1B Methods 1

1B METHODS
LECTURE NOTES
Richard Jozsa, DAMTP Cambridge
[email protected]
October 2010

PART I:
Self adjoint ODEs
1B Methods 2

PREFACE

These notes (in four parts) cover the essential content of the 1B Methods course as it
will be presented in lectures. They are intended be self-contained but they should not be
seen as a full substitute for a good textbook, which will contain further and alternative
explanations, further related material and more worked examples. I am grateful to the
previous lecturer of this course, Dr Colm-Cille Caulfield, for making his notes available
to me; they were very useful in forming a basis for the notes given here.
The term “mathematical methods” is often understood to imply a kind of pragmatic
attitude to the mathematics involved, in which our principal aim is to develop actual ex-
plicit methods for solving real problems (mostly solving ODEs and PDEs in this course),
rather than careful development of abstract mathematical theory. With this focus on
applications, we will not give proofs of some of the theorems on which our techniques
are based (being satisfied with just reasonably accurate statements). Indeed in some
cases these proofs would involve a formidable foray into subjects such as functional anal-
ysis and operator theory. This “mathematical methods” attitude is sometimes frowned
upon by pure-minded mathematicians but in its defence I would make two points: (i)
developing an ability to apply the techniques effectively, provides a really excellent basis
for later appreciating the subtleties of the pure mathematical proofs, whose considerable
abstractions and complexities if taken by themselves, can sometimes obfuscate our un-
derstanding; (ii) much of our greatest mathematics arose in just this creatively playful
way – of cavalierly applying not-fully-rigorous techniques to obtain answers, and only
later, guided by gained insights, developing an associated rigorous mathematical theory.
Examples include manipulation of infinite series, use of infinitesimals in the early devel-
opment of calculus, even the notion of a real number itself, the use of the Dirac delta
function and many more. Thus I hope you will enjoy the content of this course in the
spirit that is intended.

Richard Jozsa
October 2010
1B Methods 3

1 FOURIER SERIES

A function f on R is periodic with period T if f (t + T ) = f (t) for all t (and conven-

tionally we take the smallest such T ). Thus f is fully specified if we give its values only
on [0, T ) or any other interval of length at least T .
Basic examples: S(t) = A sin wt and C(t) = A cos wt. A is the amplitude. w is
the frequency. The period is T = 2π/w. If the variable is a spatial dimension e.g.
f (x) = A sin kx then the period 2π/k is also called the wavelength and k is called the
wave number (being the number of waves per 2π length).
In applications we often encounter periodic functions that are not continuous (especially
at 0, T, 2T, . . .) e.g. the sawtooth f (x) = x for 0 ≤ x < 1 with period 1; or the square
wave f (x) = 1 for 0 ≤ x < 1 and f (x) = 0 for 1 ≤ x < 2 with period 2.

1.1 Orthogonality of functions

Recall that for vectors, say 2-dimensional real vectors

u = ai + bj v = ci + dj

we have the notion of orthonormal basis i, j and inner product u · v which we’ll write
using a bracket notation (u, v) = ac + bd (not to be confused with an open interval!
- the meaning should always be clear from the context!). For complex vectors we use
(u, v) = a∗ c+b∗ d where the star denotes complex conjugation. u is normalised if (u, u) = 1
and u and v are orthogonal if (u, v) = 0.
Consider now the set of all (generally complex-valued) functions on an interval [a, b].
These are like vectors in the sense that we can add them (pointwise) and multiply them
by scalars. Introduce the inner product of two functions f, g : [a, b] → C as follows:
Z b
(f, g) = f ∗ (x)g(x)dx. (1)
a

Note that this even looks rather like an ‘infinite dimensional’ version of the standard inner
product formula, if we think of the function values as vector components parameterised
by x i.e. we multiply corresponding values and “sum” (integrate) them up. (f, f ) is
called the squared norm of f and two functions are orthogonal if (f, g) = 0. Note
that (f, g) = (g, f )∗ so if either is zero then the other is too.
Important example: Consider
mπx nπx
Sm (x) = sin Cn (x) = cos on [0, 2L].
L L
To calculate their inner products we can use the standard trig identities
1
cos A cos B = [cos(A − B) + cos(A + B)]
2
1B Methods 4

1
sin A sin B = [cos(A − B) − cos(A + B)]
2
1
sin A cos B = [sin(A + B) + sin(A − B)].
2

We get for example,

Z 2L
mπx nπx
(Sm , Sn ) = sin sin dx
0 L L
1 2L (m − n)πx 1 2L
Z Z
(m + n)πx
= sin dx − sin dx
2 0 L 2 0 L

0 if m 6= n.
=
L if m = n.
More concisely, using the Kronecker delta

(Sm , Sn ) = Lδmn (2)

so the rescaled set of functions √1 Sn (x) for n = 1, 2, . . . is an orthonormal set.

L

Similarly you can easily derive that for all m, n = 1, 2, . . . we have

(Cm , Cn ) = Lδmn
(3)
(Cm , Sn ) = (Sm , Cn ) = 0.
Finally consider the cases with m = 0 or n = 0: S0 is identically zero so we exclude it
whereas C0 (x) = 1 and it is easy to see

(C0 , C0 ) = 2L (C0 , Cm ) = (C0 , Sm ) = 0 for all m = 1, 2, . . .

Putting all this together we see that the infinite set of functions B = { √12 C0 , S1 , C1 , S2 , C2 , . . .}
√
is an orthogional set with each function having norm L. Indeed it may be shown (but
not proved in this methods course...) that these functions constitute a complete orthog-
onal set, or an orthogonal basis for the space of all functions on [0, 2L] (or functions on R
with period 2L) in the same sense that i, j is a complete orthogonal set for 2-dimensional
vectors – it is possible to represent any (suitably well behaved) function as a (generally
infinite) series of functions from B. Such a series is called a Fourier series.

1.2 Definition of a Fourier series

We can express any ‘suitably well-behaved’ (cf later for what this means) periodic function
f (x) with period 2L as a Fourier series:
∞ h
f (x+ ) + f (x− ) 1 X  nπx   nπx i
= a0 + an cos + bn sin , (4)
2 2 n=1
L L

where an and bn are constants known as the Fourier coefficients, and f (x+ ) and f (x− )
are the right limit approaching from above, and the left limit approaching from below
1B Methods 5

respectively. (These expressions apply even if f is a complex-valued function in which

case the coefficients an and bn are themselves complex numbers – we can just treat the
real and imaginary parts of any such complex-valued f as separate real functions).
If f is continuous at the point x, then the left hand side is of course just f (x). However, if
f has a bounded discontinuity at the point xd (i.e. the left limit approaching from below
f (xd− ), is not equal to f (xd+ ), the right limit approaching from above, but |f (xd+ ) −
f (xd− )| is bounded) then the left hand side is the average of these two limits. In our
present context this provides a convenient way to re-define the value of function at a
bounded discontinuity e.g. we would replace the step function “f (x) = 0 if x ≤ 0 and
f (x) = 1 if x > 0” by the function “g(x) = 0 if x < 0, g(0) = 21 and g(x) = 1 if x > 0”.
Often, this subtlety will be glossed over, and the left hand side will just be written as
f (x), with the behaviour at a bounded discontinuity being understood.
Determining the an and bn is easy by exploiting the orthogonality of the sines and cosines.
In terms of our previous notation of Sm and Cn we can write eq. (4) as
∞
1 X
f (x) = a0 ( C0 ) + an Cn (x) + bn Sn (x). (5)
2 n=1

Consider now Z 2L
mπx
(Sm , f ) = f (x) sin dx.
0 L
Substituting the RHS of eq. (5) and assuming it is okay to swap the order of summation
and integration, we get
∞
1 X
(Sm , f ) = a0 (Sm , C0 ) + an (Sm , Cn ) + bn (Sm , Sn ).
2 n=1

According to the orthogonality relations eqs. (2,3) all inner products on RHS are zero
except for (Sm , Sm ) which is L. Hence we get (Sm , f ) = bm L i.e.

1 2L
Z
mπx
bm = f (x) sin dx m = 1, 2, . . . (6)
L 0 L
Similarly by taking inner products of eq.(5) with Cm we get

1 2L
Z
mπx
am = f (x) cos dx m = 0, 1, 2, . . . (7)
L 0 L

The factor of 12 in the a0 term of eq. (4) conveniently makes the am formula eq. (7) valid
for m both zero and nonzero (recalling that (C0 , C0 ) = 2L but (Cm , Cm ) = L for m 6= 0).
Remarks R 2L
1
(i) The constant term a0 /2 equals the average value hf i = 2L 0
f (x) dx of f over its
period and then subsequent sine and cosine terms ‘build up’ the function by adding in
terms of higher and higher frequency. Thus the Fourier series may be thought of as the
decomposition of any signal (or function) into an infinite sum of waves with different but
discrete wavelengths, with the Fourier coefficients defining the amplitude of each of these
1B Methods 6

countably-many different waves.

(ii) The range of integration in the above formulas can be taken to be over any single
RL
period. Often it’s more convenient to use the symmetrical range −L .
(iii) Warning – if we start with a function having period T , be careful to replace L in
the above formulas by T /2 (since above, we wrote the period as 2L!)
Dirichlet conditions
So, what is meant by a ‘well-behaved’ function in the definition of a Fourier series? Here
it is defined by the Dirichlet conditions: a periodic function f (x) with period T is said
to satisfy the Dirichlet conditions if f is bounded and
R Thas a finite number of maxima,
minima and discontinuities on [0, T ) (and hence also 0 |f (x)|dx is well-defined). Then
we have the theorem (not proved in this course):
Basic theorem: If f satisfies the Dirichlet conditions then f has a unique Fourier series
as in eq. (4) with coefficients given by eqs. (7,6). This series converges to f (x) at all
points where f is continuous, and converges to the average of the left and right hand
limits at all points where f is discontinuous.
Smoothness and order of Fourier coefficients
According to the Dirichlet conditions, it is possible to establish a Fourier series represen-
tation of a certain kind of discontinuous function. More generally it can be shown that
the amount of non-smoothness is reflected by the rate at which the Fourier coefficients
decay with n, as follows.
Theorem. Suppose that the pth derivative of f is the lowest derivative that is discon-
tinuous somewhere (including the endpoints). Then the Fourier coefficients for f fall off
as O[n−(p+1) ], as n → ∞. Thus smoother functions (i.e. larger p) have coefficients falling
off faster, and hence better convergence properties of the Fourier series.
Example
Consider the sawtooth function having period 2L, given by:
f (x) = x for −L ≤ x < L, and repeating periodically outside [−L, L).
Since the function is odd, we immediately get an = 0 for all n.
Integration by parts shows (as you can check), that

2L
bm = (−1)m+1 ,
mπ     
2L  πx  1 2πx 1 3πx
f (x) = sin − sin + sin + ... . (8)
π L 2 L 3 L

This series is actually very slowly convergent – the smoothness parameter p above is zero
in this example and indeed the coefficients fall off only as O(1/n), as expected. In the
figure we plot a few of its partial sums fN (x):
N
X  nπx 
fN (x) ≡ bn sin .
n=1
L

Note that the series converges to 0 at x = ±L i.e. to the average value across these jump
discontinuities.
1B Methods 7

1.5

0.5

−0.5

−1

−1.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

Figure 1: Plots (with L = 1) of the sawtooth function f (x) = x (thin solid line) and the
partial sums f1 (x) (dots); f5 (x) (dot-dashed); f10 (x) (dashed); and f20 (x) (solid).

Also note that there is a persistent overshoot at x = L – an example of the so-called

‘Gibbs phenomenon’, as considered in more detail on the example sheet.
The integral of the sawtooth function: f (x) = x2 /2, −L ≤ x ≤ L
As an exercise, show that the Fourier series representation of this function is
" ∞
#
x2 1 X (−1) n  nπx 
= L2 +2 cos . (9)
2 6 n=1
(nπ)2 L

Note that the coefficients fall off as O(1/n2 ), consistent with the fact that f is continuous
but has discontinuous first derivative.
If we substitute x = 0 and L = 1 into the series we get
π2 1 1 1 1
=1− + − + − ...
12 4 9 16 25
Such interesting formulae are commonly constructed using Fourier series!
Finally, notice the coincidence of the term-by-term derivative of this Fourier series eq.
(9) and the series in (8). We now look into this property more carefully.
1B Methods 8

1.3 Integration and differentiation of Fourier series

Integration is always ok
Fourier series can always be integrated term-by-term. Suppose f (x) is periodic with
period 2L and satisfies the Dirichlet conditions so it has a Fourier series for −L ≤ x ≤ L:
∞
a0 X h  nπx   nπx i
f (x) = + an cos + bn sin .
2 n=1
L L

It is then always valid to integrate term by term to get (here −L ≤ x ≤ L and F is

extended to all of R by periodicity)
Z x ∞
a0 (x + L) X an L  nπx 
F (x) ≡ f (u)du = + sin
−L 2 n=1
nπ L
∞
X bn L h n
 nπx i
+ (−1) − cos ,
n=1
nπ L
∞
a0 L X bn
= +L (−1)n
2 n=1
nπ
∞
X bn  nπx 
−L cos
n=1
nπ L
∞
an − (−1)n a0
X    nπx 
+L sin ,
n=1
nπ L

where we have used eq. (8).

Note that the first infinite series on RHS of the last equality above, forms part of the
constant term in the Fourier series for F (x). This infinite series is always guaranteed to
converge – since bn comes from a Fourier series weP know2 that bn is at worst O(1/n) so
(−1)n bn /n converges by comparison test with
P
M/n for a suitable constant M .
It is to be expected that the convergence of the Fourier series for F (x) will be faster
(i.e. fewer terms will give a certain level of approximation) than for f (x) due to the
extra factor of 1/n making the coefficients decrease faster. This is unsurprising since
integration is naturally a smoothing operation. Recall also that the Dirichlet conditions
allow for finite jump discontinuities in the underlying function. But integration across
such a jump leads to a continuous function, and F (x) will always satisfy the Dirichlet
conditions if f (x) does.
Differentiation doesn’t always work!
On the other hand, term-by-term differentiation of the Fourier series of a function is not
guaranteed to yield a convergent Fourier series for the derivative! Consider this counter-
example. Let f (x) be a periodic function with period 2 such that f (x) = 1 for 0 < x < 1
and f (x) = −1 for −1 < x < 0, as shown in the figure.
1B Methods 9

1.5

0.5

−0.5

−1

−1.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

Figure 2: Plots of f (x) = 1 for 0 < x < 1 and f (x) = −1 for −1 < x < 0 (thin solid
line) and the partial sums f1 (x) (dots); f5 (x) (dot-dashed); f10 (x) (dashed); and f20 (x)
(solid).
1B Methods 10

You can readily calculate (exercise!) its Fourier series to be

∞
4 X sin([2n − 1]πx)
f (x) =
π n=1 2n − 1

and formally differentiating term by term we get

∞
?
X
0
f (x) = 4 cos([2n − 1]πx)
n=1

which is clearly divergent even though for our actual function we have f 0 (x) = 0 for all
x 6= 0! The latter may look like a rather harmless function, but what about f 0 (0)? f 0 is
not defined at x = 0 so f 0 does not satisfy the Dirichlet conditions. Why not just put in
some value f 0 (0) = c at the single point x = 0? e.g. the average R x of left and right limits,
c = 0? Well, then consider the desired relationship f (x) = −1 f (t)dt. For any finite
c, f (x) will remain at zero as x crosses x = 0 from below. To get the jump in f (x) at
x = 0, intuitively we’d need f 0 (0) to introduce a finite area under the curve, but only over
x = 0 with zero horizontal extent! i.e. we’d need f 0 (0) = ∞ with “0 · ∞ = 1”! Thus the
notion of differentiation behaves very badly (or rather most interestingly, cf later when
we discuss the so-called Dirac delta function!) when we try to differentiate over a jump
discontinuity, even if we have nice differentiable pieces on both sides.
So, when can we legitimately differentiate the Fourier series of a function term by term?
Clearly it is not enough for f to satisfy the Dirichlet conditions (merely guaranteeing a
Fourier series for f itself). It suffices for f to also not have any jump discontinuities (on
R) and we have the following result.
Theorem: Suppose f is continuous on R and has period 2L and satisfies the Dirichlet
conditions on (−L, L). Then the Fourier series for f 0 can be obtained by term-by-term
differentiation of the Fourier series for f . 
To see this, note that the conditions on f imply that f 0 satisfies the Dirichlet conditions
so both f and f 0 have Fourier series:
∞ h
1 X  nπx   nπx i
f (x) = a0 + an cos + bn sin ,
2 n=1
L L
∞ h
0 1 X  nπx   nπx i
f (x) = A0 + An cos + Bn sin .
2 n=1
L L

and so
1 2L 0 f (2L) − f (0)
Z
A0 = f (x) dx = = 0 by periodicity,
L 0 L
1 2L 0
Z  nπx 
An = f (x) cos dx,
L 0 L
1h  nπx i2L nπ Z 2L  nπx 
= f (x) cos + 2 f (x) sin dx,
L L 0 L 0 L
nπbn
= 0+ . (10)
L
1B Methods 11

where we have again used periodicity and eqs. (7,6) for the Fourier coefficients. Similarly
−nπan
Bn =
L
so the series for f 0 is obtained by term-by-term differentiation of the series for f . Note
that the differentiation of f has been reduced to just simple multiplication of the Fourier
coefficients by nπ
L
(together with cross-relating the roles of an and bn and adding in a
minus sign for the Bn ’s).

1.4 Complex form of Fourier series

When dealing with sines and cosines it is often easier and more elegant to use complex
exponentials via de Moivre’s theorem

eiθ = cos θ + i sin θ

so
 nπx  1  inπx inπx

cos e L + e− L ,
=
L 2
 nπx  1  inπx inπx

sin = e L − e− L
L 2i
and our Fourier series becomes
∞  
f (x+ ) + f (x− ) a0 X an  inπx − inπx
 b  inπx
n − inπx
= + e L +e L + e L −e L ,
2 2 n=1
2 2i
∞
inπx
X
= cn e L , (11)
−∞

where
cn = (an − ibn )/2 n > 0;
c−n = (an + ibn )/2 n > 0; (12)
c0 = a0 /2.
This is a neater (though completely equivalent) formulation. (These formulas all remain
valid even if f is complex-valued, in which case the an ’s and bn ’s are themselves complex).
We can work directly in this complex formulation by noting that the relevant complex
exponentials are orthogonal functions:
Z 2L
imπx inπx inπx −imπx
(e L , e L ) = e L e L dx = 2Lδnm for m, n ∈ Z. (13)
0

Note the signs (i.e. complex conjugation) in the integral here! – in accordance with our
definition of inner products for complex valued functions.
1B Methods 12

Using orthogonality, in the by now familiar way, we get from eq. (11):
Z 2L
1 −imπx
cm = f (x)e L dx m ∈ Z.
2L 0
For real-valued functions f (most functions in this course) we immediately get c−m = c∗m
so we need only compute c0 (which is real) and cm for m > 0.
Example. (Differentiation rule revisited). Assuming we can differentiate the Fourier
series term by term, in the complex representation we write
∞
inπx
X
f (x) = cn e L ,
n=−∞
∞
df X inπx
= Cn e L
dx n=−∞

and the differentiation rule then gives the single simple formula:
inπ
Cn = cn holding for all n ∈ Z.
L

1.5 Half-range series

Consider a function f (x) defined only on the “half range” 0 ≤ x ≤ L. It is possible to

extend this function to the full range −L ≤ x ≤ L (and then to a 2L-periodic function)
in two natural different ways, with different symmetries.
Fourier sine series: odd functions
The function f (x) can be extended to be an odd function f (−x) = −f (x) on −L ≤ x ≤
L, and then extended as a 2L-periodic function. In this case, from eq. (7), an = 0 for all
n and we get a Fourier sine series for f (note the range of integration):
∞
f (x+ ) + f (x− ) X  nπx 
= bn sin ;
2 n=1
L
2 L
Z  nπx 
bn = f (x) sin dx. (14)
L 0 L
i.e. f (x) on [0, L] has been represented as a Fourier series with only sine terms.
Fourier cosine series: even functions
Alternatively, the function f (x) can be extended to be an even function f (−x) = f (x)
on −L ≤ x ≤ L, and then extended as a 2L-periodic function. In this case, from eq. (6),
bn = 0 for all n and we get a Fourier cosine series (note again the range of integration):
∞
f (x+ ) + f (x− ) a0 X  nπx 
= + an cos ;
2 2 n=1
L
2 L
Z  nπx 
an = f (x) cos dx. (15)
L 0 L
1B Methods 13

which again represents f on [0, L] but now as a Fourier series with only cosine terms.

1.6 Parseval’s theorem for Fourier series

The integral of a squared periodic function (or squared modulus for complex functions)
is often of interest in applications, e.g. representing the energy of a periodic signal
Z 2L
E= |f (x)|2 dx = (f, f ). (16)
0

Substituting the complex form of the Fourier series

∞
inπx
X
f (x) = cn e L

n=−∞

and using the orthogonality property eq. (13) of the complex exponentials we immedi-
ately get
Z 2L X∞
2
|f (x)| dx = 2L |cn |2 .
0 n=−∞

This result is called Parseval’s theorem. Equivalently this can be expressed in terms
of the an and bn using eq. (12) as
Z 2L " ∞
#
2
|a 0 | X
[f (x)]2 dx = L + (|an |2 + |bn |2 ) (17)
0 2 n=1

i.e. the energy is obtained by adding together contributions from separate sinusoidal
harmonics whose energies are proportional to their squared amplitudes. If f is a real-
valued function, we can remove all the modulus signs in the above formula.
Remark: Parseval’s formula can be seen as a kind of infinite dimensional version of
Pythagoras’ theorem (that the squared length of a vector is the sum of the squared
components in any orthonormal basis). Indeed on [−L, L] the following functions form
an orthonormal set (i.e. pairwise orthogonal and each having norm 1):
 √
 c0 = 1/ 2L
fn (x) = √1L sin nπx
L
for n = 1, 2, . . .
1 nπx
gn (x) = L cos L for n = 1, 2, . . .
 √

The Fourier series eq. (4) with these slightly rescaled basic functions becomes
∞
√ √
r
L X
f (x) = ( a0 ) c 0 + Lan fn (x) + Lbn gn (x)
2 n−1

and then Parseval’s theorem eq. (17) is formally just Pythagoras’ theorem in this infinite
dimensional setting.
1B Methods 14

Example. Consider again the sawtooth function f (x) = x for −L ≤ x ≤ L. If we

substitute eq. (8) into Parseval’s formula eq. (17) we get
L ∞
2L3 4L2
Z X
2
x dx = =L
−L 3 m=1
m2 π 2

giving the nice formula

π2 1 1 1
=1+ + + + ...
6 4 9 16
Parseval’s theorem is indeed commonly used to construct such tantalising equalities. As
another example (exercise) Parseval’s formula can be applied to eq. (9) to obtain
∞
X 1 π4
= . (18)
m=1
m4 90
1B Methods 15

2 STURM-LIOUVILLE THEORY

Sturm-Liouville (SL) theory is about the properties of a particular class of second order
linear ODEs that arise very commonly in physical applications (as we’ll see more later).
Recall that in our study of Fourier series we intuitively viewed (complex-valued) functions
on [a, b] as vectors in an infinite dimensional vector space equipped with an inner product
defined by Z b
(f, g) = f ∗ (x)g(x)dx. (19)
a

A fundamentally important feature was that the basic Fourier (trig or complex exponen-
tial) functions were orthogonal relative to this inner product and the set of them was
complete in the sense that any (suitably well behaved) function could be expressed as an
infinite series in terms of them.
In finite dimensional linear algebra of vectors with an inner product we have a very
nice theory of self-adjoint or Hermitian matrices (that you saw in first year!) viz. their
eigenvalues are real, eigenvectors belonging to different eigenvalues are orthogonal and
we always have a complete set (i.e. a full basis) of orthonormal eigenvectors. SL theory
can be viewed as a lifting of these ideas to the infinite dimensional setting, with vectors
being replaced by functions (as before), matrices (i.e. linear maps on vectors) by linear
second order differential operators, and we’ll have a notion of self-adjointness for those
operators. The basic formalism of Fourier series will reappear as a simple special case!

2.1 Revision of second order linear ODEs

Consider the general linear second-order differential equation

d2 d
Ly(x) = α(x) 2 y + β(x) y + γ(x)y = f (x), (20)
dx dx
where α, β, γ are continuous, f (x) is bounded, and α is nonzero (except perhaps at a
finite number of isolated points), and a ≤ x ≤ b (which may tend to −∞ or +∞).
The homogeneous equation Ly = 0 has two non-trivial linearly independent solutions
y1 (x) and y2 (x) and its general solution is called the complementary function

yc (x) = Ay1 (x) + By2 (x).

Here A and B are arbitrary constants. For the inhomogeneous or forced equation
Ly = f (f (x) describes the forcing) it is usual to seek a particular integral solution
yp which is just any single solution of it. Then the general solution of eq. (20) is

y(x) = yc (x) + yp (x).

Finding a particular solutions can sometimes involve some inspired guesswork e.g. substi-
tuting a suitable guessed form for yp with some free parameters which are then matched
1B Methods 16

to make yp satisfy the equation. However there are some more systematic ways of con-
structing particular integrals: (a) using the theory of so-called Green’s functions that
we’ll study in more detail later, and (b) using SL theory, which also has other important
uses too – later we will see how it can be used to construct solutions to homogeneous
PDEs, especially in conjunction with the method of separation of variables, which reduces
the PDE into a set of inter-related Sturm-Liouville ODE problems.
In physical applications (modelled by second order linear ODEs) where we want a unique
solution, the constants A and B in the complementary function are fixed by imposing
suitable boundary conditions (BCs) at one or both ends. Examples of such conditions
include:
(i) Dirichlet boundary value problems: we specify y on the two boundaries e.g. y(a) = c
and y(b) = d;
(ii) Homogeneous BCs e.g. y(a) = 0 and y(b) = 0 (homogeneous conditions have the
feature that if y1 and y2 satisfy them then so does y1 + y2 );
(iii) Initial value problems: y and y 0 are specified at x = a;
(iv) Asymptotic conditions e.g. y → 0 as x → ∞ for infinite domains;
etc.

2.2 Properties of self-adjoint matrices

As a prelude to SL theory let’s recall some properties of (complex) N -dimensional vectors

and matrices. If u and v are N -dimensional complex vectors, represented as column
vectors of complex numbers then their inner product is

(u, v) = u† v

where the dagger denotes ‘complex conjugate transpose’ (so u† is a row vector of the
complex conjugated entries of u).
If A is any N ×N complex matrix, its adjoint (or Hermitian conjugate) is A† (i.e. complex
conjugate transposed matrix) and A is self-adjoint or Hermitian if A = A† . There is a
neater (more abstract..) way of defining adjoints: B is the adjoint of A if for all vectors
u and v we have:
(u, Av) = (Bu, v) (21)
(as you can easily check using the property that (Bu)† = u† B † ). Note that this charac-
terisation of the adjoint depends only on the notion of an inner product so we can apply
it in any other situation where we have a notion of inner product (and you can probably
imagine where this is leading!...)
Now let A be any self-adjoint matrix. Its eigenvalues λn and corresponding eigenvectors
vn are defined by
Avn = λn vn (22)
and you should recall the following facts:
If A is self-adjoint then
(1) the eigenvalues λn are all real;
1B Methods 17

(2) if λm 6= λn then corresponding eigenvectors are orthogonal (vm , vn ) = 0;

(3) by rescaling the eigenvectors to have unit length we can always find an orthonormal
basis of eigenvectors {v1 , . . . , vN } so any vector w in CN can be written as a linear com-
bination of eigenvectors.
Note: it is possible for an eigenvalue λ to be degenerate i.e. having more than one
linearly independent eigenvector belonging to it. For any eigenvalue, the set of all as-
sociated eigenvectors forms a vector subspace and for our orthogonal basis we choose
an orthonormal basis of each of these subspaces. If λ is non-degenerate, the associated
subspace is simply one-dimensional.
(4) A is non-singular iff all eigenvalues are nonzero.
The above facts give a neat way of solving the linear equation Ax = b for unknown
x ∈ CN , when A is nonsingular and self-adjoint. Let {v1 , . . . , vN } be an orthonormal
basis of eigenvectors belonging to eigenvalues λ1 , . . . , λN respectively. Then we can write
X X
b= βi vi x= ξi vi

where the βj = (vj , b) (by orthonormality of the vi ) are known and ξj are the unknowns.
Then X X X X
Ax = A ξvi = ξi Avi = ξi λi vi = b = βi vi . (23)
Forming the inner product with vj (for any j) gives ξj λj = βj so ξj = βj /λj and we get
P βj
our solution x = v . For this to work, we need that no eigenvalue is zero. If we
λj j
have a zero eigenvalue i.e. a nontrivial solution of Ax = 0 then A is singular and Ax = b
either has no solution or a non-unique solution (depending on the choice of b).

2.3 Self-adjoint differential operators

Consider the general second order linear differential operator L:

d2 d
Ly = α(x) 2
y + β(x) y + γ(x)y
dx dx
for 0 ≤ x ≤ b (and α, β, γ are all real valued functions). In terms of the inner product
eq. (19) of functions, we define L to be self-adjoint if

(y1 , Ly2 ) = (Ly1 , y2 )

Rb ∗
Rb (24)
i.e. y (x) Ly2 (x) dx
a 1
= a (Ly1 (x))∗ y2 (x) dx
for all functions y1 and y2 that satisfy some specified boundary conditions. It is important
to note that self-adjointness is not a property of L alone but also incorporates a specifi-
cation of boundary conditions restricting the class of functions being considered i.e. we
are also able to vary the underlying space on which L is being taken to act. This feature
arises naturally in many applications.
Note that (Ly)∗ = L(y ∗ ) since we are taking L to have real coefficient functions α, β, γ.
Furthermore if we work with real-valued functions y then the complex conjugations in
eq. (24) can be omitted altogether.
1B Methods 18

Eigenfunctions of L and weight functions

Let w(x) be a real-valued non-negative function on [a, b] (with at most a finite number
of zeroes). A function y (satisfying the BCs being used) is an eigenfunction for the
self-adjoint operator L with eigenvalue λ and weight function w if

Ly(x) = λw(x)y(x). (25)

Note that this is formally similar to the matrix eigenvector equation eq. (22) but here
we have the extra ingredient of the weight function. Equations of this form (with various
choices of w) occur frequently in applications. (In the theory developed
√ below, the
appearance of w could be eliminated by making the substitution ỹ = wy and replacing
Ly by √1w L( √ỹw ) but it is generally simpler to work with w in place, as done in all
textbooks, and express our results correspondingly).
Eigenvalues and eigenfunctions of self-adjoint operators enjoy a series of properties that
parallel those of self-adjoint matrices.
Property 1: the eigenvalues are always real.
Property 2: eigenfunctions y1 and y2 belonging to different eigenvalues λ1 6= λ2 are
always orthogonal relative to the weight function w:
Z b
w(x)y1∗ (x)y2 (x) dx = 0. (26)
a

Thus by rescaling the eigenfunctions we can form an orthonormal set

s
Z b
Y (x) = y(x)/ w|y|2 dx.
a

Remark: Note that the inner product eq. (24), used to define self-adjointness, has no
weight function (i.e. w = 1 there) whereas the eigenfunctions are orthogonal only if we
incorporate the weight w from the eigenvalue √equation eq. (25) into the inner product.
Alternatively we may think of the functions wyi as being orthogonal relative to the
unweighted inner product.
Remark: We may always take our eigenfunctions to be real-valued functions. This is
because in eq. (25), λ, w and the coefficient functions of L are all real. Hence by taking
the complex conjugate of this equation we see that if y is an eigenfunction belonging to λ
then so is y ∗ . Hence if the eigenvalue is nondegenerate, y must be real (i.e. y = y ∗ ). For
degenerate eigenvalues we can always take the two real functions (y + y ∗ ) and (y − y ∗ )/i
as our eigenfunctions, with the same span.
In this course we will always assume that our eigenfunctions are real-valued, so we can
omit the complex conjugation in the weighted inner product expressions such as eq. (26).
Property 3: There is always a countable infinity of eigenvalues λ1 , λ2 , λ3 , . . . and the
corresponding set of (normalised) eigenfunctions Y1 (x), Y2 (x), . . . forms a complete basis
for functions on [a, b] satisfying the BCs being used i.e. any such function f can be
1B Methods 19

expressed as
∞
X
f (x) = An Yn (x)
n=1

and property 2 gives Z b

An = w(x)f (x)Yn (x) dx.
a

(Don’t forget here to insert the weight function into the “inner product” integral!)
Remark: the discreteness of the series of eigenvalues is a remarkable feature here. The
eigenvalue equation itself appears to involve no element of discreteness, and this can be
intuitively attributed to the imposition of boundary conditions, as illustrated in the next
example below. 
Demonstration of the completeness property 3 is beyond the scope of this course, but
properties 1 and 2 can be seen using arguments similar to those used in the finite dimen-
sional case, for self-adjoint matrices. Suppose

Lyi = λi yi w Lyj = λj yj w

so (including here the complex conjugations for full generality)

Z b Z b
(yj , Lyi ) = yj∗ Lyi dx = λi yj∗ yi w dx (27)
a a
Z b Z b
(yi , Lyj ) = yi∗ Lyj dx = λj yi∗ yj w dx (28)
a a
The self-adjointness of L gives

(yi , Lyj ) = (Lyi , yj ).

Now for any f, g we have (f, g) = (g, f )∗ so (Lyi , yj ) = (yj , Lyi )∗ . Hence

(yj , Lyi )∗ − (yi , Lyj ) = 0

and substituting eqs. (27,28) gives

Z b
(λ∗i − λj ) yi∗ yj w dx = 0.
a
Rb
Now if i = j we get (noting a
yi∗ yi w dx is never zero as the integrand is non-negative)

λ∗i − λi = 0 i.e. λi is real.

Since all λ’s are thus real, if now i 6= j we have λ∗i − λj = λi − λj 6= 0 so then
Z b
yi∗ yj w dx = 0 for i 6= j,
a
1B Methods 20

completing the proof of properties 1 and 2. 

Let’s now illustrate these ideas with the simplest example.
Example (Fourier series again!) Consider

d2
L= on 0 ≤ x ≤ L
dx2
i.e. the coefficient functions are α(x) = 1, β(x) = γ(x) = 0. We impose the homogeneous
boundary conditions:
y(0) = 0 y(L) = 0
and we take the weight function to be simply

w(x) = 1.

We will work only with real-valued functions (and hence omit all complex conjugations).
Is L self-adjoint? Well, we just need to calculate
Z L Z L
00
(y1 , Ly2 ) = y1 y2 dx and (Ly1 , y2 ) = y100 y2 dx
0 0

Integrating by parts twice we get

Z L Z L Z L
00 0 L
y2 y1 dx = [y2 y1 ]0 − y20 y10 dx = [y2 y10 − y20 y1 ]L0 + y200 y1 dx
0 0 0

so
(Ly1 , y2 ) − (y1 , Ly2 ) = [y2 y10 − y20 y1 ]L0 .
With our BCs we see that RHS = 0 so with this choice of BCs L is self-adjoint.
Let’s calculate its eigenvalues and eigenfunctions:

Ly = −λwy = −λy

(the minus sign on RHS being for convenience, just a relabelling of the λ values) i.e.

y 00 = −λy with y(0) = y(L) = 0.

The solutions are well known to be

nπx n2 π 2
yn = sin λn = .
L L2
Properties 2 and 3 then reproduce the theory of half-range Fourier sine series. You
can easily check that if we had instead taken the same L but on [−L, L] with periodic
boundary conditions y(−L) = y(L) and y 0 (−L) = y 0 (L) (and weight function w(x) = 1
again) then sin nπx/L and cos nπx/L would be (real) eigenfunctions belonging to the
(now degenerate) eigenvalues n2 π 2 /L2 and properties 2 and 3 give the formalism of full
range Fourier series (at least as applicable to suitably differentiable functions).
1B Methods 21

2.4 Sturm-Liouvlle theory

The above example is the simplest case of a so-called Sturm-Liouville equation. Con-
sider again the general second order linear differential operator (with new names for the
coefficient functions, as often used in texts)

Ly ≡ p(x)y 00 + r(x)y 0 + q(x)y (29)

where p, q, r are real-valued functions on a ≤ x ≤ b. How can we choose the func-

tions p, q, r (and also associated BCs) to make L self-adjoint? An important way is the
following. We will require that
dp
r(x) = (30)
dx
so we can write L as
Ly = (py 0 )0 + qy.
Can L be self-adjoint? (recalling that we still need to specify some BCs!). Well, inte-
grating by parts twice (as in the above example, and taking all functions to be real) you
can readily verify that
Z b
(y1 , Ly2 ) − (Ly1 , y2 ) = y1 [(py20 )0 + qy2 ] − y2 [(py10 )0 + qy1 ] dx
a
= [p(y2 y10 − y1 y20 )]ba .

L will be self-adjoint if we impose BCs making the above boundary term combination
zero. The following are examples of such BCs:
(i) y = 0 at x = a, b;
(ii) y 0 = 0 at x = a, b;
(iii) y + ky 0 = 0 at x = a, b (for any k constant);
(iv) periodic BCs: y(a) = y(b) and y 0 (a) = y 0 (b);
(v) p = 0 at x = a, b (the endpoints of the interval are then singular points of the ODE).
If w(x) is any weight function then the corresponding eigenfunction equation

Ly = (py 0 )0 + qy = −λwy

(with any choice of BCs making L self-adjoint) is called a Sturm-Liouville equation.

Such equations arise naturally in many applications (we’ll see some in section 2) and the
eigenfunctions/values are then guaranteed to satisfy all the extremely useful properties
1,2,3 above. The Fourier series example above corresponds to the simplest non-trivial
case of an SL equation (with p(x) = 1, q(x) = 0 and w(x) = 1).
Reducing general second order L’s to SL form
The SL condition eq. (30) viz. that r = p0 , on the coefficients of a general second order
operator appears to be a nontrivial restriction, but in fact, any second order differential
operator L as in eq. (29) can be re-cast into the SL form as follows.
Consider the general eigenfunction equation

py 00 + ry 0 + qy = −λwy a≤x≤b (31)

1B Methods 22

(with weight w non-negative but r not necessarily equal to p0 ). Consider multiplying

through by some function F (x). The new coefficients of y 00 and y 0 are F p and F r respec-
tively and we want to choose F (then called the integrating factor) to have

(F r) = (F p)0 i.e. pF 0 = (r − p0 )F

so x
r − p0
Z  
F (x) = exp dx.
p
Then eq.(31) takes the SL form

[F (x)p(x)y 0 ]0 + F (x)q(x)y = −λF (x)w(x)y

with a new weight function F (x)w(x) (which is still non-negative since F (x) is a real
exponential and hence always positive) and new coefficient functions F p and F q.
Example. (An SL equation with integrating factor and non-trivial weight function)
Consider the eigenfunction/eigenvalue equation on [0, π]:
1
Ly = y 00 + y 0 + y = −λy
4
with boundary conditions

y = 0 at x = 0 and y − 2y 0 = 0 at x = π.

This is not in SL form since p(x) = 1 and r(x) = 1 6= p0 (x). But the integrating factor is
easily calculated: Z x
r − p0
F = exp dx = ex .
p
Multiplying through by this F gives the self-adjoint form (noting also the form of the
given BCs!):
ex
 
d x dy
e + y = −λex y
dx dx 4
(and we can view −λ as the eigenvalue).
To solve for the eigenfunctions it is easier to use the original form of the equation (second
using standard methods (i.e. substitute y = eσx giving
order linear, constant coefficients) √
1 1
σ 2 + σ + 4 + λ = 0 so σ = − 2 ± i λ) to obtain

y(x) = Ae−x/2 cos µx + Be−x/2 sin µx

√
where we have written µ = λ (with µ ≥ 0) and A, B are arbitrary constants of integra-
tion.
The first BC requires A = 0 and then the second BC gives the transcendental equation
(as you should check):
tan µπ = µ. (32)
1B Methods 23

Now to study the latter condition, in the positive quadrant of an xy plane imagine
plotting the 45◦ line y = x and the graph of y = tan πx. The line crosses each branch
of the tan function once giving an infinite sequence µ1 , µ2 , . . . of increasing solutions of
eq. (32). As n → ∞ (i.e. large x and y values in the plane) these crossing points
approach the vertical asymptotes of the tan πx function, which are at xπ = (2n + 1)π/2
so we see that µn → (2n + 1)/2 as n → ∞ i.e. the eigenvalues have asymptotic form
λn ≈ (2n+1)2 /4. The associated eigenfunctions are proportional to yn (x) = e−x/2 sin µn x.
They are orthogonal if the correct weight function ex is used:
Z π
ex ym (x)yn (x) dx = 0 if m 6= n
0

as you could verify by a (tedious...) direct integration (and you will need to use the
special property eq. (32) of the µn values.) 

2.5 Parseval’s identity and least square approximations

Looking back at Parseval’s theorem for Fourier series we see that its derivation depends
only on the orthogonality of the Fourier functions and not their particular (e.g. trig)
form. Hence we can obtain a similar Parseval formula associated to any complete set of
orthogonal functions, such as our SL eigenfunctions. Indeed let {Y1 (x), Y2 (x), . . .} be a
complete orthonormal set of functions relative to an inner product with weight function
w Z b
wYi Yj dx = δij (33)
a
and suppose that
∞
X Z b
f (x) = An Yn (x) An = Ym f dx (34)
n=1 a

(for simplicity we’re assuming here that f and all Yn ’s are real functions). Using the
series for f and the orthogonality conditions we readily get
Z b Z b X X ∞
X
2
wf dx = w( Ai Yi )( Aj Yj ) dx = A2n .
a a n=1

Finally, it is possible to establish that the representation of a function in an eigenfunction

expansion is the “best possible” representation in a certain well-defined sense. Consider
the partial sum
N
X
SN (x) = An Yn (x),
n=1

The mean square error involved in approximating f (x) by SN (x) is

Z b
N = w[f − SN (x)]2 dx.
a
1B Methods 24

How does this error depend on the coefficients Am ? Viewing the Am ’s as variables we
have
Z b N
∂ X
N = −2 w[f − An Yn ]Ym dx,
∂Am a n=1
Z b N
X Z b
= −2 wf Ym dx + 2 An wYm Yn dx,
a n=1 a

= −2Am + 2Am = 0,

once again using eqs. (33,34). Therefore the coefficients extremize (minimize) the error
in a mean square sense, and so give the ‘best’ partial sum representation of a function is
in terms of any (partial) eigenfunction expansion. This property is important computa-
tionally, where we want the best approximation within given resources.

2.6 SL theory and inhomogeneous problems

For systems of linear equations Ax = b whose coefficient matrix was self-adjoint, we

described a method of solution that utilised the eigenvectors of A – look back at eq. (23)
et. seq. The same ideas may be applied in the context of self-adjoint differential operators
and their eigenfunctions. Consider a general inhomogeneous (or forced) equation

Ly(x) = f (x) = wF (x) (35)

where L on [a, b] (with specified BCs) is self-adjoint, and we have also introduced a weight
function w.
Mimicking the development of eq. (23), let {Yn (x)} be a complete set of eigenfunctions
of L that are orthonormal relative to the weight function w:
Z b
LYn = λn wYn wYm Yn dx = δmn
a

and write X X
F (x) = An Yn y(x) = Bn Yn
Rb
where An = a wYn F dx are known and Bn are the unknowns. Substituting these into
eq. (35) gives X X
Ly = Bn λn wYn = w An Yn .
Multiplying by Ym and integrating from a to b immediately gives Bm λm = Am (by
orthonormality of the Yn ’s) and so we obtain the solution
∞
X An
y(x) = Yn (x). (36)
n=1
λn

Here we must assume that all eigenvalues λn are non-zero.

1B Methods 25

Example. In some applications, when a system modelled by a homogeneous equation

Ly = (py 0 )0 + qy = 0

is subjected to forcing, the function q develops a weighted linear term and we get

L̃y = (py 0 )0 + (q + µw)y = f

where w is a weight function and µ is a given fixed constant. This occurs for example
in the analysis of a vibrating non-uniform elastic string with fixed endpoints; p(x) is the
mass density along the string and µ, f depend on the applied driving force.
The eigenfunction equation for L̃ (with weight function w, eigenvalues λ) is

L̃y = Ly + µwy = λwy.

Hence we easily see that the eigenfunctions Yn of L̃ are those of L but with eigenvalues
(λn − µ) where λn are the eigenvalues of L, and our formula eq. (36) above gives
∞
X An
y(x) = Yn (x).
n=1
(λn − µ)

This derivation is valid only if µ does not coincide with any eigenvalue λn . If µ does
coincide with an eigenvalue then this method fails and (as in the finite dimensional
matrix case) the solution becomes either non-unique or non-existent, depending on the
choice of RHS function f .

2.7 The notion of a Green’s function

Let us look a little more closely at the structure of the solution formula eq. (36). Sub-
Rb
stituting An = a w(ξ)Yn (ξ)F (ξ) dx and interchanging the order of summation and inte-
gration we get
Z bX ∞
Yn (x)Yn (ξ)
y(x) = w(ξ)F (ξ) dξ (37)
a n=1 λn
i.e. Z b
y(x) = G(x; ξ)f (ξ) dξ (38)
a
where we have reinstated f = F w and introduced the Green’s function G defined by
∞
X Yn (x)Yn (ξ)
G(x; ξ) = . (39)
n=1
λn

Note that the Green’s function depends only on L (and not the forcing function f ). It
also depends the boundary conditions, that are needed to make L self-adjoint and used
in the construction of the eigenfunctions. Via eq. (38), it provides the solution of Ly = f
for any forcing term.
1B Methods 26

By analogy with “Ax = b ⇒ x = A−1 b” we can think of “integration against G” in eq.

(38) as an expression of a formal inverse “L−1 ” of the linear differential operator L:
Z b
Ly = f ⇒ y= G(x; ξ)f (ξ) dξ.
a

This notion of a Green’s function and its associated integral operator inverting a dif-
ferential operator, is a very important construct and we’ll encounter it again later in
more general contexts, especially in solving inhomogeneous boundary value problems for
PDEs.