Вы находитесь на странице: 1из 59

B5a: Techniques of Applied Mathematics

Derek E. Moulton
based on notes by Andreas M
unch
November 21, 2013

Contents
1 Introduction

2 Delta function and distributions


2.1 Delta function . . . . . . . . . . .
2.2 Approximating the delta function
2.3 Properties of delta function . . .
2.4 Distributions . . . . . . . . . . .
2.5 Operations on distributions . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

3
3
3
3
5
9

3 Greens Function Approach


3.1 Point heat source . . . . . . . . . . . .
3.2 Greens function . . . . . . . . . . . .
3.3 General linear BVP . . . . . . . . . .
3.3.1 A note on boundary conditions
3.4 General Greens Function . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

10
10
11
12
13
13

4 Eigenfunction expansion approach


4.1 Eigenvalue problems . . . . . . . . . . .
4.1.1 Eigenvalue problems from PDEs
4.1.2 Some simple solutions . . . . . .
4.2 Adjoint . . . . . . . . . . . . . . . . . .
4.2.1 Eigenfunction Properties . . . . .
4.3 Inhomogeneous solution process . . . . .
4.4 Another note on boundary conditions .
4.5 Example . . . . . . . . . . . . . . . . . .
4.6 Connection with Greens Function . . .
4.7 Connection with linear algebra . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

14
15
15
15
16
17
18
18
19
20
22

.
.
.
.
.

.
.
.
.
.

5 Sturm-Liouville theory
5.1 Transforming an operator to SL
5.2 Inhomogeneous SL problems . .
5.3 Further properties . . . . . . .
5.3.1 Other useful tidbits . .

form
. . .
. . .
. . .

6 Fredholm Alternative
6.1 A closer look at SL solution . . . . .
6.2 Zero eigenvalue . . . . . . . . . . . .
6.3 FAT . . . . . . . . . . . . . . . . . .
6.4 Homogeneous vs inhomogeneous BC
6.5 Examples . . . . . . . . . . . . . . .
6.6 FAT - Linear algebra version . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

23
24
24
25
26

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

27
27
28
28
30
30
31

7 Integral equations
7.1 Fredholm Integral Equation . . . . . . . . . . . . . . .
7.2 Degenerate (separable) kernels . . . . . . . . . . . . .
7.2.1 Adjoint for Integral operators . . . . . . . . . .
7.3 Inhomogeneous equations . . . . . . . . . . . . . . . .
7.4 2nd kind FIE . . . . . . . . . . . . . . . . . . . . . . .
7.5 Imhomogeneous 2nd kind FIE with degenerate kernel
7.6 Non-degenerate/nonseparable kernels . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

33
34
35
37
38
40
41
43

8 Singular points of differential equations


8.1 Ordinary points . . . . . . . . . . . . . .
8.2 Singular points . . . . . . . . . . . . . .
8.3 Frobenius method for 2nd order ODEs .
8.4 Derivative method . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

45
45
46
48
51

.
.
.
.
.

52
52
54
56
56
58

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.

9 Special functions
9.1 Bessel Functions . . . . . . . . . . . . . .
9.2 Bessel functions of first and second kind .
9.3 Legendre equation and Legendre functions
9.3.1 Properties . . . . . . . . . . . . . .
9.4 Generalisation: Orthogonal Polynomials .

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

Introduction

In this course, we will explore various techniques for solving differential


equations, building on the basic techniques you have learned from your 2nd
2

year Differential Equations courses. A primary concern will be finding ways


to solve and understand linear boundary value problems (BVPs), that is an
ordinary differential equation (ODE)
Lu(x) = f (x), a < x < b

(1)

where L is a linear differential operator of the form


Lu = an u(n) (x) + an1 u(n1) (x) + + a1 u (x) + a0 u(x)

(2)

and the function f (x) on the right hand side (RHS) is a forcing function in
the system. Along with equation (1) are extra boundary conditions at x = a
and x = b.
Some questions we will consider:
1. How do we solve the system for an arbitrary function f (x)?
2. Is there always a solution? If so, is it unique?
3. What is the effect of the boundary conditions?
4. Can we solve if the ak = ak (x) are functions of x?
The general system (1) is very common in applied mathematics. Some examples of physical systems that can be modelled using a linear BVP include
the deformation of an elastic beam, the vibrations of a drum, or the diffusion
of a chemical. An intuitive example that we will consider more closely is
heat conduction in a rod.
Stationary heat conduction in a rod

x=0

111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
q(x+ x)
q(x)
000000000
111111111
000000000
111111111
000000000
111111111
000000000
x111111111
x+ x
Figure 1: Heat conduction in a rod

Defining:
u(x) temperature field (unknown)
f (x) heat source density (given)
3

x=1

q(x) heat flux,


energy conservation applied to a segment of the rod gives
x+x

q(x + x) q(x) =

f (s)ds
x

Divide by x, then let x 0, we get


q (x) = f (x)

( = d/dx)

Now using Fouriers law (with heat conductivity cross sectional area =
unity)
q(x) = u (x)
Supposing that the temperature at the ends of the rod is held fixed, the
temperature field satisfies
u (x) = f (x)

0<x<1

u(0) = 0, u(1) = 0

(3)
(4)

Delta function and distributions

Lets consider more closely at the example of heat conduction in a rod, and
try to find a solution to (3)-(4). As a simple case, we might consider a point
heat source, say located at the middle of the rod. How would we describe
such a situation mathematically? What should we use for the function f (x)?

2.1

Delta function

The notion of a point source is described by the delta function , characterised by properties

(x) dx = 1.

(x) = 0 x = 0,

(5)

The first property captures the notion of a point function. The second
property constrains the area under the curve (which you might think of as
infinitely thin and infinitely high). This is an idealized point source at x = 0,
a point source at x = a would be given by (x a).
The problem is that no classical function satisfies (5) (think: any function
that is non-zero only at a point is either not integrable or integrates to zero).
4

2.2

Approximating the delta function

One way around this is to replace by an approximating sequence of increasingly narrower functions with normalized area, i.e. fn (x) where

fn (x)dx = 1 n,

lim fn (x) = 0 x = 0.

Increasing n

n/2

-1/n

1/n

Figure 2: Hat functions, see eqn. (6)


Example: hat functions
fn (x) =

0
n/2

for |x| > 1/n


for |x| 1/n

(6)

You can verify the fn (x) approach (x) as n .

2.3

Properties of delta function

We have defined by (5). We can use the approximating functions to obtain


further properties.
Sifting property. What happens when is integrated against another
function?
5

Let f (x) be a continuous function, and F (x) =


Now consider approximating sequences:

f (s)ds its antiderivative.

(x a)f (x)dx = lim

fn (x a)f (x)dx,

and if fn are the hat functions (6),


a+1/n

= lim

n a1/n

n
F (a + (1/n)) F (a (1/n))
f (x)dx = lim
n
2
2/n
F (a + s) F (a s)
= F (a) = f (a).
s0
2s

= lim
Thus, we have

(x a)f (x)dx = f (a)

if f is continuous at a.

(7)

if f is continuous at x = 0.

(8)

In particular,

(x)f (x)dx = f (0)

Thus, the delta function can be seen to sift out the value of a function at a
particular point.
Antiderivative of (x). The antiderivative of the delta function is the
so-called Heaviside function,
x

(s)ds = H(x)

0
1

x<0
x > 0.

(9)

Note that (9) follows by integrating the sequence of approximating functions and showing that the limit is the Heaviside function. That is, if
x
Hn (x) = fn (s)ds, then limn Hn (x) = H(x). (We leave this detail as
an exercise!)
Back to the BVP. One approach is to solve the BVP using the approximating functions. Of course, we dont know how to do this yet, and we
would still be left with the question of whether the solution to the BVP is
dependent upon our approximating sequence.
A more elegant approach is to deal directly with the delta function. But
to do this properly, we must generalise the notion of functions. For this we
need the theory of distributions.
6

2.4

Distributions

Perhaps the most important feature of the -function: when integrated


against a continuous function, it sifts out the value at x = 0:

(x)f (x)dx = f (0).

It is the operation of on another function that defines the property. This


is a key idea in the theory of distributions, in which a generalized function
is only thought of in relation to how it affects other functions when integrated against them.
We define the delta distribution such that when it operates on a test function
, it sifts out the value (0) R.
We write this as
, (0),
where is the -distribution and is the test function. , reads as
applied to .
In general, we define distributions to be functionals, i.e. they map test functions to real numbers and have additional properties.
Test functions : R R
C0 (R), which is short for:
C (R) differentiable any number of times
has compact support, i.e. supp [X, X] for some X > 0, i.e.
(x) = 0 x
/ [X, X].
Example (see figure):
Let C > 0, > 0
C; =

exp
0

C
2 (xa)2

otherwise

for a < x < a +

One can show (for all integer n 0):


d
C; (x) = 0
xa+ dxn
lim

d
C; (x) = 0
xa dxn
lim

(10)

C;(x)

Matches zero
infinitely smoothly

Matches zero
infinitely smoothly

a-

a+

Figure 3: Sample test function, corresponding to (10)


.
Definition: Distribution
A distribution u is a functional, i.e. it maps test functions C0 (R) onto
real numbers,
u : C0 (R) u, R

( u, instead of u())

(11)

where the mapping needs to satisfy the following properties:


(i) It is linear (linear functional)
u, + = u, + u,

, R

, C0 (R)

(ii) It is continuous, in the following sense: if n is a sequence of test


functions, such that
n (x) 0

as

then
u, n 0
as a sequence of real numbers.

More on (ii): To show continuity, what we really need is to be able to switch


the order of the action of the distribution (integration) and the limit, that
is (ii) will hold if
lim u, n = u, lim n .
n

It turns out that we can do this if the following holds:


(**) X > 0 there exists C > 0, and integer N 0, such that
| u, | C

max

mN

dm
dxm

with support in [X, X].


Hence in practice we will want to show (**) to show continuity.
Examples
(1) Delta distribution
, = (0)
(i)
(**) | , | = |(0)|

max

X<x<X

|(x)|

with support of in [X, X].

i.e. condition (ii) is satisfied with C = 1, N = 0.


Generalisation
Let a R, n 0. Define Dn , = (n) (a) (nth derivative).
This is a distribution (see Problem Sheet 1!)
(2) Functions as distributions
(a) let f be a continuous function, claim:

f (x)(x)dx

f, =

defines a distribution. Check:


(i) Well-defined, f, R C0 (R) and linear
(**) Let X > 0 be given. Claim (ii) holds for
X

C = C(X) =
X

|f (x)|dx and N = 0 :

| f, | = |

f (x)(x)dx|

f (x)(x)dx| = |

which by the estimation lemma


X

|f (x)|dx

max (|(x)|) = C

X<x<X

max (|(x)|)

<x<

Remark: Different continuous functions induce different distributions. Think: how would you show this?
(b) Integrable f , e.g. Heaviside function H(x)

(x)dx

H(x)(x)dx =

H, =

Check (i),(**) as in (a).


Remark: Different functions can lead to the same distribution.
Distributions induced by integrable functions are called regular
distributions; singular distributions if not. The -distribution is
an example of a singular distribution.

2.5

Operations on distributions

Now we consider some operations that can be performed on distributions.


Let u1 , u2 , u be distributions, and f1 , f2 , f be integrable functions (or the
regular distributions induced by them)
Linear combinations of distributions. Let 1 , 2 R.

(1 f1 (x) + f2 (x))(x)dx

1 f1 + 2 f2 , =

f2 (x)(x)dx

f1 (x)(x)dx + 2

= 1

= 1 f1 , + 2 f2 ,
Thus, define 1 u1 + 2 u2 for general distributions u1 , u2 via
1 u 1 + 2 u 2 , 1 u 1 , + 2 u 2 ,

C0 (R)

If u1 , u2 are distributions, is 1 u1 + 2 u2 a distribution? Need to check


(i),(**), but well skip this here.

10

Differentiation of distributions. Start with a continuously differentiable


function f .

f , =
int by parts

f (x)(x)dx

f (x) (x)dx = f,

Thus, for general distributions u, define


u , u,

C0 (R)

If u is distribution, can we be sure that u : u, is also a distribution? (try it as an exercise!)


Example. Let H be the Heaviside function, or the distribution it induces,
i.e.

(x)dx
H(x) (x)dx =
,
H

H -distribution

Hfunction

Show that H = .
H ,

= H,
Def. of derivative of a distribution

= (x)dx
see earlier example
x=
= |x=0
= (0)
has compact support
= ,
Def. of -distribution

Translation: similar considerations as before, upshot (a R, u distr):


u(x a), (x)

chg of var

u(y), (y + a) = u(x), (x + a)

Example: (x a), (x) = (x), (x + a) = (a)


Convergence of a sequence of distributions u, u1 , u2 , . . . distributions.
Convergence uj u as j means:
lim uj , = u,

C0 (R)

Similarly: if u() is a family of distributions with a continuous parameter


, then
convergence u() u(0 ) for 0 means:
lim u(), = u(0 ),

11

C0 (R)

Greens Function Approach

Armed with the delta function as a tool, we return to the task of solving our
BVP. In this section we will derive a method for solving using the so-called
Greens function, an approach that relies on the delta function.

3.1

Point heat source

Lets consider the heat conduction BVP with a point heat source of unit
strength at the centre of the rod:
u (x) = (x 1/2),

0<x<1

u(0) = u(1) = 0.

(12)
(13)

Since (x 1/2) = 0 x = 1/2, this implies


u (x) = 0,

0 < x < 1/2, 1/2 < x < 1

(14)

We can easily solve (14) in each of the two separate domains [0, 1/2) and
(1/2, 1] and then apply the BC (13). But be careful: there are two constants
of integration for each domain, meaning four unknown constants total, and
only two boundary conditions.
As you might expect (since (x 1/2) has vanished from (14)), the extra
two conditions come in at x = 1/2. To derive the extra conditions, imagine
integrating equation (12) across x = 1/2:
1/2+
1/2

u (x) dx =

1/2+
1/2

(x 1/2) dx,

(15)

where 1/2 (1/2+) signifies just to the left (right) of 1/2. Using property
(5) of the delta function, we have
1/2+

u ]1/2 = 1

u (1/2+) u (1/2) = 1.

(16)

That is, the presence of the delta function defines a jump condition on u . 1
The other extra condition needed comes as a requirement that u(x) is continuous across the point source, that is
1/2+

u]1/2 = 0.
1

Here, u() = limx u(x), and u(+) = limx u(x)

12

(17)

More on this condition below. Solving Equations (14), (13) along with extra
conditions (16) and (17), we obtain the solution
u(x) =

3.2

x
2
x
2

0 < x < 1/2


1/2 < x < 1.

1
2

(18)

Greens function

To motivate the idea of the Greens function, consider the heat conduction
problem with an arbitrary heat source:
u (x) = f (x),

0<x<1

u(0) = u(1) = 0

(19)
(20)

Imagine now describing f by a distribution of point heat sources with varying strength; that is at point x = we imagine placing the point source
f ()(x ).
The idea of the Greens function is to introduce such an extra parameter ,
and consider the system
g (x, ) = (x ),

0<x<1

g(0, ) = g(1, ) = 0.

(21)
(22)

Note that prime denotes differentiation with respect to x, while is more


like a place-holding variable. So, we have replaced f (x) by a delta function,
in order to solve for the Greens function g(x, ).
We have seen how to solve (21), (22) in the last section. The Greens function
is
(1 )x
0<x<
g(x, ) =
(23)
(1 x)
< x < 1.

How to get back to the solution of (19), (20)? For each , the Greens
function gives the solution if a point heat source of unit strength were placed
at x = . Conceptually, then, to get the full solution we must add up the
point sources, scaled by the value of the heat source at each point:
1

g(x, )f () d.

u(x) =

(24)

To verify that this is indeed a solution, we can plug (24) into (19):
u (x) =

1
0

g (x, )f () dx =
13

1
0

(x )f () dx = f (x)

(25)

3.3

General linear BVP

We now consider a general nth order linear BVP with arbitrary continuous
forcing function,
Lu(x) = an u(n) (x) + an1 u(n1) (x) + + a1 u (x) + a0 u(x) = f (x) (26)
for a < x < b, where each ai = ai (x) is a continuous function, and moreover
an (x) = 0 x2 . Along with (26) are n boundary conditions, each a linear
combination of u and derivatives up to u(n1) , evaluated at x = a, b. For
instance, in the case n = 2, the general form is:
B1 u 11 u(a) + 12 u (a) + 11 u(b) + 12 u (b) = 1

B2 u 21 u(a) + 22 u (a) + 21 u(b) + 22 u (b) = 2 .

3.3.1

(27)

A note on boundary conditions

Before proceeding, we make an important observation regarding the form of


BVP. In the general case of an inhomogeneous system with inhomogenous
boundary conditions,
Lu = f (x)
Bi u = i

(28)

a useful technique is to split the system in two, i.e. solve both


Lu1 = f (x),

Bi u 1 = 0

(29)

and
Lu2 = 0,

Bi u 2 = i .

(30)

Here, solving for u1 (x) has the difficulty of the forcing function but with
zero BCs while the other equation is homogeneous but has the non-zero
BCs. Due to linearity, it is easy to see that u(x) = u1 (x) + u2 (x) solves the
full system (28).
Since this decomposition can always be performed, and since solving (30)
tends to be an easier matter (for linear systems!), it is safe for us to primarily focus on the technique of solving the system (29), i.e. homogeneous
boundary conditions. But keep in mind what to do if confronted with imhomogeneous BC!
2

Well return to the case where an (x) = 0 somewhere in the domain later in the course.

14

3.4

General Greens Function

In the same way as in Section 3.1, to solve (26) with homogeneous BC


Bi u = 0, i = 1 . . . n 1,
we first determine the Greens function by solving
Lg(x, ) = (x ),

a<x<b

(31)

Bi g = 0.
As before,

Lg(x, ) = (x )
implies
Lg(x, ) = 0

on a < x < , < x < b,

i.e. we have a homogeneous problem to solve on two separate domains. As


before, we require extra conditions, which come by integrating Lg(x, ) =
(x ) across x = :
+

an g

(n)

(x, ) + + a0 g(x, ) d =

(x ) d.

(32)

The right hand side clearly integrates to one. If we were to perform an


integration by parts on the first term of the left hand side, we would obtain
an (x)g

(n1)

(x, )]+

(an1 an )g (n1) + + a0 g(x, ) d = 1.

This equation is balanced by setting a jump condition on the n 1st derivative:


g (n1) (x, )]+
= 1/an (),
and taking all lower derivatives to be continuous across x = :
g (j) (x, )]+
= 0,

j = 0, 1, . . . n 2.

Once the Greens function is determined, the solution to the BVP is given
by
b

g(x, )f () d.

u(x) =
a

15

(33)

Eigenfunction expansion approach

In this section we consider an alternative approach to solving linear BVPs,


using an eigenfunction expansion. We will restrict to 2nd order operators3
Ly = a(x)y (x) + b(x)y (x) + c(x)y(x) = f (x),

a < x < b.

The general idea is to seek a solution in the form


y=

ci yi (x),
i

where the yi (x) are eigenfunctions.

4.1

Eigenvalue problems

A 2nd order BVP eigenvalue problem is of the form


Ly = y,

a<x<b

(34)

where is an unknown eigenvalue corresponding to eigenfunction y(x)4 . To


this we add a homogeneous boundary condition at x = a and x = b, which
we will denote by B1 (a) = 0, B2 (b) = 0. Note that y(x) 0 is a solution for
all values of , however this is boring and we are interested in non-trivial
solutions.
4.1.1

Eigenvalue problems from PDEs

Eigenvalue problems are very common. They arise, e.g., from PDE problems
utt = Lu, by putting u(x, t) = eit y(x) and then = 2 .
4.1.2

Some simple solutions

Lets look at some simple cases, solvable using techniques youve learned
before:
3
We have switched from u(x) to y(x) for no particular reason at all. Lets say u was
tired.
4
The minus sign has been added for consistency with other ODE problems and textbooks on ODEs. I will try to keep this convention in this class. A minus sign does not
typically appear in eigenvalue problems in linear algebra, it is sort of a cultural difference
between the two fields.

16

Constant coefficients

Ly ay + by + cy = y

Try y = emx , then:


am2 + bm + (c + ) = 0
Then:
1.
2.
3.
4.

Find roots mi of the quadratic.


the general solution is: y = A1 em1 x + A2 em2 x
apply first BC to relate A1 and A2 .
apply second BC to determine values for .

Cauchy-Euler
Try y = xm , then:

Ly ax2 y + bxy + cy = y
am(m 1) + bm + (c + ) = 0

Then y = A1 xm1 + A2 xm2 .

4.2

Adjoint

As we shall see, to solve a BVP using an eigenfunction expansion will require


the adjoint problem. For operator L with homogenous BC, the adjoint
problem (L BC*) is defined by the inner product relation
Ly, w = y, L w .

(35)

Example
d2 y
Let Ly =
with a
dx2

find L w, such that

x
b

b, y(a) = 0 and y (b) 3y(b) = 0. We wish to


b

(w)(y )dx =

(y)(L w)dx

To do this, we need to shift the derivatives from y to w using integration by


parts:
b
a

wy dx = wy |ba

w y dx

= wy w y|ba +
17

b
a

yw dx

The integral gives the formal part so:


L w =

d2 w
.
dx2

The inner product only includes integral terms, so the boundary terms must
vanish, which will define boundary conditions on w. Thus, we require
w(b)y (b) w (b)y(b) w(a)y (a) + w (a)y(a) = 0.
Using the BCs y (b) = 3y(b) and y(a) = 0, gives:
0 = y(b) 3w(b) w (b) w(a)y (a) + w (a)y(a)
=0

As these terms need to vanish for all values of y(b) and y (a), we can infer
two boundary conditions on w:
y(b): 3w(b) w (b) = 0
y (a): w(a) = 0
If L = L and BC = BC then the problem is self-adjoint. If L = L
but BC = BC we still call the operator self-adjoint. (Some books use
the terminology formally self-adjoint if L = L and fully self-adjoint if both
L = L and BC = BC .
4.2.1

Eigenfunction Properties

There are two fundamental properties of eigenfunctions, which will be vital


to our BVP solution process.
1. Eigenfunctions of the adjoint problem have the same eigenvalues as the original problem
That is,
Ly = y L w = w.
Proof follows immediately from (35), and is left as an exercise.
2. Eigenfunctions corresponding to different eigenvalues are orthogonal
That is, if Lyj = j yj (so L wj = j wj ) and Lyk = k yk (L wk =
k wk ), then for j = k , yj , wk = 0.
18

Proof
j yj , wk

j yj , wk

Lyj , wk

yj , L w k

yj , k wk

= k yj , wk .
But j = k so yj , wk = 0. (The proof is exactly as for matrix
problems.)

4.3

Inhomogeneous solution process

We are now in a position to outline the construction of solution to the BVP


Ly = f (x)
with BC1 (a) = 0, BC2 (b) = 0.
Step 1: Solve the eigenvalue problem
Ly = y,

BC1 (a) = 0, BC2 (b) = 0

to obtain the eigenvalue-eigenfunction pairs (j , yj ).


Step 2: Solve the adjoint eigenvalue problem
L w = w,

BC1 (a) = 0, BC2 (b) = 0

to obtain (j , wj ).
Step 3: Assume a solution to the full system Ly = f (x) of the form y =
i ci yi (x). To determine the coefficients ci , start from Ly = f and
take an inner product with wk :
Ly = f (x)
Ly, wk = f, wk

y, L wk = f, wk

y, k wk = f, wk
k

ci yi , wk = f, wk
i

k ck yk , wk = f, wk
19

(36)

We can solve the last equality for the ck , and we are done! (Do you see
why we needed the adjoint?) Note that in the last step we have used
convergence of y to switch the order of summation and integration,
and used the orthogonality property to kill all terms yj , wk , j = k.

4.4

Another note on boundary conditions

As discussed in Section 3.3.1, we can always remove an inhomogeneous


BC by decomposing the solution. For simplicity, that has been our approach here. However, for completeness it is worth noting that one can
solve BVPs with inhomogeneous BC using an eigenfunction expansion and
without needing a decomposition. The keys are:
1. The eigenfunctions are always determined using homogeneous boundary conditions. Thus, eigenfunctions wont change whether you decompose or not. The difference comes in:
2. In going from Line 2 to 3 of (36), care must be taken in the integration
by parts, as boundary terms will generally still be present. (Can you
see why?) These extra boundary terms then carry through to the
formula for the ck .

4.5

Example

Let y = f (x) for x with 0

1, y(0) = and y(1) = . Then:

BCs Incorporated Solution Route


1. Solve y = y, with y(0) = 0 and y(1) = 0.
We get yk (x) = sin(kx) and k = k 2 2 with k = 1, 2, 3, . . ..
The problem is self-adjoint (show this as an exercise), so wk = yk =
sin(kx) and wk = k wk .

20

2.
y = f (x)
1

wk y dx =

wk f dx
0

0
1

(y wk ywk )|10 +

(y wk ywk )|10 k 2 2 ck
3. Now

1
2
0 sin (kx)

wk f dx
0
1

(y wk ywk )|10 k
(y wk ywk )|10 k ck

wk ydx =

wk f dx

wk ydx =
0

0
1

wk f dx

wk yk dx =
0

sin2 (kx)dx =

wk f dx
0

dx = 1/2, and wk = sin(kx), hence

y wk ywk |10 = k cos(k)y(1) + k cos(0)y(0)


1
1
k(1)k + k k 2 2 ck =
f (x) sin kxdx
2
0

Solving for ck gives us y(x) as a Fourier series.


Decomposed Solution Route
1. Solve two systems separately:
y = f (x),

u = 0,

y(0) = y(1) = 0

u(0) = , u(1) =

2. To solve for y, since BC=0 we can jump straight to the formula


2
f, wk
ck =
=
k yk , w k

1
0 f (x) sin(kx)
k2 2

3. The solution for u is easily obtained as


u = ( )x +
21

dx

4. The full solution is y(x) + u(x).


Although they look different, both approaches give the same solution. Either way, we see that self-adjoint problems are great: they are less work
since the wk s are the same as the yk s. If only there was an easier way to
check if (L, BC) is self-adjoint... We return to this in Section 5

4.6

Connection with Greens Function

Return to the general BVP Ly = f (x). As weve seen from Section 4.3, in
particular taking (36) one step further, its solution can be written as:

y(x) =
k=1

f, wk
yk (x)
k yk , wk

where k = 0 (! Well treat the case of a zero eigenvalue in Section 6). Let
nk = yk , wk (normalisation), then:

y(x) =
k=1
b

=
a

1
k nk
k=1

f (t)wk (t)dt yk (x)


a

1
wk (t)yk (x) f (t)dt
k nk

k(x, t)f (t)dt


a

where

k(x, t) =
k=1

wk (t)yk (x)
.
k nk

This is an eigenfunction expansion of the Greens function:


b

Ly = f y(x) =

g(x, t)f (t)dt


a

as defined and discussed in Section 3, with k(x, t) = g(x, t).


If L = L , then wk = yk and:
g(x, t) =

1
yk (t)yk (x)
k n k

Notice that we easily see from this expansion that if the ODE-BPV is selfadjoint, then the Greens function is symmetric g(x, t) = g(t, x).
22

4.7

Connection with linear algebra

There are direct parallels between linear algebra and linear BVPs:
Linear BVP

Linear algebra
vector v

Rn

function y(x) for a

vw =

k1

vk wk

=vv

norm

f (x)g(x)dx

f, g =
a

inner product

dot product
v

= f, f

norm

vector v w = 0 orthogonal function f, g = 0


Matrix A Linear Differential Operator L
Given a vector v, then the product Av is a new vector. Similarly, given a
function y(x),
dy
d2 y
+ cy
Ly = a 2 + b
dx
dx
evaluates to a new function on a x b.
In linear algebra, a common problem is to solve the equation
Av = b
for v, given matrix A and vector b. Compare that to our general task of
solving Ly = f for y, given operator L and RHS f .
Eigenvalue problems
Linear algebra

Linear BVP

Av = v Ly = y

23

How many eigenvalues?


Linear algebra

Linear BVP

A is n n

L is order n

Solve |A I| = 0

n eigenvalues

eigenvalues

To think about: why the discrepancy??


Adjoint

Self adjoint if

Linear algebra

Linear BVP

A AT

L L
BCs BC s

A = AT

L = L , BC=BC*

Sturm-Liouville theory

SturmLiouville (SL) theory of second order concerns problems of the form:


Ly = r(x)y
where r(x) 0 is a weighting function (previously we had r 1), and the
operator L is of the form
Ly =

d
dx

p(x)

dy
dx

+ q(x)y, a x b

(37)

Lets add the BCs:


1 y(a) + 2 y (a) = 0
3 y(b) + 4 y (b) = 0,
Whats so great about SL? For one, the operator is self-adjoint, L = L , so
each wk = yk .

24

5.1

Transforming an operator to SL form

The key to the SL form is that the first and second derivative terms are
grouped together in the (py ) term. Suppose we are considering the problem
Ly = y (x) + a1 (x)y (x) + a0 (x)y(x) = f (x).
We can transform to a self-adjoint SL operator by multiplying by an integrating factor function (x):
y (x) + a1 y (x) + a0 y = f.
We then choose so that the equation takes the form
d
(py ) + qy = f,
dx
and we can use all of the nice properties of Sturm-Liouville. But wait, is
the problem self-adjoint or isnt it?? The key observation is that we are no
longer solving the same problem. We have transformed to a new operator
= d (py ) + qy
Ly
dx
which does not satisfy the same equation as the original, that is Ly = f
= f . They are both valid, and must ultimately lead to the same
while Ly
answer in the end. We will explore this in Problem Sheet 3.

5.2

Inhomogeneous SL problems

Let L be a SL operator and consider


Ly = f (x)
with homogeneous BCs. The system can be solved with an eigenfunction
expansion in the same manner as in Section 4.3:
Ly = f (x)
Ly, yk = f, yk
y, Lyk = f, yk

y, k ryk = f, yk

k ck yk , ryk = f, yk .
25

(since L = L, wk = yk )

(38)

Thus we obtain the formula


ck =

f, yk
k yk , ryk

(39)

and the full solution is given by


c k yk .

y=
k

5.3

Further properties

Orthogonality.
Due to the presence of the weighting function, the orthogonality relation is
(proof left as an exercise!):
b

yk (x)yj (x)r(x)dx = 0.

(40)

Eigenvalues.
Functions p, q, r are real, so L = L. Thus, taking the conjugate of both sides
of Lyk = k yk gives
L yk = k r yk

yk , L yk = k yk , ryk

also,1 yk , L yk = Lyk , yk = k ryk , yk = k yk , ryk

(41)

k = k
Thus, all eigenvalues are real.
Moreover, if a x b is a finite domain, then s are discrete and countable:
1 < 2 < 3 < < k <
with limk k = .
Eigenfunctions.
The {yk } are a complete set, that is all h(x) with
expanded as
h(x) =
ck yk (x).

h2 r dx < can be

Take an inner product with r(x)yj (x):


ryj , h = ryj ,

c k yk =
26

ck ryj , yk = cj ryj , yj

b
a h(x)yj (x)r(x) dx
b 2
a yj (x)r(x) dx

cj =

Note: Ive used h(x) to make clear that were not talking about the solution
to the BVP, rather we are expanding any function that is suitably bounded
on the same domain.
5.3.1

Other useful tidbits

Here are some other properties that we wont worry too much about, but
might come in handy somewhere in life:
Regular Sturm-Liouville Problems. If the system satisfies all of the
above and the additional conditions
p(x) > 0 and r(x) > 0 on a x b.

q(x) 0 on a x b.

BCs have 1 2 0 and 3 4 0,

then all k 0
Proof: Using yk , Lyk + k ryk = 0,
b
a

a
b

=
a

yry dx = 0
a

y(py ) dx +
b

p(y ) dx

qy 2 dx +

ry 2 dx = 0

ry 2 dx = 0

a
b

p(y )2 dx +

qy 2 dx +

a
b

yqy dx +
a

pyy

y(py ) dx +

b
2

qy dx + pyy
a

ry 2 dx 0

As a side note, the Rayleigh quotient, R[y] = y, Ly / y, ry , is used extensively in analysis.


Comparison theorem: If j > k then the zeroes of yk (x) on a < x < b
lie between successive zeroes of yj (x).
Oscillation theorem [Simplest version]: The k th eigenfunction will have k zeroes on a < x < b (k = 0, 1, 2, ).
27

Monotonicity theorem: Comparing two SL problems, SL and SL, with


k > k if
the same boundary conditions, the eigenvalues will satisfy
p(x) > p(x)

or

q(x) > q(x)

or r(x) < r(x)

or

(
a, b) (a, b)

Fredholm Alternative

We have now seen two different ways to solve linear BVPs. So far, we have
been happily going along, assuming that the solutions we construct are the,
one and only, solutions. But are we sure? Now it is time to address the
important question of existence and uniqueness.
Existence: Does a solution actually exist? Will the method we employ
work, or will it for instance lead to contradictions?
Uniqueness: Weve found a solution, it exists, great. Is it the only one?
From a mathematical point of view, these are incredibly important questions. They can be even more important from a physics/applied mathematics standpoint. If youre modelling a physical situation with a particular
differential equation, youd like to think that you can find a single answer
that has physical meaning and that would agree with experiment.
To think about: What would non-existence mean in a physical context?
What about non-uniqueness?
These are important questions, and whole branches of mathematics have
been developed to think about them. Here, well just look at one very useful
theorem in this arena: the Fredholm Alternative Theorem (FAT).

6.1

A closer look at SL solution

Lets look more closely at the solution we obtained in Section 5.2 for a
SL-BVP. The last step in obtaining the coefficients ck of
c k yk

y=
k

was
k ck yk , ryk = f, yk .

28

(42)

But what happens if there is an eigenvalue equal to zero, say 0 = 0? We


potentially have a problem, as this equation would read
0 c0 = f, y0 .
If so, there are two possibilities:
1. f, y0 = 0. In this case, we have a contradiction, and we are forced to
conclude that no solution exists.
2. f, y0 does actually equal zero. In this case, no contradiction, but we
dont get any way to solve for c0 either. Hence

ck yk (x)

y = c0 y0 (x) +
k=1

where ck are calculated values for k = 1, 2, . . ., but c0 is arbitrary,


leading to an infinite set of solutions.
So, if there is a zero eigenvalue (meaning 0 = 0 yields a non-trivial solution), it seems that either we get non-existence, or we get existence but
non-uniqueness.
And if there isnt a zero eigenvalue? If there is no eigenfunction corresponding to 0 = 0, then (42) would never yield any problems, we could always
divide by k to get well defined ck , and the expansion would have no issues.
Thus, we would have both existence and uniqueness.

6.2

Zero eigenvalue

The question of the zero eigenvalue is a special case, as we are really asking
whether the homogeneous system
Ly = 0
has a nontrivial solution y0 (x).

6.3

FAT

What weve just seen above for a SL-BVP in fact holds much more widely.
For general (not necessarily self-adjoint) ODE-BVPs, the statement of FAT
reads:

29

Exactly one of the two alternatives holds:


I. EITHER the homogeneous adjoint problem
L w0 = 0, BC1 = 0, BC2 = 0
has a non-trivial solution
II. OR the inhomogeneous problem
Ly = f, BC1 = 1 , BC2 = 2
has a unique solution for any f, 1 , 2
Notes:
1. Note the exclusive or. Exactly one of the alternatives is true.
2. Youll see Ive put in inhomogeneous boundary conditions. If we are
in Case II, with no homogeneous solution, then we can construct a
solution with inhomogeneous BC either by decomposing or by putting
BC directly into the eigenfunction construction.
In Case I, there are two subcases. Suppose we have inhomogeneous BCs,
and we solve the problem by putting the BCs directly into the eigenfunction
expansion. Since there is an eigenvalue 0 = 0, we will have an equation
that looks like:
0 c0 = f, w0 + stuff0
where stuff0 has come from the inhomogeneous boundary conditions. The
two possibilities are:
1. If the RHS is non-zero, then we have a contradiction, and thus no
solution exists.
2. If the RHS=0, there is no contradiction, but we have no information
on c0 either, thus the solution is valid for any constant c0 , and we have
existence but non-uniqueness.

30

6.4

Homogeneous vs inhomogeneous BC

With homogeneous boundary conditions, FAT has a very nice form, easily
stated in words. It says that to have a unique solution, the adjoint homogeneous problem must have only the trivial solution; otherwise, if there is
a non-trivial solution w0 , then the solvability condition to have a solution
(but non-unique!) is that the forcing function f (x) must be orthogonal to
the homogeneous solution w0 .
With inhomogeneous BC, the criteria for a unique solution is the same,
but when the non-trivial w0 exists, the solvability condition is not so cute,
due to the stuff0 arising from the BCs. Note that if we had tried to decompose the solution, i.e. separate the boundary conditions into a problem
Ly = 0, BC = 0, then in looking at Ly = f, BC = 0, we would arrive at the
wrong solvability condition, since there would be no mention of boundary
conditions! In other words, decomposing the solution could lead to incorrect
conclusions! Which leads to the following...
Health warning: If you have a problem with inhomogeneous boundary
conditions AND there is a zero eigenvalue, do not try to decompose the solution. Incorporate the boundary conditions directly into the eigenfunction
expansion!

6.5

Examples

Ex. 1
Solve y + y = f , y(0) = 0, y() = 0. This is self-adjoint and has the zero
eigensolution y0 = sin x. Then
y , sin x

y sin xdx

y sin xdx

(by parts, twice)

= y, sin x
There is a solution only if
f, sin x = 0,
in which case y + c sin x is a solution for all c.
Ex. 2a
Solve y = f (x) with 0 < x < 1, y(0) = 0 and y (1) = 7. The zero eigenvalue
31

adjoint problem is L w0 = w0 (x) = 0 (this problem is also self-adjoint), with


w0 (0) = w0 (1) = 0. This only has the trivial solution w0 0, so the full
problem has a unique solution for any f (x).
Ex. 2b
Same problem, but change the BC to y (0) = 0 and y (1) = , and let
f (x) = 3. This time we get w0 (x) = 1. Thus:
y , w0
1

f, w0

y dx = 3

y |10 = 3
The BCs give that y (1) y (0) = , and thus if = 3, we have a contradiction and no solution exists, while if = 3, we have a non-unique
solution.

6.6

FAT - Linear algebra version

The Fredholm alternative can also be expressed as a theorem of linear algebra. It addresses the question: when does Ax = b have a unique solution?,
where
A Rn,n : n n matrix
b Rn : n dim, column vector

given

x Rn : n dim column vector (unknown)


FAT says:
Exactly one of the two alternatives holds:
I. Either: AT y = 0 has a non-trivial solution y = 0.
II. Or: Ax = b has a unique solution (has a solution, and it is unique).
If AT y = 0 has solutions y = 0, then Ax = b has either no or multiple
solutions.
Distinguish cases by solvability condition.
Let y1 , y2 , . . . , yN be a basis of AT y = 0.
(a) If ykT b = 0 for all k, then Ax = b is solvable, and the solution space
has dimension N .
(b) If ykT b = 0 for one or more k, then Ax = b has no solutions.
32

Example
(1)
1 2
3 4

A=
Look at AT y = 0

1
2

b=

(GE=Gaussian Elimination)

1 3 0
2 4 0

1 3 0
0 2 0

GE

y1 + 3y3 = 0
2y2
=0

y = 0 i.e. AT y = 0 only has a trivial solution. Therefore, FAT I is


false, thus FAT II is true i.e. Ax = b has a unique solution (for any b).
(2)
A=
AT y = 0 :

1 2
3 6

1 3 0
2 6 0

GE

b1
b2

b=
1 3 0
0 0 0

y1 + 3y2 = 0

Thus, a non-trivial solution for AT y = 0 is: y0 =

3
.
1

Therefore I is true, thus II of FAT is false:


Ax = b has either none or multiple solutions. Which of these two
cases applies? That depends on b. Since y0 forms a basis for the one
dimensional null space AT y = 0, the solvability condition is y0T b = 0.
Alternatively, we could see this by GE for Ax = b.
1 2 b1
3 6 b2

GE

1 2
b1
0 0 b2 3b1

x1 + 2x2 = b1
0
= b2 3b1

So, no solution if b2 3b1 = 0, and a one-dimensional space of solutions


if b2 3b1 = 0.
Note:
b
b2 3b1 = (3, 1) 1 = y T b
b2

33

Integral equations

An integral equation (IE) is an equation in which an unknown function


occurs inside an integral. For example,
b

k(x, t)(t) dt = f (x),

(43)

where (x) is the unknown function we wish to determine, while f (x) and
the so-called kernel k(x, t) are given functions.
Any ODE problem plus boundary (or initial) conditions can be rewritten as
an IE. For example, in the initial value problem (IVP)
dy
= f (x, y(x)), y(a) = c,
dx
integrating both sides from a to x yields the integral equation
x

y(x) = c +

f (x, y(t)) dt
a

How about a second order problem? For instance, consider


y + x2 y = 0, y(a) = , y (a) = .
Integrating from a to x once gives
y (x) +

t2 y(t) dt = 0.

We can then integrate again, and eliminate the double integral by noticing
that

(xt)f (t) dt

ds) dt =

f (t)(

f (t) ds dt =

f (t) dt ds =

a simple trick of switching the order of integration in the triangular region.


Thus, our ODE becomes the IE
x

y(x) = + +
a

t2 y(t)(x t) dt.

Think about: what if there were also a 1st order derivative in the ODE? How
do we avoid having y in the integral?
34

7.1

Fredholm Integral Equation

The IEs above came from initial value problems (IVP). IVPs always lead to
integral equations called Volterra Integral Equations, characterised by the
variable x as a limit in the integral.
If instead, we have a BVP on a x b with boundary conditions at a and
b, the tricks above wont work. But ODE BVPs can also be converted to
IEs, using Greens functions. Consider the equation
Lu = r(x)u, a x b,

(44)

with homogeneous boundary conditions at a and b, where L is a linear


differential operator. If we were to find the Greens function for the operator,
i.e. g(x, ) such that
Lg = (x )
then we could construct the solution to (44) as
b

g(x, )r()u() d.

u(x) =

(45)

Integral equations such as (45), that come from BVPs, are called Fredholm
Integral Equations. To summarize:
1. IVPs correspond to Volterra equations
2. BVPs correspond to Fredholm Integral Equations (FIE)
3. Boundary (or initial) conditions are implicitly included in the IE
4. The ODE and IE are equivalent, solving one is equivalent to solving
the other
In this course, we will primarily focus on FIEs, of which there are two types:
1st kind FIE:

Ly =

k(x, t)y(t) dt = f (x)


a

2nd kind FIE:


b

k(x, t)y(t) dt = f (x)

Ly = ry(x) +
a

35

where the kernel function k(x, t), f (x), the constant r are all known, and
y(x) is to be determined.

7.2

Degenerate (separable) kernels

The difficulty of solving the integral equation depends on the complexity of


the kernel. The simplest problems are those for which the kernel is made
up of a finite sum of terms, each of which is a product of functions of x and
functions of t, that is5
n

k(x, t) =

j (x)j (t).
j=1

In this case the kernel is called degenerate, or separable. We will generally


assume that the functions in each set {j } and {j } are linearly independent
(otherwise, the kernel can be rewritten in this form with a smaller n).
Example:
k(x, t) = 4xt 5x2 t2

(n = 2, 1 = x1 , 2 = x2 , 1 = 4t, 2 = 5t2 )

How do we find eigenvalues/eigenfunctions of Ly = y, in the case of a


separable kernel?6
For the example given, we must solve
1

k(x, t)y(t)dt = y(x)

(46)

0
1

4xtydt +
0

5x2 t2 ydt = y

The great thing about degenerate kernels is that the functions of x can be
pulled out of the integral. Whats left integrates to a constant, and thus we
are left with a perfect visual of the requisite form of y(x). In steps:
5

You might think of k(x, t) as a finite series of separation of variables products.


Youll notice we are switching to writing Ly = y now, instead of Ly = y. Oddly
enough, while BVPs often use the , integral equations almost never do. Lesson
learned: the world is an inconsistent place.
6

36

Factor out j (x)

ydt integrals
1

4tydt x +
0

constant1

5t2 ydt x2 = y(x)

constant2

Determines form: y(x) = C1 x + C2 x2 . For = 0 all eigenfunctions


will be linear combinations of j (x) functions.
Plug in y(t), evaluate integrals
5
4
C1 + C2 x + C1 C2 x2 = C1 x + C2 x2
3
4
Equate coeffs
x:
x2 :

4
3

54

1
1

C1
C2

C1
C2

matrix eigenvalue problem


1 = 21 y1 (x) = 6x + 5x2
2 = 16 y2 (x) = 2x 3x2
Solving the algebraic system gives a finite set of k s. Usually, we find one
yk for each k multiplicity 1.
But what about zero eigenvalues? Is there a function y for which Ly = 0?
Actually, there are infinite! To see this, note that Ly = 0 if
1

j (t)y(t)dt = 0

for j = 1, 2

that is if each separate integral evaluates to zero. Since the s are polynomials of degree 1 and 2, we can build lots () of functions orthogonal
to both i by taking polynomials of the same form plus one non-matching
degree7 . That is, we look for y0,k (x) of the form
y0,k = c1 x + c2 x2 + xk
7

k = 0, 3, 4, . . .

think: what might you do if the s were trig functions?

37

(47)

Now, plug in and choose c1 , c2 so that y0,k is orthogonal to 1 , 2


10 2
x +1
3
2
4
y0,2 (x) = x x2 + x3
5
5
y0,3 (x) = . . .

y0,1 (x) = 4x +

We would say that = 0 is an eigenvalue of multiplicity .


Next we want to solve Ly = f . FAT will be a useful tool, but first we need
to define the adjoint.
7.2.1

Adjoint for Integral operators

We use the same definition as for differential operators: w, Ly = L w, y ,


with inner product
b

u(x)v(x)dx

u, v =
a

Let

k(x, t)y(t) dt.

Ly =

(48)

Then

k(x, t)y(t)dtdx

w(x)

w, Ly =

a
b

k(x, t)w(x)y(t)dtdx

double integral

k(x, t)w(x)dxdt

y(t)

=
a

switch order

= y, L w
L w =

k(x, t)w(x)dx
a

Note that we integrate with respect to x! Thus, the kernel for the adjoint
is k(t, x) (as opposed to k(x, t))!
We thus have
L w = Lw k(x, t) = k(t, x).
Real symmetric kernel self adjoint, an instant visual check!
38

7.3

Inhomogeneous equations

Lets consider Ly = f (x) for a 1st kind FIE with a degenerate kernel, i.e.
n

j (t)y(t) dt

j (x)

Ly =

= f (x)

(49)

j=1

Lets begin with the question of existence and uniqueness. Using similar
ideas as above (Eq (47)), we can always construct non-trivial solutions to
Ly = 0 (and to L y = 0). Thus, we are not going to have uniqueness. What
about existence?
There are two ways we might go about this:
Visual route.
Since each j y dt will integrate to a constant, the equation is of the form
c1 1 (x) + c2 2 (x) + + cn n (x) = f (x)
It is clear that this can only be satisfied if f (x) is in the span of the j s, i.e.
the solvability condition is f span{j }
That seemed too easy. Lets make sure it agrees with all this FAT stuff we
know...
FAT route.
1. As weve seen, the adjoint operator has the same kernel, but with the
variables switched, so that
n

L w =

(t)

(x)y(x) dx

(50)

j=1

2. Thus, L w = 0 has solutions w0,k that are perpendicular to each of


the j
3. FAT says the solvability condition is that f must be orthogonal to all
the w0,k
4. Since the w0,k are orthogonal to all of the j s, for f to be orthogonal
to all the w0,k we would need f to be in the span of the j ; that is we
again conclude
the solvability condition is f span{j }
39

To summarize, weve seen that a 1st kind FIE will never have a unique
solution, and existence breaks down into whether or not we can write f as
n

f=

di i (x)
i=1

for some constants di . If not, no solution is possible.


If so, if f span{j }, then solutions exist, but they are not unique, because
we can add arbitrary solutions of the homogeneous problem Ly = 0.
How to find solutions?
Writing f = ni=1 di i (x) (where the di are known) and noting that the
b
coefficient of j (x) in Ly is a j y dt, we see that
b

j (t)y(t) dt = dj

(51)

must hold for all j = 1, 2, . . . n


Next, we assume that y span{j }. How? Well, any part of y that were
not in the span would be orthogonal to the j and would thus vanish from
(51). Hence, we make the ansatz
n

si i (x)

y(x) =

(52)

i=1

and our task is completed if we can find the si . To do this, plug and chug.
Stick y from (52) into (51), and we get an n n linear system for the si :
n

j (t)i (t)dt = dj

si
i=1

j = 1, . . . , n

Moreover, one can show that the coefficient matrix (


if {1 , . . . , n } are linearly independent functions.

b
a j i )j,i

Example
1

Ly =
0

k(x, t)y(t)dt = 8x 7x2


40

is invertible

with
k(x, t) = 4xt 5x2 t2

(1 = x, 2 = x2 , 1 = 4t, 2 = 5t2 )

Ansatz: y(x) = s1 x + s2 x2 (can absorb the constants 4, -5 into s1 , s2 )


1

so
0

4
1 (t)y(t) dt = d1 s1 + s2 = 8
3

and
0

2 (t)y(t) dt = d2 s1

5
4

s2 = 7

s1 = 12, s2 = 8

y(x) = 12x 8x2

7.4

2nd kind FIE

Fredholm Integral Equations of the 2nd kind:


b

k(x, t)y(t) dt = f (x)

Ly = ry(x) +
a

A 2nd kind FIE only differs from a 1st kind FIE by the extra term ry(x)
b

L1 y =

k(x, t)y(t)dt

1st kind

L2 y = L1 y + ry

2nd kind

Thus, if 1 is an eigenvalue of L1 with eigenfunction y, so that L1 y = 1 y


then
L2 y = L1 y + ry = (1 + r)y
which implies that 2 = 1 + r is an eigenvalue of L2 . In other words,
The 2nd kind FIE has the same eigenfunctions, with eigenvalues
shifted by r
Example
k(x, t) = 4xt 5x2 t2 ,
b

Ly

r=

1
k(x, t)y(t)dt + y
2
41

1
2

We already found the eigenvalues to the equivalent 1st kind FIE to be 1/2
and 1/6, plus the zero eigenvalue with infinite multiplicity. Thus the 2nd
kind with r = 1/2 has eigenvalues 1, 13 (mult. 1 each), 21 (mult. !)
There is no zero eigenvalue, so FAT tells us that Ly = f has a unique
solution for all f .

7.5

Imhomogeneous 2nd kind FIE with degenerate kernel

How do we solve Ly = f for a 2nd kind FIE for separable k =


j j ?
Here, we can get an ansatz immediately by playing the visual game:
b

j (t)y(t) dt + ry(x) = f (x)

j (x)

(53)

y(x) has the form


1
1
y(x) = f (x)
r
r

sj j (x).

(54)

j=1

Notice that the ansatz for is different from the ansatz for a 1st kind FIE.
To obtain a linear system of equations for the sj , plug (54) into (53), evaluate
the integrals and compare the coefficients of the j .
b

k (t)f (t)dt

k (t)j (t)dt + rsk =

sj

for k = 1, . . . , n. This is a n n system for vector s (s1 , . . . , sn ), which


we can write in matrix form
(A + rI)s = b,
b

A = (akj ),

k (t)j (t)dt,

with akj =
a

b = (b1 , . . . , bn ),

with bk =

k (t)f (t)dt.
a

Example
K(x, t) = 4xt 5x2 t2
42

j, k = 1, . . . , n,

1
0

1
Kydt + y = 8x 7x2 ;
2

r=

f (x)

1
2

Y = 2 (8x 7x ) 2 (s1 x + s2 x2 )
1
r

1
r

s j j

1 = x, 2 = x2 , 1 = 4t, 2 = 5t2
4
11
1
1
1
3
6
A + I = 5
A = 5
1
1
2
4
4
2
1

b1 =
0
1

b2 =
0

4t(8t 7t2 )dt =

11
3

5t(8t 7t2 )dt = 3

s1 =

7
2

s2 =

y = 16x 14x2 7x +

11
4

11 2
17
x = 9x x2
2
2

And this y(x) is unique!


Summary:
I 1st kind F IE
The homogeneous equation Ly = 0 has a large solution space
Thus Ly = f has a (very) non-unique solution for very restricted
f (x), otherwise non-solvable.
II 2nd kind F IE:
The eigenvalues are just shifted from the eigenvalues of the equivalent 1st kind FIE.
If the homogeneous equation Ly = 0 has only trivial solutions,
then there is a unique solution for all f .
If Ly = 0 has non-trivial solutions, it is still a finite-dimensional
solution space, say of dimension N . Then there is no unique
solution to Ly = f , i.e. depending on f , there is either no solution
or an N -dimensional subspace of solutions.
The N solvability conditions are obtained by forming N inner
products of f with a basis of the null-space of the adjoint operator.
43

7.6

Non-degenerate/nonseparable kernels

We conclude this section with a brief look at some non-denerate kernels.


The following kernel has separable terms, but infinite of them:

k(x, t) =

j (x)j (t)
j=n

One way to obtain such an infinite series is by expressing the kernel as a


Fourier series.

1
Ex. k(x, t) = ext , ln x + t,
x+t
In this case,
We would seek an infinite series solution for y(x)
No
More systematic.
Hilbert-Schmidt kernels are a common class of nonseparable kernels. IE
will have good properties for any k(x, t) satisfying the Hilbert-Schmidt (HS)
condition
b

k 2 (x, t)dxdt <

An integral operator

k(x, t)y(t) dt
a

in which the kernel satisfies the HS condition is called a Hilbert-Schmidt


integral operator. HS theory for nondegenerate self-adjoint FIE is similar
to Sturm-Liouville theory for self-adjoint ODE-BVP.
Properties of HS-FIE
(1) All eigenvalues are real and non-zero;
(2) Eigenvalues form a discrete infinite set {j }, j = 1, 2, . . .
|1 | |2 | . . .
limj j = 0; each j = 0 has finite multiplicity.

44

(3) Eigenfunctions for j = k are orthogonal


b

yj (x)yk (x)dx = 0
a

(4) Eigenfunctions are a complete set, i.e. any h can be expanded as


h, yk
yk (x)
yk , yk

h(x) =

(5) Solutions to Ly = f can be understood via FAT: existence, uniqueness/nonuniqueness.

Example:
Let g(x, t) be the function
(1 t)x
(1 x)t

g(x, t) =

xt
x>t

This is a HS-kernel, thus


1

g(x, t)y(t)dt.

Ly

is a HS-operator. We can infer its eigenvalues/functions by noticing that g


is the Greens function of the SL operator (refer back to Equation (23)).
y (x),
Ly

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

= y,
The eigenvalue problem Ly
y + y = 0 on 0 < x < 1,

y(0) = 0 = y(1)

is easily solved, and the eigenvalues/eigenfunctions are


k = k 2 2 ,

yk = sin kx

k = 1, 2, . . .

= y in terms of the Greens


Alternatively, we can write solutions of Ly
function (with y playing the role of f (x)):
1

y =

g(x, t)y(t)dt
0

45

and of course we get the same eigenvalues and eigenfunctions.


Thus, considering the eigenvalue problem for the HS-Operator L:
1

g(x, t)y(t)dt = y,
0

we can infer the eigenvalues/functions:


k =

1
1
= 2 2,
k
k

yk = sin kx

k = 1, 2, . . . .

Moreover, the operator L is self-adjoint, so the yk are also the adjoint eigenfunctions. Thus we have all we need to set up solvability conditions for the
second kind IEs
Ly + ry = f, with r = 0,
for the L defined in this example.

Singular points of differential equations

In this section we will seek solutions of the nth order linear differential
equation
Ly = y (n) (x) + pn1 (x)y (n1) (x) + + p1 (x)y (n) (x) + p0 (x)y(x) = 0, (55)
in the form a series expansion in the neighbourhood of x = x0 . How we
proceed, and the nature of the solution, depends on how well-behaved the
functions pj (x) are around x0 .

8.1

Ordinary points

The point x0 is an ordinary point of the ODE if all pj (x) are analytic there,
i.e. they can be expanded as a convergent power series. The procedure in
k
this case is pretty straightforward: (i) write y(x) =
0 ak x as a power
series, (ii) plug into (55), using the power series expansions of each of the
pj , then (iii) obtain a sequence of equations for the coefficients ak that can
be solved recursively.
This is the simplest (and least interesting) case, so we wont really spend
any time on it, but it is worth noting a few things about ordinary points:
All n linearly independent solutions of (55) are analytic at x0 .

46

The radius of convergence of the series solution distance (in C) to


next singular point.
Example:
(x2 + 1)y + 2xy = 0
Here x0 = 0 is an ordinary point. Nearest singular points are x = i,
distance 1 from 0 radius of convergence 1.
1
Here we can obtain the solution (1+x
2 ) via easier routes, but we note that
1
= 1 x2 + x4 . . .
(1 + x2 )
with a radius of convergence = 1, is the series solution one would obtain.

8.2

Singular points

The point x0 is called a singular point of the ODE if one of the pj (x) is
not analytic there.
In this case, the general solution y may have a singularity at x = x0
(but not necessarily). This means that the general solution y may not be
analytic at x0 : y or its derivatives can blow-up as x x0 .
Where do ODEs with singular points arise?
As we will see later, equations with singular points at x = 0 commonly
arise from linear PDEs in polar/spherical co-ordinates (where x = 0
corresponds to radius r = 0).
SL-BVPs with a singular point at the boundary - such problems often
require the solution to be bounded at the boundary point (rather than
prescribing a specific value).
Note: x0 = can also be classified as an ordinary or singular point by
changing the independent variable via the substitution t = 1/x, u(t) = y(x),
(i.e. dy/dx = t2 du/dt, etc.) and classifying the point t = 0 for the resulting ODE for u(t).
Some Examples
(a) y = ex y: every x0 is an ordinary point

47

(b) x5 y = y: x0 = 0 is a singular point, every x0 = 0 is an ordinary


point.
Lets consider more closely a simple example to see the types of solutions
near singular points. Consider the equation:
y + xm y = 0, m 0 an integer
The general solution can be found via separation of variables, and depends
on the value of m:
1. For m = 0, the point x = 0 is ordinary and y(x) = C exp(x). This
function can be expanded into a power series at x = 0 which converges
for all x C.
2. For m = 1, y(x) = C/x. This solution clearly has a singularity at
x = 0, but a rather benign one (a simple pole).
3. For m = 2 (and similarly for larger m), y(x) = C exp(1/x); this solution has a very strong singularity: x = 0 is an essential singularity of
exp(1/x) in the complex plane.
This example suggests that the solution at a singular point of an ODE tends
to have a stronger singularity the higher the order of the poles in the coefficients in front of the lower order terms of the ODE. In fact, this is the key
idea behind the classification of singular points:
Regular singular points: x0 is a regular singular point, if all pj (x)
pj (x)(x x0 )nj are analytic at x = 0 (for j = 0, . . . , n 1).
Irregular singular points: all other singular points.
For regular singular points, you might think of it as the singularities are not
too bad, we are essentially able to remove the trouble in pj by multiplying
by a power of (x x0 ) decreasing with the degree of derivative. In this case,
a modification of the power series approach can be used.
For irregular singular points, though, there is no general theory!
Cauchy-Euler. One of the simplest and most instructive examples is the
Cauchy-Euler equation
x2 y + axy + by = 0,
48

which clearly has a regular singular point at x = 0. As youve seen before,


the general solution can be found via the ansatz y = xm . The characteristic
equation for m is m(m 1) + am + b = 0. There are two cases:
(i) The characteristic equation has two distinct roots m1 and m2 . Then, the
general solution is:
y(x) = C1 xm1 + C2 xm2 .
(ii) The characteristic equation has a double root m. Then, the general
solution is:
y(x) = C1 xm + C2 xm ln(x).
Note that if the roots are two distinct non-negative integers, then the general
solution in (i) is analytic (even though the ODE has a singular point). In
general, however, the behaviour as x 0 is a fractional or even complex
power of x.
This behaviour carries over to the general situation for regular singular
points, except that the functions xm are multiplied by an analytic function (i.e. a power series in x). Next, well look at the general theory for
regular singular points.

8.3

Frobenius method for 2nd order ODEs

From now on, well restrict to 2nd order equations. If the ODE
Ly y + P (x)y + Q(x)y = 0

(56)

has a regular singular point at x = x0 , then


p(x) := P (x)(x x0 )
and
q(x) := Q(x)(x x0 )2
are analytic, i.e.

p(x) =
n=0

q(x) =
n=0

pn (x x0 )n

(57)

qn (x x0 )n

(58)

49

The idea is to seek a solution in the form of a Frobenius series

y(x) = (x x0 )

n=0

an (x x0 )n

(59)

(In terms of the Cauchy-Euler example, is playing the role of m and


an (xx0 )n is the analytic function with coefficients an to be determined.)
We may assume that a0 = 1 (by choosing appropriately and normalizing).
Now plug in and equate coefficients. At the lowest power ((x x0 )2 , we
find
( 1) + p0 + q0 = 0
This polynomial plays an important role, lets give it a name:
F () = ( 1) + p0 + q0 .
The equation F () = 0 is called the indicial equation, it determines the
possible indicial exponents 1 , 2 . Note that these exponents can be complex! Well order them such that (1 ) (2 ).
Lets carry on with equating coefficients of powers of (x x0 ). We find after
some algebra that the coefficients of (x x0 )n+2 satisfy
n1

F ( + n)an =

[( + k)pnk + qnk ]ak

(60)

k=0

Setting = 1 , we know that F (1 + n) = 0 for any integer n 1 (do you


see why?), thus we can use (60) to solve for all the coefficients an , and we
obtain one solution

y1 (x) = (x x0 )1

n=0

an (x x0 )n .

(61)

A(x)

Thus the first solution can always be expressed as a Frobenius series with
indicial exponent = 1 .
For the second solution, we have to distinguish between several cases and
sub-cases8 .
8
and sub-sub-cases. Its a bit of a headache, you might want to have some ibuprofen
handy.

50

Case I: 1 2 is not an integer (in particular = 0).


In this case, F (2 + n) = 0 for all n 1. Thus we can solve (60) for all
coefficients lets call them bn to distinguish from previous coefficients.
Thus, we obtain with no problems a second solution also as a Frobenius
series, with indicial exponent 2 ,

y2 (x) = (x x0 )2

n=0

bn (x x0 )n
B(x)

Case IIa: 1 = 2
In the case of a double root we clearly only get one solution with the Frobenius method, and we have to multiply by logs to get a second solution
(similar to the case of a double root in Cauchy-Euler). In particular, the
second solution is of the form

y2 (x) = y1 (x) ln(x x0 ) + (x x0 )

1
n=0

cn (x x0 )n ,

where y1 is the first solution. We can determine the cn in the usual manner.
A derivation of this form can be done using the so-called derivative method,
which I include in Section 8.4 for completeness9 .
Case IIb: 1 2 = N , where N > 0 is an integer. In this case, we will
potentially run into trouble in (60) at n = N . There are two possibilities:
(i) For n = N , RHS = 0 in (60). Then we have a contradiction, and the
solution method doesnt work. To get a second solution, we use the same
form as Case IIa:

y2 (x) = y1 (x) ln(x x0 ) + (x x0 )2

n=0

cn (x x0 )n

and determine the cn by substituting into the ODE. Note that the indicial
exponent for the second term is 2 (whereas y1 is given by the Frobenius
series using the exponent 1 ).
9

This is code for I will not test you on this, but your life just may be sweetened by
tasting of such fruits of knowledge.

51

(ii) For n = N , RHS = 0 in (60).


There is no contradiction, but any choice for an (or bn ) will satisfy (60)
2nd solution has Frobenius form

y2 (x) = (x x0 )2

n=0

bn (x x0 )n .

(62)

B(x)

where b0 can be chosen to be b0 = 1 and bN is also arbitrary. Notice that


changing bN changes (62) by multiples of y1 .

8.4

Derivative method

Suppose 1 is a double root of F (). Let a0 = 1 and solve (60) for a1 , a2 . . .


with arbitrary (i.e. F () not generally = 0. Thus, the an = an (), and
we can think of as a parameter in the series

y(x; ) (x x0 ) +

n=1

an ()(x x0 )n+

Ly = L(x x0 ) = (x x0 )2 F ()

(63)

We know that y(x; 1 ) is a solution. But 1 is double root, which implies


d
F |=1 = 0
d
The idea is to differentiate (63), then set = 1 . Since L has no dependence
on ,

(Ly)|=1 = 0 = L

y2 =

|=1

|=1

is also a solution. Specifically, to get a more concise form:

52

|=1 =

n=0

n=0

an (1 )(x x0 )n+1 ln(x x0 )+

dan
|=1 (x x0 )n+1
d

(64)

= y1 (x) ln(x x0 ) +

n=0

bn (x x0 )n+1

=(xx0 )1 C(x)

Derivative method can be used to determine bn , however, a closed form


for an () for general is required for this! It is usually easier to just
use the appropriate form of the series.
Plug it into the equation, and compare coefficients (as you would do
for a power series expansion). After plugging in, the terms containing
the log terms should cancel.

Special functions

We have seen in the previous section a method for constructing solutions


to ODEs with non-constant coefficients and singular points. For any given
problem, the success of the method and the utility of the solution depends
on whether one can obtain a direct formula for the series coefficients.
In this section, we explore several special functions, which occur commonly
enough to have a name, and for which the hard work, the series solution
method, has already been done.

9.1

Bessel Functions

These particularly common functions can be motivated by considering the


vibrating membrane of a circular drum. Let U (x, y, t be the position of the
membrane at time t and position (x, y), compared to the flat horizontal rest
state.
The governing equations for the membrane are (wave equation):
Utt = c2 U
U =0

for x2 + y 2 < 1
at x2 + y 2 = 1

(Newtons second law & elastic stresses)

(Membrane pinned at boundary.)


53

Separation of variables U (x, y, t) = v(t)u(x, y) yields


u
vtt
=
= const ,
v
u
i.e.
u = u.

Switch to polar coordinates: u now depends on r, :


1
r r

u
r

1 2u
+ u = 0
r2 2
u=0

0 < r < 1,
r = 1,

u periodic in .

0 2,

0 2

This is a PDE eigenvalue problem, we need to find for which there are
non-trivial solutions u(r, ).
What about r = 0? Note that in
1
r r

u
r

2 u 1 u
+
r2
r r

the first derivative of u has a singular coefficient. As for ODEs with singular
points, we may therefore expect that this will give rise to singularities in the
solutions of the PDE and thus we impose the condition
u bounded as r 0,
since we expect the solution of the governing equations for the drum membrane to be bounded.
Since u is periodic in we can expand u into a Fouries series in :

Un (r) cos n + Vn (r) sin n;

u(r, ) = U0 (r) +
n=1

substituting this into the previous set of equations gives ( = d/dr)


1
n2

rUn + 2
r
r

Un = 0,

for 0 r < 1,

(65a)

Un = 0 at r = 1,

(65b)

Un bounded as r 0.

(65c)

The same
equations hold for the Vn . Now eliminate by rescaling: Un (r) =
y(x), x = r,
x2 y + xy + (x2 n2 )y = 0,

and we arrive at Bessels equation (for integer n 0.).


54

9.2

Bessel functions of first and second kind

Bessels equation (BE) has a regular singular point at x = 0, with indicial


equation ( 1) + n2 = 0, the solutions of which are 1 = n, 2 = n
(double root for n = 0).
The general solution is given as a linear combination of two linearly independent solutions. A detailed discussion along the lines of Section 8 reveals
that one is locally given at x = 0 by a Frobenius series with the exponent
1 = n and the other by a Frobenius series with exponent 2 = n plus
ln(x) times the first solution (case II(b)(i) in our general discussion of ODEs
with singular points).
The first Frobenius series with a specific normalization of the leading coefficient of the expansion defines the Bessel functions of first kind
Jn (x) =

x
2

k=0

(1)k
x
k!(k + n)! 2

2k

for integer n 0.
Similarly, a specifically normalized choice for the second expansion defines
the Bessel functions of second kind
Yn (x) =

(x/2)n
2
ln(x/2)Jn (x)

(x/2)n

k=0

n1
k=0

(n k 1)! 2
x /4
k!

x2 /4
[(k + 1) + (n + k + 1)]
,
k!(n + k)!

m1 1
k , (m 1) and = 0.5772 . . . is the Eulerwhere (m) = + k=1
Mascheroni constant. (Details regarding these expansions will be studied in
the problem sets on problem sheet 7.)

Some Bessel function fun facts:


Since Bessels equation has only one singular point for finite x, the
series for Jn and in Yn have an infinite radius of convergence.
Also, Jn and Yn are oscillating functions that decay as x . They
have an infinitude of discrete zeros for x 0, which are quite important
and have therefore been tabulated (for example in Abramowitz and
Stegun).
55

At x = 0, the behaviour of the two kinds of Bessel functions is quite


different. For Jn , we have Jn (0) = 0 if n > 0, and J0 (0) = 1, while
Yn as x 0.
Two recursion relations, which can be derived from the local expansions:
2n
Jn (x) Jn1 (x),
Jn+1 (x) =
x
Jn+1 (x) = 2Jn (x) + Jn1 (x).
The same relations also hold for the Yn s. In principle, these relations
can be used to compute the Bessel functions, however, the straightforward way calculating the values for larger n from those for smaller n
is usually numerically unstable and is therefore not recommended.
Many more relations as well as theory exist. A vast collection of results for
the Bessel functions can be found in particular in Abramowitz and Stegun;
some derivations in Riley et al. But for now, back to the vibrating drum. We
can now express the general solution for (65a) in terms of Bessel functions
as

Un (r) = aJn
r + bYn
r .
The boundedness condition (65c) requires b = 0. A non-trivial solution
therefore requires that we set a = 0, without loss of generality: a = 1.
Thus, the boundary condition (65b) at r = 1 leads to

Jn
= 0,

i.e. has to be one of the zeros of Jn . If we label the zeros of Jn by m


m = 1, 2, . . ., (sorted, for example, in ascending order), then the eigenvalues
for (65) are
2
= m
m = 1, 2, . . .
with corresponding eigenfunctions Jn (m r), m = 1, 2, . . .. (Note that n is
kept fixed, and that the zeros depend on n!)
The differential equation (65a) can be written in Sturm-Liouville form by
multiplying through with r; i.e. (65) is a singular SL problem with weighting function r, and we therefore have the following orthogonality relations
between the eigenfunctions
1

rJn (l r) Jn (m r) dr = 0
0

56

for l = m;

a separate calculation is required for l = m:


1
0

9.3

rJn2 (m r) dr =

1
J ()
2 n

Legendre equation and Legendre functions

Legendre equations/functions arise from studying eigenvalue problems for


the 3D Laplace operator in spherical coordinates.
The associated Legendre equation is given by
(1 x2 )y 2xy + l(l + 1)

m2
1 x2

y = 0,

or in self-adjoint form

(1 x2 )y + l(l + 1)

m2
1 x2

y = 0.

The numbers m and l can in general be complex; here, we will focus on the
case where m and l are non-negative integers. The solutions of the associate
Legendre equation are the associated Legendre functions and are denoted by
Plm ; for m = 0, we drop the associated and speak of the Legendre equation
and functions, usually denoted by Pl .
9.3.1

Properties

1. The points x = 1 and x = are regular singular points of the


associated Legendre equation. The indicial exponents for x = 1 are
m/2 and m/2. Thus, the local expansion yields one bounded and
one unbounded solution at x = 1. The same is true for x = 1.
2. If we replace l(l +1) in the self-adjoint form of the associated Legendre
equation by ,

(1 x2 )y +

m2
1 x2

y + y = 0,

(66)

and consider bounded solutions on 1 < x < 1, we see that boundedness imposes two conditions, one at each end of the interval. This
suggests that (66) is a singular Sturm-Liouville problem (with coefficient functions p = 1 x2 , q = m2 /(1 x2 ), r = 1) with discrete
eigenvalues. Indeed, the eigenvalues are exactly of the form = l(l+1)
57

with integer l m. The eigenfunctions are the corresponding associated Legendre functions, i.e. Plm . From Sturm-Liouville theory, we
infer the orthogonality relation
1
1

Plm (x)Pnm (x)dx = 0

for l = n.

The case l = n requires explicit calculation, see problem sheet.


3. For m = 0 the Legendre functions (without associated) are polynomials, and are given explictly by a so-called Rodrigues formula:
Pl (x) =

1 dl
(x2 1)l .
2l l! dxl

Of course, to find the general solution of the Legendre equation we


need a second, linearly independent solution, and this is given by the
Legendre function of 2nd kind, denoted by Qn . These solutions are
unbounded at x = m. For the case n = 0, the solution Q0 was
stated on problem sheet 4:
Q0 (x) =

1
ln
2

1+x
1x

4. For the general case, 0 m l, the associated Legendre functions of


first and second kind are given by
Plm (x) = (1)m (1 x2 )m/2

dm Pl
dxm

m
2 m/2
Qm
l (x) = (1) (1 x )

d m Ql
dxm

and

respectively. The associated Legendre functions Plm are polynomials,


if and only if m is even.
5. There are several recurrence relations for (associated) LF, e.g.: These
relations, and further properties can be found in Abramowitz and Stegun, and some derivations in Riley at al.

58

9.4

Generalisation: Orthogonal Polynomials

There are other second order linear ODEs with families of orthogonal polynomials as solutions. Often, the orthogonality relations
b

pm (x)pn (x)(x)dx = 0

m=n

with a fixed weighting function (x) (in general non-trivial, i.e. 1) can by
inferred by formulating appropriate Sturm-Liouville (eigen)problems. Orthogonal polynomials play an important role in approximation theory, for
the construction of numerical methods to discretized differential equations,
and in many applications, e.g. from physics.
One can in fact give a complete classification of all infinite families of orthogonal polynomials that can arise from second order linear differential
equations (we omit here some specific conditions that are needed to make
this a precise statement). The most important ones include (n is a nonnegative integer):
1. The Jacobi-like polynomials, to which the Legendre, the Chebychev,
and the Gegenbauer polynomials belong. These arise from DEs of the
type
x(1 x)y + (a + bx)y + y = 0,
with constants a and b and an appropriate discrete set of .
2. The Laguerre and associated Laguerre polynomials, which are solutions of
xy + (k + 1 x)y + y = 0,
for k = 1, 2, . . . and an appropriate discrete set of . (For the
Laguerre polynomials, i.e. without associated, we have k = 0, and
= n.)
3. Hermite polynomials, which are solutions of the Hermite equation
y 2xy + y = 0

with = 2n.

These families share many similar structural properties, e.g. they are given
explicitly by Rodrigues formulae and have similar recurrence relations, see
Abramowitz and Stegun.

59

Вам также может понравиться