Sobolev Spaces

Georey R. Burton
December 16, 2013
Contents
1 Preliminaries 1
1.1 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Partial Derivatives and Distributions . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Sobolev spaces 9
2.1 More spaces and boundary values . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Linear Partial Dierential Operators with Constant coecients. . . . . . . . . . 14
2.3 Sobolev embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 Regularisation and approximation 29
3.1 Localisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 Embeddings on Smooth Bounded Domains 41
i
Recommended literature.
1) L. C. Evans, Partial Dierential Equations, Graduate Studies in Mathematics 19, Amer-
ican Mathematical Society, 2nd Edn. 2010;
2) R. A. Adams & J.J.F. Fournier, Sobolev Spaces, Elsevier, 2nd Edn. 2003;
3) V. G. Mazya & T.O. Shaposhnikova, Sobolev spaces, Springer, 2nd Edn. 2011. (For
reference.)
1 Preliminaries
1.1 Inequalities
will always denote a positive measure - think of Lebesgue measure /
N
, possibly with a positive
density function.
Conjugate exponents. p, q [1, ] with
1
p
+
1
q
= 1 or p = 1 and q = , or p = and
q = 1, are called conjugate exponents.
Youngs inequality. If 1 < p, q < with
1
p
+
1
q
= 1 then
xy
1
p
x
p
+
1
q
y
q
(1)
for all x 0 and y 0.
Proof. Write (1) as xy
1
p
x
p

1
q
y
q
. Then maximise LHS over x for xed y.
Holders Inequality. Let p, q [1, ] be conjugate exponents, f L
p
(X, ), g L
q
(X, ).
Then
_
X
[fg[d |f|
p
|g|
q
.
Proof. Put x =
[f(z)[
|f|
p
, y =
[g(z)[
|g|
q
in Youngs inequality and integrate. p = 1, q = is
trivial.
1
Minkowskis Inequality. For 1 p , f, g L
p
(X, ),
|f + g|
p
|f|
p
+|g|
p
.
Proof. Firstly consider the case p = 1. We have [f + g[ [f[ +[g[ a.e., hence
_
X
[f + g[d
_
X
[f[d +
_
X
[g[d.
Secondly consider the case 1 < p < . For x X we have
[f(x) + g(x)[
p
(2 max[f(x)[, [g(x)[)
p
2
p
([f(x)[
p
+[g(x)[
p
),
hence
_
X
[f + g[
p
d 2
p
__
X
[f[
p
d +
_
X
[g[
p
d
_
< .
Thus f + g /
p
(X, , ). Let q be the conjugate exponent of p. Then
_
X
[f + g[
p
d =
_
X
[f + g[ [f + g[
p1
d

_
X
[f[ [f + g[
p1
d +
_
X
[g[ [f + g[
p1
d

_
__
X
[f[
p
d
_
1/p
+
__
X
[g[
p
d
_
1/p
_
__
X
[f + g[
(p1)q
d
_
1/q
(by Holders inequality)
= ([[f[[
p
+[[g[[
p
)
__
X
[f + g[
p
d
_
1/q
.
If [[f + g[[
p
> 0 we can divide by [[f + g[[
p/q
p
to obtain
[[f + g[[
p
[[f[[
p
+[[g[[
p
,
whereas if [[f + g[[
p
= 0 the result is trivial.
Finally consider the case p = . For almost every x X we have
[f(x) + g(x)[ [f(x)[ +[g(x)[ esssup[f[ + esssup[g[ = [[f[[

+[[g[[

.
Thus f + g /

(X, , ) and
[[f + g[[

[[f[[

+[[g[[

.
2
Generalised Holders Inequality. Suppose p
1
, . . . , p
n
(1, ),
1
p
1
+ . . . +
1
p
n
= 1, u
i

L
p
i
(X, ), i = 1, . . . , n. Then
_
X
[u
1
u
2
u
n
[d |u
1
|
p
1
|u
n
|
pn
.
Proof. Exercise.
Interpolation Inequality. Suppose 1 p < q < r < and choose 0 < < 1 such that
1
q
=

p
+
1
r
. Suppose u L
p
(X, ) L
r
(X, ). Then
|u|
q
|u|

p
|u|
1
r
Proof. Exercise.
Jensens Inequality for sums. Let I R be an open interval, let : I R be a convex
function, let x
1
, . . . , x
n
I and let
i
0 for 1 i n with
1
+
n
= 1. Then

_
n

i=1

i
x
i
_

n

i=1

i
(x
i
).
Proof. By induction from the denition of convexity.
Jensens Inequality for functions. Let I R be an open interval, let : I R be a convex
function and let be a probability measure on X ( 0, (X) = 1). Then for u L
1
(X, )
taking values in I we have

__
X
u(x)d(x)
_

_
X
(u(x))d(x).
Proof. Recall that is everywhere subdierentiable, that is, for every x I there is at least
one real such that
(y) (x) + (y x) y I,
and so is the pointwise supremum of all the ane functionals on R dominated by .
Suppose rstly that , R s.t.
(s) = s + (s) s R.
Then

__
ud
_
=
_
X
ud + =
_
X
(u + )d
_
X
u d.
Taking the supremum over all such ane functionals dominated by , we obtain

__
X
ud
_

_
X
u d.
3
The AM-GM inequality. (x
1
x
2
x
n
)
1/n
(x
1
+ +x
n
)/n for positive x
1
, . . . , x
n
follows
by applying Jensens inequality for sums to the convex function log on (0, ).
1.2 Partial Derivatives and Distributions
Integrals are with respect to /
N
.
Denition. The support of a real-valued function f, supp f = x [ f(x) ,= 0.
Notation for partial derivatives on R
N
.
For 1 i N write D
i
=

x
i
. Write N
0
= N 0. Any = (
1
, . . . ,
N
) N
N
0
is called a
multi-index of degree [[ =
1
+ +
N
. Write ! =
1
!
N
! and D

= D

1
1
D

N
N
.
Note. 0 = (0, . . . , 0) N
N
0
, [0[ = 0, and D
0
u = u.
If u has continuous partial derivatives of order m, we have equality of cross-derivatives for
orders up to m, so the order of dierentiation in D

u for [[ m is unimportant.
Leibnizs Theorem. If u, v are m-times continuously dierentiable functions of N real vari-
ables then, for 0 [[ m,
D

uv =

0
_

_
D

uD

v
where
_

_
=
!
!( )!
and signies
i

i
for i = 1, . . . , N.
For R
N
open, C

() denotes the set of real functions on that have continuous

partial derivatives of all orders.
D() = C

c
() denotes the set of all u C

() such that supp u is a compact subset of

. Elements of D() are called test functions.
Example.
J(x) =
_
ke

1
1|x|
2
[x[ < 1
0 [x[ 1,
k is chosen such that
_
R
N
J = 1.
J

(x) =
N
J(
1
x), x R
N
and > 0.
Then J

D(R
N
) and is known as the standard mollier.
4
Convention. If D() then = 0 on R
N
.
Convergence of test functions.
We say that
n

0
in D() as n if there exists a compact set K such that
supp
n
K for all n N and D

n
D

0
uniformly on K as n , for every N
N
0
.
Denition of Distributions.
Given R
N
open, a distribution on is a real linear functional on D() (sequentially)
continuous with respect to convergence of test functions. u, denotes the value taken by the
distribution u at the test function . The set of distributions on is denoted D

().
Remarks
1. See Walter Rudins Functional Analysis for an account of a topology on D() that gives rise
to this notion of convergence of test functions. Linear functionals are shown to be continuous
i they are sequentially continuous.
2. If D() and N
N
0
then D

is also a test function.

3. D

: D() D() is linear and sequentially continuous.

Examples
1. We call a measurable function u on locally integrable (u L
1
loc
()) if
_
K
[u[ < for every
K compact. A locally integrable u gives rise to a distribution by
u, =
_

u D().
This is well dened since has compact support and u is integrable on compact sets, and
linear. If
n
in D() choose a compact K containing the supports of all the
n
.
Then
[u,
n
u, [ |u|
L
1
(K)
|
n
|

0
by uniform convergence on K. Later well show that dierent u give rise to dierent distribu-
tions.
2. Fix z and dene

z
, = (z) D().
This is well-dened and linear. If
n
in D() then
n
(z) (z), hence
z
is continuous
(Dirac -function).
3. Let R
1
, z ,

z
, =

(z) D()
denes a distribution, called a dipole.
5
Lemma 1.1. Suppose R
N
is open and u C
1
(). Then
_

(D
i
u) =
_

uD
i
D().
Proof. Assume u, are zero outside . Then u C
1
(R
N
) even if u is not in C
1
(R
N
). So
_

(D
i
u) =
_
R
N
(D
i
(u) uD
i
) =
_
rB(0)
div (u e
i
) uD
i
) = 0
_
R
N
uD
i
=
_

uD
i

where e
i
is the unit vector in the positive x
i
direction and we have applied the Divergence
Theorem on a large ball rB(0) whose interior contains the support of .
Note. If [[ = m and u C
m
() then
_

u = (1)
m
_

uD

D().
Denition. Let u D

() and N
N
0
. Dene D

u by
D

u, = (1)
||
u, D

D().
Lemma 1.2. (i) If u D

() and N
N
0
then D

u D

().
(ii) If , N
N
0
, u D

() then
D

u = D
+
u = D

u.
Proof. (i) D

u is the composition of D

: D() D() which is linear and sequentially

continuous with u : D() R which is linear and sequentially continuous. So D

u D

().
(ii) Consider D(). Then
D

u, = (1)
||
D

u, D

= (1)
||+||
u, D

= (1)
||+||
u, D
+
by equality of cross-derivatives for smooth functions
= D
+
u, .
So D

u = D
+
u = D
+
u = D

u.
6
Examples.
1. Let
u(x) = x
+
=
_
0, x 0
x, x > 0
x R.
For D(R)
u

, = u,

=
_

u(x)

(x)dx
=
_

0
x

(x)dx
=
_
x(x)

0
+
_

0
1(x)dx (integrating by parts)
= 0 +
_

0
1(x)dx ( has compact support)
=
_

H(x)(x)dx,
where
H(x) =
_
0 x 0
1 x > 0
(Heaviside Step function).
So u

= H.
2. Dierentiate H. For D(R)
H

, = H,

=
_

H(x)

(x)dx
=
_

0
1

(x)dx
=
_
(x)

0
= (0).
Thus H

, = (0) =
0
(). So H

=
0
(Dirac delta function).
3. Dierentiate
0
. For D(R)

0
, =
0
,

(0) Dipole.
4. Le be a Radon measure on (Borel measure that assigns nite measure to compact sets).
Dene
, =
_

d D().
7
Then gives rise to a distribution, for if f
n
in then there is a compact K that
contains the supports of the
n
and and
n
uniformly, so
_

n
d
_

d
and the linearity follows from properties of the integral.
Connections with classical derivatives.
1. Let f L
1
loc
(a, b), x
0
(a, b),
F(x) =
_
x
x
0
f(x)dx, a < x < b.
(a) Then F is continuous and F

= f in the sense of distributions (proved later Proposition

1.4).
(b) F

= f classically a.e. in (a, b) (tricky - see W. Rudins Real and Complex Analysis,
Ch. 8).
2. Let F be continuous on (a, b).
(a) If F

= f L
1
loc
(a, b) in the sense of distributions, then
F(x) =
_
x
x
0
f + c (x
0
(a, b))
for some c R (to be proved later).
(b) If F

= f L
1
loc
(a, b) classically a.e., we cannot conclude F(x) =
_
x
x
0
f +c. See Cantor
Function (Devils Staircase) in Rudins Real and Complex Analysis, Ch. 8.
Lemma 1.3. Let R
N
be open. Then
(i) D() is dense in L
p
() for 1 p < .
(ii) If u L
1
loc
() and
_

u = 0 D()
then u = 0 a.e.
(Hence dierent locally integrable functions u give dierent distributions.)
Proof. Later.
Proposition 1.4. Let f L
1
loc
(a, b) and F(x) =
_
x
x
0
f, some x
0
(a, b). Then F

= f in the
sense of distributions.
8
Proof. If f is continuous then F is continuously dierentiable with F

= f and the result follows

from Lemma 1.1.
Now consider the general case. Firstly, choose a sequence f
n

n=1
in D(a, b) converging to
f in L
1
([, ]) for all [, ] (a, b) (by Lemma 1.3(i)). Dene F
n
(x) =
_
x
x
0
f
n
for some xed
x
0
(a, b). Then F

n
= f
n
both classically and in the sense of distributions. For D(a, b)
F

n
, = f
n
, =
_
b
a
f
n

_
b
a
f (by Holders ineq. on compact set supp )
= f, .
Also,
F

n
, = F
n
,

=
_
b
a
F
n

_
b
a
F

(since F
n
F uniformly on supp )
= F

, .
Thus,
F

, = f, D(a, b).
So F

= f as distributions.
2 Sobolev spaces
Denition. Let R
N
be open, 1 p , m N. Dene the Sobolev space W
m,p
() by
W
m,p
() =
_
u [ D

u L
p
() for all N
N
0
s.t. 0 [[ m
_
.
With the obvious real vector space structure, dene the norm on W
m,p
() by
|u|
m,p
=
_
_

0||m
_

[D

u[
p
_
_
1
p
1 p <
|u|
m,
= max
0||m
|D

u|

.
Theorem 2.1. Let R
N
be open, 1 p , m N. Then W
m,p
() is a Banach space.
9
Proof. Let u
n

n=1
be a Cauchy sequence in W
m,p
(). For each N
N
0
, 0 [[ m,
|D

u
n
D

u
k
|
p
|u
n
u
k
|
m,p
.
Hence D

u
n

n=1
is a Cauchy sequence in L
p
(), and converges to some v

L
p
().
Now u
n
v
0
, and for D()
u
n
, D

=
_

u
n
D

Holder

v
0
D

= v
0
, D

and
D

u
n
, =
_

u
n

Holder

.
Since
D

u
n
, = (1)
||
u
n
, D

we obtain
v

, = (1)
||
v
0
, D

= D

v
0
, .
So, by uniqueness of function representing D

v
0
(Lemma 1.3(ii)),
D

v
0
= v

.
Thus D

u
n
D

v
0
in | |
p
for all 0 [[ m, that is, u
n
v
0
in W
m,p
().
Theorem 2.2. Let R
N
be open, m N.
(i) If 1 p < then W
m,p
() is separable.
(ii) If 1 < p < then W
m,p
() is reexive.
Proof. Write A = N
N
0
[ 0 [[ m, and write Y = L
p
()
A
, that is the set of maps
from A to L
p
(), whose members we think of as vectors (v

)
A
whose components belong to
L
p
() and are indexed by elements of A. For v Y set
|v|
Y
=
_

0||m
|v

|
p
p
_1
p
(note the sum is over A)
which makes Y into a Banach space. The map T : W
m,p
() Y
(Tu)

= D

u A, for u W
m,p
(),
is a linear isometry of W
m,p
() to a linear subspace X of Y . Moreover, W
m,p
() is complete,
so X is complete with | |
Y
, so X is closed in Y .
10
(i) If 1 p < then L
p
() is separable, so Y is separable, so X is separable, so W
m,p
()
is separable.
(ii) Suppose 1 < p < . Then L
p
() is reexive (L
p
() is isometric under the natural
map onto L
p
()

; equivalently, the closed unit ball of L

p
() is compact in the weak topology).
Hence Y is reexive, hence X is reexive (closed linear subspace of a reexive space), hence
W
m,p
is reexive.
2.1 More spaces and boundary values
A proper theory of boundary values for Sobolev functions requires smoothness assumptions on
; see Trace Theorem later on. A rough-and-ready denition of W
m,p
() functions whose
derivatives of orders 0, 1, . . . , m1 vanish on the boundary, is as follows:
Denition. For R
N
open, m N, dene W
m,p
0
() to be the closure of D() in W
m,p
().
This is frequently a convenient space for studying Dirichlet problems for PDE.
Denition. H
m
() = W
m,2
() is a Hilbert space with scalar product
u, v
m
=

0||m
_

uD

v u, v H
m
().
Write H
m
0
() = W
m,2
0
().
Theorem 2.3. Let R
N
be open, 1 p . Then there is a constant c = c(p, N), such
that for u W
1,p
(),
c
1
|u|
1,p,
|u A|
1,p,A
1

c|u|
1,p,
for all A O(N).
Proof. Consider A = O(N), u W
1,p
() and v = u A
T
= u A
1
L
p
(). Let D(A)
11
and set and = A D(). Let e
i
denote the unit vector in the positive x
i
direction. Then
D
i
v, =
_
A
v(y)D
i
(y)dy
=
_
A
u(A
T
y)(D
i
)(AA
T
y)dy
=
_

u(x)D
i
(Ax)dx (Ax = y, [ det A[ = 1)
=
_

u(x)D
i
( A
T
)(Ax)dx
=
_

u(x)D( A
T
)(Ax)e
i
dx (D = derivative)
=
_

u(x)D( A
T
)(Ax)e
i
dx
=
_

u(x)D(x)DA
T
(Ax)e
i
dx (Chain rule)
=
_

u(x)D(x)A
T
e
i
dx
=
_

u(x)e
T
i
A(x)dx (transposing real integrand)
=
_

e
T
i
Au(x)(x)dx
=
_
A
e
T
i
Au(A
T
y)(A
T
y)dy
so
v = A(u) A
T
.
So v L
p
(A) and
|v|
p
c|u|
p
,
where
c = sup
AO(N),||p=1
[A[
p
from which the result follows.
Remark. This shows we are free to rotate axes, at the cost of replacing the Sobolev norm by
an equivalent norm, bounded by a constant independent of the rotation. Recall - two norms
| |
1
and | |
2
are equivalent if there is a constant c > 0 such that
c
1
|x|
1
|x|
2
c|x|
1
for all x X. Two norms are equivalent if and only if they give rise to the same convergent
sequences.
12
Theorem 2.4 (Poincares Inequality). Let R
N
be open, suppose lies between two parallel
hyperplanes a distance l > 0 apart and let 1 p , m N. Then there exists c =
c(l, p, m, N) > 0 such that
|u|
m,p
c
_

||=m
|D

u|
p
p
_1
p
when 1 p <
and
|u|
m,
c max
||=m
|D

u|

when p =
for all u W
m,p
0
().
Proof. Firstly suppose m = 1. Consider u D(). Using Theorem 2.3 we can assume the axes
to be chosen in such a way that (x
1
, . . . , x
N
) [ 0 < x
N
< l. Then for x
u(x) =
_
x
N
0
D
N
u(x

,
N
)d
N
x = (x

, x
N
) R
N1
R,
so
[u(x)[
Holder
|1
[0,x
N
]
|
q
|D
N
u(x

, )|
p
q conjugate to p.
Case 1 p < . Then
[u(x)[ x
1
1
p
N
__
l
0
[D
N
u(x

,
N
)[
p
d
N
_
1
p
So
_

[u(x)[
p
dx
__
l
0
x
p1
N
dx
N
___
R
N1
_
l
0
[D
N
u(x

,
N
)[
p
d
N
dx

_
(u = 0 outside )
thus
|u|
p
p

l
p
p
|D
N
u|
p
p
so
|u|
p

l
p
1/p
|D
N
u|
p
.
Case p = . We have
[u(x)[ x
N
|D
N
u(x

, )|

,
so taking sup over x
|u|

l|D
N
u|

.
In either case,
|u|
p
lc(p, N)|u|
p
.
13
Applying repeatedly, we obtain
|u|
1,p
const |u|
p
.
.
.
|u|
m,p
const
_

||=m
|D

u|
p
p
_1
p
(1 p < )
|u|
m,
const max
||=m
|D

u|

,
for all u D(). By density the inequality holds for all u W
m,p
0
(), since both the LHS and
RHS are continuous in | |
m,p
.
Remark. Poincares inequality enables us to dene an equivalent norm on W
m,p
0
() when
has nite width (in particular when is bounded).
|u| =
_

||=m
|D

u|
p
p
_1
p
(1 p < )
|u| = max
||=m
|D

u|

(p = ).
In particular
u, v =

||=m
_

uD

v
denes an equivalent scalar product on H
m
0
().
2.2 Linear Partial Dierential Operators with Constant coecients.
.
L =

0||m
a

,
where a

are constants, is a linear partial dierential operator of order (at most) m with
constant coecients.
If f D

() then u D

() is a solution of Lu = f if
u,

0||m
(1)
||
a

= f, for all D().

The operator
L

0||m
(1)
||
a

is the adjoint of L.
14
Example.
=
N

i=1
D
2
i
.
For a distribution u and test function
u, =
N

i=1
D
i
u, D
i
= u, .
Application. Suppose R
N
is a bounded open set, f L
2
(). Show that the boundary
value problem
u = f
u H
1
0
()
_
(BVP)
has exactly one solution.
Write H = H
1
0
() and set
u, v
H
=
_

u v u, v H
which denes an equivalent scalar product on H.
For u H,
u = f
if and only if
_

u =
_

f D()
if and only if
_

u v =
_

fv v H
by density of D() in H, since LHS is the scalar product of H, and the RHS denes a bounded
linear functional of v H; to see this, put
(v) =
_

fv v H
then
[(v)[
_

[f[[v[ |f|
2
|v|
2
const |f|
2
|v|
H
by Poincares inequality. So H

, and the Riesz Representation Theorem for Hilbert spaces

shows
(v) = u
0
, v
H
v H
for exactly one u
0
H. Now u
0
is the unique solution of the BVP.
15
Remark. is a second order partial dierential operator, but u
0
H
1
0
() at rst sight only
has rst order derivatives. The question Does u
0
have second order derivatives? belongs
to Regularity Theory. In fact u
0
H
2
loc
() in general, and u
0
H
2
() if the boundary is
suciently smooth. This is typical of elliptic PDE. The situation is not so good for hyperbolic
PDE (e.g. the wave equation).
2.3 Sobolev embeddings
Theorem 2.5. Let < a < b < , 1 p . Then every element of W
1,p
0
(a, b) has a
continuous representative, and the following embeddings are well-dened bounded linear maps:
W
1,1
0
(a, b) C([a, b])
W
1,
0
(a, b) C
0,1
([a, b]) (Lipschitz continuous functions)
W
1,p
0
(a, b) C
0,
([a, b]) (Holder continuous functions), = 1
1
p
, 1 < p < .
Proof. Case p = 1.
For D(a, b)
[(x)[ =

_
x
a

|
1
(a < x < b)
so ||
sup
||
1,1
.
Case p = .
For D(a, b)
[(x) (y)[ =

_
y
x

[y x[|

so
||
C
0,1 = ||
sup
+ sup
a<x<y<b
[(x) (y)[
[x y[
||

+|

2||
1,
.
Case 1 < p < .
For D(), a < x < y < b,
[(x) (y)[
_
y
x
[

[
Holder
[x y[
1
q
|

|
p
[(x)[ (b a)
1
q
|

|
p
so with =
1
q
= 1
1
p
||
C
0, = ||
sup
+ sup
a<x<y<b
[(x) (y)[
[x y[

const ||
1,p
.
16
We have proved inequalities of the form
||
X
c||
1,p
D(a, b)
with
X = C([a, b]) when p = 1
X = C
0,1
([a, b]) when p =
X = C
0,
([a, b]) when 1 < p < and = 1
1
p
.
For general u W
1,p
0
(), choose a sequence
n
in D() converging to u in | |
1,p
. Then
|
n
u|
p
0, so passing to a subsequence
n
u a.e. Also
n
is Cauchy in | |
1,p
,
and by above inequalities Cauchy in | |
X
. Then by completeness
n

n=1
converges in X to
v say. Then v X, so v is (uniformly) continuous, and
n
v uniformly, so v = u a.e.
Thus v is a continuous representative for u, hence W
1,p
0
(a, b) X. Finally, | |
X
and | |
1,p
are
continuous functions in | |
X
and | |
1,p
respectively and
n
u in both norms, so the inequality
| |
X
c| |
1,p
holds on the whole of W
1,p
0
(a, b).
In the results proved above for N = 1 the restrictions to bounded intervals and W
m,p
0
can
be avoided.
In higher dimensions we dont generally get continuous functions;
The embeddings are bounded linear operators, which for certain domains, and for certain
values of p, are compact.
Some results in higher dimensions require regularity assumptions on the boundary.
Some results require boundedness of the domain.
We now consider the higher-dimensional cases.
17
Theorem 2.6 (Sobolevs Inequality). Let m 1, N 2, p 1, mp < N, p

=
Np
N mp
.
Then
|u|
p
c
_

||=m
|D

u|
p
p
_1
p
for all u C
m
c
(R
N
) ( D(R
N
)).
Proof of the inequality. We consider the following cases:
Case m = 1, p = 1, p

= N/(N 1). For u C

1
c
(R
N
), x R
N
,
u(x) =
_
x
j

D
j
u(x
1
, . . . ,
j
, . . . , x
N
)d
j
so
[u(x)[
_

[D
j
u(x)[dx
j
so
[u(x)[
N
N1

1jN
__

[D
j
u(x)[dx
j
_ 1
N1
.
The rst term of the product is independent of x
1
and the remaining terms are each functions
of N 1 variables including x
1
. So
_

[u(x)[
N
N1
dx
1

__

[D
1
u(x)[dx
1
_ 1
N1

j=1
__

[D
j
u(x)[dx
j
_ 1
N1
dx
1
.
On the RHS the second term is the integral of a product of N 1 functions. Applying the
generalised Holder inequality
_
v
1
v
N1
|v
1
|
N1
|v
N1
|
N1
we obtain
_

[u(x)[
N
N1
dx
1

__

[D
1
u(x)[dx
1
_ 1
N1

j=1
__

[D
j
u(x)[dx
j
dx
1
_ 1
N1
,
thus we have taken the product outside the integral. We repeat this process over all values of
j; at each step one factor in the RHS is independent of x
j
, and we apply the generalised Holder
inequality to the integral of the product of the remaining N 1 factors. We end up with
_
R
N
[u[
N
N1

j=1
__
R
N
[D
j
u[
_ 1
N1
so taking the (N 1)/N-th power yields
|u|
p

N

j=1
__
R
N
[D
j
u[
_1
N
.
18
Now by the AM-GM inequality
|u|
p

1
N
N

j=1
|D
j
u|
1
.
This proves the case m = 1, p = 1.
Case m = 1, 1 < p < N, p

=
Np
N p
. Let u C
1
c
(R
N
). Let v = [u[
s
where s > 1 is to be
chosen later; note that v C
1
c
(R
N
). Applying the above inequality to v,
__
R
N
[u[
sN
N1
_N1
N
c.
_
R
N
[v[
1
= c.
_
R
N
[u[
s1
[u[
1
c|u|
p
__
R
N
[u[
(s1)q
_1
q
where
1
q
+
1
p
= 1. We choose s so that
sN
N 1
= (s 1)q, which yields
s =
(N 1)p
N p
.
Thus
__
R
N
[u[
sN
N1
_N1
N

1
q
c|u|
p
.
Then
sN
N 1
=
Np
N p
= p

and
N 1
N

1
q
=
N p
Np
=
1
p
so
|u|
p
c|u|
p
.
This completes the case m = 1, 1 < p < N.
General case. Induction on m. The initial case m = 1 is done. Assume true for m 1.
Consider N
N
0
, [[ = m1, u C
m
C
(R
N
). Then by the initial case
|D

u| Np
Np
c|D

u|
p
.
Thus by the inductive hypothesis
|u| NNp/(Np)
N(m1)Np/(Np)
c

||=m1
|D

u| Np
Np
c

||=m
|D

u|
p
that is
|u| Np
Nmp
c

||=m
|D

u|
p
as required, since all norms on a Euclidean space are equivalent. This completes the inductive
step and we are done.
Corollary 2.7. Let N 2, m 1, mp < N, p

=
Np
N mp
, ,= R
N
open. Then
W
m,p
0
() is embedded in L
p

() and the embedding is a bounded linear map.

19
Proof. Let c be the constant in the Sobolev inequality for the given N, m, p. Thus
||
p
c
_
_

||=m
|D

|
p
p
_
_
1
p
c||
m,p
D().
Consider u W
m,p
0
(); so u is the limit in | |
m,p
of a sequence
n
of test functions. We can
also assume
n
u a.e. Now
n
is Cauchy in | |
m,p
and therefore Cauchy in | |
p
, so
n

converges in L
p

, and the limit must equal u a.e. Thus u L

p

. Continuity of | |
p
on L
p

and
| |
m,p
on W
m,p
now ensure
|u|
p
c|u|
m,p
.
Denitions. Let ,= R
N
be open, m N, 0 < 1.
C() = continuous functions on
C
m
() = functions m-times continuously dierentiable on
C
B
() = bounded continuous functions on
C
m
B
() = u [ D

u C
B
(), 0 [[ m
C() = bounded uniformly continuous functions on
C
m
() = u [ D

u C(), 0 [[ m
C
0,
() = functions on of Holder class
C
m,
() = u [ D

u C
0,
(), 0 [[ m
Then
C
m,
() C
m
() C
m
B
() C
m
(),
functions u C
m
() have D

u, 0 [[ m, continuously extendable to , Holder continuous

in case C
m,
. Note C(R
N
) ,= C(R
N
). Then C
m
B
() and C
m
() are Banach spaces with
|u|
C
m
B
()
= max
0||m
|D

u|
sup
.
C
m,
() is a Banach space with
|u|
C
m,
()
= max
0||m
|D

u|
sup
+ max
0||m
sup
x,y,x=y
[D

u(x) D

u(y)[
[x y[

.
Theorem 2.8 (Morreys Inequality). Suppose N 2, N < p < . Then there is a constant
c = c(N, p) such that
|u|
C
0, c|u|
1,p
u C
1
(R
N
) W
1,p
(R
N
),
where = 1
N
p
.
20
Proof. Step 1. We show that

_
B(x,r)
[u(y) u(x)[dy c
_
B(x,r)
[u(y)[
[y x[
N1
dy
where
_
denotes the mean.
Preliminary calculation
_
B(0,1)
[u(x + sw) u(x)[dw =
_
B(0,1)

_
s
0
d
dt
u(x + tw)dt

dw

_
B(0,1)
_
s
0
[u(x + tw) w[dtdw

_
B(0,1)
_
s
0
[u(x + tw)[dtdw
=
_
s
0
_
B(0,1)
[u(x + tw)[dwdt
=
_
s
0
_
B(0,t)
[u(x + w)[
1
t
N1
dwdt
=
_
B(0,s)
[u(x + w)[
[w[
N1
dt
=
_
B(x,s)
[u(y)[
[x y[
N1
dy.
Now
_
B(x,r)
[u(y) u(x)[dy =
_
r
0
_
B(0,s)
[u(x + w) u(x)[dwds
=
_
r
0
s
N1
_
B(0,1)
[u(x + sw) u(x)[dwds

_
r
0
s
N1
_
B(x,s)
[u(y)[
[y x[
N1
dyds

__
r
0
s
N1
ds
___
B(x,r)
[u(y)[
[y x[
N1
dy
_
=
r
N
N
_
B(x,r)
[u(y)[
[y x[
N1
dy.
So dividing by r
N

_
B(x,r)
[u(y) u(x)[dy c
_
B(x,r)
[u(y)[
[y x[
N1
dy.
Step 2. Estimate |u|
sup
.
21
For x R
N
[u(x)[
_
B(x,1)
[u(x) u(y)[dy +
_
B(x,1)
[u(y)[dy
c
_
B(x,1)
[u(y)[
[x y[
N1
dy +[B(x, 1)[
1
q
1
|u|
p
(by Step 1 and Holder
1
q
+
1
p
= 1)
|u|
p
__
B(x,1)
[x y[
(N1)q
dy
_1
q
+ c|u|
p
(since (N 1)q < N)
c|u|
p
+ c|u|
p
c|u|
1,p
.
Step 3. Holder estimate for [u(x) u(y)[.
Consider x, y R
N
, [x y[ = r > 0. For any z R
N
[u(x) u(y)[ [u(x) u(z)[ +[u(z) u(y)[.
So averaging over a region W of nite positive measure
[u(x) u(y)[
_
W
[u(x) u(z)[dz +
_
W
[u(z) u(y)[dz.
Choose W
r
= B(x, r) B(y, r) (c.f. Figure 1).
Figure 1: W
r
= B(x, r) B(y, r)
Notice that W
r
is similar to W
1
= B(0, 1) B(e, 1) where e is any unit vector. So
[W
r
[ = r
N
[W
1
[.
Now
[W
r
[
_
Wr
[u(x) u(z)[d [B(x, r)[
_
B(x,r)
[u(x) u(z)[dz.
So

_
Wr
[u(x) u(z)[dz
[B(x, r)[
[W
r
[

_
B(x,r)
[u(x) u(z)[dz
22
thus

_
Wr
[u(x) u(z)[dz c
_
B(x,r)
[u(x) u(z)[dz (since
|B(x,r)|
|Wr|
is independent of r).
Now using Step 1,

_
Wr
[u(x) u(z)[dz const
_
B(x,r)
[u(z)[
[x z[
N1
dz c|u|
p
__
B(x,r)
[x z[
(N1)q
dz
_1
q
.
Now
_
B(x,r)
[x z[
(N1)q
dz =
_
r
0
_
B(x,s)
s
(N1)q
dzds = c
_
r
0
s
N1
s
(N1)q
ds
= cr
(N1)(1q)+1
= cr
pN
p1
,
since
(N 1)(1 q) + 1 = (N 1)
_
1
p
p 1
_
+ 1 = (N 1)
(1)
p 1
+ 1 =
p N
p 1
.
So

_
Wr
[u(x) u(z)[dz c|u|
p
r
pN
p1

1
q
= c|u|
p
r
1
N
p
,
since
p N
p 1

1
q
=
p N
p 1
_
1
1
p
_
=
p N
p
= 1
N
p
.
Similarly

_
Wr
[u(z) u(y)[dz c|u|
p
r
1
N
p
so
[u(x) u(y)[ c|u|
p
[x y[
1
N
p
.
That is,
[u(x) u(y)[
[x y[

c|u|
p
.
Now
|u|
C
0, = |u|
sup
+ sup
x=y
[u(x) u(y)[
[x y[

c|u|
W
1,p
(R
N
)
.
Theorem 2.9. Let N 2, m N, m < N, mp = N, 1 q < . Then there exists a constant
c = c(N, m, q) such that
(i) |u|
q
c[[
1/q

||=m
|D

u|
p
for all u C
m
c
(R
N
), where = x R
N
[ u(x) ,= 0;
(ii) |u|
q
c|u|
m,p
, for all u C
m
(R
N
) W
m,p
(R
N
) and q > p.
23
Proof. (i) Case
N
Nm
q < . Choose r, 1 r < p, such that r

=
Nr
Nmr
= q. Then by the
Sobolev inequality, for u C
m
c
(R
N
)
|u|
q
c

||=m
|D

u|
r
.
Now
_
R
N
[D

u[
r
|1

|
s

__
R
N
[D

u[
rs
_
1/s
where rs = p, 1/s

+ 1/s = 1, rs

= pr/(p r) = (Nr/m)/(N/m r) = (Nr)/(N mr) = q

and = x [ u(x) ,= 0. So
|D

u|
r
[[
1/rs

|D

u|
p
= [[
1/q
|D

u|
p
.
Thus
|u|
q
c[[
1/q

||=m
|D

u|
p
.
(i) Case 1 q <
N
Nm
= t. We have
_
R
N
[u[
q
|1

|
s
|[u[
q
|
s
(where qs = t and 1/s

+ 1/s = 1)
so
|u|
q
[[
1/s

q
|u|
t
= [[
1/q1/t
|u|
t
.
Using the previous case to estimate |u|
t
we get
|u|
q
c[[
1/q1/t
[[
1/t

||=m
|D

u|
p
= c[[
1/q

||=m
|D

u|
p
.
(ii) Suppose u C
m
(R
N
) W
m,p
(R
N
), q > p. Construct a partition of unity as follows. Let
C

(R) satisfy () > 0 for 1 < < 1 and () = 0 for [[ 1. For k = (k

1
, . . . , k
N
)
Z
N
let

k
(x) =
N

i=1
(x
i
k
i
)
which lives on
Q
k
= (1 + k
1
, 1 + k
1
) (1 + k
N
, 1 + k
N
).
Note almost every x R
N
belongs to 2
N
of the cubes Q
k
, and all points belong to at least one.
Dene

k
=

k

lZ
N

l
24
which are smooth functions adding up to the constant function 1, and all but nitely many
vanish outside any bounded set. Thus
u =

kZ
N

k
u.
Now
[u(x)[
q
=

k
(x)u(x)

q
= 2
Nq

k
2
N

k
(x)u(x)

q
2
Nq

k
2
N
[
k
(x)u(x)[
q
by Jensens inequality, so
_
R
N
[u[
q
2
N(q1)
_
R
N

k
[
k
u[
q
.
Now
_
Q
k
[
k
u[
q
c
_
_

||=m
_
Q
k
[D

(
k
u)[
p
_
_
q/p
(by (i))
c
_
_

0||m
_
Q
k
_

0
[D

u[
_
p
_
_
q/p
c
_
_

0||m
_
Q
k
[D

u[
p
_
_
q/p
where we have dierentiated by Leibnizs theorem and used the independence of |D

k
|
sup
from k for each , , then applied Jensens inequality.
Then
_
R
N
[u[
q
c

kZ
N
_
_

0||m
_
Q
k
[D

u[
p
_
_
q/p
c
_
_

kZ
N

0||m
_
Q
k
[D

u[
p
_
_
|u|
qp
m,p,R
N
= c
_
_
2
N

0||m
_
R
N
[D

u[
p
_
_
|u|
qp
m,p,R
N
= c|u|
q
m,p,R
N
since the family Q
k

kZ
N forms a 2
N
-fold covering of R
N
except for a set of zero measure.
Thus
|u|
q
c|u|
m,p
.
Theorem 2.10. Let I = (a
1
, b
1
) (a
N
, b
N
) be a rectangle in R
N
. Then there is a constant
c, depending only on the edge-lengths of I, such that
(i) |u|
sup
c|u|
N,1
for all u C
N
(I) W
N,1
(I);
25
(ii) |u|
sup
c|u|
N,1
and u(x) 0 as [x[ for all u C
N
(R
N
) W
N,1
(R
N
).
Proof. Case N = 1. For u C
1
(a
1
, b
1
) W
1,1
(a
1
, b
1
), x, y I = (a
1
, b
1
)
[u(x) u(y)[
_
b
1
a
1
[u

[
and, by continuity,there exists x I such that
u( x) = (b
1
a
1
)
1
_
b
1
a
1
u.
So
[u(y)[ [u( x)[ +
_
b
1
a
1
[u

[ (b
1
a
1
)
1
_
b
1
a
1
[u[ +
_
b
1
a
1
[u

[ max(b
1
a
1
)
1
, 1|u|
1,1
.
Inductive step. Assume true in dimension N 1. Consider x, y I and suppose initially
that x and y dier in one coordinate only, say the last. Write x = (x

, x
N
), y = (x

, y
N
) where
x

R
N1
and x
N
, y
N
(a
N
, b
N
). Then
[u(x) u(y)[
_
b
N
a
N
[D
N
u(x

, )[d c(
1
, . . . ,
N1
)
_
b
N
a
N
|D
N
u(, )|
N1,1
d
c(
1
, . . . ,
N1
)|u|
N,1
where the W
N1,1
-norm is taken over an (N 1)-dimensional rectangle and
j
= b
j
a
j
.
In the general case we can choose points x = x
0
, x
1
, . . . , x
N
= y such that x
i
x
i1
is parallel
to the i-th coordinate axis, and apply the above calculation to obtain
[u(x) u(y)[
N

i=1
[u(x
i
u(x
i1
)[ c(
1
, . . . ,
N
)|u|
N,1
.
We can choose x I such that u( x) = [I[
1
_
I
u. Then, for all y I,
[u(y)[ [u( x)[ +[u(y) u( x)[ [u( x)[ + c(
1
, . . . ,
N
)|u|
N,1
(
1

2

N
)
1
_
I
[u[ + c(
1
, . . . ,
N
)|u|
N,1
c(
1
, . . . ,
N
)|u|
N,1
.
This completes the inductive step.
The remaining parts of Theorem 2.10 are an exercise.
Lemma 2.11. If R
N
is open and 0 < < 1, then the embedding C
0,
() C
0,
()
is bounded. [Exercise]
26
Theorem 2.12 (Sobolev Embedding Theorem for W
m,p
0
()). Suppose N 2, R
N
is open,
1 p < , m N. Then the following embeddings are well-dened bounded linear maps:
(i) W
m,p
0
() L
q
(), mp < N, p q p

, p

=
Np
N mp
;
(ii) W
m,p
0
() L
q
(), mp = N, 1 < p q < ;
(iii) W
N,1
0
() C() and W
N,1
0
() L
q
() for 1 q ;
(iv) W
m,p
0
() C
0,
(), mp > N > (m1)p, 0 < m
N
p
;
(v) W
m,p
0
() C
0,
(), (m1)p = N, 0 < < 1.
Proof. First check for test functions u.
(i) mp < N. We have |u|
p
c|u|
m,p
by the Sobolev inequality and |u|
p
|u|
m,p
. We
get L
p

L
q
by interpolation: assume p < q < p

and write
1
q
=

p
+
1
p

for some 0 < < 1;

then
|u|
q
|u|

p
|u|
1
p

hence
|u|
q
c|u|

p
|u|
1
m,p
c|u|
m,p
for u D().
(ii) mp = N, 1 < p < q < . Theorem 2.9 shows |u|
q
c|u|
m,p
for u D() C
m
().
(iii) m = N, p = 1.
Theorem 2.10(ii) shows |u|
sup
c|u|
N,1
for u D() ( W
N,1
(R
N
) C
N
(R
N
)) (and hence
|u|

c|u|
N,1
).
If 1 < q < then |u|
q
q
|u|
1
|u|
q1
sup
c|u|
q
N,1
.
(iv) mp > N > (m1)p. We want to use the Morrey inequality
|v|
C
0,
0
c|v|
1,p
0
for v D()
where
p
0
=
Np
(N (m1))p
(> N since mp > N) and
0
= 1
N
p
0
and the Sobolev inequality
|v|
p
0
c|v|
m1,p
(since (m1)p < N). (2)
27
So for test functions u
|u|
p
0
c|u|
m1,p
(from (2))
c|u|
m,p
and
|u|
p
0
c|u|
m,p
(from (2))
so
|u|
1,p
0
c|u|
m,p
.
Morrey now gives
|u|
C
0,
0
c|u|
m,p
.
By Lemma 2.11, for 0 <
0
|u|
C
0, c|u|
m,p
.
Finally, note that

0
= 1
N
p
0
= m
N
p
.
(v) N = (m1)p. Then, for p q <
|u|
q
c|u|
m1,p
(Theorem 2.9)
c|u|
m,p
|u|
q
c|u|
m1,p
c|u|
m,p
so
|u|
1,q
c|u|
m,p
. (3)
For q > N
|u|
C
0, c|u|
1,q
(Morrey) (4)
where = 1
N
q
; by varying q in the range N < q < we can make take any value,
0 < < 1. Thus from (3) and (4) we have
|u|
C
0, c|u|
m,p
if 0 < < 1 and u D().
So in each of the above cases we have an inequality
|u|
X
c|u|
m,p
for all u D() (5)
where X is L
q
(), C(), or C
0,
() as appropriate.
28
For general u W
m,p
0
() choose a sequence u
n
of test functions converging in | |
m,p
to
u. Then u
n
is Cauchy in | |
m,p
, and therefore Cauchy in | |
X
, so u
n
converges in X to
u say. Passing to a subsequence, u
n
u a.e., and either u
n
u uniformly, or after passing
to a subsequence u
n
u a.e. So u = u a.e. Each side of (5) is continuous on X or W
m,p
as
appropriate. So (5) also holds for u.
3 Regularisation and approximation
Denition. The convolution of two measurable functions u, v on R
N
u v(x) =
_
R
N
u(y)v(x y)dy
when this exists.
Lemma 3.1. (i) If u, v L
1
(R
N
) then u v is dened a.e. on R
N
and
|u v|
1
|u|
1
|v|
1
.
(ii) If u L
1
loc
(R
N
) and v L
1
(R
N
) has compact support, then u v and v u exist a.e. and
u v = v u a.e., and is locally L
1
.
(iii) If u L
1
loc
(R
N
), v, w L
1
(), v, w have compact support then
(u v) w = u (v w) a.e.
(iv) If u L
1
loc
(R
N
) and v C
c
(R
N
) then u v is continuous.
Proof. Not given, by Fubini. Part (iv) exercise.
Lemma 3.2. Suppose u L
1
loc
(R
N
) and D(R
N
). Then
(i) D

(u ) = u D

for all N
N
0
and is continuous, hence u C

(R
N
).
(ii) If D

u L
1
loc
(R
N
) then
D

(u ) = (D

u) .
Proof. (i) Consider rst order partial derivatives; say e is the unit vector in the x
i
direction for
some i. If 0 < [h[ < 1, then
(u )(x + he) (u )(x)
h
=
_
R
N
u(y)
(x + he y) (x y)
h
dy
=
_
U
u(y)
(x + he y) (x y)
h
dy
29
where U = B(x, 1) supp . The integrand converges pointwise to uD
i
, and is dominated by
[u(x)[|D
i
|
sup
which is integrable on the compact set U so we can pass to the limit using the
Dominated Convergence Theorem to get
D
i
(u ) = u D
i
.
Repeated applications give the result for D

.
D

is continuous and has compact support, hence uD

is continuous for every N

N
0
,
hence u C

(R
N
).
(ii) Assume D

u L
1
loc
(R
N
), for some N
N
0
. For D(R
N
)
D

(u )(x) =
_
R
N
u(y)(D

)(x y)dy (by (i))

=
_
R
N
u(y)(1)
||
D

x
(y)dy (
x
(y) = (x y))
= (1)
||
u, D

= D

u,
x

=
_
R
N
D

u(y)(x y)dy (since D

u L
1
loc
)
= (D

u) (6)
Reminder.
J(x) =
_
ke

1
1|x|
2
[x[ < 1
0 [x[ 1
x R
N
with k > 0 chosen so that
_
R
N
J = 1. Take
J

(x) =
N
J(
1
x) (Standard mollier)
so
_
R
N
J

= 1 and supp J

= B(0, ).
Denition. If u L
1
loc
(R
N
) we call J

u the mollication or regularisation of u.

Note.
(1) J

u C

(R
N
).
(2) J

u(x) =
_
R
N
u(y)J

(x y)dy =
_
B(x,)
u(y)J

(x y)dy.
So J

u(x) is a weighted mean of u over B(x, ).

30
(3) D

(J

u) = (D

) u.
(4) If D

u L
1
loc
(R
N
) then D

(J

u) = J

u.
Theorem 3.3. Let 1 p .
(i) |J

u|
p
|u|
p
for u L
p
(R
N
) and |J

u|
m,p
|u|
m,p
for u W
m,p
(R
N
).
(ii) If 1 p < then |J

u u|
p
0 as 0 for u L
p
(R
N
) and |J

u u|
m,p
0
as 0 for u W
m,p
(R
N
).
(iii) if 1 p < then D(R
N
) is dense in L
p
(R
N
) and in W
m,p
(R
N
) so W
m,p
0
(R
N
) =
W
m,p
(R
N
).
Proof. (i) Case 1 p < .
[J

u(x)[
p
=

_
R
N
u(y)J

(x y)dy

_
R
N
[u(y)[
p
J

(x y)dy (Jensens Inequality).

So
_
R
N
[J

u(x)[
p
dx
_
R
N
_
R
N
[u(y)[
p
J

(x y)dydx
=
_
R
N
_
R
N
[u(y)[
p
J

(x y)dxdy
=
_
R
N
[u(y)[
p
dy (since
_
R
N
J

= 1)
thus |J

u|
p
|u|
p
.
Case p = .
[J

u(x)[ =

_
R
N
u(y)J

(x y)dy

|u|

_
R
N
J

(x y)dy = |u|

.
In either case, the inequality |J

u|
m,p
|u|
m,p
follows by applying the above to each D

u.
(ii) Suppose 1 p < .
Simple case. u = 1
Q
where Q is a rectangle. Then
J

u(x) =
_
1 for all small > 0 if x Q

0 for all small > 0 if x R

N
Q
0 J

u(x) 1 for all x, and J

(x) = 0 if x , Q+ B(0, 1), < 1.

31
So J

u u a.e. and 0 J

u 1
Q+B(0,1)
. So the Dominated Convergence Theorem shows
_
R
N
[J

u u[
p
=
_
Q+B(0,1)
[J

u u[
p
( < 1)
0 as 0.
General case. Let u L
p
(R
N
) and > 0. We can choose rectangles Q
1
, . . . , Q
k
and constants
c
1
, . . . , c
k
such that |u u
0
|
p
< where
u
0
=
k

n=1
c
k
1
Q
k
.
Now J

u
0
u
0
as 0 by the above case plus the triangle inequality; choose
0
> 0 such
that |J

u
0
u
0
|
p
< for 0 < <
0
. Then
|J

u u|
p
|u u
0
|
p
+|u
0
J

u
0
|
p
+|J

u
0
J

u|
p
+|u
0
J

u
0
|
p
+ (by (i))
< 3 provided <
0
.
Thus |J

u u|
p
0 as 0.
For u W
m,p
(R
N
) we obtain |J

u u|
m,p
0 by applying the above to each D

u.
(iii) Let 1 p < .
Density of D(R
N
) in L
p
(R
N
) actually follows from the proof of (ii) since J

u
0
D(R
N
).
Let > 0. Using a result from the Problem Sheets (Sheet 4 Q2), given u W
m,p
we can
choose v W
m,p
with compact support such that |u v|
m,p
< . Now J

v v in W
m,p
by
(ii), and J

v D(R
N
), so we can choose with |J

v v|
m,p
< , so |u J

v|
m,p
< 2.
Therefore W
m,p
0
(R
N
) = W
m,p
(R
N
).
Remark. In general W
m,p
() ,= W
m,p
0
().
3.1 Localisation
We want analogues of the results of Theorem 3.3 for a general domain . For a function
u L
1
loc
(),
u J

(x) =
_
R
N
u(y)J

(x y)dy
which requires values of u at points of B(x, ) which might be outside . If we set u = 0 outside
then the resulting discontinuity of u will be reected in large derivatives of J

u near .
This necessitates restricting attention to subsets of , typically compact ones.
32
Lemma 3.4. Let R
N
be open and nonempty,
0
open,
0
a compact subset of . Let
0 < < dist(
0
, R
N
), then
(i) For 1 p we have |J

u|
p,
0
|u|
p,
0
+B

(0,)
if u L
p
() and
|J

u|
m,p,
0
|u|
m,p,
0
+B

(0,)
if u W
m,p
().
(ii) For 1 p < we have |J

u u|
p,
0
0 as 0 if u L
p
() and
|J

u u|
m,p,
0
0 as 0 if u W
m,p
().
Proof. (i) For 0 < <

< dist(
0
, R
N
) we can choose D() with 0 1
everywhere, = 1 on
0
+ B

(0, ) and = 0 outside

0
+ B

(0,

). Then by Theorem 3.3

|J

u|
p,
0
|J

(u)|
p,R
N |u|
p,R
N |u|
p,+B

(0,

)
|u|
p,+B

(0,)
as

.
The remaining parts use similar arguments, applied to the partial derivatives where necessary.
Remark. If u W
m,p
0
() then u W
m,p
(R
N
) (take u = 0 outside ). So J

u makes sense
and we have |J

u|
p,
|u|
p,R
N = u|
p,+B

(0,)
etc.
Theorem 3.5 (Fundamental Theorem of Calculus). Suppose R
N
is a nonempty, connected
open set, u W
1,1
loc
(), and D
i
u = 0 a.e. in for i = 1, . . . , N. Then u is essentially constant
on .
Proof. Consider a ball B such that B . Then, for all small > 0,
D
i
(J

u)(x) = J

D
i
u(x) = 0 for all x B,
for i = 1, . . . , N. Hence J

u is constant in B. As 0, J

u u in L
1
(B), so u = const.
a.e. in B.
Take S(c) to be the union of all the open balls B such that B and u = c a.e. on B, for
c R. Then =

cR
S(c), and the S(c) are open and disjoint, so by connectedness = S(c)
for one particular value of c.
We are now in a position to prove Lemma 1.3:
Lemma 3.6. Let R
N
be open.
(i) Let 1 p < . Then D() is dense in L
p
().
(ii) Let u L
1
loc
() with
_

u = 0 for all D(). Then u = 0 a.e. in .

33
Proof. (i) We addressed the case = R
N
in Theorem 3.3. For general , choose bounded open

0
such that
0
, and such that |1

0
u u|
p
< (u L
p
(), > 0, having been given).
Then, for > 0 small enough, J

(1

0
u) is a test function on and |J

(1

0
u) 1

0
u|
p
< .
So |u J

(1

0
u)|
p
< 2.
(ii) Consider bounded open
0
with
0
and take 0 < < dist(
0
, R
N
). Then
J

u(x) =
_
R
N
u(y)J

(x y)dy = 0 x
0
since y J

(x y) is a test function on . Letting 0 we get J

u u in L
1
(
0
), so
u = 0 a.e. in
0
. Hence u = 0 a.e. in .
Remark. This shows that dierent locally integrable functions represent dierent distributions
and in particular, if D

u L
1
loc
, then the function representing D

u is unique.
Lemma 3.7. Let R
N
be open, u C(). Then J

u u uniformly on compact subsets

of . [Exercise]
Lemma 3.8. Let ,= R
N
be open. Then there is a sequence
n
in D() such that:
(i) 0
n
1 for every n, and

n=1

n
= 1 on (partition of unity);
(ii) every point of has a neighbourhood on which all except nitely many
n
vanish identi-
cally (local niteness);
(iii) local niteness has the consequence that any compact subset of intersects the supports
of only nitely many
n
.
Proof. For n N dene

n
=
_
x

[x[ < n and dist(x, R

N
) > 2/n
_
.
Then
n
is open and bounded,
n
,
n

n+1
and

nN

n
= . Set S
n
=
n

n1
for
n 2 with S
1
=
1
, and write

n
= J
1/n
1
Sn
,
so that
n
D() and supp(
n
) = S
n
+ B(0, 1/n).
Consider x , so B

(x, r)
n
for some n N and r > 0. If k > n then B

(x, r)S
k
=
so B

(x, r/2) (S
k
+B(0, r/2)) = . Hence B

(x, r/2) supp(

k
) = if k > maxn, 2/r. It
follows by covering that if K is compact then K meets the supports of only nitely many

n
.
34
Let x ; we claim
n
(x) > 0 for some n N. We have x
m
for some m N and
then
m
S
1
S
m
. We can choose r, 0 < r < 1/m, such that B

(x, r)
m
and then
S
n
B

(x, r) has positive measure for some n 1, . . . , m so

n
(x) =
_
Sn
J
1/n
(x y)dy
_
SnB

(x,r)
J
1/n
(x y)dy > 0.
Set

n
=

n

kN

k
.
Then every point of has a neighbourhood on which the above sum involves only nitely many
functions, hence
n
is smooth and

nN

n
= 1.
If K is compact, then K intersects the supports of only nitely many v
n
. For, each
point of K is the centre of an open ball that intersects only nitely many supp v
n
, and K can
be covered by nitely many such balls.
Theorem 3.9 (Meyers-Serrin H = W Theorem). Let ,= R
N
be open, m N,
1 p < . Then C

() W
m,p
() is dense in W
m,p
().
Proof. Choose a locally nite, countable partition of unity into test functions on ,
n

n=1
,
as provided by Lemma 3.8. Consider > 0, u W
m,p
().
For each n N choose 0 <
n
< 1/n such that
n
< dist(supp
n
, R
N
) so v
n
=
J
n
(
n
u) D(), and such that |v
n

n
u|
m,p
< 2
n
.
Consider x . Then r > 0 can be chosen such that B

(x, r) supp
n
= for all except
nitely many n, so B

(x,
1
2
r)
_
supp
n
+ B(0,
1
2
r)
_
= except for nitely many n, hence
B

(x,
1
2
r) supp v
n
= for all suciently large n. Thus the family v
n

n=1
is locally nite.
Take v =

k=1
v
k
C

() by local niteness.
Choose
n

n=1
to be an increasing family of bounded open sets with
n
, and

n=1

n
= .
By local niteness, each
n
intersects the supports of only nitely many
k
and v
k
and
since u =

k

k
u we have
|v u|
m,p,n
=
_
_
_
_
_

k=1
(v
k

k
u
_
_
_
_
_
m,p,n
;
the above sum involves only nitely many functions so there are no convergence problems.
35
Therefore
_
n

0||m
[D

v D

u[
p
=
_
_
_
_

k=1
(v
k

k
u)
_
_
_
_
p
m,p,n

k=1
|v
k

k
u|
m,p,n
_
p
<
_

k=1
2
k
_
p
=
p
and we can let n and apply the Monotone Convergence Theorem to LHS to get
_

0||m
[D

v D

u[
p

p
i.e.
|v u|
m,p,
.
Now v = u + (v u) W
m,p
() so v C

() W
m,p
().
Remarks. This result says nothing about the behaviour of the approximating smooth functions
near the boundary, so it cannot be used to dene boundary values of Sobolev functions.
Note that p < cannot be avoided.
Theorem 3.10. Let , be nonempty, bounded, open sets in R
N
and suppose F : is
a bijection satisfying F C
1
() and F
1
C
1
(). Then, for 1 p < , the map v v F
is an invertible bounded linear operator from W
1,p
() onto W
1,p
().
Proof. First consider u = v F, v C
1
() W
1,p
(). Then
_

[D
j
u(x)[
p
dx =
_

x
j
v(F(x))

p
dx =
_

k=1
D
k
v(F(x))D
j
F
k
(x)

p
dx
const
_

k=1

D
k
v(F(x))

p
dx
= const
_

k=1

D
k
v(y)

p
[JF
1
(y)[dy (JF
1
Jacobian)
const
_

k=1
[D
k
v(y)[
p
dy
hence
|u|
1,p,
const |v|
1,p,
.
A similar inequality holds in the reverse direction, and by density (Meyers-Serrin) these in-
equalities hold throughout W
1,p
().
36
Lemma 3.11. Let B = B

(0, r) R
N
, B

= (x

, x
N
) B [ x
N
> 0, u W
1,1
(B
+
). Then
(i)
_
B
+
(D
j
u) =
_
B
+
u(D
j
) for all D(B), 1 j N 1;
(ii)
_
B
+
(D
N
u) =
_
B
+
u(D
N
) for all D(B) such that (x

, 0) = 0 if x

B
N1
;
(iii) dening u(x

, x
N
) = u(x

, [x
N
[) we have u W
1,1
(B) with D
j
u(x

, x
N
) = D
j
(x

, [x
N
[) for
1 j N 1 and D
N
u(x

, x
N
) = sgn(x
N
)D
N
u(x

, [x
N
[) a.e.
Figure 2: 1
[1,)
1
[
1
2
,)
Proof. For (i) and (ii) choose increasing C

(R) such that 1

[1,)
1
[
1
2
,)
(c.f. Fig-
ure 2), e.g. = J
1/4
1
[
3
4
,)
. Dene

(s) =
_
s

_
for s R. Thus

(x
N
)(x

, x
N
) denes
an element of D(B
+
).
(i) For 1 j N 1
_
B
+
(D
j
u)(x)

(x
N
)(x

)dx =
_
B
+
u(x)D
j
(

(x
N
)(x))dx
=
_
B
+
u(x)

(x
N
)D
j
(x)dx.
For x
N
> 0 we have 0

(x) 1 and

1 as 0, so we can apply the Dominated

Convergence Theorem to deduce (i).
(ii) For j = N,
_
B
+
(D
N
u(x))

(x
N
)(x)dx =
_
B
+
u(x)
_

_
x
N

_
(x) +

(x
N
)D
N
(x)
_
dx.
Now on B
+
we have [(x)[ c
1
x
N
where c
1
= |D
N
|
sup
, since (x) = 0 when x
N
= 0, hence

_
x
N

_
(x

, x
N
)

c
1
_
x
N

_
x
N

_
,
37
which is bounded above by c
1
c
2
when 0 < x
N
< and vanishes for x
N
, where c
2
= |

|
sup
.
Therefore
1

_
x
N

_
(x

, x
N
) is uniformly bounded and tends to 0 pointwise as 0. We
now deduce (ii) using the Dominated Convergence Theorem.
(iii) If 1 j N 1 and D(B) then
_
B
u(x)D
j
(x)dx =
_
B
+
u(x

, x
N
)D
j
(x

, x
N
)dx +
_
B

u(x

, x
N
)D
j
(x

, x
N
)dx
=
_
B
+
u(x

, x
N
) [D
j
(x

, x
N
) + D
j
(x

, x
N
)] dx (by (i) with )
=
_
B
(D
j
u(x

, [x
N
[)) (x

, x
N
)dx.
For j = N we have
_
B
u(x)D
N
(x)dx =
_
B
+
u(x

, x
N
)D
N
(x

, x
N
)dx +
_
B

u(x

, x
N
)D
N
(x

, x
N
)dx
=
_
B
+
u(x

, x
N
)
_
D
N
(x

, x
N
)

x
N
(x

, x
N
)
_
dx
=
_
B
+
(D
N
u(x)) [(x

, x
N
) (x

, x
N
)] dx
(by (ii) since, if x
N
= 0 then (x

, x
N
) (x

, x
N
) = 0)
=
_
B
sgn(x
N
) (D
N
u(x

, [x
N
[)) (x)dx
Terminology. Let ,= R
N
, N 2. A C
1
chart for is an open set U = rB
N1

(a, a) (with respect to some local Cartesian coordinates in R
N
) and f C
1
(rB
N1
) such that
|f|
sup
< a and such that
U = (x

, x
N
) U [ x
N
< f(x

) .
We say is of class C
1
if there is a C
1
chart for in a neighbourhood of every point.
Figure 3: C
1
chart
38
Theorem 3.12 (Extension Theorem). Let ,= R
N
(N 2) be bounded with C
1
boundary,
and 1 p < . Then there exists a bounded open set V and a bounded linear operator
E : W
1,p
() W
1,p
0
(V )
such that Eu = u almost everywhere in for all u W
1,p
() and Eu C(V ) for all u
C() W
1,p
().
Proof. Consider a chart (U, f) where U = rB
N1
(a, a). Dene
F(x

, x
N
) = (x

, x
N
f(x

)),
(c.f. Figure 4) which is a bijection from U to an open set W such that F C
1
(U) and
F
1
C
1
(W) given by
F
1
(y

, y
N
) = (y

, y
N
+ f(y

)).
Figure 4: F C
1
(U)
Choose a ball B with B W, centre O (which lies on F(() U) and set
B

= (y

, y
N
) B [ y
N
> 0 .
Then Lemma 3.11 provides an extension operator T : W
1,p
(B
+
) to W
1,p
(B). Note that if
u W
1,p
(B
+
) C(B
+
) then by construction Tu W
1,p
(B) C(B).
Now, assuming p < , the operator
L : W
1,p
(D) W
1,p
(B), dened by Lu = u F
1
,
where D = F
1
(B), D

= F
1
(B

), is bounded and has a bounded inverse. Dene

K : W
1,p
(D

) W
1,p
(D) by Ku = L
1
TLu, u W
1,p
(D

).
39
Then K is bounded and is an extension operator of the desired form for D

.
Now cover with nitely many bounded open sets D
1
, . . . , D
n
, each having an extension
operator K
i
: W
1,p
(D
i
) W
1,p
(D
i
). Choose an open set D
0
, D
0
, such that D
0
, . . . , D
n
cover .
Choose
i
D(D
i
) such that
n

i=0

i
1 on a set whose interior contains . Dene
Eu(x) =
n

i=1

i
(x)K
i
(u[
D
i
)(x) +
0
u(x).
Then E(x) =

n
i=0

i
(x)u(x) = u(x) for x , and Eu W
1,p
0
(V ) where V =

n
i=0
D
i
.
Moreover if u C() W
1,p
() then Eu C
c
(V ).
Remark. For smoother boundaries, extension operators for W
m,p
can be dened.
Theorem 3.13 (Trace Theorem). Let ,= R
N
be a bounded domain with C
1
boundary.
Let 1 p < . Then there is a bounded linear operator Tr : W
1,p
() L
p
() such that if
u C() W
1,p
() and u denotes the uniformly continuous extension of u to then
Tr u(x) = u(x) for all x .
Proof. Consider u W
1,p
() C
1
(). Consider a chart for , say (U, f) where U = rB
N1

(a, a). Consider C

(R) such that

(a) = 0 and (s) = 1 for s >
1
2
(a |f|
sup
) .
Then
_
U
[u[
p
=
_
U
(x
N
)[u(x)[
p
dS(x)
=
_
B
N1
(f(x

))[u(x

, f(x

))[
p
(1 +[f(x

)[
2
)
1
2
dx

=
_
B
N1
_
f(x

)
a
D
N
_
(x
N
)[u(x

, x
N
)[
p
_
(1 +[f(x

)[
2
)
1
2
dx
N
dx

if p > 1 (*)
c
_
U
[

(x
N
)[[u(x)[
p
+ p[(x
N
)[[u(x)[
p1
[D
N
u(x)[dx
c
_
U
[u[
p
+[u[
p1
[D
N
u[dx
c
_
u
[u[
p
+[D
N
u[
p
dx
40
where we have used Youngs inequality in the last line to obtain
[u[
p1
[D
N
u[
[u[
(p1)q
q
+
[D
N
u[
p
p
with 1/p + 1/q = 1, so (p 1)q = p, hence
_
U
[u[
p
c|u|
W
1,p
(U)
.
The case p = 1 is similar, using at (*) the inequality
[(x

, f(x

))u(x

, f(x

))[
_
f(x

)
a
[D
N
((x

,
N
)u(x

,
N
))[dx
N
.
Now, covering with nitely many charts (U
k
, f
k
)
n
k=1
then
|u|
L
p
()

n

k=1
|u|
L
p
(U
k
)

n

k=1
c
k
|u|
W
1,p
(U
k
)
c|u|
W
1,p
()
.
To deal with the case of general u W
1,p
(), it must be shown that u can be approximated
in | |
1,p
by such functions. First extend u to Eu W
1,p
0
(V ), then use density to approximate
Eu in | |
1,p
by a sequence u
n
in D(V ). The restrictions of the u
n
to form the desired
approximating sequence. If now u
n
is any sequence in C
1
() W
1,p
() converging in | |
1,p
to u, then their boundary traces form a Cauchy sequence in L
p
() converging to a limit,
denoted Tr(u), which is independent of the choice of approximating sequence (any two such
sequences can be interlaced to give another one, whose boundary traces must also converge).
Finally, let u C() W
1,p
(); we have to check that the above denition agrees with
u[

. Note that Eu[

is a uniformly continuous extension of u to , so u = Eu[

and therefore
u[

= Eu[

. Now, as 0, we have J

Eu Eu on V both uniformly and in | |

1,p
, so
J

Eu[

converges uniformly to Eu[

and converges in L
p
() to some limit which must
therefore be Tr u. Hence Tr(u) = Eu[

= u[

as required.
4 Embeddings on Smooth Bounded Domains
Theorem 4.1 (Sobolev Embedding Theorem for smooth bounded domains). Let N 2,
,= R
N
be open and bounded with C
1
boundary, m N and 1 p < . Then the
following embeddings are bounded:
(i) W
m,p
() L
q
() for p q p

:=
Np
N mp
if mp < N;
(ii) W
m,p
() L
q
() for p q < if m < N, mp = N;
41
(iii) W
N,1
() C();
(iv) W
m,p
() C
0,
() for 0 < m
N
p
if mp > N > (m1)p;
(v) W
m,p
() C
0,
() for 0 < < 1 if (m1)p = N.
Proof. When m = 1 cases (i), (ii), (iv), (v) follow by using Theorem 3.12 to choose an extension
operator E : W
1,p
() W
1,p
0
(V ) for some bounded open V , and applying the embedding
theorem for W
1,p
0
() (Theorem 2.12). We leave case (iii) to the end, and proceed to describe
the inductive step in the other cases.
(i) Suppose the result holds for some m 1 and all p with mp < N. Let p satisfy (m+1)p < N.
Let p
1
=
Np
N mp
and p
2
=
Np
N (m+ 1)p
=
Np
1
N p
1
. For u W
m+1,p
() we now have
u W
m,p
() and thence
|u|
p
1
c|u|
m,p
c|u|
m+1,p
|u|
p
1
c|u|
m,p
|u|
m+1,p
,
so |u|
1,p
1
c|u|
m+1,p
,
so |u|
p
2
c|u|
m+1,p
from the initial case W
1,p
1
L
p
2
. The case q < p

follows by interpolation, completing the

inductive step.
(ii) Suppose the result holds for some m 1 with m < N. Suppose m + 1 < N and let
p = N/(m + 1); then mp < N and Np/(N mp) = N. For u W
m+1,p
() we have
u W
m,p
() hence using (i)
|u|
N
c|u|
m,p
c|u|
m+1,p
,
|u|
N
|u|
m,p
c|u|
m+1,p
,
|u|
q
c|u|
1,N
c|u|
m+1,p
,
where the rst inequality of the last line comes from the initial case of (ii). This completes the
inductive step of (ii).
(iv) Suppose m 2 and mp > N > (m1)p. Consider u W
m,p
(). Then from (i) we have,
writing p
0
=
Np
N (m1)p
> N,
|u|
p
0
c|u|
m1,p
c|u|
m,p
,
|u|
p
0
c|u|
m1,p
c|u|
m,p
,
so |u|
1,p
0
c|u|
m,p
.
42
Now apply the initial case of (iv) together with the above inequality to obtain, writing
0
=
1
N
p
0
= m
N
p
,
|u|
C
0,
0
c|u|
1,p
0
c|u|
m,p
.
When 0 < <
0
we can apply the embedding C
0,
0
() C
0,
() (Lemma 2.11) to obtain
|u|
C
0, c|u|
m,p
establishing the higher-order cases of (iv).
(v) Suppose N = (m1)p with m 2, let q > p and let u W
m,p
(). Then (ii) yields
|u|
q
c|u|
m1,p
c|u|
m,p
,
|u|
q
c|u|
m1,p
c|u|
m,p
,
so |u|
1,q
c|u|
m,p
.
When q > N (so q > p) and (q) = 1
N
q
, the preliminary case of (v) yields
|u|
C
0,(q) c|u|
1,q
,
and for 0 < < 1 we can apply this inequality with q >
N
1
together with the embedding
C
0,(q)
() C
0,
() to deduce
|u|
C
0, c|u|
C
0,(q) c|u|
m,p
,
establishing the higher-order cases of (v).
(iii) Recall the estimate
|u|
sup
c|u|
N,1
u W
N,1
(Q)
where the constant c depends on the dimensions of the rectangle Q but not on its position or
orientation; this holds for u C
N
(Q) W
N,1
(Q) by Theorems 2.10 and 2.3, and follows for
general u W
N,1
(Q) by Meyers-Serrin.
Consider a chart (U, f) for , where U = rB
N1
(a, a) and f C
1
(rB
N1
). Let
n = |n| n = (f(0), 1) which is the inward normal to at (0, f(0)). Let
g() = f(0) [[ for R
N1
,
where > 0 is to be chosen later. Let v
i
= (v

i
, v
i,N
) , i = 1, . . . , N be the vertices adjacent
to 0 of a (unit) cube Q with diagonal [0,

N n]. Then the vertices (0, f(0)) + v

i
lie below the
43
tangent hyperplane to at (0, f(0)), so f(0)v

i
> v
i,N
. We choose r and small enough
that g(v

i
) > v
i,N
and f(x

) > g() for all x

rB
N1
and 0 ,= R
N1
.
If x

,= x

+ both belong to rB
N1
then for 0 < t < 1 the forward directional derivative
satises
d
dt+
(f(x

+ t) g(t)) = f(x

+ t) g() > 0
and it follows that
f(x

+ ) > f(x

) + g().
It now follows that
((0, f(x

)) + Q) U for all x

rB
N1
.
Hence x + Q for all x within distance > 0 of (0, f(0)) provided that is chosen
suciently small. Then every point of suciently close to (0, f(0)) lies in a cube of edge
contained in .
A compactness argument now shows that, for some > 0, every point x lies in a
(closed) cube Q
x
of side contained in , and so
[u(x)[ |u|
sup,Qx
c|u|
N,1,Q

x
c|u|
N,1,
.
Remarks.
1) Boundedness of can be avoided, at the expense of a more complicated proof and carefully
chosen regularity assumptions on .
2) The smoothness of can be weakened somewhat. See Adamss book.
Theorem 4.2. Let ,= R
N
be open, 1 p < , and K L
p
(). Then K is relatively
compact in L
p
() if and only if K is bounded in L
p
() and > 0 > 0 and G compact
such that
(i)
_
\G
[u[
p
<
p
for all u K, and
(ii)
_

[u(x + h) u(x)[
p
dx <
p
(taking u = 0 outside ) for all u K and all h R
N
,
satisfying [h[ < .
Proof. ( only will be proved.) It is enough to suppose = R
N
, extending u = 0 outside .
Claim 1. If > 0 and G R
N
then K(G, ) := 1
G
J

u [ u K is relatively compact in
C(G), and therefore in L
p
(G).
44
For, writing B = B(0, 1) and taking q to be conjugate to p,
|J

u|
C(G)
|J

|
sup
|u|
L
1
(G+B)
|J

|
sup
|u|
p
|1
G+B
|
q
,
so K(G, ) is uniformly bounded on G. Further, if x G and [h[ < 1 then
[J

u(x + h) J

u(x)[
_
R
N

(x + h y) J

(x y)

[u(y)[dy
[h[|J

|
sup
|u|
L
1
(G+(1+)B)
,
whence K() is equicontinuous. Relative compactness in C(G) follows by Arzel`a-Ascoli, and
relative compactness in L
p
(G) follows from this.
Claim 2. |J

u u|
p
0 as 0 uniformly over u K. For
|J

u u|
p
=
_
R
N

_
R
N
J

(h)(u(x h) u(x))dh

p
dx

_
R
N
_
R
N
J

(h)

(u(x h) u(x))

p
dhdx (Jensens inequality)
=
_
R
N
_
R
N
J

(h)

(u(x h) u(x))

p
dxdh
sup
|h|<
_
R
N

(u(x h) u(x))

p
dx 0 as 0, uniformly over u K by (ii).
Claim 3. > 0 G R
N
compact and > 0 such that u K |u 1
G
J

u|
p
< .
For, let > 0 and let G

be the compact set provided by (i). Then, for G = G

+ B, u K,
|u 1
G
J

u|
p
|u 1
G
|
p
+|1
G
(u J

u)|
p
+ |(1
G
1
G
)J

u|
p
< +|u J

u|
p
+ |(1
G
1
G
)|
p
(if 0 < < 1)
< 3
for > 0 small enough, independent of u K.
Claim 4. K is totally bounded.
For, let > 0 and choose > 0 and compact G R
N
such that
|u 1
G
J

u|
p
< for all u K.
By Claim 1,
K(G, ) := 1
G
J

u [ u K
is relatively compact in L
p
; let S
1
, . . . , S
n
be sets of diameter less than covering K(G, ).
Then S
k
+ B(0, )
n
k=1
is a nite collection of sets of diameter less than 2 covering K.
45
Lemma 4.3. Let ,= R
N
be bounded and open and let 0 < < 1. Then the
embedding C
0,
() C
0,
() is compact.
Proof. Exercise on Sheet 5.
Theorem 4.4 (Rellich-Kondrachov Compact Embedding Theorem). Let N 2, ,= R
N
be open and bounded, m N and 1 p < . Then the following embeddings are compact:
(i) W
m,p
0
() L
p
() for p q < p

:=
Np
N mp
, if mp < N;
(ii) W
m,p
0
() L
q
() for p q < , if m < N, mp = N;
(iii) W
m,p
0
() C
0,
() for 0 < < m
N
p
, if mp > N > (m1)p;
(iv) W
m,p
0
() C
0,
() for 0 < < 1, if (m1)p = N.
If is of class C
1
then W
m,p
0
() can be replaced by W
m,p
().
Proof. We rstly assume 1 p N and show W
1,p
0
() L
1
() is compact. Let S denote the
unit ball in W
1,p
0
(). Fix q
0
, p < q
0
<
Np
N p
if p < N, or p < q
0
< if p = N. Let > 0,
G a measurable set; then
_
\G
[u[
__
\G
[u[
q
0
_ 1
q
0

1
1
q
0
.
We can now choose compact G such that
_
\G
[u[ u S,
since S is bounded in L
q
0
().
Consider u D(), h R
N
. Then
_

[u(x + h) u(x)[dx
_

_
1
0

d
dt
u(x + th)

dtdx

_
1
0
[u(x + th)[[h[dtdx
=
_
1
0
_

[u(x + th)[[h[dxdt
[h[
_

[u[ c[h[|u|
1,p
.
By density this inequality holds for all u W
1,p
0
(). We can now choose > 0 such that
_

[u(x + h) u(x)[dx < u S, [h[ < .

46
If is C
1
, using the Extension Theorem we can prove the above for u W
1,p
(). Using
Theorem 4.2 it follows that S is relatively compact in L
1
().
Case (i) Choose q
1
, q < q
1
< p

. Choose , 0 < < 1, such that

1
q
=

1
+
1
q
1
. Then
|u|
q
|u|

1
|u|
1
q
1
(7)
for u L
q
1
() ( L
1
()), and so for u W
m,p
0
(). Consider a bounded sequence u
n
in
W
m,p
0
(). Then u
n
is bounded in W
1,p
0
(), and by above has a subsequence, also denoted
u
n
, converging in L
1
(). From (7), together with boundedness of u
n
in L
q
1
(), we deduce
that u
n
converges in L
q
(). Hence compactness of the embedding W
m,p
0
() L
q
().
Case (ii) Essentially the same; choose q
1
such that q < q
1
< .
Case (iii) Consider , 0 < < m
N
p
, and choose , < < m
N
p
. Then W
m,p
0
()
C
0,
() is bounded and C
0,
() C
0,
() is compact by Lemma 4.3, hence W
m,p
0
()
C
0,
() is compact. Case (iv) Similar: given 0 < < 1, choose < < 1. When is of
class C
1
, identical arguments apply except at the stage indicated (compactness into L
1
).
Remarks
1) The assumption that is bounded is unavoidable. For example, consider u
0
W
m,p
0
(R
N
),
u
0
,= 0. Set
u
n
(x) = u
0
(x + nh)
where h is a xed unit vector. Then u
n
is a bounded sequence in W
m,p
0
(R
N
), and
_

[u
n
[
q
0 a.e. as n
for every bounded domain R
N
, so no subsequence of u
n
can converge to a nonzero limit
in | |
q
. But |u
n
|
q
= |u|
q
so no subsequence of u
n
tends to 0 in L
q
. So W
m,p
0
(R
N
) is not
compactly embedded in L
q
(R
N
) for all q.
2) W
N,1
() C() is not compact. For the case N = 1 see Problem Sheet 9 Q1.
3) W
m,p
0
() L
p

(), p

=
Np
N mp
, mp < N is not compact (Problem Sheet 4 Q3). For
suppose B(0, 1) , choose 0 ,= D(B(0, 1)) and let

(x) =

Np
p

_
x

_
, 0 < < 1.
Then

is bounded in W
1,p
0
(), and

(x) 0 as 0 for all x ,= 0, but |

|
p
0 as
0 through any subsequence.
4) In dimension 2, if p > 2 then > 0, > 0 can be chosen such that u(x, y) = x

and
=
_
(x, y) [ 0 < x < 1 and 0 < y < x

_
47
satisfy u W
1,2
() L
p
(), showing p

= 2 is best possible for this case of the embedding

theorem when the boundary is not assumed smooth (Problem Sheet 6 Q2).
48