DISTRIBUTION THEORY

GENERALIZED FUNCTIONS
NOTES

Ivan F Wilde

Contents

1.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2.

The spaces S and S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.

The spaces D and D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.

The Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.

Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

6.

Fourier-Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7.

Structure Theorem for Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

8.

Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Chapter 1
Introduction

The so-called Dirac delta

x 6= 0 but
R function (on R) obeys (x) = 0 for all
d
is supposed to satisfy (x) dx = 1. (The function on R is similarly
described.) Consequently,
Z
Z
Z
(x) dx = f (0)
(f (x) f (0)) (x) dx + f (0)
f (x) (x) dx =

because (f (x)f (0)) (x) 0 on R. Moreover, if H(x) denotes the Heaviside

step-function
(
0, x < 0
H(x) =
1, x 0 ,

then we see that H = , in the following sense. If f vanishes at infinity,

then integration by parts gives
Z
Z



f (x) H (x) dx = f (x) H(x)

f (x) H(x) dx

Z
=
f (x) H(x) dx
Z

=
f (x) dx
0


= f (x) 0
= f (0)
Z
=
f (x) (x) dx .

Of course,R there is no such function with these properties and we cannot

interpret f (x) (x) dx as an integral in the usual sense. The function

is thought of as a generalized function.
However, what does make sense is the assignment f 7 f (0) = h, f i, say.
Clearly h, f + gi = h, f i + h, gi for functions f , g and constants
and . In other words, the Dirac delta-function can be defined not as a
function but as a functional on a suitable linear space of functions. The
development of this is the theory of distributions of Laurent Schwartz.
1

Chapter 1

One might think of (x) as a kind of limit of some sequence of functions

whose graphs become very tall and thin, as indicated in the figure.

Figure 1.1: Approximation to the -function.

The Dirac function can be thought of as a kind of continuous version of
the discrete Kronecker and is used in quantum mechanics to express the
orthogonality properties of non square-integrable wave functions.
Distributions play a crucial r
ole in the study of (partial) differential equations. As an introductory remark, consider the equations
2u
=0
x y

2u
= 0.
y x

and

These ought to be equivalent. However, the first holds for any function u
independent of y, whereas the second may not make any sense. By (formally)
integrating by parts twice and discarding the surface terms, we get
Z
So we might interpret

2u

dx dy =
x y

2
dx dy .
x y

2u
= 0 as
x y
Z

2
dx dy = 0
x y

for all in some suitably chosen set of smooth functions. The point is that
this makes sense for non-differentiable u and, since is supposed smooth,
Z

2
u
dx dy =
x y

2
dx dy ,
y x

2u
2u
=
in a certain weak sense. These then are weak or
x y
y x
distributional derivatives.
Finally, we note that distributions also play a central r
ole in quantum field
theory, where quantum fields are defined as operator-valued distributions.
that is,

ifwilde

Notes

Introduction

Bibliography
I. M. Gelfand and G. E. Shilov, Generalized Functions, Academic Press,
Inc., 1964.
J. Lighthill, Introduction to Fourier Analysis and Generalized Functions,
Cambridge University Press, 1958.
M. Reed and B. Simon, Methods of Mathematical Physics, Volume II,
Academic Press, Inc., 1975.
W. Rudin, Functional Analysis, McGraw-Hill, Inc,. 1973.
L. Schwartz, Theorie des distributions, Hermann & Cie, Paris, 1966.

The proper approach to the theory is via topological vector spacessee

Rudins excellent book for the development along these lines, as well as
much background material. The approach via approximating sequences of
functions is to be found in Lighthills book.
For the preparation of these lecture notes, extensive use was made of the
books of Rudin and Reed and Simon.

November 9, 2005

ifwilde

Chapter 1

Notes

Chapter 2
The spaces S and S

Let Zn+ denote the set of n-tuples (1 , . . . , n ) where each

Pni is a non1
n
negative integer and write Z+ for Z+ . For I+ , let || = i=1 i and let
||
D denote the partial differential operator
.
x1 1 . . . xnn
Finally, if x = (x1 , . . . , xn ) Rn , let x denote the product x1 1 . . . xnn .
Definition 2.1. The complex linear space of bounded continuous complexvalued functions on Rn is denoted Cb (Rn ). It is equipped with the norm
kf k = sup |f (x)| .
xRn

Theorem 2.2. For any d N, Cb (Rd ) is a complete normed space with respect
to the norm k k .
Proof. Suppose (fn ) is a Cauchy sequence in Cb (Rd ), that is, kfn fm k 0
as m, n . We must show that there is some f Cb (Rd ) such that
kfn f k 0 as n . To see where such an f comes from, we
note that the inequality |g(x)| kgk implies that for each x Rd , the
sequence (fn (x)) is a Cauchy sequence in C and therefore converges. Let
f (x) = limn fn (x).
We claim that f Cb (Rd ) and that kfn f k 0. Let > 0 be given.
Since (fn ) is a Cauchy sequence in Cb (Rd ) there is N N such that
kfn fm k <

1
2

()

for all n, m N . But then, for any x Rd ,

|fN +k (x)| |fN +k (x) fN (x)| + |fN (x)| <
by (). Letting k gives
|f (x)|

1
2

+ kfN k

which shows that f is bounded on Rd .

5

1
2

+ kfN k

Chapter 2

Next, we note that for any x Rd , |fn (x) fm (x)| kfn fm k and so
the inequality () implies that
|fn (x) fm (x)| <

1
2

provided n, m N . Letting m , we see that

|fn (x) f (x)|

1
2

()

for any x Rd , provided n N . In other words, fn (x) f (x) uniformly

on Rd . However, each fn is continuous and so the same is true of f . But
then this means that f Cb (Rd ). The inequality () gives
kfn f k

1
2

<

whenever n N and so fn f with respect to k k and the proof is

complete.
Definition 2.3. The linear space of infinitely-differentiable bounded functions
on Rn is denoted by Cb (Rn ). Evidently Cb (Rn ) Cb (Rn ).

The space S (Rn ) is the linear subspace of Cb (Rn ) formed by the set of
functions f on Rn such that x D f (x) is bounded on Rn for each , Zn+ .
S (Rn ) is equipped with the family of norms

|||f |||, = sup | x D f (x) |

xRn

for , Zn+ . The elements of S (Rn ) are said to be rapidly decreasing

functions.
2

Example 2.4. Evidently, the function f (x) = xm ex belongs to S (R) for

any m Z+ . Indeed, S (R) contains all the Hermite functions.
2
2
For any polynomial p(x1 , . . . , xn ) on Rn , the function p(x1 , . . . , xn )e(x1 ++xn )
belongs to S (Rn ).
Definition 2.5. We say that a sequence (fn ) in S (Rd ) converges to f in
S (Rd ) if, for each , Zd+ , |||fn f |||, 0 as n .

The sequence (fn ) in S (Rd ) is said to be a Cauchy sequence in S (Rd ) if

|||fn fm |||, 0 as n, m , for each , Zd+ .
Theorem 2.6. S (Rd ) is complete, that is, every Cauchy sequence in S (Rd )
converges in S (Rd ).
Proof. First consider the case d = 1. So suppose that (fn ) is a Cauchy
sequence in S (R). Fix , Z+ . Then we know that
|||fn fm |||, 0 as n, m .
ifwilde

Notes

The spaces S and S

In other words, the sequence x D fn (x) is a Cauchy sequence with respect

to the norm k k and so converges to some function g, , say.
We shall show that x D g = g, . This follows from the equality
Z x
fn (x) = fn (0) +
fn (t) dt .
0

Indeed, fn = D1 fn g0,1 on R and so, letting n , we may say that

Z x
g0,0 (x) = g0,0 (0) +
g0,1 (t) dt .
0

Hence g0,0 is continuously differentiable and g0,0

= g0,1 . Repeating this
argument, we see that g0,0 is infinitely-differentiable and that D g0,0 = g0, .

Now, D fn g0, = D g0,0 uniformly and so x D fn (x) x D g0,0 (x)

pointwise. But we also know that x D fn g, uniformly and so it follows
that g, = x D g0,0 . We note that g, is bounded and so g0,0 S (R).
Hence fn g0,0 in S (R) and we conclude that S (R) is complete.
For the general d-dimensional case, suppose that (fn ) is a Cauchy sequence
in S (Rd ). Then for each , Zd+ the sequence x D fn is a Cauchy
sequence in Cb (Rd ) and so converges; x D fn g, uniformly on Rd ,
for some g, Cb (Rd ).
Fix x2 , x3 , . . . , xd . Then as in the 1-dimensional argument above, we know
that for any 1 , 1 Z+ , (all relevant partial derivatives exist and)
x1 1 x11 fn (x1 , x2 , . . . , xd ) g (1 ,0,...,0),(1 ,0,...,0) (x1 , x2 , . . . , xd )
= x1 1 x11 g0,0 (x1 , x2 , . . . , xd ) .

Considering now the function x2 7 x1 1 x11 fn (x1 , x2 , . . . , xd ), we similarly

see that
x2 2 x22 x1 1 x11 fn (x1 , . . . , xd ) g (1 ,2 ,0,...,0),(1 ,2 ,0,...,0) (x1 , . . . , xd )
= x2 2 x22 x1 1 x11 g0,0 (x1 , . . . , xd )

for any 2 , 2 Z+ .
Repeating this for the function x3 7 x2 2 x22 x1 1 x11 fn (x1 , x2 , . . . , xd ), we
find that
g (1 ,2 ,3 ,0,...,0),(1 ,2 ,3 ,0,...,0) (x1 , . . . , xd )
= x3 3 x33 x2 2 x22 x1 1 x11 g0,0 (x1 , . . . , xd )
for 3 , 3 Z+ . Continuing this way, we conclude that
g, = x D g0,0
for any , Zd+ . It follows that fn g0,0 in S (Rd ).
November 9, 2005

Chapter 2

Definition 2.7. Continuous linear functionals on S (Rd ) are called tempered

distributions. The linear space of tempered distributions is denoted by
S (Rd ). Thus T S (Rd ) if and only if T : S (Rd ) C is linear and
fn f in S (Rd ) implies that T (fn ) T (f ) in C.
Example 2.8. For fixed a Rd , let a be the map on S (Rd ) given by the
prescription a : f 7 f (a). Evidently a S (Rd ). a is called the Dirac
delta function (at a Rd ).
Remark 2.9. Since T (fn ) T (f ) = T (fn f ) and fn f in S if and only
if (fn f ) 0 in S , we see that a linear map on S is continuous if and
only if it is continuous at 0 S .
Proposition 2.10. Suppose that T : S S is linear and that there are
, Zd+ such that
|T (f )| |||f |||,
for all f S (Rd ). Then T S (Rd ).
Proof. According to the remark above, we need only verify the continuity
of T at 0. But if fn 0 in S , it follows, in particular, that |||fn |||, 0
and so
|T (fn )| |||fn |||, 0
as n . This means that T is continuous at 0, as required.
To establish a converse, we shall introduce another family of norms on S .
Definition 2.11. For each k, m Z+ and f S (Rd ), set
kf kk,m =

||k
||m

|||f |||, .

These norms on S have the property of being directed, that is, for any
and (k , m ) there is (k, m) such that

(k , m )

max{ kf kk ,m , kf kk ,m } kf kk,m
for all f S (Rd ). (Any (k, m) with k max{ k , k } and m max{ m , m }
will do).
Remark 2.12. It is clear that |||fn f |||, 0 for each , Zd+ if and only
if kfn f kk,m 0 for each k, m Z+ . It follows that a linear functional
T on S (Rd ) is a tempered distribution if and only if T (fn ) 0 whenever
kfn kk,m 0 for all k, m Z+ .
ifwilde

Notes

The spaces S and S

Theorem 2.13. A linear functional T on S (Rd ) is a tempered distribution if

and only if there is C > 0 and some k, m Z+ such that
|T (f )| Ckf kk,m

for all f S (Rd ).

Proof. If such a bound exists, it is clear that T S (Rd ). For the converse,
suppose that T S (Rd ) but that there exists no such bound. Then for
any n N, it is false that
|T (f )| n kf kn,n

for all f S (Rd ). In other words, there is a sequence (gn ) in S such that
|T (gn )| > n kgn kn,n .
Set fn = gn /n kgn kn,n so that |T (fn )| > 1. However,
kfn kk,m =

kgn kk,m
1

n kgn kn,n
n

whenever n max{ k, m }. It follows that fn 0 in S (Rd ). This is a

contradiction because it is false that T (fn ) 0. The result follows.
Proposition 2.14. Let g L2 (Rd ). Then the linear map
Z
Tg : f 7 g(x) f (x) dx
on S (Rd ) defines a tempered distribution.

R



Proof. For f S (Rd ), we have |Tg (f )| =  g(x) f (x) dx  kgkL2 kf kL2 .
But
Z
2
kf kL2 = |f (x)| |f (x)| dx
Z
kf k0,0
|f (x)| dx
= kf k0,0

Z Y
d

j=1

kf k0,0 kf k2d,0

(1 + x2j )
Z

= d kf k0,0 kf k2d,0

Qd

k=1 (1

|f (x)| Qd

k=1 (1

+ x2k )

+ x2k )

dx

dx1 dx2 . . . dxd

d kf k22d,0 .

This leads to the estimate

|Tg (f )| kgkL2 d/2 kf k2d,0

which shows that T S (Rd ), as claimed.

November 9, 2005

10

Chapter 2

The next result indicates that polynomially bounded functions determine

tempered distributions, via integration.
TheoremQ2.15. Suppose that g(x) (is measurable and) is such that for some
m N, dj=1 (1 + x2j )m g(x) is bounded on Rd . Then the map
Tg : f 7

g(x) f (x) dx

is a tempered distribution.
Q
Proof. Let p(x) = dj=1 (1 + x2j ). Then the hypotheses mean that for any
f S (Rd )
|g(x) f (x)| = p(x)m |g(x)| p(x)m |f (x)|
< M p(x)m |f (x)|

for some M > 0. It follows that g(x)f (x) is integrable and so Tg is welldefined on S (Rd ).
To show that T S (Rd ), we estimate
Z
|Tg (f )| < M p(x)m |f (x)| dx
Z
1
dx
= M p(x)m+1 |f (x)|
p(x)
Z
1
dx
M kf k2d(m+1),0
p(x)
= M kf k2d(m+1),0 d .

It follows that Tg S (Rd ).

Theorem 2.16 (Cauchy Principal Part Integral).
The functional
Z
1
P ( x1 ) : f 7 lim
f (x) dx
0 |x| x
belongs to S (R).
Proof. We first show that P ( x1 ) is well-defined on S (R). For f S (R)
Z

|x|

1
f (x) dx =
x

f (x) f (x)
dx .
x

f (x) f (x)
f (x) f (x)
2f (0) as x 0 and therefore
is
x
x
integrable on [0, ) and P ( x1 ) is indeed well-defined.
However,

ifwilde

Notes

The spaces S and S

11

Clearly P ( x1 ) is linear and so we need only verify its continuity on S (R).

To do this, we observe that for x > 0
 f (x) f (x)   1 Z x


 


f (t) dt 

=
x
x x
Z x
 
1
f (t) dt

x x
2 kf k .

Therefore
Z

f (x) f (x)
f (x) f (x)

dx +
dx 
x
x
1
0
Z 1
Z
dx

2 kf k dx +
{ |f (x)| + |f (x)| } x 2
x
0
1
Z
dx
2 kf k + 2 kxf (x)k
x2
1
= 2 |||f |||0,1 + 2 |||f |||1,0 .

Z
 
 1
P ( )(f ) = 
x

The result follows.

Definition 2.17. A sequence (Tn ) in S (Rd ) is said to converge in S (Rd )
if Tn (f ) T (f ) for each f S (Rd ). One also says that Tn converges to T
in the sense of distributions.
Example 2.18. Lebesgues Dominated Convergence Theorem implies that (if
each gn is measurable and) if gn (x) Rg(x) pointwise andR if |gn (x)| (x)
for some integrable function , then gn (x)f (x) dx g(x)f (x) dx for
each f S . In other words, the sequence Tgn of tempered distributions
converges to Tg in S (Rd ).
We wish to discuss the well-known formula
lim
0

1
1
= P ( xx
) i (x x0 ) .
0
x x0 + i

We will see that this holds with convergence in the sense of distributions.
Theorem 2.19. Let g (x) =

x2

x
. Then Tg P ( x1 ) in S (R) as 0.
+ 2

Proof. Let f S (R) and let > 0. Then

Z


 
x f (x)


 
1
1
dx 
 P ( x )(f ) Tg (f )  =  P ( x )(f )
2
2
x +
Z
 Z f (x) f (x)

xf (x) xf (x)


=
dx
dx

2 + 2
x
x
0
0

November 9, 2005

12

Chapter 2

Z

f (x) f (x)
2


=
dx

2 + 2 )
(x
x
0


Z

f (x) f (x)
2



dx

2 + 2 )
(x
x
0



Z
f (x) f (x)
2


dx
+

(x2 + 2 )
x

Z 

 f (x) f (x) 

 dx

x
0
Z 2 
 f (x) f (x) 
+

 dx .
2
x

(x)
2f (0) as x 0 and so the first term on the right hand
Now, f (x)f
x
side can be made arbitrarily small by choosing sufficiently small. But for
(x)
fixed > 0, the fact that f (x)f
is integrable means that the second term
x

approaches 0 as 0. [Alternatively, one can set = in the discussion

above. Another proof is to use Lebesgues
Monotone
Convergence Theorem


2  f (x)f (x) 
together with the fact that x2 +2
0 on (0, ) as 0.]
x

The next theorem tells us that the Dirac delta function is the limit, in
the sense of distributions, of a sequence of functions whose graphs become
thin, tall peaks around x = 0.
Theorem 2.20. Let (n ) be a sequence of functions on R such that
(i) n (x) 0 for all x R.
R

(ii)

n (x) dx = 1 for all n.

(iii) For any a > 0,

|x|a n (x)

dx 0 as n .

Then n in S (R) as n (that is, Tn in S (R)).

R
Proof. Fix f S (R) (with f
6 0). To show that n (x) f (x) dx f (0)
as n , let > 0 be given. Then, for any > 0,

 Z
Z

 

 n (x) f (x) dx f (0)  =  n (x) (f (x) f (0)) dx 
Z

n (x) |f (x) f (0)| dx

Z
+
n (x) |f (x) f (0)| dx .
x

ifwilde

Notes

The spaces S and S

13

Fix > 0 such that |f (x) f (0)| < 12 for all |x| . Then we can estimate
the first term on the right hand side by
Z
Z
n (x) dx
n (x) |f (x) f (0)| dx 21

Z
1
n (x) dx
2

1
2

for all n. Furthermore, by hypothesis, there is N N such that if n > N

then
Z

n (x) dx <
.
4kf k
|x|
So for n > N , the second term on the right hand side above is estimated
according to
Z
Z
n (x) dx
n (x) |f (x) f (0)| dx 2kf k
x

1
2

Hence, for all n > N , we find that

Z



 n (x) f (x) dx f (0)  <
as required.

Remark 2.21. If we replace (iii) by the requirement that

0 as n , then one sees that n x0 in S (R).
Corollary 2.22. For > 0, let g =
S (R) as 0.

|xx0 |a n (x)

dx

. Then g x0 in
(x x0 )2 + 2

R
Proof. Clearly g (x) 0 for all x R and g (x) dx = . Also, for any
a > 0,
Z
Z

g (x) dx = 2
dx
2
x + 2
|xx0 |a
a
h
i
= 2 tan1 x
a


1 a

= 2 2 tan
0

as 0. The result now follows from the theorem (applied to

g ).

November 9, 2005

14

Chapter 2

Theorem 2.23. For > 0, let h =

1
. Then
x x0 + i

1
) i x0
h P ( xx
0

in S (R), as 0.
Proof. We have
x x0 i
1
=
x x0 + i
(x x0 )2 + 2
(x x0 )
i
=

(x x0 )2 + 2 (x x0 )2 + 2
1
) i x0
P ( xx
0

h (x) =

in S (R) as 0, by the previous theorems.

To motivate the next definition, consider the integral
where f, g S (R). Integrating by parts, we find that
Z
Z

g (x) f (x) dx = g(x) f (x) dx .

g (x) f (x) dx

Using our notation introduced earlier, identifying a function g with the

distribution Tg , this equality becomes
Tg (f ) = T (g)(f ) .
If we think of Tg as the derivative of Tg , then the following definition is
quite natural.
Definition 2.24. Let T S (Rd ) and Zd+ . The weak derivative D T
(or the derivative in the sense of distributions) is defined by
(D T )(f ) = (1)|| T (D f )
for f S (Rd ).
This corresponds to D Tg = TD g . Note that a distribution always has
a weak derivative. Of course, we should verify that the weak derivative of a
tempered distribution is also a tempered distribution. We do this next.
Theorem 2.25. For any Zd+ , D : S (Rd ) S (Rd ) is continuous. In
particular, for any T S (Rd ), D T S (Rd ) .
Proof. Suppose that fn 0 in S (Rd ). Let , Zd+ . Then
|||D fn |||, = kx D D fn k
= kx D+ fn k

ifwilde

Notes

The spaces S and S

15

= |||fn |||,+
0
as n . Hence D : S (Rd ) S (Rd ) is continuous.

Now let T S (Rd ). Evidently D T is well-defined and is a linear map on

S (Rd ). If fn 0 in S (Rd ), then D fn 0 in S (Rd ), by the first part.
Hence
(D T )(fn ) = (1)|| T (D fn ) 0
and so D T S (Rd ), as required.

Examples 2.26.
1. We find that a (f ) = a (f ) = f (a).
(
x, x > 0
2. Let g(x) =
0, x 0.
Then we know that Tg S (R) and

Tg (f ) = Tg (f )
Z
=
x f (x) dx
0
Z

= [xf (x)]0 +
f (x) dx
0
Z
=
f (x) dx
0
Z
=
H(x) f (x) dx

where H(x) =

1,
0,

x0
x<0

is the Heaviside step-function.

So Tg = TH . Moreover,
TH (f ) = TH (f )
Z
=
f (x) dx
0

= [f (x)]
0
= f (0)

= (f )
and so TH = . Therefore Tg = TH and Tg = TH = . We say g = H
and g = H = , in the sense of distributions.
November 9, 2005

16

Chapter 2

Remark 2.27. We notice that although is not a function, it is the second

distributional derivative of a continuous function, namely g. We will see
that every tempered distribution is the weak derivative (of a suitable order)
of some continuous function.
Definition 2.28. The support of a function f on Rd , denoted by supp f , is
the closure of the set where f does not vanish;
supp f = { x Rd : f (x) 6= 0 } .
Let C0 (Rd ) denote the linear subspace of C (Rd ) of those functions with
compact support. Clearly C0 (Rd ) S (Rd ).
Example 2.29. For x R, let

(
2
e1/(1x ) , |x| 1
h(x) =
0,
|x| > 1 .

Then h is infinitely-differentiable and one finds that its nth derivative has
the form h(n) (x) = pn (x, 1/(1 x2 )) h(x) for some polynomial pn (s, t) and
therefore h C0 (R).
Rx
Let g(x) = h(t) dt. Then g is infinitely-differentiable, g(x) = 0 for
x < 1 and g is constant for x > 1. Evidently g
/ S (R).
Let g (x) = g(x), where > 0. Then g is zero for x < 1/ and constant
for x > 1/. Now let g,a (x) = g((x a)). Then g,a vanishes for x < a 1
and is constant when x > a + 1 .
Let a < b and suppose that , are such that a < a + 1 < b 1 < b. Let
f (x) = g,a (x) g,b (x). Then f C (R) and we see that f (x) = 0 for
x < a 1 , f (x) = 0 for x > b + 1 and f is constant for a + 1 < x < b 1 .
Evidently f C0 (R) and supp f [a 1 , b + 1 ].
g(x)
f (x)

b+

Figure 2.1: The functions g(x) and f (x).

In d-dimensions, set f (x1 , . . . , xd ) = g(|x|2 ) where g C (R) is such that
g(t) = 1 for 0 t a and g(t) = 0 for x b, where 0 < a < b. Then
f C0 (Rd ), supp f { x Rd : |x|2 b2 } and f (x) = 1 for |x|2 a2 .
ifwilde

Notes

The spaces S and S

17

In particular, if Nr (x0 ) denotes the open ball of radius r, centred at x0

in Rd , then there exist functions f C0 (Rd ) such that f = 1 on Nr/2 (x0 )
and f vanishes outside Nr (x0 ).
Theorem 2.30. Let K be compact and let A be an open set in Rd with
KsubsetA. Then there exists a C -function such that 0 (x) 1
for all x Rd , (x) = 1 for all x K and (x) = 0 for x
/ A.
Proof. For each x K, there is r(x) > 0 such that Nr(x) (x) A. The
collection { Nr(x)/2 (x) : x K } is an open cover of the compact set K and
so has a finite subcover, that is, there is x1 , . . . , xm K such that
K Nr1 /2 (x1 ) Nrm /2 (xm )
where ri = r(xi ).
Let i C (Rd ) be such that 0 i (x) 1, i (x) = 1 for x Nri /2 (xi )
and i (x) = 0 if x
/ Nri (x). (Such functions can be constructed as in the
previous example.)
Set
(x) = 1 (1 1 (x))(1 2 (x)) . . . (1 m (x)) .
Then C (Rd ) and obeys 0 (x) 1 for all x Rd . Furthermore,
for any x K, there is some 1 i m such that x Nri /2 (xi ) and so
i (x) = 1 and therefore (x) = 1.
On the other hand, for any x
/ A, it is true that x
/ Nri (xi ) for all
1 i m (since Nri (xi ) A). Hence i (x) = 0 for all 1 i m and so
(x) = 1 1 = 0. Therefore satisfies the requirements and the proof is
complete.

Theorem 2.31. C0 (Rd ) is dense in S (Rd ).

Proof. Let S (Rd ) and let fn C0 (Rd ) be a sequence of functions
such that supp fn { x Rd : |x| < n + 1 }, fn (x) = 1 for |x| n 1
and such that the shape of the graph of fn for |x| between n 1 and n + 1
is independent of n. This means that for any given multi-index Zd+ ,
D fn (x) is bounded uniformly in n. (Such functions can be constructed as
in example 2.29).
Let n = fn . Then n C0 (Rd ) and
|||n |||, = |||(fn 1)|||,

= sup | x D ((x)(fn (x) 1)) | .

xn1

But for each Zd+ , kD (fn (x) 1)k is bounded uniformly in n and so it
follows by Leibnitz formula and the fact that S that |||n |||, 0
as n for each , Zd+ .
November 9, 2005

18

Chapter 2

Definition 2.32. Let G be an open set in Rd . We say that a distribution T

vanishes on G if T () = 0 for each S (Rd ) with suppR G.
If g S vanishes on G (as a function) then evidently g(x) (x) dx = 0
for all S with supp G, that is, Tg vanishes on G as a distribution.
Remark 2.33. Suppose that T S (Rd ) vanishes on the open sets G1 and
G2 where G1 G2 = . Then T also vanishes on G1 G2 . To see this, let
S with supp G1 G2 .
Now supp G1 is a closed set in G1 and so there is an infinitelydifferentiable function f1 such that f1 = 1 on supp G1 and f1 = 0
outside some closed set F1 containing supp G1 . Hence f1 has support
in G1 . Similarly, there is some f2 such that f2 has support in G2 .
But = f1 + f2 and therefore T () = T (f1 ) + T (f2 ) = 0 since the
distribution T vanishes on both G1 and G2 .
This result has a satisfactory generalization, as follows.
Theorem 2.34. Suppose that T S (Rd ) vanishesSon each member of a
collection { G } of open sets. Then T vanishes on G .

Proof. A proof of this result may be found in Rudins book1 .

Thanks to this theorem, the following (desirable) definition makes sense.

Definition 2.35. For any T S (Rd ), let W denote the union of all open
sets on which T vanishes. The support of T is defined to be supp T = W c ,
the complement of W in Rd .
Examples 2.36.
1. One sees that supp a = { a }, for any a Rd .
2. If H is the Heaviside function, then we see that supp H = [0, ).

[Functional Analysis, by Walter Rudin, Tata McGraw-Hill, 1973]

ifwilde

Notes

Chapter 3
The spaces D and D

In this section, we consider another space of functions and the associated

collection of continuous linear functionals.
Definition 3.1. Let Rd be an open subset of Rd . C0 () denotes the
linear subset consisting of those functions in C0 (Rd ) which have support
in . Suppose that (n ) is a sequence in C0 () and let C0 (Rd ). We
say that n in C0 () if
(i) there is some compact set K such that supp n K for all n,
and
(ii) D n D uniformly as n , for each Zd+ .
Note that it follows immediately that supp K.

D() is the space C0 () equipped with this notion of convergence and we

say that n in D().

Example 3.2. Let C0 (R) be such that (x) = 0 for |x| > 1. For each
n N, let n (x) = (x 1) + 21 (x 2) + + n1 (x n) (so that n
comprises n smaller and smaller smooth bumps).
Evidently, n C0 (R) and (n ) is a Cauchy sequence
Pwith1 respect to
the norm k k . Indeed, kn k 0 where (x) =
k=1 k (x k).

Clearly C (R) but does not converge to in D(R) because the

supports of the n are not all contained in a compact set (and
/ C0 (R),
anyway).
Continuing with this notation, let hn (x) = n1 (x n). Then hn C0 (R)
and khn k 0 but (hn ) does not converge to 0 in D(R) (because there is
no compact set K such that supp hn K for all n).
The notion of convergence in D ensures its completeness, as we show next.
Definition 3.3. We say that (n ) is a Cauchy sequence in D() if there is
some compact set K such that supp n K for all n and such that
kD (n m )k 0 as n, m for every Zd+ .
19

20

Chapter 3

Theorem 3.4. D() is complete.

Proof. Exactly as in the proof of the completeness of S , we see that if (n )
is a Cauchy sequence in D(), then there is some f C (Rd ) such that
kD (n f )k 0 for all Zd+ . But if supp n K for all n, then it
follows that supp f K also. Hence f C0 () and n f in D().
Definition 3.5. A linear functional u : D() C is said to be continuous
if u(n ) u() whenever n in D() as n . Such a continuous
linear functional is called a distribution. The linear space of distributions is
denoted D ().
The derivatives of a distribution are defined as for tempered distributions,
namely by the formula
D u() = (1)|| u(D )
for D() and Zd+ .
Example 3.6. Clearly the map a : 7 (a) (Dirac delta function) is a
distribution (i.e., belongs to D (Rd )) for any a Rd .
Example 3.7. Suppose u is a locally integrable function (that
is, u L1 (K)
R
d
for each compact set K R ). Then the map Tu : 7 Rd u(x) (x) dx is
2
a distribution. (In particular, u(x) = ex defines a distribution, Tu D (R)
but Tu
/ S (R). Indeed, Tu is not defined on every element of S (R).)
To see this, we first note that Tu is well-defined because has compact
support if it belongs to D(Rd ). Furthermore,
Z



u(x) (x) dx
|Tu ()| =
Z
|u(x)| |(x)| dx ,

K
Z
kk
|u(x)| dx .

where K = supp ,

From this, we see that if n in D(Rd ), then certainly Tu (n ) Tu ()

so that Tu D (Rd ), as claimed.
Remark 3.8. It is sometimes convenient to identify u with Tu and to consider
the function u as being a distribution (namely that given by Tu ).
The next result tells us that every tempered distribution is a distribution.
(If this were not true then the terminology would be most inappropriate.)
ifwilde

Notes

21

The spaces D and D

Theorem 3.9. Let T S (Rd ). Then T C0 (Rd ) D (Rd ).

Proof. Suppose that n in D(Rd ). Then there is some compact set
K Rd such that supp n K for all n. So for any , Zd+
|||n |||, = kx D (n )k

= sup | x D (n ) |
xK

C sup | D (n ) |
x

0,

as n ,

where C is some constant such that |x | < C for all x K. It follows that
n in S (Rd ) and so T (n ) T (). Hence 7 T () is continuous
on D(Rd ).
Theorem 3.10. A linear functional u on C0 () is a distribution if and only
if for each compact subset K there is a constant C and an integer N
such that
|u()| C ||||||N , for all C0 (K),
P
where ||||||N ||N kD k .

Proof. Clearly u is continuous on D() if such bounds hold. Conversely,

suppose that u D () but that no such bounds exist. Then there is some
compact set K0 and a sequence (n ) in C0 (K0 ) such that
|u(j )| > j |||j |||j ,

for all j N.

Set fj = j /u(j ) so that u(fj ) = 1 for all j. However, for any Zd+ ,
kD j k
|u(j )|
||||||j

, for all j > ||,

|u(j )|

kD fj k =

<

1
j

0,

as j .

Hence fj 0 in D() which forces u(fj ) 0. This contradicts the fact

that, by construction, u(fj ) = 1 for all j. The result follows.
Remark 3.11. Note that the same proof works if the norm ||||||N above is
redefined to be max||N kD k .
Definition 3.12. If the integer N in theorem 3.10 can be chosen independently
of K, then the distribution u is said to of finite order. The smallest such N
is called the order of u.
November 9, 2005

22

Chapter 3

Example 3.13. Let u D (R) be given by u = , that is, u : 7 (0).

P (n)
Then we see that u has order 1. On the other hand, if u =
n=0 n
(n)
(where n () = (1)n (n) (n)), then u is an element of D (R) but its order
is infinite.
Note that in this latter case, supp u = { 0, 1, 2, . . . }. We show next that
distributions with compact support must have finite order.
Theorem 3.14. Suppose that u D () and that supp u is compact. Then u
has finite order. In fact, there is C > 0 and N Z+ such that
|u()| C ||||||N
for all C0 () (i.e., C does not depend on nor on supp ).
Proof. Suppose that u D () and that supp u is compact. Let W be an
open set with supp u W and let C0 () be such that = 1 on W .
For any C0 (), we have
u() = u( + (1 ) ) = u( ) + u((1 ) ) .
But (1) = 0 on W and so u((1) ) = 0 and therefore u() = u( )
for all C0 (). It follows by theorem 3.10 that there is some C > 0
and N Z+ such that
|u()| = |u( )| C ||| |||N
for all C0 () (since supp supp ). Note that C does not depend
on nor supp but may depend on supp . An application of Leibnitz
formula implies that
|u()| = |u( )|

C ||| |||N
C ||||||N

for all C0 () for some C > 0 which may depend on but does not
depend on nor on supp . The result follows.
Remark 3.15. The converse is false. Indeed, suppose that u(x) is a non-zero
constant, say u(x) = c 6= 0, for all x R. Then supp u = R which is not
compact. However, for any compact set K, and any C0 (K), we have
Z



|u()| =
c (x) dx 
|c| diam K kk .

So u has order zero but its support is not compact.

ifwilde

Notes

23

The spaces D and D

Theorem 3.16. Let T S (Rd ). Then T D(Rd ) is a distribution of finite

order.
Proof. We have already seen that T D(Rd ) is a distribution, so we need
only show that it has finite order. Since T S (Rd ), there is C0 > 0 and
integers n, k Z+ such that
|T (f )| C0 kf kn,k

for all f S (Rd ).

|T (f )| C kf kn,k ,

for all f C0 (K).

But if K Rd is compact, there is M > 0 such that |x | M for all Zd+

with || n and all x K. Hence there is C > 0 such that

It follows that
|T ()| C ||||||k

for all C0 (K), where k does not depend on K and so we see that the
order of T on D(Rd ) is finite (no larger than k).
Remark
Again, the converse is false. For example, the linear map
R 3.17.
2
7 ex (x) dx defines an element of D (R) which does not extend to a
continuous functional on S (R). However, for any compact set K R,
|u()| C ||||||0
for all C0 (K) so that u has order 0. There is no T S (R) such that
T D(R) = u.
Theorem 3.18. Let u D (Rd ) and suppose that supp u is compact. Then
u S (Rd ), that is, there is a unique T S (Rd ) such that T D(Rd ) = u.
Proof. Let W be an open ball in Rd with supp u W and let C0 (Rd ) be
such that W supp and = 1 on W . For any f S (Rd ), f C0 (Rd )
and so we may define the linear map T on S (Rd ) by
T (f ) = u(f )

for f S (Rd ).

Now, for any C0 (Rd ), = on W and so supp( ) W c and

therefore u( ) = u(). It follows that T () = u() for any C0 (Rd ).
Suppose that fn 0 in S (Rd ). Then (using Leibnitz formula) it follows
that fn 0 in S (Rd ). But supp fn supp for all n and so fn 0
in D(Rd ). This means that T (fn ) = u(fn ) 0 as n and so T is
continuous on S (Rd ), that is, T S (Rd ).
To see that T is unique, suppose that S S (Rd ) and S D(Rd ) = u.
Then S T S (Rd ) and vanishes on C0 (Rd ) which is dense in S (Rd ).
By continuity, it follows that S = T on S (Rd ).

November 9, 2005

24

Chapter 3

We know that and all its derivatives are distributions whose support is
the singleton set { 0 }. The next theorem gives a precise converse.
Theorem 3.19. Let u D (Rd ) be such that supp u = { 0 }. Then there is
N Z+ and constants aj such that
X
u=
aj Dj .
|j|N

Proof. We will only give the proof for d = 1. The general case is similar.
So suppose that u D (R) with supp u = { 0 }. Since { 0 } is compact, it
follows that u has finite order, N , say. Then there is C > 0 such that
|u()| C ||||||N ,

for all D(R).

Claim: if the derivatives (j) (0) = 0 for all 0 j N , then u() = 0.

Proof of Claim. Let h C0 (R) be such that h(x) = 1 for |x| 1 and
such that h(x) = 0 if |x| 2. Set hn (x) = h(nx) so that hn (x) = 1 if
|x| n1 but hn (x) = 0 if |x| n2 . Let n (x) = hn (x) (x). Evidently,
supp n { x : |x| 2 }.
We will show that |||n |||N 0 as n . Let k Z+ be fixed such that
0 k N and let > 0 be given. Since (k) (0) = 0, there is > 0 such
that |(k) (x)| < for all |x| < . But then
Z x



(k1)
(k)
|
(x) | = 
(t) dt  |x|
0

for all |x| < . Continuing in this way, we obtain

|(k2) (x)|

|x|2

2!

|(kr) (x)|

|x|r

r!

..
.

for all |x| < . Using Leibnitz formula, we find that

| Dk n (x) | = | Dk (hn (x)(x)) |
k  

X
k


Dr hn (x) Dkr (x) 
=
r
r=0
k  
X

k


r (r)
(kr)
=
n h (nx)
(x) 
r
r=0
k
X k  |nx|r
|h(r) (nx)| .

r!
r
r=0

ifwilde

Notes

25

The spaces D and D

Hence
sup | Dk n (x) | = sup | Dk n (x) |
x

|x|2/n

C
for some constant C > 0 (which may depend on k but not on n ) provided
2/n < , that is, n > 2/. It follows that supx | Dk n (x) | 0 as n
for each 0 k N and so |||n |||N 0, as required.
To complete the proof of the claim, we note that the bound for u implies
that u(n ) 0. However, (x) n (x) = 0 if |x| < 1/n, and therefore
supp( n ) { x : |x| 1/n }. Since supp u = { 0 }, it follows that
u( n ) = 0 so u() = u(n ) for all n which forces u() = 0 and the claim
is proved.
To continue with the proof of the theorem, let C0 (R) be given and
define by
(x) = (0) + x (0) +
|

x2
2!

(2) (0) + +
{z
=p(x)

xN
N!

(N ) (0) + (x) .
}

Then is infinitely-differentiable and (k) (0) = 0 for all 0 k N . Note,

however, that
/ C0 (R) indeed, is not even bounded.
Let g C0 (R) be such that g(x) = 1 if |x| 1. Then u() = u( g) and
g = pg + g.
Now, g C0 (R) and (by Leibnitz formula) we see that ( g)(k) (0) = 0
for all 0 k N and so u( g) = 0, according to the claim above. Hence
u() = u( g)
= u( p g ) + u( g)
= u( p g )
= (0) u(g) + (0) u(xg) + + (N ) (0) u(xN g/N !)
=

N
X

ak Dk ()

k=0

where ak = (1)k u(xk g)/k!.

Definition 3.20. For any C0 () and u D (), the product u is the
linear functional
u : 7 u( ) ,

for D().
November 9, 2005

26

Chapter 3

Theorem 3.21.
(i) For any C0 (), the product u D ().
(ii) Suppose that and its derivatives are polynomially bounded, that
is, for each multi-index Zd+ , there is some integer N N and
constant C > 0 such that | D (x) | C (1 + |x|2 )N for all x Rd .
Then for any f S (Rd ), the function f S (Rd ) and the map
T : f 7 T ( f ) defines a tempered distribution.
Proof. (i) For any C0 (), the product C0 () also and so u is
a well-defined linear functional on D(). Now if n in D(), it follows
from Leibnitz formula that n in D() and so u D (), as
claimed.
(ii) If and its derivatives are polynomially bounded, then f S (Rd )
for any f S (Rd ). Furthermore, again by Leibnitz formula, we see that if
fn f in S (Rd ), then also fn f in S (Rd ) and so T S (Rd ).

ifwilde

Notes

Chapter 4
The Fourier transform

We begin with the definition of the Fourier transform and the inverse Fourier
transform for smooth functions.
Definition 4.1. The Fourier transform of the function f S (Rd ) is the
function Ff given by
Z
1
Ff () = (2)d/2
ei x f (x) dx
Rd

where Rd and x =

Pd

j=1 j

xj for x Rd .

The inverse Fourier transform of f S (Rd ) is the function F 1f given by

Z
1
1
ei x f (x) dx .
F f () = (2)d/2
Rd

It is often convenient to also use the notation fb for Ff .

Of course, the terminology must be justified, that is, we must show that
these transforms really are inverses of each other.
Proposition 4.2. Let f S (Rd ). Then fb C (Rd ) and for any , Zd+


(i) D fb () = ( D ((ix) f (x)) )b () .

In particular, ij fb () = (Dj f )b () and (Dj fb )() = (ixj f (x))b ().

Proof. Let e1 = (1, 0, . . . , 0), e2 = (0, 1, 0, . . . , 0), . . . , ed = (0, 0, . . . , 0, 1)

denote the standard basis vectors for Rd . For f S (Rd ) and h 6= 0,
Z
 fb ( + h e ) fb ()



j
1
+ (2)d/2
i xj ei x f (x) dx 

h
Rd
Z n i(+hej ) x


o
(e
ei x )


+ i xj ei x f (x) dx
=  (2)1d/2
h
d
R
0 as h 0
27

28

Chapter 4

since f S (Rd ). In other words, differentiation under the integral sign is

justified. Repeated differentiation (since xj f (x) S (Rd )) shows that
Z


b
1
(ix) ei x f (x) dx = (ix) f (x) b () .
(D f )() = (2)d/2
Rd

Furthermore,

( D fb )() =
=

=
so that

1
(2)d/2
1
(2)d/2
(1)
(2)d/2

(i)
(2)d/2

(ix) ei x f (x) dx

Rd

Rd

Rd

(i) ( Dx ei x ) (ix) f (x) dx



(i) ei x Dx (ix) f (x) dx
ei x

Rd

(integrating by parts)


Dx (ix) f (x) dx

( (i) D fb )() = ( D ( (ix) f (x) ) )b ()

and the proof is complete.

Remark 4.3. Clearly, similar formulae also hold for the inverse Fourier
transform F 1f (replacing i by i).
Theorem 4.4. Both F and F 1 are continuous maps on S (Rd ).
Proof. We first show that if f S (Rd ), then so are Ff and F 1f . For any
, Zd+ , we have
Z





 (i)

ix
b
e
D (ix) f (x) dx 
| ( D f )() | =  (2)d/2
d
Z R


| D x f (x) | dx .
(2)1d/2
Rd

It follows that |||fb |||, = supRd | ( D fb )() | is finite for each pair of
multi-indices , Zd+ and therefore fb S (Rd ). A similar proof holds for
F 1f (or one can simply note that (F 1f )() = (Ff )()).

To show that F : S (Rd ) S (Rd ) is continuous, we use the estimate

obtained above. We have
Z


|||fb |||, (2)1d/2
| D x f (x) | dx
Rd
Z


x21 x22 x2d
| D x f (x) | dx
(2)1d/2
2
2
2
Rd (1 + x1 )(1 + x2 ) (1 + xd )
ifwilde

Notes

29

The Fourier transform

sup | x21 x2d D

x

C kf km,n

x f (x) |

1
(2)d/2

Z

d
1
dt
(1 + t2 )

by Leibnitz formula, for some constant C > 0 and integers m, n Z+

depending on and . (In fact, we can take m = || + 2d and n = ||.)
d
b
From this, it follows that if fn f in S (Rd ), then fc
n f in S (R ), that
d
d
is, the map F : S (R ) S (R ) is continuous.
Similarly, one sees that F 1 is continuous on S (Rd ).
The next theorem justifies the terminology.
Theorem 4.5 (Fourier Inversion Theorem). For any f S (Rd ),
F 1(Ff ) = f = F(F 1f )
(so that the Fourier transform F is a linear bicontinuous bijection of S (Rd )
onto S (Rd ) with inverse F 1).
Proof. For any f, g S (Rd ), we have
Z
Z
Z
o
n
iy
1
b
eix f (x) dx eiy d
g() f () e d =
g() (2)d/2
Rd
Rd
Rd
Z nZ
o
= (2)1d/2
g() ei(xy) d f (x) dx
Rd
Rd
Z
gb (x y) f (x) dx
=
d
ZR
=
gb (x) f (y + x) dx .
Rd

Let g () = g(), so that

gb (x) =
=

1
(2)d/2 )

1
(2)d/2 )

=
Therefore
Z

Rd

Rd

eix g() d
Z
eixu/ g(u) du
Rd

gb (x/) .

g() fb () eiy d =
=

Rd

d
ZR

Rd

gb (x) f (y + x) dx

gb (x/) f (y + x) dx/d

gb (x) f (y + x) dx .

November 9, 2005

30

Chapter 4

Letting 0, we obtain
Z
Z
iy
b
f () e d = f (y)
g(0)

Rd

Rd

gb (x) dx .

NowRset g(x) = ex /2 . Then g(0) = 1 and one knows that gb (u) = eu /2

and Rd gb (u) du = (2)d/2 . Substituting this into the equation above gives
(F 1fb )(y) = f (y) ,

that is, F 1(Ff ) = f . Similarly, one shows that F(F 1f ) = f and the
result follows.
Remark 4.6. We see that
(F 2 f )(x) = (Ffb )(x) = (F 1fb )(x) = f (x) .

It follows that F 4 f = f so that F and F 1 have period 4.

Furthermore, writing the identity ij fb () = (Dj f )b () obtained earlier as
ij (Ff )() = (FDj f )() and replacing f by F 1f , we get the formula
ij f () = (FDj F 1f )(). This gives the identity
ij = FDj F 1
as operators on S (Rd ).
We also find that F 1(ij )F = Dj and F(ij )F 1 = F2 Dj F 2 = Dj
on S (Rd ).
Corollary 4.7 (Parsevals formula). For any f, g S (Rd )
Z
Z
fb (x) gb (x) dx =
f (x) g(x) dx .
Rd

Rd

In particular, kfb kL2 = kf kL2 (Plancherals formula).

Proof. We have seen that

Z
Z
iy
b
g() f () e d =
Rd

Setting y = 0, we get
Z

Rd

g() fb () d =

Rd

Rd

gb (x) f (y + x) dx .
gb (x) f (x) dx .

Replacing f by F 1f and using the identity F 1f (x) = fb (x), we obtain

Z
Z
g(x) f (x) =
gb (x) fb (x) dx .
Rd

ifwilde

Rd

Notes

31

The Fourier transform

However, fb (x) = (f )b (x), so putting h = f , we see that fb (x) =

h (x). Hence
(h)b (x) = b
Z

g(x) h(x) dx =

Rd

Rd

as required.

h (x) dx ,
gb (x) b

It is now easy to see that the Fourier transform is a unitary operator on

the Hilbert space L2 (Rd ).
Theorem 4.8 (Plancherel). The Fourier transform F extends from S (Rd )
to a unitary operator on L2 (Rd ).
Proof. We have seen above (Plancherels formula) that the Fourier transform
F : S (Rd ) S (Rd ) is isometric with respect to k k2 (and maps S (Rd )
onto S (Rd )). However, S (Rd ) is dense in L2 (Rd ) and so the result follows
by standard density arguments.
[ The details are as follows. Let h L2 (Rd ). Then there is a sequence (fn ) in S (Rd )
such that kfn hk2 0. In particular, (fn ) is L2 -Cauchy. But k
b k2 = kk2 for
2
S (Rd ) and so fc
is
also
L
-Cauchy
and
therefore
converges
to
some element,
n
F , say, in L2 (Rd ). We define Fh = F . Then
kFhk2 = lim kfc
n k2 = lim kfn k2 = khk2 .
n

To see that that F is independent of the particular sequence (fn ), suppose that
(gn ) is any sequence in S (Rd ) such that kgn hk2 0. Define a new sequence
(n ) in S (Rd ) by setting
(
fn , n odd
n =
gn , n even.
Arguing as above (but with n rather than fn ), we see that the sequence (
cn )
converges in L2 (Rd ). But then
2
cn = L2- lim gc
Fh = L2- lim fc
n = L - lim
n
n

so that Fh is well-defined and kFhk2 = khk2 .

A similar argument holds for the inverse Fourier transform F 1. Moreover,
FF 1 and F 1F are both equal to the identity operator on S (Rd ) which is dense
in L2 (Rd ) and so FF 1 = F 1F=1l on L2 (Rd ). ]

Corollary 4.9. For any f, g L2 (Rd ), we have

Z

Rd

fb (x) g(x) dx =

Rd

f (x) gb (x) dx .

November 9, 2005

32

Chapter 4

Proof. The fact that F is unitary on L2 (Rd ) means that

Z
Z
b
f gb dx =
f g dx .
Rd

Rd

Replacing f by f and g by gb and using the facts that (Ff )(x) = (Ff )(x)
and (FFg)(x) = g(x) we see that
Z
Z
b
f (x) g(x) dx =
f (x) gb (x) dx
Rd

Rd

and the result follows.

Remark 4.10. The unbounded self-adjoint operator iDj on the Hilbert

space L2 (Rd ) is unitarily equivalent to the operator of multiplication by xj .
This follows because F(i)Dj F 1 = xj on S (Rd ) which is a core for the
multiplication operator xj .
Definition 4.11. The Fourier transform FT of the tempered distribution
T S (Rd ) is given by
FT (f ) = T (Ff )

for f S (Rd ).

We often write Tb for FT . Similarly, the inverse Fourier transform F 1T is

given by
F 1T (f ) = T (F 1f ) for f S (Rd ).

Remark 4.12. Note that F:S (Rd ) S (Rd ) is continuous and so the Fourier
transform F maps S (Rd ) into S (Rd ). Similarly, F 1T S (Rd ) for
every T S (Rd ). Evidently, F 1FT = T = FF 1T .
If T is given by some element g of L2 (Rd ), so T = Tg , then
Z
Z
b
b
FTg (f ) = Tg (f ) =
g(x)f (x) dx =
gb (x) f (x) dx = Tgb (f ) .
Rd

Rd

This means that we can think of the Fourier transform on S (Rd ) as an

extension of that on L2 (Rd ).
Examples 4.13.
1. We compute bb for b Rd . For any S (Rd )
bb () = b (
b)
=
b (b)
=

1
(2)d/2

Rd

eibx (x) dx

Rd

eibx
(x) dx
(2)d/2

so that bb = T where (x) = eibx /(2)d/2 . In particular, with b = 0,

we find that b = (2)d/2 .

ifwilde

Notes

33

The Fourier transform

2. We shall determine the Fourier transform F(b ). As above,

d
(
b)
b ) () = b (

= b (
b )

=
b (b)
Z
ix eibx
(x) dx
=
d/2
Rd (2)

and therefore bb = ix eibx /(2)d/2 .

Theorem 4.14. F and F 1 are continuous on S (Rd ).

Proof. Suppose that Tn T in S (Rd ). Then Tn () T () for every
S (Rd ). Hence
cn () = Tn (
T
b ) T (
b ) = Tb ()

cn Tb in S (Rd ). A similar argument holds for the inverse Fourier

so T
transform.

Theorem 4.15. For any T S (Rd ) and multi-indices , Zd+ ,

In general,

(ix) Tb = (D T )b and

D Tb = ( (ix) T )b .

(ix) D Tb = {D ( (ix) T )}b .

Proof. Let S (Rd ). Then

(ix) Tb () = Tb ( (ix) )


= T ((ix ) )b


= T (D)
b
Similarly,

= D T (
b)
= (D T )b () .

D Tb () = Tb ( (D) )
= T ( ((D) )b )
= T ( (ix)
b)

= (ix) T (
b)
= ( (ix) T )b () .

The general case is proved in the same way.

November 9, 2005

34

ifwilde

Chapter 4

Notes

Chapter 5
Convolution

Definition 5.1. Let f, g S (Rd ). The convolution of f and g, denoted by

f g, is the function
Z
(f g)(y) =
f (y x) g(x) dx .
Rd

Theorem 5.2. f g S (Rd ) for any f, g S (Rd ). Moreover,

(2)d/2 fcg = fb gb

and

(2)d/2 fb gb = f[
g .

Furthermore, f g = g f and f (g h) = (f g)h for any f, g, h S (Rd ).

Proof. We have
(2)d/2 fcg (y) =

eiyx f (x) g(x) dx .

Rd

R
R
But we know that
b dx = b dx, for any , S (Rd ). If we let
R
R
denote F 1, then, replacing by ,
we get dx = b dx.
Now, for fixed y Rd , let (x) = eiyx f (x) and set (x) = g(x). Then
we have
Z
Z
(x) (x) dx =
(x)

b (x) dx
Rd
Rd
Z n
Z
o
d/2
=
(2)
eixt eiyt f (t) dt gb (x) dx
d
Rd
ZR
=
fb (y x) gb (x) dx
Rd

giving
as required.

= (fb gb )(y)

(2)d/2 fcg (y) = (fb gb )(y) ,

35

36

Chapter 5

Replacing f by F 1f and g by F 1g in this identity, we find that

(2)d/2 F(F 1f F 1g) = f g
which shows that f g S (Rd ) (because this is true of the left hand side).
The left hand side is unaltered if we interchange f and g and therefore
f g =gf.
Next, taking the Fourier transform once again gives
F(f g)(y) = (2)d/2 FF((F 1f )(F 1g))(y)
= (2)d/2 ((F 1f )(F 1g))(y)

= (2)d/2 F 1f (y) F 1g(y)

= (2)d/2 Ff (y) Fg(y)
and so F(f g) = (2)d/2 (Ff )(Fg), as claimed.
Finally, we have
F(f (g h)) = (2)d/2 fb g[
h
= (2)d fb gb b
h
= F((f g) h)

and therefore (taking the inverse Fourier transform) f (gh) = (f g)h.

Corollary 5.3. If , C0 (Rd ), then C0 (Rd ). Moreover,
supp supp + supp .
Proof. It follows from the theorem that S (Rd ) for , C0 (Rd ).
Now
Z
( )(y) =
(y x) (x) dx
Rd

which certainly vanishes if it is false that y x supp for some x supp ,

that is, if it is false that y = x1 + x2 for some x1 supp and x2 supp .
Hence supp supp + supp , as required.
Moreover, each of supp and supp is compact and so therefore is
supp . We conclude that C0 (Rd ).
Corollary 5.4. For fixed f S (Rd ), the mapping g 7 f g is continuous
from S (Rd ) into S (Rd ).
Proof. Fix f S (Rd ) and suppose gn 0 in S (Rd ). Then also gbn 0
in S (Rd ) and so (by Leibnitz formula) fb gbn 0 in S (Rd ). But then
f gn = (2)d/2 F 1(fb gbn ) 0 in S (Rd ) and the result follows.
ifwilde

Notes

37

Convolution

Definition 5.5. For any function u on Rd , we define the translation x u and

the inversion u
e by the formulae
(x u)(y) = u(y x)

and u
e (y) = u(y) .

Then (x u
e )(y) = u
e (y x) = u(x y) and for u, v S (Rd ), we have
Z
Z
(u v)(y) =
u(x) v(y x) dx =
u(x) (y ve )(x) dx .
Rd

Rd

One readily checks that for fixed x Rd , x and e are continuous maps from
D(Rd ) onto D(Rd ) and from S (Rd ) onto S (Rd ).

Definition 5.6. For u D (Rd ) and D(Rd ), the convolution u is the

function
(u )(x) = u(x
e ) = u((x )) .
For T S (Rd ) and f S (Rd ), the convolution T f is the function
(T f )(x) = T (x fe ) = T (f (x )) .

Note that u() can be expressed as a convolution. Indeed, we see that

= (
e )e = 0 ((
e )e). Hence u() = u(0 ((
e )e)) = (u
e )(0). Similarly,
T (f ) = (T fe )(0).
f (x + ej ) f (x)

where e1 = (1, 0, . . . , 0), . . . , ed = (0, 0, . . . , 0, 1) are the standard basis

vectors of Rd . Then f j f in S (Rd ) as 0.

Lemma 5.7. For f S (Rd ) and 6= 0, set f (x) =

Proof. We shall show that |||f j f |||, 0 for each , Zd+ . Since
D (f j f ) = g j g where g = D f , we may assume that = 0.
P
For notational simplicity, let h = j f S (Rd ) and let kxk1 di=1 |xi |
for x Rd . Let || < 1. By the Mean Value Theorem, for each x Rd there
is some R (depending on x) with || < 1 such that f (x) = h(x + ej )
and so
sup | x (f (x) j f (x)) | = sup | x (h(x + ej ) h(x)) |

xRd

xRd

sup | x (h(x + ej ) h(x)) |

kxk1 M

+ sup | x h(x + ej ) |
kxk1 >M

+ sup | x h(x) | .

()

kxk1 >M

We shall estimate each of the three terms on the right hand side. Let > 0
be given. Since h S (Rd ), the third term on the right hand side is smaller
than 13 for all sufficiently large M .
November 9, 2005

38

Chapter 5

To estimate the middle term, note that |xj | |xj + |+|| |xj + |+1
and so

Y
|xi |i (1 + |xj + | )j .
| x |
i6=j

Furthermore, if kxk1 > M , then kx + ej k1 > M 1. Combining these

remarks, it follows that
sup | x h(x + ej ) |

kxk1 >M

sup
kxk1 >M

Y
i6=j

sup
kxk1 >M 1

<

1
3

|xi |i

Y
i6=j

|xi |i

(1 + |xj + | )j | h(x + ej ) |


(1 + |xj | )j | h(x) |

provided M is sufficiently large (again because h S (Rd )).

Fix M sufficiently large (according to the discussion above) so that each of
the second and third terms on the right hand side of the inequality () is
smaller than 31 . The function h is uniformly continuous on the compact
set { x Rd : kxk1 M } and so the first term on the right hand side of ()
is smaller than 13 for all || sufficiently small. The result follows.
Lemma 5.8. For any f S (Rd ), a f f in S (Rd ) as a 0 in Rd .
Proof. Fix Zd+ and suppose that kak1 < 1. Then
sup | x (f (x a) f (x)) |
x

sup | x (f (x a) f (x)) |

kxk1 M

+ sup | x f (x a) |
kxk1 >M

+ sup | x f (x) |
kxk1 >M

Since f S (Rd ), for any given > 0, we may fix M sufficiently large that
the second and third terms on the right hand side are each smaller than 13 .
But then for all sufficiently small a, the first term is also smaller than 13
because f is uniformly continuous on { x Rd : kxk1 M }. It follows that
supx | x (f (x a) f (x)) | 0 as a 0. Replacing f by D f , we see that
|||a f f |||, 0 as a 0 for any , Zd+ , that is, a f f in S (Rd ) as
a 0.
ifwilde

Notes

39

Convolution

Corollary 5.9. Suppose that D(Rd ). Then

(i) j in D(Rd ) as 0, and
(ii) a in D(Rd ) as a 0.
Proof. We have seen that j in S (Rd ) as 0 and that a in
S (Rd ) as a 0. However, for all || < 1, say, there is some fixed compact
set K such that each supp K (and supp j K) which means that
j in D(Rd ) as 0.
Similarly, for all kak1 < 1, say, the supports of and a all lie within
some fixed compact set and so a in D(Rd ) as a 0.
Theorem 5.10. Let u D (Rd ) and let D(Rd ). Then u C (Rd )
and
D (u ) = (D u) = u (D )
for any Zd+ . Furthermore, supp(u ) supp u + supp .
Proof. By the corollary, it follows that y
e x
e in D(Rd ) if y x in Rd .
Hence u(y
e ) u(x
e ), that is, (u )(y) (u )(x) if y x which
shows that u is continuous on Rd .
Again, using the corollary, we see that for fixed x Rd ,
u( x+ej
e x
e)
(u )(x + ej ) (u )(x)
=

 (
e) 
x ej e
=u

u( x (Dj
e ) ) , as 0,
g
= u( x (D
j ) )
= (u Dj )(x) .

Hence the partial derivative Dj (u )(x) exists at each x Rd and it is

equal to (u Dj )(x). Furthermore, for fixed x,
u( x (Dj
e ) ) = u(Dj x
e)
and therefore

= (Dj u)(x
e)


= (Dj u) (x)

Dj (u ) = u Dj = (Dj u) ,
as required. The general case follows by induction.
November 9, 2005

40

Chapter 5

For the last part, we note that (u )(x) = 0 if supp u supp x

e = , that
is to say, if supp u supp (x ) = . Hence
supp(u ) { x Rd : supp u supp (x ) 6= }

= { x : there is y supp u such that x y supp }

= { x : x supp u + supp }
and the proof is complete.

Corollary 5.11. For u D (Rd ) and D(Rd ), u D (Rd ).

Proof. The function u belongs
to C and so is bounded on each compact
R
subset of Rd . Hence 7 Rd (u )(x) (x) dx is a continuous linear map
on D(Rd ), that is, it is a distribution.
There is an analogous result for S and S .
Theorem 5.12. Let u S (Rd ) and f S (Rd ). Then u f C and
D (u f ) = (D u) f = u (D f )
for any Zd+ . Furthermore, u f is polynomially bounded and hence
determines a tempered distribution.
Proof. The first part is just as for D(Rd ). We need to show that u f
is polynomially bounded. Since u S (Rd ), it follows that there is some
constant C > 0 and integers k, n such that
|u(g)| C kgkk,n
for all g S (Rd ). Hence, for x Rd ,
|(u f )(x)| = | u(x fe ) |
C kx fe kk,n
X
=C
sup |y | | Dy f (x y) |
y

||k
||n

=C

||k
||n

sup |(y + x) | | D f (y) |

y

which is polynomially bounded in x.

ifwilde

Notes

41

Convolution

Proposition 5.13. Let T S (Rd ) and suppose that fn f in S (Rd ) as

n . Then (T fn )(x) (T f )(x) uniformly on compact sets in Rd .
Proof. By replacing fn by fn f , we may assume that f = 0. Now, since
T S (Rd ), there is C > 0 and k, n Z+ such that
X
| T (g) | C
|||g|||,
||k
||n

for all g S (Rd ). Therefore

| (T fn )(x) | = | T (x ff
n )| C

||k
||n

|||x ff
n |||, .

However, if kxk1 M , say, then

|||x ff
n |||, = sup | y Dy fn (x y) |
y

sup |(y + x) | |D fn (y)|

y

sup
y

d
Y
i=1

(M + |yi |)i

|D fn (y)|

as n (because fn 0 in S (Rd )). It follows that (T fn )(x) 0

uniformly on { x : kxk1 M } for any fixed M > 0 which establishes the
result.
Theorem 5.14. Let u D (Rd ) and let , D(Rd ). Then
(u ) = u ( ) .
Proof. We first observe that the statement of the theorem makes sense
because u C and C0 (Rd ).
Let > 0 and consider the Riemann sum
X
f (x) = d
(x ) () .
Zd

This is always a finite sum because the functions and have compact
support. Furthermore, supp f supp + supp and f D(Rd ). Now,
using the uniform continuity of D , we see that
X
D f (x) = d
D (x ) () ( (D ) )(x) = D ( )(x)
Zd

November 9, 2005

42

Chapter 5

uniformly as 0. Hence f in D(Rd ) as 0 and therefore

x fe x (]
) in D(Rd ) as 0. It follows that
u ( )(x) = u( x (]
))
= lim u(x fe )
0
X
= lim d
u( (x ) () )
0

= lim d
0

Zd

Zd

( u )(x ) ()

= ( (u ) )(x)
and the proof is complete.
Corollary 5.15. Let T S (Rd ) and let f, g S (Rd ). Then
(T f ) g = T (f g).

Proof. Since C0 (Rd ) is dense in S (Rd ), there are sequences (n ) and (n )

in C0 (Rd ) such that n f and n g in S (Rd ). Furthermore, since
T C0 (Rd ) D (Rd ), it follows from the theorem that
( T n ) k = T ( n k ) .
However, we know that (T n )(x) (T f )(x) uniformly on compact sets
in Rd and so, for fixed y Rd ,
fk )
( (T n ) k )(y) = (T n )(y
Z
=
(T n )(x) k (y x) dx ,
Rd

since T n C (Rd ) (and is polynomially bounded),

Rd

(T f )(x) k (y x) dx ,

since supp k is compact,

= (T f ) k (y) .
On the other hand, n k f k as n in S (Rd ) and so
(T n ) k = T (n k ) T (f k ) .

It follows that (T f ) k = T (f k ) for each k. But k g in S (Rd ),

so

and

fk ) (T f )(y ge ) = (T f ) g(y)
(T f ) k (y) = (T f )(y

T (f k )(y) = T ( y (f^
k ) ) T ( y (f]
g ) ) = T (f g)(y) .
Hence (T f ) g = T (f g), as required.
ifwilde

Notes

43

Convolution

R
Theorem 5.16. Let g S (Rd ) be such that Rd g(x) dx = 1 and for 6= 0
set g (x) = d g(x/). Then for any f S (Rd ), g f f in S (Rd ),
as 0.
Proof. It is enough to show that (g f )b fb in S (Rd ) (the result then
follows by taking the inverse Fourier transform). However, we know that
(g f )b = (2)d/2 gb fb so we must show that ((2)d/2 gb 1) fb 0 in
S (Rd ), as 0.
Using Leibnitz formula together with the fact that D fb () S (Rd )
for any , Zd+ , it is enough to show that


D ( (2)d/2 gb 1 ) 0

uniformly on Rd , as 0,

for any S (Rd ) and Zd+ . Note that

gb () =

1
(2)d/2

Rd

eix g( x )

1
d

dx = gb () .

(So we see that gb () gb (0) = (2)d/2 as 0.)

We consider two cases.

(i) Suppose || = 0. Fix S (Rd ). Then

| ( (2)d/2 gb () 1 ) () | = | ( (2)d/2 gb () 1 ) () |
Z



=
g(x) ( eix 1 ) dx () 
d
Z R

|g(x)| |x| dx |()|

Rd
Z
=
|xg(x)| dx | ()|
Rd

< M

for some constant M > 0 independent of . So ( (2)d/2 gb ()1 ) () 0

uniformly in , as 0.
(ii) Suppose that || > 0. For fixed S (Rd ), we have





 D ( (2)d/2 gb 1 ) () ()  = | (2)d/2 || (D gb )() () |
< || M

for some constant M > 0, since both D gb and are bounded on Rd .

The result follows.

November 9, 2005

44

Chapter 5

R
Theorem 5.17. Let T S (Rd ) and let g S (Rd ) with Rd g(x) dx = 1.
Then T g T in S (Rd ) as 0, where g (x) = d g(x/).
R
If u D (Rd ) and D(Rd ) with Rd (x) dx = 1, then u u in
D (Rd ) as 0, where (x) = d (x/).
Proof. Fix f S (Rd ). Then
(T g )(f ) = (T g ) fe (0)
= T (g fe )(0)
T fe (0) ,

by the previous theorem,

= T (f )

which proves the first part.

Next, we note that for given C0 (Rd ) and 0 < < 1, say, the supports
of and all lie in some fixed compact set (independently of ). Hence
in D(Rd ) as 0. Arguing now as above, we deduce that for any
u D (Rd ) and D(Rd ), (u )() u() as 0, that is, u u
in D (Rd ).
Remark 5.18. The infinitely-differentiable function T g is called the regularization of T . This is easier to deal with than T itself, but of course, one
must eventually take the limit 0 in order to recover the distribution T .
Theorem 5.19.
(i) C0 (Rd ) is dense in D (Rd ), that is, for any u D (Rd ) there is
some sequence (fn ) in D(Rd ) such that fn u in D (Rd ).
(ii) C0 (Rd ) is dense in S (Rd ), that is, for any T S (Rd ) there is
some sequence (fn ) in S (Rd ) such that fn T in S (Rd ).
R
Proof. (i) Fix u D (Rd ) and let C0 (Rd ) be such that (x) dx = 1
and set (x) = d (x/). For n N and t 0, let the function n (t) be
as shown in the diagram.
(t)
1

n+1

Figure 5.1: The function (t).

ifwilde

Notes

45

Convolution

For x Rd , let n (x) = n (|x|). Evidently, n C0 (Rd ) and n (x) = 1 for

|x| n and n (x) = 0 when |x| n + 1. Now, n u has compact support
and so n u 1/n C0 (Rd ). We claim that n u 1/n u in D (Rd ) as
n .
Indeed, for D(Rd ), we have
( n u 1/n )() = ( n u 1/n ) e (0)
= u (1/n e )(0)
n

= n u(
e 1/n ))

= u( n (
e 1/n ))

= u(
e 1/n ) ,

for all sufficiently large n,

= u( (1/n e ) e )
u((e ) e )
= u() ,

as required.
(ii) The functions n are supposed to be smooth and obey the requirements
that n (t) = 1 when 0 t n, n (t) = 0 for t n + 1 and n+1 (t) =
n (t 1) for n + 1 t n + 2. As n increases, so the graph of n extends
out but maintains its general shape as it decreases from 1 to 0. The point is
that the n s and any derivatives are bounded independently of n. That is,
(k)
for any k = 0, 1, 2, . . . , there is Mk > 0 such that supn supt0 |n (t)| < Mk .
Let T S (Rd ) and let f S (Rd ). Then, with notation as in part (i), we
note that n T 1/n C0 (Rd ). We claim that n T 1/n T in S (Rd )
as n .
To see this, first we observe that n (1/n f ) f in S (Rd ). This follows
from the inequalities
k n (1/n f ) f k, k n (1/n f f )k, + k n f f k,
for , Zd+ , together with (Leibnitz formula and) the bounds on kD n k
for each fixed Zd+ uniformly in n.
Similarly, n (1/n fe )e f in S (Rd ) and therefore
( n T 1/n )(f ) = T ( n (1/n fe )e ) ,

as in part (i),

T (f ) ,

as n , that is, n T 1/n T in S (Rd ) and the proof is complete.

November 9, 2005

46

Chapter 5

Remark 5.20. This result ties up the approach to distributions taken here
with the generalized function approach in which distributions are defined
via sequences of functions in S (Rd ). If T S (Rd ) and fn S (Rd ) is
such that fn T in S (Rd ), then the generalized function approach would
be to consider T to be the sequence (fn ) (or strictly speaking, equivalence
classes of such sequences so as to allow for different sequences in S (Rd )
which converge in S (Rd ) to the same distribution). This is in the same
spirit as defining real numbers via Cauchy sequences of rational numbers.
The next result tells us that, under the Fourier transform, convolution
becomes essentially multiplication.
Theorem 5.21. For any T S (Rd ) and f S (Rd ),
(i) F(T f ) = (2)d/2 Ff FT and
(ii) FT Ff = (2)d/2 F(f T ).
Proof. (i) We know that there is a sequence (n ) in C0 (Rd ) such that
n T in S (Rd ). So for given g S (Rd ), we have
(T f )b (g) = (T f )( gb ) = (T f ) f
b
g (0)
= T (f f
b
g )(0) = T ( ( f f
b
g )e )
= T ( fe gb )

= lim n ( fe gb ) = lim n ( f f
b
g )(0)
n

= lim ( n f )( gb ) = lim ( n f )b (g)

n
n
Z
d/2
b
= (2)
lim (
bn f )(g) = lim

bn (x) fb (x) g(x) dx

n
n
Rd
Z
n (x) (fb g)b (x) dx
= lim
n

Rd
d/2

T ( (fb g)b ) = (2)d/2 Tb ( fb g )

= (2)d/2 (fb Tb )(g) ,

= (2)

that is, (T f )b = (2)d/2 fb Tb , as required.

The second part can be established in a similar way.

ifwilde

Notes

Chapter 6
Fourier-Laplace Transform

We have seen that any tempered distribution has a Fourier transform which
is also a tempered distribution. We will see here, however, that the Fourier
transform of a distribution with compact support is actually given by a
function. By way of motivation, we note that the Fourier transform of an
integrable function u, say, is given by
d/2

u
b (y) = (2)

Rd

ey (x) u(x) dx

where ey denotes the function x 7 eiyx for x Rd . We cannot write this in

the distributional sense as Tu (ey ) because the function ey does not belong to
D(Rd ) (or S (Rd )). However, if u has compact support, then we can write
u
b (y) = (2)d/2

Rd

ey (x) u(x) dx = (2)d/2

Rd

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

where C0 (Rd ) is chosen such that = 1 on some open set containing

supp u. In this case, we see that
u
b (y) = (2)d/2 Tu (ey ) .

This formula makes sense if y Rd is replaced by any z Cd . The following

definition seems appropriate.
Definition 6.1. Suppose that u D (Rd ) and that u has compact support.
The Fourier-Laplace transform of u is the function u
b () on Cd given by
u
b (z) = (2)d/2 u(ez )

where ez (x) = eizx and C0 (Rd ) is such that = 1 on some open

set W with supp u W .
47

48

Chapter 6

We note straightaway that ez D(Rd ) and that if W1 and W2 are

open sets with supp u W1 and supp u W2 and if 1 C0 (Rd ) and
2 C0 (Rd ) are such that 1 = 1 on W1 and 2 = 1 on W2 , then it
follows that u(ez 1 ) = u(ez 2 ) so that u
b is a well-defined function on Cd .
Furthermore, if u is given by an integrable function, then, as discussed above,
u
b (y) is the Fourier transform of u.
Now, any element of D (Rd ) with compact support determines a tempered
distribution and for any u S (Rd ) we have already defined its Fourier
transform as the tempered distribution u
b : f 7 u(fb ). We can then ask
whether there is any relationship between the tempered distribution u
b and
d
the function u
b (x), x R . If u is given by a square-integrable function, then
by Plancherels Theorem, Corollary 4.9, we find that for any f S (Rd ),
u(fb ) =

u(x) fb (x) dx =

u
b (x) f (x) dx .

This shows that in this case, the tempered distribution u

b is indeed given by
the function u
b (x). This is true in general, as we now show.

Theorem 6.2. Let u D (Rd ) and suppose that u has compact support. Then
the Fourier-Laplace transform u
b (z) is an entire function whose restriction
to Rd determines the tempered distribution u
b , that is, u
b S (Rd ) is given
by the function u
b () Rd .

Proof. Let C0 be such that = 1 on some open set W Rd with

supp u W . Then u
b (z) = (2)d/2 u(ez ) for z Cd and
ez (y) (y) = eizy (y) =

(i)
!

z y (y)

Zd+

for any y Rd . Since supp is compact, it follows that for fixed z Cd

the partial sums converge uniformly in y and the same is true of any partial
derivatives (with respect to y). In other words, the partial sums converge
in D(Rd ) and so
X (i)

u(ez ) =
! z u(g )
Zd+

where g D(Rd ) is the function g (y) = y (y).

Moreover, u has finite order by Theorem 3.14, so that |u()| C ||||||N
for some constants C > 0 and N Z+ and any D(Rd ). It follows that
|u(g )| C R|| for constants C and R (depending, of course, on u and
supp u). The series above for u(ez ) therefore converges absolutely for all
z Cd and so u(ez ) is entire.
ifwilde

Notes

49

Fourier-Laplace Transform

We must now show that u

b , the tempered distribution, is given by the
function u
b (x), x Rd .

Now, u = u as distributions, that is,

u() = ( u)() = u( )
for all S (Rd ) and so

It follows that

u
b = ( u)b = (2)d/2 u
b b .
u
b () =

(2)d/2 (b
u b )(y) (y) dy ,

that is, u
b is determined by the C -function (2)d/2 (b
u b )(x). Our aim
is to show that (2)d/2 (b
u b )(x) = u
b (x). To see this, let S (Rd ) be
such that
b = . Then
b )(x)
(b
u b )(x) = (b
u
b
= (b
u
e )(x)
=u
b (x )

= u( (x )b )

= u(ex
b)

= u(ex )

and the result follows.

= (2)d/2 u
b (x)

Next we consider the relationship between the support of u and growth

properties of the function u
b (z). We need a preliminary result.

Lemma 6.3. Suppose that f is analytic on Cd and vanishes on Rd . Then

f = 0 everywhere.
Proof. Let z1 , z2 , . . . , zd Cd . Fix a2 R, . . . , ad R and consider the
map z 7 f (z, a2 , . . . , ad ). This is entire and vanishes on R and so vanishes
everywhere on C, by the Identity Theorem. Since a2 R is arbitrary,
we may say that f (z1 , a2 , . . . , ad ) = 0 for all a2 R. But then the map
z 7 f (z1 , z, a3 , . . . , ad ) is entire and vanishes on R and so vanishes on C. In
particular, f (z1 , z2 , a3 , . . . , ad ) = 0 for any a3 R.
Continuing in this way, we see that f (z1 , z2 , . . . , zd ) = 0 and the result
follows.
For the following, let Kr denote the closed ball Kr = { x Rd : |x| r }.
November 9, 2005

50

Chapter 6

Theorem 6.4.
(a) Suppose that C0 (Rd ) and that supp Kr .
Then
(1)

f (z) = (2)d/2

Rd

eizt (t) dt, for z Cd , is entire

and there are constants N such that

(2)

|f (z)| N (1 + |z|)N er |Im z| for z Cd and N = 0, 1, 2, . . . .

(b) Conversely, if an entire function f satisfies (2), then there is some

C0 (Rd ) with supp Kr such that (1) holds.
Proof. (a) First note that if z = x + iy Cd , then
|eizt | = eyt e|y| |t| er| Im z|
for all t Rd with |t| r (and where | Im z| = (y12 + + yd2 )1/2 ). It
follows that eizt (t) is integrable for each z Cd . Moreover, for each fixed
z Cd , the power series expansion for eizt converges uniformly for t Kr
and therefore
Z
X (iz) t
d/2
dt
f (z) = (2)
(t)
!
Rd
d
Z+
Z
X
(iz) t
(2)d/2
(t)
=
dt .
!
Rd
d
Z+

Now
Z
 Z



(t)
t
dt

 |(t)| r1 . . . rd dt
= r|| kkL1

and so we see that the series expression for f (z) converges absolutely for all
z Cd and so f is analytic on the whole of Cd .
Next, integrating by parts, we find that
Z
i z f (z) = (2)d/2
(t) (iz) eizt dt
Z
= (2)d/2 (D )(t) eizt dt
and so
|z | |f (z)| (2)d/2 kD kL1 er| Im z| .
ifwilde

Notes

51

Fourier-Laplace Transform

This, together with the inequality (1 + |z|)N (1 + |z1 | + + |zd |)N implies
that
(1 + |z|)N |f (z)| N er| Im z|

for a suitable constant N , which is (2).

(b) Suppose that f is entire and satisfies the inequalities (2). For t Rd , let
Z
d/2
(t) = (2)
f (x) eitx dx .
Rd

Since (1 + |x|)N f (x) is bounded on Rd for any N (by (2)), it follows that
is a well-defined function and that C (Rd ).
We wish to show that it is possible to replace x by x + iy in this formula
for without any other changes. To see this, let
Z
f (x + i, z2 , . . . , zd ) e(t1 (x+i)+t2 z2 ++td zd ) dx
I() =

where t1 , . . . , td R, z2 , . . . zd C and R.
We shall show that I() = I(0) which shows that, in fact, I does not
depend on . Let be the (closed, simple) rectangular contour in C with
vertices at the points X and X + i, where X > 0. Since the function
7 f (, z2 , . . . , zd ) e(t1 +t2 z2 ++td zd ) is analytic, it follows that
Z
f (, z2 , . . . , zd ) e(t1 +t2 z2 ++td zd ) d = 0 ,
()

by Cauchys Theorem. Now we use (2) to estimate the integrand along the
vertical sides of the rectangular contour .
We have, with z = (X + iy, z2 , . . . , zd ),


 f (X + iy, z2 , . . . , zd ) ei(t1 (X+iy)+t2 z2 ++td zd ) 

N er| Im z| et1 y et2 | Im z2 | etd | Im zd |

N
1 + ( | X + iy|2 + |z2 |2 + + |zd |2 )1/2
N
er| Im z| et1 y et2 | Im z2 | etd | Im zd |

N
( 1 + | X + iy| )
N

er| Im z| et1 y et2 | Im z2 | etd | Im zd |

N
(1 + X)
N
er| Im z| et1 y et2 | Im z2 | etd | Im zd |

N
(1 + X)
0

as X for all |y| ||. It follows that the part of the contour integral
along the vertical sides of converges to zero, as X and so from ()
we conclude that I(0) I() = 0, as required.
November 9, 2005

52

Chapter 6

Repeating this argument coordinate by coordinate, we deduce that

Z Z
d/2
(t) = (2)
...
f (x) eitx dx1 dx2 . . . dxd

Z Z
d/2
= (2)
...
f (x + iy) eit(x+iy) dx1 dx2 . . . dxd

for any y Rd .

Now, let N N be such that (1 + |x| )N L1 (Rd ). Then, using (2),

together with the inequality (1 + |x + iy| )N (1 + |x| )N , we find that
Z
(1 + |x|)N er|y| ety dx
|(t)| (2)d/2 N
d
R
Z
d/2
(r|y|ty)
(2)
N e
(1 + |x|)N dx
Rd

for all y Rd . That is, there is some constant C > 0 such that
|(t)| C er|y|ty
for any y Rd . We shall show that this implies that supp Kr . Indeed,
let t Rd be fixed such that |t| > r. Setting y = t with > 0, it follows
that ty = |t|2 and we see that
|(t)| C e|t|(r|t|)
for any > 0. Letting , it follows that (t) = 0 and so we conclude
that supp Kr , as claimed.
By the Fourier inversion theorem, we have
Z
d/2
f (x) = (2)
(t) eixt dt .

R
It follows that the entire function z 7 g(z) (t) eizt dt agrees with f
on Rd and so, by the Lemma, f = g on Cd , that is, f is given as in (1) and
the proof is complete.
There is a version of this result for distributions, as follows.
Theorem 6.5.
(a) Suppose that u D (Rd ) and that supp u Kr . Then the FourierLaplace transform f (z) = u
b (z) is entire, its restriction to Rd is the
Fourier transform of u, and there is a constant > 0 such that
|f (z)| (1 + |z| )N er|Im z|

()

where N is the order of u.

ifwilde

Notes

53

Fourier-Laplace Transform

(b) Conversely, if f is entire and satisfies () for some N N, some > 0

and r > 0, then f (z) = u
b (z) for some u D (Rd ) with supp u Kr .

Proof. (a) We have already proved everything except for the estimate ().
Let h C (R) be such that h(s) = 1 when s 1 and h(s) = 0 for all s 2.
For given z Cd (with z 6= 0), let
(x) = h( |x| |z| r |z| ) .

Then C (R) and (x) = 0 whenever |x| |z|r |z| > 2, that is vanishes
2
+ r. So C0 (R) with supp Kr+2/|z| . Furthermore,
whenever |x| > |z|
1
, then (x) = 1 and so u
b (z) = u(ez ) (since = 1 on some
if |x| < r + |z|
open set W with sup u W ).
By hypothesis, supp u is compact and so u has finite order, say N . Then
there is C > 0 such that
X
| u(ez ) | C
kD (eizx (x))k .
||N

Now h/xj = |z|

xj
h (|x| |z| r |z|) and so for any , Zd+ , we have
|x|

 izx
 
 (D e
)(D (x))  =  (iz) eizx |z|||

x
|x|||


D|| 

|z|||+|| k D|| h k e|Im z|(r+2/|z|)

(since |zj j | |z|j and if x supp then |x| r + 2/ |z|)

|z|||+|| k D|| h k er|Im z|+2

(since |Im z| |z|) .

Hence, by Leibnitz formula, we conclude that

|b
u (z)| (1 + |z| )N er|Im z|
for a suitable constant .
(b) Suppose that f satisfies |f (x)| (1 + |x| )N er Im z for some constants
, r > 0 and some N N. Then the map from S C given by
Z
7 f () = f (x) (x) dx
is a tempered distribution. Hence there is some T S (Rd ) such that
Tb = f (namely, T = F 1f ). We will show that supp T Kr and that
f (z) = Tb (z). The idea is to regularize T and apply the Paley-Wiener
theorem for functions.
November 9, 2005

54

Chapter 6

R
Let g C0 (Rd ) be such that supp g K1 and g(x) dx = 1. For > 0,
set g (x) = ed g(x/). Then supp g K and we have seen that for any
tempered distribution S S (Rd ), it is true that S g S in S (Rd ) as
0. Now
T g = (F 1f ) g = F 1((2)d/2 gb f )
since (S )b = (2)d/2
b Sb , for any S (Rd ) and S S (Rd ). Since
supp g K , the Paley-Wiener Theorem implies that gb is entire and that
for any integer k there is a constant k such that
Z
d/2
| gb (z) | = | (2)
g (y) eizy dy |
k (1 + |z| )k e |Im z|

for any z Cd . But, by hypothesis,

|f (z)| (1 + |z| )N er |Im z|
and so, for any integer m,
| gb (z) f (z) | N +m (1 + |z| )N m (1 + |z| )N e(r+) |Im z|
= N +m (1 + |z| )m e(r+) |Im z|

for z Cd .
Again, by the Paley-Wiener Theorem, it follows that there is C0 (Rd )
with supp Kr+ such that
Z
gb (z) f (z) = (2)d/2
eizt (t) dt .

In particular, gb (x) f (x) S (Rd ) and is the Fourier transform of . That

is, F 1(gb f ) = as functions and so therefore as tempered distributions.
Now, let S (Rd ) be such that supp Kr =R. Then, for all > 0
sufficiently small, supp Kr+ = and therefore (x) (x) dx = 0 for
small . It follows that
T () = lim T g ()
0
Z
d/2
(x) (x) dx
= lim (2)
0

=0

and so supp T Kr , as required.

It remains to verify that Tb (z) = f (z) on Cd . However, Tb is entire and
Tb (x) is the Fourier transform of T , Tb (x) = f (x) for x Rd . But, by
hypothesis, f is entire and so Tb = f on Rd means that Tb = f on Cd and
the proof is complete.
ifwilde

Notes

Chapter 7
Structure Theorem for Distributions

Any polynomially bounded continuous function certainly determines a tempered distribution. It turns out that such functions, together with their
distributional derivatives exhaust S (Rd ), as we discuss next.
Theorem 7.1. Let T S (Rd ). Then there is a continuous polynomially
bounded function F (x) and a multi-index Zd+ such that T = D F .
Proof. For notational convenience, we shall only consider the one-dimensional
case, d = 1. So let T S (R) be given. Then we know that there exist
k, m N and C > 0 such that
X
|||f |||,
|T (f )| Ckf kk,m = C
k
m

for all f S (R). It follows that there is C > 0 such that

X
sup | (1 + x2 )k D f (x) |
|T (f )| C
m

Rx
for all f S (R). Now, if C0 (R), then (x) = (t) dt and so
Z x
Z
 
 


 (t) dt = k kL1 .
|(x)|
(t) dt

Hence, for any C0 (R), supx |(x)| k kL1 and so

X


k D (1 + x2 )k D (x) kL1
|T ()| C
m

since (1 + x2 )k D C0 (R). Using the inequality

D(1 + x2 )k = k 2x(1 + x2 )k1 k(1 + x2 )k
we get


| D (1 + x2 )k D (x) | k | (1 + x2 )k D (x) | + | (1 + x2 )k D+1 (x) |
55

56

Chapter 7

and so there is C > 0 such that

X
k (1 + x2 )k Dj (x) k
|T ()| C

()

jm+1

for C0 (R).
Let J : C0 (R) L1 (R) L1 (R) (with (m + 2) terms) be the map
given by
J() = (1 + x2 )k (x) (1 + x2 )k D(x) (1 + x2 )k Dm+1 (x) .
Note that J is one-one, because if J() = J(), then, in particular, the
first components agree and so (1 + x2 )k (x) = (1 + x2 )k (x) which means
that = . It is also clear that J is a linear map. Now we define the map
: J(C0 (R)) C by setting
(J()) = T () .
Then is well-defined, since J() is uniquely determined by , and is linear.
Moreover, the bound () implies that is a bounded linear functional on
J(C0 (R)) when considered as a subspace of (L1 (R))m+2 .
By the Hahn-Banach theorem, has an extension to a bounded linear
functional on the whole of (L1 (R))m+2 . But then must have the form
(f0 f1 fm+1 ) =

m+1
XZ
j=0

gj (x) fj (x) dx
R

for suitable g0 , . . . , gm+1 L (R). Hence

T () = (J) =

m+1
XZ
j=0

gj (x) (1 + x2 )k Dj (x) dx .

and so, as distributions,

T =

m+1
X
j=0

(1)j Dj (1 + x2 )k gj (x)

on C0 (R). But C0 (R) is dense in S (R) and therefore T has this same
form on S (R).
For each j, set
Z x
hj (x) =
(1 + t2 )k gj (t) dt .
0

Evidently hj is continuous and polynomially bounded (because gj is bounded)

and
m+1
X
(1)j Dj+1 hj .
T =
j=0

ifwilde

Notes

57

Structure Theorem for Distributions

To obtain the stated form for T , define fj by

fj (x) =

dtm+1j . . .

t2

dt1 hj (t1 )

i.e., by integrating hj (m + 1 j) times. Then fj is continuous, polynomially

bounded and Dm+2 fj (x) = Dj+1 hj and so
X
T =
(1)j Dm+2 fj .
jm+1

P
Set F (x) = jm+1 (1)j fj (x). Then F is continuous, polynomially bounded
and T = Dm+2 F .
Theorem 7.2. Let u D (). Then for any compact set K , there exists
a continuous function F and a multi-index such that u = D F on C0 (K).
Proof. Let u D () and let K with K compact. Let C0 ()
be such that K W supp for some open set W with = 1 on W .
Then u = u on C0 (K). However, u has compact support and so defines
a tempered distribution and therefore has the form u = D F for some
continuous function F and Zd+ . Hence, for all C0 (K),
u() = u() = D F () ,
as required.

November 9, 2005

58

ifwilde

Chapter 7

Notes

Chapter 8
Partial Differential Equations

We will make a few sketchy remarks in this last chapter. We have seen that
every distribution is differentiable (in the distributional sense) and so we can
consider (partial) differential equations satisfied by distributions. Indeed, we
might expect a differential equation to have a distribution as solution rather
than a function.
Example 8.1. Consider the differential equation u = H where
(
0, x < 0
H(x) =
1, x 0

is the Heaviside step-function. For x < 0, u (x) = 0, so u(x) = a and

for x > 0, u (x) = 1 giving u(x) = x + b, for suitable constants a and b.
However, continuity of u at x = 0 would require u(0) = a = b and so
(
a, x < 0
u(x) =
1, x + a 0.
But such u is not differentiable at x = 0 and so cannot satisfy the original
differential equation at this point. However, for any S (R), we see
that u () = H(), that is, u is a distributional solution to the differential
equation. Indeed, integrating by parts, we get
Z
u(x) (x) dx
u () = u( ) =

Z
(x + a) (x) dx
a (x) dx

Z 0
= a(0) + a(0)
x (x) dx
0
Z
=
(x) dx
0
Z
=
H(x)(x) dx

= H() .

59

60

Chapter 8

Notation. Suppose that P (x1 , . . . , xd ) is a polynomial in d-variables. Then

the symbol P (D) denotes the partial differential operator obtained after the

for xj in the polynomial expression P (x, . . . , xd ). For

substitution Dj = x
j
example, if P (x1 , . . . , xd ) = x31 + x3 x4 , then P (D) =

3
x31

2
x3 x4

For a given function g on Rd and polynomial P , a distributional solution

u D (Rd ) to the partial differential equation P (D)u = g is called a weak
solution. We have seen above that a partial differential equation may possess
a weak solution but no solution in the classical sense. Notice that a partial
differential equation such as P (D)u = g is meaningful even if g D (Rd ).
Of particular interest is the case for which g = .
Definition 8.2. A distribution E D (Rd ) satisfying the partial differential
equation P (D)E = is said to be a fundamental solution for the partial
differential operator P (D).
The importance of fundamental solutions is their part in the solution of
inhomogeneous partial differential equations of the form P (D)u = g, with
g C0 (Rd ). Indeed, if we set u = E g, where E is a fundamental solution
for P (D), then u C0 (Rd ) and we find that
P (D)u = P (D)(E g)

= (P (D)E ) g
=g

= g,

that is, P (D)u = g. So E g is a solution in the classical sense.

Example 8.3 (Poissons Equation). Consider u = g (in R3 ).
1
. To see this,
We claim that a fundamental solution for is E = 4|x|

3
let C0 (R ). Then
(E)() = E()
Z
1
p
=
dx1 dx2 dx3
2
2
2
3
R 4 x1 + x2 + x3
Z
1
r2 cos sin dr d d
=
4r
3
R
Z
1
= lim
r2 dr dS
0
4r
r
where dS = cos sin d d. Integrating by parts with respect to r, one
finds (after some manipulation) that
Z
Z




2
2
2
2
E(r) r dr =
E(r) r dr r E(r) r 
+ r r E(r)
.
r

ifwilde

r=

r=

Notes

61

Partial Differential Equations

But E = 0 for x 6= 0, so the first term on the right hand side above
vanishes and we have
ZZ
o
nZZ
2
r E(r) r2 dS .
E(r) r r dS
(E)() = lim
0

r=

Now, the first term in brackets gives zero in the limit 0 (r is bounded)
and the second can be written as
ZZ
1
r2 d d
lim
(r, , )
0
4r2
r=
ZZ
ZZ

1
1
=
(0) 4 d d + lim
(, , ) (0) 4
d d
0

= (0) + 0

since is continuous at 0. We have shown that (E)() = (0) for any

C0 (R3 ), so that E = as required.
Example 8.4. The heat operator (or diffusion operator) is P (D) = t
where is the Laplacian in Rd (so we are working with (x, t) Rd R).
One checks that
E(x, t) = H(t)

2
1
1
e|x| /4t
d
d/2
2 (t)

satisfies P (D)E = 0 on Rd (R \ { 0 }). We claim that E is a fundamental

solution for P (D).
To see this, let C0 (Rd R). Then
(P (D)E)() = ( (t )E )()

= E((t + ))
Z
= E(x, t) (t + ) dx dt

noting that E is locally integrable,

Z
E(x, t) (t + ) dx dt
= lim
0 t
Z
= lim
E(x, ) (x, ) dx
0

integrating by parts and using P (D)E = 0

on Rd (R \ { 0 }),
Z
2
lim (21/2 y, ) e|y| dy

d/2 0

changing variable, x = 21/2 y,

November 9, 2005

62

Chapter 8

1
d/2

= (0) ,

(0, 0) e|y| dy

so P (D)E = , as claimed.
Of course, one might enquire as to the existence of fundamental solutions.
We state the following theorem, without proof.
Theorem 8.5 (Malgrange-Ehrenpreis). For every constant coefficient partial
differential operator P (D) on Rd , there is a distribution E D (Rd ) such
that P (D)E = .
Suppose now that we can show that a particular partial differential equation
has a weak solution. Is it possible to show that under certain circumstances
this solution is actually a solution in the classical sense? A result in this
vein is the following.
Theorem 8.6. Let Rd be an open set and suppose that u and f are
continuous on and that Dj u = f as distributions. Then Dj u = f in the
classical sense, that is, Dj u exists (as a function) and is equal to f on .
Proof. Let W be any open ball in and let C0 () be such that = 1
on W . Then u is continuous and has compact support. Also, for any
C0 (),
(Dj (u))() = (u)(Dj )
= u(Dj )

= u(Dj ()) + u(Dj )

= (Dj u)() + u(Dj )

= (Dj u)() + ((Dj )u)() ,

that is, Dj (u) = (Dj )u + (Dj u), as distributions.
Let v = u and g = Dj v = (Dj )u + (Dj u). Then g = (Dj )u + f
since Dj u = f . So g is continuous, as is v and both v and g have compact
support. We have shifted the problem to theR case of compact support.
Let C0 (Rd ) be such that 0 and (y) dy = 1. For > 0 let
Z
v (x) = v(x y) (y) dy
Z


1
= d v(y) xy
dy .

Then v C (Rd ) and v v uniformly as 0. Also,

Z
Dj v (x) = 1d v(y) Dxj ( xy
) dy

ifwilde

Notes

63

Partial Differential Equations

1d

1
d

1
d

v(y) Dyj ( xy
) dy

(Dj v)(y) ( xy
) dy
integrating by parts,
g(y) ( xy
) dy

g .

It follows that g = Dj v g uniformly as 0. Let e1 , . . . , ed denote the

usual orthonormal basis vectors for Rd (so that ej = (0, 0, . . . , 0, 1, 0, . . . , 0),
with the 1 in the j th coordinate position). Then
v (x + ej ) v (x) =
=

xj +

Dj v (y) dyj

xj
Z xj +

g (y) dyj .

xj

Letting 0, we deduce that

v(x + ej ) v(x) =

xj +

g(y) dyj

xj

and hence Dj v exists and Dj v(x) = g(x). But for any x W , we have
v(x) = (x)u(x) = u(x) and Dj v(x) = Dj u(x) and g(x) = f (x). It follows
that Dj u exists on W and Dj u = f on W . Since W is arbitrary, the result
follows.
P
Definition 8.7. Let P (D) = ||m a D be a linear differential operator
of orderP
m, with constant coefficients, defined on Rd . Then the polynomial
P () = ||m a is called the symbol of P . The sum of those terms of
order m in P () is called the principal symbol of P , denoted P , that is,
P () =

a .

||=m

Note that P is homogeneous of degree m. The differential operator P (D)

is said to be elliptic if P () 6= 0 for all 0 6= Rd .
Examples 8.8.
1. is elliptic on Rd .
2. 1 + i2 is elliptic on R2 .
November 9, 2005

64

Chapter 8

Theorem 8.9 (Elliptic Regularity). Let Rd be an open set and let P (D)
be an elliptic differential operator (constant coefficients). If v C () and
u D () is a weak solution to P (D)u = v, then u C ().
In particular, every weak solution to the homogeneous partial differential
equation P (D)u = 0 belongs to C ().
We will not prove this here.
Example 8.10. Suppose that P (D) is elliptic (constant coefficients) and that
E is a fundamental solution: P (D) = . Then P (D)E = 0 on Rd \ { 0 },
that is (P (D)E)() = 0 for all C0 (Rd \ { 0 }). So E C0 (Rd \ { 0 }).
Example 8.11. Let be an open set in R2 and suppose that u D ()
satisfies (1 + i2 )u = 0 on . Since 1 + i2 is elliptic, the theorem tells
us that u C0 (). But then (1 + i2 )u = 0 is just the Cauchy-Riemann
equations and so we conclude that u is analytic in . (Note that 1 u and 2 u
are continuous because u C .) In other words, an analytic distribution
is an analytic function.
Definition 8.12. For s R, the Sobolev space Hs (Rd ) is defined to be the
set of temperedR distributions T S (Rd ) such that Tb is a function with the
property that |Tb ()|2 (1 + ||2 )s d < . Evidently, Hs Ht if s t.
Hs is a Hilbert space with respect to the inner product
Z
c2 () (1 + ||2 )s d .
c1 () T
(T1 , T2 ) = T

Definition 8.13. Let Rd be an open set. The local Sobolev space Hs (),
for s R, is the set of distributions u D () such that u Hs (Rd ) for
all C0 ().
The Sobolev spaces are used in the proof of the elliptic regularity theorem.
In fact, one can prove the following stronger version.
Theorem 8.14 (Elliptic Regularity). Let Rd be open and let P (D) be an
elliptic operator of order N . Suppose that P (D)u = v where v Hs () for
some s R. Then u Hs+N ().
The theorem says that u is better behaved than v (by order N ).
If v C (), then v C0 () for all C0 (). Hence v S (Rd )
and so (v)b S (Rd ) and therefore v Hs () for all s. Hence u Hs ()
for all s. One then shows that this implies that u C () (Sobolevs
Lemma).

ifwilde

Notes