Sunteți pe pagina 1din 25

Fourier and Sobolev

Jorge Aarao
June 2005
These notes are a companion to the lectures I will give in Las Cruces,
preparatory for the CBMS conference lectures by Terry Tao. I will not cover all
the material in these notes; in particular we will spend little time on sections
14. Please read these notes beforehand, concentrating on sections 5, 7, 8, 10,
11, 12, and 13. Most proofs are omitted from this text, since the objective here
is to get you acquainted with this material from a users perspective. On the
other hand, except for the proofs of the main theorems, most proofs are quite
straightforward, and you may try your hand at them yourself. Or, consult any
book on this subject; I recommend
. Topics in functional analysis and applications, by S. Kesavan;
. Partial dierential equations, by L.C. Evans;
. Analysis, by Lieb and Loss.
1
1 Introduction
Relatively speaking, Algebra is a new branch of mathematics (as compared to
Geometry, Analysis, and Number Theory), and since the emergence of Algebra
mathematicians have tried to use its power to serve the other areas. It can
be argued that the material we will cover here is part of that program, trying
to view spaces of functions in a more algebraic way, in order to solve partial
dierential equations. Here is a parallel: we want to solve x
2
1 = 0, and the
solutions are real numbers. But what of x
2
+ 1 = 0? We have two options.
We can declare that the equation has no solutions, or we can declare that the
equation has solutions, only these solutions are not real numbers. With that
we create a new number system, dene its (algebraic) rules, and show that the
old number system (R) is somehow included in the new one (C). Likewise with
dierential equations. Sometimes, when we take the Fourier transform of a
dierential equation, we may obtain that its solution u is such that, say, u = 1,
an equation that is not satised by any function. We have two options. We can
declare that the dierential equation has no solutions, or we can declare that it
has solutions, only those solutions are not functions anymore. It is then our job
to provide a framework for that new space of solutions, complete with algebraic
structure, and making explicit how to include the old structure into the new
structure. We start from the beginning.
2 Smooth functions, and L
p
functions
We let be a non-empty open set of R
n
; K will always denote a compact
set. We dene the multi-index = (
1
, . . . ,
n
) to be an n-tuple of non-
negative integers (n is the dimension of R
n
!). The order of is the integer
[[ =
1
+ +
n
. If x = (x
1
, . . . , x
n
) R
n
, we write x

= x
1
1
. . . x
n
n
, and
D

=

||
x
1

1
. . . x
n

n
.
A function u : C is said to be smooth if all of its partial derivatives,
of any order, exist and are continuous. This space is denoted by C

(). A
smooth function u : R
n
C is said to be in the Schwartz space o(R
n
) if for
any multi-indices and there is a constant C = C(u, , ) such that
sup
xR
n

u(x)

< C.
Exercise 1 Show that f(x) = exp([x[
2
) is in o(R
n
).
A function f : C is said to be compactly supported if there is
a compact K (where K may depend on f) such that f(x) = 0 for all
x K. The support of f is the intersection of all compact sets K as above,
and is denoted by supp f. The set of all such smooth, compactly supported
functions (dened over the same ) is denoted by C

c
().
2
Exercise 2 Show that C

c
() contains more than one function.
From now on, we let p be a real number with 1 p < ; on occasion we
will allow p = as well.
For 1 p < , the space L
p
() consists of those functions f : C such
that
(|f|
p
)
p
=
_

[f(x)[
p
dx < .
The equality denes the symbol |f|
p
, the norm of f in L
p
(). The integral
we used is the Lebesgue integral.
If two functions f and g in L
p
() are such that |f g|
p
= 0, then we just
identify f and g. In that situation it can be shown that there is a set A ,
of measure zero, such that f(x) = g(x) for x A; in this case we say that
f(x) = g(x) almost everywhere (a.e.).
When p = we dene the space L

() as consisting of those functions f :


C which are equal almost everywhere to a bounded function g; equivalently
there is a constant C = C(f) and a set A of measure zero such that
|f|

= sup
x\A
[f(x)[ < C.
Exercise 3 Let be bounded, containing the origin; let 1 p < . Show that
f(x) = [x[
k
is in L
p
() if and only if k > n/p.
Exercise 4 Show that if 1 p < , then C

c
() is dense in L
p
(). What
happens if p = ?
Exercise 5 Show that if 1 p , then o(R
n
) is contained in L
p
(R
n
).
Exercise 6 Let 1 p < q . Show that, in general, neither of L
p
() and
L
q
() is contained in the other. On the other hand, show that if q < , then
L
p
() L
q
() is dense in both. What happens if q = ?
Exercise 7 Suppose is bounded, and that 1 p < q < . Show that L
q
()
is contained in L
p
(). Let i(f) = f be the inclusion map. Is it continuous?
That is, is there a constant C (independent of f) such that |f|
p
C|f|
q
?
Prove, or give counter-examples. What happens if q = ?
Each L
p
-space comes with a local version. Let 1 p < . We say that f
belongs to L
p
loc
() if for each compact K in there is a constant C = C(K, f)
such that
_
K
[f(x)[
p
dx < C. Clearly L
p
() is contained in L
p
loc
().
Exercise 8 If 1 p < q < , then L
q
loc
() L
p
loc
().
The above result shows that L
1
loc
() is, in a sense, the most important of
the local integrability spaces. We will return to this space in the future.
Exercise 9 Is it true that

p>1
L
p
loc
() = L
1
loc
()?
3
We have the following notions of convergence in L
p
loc
(): we say that f
m
converges strongly to f in L
p
loc
() if and only if for every compact K in
we have that f
m
converges to f in L
p
(K) (in the usual L
p
sense).
We say that f
m
converges weakly to f in L
p
loc
() if and only if for every
compact K in f
m
converges to f weakly in L
p
(K). In practical terms, this
means that if 1 < p, then
_
K
f
m
(x) g(x) dx
_
K
f(x) g(x) dx
for every g in L
p

(K) (where
1
p
+
1
p

= 1), and if p = 1, then the same integral


converges for every bounded g.
Exercise 10 Let f
m
(x) = sin(mx), f(x) 0. Show that f
m
converges weakly,
but not strongly, to f in L
2
loc
(R).
3 Some tools of the trade
Let 1 < p < , and let p

be the conjugate exponent of p, that is,


1
p
+
1
p

= 1.
If f is in L
p
() and g is in L
p

(), then fg is in L
1
(), and we have Holders
inequality
_

[f(x) g(x)[ dx
__

[f(x)[
p
dx
_1
p
__

[g(x)[
p

dx
_ 1
p

,
or simply |fg|
1
|f|
p
|g|
p
.
If 1 p , and both f and g are in L
p
(), Minkowskis inequality
states that
|f + g|
p
|f|
p
+|g|
p
.
Now let be R
n
, and let p and p

be conjugate exponents. Let f be in


L
p
(R
n
), and g be in L
p

(R
n
). Note that, if x R
n
is xed, then the function
y f(x y) is such that |f|
p
= |f(x )|
p
.
The convolution of f and g, given by
(f g)(x) =
_
R
n
f(x y)g(y) dy,
is well-dened, and by Holders inequality we have |f g|

|f|
p
|g|
p
.
Observe that f g = g f.
Exercise 11 Show that supp (f g) supp f + supp g, where the sum of two
sets is dened in the standard way. In particular, if f and g are compactly
supported, then so is f g.
Exercise 12 Show that if f L
p
(R
n
) and g L
p

(R
n
), then fg is continuous.
Is it true that lim
|x|
(f g)(x) = 0?
4
The convolution of f and g is dened for every x R
n
under the above
hypotheses. We may extend this denition in many ways, and pay only a small
price.
Exercise 13 If f and g are in L
1
(R
n
), then |f g|
1
|f|
1
|g|
1
. In particular,
(f g)(x) is nite a.e. Is it true that lim
|x|
(f g)(x) = 0?
And here is Youngs inequality:
Theorem 1 If f L
p
(R
n
), g L
q
(R
n
) and
1
p
+
1
q
= 1 +
1
r
, then
|f g|
r
|f|
p
|g|
q
.
In particular, if q = 1 then |f g|
p
|f|
p
|g|
1
.
One of the nice things about the convolution is that it inherits the good
dierentiability properties of f and g.
Exercise 14 Let 1 p . If f L
p
(R
n
) and g o(R
n
), then f g
C

(R
n
), and D

(f g) = f D

g. In addition, if D

f is in L
p
(R
n
), then
D

(f g) = (D

f) g.
An approximation of the identity, or mollier, is a one-parameter fam-
ily of functions
t
: R
n
R with the following properties:
1) For all t > 0 we have
_
R
n

t
(x) dx = 1;
2) For all > 0
t
converges uniformly to zero outside a ball of radius
centered at the origin, as t goes to zero;
3)
t
(x) 0 for all x.
For example, if 0 is a radially symmetric function decaying to zero as x
approaches innity, and with total area equal to 1, then
t
(x) =
1
t
n
(x/t) is a
mollier. More specically, we may take (x) = exp([x[
2
/2)/(2)
n/2
, and in
many cases we want to take to be in C

c
().
Theorem 2 Let 1 p < , let f belong to L
p
(R
n
), and let
t
C

(R
n
) be a
mollier. Then f
t
= f
t
converges to f in L
p
(R
n
), as t goes to zero.
The result in the above theorem is very useful; for instance, you might want
to revisit some of the previous exercises in light of this theorem. On the other
hand, approximation in L
p
in principle only tells us that for some sequence
t
m
0 we have f
t
m
(x) f(x) a.e. It can be shown that more is true, namely
that f
t
(x) f(x) a.e. (if the mollier is nice), but we wont do it here.
Exercise 15 Show that if f is continuous and bounded, then f
t
converges to f
uniformly on compact sets, as t goes to zero. In particular f
t
(x) converges to
f(x), for every x.
5
4 The Fourier Transform
Let f be a function in L
1
(R
n
). The Fourier Transform of f is a function

f : R
n
C, given by

f() =
_
R
n
f(x) e
2ix
dx.
Here x is the usual dot product in R
n
. Notice that

f() is dened for every ,
and in fact |

f |

|f|
1
. So the Fourier Transform is a continuous operator
from L
1
(R
n
) to L

(R
n
).
Exercise 16 (Riemann-Lebesgue Lemma) If f L
1
(R
n
), then

f is con-
tinuous, and lim
||

f() = 0.
Exercise 17 Let 1
[1,1]
be the function that is equal to 1 if 1 x 1, and
zero otherwise. Compute its Fourier transform, and conclude that

1
[1,1]
is not
in L
1
(R).
Theorem 3 For t > 0 let g
t
(x) = exp(t[x[
2
), for x R
n
. Then g
t
() =
t
n/2
exp([[
2
/t).
Observe that g
t
in Theorem 3 is a mollier.
Theorem 4 If f and g are in L
1
(R
n
), then

(f g)() =

f() g().
Theorem 5 If f is in o(R
n
), then

f() = (2i)
||


f(), D

f() = (2i)
||

x

f().
As a consequence

f is in o(R
n
).
Theorem 6 (Fourier Inversion Theorem) If f is in o(R
n
), then for all x
f(x) =
_
R
n

f() e
2ix
d.
Proof : With the same g
t
as in Theorem 3, and using the result of Exercise 15,
we have
_
R
n

f() e
2ix
d =
_
R
n
__
R
n
f(y) e
2iy
dy
_
e
2ix
d
= lim
t0
+
_
R
n
__
R
n
f(y) e
2iy
dy
_
e
2ix
e
t||
2
d
= lim
t0
+
_
R
n
f(y)
__
R
n
e
t||
2
e
2i(yx)
_
dy
= lim
t0
+
_
R
n
f(y)t
n/2
e
|xy|
2
/t
dy
= lim
t0
+
(f g
t
)(x) = f(x).
6

Exercise 18 (Fourier Inversion Theorem, 2) Show that if f is in L


1
(R
n
),
then for almost every x we have
f(x) = lim
t0
+
_
R
n

f() e
2ix
e
t||
2
d.
Theorem 7 If f and g are in o(R
n
), then
_
R
n
f(y) g(y) dy =
_
R
n

f(y) g(y) dy.


In particular |f|
2
= |

f|
2
.
Theorem 8 Let f
m
be a sequence in o(R
n
), let f be in L
2
(R
n
), and suppose
that |f
m
f|
2
0. Then

f
m
is a Cauchy sequence in L
2
(R
n
), and therefore
has a limit in L
2
(R
n
). We extend the denition of the Fourier Transform to
L
2
(R
n
) by dening

f to be this limit. From the results of previous theorems and
exercises the Fourier Transform is an invertible isometry from L
2
(R
n
) to itself.
Moreover

f() = lim
R
_
|x|R
f(x) e
2ix
dx,
f(x) = lim
R
_
||R

f() e
2ix
d,
where both limits are in the L
2
sense.
It can be shown (but it is much harder) that the equations above hold for
almost every x.
Exercise 19 Let 1 < p < 2. Show that any function f in L
p
(R
n
) may be
written (non-uniquely) as f = f
1
+ f
2
, with f
j
in L
j
(R
n
), j = 1, 2. Dene

f =

f
1
+

f
2
. Show that

f is well dened (that is, its denition is independent
of the particular way of splitting f into f
1
and f
2
).
The previous exercise shows that it is possible (and easy) to dene the Fourier
Transform over any L
p
(R
n
), 1 p 2. It is possible to obtain a better (and
harder) result, namely that the Fourier Transform maps L
p
(R
n
) continuously
into L
p

(R
n
), where 1 p 2 and p

is the conjugate exponent of p.


What if p > 2? The following is a standard example.
Exercise 20 Let t = a+ib, with a > 0, and dene g
t
(x) = exp(t[x[
2
). Show
that if p > 2 and 1 < q < 2, then | g
t
|
q
/|g
t
|
p
is not bounded (in t).
As a consequence the Fourier Transform cannot map L
p
(R
n
) boundedly into
any L
q
(R
n
), if 1 < q < 2 < p. As mentioned in the Introduction, we can give up
7
now, or we can try to enlarge the domain of denition of the Fourier Transform
(and its range, as well).
Lets look at an application of the Fourier Transform. Suppose f is smooth
and compactly supported, and we want to solve the equation u = f on R
n
,
n 3. Assuming the solution u is a well-behaved function, we take the Fourier
Transform of both sides to obtain [2[
2
u =

f, and so u =

f/[2[
2
. Because
f is in o(R
n
), we know that

f is also in o(R
n
), and as a consequence the function

f/[2[
2
is in L
1
(R
n
) (n 3!). Consequently, we may now dene u to be
u(x) =
_
R
n

f()
[2[
2
e
2ix
d,
and go on to show that this u indeed solves the problem u = f. But let me
call your attention to Theorem 4. Since u =

f/[2[
2
, it is natural to think
that the solution u will be given by u = f G, where G is a function such that

G = 1/[2[
2
, except that

G is not in any L
p
(R
n
). Again, we can either give
up, or . . .
5 Distributions
Here is what we want to accomplish: we want to enlarge the spaces L
p
() to
include objects that are not necessarily in L
p
(). We wrote objects because in
principle they may not even be functions but we want our old L
p
-spaces to be
naturally included in this new space. The motivation we presented in the last
section was that this new space would allow us to broaden the denition of the
Fourier Transform, but we will accomplish much more than that. Here is some
extra motivation.
Given any f in L
2
() the mapping
_

f dx is a bounded linear func-


tional. The converse is true.
Theorem 9 (Riesz Representation) If T : L
2
() C is a bounded linear
functional, then there is a unique f L
2
() such that
T() =
_

(x)f(x) dx
for every in L
2
().
So we can think of the bounded linear functionals on L
2
() as being exactly
the very functions in L
2
().
Exercise 21 Take = [0, 1] R. Show that if is in L
3
([0, 1]), then it is in
L
2
([0, 1]), and ||
2
||
3
. Show that there is a function f not in L
2
([0, 1])
such that T() =
_
1
0
(x)f(x) dx is a bounded linear functional on L
3
([0, 1]).
8
What this exercise shows is that by restricting our function space (from
L
2
to L
3
), we enlarged our dual space (that is, the space of bounded linear
functionals). But in any case this enlarged space of functionals still contained,
in a sense, the original space L
2
. So here is the idea: we will restrict the
function space as much as possible, and correspondingly enlarge the dual space.
The elements of this enlarged dual will be our distributions. We give yet another
example, then move to the theory itself.
Exercise 22 Let H = C

c
(R
n
) L
2
(R
n
) be endowed with the following norm:
if is in H, then ||
2
H
=

||1
|D

|
2
2
. Show that there is a bounded linear
functional T on H (that is, [T[ C||
H
) that is not representable by any
function f in L
2
(R
n
).
A function : C will be called a test function if it belongs to C

c
().
We introduce a notion of convergence on this space: A sequence of test functions

m
will converge to a test function if and only if there is some xed compact
set K in such that the supports of all
m
and of are all contained in K,
and moreover for every multi-index we have that D

m
converges uniformly
(in K) to D

. With this notion of convergence, the space of test functions is


denoted by T().
Notice that T() is contained in every space L
p
(), and moreover, if
m
converges to in T(), then
m
converges to in L
p
() as well. (And all of
its derivatives converge in L
p
() as well, to the corresponding derivative of .)
As a consequence we expect that, in some sense, all spaces L
p
() should be
contained in the dual of T(), and this inclusion map should be continuous.
The (topological) dual of T() is the space of distributions, and is denoted
by T

(). Its elements (the distributions themselves) are continuous linear


functionals T : T() C; that is:
1) T(
1
+
2
) = T(
1
) + T(
2
),
2) T() = T(),
3) T(
m
) T() whenever
m
in T().
Exercise 23 (Dirac delta-function) If x
0
, show that
x
0
() = (x
0
) is
a distribution. Show that there is no function f in L
1
loc
() such that for all
in T()

x0
() =
_

(x) f(x) dx.


(Since L
1
loc
contains all other types of L
p
, we are saying that this distribution
cant be realized as an integrable function.) If x
0
= 0, we simply write =
0
.
Exercise 24 If x
0
and is any multi-index, show that T
x
0
,
() = D

(x
0
)
is a distribution.
Exercise 25 Let f be in L
1
loc
(). Show that T
f
() =
_

fdx is a distri-
bution.
9
Theorem 10 If f and g are in L
1
loc
() and they determine the same distribu-
tions, then f(x) = g(x) a.e.
Functions in L
1
loc
are not the most general objects that can be viewed as
distributions.
Exercise 26 Let be a measure with bounded total variation. Then T

() =
_
d denes a distribution.
We say that the distributions T
m
converge to a distribution T if for each
in T() we have T
m
() T().
Exercise 27 Show that the map f T
f
from L
1
loc
() to T

() is continuous,
that is, if f
m
f weakly, then T
f
m
T
f
.
Exercise 28 Let
t
be a mollier. Show that T
t
converges to .
Next we dene what we mean by the derivative of a distribution. The
denition is such that the formula for integration by parts is still valid.
Let T be a distribution, and be a multi-index. The derivative D

T is
itself a distribution, given by
(D

T)() = (1)
||
T(D

).
The key to understanding this denition is to see that the integration by
parts formula still works in some sense, as we pointed out before. So, for in-
stance, if f is a locally integrable function such that f

is also a locally integrable


function, then for any in T(R) we have

_
R
f

dx =
_
R
f

dx.
Now, if f

is not locally integrable, the left hand side makes no sense, but the
right hand side still does. So we extend the meaning of the left hand side:
instead of f

we write (T
f
)

(the derivative of the distribution determined by f),


and the formula becomes
(T
f
)

() =
_
R
f

dx = T
f
(

).
Exercise 29 Let f be a smooth function on . Justify the equality
(1)
||
_

(D

) f dx =
_

(D

f) dx,
and thus verify that D

T
f
= T
D

f
.
Exercise 30 Show that D

T is indeed a distribution; that is, show it is linear


and continuous.
10
Exercise 31 Working over R, let H(x) be 1 if x 0, and 0 if x < 0. (H is
called the Heaviside function.) Let T
H
be the distribution determined by H.
Show that the derivative of the distribution T
H
is the delta function
0
.
From now on we will abandon the cumbersome notation T
f
to represent
the distribution determined by the function f, and simply write f, unless it
is imperative to distinguish between the two. We will also write distributional
derivative to indicate that the derivative is being viewed as a distribution. The
previous exercise, rephrased, says that the distributional derivative of H is
0
.
Next we want to dene the Fourier Transform of a distribution, but it turns
out that not every distribution will have a Fourier Transform. The reason lies
in the following result.
Exercise 32 If f and g are in o(R
n
), then
_
R
n

f(z) g(z) dz =
_
R
n
f(z) g(z) dz.
Interpreted in terms of distributions determined by functions, this exercise wants
to say that if f is a distribution, then

f is a distribution given by

f(g) = f( g).
This would indeed be the case, except for the fact that we cant have both g and
g in T(R
n
) (by Heisenbergs principle). So instead we look at distributions that
are dened on o(R
n
); those will be called tempered distributions. Here is how
it is done.
Dene the following notion of convergence in o(R
n
): f
m
o(R
n
) converges
to f o(R
n
) if and only if for any compact set K and any multi-index
the sequence D

f
m
converges uniformly to D

f in K. With this notion of


convergence the inclusion T(R
n
) o(R
n
) is continuous. We denote by o

(R
n
)
the (topological) dual space of o(R
n
), that is, continuous linear functionals
f : o(R
n
) C. The elements of o

(R
n
) are called tempered distributions;
they are distributions since there is a continuous inclusion o

(R
n
) T

(R
n
).
Let T be a tempered distribution. We dene the Fourier Transform of T
as being the (tempered) distribution

T given by

T() = T(

),
for all in o(R
n
).
Exercise 33 Show that

T is a tempered distribution.
Thanks to Exercise 32, we know that

T
f
() = T
f
(

), showing that the


denition of

T
f
is compatible with the denition of

f.
Exercise 34 Show that the delta function
0
is a tempered distribution, and
that

0
= 1 (the distribution determined by the constant function 1). Also,

1 =
0
.
11
Exercise 35 Show that if f is locally integrable, and grows no faster than a
polynomial at innity, then T
f
is a tempered distribution.
Exercise 36 Give an example of a distribution that is not a tempered distribu-
tion.
Exercise 37 Show that if T is a tempered distribution, then D

T is a tempered
distribution.
6 A bit of fun
In this short section lets pretend we are Euler. Say f : R C is a 2-periodic
function (it is your job to put extra hypotheses here). Then

f() =
_
R
f(x) e
2ix
dx.
Of course, the integral is not convergent, but that never stopped Euler, and it
shouldnt stop us. We compute

f() =

j=
_
2(j+1)
2j
f(x) e
2ix
dx
=

j=
_
2
0
f(y) e
2iy
dy e
2i(2j)
=
_
2
0
f(y) e
2iy
dy

j=
e
2i(2j)
= g()

j=
e
2i(2j)
.
Well, g is a function, and we want to know what that sum is. Writing = 4
2
,
we get

j=
e
ij
= lim
r1

j=
r
|j|
e
ij
.
Since r approaches 1 from below, we can actually sum the last series (it is the
sum of two convergent geometric series).
Exercise 38 Show that

j=
r
|j|
e
ij
=
1 r
2
1 2r cos + r
2
= 2P
r
(),
where the last equality denes P
r
, the Poisson kernel. Show that P
r
has the
following properties:
12
1. P
r
() 0;
2. P
r
is 2-periodic in ;
3.
_

P
r
() d = 1;
4. For all (0, ), P
r
converges uniformly to zero on the set < [[ < ;
5. P
r
(0) as r 1

.
In other words, as r goes to 1, P
r
behaves like an approximation of the
identity on each interval of the form [2j , 2j + ]. Therefore

j=
e
ij
= 2

j=
( 2j).
Thus we obtain

f() = 2 g()

j=
(4
2
2j).
Using the inversion formula, we have
f(x) =
_
R

f() e
2ix
d
=

j=
2
_
R
g()(4
2
2j) e
2ix
d
=

j=
1
2
_
R
g
_
z + 2j
4
2
_
(z) e
ixj
e
ixz
2
dz
=

j=
1
2
g(j/2) e
ixj
.
In other words, we obtain the Fourier series expansion of f, by way of the
Fourier transform of f, and distributions.
Exercise 39 Justify the steps above.
7 Sobolev Spaces
Some distributions are representable by functions, and when that happens the
function is uniquely dened a.e., as we saw. We are now concerned with func-
tions whose distributional derivatives are representable by functions. Here is an
example.
Exercise 40 On R, let f(x) = max0, x. Show that the distributional deriva-
tive f

is representable by the Heaviside function.


13
When a distributional derivative D

T is representable by a function in L
p
()
we will simply write D

T L
p
(), identifying the distribution with its repre-
sentation, and denoting both by the same symbol. In this situation the symbol
|D

T|
p
will denote the L
p
-norm of the function which represents D

T.
Let 1 p . The Sobolev space W
k,p
() consists of those functions f
in L
p
() such that all the distributional derivatives of f of order at least k are
also in L
p
(), or
W
k,p
() = f L
p
() ; D

f L
p
() for all [[ k .
For 1 p < , we put the following norm on W
k,p
():
|f|
W
k,p
()
=
_
_
_

||k
|D

f|
p
p
_
_
_
1/p
.
We have that T() W
k,p
() L
p
(), and |f|
p
|f|
W
k,p
()
. Moreover,
the rst inclusion (as well as the second) is continuous.
Exercise 41 (Alternative denition) Denote by H the subspace of L
p
()
consisting of smooth functions f such that for all multi-indices we have D

f
also in L
p
(). Endow H with the norm |f|
H
= (

||k
|D

f|
p
p
)
1/p
. Show
that the completion of H in this norm is W
k,p
().
We may also dene local Sobolev spaces W
k,p
loc
(); for these spaces we
have notions of convergence (both strong and weak), but not a norm.
For 1 p < , the space W
k,p
() has an important subspace, denoted by
W
k,p
0
(). This space is the completion of T() in W
k,p
() with respect to the
norm of W
k,p
(). Hence f W
k,p
() is in W
k,p
0
() if and only if there is a
sequence of functions f
m
in T() such that |f
m
f|
W
k,p
()
0.
Exercise 42 If 1 p < and ,= R
n
, then W
k,p
0
() W
k,p
().
Theorem 11 If 1 p < , then W
k,p
0
(R
n
) = W
k,p
(R
n
).
Proof : We want to show that for any f W
k,p
(R
n
) there is a sequence f
m
in
T(R
n
) such that f
m
f in W
k,p
(R
n
).
It is easy to obtain a sequence of smooth functions in L
p
with all the con-
vergence properties we want, except that they are not compactly supported.
Indeed, let
t
be a smooth mollier; then
t
is in L
1
(R
n
), and by Youngs in-
equality f
t
= f
t
is in L
p
(R
n
). Moreover, for all multi-indices with [[ k
we have D

f
t
= (D

f)
t
L
p
(R
n
). Using standard properties of molliers,
f
t
f and D

f
t
D

f in L
p
(R
n
), so that f
t
converges to f in W
k,p
(R
n
). As
mentioned before, these f
t
are not compactly supported.
14
Now pick a function T(R
n
) such that 0 (x) 1 for all x, (x) 1
if [x[ 1, and (x) 0 for [x[ 2. Dene g
t
(x) = f
t
(x)(tx). For any multi-
index we have D

g
t
(x) = D

f
t
(x) if [x[ 1/t, and D

g
t
(x) 0 if [x[ 2/t.
Thus, with [[ k,
|D

f
t
D

g
t
|
p
p
=
_
|x|>1/t
[D

f
t
(x) D

g
t
(x)[
p
dx.
Now comes the technical part of the proof. We may write D

g
t
in the form
D

g
t
=

||||

,t
(x) D

f
t
(x) = (tx) D

f
t
+

||<||

,t
(x) D

f
t
(x),
where the functions
,t
depend on as well. For example, if = (1, 0, 1, 1),

1
= (0, 1, 0, 0), and
2
= (1, 0, 1, 0), then

1
,t
(x) 0, and

2
,t
(x) = t

x
4
(tx).
In any case, it is clear that the functions
,t
(and there are nitely many of
them) are all bounded by the same constant, and in fact since t will be going
to zero, they will converge to zero as well. Thus when we write
D

f
t
(x) D

g
t
(x) = D

f
t
(x) (1 (tx))

||<||

,t
(x) D

f
t
(x),
we obtain, for some constant C which depends only on p, that
|D

f
t
D

g
t
|
p
p
C
_
|x|>1/t
2
p
[D

f
t
(x)[
p
+

||<||
[
,t
(x)[
p
[D

f
t
(x)[
p
dx.
As t goes to zero, this integral goes to zero.
8 H
k
(), particularly H
k
(R
n
)
The case p = 2 is special, since W
k,2
() is a Hilbert space; for that special case
we use the notation W
k,2
() = H
k
().
Now let = R
n
. Thanks to the fact that the Fourier Transform is an
isometry on L
2
(R
n
), we have that
|D

f|
2
= |

f|
2
= (2)
||
|

f|
2
,
and therefore
|f|
2
H
k
(R
n
)
=

||k
(2)
||
_
R
n
[

[
2
[

f()[
2
d
=
_
R
n
_
_

||k
(2)
||
[

[
2
_
_
[

f()[
2
d.
15
Exercise 43 There are constants M
1
and M
2
, depending only on n and k, such
that
M
1
(1 +[[
2
)
k

||k
(2)
||
[

[
2
M
2
(1 +[[
2
)
k
.
Theorem 12 (Equivalence of norms) The norm
__
R
n
(1 +[[
2
)
k
[

f()[
2
d
_
1/2
is equivalent to the norm of H
k
(R
n
).
Henceforth we will use both norms interchangeably, and all bounds will be
up to a multiplicative constant when translating into the other norm.
Theorem 13 (Transform characterization of H
k
) Suppose that f L
2
(R
n
),
and that _
R
n
(1 +[[
2
)
k
[

f()[
2
d
is nite. Then f is in H
k
(R
n
).
Proof : For each multi-index with [[ k we consider the distributional
derivative D

f. From the hypotheses, |



f|
2
is nite, and so there is a function
g L
2
(R
n
) such that g = (2i)
||


f. We claim that T
g
= D

f:
(D

f)() = (1)
||
T
f
(D

) = (1)
||
_
R
n
f(x) D

(x) dx
= (1)
||
_
R
n

f()

() d = (1)
||
_
R
n

f()(2i)
||

() d
=
_
R
n
g()

() d =
_
R
n
g(x) (x) dx.
This theorem suggests the denition of Sobolev spaces for non-integer in-
dices. We dene, for s 0,
H
s
(R
n
) = f L
2
(R
n
) ; (1 +[[
2
)
s/2

f() L
2
(R
n
) ,
with norm
|f|
H
s
(R
n
)
=
__
R
n
(1 +[[
2
)
s
[

f()[
2
d
_
1/2
.
Notice that s < t implies that H
s
H
t
. We will come back to these spaces
later.
16
9 A word from the wise
At this point I cant do better than to give you the gist of what is on pages
5657 in Kesavans book Topics in Functional Analysis and Applications.
Approximation theorems. The idea is to prove results for smooth functions,
and obtain these results for Sobolev functions via density, so one object of study
is how to approximate functions in W
k,p
() by smooth functions. We got a taste
of that in Theorem 11.
Extension theorems. Some results are easier to prove over the whole of R
n
;
to prove them for extend the setting from to R
n
, use the result there, then
restrict it back again to . For this to work we need to know if we can perform
the extension, and if the extension, as a map from W
k,p
() to W
k,p
(R
n
), is
continuous.
Imbedding theorems. It turns out that many Sobolev spaces are naturally
contained in other well-known spaces, the reason being that the more integra-
bility we require for the derivatives of a function f, the smoother f should
be.
Compactness theorems. Even better, some of the inclusions of Sobolev spaces
into other spaces are compact (in the functional analysis sense).
Trace theory. Sobolev spaces are often the natural spaces where to nd so-
lutions to partial dierential equations, but these equations often come with
boundary data. It becomes important to understand what is meant by restrict-
ing a Sobolev function to the boundary of the domain. This restriction is called
the trace of the function, and the goal of trace theory is to give meaning to
this restriction.
10 A warm-up
From now to the end our guiding light is the following principle: If a function
f has many distributional derivatives in L
p
(that is to say, if f is in W
k,p
for
large k), then f itself has to have a decent degree of smoothness (in the classical
sense). As a warm-up we have a famous result.
Theorem 14 Let I R be an open interval, and suppose f W
1,1
(I). Then
f is absolutely continuous.
Proof : We should have written that there is an absolutely continuous function
g such that g(x) = f(x) a.e., but we will gloss over such technicalities, unless
they are really necessary. So x x
0
I and dene
v(x) =
_
x
x
0
f

(x) dx,
which is absolutely continuous by denition. The Fundamental Theorem of Cal-
culus asserts that v

(x) = f

(x) a.e., and therefore in the sense of distributions


17
(f v)

= 0. After you do the next exercise, we will be able to conclude that


f(x) = v(x) + c a.e., for some constant c.
Exercise 44 If T is a distribution on T(I), and T

= 0, then T = T
c
for some
constant c.
Exercise 45 Let I R be an open interval, and suppose f W
k,1
(I). Then
f C
k1
(I), and f
(k1)
is absolutely continuous.
A consequence of Theorem 14 is that W
1,1
(I) C(I), the continuous func-
tions on I. What sort of inclusion is that? Continuity of this inclusion would
mean that |f|

C |f|
W
1,1
(I)
, but for this to make sense wed better take I
to be a bounded interval. In that case W
1,p
(I) W
1,1
(I) for 1 p < , so
lets look at the inclusion j : W
1,p
(I) C(I). Fix y I. For any f W
1,p
(I)
we have
f(x) = f(y) +
_
x
y
f

(t) dt,
and so
[f(x)[ [f(y)[ +
_
I
[f

(t)[ dt.
Integrating in y over I, and with [I[ denoting the length of I, we obtain
[I[ [f(x)[
_
I
[f(y)[ dy +[I[
_
I
[f

(t)[ dt.
For p = 1 we immediately obtain, for a constant C depending only on I,
that |f|

C|f|
W
1,1
(I)
. For p > 1 we rst apply Holders inequality to the
right-hand-side integrals with p and p

to obtain |f|

C|f|
W
1,p
(I)
, with a
constant C that depends on I and p. In any case the inclusion j is continuous.
But we can say more about this inclusion. Look at the unit ball in W
1,p
(I),
B = f W
1,p
(I) ; |f|
W
1,p
(I)
1 .
From the continuity of j we conclude that j(B) is a uniformly bounded set in
C(I). Now, if p > 1 then [f(x) f(y)[ |f|
W
1,p
(I)
[x y[
1/p

C[x y[
1/p

.
We conclude that j(B) is equicontinuous in C(I). Applying the Arzela-Ascoli
Theorem, we have just proved the following:
Theorem 15 If 1 < p < , then the inclusion j : W
1,p
(I) C(I) is compact.
Thus any sequence f
m
bounded in W
1,p
(I) has a subsequence f
m
l
which
converges uniformly to a function f C(I). Pretty neat.
If p = 1 the result is not true, as the example f
m
(x) = x
m
on I = (0, 1)
shows. Here equicontinuity fails to hold. Also, notice that for p > 1 this
sequence f
m
is not bounded in W
1,p
(I).
Notice another by-product of Theorem 14: If I = (a, b) and f W
1,1
(I),
then it makes sense to talk about f(a) and f(b). This is an example of a result
18
in trace theory. A consequence of this result is that f may be extended from
I to the whole line; if we extend it carefully, the extension procedure will give
us a continuous map from W
1,p
(I) to W
1,p
(R). Here is how this can be done.
Given f W
1,p
(I), I = (a, b), we dene an extension

f W
1,p
(R) as follows:

f(x) =
_

_
0, x a 1;
u(a) (x a + 1), a 1 x a;
u(x), a x b;
u(b) (b + 1 x), b x b + 1;
0, x b + 1.
Exercise 46 Show that f

f denes a continuous map W
1,p
(I) W
1,p
(R).
Exercise 47 Show that extension operators as dened above are not unique.
Finally, here is another way of looking at Theorem 14. It says that if we
start with a function f that has a certain level of integrability p > 1, and for
which f

also has that level of integrability, then f indeed has a higher level of
integrability q > p. (But the theorem says nothing about improved integrability
for f

.) This point of view will be a feature of the theory, that is, a function
with a certain integrability p, and whose derivatives up to order k also have
that integrability, actually has a level of integrability q > p. This q depends on
many factors, including p, k, and the dimension n. So there is a trade-o: If
we are willing to give up some dierentiability of f, we may view f as being in
a higher integrability space, in the sense that if f W
k,p
, then f W
0,q
for
some q > p.
11 Imbedding theorems
In contrast with other sections, here we state the main results rst. In what
follows, if 1 p < n, we dene the exponent p

by
1
p

=
1
p

1
n
or p

=
np
n p
,
and observe that p

> p.
Theorem 16 (Sobolevs inequality) Let 1 p < n. Then the inclusion
W
1,p
(R
n
) L
p

(R
n
) is continuous. In other words, there is a positive constant
C = C(p, n) such that for any f W
1,p
(R
n
) we have
|f|
p
C |f|
W
1,p
(R
n
)
.
Theorem 17 If p = n and q n, then W
1,n
(R
n
) is contained in L
q
(R
n
).
19
Theorem 18 If n > p then we have the continuous inclusion W
1,p
(R
n
)
L

(R
n
). Moreover, there is a positive constant C = C(p, n) such that
[f(x) f(y)[ C |f|
W
1,p
(R
n
)
[x y[
1(n/p)
,
for all f in W
1,p
(R
n
). (So f is not only bounded, but also Holder continuous.)
Theorems 16, 17, and 18 oer a panorama of the imbedding situation when
= R
n
and k = 1. When is not R
n
we resort to extension operators
(when feasible): start with f in W
1,p
(), extend to

f in W
1,p
(R
n
), apply the
appropriate theorem for that case, then restrict back to the original setting.
The case when k > 1 is obtained more or less by iterating the case when k = 1.
Being able to extend functions in W
k,p
() depends very much on the geom-
etry of the boundary of ; here, smooth, bounded boundaries are best. Nev-
ertheless, when is uncooperative, we can still get good results for the class
W
k,p
0
(), in a trade-o between good boundaries and functions which are zero
at the boundary. As such, these functions admit easy extensions to the whole
space. With this in mind, we state a more general imbedding theorem for
= R
n
, pointing out at the end some instances when the theorem is valid if
,= R
n
(because in those instances there are extension theorems).
Theorem 19 Let k 1 be an integer, and let 1 p < .
i. If
1
p

k
n
> 0, then W
k,p
(R
n
) L
q
(R
n
), for
1
q
=
1
p

k
n
;
ii. If
1
p

k
n
= 0, then W
k,p
(R
n
) L
q
(R
n
), for q [p, );
iii. If
1
p

k
n
< 0, then W
k,p
(R
n
) L

(R
n
).
In this last case k > (n/p). Let m be the integer part of k (n/p), and =
k (n/p) m. For each multi-index with [[ m we have
|D

f|

C |f|
W
k,p
(R
n
)
.
If [[ = m, then
[D

f(x) D

f(y)[ C |f|
W
k,p
(R
n
)
[x y[

.
If [[ < m, then
[D

f(x) D

f(y)[ C |f|
W
k,p
(R
n
)
[x y[.
Thus, for k > (n/p) we have the continuous inclusion W
k,p
(R
n
) C
m
(R
n
).
All of the above results are valid for general if we consider the spaces
W
k,p
0
() instead of W
k,p
().
If is the half-space R
n
+
, or if the boundary of is of class C
1
and bounded,
then the above results are valid for W
k,p
().
20
12 Compactness theorems
Exercise 48 Fix 1 p < . Give an example of a sequence of functions
f
m
W
1,p
(R) such that for all m we have |f
m
|
W
1,p
(R)
= 1, and such that for
any 1 q , and any j ,= k, we have |f
j
f
k
|
q
= c(q) > 0, where c(q) is
a constant depending on q and the sequence f
m
. Conclude that no imbedding of
W
1,p
(R) into L
q
(R) can be compact. Generalize this result to = R
n
.
In view of this exercise we restrict our attention to bounded domains, so from
now on will denote a bounded domain in R
n
. We wish to determine which
imbeddings in Theorem 19 are compact. The target spaces in that theorem
are either L
q
() or C(), so it is important to understand compactness in
these spaces. The Arzela-Ascoli theorem, which we have used before, provides
compactness information in C(). We turn our attention to compactness in
L
q
().
We say that

is relatively compact in if

, and we write

.
In this case there is a > 0, depending only on and

, such that for all h R


n
with [h[ < and all x

, we have x h . (So, moving the whole of

by a quantity wont leave .) We denote by


h
the shift by h, and dene

h
f(x) = f(x h). We will also write

) to indicate the fact that


shifts by keep

inside .
If f L
q
(), we will still denote by f its restriction to

. As such, |f|
L
q
(

)
will make perfect sense. Moreover, if we take h R
n
with [h[ < , then the
restriction to

of
h
f also makes perfect sense, and we may compute the
quantity
|
h
f f|
L
q
(

)
.
Theorem 20 Let R
n
be a bounded open set, and let 1 q < . Let T be
a bounded set in L
q
(). Assume both conditions below:
(i) For every > 0 and every

, there is a > 0 such that

) ,
and for all f T,
[h[ < = |
h
f f|
L
q
(

)
< .
(ii) For every > 0 there is

such that
f T = |f|
L
q
(\

)
< .
Then T is relatively compact in L
q
(); that is, every sequence f
m
T has a
subsequence that converges in L
q
().
Using this criterion, we can prove the following.
Theorem 21 (Rellich-Kondrasov) Let R
n
be a bounded open set of
class C
1
. Then the following inclusions are compact:
i. If p < n, W
1,p
() L
q
(), 1 q < p

;
ii. If p = n, W
1,n
() L
q
(), 1 q < ;
iii. If p > n, W
1,p
() C().
21
13 Once more, with feeling
In the introduction we promised that distributions would enlarge our functional
spaces, allowing us to nd solutions to dierential equations. How well did this
promise stack up? Consider again, for n 3, Poissons equation
u = f,
where f : R
n
C is a given function. As we saw at the end of Section 4, as
long as we can take the Fourier Transform on both sides of this equation, we
obtain
u() =
1
[2[
2

f(),
and it is natural to ask ourselves whether there is a tempered distribution G
such that

G = 1/[2[
2
. Notice that 1/[2[
2
is in L
1
loc
(R
n
), and therefore
can be viewed as a tempered distribution. Therefore there is some tempered
distribution G such that

G = 1/[2[
2
.
Exercise 49 Let M be a square matrix and a function, both dened on R
n
.
We dene the symbol (M)(x) = (Mx). Notice that if is in o(R
n
), then
M is also in o(R
n
). A tempered distribution T is radial if for every orthogonal
matrix M we have T(M) = T(). Show that T is radial if and only if

T is
radial.
Exercise 50 For every > 0 we dene ( )(x) = (x). Again, notice
that if is in o(R
n
), then is in o(R
n
). A tempered distribution T is
homogeneous of degree k if for every > 0 we have T( ) =
k
T(). Show
that T is homogeneous of degree k if and only if

T is homogeneous of degree
(n + k).
Since our

G is radial and homogeneous of degree 2, then G is radial and
homogeneous of degree 2 n.
Exercise 51 Dene G(x) = c
n
[x[
2n
for some constant c
n
. Determine the c
n
for which

G =
1
|2|
2
.
Looking at the above, theres nothing to prevent f from being a distribution.
In particular if we take f = , then u = G is a solution to the equation. In
general, if L is a linear partial dierential operator, a solution G to LG =
is called a fundamental solution to L, and if we are looking for solutions to
Lu = f, then u = f G is one such solution. Thus nding fundamental solutions
is a useful thing, although unfortunately not always easy.
Theorem 22 (Malgrange-Ehrenpreis) If L is a constant-coecient linear
partial dierential operator, then L has a fundamental solution.
22
This fundamental solution is a distribution, and doesnt necessarily corre-
spond to a function in L
1
loc
.
We now turn to the other motivation we oered for the denition of distri-
butions, that of distributions as continuous linear functionals dened over all
spaces L
p
. In general terms, if V
1
V
2
, then the reverse relation is true for their
duals, and V

1
V

2
. Since T() is contained in all L
p
(), it follows that T

()
contains all L
p

(). In particular, it contains the duals of all W


k,p
. This is all
very ne, but also very abstract, and it would be good to have more concrete
representations for these dual spaces. For example, in Exercise 22 we asked for
an explicit element in the dual of H
1
(R
n
). Lets denote by H
s
(R
n
) the dual
of H
s
(R
n
). We have the following characterization of H
1
(R
n
).
Theorem 23 If T is in H
1
(R
n
), then there are functions f
0
, f
1
, . . . , f
n
in
L
2
(R
n
) such that for any in H
1
(R
n
) we have
T() =
_
R
n
_
f
0
+ f
1

x
1
+ . . . + f
n

x
n
_
dx.
The converse is also true.
Proof. The space H
1
(R
n
) is a Hilbert space with inner product
, ) =
_
+ dx.
According to the Riesz Representation Theorem, there is a unique element u in
H
1
(R
n
) such that T() = , u). In particular
T() =
_
R
n
_
u + u
x
1

x
1
+ . . . + u
x
n

x
n
_
dx,
which is what we wanted to prove, with f
0
= u, f
j
= u
x
j
. The converse is left
as an exercise.
The theorem is true (with the same proof) if we replace H
1
(R
n
) by H
1
0
().
Exercise 52 Let n = 1. Find two functions f
0
and f
1
in L
2
(R) such that if
H
1
(R), then (0) =
_
R
f
0
(x)(x) + f
1
(x)

(x) dx. Conclude that the delta


function is an element of H
1
(R).
So, is in the dual of H
1
(R). We could have concluded the same by other
means. Consider Theorem 19 with k = 1, n = 1, p = 2. That is case iii, and we
conclude that H
1
(R) is contained in L

(R) C(R). Moreover, the inclusion is


continuous. As a consequence
[()[ = [(0)[ ||

C ||
H
1
(R)
.
Again, we conclude that is in the dual of H
1
(R).
The following theorem justies our notation for the dual of H
s
(R
n
).
23
Theorem 24 Let s > 0. The elements of H
s
(R
n
) are those tempered distribu-
tions T such that the tempered distribution U = (1 +[[
2
)
s/2

T is representable
by a function in L
2
(R
n
).
Here, U() =

T((1 +[[
2
)
s/2
). We present a proof for s = 1.
Proof (for s = 1). First, assume T H
1
(R
n
). Then
T() =
_
f
0
+

f
j

x
j
dx,
with f
j
L
2
(R
n
). Taking in o

(R
n
), we see that T is a tempered distribution.
Also,

T() = T(

) =
_
f
0

f
j

j
d
=
_

f
0
+

f
j
(2i)

j
d
=
_

f
0

(2i
j

f
j
) d.
As a result
U() =
_
(1 +[[
2
)
1/2
(

f
0

2i
j

f
j
) d,
showing that U is representable by a function in L
2
(R
n
).
Now suppose T is a tempered distribution, and U = (1 + [[
2
)
1/2

T is
represented by some f in L
2
(R
n
). Let be in T(R
n
), and o(R
n
) be such
that

= . (Notice that () =

().) Then
T() = T(

) =

T() = U((1 +[[
2
)
1/2
)
=
_
R
n
f() (1 +[[
2
)
1/2
() d
=
_
R
n
f() (1 +[[
2
)
1/2

() d
=
_
R
n
f() (1 +[[
2
)
1/2

() d.
Since is in H
1
(R
n
), we use Holders inequality to obtain
[T()[ |f|
2
|(1 +[[
2
)
1/2

|
2
= |f|
2
||
H
1
(R
n
)
.
This shows that T is in H
1
(R
n
).
Notice that since

= 1, we have shown in yet another way that is in
H
1
(R).
The nal technical tool we want to mention is the famous Poincare inequality,
or perhaps we should say inequalities, because there are several of them. In
essence, it says that on bounded domains the derivative of a function controls
the function. The most common statement of the inequality is the following.
24
Theorem 25 (Poincares inequality) Let be a bounded open set in R
n
.
Then there is a positive constant C = C(, p) such that for every f W
1,p
0
()
|f|
p
C

||=1
|D

f|
p
.
A few things to note about this theorem. First, the right hand side contains
exactly rst order derivatives, and the left hand side contains zero order deriva-
tives. Second, a consequence of this theorem is that the right hand side may be
used to dene a norm in W
1,p
0
(). Third, the result is not true for f in W
1,p
(),
since in that case we can always add a constant to f to falsify the inequality. In
a sense, requiring f to be zero at the boundary anchors the function. There
are other ways to do this; see Exercise 53.
Proof. First we do the case = (a, a)
n
. Take f in T(). Then, with
x = (x

, x
n
), we have
f(x) =
_
x
n
a
f
x
n
(x

, t) dt.
Using Holders inequality, raising to the p-th power, and then increasing the
interval of integration produces
[f(x)[
p
[x
n
+ a[
p/p

_
a
a

f
x
n
(x

, t)

p
dt.
Now integrate in x to get
_

[f(x)[
p
dx (2a)
1+p/p

f
x
n
(x)

p
dx.
In particular |f|
p
C |f/x
n
|
p
C

||=1
|D

f|
p
, with a constant de-
pending only on and p. The result follows because T() is dense in W
1,p
0
().
For a general bounded , enclose it in a box of the form (a, a)
n
, and extend
f to be zero on (a, a)
n
. This extension preserves all norms of f (since f
and its derivatives are identically zero on the boundary of ), and the theorem
follows.
The attentive reader will notice that it is enough to have bounded in just
one direction for the above theorem to hold.
The nal exercise is not just a simple extension (so to speak) of Poincares
inequality.
Exercise 53 (Poincare-Wirtinger inequality) Let be bounded, connected,
and assume the boundary of is C
1
. Denote by f the average of f over ; in
other words f = (measure())
1
_

f(x) dx. Let 1 p . Then there is a


constant C > 0, depending on , n, and p, such that for all f W
1,p
() we
have
|f f|
p
C

||=1
|D

f|
p
.
25

S-ar putea să vă placă și