NPTEL COURSE

SELECTED TOPICS IN MATHEMATICAL PHYSICS

V. Balakrishnan
Department of Physics, Indian Institute of Technology Madras
Chennai 600 036, India

Note: All figures have been omitted in this write-up. The reader will find it an instruc-
tive exercise to fill in all necessary figures.

Table of contents

1. Analytic functions of a complex variable: Complex numbers. Equations to

curves in the plane in terms of z and z ∗ . The Riemann sphere and stereographic
projection. Analytic functions of z and the Cauchy-Riemann conditions. The
real and imaginary parts of an analytic function. The derivative of an analytic
function. Power series as analytic functions. Convergence of power series.

2. Calculus of residues: Cauchy’s integral theorem. Singularities— removable

singularity, simple pole, multiple pole, essential singularity. Laurent series. Sin-
gularity at infinity. Accumulation point of poles. Meromorphic function. Cauchy’s
integral formula. Contour integration. Residue at infinity. Summation of series
using contour integration. Solving difference equations using contour integration.
Mittag-Leffler expansions of meromorphic functions.

3. Linear response; dispersion relations: Causal, linear, retarded response.

Dynamic susceptibility. Symmetry properties of the dynamic susceptibility. Dis-
persion relations. Admittance of an LCR circuit. Subtracted dispersion relations.
Hilbert transform pairs. Discrete and continuous relaxation spectra.

4. Analytic continuation and the gamma function: The gamma function.

Connection with the gaussian integral. Analytic properties of the gamma func-
tion. Mittag-Leffler expansion of Γ(z). The logarithmic derivative of the gamma
function. Infinite product representation for the gamma function. The beta
function. Reflection formula for Γ(z). Legendre’s duplication formula for Γ(z).

5. Möbius transformations: Conformal mapping. Definition of a Möbius trans-

formation. Fixed points of a Möbius transformation. The cross-ratio. Normal
form of a Möbius transformation. Iterates of a Möbius transformation. Classi-
fication of Möbius transformations. The isometric circle. Group properies; the
Möbius group. The Möbius group over the reals. The modular group. The
invariance group of the unit circle. Connection with the pseudo-unitary group
SU(1, 1). The group of cross-ratios.

1
 6. Multivalued functions; integral representations: Branch points and branch
cuts. Algebraic and logarithmic branch points, winding point. Contour integrals
in the presence of branch points. An integral involving a class of rational func-
tions. Contour integral representations for the gamma, beta, and Riemann zeta
functions. Connection with Bernoulli numbers. The statement of the Riemann
hypothesis. Contour integral representations of the Legendre functions Pν (z) and
Qν (z). Singularities of functions defined by integrals. End-point and pinch singu-
larities, examples. Singularities of the Legendre functions. Dispersion relations
for the Legendre functions.

7. Laplace transforms: Definition of the Laplace transform. The convolution

theorem. Laplace transforms of derivatives. The inverse transform, Mellin’s
formula. Laplace transform of the Bessel and modified Bessel functions of the
first kind. The LCR series circuit. Laplace transforms and random processes: the
Poisson process, biased random walk on a linear lattice and on a d-dimensional
lattice.

8. Fourier transforms: Fourier integrals. Parseval’s formula for Fourier trans-

forms. Fourier transform of the δ-function. Relative ‘spreads’ of a Fourier trans-
form pair. The convolution theorem. Generalization of Parseval’s formula. Iter-
ates of the Fourier transform operator F . Eigenvalues and eigenfunctions of F .
Unitarity of the Fourier transformation. The Fourier transform in d dimensions.
The Poisson summation formula. Some illustrative examples. Generalization to
higher dimensions.

9. Quiz 1; solutions.

10. The fundamental Green function for ∇2 : Green functions. Poisson’s equa-
tion; the fundamental Green function for ∇2 . Solution for a spherically symmet-
ric source. The Coulomb potential in d dimensions. Ultraspherical coordinates.
A divergence problem. Dimensional regularization. A direct derivation using
Gauss’ Theorem. The Coulomb potential in d = 2 dimensions. A direct deriva-
tion once again.

11. The diffusion equation: Fick’s laws of diffusion. The fundamental solution
in d dimensions. Solution for an arbitrary initial distribution. Moments of the
distance travelled in time t. Diffusion in one dimension: continuum limit of a ran-
dom walk. The Smoluchowski equation. Sedimentation. Absorbing and reflecting
boundary conditions. Free diffusion on a semi-infinite line. Finite boundaries:
solution by the method of images. Solution by separation of variables. Survival
probability. First-passage-time distribution and mean first-passage time. Con-
nection with the Schrödinger equation for a free particle. Spreading of a quantum
mechanical wave packet.

2
12. The Green function for (∇2 +k 2 ); nonrelativistic scattering: The Helmholtz
operator. The scattering amplitude; differential and total cross-sections. Inte-
gral equation for scattering. Green function for the Helmholtz operator. Exact
formula for the scattering amplitude. Scattering geometry and the momentum
transfer. Born series and the Born approximation. The Yukawa and Coulomb
potentials. The Rutherford scattering formula.

13. The wave equation: Formal solution for the causal Green function. The
solution in (1 + 1), (2 + 1) and (3 + 1) dimensions. Retarded solution of the wave
equation. Remarks on propagation in spatial dimensions d > 3.

14. The rotation group and all that: Rotations of the coordinate axes. Orthog-
onality of rotation matrices. Proper and improper rotations. Generators of
infinitesimal rotations in 3 dimensions. The general rotation matrix in 3 dimen-
sions. The finite rotation formula for a vector. Relation between the groups
SO(3) and SU(2). Rotation generators in 3 dimensions transform like a vector.
Derivation of Hadamard’s lemma. The general form of the elements of U(2) and
SU(2). The parameter spaces of SU(2) and SO(3). Tensor and spinor represen-
tations. The universal covering group of a Lie group. The group SO(2) and its
covering group. The groups SO(n) and Spin (n). Parameter spaces of U(n) and
SU(n). A bit about the first homotopy group of a space.

15. Quiz 2; solutions.

3
1 Analytic functions of a complex variable
We begin this course with the study of analytic functions of a complex variable. The
theory of functions of a complex variable (and its generalization, the theory of functions
of several complex variables) is one of the richest and most beautiful branches of math-
ematics, with deep results and far-reaching implications and applications. It is also a
vast subject. Only its most rudimentary aspects will be dealt with in what follows.
No formal proofs of theorems will be given, as usual, and the treatment of standard
material will be heuristic rather than rigorous. I shall assume that you already have
some familiarity with analytic functions of a complex variable, from an earlier course
on mathematical methods in physics.

Complex numbers: Let’s recapitulate some elementary properties of complex num-

bers. Given a complex number z = x + iy = reiθ , its real and imaginary parts are x
and y, respectively; its modulus and argument are r and θ, respectively. Its complex
conjugate is z ∗ = x − iy = re−iθ . It is important to remember that the specification of
a complex number implies the specification of two independent pieces of information,
namely, x and y, or r and θ. be z and z ∗ themselves. Recall also the basic relations
(z + z ∗ ) (z − z ∗ )
x = r cos θ = , y = r sin θ = ,
2 2i
as well as

y 1
z
r = (x2 + y 2 )1/2 = zz ∗ = |z|2 , θ = tan−1 = ln .
x 2i z∗

1. Show that the real and imaginary parts of the complex numbers listed below are as
indicated. Remember the standard phase convention, according to which i = eiπ/2 and
−i = e−iπ/2 = e3πi/2 .
   
(a) (i)i = cos 12 πe−π/2 + i sin 12 πe−π/2 .
i

∞
(iπ)2n+1
(b) = 0.
n=0
(2n + 1)!
 n

 π
π
π
π
∞
1 1+i
(c) √ = cosh cos cos sin + i sinh cos sinh sin .
n=0
(2n)! 2 8 8 8 8

(d) The iterated square root

 √ √

17 − 1
√ 1 17 + 1
i + i + i + . . . ad inf. = + √ +i √
2 2 2 2 2
 1.3002 + 0.6248 i.
(The other root is  −0.3002 − 0.6248 i.)

4
 (e) The infinite continued fraction
√ √
i i i 1 17 + 1 17 − 1
· · · ad inf. = − + √ +i √
1+ 1+ 1+ 2 2 2 2 2
 0.3002 + 0.6248 i.

(The other root is  −1.3002 − 0.6248 i.)

(f) The infinite continued fraction

√
1 1 1 3−i
· · · ad inf. = e−iπ/6 = .
i+ i+ i+ 2
√
(The other root is e−5πi/6 = − 12 ( 3 + i).)

Hint: In (d), (e) and (f), calling the left-hand side z, you have to solve the respec-
tive equations z = (i + z)1/2 , z = i/(1 + z) and z = 1/(i + z). Each of these reduces
to a quadratic equation, and you must take care to identify the correct root in each case.

Equations to curves in the plane in terms of z and z ∗ : We know that the Carte-
sian coordinates x and y labeling points on a plane are linearly independent of each
other. It is important to recognize that, equivalently, the complex variable z = x + iy
and its complex conjugate z ∗ = x − iy are linearly independent. Once you bear this
in mind, it becomes much easier to understand the concept of analytic functions of a
complex variable z.

2. Equations to curves in the xy-plane are often very conveniently expressed in terms
of z and z ∗ . Here are some examples of familiar curves, thus expressed:
(a) In terms of the complex variable z, write down the equation of the ellipse in the
xy-plane whose foci are at the points ±x0 on the x-axis, and whose semi-major
axis is equal to a (where a > x0 ).

(b) Find the locus of points in the z-plane given by the following conditions :
     2
(i)  |z − 1| − |z + 1|  = 1 (ii) (z − 1)/(z − 2) = 1 (iii) ez  = e4 .

The Riemann sphere and stereographic projection: We know that the number
line extends from −∞ on the left to +∞ on the right. In the complex plane, z may tend
to infinity along any of an infinite number of directions. It is convenient to ‘compactify’
the plane by bringing together all these points at infinity and ‘glueing’ them together
into a single point. This can be done by a mapping between the complex plane and
the surface of a sphere. Consider a sphere of unit radius with its center at the origin

5
of coordinates. The coordinates of any point on the surface are given by (ξ1 , ξ2 , ξ3 ),
where
ξ12 + ξ22 + ξ32 = 1.
In terms of spherical polar coordinates on the unit sphere, we have of course
ξ1 = sin θ cos ϕ , ξ2 = sin θ sin ϕ , ξ3 = cos θ .
The coordinates of the ‘north pole’ N are (0, 0, 1), while those of the ‘south pole’ S are
(0, 0, −1).

Now consider the equatorial plane of the sphere. Let the Cartesian coordinates
on this plane be (x , y), with the same origin as the sphere, and the x and y axes
running along the ξ1 and ξ2 axes, respectively. The stereographic projection of any
point P on the sphere onto the equatorial plane is obtained by joining N and P by a
straight line. The projection of P is the point P  where the line cuts the equatorial
plane. This plane is regarded as the complex plane, while the sphere is called the
Riemann sphere. The coordinates x and y of P  in the complex plane are related to
the coordinates (ξ1 , ξ2 , ξ3 ) on the Riemann sphere by
ξ1 ξ2
x= = cot 12 θ cos ϕ and y = = cot 12 θ sin ϕ.
1 − ξ3 1 − ξ3
Therefore
ξ1 + iξ2
z = x + iy = = cot 12 θ eiϕ .
1 − ξ3
It is evident that points of the northern hemisphere are mapped to points outside the
unit circle in the complex plane; points in the southern hemisphere are mapped to
points inside the unit circle; and points on the equator are mapped to themselves as
points on the unit circle |z| = 1.

As P gets closer to the point of projection N, it is clear that the point P  moves
farther and farther away from the origin in the complex plane. The point at infinity
in the complex plane is defined as the image of the north pole under the projection,
and denoted by ∞ . The finite part of the complex plane, together with the point at in-
finity, is called the extended complex plane. In what follows, by the term ‘complex
plane’ we shall generally mean the extended complex plane. It is very advantageous
to have ‘infinity’ identified with a single point in this fashion, and to be able to specify
∞ as a specific value for the complex number z.

2. Some of the properties of this stereographic projection are the following.

(a) The mapping from the Riemann sphere to the complex plane, is invertible. Show
that the inverse relations expressing (ξ1 , ξ2 , ξ3 ) in terms of z and z ∗ are
z + z∗ z − z∗ |z|2 − 1
ξ1 = , ξ2 = , ξ3 = .
|z|2 + 1 i(|z|2 + 1) |z|2 + 1

6
 (b) What is the projection in the complex plane of a circle of latitude on the Riemann
sphere? Of a meridian of longitude?

• In general, circles on the Riemann sphere are mapped into circles or straight lines
in the complex plane by stereographic projection.

(c) Let P and Q be any two points on the Riemann sphere, whose projections on the
complex plane are given by z1 and z2 . Show that the chordal distance between
P and Q (i.e., the length of a chord drawn from P to Q) is given by
2 |z1 − z2 |
d(z1 , z2 ) =  .
(|z1 |2 + 1)(|z2 |2 + 1)

It is obvious that d(z1 , z2 ) = d(z2 , z1 ) and that d(z1 , z2 ) = 0 iff z1 = z2 . Further, for

any three points z1 , z2 and z3 , we have

d(z1 , z3 ) ≤ d(z1 , z2 ) + d(z2 , z3 ).

With this definition of the distance, the exceptional role played by the point at infinity
causes no trouble, as it does in the case of the ordinary Euclidean definition |z1 − z2 | .
We have
2
d(z, ∞) =  .
|z|2 + 1
Clearly, d(z1 , z2 ) ≤ 2 for any two points z1 and z2 in the extended complex plane.

Analytic functions of z ; the Cauchy-Riemann conditions: The function f (z)

is said to be analytic (more precisely, holomorphic) in some region of the complex
z-plane if its real and imaginary parts satisfy the Cauchy-Riemann conditions at every
point in this region. Let u(x , y) and v(x , y) be the real and imaginary parts of such
a function. The Cauchy-Riemann conditions are
∂u ∂v ∂u ∂v
= and =− .
∂x ∂y ∂y ∂x
The region in which these conditions are satisfied is the region of analyticity, or domain
of holomorphy, of the function concerned. If a function f (z) is holomorphic in the
whole of the finite part of the complex plane (that is, for all |z| < ∞), then it is called
an entire function. There exists a theorem (Liouville’s Theorem) that states that
• the only function analytic at every point in the extended complex plane (that
is, at all points in the finite part of the complex plane as well as the point at
infinity) is a constant.

7
 • Hence an entire function that is not just a constant cannot be analytic at the
point at infinity.1
One way to understand what is meant by an analytic function of z is as follows.
Recall that, just as x and y are linearly independent coordinates on the plane, so are
the linear combinations z = x + iy and z ∗ = x − iy. Hence any function of x and y can
equally well be written as a function of z and z ∗ . An analytic function of z is then a
sufficiently well-behaved function2 of x and y that depends on the combination x + iy
alone, and that does not involve the other combination, namely, x − iy. Thus, f is an
analytic function of z if ∂f /∂z ∗ ≡ 0. But, by the chain rule of differentiation,
∂f ∂f ∂f
=0 =⇒ −i = 0.
∂z ∗ ∂x ∂y
The real and imaginary parts of the last equation yield precisely the Cauchy-Riemann
conditions. These conditions may therefore be written compactly as ∂f /∂z ∗ ≡ 0.
Writing any function of x and y in terms of z and z ∗ enables us to identify functions
that cannot be analytic functions of z, merely by checking whether z ∗ makes its ap-
pearance in the function. If it does so, the function cannot be an analytic function of z.

4. Identify which among the following are analytic functions of z in some region, and
which are not. In the former case, identify the region of analyticity.
(a) x (b) ix − y (c) r
(d) eiθ
(e) x − iy
2 2
(f) x − iy
(g) (x + iy) 2
(h) x − y − 2ixy
2 2
(i) i tan−1 (y/x)
(j) (iy + x2 + y 2) (k) [(x + i)2 − y 2 ]1/2 (l) (x − iy)/(x2 + y 2 )
(m) x4 + 2ix2 y 2 − y 4 (n) i ex sin y (o) x2 + x + 1 − y 2 + iy (2x + 1).

The real and imaginary parts of an analytic function : Let u and v be the real
and imaginary parts of an analytic function f (z). It follows from the Cauchy-Riemann
conditions that both u and v individually satisfy Laplace’s equation in two dimensions,
i.e., ∇2 u = 0 and ∇2 v = 0.
• The real and imaginary parts of an analytic function are thus harmonic functions.
The combination u + iv constitutes an analytic function in the intersection of the
regions in which they are individually harmonic.
• It is also evident from the Cauchy-Riemann conditions that an analytic function
cannot be identically equal to a purely real or purely imaginary function, except
in the trivial case when the function is just a constant.
1
In other words, it must be singular at z = ∞. We will discuss singularities subsequently.
2
By the phrase ‘sufficiently well-behaved’, we mean that the real and imaginary parts of the
function have continuous partial derivatives with respect to x and y.

8
 • An analytic function that is real when its argument is real, i.e., a function such
that f (x) = u(x, 0) is real (or v(x, 0) vanishes identically), is called a real ana-
lytic function.

• It follows from the Cauchy-Riemann conditions that the scalar product (∇u ·
∇v) = 0. The curves u = constant and v = constant thus constitute two mutu-
ally orthogonal families of curves in the complex plane, for any analytic function.

5. Given u (or v) as indicated below, find the corresponding v (respectively, u), either
by inspection or by integrating the Cauchy-Riemann conditions. Indicate also the
region of analyticity of f (z) in each case.

(a) u = x2 − y 2 (b) u = ex cos y (c) u = ln (x2 + y 2 )1/2

(d) u = cos x cosh y (e) v = e2x sin (2y) (f) v = 2xy
2 −y 2
(g) v = sin x sinh y (h) v = tan−1 (y/x) (i) v = ex sin (2xy).

6. Cauchy-Riemann conditions in polar form : Let z = r eiθ and f (z) = R eiψ .

Thus R = |f (z)|, while ψ = arg f (z). Show that the Cauchy-Riemann conditions now
read
∂ ln R ∂ψ ∂ ln R ∂ψ
= , =− .
∂ ln r ∂θ ∂θ ∂ ln r
The derivative of an analytic function : The condition of analyticity is so strong
that an analytic function is guaranteed to have a derivative that is, moreover, itself
an analytic function. It then follows at once that it has derivatives of arbitrarily high
order that are also analytic functions. Clearly, this is in marked contrast to the case of
functions of a real variable, where a function may be once differentiable but not twice
differentiable, or, in general, differentiable r times, but not (r + 1) times. On the other
hand,

• any analytic function of a complex variable is infinitely differentiable.

Defining the derivative of an analytic function helps us understand the Cauchy-Riemann

conditions and the meaning of analyticity in yet another (albeit related) way. In anal-
ogy with the definition of the derivative of a function of a real variable, we may define
the derivative as
df (z) f (z + δz) − f (z)
= lim ,
dz δz→0 δz
where δz is an infinitesimal quantity of magnitude . The question that arises is : in
what direction should the point z + δz be taken, relative to the point z? That is, what
should the argument (or phase angle) of δz be? Suppose the complex number δz has
an argument α, i.e., δz =  eiα . Then, provided the real and imaginary parts of f (z)

9
have continuous partial derivatives, the definition of the derivative given above yields,
in the limit  → 0,
   
df (z) −iα ∂u ∂v ∂v ∂u
=e cos α + i sin α + i cos α − i sin α .
dz ∂x ∂y ∂x ∂y
The remarkable fact is that this expression becomes completely independent of α if
and only if
∂u ∂v ∂u ∂v
= and =− ,
∂x ∂y ∂y ∂x
which are precisely the Cauchy-Riemann conditions.
• Analyticity may therefore be understood as equivalent to the requirement that
the limit defining the derivative of a complex function be unique, independent of
the direction along which z + δz approaches z in the complex plane.

7. In its region of analyticity, and analytic function f (z) may also be regarded as a
map from (a region of) the complex plane to (a region of) the complex plane.
(a) Show that (the determinant of) the Jacobian of the transformation (x , y)

(u , v) is just |f  (z)|2 , where f  (z) denotes the derivative of f (z).
(b) Show that ∇2 (|f (z)|2 ) = 4 |f  (z)|2 .

Power series as analytic functions : Let f (z) be analytic inside a region R. If z0

is a point in the region, then there generally exists a neighborhood of z0 such that f (z)
can be expanded in an absolutely convergent power series in (z − z0 ), called a Taylor
series :

∞
f (z) = an (z − z0 )n .
n=0
Moreover, this neighborhood is the interior of a circle centered at z0 , i.e., it is given by
|z − z0 | < R, where R is called the radius of convergence of the series. The circle
|z − z0 | = R is called the circle of convergence of the series. It lies within R, the
region of analyticity of f (z). (Its boundary may coincide with that of R at one or more
points.) Absolute convergence means that the sum of the magnitudes of the terms of
the series is also finite. Loosely speaking, such convergence permits us to manipulate
the series terms by term, as long as z remains inside the circle of convergence : that
is, the series may be re-arranged without affecting the sum of the series, or integrated
or differentiated term by term, and so on.

If f (z) and z0 are specified, the coefficients an in the Taylor series above are uniquely
determined. We have 
1 dn f (z)
an = ,
n! dz n z=z0

10
the existence of derivatives of all orders being guaranteed by the analyticity of f (z).
The radius of convergence itself is determined by the coefficients {an } —more precisely,
by the asymptotic (or n → ∞) behavior of an . The radius of convergence is given by
 
 an 
R = lim  ,
n→∞ an+1 

provided the limit concerned exists. More generally,

R = lim sup |an |1/n ,

where lim sup (or supremum) stands for the least upper bound of the quantity con-
cerned. It is important to keep in mind that

• a power series is a representation of an analytic function inside the circle of con-

vergence of the series. The function may have other representations as well, each
with its own region of validity.

8. Convergence of power series : Find the region of absolute convergence of each

of the following power series :


∞
zn ∞
(z − 3)n ∞
(z/2)n 
∞
zn
(a) (b) √ (c) (d)
n=0
(n + 1)3 n=0 n! n=1
n(n + 1) n=1
ln (n + 1)


∞
zn 
∞
(z − 2)2n 
∞
zn 
∞
2
(e) (f) (g) (h) zn
n=0
(n + 1)n n=0
(2n)! n=1
ln ln (n + 1) n=0

√

∞ 
∞
zn ∞
(z − i 2)n 
∞
n!
(i) z (j) (k) (l) n7 z n
n=0 n=0
(n!)1/6 n=0
n! n=0


∞ ∞ ∞ ∞ 
 n
2n (ln n) z n (ln ln n) z n 1−z
(m) (z) (n) (o) (p)
n=0 n=1
n n=2
ln n n=0
1+z

∞ ∞  2 n ∞ ∞
(2z)n z −1 (z/4)n Γ(n − 12 ) n
(q) (r) (s) (t) z .
n=0
n2 + 1 n=0
z2 + 1 n=0
n+2 n=1
Γ(n)

If R is infinite for a power series centered about any finite point z0 , this means
that the function represented by the power series is an entire function. A power series
diverges outside its circle of convergence. On the circle of convergence itself, its behavior
can be quite complicated, and change from one point on the circle to another. The
series may diverge at one or more points on the circle; or it may converge at some

11
points, diverge at others, and oscillate at others; it may even converge at all points.
What is guaranteed, however, is that the function that the power series represents
inside the circle of convergence has at least one singularity on that circle. Remarkably
enough, this remains true even if the power series itself converges absolutely at all
points on the
∞circle of convergence. An example of this phenomenon is provided by
the series n=1 z /n2 . This series converges absolutely at all points on its circle of
n

convergence, which is just the unit circle |z| = 1. (It attains its largest value on the
circle at the point z = 1, where the sum is equal to ζ(2) = π 2 /6.) And yet, the function
represented by the series is singular at z = 1. At the other extreme, we have cases in
which the function represented by the series has a dense setof singularities
∞everywhere
∞ n
on the circle of convergence. Examples of such series are n=0 z and n=1 (z)2 , for
n!

each of which the unit circle |z| = 1 is the circle of convergence. We shall return to
some of these aspects later on.

12
2 Calculus of residues
Cauchy’s integral theorem : As you know, the property of analyticity of a function
f (z) implies that the function is infinitely differentiable, and that each derivative is also
an analytic function. What does analyticity imply for the integral of a function? Once
again, the Cauchy-Riemann condition can be used to arrive at the following conclusion:
Suppose f (z) is analytic in some region R. Then, if z1 and z2 are two points lying in
R, the line integral  z2
dz f (z)
z1

of the function is independent of the actual path between the end-points, provided of
course that the path lies entirely in R. The value of the line integral is dependent only
on the end points.3 As a consequence, the path connecting the end-point z1 and z2 can
be distorted like a flexible and stretchable piece of string, and the value of the integral
does not change.

Cauchy’s integral theorem follows at once : Let C be an oriented closed contour

lying entirely in R. Then 
dz f (z) = 0.
C
It is evident that the contour C may be distorted like a rubber band without changing
the property above, as long as it does not leave R. These properties are responsible for
much of the power of contour integration in evaluating integrals, and hence in solving
a variety of problems that can be reduced to the evaluation of such integrals.

Singularities: We now come to the singularities of an analytic function. These are

the most interesting features of analytic functions of one or more complex variables.
Even in the case of a function of a single complex variable, singularities come in a
remarkable number of varieties. But we shall only be concerned with the simplest (and
most frequently encountered kinds) of singularities.

Removable singularity: The case of a removable singularity may be disposed of

first, with the help of an example. The function
sin z
f (z) =
z
3
This should remind you of the corresponding property of a conservative vector field u(r), namely,
one that can be expressed as the gradient of a scalar field ∇ φ(r). In that case, too, the value of the
line integral of the vector field from r1 to r2 is independent of the actual path between the two points,
and is given by  
r2 r2
u · dr = ∇ φ · dr = φ(r2 ) − φ(r1 ).
r1 r1

13
is not defined at z = 0, and appears to have a singularity at that point. However,
limz→0 f (z) = 1, and we may define f (0) = 1 so as to preserve continuity. This gets
rid of the ‘removable’ singularity and ensures the continuity of f (z) at z = 0. In
all that follows, it will be assumed that functions are defined at such removable sin-
gularities in accord with the limiting values concerned, so that continuity is guaranteed.

Simple pole: The simplest nontrivial singularity is a simple pole. f (z) has a simple
pole at the point z = a if, in a neighborhood of that point, it can be expressed in the
form
c−1 ∞
f (z) = + cn (z − a)n .
(z − a)
   n=0  
singular part regular part

The first term on the right is called the ‘singular part’ of f (z) at z = a. It becomes
infinite at z = a. The coefficient c−1 is a non-zero, finite constant (in general, a com-
plex number) that is called the residue of f (z) at z = a. The notation c−1 is used to
remind us that it is the coefficient of (z − a)−1 in the expansion of f (z) in powers of
(z − a). The ‘regular part’ stands for a function that is analytic in the neighborhood
concerned, and hence is expandable in a convergent power series in (non-negative)
powers of (z − a), in the neighborhood concerned. In some cases the series may, of
course, terminate after a finite number of terms.

The crucial point is that, at a simple pole, the singular part comprises just the
single term c1 (z − a)−1 . This term may or may not be explicit. Consider, for example,
the function (sin z)/z 2 . In the neighborhood of the origin, it has the expansion
sin z 1  (−1)n z 2n+1
∞
1
z2 z4 
= = + − + − . . . .
z2 z 2 n=0 (2n + 1)! z
  3! 5!
 
sing. part reg. part

This function therefore has a simple pole at z = 0, with residue equal to 1. Incidentally,
it has no other singularities in the finite part of the complex plane (remember that sin z
is an entire function). As a second example, consider the function cosec z = 1/(sin z).
Since sin z has simple zeroes at all integer multiples of π, i.e., at z = nπ where n ∈ Z,
it follows that cosec z has simple poles at those points. Using the fact that sin nπ = 0
and cos nπ = (−1)n , the Taylor expansion of sin z near z = nπ is given by
sin z = (−1)n (z − nπ) − 16 (−1)n (z − nπ)3 + . . . .
Hence
(−1)n (−1)n
cosec z = + 16 (−1)n (z − nπ) + . . . = + regular part.
(z − nπ) (z − nπ)
Thus the residue of cosec z at its simple pole at z = nπ is (−1)n .

14
 In general, if f (z) has a simple pole at z = a, then its residue at the pole is given
by 

c−1 = Res f (z) = lim [(z − a) f (z)].
z=a z→a

If f (z) is of the form g(z)/h(z) where g(z) and h(z) are analytic at z = a, and h(z)
has a simple zero at z = a while g(a)
= 0, then f (z) has a simple pole at z = a. We
have 
g(z)  g(a)
f (z) = =⇒ Res f (z) =  ,
h(z) z=a h (a)
where the prime denotes the derivative of the function concerned.

1. Verify the following: for every n ∈ Z ,

(a) cosech z has a simple pole at z = inπ, with residue (−1)n .

(b) sec z has a simple pole at z = (n + 12 )π, with residue (−1)n+1 .

(c) sech z has a simple pole at z = i(n + 12 )π, with residue i (−1)n+1 .

(d) tan z has a simple pole at z = (n + 12 )π, with residue −1.

(e) tanh z has a simple pole at z = i(n + 12 )π, with residue +1.

(f) cot z has a simple pole at z = nπ, with residue +1.

(g) coth z has a simple pole at z = inπ, with residue +1.

Note, in particular, how the residue becomes independent of n in the last four cases.
This property will be useful when we deal with the summation of certain series.

Multiple pole: The next case is that of a pole of higher order, or a multiple pole.
The function f (z) has a pole of order m at z = a if, in the neighborhood of that point,
it can be expressed in the form

c−m c−1 
∞
f (z) = +···+ + cn (z − a)n ,
(z − a)m (z − a) n=0
     
singular part regular part

where m is a positive integer. The residue of f (z) at the multiple pole is still the
coefficient c−1 . The significance of the residue will become clear shortly, when we
consider Cauchy’s integral theorem. For the present, note that the residue at a multiple
pole of order m is given by
 dm−1
 1
c−1 = Res f (z) = lim [(z − a)m f (z)].
z=a z→a (m − 1)! dz m−1

15
This formula is a generalization of that for the residue at a simple pole.

Essential singularity; Laurent series: Going a step further, the singular part may
involve all negative integral powers of (z − a). The function f (z) has an (isolated)
essential singularity at z = a if, in the neighborhood of that point (see below), it
can be expressed in the form

∞
c−n 
∞
f (z) = + cn (z − a)n .
(z − a) n
n=1   n=0  
singular part regular part

Once again, the coefficient c−1 is the residue of f (z) at the singularity. A representation
such as the foregoing in positive as well as negative powers of (z−a) is called a Laurent
series. The question now arises as to the region in which such a representation is valid.
The regular part, as we know, is convergent inside some circle of convergence centered
at z = a, and of radius r1 , say. By setting w = (z − a)−1 , it follows that the singular
part is convergent inside some circle of convergence in the complex w-plane that is
centered at the origin. In other words, it is convergent outside some circle centered at
z = a, of radius r2 , say. Then, provided r2 < r1 , there is an annular overlap region
of inner radius r2 and outer radius r1 , in which both the infinite series are absolutely
convergent. This annulus is the region in which the Laurent series is valid. It may so
happen, of course, that r2 = 0. Then the Laurent series is valid in the punctured disc
of radius r1 centered at z = a, with only the point z = a left out. For example, the
function
∞
1
1/z
e = + 
1
n! z n
  
n=1 reg. part
sing. part

has an essential singularity at the origin z = 0. In this case r1 = ∞ while r2 = 0.

Hence the Laurent series on the right-hand side of the equation above is valid for all
z
= 0, including the point at infinity in the z-plane. That is, it is valid in the region
|z| > 0. In general, a Laurent series is convergent (i.e., provides a valid representation
of an analytic function) in some annular region.

Singularity at infinity: What about possible singularities at the point at infinity,

i.e., at z = ∞? The nature of such a singularity, if any, can be deduced by first
changing variables to w = 1/z. This maps the point z = ∞ to the point w = 0. The
singularity of the corresponding transformed function f (1/w) ≡ φ(w) at the origin
w = 0 then determines the nature of the singularity of f (z) at z = ∞. For example, if
f (z) = z itself, then φ(w) = 1/w, which has a simple pole at w = 0. Hence f (z) = z
has a simple pole at z = ∞. Likewise, it follows that:

— A polynomial of order n in z has a pole of order n at z = ∞.

16
 — The function ez has an essential singularity at z = ∞. Hence each of the functions
cos z, sin z, cosh z and sinh z also has an essential singularity at ∞.

The function

∞
1 ∞
zn
e1/z + ez = + 1 +
n=1
n! z n n=0
n!
has essential singularities at z = 0 as well as z = ∞. The Laurent series on the right-
hand side of the equation above is valid in the annular region 0 < |z| < ∞.

The residue at infinity of a function f (z), however, is a little different from the
usual idea of a residue at a singularity at any point in the finite part of the complex
plane. I’ll return to this question shortly.

Accumulation point of poles: We have seen that the function cosec z has simple
poles at all integer multiples of π, i.e., at the set of points {z = nπ, n = 0, ±1, ±2, . . .}.
All these points lie on the x-axis. What is the nature of the singularity, if any, of cosec z
at z = ∞? The answer is provided by considering the Riemann sphere. Recall that
the mapping from the complex plane to the Riemann sphere is given by

z + z∗ z − z∗ |z|2 − 1
ξ1 = , ξ2 = , ξ3 = .
|z|2 + 1 i(|z|2 + 1) |z|2 + 1
The positive and negative segments of the x-axis on the complex plane map onto the
lines of longitude 0◦ and 180◦, respectively, on the Riemann sphere. Hence the poles
of cosec z lie on these lines of longitude, at the coordinates

2nπ n2 π 2 − 1
ξ1 = , ξ2 = 0 , ξ3 = , where n ∈ Z.
n π2 + 1
2 n2 π 2 + 1
Except for the south pole (corresponding to n = 0), all the poles lie in the northern
hemisphere (since π is larger than unity). It is evident that they get more and more
crowded together as n → ±∞, and approach the point (0 , 0 , 1), i.e., the north pole.
They become dense on the two longitude lines in an infinitesimal neighborhood of the
north pole. On the extended complex plane, z = ∞ is therefore an accumulation
point of poles. The singularity is no longer an isolated singularity.

Meromorphic function: An analytic function whose only singularities (if any) in the
finite part of the complex plane are isolated poles is called a meromorphic function.
Entire functions are trivially also meromorphic (as they have no singularities at all in
the finite part of the complex plane). Any rational function of z, i.e., a ratio of two poly-
nomials in z, is a meromorphic function. So are the functions cosec z, sec z, cosech z
and sech z. The function cosec (1/z) is not meromorphic, because it has an accumula-
tion point of poles at z = 0.

17
2. Identify the regions of the extended complex plane in which the functions listed be-
low are holomorphic. Also find the location, order and residue of each pole or essential
singularity of the functions, including the nature of the singularity (if any) at z = ∞.

2
(a) e−z (b) (1 − z 100 )/(1 − z) (c) (2z − 1)/z
(d) coth z − z −1 (f) (1 − cos z)/z 2
z
(e) (e)e
(g) 1/(ez + 1) (h) z cosec z (i) z cosec2 z
(j) (sin z − z cos z)/z 4 (k) 1/(2z − 1) (l) cosec (z 2 )
(m) 1/(sin z − cos z) (n) 1/(cos z − cosh z) (o) 1/(3z − 2z )
2
(p) 1/(ez − 1) (q) 1/(e2z − ez ) (r) exp [1/(z 2 − 1)]

∞
(s) cos (1/z) (t) (ez − 1)−2 (u) e−nz
n=0

∞ 
∞ 
∞
(v) z n /nn (w) z n /(n!)2 (x) [(1 − z)/(1 + z)]n
n=1 n=0 n=0

3. Obtain the Laurent series expansions of the following functions, valid in the regions
indicated. In each case, the expansion is understood to be in powers of (z − a), where
a is the center of the annular region concerned.

(a) (z + z −1 )1000 (0 < |z| < ∞) (b) z 2 e1/z (0 < |z| < ∞)
−1/z 2
(c) z −1 e (0 < |z| < ∞) (d) (z 2 + z −2 )100 (0 < |z| < ∞)
(e) e1/z (1 − z)−1 (0 < |z| < 1) (f) z 2 e1/(z−2) (0 < |z − 2| < ∞)
(g) 21/(z−1) z −2 (0 < |z − 1| < 1) (h) (z − 1)−1 (z − 2)−2 (0 < |z − 1| < 1).


Cauchy’s integral formula : Consider the contour integral C z n dz where n =
0, 1, . . . and C is a simple closed contour that encircles the origin once in the posi-
tive sense. As z n is an entire function, C may be distorted to a circle of arbitrary
radius r centered at the origin, and encircling it once in the positive sense. Since
dz = r eiθ i dθ on this circle, we have
  2π
n n
z dz = i r dθ e(n+1)iθ = 0.
C 0

The vanishing of the contour integral is in any case immediately evident, even without
writing it out in the form above: the integrand z n is holomorphic everywhere inside
and on the contour C, so that C can be distorted till it shrinks to a point (i.e., you

18
can let r → 0), and the integral vanishes.

Now consider the contour integral C dz/z n+1 where n = 0, 1, . . . and C is again
a simple closed contour that encircles the origin once in the positive sense. As the
integrand z −n−1 has a pole of order (n + 1) at the origin, and no other singularities, C
can again be deformed to a circle of some radius r centered at the origin. Then
  2π
dz i
n+1
= n dθ e−niθ .
C z r 0

As before, this integral vanishes identically an all cases except one: namely, when n = 0.
(Alternatively, we can now expand the circle outwards, and pass to the limit r → ∞.)
Only when n = 0, and in that case alone, do we get  the nonzero value C dz/z = 2πi.
−1
Hence we have, for all integers n ∈ Z, (2πi) C
dz/z n+1
= δn , 0 . A little more
generally, if a is any complex number and C is a closed contour encircling a once in
the positive sense, then 
1 dz
= δn , 0 .
2πi C (z − a)n+1
This is a most useful representation of the Kronecker delta. Cauchy’s Integral Theorem
and the whole of the so-called calculus of residues is based upon it.

Cauchy’s integral formula follows from the result above. Omitting the details,
the theorem states : Let C be a closed contour lying entirely in a region in which the
function f (z) is analytic, except for isolated poles and essential singularities at the
points zk , k = 1, 2, . . . . No singularities must lie on C itself. Then, if C encircles the
singularities once in the positive sense,
 
dz f (z) = (2πi) Res f (z).
C {poles zk }

The sign of the right-hand side is reversed if C encircles the singularities in the negative
or clockwise sense. If a singularity is encircled r times, the contribution to the integral
is ±r times the residue at that singularity, the negative sign being applicable when the
contour winds around the singularity in the negative sense. (r is called the winding
number of the contour at the singularity.)

4. Contour integration : Evaluate the following integrals by contour integration.

Here a , b , etc. denote positive constants, and n is a non-negative integer. (The values
of the integrals have also been given.)
 ∞
dx π
(a) 2 2 2 2
= .
−∞ (x + a )(x + b ) ab(a + b)

19
Consider the integral from −R to R on the x-axis, and attach a semicircle of radius R
to close the contour either in the upper or lower half-plane. You now have the integral
of the function [(z 2 + a2 )(z 2 + b2 )]−1 over a closed contour C. By Cauchy’s theorem,
this integral is (apart from a factor ±2πi) the sum of the residues of the function at the
poles enclosed by C. Let R → ∞. The contribution from the semi-circle vanishes as
R → ∞, while the integral along the x-axis becomes the integral you want to evaluate.

 ∞
dx x2 π
(b) 2 2 3
= .
0 (x + a ) 16a3

Write the integral as 12 times the integral from −∞ to ∞ along the x-axis, and proceed
as in the preceding case. Note that the pole enclosed by C is a multiple pole in this case.

 ∞
dx x sin x
(c) 2 2
= π e−a .
−∞ (x + a )

In this case, the contour can’t be closed in either the upper or the lower half-plane
because sin z is a linear combination of eiz as well as e−iz , and one of the exponentials
will diverge as |z| → ∞ in each of the two half-planes.
 ∞ Therefore first write sin x as the
imaginary part of e , and consider the integral −∞ dx x e /(x + a2 ). It is clear that
ix ix 2

you can now close the contour in the upper half-plane. After evaluating the integral,
take its imaginary part. Alternatively, retain sin x in the integrand, and note that you
have to close the contour in different half-planes for the two exponential factors. Do
the integral both ways to check that the final results match.

 ∞
dx (x2 − x + 2) 5π
(d) 4 2
= .
−∞ x + 10x + 9 12
 ∞
dx sin x π
(e) = .
0 x 2

Extend the integral to one from −∞ to ∞, and consider the integrand eiz /z over a
contour that is closed in the upper half-plane. The pole at the origin must be avoided
by a small semi-circular indentation around z = 0. The rest of the integral along the
real axis is now a Cauchy principal value integral. After evaluation of the integral,
equating imaginary parts yields the result sought. Note that the principal value re-
striction can be dropped at this stage, because (sin x)/x does not have a singularity
at x = 0.

 2π
dθ 2π
(f) =√ (a > b) .
0 a − b cos θ a2 − b2

20
Let z = eiθ . Then cos θ = 12 (z + z −1 ). The integration over θ becomes an integral over
z around the unit circle traversed once in the positive sense. Pick up the residue at
the pole of the integrand that lies inside the unit circle.

 2π
2π
(g) dθ ecos θ cos (nθ − sin θ) = (n = 0, 1, . . .) .
0 n!
 2π
The integral required is the real part of 0 dθ ecos θ ei(nθ−sin θ) . Set eiθ = z to convert
the integral to one around the unit circle in the z-plane. Pick up the residue of the
integrand at the singularity enclosed by the contour.

Residue at infinity : The residue of an analytic function f (z) at infinity is defined

via a contour integral in the same way that it is defined at any finite point z = a,
namely, as (2πi)−1 times the contour integral of the function over a sufficiently small
circle enclosing (in the positive sense) no singularities other than the possible one at
z = a. In the case of the point at infinity, we must do this on the Riemann sphere:
the contour is a sufficiently small circle going around the point at infinity once in the
negative sense. On the complex plane, this contour looks like a circle C of sufficiently
large radius R traversed in the negative sense. On the other hand, this circle now
encloses all the singularities of f (z) located in the finite part of the plane. Therefore
 
def. 1
Res f (z) = dz f (z) = − Res f (z),
z=∞ 2πi C j
z=aj

where the sum runs over the singularities at the points aj in the finite part of the
complex plane. Two noteworthy points follow at once.
(i) f (z) may have a nonzero residue at z = ∞ even if it is not singular at the point
at infinity!
For instance, the function (z − 1)−1 + (z − 2)−1 has simple poles at z = 1 (residue = 1)
and z = 2 (residue = 1), and is regular at z = ∞. And yet it has a nonzero residue
at infinity that is equal to the negative of the sum of its residues at z = 1 and z = 2,
namely, −2.
(ii) On the other hand, an entire function must have zero residue at infinity, even
though it is, in general, singular at that point!
Thus, a polynomial of order n, which has a pole of order n at z = ∞, must have
zero residue at that point. Likewise, an exponential such as ez , which has an essential
singularity at ∞, must have zero residue at that point.

There is an equivalent way of finding the residue at infinity. Changing variables

from z to w = 1/z converts the contour integral over the circle C of radius R to a

21
circle c around the origin in the w-plane, with an infinitesimal radius 1/R, traversed
in the positive sense. Remembering that dz = −dw/w 2, this yields the formula

1 dw
Res f (z) = − f (1/w).
z=∞ 2πi c w 2

It follows that the residue at infnity of f (z) is the coefficient of w −1 in the Laurent
expansion of −(1/w 2)f (1/w) in powers of w. It is now easily checked that the residue
at infinity of (z − 1)−1 + (z − 2)−1 is indeed equal to −2, as found earlier. Similarly, it is
easy to verify that there is no term proportional to w −1 in the power series expansion
of −(1/w 2 )f (1/w) when f (z) is any polynomial or exponential like ez .

5. A practical use of the concept of the residue at infinity is in the evaluation of

certain contour integrals involving rational functions. Let pn (z) = an z n + . . . + a0 be a
polynomial of order n, and let qn+1 (z) = bn+1 z n+1 + . . . + b0 be a polynomial of order
(n + 1). Let C be a simple closed contour that encloses all the roots of qn+1 (z) = 0
once in the positive sense. Show that

pn (z) 2πian
dz = .
C qn+1 (z) bn+1

Note that this result does not depend on the details of the locations or the multiplici-
ties of the zeroes of qn+1 (z).

6. Summation of series using contour integration: The fact that π cot πz has a
simple pole at each integer z = n, with residue equal to +1, can be used to sum certain
infinite series. Similarly, the fact that π cosec πz has simple poles at all integers z = n,
with residue equal to (−1)n , can be used to sum certain series in which the signs of
the terms alternate. The method works when the summand is an even function of n,
so that the series can be written as a sum over both positive and negative integers.
Evaluate the following infinite sums by the contour integration method. (The values
of the sums are also given, so that you can check your answers.)


∞  
1 π 1
(a) S(a) = 2 2
= coth πa − (a > 0).
n=1
n +a 2a πa

First write the sum required as 12 times the sum over positive as well as negative integers
n. Each terms in the sum is just the residue of the function f (z) = π cot (πz)/(z 2 + a2 )
at its simple pole at z = n. Hence it is (2πi)−1 times the contour integral of f (z) over a
small circle encircling z = n in the positive sense. Merge all these little circles together,
to get two ‘hairpin’ contours straddling the x-axis from −∞ to −1 and from 1 to ∞,
respectively. These can be made part of a single closed contour by attaching large
semi-circles in the upper and lower half-planes. This contour now encircles just three

22
poles of f (z), namely, those at z = 0 and z = ±ia, but in the clockwise or negative
sense. Pick up the residues at these poles and simplify the resulting expression to
obtain the final result given above.

∞
1 π2
(b) = .
n=1
n2 6

This sum is equal to ζ(2), where ζ(s) = ∞ s
n=1 1/n is the Riemann zeta function.
The same method as that used in part (a) above will lead to the result quoted. Note
that there is a pole of order 3 at z = 0 in this instance.

Verify the result above by passing to the limit a → 0 in the sum S(a) found in part
(a). You will need to use the fact that cot z = z −1 + 13 z + O(z 3 ) in the neighborhood
of z = 0.

1
(c) Use the fact that ζ(2) = 6
π 2 to show (by elementary means) that


∞
(−1)n−1 π2 
∞
1 π2
= and = .
n=1
n2 12 n=1
(2n − 1)2 8


∞  
(−1)n−1 π 1 1
(d) 2 2
= − .
n=1
n +a 2a πa sinh πa

Repeat the procedure used in part (a), but with cosec (πz) instead of cot (πz) in the
definition of the integrand f (z). Once again, verify the result you obtain by noting
that
∞
(−1)n−1 ∞
1
2 2
= S(a) − 2 S( 2 a) where S(a) =
1 1
.
n=1
n +a n=1
n + a2
2


Finally, pass to the limit a → 0 to check that ∞ n=1 (−1)
n−1 1 2
/n2 = 12 π .


∞  
1 π πa 2
(d) 2 2 2
= 3 coth πa + 2 − .
n=1
(n + a ) 4a sinh πa πa

Use the same procedure as in part (a) to establish this result. Verify the answer directly
from S(a) by observing that


∞
1 1 dS(a)
=− .
n=1
(n2 2
+a ) 2 2a da

23
 (e) Pass to the limit a → 0 in the foregoing result to establish that
∞
1 π4 
∞
(−1)n−1 7π 4
ζ(4) = 4
= , and hence = .
n=1
n 90 n=1
n4 720


∞
1 π 1
(f) = (b coth πa − a coth πb) − .
n=1
(n2 + a2 )(n2 + b2 ) 2ab(b2 − a2 ) 2a2 b2

Verify the result by noting that the sum required is [S(a) − S(b)]/(b2 − a2 ).


∞
1
(g) Find the sum (α > 0).
n=1
n4 + α4

It is evident that we must now consider the contour integral of the function π cot (πz)/(z 4 +
α4 ). Check your answer by writing (n4 + α4 ) as (n2 + a2 )(n2 + b2 ), where a = α eiπ/4
and b = α e3πi/4 = −α e−iπ/4 , and using (the analytic continuation of) the result of
part (f).

7. Solving difference equations using contour integration: Solve the following

recursion relations (that is, find cn as a function∞ of n). In each case, first define
n
the corresponding generating function f (z) = n=0 cn z . Use the recursion relation
to obtain the generating function in closed form. It will, in general, turn out to be
a rational function in the cases to be considered. The infinite series defining f (z)
converges absolutely inside its circle of convergence, and defines an analytic function
of z in that region. The inversion formula is of course cn = (1/n!)[dn f /dz n ]z=0 . But
what we need here is the more convenient formula

1 f (z) dz
cn = ,
2πi C z n+1
where C is a simple closed contour enclosing (in the positive sense) only the pole of
the integrand at the origin. Evaluating the residue at this (n + 1)th order pole gets you
back to the original formula for cn . But you can also open out the contour to infinity
in all directions, picking up the residues at the poles of f (z) (with a minus sign). This
will yield the expression sought for cn much more easily.

(a) cn+2 − cn+1 − cn = 0 , n ≥ 0, c0 = 1 , c1 = 1.

The set {cn } in this case is a Fibonacci sequence of numbers. (Each member of the
sequence is the sum of the preceding two numbers.) The answer for cn is, in this case,
1 √ √ 
cn = √ ( 5 + 1)n − (−1)n ( 5 − 1)n , n = 0, 1, . . . .
2n 5

24
 √
As you know, the irrational number τ = 12 ( 5 + 1) is called the golden mean, and
has a large number√of interesting properties. So do the Fibonacci numbers, since
cn = [τ n − (−τ )−n ]/ 5. The result above shows that
• the asymptotic (large-n) growth of the Fibonacci sequence is exponential, ∼ eλn ,
where λ = ln τ  0.4812 .
(b) cn+2 − cn+1 − 2cn = 0 , n ≥ 0, c0 = 1 , c1 = 2.

(c) cn+2 − 2cn+1 + cn = 0 , n ≥ 0, c0 = 1 , c1 = 1. (This recursion relation can be

solved by inspection!)
(d) cn = c0 cn−1 + c1 cn−2 + c2 cn−3 + · · · + cn−1 c0 , c0 = 14 .
Unlike the preceding examples, the last case is a nonlinear difference equation. How-
ever, it can be solved quite easily by elementary means. Define the generating function
for cn and obtain a functional equation for it (by examining its square!) Use this to
show that
(2n)!
cn = n+1 .
4 (n + 1) (n!)2
When c0 = 1, we find  
(2n)! 1 2n
cn = 2
= .
(n + 1) (n!) n+1 n
Hence c0 = 1, c1 = 1, c2 = 2, c3 = 5, c4 = 14, c5 = 42, . . . . These are called Catalan
numbers. They appear in an enormous number of combinatorial problems.

Mittag-Leffler expansions of meromorphic functions: We have seen that ana-

lytic functions can be represented by Taylor series in their regions of holomorphy. In
the neighborhood of poles and isolated essential singularities, Taylor series are replaced
by Laurent series.

A further generalization is possible for meromorphic functions. Recall that a mero-

morphic function is one that has, at best, only poles in the finite part of the complex
plane. Suppose we are given the locations of all the poles of such a function, and the
singular parts at these poles. Can we find a representation for the function that in-
volves a ‘sum over poles’ ? The Mittag-Leffler expansion of a meromorphic function
provides such a representation. In general, suppose the meromorphic function f (z) has
poles at the points aj (j = 1, 2, . . .), with singular parts sj (z). For example, if aj is a
pole of order m, then the singular part4 is
(j) (j)
c−m c−1
sj (z) = +···+ .
(z − aj ) m (z − aj )
4
The singular part is also called the principal part, but I prefer the former term because it is
more suggestive.

25
The Mittag-Leffler expansion of f (z) is then of the form

f (z) = sj (z) + g(z),
j

where g(z) is an entire function. When the number of poles is finite, the sum over
j presents no problem. This is the case when f (z) is a rational function, i.e., the
ratio of two polynomials. In such cases the Mittag-Leffler expansion is nothing but
the resolution of the function into partial fractions. For instance, the Mittag-Leffler
expansion of the function f (z) = (z − a)/(z − b)2 is just

z−a b−a 1
= + .
(z − b) 2 (z − b) 2 z−b

When the summation over j involves an infinite number of terms, however, the ques-
tion of the convergence of the sum arises.

Mittag-Leffler expansion of cot πz: We have already encountered an example of

a Mittag-Leffler expansion that involves an infinite number of terms, or at least an
expansion that is related to such an expansion. Recall that
 π
1 
∞
1
S(a) = = coth πa − .
n=1
n2 + a2 2a πa

Although the infinite series was originally summed for real positive values of a, we may
regard the equality above as a relation between two analytic functions of the complex
variable a, by virtue of the principle of analytic continuation. Set a = iz and note
that coth (iπz) = −i cot (πz). The relation above can be written as

1 ∞
z ∞
z
π cot (πz) = + 2 = .
z n=1
z −n
2 2
n=−∞
z − n2
2

We’re almost there, because this expansion seems to express the meromorphic function
π cot (πz) as a sum over the singular parts at its poles (which are located at all integer
values of z). But the poles of cot (πz) are simple poles, whereas the summand in the
last equation involves a quadratic function of z in the denominator. It is a simple
matter to write ∞  
 ∞
2z  1 1
2 − n2
= + .
n=1
z n=1
z − n z + n
But we cannot go on to split the right-hand side into two sums and write it as


∞
1 
∞
1 
∞
1 
−1
1
+ = + ,
n=1
z − n n=1 z + n n=1 z − n n=−∞ z − n

26
because each of the two individual infinite series in the last line above diverges! To
find the correct Mittag-Leffler expansion, let’s go back to

∞
z 1 
∞
z
π cot (πz) = = + .
n=−∞
z −n
2 2 z n=−∞ z − n2
2
n=0

But
∞

   1 
∞ ∞
z 1 2z 1 1
= = +
n=−∞
z − n2
2 2
n=−∞
z − n2
2 2
n=−∞
z−n z+n
n=0 n=0 n=0
∞

 1 1
1 1
= 1
+ + −
2
n=−∞
z−n n z+n n
n=0
∞

 1 1
= + ,
n=−∞
z−n n
n=0

where we have used (in writing the last line) the fact that the sum over all nonzero
integers remains unchanged if n → −n. Hence we obtain
∞

1  1 1
π cot (πz) = + + .
z n=−∞ z − n n
n=0

This is the Mittag-Leffler expansion of the meromorphic function π cot (πz). Note
that the term 1/n inside the large brackets in the summand on the right-hand side
is necessary. Without it, the sum over n of 1/(z − n) would diverge. With the 1/n
term present, the summand is sufficiently well-behaved5 to guarantee convergence of
the series.

Moreover, the series can be differentiated term by term with respect to z. The
result is the Mittag-Leffler expansion of the function cosec2 (πz), which reads

π2 ∞
1
= .
sin πz n=−∞ (z − n)2
2

Now step back and admire this remarkably beautiful formula!

5
It falls off like 1/n2 as |n| → ∞.

27
3 Linear response; dispersion relations
Causal, linear, retarded response: Under very general conditions, the physical re-
sponse of a system to an applied time-dependent stimulus satisfies certain basic criteria.
Let us denote the stimulus (or ‘force’) by F (t), and the response by R(t). Examples of
such stimulus-response pairs are : mechanical force and displacement; electric field and
polarization; magnetic field and magnetization; stress and strain; electromotive force
and current; and so on. For notational simplicity, I have suppressed the appropriate
indices in the stimulus and the response when these are vectors, tensors, etc. in what
follows.

Under fairly general conditions, the response at any instant of time t may be as-
sumed to meet three basic requirements:

(i) Causality: The response at time t depends on the force history at all earlier
instants of time, but not later ones. This is the principle of causality, which says that
the effect cannot precede the cause.

(ii) Linearity: The response is assumed to be linear in the applied force. This implies
at once that the superposition principle holds good.

(iii) Retarded response: The effect of a force F (t  ) applied at the instant t  on the
response R(t) at a later instant t depends only on the elapsed time interval (t − t  ),
rather than on both t and t  individually. This means that the location of the actual
origin of the time coordinate, the particular instant at which we set t = 0, does not
matter.

Taking the force to be applied from t = −∞ onward (this will automatically incorporate
all other starting times), the most general linear functional of F (t) subject to these
requirements is given by
 t
R(t) = dt  φ(t − t  ) F (t ).
−∞

Note that the upper limit of integration over t  is t, in accordance with the require-
ment of causality. The quantity φ(t − t  ) is called the response function. The fact
that it is a function of the time difference (t − t  ) indicates that we are dealing with a
retarded response. The quantity φ(t) represents the ‘weight’ with which a force applied
at t = 0 contributes to the response R(t) at any later time t. We may expect it to be
a decreasing (or at least a non-increasing) function of its argument, such as a decaying
exponential. But it could oscillate as a function of t, although we would expect it to
do so with decreasing amplitude, in most instances. These are not strict requirements,
of course, but they seem to be plausible on physical grounds.

28
Dynamic susceptibility: It is natural to decompose general time-dependent func-
tions such as F (t) and R(t) into their Fourier (or frequency) components. We have
 ∞  ∞
−iωt  
F (t) = dω e F (ω) and R(t) = dω e−iωt R(ω).
−∞ −∞

The inverse transforms are

 ∞  ∞
1 1
F(ω) = iωt 
dt e F (t) and R(ω) = dt eiωt R(t).
2π −∞ 2π −∞

1. Use these relations in the foregoing relation between R and F to show that
 ∞

R(ω) = χ(ω) F(ω), where χ(ω) = dt eiωt φ(t).
0

The function χ(ω) is called the frequency-dependent susceptibility or dynamic

susceptibility (corresponding to this particular stimulus-response pair). It measures
the response of the system to a sinusoidally varying ‘force’ of unit amplitude and fre-
quency ω. Observe that χ(ω) is not the Fourier transform of φ(t). The lower limit
of integration in its definition is t = 0 rather than −∞. It is important to note that
this lower limit is a direct consequence of the fact that t is the upper limit of inte-
gration in the basic expression for R(t). In turn, this follows directly from causality.
Thus, the dynamic susceptibility is the one-sided Fourier transform (sometimes also
called the Fourier-Laplace transform) of the response function φ(t) because of causality.

Physical examples of the dynamic susceptibility include the AC susceptibility (or

frequency-dependent susceptibility) of a magnetic substance, the electric polarizability
of a dielectric material, the dynamic mobility of a fluid, the elastic compliances (in-
verses of the elastic moduli) of a solid, and so on.

Some symmetry properties of the dynamic susceptibility: An important prop-

erty of the dynamic susceptibility can be extracted from its defining equation above :
since φ(t) is a real quantity, we find that

χ(−ω) = χ∗ (ω) for all real ω,

where ∗ denotes the complex conjugate. I have added the phrase ‘for all real ω’ has
been added because it makes sense, as you’ll see shortly, to regard of χ(ω) as an
analytic function of the complex variable ω, although the physically accessible values
of the frequency are of course real and non-negative. It follows at once that, for all
real ω,
Re χ(−ω) = Re χ(ω) and Im χ(−ω) = −Im χ(ω).
The real part of the dynamic susceptibility is an even function of the frequency, while
the imaginary part is an odd function of the frequency. You can, of course, write down

29
these properties directly by setting eiωt = (cos ωt + i sin ωt) in the definition of χ(ω).

Consider the formula for χ(ω) as the one-sided Fourier transform of φ(t). We have
assumed that the integral exists for physical (i.e., real, non-negative) values of the
frequency ω. It is obvious that it will continue to do so even if ω is complex, pro-
vided Im ω is positive. This is because the factor eiωt then leads to an extra damping
factor e−(Im ω) t . Such a factor can only improve the convergence of the integral, since
the integration over t is restricted to non-negative values. We may therefore conclude
that the dynamic susceptibility can be analytically continued to the upper half-plane6
(UHP for short) in ω. The defining expression for χ(ω) provides a representation for
this analytic function for all Im ω ≥ 0.

Given that the dynamic susceptibility is analytic in the UHP, we can extend the
relation χ(−ω) = χ∗ (ω) (that holds good for real ω) to complex values of ω. We have

χ(−ω ∗ ) = χ∗ (ω), Im ω ≥ 0.

Note that if ω lies in the UHP, so does −ω ∗ . Hence the arguments of the functions on
both sides of the equation above do lie in the region in which we are guaranteed that
the dynamic susceptibility is analytic. On general grounds, and without further input,
we really cannot say very much about its possible behavior in the lower half of the
complex ω-plane. But we know that an analytic function of a complex variable cannot
be holomorphic at all points of the extended complex plane unless it is just a constant,
as we know from Liouville’s Theorem. In general, therefore, the dynamic susceptibility
will have singularities in the lower half-plane in ω.

Dispersion relations : The analyticity of χ(ω) in the UHP enables us to derive cer-
tain relations between its real and imaginary parts. Let us assume that |χ(ω)| → 0
as ω → ∞ along any direction in the UHP. This is a physically plausible assumption
to make in most circumstances, for the following reason. Even for real ω, we expect
the susceptibility to vanish as the frequency becomes very large, because the inertia
present in any system will not permit it to respond to a sinusoidal applied force os-
cillating at a frequency much higher than all the natural frequencies present in the
system. And when ω is a complex number with a positive imaginary part, the factor
eiωt in the formula for the susceptibility lends an extra damping factor e− (Im ω) t . The
assumption is thus a reasonable one to make. However, it is not absolutely essential,
6
Causality ensures that the Fourier-Laplace transform of a causal response function is analytic in a
half-plane in ω. Whether this is the upper half-plane (UHP) or lower half-plane (LHP) depends on the
Fourier transform convention chosen. With our particular convention in which a function f (t) of the
∞
time is expanded as −∞ dω f(ω) e−iωt , it is the upper half of the ω-plane in which the susceptibility
is analytic. Had we chosen the opposite sign convention and used the factor e+iωt in the expansion
above, the region of analyticity would have been the lower half-plane in ω. The convention I have
adopted is the more commonly used one, at least in physics.

30
and can be relaxed, as you’ll see further on.

Let ω be a fixed, real, positive frequency. Consider the quantity

 χ(ω  )
f (ω ) = 
ω −ω
as a function of the complex variable ω  . This function is analytic everywhere in the
upper half-plane in ω  , as well on the real axis in that variable, except for a simple
pole at ω  = ω located on the real axis. By Cauchy’s integral theorem its integral
over any closed contour C  lying entirely in the upper half-plane is identically equal to
zero Without changing the value of the integral (namely, zero), the contour C  can be
expanded to the contour C that comprises the following:
(i) a large semicircle of radius R in the UHP,
(ii) a line integral on the real axis running from −R to ω − ,
(iii) a small semicircle, from ω −  to ω + , lying in the UHP so as to avoid the simple
pole of the integrand, and, finally,
(iv) a line integral from ω +  to R.
Thus  
  χ(ω  )
dω f (ω ) = dω  = 0.
C C ω−ω
In the limit R → ∞, the contribution from the large semicircle vanishes, because f (ω  )
vanishes faster than 1/ω  as ω  → ∞ along all directions in the UHP: this is ensured by
the fact that χ(ω  ) → 0 as |ω  | → ∞. On the small semicircle, we have ω  = ω +  eiθ ,
where θ runs from π to 0. Therefore, in the limit  → 0, the contribution from the
small semicircle tends to −iπχ(ω). Hence we get
 ω−
  ∞ 

 χ(ω )  χ(ω )
lim dω + dω − iπχ(ω) = 0.

→0 −∞ ω−ω ω+
ω−ω
But the limit on the left-hand side of this equation is just the Cauchy principal value
integral of the integrand χ(ω  )/(ω  − ω). Recall that the Cauchy principal value is a
specific prescription for avoiding the divergence or infinity that would otherwise arise,
owing to the singularity (a simple
 pole) of the integrand at ω  = ω. Denoting this
principal value integral by P (· · · ), we get
 ∞
i χ(ω  )
χ(ω) = − P dω   .
π −∞ ω −ω
Note that this equation expresses the susceptibility at a real frequency as a certain
weighted sum of the susceptibility over all other real frequencies. No complex frequen-
cies appear anywhere in this formula. We made an excursion into the upper half of

31
the complex ω-plane. This was made possible by the analyticity properties of the sus-
ceptibility. But we have returned to the real axis, bringing back the last equation with
us. Equating the respective real and imaginary parts of the two sides of this equation,
we get
 ∞   ∞ 
1  Im χ(ω ) 1  Re χ(ω )
Re χ(ω) = P dω and Im χ(ω) = − P dω .
π −∞ ω−ω π −∞ ω −ω
These formulas are called dispersion relations (the origin of the term is given below).
They imply that the two real functions of a real argument, Re χ(ω) and Im χ(ω) , form
a Hilbert transform pair.

It should also be evident that the dynamic susceptibility cannot, in general, be a

real analytic7 function of ω. For, if that were so, the imaginary part Im χ(ω) would
be zero for all real frequencies, and then so would the real part Re χ(ω). Thus χ(ω)
would itself have to vanish identically.

2. The dispersions relations above still involve integrals over negative as well as pos-
itive frequencies, whereas physically accessible frequencies are non-negative. But the
symmetry properties of Re χ(ω) and Im χ(ω) can be used to restrict the range of
integration to physically accessible frequencies, namely, 0 ≤ ω  < ∞. Show that
 ∞    ∞ 
2  ω Im χ(ω ) 2ω  Re χ(ω )
Re χ(ω) = P dω and Im χ(ω) = − P dω .
π 0 ω  2 − ω2 π 0 ω  2 − ω2

The term ‘dispersion relation’ originates from the fact that they were first derived
in the context of optics : the real and imaginary parts of the frequency-dependent re-
fractive index of an optical medium are, respectively, measures of the dispersion and
absorption of radiation by the medium. These are not independent quantities, but are
related to each other via dispersion relations, which are also called Kramers-Kronig
relations in physics. As I have already mentioned, they are a direct consequence of
the principle of causality as applied to response functions.

Admittance of an LCR circuit : Consider a series LCR circuit. If the applied emf
is V (t), the current I(t) is given by the equation

dI 1 t
L + RI + I(t  ) dt  = V (t).
dt C −∞
Now use the Fourier expansions
 ∞  ∞
−iωt  
V (t) = dω e V (ω) and I(t) = dω e−iωt I(ω).
−∞ −∞
7
A function g(z) is a real analytic function of z = x + iy if it is (i) an analytic function of z,
and (ii) real when its argument is real, i.e., g(x) is real.

32
Inserting these in the equation for I(t), we find that the last term on the left-hand side
becomes (after a change of integration variable from t  to τ = t − t  )
  ∞
1 ∞ −iωt 
dω e I(ω) dτ eiωτ .
C −∞ 0

The problem is that the last integral above diverges, i.e., it is formally infinite. We
should really have used Laplace transforms instead of Fourier ∞transforms to tackle the
iωτ
problem under ∞consideration! Had we done so, the integral 0 dτ e would have been
−sτ
replaced by 0 dτ e , with s in a region such that Re s large enough to guarantee
the convergence of the integral involved. In this case this is simply Re s > 0, and
the value of the integral
∞ is just 1/s. Let’s suppose that this has been done, so that
we may interpret 0 dτ eiωτ as standing for the analytic continuation to s = −iω of
the corresponding Laplace transform. This gives the value 1/(−iω) = i/ω for the
apparently divergent integral. With this small technicality taken care of, the Fourier
transforms of the current and the voltage are related to each other according to
 
i 
−iωL + R + I(ω) = V (ω).
ωC

The quantity in the brackets on the left-hand side is of course the complex impedance
of the circuit, usually denoted by Z(ω). The corresponding dynamic susceptibility
is nothing but the reciprocal of the impedance, namely, the complex admittance,
customarily denoted by Y (ω). We have


I(ω) = Y (ω) V (ω),

where  −1
i iω
Y (ω) = −iωL + R + = .
ωC L (ω + iγ ω − ω02)
2
√
You will recognize γ −1 = L/R as the time constant of an LR circuit, and ω0 = 1/ LC
as the natural frequency of an undamped LC circuit. Observe that Y (ω) is analytic in
the upper half-plane in ω, as required. Moreover, in this simple case we know its form
explicitly. It has two simple poles in the lower half-plane, at

ω± = − 2 iγ ± ω02 − 14 γ 2
1

respectively. Further, Y (ω) vanishes like ω −2 as |ω| → ∞ in the upper half-plane. Its
real and imaginary parts must therefore satisfy (unsubtracted) dispersion relations.

3. Verify explicitly that the functions

γω 2 iω(ω 2 − ω02 )
Re Y (ω) = and Im Y (ω) =
(ω 2 − ω02 )2 + ω 2 L2 (ω 2 − ω02)2 + ω 2 L2

33
satisfy the dispersions relations
 ∞  ∞
2 ω  Im Y (ω  ) 2ω Re Y (ω  )
Re Y (ω) = P dω  , Im Y (ω) = − P dω   2 .
π 0 ω  2 − ω2 π 0 ω − ω2

4. Subtracted dispersion relations: It may so happen that a dynamic susceptibility

χ(ω) does not vanish as ω → ∞. Instead, it may tend to a constant as ω → ∞ along
some direction or directions in the region Im ω ≥ 0. (It is clear that this is most likely
to happen as ω → ∞ along the real axis itself.) In this case, when we try to derive
dispersion relations for the real and imaginary parts of χ(ω), we find that the contribu-
tion from the large semicircle of radius R no longer vanishes as R → ∞. If χ(ω) tends
uniformly to a constant as |ω| → ∞ along the real axis and along all directions in the
upper half-plane, we could go ahead by including iπχ∞ as the extra contribution from
the large semicircle. But there is no guarantee that this will be the case. In fact, it
does not happen, in general.

To get over the problem, we assume that the value of χ(ω) at some particular real
value of ω0 of the frequency is known. Then, instead of the function χ(ω  )/(ω  − ω),
consider
χ(ω  ) − χ(ω0 )
f (ω  ) = .
(ω  − ω0 )(ω  − ω)
It is evident that this function does vanish faster than 1/ω  as |ω  | → ∞, owing to
the extra factor (ω  − ω0 )−1 . Moreover, it is analytic everywhere in the upper half
ω  -plane and on the real axis, except for a simple pole at ω  = ω, as before. Observe
that it does not have any singularity at ω  = ω0 . This is the reason for subtracting
χ(ω0 ) from χ(ω  ) in the numerator. Repeat the earlier derivation to show that
 ∞
i χ(ω  ) − χ(ω0 )
χ(ω) = χ(ω0 ) − (ω − ω0 ) P dω  .
π −∞ (ω  − ω0 )(ω  − ω)
Hence obtain the dispersion relations satisfied by the real and imaginary parts of χ(ω)
in this case. These are called once-subtracted dispersion relations, and ω0 is the
‘point of subtraction’. From the mathematical point of view, it should be evident
that the procedure given above can be extended to cover situations when the analytic
function χ(ω) actually diverges as ω → ∞ in the upper half-plane. For instance, if
χ(ω) ∼ ω n−1 asymptotically, where n is a positive integer, we can write down an n-fold
subtracted dispersion relation for it. The latter will require n constants as inputs.
These could be, for instance, the values of χ(ω) at n different frequencies or points of
subtraction.

Hilbert transform pairs : Given a real function g(x) of the real variable x, its
Hilbert transform is defined as
 ∞
1 g(x  )
h(x) = P dx   .
π −∞ x −x

34
The inverse transform is given by
 ∞
1 h(x  )
g(x) = − P dx   .
π −∞ x −x

Eliminating h, we have
 ∞  ∞
1  g(x )
g(x) = − 2 P dx P dx 
π −∞ −∞ (x − x)(x − x  )

This may be re-written as

 ∞   ∞ 
 1 dx 
g(x) = dx 2
P  − x)(x  − x )
g(x ).
−∞ π −∞ (x

Hence the quantity in the curly brackets in the equation above must be the unit
operator in function space, i.e., we must have
 ∞
1 dx 
P = δ(x − x ).
−∞ (x − x)(x − x )
π 2   

5. Verify this identity explicitly. First consider x

= x , and show that the integral on
the left-hand side vanishes identically. Next, for any sufficiently well-behaved function
h(x), show that
 ∞   ∞ 
 1 dx 
dx P h(x ) = h(x),
−∞ (x − x)(x − x )
π 2   
−∞

thereby establishing the required representation of the δ-function.

6. Discrete relaxation spectrum : As we have seen, the real and imaginary parts
of the dynamic susceptibility constitute a Hilbert transform pair. Given that

n
σj
Re χ(ω) =
j=1
1 + ω 2 τj2

where σj and τj are positive constants, use the dispersion relation to show that

n
ω σj τj
Im χ(ω) = 2 τ2
.
j=1
1 + ω j

Hence

n
σj
χ(ω) = .
j=1
1 − iω τj

35
Thus χ(ω) has a set of poles in the lower half-plane in ω, at ω = −i/τj , 1 ≤ j ≤ n.
Physically, this corresponds to a discrete relaxation spectrum comprising the re-
laxation times τj , contributing with corresponding weights σj . The response function
φ(t) itself is a weighted sum of decaying exponentials e−t/τj . In a physical situation in
which there is essentially just a single relaxation time τ (or when it is a good approxi-
mation to assume that this is so), we have the well-known case of Debye relaxation.
In this instance, the so-called Cole-Cole plot or the Argand diagram of χ(ω), i.e., a
plot of (the suitably scaled variables) Im χ versus Re χ, obtained by eliminating ω, is
the upper half of semi-circle of radius 12 centered at ( 12 , 0). Departures of the Cole-Cole
plot from this curve serve as an indication that the relaxation is not of the single-
exponential form.

7. Continuous relaxation spectrum: If the number of contributing relaxation modes

is infinite, or if there is a continuum of such modes, much more complex dynamical
behavior may result. I will not go into these aspects here. Given that
 τmax
σ(τ ) dτ
Re χ(ω) = 2 2
,
τmin 1 + ω τ

show that  τmax

ωτ σ(τ ) dτ
Im χ(ω) = .
τmin 1 + ω2 τ 2
Hence  τmax
σ(τ ) dτ
χ(ω) = .
τmin 1 − iω τ
The point is that the singularity structure of χ(ω) in the lower half-plane in ω may
now be quite intricate. (It continues to remain holomorphic in the UHP, of course.)
In general, a functional form such as that given by the last equation above will have
logarithmic branch points at ω = −i/τmax and ω = −i/τmin . (For instance, take
σ(τ ) to be a constant independent of τ .) Even more intricate possibilities arise when
(τmin , τmax ) = (0 , ∞). These, in turn, are associated with physically interesting kinds
of non-Debye relaxation of response functions, including ‘glassy’ or ‘slow dynamics’,
‘stretched-exponential decay’, etc., in various condensed matter systems, notably those
possessing some kind of frozen or ‘quenched’ disorder.

36
4 Analytic continuation and the gamma function
The gamma function : In practice, there are numerous methods of analytically
continuing a function from a region in which a local representation of the function is
available, to a larger region. Some of these are: chains of rearranged power series; the
Schwarz reflection principle (for real analytic functions); the use of the ‘permanence’
of functional equations; summability methods for power series; integral representations
for functions in terms of contour integrals, using the distortability of the contours; and
so on.

A good illustration is provided by the gamma function, which may be defined (to
start with) by the integral
 ∞
Γ(z) = dt tz−1 e−t , Re z > 0.
0

As you know, Γ(n + 1) = n! for positive integer values of n. The integral above serves
also to define 0! as unity. The question is, how do we analytically continue the gamma
function to the region Re z ≤ 0 in the complex z-plane?

It is obvious that the problem with the convergence of the integral above for Re z ≤
0 arises from the lower limit of integration in t. The behavior of the integrand at this
end-point can be improved by integration by parts. Keeping Re z > 0, integrate by
parts to obtain
 ∞
Γ(z) = dt tz−1 e−t (Re z > 0)
0
z −t t=∞ 
t e 1 ∞ z −t
= + dt t e (Re z > 0)
z z 0
  t=0

=0 because Re z>0

But the final integral on the right-hand side converges in the extended region Re z >
−1, while the factor 1/z multiplying it suggests that the gamma
 ∞ function itself has a
z −t
simple pole at z = 0. And indeed it does, because the integral 0 dt t e is finite and
nonzero at z = 0. It is equal to 1 when z = 0, so that we may conclude that Γ(z) has
a simple pole at z = 0 with residue equal to unity. We have thus extended the region
in which we have an explicit representation of the gamma function by the extra strip
−1 < Re z ≤ 0, and we may write

1 ∞ z −t
Γ(z) = dt t e , Re z > −1.
z 0
Continuing this procedure of integrating by parts, it is easy to see that we have
 ∞
1
Γ(z) = dt tz+n e−t , Re z > −n − 1.
z(z + 1) · · · (z + n) 0

37
The integral on the right-hand side converges for Re z > −n−1, for any n = 0, 1, 2, . . . ,
while the factor multiplying it makes the poles at z = 0, 1, . . . , −n explicit. We have
thus achieved an analytic continuation of the gamma function, strip by strip, to the
region Re z > −n − 1, starting with a representation for the function that was only
valid in the region Re z > 0. In principle, this can be done for arbitrarily large value
of n. The underlying mechanism is summarized in the functional equation satisfied by
the gamma function, namely,
1
Γ(z + 1) = z Γ(z) or Γ(z) = Γ(z + 1).
z
The ‘permanence’ of this equation (i.e., its validity for all z) can be used to analytically
continue the gamma function arbitrarily far to the left of the imaginary axis in the
complex z-plane.

1. Connection with the Gaussian integral: The gamma function helps us express
a number of very useful integrals in closed form.
(a) Show, by a simple change of variables, that
 ∞
2  
du un e−au = 12 Γ 12 (n + 1) a−(n+1)/2 , a > 0, n > −1.
0

Note, in particular, that n need not be an integer in this formula. In fact, as you can
readily guess, the formula continues to be valid for complex values of n and a, provided
Re n > −1 and Re a > 0. Setting n = 0 and a = 1 in the formula, we may turn it
around to obtain  ∞
1 2 √
Γ 2 =2 du e−u = π,
0
where we have made use of the known result for the basic Gaussian integral.

(b) It follows at once that the value of the gamma function of a half-odd-integer can
be written down explicitly. Use the functional equation for the gamma function
to show that
√ √
  (2n)! π   2n
n 2 n! π
Γ n+ 2 =1
2n
and Γ −n + 2 = (−1)
1
, n = 0, 1, 2, . . .
2 n! (2n)!

Analytic properties of the gamma function: It is clear from the analytic con-
tinued form of the gamma function (obtained either by integration by parts, or from
the functional equation) that Γ(z) has simple poles at all the non-positive integers,
n = 0, −1, −2, . . . . Its residue at z = −n is equal to (−1)n /n! . Other than these

38
simple poles, it has no other singularities in the finite part of the complex plane. It
is therefore a meromorphic function. It is then natural to ask for the Mittag-Leffler
expansion of Γ(z)—namely, a representation that explicitly separates out the contri-
bution from all the poles of the function from the part that is an entire function.

2. Mittag-Leffler expansion of Γ(z): Note that the pole at z = −n arises from the
behavior of the factor tz+n−1 at the lower limit of integration, t = 0, in the integral
representation of Γ(z). Write
 ∞  1  ∞
z−1 −t z−1 −t
Γ(z) = dt t e = dt t e + dt tz−1 e−t .
0 0 1

The second integral on the right-hand side in the last equation above does not have
any singularities in the region |z| < ∞, and is therefore an entire function of z. In the
integral from t = 0 to t = 1, expand e−t in powers of t and integrate term by term, to
obtain the Mittag-Leffler expansion of the gamma function :
 ∞  ∞
(−1)n
Γ(z) = + dt tz−1 e−t .
n! (z + n)
   
1
n=0 
entire function
sum over pole terms

3. The logarithmic derivative of Γ(z) is also known as the digamma function,

and is denoted by ψ(z). We have

d Γ  (z)
ψ(z) = Γ(z) = .
dz Γ(z)

The functional relation Γ(z + 1) = z Γ(z) leads at once to the difference equation
1
ψ(z + 1) − ψ(z) = .
z
From the fact that
(−1)n
Γ(z) = + regular part
n! (z + n)
in the neighborhood of its simple pole at z = −n, it follows that Γ  (z) has a double
pole at z = −n, and has the behavior
(−1)n
Γ  (z) = − + regular part.
n! (z + n)2

Hence show that ψ(z) has simple poles at n = 0, −1, −2, . . . , with residue equal to
−1 at each pole.

39
Infinite product
 representation for the gamma function: Recall that the har-
−1
monic series N n=1 n diverges logarithmically as N → ∞. That is, the sum has a
leading asymptotic behavior ∼ ln
N for large
 values of N. The question that arises
N −1
naturally is: does the difference n=1 n − ln N tend to a finite limit as N → ∞?
It turns out that

N  
lim n−1 − ln N = γ,
N →∞
n=1

where γ is a number (conjectured to be irrational8 ) called the Euler-Mascheroni

constant. Its numerical value is 0.5772 . . . . This constant has many close links with
the gamma function. For instance, we know that Γ(z) has a simple pole at z = 0, with
residue equal to 1. If we subtract out this singular part, and then we pass to the limit
z → 0, we get  1
lim Γ(z) − = −γ.
z→0 z
In other words, the behavior of the gamma function near in the neighborhood of z = 0
is given by
1
Γ(z) = − γ + O(z).
z
Another relationship is
ψ(1) = Γ  (1) = −γ.
The constant γ occurs explicitly in the following representation of the gamma function.
Γ(z) is a meromorphic function of z. Moreover, it has no zeroes. Hence its recipro-
cal is an entire function of z. These properties are explicit in the infinite product
representations
e−γz 
n  z/n
∞
Γ(z) = e
z n=1 z + n
and hence ∞

1 γz
 z  −z/n
= ze 1+ e .
Γ(z) n=1
n

Taking logarithms and differentiating with respect to z, this gives

1 ∞
z
ψ(z) = − − γ + .
z n=1
n(n + z)

Setting z = 1 in this expression, we recover the result ψ(1) = −γ.

The beta function B(m, n) is defined as

 1
B(m, n) = dt tm−1 (1 − t)n−1 ,
0
8
In fact, a transcendental number, rather than an algebraic irrational.

40
where m and n are positive integers. It is clear, however, that the same integral
representation can be used to define the beta function as a function of two complex
variables z and w, say, according to
 1
B(z, w) = dt tz−1 (1 − t)w−1 , Re z > 0 and Re w > 0.
0

B(z, w) is an analytic function of its arguments in the region indicated. As in the

case of the gamma function, we can try to extend the region of analyticity to the left
in the z-plane by integrating by parts with respect to t. However, while integrating
the factor tz−1 will improve matters in the z-plane, it will also worsen matters in the
w-plane, because the factor (1 − t)w−1 will then get differentiated, and lead to the
factor (1 − t)w−2 . The opposite will happen if we integrate the factor (1 − t)w−1 and
differentiate the factor tz−1 . In fact, a change of variables from t to (1 − t) easily
establishes the symmetry property

B(z, w) = B(w, z).

The actual structure of the beta function is clarified by relating it to the gamma func-
tion as follows.

4. Consider the double integral

 ∞  ∞
2 +v 2 )
I(z, w) = du dv e−(u u2z−1 v 2w−1 ,
0 0

which converges in the region Re z > 0 and Re w > 0. Using the formula connecting
the Gaussian integral to the gamma function,

I(z, w) = 14 Γ(z) Γ(w).

Now go over to plane polar coordinates u = r cos θ and v = r sin θ, to get

 π/2
1
I(z, w) = 2 Γ(z + w) dθ (cos θ)2z−1 (sin θ)2w−1 .
0

Set t = cos2 θ to show that

 π/2
dθ (cos θ)2z−1 (sin θ)2w−1 = 12 B(z, w).
0

Hence, equating the two expressions for I(z, w), arrive at the basic relationship
Γ(z) Γ(w)
B(z, w) = .
Γ(z + w)
As always, this equation must be regarded as one between two analytic functions, valid
for all values of the arguments z and w. It leads to a number of important properties

41
of the gamma function.

Reflection formula for Γ(z) : Set w = 1 − z in the foregoing relation. Hence

Γ(z) Γ(1 − z) = B(z, 1 − z). In the strip region given by 0 < Re z < 1, we may use the
defining representation of the beta function to write
 1
tz−1
Γ(z) Γ(1 − z) = dt (0 < Re z < 1).
0 (1 − t)z

Changing variables to u = t/(1 − t), this becomes

 ∞
uz−1
Γ(z) Γ(1 − z) = du (0 < Re z < 1).
0 u+1

The integral on the right-hand side can be evaluated by contour integration.9 The
result is
π
Γ(z) Γ(1 − z) = .
sin πz
This is the formula sought. Once again, it is a relation between analytic functions.
Therefore, by analytic continuation, it is valid for all z. Note that cosec (πz) has a
simple pole at every integer value of z. On the left-hand side, Γ(z) supplies the poles
at zero and the negative integers, while Γ(1 − z) has poles at the positive integers.
Taking logarithmic derivatives on both sides, it follows that

ψ(1 − z) − ψ(z) = π cot (πz).

5. Using the fact that Γ(z) is a real analytic function of z, i.e., it is real when z is real,
show that  1 
Γ + iy 2 = π sech (πy).
2

6. Legendre’s duplication formula

 for
 Γ(z) is a very useful identity that expresses
Γ(2z) as a product of Γ(z) and Γ z + 12 . Start with
 1
B(z, z) = dt [t (1 − t)]z−1 , Re z > 0.
0

But the integrand in the above is symmetric about the mid-point of integration, t = 12 .
Hence  1/2
B(z, z) = 2 dt [t (1 − t)]z−1 .
0
9
You’ll see how this can be done after we discuss multivalued functions and branch points.

42
Change variables to u = 4t(1 − t), to get
 
B(z, z) = 21−2z B z, 12 .

Now use the relation between the beta and gamma functions to obtain Legendre’s
duplication formula, namely,

22z−1  
Γ(2z) = √ Γ(z) Γ z + 12 .
π

The formula is actually a special case of a multiplication theorem for the gamma
function, which reads


n−1
j
(1−n)/2 nz− 21
Γ(nz) = (2π) n Γ z+ , n = 1, 2, . . . .
j=0
k

7. Taking logarithmic derivatives in the duplication formula, we get

 
ψ(2z) = ln 2 + 12 ψ(z) + 12 ψ z + 12 .

Use this relation, and the difference equation for ψ(z), to establish the following:

n
1
(a) ψ(z + n) = ψ(z) + .
j=1
(z + j − 1)


n
1
(b) ψ(n + 1) = −γ + .
j=1
j
1
(c) ψ 2
= −γ − 2 ln 2 .

  
n
1
(d) ψ n + 1
= −γ − 2 ln 2 + 2 .
2
j=1
(2j − 1)

43
5 Möbius transformations
Conformal mapping: Every analytic function f (z) is a map of some region R of the
complex plane to some region R of the complex plane. In other words, given a complex
number z ∈ R, the map f : z → w produces another complex number w ∈ R . Under
such a mapping, points in R map to points in R , curves to curves, and areas to areas,
in general.
• The special feature of an analytic function f (z) is that this map is conformal.
That is, it preserves the angles between curves.
Conformal mapping is a standard and very well-studied topic in complex analysis. It
is also one that has a number of practical uses—for instance, in fluid dynamics and
aerodynamics. We will not go into conformal mapping per se here. Instead, in this
chapter, the focus is on one specific kind of conformal map, because it is a very special
one in a sense that will become clear shortly. It is also of fundamental importance in
many areas of mathematical physics.

A natural question that arises is whether the mapping is one-to-one: namely, is

there a unique w = f (z) for every given z, and vice versa? If the function f (z) is
single-valued, then by definition there is a unique w for every z. The converse is not
necessarily true, of course. It is clear that a nonlinear function such as f (z) = z 2 does
not satisfy this requirement. While there is a unique w = z 2 for every z, we cannot
find z uniquely if we are given w. We can only do so up to an overall sign. Or else, as
you know, we agree to map the complex w-plane onto a two-sheeted Riemann surface,
on which z = w 1/2 is single-valued. On the other hand, it may well be possible to
map of some specific, restricted part R of the complex plane to another such part R ,
in a one-to-one manner. Indeed, in most of the standard applications of conformal
mapping, this is all that is sought.

The more general question of whether there is a one-to-one conformal mapping

C → C has a simple and rather trivial answer: The most general one-to-one mapping
that takes the whole of the finite part of the complex plane to itself is the linear map

f (z) = az + b,

where a and b are complex constants. Further, the point at ∞ is mapped to itself under
this map, i.e., ∞ is a fixed point of this map. In terms of the real and imaginary
parts of z = x + iy, this map amounts to the linear transformations

x → a1 x − a2 y + b1 , y → a2 x + a1 y + b2 ,

where a = a1 + ia2 and b = b1 + ib2 . In geometrical terms, this map is made up (in
general) of a rotation, dilation and shear in the xy-plane, followed by a shift of the
origin. Interesting as it is, the map is quite simple, and the matter appears to end

44
here. But this is not quite so, as you will see below.

Definition of a Möbius transformation: Things become much more interesting

when we consider the extended complex plane Ĉ (that is, we include the point at infin-
ity). The linear transformation just discussed leaves the point at infinity unaffected,
i.e., it maps it to itself. What if we consider transformations that are not restricted in
this manner? Remember that Ĉ is essentially equivalent to the Riemann sphere S. A
nontrivial result emerges now.
• The most general one-to-one, conformal map Ĉ → Ĉ of the extended complex
plane to itself (or of the Riemann sphere, S → S) is of the form
az + b
w = f (z) = , (ad − bc
= 0)
cz + d
where a, b, c, d are finite, complex constants. This is called a fractional linear
transformation or a Möbius transformation.10
Note that the condition ad − bc
= 0 is required. (If ad − bc = 0, then w = f (z)
trivially reduces to a constant for every z, which is not very interesting.) It is obvious
that we can divide all four of the four constants a, b, c and d by any constant without
affecting the ratio (az + b)/(cz + d) that defines the transformation. In particular, we
can divide through by (ad − bc)1/2 . In effect, therefore, (ad − bc) can always be taken
to be equal to 1, without loss of generality. We’ll assume henceforth that this has been
done, unless specified otherwise.

The inverse of a Möbius transformation is also a Möbius transformation. We have

dw − b
z = f −1 (w) = .
−cw + a
The same determinant condition, ad−bc = 1, remains valid for this transformation too.

Every Möbius transformation maps a certain point in the z-plane to ∞. If c

= 0,
this point is z = −d/c. Similarly, the point at infinity maps to a/c. That is,
d a
z=− → w = ∞ and z = ∞ → w = .
c c
If c = 0, we have just a linear transformation, and z = ∞ maps to w = ∞.

The fixed points of a Möbius transformation, i.e., points that map to themselves
under the transformation, are of fundamental importance. They help classify the
10
The latter name is more commonly used when the transformation is regarded as a mapping of the
Riemann sphere, and I will use this term. There are several other names (!) for this transformation—
linear fractional transformation, homographic transformation, projective transformtion, and bilinear
transformation.

45
transformations into distinct types, as you’ll see later. The fixed points are the roots
of the equation f (z) = z, or

cz 2 + (d − a)z − b = 0.

It is immediately clear that a Möbius transformation can have at most two fixed points,
which I’ll denote by ξ1 and ξ2 . (The only exception is the identity transformation,
which of course maps every point to itself.) Two cases arise, each with a sub-case, so
that there are four possibilities:

(i) Finite, distinct fixed points: In the general case, when c

= 0 and a + d
= ±2,
there are two distinct, finite, fixed points given by

a − d ± (a + d)2 − 4
ξ1,2 = ,
2c
using the fact that ad − bc = 1.

(ii) Finite, coincident fixed points: If c

= 0 but a + d = ±2, the two fixed points
coincide, and we have
a−d a∓1
ξ1 = ξ2 = = .
2c c

(iii) One fixed point at ∞: If c = 0, the Möbius transformation reduces to the

linear transformation
az + b
w= = a2 z + ab,
d
since ad = 1 in this case. The fixed points are then at
b ab
ξ1 = = and ξ2 = ∞.
d−a 1 − a2

(iv) Both fixed points at ∞: If c = 0 and further a = d, then the linear transfor-
mation is merely a shift of the origin,
b
w=z+ = z ± b,
a
since a2 = 1 in this case. The fixed points then coincide at infinity,

ξ1 = ξ2 = ∞.

We’ll return to the role played by these fixed points shortly.

46
The cross-ratio of four points is a basic concept in geometry. It plays a fundamen-
tal role in the subject of projective geometry. It is extremely useful in the context of
Möbius transformations, as many results can be established quite easily with the help
of its properties.

Let z1 , z2 , z3 and z4 be an ordered set of four distinct points in the complex plane.
The cross-ratio of these points is defined as

def. (z1 − z3 )(z2 − z4 )

[z1 , z2 ; z3 , z4 ] = .
(z1 − z4 )(z2 − z3 )

(Draw a figure showing the distances involved in the cross-ratio, for a general configu-
ration of points.) It is obvious that the cross-ratio of four points depends on the order
in which they are specified. Hence there are 4! = 24 possible cross-ratios that can be
associated with an unordered set of four points. However, there are several obvious
symmetries among these, and the 24 cross-ratios are related to each other. Let’s write
[zi , zj ; zk , zl ] as [ij; kl], for brevity. It is easy to see that

[ij; kl] = 1/[ij; lk] = 1/[ji; kl] = [ji; lk] = [kl; ij].

The identities

[ik; jl] = 1 − [ij; kl] and [il; jk] = 1 − 1/[ij; kl]

are also easily established. Using these, it is possible to express all 24 cross-ratios
among any set of 4 points in terms of any one cross-ratio (see below). If one of the four
points in a cross-ratio is the point at infinity, the cross-ratio is defined by a limiting
process. For instance, if z1 = ∞, we have

(z2 − z4 )
[∞ , z2 ; z3 , z4 ] = lim [z1 , z2 ; z3 , z4 ] = .
z1 →∞ (z2 − z3 )

Numerous results can be derived from the following fundamental property:

• A Möbius transformation leaves the cross-ratio of any four arbitrary points un-
altered in value.

1. The properties of the cross-ratio stated above are easily established.

(a) Find the relations between the 24 cross-ratios associated with any four distinct
points (z1 , z2 , z3 , z4 ) ∈ C.

(b) Verify that [z1 , z2 ; z3 , z4 ] = [w1 , w2 ; w3 , w4 ], where wi = (azi +b)/(czi +d), i =

1, 2, 3, 4.

47
From the invariance of the cross-ratio under a Möbius transformation, the following
property can be established:

• Given two sets of three distinct points (z1 , z2 , z3 ) and (w1 , w2 , w3 ) on the Rie-
mann sphere, there exists a unique Möbius transformation that maps zi to wi for
i = 1, 2 and 3.

There are several ways to prove this assertion. We won’t go into the complete proof
here, but it is possible to construct the Möbius transformation concerned quite easily:

2. Assume that such a Möbius transformation exists. Construct it explicitly.

Hint: Consider the cross-ratio [z , z1 ; z2 , z3 ]. If z → w under the transformation, then

this cross-ratio must be equal to [w , w1 ; w2 , w3 ]. Solve this (simple) equation for w,
to show that it is in the form of a Möbius transformation of z.

And now we come to a crucial geometrical property of Möbius transformations. Since

a circle can be drawn through any three distinct points, the result stated above has
the following implication:

• A Möbius transformation maps circles to circles on the Riemann sphere.

• In the complex plane, a Möbius transformation maps circles and straight lines
to circles and straight lines. (Recall that a straight line in the complex plane is
a circle passing through the point of projection on the Riemann sphere.)

3. Let z1 , z2 , z3 be three arbitrary finite points in the complex plane. Show that the
Möbius transformation that maps the circle passing through these three points (or
the straight line on which they lie, if they are collinear) to the real axis such that
z1 → 0, z2 → 1, z3 → ∞ is

(z − z1 )(z3 − z2 )
w= .
(z − z3 )(z1 − z2 )

Hint: Use the fact that we must have [z , z1 ; z2 , z3 ] = [w , 0 ; 1 , ∞] under the trans-
formation sought. Solve this equation to find w as a function of z.

Normal form of a Möbius transformation: The original form of a Möbius trans-

formation, z → w = (az + b)/(cz + d) with ad − bc = 1, is not the most convenient way
to express the transformation for some purposes. The so-called normal form provides
a better representation. It involves the fixed points of the transformation, comprising
the four cases listed earlier.

48
(i) Finite, distinct fixed points ξ1 and ξ2 : Consider the four points z, ∞, ξ1 and
ξ2 , where z is variable. Under the Möbius transformation, these points are mapped
according to
a
z → w, ∞ → , ξ1 → ξ1 , ξ2 → ξ2 .
c
Since the cross-ratio does not change under the mapping, we have

[z, ∞ ; ξ1 , ξ2 ] = [w, a/c ; ξ1 , ξ2 ].

Writing this out explicitly, we have the following result. Define the constant

def. a − c ξ1 a + d − (a + d)2 − 4
K = =  .
a − c ξ2 a + d + (a + d)2 − 4

Then the transformation can be expressed in the form

 
w − ξ1 z − ξ1
=K .
w − ξ2 z − ξ2

This is the normal form of the Möbius transformation. The constant K is called the
multiplier of the Möbius transformation. Observe that it depends solely on the sum
(a + d). The importance of this fact will become clear subsequently.

(ii) Finite, coincident fixed points: Recall that this case corresponds to c
= 0 and
(a + d) = ±2. The fixed point is given by ξ1 = ξ2 = ξ = (a − d)/(2c). The multiplier
reduces to K = 1. Starting with w = (az + b)/(cz + d), a little algebra leads to the
normal form in this case, which is
1 1
= ± c, depending on whether a + d = ±2.
w−ξ z−ξ

(iii) One fixed point at ∞: This is the case c = 0, a

= d. As you know, the Möbius
transformation now reduces to the linear transformation z → w = (az + b)/d, with
fixed points at ξ1 = b/(d − a) and ξ2 = ∞. It is easy to see that we now have the
normal form
w − ξ1 = K(z − ξ1 ) where K = a/d.

(iv) Both fixed points at ∞: Now c = 0 and a = d. The transformation reduces

to a shift of the origin,

w = z ± b, depending on whether a = d = ±1.

49
Again, this is the normal form in the present case. The multiplier K remains equal to
unity, of course.

4. Verify that the normal forms in the four cases above are as given above.

Iterates of a Möbius transformation: It is easy to check that the result of making

two Möbius transformations in succession is again a Möbius transformation. Hence
the effect of an arbitrary number of them performed in succession is also a Möbius
transformation. In particular, we may ask for the result of a given transformation,
z → w = (az + b)/cz + d), iterated an arbitrary number of times. Let z → z (n) after n
repeated applications of a Möbius transformation, i.e.,
az + b
z (n) = f (f (· · · f (z)) · · · ) where f (z) = .
   cz + d
n−fold iteration

In its original form, the final expression for z (n) will obviously be quite complicated.
But if you use the normal form, you can write down the answer by inspection, in each
of the four cases listed above.

(i) Finite, distinct fixed points: In this case,

 
z (n) − ξ1 n z − ξ1
=K .
z (n) − ξ2 z − ξ2

(ii) Finite, coincident fixed points: In this case,

1 1
= ± n c, for a + d = ±2.
z (n)−ξ z−ξ

(iii) One fixed point at ∞: In this case,

z (n) − ξ1 = K n (z − ξ1 ), where K = a/d.

(iv) Both fixed points at ∞: In this case,

z (n) = z ± n b, for a = d = ±1.
The relations above are trivially valid for n = 0 as well, with z (0) ≡ z. Interestingly,
they are also valid for negative integer values of n, with z (−1) interpreted as the inverse
map of z, and z (−n) as the n-fold iterate of the inverse map of z:
dz − b  
z (−1) = f −1 (z) = , z (−n) = f −1 z −(n−1) .
−cz + a

50
It is often useful to re-write the formula for z (n) in the general case (i) in the form

(ξ1 − K n ξ2 ) z + ξ1 ξ2 (K n − 1)
z (n) = .
(1 − K n ) z + (K n ξ1 − ξ2 )

This form shows explicitly that the result of an n-fold iteration of a Möbius trans-
formation is again such a transformation.11 The relation above is easily inverted to
express z ≡ z (0) in terms of z (n) . All you have to do is to replace K n by K −n . It is
easy to see that this merely entails an interchange of ξ1 and ξ2 . Thus,

(ξ2 − K n ξ1 ) z (n) + ξ1 ξ2 (K n − 1)
z (0) = .
(1 − K n ) z (n) + (K n ξ2 − ξ1 )

Classification of Möbius transformations: We are ready, now, to consider the

systematic classification of Möbius transformations. The discussion will be restricted
to some introductory remarks on the topic.

For any value of z, the set of points {z (n) |n ∈ Z} is an orbit under the map rep-
resented by a Möbius transformation. Such an orbit is analogous to the orbit of a
point in the phase space of a dynamical system in discrete time, n playing the role
of time. The collection of orbits corresponding to different values of z = z (0) (or the
initial conditions, in the case of a dynamical system) gives us a flow in the complex
plane Ĉ (analogous to a phase portrait in phase space). The nature of this flow is
essentially determined by the fixed points, which are the analogs of the critical points
or equilibrium points in phase space. This is the starting point of the classification of
Möbius transformations.

Let’s consider the general case in which c

= 0 (so that the transformation does not
reduce to a mere linear transformation), and there are two distinct, finite fixed points.
It is convenient to write z (n) in the form
  (0)    (0) 
n z − ξ1 n z − ξ1
z = ξ1 − ξ2 K
(n)
1−K .
z (0) − ξ2 z (0) − ξ2

It is now easy to see, at least in most cases, what happens as n → ∞.

(i) If |K| > 1, the factor K n will grow in magnitude with increasing n. Therefore
z (n) → ξ2 as n → ∞, for all initial points z (0) (other than ξ1 , of course). In the spirit
of dynamical systems, we may regard ξ2 as an asymptotically stable fixed point, or
attractor; the flow is generally toward this point. On the other hand, ξ1 acts like an
11 th
Incidentally, this expression should
 a b suggest to you the possibility of a simple ‘formula’ for the n
power of a general (2 × 2) matrix c d with unit determinant. Check this out.

51
unstable fixed point, or repellor; the flow is generally away from this point.

(ii) If |K| < 1, the factor K n tends to zero with increasing n. Hence z (n) → ξ1 as
n → ∞, for all initial points z (0) (other than ξ2 ). The attractor is now ξ1 , and the flow
is toward this point. The repellor is ξ2 , and the flow leads away from this point.

(iii) When |K| = 1, we have the analog of marginal fixed points, and the flow is neither
toward nor away from the fixed points.

(iv) The case when the two fixed points coincide at ξ = (a − d)/(2c) occurs when
(a + d)2 = 4, and corresponds to K = 1. This is a special case of |K| = 1.

These statements can be corroborated as follows. Under the transformation from z

to w = f (z), we have dw = f  (z)dz. The Jacobian of the transformation is therefore
f  (z). Its magnitude |f  (z)| is the local stretch factor (or contraction factor). It is this
quantity that essentially characterizes the local flow in the neighborhood of any point.
(Remember that we are considering the nontrivial case in which c
= 0.) Now,
az + b 1
f (z) = ⇒ |f  (z)| = ,
cz + d |cz + d|2
on using the fact that ad − bc = 1. At the fixed points ξ1 and ξ2 , the respective stretch
factors become
4 
|f  (ξ1 )| =    = |K| 2
, 

 a + d + (a + d)2 − 4 2 




. 
4 1
|f  (ξ2 )| =   2 = 

 a + d − (a + d)2 − 4  |K|2

It can be shown that a fixed point is

— unstable if the stretch factor at that point is greater than unity;
— asymptotically stable if it is less than unity; and
— marginal when it is equal to unity.
The statements made above then follow.

 5. Verify the foregoing.

The crucial point is that, in all cases, K depends only on the trace T ≡ (a + d) of the
matrix ( ac db ) made up of the coefficients of the transformation. The multiplier is given
in terms of T by √
T − T2 − 4
K= √ .
T + T2 − 4

52
This is the starting point for the classification of Möbius transformations into different
types. Four types of transformations are possible, depending solely on the value of
T . I merely list them here, without going into further detail. The first three types
correspond to real values of the trace T .

Type 1: Elliptic transformation, when T is real and −2 < T < 2. Then |K| = 1,
but K
= 1.

Type 2: Parabolic transformation, when T = ±2. The two fixed points now
coincide, and we have K = 1.

Type 3: Hyperbolic transformation, when T is real and |T | > 2. In this case K

is a positive number other than unity (i.e., K
= 1).

Type 4: Loxodromic transformation, when T is any complex number such that

T ∈/ [−2, 2]. A loxodrome is a curve drawn on the surface of a sphere that intersects all
lines of longitude at a given, constant angle ψ. It is obvious that ψ = 0 and ψ = 12 π
correspond, respectively, to a line of longitude and a latitude, respectively. For in-
termediate values of ψ, the loxodrome is a spiral. As you might expect, loxodromes
first arose in the context of navigation on the high seas. Sailing on a path that makes
a constant angle with the northern direction takes you on a loxodrome. Hyperbolic
transformations are a special case of this type, corresponding to real values of T .

The isometric circle: As stated already, the local stretch or contraction factor
associated with the mapping of an infinitesimal neighborhood of any point z by the
Möbius transformation z → (az + b)/(cz + d) is |f  (z)| = 1/|cz + d|2 . It follows that
there is no distortion at all in the map of the circle given by

|cz + d| = 1, or |z + (d/c)| = 1/|c|.

(Recall, once again, that we are considering the case c

= 0.) This circle with center
at −d/c and radius equal to 1/|c| is called the isometric circle corresponding to the
transformation concerned. It plays an important role in the theory of Möbius trans-
formations.

Points inside the isometric circle satisfy |cz + d| < 1, and hence infinitesimal area
elements inside the circle are magnified by the mapping, by a factor |cz + d|−4 . Simi-
larly, points outside the isometric circle satisfy |cz + d| > 1. Hence infinitesimal area
elements in this region are shrunk by the transformation, by a factor |cz + d|−4.

6. Consider the Möbius transformation z → w = (az + b)/(cz + d) (where c

= 0).
(a) Show that the transformation maps its isometric circle to the isometric circle of
the inverse transformation w → z = (dw − b)/(−cw + a). Further, the interior

53
 (respectively, exterior) of the isometric circle in the z-plane is mapped to the
exterior (respectively, interior) of its image in the w-plane. (As you know, the
center −d/c of the isometric circle is mapped to ∞, while ∞ is mapped to the
center a/c of the image of the isometric circle under the mapping.)

(b) Consider the case of finite, distinct fixed points ξ1 and ξ2 , and (i) |K| > 1 and
(ii) |K| < 1, respectively. What is the isometric circle of the nth iterate of the
transformation? What happens as n → ∞?

Group properties; the Möbius group: An important feature of Möbius transfor-

mations or maps f : z → w is the fact that they form a group, the so-called Möbius
group Möb (2, C). A few salient properties of this group will be discussed here.

Möbius transformations are easily seen to satisfy the axioms that define a group:

(a) The successive application of two Möbius transformations is another Möbius

transformation. More formally, if f and g are maps, so is their composition f ◦ g.

(b) The composition of transformations is associative. If f, g and h are maps, then

f ◦ (g ◦ h) = (f ◦ g) ◦ h.

c) There is an identity transformation under which every z is mapped to itself.

(d) Each transformation f : z → w has an inverse f −1 : w → z.12

Here’s how the transformation obtained by composing two successive transformations

can be read off easily:

• The transformation z → w = (az + b)/(cz + d) can be associated with the matrix

of coefficients, ( ac db ).

• The composition of transformations then corresponds to the product of the as-

sociated matrices (in the right order).

Thus, if
a1 z + b1 a2 z + b2
z → z  = is followed by z  → z  = ,
c1 z + d 1 c2 z  + d 2
then
(a2 a1 + b2 c1 )z + a2 b1 + b2 d1
z  = .
(c2 a1 + d2 c1 )z + c2 b1 + d2 d1
But the set of all nonsingular (2 × 2) matrices with complex elements constitutes a
group. The group composition law is just matrix multiplication. The group is called
the general linear group over the complex numbers in 2 dimensions, and is denoted
by GL(2, C). As you know, however, the coefficients of a Möbius transformation can
12
That is, if w = f (z) = (az + b)/(cz + d), then z = f −1 (w) = (dw − b)/(−cw + a).

54
always be chosen such that ad − bc = 1. What is relevant here, therefore, is the set of
(2×2) matrices with complex elements and determinant equal to unity. These matrices
form a subgroup of GL(2, C): the unimodular or special linear group over the complex
numbers, denoted by SL(2, C). We might therefore expect Möb (2, C) to be essentially
the same as SL(2, C). But there is one more point to be taken into account.

Changing the sign of all the four coefficients in a Möbius transformation does not
alter it, because (az + b)/(cz + d) ≡ (−az − b)/(−cz − d). Hence, to each Möbius
transformation
az + b
z → w = with ad − bc = 1,
cz + d
there correspond the two matrices
   
a b −a −b
and ∈ SL(2, C).
c d −c −d

The identity transformation z → z thus corresponds to the (2 × 2) identity matrix I

as well as its negative, −I. Hence:

• There is a two-to-one correspondence, or homomorphism, from the group SL(2, C)

to the group Möb (2, C).

• The matrices I and −I form the kernel of the homomorphism. That is, {I, −I}
is the set of matrices in SL(2, C) whose image in Möb (2, C) is the identity trans-
formation.

But the two matrices I and −I themselves form a group, under matrix multiplication.
This group is just the cyclic group of order 2, denoted by Z2 . It is the same as the
group formed by the integers 0 and 1 under binary addition, i.e., addition modulo 2.
One now forms the quotient group SL(2, C)/Z2 . This is read as ‘SL(2, C) modulo Z2 ’,
and refers, in effect, to the group of unimodular 2 × 2 matrices up to the overall sign of
the matrices. In other words, it is a group of these matrices, such that any two matrices
differing only in sign are identified with each other and regarded as a single element of
the group. It is this group, also known as the projective special linear group P SL(2, C),
with which the group of Möbius transformations is in one-to-one correspondence.

• Möb (2, C) is isomorphic to SL(2, C)/Z2 . This isomorphism is written as

Möb (2, C) ∼
= SL(2, C)/Z2 ≡ P SL(2, C).

• The group SL(2, C) is the so-called universal covering group of Möb (2, C).
The latter is a subgroup of SL(2, C).

Finally, I mention the following remarkable fact that connects Möbius transformations
to special relativity. The special linear group SL(2, C) is also the so-called universal

55
covering group of the group of homogeneous, proper, orthochronous Lorentz trans-
formations, SO(3, 1), in the usual spacetime comprising three spatial dimensions and
one time dimension. There is again a 2-to-1 homomorphism between SL(2, C) and
SO(3, 1), and we have

Möb (2, C) ∼
= SL(2, C)/Z2 ∼
= SO(3, 1).

That is, the group of Möbius transformations is isomorphic to the Lorentz group!

The Möbius group over the reals: We’ve seen that the Möbius group Möb (2, C)
is isomorphic to certain other important groups such as the projective linear group
P SL(2, C) and the homogeneous Lorentz group SO(3, 1). Further, it is a subgroup of
the special linear group SL(2, C), and hence of the general linear group GL(2, C). In
turn, the Möbius group itself has some important and interesting subgroups.

Möbius transformations with real parameters, in which the coefficients a, b, c and d

are restricted to real numbers, comprise a group on their own, Möb (2, R). This group
is isomorphic to the projective linear group P SL(2, R) over the reals, a subgroup of
P SL(2, C). It is also a quotient group, being the special linear group SL(2, R) modulo
Z2 : we have
Möb (2, R) ∼
= SL(2, R)/Z2 ≡ P SL(2, R).
Möbius transformations with real parameters have some additional properties that are
of importance. It is easy to show that

• a real Möbius transformation maps the upper (respectively, lower) half of the
complex plane to the upper (respectively, lower) half-plane.

7. Verify the foregoing statement. Hint: Consider the Möbius transformation z → w =

(az + b)/(cz + d) where a, b, c and d are real numbers, and ad − bc = 1. Let z = x + iy
and w = u + iv. Show that v = y/|cz + d|2, so that Im w ≷ 0 according as Im z ≷ 0.

The modular group: Among Möbius transformations with real parameters (a, b, c, d)
where ad − bc = 1, those with integer values of the parameters form a group by them-
selves! This is the group P SL(2, Z), called the modular group Similarly, the set of
Möbius transformations in which (a, b, c, d) are Gaussian integers, i.e., each parameter
is of the form m + ni where m and n are integers, also forms a group. The modular
group has numerous remarkable properties, and it plays a role in many areas of math-
ematics including hyperbolic geometry, number theory, elliptic functions, etc. It also
appears in certain topics in theoretical physics such as conformal field theory.

The invariance group of the unit circle: Here is another important example of a
continuous subgroup of the Möbius group. Consider all Möbius transformations that

56
leave the unit circle unchanged, i.e., those that map the circle |z| = 1 in the z-plane
to the circle |w| = 1 in the w-plane. These transformations are either of the form
 
az + b a b
z → w = ∗ , where det ∗ ∗ = |a|2 − |b|2 = 1,
b z + a∗ b a
or of the form
 
az + b a b
z→
 w= , where det = −|a|2 + |b|2 = 1.
−b∗ z − a∗ −b −a∗
∗

Let’s call these type (i) and type (ii) transformations, for convenience. They form a
subgroup of the group Möb (2, C) of Möbius transformations. Transformations of types
(i) and (ii) are distinguished by the following property:

— A type (i) transformation maps the interior of the unit circle in the z-plane to
its interior in the w-plane; and the exterior of the unit circle in the z-plane to
the exterior of the unit circle in the w-plane.
— A type (ii) transformation maps the interior of the unit circle in the z-plane to
its exterior in the w-plane, and vice versa.
8. Show that
(a) the general form of a Möbius transformation z → w that maps the unit circle to
the unit circle is as given above;
(b) all such transformations form a group;
(c) |z| ≶ 1 ⇒ |w| ≶ 1 for type (i) transformations, and |z| ≶ 1 ⇒ |w| ≷ 1 for type
(ii) transformations.

Connection with the pseudo-unitary group SU(1, 1): The matrices correspond-
ing to Möbius transformations of types (i) and (ii) above are, respectively, of the general
form  
a b
U+ = ∗ ∗ where det U+ = |a|2 − |b|2 = 1,
b a
and  
a b
U− = where det U− = −|a|2 + |b|2 = 1.
−b −a∗
∗

But these are precisely the general forms of the matrices belonging to the indefinite
unitary group (or pseudo-unitary group) SU(1, 1). This is the group of (2 × 2)
matrices with complex entries satisfying the following conditions: if U ∈ SU(1, 1), then
 
† 1 0
det U = +1 and U g U = g, where g = .
0 −1

57
Now, SU(1, 1) turns out to be isomorphic to the special linear group SL(2, R) of
unimodular (2 × 2) matrices over the real numbers. Further, the group SL(2, R) is
also isomorphic to the symplectic group Sp(2, R), which is the group of canonical
transformations of a Hamiltonian system with one degree of freedom! We have

SU(1, 1) ∼
= SL(2, R) ∼
= Sp(2, R),
and the group of Möbius transformations that leaves the unit circle invariant is there-
fore isomorphic to SL(2, R)/Z2 , namely, Möb (2, R) once again.

The group of cross-ratios: An interesting discrete subgroup of Möb (2, C) is made

up of the following set of six Möbius transformations, which I will denote by ei where
i = 1, 2, . . . , 6:

e1 : z → z (the identity transformation) 



e2 : z → 1/z (inversion) 


e : z → 1 − z
3 (rotation about the point through an angle π) 
1
2
e4 : z→
 1 − (1/z) (inversion followed by rotation) 



e5 : z→ 1/(1 − z) (rotation followed by inversion) 



e6 : z→  z/(z − 1) (inversion-rotation-inversion).
The name ‘group of cross-ratios’ arises from the fact that these transformations are
precisely the ones leading to the identities (already listed) satisfied by the cross-ratio
of any four distinct points in Ĉ.

The transformation e3 can also be regarded as successive reflections about the line
Re z = 12 and about the real axis, performed in either order. Let ei ej stand for the
transformation ej followed by the transformation ei . It is obvious that e2 e2 = e1 =
e3 e3 , so that e−1
2 = e2 and e−13 = e3 . Further, e4 ≡ e3 e2 and e5 ≡ e2 e3 . The
last element e6 = e2 e3 e2 (inversion-rotation-inversion) or, alternatively, e6 = e3 e2 e3
(rotation-inversion-rotation).

9. Verify that the values of the parameters a, b, c, d corresponding to the foregoing

transformations (such that ad − bc = 1 in each case), and the ‘multiplication table’ for
the group, are as follows:
a b c d e1 e2 e3 e4 e5 e6
e1 1 0 0 1 e1 e1 e2 e3 e4 e5 e6
e2 0 −i −i 0 e2 e2 e1 e5 e6 e2 e4
e3 i −i 0 −i e3 e3 e4 e1 e2 e6 e5
e4 1 −1 1 0 e4 e4 e3 e6 e5 e1 e2
e5 0 1 −1 1 e5 e5 e6 e2 e1 e4 e3
e6 i 0 i −i e6 e6 e5 e4 e3 e2 e1

58
Note that this group is not an abelian or commutative group, since ei ej
= ej ei in
general. Several identities follow at once from the multiplication table above: for
instance, (e4 )3 , (e5 )3 , (e6 )2 , (e3 e4 )2 , (e4 e3 )2 , (e3 e5 )2 , (e5 e3 )2 , . . . are all equal to the
identity transformation.

59
6 Multivalued functions; integral representations
Branch points and branch cuts: An analytic function f (z) = w should be regarded
as a map f : z → w of some region R of the extended complex plane Ĉ to some region
R of Ĉ. If f (z) is single-valued, this map yields a unique value of w for every value
of z. All the functions we have considered so far satisfy this property. In general,
however, we have to deal with multivalued functions.

A simple example is provided by the function f (z) = z 1/2 . It is obvious that the
whole of the complex z-plane is mapped to just the upper half-plane in w, because
z = r eiθ maps to w = r 1/2 eiθ/2 . In order to cover the whole of the w-plane, the
z-plane has to be covered twice: the argument θ of z has to run from √ 0 to 4π, rather
than just 2π. As θ increases from 0 to 2π, we obtain the branch + z of the function √
f (z) = z 1/2 . As it increases further from 2π to 4π, we obtain the branch eiπ z 1/2 = − z
of the square root function.

In order to keep track of the correct branch, we therefore need two copies of the
z-plane. The two copies may be imagined to be two sheets, one lying below the other,
such that we descend to the second sheet by starting just above the positive real axis
on the first sheet and traversing a path that encircles the origin once in the positive
sense. On encircling 0 once again, this time on the second sheet, we ascend back to
the first sheet. The two sheets (labelled I and II, say) are called Riemann sheets,
and they are supposed to be connected to each other along a slit in each sheet running
from 0 to ∞ on the positive real axis. The top sheet, on which the phase of z runs from
0 to 2π, is called the principal sheet of the function concerned. The points where
the two branches of the two-valued function z 1/2 coincide in value, namely, z = 0 and
z = ∞ (recall that there is only one point at infinity in Ĉ), are called branch points.
They are connected by a branch cut. The two sheets pasted together as described
above form the Riemann surface of the function z 1/2 .

The branch cut joining z = 0 and z = ∞ may be taken to run along any curve
running between these two points. It is most convenient to take it to run along a
straight line. We have chosen the positive real axis in the foregoing. But this need not
always be so. All that is needed is a specification of the phase (or argument) of the
function concerned just above and just below the branch cut, so that we can calculate
the discontinuity (or the jump in the value) of the function across the branch cut.
Label the function on the sheets I and II as fI (z) and fII (z), respectively. By continuity,
the value of the function on sheet I, as you approach the positive real axis from below,
is the same as the value of the function on sheet II, as the real axis is approached from
above. That is,
lim fI (x − i) = lim fII (x + i), x > 0.

→0
→0

The discontinuity across the cut is then easily determined. In the present instance, it

60
is given by


disc f (z) = lim [fI (x + i) − fI (x − i)]
x>0
→0

= lim [fI (x + i) − fII (x + i)]

→0
√ √ √
= x − (− x) = 2 x.
Algebraic branch point: The branch points of z 1/2 at z = 0 and ∞ are algebraic
branch points. So are those of the function z 1/3 , for instance. In this case the Riemann
surface has three sheets. More generally, the function z p/q , where p and q are integers
with no common factors, has algebraic branch points at z = 0 and z = ∞. Its Riemann
surface comprises q sheets. Each sheet descends smoothly to the one below it as we
cross the branch cut; crossing the cut on the lowest sheet brings us back to topmost
sheet.

Winding point: The function z α , where α is not a rational real number, also has
branch points at z = 0 and z = ∞. These are called winding points. The Riemann
surface of this function has an infinite number of sheets, because e2πniα is never equal
to unity for any integer value of n. These sheets are labelled by an integer n ∈ Z. On
the principal sheet, the phase of z runs from 0 to 2π, as usual, and n = 0.

Logarithmic branch point: The function ln z has logarithmic branch points at

z = 0 and z = ∞. The Riemann surface is again infinite-sheeted, the sheets being
numbered by the full set of integers. On the principal sheet, ln z = ln r + iθ, where
0 ≤ θ < 2π. On the nth sheet, ln z = ln r + iθ + 2πni, where n ∈ Z and 0 ≤ θ < 2π.

Branch cuts run from one branch point to another. No function can have just one
branch point; the smallest number of branch points that a function can have is two.
When a branch point is encircled in the z-plane, the function does not return to its
original value. In order to obtain a closed contour that encircles a branch point, the
contour must encircle the branch point (or other branch points) as many times as is
necessary to return the function to its original value at the starting point.

Consider the functions (z − a)1/2 (z − b)1/2 and (z − a)1/2 /(z − b)1/2 , where a and b
are any two finite complex numbers. They have algebraic (square-root) branch points
at z = a and z = b, but their behavior at z = ∞ is regular. Therefore their branch
cuts can be chosen to run over the finite segments from a to b.

For any arbitrary non-integer value of α, including complex values, the cuts of the
function (z − a)α /(z − b)α can also be chosen to run over the finite segment from z = a
to z = b alone. But when α is not a half-odd-integer, this is not possible for the product
function (z − a)α (z − b)α . This function has branch points at z = a, z = b, as well as
z = ∞, for all α that is not an integer or half-odd-integer. The branch cut structure

61
of this function must necessarily run up to ∞.

The singularity structure of multivalued functions can be quite intricate and inter-
esting. Here’s a simple example. The function
1
f (z) = ln (1 − z)
z
has logarithmic branch points at z = 1 and z = ∞. On the principal sheet of the
logarithm, we have ln 1 = 0; hence ln (1 − z)  −z in the infinitesimal neighborhood
of the origin on this sheet. The simple pole at z = 0 of the factor z −1 is therefore ‘can-
celled’ by the simple zero of the logarithm, and f (z) only has a removable singularity
at z = 0. On every other sheet of the logarithm, however, f (z) does have a simple pole
at z = 0, because ln 1 = 2πni
= 0 on all these sheets.

Contour integrals in the presence of branch points: As you have seen, the eval-
uation of integrals via contour integration relies, ultimately, on integrating analytic
functions over closed contours. In the case of multivalued functions, however, if the
contour starts at some point z and encircles a branch point before returning to the
starting point, the function does not return to its original value. Hence the contour is
not really a closed one—the final point is actually on some other Riemann sheet of the
function!

In order to apply the theorems pertaining to integrals over closed contours, you
must ensure that the function has returned to its starting value (or, equivalently, that
z has returned to the starting point on the Riemann surface of the function). In gen-
eral, this involves encircling more than one branch point, or the same branch point
more than once, and so on, as will become clear from the examples that follow.

1. Let a and b be arbitrary real numbers, where a < b. Use contour integration to show
that the integral  b
dx
I=  = 2π,
a (b − x)(x − a)
independent of a and b.

Hint: Consider the function f (z) = (z − a)−1/2 (z − b)−1/2 . Note that the behavior
of the function at ∞ is regular. Choose the branch cut to run from a to b, and write
down the phases of the function above and below the cut. Relate the line integral I to
a contour integral surrounding the branch cut. Now open out the contour to make it a
large circle whose radius tends to infinity, and pick up the contribution from the circle
to arrive at the result. This method makes it evident why the value of the integral is
independent of the actual values of a and b.

62
2. Use a similar method to show that

dz
√ = 2πi,
C 1 + z + z2
where C is the circle |z| = 2 traversed once in the positive sense.

Hint: The branch points of the integrand are at two of the cube roots of unity, given
by z = ω = e2πi/3 and z = ω 2 = e4πi/3 . The branch cut of the integrand may be taken
to run between these points. As |z| → ∞ along any direction, the integrand tends to
1/z.

An integral involving a class of rational functions: Consider the definite integral

of a rational function of the form
 ∞
p(x)
I= dx ,
0 q(x)

where p(x) and q(x) are polynomials in x satisfying two conditions:

(i) The degree of q(x) exceeds that of p(x) by at least 2. Hence the integrand decays
at least as rapidly as 1/x2 as x → ∞, and the convergence of the integral is
guaranteed.

(ii) q(x) does not have any zeroes for x ≥ 0. Hence there is no non-integrable
singularity on the path of integration, and the integral exists.

As you know, such an integral can be evaluated by elementary means—for instance, by

resolving the integrand into partial fractions.13 But here is how you can evaluate I by
contour integration, using a simple trick. Consider, instead of I, the contour integral

 1 p(z) ln z
I =− dz ,
2πi C q(z)

where C is the familiar hairpin contour that comes in from infinity just below the
positive real axis, encircles zero from the left, and runs just above the positive real
axis. The branch cut of ln z is taken to run from 0 to ∞ along the positive real axis
in the z-plane. Since ln z = ln x just above the cut, and ln z = ln x + 2πi just below
the cut, we have
 0  ∞
 1 p(x) (ln x + 2πi) 1 p(x) ln x
I =− dx − dx = I.
2πi ∞ q(x) 2πi 0 q(x)
13
Obviously, you must then take care to avoid the appearance of any spurious logarithmic singularity
owing to the upper limit of integration. Evaluate the integral up to some upper limit L, and take the
limit L → ∞ after the different terms are recombined properly.

63
On the other hand, we can evaluate I  by completing the contour C with the addition
of a large circle (that does not cross the cut). The contribution of the large circle
vanishes as its radius R of the circle tends to infinity, because the integrand vanishes
at least as fast as (ln R)/R on this circle. But the closed contour can now be shrunk
to pick up the residues of all the poles of the integrand (located at the zeroes of q(z)).
The integral I is thus evaluated quite easily.

3. Use the method just described to evaluate the following standard integrals:
 ∞
dx π π
(a) n
= cosec (n = 2, 3, . . . ).
0 x +1 n n
 ∞
dx π 2π
(b) = cosec (n = 3, 4, . . . ).
0 xn−1 +xn−2 +···+ x+1 n n
Hint: (a) The poles of the integrand that are enclosed by C are at the roots of
z n + 1 = 0, namely, at z = eiπ/n ω r , where ω = e2πi/n and r = 0, 1, . . . , n − 1.

(b) The poles of the integrand that are enclosed by C are at the roots of unity other
than 1 itself, i.e., at z = ω r , where r = 1, . . . , n − 1.

Contour integral representations for various special functions are easily constructed
using the properties of multivalued functions. Such representations often provide ‘mas-
ter representations’ or analytic continuations of the functions concerned that are valid
for all complex values of the arguments of these functions. Let’s consider some exam-
ples.

The gamma function: Recall the defining representation of this function, namely,
 ∞
Γ(z) = dt tz−1 e−t , Re z > 0.
0

We’ve seen that the functional equation z Γ(z) = Γ(z + 1) helps us extend the region
of analyticity to the left of the region Re z > 0, stripwise. But we are now in a position
to write down a single representation for the gamma function that is valid throughout
the complex plane.

Consider the integrand tz−1 e−t in the defining representation as an analytic function
of t, for any arbitrary complex value of z. The factor tz−1 has a branch points at t = 0
and t = ∞, with a cut running between them. Take the branch cut to run along the
positive t-axis. The phase of tz−1 just above the cut is 0, while
 it is 2πiz just below
−2πi z−1 −t
the cut (the factor e is just unity). Consider the integral C dt t e , where the
contour C comes in from ∞ to  just below the cut, encircles the origin in the negative
sense in an arc of a circle of radius , and runs just above the cut from  to ∞. As
long as Re z > 0, the contribution from the arc of the small circle vanishes as  → 0.

64
Moreover, the contributions from the two line segments equal −e2πiz Γ(z) and Γ(z),
respectively. Hence, in the region Re z > 0, we have

1
Γ(z) = dt tz−1 e−t .
(1 − e2πiz ) C

But C does not pass through the point t = 0, and may be deformed to stay clear of
the origin, in the form of a hairpin-shaped contour straddling the branch cut of tz−1 .
The contour integral is thus defined for all finite values of z. On the other hand, the
factor (1 − e2πiz )−1 has simple poles at all integer values of z. The product of these
two factors is a meromorphic function of z. It represents Γ(z) for all z. Note that the
hairpin contour C can be partially ‘straightened out’, but we must always ensure that
Re t → +∞ asymptotically at both ends of the open contour, so that the damping
factor e−t ensures the convergence of the integral.

4. It can be checked that the known analytic properties of Γ(z) follow from the integral
representation derived above.

(a) Owing to the factor 1/(1 − e2πiz ), it might appear that Γ(z) has a simple pole
at every integer value of z. But we know that Γ(n) is finite (= (n − 1)!) when
n is a positive integer. What happens is that the contour integral C dt · · · also
vanishes when z = 1, 2, . . . : the integrand becomes single-valued because tn−1
does not have any branch points, and the two straight segments of the contour
cancel each other out. Verify that, as z tends to any positive integer n, the
limiting value of the right-hand side is precisely (n − 1)!.

(b) On the other hand, when z = −n where n = 0, −1, −2, . . . , the branch cut
disappears, but a pole of order (n + 1) is left at the origin in the t-plane. The
contour integral is then evaluated easily. Verify that Γ(z) has a simple pole at
z = −n with residue equal to (−1)n /n!.

The beta function: Recall the original definition of the beta function,
 1
B(z, w) = dt tz−1 (1 − t)w−1 , where Re z > 0, Re w > 0.
0

You have also seen that, in this case, it is not possible to extend the region of analyt-
icity in z and w simultaneously by using integration by parts.

Once again, we can continue B(z, w) analytically to all values of z and w by suit-
ably exploiting the discontinuity of the integrand tz−1 (1 − t)w−1 across the branch cut
running between the branch points at t = 0 and t = 1. For general values of z and w,
there is a branch point at t = ∞ as well. The branch cut structure of the integrand
is therefore more complicated in this case, because the cut actually runs all the way
to infinity. But all that we need to keep track of is the following fact: starting on the

65
principal sheet, the integrand acquires a factor e2πiz (respectively, e−2πiz ) whenever
the contour encircles the branch point at t = 0 in the positive (respectively, negative)
sense. Similarly, a factor e2πiw or e−2πiw results when the branch point at t = 1 is
encircled in the positive or negative sense.

What is required in order to return to the original value of the function, and hence
to close the contour, is a double encircling of each of the branch points t = 0 and
t = 1, once in the positive sense and once in the negative sense. The outcome is a
Pochhammer contour that is written symbolically as C : (1−, 0−, 1+, 0+).14 The
contributions from the infinitesimal circles around 0 and 1 vanish as long as Re z > 0
and Re w > 0. Therefore
 1 
1
B(z, w) = dt t (1 − t)
z−1 w−1
= dt tz−1 (1 − t)w−1 .
0 (1 − e−2πiz )(1 − e−2πiw ) C

But the contour of integration C does not pass through the singularities of the integrand
at t = 0 and t = 1. The contour can be distorted away from these points without
changing the value of the integral. The contour integral thus provides an analytic
continuation of the beta function for all finite values of z and w:

1
B(z, w) = dt tz−1 (1 − t)w−1 .
(1 − e−2πiz )(1 − e−2πiw ) C

The Riemann zeta function: Recall that the Riemann zeta function is defined as
∞
1
ζ(z) = (Re z > 1).
n=1
nz

Once again, an analytic continuation of ζ(z) to the whole of the complex z-plane is
easily derived, in terms of a contour integral. Consider the product
∞  ∞
1
ζ(z) Γ(z) = z
du uz−1 e−u , where Re z > 1.
n=1
n 0

Change variables of integration by setting u = nt. Then

∞ 
 ∞  ∞ 
∞
z−1 −nt
ζ(z) Γ(z) = dt t e = z−1
dt t e−nt
n=1 0 0 n=1
 ∞
tz−1
= dt , Re z > 1. (1)
0 (et − 1)
14
The term Pochhammer contour is used for the class of such contours that encircle several branch
points, each in a specific (positive or negative) sense.

66
 
The divergence of the sum ∞ n=1 n
−z
that occurs when Re z = 1 now becomes a diver-
gence of the integral: near t = 0, the integrand behaves like tz−2 , an extra factor of t−1
coming from the denominator (et − 1)−1 . Hence the integral diverges unless Re z > 1.

It is obvious that we cannot achieve convergence of the integral to the left of Re z =

1 by integration by parts, in this case. While tz → tz+1 upon integration, (et − 1)−1 →
(et − 1)−2 upon differentiation. The behavior of the integrand near t = 0 does not
improve. But we may exploit (once again) the branch cut of tz−1 running from t = 0 to
t = ∞ to convert the integral to a contour integral over a hairpin contour C straddling
the branch cut, encircling the origin in the negative sense. The contour is the same as
the one used for the gamma function. This gives

1 tz−1
ζ(z) = dt .
Γ(z) (1 − e2πiz ) C (et − 1)
The path of integration no longer passes through t = 0. This integral representation
is therefore valid for all z. Using the reflection formula Γ(z) Γ(1 − z) = π cosec (πz), it
is convenient to re-write it as

i −iπz tz−1
ζ(z) = e Γ(1 − z) dt t .
2π C (e − 1)
A number of interesting properties of the zeta function can now be deduced easily.

5. The contour integral in the formula above is an entire function of z. It follows that
the only possible singularities of ζ(z) in the finite part of the complex plane must come
from the poles of the factor Γ(1 − z) at z = 1, 2, . . . .
(a) Show that ζ(z) has a simple pole at z = 1, with residue equal to 1.
(b) Show that ζ(z) has no singularities at z = n where n = 2, 3, . . . .


Hint: (a) At z = 1, the contour integral reduces to dt/t = −2πi.

(b) The contour integral is an analytic function of z in the neighborhood of z = n

(where n = 2, 3, . . . ), and vanishes at those points. To find the coefficient of (z − n),
use the Taylor series tz−1 = tn−1 + (z − n) tn−1 ln t + · · · . The limiting value of ζ(z) as
z → n then works out to be

i tn−1 ln t
ζ(n) = dt t , n = 2, 3, . . . .
2π(n − 1)! C (e − 1)
Since the discontinuity of ln t across the cut on the positive real axis is 2πi, this further
simplifies to  ∞
1 tn−1
ζ(n) = dt t , n = 2, 3, . . . .
(n − 1)! 0 (e − 1)
Hence:

67
 • ζ(z) is a meromorphic function of z, with a simple pole at z = 1. The residue at
this pole is equal to unity.

Connection with Bernoulli numbers: We have seen earlier that ζ(2n), where n is
a positive integer, is π 2n multiplied by a rational number less than unity. I have also
mentioned there that no such simple closed-form expression exists for ζ(2n + 1). In
contrast, the value of ζ(z) at 0 and at the negative integers can be determined quite
easily, as follows.

The function 1/(et − 1) has a simple pole at t = 0, with residue equal to unity. The
function t/(et − 1) is analytic in the neighborhood of the origin.15 Its Taylor expansion
about t = 0 is
t ∞
tn
= Bn ,
(et − 1) n=0 n!
where the constants Bn are certain rational numbers, called the Bernoulli numbers.
These are defined by the expansion above. Thus t/(et − 1) is the generating function
for the Bernoulli numbers. The first few numbers are found to be

B0 = 1, B1 = − 12 , B2 = 16 , B4 = − 30
1
, B6 = 1
42
, ... .

Interestingly, all the odd Bernoulli numbers B2n+1 are equal to zero, except for B1 .

The sequence |B2 |, |B4 |, |B6 |, . . . appears to be a decreasing sequence. You might
therefore think that the series ∞ n
n=0 Bn t /n! ought to converge at least as well as the
series for the exponential et . The radius of convergence would then be infinite, and
the series would represent an entire function. But t/(et − 1), the function represented
by the series, has poles at t = 2πni, where n = ±1, ±2, . . . . Hence the radius of
convergence of the series must be the distance from the origin to the nearest of these
singularities, i.e., 2π. After the first few numbers, the Bernoulli numbers B2n actually
start increasing in magnitude quite rapidly with increasing n.

6. Use the expansion above in the representation of the zeta function to establish the
following results:

(a) ζ(0) = − 12 .

(b) ζ(−2n) = 0, where n = 1, 2, . . . .

B2n
(c) ζ(1 − 2n) = − , where n = 1, 2, . . . .
2n
15
The function has a removable singularity at the origin, and tends to unity as t → 0. We take this
to be the value of the function at t = 0, as usual.

68
Hint: When z is zero or a negative integer, the factor tz−1 in the integrand does not
have any branch points. Instead, there is a pole at t = 0. The contour C then collapses
to a small circle encircling the pole once in the clockwise sense. The integral is easily
evaluated by the residue theorem.

The Riemann Hypothesis: The zeroes of ζ(z) at even negative integer values of z
are called the trivial zeroes of the zeta function. ζ(z) also has an infinite number of
other zeroes in the strip 0 < Re z < 1.

• The famous Riemann hypothesis asserts that all of these nontrivial zeroes lie
on the so-called critical line Re z = 12 .

It is known that an infinite number of zeroes do lie on the critical line. It is also known
that all nontrivial zeroes lie on that line ‘almost surely’ (in the sense of probability
theory, with probability equal to 1). The first 13 billion or so zeroes have indeed been
verified explicitly to lie on the critical line.16 The Riemann conjecture is perhaps the
most important unsolved problem in mathematics. It has resisted a rigorous and com-
plete proof for over a century and a half.17 A very large number of other results in
mathematics rest on the validity of the hypothesis. There are also several intriguing
connections between the distribution of the zeros of ζ(z) on the critical line, on the
one hand, and physical problems, on the other—for instance, the level-spacing of the
energy eigenvalues of quantum mechanical systems whose classical counterparts are
chaotic, of the eigenvalues of certain classes of random matrices, and of the energy
eigenvalues of complex nuclei. It is clear that the zeta function and its counterparts
and analogs in classical and quantum dynamical systems hide many secrets yet to be
discovered.

The Legendre functions Pν (z) and Qν (z): As another example of integral repre-
sentations of special functions, let’s consider the Legendre functions of the first
and second kinds, Pν (z) and Qν (z), for complex z and ν. These functions appear
in the solution of a very large number of physical problems. They are the linearly
independent solutions of Legendre’s differential equation,
 2 
2 d d
(1 − z ) 2 − 2z + ν(ν + 1) φ(z) = 0.
dz dz
They turn out to have the contour integral representations given below, for complex
values of both the argument z and the order ν.

Consider the function f (t ; z, ν) = [(t2 − 1)/2(t − z)]ν of the complex variable t, for
general complex values of z and ν. There are branch points (winding points, in general)
16
ζ(z) can be shown to be a real analytic function. Therefore if ζ( 12 + iy) = 0 for any real value of
y, it follows that ζ( 12 − iy) = 0.
17
You might want to give it a try!

69
at t = 1, t = −1 and t = z. Suppose we start at some point on the principal sheet of
f (t; z, ν), and move in a path that encircles one or more of the branch points. In how
many different ways can this be done, such that the function returns to its original
value when we return to the starting point? In each case, the path then becomes a
closed contour over which f (t; z, ν) can be integrated.

If either t = 1 or t = −1 is encircled once in the positive sense, the function ac-

quires a factor e2πiν . If t = z is encircled once in the positive sense, it acquires a
factor e−2πiν . This shows that there are essentially three such independent paths. Us-
ing the Pochhammer notation, the contours are as follows. (i) C1 = (1+, z+); (ii)
C2 = (1+, −1−); and (iii) C3 = (−1+, z+). More complicated paths can be decom-
posed into linear combinations of these basic paths. (Convince yourself that this is
so.) The branch cut structure of f (t; z, ν) in each of the three cases is implicit in the
statement that Ci is a closed contour for the function. In case (i), there is a cut of
finite length running between t = 1 and t = z, and a cut from −1 to −∞ along the
negative real axis on the t-plane (say). In case (ii), there is a cut running from each
of the three branch points to infinity. Case (iii) is similar to case (i), with the roles of
the branch points at 1 and −1 interchanged.

The Legendre functions of the first and second kind are then given by the formulas
 2 ν
1 dt (t − 1)
Pν (z) =
2πi C1 (t − z) 2(t − z)

and  2 ν
1 dt (t − 1)
Qν (z) = .
π(e2πνi − 1) C2 (t − z) 2(t − z)
The analytic properties of Pν (z) and Qν (z) can be deduced from these representations.
Of the extensive set of these properties, I mention the following:

Pν (z) for non-integral order ν: For general non-integer values of the order ν, the
function Pν (z) is no longer a polynomial in z. It has branch points at z = −1 and
z = ∞. It is conventional to choose the branch cut to run from −1 to −∞ on the real
axis (the x-axis) in the z-plane. The discontinuity across the cut at any any point x is
again proportional to Pν (|x|) itself.

Qν (z) for non-integral order ν: Similarly, for general non-integer values of the
order ν, the function Qν (z) has branch points at z = 1, z = −1 and z = ∞. It is
conventional to choose the branch cut to run from 1 to −∞ on the x-axis. The dis-
continuity of Qν (x) for x ∈ (−1, 1) is proportional to Pν (x), while the discontinuity at
x < −1 is proportional to Qν (|x|). Based on these properties, it is possible to write
down dispersion relations for the functions Pν (x) and Qν (x).

70
Pl (z) for integer values of the order l: Consider Pν (z) when ν = l, where
l = 0, 1, . . . . The contour C1 now encloses no branch points, but only a pole of
order (l + 1) at t = z. The integral is then evaluated easily, and the Rodrigues
formula for Pl (z) is recovered:

1 dl 2 
l 1 dl 2
Pl (z) = l (t − 1)  ≡ (z − 1)l .
2 l! dtl t=z 2 l l! dz l

You would have encountered this formula in the case when z is a real variable lying
in the range [−1, 1]. We see now that this formula is also valid for complex values of
z. Note that Pl (z) continues to remain a polynomial of order l in the complex variable z.

When ν is a negative integer −(l + 1), the contour C1 encloses a pole of order (l + 1)
at t = 1. Once again, the integral can be evaluated. The symmetry property

P−l−1 (z) = Pl (z)

can then be deduced. This is just a special case of the more general reflection symmetry

P−ν−1 (z) = Pν (z)

that is valid for all complex values of ν.

Ql (z) for integer values of the order l: Turning to Qν (z), when ν = l (= 0, 1, . . . ),

the contour C2 encloses no singularity at all. Hence the contour integral vanishes. But
so does the factor (e2πνi − 1) in the denominator of the integral representation for
Qν (z). Their ratio has a finite limit as ν → l. The outcome is the function Ql (z),
which turns out to have logarithmic branch points at z = ±1. For instance,
 
1 1+z
Q0 (z) = ln .
2 1−z

In general, for positive integer values of l, we have

 
1 1+z
Ql (z) = Pl (z) ln + Rl−1 ,
2 1−z

where Rl−1 is a polynomial of order l−1. There is a simple and very useful formula that
connects Ql (z), for any arbitrary complex value of the argument z, to the Legendre
polynomial of order l. It is

1 1 Pl (t)
Ql (z) = dt , l = 0, 1, . . . .
2 −1 (z − t)

71
 When ν is a negative integer −(l + 1), the contour C2 encloses poles of order (l + 1)
at z = −1 and z = 1. The contour integral makes a finite, nonzero contribution. The
factor (e2πνi −1) in the denominator then leads to a simple pole of Qν (z) at ν = −(l+1).
The residue at the pole turns out to be Pl (z) itself.

Finally, I mention that such contour integral representations exist for all other
special functions as well. In most cases, they serve as analytic continuations of the
corresponding functions to the largest possible domain of their arguments, including
parameters such as the order, degree, and so on.

Singularities of functions defined by integrals: As I’ve just stated, integral

representations of functions are most useful for exhibiting their analytic properties.
The natural question that arises is the following:

• How does a function defined by an integral become singular, and what are the
possible singularities?

As you might expect, the answer to such a general question involves whole branches of
mathematics (such as homology theory). Let’s narrow the question down very con-
siderably, and examine the simplest possibilities. The treatment below is elementary,
heuristic, and essentially based on a few simple examples.

Consider functions of the form

 b
f (z) = dt φ(t, z),
a

where the path of integration is an open contour running from some point a to some
other point b in the complex t-plane. The integrand φ(t, z) is assumed to be analytic
in t and z, with some singularities. In general, when the integral exists, it defines an
analytic function of z. Let’s start with a value of z in the region in which f (z) is
holomorphic, and examine how the function could develop a singularity as z moves out
of this region. This is determined by the behavior of the singularities in the t-plane of
the integrand φ(t, z).

These singularities are of two kinds: they could be fixed or z-independent, such as
a pole at some point t = c (not lying on the original path of integration); or they could
be movable, or z-dependent, such as a pole at t = z. As z changes, one or more of
the latter could approach the path of integration in the t-plane. If the contour can be
distorted away so as to avoid the singularity (keeping it pinned down at the end-points
a and b, of course), we have an analytic continuation of f (z), defined by the integral
over the distorted contour. There are two cases in which this simple device will not
work. Each of them leads to a singularity of f (z).

72
(i) End-point singularity: If a moving singularity approaches one of the end-points
of integration, a or b, the contour cannot be moved away and an end-point singu-
larity of f (z) ensues. Consider the (extremely!) elementary example
 1
dt
f (z) = .
0 t−z

The integrand has a moving pole at t = z. As it stands, the integral exists for all
z
∈ [−1, 1]. If z approaches the real axis in the t-plane either from above or from below,
the contour of integration can be moved away ahead of the pole, and the integral will
continue to exist in each case.18 As z → 0 or 1 (in the z-plane), the pole in the t-
plane approaches one of the end-points of the contour, and a singularity of f (z) occurs.
This is corroborated by the explicit form for f (z) obtained by carrying out the trivial
integral above, to get
z − 1
f (z) = ln .
z
Clearly, f (z) has logarithmic branch points at z = 0 and z = 1, confirming our expec-
tation.

A little more generally, consider the function

 1
φ(t)
f (z) = dt ,
0 (t − z)
where φ(t) is a well-behaved function (e.g., a polynomial in t) such that the integral
exists as long as z
∈ [0, 1]. Once again, f (z) has end-point singularities at z = 0 and
z = 1, and these are again logarithmic branch points. Note the important point that
it is not necessary to be able to carry out the integral defining f (z) explicitly in order
to reach this conclusion! The discontinuity of f (z) across the branch cut running from
z = 0 to z = 1 on the real axis in the z-plane is also computed easily. All you have to
do is to apply the formula we have already encountered (implicitly) when discussing
dispersions relations: namely,
1 1
=P ± iπ δ(t − x).
t − x ∓ i t−x
Here t and x are real variables, P stands for the Cauchy principal value, and the formula
is to be understood in the following sense: multiply both sides by a smooth function
of t, and integrate with respect to t over a range that includes the point x. The result
is  % &
 def.
disc f (z) = lim f (x + i) − f (x − i) = 2πi φ(x).
z=x∈(−1,1)
↓0

1
7. Apply this result to the formula Ql (z) = 1
2 −1
dt Pl (t)/(z − t), to show that
18
As you might guess, however, the analytic continuations of f (z) will differ in the two cases,
suggesting already that we’re dealing with different branches of f (z).

73
 (a) the Legendre function Ql (z) (where l = 0, 1, . . . ) has logarithmic branch points
at z = ±1 (as stated earlier);
(b) the discontinuity across the branch cut running from −1 to 1 on the real axis is
given by
 % &

disc Ql (x) = lim Ql (x + i) − Ql (x − i) = −iπ Pl (x).
−1<x<1
↓0

(ii) Pinch singularity: If two moving singularities approach the same point on
the path of integration, but from opposite sides of the path, the contour gets trapped
between them, and cannot ‘escape’. A pinch singularity of f (z) may then occur.
The same thing happens if a moving singularity traps the contour between itself and a
fixed singularity of the integrand lying on the other side of the contour. Consider the
function  1
dt
f (z) = .
−1 (z − t )
2 2

The integral exists and defines an analytic function of z, as long as z

∈ [−1, 1]. The
integrand has moving poles at t = ±z. We may expect end-point singularities of f (z)
at z = ±1. Further, the poles at t = z and t = −z pinch the contour of integration
from opposite sides as z → 0, and so a pinch singularity may be expected at z = 0.
Explicit evaluation of the integral gives
1
z + 1
f (z) = ln .
z z−1
Thus, f (z) does have singularities (logarithmic branch points) at z = ±1. It also has
a pole at z = 0, on every sheet of the logarithm except the principal sheet (on which
ln 1 = 0).19 As before, a similar analysis is applicable to the more general integral
 1
φ(t)
f (z) = dt 2 ,
−1 z − t2
where φ(t) is a well-behaved function such as a polynomial.

A slightly more complicated example: The nature of an end-point or pinch

singularity depends also on the kind of moving singularities involved. In the preceding
examples, these were simple poles. As an example of what can happen when two
branch-points pinch the contour of integration, consider the integal
 1
dt
f (z) = √ .
−1 z 2 − t2
Again, the integral is very easily evaluated. But let’s first list the possible singularities
of the integral as it stands. The integrand has square-root branch points at t = z and
19
Recall that you have already encountered an example of this feature.

74
t = −z. End-point singularities may be expected at z = ±1. We may expect these to
be ‘mild’ singularities, since (t ± z)−1/2 is an integrable singularity. (It is trivially seen
that f (z) does not diverge at z = ±1, but rather has the finite value π.) Further, as
z → 0, the contour of integration is pinched between these branch points. Hence z = 0
must also be a singularity of f (z). Note that if we simply set z = 0 in the expression
for f (z), the integral diverges.

In order to find f (z) explicitly, let’s start with z on the positive real axis, z = x > 1.
Then a simple change of the variable of integration yields

f (x) = 2 sin−1 (1/x).

This is a multivalued function, and you must be careful about its branch structure
when writing down its analytic continuation to the rest of the z-plane. It is convenient
to re-write the arcsine function as a logarithmic function, using the identity
% &
sin−1 u = −i ln iu ± (1 − u2 )1/2 .

In order to choose the right sign before the radical, note the following: When 1 ≤
x < +∞, the integral representing f (x) is real positive; and it decreases monotonically
from π to 0 as x → ∞. It follows that

x 
f (x) = 2i ln √ , 1 ≤ x < ∞.
i + x2 − 1
In this form, the function is (trivially) analytically continued to

z 
f (z) = 2i ln √ .
i + z2 − 1
It is now easy to see that f (z) has
(i) a logarithmic branch point at 0, as well as

(ii) square-root branch points at z = 1 and z = −1,

corroborating our earlier conclusions.

Singularities of the Legendre functions: Finally, let’s apply the foregoing to

the integral representations written down earlier for the Legendre functions Pν (z) and
Qν (z). It should now be obvious that, for general values of the index ν, this is what
happens:
• As z → −1, the contour C1 gets pinched between the moving singularity of the
integrand at t = z and the fixed singularity at t = −1. Hence Pν (z) has a
singularity at z = −1. The branch cut of Pν (z) is customarily taken to run from
−1 to −∞ along the negative real axis.

75
 • As z → ±1, the contour C2 gets pinched between the singularities of the inte-
grand at t = z and at t = ±1, respectively. Hence Qν (z) has a singularities at
z = 1 and z = −1. The branch cut of Qν (z) is customarily taken to run from 1
through −1 to −∞ along the real axis.

8. With a little effort, you can show that the discontinuities of Pν (z) and Qν (z) across
their branch cuts on the real axis are as follows:
disc Pν (x) = 2i sin (πν) Pν (−x) for −∞ < x < −1
disc Qν (x) = 2i sin (πν) Qν (−x) for −∞ < x < −1
disc Qν (x) = −iπ Pν (x) for −1 < x < 1.

Hint: In order to find the discontinuity of Pν (z) across the cut from −1 to −∞, con-
sider its integral representation in the respective cases when z = x + i and z = x − i,
where x < −1. Note the configuration of the contour C1 in the two cases, along
with the branch cuts of the integrand. Write down the contour integrals segment by
segment, keeping careful track of the phases of the various factors in the integrand.
Compare the result with the contour integral for Pν (−x) (where −x > 1), to arrive
at the result quoted. A similar procedure will yield the corresponding results for Qν (x).

As expected, the discontinuity of Pν (z) vanishes when ν = l, an integer. Pl (z) is a

polynomial in z, and it has no singularity at z = −1 or at any other point in C, the
finite part of the complex z-plane. Similarly, Qν (z) has no branch cut running from −1
to −∞ when ν = l; the discontinuity across this cut vanishes identically. The discon-
tinuity of Qν (z) between −1 and 1 is nonzero even when ν = l, as you have seen already.

Dispersion relations for the Legendre functions: Finally, based on the disconti-
nuities found above, we can derive dispersions relations for the Legendre functions of
both kinds. As you might expect, the asymptotic (|z| → ∞) behaviors of Pν (z) and
Qν (z) are also required for this purpose. I do not go into this here, but merely quote
the relations, for completeness:

sin πν ∞ Pν (t)
Pν (z) = dt (ν
= integer),
π 1 (z + t)

and  
1 1 Pν (t) sin πν ∞ Qν (t)
Qν (z) = dt + dt .
2 −1 (z − t) π 1 (z + t)
1
The relation Ql (z) = 12 −1 dt Pl (t)/(z − t) when ν = l (= 0, 1, . . . ) is a special case of
the last formula above.

76
7 Laplace transforms
In physical applications, we are often concerned with functions that are only defined
on a half-line, say [0, ∞). An example is a causal response function φ(t), where t
denotes the time variable. More generally, the whole class of initial value problems
involves functions of this kind. Such functions may vanish, or tend to a constant, or
even diverge as t → ∞. It is very helpful to define an integral transform of the
functions concerned, that turns differentiation with respect to t into multiplication by
the variable ‘conjugate’ to t.

Definition of the Laplace transform: The Laplace transform of a function f (t)

(where 0 ≤ t < ∞) is defined as
 ∞
 def.
L[f (t)] = f (s) = dt e−st f (t).
0
−st
It is clear that the factor e provides a convergence factor if Re s > 0 : the integral
above converges even if f (t) increases like any arbitrary power of t for large values of
t. In fact, even if f (t) increases exponentially with t, so that f (t) ∼ ect (where c is
a positive number) as t → ∞, its Laplace transform as given by the integral above is
well-defined, as long as s is held in the region Re s > c. Thus :
• The Laplace transform f(s) of a function f (t), as given by its defining integral
representation, is an analytic function of s for a sufficiently large positive value
of Re s, i.e., in some right-half plane in s.
In general, we may then expect to be able to define it to the left of this region by
analytic continuation.20 It is evident that f(s) would have one or more singularities in
the left half-plane, in general.

The Laplace transforms of simple functions are easily written down. Consider, for
instance, the function
f (t) = tα e−at .
It is clear that its Laplace transform
 ∞
f(s) = dt tα e−(s+a)t
0

exists as long as Re α > −1 (so that the lower limit of integration, t = 0, does not
pose any problem), and also Re s > −Re a (so that there is no divergence owing to the
upper limit of integration, t = ∞). The integral is easily seen to be a gamma function.
We have
Γ(α + 1)
f (t) = tα e−at (Re α > −1) =⇒ f(s) = .
(s + a)α+1
20
Note that there do exist functions which do not possess a Laplace transform—e.g., if f (t) ∼
exp (t1+α ) (where α > 0) as t → ∞, then f(s) does not exist for any value of s.

77
f(s) may now be continued analytically to all values of s, using the explicit representa-
tion above. Observe that f(s) has a singularity at s = −a, in accord with the general
statement made earlier.

A number of simpler cases may be read off from this result. For instance,
1 n! s a
L[e−at ] = , L[tn ] = n+1 , L[cos (at)] = 2 2
, L[sin (at)] = 2 .
(s + a) s s +a s + a2

In general, if L[f (t)] = f(s), then L[f (t) e−at ] = f(s + a).

1. It is easy to see that   ∞

∞
dt f (t)
= ds f(s),
0 t 0
provided both integrals exist. Numerous definite integrals may be evaluated with the
help of this identity. Use it to recover the known result for the Dirichlet integral,
 ∞
sin (at)
dt = 12 π sign (a)
0 t
where a is a real constant.

2. The convolution theorem for Laplace transforms:

(a) If f (t) and g(t) have Laplace transforms f(s) and  g (s), show that their convolu-

tion has the Laplace transform f (s)  g(s). That is,
 t   t 
L   
dt f (t ) g(t − t ) = L dt f (t − t ) g(t ) = f(s) 
  
g(s).
0 0
 
dt f (t ) = f(s)/s.
t
Hence note that L 0

(b) Show that

 t  tn  t2   n
L dtn dtn−1 · · · dt1 f (t − tn ) f (tn − tn−1 ) . . . f (t2 − t1 ) = f(s) .
0 0 0

3. Laplace transforms of derivatives : Using integration by parts, it is easy to see

that
L[df (t)/dt] = s f(s) − f (0).
Let f (n) (t) ≡ dn f (t)/dtn , with f (0) (t) ≡ f (t). Check that

L[f (n) (t)] = s L[f (n−1) (t)] − f (n−1) (0), n ≥ 1.

78
Hence show that
L[f (n) (t)] = sn f(s) − sn−1 f (0) − sn−2 f (1) (0) − . . . − s f (n−2) (0) − f (n−1) (0)
n
n 
= s f (s) − sn−j f (j−1) (0).
j=1

Thus the Laplace transform essentially converts differentiation (with respect to t)

to multiplication (by s). This fact is of great use in the applications of Laplace
transforms—among others, in the solution of linear differential equations with constant
coefficients. Note the occurrence of the ‘initial data’ (the values of the function and
its first (n − 1) derivatives at t = 0) in the expression for the transform of dn f (t)/dtn .

The inverse transform: The inverse Laplace transform that expresses f (t) in terms
of f(s) is given by the Mellin formula
 c+i∞
1
f (t) = ds est f(s),
2πi c−i∞
where the contour runs parallel to the imaginary axis in the s-plane, cutting the real
axis at a point c such that the contour stays to the right of all the singularities of f(s).
Hence the contour lies entirely in a region in which f(s) is analytic, as required to make
the integral well-defined. (Recall that f(s) generally has singularities in a certain left
half-plane.) Using the Mellin formula, it is easy to invert a Laplace transform whenever
f(s) is a rational function of s. The contour over s can then be closed by adding a
large semi-circle in the left half-plane, because the contribution of the semi-circle to
the integral vanishes as its radius tends to infinity: the factor est in the Mellin formula
ensures this. (Remember that t ≥ 0.) The closed contour now encircles all the poles
of f(s). The residue theorem then gives the value of the contour integral, and hence
f (t) is determined.

4. Show that the inverse transform of f(s) = (s2 + a2 )−2 is given by

1
f (t) = 3 [sin (at) − at cos (at)].
2a

5. The Bessel function of the first kind and of order ν may be defined by the power
series
∞
(−1)n  1 2n+ν
Jν (z) = 2
z .
n=0
Γ(ν + n + 1) n!
The order ν does not necessarily have to be an integer, in this definition. Jν (z) is an
entire function of z. When ν = 0, we have
∞
(−1)n  1 2n
J0 (z) = 2 2
z .
n=0
(n!)

79
Now consider the function f(s) = (s2 +a2 )−1/2 , where a is a real constant (say). Expand
it in a binomial series in inverse powers of s, and invert the transform term by term
using the fact that the inverse transform of 1/s2n+1 is t2n /(2n)! Hence show that

−1 1
L √ = J0 (at).
s2 + a2
Since  ∞
1
dt e−st J0 (at) = √ ,
0 s2 + a2
∞
it follows that 0
dt J0 (at) = 1/a for a > 0.

6. The iterate of the Laplace transform: Show that, provided the integrals
concerned exist,
 ∞  ∞  ∞
−us −st f (t)
L [f (t)] =
2
ds e dt e f (t) = dt .
0 0 0 (t + u)

The last integral is the so-called Stieltjes transform of f (t).

LCR series circuit: As another standard example from elementary physics, consider
an LCR series circuit under a sinusoidal applied voltage of amplitude V0 and (angular)
frequency ω. Recall that we have already determined the complex admittance of the
system, when discussing linear response and dispersions relations.

The differential equation satisfied by the charge q(t) on the capacitor is

L q̈ + R q̇ + (1/C) q = V0 sin ωt.

As you know, this equation is precisely the equation of motion of a sinusoidally forced
damped simple harmonic oscillator. The damping constant is R/L = γ, which is just
the reciprocal of the time constant
√ of an LR circuit. The natural frequency of the circuit
in the absence of the resistor 1/ LC = ω0 . The condition ω0 > 12 γ corresponds to the
underdamped case. Let’s consider this case first , for definiteness.21 It is convenient,
in this case, to work in terms of the shifted frequency
 1/2
ωu = ω02 − 14 γ 2 .

7. Suppose the initial conditions are such that both the initial charge and the initial
current are equal to zero, i.e.,

q(0) = 0 and q̇(0) = 0.

21
The critically damped and overdamped cases will be considered subsequently.

80
The solution for q(t) can be written as the sum of a transient part and a steady state
part,
q(t) = q tr (t) + q st (t).
Show that these are given, respectively, by
  2 
tr V0 ω (ω − ω02 + 12 γ 2 ) sin (ωu t) + ωu γ cos (ωu t) −γt/2
q (t) = e
L ωu (ω 2 − ω02 )2 + ω 2γ 2
and   
V0 (ω 2 − ω02 ) sin (ωt) + ω γ cos (ωt)
q (t) = −
st
.
L (ω 2 − ω02 )2 + ω 2 γ 2

Hint: Take the Laplace transform of both sides of the differential equation. With the
initial conditions q(0) = 0 and q̇(0) = 0, we get
(V0 ω/L)
q(s) = .
(s2 + ω )(s2 + γs
2 + ω02)
Resolve the right-hand side into partial fractions, and invert the Laplace transform.
All you need to use is the fact that the inverse Laplace transform of (s + a)−1 is e−at .

Observe that the transient component of the solution, q tr (t), is characterized by the
frequency ωu that depends on the circuit parameters L, C and R. This part decays to
zero exponentially in time, owing to the dissipation in the system. The steady state
component q st (t), on the other hand, oscillates with the same frequency ω as the ap-
plied voltage. These statements apply equally to the current in the circuit, given by
I(t) = q̇(t).

8. The remarks just made are, in fact, applicable to all initial conditions. In order to
check this out, consider general values q(0) and q̇(0) = I(0), respectively, of the initial
charge on the capacitor and the initial current in the circuit. Show that the complete
solution for q(t) is now given by the one already found above, plus an extra term added
to q tr (t), namely,
  
I(0) + 12 γ q(0)
sin (ωu t) + q(0) cos (ωu t) e−γt/2 .
ωu

Complementary function and particular integral: Note, incidentally, that the

last expression above is precisely the solution to the homogeneous differential equation
L q̈ + R q̇ + (1/C) q = 0 that is satisfied by the charge on the capacitor in the absence of
any applied voltage. With reference to the inhomogeneous differential equation for q(t),
the full solution given above represents the particular integral, while the solution of
the homogeneous equation represents the complementary function.

81
 • In the solution of an inhomogeneous differential equation, the purpose of adding
‘the right amount’ of the complementary function to the particular integral is
to ensure that the boundary conditions (in this case, the initial conditions) are
satisfied by the solution.

The example just considered ought to help you see this quite clearly.

9. By now, it should be obvious to you that the solutions in the critically damped
(ω0 = 12 γ) and overdamped (ω0 < 12 γ) cases may be written down by simple analytic
continuation of the solution in the underdamped case.22 Do so.

Laplace transforms and random processes : The master equations for certain
Markov processes lead to simple differential equations for the generating functions of
these processes. Such equations are solved very easily using Laplace transforms. Here
are a few examples.

10. The Poisson process: Radioactive decay of an unstable isotope provides a phys-
ical example of a Poisson process. If Pn (t) is the probability that exactly n events of
the process take place in a time interval t, and λ > 0 is the mean rate at which events
take place, then

dP0 (t) dPn (t) % &

= −λ P0 (t) and = λ Pn−1 (t) − Pn (t) , n ≥ 1.
dt dt
This coupled set of ordinary differential equations is to be solved with the initial con-
−λt
0∞for n ≥ 1.n It follows at once that P0 (t) = e23 . The
ditions P0 (0) = 1 and Pn (0) =
generating function f (z, t) = n=1 Pn (t) z satisfies the differential equation

∂f
+ λ(1 − z) f = λz P0 = λz e−λt ,
∂t
with the initial condition f (z, 0) = 0. Use Laplace transforms to obtain the solution
 
f (z, t) = e−λt eλzt − 1 .

Picking out the coefficient of z n in the power series expansion of the exponential we
get the Poisson distribution
e−λt (λt)n
Pn (t) = .
n!
The mean number of events in a time interval t is therefore λt. Every cumulant of the
distribution is also equal to λt.

22
Typically, trigonometric functions will become hyperbolic functions.
23
Note that the generating function has been defined as a sum from n = 1 rather than n = 0.

82
 The Poisson process is an example of what is known as a birth process, because
the random variable n(t) never decreases as t increases. The next example shows what
happens when the value of the random variable can both increase as well as decrease
as time elapses (a birth-and-death process).

11. A biased random walk on a linear lattice : Consider a linear lattice, with its
sites labelled by an integer j ∈ Z. A walker on this lattice jumps from any site to one
of the two neighboring sites with a mean jump rate λ. The probability of a jump from
the site j to the site (j + 1) is p, and that of a jump to the site (j − 1) is q = (1 − p),
where 0 < p < 1. Successive jumps are statistically independent of each other. Let
Pj (t) denote the probability that the walker is at site j at time t. (The random variable
in this problem is j). The master equation satisfied by the set of probabilities {Pj (t)}
is
dPj (t) % &
= λ p Pj−1(t) + q Pj+1(t) − Pj (t) , j ∈ Z.
dt      
‘gain’ terms ‘loss’ term

Without loss of generality,24 we may take the site 0 to be the starting point of the
random walk (RW for short). The initial condition is then

Pj (0) = δj,0 .

Define the generating function


∞
f (z, t) = Pj (t) z j .
j=−∞

Note that the summation is over all integers j. Hence the equation above is to be
regarded as a Laurent series expansion of f (z, t).

(a) From the differential equation satisfied by Pj (t) and its initial condition, show
that f (z, t) satisfies the equation

∂f
= λ (pz − 1 + qz −1 ) f ,
∂t
with the initial condition f (z, 0) = 1 .

(b) Solve this equation using Laplace transforms, to get

−1 )
f (z, t) = e−λt eλt (pz+qz .
24
Because the lattice is of infinite extent in both directions! Finite boundaries would obviously spoil
this translation invariance.

83
 (c) The probability we seek, Pj (t), is the coefficient of z j in the Laurent expansion
of f (z, t) in powers of z. It is helpful to write

−1 √
 −1
 
pz + qz = pq z p/q + z q/p .

Now expand the second exponential factor in the solution for f (z, t), and pick
out the coefficient of z j . Consider j ≥ 0 first. Show that

∞
1 √
−λt
Pj (t) = e (p/q) j/2
(λt pq)2n+j , j ≥ 0.
n=0
n! (n + j)!

The modified Bessel function of the first kind and of order ν may be defined by
means of its power series, namely,

∞
1  1 2n+ν
Iν (z) = z .
n=0
Γ(ν + n + 1) n! 2

Compare this series with that for the ordinary Bessel function Jν (z) given earlier.
Apart from the factor (−1)n in the summand in the case of Jν (z), the series are the
same.25 Like the Bessel function Jν (z), the modified Bessel function Iν (z) is also an
entire function of z. The generating function for the modified Bessel function of integer
order is26

∞
t (z+z −1 )/2
e = Ij (t) z j .
j=−∞

Owing to the symmetry of the generating function under the exchange z ↔ z −1 , it is

obvious that I−j (z) = Ij (z) for every integer value of j.

(d) Compare the series expansion obtained above for Pj (t) with the power series for
the modified Bessel function, to obtain
√
Pj (t) = e−λt (p/q)j/2 Ij (2λt pq) .

Verify that exactly the same expression is valid for negative integer values of j
as well, using the fact that I−j = Ij for any integer j.

25
The two functions are related to each other according to Jν (z) = eiπν/2 Iν (−iz), or Iν (z) =
−iπν/2
e Jν (iz).
26
For completeness, I mention that the generating function of the Bessel function Jk is given by
∞

−1
et (z−z )/2
= Jj (t) z j .
j=−∞

84
 An RW in which p
= q is called a biased random walk. When p > q [respectively,
q > p], there is a bias to the right [left], and the factor (p/q)j/2 in the expression for
Pj (t) shows that the probability of positive [negative] values of j is enhanced, at any
t > 0.

In the case of an unbiased random walk, p = q = 12 . The probability distribution

then simplifies to

Pj (t) = e−λt Ij (λt) (unbiased RW in 1 dimension).

Since Ij = I−j , we have in this case Pj (t) = P−j (t) at any time t, as expected on
physical grounds.

The asymptotic or long-time behavior of the RW is of importance. In the problem

at hand, the characteristic time scale is set by λ−1 . Now, for λt  1, the leading
asymptotic behavior of the modified Bessel function is given by

eλt
Ij (λt) ∼ √ ,
2πλt
independent of j. Therefore, in an unbiased RW, Pj (t) has a leading asymptotic time-
dependence ∼ t−1/2 , which is characteristic of purely diffusive behavior.27 In marked
contrast, the leading asymptotic behavior of Pj (t) for a biased random walk is given
√
−1/2 −λt(1−2 pq) √
by Pj (t) ∼ t e . Since 1 > 2 pq when p
= q, this shows that Pj (t) decays
exponentially with time for all j.

12. Unbiased random walk in d dimensions: Consider an unbiased simple RW

on an infinite hypercubic lattice in d dimensions. Each site is labelled by a set of d
integers, (j1 , . . . , jd ) ≡ j. The walker jumps from any given site to any one of the 2d
nearest-neighbor sites with a probability 1/(2d), with a mean transition rate λ. Let
P (j, t) be the probability that the walker is at j at time t, given that the walk started
at the origin 0 at t = 0. The master equation satisfied by P (j, t) is

dP (j, t) λ 
= P (j + δ, t) − λP (j, t) ,
dt 2d δ   
   loss term
gain term

where δ stands for a nearest-neighbor vector of the site j. (That is, δ has one of its d
components equal to ±1, and all other components equal to 0.) The initial condition
is 
1, if j = 0
P (j, 0) =
0, if j
= 0.
27
We’ll study the diffusion equation and its properties later on.

85
Define the generating function

f (z1 , . . . , zd , t) = P (j, t) z1j1 z2j2 · · · zdjd ,
j

where each component of j is summed over all the integers.

(a) Show that the equation satisfied by this generating function is

∂f  z + z −1 + · · · + z + z −1 
1 d
=λ 1 d
− 1 f,
∂t 2d
with the initial condition f = 1 at t = 0.

(b) The solution to this equation is clearly an exponential that factors into a product
of generating functions of modifed Bessel functions. Show that

P (j, t) = e−λt Ij1 (λt/d) Ij2 (λt/d) · · · Ijd (λt/d).

The leading asymptotic (t → ∞) behavior of P (j, t) follows from that of the modified
Bessel function. We find P (j, t) ∼ t−d/2 . This specific power-law decay of the proba-
bility is characteristic of purely diffusive motion.

13. The generalization of the solution above to include any directional bias is straight-
forward. Let pi and qi be the
 probabilities of a jump in which ji changes to ji + 1 and
ji − 1, respectively. Then d1 (pi + qi ) = 1.

(a) Write down the master equation satisfied by P (j, t).

(b) Show that the solution for an RW starting at the origin at t = 0 is


d
 √ 
−λt
P (j, t) = e (pi /qi )j/2 Iji 2λt pi qi .
i=1

As in the one-dimensional case, the long-time behavior of P (j, t) is now a decaying

exponential in t, rather than a pure power-law fall-off.

86
8 Fourier transforms
Fourier integrals: As you know, periodic functions can be expanded in Fourier
series. Recall that, if the fundamental interval of the function f (x) is (a , b) so that
the period is L = (b − a), the expansion and inversion formulas are

∞  b
f (x) = (1/L) fn e 2πnix/L
and fn = dx f (x) e−2πnix/L .
n=−∞ a

What happens if the function is not periodic? By letting a → − ∞ and b → ∞ ,

and hence passing to the limit L → ∞ , we can extend the idea of the expansion of
an arbitrary periodic function in terms of elementary periodic functions (sines and
cosines) to functions that need not be periodic. The number of ‘harmonics’ required
now becomes uncountably infinite. Therefore, instead of a summation over the integer
index n, we require an integration over the continuous variable k to which 2πn/L tends
in the limit. The general coefficient in the original expansion, fn , becomes a function
of the continuous variable k. In order to avoid confusion with f (x), we may denote
this function by f(k). The ‘dictionary’ to go from Fourier series to Fourier integrals is
 ∞
1 
∞
2πn 1
−→ k , −→ dk , fn −→ f(k) .
L L n=−∞ 2π −∞

We then have the expansion formula

 ∞
1
f (x) = dk eikx f(k) (expansion formula).
2π −∞
The inversion formula is
 ∞
f(k) = dx e−ikx f (x) (inversion formula).
−∞

The function f(k) is obtained from the function f (x) by acting upon the latter with
an integral operator whose kernel is exp (−ikx). Similarly, the function f (x) is ob-
tained from the function f(k) by acting upon the latter with an integral operator whose
kernel is (2π)−1 exp (ikx). The functions f (x) and f(k) are Fourier transforms of
each other.28 A Fourier transform is an example of an integral transform. A func-
tion f (x) has a Fourier transform if it satisfies conditions analogous to the Dirichlet
conditions for periodic functions. Broadly speaking, if
(i) f (x) has at most a finite number of finite discontinuities or jumps, and
28
Which of the two we call the transform, and which we call the inverse transform, is a matter of
convention.

87
 ∞
(ii) f (x) is absolutely integrable in (−∞ , ∞), so that −∞
dx |f (x)| < ∞ (or f (x) ∈
L1 (−∞ , ∞)),

then its Fourier transform exists.

Note also the Fourier transform conventions used here:

— When integrating over k, the kernel is (2π)−1 e+ikx .

— When integrating over x, the kernel is e−ikx .

I repeat that these are just matters of convention,29 but it is quite important to choose
some specific convention and stick to it consistently, in order to avoid errors.

Parseval’s formula for Fourier transforms: If f (x) is not only integrable but is
also square-integrable, i.e., if f (x) ∈ L2 (−∞ , ∞), it follows on applying Parseval’s
formula that its Fourier transform f(k) is also an element of L2 (−∞ , ∞): we have
 ∞  ∞
1
dx |f (x)| =
2
dk |f(k)|2 .
−∞ 2π −∞

(This relation is just the continuum analog of the discrete version of Parseval’s formula
for Fourier series.) The connection between L2 functions and Fourier tranforms should
be familiar to you from elementary quantum mechanics. The Fourier transform, as
applied to elements of the function space L2 , may be regarded as a change of basis.
In the context of quantum mechanics, Parseval’s formula implies that, if the position-
space wave function of a particle is normalizable, then so is its momentum-space wave
function, because these two wave functions form a Fourier transform pair.

Fourier transform of the δ-function: The Dirichlet conditions and the conditions
stated above are sufficient conditions for a function f (x) to have a Fourier series ex-
pansion or a Fourier transform, as the case may be. They are by no means necessary
conditions. For instance, numerous functions that are integrable, but not absolutely
integrable (i.e., their absolute values are not integrable) have Fourier transforms. Func-
tions that are more singular than permitted by the Dirichlet conditions may also have
Fourier representations. The theory of generalized functions or distributions is, in fact,
very closely linked with the Fourier transforms of these objects. There is a highly-
developed area of mathematics, harmonic analysis, dealing with these matters and
their generalizations.

29
For instance, as you’ll see below, the kernels (2π)−1/2 eikx and (2π)−1/2 e−ikx are often used in
the expansion formula and the inversion formula, instead of the kernels (2π)−1 eikx and e−ikx that we
have used.

88
 The Dirac δ-function, which is so useful in applications, also has a Fourier transform.
Since we know that  ∞
1
δ (x) = dk eikx ,
2π −∞
it follows that the Fourier transform of δ (x) is just unity, i.e., δ (k) = 1.

1. Sketch the functions f (x) listed below, and show that their Fourier transforms are
as given. Here a and σ are positive constants, while b and µ are real constants. θ(x)
is the unit step function.
−4iabk
(a) f (x) = e−a|x| sin bx =⇒ f(k) = .
(a2 + b2 + k 2 )2 − 4b2 k 2
2a(a2 + b2 + k 2 )
(b) f (x) = e−a|x| cos bx =⇒ f(k) = .
(a2 + b2 + k 2 )2 − 4b2 k 2
[θ (x + a) − θ (x − a)] sin ka
(c) f (x) = =⇒ f(k) = .
2a ka
sin ax
(d) f (x) = =⇒ f(k) = π [θ (k + a) − θ (k − a)].
x
1
f(k) = e−iµk− 2 k σ .
2 /(2σ 2 ) 1 2 2
(e) f (x) = √ e−(x−µ) =⇒
2π σ 2

The functions in (c) and (e) above can be regarded as normalized probability density
functions corresponding to a random variable x. The first of these corresponds to
a uniform distribution, while the second corresponds, of course, to the normal or
Gaussian distribution. The Fourier transform of the probability density of a random
variable is called its characteristic function. Knowing the characteristic function of
a random variable is equivalent to knowing all its moments and cumulants.

2. Sketch the functions f(k) listed below, and show that their inverse Fourier trans-
forms f (x) are as given. (σ, k0 and λ are positive constants.)
1 1
(a) f(k) = e− 2 (k−k0 ) σ =⇒ f (x) = √
2 2 2 2
eik0 x e−x /(2σ ) .
2πσ 2

k i(1 − cos k0 x)
(b) f(k) = θ (k0 − |k|) =⇒ f (x) = .
|k| πx
1 − cos (k0 x) + k0 x sin (k0 x)
(c) f(k) = θ (k0 − |k|) (k0 − |k|) =⇒ f (x) = .
2πx2
λ
(d) f(k) = e−λ|k| =⇒ f (x) = .
π (x2 + λ2 )

89
The function f (x) in case (d) is a ‘Lorentzian’. It is is the probability distribution
function corresponding to the Cauchy distribution.

Relative ‘spreads’ of a Fourier transform pair: A general feature of great impor-

tance emerges from the examples above. Roughly speaking, when the function f (x) is
of compact support, i.e., it is nonzero in only a finite interval in x and is zero outside
it, or is mostly concentrated in a finite interval and rapidly decreases to zero outside
it, its transform f(k) is spread out in k. This is why the rectangular pulse considered
earlier, which is strictly zero for |x| > a, has a Fourier transform that decays to zero
relatively slowly, like |k|−1. The same statement is applicable with the roles of f(k)
and f (x) interchanged. The compact pulses represented by f(k) in the examples above
have transforms that decay like |x|−1 for large |x|. When one member of the transform
pair decays exponentially, which means it falls off faster than any inverse power of the
argument, the other member decays like the inverse square of the argument. This is
demonstrated by case of the Lorentzian. Finally, when one of the pair is a Gaussian
(with possible linear terms in the exponent), so is the other member of the pair. This
result is so useful that I’ll write it down again for ready reference:
1
f(k) = e−iµk− 2 k σ .
2 /(2σ 2 ) 1 2 2
f (x) = √ e−(x−µ) ⇐⇒
2π σ 2
The feature under discussion has a far-reaching physical consequence. This sort of
‘duality’—compactness in one variable, spreading in the ‘conjugate’ variable—is at the
very heart of the Heisenberg Uncertainty Principle. The fact that

≡ 0 necessitates
a quantum mechanical description of systems, in terms of state vectors (or wave func-
tions). In turn, the ‘fuzziness’ implicit in such a description is inevitably subject to
the duality mentioned above.

The convolution theorem for Fourier transforms is a basic and most useful rela-
tionship. Given two functions f (x) and g(x), the function
 ∞  ∞
  
h(x) = dx f (x ) g(x − x ) = dx  f (x − x  ) g(x )
−∞ −∞

is called the convolution of the functions f (x) and g(x). This relationship is sometimes
written as h = f ∗ g . (This should not be confused with ordinary multiplication of the
functions concerned.)

3. Establish the convolution theorem for Fourier transforms: If f (x) and g(x) have
Fourier transforms f(k) and 
g (k) respectively, then the Fourier transform of their con-
volution h(x) is given by h(k) = f(k) 
 g (k).

Thus, the Fourier transform of the convolution of two functions is just the product of
the Fourier transforms of the two functions. The Fourier transform operation converts

90
convolution to an ordinary multiplication.

4. It should be clear from the foregoing that this works in reverse, too. Consider the
Fourier transform of the convolution of f and g, namely, of the function of k given by
 ∞  ∞

H(k) = dk f(k ) 
  
g (k − k ) ≡ dk  f(k − k  ) 
g (k  ).
−∞ −∞


Check that the inverse Fourier transform of H(k) is simply the product H(x) =
f (x) g(x).

5. Generalization of Parseval’s formula: Let f (x) and F (x) be two good func-
tions30 of x, i.e., they have derivatives of all orders for all x, and vanish (along with
their derivatives of all orders) as x → ∞. Show that
 ∞  ∞
1
dx f (x) F (−x) = dk f(k) F(k) .
−∞ 2π −∞

Note that (i) neither side of this equation is in the form of a convolution of two func-
tions; (ii) there is a minus sign present in the argument of one of the functions in the
integrand; and (iii) each side of the equation is a number, and not a function. We’ll
use this result in the sequel, in the derivation of the Poisson summation formula.

Note that Parseval’s formula itself follows as a special case of the result above. All you
have to do is to identify F (−x) with f ∗ (x), and observe that

F (−x) = f ∗ (x) ⇐⇒ F(k) = f∗ (k).

Iterates of the Fourier transform operator: We’ve seen that the Fourier trans-
form f(k) of a function f (x) may be regarded as the result of applying a certain integral
operator, namely, the Fourier transform operator F , to the function. The kernel of
the operator is just e−ikx . In order to make the notion precise, we need to specify a
function space, such that both the function f (x) and its transform f(k) belong to the
same space. As we know, the space L2 (−∞ , ∞) satisfies this requirement. Let us
therefore restrict ourselves to this function space in what follows.

Given a function f ∈ L2 (−∞ , ∞), we have

 ∞

[F f ](x) ≡ f (x) = dy e−ixy f (y) .
−∞

30
I believe this simple but expressive terminology is due to M. J. Lighthill.

91
(Observe the notation in this equation! I have written x for the argument of the output
function after the Fourier transform operation is performed.) Something very interest-
ing happens when the operator F is iterated i.e., applied repeatedly to a function. We
have
 ∞

[F f ](x) ≡ [F F f ](x) = [F f ](x) =
2
dz e−ixz f(z)
−∞
 ∞  ∞  ∞  ∞
−ixz −izy
= dz e dy e f (y) = dy f (y) dz e−i(x+y)z
−∞
∞
−∞ −∞ −∞

= dy f (y) 2π δ (x + y) = 2π f (−x) .
−∞

Now, we know that the parity operator P changes the sign of the argument when
acting on any function of x. That is, (Pf )(x) ≡ f (−x) . The result above shows that
the square of the Fourier transform operator is just 2π times the parity operator. In
symbols,31
F 2 = 2π P.
Moreover, since the square of the parity operator is obviously just the unit operator I,
we have the operator relationship

F 4 = (2π)2 P 2 = (2π)2 I.

The Fourier transform operator is thus proportional to a ‘fourth root’ of the unit op-
erator! These results find applications in Fourier optics.

Eigenvalues and eigenfunctions of F : The fact that F 4 = (2π)2 I seems to suggest

that the eigenvalues of the Fourier transform operator in L2 are given by (2π)1/2 times
the fourth roots of unity, namely, the four numbers ±(2π)1/2 and ±i (2π)1/2 . Can we
find the eigenvalues and the corresponding eigenfunctions explicitly?

Recall that the Hamiltonian of the quantum mechanical linear harmonic oscillator
has exactly the same form in the position basis and in the momentum basis: In units
such that m, ω and
are all equal to unity, the Hamiltonian is represented by the
operators
1
d2  1
d2 
− 2 + x2 and − 2 + p2 ,
2 dx 2 dp
respectively, in the position basis and the momentum basis. As a result, the normalized
position-space wave functions Φn (x) and the momentum-space wave functions Φ  n (p)
representing the eigenstates of this Hamiltonian must be identical in form. But we also
know that Φn (x) and Φ  n (p) are elements of L2 (−∞ , ∞), and that they are Fourier
transform pairs. It follows that these are precisely the eigenfunctions of F that we’re
31
The factor of 2π is a consequence of the particular Fourier transform convention I have used. It
should be obvious that the factor can be made equal to unity by a suitable choice of convention.

92
looking for. It remains to prove this assertion directly, and also to find the eigenvalues
of F explicitly. Let’s now do so.

We want to show that the functions

1 −x2 /2
Φn (x) =  √ 1/2 e Hn (x) (n = 0, 1, . . .),
2n n! π

where Hn (x) is the Hermite polynomial of degree n, are eigenfunctions of F . The

n-dependent constant of proportionality is irrelevant in an eigenvalue equation, and so
2
let’s consider the function e−x /2 Hn (x). Now, the generating function for the Hermite
polynomials is given by
2
∞
tn
e2tx−t = Hn (x) .
n=0
n!
Such an equation must be regarded as an equation between two analytic functions of
the complex variable t. The left-hand side is an entire function of t. The series on the
right-hand side must therefore converge for all finite values of |t|.

Multiply both sides of the last equation by exp (−ikx − 12 x2 ), and integrate over x
from −∞ to ∞. The left-hand side becomes a Gaussian integral that can be evaluated.
The result is
∞  ∞
tn  % −x2 /2
∞
1/2 −k 2 /2 −2ikt+t2 −ikx −x2 /2
& tn
(2π) e e = dx e e Hn (x) = F e Hn (x) .
n=0 −∞
n! n=0 n!

But the second exponential factor on the left-hand side is again a generating function
for Hermite polynomials. We have

∞
(−it)n
−2ikt+t2 2k(−it)+(−it)2
e =e = Hn (k) .
n=0
n!

Equating the coefficients of tn of the two absolutely convergent power series in t, we

finally get
 ∞
2 % 2 &  2 
dx e−ikx e−x /2 Hn (x) = F e−x /2 Hn (x) = (2π)1/2 (−i)n e−k /2 Hn (k) .
−∞

2
Hence the function
√ e−x /2 Hn (x), or this function multiplied by the normalization con-
stant (2n n! π )−1/2 , is an eigenfunction of the Fourier transform operator F in the
space L2 (−∞ , ∞). Here n runs over the values 0, 1, 2, . . . . The corresponding eigen-
value is (2π)1/2 (−i)n .
• The eigenvalues of the operator (2π)−1/2 F are (−i)n , where n = 0, 1, 2, . . . .

93
6. Work through the steps of this derivation.

Unitarity of the Fourier transformation: Since (−i)n is always equal to 1, −i, −1

or i, the operator (2π)−1/2 F has just these four distinct eigenvalues. Each of these is
infinitely degenerate, i.e., there is a countable infinity of linearly independent eigen-
functions corresponding to each eigenvalue. These eigenfunctions are, respectively,

Φ4n (x), Φ4n+1 (x), Φ4n+2 (x) and Φ4n+3 (x) (n = 0, 1, 2, . . . ),

where Φn (x) is as given above.

As I have mentioned already, the factors of (2π)1/2 in the eigenvalues of F arises

from the Fourier transform convention we have adopted. Let me repeat this for clar-
ity: we have defined
 ∞  ∞
1
[F f ](k) = f(k) = dx e−ikx
f (x), [F f ](x) = f (x) =
−1
dk eikx f(k).
−∞ 2π −∞

Had we used the alternative (and more symmetrical) Fourier transform convention
 ∞  ∞
1 −1  1
[F f ](k) = √ dx e−ikx
f (x), [F f ](x) = √ dk eikx f(k),
2π −∞ 2π −∞

we would have found that F 4 = I, and the eigenvalues of F would have been simply ±1
and ±i. Indeed, the customary normalization of quantum mechanical wave functions
uses precisely this convention. But the convention chosen here is used in many other
applications of Fourier analysis, and so let’s stay with it.

Taking the Fourier transform of a function involves applying an integral operator

to the function. What is the adjoint of this operator? Let us digress for a moment to
find the adjoint of a general integral operator.

The adjoint of an integral operator: Let f (x) be the function representing the
vector | f  in the given function space (here, L2 (−∞ , ∞)). If
∞ K is the integral operator
concerned, the vector K | f  is represented by the function −∞ dy K(x, y) f (y), where
K(x, y) is the kernel of the operator. The adjoint of K, denoted by K† , is of course to
be identified by applying the condition g | Kf  = K† g | f  for every pair of elements

94
| f , | g  in the function space. But
 ∞  ∞
∗
g | Kf  = dx g (x) dy K(x, y) f (y)
−∞ −∞
 ∞  ∞
= dy dx g ∗(x) K(x, y) f (y) (changing the order of integration)
−∞∞ −∞∞
= dx dy K(y, x) g ∗(y) f (x) (re-labeling x ↔ y)
−∞∞ −∞∞
= dx dy [K ∗ (y, x) g(y)]∗ f (x) = K† g | f .
−∞ −∞

It follows that the kernel of the integral operator K† is K ∗ (y, x).32

Returning to the case at hand, we see that (2π)−1/2 F has the kernel K(k, x) =
(2π)−1/2 e−ikx , while its inverse [(2π)−1/2 F ]−1 has the kernel (2π)−1/2 eikx . But the
latter quantity is precisely K ∗ (x, k). We therefore have the operator identity
[(2π)−1/2 F ]−1 = [(2π)−1/2 F ]† .
In other words, the operator (2π)−1/2 F is a unitary operator. It is not surprising,
then, that all its eigenvalues lie on the unit circle in the complex plane, just as those
of a unitary matrix do!

Hence the Fourier transform in L2 (−∞ , ∞) is not only a change of basis in the
space, but also a unitary transformation. This fact has implications in quantum
mechanics. For instance, it guarantees that the description of a particle (or a system
of particles) in position space and in smomentum space are unitarily equivalent—i.e.,
you can use either description without altering the underlying physics.

The Fourier transform in d dimensions: The Fourier transform and most of the
results in the foregoing are generalized in a straightforward manner to functions of
r ∈ Rd , where d = 2, 3, . . . . We have
 
1
f (r) = d k e f (k) ⇐⇒ f(k) = dd r e−ik·r f (r).
d ik·r 
(2π)d
Here dd k and dd r denote the volume elements in d-dimensional k-space and r-space,
respectively. The Fourier representation of the d-dimensional δ-function is

1
(d)
δ (r) = dd k ei k·r .
(2π)d
32
Observe how reminiscent this is of the situation in the case of matrices. If the (ij)th element of
a matrix A is aij , the (ij)th element of its hermitian adjoint A† is a∗ji . The close analogy between
integral operators and matrices (as operators in an LVS) plays a crucial role in the theory of integral
equations.

95
 Fourier expansions of vector-valued functions can also be written down in an anal-
ogous manner. Thus, if u(r) is a vector field in Rd , we have
 
1
u(r) = d
d ke uik·r
 (k) ⇐⇒ u (k) = dd r e−ik·r u(r) .
(2π)d
One of the great advantages of the Fourier expansion of functions now becomes evident.
Consider the usual case, d = 3. A vector field is generally specified by differential
equations for its divergence and curl, respectively.33 We have

1
∇ · u(r) = 3
 (k)
d3 k eik·r ik · u
(2π)
and 
1
∇ × u(r) = (k).
d3 k eik·r ik × u
(2π)3
' (
Further, the set of functions eik·r | k ∈ R3 forms an orthonormal basis for integrable
functions of r ∈ R3 . Hence partial differential equations for the divergence and curl of
a vector field u(r) reduce to algebraic equations for its Fourier transform u  (k). The
utility of this result should now be obvious to you.

The Poisson summation formula is a very useful result that has many applications.
For instance, it can be used to sum many infinite series. It is also related to much deeper
mathematical results (which we do not go into here).
Consider the infinite periodic array of δ-functions (a Dirac comb) given by

∞
F (x) = δ (x − nL),
n=−∞

where L is a positive constant. It is trivial to verify that F (x) = F (−x). If we

regard F (x) as a periodic function of x with fundamental interval (− 12 L, , 12 L), we
have F (x) = δ(x) in the fundamental interval. It can then be expanded in a Fourier
series according to
 1L
1 
∞
2
F (x) = Fn e2πnix/L
, where Fn = e−2πnix/L δ (x) = 1.
L n=−∞ − 12 L

Therefore F (x) can be written in two different ways—either as an infinite sum of

exponentials, or as an infinite sum of δ-functions:

∞
1  2πnix/L
∞
F (x) = δ (x − nL) = e .
n=−∞
L n=−∞
33
Maxwell’s equations for electromagnetic fields provide a prominent example of this statement.

96
But
1
x
δ n−
δ (x − nL) = δ (nL − x) = .
L L
We thus have a useful identity relating a sum of equally-spaced δ-functions to a sum
of exponentials, namely,

 
 x 
∞ ∞
2πnix
δ n− = exp .
n=−∞
L n=−∞
L

It is worth pausing once again for a moment to think about this most remarkable for-
mula. Note that neither side of this equation is anything like an absolutely convergent
series!

Next, consider the Fourier transform of F (x). This is

 ∞
F(k) = dx e−ikx F (x)
−∞
∞ 
 ∞ 
∞
−ikx
= dx e δ (x − nL) = einkL .
n=−∞ −∞ n=−∞

But, as we have seen already, an infinite sum of exponentials of this kind can be written
as an infinite sum of δ-functions. Using this relationship, we get
∞  
 kL
F (k) = δ n− .
n=−∞
2π

The final step is to insert the expressions for F (x) and F(k) in the identity34
 ∞  ∞
1
dx f (x) F (−x) = dk f(k) F(k).
−∞ 2π −∞

Using the fact that the function F (x) in the present instance is an even function of x,
we get
 ∞ ∞  ∞ ∞  
1  kL
dx f (x) δ (x − nL) = dk f (k) δ n− ,
−∞ n=−∞
2π −∞ n=−∞
2π

where f (x) is any arbitrary function with a Fourier transform f(k). Finally, therefore,

∞ 
∞
f (nL) = (1/L) f(2πn/L) .
n=−∞ n=−∞

34
The identity was derived for good functions, but we’re using it here for singular functions like the
δ-function. I’ll merely state here without proof that this step can be justified.

97
Here L is an arbitrary positive parameter. This is the famous Poisson summation
formula It is helpful in the summation of series in cases when the left-hand side is
difficult to evaluate, but the right-hand side is more tractable; or vice versa. Note
that the parameter L occurs in the numerator of the argument of the function on the
left-hand side, but on the right-hand side it occurs in the denominator. Hence one of
the two representations may be useful for small values of L, while the other is useful
for large values of L, in physical applications.

Some illustrative examples: Here are some instances of the use of Poisson’s sum-
mation formula to derive useful identities and to sum certain infinite series.

7. If f (x) is chosen to be a Gaussian function, the formula immediately gives


∞
1 
∞
exp (−π n2 λ2 ) = exp (−π n2 /λ2 ) , (λ > 0).
n=−∞
λ n=−∞

Establish this result using the Poisson summation formula.

This identity, due to Jacobi, is so useful and important that it is sometimes called the
Poisson summation formula itself!35 For example, in the context of the phenomenon of
diffusion, the parameter λ2 is proportional to Dt, where D is the diffusion coefficient
and t is the time. The Poisson summation formula can then be used to obtain valuable
insight into the behavior of the solution to the diffusion equation for both small t and
large t.

Another useful case is that of the function

cos bx
f (x) = 2 (a > 0, 0 ≤ b < 2π).
x + a2
The Fourier transform of this function is easily found by contour integration. You can,
however, write down f(k) using the information that is already available. Recall that
the inverse Fourier transform of e−λ|k| is a Lorentzian. That is,
λ
f (x) = ⇐⇒ f(k) = e−λ |k| .
π(x2 + λ2 )
Hence the Fourier transform of the function f (x) is given by
 ∞  ∞ % −i(k−b)x −i(k+b)x
&
cos bx 1 e + e
f(k) = dx 2 e−ikx
= dx
−∞ x + a2 2 −∞ x2 + a2
π % −a |k−b| &
= e + e−a |k+b| .
2a
35
It is an example of a Gaussian sum, i.e., a sum over the exponentials of the squares of the integers.
A vast literature exists on this subject, that has ramifications in several parts of mathematics and
mathematical physics.

98
Setting L = 1 in the Poisson summation formula, we therefore have
∞
cos (nb) ∞
π % −a |2πn−b| &
−a |2πn+b|
2 + a2
= e + e .
n=−∞
n n=−∞
2a

Some simplification then yields the identity


∞  
cos (nb) π cosh a(π − b) 1
= − .
n=1
n2 + a2 2a sinh (πa) πa

Setting b = 0 in this relation, we recover an important and useful result that we have
already derived using contour integration, namely,

∞  
1 π 1
S(a) ≡ 2 2
= coth πa − .
n=1
n +a 2a πa

As usual, a relation of this sort may be regarded as a relation between two analytic
functions—in this case, of the variable a. Recall that the quantity in curly brackets
on the right-hand side is the Langevin function. It occurs, for instance, in the el-
ementary
∞ theory of paramagnetism. Passing to the limit a → 0 once again gives us
n=1 1/n 2
≡ ζ(2) = π 2 /6. Repeatedly differentiating S(a) with respect to a and pass-
ing to the limit a → 0 enables us to find ζ(4) = π 4 /90, ζ(6) = π 6 /945, . . . .

Generalization to higher dimensions: The Poisson summation formula is readily

generalized to higher dimensions. If f(k) is the Fourier transform of f (r) where r ∈ Rd ,
then  
f (nL) = (1/Ld ) f(2πn/L) .
n∈Zd n∈Zd

Here Zd is the set of d-tuples of integers, in an obvious notation. Further generaliza-

tions are also possible, such as the counterpart of the formula above for f (nL + r0 ),
where r0 is any given vector in Rd .

Finally, it turns out that the generalizations of the Poisson summation formula
are related to deep results in advanced mathematics. I merely mention some of these
results, to whet your curiosity: the asymptotic behavior of the heat kernel, the spectra
of the Laplacian operator and general elliptic operators on manifolds, Eisenstein series,
automorphic functions, the Selberg trace formula, and so on.

99
9 QUIZ 1
1. Are the statements in quotation marks true or false?

(a) “Every derivative of an analytic function of a complex variable is also an

analytic function.”

(b) Let u and v denote the real and imaginary parts of an analytic function of
z = x + iy.
“The curves u(x, y) = constant and v(x, y) = constant intersect each other
at right angles.”

(c) “An entire function must necessarily be singular at z = ∞, unless it is just

a constant.”

(d) “A meromorphic function cannot have an essential singularity at the point

at infinity.”

∞
(e) “The radius of convergence of the power series n=1 n1/n z n is zero.”

(f) “The function sin (π/z) has an accumulation point of poles at z = 0.”

(g) “The relation Γ(z) Γ(1 − z) = π cosec πz is only valid in the region 0 <
Re z < 1.”

(h) Let a be a positive constant.

∞
“f (z) = a dt tz e−t is an entire function of z.”


(i) “The power series ∞ n 4
n=1 z /n is absolutely convergent at all points inside
and on the unit circle |z| = 1.”


∞
(j) “The series (n + 1)z−1 converges in the region Re z > 0 .”
n=0
√ √
(k) Consider the Möbius transfomation z → w = (2z + 3 )/( 3 z + 2).
“This is a hyperbolic Möbius transformation.”

√ √
(l) Consider z → w = (2z + 3 )/( 3 z + 2) once again.
“The transformation maps the circle |z| = 1 to the circle |w| = 1.”

100
 (m) Legendre’s differential equation is (1 − z 2 )φ − 2zφ + ν(ν + 1)φ = 0.
“Since this equation is invariant under the interchange ν ↔ −ν − 1, all its
solutions must also be invariant under this interchange.”

3/2
(n) “The function f (t) = et , where t ≥ 0, has no Laplace transform.”

∞ ∞
(o) “If [Lf ](s) = 0
dt e−st f (t), then [L2 f ](s) = 0
dt f (t)/(s + t).”

(p) “The product Γ(z) ζ(z) tends to a finite, nonzero limit as z → −2n, where
n = 1, 2, . . . .”

 its Laplace
(q) Let φ(t) be a linear, causal, retarded response function, and φ(s)
transform.
“The corresponding dynamic susceptibility χ(ω) is the analytic continuation
 to s = −iω.”
of φ(s)

(r) “The logarithmic derivative of the Riemann zeta function, ζ  (z)/ζ(z), has a
simple pole at z = 1 with residue equal to −1.”

(s) “The only pole of the logarithmic derivative of the Riemann zeta function
is at z = 1.”

(t) Consider the function space L2 (−∞ , ∞).

“Any eigenfunction of the Fourier transform operator is also an eigenfunc-
tion of the parity operator, but the converse is not necessarily true.”

2. Fill in the blanks in the following.

(a) The real part of an entire function f (z) is given by u(x, y) = (cosh x) (cos y).
Hence the function is f (z) = · · · .

(b) coth z is a periodic function of z, with a period equal to · · · .

(c) The singularity of the polynomial p(z) = z n + a1 z n−1 + · · · + an at z = ∞ is

· · · . (Select one from the following: (i) a removable singularity (ii) a simple
pole (iii) a pole of order n (iv) an essential singularity.)

101
(d) The residue at z = ∞ of the polynomial p(z) = z n + a1 z n−1 + · · · + an is
Res p(z) = · · · .
z=∞

(e) Let
 C denote the circle |z| = 2 traversed once in the positive sense. Then
C
dz/(z 4
− 1) = ···.

(f) Let a and b be two different complex numbers, each with nonzero real and
imaginary parts. The radius of convergence of the power series

∞
Γ(n + a) Γ(b) z n
f (z) =
n=0
Γ(a) Γ(n + b) n!

is R = · · · .


∞ 
∞
(g) Given that 1/n4 = π 4 /90, it follows that 1/(2n + 1)4 = · · · .
n=1 n=0
∞
 ∞ that 0 dx (sin
(h) Given kx)/x = 12 π (where k > 0), the value of the integral
0
dx (1 − cos x)/x = · · · . (Hint: Integrate k over a suitable range.)
2

(i) The numerical value of the product

               
Γ − 54 Γ − 34 Γ − 14 Γ 14 Γ 34 Γ 54 Γ 74 Γ 94 = · · · .

1
(j) The value of the integral 0
dt t−1/2 (1 − t)−1/2 = · · · .

(k) Let an arbitrary initial point z (0) in the complex plane be mapped to the
point z (n) under n iterations of the Möbius transfomation
√ √
z → (2z + 3 )/( 3 z + 2).

As n → ∞, z (n) → · · · for all z (0) , with one exception.

(l) Previous question continued: The exceptional point that does not tend to
the limit point above is z = · · · .

(m) Under the Möbius transformation z → w = (2z + 3)/(z + 2), the circle
|z + 2| = 1 is mapped to the circle · · · .

102
 √
(n) The function f (z) = z ln [(z − 1)/(z + 1)] has branch points at z = · · · .

(o) Let α and β be arbitrary complex numbers. The function

f (z) = (z 2 − 1)α /(z 2 + 1)β

has branch points at z = · · · .

(p) The residue of f (z) = exp (z + z −1 ) at z = 0 is · · · . (Express your answer

in terms of a modified Bessel function.)

(q) Given that the Laplace transform of sin t is 1/(s2 + 1), it follows that the
Laplace transform of sinh t is · · · .

(r) The generating function for the Hermite polynomial Hn (z) is


∞
tn
2tz−t2
e = Hn (z) .
n=0
n!

It follows that the Rodrigues formula for Hn (z) is Hn (z) = · · · .

(s) Let 
1, |x| ≤ 1
f (x) =
0 |x| > 1.
If f(k) denotes the Fourier transform of f (x), the value of the integral
∞
−∞
dk |f(k)|2 = · · · .

(t) Consider a random walk on an infinite linear lattice whose sites are labelled
by the integers. The walker jumps from any site j to j −1 with a probability
per unit time given by λq, and from j to j + 1 with a probability per unit
time given by λp; further, the walker stays at the site j with probability
per unit time given by λr. Here p, q and r are positive constants satisfying
p + q + r = 1, and λ is a positive constant with the physical dimensions of
(time)−1 . Let P (j, t) be the probability that the walker is at the site j at
time t. The differential equation satisfied by P (j, t) is dP (j, t)/dt = · · · .

103
 Quiz 1: Solutions
1. (a) T
(b) T
(c) T
(d) F
(e) F
(f) F
(g) F
(h) T
(i) T
(j) F
(k) T
(l) T
(m) F
(n) T
(o) T
(p) T
(q) T
(r) T
(s) F
(t) T

2. (a) f (z) = cosh z

(b) iπ
(c) a pole of order n
(d) 0
(e) 0
(f) ∞
(g) π 4 /96
1
(h) 2
π
4
(i) 4π
(j) π

104
(k) 1
(l) −1
(m) |w − 2| = 1
(n) −1, 0, 1 and ∞
(o) 1, i, −1, −i and ∞
(p) I1 (2)
(q) 1/(s2 − 1)
 dn  n
2tz−t2 n z2 d 2
(r) Hn (z) = n
e , which simplifies to H n (z) = (−1) e n
e−z .
dt t=0 dz
(s) 4π
dP (j, t)  
(t) = λ p P (j − 1, t) + q P (j + 1, t) − (p + q) P (j, t)
dt

105
10 The fundamental Green function for ∇2
The Laplacian operator ∇2 appears in all the standard and important linear partial dif-
ferential equations of elementary mathematical physics: Laplace’s equation, Poisson’s
equation, Helmholtz’s equation, the diffusion or heat equation, and the wave equation.
We’ll now consider Poisson’s equation, which requires the determination of the Green
function of the Laplacian operator. Let’s start with a brief recapitulation of the essen-
tials of the Green function method.

Green functions: Consider an inhomogeneous, linear, ordinary differential equation

of the form
Dx f (x) = g(x)
in some interval (a, b) in the real variable x. Here Dx is a differential operator involving
derivatives of various orders with respect to x, and g(x) is a given function. It is
required to find f (x). Let’s write the equation in abstract form in terms of the elements
| f  and | g  of a linear space that are represented by f (x) and g(x) in a suitable
function space. Recall that x | f  ≡ f (x) and x | g  ≡ g(x). Let D denote the
abstract operator that is represented by Dx in function space. Then

D | f  = | g .

The formal, general solution of this equation is given by


| f  = D−1 | g  + ci |hi ,
i

where the ci are constants, and the kets |hi  are the linearly independent solutions of
the homogeneous equation

D |hi  = 0 or Dx hi (x) = 0.

In function space, the inverse of the operator D is represented by the inverse of the
differential operator Dx . In general, the latter is an integral operator. To see this,
take the scalar product of both sides of the formal solution for | f  above with  x |.
We have

x | f  ≡ f (x) =  x |D−1 | g  + ci  x |hi 
i
 b 
= dx   x |D−1| x  x  | g  + ci hi (x)
a i
 b 
= dx  G(x, x  ) g(x ) + ci hi (x),
a i

106
where
G(x, x  ) ≡  x |D−1|x  .
G(x, x  ) is the Green function of the differential operator Dx . It is just the ‘matrix
element’ of the operator D−1 between the states  x | and | x  , and is called the kernel
of the integral operator D−1 .

We learn in elementary treatments of differential equations that the general solu-

tion of an inhomogeneous differential equation is made up of two parts: a particular
integral (PI) that depends on g(x), and a complementary function (CF) that does
not. The first term on the right-hand side in the solution above is the PI, while the
second is the CF. The right combination of the two is determined by fixing the values
of the constants ci using the boundary conditions.

The fact that G(x, x  ) represents (in function space) the inverse of the abstract
operator D−1 means that it, too, satisfies the same differential equation as f (x), but
with a δ-function as the inhomogeneous term. To see this, we start with

D D−1 = I =⇒  x |D D−1| x   =  x |I| x   = δ(x − x  ).

But
 b
−1 
 x |D D | x  =  x |D dy | y  y | D−1 | x  
a
 b 
= dy  x |D| y G(y, x ) = dy Dx x| y G(y, x )


a
= dy Dx δ(x − y)G(y, x ) = Dx G(x, x  ).

Therefore the Green function satisfies the differential equation

Dx G(x, x  ) = δ(x − x  ).

In order to write down the solution for f (x) explicitly, we need to find the Green
function G(x, x  ). This must be done by solving the differential equation for G, using
appropriate boundary conditions and adjusting the values of the constants ci . In this
manner, we arrive at the unique solution that satisfies the given boundary conditions
on f (x).

1. Here’s a simple example that shows how the Green function method works. Con-
sider the ordinary differential equation

d2
f (x) = g(x),
dx2

107
where x ∈ [0, 1], and g(x) is a given function of x. It is required to find the solution
f (x) that satisfies the general linear boundary conditions

f (0) + af  (0) = b and f (1) + cf  (1) = d,

where a, b, c and d are given constants. Show that


 A1 x + A2 for 0 ≤ x < x 
G(x, x ) =
A3 x + A4 for x  < x ≤ 1,
where

x − 1 − b − c + d −ax  + (a + b)(1 + c) − ad 
A1 = , A2 = 

1−a+c 1−a+c 

x − a − b + d −(1 + c)x  + (a + b)(1 + c) − ad 


A3 = , A4 = . 
1−a+c 1−a+c
Hence write down the solution for f (x).

In essence, this is the Green function method. All that has been said above can be
generalized to linear ordinary differential equations in a complex variable z, and fur-
ther, to the case of linear partial differential equations in several variables—e.g., when x
is replaced by r in any number d of spatial dimensions, or by (r, t) in (d+1) dimensions.

Poisson’s equation; the fundamental Green function for ∇2 : Let’s turn now to
a problem of great physical importance: the determination of the Green function of the
Laplacian operator, required for the solution of Poisson’s equation. In effect, we seek
the inverse of the ∇2 operator. To start with, consider the standard three-dimensional
case. Subsequently, we’ll go on to the cases d = 2 and d ≥ 4.

As you know, the electrostatic potential φ(r) in the presence of a given static charge
density ρ(r) satisfies Poisson’s equation, ∇2 φ(r) = −ρ(r)/0 . The solution satisfying
natural boundary conditions,
 3  namely, φ(r) = 0 as r → ∞ along any direction in space,
−1
is φ(r) = (4π0 ) d r ρ(r )/|r − r  |. This is just Coulomb’s Law together with the


superposition principle. This solution can be derived systematically, as follows.

Consider the general form of Poisson’s equation together with natural boundary
conditions, namely,

∇2 f (r) = g(r), with f (r) → 0 as r → ∞.

g(r) is a given function that acts as the ‘source’ term for the scalar field f (r). We’re
interested here in the PI, given by

f (r) = d3 r  G(r, r ) g(r ),

108
where the Green function G satisfies the differential equation

∇2r G(r, r ) = δ (3) (r − r  ).

I’ve used a subscript on the gradient operator in order to indicate the variable with
respect to which the differentiation is to be performed. We require the solution that
vanishes as r → ∞. Note that G may be interpreted (apart from a constant of
proportionality) as the Coulomb potential due to a point charge at the source point
r  . Observe that
(i) the operator ∇2 is translationally invariant—shifting r by r  does not change the
operator;

(ii) the δ-function on the right-hand side is a function of the difference r − r  ; and
finally,

(iii) the boundary condition imposed involves r → ∞, which is the same as |r − r  | →

∞.
Together, these facts ensure that the solution G(r, r ) is a function of the difference
r − r  alone. Let’s therefore set
R = r − r ,
and write the Green function as G(R). Since R is just a shift of the variable r, we
have ∇2r = ∇2R . Then
∇2R G(R) = δ (3) (R).
Now define the Fourier transform pair
 
1  
G(R) = 3
d3 k eik·R G(k) ⇐⇒ G(k) = d3 R e−ik·R G(R).
(2π)
The Fourier representation of the δ-function is

(3) 1
δ (R) = d3 k eik·R .
(2π)3

Using the fact that ∇2R (eik·R ) = −k 2 (eik·R ), we get


G(k) = −1/k 2 .

Inverting the Fourier transform,


1 eik·R
G(R) = − d3 k .
(2π)3 k2
2. Use spherical polar coordinates to evaluate this integral. Since the integral is a
scalar, it is rotationally invariant. You may choose the polar axis in k-space along the

109
 ∞
vector R. You will need the Dirichlet integral 0 dk (sin kR)/k = 12 π sgn R = 12 π. The
final result is
1 1
G(r − r  ) = − =− .
4πR 4π|r − r  |
Note that the Green function is actually a function of R alone. The PI in the solution
of Poisson’s equation is then

1 g(r )
f (r) = − d3 r  .
4π |r − r  |

3. Solution for a spherically symmetric source: When the source function

g(r) = g(r), i.e.,when it is spherically symmetric, the solution above can be simplified
further. For this purpose you need the expansion of the Coulomb kernel 1/|r − r |
in spherical harmonics. Let r = (r, θ, ϕ) and r  = (r  , θ  , ϕ  ), in spherical polar
coordinates. Then
 l 
1 
∞ l
1 1 rs ∗
= Ylm (θ, ϕ) Ylm (θ  , ϕ  ),
|r − r | rg l=0 (2l + 1) rg m=−l

where rs = min (r, r  ) and rg = max (r, r  ). Insert this expansion in the solution for
f (r), and and interchange the order of summation and integration. The integrations
over the angular variables θ  and ϕ  can then be carried out immediately by using the
fact that  √
dΩ  Ylm
∗
(θ  , ϕ  ) = 4π δl,0 δm,0 .

Because of the Kronecker deltas, the dependence of f (r) on the angles θ and ϕ also
disappears, so that f (r) = f (r). Show that the solution simplifies to
 ∞ 2   ∞
 r g(r  ) 1 r  2 
f (r) = − dr =− dr r g(r ) − dr  r  g(r  ).
0 r g r 0 r

4. Check that the solution above satisfies Poisson’s equation, ∇2 f = g. You will need
the Laplacian in spherical polar coordinates, and also the formula for differentiation
under the integral sign.

The Coulomb potential in d dimensions: The fundamental Green function of

the ∇2 operator in 3-dimensional Euclidean space, −1/(4π|r − r  |), is essentially the
Coulomb potential due to a point charge. This connection lends a deeper significance
to the Coulomb potential, and hence to the inverse-square central force. As you know,
the inverse-square force law leads directly to Gauss’ Law in electrostatics: The flux of
the electrostatic field over a closed surface is equal to the total charge enclosed by

110
the surface, apart from a multiplicative constant (= 1/0 , in SI units). A counterpart
obviously exists in gravitation as well. The inverse-square central force is the only
force law in three-dimensional space for which such a property holds good. Interest-
ingly enough, in spaces of other dimensionalities (d = 2, 4, . . .) too, there exist force
laws with an analogous property. They arise from the counterparts of the Coulomb
potential in those spaces. The case d = 2 is somewhat exceptional, as you will see, and
will be dealt with after we discuss the case d ≥ 3.

What is the analog of the Coulomb potential (or the inverse-square central force) in
a space of an arbitrary number of dimensions? The connection that I have just pointed
out leads to the consistent way to define such a potential, via Poisson’s equation for
the potential due to a point source (or charge):

• The Coulomb potential in d-dimensional space is (apart from a constant of pro-

portionality) the fundamental Green function of the Laplacian operator in d
dimensions.

We therefore look for the fundamental solution of the equation

∇2R G(d) (R) = δ (d) (R), with G(d) (R) −→ 0.

R→∞

The superscript in G(d) is to remind us that we are concerned with the Green function
in d-dimensional space. We’ll find the solution in two different ways: first, by using
Fourier transforms as in the d = 3 case derived above, and evaluating the resulting
integral the ‘hard’ way; second, by using a much simpler argument that will get us to
the same answer. This exercise is worth it because it is instructive.

Fourier-transforming the equation for G(d) (R) gives G(d) (k) = −1/k 2 , as before.
Remember that we are now working in d dimensions. Therefore
 d
1 d k ik·R
G (R) = −
(d)
d
e .
(2π) k2

That is, the Coulomb potential in d dimensions, i.e., the fundamental Green function
of the Laplacian, is essentially the inverse Fourier transform of −1/k 2 .

5. The next task is to evaluate the d-dimensional above for G(d) (R). Once again,
This is most conveniently done in terms of the analog of spherical polar coordinates
in d dimensions, in k-space. These ultraspherical coordinates consist of (i) the
magnitude |k| = k, (ii) a set of (d − 2) ‘polar’ angles θ1 , θ2 , . . . , θd−2 , and (iii) an
‘azimuthal’ angle ϕ. The ranges of these variables are given by

0 ≤ k < ∞, 0 ≤ θj ≤ π (1 ≤ j ≤ d − 2), and 0 ≤ ϕ < 2π.

111
The Cartesian components k1 , k2 , . . . , kd are related to the polar coordinates by

k1 = k cos θ1
k2 = k sin θ1 cos θ2
k3 = k sin θ1 sin θ2 cos θ3
··· = ························
kd−1 = k sin θ1 sin θ2 . . . sin θd−2 cos ϕ
kd = k sin θ1 sin θ2 . . . sin θd−2 sin ϕ.

The volume element dd k is then given by

dd k = dk1 dk2 · · · dkd

= k d−1 (sind−2 θ1 ) (sind−3 θ2 ) . . . (sin θd−2 ) dk dθ1 . . . dθd−2 dϕ.

We can choose the orientation of the axes such that R is along the direction of the
first coordinate k1 , so that k · R = kR cos θ1 . The integration over ϕ gives a factor
of 2π. Carry out the integration over each of the ‘polar’ angles θ2 , . . . , θd−2 using the
integral √  
 π
π Γ 12 (r + 1)
r
dθ sin θ =   .
0 Γ 1 + 12 r
But the integral over the last angular coordinate, θ1 , is more complicated: It turns out
to be essentially a representation of the Bessel functionof the first kind. The formula
you require is
 1 ν  π
z
Jν (z) = √ 2  dθ e±iz cos θ sin2ν θ, Re ν > − 12 .
π Γ ν + 12 0

Applying this formula,

 π
√  
dθ1 eikR cos θ1 sind−2 θ1 = π Γ 12 (d − 1) (2/kR)(d−2)/2 J d −1 (kR).
2
0

Insert these results and simplify the expression obtained, to arrive at the following
expression:  ∞
1
G (R) = −
(d)
d/2 (d−2)/2
dk k (d−4)/2 J d −1 (kR).
(2π) R 0
2

Thus G(d) (R) has been reduced to a single integral. It is obvious from the last equation
that the Green function is only a function of the magnitude R of the vector R. Let’s
therefore write it as G(d) (R) from now on.

A divergence problem: A difficulty crops up now. The integral in the last equation
above does not converge for arbitrary values of the dimensionality
∞ d. As you have
seen, the case d = 3 already involves the Dirichlet integral 0 dk (sin kR)/k, which

112
is convergent, but not absolutely convergent. It is finite only because the sine func-
tion oscillates in sign, and there is some cancellation between positive and negative
contributions
∞ to the integral. If the integrand is replaced by its magnitude, we find
that 0 dk | sin kR|/k diverges logarithmically owing to the slow (∼ k −1 ) decay of the
integrand as k → ∞. For d > 3, you may expect the divergence of the relevant integral
to get worse, essentially because the volume element dd k in d dimensions involves a
factor k d−1 . This factor increases rapidly with k for larger values of d.

This sort of divergence, arising from the behavior of the integrand as k → ∞, is

an example of what is known as an ultraviolet divergence in physics, especially
in the context of quantum field theory. The way to deal with such divergences in
physical problems, and to extract meaningful results for physical quantities, is called
regularization. There are many possible regularization methods. Here, I choose one
that relies on the idea of analytic continuation, because you are already familiar with
analytic continuation. It is called dimensional regularization. This is a powerful
method of regularization, and is used quite commonly in modern quantum field theory.

Dimensional regularization: Consider, first, the conditions under which the integral
 ∞
dk k (d−4)/2 J d −1 (kR)
2
0

actually converges. The argument is called power-counting and is quite simple in

this instance. We must look at the behavior of the integrand at both the end-points of
integration, and make sure that it is not too singular to be integrable at these points.
Consider the upper limit first. As you know,
 ∞
dk k r < ∞ provided r < −1, or, more generally, Re r < −1.

Now, as kR → ∞, the leading asymptotic behavior of the Bessel function Jν (kR) is

given by
Jν (kR) ∼ (kR)−1/2 ,
independent of the order ν of the Bessel function.36 The integrand therefore behaves
like k (d−5)/2 for large k. Hence the integral converges at the upper limit of integration
only if
1
2
(Re d − 4) − 12 < −1, that is, if Re d < 3.
This is precisely as expected: I have already pointed out that d = 3 is a marginal case
in which the integral barely manages to converge to a finite value.

36
The actual asymptotic form is [2/(π kR)]1/2 times a cosine function whose magnitude does not
exceed unity.

113
 At the lower limit of integration, the situation is different. We know that

dk k r < ∞ provided Re r > −1.
0

The leading behavior of the Bessel function Jν (kR) as kR → 0 is given by

Jν (kR) ∼ (kR)ν .

The integrand therefore behaves like k d−3 in the neighborhood of k = 0, and the
integral converges if
1
2
(Re d − 4) + 12 Re d − 1 > −1, that is, if Re d > 2.

For Re d ≤ 2, there is an infrared divergence. The terms ‘ultraviolet divergence’ and

‘infrared divergence’ originate in quantum field theory. The variable k is associated
with momentum (recall that it is the Fourier conjugate of a position variable). Large
k implies a small de Broglie wavelength (‘ultraviolet’), while small k corresponds to a
large de Broglie wavelength (‘infrared’).

Thus, the integral representing G(d) (R) converges in the range 2 < Re d < 3 of
the parameter d. For other values of d, the integral is infinite. Let’s now see how
dimensional regularization helps us extract sensible results from the formally divergent
integral we are faced with. The underlying idea is as follows.

(i) The dimension d itself is treated as a complex variable. In the region of the
complex d-plane in which the integral converges, it defines an analytic function
of d.

(ii) Analytic continuation is then used to go outside this region in the d-plane.

(iii) There are likely to be singularities (specifically, poles) present at the boundaries
of the original region of convergence. If such a singularity occurs at a physical
value of d, the regular part of the function (i.e., the function with the singular part
subtracted out) is supposed to represent the value of the function at the point
concerned.

Let’s now see how this works in the case at hand.

6. First of all, keeping d in the region 2 < Re d < 3, we must evaluate the integral
in the formula for G(d) (R). It turns out to be expressible in closed form, on using the
following formula:
 ∞  
2µ Γ 12 (ν + 1 + µ)  
dx x Jν (ax) = µ+1  1
µ
 , Re µ < − 12 , Re (µ + ν) > −1
0 a Γ 2 (ν + 1 − µ)

114
where a is a positive constant. First check that, when applied to the integral at
hand, the conditions on the parameters µ and ν amount to precisely the restriction
2 < Re d < 3. Next, use the formula above to show that
 
Γ 12 d − 1
G (R) = − d/2 d−2 .
(d)
4π R
But this expression for G(d) (R) is now in an explicit closed form that is analytic even
to the right of Re d = 3. Setting d = 3, it trivially checked that G(3) (R) = −1/(4πR),
as we have found already. You can now proceed to set d = 4, 5, . . . in the final result
to write down the fundamental Green function of the Laplacian in these dimensions.
We conclude that:

• the fundamental Green function of the Laplacian (i.e., the Coulomb potential)
in d ≥ 3 spatial dimensions is proportional to 1/Rd−2 .

7. In this simple example, there is actually no singular part in the region Re d ≥ 3

that has to be extracted and discarded. The ultraviolet divergence in this problem is
not a serious one. The infrared divergence, however, does exist. Recall that the leading
(right-most) pole of the gamma function Γ(z) is at z = 0. Therefore the expression
found above for G(d) (R) has a pole at d = 2. Make this explicit by using the identity
Γ(z − 1) = Γ(z)/(z − 1) to express G(d) (R) in the form
 
Γ 12 d
G (R) = − d/2
(d)
.
2π (d − 2) Rd−2

A direct derivation using Gauss’ Theorem: As I’ve mentioned already, there

is a simple way to arrive at the final answer found above for G(d) (R), without going
through the Fourier transform, its inversion, and analytic continuation in d. Let’s go
back to the equation
∇2R G(d) (R) = δ (d) (R),
and regard it as Poisson’s equation for the electrostatic potential due to a unit charge
(in suitable units) at R = 0. Let F ≡ ∇G(d) denote the ‘field’ due to this charge.
Integrate both sides of Poisson’s equation over a hypersphere of radius R. Hence
  
dV ∇R G (R) =
2 (d)
dV ∇ · F = dV δ (d) (R) = 1.

Now apply Gauss’ Theorem in vector calculus37 , to obtain

 
dV (∇ · F) = F · dS = 1.

37
This theorem is not restricted to three-dimensional space! It is valid in any number of dimensions,
as it is really a form of the fundamental theorem of the calculus.

115
But the field required is spherically symmetric, and has only a radial component FR
which, moreover, depends only on the magnitude R. Hence
 
F · dS = FR dS = Sd (R) FR = 1,

where Sd (R) is the surface ‘area’ of a hypersphere of radius R in d-dimensional space.

Hence FR = 1/Sd (R).

8. Determine Sd (R). You can do so, for instance, by integrating the ‘solid angle’
element
def.
dΩ(d) = (sind−2 θ1 ) (sind−3 θ2 ) . . . (sin θd−2 ) dθ1 . . . dθd−2 dϕ,
over the full ranges of the angular variables, and multiplying the result by Rd−1 . The
answer is
2π d/2 Rd−1
Sd (R) =   .
Γ 12 d

Returning to our problem, we have

1 
(d) Γ d
dG 1
FR = (∇G(d) )R = = = d/22 d−1 .
dR Sd (R) 2π R

Integrating and imposing the boundary condition G(d) → 0 as R → ∞, we get

 ∞   ∞
dG(d) (R  )  Γ 12 d dR 
dR = −G (d)
(R) = .
R dR  2π d/2 R R  d−1
The integral on the right-hand side converges provided Re d > 2. Keeping d in this
region and evaluating the integral, we get
 
Γ 12 d
G (R) = − d/2
(d)
.
2π (d − 2) Rd−2

This is precisely the expression found earlier. As already pointed out, it is singular at
d = 2. The singularity is a simple pole.

The reason why an inverse-square central force (in three dimensions) leads to an
integral theorem like Gauss’ Law is also obvious now. The surface area of a sphere
increases with its radius like R2 . If the field drops off like 1/R2 , there is obviously an
exact compensation in the flux of the field across a sphere centered at the origin, and
the total flux becomes independent of the radius R. Precisely the same thing happens
in d dimensions, provided the field drops off like 1/Rd−1 , because the ‘surface’ of a hy-
persphere increases like Rd−1 . The potential must then decrease like 1/Rd−2 , exactly

116
as we have deduced.

9. The Coulomb potential in d = 2 dimensions: The case d = 2 requires a sep-

aratetreatment. The reason can be traced, ultimately, to a very simple fact: namely,
that dR/R = ln R, rather than a power of R. As we’ve found, the analytic formula
for G(d) (R) has a simple pole at d = 2. This is a reflection of the fact that the origi-
nal integral representation for G(d) (R) does indeed have an infrared divergence when
Re d ≤ 2. The prescription of dimensional regularization, as applicable to the problem
at hand, is as follows:

(i) Write G(d) (R), which is an analytic function of d, in the form of a Laurent series
about the point d = 2, i.e.,

residue 
∞
(d)
G (R) = + cn (d − 2)n .
(d − 2)
   n=0  
singular part regular part

(ii) Subtract out the singular part. Setting d = 2 in the regular part leaves behind
just the coefficient c0 , which is guaranteed to be the Green function G(2) (R) that we
seek. To find it, expand each of the d-dependent factors in the expression for G(d) (R),
except the pole factor (d − 2)−1 , in a Taylor series about d = 2. Retain only terms up
to the first order in (d − 2), to obtain
1 1 1
G(d) (R) = − + ln R + (ln π + γ) + O(d − 2),
2π(d − 2) 2π 4π

where γ is the Euler-Mascheroni constant. Therefore, according to the prescription de-

scribed above, the true fundamental Green function of the Laplacian in two dimensions
is
1
G(2) (R) = ln R + constant.
2π
The constant is actually arbitrary, and is fixed by specifying a boundary condition.

Note that the boundary condition G(2) (R) → 0 as R → ∞ is not possible in d = 2,

owing to the logarithmic R-dependence of the potential. You will recognize that this
logarithmic potential is essentially the same as the electrostatic potential φ due to
an uniformly charged, infinitely long straight line in three-dimensional space, with R
replaced by , the axial distance from the line. Recall that, in this problem too, the
potential does not vanish, but instead diverges, as  → ∞. What is done then is to
specify the value of the potential at some axial distance a to be equal to a given value
φa . The potential difference φ() − φa is then (λ/2π0 ) ln (/a), where λ is the line
charge density.

117
A direct derivation, once again: Should we believe the result obtained above, as
it seems to have been derived using a prescription that appears to be arbitrary? The
answer is ‘Yes’. Corroboration comes from the same direct physical argument as was
given for the case d ≥ 3. As in that case, we go back to the differential equation

∇2R G(2) (R) = δ (2) (R).

Regard this as Poisson’s equation for the electrostatic potential due to a unit charge
(in suitable units) at R = 0. Let F ≡ ∇G(2) denote the planar vector field due to
this charge. Integrate both sides of Poisson’s equation over a circle of radius R. The
right-hand side of the equation becomes unity, of course. Gauss’ Theorem, applied to
the left-hand side, gives (using the circular symmetry of the field)

2πR FR = 1, so that FR = 1/(2πR).

But FR = dG(2) /dR. Integrating with respect to R gives precisely the expression found
earlier, namely, G(2) (R) = (1/2π) ln R + constant.

• The fundamental Green function of the Laplacian (i.e., the Coulomb potential)
in 2-dimensional Euclidean space is proportional to the logarithm of R.

This seemingly simple fact has profound consequences in diverse areas of physics, such
as condensed matter physics and quantum field theory, among others. It even seems
to have a bearing on the phenomenon of quark confinement!

118
11 The diffusion equation
Fick’s laws of diffusion: Diffusion is the process by which an uneven concentration
of a substance gets gradually smoothed out spontaneously—e.g., a concentration of a
chemical species (like a drop of ink or dye) in a beaker of water spreads out ‘by itself’,
even in the absence of stirring. The microscopic mechanism of diffusion involves a very
large number of collisions of the dye molecules with those of the fluid, which cause the
dye molecules to move essentially randomly and disperse throughout the medium, even
without any stirring of the fluid. A macroscopic description of the process is based
on Fick’s Laws, and leads to a fundamental partial differential equation, the diffusion
equation. This equation serves as a basic model of phenomena that exhibit dissipa-
tion, a consequence of the irreversibility of macroscopic systems in time.

The local, instantaneous concentration ρ(r, t) of dye molecules satisfies the equation
of continuity, which is called Fick’s first law in this context:
∂ρ
+ ∇ · j = 0, (Fick’s I Law)
∂t
where j the ‘diffusion current density’. The crucial physical input is the specification of
this quantity. We assume that j is proportional to the local difference in concentrations,
i.e., to the gradient of the concentration itself. Thus

j(r, t) = −D ∇ ρ(r, t) (Fick’s II Law)

The positive constant D called the diffusion coefficient. It has the physical dimen-
sions of (length)2 /time. The minus sign on the right-hand side signifies the fact that
the diffusion occurs from a region of higher concentration to a region of lower con-
centration: That is, the diffusion current tends to make the concentration uniform.
Eliminating j, we get the diffusion equation for the concentration ρ(r, t):

∂
ρ(r, t) = D ∇2 ρ(r, t).
∂t
This equation is of first order in the time variable, and second order in the spatial
variables. It is a parabolic equation in the standard classification of second-order
partial differential equations. In order to find a unique solution to it, you need an ini-
tial condition that specifies the initial concentration profile ρ(r, 0), as well as boundary
conditions that specify ρ(r, t), for all t ≥ 0, at the boundaries of the region in which the
diffusion is taking place. The presence of the first-order time derivative in the diffusion
equation implies that the equation is not invariant under the time reversal transfor-
mation t → −t. Irreversibility is thus built into the description of the phenomenon.

Fick’s II Law is an example of a very general feature of diverse physical systems

called linear response, that we discussed when we derived dispersion relations for

119
the generalized susceptibility. In the present case, linear response implies that the
diffusion current that is set up in the medium as a result of the unequal concentrations
at different points is proportional to the gradient of ρ, rather than the gradient of some
nonlinear function of ρ (such as a power of ρ other than unity). Another example of
linear response, leading to an exact analog of the diffusion equation, is provided by
the phenomenological description of heat conduction. Given the initial temperature
distribution T (r, 0) of a body, the problem is to find the temperature distribution
T (r, t) at any later time t > 0. Analogous to Fick’s second law, it is assumed that
the heat flux is proportional to the negative of the temperature gradient (a linear
response). The constant of proportionality in this case is the thermal conductivity
of the body, κ. The equation for T (r, t) reads
∂
T (r, t) = κ ∇2 T (r, t).
∂t
This is why the diffusion equation is also known as the heat equation.

1. The fundamental solution in d dimensions: At the level of individual particles,

it turns out that the positional probability density function (or PDF) p(r, t) of a
particle satisfies exactly the same diffusion equation as the concentration ρ(r, t) does
in the macroscopic description of the diffusion process. Consider this equation in a
(Euclidean) space of an arbitrary number of spatial dimensions, d. We can subsequently
set d = 1, 2, 3, . . . in the solution, as required. Let’s therefore begin with
∂
p(r, t) = D ∇2 p(r, t)
∂t
where p(r, t) satisfies natural boundary conditions, i.e., p(r, t) → 0 as r → ∞ along
any direction. We may start with the initial condition

p(r, 0) = δ (d) (r),

where δ (d) (r) is the d-dimensional δ-function. This means that the diffusing particle
starts at the origin at t = 0. In the context of the diffusion equation for the concen-
tration ρ(r, t), such an initial condition represents a point source of unit concentration
at the origin. In a space of infinite extent, we may take the starting point to be the
origin of coordinates without any loss of generality. The solution thus obtained is
the fundamental solution (or Green function) of the diffusion equation. As you’ll see,
it can be used to write down the solution corresponding to an arbitrary initial PDF
p(r, 0) = pinit (r) (or an initial concentration profile ρinit (r); all the results for p(r, t)
that follow are applicable, as they stand, to the case of ρ(r, t).)

The diffusion equation presents an initial value problem. Moreover, it is a linear

equation in the unknown function p(r, t). It is therefore well-suited to the application
of Laplace transforms (with respect to the time variable). As far as the spatial variable

120
r is concerned, it is natural to use Fourier transforms. To avoid confusion, let’s adopt
the following notation:

p(r, t) = the PDF of the position at time t;

 ∞
p(r, s) = dt e−st p(r, t), the Laplace transform of p(r, t);
0
φ(k, t) = dd r e−ik·r p(r, t), the Fourier transform of p(r, t);
 ∞
 s) =
φ(k, dt e−st φ(k, t), the Laplace transform of φ(k, t)
0

= dd r e−ik·r p(r, s), the Fourier transform of p(r, s).

(You can tell which function we’re dealing with by looking at the arguments of the
functions concerned.) Taking the Laplace transform of both sides of the diffusion
equation, we get
s p(r, s) − p (r, 0) = D∇2 p(r, s).
Therefore
(s − D∇2 ) p(r, s) = δ (d) (r).
Now expand p(r, s) in a Fourier integral with respect to the spatial variable r, according
to 
1  s).
p(r, s) = dd k eik·r φ(k,
(2π)d
The δ-function has the familiar Fourier representation

(d) 1
δ (r) = dd k eik·r .
(2π)d
Use these expressions in the equation for p(r, s) and equate the coefficients of the basis
vector eik·r in the space of functions of r. We must then have, for each k,

 s) = 1,  s) = 1
(s + Dk 2 ) φ(k, or φ(k, .
s + Dk 2
We thus obtain a very simple expression for the double transform φ(k,  s). The trans-
forms must be inverted to find the PDF p(r, t). It is easier to invert the Laplace
transform first, and then the Fourier transform.38 The Laplace transform is trivially
inverted: recall that L−1 [1/(s + a)] = e−at . Therefore

−Dk 2 t 1 2
φ(k, t) = e , and hence p (r, t) = d
dd k eik·r e−Dk t .
(2π)
38
You can also invert the Fourier transform first, and then the Laplace transform. This procedure
leads to the same result, as it ought to, but it is needlessly complicated.

121
This d-dimensional integral factors into a product of d integrals upon writing r and k
in Cartesian coordinates. Each of the factors is the familiar ‘shifted’ Gaussian integral.
Show that the final result is
1 2
p (r, t) = d/2
e−r /(4Dt) .
(4πDt)

This is the fundamental Gaussian solution to the diffusion equation in d spatial dimen-
sions.

Observe that the solution is spherically symmetric: the coordinate dependence of

p (r, t) is restricted to the radial coordinate r, with no dependence on the direction of
r. It is important to understand why this comes about. The basic reason, of course,
is that the diffusion equation involves the scalar operator ∇2 , which is rotationally
invariant. But it is also necessary for the boundary conditions and the initial condition
to be spherically symmetric. These requirements are satisfied in the present instance.

Solution for an arbitrary initial distribution: We can now write down the par-
ticular integral that solves the diffusion equation for any specified initial probability
density function. As a trivial generalization, we could take the initial instant to be any
t  . Then, given the PDF pinit (r, t  ) we have, for all t > t  ,
 
1 d  (r − r  )2 
p(r, t) = d r exp − pinit (r , t  ).
[4πD(t − t )]
 d/2 4D(t − t ) 

Thus, the PDF at any time t > t  is an integral transform of the initial PDF. You will
recognize that the kernel of the transform is just the Green function of the operator
(∂/∂t − D∇2 ), and is given by

1  (r − r  )2 
G(r, t ; r  , t  ) = exp −
[4πD(t − t  )]d/2 4D(t − t  )

As I have mentioned already, the heat conduction equation has exactly the same form
as the diffusion equation. For this reason, the Gaussian kernel above is called the heat
kernel in the mathematical literature.

2. Moments of the distance travelled in a time interval t: Let’s revert to


t = 0 as the initial instant. Since p(r, t) is a function of r alone, the mean displacement
of the diffusing particle vanishes for all t ≥ 0, as you would expect:

r(t) = dd r r p(r, t) = 0.

(Integration over the angular coordinates of the vector r yields zero.) The mean dis-
tance travelled by the particle does not vanish, of course, for any t > 0. In fact, all

122
the moments of the distance travelled in a given time interval t can be written down
easily. We find
  ∞ 
r (t) = d r r p(r, t) =
l d l
dr dΩd r d+l−1 p(r, t),
0

on going over to polar coordinates in d dimensions. Since the PDF  depends on r

alone, the
 angular
 integration can be carried out at once. It gives dΩd = Sd (1) =
2π d/2 /Γ 12 d , the surface ‘area’ of a sphere of unit radius in d-dimensional space. The
integration over r involves a Gaussian integral. Verify that
 
Γ 12 (d + l)
r (t) =
l
  (4Dt)l/2 .
Γ 12 d

The noteworthy point is that the moments r l (t) increase with time like tl/2 . Moreover,
this is independent of the dimensionality d of the space in which the diffusion takes
place! In particular, the variance of the displacement is obtained by setting l = 2, and
is given by
Var (r) = r2  − r2 = r 2 = 2dDt.

3. Diffusion in one dimension: continuum limit of a random walk: Let’s turn

now to the case d = 1, i.e., diffusion in one spatial dimension. The calculations are
simpler in this instance, and at the same time we obtain a number of useful insights
into the nature of the diffusion problem. In particular, the effects of boundary condi-
tions can be examined in some detail.

In order to see how diffusion arises as the continuum limit of a random walk,
consider a random walk on a linear lattice. A little more generality is achieved by
considering a biased random walk, with different probabilities for jumps to the right
and left, respectively. on the lattice. This leads to a diffusion equation that is appro-
priate an important physical situation: namely, diffusion on a line in the presence of a
constant external force field. The force-free case is a special case of this equation. The
random walker starts at the site j = 0 at time n = 0. The probability of a step to the
right is α, while that of a step to the left is β = 1 − α. The probability that the walker
is at the site j (that is, at the point ja on the line) at time step n (that is, at time nτ )
is
P (ja, nτ ) = α P (ja − a, nτ − τ ) + β P (ja + a, nτ − τ ).
I have explicitly introduced the lattice constant a and the time step τ , so that it
becomes easy to see how the continuum limit (in both space and time) arises. This
happens when the quantities a, τ and α − β tend to zero simultaneously, as follows.
You will find it instructive to work through the steps outlined below.

123
 Subtract P (ja, nτ − τ ) from each side of the equation for P (ja, nτ ). On the
right-hand side, re-write β as α − (α − β), and the coefficient of P (ja, nτ − τ ) as
−1 = −2α + (α − β). Collect terms suitably to get

P (ja, nτ ) − P (ja, nτ − τ )
 
= α P (ja − a, nτ − τ ) − 2P (ja, nτ − τ ) + P (ja + a, nτ − τ )
 
− (α − β) P (ja + a, nτ − τ ) − P (ja, nτ − τ ) .

Divide both sides by τ . Multiply and divide the first term on the right-hand side by
a2 , and the second term by a. Now let a → 0, τ → 0 and α − β → 0 (that is, let
α → 12 , β → 12 ), such that

a2 α a(α − β)
lim =D and lim = c,
τ τ
where D and c are finite, nonzero constants. Since α → 12 , the constant D is essentially
lim a2 /(2τ ). D has the physical dimensions of (length)2 /(time), while c has the
a,τ →0
physical dimensions of (length)/(time), i.e., a velocity. Further letting j → ∞ and
n → ∞ such that ja and nτ tend to the continuous variables x and t respectively, the
difference equation above for the probability P (ja, nτ ) reduces to the following partial
differential equation for the probability density function (PDF) p(x, t) of the position:

∂p(x, t) ∂p (x, t) ∂ 2 p (x, t)

= −c +D .
∂t ∂x ∂x2
This is called the Smoluchowski equation. It describes diffusion in the presence
of a drift—for instance, the diffusion of colloidal particles in one dimension under the
influence of a constant field of force. The first term on the right-hand side is the drift
term and the second is the diffusion term. The parameter c represents the mean
drift velocity, while D is the diffusion coefficient, as usual. I will refer to diffusion
in the absence of a drift as free diffusion. Observe that the Smoluchowski equation
above can be written in the form of a continuity equation, according to
∂p ∂j
+ = 0,
∂t ∂x
where the current density j(x, t) is given by

∂p (x, t)
j(x, t) = c p(x, t) − D .
∂x
The ratio D/c is a natural length scale in the problem, with the following physical
significance: it is a measure of the relative importance of the drift and diffusion contri-
butions to the motion of the colloidal particles. When the ratio c/D is multiplied by

124
a characteristic length L (such as that used in the definition of the Reynold’s number
in fluid flow), we get a dimensionless number Lc/D, called the Péclet number: it
quantifies the relative strengths of the advective transport rate and the diffusive
transport rate in fluid flow.

An important physical application of the Smoluchowski equation is provided by the

phenomenon of sedimentation, the constant field of force being provided by gravity.
As this phenomenon involves a medium with a boundary, we’ll turn our attention,
next, to the solution of the diffusion equation in the presence of finite boundaries.

Absorbing and reflecting boundary conditions: In general, in physical applica-

tions the region in which diffusion occurs is finite rather than infinite. We then need
to specify appropriate boundary conditions at the end-points of the region. It is im-
portant to recognize that the solutions of a given PDE with the same initial conditions
but different boundary conditions may be very different functions. Let’s consider, for
simplicity and ease of illustration, diffusion in one spatial dimension (the x-axis, say),
inside the ‘box’ given by b1 ≤ x ≤ b2 . If the diffusing particle gets absorbed at the
ends of the box (or can leak out through the ends and thus leave the region of physical
interest), we must impose the so-called absorbing boundary conditions, for which
I’ll use the abbreviation ABC. I’ll also write the corresponding PDF as pabs (x, t), to
avoid any confusion.

pabs (b1 , t) = 0 and pabs (b2 , t) = 0. (ABC)

On the other hand, if the substance cannot leak out of the ends and stays confined in
the region b1 ≤ x ≤ b2 at all times, the diffusion current must vanish at the end-points.
Hence we must impose the so-called reflecting boundary conditions, for which I’ll
use the abbreviation RBC. The corresponding PDF will be written as pref (x, t). These
conditions imply that the flux or current, rather than the PDF itself, vanishes at the
boundaries. In the case of free diffusion, this means that
∂pref
= 0 at x = b1 and x = b2 , for all t ≥ 0. (RBC)
∂x
RBCs are a little more involved when both diffusion and drift are present. Recall that
the Smoluchowski equation for the PDF p(x, t) in the case of diffusion with drift in
one dimension is given by

∂p(x, t) ∂p (x, t) ∂ 2 p (x, t)

= −c +D .
∂t ∂x ∂x2
As mentioned already, we can write this in the form of an equation of continuity,

∂p ∂j ∂p (x, t)
+ = 0, where j(x, t) = c p(x, t) − D .
∂t ∂x ∂x

125
Hence RBCs in this case are given by
∂pref
cpref − D = 0 at x = b1 and x = b2 . (RBC)
∂x
Other boundary conditions, or combinations of these, may be imposed, depending on
the physical problem.

There is a drastic difference in the long-time behavior of the respective solutions

for reflecting and absorbing boundary conditions. In the case of reflecting boundaries,
the particle never leaves the region between the boundaries. (Alternatively, the total
amount of the diffusing substance remains conserved.) Therefore the normalization
b
condition, b12 dx pref (x, t) = 1, remains valid for all t ≥ 0. In the case of absorbing
boundaries, however, the particle is absorbed when it happens to hit either bound-
ary. The diffusion process then comes to an end (or the diffusing substance leaks out
through the end points). In this case it is clear that pabs (x, t) → 0 as t → ∞, at any
b
point x inside the box. The total probability b12 dx pabs (x, t) = S(t) then represents
the survival probability inside the box. It starts with an initial value equal to unity,
and decreases to zero with increasing time. The question of precisely how S(t) → 0
as t → ∞ is of interest in its own right. We’ll consider it shortly, in the case of free
diffusion.

4. Free diffusion on a semi-infinite line: Consider free diffusion in the semi-

infinite region −∞ < x < b, where b is a positive constant. On the left, we have the
natural boundary condition p(x, t) → 0 as x → −∞. On the right, we have either
(i) pabs (b, t) = 0 if the barrier at x = b is an absorber, or (ii) (∂pref /∂x)x=b = 0 if the
barrier is a reflector. Verify that, for the initial condition p(x, 0) = δ(x), the solutions
in the two cases are give by
1  2 
−x /(4Dt) −(2b−x)2 /(4Dt)
pref (x, t) = e + e
(4πDt)1/2

and  2 
1 −x /(4Dt) −(2b−x)2 /(4Dt)
pabs (x, t) = e − e ,
(4πDt)1/2
respectively.

5. Finite boundaries: Solution by the method of images: The solutions above

have a striking interpretation. They are superpositions (the sum and the difference,
respectively) of (Gaussian) solutions of the diffusion equation with natural boundary
conditions, but with x and (2b − x), respectively, as the coordinate argument. Now,
(2b − x) is precisely the coordinate of the image of the point x reflected in a mirror
located at the boundary b. This is no coincidence! The solutions above can indeed be
obtained by a general technique called the method of images. You have come across

126
this method in the context of electrostatics, where it yields the solution to Poisson’s
equation in the presence of specific boundary conditions, provided the problem has
a convenient symmetry. However, the method is quite general, and is applicable in
other instances as well. Broadly speaking, the method of images is a technique to find
the Green function of a linear differential operator with specific boundary conditions,
provided the problem has some symmetry that can be exploited. It is based on the
fact that the differential equation concerned, taken together with appropriate initial
and boundary conditions, has a unique solution. My purpose in using the method
of images in the present context of the diffusion equation is two-fold. The first is to
demonstrate how powerful the method is, when the conditions are right. The second
is to show that the method is also applicable to time-dependent problems, and not just
time-independent problems as in electrostatics.

Consider free diffusion in the line segment region −b ≤ x ≤ b, where b is a positive

constant. For simplicity, we assume once again that the diffusing particle starts at the
origin at t = 0, so that p(x, 0) = δ(x). We want the solution to the diffusion equation
subject to this initial condition and
% &
(i) ∂pref (x, t)/∂x x=±b = 0 (RBC); or (ii) pabs (±b, t) = 0 (ABC).

Imagine placing mirrors at both boundaries, facing each other. A source point x
therefore has an infinite number of images, at the points −2b − x, 2b − x, −4b + x, 4b +
x, . . ., as we move away from the source point on both sides of the x-axis. (Draw a
figure and mark the successive image points.) The solution is a superposition of the
fundamental Gaussian solution with each of these points as the coordinate argument.
The boundary conditions above are incorporated by the following simple prescriptions:

(i) For RBC, the coefficient of the contribution from each image is just 1.

(ii) For ABC, the coefficient of the contribution from an image arising from n reflec-
tions is (−1)n .

The complete solutions in the two cases may therefore be written down. Show that,
after a bit of simplification, they are given by

1 
∞
2
pref (x, t) = 1/2
e−(x+2nb) /(4Dt) (RBC at x = ±b)
(4πDt) n=−∞

and
1 ∞
n −(x+2nb)2 /(4Dt)
pabs (x, t) = (−1) e (ABC at x = ±b).
(4πDt)1/2 n=−∞

127
6. You may now ask: can each of these solutions be written as the fundamental Gaus-
sian solution on the infinite line, plus an ‘extra’ piece arising from the presence of the
finite boundaries? Show that the solutions above can be re-written in the form
) *
2
e−x /(4Dt) ∞
2 2

nbx 
pref (x, t) = 1+2 e−n b /(Dt) cosh (RBC)
(4πDt)1/2 n=1
Dt

and
) *
2
e−x /(4Dt) 
∞
nbx 
−n2 b2 /(Dt)
pabs (x, t) = 1+2 (−1)n e cosh (ABC),
(4πDt)1/2 n=1
Dt

respectively. It is important to remember, however, that the physical region in which

these solutions are valid is restricted to the line interval [−b, b].

• It is remarkable that the only difference between the solutions for reflecting and
absorbing boundary conditions is the extra factor (−1)n in the summand in the
latter case. But this is sufficient to alter completely the long-time behavior of
the PDF p(x, t), as you will see below.

7. Finite boundaries: Solution by separation of variables: You are undoubt-

edly familiar with an elementary method for the solution of partial differential equa-
tions, namely, the method of separation of variables The diffusion equation, with an
initial condition p(x, t) = δ(x) and either reflecting or absorbing boundary conditions
at x = ±b, is tailor-made for solution by this method. Set p(x, t) = T (t) X(x), as
usual, in the diffusion equation. It follows that (1/DT )dT /dt = (1/X)dX/dx = −C 2 ,
a constant. The boundary conditions yield the allowed values of C. The general solu-
tion is a superposition of solutions for the various allowed values of C, as the diffusion
equation is a linear equation. The calculations are simplified by noting that the diffu-
sion equation, the set of boundary conditions, as well as the initial condition, are all
unchanged under the transformation x → −x. As a consequence, the solution p(x, t)
remains an even function of x at all times. This fact will help you select the correct
solution in each case. You will also need the following representation of the initial
PDF,
1
∞
1 nπx
δ(x) = + cos ,
2b b n=1 b
  
∞
which follows from the relation δ x/(2b) = eπnix/b . Show that
n=−∞

1  − n2 π2 Dt/b2
∞
1 nπx
pref (x, t) = + e cos (RBC)
2b b n=1 b

128
and
1  −(2n+1)2 π2 Dt/(4b2 )
∞
(2n + 1)πx
pabs (x, t) = e cos (ABC).
b n=0 2b

8. Survival probability: The representations of p(x, t) just found are useful in

reading off the long-time behavior of the PDF, because they are superpositions of
decaying exponential functions of time. It follows by inspection that pref (x, t) → 1/(2b)
as t → ∞ in the case of reflecting boundaries. This uniform distribution in the interval
[−b , b] is precisely what we would expect on physical grounds as the asymptotic PDF.
On the other hand, in the case of absorbing boundaries, pabs (x, t) → 0 as t → ∞ for
every value of x. This is an indication of the fact that absorption at either one boundary
or the other is a sure event for the random process concerned—that is, it will occur
with probability 1. The survival probability S(t, ±b | 0) is the probability that,
starting from the origin at t = 0, the diffusing particle survives in the open interval
(−b, b) till time t, without hitting either of the boundary points.39 Its definition is
obvious:  b
S(t, ±b | 0) = dx pabs (x, t).
−b

(a) Using the solution for pabs (x, t) with ABC, check that

4  (−1)n − (2n+1)2 π2 Dt/(4b2 )

∞
S(t, ±b | 0) = e .
π n=0 (2n + 1)

verify that S(0, ±b | 0) = 1, as required. You will need the value40 of the
(b) Hence 
series ∞ n
0 (−1) /(2n + 1).

S(t, ±b | 0) is a superposition of an infinite number of decaying exponentials that de-

2 2
cays monotonically to zero, with a leading asymptotic behavior ∼ e−π Dt/(4b ) .

9. First-passage-time distribution & mean first-passage time: A diffusing

particle starting from x = 0 will hit one of the end-points ±b for the first time at some
random instant of time, which may be called the first-passage time.41 What is the
distribution of this time? This sort of question occurs in numerous applications—for
instance, in reaction-diffusion problems in chemical physics.

39
More generally, the survival probability in any region should be computed by averaging over all
possible starting points as well, but I do not go into this detail here.
40
This was first deduced by the father of mathematical analysis, Madhava of Sangamagrama (1350-
1425), as a special case of the power series for tan−1 x discovered by him. The latter was rediscovered
a few centuries later by Gregory (1638-1675), and is now known as the Madhava-Gregory series.
41
Also called the escape time, or hitting time, or exit time, depending on the application.

129
 Let Q(t, ±b | 0) dt be the probability for a particle starting from x = 0 at t = 0 to
reach either b or −b for the first time in the time interval (t, t + dt), without ever having
hit either of the end-points at any earlier time. It is obvious that the probability density
Q(t, ±b | 0) is nothing but the rate of decrease of the survival probability S(t, ±b | 0).
That is,

πD 
∞
d 2 2 2
Q(t, ±b | 0) ≡ − S(t, ±b | 0) = 2 (−1)n (2n + 1) e−(2n+1) π Dt/(4b ) .
dt b n=0
∞
(a) Verify that Q(t, ±b | 0) is normalized to unity, i.e., 0
dt Q(t, ±b | 0) = 1.
Hence a first passage to either one or the other of the end-points is a sure event, as
already stated.

The mean first-passage time (MFPT) t(0 → ±b) is also easily determined. By
dimensional considerations, it must be proportional to b2 /D, which is the only time
scale in the problem.

(b) Show that the MFPT t(0 → ±b) = b2 /(2D), justifying the
term ‘diffusion time’
for this mean value. You will need the value of the sum ∞ n 3
0 (−1) /(2n + 1) ,
3
which is π /32.

In many instances, first-passage-time problems are conveniently handled by considering

the Laplace transforms of the corresponding first-passage-time densities. In the present
 ±b | 0) denote the Laplace transform of Q(t, ±b | 0). First passage
instance, let Q(s,
(from the starting point 0 to either of the points ±b) is a sure event if and only if
 ±b | 0) = 1.
Q(0,

Further, the mean first-passage time is given by

 dQ

t(0 → ±b) = − .
ds s=0

Connection with the Schrödinger equation for a free particle: Let’s return
to the fundamental solution of the diffusion equation in d-dimensional space, for an
arbitrary initial PDF. This solution enables us to write down a similar solution for
another very important equation, namely, the Schrödinger equation for the position-
space wave function ψ(r, t) of a nonrelativistic free particle of mass m moving in d-
dimensional space:
∂ψ(r, t) i
2
= ∇ ψ(r, t).
∂t 2m
130
It is obvious that this equation has exactly the same form as the diffusion equation,
with the physically and mathematically important difference that the real diffusion
constant D in the latter is replaced by the pure imaginary constant i
/(2m). We can
now write down the formal solution to the Schrödinger equation by analogy with the
solution of the diffusion equation. Given that the wave function is ψ(r, t  ) at some
instant of time t  , the solution at any time t > t  is given by
d/2   
m d  im(r − r  )2
ψ(r, t) = d r exp ψ(r , t  ).
2πi
(t − t )
 2
(t − t )


But we know from quantum mechanics that the solution must have the form

ψ(r, t) = dd r  K(r, t ; r , t  ) ψ(r , t  ), (t > t  )

where K(r, t ; r  , t  ) is the free-particle Feynman propagator. This quantity is de-

fined as +   ,
K(r, t ; r , t  ) = re−iH(t−t )/
r  ,


where H is the free-particle Hamiltonian p2 /(2m), and |r is the position eigenstate
of the particle corresponding to the position eigenvalue r. Hence the explicit form
of the free-particle propagator for a nonrelativistic particle moving in d-dimensional
Euclidean space is
d/2  
  m im(r − r )2
K(r, t ; r , t ) = exp (t > t  ).
2πi
(t − t  ) 2
(t − t  )

The replacement of the real constant D by the pure imaginary constant i
/(2m) actu-
ally represents a drastic change. The kernel exp [−(r−r  )2 /4D(t−t  )] decays to zero as
|r−r  | → ∞. On the other hand, the kernel exp [im(r−r  )2 /2
(t−t  )] is an oscillatory
function with a modulus equal to unity. As a result, questions of convergence arise, and
these require careful handling. I do not go into these aspects here, but merely mention
that the formal similarity between the diffusion equation and the Schrödinger equation
has important consequences. The solution for ψ(r, t) written down above serves as the
starting point for the path integral formulation of quantum mechanics.

10. Spreading of a quantum mechanical wave packet: As you know from ele-
mentary quantum mechanics, the wave packet representing a free particle undergoes
dispersion, i.e., it broadens or spreads in time, even though the particle moves in a
vacuum. The physical reason for this fact is easy to see. Let ε and p denote the energy
and the magnitude of the momentum of the particle, respectively. Wave-particle dual-
ity is expressed by the Einstein-de Broglie relations ε =
ω and p =
k, where ω is the
angular frequency and k is the wave number. Hence the relation ε = p2 /(2m) becomes
ω =
k 2 /(2m). This nonlinear dispersion relation immediately implies that the phase
velocity ω/k is not equal to the group velocity dω/dk, i.e., there is dispersion. In other

131
words, an initial wave packet will change shape and spread with time, even though the
particle is free, and not under the influence of any force.

In order to see this quantitatively, let’s consider the problem in one spatial dimen-
sion. Setting t  = 0 for simplicity, we have

m 1/2  ∞  2
ψ(x, t) = dx eim(x−x ) /(2
t) ψ(x  , 0).
2πi
t −∞

We need an initial state that represents the quantum mechanical counterpart of a

classical free particle moving with some constant momentum p0 . We could start with
|p0 , which is an eigenstate of the momentum of the particle, corresponding to the
eigenvalue p0 . The position-space wave function corresponding to this state is x | p0.
This is a plane wave proportional to exp (ip0 x/
). But such a wave function has a
constant modulus, and is clearly not normalizable in (−∞, ∞), whereas we would like
to work with square-integrable functions throughout. This is achieved by modulating
the plane wave with a Gaussian, centered at the origin, say. The initial wave function
is then given by
2 2
ψ(x, 0) = (2πσ 2 )−1/4 eip0 x/
e−x /4σ ,
where σ is a positive constant with the physical dimensions of a length. The wave
∞
function is normalized according to −∞ dx |ψ(x, 0)|2 = 1.
(a) Show that
 ∞  ∞
x(0) = dx x |ψ(x, 0)| = 0, x (0) =
2 2
dx x2 |ψ(x, 0)|2 = σ 2 .
−∞ −∞

Therefore the initial uncertainty (i.e., the standard deviation) in the position of the
particle is ∆x(0) = σ.
(b) Show also that
 ∞
d
p(0) = dx ψ ∗ (x, 0) (−i
) ψ(x, 0) = p0 ,
−∞ dx
while  ∞
d2
2
p (0) =
2
dx ψ ∗ (x, 0) (−i
)2 ψ(x, 0) = p 2
0 + .
−∞ dx2 4σ 2
Hence the initial uncertainty (or standard deviation) in the momentum is ∆p(0) =

/(2σ). It follows that ∆x(0) ∆p(0) = 12
. In other words, the initial state is a mini-
mum uncertainty state.

The Hamiltonian of a free particle is given by just the kinetic energy term, p2 /(2m).
As this is quadratic in the dynamical variable concerned, Ehrenfest’s Theorem im-
plies that the expectation value p(t) will remain equal to p0 for all t, while x(t) will
be given by p0 t/m: these are the expressions that would obtain for a classical particle.
These statements can be checked out directly, as follows.

132
 (c) The integral representing ψ(x, t) now becomes a shifted Gaussian integral, and
can be evaluated, to find the wave function at any time t. Show that
eφ(x,t)
ψ(x, t) = ,
(2πσ 2 )1/4 [1 + (i
t/2mσ 2 )]1/2
where the exponent φ(x, t) is given by
imx2 im(x − p0 t/m)2 (x − p0 t/m)2
φ(x.t) = − − .
2
t 2
t[1 + (
2 t2 /4m2 σ 4 )] 4σ 2 [1 + (
2 t2 /4m2 σ 4 )]

The first two terms on the right-hand side of the last equation are purely imaginary,
and hence only contribute phase factors to ψ(x, t). They do not, therefore, affect the
probability density function |ψ(x, t)|2 . This quantity is given by
 1/2  
1 (x − p0 t/m)2
|ψ(x, t)| =
2
exp − .
2π[σ 2 + (
2 t2 /4m2 σ 2 )] 2[σ 2 + (
2 t2 /4m2 σ 2 )]
Thus, the PDF of the position of the particle at any time t > 0 is also a normalized
Gaussian, with its peak at x = p0 t/m, and a variance that increases with time.
(d) Show that
p0 t 
x(t) = , 

m
 2
2 t2 
x2 (t) − x(t)2 ≡ ∆x(t) = σ 2 + . 
4m2 σ 2
(e) Show also that
 ∞ 
p(t) = ∗ d 

dx ψ (x, t) (−i
) ψ(x, t) = p0 , 

−∞ dx 
 
∞
d2
2 

p (t) =
2
dx ψ ∗ (x, t) (−i
)2 ψ(x, t) = p 2
0 + . 

−∞ dx2 4σ 2

We thus arrive at the following conclusions:

(i) The initial Gaussian wave packet in position space remains a Gaussian wave
packet, but broadens as a function of time.
(ii) The uncertainty in the position of the particle increases with time according to
 1/2

2 t2
∆x(t) = σ 1 + .
4m2 σ 4
At very long times (t  2mσ 2 /
), the uncertainty in the position increases
linearly with time.

133
(iii) The uncertainty in the momentum at any time t remains equal to its initial value,

/(2σ). Thus the state of the particle is no longer a minimum uncertainty state
for any t > 0.

Observe, incidentally, that the energy of the particle (i.e., the expectation value of its
Hamiltonian) in the state under consideration is not p20 /(2m), but rather

1 p2
2
E ≡ H = p2  = 0 + .
2m 2m 8mσ 2
H, of course, is constant in time.

(f) The particle is not in a momentum eigenstate either at t = 0 or at any time

t > 0. Why, then, do the expectations values of p and p2 remain unchanged from
their initial values, while those of x and x2 are functions of time?

134
12 The Green function for (∇2 + k 2); nonrelativistic
scattering
The Helmholtz operator: After the Laplacian operator ∇2 and the diffusion op-
erator (∂/∂t − D∇2 ), we turn now to the Helmholtz operator ∇2 + k 2 , where k is a
constant. As you know, the normal modes of vibration of a region R are given by the
solutions of the Helmholtz equation (∇2 + k 2 )u(r, t) = 0, i.e., the eigenvalue equation
∇2 u = −k 2 u, with appropriate boundary conditions on the ‘displacement’ u at the
boundary S of the region. The most common boundary condition is the Dirichlet
boundary condition u(r, t) = 0 for r ∈ S. A vast literature exists on this problem
and its solutions.

Here, we shall not be concerned with the foregoing eigenvalue problem, but rather,
with the inhomogeneous equation (∇2 + k 2 ) f (r, t) = g(r, t). More specifically, we are
interested in finding the fundamental Green function for the operator ∇2 + k 2 . As
it is useful to work in the context of a specific physical application, I will consider
quantum mechanical scattering theory. For simplicity, we’ll focus on the elastic scat-
tering of a nonrelativistic particle of mass m from a static central potential V (r) that
vanishes as r → ∞. It is convenient (and conventional) to introduce a parameter λ as
a measure of the ‘strength’ of the potential, so that various physical quantities can be
expressed as power series in λ, and approximations to different orders in λ can be made.

The scattering amplitude; differential and total cross-sections: The time-

independent Schrödinger equation for the position-space wave function of the particle
is given by

2 2
− ∇ ψ(r) + λV (r) ψ(r) = E ψ(r).
2m
We are interested here in the scattering states of the particle, i.e., states belonging to
the continuous part of the spectrum of the Hamiltonian, with energy eigenvalue E > 0.
These states are not normalizable, i.e., the corresponding wave functions are not square-
integrable.42 The initial state of the particle is taken to be a momentum eigenstate,
with eigenvalue p =
k. It is therefore represented by the position-space wave function
ψinc (r) ≡ r|p = eip·r/
= eik·r . Thus k is the direction of the momentum of the
incident particle. After scattering, the wave vector changes direction without changing
its magnitude, because the scattering is elastic: the energy remains

E =
2 k 2 /(2m)

throughout. The scattered wave vector can be directed along any direction in space.
The quantity of interest is the probability of scattering in any particular direction. This
42
This technical difficulty can be overcome by working with normalizable wave packets rather than
plane waves as wave functions. But we shall not enter into this aspect here.

135
is measured by the differential cross-section dσ/dΩ, defined as
dσ flux of particles per unit solid angle in the direction concerned
= .
dΩ incident flux

Sufficiently far away from the scattering center (i.e., as r → ∞, or more precisely,
for kr  1), the scattered wave has the form of an outgoing spherical wave. The radial
coordinate dependence of this wave is (in three-dimensional space) eikr /r, and it is, in
general, modulated by an amplitude f relative to the unit amplitude of the incident
plane wave. The amplitude f is, in general, a function of the angular coordinates θ
and ϕ, as well as the energy E (or the wave number k). Without loss of generality, we
may choose spherical polar coordinates such that the polar axis is along the direction
of the incident wave vector k. Then, the spherical symmetry of a central potential
clearly implies that there can be no ϕ-dependence in the amplitude f . There remains
a nontrivial θ-dependence, because the presence of an incident wave vector k breaks
the spherical symmetry, leaving an axial symmetry about the direction of k. The
total wave function is a superposition of the incident and scattered wave functions. Its
asymptotic form given by
eikr
ψ(r) = ψinc (r) + ψsc (r) −−−→ eik·r + f (k, θ) ,
kr 1 r
The energy and angle-dependent factor f (k, θ) is called the scattering amplitude.

To repeat: The scattering amplitude f is independent of the azimuthal angle

ϕ because the scattering potential is spherically symmetrical. But it does have θ-
dependence, because the incident wave vector k singles out a special direction. The
potential is spherically symmetric, but the symmetry of the scattered wave function is
reduced to an axial or cylindrical symmetry about the direction of the initial momen-
tum of the particle.

1. There is a direct relation between the differential cross-section and the scattering
∗ ∗
amplitude. incident current density is
(ψinc ∇ψinc − ψinc ∇ψinc )/(2mi), which is easily
simplified to
k/m. Its magnitude
k/m is the incident flux. The scattered current
∗ ∗
density is given by
(ψsc ∇ψsc − ψsc ∇ψsc )/(2mi). What we need here is the radially
outward scattered flux through a cone of solid angle dΩ located at any angular position
(θ, ϕ). This is given by
 ∗


∗ ∂ψsc ∂ψsc
ψsc − ψsc r 2 dΩ.
2mi ∂r ∂r
Using the asymptotic form above for ψsc , show that this quantity simplifies to

(
k/m) |f (k, θ)|2 dΩ.

136
Hence the differential cross-section is given by

dσ/dΩ = |f (k, θ)|2 .

The total cross-section for scattering is then

   1
def.
σ = dσ = dΩ |f (k, θ)| = 2π
2
d (cos θ) |f (k, θ)|2.
−1

σ is obviously a function of k alone, i.e., a function of the energy E of the incident

particle.

Integral equation for scattering: In order to find the scattering amplitude, we

need the asymptotic form of the solution for the wave function. For this purpose, it
is convenient to convert the time-independent Schrödinger equation from a differential
equation to an integral equation. First write the differential equation in the form

(∇2 + k 2 ) ψ(r) = λ U(r) ψ(r), where U(r) = (2m/
2 ) V (r).

Suppose, for a moment, we treat the right-hand side λU(r)ψ(r) as a ‘source’ term
in an inhomogeneous differential equation. The general solution will be the sum of
the complementary function and the particular integral. The former is a solution of
the homogeneous equation (∇2 + k 2 ) ψ(r) = 0, chosen so as to satisfy the boundary
conditions. In the present case, the complementary function is just the incident wave
eik·r . The particular integral involves G(r, r ), the Green function for the Helmholtz
operator (∇2 + k 2 ). Thus

ψ(r) = e + λ d3 r  G(r, r ) U(r  ) ψ(r ).
ik·r

The presence of the unknown function ψ in the integrand in the last term makes this
an inhomogeneous integral equation for the wave function, rather than a solution for
this quantity.

2. Green function for the Helmholtz operator: The Green function satisfies the
equation
(∇2r + k 2 )G(r, r ) = δ (3) (r − r ).
Now, the operator ∇2r + k 2 , the the δ-function on the right-hand side, as well as the
free boundary condition (the vanishing of G as r → ∞) are all translation invariant
(i.e., invariant under the shift r → r − r ). Hence G will once again turn out to be a
function of R = r − r  . Introducing the Fourier transform of G(R) and following the
same steps as in the case of ∇2 , we now get

1 eiq·R
G(R) = − d 3
q .
(2π)3 q2 − k2

137
I have used q for the Fourier transform variable conjugate to R, since k has already
been used for the incident wave vector. As usual, go over to spherical polar coordinates
in q-space, and choose the polar axis along the direction of the vector R. Carry out
the angular integrations to get
 ∞  ∞
1 q sin qR 1 q (eiqR − e−iqR )
G(R) = − 2 dq 2 = − dq .
2π R 0 q − k2 8π 2 iR −∞ q2 − k2

Note that the answer is a function of R alone, which is why G(R) has been written
as G(R). We must now use an appropriate i-prescription to avoid the singularities of
the integrand at the points q = ±k on the path of integration.

• This prescription will also help us select the correct Green function, i.e., the
one that corresponds to an outgoing spherical wave with an asymptotic behavior
∼ eikr /r.

The next step is to close the contour of integration in the complex q-plane by adding
a semicircle to the line integral. The contribution of this semicircle to the integral must
vanish as its radius tends to infinity. It must therefore lie (i) in the upper half-plane for
the term involving eiqR , and (ii) in the lower half-plane for the term involving e−iqR .
Moreover, we want G(R) to have an asymptotic behavior ∼ eikr /r as r → ∞. Hence:

(i) Only the pole at q = k must contribute to the term proportional to eiqR . There-
fore, the pole at q = k must be displaced into the upper half-plane.

(ii) Only the pole at q = −k must contribute to the term proportional to e−iqR .
Therefore, the pole at q = −k must be displaced into the lower half-plane.

The correct i-prescription is thus implemented by setting

 ∞
1 q (eiqR − e−iqR )
G(R) = − 2 lim dq 2 ,
8π iR
→0 −∞ q − (k + i)2

where  is an infinitesimal positive number. Close the contour as prescribed above

(draw the corresponding figures) and use the residue theorem to show that

eikR eik|r−r |
G(R) = − =− .
4πR 4π|r − r  |

3. It is easy to work out the Green functions obtained by adopting the three other
ways in which the poles at q = ±k can be given infinitesimal imaginary displacements.
Let  be a positive infinitesimal, as usual.

138
 (a) Suppose we displace the pole at +k into the lower half-plane in q, and the pole
at −k into the upper half-plane. (Draw a figure.) Verify that the Green function
then vanishes identically, i.e.,

1 eiq·R
− lim d 3
q = 0.
(2π)3
→0 q 2 − (k − i)2

(b) Similarly, suppose both poles are displaced into either the upper or the lower half-
plane. (Draw the corresponding figures.) Show that the corresponding Green
function is 
1 eiq·R cos kR
− lim d 3
q =− .
3
(2π)
→0 (q ± i) − k
2 2 4πR

Exact formula for the scattering amplitude: Using the correct Green function,
the integral equation for scattering becomes
 ik|r−r  |
λ 3  e
ψ(r) = e −
ik·r
dr U(r  ) ψ(r ).
4π |r − r |

We need to extract the asymptotic behavior of the wave function from this integral
equation, and identify the scattering amplitude. As r → ∞ (more precisely, for kr 
1), we may replace the factor 1/|r − r | in the integrand by just r −1 . The implicit
assumption here is that the potential V (r  ) decays rapidly for very large values of r  ,
so that the contribution to the integral from the region in which both r and r  are large
is quite negligible. In the pure phase factor exp (ik|r − r |), however, we need to be
more careful. The general field point r at which the wave function is being calculated
is the point at which the particle detector is located, in order to measure the flux per
unit solid angle in that direction. Thus the wave vector of the scattered particle, k  ,
is in the same direction as r: that is, k er = k  . We then have
 1/2
2r · r  1/2
ik |r − r | = ik r 2 − 2r · r  + r  2  ikr 1 −
r2

r·r 
 ikr 1 − 2 = ikr − ik er · r  = ikr − ik  · r .
r
Inserting these approximations in the integral equation for ψ(r), we get

eikr λ  
ψ(r) → e −
ik·r
d3 r  e−ik ·r U(r  ) ψ(r ).
r 4π
Comparing this result with the asymptotic form of the wave function, it follows at once
that the scattering amplitude is given by the formula

λ  
f (k, θ) = − d3 r  e−ik ·r U(r  ) ψ(r ).
4π

139
While this formula is exact, it is not of much use unless we already know the full wave
function ψ(r) at all points in space.

Scattering geometry and the momentum transfer: For scattering in any given
direction, the vector

Q =
(k  − k)
is the final momentum of the particle minus its initial momentum. That is, Q is the
wave vector corresponding to the momentum transfer associated with the scattering
process. Recall that the energy E is related to k = |k| = |k  | according to E =

2 k 2 /(2m). The relation between k, Q and θ is also easy to derive. (Draw a figure.) It
is easily seen that
   
Q2 = 2k 2 (1 − cos θ) = 4k 2 sin2 12 θ , so that Q = 2k sin 12 θ .

θ = 0 corresponds to forward scattering, while θ = π corresponds to backward

scattering.

Born series and the Born approximation: Going back to the inhomogeneous
integral equation for the wave function, we can solve it iteratively for sufficiently small
values of the ‘strength’ |λ| of the potential. We find
 ik|r−r  |
3  e 
ψ(r) = e − (λ/4π) d r
ik·r
U(r  ) eik·r
|r − r | 
  ik|r−r  | ik|r  −r  |
3  3  e  e 
2
+ (λ/4π) d r d r U(r )  U(r  ) eik·r
|r − r | |r − r | 

+ ···

The precise conditions on the potential under which this is a convergent series are
discussed in textbooks on quantum mechanics, and I shall not go into this aspect here.
Very roughly speaking, the series solution is valid when the effect of the potential is
weak as compared to the kinetic energy of the incident particle. Substitution of the so-
lution above in the formula for f (k, θ) gives an expression for the scattering amplitude
as a power series in λ, called the Born series.

It is trivial to write down the scattering amplitude to first order in λ. All we have
to do is to approximate ψ(r  ) in the integrand by the incident wave exp (ik · r ) itself.
This is called the (first) Born approximation. Denoting the scattering amplitude in
this approximation by fB (k, θ), we have


fB (k, θ) = −(λ/4π) d3 r  e−iQ·r U(r  ), where Q = k  − k.

Hence the scattering amplitude in the Born approximation is, up to a constant fac-
tor, the (three-dimensional) Fourier transform of the potential, with the momentum

140
transfer wave vector Q playing the role of the variable conjugate to r. Now, eik·r =
eip·r/
= r | p is the position space wave function corresponding to the initial momen-
−ik  ·r 
tum eigenstate |p of the particle.  3 Similarly, e = e−ip ·r
= r | p ∗ = p  | r.
Using the completeness relation d r |r r| = I, we see that

λ mλ
fB (k, θ) = − p  |U|p = − 2
p  |V |p.
4π 2π

In other words:

• The scattering amplitude in the Born approximation is essentially the matrix el-
ement of the potential energy operator between the initial and final free-particle
momentum eigenstates.

4. The angular integration in the expression for fB (k, θ) is easily carried out. Show
that 
2mλ ∞
fB (k, θ) = − 2 dr r sin (Qr) V (r),

Q 0
where Q is the magnitude of the momentum transfer vector Q.

5. For a finite potential barrier given by

)
V0 for r ≤ a
V (r) =
0 for r > a

show that
2mλV0 (sin Qa − Qa cos Qa)
fB (k, θ) = .

3 Q3
Hence show that the forward scattering amplitude in the Born approximation is fB (k, 0) =
(2mλV0 a3 )/(3
2 ).

6. The Yukawa potential: According to quantum field theory, the forces between
elementary particles arise from the exchange of other particles which are the quanta
of gauge fields. In the nonrelativistic limit of scattering from a static potential, such
a force generically reduces to a form called the Yukawa potential, given by

e−r/ξ
V (r) = ,
r
where ξ is a positive constant with the physical dimensions of length. It represents the
‘range’ of the potential, and is essentially the Compton wavelength of the exchanged
particle. The latter is given by h/(µc), where µ is the (rest) mass of the exchanged
particle. The functional form of the Yukawa potential arises in other physical contexts
as well. For instance, the screened Coulomb potential in a dielectric medium has

141
the Yukawa form.

Show that, in the Born approximation, the scattering amplitude for the Yukawa
potential is given by
2mλξ 2
fB (k, θ) = − 2 2 2 ,

(Q ξ + 1)
where Q is the magnitude of the momentum transfer vector, as usual. Hence show
that the total cross-section in the Born approximation is
 
4π 2 mλ2 ξ 2 4mEξ 2 +
2
σB (E) = .
E
2 8mEξ 2 +
2

For very large values of the incident energy E, the total cross-section falls off like 1/E.
This is a general feature of scattering from a wide class of potentials. (This class,
however, does not include the Coulomb potential!)

The Coulomb potential corresponds to the ξ → ∞ limit of the Yukawa potential.43

It follows that the scattering amplitude in the Born approximation is now fB (k, θ) =
−(2mλ)/(
2 Q2 ). The differential cross-section is therefore

dσ λ2  
4 1
= cosec 2
θ .
dΩ B 16E 2
This is precisely the famous Rutherford scattering formula. There are several re-
markable features about this result:

(i) It is also the exact expression for the differential cross-section for scattering in the
Coulomb potential λ/r, because all the higher-order corrections to the first Born ap-
proximation happen to vanish in this instance.

(ii) Moreover, the expression is actually independent of Planck’s constant, and is ex-
actly the same as the differential scattering cross-section of classical particles in a
Coulomb potential.

(iii) The differential cross-section for forward scattering (θ = 0) diverges. So does the
total cross-section, as the divergence at θ = 0 is not integrable.

All these ‘minor miracles’ and drawbacks are related to the long-range nature of the
Coulomb potential. It turns out that the assumption of an initial plane wave state
(ψinc = eik·r ) is incompatible with a potential that decays as slowly as 1/r. The
scattering problem has to be re-done in this case. The Schrödinger equation in the
43
If we set ξ = h/(µc), then ξ → ∞ corresponds to µ = 0. Recall that electromagnetic interactions
are mediated by the exchange of photons, which are massless quanta.

142
presence of a Coulomb potential can be solved exactly, using parabolic cylindrical
coordinates. The solution involves confluent hypergeometric functions. The initial
state is found to be a plane wave that is amplitude-modulated by a factor of 1/r, and
phase-modulated by a term involving ln r.

143
13 The wave equation
The last standard equation of mathematical physics to which we turn our attention in
this course is the wave equation. To be specific, we’ll consider the fundamental Green
function of the wave operator, that corresponds to the causal, retarded solution of
the wave equation. This is the solution that is of direct physical significance for the
propagation of signals.

Formal solution for the causal Green function: The wave equation for a scalar
function (or signal) f (r, t) is given by

1 ∂2 
− ∇ f (r, t) = g(r, t),
2
c2 ∂t2
where g(r, t) is a specified function of space and time that represents the source of
the signal, and c is the speed of the signal. Although I have used c to denote this
speed, what follows is not restricted to the propagation of light in free space. We’ll
first consider, formally, wave propagation in a general d-dimensional Euclidean space
of infinite extent,44 and then work out the explicit solutions for the cases d = 1, 2 and
3 that are of direct physical interest. Subsequently, I’ll also comment on what happens
when d > 3.

The crucial difference between the wave equation and Poisson’s equation, is, of
course, the relative minus sign between the time derivatives and the Laplacian. This
makes all the difference in the world (literally!), and is a reflection of the fact that the
spacetime we consider has d space-like dimensions and 1 time-like dimension. From
the mathematical point of view, the wave equation is a hyperbolic partial differential
equation, while Poisson’s equation is an elliptic equation, and the diffusion equation
is a parabolic equation.

The particular integral of the inhomogeneous wave equation is given by the integral
representation   ∞
f (r, t) = dd r  dt  G(r, t ; r  , t  ) g(r , t  ),
−∞
d 
where d r is the volume element in d dimensions. The Green function G satisfies the
equation

1 ∂2 
− ∇ 2
G(r, t ; r  , t  ) = δ (d) (r − r  ) δ(t − t  ),
c2 ∂t2
where δ (d) denotes the d-dimensional δ-function. In physical terms, this Green function
represents the signal at time t at the point r, arising when a sharply-pulsed point
source of unit strength is switched on at the point r  at the instant of time t  . We are
interested in the Green function that satisfies natural boundary conditions, i.e., G → 0
44
So that ∇2 stands for the Laplacian in d-dimensional Euclidean space.

144
as |r − r  | → ∞. Further, we seek the causal Green function that vanishes identically
for all t < t  , so that there is no signal anywhere before the source is switched on. This
is the principle of causality. Hence G must satisfy the conditions

G = 0 and ∂G/∂t = 0 for all t < t  , at all points.

As a consequence of the translational invariance in space and time of the differential

equation as well as the boundary and initial conditions, we expect to find that the
Green function has the form

G(r, t ; r  , t  ) ≡ G(r − r  , t − t  ) = θ(t − t  ) K(r − r  , t − t  ).

(In a region of finite extent, in the presence of boundary conditions at finite values of
r, the dependence of G on r and r  cannot be reduced in general to a dependence on
the difference r − r  alone.) The quantity K(r − r  , t − t  ) is sometimes referred to as
the propagator. As you will see, causality imposes an even stronger constraint. Since
the propagation of the signal occurs with a finite speed c, a signal emanating from the
point r  at time t  cannot reach the point r till time t  + |r − r  |/c. This feature will
also emerge automatically in the solution for G.

It is obviously convenient to shift variables from r and t to

R = r−r and τ = t − t  ,

respectively. Then

1 ∂2 
− ∇ 2
G(R , τ ) = δ (d) (R) δ(τ ),
c2 ∂τ 2
where ∇2 now stands for the Laplacian operator with respect to R. We must impose
the conditions G = 0 and ∂G/∂τ = 0 for all τ < 0, and the boundary condition G → 0
as R → ∞. Now express G as a Fourier integral, according to
  ∞
dd k dω i(k·R−ωτ ) 
G(R , τ ) = d
e G(k, ω).
(2π) −∞ 2π

Observe that we use a Fourier transform in both space and time, rather than a Fourier
transform with respect to r and a Laplace transform with respect to t. This is helpful
in imposing the initial conditions on G, as you will see. The inverse relation is of course
  ∞

G(k, ω) = d R d
dτ e−i(k·R−ωτ ) G(R , τ ).
−∞

Further,   ∞
(d) dd k dω i(k·R−ωτ )
δ (R) δ(τ ) = e .
(2π)d −∞ 2π

145
Using these expressions in the partial differential equation satisfied by G(R , τ ), we
get   ∞
dd k dω i(k·R−ωτ )  2 2 2 

e (ω − c k ) G(k, ω) + c = 0.
2
(2π)d −∞ 2π
But the set of functions {ei(k·R−ωτ ) }, where ω and the Cartesian components of k
run over all real values, forms a complete orthonormal basis in the space of integrable
functions of τ and R. Therefore the expression in curly brackets must vanish, for every
value of ω and the components of k. Hence

 ω) = − c2
G(k, , where k 2 = |k|2 = k12 + . . . + kd2 .
(ω 2 − c2 k 2 )

As expected, the introduction of the Fourier transform has converted the partial differ-
ential equation for G into a trivially-solved algebraic equation for its Fourier transform
 ω). Inverting the Fourier transforms, we obtain the formal solution for G(R , τ ),
G(k,
namely,  
dd k ik·R ∞ dω e−iωτ
G(R , τ ) = −c 2
e .
(2π)d −∞ 2π (ω − c k )
2 2 2

We now encounter a familiar difficulty: the formal solution above does not make sense
as it stands! The integral over ω diverges because the integrand has poles at ω = −ck
and ω = ck that lie on the path of integration.

The difficulty is overcome, as usual, by an appropriate i-prescription that also

takes care of the physical requirement of causality. This condition enables us to select
the correct Green function unambiguously. Once the poles are displaced off the real
axis in the ω-plane, the integration over ω can be carried out using contour integration.
Let Ω be a large positive constant. Consider a closed contour C+ [respectively, C− ]
comprising a straight line from −Ω to +Ω along the real axis in the ω-plane, and a
semicircle of radius Ω that takes us back from +Ω to −Ω in the upper [respectively,
lower] half-plane. The limit Ω → ∞ is to be taken after the contour integral is evalu-
ated. Provided the contribution from the semicircle vanishes in the limit Ω → ∞, the
original line integral from −∞ to +∞ over ω is guaranteed to be precisely equal to
the integral over the closed contour.45

Now, for τ < 0, this semicircle must lie in the upper half-plane in ω, because it
is only in this region that the factor e−iωτ in the integrand vanishes exponentially as
45
You will recall that we’ve already used this sort of contour completion more than once—for in-
stance, in the derivation of dispersion relations for the generalized susceptibility in linear response
theory, and of the Green function corresponding to outgoing spherical waves for the Helmholtz oper-
ator.

146
Ω → ∞. The addition of the semicircle to the contour would then simply add a van-
ishing contribution to the original line integral that we want to evaluate. Therefore,
provided no singularities of the integrand lie on the real axis or in the upper half-plane
in ω, the integral over the contour C+ is guaranteed to vanish identically for τ < 0.
But this is precisely what is required by causality: namely, that G(R , τ ) be equal to
0 for all τ < 0.

On the other hand, for τ > 0, we do expect to have a signal that does not vanish
identically. But now the semicircle closing the contour must lie in the lower half-plane,
because it is only then that the factor e−iωτ in the integrand vanishes exponentially
as Ω → ∞; and hence so does the contribution from the semicircle to the integral
over the contour C− . Therefore, provided all the singularities of the integrand are
in the lower half-plane, all our requirements are satisfied. This is ensured by displac-
ing each of the poles of the integrand at ω = −ck and ω = +ck by an infinitesimal
negative imaginary quantity −i where  > 0, and then passing to the limit  → 0
after the integral is evaluated. Equivalently, we may replace ω by ω + i in the de-
nominator of the integrand, and take the limit  → 0 after carrying out the integration.

The causal Green function we seek is therefore given by

 
dd k ik·R ∞ dω e−iωτ
G(R , τ ) = −c lim
2
e ,
(2π)d −∞ 2π (ω + i) − c k

→0 2 2 2

where  is a positive infinitesimal. But, as discussed above,

 ∞  Ω
dω e−iωτ dω e−iωτ
= lim
−∞ 2π (ω + i) − c k Ω→∞ −Ω 2π (ω + i)2 − c2 k 2
2 2 2


dω e−iωτ
= lim .
Ω→∞ C 2π (ω + i)2 − c2 k 2
±

As explained in the foregoing, we must use C+ for τ < 0 and C− for τ > 0.

1. Show that 
dd k sin cτ k ik·R
G(R , τ ) = c θ(τ ) e .
(2π)d k
Note how the step function θ(τ ) required by causality emerges automatically in the
solution for G(R , τ ).

2. Each of the two poles of the integrand in the formal solution for G(R , τ ) the
we started with can be displaced so as to lie either in the upper or lower half-plane.
This leads to four possible ways of making the divergent integral finite. The partic-
ular i-prescription we have used above is tailored to ensure that the correct causal
solution is picked up from among the set of possible solutions. Find the other solutions.

147
 We are now in a position to consider the solutions for different values of d. In order
to make this explicit, I’ll denote the Green function by G(d) (R , τ ) instead of G(R , τ ),
from now on.

3. The solution in (1 + 1) dimensions: The case of one spatial dimension (d = 1)

is somewhat distinct from the others, and simpler too. Note that the symbol k in the
factor (sin cτ k)/k above stands for |k|. When d = 1, therefore, we should remember
to write |k| instead of just k in this factor. Further, k · R is just kX in this case, where
X = x − x  . Therefore
 ∞  ∞
(1) dk sin cτ |k| ikX dk sin cτ k ikX
G (X, τ ) = c θ(τ ) e = c θ(τ ) e .
−∞ 2π |k| −∞ 2π k

It is obvious from this expression that G(1) (X, τ ) is a symmetric function of X, i.e.,

G(1) (−X, τ ) = G(1) (X, τ ).

(Change the variable of integration from k to −k, and the result follows.) As you will
see shortly, G(1) (X, τ ) is actually a function of |X|. Show that

G(1) (X, τ ) = 14 c θ(τ ) [ε(cτ + X) + ε(cτ − X)] ,

where ε(x) = 1 for x > 0, and −1 for x < 0. Simplify this result to get

G(1) (X, τ ) = 12 c θ(τ ) θ(cτ − |X|).

The factor θ(cτ − |X|) ensures that the signal arising from the source point x  at time
t  does not reach any point x until time t  + |x − x  |/c, as required by causality and the
finite speed of signal propagation. The presence of this step function makes the other
step function, θ(τ ), redundant from a physical point of view. However, it is present in
the formal mathematical solution for the quantity G(1) (X, τ ).

Another aspect of the solution is also noteworthy. An observer at x does not re-
ceive a pulsed signal, even though the sender sent out such a signal. (Recall that the
initial disturbance was a Dirac δ-function.) In fact, the signal received at any x persists
thereafter for all time. And it does so without diminishing in strength, a feature that
is unique to the case d = 1.

Let’s return to the general formula for G(R , τ ) in more than one spatial dimension,
i.e., d ≥ 2. The integrand on the right-hand side is a scalar, and so is the volume ele-
ment dd k. Further, the region of integration (namely, all of k-space) is invariant under
rotations of the coordinate axes. Hence G(d) (R , τ ) is a scalar, i.e., it is unchanged un-
der rotations of the spatial coordinate axes about the origin. This remains true for all
integer values of d ≥ 2. As a result of this rotational invariance, G(d) (R , τ ) is actually
a function of R and τ (where R = |R|, as usual). This means that we can choose the

148
orientation of the axes in k-space according to our convenience, without affecting the
result. This fact is of help in evaluating the integral. We’ll write the Green function
as G(d) (R, τ ), henceforth.

4. The solution in (2 + 1) dimensions: Let’s now consider the case d = 2. It is

evidently most convenient to work in plane polar coordinates in k-space, choosing one
of the axes (say, the k1 -axis) along the vector R. Then
 ∞ 
(2) k dk sin cτ k 2π
G (R, τ ) = c θ(τ ) 2
dϕ eikR cos ϕ
(2π) k
0 ∞ 0
c θ(τ )
= dk sin (cτ k) J0 (kR),
2π 0
 2π
on using the representation 0 dϕ eiz cos ϕ = 2πJ0 (z) for the Bessel function (of the first
kind) of order zero. The integral on the right-hand side is a little tricky to evaluate.
As you know, J0 (kR) decays quite slowly, like k −1/2 , as k → ∞. However, it is an
oscillatory function that changes sign. So is the factor sin (cτ k). As a result of the
partial cancellation of positive and negative contributions, it turn out that the definite
integral above has a finite value. (You should be reminded of the case of the Dirichlet
integral.) Here is a quick way to find it. Since sin (cτ k) is the imaginary part of eicτ k ,
and J0 (kR) is real for real values of the argument kR, we have

(2) c θ(τ )  ∞ 
G (R, τ ) = Im dk eicτ k J0 (kR) .
2π 0

But we may regard this integral as the analytic continuation, to s = −icτ , of the
Laplace transform of J0 (kR), which is given by
 ∞
1
dk e−sk J0 (kR) = √ .
0 s2 + R 2
This analytic continuation can be justified properly, but I’ll not digress to do so here.
It is equivalent to putting in a decaying exponential factor (i.e., a regulator) like e−λk
in the integrand, and then passing to the limit λ → 0 after the integral is evaluated.
Show from the result above that
)
 ∞  0 for c2 τ 2 < R2
Im dk eicτ k
J0 (kR) = √
0 1/ c2 τ 2 − R2 for c2 τ 2 > R2 .

A little care is needed in getting the overall sign right in the result. Note that the
Laplace transform, regarded as an analytic function of s, has branch points at s = ±iR,
and a branch cut running between these points. Identify the phases of the analytic
function (s2 + R2 )−1/2 at different points in the complex s-plane, and use this informa-
tion to arrive at the result quoted.

149
 Since we are only interested in non-negative values of τ and R, the condition c2 τ 2 >
R yields a factor θ(cτ − R). We get, finally,
2

c θ(τ ) θ(cτ − R)
G(2) (R, τ ) = √ .
2π c2 τ 2 − R 2
The sharply-pulsed signal emanating from r  at time t  thus reaches any point r only
at time t  + |r − r |/c, in accordance with causality and the finite velocity of propaga-
tion of the disturbance. But once again, the signal received at any field point r is no
longer a sharply pulsed one : it persists for all t > t  + |r − r  |/c, although its strength
slowly decays as t increases, like t−1 at very long times. Thus, both in d = 1 and in
d = 2, there is an after-effect even for a sharply-pulsed initial signal emanating from
the source.

5. The solution in (3 + 1) dimensions: Something entirely different happens in the

most important case of three-dimensional space. We have

(3) d3 k sin cτ k ik·R
G (R, τ ) = c θ(τ ) e .
(2π)3 k
It is immediately clear that we should use spherical polar coordinates (k , θ , ϕ), and
exploit rotational invariance to choose the polar axis in k-space along the vector R.
Show that this leads to
 ∞
c θ(τ ) ' (
(3)
G (R, τ ) = dk cos (cτ − R)k − cos (cτ + R)k .
2(2π)2 R −∞
Each cosine in the integrand can be replaced by the corresponding exponential, since
the contribution from the sine function vanishes by symmetry. Hence show that
θ(τ ) δ(τ − R/c)
G(3) (R, τ ) = .
4πR
Note that the other δ-function, namely, δ(cτ + R), can be dropped, because we are
only concerned with non-negative values of τ and R.

The crucial point about the solution in three-dimensional space is this: If the source
pulse is a δ-function impulse emanating from r  at time t  , the signal at any field point
r is also a δ-function pulse that reaches (and passes) this point at precisely the instant
t  + |r − r  |/c. Hence there is no after-effect that lingers on at r, in stark contrast to
the situation in d = 1 and d = 2. Moreover, the amplitude of the pulse at r drops with
the distance from the source like 1/R, in exactly the way the Coulomb potential falls
off. These features are unique to three-dimensional space.

Another interesting reduction takes place in the case d = 3. Formally, if the limit
c → ∞ is taken in the wave equation, the wave operator reduces to the negative of the

150
Laplacian operator. We might therefore expect the solution for G(3) (R, τ ) to reduce to
the corresponding Green function for −∇2 . And indeed it does so, because the latter
Green function is precisely 1/(4πR) in three spatial dimensions.

Retarded solution of the wave equation: Substituting the expression obtained

above for G(3) (R, τ ) in the formal solution of the inhomogeneous wave equation in
three dimensions yields the particular integral in the absence of boundaries (or natural
boundary conditions in infinite space):

1 d3 r  %  
&
f (r, t) = g(r , t ) ,
4π |r − r  | ret

where
|r − r | 
[g(r , t  )]ret = g r , t −
def.
.
c
This is called the causal or retarded solution of the wave equation.

6. There is an immediate application to electromagnetism in the Lorenz gauge. In

this gauge, the electromagnetic scalar and vector potentials in free space satisfy the
inhomogeneous wave equations

1 ∂2φ ρ(r, t) 1 ∂2A

2 2
− ∇2 φ = and 2 2
− ∇2 A = µ0 j(r, t).
c ∂t 0 c ∂t
The retarded solutions to these equations with natural boundary conditions are there-
fore given by 
1 d3 r  % &
φ(r, t) = ρ(r  , t  ) ret
4π0 |r − r | 

and 
µ0 d3 r  %   &
A(r, t) = j(r , t ) ret .
4π |r − r |
Hence write down the formal solutions for the electric field E(r, t) and magnetic
field B(r, t) arising from arbitrary sources ρ(r, t) and j(r, t), using the relations E =
−∂A/∂t − ∇φ and B = ∇ × A. You must bear in mind that the solutions given
above for φ and A are, of course, only valid in the Lorenz gauge. But the expressions
obtained for the physical fields E and B are gauge-invariant, and hold good in any
gauge.

Remarks on propagation in spatial dimensions d > 3: The results derived above

essentially imply that the basic, linear wave equation permits the propagation of sharp
pulses in three-dimensional space, but not in one- or two-dimensional space. Here are
some comments on what happens in a general number of dimensions d > 3.

151
 Interestingly enough, it turns out that the propagation of sharp signals is possible
in all odd-dimensional spaces with d ≥ 3, while it fails for all even values of d. This is
yet another manifestation of the fundamental differences that exist between Euclidean
spaces of even and odd dimensionalities, respectively. If a sharp δ-function pulse is
emitted at the origin at t = 0, then:

(i) The signal received at any point r is sharply pulsed, arriving and passing on at
time t0 + |r − r  |/c with no after-effect, in spaces of dimension d = 3, 5, . . . .

(ii) In contrast, the signal lingers on for all t > t0 + |r − r  |/c in spaces of dimension
d = 2, 4, . . . . We have already seen an example of such an after-effect in the case d = 2.

(iii) There is, however, one feature that is quite unique to d = 3: this is the only case
in which the original δ-function pulse is transmitted without any distortion, namely,
as a pure δ-function pulse.

There is an elegant and powerful way to solve the general problem. It is based on
the relativistic invariance of both the wave operator and the solution sought, i.e.,
on the fact that these remain invariant under rotations of the coordinate axes as well
as boosts to other inertial frames. (Here we have identified c with the speed of light in
a vacuum). I shall merely write down the final answer here. The causal Green function
turns out to be proportional to a derivative of a δ-function, according to
d(d−3)/2
G (R, τ ) ∝ (d−3)/2 δ(ξ 2 ),
(d)
where ξ = (c2 τ 2 − R2 )1/2 .
dξ
The order of the derivative, 12 (d−3), is an integer when d is an odd number. The Green
function remains a sharply-pulsed quantity in this case, although it is only for d = 3
that you get just a δ-function. For larger odd values of d (= 5, 7, . . . ), the fundamental
solution is given by higher and higher derivatives of the δ-function. These are increas-
ingly singular quantities. When d is an even integer, on the other hand, the solution is
a fractional derivative of a δ-function. Fractional derivatives are nonlocal objects,
defined in terms of suitable transforms (such as the Fourier transform). This is how
the extended nature of the Green function arises in the case of wave propagation in
even-dimensional spaces, leading to the after-effects mentioned earlier.

There is another interesting connection between the Green functions in spaces of

different dimensions. The solution in (d + 2) spatial dimensions is related to that in d
space dimensions by
1 ∂ 2 G(d)
G(d+2) = − .
2πd ∂ξ 2
This relationship shows how the solutions in d = 5, 7, . . . can be generated from the
solution in d = 3, while those in d = 4, 6, . . . can be generated from that in d = 2.

152
Dispersion and nonlinearity: Finally, I mention that there are two important ad-
ditional aspects of wave or signal propagation that can be adjusted so as to modify the
fundamental solution considered here. The first is dispersion, which occurs because
waves of different wavelengths propagate with different speeds in a medium. The cor-
responding dispersion relation (or frequency-wave number relationship) can be quite
complicated. The second aspect is nonlinearity. While the simple wave equation we
have considered here is linear in the signal f (r, t), physical situations often call for non-
linear equations. The interplay between dispersion and nonlinearity can be extremely
intricate and interesting, and a vast variety of new phenomena can arise as a result.
Among these are the so-called solitary waves and propagating solitons which rep-
resent very robust pulsed disturbances. These entities comprise a whole subject in
their own right.

153
14 The rotation group and all that
Rotations of the coordinate axes: Consider n-dimensional Euclidean space, where
n ≥ 2. A rotation of the coordinate axes about the origin of coordinates takes a
general point r to another point r  , such that (i) the origin remains unchanged, and
(ii) the distance between any two points remains unchanged. Therefore each rotation
of the coordinate axes about the origin is a linear, homogeneous transformation of the
Cartesian coordinates. Such a transformation is specified by a (n × n) orthogonal
matrix R, as you’ll see shortly. That is, R satisfies the condition R RT = I, where
the superscript T denotes the transpose, and I is the unit matrix. Here’s one of the
simplest examples of a rotation matrix in 3-dimensional Euclidean space. The matrix
corresponding to a rotation of the coordinate axes about the origin, in the xy-plane,
and through an angle ψ, is given by
 
cos ψ sin ψ 0
R(ψ) = − sin ψ cos ψ 0 .
0 0 1

It is easy to check that R(ψ) is an orthogonal matrix. We could also have said, “a ro-
tation about the z-axis through an angle ψ”, because it is a rotation in 3-dimensional
space. In general, however, the correct way to specify a rotation in any number of
dimensions is to specify the plane in which the rotation takes place, rather than the
axis about which it occurs. This is because no such axis may exist in general, although
in n = 3 it so happens that a unique axis of rotation always exists for every rotation.
Basically, this is because the number of independent, mutually orthogonal axes (= n)
becomes equal to the number of independent, mutually orthogonal planes (= 12 n(n−1))
only for n = 3.

Under a rotation, the components xi of a point r change to the components x i of

the vector r  , given by x i = Rij xj , where the indices run from 1 to n. (Summation
over repeated indices is to be understood.) The definition of any vector a now follows:
The ordered n-tuple a = (a1 , a2 , . . . , an ) is a vector if, under a rotation R of the
coordinate axes, the new components are given by a i = Rij aj . Tensors of rank
2, 3, . . . are sets of quantities that have transformation properties generalizing that for
a vector. For example, tensors of rank 2 and 3 transform like T ij = Rik Rjl Tkl and
S ijk = Ril Rjm Rkn Slmn , respectively. The transformation rule for a tensor of rank 2
is of special interest. It can be written as

T ij = Rik Tkl (RT )lj = Rik Tkl (R−1 )lj = (RT R−1 )ij ,

because the orthogonality condition on R implies that RT = R−1 . But we may also
regard components Tij of a tensor of rank 2 in n dimensions as the elements of a
(n × n) matrix T . Thus the transformation rule can be written in the compact form
T  = R T R−1 . In other words, T  is obtained from T by a similarity transformation

154
involving R.

Rotations of the coordinated axes (equivalently, the matrices representing them) in

n dimensions form the rotation group: Two rotations in succession are equivalent
to a single ‘resultant’ rotation; no rotation at all corresponds to the identity element
of the group; and for every rotation there is an ‘inverse’ rotation that takes us back to
the original orientation of the axes.

Orthogonality of rotation matrices: It is now very easy to see why an arbitrary

rotation of the coordinate axes about the origin is specified by an orthogonal matrix.
Under such a rotation, the distance from the origin to any point remains unchanged.
Therefore r  2 = r 2 , or
x i x i = xj xj = δjk xj xk .
But we also have x i = Rij xj . Therefore

x i x i = Rij xj Rik xk = (RT )ji Rik xj xk = (RT R)jk xj xk .

The two expressions for r  2 must be equal to each other for every point in space.
Therefore we must have

(RT R)jk = δjk , or RT R = I.

For finite-dimensional square matrices, the left and right inverses are the same. Hence
RT R = I ⇒ R RT = I.

Proper and improper rotations: As we have just seen, a rotation of the coordi-
nate axes about the origin is specified (in n-dimensional Euclidean space) by an (n×n)
orthogonal matrix. These matrices form a group, denoted by O(n). Let R be a rota-
tion matrix, i.e., a matrix whose elements tell you what linear combinations of the old
coordinates yield the new coordinates. The orthogonality condition R RT = I on the
matrix R implies that (det R)2 = 1. Therefore det R = ±1.

Rotations for which det R = +1 are called continuous or proper rotations. They
are obtainable “continuously from the identity transformation”—that is, they can be
built up by a succession of infinitesimal rotations, starting from the identity transfor-
mation (or no rotation at all). For this reason, the set of proper rotations is called the
connected component of the rotation group. Proper rotations constitute a group
of their own, denoted by SO(n). (The ‘S’ stands for special, which means ‘unimodu-
lar’ or ‘with unit determinant’, in this context.) This is a subgroup of O(n). Proper
rotations preserve the orientation or handedness of the coordinate system. That is,
after a proper rotation, a right-handed coordinate system remains right-handed, and
a left-handed coordinate system remains left-handed.

155
 In contrast, transformations with det R = −1 are called discontinuous or improper
rotations. They cannot be built up continuously from the identity transformation:
in general, they involve proper rotations together with reflections, such that a right-
handed coordinate system transforms to a left-handed one or vice versa. Examples of
such orientation-reversing transformations in n dimensions are:
(i) Reflection about any plane in space. (The plane need not be one of the planes
normal to the Cartesian axes.) In three dimensions, for example, a reflection
about the yz-plane corresponds to a transformation under which x → −x, y →
y, z → z.

(ii) The parity transformation r → −r, i.e., xi → −xi , , provided n is odd.46 In

three dimensions, for example, the transformation x → −x, y → −y, z → −z is
an improper rotation, while reversing the signs of any two of the three coordinates
is actually a proper transformation: The determinant of the corresponding matrix
remains equal to +1.
Note also that improper rotations cannot form a subgroup of their own, because they
do not include the identity transformation. (Moreover, the product of two matrices,
each with determinant equal to −1, is a matrix with determinant equal to +1.) The
group O(n) is thus made up of 2 disjoint sets: (n × n) orthogonal matrices with de-
terminant equal to +1, and that constitute the group SO(n); and (n × n) orthogonal
matrices with determinant equal to −1. These are obtained by multiplying the ele-
ments of SO(n) by the matrix corresponding to a parity or reflection transformation
with determinant = −1.

The most important aspect of rotations in n ≥ 3 dimensions is the following: Suc-

cessive rotations do not commute with each other, unless they are rotations in the
same plane. In other words, the net result of two successive rotations depends on the
order in which the two are carried out. This non-commutativity is crucial to the
understanding of rotations. It has truly profound consequences for the way the physi-
cal universe is. The rotation group in every dimension n ≥ 3 is a noncommutative or
non-abelian group.

Generators of infinitesimal rotations in 3 dimensions : Let’s turn to the im-

portant special case of rotations in 3-dimensional space. There are many ways of
parametrizing rotations in three dimensions. A very useful way in applications is via
three Euler angles. These are used, for instance, in studying the rotational motion
of a rigid body. Any given orientation of the triad of coordinate axes may be reached
from an initial reference orientation by a succession of three rotations about a pre-
scribed set of three different axes.47 Our present objective is somewhat different: we
46
When n is even, it is obvious that r → −r is a proper rotation, with det R = +1.
47
There are as many as twelve different conventions for defining Euler angles. I will not digress into
these here.

156
are interested in the rotation matrices per se and in their algebraic properties. We
are also interested in finding the explicit transformation formula of a vector under an
arbitrary rotation of the coordinate axes. As you’ll see below, there’s quite an easy
way to arrive at the exact answer without any tedious algebra.

A convenient way of parametrizing any given rotation is to specify the axis of

rotation, i.e., the direction in space about which the triad of Cartesian coordinate axes
is rotated, and the amount or angle of rotation about this axis. We may therefore
denote a general rotation matrix by R(n, ψ), where n is the unit vector along the axis
of rotation, and ψ is the angle of rotation about this axis. All you need in order to
write down the rotation matrices corresponding to rotations about the three Cartesian
axes is to recall a result from elementary coordinate geometry. A rotation by an angle
ψ of the coordinate axes about the origin in the xy-plane gives the new coordinates
x  = x cos ψ+y sin ψ and y  = −x sin ψ+y cos ψ. The z coordinate is left unchanged.
Hence, as we have already written down,
 
cos ψ sin ψ 0
R(ez , ψ) = − sin ψ cos ψ 0 .
0 0 1

By cyclic permutation of xyz, we may write down the other two matrices
   
1 0 0 cos ψ 0 − sin ψ
R(ex , ψ) = 0 cos ψ sin ψ  and R(ey , ψ) =  0 1 0 .
0 − sin ψ cos ψ sin ψ 0 cos ψ

It is easily checked that each of these matrices is orthogonal, and has a determinant
equal to +1. Hence each of them can be built up from the identity matrix by a succes-
sion of infinitesimal rotations about the axis concerned. We can work backwards from
the matrices written down above to see how this is done.

Consider, for definiteness, R(ez , ψ). We could implement such a rotation by n

successive rotations about the z-axis, each through an infinitesimal angle δψ, such that
n δψ = ψ. The matrix R(ez , δψ) is easily written down: use the fact that sin δψ  δψ
and cos δψ  1 to first order in δψ. Separating out the (3 × 3) unit matrix, which
corresponds to the identity transformation (or zero rotation), we get
 
0 −i 0
R(ez , δψ) = I + i (δψ) J3 , where J3 =  i 0 0 .
0 0 0

The parameter δψ has been factored out in the expression above. This makes the
elements of the matrix J3 pure numbers that are independent of the angle of rotation.
The reason for separating out the factor i in the definition of J3 is to ensure that J3 is

157
a hermitian matrix.48 The finite-angle rotation matrix R(ez , ψ) is then given by

R(ez , ψ) = R(ez , δψ) · · · R(ez , δψ) = [R(ez , δψ)]n = [I + i (δψ) J3 ]n .

  
n factors

Setting δψ = ψ/n and passing to the limit n → ∞,

i ψ J3  n
R(ez , ψ) = lim I + = ei ψ J3 .
n→∞ n
Repeat the procedure above for the matrices R(ex , ψ) and R(ey , ψ), to get

R(ex , ψ) = ei ψ J1 and R(ey , ψ) = ei ψ J2 ,

where    
0 0 0 0 0 i
J1 = 0 0 −i and J2 =  0 0 0 .
0 i 0 −i 0 0
The form R(ez , δψ) = I + i (δψ) J3 makes it quite clear why the matrix J3 is called
the generator of an infinitesimal rotation about the z-axis (i.e., a rotation about the
origin, in the xy-plane). Similarly, J1 and J2 may be identified as the generators of
infinitesimal rotations about the x and y axes (i.e., about the origin, in the yz-plane
and zx-plane), respectively.

The matrix corresponding to rotation by a finite angle is obtained by exponentiating

the corresponding generator. This is a general feature for groups of transformations
(more generally, Lie groups). The hermitian matrices J1 , J2 and J3 satisfy the com-
mutation relations
[Jk , Jl ] = iklm Jm ,
where klm is the Levi-Civita totally antisymmetric symbol in three dimensions. These
commutation relations comprise the Lie algebra so(3) of the rotation group SO(3).
In general, the Lie algebra of the infinitesimal generators of a Lie group determine the
‘structure’ of the group in the neighborhood of the identity element, and, by continuity,
elsewhere too (at least in the part of the group that is continuously connected to the
identity element).
• In this sense, the Lie algebra corresponding to a Lie group contains all the infor-
mation about the group except the global topological properties of its parameter
space.

1. It is instructive to check out the statements made in the foregoing.

48
This is the convention generally used in physics, because quantities like J3 will be associated with
operators that, in turn, represent physical observables which must have real values. Recall that the
eigenvalues of hermitian matrices are guaranteed to be real.

158
 (a) Work through the steps outlined above.
(b) Using the expressions given above for the matrices Jk , directly calculate the expo-
nentials eiψ Jk for k = 1, 2 and 3. (That is, sum the corresponding exponential se-
ries.) Verify that you recover the finite-angle rotation matrices R(ex , ψ), R(ey , ψ)
and R(ez , ψ).
(c) Verify that the generators Jk satisfy the commutation relations [Jk , Jl ] = iklm Jm .

2. The general rotation matrix in 3 dimensions: We’ve seen that the matrices
corresponding to rotations about the three Cartesian axes can be written as expo-
nentials of the corresponding infinitesimal generators. What about a general rotation
R(n, ψ) about an axis n pointing in an arbitrary direction in space?

It turns out49 that the three generators of infinitesimal rotations in 3-dimensional

space, (J1 , J2 , J3 ), themselves transform under rotations like the components of a
vector. It is therefore natural to denote the triplet by the vector symbol J. Then, if
the components of the direction vector n are given by (n1 , n2 , n3 ), we are guaranteed
that
R(n, ψ) = ei (J1 n1 +J2 n2 +J3 n3 ) ψ ≡ ei (J·n) ψ .
Since the different matrices Jk do not commute with each other, however, the right-
hand side of this equation is not equal to the product of exponentials, i.e.,

ei (J1 n1 +J2 n2 +J3 n3 ) ψ

= eiJ1 n1 ψ eiJ2 n2 ψ eiJ3 n3 ψ .

In spite of this problem, it turns out to be possible to compute the exponential of the
(3 × 3) matrix i (J · n) ψ exactly, and in closed form. Here’s how this is done.

We want to find eM ψ , where M = i (J · n). Using the definitions of the matrices Jk

above, we find  
0 n3 −n2
M = i (J · n) = −n3 0 n1  .
n2 −n1 0
In order to find the powers of M explicitly, it is helpful to note that the elements of
M are given by
Mij = ijk nk .
It follows that

(M 2 )ij = ni nj − δij , and hence (M 3 )ij = −ijk nk = −Mij .

The fact that M 3 = −M immediately enables us to simplify the exponential eM ψ . The

result is just a linear combination of the three matrices I, M and M 2 . The final answer
49
As you’ll see subsequently.

159
for the matrix elements of the rotation matrix R(n, ψ) is both simple and elegant. It
reads
Rij (n, ψ) = δij cos ψ + ni nj (1 − cos ψ) + ijk nk sin ψ.
Even more explicitly, if the spherical polar angles of the unit vector n are given by θ
and ϕ, we have

n1 = sin θ cos ϕ, n2 = sin θ sin ϕ, n3 = cos θ.

Using these expressions, you can write down the complete rotation matrix for an ar-
bitrary rotation R(n, ψ). Work out the steps outlined above, and find the explicit
expression for the rotation matrix R(n, ψ).

3. Find the eigenvalues and eigenvectors of the general rotation matrix R(n, ψ).

The finite rotation formula for a vector: Once we have R(n, ψ) explicitly, it is
straightforward to apply it to an arbitrary position vector r, to obtain

x i = Rij xj = xi cos ψ + ni xj nj (1 − cos ψ) + ijk xj nk sin ψ.

Expressing this formula back in terms of the vectors r and n helps us understand it in
physical terms:

r  = (cos ψ) r + (1 − cos ψ) (r · n) n + (sin ψ) (r × n).

It follows at once that, by its very definition, any vector a transforms under a general
rotation R(n, ψ) according to
R(n,ψ)
a −−−−→ a  = (cos ψ) a + (1 − cos ψ) (a · n) n + (sin ψ) (a × n).

This is sometimes called the finite rotation formula for a vector. a  has a com-
ponent along the original vector a, a component along the axis of rotation n, and a
component in the direction normal to the plane formed by a and n.

Relation between the groups SO(3) and SU(2): Orthogonal matrices and uni-
tary matrices are foremost among the sets of special kinds of matrices that form groups
under matrix multiplication, and that are of great importance in physics. All (n × n)
orthogonal matrices with real elements form the orthogonal group O(n), while all
(n × n) unitary matrices with complex elements form the unitary group U(n). When
the determinants of the corresponding matrices are further restricted to the value +1,
we have the special groups SO(n) and SU(n). It turns out that there is a very im-
portant connection between the rotation group SO(3) and the group of unimodular,
unitary (2 × 2) matrices, denoted by SU(2). This relationship has profound physical
consequences.

160
 Let’s begin by observing that the three Pauli matrices σ1 , σ2 , σ3 (abbreviated as
σ) satisfy the same Lie algebra as the generators of rotations: if we set

Jk = 12 σk (where k = 1, 2, 3), then [Jk , Jl ] = iklm Jm .

Therefore the (2 × 2) matrix

U(n, ψ) = ei(σ·n)ψ/2
also represents a rotation of the coordinate axes in 3-dimensional space about the
direction n, and through an angle ψ, just as the (3 × 3) orthogonal matrix R(n, ψ)
does. But what does the matrix U matrix act on? Clearly, we need a representation of a
point r in 3-dimensional space that is compatible with the (2×2) matrix representation
of a rotation of the coordinate axes. This is obtained by noting that the point r =
(x1 , x2 , x3 ) in 3-dimensional space can also be represented as a (2 × 2) matrix given
by  
x3 x1 − ix2
r · σ ≡ x1 σ1 + x2 σ2 + x3 σ3 = .
x1 + ix2 −x3
Similarly, any vector a = (a1 , a2 , a3 ) in 3-dimensional space can also be represented
as a (2 × 2) matrix given by
 
a3 a1 − ia2
a · σ ≡ a1 σ1 + a2 σ2 + a3 σ3 = .
a1 + ia2 −a3

Note that the matix (a · σ) uniquely determines the components (a1 , a2 , a3 ), and vice
versa. The corresponding transformation rule under a rotation, when we choose to
represent vectors by (2 × 2) matrices as above, is given by

(r  · σ) = U(n, ψ) (r · σ) U −1 (n, ψ),

and more generally,

(a  · σ) = U(n, ψ) (a · σ) U −1 (n, ψ).

4. The matrix U(n, ψ) can be determined explicitly.

(a) The square of any Pauli matrix is the unit matrix, i.e., σj2 = I. Moreover, any
two different Pauli matrices anticommute with each other, i.e., σj σk + σk σj =
0, j
= k. Using these properties, show that
 
U(n, ψ) = exp 12 i σ · n ψ = I cos 12 ψ + i (n · σ) sin 12 ψ
1 2
cos 12 ψ + i sin 12 ψ cos θ i sin 12 ψ sin θ e−iϕ
= ,
i sin 12 ψ sin θ eiϕ cos 12 ψ − i sin 12 ψ cos θ

where θ and ϕ, respectively, denote the polar and azimuthal angles specifying
the unit vector n.

161
 (b) Hence verify that U is unitary and unimodular, i.e., UU † = I = U † U, and
det U = +1.

In other words, U(n, ψ) is an element of the special unitary group SU(2). Now, it is
obvious that replacing the matrix U(n, ψ) by −U(n, ψ) does not alter these properties.
Moreover, given any vector a, a rotation transformation made using −U in the place
of U leads to exactly the same a  .

• Hence there are two distinct elements U(n, ψ) and −U(n, ψ) of SU(2), differing
only by an overall sign, corresponding to every rotation matrix R(n, ψ) ∈ SO(3).

• This is the famous 2-to-1 homomorphism from SU(2) to SO(3).

Both the (2 × 2) unit matrix I and its negative −I are elements of SU(2). By them-
selves, these two matrices form a group, namely, the cyclic group of order 2, denoted by
Z2 .50 This group (Z2 ) is said to be the center of SU(2): its elements (I and −I) are
the only elements of SU(2) with which all the elements of the group commute. I and
−I constitute the kernel of the homomorphism from SU(2) to SO(3) : that is, they
are the only elements of SU(2) that map into the unit element of SO(3), the (3 × 3)
unit matrix. Thus there is an isomorphism between the quotient group SU(2)/Z2
and SO(3), written symbolically as

SU(2)/Z2  SO(3).

We’ll resume our discussion of other group theoretical properties after completing
an important bit of unfinished business: namely, the proof that the three generators
of rotations in 3-dimensional space themselves transform like a vector.

Rotation generators in 3 dimensions transform like a vector: We have used

the symbol J to denote the triplet (J1 , J2 , J3 ) of the generators SO(3), implying that
these quantities themselves transform like the components of a vector under rotations
of the coordinate axes. We must now establish that this is indeed so.

Recall that the actual representation of each Jk depends on what it is required to act
on. You’ve already come across two such representations: the defining representation
in terms of (3 × 3) matrices, and a (2 × 2) matrix representation in which Jk = 12 σk .
But the generators Jk operators can, and do, have an infinite number of other repre-
sentations, including infinite-dimensional ones—for example, when they act on state
vectors in infinite-dimensional Hilbert spaces in quantum mechanics. A rotation R of
the coordinate axes induces a transformation in the Hilbert space. This is a unitary
50
Z2 is also the group of integers under addition modulo 2: all the even integers are represented by
the identity element of Z2 , while all the odd integers are represented by the only other element of the
group.

162
transformation given by U(R) = ei (J·n) ψ , where the generators Jk (k = 1, 2, 3) have
the representation appropriate to the Hilbert space. What we need to do is to establish
that the triplet (J1 , J2 , J3 ) transforms like a vector under rotations of the coordinate
axes, independent of any particular representation for the generators. In effect, this
means that the only input we can use is the algebra satisfied by the three operators,
namely, [Jk , Jl ] = i klm Jm .

Let’s consider the general case of a rotation about an arbitrary axis n through an
angle ψ. (All special cases can then be read off from it.) The task is to show that J
satisfies the finite rotation formula, which is
e−i (J·n) ψ Ji ei (J·n) ψ = (cos ψ) Ji + (1 − cos ψ) (Jj nj ) ni + (sin ψ) ijk Jj nk .
Observe that the transformation rule for the operator J, given by the left-hand side of
this equation, is not just J  = R(n, ψ) J. This is because J is not an ordinary vector,
but an operator-valued vector. I reiterate that the components of J are not necessarily
(3 × 3) matrices, nor do we care what the actual representation is. In vector form, we
need to show that
e−i (J·n) ψ J ei (J·n) ψ = (cos ψ) J + (1 − cos ψ) (J · n) n + (sin ψ) (J × n).
The right-hand side of this equation has precisely the form of the finite rotation for-
mula for the coordinate r of a point in three-dimensional space, with J replacing r.
Therefore, once the equation above is proved, we may assert that J itself transforms
like a vector under rotations.

What you will need for the purpose is a remarkable and extremely useful operator
identity called Hadamard’s Lemma. This identity expresses the operator eλA B e−λA ,
where A and B are (in general, non-commuting) operators acting on some linear vector
space and λ is a scalar constant (real or complex), as an infinite sum involving multiple
nested commutators of A with [A , B]. The formula is

−λA λ2 % & λ3  % &

λA
e Be = B + λ [A , B] + A , [A , B] + A , A , [A , B] + · · ·
2! 3!
If any of the multiple commutators on the right-hand side should happen to be zero,
the series terminates. The derivation of this identity is quite straightforward, and will
be dealt with shortly.

5. Establish the transformation rule for J given above, using Hadamard’s Lemma.
Here’s an outline of the steps involved. Start with the identifications
λ = −iψ, A = J · n = Jk nk , B = Ji .
For ease of notation, write the r-fold multiple commutator in Hadamard’s Lemma as
Cr . Then
[A , Cr ] = Cr+1 , where C1 = [A , B] and r = 1, 2, . . . .

163
Using the commutator algebra [Jk , Jl ] = i klm Jm , it is straightforward to show that

C1 = i ilk Jl nk and C2 = [A , C1 ] = Ji − ni nk Jk .

Further,

C3 = [A , C2 ] = i ilk Jl nk = C1 , and hence C4 = C2 , C5 = C1 , . . . .

Thus, C2r+1 = C1 and C2r = C2 . The infinite series in Hadamard’s lemma can now be
summed easily. The transformation formula for J follows upon simplification.

There’s a small subtlety involved in the left-hand side of the transformation for-
mula. The factor ei (J·n) ψ , which we have identified with R(n, ψ), appears on the right
of Ji , while its adjoint e−i (J·n) ψ appears on its left. You might have expected the
reverse, based perhaps on the transformation rule for a second-rank tensor. But the
order given here is correct. It is related to the fact that J itself is an operator (e.g.,
in a Hilbert space), and this is the transformation rule for an operator, as opposed to
that for a state vector.

6. Derivation of Hadamard’s Lemma: Show that, if A and B are operators acting

on some linear vector space, and λ is a scalar parameter,

λ2 % & λ3  % &
eλA B e−λA = B + λ [A , B] + A , [A , B] + A , A , [A , B] + · · · .
2! 3!
Here’s how the formula is derived. Define the operator-valued function the parameter
λ according to
F (λ) = eλA B e−λA .
Differentiate both sides with respect to λ, to get the differential equation satisfied by
F (λ):
F  (λ) = A F (λ) − F (λ) A.
Note the ordering of the operators. A and F do not commute with each other, in
general. Now, F (λ) is a regular function of λ that can be expanded in a Taylor series
in powers of λ, with coefficients that are operators, of course. Thus

λ2 
F (λ) = F (0) + λ F  (0) + F (0) + · · · .
2!
The existence of such a power series expansion for F (λ) implies the existence of an
analogous series for its derivative F  (λ) as well. This assertion follows from the fact
that F (λ) is an analytic function of the complex variable λ—in this instance, for all
finite values of |λ|. Hence

λ2 
F  (λ) = F  (0) + λ F  (0) + F (0) + · · · .
2!
164
Insert these series for F (λ) and F  (λ) in the differential equation, and equate the coef-
ficients of like powers of λ. Note that F (0) = B. Determine the successive derivatives
F (n) (0) recursively, to arrive at Hadamard’s formula. This formula is the starting point
for the derivation of a number of extremely useful and important operator identities.

7. The general form of the elements of U(2) and SU(2): Let’s return to our
discussion of group theoretical aspects. You have seen that there is a homomorphism
between the rotation group SO(3) and the special unitary group SU(2). This is a
convenient place to determine the general forms of the elements of the unitary group
U(2) and the special unitary group SU(2).

Consider an arbitrary (2 × 2) matrix

   ∗ ∗
α β † α γ
M= , so that M = ,
γ δ β ∗ δ∗
where α, β, γ and δ are complex numbers. Hence 8 real parameters are required to
specify such a matrix. Show that, if the condition that M be unitary (i.e., MM † = I)
is imposed, the matrix must be of the form
 
α β
M= , where |α|2 + |β|2 = 1,
−eiθ β ∗ eiθ α∗
and θ is any real number. This is the form of a general element of the unitary group
U(2). It is specified by 4 independent real parameters, namely: θ, and 3 out of the 4
numbers α1 , α2 , β1 and β2 , where α1 + iα2 = α, β1 + iβ2 = β, and

α12 + α22 + β12 + β22 = 1.

If the further condition det M = +1 is imposed, then θ must be 0 or an integer multiple

of 2π; and M must then be of the form
 
α β
M= , where |α|2 + |β|2 = 1.
−β ∗ α∗
This is the form of a general element of SU(2). The four real parameters occurring in
it satisfy a single constraint, reducing the number of independent parameters to three.

More generally, an (n × n) matrix with complex elements has 2n2 independent real
parameters. Requiring that the matrix be unitary reduces this number to n2 . Thus
U(n) is an n2 -parameter group. The determinant of any element of U(n) is a complex
number of unit modulus, i.e., a number like eiθ . If, further, we require that the matri-
ces be unimodular, the number of independent real parameters is reduced further to
n2 − 1. Such matrices comprise the special or unimodular unitary group SU(n), which
is a subgroup of U(n).

165
The parameter spaces of SU(2) and SO(3): The parameter space of SU(2) follows
immediately: it is just the ‘surface’ of the unit (hyper)sphere
α12 + α22 + β12 + β22 = 1.
This is the 3-sphere S3 , which is simply connected: that is, any closed path in it can
be continuously deformed and shrunk to a point without leaving the space.

In contrast, the parameter space of SO(3) is more complicated. As we’ve seen, a

rotation in 3-space can be parametrized by a unit vector n and an angle of rotation
ψ. The unit vector n is specified a polar angle 0 ≤ θ ≤ π and an azimuthal angle
0 ≤ ϕ < 2π. You might expect, at first sight, that the angle ψ can run from 0 to
2π. But it is sufficient to let it take values in the range 0 ≤ ψ ≤ π, because of a
simple but profound fact: rotation about the axis n through an angle π leads to the
same end result as a rotation about the oppositely-directed axis −n through an an-
gle π, as is easily verified. This is ‘a fact of life’ in three-dimensional space! As a
result, the parameter space spanned by θ , ϕ and ψ can be represented by a solid (or
3-dimensional) sphere of radius π (rather than 2π). The direction vector of any point
in this sphere specifies the direction n of the corresponding rotation, while the distance
of that point from the origin O represents the amount of rotation, ψ. The origin itself
corresponds to the identity transformation, i.e., to the case of no rotation at all. The
noteworthy point is that every point on the surface of this sphere is mathematically
identical to its antipodal point on the surface. (You may imagine that every pair of
antipodal points is connected by an invisible string of zero length!) This identification
ensures that a rotation through π about any axis n is the same as a rotation through
π about the oppositely directed axis, −n. (Unfortunately, 3-dimensional Euclidean
space is insufficient to embed this parameter space, precisely because it cannot exhibit
the mathematical fact that every pair of antipodal points on the surface of the solid
sphere is really a single point. That’s why we can’t display an actual real-life model of
this space. The space is termed the real projective space RP3 .)

As a consequence of this identification of antipodal points, the parameter space be-

comes doubly-connected. This means that there are two distinct classes of oriented
closed paths in such a space:
(i) Closed paths that can be shrunk continuously to a point (without leaving the
space).
(ii) Closed paths that can be shrunk to a point continuously only when traversed
twice.
The existence of these two classes of closed paths leads, respectively, to the so-called
single-valued or tensor representations of the rotation group, and the double-valued
or spinor representations of the group. These correspond, in the language of quan-
tum mechanics, to integral and half-odd-integral values of the angular momentum

166
quantum number j. Tensors return to their original values under a 2π rotation of the
coordinate axes about any direction, while spinors change sign under such a rotation.
A rotation of 4π is needed to return a spinor to its original value.

The universal covering group of a Lie group: It turns out that every multiply-
connected Lie group G (that is, a Lie group with a parameter space that is not simply-
connected) has a unique universal covering group G  whose parameter space is
simply connected. There is a homomorphism or many-to-one mapping from G  to G.

• The Lie algebras of a Lie group and its covering group are, however, identical.

This means that the two groups are essentially identical in local neighborhoods (in
parameter space), although they may be very different from each other globally. For
one thing, the topologies of the parameter spaces are quite different from each other
(their connectivities are different). In the case of the rotation group SO(3), the param-
eter space (or group manifold) is doubly connected. The universal covering group
of SO(3) is the special unitary group SU(2). The latter is simply connected: its pa-
rameter space is the 3-sphere S3 . And, as I have already pointed out, there is a 2-to-1
homomorphism from SU(2) to SO(3): there are two matrices in SU(2), differing by
a sign, that correspond to each matrix in SO(3).

The group SO(2) and its covering group: These ideas are even more easily illus-
trated in the case of the group SO(2) of rotations 2-dimensional space. The parameter
space of SO(2) is a single angle modulo 2π. In other words, it is the 1-sphere S1 (the
circumference of a circle, in plain language). This space has an infinite number of
distinct classes of closed paths, labelled by the number of times the (directed) path
winds completely around the circle. When S1 is ‘rolled out’ on a line, the latter covers
the circle an infinite number of times, the copies being labeled by the elements of the
additive group of integers, Z. The universal covering group of SO(2) is therefore the
real number line R (which is also a Lie group!). We have the isomorphism

R/Z  SO(2),

so that SO(2) is also a quotient group. Note also that SO(2) is equivalent to U(1),
the multiplicative group of complex numbers of unit modulus (i.e., complex numbers
of the form eiα , where α is real).

The groups SO(n) and Spin (n): What about the rotation group SO(n) for n =
4, 5, . . . ? It turns out that every one of these groups is also doubly connected, just as
SO(3) is. The universal covering group of SO(n) for n ≥ 3 is called Spin (n). There is
again a 2-to-1 homomorphism from Spin (n) to SO(n), and SO(n) is a quotient group
according to
Spin (n)/Z2  SO(n), n ≥ 3.

167
We’ve seen that Spin (3)  SU(2). For higher values of n the covering group does not
reduce to anything so readily identifiable, except for the case of SO(6). In that case
it turns out that Spin (6)  SU(4). These are the only two cases when Spin (n) is a
special unitary group.

Parameter spaces of U(n) and SU(n): For completeness, let me touch upon the
parameter spaces of the unitary group U(n) and the special unitary group SU(n),
where n ≥ 2.51 The parameter space of SU(n) is simply connected. U(n) is connected,
but not simply connected. The determinant of each member of U(n) is a complex
number of unit modulus. U(1) is a subgroup of U(n): matrices of the form eiα I
are included among the elements of U(n), and these matrices form a subgroup that
is isomorphic to U(1). The presence of this U(1) subgroup, whose parameter space
is the infinitely connected 1-sphere or circle S1 , makes the parameter space of U(n)
also infinitely connected. In technical terms, U(n) is the semidirect product of its
normal subgroup52 U(1) and its subgroup SU(n), and is written as

U(n)  U(1)  SU(n).

Conversely, the special unitary group SU(n) is just the unitary group U(n) ‘modulo’
U(1), i.e., it is a quotient group according to

U(n)/U(1)  SU(n).

The universal covering group of U(n) is R ⊗ SU(n), which you can understand heuris-
tically because R is the universal covering group of U(1), and SU(n) is already simply-
connected.

A bit about the first homotopy group of a space: The connectivity of a space
V is formally determined by finding the different equivalence classes of directed
closed paths (or loops) in V. (Two such paths are equivalent if either of them can
be continuously and smoothly deformed into the other without leaving the space and
without cutting open the closed path.) More formally, each closed path is a map
S1 → V. The equivalence classes of such maps form a group, called the fundamental
group or the first homotopy group of V, denoted by π1 (V). If every possible closed
path in V can be shrunk continuously to a point, then π1 (V) reduces to the trivial group
with one element (this is often written as π1 (V) = 0), and V is a simply-connected
space. The converse is also true. Thus

V is a simply-connected space ⇐⇒ π1 (V) = 0.

51
Recall that U (1) is isomorphic to the group of complex numbers {eiα }, where α is real, under
multiplication. It should be obvious that the group SU (1) is the trivial group with just one element
(the unit element).
52
H is a normal subgroup of a group G if the following condition is satisfied: for any element h ∈ H
and any g ∈ G, the element ghg −1 ∈ H.

168
 The fundamental group of the 1-sphere S1 itself is the set of equivalence classes
of maps of S1 to S1 . This is π1 (S1 )  Z, the group of integers under addition. This
follows from the fact that the class of a directed closed path on S1 is labeled by the
winding number of the path (the number of times it goes around S1 completely). It is
intuitively clear that π1 (S2 ) = 0.53 It can be shown rigorously that π1 (Sn ) = 0 for all
n ≥ 2. As you might expect, since the 2-torus T2 is equivalent to the direct product
space S1 ⊗ S1 , we have π1 (T2 ) = Z ⊗ Z. This generalizes in an obvious fashion to the
n-torus Tn = S  ⊗ ·
1
· · ⊗ S1 .
n factors

The fundamental groups of various group manifolds (and of other spaces as well,
of course) are of great interest. Based on what has been mentioned in the foregoing,
here are some of the results concerned.54

π1 (Rn ) = 0, n ≥ 1.
π1 (S1 ) = π1 (T1 ) = Z.
π1 (Sn ) = 0, n ≥ 2.
π1 (Tn ) = π1 (S1 ⊗ · · · ⊗ S1 ) = Z ⊗ · · · ⊗ Z.
π1 (SO(2)) = π1 (U(1)) = π1 (S1 ) = Z.
π1 (Spin (n)) = 0, n ≥ 3.
π1 (SO(n)) = π1 (Spin (n)/Z2 ) = Z2 , n ≥ 3.
π1 (SU(2)) = π1 (S3 ) = 0.
π1 (SU(n)) = 0.
π1 (U(n)) = π1 (U(1)  SU(n)) = Z.

53
A fact that is picturesquely expressed by the statement, “You can’t lasso a basketball!”
54
When we write π1 (G) where G is a Lie group, we mean the fundamental group of the parameter
space of G.

169
15 QUIZ 2
I. Are the statements in quotation marks true or false?

1. “The function sin (1/z) does not have a Taylor series expansion in the neighbor-
hood of z = 0.”


∞
2. A function f (z) is defined by the power series z 2n+1 /[n! (n + 1)!] about the
n=0
origin.

“f (z) is an entire function of z.”

3. “The only singularity of 1/Γ(z) is a simple pole at z = 0.”


∞
4. “The Mittag-Leffler expansion of the gamma function is given by Γ(z) = (−1)n /[(z+
n=0
n) n!].”

5. Let f (z) = z + z 3 + z 9 + z 27 + · · · ad infinitum, for |z| < 1.

“f (z) cannot be analytically continued outside the unit circle.”


∞
6. “The power series (ln n) z n /n converges at all points on its circle of conver-
n=1
gence.”

7. “The function f (z) = 1/(ez − 1) is a meromorphic function of z.”

8. “Dispersion relations for the real and imaginary parts of a generalized suscepti-
bility χ(ω) can be derived only if the corresponding response function φ(t) decays
to zero faster than any negative power of t, as t → ∞.”

9. “The derivative of the gamma function, Γ  (z), has zero residue at each of its
poles.”

170
10. “The Legendre function of the second kind, Qν (z), has branch points in the z-
plane even when ν is a positive integer.”

11. “The Laplace transform of the function f (t) = cosh πt has no singularities in the
region Re s > π.”

 d2 d 
12. Bessel’s differential equation is z 2 2 + z + (z 2 − ν 2 ) f (z) = 0, where ν is
dz dz
a parameter.

“If φν (z) is any solution of this equation, then φ−ν (z) must be equal to φν (z),
apart from a possible multiplicative constant.”

13. “The group Möb (2, C) of Möbius transformations of the complex plane has con-
tinuous subgroups, but no discrete subgroups.”

14. “The group Möb (2, C) of Möbius transformations of the complex plane is iso-
morphic to the group SO(3, 1) of homogeneous proper Lorentz transformstions
in (3 + 1)-dimensional spacetime.”

15. “The Riemann surface of the function f (z) = z 1/2 (z − 1)−1/3 has 6 sheets.”

16. “It is possible to find a contour integral representation of the beta function
B(z, w) that is valid for all complex values of both z and w.”

17. “The Riemann zeta function ζ(z) cannot be continued analytically to the left of
the line Re z = 12 , because it has an infinite number of zeroes on that line.”

18. “The Fourier transform operator in L2 (−∞ , ∞) has a finite number of eigenval-
ues, each of which is infinitely degenerate.”

19. Let G(x, x  ) denote the Green function of the differential operator d2 /dx2 where
x ∈ [−1, 1].

“As a function of x, G is continuous at x = x  , but its derivative ∂G/∂x has a

finite discontinuity at x = x  .”

171
20. “The fundamental Green function of the Laplacian operator ∇2 in four-dimensional
Euclidean space is G(r, r ) = −1/[4π 2 (r − r )2 ].”

21. Consider the diffusion equation in d-dimensional space, ∂f /∂t = D∇2 f with
boundary condition f (r, t) → 0 as r → ∞ and initial condition f (r, 0) = δ (d) (r).

“The fundamental solution to this equation is a Gaussian in each Cartesian com-

ponent of r, for all positive integer values of the dimension d.”

22. The scattering amplitude for the scattering of a nonrelativistic particle of mass
m in a central potential λV (r) is given by

mλ 
f (k, θ) = − 2
d3 r e−ik · r V (r) ψ(r),
2π

where k  is the scattered wave vector.

“This formula is valid only if the potential V (r) decays to zero as r → ∞ more
rapidly than any inverse power of r.”

23. Continuation: “In the Born approximation, the scattering amplitude in the for-
ward direction (θ = 0) vanishes identically.”

24. Continuation: “In the Born approximation, the imaginary part of the scattering
amplitude vanishes identically.”

25. Consider the Helmholtz operator ∇2 + k2 in three-dimensional space.

“The fundamental Green function of this operator, corresponding

 to outgoing
 ik·(r−r  ) 
spherical waves, is G(r − r ) = −e / 4π|r − r | .”

26. Consider the wave operator (1/c2 ) ∂ 2 /∂t2 − ∇2 in (d + 1)-dimensional spacetime,

where c is the speed of light in a vacuum. Let G(d) (R, τ ) denote the causal re-
tarded Green function of the operator.

“G(d) (R, τ ) vanishes identically when (cτ, R) is a time-like four-vector.”

27. Continuation: “G(d) (R, τ ) is singular when (cτ, R) is a light-like four-vector.”

172
 28. Let J = (Ji , J2 , J3 ) be the generators of rotations in three-dimensional space,
satisfying the Lie algebra [Jj , Jk ] = ijkl Jl .

“The lowest-dimensional, non-trivial, unitary representation of the generators is

in terms of (2 × 2) matrices with complex elements.”

29. “The parameter space of the group SU(n) is doubly connected.”

30. “The first homotopy group of the parameter space of the special orthogonal group
SO(n), for every n ≥ 3, is Z2 .”

II. Fill in the blanks in the following.

1. Given that the imaginary part of an entire function f (z) is

2 −y 2 )
v(x, y) = e(x sin (2xy),

the function is f (z) = · · · .

2. The meridian of longitude ϕ on the Riemann sphere is mapped into a straight

line in the complex plane. The equation of this straight line is y = mx + c, where
m = · · · and c = · · · .

3. The region of absolute convergence in the complex z-plane of the power series
∞ %
 &
(n + 1)/(n2 + 1) ( 12 z)n is · · · .
n=0

4. The residue at infinity of the function f (z) = (z − z −1 )3 is Res f (z) = · · · .

z=∞

5. Let [z1 , z2 ; z3 , z4 ] denote the cross-ratio of the four points z1 , z2 , z3 and z4 in the
complex plane. Then [z1 , z2 ; z3 , z4 ] + [z1 , z3 ; z2 , z4 ] = · · · .

6. The Möbius transformation z → w such that three given points z1 , z2 , z3 are

mapped respectively into three other given points w1 , w2 , w3 is expressed by a
relation between w and z that reads · · · .

7. Under the Möbius transformation z → w = (z + 1)/(z + 2), an infinitesimal area

element δA centered at the point z = −3/2 is mapped to an element of area
λ δA, where the value of λ is · · · .

173
 
∞
8. The Bernoulli numbers Bn are defined via the expansion z/(ez −1) = Bn z n /n!.
n=0
Therefore Bn is given by the contour integral Bn = · · · . (You must specify both
the integrand and the contour.)

9. The Chebyshev polynomial of the second kind, Un (cos θ), has the generating
function
1 ∞
= Un (cos θ) tn ,
1 − 2t cos θ + t2
n=0

where θ ∈ [0, π]. Therefore Un (cos θ) can be expressed as a contour integral in

the t-plane given by Un (cos θ) = · · · . (You must specify both the integrand and
the contour.)

10. Continuation: Evaluating the contour integral and simplifying the result, the
final expression for Un (cos θ) is Un (cos θ) = · · · . (You must express your answer
in terms of trigonometric functions of θ.)

11. Continuation: Hence the polynomial U1 (cos θ) reduces to U1 (cos θ) = · · · .

12. The function f (z) = (z 2 + 2)1/3 has branch points at z = · · · .

13. Express your answer in terms of a Bessel function:

The residue of f (z) = exp (z − z −1 ) at z = 0 is Res f (z) = · · · .
z=0

14. The inverse Laplace transform of f(s) = 1/(s2 − 2s + 1) is f (t) = · · · .

15. Let λ be a positive constant. The Laplace transform of the function

 t  tn  t2
f (t) = dtn dtn−1 · · · dt1 e−λ(t−t1 )
0 0 0

is f(s) = · · · .

16. Let 
1 − |x|, |x| ≤ 1
f (x) =
0, |x| > 1.
If f(k) denotes the Fourier transform of f (x), the value of the integral
 ∞
dk |f(k)|2 = · · · .
−∞

17. The positional probability distribution at any time t ≥ 0 of a random walker on

a square lattice with sites labelled by the integers (, m) is given by
 √   √ 
P (, m, t) = e−λt (p1 /q1 )/2 (p2 /q2 )m/2 I 2λt p1 q1 Im 2λt p2 q2 ,

174
 where λ is the mean jump rate and pi , qi are directional probabilities such that
p1 + q1 + p2 + q2 = 1. The leading asymptotic behavior of P (, m, t) at very long
times (λt  1) is given by P (, m, t) ∼ · · · .

18. The diffusion equation for the positional probability density of a particle diffusing
on the x-axis in the region −∞ < x ≤ a, in the presence of a constant force, is
given by
∂p(x, t) ∂p (x, t) ∂ 2 p (x, t)
= −c +D .
∂t ∂x ∂x2
Here c and D are positive constants denoting the drift velocity and diffusion
constant, respectively. p(x, t) is normalized to unity for all t ≥ 0. There is a
reflecting boundary at the point x = a. The boundary condition satisfied by
p(x, t) at x = a is then given by · · · .

19. Continuation: As t → ∞, p(x, t) tends to the stationary probability density

pst (x). This quantity satisfies the ordinary differential equation · · · .

20. Continuation: The normalized solution for pst (x) is pst (x) = · · · .

21. A quantum mechanical particle of mass m moving in one dimension has the
Hamiltonian H = p2 /(2m), where p is the momentum operator of the particle.
Its momentum-space wave function at t = 0 is given to be φ(p). Therefore its
momentum-space wave function at any time t ≥ 0 is given by ψ(p, t) = · · · .

22. The scattering amplitude for a nonrelativistic particle of mass m in a central

potential λV (r) is given, in the Born approximation, by

2mλ ∞
fB (k, θ) = − 2 dr r sin (Qr) V (r),

Q 0
where Q is the magnitude of the momentum transfer vector Q. The forward scat-
tering amplitude in the Born approximation is therefore given by the expression
fB (k, 0) = · · · .

23. Continuation: The backward scattering amplitude in the Born approximation is

therefore given by the expression fB (k, π) = · · · .

24. Let R = r − r  and τ = t − t  , as usual. Let G(d) (R, τ ) denote the fundamen-
tal Green function of the Klein-Gordon operator  + µ2 , where µ is a positive
constant and  = (1/c2 )(∂ 2 /∂t2 ) − ∇2 , in (d + 1)-dimensional spacetime. Then
G(d) (R, τ ) can be expressed in the form

(d) 1
G (R, τ ) = dd k φ(k, R, τ ),
(2π)d

where φ(k, R, τ ) = · · · .

175
25. Let J = (J1 , J2 , J3 ) denote the generators of rotations in three-dimensional
space, and let ψ be an arbitrary angle. The quantity e−iJ1 ψ J2 eiJ1 ψ , expressed
as a linear combination of the generators, is · · · .

26. Continuation: Let n = (n1 , n2 , n3 ) be an arbitrary unit vector. Then the

commutator [J · n , J2 ] = .

27. The number of generators of the orthogonal group O(n) and the special orthog-
onal group SO(n) are, respectively, · · · and · · · .

28. The number of generators of the unitary group U(n) and the special unitary
group SU(n) are, respectively, · · · and · · · .

29. Let x and p denote the position and momentum operators of a quantum mechan-
ical particle moving in one dimension, so that their commutator [x , p] = i
I,
where I is the unit operator. Let a be a real constant with the physical dimen-
sions of length. Using Hadamard’s Lemma, the operator eiap/
x e−iap/
simplifies
to
eiap/
x e−iap/
= · · · .
This result tells us why the momentum operator p is the generator of translations
in position space.

30. Continuation: Let b be a real constant with the physical dimensions of linear
momentum. Once again, using Hadamard’s Lemma, the operator e−ibx/
p eibx/

simplifies to
e−ibx/
p eibx/
= · · · .
This result tells us why the position operator x is the generator of translations
in momentum space.

176
 Quiz 2: Solutions
I. True or false:

1. T

2. T

3. F

4. F

5. T

6. F

7. T

8. F

9. T

10. T

11. T

12. F

13. F

14. T

15. T

16. T

17. F

18. T

19. T

20. T

21. T

22. F

23. F

24. T

177
 25. F

26. F

27. T

28. T

29. F

30. T

II. Fill in the blanks:

2
1. f (z) = ez .

2. m = tan ϕ and c = 0.

3. |z| < 2.

4. Res f (z) = −3.

z=∞

5. 1.
(w − w2 )(w1 − w3 ) (z − z2 )(z1 − z3 )
6. = .
(w − w3 )(w1 − w2 ) (z − z3 )(z1 − z2 )
7. λ = 16.

n! dz
8. Bn = , where C encloses the origin once in the positive sense.
2πi C (ez − 1)
zn

1 dt
9. Un (cos θ) = , where C encloses the origin once in
2πi C tn+1 (1 − 2t cos θ + t2 )
the positive sense.
sin (n + 1)θ
10. Un (cos θ) = .
sin θ
11. U1 (cos θ) = 2 cos θ.
√ √
12. i 2, −i 2, ∞.

13. −J1 (2).

14. t et .

15. 1/[s(s + λ)n ].

16. 4π/3.

178
 %  √ √ &
(p1 /q1 )/2 (p2 /q2 )m/2 exp − λt 1 − 2 p1 q1 − 2 p2 q2
17. P (, m, t) ∼ .
4πλt (p1 q1 p2 q2 )1/4
 ∂p 
18. D −cp = 0.
∂x x=a

d2 pst (x) dpst (x)

19. D − c = 0.
dx2 dx
20. pst (x) = (c/D) ec(x+a)/D .
2 t/(2m
)
21. ψ(p, t) = e−ip φ(p).

2mλ ∞ 2
22. fB (k, 0) = − dr r V (r).

2 0

mλ ∞
23. fB (k, π) = − 2 dr r V (r) sin (2kr).

k 0
sin cτ k ik·R
24. φ(k, R, τ ) = c θ(τ ) e .
k
25. J2 cos ψ + J3 sin ψ.

26. 0.

27. 1
2
n(n − 1) and 12 n(n − 1).

28. n2 and n2 − 1.

29. x + a.

30. p + b.

179