Fourier Transform and Plancherel Theorem

Oliver Daz
University of Texas at Austin
Email: odiaz@math.utexas.edu
September 27, 2000
Here we present the L
1
(R
n
) and L
2
(R
n
) theory of the Fourier transform. We make
a brief treatment of tempered distributions and use it to classify the operators in
L
1
(R
n
) and in L
2
(R
n
) that commute with translations.
1 L
1
(R
n
) theory of the Fourier transform
We begin by dening the fourier transform in L
1
(R
n
).
Denition 1 If f L
1
(R
n
) , the Fourier transform of f is the function

f
dened by letting

f(t) =
_
n
f(x)e
2ixt
dx
The following properties of the Fourier transform are easy to obtain.
Proposition 1 Suppose f L
1
(R
n
) and h R
n
, R. Then,
(a) If g(x) = f(x)e
2ixh
then g(t) =

f(t h).
(b) If g(x) = f(x h) then, g(t) =

f(t)e
2iht
.
(c) If g L
1
(R
n
) and = f g then (t) =

f(t) g(t).
(d) If g(x) = f(x), then g(t) =

f(t).
(e) If g(x) = f(x/) and > 0, then g(t) =
n

f(t).
1
Proof: We only prove (c). The rest are straight forward.
(t) =
_
n
e
2ixt
_
_
_
_
n
f(x y)g(y) dy
_
_
_
dx
=
_
n
e
2iyt
g(y)
_
_
_
_
n
f(x y)e
2i(xy)t
dx
_
_
_
dy
=
_
n
e
2iyt
g(y)
_
_
_
_
n
f(x)e
2ixt
dx
_
_
_
dy
=

f(t) g(t)
For any positive number a and any vector h we dene the dilation by a,
a
,
and the translation by h,
h
, as the operators mapping any function g(x)
into g(ax) and g(x h) respectively. Proposition 1 formulated in terms of
these operators says
(a) (e
2ixh
f(x))

(t) = (
h

f)(t).
(b) (
h
g)

(t) = e
2ith

f(t).
(e) (
a
f)

(t) = a
n

f(a
1
t).
The following lemma will be needed:
Lemma 1 If 1 p < , then the mapping from R to L
p
(R
n
)given by
h
h
f is uniformly continuous.
Proof: We rst prove tihs lemma for continuous functions of compact sup-
port. Suppose that g is such a function and that supp(g) B(0, a) then,
g is uniformly continuous. Given > 0, by uniform continuity of there is a
0 < < a such that [s t[ < implies
[g(t) g(t)[ < (m(B(0, 3a)))
1/p

Thus _
n
[g(x t) g(x s)[
p
dx <
p
2
so that |
t
g
s
g|
p
= |
ts
g g[
p
< . The conclusion follows from the
density of (
c
(R
n
) in L
p
(R
n
).
Another useful result of this nature, which we will use until the next section,
has to do with convolution of functions.
Theorem 1 If 1/p + 1/q = 1 and f L
1
(R
n
) and g L
q
(R
n
) then f g is
uniformily continuous. If 1 < p < then we also have lim
|x|
f g(x) = 0.
Proof: Without lost of generality, we might assume that 1 p < , then
by H olders inequality and translation invariance of Lebesgue measure we
have
[(f g)(x +h) (f g)(x)[ |
h
(f g) f g|
p
|g|
q
and lemma 1 implies uniform continuity. To prove that f g vanishes at
innity, we use sequences f
k
and g
k
of compact supported functions aprox-
imating f and g. If suppf
k
suppg
k
B
a
k
(0), then f
k
g
k
is continuous
and of compact support, supp (f
k
g
k
) B
2a
k
(0). Using H olders inequality
we get that
|f g f
k
g
k
|

|f f
k
|
p
|g|
q
+|f
k
|
p
|g g
k
|
q
This clearly nishes the proof.
It is easy now to establish and prove the following result:
Theorem 2 (Riemann -Lebesgue) (a) The mapping f

f is a bounded
linear transformation from L
1
(R
n
) to L

(R
n
). In fact |

f|

|f|
1
. (b)
If f L
1
(R
n
) then

f is uniformly continuous and

f(t) 0 as [t[ .
Proof: We only prove (b). Note that
[

f(t)

f(s)[

_
n
f(x)e
2ixs
_
e
2ix(ts)
1
_
dx

Dominated convergence does the rest. To prove that

f vanishes at innity,
note that since e
i
= 1 then

f(t) =
_
f(x)e
2i x+
t
2|t|
2
t
dx =
_
f
_
x
t
2|t|
2
_
e
2ixt
dx
3
Hence
2

f(t) =
_
_
f(x) f
_
x
t
2|t|
2
__
e
2ixt
dx
so that
2

f(t) |f
h
f|
1
with h =
t
2|t|
2
. Now, lemma 1 implies that

f(t) 0 as [t[ .
Dierentiation and fourier transformation are related in the following way:
Theorem 3 Suppose f L
1
(R
n
) and x
k
f(x) L
1
(R
n
) , where x
k
is the
r-th coordinate function. Then

f is dierentiable with respect to t
k
and

f
t
k
(t) = (2ix
k
f(x))

(t).
Proof: Letting h = (0, . . . , h
k
, . . . , 0) be a nonzero vector along the k-th
axis. Since [e
2it
k
h
k
1[ t
k
h
k
, we have that

f(t +h)

f(t)
h
k
=
_
e
2ith
1
h
k
f(x)
_

(t) (2it
k
f(x))

(t)
as h
k
0 by dominated convergence.
It is also true that we can take fourier transforms of partial derivatives of
funtions. To make precise this statement, we introduce the following
Denition 2 We say that f L
p
(R
n
) is dierentiable in the L
p
(R
n
) norm
with respect to x
k
, if there exists g L
p
(R
n
) such that
_
_
_
_
f(x +h) f(x)
h
k
g(x)
_
_
_
_
p
0
as h
k
0
Applying proposition 1 and part (a) of theorem 2 leads to

e
2ith
1
h
k

f(t) g(t)

_
_
_
_
f(x +h) f(x)
h
k
g(x)
_
_
_
_
1
and taking h
k
0 proves the following
4
Theorem 4 If f L
1
(R
n
) and g is the partial derivative of f with respect
to x
k
in the L
1
(R
n
) norm then
g(t) = 2it
k

f(t)
everywhere.
For n = 1 we can use integration by parts to get a simpler result.
Proposition 2 If f and f

belong to L
1
(R) and f is the indenite integral
of f

, then (f

(t) = 2it

f(t)
Proof: Since f(x) =
x
_

(t) dt, it follows that lim

x
f(x) = 0 = lim
x
f(x).
Integration by parts does the rest.
The last two theorems can be extended to higher derivatives. Without going
into the details, we note the following formulas:
(i) P(D)

f(t) = (P(2ix)f(x))

(t)
(ii) (P(D)f)

(t) = P(2it)

f(t)
(1)
where, for a n n-tuple = (
1
, ,
n
) of nonnegative integers we let
x

= x

1
1
x

2
2
x
n
n
, D

1
+
2
+...+n
/x

1
1
x

2
2
x
n
n
, P is a polyno-
mial in the n variables x
1
, x
2
, . . . , x
n
and P(D) is the associated dierential
operator.
The main problem in the L
1
(R
n
) thoery of Fourier transform is to obtain
a function f back from its Fourier transform

f. Our next task is to answer
this question. Here we introduce certain summability methods for integrals.
Denition 3 For each > 0 and f a locally integrable function, we dende
the Abel mean A

f to be the integral
A

f =
_
n
f(x)e
|x|
dx (2)
In the same way, we dene the Gauss-Weierstrass mean W

f as
W

f =
_
n
f(x)e
|x|
2
dx. (3)
Whenever the limit lim
0
A

f or lim
0
W

f exists, we say that

_
n
f(x) dx is
Abel or Weierstrass summable.
5
Clearly, if f is integrable, then lim
0
A

f =
_
n
f(x) dx.
It is noteworthy to observe that a function f might be Abel or Weierstrass
summable even if f is not integrable itself, as the next example shows:
Example 1 Let f be the function
f(x) =
sinx
x
1
(0,)
(x)
We claim that Af = lim
0
A

f = /2 For any > 0 there exist P > 0 such

that

p
_
0
sin x
x
dx

2

<

3
and

_
p
e
x
dx <

3
Since f is integrable on the interval [0, P] then there exists > 0 such that
whenever 0 < < then

p
_
0
(e
x
1)
sin x
x
dx

<

3
Combinig these three inequalies we prove our statement.
Equations (2) and (3) can be put in a more general framework
M
,
f = M

f =
_
n
(x)f(x) dx
where (
0
and (0) = 1. Then
_
n
f(x) is summable to l if lim
0
M

f = l.
We shall call M
,
f the means of this integral.
The following computational results will be needed in our pursue of invert-
ing the Fourier transform.
Lemma 2 If (x) = e
x
2
, then (t) = (t)
Proof: We give a simple ODE proof of this fact. First note that
(0) =
_
e
x
2
dx = 1
6
It is easy to see that satises the following initial value problem
/(x) + 2x(x) = 0 (0) = 1
If we apply theorem 3 and proposition 2, then by taking Fourier transform
on the last equation we get that satises same initial value problem as .
Therefore (t) = e
t
2
.
Next, we will need to establish the following identity:
e

=
1

_
0
e
u

u
e

2
/4u
du (4)
To do so, we will use the following identities
e

=
2

_
0
cos x
1 +x
2
dx, with > 0
and
1
1 +x
2
=

_
0
e
(1+x
2
)u
du
The second of this equalities is obvious, while the rst is an easy application
of the theory of residues to the function e
iz
/(1 +z
2
). Then
e

=
2

_
0
cos x
1 +x
2
dx =
2

_
0
cos x
_
_
_

_
0
e
u
e
ux
2
du
_
_
_
dx
=
2

_
0
e
u
_
_
_

_
0
e
ux
2
cos xdx
_
_
_
du
=
1

_
0
e
u
_
_
_

e
ux
2
e
ix
dx
_
_
_
du
=
1

_
0
e
u

u
e

2
/4u
du
Now we are in conditions to state and prove the following
7
Theorem 5 For all > 0 we have that
_
n
e
2iyt
e
|y|
2
dy =
n/2
e
|t|
2
/
(5)
and
_
n
e
2iyt
e
2|y|
dy = c
n

(
2
+[t[
2
)
(n+1)/2
(6)
where c
n
= [(n + 1)/2]/(
(n+1)/2
).
Proof: By a change of variables, it suces to consider the case = 1.
Equality (5) follows easily from lemma 2. Since
_
n
e
2iyt
e
|y|
2
dy =
n

j=1

e
2iy
j
t
j
e
y
2
j
dy
j
=
n

j=1
e
t
2
k
= e
|t|
2
Equation (6) is harder to obtain. Using equation (4) we get
_
n
e
2iyt
e
2|y|
dy =
_
n
_
_
_
1

_
0
e
u

u
e

2
|y|
2
/u
du
_
_
_
e
2iyt
dy
=
1

_
0
e
u

u
_
_
_
_
n
e

2
|y|
2
/u
e
2iyt
dy
_
_
_
du
=
1

_
0
e
u

u
_
__
u

_
n/2
e
u|t|
2
_
du
=
1

(n+1)/2

_
0
e
u
u
(n1)/2
e
u|t|
2
du
=
1

(n+1)/2
1
(1 +[t[
2
)
(n+1)/2

_
0
e
s
s
(n1)/2
ds
=
[(n + 1)/2]

(n+1)/2
1
(1 +[t[
2
)
(n+1)/2
8
For > 0, let us denote the Fourier transform of the functions

(x) =
e
4|x|
2
and

(x) = e
2|x|
by W(t, ) and P(t, ) respectively. That is,
W(t, ) =
1
(4)
n/2
e
|t|
2
/4
P(t, ) = c
n

(
2
+[t[
2
)
(n+1)/2
W(t, ) is called the GaussWeierstrass kernel. P(t, ) is called the Poisson
kernel.
We will show among other things that the Abel and Gaussian means of the
integral
_
n

f(t)e
2itx
dt (7)
converges almost everywhere to f(x). The idea will be to expressed those
means in tearms of convolutions with the Poisson and Gauss kernels and then
use the theory of aproximations to the identity. for the rst step towards
this, we will use the following
Theorem 6 If f anf g belong to L
1
(R
n
) then
_
n

f(t)g(t) dt =
_
n
f(t) g(t) dt
Proof: Applying Fubinis theorem, we get
_
n

f(t)g(t) dt =
_
n
_
_
_
_
n
f(x)e
2ixt
dx
_
_
_
g(t) dt
=
_
n
_
_
_
_
n
g(t)e
2ixt
dt
_
_
_
f(x) dx
=
_
n
f(x) g(x) dx
For any given function we let

(x) =
n
(
1
x). In this notation,
proposition 1 says that (

(t) =

(t).
9
Example 2 Consider the function
(x) = e
4
2
|x|
2
its Fourier transform is given by
(t) = (4)
n/2
e
|t|
2
/4
Thus
(

)(x))

(t) = (4
2
)
n/2
e
|t|
2
/4
2
= W(t,
2
)
Example 3 Consider the function
(x) = e
2|x|
its Fourier transsform is given by
(t) = c
n
1
(1 +[t[
2
)
(n+1)/2
Thus
(

)(x))

(t) = c
n

(
2
+[t[
2
)
(n+1)/2
= P(t, )
These examples together with theorem 6 lead to the following
Theorem 7 If f and belong to L
1
(R
n
) then
_
n

f(x)e
2itx
(x) dx =
_
n
f(x)

(x t) dx
for all > 0. In particular,
_
n

f(x)e
2itx
e
2|x|
dx =
_
n
f(x)P(x t, ) dx
and
_
n

f(x)e
2itx
e
4
2
|x|
2
dx =
_
n
f(x)W(x t, ) dx
10
We have almost everything set up to use the theory of aproximations to
the identity to show that integral (7) is sumable to f for a large class of
methods that includes both the Able and Guass summability. The task will
be to prove that the means
_
n

f(t)e
2itx
(t) dt converge to f in L
1
(R
n
)
provided that both and are integrable and
_
n
(t) dt = 1. The following
lemma shows that the Gauss and Poisson kernels satisfy that condition.
Lemma 3 For all > 0
_
n
W(x, ) dx = 1
and _
n
P(x, ) dx = 1
Proof: For > 0, a simple change of variable gives that
_
n
W(x, ) dx =
_
n
W(x, 1/4) dx =
_
n
e
|x|
2
dx = 1
Similarly, we have that
_
n
P(x, ) dx =
_
n
P(x, 1) dx
. The following geometrical facts will be at handy in our calculations:
(i) The surface element d
n1
of the sphere S
n1
on the space R
n
is given
by:
d
n1
= sin
n2

1
d
1
d
n2
= sin
n2

1
sin
n3

2
sin
n2
d
1
d
2
d
n1
where 0
k
< for 1 k n 2 and 0
n1
< 2
(ii) The volume
n
of the unit ball in R
n
and the surphace
n1
of its
boundary are:

n
=
_
|x|1
dx =

n1
n

n1
=
_
S
n1
d
n1
(u) =
2
n/2
[n/2]
11
With this in mind, if we use polar coordinates and then the change of variable
r = tan and we will have
_
n
1
(1 +[x[
2
)
(n+1)/2
dx =

_
0
_
S
n1
r
n1
(1 +r
2
)
(n+1)/2
d
n1
(u) dr
=
n1

_
0
r
n1
(1 +r
2
)
(n+1)/2
dr
=
n1
/2
_
0
sin
n1
d
=
1
2

n
=

(n+1)/2
[(n + 1)/2]
This nishes the proof.
Now we state and proof a theorem concerning approximations to the identity
Theorem 8 (Aproximation to the Identity) Suppose L
1
(R
n
) with
a =
_
n
(x) dx, and for > 0 let

(x) =
n
(
1
x). If f L
p
(R
n
) or
f (
o
(R
n
) L

(R
n
), then |f

a f|
p
0 as 0. In particular,
u(x, ) =
_
n
f(t)P(x t, ) dt and s(x, ) =
_
n
f(t)W(x t, ) dt converge
to f on L
p
(R
n
) norm as 0.
Proof: By a simple change of variables we have
_
n

(s) dx =
_
n

n
(
1
x) dx =
_
n
(x) dx = a
Hence,
(f g)(x) af(x) =
_
n
[f(x t) f(x)]

(t) dt
=
_
n
[f(x t) f(x)](t) dt
For 1 p < let S = h L
q
(R
n
):|h|
q
= 1, where 1/p +1/q = 1. Using
Minkowskis duality equation, Fubinis theorem and H older inequality we
12
get
|f g af|
p
= sup
S
_
_
_

_
n
h(x)
_
_
_
_
n
[f(x t) f(x)](t)dt
_
_
_
dx

_
_
_
sup
S
_
_
_
_
n
[h(x)[
_
_
_
_
n
[f(x t) f(x)[[(t)[dt
_
_
_
dx
_
_
_
sup
S
_
_
_
_
n
[(x)[
_
_
_
_
n
[f(x t) f(x)[[h(t)[dx
_
_
_
dt
_
_
_

_
n
[(t)[|h|
q
|
t
f f|
p
dt
=
_
n
[(t)[|
t
f f|
p
dt
For f (
0
(R
n
), the same relationship takes place with p = . We in-
troduce the function modulus of continuity of f,
f,p
, dened as
f,p
(h) =
(h) = |
h
f f|
p
. It is clear that [(h)[ 2|f|
p
for all h. Continuity of
translations, lemma 1, implies that (h) 0 as h 0. Therefore f

converges to f in L
p
(R
n
) norm as 0.
The same argument applies to give a more general result:
Corollary 1 Suppose A() is a familly of invertible matrices such that
lim
0
A() = 0. Let us dene
A()
(x) = (det A())
n
(A
1
()x). Then
|f
A()
a f|
p
0.
The strength of theorems 7 and 8 is reected in the solution to the Fourier
inverse problem
Theorem 9 If and its Fourier transform are integrable and
_
n
(x) dx =
1, then the means of the integral
_
n

f(t)e
2ixt
dt converges to f(x) in
L
1
(R
n
) norm. In particular, the Abel and Gauss means of this integral
converges to f in L
1
(R
n
).
Since s(x, ) =
_
n

f(t)e
2ixt
e
4
2
|t|
2
dt converges to f(x) in L
1
(R
n
) as
0 there is a sequance
k
0 such that s(x,
k
) f(x) for a.e. x. If
13

f happens to be integrable as well, then by dominated convergence we get

the following result:
Corollary 2 If both f ad

f are integrable then
f(x) =
_
n

f(t)e
2itx
dt
for almost every x.
Remember that

f is continuous, and if

f is integrable then the integral
_
n

f(t)e
2itx
dt will be continuous as well (in fact, it is (

f)

(x)). Hence,
by changing f on a set of measure zero we can obtain equality in corolary 2.
In other words, f can be make into a continuous function by changing its
values on a set of measure zero.
If

f(t) = 0 then f(x) = 0 a.e. Applying this to f
1
f
2
we obtain the
following result on uniqueness of the Fourier transform.
Corollary 3 If f
1
and f
2
belong to L
1
(R
n
) and

f
1
(t) =

f
2
(t) for all t R
n
then f
1
(x) = f
2
(x) a.e.
The Fourier inversion problem has a pointwise solution as well. The next
result is a pointwise version of the theorem of aproximations to the identity.
for any given measurable function , we dened the least decreasing radial
majorant of as the function
(x) = ess. sup
|y||x|
[(y)[
Theorem 10 Suppose L
1
(R
n
) with a =
_
n
(x) dx. If L
1
(R
n
)
and f L
p
(R
n
) 1 p < , then
(i) lim
0
(f

)(x) = af(x) whenever x is a Lebesgue point of f.

(ii) sup
>0
[(f

)(x)[ ||
1
(M(f))(x), where M(f) is the Hardys maximal
function of f
In particular
14
(a) The Poisson integral of f
u(t, ) =
_
n
f(x)P(t x, ) dx
and the GaussWeierstrass integral of f
s(t, ) =
_
n
f(x)W(t x, ) dx
converge to f(x) as 0 for almost every x R
n
.
(b)
sup
>0
[u(t, )[ (M f)(t)
sup
>0
[s(t, )[ (M f)(t)
Proof: Let x be a Lebesgue point of f. This means that for any > 0, one
con nd > 0 such that r < implies that
1
m(B
r
(x))
_
Br(x)
[f(y) f(x)[ dy <
Changing to polar coordinates we can reinterpret the former condition as:
if r < then
G(r) =
r
_
0
s
n1
g(s) ds <
n
r
n
(8)
where
g(s) =
_
S
n1
[f(x s u) f(x)[ d
n1
(u)
15
On the other hand, for all > 0 we have
[f

(x) a f(x)[ =

_
n
[f(x t) f(x)]

(t) dt

_
|t|<
[f(x t) f(x)]

(t) dt

_
|t|
[f(x t) f(x)]

(t) dt

= I
1
+I
2
To do estimates on I
1
we make the following observations. First, note that
is a radial function, i.e. (x
1
) = (x
2
) if [x
1
[ = [x
2
[. Let us dene

0
(r) = (x) where r = [x[. Clearly
0
is a non deacrising function. Using
polar coordinates we get that
_
r/2|x|r
(x) dx =
r
_
r/2
_
S
n1

0
(s)s
n1
d
n1
(u) ds
n
2
n
1
2
n
r
n

0
(r)
thus, r
n

0
(r) 0 as r , and as r 0. Also, for some constant B
n
depending on dimension n, r
n

0
(r) ||
1
B
n
= A for all r. Using these
16
observations on equation 8 we get
I
1

_
|y|<
[f(x y) f(x)[[

(t)[ dt

_
0
s
n1
g(s)
n

0
(
1
s) ds
= G(s)
n

0
(
1
s)

_
0
G(s)
n
d
0
(
1
s)

n
s
n

0
(
1
s)

0

/
_
0
G(s)
n
d
0
(s)

n
A
n

/
_
0
s
n
d
0
(s)
n

_
_
A

_
0
s
n
d
0
(s)
_
_
=
n

_
_
A+n

_
0
s
n1

0
(s) ds
_
_
= (
n
A+||
1
)
To estimate I
2
, note that

L
q
(R
n
) for any 1 q . Therefore, if
1/p + 1/p

= 1 then by H olders inequality

I
2
|f|
p
|

|
p
+[f(x)[ |[

|
1
Since
|

|
1
=
_
|x|

(x) dx =
_
|x|/
(x) dx
we see that the second summand tends to 0 as 0. For the rst summand,
since p

= 1 +p

/p we have that
|

|
p
=
_
_
_
_
|x|
[

(x)]
p

dx
_
_
_
1/p

=
_
_
_
_
|x|

(x)[

(x)]
p

/p
dx
_
_
_
1/p

_
_
_|

|
p

/p

_
|x|

(x) dx
_
_
_
1/p

= |

|
1/p

|
1/p

1
17
But |

=
n
(
1
) =
n
(/)
n

0
(/) 0 as 0 as pointed
out before, and (i) follows.
As for part (ii), note that [(f

)(x)[ ([f[ [

[)(x) ([f[

)(x). Thus
we only need to consider f 0 and show that
sup
>0
[(f

)(x)[ ||
1
(M(f))(x)
Also, since
((
x
f)

)(0) = (f

)(x)
M(
x
f)(0) = (M(f))(x)
(f

)(0) = ((

f) )(0)
it suces to prove that
(f )(0) ||
1
(M(f))(0) for all f L
p
(R
n
)
Let
(r) =
_
S
n1
f(ru) d
n1
(u)
(r) =
r
_
o
(s)s
n1
ds =
_
|x|r
f(x) dx
(f )(0) =
_
f(y)(y) dy =
_
f(y)(y) dy =
_
f(y)(y) dy. Thus,
using polar coordinates, the radial property of and integration by parts
we obtain
(f )(0) = lim
0
N
_
_

0
(r)(r)]
N

+
N
_

(r) d(
0
(r))
_
_
For the rst term converges to 0 as the following estimate shows it:
[
0
(N)(N)
0
()()[
n
(M f)(0)(
0
(N)N
n
+
0
()
n
)
For the second term we have the following estimate
N
_

(r) d(
0
(r))
n
(M f)(0)

_
0
r
n
d(
0
(r))
= ||
1
(M f)(0)
18
This concludes our proof.
Any point of continuity of f is of course a Lebesgue point of f. Thus if f is
continuous at 0, then by theorems (7) and (10) we have that
lim
0
_
n

f(x)e
2|x|
dx = lim
0
_
n
f(x)P(x, ) dx = f(0)
If we further assume that

f 0 it follows from monotone convergence that

f L
1
(R
n
). In this way we obtain the following simple and neat result:
Corollary 4 Suppose f L
1
(R
n
) and

f 0. If f is continuous at 0 then

f L
1
(R
n
) and
f(x) =
_
n

f(t)e
2itx
dt
almost everywhere (i.e. at every Lebesgue point of f). In particular,
f(0) =
_
n

f(t) dt
Applying this result ot the Poisson ans GaussWeierstrass kernels we get:
Corollary 5
_
n
W(x, )e
2itx
dx = e
4
2
|t|
2
_
n
P(x, )e
2itx
dx = e
2|t|
for all > 0.
We inmidiately obtain the semigroup properties of the Poisson and Gauss
Weierstrass kernels:
Corollary 6 If
1
and
2
are positive real numbers then
(a) W(x,
1
+
2
) =
_
n
W(x t,
1
)W(t,
2
) dt
(b) P(x,
1
+
2
) =
_
n
P(x t,
1
)P(t,
2
) dt
19
2 L
2
(R
n
) Theory and the Plancherel Theorem
We know that the space L
1
(R
n
)L
2
(R
n
) is dense in L
2
(R
n
). Here we will
extend the fourier transform from the former space to the latter. One of the
nice properties of this extension is that it turns out to be a unitary linear
tranformation.
Theorem 11 If f L
1
(R
n
)L
2
(R
n
) then |

f|
2
= |f|
2
Proof: Let g(x) = f(x). Then by theorem 1 h = f g is conitnuous.
Being the convolution of functions in L
1
(R
n
) h is also integrable. Since
g =

f then

h =

f g = [

f[
2
. By corollary 4 it follows that

h is integrable and
h(0) =
_
n

h(t) dt. Thus

_
n
[

f(t)[
2
dt =
_
n

h(t) dt = h(0) =
_
n
f(y)g(0 y) dy
=
_
n
f(y)f(y) dy =
_
n
[f(y)[
2
dy
This theorem asserts that the Fourier transform is a unitary operator dened
on the dense subspace L
1
L
2
of L
2
(R
n
) into L
2
(R
n
). Thus by Carathe odory
extension, there exists a unique bounded extension, T on the whole space
L
2
(R
n
). We will also use the notation

f = Tf. Clearly, for any function
f L
2
(R
n
)

f(t) = Tf(t) = lim

k
_
|x|k
f(x)e
2ixt
dx
It is also clear that T is a unitary operator from L
2
(R
n
) into L
2
(R
n
). In
fact this operator is onto:
Theorem 12 (Plancherel)
(i) The Fourier transform is a unitary operator on L
2
(R
n
).
(ii) The inverse Fourier transform, T
1
, can be obtained by letting (T
1
g)(x) =
(Tg)(x) for all g L
2
(R
n
).
20
Proof: Since T is a unitary operator, T(L
2
(R
n
)) is a closed subspace of
L
2
(R
n
). Let g
_
T(L
2
(R
n
))
_

. A simple density argument extends theo-

rem 6 to functions in L
2
(R
n
). Thus
_
n

f(t)g(x) dx =
_
n
f(x) g(x) dx = 0
for all f L
2
(R
n
). This implies that |g|
2
= | g|
2
= 0.
For the second part of the proof, note that since T is a unitary operator,
then it preserves the inner product: (u[v) =
_
n
uv dx. Let g be any function
in L
1
L
2
and f L
2
(R
n
). Let f
k
be the sequence given by
f
k
(x) =
_
|t|k

f(t)e
2ixt
dt
Clearly each f
k
L
1
L
2
, and f
k
(x) converges in L
2
to the function

f given
by

f(x) = T(

f)(x).Then
(g[

f) = lim
k
(g[f
k
) = lim
k
_
n
g(x)
_

_
_
|t|k

f(t)e
2itx
dt
_

_
dx
= lim
k
_
|t| k

f(t)
_
_
_
_
n
g(x)e
2itx
dx
_
_
_
dt
= lim
k
_
|t|k
g(t)

f(t) dt = ( g[

f) = (g[f)
This clearly implies that f(x) = (T
1

f)(x) =

f(x) = (T

f)(x) for all
f L
2
(R
n
).
3 Tempered Distributions and a Characterization
of Operators that Commute with Translations
The basic idea in the theory of distributions is to consider them as linear
functionals on a space of regular functionsoften called testing functions.
This space is assumed to be well behave with respect to operations we
have studied so far, namely: dierentiation, Fourier transform, convolution,
translation, dilation, etc. This will be reected, in return, on the properties
of distributions. We are after a space of testing functions o on whichthese
21
operations are dened and preserve the space. Motivated by theorems 3, 4
and formulas 1, the space o must consist of innitely dierentiable functions
such that when they or their derivatives are multiplied by polinomials, they
must still be integrable.
Denition 4 The space o of testing functions is dened to be the class of
all functions (

on R
n
such that
sup
x
n
[x

(D

)(x)[ <
for all ntuples = (
1
, . . . ,
n
) and = (
1
, . . . ,
n
) of nonnegative inte-
gers.
Clearly that space is not void: the family of functions

(x) = e
|x|
2
belongs to o. o also contains the space T of all (

functions with compact

support. The later space is not void. To see this, we rst consider the case
n = 1. Let f(t) = e
1/t
if t > 0 and f(t) = 0 when t 0. It is an easy
exercise to see that f (

and that f and all its derivatives are bounded.

Let (t) = f(1 + t)f(1 t); then (t) = e
2/(1|t|
2
)
if [t[ < 1 and zero
otherwise. Thus T(R). We can easily obtain ndimensional variants
from :
(a) For x R
n
dene (x) = (x
1
) (x
n
); then T(R
n
)
(b) For x R
n
dene (x) = e
2/(1|x|
2
)
, if [x[ < 1 and zero otherwise.
Then T(R
n
)
(c) If (

and is the function in (b) then (x)(x) denes a function

in T(R
n
).
The spaces (
o
(continuous functions that vanish at innity with the sup
norm) and L
p
(R
n
), with 1 p < contain o. Using the desity of (
c
on
both spaces and the StoneWeierstrass theorem we can show that T is in
fact dense in both spaces and therefore so would o. Another interesting
observation is given by the following
Lemma 4 The L
p
norm of a function o is bounded by a linear combi-
nation of L

norms of the form x

(x), The coecients and exponents of

these terms depend only on p and the dimension n.
22
Proof: Let A = ||

and B = sup
x
n
[x[
2n
[(x)[ then
_
_
_
n
[(x)[
p
dx
_
_
1/p

_
_
_
_
|x|1
[(x)[
p
dx
_
_
_
1/p
+
_
_
_
_
|x|>1
[(x)[
p
dx
_
_
_
1/p
A
_

n1
n
_
1/p
+B
_
_
_
_
|x|>1
[x[
2np
dx
_
_
_
1/p
= A
_

n1
n
_
1/p
+B
_

n1
(2p 1)n
_
1/p
Theorem 3 implies that (

whenever o. Combinig this with

theorem 4 gives that o. In other words,
Theorem 13 If o then o.
Remember that ()

. Since o is closed under multiplication, then

the inverse Fourier transform theorem shows that
Theorem 14 If and are in o then so is .
We introduce now a metric on o that makes this sapce a topological vector
space. There is a natural countable family of norms

in o, indexed by
ordered pairs (, ) of n-tuples of nonnegative integers, given by:

() = sup
x
n
[x

(x)[
These norms induce the distances d

(, ) =

(). Let d
1
, . . . , d
n
, . . .
be an ordering of these metrics. Thus, we dene a metric d in o as
d(, ) =

k=1
2
k
d
n
(, )
1 +d
n
(, )
It is clear that
m
with respect to d, if and only if
m
with respect
to each d
k
(as m ). From this observation, it follows that the space
operations (, ) + and (a, ) a are continuous. Thus o with the
topology induced by d becomes a topological vector scape. The following
properties of o are not dicult to be checked:
23
Theorem 15 (o, d) has the following properties:
(1) The mapping (x) x

(x) is continuous
(2) If o then lim
h0

h
=
(3) Suppose o and h = (0, . . . , h
i
, . . . , 0) lies on the i-th coordinate axis
of R
n
, then the dierence quotient [
h
]/h
i
tend to /partialx
i
as h 0
(4) o is a complete metric space
(5) The Fourier transform is a homeomorphism of o onto itself.
(6) T is dense subset of o
(7) o is separable.
The collection o

of all continuous linear functionals on o is called the space

of tempered distributions. Thus to show that a linear functional L on o is
a tempered distributionm it suces to show continuity at the origin, i.e,
lim
k
L(
k
) = 0 whenever
k
0 in o. Here we present some examples:
Example 4 Let f L
p
(R
n
), 1 p , and let L = L
f
dened as
L() = L
f
() =
_
n
f(x)(x) dx
for o. Lemma 4 implies that |
k
|
q
0 if
k
0 in o. By H olders
inequality
[L()[ |f|
p
||
q
Thus, L is a tempered distribution
Example 5 If is a nite Borel measure, then the linear functional L = L

dened by
L() = L

() =
_
n
(x) d(x)
is a tempered distribution.
24
Example 6 Let f be a measurable function such that f(x)(1 + [x[
2
)
k
belongs to L
p
(R
n
), 1 p for some integer k. Dene L = L
f
just as in
example 4: L() =
_
n
f(x)(x) dx. Since
L() =
_
n
_
(1 +[x[
2
)
k
(x)
_ _
f(x)/(1 +[x[
2
)
k
_
dx
it follows that L is a tempered distribution. f is called a tempered L
p
func-
tion. When p = , f is said to be slowly increasing.
Example 7 A Borel measure is called a tempered measure if
_
n
(1 +[x[
2
)
k
d[[(x) <
for some integer k. If we dene L() = L

() =
_
n
(x) d(x), then we
get a tempered distribution.
Example 8 For a xed point x
0
R
n
, and an n-tuple of nonnegative
integers, we dene the linear functional L() = D

(x
0
). The continuity of
the metric
0
in o implies that L is a tempered distribution. In particular,
for x
0
= 0 and = (0, . . . , 0) we get the Diracs delta distribution: L() =
(0). The latter can be obtained also, by considering the Borel measure of
mass 1 concentrated at the origin.
There is a simple characterization of tempered distributions.
Theorem 16 A linear functional L on o is a tempered distribution if and
only if there exists a constant C > 0 and integers m and l such that
[L()[ C

||l,||m

()
for all o.
Proof: The existence of C, l, m clearly implies the continuity of L.
For the converse, note that the family of sets of the form
N
,l,m
=
_
_
_
:

||l,||m

() <
_
_
_
25
where > 0, and l and m are nonnegative integers, forms a basis of neigh-
borhoods arround 0 in o. To see that, let k = lm and consider the sets
U
,
= :

() < /k
Clearly, these are open sets in o. Also

||l,||m
U

N
,l,m
.
For the ball B(0, ) in o, pick K > 0 such that

n=K
2
n
< /2. Take l
and m big enough such that the norms in the list
1
, . . . ,
K
are included
among the norms

with [[ l, [[ m. Then
N
/2,l,m

K

n=1
:
n
() < /2 B(0, )
Ith follows that there is a set N
,l,m
such that [L()[ < 1 for all N
,l,m
.
For ,= 0, let
|| =

||l,||m

()
it is easy to see now that C = 2/ will do the job:
[L()[ =
2||

L
_

2||

2||

Many important operations on functions, i.e: diferentiation, translation,

dilation, reections, Fourier transform, convolution, etc, can be extended in
a natural way to distributions. Here we motivate some of these extension.
Derivative:
If f L
p
(R
n
) admits an L
p
(R
n
) derivative g = D

f, then by integration
by parts we have that
L
g
() =
_
n
(D

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

(x) dx = (1)
||
L
f
(D

)
26
for all o. Thus, for any distribution u, we dene D

u as the linear
functionalv on o such that
v() = (1)
||
u(D

)
Clearly D

u is a distribution.
Translation:
If f L
p
(R
n
) then
L

h
f
() =
_
n
f(x h)(x) dx =
_
n
f(x)(x +h) dx = L
f
(
h
)
Thus, for any distribution u, we dene
h
u as the linear functional on o
such that

h
u() = u(
h
)
It is clear that
h
u is a distribution since it is the composition of continuous
functions on o.
Fourier transform:
If f L
1
(R
n
) then from theorem 6 we have
L

f
() =
_
n

f(x)(x) dx =
_
n
f(x) (x) dx = L
f
( )
thus, for any given distribution u we dene u by
u() = u( )
Multiplication by a test function:
If o and u o

then we dene u as the linear functional

(u)() = u()
for all o. Clearly this is a distribution on o.
Convolution:
If f L
p
(R
n
) and and belong to o. Let (x) = (x). A simple
application of Fubinis theorem shows that
L
f
() =
_
n
(f )(x)(x) dx =
_
n
f(x)( )(x) dx = L
f
( )
27
Thus, for any distribution u and test function we dene u as the linear
functional
(u )() = u( )
The right side of the equation above is the composition of continuous func-
tions in o. Therefore u is a distribution.
The following theorem gives a better characterization of the convolution
distribution just described:
Theorem 17 If u o

and o then the convolution u is the function

f(x) = u(
x
). Moreover, f (

and it is slowly increasing.

Proof: First we show that (u )() =
_
n
f(t)(t) dt:
(u )() = u( ) = u
_
_
_
n
(x t)(t) dt
_
_
= u
_
_
_
n
(
t
)(x)(t) dt
_
_
The Reimann sums of the last integral converge in the topology of o, thus
u
_
_
_
n
(
t
)(x)(t) dt
_
_
=
_
n
u(
t
)(t) dt =
_
n
f(t)(t) dt
Continuity follows from theorem 15. Let h = (0, . . . , h
j
, . . . , 0), then

x+h

x

h
j

x
_

x
j
_
in the o topology. Thus, by continuity of u,
f(x +h) f(x)
h
j
u
_

x
_

x
j
__
By an iterative argument we have (D

f)(x) = (1)
||
u(
x
D

). The fact
that f is slowly increasing follows from theorem 16: There exist C > 0 and
nonnegative integers l and m such that
[f(x)[ = u(
x
) C

||l,||m

(
x
)
28
But

(
x
) = sup
w
n
[(w +x)

(D

)(w)[
whis is clearly a polynomial in x.
In what follows, we will apply the theory of distributions to the study of a
class of operators that occur in the theory of singular integrals: those that
commute with translations. Suppose B is an operator mapping a linear
space V of functions in R
n
into another space of the like. We say that B
commutes with translations if
h
B = b
h
for all h R
n
.
Example 9:
Fix f L
p
(R
n
) and let be B the operator on L
1
(R
n
) to L
p
(R
n
) dened
by convolution: g f g. We have that |Bg|
p
|f|
p
|g|
1
. Clearly it
commutes with translations.
Example 10:
Consider the Banach space of all complex measures on R
n
with norm ||.
The Fourier transform of f as a distribution concides with that of a measure:
(t) =
_
n
e
2ixt
d(x)
Let be the map:
: R
n
R
n
R
n
(x, y) x +y
Let be the measure on R
n
induced by , i.e:
( )(E) = ( )
_

1
(E)
_
It follows stright forward that
| | ||||
From the Rieszrepresentation theorem, it follows that is the unique
measure such that
_
n
f(x) d(x) =
_
n

n
f(x +y) d( )(x, y)
29
for every f (
c
(R
n
). Application of Fubunis theorem implies
( )(E) =
_
n
(E t) d(t) =
_
n
(E s) d(s)
and if d = f dx, where f L
1
(R
n
) then d( )(y) = (f )(y) dy:
( )(E) =
_
n

n
1
E
(x +y) d( )(x, y)
=
_
n
_
_
_
_
n
1
E
(x +y)f(y) dy
_
_
_
d(x)
=
_
n
_
_
_
_
n
1
E
(y)f(y x) dy
_
_
_
d(x)
=
_
E
_
_
_
_
n
f(y x) d(x)
_
_
_
dy =
_
E
(f )(y) dy
It is clear that the map on L
1
(R
n
) given by f f is continuous:
|f |
1
= | | |f|
1
|| = ||||. Clearly it also commutes with
translation.
We will show that all bounded operators in L
p
(R
n
)are of the similar convo-
lution type. The following lemma will imply the former statement:
Lemma 5 If f L
p
(R
n
) has derivatives in the L
p
norm of all orders
n + 1, then f equals almost everywhere a continuous function g satisfying
[g(0)[ C
n,p

||n+1
|D

f|
p
where C
n,p
depends only on the dimension n and the exponent p.
Proof: Let x R
n
, then
(1 +[x[)
(n+1)/2

_
_
1 +
n

j=1
[x
j
[
_
_
n+1
C

||n+1
[x

[
30
For p = 1 we will prove that

f is integrable and use corollary 2:
[

f(t)[ C

(1 +[t[
2
)
(n+1)/2

||n+1
[t[

[[

f(t)[
= C

(1 +[t[
2
)
(n+1)/2

||n+1
[((2)
||
D

f)

(t)[
C

(1 +[t[
2
)
(n+1)/2

||n+1
|D

f|
1
Therefore
|

f|
1

C

n
2

||n+1
|D

f|
1
thus f is equal to a continuous function g almost everywhere and
[g(0)[ |g|

= |f|

f|
1

C

n
2

||n+1
|D

f|
1
For p > 1, choose 0 1 in T such that (x) = 1 if [x[ 1 and (x) = 0
if [x[ 2. Then f satises the hypothesis of p = 1. Thus varphif equals
a continuos function h such that
[h(0)[ C

||n+1
|D

f|
1
Since D

(f) =

+=
C

(D

f)(D

) and using H olders inequality we

have
|D

(f)|
1

_
|x|2

+=
C

[D

[ [D

f[ dx
K

+=

0
()
_
|x|2
[D

f[ dx
K
_
2
n

n1
n
_
1/q

+=

0
()|D

f|
p
K
_
2
n

n1
n
_
1/q
max
||||

0
()

||||
|D

f|
p
31
Thus, there is a constant C
n,p
depending only on n and p such that
h(0) C
n,p

||n+1
|D

f|
p
(9)
Since (x) = 1 on the unit ball we see that f is equal to a continuous
function h almost everywhere on the unit ball. Choosing
k
(x) = (x/k)
shows that f is almost everywhere a continuous function h
k
on the ball of
radius k arround 0. Note that h
k
(0) = h(0), hence equation (9) is preserved.
This nishes the proof of the lemma.
The following theorem characterizes all continuous operators on L
p
(R
n
) that
commute with translations:
Theorem 18 Suppose B : L
p
(R
n
) L
q
(R
n
), 1 p, q , is a linear
bounded operator that commutes with translations; then, there exists a unique
tempered distribution u such that B = u for all o.
Proof: Let h = (0, . . . , h
j
, . . . , 0) be a nonzero vector along the jth
axis. Since

h

h
j

x
j
in o, it converges in L
p
as well. Therefore,
B
_

h
j
_
=

h
BB
h
j
converges to B
_

x
j
_
=
B
x
j
in L
q
by continuity
of B. By an iteration argument, it follows that B has derivatives in L
q
of all orders and D

(B) = B(D

). Hence by the previous lemma, B

equals a continuos function g

almost everywhere and

[g

(0)[ C
n,q

||n+1
|D

(B)|
q
= C
n,q

||n+1
|B(D

)|
q
|B|C
n,q

||n+1
|D

|
q
This shows that the linear functional g

(0) is in fact a distribution

u
1
on o. Letting u = u
1
ge obtain the distribution we are looking for:
(u)(x) = u(
x
) = u((
x

)) = u(
x
) = u
1
(
x
) = (B(
x
))(0) =
(
x
B)(0) = (B)(x). Uniqueness follows inmidiately from the construc-
tion.
Let us denote by (L
p
, L
q
) the space of those tempered distributions u for
which there exists A > 0 such that |u |
q
A||
p
for all o. Last
32
theorem shows that there is a onetoone relationship between this space,
and that of continuous linear operartors from L
p
(R
n
) to L
q
(R
n
) that com-
mute with translations. For the case p = 2 = q there is a much better
characterization:
Theorem 19 The distribution u belongs to (L
2
, L
2
) if and only if there
exists b L

(R
n
) such that u = b. In this case, |b|

is the norm of
the operator B : L
2
o L
2
dened by putting B = u ; moreover
(u )

= u .
Proof: By the Fourier inverse theorem, we have that

is the Fourier
transform of . Then, by denition of Fourier transform and convolution
in the distributional sence we have that
(u )

() = (u )(

) = u(

) = u(( )

) = ( u )()
This shows that (u )

= u in general.
Consider the function
0
(x) = e
|x|
2
in o. Since u (L
2
, L
2
), Plancherels
theorem implies that
0
= (u
0
)

= u
0
= u
0
belongs to L
2
(R
n
). Let
b(x) = e
|x|
2

0
(x) =
0
(x)/
0
(x). Let o and T, then
( u )() = u( ) = u(
0
/
0
) = ( u)( /
0
)
=
_
n

0
(x) (x)e
|x|
2
(x) dx
=
_
n
b(x) (x)(x) dx
= (b )()
Taking such that the support of its Fourier transform contains the support
of we get that u() = b() for all T. The density of T in o implies
that u = b.
Since u (L
2
, L
2
) there exists a constant A > 0 such that
|b |
2
= |(u )

|
2
= |u |
2
A||
2
for all o. Since o is dense in L
2
(R
n
), it is now easy to show that
b L

(R
n
) with |b|

A.
The converse is easy: if u = b L

(R
n
) then
|u |
2
= | u |
2
= |b |
2
|b|

| |
2
= |b|

||
2
33
This nishes the proof.
For p = 1 = q there is a nice characterization as well.
Theorem 20 The distribution u belongs to (L
1
, L
1
) if and only if it is a
nite Borel measure. In this case, the total variation of u equals the norm
of the operator B : L
1
o L
1
dened by B = u for o
Proof: Consider the space (
o
(R
n
) of continuous functions that vanish
at innity with the sup norm. Its dual space, according to the Riesz
representation theorem, is the normed space of nite Borel measures M =
M(R
n
) on R
n
: mapping the measure to the linear functional that assings
the value
_
n
(x) d(x). Note that L
1
(R
n
) is naturally embeded into M
by the map f f dx.
On the other hand, we can topologize M with the weak

topology and by
Alaoglus theorem, the unit ball B = : || 1 in M is compact with
respect this topology; moreover, since (
0
is separable, the unit ball is in fact
metrizable (with the weak

topology), thus sequantially compact.

Consider the family of L
1
functions u

= u W(, ), > 0, where W is

the GaussWeierstrass kernel. By hypothesis, there exists A > 0 such that
|u

|
1
A|W(, )|
1
= A. Thus, by the previous observations, there exists
M and a sequence
k
coverging to 0 such that u

k
as k in
the weak

topology. That is
lim
k
_
n
(x) du

k
(x) =
_
n
(x) d(x)
for any o. Let

(x) =
_
n
(x t) W(t, ) dt. By dominated conver-
gence we have that (D

)(x) =
_
n
(D

)(x t)W(t, ) dt = (D

for
any ntuple of nonnegative numbers. By the theorem of aproximations to
the identity, (case p = ) we have that D

(x) converges uniformily in x

to D

(x) as 0. Thus

in o as 0. Hence, by continuity of
u, u(

) u(). Since W(, ) =

W(, ) we have
u(

) = u(W(, ) ) = (u W(, ))() =

_
n
(x)u

(x) dx
Taking =
k
and letting k 0 we obtain
u() =
_
n
(x) d(x)
This shows that u = M(R
n
). The rest of the proof is easy.
34