Вы находитесь на странице: 1из 24

Chapter 6 Fourier analysis on Rn

6.1

Fourier transform on Rn
In this section, we define the Fourier transform on Rn and study some of its

properties. For f L1 (Rn ), its Fourier transform, denoted by fb, is defined as

fb() = (2)

n
2

f (x) eix dx,

Rn .

Rn

We warn the readers that the defintion taken here is different from the one defined
in one dimension. As mentioned in Chapter 3, we take in the remaining chapter the
definition of Fourier transform as the extension of the definition given in 3.2.
Clearly, fb is a bounded function on Rn with kfbk

(2)n/2 kf k1 . The

Riemann-Lebesgue lemma states that, for f L1 (Rn ), fb is a continuous function


n
Q
1 2
1
2
e 2 xj , then
vanishing at infinity. If we take (x) = e 2 |x| , x Rn , i.e., (x) =
j=1

b = e 21 ||2 = ().
()
We now introduce some multi-index notation. For x Rn , Nn , we
 j
n 
Q

. Given a
define x = x1 1 xnn ; || = 1 + 2 + + n and =
xj
j=1
P
polynomial p(x) =
a x , the corresponding differential operator is defined by
||m
P
p(D) =
a . An easy computation shows that
||m

p(D)()
b = (2)

n/2

Z
(p(D)e

ix

) (x) dx = (2)

n/2

Rn

Z
Rn

p(D)()
b = (p(ix)(.))b()
147

eix p(ix) (x) dx,

Similarly, using the formula (i)|| x (eix ) = eix , and integrating by parts, we
get
(p(D))b() = p(i)().
b
In general, when f L1 (Rn ), fb need not be integrable. For instance, f (x) =
(1,1) (x), is an integrable function on R, whose Fourier transform is not so. We are
interested in obtaining an inversion formula for the Fourier transform.
For this purpose, we need to study subspaces W of L1 (Rn ), for which fb
L1 (Rn ) whenever f W . Such a subspace of L1 (Rn ) is provided by the subspace of
smooth functions on Rn which is invariant under differentiation and multiplication
by polynomials. We denote by S(Rn ), the space of all functions f C (Rn ) such
that x f L (Rn ) for every , Nn .
Clearly, S(Rn ) L1 (Rn ), Cc (Rn ) S(Rn ), and so S(Rn ) is dense in L1 (Rn ).
By definition, S(Rn ) is invariant under the Fourier transform, i.e.,

f S(Rn )

implies fb S(Rn ). Indeed, for f S(Rn ), fb() = ((ix) f ()) b () =


(i)|| (x (ix) f (.))b(). Also note that if f Cc (Rn ) then fb S(Rn ). For this
space S(Rn ) we obtain the inversion formula:

Figure 6.1: Two dimension Gaussian.

148

Theorem 6.1.1. For every f S(Rn ), we have


Z

n
2

f (x) = (2)

fb() eix d, x Rn .

Rn

Proof. It is enough to prove the formula for x = 0. For, we can apply it to the
function x f ( where (x f )(y) := f (y x)) to obtain the required formula. We now
prove that
f (0) = (2)

n/2

Z
fb()d.
Rn

1
2
For this, we make use of the fact that b = when (x) = e 2 |x| . For t > 0, we define
1 2
2
b
= e 2 t || . Also, by
t (x) = tn ( xt ). An easy calculation gives that bt () = (t)

Fubinis theorem, we have


Z

Z
f (x)b
g (x) dx =

Rn

fb() g() d,

for f, g S(Rn ).

Rn

Taking gb = t , we get
Z
Z
x
t2
2
n
t
f (x)
dx = fb() e 2 || d
t
or

Rn

Rn

Z
f (tx)(x) dx =

Rn

t
2
fb()e 2 || d.

Rn

Here, f S(Rn ), L (Rn ) and f (tx) f (0) as t 0. By the dominated


convergence theorem, taking limit as t 0, we get
Z
f (0)

Z
(x) dx =

Rn

fb() d.
Rn

149

As (2)n/2

e 2 |x| dx = 1, the above gives f (0) = (2)n/2

Rn

fb() d, proving the

Rn

result.
From the inversion formula for S(Rn ), it is clear that f S(Rn ) if and only
b
if fb S(Rn ) ( since fb() = f ()). Denoting F f = fb, we observe that F 4 =Id on
S(Rn ), where Id denotes the identity operator. Using the density of S(Rn ) in L1 (Rn ),
we prove the following inversion formula:
Theorem 6.1.2. (Inversion formula) Assume that both f, fb L1 (Rn ). Then
n/2

fb() eix d,

f (x) = (2)

for a.e. x Rn .

Rn

Proof. We start with


Z

Z
fb() g() d =

Rn

f (x)b
g (x) dx,
Rn

for any g S(Rn ). By the inversion formula for S(Rn ), we have

g() = (2)

n/2

gb(x)eix dx, Rn .

Rn

Using this in the earlier equation, we see that



Z
Z
Z 
n/2
ix
b
(2)
f ()e
d gb(x)dx = f (x) gb(x) dx.
Rn

Rn
n/2

Taking h(x) = (2)

Rn
ix

fb() e
d, gives
Z
(f (x) h(x))b
g (x) dx = 0, for all g S(Rn ).

Rn

Rn

150

As Cc (Rn )b S(Rn ), by the inversion formula for S(Rn ),


Z

(f (x) h(x))(x)dx = 0, for all Cc (Rn ).

Rn

This shows that

(f (x) h(x)) (x) dx = 0 for any Cc (Rn ) with support

contained in a compact set K. Note that f h L1 (K) (since f L1 (Rn ) and


h L (Rn )). Since Cc is dense in L1 (K), we get (f h)(x) = 0 for a.e x K.
As K is an arbitrary compact set, f (x) = h(x) for a.e. x Rn , which means
R
f (x) = (2)n/2 fb() eix d, for a.e. x Rn .
Rn

For functions f, g L1 (Rn ), their convolution is defined by


Z
(f g)(x) =

f (x y) g(y) dy, x Rn .

Rn

Fubinis theorem gives that f g is also integrable on Rn with kf gk1 kf k1 kgk1 . It


n
is easy to check that (f g)b() = (2) 2 fb() gb(). If we define f (x) := f (x) then

we can check that (f )b() = fb(). For f, g S(Rn ), applying the inversion formula
(at 0) to the function f g, we get

f g(0) = (2)

n/2

Z
(f g)b() d =

Rn

fb() gb() d.
Rn

Taking g = f gives

f f (0) =

|f (x)| dx i.e.
Rn

Rn

151

|fb()|2 d =

Z
Rn

|fb()|2 dx.

Theorem 6.1.3. (Plancherel) The Fourier transform, defined on S(Rn ) extends to


L2 (Rn ) as a unitary operator. For every f, g L2 (Rn ), we have
Z

fb() gb() d.

f (x) g(x) dx =
Rn

Rn

Proof. We have already proved that

|f (x)|2 dx =

Rn

|fb()|2 d, for f S(Rn ).

Rn

Given f L2 (Rn ) choose a sequence {fk } S(Rn ) such that fk f in L2 (Rn ). As


kfk fj k22 = kfbk fbj k22 , we see that {fbk } is Cauchy in L2 (R).
So we define fb to be the L2 -limit of {fbk }. It is easy to see that fb is well defined.
As kfk k22 = kfbk k22 , the continuity of norm implies that kf k22 = kfbk22 . By polarisation,
we obtain
Z

Z
f (x) g(x) dx =

Rn

fb() gb() d.
Rn

Writing F f = fb, f L2 (Rn ) the above argument gives that F : L2 (Rn ) L2 (Rn )
is an isometry. To show that F is unitary, we are left with checking that F is onto.
For this, given g L2 (Rn ), we choose a sequence {gk } S(Rn ) such that gbk g in
L2 (Rn ). Then {gk } will converge to some f L2 (Rn ) for which fb = g.
Remark 6.1.4. As F 4 = Id on S(Rn ), we see that F 4 = Id on L2 (Rn ) as well.
Remark 6.1.5. The definition of Fourier transform can be extended to Lp (Rn ), 1
p 2. Given f Lp (Rn ) we can write f = g + h, where g = f |f |>1 and h = f |f |1 .
Note that g L1 (Rn ), h L2 (Rn ), and so we can define fb := gb+b
h L (Rn )+L2 (Rn ).
To see that fb is well-defined, suppose f = g1 +h1 where g1 L1 (Rn ) and h1 L2 (Rn ).
Then g g1 = h1 h so that (g g1 )b = (h1 h)b or gb + b
h = gb1 + b
h1 . Thus fb is
well-defined.
We shall make use of the following interpolation theorem.
152

Theorem 6.1.6. (Riesz-Thorin Interpolation). Let 1 p0 , p1 , q0 , q1 , and


for 0 < t < 1 define p and q by
1t
t
1
=
+
,
p
p0
p1

1
1t
t
=
+
.
q
q0
q1

If T is a linear operator from Lp0 + Lp1 to Lq0 + Lq1 such that kT f kq0 M0 kf kp0 ,
for f Lp0 , and kT f kq1 M1 kf kp1 , for f Lp1 , then kT f kq M01t M1t kf kp , for
f Lp .
Theorem 6.1.7. (Hausdorff-Young Inequality). Let 1 p 2. If f Lp (Rn ),
0
then fb Lp (Rn ), and kfbkp0 ckf kp . (Here,

1
p

1
p0

= 1.)

n
Proof. We already have kfbk (2) 2 kf k1 and kfbk2 = kf k2 . Applying the previous

theorem gives the required inequality.


We have already observed that S(Rn ) is invariant under the Fourier transform.
This property utterly fails for Cc (Rn ), as the following result states.
Theorem 6.1.8. If f Cc (Rn ), then fb
/ Cc (Rn ) unless f = 0.
This will follow from the properties of holomorphic functions once we have the
following result.
Theorem 6.1.9. (Paley-Wiener). If f Cc (Rn ), then fb can be extended to Cn
as an entire function which satisfies |fb()| CN (1 + ||)N eR|Im | , for every N N
and Cn . Conversely, if F is an entire function satisfying the above estimate, then
there is a function f Cc (Rn ) such that F = fb.
Proof. Let f Cc (Rn ) be supported in |x| R. Let
F () = (2)

n
2

eix f (x)dx,

Rn

153

= + i Cn .

Since f is compactly supported, F () is well-defined for any Cn , and F () = fb().


R ix x
n
e
e f (x) dx, and hence F is a continuous function.
Also, F () = (2) 2
|x|R

By Moreras theorem, it follows that F is actually holomorphic. Moreover,


n
2

|F ()| (2)

|f (x)| ex dx

|x|R

C exp R|| (Since |x | |x||| R|| as |x| R.)

Hence |F ()| CeR|Im | . For any Nn , we have

||

n
2

f (x)x (eix ) dx.

F () = (i) (2)

|x|R

Integration by parts gives

n
2

||

F () = (i) (2)

(x f )(x) eix dx.

|x|R

(Note that the boundary terms vanish.) This gives | F ()| C eR|Im| . As this is
true for any , we get
|F ()| CN (1 + ||)N eR|Im

for any N .

Conversely, let F be an entire function satisfying


|F ()| CN (1 + ||)N eR|Im| for all N.
In particular, |F ()| CN (1 + ||)N for any N , and hence F L1 (Rn ). So we can
define
n
2

f (x) = (2)

Rn

154

eix F () d.

If we can show that f is compactly supported in |x| R, then by Fourier Inversion,


we would get fb = F , proving the theorem. Since the function f C (Rn ) and
F () L1 (Rn ) for any , we have
x f (x)

n
2

= (2)

eix (i) F () d.

Rn

It remains to show that f is compactly supported. We prove this using Cauchys


theorem.
Look at the contour integral

R
1

eix F () d1 dn , where j is the

contour bounded by the real axis, the lines |Re z| = r and the line Im z = j .
By Cauchys theorem, this integral is zero, and so we have
Zr

Zr

ix

Zr

F ()d =

Zr

r
Z1

r
Zn

Z1

Zn

+
0

n
P

xj (j +j )

j=1

n
P

xj (r+ij )

n
P

F (r + i1 , r + in ) (i)n d1 dn

xj (r+ij )

j=1

F (1 + i1 , , n + in ) d1 dn

F (r + i1 , r + in ) d1 dn .

The second integral on the right is bounded by


Z1

Zn

e|x||| (1 + nr2 + ||2 )N eR|| d1 dn

C(1 + r2 )N

Z1

155

Zn
0

eR|| e|x||| d1 dn .

This goes to zero as r and so does the third integral on the right. Thus we are
left with
Z

ix

eix(+i) F ( + i) d.

F () d =
Rn

Rn

We claim that above integral is zero whenever |x| > R. For this, let x = r, S n1 ,
and consider
n
2

f (x) = (2)

eir er F ( + i) d.

Rn

Choose = t, where t > 0. Then


Z
|f (x) CN ert (1 + ||)N eRt d CN0 e(rR)t .
Rn

If r > R, we can take t to conclude that f (x) = 0 for |x| > R. Hence
f Cc (Rn ) with support in |x| R. Note that when f (6 0) Cc (Rn ), fb cannot
vanish on any open set in Cn . We know that fb S(Rn ) whenever f Cc (Rn ).
Hence
|fb()| CN (1 + ||)N for any N .
The above is the best possible decay we can get for fb when f Cc (Rn ). For
instance, if f Cc (Rn ) and |fb()| Cea|| for some a > 0, then f = 0. To see this,
R
n R
observe that f (x) = (2) 2 fb()eix d. When |fb()| Cea|| , fb() ei(x+iy) d
Rn

Rn

converges and defines a holomorphic function on a = {x + iy : |y| < a}. Since


f Cc (Rn ), this is possible only if f 0.
If fb S(Rn ) and |f (x)| Cea|x| , then fb has a holomorphic extension to
the tube domain a = {x + iy : |y| < a}. Let us look at a function f on Rn having
2

Gaussian decay, i.e., |f (x)| Cea|x| , for some a > 0.


156

Then fb extends to Cn as an entire function. Let us look at


fb() = (2)

n
2

f (x) eix. dx,

= + i Cn

Rn

= (2)

n
2

f (x) ex. eix. dx

Rn

= fb (), where f (x) = f (x)ex..


Then f L2 (Rn ), and so by Plancherel theorem,
Z

|fb ()|2 d =

Rn

Z
Rn

|fb( + i)|2 d =

i.e.,

|f (x)|2 dx

Rn

|f (x)|2 e2x dx

Rn

Integrating the above against () d for a sufficiently nice function gives


Z Z

Z Z

|fb( + i)| () d d =
Rn Rn

|f (x)|2 e2x () dx d.

Rn Rn
2

Taking () = e|| for some > 0, we get


Z Z

|f (x)|2 e2x ()dx d < .

Rn Rn

But we do not know yet if


Z Z

|fb( + i)|2 () d d < .

Rn Rn

157

n
Recall that fb( + i) = (2) 2

eix f (x) ex. dx, which gives

|f (x)|2 e2x dx.

|fb( + i)| d =
Rn

Rn

Since |f (x)| < C e

a|x|2

, we get
Z
Z
||2
2
2
e2x e2a|x| dx C e a .
|fb( + i)| d C
Rn

Rn

Hence
Z Z

1
|fb()|2 () d d < provided > .
a

Rn Rn

We have
Z Z

2
|fb( + i)|2 e|| d d =

Rn Rn

Z Z

|f (x)|2 e2x e|| dx d.

Rn Rn

We know that
(2)

n
2

n Z
Y

e2xj j e|| dj = e

|x|2

j=1

Thus we have
Z Z

2
|fb( + i)|2 e|| d d =

|f (x)|2 e

|x|2

dx, > 1/a.

2
The above argument gives that if |fb()| Cea|| , then f extends to Cn as an entire

function, and
Z Z

|f (x + iy)| e

|y|2

Z
dx dy =

Rn Rn

|fb()|2 e

||2

1
d , > .
a

Rn

Given a function g L1 (Rn ) can we find f (related to g) satisfying the inequality


2

|fb()| Cea|| ? The inequality |b


g ()| Cea|| does not hold for all g L1 (Rn ).
For example, if we take g(x) = (1,1) (x), then gb does not have Gaussian decay. We
158

know that (f g)b() = (2)n/2 fb() gb(). If we start with a function f L1 (R) and
n

consider f pa , where pa is given by pa (x) = (4a) 2 e

|x|2
4a

, then |(f pa ) b ()|

C ea|| . Actually, we have


n
n
n
2
2
(f pa )b() = (2) 2 (4a) 2 fb() ea|| = (2) 2 fb() ea|| .

Suppose we start with f L2 (Rn ) (instead of f L1 (Rn )), then we can extend f pa
as an entire function. Note that the function (f pa )b L2 L1 (L2 by H
olders
inequality). By the inversion formula, we have
n
2

f pa (x) = (4a)

2
eix. fb() ea|| d.

Rn

Can we holomorphically extend f pa to Cn ? Even if we start with f L2 (Rn ) the


same is true for the following reason. The integral
Z
f (u) pa (z u) du = (4a)
Rn

(where (z u)2 =

n
2

f (u)e 4a (zu) du

Rn
n
P

(zj uj )2 , z Cn ), makes sense and defines a holomorphic

j=1

function, giving the required extension of f pa to Cn . We also have


f pa (x) = (2)

n/2

2
fb() ea|| eix. d

Rn

so that the extension is also given by


Z
2
n/2
f pa (z) = (2)
fb() ea|| ei(x+iy) d, z = x + iy.
Rn

159

By Plancherel theorem,
Z

2
Z Z


2
a||
y
ix
e
e
d dx
|f pa (x + iy)| dx = fb() e
2

Rn Rn

Rn

Z
=

|fb()|2 e2a|| e2y d.

Rn

We can integrate both sides against p a2 (y) = (2a) 2 e


Z Z

|y|2
2a

, to get

|f pa (x + iy)|2 p a2 (y) dx dy

Rn Rn
n/2

Z Z

= (2a)

|fb()|2 e2a|| e2y. pa/2 ()ddy.

Rn Rn

Since (2a) 2

e2y. e

|y|2
2a

dy = e2a|| , we get

Rn

Z Z

|f pa (x + iy)| p (y) dx dy =
a
2

Rn Rn

2
2
|fb()|2 e2a|| e2a|| d =

Rn

|f (x)|2 dx.

Rn

Thus we have proved.


Theorem 6.1.10. Let f L2 (Rn ). Then f pa extends to Cn as a holomorphic
function and satisfies
Z Z

|f pa (x + iy)| pa/2 (y) dx dy =

|f (x)|2 dx.

Rn

Rn Rn

The transform which takes f into the entire function f pa (z) is called the
Segal-Bargmann transform. We define Ba (Cn ) to be the space of all entire functions

160

F satisfying
kFa k2Ba

Z Z
:=

|F (x + iy)|2 p a2 (y)dx dy <

Rn Rn

This forms a Hilbert space with the inner product given by


Z Z
hF, Gi =

F (x + iy) G(x + iy) pa/2 (y) dx dy.


Rn Rn

The theorem says that the Segal-Bargmann transform takes L2 (Rn ) into Ba (Cn ).
Actually, the Segal-Bargmann transform is onto Ba (Cn ). In order to prove this,
without loss of generality, we consider the case a = 21 . If we take F B1/2 (C n ) then
1 2

the function G(z) = F (z)e 4 z , where z 2 = z12 + + zn2 , satisfies


Z

21 |z|2

|G(z)| e

Z Z
dz =

Cn

|F (z)|2 e 2 |x| e 2 |y| e 2 |x| e 2 |y| dx dy

Rn Rn

Z Z
=

|F (z)|2 e|y| dx dy < .

Rn Rn
1 2

Thus we see that F B1/2 (Cn ) if and only if F (z)e 4 z belongs to the space F(Cn ),
R
1
2
consisting of all entire functions G which satisfy
|G(z)|2 e 2 |z| dz < . Also,
Cn

the Segal-Bargmann transform f 7 f p 1 gives rise to the transform Bf (z) =


2

1 2
z
4

f p 1 (z), which takes L (R ) into F(Cn ). In order to show that the Segal2

Bargmann transform is onto B 1 (Cn ), it is enough to show that the transform B :


2

L (R ) F(C ) is onto.
Theorem 6.1.11. B is a unitary operator from L2 (Rn ) onto F(Cn ).

161

We already know that B is an isometry. So it is enough to prove that B is

1
onto. In F(Cn ), consider the functions (z) = (2|| ! ) 2 z , Nn . Then
Z

(z) (z) e1/2|z| dz

h , iF(Cn ) =
Cn

1
= (2 ! ) 2 (2|| ! )1/2
||

z z e 2 |z| dz.

Cn

Integrating in polar coordinates, we can check that h , i = . In fact, { :


Nn } is an orthonormal system in F(Cn ). If F F(Cn ) then we have its Taylor
P
expansion F (z) =
c z , which converges uniformly over compact subsets of Cn .
N nP
It can be shown that
c z actually converges in F(Cn ). Hence { : Nn } is
Nn

an orthonormal basis for F(Cn ). The surjectivity of B will be proved if we can find
n
Q
functions such that B = . We claim that if we take (x) =
hj (xj ),
j=1

where hj (xj ) are Hermite functions, then B = . To prove this we can assume
that n = 1. All we need is

1
1 2
Proposition 6.1.12. If hk (x) = (2k k! ) 2 Hk (x)e 2 x , where Hk is the k th Hermite polynomial, then Bhk (z) = k (z).

1
1
2
Proof. We need to show that e 4 |z| (hk p 1 )(z) = (2k k! ) 2 z k , or equivalently,
2

Hk (z u) e 2 (zu) e 2 u du = z k e1/4z .

The above can be simplified to


1

1 2

Hk (z u) eu ezu du = z k e 4 z .

162

Since both sides are holomorphic, it is enough to prove


1

1 2

Hk (x u) eu exu du = xk e 4 x , x R

It is enough to prove that


Z

1 2

Hk (u) eu eixu du = (i)k xk e 4 x , x R.

From the definition, we have (by integration by parts)


1

Z 

d
du

k

u2

ixu

(i)k k
x
du =
2

1 2

eu eixu du = (i)k xk e 4 x

Thus we have proved that B = , where (x) =

n
Q

hj (xj ). Since { :

j=1

Nn } is an orthonormal basis for L2 (Rn ), we see that for F F(Cn ),


B 1 F (x) =

hF, iF (x)

Nn

defines a function f in L2 (Rn ) and Bf = F . This proves the surjectivity of B.

6.2

Fourier expansion of Fourier transform


In this section, we consider the Fourier transform of functions on R2 , which

we identify with C. For an integrable function f on C, we can write the Fourier


transform in the form
1
fb(w) =
2

f (z) eiRe(zw)
dz, w C.

163

In polar coordinates, we write w = ei , > 0, R, so that


1
fb(ei ) =
2

f (z) eiRe(ze

i )

dz.

For any > 0, 7 fb(ei ) is a 2- periodic function of . Hence we can expand


fb(ei ) into a Fourier series
fb(ei ) =

Fm () eim ,

m=

where
1
Fm () =
2

Z2

fb(ei ) eim d.

Using the definition of fb(w), we have


 2 Z2 Z

1
iRe(zei )
Fm () =
f (z) e
dz eim d.
2
0

Writing z = rei , we have


 2 Z Z2 Z2
1
i() )
Fm () =
f (rei ) eirRe(e
eim d d r dr.
2
0

Look at
1
2

Z2

irRe(ei() )

eim

1
d = eim
2

Z2
0

164

eir cos eim d = eim Km (r)

Thus we have

Z
Fm () =

fm (r) Km (r) rdr,


0

where

Z2

1
fm (r) =
2

f (rei )eim d.

We have
1
2

Z2

fb(ei ) eim d =

Z 

Z2

1
2

im

f (re ) e


d

Km (r) rdr.

By the above calculations, we have proved


Theorem 6.2.1. If f satisfies f (zei ) = eim f (z), then fb also satisfies the same


R
i
im
b
equation, and we have f (e ) = e
f (r) Km (r) rdr . In other words, the
0

set of all functions f satisfying f (zei ) = eim f (z) is invariant under the Fourier
transform.
If we denote by ,m (z) the function
1
,m (z) =
2

Z2

eiRe(ze

i )

eim d,

then we have
1
Fm () =
2

Z2

fb(ei ) eim d =

Z
f (z) ,m (z) dz.
Cn

Note that ,m (zei ) = eim ,m (z), so that


,m (rei ) = eim ,m (r) = eim Km (r).
165

Consider the function


1
,m (z) =
2

Z2

eiRe(ze

i )

eim d.

Clearly,
1
,m (r) =
2

Z2

eir cos eim d

has a zero of order m at the origin. So we can define Jm (r) = (r)m ,m (r). When
Z2
Z1
1
1
1
m = 0, J0 (r) =
eir cos d =
eirt (1 t2 ) 2 dt. We claim that
2
2
1

(r/2)m

Jm (r) = m
( 2 + 1)

Z1

eirt (1 t2 )m 2 dt.

In order to prove this, let us look at


d
dr

Jm (r)
rm

1
Z

1
d
irt
2 m 21
e
(1

t
)
dt
= m

2
(m + 12 ) dr
1
1

1
d
2 m 12
= m
(cos
rt)
(1

t
)
dt

2
(m + 12 ) dr
1

2m (m + 12 )

Z1

(sin rt) t (1 t2 )m 2 dt.

Integrating by parts, we get




Z1
1
1
1
d Jm (r)
=
(cos rt)(1 t2 )m+ 2 dt
1
m
m+1
dr
r
r 2
(m + 1 + 2 )
1

Jm+1 (r)
rm
166

We now consider




Z2
d 1,m
1 d
1
ir cos im
=
e
e
d
dr rm
2 dr rm
0

m 1
2 rm+1

Z2

eir cos eim d

Z2

1
+
(i)
2rm

eir cos (cos ) eim d.

1,m+1 (r)
d  1,m 
=
, then we can conclude that 1,m (r) =
m
dr r
rm
c Jm (r). The function Jm (r) is called the Bessel function of order m. We have
If we can show that

1
2

Z2

fb(ei ) eim d =

Z
0

1
2

Z2

f (rei ) eim d Jm (r) rdr.

We can write the above equation as

m Fm () =

rm fm (r)

Jm (r) 2m+1
r
dr
(r)m

or

Fm () = Hm fm (),

where

Z
Hm g(x) =

g(r)

Jm (r) 2m+1
r
dr.
(r)m

Hm is called the Hankel transform of order m. By writing down the Fourier series of
f as
i

f (re ) =

rm fm (r) eim ,

m=

167

we see that
fb(ei ) =

m Hm fm () eim

m=

Hence we can reduce the study of Fourier transform to the study of Hankel transforms.
We can generalise the definition of Bessel functions Jm to define
(r/2)

J (r) =
( + 12 )

Z1

eirt (1 t2 ) 2 dt

for any > 21 . Then we can also define


Z
H g() =

g(r)

J (r) 2+1
r
dr,
(r)

g L2 (R+ , r2+1 dr).

6.3

Exercises

6.3.1. We can define the space of tempered distribution S 0 (Rn ) to be the dual space
of S 0 (Rn ) as in the case of R. Show that Lp (Rn ), 1 p is a subspace of S 0 (Rn ).
6.3.2. Let s R. Let H s (Rn ) denote the collection of all f S 0 (Rn ) such that
s

(1 + ||2 ) 2 fb() L2 (Rn ). Then H s (Rn ) is called the Sobolev space of order s. Show
that H s (Rn ) is a Hilbert space with the inner product given by
Z
hf, gi =

1 + ||2

s

fb() gb() d.

Rn

6.3.3. Show that for m N0 , H m (Rn ) can be written as the collection of all f
L2 (Rn ) for which D f L2 (R), Nn0 with || m. Here, if = (1 , 2 , . . . , n ),

168

then || = 1 + 2 + + n and D f =
x1
derivative in the sense of tempered distribution.

6.3.4. Show that S(Rn ) is dense in H s (Rn ).

169

1 

x2

2

xn

 n
denotes

170

Вам также может понравиться