, 64 Parseval’s identity, 69

Fourier transform Plancherel’s theorem, 70
adjoint, 63
inversion formula, 65, 66 Riemann-Lebesgue theorem, 63
on S, 69
on L1 , 63
on L2 , 70 Schwartz test function, 61
function
rapidly decreasing, 61 theorem
Schwartz test function, 61 Parseval, 69
Plancherel, 70
inversion formula, 65, 66 Riemann-Lebesgue, 63

79