Beruflich Dokumente
Kultur Dokumente
1. L2 (Rn ) =
Vm .
m=0
First we look the action of the Fourier transform on tha class of radial functions on
Rn .
7.1
all k SO(n), the special orthogonal group consisting of orthogonal matrices with
determinant 1.
Lemma 7.1.2. If f L1 (Rn ) is radial, so is its Fourier transform.
Proof. For Rn ,
Z
n
f() = (2) 2
f (x) eix dx
Rn
f (x) eix.k dx
Rn
Making a change of variables x 7 kx and noting that kxk = x and f (kx) = f (x),
we get f(k) = f().
Let V0 = {f L2 (Rn ) : f (kx) = f (x) for all k SO(n)}, i.e., V0 is the
radial subspace of L2 (Rn ). The above lemma states that Fn V0 = V0 . Let W0 be
the orthogonal complement of V0 in L2 (Rn ). Given f L2 (Rn ), define f0 (x) =
R
f (kx) dk, which belongs to V0 . Let us look at hg, f0 i where g = f f0 .
S0(n)
Z
hg, f0 i =
f (x)f0 (x)
Rn
|f0 (x)|2 dx
Rn
But
Z
Z
f (x)f0 (x) =
Rn
SO(n)
Z
f (kx)f 0 (x) dx dk.
Rn
Z
Z
Z
Z
f0 (x)
f (kx) dk dx = |f0 (x)|2 .
Rn
SO(n)
Rn
Rn
172
fe (x) =
f (x) + f (x)
,
2
and fo (x) =
f (x) f (x)
2
are respectively the even and odd parts of the function f. As can be easily checked
these two subspaces are inviariant under the Fourier transform. Moreover, for even
functions
1
f() = (2) 2
Z
f (x) cos(x) dx
R
is the cosine transform. Similarly, for odd functions the Fourier transform reduces to
the sine transform. Now we proceed to obtain a decomposition of L2 (R2 ). For this,
we identify a point (x, y) R2 with the complex number z = x + iy = rei . Let
V0 = {f L2 (C) : f (ei z) = f (z), for all z C, ei S 1 }
be the space of all radial functions on R2 . Let
Vm = {f L2 (C) : f (ei z) = eim f (z), for all ei S 1 }.
Then for each m, the space Vm is invariant under the Fourier transform, i.e., if f Vm ,
then fb Vm . For an integrable function f on C, we can write its Fourier transform
173
in the form
1
f(w) =
2
f (z) eiRe(zw)
dz, w C.
f (z) eiRe(ze
i )
dz.
For any > 0, the map 7 f(ei ) is a 2- periodic function of . Hence we can
expand f(ei ) into a Fourier series
fb(ei ) =
Fm () eim ,
m=
where
1
Fm () =
2
Z2
f(ei ) eim d.
Z2 Z
f (z) e
0
iRe(zei )
dz
eim d.
Writing z = re , we have
Z Z2 Z2
1
i() )
Fm () =
f (rei ) eirRe(e
eim d d r dr.
2
(2)
0
Look at
Z2
Z2
1
1
i()
) im
eirRe(e
e
d = eim
eir cos eim d
2
2
0
= eim Km (r)
174
Thus we have
Fm () =
1
fm (r) =
2
where
Z2
f (rei )eim d.
We have
1
2
Z2
f(ei ) eim d =
Z
1
2
Z2
f (re ) e
im
d
Km (r) rdr.
set of all functions f satisfying f (zei ) = eim f (z) is invariant under the Fourier
transform.
1
By defining ,m (z) =
2
Z2
eiRe(ze
i )
1
Fm () =
2
Z2
fb(ei ) eim d =
Z
f (z) ,m (z) dz.
(7.1)
Cn
Note that ,m (zei ) = eim ,m (z), so that ,m (rei ) = eim ,m (r) = eim Km (r).
Consider the function
1
,m (z) =
2
Z2
i
eiRe(ze ) eim d.
0
175
Clearly,
1
,m (r) =
2
Z2
1
d = (i)m
2
Z2
Jm (r) =
2m
rm
(m + 1/2) (1/2)
Z1
eirs (1 s2 )m 2 ds.
Proof. Let
rm
Jm
(r) = m
2 (m + 1/2) (1/2)
Z1
eirs (1 s2 )m 2 ds.
By definition,
1
J0 (r) =
2
Z2
eir cos s
1
ds =
2
Z2
eirs (1 s2 ) 2 ds = J0 (r).
Let us compute
d
dr
Jm (r)
rm
= rm
m
r
0
(r)
Jm (r) Jm
176
= rm
m
2r
Z2
i
2r
rm
=
2
=
rm
2
Z2
Z2
d eir sin eim
ir sin im
ir sin im
i (
) + cos e
e
i sin e
e
d
d
r
0
Z2
= r
Jm+1 (r).
1
d
= m
1
2
(m + 2 ) dr
=
2m
(m + 21 )
Z1
Z1
2 m 12
(cos rt) (1 t )
dt
1
1
Jm
(r)
m
r
1
1
=
m+1
r 2
(m + 1 + 21 )
=
Z1
J (r)
m+1
rm
Jm = Jm
for all non-negative integers m.
177
makes sense even for non-integral values of m. Thus
The integral defining Jm
r
J (r) :=
2 (r + 1/2) (1/2)
eirs (1 s2 ) 2 ds, r 0.
Z2
f(ei ) eim d =
Z
0
1
2
Z2
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
f (rei ) =
rm fm (r) eim ,
m=
we see that
f(ei ) =
m Hm fm () eim
m=
178
Hence we can reduce the study of Fourier transform on radial funcions on R2 to the
study of Hankel transforms.
For any > 1/2, we can define Hankel tranforms H by the equation
Z
H g() =
g(r)
J (r) 2+1
r
dr,
(r)
n
2
f (x) eix dx
F () = (2)
Rn
Z
i.e., F () =
f0 (r)
(2)
n
2
ix
d
rn1 dr,
S n1
n1
eirx d(x0 ).
S n1
179
Z1
=C
Cn J n2 1 (r)
n
(r) 2 1
f0 (r)
J n2 1 (r)
n
(r) 2 1
rn1 dr
= H n2 1 f0 ()
Thus the Fourier transform of a radial function reduces to a Hankel transform.
7.2
Spherical harmonics
Now that we have studied the action of the Fourier transform on radial func-
tions in all dimensions, we would like to obtain the decomposition stated in the
L
beginning. In the case of L2 (C), we have the decomposition L2 (C) =
Vem , where
m=0
L
Vem = Vm Vm . Note that the elements of Vem are functions of the form
f (z) = eim g(r) + eim h(r),
where g and h are functions on R+ = [0, ). We can think of eim and eim as
restrictions of z m and zm to S 1 . Note that z m and zm are harmonic and homogeneous
of degree m. Taking this as a clue, we want to look at the higher dimensional analogues
of z m and zm . Before proceeding, we recall some multi-index notation. For a point
180
> 0.
Let Pm stand for the set of all polynomials which are homogeneous of degree
P
m. By definition, an element of Pm has the form P (x) =
c x . The set {x :
||=m
array of (n+m1) boxes. Label each of the remaining unmarked boxes as W. Writing
the labels of the boxes, we get a sequence of the letters B1 , B2 , , Bn1 and W.
Let 1 be the number of W 0 s which precede B1 . Let 2 be the number of W 0 s
which are between B2 and B3 , and so on. Let n be the number of W 0 s following
Bn1 . Thus, we obtain an n-tuple = (1 , 2 , , n ) with || = m. The number of
such n-tuples is the number of ways we can select (n 1) boxes among the (n + m 1)
boxes, which is given by
dm =
n+m1
n1
n+m1
= (n + m 1)! .
(n 1)! m!
||=m
181
||
. For P, Q Pm , define hP, Qi = P (D)Q. Then h, i satisfies the
x11 x22 xnn
following properties.
=
(i) As P and Q are homogeneous of the same degree, hP, Qi C, for all polynomials
P, Q Pm . For n-tuples = (1 , 2 , , n ), (1 , 2 , , n ) Nn , we have
n
1 2
x11 x22
xnn
n
P
x1 1 x2 2 xnn = 1 !2 ! n !.
j=1
solid harmonics of degree m. Note that Pm and Hm are invariant under the action
SO(n), i.e., if p Pm (resp. p Hm ), then for every S0(n), the polynomial
em = {p|S n1 : p Hm }. An
p (x) := p(.x) belongs to Pm (resp. to Hm ). Let H
em is called a (surface) spherical harmonic of degree m. Every polynomial
element of H
em , there exists
p Hm is uniquely determined byits restriction to S n1 . For Y H
x
x
a polynomial P Hm such that Y = P |S n1 . Then, Y ( |x|
) = P ( |x|
) = |x|m P (x). If
e m onto H
em .
Y = 0, then P = 0, and so the map P 7 P |Sn1 is an isomorphism of H
Thus,
em = dim Hm = dim Pm dim Pm2 = dm dm2 .
dim H
182
em H
ek if m 6= k.
Lemma 7.2.2. H
em , then p(x) = rm p(),
Proof. For, if p H
x = r, r > 0, S n1 . Since
p = 0, we get
n1
1
+
+ 2 S
2
r
r r
r
(rm p()) = 0,
1
n1
m rm1 p() + 2 rm S p() = 0
r
r
183
Proof. As any polynomial of degree less than 2 is harmonic, we assume that k 2. Let
n
P
2
=
be the Laplacian on Rn . Then the map , defined by (P ) = P maps
2
j=1
xj
Pk onto Pk2 . For if is not onto, there exists a polynomial Q Pk2 such that Q is
orthogonal to P for all P Pk , i.e., hP, Qi = hQ, P i = 0. In particular, taking
P (x) = |x|2 Q(x), we get 0 = hQ, P i = Q(D)P = P (D)P = hP, P i, a contradiction, since P 6= 0. Let Bm = {P (x) Pm : P (x) = |x|2 Q(x), for some Q Pm2 }.
Claim: Pm is the orthogonal direct sum of Hm and Bm .
For, if R(x) = |x|2 Q(x) for Q Pm2 , then hR, P i = 0 for all Q Pm2 if
and only if Q(D)P = 0 for all Q Pm2 . This happens if and only if hQ, P i = 0
for all Q Pm2 , which is true if and only if P = 0. In particular, if m = k and
P Pk , then P (x) can be written as P (x) = P0 (x) + |x|2 Q(x), where P0 is harmonic
and Q Pk2 .
Applying the above claim for m = k 2, there exist a harmonic polynomial P1
and Q Pk4 such that Q(x) = P1 (x) + |x|2 Q1 (x) = P0 (x) + |x|2 P1 (x) + |x|4 Q1 (x).
Thus the result follows by induction.
The following are some of the interesting consequences of the above theorem:
Corollary 7.2.4. The restriction to the surface of the unit sphere S n1 of any polynomial of nvariables is a sum of restrictions to S n1 of harmonic polynomials.
e be the collection of all finite linear combinations of the
Corollary 7.2.5. Let H
em . Then
elements of cup H
k=0
e is dense in L2 (S n1 ).
(ii) H
Proof. (i) By the Weierstrass Approximation theorem, any continuous function g on
S n1 can be uniformly approximated by polynomials restricted to S n1 . Corollary
e thus
7.2.5 gives these restrictions as finite linear combinations of elements of H,
proving (i).
(ii) For f L2 (S n1 ) and > 0, let g be a continuous functions on S n1 satisfying
e such
kf gk2 < /2. By (i), there is a finite linear combination h of elements of H
that kg hk <
.
2 n1
Therefore,
7.3
+ n1 kg hk < /2 + /2 = .
2
Zonal harmonics
For a fixed x0 S n1 , the mapping Y 7 Y (x0 ) defines a continuous linear
Y (x ) =
(k)
ek .
Y (t0 ) Zx0 (t0 ) dt0 , Y H
S n1
(k)
The function Zx0 is called the zonal harmonic of degree k with pole x0 . Some properties
of zonal harmonics are given in the next lemma.
0
ek , then Z (k)
Lemma 7.3.1. (a) If {Y1 , Y2 , , Yak } is an orthonormal basis of H
x0 (t ) =
ak
P
Ym (x0 )Ym (t0 ).
m=1
(k)
(k)
(k)
(k)
ek , Z (k)
Proof. (a) As {Y1 , Y2 , , Yak } is an orthonormal basis for H
x0 has the form
ak
P
(k)
(k)
(k)
(Zx0 , Ym ) Ym . By the definition of Zx0 , we have
Zx0 =
m=1
(k)
(Zx0 , Ym )
Z
=
(k)
S n1
(k)
Zx0 (t0 )
Y (t ) dt =
(k)
S n1
S n1
By the definition of zonal harmonic, the above integral is Y (1 (x0 )) = Y (x0 ). By the
(k)
(k)
uniqueness of the representation of linear functionals, we get Zx0 (t0 ) = Zx0 (t0 ).
(k)
1
Corollary 7.3.2. (a) Zx0 (x0 ) = ak n1
, for all x0 S n1 , where ak is the dimension
ak
P
1
|Ym (x0 )|2 = ak n1
, for all x0 S n1 . This equality is independent of the
m=1
ek .
choice of the orthonormal basis {Y1 , Y2 , , Yak } of H
1
(c) |Zt0 (x0 )| ak n1
, for all x0 , t0 S n1 .
Proof. (a) Fix x01 S n1 . For any x02 S n1 , there exists a rotation such that
x02 = x01 . By Lemma 7.3.1(c), Zx02 (x02 ) = Zx01 (x01 ). Thus Zx0 (x0 ) is independent of the
point x0 S n1 , say Zx0 (x0 ) = c. By Lemma 7.3.1(a), we have
186
(k)
Zx0 (x0 )
ak
X
Ym
(x0 )Y
m (x
ak
X
)=
m=1
m=1
ak Z
X
m=1 n1
S
Z X
ak
S n1
Z
=
m=1
(k)
S n1
Thus
(k)
Zx0 (x0 ) = ak n1 .
ak
P
1
|Ym (x0 )|2 = ak n1
, thus proving
m=1
(b).
ek , by Lemma 7.3.1(a), we have
(c) For {Y1 , Y2 , , Yak } an orthonormal basis for H
(k)
kZu0 k22
Z
=
(k)
|Zu0 (w0 )|2 dw0
(k)
m=1
S n1
So
ak
X
(k)
(k)
(k)
(k)
(k)
1
i.e., |Zt0 (x0 )| ak n2
, for all x0 , t0 S n1 .
187
m
X
k=0
P (x) =
j
P
l=0
Pl (x) =
l=0
j
P
Pl (x) =
j
P
l=0
l=0
Let F (x) = |x|l Pl (x). Then F (x) = F (x) for all > 0, i.e., F is homogeneous of
degree 0. Also, F (x) = |x|l Pl (x) = |x|l Pl (x) = F (x), i.e., F is invariant under
rotations. This together with its homogeneity forces F to be a constant function,
F (x) = cl 0 , for all x Rn . This implies that Pl (x) = cl 0 |x|l . As Pl is a polynom
P
mial, cl 0 6= 0 only when l is even. Thus P (x) =
ck |x|2k , where ck = c02k , for
k=0
(k)
for any x0 S n1 , we have Ze (x0 ) = Ze (x0 ). Since the point e is fixed under the
action of , the parallel of S n1 orthogonal to e is left invariant by . Let x01 , x02 belong
to this parallel of S n1 . Choose a rotation which fixes e and maps x01 to x02 , i.e.,
e = e and x01 = x02 . Then
Ze(k) (x02 ) = Ze(k) (x01 ) = Ze(k) (x01 ),
(k)
i.e., Ze
fixes e1 . Then P (x) = P (x) for all x Rn . Also, all polynomials of the form
xm
1 , for m = 0, 1, 2, , are invariant under the action of . We write P (x) =
k
P kj
x1 Pj (x2 , , xn ). Then we get that the rotation maps (x1 , x2 , , xn ) 7
j=0
(x1 , x02 , , x0n ), where (x2 , x3 , xn ) 7 (x02 , , x0n ) is a rotation of the (n1)dimensional
space x1 = 0. As varies over all rotations fixing x1 , these mappings give rise to all
rotations of the space x1 = 0. Since P is invariant under the rotation , so are all the
polynomials P0 , P1 , , Pk . This gives by Lemma 7.3.4 that Pj = 0 for j odd, and
189
that Pj (x2 , , xn ) = cj (x22 + + x2n ) 2 , for j even. Let R = (x22 + + x2n ) 2 . Then
we have
2
k2l 2l
P (x) = c0 xk1 + c2 xk2
R .
1 R + + c2l x1
(7.1)
k2(j+1)
[c2j aj + c2(j+1) bj ] x1
R2j = 0
j=0
where aj = (k 2j)(k 2j 1), and bj = 2(j + 1)(n + 2j 1). This gives c2(j+1) =
a
bjj c2j , j = 0, 1, , l 1, i.e., P depends only on the coefficient c0 , and so any two
polynomials of the form (7.1) differ by a constant multiple. But we have proved that
any homogeneous polynomial of degree k whose restriction to S n1 is constant on the
(k)
x
parallels of S n1 orthogonal to e1 has the form (7.1). We also know that Ze1 ( |x|
)|x|k
Y (y 0 ) = W ( 1 y 0 ) = cZe(k)
( 1 y 0 ) = cZ 1 e ( 1 y 0 ) = cZe(k) (y 0 ).
1
Corollary 7.3.6. Suppose Fy0 (x0 ) is defined for all x0 , y 0 S n1 and satisfies the
following properties:
(a) Fy0 is a spherical harmonic of degree k for each y 0 S n1 ,
(b) Fy0 (x0 ) = Fy0 (x0 ), for all rotations .
190
(k)
Then, there is a constant c such that Fy0 (x0 ) = cZy0 (x0 ), for all x0 , y 0 S n1 .
Proof. Let y 0 S n1 and be a rotation that fixes y 0 . Then, by hypothesis (b), for
any x0 S n1 , we have Fy0 (x0 ) = Fy0 (x0 ) = Fy0 (x0 ), ie., Fy0 is a spherical harmonic
which is constant on parallels to S n1 at y 0 S n1 . Applying Theorem 7.3.5, we get
(k)
a constant c(y 0 ) such that Fy0 = c(y 0 )Zy0 . Let y10 , y20 , S n1 , and be a rotation such
that y10 = y20 . Again by hypothesis (b), we get
(k)
(k)
c(y20 )Zy0 (x0 ) = Fy20 (x0 ) = Fy10 (x0 ) = Fy10 (x0 ) = c(y10 )Zy0 (x0 ),
(k)
(k)
(7.2)
(k)
(k)
(7.3)
From (7.2) and (7.3), we get that c(y10 ) = c(y20 ), proving the result.
For k = 0, 1, 2, , let hk be the subspace of L2 (Rn ) generated by the functions
of the form f (r)P (x), where f is a radial function and P is a solid harmonic of degree
k. The following result describes the decomposition we are looking for:
L
hk holds in the sense
Theorem 7.3.7. The direct sum decomposition L2 (Rn ) =
k=0
that
(d) for each k = 0, 1, 2, , the space hk is invariant under the Fourier transform.
ek ,
Proof. (a) The space of solid harmonics of degree k is isomorphic to the space H
and so is of dimension ak . If P1 , P2 , , Pak is an orthonormal basis of this space, then
ak
R
P
fj (r) Pj (x), with |f (x)|2 dx =
every element f of the space hk has the form
Rn
j=1
ak R
P
j=1 0
(b) Integrating in polar coordinates and using Lemma 7.2.2 gives (b).
(c) Suppose there is a function f L2 (Rn ) which is orthogonal to hk , for all k. By
Corollary 7.2.5, the function f vanishes almost everywhere on almost every sphere
centered at the origin, and so f = 0 in L2 (Rn ).
(d) Let f L1 L2 (Rn ) be of the form f (u) = f0 ()P (u) = k f0 ()Y (u0 ), with
ek , where = |u|, u = u0 . Suppose we prove that fb hk whenever the function
Y H
f is of the above form. Since finite linear combinations of such functions are dense in
hk , this will prove that hk is invariant under the Fourier transform.
For r = |x|, x = rx0 , we have
n
fb(x) = (2) 2
= (2) 2
Z
Rn
Z
eixu f (u) du
f0 () k+n1
S n1
For s 0,
Z
isx0 u0
e
S n1
Y (u ) du
isx0 u0
=
S n1
S n1
192
(k)
Y (v 0 ) Zu0 (v 0 ) dv 0 du0
(7.4)
Z
=
Y (v )
S n1
Let Fx0 (v 0 ) =
(k)
(7.5)
S n1
(k)
R
S n1
Fx0 (v ) =
(k)
S n1
Z
=
(k)
S n1
Z
=
(k)
S n1
(k)
By Corollary 7.3.6, there is a number c = (s) such that Fx0 (v 0 ) = cZx0 (v 0 ), for all
x0 , v 0 S n1 . Therefore (7.5) becomes
Z
e
isx0 u0
Y (u ) du =
S n1
Y (v 0 ) Fx0 (v 0 ) dv 0
S n1
Z
i.e.,
isx0 u0
Y (u ) du =
S n1
(k)
S n1
n
fb(x) = (2) 2
Z
0
193
7.4
radial. In the two dimensional case, the Fourier transform of a radial function was
expressed in terms of the Bessel functions. We proved that this was the case in higher
dimensions too. We now study the action of the Fourier transform on the invariant
subspaces hk , defined in the previous section.
2
Theorem 7.4.1. Suppose f (u) = e|u| P (u) for u Rn , where P (u) is a solid
harmonic of degree k, then fb(v) = ik f (v), for all v Rn .
Proof. Since P is harmonic, its satisfies the mean value property. This gives, for a
fixed t Rn ,
Z
e|u| P (u + t) du =
Rn
2
rn1 er
P (t + ru0 ) du0 dr
n1
rn1 er dr n1 P (t)
Z
= P (t)
n1
r2
i.e., P (t) =
du
Z
dr = P (t)
S n1
e|x| dx
Rn
e|u| P (u + t) du
Rn
Z
P (z) =
e|u| P (u + z) du, z = x + iy Cn .
Rn
194
Rn
homogeneous of degree k, P (iv) = (i)k P (v). Using this and applying Cauchys
theorem n times, we get
Z
(u+iv)(u+iv)
Z
P (u) du =
Rn
Rn
2
Multiplying both sides by e|v| , we get fb(v) = (i)k f (v) = ik f (v).
Let Dn+2m denote the space of all such functions with the obvious norm. We claim
2
195
As the above equation is true for all k N, the function g(r)rn+2m1 e 2 r is or1
thogonal to all functions of the form P (r2 )e 2 r , as P runs through all polynomials.
Since these functions form a dense class in L2 ((0, ), dr) we get that g = 0. So it
suffices to prove that W is invariant under the Fourier transform.
2
2
For t > 0, let ft (x) = tn f (tx), so that fbt () = fb(t1 ). If f (x) = p(x)et |x| ,
2
take g(x) = p(x)e|x| , and consider fb() = tnm gbt () = tnm gb(t1 ). Since gb() =
196