Sei sulla pagina 1di 9

Bochners Theorem on the Fourier Transform on R

Yitao Lei
October 2013

Introduction

Typically, the Fourier transformation sends suitable functions on R to functions on R.


This can be defined on the space L1 (R) + L2 (R), i.e. functions which can be written
as the sum of a function in L1 (R) and a function in L2 (R). A celebrated result (the
HausdorffYoung inequality) states that the Fourier transform takes functions in Lp (R)
to Lq (R) for 1 p 2, where p1 + 1q = 1.
However, this does not extend to the case when p > 2. In addition, the maps F : Lp (R)
Lq (R) are not surjective when p < 2. Therefore it seems natural to try to extend the
Fourier transform to objects other than functions. The most complete method of doing
this is extending the Fourier transform to the space of tempered distributions, i.e. the
space of linear functionals on the Schwartz functions.
Instead, we will study the FourierStieltjes transform, a slight generalisation of the
Fourier transform. We now transform complex finite Borel measures rather than functions, and output a function. Bochners Theorem answers the question of which functions
are the FourierStieltjes transform of some positive Borel measure. It states that the
function is continuous and positivedefinite is a necessary and sufficient condition for it
to be a FourierStieltjes transform.
We shall first explore the analogous situation on the torus (or the circle here, when the
dimension is one). FourierStieltjes coefficients will be examined, and are related to
Fourier coefficients. There is much similarity between FourierStieltjes coefficients and
the FourierStieltjes transform. However, the theory building up to a Bochnertype
result on the torus is clearer and simpler than going directly to Bochners theorem on R.

Preliminaries

Let T be the torus [0, 2] with the points 0 and 2 identified, i.e. the same point. In this
paper, we shall be working with two spaces of continuous functions, both equipped with
the supremum norm: C(T), continuous functions on T, and C0 (R), continuous functions
that vanish at infinity.
We will also consider two spaces of finite complex Borel measures on T and R, namely
M (T) and M (R). For clarity, will denote a measure in M (T), while will denote
1

a measure in M (R). The norm on both spaces are given by kkM (T) = ||(T) and
kkM (R) = ||(R).
Recall that if V is a Banach space, then the dual space V is the set of linear functionals
: V C which are continuous/bounded. The following theorem is very useful in
multiple ways; the proof can be found in [3].
Theorem 1 (Riesz representation theorem). (1) Any linear functional R (C(T)) can
1
f (t) d(t). In
be identified with a unique measure M (T) such that (f ) = 2
T
addition, kkC(T) = kkM (T) .
(2) Any linear functional
(C0 (R)) can be identified with a unique measure M (R)
R
such that (g) = R g(x) d(x). Furthermore, kkC(R) = kkM (R) .

the Fourier series on T of an integrable function f is given


=
RRecall that
PNby f (n) int
int
f (t)e
dt. Trigonometric polynomials are functions of the form P (t) = n=N an e .
T
P
int

It can be easily verified that P (t) = N


n=N P (n)e . A basic result in Fourier analysis is
PN
that the partial sums n=N f(n)eint do not necessarily converge to the function itself.
Nevertheless, we have convergence in a related series. A proof of this statement can be
found in [2].
Lemma 2 (Fejer). For positive integer N , define the Fej
er kernel by
1 sin2 [(N + 1)x/2]
.
FN (x) =
N +1
sin2 [x/2]
Then for a continuous function f on T, we have the following convolution:

N
X
(FN f )(t) =
1
n=N

|n|
N +1

f(n)eint .

Furthermore, (FN f )(t) converges uniformly to f as N .


As the convergents in the previous lemma are trigonometric polynomials, any function
can be approximated by trigonometric polynomials. Hence trigonometric polynomials
are dense in C(T).

FourierStieltjes Coefficients

FourierStieltjes coefficients are an extension of Fourier coefficients, defined with measures


in M (T):
Definition 3. The FourierStieltjes coefficients of a measure M (T) is a function
on Z given by the expression
Z
1
eint d(t).

(n) =
2 T

Note that if we have an integrable function f , we can identify it with a measure d =


f (t) dt (dt represents the usual
Lebesgue measure).
Computing the FourierStieltjes
R int
R int
1
1
coefficients gives
(n) = 2 T e
d = 2 T e
f (t) dt = f(n), so we really do have
an extension of Fourier coefficients.
2
For
R of functions in L (R), an important result is Parsevals formula:
R Fourier transforms
1
g () d. An analogue holds in the context of FourierStieltjes
f (x)g(x) dx = 2 R f()
R
coefficients in the following sense:

Proposition 4 (Parsevals Formula). If M (T) and f C(T), then


1
2

Z
f (t)d(t) = lim


N
X
n=N

|n|
1
N +1

f(n)
(n).

Proof. Suppose f is any


 continuous
 function. Then by using lemma 2, we can approximate
|n|
f uniformly as lim 1 N +1 f(n)eint . As is a finite measure, we can exchange the
N

limit and integral to yield



N
X
1



|n|
int
f(n)e
f (t)d(t) = lim
d(t)
N T
N
+
1
T
n=N

Z
N
X 
|n|

= lim
f (n) eint d(t)
1
N
N
+
1
T
n=N


N
X
|n|
f(n)
= lim
(n).
1
N
N
+
1
n=N
Z

We remark that without the 1


converge.

|n|
N +1

factor, the right hand side may not necessarily

Bochners theorem is about trying to determine which sequences are the FourierStieltjes
coefficients of a measure. As a first step, we give a necessary and sufficient condition for
a sequence to the FourierStieltjes coefficients of a measure :
Proposition 5. Let {an }
n= be a sequence of complex numbers. Then the following
are equivalent:
(a) There exists M (T) with kk C and
(n) = an for all n.

P


(b) For all trigonometric polynomials P , P (n)an C sup P (t).
tT

Proof. (a) = (b): Given any trigonometric polynomial P , use Parsevals formula to get
Z
N


X



P (n)an = P (t) d kk sup P (t)



tT
T
n=N

where we used the Riesz representation theorem to show that |(f )| kk kf kC(T) =
kk suptT f (x).
3

P
(b) = (a): The linear map P 7
P (n)an defines a linear functional on the space
of trigonometric polynomials. As trigonometric polynomials are dense in C(T), we can
extend our linear functional to C(T).
R
1
f d
By the Riesz representation theorem our linear functional is of the form f 7 2
T
int
for
R M (T) of norm C. Substituting f = e into both P 7
P a suitable measure
1
(n) = an or
(n) = an .
P (n)an and f 7 2
f d immediately gives
T
Consider a sequence {an }nZ which eventually vanishes,
= 0 when |n| > K. Then
PK i.e. an int
it can be checked that the measure given by dN := n=K an e dt satisfies c
N (n) = an
for all n. This allows us to make sense of the following corollary. The proof involves
algebraic manipulations and Parsevals formula only, and hence will not be presented
here.
Corollary 6. A sequence {an }nZ is the FourierStieltjes coefficients of some with
kk C, if, and only if, kN kM (T) C for all N . Here, N is the measure such that
|n|
c
N (n) = (1 N +1 )an whenever |n| N , and zero when |n| > N .

Hergoltzs Theorem

Hergoltzs theorem is the analogue of Bochners theorem on the torus, as in it gives


necessary and sufficient conditions for a sequence to be the FourierStieltjes coefficients
of a positive measure. To prove this, we first need the following lemma:
Lemma 7. A sequence {an }nZ is the FourierStieltjes series of a positive measure if,
and only if, for all N and t T,

N
X
N (t) :=
1

|n|
N +1

n=N

an eint 0.

Proof. First suppose


(n) = an for some positive measure M (T). Let f C(T) be
an arbitrary non-negative function. Then
1
2

#

|n|
f (t)
an eint dt
1
N
+
1
T
n=N


N
X
|n|
=
f(n)
(n)
1
N
+
1
n=N


Z
N
X
|n|

=
1
f (n) eint d(t)
N
+
1
T
n=N

Z 
|n|
f(n)eint d
=
1
N
+
1
ZT
= (FN f )(t) d
(from lemma 2).

1
f (t)N (t) dt =
2
T

"


N
X

Now note that the Fejer kernel is non-negative. As f is non-negative, the convolution
(FN f ) is also non-negative. From the positivity of the measure , our quantity is
non-negative. As this is true for any non-negative function f , we get that N (t) 0.


P
|n|
1

an eint 0. As before, denote N as the measure with


Now suppose N
n=N
N +1


|n|
FourierStieltjes coefficients 1 N +1 an . Then we want to compute kN kM (T) . By
the Riesz representation theorem we just need to compute
 the norm
 of the functional
R
|n|
1
f 7 2 T f dN . However, from earlier discussion, dN = 1 N +1 an dt. Therefore
kN kM (T)

1
= sup
kf kC(T) =1 2

1
f (t) dN = sup
kf k=1 2
T

Z
f (t) N (t) dt.
T

As N (t) is positive, N is a positive measure and the above quantity is maximised when
f (t) = 1. This means that
kN kM (T)

1
=
2


N
X
1

T n=N

|n|
N +1

an eint dt = a0 .

This must be true for all N . Therefore, kN kM (T) are uniformly bounded by a0 . By
corollary 6, we can find a M (T) such that
(n) = an for all n. To show that is a
positive measure, take an arbitrary non-negative function f . Then
Z
Z
f d = lim
f dN 0
N

as this convergence comes from taking the limit in Parsevals formula. Thus is positive.

We remark in the above proof that one can replace the condition N (t) 0 for all N
with N (t) 0 for infinitely many N . The proof remains the same, except that we only
take a subsequence Nj where Nj (t) 0.
Definition 8. A sequence {an }nZ is positive definite if for all sequences of complex
numbers {zn }nZ which have all but a finite number of terms zero, we have
X
anm zn zm 0.
n,mZ

The work from the last few pages can now be combined to prove Herglotzs theorem:
Theorem 9 (Herglotz). A sequence {an }nZ is the FourierStieltjes transform of a positive measure M (T) if, and only if, the sequence is positive definite.
Proof. If is a positive measure with
(n) = an , then
X
n,mZ

anm zn zm =

X Z
n,mZ

2
Z X



eint eimt zn zm d =
zn eint d 0.


nZ

On the(other hand, suppose {an } is positive definite. Fix values t, N . Define numbers
P
eint , |n| N
zn =
. By direct substitution, one can verify that n,m anm zn zm =
0,
else
P
ijx
, where Cj,N = max(0, 2N + 1 |j|). Therefore by positivedefiniteness,
j Cj,N aj e
2N (t) =

2N 
X
j=2N


2N
X
|j|
1
ijx
1
aj e =
Cj,N aj eijx
2N + 1
2N + 1 j=2N
X
anm zn zm 0.

1
2N + 1 n,mZ

We proved that N (t) 0 for even N . The result then follows from lemma 7.

The FourierStieltjes Transform

We define the main object of interest here, the FourierStieltjes transform.


Definition 10. The FourierStieltjes transform of a measure M (R) is a function
on R given by the expression
Z
() =
eix d(x).
R

Like FourierStieltjes coefficients, there is consistency between the definition here, and
the Fourier transform of L1 functions.
If g is an integrable function, identify it with a
R ix
measure d = g dx. Hence () = R e
g(x) dx = g(), which shows consistency in the
transforms.
The RiemannLebesgue lemma on the ordinary Fourier transform states that if g L1 (R),
then g() is uniformly continuous and goes to 0 as || . The uniform continuity still
holds for the FourierStieltjes transform, but () does not necessarily go to 0. A simple
example is = 0 , the measure with mass at 0. It can be easily verified that () = 1 for
all .
We can deduce another version of Parsevals formula in this new context:
Proposition 11 (Parsevals Formula). If M (R) and both g, g are in L1 (R), then
Z
Z
1
g()
() d.
g(x)d(x) =
2 R
R
Proof.
g and g are integrable, then the Fourier inversion formula holds: g(x) =
R If both
1
ix
g

()e
d.
Therefore,
2 R
Z
ZZ
Z
1
1
ix
g(x)d(x) =
g()e d(x) d =
g()
(),
2
2 R
R
R
where we used the integrability of g() to justify the usage of Fubinis theorem.
6

The following proposition gives a necessary and sufficient statement for a function to be
the FourierStieltjes transform of a measure, and is the first step to Bochners theorem
on R.
Proposition 12. If is a continuous function defined on R, then it is the Fourier
Stieltjes transform of a measure if, and only if, there exists a constant C such that
Z

1


C sup |g(x)|
g

()()
d
2

xR
R
for every continuous function g L1 (R) such that g has compact support.
Proof. Firstly, suppose that = . Then the statement follows directly from Parsevals
formula by setting C = kkM (R) .
On the other hand,
suppose our inequality is valid. Then the linear functional which
R
1
maps g to 2 R g()() d is a bounded, continuous linear functional on the set of
continuous functions g such that g is compactly supported. This is a dense subset of
C0 (R). Therefore we can extend to a bounded functional on C0 (R).
By the Riesz representation theorem, can beR represented by a measure M (R),
where kkM (R) C.
R Therefore maps
Rg to R g(x) d(x). Using Parsevals formula
1
1
() d. This must hold for all g, so
here yields that 2 R g()() d = 2 R g()
= .

Connection to FourierStieltjes Coefficients

We have a natural covering map p : R T sending x to x mod 2. This can be used to


pushforward a measure M (R) to a measure M (T). If we extend a continuous
function f on T to a 2-periodic function g on R (which do not necessarily go to zero),
then we get that
Z
Z
g(x) d(x) = f (t) d(t).
R

An immediate consequence is that (n) =


(n) for all integers n. This allows us to relate
the original problem on R to the problem on T. The following theorem is a key step in
doing this:
Theorem 13. If is a continuous function defined on R, then it is the FourierStieltjes
transform if, and only if, there exists a constant C > 0 such that for any choice of > 0,
{(n)}
n= are the FourierStieltjes coefficients of a measure of norm C on T.
Proof. First suppose = . Then (n) = (n) =
(n), where is the pushforward
measure of with respect to the covering. Note that kkM (T) kkM (R) . Denote (x/)
the measure on R which satisfies the following equation for all g:
Z
Z
x
g(x)d( ) =
g(x) d(x).
R

\
We get that k(x/)kM (R) = kkM (R) and (x/)()
= (). Setting = n yields that
(n) = (x/)(n), so {(n)}nZ form the FourierStieltjes coefficients of norm at most
kkM (R) .
To show the converse, we shall use theorem 12. Let g be continuous such that
g is continR
1
g

()()
d
uous and compactly supported. Then by those conditions the integral 2
R
can be approximated by its Riemann sums (where the width of each rectangle is ). For
arbitrary , choose sufficiently small to obtain

X
Z


1

<

g

()()
d
g

(n)(n)
+ .
2
2

R
n
Note that

g(n)
2

is the Fourier coefficient for the function G (t) =

g((t + 2m)/)

mZ

on T. If is sufficiently small, then by the decay of g, we have


sup |G (t)| sup |g(x)| + .
tT

xR

By assumption, (n) =
c (n) for some M (T), with k kM (T) C. Then Parsevals
formula for T gives



X

X


g(n)(n) =
G (n) (n) C sup |G (t)|.



2
tT
n
n
Combining all inequalities gives


X



g(n)(n) < C sup |g(x)| + (C + 1).

2

xR
n

As  > 0 was arbitrary, the condition for proposition 12 is satisfied.


Now here is a necessary and sufficient condition for a function to be a FourierStieltjes
transform of a positive measure. The proof is very similar to theorem 13, so it will not
be presented here.
Proposition 14. If is a continuous function defined on R, then it is the Fourier
Stieltjes transform of a positive measure if, and only if, there exists a constant C > 0
such that for any choice of > 0, {(n)}
n= are the FourierStieltjes coefficients of
a positive measure on T.

Bochners Theorem

Definition 15. A function defined on R is positive definite if for all 1 , . . . , N R


and z1 , . . . , zN C, we have
N
X

(j k )zj zk 0.

j,k=1

Finally, we have the machinery to prove the main subject of this paper.
Theorem 16. A function defined on R is the FourierStieltjes transform of a positive
measure if, and only if, it is continuous and positive definite.
Proof. First assume that = for a positive measure M (R). Given numbers
1 , . . . N R and z1 , . . . zN C, we get
N
X
j,k=1

(j k )zj zk =

Z X
N

eij x zj eik x zk d(x)

R j,k=1

2
Z X
N


ij x
=
zj e

d(x) 0.

R
j=1

For the other direction, if is positive definite, then by definition, {(n)} is a positive
definite sequence for all . Herglotzs theorem guarantees a measure , such that
c (n) =
(n). By proposition 14, there must exist such that = .

References
[1] Katznelson, Y. (2004). An Introduction to Harmonic Analysis (3rd edn), Cambridge
Mathematical Library.
[2] Grafakos, L. (2003). Classical and Modern Fourier Analysis, Prentice Hall.
[3] Rudin, W. (1987). Real and Complex Analysis (3rd edn), McGrawHill Book Company.