CH 8 Aaaaaaaaaaaaaaaaaaaaaaaaaa

Chapter 8
Special Functions
Many of the important dierential equations of mathematical physics contain the Laplacian
operator and often other terms such as rst or second derivatives with respect to time. When
solving such an equation using separation of variables one chooses a coordinate system that
reects the symmetry of the problem so that one can satisfy easily the boundary conditions.
This procedure leads to ordinary dierential equations that often have nonconstant coecients
and whose solutions therefore cannot be written in terms of elementary functions.
In all of the problems we will consider here, we will nd that the ODEs that result from
separation of variables are special cases of the Sturm-Liouville equation. We will therefore
immediately know that the solutions have certain important properties, e.g., that they are
orthogonal and form a complete basis. For these problems we will nd that one can write down
a solution as a series, and the corresponding function is often named after the person who was
clever enough to do this rst.
1
In this chapter we will look at some of the named special functions of mathematical
physics, including Bessel functions, Legendre polynomials, spherical harmonics and Hermite
polynomials. We will take look at the most important properties of these functions but we
will only be scratching the surface of a deep and broad topic. There are many more properties
and further special functions we will not have time to cover (Laguerre polynomials, Chebyshev
polynomials, Mathieu functions, . . . ); more information can be found in standard references such
as [1, 2, 7, 8]. Finally in Sec. 8.4 we include a brief description of the Euler gamma function.
8.1 Bessel functions
In this section we will write down and solve the dierential equation that describes the vibrations
of a circular drum. This will give us the opportunity to look at one of the most important
equations in mathematical physics, the wave equation. When we separate variables in terms of
polar coordinates and time, we will nd solutions to the radial equation related to an important
family of functions called Bessel functions. Solving this problem will also give us a chance to
practice the Frobenius method that we saw in Chapter 7.
1
History is of course not so simple. Bessel functions, for example, were developed by Daniel Bernoulli some 80
years before Bessels related work on the subject.
97
98 Lecture Notes on Mathematical Methods
8.1.1 The wave equation
Consider a circular drum of radius a. The lateral displacement of the drums membrane, u, can
be modeled as a function of position and time by the wave equation,
2
u
x
2
+

2
u
y
2
=
1
v
2
2
u
t
2
, (8.1)
where the parameter v represents the speed with which waves propagate across the surface. A
derivation of the analogous one-dimensional wave equation is given in Appendix D, and the
generalization to two or three dimensions is discussed in many texts, such as Ref. [4]. For
purposes of the present discussion we will assume that Eq. (8.1) provides an appropriate model
of our drum and we will take as our goal to investigate its solutions.
Since our drum is circular, we will nd it easiest to use polar coordinates (see Sec. 6.2) so
that the displacement as a function of position and time can be written u(r, , t). Using the
Laplacian for polar coordinates from Sec. 6.2, we can write the wave equation as
2
u
r
2
+
1
r
u
r
+
1
r
2
2
u
2
=
1
v
2
2
u
t
2
. (8.2)
The boundary conditions are
u(a, , t) = 0 , (8.3)
u(r, , t) = u(r, + 2, t) , (8.4)
|u(0, , t)| < . (8.5)
The rst equation above says that the displacement of the drum is zero at its edge, r = a. The
second condition expresses the fact that and + 2 represent the same place on the drum.
The third says the solution must be nite at the origin, which must clearly hold for a physical
drum (we will see that one of the possible solutions to the wave equation diverges at r = 0 so it
it will be necessary to impose this condition).
In addition to the boundary conditions we need to supply initial conditions to determine
a unique solution. Since the wave equation is second order in time, we need to provide two
functions of r and , which we can take as the initial position and speed:
u(r, , 0) = f(r, ) , (8.6)
u
t
(r, , 0) = g(r, ) . (8.7)
Using separation of variables, we seek product solutions (r, , t) of the form
(r, , t) = R(r)Q()T(t) . (8.8)
Special Functions 99
Substituting this into the wave equation (8.2) gives
QTR
+
1
r
QTR
+
1
r
2
RTQ
=
1
v
2
RQT
. (8.9)
Dividing both sides by RQT we nd
R
R
+
1
r
R
R
+
1
r
2
Q
Q
=
1
v
2
T
T
. (8.10)
The left-hand side of (8.10) is only a function of r and and the right-hand side is only a
function of t, so therefore both sides must be equal to a constant. We will see later in Sec. 8.1.6
that this constant must be negative so for convenience we call it
2
, i.e.,
R
R
+
1
r
R
R
+
1
r
2
Q
Q
=
1
v
2
T
T
=
2
. (8.11)
We can therefore write the T equation as
T
+v
2
2
T = 0 , (8.12)
which has the general solution
T(t) = Acos(vt) +Bsin(vt) . (8.13)
We can take the part of Eq. (8.11) that depends on r and and separate it further by
multiplying both sides by r
2
, so as to isolate the dependence of the Q
/Q term. This gives

r
2
R
R
+r
R
R
+
Q
Q
=
2
r
2
. (8.14)
The dependence is entirely contained in Q
/Q and all of the r dependence is in the other terms.

So Q
/Q must be equal to a separation constant, which gives a situation identical to the one
found in Sec. 6.2.2 when solving the Laplace equation in polar coordinates. There by imposing
the periodicity requirement Q( + 2) = Q() we concluded that the separation constant must
satisfy Q
/Q = n
2
where n is an integer. This reasoning leads to the same conclusion here.
This term plus the rst two terms in Eq. (8.14) must be equal to the right-hand side,
2
r
2
, so
we therefore nd
Q
Q
= n
2
, (8.15)
r
2
R
R
+r
R
R
= n
2
2
r
2
, (8.16)
The equation for Q,
Q
+n
2
Q = 0 , (8.17)
has the general solution
Q() = C cos(n) +Dsin(n) . (8.18)
Note that cos(n) and sin(n) with negative n are not linearly independent from the
corresponding terms with positive n and so we will only need to consider n = 0, 1, 2, . . . .
8.1.2 Sturm-Liouville form of the radial equation
Equation (8.16) for the radial part of the solution R(r) can be written
r
2
R
+rR
+ (
2
r
2
n
2
)R = 0 . (8.19)
This is a second-order ODE with nonconstant coecients, which is similar to the Euler
equation (6.88) that we found as the radial part of the Laplace equation in Sec. 6.2.2; it in fact
reduces to Eq. (6.88) if we have = 0. We were able to convert the Euler equation to one
with constant coecients by means of a clever variable transformation. In the present case this
will not work, and in fact it is not possible to express the solution to (8.19) in terms of usual
elementary functions such as sines or exponentials, and the best we will manage is to write the
solution as an innite power series.
Before we proceed to that step, however, we can already say something about the solutions
by noting that Eq. (8.19) is a special case of the Sturm-Liouville equation that we studied in
Chapter 5,
d
dx
_
p
d
dx
_
q = w . (8.20)
This is not immediately obvious when Eq. (8.19) is written in the form given, but if we divide
both sides by r and rearrange terms it becomes
rR
+R
n
2
r
R =
2
rR , (8.21)
or equivalently
d
dr
_
r
dR
dr
_
+
n
2
r
R =
2
rR . (8.22)
We can therefore identify Eq. (8.21) for R(r) with Eq. (8.20) for (x) with
p(r) = r , (8.23)
q(r) =
n
2
r
, (8.24)
w(r) = r . (8.25)
The functions p = r and w = r satises the requirements mentioned in Chapter 5, namely, we
have p > 0, p
is continuous, and w > 0 except at the isolated point r = 0. From Sec. 5.3 we
therefore know that there will be an innite number of real eigenvalues
2
each corresponding
to a distinct eigenfunction, and that the eigenfunctions will be orthogonal and constitute a
complete set. We will examine some of these properties further after we nd a series solution
for R(r).
8.1.3 Bessels equation
Before we nd the series solution for R(r) it is convenient to transform the radial equation (8.16)
into one for a function y of a dimensionless variable x that we dene as
x = r , (8.26)
y(x) = R(r) = R
_
x
_
. (8.27)
We need the derivatives R
and R
, which we can nd using the chain rule,

R
=
dR
dr
=
dy
dx
dx
dr
=
dy
dx
, (8.28)
R
=
dR
dr
=
d
dr
_
dy
dx
_
=
d
2
y
dx
2
dx
dr
=
2
d
2
y
dx
2
. (8.29)
Equation (8.16) can therefore be written
_
x
_
2
2
y
+
x
+ (x
2
n
2
)y = 0 , (8.30)
where the prime now refers to dierentiation with respect to x. Cancelling the factors of gives
x
2
y
+xy
+ (x
2
n
2
)y = 0 . (8.31)
This is called Bessels equation of order n, where the order here is however not the same as
that of the ODE, which is second order. Although in our problem of the vibrating drum we
have seen that n must be an integer, Bessels equation appears in other contexts with noninteger
values for this parameter. We will follow the convention of writing n for an integer and if it
is noninteger. Notice that no longer appears explicitly in the equation when it is written in
terms of x, and therefore only appears in the solution R(r) through the value of x, i.e., through
the product r.
8.1.4 The Bessel functions J
n
and Y
n
We can divide both sides of Bessels equation by x
2
to put it in the standard form where the
coecient of the highest derivative is equal to unity:
y
+
1
x
y
+ (1
n
2
x
2
)y = 0 . (8.32)
Comparing with the standard form from Eq. (7.22),
y
+P(x)y
+Q(x)y = 0 , (8.33)
we see that P(x) = 1/x and Q(x) = 1 n
2
/x
2
both have singularities at x = 0, so this is a
singular point of the dierential equation. However, the functions
xP(x) = 1 , (8.34)
x
2
Q(x) = x
2
n
2
(8.35)
are both nite at x = 0 and thus x = 0 is a regular singular point. Therefore we know that the
Frobenius method will lead to at least one solution of the form
y(x) =
k=0
a
k
x
k+
. (8.36)
Going back to the form of Eq. (8.31) we can compute the required terms as
x
2
y
k=0
a
k
(k +)(k + 1)x
k+
, (8.37)
xy
k=0
a
k
(k +)x
k+
, (8.38)
x
2
y =
k=0
a
k
x
k++2
=
k=2
a
k2
x
k+
, (8.39)
n
2
y =
k=0
n
2
a
k
x
k+
. (8.40)
Notice that when written in terms of x
k+
, the sum in Eq. (8.39) starts at k = 2, whereas the
others all start at k = 0. Substituting these terms into Bessels equation (8.31) gives
1
k=0
_
a
k
(k +)(k + 1) +a
k
(k +) n
2
a
k
_
x
k+
+ (8.41)
k=2
_
a
k
(k +)(k + 1) +a
k
(k +) +a
k2
n
2
a
k
_
x
k+
= 0 . (8.42)
Equating the coecients of x
k+
to zero gives
a
0
(
2
n
2
) = 0 , (8.43)
a
1
_
(1 +)
2
n
2
_
= 0 , (8.44)
a
k
_
(k +)
2
n
2
_
+a
k2
= 0 . (8.45)
The coecient of a
0
gives the indicial equation, which has the two solutions
= n . (8.46)
If n is an integer and = n, then Eq. (8.44) can only be satised if a
1
= 0. The recurrence
relation for the coecient a
k
from a
k2
is therefore
a
k
=
1
n
2
(k +)
2
a
k2
(8.47)
with = n. Since this connects even and odd numbered coecients separately and as we have
a
1
= 0, all of the a
k
are zero for k odd.
From our discussion of the Frobenius method in Sec. 7.2, we know that if the two solutions
to the indicial equation (8.43) do not dier by an integer, then both values for give linearly
independent solutions to the dierential equation. In some problems, Bessels equation may
arise with noninteger n, and in such a case the two linearly independent solutions are given by
the Frobenius series, one with = n and the other with = n.
In our problem of the circular drum, however, the periodicity requirement on the angular
solution Q means that n has to be be an integer, and therefore the two solutions for = n are
not linearly independent. One may take one of the solutions to correspond to = |n|, which is
called a Bessel function of the rst kind, J
n
(x). Using the recurrence relation for the coecients
one nds that J
n
(x) can be written
J
n
(x) =
k=0
(1)
k
k!(k +n)!
_
x
2
_
2k+n
. (8.48)
Here the (arbitrary) value of a
0
has been set to
a
0
=
1
2
n
n!
, (8.49)
which is chosen so that one has
_

0
J
n
(x) dx = 1 . (8.50)
To nd the second linearly independent solution to the Bessel equation we can use reduction of
order, as described in Sec. 1.3.3. The resulting solutions are called called Bessel functions of
the second kind and denoted Y
(x). (In some references they are called Neumann functions and
denoted N
(x).) For noninteger order one can show that they are related to the J
by
Y
(x) =
J
(x) cos() J
(x)
sin()
. (8.51)
For integer order n one can then use
Y
n
(x) = lim
n
Y
(x) . (8.52)
The general solution to Bessels equation can therefore be written
y(x) = AJ
n
(x) +BY
n
(x) . (8.53)
Plots of the Bessel functions of the rst and second kind are shown in Figs. 8.1. Notice that
functions Y
n
(x) diverge to for x 0.
x
0 2 4 6 8 10
(
x
)
n
J
-1
-0.5
0
0.5
1
0
J
1
J
2
J
3
J
x
0 2 4 6 8 10
(
x
)
n
Y
-1
-0.5
0
0.5
1
0
Y
1
Y
2
Y
3
Y
(a) (b)
Figure 8.1: Plots of Bessel functions of (a) the rst kind J
n
(x) and (b) the second kind Y
n
(x) for integer
order n.
Like sine and cosine, the Bessel functions J
n
and Y
n
have an oscillatory behaviour. But
unlike sine and cosine, the spacing between oscillations is not constant. It is not possible to
solve explicitly for the values of x where the functions cross zero, but these can be determined
numerically. Table 8.1 gives values of z
nm
, the mth positive zero of the function J
n
(x).
Table 8.1: Values of z
nm
, the mth zero of J
n
, for n = 0, 1, 2, 3 (computed with the GSL routine
gsl sf bessel zero Jnu [9]).
m z
0m
z
1m
z
2m
z
3m
1 2.40483 3.83171 5.13562 6.38016
2 5.52008 7.01559 8.41724 9.76102
3 8.65373 10.1735 11.6198 13.0152
4 11.7915 13.3237 14.796 16.2235
5 14.9309 16.4706 17.9598 19.4094
6 18.0711 19.6159 21.117 22.5827
7 21.2116 22.7601 24.2701 25.7482
8 24.3525 25.9037 27.4206 28.9084
9 27.4935 29.0468 30.5692 32.0649
10 30.6346 32.1897 33.7165 35.2187
8.1.5 More Bessel functions
Before returning to the problem of the drum we should remark that Bessel functions arise in many
contexts and in a number of dierent forms. They often come up in problems with cylindrical
symmetry (like the drum). An alternative way of expressing the two linearly independent
solutions to Bessels equation is to use the Hankel functions, dened as
H
(1)
(x) = J
(x) +iY
(x) , (8.54)
H
(2)
(x) = J
(x) iY
(x) . (8.55)
The modied Bessel equation,
x
2
y
+xy
(x
2
+
2
)y = 0 . (8.56)
is similar to the Bessel equation (8.31) but with the opposite sign for the x
2
y term. Its two
linearly independent solutions are called the modied Bessel functions, I
(x) and K
(x). These
functions are exponentially growing and decaying, in contrast to the oscillating behaviour of
J
(x) and Y
(x). They are shown for integer order in Fig. 8.2.

If the Helmholtz equation,
2
+k
2
= 0 , (8.57)
is separated in spherical polar coordinates, then the radial equation reduces to
x
2
y
+ 2xy
+
_
x
2
n(n + 1)
_
y = 0 . (8.58)
x
0 1 2 3 4 5
(
x
)
n
I
0
1
2
3
4
5
0
I
1
I
2
I
3
I
x
0 1 2 3 4 5
(
x
)
n
K
0
1
2
3
4
5
0
K
1
K
2
K
3
K
(a) (b)
Figure 8.2: Plots of modied Bessel functions of (a) the rst kind I
n
(x) and (b) the second kind K
n
(x)
for integer order n.
The two linearly independent solutions are called spherical Bessel functions, j
n
(x) and y
n
(x),
which are related to the ordinary Bessel functions J
n
(x) and Y
n
(x) by
j
n
(x) =
_

2x
J
n+1/2
(x) , (8.59)
y
n
(x) =
_

2x
Y
n+1/2
(x) . (8.60)
Plots of spherical Bessel functions j
n
and y
n
are shown in Fig. 8.3 for several values of the order
n
x
0 2 4 6 8 10
(
x
)
n
j
-1
-0.5
0
0.5
1
0
j
1
j
2
j
3
j
x
0 2 4 6 8 10
(
x
)
n
y
-1
-0.5
0
0.5
1
0
y
1
y
2
y
3
y
(a) (b)
Figure 8.3: Plots of spherical Bessel functions of (a) the rst kind j
n
(x) and (b) the second kind y
n
(x)
for integer order n.
The properties of Bessel functions have been investigated for almost 200 years resulting in
a vast literature on the topic. One can derive asymptotic expressions for x 0, x ,
recurrence relations that connect Bessel functions of dierent order, etc. Details can be found
in standard references such as [1, 2, 7, 8] as well as on the excellent Wikipedia page for Bessel
functions.
8.1.6 Justication of the sign of
2
We are now in a position to justify the assumption we made earlier concerning the separation
constant introduced for Eq. (8.10). We said we would show this must be negative and so we
called it
2
. In this section we will show that zero or positive values would not allow us to
satisfy the boundary conditions.
If we had allowed a value of zero, then time solution T(t) from Eq. (8.13) is a constant,
i.e., there are no vibrations, and we would not be able to simultaneously satisfy the boundary
condition (8.3) and the initial conditions (8.6) and (8.7).
If we had allowed a positive value for the separation constant, i.e.,
2
in place of
2
, then
instead of (8.19) for the radial equation we would have obtained
r
2
R
+rR
(
2
r
2
+n
2
)R = 0 . (8.61)
After applying the same transformations from Eqs. (8.28) and (8.29) as used before, x = r and
y(x) = R(r), we nd
x
2
y
+xy
(x
2
+n
2
)y = 0 , (8.62)
which we identify from Eq. (8.56) as the modied Bessel equation. Its solutions are the modied
Bessel functions I
n
and K
n
, so we would nd the general solution to the radial equation from
Eq. (8.29) as R(r) = y(x) = y(r), i.e.,
R(r) = EI
n
(r) +FK
n
(r) . (8.63)
As we saw in Fig. (8.2), the functions I
n
and K
n
are are exponentially rising and falling,
respectively. So with these we would not be able to satisfy the boundary condition that the
solution go to zero at the edge of the drum, and we therefore conclude that the separation
constant which we took as
2
must indeed be negative.
8.1.7 Eigenmodes of a circular drum
Returning now to the problem of the vibrations of a circular drum, we can relate the solutions
R(r) of the radial equation to the solution y(x) of Bessels equation as R(r) = y(x) = y(r),
R(r) = EJ
n
(r) +FY
n
(r) , (8.64)
where E and F are constants that we will x using the boundary conditions. The boundary
condition given by Eq. (8.5) states that the solution must be nite at r = 0. Since all of the Y
n
diverge for r 0, we must have F = 0.
The boundary condition of Eq. (8.3) says that
R(a) = EJ
n
(a) = 0 . (8.65)
We cannot have E = 0 or else we are left with R = 0 (the trivial solution) an therefore the
eigenvalue must take on values such that a is one of the zeros of J
n
. That is, the allowed
values of are
nm
=
z
nm
a
, (8.66)
where z
nm
is the mth zero of J
n
, values of which were shown in Table 8.1.
Collecting together all of the ingredients for the product solutions RQT, we have
R(r) = EJ
n
(z
nm
r/a) , (8.67)
Q() = C cos(n) +Dsin(n) , (8.68)
T(t) = Acos(vz
nm
t/a) +Bsin(vz
nm
t/a) . (8.69)
If we choose the initial speed of the drum to be zero but give it a nonzero displacement, then
we have B = 0. The full solution for the motion of the surface of the drum will therefore be a
linear combination of the form
u(r, , t) =
n=0
m=1
J
n
(z
nm
r/a) cos(z
nm
vt/a) [a
nm
cos(n) +b
nm
sin(n)] . (8.70)
Each term in the sum (8.70) corresponds to specic eigenmodes of vibration, whose form is
determined by the values of n and m. We can write the time dependence of (8.70) as
cos(vz
nm
t/a) = cos(2
nm
t) , (8.71)
where the frequency of a specic eigenmode is given by
nm
=
vz
nm
2a
. (8.72)
Because the spacing of the zeros of the Bessel functions z
nm
is irregular (see Table 8.1), the
mixture of frequencies from a drum results in a dissonant sound.
We can compare our result for the drum to the familiar case of a piano string of length L.
Recall (see, e.g., Ref. [10]) that the lateral displacement of a vibrating string is of the form
u(x, t) =
n=1
a
n
sin
_
nx
L
_
sin(2
n
t) , (8.73)
where the allowed frequencies
n
are related to the length of the string L and the speed of waves
v by
n
=
nv
2L
. (8.74)
The corresponding wavelengths
n
for the vibrations are such that an integer number of half-
waves can t in the length of the string, i.e.,
n
/2 = L/n.
That is, for a vibrating string, the sound one hears is a combination of the fundamental
tone n = 1 and the higher harmonics with n = 2, 3, . . .. Depending on how the string is struck,
dierent combinations of harmonics are present. The frequencies of the overtones are always an
integer multiple of the fundamental frequency, and this is presumably why our brains perceive
a piano string as sounding harmonious.
For the drum, however, the wavelengths of the dierent eigenmodes are not related by ratios
of small integers, but are determined by the irregular spacing of the zeros of the Bessel functions.
This is the reason why a drum does not play clear musical notes in the same way that a piano
does.
The mixture of eigenmodes produced in a drum will depend on how it is struck. If, for
example, one hits the drum in the centre, then by symmetry the solution cannot depend on ,
so for the angular solution Q() one must have n = 0. The resulting modes will have wavelengths
corresponding to the zeros of J
0
, i.e., z
01
, z
02
, . . .. The eigenmodes for n = 0 and m = 1, 2, 3 are
shown in Fig. 8.4.
(a) (b) (c)
Figure 8.4: Eigenmodes of a circular drum corresponding to axially symmetric vibrations (n = 0) and
m = 1, 2, 3.
If, on the other hand, the drum is struck at some point between the centre and edge, then
this will excite eigenmodes corresponding to nonzero values of n. Some eigenmodes for nonzero
n are shown in Fig. 8.5.
Good timpanists all know (see Ref. [11] and references therein) that hitting a drum in
the centre will produce a hollow sound because this only excites axially symmetric (n = 0)
eigenmodes. The best place to strike a timpani is roughly 1/6 of the way across its diameter,
which excites mainly the n = 1, m = 1 mode.
(a) (b) (c)
Figure 8.5: Eigenmodes of a circular drum corresponding to m = 1 and n = 1, 2, 3.
8.1.8 Orthogonality of Bessel functions
From the standpoint of a musician it may be sucient to know what eigenmodes can be present.
But by specifying initial conditions, Eqs. (8.6) and (8.7), we can determine the coecients a
nm
in Eq. (8.70), which will tell us the drums exact motion as a function of time.
Let us suppose that the drums membrane is initially at rest,
u
t
(r, , 0) = 0 , (8.75)
and that its initial position is given by a specied function of r and :
u(r, , 0) = f(r, ) . (8.76)
We have already seen that Bessels equation is a special case of the Sturm-Liouville
equation, and therefore we know that the the solutions corresponding to dierent eigenvalues
are orthogonal. That is, for any given order of the Bessel functions n and for any two distinct
zeros z
nl
and z
nm
, we must have
_
a
0
J
n
(z
nl
r/a)J
n
(z
nm
r/a)r dr = 0 (l = m) . (8.77)
Remember that here the inner product is dened using the weight function w(r) = r and that
the value of n (the order of the Bessel function) is the same for both terms in (8.77). It is the
eigenvalues of the corresponding Sturm-Liouville operator,
nm
= z
nm
/a and
nl
= z
nl
/a, that
are dierent.
For the case l = m, i.e., where the two zeros of J
n
are equal, the integral (8.77) is nonzero.
We will not derive this explicitly but simply state the result. By using the Kronecker delta one
can express the result for both cases, l = m and l = m as
_
a
0
J
n
(z
nl
r/a)J
n
(z
nm
r/a)r dr =
lm
a
2
2
J
2
n+1
(z
nm
) . (8.78)
We can use this orthogonality relation for Bessel functions together with the corresponding
formulae for sines and cosines to solve for the coecients a
nm
and b
nm
, and this will then
determine the full solution (8.70). For this we follow the same basic procedure as in previous
problems such as the heated disc in Sec. 6.2.2. We had already eectively satised the initial
condition (8.75) by only including a cos(vz
nm
at) term in the solution (8.70). This corresponds
to sine terms in the derivative u/t, so we have zero speed at t = 0.
We now need to impose the initial condition (8.76) on our general solution (8.70), i.e.,
f(r, ) = u(r, , 0) =
n=0
m=1
J
n
(z
nm
r/a) [a
nm
cos(n) +b
nm
sin(n)] . (8.79)
The right-hand side involves two dierent types of orthogonal functions: the trigonometric
functions cos(n) and sin(n) and the Bessel functions J
n
(z
nm
r/a). For the Bessel functions we
have the orthogonality relation given by Eq. (8.78) and for the sines and cosines recall that we
have
_
2
0
cos(n) cos(l) d = N
l
nl
, (8.80)
_
2
0
sin(n) sin(l) d = M
l
nl
, (8.81)
_
2
0
cos(n) sin(l) d = 0 , (all n, l) , (8.82)
where N
0
= 2, M
0
= 0, and N
l
= M
l
= 1 for l = 1, 2, . . ..
We can solve for the coecients by using essentially the same approach as with the Fourier
series in Sec. 4.1. First we can multiply both sides of Eq. (8.79) by cos(l) and integrate from
0 to 2, which gives
a
l
(r)
_
2
0
f(r, ) cos(l) d (8.83)
=
n=0
m=1
J
n
_
z
nm
r
a
__
a
nm
_
2
0
cos(n) cos(l) d +b
nm
_
2
0
sin(n) cos(l) d
_
.
The rst integral gives us N
l
nl
and the second one is zero for all l. We can then carry out the
sum over n which gives
a
l
(r) = N
l
m=1
a
lm
J
l
_
z
lm
r
a
_
(8.84)
By multiplying both sides of Eq. (8.79) by sin(l) and integrating we nd in a similar way
b
l
(r)
_
2
0
f(r, ) sin(l) d =
m=1
b
lm
J
l
_
z
lm
r
a
_
(8.85)
for l = 1, 2, . . . and zero for l = 0.
Having dealt with the sine and cosine functions we can now carry out the corresponding
procedure for the Bessel function by multiplying both sides of Eqs. (8.84) and (8.85) by J
l
(z
lk
r/a)
times the weight function r and integrating from 0 to a. This gives,
_
a
0
a
l
(r)J
l
_
z
lk
r
a
_
r dr = N
l
m=1
a
lm
_
a
0
J
l
_
z
lm
r
a
_
J
l
_
z
lk
r
a
_
r dr . (8.86)
For the integral on the right-hand side we can use the orthogonality relation for Bessel functions
from Eq. (8.78), which will give a Kronecker delta
km
. This allows us to carry out the sum over
m, which is nonzero only for m = k. We can then solve for a
lk
,
a
lk
=
1
N
l
2
a
2
1
J
2
l+1
(z
lk
)
_
a
0
a
l
(r)J
l
_
z
lk
r
a
_
r dr , l = 0, 1, . . . , (8.87)
where recall N
0
= 2 and N
l
= 1 for l = 1, 2, . . .. In a similar way we nd
b
lk
=
2
a
2
1
J
2
l+1
(z
lk
)
_
a
0
b
l
(r)J
l
_
z
lk
r
a
_
r dr , l = 1, 2, . . . . (8.88)
The coecients from Eqs. (8.87) and (8.88) along with Eqs. (8.84) and (8.85) for a
l
(r) and
b
l
(r), with l and k relabeled back to n and m, can then be used together with Eq. (8.70) to
provide the full solution for the motion of the drum.
8.1.9 Fourier-Bessel series
The Bessel functions J
n
(z
nm
r/a) and Y
n
(k
nm
r/a) for m = 1, 2, . . ., where z
nm
and k
nm
are the
mth zeros of J
n
and k
nm
, respectively, constitute a complete set of functions on the interval
0 r a for any order n. This fact allows us to express any function dened on [0, a] as a
linear combination
f(r) =
m=1
[a
m
J
n
(z
nm
r/a) +b
m
Y
n
(k
nm
r/a)] . (8.89)
This is analogous to the Fourier series we saw in Ch. 4 and is called a Fourier-Bessel series.
If the function is nite on the interval 0 r a then all of the b
m
must be zero, since the
functions Y
n
diverge for r 0. We are usually dealing with nite functions and so the Y
n
are
rarely used.
The coecients are found as usual by taking the inner product of both sides with the
orthogonal basis function J
n
(z
nl
r/a) (with the weight function w(r) = r),
_
a
0
f(r)J
n
(z
nl
r/a)r dr =
m=1
_
a
0
a
m
J
n
(z
nl
r/a)J
n
(z
nm
r/a)r dr
=
m=1
lm
a
2
2
J
2
n+1
(z
nm
)
= a
l
a
2
2
J
2
n+1
(z
nl
) , (8.90)
where we have used the orthogonality relation (8.78). Solving this for the coecient a
l
and
relabeling gives
a
m
=
_
a
0
f(r)J
n
(z
nm
r/a)r dr
a
2
2
J
2
n+1
(z
nm
)
. (8.91)
In general the integrals (8.91) must be done numerically.
As an example consider the sawtooth function
f(x) = x 0 x 1 . (8.92)
We can expand this in a Fourier-Bessel series using the functions J
0
(we could use any order). To
compute the coecients (8.91) we can calculate the integral in the numerator numerically using,
say, the trapezium rule. In fact many software packages such as the GNU scientic library [9]
include better methods such as Gaussian quadrature (see, e.g., [12]). This was used to compute
the coecients shown in Table 8.2. The resulting series based on sums of the rst N terms for
N = 1, 3, 10, 20, 50 and 100 are shown in Fig. 8.6.
Table 8.2: Values of the rst 30 coecients a
m
for the Fourier-Bessel expansion of the sawtooth function
(8.92) (Bessel functions from the GSL routine gsl sf bessel Jn; integrals computed numerically with
gsl integration qags [9]).
m a
m
m a
m
m a
m
1 0.817455 11 0.428153 21 0.309639
2 -1.13349 12 -0.414533 22 -0.303839
3 0.798285 13 0.393824 23 0.295819
4 -0.747008 14 -0.382839 24 -0.290696
5 0.631543 15 0.366574 25 0.283695
6 -0.597359 16 -0.357468 26 -0.279125
7 0.535993 17 0.344273 27 0.272946
8 -0.512335 18 -0.336561 28 -0.268836
9 0.473225 19 0.325587 29 0.263331
10 -0.455716 20 -0.318946 30 -0.259608
Notice from Table 8.2 that the coecients decrease in magnitude very slowly, having dropped
by only about a factor of four after 30 terms. Although one can see in Fig. 8.6 that the series
is clearly converging to the desired function, noticeable dierences are present even after 100
terms. A clear overshoot is also visible at x = 1, which is the analogue of the Gibbs phenomenon
that we saw in the Fourier series of a square wave in Fig. 4.1.
We have now seen a large number of ways of expressing a function f(r) on an interval
0 r a. We can expand in terms of sin(nr/a), cos(nr/a), J
n
(z
nm
r/a), or simply in powers
of r. Which do we choose? And if we choose the Bessel functions, which order n do we use?
Although we are guaranteed in all cases the series will converge (in the mean) to the desired
function, there can be substantial dierences between the dierent choices in terms of the rate of
convergence. For example, if we had simply expanding the sawtooth function (8.92) in a power
series in x,
f(x) =
n=0
a
n
x
n
, (8.93)
then we would nd the series would be exact with only a single nonzero coecient, a
1
= 1. So if
we had reason to believe that the function we are trying to represent is close to linear in x, then
a power series in x is likely to provide a far more accurate representation for a given number of
terms than expansion in Bessel functions.
Furthermore, some choices may be particularly convenient in that certain boundary or
periodicity requirements are automatically satised. As an example consider a function not
only of r but dened on a circular disc, i.e, a function f(r, ) with 0 r a and 0 < 2.
The original function will satisfy
f(r, + 2) = f(r, ) , (8.94)
since and +2 refer to to the same place on the circle. Furthermore the functions one often
considers are continuous at the origin, so that two points at some small radius but on opposite
sides of the origin, and + , should be close in value, and the function and its derivatives
should be continuous as one passes from one such point to the other.
The choice of basis function for an expansion is often guided by some physical insight into
what the function is supposed to represent. Suppose we have a model for the membrane of a
drum which is not quite the wave equation we have seen above, but perhaps something more
sophisticated that allows for energy dissipation or slight variations in the tension of the surface.
The motion of the surface will therefore not be exactly described by the solution that we have
derived above, but it may be close if the departure from the idealised drum is small.
In such a case one might try to express the spatial conguration of the drum f(r, ) in a
expansion similar to the one we saw in Eq. (8.79). That is, we can rst use a Fourier series for
the dependence, where the coecients will depend in general on r and t,
f(r, , t) =
n=
a
n
(r, t)e
in
. (8.95)
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 1
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 3
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 10
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 20
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 50
x
0 1
f
(
x
)
0
0.5
1
1.5
N = 100
(a) (b)
(c)
(d)
(e) (f)
Figure 8.6: Fourier-Bessel expansion for a sawtooth function (8.92) based on the Bessel function J
0
for
a partial sum of terms up to the value of N indicated.
Here n is an integer and we have used the exponential form of the Fourier series (we could also
use the sine and cosine version). The coecients a
n
(r) are found as usual (see Sec. 4.3) by
a
n
(r, t) =
1
2
_
2
0
f(r, , t)e
in
d . (8.96)
The coecients a
n
(r, t) are all functions of r and t, and these can be expanded in Bessel
functions of the corresponding order n with the time dependence contained in the coecients:
a
n
(r, t) =
m=1
b
nm
(t)J
n
(z
nm
r/a) . (8.97)
The coecients b
nm
(t) are found from
b
nm
(t) =
_
a
0
a
n
(r, t)J
n
(z
nm
r/a)r dr
a
2
2
J
2
n+1
(z
nm
)
. (8.98)
If we only include a nite number of terms in the expansion, then we are still guaranteed that
it will be periodic in and continuous at the origin.
From a mathematical standpoint the coecients b
nm
(t) could be any functions of time, but
from a physical perspective we may believe the surface will be similar to an actual drum, and so
the series may provide a good approximation with a relatively small number of terms and the
b
nm
for higher n and m can be neglected. Furthermore a more accurate model for a drum that
includes complications such as energy dissipation, etc., will result in only a relatively slow and
simple time variation of the b
nm
(t) to the extent that the complications are themselves small
eects.
8.2 Legendre polynomials and spherical harmonics
In this section we will look at solutions to the three-dimensional Laplace equation in spherical
polar coordinates. Following the by now well established pattern we will carry out separation of
variables, which will result in ordinary dierential equations for r, and . We will be interested
in particular in the angular part of the solution, which will lead to functions called Legendre
polynomials, associated Legendre polynomials and spherical harmonics.
8.2.1 Separating the Laplace equation in spherical polar coordinates
Consider a function u that is a solution to the Laplace equation. This could represent the
electrostatic potential in a charge free region or a solution to the steady-state heat equation.
Suppose the boundary conditions of the problem have a spherical symmetry so that we choose to
separate variables in spherical polar coordinates. The Laplace equation is then (see Sec. 6.3.2)
2
u =
1
r
2
sin
_
sin

r
_
r
2
u
r
_
+

_
sin
u
_
+
1
sin
2
u
2
_
= 0 . (8.99)
We seek product solutions of the form
(r, , ) = R(r)P()() . (8.100)
Substituting this into the Laplace equation, multiplying by both sides by sin
2
and dividing by
RP gives
sin
2
R
d
dr
_
r
2
dR
dr
_
+
sin
P
d
d
_
sin
dP
d
_
+
1
d
2
d
2
= 0 . (8.101)
In this way we have succeeded in isolating the dependence in the third term, which therefore
must be equal to a constant . As we have seen in previous examples (see, e.g., Sec. 6.2.2) the
periodicity requirement (+2) = () means we must have = m
2
where m is an integer.
The general solution for is then
() = Acos(m) +Bsin(m) , (8.102)
or if we regard the solution as complex, we can write (with redenition of the constants A and
B)
() = Ae
im
+Be
im
. (8.103)
We will return in Sec. 8.2.5 to the general solution to the problem, but for the moment let
us suppose that the system is symmetric in , i.e., () is constant so we must have m = 0. In
this case we can separate the terms that depend on r and by dividing Eq. (8.101) by sin
2
,
which gives
1
R
d
dr
_
r
2
dR
dr
_
+
1
P sin
d
d
_
sin
dP
d
_
= 0 . (8.104)
The r dependence is all in the rst term and the dependence in the second, and the two sum
to zero. So the terms are equal and opposite constants, i.e.,
1
R
d
dr
_
r
2
dR
dr
_
= , (8.105)
1
P sin
d
d
_
sin
dP
d
_
= . (8.106)
8.2.2 Legendres equation
We are mainly interested in the solution to the angular equation (8.106), which we can write
d
d
_
sin
dP
d
_
= P sin . (8.107)
In spherical polar coordinates the angle takes on values in the range 0 . To solve the
equation it is convenient to express the angle using the equivalent variable
x = cos , (8.108)
which then takes on values in 1 x 1. To convert the derivatives with respect to to the
new variable we use the chain rule,
d
d
=
dx
d
d
dx
= sin
d
dx
. (8.109)
We therefore have
dP
d
= sin
dP
dx
, (8.110)
and the left-hand side of (8.107) can be written
d
d
_
sin
dP
d
_
= sin
d
dx
_
sin
_
sin
dP
dx
__
. (8.111)
Using this in Eq. (8.107) and cancelling a factor of sin from both sides gives
d
dx
_
sin
2
dP
dx
_
= P . (8.112)
By using sin
2
= 1 cos
2
= 1 x
2
we therefore have
d
dx
_
(1 x
2
)
dP
dx
_
= P , (8.113)
which is called Legendres equation.
By comparison of Eq. (8.113) with Eq. (5.2), we see that Legendres equation is a special
case of the Sturm-Liouville equation with
p(x) = 1 x
2
, (8.114)
q(x) = 0 , (8.115)
w(x) = 1 , (8.116)
and furthermore the separation constant is the eigenvalue of the SL operator. This tells us
that the solutions corresponding to dierent eigenvalues will be orthogonal and that they will
constitute a complete set of eigenfunctions.
8.2.3 Legendre functions
To solve Legendres equation (8.113) we can rst use the product rule for the term in square
brackets to write
(1 x
2
)
d
2
P
dx
2
2x
dP
dx
+P = 0 . (8.117)
If we divide through by 1 x
2
to put the equation in standard form we see that x = 1 are
regular singular points but x = 0 is an ordinary point. We therefore try a series solution about
x = 0 of the form
P(x) =
n=0
a
n
x
n
. (8.118)
Computing the required derivatives and substituting into (8.117) gives
(1 x
2
)
n=0
a
n
n(n 1)x
n2
2
n=0
a
n
nx
n
+
n=0
a
n
x
n
= 0 . (8.119)
Note that in the rst sum the terms for n = 0 and 1 are in fact zero as is the n = 0 term in the
second sum. Using the standard technique of replacing n 2 by a new index m in the rst sum
and then relabeling m back to n, Eq. (8.119) becomes
n=0
a
n+2
(n + 2)(n + 1)x
n
n=2
a
n
n(n 1)x
n
2
n=1
a
n
nx
n
+
n=0
a
n
x
n
= 0 . (8.120)
Equating the sum of coecients of x
n
to zero gives
a
n+2
(n + 2)(n + 1) a
n
n(n 1) 2a
n
n +a
n
= 0 , (8.121)
which can be rewritten as the recurrence relation
a
n+2
=
n(n + 1)
(n + 1)(n + 2)
a
n
. (8.122)
As one expects for a second order dierential equation, there are two arbitrary constants which
we can take as a
0
and a
1
. The recurrence relation (8.122) then connects the even-numbered
ones to a
0
and the odd-numbered ones to a
1
. We can therefore construct two independent series
solutions, one where we take a
0
= 0 and a
1
= 0, which then contains only even powers of x, and
one with a
0
= 0 and a
1
= 0 that contains only odd powers of x.
One may now ask whether these series will converge. In the limit of large n, the ratio of
successive coecients becomes
lim
n
a
n+2
a
n
= lim
n
n(n + 1)
(n + 1)(n + 2)
= 1 , (8.123)
so by the ratio test (see, e.g., Ref. [2]) the series will converge for |x| < 1.
But we are also interested in the points x = 1 since these correspond to the physically
accessible angles = 0 and = , and here the ratio test is inconclusive. One can in fact show
using other convergence tests such as the integral test (see, e.g., Refs. [7, 13]) that in general the
series will diverge for x = 1. The only way that it can converge is if the coecients take on
the value of zero for some nite n, because then all coecients for higher n are necessarily zero
owing to the recurrence relation (8.122). And the only way to arrange for one of the coecients
to become zero is if the eigenvalue satises
n(n + 1) = , (8.124)
where n is the (integer) order of the term in the series expansion. If this holds, then the series
truncates after reaching the nth term and then it obviously converges.
Following convention we will write the eigenvalue as l(l + 1) where l = 0, 1, 2, . . ., and we
will write the solution P with a subscript l to label the eigenvalue to which it corresponds, and
which also gives therefore the highest power of x in the series before it cuts o. We can then
set the arbitrary constants a
0
and a
1
such that P
l
(1) = 1 for all l, and such that a
1
= 0 for
l even and a
0
= 0 for l odd. In this way we obtain the family of Legendre polynomials P
l
(x),
with l = 0, 1, 2, . . .. These are also sometimes called the Legendre functions of the rst kind. For
l = 0 to 5 they are
P
0
(x) = 1 , (8.125)
P
1
(x) = x , (8.126)
P
2
(x) =
1
2
(3x
2
1) , (8.127)
P
3
(x) =
1
2
(5x
3
3x) , (8.128)
P
4
(x) =
1
8
(35x
4
30x
2
+ 3) , (8.129)
P
5
(x) =
1
8
(63x
5
70x
3
+ 15x) . (8.130)
For a given value of = l(l + 1), the Legendre polynomial P
l
(x) constitutes only one of the
two linearly independent solutions to Legendres equation. This is analogous to the situation
with Bessels equation where the Frobenius method only led to the solution J
n
, but not Y
n
.
As with Bessels equation, one can use reduction of order to nd the second solution. These
are called Legendre functions of the second kind and denoted Q
l
(x). They are related to the
Legendre polynomials (i.e., the Legendre functions of the rst kind), by
Q
l
(x) =
1
2
P
l
(x) ln
_
1 +x
1 x
_
. (8.131)
Figure 8.7 (a) and (b) show the rst few Legendre functions of the rst and second kind,
respectively. Notice that the P
l
(x) are all nite, even at x = 1, whereas the Q
l
(x) all diverge
at x = 1.
x
-1 0 1
(
x
)
l
P
0
P
1
P
2
P
3
P
-1
1
x
-1 0 1
(
x
)
l
Q
0
Q
1
Q
2
Q
3
Q
-1
1
(a) (b)
Figure 8.7: Plots of Legendre functions of (a) the rst kind P
l
(x) and (b) the second kind Q
l
(x) for
l = 0 to 3.
8.2.4 Expansion in Legendre polynomials
Because the Legendre polynomials are solutions to a Sturm-Liouville equation we know that the
functions corresponding to dierent eigenvalues
a
= a(a +1) and
b
= b(b +1) are orthogonal,
i.e.,
P
a
(x), P
b
(x) =
_
1
1
P
a
(x)P
b
(x) dx = 0 (a = b) , (8.132)
and the inner product for Legendre polynomials of equal order l is found to be
P
l
(x), P
l
(x) =
_
1
1
[P
l
(x)]
2
dx =
2
2l + 1
. (8.133)
Together the Legendre functions P
l
(x) and Q
l
(x) constitute a complete set of functions in
the interval 1 x 1, and any function f(x) in this interval can be expressed as a linear
combination of the form
f(x) =
l=0
[a
l
P
l
(x) +b
l
Q
l
(x)] . (8.134)
If the function is nite in 1 x 1, then all of the coecients b
l
must be zero, since the
Q
l
(x) diverge at x = 1. The functions we usually consider are nite and so one typically deals
with an expansion in Legendre polynomials P
l
(x); the Q
l
(x) do not enter. The coecients a
l
are found as usual by taking the inner product of P
l
(x) and f(x) (see Sec. 3.3):
a
l
=
P
l
, f
P
l
2
=
2l + 1
2
_
1
1
P
l
(x)f(x) dx . (8.135)
8.2.5 Associated Legendre polynomials
We can now return to the Laplace equation and consider the case where the solution is not
necessarily symmetric in . Then the separation constant m that appears in Eq. (8.103) for the
solution () can take on a nonzero integer value. Consider again the Laplace equation (8.101)
into which we had substituted the product solution RP. If we substitute the separation
constant m
2
for the third term
/ and divide through by sin

2
this becomes
1
R
d
dr
_
r
2
dR
dr
_
+
1
sinP
d
d
_
sin
dP
d
_
m
2
sin
2
= 0 . (8.136)
As in the m = 0 case, the r dependence is all contained in the rst term (which is the same
as before) and the dependence all in the second and third. So the rst term must be equal to
a separation constant and the second and third equal to . In essentially the same manner
as for the m = 0 case one can show that nite solutions exist for 1 x 1 only if is equal
to l(l + 1) where l is a nonnegative integer.
If we then use the same change of variable as before, x = cos , we can rewrite the -dependent
part of the equation as
(1 x
2
)
d
2
P
dx
2
2x
dP
dx
+
_
l(l + 1)
m
2
1 x
2
_
P = 0 , (8.137)
where we have used l(l + 1) for the separation constant . This is called Legendres associated
equation. It reduces to Legendres equation for m = 0. It can be rearranged into the equivalent
form
d
dx
_
(1 x
2
)
dP
dx
_
+
m
2
1 x
2
P = l(l + 1)P , (8.138)
from which we see that it is also a special case of the Sturm-Liouville equation Eq. (5.2) with
eigenvalue = l(l + 1) and
p(x) = 1 x
2
, (8.139)
q(x) =
m
2
1 x
2
, (8.140)
w(x) = 1 . (8.141)
Therefore the solutions have the usual properties of orthogonality, completeness, etc., as we will
see below.
One can show that two linearly independent solutions to Legendres associated equation for
m 0 can be written
P
m
l
(x) = (1 x
2
)
m/2
d
m
dx
m
P
l
(x) , (8.142)
Q
m
l
(x) = (1 x
2
)
m/2
d
m
dx
m
Q
l
(x) . (8.143)
The functions P
m
l
(x) and Q
m
l
(x) are called associated Legendre functions of the rst and second
kind, respectively. Like the Q
l
, the associated versions Q
m
l
diverge at x = 1. So if the boundary
conditions of the problem require a nite solution, as is usually the case, then the Q
m
l
will not
appear. The (associated) Legendre functions of the second kind are therefore rarely seen in
practice and we will not need them further in this course.
Notice that the function P
m
l
is obtained by dierentiating P
l
m times. But the highest power
of x in P
l
is x
l
, so if we dierentiate l + 1 times or more, we get zero. Therefore we only nd
nonzero solutions for m l.
Solutions with negative m for the P
m
l
also exist. One can show (see, e.g., Ref. [13])
P
m
l
(x) = (1)
m
(l m)!
(l +m)!
P
m
l
(x) . (8.144)
Since P
m
l
and P
m
l
are proportional to each other one must therefore have l m l.
As the associated Legendre equation is a special case of the Sturm-Liouville equation, we
know that solutions corresponding to distinct eigenvalues are orthogonal. One nds
_
1
1
P
m
k
(x)P
m
l
(x) dx =
2
2l + 1
(l +m)!
(l m)!

kl
. (8.145)
Finally we give a useful relation called Rodriguess formula (for a derivation see, e.g.,
Ref. [13]),
P
l
(x) =
1
2
l
l!
d
l
dx
l
(x
2
1)
l
. (8.146)
If this is combined with Eq. (8.142) for the associated Legendre functions then one nds
P
m
l
(x) =
(1)
m
2
l
l!
(1 x
2
)
m/2
d
l+m
dx
l+m
(x
2
1)
l
, (8.147)
From here one sees directly how negative values of m can be used as long as m l.
8.2.6 Spherical harmonics
The associated Legendre polynomials P
m
l
(x) with x = cos give the dependence of the product
solution to Laplaces equation, but if m = 0 then we know from Eq. (8.103) that the solution
depends on as well. These parts of the solution are often combined into a family of functions
of and called spherical harmonics, dened as
Y
lm
(, ) =
(2l + 1)
4
(l m)!
(l +m)!
P
m
l
(cos )e
im
, l 0, l m l . (8.148)
Several dierent conventions for the normalization of spherical harmonics can be found in the
literature. The one used above is common in physics and is chosen that so that the Y
lm
form
an orthonormal set. That is, the orthogonality relation for spherical harmonics is
Y
lm
, Y
kn
=
_
2
0
_

0
Y
lm
(, )Y
kn
(, ) sin d d =
kl
mn
. (8.149)
The factor of sin in (8.149) comes from the fact that x = cos and therefore dx = d cos =
sind. That is, when integrating a function f over x = cos one has
_
1
1
f(x) dx =
_

0
f(cos ) sin d , (8.150)
since the lower limit of integration in the rst integral, x = 1, corresponds to the upper limit
= in the second and vice versa. So by arranging the order of the limits of integration as
done here we lose the minus sign and simply replace dx by sin d. Equivalently we can write
Eq. (8.149) as
_
Y
lm
(, )Y
kn
(, ) d =
kl
mn
. (8.151)
Here the solid angle element d is
d = sin d d , (8.152)
and an integral over the entire range of d implies 0 and 0 2.
Some of the spherical harmonics Y
lm
are
Y
0,0
(, ) =
1
4
, (8.153)
Y
1,1
(, ) =
_
3
8
sin e
i
, (8.154)
Y
1,0
(, ) =
_
3
4
cos , (8.155)
Y
1,1
(, ) =
_
3
8
sin e
i
, (8.156)
Y
2,2
(, ) =
_
15
32
sin
2
e
2i
, (8.157)
Y
2,1
(, ) =
_
15
8
sin cos e
i
, (8.158)
Y
2,0
(, ) =
_
5
16
(3 cos
2
1) , (8.159)
Y
2,1
(, ) =
_
15
8
sin cos e
i
, (8.160)
Y
2,2
(, ) =
_
15
32
sin
2
e
2i
. (8.161)
Plots of the absolute value of spherical harmonics for several values of l and m are shown in
Fig. 8.8.
8.2.7 Combining spherical harmonics with the radial equation
In the last section we found the angular part of the solution to the Laplace equation in spherical
polar coordinates, namely, the spherical harmonics Y
lm
(, ), and this was the main goal of this
chapter. But before leaving the Laplace equation we should for completeness nd the radial part
of the equation R(r) and write down the full solution to the original problem. From Eq. (8.105)
combined with the fact that the eigenvalue was found to be l(l+1), we can write the dierential
equation for R(r) as
r
2
R
+ 2rR
l(l + 1)R = 0 . (8.162)

We can recognize this as an Euler equation of the type we investigated in Sec. 7.2.2. We have
seen that the solution is of the form
R = r
. (8.163)
Substituting this into the dierential equation (8.162) leads to the indicial equation
(a) (b)
(c) (d)
(e) (f)
l = 0, m = 0 l = 1, m = 1
l = 1, m = 0 l = 2, m = 2
l = 2, m = 1 l = 2, m = 0
Figure 8.8: Plots of the absolute values of the spherical harmonics Y
lm
(, ) for several values of l and
m. The distance of the surface from the origin gives the value of |Y
lm
(, )|.
( + 1) l(l + 1) = 0 , (8.164)
which has solutions = l and = (l + 1). So we can write the general solution as
R(r) = Ar
l
+Br
(l+1)
. (8.165)
The general solution to the Laplace equation is obtained by combining this with the angular
part Y
lm
(, ),
u(r, , ) =
l=0
l
m=l
_
A
lm
r
l
+B
lm
r
(l+1)
_
Y
lm
(, ) . (8.166)
The coecients A
lm
and B
lm
are determined as usual by imposing the boundary conditions.
Often one is interested in the region to the exterior of some surface, in which case one of the
boundary conditions will be that the solution goes to zero in the limit r , which means that
all the A
lm
= 0 (except possibly A
00
, which gives a constant solution). In other problems one
may be interested in an interior region, in which case the solution should be nite for r 0, so
one has B
lm
= 0. If the region of interest excludes both innity and the origin then both the
A
lm
and B
lm
may be nonzero.
In any case one then nds the nonzero coecients by exploiting the orthogonality of the
spherical harmonics. For example, if we want the solution to the interior of a surface of a sphere
of radius a on which we have the boundary condition
u(a, , ) = f(, ) , (8.167)
then the coecients A
lm
are found by exploiting the orthogonality of the spherical harmonics.
Assuming the B
lm
are zero (for a nite interior solution), evaluating Eq. (8.166) at r = a and
taking the inner product of both sides with Y
kn
(, ) gives
Y
kn
, f =
l=0
l
m=l
A
lm
a
l
Y
kn
, Y
lm
l=0
l
m=l
A
lm
a
l
kl
mn
= A
kn
a
k
. (8.168)
Note that the usual factor Y
lm
2
is not needed because the spherical harmonics are orthonormal,
i.e., Y
lm
, Y
lm
= 1. Solving for A
kn
and relabeling indices gives
A
lm
=
Y
lm
, f
a
l
=
1
a
l
_
2
0
_

0
Y
lm
(, )f(, ) sin d d . (8.169)
The coecients A
lm
together with Eq. (8.166) constitute the nal answer to the problem.
8.2.8 Using spherical harmonics
You will have plenty of opportunities to use spherical harmonics in your electromagnetism and
quantum mechanics courses, solving for electrostatic potentials and nding wave functions. Here
we will show a somewhat dierent use of spherical harmonics to represent the temperature of
the cosmic microwave background (CMB) as a function of direction in space. This example does
not rely on the spherical harmonics representing the solution to a physically relevant dierential
equation, but rather only exploits the fact that they constitute a complete set of functions with
which we can expand any function of and .
The wavelength spectrum of microwave radiation from space is found to be an almost exact
blackbody distribution at a temperature close to T = 2.725 K. But the temperature is found to
vary slightly as a function of direction, and these variations can be related to density uctuations
in the very early universe that eventually grew into galaxies.
The space-based Wilkinson Microwave Anisotropy Probe (WMAP) has measured the CMB
temperature as a function of direction in the sky and in 2003 published the temperature map in
Fig. 8.9 [14]. Each point on the map corresponds to a direction in space with the plane of the
galaxy running horizontally through the middle. The colours on the map indicate the deviations
from the mean temperature, which are at the level of 1 part in 10
5
.
Figure 8.9: CMB temperature
map of the universe from WMAP
[14] (see text).
As the temperature map T(, ) represents a function of direction in space, it can be
expanded in spherical harmonics as
T(, ) =
l=0
l
m=l
a
lm
Y
lm
(, ) . (8.170)
Terms with higher l correspond to ner angular separation. The angular resolution of the
measurement was limited to a few tenths of a degree, so the sum only includes terms up to
around 1000.
Dierent models of the early universe predict dierent amounts of variation between hot
and cold at characteristic angular separations. To express how much structure there is as a
function of angle one can compute the angular power spectrum. This is a set of numbers c
l
with
l = 0, 1, 2, . . ., dened as
c
l
=
1
2l + 1
l
m=l
|a
lm
|
2
. (8.171)
Since the Legendre polynomial of order l has l zeros in the region from 1 cos 1 (i.e.,
0 ), the value of c
l
is a measure of the amount of angular structure at a separation of

180
l
. (8.172)
Measures of the angular power spectrum of the CMB from WMAP are shown in Fig. 8.10. The
large peak at l 200 corresponds to an angular separation of somewhat less than a degree.
Figure 8.10: Angular power
spectrum of the CMB from
WMAP [15] (see text).
The curve in Fig. 8.10 drawn through the points is the prediction of what is called the CDM
cosmological model. By adjusting the parameters of the model so that the curve passes through
the measurements, one is able to determine a number of quantities of cosmological interest, such
as the density of the universe, the Hubble constant, the fraction of ordinary (baryonic) matter,
dark matter, dark energy, etc. (for more details see Refs. [14, 15, 16]).
8.2.9 Spherical harmonics in quantum mechanics
Before leaving the topic of spherical harmonics we should note an important reason why they
come up in so many problems, particularly in quantum mechanics. We saw that they emerged as
the angular part of the solution to the Laplace equation when it was separated in spherical polar
coordinates. But suppose our equation is not the Laplace equation but rather the eigenvalue
problem
2
u +f(r)u = u , (8.173)
where f(r) is an arbitrary function of the radial coordinate r, but is independent of the angles
and . This is the general form of the Schrodinger equation for a central potential, i.e., one
that only depends on r.
We can use separation of variables in this case, just as with the Laplace equation, and seek
product solutions of the form
(r, , ) = R(r)P()() . (8.174)
It will be left as an exercise to show that for Eq. (8.173), the angular part of the solution,
P()(), is exactly the same as in the case of the Laplace equation, namely, the spherical
harmonic functions Y
lm
(, ). The radial equation and its solution will depend on f(r), and
this is in general dierent from what was found for the Laplace equation. But many important
problems, for example the Schrodinger equation for an electron in an atom, are of the form of
Eq. (8.173), and so the spherical harmonics gure prominently in atomic physics where they
give the angular part of the wave function for dierent values of the quantum numbers l and m.
8.3 Hermite polynomials
One of the most important problems in quantum mechanics is the harmonic oscillator. In this
section we will write down and solve the Schrodinger equation for this system, which, after some
manipulation will turn out to be yet another special case of the Sturm-Liouville problem. We
will nd solutions using the Frobenius method, which will lead to a family of functions called
Hermite (pronunciation: air-meet) polynomials.
8.3.1 The quantum harmonic oscillator and Hermites equation
The time-independent Schrodinger equation for a one-dimensional quantum harmonic oscillator
can be written
_
h
2
2m
d
2
dx
2
+
1
2
m
2
x
2
_
(x) = E(x) (8.175)
As is often the case in physical problems it is useful to transform the equation into a dimensionless
form. We will use a shorthand notation using square brackets to give the dimension of a quantity
labeled by the corresponding SI units:
[x] = m , (8.176)
[m] = kg , (8.177)
[h] = J s = kg m
2
s
1
, (8.178)
[] = s
1
. (8.179)
We would like to construct a new dimensionless variable y of the form
y = x/a . (8.180)
This means the proportionality constant a should have units of metres. Looking at the units
of the available quantities, we see that only h has anything related to length, namely, metres
squared, so a must be proportional to
h. We then need to bring in other quantities to cancel

the other units that entered through h, which leads to
a =
h(kg
1/2
ms
1/2
)
(s
1/2
)
m(kg
1/2
)
. (8.181)
All of the units on the right-hand side cancel except metres in the numerator, so we can therefore
dene the dimensionless variable
y =
x
a
=
_
m
h
x . (8.182)
We now need to write the solution to the Schrodinger equation (x) in terms of the new variable
y and so we dene u(y) such that
(x) = u(y(x)) . (8.183)
That is, (x) is not the same function of x as u is of y, but the value of (x) is the same as
the value of u(y) if we use the y corresponding to y(x) = x/a.
2
With these substitutions
Eq. (8.175) becomes
u
+ ( y
2
)u = 0 , (8.184)
where we have dened
=
2E
h
. (8.185)
We can rearrange (8.184) as
u
+y
2
u = u , (8.186)
Comparing with (5.2) we see this is the Sturm-Liouville equation with
p(y) = 1 , (8.187)
q(y) = y
2
, (8.188)
w(y) = 1 . (8.189)
Therefore we know that the solutions v(y), and hence also the wavefunctions (x), as well
as the related functions u(y), have the usual properties as solutions to the SL equation (real
eigenvalues, orthogonality, completeness, etc.).
2
Sometimes we are not so rigorous in the notation. We could have simply written (y) and used the argument
y as a way to identify which function we are talking about. We did this, for example, in Sec. 8.2.7 with R(r) and
R(z). The notation with u(y) is more correct.
As Eq. (8.184) is already in the standard form of Eq. (7.22) we immediately see that no
points in < y < are singular and so we can solve with the series method. It turns out
to be easier, however, to rst examine how we expect the solution to behave for y . As y
becomes very large one as y
2
and therefore
u
y
2
u . (8.190)
If we try
u = e
y
2
/2
, (8.191)
then we see that this almost works, since
u
= e
y
2
/2
(y
2
1) y
2
u , (8.192)
where the approximation holds for y 1. So if we write u(y) as
u(y) = e
y
2
/2
v(y) , (8.193)
then we expect the exponential term to at least give us roughly the right behaviour for y .
Substituting (8.193) into the dierential equation (8.184) gives
u
+ ( y
2
)u = e
y
2
/2
_
v
2yv
+ ( 1)v
= 0 , (8.194)
Canceling the exponential term, which can never be zero, we nd
v
2yv
+ ( 1)v = 0 , (8.195)
Equation (8.184) is called Hermites equation. The boundary condition is that the solution
u(y) must go to zero at y , which here will be sucient to ensure that the wave function
be normalizable.
8.3.2 Series solution for Hermites equation
We can solve Hermites equation with the Frobenius method by seeking a solution of the form
v(y) =
n=0
a
n
y
n+
. (8.196)
First we compute the required ingredients,
v
n=0
a
n
(n +)y
n+1
, (8.197)
v
n=0
a
n
(n +)(n + 1)y
n+2
=
n=2
a
n+2
(n + 2 +)(n + 1 +)y
n+
. (8.198)
Substituting these into Hermites equation (8.195) gives
a
0
( 1)y
2
2a
1
( + 1)y
1
+
n=0
[a
n+2
(n + 2 +)(n + 1 +)
2a
n
(n +) + ( 1)a
n
] y
n+
= 0 . (8.199)
Setting the coecients of powers of y to zero gives the three relations
a
0
( 1) = 0 , (8.200)
a
1
( + 1) = 0 , (8.201)
a
n+2
(n + 2 +)(n + 1 +) + [2(n +) + 1] a
n
= 0 . (8.202)
From the indicial equation (8.200) we nd = 0, in which case a
1
can be nonzero, or = 1
in which case a
1
must be zero (in any case we assume a
0
is nonzero). Taking = 0 gives from
(8.202) the recurrence relation
a
n+2
=
2n + 1
(n + 2)(n + 1)
a
n
. (8.203)
Because the recurrence relation connects the coecients a
n+2
and a
n
, one obtains two
independent series, one with even-numbered and one with odd-numbered coecients, and the
full series is the sum of the two.
The next step is to investigate under what circumstances the series converges, and to do this
we consider the even and odd-numbered parts seprately. The ratio of neighbouring coecients
for the even-numbered series is
a
n+2
a
n
=
2n + 1
(n + 2)(n + 1)
. (8.204)
For large n this becomes
a
n+2
a
n
2
n
. (8.205)
For the odd-numbered series one nds the the same ratio for large n, and this goes to zero in
the limit n . So from the ratio test we conclude that both the even- and odd-numbered
series both converge. But how does the resulting function behave for large y? We can write the
Taylor series for e
y
2
as
e
y
2
= 1 +y
2
+
y
4
2!
+
y
6
3!
+ =
n=0,2,4,...
y
n
(n/2)!
, (8.206)
that is,
e
y
2
=
n=0,2,4,...
c
n
y
n
(8.207)
with c
n
= (2/n)!. The ratio of neighbouring coecients is therefore for large n
c
n+2
c
n
2
n
, (8.208)
which is the same as what we found for the Frobenius series above. So for large y the behaviour
of the series is like e
y
2
, so for u(y) we therefore have
u(y) = e
y
2
/2
v(y) e
y
2
/2
. (8.209)
This means that the function u(y) will diverge for large y and we will not be able to satisfy the
boundary condition that the function go to zero for y . The only way to get u(y) to go
to zero is if the series for v(y) terminates after a nite number of terms. Looking back at the
recurrence relation (8.204) we see this will happen if the quantity takes on the value
= 2n + 1 . (8.210)
If (8.210) holds then all terms higher than n are zero and v(y) is a polynomial of order n. The
product of a polynomial and e
y
/
2
goes to zero for y and results in a square integrable
u(y). If we convert back into the energy E using Eq. (8.185) then we nd the well-known
formula for the quantum harmonic oscillator,
E = (n +
1
2
) h , n = 0, 1, 2, . . . . (8.211)
The solutions v(y) are called Hermite polynomials of order n, written H
n
(y), and are dened
by choosing the constants a
0
and a
1
so that H
0
(y) = 1 and H
1
(y) = 2y. Some of the Hermite
polynomials are
H
0
(y) = 1 , (8.212)
H
1
(y) = 2y , (8.213)
H
2
(y) = 4y
2
2 , (8.214)
H
3
(y) = 8y
3
12y , (8.215)
H
4
(y) = 16y
4
48y
2
+ 12 , (8.216)
H
5
(y) = 32y
5
160y
3
+ 120y . (8.217)
A plot of the Hermite polynomials for n = 0, 1, . . . , 3 is shown in Fig. 8.11.
x
-2 -1 0 1 2
(
x
)
n
H
-20
-10
0
10
20
0
H
1
H
2
H
3
H
Figure 8.11: Plot of Hermite
polynomials H
n
(x).
One can show that application of the recurrence relation (8.204) is equivalent to
H
n
(y) = (1)
n
e
y
2 d
n
dy
n
e
y
2
, (8.218)
which is the generalisation of Rodriguess formula (8.146) for Hermite polynomials.
8.3.3 Orthogonality of Hermite polynomials
Hermites equation (8.195) does not appear to be in the Sturm-Liouville form, but if we multiply
it by e
y
2
,
e
y
2
v
2ye
y
2
v
+ ( 1)e
y
2
v = 0 , (8.219)
then this can be rewritten as
d
dy
_
e
y
2 dv
dy
_
+e
y
2
v = e
y
2
v . (8.220)
Comparing with (5.2) we see this is the Sturm-Liouville equation with
p(y) = e
y
2
, (8.221)
q(y) = e
y
2
, (8.222)
w(y) = e
y
2
. (8.223)
Therefore we know that the Hermite polynomials, as well as the related functions u(y), have the
usual properties as solutions to the SL equation (real eigenvalues, orthogonality, completeness,
etc.). For the orthogonality relation one nds
H
n
, H
m
=
_

H
n
(y)H
m
(y)e
y
2
dy =
2
n
n!
nm
. (8.224)
Notice that here the inner product is dened using the weight function w(y) = e
y
2
.
Further properties of Hermite polynomials can be found in standard references such as [1, 2, 7]
or from the Wikipedia page [17].
8.3.4 Wavefunction of the quantum harmonic oscillator
Having solved the dierential equation for v(y), we therefore also nd u(y) = e
y
2
/2
v(y). We can
convert this back into the wavefunction (x) for the harmonic oscillator by using Eq. (8.182),
i.e., y = x/a. Because the dierential equation for u(y) (8.186) was a special case of the Sturm-
Liouville equation (5.2) we know that the eigenvalues are real, which must be true for us to
interpret the values as possible energies of the system.
Furthermore we know that the wavefunctions will be orthogonal. The normalization constant
can be chosen so that they are orthonormal, so that we have
_

n
(x)
m
(x) dx =
nm
. (8.225)
Here the subscript on refers to the order of the Hermite polynomial that it contains, and
also determines the energy of the oscillator through (8.211). Notice that here the inner product
has the weight function w = 1, as we saw from the dierential equation for u(y) in the Sturm-
Liouville form, Eq. (8.186).
This gives for the wavefunctions
n
(x) =
1
2
n
n!(a
2
)
1/4
H
n
(x/a)e
x
2
/2a
2
, n = 0, 1, . . . . (8.226)
Some plots of wavefunctions
n
(x) using a = 1 are shown in Fig. 8.12 for several values of n.
x
-5 0 5
(
x
)
n
-1
0
1
0

1

2

3

Figure 8.12: Plot of wavefunctions
n
(x) for the quantum harmonic
oscillator for several values of n.
8.4 The gamma function
In this section we will take a look at the Euler gamma function, usually just called the gamma
function. This is dierent from the functions that we have seen earlier in this chapter in that
it is not the solution of a Sturm-Liouville equation, but it is nevertheless very useful and comes
up in a number of problems of mathematical physics.
The gamma function can be dened through the integral relation
(x) =
_

0
e
t
t
x1
dt . (8.227)
One can easily work out the special value (1) as
(1) =
_

0
e
t
dt = e
t
0
= 1 . (8.228)
Using Eq. (8.227) we can write down (x + 1) and integrate it by parts with u = t
x
and
dv = e
t
dt, which gives
(x + 1) =
_

0
e
t
t
x
dt = t
x
e
t
0
+
_

0
e
t
xt
x1
dt = x(x) , (8.229)
since the rst term is zero and the factor of x in the second term can be pulled outside the
integral (which is over t). But since (1) = 1 we nd
(2) = 1(1) = 1 , (8.230)
(3) = 2(2) = 2 1 , (8.231)
(4) = 3(3) = 3 2 1 , (8.232)
(5) = 4(4) = 4 3 2 1 , (8.233)
and so forth. That is, for integer x 1 one has
(x) = (x 1)! (8.234)
The denition of the gamma function is, however, perfectly valid for noninteger and even
negative values of the argument x. For example, one nds
(1/2) =
. (8.235)
A plot of the gamma function for both positive and negative x is shown in Fig. 8.13.
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
(
x
)
-20
-10
0
10
20
Figure 8.13: Plot of the gamma function.
As can be seen in the gure, for x 1, the gamma function (x) essentially connects the
dots between the values of (x 1)! that one has for integer x. But for x < 1 the behaviour is
nothing like a smooth extrapolation of the behaviour for x 1. The gamma function diverges
to either + or for all integer values of the argument less than or equal to zero, depending
on the direction from which one approaches the singularity.
As an example of the use of the gamma function, consider again Bessels equation
x
2
y
+xy
+ (x
2
2
)y = 0 , (8.236)
but suppose now that the parameter is not required to be an integer. In this case one nds
the series solution of the form
J
(x) =
k=0
(1)
k
k!(k + + 1)
_
x
2
_
2k+
. (8.237)
Comparing with Eq. (8.48) we see this is the same as what we found before for integer , but
the series (8.237) is valid for equal to any real value.

CH 8 Aaaaaaaaaaaaaaaaaaaaaaaaaa

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

CH 8 Aaaaaaaaaaaaaaaaaaaaaaaaaa

Diunggah oleh

Hak Cipta:

Format Tersedia

Chapter 8

/Q term. This gives

/Q and all of the r dependence is in the other terms.

, which we can nd using the chain rule,

(x). They are shown for integer order in Fig. 8.2.

/ and divide through by sin

l(l + 1)R = 0 . (8.162)

h. We then need to bring in other quantities to cancel

Anda mungkin juga menyukai