Chapter 5

Fourier Expansions and Separation of

Variables in Rectangular Coordinates

Eigenfunctions come naturally in the treatment of vibrations in a confined space. Electro-

magnetic eigenmodes arise in the same way as their mechanical counterpart, such as sound
waves in organ pipes. The basic equation for waves of speed c is

∂ 2ψ 1 ∂2ψ
= 2 2 (5.1)
∂x2 c ∂t
in one dimension. An eigenmode is sinusoidal in time, so that the wave equation becomes

∂2ψ
+ k 2 ψ = 0, (5.2)
∂x2
where c |k| is the angular frequency (usually called ω). The solutions are plane waves
exp(ikx), with positive and negative k.

5.1 Convergence
If the waves are confined to a ring of length L, the eigenfunctions are simply plane waves
with k taking the special values (eigenvalues)

kn = 2πn/L, n integer. (5.3)

The expansion
∞
X
f (x) = Cn exp(ikn x), (5.4)
n=−∞

with Z L/2
Cn = exp(−ikn x)f (x)dx/L, (5.5)
−L/2

32
is the basic Fourier expansion. It converges uniformly for any continuous periodic function
of period L. If f (x) is discontinuous or not periodic (i.e. if f (0) 6= f (L)), or worse if it is
singular, the convergence is slow, not uniform, and likely to cause problems in numerical
evaluations. At a simple jump of f from 0 to 1, for instance, any partial sum S(x) of
the Fourier series will “overshoot” up to 1.179 and show some wiggles. The overshoot and
wiggles occur over a narrower range as more terms are included in S, but never die out—this
is known as the “Gibbs’ phenomenon.” One should be wary when using Fourier expansions
near jumps and even near cusps. On the other hand for any square-integrable f there is
convergence in the mean, i.e. the mean square error
Z L/2
M= |f (x) − S(x)|2 dx (5.6)
−L/2

can be made arbitrarily small. In practice, this means that most integrals involving f can
also be evaluated with arbitrarily small error.

5.2 Normalization
The eigenfunctions exp(ikn x) are mutually orthogonal but are normalized to L:

Z L/2 Z L/2
∗
[exp(ikm x)] exp(ikn x) dx = exp(2πi(n − m)x/L) dx = Lδnm . (5.7)
−L/2 −L/2

This is not a problem, it simply entails that we must use dx/L in integrals such as (5.5). On
the other hand, we may prefer to introduce normalized eigenfunctions
1
un (x) = √ exp(ikn x) (5.8)
L
and obtain, in place of (5.4) and (5.5), the more symmetric pair
∞
X
f (x) = an un (x), (5.9)
n=−∞

Z L/2
an = u∗n (x)f (x) dx. (5.10)
−L/2

Often one deals with functions defined over all space, i.e. with the limit a → ∞.
The limit can be tricky and sometimes it is better to keep a finite until the very end of the
calculation. But usually one can simply make the replacements kn → k and
∞
X L Z∞
→ dk. (5.11)
n=−∞ 2π −∞

33
The most usual way to write the resulting Fourier-integral transform pair is
Z ∞ dk
f (x) = F (k) eikx , (5.12)
−∞ 2π
Z ∞
F (k) = f (x) e−ikx dx, (5.13)
−∞
where f (k) corresponds to LC √n of Eq. (5.5). Mathematicians (and Jackson, at times) have
the habit of using a factor 1/ 2π in front of each integral, instead of leaving the entire 1/2π
attached to the k integral.

5.3 Other boundary conditions

In place of the basis functions exp(ikn x) one can use cos(kn x) and sin(kn x) with kn ≥ 0.
This serves only to produce messier formulas.
On the other hand, there are cases where the functions of interest are required to
vanish at the boundaries (−L/2, L/2) or (0, L). The relevant normalized eigenfunctions in
(0, L) are s
2
Un (x) = sin(qn x), (5.14)
L
with
qn = nπ/L, n positive integer. (5.15)
The transform pair (5.9), (5.10) is still valid in this case and is in fact entirely general
for expansions in orthonormal eigenfunctions, with appropriate modifications. One obvious
modification is that for fixed boundary conditions the sum over n runs from 1 to ∞, while
for periodic boundary conditions it runs from −∞ to ∞. The comments about convergence
apply to the general case as well.

5.4 Higher dimensions

In 3 dimensions, the wave equation leads to the Helmholtz equation
∇2 ψ = −k 2 ψ, (5.16)
which has plane wave solutions exp(ik · x). Periodic boundary conditions on a box of sides
Lx , Ly , Lz are not physically attainable in this case, but they are mathematically convenient.
They lead to the set of orthogonal functions exp(ik · x), with the j Cartesian component of
k given by
(k)j = 2πnj /Lj , nj integer. (5.17)
The expansion of a function f (x) is
∞
X ∞
X ∞
X
f (x) = Cnx ,ny ,nz exp(ik · x), (5.18)
nx =−∞ ny =−∞ nz =−∞

34
with Z Lx /2 dx Z Ly /2 dy Z Lz /2 dz
Cnx ,ny ,nz = exp(−ik · x)f (x). (5.19)
−Lx /2 Lx −Ly /2 Ly −Lz /2 Lz
In the continuum limit we obtain the Fourier integral representation
Z
d3 k
f (x) = F (k) eik·x , (5.20)
(2π)3
Z
F (k) = f (x) e−ik·x d3 x, (5.21)
by the replacements
∞
X ∞
X ∞
X V Z 3
→ d k, (5.22)
nx =−∞ ny =−∞ nz =−∞ (2π)3
V Cn → f (k), (5.23)
where V = Lx Ly Lz is the volume of the “quantization box” and the integrals are over all
space. Similar formulae apply in two dimensions.

5.5 The Dirac delta function

This is a particularly important singular function. In one dimension, it has the general
expression X
δ(x − x0 ) = u∗n (x0 )un (x) (5.24)
n

for any complete set of eigenfunctions Un . For the plane waves exp(ikn x) this becomes
∞
1 X
0
δ(x − x ) = exp [ikn (x − x0 )] , (5.25)
L n=−∞

and in the continuum limit Z

0 dk 0
δ(x − x ) = . eik(x−x ) (5.26)
2π
The three-dimensional generalization is the usual δ(x) = δ(x)δ(y)δ(z), so that, for instance
Z
0 d3 k
δ(x − x0 ) = eik·(x−x ) . (5.27)
(2π)3

It is worth noting that the partial sum SN of (5.25) has a central peak that becomes higher
and narrower for larger N , but also a lot of side wiggles that never go away:

1 e−i2πN x/L − ei2πN x/L sin [(2N + 1)πx/L]

SN (x) = = . (5.28)
a 1−e i2πx/L L sin(πx/L)

35
5.6 Application to electrostatics
How do we use all this math in electrostatics, where in the simplest case we need solutions
of Laplace’s equation, corresponding to k = 0 in the Helmholtz equation, with appropriate
boundary conditions?

5.6.1 Rectangular box

Suppose the potential is assigned on the sides of an empty conducting rectangular box of
sides Lx , Ly , Lz . It is enough to consider the case where Φ is not zero on one face only:
the general case can be built by superposition. Suppose then that Φ = V (x, y) on the face
z = Lz , and vanishes on the other faces. We Fourier-expand in the dimensions x and y,
cleverly picking eigenfunctions that vanish on the lateral faces:
X
Φ(x, y, z) = sin(nx πx/Lx ) sin (ny πy/Ly ) Anx ny (z). (5.29)
nx ny

The expansion coefficients are now functions of z, to be determined by using Laplace’s

equation and the remaining boundary conditions on the top and bottom faces.

• From ∇2 Φ = 0 for all x and y it follows that

 !2 
 2
−
nx π ny π ∂2
− + 2  Anx ny (z) = 0, (5.30)
Lx Ly ∂z

which has the solutions exp(±γnx ny z), with

v
u !2
u nx π 2 ny π
t
γnx ny = + . (5.31)
Lx Ly

• To satisfy Φ = 0 on the z = 0 face, we pick the combination

Anx ny (z) = Anx ny sinh(γnx ny z). (5.32)

• To satisfy the condition on the z = Lz face, which is

X
sin(nx πx/Lx ) sin (ny πy/Ly ) Anx ny sinh(γnx ny Lz ) = V (x, y), (5.33)
nx ny

we use Fourier’s theorem and find

Z Lx Z Ly
4
Anx ny = dx dy V (x, y) sin(nx πx/Lx ) sin (ny πy/Ly ) .
Lx Ly sinh(γnx ny Lz ) 0 0
(5.34)

36
 • The complete solution was easily obtained because:

– The starting equation (Laplace’s equation in this case) is completely separable

in the chosen variables x, y, z. That means that there are solutions of the form
X(x)Y (y)Z(z). In our case the differential operator is the sum of three separate
pieces that depend only on x, y and z respectively. This is the simplest case of
separability.
– The boundary conditions are also separable in the same variables x, y, z.

• Even partial separability may help. For instance, we may be able to expand in one
dimension only, thus reducing the problem to a two-dimensional one.

• We are still left with an infinite double sum that must usually be evaluated numerically,
with all the caveats about the behavior at discontinuities. It is worthwhile to look for
approximations that permit the series to be evaluated analytically near singular points.

5.6.2 Rectangular trough

This is a two-dimensional problem: nothing depends on z. The boundary conditions are
that Φ vanishes at x = 0 and x = L, as well as at y = ∞, and it has an assigned value V (x)
at y = 0 (Fig. 2.10). The solution is
∞
X
Φ(x, y) = An sin(nπx/L) exp(−nπy/L), (5.35)
n=1

with
2ZL
An = V (x) sin(nπx/L) dx. (5.36)
L 0
For instance, if V (x) is constant, we have

4V
An = for odd n, (5.37)
nπ
and 0 otherwise. In simple cases like this the series can be summed. As in most two-
dimensional problems, it is convenient to work with the complex variable

z = x + iy (5.38)

instead of x and y. In our case we note that

4V X 1
Φ= Im exp(inπz/a). (5.39)
π n odd n

37
We put Z = eiπz/L and recall that
∞
X 1
ln(1 + Z) = (−1)n+1 Z n ,
n=1 n
X∞
1 n
− ln(1 − Z) = Z , (5.40)
n=1 n

so that by subtracting,
1+Z X 1
ln =2 Z n. (5.41)
1−Z n odd n
Then   !
2V 1+Z 2V 2 Im Z
Φ= Im ln = arctan . (5.42)
π 1−Z π 1 − |Z|2
With Im Z = e−πy/L sin(πx/L) and |Z|2 = e−2πy/L , we obtain finally
" #
2V sin(πx/L)
Φ= arctan . (5.43)
π sinh(πy/L)

Plots of the equipotentials and field lines are shown in the figure.

Φ=0 Φ=0

x=0 x=L
Φ= V

In the corner near the origin, Φ ' V (1 − 2ϕ/π), where ϕ = arctan(y/x) as usual.
The field is azimuthal, Eϕ = 2V /πρ. The energy density diverges as 1/ρ2 and the total
energy for ρ > ρ0 also diverges as ρ0 → 0, although only logarithmically. More generally, it
costs infinite energy to join two plates at any angle α when there is a potential difference V
between them. In practice, a small gap must be left and the energy of the adjoining plates

38
goes like (V /α)2 ln w per unit length of the common edge, where w the width of the gap.
This logarithmic divergence also occurs for adjoining edges that are not straight.
Jackson has many problems and examples of adjoining plates and shells at different
potentials. Such problems are not very realistic, for the reason discussed above.

5.6.3 Polar coordinates

Some problems in electrostatics are best solved in polar coordinates (ρ, φ). Laplace’s equation
in polar coordinates is !
1 ∂ ∂Φ 1 ∂ 2Φ
∇ Φ=
2
ρ + 2 2 = 0. (5.44)
ρ ∂ρ ∂ρ ρ ∂φ
Let’s assume that the solution is separable in these coordinates, so that Φ(ρ, φ) = R(ρ)Q(φ).
Substituting this into Laplace’s equation, and multiplying by ρ2 /RQ, we obtain
!
ρ d dR 1 d2 Q
ρ + = 0. (5.45)
R dρ dρ Q dφ2
The first term is only a function of ρ and the second term is only a function of φ, so they
must both be equal to a constant which we will call −ν 2 . Therefore,
1 d2 Q
2
= −ν 2 , (5.46)
Q dφ
with the solution
Q(φ) = A cos(νφ) + B sin(νφ). (5.47)
The radial function R(ρ) is a solution of
!
ρ d dR
ρ − ν 2 = 0. (5.48)
R dρ dρ
To solve this, seek solutions of the form R(ρ) = aρα . Substituting into Eq. (5.6.3), we find
solutions for α = ±ν. We must treat ν = 0 separately; in this case Q(φ) = A0 + B0 φ, and
R(ρ) = a0 + b0 ln(ρ/R), with R some scale factor. Therefore, our general solution is
Φ(ρ, φ) = [a0 + b0 ln(ρ/R)](A0 + B0 φ) + (aν ρν + bν ρ−ν )(Aν cos νφ + Bν sin νφ). (5.49)
This result is quite general. Suppose that we now permit the full range of φ, so that Φ(ρ, φ +
2π) = Φ(ρ, φ); in this case B0 = 0, and ν = n, with n a non-negative integer:
∞ 
X 
Φ(ρ, φ) = a0 + b0 ln(ρ/R) + an ρn + bn ρ−n (An cos nφ + Bn sin nφ) . (5.50)
n=1

As an example, let’s consider a grounded conducting cylinder placed in a uniform

electric field perpendicular to its axis. For ρ → ∞, to produce a uniform field we must have
Φ(ρ, φ) = −E0 ρ cos φ; the solution is then of the form
∞
X
Φ(ρ, φ) = a0 − E0 ρ cos φ + ρ−n (An cos nφ + Bn sin nφ) . (5.51)
n=1

39
On the cylinder we have Φ(ρ = a, φ) = 0, so that a0 = 0 and A1 = E0 a2 , with all of the
other An ’s and Bn ’s zero. The solution is therefore
!
a2
Φ(ρ, φ) = −E0 ρ− cos φ, (5.52)
ρ

which we also obtained using the method of images.

40