Você está na página 1de 93

Chapter 1: Sobolev Spaces Introduction

In many problems of mathematical physics and variational calculus it is not sucient to deal with the classical solutions of dierential equations. It is necessary to introduce the notion of weak derivatives and to work in the so called Sobolev spaces. Let us consider the simplest example the Dirichlet problem for the Laplace equation in a bounded domain Rn : u = 0, x u(x) = (x), x , ()

where (x) is a given function on the boundary . It is known that the Laplace equation is the Euler equation for the functional
n

l(u) =
j=1

u xj

dx.

We can consider () as a variational problem: to nd the minimum of l(u) on the set of functions satisfying condition u| = . It is much easier to minimize this functional not in C 1 (), but in a larger class. 1 Namely, in the Sobolev class W2 (). 1 W2 () consists of all functions u L2 (), having the weak derivatives j u L2 (), j = 1, . . . , n. If the boundary is smooth, then the trace of u(x) on is well dened and relation u| = makes sense. (This follows from the so called boundary trace theorem for Sobolev spaces.) 1 If we consider l(u) on W2 (), it is easy to prove the existence and uniqueness of solution of our variational problem. 1 The function u W2 (), that gives minimum to l(u) under the condition u| = , is called the weak solution of the Dirichlet problem ().

Well study the Sobolev spaces, the extension theorems, the boundary trace theorems and the embedding theorems. Next, well apply this theory to elliptic boundary value problems.

1: Preliminaries
Let us recall some denitions and notation. Denition An open connected set Rn is called a domain. By we denote the closure of ; is the boundary. Denition We say that a domain Rn is a strictly interior subdomain of and write , if . If is bounded and , then dist { , } > 0. We use the following notation: u x = (x1 , x2 , . . . , xn ) Rn , j u = , xj = (1 , 2 , . . . , n ) Zn + || = 1 + 2 + . . . + n , Next, Denition u = (1 u, . . . , n u) , is a multiindex || u x1 x2 ...xn n 1 2
n j=1

u =

| u| =

|j u|2

1/2

Lq (), 1 q < , is the set of all measurable functions u(x) in such that the norm
1/q

u is nite.

q,

|u(x)|q dx

Lq () is a Banach space. Well use the following property: Let u Lq (), 1 q < . We denote
|z|

Here u(x) is extended by zero on Rn \. J (u; Lq ) is called the modulus of continuity of a function u in Lq (). Then J (u; Lq ) 0 as 0. 2

J (u; Lq ) = sup

|u(x + z) u(x)|q dx

1/q

Denition Lq,loc (), 1 q < , is the set of all measurable functions u(x) in such that |u(x)|p dx < for any bounded strictly interior subdomain . Lq,loc () is a topological space (but not a Banach space). We say that uk u in Lq,loc (), if uk u
k k q,

0 for any bounded

Denition L () is the set of all bounded measurable functions in ; the norm is dened by u , = ess sup |u(x)|
x

Denition C l () is the Banach space of all functions in such that u(x) and u(x) with || l are uniformly continuous in and the norm u
C l ()

=
||l

sup | u(x)|

is nite. If l = 0, we denote C 0 () = C(). Remark If is bounded, then u ty of u, u, || l < follows from the uniform continui-

C l ()

Denition C l () is the class of functions in such that u(x) and u, || l, are continuous in . Remark Even if is bounded, a function u C l () may be not bounded; it may grow near the boundary. Denition C0 () is the class of the functions u(x) in such that a) u(x) is innitely smooth, which means that u is uniformly continuous in , ; b) u(x) is compactly supported: supp u is a compact subset of . 3

2: Mollication of functions
1. Denition of mollication
The procedure of mollication allows us to approximate function u Lq () by smooth functions. Let (x), x Rn , be a function such that
C0 (Rn ), (x) 0, (x) = 0 if |x| 1, and

(x)dx = 1.

(1)

For example, we may take


1 c exp 1|x|2

(x) =

where constant c is chosen so that condition (1) is satised. For > 0 we put (x) = n
Then C0 (Rn ), (x) 0,

(x) = 0 if

Denition w is called a mollier. Let Rn be a domain, and let u Lq () with some 1 q . We extend u(x) by zero on Rn \ and consider the convolution u =: u

u (x) =

In fact, the integral is over {y : |x y| < }. Denition u (x) is called a mollication or regularization of u(x). 4

if if

|x| < 1 |x| 1

, x Rn .

(2)

|x| ,

(3) (4)

(x)dx = 1.

(x y)u(y)dy.

(5)

2. Properties of mollication
1) u C (Rn ), and u (x) = n x (x y)u(y)dy. This follows from C . 2) u (x) = 0 if dist {x; } , since (x y) = 0, y . 3) Let u Lq () with some q [1, ]. Then
q,

In other words, the operator Y : u u is a linear continuous operator from Lq () to Lq (Rn ) and Y Lq ()Lq ( n ) 1. Proof: Case 1 : 1 < q < . 1 o Let 1 + q = 1. By the Hlder inequality and (4), we have q |u (x)| = (x y)1/q (x y)1/q u(y)dy
1/q
n

=1

By (4), we obtain

|u (x)|q dx =

dx
n

Case 2 : q = . We have

|u (x)|

|u(y)|q dy

(x y)|u(y)|dy

(x y)dy
=1

dy |u(y)|q

|u (x)|q

(x y)|u(y)|q dy

(x y) |u(y)|q dy
n

(x y)

q, .

(6)

1/q
n

(x y)|u(y)|q dy

(x y)dx
=1

Case 3 : q = 1. We integrate the inequality |u (x)| and obtain:


n n n

(x y)|u(y)|dy

4) Let u Lq (), 1 q < . Then


q,

Consequently, u u
q,

Proof The proof is based on the following property: if u Lq () (and u(x) is extended by 0), then sup
|z|

|u(x + z) u(x)|q dx

(J (u; Lq ) is called the modulus of continuity of u in Lq .) Case 1 : 1 < q < . By (4) and (5) we have

|u (x) u(x)| =

(x y) (u(y) u(x)) dy (x y)1/q (x y)1/q (u(y) u(x)) dy

Then, by the Hlder inequality, it follows that o


1/q
n

=1

|u (x) u(x)|

(x y)dy

u u

0 as 0 as

0. 0.

1/q

=: J (u; Lq ) 0

(x y) |u(y) u(x)|q dy

|u (x)| dx

dx

(x y)|u(y)|dy =

|u(y)|dy

(7)

as

0.

1/q

Hence,

xy=z

|z|<

|z|

(J (u; Lq ))q .

q,

Case 2 : q = 1. We have |u (x) u(x)| (x y) |u(y) u(x)| dy

xy=z

|z|<

J (u; L1 ) 0 as 0.

Remark If q = , there is NO such property, since L limit of smooth functions u (x) must be a continuous function. If u C() and we extend u(x) by zero, then we may loose continuity. 0 as 0. In general, u u C() However, we have the following property: 5) If u C() , and is bounded, then u u
C( )

as

dz (z)

|u (x) u(x)| dx

u u

J (u; Lq ) 0 as

sup

|u(y + z) u(y)|q dy

dx

(x y) |u(y) u(x)| dy
n

|u (x) u(x)|q dx

dx
n

(x y) |u(y) u(x)|q dy
n

dz (z)

|u(y + z) u(y)|q dy dz
|z|< =1

(z)

0.

|u(y + z) u(y)| dy

Proof Let < dist { ; }. Then

u (x) u(x)

=
xy=z

(x y) (u(y) u(x)) dy (z) (u(x z) u(x)) dz 0

sup |u (x) u(x)|


x

x |z|

sup sup |u(x z) u(x)|

0 as

(since u(x) is continuous in ).

3: Class C0 ()
By C0 () we denote the class of innitely smooth functions in with compact support: u C0 ()

u C () and supp u .

Theorem 1
C0 () is dense in Lq (), 1 q <

Proof Let u Lq () and > 0. Let be a bounded domain, , and u We put u() (x) = Then u u()
() q, q,\

. 2 if x if x \
2

u(x) 0
()

property 4) of mollication, u u() Hence, u u Note that u


() q, C0 ()

2.

Let u (x) be the mollication of u() (x). By


() q,

for suciently small .

for suciently small . if < dist { , } .

Theorem 2 Let u L1,loc (), and suppose that u(x)(x)dx = 0,


C0 ().

(8)

Then u(x) = 0,

a. e. x .

Theorem 2 is an analog of the Main Lemma of variational calculus. Proof 1) First, let us prove that u(x)(x)dx = 0

for any L () with compact support supp . Suppose that supp , where is a bounded domain and . Then Let 0 = {x : dist {x; } < 0 }, and let 0 (x) = By (8), u(x) (x)dx = 0,
C0 () if < dist ; =: 20 .

1 0

if x 0 otherwise

< 0 .

(9)

Since L1 (), by property 4) of mollication, 1, 0 as 0. Then there exists a sequence {k }k , k 0, k < 0 , such that k (x) (x)
k k

for almost every x . for a. e. x . ), we have


,

Then also k (x)u(x) (x)u(x) Using property 3) (that

and the right-hand side in (10) belongs to L1 (). Then, by the Lebesgue Theorem,

|u(x)k (x)| 0 (x)|u(x)|

u(x)k (x)dx

By (9), the left-hand side is equal to zero. Hence, u(x)(x)dx = 0. 2) Now, let be a bounded domain such that . We put (x) = Then u(x)(x) =
u(x) |u(x)|

, if u(x) = 0, x , otherwise |u(x)| 0 ,x , x \

Since (x) is L function with compact support supp , then, by part 1), 0=

u(x)(x)dx =

It follows that u(x) = 0 for a. e. x . Since is an arbitrary bounded domain such that , then u(x) = 0, a.e. x

10

(10)

u(x)(x)dx.

|u(x)|dx.

4: Weak derivatives
1. Denition and properties of weak derivatives
Denition 1 Let be a multiindex. Suppose that u, v L1,loc (), and u(x) (x)dx = (1)||
v(x)(x)dx, C0 ().

(11)

Then v is called the weak ( or distributional ) partial derivative of u in , and is denoted by u. If u(x) is suciently smooth to have continuous derivative u, we can integrate by parts: u(x) (x)dx =

(1)|| u(x)(x)dx.

Hence, the classical derivative u is also the weak derivative. Of course, u may exist in the weak sense wihout existing in the classical sense. Remark 1) To dene the weak derivative u, we dont need the existence of derivatives of the smaller order (like in the classical denition). 2) The weak derivative is dened as an element of L1,loc (), so we can change it on some set of measure zero.

Properties of u
1) Uniqueness Proof Uniqueness of the weak derivative follows from Theorem 2. Suppose that u L1,loc () and v, w L1,loc () are both weak derivatives of u. Then, by (11), (v(x) w(x)) (x)dx = 0, v(x) = w(x),
C0 ().

By Theorem 2,

a.e. x .

11

2) Linearity If u1 , u2 L1,loc () and there exist weak derivatives v1 = u1 , v2 = u2 L1,loc (), then there exists (c1 u1 + c2 u2 ) and (c1 u1 + c2 u2 ) = c1 u1 + c2 u2 , Proof Obviously, (c1 u1 + c2 u2 ) dx = c1

c1 , c2 C.

u1 dx + c2

u2 dx v2 dx

= (1)|| c1

v1 dx + (1)|| c2 (c1 v1 + c2 v2 ) dx.

= (1)||

= (c1 u1 +c2 u2 )

3) If v = u in , then v = u in for any . Obvious 4) Mollication of the weak derivative Derivative of mollication is equal to mollication of derivative. This is true in any bounded strictly interior domain . Suppose that u, v L1,loc () and v = u. Then v (x) = u (x) if < dist {x, } . (12) The functions u and v are smooth; the derivative u in (12) is understood in the classical sense. Proof Let < dist {x, }. We have u (x) =

(x y)u(y)dy

Then u (x) = x (x y)u(y)dy. Note that x (x y) = (1)|| y (x y). Hence,

u (x) = (1)||

y (x y)u(y)dy.

Since < dist {x, }, then for (y) := (x y) we have C0 (). u = v, we obtain By denition of the weak derivative

u (x) =

(x y)v(y)dy = v (x).

12

5) Suppose that u L1,loc () and there exists the weak derivative u such that u Lq (), 1 q < . Then u u q, 0 as 0, for any bounded strictly interior domain .

Proof This follows from property 4) of mollication and property 4) of weak derivatives: u = v Lq (); u = v in v v
q,

( for suciently small ) ; 0.

0 as

Remark If we extend u(x) by zero on Rn \, then, in general, the weak derivative u in Rn does not exist. Hence, we have convergence 0 u u in Lq ( ) only for bounded strictly interior domain . Exclusion: if u(x) = 0, if dist {x; } < o , then u u
q,

0.

and u Lq (),

2. Another denition of the weak derivative


Denition 2 Suppose that u, v L1,loc () and there exists a sequence um C l (), m m m N, such that um u and um v in L1,loc (). Here is a multiindex and || = l. Then v is called the weak derivative of u in : u = v. Denition 1 Denition 2 Proof 1) Denition 1 Denition 2. Since um C l (), then um dx = (1)||

um dx,

C0 ().

(13)

13

For xed, the lefthand side of (13) tends to (um u) dx max | |

u dx as m :
m

supp

|um u| dx 0.

Similarly, the righthand side of (13) tends to (1)|| quently, u dx = (1)||


vdx. Conse-

vdx,

C0 ().

It means that v = u in the sense of Denition 1. 2) Denition 1 Denition 2. Let u, v L1,loc (), and let v = u in the sense of Denition 1. m We want to nd a sequence um C () such that um u and m um v in L1,loc (). Let {m } , m N, be a sequence of bounded domains such that

(m)

m , We put

m m+1 and

m = .
m

u(m) (x) =

u(x) 0

if x m otherwise

Then u(m) L1 (). Consider the mollication of u(m) : u C (). Let {m }m be a sequence of positive numbers such that m 0 as m . We put (m) um (x) = um (x), x . Then um C () and um u in L1,loc (). Prove this yourself, using property 4) of mollication. m Next, by property 5) of u, prove that um v in L1,loc (). Thus, v = u in the sense of Denition 2.

14

Theorem 3 Let um L1,loc () and um u in L1,loc (). Suppose that there m exist weak derivatives um L1,loc () and um v in L1,loc (). Then v = u. In other words, the operator is closed. Proof By Denition 1, for um we have um dx = (1)||
m

um dx, m vdx,

C0 ()

u dx = (1)||

C0 ()

v = u in the sense of Denition 1.

Remark The conclusion of Theorem 3 remains true under weaker assumptions that um dx
m

udx

and

um dx

vdx,

C0 ().

(It means that um u and um v in D ().)

3. Weak derivatives of the product of functions


Proposition If u, j u Lq,loc (), and v, j v Lq ,loc () with some 1 1 < q < , 1 + q = 1 or if u, j u L1,loc () and v, j v C(), then q j (uv) = (j u) v + u (j v) . Proof 1) Case 1: 1 < q < Let us x C0 (). Let be a bounded domain such that supp . We put u(x) = u(x) , x 0 otherwise 15 v (x) = v(x) , x 0 otherwise

Then u Lq ( ), v Lq ( ). By property 4) of mollications, u u


q,

0,

v v

q ,

0 as 0

Next, j u = j u in , j v = j v in (it is clear from Denition 1). So, j u Lq ( ), j v Lq ( ). By property 5) of weak derivatives, j v j v u v j dx = = Let us show that

j u j u

q,supp q ,supp

0,

0,

as 0,

as 0.

Since u , v are smooth functions, we have j ( v ) dx u


(j u ) v dx uv j dx =

u (j v ) dx. (14)

u v j dx +

uvj dx

(15)

We have

( v uv ) j dx u

( u) v j dx + u
q,

u ( v ) j dx v

u u
0

q ,

max |j | +

bounded

q,

v v
0

q ,

max |j |

0 Similarly, we can show that ((j u ) v + u (j v )) dx

as

0 ((j u) v + u (j v )) dx

((j u) v + u (j v)) dx (16)

From (14) - (16) it follows that

uvj dx =

((j u) v + u (j v)) dx

This identity is proved for any C0 (). It means (by Denition 1) that there exists the weak derivative j (uv) and

j (uv) = (j u) v + u (j v) 2) Case q=1. Prove yourself

16

4. Change of variables
Suppose that u L1,loc () and there exist weak derivatives j u L1,loc (), j = 1, . . . , n. Let y = f (x) be a dieomorphism of class C 1 and f () = . 1 (y)). Then u L We put u(y) = u(f 1,loc (). Let us show that there exist u weak derivatives y , = 1, . . . , n, and u = yk Proof Since there exist weak derivatives C 1 ()
n j=1

u xj xj y

u xj

u exists a sequence um such that um u and um xj in xj L1,loc () for all j = 1, . . . , n. (We can construct this sequence like in the proof, that Denition 1 and Denition 2 are equivalent ). We denote um (y) = um (f 1 (y)). Then um C 1 (), and, by usual rule (for classical derivatives), n um xj um = yk xj y j=1 m Let us check that um u in L1,loc (). Indeed, for every bounded domain we have

L1,loc (), j = 1, . . . , n, then there


m m

|m (y) u(y)| dy u

= =

um (f 1 (y)) u(f 1 (y)) dy |um (x) u(x)| |J(x)| dx as m

Here = f 1 ( ) and J(x) is the Jacobian of the transformation f (x) y J(x) = det x . Here the right-hand side tends to zero, since |J(x)| is bounded in ; m is a bounded domain such that ; and um u in L1,loc (). m u Similarly, using that um xj in L1,loc (), one can show that xj um yk
n

=
j=1

um xj xj y

n j=1

u xj xj y

in L1,loc ()

Then, by Denition 2, there exist weak derivatives u yk and u yk 17


n

=
j=1

u xj xj yk

Thus, for weak derivatives we have the usual rule of change of variables. The same is true for derivatives of higher order.

5.
For ordinary derivatives we have the following property: u if xj = 0 in , j = 1, . . . , n, then u = const. The same is true for weak derivatives. Theorem 4 Suppose that u L1,loc () and there exist weak derivatives u for any multi-index such that || = l (l N) and u = 0 in , || = l. Then u(x) is a polynomial of order l 1 in . Proof 1) Let be a bounded domain such that . Let be another bounded domain such that . We put u(x) = u(x) , x 0 otherwise

Then u L1 ( ) , and u = u = 0, || = l, in . Consider the mollication u (x). If < dist { , }, then, by property 4) of u, u (x) = ( u) (x), x , || = l. Hence, u = 0 in . Thus, u (x) is a smooth function in and all its derivatives of order l are equal to zero. It follows that () () u (x) = Pl1 (x), x , where Pl1 is a polynomial of order l 1. By property 4) of mollication, u u
1,

as

0, i. e. ,

Pl1 u

() 0

in

L1 ( ).

The set of all polynomials in of order l 1 is a nite-dimensional (and, so, closed!) subspace in L1 ( ). Therefore, the limit u(x) must be also a polynomial of order l 1: u(x) = Pl1 (x), x.

2) Now it is easy to complete the proof by the standard procedure. Let {k }k be a sequence of bounded domains such that

k ,

k k+1 , 18

and
k

k = .

We have poved that for each domain k u(x) = Pl1 (x),


(k+1) (k) (k)

x k

Then Pl1 (x) is continuation of Pl1 (x), but continuation of a polynomial is unique. There exists a polynomial Pl1 (x) such that u(x) = Pl1 (x), x .

6. Absolute continuity property


The existence of the weak derivative is related to the absolute continuity property. Recall the denition of absolute continuity for function of one variable. Denition Function u : [a, b] R is called absolutely continuous, if for any > 0 there exists > 0 such that for any nite set of disjoint intervals x1 , x1 , with
m j=1

x2 , x2 , . . . , xm , xm one has u(xj ) u(xj ) < .

( [ a, b ])

xj xj < ,
m j=1

Well use the following facts: 1) u : [a, b] R is absolutely continuous if and only if there exists a function v L1 (a, b) such that
x

u(x) = u(a) +
a

v(t)dt,

x [a, b] .

2) If u : [a, b] R is absolutely continuous, then there exists derivative du du dx for almost every x (a, b) and dx = v ( L1 (a, b)). Theorem 5 Let n = 1. A measurable function u(x) is absolutely continuous on [a, b] if and only if there exists the weak derivative du L1 (a, b). The weak dx derivative coincides with the classical derivative almost everywhere. 19

Remark When we speak about measurable functions, we mean not just one function but a class of functions, that are equal to each other almost everywhere. So, when we say that a measurable function u(x) is absolutely continuous, it means that in the class of functions equivalent to u, there exists an absolutely continuous representative. Proof 1) u(x) is absolutely continuous. weak derivative du L1 . If u(x) is a. dx c., then there exists the classical derivative du = v almost everywhere dx and v L1 (a, b). Next, let C0 (a, b). Then the product u is also absolutely continuous. There exists the classical derivative d( u) for dx almost every x (a, b). We have the usual rule: d d ( u) = v + u . dx dx Integrate this identity over (a, b). Then (Since (x) = 0 near a and b). Hence,
b b d( u) a dx dx

= 0.

v + u
a

d dx

dx = 0.

The obtained identity


b

u
a

d = dx

vdx,
a

C0 (a, b), du dx

by Denition 1, means that v is the weak derivative

2) weak derivative v = du L1 (a, b) u is a. c. dx x Consider w(x) = a v(t)dt. Then w(x) is absolutely continuous. There exists classical derivative dw dx = v, a.e. x (a, b). By statement 1) (already proved), there exists the weak derivative dw dx which coincides with the classical one and with v. Thus, du dw d (u w) = , i. e. = 0. dx dx dx (the weak derivative is equal to zero.) By Theorem 4, u w = const. Since w(x) is absolutely consinuous, then u = c + w is also absolutely continuous. If x = a, we have u(a) = c + w(a) = c. Thus,
=0 x

u(x) = u(a) +
a

v(t)dt.

20

u(x) is absolutely continuous; it has classical derivative for a. e. x (a, b); classical derivative = weak derivative = v L1 (a, b).

Theorem 6 Let Rn , n > 1. We denote x = (x1 , . . . , xj1 , xj+1 , . . . , xn ) and write x = {x , xj }. Suppose that [a(x ), b(x )] are some intervals such that {x } [a(x ), b(x )] . u Let u L1,loc () and there exists the weak derivative xj L1,loc (). Then for almost every x the function u(x , xj ) is absolutely continuous on interval [a(x ), b(x )] (as a function of one variable xj ). Exercise: Prove Theorem 6.

7. Examples
1) Let = (0, 1)2 and u(x1 , x2 ) = (x1 ) + (x2 ), where and are not absolutely continuous on [0, 1], but , L1 (0, 1). u u Then, by Theorem 6, u(x1 , x2 ) does not have weak derivatives x1 , x2 in . (Since, if they exist, then u(x) must be also absolutely continuous in x1 for x2 xed, and in x2 for x1 xed.) 2u However, there exists the weak derivative x1 x2 = 0. Indeed, for C0 (), u

2 dx = x1 x2 =

1 2 dx1 dx2 + x1 x2 0 0 0 1 1 2 dx2 + dx1 (x1 ) 0 0 x1 x2

(x1 )

2 dx1 dx2 x1 x2 0 1 1 2 dx1 dx2 (x2 ) 0 0 x1 x2 (x2 )


=0

=0

= 0. By Denition 1 of weak derivative, it means that there exists weak 2u 2u derivative x1 x2 and x1 x2 =0. This example shows that functions may have derivative of higher order, not having derivatives of lower order. 2) Suppose that the domain Rn is devided by a smooth (n-1)dimensional surface into two parts 1 and 2 . So, = 1

21

2 . Let u1 C 1 (1 ), u2 C 1 (2 ), u(x) =

If u1 | = u2 | , then, in general, weak derivatives do not exist. Let n(x) be the unit normal vector to exterior with respect to 1 . Since uk (x), k = 1, 2, is a C 1 function in k , we can integrate by parts in k : for C0 () we have: u

u1 (x) , x 1 u2 (x) , x 2

dx = xj

u1
1

= +

dx + dx u2 xj xj 2 u2 u1 dx dx + xj 2 xj

(u1 (x) u2 (x)) cos ((n, 0xj )) dS(x)

If u1 = u2 on (we have NO jump on ), then the integral over is u equal to zero. In this case, there exists the weak derivative xj and u = xj
u1 xj u2 xj

in 1 in 2

u Also, if cos ((n, 0xj )) = 0, then there exists xj . For example, if is parallel to the axis 0xj , then cos ((n, 0xj )) = 0. Even if u1 | = u2 | , the tangential derivative exists. u If (u1 (x) u2 (x)) cos ((n, 0xj )) dS(x) = 0, then xj does not exist.

Exercise Let = {x Rn : |x| < 1}, and let u(x) = |x| , > n + 1. Prove that there exist the weak derivatives u u and = xj |x|2 , j = 1, . . . , n. xj xj For this consider continuous functions |x| , |x| > u() (x) = , |x| u() = xj xj |x|2 , |x| > 0 , |x| vj := xj |x|2 .

From the previous example we know that


L1 ()

Check that u() u and

u Then, by Theorem 3, it follows that xj = vj

u() L1 () xj

22

l l 5: The Sobolev spaces Wp () and Wp ()


l 1. Denition of Wp ()

(1 p < , l Z+ )

Denition Suppose that u Lp () and there exist weak derivatives u for any with || l (all derivatives up to order l), such that u Lp (), || l.
l Then we say that u Wp (). l We introduce the (standard) norm in Wp ():

l Wp ()

||l

| u|p dx

1/p

Remark 1) The norm


||l

p,

is equivalent to the standard norm.

0 2) Wp () = Lp ().

Proposition
l Wp () is complete. l In other words, Wp () is a Banach space.

Proof l Let {um } be a fundamental sequence in Wp (). It is equivalent to the fact u } for || l are fundamental sequences in L (). that all sequences { m p Since the space Lp () is complete , there exist functions u, v Lp () such that Lp () Lp () um u, um v as m . Then also um u, um v in L1,loc (). By Theorem 3, v = u. Hence, um u
l Wp ()

as m .

l If p = 2, the space W2 () is a Hilbert space with the inner product

(u, v)W l () =
2

u(x) v(x)dx.
||l

23

l l For W2 () another notation H l () is often used: W2 () = H l ().

Using the properties of weak derivatives (see section 4 Change of varia l bles in 4), we can show that the class Wp () is invariant with respect to smooth (C l -class) change of variables. Theorem 7 Let f : be a dieomorphism of class C l , so that l (), f 1 C l (). f C l () , then u = u f 1 W l (), and Then, if u Wp p c1 u
l Wp ()

l Wp ()

c2 u

l Wp ()

(17)
C l ()

The constants c1 , c2 do not depend on u; they depend only on f and f 1 C l () . Proof: For simplicity, let us prove Theorem 7 in the case l = 1. We have By section 4 in 4, there exist the weak derivatives u = yk Let us check that u yk
p u yk 1/p n j=1 1 u Wp (),

u(y) = u(f 1 (y)). u xj . xj yk

dy

Lp (): =
n

j=1

u xj xj yk

max
j=1 n j=1 x

xj |J(x)|1/p yk u xj
p 1/p

|J(x)| dx

1/p u xj
p 1/p

dx

dx

.
xj yk

Here J(x) = det f (x) and the constant c := maxj,k maxx depends only on the norms f Also we have
C 1 ()

and f 1

C 1 ()

|J(x)|1/p

|(y)|p dy = u u
1 Wp () f 1

|u(x)|p |J(x)| dx (max |J(x)|)


1 Wp ()

|u(x)|p dx.

Thus,

f C 1 and . Prove the lower estimate in (17) yourself (for this C1 change the roles of u and u in the argument).

c2 u

with the constant c2 depending only on

24

l 2. Denition of Wp ()

Denition
l l The closure of C0 () in the norm of Wp () is denoted by Wp ().

l l So, Wp () is a subspace in the space Wp ().

Proposition
l Let u Wp (), and let

u(x) =

u(x) 0

x x Rn \.

l Then u Wp (1 ) for any 1 such that 1 . In particular, l u Wp (Rn ).

Proof
l Wp ()

l By denition of Wp (), there exists a sequence um C0 () such that

um u as m . We put um (x) = um (x) 0 x otherwise .

Then um C0 (1 ) and um u as m (since um u Wp (1 ) = um u Wp () ). l l l Hence, u Wp (1 ).

l Wp (1 )

Theorem 8
l Let u Wp () and let

u(x) =

u(x) 0

x x Rn \.
0

l Then for mollications u (x) we have u u in Wp ().

25

Proof l We have already proved that u Wp (Rn ). Then u Lp (Rn ), || l. By u (mollication of the weak derivative), property 4) and 5) of u u in Lp (), It means that
0 0

|| l.
0

l of mollication, u = u in , and u = u . So, u u in Wp ().

l u u in Wp (). But, by denition of u and denition

Remark l If u(x) is an arbitrary function in Wp (), and u(x) is dened as above, then, in general, u(x) does not have weak derivatives in Rn . (See example 2
l l in Section 7 of 4). So, in general, Wp () = Wp ().

3. Integration by parts
Proposition
l l Let u Wp () and v Wp (), where 1 p

1 p

= 1. Then || l. (18)

uvdx = (1)||

u vdx,

Proof

l Let vm C0 () and vm v as m in Wp (). By Denition 1 of the u, we have weak derivative

uvm dx = (1)||

u vm dx.

(19)

Let us show that u vm dx


m

u v dx,

u vm dx

u v dx.

We have u (vm v) dx u | u|p dx


1/p
l Wp

|vm v|p dx

1/p

l Wp ()

vm v

()

0 as m ; 26

u ( vm v) dx

|u|p dx

1/p

| vm v|p dx

1/p

l Wp ()

vm v

l Wp ()

0 as m . Tending to the limit in (19) as m , we obtain (18).

4. Separability
By Vpl () we denote the linear space of all vectorvalued functions v = {v }||l such that v Lp (), || l. We introduce the norm in Vpl (): v
l Vp ()

=
||l

p, .

Then Vpl () is the direct product of a nite number (equal to the number of multiindices with || l) of Lp (). We know that Lp () is a separable Banach space if 1 p < . Then so is Vpl (). l Now, consider the transformation J from Wp () (equipped with the norm ||u||Wp () = ||l u p, , which is equivalent to the standard norm) to l l (): Vp l J : Wp () Vpl (), Ju = { u}||l . Then J is a linear operator; it preserves the norm: Ju Vpl () = ||u||Wp () ; l and J is injective. Such an operator is called isometric. The range Ran J = Vpl () is a linear set in Vpl () consisting of l vectorvalued functions v of the form v = { u}||l , u Wp (). From Theorem 3 it follows that Vpl () is a closed subspace of Vpl (). Hence, l () is separable together with V l (). (Since any subspace of some sepa Vp p rable space is also separable.) l l Since J is isometric, we can identify Wp () with Vpl (). It follows that Wp () is separable if 1 p < .
l 5. The space Wp (Rn )

Proposition
l l l Wp (Rn ) = Wp (Rn ) . In other words, C0 (Rn ) is dense in Wp (Rn ).

27

Proof Let C (R+ ) be such that 0 (t) 1, (t) = 1 if 0 t 1,


|x| R

(t) = 0 if t 2. . Then

l Let u Wp (Rn ). We put u(R) (x) = u(x)

u(R) (x) = u(x) if |x| R,


Note that derivatives x |x| R

u(R) (x) = 0 if |x| 2R.

are uniformly bounded with respect to

R 1. Calculating the derivatives of u(R) (x), we obtain the inequality u(R) (x) c Then for || l we have u(R) u u(x) ,
||||

a.e. x Rn .

p,

=
|x|>R

c
|||| |x|>R

0 as R
l This expression tends to zero as R , since u Wp (Rn ), and so, u p L . Thus, (R) u as R in W l (Rn ). u 1 p (R)

Now, we consider mollication u of u(R) . (R) (R) l Then u C0 (Rn ) and u u(R) as 0 in Wp (Rn ). l l It follows that C0 (Rn ) is dense in Wp (Rn ). Indeed, let u Wp (Rn ) and let > 0. We nd R so large that u(R) u W l ( n ) < 2 . Next, we nd so
p

l Wp (

n)

l Wp (

28

small that u

(R)

u(R)

< 2 . Then u

(R)

p u(R) (x) u(x) dx =0 for |x|R

1/p
1/p

(R)

(x) u(x) dx u(x) dx


p 1/p

n)

< .

6. The Friedrichs inequality


Theorem 9
l If is a bounded domain in Rn , then for any function u Wp () we have

u Here

p,

(diam)l up,l, . 1/p

(20)

up,l, = Proof

||=l

p p,

(21)

l Since C0 () is dense in Wp (), it suces to prove (20) for u C0 (). 1) So, let u C0 (). Let Q be a cube with the edge d = diam, such that Q. We extend u(x) by zero to Q\. We can choose the coordinate system so that Q = {x : 0 < xj < d, j = 1, . . . n}. Obviously, xn u u(x) = (x , y)dy, x Q. xn 0

Here x = (x1 , . . . , xn1 , xn ) = (x , xn ). Then, by the Hlder inequality, o |u(x)|p


xn 0 p u (x , y) dy xn xn p/p x

1dy
0 dp/p

dp/p

d 0

u(x , xn ) xn

dxn

1 Here 1 + p = 1. p We integrate both sides of this inequality:

|u|p dx =

|u|p dx
p p

+1=p

dp/p
0

dxn
Q p

u xn

dx

u xn

dx.

We have proved that u


p,

(diam)

u xn

1/p

dx

(diam)up,1, .

(22)

This is inequality (20) for l = 1. 29

2) In order to prove (20) with l > 1, we iterate (22): u xn


p p

dx d
lp

2u x2 n
p

dx,

etc.

|u| dx d

lu xl n

dx dlp up . p,l,

Remark
l Inequality (20) is not valid for all u Wp ().

Example If is a bounded domain and u(x) = Pl1 (x)(= 0) is a polynomial of order l 1, then up,l, = 0, but u p, = 0.

30

6. Domains of star type


A natural question: l Can we approximate functions in Wp () by smooth functions? The answer depends on domain . Well consider the class of domains for which the answer is YES, we can. Denition We say that a bounded domain is of star type with respect to a point 0, if any halfline starting at point 0 intersects only in one point. Theorem 10 Let be a bounded domain of star type with respect to a point 0. l Then C () is dense in Wp (). Proof Let us use the coordinate system with origin 0. Consider a sequence of domains k = x : k1 x , k N. k Then k+1 k and k . l Let u Wp (). We put uk (x) = u( k1 x). k l Clearly, uk Wp (k ). Let us show that uk u Wp () 0 as k . l We have uk u
p,

p k1 x u(x) dx k

1/p

0 as k .

This follows from the property of Lp functions: if u Lp (), then


c |z(x)| k

sup

|u(x + z(x)) u(x)|p dx 0.

d (In our case z(x) = x and |x| diam = d |z(x)| k .) k Let be a multiindex with || l. Then

uk u

p,

k1 k k1 k

||

u
||

k1 x u(x) dx k u

1/p

k1 x k
c u
l Wp ()

1/p

dx

0 as k

p k1 x u(x) dx k 0 as k

1/p

31

l Hence, uk u in Wp (). Consider mollications uk, (x). Then 0

l uk, C () and uk, uk in Wp () (since is bounded and k ). We can choose a sequence {k }, so that k 0 as k , and a sequence l uk (x) := uk,k (x) tends to u(x) in Wp (): () and uk C

uk u

l Wp ()

0.

Remark Let = {x : |x| < 1, xn > 0} be a halfball. is of star type with l respect to any interior point 0 . Suppose that u Wp () and u(x) = 0 if |x| > 1 . Then uk C () and uk (x) = 0 if |x| > 1 2 for suciently large k.

32

7: Extension theorems
l We can always extend a function u Wp () by zero and the extended l () in ( ). It is a natural question if we can extend function Wp l (). We start with the case l = 1. functions of class Wp

Theorem 11 Suppose that Rn is a bounded domain such that is a compact manifold of class C 1 . Let be a domain in Rn such that . Then there exists a linear bounded extension operator
1 1 : Wp () Wp () such that (u) (x) = u(x),

x .

Proof We proceed in three steps: Step 1 1 Let = K+ = {x : |x| < 1, xn > 0} be a halfball, and let u Wp (K+ ) and u(x) = 0 near + = {x K+ : |x| = 1}. We extend u to the left halfball K = {x : |x| < 1, xn < 0} as follows: v(x) = u(x) x K+ u(x , xn ) x K .

1 Let us show that v Wp (K) and

1 Wp (K)

= 21/p u

1 Wp (K+ ) .

(23)

Here K = {x : |x| < 1}. Using construction of Theorem 10 (and Remark after Theorem 10), we can nd a sequence um (x) such that um C (K+ ), um (x) = 0 near + , and um u Wp (K+ ) 0 as m . We put 1 vm (x) = um (x) x K+ um (x , xn ) x K .

1 follows that vm Wp (K) (see 4, Subsection 7, Example 2). For the norm of vm we have:

Then vm C(K), vm (x) = 0 near K, vm C (K+ ), vm C (K ). It

vm

p 1 Wp (K)

=
K

(|vm (x)|p + | vm (x)|p ) dx


K+

= 2

(|um (x)|p + | um (x)|p ) dx = 21/p um 33


1 Wp (K+ ) .

vm

1 Wp (K)

(24)

Next, vm (x) = xj
m xj

x K+ (um (x , xn )) x K .
m

um (x) xj

1 Since um u in Wp (K+ ), it follows that vm v in Lp (K) and m vm xj wj in Lp (K), where

wj (x) =

u xj (x

u(x) xj

x K+ , , xn ) x K
u(x) xn u xn (x

j = 1, . . . , n 1; x K+

wn (x) =

, xn ) x K .
v xj

By Theorem 3, there exist weak derivatives vm v in (as m ).


m 1 Wp (K).

in K and

v xj

= wj . Thus,

Relation (23) follows from (24) by the limit procedure

Step 2 1 Suppose that u Wp () and supp u U , where U is a neighbourhood of 0 , such that U and dieomorphism x 1 (U ), f : U K, f C f 1 C 1 (K), f (U ) = K, f (U ) = K+ , f (U ) = K+ \+ . 1 We consider the function u(y) = u(f 1 (y)), y K+ . Then u Wp (K+ ) and u(y) = 0 near + . We extend u(y) on K like in step 1: v (y) = u(y) y K+ u(y , yn ) y K .

1 As it was proved in step 1, v Wp (K), and

1 Wp (K)

= 21/p u

1 Wp (K+ ) .

1 v(x) by zero on \U . Then v Wp (), v| = u, and

1 Consider the function v(x) = v (f (x)), x U . Then v Wp (U ). We extend

1 Wp ()

= v

1 Wp (U )

c1 v

1 Wp (K)

c1 21/p u

1 Wp (K+ )

c2 c1 21/p u
=c

1 Wp () .

The constant c depends on f C 1 , f 1 C 1 and on p. Step 3 (general case) Let be a bounded domain such that is a compact manifold of class 34

N j C0 (Rn ), supp j Uj , j=1 j (x) = 1, x . N 1 (). We represent u(x) as u(x) = Let u Wp j=1 uj (x), where uj (x) = j (x)u(x).

C 1 with boundary . Then (by denition of such manifolds) there exists a nite number of open sets U1 , U2 , . . . , UN such that either Uj or Uj is a neighbourhood of some point x(j) , and a dieomorphism 1 fj C 1 (Uj ), fj C 1 (K), fj (Uj ) = K, fj (Uj ) = K+ , fj (Uj ) = K+ \+ . Finally, N Uj . j=1 We can choose the sets U1 , U2 , . . . , UN so that N Uj . j=1 There exists a partition of unity {j (x)}j=1,...,N such that

1 If Uj , then uj Wp () (since supp j Uj ) and we can extend uj (x) by zero to \ : uj (x) x vj (x) = 0 x \. 1 of step 2, we can extend uj (x) to some function vj Wp () such that vj (x) = uj (x), x , and vj W 1 () cj uj W 1 () .
p p

If U = , then uj (x) satises assumptions of step 2. By the result

The constant cj depends on fj

C1

1 fj N

C1

and on p. We put

u = v =
j=1 1 Then v Wp (), N N j=1 vj (x) N

vj .

v(x) = vj
j=1

N j=1 uj (x)

= u(x),
N

x , and
1 Wp () ,

1 Wp ()

1 Wp ()

cj uj
j=1

1 Wp ()

uj
j=1

where c = max1jN {cj }. Finally, uj


1 Wp ()

= j u
C1 .

1 Wp ()

cj u

1 Wp () .

The constant cj depends on j v

Hence,
N
1 Wp () ,

1 Wp ()

c u c

c=
j=1

cj .

Thus, we constructed the linear continuous extension operator 1 1 : Wp () Wp ().

35

Remark It is clear from the proof that for the constructed extension operator we have v Lp () c u Lp () , v = u. The constant c depends on p, and .

2.
Similar extension theorem is true for unbounded domain Rn satisfying the following condition. Suppose that there exist bounded open sets {Uj } , j N, such that Uj . Here either Uj or Uj is a neighj=1 (j) and a dieomorphism f C 1 (U ), bourhood of a point x j j 1 fj C 1 (K), fj (Uj ) = K, fj ( Uj ) = K+ , fj ( Uj ) = K+ \+ . Moreover, suppose that the norms fj
C 1 (Uj ) 1 and fj C 1 (K)

are uniformly

bounded for all j N. Suppose also that each point x belongs only to a nite number N (x) of sets Uj , and that N (x) N < , x . (This means that the multiplicity of covering is nite.) Theorem 12 Under the above conditions on Rn , let be a domain in Rn such that j=1 Uj . Then there exists a linear bounded extension operator
1 1 : Wp () Wp ()

such that (u) (x) = u(x), We omit the proof.

x .

3. Now we consider the case l > 1


Theorem 13 Suppose that Rn is a bounded domain such that is a compact manifold of class C l . Let be a domain in Rn such that . Then there exists a linear bounded extension operator
l l : Wp () Wp (), i. e., (u) (x) = u(x), for x

and Besides,

The constants c1 , c2 depend on l, p, and . 36

c1 u Wp () . l u Lp () c2 u Lp () .
l Wp ()

Proof Like in the proof of Theorem 11, the question reduces to the case, where = K+ and u(x) = 0 near + . Moreover, it suces to consider smooth functions u C (K+ ). So, let u C (K+ ) and u(x) = 0 near + . We extend u(x) by zero to Rn \K+ . We put + v(x) =
l1 j=0 cj

u(x) x K+ jx ) x K . u(x , 2 n

The constants cj , j = 0, . . . , l 1, are chosen so that mv mv (x , +0) = (x , 0), xm xm n n m = 0, . . . , l 1.

These conditions are equivalent to the following system of linear equations for c0 , c1 , . . . , cl1 :
l1 j=0

2j

cj = 1,

m = 0, 1, . . . , l 1.

The determinant of this system is not zero. l Hence, such constants c0 , . . . , cl1 exist. It is easy to check that v Wp (K), and v p,K c u p,K+ , || l. (We use that v C (K+ ), v C (K ) and v C l1 (K). l Then v Wp (K).) Hence , v Wp (K) c1 u Wp (K+ ) . l l
l Obviously, v(x) = 0 near K. So, v Wp (K). Next, for arbitrary domain , we use the covering N Uj and the j=1 partition of unity. The argument is the same as in proof of Theorem 11. The only dierence is that we consider dieomorphisms of class C l .

Remark 1) The conclusion of Theorem 13 remains true under weaker assumptions on the domain . It suces to assume that is domain of class C 1 (for arbitrary l!) or, even that is Lipschitz domain (it means that 1 dieomorphisms fj , fj Lip1 ).
l 2) Extension theorems allow us to reduce the study of functions in Wp () l to the study of functions in Wp (). In particular, from the fact that l l C0 () is dense in Wp () it follows that C () is dense in Wp (), if domain satises conditions of Theorem 13.

37

Chapter 2: Embedding Theorems Introduction


Embedding theorems give relations between dierent functional spaces. Denition Let B1 and B2 be two Banach spaces. We say that B1 is embedded into B2 and write B1 B2 , if for any u B1 we have u B2 and u B2 c u B1 , where the constant c does not depend on u B1 . We dene the embedding operator J : B1 B2 , which takes u B1 into the same element u considered as an element of B2 . The fact that B1 B2 is equivalent to the fact that the embedding operator J : B1 B2 is continuous linear operator. If u B2 c u B1 , u B1 , then J B1 B2 c. Denition If B1 B2 and the embedding operator J : B1 B2 is a compact operator, then we say that B1 is compactly embedded into B2 . The compactness of operator J is equivalent to the fact that any bounded set in B1 is a compact set in B2 . Some embeddings are obvious. l l For example, it is obvious that Wp1 () Wp2 (), if l1 > l2 . In particul () L (), l > 0. But the fact that for bounded domain , these lar, Wp p embeddings are compact, is nontrivial. (This is the Rellich embedding theorem.) l r More general is the Sobolev embedding theorem : Wp () Wq () under some conditions on p, l, q, r (with q > p and r < l). l Another embedding theorem is that, if pl > n, then a function u Wp () is continuous (precisely, u(x) coincides with a continuous function for a. e. x ). l The trace embedding theorems show that functions in Wp () have traces on some surfaces of lower dimension. The embedding theorems are very important for the modern analysis and boundary value problems.

38

1: Integral operators in Lp ()
In order to prove embedding theorems, we need some auxiliary material about integral operators.

1.
Let Rn and D Rm be some bounded domains. We consider the integral operator (Ku) (x) = v(x) =

K(x, y)u(y)dy,

x D,

u Lp ()

(1 p < ).

Well show that under some conditions on the kernel K(x, y), the operator K is continuous or, even, compact from Lp () to Lq (D), or from Lp () to C(D). We always assume that K(x, y) is a measurable function on D , and K satises one or several of the following conditions: a)

|K(x, y)|t dy M |K(x, y)|s dx N sup


xD,y

for a. e. x D, where t 1.

b)
D

for a. e. y , where s > 0. (K is bounded).

c) ess d)
x,zD

|K(x, y)| L <

sup sup |K(x, y) K(z, y)| () 0 as 0


y

|xy|

(K is continuous in x). Lemma 1 If K(x, y) satises conditions c) and d) , then K : Lp () C(D) is compact operator. Here 1 p < . Proof Let u Lp () and v(x) = (Ku) (x). Then, by condition c), |v(x)| L |u(y)| dy L |u(y)| dy 39
p 1/p 1/p

1 dy

= L||1/p u

p, ,

(1)

1 1 where p + p = 1. (If p = 1 then (1) is also true with p = , ||1/p = 1.) Next, if |x z| (x, z D), then, by condition d),

|v(x) v(z)| =

(K(x, y) K(z, y)) u(y)dy


1/p

()

|u(y)| dy u
p, .

()||

(2)

From (1) and (2) it follows that, if u belongs to some bounded set in Lp (): u p, c, then the set of functions {v} is uniformly bounded ( v C(D) L||1/p c) and equicontinuous (|v(x) v(z)| ()||1/p c, if |x z| ). By the Arzela Theorem, this set is compact in C(D). It means that the operator K : Lp () C(D) is compact.

Lemma 2
1 1) If p > 1, 1 + p = 1, and K(x, y) satises conditions a) and b) with p s t some t < p and p + p 1, then v = Ku Lq (D) (for u Lp ()), s t where q p is dened from the relation q + p = 1. We have

q,D

M 1/p N 1/q u

p, ,

u Lp ().

(3)

2) If p = 1, and K(x, y) satises condition b) with s = q 1. Then v = Ku Lq (D) and v


q,D

N 1/q u

1, ,

u L1 ().

(4)

3) If p > 1, and K(x, y) satises condition a) with t = p , then v = Ku L (D) and v


,D

M 1/p u

p, ,

u Lp ().

(5)

4) If p = 1, and K(x, y) satises condition c), then v = Ku L (D) and v


,D

L u

1, ,

u L1 ().

(6)

40

Proof 1) Let p > 1. Using that


s q

t p

= 1, we obtain:
p

|K(x, y)u(y)| = |K(x, y)|s/q |u(y)|p/q |u(y)|1 q |K(x, y)|t/p We apply the Hlder inequality for the product of three functions: o |f1 (y)f2 (y)f3 (y)| dy
1 p1
1 p1 1 p2 1 p3

|f1 | dy

p1

|f2 | dy

p2

|f3 | dy

p3

with

1 p2

1 p3

= 1.
pq qp ,

We take p1 = q, p2 = |v(x)|

p3 = p . Then

|K(x, y)u(y)| dy

|K(x, y)| |u(y)| dy


1 p q p,

1 q

|u(y)| dy
1 q

1 1 q p

|K(x, y)| dy

1 p

M 1/p ( by cond. a) )

1 p

|K(x, y)| |u(y)| dy

Note that in the case q = p, we simply apply the ordinary Hlder o inequality and obtain the same result. We have |v(x)|q M q/p u
D qp p, qp p, qp p,

|K(x, y)|s |u(y)|p dy dx

|v(x)|q dx M q/p u = M q/p u

|K(x, y)|s |u(y)|p dy


D

|u(y)|p dy

|K(x, y)|s dx

N ( by cond b ))

NM This gives estimate (3).

q/p

q p, .

2) Let p = 1 and s = q 1. If q > 1, we have |K(x, y)u(y)| = (|K(x, y)|q |u(y)|)1/q |u(y)|1/q , Then, by the Hlder inequality, o |v(x)|

1 1 + = 1. q q

|K(x, y)u(y)| dy

|K(x, y)|q |u(y)| dy 41

1/q

1/q

|u(y)| dy

|v(x)|q dx =

dx
D

|K(x, y)q | |u(y)| dy


D

q/q 1,

|u(y)| dy
q 1,

|K(x, y)|q dx

q/q 1,

N ( by cond b) with s=q)

N u

This implies (4) (in the case q > 1). If q = 1, then |v(x)|
D

|K(x, y)| |u(y)| dy dx

|v(x)| dx =

|K(x, y)| |u(y)| dy


D

|u(y)| dy
1, .

|K(x, y)| dx
N

N u This implies (4) (in the case q = 1).

3) Let p > 1, and condition a) is satised with t = p . Then, by the Hlder inequality, o |v(x)| |K(x, y)u(y)| dy

|K(x, y)|p dy
M 1/p p, .

1/p

|u(y)|p

1/p

M 1/p u This yields (5).

4) Let p = 1 and K satises condition c). Then |v(x)| which gives (6). |K(x, y)| |u(y)| dy L |u(y)| dy = L u
1, ,

42

Remark Statements 1) and 2) mean that the operator K : Lp () Lq (D) is continuous, and K K
Lp ()Lq (D) Lp ()Lq (D)

N 1/q ,

M 1/p N 1/q ,

p > 1;

(7) (8)

p = 1.

Under conditions of 3) and 4) the operator K : Lp () L (D) is continuous and K


Lp ()L (D) Lp ()L (D)

L,

M 1/p ,

p > 1;

(9) (10)

p = 1.

2.
Now, well show that under some additional assumptions on K(x, y), the operator K is compact. Well assume that K(x, y) can be approximated by Kh (x, y) (as h 0) and Kh (x, y) are bounded and continuous in x. Suppose that Kh (x, y), 0 < h < h0 , satises conditions c) and d) (where L = L(h) and () = (; h) depend on h). 1) Suppose that K(x, y) and Kh (x, y), 0 < h < h0 , satisfy conditions of Lemma 2(1) with common t, s, M, N , and |Kh (x, y) K(x, y)|t dy mh 0, for a. e. x D; (11) |Kh (x, y) K(x, y)|s dx nh 0, for a. e. y .
h0 h0

Lemma 3

(12)

Then the operator K : Lp () Lq (D) is compact. 2) Suppose that K(x, y), Kh (x, y), 0 < h < h0 , satisfy conditions of Lemma 2(2) with common s = q, N , and |Kh (x, y) K(x, y)|q dx nh 0, for a. e. y . (13)
h0

Then the operator K : L1 () Lq (D) is compact. 3) Suppose that K(x, y), Kh (x, y) satisfy conditions of Lemma 2(3) with common t = p , M , and condition (11) is satised with t = p . Then the operator K : Lp () C(D) is compact. (Here p > 1.) 43

Proof We denote (Kh u) (x) = vh (x) =

Kh (x, y)u(y)dy.

By Lemma 1, the operator Kh : Lp () C(D) is compact. Obviously, the embedding C(D) Lq (D) (for a bounded domain D) is continuous. Hence, the operator Kh : Lp () Lq (D) is also compact. 1) From conditions (11), (12) and the estimate (7) it follows that Kh K
Lp ()Lq (D)

mh

1/p

nh

1/q

0 as h 0.

Thus, K is the limit in the operator norm of compact operators Kh . Hence, K : Lp () Lq (D) is compact. 2) Similarly, if p = 1, from condition (13) and estimate (8) it follows that Kh K
L1 ()Lq (D) 1/q

nh

0 as h 0.

It follows that K : L1 () Lq (D) is compact. 3) From condition (11) with t = p and estimate (5), it follows that vh v
,D 1/p

mh

p, ,

u Lp ().

Hence, vh v ,D 0 as h 0. Since (Kh u) (x) = vh (x) is uniformly continuous, then v(x) is also uniformly continuous: v C(D). Thus, the operator K maps Lp () into C(D), and Kh K
Lp ()C(D)

mh

1/p

0 as h 0.
h0

Since Kh : Lp () C(D) is compact and Kh K in the operator norm, then K : Lp () C(D) is also compact operator.

3.
Now we apply Lemmas 13 to the study of the operator (Kj u) (x) =

xj y j u(y)dy, |x y|n

j = 1, . . . , n.

Here x or x m , where m is some section of by mdimensional hyperplane (m < n). So, either D = or D = m . If m = n, we agree that n . 44

Lemma 4 1) Suppose that 1 p n, n p < m n, q 1 and 1 n + m > 0. p q Then the operator Kj : Lp () Lq (m ) is compact. (If m = n, then m = ; if m < n, then m is arbitrary section of by mdimensional hyperplane.) 2) If p > n, then the operator Kj : Lp () C(m ) is compact. In particular, Kj : Lp () C() is compact. Proof The proof is based on Lemmas 2 and 3. 1) Case 1 < p n Suppose that conditions 1) are satised and, moreover, that q p > 1. We put n m n n =1 + , t= = n m; p q n1+ p + q s= m = n1+
n p

m t s + = 1. m . Then q p + q

s t Since q p, then p + p 1. Clearly, t < p . Since m n, q p, then 1. Hence, t 1. Thus, the numbers t and s satisfy conditions of Lemma xj y 2(1). Let us check that Kj (x, y) = |xy|j satisfy conditions a) and b) n with these t and s. Note that t(n 1) < n (since t(n 1) < t(n 1 + ) = n) and s(n 1) < m (since s(n 1) < s(n 1 + ) = m). We have

|Kj (x, y)|t dy =

|xj yj |t dy |x y|tn

dy

|x y|t(n1)

This integral converges since t(n 1) < n. Let d = diam, B(x) = {y Rn : |x y| d}. Obviously, B(x). Then

|Kj (x, y)|t dy

y=x+r n1

B(x)

dy |x y|t(n1)
d

n , n t(n 1)

0 dnt(n1)

r n1 dr r t(n1)

where n is the square of the unit sphere Sn1 in Rn . Thus, condition a) is satised with M= n dnt(n1) < . n t(n 1) 45

Let us check condition b): |Kj (x, y)|s dx dx |x y|s(n1) dx , |x y |s(n1)

where y is the projection of point y onto the hyperplane m (which contains m : m m ). The integral is nite, since s(n 1) < m. Consider the mdimensional ball Bm (y ) = {x m : |x y | d}. Clearly, m Bm (y ). Then
m

|Kj (x, y)|s dx

Bm (y )

dx m dms(n1) . = m s(n 1) |x y |s(n1)

Hence, condition b) is satised with N= m dms(n1) < . m s(n 1)

The constant N depends on m and on d = diam, but it does not depend on m (it is one and the same for all sections m of dimension m). Thus, conditions of Lemma 2(1) are satised and, therefore, Kj : Lp () Lq (m ) is continuous. We want to prove that this operator is compact. For this, we want to nd the operators Kjh satisfying conditions of Lemma 3. Let (r), r [0, ), be a smooth function such that C ([0, )), 1 (r) = 0 if 0 r 2 , (r) = 1 if r 1, and 0 (r) 1, r. r We put h (r) = ( h ). Then h (r) = 0 if 0 r h , 2 h (r) = 1 if r h. Consider the kernels Kjh (x, y) = xj y j h (|x y|). |x y|n

Obviously, |Kjh (x, y)| |Kj (x, y)| , x, y. Hence, Kjh satisfy conditions a) and b) together with K with the same constants t, s, M and N . Clearly, Kjh (x, y) are bounded: |Kjh (x, y)| h (|x y|) |x y|n1 1
h n1 2

= L(h).

So, Kjh satisfy condition c). And, nally, Kjh is uniformly continuous in both variables. So, condition d) for Kjh is also satised. Let us check

46

condition (11): |Kjh (x, y) Kj (x, y)|t dy =


|xy|<h

|xj yj |t (1 h (|x y|))t dy tn |x y|


1

|xy|<h n hnt(n1)

dy |x y|t(n1)

n t(n 1) 0 as h 0. Thus, (11) is true. Next, |Kjh (x, y) Kj (x, y)|s dx = = |xj yj |s (1 h (|x y|))s dx sn |xy|<h |x y| dx |x y |s(n1) {xm :|xy |<h}
m

m hms(n1) m s(n 1) 0 as h 0.

Thus, (12) is satised. Then all conditions of Lemma 3(1) are satised. Hence, the operator Kj : Lp () Lq (m ) is compact. (Recall that we assumed q p > 1). 2) If 1 q < p, then we apply the result that Kj : Lp () Lp (m ) is compact (i. e. , we apply 1) with p = q; condition 1 n + m > 0 is p p true, since m > n p). Since m is a bounded domain, then Lp (m ) Lq (m ) (if q < p), and any compact set in Lp (m ) is also compact in Lq (m ). It follows that the operator Kj : Lp () Lq (m ) is compact. 3) Case p=1 Condition n p < m n means that n 1 < m n. Then m = n, so, now m = . Next, condition 1 n + m > 0 means that p q n 1 n + n > 0 1 q < n1 . Let us check that condition b) with q s = q is true: |Kj (x, y)|q dx = |xj yj |q dx |x y|nq dx . |x y|(n1)q

The integral is nite since (n 1)q < n

|Kj (x, y)|q dx 47

n dnq(n1) . n q(n 1)

Also, condition (13) is satised: |Kjh (x, y) Kj (x, y)|q dx n hnq(n1) 0 as h 0. n q(n 1)

By Lemma 3(2), the operator Kj : L1 () Lq () is compact. 4) Case p n Let us check that conditions of Lemma 3(3) are satised. Indeed, the kernels Kj (x, y) (and Kjh (x, y) with it) satisfy condition a) with t = p : |Kj (x, y)|p dy n dnp (n1) dy < . n p (n 1) |x y|p (n1)
1 p

1 1 1 Since p > n, then p < n , p = 1 Hence n > p (n 1). Next,

>1

1 n

n1 n .

|Kjh (x, y) Kj (x, y)|p dy

|xy|h

n hnp (n1) n p (n 1) 0 as h 0.

dy |x y|p (n1)

Thus, conditions of Lemma 3(3) are satised. It follows that the operator Kj : Lp () C() is compact.

Lemma 5 If 1 < p < n, n p < m n, then the operator Kj : Lp () Lq (m ) is continuous (but not compact), where 1 Without proof. n m =0 + p q q = mp np .

48

1 2: Embedding theorems for Wp ()

1 1. The integral representation for functions in Wp ()

Lemma 6
1 Let Rn be a bounded domain. Let u Wp (). Then

u(x) =

1 n

n j=1

xj yj u dy, |x y|n yj

for a. e. x .

(14)

Proof
1) First, assume that u C0 (). Consider the fundamental solution of the Poisson equation E(z) = (z):

E(z) =

1 n (n2)|z|n2 1 2 ln |z|

n>2 n = 2.

Then for any u C0 () we have

u(x) =

E(x y)( u)(y)dy.

The function E(z) has weak derivatives 1 zj E(z) = . zj n |z|n


j j Then, yj E(x y) = 1 |xy|n , j = 1, . . . , n. n By Denition 1 of weak derivatives, we have

x y

E(x y)( u)(y)dy =


n j=1

j=1

E(x y) u dy. yj yj

Then, 1 u(x) = n

xj yj u dy, |x y|n yj

u C0 ().

1 1 2) Now, let u Wp (), and let uk C0 (), uk u in Wp (). For uk we have n 1 xj yj uk uk (x) = dy. n n |x y| yj j=1

49

k Thus, uk = 1n n Kj uj . j=1 y By Lemma 4, each operator Kj is compact from Lp () to Lp (). We

k Kj ( uk ) yj

u Kj ( yj ) in Lp (). (since Kj : Lp () Lp () is continuous operator.) Hence, by the limit procedure as k we obtain:

know that uk u in Lp () and

uk k u yj yj

in Lp (). Then

u(x) =

1 n

n j=1

xj yj u dy, |x y|n yj

1 u Wp ().

1 2. Embedding theorems for Wp ()

Theorem 1 Let Rn be a bounded domain.

1 1) If 1 p n, m > n p, q < and 1 n + m 0, then Wp () p q is embedded into Lq (m ), where m = (if m = n ) and m is any section of by mdimensional plane (if m < n). In the case 1 n + m > 0, this embedding is compact. p q 1 2) If p > n, then Wp () is compactly embedded into C().

Comments 1) Let us distinguish the case m = n (m = ):

np 1 If 1 p n, q < and q np = q , then Wp () Lq (). If , then this embedding is compact. q<q 1 2) What does it mean that Wp () Lq (m ) in the case m < n ? 1 A function u Wp () is a measurable function in ; it can be changed on any set of measure zero; m is a set of measure zero. First we consider u C0 (), and put T u = u|m . Then T : C0 () C0 (m ) is a linear operator. This linear operator can be extended by continuity to a continuous operator 1 T : Wp () Lq (m ). We have the estimate

Tu

q,m

c u

1 Wp () ,

u C0 ().

50

1 Let u Wp (). Then uk C0 (),

uk u

Then T uk T uj q,m c uk uj W 1 () 0. p Hence {T uk } is a Cauchy sequence in Lq (m ). There exists limit T uk w in Lq (m ). By denition, w = T u.


k

1 Wp () k,j

0.

Proof of Theorem 1

u 1 By Dj we denote operators Dj u = xj . Then Dj : Wp () Lp () is continuous operator, j = 1, . . . , n. Then representation (14) can be written as

u=

1 n

n j=1

Kj Dj u.

(15)

1) Suppose that 1 p n, m > n p, q < and 1 n + m > 0. Then p q conditions of Lemma 4(1) are satised. So, operator Kj : Lp () Lq (m ) is compact. Hence, the embedding operator
n

J=

1 n
j=1

1 Kj Dj : Wp () Lq (m )

is compact. (We use the fact that if A1 : B1 B2 is continuous operator and A2 : B2 B3 is compact operator, then A2 A1 : B1 B3 is compact. Here B1 , B2 , B3 are Banach spaces.) If p > 1 and 1 n + m = 0 (i. e., q = q ), then, by Lemma 5, the p q operator Kj : Lp () Lq (m ) is continuous. Hence, the embedding operator
n

J = 1 n
j=1

1 Kj Dj : Wp () Lq (m )

is continuous. For p = 1 without proof. 2) Let p > n. Then, by Lemma 4(2), operators Kj : Lp () C() are compact. Hence, the embedding operator
n

J= is compact.

1 n
j=1

1 Kj Dj : Wp () C()

51

Remark 1) Under conditions of Theorem 1(1), we have the estimate


n

q,m

1 n
j=1 n

Kj

Lp ()Lq (m )

Dj u

Lp ()

j u
j=1

p,

c u

1 Wp () ,

1 u Wp ().

(16)

2) Under conditions of Theorem 1(2), we have


n

C()

1 n
n

j=1

Kj

Lp ()C()

j u

p,

j u
j=1

p,

c u

1 Wp () ,

1 u Wp ().

(17)

Using the estimates from Lemma 4 it is easy to see that the constants in estimates (16), (17) depend only on diam, n, m, p, q, but they do not depend on m (they are one and the same for any section m ).
1 3. Embedding theorems for Wp ()

Theorem 2 Let Rn be a bounded domain of class C 1 . Then both statements 1 of Theorem 1 are true for Wp (). 1) If 1 p n, m > n p, q < and 1 n + m 0, then p q 1 Wp () is embedded into Lq (m ). In the case 1 n + m > 0, this p q embedding is compact.
1 2) If p > n, then Wp () is compactly embedded into C().

Proof Let Rn be a bounded domain such that . (For example, is a ball of suciently large diameter.) By Theorem 11 (Chapter 1), there exists a linear continuous extension ope1 1 1 1 rator : Wp () Wp (). If u Wp (), then v = u Wp (), and

1 Wp ()

c u

1 Wp ()

c= .

52

1 1) Under conditions of part 1), by Theorem 1,Wp () Lq (m ); m n if 1 p + q > 0, this embedding is compact. (Here m m , where m is mdimensional plane, and m is the section of by the same m .)

1 Then = {v = u : u } is a bounded set in Wp (). Then, by Theorem 1, this set is compact in Lq (m ). Then is compact in Lq (m ) (because functions in are restrictions of functions in 1 back to ). Hence, Wp () compactly embedded into Lq (m ). 1 b) Let 1 n + m = 0. In this case embedding Wp () Lq (m ) is p q continuous (but not compact). By similar arguments, we show that 1 embedding Wp () Lq (m ) is also continuous. 1 2) Under condition p > n, by Theorem 1(2), Wp () is compactly embed ded into C(). If is a bounded set in W 1 (), then is bounded 1 set in Wp ();

is compact in C(). Hence,

Comments
1 1) Under conditions of Theorem2(1), let J : Wp () Lq (m ) be 1 the embedding operator and let J : Wp () Lq (m ) be the em

1 1 bedding operator; : Wp () Wp () is the extension operator; R : Lq (m ) Lq (m ) is the restriction operator. Then J = RJ . 1 We have the estimate for all u Wp ().

q,m

RJ u J u J

q,

q,m

1 Wp ()Lq (m )

1 1 Wp ()Wp ()

=c1

=c2 1 u Wp ().

q,m

c u

1 Wp () ,

1 Compare (18) with estimate (16): in the case u Wp () we can n estimate u q,m by the norms of derivatives j=1 j u p, . Now it is impossible. (It is clear for u = const = 0 : u q,m = 0, but j u 0.)

53

is compact in C().

1 Wp ()

(18)

a) Let 1

n p

m q

> 0. Let

1 be some bounded set in Wp ().

2) Similarly, under conditions of Theorem 2(2), we have the estimate u


C()

C u

1 Wp () ,

1 u Wp ().

(19)

The constants in estimates (18), (19) depend on and, so, on the properties of . (While constants in estimates (16), (17) depend only on diam and on p, q, m, n.) Let us formulate the analog of Theorem 2 for unbounded domain. Theorem 3 Suppose that Rn is unbounded domain satisfying conditions of Theorem 12 (Chapter 1). Then 1) If p 1, m > n p, p q < and 1 1 Wp () Lq (m ).
1 2) If p > n, then Wp () C(). n p

m q

0, then

Remark 1) In Theorem 3 embeddings are continuous, but not compact. 2) In part 1) we have condition q p (we dont need this condition in Theorem 2.). 3) If is bounded and p > n, then 1) follows from 2). Now 1) does not follow from 2).

4. Comments. Examples.
All conditions in Theorems 2, 3 are precise.
1 1) If 1 n + m < 0, then Wp () Lq (m ). p q Example. Let = {x Rn : |x| < 1}. Let u(x) = |x| with 1 1 Then u Wp (), but u Lq (). / 1 Indeed, | u| c |x| ,

n p

< < m. q

| u|p dx c

|x|p(1) dx
1

= cn
0

r n1+p(1) dr n . p

< ,

since n 1 + p( 1) > 1 > 1

54

Also,

|u|p dx < . However, |u(y)|q dy =


m

|y|q dy
1

= m
0

r m1+q dr m ). q

= ,

since m 1 + q < 1( <

Here m is a section of by some mdimensional plane m such that point 0 m .


1 2) For unbounded domains, if p < q, then Wp () Lq (). Example Let = {x Rn : |x| > 1} , u(x) = |x| . Let n < < n . Then q p 1 u Wp (), but u Lq (). Check yourself. / m 3) The critical exponent q is dened by the relation 1 n + q = 0. p mp (q = np ). Here p < n. We have q > p, since m > n p. 1 Wp () Lq (m ) for q q , but not for q > q . If p n, then 1 Wp () Lq (m ) for all q < (if is bounded) and all p q < 1 (if is unbounded). If p > n, then Wp () C(). 1 () C() and even W 1 () L () . But for p = n > 1, Wn n (Here q = .) Example Let = x Rn : |x| < 1 . Consider u(x) = ln | ln |x||. e 1 1 Then u Wn (), but u L (). Indeed, | u(x)| |x|| ln |x|| . Then /

| u(x)|n dx

dx

|x|n | ln |x||n
1/e 0 1/e

= n = n < . Also,
0

r n1 dr r n | ln r|n dr r| ln r|n

1 |u(x)|n dx < . Then u Wn ().

1 4) If p = n = 1, = (a, b), then any function u W1 () is absolutely continuous. This follows from Theorem 5 (Chapter 1).

5) For unbounded domains embeddings from Theorem 3 are not compact. Example Let u C0 (Rn ) and let x(k) be a sequence of points x(k) Rn such 55

that x(k) as k . We put uk (x) = u(x x(k) ). Then the set 1 {uk } is bounded in Wp (Rn ). Here p > n. (Obviously, uk Wp ( n ) = 1 u Wp ( n ) = const.) But the set {uk } is not compact in C(Rn ). Indeed, 1 C(Rn ). Since ukj 0 in C() for any bounded domain (simply ukj 0 in for suciently large j), then u0 (x) 0. But ukj C( n ) = u C( n ) = 0. Contradication. Example
1 Let u C0 (Rn ), and vk (x) = k p u x . Then vk Wp (Rn ) and {vk } k 1 (Rn ). But {v } is not compact in L (Rn ). Thus, the is bounded in Wp p k 1 embedding Wp (Rn ) Lp (Rn ) is not compact. 1 6) For bounded domains and q = q embedding Wp () Lq () is not compact. Example = {x : |x| < 1}, u C0 (Rn ), wk (x) = k p 1 u(kx), p < n. Then 1 {wk } is bounded in Wp (), but {wk } is not compact in Lq (). Check this yourself.
n

suppose that there exists a subsequence ukj such that ukj u0 in


j

5. Embeddings on submanifolds
Instead of the section of by mdimensional planes we can consider sections of by some mdimensional manifolds. Theorem 4 Let Rn be a bounded domain of class C 1 . Let 1 p n, m > n p, 1 q < and 1 n + m 0. Let Rn be a manifold p q 1 of class C 1 , dim = m. Let = . Then Wp () Lq ( ). If n m 1 p + q > 0, then this embedding is compact. Without proof (The proof is based on Theorem 2 and using of covering fj and patition of unity.) Uj , diemorphisms

Important case = (then also = ). dim = n 1. Conditions: m = n 1 > n p p > 1, 1 n + n1 0 q (n1)p = q . p q np 1 If q < (1 < p < n), then Wp () Lq (), q q . For q < q this embedding is compact. If n = p > 1, then q = , 1 Wn () Lq (), q < .

56

l 3: Embedding theorems for Wp ()

Theorem 5 Let Rn be a bounded domain of class C 1 . 1) If p 1, 1 q < , 0 r < l, l r n + n 0, then p q l r Wp () Wq (). If l r n + n > 0, then this embedding is p q compact.
l 2) If p(l r) > n, then Wp () C r () and this embedding is compact.

Proof 1) We put s = l r and x the numbers q0 , q1 , . . . , qs such that qj 1, n n q0 = p, qs = q and 1 qj + qj+1 0. Such numbers exist due to condition s n + n 0. If l r n + n = 0, then q0 , . . . , qs are dened p q p q n n uniquely from the equations 1 qj + qj+1 = 0, j = 0, . . . , s 1. If n n = s p + q > 0, such numbers exist (but they are not unique). 1 By Theorem 2(1), Wqj () Lqj+1 ().
lj lj1 lj It follows that Wqj () Wqj+1 (). Indeed, let u Wqj (). 1 1 Then u Wqj () for || l j 1. Since Wqj () Lqj+1 (), u L then qj+1 () , || l j 1, and

qj+1 ,

c u

1 Wqj ()

c u

lj Wqj ()

for all with || l j 1.


lj1 u Wqj+1 () and

lj1 Wqj+1 ()

c u

lj Wqj ()

We denote the embedding operator by Jj ,


lj lj1 Jj : Wqj () Wqj+1 (), j = 0, 1, . . . , s 1.

Jj is a continuous operator. We have:


3 1 0 r ls l2 l1 l l Wp () = Wq0 () Wq1 () Wq2 () . . . Wqs () = Wq ().

Js1

l r The embedding operator J : Wp () Wq () is represented as J = Js1 . . . J1 J0 . Each operator Jj is continuous , then J is also continuous. If > 0, then at least one of Jj is compact (at least for n n one index j we have 1 qj + qj+1 > 0). In this case J is also compact.

57

2) Let p(l r) > n l r n > 0 p Case a) r = l 1 l l + 1 n > 0 p > n. By Theorem 2(2), the embedding p 1 l Wp () C() is compact. It follows that Wp () C l1 () and this l (), then u W 1 () C() for embedding is compact. (If u Wp p || l 1.) Case b) r < l 1 Then there exists a number q such that q > n and l(r+1) n + n > 0. p q (Indeed, l r n =: > 0. We can nd q > n such that 1 n < , p q n i. e., n < q < 1 .) l Then we can represent the embedding operator J : Wp () C r () l r+1 as J = J2 J1 , where J1 : Wp () Wq () (J1 is compact by part r+1 C r () (J is compact by case a), 1) of Theorem 5) and J2 : Wq 2 since q > n). l Hence, J : Wp () C r () is compact.

Particular cases 1) Let r = 0, pl < n. The critical exponent q is dened from the conditinp n on l n + q = 0 q = nlp . Since pl < n, then q < . Embedding p l Wp () Lq () is compact for q < q , and continuous for q = q .
l 2) If pl = n, then q = . In this case Wp () Lq () q < (and this embedding is compact). l But Wp () L (). l 3) If pl > n, then Wp () C() and this embedding is compact. l r 4) Let q = p, r < l. Then embedding Wp () Wp () is compact. In l () L () (for l 1) is compact. particular, embedding Wp p

Remarks
1 1) The embedding theorem for m with m < n (Wp () Lq (m )) can l be also generalized for Wp (). However, for the proof we need another l integral representation for u Wp () (including derivatives of higher order). l 2) The embedding theorems for Wp () can be also generalized for the case of unbounded domains.

58

l Equivalent norms in Sobolev spaces Wp ()

(lecture by prof. M. Birman)

1. Finitedimensional linear spaces and norms in these spaces


Let X be a linear space, dim X = N < . It means that there exists a system of linear independent elements x1 , . . . , xN X, such that any x X can be represented as a linear combination of x1 , . . . , xN :
N

x=
k=1

k xk , , k C , k = 1, . . . N.
2 1/2

(1)

N k = k k=1 . We denote x = . Check yourself, that this k=1 functional has all properties of the norm. X is a Banach space with respect to this norm (i. e., the space X with this norm is complete). The mapping x is an isometric isomorphism of X and CN (with the standard norm). Proposition Any other norm x on X is equivalent to x . Therefore, all norms on X are equivalent to each other.

There is a onetoone correspondence of elements x X and coordinates


N

Proof From (1) it follows that


N N

k=1

xk

k=1

1/2

1/2

xk
k=1 N 2

1/2

i. e. , x x ,

=
k=1

xk

> 0.

(2)

Now, let us prove the opposite inequality. Let us check that the function x is continuous on X with respect to x . From (2) and from the triangle inequality it follows that x x xx xx .

Now we restrict the continuous function x to the unit sphere x = 1. Then x is a continuous function of on the closed bounded set CN : || = 1 in CN . Since x > 0, then by the Weierstrass Theorem, x > 0 for x = 1. Then y = y y y y , y X. Thus, x x , x X. (3)

59

l 2. Trivial equivalent norms in Wp (). l The standard norm in Wp (), l N, 1 p < , is

l Wp ()

Let N be the number of all multiindices with || l. In CN we introduce the norm of lp type by the formula
N

||l

p Lp ()

1/p

Rd .

(4)

p
N

=
s=1

| s |p ,

CN .

(5)

Then we can rewrite (4) as

p l Wp ()

p
n

(6)

where = u Lp () , || l. If we replace the norm (5) in relation (6) by any other (equivalent!) norm l of vector in CN , then (6) will automatically dene some norm in Wp (), l which is equivalent to the standard one. Such new norms in Wp () are trivial. Example The norm u Lp () + max1||l u Lp () is equivalent to the standard l norm in Wp (). Give yourself several examples of new trivial norms in l Wp ().

3. The notion of seminorm.


Denition A functional on a linear space X is called a seminorm on X, if 1) 0 (x) < , x X,

2) (cx) = |c|(x),

3) (x1 + x2 ) (x1 ) + (x2 ).

x X, c C,

Thus, a seminorm has all properties of the norm besides one: from (x) = 0 it does not follow x = 0. Example l X = Wp (), (x) = u(x)dx . This functional is equal to zero for any l u Wp () with zero mean value. 60

l 4. General theorem about equivalent norms in Wp ().

Assume that Rn is bounded and C 1 . By Pl we denote the class l of all polynomials in Rn of order l 1. Let be a seminorm on Wp () which satises properties: 4) (u) c u nuous in
l Wp () l ().) Wp

(It means that is bounded, and, therefore, conti-

5) If u Pl and (u) = 0, then u = 0 ( is nondegenerate on the l subspace Pl W2 ()). Theorem l Let be a functional on Wp () satisfying conditions 1) 5). Then the functional 1/p
l denes the norm in Wp (), which is equivalent to the standard norm.

uWp () = l

p Lp ()

||=l

+ (u)p

(7)

Proof Obviously, functional (7) is homogeneous and satises the triangle inequality. Next, if uWp () = 0, then u = 0 for with || = l. Then it follows l that u Pl . Besides, (u) = 0, and, by property 5), u=0. Thus, functional l (7) is a norm on Wp (). Taking account of property 4), it suces to check that u
l Wp ()

CuWp () , l

l u Wp ().

(8)

Suppose the opposite. Then for any C > 0, (8) is not true. Then there exists l a sequence {um }, um Wp () such that mum Wp () um l We put vm =
um um W l () .
p l Wp () .

(9)

Then, by (9), vm
l Wp ()

= 1,

(10) (11)

vm Wp () l

1 0 as m . m

l l1 Since the embedding Wp () Wp () is compact it follows from (10) l1 that there exists a subsequence vmj , which converges in Wp () to some l1 v0 Wp ():

vmj v0

l1 Wp ()

0 as j

(12)

61

From (11) it follows that vmj


j Lp ()

0 for with || = l.

(13)

Since the operator is closed in Lp (), v0 = 0 for with || = l. Then by (12) and (13), we have vmj v0 as j ,
l Wp ()

v0 Pl .

(14)

From (11) it follows that (vmj ) 0 as j . By (14) and property 4), (vmj ) (v0 ) as j . Thus, (v0 ) = 0, v0 Pl . By property 5), v0 = 0. Together with (14) this contradicts to (10).

Mention that, in the proof of inequality (8), we did not use any explicit construction and we did not obtain any upper bound for the constant C. However, we have proved rather general theorem, which in particular cases implies a number of concrete inequalities (proved before by special tricks). Control question Where did we use that is bounded and C 1 ?

5. Examples. Additions.
1. Let l 2 and (u) = u Theorem, the norm
Lp () .

Conditions 1) - 2) are obviously satises. By 1/p

is equivalent to the standard one. It follows that u Lp () , 0 < || < l, is estimated by the norm (15). Exercise In the case p = 2, l = 2, prove this estimate using Fourier transform. 2. Let l = 1, , is a measurable set such that mesn > 0. Now Pl consists of constants. Let (u) = u(x)dx . Clearly, conditions 1) 5) are satised. Then the Theorem implies that u
p Lp ()

uWp () = l

p Lp ()

+ u

||=l

p Lp ()

(15)

| u|p dx +

u(x)dx

For = and p = 2 this is the classical Poincare inequality. 62

3. Let l = 1, , mesd1 > 0. We put (u) = udS . Properties 1) 5) are satised. Condition 4) follows from the estimate |u|p dS C u
p , l Wp ()

i. e. , from the trace embedding theorem. By Theorem (on equivalent norms) we obtain p u
p Lp ()

| u|p dx +

udS

This generalizes and strengthens the Friedrichs inequality |u|2 dx C | u|2 dx + |u|2 dS .

4. Let l = 2. P2 consists of linear functions, i. e. , of linear combinations of the basis funcrtions 1, x1 , . . . , xn . Let be a measurable set, mesd > 0. We put
n

(u) =

u(x)dx +
k=1

xk u(x)dx .

(16)

We have to check condition 5). Consider P2 as a nitedimensional subspace in L2 (). If (u) = 0, then u is orthogonal in L2 () to the basis in P2 . Then, if u P2 , it follows that u = 0. Thus, the norm (7) with l = 2 and such (u) is equivalent to the 2 standard norm in Wp (). 5. Let l = 2, , mesd1 > 0. We put (u) =

|u|dS.

(17)

Condition 4) follows from the trace embedding theorem. Let us check 5): (u) = 0 u| = 0. If u P2 (u(x) is a linear function), then condition u| = 0 and u = 0 is equivalent to the fact that is a plane part of the boundary, and u(x) = 0 is equation of this plane. In the case where does not lie in some plane, from u P2 , u| = 0, it follows that u = 0. Then the norm (7) with l = 2 2 and (u) given by (17) is equivalent to the standard norm in Wp (). In particular, it is always so, if = .

63

6. In conclusion, we discuss one example, which does not follow from Theorem. The norm
n

uWp () = l
k=1

lu (xk )l

1/p

+ u
Lp ()

p Lp ()

is equivalent to the standard one. 2 For example, in W2 (Rd ) = H 2 (Rd ), this fact follows from the inequality j k | | j |2 + | k |2 . 2|

64

Chapter 3: Sobolev spaces H s (Rn ) 1: Classes S(Rn ) and S (Rn ). Fourier transform.
Denition S(Rn ) is a class of functions C (Rn ) such that for any multiindex and any k N,
x

sup (1 + |x|)k | (x)| < .

S(Rn ) is called the Schwartz class. For S(Rn ) all derivatives (x) are rapidly decreasing as |x| . We can introduce topology in S(Rn ). Denition We say that m in S(Rn ), if
x m m

sup (1 + |x|)k | m (x) (x)| 0, , k.

S(Rn ) is a topological space, but not Banach space. Denition Let f S(Rn ). We dene the transformation F : f f ,

f () = (2)

n 2 n

f (x)eix dx.

F is called the Fourier transformation. It is known that f S(Rn ), if f S(Rn ). So, F : S(Rn ) S(Rn ) is a linear operator. The inverse transformation F 1 is given by the formula

f (x) = (2)

n 2 n

f ()eix d,

F 1 : S(Rn ) S(Rn ). It is known that the Fourier transform F can be extended by continuity to L2 (Rn ), and F is unitary operator in L2 (Rn ): F : L2 (Rn ) L2 (Rn ),
n

|f (x)|2 dx =

f () d, f L2 (Rn ). 65

Denition By S (Rn ) we denote the dual space to S(Rn ) (i. e., the space of linear continuous functionals on S(Rn )). Sometimes, S (Rn ) is called the space of slowly increasing distributions. If v S (Rn ), S(Rn ), by v, we denote the meaning of functional v on function . The Fourier transformation is extended to the class S (Rn ). Denition Let f S (Rn ). A functional f S (Rn ) is called the Fourier image of f , if f , := f, , S(Rn ). It is known that F : S (Rn ) S (Rn ), F 1 : S (Rn ) S (Rn ).
onto

66

2: Spaces H s (Rn )
1. Denition of H s (Rn )
l l We know that the spaces W2 () (l N) are Hilbert spaces : W2 () = H l (). n . We can use the Fourier transform and express the norm in Let = R l W2 (Rn ) = H l (Rn ) in terms of the Fourier image. Let u H l (Rn ). Consider the Fourier image

Then

u(x) = (2)

n 2 n

For the derivatives u(x), we have u() = (i) u() = i|| u(). Then

||l

Since c1 (1 + ||2 )l

||l

| |2 c2 (1 + ||2 )l (prove this!), then



2 Hl(
n)

c1

(1 + ||2 )l |u()|2 d u

Thus, the norm

l norm in W2 (Rn ). We introduce the space with this norm; now we consider arbitrary l (not only l N). Denition

(1 + ||2 )l |u()|2 d

H s (Rn ) = u S (Rn ) : n (1 + ||2 )s |u()|2 d < , The inner product in H s (Rn ) is dened by

Theorem 1
H s (Rn ) is the closure of C0 (Rn ) with respect to the norm u Hs .

67

(u, v)H s (

n)

(1 + ||2 )s u()v()d.

2 Hl(

n)

| u|2 dx =

u() = (2)

n 2 n

u(x)eix dx.

u()eix d.

||l

| |2 |u()|2 d.

c2
1/2

(1 + ||2 )l |u()|2 d.

is equivalent to the standard

s Rn .

Proof 1) Let us show that any u H s (Rn ) can be approximated by functions in C0 (Rn ). If u H s (Rn ), then u () = u()(1 + ||2 )s/2 L2 (Rn ). Since C0 (Rn ) is dense in L2 (Rn ), there exists a sequence vk C0 (Rn ) such that vk () u () in L2 (Rn ). We put vk ()(1 + ||2 )s/2 = wk (). Then wk C0 (Rn ) and wk ()(1 + ||2 )s/2 u () in L2 (Rn ). Obviously, wk S(Rn ). We put uk = F 1 wk . Then also uk S(Rn ) L2 (Rn ), then uk u in H s (Rn ). It remains to approximate functions uk S(Rn ) by functions ukj C0 (Rn ) (in the H s norm). For this, we x h C0 (Rn ) such that h(x) = 1 for |x| 1. We put ukj (x) = uk (x)h C0 (Rn ) and

x j k

and wk = uk . Since uk ()(1 + ||2 )s/2 u () = u()(1 + ||2 )s/2 in


k

. Then ukj
s

ukj uk

2 Hs

|ukj () uk ()|2 1 + ||2

ls,l

|ukj () uk ()|2 1 + ||2 d


2 Hl

ukj uk

For l N we can use another norm (which is equivalent to the standard one): ukj uk
2 Hl

||l

c
||l |x|>j

0 as j . It follows that ukj uk in H s (Rn ).


2) Let us show that each element of the closure of C0 (Rn ) in H s norm belongs to H s (Rn ). Suppose that um C0 (Rn ) and {um } s (Rn ), i. e. , u u is the Cauchy sequence in H m l H s ( n ) 0 as 2 )s/2 =: u () is a fundamenm, l . It means that um ()(1 + || m tal sequence in L2 (Rn ). Since L2 (Rn ) is complete, there exists a limit m u () u () in L2 (Rn ). We put w() = u ()(1 + ||2 )s/2 . Then m w S (Rn ), and, therefore, F 1 w = u S (Rn ). We have: j

w() = u(),

u () = u()(1 + ||2 )s/2 L2 (Rn ),


m

um ()(1 + ||2 )s/2 u()(1 + ||2 )s/2 in L2 (Rn ). 68

uk (x) 1 h | uk (x)|2 dx

x j

dx x j = 1 for |x| j.

since h

It means that um u in H s (Rn ). Thus, each element of the closure of C0 (Rn ) in H s belongs to H s (Rn ).

2. Duality of H s and H s .
Theorem 2
Let u H s (Rn ), v H s (Rn ), and let uj , vj C0 (Rn ), uj u in j

H s (Rn ), vj v in H s (Rn ). Then there exists the limit


j

uvdx u

Proof We have
n

Since uj u in H s (Rn ), it follows that uj ()(1 + ||2 )s/2 u()(1 + ||2 )s/2 =: (As u) () in L2 (Rn ). The fact that vj v in H s (Rn ) means that vj ()(1 + ||2 )s/2 v()(1 + ||2 )s/2 =: (As v) () in L2 (Rn ). Then, by (1), we have
n n

It is clear that the limit limj n uj vj dx does not depend on the choice of the sequences {uj } and {vj }. We have:
n

L2 (

n)

L2 (

Hs

H s .

69

As u

uvdx

(As u) ()(As v) ()d As v


n)

uj (x)vj (x)dx

uj (x)vj (x)dx =

(As u) ()(As v) ()d =:

We denote this limit by

u(x)v(x)dx. We have
Hs

uj ()vj ()d uj ()(1 + ||2 )s/2 vj ()(1 + ||2 )s/2 d. (1)

lim

uj (x)vj (x)dx.

H s

u(x)v(x)dx.
n

Theorem 3 If v H s (Rn ), then


0=uH s (

0=uC0 (

Proof 1) The mapping As : H s (Rn ) L2 (Rn ), (As u)() = u() 1 + ||2 is a onetoone isometric mapping. Indeed, As u L2 = u H s . The inverse mapping A1 : L2 (Rn ) H s (Rn ) is dened as follows: for s u () u L2 consider w() = (1+||2 )s/2 , and put u = F 1 w.
s/2

Then u() = w() and u () = u()(1 + ||2 )s/2 = (As u)(). Thus, A1 u = F 1 w = u. The mapping As : H s L2 is dened s similarly.

2) Let v H s and v () = (As v)(). Then v L2 . It is known that in L2 we have


n

L2

= sup
0=gL2

We put u = A1 g. Then g() = (As u)(), g L2 = u H s . If g runs s over L2 , then u runs over H s . Thus, for v = As v we have v
H s

= =

L2
n

0=uH s

sup

(As u) ()(As v) ()d u


Hs

u(x)v(x)dx u
Hs

0=uH s

sup

From Theorems 2 and 3 it follows that l(u) = n uvdx is a linear continuous functional on u H s (Rn ) (if v H s (Rn )) and the norm of this functional is equal to v H s : l = sup

0=uH s

|l(u)| = sup u Hs 0=uH s

70

sup

n)

g()v ()d g
L2

uvdx
Hs

H s

sup

uvdx
Hs

n)

u
n

uvdx
Hs

(2)

= v

H s

Riesz Theorem Let H be a Hilbert space and l(u), u H, be a continuous linear functional on H. Then there exists such element v H that l(u) = (u, v)H . This element v is unique and l = v H . Proof 1) Let N = Ker l = {z H : l(z) = 0}. Then N is a closed subspace in
j j

H. Indeed, if zj N and zj z in H, then l(zj ) l(z). Since l(zj ) = 0, it follows that l(z) = 0, i. e. , z N .

2) If N = H, then l(u) = 0, u H. In this case v = 0. If N = H, then N = {0} ( where N is the orthogonal complement of N ). So, there exists v0 N , v0 = 0. Then, l(v0 ) = 0. 3) For u H consider u Since v0 N , we have u Denote v= Indeed, l u
l(u) l(v0 ) v0 l(u) l(v0 ) v0

N.

= l(u)

l(u) l(v0 ) l(v0 )

= 0. v0 2 . l(v0 )

l(u) v0 , v0 l(v0 )

= 0 (u, v0 ) = l(u)

l(v0 ) v . v0 2 0

Then l(u) = (u, v).

4) Uniqueness If (u, v) = (u, v ), u H, then v v H v v = 0. 5) The norm of l. l = sup


0=uH

|(u, v)| |l(u)| = sup = v u H u H 0=uH

Indeed, |(u,v)| v u
H

for 0 = u H, and for u = v we have

|(u,v)| u H

= v

H.

Let l(u) be a continuous linear functional on H s (Rn ). It means that l : H s C, a) l(c1 u1 + c2 u2 ) = c1 l(u1 ) + c2 l(u2 ), u1 , u2 H s , c1 , c2 C, b) |l(u)| c u
Hs ,

u H s (Rn ). |l(u)| . u Hs

The norm l of a functional l is dened by the formula l =


0=uH s

sup

71

Theorem 4 Let l(u) be a linear continuous functional on H s (Rn ). Then there exists unique element v H s (Rn ), such that

l(u) = and

uvdx,
n

u H s (Rn ),

(3)

l = v Proof Consider the mapping

H s

(4)

As : H s (Rn ) L2 (Rn ), (As u) () = u () = u() 1 + ||2 ) = l(A1 u ) = l(u). l(u s

s/2

Then u = A1 u . We dene the functional ) on L2 (Rn ) by the formula l(u s

Then is a linear continuous functional on L2 (Rn ). l By the Riesz theorem for the functional there exists unique function l n ) such that w L2 (R

) = l(u

u ()w()d, and

= w l

L2

s/2 Then l(u) = ) = n u() 1 + ||2 l(u w()d. 1 w()(1 + ||2 )s/2 . We denote v(x) = F Then

So, v H s , and v H s = w L2 . We have w() = (1 + ||2 )s/2 v() = (As v) (), l(u) = ) = n (As u) ()(As v) ()d = n uvdx. l(u For the norm of the functional l we have: l = sup ) l(u |l(u)| l = = w = sup u Hs u L2 0=u L2

0=uH s

L2

= v

Remark Theorem 4 means that H s (Rn ) is dual to H s (Rn ) with respect to L2 duality. 72

v() = w()(1 + ||2 )s/2 ;

|v()|2 (1 + ||2 )s d =

|w()|2 d.

H s

3. Mollications in H s (Rn )
Let (x) = n
x

be a mollier.

Recall that (x) 0, n (x)dx = 1. For u H s (Rn ) consider mollications: u (x) = ( u)(x), > 0. Theorem 5 If u H s (Rn ), then u u
Hs

C0 (Rn ),

0 as 0.

Proof For the Fourier transform of the convolution u = u we have u () = (2)n/2 ()u(). Next, () =
x =y

().

Since C0 (Rn ), then () belongs to the Schwartz class S(Rn ). Hence,

a) |()| c, Rn .

b) lim0 () = (0) = (2)n/2

Let us estimate the norm u u

Hs .

u () u() = (2)n/2 () 1 u();

u u

2 Hs

(1 + ||2 )s |u()|2 (2)n/2 () 1 d.


0 as 0

The function under the integral is estimated by C(1 + ||2 )s |u()|2 , which is summable since u H s . By the Lebesgue Theorem, u u H s 0 as 0.

73

(2)n/2

(2)n/2

n
n

eix dx

(y)eiy dy
n

(y)dy = (2)n/2 .
n

=1

We have

4. Embedding H s C r
Theorem 6 Let s > r + n . Then H s (Rn ) C r (Rn ). 2 Proof
1) Let u C0 (Rn ). We have

Then, by the Hlder inequality, o

| u(x)|

| | |u()| d
n

||r x

max max | u(x)| C u n i. e. , u

Cr

C u

2) Let u H s (Rn ). Then there exists a sequence uj C0 (Rn ), such that

u C r (Rn ): uj u C r 0. In fact, u(x) = u(x), for a.e. x Rn (check this!). We identify u = u. We have proved that H s C r and u
Cr

uj u in H s . By (5), uj ul C r C uj ul H s 0. So, {uj } is the Cauchy sequence in C r (Rn ). There exists a limit
j

C u

Remark Theorem 6 is generalization of the embedding theorem: l W2 C r if 2(l r) > n.

74

If || r, and s r > n , then 2

| |2 d (1+||2 )s Hs Hs

< . Thus, , ,
u C0 (Rn ), u C0 (Rn ).

Hs

, u H s.

| |2 (1 + ||2 )s d

1/2
n

|u()|2 (1 + ||2 )s d

j,l

u(x) = (2)n/2

u()eix d,

u(x) = (2)n/2

(i) u()eix d, .

1/2

(5)

5. Equivalent norm in H s with fractional s > 0


Theorem 7 If 0 < s < 1, the norm u u
Hs

Hs

is equivalent to the norm |u(x) u(y)|2 dxdy |x y|n+2s


1/2

Proof Note that F : u(x) u(x + z) u()(1 eiz ). We have


n n n

P arseval
n

| where g() = n |1en+2sdz . |z| The function g() is homogeneous in of order 2s:

iz 2

The function g() depends only on ||:

g() =

where the axis 0z1 has direction of vector . It follows that g() = A||2s , A > 0. Then

Hs

|u()|2 (1 + A||2s )d

Obviously, c1 (1 + ||2 )s 1 + A||2s c2 (1 + ||2 )s , Rn . Then u H s u Hs . Corollary If s > 0 , [s] = k , {s} > 0, then the norm
||k
n

||=k

is equivalent to u

Hs .

75

Hs

| u|2 dx +

g(t) =

|1 eitz |2 dz = t2s |z|n+2s

|1 eitz |2 d(tz) = t2s g(), t > 0. |tz|n+2s

|1 eiz1 || |2 dz , |z|n+2s

g()|u()|2 d,
n

|u(x) u(y)|2 dxdy |x y|n+2s

y=x+z

|u| dx +

|u(x) u(x + z)|2 dxdz |z|n+2s |u()|2 |1 eiz |2 ddz |z|n+2s

1/2

1/2 | u(x) u(y)|2 dxdy |x y|n+2{s}

Proof u H s u H k and u H {s} with || = k. It is easy to check that u


2 Hs check this!

2 Hk

+
||=k

2 H {s}

by Theorem 7 ||k

||=k

6. inequalities
Obviously, H s1 (Rn ) H s2 (Rn ) for s1 > s2 . Proposition Let s1 < s < s2 . Then for > 0 C() > 0 such that u
2 Hs

2 H s2

+ C() u

2 H s1

Proof (6) is equivalent to the inequality 1 + ||2


s

s s2 + C()s1 , 1
s2 s

1 + ||2

s2

+ C() 1 + ||2
(ss1 )

+ C()

1. ,

We denote = s2 s > 0, and put C() = (ss1 ) =

s2 s + C()(ss1 ) = ()s2 s + ()(ss1 )

76

| u|2 dx +

| u(x) u(y)|2 dxdy . |x y|n+2{s}

(6)

s1

1.
s
ss1 2 s

. Then 1

obviously

3: Trace embedding theorems


We write x Rn as x = (x , xn ), x = (x1 , . . . , xn1 ). Consider the traces of functions on the hyperplane xn = 0. We dene the trace operator
0 : C0 (Rn ) C0 (Rn1 ),

(0 u) (x ) = u(x , 0).

Theorem 8 Let s > 1 . Then the trace operator 0 : C0 (Rn ) C0 (Rn1 ) can be 2 extended by continuity to the linear continuous operator 1 0 : H s (Rn ) H s 2 (Rn1 ). We have
H s 2 (

Proof
1) Let u C0 (Rn ). Then

=0 u( )

Then 0 u( )
2

1 0 u( ) = 2

u( , n )dn .

1 Here the second integral is nite, since s > 2 . We have

|u()|2 (1 + ||2 )s dn

2 (1 + | |2 +n )s dn . (8) =a2

dn 2 + 2 )s (a n

a a2s

d 1+

n a n a 2 s

= a12s
12s

dt (1 + t2 )s
=cs

= cs a dn 2 (1 + | |2 + n )s 77

= cs (1 + | |2 ) 2 s .

(0 u) (x ) = u(x , 0) = (2)

n1 2 n1

u(x) = u(x , xn ) = (2)n/2

n1 )

Hs(

u( , n )eixn n eix d dn ; d eix 1 2

0 u

C u

n)

(7)

u( , n )dn

(9)

Thus, from (8) and (9) it follows that 1 + | |2


s 1 2

0 u( )

cs

|u()|2 (1 + ||2 )s dn .

Integrate over Rn1 :


n1

n1 )

2) C0 (Rn ) is dense in H s (Rn ). Hs(

n1 )

By denition , v = 0 u. By the limit procedure, the estimate (10) is extended to all u H s (Rn ). Corollary
1 Let k N and s > k + 2 . Then the trace operators 1 j j = 0 xn : H s (Rn ) H sj 2 (Rn1 ) are continuous for j = 0, 1, . . . , k. We have H sj 2 (

So, {0 uj } is a Cauchy sequence in H s 2 (Rn1 ). Then there exists a limit: 1 1 j 0 uj v H s 2 (Rn1 ) in H s 2 (Rn1 ).

Theorem 9 (extension theorem) Denote H s 2 (Rn1 ) = H s 2 (Rn1 )H s 2 (Rn1 ). . .H sk 2 (Rn1 ). There exists a linear continuous operator P :H
1 s 2

1 Let k Z+ , s > k + 2 .
1

(Rn1 ) H s (Rn ),
1

such that, if = (0 , 1 , . . . , k ) H s 2 (Rn1 ), u = P ( H s (Rn )), then j = j u, j = 0, 1, . . . , k. We have


k H (
n1 )

j=0

78

2 Hs(

n)

1 s 2

n1 )

Hs(

=c

j u

c u

n)

0 uj 0 ul

2 1 H s 2 (

cs uj ul

2 Hs(

j,l
n)

0.

2 1 H sj 2 (

Let u H s (Rn ). Then {uj }, uj C0 (Rn ), uj u By (10),

i. e.

0 u

2 1 H s 2 (

cs u

2 Hs(

1 + | |2

1 s 2

0 u( ) d

cs

|u()|2 (1 + ||2 )s d,
n)

, u C0 .

(10)

j
n)

0.

n)

Proof Let h C0 (R), h(t) = 1 for |t| 1, 0 h(t) 1. We put


k

V ( , xn ) =
j=0

1 j x j ( )h xn j! n

1 + | |2 , Rn1 , xn R.

Here j ( ) is the Fourier image of j (x ). Clearly, V ( , 0) = 0 ( ), j xn V ( , 0) = j ( ), j = 1, . . . , k. Let us show that V ( , xn ) is the Fourier image of the function u(x , xn ) such that u H s (Rn ). We put u( , n ) = V ( , n ), where V ( , n ) is the Fourier image (in one variable xn n ) of V ( , xn ). Note that xj g(xn ) ij g(j) (n ) = ij n g(xn )
F F F

dj
j dn

g(n ),

1 g 1

, n .

xj h(xn ) ij n Then u() =


j=0 k

j+1

h(j)

j+1 ij j ( )(1 + | |2 ) 2 h(j) j!

n 1 + | |2

We have:

k j=0

j=0

n1

1+| |

2 1 + ||2 = 1 + | |2 + n = (1 + | |2 )(1 + 2 );

n1

Since h C0 (R), then h S(R) and, so,

h(j) ( ) (1 + 2 )s d = C(j, s) < .

79

2 Hs(

n)

d j ( ) (1+| |2 )sj 2

We write the integral as change variable: = n

2 Hs(

n)

|u()|2 (1 + ||2 )s d j ( ) (1 + | | ) d . Then 2


2 2 j1

(j)

n 1 + | |2

(1 + ||2 )s d.

dn . . ., and in the internal integral

h(j) ( ) (1+ 2 )s d.

k j=0 k j=0

So, the operator P : = (0 , 1 , . . . , k ) u is a linear continuous operator 1 from H s 2 (Rn1 ) to H s (Rn ) , and j u = j , j = 0, . . . , k.

80

= c

2 Hs(

n)

n1

j ( ) (1 + | |2 )sj 2 d . .

2 1 H sj 2 (

n1 )

4: Spaces H s () (survey)
1. Denition of H s ()
Let Rn be a domain. There are dierent ways of denition of the Sobolev spaces H s (). Approach I. Denition 1 H s () is the class of restrictions to of functions in H s (Rn ): u H s () v H s (Rn ), v| = u. Approach II. Case s 0. Denition 2 H s () is the set of functions in L2 (), such that their weak derivatives up to order k = [s] also belong to L2 (), and the following norm is nite: u H s < , 2 if s = [s] ||s | u| dx, 2 def | u(x) u(y)|2 dxdy 2 u Hs = , (11) ||k | u| dx + ||=k |xy|n+2{s} if s = [s] = k, {s} = s k.

Comments

1) If Rn is a bounded domain of Lipschitz class, then both denitions give the same space: Def 1 Def 2. If H s (), s 0, is the Sobolev space in the sense of Def 2, there exists a linear continuous extension operator : H s () H s (Rn ).
s 2) The spaces Wp () with fractional s 0 and p = 2 can be dened by analogy with Def 2 (with 2 replaced by p).

3) The embedding theorems can be generalized for spaces of fractional order. Next, the space H s () is dened. Denition 3
H s () is the closure of C0 () with respect to the norm (11).

81

Denition 4 Let s > 0. Then, by denition, H s () = H s ()


, i. e. , H s ()

is the space of linear continuous functionals on H s () with the norm u


H s

sup
0=H s ()

| u, | . H s ()

By analogy with H s (Rn ) and H s (Rn ), for u H s (), H s (), we denote u, = u(x)(x)dx.

Comments
1 1) H s () = H s () for s < 2 .

2) Let u H s (), P0 u(x) =

1 Then P0 : H s () H s (Rn ) is continuous, if s = m + 2 , m Z+ .

u(x), x . 0, x Rn \

3) Let Rn be a bounded domain with Lipschitz boundary. Then H s () coincides with the space of restrictions to of distributions H s (Rn ), if s = m + 1 , m Z+ . 2 4) H s () is invariant with respect to dieomorphisms of class C l , l |s|, l N.

2. Trace embedding theorems


Theorems 8 and 9 can be extended to the case of bounded domain with smooth boundary. Let Rn be a bounded domain of class C l . Then there exists a covering {Uj }j=1,...,N such that N Uj , either Uj j=1 1 , or Uj = , then dieomorphism fj : Uj K, fj , fj C l , fj (Uj ) = K+ , fj (Uj ) = K+ \+ = . Suppose that domains U1 , U2 , . . . , UM are of second kind, and UM +1 , . . . , UN are strictly interior. There exists a partition of unity {j }, such that j C0 (Rn ), supp j Uj , N j=1 j (x) = 1, x . For x we have M j (x) = 1. (This is true in some neighbourhood j=1 1 of .) Let u C l (), uj (x) = u(x)j (x), vj = uj fj . Then vj C l (), supp vj . We extend vj by zero to Rn1 \: vj (y) = vj (y), y . 0, y Rn1 \ 82

Then vj C l (Rn1 ), supp vj . Consider 1/2


M j=1

Denition

H s () is the closure of C l () with respect to the norm (12). This norm depends on the choice of covering {Uj }, dieomorphisms {fj }, and partition of unity {j }. It can be proved that all such norms (for dierent {Uj } , {fj } , {j }) are equivalent to each other. So, the class H s () is well dened. Theorem 10 (trace embedding theorem) Let Rn be a bounded domain of class C l , l N. Let k Z+ , 1 s > k + 2 , s l. Let j : C l () C lj () be the trace operator: are normal derivatives of u). Then the operator j can be extended (uniquely) to linear continuous 1 operator j : H s () H sj 2 (), j = 0, 1, . . . , k.
j u , j

j u =

j = 0, . . . , k (where

The proof is based on Theorem 8, and using covering {Uj }, dieomorphisms {fj } and partition of unity {j }. Theorem 11 (extension theorem)
1 Let Rn be a bounded domain of class C l , s l, s > k + 2 , where k Z+ . We denote

1 s 2

() = H s 2 () H s 2 () . . . H sk 2 ().
s 1 2

There exists a linear continuous operator P : H () H s ()


1

such that, if = (0 , 1 , . . . , k ), j H sj 2 (), and u = P , then j u = j , j = 0, 1, . . . , k, and


k

u Theorem 12

2 H s ()

Let Rn be a bounded domain of class C l , l N. Then u H l () =


l W2 () if and only if u H l () and 0 u = 1 u = . . . = l1 u = 0.

vj

2 Hs(

n1 )

def

= u

H s () ,

s l.

(12)

j j

j
j=0

2 1 H sj 2 ()

=c

2 H

1 s 2

()

83

Proof For simplicity we prove Theorem 12 in the case l = 1. u H 1 ()

u H 1 () and 0 u = 0.

Obvious. Using covering, dieomorphisms and partition of unity, we reduce the question to the following. Let K+ = {x Rn : |x| < 1, xn > 0} be the halfball. Suppose that u H 1 (K+ ), u(x) = 0 near + , 0 u = u| = 0. We have to prove that u H 1 (K+ ). We have the following representation for u(x):
xn

u(x , xn ) =
0

u (x , t)dt, for a. e. (x , xn ) K+ . xn

(13)

1 We x a cuto function h(t) such that h C (R+ ), h(t) = 0, 0 t 2 , h(t) = 1 for t 1, and 0 h(t) 1. We put hm (t) = h(mt), then hm (t) = 1 1 for t m . Consider um (x) = u(x , xn )hm (xn ). Then um (x) = 0 near K+ , 1 (K ). um H + m Let us check that um u H 1 (K+ ) 0. We have:

u(x , xn ) um (x , xn ) = (1 hm (xn ))u(x , xn ), u(x) (u(x) um (x)) = (1 hm (xn )) , j = 1, . . . , n 1, xj xj u(x) hm (u(x) um (x)) = (1 hm (xn )) u(x). xn xn xn Then |u(x) um (x)|2 dx =
K+

K+

|1 hm (xn )|2 |u(x)|2 dx


1

(u um ) dx xj
2

0 as m ; =
K+

K+ {

1 0<xn < m

|u(x)|2 dx u xj
2 2

K+

|1 hm (xn )|2

dx

u 1 K+ {0<xn < m } xj 0 as m ;

dx

84

K+

2 (u um ) dx xn

1/2

=
K+

|1 hm (xn )|

u xn

1/2

dx

+
K+

0 as m 1/2 hm 2 2

xn

|u| dx

=Jm [u]

It remains to show that Jm [u] 0 as m . We have: hm (x) = (h(mxn )) = mh (mxn ). xn xn Using (13), we obtain: Jm [u] = m2
K+

h (mxn ) h (mxn )
K+ C

2 0 2

xn

2 u (x , t)dt dx xn 2 u (x , t) dt xn 2 xn 0 =xn xn

m2

xn 0

12 dt dx

cm2 cm2

u 1 K+ {0<xn < m } xn 0 as m . 1 m2

K+ {

1 0<xn < m

xn
0 2

u (x , t) dt dxn dx xn dx

Next step: consider mollications of um (x): (um ) , > 0. Then (um )


C0 (K+ ) for suciently small and (um ) um H 1 (K+ ) 0. Thus, we can approximate function u(x) by functions (um ) C0 (K+ ) in 0

H 1 (K+ )norm u H 1 (K+ ).

85

5: Application to elliptic boundary value problems


1. Dirichlet problem for the Poisson equation
Let Rn be a bounded domain. Consider the classical Dirichlet problem: u = F, x u| = g. (1)

If (x) is arbitrary function in such that | = g, then the function v(x) = u(x) (x) is solution of the problem v = f, x v| = 0 (2)

where f (x) = F (x) + (x). First, well study problem (2) with homogeneous boundary condition. In the classical setting of problem (2), the boundary is suciently smooth, f C() and solution v C 2 (). Now we want to dene weak solution of problem (2) under wide conditi ons on and f . Let us formally multiply equation v = f by the test function C0 () and integrate over . Then v(x) satises the integral identity v dx =

f dx,

C0 ().

(3)

1 The lefthand side is welldened for any v H 1 () = W2 (), H 1 ();

and the righthand side is welldened for f H 1 (), H 1 () (since

H 1 () is the dual space to H 1 () with respect to L2 duality). The boun-

dary condition v| = 0 we understand in the sense that v H 1 (). Then we can consider arbitrary domains. Denition Let Rn be arbitrary bounded domain. A function v H 1 () is called a weak solution of the Dirichlet problem (2) with f H 1 (),

if v satises the identity (3) for any H 1 (). Theorem 1

Let Rn be a bounded domain. Then, for any f H 1 (), there exists unique (weak) solution v H 1 () of the Dirichlet problem (2). We have v H 1 () C f H 1 () . 86

Proof 1) The form [v, ] :=


v dx,

v, H 1 (),
1/2

denes an inner product in the space H 1 (). The corresponding norm [v, v]1/2 is equivalent to the standard norm v H 1 () = |v|2 + | v|2 dx This follows from the Friedrichs inequality: |v|2 dx C | v|2 dx, v H 1 ()

(here it is important that is bounded). 2) The righthand side of (3) is lf () =

f dx.

lf () is antilinear continuous functional on H 1 (): |lf ()| f


H 1 ()

H 1 () .

We rewrite (3) in the following form: [v, ] = lf ().

(4)

By the Riesz Theorem, for antilinear continuous functional lf on H 1 () there this functional is equal to the norm of v. (Now we consider H 1 () as the Hilbert space with the inner product [, ].) Then, by the Riesz Theorem, lf = sup
0=H 1 ()

exists unique function v H 1 () such that lf () = [v, ], and the norm of

[, ]1/2

|lf ()|

= [v, v]1/2 .

(5)

Thus, v is the unique solution of (4) ( (3)). Since, by denition of the class H 1 (), |lf ()| f H 1 () = sup , H 1 ()
0=H 1 ()

and

H 1 ()

[, ]1/2 , it follows from (5) that v


H 1 ()

C f

H 1 ()

87

2.
Now we return to the problem (1) with nonhomogeneous boundary condition u| = g. Suppose that Rn is a bounded domain of class C 1 . Then, by Theorem 10 (trace embedding theorem) the trace operator 0 (0 u = u| ) is continuous from H 1 () onto H 1/2 (): 0 : H 1 () H 1/2 (). Consider the problem u = F, x , 0 u = u| = g(x), for given F H 1 () and g H 1/2 (). We look for solution u H 1 (). Equation u = F in is understood in the sense of distributions: u(x) is a weak solution of (1), if u H 1 (), u(x) satises the identity u dx =

F dx,

C0 (),

and 0 u = g. Theorem 2 Let Rn be a bounded domain of class C 1 . Let F H 1 (), g H 1/2 (). Then there exists unique weak solution u H 1 () of problem (1). We have u Proof 1) By Theorem 11 (extension theorem), for g H 1/2 (), there exists extension G = P g H 1 () such that 0 G = g and G
H 1 () H 1 ()

H 1 ()

+ g

H 1/2 ()

(6)

C1 g

H 1/2 ()

(7)

If u H 1 () and 0 u = g. Then v = u G H 1 () and 0 v = 0. This is equivalent to the fact that v H 1 (). Function v is solution of the problem v=f v| = 0,

(8)

where f = F + G. From G H 1 () it follows that G H 1 () and G H 1 () C2 G H 1 () . Then f H 1 () and f


H 1

H 1

+ C2 G

H 1 ()

H 1

+ C1 C2 g

H 1/2

88

By Theorem 1, there exists unique solution v H 1 () of the problem (8), and v H 1 () C3 f H 1 . Then u = v + G is unique solution of the problem (1), and u
H1

H1

+ G
H 1 H 1

H1 H 1/2 () H 1/2 () .

C3 f

+ C1 g

C3 F

+ (C1 C2 C3 + C1 ) g

3. Dirichlet problem with spectral parameter


Now we consider the problem u = u + f (x), x u| = 0, with spectral parameter . Here is bounded. Denition Let Rn be arbitrary bounded domain. Let f H 1 (). A function u H 1 () satisfying identity u dx =

(9)

udx +

f (x)dx,

H 1 (),

(10)

is called a weak solution of problem (9). As before, we denote [u, ] =

u dx. This is inner product in H 1 ().

The form udx, u, is continuous sesquilinear form in H 1 (). By the Riesz theorem for such forms it can be represented as [Au, ], where A is a linear continuous operator in H 1 (). Obviously, udx = udx , so [Au, ] = [A, u] = [u, A], u, H 1 (). It follows that A = A . Next, [Au, u] = |u|2 dx > 0 if u = 0. So, A > 0. Lemma The operator A is compact operator in H 1 (). Proof

H 1 ()

This follows from the embedding theorem: H 1 () is compactly embedded in L2 (). 89

Well use the following property of compact operators: T is a compact operator in the Hilbert space H, if and only if for any sequence {uk } which converges weakly in H, the sequence {T uk } converges strongly in H.

operator J : H 1 () L2 () is compact, {uk } converges strongly in L2 (). We want to check that {Auk } converges strongly in L2 (). Since {uk } weak

Let {uk } be a weakly convergent sequence in H 1 (). Since the embedding

ly converges in H 1 (), it follows that uk H 1 () is uniformly bounded. A is a continuous operator; then also Auk H 1 () is uniformly bounded. We have [A(uk ul ), A(uk ul )] =

(uk ul )(Auk Aul )dx


L2 ()

u k ul
0

Auk Aul
C

L2 ()

as k, l .

{Auk } converges strongly in H 1 (). It follows that A is compact operator.

As before, the functional lf () =

unique element v H 1 () such that f dx = [v, ], H 1 (), and f H 1 () v H 1 () . Now, we can rewrite identity (10) in the form [u, ] = [Au, ] + [v, ] , which is equivalent to the equation u Au = v,

antilinear functional on H 1 (). By the Riesz Theorem, there exists

f dx (where f H 1 ()) is continuous

H 1 (),

(11)

(12)

where v H 1 () is given, and we are looking for solution u H 1 (). Thus, we reduced the problem (9) to the abstract equation (12) with compact operator A in the Hilbert space H 1 (). We analyse equation (12), using the properties of compact operators. The case v = 0 (which corresponds to f = 0): u = u x u| = 0 90

(13)

1 ) It is known that the spectrum of a compact operator is discrete: it consists of eigenvalues j , j N, that may accumulate only to point = 0; each eigenvalue is of nite multiplicity (i. e. , dim ker(A j I) < ). In our case A = A > 0, then all eigenvalues j are positive: j > 0. We enumerate eigenvalues in nonincreasing order counting multiplicities 1 2 . . . Then each eigenvalue corresponds to one eigenfunction uj : Auj = j uj , j N. Eigenfunctions {uj } are linearly independent. We have : j 0 as j . 1 Then for the eigenvalues j = j of the Dirichlet problem (13) we have the following properties: 0 < 1 2 . . ., j as j . Thus, we have the following theorem. Theorem 3 u Au = 0 Au = u (where = The spectrum of the Dirichlet problem (13) is discrete. There exists nontrivial solution only if = j , j N. All eigenvalues are positive and have nite multiplicities. The only accumulation point is innity: j as j . The case v = 0 (f = 0) u = u + f u| = 0 x

(14)

1 For compact operator A it is known that, if = j (= j ), j N, then the operator (I A)1 is bounded. We can nd unique solution

u Au = v

u = (I A)1 v, and u Since v H 1 () Theorem 4 f


H 1 ()

(I A)1
=C

H 1 () .

H 1 () ,

we arrive at the following theorem.

If {j }j ( is not eigenvalue), then for any f H 1 () there / exists unique (weak) solution u H 1 () of the problem (14), and u
H 1 ()

C f

H 1 () .

91

Now, suppose that = j , and v = 0 (f = 0). Then, solution of the equation u j Au = v exists, if v satises the solvability condition: vker(I j A). It means that v is orthogonal (with respect to the inner product [, ]) in
(k)

H 1 () to all eigenfunctions j , k = 1, . . . , p, corresponding to the eigenvalue j (here p is the multiplicity of j ). Since [v, ] = f (x)(x)dx, this solvability condition is equivalent to: f (x)j (x)dx = 0,
(k)

k = 1, . . . , p.

(15)
(k) p j=1 cj j

The solution u(x) is not unique, but is dened up to a summand with arbitrary constants cj . Theorem 5
(k)

If = j is eigenvalue of the Dirichlet problem, and j , k = 1, . . . , p, are corresponding (linearly independent) eigenfunctions, then problem (14) has solution for any f H 1 (), which satises the solvability conditions (15). Solution is not unique and is represented as
p

u = u0 +
j=1

cj j ,

(k)

where u0 is a xed solution, and cj are arbitrary constants.

4. HilbertSchmidt Theorem
Finally, we can apply the HilbertSchmidt Theorem for compact operators and obtain the following result. Let 1 2 . . . be eigenvalues of the Dirichlet problem. Here we repeat each j according to its multiplicity. There exists an orthogonal system of eigenfunctions {j }j : j j Aj = 0, j N,

By the HilbertSchmidt Theorem, {j }j is orthogonal basis in H 1 (), i. e. , for any F H 1 (),

F =
j=1

[F, j ] j [j , j ]

It is important that j l also in L2 (). Indeed, [Au, ] = udx (by 1 Denition of operator A). Next, Aj = j j (where j = j ). Thus, [Aj , l ] =

j l dx = j [j , l ] = 0, j = l. 92

[j , l ] = 0, j = l.

We have [F, j ] = j [F, Aj ] = j

F j dx, |j |2 dx.

[j , j ] = j [Aj , j ] = j

Then [F, j ] = [j , j ]
F j dx 2 |j | dx

(F, j )L2 () j
2 L2 () 2 L2 ()

, and

F =
j=1

(F, j )L2 () j

j .

The last fomula can be extended to all F L2 (). Theorem 6 Let Rn be a bounded domain. Then there exists an ortogonal system of eigenfunctions {j }j of the Dirichlet problem. This system forms an orthogonal basis in L2 () and in H 1 () (with respect to the inner product [, ]).

93