Chapter 2

A Compact Symmetric Space: The Sphere

It is important that you establish a steady, rhythmic tempo while doing the exercises and
that you rest as little as possible between them. Also remember not to strain yourself.
From Jane Fondas Year of Fitness and Health/1984, Desk Diary, Simon and Schuster,
NY, 1983. Reprinted by permission.

2.1 Fourier Analysis on the Sphere

2.1.1 The Sphere as a Symmetric Space
Whenever there is a large earthquake the Earth vibrates for days afterwards. The vibrations
consist of the superposition of the elasticgravitational normal modes of the Earth that are
excited by the earthquake.
From F. Gilbert [212, p. 107].

A (surface or Laplace) spherical harmonic is an eigenfunction of the Laplacian

on the sphere. These are the analogues of exponentials for Fourier analysis on
the sphere. Laplace and Legendre introduced these functions in order to study
gravitational theory in the 1780s. Spherical harmonics are necessary for the analysis
of any phenomena with spherical symmetry; e.g., earthquakes, the hydrogen atom,
and the solar corona. Some of these topics will be discussed later in this section.
Since our treatment of harmonic analysis on the sphere is rather condensed, the
reader may want to consult some of the following references for more information:
Coifman and Weiss [98], Courant and Hilbert [111], Dym and McKean [147],
Erdelyi et al. [164], Lebedev [401], Muller [481], Sugiura [648], Talman [655],
Wawrzynczyk [722], and Vilenkin [704]. For the history of the subject, see
Wangerin [716].
Before discussing spherical harmonics, we need to understand something about
the geometry of the sphere. This symmetric space is closely related to the orthogonal group O(n) of real n n matrices U such that t UU = I, where t U denotes
the transpose of U and I denotes the n n identity matrix. The special orthogonal
A. Terras, Harmonic Analysis on Symmetric SpacesEuclidean Space, the Sphere,
and the Poincare Upper Half-Plane, DOI 10.1007/978-1-4614-7972-7 2,
Springer Science+Business Media New York 2013

107

108

2 A Compact Symmetric Space: The Sphere

Fig. 2.1 Angular coordinates

on the unit sphere

group, SO(n), is the subgroup of matrices U in O(n) such that the determinant of
U is one. You can regard U in SO(n 1) as an element of SO(n) by forming


U0
0 1


in SO(n).

The group O(n) is a compact group and the sphere is a compact symmetric space,
making some of the analysis on it somewhat easier.
Exercise 2.1.1 (The Sphere as a Quotient or Homogeneous Space). Consider
the sphere Sn1 = {x Rn | x = 1}. Show that Sn1 can be identified with the
quotient space SO(n)/SO(n 1), using the preceding identification of SO(n 1) as
a subgroup of SO(n).
Hint. Map the coset gSO(n 1) to the vector gen for g in SO(n 1) and en =
t (0, . . . , 0, 1).
You may also want to read the discussion of the topology of spheres and
orthogonal groups in Chevalley [85, pp. 5267]. In particular, the fundamental (or
Poincare) group of SO(2) is isomorphic to Z, while that of SO(n), n 3, has order 2.
From now on we shall consider only the sphere S2 . See the references for the
general case. Now the sphere S2 is a differentiable manifold. This means that
locally it looks like two-dimensional Euclidean space. To make this precise, we use
the usual angular coordinates ( , ), 0 < < 2 , 0 < < , pictured in Fig. 2.1,
to parameterize S2 except for the semi-circle C through the poles and (1, 0, 0).
A similar coordinate patch can be constructed to cover the rest of the sphere. The
equations for the rectangular coordinates (x, y, z) of a point on S2 in terms of the
angular coordinates are

x = sin cos
y = sin sin

z = cos

(2.1)

2.1 Fourier Analysis on the Sphere

109

More information on differentiable manif olds can be found in Yvonne ChoquetBruhat et al. [87], Helgason [276278], Loomis and Sternberg [428], Singer and
Thorpe [604], or Spivak [614], for example.
Before proceeding further with our discussion of spherical geometry, let us
review how to change variables in the Laplacian, using the method of Courant
and Hilbert [111, pp. 224225]. Assume that the substitution mapping (x, y, z) to
(u1 , u2 , u3 ) is differentiable with differentiable inverse. Let the Jacobian matrix of
the change of variables be

(x, y, z)
.
(u1 , u2 , u3 )

A=

(2.2)

Then the volume element is

dx dy dz = A du1 du2 du3 ,

A = absolute value of determinant of A. (2.3)

The Euclidean arc length element is

dx
du1
ds2 = (dx dy dz) dy = (du1 du2 du3 ) t AA du2 .
du3
dz

Thus we obtain
ds2 =

gi j dui du j

where

G = t AA = (gi j )1i, j3 .

(2.4)

i, j=1

Similarly, one uses the fact that



f f f
,
,
x y z


=

f f f
,
,
u1 u2 u3

A1 ,

to see that


f
x

2
+

f
y

2
+

f
z

2
=

i, j=1

gi j

f f
where G1 = (gi j )1i, j3 . (2.5)
ui u j

To see what happens to the Laplacian in the new coordinate system, one can use
the calculus of variations (see Courant and Hilbert [111, Chap. 4]). Suppose that f
minimizes the integral
J( f ) =
subject to the constraint

( fx2 + fy2 + fz2 ) dx dy dz

110

2 A Compact Symmetric Space: The Sphere

K( f ) =

f 2 dx dy dz = constant.

Then f must satisfy the EulerLagrange equation

fxx + fyy + fzz = f = f
(see Courant and Hilbert [111, p. 192]). That is, f must be an eigenfunction of the
Euclidean Laplacian. Suppose now that we change variables in the integrals J and
K. The new constrained minimization problem should lead to a differential equation
involving the transformed version of the Laplacian. And this is indeed the case.
From (2.3) and (2.5), one obtains
J( f ) =

gi, j fui u j


|G| du1 du2 du3 ,

with

|G| = det G,

i, j=1

and
K( f ) =

f2


|G| du1 du2 du3 .

The EulerLagrange equation for the transformed problem is



f
|G| gi,k
= f |G|.
uk
k

ui
i

Thus the Laplacian in the new coordinate system is

1

f = fxx + fyy + fzz = |G| 2

ui |G| 2 gi,k uk .

i=1

(2.6)

k=1

There are other ways to obtain these formulas. See, for example, the following
references: Arfken [13, Chap. 2], Yvonne Choquet-Bruhat et al. [87, p. 307],
Churchill and Brown [89, pp. 1619], or Helgason [277, pp. 386387]).
Now we can show that the sphere is a Riemannian manifold, meaning that there
is a notion of arc length defined from an inner product on the tangent space to the
surface at each point. In order to obtain the arc length element on S2 , we shall use
the preceding discussion to write the Euclidean arc length element in spherical polar
coordinates. This is done in the next exercise.
Exercise 2.1.2 (Spherical Polar Coordinates). Spherical polar coordinates for a
point (x, y, z) in R3 are defined by
x = r sin cos , y = r sin sin , z = r cos ,

2.1 Fourier Analysis on the Sphere

111

where 0 2 , 0 , 0 r. Use formulas (2.3), (2.4), and (2.6) above

to transform the Euclidean volume, arc length, and Laplacian to spherical polar
coordinates.
Answer.
d = dx dy dz = r2 sin dr d d ,
d 2 + r2 
sin2 d 2 ,
ds2 = dx2 +dy2 + dz2 = dr2 + r2 
1

(r2 sin rf ) + sin f + sin1

f = r2 sin
r



It follows from Exercise 2.1.2 that the element of arc length on the unit sphere
S2 is
ds2 = d 2 + sin2 d 2 .

(2.7)

This element of arc length is invariant under O(3) as is the corresponding volume
or area element:
d = sin d d .

(2.8)

Finally the Laplacian on S2 is

=

1
sin


sin


+

1 2
sin 2


.

(2.9)

Exercise 2.1.3. Show that the Euclidean Laplacian f = fxx + fyy + fzz is invariant
under O(3) (i.e., that commutes with rotation). Deduce that the spherical
Laplacian (2.9) is also invariant under O(3).
Hint. Dym and McKean [147, p. 243], have a proof using the Fourier transform.
Because the sphere is a Riemannian manifold, we can consider geodesics on the
sphere. The geodesic through two points on the sphere is the curve through these
two points that minimizes distance. Airplane pilots know that geodesics in S2 are
great circles; i.e., the intersection of S2 with a plane through the origin and the two
given points on S2 . To see this, suppose that you are given points p and q on S2 . You
can rotate S2 so that both points p and q have -coordinate in (2.1) equal to zero.
See Exercise 2.1.11 if you do not believe this. Then the distance between p and q
on S2 is
 q
 q
d 2 + sin2 d 2
d = (q) (p).
p

Thus the great circle route minimizes distance on the sphere (assuming that you go
in the proper direction around the circle).

112

2 A Compact Symmetric Space: The Sphere

Exercise 2.1.4 (The Sphere Is a Symmetric Space). Show that S2 is a symmetric

space. To do this, it only remains to show that for each point p in S2 there is a
diffeomorphism f p : S2 S2 that preserves arc length (i.e., f p is an isometry of the
Riemannian manifold) and reverses geodesics.
Hint. At p = (0, 0, 1), use ( , )  ( , ).
Spherical geometry is non-Euclidean because the geodesics cannot be extended
indefinitely, which violates Euclids second postulate. In fact, Euclids fifth postulate
fails as well. And Klein noticed that if you identify antipodal points on the sphere,
then any two geodesic lines have a unique point in common. The geometry obtained
by identifying antipodal points on the sphere is called elliptic geometry and the
resulting compact surface is also called the projective plane. The term elliptic is
due to Klein (from a Greek work elleipein, to fall short). It is motivated by the
following picture. Suppose that two geodesic rays, R1 and R2 , emanate from the
ends of a geodesic segment perpendicular to them both. The distance between the
rays R1 and R2 will decrease or fall short.
In elliptic or spherical geometry, when A, B, C are the three angles of a geodesic
triangle, then the area of the triangle is A + B + C , measuring the angles in
radians, of course. This result goes back to Albert Girard (15951632) for the
sphere. References for non-Euclidean geometry are Coxeter [113] and Hilbert and
Cohn-Vossen [297].
Exercise 2.1.5. Prove Girards formula for the area of a spherical triangle, which
was mentioned above.
Note. This could be proved using the GaussBonnet formula (see Singer and
Thorpe [604]) and, in fact, it was the first known case of the GaussBonnet theorem.

2.1.2 Spherical Harmonics

It is now possible to make sense of the definition of a surface spherical harmonic,
which was given at the beginning of this section. A surface spherical harmonic Y
is a function Y : S2 C such that


1
1
Y = Y
(2.10)
(sin Y ) +
Y =
sin
sin
for some eigenvalue C. The theorem that follows characterizes these spherical
harmonics very explicitly in terms of (associated) Legendre polynomials. Legendre
had discussed these polynomials before Laplaces work, but Legendre did not
manage to publish his results until after Laplace did (see Legendre [404, 405] and
Laplace [396]). Spherical harmonics also played a role in Laplaces long treatise on
celestial mechanics, as well as in Gausss work on terrestrial magnetism.

2.1 Fourier Analysis on the Sphere

113

The following exercises give some of the properties of Legendre polynomials

Pn (x) defined by
2n n!Pn(x) =


dn
2
(x 1)n ,
n
dx

for n = 0, 1, 2, . . . ,

(2.11)

and associated Legendre functions Pnm (x) defined by

Pnm (x) = (1 x2)m/2

d m Pn (x)
.
dxm

(2.12)

Note that some authors replace (1 x2 ) by (x2 1) in formula (2.12). We follow

Courant and Hilbert [111] and Arfken [13] rather than Lebedev [401] in our choice
of notation.
Exercise 2.1.6 (Legendre Polynomials and Associated Legendre Functions).
(a) Show that the associated Legendre functions are solutions of the SturmLiouville
equation:



m2
(1 x2 )u 2xu + n(n + 1)
u = 0.
1 x2
(b) If n = 0, 1, 2, . . . , show that Pnm (x) = 0 for all x, unless m = 0, 1, 2, . . ., n.
Note that we can allow m to be negative by writing (d/dx)1 (d/dx) = identity.
Show also that
Pnm (x) = (1)m

(n m)! m
P (x).
(n + m)! n

(c) Show that for fixed m in {0, 1, 2, . . .}, the set

{ Pnm | n = |m|, |m| + 1, |m| + 2, . . .}
is a complete orthonormal set for L2 [1, +1].
Hint. The orthogonality is easy. For completeness when m = 0, see Courant and
m (x) = (1 x2 )m/2 f m (x).
Hilbert [111, pp. 8283]. In the general case, write Pm+k
k
Then { fkm , k = 0, 1, 2, . . .} is a complete set of orthogonal polynomials for
L2 ([1, +1], (1 x2)m ). This is the weighted L2 -space with inner product given by
( f , g) =

 +1
1

f (x)g(x) 1 x2 dx.

This means that you must show that fkm is orthogonal to xr when r = 0, 1, . . . , k 1.
Integration by parts, using the case m = 0, will do the trick.

114

2 A Compact Symmetric Space: The Sphere

Exercise 2.1.7 (Integral Formula for Associated Legendre Functions). Show

that the associated Legendre Pnm is represented by the following integral:
Pnm (x) = im

(n + m + 1)
(n + 1)

 
0

x+



x2 1cos n cos(m ) d

for Re x > 0, m = 0, 1, 2, . . ..
Hint. See Courant and Hilbert [111, p. 505].
Exercise 2.1.7 has been generalized to symmetric spaces by Harish-Chandra.
This will be discussed in later chapters.
Theorem 2.1.1 (Spherical Harmonics).
(a) The only eigenvalues of the spherical Laplacian defined in (2.10)
are = n(n + 1), n = 0, 1, 2, . . .. The vector space of eigenfunctions of
corresponding to the eigenvalue = n(n + 1) has dimension 2n + 1.
A complete orthogonal set of eigenfunctions of is
Y ( , ) = exp(im )Pnm (cos )

for n = 0, 1, 2, . . . , |m| n.

(2.13)

Here Pnm is the associated Legendre function (2.12) and the coordinates ( , )
are defined by (2.1). We call cY , with Y given by formula (2.13), a (surface)
spherical harmonic of degree n and order m, where c is a normalizing
constant.
(b) Let f (x, y, z) = rnY ( , ), using spherical polar coordinates (r, , ) corresponding to the rectangular coordinates (x, y, z) in Exercise 2.1.2. Then Y is
a surface spherical harmonic satisfying
Y = n(n + 1)Y
if and only f (x, y, z) is a homogeneous harmonic polynomial of degree n. When
we say that f is harmonic, we mean that f satisfies Laplaces equation
f = fxx + fyy + fzz = 0.
Exercise 2.1.8. Prove part (a) of Theorem 2.1.1, using separation of variables on
Y = Y and Exercise 2.1.6.
Note. There are other proofs that { = n(n + 1), n = 0, 1, 2, . . .} are the only
possible eigenvalues of on S2 . For example, see Dym and McKean [147, pp.
252253], Kirillov [353, pp. 271274], or Van der Waerden [691, p. 21].
Exercise 2.1.9. Use separation of variables on f = 0 to prove part (b) of
Theorem 2.1.1. Note that Exercise 2.1.2 says that if f (x, y, z) = R(r)Y ( , ), then

2.1 Fourier Analysis on the Sphere

115

Fig. 2.2 Zero set of

Pn (cos )

f =

1 2
R
(r R (r)) Y + 2 Y
r2
r

where

d
.
dr

If f = 0, then Y = Y and (r2 R (r)) = R, for some constant .

You should not be surprised that the eigenvalues of the Laplacian are real since
the operator is self-adjoint. The next exercise explains why they are negative.
Exercise 2.1.10 (Why the Eigenvalues of the Laplacian on the Sphere Are
Negative). Suppose that f and f are in L2 (S2 ). Show that
( f , f ) 0.
Use Greens theorem. Then show that the eigenvalues of are all negative or zero.
According to part (a) of Theorem 2.1.1, separation of variables in Y = Y
leads from functions Y ( , ) of the two angle variables in (2.1) to functions
exp(im )Pnm (cos ). If we assume that Y is constant with respect to the angle ,
then Y = Y ( ) = Pn (cos ). Such a spherical function Y = Y ( ) is called a zonal
spherical function. The name results from the fact that the zeros of Y cut the sphere
up into zones, as in Fig. 2.2. In general, Pn (cos ) has n distinct zeros in 0
which are positioned symmetrically about = /2 (see Fig. 2.2). There is a more
group theoretical definition of zonal spherical function, which is developed in the
following exercise.
G/ K , where G = S O ( 3 ) , K =
Exercise 2.1.11 (Zonal Spherical Functions on K \G
S O ( 2 ) ). Let G = SO(3) and K = SO(2) be considered as a subgroup of G as in
Exercise 2.1.1. We know from Exercise 2.1.2 that the sphere can be identified with
G/K. Show that the space K\G/K of double cosets KgK, g G, can be represented
by the cosets

Kg K,

with

cos
g = 0
sin

0 sin
1
0 ,
0 cos

0 < .

116

2 A Compact Symmetric Space: The Sphere

Table 2.1 Surface spherical harmonics

(2n+1)(nm)! m
m
Ym
n ( , ) = (1)
4 (n+m)! Pn (cos ) exp(im )
s
p

Y00 = 14


Y 1 = 3 sin exp (i )
1
 8
3
Y0 =
cos
1
4
15

sin2 exp (2i )

Y 2 = 32

2


Y21 = 815 sin cos exp (i )

5 3
Y0 =
cos2 12
2
4

Y33 = 6435 sin3 exp (3i )

2
105

Y3 = 32
sin2 cos exp (2i )



1
21
2

Y3 = 64 sin 5 cos 1 exp (i )

7 5
Y0 =
cos3 3 cos
3

When do two of these cosets coincide? The corresponding points to gK on the sphere
are ge3 , where e3 = t (0, 0, 1). So if

cos sin 0
k = sin cos 0 ,
0
0 1
we find that k g e3 = t (cos sin , sin sin , cos ). Thus a function on K\G/K
is a function only of the angle .
Let A = {g , |0 }, with g from Exercise 2.1.11. Then the decomposition
SO(3) = KAK, which follows from the preceding exercise, is the Euler angle
decomposition of SO(3). See Volume II [667] for a generalization of this decomposition and its application to harmonic analysis on general symmetric spaces.
Exercise 2.1.12. Check Table 2.1 which lists the first few surface spherical harmonics. The table uses the CondonShortley [101] convention.
We can use Mathematica to visualize spherical harmonics using density plots on
7
the sphere. See Fig. 2.3 for two such pictures. On the left is that of ReY14
( , )
obtained using the Mathematica command:
ParametricPlot3D[ {Cos[ ] Sin[ ], Sin[ ] Sin[ ], Cos[ ]}, { , 0, 2 }, { , 0, },
PlotPoints ->300,Mesh->None,ColorFunction->Function[ {x, y, z, , },
Hue[Re[SphericalHarmonicY[14,7, , ]]]],ColorFunctionScaling->

False].
7 ( , )+2ReY 3 ( , ) obtained using a similar
On the right is that of ReY14
7
command.

2.1 Fourier Analysis on the Sphere

117

7 ( , ) on the left and for ReY 7 ( , )+2ReY 3 ( , )

Fig. 2.3 Mathematica density plot for Y14
14
7
on the right

These should be compared with analogous plots for the analogous functions on
a unit square in the plane in Fig. 1.4, which we considered in Sect. 1.3.
Exercise 2.1.13. (a) Make some similar plots to those in Fig. 2.3 using the Mathematica command
Re[SphericalHarmonicY [m, n, , ]].
(b) Solve the wave equation on the sphere with initial condition modelling what
would happen if the sphere were struck with a hammer at a point. Then make
a Mathematica movie of your solution.
(c) Consider the effect of forcing a sphere to vibrate at some frequency near an
eigenfrequency of .
Corollary 2.1.1 (Harmonic Analysis on the Sphere). Every function f : S2 C
with continuous second-order derivatives can be expanded in an absolutely and
uniformly convergent series of spherical harmonics. If we take Ynm , |m| n, to be
an orthonormal basis for L2 (S2 , d ), with d as in Exercise 2.1.2, then
f ( , ) =

n0 |m|n

f(n, m)Ynm ( , ),

where
f(n, m) =


S2

f Ynm d .

118

2 A Compact Symmetric Space: The Sphere

Exercise 2.1.14. Prove the preceding corollary of Theorem 2.1.1.

Hint. The corollary can be proved from the existence of the Greens function for
the problem (see Courant and Hilbert [111, p. 369]) or you might try to relate this
problem to Theorem 1.3.1 of Sect. 1.3.
Exercise 2.1.15 (The Gibbs Phenomenon for the Sphere). Find an analogue of
Exercises 1.3.1 and 1.3.2 of Sect. 1.3 for the sphere.
Hint. There is a discussion of the Gibbs phenomenon as well as summability
procedures in Weyl [730, Vol. I, pp. 305353, 376389 and Vol. IV, pp. 432456].
There are many facts about Fourier series that would be worthwhile to extend
to Laplace series of spherical harmonics. For example, it is possible to obtain the
analogue of the results of Slepian and Pollak (see Grunbaum et al. [234]). And it
is possible to obtain a central limit theorem for the sphere or even SO(3) (see Clerc
and Roynette [92]).
Expansions in spherical harmonics are useful in many areas. For example,
Backus and Gilbert [19], among others, have used such expansions relative to the
vibration of the earth due to large earthquakes to study the structure of the interior of
the earth. Some of the flavor of their work can be derived from the following quote.
The Chilean earthquake of May 22, 1960 excited oscillations of the earth with periods
from six to one hundred minutes which were observed continuously on strain gauges for
265 hours after the main shock and at one-minute intervals on a gravimeter for 110 hr. In
the Fourier spectra of these records, peaks were observed which corresponded well with
the characteristic seismic oscillations computed by Pekeris and Gilbert and MacDonald for
certain earth models. Some of the observed peaks, however, differed in period from the
theoretical spectral lines, and some of the observed peaks appeared to be double or triple,
the components being separated by as much as two minutes and the theoretical line falling
approximately at their mean position.
The displacements of observed peaks relative to theoretical frequencies presumably
reflect slight errors in the earth model used in the theory and will not be discussed further
here. The line splitting cannot be explained thus; we propose to explain it quantitatively as
an effect of the diurnal rotation of the earth.

2.1.3 Quantum Mechanics: The Hydrogen Atom

According to the most simplistic version of quantum mechanics (i.e., neglecting
spin and relativistic effects), the wave function for the hydrogen atom satisfies
the Schrodinger equation


2
e2

+
2m
r

= E,

r = (x2 + y2 + z2 )1/2 .

(2.14)

Here e is the charge of the electron,  = Plancks constant divided by 2 , is

the Laplacian ( f = fxx + fyy + fzz ), m = me m p /(me + m p ) is the reduced mass

2.1 Fourier Analysis on the Sphere

119

of the system if me is the mass of the electron and m p the mass of the proton. The
eigenvalue E is the energy level of the system. Some references for these matters
are Biedenharn and Louck [39], Castellan [80], Condon and Odabasi [100], Courant
and Hilbert [111, pp. 341343], Eyring, Walter, and Kimball [170], Mackey [442,
pp. 159271], Messiah [465, especially pp. 362 and 412], Sommerfeld [610, pp.
200206], B. Thaller [676], Van der Waerden [691, Chap. 1], Dorothy Wallace and
J. BelBruno [713], Weyl [731, pp. 4170], and Wigner [738].
The eigenvalues E that lie in the discrete spectrum of the Schrodinger operator in
equation (2.14) will be the only ones of interest for our discussion. Such eigenvalues
are negative (in fact, there is also a continuous spectrum of positive real numbers).
Physicists interpret the values of E that lie in the discrete spectrum as energy levels
of bound states of hydrogen. Spectroscopists going back to Balmer in the 1880s
have found various series of lines in the spectrum of hydrogen corresponding to
these energy levels. We can obtain some understanding of these spectral lines by
finding the discrete spectrum of the Schrodinger operator very explicitly.
We shall use spherical harmonics to separate variables via (x, y, z) =
w(r)Y ( , ) in equation (2.14) for the purpose of obtaining the discrete spectrum
eigenvalues E. Then Exercise 2.1.2 shows that (2.14) becomes
Y
(r2 w ) + (2mr2 /2 )(E + e2/r)w
=
,
w
Y

where

d
.
dr

Both sides of this equation must be constant. Thus, by Theorem 2.1.1,

Y = n(n + 1)Y, n = 0, 1, 2, . . . ,


2mr2
e2
2
(r w ) + 2
E+
w = n(n + 1)w.

r
Using Exercise 2.1.16 on Laguerre polynomials we find that
(2n+1)

w(r) = rn exp(cr/(2k))Lk+n

(cr/k),

k Z, k > n, c = 2me2 /(2 ).

Furthermore, the corresponding eigenvalues are

Ek =

me4
.
22 k2

(2.15)

Physicists call k the principal quantum number, n the azimuthal quantum

number. If Y = Ynp , |p| n, from Theorem 2.1.1, then p is called the magnetic
quantum number. For each k, there are
k2 =

k1

(2n + 1)

n=0

(2.16)

120

2 A Compact Symmetric Space: The Sphere

linearly independent eigenfunctions of (2.14) with eigenvalue Ek . Physicists call this

k2 -fold degeneracy. For example, spectroscopists use the letters (s, p, d, f , q, . . .) to
indicate the value of n corresponding to the state of hydrogen. They would say the
ground state corresponds to k = 1 and is a 1s state. The first excited state corresponds
to k = 2. It is said to be fourfold degenerate because it contains one 2s state and three
2p states.
Thus for each value k, n of the principal and azimuthal quantum numbers there
are (2n + 1) experimentally indistinguishable states of the hydrogen atom. This
mathematically explains the Zeeman effect in which spectral lines of hydrogen
split into an odd number of lines when a magnetic field is switched on. The effect
was first observed by Zeeman in 1896.
Note that Ek given by formula (2.15) must equal h , where is the frequency of
a spectral line. And we obtain

= Rk2 ,
where R is the Rydberg constant, R = 2 2 me4 h3 . When an atom moves from initial
state 1 with principal quantum number k1 to final state 2 with principal quantum
number k2 , the frequency of the associated spectral line is

= R(k22 k12 ).
The series of spectral lines of hydrogen observed by Balmer in 1885 has k2 = 2.
Thus the Balmer series lines have frequencies

= R(22 k2 ),

k = 3, 4, 5, . . . .

These lines are in the visible range. Lyman found a series of spectral lines in the
ultraviolet range in 1909 with k2 = 1. The series with k2 = 3, 4, are in the infrared
range. The Balmer series is not obtained in the laboratory because it requires the
temperature of stellar atmospheres.
The theory just described does not account for the fine structure of the spectrum.
Relativistic effects and spin must be considered. Even then, agreement between
theory and experiment is not perfect. See Messiah [465, Vol. II, pp. 930933], for a
discussion of the Lamb shift.
There is an additional degeneracy of (n + 1)2 rather than 2n + 1, which was
explained by Fock, who showed that if one formulates the Schrodinger equation
for the hydrogen atom as an integral equation, then it is also invariant under SO(4).
See Louck and Metropolis [430] for a discussion of a related problem which was
motivated by Focks result.
Spectroscopists daily analyze all sorts of materialsatoms, molecules, crystals,
solids, gases. But, of course, the theory becomes much more complicated for the
many-body problem. Group theory helps, as many of the references mentioned at
the beginning of this discussion show.

2.1 Fourier Analysis on the Sphere

121

Exercise 2.1.16 (Laguerre Polynomials). Set

L p = ez

d p z p
(e z ) = the pth Laguerre polynomial.
dz p
(2n+1)

Show that w(r) = rn exp(cr/(2k))Lk+n

2mr2
(r w ) + 2

2

(cr/k) satisfies



e2
E+
w = n(n + 1)w,
r

for E given by equation (2.15), k Z, k > n, c = 2me2 /(2 ). Then show that we
have found the only continuous bounded eigenfunctions w(r) for the differential
operator involved here.
Hint. See Arfken [13, Chap. 13], or Courant and Hilbert [111, p. 330].
We have found a wave function solving equation (2.14) of the form
(2n+1)

(r, , ) = Ynm ( , )rn exp(cr/2k)Lk+n

(cr/k),
(b)

where Ynm denotes a spherical harmonic as in Theorem 2.1.1 and La is the bth
derivative of the Laguerre polynomial La . Then | |2 can be interpreted as the
probability density of the electron for the various states of the hydrogen atom.
Fig. 2.4 from White [733, p. 71], illustrates the first few states.

2.1.4 The Suns Magnetic Field

Using the measured magnetic field of the suns surface, the problem is to determine
the magnetic field of the corona, assuming that the latter is current-free. The corona
itself was not recognized until the last half of the nineteenth century. A new green
spectral line was found to appear only in the corona and just above the thin layer
next to the solar surface which emits the bright red Balmer line of hydrogen. By
1930, eighteen spectral lines had been discovered in the solar corona. And by the
early 1940s it was realized that the green line comes from forbidden transitions in
highly ionized metal atoms. The lack of spectral lines of hydrogen and helium in
the corona implies that these elements are completely ionized. The corona is mostly
hydrogen plasma at one to two million degrees K. There are also cooler regions and
warmer regions. Since the corona is too hot to be in thermal equilibrium, there is
a solar wind. And observed coronal holes are sources of streams of plasma, which
ultimately give rise to terrestrial phenomena such as aurorae and disturbances in
radio transmission. A more complete discussion of these basic facts can be found in
Altschulers article in Herman [290, pp. 105145]. A fascinating account of solar
theory can be found in Zirin [755].

122

2 A Compact Symmetric Space: The Sphere

Fig. 2.4 Photographs of the electron cloud for various states of the hydrogen atom as made from
a spinning mechanical model. The probability-density distribution is symmetrical about the
-axis, which is vertical and in the plane of paper. (From H.E. White, [733, p. 71]. Reprinted by
permission of McGraw-Hill)

2.1 Fourier Analysis on the Sphere

123

The idealized mathematical problem involved in determining the magnetic field

of the solar corona using measurements from the solar surface can be posed as
follows: find , where
= 0,
R < x < Rw ,
(x) = f (x), if x = R, the suns surface,
(x) = 0,
if x = Rw , sufficiently far from the sun.
The function f (x) is obtained from the measured magnetic field of the suns surface.
The relation B = grad gives the magnetic field itself. The assumption is that
there are no magnetic monopoles and thus must be harmonic. So we can use
Exercise 2.1.9 and Corollary 2.1.1 to Theorem 2.1.1 to obtain

(r, , ) = R

n0
|m|n

1
1 a2n+1

 
  
R n+1
2n+1 r n
a
f(n, m)Ynm ( , ),
r
R

where a = R/Rw < 1 and

f(n, m) =


S2

f Ynm d .

This formula is satisfactory as a solution in a textbook, but the numerical problem

remains to be solved. See Altschulers article mentioned above for a description of
the numerical work that led to the illustrations in Fig. 2.5.
Another interesting application of spherical harmonics can be found in
Chandrasekhar [83, Chap. VI]. The same sort of picture that gave us an idea of
the shape of the probability distribution of the electron in a hydrogen atom should
give an idea of the distribution of heat in a uniform sphere. Several applications
of spherical harmonics to geophysics can be found in Gilbert [212], Kato [341],
Stacey [616], and Trefil [683, pp. 106111].
It is possible to generalize Theorem 2.1.1 to the unit sphere Sn in Rn+1 . In
fact, spherical harmonics for such higher-dimensional spheres have been of use in
quantum mechanics, as have the representations of many higher-dimensional Lie
groups. In the Lecture Note-Reprint collection Dyson [148] gives an interesting
glimpse into the arguments concerning the choice of symmetry groups for nuclear
and particle physics. The compact groups SU(n), n = 3, 6, 12, seem to have attracted
the most attention. Here SU(n) is the special unitary group of n n complex
matrices U such that t UU = I and |U| = 1. Some of the papers in the Dyson
collection are devoted to showing that the representations of SU(6) are somehow
incompatible with those of the Lorentz group O(3, 1) which leaves the Maxwell
equations invariant. See R. Hermann [291, pp. 144149] and Talman [655, pp.
186187] for discussions of Focks work on applications of four-dimensional
spherical harmonics to quantum mechanics. The standard model in particle physics
has as its local gauge group SU(3) SU(2) SU(1). Representations of the group

Fig. 2.5 Open magnetic structures on the sun. (From Altschulers article in Herman [290, p. 125]. Reprinted by permission of Springer-Verlag)

124
2 A Compact Symmetric Space: The Sphere

2.1 Fourier Analysis on the Sphere

125

correspond to the fields. There are many books on the subject. We leave it to you
to Google them, or see Wikipedia. There is also the Grand Unified Theory, or GUT.
This involves SU(5), SO(10) and even the exceptional Lie group E8 . Again there is
much on the Internet including some beautiful movies related to E8 .
Hecke and others have made use of these higher-dimensional spherical harmonics in number theory and the kinetic theory of gases (see Hecke [258, pp. 361373,
849854]) and Ogg [505, Chap. VI, p. 6]. We shall see in the next section that
spherical harmonics also give a new description of harmonic analysis on Rn .

2.1.5 Connections with Group Representations

It is not difficult to connect spherical harmonics with representations of SO(3).
Before doing so, let us sketch some of the basic facts about group representations.
We will be sketchy because we do not intend to emphasize the representationtheoretic point of view. It would be useful for the reader to study these things
however, and the following references will provide plenty of food for thought: Barut
and Raczka [28], Boerner [45], Gelfand, Minlos, and Shapiro [205], Gurarie [238],
Hamermesh [246], Knapp [356], Mackey [442444], Maurin [458], Sugiura [648],
Talman [655], Van der Waerden [691], Varadarajan [692], Michelle Vergne [697],
Vilenkin [704], Wallach [714,715], Warner [717], Wawrzynczyk [722], Weyl [730
732], Wigner [738], and Williams [741]. Mackeys introduction to Biedenharn and
Louck [39] provides a translation of physicists terminology for mathematicians (or
vice versa).
Suppose that H is a separable Hilbert space and let GL(H) be the group of
invertible continuous linear maps from H to H. We want to consider representations
of a topological group G by elements of GL(H). By a topological group G we
mean a group with a topology such that multiplication (x, y) xy and inversion
x  x1 are continuous maps. A topological group is locally Euclidean if there is
a neighborhood of the identity e in G which is homeomorphic to an open subset of
Rn . A Lie group is a locally Euclidean topological group with group operations that
are infinitely differentiable maps. Most, if not all, of the Lie groups considered here
are groups of matrices like O(n).
A representation of a topological group G is a pair (T, H) where H is a separable
Hilbert space and T : G GL(H) is a continuous group homomorphism, using the
strong operator topology on GL(H). Actually a stronger definition is made when G
is not locally compact (see Kirillov [353, p. 111]). You might wonder why we want
to represent groups of matrices by their infinite-dimensional analogues. The answer
is that this cannot be avoided if you wish to do Fourier analysis on a noncompact
group.
Let (v, w), v, w H, denote the Hilbert space inner product on H. We say that
the representation (T, H) is unitary if (T (g)v, T (g)w) = (v, w) for all v, w in H and
g in G. The representation (T, H) is said to be finite-dimensional if H is finitedimensional and then dim H = degree of the representation.

126

2 A Compact Symmetric Space: The Sphere

Suppose that v, w are elements of H. The function g  (T (g)v, w) = mw,v (g),

for g G, is called a matrix entry of the representation (T, H). Many of the
special functions which are useful in applied mathematics are either matrix entries
of representations of well-known Lie groups or are part of an orthogonal basis for
the Hilbert space of such a representation.
The theory of group representations is closely connected with the search for
eigenfunctions of differential operators. For if a group G leaves a differential
operator L invariant, then G takes eigenfunctions of L into eigenfunctions of L with
the same eigenvalue. So G acts on the vector space H of eigenfunctions of L for a
fixed eigenvalue to give a representation of G.
We say that a representation (T, H) of G is irreducible if there is no closed
subspace W of H with {0}  W  H such that T (g)W W for all g G.
Two representations (T1 , H1 ) and (T2 , H2 ) of a Lie group G are said to be
equivalent if there is a linear bicontinuous bijection A : H1 H2 such that T2 (g)A =
AT1 (g) for all g in G. Then A is called an intertwining operator. Or even more
generally, if A is only continuous linear, it is still called an intertwining operator.
If the group is Rm under addition, the irreducible unitary representations are onedimensional and a complete list of inequivalent irreducible unitary representations
of Rm under addition consists of the exponentials




m = ea (x) = exp t ax , for x Rm | a Rm
R

(2.17)

Exercise 2.1.17 (Schurs Lemma).

(a) Suppose that (T1 , H1 ) and (T2 , H2 ) are finite-dimensional irreducible representations of G that are inequivalent. If A : H1 H2 is a linear map such that
AT1 (g) = T2 (g)A, for all g in G, then A = 0. Prove this.
(b) Suppose that (T, H) is an irreducible representation of G on a finite-dimensional
complex vector space H. If A : H H is a linear map such that AT (g) = T (g)A,
for all g in G, show that there is a complex number c such that A = cI, where I
is the identity operator; i.e., Iv = v, for all v in H.
Note. The spectral theorem for unitary operators allows you to extend the result
of the preceding exercise to unitary representations on infinite-dimensional Hilbert
spaces.
We say that a finite-dimensional representation (T, H) of G is completely
reducible if H = H1 Hn , with T (g)Hi Hi , for all g G, provided that
the representations (Ti , Hi ) of G obtained by restricting T (g) to Hi are irreducible
for all i = 1, . . . , n. We say T = T1 Tn = the direct sum of the Ti .
Exercise 2.1.18. Given a basis e1 ,. . . , em of the representation space H of a finitedimensional representation (T, H) of G, you can form the matrix corresponding to
T (g). If e1 , . . . , em is an orthonormal basis of H, then this matrix has (i, j) entry
(T (g)ei , e j ) = m j,i (g), for g G.

2.1 Fourier Analysis on the Sphere

127

(a) Show that two finite-dimensional representations (T1 , H) and (T2 , H) are equivalent if and only if the matrix of T1 is obtained from the matrix of T2 by changing
the basis of H that is used.
(b) Show that a finite-dimensional representation (T, H) of G is completely
reducible if and only if you can find a basis of H which puts the matrix of
T in block diagonal form:

T1
.. . .
. .

0
..
.

0 Tn
so that the representation determined by the matrix Ti is irreducible for each
i = 1, . . . , n.
We will often need to be able to integrate functions on topological groups. The
integral involved is an analogue of the Lebesgue integral on Rn and is supposed to
come from a countably additive positive measure on the Borel sets in the group.
Such an integral on a topological group G is right invariant if

G

f (x) dx =


G

f (xa) dx

for all a in G.

The invariant integral on a compact Lie group was used by Hurwitz, Schur, and
Weyl beginning in the 1890s. Haar proved the existence of a invariant integral on
any locally compact topological group in 1933 and thus the integral is called the
Haar integral. If G = Rm , under addition, then the Haar integral is the standard
Lebesgue integral. Proofs of the existence of the Haar integral may be found in
Helgason [277, p. 365], Lang [387], Pontryagin [518], or Weil [725], for example.
The invariant integral can be used to show that any irreducible unitary representation of a compact group must be finite-dimensional (see Kirillov [353, p. 135]),
for example. And you will need the invariant integral to do Exercise 2.1.20. In fact,
integration on topological groups will be a necessary tool for the rest of this book.
Of course, our groups will be matrix groups and thus the Haar integral for these
groups is not so mysterious.
The right-invariant Haar measure dx is unique up to a positive constant multiple.
This means that for each g in G, there is a positive constant (g) (not to be confused
with the like-named distribution) defined by


f (gx) dx = (g)

f (x) dx.

Then : G R+ is a continuous homomorphism, which is called the modular

function of G (not to be confused with the modular functions that appear in
Chap. 3). Clearly this function relates the right- and left-invariant Haar integrals.
In particular, one has

128

2 A Compact Symmetric Space: The Sphere

f (x) (x) dx =

f (x1 ) dx.

The group G is said to be unimodular when (g) = 1 for all g in G. This means
that right-invariant Haar integrals are also left invariant. Many groups that we shall
consider are unimodular. In particular, all compact or abelian groups are unimodular.
Most of the Lie groups that we shall consider in this book are what is called
semisimple; for example, the special linear group of all n n real matrices of
determinant one. It can be shown that all semisimple Lie groups are unimodular (see
Helgason [277, p. 366]). However, not all groups are unimodular. For example, the
group of upper triangular matrices with positive diagonal entries is not unimodular
if n 3 (Exercise).
Exercise 2.1.19. Show that any finite-dimensional unitary representation is completely reducible and that its decomposition into irreducible representations is
unique up to equivalence.
Exercise 2.1.20 (The Haar Integral on a Compact Group).
(a) Show that any compact group is unimodular; i.e., right Haar measure = left
Haar measure.
(b) Use Haar measure to show that any representation of a compact group is
equivalent to a unitary representation.
Suppose that (T1 , H1 ) and (T2 , H2 ) are representations of G. Define the tensor
product representation (T1 T2 , H1 H2) by
(T1 T2 )(g)(v1 v2 ) = (T1 (g)v1 ) (T2(g)v2 ) for g G, vi Hi , i = 1, 2.
See the references mentioned above for more information on tensor products. The
decomposition of tensor products of representations into direct sums of irreducible
representations has much importance for quantum mechanics. For example, if
one ignores the interaction between two electrons in the field of a positive nucleus,
the Schrodinger operator of the system has eigenfunctions which are products 1 2
of eigenfunctions i corresponding to representations Ti of O(3), i = 1, 2. And the
product 1 2 corresponds to the representation T1 T2 . The ClebschGordon series
breaks T1 T2 up into its irreducible components. Thus one can conclude what sort
of spectral lines should occur for such a situation. See the references mentioned in
Sect. 2.1.3 on quantum mechanics for more details.
There are many other useful topics in representation theory such as induced
representations, Frobenius reciprocity, Cartans theorem on the highest weight,
and Weyls character formula. See the references on group representations for a
discussion of these matters.
The systematic study of group representations of finite groups began in the
1890s with work of Frobenius, Schur, and others. In the 1920s, Weyl obtained the
irreducible representations of the compact simple Lie groups such as G = SO(3) (see
Weyl [730, Vol. II, pp. 543647]). To do this, Weyl used his formula for the character

2.1 Fourier Analysis on the Sphere

129

of a representation T , which is the trace of T (g), g G. The representations of

compact Lie groups G are studied by restricting them to a maximal abelian subgroup
A or torus in G. Any representation of A decomposes into a direct sum of onedimensional representations called weights. Among the weights there is a highest
one and this characterizes the original representation of G, up to equivalence. We
have actually seen an example of the theorem on the highest weight in our study
of spherical harmonics, but this will not be developed here. Another result called
the BorelWeilBott theorem realizes the representations of compact Lie groups on
sheaf cohomology groups (see Warner [717]).
The character T of a finite-dimensional representation T is defined by T (g) =
Trace(T (g)) for g G. Define G = the dual of G to be the set of equivalence classes
of irreducible unitary representations of G. For example, since the complete list of
m can be identified
irreducible unitary representations is given by formula (2.17), R
with Rm . It is said that Rm is self-dual. If G is compact, describing G is equivalent
to describing the characters of G.
When infinite-dimensional representations are required, the character is defined
as a distribution when possible. If f L1 (G), define
T(f) =


G

f (g)T (g) dg,

which means that for x, y H, we have

(T ( f )x, y) =


G

f (g)(T (g)x, y) dg,

with (, ) = the Hilbert space inner product on H. If H has an orthonormal basis ei ,

then we can often define the character of T via
Tr T ( f ) = (T ( f )ei , ei) where Tr = Trace,
i

considering the map f  Tr T ( f ) as a distribution. When H is finite-dimensional,

Tr T ( f ) =

(Tr T (g)) f (g) dg.

When G is a tame unimodular Lie group the abstract Plancherel theorem (or
Fourier inversion theorem) says there is a measure d on G called the Plancherel
measure such that
f (e) =

Tr T ( f ) d (T ),

for infinitely differentiable functions f on G with compact support. See Michelle

Vergne [697] for a interesting discussion of this subject with many examples.
Harish-Chandra obtained the Plancherel measure for real semisimple Lie groups,
for which the Plancherel inversion formula does not always involve an integral over

130

2 A Compact Symmetric Space: The Sphere

In Chap. 1 we saw the result for Rm and Rm /Zm . In this chapter

all of the dual G.
we have considered the result for functions on SO(3)/SO(2).
One of the main goals of these volumes is the explicit description of the
Plancherel measurenot for functions on the group GL(n, R) but instead for
functions on the symmetric space GL(n, R)/O(n). See Volume II (i.e., [667]) for
this result and its history.
Formula (2.18) below gives the Plancherel theorem for the group SO(3).
Exercise 2.1.21 (The Irreducible Unitary Representations of SO(3)). Let
{Ynm , |m| n, n = 0, 1, 2, . . .} denote a complete orthonormal set of spherical
harmonics as in Theorem 2.1.1. Define a (2n + 1) (2n + 1) matrix An (g) for g in
SO(3) by
Ynm (gx) =

(An )m,kYnk (x).

|k|n

(a) Justify this formula for An (g) and then show that An (g) defines a representation
of SO(3).
(b) Show that An (g) is a unitary representation using


S2

Ynm (gx)Ynk (gx) dx =

S2

Ynm (x)Ynk (x) dx.

(c) Show that the representation An (g) is irreducible.

In fact, the representations An (g) form a complete set of inequivalent irreducible
unitary representations of SO(3). This means that any function f in L2 (SO(3)) with
respect to Haar measure has a Fourier series expression, converging in the L2 norm

f (g) =
n=0 (2n + 1) Trace [ f (n)An (g)], where
f(n) =


SO(3)

f (g)

tA

n (g) dg, dg = Haar measure on G.

(2.18)

This is proved, for example, in Dym and McKean [147, pp. 256261]. Vilenkin
[704, pp. 440457], examines the representations of SO(n) defined in an analogous
way to that used for n = 3 in Exercise 2.1.21 and shows that if n > 3, then
these representations do not exhaust all of the irreducible unitary representations
of SO(n). Only the class-one representations are obtained in this way. A class-one
representation (A, H) of G has a vector v in H such that A(k)v = v for all k in K.
Here G = SO(n) and K = SO(n 1) (embedded in G as in Exercise 2.1.1).
Formula (2.18) giving Fourier analysis on the group SO(3) is a special case of the
PeterWeyl theorem for any compact group (see Weyl [730,
+ Vol. III, pp. 5875],
,
Pontryagin [518], or Weil [725]). The theorem says that if T = (mi j ), A
is a complete system
of irreducible unitary representations of a compact topological

group G, then d mi j forms a complete orthonormal set in L2 (G), where d =

degree of T . The proof is not hard using the spectral theorem for compact self-

2.1 Fourier Analysis on the Sphere

131

adjoint operators (Theorem 1.3.7 of Sect. 1.3). The compact self-adjoint operator
used in the proof of the PeterWeyl theorem is the convolution operator T f = g f ,
where g is a fixed nonzero continuous function on G such that g(x) = g(x1 )
for all x in G. The convolution is taken with respect to Haar measure on G. The
eigenfunctions for a fixed eigenvalue of T form a finite-dimensional vector space on
which G operates by sending f (x) to f (xa) for a G. This gives a representation of
G. The eigenfunctions for T form a complete orthonormal set in L2 (G) and thus so
do the matrix entries of the unitary irreducible representations of G. It follows that
the analogue of formula (2.18) replaces An by the irreducible unitary representations
of G and (2n + 1) by the degree of the irreducible representation.
The fact that special functions come from representations often leads to an easier
understanding of the myriads of formulas listed in books such as Erdelyi et al. [164,
165].
Exercise 2.1.22 (Addition Formulas for Matrix Entries). Suppose that (T, H) is
a finite-dimensional representation of G and that e1 , . . . ,en is an orthonormal basis
of H. Let the i, j th matrix entry of T be mi j (g) = (T (g)ei , e j ) for g in G. Here (, )
denotes the inner product on H. Show that T (gh) = T (g)T (h) implies the following
addition formula for the matrix entries:
mi j (gh) =

mik (h)mk j (g).

k=1

Exercise 2.1.23 (Addition Formula for Spherical Harmonics). For x, y in S2 ,

prove that

|m|n

Ynm (x)Ynm (y) =

2n + 1 t
Pn ( xy).
4

Note that both sides are unchanged if you replace x by gx and y by gy for g in
SO(3). So the right and left sides of the equation can only differ by a constant. Find
the constant by setting x = y and integrating over S2 .

2.1.6 Integral Equations for Spherical Harmonics

It is possible to characterize spherical harmonics by integral equations rather than
differential equations. This method often simplifies the theory, for the same reason
that Greens functions simplify the theory of differential operators (see Sect. 1.3
and the discussion surrounding equation (1.16)). Examples of the method appear
in Weyls work on spherical harmonics (see Weyl [730, Vol. III, pp. 386399]),
Selbergs work on the trace formula (see Selberg [569]), and Harish-Chandras work
on group representations (see Harish-Chandra [253, 254]). The method of integral
equations also makes it easier to extend results to finite, p-adic and adelic groups

132

2 A Compact Symmetric Space: The Sphere

(see Gelbart [200], Godement [213], Macdonald [440], Tamagawa [656]). This
should not, however, cause us to forget the differential equations and the contact
with applied mathematics.
The following theorem was proved in 1916 by Funk [187] and in 1918 by Hecke
[258, pp. 208214].
Theorem 2.1.2 (The FunkHecke Theorem). Suppose that Yn is a spherical
harmonic of degree n. Let f : [1, +1] C be continuous. Then

S2

f (t ux)Yn (u) du = 2 Yn (x)

 +1
1

f (t)Pn (t) dt.

Proof. Since every continuous function on [1, +1] can be uniformly approximated
by zonal spherical functions Pn (x), it suffices to do the following exercises.


Exercise 2.1.24 (Properties of Spherical Harmonics).
(a) Show that

S2

Ym (x)Pn (t xy) dx =

4
Yn (y)mn .
2n + 1

Here mn = 0 if m = n and mn = 1 if m = n.
Hint. Use the addition formula of Exercise 2.1.23.
(b) Prove that
 +1
1

Pn (x)Pm (x) dx =

2
mn .
2n + 1

The FunkHecke theorem actually characterizes spherical functions. This result

is generalized in Helgason [277, p. 439, Cor. 7.4]. And we shall also use this
approach in later chapters. In order to generalize the FunkHecke theorem, one
should view the spherical harmonic Y as a function on G = SO(3) and replace the
formula in Theorem 2.1.2 by
Y (I)


K

Y (xky) dk = Y (x)


K

Y (ky) dk

for all x, y in G,

(2.19)

where dk is Haar measure on K = SO(2) embedded in G as in Exercise 2.1.1, and

I = the identity matrix. Furthermore zonal spherical harmonics P are characterized
by the integral equation


P(I)
K

P(xky) dk = P(x)P(y) for all x, y in G.

(2.20)

2.2 O(3) and R3 : The Radon Transform

133

This characterization of zonal spherical harmonics is due to Gelfand [201] in the

general case.
Exercise 2.1.25 (The Group-Theoretical Version of the FunkHecke Theorem).
Show that formula (2.19) implies the FunkHecke theorem by integrating (2.19)
times f (y) where f : K\G/K C.
Exercise 2.1.26 (The Convolution Theorem for SO(3)). Let f , g be integrable
functions on SO(3) with respect to Haar measure. Define the convolution of f and
g by
( f g) (x) =


SO(3)

f (u)g(xu1 ) du

for x in SO(3).

Define the Fourier transform f(n) for n = 0, 1, . . . as in formula (2.18). Show that

( f g)(n) = f(n)g(n)

for all n = 0, 1, 2, . . . .
Hint. Use the fact that An is a representation.
Exercise 2.1.27 (The Fundamental Solution to the Heat Equation on the Sphere
S2 R3 ). Given the initial heat distribution f ( ), solve the following initial value
problem:
u( , ,t) = ut ,

u( , , 0) = f ( ),

t > 0.

Here is the Laplacian on the sphere given by formula (2.10).

Answer. u = Gt f , where
Gt ( , ) = Gt ( ) =

cn Pn(cos ) exp[n(n + 1)t].

n>0

The constants cn are chosen so that Gt as t 0+ .

Note. G. Watson [720] considers various analogues of the Gaussian distribution for
the sphere, in connection with various statistical problems such as that of studying
the distribution of the unit normal vectors to the planes of all known comets.
Another reference for group theory and statistics is the book by Diaconis [132]
which includes an interesting story about testing for uniformity, Fisher, Jeffries, and
continental drift on pages 170171.

2.2 O(3) and R3 : The Radon Transform

It is interesting to consider the method by which algorithms have been justified mathematically in this field [computerized tomography]. While this method consists of mathematical

134

2 A Compact Symmetric Space: The Sphere

reasoning in a certain sense, the reasoning is far from rigorous. Approximations are
introduced at many steps with only intuition as a guide to the error involved. We do not
know of a single instance in which a tomographic algorithm has been justified in a truly
rigorous sense. Thus, in contrast to some other workers in this field, we do not feel that one
derivation is more rigorous than another, whether it is based on Radons inversion formula,
the Fourier inversion formula or any other foundation.
From Shepp and Kruskal [588, p. 421].

2.2.1 Harmonic Analysis on R3 in Spherical Polar

Coordinates and Spectral Measures
It is easy to see (cf. Exercise 2.1.1) that the Fourier transform on Rn commutes with
rotation. This leads to a new formulation of harmonic analysis on Rn in terms of
spherical harmonics. For simplicity, we shall consider only the case n = 3 here. The
generalization to Rn , n arbitrary, can be found in Coifman and Weiss [98] as well
as in Stein and Weiss [636]. These results are due to Cauchy and Poisson for radial
functions, and Bochner and Hecke in general (see Bochner [43, p. 235], and the
article of Stein in J. M. Ash [15, pp. 104105]).
Exercise 2.2.1 (The Fourier Transform on Rn Commutes with Rotation). If g

O(n) and f : Rn C, set (L(g) f )(x) = f (gx) for x Rn . Show that L(g) f = L(g)
f,
when f is a sufficiently nice function, that the Fourier transform of f exists (as
defined in Sect. 1.2 for various classes of functions).
Exercise 2.2.2 (J-Bessel Functions). Define the Bessel function of the first kind
by the power series
J (z) =

(1)k (z/2) +2k

(k + 1)(k + + 1) ,

for |arg z| < .

k=0

(a) Show that y(z) = J (z) satisfies Bessels equation:



y + (1/z)y + 1 ( /z)2 y = 0.
(b) Show that J (z) is represented by the integral formula:
J (z) =

(z/2)
(1/2)( + 1/2)

 +1
1

when Re > 12 and |arg z| < .

(c) Show that J 1 (z) + J +1(z) = 2z J (z).
(d) Prove that

(1 t 2) 1/2 exp(izt) dt,

2.2 O(3) and R3 : The Radon Transform

Jn+ 1 (z) = (1)

135

2 n+ 1
z 2

1 d
z dz

n

sin z
, n = 0, 1, 2, . . . .
z

These functions Jn+ 1 are often called spherical Bessel functions.

2

Exercise 2.2.3 (More Properties of Bessel Functions).

(a) Asymptotics. Show that
1

J (z) (1
 + )
(z/2) ,  as z 0;

J (z) 2z cos z
2 4 , as z .

(b) Show that Jn (z) = (1)n Jn (z), when n = 0, 1, 2, . . .. Then show that if is not
an integer, J and J are linearly independent.
(c) Functional Equation. Prove that J (z) = exp( i)J (z).
(d) Addition Formula. Set R = (r12 + r22 2r1 r2 cos )1/2 and prove that
J 1 (R) =
2

  

1
1
2R

m + 2 Jm+ 12 (r1 )Jm+ 12 (r2 )Pm (cos ).

r1 r2
2 m=0

Theorem 2.2.1 (The BochnerHecke Formula). Suppose that Y (u), u S2 is a

surface spherical harmonic of degree n (as in Theorem 2.1.1 of Sect. 2.1) and let
f (ru) = g(r)Y (u) for r R+ , u S2 , where g : R+ C, is such that

0

|g(r)|r2 dr < .

Then the Euclidean Fourier transform as defined in Sect. 1.2 of f (assuming that f
is nice enough for the transform to exist), is given by
2
f(ru) = n Y (u)
i r


sR+

g(s)s3/2 Jn+ 1 (2 rs) ds.

2

Proof. Let dv denote the element of surface area on S2 . Then, using Theorem 2.1.2
of Sect. 2.1, we have
f(ru) =

sR+ vS2

= 2 Y (u)

exp(2 irs t uv)g(s)Y (v)s2 ds dv

sR+

g(s)s2

 +1
1

exp(2 rst)Pn(t) dt ds.

The proof of this theorem is completed in Exercise 2.2.4.

Exercise 2.2.4. Show that Pn (x) =

1
2n n!

dn
dxn

(x2 1)n implies that

136

2 A Compact Symmetric Space: The Sphere

 +1
1

(i)n
exp(2 irst)Pn(t) dt = Jn+ 1 (2 rs).
2
rs

Use integration by parts and Exercise 2.2.2. Then complete the proof of
Theorem 2.2.1.
How does the formula of Bochner and Hecke fit into the general scheme of
harmonic analysis on symmetric spaces? Why did the Bessel functions suddenly
appear out of the blue? To understand this, one must view R3 as a symmetric space
rather than just as an additive group. We explained this at the end of Chap. 1. It
leads one to think of the full group of isometries of R3 , namely, the Euclidean
group M(3, R) of rigid motions of R3 . If we consider only the orientation-preserving
motions G, then G is the semidirect product of the groups T and K, where T consists
of translations and is isomorphic to R3 , 
and K= SO(3). One can view M(3, R) as
Ab
a group of 4 4 matrices of the form
, A SO(3), b R3 . The matrix
01


Ab
sends x R3 to Ax + b. Clearly one can identify R3 with G/K by mapping
01
a coset gK, g G, to g0 = the result of applying the motion g to the origin 0 in R3 .
Exercise 2.2.5. Do an analogue of the BochnerHecke theorem for R2 for rotationinvariant functions f on R2 .
Answer. For nice rotation-invariant functions f on R2 , one has
f( u) = 2

f (r)J0 (2 r)rdr.

The connection of Theorem 2.2.2 below with the representations of

the Euclidean group (as well as the generalization to Rn ) is made in
Vilenkin [704, Chap. 11] and Wawrzynczyk [722]. The case of R2 is
discussed in Dym and McKean [147, pp. 263273]. See also Helgason
[277, pp. 402403], and Talman [655, Chap. 12].
Exercise 2.2.6 (The Eigenfunctions of the Euclidean Laplacian in Spherical
Polar Coordinates). Use Exercise 2.1.2 of Sect. 2.1 to show that if f = fxx +
fyy + fzz = f , and f (ru) = g(r)Y (u) for r R+ , u S2 , and Y is a surface spherical
harmonic of degree n, then
1
g(r) = Jn+ 1 (2 rt), = (2 t)2 .
2
rt
Theorem 2.2.2 (Harmonic Analysis on R3 in Spherical Polar Coordinates). Let
en,m,t (ru) = 2 Ynm (u)(rt)1/2 Jn+ 1 (2 rt) for r R+ , u S2 , where {Ynm | n =
2
0, 1, 2, . . . , |m| n} denotes a complete orthonormal set of surface spherical

2.2 O(3) and R3 : The Radon Transform

137

harmonics of degree n (as in Theorem 2.1.1 of Sect. 2.1). Then any f in L2 (R3 )
has a Fourier expansion (converging in the L2 -norm) of the form
f (x) =

n0 |m|n t>0

where
f(n, m,t) =


R3

f(n, m,t)en,m,t (x)t 2 dt,

f (y)en,m,t (y) dy.

Proof. It suffices to assume that f (ru) = g(r)Ynm (u) for r R+ , u S2 . Then we

must show that



1
1
g(r) = (2 )2
g(t) Jn+ 1 (2 st)t 2 dt Jn+ 1 (2 rs)s2 ds. (2.21)
2
2
rs
st
s>0
t>0



This is done in the following exercise.

Exercise 2.2.7. (a) Prove formula (2.21) by writing down the inversion formula
for the ordinary Fourier transform (from Theorem 1.2.1 of Sect. 1.2) in spherical
polar coordinates. Then use Theorem 2.2.1 to evaluate the inner Fourier
transform. Finally, Theorem 2.1.2 of Sect. 2.1 and Exercise 2.2.4 should be used
to complete the proof.
(b) Show that formula (2.21) is equivalent by change of variables to a special case
of Hankels inversion formula: (assuming > 12 and r > 0) :

f (r) =


0

yJ (yr)


0

xJ (xy) f (x) dx dy.

(2.22)

Hankels inversion formula (2.22) gives the spectral resolution of the singular
SturmLiouville operator
(L f ) (x) =



2
1
f
(x f ) +
x
x

for x (0, )

(2.23)

This spectral resolution can be derived from a formula of Stieltjes, Stone, Kodaira,
and Titchmarsh, which is itself a corollary of the Von Neumann Spectral Theorem
for unbounded operators. Let us summarize the theory briefly, following Lang [387],
Reed and Simon [540, Chap. 7], Stakgold [618, 619], and some unpublished notes
of J. Korevaar. More details can be found in these references as well as Dunford and
Schwartz [145], Gelfand and Vilenkin [207], Levitan and Sargsjan [415], Maurin
[459], Naimark [490], Titchmarsh [679], Weyl [730, Vol. I, pp. 195297], and
Yosida [748]. In particular, Dunford and Schwartz [145, Vol. II, pp. 13331392,
15321533], provides an extremely careful discussion of spectral theory for singular
differential operators, including the Bessel operator (2.23).

138

2 A Compact Symmetric Space: The Sphere

We want to consider linear, possibly unbounded, operators T : D V , where D

is a dense subspace of a Hilbert space V . Define
D = {y V | there is a z in V with (T x, y) = (x, z), for all x D}.
Define T = the adjoint of T by (T x, y) = (x, T y). We say that T is self-adjoint if
D = D and T = T . The spectrum (T ) of the operator T is defined to be



(T ) = C  (T I)1 does not exist as a bounded linear operator on V .
The point or discrete spectrum of T (or the set of eigenvalues of T ) consists of
C such that (T I) is not one-to-one. The continuous spectrum of T is the
set of C such that (T I) is 1 - 1 and D = range (T I) is dense in V but
(T I)1 is an unbounded linear operator with domain D.
The von Neumann Spectral Theorem says that if T is any self-adjoint operator
(bounded or not), there is a Stieltjes integral representation
I=

dE ,

T=

dE ,

(2.24)

for some family of projection-valued measures dE (cf. Reed and Simon [540,
Chap. 7]). The integrals are over the spectrum of T , which is real, because T
is
self-adjoint. The spectral theorem implies that for polynomials p( ), p(T ) =

p( )dE . It is possible to define f (T ) for continuous functions f on the spectrum
of T and then this same formula holds upon replacing p by f .
In order to obtain the formula of Stieltjes, Stone, Kodaira, and Titchmarsh one
must be able to compute the Greens function or resolvent kernel G( ; x, y)
defined by
(T I)1v(x) =

G( ; x, y)v(y) dy.

(2.25)

For if we have a convergent integral of the form

f ( ) =

 +

( )1g( )d ,

then at any point c of continuity of g( ), we have

1
2 i

lim 0+ ( f +
(c + i ) f (c i ))

= lim 0+

1  +

( c)2 + 2

g( ) d

= g(c).

This is Poissons integral formula for a half-plane (see Ahlfors [3, p. 171]). It follows
that the spectral measure for the operator T is given by the formula of Stieltjes,
Stone, Kodaira, and Titchmarsh:

2.2 O(3) and R3 : The Radon Transform

2 i

139

dE
= (T ( + i0)I)1 (T ( i0)I)1 ,
d

(2.26)

assuming that E is well-behaved. This formula is proved in Dunford and Schwartz

[145, Vol. II], Lang [387, pp. 412413], and Reed and Simon [540, p. 237]. Stieltjes
[639] found the result in 1894. We will usually refer to formula (2.26) as the
KodairaTitchmarsh formula for brevity.
In order to apply formula (2.26), we need a way to compute the Greens function
G( ; x, y). We use the method given in Stakgold [618, 619]. Consider a Sturm
Liouville operator which is singular at a and b:
Lf =

(p f ) + q f ,
w

a < x < b.

(2.27)

Here singular means that either the interval is infinite or the functions w(x) or
q(x) blow up, or p(x) vanishes at some point in [a, b]. The Hilbert space associated
to (2.27) is L2 ([a, b], w) consisting of Lebesgue measurable functions f on [a, b]
such that
 b
a

| f (x)|2 w(x) dx < .

The inner product for this weighted L2 -space with weight w is

( f , g) =

 b
a

f (x)g(x)w(x) dx.

If one makes the correct assumptions about p, q, w and if one imposes the correct sort
of boundary conditions at a and b, the operator L will be self-adjoint. In particular,
p, q, w should be real and w should be positive.
The Greens function G( ; x, y) of formula (2.24) must satisfy


p

G( ; x, y) + (q w)G( ; x, y) = (x y).
x

Pick c in (a, b). We want to choose two linearly independent solutions and
of Lu = u, as follows. Suppose that solves Lu = u and in addition lies in
L2 ([a, c], w). Let solve Lu = u and L2 ([c, b], w). Then
G( ; x, y) =
and

1
c

(x) (y), a < x < y < b

(y) (x), a < y < x < b

(2.28)

140

2 A Compact Symmetric Space: The Sphere



 (x) (x) 
.

c = 
p (x) p (x) 
Example 2.2.1 (Spectral Measure for the Hankel Transform Using the Kodaira
Titchmarsh Formula). Consider the SturmLiouville operator given by formula
(2.23). Then formula (2.28) becomes

i J ( 1/2 x)H(1) ( 1/2 y), 0 < x < y <
G( ; x, y) =
2 J ( 1/2 y)H(1) ( 1/2 x), 0 < y < x < .
(1)

Here J is the Bessel function of the first kind from Exercise 2.2.2 and H is the
Hankel function (or Bessel function of the third kind) defined by
H1 (z) =

J (z) exp( i)J (z)

i sin( )
(1)

when is not an integer. Take limits as n to define Hn when n is an integer. It

follows from Lebedev [401, pp. 112113], that the determinant c = 2i/ in this
case.
In order to check this formula for G( ; x, y), one needs the following asymptotic
properties of the Bessel functions from Lebedev [401, Chap. 5], ( > 0):
1
J (x) ( 2x ) (1+
as
) ,



J (x) 2x cos x
as
2 4 ,

(1)
H (x) i 2x ( ),
as




(1)

H (x) 2x exp i x
, as
2 4

x 0,
x ,
x 0,

(2.29)

x .

(1)

Note that if 0 < < 1, both J ( 1/2 x) and H ( 1/2 x) are in L2 ([0, 1], x). One says
that 0 is then in the limit-circle case in Weyls theory (cf. Stakgold [618,619]). If
(1)
1, then J ( 1/2 x) is in L2 ([0, 1], x) and H ( 1/2 x)
/ L2 ([0, 1], x). One says that 0
(1)
1/2
/ L2 ([1, ], x) and H ( 1/2 x)
is then in the limit-point case. Note that J ( x)
2
L ([1, ], x) so that infinity is always in the limit-point case.
To compute the jump of G( ; x, y) required by formula (2.26), assume that x < y.
Then for > 0,
G( + i0;
 x, y) G( i0; x, y)
=

i
2


(1)
(1)
J ( 1/2 x)H ( 1/2 y) J ( 1/2 x)H ( 1/2 y) .

So we need the functional equations

2.2 O(3) and R3 : The Radon Transform

141

J (x) = exp( i)J (x),




(1)
(1)
H (x) = exp( i) H (x) 2J (x) .
These formulas imply
G( + i0; x, y) G( i0; x, y) = i J ( 1/2 x)J ( 1/2 y).
Thus we obtain from (2.24) and (2.26),

(x y) 1
=
x
2


0

J ( 1/2 x)J ( 1/2 y) dy =


0

J ( x)J ( y) d .

This is Hankels integral theorem.

For higher-rank symmetric spaces, the method just used to compute the spectral
measure proves rather unwieldy because the Greens functions are much more
complicated. Thus we will formulate another method in Chap. 3 and Volume
II [667]. For this method, one needs only the asymptotics and functional equations
of the eigenfunctions appearing in the spectral expansion of the operator L, rather
than the expression for the resolvent kernel, involving other eigenfunctions as well.
The method is due to Harish-Chandra and it would be very interesting to derive
Harish-Chandras Plancherel measure from the spectral theorem in a similar way to
that which gives formula (2.26).
One can also ask for the spectral decomposition of the operator in (2.23) on
the finite interval (0, a). The problem still has a singularity at 0. But on (0, a), the
spectral resolution of the operator L in (2.23) is given by a series rather than an
integral, when > 1:
f (x) =

cm J ( m x/a),

0 x a,

> 1,

(2.30)

m=1

where { m }m1 is the set of all positive roots of J ( m ) = 0 and

cm = 2a2 [J +1 ( m )]2

 a
0

f (r)J ( m r/a)r dr.

The FourierBessel series expansion (2.30) is derived in Titchmarsh [679, Vol. 1,

pp. 8186]. See also Stakgold [619, pp. 305308, 313315]. It is quite interesting
to let a approach infinity in (2.30) and watch (2.30) approach (2.22) (cf. Morse and
Feshbach [479, Vol. I, pp. 762766]).
Exercise 2.2.8 (The KontorovichLebedev Transform). Consider Bessels equation with the role of the parameters interchanged:
(xw ) + xw w/x = 0,

0 < x < .

142

2 A Compact Symmetric Space: The Sphere

Here > 0 and is the eigenvalue. Show that the Greens function is

G( ; x, y) =

Ii ( 1/2 x)Ki ( 1/2 y), 0 < x < y < ,

Ii ( 1/2 y)Ki ( 1/2 x), 0 < y < x < ,

where I and K are the Bessel functions of imaginary argument. The definitions
are
I (z) = exp( i/2)J (z exp( i/2)),
K (z) =

< arg z < /2,

i
(1)
exp( i/2)H (z exp( i/2)),
2

< arg z < /2.

See Chap. 3 and Lebedev [401] for more information on these functions.
Show that K (x) = K (x) and I (x) I (x) = 2 sin( )K (x)/ . Then prove
the KontorovichLebedev inversion formula:
x (x y) =

2
2

Ki ( x)Ki ( y) sinh ( ) d .

This will be a central result in Sect. 3.1.

2.2.2 CAT Scanners and the Radon Transform

Modern X-ray scanners can reproduce the tissue density function f (x), x R2 , for
a plane slice of a persons head, for example. They operate by inverting the Radon
transform, defined for k R, u S1 , by
R f (k, u) =


t xu=k
xR2

f (x) ds,

(2.31)

where the integral is over a line in the plane and ds is the element of arc length. The
inversion formula for this transform goes back to Radon [528] in 1917. It says that
f (x) =

1
d
q>0 q

1
2


u S1


R f (q + t xu, u) du .

(2.32)

Here the integral over q in R+ can be viewed as a Stieltjes integral or as a Cauchy

principal value integral, assuming that f is continuous with compact support.
References for the Radon transform include Bracewell [62], Dym and McKean
[147], Gelfand, Graev, and Vilenkin [204], Herman [290], Helgason [279], Louis
[431, 432], Ludwig [435], Elena Prestini [522], Quinto [524], and Shepp and
Kruskal [588]. In fact, Funk [187] proved the analogue of (2.32) for S2 rather than

2.2 O(3) and R3 : The Radon Transform

143

S1 one year earlier than Radon proved his result. Helgason [279] shows that a vast
generalization of this theory is possible, viewing the Radon inversion formula as
involving two dual integrationsone over the points in a hyperplane and the other
over the hyperplanes through a point. The Radon transform appears in so many
applications to signal processing that Matlabs Signal Processing Toolbox has a
Radon transform command. We look at finite analogues of the Radon transform
in [668, pp. 7172]
The next sequence of exercises presents a derivation of Radons inversion
formula.
Exercise 2.2.9. (a) Suppose that f : R2 C is a Schwartz function. Show that the
Fourier inversion formula can be written in the form
f (x) =


uS1

E f ( t ux, u) du,

where
E f (t, u) =

1
2


rR kR

|r|R f (k, u) exp[2 ir(t k)] dk dr.

In the first formula du denotes the angle measure on the unit circle S1 . The
Radon transform R f (k, u) is defined by (2.31).
(b) Let abs(r) = |r|, r R. Show that abs(r) is not the Fourier transform of a
Lebesgue integrable function.
 were a function,
Hint. Recall the RiemannLebesgue lemma. Note that if abs(r)
then you could write:
E f (t, u) =

1
2


kR


abs(t
k)R f (k, u) dk.

Exercise 2.2.10 (Derivatives and Fourier Transforms of Some Distributions).

Define the Cauchy principal value integral by

PV


f (x) dx = lim


|x|>


f (x) dx .

Define
(x1 , ) = PV

x1 (x) dx

and
2

(x , ) = PV

x2 ( (x) (0)) dx

144

2 A Compact Symmetric Space: The Sphere

for test functions as in Sect. 1.1. Prove that in the sense of distributions (see
Sect. 1.1) the following formulas are valid:
(a)
(b)
(c)
(d)

x1 = (log |x|) .
x2 = (x1 ) .
1 = i sgn(x), where sgn(x) = abs(x)/x.
x
 = (2 2)1 x2 .
abs(x)

Hint. See Vladimirov [706, pp. 75, 86, 134], or Bracewell [61, p. 130].
Exercise 2.2.11. Derive Radons inversion formula from Exercises 2.2.9
and 2.2.10, using properties of Fourier transforms of distributions.
Note. The Hilbert transform of a function is
H f (x) =

 +
f (t)

(x t)

dt =

1
f (x1 ).

Thus part (c) of Exercise 2.2.10 implies that


H
f (x) = i sgn(x) f.
Because Radons inversion formula (2.32) contains a derivative, it does not
appear to be as useful for numerical calculation as the formula of Exercise 2.2.9.
So those that have computed these transforms in practice have used approximations
 (cf. Herman [290, pp. 19ff], and Shepp and Kruskal [588]).
to abs(x)
Exercise 2.2.12. Compute the Fourier transform of the following approximation to
abs(x) for x R2 :

f (x) =

|x|, if |x| < A,

0, otherwise.

Compare the decay at infinity with that of |x|2 from part (d) of Exercise 2.2.10.
Note. You can write the Fourier transform in the plane as a Bessel transform, much
as we did in Theorem 2.2.1 for the Fourier transform in 3-space. See Exercise 2.2.5.
Some History. In 1956, R. Bracewell used a method analogous to the Radon
transform to study solar radiation. In the 1960s and 1970s CT scanners for medicine
were developed independently by A. M. Cormack and G. N. Hounsfield. They were
not based on the Radon transform. The Hounsfield algorithm was used in the first
commercial CAT scanner made by EMI Central Research Labs. in the UK. The first
patient brain scan was done in 1971. A. M. Cormack and G. N. Hounsfield shared
the Nobel Prize in Medicine in 1979.
The four illustrations in Figs. 2.6 and 2.7 from the 1978 paper of Shepp and
Kruskal [588] show a mathematical phantom on the bottom in Fig. 2.6 representing

2.2 O(3) and R3 : The Radon Transform

145

Fig. 2.6 On the bottom is a simulation of a human head using 11 ellipses. On the top is a
reconstruction using the algorithm embodied in the first commercial machine (EMI Ltd.) from
180 160 strip projection data obtained by exact calculation from the image on the bottom.
(From Shepp and Kruskal [588, pp. 422423]. Reprinted by permission of American Mathematics
Monthly)

146

2 A Compact Symmetric Space: The Sphere

Fig. 2.7 On the bottom is a reconstruction from the bottom image in Fig. 2.6 using the Fourier
based algorithm of Shepp. On the top is a reconstruction using the algorithm used in the 1970s
EMI machine from 180 239 strip projection data obtained by exact calculation from the bottom
image in Fig. 2.6. (From Shepp and Kruskal [588, p. 424]. Reprinted by permission of American
Mathematics Monthly

a slice of a human head and reconstructions of this phantom by three different

algorithms. The Hounsfield algorithm was used to produce the top part of Fig. 2.6.
The bottom part of Fig. 2.7 was produced with a Fourier-based algorithm of Shepp.
The top part of Fig. 2.7 was produced with an algorithm used by EMI Ltd. in the
1970s. Clearly progress was made.

2.2 O(3) and R3 : The Radon Transform

147

A description of the progress made between the 1970s and 1980s is contained in
the following quotation from Science 214 (1981), p. 1327:
When CT scanners first became commercially available about 8 years ago it took 5 minutes
to scan a patients head and 5 minutes for each computerized reconstruction of an image
from the x-ray data. Now, because of advances in the design of the scanners and in computer
technology, the newest machines can scan a head just 10 seconds and can reconstruct an
image virtually instantaneously. According to Jay Thomas Payne of Abbott Northwestern
Hospital in Minneapolis, the Mayo Clinics first CT scanner, which is only 5 years old, has
been relegated to the clinics historical museum.

Of course, despite the progress in speed and accuracy of CT scanners, they still
expose people to radiation. In a New York Times article from August 21, 2012, Jane
E. Brody wrote the following.
But it [radiation] also has a potentially serious medical downside: the ability to damage
DNA and, 10 to 20 years later, to cause cancer. CT scans alone, which deliver 100 to
500 times the radiation associated with an ordinary X-ray and now provide three-fourths of
Americans radiation exposure, are believed to account for 1.5 percent of all cancers that
occur in the United States.

Thus nuclear magnetic resonance tomography (NMR alias MRI) may be destined
to be the tomography of the future. It uses magnetic fields rather than x-rays and thus
is presumably less damaging to the body. Louis [431,432] considers applications of
the three-dimensional Radon transform to NMR tomography. Spherical harmonics
are used to improve the algorithm by studying the kernel (i.e., inverse image of
0) of the transform (ghosts). In 2003, the Nobel Prize for Physiology or Medicine
went to Paul Lauterbur and Sir Peter Mansfields for their work on MRI. See Elena
Prestini [522], Chap. 8, for more details on the subject. She begins the chapter
with a quotation from a New York Post article from 1939 about a talk by I. I. Rabi
explaining why NMR scanning works. The quote is: We are all radio stations.
There are many other applications of the Radon transform; for example, in radio
astronomy (see Bracewells article in Herman [290, pp. 81104]). And there are
applications to partial differential equations; e.g., to find solutions of the wave
equation (see Dym and McKean [147, pp. 137139], and Helgason [279].
This concludes our discussion of harmonic analysis on the sphere. It would be
possible to consider spherical analogues of many more results from Chap. 1. We
shall leave this to the interested reader. For example, the group of isometries of the
sphere is just O(3). A discontinuous subgroup O(3) has the property that any
domain in the sphere can contain only finitely many points equivalent under to
any given point. There are very few such discontinuous for the sphere or elliptic
plane. They correspond to the regular polyhedra (see Hilbert and Cohn-Vossen [297,
p. 242]). These are the analogues of the space groups considered in Sect. 1.4. It is
also possible to discuss the analogue of the Poisson summation formula for \S2
=
\G/K, G = SO(3), K = SO(2). We shall not do this here, since our main interest
is the Selberg trace formula for noncompact fundamental domains (see Chap. 3). In

148

2 A Compact Symmetric Space: The Sphere

fact, we never examined the Selberg trace formula when G is the Euclidean group
either (but see Hejhal [265]).
There is work on fast computation of Fourier transforms on the sphere by Driscoll
and Healy [140], for example. There are also finite analogues of Radon transforms
(see [668]).

http://www.springer.com/978-1-4614-7971-0