PDEs, Special Functions and Basic Fourier Analysis

Notes by G. K. Srinivasan

Department of Mathematics, IITBombay

Dieses Buch widme ich meiner Frau in Dankbarkeit

1
 Table of Contents

Preface

Introducing the players 2

0 Preparatory results from infinite series

1 Power series solutions of ODES - Legendre polynomials

2 Differential equations with regular singular points

3 Sturm-Liouville problems and PDEs

4 Fourier series and PDEs

5 Fourier transforms and PDEs

6 Applications to celestial mechanics

2
 Preface

The development of special functions and differential equations have over three centuries occurred
in perfect harmony and with a view towards applications to physics. The course tries to keep this
synergy in focus, to unfold these development and present them naturally keeping technicalities at
bay.
Over the years surprising connections with many other branches of mathematics - particularly
number theory, probability theory and combinatorics have been unearthed. The instructor hopes that
students would be stimulated enough to take an occasional look at the many references to further
education and explorations into interesting by-lanes of mathematical physics. A subject that has
developed in the hands of giants such as Gauss, Euler, Riemann, Laplace and Legendre to name a few,
cannot fail to whet the appetite of the mathematically inclined students.

Happy reading !!

3
 Introducing the players - Equations of mathematical physics

In this course we shall focus on the fundamental partial differential equations that arise in mathe-
matical physics and their ODE reductions notably the Legendre and Bessel’s equations. The evolution
of a physical system is governed by partial differential equations such as
1. The wave equation
2. The heat equation
3. Maxwell’s equations of electromagnetism
4. The Schrödinger equation.
In the heart of each of these lies the Laplace operator. In the case of Maxwell’s equations, each
component of electric and magnetic field satisfies the wave equation. The steady state behavior is
governed by the Laplace’s equation.

Ubiquity of the Laplacian Why does nature favour the Laplacian? Well, if the physical system
under consideration is homogeneous (translational symmetry), then the governing equations are con-
stant coefficient equations. Now, if the solutions exhibit rotational symmetry (in the space variables)
which would be so if the physical phenomenon under consideration is isotropic about the origin, then
the governing equations would be such that the partial derivatives in space variables would always
occur as polynomials in the Laplacian. This can be established mathematically 1 .
Thus the Laplace’s operator is the most important operator in mathematical physics and its ubiq-
uity is a reflection of the symmetry of space with respect to the group of spacial rotations. One must
take advantage of this symmetry and choose coordinate systems adapted to it.

The ODE reductions of the basic equations When the PDEs of mathematical physics are writ-
ten in spherical polar coordinates and the variables are separated one immediately sees the appearance
of the Legendre and associated Legendre equations:
1. Legendre’s equation:
((1 − x2 )y 0 )0 + p(p + 1)y = 0.

2. The Associated Legendre’s equation:

((1 − x2 )y 0 )0 − m2 y/(1 − x2 ) + p(p + 1)y = 0.

In problems such as diffraction of light through a circular aperture it is more expedient to use cylindrical
coordinates and we see the emergence of the differential equation of Bessel which is
1
See for example G. B. Folland, Introduction to Partial Differential Equations, Prentice Hall, New Delhi

4
 3. Bessel’s equation
x2 y 00 + xy 0 + (x2 − p2 )y = 0.

Thus it is no surprise that the radii of Newton’s rings are expressible in terms of the zeros of Bessel
functions. These ODEs are variable coefficient equations and as such their solutions are not express-
ible in terms of the elementary transcendental functions. Indeed they define higher transcendental
functions.

Methods for representing solutions There are two basic methods of representing the solutions
of these ODEs.

1. Power series: These are amenable to algebraic manipulations and are easy to obtain from the
ODEs. But it is difficult to obtain information such as growth of solutions, zeros of solutions
and estimates.

2. Integral representations: These are usually harder to arrive at but they are more useful
to estimate the solutions, their derivatives and growth/decay properties when the parameters
involved are small/large. One such method has already been seen in the earlier course namely
the method of Laplace transforms. The ODEs studied were such that the solutions could be
“read off” from their Laplace transforms but a systematic inversion of the Laplace transform
would lead to an integral representation of solutions.
Another useful technique to obtain integral representation is the Fourier transform. We shall see
some of this later in the course.

5
 0 Preparatory results on infinite series

We begin with some generalities on infinite series which are important inasmuch as power series
and trigonometric series would appear in great profusion in these lectures.
Given an infinite sequence a0 , a1 , a2 , . . . we wish to assign a meaning to

a0 + a1 + a2 + . . .

The infinite sum above is called an infinite series and is defined to be

lim (a0 + a1 + a2 + · · · + an ).
n→∞

∞
X
The limit (if it exists) is denoted by an and we say the series CONVERGES. If the limit doesn’t
n=0
exits the series is said to be DIVERGENT. The finite sum a0 + a1 + a2 + · · · + an is called the n-th
partial sum of the infinite series.
The simplest example is that of a geometric series where it is possible to find a closed expression
for the sum of the first n + 1 terms:
c(rn+1 − 1)
c + cr + · · · + crn =
r−1
Passing to the limit as n → ∞ we conclude

c + cr + cr2 + · · · = c/(1 − r), |r| < 1

If |r| ≥ 1 the limit doesnt exist except in the trivial case when c = 0. Some easy consequences of the
definition
P
Theorem 0.1: If the series an converges then limn→∞ an = 0.

Proof: Let Sn = a0 + an + · · · + an and l be the limit of Sn as n → ∞. Then

lim an = lim (Sn − Sn−1 ) = l − l = 0.

n→∞ n→∞

The converse of this result is not true. It is not difficult to see that
1 1
1+ + + ...
2 3
diverges but the n-th term goes to zero as n → ∞. In fact we have the following general result:

6
Theorem 0.2 (Improper integral test): Suppose f (x) is a positive monotone decreasing function
defined on [1, ∞) then
f (1) + f (2) + f (3) + . . .
converges if and only if the improper integral
Z ∞
f (x)dx
1

converges.
¿From the improper integral test we see at once that 1 + 21 + 13 + . . . diverges.

Exercises:
1. Show that
1 1
1+ p
+ p + ...
2 3
converges if p > 1 and diverges if p ≤ 1.
P
A series an is said to absolutely converge if

|a0 | + |a1 | + |a2 | + . . .

converges.

Question : Suppose
|a0 | + |a1 | + |a2 | + . . .
converges, what can one say about
a0 + a1 + a2 + . . .?

Theorem 0.3: An absolutely convergent series is convergent.

The proof is not difficult but we shall not discuss it here.
What is the significance of the concept of absolute convergence?? Well, in general rearranging
the terms of an infinite series may alter the character of the series (a convergent one can turn into
a divergent one) and even it the rearranged series remains convergent the sum may change. Before
investigating this we insert another notion. A series that converges but not absolutely is said to be
conditionally convergent

Theorem 0.4 (The Alternating series test): Suppose a0 , a1 , a2 , . . . is a monotone decreasing

sequence of positive reals such that an → 0 as n → ∞ then the series

a0 − a1 + a2 − a3 + a4 − . . .

converges.
¿From this we infer immediately that
1 1 1
1− + − + ...
2 3 4
converges but not absolutely, that is to say it is a conditionally convergent series.

7
Example forbidding rearrangements: Let us consider the two series
1 1 1 1
S =1− + − + − ...
2 3 4 5
and
1 1 1 1 1 1
T =1− − + − − + − ...
2 4 3 6 8 5
Well, let us examine T closely:
1 1 1 1 1 1 1 1
T = (1 − ) − + ( − ) − + ( − ) − + ...
2 4 3 6 8 5 10 12
The last one equals
1 1 1 1 
1 − + − + . . . = S/2.
2 2 3 4
6 S. It turns out that S = ln 2.
Which means T =

Theorem 0.5 (Riemann’s rearrangement theorem): A conditionally convergent series may

through a suitable rearrangement made to converge to ANY PREASSIGNED real value.
In the light of this the following result is remarkable.

Theorem 0.6 (Dirichlet’s theorem): Rearranging the terms of an absolutely convergent series
has no effect on its behaviour - that is to say it remains an absolutely convergent series. Moreover,
the sum is also unaffected.
Riemann’s theorem appeared in his Habilitationschrift written in 1854 and published posthumously
in 1867 by Richard Dedekind, a year after Riemann’s death. This work, originating from the study of
Fourier series, also lays the foundations of the theory of integration.

The Cauchy product of two series: Given two convergent series

A = a0 + a1 + a2 + . . . , B = b0 + b1 + b2 + . . . ,

Their Cauchy product is defined to be the series

C = a0 b0 + (a0 b1 + a1 b0 ) + (a0 b2 + a1 b1 + a2 b0 ) + . . .

The general term of the Cauchy product is the finite sum

a0 bn + a1 bn−1 + · · · + an b0 .

Exercises:
2. Compute the Cauchy product of a geometric series with itself.

Question: Is C a convergent series and if so is it the case that C = AB?

Theorem 0.7 (The product theorem of Cauchy): Suppose A and B displayed above are both
absolutely convergent then their Cauchy product C is also absolutely convergent and C = AB.
We shall have occasion to use this result frequently.

8
Remark: It is not necessary that both series be absolutely convergent. Mertens proved the following
result:

Theorem 0.8 (Mertens): If the two series converge and at least one of them converges absolutely
then the Cauchy product converges and C = AB.

Tests for absolute convergence We now discuss two useful criteria for absolute convergence of
series.
P
Theorem 0.9 (D’Alembert’s Ratio Test): Let an be a series of non zero numbers and
an+1
L = lim | |
n→∞ an
P
The series an converges absolutely if L < 1. The series diverges if L > 1 or L = +∞. If L = 1 the
test is inconclusive.
P
Theorem 0.10 (Cauchy’s Root Test): Let an be a series and

L = lim |an |1/n

n→∞
P
The series an converges absolutely if L < 1. The series diverges if L > 1 or L = +∞. If L = 1 the
test is inconclusive.

Power Series: The tests of D’Alembert and Cauchy are particularly useful while dealing with series
of the form
c0 + c1 x + c2 x 2 + c3 x 3 + . . .
These series are called power series.
As an example let us take up
x 2 x3
1+x+ + + ....
2! 3!
Here the general term is an = xn /n!. We apply D’Alembert’s Ratio test
an+1
| | = |x|/(n + 1)
an
Proceeding to the limit as n → ∞ we see that
an+1
L = lim | |=0
n→∞ an
which is strictly less than one and so the series converges absolutely for any value of x real or complex.

The Exponential Series: Let E(x) denote the sum of the above series namely,

x2
E(x) = 1 + x + + ....
2!
The function E(x) is defined for all complex values of x.

9
Sine and Cosine series: Show using the ratio test that both the series
∞ ∞
X (−1)n x2n+1 X (−1)n x2n
S(x) = , C(x) =
n=0
(2n + 1)! n=0
(2n)!

Converge absolutely for all values of x real or complex.

Exercises:
∞
X
3. Show that the series xn /n2 converges absolutely for |x| < 1 and diverges when |x| > 1. What
n=1
happens when |x| = 1?
∞
X √
4. Discuss the series xn / n.
n=1

∞ 
X 1 n2 n
5. Discuss the series 1+ x .
n=1
n
∞
X
6. Discuss the series nn xn
n=1

an xn there exists in general an R ≥ 0 such that

P
Differentiation theorem: Given a power series
the series converges absolutely for |x| < R and diverges for |x| > R. Along the circle |x| = R nothing
can be said in general. It may happen that R = +∞. This R is called the radius of convergence of
the power series.

Theorem 0.11: The sum of the power series is differentiable infinitely often in the disc {x : |x| <
R}. Furthermore if f (x) denotes the sum of the power series then
∞
X
f 0 (x) = nan xn−1
n=1

In other words the power series may be differentiated term by term within the disc of convergence.

Examples on Differentiation theorem:

Exercises:

7. Prove that
x3 x 5
tan−1 x = x − + − ..., |x| < 1.
3 5
8. Prove that
x2 x3
log(1 + x) = x − + − ..., |x| < 1.
2 3

10
Question: Are we justified in setting x = 1 in order to derive
1 1
log 2 = 1 −+ − ...
2 3
1 1
π/4 = 1 − + − . . .?
3 5

As such it is not justified but to legitimize it we need to invoke Abel’s limit theorem.

Exponential Addition Theorem: As an illustration on the use of the Cauchy Product theorem
let us compute the Cauchy product of the series for E(x) and E(y) with general terms xm /m! and
y l /l! respectively. The n-th term of the Cauchy product is
X xm y l n
X xm y n−m
=
m+l=n
m! l! m=0
m! (n − m)!
n
1 X n! m n−m (x + y)n
= x y =
n! m=0 m!(n − m)! n!

Thus, the Cauchy product of E(x) and E(y) is

∞
X (x + y)n
= E(x + y)
n=0
n!

We have proved
E(x + y) = E(x)E(y).
We shall see an analogue of this result in chapter - II for the Bessel’s Functions.

Power Series with Arbitrary Center The series

∞
X
cn (x − p)n
n=0

is said to be a power series with center p. The theory of this parallels the theory of power series with
center at the origin and so need not be repeated here. In particular the power series has a radius of
convergence R which may be zero, a positive real number or infinity. The disc of convergence is given
by
|x − p| < R
The series converges absolutely in the disc of convergence and diverges on |x − p| > R.

Analytic Functions The sum of the power series f (x) is differentiable infinitely often in the disc
of convergence and the derivative is given by
∞
X
f 0 (x) = ncn (x − p)n−1
n=1

11
A function f (x) defined on an open set is said to be analytic at a point p if the function admits a
power series representation
X∞
f (x) = cn (x − p)n
n=0

valid in a disc |x − p| < r. Clearly the coefficients cn would depend on the chosen point. For most
familiar functions the Taylor series at a point gives the power series representation at the point. The
function is said to be analytic throughout its domain it is so at each point of the domain.

Exercises:

9. Show that x cot x is analytic everywhere except at ±π, ±2π, ±3π, . . . .

10. Show that −2x/(1 − x2 ) is analytic everywhere except at ±1.

11. Examine whether sin πx/ log(1 + x) is analytic at the origin. If it is analytic determine the first
three terms of the power series expansion.
Hint: Suppose 1 + c1 x + c2 x2 + . . . is a power series with positive radius of convergence, then
seek a power series for the reciprocal:
1
2
= d0 + d1 x + d2 x2 + . . .
1 + c1 x + c2 x + . . .
Multiply out, equate coefficients and check that d0 , d1 , d2 , . . . are uniquely determined. The series
d0 + d1 x + d2 x2 + . . . has positive radius of convergence but we shall not prove this.

12
 I - Power Series Solutions of ODEs - Legendre Polynomials

We shall solve second order differential equations

y 00 + A(x)y 0 + B(x)y = C(x)

where A(x), B(x) and C(x) are analytic in a neighborhood of a point x0 in the form of a power series
centered at x0 . Thus we look for solutions of the form
∞
X
y(x) = an (x − x0 )n ,
n=0

where coefficients an are determined through a recurrence relation.

The procedure is best illustrated through an example. Let us consider the Legendre’s equation:

(1 − x2 )y 00 − 2xy 0 + p(p + 1)y = 0. (1)

The coefficient (1 − x2 ) does not vanish at the origin and so origin is not a singular point. We seek a
solution of (1) in the form of a power series
∞
X
y(x) = an x n (2)
n=0

Differentiating,
∞
X
−2xy 0 (x) = −2nan xn (3)
n=0
2 00
and for the term x y (x) we have
∞
X
2 00
x y (x) = n(n − 1)an xn (4)
n=0

The term y 00 in the ODE is to be dealt with in the following manner:

∞
X ∞
X
00 n−2
y = n(n − 1)an x = n(n − 1)an xn−2 .
n=0 n=2

It is important to arrange it so that the exponent of the general term in the series is n. Here it is n − 2.
So we set n − 2 = N in the series and we get
∞
X
y 00 = (N + 2)(N + 1)aN +2 xN .
N =0

13
Changing the dummy index N to n we see that
∞
X
00
y = (n + 2)(n + 1)an+2 xn . (5)
n=0

Consolidating we get

(1 − x2 )y 00 − 2xy 0 + p(p + 1)y =

∞
X  
xn (n + 2)(n + 1)an+2 − n(n − 1)an − 2nan + p(p + 1)an = 0
n=0

Equating the coefficient of xn to zero we get the recurrence relation:

−an (p − n)(p + 1 + n)
an+2 = , n = 0, 1, 2, . . .
(n + 2)(n + 1)
Setting n = 0, 1, 2, . . . in succession we find the list of coefficients:

a2 = −a0 p(p + 1)/2!, a3 = −a1 (p − 1)(p + 2)/3!

a0 p(p − 2)(p + 1)(p + 3) a1 (p − 1)(p − 3)(p + 2)(p + 4)

a4 = , a5 =
4! 5!
The law of formation is clear:
p(p − 2) . . . (p − 2n + 2)(p + 1)(p + 3) . . . (p + 2n − 1)
a2n = (−1)n a0
(2n)!
and

a2n+1 =
(p −1)(p −3) . . . (p − 2n + 1)(p + 2)(p + 4) . . . (p + 2n − 1)
(−1)n a1
(2n + 1)!

The general solution is thus given by

 p(p + 1)x2 p(p − 2)(p + 1)(p + 3)x4 
a0 1 − + − ... +
2! 4!
3
 (p − 1)(p + 2)x (p − 1)(p − 3)(p + 2)(p + 4)x5 
a1 x − + − ... .
2! 4!

The coefficients a0 and a1 may be assigned arbitrary values and assigning the values a0 = 1, a1 = 0 we
get one solution and another linearly independent one by setting a0 = 0, a1 = 1. Both are power series
with unit radius of convergence (Exercise).
We see that if p is an integer then exactly one of the two series terminates and we have a polynomial
solution of the Legendre equation. With a suitable normalization that we shall presently specify, these
polynomials are called Legendre Polynomials.

14
The Legendre Polynomials Assume that p = k is an integer and one of the two series described
above terminates into a polynomial solution f (x). One can show without much difficulty that f (1) 6= 0.
Try this out. Suppose not. Set x = 1 in the Diff. Eq. to conclude that f 0 (1) = 0 as well. Assume
f (n) (1) = 0. Differentiate the differential equation n-times using Leibnitz rule and then put x = 1.
Apply induction on n.
Now that f (1) 6= 0 we can normalize our solution f (x) by dividing by f (1) and consider the solution
f (x)
.
f (1)
This special solution is called the k−th Legendre Polynomial. It is customary to denote this as Pk (x).
We record here the three defining properties of Pk (x). Note that k = 0, 1, 2, . . . .
1. Pk (x) is a polynomial of degree k
2. Pk (x) satisfies the Legendre Equation:
(1 − x2 )Pk00 (x) − 2xPk0 (x) + k(k + 1)Pk (x) = 0

3. Pk (1) = 1.
It is clear that Pk (x) is an odd function if k is odd and an even function if k is even. Also
P0 (x) = 1

Orthogonality Properties of Legendre Polynomials:

Theorem 1.1: If k 6= l then Pk (x) and Pl (x) are orthogonal in the following sense:
Z 1
Pk (x)Pl (x)dx = 0.
−1

Proof: To prove this we begin with the differential equations

(1 − x2 )Pk00 − 2xPk0 + k(k + 1)Pk = 0 (1)
and
(1 − x2 )Pl00 − 2xPl0 + l(l + 1)Pl = 0 (2)
We shall write these equations can be written in a more following convenient form known as self-adjoint
form.
d 
(1 − x2 )Pk0 + k(k + 1)Pk = 0 (3)
dx
and
d 
(1 − x2 )Pl0 + l(l + 1)Pl = 0 (4)
dx
Multiply (3) by Pl , (4) by Pk , subtract and integrate over [−1, 1]. Integration by parts would confirm
that Z 1
(k(k + 1) − l(l + 1)) Pk (x)Pl (x)dx = 0.
−1
Since k 6= l and are non-negative integers, the factor k(k + 1) − l(l + 1) 6= 0 and the proof is complete.

15
Exercise: Explain what happens if k and l are not non-negative integers.

Theorem 1.2 (Fundamental Orthogonality Lemma): Suppose V is a vector space endowed

with inner product with respect to which {v0 , v1 , v2 , . . . } and {w0 , w1 , w2 , . . . } are two orthogonal
systems of non-zero vectors. Further assume that
span(v0 , v1 , . . . , vk ) = span(w0 , w1 , . . . , wk ), for every k = 0, 1, 2 . . .
Then, for certain scalars ck (k = 0, 1, 2 . . . ),
vk = ck wk , for every k = 0, 1, 2 . . .
Proof is an Exercise. First think of what happens in ordinary Euclidean spaces like Rn . Geometrical
considerations suggests the proof.

Exercise: Consider the sequence of polynomials

dn
Qn (x) = n (x2 − 1)n .
dx
Show that Qn (x) has degree n for every n. Further show that the sequence is orthogonal with respect
to weight 1 namely, Z 1
Qn (x)Qm (x)dx = 0, m 6= n.
−1
¿From this infer the following result:
Pn (x) = cn Qn (x), for every n = 0, 1, 2 . . .
for a certain sequence of constants {cn }.

Rodrigues’ Formula: Compute the constants cn in the last slide by evaluating Qn (1). Deduce the
following formula due to Olinde Rodrigues.
1 dn 2
Pn (x) = n (x − 1)n
2 n! dxn
For more information on the work of Rodrigues see the Book Review by W. P. Johnson, in the American
Math. Monthly, Volume 114, Oct 2007, 752-7582

Exercises:
R1
1. Compute −1
(Pn (x))2 dx
R1
2. Show that −1 (1 − x2 )(Pn0 (x))2 dx = 2n(n + 1)/(2n + 1). Hint: Multiply the Diff. Eqn by Pn and
integrate by parts.
3. Use Rodrigues formula to prove that the Legendre polynomial of degree n has precisely n distinct
roots in the open interval (−1, 1). Use Rolle’s theorem. Note: The roots were used by Gauss in
1814 in his famous quadrature formula3 .
4. Show that the Legendre polynomials satisfy the three term recursion formula
(n + 1)Pn+1 − x(2n + 1)Pn + nPn−1 = 0.
2
Available in the link: http://www.jstor.org/stable/27642326
3
See the discussion on pp. 56-69 of S. Chandasekhar, Radiative transfer, Dover Publications, Inc., New York, 1960.

16
More Problems on Legendre Polynomials

5. Prove that Pn0 (1) = 12 n(n + 1)

0
6. Prove that Pn+1 − xPn0 = (n + 1)Pn

7. Prove that (x2 − 1)Pn0 − nxPn + nPn−1 = 0.

0 0
8. Prove that Pn+1 − Pn−1 − (2n + 1)Pn = 0.
0
9. Prove that xPn0 − Pn−1 = nPn .

10. Suppose xn = nj=0 cj Pj (x) show that cn = 2n (n!)2 /(2n)!.

P

11. Use the method of series solutions to find the power series expansion of (1 + x)a where a is any
real number. Hint: Find an ODE satisfied by the function.

Generating function for the Legendre Polynomials: Given a sequence {an } of real or complex
numbers, their generating function is by definition the power series
∞
X
an tn
n=0

Theorem 1.3: ∞
X 1
tn Pn (x) = √
n=0
1 − 2xt + t2
For connections to potential theory4 . There are several proofs of this important theorem and we select
the one from Courant and Hilbert’s monumental treatise5 . Only a sketch of the proof is provided and
the student can work out the details. According to the last exercise in the previous paragraph, the
function V (x, t) = (1 − 2xt + t2 )−1/2 can be expanded in an absolutely convergent series
∞
X
V (x, t) = Rn (x)tn
n=0

where each Rn (x) is a polynomial of degree exactly n (why?) and the series is valid for x ∈ [−1, 1]
and for a certain interval |t| < ρ. Indeed ρ = 0.4 would suffice (why?).
In particular
∞ X ∞
1 X
√ √ = Rj (x)Rk (x)uj v k
1 − 2xu + u2 1 − 2xv + v 2 j=0 k=0

Integrating both sides with respect to x over the range [−1, 1] we get

1  1 + √uv  X ∞ X ∞ Z 1
j k
√ ln √ = uv Rj (x)Rk (x)dx.
uv 1 − uv j=0 k=0 −1

4
see for instance, A. S. Ramsey, Newtonian Attraction, Camb. Univ. Press, p. 131 ff., or pp 121 - 134 of the more
comprehensive and classic treatise of O. D. Kellog, Foundations of potential theory, Dover, New York, 1953.
5
Methods of mathematical physics - I.

17
But the left hand side is (using the logarithmic series we have encountered earlier) given by

1  1 + √uv  X ∞
2un v n
√ ln √ =
uv 1 − uv n=0
2n + 1

Comparing the coefficient of uj v k in the last two expressions we get

Z 1
Rj (x)Rk (x)dx = 0, if j 6= k.
−1

Whereas if j = k then Z 1
(Rj (x))2 dx = 2/(2j + 1).
−1

The fundamental orthogonality lemma now implies that Rj (x) = Pj (x) for every j = 0, 1, 2, . . . .

Remark: The student is not expected to remember this proof for the examinations.
The function V (x, t) is the potential due to a point mass P placed at unit distance from the origin
at a point X at distance t from the origin, where x is the cosine of the angle between OX and OP .

Bernoulli Polynomials and Bernoulli Numbers: This is just intended as an example. We

inductively define the sequence of Bernoulli polynomials {Bn (x)} as follows:

Bn0 (x) = nBn−1 (x), n≥1

and B0 (x) = 1. Further we demand

Z 1
Bn (x)dx = 0, n ≥ 1.
0

The student can easily determine the first few Bernoulli polynomials. The numbers {Bn (0)} are called
the Bernoulli Numbers.

Easy exercise: Show that

∞
X Bn (x)tn
= C(t)ext (1)
n=0
n!
However it is not so easy to show C(t) = t/(et − 1). Here is an interesting suggestion due to Mr.
Sanket Sanjay Barhate. Integrate both sides of (1) with respect to x from 0 to 1. From the LHS
we get 1 and on the RHS we get C(t)(et − 1)/t. At the formal level this computation is correct but
the term by term integration needs justification. We need some decent estimate on |Bn (x)| valid over
[0, 1]. It may be noted that the Bernoulli numbers rapidly increase to infinity.
Thus ∞
X Bn (x)tn text
= t
n=0
n! e −1
So text /(et − 1) is the generating function for the sequence {Bn (x)/n!}.
It would not have been easy to guess P the above formula. So the important thing was to obtain the
closed form for the generating function ∞ n
n=0 Bn (x)t /n! (never mind it being formal). Let us now

18
abandon the task of justifying the interchange of integration and infinite sums but verify the result
obtained in an entirely different way. Let us define
 d n etx t
βn (x) =
dt et − 1 t=0
Then β0 (x) = 1. Now,
d  d n etx t2
(βn (x)) =
dx dt et − 1 t=0
Using Leibnitz rule we see immediately that βn (x) = nβn−1 (x). It is easy to justify (how?)
Z 1  n tx  d n Z 1 etx t
d e t
dx = dx.
0 dt et − 1 dt t
0 e −1

Setting t = 0 in the last equation we get,

Z 1  d n
βn (x)dx = 1 = 0, n = 1, 2, . . . .
0 dt
Thus we conclude that βn (x) = Bn (x) and the proof is complete.

Some Classical Formulas

James Bernoulli - Ars Conjectandi (1713):

p p p 1  
1 + 2 + ··· + n = Bp+1 (n + 1) − Bp+1 (0) , p = 1, 2, 3, . . .
n+1

Euler:
1 1 (−1)p+1 (2π)2p B2p (0)
1+ + + · · · = , p = 1, 2, 3, . . .
22p 32p 2 · (2p)!
James and John Bernoulli tried in vain to obtain the latter for p = 1. It was discovered by Euler in
1736. However James Bernoulli did not live to see the last displayed formula in which the numbers
that bear his name feature so prominently.

Tchebychev’s Differential Equation:

12. Discuss the series solutions of the Tchebychev’s differential equation:

(1 − x2 )y 00 − xy 0 + p2 y = 0

Show that if p is an integer, of the two linearly independent solutions exactly one of them
terminates into a polynomial solution which after suitable renormalization is denoted by Tn (x).
13. Rewrite the ODE in self-adjoint form and show that if k 6= l
Z 1
Tk (x)Tl (x)(1 − x2 )−1/2 dx = 0.
−1

In other words the Tchebychev’s polynomials form an orthogonal system with respect to the
weight function (1 − x2 )−1/2 .

19
 14. Show that sin(p sin−1 (x)) and cos(p cos−1 x) satisfy the Tchebychev’s equation.

15. Show that Tn (x) = cos(n cos−1 x). This means you need to prove first that the function on the
right is a polynomial. Try this by induction. Then invoke uniqueness of Tn as a polynomial of
degree n satisfying the ODE with appropriate normalization.

For more problems on Tchebychev’s polynomials, the student is referred to pp 177-187 of L. Sirovich,
Introduction to Applied Mathematics, Springer Verlag, 1988.

Hermite’s Equation - Appears in Quant. Mech. See A. Beiser, Perspectives of Modern Physics,
1969, pp 180-187. The differential equation appears on p. 183.

16. Discuss the series solutions of Hermite’s equation

y 00 − 2xy 0 + 2λy = 0

17. Write the equation in self-adjoint form. Hint: Multiply by exp(−x2 ).

18. If n is a non-negative integer show that one of the series solutions terminates and we get a
polynomial solution. After suitable normalization these are called the Hermite Polynomials
Hn (x).

19. Show that the Hermite polynomials are orthogonal on (−∞, ∞) with respect to the weight
function exp(−x2 ).

Orthogonal Polynomials in general The properties established for the Legendre Polynomials are
rather typical of classical orthogonal systems of polynomials.

1. It is a general fact that if {fn (x)} is a sequence of orthogonal polynomials then the zeros of fn (x)
are real distinct and lie in the interval of orthogonality.

2. The sequence {fn } satisfies a three term recursion formula.

3. The zeros of fn and fn+1 interlace (we have not proved this for the Legendre polynomials).

4. There is an analogue for the Rodrigues formula for the system of Hermite, Tchebychev, Laguerre
and other classical systems of polynomials. Note: The Laugerre polynomials arise when the
sequence 1, x, x2 , . . . is subjected to the Gram-Schmidt process with respect to the inner product
on [0, ∞) with weight function e−x . These also arise from the Laguerre ODE we shall see later.

The book by Ian. Sneddon, Special functions of mathematical physics and chemistry, Longman Math-
ematical Texts, Longman, New York, 1980, contains very instructive list of problems on Legendre
Polynomials (see pp. 96-105).

20
Additional Problems on Tchebychev’s polynomials: These examples are all from chapter 6 of
L. Sirovich, Introduction to applied mathematics, Springer Verlag, 1988.

20. Recall that Tn (x) = cos(n cos−1 (x)). Use this to determine the three term recursion formula for
the sequence {Tn (x)}.

21. Compute the integral

1
(Tn (x))2 dx
Z
√
−1 1 − x2
22. As for the case of Legendre polynomials, show that the Tchebychev’s polynomial Tn (x) has n
distinct roots in (−1, 1). Determine these roots.

23. Show that

1n √ √ o
Tn (x) = (x − i x2 − 1)n + (x + i x2 − 1)n
2
Hint: Write cosine in exponential form.

24. Use the previous result to prove that the generating function for the sequence {Tn (x)} is
1 − tx
G(x, t) = .
1 + t2 − 2tx

25. Use trigonometry to show that 2Tm (x)Tn (x) = Tm+n + Tm−n .

26. Show that Tn (Tm (x)) = Tmn (x).

27. Prove that

 d n−1 2n−1 1 · 3 · 5 . . . (2n − 1)
sin θ = (−1)n−1 sin nθ
d cos θ n
This formula is due to C. G. J. Jacobi (1836). See p. 26 ff. of G. N. Watson, Treatise on the
theory of Bessel functions to understand its immense use in special functions.
Hint: Put t = cos θ and show that
 d n−1 1
f (t) = (1 − t2 )n− 2
dt
is a solution of Tchebychev’s ODE whereby

f (t) = cn sin(n cos−1 t).

√
To determine cn divide both sides by 1 − t and let t → 1.

21
 II - Differential Equations with regular singular points

We shall discuss in detail the Bessel’s functions of the first kind but as a preparation we shall
have to begin with some basic notions on the gamma function. The gamma function was introduced
into analysis by L. Euler in 1729 in a letter to Goldbach seeking an analytic interpolation of the
factorial function. Since then the function has been subjected to intense research and still interests
mathematicians. For details on the historical developments and further references see the article by G.
K. Srinivasan, The gamma function - an eclectic tour, American Math. Monthly, Volume 114 (2007)
297-315. A. M. Legendre introduced the notation Γ(p) and since then the function is referred to as
the gamma function

Definition: The gamma function is defined as an integral

Z ∞
Γ(p) = e−t tp−1 dt, p > 0.
0

We can also allow p to be a complex number with positive real part. In this case the factor tp−1 is
defined as
tp−1 = exp((p − 1) ln t),
where it must be recalled that the exponential function was defined earlier as an infinite series. The
integral makes sense and defines an analytic function in the right half plane.

Basic Properties of the Gamma function: A simple integration by parts gives:

Γ(p + 1) = pΓ(p), Γ(1) = 1.

¿From this it immediately follows

Γ(n + 1) = n!, n = 0, 1, 2, . . . .

It is easy to verify that Z ∞ √

2
Γ(1/2) = e−x dx = π.
−∞

The Stirling’s Approximation Formula: It would be inappropriate not to mention at least briefly
one of the most remarkable results of classical analysis obtained by James Stirling in his Methodus
Differentialis in 1730.

22
Theorem 2.1 (Stirling’s approximation formula) As n → ∞,
√
n! ∼ nn e−n 2nπ

in the sense that the ratio tends to 1 as n → ∞.

The proof may be found for instance in the appendix to the first chapter of W. Feller: Introduction
to the theory of probability, Volume - I, Wiley.
The theorem can be extended to the gamma function. As x → ∞,
√
Γ(x + 1) ∼ xx e−x 2xπ

in the sense that the ratio tends to 1 as x → ∞.

We shall consider differential equations of the form

d2 y dy
A(x) + B(x) + C(x)y = 0
dx2 dx
where A(x), B(x) and C(x) are analytic functions - except possibly for isolated singularities.

Definition (Singular point): In general any point where B(x)/A(x) or C(x)/A(x) fails to be
analytic is said to be a singular point of the ODE. Suppose A(x0 ) = 0 then x0 may be a singular point
of the differential equation since B(x)/A(x) and C(x)/A(x) may fail to be analytic at the point x0 .
However there may be singularities other than the zeros of A(x) as the following example indicates.

xy 00 + (cot x)y 0 − (sec x)y = 0

has singular points at 0, ±π, . . . as well as π/2 ± π, π/2 ± 2π, . . . .

We say that x0 is a regular singular point if (x − x0 )B(x)/A(x) and (x − x0 )2 C(x)/A(x) are both
analytic at x0 . Otherwise the singular point is said to be irregular.

Definition After the substitution x = 1/t if the origin is a ordinary/reg. sing./irr. sing. point then
we say that the point at infinity is an ordinary/reg. sing/irr. sing. point of the original equation.
Check that the point at infinity is an irregular singular point of Airy’s Equation y 00 − xy = 0.

Examples of ODEs with regular singular points (except possibly ∞)

1. Legendre: (1 − x2 )y 00 − 2xy 0 + p(p + 1)y = 0.

2. Tchebychev: (1 − x2 )y 00 − xy 0 + p2 y = 0.

3. Bessel: x2 y 00 + xy 0 + (x2 − p2 )y = 0.

4. Laguerre: xy 00 + (1 − x)y 0 + py = 0.

5. Hypergeometric: x(1 − x)y 00 + (c − (a + b + 1)x)y 0 − aby = 0.

Discuss the nature of the point at infinity for each of the above.

23
 d y dy 2
Theorem 2.2 (Fuchs-Frobenius): Suppose the ODE A(x) dx 2 + B(x) dx + C(x)y = 0 has a regular

singular point at x = x0 . Then the differential equation admits at-least one solution of the form
∞
X
ρ
y(x) = (x − x0 ) an (x − x0 )n
n=0

an (x − x0 )n has positive radius of convergence.

P
where the associated power series
The theorem was discovered by Lazarus Fuchs in 1866 (Crelle’s Journal) and in 1876 (Crelle’s
Journal) Frobenius described the algorithm for determining the complete system of solutions.
We shall not prove the theorem here but as a motivation look at the case of Bessel’s ODE rewritten
as
x2 y 00 + xy 0 − p2 y = −x2 y
So the Bessel’s equation appears as a perturbed Cauchy-Euler equation with “forcing function” −x2 y.
For small values of x one hopes the solution would approximate the solution of the associated homoge-
neous equation
x2 y 00 + xy 0 − p2 y = 0
which has solutions
xp and x−p .
This suggests the Ansatz
 
y(x) = x±p a0 + a1 x + a2 x2 + . . . , , a0 6= 0. (1)

Series of the form (1) are called Frobenius Series. The Fuchs-Frobenius theorem assures us of the
correctness of the Ansatz (1). The reason for assuming a0 6= 0 is that if a0 = 0 and a1 6= 0 we can
factor out a x from the series and write
 
1±p 2
x a1 + a2 x + a3 x + . . . , ,

We can rename the exponent as well as the coefficients and get back to the Ansatz (1). Thus it is
legitimate to assume that this has been done at the outset. Hereafter we shall always assume that
a0 6= 0.

Determining the Frobenius series for an ODE This is done by substituting the Ansatz
∞
X
ρ
x an x n (1)
n=0

into the ODE. First we must determine the exponent ρ known as the Frobenius exponent and then
successive the coefficients a1 , a2 , . . . . Let us illustrate this for the Bessel’s equation

x2 y 00 + xy 0 + (x2 − p2 )y = 0. (2)

Well,
∞
X ∞
X
x2 y 00 = xρ (n + ρ)(n + ρ − 1)an xn , xy 0 = xρ (n + ρ)an xn
n=0 n=0

24
¿From x2 y 00 + xy 0 + (x2 − p2 )y = 0 we infer
∞
X   ∞
X
ρ 2 n
x an (n + ρ)(n + ρ − 1) + (n + ρ) − p x + an xn+ρ+2 = 0.
n=0 n=0

The last series must be modified by the substitution n + 2 = N . The result is

∞
X   ∞
X
ρ 2 2 n
x an (n + ρ) − p x + an−2 xn+ρ = 0.
n=0 n=2

Equating the n = 0 coefficient to zero we get (since a0 6= 0)

ρ2 − p2 = 0.

This quadratic which determines ρ is called the indicial equation. Thus

ρ = ±p.

We shall assume p ≥ 0 and work with the larger root p.

Equating the coefficient corresponding to n = 1 we get

a1 ((p + 1)2 − p2 ) = 0

which forces a1 = 0.
Finally we have the recursion formula

n(n + 2p)an = −an−2 , n = 2, 3, . . .

We see that all odd coefficients vanish and for the even coefficients which simplifies to
(−1)n a0
a2n = .
22n n!(1 + p)(2 + p) . . . (n + p)
Exercise: Show that the series
∞
X (−1)n x2n
n=0
22n n!(1 + p)(2 + p) . . . (n + p)

has infinite radius of convergence. So we get the solution

∞
X (−1)n x2n+p
a0 .
n=0
22n n!(1 + p)(2 + p) . . . (n + p)

Bessel’s functions of the first kind: In view of the properties of the Gamma function,

Γ(n + p + 1) = (n + p)Γ(n + p) = · · · = (n + p)(n + p − 1) . . . (1 + p)Γ(1 + p).

and the displayed expression in the last slide can be recast as

∞
X (−1)n x2n+p
a0 2p Γ(1 + p)
n=0
22n+p n!Γ(n + p + 1)

25
The Bessel’s function of the first kind denoted by Jp (x) is the solution with a specific value of
1
a0 = .
2p Γ(1 + p)
Thus ∞
X (−1)n (x/2)2n+p
Jp (x) =
n=0
n!Γ(n + p + 1)
It is clear that if p = 0 we can only get one solution by this method. Also note that if p is not an
integer then we can repeat the above with the smaller root −p as well and get a second solution. These
two are linearly independent (how?).
Show that r r
2 2
J1/2 (x) = sin x, J−1/2 (x) = cos x
πx πx
What about J3/2 (x)? One can determine this directly from the definition but it would be easier to
use the theorem that follows. The Bessel’s functions of half-integer order are expressible in terms of
elementary functions. We shall see later how this is reflected in the different behaviour of the Fourier
transform of radial functions of two and three variables.

Basic Properties of Bessel’s functions of the first kind For k ∈ N, we DEFINE J−k (x) =
(−1)k Jk (x).

Theorem 2.3: The following relations hold:

d p  d  −p 
x Jp (x) = xp Jp−1 (x), x Jp (x) = −x−p Jp+1 (x).
dx dx

Proof: Differentiation theorem which permits term by differentiation of the series:

∞
d p  d X (−1)n x2n+2p
x Jp (x) =
dx dx n=0 22n+p n!Γ(n + p + 1)

Assume p > 0 and we get

∞
 X (−1)n x2n+2p−1
d p
x Jp (x) =
dx n=0
22n+p−1 n!Γ(n + p)
and the RHS is precisely xp Jp−1 (x).
In case p = 0 then after differentiation the series would begin with the n = 1 term. Well, with
p = 0,
∞
d  X (−1)n (x/2)2n−1
J0 (x) =
dx n=1
n!(n − 1)!
Put n − 1 = N and we get
∞
d  X (−1)N (x/2)2N +1
J0 (x) = − = −J1 (x) = J−1 (x).
dx N =0
N !(N + 1)!

as asserted. The other formula is left as an exercise.

26
 We rewrite the formulas obtained in the previous theorem in a different form:

xJp0 + pJp = xJp−1 , xJp0 − pJp = −xJp+1

Hence
2pJp = x(Jp−1 + Jp+1 ), 2Jp0 = Jp−1 − Jp+1

Exercises:
1. Assume p ≥ 1. Show that between two successive positive zeros of Jp (x) there is a unique zero
of Jp−1 (x) and a unique zero of Jp+1 (x).

2. Deduce an exact expression for J±3/2 (x) in terms of the sine/cosine functions.

3. Use induction to show that

r
n 2 n+ 1  1 d n sin x
Jn+ 1 (x) = (−1) x 2
2 π x dx x

4. Prove that |Jm (x)| ≤ |x|m e|x| /m!.

A bilateral series ∞
X
an
n=−∞

is said to converge it each to the two series ∞

P P∞
n=0 an and n=1 a−n converges. The sum of the bilateral
series is then the sum to the last two. Likewise one defines the notion of absolute convergence. Show
that the bilateral series ∞
X
Jn (x)
n=−∞

converges absolutely. What is the coefficient of xt when each Jn (x) is written as a power series and
the resulting expression is rearranged? Ans: zero except when t = 0. Thus the sum of the bilateral
series is 1. Hint: First prove that
t    
X
n 2t 1 2t
(−1) =
n=0
t−n 2 t

Here is an alternate solution that avoids the tedious algebra involved in the previous slide. Suppose
the sum is denoted by C(x) and we assume term by term differentiation is valid. In fact it is valid but
the proof needs more background from analysis such as notions of uniform convergence. Then
∞
X
0
2C (x) = (Jn−1 (x) − Jn+1 (x)) = 0.
n=−∞

Deduce that C(x) = 1.

The Schlömilch’s Formula This is the formula giving explicitly the generating function for the
bilateral sequence {Jn (x) : −∞ < n < ∞}.

27
Theorem 2.4 (Schlömilch):
∞  tx
X x
Jn (x)tn = exp − , t 6= 0.
n=−∞
2 2t

To prove this we denote the sum of the series on the left hand side by F (x, t). Show that the series
converges for all values of t 6= 0. The differentiation theorem gives
∞
∂F (x, t) 1 X
= 2nJn (x)tn−1
∂t 2 n=−∞

But 2nJn = x(Jn−1 + Jn+1 ). Hence

∞ ∞
∂F (x, t) x X x X
= Jn−1 (x)tn−1 + Jn+1 (x)tn−1
∂t 2 n=−∞ 2 n=−∞

So finally we get the Diff. Eqn.

∂F (x, t)  x x 
= + F
∂t 2 2t2
Integrating we get  xt x
F (x, t) = exp − C(x)
2 2t
where we need to determine C(x). Put t = 1 and we get
∞
X
C(x) = Jn (x) = 1.
n=−∞

The proof is complete.

Integral Representation for the Bessel’s function:

Theorem 2.5: For m = 0, 1, 2, . . . we have the integral representation:

1 π
Z
Jm (x) = cos(x sin θ − mθ)dθ.
π 0

Proof: Put t = exp(iθ) in Schlömilch’s formula and we get

∞
X
Jn (x)einθ = exp(ix sin θ)
n=−∞

Multiply by exp(−imθ) and integrating over [−π, π] we get

Z π
2πJm (x) = eix sin θ e−imθ dθ
−π

28
Using eiy = cos y + i sin y we get
Z π Z π
2πJm (x) = cos(x sin θ − mθ)dθ + i sin(x sin θ − mθ)dθ
−π −π

The second integral is zero (why?) and we get the desired result:
Z π
1
Jm (x) = cos(x sin θ − mθ)dθ
2π −π

This was the form in which Bessel originally presented in 1824 the functions that bear his name. Bessel
was an astronomer in Königsberg. We shall later see an application to a famous problem in celestial
mechanics namely, inverting the famous Kepler Equation.

Legendre Polynomials Again!! Laplace’s Integral Representation An integral representation

very similar to the one we have just obtained for Bessel’s function was given by Pierre Simon Marquis
de Laplace in his great work on Celestial Mechanics 6

Theorem 2.6:
1 π
Z √
Pn (x) = (x + x2 − 1 cos φ)n dφ.
π 0
1 π
Z √
Pn (x) = (x + x2 − 1 cos φ)−n−1 dφ.
π 0

We immediately deduce that |Pn (x)| ≤ 1 when |x| ≤ 1.

Exercises: We prove the first formula. Call the integral Qn (x).

1. Why is Qn (x) a polynomial of degree n?

2. Verify that Q0 (x) = 1 = P0 (x) and Q1 (x) = x = P1 (x).

It suffices to show that the sequence Qn (x) satisfies the same three term recursion as Pn (x) namely

(n + 1)Qn+1 − x(2n + 1)Qn + nQn−1 = 0.

√
Write A for x + x2 − 1 cos φ and
√ d
An+1 = xAn + An x2 − 1 sin φ
dφ
Integrate by parts. We get
Z π
n
Qn+1 = xQn + An−1 (x2 − 1) sin2 φdφ.
π 0
6
Traité de méchanique célèste

29
Now write
√
(x2 − 1) sin2 φ = (x2 − 1) − ( x2 − 1 cos φ)2
= (x2 − 1) − (A − x)2 = 2Ax − A2 − 1.

Thus we get Z π
n
Qn+1 = xQn − (An+1 + An−1 − 2xAn )dφ
π 0
which readily translates to

(n + 1)Qn+1 − x(2n + 1)Qn + nQn−1 = 0.

For a different proof of this see Byerly, pp. 165-167. The second formula is easily deduced from the
first as will be indicated in the exercises.

Lots of Exercises ! Have Fun!!

1. Show that
1 d 2 2 n n+1 2 d
[Jn + Jn+1 ] = Jn2 − Jn+1 , [xJn Jn+1 ] = x(Jn2 − Jn+1
2
).
2 dx x x dx
and deduce from these the following
∞
X ∞
X
J0 (x)2 + 2 (Jn (x))2 = 1, (2n + 1)Jn Jn+1 = x/2.
n=1 n=0
P
Hint: Differentiate x (2n + 1)Jn Jn+1 .

2. Prove using Schlömilch’s formula:

∞
X
J0 (x)2 + 2 (Jn (x))2 = 1.
n=1

Hint: Replace t by −t is Schlömilch’s formula.

√
3. Deduce that |J0 (x)| ≤ 1 and Jn (x)| ≤ 1/ 2.

4. The last exercise shows in particular that the Laplace transform of J0 (x) exists. Find the Laplace
transform of J0 (x). Do this by computing term by term Laplace transform from the series. Do
this using the integral representation. Do this from the ODE - you will have some trouble
determining the arbitrary constant.

5. Recall the definition of convolution from MA 108. Determine the convolution pf J0 (x)u(x) with
itself where u(x) is the Heaviside unit step function. Hint: Use the convolution theorem.

6. Prove the Bessel addition theorem.

∞
X
Jn (x + y) = Jk (x)Jn−k (y).
k=−∞

30
 √  
7. Show that xJ±1/3 2 3/2
3
x satisfy the Airy’s equation y 00 + xy = 0.

8. Show that xJ0 (x) is a solution of the ODE y 00 + y = −J1 (x). Deduce that
Z x
xJ0 (x) = cos(x − t)J0 (t)dt.
0

9. Prove the following: Z 1

1
J0 (x) = eitx (1 − t2 )−1/2 dt
π −1

and more generally for non-negative integer values of k,

Z 1
xp
Jk (x) = p √ 1 eitx (1 − t2 )k−1/2 dt
2 πΓ(p + 2 ) −1

10. Starting from the series for J0 (x) and the formula
Z π
π(2n)!
sin2n θdθ = n
0 2 n!
re-derive the integral representation for J0 (x). By differentiating derive it for J1 (x) as well.

11. Attempt a Frobenius series solution y(x) = xρ ∞ n 2 00 0

P
n=0 an x of the ODE x y + (3x − 1)y + y = 0.
Why does the method break down?

12. Determine the indicial equation of the Hypergeometric ODEs. Show that it admits a series
solution as well as a Frobenius series solution if c is not an integer.

13. Show that the Laguerre ODE has a power series solution. What is its indicial equation? Show
that for a discrete set of parameter values it admits polynomial solutions and that these polyno-
mials are orthogonal on [0, ∞) with respect to the weight function e−x .

14. Show that the sequence of polynomials

Ln (x) = ex Dn (xn e−x )

forms an orthogonal system with respect to the weight function e−x . Determine the inner product
of Ln (x) with itself. Do the polynomials Ln (x) satisfy the Laguerre ODE?

15. Show that the substitution t = 21 (1 − x) reduces the Legendre’s equation to the Hypergeometric
equation with a = p + 1, b = −p and c = 1.

Some points on Frobenius’s algorithm

1. Suppose the indicial equation has roots ρ1 and ρ2 that DO NOT differ by an integer then each of
them gives rise to a Frobenius series solution of the ODE and the two are linearly independent.

2. Suppose the roots ρ1 and ρ2 differ by an integer, the one with larger real part always leads to
a Frobenius series solution. However if we attempt to use the root with smaller real part three
possibilities occur as described below

31
 A The procedure goes through and we do get a second linearly independent solution (such as
in the case of Bessel’s ODE of order 1/2, 3/2, . . . ).
B The recurrence relation breaks down at some stage and its validity ceases.
C The recurrence relation does not break down but the resulting solution is a multiple of the
solution obtained earlier. The trivial example being the case of equal roots.
In sub cases (B) and (C) the second solution has a logarithmic term. In the next slide we look at two
examples that illustrate sub-cases (B) and (C).

Two illustrative examples - From the book of Earl. Rainville, Elem. Diff. Eqns pp.
338-341
16. Determine the indicial equation for the ODE

xy 00 − (4 + x)y 0 = 2y = 0.

The roots differ by an integer and using either root one may proceed but the one obtained from
the smaller root is a multiple of the solution obtained by using the larger root.
17. Consider the ODE
x2 y 00 + x(1 − x)y 0 − (1 + 3x)y = 0.
Check that the indicial equation has roots ±1.
We describe the procedure for finding the two solutions. Find the recurrence relation without putting
in the value of c and call the resulting Frobenius series y(x, c). Denote the LHS of the ODE by Ly.
Then
L(y(x, c)) = a0 (c − 1)(c + 1)y(x, c).
Show that the choice a0 = c + 1 leads to the equation

L(y(x, c)) = (c − 1)(c + 1)2 y(x, c).

Differentiate and put c = −1 and we get the second solution. What happens if we put c = −1 in
y(x, c)?

Second Solution of Bessel’s Eqn. of Order Zero Find the second linearly independent solution
of the Bessel’s equation of order zero. Well, first substitute the Ansatz
∞
X
y(x, ρ) = an xn+ρ
n=0

into the ODE and determine the successive coefficients without setting ρ = 0 so we get a function of
x as well as ρ with an a0 floating around. When substituted into the ODE we get

L(y(x, ρ)) = a0 xρ ρ2 . (1)

If we take ρ = 0 we get the good old J0 (x). But we can differentiate both sides with respect to ρ and
then put ρ = 0. We would get a second solution. Find this second solution which involves a logarithm.
It may be useful to rethink about the Cauchy Euler equation with repeated roots and the cause for
the appearance of logarithms,

32
 18. Prove the mean value theorem for integrals: Suppose f, g are continuous on [a, b] and g > 0 on
(a, b) show that there is a c ∈ (a, b) such that
Z b Z b
f (x)g(x)dx = f (c) g(x)dx.
a a

Hint: First prove that if f is continuous

Z b
f (x)dx = f (c)(b − a), for some c ∈ (a, b).
a

Now use the integral of g over [a, x] as a variable of integration.

√
19. Let u(x) = kxJn (kx). Show that u satisfies the ODE
  n2 − 1 
u00 = − k 2 − 4
u
x2

The last equation suggests that√when x is very large u(x) must behave like the sine function and
Jn (kx) must behave like sin kx/ kx and as such must have infinitely many zeros. We shall see that
this is indeed so if k > 1. The last condition can be removed later.
20. Let v(x) = sin(x − a). Show that

d   n2 − 1 
(vu0 − uv 0 ) = −uv k 2 − 1 − 4
dx x2

21. Let a be so large that k 2 − 1 − (n2 − 1/4)/x2 > 0 on [a, a + π]. Integrate the equation obtained
in the previous exercise over [a, a + π] and use the MVT for integrals. So for some c ∈ (a, a + π)
we have Z a+π   n2 − 1 
2 4
−(u(a + π) + u(a)) = u(c) v(x) k − 1 − 2
dx
a x

Zeros of Bessel’s Functions: Thus we see that u(a), u(c) and u(a + π) cannot all have the same
sign. Thus u must have a zero in every interval (a, a + π) for all a >> 1. We have proved,

Theorem 2.7: For k > 1, the function Jn (kx) has infinitely many zeros for each n ≥ 0.

Question: Explain why the condition k > 1 can be replaced by k = 1 or even k > 0? We shall see
an application of this theorem to the theory of wave propagation. Another interesting proof via the
integral representation is on pp. 76 - 78 of D. Jackson, Fourier series and orthogonal polynomials,
Dover, New York, 2004. See also G. N. Watson, Treatise on the theory of Bessel functions, p. 500 ff
for a discussion of the techniques used by L. Euler and Lord Rayleigh to compute the zeros of Jp (x).
22. Prove Laplace’s second integral representation:
1 π
Z √
Pn (x) = (x + x2 − 1 cos θ)−n−1 dθ
π 0
Hint: A suitable change of variables in the first integral formula would do the job.

33
 23. Use Laplace’s integral representation to prove the following result of C. Neumann (1862)
  x 
lim Pn cos = J0 (x)
n→∞ n

24. Writing the product of the integrals for J0 (x) as a double integral show that (I. N. Sneddon, p.
145)
1 π
Z
2
(J0 (x)) = J0 (2x sin θ)dθ.
π 0
Deduce the power series expansion of (J0 (x))2 .

Sneddon also gives a formula for Jn (x)Jm (x).

Hint: The integral over the square [0, π]×[0, π] equals twice the integral over the triangle 0 ≤ θ +φ ≤ π
and θ ≥ 0, φ ≥ 0. Now use further symmetries and write the integral over this triangle as one-fourth of
the integral along the square with vertices (π, 0), (0, π), (−π, 0) and (0, −π). Now change coordinates.
See Watson, pp. 31-32.

34
 III - Sturm-Liouville Problems and PDEs. The circular membrane:

As our first example of a boundary value problem, let us consider a circular membrane clamped
along its rim and set into vibration. The mean position being along the x − y plane and the origin at
the center of the membrane. At time t let the displacement from the mean position be z(x, y, t). It is
well known (see Kreyszig, pp 616-618) that z satisfies the wave equation
 ∂ 2z ∂ 2z  ∂ 2z
c2 + = 2 (1)
∂x2 ∂y 2 ∂t
where c denotes the wave speed. We seek a special solution of the form

z = (A cos pt + B sin pt)u(x, y). (2)

More general solutions can then be determined by superposition. Substituting the Ansatz (2) in the
PDE we get
c2 (A cos pt + B sin pt)∆u = −p2 (A cos pt + B sin pt)u
from which we conclude that u must satisfy the equation

Helmholtz Equation or the Reduced Wave Equation

∆u + k 2 u = 0,

where k = p/c. This equation is known as the Helmholtz’s equation or the reduced wave equation.

Exercises:

1. Write the Laplace operator ∆ in plane polar coordinates.

2. Write the Laplace operator in R3 in spherical polar coordinates. Computation gets very UGLY
unless you use some cleverness.

The Helmholtz’s equation in polar coordinates reads

∂ 2 u 1 ∂u 1 ∂ 2u
+ + + k 2 u = 0.
∂r2 r ∂r r2 ∂θ2
We now separate the radial and angular variables by setting

u(r, θ) = v(r) cos nθ,

where n must be an integer owing to 2π periodicity.

35
Exercise:
3. The function v(r) satisfies
r2 v 00 + rv 0 + (k 2 r2 − n2 )v = 0
Check that this is the Bessel’s equation after a rescaling of the variable r.
The indicial for the equation is ρ2 − n2 and only the positive index gives a solution which is finite
at the origin. Since r = 0 corresponds to the center of the membrane which always remains at finite
distance, the only physically tenable solution is Jn (kr).
Thus we see that our special solution is (recalling p = ck),

z(x, y, t) = Jn (kr)(A cos ckt + B sin ckt) cos nθ.

Since the membrane is clamped along the rim we see that the solution vanishes along r = 1 for all
values of θ and t. Thus the following boundary condition must be satisfied:

Jn (k) = 0.

Thus the frequency k must be a zero of the Bessel’s function Jn and we have seen that there is an infinite
list of them. The frequencies therefore form a discrete set of values. The most general solutions are
then obtained from superpositions whose coefficients are determined via initial conditions and Fourier
Analysis. We illustrate this by means of an example where the oscillations are radial.

Radial vibrations of the circular membrane Suppose the initial p conditions, the value of z(x, y, 0)
as well as zt (x, y, 0), are radial functions (that is depends only on x2 + y 2 then so would the solutions.
Thus the term cos nθ would disappear (that is n = 0) and we have a sequence of solutions

J0 (kr)(A cos ckt + B sin ckt),

where k runs through the discrete set of zeros of zeros of J0 (x) say ζ1 , ζ2 , ζ3 , . . . . The most general
solution is then ∞
X
z(r, t) = J0 (ζj r)(Aj cos cζj t + Bj sin cζj t)
j=1

Setting t = 0 in z(x, y, t) as well as zt (x, y, t) we get the following pair of equations for determining
the coefficients Aj and Bj :
∞
X
z(r, 0) = Aj J0 (ζj r)
j=1
X∞
zt (r, 0) = Cj J0 (ζj r)
j=1

where Cj = jBj ζj c.
In order to proceed further we need a result from Analysis called the Bessel expansion theorem.
We are obviously not equipped to prove this. The result is available for example in chapter 18 of
the authoritative work G. N. Watson, Treatise on the theory of Bessel functions, Second edition,
Cambridge University Press, 1958. See the historical introduction on pp. 577 - 579.

36
Theorem 3.1 (Bessel expansion theorem): Suppose f (r) is a smooth function on [0, 1] then it
admits a Fourier-Bessel expansion
∞
X
f (r) = Aj J0 (ζj r)
j=1

The coefficients Aj are uniquely determined by the formula (due to Lommel)

Z 1
2
Aj = rf (r)J0 (ζj r)dr.
(J1 (ζj ))2 0

So the coefficients Aj and Bj appearing in the solution of the vibration problem can be recovered from
Lommel’s formula applied to the initial conditions z(r, 0) and zt (r, 0).

Orthogonality properties of Bessel’s functions:

Exercises:
d
4. Write the Bessel’s ODE in self adjoint form. Check that the operator x dx is scale invariant.

5. Put φu (x) = Jp (xu) and check that

 d  d 
x x φu (x) + (x2 u2 − p2 )φu (x) = 0.
dx dx

6. Fix p ≥ 0 and ζ1 , ζ2 , ζ3 , . . . be the list of zeros of Jp (x). Show that the family {Jp (ζj x) : n =
1, 2, 3, . . . } is orthogonal over the interval [0, 1] with respect to the weight function x. Warning:
The cases p = 0 and p > 0 have to be dealt with separately.

There remains the computation of Z 1

x(Jp (ζj x))2 dx
0

7. Let ζ be a zero of Jn (x). Multiply by 2xζJn0 (ζx) the ODE satisfied by Jn (ζx) and deduce that
Z 1
2 (Jn (ζx))2 xdx = (Jn0 (ζ))2 = (Jn+1 (ζ))2
0

8. Deduce the formula of Lommel. We are not proving the Bessel expansion theorem. Only that if
the expansion exists and the functions involved are smooth we are deriving the formula for the
coefficients in a formal way.

9. Determine the Bessel expansion for the constant function 1. Hint: Use (xp Jp (x))0 = xp Jp−1 (x).
∞
X 2Jn (ζj x)
10. Show that xn = .
j=0
ζj J n+1 (ζj )

A very interesting proof of the orthogonality property of the Bessel functions suggested by physical
considerations is available on pp 324-325 of Lord Rayleigh, Theory of Sound, Vol - I, Dover, 1945. We
shall see yet another proof at the end of this chapter.

37
Fourier-Legendre Series This is a result similar in spirit to the Fourier Bessel expansion. Rather
than state the theorem we give an example due to Lord Rayleigh, Theory of sound, Volume - II, p. 273.
∞ r
itx
X
n π
e = (2n + 1)i J 1 (t)Pn (x).
n=0
2t n+ 2

Prove by induction
1  t  12 +n 1
Z
Jn+ 1 (t) = √ (1 − x2 )n cos txdx
2 πn! 2 −1
Formally deduce the result of Lord Rayleigh. This expansion appears in connection with scattering of
plane waves by a spherical obstacle.

Uniqueness of the solution obtained We shall now demonstrate that the solution to the problem
of vibrating membrane is unique. Thus the solution we have obtained completely settles the matter.
Suppose z1 and z2 are two solutions satisfying the same initial conditions and boundary conditions
(namely vanishing along the boundary of the membrane r = 1). Then the difference Z = z1 − z2 also
satisfies the PDE with zero initial and boundary conditions:
∂ 2Z
c2 ∆Z − = 0,
∂t2
Z(x, y, 0) = 0, Zt (x, y, 0) = 0,
Z(cos θ, sin θ, t) = 0.

The Energy Method The idea of proof is well-known under the name of energy method. The
method is frequently employed in the analysis of PDEs. Multiply the differential equation by Zt and
integrate over the disc D : x2 + y 2 ≤ 1. Integration by parts gives (how?)
ZZ ZZ
d 2 2
Z dxdy + c (Zxt Zx + Zyt Zy )dxdy = 0.
dt D t D

¿From this we infer that the energy

ZZ
E(t) = (c2 Zx2 + c2 Zy2 + Zt2 )dxdy = 0.
D

is constant in time. Since E(0) = 0 we see that Z is a constant and so is zero. The proof is complete.
11. Imitate the energy method to show that the twice continuously differentiable solution to the
initial-boundary value problem
ut = ∆u, u(x, y, 0) = f (x, y), u(cos θ, sin θ, t) = 0
for the heat equation is unique. Hint: Here the energy function is monotone decreasing in time.
12. Show that the twice continuously differentiable solution to the boundary value problem
∆u = 0, on D
u(x, y, z) = f (x, y, z), on ∂D

38
 is unique where D is a region in R3 with a smooth boundary ∂D and f is a function prescribed
along ∂D.

For more detailed discussion on these types of wave phenomena see

1. Courant and Hilbert, Methods of Mathematical Physics, Volume I. We have already cited this.
For the discussion of vibration of a circular plate (which is more involved than the membrane)
see pp. 307-308. These involve Bessel’s functions of imaginary orders ip (p > 0), known as the
modified Bessel’s functions.

2. Lord Rayleigh, Theory of Sound, Volume - I. This is a comprehensive account of the theory of
vibrations. See particularly the long and detailed discussion on vibrating plate. This is still the
best source on the Physics of Vibrations.

3. For other applications such as the skin effect. see F. Bowman, Introduction to Bessel’s functions,
Dover.

4. The Bessel functions also appear in optics. The radii of the successive interference fringes due
to diffraction from a circular aperture are given in terms of the zeros of Bessel’s function.

5. Besides the authoritative treatise of G. N. Watson mentioned earlier, the book of Byerly is
particularly recommended especially the historical sketch at the end. The book is available on
line.

The Vibrating String and regular Sturm Liouville Problems: Suppose we have a string of
length l and density ρ(x) stretched along the line segment [0, l] and clamped at its two ends. The
string is set into vibrations and the governing differential equation my be reduced to

y 00 + λρ(x)y = 0 (1)

where λ is a parameter. Since the string being clamped at its ends we have the boundary conditions
(BC)
y(0) = y(l) = 0. (2)
A λ for which a non-trivial solution exists is called the eigen-value of the B. V. Problem and the
corresponding non-trivial solution the eigen-functions. The BC (2) is referred to as the Dirichlet
boundary conditions. Other physical conditions lead to BC of the form

y 0 (0) = y 0 (l) = 0. (3)

known as the Neumann BC. Later we shall see another BC that has some different features. We shall
assume that ρ(x) is continuous and positive.
As a simple example let us consider a uniform string of unit density and unit length. The problem
to be solved is
y 00 + λy = 0, y(0) = 0 = y(1).
Solving the ODE (λ = 0 doesnt give non-trivial solutions),
√ √
y(x) = A cos λx + B sin λx

39
The condition y(0) = 0 immediately gives A = 0 and the condition y(1) = 0 gives
√
B sin λ = 0.

Since the solution in question is non-trivial, B 6= 0 which means

√
λ = ±π, ±2π, . . .

Thus the solution to the boundary value problem exists only for a discrete set of values of the parameter
and the eigen-values form a discrete set.

Theorem 3.2 (Orthogonality of Eigen Functions): For the boundary value problem on [0, l]

y 00 + λρ(x)y = 0

with either Dirichlet or Neumann boundary conditions the eigen-functions corresponding to distinct
eigen-values are orthogonal on [0, 1] with respect to the weight function ρ(x) namely,
Z l
u(x)v(x)ρ(x)dx = 0.
0

Proof: Suppose λ and µ are distinct eigen-values with eigen-functions u and v. Proof is similar to
the orthogonality of Legendre Polynomials. We have the pair of equations:

u00 + λρ(x)u = 0, u(0) = 0 = u(l)

v 00 + µρ(x)v = 0, v(0) = 0 = v(1).

Multiply the first by v, second by u, integrate over [0, l] by parts and subtract. Details left to the
student.

Exercise:
13. Discuss the orthogonality of eigen-functions for the Neumann BC.

Theorem 3.3 (Simplicity of the eigen-values): To each eigen-value of the BVP (1)-(2) with
Dirichlet BC, there is only one eigen-function upto scalar multiples.

Proof Assume that u and v are two linearly independent eigen-functions with the same eigen value
then the linear span of u and v is the set of all solutions of the homogeneous ODE

y 00 + λρ(x)y = 0

and since both u and v vanishes at 0, it follows that every solution of the ODE vanishes at 0. This is
plainly false since the solution of the initial value problem for this ODE with initial conditions

y(0) = 1, y 0 (0) = 0

which exists by Picard’s theorem, does not vanish at 0. Contradiction.

40
Completeness of eigen-functions We now state a theorem on the completeness of the set of
eigen-functions of the Sturm-Liouville problem

y 00 + λρ(x)y = 0, y(0) = 0 = y(l).

We assume that the function ρ(x) is positive and continuous on [0, l].
The theorem is often stated under stronger hypothesis 7 . One can also discuss mean convergence but
we shall not do so here.

Theorem 3.4: For the Sturm Liouville problem in question, there is an infinite sequence of eigen-
values
λ1 < λ2 < λ3 < . . .
tending to infinity. The set of eigen-functions {φn : n = 1, 2, . . . } is complete in the following sense.
Each Lipschitz function f (x) on [0, l] can be expanded as a series
∞
X
f (x) = cj φj (x)
j=1

which converges for each x ∈ (0, l). The coefficients are given by
Z l  Z l −1
cj = f (x)φj (x)ρ(x)dx φ2j (x)ρ(x)dx
0 0

Exercises:

14. Show that the eigen-values of

y 00 + λρy = 0
with boundary conditions y(0) = 0 = y(1) are positive real numbers.

15. Determine the eigen-values and eigen-functions of the Sturm-Liouville problem

y 00 − 2y 0 + (1 + λ)y = 0

with boundary condition y(0) = 0, y(1) = 0.

16. Determine the eigen-values and eigen functions of the Sturm-Liouville problem

x2 y 00 + xy 0 + λy = 0

on [e, 1/e] with the periodic boundary conditions:

y(1/e) = y(e), y 0 (1/e) = y 0 (e).

7
R. Courant and D. Hilbert, Methods of Mathematical Physics, Volume - I, p. 293. However E. C. Titchmarsh in
his Eigen-function expansions associated with second order differential equations, Volume - I, Oxford, Clarendon Press,
1969, proves it with substantially weaker hypothesis (see page 12).

41
 17. Show that the eigen-values of the boundary value problem

y 00 + λy = 0, y(0) = 0, y(1) + y 0 (1) = 0

are given by λ = k 2 where k satisfies tan k + k = 0. Graphically show that there are infinitely
many roots.
To see how this type of BC appears in physical problem, see the example on p. 117, §67 of
Byerly’s text.

18. A rigid body is rotated with uniform and fixed angular speed about an axis that is not specified.
How would one choose the axis of rotation so as to maximize the rotational Kinetic Energy?
Formulate the problem mathematically.

19. Variational principles underlying eigen-values and eigen vectors/functions. Suppose A is a n × n

real symmetric matrix, show that the maximum and the minimum of the quadratic function

hAx, xi, x21 + x22 + · · · + x2n = 1

are both attained at eigen-vectors and the maximum and minimum values are the largest and
smallest eigen-values of A.

Proof of the Min/Max properties of Eigen-values Let Q(v) = vT Av. This is a quadratic
polynomial in n-variables and as such attains its minimum value at some point say v1 on the unit
sphere in Rn . We now perturb the vector v1 to say

w = (v1 + h)/kv1 + hk

and compare the values of the quadratic at the two points namely

Q(v1 ) ≤ Q(w), for all  small enough.

Thus, we get after cross multiplying by kv1 + hk2 , the following inequality valid for all small  positive
or negative !
Q(v1 )kv1 + hk2 ≤ (v1 + h)T A(v1 + h).
Expanding and canceling off Q(v1 ) we get

2Q(v1 )hT v1 ≤ 2hT Av1 + 2 (hT Ah − Q(v1 )khk2 ).

We now divide by  and let  → 0. Since  may have either sign we get the pair of inequalities

hT (Q(v1 )v1 − Av1 ) ≤ 0, and hT (Q(v1 )v1 − Av1 ) ≥ 0.

Thus we conclude hT (Av1 − Q(v1 )v1 ). Since h is arbitrary we see that

Av1 = Q(v1 )v1

In other words the minimum is attained at an eigen-vector and the minimum value is the corresponding
eigen-value.

42
Exercise: Examine carefully the computations and explain where have we used the fact that A is a
symmetric matrix?
To proceed further, let S be the intersection of the unit sphere in Rn with the hyperplane

v · v1 = 0.

This is also a closed bounded set and the minimum of Q(v) on S is attained at say v2 and the
corresponding Q(v2 ) is not less than Q(v1 ) (why?).
Now we perturb v2 to
w = (v2 + h)/kv2 + hk
where h is chosen such that h · v1 = 0. Again Q(v2 ) ≤ Q(w). Multiplying the inequality through by
kv1 + hk2 , expanding out and canceling Q(v2 ) we get as before

2hT (Q(v2 )v2 − Av2 ) ≤ 2 (hT Ah − Q(v2 )khk2 ).

Dividing by  and letting  → 0 we get the pair of inequalities from which we again deduce

hT (Q(v2 )v2 − Av2 ) = 0.

This is now valid for all h such that h · v1 = 0. But this is also holds for h parallel to v1

Exercise: Verify the last statement.

Hence Q(v2 )v2 − Av2 = 0.
Thus Q(v) attains its minimum over S at an eigen-vector v2 . The minimum value is the corre-
sponding eigen-value. By construction v2 is orthogonal to v1 . Now minimize Q(v) over the intersection
of the unit sphere in Rn with the pair hyperplanes

v · v1 = 0, v · v2 = 0

and the rest of the proof simply writes itself out. The process terminates after we have n orthogonal
eigen-vectors of our matrix A. We have proved

Theorem 3.5 (Spectral Theorem): A real symmetric matrix has an orthonormal basis of eigen-
vectors.
The analogue of the above result for self-adjoint Diff. Eqns with Dirichlet BC is a serious matter
that has lead to a huge corpus of mathematical research - for these mathematical development against
a historical back-drop see the introductory parts of R. Courant: The Dirichlet’s principle, conformal
mappings and minimal surfaces, Dover Reprint, 2005. See the free preview of the first three pages of
introduction on the Internet!
Let us consider the problem of minimizing the “energy”
Z 1
(y 0 (t))2 dt (1)
0

subject to the condition Z 1

y(t)2 ρ(t)dt = 1. (2)
0
where y(t) ranges over continuous piecewise smooth functions with y(0) = y(1) = 0 and ρ(x) is a
positive continuous function on [0, 1].

43
The Dirichlet Principle: The minimization problem (1)-(2) has a twice continuously differentiable
solution which corresponds to the smallest eigen-value of the Sturm-Liouville problem

y 00 + λρ(x)y = 0, y(0) = 0 = y(1). (3)

The principal difficulty is in showing

(i) That the minimizer exists and
(ii) The minimizer is twice continuously differentiable.
These are rather deep waters. Riemann used these ideas motivated by potential theoretic consid-
erations to prove his celebrated theorem in complex analysis known today as The Riemann Mapping
Theorem.
20. Verify this formally. That is to say, if y0 (x) is a twice continuously differentiable minimizer then
y0 (x) is an eigen-function of the Sturm-Liouville problem.

21. Try to give an intuitive (non-rigorous) geometric argument that the eigen-function corresponding
to the smallest eigen value has no zeros in (0, 1). In physics books this is referred to as the
Fundamental Mode.

22. Assuming that y0 is the fundamental mode, we seek to minimize (1) subject to the condition (2)
as well as Z 1
y(t)y0 (t)ρ(t)dt = 0. (20 )
0
then we get an eigen function with eigen value larger than the first one. Why is this so? Why is
it the case that this eigen-function has at least one zero in (0, 1)? Is it geometrically clear that
there is precisely one zero in (0, 1)?

23. How does one continue to construct the successive eigen functions?

Sturm’s comparison theorem

24. Suppose ρ and σ are both continuous positive functions on [0, 1] and ρ(x) > σ(x) for every
x ∈ [0, 1]. Let y(x) and z(x) be solutions of the pair of ODEs

y 00 + ρ(x)y = 0, z 00 + σ(x)z = 0.

Prove that between two successive zeros of z(x) there is at least one zero of y(x).
Hint: Suppose a and b are successive zeros of z(x) in [0, 1] and we may assume z(x) > 0 on
(a, b). Suppose that y(x) does not vanish on (a, b) and is positive there. Imitate the proof of
orthogonality of eigen-functions but on [a, b] and show that a sum of positive things adds up to
zero.
See the commentary in Lord Rayleigh: Theory of Sound - I, pp. 217-222.
Returning to the Sturm-Liouville problem y 00 + λρ(x)y = 0 with Dirichlet BC at 0 and 1, let us
consider the solution y(x, λ) of the ODE with initial conditions

y(0) = 0, y 0 (0) = 1.

For small values of λ the solution y(x, λ) has no zeros on (0, 1] and for a suitably large value of λ there
is a unique zero in this interval. Use the Sturm’s comparison theorem.

44
 We must now select the value of λ such that the boundary condition at x = 1 is satisfied. In other
words we must select λ such that
y(x, λ) = 0.
We need to show that the solution to this equation for x as a function of λ exists and is continuous.
The proof of these assertions rely on two theorems of analysis namely the existence uniqueness of IVP
and the implicit function theorem. Thus we see that there is a value λ1 such that

y(1, λ1 ) = 0

and we have the first eigen value and eigen function. Increasing the value of λ and invoking the same
principles again we get a λ2 > λ1 such that y(x, λ2 ) is the second eigen function with eigen value λ2
and this has exactly one zero in (0, 1).

Synopsis of chapter III We started off by solving the wave equation on a disc of unit radius
governing the vibrations of a circular membrane. The membrane is clamped along the rim which
means the solution is zero on boundary. Separating off the time variable results in the Helmholtz’
PDE. The circular symmetry makes it natural to employ polar coordinates and after separating radial
and angular variables we see that the radial part satisfies the Bessel’s equation of integer order. The
fact that the membrane remains at finite distance singles out Jn (kr) as the relevant solution. The
other boundary condition forces the parameter k to run through a discrete set of values, the zeros of
Jn (x). We prove the infinitude of these zeros and we get a sequence of basic solutions from which
one can obtain general solutions by superposition. Thus we have solved a singular Sturm-Liouville
problem. For simplicity we assumed that the vibrations are radial which fixes n = 0.
We next take up the case of a vibrating string clamped at its ends. The wave equation in question
is
ρ(x)utt − uxx = 0
Substituting the Ansatz u(x, t) = y(x)φ(t) in the equation results in

φ00 (t) y 00 (x)

= .
φ(t) ρ(x)y(x)

Since the left hand side is a function of t alone and the right hand side is a function of x alone both
sides are constant say −p2 . Thus we get φ(t) = A cos pt + B sin pt whereas for y(x) we get

y 00 + p2 ρ(x)y = 0.

The physical condition that the string is clamped at the ends leads to the Dirichlet Boundary conditions
y(0) = 0 = y(l). This BVP is called a regular Sturm-Liouville problem. As in the case of a circular
membrane, the set of values of the parameter p for which a non-trivial solution exists (known as
eigen-functions) is discrete and these are called eigen-values.
We prove the simplicity of eigen-values and the orthogonality of eigen-functions corresponding
to distinct eigen-values. The Sturm-Liouville problem is then given a variational formulation. The
underlying principle is quite general and has wide applicability. The existence of the minimizer and its
regularity is known as the Dirichlet principle. We also discuss an alternative approach to the existence
of eigen-values and eigen-functions for the one-dimensional problem by using the Sturm’s comparison
theorem. The completeness of eigen-functions is a deep result in analysis and we can only state it here

45
(both for the one-dimensional case of a vibrating string and the two dimensional case of a vibrating
circular membrane).
I mention in passing that the case of an elliptical membrane has been considered by Émile Léonard
Mathieu in 1868 and the resulting ODE known as the Mathieu equation:

y 00 + (a + b cos 2x)y = 0.

where b is given and a is an eigen-parameter. The equation has led to a long and rich chapter in the
theory of analytic ODEs, generalized and studied by G. W. Hill in 1886 in his researches on Lunar
motion. Unfortunately we are not in a position to say anything about these exciting theory in this
elementary course!
Hope this thumb-nail sketch helps!!

46
 IV - Fourier series and partial differential equations

The boundary value problem on [−π, π] with periodic boundary conditions:

y 00 + λy = 0, y(−π) = y(π), y 0 (−π) = y 0 (π).
leads to the theory of Fourier series. The study began long before Joseph Fourier in connection with
a physical problem by Daniel Bernoulli in 1747 of a vibrating string. The method became popular
after Fourier’s work on heat conduction (circa 1807). Fourier Analysis has been the subject of intense
research and occupies today a distinguished position in Mathematics.
Study of the paper by Roger Cooke, Uniqueness of trigonometric series and descriptive set theory
1870-1985, Archiv for the history of exact sciences 45 281-334 (1993) would be an excellent project
for any serious mathematically inclined student.
For the boundary value problem displayed in the previous paragraph, it is evident that the eigen-
values are 0, 1, 4, . . . . Each eigen value other than 0 has two linearly independent eigen-functions
namely
sin nx, cos nx.
There is only one eigen-function with eigen-value 0 namely the constant function 1.

The Basic Issue: To develop a “fairly general function” f (x) as a series:

∞
X
f (x) = a0 + (an cos nx + bn sin nx) (1)
n=1

We call the partial sum a0 + N

P
n=1 (an cos nx + bn sin nx) as a trigonometric polynomial. If in addition
|aN | + |bN | =
6 0 we say it is of degree N .

Formula for Fourier coefficients Proceeding formally let us multiply equation (1) by cos mx and
integrate term by term over [−π, π] and we find
1 π
Z
am = f (x) cos mxdx.
π −π
Likewise we find Z π
1
bm = f (x) sin mxdx.
π −π
For the case of a0 we have Z π
1
a0 = f (x)dx.
2π −π
The Fourier series is by definition the series in (1) with the coefficients given by (2), (3) and (4).
Question is now the validity of the equation (1).

47
Dirichlet’s theorem The completeness theorem of the previous chapter has an analogue known as
Dirichlet’s theorem. Before stating the result we shall insert here a definition.

Definition A function f : [a, b] −→ R is said to be piecewise continuously differentiable if there

exists a partition a = t0 < t1 < · · · < tn = b such that f is continuous on each [tj , tj+1 ], differentiable
on (tj , tj+1 ) and the derivatives have finite limits at the ends tj and tj+1 , j = 0, 1, 2, . . . , n − 1.
The Heaviside unit step function u(t) is piecewise continuously differentiable and so is

u(t) cos t + u(1 − t) sin t.

Any Lipschitz function is piecewise continuously differentiable.

Theorem 4.1 (Dirichlet (1829)): Suppose f : [−π, π] −→ R is piecewise continuously differen-

tiable then we have the development
∞
X
f (x) = a0 + (an cos nx + bn sin nx), x 6= ±π
n=1

at points of continuity, where a0 , an and bn are given by

Z π
1 π 1 π
Z Z
1
a0 = f (x)dx, an = f (x) cos nxdx, bn = f (x) sin nxdx.
2π −π π −π π −π

At a point of discontinuity x0 or x0 = ±π, the series converges to

1
(f (x0 +) + f (x0 −)).
2
The partial sums of the displayed series are trigonometric polynomials.
The series displayed in the last slide is called the Fourier Series of the function f (x) and the
sequences an , bn the Fourier coefficients of f (x). The proof of this theorem cannot be given here.
Proof can be found on pp 71-73 of Richard Courant and David Hilbert, Methods of Mathematical
physics, Volume - I, Wiley (Indian Reprint), 2008. Or p 56 ff. of the old but nevertheless excellent
book Byerly, See particularly the brief commentary on §38 on page 62 and the detailed summary on
pp. 267-274 and further references therein.

The issue with continuous functions Question: If f (x) is merely assumed to be continuous does
the above result hold ? This was believed to be so by several mathematicians including Dirichlet until
Paul Du Bois Reymond after several abortive attempts at proving it, produced a counter example in
1875 ! Using ideas from set topology one can show that a majority of continuous functions display
such errant behaviour.The simplified proof given by Stephan Banach is available in most texts.
The assumption of piecewise smoothness may be slightly weakened though. In applications one
takes a periodic extension of f (x) with period 2π:

f (x + 2π) = f (x), x ∈ R.

In case f (π) 6= f (−π) then we shall always redefine the value at these points as the arithmetic mean.
This convention w ill be followed throughout.

48
 1. Compute the Fourier coefficients of the saw-tooth function f (x) = x, −π < x < π. When
extended as a 2π periodic function you see a jump discontinuity of 2π at each of the points
±π, ±3π, . . . . Draw a graph of the function.
2. Determine the Fourier coefficients of the triangular wave train
f (x) = π − |x|, −π ≤ x ≤ π
and sketch the graph of its 2π periodic extension. Use Dirichlet’s theorem to deduce the value
of
1 1
1 + 2 + 2 + ....
3 5
Deduce the value of
1 1
1 + 2 + 2 + ...
2 3
3. Determine the Fourier coefficients of the square wave train which is given by f (x) = π on the
range (0, π), f (0) = 0, extended as an odd function on (−π, π) and then as a 2π periodic function
on the whole real line.

Partial fraction expansion of cosecent and cotangent

4. Determine the Fourier coefficients of f (x) = cos ax where a ∈
/ Z and deduce that
∞
1 X 2a(−1)n
πcosec(πa) = +
a n=1 a2 − n2
∞
1 X 2a
π cot(πa) = +
a n=1 a2 − n2

5. Let ζ(2k) = ∞ −2k

P
m=1 m . Find the generating function f (z) for the sequence {ζ(2k)} and write
the result in terms of exponentials.
6. By comparing 2f (z) − 1 + iπz with the generating function for Bernoulli numbers derive the
formula of Euler for
1 1
1 + 2k + 2k + . . .
2 3
Remark: See pp. 277-281 of Alain Robert, Advanced calculus for users, North Holland, 1989.

The Riemann Lebesgue Lemma This important result states that the Fourier coefficients of a
continuous function tends to zero as n → ∞. The continuity assumption can easily be weakened but
we shall not strive for the most general results.
7. Suppose f (x) = xk on the interval [−π, π]. Show that the Fourier coefficients an and bn tend to
zero as n → ∞.
8. Deduce that for a polynomial P (x), the Fourier coefficients tend to zero as n → ∞.
9. Suppose f (x) is continuous on [−π, π] extend it continuously as a constant on (−∞, −π] and
[π, ∞). In the integral defining an perform the change of variables x = y + πn and prove that
an −→ 0 as n → ∞. The proof that bn −→ 0 as n → ∞ is similar.

49
 Remark: This elegant argument goes back to Riemann himself. Modern proofs take a different
approach via approximations as we shall see presently.

Question: Suppose we have sequences {an } and {bn } such that they both tend to zero as n → ∞ is
it true that there is an integrable function whose Fourier coefficients are precisely these?
This question was answered relatively late in the development of Fourier series. An example appears
in a letter dated 1910 by N. Luzin to Florenskii.

Example: The series

∞
X sin nx
n=1
log(n + 1)
is NOT the Fourier series of any integrable function on [−π, π]. See pp. 309-310 of Roger Cooke’s
article.

The Weierstrass’s Approximation theorem We now state one of the most fundamental theorems
in analysis whose proof may be found in Courant and Hilbert’s book cited earlier (pp. 65-68). Since
its publication in 1885, Several different proofs have been supplied for this result.

Theorem 4.2 (Weierstrass’s Approximation theorem): Suppose f : [a, b] −→ R is continuous

then given any  > 0 there exists a polynomial P (x) such that

|f (x) − P (x)| < , x ∈ [a, b].

A charming commentary on the life of Weierstrass and the impact of this theorem on modern analysis
is the article A. Pinkus, Weierstrass and approximation theory, Journal of Approximation Theory 107
1-66 (2000)

10. Prove the Riemann-Lebesgue Lemma using the Weierstrass’s approximation theorem.

11. Show that if f (x) is continuously differentiable and the second derivative has only finitely many
jump discontinuities then the Fourier coefficients decay to zero faster than 1/n2 . One can then
show (using analysis) that once term by term differentiation is valid.

12. Imitate the proof of Riemann Lebesgue Lemma to show that

√
|J0 (x)| ≤ c/ x,

for some constant

√ C. What about Jn (x)? Deduce that the Bessel’s functions of integer orders
decay like 1/ x for large x.

Mean Approximation Suppose f : [a, b] −→ R is a continuous function then its L2 norm is defined
to be s
Z b
1
kf k2 = |f (t)|2 dt
b−a a
Engineers are accustomed to calling it the Root Mean Square. We can define this for Riemann integrable
functions as well but the word “norm” would be somewhat inappropriate in this context. The reason

50
is that if we look at the function g(x) which is zero at all but finitely many points then kgk2 = 0 but
g is not the zero function.
There is a way to get around this annoyance but the correct setting for this is the theory of Lebesgue
integrals. Indeed Lebesgue theory is the right backdrop for discussing Fourier Analysis and whatever
we discuss without its aid would necessarily be incomplete and at places would result in abuse of
language and notation. With this disclaimer we now continue.

Definition: We say that a sequence fn of square integrable functions converges to f in mean if

kfn − f k2 −→ 0, as n → ∞.

Exercise: Show that if f, g and h are Riemann integrable then

kf − gk2 ≤ kf − hk2 + kh − gk2 .

Theorem 4.3 (Least Square Approximation): Suppose f (θ) is a Riemann integrable function
on [−π, π] and Sn (f ) is the n−th partial sum of its Fourier series. For any arbitrary trigonometric
polynomial P (θ) of the degree at most n,
kf (θ) − sn (f )k2 ≤ kf (θ) − P (θ)k2
with equality if and only if P (θ) = sn (θ).

Exercises:
13. Prove the following theorem in Linear Algebra. Suppose V is a vector space endowed with
an inner product hu, vi and a, b are a pair of orthogonal vectors then we have the theorem of
Pythagoras:
kak2 + kbk2 = ka + bk2
14. Now show that f − sn and sn − P are orthogonal with respect to the inner-product
Z π
1
hf, gi = f (t)g(t)dt
2π −π
Hint: First compute hf − sn (f ), cos mxi for m ≤ n. Use Pythagoras’ theorem to prove the Least
square approximation theorem.

Theorem 4.4 (Bessel’s Inequality): Suppose f : [−π, π] −→ R is a Riemann integrable function

then n Z π
2 1X 2 2 1
a0 + (|aj | + |bj | ) ≤ |f (t)|2 dt, n ∈ N.
2 j=1 2π −π
15. Prove the above theorem. Hint: We have proved (f − sn (f )) ⊥ (sn (f ) − P ). Take P = 0.
Letting n → ∞ we conclude
∞ Z π
1X 1
a20 + (|aj |2 + |bj |2 ) ≤ |f (t)|2 dt.
2 j=1 2π −π

Now we shall invoke the Weierstrass’ approximation theorem to make the inequality into an equality
knows as The Parseval Formula. The transition is highly non-trivial using analysis.

51
Some preliminaries for the main result
16. Suppose that f (θ) is an even continuous function on [−π, π] show that given any  > 0, there is
a trigonometric polynomial P (θ) such that
|f (θ) − P (θ)| < , −π ≤ θ ≤ π.

17. Suppose g(θ) is an odd continuous function on [−π, π] vanishing at ±π show that there is a
trigonometric polynomial Q(θ) such that
|g(θ) − Q(θ)| < , −π ≤ θ ≤ π.
Hint: Apply the previous to g(θ) sin θ and then we see that
|g(θ)(1 − cos2 θ) − sin θQ1 (θ)| < .
Now repeat the above with g(θ) cos2 θ and we see that
|g(θ) − cos4 θg(θ) − Q2 (θ)| < .
Iterate sufficiently often.
Finally we have the ingredient we need:
18. Suppose f (θ) is a continuous function on [−π, π] such that f (π) = f (−π). Given any  > 0,
there is a trigonometric polynomial R(θ) such that
|f (θ) − R(θ)| < , −π ≤ θ ≤ π.

We are now ready to surge ahead and prove one of the fundamental results on Fourier Analysis.

Convergence in Mean and Parseval Formula

Theorem 4.5 (The Parseval Formula): Suppose f, g : [−π, π] −→ C are Riemann integrable
functions and sn (f ) is the n−th partial sum of its Fourier series for f then
1.
kf (θ) − sn (θ)k2 −→ 0, n → ∞.
2. ∞
Z π
1 1 X 0 00
f (θ)g(θ)dθ = a00 a000 + (a a + b0n b00n )
2π −π 2 n=1 n n
where a0n , b0n and a00n , b00n are the Fourier coeff. of f and g.
3. Taking f = g in particular (and so a0n = a00n = an etc.,) we have
Z π ∞
1 2 2 1X
|f (θ)| dθ = |a0 | + (|an |2 + |bn |2 ).
2π −π 2 n=1

We shall prove this for continuous functions f (θ) such that f (π) = f (−π). The transition to the
general case is a fairly routine matter that shall be indicated as a list of three (optional) exercises:
1. The main point in this transition is that an arbitrary Riemann integrable function can always
be approximated by continuous functions - in fact piecewise linear functions.
2. The continuous function can be altered at the two ends to ensure f (π) = f (−π) such that the
integrals (areas) are change by arbitrarily small amounts.

52
Proof of Parseval’s formula: Let  > 0 be arbitrary. Choose a trigonometric polynomial R(θ)
such that
kf (θ) − R(θ)k2 < .
Let n0 be the degree of this trigonometric polynomial. Then by the least square approximation
principle,
kf (θ) − sn (f )k2 ≤ kf (θ) − R(θ)k2 < , n ≥ n0 .
This proves the first part of the theorem. To prove the third part note that f (θ) − sn (θ) is orthogonal
to sn (θ) and so by theorem of Pythagoras,

kf k22 = kf − sn (f )k22 + ksn (f )k22

and letting n → ∞ we conclude

lim ksn k22 −→ kf k22
n→∞
Pn
But ksn k22 = |a0 |2 + 21 k=1 (|ak |
2
+ |bk |2 ) and the result is proved. The second part will be an exercise.

19. Prove the second part of Parseval’s theorem applying the simple formula

4AB = |A + B|2 − |A − B|2

to A = f + g and B = f − g

20. Determine the Fourier series for the function f (x) = x2 on [−π, π]. Use Parseval formula to show
that
1 1 π4
1 + 4 + 4 + ··· =
2 3 90
21. Using f (x) = π 2 x − x3 , determine the sum of the series
1 1
1+ + + ...
26 36

22. Verify the results of the last two exercises with the Formula obtained by Euler in terms of the
Bernoulli numbers.

Extending Parseval’s theorem to Riemann integrable functions Here are three OPTIONAL
exercises which shows how to get Parseval’s theorem for Riemann integrable functions.

23. Show that if f is a Riemann integrable function on [a, b] then given any  > 0 there is a continuous
function g on [a, b] such that kf − gk2 < .
Well, select a partition P : a = t0 < t1 < · · · < tn = b such that

U (f, P ) − L(f, P ) < .

Now let g be the continuous function whose graph on [tj , tj+1 ] is obtained by joining the points
(tj , f (tj )) and (tj+1 , f (tj+1 )).

24. Suppose g : [−π, π] −→ R is continuous, then given any  > 0, we can select a continuous
function h on [−π, π] such that h(π) = h(−π) and kg − hk2 < 

53
 25. Use triangle inequality

kf − sn (f )k2 ≤ kf − gk2 + kg − hk2 + kh − sn (h)k2 +

ksn (h) − sn (g)k2 + ksn (g) − sn (f )k2

and the Bessel’s inequality to complete the proof of the mean convergence of Fourier series.
Now use the fact that f − sn (f ) is orthogonal to sn (f ) and the Pythagoras’s theorem to deduce
the Parseval’s formula.

Descarte’s isoperimetric problem This is one more important classic variational principle which
goes back at least to René Descartes:

Theorem 4.6: Of all piecewise smooth closed curves with a given perimeter the circle encloses
maximum area.
One can turn the problem around by fixing the area and minimizing the perimeter.

Theorem 4.7: Of all piecewise smooth closed curves enclosing a given area the circle has the least
perimeter.
The theorem generalizes to higher dimensions in an obvious way.

Spacial isoperimetric theorem “With a little knowledge of the physics of surface tension, we could
learn the isoperimetric theorem from a soap bubble.
Yet even if we are ignorant of serious physics, we can be led to the isoperimetric theorem by quite primitive
considerations. We can learn it from a cat. I think you have seen what a cat does when he prepares himself
for sleeping through a cold night: he pulls in his legs, curls up, and, in short, makes his body as spherical as
possible. He does so obviously, to keep warm, to minimize the heat escaping through the surface of his body.
The cat who has no intension of decreasing his volume, tries to decrease his surface. He solves the problem
of a body with a given volume and minimum surface in making himself as spherical as possible. He seems to
have some knowledge of the isoperimetric theorem.” Quotation from p. 170 of
G. Polya, Mathematics and plausible reasoning, Princeton University Press, Princeton, 1954.

Hurwitz’ Proof of isoperimetric theorem (1902): Let the piecewise smooth closed curve be
parametrized by arc-length (x(s), y(s)), 0 ≤ s ≤ L and the curve is traced counter-clockwise. The
area A is given by I Z L
dy
A = xdy = x ds (How?)
0 ds
Let t = (2πs/L) − π so that t runs over the interval [−π, π]. Then
Z π
dy
A= x dt (1)
−π dt

For the perimeter we have

 dx 2  dy 2  ds 2 L2
+ = = .
dt dt dt 4π 2

54
which is conveniently rewritten as
π
L2
Z
1 n dx 2  dy 2 o
+ dt = 2 . (2)
2π −π dt dt 4π
We now apply the Parseval’s formula to (1) and (2). Let the n−th Fourier coefficients of x(t) be an , bn
and those of y(t) be cn , dn . For the area integral we get
∞
X
A=π n(an dn − bn cn ).
n=1

For second,
∞
X
2 2
L = 2π n2 (a2n + b2n + c2n + d2n )
n=1

Thus we see
∞ n
X o
2 2
L − 4πA = 2π n2 (a2n + b2n + c2n + d2n ) − 2n(an dn − bn cn )
n=1
X∞ n o
2 2 2 2
= 2π (nan − dn ) + (nbn + cn ) + (n − 1)(c2n + d2n )
n=1

Thus
L2 ≥ 4πA
and the maximum value of the enclosed area equals L2 /4π. To determine the curve that achieves this,
equality must hold which is so if and only if

nan − dn = nbn + cn = cn = dn = 0, n = 2, 3, . . .

and a1 = d1 , b1 = −c1 . Thus

x(t) = a0 + a1 cos t + b1 sin t, y(t) = c0 − b1 cos t + a1 sin t.

which represents a circle. The proof is complete.

Question: How do you know that the Fourier series of x0 (t) and y 0 (t) are nan , nb and ncn , ndn
respectively? Do you need to differentiate these Fourier series term by term? Is this justified? Can
you prove the stated result without term by term differentiation??

Laplace’s Equation on a disc We shall now apply the theory of Fourier series for solving some
classical PDEs. We take up the Dirichlet problem for the Laplace equation on the unit disc D =
{(x, y) : x2 + y 2 ≤ 1}. The problem seeks a twice continuously differentiable function u such that

∆u = 0, on D, u(cos θ, sin θ) = f (θ)

where we assume that f is Lipschitz and 2π periodic on R. First we write the equation in polar
coordinates:
∂ 2 u 1 ∂u 1 ∂ 2u
+ + = 0.
∂r2 r ∂r r2 ∂θ2
55
Seeking special solutions in the form u(x, y) = v(r)g(θ) where g(θ) is 2π periodic,

(r2 v 00 + rv 0 )/v = −g 00 /g

Since g is a function of θ alone and v is a function of r alone, either side must be a constant say k 2
and we get the pair of ODEs

r2 v 00 (r) + rv 0 (r) − k 2 v(r) = 0, g 00 (θ) + k 2 g(θ) = 0.

These have solutions

v(r) = Ark + Br−k , g(θ) = C cos kθ + D sin kθ.
Now since g(θ) is 2π-periodic, we must have k ∈ Z (how?). Also since the solution is continuous at
the origin, k ≥ 0. Thus we get the solution in the form
∞
X
u = a0 + rn (an cos nθ + bn sin nθ)
n=1

To determine the coefficients of this Fourier expansion we must use the boundary condition. Setting
r = 1 we get
∞
X
f (θ) = a0 + (an cos nθ + bn sin nθ)
n=1

from which we deduce the values of a0 , a1 , . . . and b1 , b2 , . . . .

Exercises:

26. Determine the solution of the Laplace’s equation in the unit disc with the prescribed boundary
value | sin θ|.

27. Show that if u is a harmonic function (that is ∆u = 0) then

Z π
1
u(r cos θ, r sin θ) = u(0, 0).
2π −π

This is called the mean value theorem for harmonic functions.

The Poisson Kernel Let us continue with the formula obtained in the last slide:
∞
X
u(reiθ ) = a0 + rn (an cos nθ + bn sin nθ)
n=1

But we know the formula for these Fourier coefficients and inserting these,
Z π ∞ Z π
iθ 1 1 X n
u(re ) = f (t)dt + r cos nθ f (t) cos ntdt +
2π −π π n=1 −π
Z π 
sin nθ f (t) sin ntdt
−π

56
Since the integrals decay to zero by Riemann Lebesgue lemma it is easy to justify exchange of sum
and integral (with 0 ≤ r < 1):
Z π ∞
iθ 1  X 
u(re ) = 1+2 rn (cos nθ cos nt + 2 sin nθ sin nt) f (t)dt
2π −π n=1

The student is invited to sum the series

∞
X
1+ 2rn cos n(θ − t)
n=1

and we get the result

π
(1 − r2 )f (t)dt
Z
1 iθ
u(re ) =
2π −π 1 + r2 − 2r cos(θ − t)
2 2
The expression Πr (θ − t) = (1 − r )/(1 + r − 2r cos(θ − t)) is called the Poisson Kernel.

Problems on Poisson Kernel We still have to show that the solution obtained in the last slide
does attain the value f (θ) on the boundary.

28. Show that the Poisson kernel is non-negative and

Z π
1
Πr (θ − t) = 1.
2π −π

29. Show that

lim |u(reiθ ) − f (θ)| = 0.
r→1−

Hint: Write f (θ) as the integral w.r.t over [−π, π] of f (θ)Πr (θ − t)/(2π). Now let  > 0 and I be
an interval of length  centered at θ. The integral over I is small for one reason and the integral
over [−π, π] − I is small for a different reason.

Poisson Formula for the Ball There is a corresponding result for the ball in R3 but to derive that
we need to spend a little time with associated Legendre equations.

Theorem 4.8: Suppose given a continuous function f (x) on the unit ball B centered at the origin
in R3 then the solution of the boundary value problem

∆u = 0 on B u =f
∂B

is given by
(1 − kxk2 )f (x)dS(x)
Z
f (y) = ,
∂B (1 + kxk2 − 2kxk cos α)3/2
where α is the angle between x and y.
For the proof see page 180 of G. B. Folland, Fourier Analysis and its applications, Amer. Math.
Soc, Indian Reprint by Universities Press, New Delhi, 2012.

57
Problems on the Heat Equation
30. Let u(x, t) be a smooth solution of the heat equation in the upperZ πhalf-plane t ≥ 0 such that
u(x, t) = u(x + 2π, t) for all x ∈ R, t ≥ 0. Show that the energy (u(x, t))2 dx is a monotone
−π
decreasing function of time. Prove the same result if the integral is over any interval of length 2π.
Deduce that a smooth 2π−periodic solution of the heat equation with a given initial condition,
if it exists, is unique.
31. Determine the periodic solution of the initial-boundary value for the heat equation ut − uxx = 0
on −π ≤ x ≤ π and 0 ≤ t ≤ T subject to
u(−π, t) = u(π, t), u(x, 0) = π 2 − x2 .

32. Solve the heat equation on the same rectangle above but with data u(−π, t) = u(π, t) =
0, u(x, 0) = π 2 − x2 . This leads to a Sturm-Liouville problem with Dirichlet boundary con-
ditions for which you would have to compute the eigen-values and eigen-functions.
33. Seek a periodic solution (with period 2π) of the heat equation ut = uxx of the form f (x)g(t) and
write down the Ansatz for the most general solutions as superpositions of these special solutions.
Given a smooth 2π−periodic initial condition u(x, 0) = φ(x), show that the solution (whose
uniqueness has been established) is given by
Z π ∞
1  X
in(x−ξ)−n2 t

u(x, t) = φ(ξ) e dξ, t > 0.
2π −π n=−∞

Call the sum of the series as G(x − ξ, t) and we see that the solution is (a convolution?)
Z π
1
u(x, t) = φ(ξ)G(x − ξ, t)dξ.
2π −π

34. Note that the series as such makes no sense when t = 0. How would you say that the solution
displayed achieves the boundary value φ when t → 0?

Wave equation
35. Solve the initial-boundary value problem for the wave equation
utt − uxx = 0,
on the rectangle −π ≤ x ≤ π and 0 ≤ t ≤ 1 with initial conditions
u(x, 0) = | sin x|, ut (x, 0) = 0,
and periodic boundary condition u(x, t) = u(x + 2π, t).
36. Do the preceding for 2π−periodic solutions but with data
u(x, 0) = π 2 − x2 , ut (x, 0) = 0.

37. Solve the preceding two problems with the same initial conditions but with Dirichlet boundary
conditions u(−π, t) = 0 = u(π, t) instead of 2π− periodicity. Again you will run into a Sturm-
Liouville problem as in the case of heat equation.

58
Additional Problems Determine the Fourier coefficients of f (x) = (π 2 − x2 )−1/2 . I. N. Sneddon,
p. 141.

59
 V - Fourier transforms and partial differential equations

Recall that in basic ODE theory where one studies equations with constant coefficients, special
solutions were sought in the form
P (x) exp(mx).
Here m is a root of the characteristic equation and P (x) is a polynomial which would be non-constant
when the characteristic equation has multiple roots.
Generalizing to the case of partial differential equations with constant coefficients (such as the
fundamental equations arising in physics), it is natural to seek plane wave solutions
exp i(x1 ξ1 + x2 ξ2 + · · · + xn ξn ) (1)
and more general solutions can be obtained by superpositions.
In the case of partial differential equations, the characteristic equation would be a polynomial in
several variables. For example taking the case of the wave equation
utt − uxx = 0, (2)
let us substitute the Ansatz (1) in the form exp i(at − bx) into (2) and we get
a2 − b2 = 0. (3)
Equation (3) has infinitely many solutions and indeed two continuous families (λ, λ) and (λ, −λ). We
would now have to take a continuous superposition of the plane waves
exp iλ(x + t), exp iλ(x − t)
which means we must consider the sum of two integrals
Z ∞ Z ∞
f (λ) exp iλ(x + t)dλ + g(λ) exp iλ(x − t)dλ. (4)
−∞ −∞

Definition of the Fourier transform We are naturally led to the following

Definition: Suppose f : R −→ C is a function for which

Z ∞
|f (t)|dt (5)
−∞

is finite then the integral Z ∞

fb(ξ) = f (t)e−itξ dt
−∞
is called the Fourier transform of f (t).
There are several conventions and we follow the one that is common in PDEs for example see p.
213 of G. B. Folland, Fourier analysis and its applications.

60
Exercises:
1. Let f (t) = 1 when |t| ≤ 1 and f (t) = 0 otherwise. Compute the Fourier transform of f .
√
2. Compute the Fourier transform of the function f (t) given by f (t) = 1/ 1 − t2 if |t| < 1 and
f (t) = 0 if |t| ≥ 1.
3. Prove the Riemann Lebesgue lemma which states that if f (t) is continuous and the integral (5)
is finite then |fb(ξ)| −→ 0 when ξ → ±∞.
4. Compute the Fourier transform of f (t) = 1/(a2 + t2 ) where a is a non-zero real number. Note
that the cases a > 0 and a < 0 have to be dealt with separately.
Hint: Let I(ξ) be the integral. Find the Laplace transform of I(ξ).
5. Compute the Fourier transform of exp(−a|t|) where a > 0.
6. Looking at the last two examples are you led to conjecture any general result?
7. Calculate the Fourier transform of f (t) = sin2 t/t2 using the ideas of exercise (4) above.
8. One can also compute the Fourier transform of f (t) = sin t/t but a careful justification would
have to wait. Why so? However proceed formally and try to arrive at the answer.
9. Try to calculate fa ∗ fb where fa (t) = a/(π(x2 + a2 )). Recall the definition of convolution. It is
an integral from −∞ to ∞. Don’t be too surprised if the computation gets pretty ugly. This
example comes up in Probability theory under the name of Cauchy distribution.

Fourier transform of the Gaussian: This is one of the most important examples in the theory of
Fourier transforms and plays a crucial role in probability theory, number theory, quantum mechanics,
theory of heat conduction and diffusive processes in general.

Theorem 5.1: Suppose a > 0. The Fourier transform of exp(−at2 ) is the function
r
π
exp(−ξ 2 /4a).
a

Proof: We shall obtain a first order ODE for I(ξ) given by

Z ∞ Z ∞
2
I(ξ) = exp(−at − itξ)dt = exp(−at2 ) cos(tξ)dt.
−∞ −∞

Differentiate the integral with respect to ξ and we get

Z ∞
0
I (ξ) = − t exp(−at2 ) sin(tξ)dt.
−∞

which can be written as Z ∞

0 1 d 
I (ξ) = exp(−at2 ) sin(tξ)dt
2a −∞ dt
Integrating by parts we get Z ∞
0 ξ
I (ξ) = − exp(−at2 ) cos(tξ)dt
2a −∞
whereby we obtain the ODE I 0 + ξ/(2a)I = 0.

61
Exercises:
10. Integrating this linear ODE and complete the argument.
11. Compute the Fourier transform of x2 exp(−ax2 ) and more generally x2k exp(−ax2 ).

The Schwartz space S of rapidly decreasing functions This is a convenient class of functions
introduced by Laurant Schwartz in his influential work Theorié des distributions, Hermann, Paris.
This function space has the advantage that it is a vector space and it closed under differentiation as
well as multiplication by polynomials. Besides it also contains enough functions so that any f : R −→ C
such that Z ∞
|f (t)|dt < ∞
−∞
can be approximated arbitrarily closely by functions in S.
This space is particularly well-suited to the study of the Fourier transform. One proves all the
results in the context of S where there are no technical obstructions to differentiation under integral
sign or switching order of integrals. To pass on to the more general case one resorts to approximation
techniques. The situation is reminiscent of the proof of Parseval formula for Riemann integrable
functions.

Definition: The space S consists of all infinitely differentiable functions f (t) such that for any
m, n ∈ N,
 n
m d
t f (t)
dtn
remains bounded.
We immediately see that for any polynomial P (t) and any a > 0, the function P (t) exp(−at2 ) lies
in S.
12. Show that (cosh at)−1 and t(sinh at)−1 lie in S for any a > 0.
13. One can show using the complex version of Stirling’s approximation formula that

|Γ(a + it)|2

lies in S for a > 0. This is an Optional exercise.

Properties of S: The space S is obviously a vector space and it has the following further properties
1. Suppose f (t) and g(t) belong to S then so does f (t)g(t).
2. It f (t) ∈ S then the derivative f 0 (t) also lies in S.
3. If f (t) ∈ S and P (t) is a polynomial then f (t)P (t) ∈ S.
4. For every f (t) ∈ S the integral of |f (t)| over R exists and so for every f (t) ∈ S the Fourier
transform is defined.
We shall now show that if f (t) ∈ S then fb(ξ) also lies in S. Thus the Fourier transform maps S into
itself as a linear transformation. Can you point out some of its eigen-values by looking at the list of
Fourier transforms we have calculated?

62
Theorem 5.2: Suppose f (t) ∈ S then fb(ξ) ∈ S.

Proof: Differentiating under the integral is permissible since f (t) decays rapidly. Thus
 d n Z ∞
f (ξ) = (−i)
b n
e−itξ (tn f (t))dt
dξ −∞
n o
Rewrite the integrand as e−itξ tn f (t)(1 + t2 ) (1 + t2 )−1 . The bracketed term is bounded in absolute
value. Multiplying the integral by (−iξ)k we see
 d n Z ∞  k
d
(−iξ) k
fb(ξ) = (−1)n
e−itξ (tn f (t))dt
idξ −∞ dt

Integrating by parts and using Dk (tn f (t)) ∈ S we conclude ξ k Dn fb(ξ) is bounded for every n, k ∈ N.
Since k is arbitrary, the result is proved.

Differentiation and multiplication

Theorem 5.3: Suppose f (t) ∈ S then

 d\ 
(f (t)) = ξ fb(ξ)
idt
[
 d  b 
tf (t)(ξ) = − f (ξ)
idξ

To prove the first part integrate by parts. To prove the second part, differentiate under the integral
with respect to ξ: Z ∞ Z ∞
d b  d −itξ
f (ξ) = e f (t)dt = −ite−itξ f (t)dt
dξ dξ −∞ −∞
Divide by −i and we get the second formula.

Hermite’s ODE and Hermite polynomials again !!

14. Transform the Hermite’s differential equation
y 00 − 2xy 0 + 2λy = 0
by the substitution y exp(−x2 /2) = u. Ans: u00 − x2 u + (2λ + 1)u = 0.
15. Show that is u is a solution for the transformed ODE in the previous exercise then u b is also a
solution of the same differential equation. That is, the transformed equation is invariant under
Fourier transform.
16. Show that at most one of the solutions of the transformed ODE lies in S. If λ ∈ N then

Hn (x) exp(−x2 /2)

where Hn (x) is the n−th Hermite polynomial, lies in S. Hint: The Abel-Liouville formula.
17. The Fourier transform is a linear transformation of S to itself. Show that Hn (x) exp(−x2 /2) are
eigen-vectors of this linear map.

63
Inversion theorem

Theorem 5.4: Suppose f (t) is a function in S then the function can be recovered from its Fourier
transform via the formula Z ∞
1
f (t) = fb(ξ)eitξ dξ. (6)
2π −∞
To prove this first try to substitute in the RHS of (6) the expression for fd
(ξ) and invert the order of
integrals. You will run into the following (hitherto meaningless) integral:
Z ∞
exp(i(t − x)ξ)dξ
−∞

Proof of inversion theorem: The exp(−x2 ) trick ! Equipped with this, we now write
Z ∞ Z ∞
itξ 2
f (ξ)e dξ = lim
b fb(ξ)eitξ−ξ dξ
−∞ →0 −∞

Now we put in the definition of fb(ξ) and we get

Z ∞ Z ∞ Z ∞ 
itξ iξ(t−x)−ξ 2
f (ξ)e dξ = lim
b f (x)e dx dξ
−∞ →0 −∞ −∞

We can now safely invert the order of integral and write

Z ∞ Z ∞ Z ∞ 
itξ 2
f (ξ)e dξ = lim
b eiξ(t−x)−ξ dξ f (x)dx
−∞ →0 −∞ −∞

The inner integral is the Fourier transform of the Gaussian that we have computed !!
Z ∞ r Z ∞
itξ π
fb(ξ)e dξ = lim exp(−(x − t)2 /4)f (x)dx
−∞ →0  −∞
√
Putting x = t + 4s we obtain
Z ∞
√
Z ∞ √ 2
itξ
f (ξ)e dξ = 2 π lim
b f (t + 4s)e−s ds = 2πf (t)
−∞ →0 −∞

The proof is complete.

Here is a bunch of exercises are collected all of which are amenable to the exp(−t2 ) trick.

Exercises:

18. Prove that Z ∞

fb(ξ)dξ = 2πf (0), f ∈ S.
−∞

19. Compute the Fourier transform of (sin at)/t.

20. Compute the Fourier transform of f (t) = (t2 + a2 )−1 by obtaining a second order ODE.

64
 21. Use the exp(−t2 ) trick to obtain a second order ODE for the function

2 ∞ (sin tx)dt
Z
√ , x > 0.
π 1 t2 − 1
Can you recognize this as a familiar function? Also prove this by taking Laplace transform.
This is from H. Weber, Die Partiellen Differential-Gleichungen der mathematischen Physik, Vol
- I, Braunschweig, 1900, based on Riemann’s lectures. The formula appears on p. 175. This
representation is due to Mehler (1872) and Sonine (1880). See p. 170 of G. N. Watson’s treatise.

Ramanujan’s Formula: Here is an interesting formula due to Ramanujan that can be deduced
from the exp(−x2 ) trick.

Theorem 5.5: Let a > 0 then

Z ∞
√ 1
|Γ(a + it)|2 e−itξ dt = πΓ(a)Γ(a + )sech−2a (ξ/2).
−∞ 2

For the details of the proof see D. Chakrabarty and G. K. Srinivasan, On a remarkable formula of
Ramanujan, Archiv der Mathematik (Basel) 99 (2012) 125-135.

The Jacobi theta function identity:

22. Show that if f ∈ S and c > 0 then

Z c
sin kt
lim f (t)dt
k→∞ 0 t

exists. Hint: (sin kt)/t is the integral of cos ut over [0, k]. Is it necessary that f ∈ S?

23. Show that if 0 < c < π then

Z c
sin kt π
lim f (t)dt = f (0), f ∈ S.
k→∞ 0 sin t 2

24. Modify the formula of the previous exercise if 0 < c < 2π and → ∞ through odd-integer values.
Hint: Break the integral into three integrals over [0, π/2], [π/2, π] and [π/2, c]. What if c = π?
What if c > 2π??

We shall see how this leads to the famous theta function identity of Jacobi.

25. With a little care one can continue from the last exercise and show Oskar Schlömilch, Analytische
Studien, Vol - II, Leipzig, 1847, p. 29, that if f ∈ S and an even function,
Z ∞
sin kt n o
lim f (t)dt = π f (0) + 2f (π) + 2f (2π) + 2f (3π) + . . .
k→∞ −∞ sin t

65
 26. Recall the Fourier transform of the Gaussian:
Z ∞ √
−a2 t2 π
e cos(2ξt)dt = exp(−ξ 2 /a2 )
−∞ a

Set ξ = 0, 1, . . . , n, add and deduce Jacobi’s theta function identity:

n o √π n o
−a2 π 2 −4a2 π 2 2 2
π 1 + 2e + 2e + ... = 1 + 2e−1/a + 2e−4/a + . . .
a
Hint: 1 + cos 2t + · · · + cos 2nt = sin(2n + 1)t/ sin t.

One can now argue in general, without additional effort, to prove:

Theorem 5.6 (Poisson summation formula): Suppose f ∈ S then

∞
X ∞
X
f (n) = fb(2nπ).
n=−∞ −∞

Optional: Try this! One can relax the requirement f ∈ S. How about f (x) = 1/(a2 + x2 )? Compute
the sum of the series
1 1 1
2
+ 2
+ + ...
1+1 1+2 1 + 32
Compare this with the partial fraction decomposition of the cosecent obtained in the last chapter.

Theorem 5.7 (Parseval formula also known as Plancherel’s theorem): Suppose f (t) and
g(t) are in S then Z ∞ Z ∞
1
f (t)g(t)dt = fb(ξ)b
g (ξ)dξ (7)
−∞ 2π −∞

Exercises:

27. (Optional) Prove this using the exp(−t2 ) trick.

28. Apply this theorem to calculate the eigen-values of the Fourier transform which is a linear
transformation from S to itself.

Theorem 5.8 (Convolution theorem): Suppose f (t) and g(t) are both in S then so is their
convolution (f ∗ g)(t). Further
∗ g(ξ) = fb(ξ)b
f[ g (ξ). (8)
We shall not prove that the convolution is in S. Verification of (8) is an easy exercise. Observe the
analogy with the corresponding result for Laplace transforms.

66
Heat equation again !! Let us now solve the initial value problem for the heat equation in the
half-plane
ut − uxx = 0, u(x, 0) = f (x).
Let us assume to begin with f (x) ∈ S and compute the Fourier transform with respect to x:

d
u(ξ, t)) + ξ 2 u
(b b = 0, u
b(ξ, 0) = fb(ξ).
dt
This is an ODE in u
b where ξ is regarded as a parameter.

b(ξ, t) = C exp(−tξ 2 )
u

Putting in t = 0 we see that C = fb(ξ). Thus

b(ξ, t) = fb(ξ) exp(−tξ 2 )

u

The Heat Kernel Observe that exp(−tξ 2 ) is the Fourier transform with respect to x, of the function
1
G(x, t) = √ exp(−x2 /4t)
4πt
Appealing to the convolution theorem,

u
b(ξ, t) = fb(ξ)G(ξ, [
b t) = G ∗ f (ξ, t)

Thus we have Z ∞
1
u(x, t) = √ f (s) exp(−(x − s)2 /4t) ds (9)
4πt −∞

The function G(x, t) called the heat kernel, plays a crucial role in Probability theory. The formula was
derived assuming that the initial data f (x) is in S. However it makes perfect sense even if f (x) is of
exponential type !

Exercises:

29. Solve the heat equation ut − uxx = 0 with initial condition u(x, 0) = x2 .

30. Solve the heat equation ut − uxx = 0 with initial condition u(x, 0) = cos(ax). What about the
solution with initial condition sin(ax).

31. Suppose the initial condition is a continuous function that is positive at say on (−1, 1) but zero
outside [−1, 1] then the solution is positive at all points u(x, t) no matter how large x is and how
small t > 0 is. Thus the effect of initial heat distribution in [−1, 1] is instantaneously propagated
throughout space. Is this physically tenable?

Philosophical Question: How is it that the equation nevertheless is used to explain physical
phenomena??

67
Jacobi theta function identity via heat equation Let us now assume that the initial data f (s)
in equation (9) is 2π− periodic. Then we write the RHS of (9) as
∞ Z (2n+1)π
1 X
√ f (s) exp(−(x − s)2 /4t)ds
4πt −∞ (2n−1)π

which in turn can be recast as (how?)

Z π nr π X ∞
1 −(x−s+2nπ)2 /4t
o
f (s) e ds
2π −π t n=−∞

Compare this with problem 33 of the previous chapter and re-derive the Jacobi theta function identity.
See Courant-Hilbert, Methods of mathematical physics, Volume - II, page 200.

D’Alembert’s solution of the wave equation At the beginning of the chapter we saw were led
to looking for solutions of the wave equation

utt − uxx = 0

in the form u(x, t) = φ(x + t) + ψ(x − t). Let us now supplement this with initial conditions

u(x, 0) = f (x), ut (x, 0) = g(x).

Then
φ(x) + ψ(x) = f (x), φ0 (x) − ψ 0 (x) = g(x),
whereby
1 1
φ0 (s) = (f 0 (s) + g(s)), ψ 0 (s) = (f 0 (s) − g(s))
2 2
We immediately get the formula
Z x+t
1 1
u(x, t) = (f (x + t) + f (x − t)) + g(s)ds. (10)
2 2 x−t

This is known as D’Alembert’s formula.

Exercises:
32. Solve the wave equation utt − uxx = 0 with initial conditions u(x, 0) = x and ut (x, 0) = 1 − x.

33. Solve the wave equation utt − uxx = 0 with initial conditions u(x, 0) = sin x and ut (x, 0) = 0.

34. If v1 is solution of the wave equation utt − uxx = 0 and v2 is the solution with the data of the
first solution altered along [2, 3]. Sketch the regions in the (x, t) plane where v1 and v2 definitely
agree and the region where they could disagree.

35. A, B, C, D are the successive vertices of a rectangle whose sides have slopes ±1. Show that if u
is a solution of the wave equation,

u(A) + u(C) = u(B) + u(D).

68
Radial functions - the Bessel transform Let us now consider the multi-dimensional Fourier
transform of a function f (x1 , x2 , . . . , xn ) which is assumed continuous and decaying rapidly enough so
that the integrals appearing exist. The Fourier transform is defined as
Z
fb(ξ1 , . . . , ξn ) = exp(−i(x1 ξ1 + · · · + xn ξn ))f (x1 , . . . , xn )dx1 . . . dxn .
Rn
p
that the function f depends only on x21 + · · · + x2n . Then the Fourier transform depends
Assume p
only on ξ12 + · · · + ξn2 . To see this let us write

hx, ξi = x1 ξ1 + x2 ξ2 + · · · + xn ξn

Now let us choose a rotation matrix A such that Akξken = ξ. Then

Z
fb(ξ1 , . . . , ξn ) = exp(−ikξkhAen , xi)f (x1 , . . . , xn )dx1 . . . dxn .
Rn

But hAen , xi = hen , AT xi. Now put x = Ay. Since A is orthogonal, its determinant is one and
applying the change of variable formula
Z
fb(ξ1 , . . . , ξn ) = exp(−ikξkyn )f (y1 , . . . , yn )dy1 . . . dyn .
Rn

We should now write it in polar coordinates. p But we shall do this for n = 2 and n = 3 only. Write
2
f (x1 , x2 , . . . , xn ) = F (r) where r = x1 + · · · + x2n .
For n = 3, y3 = r cos φ and
Z ∞ Z π Z 2π
2
f (ξ1 , . . . , ξn ) =
b F (r)r dr exp(−irkξk cos φ) sin φdφ dθ
0 0 0

Writing cos φ = s, we get

Z ∞ Z 1 Z ∞
2 −irskξk 4π
fb(ξ1 , ξ2 , ξ3 ) = 2π F (r)r dr e ds = F (r)r sin kξkrdr
0 −1 kξk 0

For the case n = 2 we get

Z ∞ Z π
fb(ξ1 , ξ2 ) = rF (r)dr cos(rkξk sin θ)dθ
0 −π

Which is a Bessel transform Z ∞

fb(ξ1 , ξ2 ) = 2π rF (r)J0 (rkξk)dr
0
More generally in even space dimensions it reduces to a Bessel transform and in odd dimension a “sine
transform”.

69
Principle of Equipartioning of Energy for the Wave Equation We derived the energy equation
for utt − uxx = 0.
36. Use Parseval’s formula to compute the Fourier transform (with respect to x−variable) of the
energy.
37. Calculate the Fourier transform with respect to the x−variable of the kinetic energy alone. Also
compute the Fourier transform of the solution in terms of the initial conditions.
38. Calculate the limit as t → ±∞ of the result obtained in the last exercise. Use the half angle
formula and Riemann Lebesgue lemma.
39. Show that when time goes to infinity, the limiting value of kinetic energy is half the total energy.
See R. Stricharz, A guide to distribution theory and Fourier transforms, CRC Press LLC, 1994.

Airy’s Function: Airy studied the function that bears his name in the course of his investigations
on the intensity of light in the neighborhood of a caustic (See G. N. Watson’s treatise, p. 188). The
work dates back to 1838. Before commencing on the discussion of Airy’s function, here is a pointer to
the interesting life of Sir George Biddell Airy:
library.mat.uniroma1.it/appoggio/MOSTRA2006/airy
Another interesting account I had read long ago was by Patrick Moore but I am unable to locate it at
the moment.
Airy’s equation is the ODE
1
y 00 (x) − xy = 0.
3
40. Formally take the Fourier transform of the Airy’s differential equation y 00 − 13 xy = 0. Use the
inversion formula to obtain a representation
Z ∞
1
I(x) = exp(iξ 3 + ixξ)dξ.
2π −∞

The integral converges only conditionally (how? maybe you should first discuss the integrals over R
of sin(x2 ) and cos(x2 )). The steps leading to the integral are suspect and so one has to justify it in
some other way.
41. Now replace ξ by z = ξ + iη where the imaginary part η is positive and consider
Z ∞
1
I(x) = exp(iz 3 + ixz)dξ..
2π −∞
The integral now along a line parallel to the real axis, converges absolutely and very rapidly.
Differentiate under the integral sign and check that I(x) satisfies the ODE.
42. The tricky business is to show that one can allow η → 0 and get the integral of the last slide.
How can one justify this when the integral converges only conditionally? Try to integrate by
parts once and then see if you can allow η −→ 0. Those who have studied complex analysis can
put it to good use. In fact the integral does not depend on η at all !! How would you verify this?
43. Determine the radius of convergence of the power series solution of Airy’s equation. We have
now both the series solution as well as the integral representation. Which one would you think
is more useful?

70
Additional Problems

1. In the last chapter you computed the Fourier coefficients of f (x) = (π 2 − x2 )−1/2 . Assuming that
the series converges on the open interval (−π, π) deduce (I. N. Sneddon, p. 141)
∞
sin x n 2
X (−1)n J0 (nx) o
J0 (x) = 1 + 2x .
x n=1
x2 − n 2 π 2

71
 VI - Application to Celestial Mechanics

In this last chapter of this course we discuss the original problem that led Bessel in 1824 to
introduce the functions that bear his name. Wilhelm Bessel was an astronomer at Königsberg and his
chief interest was the study of orbits of comets.
We must begin by recalling the three basic laws of planetary mechanics motion enunciated by
Johannas Kepler. The discovery of these laws forms an interesting culmination of classical astronomy.
Equipped with his calculus, Issac Newton, with his laws of dynamics able to explain:

1. The motion of planets.

2. The precession of equinoxes.

3. The formation of tides.

Astronomy hitherto an empirical science transformed into a dynamical science.

Kepler’s laws of planetary motion The first two laws were enunciated in 1609 in De Motibus
Stellae Martis:

Kepler’s First Law: The planets move around the sun in elliptical orbits with the sun at one of
the foci.

Kepler’s Second Law: The radius vector joining the sun and the planet sweeps out equal areas in
equal intervals of time. This is just a restatement of the law of conservation of angular momentum.
The third law (which is an approximate law neglecting the masses of planets) appeared much later
in 1619 in his Harmonices Mundi.

Kepler’s Third Law: The square of the period T is proportional to the cube of the semi-major
axis a of the orbit.
The constant of proportionality was determined by Gauss in his famous book on Astronomy8 and
is known as Gaussian gravitational constant.

The Kepler problem Suppose the planet passes the perihelion Π at time t = 0. Placing the sun
at the F, we measure all angles from the radius vector F Π. Let X(t) be the position of the planet at
time t and θ(t) be the angle between F Π and F X. In astronomy one calls the function θ(t) the True
Anomaly.
8
Theoria Motus Corporum Coelestium

72
Kepler Problem: Find the function θ(t) as explicitly as possible.
PICTURE

A basic lemma: Let P and P 0 be corresponding points on the ellipse and the auxiliary circle. Then
Area(AF P ) refers to the area of the sector of the ellipse and Area(AF P 0 ) the corresponding sector of
the auxiliary circle.
Area(AF P ) b Area(Ellipse)
0
= = (1)
Area(AF P ) a Area(Circle)
Exercise:
1. Prove this using integral calculus to compute the indicated areas.
We now use Kepler’s second law and write
Area(AF P ) t Area(AF P 0 )
= = (2)
Area(Ellipse) T Area(Circle)
But if C denotes the center of the ellipse (and the circle) then
Area(AF P 0 ) = Area(ACP 0 ) − Area(∆F CP 0 ) (3)
The last two areas are readily described and we get
1
Area(AF P 0 ) = (a2 φ − a2  sin φ) (4)
2
where φ is the eccentric angle of P and  the eccentricity of the ellipse.
Substituting (4) in (3) we get
Area(AF P 0 ) φ −  sin φ
=
Area(Circle) 2π
or, using (2) we get
2πt
φ −  sin φ = (5)
T
Equation (5) is the famous Kepler Equation. The number 2πt/T is called the Mean Anomaly.

Inverting the Kepler equation

Exercises:
2. Show that the function φ −  sin φ is strictly increasing and maps R onto R. The inverse function
φ(t) is a strictly increasing infinitely differentiable function.
3. Explain why the function φ(t) is an odd function. Use both geometrical reasoning as well as
mathematical analysis.
4. Show that φ(0) = 0 and φ(T /2) = π.
The problem of inverting the Kepler equation has been studied by many eminent mathematicians
such as J. L. Lagrange, Memoirs of the Berlin Academie 1768-69. Also volume - II, p. 22 ff. of his
Méchanique Analytique 1815. In this connection Lagrange discovered the inversion formula that bears
his name. The Lagrange inversion formula has important applications in quite un-related fields such
as combinatorics.

73
Bringing in the periodicity Let us now look at φ(t + T ). The Kepler equation gives
2π(t + T ) 2πt
φ(t + T ) −  sin φ(t + T ) = = + 2π.
T T
which can be written as
φ(t + T ) −  sin φ(t + T ) = (φ(t) −  sin φ(t)) + 2π
= (φ(t) + 2π) −  sin(φ(t) + 2π).

By injectivity of the function λ −  sin λ we conclude

φ(t + T ) = φ(t) + 2π. (6)

Exercises:
5. Could you derive this directly from physical considerations?
6. Verify using (6) that the function ψ(t) = φ(t) − 2πt
T
is a periodic function with period T .
   
tT tT
Clearly then ψ 2π is periodic with period 2π. Let us write the Fourier series for ψ 2π :
 tT  ∞
X
ψ = bn sin(nt),
2π n=1
where
2 π  tT  2 π n  tT 
Z Z o
bn = ψ sin(nt)dt = φ − t sin(nt)dt.
π 0 2π π 0 2π
Integrating by parts and recalling φ(T /2) = π, φ(0) = 0, we get
Z π n  
2 d tT o
bn = φ − t cos(nt)dt.
nπ 0 dt 2π
This simplifies to
Z π Z T /2
2 d  tT  2
bn = φ cos(nt)dt = φ0 (s) cos(2πns/T )ds.
nπ 0 dt 2π nπ 0
Using the Kepler equation again, the argument of the cosine can be re-written resulting in:
Z T /2
2
bn = φ0 (s) cos(nφ(s) − n sin φ(s))ds.
nπ 0
The change of variables φ(s) = λ now gives
Z π
2 2Jn (n)
bn = cos(nλ − n sin λ)dλ = .
nπ 0 n
The Fourier series now reads
 tT   tT  ∞
X 2Jn (n)
ψ =φ −t= sin nt.
2π 2π n=1
n
So the eccentric angle φ(t) (also known in astronomy as the Eccentric Anomaly ) can be written as a
Kaypten series:
∞
2πt X 2Jn (n)
φ(t) = + sin(2πnt/T ). (7)
T n=1
n

74
The true anomaly

7. Using elementary trigonometry, find a relation between the true anomaly and the eccentric
anomaly.

For a short but quick historical survey see C. A. Ronan, Science, its history and development among
world’s culture, pp. 336-337.
For more on the Kepler problem and mathematical principles underlying celestial mechanics, the book
by J. M. Danby, Celestial Mechanics and dynamical astronomy, Kluwer Academic, 1991, is HIGHLY
recommended. This second edition contains computer experiments.
A more ambitious project would be to read the comprehensive two volumes
D. Boccaletti and G. Puccaco, Theory of Orbits, Vol- I, II, Springer Verlag, 2004.
One can try to expand the Bessel functions appearing in the Kaypten series and rearrange terms to
get a power series in e. however it was known that the resulting series converges only when e < 0.667.
The orbits of most comets exceed this number - Orbit of Halley’s comet is 0.96 !!
It seems an investigation into why the series fails to converge beyond this threshold led Cauchy
to develop the theory of functions of one complex variable. There is an imaginary singularity that
prevents the power series from converging beyond the threshold value.

Further reading: The classic work of R. Courant and D. Hilbert has been cited several times.
Besides this, there are several others. The serious student can scarcely do better than to begin a
systematic study of the following three books which also have been referred in the slides.

1. G. B. Folland, Fourier Analysis and its applications,

2. A. Robert, Advanced calculus for users, North-Holland, 1989.

3. R. Strichartz, Guide to distribution theory and Fourier transforms.

75
 VI - Acknowledgment

The instructor wishes to thank the many students who have pointed out errors as well as suggested
alternate solutions to the problems. The response of students has been gratifying. Thanks also for
patiently listening to the lectures and brooking the instructor’s pestilence.

76
 Solutions and hints to problems in chapter - I
Z 1
1. Compute (Pn (x))2 dx
−1

Solution: We use Rodrigues’s formula and write

Z 1 Z 1
2 1
(Pn (x)) dx = 2 n Dn (x2 − 1)n Dn (x2 − 1)n dx.
−1 n! 4 −1

Integration by parts gives, calling 1/(n!2 4n ) as cn ,

Z 1 1
−cn Dn−1 (x2 − 1)n Dn+1 (x2 − 1)n dx + cn Dn−1 (x2 − 1)n Dn (x2 − 1)n .
−1 −1

The boundary terms vanish since ±1 being n−fold roots of (x2 − 1)n ,

Dk (x2 − 1)n = 0, k = 0, 1, 2, . . . , n − 1.
±1

Further integration by parts gives

Z 1 Z 1
2 n
(Pn (x)) dx = (−1) cn (x2 − 1)n D2n (x2 − 1)n dx
−1 −1
Z 1
= cn (2n)! (1 − x2 )n dx
−1
Z π/2
= cn (2n)! cos2n+1 xdx
−π/2
π/2
2n 2n − 2
Z
2
= cn (2n)! ... cos xdx.
2n + 1 2n − 1 3 −π/2
2
=
2n + 1

as desired.

Use Rodrigues formula to prove that the Legendre polynomial of degree n has precisely n distinct
roots in the open interval (−1, 1).

77
Solution: Suppose f (x) is a polynomial with a double root at a and b say a < b. Then by
Rolle’s theorem there is a c ∈ (a, b) such that f 0 (c) = 0. But we know that

f 0 (a) = f 0 (b) = 0.

So we can apply Rolle’s theorem to f 0 (x) on each of the two intervals (a, c) and (c, b) to conclude
that f 00 (x) has a root in each of the intervals (a, c) and (c, b) and thus f 00 (x) has at least two
distinct roots in (a, b). Repeating the same argument one shows that if f (x) has a triple root at
a and b then f 000 (x) has at least three distinct roots in (a, b). The general situation is quite clear.
Now the polynomial (x2 − 1)n has an n−fold root at ±1 and so its n−th derivative Dn (x2 − 1)n
has at least n distinct roots in (−1, 1). But being a polynomial of degree n it cannot have more
than n roots and so has exactly n distinct roots in (−1, 1).

Show that the sequence of Legendre polynomials Pn (x) satisfies the three term recursion formula

(n + 1)Pn+1 − (2n + 1)xPn (x) + nPn−1 (x) = 0.

Solution: Coefficient of xn+1 in (n + 1)Pn+1 − (2n + 1)xPn (x) equals:

(n + 1)(2n + 2)! (2n + 1)(2n)!

− = 0.
2n+1 ((n + 1)!)2 2n (n!)2

So (n + 1)Pn+1 − (2n + 1)xPn (x) is a polynomial of degree at most n and so can be written as
n
X
(n + 1)Pn+1 − (2n + 1)xPn (x) = ak Pk (x). (1)
k=0

Since Pn+1 (x) and xPn (x) has the same parity and Pn (x) has parity opposite to these, we see
that an = 0. Let us now show ak = 0 when k = 0, 1, . . . , n − 2.
Using orthogonality of the Legendre polynomials we get
Z 1 Z 1
((n + 1)Pn+1 − (2n + 1)xPn (x))Pk (x)dx = ak (Pk (x))2 dx.
−1 −1

Again by orthogonality the integral on the LHS is simply

Z 1
−(2n + 1) Pn (x)(xPk (x))dx (2)
−1

Now xPk (x) = b0 P0 (x) + · · · + bk+1 Pk+1 (x) say. But k + 1 ≤ n and so by orthogonality again,
the integral (2) is zero. Thus we see that ak = 0 for k = 0, 1, 2, . . . , n − 2. We are left with

(n + 1)Pn+1 − (2n + 1)xPn (x) = an−1 Pn−1 (x).

To compute an−1 put x = 1 and recall that Pk (1) = 1 for all k.

Prove that Pn0 (1) = 21 n(n + 1).

78
 Solution: Using Rodrigues’ formula
1
Pn0 (1) = Dn+1 (x − 1)n (x + 1)n .
n!2n x=1

Now using Leibnitz’s rule for higher order derivatives of a product,

n+1 1
Pn0 (1) = n
n!n(2n−1 ) = n(n + 1).
n!2 2
as desired.
0
Prove that Pn+1 (x) − xPn0 (x) = (n + 1)Pn (x).

0
Solution: The polynomial Pn+1 (x) − xPn0 (x) of degree n can be written as
n
X
0
Pn+1 (x) − xPn0 (x) = ak Pk (x).
k=0

Multiply both sides by Pk (x) and integrate over [−1, 1]. Assume k ≤ n − 1. A single integration
by parts gives for the LHS:
Z 1 Z 1
0
− Pn+1 (x)Pk (x)dx + Pn (x)(xPk )0 (x). (1)
−1 −1

The boundary terms cancel out since both Pn+1 (x)Pk (x) and (xPk (x)Pn (x)) equal 1 when x = 1.
Thus
Pn+1 (x)Pk (x) = (xPk (x)Pn (x)) = (−1)n+k+1 , when x = −1.
Since Pk0 (x) and (xPk (x))0 are both polynomials of degree less than or equal to n − 1, we have
by orthogonality the integral (1) equals zero and so ak = 0 when k ≤ n − 1.

Solution to problem 27: Proof of Jacobi’s formula. Following the hint given, let f (t) = Dn−1 (1 −
t2 )(2n−1)/2 . Then
1
(1 − t2 )f 00 (t) = (1 − t2 )Dn+1 (1 − t2 )n− 2
1 1
= Dn+1 (1 − t2 )n+ 2 + 2(n + 1)tDn (1 − t2 )n− 2
1
+ (n2 + n)Dn−1 (1 − t2 )n− 2

Thus (1 − t2 )f 00 − tf 0 (t) + n2 f equals:

1 1 1
Dn+1 (1 − t2 )n+ 2 + t(2n + 1)Dn (1 − t2 )n− 2 + (2n2 + n)Dn−1 (1 − t2 )n− 2 .

The first term can be rewritten as

1 1
Dn+1 (1 − t2 )n+ 2 = −(2n + 1)Dn (1 − t2 )n− 2 t.

Apply Leibnitz rule and things cancel out. The rest of the problem is clear.

79
Solutions for selected problems in chapter - II Prove Laplace’s second integral representation
1 π
Z
dθ
Pn (x) = √
π 0 (x + x2 − 1 cos θ)n+1

Solution: We assume |x| ≥ 1 and use the first formula of Laplace

1 π
Z √
Pn (x) = (x + x2 − 1 cos θ)n dθ
π 0
Perform the change of variables
√ √
x+ x2 − 1 cos θ = 1/(x + x2 − 1 cos φ)

Check that this change of variables is licit and complete the solution.

80
 Use Laplace’s integral representation to prove the following result of C. Neumann (1862)
lim Pn (cos(x/n)) = J0 (x).
n→∞

Solution: Substituting into Laplace’s formula

Z π
n1
Pn (cos(x/n)) = (cos(x/n)) (1 + i tan(x/n) cos θ)n dθ (1)
π 0

The factor (cos(x/n))n tends to one as n → ∞. Let us then focus on the integral. Note that if
a(λ) → α then by L’Hospital’s rule,
(1 + λa(λ))1/λ −→ eα
Passing to the limit in (1) we get
Z π
1
lim Pn (cos(x/n)) = eix cos θ dθ
n→∞ π 0

Solutions and hints to problems of chapter - III

1. Show that the eigen-values of the boundary value problem
y 00 + λy = 0, y(0) = 0, y(1) + y 0 (1) = 0
are roots of k + tan k = 0, where k 2 = λ. Graphically demonstrate that there is an eigen value
in every interval of length π.

Solution: If λ = 0 the solution of the ODE is A + Bx and the boundary conditions force
A = 0 = B. Thus 0 is not an eigen-value. If λ 6= 0, put λ = k 2 and the solution of the ODE is
y(x) = A cos kx + B sin kx
Setting y(0) = 0 we see A = 0. The other boundary condition gives
B sin k + kB cos k = 0,
which is k + tan k = 0. The graphical argument is left to the student. The eigen functions are
sin kn x, where k1 , k2 , k3 , . . . are the roots of k + tan k = 0.
2. Suppose y0 (x) is a minimizer for Z 1
Q[y] = (y 0 (x))2 dx
0
subject to the constraint Z 1
(y(x))2 ρ(x)dx = 1,
0
over the class S of all continuous piecewise once differentiable functions vanishing at 0 and 1.
Assume that the minimizer y0 (x) is twice differentiable then show that it satisfies the BVP
y 00 + λρ(x)y(x) = 0, y(0) = 0 = y(1).
where λ equals Q[y0 ].

81
 Solution: Since the minimum is attained at y0 (x) we have to begin with
Z 1
(y0 (x))2 ρ(x)dx = 1. (1)
0

Let us perturb y0 (x) to y0 (x) + φ(x) where φ(x) is an arbitrary smooth function vanishing at 0
and 1. We normalize by dividing by
s
Z 1
N= (y0 (x) + φ(x))2 ρ(x)dx
0

and consider the perturbation

w(x) = (y0 (x) + φ(x))/N
Then
Q[y0 ] ≤ Q[w] (2)
for all small values of  positive or negative.
Writing out (2) after multiplying through by N 2 and writing λ in place of Q[y0 ] we get
Z 1 Z 1
λ 2
(y0 (x) + φ(x)) ρ(x)dx ≤ (y00 (x) + φ0 (x))2 dx.
0 0

Expanding out, using (1) and canceling out Q[y0 ] or λ from both sides we get
Z 1
2 (λy0 (x)ρ(x) − y00 (x)φ0 (x))dx + 2 (Stuff) ≤ 0.
0

Dividing by  and letting  → 0, separating the cases when it tends to zero through positive and
negative values, we get a pair of opposite inequalities that finally results in
Z 1
(λy0 (x)φ(x)ρ(x) − y00 (x)φ0 (x))dx = 0.
0

Integrating by parts the second term we get

Z 1
(λy0 (x)φ(x)ρ(x) + y000 (x))φ(x)dx = 0.
0

Since φ is arbitrary, it follows that

λy0 (x)φ(x)ρ(x) + y000 (x) = 0.

3. Explain why the the eigen-function corresponding to the first eigen-value of

y 00 + λρ(x)y = 0, y(0) = 0 = y(1).

has no zeros on (0, 1). Give an intuitive non-rigorous argument.

82
Solution: Suppose there is a zero p ∈ (0, 1). Assume the eigen-function y0 (x) is positive on (0, p)
and negative on (p, 1). Consider the function v(x) which equals y0 (x) on (0, p) and −y0 (x) on (p, 1).
The student is advised to draw pictures. Then
Z 1 Z 1
0 2
λ= (y0 (x)) dx, (y0 (x))2 ρ(x)dx = 0.
0 0

Replacing y0 (x) by v(x) would not alter either of these. At this stage one can give a precise proof
using mathematical analysis by appealing to the Dirichlet principle which asserts that v(x) is twice
differentiable and satisfies
v 00 + λρ(x)v = 0.
It immediately follows (exercise) that v 0 (p) = 0 and so by Picard’s theorem v(x) is identically zero
which is a contradiction. The geometrical argument runs as follows. One tweaks the function v(x)
making it flat and as a result the minimum value of the integral
Z 1
(y 0 (x))2 dx
0

would get strictly below the smallest eigen-value leading to a contradiction.

1. Prove the Sturm’s comparison theorem. Suppose ρ(x) and σ(x) are two continuous functions on
an interval I such that ρ(x) > σ(x) throughout I and z(x), w(x) are solutions of the ODEs
z 00 (x) + ρ(x)z(x) = 0, w00 (x) + σ(x)w(x) = 0.
Between two successive zeros of w(x) there is a zero of z(x).

Solution: Let a, b be successive zeros of w(x) and assume z(x) has no zeros in (a, b). We may assume
that both z(x) and w(x) are positive in (a, b) (why?). Multiply the w equation by z, the z equation
by w, integrate over [a, b] and subtract. After a single integration by parts we get
b b
Z b
0 0
w(x)z (x) − z(x)w (x) + (ρ(x) − σ(x))y(x)z(x)dx = 0.
a a a

Using the fact that w(a) = w(b) = 0, we are left with

Z b
0 0
z(a)w (a) − z(b)w (b) + (ρ(x) − σ(x))y(x)z(x)dx = 0.
a

First two terms are non-negative, the integral is strictly positive. Contradiction.

Solutions and hints to √ problems of chapter - IV Imitate the proof of Riemann-Lebesgue lemma
to show that |J0 (x)| ≤ c/ x for some constant c. We use the integral representation and look at the
integral Z 1
eitx
I= √ .
−1 1 − t2
Performing the change of variables s = t + πx ,
Z π +1
x eitx
I=− .
1 − (t − πx )2
p
π
x
−1

83
Adding we get π
Z 1 Z
x
+1
itx
2I = e A(x, t)dt + eitx B(x, t)dt (1)
π
x
−1 1

The student is invited to write out the expression for A(x, t) and B(x, t). Further,
π π
Z
x
+1 Z
x
+1
itx
e B(x, t)dt ≤ |B(x, t)|dt.
1 1

The MVT for integrals is not directly applicable

√ (why?) but the integral can be evaluated in terms of
−1
sin and LMVT can be applied to get the c/ x estimate. Equivalently the MVT for integrals can be
applied after shrinking the intervals by arbitrarily small amount. We now have to deal with the first
integral.  1 1 
A(x, t) = √ −p ≥ 0.
1 − t2 1 − (t − πx )2
So estimating the integral we again evaluate it in terms of sin−1 and get the value
 π 
2 sin−1 1 − sin−1 ( − 1)
x
which behaves like c/x for large x. The proof is complete.

Euler’s formula for ζ(2k).

1. We begin by computing the Fourier series of cos(ax) where a ∈/ Z. Well, bn = 0 for all n ∈ N.
Now,
1 π
Z
sin aπ
a0 = cos(ax)dx = .
π 0 aπ
For n ≥ 1,
π
(−1)n−1 2a sin πa
Z
2
an = cos(ax) cos(nx)dx =
π 0 n 2 − a2
We apply Dirichlet’s theorem and divide by sin aπ:
∞
π cos ax 1 X (−1)n−1 cos nx
= + , x ∈ [−π, π]
sin aπ a n=1 n2 − a2

Putting x = 0 and x = π we get

∞
1 X (−1)n−1 2a
πcosec(aπ) = + ,
a n=1 n2 − a2
∞
1 X 2a
π cot(aπ) = + ,
a n=1 a2 − n2

2. Let ζ(2k) = ∞ −2k

P
m=1 m . Find the generating function f (z) for the sequence {ζ(2k)} and write
the result in terms of exponentials.

84
Solution: We have ∞ ∞
X X
2k 1
f (z) = z .
k=1 m=1
m2k

It is easy to convince yourself that ζ(2k) has reasonable bound (for example 2−2k + 3−2k < 2 · 2−2k
etc.,) which makes it easy to interchange the order of summation and write
∞ X ∞
X z 2k
f (z) = .
m=1 k=1
m2k

The inner sum is a geometric series which can be readily summed:

∞ ∞
X z2 zX 2z z 1
f (z) = = = − π cot πz −
m=1
m2 − z 2 2 m=1 m2 − z 2 2 z

Writing the result in exponential form we get

1 iπz iπz
f (z) = − − 2πiz
2 2 e −1
Replace z by iz and we get
∞
X
k 2k 1 πz πze2πz 1 πz πz
(−1) z ζ(2k) = + − 2πz = − − 2πz .
m=1
2 2 e −1 2 2 e −1

We now compare it with the generating function for Bernoulli numbers

∞
X Bm (0)z m z
= z .
m=0
m! e − 1

Replacing z by 2πz and dividing by 2 and remembering that B0 (0) = 1, B1 (0) = −1/2,
∞
1 πz X Bm (0)z m (2π)m z
− + = 2πz .
2 2 m=2
m! e −1

We conclude that B2m+1 (0) = 0 and further

∞ ∞
X
k 2k 1 X B2m (0)z 2m (2π)2m
(−1) z ζ(2k) = − .
m=1
2 m=1 (2m)!

Comparing the coefficient of z 2k on both sides we derive Euler’s formula

(−1)k−1 B2k (0)(2π)2k

ζ(2k) = .
2 · (2k)!

Determine the Fourier series for the function f (x) = x2 on [−π, π]. Use Parseval formula to show that

1 1 π4
1+ + + · · · = .
24 34 90
85
The function being an even function bn = 0 for every n and
Z π
1 2 π2
a0 = x dx =
2π −π 3

and π
4(−1)n
Z
1
an = x2 cos nxdx =
π −π n2
Appealing to Parseval’s formula
π ∞
π4
Z
1 2 1X
|f (x)| dx = = |a0 |2 + (|an |2 + |bn |2 )
2π −π 5 2 n=1

Putting in the computed values and simplifying we get the stated result.

86
 Find a harmonic function u in the unit disc in the plane satisfying the boundary condition

u(cos θ, sin θ) = | sin θ|.

Solution: Use the Poisson formula

π
(1 − r2 ) sin t
Z
1
u(r cos θ, r sin θ) = .
2π 0 1 + r2 − 2r cos(t − θ)

Put t − θ = s and compute the resulting integral. With a little patience you will get

(1 − r2 ) cos θ  1 + r2 + 2r cos θ  (1 − r2 ) sin θ

log − +
4πr 1 + r2 − 2r cos θ 4r
1 + r2 n 1 + r  1 + r o
tan−1 cot(θ/2) + tan−1 tan(θ/2)
2πr 1−r 1−r

Solve the problem using Fourier series. This time the solution will be in the form of an infinite series.

87
 Let u(x, t) be a smooth solution of the heat equation in the upper half-plane t ≥ 0 such that
u(x, t) = u(x + 2π, t) for all x ∈ R and for all t ≥ 0. Show that the energy
Z π
E(t) = (u(x, t))2 dx
−π

is a monotone decreasing function of time. Prove the same result if the integral is over any interval
of length 2π. Deduce that a smooth 2π−periodic solution of the heat equation with a given initial
condition, if it exists, is unique.

Solution: Multiply the PDE by u and integrate over [−π, π]. We get
1d π
Z Z π
2
(u(x, t)) dx − uuxx dx = 0.
2 dt −π −π

Integrating by parts once we get

1d π
Z Z π
2
(u(x, t)) dx + (ux (x, t))2 dx + uux − uux = 0. (1)
2 dt −π −π −π π

Now u(x + 2π, t) = u(x, t) for all x ∈ R and for all t ≥ 0 so differentiating with respect to x we see that
ux is also periodic with period 2π and so the two boundary terms in (1) cancel out and we conclude
d π
Z
0
E (t) = (u(x, t))2 dx ≤ 0,
dt −π
proving that the energy is monotone decreasing. Now if u1 (x, t) and u2 (x, t) are two 2π−periodic
solutions then v(x, t) = u1 (x, t) − u2 (x, t) is also a 2π−periodic solution of the heat equation with
zero initial condition. For this the energy function iz zero at time t = 0. Since the energy is non-
negative and monotone decreasing, it is identically zero. Thus v(x, t) = 0 for all t ≥ 0 which proves
the uniqueness assertion.

Exercise Solve the initial-boundary value for the heat equation on the rectangle:

ut = uxx , u(−π, t) = u(π, t) = 0, u(x, 0) = π 2 − x2 . (1)

Solution: We seek special solutions of the form f (x)g(t) and substituting this Ansatz into the PDE
we get the pair of ODEs:
f 00 (x) + p2 f (x) = 0, g 0 (t) + p2 g(t) = 0.
Thus the special solution is αx + β (corresponding to p = 0) or
2
e−p t (A cos px + B sin px), p 6= 0.

Putting in the boundary conditions u(±π, t) = 0 results in the Sturm Liouville problem

f 00 (x) + p2 f (x) = 0, p 6= 0.

with Dirichlet boundary conditions

A cos pπ + B sin pπ = 0, A cos pπ − B sin pπ = 0.

88
Note that p = 0 is not an eigen-value. Since we are looking for non trivial solutions A and B cannot be
both zero so the determinant of this system must be zero which means cos pπ sin pπ = 0 which implies
2p ∈ Z. We have the list of eigen values

n, (2n − 1)/2, n = 1, 2, . . .

and the corresponding eigen functions

1
sin nx, cos(n − )x, n = 1, 2, 3, . . . .
2
So the solution is of the form
∞ ∞
X 2
X 2 t/4 1
u(x, t) = e−n t Bn sin nx + e−(2n−1) An cos(n − )x
n=1 n=1
2

To determine the sequence of constants, put in the initial condition t = 0 and we get
∞ ∞
2 2
X X 1
π −x = Bn sin nx + An cos(n − )x
n=1 n=1
2

Using orthogonality of eigen-functions the student can readily compute the sequence of coefficients An
and Bn .

Question: It is easy to see that Bn = for every n (how?) and

∞
X 1
π 2 − x2 = An cos(n − )x
n=1
2

In the case of usual Fourier series of an even function, one expects a constant term. Isn’t it strange
that in this case there is no A0 term in the series? Is there some contradiction of some sort??

Problem 33:

Solution: Following the standard procedure which should by now be clear to the students, the
solution can be expressed as
∞
2
X
u(x, t) = A0 + e−n t (An cos nx + Bn sin nx).
n=1

Putting in the initial condition

∞
X
φ(x) = A0 + (An cos nx + Bn sin nx).
n=1

Using the formulas for the Fourier coefficients,

Z π ∞
1 X 2

u(x, t) = 1+2 e−n t (cos nx cos ns + sin nx sin ns) φ(s)ds
2π −π n=1

89
Solutions to selected problems in chapter - V Prove the Riemann Lebesgue Lemma for Fourier
transforms. Suppose f (t) is an absolutely integrable continuous function then its Fourier transform
fb(ξ) tends to zero as ξ → ±∞.

Solution: Let  > 0 be arbitrary. Then there is an M > 0 such that

Z
|f (t)|dt < /4
|t|≥M

π
Let us make the change of variables t = s + ξ
in the integral defining the Fourier transform and add.
Z Z M
π π −itξ
2b
t(ξ) = (f (t) − f (t + ))e−itξ dt + (f (t) − f (t + ))e dt.
|t|≥M ξ −M ξ

The first piece is less than /2 in absolute value. To deal with the last piece note that the function is
uniformly continuous on [−M, M ] and so there is a η > 0 such that
π π
|f (t) − f (t + )| < /4M, | | < η.
ξ ξ
So the last integral in absolute value is less than /2 as soon as |ξ| > π/η. The proof is complete.

Solution to problem 4: Let I(ξ) denote the Fourier transform of 1/(a2 + x2 ) (a > 0). so that
Z ∞
cos ξxdx
I(ξ) = 2 2
−∞ x + a

The function I(ξ) is even (why?) and so we may assume that ξ > 0. We take the Laplace transform
of both sides with respect to ξ and we find
Z ∞
dx
L(I) = s 2 2 2 2
−∞ (x + a )(x + s )

Using partial fractions we can complete the integration and we get π/(a(s + a)). Thus
π −a|ξ|
I(ξ) = e
a

Solution to problem 7: We shall see that in the course of the computation we arrive at the following
integral that we shall determine first.
Z ∞
(1 − cos(bt))dt
φ(b) =
−∞ t2

To compute φ(b) we may assume b > 0 and take Laplace transform with respect to b and we get
Z ∞n
1 ∞ dt
Z
1 s o
Lφ = 2
− 2 2 2)
dt = 2 + t2
= π/s2 .
−∞ st t (s + t s −∞ s

Thus the integral is π|b|. We are now ready to compute the Fourier transform of sin2 t/t2 .

90
Exercise: Compute the Fourier transform of f (t) = (sin2 t)/t2 .

Solution: Let us denote the Fourier transform by I(ξ) and take the Laplace transform of I(ξ) with
respect to the ξ variable. Denoting by L(s) the Laplace transform we get

s ∞ (1 − cos 2t)dt
Z
L(s) =
2 −∞ t2 (s2 + t2 )

The integral can be decomposed using partial fractions into three integrals namely,
Z ∞ Z ∞ Z ∞
cos(2t)dt (1 − cos 2t)dt dt
2 2
, 2
, 2 2
.
−∞ s + t ∞ t −∞ s + t

The first two have been determined and the last is elementary. Thus
π π π
L(s) = − 2 + 2 e−2s .
s 2s 2s
Now using the shifting theorem for Laplace transforms we get I(ξ) = 0 if ξ ≥ 2 and I(ξ) = π − πξ/2
if 0 ≤ ξ ≤ 2. Extend as an even function. The student is invited to draw a picture. Later we shall
see another simpler way to get this result. Compute the convolution f ∗ f where f (t) = 1 if |t| ≤ 1
and zero outside. Draw the graph of this convolution and compare it with the graph of the Fourier
transform.
For the formal computation of the Fourier transform of sin t/t. Denoting by I(ξ) the Fourier
transform we get assuming a > 0 and ξ > 0,
Z ∞
sin at cos ξtdt
I(ξ) = .
−∞ t

The same method would work. Take the Laplace transform with respect to ξ and we get
Z ∞
sin tdt
s 2 2
.
−∞ t(s + t )

So we now have to calculate an integral of the form

Z ∞
sin xtdt
2 2
−∞ t(t + s )

The value of this integral is π(1 − e−s )/s2 (How?) and so the Fourier transform fb(ξ) equals πξ if
0 ≤ ξ ≤ 1 and zero if ξ ≥ 1. Extend as an even function.

Problems 14-16:

14 An easy calculation gives the following ODE for u:

u00 − x2 u = −(1 + 2λ)u.

15 Now using theorem 41 we see at once that if u is a solution of this ODE that lies in S then u
b
also satisfies the same ODE. So the transformed ODE is invariant under Fourier transform.

91
 16 If λ is an integer n we know that the original ODE has a polynomial solution Hn (x) and so
we have a solution Hn (x) exp(−x2 /2) for the transformed ODE and this solution lies in S. The
Abel-Liouville formula for the Wronskian of two linearly independent solutions is a non zero
constant and so only one solution can lie in S.
Well, otherwise the Wronskian would tend to zero as x −→ ∞ and being a constant would have to
be identically zero. Now since u(x) = Hn (x) exp(−x2 /2) is a solution lying in S we see that u
b is also
a solution of the same equation in S and these two must be linearly dependent by the last exercise
which means
u
b = cu
for some constant and u is an eigen-vector for the Fourier transform as a linear map from S to itself.
One can show that there is an orthonormal basis of eigen-vectors but that uses more analysis.

Solution to problem 22: Following the hint provided,

Z c Z k Z c Z k
du c
Z
sin kt
f (t)dt = du (cos ut)f (t)dt = (sin ut)0 f (t).
0 k 0 0 0 u 0
Hence, Z c Z k n Z c
sin kt du o
f (t)dt = f (c) sin uc − (sin ut)f 0 (t)dt.
0 k 0 u 0
The first piece on the RHS has limit f (c)π/2 as k → ∞ (why?). The second piece can be written as
Z c Z kc
π c 0
Z
0 sin v
f (t)dt dv −→ f (t)dt (How?)
0 0 v 2 0
The rest is clear.

Solution to problem 23: Suppose 0 < c < π. We simply write the integrand as
 sin kt  tf (t)  sin kt
= g(t)
t sin t t
where g(t) = tf (t)/ sin t. We can apply the previous result with g(t) in place of f (t).

Solution to problem Assume π < c < 2π. Then as before

Z π/2
sin kt π
lim f (t)dt = f (0).
k→∞ 0 sin t 2
We now take up the remaining two pieces. The substitution π − t = s transforms the integral
Z π
sin kt
f (t)dt
π/2 sin t

to (remembering that k is odd),

Z π/2
sin ks
f (π − s)ds
0 sin s
which has limit f (π)π/2. Finally for the third piece with s = t − π we get
Z c Z c−π
sin kt sin ks
f (t)dt = f (π + s)ds
π sin t 0 sin s
which has limit f (π)π/2 as k → ∞ through odd values. It is clear what happens when c = π.

92
To discuss the convergence of the integral in problem 40. Let us denote the integral without
the (2π)−1 factor as I(x). Assume to begin with x > 0. We make the change of variables

ξ 3 + xξ = t

and note that the derivative being positive makes this licit. Thus we get
Z ∞ Z ∞ Z ∞
cos tdt 1 d sin t ξ sin tdt
I(x) = 2
= 2
dt = −6 2 2
.
−∞ (3ξ + x) −∞ (3ξ + x) dt −∞ (3ξ + x)

Now, the integral over [−M, M ] is anyway finite and we need merely focus on what happens when
ξ is large. The integrand is easily estimated and convergence can be established. Explain why the
condition x > 0 is not essential for the argument.

93