You are on page 1of 26

Chapter 5 Some topics on Fourier trans-

form
5.1 Poisson summation formula
This formula due to S.D.P oisson gives a connection between the series of
functional values evaluated at certain period and the series of values of the function’s
Fourier transform, evaluated at the same period. This is obtained as a corollary of
the following theorem.
∞ 1 P ∞
n

einx .
P
Theorem 5.1.1. Let f ∈ S(R). Then f (x + 2πn) = fb 2π
n=−∞ 2π n=−∞

P
Proof. Define g(x) = f (x + 2πn). Then
n=−∞


X
g(x + 2π) = f (x + 2π(n + 1))
n=−∞
X∞
= f (x + 2πm)
m=−∞

= g(x).

Thus g is a periodic function with period 2π. Consider


1
gb(n) = g(x)e−inx dx

−π
Zπ ∞
1 X
= f (x + 2πk)e−inx dx

−π k=−∞

121
∞ Z π
1 X
= f (x + 2πk)e−i (x+2πk)n dx
2π k=−∞
−π
(2k+1)π
∞ Z
1 X
= f (y)e−iyn dy
2π k=−∞
(2k−1)π
Z∞
1 n 1 b
f (y)e−2πi( 2π )y dy = n

= f 2π
.
2π 2π
−∞

Here, the interchange of sum and integral is permissible as f is assumed to be in


S(R). Since f ∈ S(R), one can show that the Fourier series of g converges absolutely
and hence

∞ ∞ ∞
X X
inx 1 X b n  inx
g(x) = f (x + 2πn) = gb(n)e = f e ,
n=−∞ n=−∞
2π n=−∞ 2π

proving our assertion.

Putting x = 0 in Theorem 5.1.1, we obtain the following:

∞ ∞
1 n
P P 
Corollary 5.1.2. f (2πn) = 2π
fb 2π
for f ∈ S(R).
n=−∞ n=−∞

5.2 Uncertainty principle


In quantum mechanics, the famous Heisenberg uncertainty principle states
that the position and the momentum of a particle cannot be determined simulta-
neously with a reasonable good degree of accuracy or certainty. From the Fourier
analysis point of view this leads to the uncertainty principle between a given function
and its Fourier transform. In general, uncertainty principle can be broadly classi-

122
fied into two types: qualitative uncertainty principle and quantitative uncertainty
principle.
In quantitative uncertainty principle, the inequalities gives information about
the relation of a function and its Fourier transform. In the case of qualitative un-
certainty principle, the theorem can establish the simultaneous behavior of both a
function and its Fourier transform. Towards quantitative uncertainty principle, we
discuss Heisenberg’s uncertainty principle and towards qualitative uncertainty prin-
ciple we discuss Hardy’s theorem named after Hardy in this section.

5.2.1 Heisenberg’s uncertainty principle


Theorem 5.2.1. Let ψ ∈ S(R) be such that kψk2 = 1. Then

Z∞ Z∞
2 2 1
x |ψ(x)| dx · ξ 2 |ψ(ξ)|
b 2 dξ ≥ .
16π 2
−∞ −∞

q
2
Further, equality holds if and only if ψ(x) = Ae−Bx , where B > 0 and A2 = 2B
π
.
Proof. Consider

Z∞ Z∞
2 ∞

xd(|ψ(x)| ) = x |ψ(x)|2 −∞ − |ψ(x)|2 dx
−∞ −∞

= −kψk22 = −1.

Hence
Z∞
1=− xd(|ψ(x)|2 )
−∞
Z∞ h 0 i
0
=− x ψ(x)ψ (x) + ψ (x)ψ(x) dx.
−∞

123
Thus ∞
Z
0 0
1 = x[ψ(x)ψ (x) + ψ (x)ψ(x)]dx


−∞
Z∞
≤2 |x||ψ(x)||ψ 0 (x)|dx.
−∞
 21   21
Z∞ Z∞

≤ 2 x2 |ψ(x)|2 dx  |ψ 0 (x)|2 dx , (5.1)


−∞ −∞

on applying Schwarz inequality. But using Plancherel formula, we get,

Z Z
0 2
|ψ (x)| = |ψb0 (ξ)|2 dξ
R R
Z
2
= 4π ξ 2 |ψ(ξ)|
b 2 dξ.
R
From (5.1), it follows that
  21   21
Z Z
1 ≤ 2 x2 |ψ(x)|2 dx 4π 2 ξ 2 |ψ(ξ)|
b 2 dξ 
R R
  21   21
Z Z
= 4π  x2 |ψ(x)|2 dx  ξ 2 |ψ(ξ)|
b 2 dξ  .
R R

Squaring both sides, we get

Z Z
b 2

2 2 2
1 ≤ 16π x |ψ(x)| dx ξ 2 ψ(ξ) dξ,
R R

proving our assertion. The equality holds if equality holds in Schwarz inequality.
This leads to the following. We can write ψ 0 (x) = βxψ(x) for some constant β,

124
βx2
ψ 0 (x)
which means ψ(x)
= βx. On integration, we get ψ(x) = Ae 2 . But ψ ∈ S(R),

r β = −2B, B > 0. As kψk2 = 1, we get


which forces that β must be negative. Let
∞ 2B
2
e−2B dx = 1 which leads to A2 =
R
A2 .
−∞ π

Definition 5.2.2. A function f : C → C is said to be of exponential type T ,


0 < T < ∞, if |f (z)| ≤ ceT |z| ∀ z ∈ C and a constant c.
In order to prove Hardy’s theorem, we make use of the following well known
theorem in complex analysis.
Theorem 5.2.3. (Phragmen-Lindelof theorem) If f is analytic and of exponential
type in a sector D with opening less than π and if it is continuous in D, |f | ≤ M on
∂D, then |f | ≤ M throughout D.
Theorem 5.2.4. (Hardy) Suppose f is a measurable function on R such that |f (x)| ≤
2 2
ce−ax and |fb(ξ)| ≤ ce−bξ , where a, b and c are positive constants. Then
(i) f (x) = 0 if ab > π 2 .
2
(ii) f (x) = c e−ax if ab = π 2 .
Proof. Assume that the result is proved for a = b = π. For r > 0, define fr (x) = f (rx).
Take r2 = πa . Then

2 2 2
|fr (x)| = |f (rx)| ≤ c e−ar x = c e−πx .
1 −bξ2 ab 2
|fbr (ξ)| = fb(ξ/r) ≤ c e r2 = c e− π2 ξ . (5.2)

r

2
If ab = π 2 , then fr (x) = c e−πx . If ab > π 2 , the inequality (5.2) is impossible unless
c = 0. Therefore, it is enough to prove for a = b = π. Let a = b = π and w = ξ + iη.

125
Then fb has a holomorphic extension to C and it satisfies

Z
−2πixw

|fb(w)| = f (x) e dx

R
Z
−2πxiξ
2πxη

= f (x) e e dx

ZR Z
2
≤ |f (x)| e2πxη dx ≤ c e−πx e2πxη dx
R R
Z
−π(x−η)2 πη 2 2
≤c e e dx = ceπη .
R

Clearly, fb is continuous. For any simple closed curve C in C, an application of Fubini’s


theorem leads to
 
Z Z Z
fb(w)dw =  f (x) e−2πixw dx dw
C C
R 
Z Z
e−2πixw dw  f (x)dx = 0.

=  (5.3)
R C

Thus, it follows from Morera’s theorem that, fb is an entire function. Now assume

cn w2n .
P
that f is an even function. Then fb is also an even function. Let fb(w) =
n=0
√ ∞
cn wn . Then h(w) is an entire function. We claim that
P
Define h(w) := fb( w) =
n=0 √
e h(w) is a constant on C. Suppose w = R > 0. Then |h(w)| = fb( w) ≤ c e−πR .
πw

But, if we take w = Reiθ , we get

√ √ 2
|h(w)| = |fb( w)| ≤ c eπ(Im w) .

126
2 θ
Hence, |h(w)| ≤ c eπR sin 2. Therefore h is of exponential type. Now consider,
( δ
)
iπw e−iδ/2 iπR ei(θ− 2 )
exp = exp

sin 2δ sin 2δ


( )
iπR[cos(θ − δ/2) + i sin(θ − δ/2)]
= exp

sin 2δ


( )
−πR sin(θ − δ/2)
= exp .
sin 2δ


If θ = 0, then
−δ
!
iπw e 2
exp h(w) = eπR |h(w)| ≤ eπR · c e−πR = c.

sin(δ/2)

If θ = δ, then
−δ
!
iπw e 2 2
exp h(w) = e−πR |h(w)| ≤ e−πR c eπR sin (δ/2) ≤ c if 0 < δ < π.

sin(δ/2)

An application of Phragmen-Lindelof theorem leads to

 
πR sin(θ − δ/2)
|h(w)| ≤ c exp for 0 ≤ θ ≤ δ.
sin(δ/2)
Letting δ → π, we get,

|h(w)| ≤ c exp(−πRcosθ) for, 0 ≤ θ ≤ π.

Similarly, h(w) will satisfy the same bound in the lower half plane. Thus

|h(w)| ≤ c exp(−πR cosθ) for all w ∈ C.

127
Hence,

|eπw h(w)| ≤ c for all w ∈ C.

By applying Liouville’s theorem, we conclude that eπw h(w) is a constant on C, which


2
implies that h(w) = c e−πw and consequently f (w) = c e−πw for w ∈ C. Now assume
that f is odd. Then fb is also odd and fb(0) = 0.
fb(ξ) 2
If g(ξ) = , then g is even. By the previous argument, g(x) = c e−πx and
ξ
−πξ 2 2
f (ξ) = c ξ e
b . But by the given hypothesis, |fb(ξ)| ≤ c e−πξ . This is impossible
unless c = 0.

5.3 Paley-Wiener theorem


Consider the analytic extension of a Fourier transform of a square integrable
function to the complex plane. This theorem named after P aley and W iener gives
a necessary and sufficient condition for such extended function to become an entire
function of exponential type. The proof of this theorem also uses complex analysis
techniques.
Theorem 5.3.1. Let f ∈ L2 (R). Then fb can be extended as an entire function of
exponential type less than or equal to 2πT if and only if f (x) = 0 for |x| > T .
Proof. Suppose f (x) = 0 for |x| > T . Then f is integrable on R. Consider for
w = ξ + iη

Z
fb(w) = f (x)e−2πiwx dx,
R
ZT
= f (x)e−2πwx dx.
−T

128
Then
ZT ZT
−2πiwx
|fb(w)| ≤ |f (x)||e | dx ≤ |f (x)|e2π|w|T dx
−T −T
ZT
≤ e2π|w|T |f (x)| dx = kf k1 e2π|w|T .
−T

Therefore, fb is an entire function of exponential type 2πT . Conversely, suppose fb


can be extended to an entire function of exponential type 2πT.
In order to show that f = 0 for |x| > T, consider the function |h(w)| =
1
R2
fb(w + y) dy. Then h(w) is an entire function and of exponential type 2πT. In fact,
− 12
we have,
1
Z2 Z12 Z2
1

h(w) = fb(w + y) dy ≤ |fb(w + y)| dy ≤ ce2π|w+y|T dy,


− 1 −1 −1
2 2 2
1
Z 2

2π|w|T 0
= ce e2π|y|T dy, = c e2π|w|T .
− 12

Now h can be viewed as follows:

1
Z∞
 
Z2
h(w) =  f (x) e−2πi(w+y)x dx dy
− 21 −∞
1
Z∞ Z2
= f (x) e−2πi(w+y)x dy dx
−∞ − 1
2

129
1
 
Z∞ Z2
= f (x) e−2πiwx  e−2πiyx dy  dx
 

−∞ − 12
Z∞  
−2πiwx sin πx
= f (x) e dx
πx
−∞
 b
sin πx
= f (x) (w)
πx
On the real line, w = ξ, we have
1
Z2
|h(w)| = |h(ξ)| ≤ |fb(ξ + y)| dy.
− 12

Applying Cauchy-Schwartz inequality, we get


 1  21  1  21
Z2 Z2
2
|h(ξ)| ≤  |f (ξ + y)| dy   dy  ,
 b   

− 21 − 12

showing that

|h(w)| ≤ kfbk2 = kf k2 , and


 1 
Z∞ Z∞ Z2
khk22 = |h(ξ)|2 dξ, = 2
 |fb(ξ + y)| dy  dξ.
 

−∞ −∞ − 12
1
Z2 Z∞
= |fb(ξ + y)|2 dξ dy
− 21 −∞
1
Z2
= kfbk22 dy = kfbk22 = kf k22 .
− 21

Let B > T, and consider the function h(w)e2πiBw . Then


0 0
|h(w)e2πiBw | ≤ |h(w)| e2πB|w| ≤ c e2π|w|T e2πB|w| ≤ c e2π(B+T )|w| .

130
Thus h(w) e2πiBw is of exponential type 2π(B + T ). On the positive real line, w = R,
we have, |h(R)e2πiBR | = |h(R)| ≤ kf k2 . On the positive imaginary axis, w = ib, b > 0,
we have,

0 0
|h(ib) e2πiB(ib) | = |h(ib)|e−2πBb ≤ e−2πBb (c e2πbT ) = c e−2π(B−T )b ,

which goes to 0 as b → ∞. Thus h(w) e2πiBw is bounded on the positive real axis
and the positive imaginary axis. Therefore by Phragmen-Lindelof lemma, we have
|e2πiBw h(w)| ≤ c in the first quadrant.
Similarly, |e2πiBw h(w)| ≤ c in the second quadrant. In other words, for w =

Reiθ , 0 ≤ θ ≤ π, we have |h(Reiθ ) e2πiBRe | ≤ c, which means |h(Reiθ )| ≤ c e2πBR sin θ,
for 0 ≤ θ ≤ π. Now for a fixed A > 0 and x > B > T , consider the integral

ZR
e2πiwx
I= h(w) dw. (5.4)
1 − Aiw
−R
e2πiwx
Now using the Cauchy theorem for the semi circle to the function h(w), we
1 − Aiw
get,

exp(2πiReiθ )h(Reiθ ) iθ
I = −i R e dθ.
1 − AiReiθ
0
Then

|h(Reiθ )|| exp(2πiRx(cosθ + i sin θ))|
|I| ≤ R dθ
|1 − AiReiθ |
0

ce2πBR sin θ e−2πRx sin θ
≤ R dθ
|1 − AiReiθ |
0

131

cR
= e−2πR(sin θ)(x−B) dθ.
AR − 1
0

Now letting R → ∞, and applying inverse Fourier transform, we can conclude from
h(w)
(5.4) that = 0. Now letting A → 0, we conclude that h(w) = 0.
1 − Aiw  b
sin πx
Recall h(w) = f (x) (w). Thus f (x) = 0 for a.e x, x > B > T . In a
πx
similar way, we can show that f (x) = 0 for x < −T a.e. Consequently f (x) = 0 for
|x| > T a.e.

5.4 Wiener’s theorem


We know that L1 (R) is translation invariant. Now start with a closed translation-
invariant subspace Y of L1 (R). How big this space can be? Wiener’s theorem says
that if the zero set of Y (defined below) is empty, then Y coincides with L1 (R).

Definition 5.4.1. Let Y be a subspace of L1 (R). The zero set of Y , denoted by


Z(Y ) is defined to be the set of all ξ ∈ R such that fb(ξ) = 0 for all f ∈ Y . In other
words, Z(Y ) = {ξ ∈ R : fb(ξ) = 0 ∀ f ∈ Y }.

Theorem 5.4.2. (Wiener) If Y is a closed translation invariant subspace of L1 (R)


and if Z(Y ) is empty, then Y = L1 (R).
In order to prove the theorem we first prove the following lemmas.
Lemma 5.4.3. Let f ∈ L1 (R) and t ∈ R. Then for a given  > 0, there exists
h ∈ L1 (R), khk1 <  such that (f + h)b(ξ) = fb(t), for every ξ in some neighborhood
of t.

132
Proof. Let ϕ ∈ L1 (R) be such that ϕ
b = 1 in some neighborhood of the origin. Fix
δ > 0. For x ∈ R, define, ϕtδ (x) = eitx Dδ ϕ(x). Using this, we define,

hδ (x) = ϕtδ (x)fb(t) − (ϕtδ ∗ f )(x). (5.5)

Then
Z Z
−ity
hδ (x) = ϕtδ (x) f (y) e dy − ϕtδ (x − y)f (y) dy
R R
Z
f (y) ϕtδ (x) e−ity − ϕtδ (x − y) dy.
 
=
R
Thus
Z
|hδ (x)| ≤ |f (y)||ϕtδ (x) e−ity − ϕtδ (x − y)| dy.
R
Consider
ϕδ (x) e−ity − ϕtδ (x − y) = eitx Dδ ϕ(x) e−ity − e−it(x−y) Dδ ϕ(x − y)
t

= eit(x−y) δ −1 ϕ xδ − eit(x−y) δ −1 ϕ x−y


 

δ

= eit(x−y) δ −1 ϕ xδ − ϕ x−y
 

δ

= δ −1 ϕ xδ − ϕ x−y
 
.
δ

Thus
Z
−1 x−y
x
 
|hδ (x)| ≤ δ |f (y)| ϕ δ
−ϕ δ
dy.
R
Hence, it follows that
 
Z Z Z
khδ k1 = |hδ (x)| dx ≤ δ −1  |f (y)| ϕ x−y
x
 
δ
−ϕ δ
dy  dx
R R R

133
 
Z Z
|f (y)| δ −1 x−y
x
 
= ϕ
δ
−ϕ δ
dx dy,
R R
by applying Fubini’s theorem. But
Z Z
−1 x−y
− ϕ δ dx = ϕ(u) − ϕ u − yδ du ≤ 2kϕk1 ,
x  
δ ϕ
δ
R R
and
Z
ϕ(u) − ϕ u − y du = kϕ − τ y ϕk1 → 0 as δ → ∞.

δ
δ
R

In other words khδ k1 → 0 as δ → ∞. Further, using (5.5)

hbδ (ξ) = ϕtδ (ξ)fb(t) − ϕ


b
bδ (ξ)fb(ξ)

b = 1 in some neighborhood of the origin, (ϕtδ ) = 1 in some neighborhood Vδ


Since ϕ
b

of t. Thus, for every ξ ∈ Vδ ,

hδ (ξ) = fb(t) − fb(ξ),


b

proving our assertion.


Definition 5.4.4. Let V be an open subset of R. We say that f ∈ S 0 (R) vanishes in
V if hϕ, f i = 0, ∀ ϕ ∈ S(R), with support in V . Let Ω be the union of all open sets
V in R in which f vanishes. The complement of Ω is called the support of f .
Lemma 5.4.5. Let ϕ ∈ L∞ (R). If Y is a subspace of L1 (R) such that f ∗ ϕ = 0 for
every f ∈ Y . Then the support of the tempered distribution ϕ
b is contained in Z(Y ).
Proof. In order to prove that support of ϕ
b is contained in Z(Y ), it is enough to show
that Z(Y )c ⊂ (supp ϕ)
b c.

134
In view of the definition of support of a tempered distribution, it is enough
to show that any t ∈ Z(Y )c will have a neighborhood in which ϕ
b will vanish. Fix
t ∈ Z(Y )c . By definition of Z(Y ), there exists f ∈ Y such that fb(t) 6= 0.
Without loss of generality, we can assume that fb(t) = 1. Then by lemma 5.4.3
for a given  > 0, there exists h ∈ L1 (R) such that khk1 <  and (f + h)b(ξ) = fb(t) = 1,
for every ξ in some neighborhood V of t. Thus, it is enough to prove that ϕ
b vanishes
in V .
In other words, we should show that hg, ϕi
b = 0 for every g ∈ S(R) with the
support of g in V . By inverse Fourier transform theorem on S(R), it is enough to
show that hψ, b = 0 for every ψ ∈ S(R) whose Fourier transform ψb has support in
b ϕi

V . But hψ, b = hψ,


b ϕi ϕi = (ϕ ∗ ψ)(0). Thus, it is sufficient to show that ϕ ∗ ψ = 0.
bb

Fix ψ ∈ S(R) such that support ψb ⊂ V . Construct a sequence of functions {un } as


follows:
Let u0 = ψ, u1 = h ∗ u0 , u2 = h ∗ u1 . Inductively, we have un = h ∗ un−1 , n ≥ 1.

P PN
Consider the series un . Let SN = un .
n=0 n=0

Consider, for N ≥ M ,

X N XM
kSN − SM k1 = un − un


n=0 n=0 1

X N
= un


n=M +1 1
N
X N
X
≤ kun k1 ≤ khkn1 kψk1
n=M +1 n=M +1
N
X
= kψk1 khkn1 .
n=M +1

135
N
khkn1 → 0 as M, N → ∞. Thus, {SN } is a Cauchy sequence
P
Since khk1 < 1,
n=M +1
in L1 (R). Since, L1 (R) is complete, {SN } converges to s in L1 (R). Thus, the series

un is summable to s in L1 (R). Since (f + h)b(ξ) = 1 ∀ ξ ∈ V , and support
P
n=0
ψb ⊆ V , we have fb(ξ) + b
h(ξ) = 1 for all ξ ∈ supp ψ.
b Thus ψ(ξ)
b fb(ξ) + ψ(ξ)
b b h(ξ)
= ψ(ξ)
b for all ξ ∈ V . Hence, [1 − b
h(ξ)]ψ(ξ)
b = ψ(ξ)
b fb(ξ). Recall that for every ξ,
1   ∞
hm (ξ)ψ(ξ)
P
|h(ξ)| ≤ khk∞ ≤ khk1 < 1. Thus, ψ(ξ) =
b b b ψ(ξ)f (ξ) =
b b b b fb(ξ).
1 − h(ξ)
b m=0

hm (ξ)ψ(ξ)
P
Thus, ψb = b b fb(ξ) = sbfb. Taking the inverse Fourier transform, we get,
m=0
ψ = s ∗ f . Consequently, ψ ∗ ϕ = s ∗ f ∗ ϕ = 0, proving our assertion.

Proof of Wiener’s theorem. By the Hahn-Banach theorem, it is enough to show


that Y ⊥ , annihilator of Y is {0}. Let φ ∈ L∞ (R) such that hf, ϕi = 0 ∀ f ∈ Y . This
R
means f ∗ ϕ(0) = f (x)ϕ(−x)dx = 0 ∀ f ∈ Y . Then it follows from Lemma 5.4.5
R
that supp ϕ
b is contained in Z(Y ). But given that Z(Y ) is empty. Thus support ϕ
b
b = 0. Since the Fourier transform from S 0 (R) into S 0 (R) is one-one,
= 0. Hence ϕ
it follows that ϕ = 0 as a tempered distribution. Thus ϕ = 0 in L∞ (R), proving
Y ⊥ = 0.
Theorem 5.4.6. (Wiener-Tauberian Theorem) Suppose ϕ ∈ L∞ (R), b
h(ξ) 6=
0 ∀ ξ ∈ R, and lim (h ∗ ϕ)(x) = ab
h(0). Then lim (f ∗ ϕ)(x) = afb(0) for all
x→±∞ x→±∞
1
f ∈ L (R).
Proof. Let Y = {f ∈ L1 (R) : lim (f ∗ ϕ)(x) = afb(0)}. Let f, g ∈ Y and α, β ∈ R.
x→±∞
1
Then clearly, αf + βg ∈ L (R). Now,

lim ((αf + βg) ∗ ϕ)(x) = α lim (f ∗ ϕ)(x) + β lim (g ∗ ϕ)(x)


x→±∞ x→±∞ x→±∞

= αafb(0) + βab
g (0) = a(αf + βg)b(0).

136
Hence, αf + βg ∈ Y proving that Y is a subspace. Let s ∈ R, f ∈ Y . Then
(τs f )b(0) = fb(0). Again lim ((τs f ) ∗ ϕ)(x) = lim τs (f ∗ ϕ)(x) = lim (f ∗ ϕ)
x→±∞ x→±∞ x→±∞
(x − s) = afb(0) = a(τs f )b(0). This proves that Y is translation invariant.
Now we claim that Y is closed. Let {fn } be a sequence in Y such that kfn −
f k1 → 0 as n → ∞ for some f ∈ L1 (R). We have to show that f ∈ Y . We have
(fn ∗ ϕ)(x) → afbn (0) as x → ± ∞.
Consider, ∀ x ∈ R,

Z Z

|fn ∗ ϕ(x) − f ∗ ϕ(x)| = fn (x − y)ϕ(y)dy − f (x − y)ϕ(y)dy


R
Z
≤ |fn (x − y) − f (x − y)| |ϕ(y)| dy
R
Z
= kϕk∞ |fn (x − y) − f (x − y)| dy
R

= kϕk∞ kfbn − f k1 → 0,

as n → ∞. Further, |fbn (0) − fb(0)| ≤ kfbn − fbk∞ ≤ kfn − f k1 → 0 as n → ∞.


Here kϕk∞ denotes the L∞ -norm of φ and kfbn − fbk∞ denotes the supremum norm of
fbn − fb. Thus fn ∗ ϕ converges to f ∗ ϕ uniformly on R. Thus

lim (f ∗ϕ)(x) = lim lim (fn ∗ϕ)(x) = lim lim (fn ∗ϕ)(x) = lim afbn (0) = afb(0),
x→±∞ x→±∞ n→∞ n→∞ x→±∞ n→∞

showing that f ∈ Y . Now, it follows from Wiener’s theorem that Y = L1 (R), thus
proving our assertion.

137
5.5 A multiplier theorem for L1(R)
Definition 5.5.1. Let T : L1 (R) → L1 (R) be continuous linear transformation. If
T commutes with translation, viz, T τs = τs T ∀ s ∈ R, then we say that T is a
multiplier for L1 (R).
From the above definition, it is not really clear why it is called a multiplier.
We shall show in (statement v of ) the following theorem that in the Fourier domains,
T acts as an multiplication operator.
Theorem 5.5.2. Let T: L1 (R) → L1 (R) be a continuous linear transformation. Then
the following are equivalent.
(i) T τs = τs T ∀ s ∈ R.
(ii) T (f ∗ g) = T f ∗ g = f ∗ T g ∀ f, g ∈ L1 (R).
(iii) There exists a unique µ ∈ M(R) such that, T f = µ ∗ f ∀ f ∈ L1 (R), where
M(R) denotes the Banach algebra of all finite Borel measures on R.
bfb for every f ∈ L1 (R), where
(iv) There exists a unique µ ∈ M(R) such that (T f )b = µ
µ
b denotes the Fourier-Stieltjes transform of µ.
(v) There exists a unique φ ∈ L∞ (R) such that (T f )b = φfb for every f ∈ L1 (R).
Proof. Assume (i). In order to show that T (f ∗ g) = T f ∗ g for every f, g ∈ L1 (R),
it is enough to show that hT (f ∗ g), hi = hT (f ∗ g), hi ∀ h ∈ L∞ (R). Let T ∗ denote
the adjoint of T . Consider

Z
hT f ∗ g, hi = h(t)(T f ∗ g)(t) dt
R
 
Z Z
= h(t)  g(s)T f (t − s)ds dt
R R

138
 
Z Z
= g(s)  h(t)T f (t − s)dt ds
R
R 
Z Z
= g(s)  h(t)τs T f (t)dt ds
R
R 
Z Z
= g(s)  h(t)T τs f (t)dt ds
R
R 
Z Z
= g(s)  T ∗ h(t)τs f (t)dt ds
R R
 
Z Z
= T ∗ h(t)  g(s)f (t − s)ds dt
R R
Z
= h(t)T (f ∗ g)(t)dt
R

= hT f ∗ g, hi ,

for every h ∈ L∞ (R), using the assumption of (i), Fubini’s theorem and change of
variables. In a similar way, we can show that

T (f ∗ g) = f ∗ T g ∀ f, g ∈ L1 (R).

Thus (i) ⇒ (ii). Now we shall assume (ii). In order to prove (iii), let {gα } be an
approximate identity in L1 (R) such that kgα k1 ≤ 1. Consider, kT gα ∗ f − T f k1 =
kT (gα ∗ f ) − T f k1 ≤ kT kkgα ∗ f − f k which goes to zero as α → ∞. As kT gα k1 ≤
kT k, kgα k1 ≤ kT k, {T gα } is a bounded subset of L1 (R). Since L1 (R) ⊂ M(R) and
kf kM(R) ≤ kf kL1 (R) for every f ∈ L1 (R), we get {T gα } is a norm bounded subset

139
of M(R). But M(R) can be viewed as a dual space of C0 (R). By applying Banach
Alaoglu theorem, we get a unique µ ∈ M(R) and a subset {T gβ } of {T gα } such that
T gβ → µ in the weak * topology and also hT f, gi = limhT gβ ∗ f, gi = hµ ∗ f, gi, for
β
all f, g ∈ Cc (R).
As Cc (R) is dense in C0 (R), we get T f = µ ∗ f as elements of M(R) and
in particular as element of L1 (R). As Cc (R) is dense in L1 (R), we get T f = µ ∗ f
for every f ∈ L1 (R), proving (iii). Now, (iv) and (v) are obvious. To prove (i) by
assuming (v), let f ∈ L1 (R). For s ∈ R, consider

−2πisξ b
(T τs f )b(ξ) = φ(ξ)(τ
b s f ) (ξ) = φ(ξ)e
b b f (ξ)
b fb(ξ)e−2πisξ = (T f )b(ξ)e−2πisξ
= φ(ξ)

= (τs T f )b(ξ).

Then (i) follows from the uniqueness of the Fourier transform.

5.6 Solved problems


5.6.1. Let 
 1 − |x| if |x| ≤ 1
g(x) =
 0 otherwise.

(a) Find the Fourier transform of g.



P 1 π2
(b) Apply the Poisson summation formula to g to show that 2
= ,
n=−∞ (n + α) (sin πα)2
where α is real, but not an integer.

P 1 π
(c) Prove as a consequence that = , where α is real, but not an
n=−∞ (n + α) tan πα
integer.

140
Solution. (a) We have

Z0 Z1
gb(ξ) = (1 + x) e−2πixξ dx + (1 − x) e−2πixξ dξ
−1 0
1 (−1)  −2πiξ
1 − e2πiξ +
  
=− 2 2
−e +1
(2πiξ) (2πiξ)
1
1 − e2πiξ − e−2πiξ + 1
 
=− 2
(2πiξ)
e2πiξ 2 e−2πiξ
= − +
(2πiξ)2 (2πiξ)2 (2πiξ)2
 πiξ 2  πiξ   −πiξ   −πiξ 2
e e e e
= −2 +
2πiξ 2πiξ 2πiξ 2πiξ
2 2
e − e−πiξ
 πiξ   
sin πξ
= = .
2πiξ πξ

(b) Let gα (x) = e2πiαx g(x). Applying Poisson summation formula to gα ,


X ∞
X
gbα (n) = gα (n)
n=−∞ n=−∞

2 X ∞
π2

sin πα 1 1 X
= 1 ⇒ = .
π2 n=−∞
(n + α)2 n=−∞
(n + α)2 (sin πα)2


P 1 π2
(c) By (b) 2
= . Integrating both sides, we obtain
n=−∞ (n + α) (sin πα)2


X 1 π
= .
n=−∞
n + α tan πα

5.6.2. (a) Suppose a < b, and f is the function such that f (x) = 0 if x ≤ a or x ≥ b
−1 −1
and f (x) = e x−a e b−x if a < x < b. Show that f is infinitely differentiable on R.

141
(b) Prove that there exists an infinitely differentiable function F on R such that
F (x) = 0 if x ≤ a, F (x) = 1 if x ≥ b and F is strictly increasing on [a, b].
(c) Let δ > 0 be so small that a + δ < b − δ. Show that there exists an infinitely
differentiable function g such that g is zero if x ≤ a or x ≥ b, g is one on [a + δ, b − δ]
and g is strictly monotonic on [a, a + δ] and [b − δ, b].
Solution. Given 
−1 −1
 e x−a e b−x a < x < b
f (x) =
 0 otherwise.

In order to prove that f is infinitely differentiable it is enough to prove that the


function g defined as 
 e− x1 x > 0
g(x) =
 0 otherwise

is infinitely differentiable. In order to prove this it is enough to verify only at x = 0.


Note that g is continuous. We prove by induction that


Pk (x) − 1
(k)

x2k
e x x>0
g (x) =
 0 otherwise,

where Pk (x) is any polynomial in x. This is clearly true for k = 0. Assume that this
is true for any k ≥ 0. By the product rule, we have

Pk0 (x) − 1 2kPk (x) − 1 Pk (x) − 1 x2 Pk0 (x) − 2kxPk (x) + Pk (x) − 1
g (k+1) (x) = e x − e x + e x = e x
x2k x2k+1 x2k+2 x2k+1

which proves our claim. Further notice that lim g (k) (x) = 0 ∀ k ≥ 0. Hence g is
x→0
differentiable.

142
(b) By (a) we have an infinitely differentiable function f given by


−1 −1
 e x−a e b−x a < x < b
f (x) =
 0 otherwise.

Rx
f (x)dx
a
Note that supp(f ) = [a, b]. Define F (x) = Rb
. This F satisfies all the require-
f (x)dx
a
ments.
(c) Define functions F and G as follows:

Rx −1 −1
e x−a e a+δ−x dx
a
F (x) = a+δ
R −1 −1
e x−a e a+δ−x dx
a
Rb −1 −1
e x−(b−δ) e b−x dx
x
G(x) = .
Rb −1 −1
e x−(b−δ) e b−x dx
b−δ

Both F and G are infinitely differentiable. Define the function g as






 F (x) a<x<a+δ


 1 a+δ ≤x≤b−δ

g(x) =



 G(x) b−δ <x<b


 0

otherwise

This g satisfies all the requirements.

143
5.6.3. Define T : R → Lp (R), 1 ≤ p < ∞ by T (x) = τx f , where f ∈ Lp (R). Show
that T is continuous on R.
Solution. Define T : R → Lp (R) as T (x) = τx f , where f ∈ LP (R), 1 ≤ p < ∞. We
first show that T is continuous on R when f ∈ Cc (R). Then there exists a compact
subset K of R such that f (x) = 0 for all x ∈
/ K. We choose M > 0 such that [−M, M ]
and f (M ) = f (−M ) = 0. Then f is uniformly continuous on K 0 = [−M, M ] choose
 > 0. There exists δ > 0 such that for all x, y ∈ K 0 with |x − y| < δ, we have

|f (x) − f (y)| < 1 . Take x, y ∈ R with |x − y| < δ. Now
(2M ) p

Z
kTx f − Ty f kpp = |τx f (u) − τy f (u)|p du
R
Z
= |f (u − x) − f (u − y)|p du
R
Z
= |f (u − x) − f (u − y)|p du.
K0

If u − x and u − y belong to K 0 for u ∈ K 0 then by uniform continuity |f (u − x) −


f (u − y| < 
1 / K 0 and u − y ∈
. If u − x ∈ / K 0 then using f (M ) = f (−M ) = 0 and
(2M ) p
uniform continuity of f we have


|f (u − x) − f (u − y)||f (M ) − f (u − y)| < 1
(2M ) p

Since |M −(u−y)| ≤ |(u−x)−(u−y)| < δ for all u ∈ K 0 , |f (u−x)−f (u−y)| < 


1 .
(2M ) p
p
Therefore kTx f −T yf k ≤ 2M
2M = p for all |x−y| < δ which implies T is continuous
when f ∈ Cc (R). Now let f ∈ Lp (R). Since Cc (R) is dense in Lp (R) there exists

144
ϕ ∈ Cc (R) such that kf − φkp < . Then by the previous result for each  > 0 there
exists δ > 0 such that x, y ∈ R with |x − y| < δ implies kτx ϕ − τy ϕkp < .

kT (x) − T (y)kp = kτx f − τy f kp

≤ kτx f − τx ϕkp + kτx ϕ − τy ϕkp + kτy ϕ − τy f kp

= kτx (f − ϕ)kp + kτx ϕ − τy ϕkp + kτy (f − ϕ)kp

= 2k(f − ϕ)kp + kτx ϕ − τy ϕkp

< 2 +  = 3

kT (x) − T (y)k <  for |x − y| < δ. Thus T is continuous on R.

5.7 Exercises
5.7.1. Let f ∈ L1 (R). If fb is an entire function of the exponential type less than or

equal to 2πT , then show that fb(w) ≤ Ce2πBR sin θ , for B > T on the closed upper

half plane.
1 x2

5.7.2. Let Ht (x) = e 4t . If f ∈ S(R), then show that u(x, t) = (f ∗ Ht )(x)
(4πt)1/2
for t > 0, then show that u(x, t) = (f ∗ Ht )(x), for t > 0, solves the heat equation
∂u ∂ 2u
= . Further show that ku(·, t) − f (·)k2 → 0 as t → 0.
∂t ∂x2
1 t
5.7.3. Let Pt (x) = , x ∈ R, t > 0.
π x + t2
2

(a) Show that the Fourier transform of Pt is e−2π|ξ|t .


(b) Let u = f ∗ Pt , for f ∈ S(R). Then show that ∆u = 0.
(c) Show that ku(·, t) − f (·)k2 → 0 as t → 0.
1 P ∞ t ∞
e−2πt|n| .
P
(d) Using Poisson summation formula show that =
π n=−∞ t2 + n2 n=−∞

145
146

You might also like