Vous êtes sur la page 1sur 13

Chapter 2 Sobolev spaces

In this chapter, we give a brief overview on basic results of the theory


of Sobolev spaces and their associated trace and dual spaces.
2.1 Preliminaries
Let be a bounded domain in Euclidean space lRd . We denote by
its closure and refer to = := \ as its boundary. Moreover, we
denote by e := lRd \ the associated exterior domain.
We consider functions u : lR and denote by
|| u
D u.=
x1 1 x2 2 ...xd d
its partial derivatives of order ||. Here, = (1 , ..., d )T lNd0 is a
P
multiindex of modulus || = di=1 i .
We define by C m (), m lN0 , the linear space of continuous functions
on whose partial derivatives D u, || m, exist and are continuous.
C m () is a Banach space with respect to the norm
kukC m () := max sup |D u(x)| .
0||m x
m,
We refer to C (), m lN0 , 0 < < 1, as the linear space of func-
tions in C m () whose m-th order partial derivatives are Holder con-
tinuous, i.e., for all lNd0 with || = m there exist constants > 0
such that for all x, y
|D u(x) D u(y)| |x y| .
We note that C m, () is a Banach space with respect to the norm
|D u(x) D u(y)|
kukC m, () := kukC m () + max sup .
||=m x,y |x y|
Moreover, C0m () and C0m, () are the subspaces of functions with
compact support in ).
Finally, C () stands for the set of functions with continuous partial
derivatives of any order and C0 () denotes the set of C () functions
with compact support in .
In the sequel, we will mainly deal with Lipschitz domains which are
defined as follows.

12
Finite Element Methods; Chapter 2 13

Definition 2.1 Lipschitz domain


A bounded domain lRd with boundary is said to be a Lipschitz
domain, if there exist constants > 0, > 0, and a finite number
of local coordinate systems (xr1 , xr2 , ..., xrd ), 1 r R, and local
Lipschitz continuous mappings
xr = (xr2 , ..., xrd ) lRd1 | |xri | , 2 i d} lR
ar : {
such that
R
[
= xr ) , |
{(xr1 , xr ) | xr1 = ar ( xr | < } ,
r=1
{(x1 , xr ) |
r
xr ) < xr1 < ar (
ar ( xr ) + , |xr | < } , 1 r R ,
{(xr1 , xr ) | xr ) < xr1 < ar (
ar ( xr ) , |
xr | < } e , 1 r R , .
In particular, the geometrical interpretation of the conditions is that
both and e are locally situated on exactly one side of the boundary
.
Definition 2.2 C m -domain and C m, -domain
A Lipschitz domain lRd is a C m -domain (C m, -domain), if the
functions ar , 1 r R, in Definition 2.1 are C m -functions (C m, -
functions).

2.2 Sobolev spaces


We refer to Lp (), p [1, ), as the linear space of p th order inte-
grable functions on and to L () as the linear space of essentially
bounded functions which are Banach spaces with respect to the norms
Z 1/p
p
kvkp, := |v(x)| dx ,

kvk, := ess supx |v(x)| .
Note that for p = 2, the space L2 () is a Hilbert space with respect to
the inner product Z
(v, w)0, := vw dx .

Sobolev spaces are based on the concept of weak (distributional) deriv-
atives:
14 Ronald H.W. Hoppe

Definition 2.3 Weak derivatives


Let u L1 () and lNd0 . The function u is said to have a weak
derivative Dw u, if there exists a function v L1 () such that
Z Z
||
uD dx = (1) v dx , C0 () .

We then set Dw u := v.
The notion weak derivative suggests that it is a generalization of the
classical concept of differentiability and that there are functions which
are weakly differentiable, but not differentiable in the classical sense.
We give an example.
Example 2.1 Example of a weakly differentiable function
Let d = 1 and := (1, +1). The function u(x) := |x|, x , is
not differentiable in the classical sense. However, it admits a weak
derivative Dw1 u given by

1 1 , x < 0
Dw u = .
+1 , x > 0

Indeed, for C0 () we obtain by partial integration

Z+1 Z0 Z+1
u(x)D1 (x) dx = u(x)D1 (x) dx + u(x)D1 (x) dx =
1 1 0
Z0 Z+1
= Dw1 u(x)(x) dx + (u)|01 Dw1 u(x)(x) dx + (u)|10 =
1 0
Z+1
= Dw1 u(x)(x) dx [u(0)](0) ,
1

where [u(0)] := u(0+) u(0) is the jump of u in x = 0. But u is


continuous and hence, [u(0)] = 0 which allows to conclude.

Definition 2.4 The Sobolev spaces W m,p (), p [1, ]


The linear space W m,p () given by

(2.1) W m,p () := { u Lp () | Dw u Lp () , || m }
Finite Element Methods; Chapter 2 15

is called a Sobolev space. It is a Banach space with respect to the


norm
X 1/p
(2.2) kvkm,p, := kDw vkpp, , p [1, ) ,
||m

(2.3) kvkm,, := max kDw vk, .


||m

Note that W m,2 () is a Hilbert space with respect to the inner product
X Z
(2.4) (u, v)m,2, := Dw uDw v dx .
||m

The associated norm will be denoted by k km,2, . We will simply write


(, )m, and k km, , if the context clearly indicates that the W m,2 ()-
inner product and the W m,2 ()-norm are meant.

A natural question is whether C m () is dense in W m,p (). To this end,


we consider the linear space
C m, () := { u C m () | kukm,p, < } .
It is a normed linear space with respect to | |m,p, , but not complete,
i.e., a pre-Banach space. We denote its completion with respect to the
k km,p, -norm by H m,p (), i.e.,
kkm,p,
(2.5) H m,p () := C m, () .
We note that for p = we have H m, () = C m () and hence,
W m, () 6= H m, (). However, a famous result by Meyers/Serrin
(cf., e.g., [1]) states that for a Lipschitz domain and p [1, ) the
spaces W m,p () and H m,p () are equivalent.
Theorem 2.1 Characterization of Sobolev spaces
Let lRd be a Lipschitz domain and m lN0 . Then, for p [1, )
there holds
(2.6) W m,p () = H m,p () .
Remark 2.1 Remark on Meyers/Serrin
The result (2.6) actually holds true for any open subset lRd . How-
ever, C m () is dense in W m,p (), if satisfies the so-called segment
property (cf., e.g., [1]). In particular, Lipschitz domains have this
property.

Theorem 2.1 asserts that for Lipschitz domains and p [1, ) the space
C m () W m,p () is dense in W m,p (). A natural question to ask is
whether or not functions in W m,p (), m 1, belong to the Banach
16 Ronald H.W. Hoppe

space Lq (), or even are continuous. As we shall see, the latter only
holds true, if m is sufficiently large (Sobolev imbedding theorem).
However, let us first give an example which shows that in general we
can not expect such a result:
Example 2.2 Example of a weakly differentiable, but not es-
sentially bounded function
Let d 2 , := {x lRd | |x| < 12 } and consider the function
u(x) := ln(|ln(|x|)) .
The function u has square-integrable first-order weak derivatives
x
D u(x) = , || = 1 ,
|x|2 ln(|x|)
since in view of
1
|D u(x)|d (|x|) := , || = 1 ,
|x|d |ln(|x|)|d
it possesses a square-integrable majorant.
On the other hand, u obviously is not essentially bounded.

Theorem 2.2 Sobolev imbedding theorems


Let lRd be a Lipschitz domain, m lN0 and p [1, ]. Then, the
following mappings represent continuous imbeddings
1 1 m d
(2.7) W m,p () Lp () , = , if m < ,
p p d p
d
(2.8) W m,p () Lq () , q [1, ) , if m = ,
p
d d d
(2.9) W m,p () C 0,m p () , if <m< +1 ,
p p
d
(2.10) W m,p () C 0, () , 0 < < 1 , if m = +1 ,
p
d
(2.11) W m,p () C 0,1 () , if m > +1 .
p

Proof. For a proof, we refer to [1].

In the following chapters, we will frequently take advantage of com-


pact imbeddings of Sobolev spaces.
Finite Element Methods; Chapter 2 17

Theorem 2.3 Kondrasov imbedding theorems


Let lRd be a Lipschitz domain, m lN0 and p [1, ]. Then, the
following mappings are compact imbeddings
1 1 m d
(2.12) W m,p () Lq () , 1 q p , = , if m < ,
p p d p
d
(2.13) W m,p () Lq () , q [1, ) , if m = ,
p
d
(2.14) W m,p () C 0 () , if m > .
p

Proof. The interested reader is referred to [1].

We finally deal with restrictions and extensions of Sobolev func-


tions:
If u W m,p (lRd ), p [1, ],, then for any domain lRd the restric-
tion Ru of u to
Ru(x) := u(x) f.a.a. x
belongs to W m,p () and R : W m,p (lRd ) W m,p () is a bounded linear
operator.
Conversely, it is in general not possible to continuously extend a func-
tion u W m,p () to a function in W m,p (lRd ) as the following example
shows:
Example 2.3 Counterexample (Extension of Sobolev functions)
Let d = 2 , ; = {x = (x1 , x2 ) lR2 | 0 < x1 < 1, |x2 | < xr1 }, r > 1,
and consider the function

u(x) := x1 p , 0<<r.
We note that has a cusp in the origin and hence, is no Lipschitz
domain.
For < r + 1 p, we have that u W 1,p (), since
XZ Z1
p
|D u| dx = C,p x1p+r dx1 .
||=1 0

On the other hand, u is not in L (). Choosing such that p > 2
is possible, we see that a Lipschitz domain is necessary for the
Sobolev imbedding theorem to hold true. Since the Sobolev imbedding
18 Ronald H.W. Hoppe

theorem is valid on lRd , we can not extend u to a function in W 1,2 (lR2 ).

However, in case of a Lipschitz domain we have the validity of the


following Sobolev extension theorem.
Theorem 2.3 Sobolev extension theorem
Let lRd be a Lipschitz domain, m lN0 , and p [1, ]. Then,
there exists a bounded linear extension operator E : W m,p ()
W m,p (lRd ), i.e., Eu = u for all u W m,p () and there exists a constant
C 0 such that for all u W m,p ()
kEukm,p, C kukm,p,lRd .
Proof. We refer to [3].

We recall that the algebraic and topological dual V of a Hilbert


space V is the linear space of all bounded linear functionals on V which
is itself a Hilbert space with respect to the norm
|`(v)|
k`kV := sup .
v6=0 kvkV

We may thus define the dual spaces of the Sobolev spaces W m,p (), p
[1, ].
Definition 2.5 Sobolev spaces with negative index
Let lRd be a bounded domain, let m be a negative integer and
suppose p [1, ]. Then, the Sobolev space W m,p () is defined as the
dual space (W m,q ()) , where q is conjugate to p, i.e., 1q + p1 = 1.
Remark 2.2 The Dirac -function
The Sobolev spaces W m,p (), m < 0, are proper subspaces of Lp ().
For instance, for m < d + dp , if p < , and m d, if p = , they
contain the Dirac -function considered as a linear functional
: W m,p () lR
u 7 x (u) ,
where x is some given point in .
Finite Element Methods; Chapter 2 19

2.3 Sobolev spaces with broken index


For = lRd , we define Sobolev spaces with broken index H s (lRd ) ,
s lR+ by using the Fourier transform v of a function v C0 (lRd )
Z
1 d/2
v() := ( ) exp(i x)v(x) dx .
2
lRd

The associated Sobolev norm is defined according to


(2.15) kvks,lRd := k(1 + | |2 )s/2 v()k0,lRd ,
and we set
kks,lRd
(2.16) H s (lRd ) := C0 (lRd ) .
If lRd is a Lipschitz domain, the space H s () can be either defined
implicity by means of
(2.17) kvks, := inf kzks,lRd ,
z=EvH s (lRd )

where Ev is the extension of v to H s (lRd ), or - for s = m + , m


lN0 , 0 < 1, - explicitly by
1 1/2
(2.18) kvks, := ( )2s kvk20, + |v|2s, ,
diam()
where | |, stands for the seminorm
(2.19) |v|2s, :=
X X Z Z |D v(x) D v(y)|2
2
kD vk0, + dx dy .
|x y|d+2
1||m ||=m

If = , we define the spaceH s () as follows


(2.20) H s () := {v L2 () | kvks, < } ,
equipped with the norm
1 1/2
(2.21) kvks, := ( )2s kvk20, + |v|2s, ,
diam()
where
(2.22) |v|2s, :=
X X Z Z |D v(x) D v(y)|2
2
kD vk0, + dx dy .
|x y|d1+2
1||m ||=m
20 Ronald H.W. Hoppe

In case and v H s (), s < 1,, we define v as the extension by


zero of v, i.e.,

v(x) , x
(2.23) v(x) := .
0 , x\
s
We define H00 () according to
s
(2.24) H00 () := {v L2 () | v H s ()} ,
equipped with the norm
1/2
(2.25) kvkH00
s () := kvk20, + |v|H00
s () ,
where
Z
v 2 (x)
(2.26) |v|2H00
s () := |v|2s, + d
dist(x, )

and
Z Z
|v(x) v(y)|2
(2.27) |v|2s, := dx dy .
|x y|d1+2s

For s < 0, we define H s (G), G {, }, as the dual of H s (G),


equipped with the negative norm
< v, w >G
(2.28) |v|s,G := sup ,
wH s (G),w6=0 kwks,G

where < , >G stands for the dual pairing.


Moreover, if 1 < s < 0, we define H s (G), G {, }, , as the
s
dual of H00 (G). For s = 1, we further define H 1 (G) as the dual of
H01 (G), whereas for = and 1 s < 0 we define H s () as the
dual of H s ().
For details we refer to [2].

2.4 Trace spaces


For a bounded domain lRd with boundary and a function
u C(), it makes sense to define the restriction of u to the bound-
ary simply by considering the pointwise restriction. In view of the
Sobolev imbedding theorem, however, the pointwise restriction does
not make sense for functions u W m,2 () unless m is sufficiently large.
Therefore, we have to find appropriate means to specify the trace u|
of a function u W m,2 (). The following example suggests how to
proceed.
Finite Element Methods; Chapter 2 21

Example 2.4 Trace of smooth functions


Let lR2 be the unit disk which in polar coordinates reads as follows

:= { (r, ) | 0 r < 1 , 0 < 2 } .

For u C 1 (), the restriction to can be expressed according to


Z1 Z1
2
u(1, )2 = r u(r, )2 dr = 2 r2 uur + ru2 (r, ) dr =
r
0 0
Z1
(x1 , x2 )
= 2 r2 uu + ru2 (r, ) dr
r
0
Z1
2 r2 |u||u| + ru2 (r, ) dr
0
Z1
2 |u||u| + u2 (r, ) r dr .
0

Integrating the previous inequality over and applying the Cauchy-


Schwarz inequality as well as Youngs inequality results in
Z Z
2 2
kuk0, = u d 2 |u||u| + u2 dx1 dx2

Z 1/2 Z 1/2 Z
2 2
2 u dx |u| dx + 2 u2 dx


4 1/2 1/2
8 kuk0, kuk1, .

The example can be easily generalized to give the following result.


Lemma 2.1 Trace of Sobolev space functions. Part I
Let lR2 be the unit disk and u W 1,2 (). Then, the trace u|
can be interpreted as a function in L2 () satisfying
1/2 1/2
kuk20, 8 kuk0, kuk1, .
4
(2.29)

Proof. Since C 1 () is dense in W 1,2 (), there exists a sequence


(uk )lN C 1 () such that ku uk k1, 1/k, k lN. Applying the
22 Ronald H.W. Hoppe

inequality derived in Example 2.4, we obtain


4 1/2 1/2
kui uk k0, 8 kui uk k0, kui uk k1,
1 1
4 4
8 kui uk k1, 8 + .
i k
Consequently, (uk )lN is a Cauchy sequence in L2 () and hence, there
exists v L2 () such that uk v (k ) in L2 ().
We define the trace of u on according to
u := v .
We have to show that u is well defined, i.e., it does not depend on
the particular choice of a sequence of C 1 ()-functions. To this end, let
(vk )lN , vk C 1 (), k lN be another sequence satisfying ku uk k1,
0 (k ). We find
kv vk k0, kv uk k0, + kuk vk k0,
4
kv uk k0, + 8 kuk vk k1,

4
kv uk k0, + 8 kuk uk1, + ku vk k1,
0 (k ) .
Finally, the validity of (2.29) follows by a density argument:
kuk0, = kvk0, = lim kuk k0,
k
4 1/2 1/2
4 1/2 1/2
lim 8 kuk k0, kuk k1, = 8 kuk0, kuk1, .
k

The previous result can be generalized to arbitrary Lipschitz domains.

Theorem 2.4 Trace of Sobolev space functions. Part II


Let lRd be a Lipschitz domain and p [1, ]. Then, the trace
mapping
W 1,p () Lp ()
u 7 u|
is a bounded linear, injective mapping.
Proof. We refer to [2].
1,p
However, the trace mapping considered as a mapping from W () in
Lp () is not surjective, i.e., there exist functions v Lp () such
that we can not find u W 1,p () with u| = v.
In particular, for p = 2 we have the following result.
Finite Element Methods; Chapter 2 23

Theorem 2.5 Trace of Sobolev space functions. Part III


Let lRd be a Lipschitz domain and 1/2 < s 1. Then there holds:

(i) There exists a unique linear bounded trace mapping


(2.30) 0 : H s () H s1/2 ()
v 7 0 (v) := v|
such that
(2.31) k0 (v)ks1/2, C kvks, ,
|0 (v)|s1/2, C |v|s, .
(ii) The trace operator 0 has a bounded right inverse
(2.32) 0 : H s1/2 () H s () ,
i.e., 0 0 w = w, w H s1/2 (), and
(2.33) k0 (w)ks, C kwks1/2, ,
|0 (w)|s, C |w|s1/2, .
(iii) In case , the results in (i) and (ii) hold true with H s1/2 ()
s1/2
replaced by H00 () and H s () replaced by Hs D (), where D :=
\ and
(2.34) Hs D () := {v H s () | 0 v = 0 on D } .
Proof. We refer to [2].

Definition 2.6 Dirichlet data


For u H 1 (), the trace 0 u is called the Dirichlet data.

Definition 2.7 Normal component trace mapping


Given a Lipschitz domain lRd with boundary , we denote by
n = (n1 , ..., nd )T
the unit exterior normal vector which is well defined at almost all
x . For a vector field q C ()d , we refer to
(2.35) n (q) := n q| .
as the normal component trace mapping. It can be extended by
continuity to a linear continuous mapping
(2.36) n : H 1 ()d H 1/2 () .
24 Ronald H.W. Hoppe

Definition 2.8 Normal derivative and Neumann data


Within the notations of Definition 2.7, for u C () the normal
component trace of u
(2.37) n u := n (u) = n u
is called the normal derivative of u.
The mapping n can be extended to a linear continuous mapping
(2.38) n : H 2 () H 1/2 () .
In particular, for u H 2 () the normal derivative n u is called the
Neumann data.

References
[1] Adams, R.A. (1975), Sobolev spaces. Academic Press, New York.
[2] Grisvard, P. (1985), Elliptic Problems in Nonsmooth Domains. Pitman,
Boston.
[3] Stein, E.M. (1970), Singular Integrals and Differentiability Properties of Fun-
crions. Princeton University Press, Princeton.