AN INTRODUCTION TO SOBOLEV SPACES

Steve Taylor, Montana State University.

Preface:
These notes were written to supplement the graduate level PDE course at
Montana State University. Sobolev Spaces have become an indispensable tool in the theory
of partial differential equations and all graduate-level courses on PDE's ought to devote
some time to the study of the more important properties of these spaces. The object of these
notes is to give a self-contained and brief treatment of the important properties of Sobolev
spaces. The main aim is to give clear proofs of all of the main results without writing an
entire book on the subject! Why did I write these notes? Much of the existing literature on
the subject seems to fall into two categories, either long treatises on the subject with the
most general assumptions possible (and thus unsuitable for part of a PDE course), or very
sketchy discussions confined to a chapter of a PDE text.
CONTENTS:
1.

The Spaces W j , p () and W0j , p ().......................................... 1

2.
3.
4.
5.
6.

Extension Theorems...........................................................
Sobolev Inequalities and Imbedding Theorems......................
Compactness Theorems......................................................
Interpolation Results.........................................................
The Spaces H k () and H0k ().............................................

7.
Trace Theorems................................................................
Appendix: Some Spaces of Continuous Functions.......................
References................................................................................

10
13
20
25
27
29
34
35

In these notes, is a domain (i.e. an open, connected set) in R n .

1.

The Spaces W j , p () and W0j , p ()

Definitions: Suppose 1 p < . Then

(i)
Lploc () = {u: u Lp (K) for every compact subset K of }
(ii)
u is locally integrable in if u L1loc () .
(iii)

Let u and v be locally integrable functions defined in . We say that v is

the th weak derivative of u if for every C0 ()


(iv)

uD dx = (1)|| v dx ,

and we say that D u = v in the weak sense.

Let u and v be in Lploc () . We say that v is the th strong derivative of u if
for each compact subset K of there exists a sequence { j} in C|| (K)
such that j u in Lp (K) and D j v in Lp (K).

THEOREM 1
sense.

If D u = v and D v = w in the weak sense then D+ u = w in the weak

P ROOF

Let C0 () and = D . Then

uD + dx = (1)|| v dx = (1)| | vD dx = (1)| |+| | w dx .

Let C0 (Rn ) be such that

Definition (mollifiers):
(i)

Supp B1 (0) , (recall that "supp" denotes the support of a function, and
Br (p) denotes an open ball of radius r and center p).

(ii)

(x) dx = 1 ,

(iii)

(x) 0 .

If > 0 then we set (provided that the integral exists)

J u(x ) =

1
xy
)u(y) dy .
n (

J u is called a mollifier of u. Note that if u is locally integrable in and if K is a

compact subset of then J u C (K) provided that < dist(K,). Suppose now that
u Lploc () . Clearly
J u(x ) = B (0) (y)u(x y) dy ,
1

so for p > 1 we have (if 1 / p + 1 / q = 1)

| J u(x)| B

1 (0)

{(y)}1/ q {(y)}1/ p |u(x y)| dy

 ( B (0) ({(y)}1/ q ) q dx)1/ q (B (0) ({ (y)}1/ p | u(x y)|) p dy)1/ p .

1

Hence | J u(x)| p B

(0)

(y)|u(x y)| p dy , and this trivially holds if p = 1 too. Integrating

this, we see that

| J u(x)| p dx B (0) (y) K |u(x y)| p dx dy

1

B (0) (y) K |u(x)| p dx dy

1

= K |u(x)| dx ,
p

where K0 is a compact subset of , K Interior(K0 ) and < dist(K,K0 ) . i.e. we have

|| J u||L p (K) ||u|| Lp (K0 ) .
LEMMA 2

(1)

If u Lploc () and K is a compact subset of then || J u u|| Lp (K ) 0 as

0.
P ROOF
Let K0 be a compact subset of where K Interior(K0 ) and let
< dist(K,K0 ) . Let > 0 and let w C (K0 ) be such that || u w|| Lp (K0 ) < . Then
applying (1) to u w , we obtain
|| J u J w|| Lp (K ) < .

(2)

But J w(x) w(x) = B (0 ) (y){w(x y) w(x)} dy , and this goes to zero uniformly on K
1

as 0. Hence, if is sufficiently small, we have

|| J w w|| Lp (K ) < .

(3)

Hence, by (2) and (3)

|| J u u|| Lp (K ) ||w u|| Lp ( K) +|| J u J w|| Lp (K ) +|| J w w|| Lp ( K) < 3 .
Since is arbitrary, || J u u|| Lp (K ) 0 as 0.

(4)

The proof of the following theorem contains some other important approximating
properties of mollifiers.
THEOREM 3

Suppose that u and v are in Lploc () . Then D u = v in the weak sense if and

only if D u = v in the strong Lp sense.

P ROOF
Suppose that D u = v in the strong Lp sense. Let C0 () and let
K = supp . Let > 0 and take C|| (K) so that || u|| Lp ( K) < and
|| D v|| Lp (K ) < . Then
| K uD dx (1)|| K v dx|| K D dx (1)|| K D dx|

+ | K (u )D dx|+| K (v D ) dx|

||u || Lp ( K) || D || Lq (K ) +||v D || Lp ( K) || || Lq (K )
(||D ||Lq (K ) +|| || Lq (K ) ),

where q is the conjugate exponent of p (if p = 1 then q = and if p > 1 then

1 / p + 1 / q = 1). But is arbitrary, so the LHS must be zero. So D u = v in the weak
sense.
Conversely, suppose that D u = v in the weak sense and let K be a compact subset
of . Then J u C (K) if < dist(K,) and we have for all x in K
xy
)u(y) dy

xy
= n (1)|| Dy (
)u(y) dy

x y
= n (
)v(y) dy

= J v(x).

D J u(x) = n Dx (

But by Lemma 2, || J u u|| Lp (K ) 0 and || D J u v|| Lp ( K) =|| J v v|| Lp ( K) 0 as 0.

Thus D u = v in the strong sense.
Definitions (i)
(ii)
(iii)

|u|j, p = (|| j | D u(x)| p dx )1/ p .

C j, p () = {u C j (): | u| < } .
j, p

j,p

() = completion of C j, p () with respect to the norm | | j, p .

H j , p () is called a Sobolev space. We will encounter other such spaces as well.

Recall that the completion of a normed linear space is a larger space in which all Cauchy
sequences converge (i.e. it is a Banach space). It is constructed by first defining a space of
equivalence classes of Cauchy sequences. Two Cauchy sequences {xm} , {ym} are said to
be in the same equivalence class if lim || xm ym ||= 0 . A member x of the old space is
m

identified with the equivalence class of the sequence {x, x,x, . . .} of the new space and in
this sense the new space contains the old space. Further, the old space is dense in its
completion. Moreover, if a normed linear space X is dense in a Banach space Y, then Y is
the completion of X.
Recall that for 1 p < , Lp () is the completion of C0 () with respect to the
usual "p norm". This knowledge allows us to see what members of H j , p () "look like".
Members of Lp () are equivalence classes of measurable functions with finite p norms,
two functions being in the same equivalence class if they differ only on a set of measure
zero.
Suppose that {um} is a Cauchy sequence in C j, p (). Then for || j , {D um} is a
Cauchy sequence in Lp (). Hence, there are members u of Lp () such that D um u
in Lp (). Hence, according to our definition of strong derivatives, u 0 is in Lp () and u
is the strong derivative of u 0 . Hence we see that
H j , p () = {u Lp (): u has strong Lp derivatives of order j in Lp () and there exists a
sequence {um} in C j, p () such that D um D u in Lp ()}.
Definition

W j , p () = {u Lp (): the weak derivatives of order j of u are in

Lp ()}

Note that by Theorem 3, an equivalent definition of W j , p () is obtained by writing "strong

derivatives" instead of "weak derivatives". Because of this, we see easily that
H j , p () W j , p () . In fact, H j , p () = W j, p (). This is not obvious because for
members of H j , p () we can find sequences of nice functions such that D um D u in
the topology of Lp (), while according to our definition of strong derivatives, such limits
exist only in the topology of Lploc () for members of W j , p () . Before proving that
H j , p () = W j, p (), we need the concept of a partition of unity.

LEMMA 4
Let E R n and let G be a collection of open sets U such that
E {U: U G} . Then there exists a family F of non-negative functions f C0 (R n ) such
that 0 f (x) 1 and
(i)
for each f F , there exists U G such that supp f U ,
(ii)
if K E is compact then supp f K is non-empty for only finitely many
f F ,
(iii)
f F f (x) = 1 for each x E (because of (ii), this sum is finite),
(iv)

if G = {1 , 2 , . ..} where each i is bounded and i E then the family

F of such functions can be constructed so that
supp f j j .

F = {f 1 , f 2 , ..} and

The family of functions F is called a partition of unity subordinate to the cover G.

P ROOF
Suppose first that E is compact, so there exists a positive integer N such that
N
E i =1Ui , where each Ui G . Pick compact sets Ei Ui such that E Ni =1 Ei . Let
gi = J i Ei , where i is chosen to be so small that supp gi Ui . Then gi C0 (Ui ) and
gi > 0 on a neighborhood of Ei . Let g = i =1 gi , and let S = supp g Ni =1Ui . If
N

< dist(E,S) then k = J S is zero on E and h = g + k C (R n ). Further, h > 0 on R n

and h = g on E. Thus F = {f i : f i = gi / h} does the job.
If E is open, let
1
Ei = E Bi (0) {x: dist(x,E) }.
i

Thus Ei is compact and E = iN=1 Ei . Let Gi be the collection of all open sets of the form
U [Interior(Ei +1 ) Ei 2 ], where U G and E0 = E1 = . The members of Gi provide
an open cover for the compact set Ei Interior(Ei1 ), so they possess a partition of unity
F i with finitely many elements. We let
s(x) = i =1 gF g(x)

and observe that only finitely many terms are represented and that s > 0 on E. Now we let
F be the collection of all functions of the form

 g(x) ,
f (x) = s(x)
0,

x E
x E

This F does the job.

If E is not open, note that any partition of unity for U is a partition of unity for
E.
For the proof of (iv), let H be the partition of unity obtained above and let f i = sum
of functions h in H such that supp h i , but supp h j , j < i . Note that each h is
represented in one and only one of these sums and that the sums are finite since each i is
a compact subset of E. Thus the functions f i provide the required partition of unity.
THEOREM 5 (Meyers and Serrin, 1964)

H j , p () = W j, p ().

We already know that H j , p () W j , p () . The opposite inclusion follows

if we can show that for every u W j ,p and for every > 0 we can find w C j, p such that
P ROOF

for || j , || D w D u||L p () < .

For m 1 let

m = {x : || x||< m, dist(x,) >

1
}
m

and let 0 = 1 = . Let { m} be the partition of unity of part (iv), Theorem 4,

subordinate to the cover { m+ 2 m} . Each u m is j times weakly differentiable and has
support in m +2 m . As in the "conversely" part of the proof of Theorem 3, we can pick

m > 0 so small that wm = J m (u m ) has support in m +3 m 1 and |wm u m | j, p < m .

2

Let w = m =1 wm . This is a C function because on each set m +2 m we have

w = wm 2 + wm 1 + wm + wm +1 + wm +2 . Further,
|| D w D u||L p () =|| m =1 D (wm u m )||L p ()
m =1 || D (wm u m )|| Lp ()
m =1 / 2 m = .

Remarks
(i)
The proof shows that in fact C () C j, p () is dense in W j , p () .
(ii)
Clearly members of C () C j, p () are not necessarily continuous on or
even bounded near . It would be very useful to have the knowledge that
C ( ) C j , p () or C j ( ) C j, p () is also dense in W j , p () . But the following
example shows that this cannot always be expected.
Problem 1 Let = {(x, y) : 1 < x 2 + y 2 < 2, y 0 if x > 0}, i.e. an annulus minus the
positive x-axis. Let w(x, y) = , the angular polar coordinate of (x,y). Clearly w is in
W1,1 () because it is a bounded continuously differentiable function. Show that we cannot
find a C1 ( ) such that |u |1,1 < 2 . (Note that is the whole annulus).
The reason for the failure of the domain in Problem 1 is that the domain is on each
side of part of its boundary. The following definition expresses the idea of a domain lying
on only one side of its boundary.
Definition A domain has the segment property if for each x there exists an
open ball U centered at x and a vector y such that if z U then z + ty for
0 < t < 1.
We will not need the following theorem, so we don't prove it. For a proof, see Adam's
book. However, see Lemma 9 for the simpler version of the result that we will need.
THEOREM 6 If has the segment property then the set of restrictions to of functions
in C0 (R n ) is dense in W m, p () .
THEOREM 7 Change of Variables and the Chain Rule.
Let V, be domains
n
1
in R and let T: V be invertible. Suppose that T and T have continuous, bounded
derivatives of order j . Then if u W j ,p () we have v = u o T W j , p (V) and the
derivatives of v are given by the chain rule.
P ROOF
Let y denote coordinates in and let x denote coordinates in V
( y = T(x) ). If f Lp () then f oT Lp (V ) because

| f T|
V

dx = | f | p J dy const. | f | p dy

(5)

(Here J is the Jacobian of T 1 ).

If u W j ,p (), let {um} be a sequence in C j, p () converging to u in W j , p () and
set v m = um oT . By the chain rule, if || j
Dx vm = (Dyum )oT R,
where the R , are bounded terms involving T and its derivatives. But for || j
Dy u Lp () (Dy u)oT Lp (V) (Dyu)oTR , Lp (V) since the R , are bounded.
Further,
|| Dx vm (Dy u)oTR , || Lp (V ) =|| (Dy um Dyu)oTR , || Lp (V )
||(Dy um Dy u)oTR, || Lp (V )
const. ||(Dyum Dy u)oT|| Lp (V )
const. || Dyum Dy u|| Lp ()
by (5). So ( = 0 case), v m v = uoT in Lp (V ) and Dx vm (Dy u)oT R, in
Lp (V ). This shows that v W j, p (V) and Dx v = (Dy u)oT R , .

Definition

W0j , p () =completion of C0 () with respect to the norm | | j, p .

Remarks

(i)
Clearly W0j , p () W j , p () because C0 () C j , p ().
(ii)
Saying that f W0j, p () is a generalized way of saying that f and its
derivatives of order j 1 vanish on . e.g. W01, p () W2, p () is a
useful space for studying solutions of the Dirichlet problem for second
order elliptic PDE's.
(iii)
C0j () W0j, p () because if f C0j (), we know that if is
sufficiently small then J f C0 () and J f f in | | j, p norm.

Problem 2 Show that W j , p (Rn ) = W0j, p (Rn ) . Hint: Why is it enough to show that
C j, p (Rn ) W j , p (Rn ) ?
0

Problem 3

Show that if is a domain in R n , f W0j, p () and if f is extended to be

zero outside then the new function is in W j , p (Rn ).

Problem 4 Show that if y C1 [0,1] and y(0) = y(1) = 0 then y W01, p (0,1) . Use this
fact to show that for any f Lp (0,1) there is a unique y W01, p (0,1) W2, p (0,1) such that
y" y = f . Hint: Solve the problem first with f C0 (0,1) and then take limits.

2.

Extension Theorems

Most of the important Sobolev inequalities and imbedding theorems that we will
derive in the next section are most easily derived for the space W0j , p () which (see
Problem 3) can be viewed as being a subspace of W j , p (Rn ). Direct derivations of these
results for the spaces W j , p () are tedious and difficult because of the boundary behavior
of the functions (Adams uses the direct derivation approach in his book). In this section we
investigate the existence of extension operators that allow us to extend functions in
W j , p () to be functions in W j , p (Rn ). This will allow us to easily deduce the Sobolev
imbedding theorems for the spaces W j , p () from the corresponding results for W j , p (Rn ).
LEMMA 8
p

Let u Rn and f Lp (R n ). Set f (x) = f (x + u) . Then lim f = f in

0

L (R ).
P ROOF

Given > 0, let C0 (R n ) be such that || f ||L p < . Since

uniformly on a sufficiently large ball containing the supports of all (say, for 1), we
can pick so small that || || Lp < . Then
|| f f || Lp || f ||L p +|| || Lp +|| f ||L p < 3 .
LEMMA 9

Let R+n = {x R n : xn > 0} . C (R+n ) C j, p (R+n ) is dense in W j , p (R+n ).

P ROOF
Suppose f is in W j , p (R+n ) let > 0 and pick C (R+n ) C j, p (R+n ) so that
|| D D f || Lp ( Rn+ ) < for all || j . We take the vector of Lemma 8 to be
u = (0,0,0, . . ,1) and define functions Lp (R n ) as
D (x) , x n > 0
(x ) =
, xn 0
0
Observe that for each > 0, C (R+n ) C j, p (R+n ) . By Lemma 8, we can pick > 0
so that, for all || j , || || Lp ( R n ) < . But this implies that || D D || Lp (R n+ ) < .
Hence
|| D D f || Lp ( Rn+ ) || D D || Lp (R n+ ) +||D D f || Lp ( R+n ) < 2 .

10

LEMMA 10

There exists a linear mapping E0 : W j, p (R+n ) W j, p (R n ) such that E0 f = f

n

in R+n and |E 0 f | Rj , p C| f | Rj,+p , where C depends on only n and p.

If f C (R+n ), define

P ROOF

f (x)
, xn 0

E0 f (x) = j +1
k =1 ck f (x1 , x 2 , . . , xn 1 , kxn ) , xn < 0
where the constants ck are chosen so that E0 f (x) C j (Rn ) , i.e.

j+1
k =1

(k)m ck = 1, m = 0,1,2, . . , j .

It is easy to check that there is a constant C depending on only n and p such that
|| D E0 f || Lp (R n ) C|| D f || Lp ( Rn+ ) .

(6)

If now f W j, p (R+n ) , take a sequence f m C (R+n ) C j, p (R+n ) converging to f in

W j , p (R+n ) (we can do this by Lemma 9). Then f m is a Cauchy sequence and (6) implies
that E0 f m is a Cauchy sequence in W j , p (Rn ). We denote the limit by E0 f . Since
|| D E0 f m || Lp (R n ) C|| D f m || Lp (R+n ) , taking limits shows that f satisfies (6).

Definition

A domain is of class C m if can be covered by bounded open sets j

such that there are mappings j : j B , where B is the unit ball centered at the origin
and
(i)

j ( j ) = B R+n

(ii)

j ( j ) = B R+n

(iii)

m
j C m ( j ) and 1
j C (B ).

(Because of (iii), all derivatives of order m of j and its inverse are bounded).
THEOREM 11 If is a bounded domain of class C m then there exists a bounded linear
extension operator E:Wm , p () Wm , p (Rn ).

11

P ROOF
Since is compact (boundaries are always closed), we might as well
assume that the number of sets j covering is a finite number N. Let U = Nj =1 j and
let d = dist(,U ). Setting 0 = {x : dist(x,) > d / 2} , we see that
0 ,1 , 2 , . ., N cover . These sets also cover , which is compact, so by the first
part of the proof of Lemma 4, there exists a finite partition of unity 0 , 1 , 2 , . . , N for
and supp j j . Recall that the support of a function is the closure of the set on which
that function is non-zero. Hence, supp j is even bounded away from j .
Let f W m, p () . Then f j W m, p ( j ), so by our chain rule theorem
m, p
(Theorem 7) w j = ( f j ) o 1
(R+n B). Clearly supp w j is bounded away from
j W

B , so we can extend w j to be a member of W m, p (R+n ) by letting it be zero in R+n B . We

can further extend w j to all of R n by use of the extension operator E0 of Lemma 10. Let
w j = E0 w . If < 1 is chosen so that supp j o 1
j B (0) , then we observe from the way
that E0 was constructed that supp w j B (0). Consequently, supp w j o j j (B (0)) is
bounded away from j . Further, again by Theorem 7, this function is in W m, p ( j ). We
extend it to be in W m, p (Rn ) by defining it to be zero outside j . If we call the extended
function g j , it is clear from our construction that g j = f j on j and that
|g j | Rm, p C| f |m , p , where C is independent of f. Finally, we let g0 denote the function
n

obtained by extending f0 to be zero outside and define Ef = Nj = 0 g j .

Remarks
The theorem can be improved in a number of ways:
(i)
We can allow to be unbounded if is bounded (e.g. is the exterior of a
bounded domain).
(ii)
We can allow to be of class C m 1,1 instead of C m (i.e. the derivatives of order
m 1 of the functions j are Lipschitz continuous. The proof of this requires a

(iii)

better version of Theorem 7 which we don't have time to prove here. Note that for
the case m = 1, the boundary could have corners.
Caldern has proved an extension theorem for domains satisfying the cone property
(see the definition below) and a few other minor assumptions. The proof is much
too time-consuming for us and it relies on the Caldern-Zygmund inequality, which
also has a very lengthy proof. (See [Ad] for this).

Definition A domain is said to satisfy the cone property if there exist positive
constants , h such that for each x there exists a right spherical cone V x with
height h and opening .

12

3.

Sobolev Inequalities and Imbedding Theorems

THEOREM 12 If R n satisfies the cone condition (with height h and opening ) and if
p > 1, mp > n then W m, p () CB () and there is a constant C depending on only , h,
n and p such that for all u Wm , p (), sup|u| C|u|m , p .
Note: does not have to be bounded as Friedman suggests in his Theorem 9.1!
P ROOF
Initially, suppose that u is in C m, p (). Let g C (R) be such that g(t) = 1
if t 1 / 2 and g(t) = 0 if t 1. Let x and let (r, ) denote polar coordinates centered
at x. Here, = (1 , 2 , . . , n1 ) denotes the angular coordinates and we can describe the
cone with vertex x in polar coordinates as V x = {(r,) : 0 r h, A}. Clearly, we have

{g(r / h)u(r,)} dr
r
(1)m h m 1 m
=
r
{g(r / h)u(r, )} dr ,
(m 1)! 0
rm
h

u(x) = 0

after m-1 integrations by parts. Next, we integrate with respect to the angular measure dS ,
noting that the left-hand-side becomes a constant times u(x).
m
u(x) = c A 0 r
{g(r / h)u(r, )} drdS
r m
h
m
= c A 0 r m n m {g(r / h)u(r, )} r n1drdS
r
m

= c V r m n m {g(r / h)u(r, )} dV .
x
r
h

m 1

Applying Hlder's inequality to this, we obtain

|u(x)| const.||r m n || Lq (V x ) || r m {g(r / h)u(r,)}|| Lp (V x )
m

const.|| rm n || Lq (V x ) |u|m, p .
But r m n is in Lq (V x ) if n 1+ (m n)q > 1, which is the case because q =

p
p1

and

mp > n . Thus, we obtain sup|u| C|u|m , p . To extend this result to arbitrary u W (),
take a sequence {uk} of functions in C m, p () converging to u in the | | m, p norm. Then
sup|u j uk | C|u j uk |m , p , showing that the sequence is a Cauchy sequence in CB ().
m, p

13

Thus u is in CB () and taking the limit of sup|u j | C|u j |m , p shows that u satisfies the same
inequality.

Problem 5. Modify the proof to show that the theorem also applies to the case p = 1,
m = n.
Problem 6. Show that a similar theorem holds for W0m, p () and the cone condition is
not required. Note that here we can even conclude that W0m, p () {u CB ( ) : u = 0 on
} .
COROLLARY 13
If R n satisfies the cone condition (with height h and opening
) and if p > 1, (m k)p > n then W m, p () CBk () and there is a constant C depending
on only , h, n, k and p such that for all u Wm , p () sup | D u| C|u|m , p .
||k

P ROOF

Apply the previous theorem to the derivatives D u for || k .

Problem 7
Problem 8

What can you conclude if p = 1 and m k = n ? See Problem 5.

What is the corresponding theorem for W0m, p () ? See Problem 6.

THEOREM 14 If R n is any domain and p > n then W01, p () C 0, ( ), where

= 1 np and there exists a constant C depending on only p and n such that for all
u W01, p ()
n
|u(x) u(y)|

C
|| Diu|| Lp () .

|| x y||
i=1

P ROOF
Let u C0 () . We might as well assume that u C0 (Rn ) . Let
d =|| x y|| , Sx = Bd (x), Sy = Bd (y) and S = Sx Sy . Then
|u(x) u(y)| vol(S) = S |u(x) u(y)| dz

S |u(x) u(z)|+|u(z) u(y)| dz

S |u(x) u(z)| dz + S |u(z) u(y)| dz

x

But if (r, ) are the polar coordinates of z in a coordinate system centered at x, we get
u
|u(x) u(z)| r0 |
| d , which implies

14

Sx

=
=
=

u
| d rn 1drdS

u
| d rn1 drdS

n
d
d
u
| | ddS

0
n

n
d
d
u
1 n | | n 1d dS

0
n

n
d
1 n u

| | dz

n Sx

n
d
u
|| 1n || Lq (S x ) || || Lp (S x )
n

where q =

|u(x) u(z)| dz 0 0 |

|
0

p
. A simple calculation shows that
p 1

||

1 n

|| Lq (S x ) = const. d

n
p

and it is easy to see that

||

n
u
|| Lp (S x ) const. || Di u|| Lp () .

i =1

Also, vol(S) = const. d n and the integral over Sy can be estimated in a similar fashion.
Putting this together yields
1

|u(x) u(y)| Cd

n
p

|| D u||
i

Lp ()

i=1

which is precisely the inequality that we wanted. Further, we know from Theorem 12
applied to R n that sup|u| C|u|1, p . Combining this with the previous inequality shows that
for u C0 () we have || u||C 0, ( ) C|u|1, p . Thus, if we now let u W01, p () and take a
sequence {um} of functions in C0 () converging to u in | |1,p norm, it follows that {um}
converges in C 0, () . Thus u C0, ( ) , and taking limits shows that u satisfies the
inequality in the statement of the theorem.

15

THEOREM 15 If R n is any domain and p < n then W01, p () Lr (), where

np
r=
and there exists a constant C depending on only p and n such that for all
n p
u W01, p ()
n

|| u|| Lr () C || Diu||L p () .
i =1

Remark
The proof relies on a simple generalization of Hlder's inequality which can
be proved by induction by using Hlder's inequality. The inequality states

where

|u1u2 u3 . . um | dx ||u1 ||L p1 ||u2 || Lp 2 . . ||um || Lp m

(7)

1
1
1
1
+
+
+... +
= 1.
p1 p2 p3
pm
It suffices to prove the result for u C01 (R n ). First we prove

P ROOF of Theorem 15.

the result for the case p = 1. For each i we have

xi

|u(x)| | Diu| dxi | Di u| dxi .

Multiplying these n inequalities together and taking the n 1 th root gives
n

|u(x)| n 1 ( | Diu| dxi ) n1

(8)

i =1

Observe that

| Diu| dxi does not depend on xi , but it does depend on all n 1 of the

remaining variables. We integrate each side of (8) with respect to x1 and use the
generalized Hlder inequality with pi = m = n 1 to obtain

1
n1

|u(x)| n1 dx1 ( | D1u| dx1 )n 1

( | D1u| dx1 )

16

i= 2

(
i=2

( | Diu| dxi )n 1 dx1

| Di u| dxi dx1 )

1
n1

The RHS is still a product of n 1 functions of x2 , so we integrate each side with respect
to x2 , again applying (7) with pi = m = n 1. Continuing in this manner, we finally obtain

Rn

|u(x)|

n
n1

dx ( Rn | Diu| dx )
i =1

i.e.

1
n1

|| u||

Ln1

( R n | Di u| dx )

1
n

(geometric mean)

i =1

1 n

n i=1

Rn

|Diu| dx

(arithmetic mean)

Here we have used the fact that an arithmetic mean is no less than a geometric mean of the
same numbers. This proves the result for the case p = 1.
(n 1)p
n(p 1)
For p > 1, let =
=1+
. Since > 1 and u C01 (R n ), it follows
np
np
that |u| C01 (R n ) . Clearly
(n 1) p
Di |u| =
|u|
n p

n( p1)
n p

( Diu) .

We apply the p = 1 case to |u| and obtain

( R n |u|

np
n p

n 1
n

n( p1)

1
(n 1)p
dx ) R n
|u| n p | Diu| dx
n
n

p
i =1
n( p1)
p
p1
(n 1)p n
n p
p1

( R n (|u|
) dx ) p || Diu|| Lp

n(n p) i =1
np
p1
(n 1)p n
n p
=
(R n |u| dx ) p || Diu||L p

n(n p) i =1
n

Hence
( R n |u|

np
n p

dx )

np
np

(n 1) p n
|| D u|| p ,
n(n p) i =1 i L

which is the desired result. As usual, to obtain the same result for a function u W01, p (),
we just take a sequence of functions in C10 converging to u.

17

Remark
According to the theorem, W01, p () Lr (), where r is given above. But
obviously W01, p () Lp (), so by the following interpolation lemma, W01, p () Lq ()
for all q satisfying p q r . If is bounded then clearly this holds for all q satisfying
1 q r.
LEMMA 16

If s q r and Ls () Lr () , then Lq () and

|| ||Lq || || Ls ||||1Lr ,

s(r q)
.
q(r s)

where =

Apply Hlder's inequality to the integral of ||q , using the facts that

P ROOF
(1 )q

||

r
(1 )q

s
q

and || L .

Modify the proof of Theorem 15 to show that if p = n > 1 then

n2
1, p
r
W0 () L () for every r p. Hint: First prove this for the case r >
by setting
n 1
n 1
n
=r
and applying the p = 1 result to |u| . The r
norm that shows up on the
n
n1
RHS after applying Hlder's inequality (as we did in our proof above) can be estimated in
terms of the n = p norm and the r norm by use of Lemma 16. Finally, obtain the result for
all r p by applying Lemma 16 to the result that you have just proved.
Problem 9

COROLLARY 17
For every domain in R n there exists a constant C depending on
only n and p such that
a)

if kp < n then W

k, p
0

np
n kp

() L

() and for each u W0k , p ()

|| u||
L

np
nkp

C|u|k , p

if kp > n then W0k, p () C m , () , where m is the integer satisfying

n
n
0 < k m < 1 and = k m . Further, if u W0k , p () then
p
p

b)

|| u||C m, ( ) C|u| k ,p .

18

a) If || k 1 and u W

k, p
0

P ROOF

np
k 1,
n p
0

by Theorem 15. Thus W0k, p W

np
k 2,
n2 p
0

W0k, p W

np
n p

then D u W , which is contained in L

1, p
0

. Iterating this process once more, we find

. Continuing the iterations culminates in the desired result.

k m 1, p
0

b) Since (k m 1)p < n , case (a) implies that W

np
n (k m1)p

n
m +1,
(1 )
0

u W0k , p and || m + 1, then D u W0k m 1, p . Hence W0k, p W

n
1,
(1 )
0

that if u W0k , p and || m then D u W

Theorem 14. Thus W0k, p () C m , () .

n
(1 )

=L

. Thus if

. But this shows

, which is contained in C 0, () by

Remarks
A few "particular cases" have been left out because they require separate
proofs (see Problem 10 below). They are:
i)
ii)

If kp = n and p > 1 then W0k, p () Lq () for all q satisfying p q < .

If kp = n and p = 1 (so that k = n) then W0k, p () CB ( ) .
n

iii)

iv)

k ,q
n
If kp > n , p > 1 and is an integer then W0k, p () W0 p () for all q satisfying
p
p q < .
n
If kp > n and p = 1 (so
is obviously an integer) then W0k, p () CBk n ( ).
p

All of the particular cases listed above have the appropriate norm inequalities associated
with them.
Problem 10 Use the results of Problems 5, 6, 7, 8 and 9 to prove the particular cases
listed above.
COROLLARY 18
If is a bounded C1 domain in R n (or any other domain such that
there exists a bounded extension operator E : W1, p () W1, p (R n ) ) then the statements
concerning the spaces W0k, p () in Corollary 17 and in the remark following the corollary
also apply to the spaces W k, p (). However, the constant C may also depend on .

19

P ROOF
The cases for k = 1 dealt with in Theorems 14 and 15 are easily seen to
have their counterparts here because of the extension operator. Inspection of the proof of
Corollary 17 shows how the results for k > 1 may be derived from the results for k = 1
without any additional assumptions on the domain.

Remark
One can show that extension operators exist for Lipschitz domains and even
domains satisfying certain cone conditions (see the remarks following the proof of
Theorem 11). This Sobolev imbedding theorem is thus valid for such domains.
Definition Let A and B be Banach spaces. If A B , we say that A is continuously
imbedded in B (in symbols, this is written A\o(\s\do3( ), )B)if there is a constant C
such that || x||B C|| x|| A for all x A .
The theorems in this section provide examples of imbeddings and are called
np
n p

Sobolev Imbedding Theorems. e.g. W () \o(\s\do3( ), ) L

1, p
0

for p < n.

It is easy to see that A\o(\s\do3( ), )B is equivalent to the identity mapping from

A into B being continuous (i.e. bounded).
4.

Compactness Theorems

Definition Suppose that A\o(\s\do3( ), )B. We say that A is compactly imbedded

in B if every sequence bounded in A has a subsequence that converges in B.
e.g. If K is compact then any bounded sequence in C1 (K) is a set of equicontinuous
functions so, by the Arzela-Ascoli Theorem, it has a subsequence that converges in C(K) .
i.e. C1 (K) is compactly imbedded in C(K) .
Recall that if A and B are Banach spaces and if M : A B is a bounded linear
mapping then M is said to be compact if for every bounded sequence {xm} in A the
sequence {Mxm} has a subsequence that converges. Thus, saying that A is compactly
imbedded in B is equivalent to saying that the identity mapping from A into B is compact. It
is easy to see that if M : A B and P : B C are bounded linear mappings and A, B and
C are Banach spaces then PM is compact if one of the mappings A or B is compact.
Consequently, we obtain the very useful result that if A\o(\s\do3( ), )B and B\o(\s\do3(
), )C then the imbedding A\o(\s\do3( ), )C is compact if one of the other two
imbeddings is compact.

20

LEMMA 19 Suppose that is a bounded domain. If

a)
0 < 1 then C m, ( ) is compactly imbedded in C m ( ).
b)
0 < < 1 then C m, ( ) is compactly imbedded in C m, () .
P ROOF
It suffices to prove the results for m = 0 because, once this is done, we can
apply this case to the derivatives of the functions and deduce the result for general m.. Let
{ f j} be a sequence in C 0, ( ) such that || f j || C0, M . But this implies
| f j (x) f j (y)| M|| x y|| , showing that the sequence is a bounded, equicontinuous set
of functions. By the Arzela-Ascoli Theorem, there exists a subsequence { f jk } that
converges in C( ). Thus C 0, ( ) is compactly imbedded in C( ).
We show below that the same subsequence also converges in C 0, ( ). Suppose
that C0, ( ). Then

[ ]0,

1
| (x) (y)|
| (x) (y)|

= sup
= sup
| (x) (y)|
|| x y||
||x y||
1

([ ]0, ) (max || )

We apply this to f jk f j r , noting that [ f j k f jr ]0, [ f jk ]0, + [ f jr ]0, 2M , and obtain

[ f j k f jr ]0, 2M (max | f jk f jr | )

showing that the subsequence is a Cauchy sequence in C 0, ( ) (because it converges in

C( )). Thus the subsequence converges in C 0, ( ).

COROLLARY 20

If is bounded, kp > n and 0 < k m

compactly imbedded in C m, () if < k m

n
< 1 then W0k, p () is
p

n
.
p

n
. Then W0k, p ()\o(\s\do3( ), )C m, ( ) \o(\s\do3(
p
m,
), )C () , and the second imbedding is compact.

P ROOF

Let = k m

21

If is a bounded C1 domain (or any other domain for which there

n
is a bounded extension operator E : W1, p () W1, p (R n ) ), kp > n and 0 < k m < 1
p
n
then W k, p () is compactly imbedded in C m, () if < k m .
p
COROLLARY 21

Let C0 (R n ) be such that supp is contained in some ball B containing

and = 1 on . Then we can define E : W1, p () W1,0 p (B) by E ( f ) = E( f ) . By
Corollary 20, W01, p (B) is compactly imbedded in C 0, (B ). Hence W1, p () is compactly
P ROOF

imbedded in C 0, ( ). The result for general k can be deduced from the k = 1 case by
considering derivatives of the functions (as in the proof of Corollary 17 (b), deduce that if
u W

k, p

() and || m then D u W

1,

n
(1 )

() , which is contained in C 0, () ).

Remark

The statements of Corollaries 20 and 21 continue to hold if we replace the

n
n
condition 0 < k m < 1 by the condition 0 < k m 1, i.e. we can include the
p
p
n
case of being an integer. The proof of this is left to the next problem.
p
Problem 11 Show that if is bounded then the statements in the Remark above hold
for the spaces W0k, p (). Hint: You need to use the particular cases (iii) and (iv) following
the proof of Corollary 17. For the case p = 1, you need to use the fact that C m +1 () is
compactly imbedded in C m, () , which follows easily from Lemma 19 and the fact that
C m +1 is continuously imbedded in C m,1 ( ). Can you see how to modify your proof so that
it deals with W k, p () if is a bounded C1 domain ?
Definition A subset E of a metric space is said to be totally bounded if for each > 0,
E can be covered by finitely many balls of radius .
The following theorem is a standard result that can be found in most books on topology
and in many books on functional analysis (e.g. Rudin's "Functional Analysis", Appendix
A4). We omit the proof.
THEOREM 22
Let E be a subset of a complete metric space X. Then the following
statements are equivalent.
(i)
E is compact.

22

(ii)
(iii)

Every sequence in E has a convergent subsequence.

E is totally bounded.

The Theorem gives us two other very useful characterizations of compact mappings and
imbeddings.
If is bounded and p < n, then W01, p () is compactly imbedded in

THEOREM 23
Lq () for all q <

P ROOF

np
.
n p

Consider first the case q = 1. Let A be a bounded set in W01, p () . We may

consider the members of A as members of W1, p (Rn ) with supports contained in . Let
Ah = {J h u : u A}. Note that we have
| Jh u(x)| h n (
and

xz
)|u(z)| dz h n (max )|| u|| L1 ()
h

| Di Jh u(x)| h n 1 | Di (

xz
)||u(z)| dz h n1 (max| Di|)||u|| L1 () .
h

Since is bounded, || u|| L1 const.||u||L p . The inequalities above show that Ah is a

bounded equicontinuous set of functions in C( ). By the Arzela-Ascoli Theorem, every
sequence in Ah has a subsequence that converges in C( ). Obviously, such subsequences
also converge in L1 (), so we see that Ah is totally bounded in L1 ().
If u A then
u(x) J h u(x) =

(z)(u(x) u(x hz)) dz

|z|1

|z|1

Thus

h||z||

(z)0

|u(x) Jh u(x)|

|z |1

Integrating this with respect to x, we find

23

z
u(x r
) dr dz .
r
|| z||
h||z || n

(z)0

| D u(x r || z|| )| dr dz .
i =1

|u(x) Jh u(x)| dx

z
)| dx dr dz
||z||

(z)0

(z)0

(z)h||z|| | Di u(x)| dx dz

|z|1

|z |1

h||z || n

|z|1
n

i =1

Rn

| Di u(x r

h||z|| n

i =1

| Diu(x)| dx dr dz

i=1

h |Di u(x)| dx
i=1

hB,

(9)

where B is a constant depending on our bound of members of A in W01, p () (again we use

the fact that the L1 norm is weaker than the Lp norm on a bounded domain).
Let > 0. Since Ah is totally bounded in L1 (), we can cover Ah by a finite

number of balls Bi of radius / 2. Let h =

. By (9), if Jh u Bi , then u is contained in
2B
a ball of radius centered at the center of Bi . Thus, A is covered by a finite number of
balls of radius . i.e. A is totally bounded in L1 (). Thus W01, p () is compactly imbedded
in L1 ().
np
np

Suppose W (). Then L

np
(with s = 1 and r =
) that
n p
1, p
0

by Theorem 15 and we get from Lemma 16

|| ||Lq || || L1 || ||1np C|| || L1 (|| || Di ||L p )1

Ln p

i =1

Now let {um} be a bounded sequence in W01, p () and assume |um |1, p M . Since W01, p ()
is compactly imbedded in L1 (), we can extract a subsequence {um j } that converges in
L1 (). Applying the inequality above to um j umk , noting that |um j umk |1, p 2M , we
obtain
|| um j umk || Lq const.||um j umk ||L1 ,
showing that the subsequence is a Cauchy sequence in Lq () . Hence the subsequence
converges in Lq () and W01, p () is compactly imbedded in Lq () .

24

If kp < n and is bounded then W0k, p () is compactly imbedded

np
in Lq () for all q <
.
n kp
COROLLARY 24

P ROOF

1,

k, p
0

() is continuously imbedded in W0
np
imbedded in Lq () if q <
, by Theorem 23.
n kp
COROLLARY 25

np
n (k 1) p

(), which is compactly

The same compactness results hold for W k, p () if is a bounded,

C1 domain (or any other type of bounded domain for which there is an extension operator
E : W1, p () W1, p (R n ) .
P ROOF

See the proof of Corollary 21.

Remark
The case kp = n is missing from the previous results. But since W k, p is
continuously imbedded in W k, r for all r < p (provided that the domain is bounded), it
follows from Corollary 25 that W k, p is compactly imbedded in Lq () for all q < . The
same applies to W0k, p ().
5. Interpolation Results
The following results are very useful in PDE theory. We make use of Theorem 26
in our proof of Grding's Inequality in our study of elliptic problems.
THEOREM 26

Let u W0k , p () . Then for any > 0 and any 0 <| |< k

|| D u|| Lp |u|k , p +C

| |
k | |

||u|| Lp

where C is a constant depending only on k.

P ROOF

We prove the result for ||= 1, k = 2. The general result is easily obtained

from this case by induction. In fact, we show that for each i

||

u
2 u
72
|| Lp || 2 || Lp + ||u|| Lp
xi
x i

25

(10)

First suppose that u C02 (R) and consider an interval (a,b) of length b a = . If
y (a,a + / 3) and z (b / 3,b), then by the Mean Value Theorem there is a
p (a,b) such that
| u( p)|=|

u(z ) u(y) 3
| (|u(z)|+|u(y)|)
zy

Consequently, for every x (a,b), we obtain

x

| u(x)|=| u( p) + p u (t) dt|

b
3
(|u(z)|+|u(y)|) + a | u(t)| dt .

Integrating with respect to y and z over the intervals (a, a + / 3) and (b / 3,b)
respectively, we obtain
b

| u(x)| a | u (t)| dt +

18 b
|u(t)| dt ,
2 a

so by Hlder's inequality and the inequality (A + B) p 2 p1 (A p + B p ),

(18) p b
p
2p {a | u(t)| dt} )

b
b
b
(18) p b
p1
p
p1
2 ({a | u (t )| dt}{a 1 dt} + 2 p {a |u(t)| p dt}{a 1 dt}p1 )

p b
b
(18)
= 2 p1 ( p1 a | u(t)| p dt + p+1 a |u(t)| p dt ).

| u(x)| p 2 p 1 ({a | u(t)| dt}p +

Integrating this with respect to x over the interval (a,b) gives

b

| u (x)|
a

dx = 2

p1

(18) p b
( a | u(t)| dt + p a |u(t)| p dt ).

We now subdivide R into intervals of length and obtain by adding all of these
inequalities that

p
p 1
p
p
| u (x)| dx 2 ( | u (t)| dt +

26

(18) p
|u(t)| p dt )
p

(11)

Suppose now that u C0 (Rn ) . Then we can apply (11) to u regarded as a function of xi
and integrate with respect to the remaining variables to obtain

R n |

u p
2 u
(18) p
| dx 2 p1 ( p R n | 2 | p dx + p R n |u| p dx )
xi
x i

Taking the pth root of this and using (A p + B p )1/ p A + B, we obtain (10). (Actually, we
don't quite obtain (10). We actually obtain the inequality (10) for 2 instead of . But
since is an arbitrary positive constant, (10) holds). Finally, (as usual) to obtain the result
for u W0 (), we take a sequence of functions in C0 converging to u.

COROLLARY 27
The interpolation inequality stated in Theorem 26 also applies to
k, p
members of W (), provided that is a bounded C 2 domain (or any other domain for
which there is a bounded extension operator E : W2, p () W 2,p (R n ). Here the constant C
may also depend on p and .
P ROOF
Because of the extension operator, an inequality of the form (10) holds for
2, p
functions in W (). The full result follows by induction from this case.

6. The Spaces H k () and H0k ().

Definitions H0k () = W0k, 2 () and H k () = W k, 2 (). These spaces are Hilbert spaces
with inner product
(u,v)k =

D u(x)D v (x) dx .

|| k

If = Rn , we get a very useful representation of such functions in terms of the

Fourier-Plancherel Transform. Recall that for functions f L2 (Rn ) , we define the FourierPlancherel transform of f as
f () = lim (2 ) n/ 2
R

||x || R

27

ix .

f (x) dx .

The limit exists in the topology of L2 (R n ), || f ||L2 =|| f|| L2 , and f can be recovered by using
the inversion formula
f (x) = lim (2 ) n/ 2
R

ix .

f ( ) d .

|| || R

Again, the limit here exists in the topology of L2 (R n ). The reason for the limits in these
formulas is that the integrands are not necessarily in L1 (R n ). Clearly, the FourierPlancherel transform is an isometric (i.e. norms are equal) isomorphism (i.e. bounded
linear mapping with a bounded inverse) from L2 (R n ) onto L2 (R n ).
f
Integration by parts shows that for f C0 (R n ), the transform of
is i j f ( ).
x j
From this, we see by induction that the transform of D f is (i ) f(). If now we let
f H k (R n ) and take a sequence of C functions converging to f, we find that (i ) f()
0

is in L (R ) for all || k and the transform of D f is (i ) f(). Thus, we see that if

f H k (R n ) then f H k (R n ), where
2

H k (R n ) = {g L2 (R n ) : (1+||||) k g( ) L2 (R n )}.
It is easy to see that C0 is dense in H k and if g is in C0 then g is the transform of an
infinitely differentiable, rapidly decaying function f (a function in the Schwarz class, to be
precise). Taking limits, we see that if g is in H k then g is the transform of a function f
belonging to H k . Further, if we define an inner product on H k as
(u,v)H k =

u( )v () d ,

| |k

we find that || f ||k =|| f|| H k . Thus, the Fourier-Plancherel transform is an isometric
isomorphism from H k onto H k . Questions about functions in H k are thus transformed
into equivalent (and often simpler) questions about functions in H k .
Problem 12 Consider the initial value problem for the wave equation
d 2u
du
k
(0) = g H k 1 .
2 = u , u(0) = f H ,
dt
dt

28

(We think of u as being a function of t taking values in H k ). Construct a candidate u for a

solution using Fourier transforms.
a)
Show that u is a continuous H k -valued function of t.
b)
Show that u is a continuously differentiable H k 1 -valued function of t.
du
c)
Show that
is a continuously differentiable H k 2 -valued function of t and that
dt
d 2u
k 2
.
2 = u in H
dt
d)
How large does k have to be in order for u to be a classical (i.e. C 2 ) solution.
Hints: Clearly it suffices to answer the equivalent questions about u . Use the Dominated
Convergence Theorem to help you answer a), b), c). For d), use the Sobolev Imbedding
Theorem.
7. Trace Theorems.
In PDE Theory, one often needs to know how functions behave on boundaries of
domains. If f is a function defined on a domain , we call the restriction of f to the
trace of f. If all we know about f is that it is in some Lp space, then the trace of f is not
well-defined because has measure zero. However, if kp > n and is a bounded C1
domain in R n , then we know by Corollary 18 that functions in W k, p () are continuous on
and thus they have well-defined traces that are bounded functions. In this section, we
concern ourselves with the important case kp < n.
In the following results, a vector x in R n is denoted by x = ( x , xn ) , where x
belongs to R n1 .
LEMMA 27 If u W1,1 (Rn ) , then for every R , the function v( x ) = u( x , ) is in
L1 (R n1 ) , and
|| v|| L1 ( R n1 ) ||u|| L1 (R n ) +|| Dn u|| L1 (R n ) .
Remark.
One needs to be careful when talking about traces of equivalence classes of
functions. The trace certainly exists for u C0 (Rn ) . For u W1,1 (Rn ) , we know that we
can find a sequence of functions in C0 (R n ) that converges to u. The norm inequality
asserted in the lemma shows that the sequence of traces of these functions converges in
L1 (R n1 ) . It is in this sense that the trace of u exists in L1 (R n1 ) .

29

P ROOF

It suffices to prove the result for the case = 0 and u C0 (Rn ) . By the

Mean Value Theorem for integrals

1

|u( x , x
0 R

|u( x ,)| dx

)| dx dxn =

n1

R n1

for some [0,1]. But

|u( x ,0)|=|u( x , ) Dn u( x ,t) dt |

0

|u( x ,)|+ | Dn u( x ,t)| dt .

0

Integrating this over R n1 gives

|| v|| L1 ( R n1 )

|u( x ,)| dx +
n1

R
1

| D u( x ,t)| dtdx
n

R n1 0
1

|u( x , )| dx dt + | D u( x ,t)| dtdx .

n

0 R n1

R n1 0

This completes the proof of the lemma.

LEMMA 28 If u W1, p (R n ) where
v( x ) = u( x , ) is in Lr (Rn 1 ), where
r=

p < n, then for every R , the function

(n 1)p
n(p 1)
= 1+
np
n p

and there is a constant C depending on only n and p such that

n

|| v|| Lr (R n1 ) C |u|1,R p .
P ROOF
We can assume that p > 1 because the p = 1 case is dealt with in the
previous lemma. We first show that if u W1, p (R n ) then w =|u|r W1,1 (Rn ) and
r
|| w||L1 const. || Du||rL1
p ||u|| p , || Di w|| 1 const. || Du|| p .
L
L
L

30

(12)

It suffices to prove this result for the case u C0 (Rn ) . Let q = p / (p 1) . Then
(r 1)q = np / (n p) , so by the Sobolev Imbedding Theorem (Th. 15),
/ (n p)
|| |u|r 1 ||qLq const.||Du||np
Lp

and combining this with Hlder's Inequality, we get the first of (12):
|| w||L1 = |u|r dx = |u| r1 |u| dx ||u|| Lp || |u|r 1 || Lq const.|| Du||rL1
p ||u|| p .
L
Since Diw = r|u| r1 Diu , we obtain the second of (12):
|| Di w|| L1 = r|| |u|r 1 ||Lq || Diu|| Lp const. || Du||rL p .
We now apply Lemma 27 to w and immediately obtain the inequality
|| v|| Lr (R n1 ) const.(|| Du|| r1
||u||L p +|| Du|| rLp )1/ r
Lp
r
const.(|| Du||11/
||u||1/Lpr +||Du||L p )
Lp

const.(||u|| Lp +||Du|| Lp ).
LEMMA 29 If u Wk , p (Rn ) where
v( x ) = u( x , ) is in Lr (Rn 1 ), where

kp < n , then for every R , the function

r=

(n 1)p
n kp

and there is a constant C depending on only n, k and p such that

n

|| v|| Lr (R n1 ) C |u|kR, p .
P ROOF
By Sobolev's Imbedding Theorem (Th. 15) applied to the first order
derivatives of u, we have u W1,np / (n(k 1)p ) (R n ) . Now apply Lemma 28.

31

Reminder:

Parametrized Surface Integrals.

If X(u) = (x1 (u1 ,u2 ), x2 (u1 ,u2 ), x3 (u1 ,u2 )) is a parametrization for a smooth surface
S in R3 , it is well-known from elementary calculus that one may integrate functions defined
on S using the formula

f (x) dS = f o X(u) K(u) du1 du2 ,

where is the domain of X and

K(u) =||

X X
(x 3 , x2 ) 2
(x ,x )
(x , x )

|| = ((
) + ( 1 3 )2 + ( 1 2 ) 2 )1/ 2 .
u1 u2
(u1 ,u2 )
(u1 ,u2 )
(u1 ,u2 )

Differential Geometry yields a generalization of this formula. Suppose now that

is a domain in R n1 and that X: Rn is a parametrization for a smooth hypersurface S.
Then surface integrals over S may be calculated using

f (x) dS = f o X(u) K(u) du ,

where
n

K(u) = ( (
k =1

(x1 ,x 2 ,K, x k ,K, xn ) 2 1/ 2

) ) .
(u1 ,u2 ,K,un )

Here the xk notation means that the xk term does not appear.
THEOREM 30 Suppose that is bounded and is of class C k . If u Wk , p () where
kp < n , then the restriction v of u to is in Lr (), where
r=

(n 1)p
n kp

and there is a constant C depending on only n, k and p and such that

|| v|| Lr () C |u|k , p .

32

Remark.
See the remark following the statement of Lemma 27 for clarification of the
phrase "restriction v of u to ". The same remark applies because we know that an
extension operator E: W k , p () Wk , p (Rn ) exists, and thus the restriction to of
functions in C0 (R n ) is dense in W k, p ().
P ROOF
Let E: W k , p () Wk , p (Rn ) be the extension operator of Theorem 11.
Since any u Wk , p () is associated with an element Eu Wk , p (Rn ), we might as well just
study the properties of the trace on of C0 (R n ) functions.
Let j and j be as in the definition of C k domains. Since is compact, we
might as well assume that there is a finite number of the j , 1 j N , covering . Let
j 1 j N , be a partition of unity for subordinate to this cover. If u C0 (Rn ) , then
k
1
( ju)o 1
to be in C0k (Rn ) by defining the
j C0 (B) and we can extend ( ju)o j

function to be zero outside B. By Lemma 29, the trace w j of ( ju)o 1

j on the hyperplane
P: yn = 0 satisfies
B
R
|| w j ||Lr (P ) C|( ju)o 1
j | k , p Cj |u| k , p ,
n

where C depends on only n,

p, and k

and

Cj

is independent of u..

1
j

X j (y) = (y1 ,K,y n1 ,0) is a parametrization for the hypersurface S j = () j and

we may estimate the trace v j = wj o j of ju on this hypersurface using this
parametrization (see the "reminder" preceding the statement of the theorem).

|v (x)|

Sj

dS =

|w (y)|

K j (y) dy Rj

P B

|w |

dy ,

P B

where Rj = max(K j ). Comparing this to the preceding inequality, we see that there is a
constant M j independent of u such that
n

|| v j || Lr (S j ) M j |u| Rk, p .
The function v satisfies a similar inequality because v = v j . Finally, it is clear that the
result holds for arbitrary u Wk , p () (see the remark following the statement of the
theorem).

33

Problem 13 Modify the proof of Lemma 27 to show that if u W1, p (R n ), then for every
R , the function v( x ) = u( x , ) is in Lp (R n1 ) , and there exists a constant K
n
depending only on n and p such that|| v|| Lp (R n1 ) K|u|1,R p .
Problem 14 Deduce from the previous problem and Lemma 28 that the function v of
Lemma 28 belongs to Lq (R n1 ) for all q satisfying p q r .
Appendix:

Some Spaces of Continuous Functions.

Here, we define the spaces of continuous functions that appear in these notes. Caution:
Notation and definitions of such function spaces vary from text to text. Recall that we
stated that is a domain in R n . The connectedness of is not needed in the following
definitions, so we need only assume that is an open subset of R n .
1.
2.
3.

C() is the set of functions continuous in .

C( ) is the set of functions continuous in .
C k () is the set of functions which have derivatives of order k that are

5.
6.
7.

continuous in .
C k ( ) is the set of functions in C( ) which have derivatives in of order k
that can be extended to be members of C( ).
C () is the set of functions in C k () for all k.
C ( ) is the set of functions in C k ( ) for all k.
C0 () is the set of functions in C() that have supports that are compact subsets

8.

of (recall that the support of a function is the closure of the set on which the
function fails to vanish). Since is open, such functions necessarily vanish in a
neighborhood of the boundary of .
C0k () is the set of functions in C k () that have supports that are compact subsets

9.

of .
C0 () is the set of functions in C0k () for all k.

10.

CB () is the set of bounded functions in C(). This is a Banach space when

11.

equipped with the "sup norm".

CB ( ) is the set of bounded functions in C( ). This is a Banach space when
equipped with the "sup norm". If is bounded, this space coincides with C( ).

4.

34

12.

CBk () is the set of functions in CB () with derivatives of order k belonging to

CB (). This is a Banach space if we define the norm of a member f of this space as
sup | D f (x)|.
|| k, x

13.

14.

CBk ( ) is the set of functions in both CBk () and C k ( ). This is a Banach space,
equipped with the same norm as in (12). If is bounded, this space coincides with
C k ( ).
C k , () , where 0 < 1, is the set of functions in CBk ( ) that have derivatives of
order k that are uniformly Hlder continuous with exponent . C k , () is a
Banach space with norm
|| f ||C k, = sup | D f (x)|+[ f ]k, ,
|| k, x

where

[ f ]k , =

| D f (x) D f (y)|
.
x , y, x y, | |= k
|| x y||
sup

References
The results stated in these notes appear in most texts on Sobolev spaces, including
those listed below. However, there are many different proofs of the results. For this
reason, the key lemmas and theorems that appear in these notes are listed below with a
reference to the source which has a proof that most resembles the proof in these notes.
Th. 1
L. 2
Th. 3
L. 4
Th. 5
Th. 6
Th. 7
Th. 11
Th. 12
Th. 14
Th. 15
L. 19

[Fr], Th. 6.1

[Fr], Th. 6.2
[Fr], Th. 6.2
[Zi], L. 2.3.1
[Fr], Th. 6.3; [Ad], Th. 3.16.
[Ad], Th. 3.18.
[Ad], Th. 3.35. (see [Zi], Th. 2.2.2 for Lipschitz changes of variables).
[Ad], Th. 4.26; [G.T.], Th. 7.25.
[Fr], Th. 9.1
[Fr], Th. 9.2
[G.T.], Th. 7.10
[Ad], Th. 1.31

35

Th. 22
Th. 23
Th. 26

[Ru], Appendix A4
[G.T.], Th. 7.22
[G.T.], Th. 7.27; [Ad], Th. 4.13

There is obviously no room here for a complete bibliography. For a more complete list of
references, the reader should refer to the bibliographies of the texts listed below.
Ad.

R. A. Adams, Sobolev Spaces, Academic Press, 1975.

Fr.

A. Friedman, Partial Differential Equations, Krieger, 1983.

G.T.

D. Gilbarg and N. S. Trudinger, Elliptic Partial Differential Equations of Second

Order, Springer-Verlag, 1983.

Ma.

V. G. Maz'ja, Sobolev Spaces, Springer-Verlag, 1985.

Ru.

W. Rudin, Functional Analysis, MacGraw-Hill, 1973.

Zi.

W. P. Ziemer, Weakly Differentiable Functions, Springer-Verlag, 1989.

36