SL

52
Sturm-Liouville Problems
Sturm-Liouville problems are boundary-value problems that naturally arise when solving certain partial differential equation problems using a separation of variables method that will be
discussed in a later chapter. It is the theory behind Sturm-Liouville problems that, ultimately,
justifies the separation of variables method for these partial differential equation problems.
The simplest applications lead to the various Fourier series, and less simple applications lead to
generalizations of Fourier series involving Bessel functions, Hermite polynomials, etc.
Unfortunately, there are several difficulties with our studying Sturm-Liouville problems:
1.
Motivation: Had we the time, we would first discuss partial differential equation problems
and develop the separation of variables method for solving certain important types of
problems involving partial differential equations. We would then see how these SturmLiouville problems arise and why they are so important. But we dont have time. Instead,
Ill briefly remind you of some results from linear algebra that are analogous to the results
we will, eventually, obtain.
2.
Another difficulty is that the simplest examples (which are very important since they lead
to the Fourier series) are too simple to really illustrate certain elements of the theory,
while the other standard examples tend to get complicated and require additional tricks
which distract from illustrating the theory. Well deal with this as well as we can.
3.
The material gets theoretical. Sorry, there is no way around this. The end results, however,
are very useful in computations, especially now that we have computers to do the tedious
computations. I hope we get to that point.
4.
Finally, I must warn you that, in most texts, the presentation of the Sturm-Liouville
theory stinks. In most introductory ordinary differential equation texts, this material is
usually near the end, which usually means that the author just wants to finish the damn
book, and get it published. In texts on partial differential equations, the results are usually
quoted with some indication that the proofs can be found in a good text on differential
equations.
12/7/2014
Chapter & Page: 522
52.1
Sturm-Liouville
Linear Algebraic Antecedents
Let us briefly review some elements from linear algebra which we will be recreating using
functions instead of finite-dimensional vectors. But first, let me remind you of a little complex
variable algebra. This is because our column vectors and matrices may have complex values.
Recollections From (Real) Linear Algebra

For the rest of this section, let N be some positive integer. Following fairly standard convention,
well let R N denote the vector space of all N1 matrices with real components. That is, saying
that v and w are vectors in R N will simply mean

v1
w1
v
w
2
2
.
v =
and
w
=
.
.
.
.
.
vN
wN
where the vk s and wk s are real numbers.
Recall that the norm (or length) of the above v , kvk , is computed by
q
kvk = |v1 |2 + |v2 |2 + |v N |2
and that, computationally, the classic dot product of the above v and w is
v w = v1 w1 + v2 w2 + + v N w N

w1
w
2
T
= [v1 , v2 , . . . , v N ]
.. = v w
.

wN
where vT w is the matrix product with vT being the transpose of matrix v (i.e., the matrix
constructed from v by switching rows with columns).
Observe that, since v 2 = |v|2 when v is real,
v v = v1 v1 + v2 v2 + + v N v N
= |v1 |2 + |v2 |2 + + |v N |2 = kvk2
Also recall that a set of vectors {v1 , v2 , v2 , . . .} is said to be orthogonal if and only if
v j vk = 0
whenever
j 6= k
And finally, recall that a matrix A is symmetric if and only if AT = A , and that, for
symmetric matrices, we have the following theorem from linear algebra (and briefly mentioned
in chapter 38):
Theorem 52.1
Let A be a symmetric NN matrix (with real-valued components). Then both of the following
hold:
Chapter & Page: 523
1.
All the eigenvalues of A are real.
2.
There is an orthogonal basis for R N consisting of eigenvectors for A .
Ultimately, we will obtain an analog to this theorem involving a linear differential operator
instead of a matrix A .
Linear Algebra with Complex Components

Complex Conjugates and Magnitudes
Weve certainly used complex numbers before in this text, but I should remind you that a complex
number z is something that can be written as
z = x + iy
where x and y are real numbers the real and imaginary parts, respectively of z . The
corresponding complex conjugate z and magnitude |z| of z are then given by
p
|z| = x 2 + y 2 .
z = x i y
and
Now if x is a real number positive, negative or zero then x 2 = |x|2 . However, you can
easily verify that z 2 6= |z|2 . Instead, we have
|z|2 = x 2 + y 2 = x 2 (i y)2 = (x i y)(x + i y) = z z
Vectors and Matrices with Complex-Valued Components

Everything mentioned above can be generalized to C N , the vector space of all N 1 matrices
with complex components. In this case, saying that v and w are vectors in C N simply means
v1
v
2
v =
..
.

vN
and
w1
w
2
w =
..
.

wN
where the vk s and wk s are complex numbers. Since the components are complex, we define
the complex conjugate of v in the obvious way,
v1
v
2
v =
..
.
v N
The natural norm (or length) of v is still given by

kvk =
q
|v1 |2 + |v2 |2 + |v N |2
Chapter & Page: 524
Sturm-Liouville
However, when we try to relate this to the classic dot product v v , we get
kvk2 = |v1 |2 + |v2 |2 + + |v N |2
= v1 v1 + v2 v2 + + v N v N
= (v ) v 6= v v
This suggests that, instead of using the classic dot product, we use the (standard vector) inner
product of v with w , which is denoted by h v | w i and defined by by
h v | w i = v w = v1 w1 + v2 w2 + + v N w N

w1
w
2
T
= [v1 , v2 , . . . , v N ]
.. = (v ) w
.

wN
Then
h v | v i = v v = kvk2
We also adjust our notion of orthogonality by saying that any set {v1 , v2 , . . .} of vectors in
C N is orthogonal if and only if
m n
v v
= 0
whenever m 6= n .
The inner product for vectors with complex components is the mathematically natural extension of the standard dot product for vectors with real components. Some easily verified (and
useful) properties of this inner product are given in the next theorem. Verifying it will be left as
an exercise (see exercise 52.4).
Theorem 52.2 (properties of the inner product)

Suppose and are two (possibly complex) constants, and v , w , and u are vectors in C N .
Then
1.
h v | w i = h w | v i ,
2.
h u | v + w i = h u | v i + h u | w i ,
3.
h v + w | u i = h v | u i + h w | u i ,
and
4.
h v | v i = kvk .
Later, well define other inner products for functions. These inner products will have very
similar properties to those in given in the last theorem.
Adjoints
The adjoint of any matrix A denoted A is the transpose of the complex conjugate of A ,

A = (A )T
(equivalently, AT ) .
Chapter & Page: 525
That is, A is the matrix obtained from matrix A by replacing each entry in A with its complex
conjugate, and then switching the rows and columns (or first switch the rows and columns and
then replace the entries with their complex conjugates you get the same result either way).
!Example 52.1:
If
A =
"
1 + 2i
6i
3 4i
7
5i
8i
then
A =
"
1 + 2i
6i
3 4i
7
5i
8i
# !T
"
1 2i
=
6i
3 + 4i
7
5i
8i
#T
1 2i
= 3 + 4i
5i
6i
7
8i
This adjoint turns out to be more useful than the transpose when we allow vectors to have
complex components.
A matrix A is self adjoint 1 if and only if A = A . Note that:
1.
A self-adjoint matrix is automatically square.
2.
If A is a square matrix with just real components, then A = AT , and A is self adjoint
means the same as A is symmetric.
If you take the proof of theorem 52.1 and modify it to take into account the possibility of
complex-valued components, you get
Theorem 52.3
Let A be a self-adjoint N N matrix. Then both of the following hold:
1.
All the eigenvalues of A are real.
2.
There is an orthogonal basis for C N consisting of eigenvectors for A .
This theorem is noteworthy because it will help explain the source of some of the terminology
that we will later be using.
What is more noteworthy is what the above theorem says about computing Av when A is
self adjoint. To see this, let
1 2 3

b , b , b , . . . , bN
be any orthogonal basis for C N consisting of eigenvectors for A (remember, the theorem says
there is such a basis), and let
{ 1 , 2 , 3 , . . . , N }
be the corresponding set of eigenvectors (so Abk = k bk for k = 1, 2, . . . , N ). Since the set
of bk s is a basis, we can express v as a linear combination of these basis vectors.
v = v1 b1 + v2 b2 + v3 b3 + + v N b N =
1 the term Hermitian is also used
N
X
k=1
vk bk
(52.1)
Chapter & Page: 526
Sturm-Liouville
and can reduce the computation of Av to

Av = A v1 b1 + v2 b2 + v3 b3 + + v N b N
= v1 Ab1 + v2 Ab2 + v3 Ab3 + + v N Ab N
= v1 1 b1 + v2 2 b2 + v3 3 b3 + + v N N b N
If N is large, this could be a lot faster than doing the basic component-by-component matrix
multiplication. (And in our analog with functions, N will be infinite.)
Finding Vector Components

One issue is finding the components (v1 , v2 , v3 , . . . , v N ) in formula (52.1) for v . But a useful
formula is easily derived. First, take the inner product of both sides of (52.1) with one of the
bk s , say, b1 and then repeatedly use the linearity property described in theorem 52.2,

b v = b1 v1 b1 + v2 b2 + v3 b3 + + v N b N

.
= v1 b1 b1 + v2 b1 b2 + v3 b1 b3 + + v N b1 b N
Remember, this set of bk s is orthogonal. That means
1 k
b b = 0
if 1 6= k
On the other hand,

2

b1 b1 = b1
Taking this into account, we continue our computation of b1 v :

b v = v1 b1 b1 + v2 b1 b2 + v3 b1 b3 + + v N b1 b N
2
= v1 b1 + v2 0 + v3 0 + + v N 0 .
So,
Solving for v1 gives us

2
b1 v = v1 b1
v1 =

b1 v
2
b1
Repeating this using each bk yields the next theorem:
Theorem 52.4
Let
b1 , b2 , b3 , . . . , b N
be an orthogonal basis for C N . Then, for any vector v in R N ,

1
v = v1 b + v2 b + v3 b + + v N b
where
vk =

bk v
2
bk
N
X
vk bk
(52.2a)
k1
for k = 1, 2, 3, . . . , N
(52.2b)
Chapter & Page: 527
If you know a little about Fourier series, then you may recognize formula set (52.2) as a
finite-dimensional analog of the Fourier series formulas. If you know nothing about Fourier
series, then Ill tell you that a Fourier series for a function f is an infinite-dimensional analog
of formula set (52.2). We will eventually see this.
About Normalizing a Basis

In the above, we assumed that
b1 , b2 , b3 , . . . , b N
is an orthogonal basis for C N . If it had been orthonormal, then we would also have had
k
b = 1
for k = 1, 2, 3, . . . , N ,
and the formulas in our last theorem would have simplified somewhat. In fact, given any orthogonal set
1 2 3

b , b , b , . . . , bN
,
we can construct a corresponding orthonormal set
1 2 3

n , n , n , . . . , nN
by just letting
bk
nk = k
b
for k = 1, 2, 3, . . . , N
Moreover, if bk is an eigenvector for a matrix A with corresponding eigenvalue k , so is nk .

When we so compute the nk s from the bk s , we are said to be normalizing our bk s .
Some authors like to normalize their orthogonal bases because it does yield simpler formulas for
computing with such a basis. These authors, typically, are only deriving pretty results and are not
really using them in applications. Those that really use the results rarely normalize, especially
when the dimension is infinite (as it will be for us), because normalizing leads to artificial
formulas for the basis vectors, and dealing with these artificial formulas for basis vectors usually
complicates matters enough to completely negate the advantages of having the simpler formulas
for computing with these basis vectors.
We wont normalize.
Inner Products and Adjoints

You should recall (or be able to quickly confirm) that
T
(AB)T = BT AT ,
AT = A , (AB) = A B
and
With these and the definition of the adjoint, you can easily verify that

(AB) = B A
and
A = A .
Also observe that, if
v1
v
2
v =
..
.
vN
and
w1
w
2
w =
..
.
wN
= A .
Chapter & Page: 528
then

v w = v1 , v2
Sturm-Liouville
w1

w2

, . . . , vN .
= v1 w1 + v2 w2 + + v N w N = h v | w i
..
wN
So, for any N N matrix A and any two vectors v and u in C N ,

v A u = v A u = v A u = (Av) u = h Av | u i
Consequently, if A is self adjoint (i.e., A = A ), then

h v | Au i = h Av | u i
for every v, u in C N
Its a simple exercise in linear algebra to show that the above completely characterizes
adjointness and self adjointness for matrices. That is, you should be able to finish proving
the next theorem. Well use the results to extend the these concepts to things other than matrices.
Theorem 52.5 (characterization of adjointness)

Let A be an N N matrix. Then,

v A u = h Av | u i
for every v and u in C N
Moreover, both of the following hold:

1.
B = A if and only if
h v | Bu i = h Av | u i
2.
A is self adjoint if and only if

h v | Au i = h Av | u i
Comments
To be honest, we are not going to directly use the material weve developed over the past several
pages. The reason we went over the theory of self-adjoint matrices and related material
concerning vectors in C N is that the Sturm-Liouville theory well be developing is a functional
analog of what we just discussed, using differential operators and functions instead of matrices
and vectors in C N . Understanding the theory and computations weve just developed should
expedite learning the theory and computations we will be developing.
Boundary-Value Problems with Parameters
52.2
Chapter & Page: 529
In each boundary-value problem of interest to us, the differential equation will involve both
the unknown function which (following tradition) we will denote by or (x) and a
parameter (basically, is just some yet undetermined constant). Well assume that each of
these equations can be written as
A(x)
d 2
d
+ [C(x) + ] = 0
+ B(x)
dx
dx 2
(52.3a)
d
d 2
+ B(x)
+ C(x) =
2
dx
dx
(52.3b)
or, equivalently, as
A(x)
where A , B and C are known functions.

A solution to such a problem is a pair (, ) that satisfies the given problem both the
differential equation and the boundary conditions. To avoid triviality, we will insist that the
function be nontrivial (i.e., not always zero). Traditionally, the is called an eigenvalue, and
the corresponding is called an eigenfuction. Well often refer to the two together, (, ) , as
an eigen-pair. The entire problem can be called an eigen-problem, a boundary-value problem or
a Sturm-Liouville problem (though we wont completely define just what a Sturm-Liouville
problem is until later).
In our problems, we will need to find the general solution = to the given differential
equation for each possible value of , and then apply the boundary conditions to find all possible
eigen-pairs. Technically, at this point, can be any complex number. However, thanks to the
foreknowledge of the author, we can assume is real. Why this is a safe assumption will be one
of the things we will later need to verify (its analogous to the fact that self-adjoint matrices have
only real eigenvalues). What we cannot yet do, though, is assume the s all come from some
particular subinterval of the real line. This means you must consider all possible real values for
, and take into account that the form of the solution may be quite different for different
ranges of these s .
!Example 52.2: The simplest (and possibly most important) example of such an boundaryvalue problem with parameter is
with boundary conditions
+ = 0
(52.4a)
and
(52.4b)
(0) = 0
(L) = 0
where L is some positive number (think of it as a given length).

The above differential equation is a simple second-order homogeneous differential equation with constant coefficients. Its characteristic equation is
with solution
r2 + = 0
r = .
In this example, the precise formula for (x) , the equations general solution corresponding to a particular value of , depends on whether > 0 , = 0 or < 0 . Lets go
through all the cases:
Chapter & Page: 5210
Sturm-Liouville
< 0 : In this case, > 0 . For convenience, let =

r =
. Then
and
(x) = c1 ex + c2 ex
where c1 and c2 are arbitrary constants, and, as already stated, =
Applying the first boundary condition:

0 = (0) = c1 e0 + c2 e0 = c1 + c2
which tells us that

c2 = c1
and thus,

(x) = c1 ex c1 ex = c1 ex ex .
Combining this with the boundary condition at x = L yields

0 = (L) = c1 e L e L .
Now e L > e0 = 1 and e L < e0 = 1 . So
e L e L > 0
and the only way we can have

0 = c1 e L e L
is for c1 = 0 . Thus, the only solution to this problem with < 0 is

(x) = 0 e L e L = 0 ,
the trivial solution (which we dont really care about).

= 0 : Instead of using the characteristic equation, lets just note that, if = 0 , the differential equation for = 0 (x) reduces to
(x) = 0
Integrating this twice yields

.
0 (x) = 0 (x) = c1 x + c2
Applying the first boundary condition gives us

0 = 0 (0) = c1 0 + c2 = c2
Combined with the boundary condition at x = L , we then have

0 = 0 (L) = c1 L + 0 = c1 L
which says that c1 = 0 (since L > 0 ). Thus, the only solution to the differential
equation that satisfies the boundary conditions when = 0 is
0 (x) = 0 x + 0 = 0
again, just the trivial solution.
> 0 : This time, it is convenient to let = , so that the solution to the characteristic
equation becomes
r = = i .
While the general formula for can be written in terms of complex exponentials, it
is better to recall that these complex exponentials can be written in terms of sines and
cosines, and that the general formula for can then be given as
(x) = c1 cos(x) + c2 sin(x)
where, again, c1 and c2 are arbitrary constants, and, as already stated, =
Applying the first boundary condition:

0 = (0)
= c1 cos( 0) + c2 sin( 0)
= c1 1 + c2 0 = c1
which tells us that

(x) = c2 sin(x) .
With the boundary condition at x = L , this gives

0 = (L) = c2 sin( L)
To avoid triviality, we want c2 to be nonzero. So, for the boundary condition at x = L

to hold, we must have
sin( L) = 0 ,
which means that
L = an integral multiple of
Moreover, since = > 0 , L must be a positive integral multiple of . Thus, we

have a list of allowed values of ,
k
L
k =
with k = 1, 2, 3, . . .
a corresponding list of allowed values for ,

2
k = (k ) =
k
L
2
with k = 1, 2, 3, . . .
(these are the eigenvalues), and a corresponding list of (x)s (the corresponding eigenfunctions),

k
k (x) = ck sin(k x) = ck sin
x
with k = 1, 2, 3, . . .
L
where the ck s are arbitrary constants. For our example, these are the only nontrivial
eigenfunctions.
Sturm-Liouville
In summary, we have a list of solutions to our Sturm-Liouville problem, namely,

for k = 1, 2, 3, . . .
(k , k )
where, for each k ,

k =
and
k
L
2

k
k (x) = ck sin(k x) = ck sin
x
L
(with the ck s being arbitrary constants).

In particular, if L = 1 , then the solutions to our Sturm-Liouville problem" are given by

for k = 1, 2, 3, . . . .
(k , k (x)) = k 2 2 , ck sin(k x)
52.3
The Sturm-Liouville Form for a Differential

Equation
To solve these differential equations with parameters, you will probably want to use form (52.3a),
as we did in our example. However, to develop and use the theory that we will be developing
and using, we will want to rewrite each differential equation in its Sturm-Liouville form

d
d
p(x)
+ q(x) = w(x)
dx
dx
where p , q and w are known functions.

!Example 52.3:
Technically, the equation

+ = 0
is not in Sturm-Liouville form, but the equivalent equation

=
is, with p = 1 , q = 0 and w = 1 .

!Example 52.4:
The equation

d
d
sin(x)
+ cos(x) = x 2
dx
dx
is in Sturm-Liouville form, with

,
p(x) = sin(x)
q(x) = cos(x)
and
On the other hand,

x
d 2
d
+ 2
+ [sin(x) + ] = 0
dx
dx 2
is not in Sturm-Liouville form.
w(x) = x 2
The Sturm-Liouville Form for a Differential Equation
Fortunately, just about any differential equation in form (52.3a) or (52.3a) can converted to
Sturm-Liouville form using a procedure similar to that used to solve first-order linear equations.
To describe the procedure in general, lets assume we have at least gotten our equation to the
form
A(x)
d 2
d
+ B(x)
+ C(x) =
2
dx
dx
To illustrate the procedure, well use the equation

x
d
d2
+ sin(x) =
+ 2
dx
dx 2
(with (0, ) being our interval of interest).

Here is what you do:
1.
Divide through by A(x) , obtaining

d 2
B(x)
C(x)
1
+
+
=
A(x)
A(x)
A(x)
dx 2
Doing that with our example yields

d2
2 d
sin(x)
1
+
+
=
x dx
x
x
dx 2
2.
Compute the integrating factor

p(x) = e
B(x)
A(x)
dx
ignoring any arbitrary constants.
In our example,
Z
B(x)
dx =
A(x)
2
d x = 2 ln x + c
x
So (ignoring c ),
p(x) = e
B(x)
A(x)
dx
= e2 ln x = eln x = x 2
(As an exercise, you should verify that

dp
B
= p
dx
A
In a moment, well use this fact.)

3.
Using the p(x) just found:

(a) Multiply the differential equation resulting from step 1 by p(x) , obtaining
p
(b)
B d
C
p
d2
+ p
+ p =
A dx
A
A
dx 2
observe that (via the product rule),

d
d
d2
B d
p
= p 2 + p
dx
dx
dx
A dx
(c)
Sturm-Liouville
and rewrite the differential equation according to this observation,

d
d
C
p
p
+ p = .
dx
dx
In our case, we have

2

2 d
sin(x)
1
2 d
2
x
+
= x
+
2
x dx
dx
x2
d 2
d
+ 2x
+ x sin(x) = x
dx
dx 2
Oh look! By the product rule,

d
d
d
d 2
d
2 d
p
=
x
= x 2 2 + 2x
dx
dx
dx
dx
dx
dx
Using this to simplify the first two terms of our last differential equation
above, we get

d
2 d
x
+ x sin(x) = x .
dx
dx
It is now in the desired form, with

p(x) = x 2
52.4
q(x) = x sin(x)
and
w(x) = x
Boundary Conditions for Sturm-Liouville

Problems
The boundary conditions for a Sturm-Liouville problem will have to be homogeneous as defined
in the previous chapter. There is an additional condition that well derive in this section that will
help allow us to view a certain operator as self adjoint. That operator is given by the left side of
the differential equation of interest when that differential equation is written in Sturm-Liouville
form,

d
dx
p(x)
d
dx
+ q(x) = w(x)
For the rest of this chapter, this operator will be denoted by L . That is, given any suitably
differentiable function ,

d
d
L[] =
p(x)
+ q(x)
dx
dx
where p and q are presumably known real-valued functions on some interval (a, b) . Note that
our differential equation can be written as
L[] = w
Boundary Conditions for Sturm-Liouville Problems
Greens Formula, and Sturm-Liouville Appropriate

Boundary Conditions
To describe the additional boundary conditions needed, we need to derive an identity of Green.
You should start this derivation by doing exercise 52.10 on page 5236, in which you verify the
following lemma :
Lemma 52.6 (the preliminary Greens formula)

Let (a, b) be some finite interval, and let L be the operator given by

d
d
p(x)
+ q(x)
L[] =
dx
dx
where p and q are any suitably smooth and integrable functions on (a, b) . Then
b
Z b
Z b
Z b
d f dg
dg
p
dx +
q f g dx
f L[g] d x = p(x) f (x)
dx
dx dx
(52.5)
whenever f and g are suitably smooth and integrable functions on (a, b) .
Now suppose we have two functions u and v on (a, b) (assumed suitably smooth and
integrable, but, possibly, complex valued), and suppose we want to compare
Z b
Z b
u L[v] d x
and
L[u] v d x .
a
(Why? Because this will lead to our extending the notion of self adjointness as characterized
in theorem 52.5 on page 528)
If p and q are real-valued functions, then it is trivial to verify that

L[u] = L u
Using this and the above preliminary Greens formula, we see that
Z b
Z b
Z b
Z b

u L[v] d x
L[u] v d x =
u L[v] d x
v L u d x
a
#
b
Z b
Z b

du dv
dv
= pu
p
dx +
qu v d x
dx a
dx dx
a
a
"
#
b
Z b
Z b
du
dv du
pv
p
dx +
qvu d x
dx
dx dx
"
Nicely enough, most of the terms on the right cancel out, leaving us with:
Theorem 52.7 (Greens formula)

Let (a, b) be some interval, p and q any suitably smooth and integrable real-valued functions
on (a, b) , and L the operator given by

d
d
L[] =
p(x)
+ q(x) .
dx
dx
Then
Sturm-Liouville
u L[v] d x

b
du
dv
L[u] v d x = p u
v

dx
dx
(52.6)
for any two suitably smooth and differentiable functions u and v .
Equation (52.6) is known as Greens formula. (Strictly speaking, the right side of the equation
is the Greens formula for the left side).
Now we can state what sort of boundary conditions are appropriate for our discussions.
We will refer to a pair of homogeneous boundary conditions at x = a and x = b as being
Sturm-Liouville appropriate if and only if

b
du
dv
p u
v
(52.7)
= 0 .
dx
dx
whenever both u and v satisfy these boundary conditions.

For brevity, we may say appropriate when we mean Sturm-Liouville appropriate. Do
note that the function p in equation (52.7) comes from the differential equation in the SturmLiouville problem. Consequently, it is possible that a particular pair of boundary conditions is
Sturm-Liouville appropriate when using one differential equation, and not Sturm-Liouville
appropriate when using a different differential equation. This is particularly true when the
boundary conditions are of the periodic or boundedness types.
!Example 52.5:
Consider the boundary conditions

and
(a) = 0
(b) = 0
If u and v satisfy these conditions; that is,
and
u(a) = 0
and
u(b) = 0
v(a) = 0
and
v(b) = 0
then, assuming p(a) and p(b) are finite numbers,

b

du
du
dv
dv
p u
v
= p(b) (u(b))
v(b)
dx
dx a
dx x=b
dx x=b

du
dv
p(a) (u(a))
v(a)

dx
dx
x=a
x=a

du
dv
= p(b) 0
0
dx x=b
dx x=b

dv
du
p(a) 0
0
dx
dx

x=a
= 0
x=a
So
(a) = 0
and
(b) = 0
are Sturm-Liouville appropriate boundary conditions, at least whenever p(a) and p(b) are
finite.
Boundary Conditions for Sturm-Liouville Problems
A really significant observation is that, whenever u and v satisfy appropriate boundary

conditions, then Greens formula reduces to
Z b
Z b
u L[v] d x
L[u] v d x = 0 ,
a
which, in turn, gives us the following lemma, which, in turn, suggests just what we will be using
for inner products and self adjointness in the near future.
Lemma 52.8
Let (a, b) be some interval, p and q any suitably smooth and integrable real-valued functions
on (a, b) , and L the operator given by

d
d
L[] =
p(x)
+ q(x) .
dx
dx
Assume, further, that u(x) and v(x) satisfy Sturm-Liouville appropriate boundary conditions
at a and b . Then
Z b
Z b
u L[v] d x =
L[u] v d x .
a
The Rayleigh Quotient

Before we stray too far from the preliminary Greens formula (lemma 52.6), lets take a look
at the integral
Z b
L[] d x
a
when is an eigenfunction corresponding to eigenvalue for some Sturm-Liouville problem

with differential equation

d
d
p(x)
+ q(x) = w
for a < x < b .
dx
dx
In terms of the operator L , this equation is

L[] = w
and, so,
Z b
Z
L[] d x =
a
[w] d x =
((x)) (x)w(x) d x =
On the other hand, using the preliminary Greens formula gives us

Z
b

d
Z b
Z b
d d
p
d
x
+
q d x
dx a
dx
dx
a
a
b
Z b 2
Z b
d
d
= p
p d x +
q ||2 d x
dx
dx
L[] d x = p
||2 w d x
Sturm-Liouville
= p
Together, the above tells us that
b
2
|| w d x =
b

d
dx a
L[] d x = p
2

d
2

p q || d x
dx
b

d
dx a
2

d
2

p q || d x
dx
Cutting out the middle and solving for gives us the next lemma.
Lemma 52.9 (Rayleighs Quotient)

Let (, ) be an eigen-pair for a Sturm-Liouville problem with differential equation

d
d
p(x)
+ q(x) = w
for a < x < b .
dx
Then
dx
b Z b 2

d
q ||2 dx
p
+
p
dx
dx a
a
=
Z b
||2 w dx
d
(52.8)
Equation (52.8) is called the Rayleigh quotient. It relates the value of an eigenvalue to any
corresponding eigenfunction. It is of particular interest in the Sturm-Liouville problems in which
the boundary conditions ensure that
p
Then the Rayleigh quotient reduces to
b
a
b

d
dx a
= 0
2

d
p q ||2 dx
dx
Z b
||2 w dx
Now look at the right side of this equation. Remember, p , q and w are all real-valued functions
on (a, b) . It should then be clear that the integrals on the right are all real valued. Thus, must
be real valued.
If, in addition,
q(x) 0
for a < x < b
(which occurs fairly often) then, in fact,
b
a
2

d
2

p q || dx
dx
0
Z b
2
|| w dx
Now if q(x) is nonzero on some subinterval of (a, b) (in addition to being nonpositive on
(a, b) ), then the integrals must be positive for any nontrivial function , which means that
must always be positive. On the other hand, if q = 0 on (a, b) (which is quite common), then
the Rayleigh quotient further reduces to
2
d
p dx
dx
|| w dx
With a little thought, you can then show that the only possible eigenfunctions corresponding to a
zero eigenvalue, = 0 , are nonzero constant functions, and, consequently, if nonzero constant
functions do not satisfy the boundary conditions, then cannot be zero; all the eigenvalues are
positive.
52.5
Full Definition
Finally, we can fully define the Sturm-Liouville problem: A Sturm-Liouville problem is a
boundary-value problem consisting of both of the following:
1.
A differential equation that can be written in the form

d
d
p(x)
+ q(x) = w(x)
dx
dx
for a < x < b
(52.9)
where p , q and w are sufficiently smooth and integrable functions on the finite interval
(a, b) , with p and q being real valued, and w being positive on this interval. (The sign
of p and w turn out to be relevant for many results. We will usually assume they are
positive functions on (a, b) . This is what invariably happens in real applications.)
2.
A Sturm-Liouville appropriate pair of corresponding homogeneous boundary conditions

at x = a and x = b .
As stated before, a solution to a Sturm-Liouville problem consists of a pair (, ) where

is a constant called an eigenvalue and is a nontrivial (i.e, nonzero) function called an
eigenfunction which, together, satisfy the given Sturm-Liouville problem. When convenient,
we will refer to (, ) as an eigen-pair for the given Sturm-Liouville problem.
Again, Ill mention that whenever we have a Sturm-Liouville problem, we will automatically
let L be the linear differential operator defined by the left side of differential equation (52.9),

d
d
L[] =
p(x)
+ q(x) .
dx
dx
Sturm-Liouville
A Important Class of Sturm-Liouville Problems

There are several classes of Sturm-Liouville problems. One particularly important class goes
by the name of regular Sturm-Liouville problems. A Sturm-Liouville problem is said to be
regular if and only if all the following hold:
1.
The functions p , q and w are all real valued and continuous on the closed interval
[a, b] , with p being differentiable on (a, b) , and both p and w being positive on the
closed interval [a, b] .
2.
We have homogeneous regular boundary conditions at both x = a and x = b . That is

a (a) + a (a) = 0
where a and a are constants, with at least one being nonzero, and
b (b) + b (b) = 0
where a and b are constants, with at least one being nonzero.
Many, but not all, of the Sturm-Liouville problems generated in solving partial differential
equation problems are regular. For example, the problem considered in example 52.2 on page
529 is a regular Sturm-Liouville problem.
So What?
It turns out that the eigenfunctions from a Sturm-Liouville problem on an interval (a, b) can
often be used in much the same way as the eigenvectors from a self-adjoint matrix to form an
orthogonal basis for a large set of functions. Consider, for example, the eigenfunctions from
the example 52.2

k
k (x) = ck sin
x
for k = 1, 2, 3, . . . .
L
If f is any reasonable function on (0, L) (say, any continuous function on this interval), then
we will discover that there are constants c1 , c2 , c3 , . . . such that
f (x) =
X
k=1
k
x
ck sin
L
for all x in (0, L)
This infinite series, called the (Fourier) sine series for f on (0, L) , turns out the be extremely
useful in many applications. For one thing, it expresses any function on (0, L) in terms of the
well-understood sine functions.
Our next goal is to develop enough of the necessary theory to discover what I just said we
will discover. In particular we want to learn how to compute the ck s in the above expression
for f (x) . It turns out to be remarkable similar to the formula for computing the vk s in theorem
52.4 on page 526. Take a look at it right now. Of course, before we can verify this claim, we
will have to find out just what we are using for an inner product.
All this, alas, will take some time.
The Eigen-Spaces
52.6
The Eigen-Spaces
Suppose we have some Sturm-Liouville problem. For convenience, let us write the differential
equation in that problem as
L[] = w
with, as usual,
d
L[] =
dx
d
p(x)
dx
+ q(x)
One observation we can immediately make is that L is a linear operator. That is, if (x)
and (x) are any two functions, and c1 and c2 are any two constants, then
L[c1 + c2 ] = = c1 L[] + c2 L[]
Now, further suppose that both (x) and (x) are eigenfunctions corresponding to the same
eigenvalue . Then,
L[c1 + c2 ] = c1 L[] + c2 L[]
= c1 [w] + c2 [w] = w[c1 + c2 ]
This shows that the linear combination c1 + c2 also satisfies the differential equation with
that particular value of . Does it also satisfy the boundary conditions? Of course. Remember,
the set of boundary conditions in a Sturm-Liouville problem is homogeneous, meaning that, if
(x) and (x) satisfy the given boundary conditions, so does any linear combination of them.
Hence, any linear combination of eigenfunctions corresponding to a single eigenvalue is also an
eigenfunction for our Sturm-Liouville problem, a fact significant enough to write as a lemma.
Lemma 52.10
Assume (, ) and (, ) are both eigen-pairs with the same eigenvalue for some SturmLiouville problem. Then any nonzero linear combination of (x) and (x) is also an eigenfunction for the Sturm-Liouville problem corresponding to eigenvalue .
This lemma tells us that the set of all eigenfunctions for our Sturm-Liouville problem corresponding to any single eigenvector is a vector space of functions (after throwing in the zero
function). Naturally, we call this the eigenspace corresponding to eigenvalue . Keep in mind
that these functions are all solutions to the second-order homogeneous linear equation

d
d
p(x)
+ [q(x) + w(x)] = 0 ,
dx
dx
and that the general solution to such a differential equation can be written as
(x) = c1 1 (x) + c2 2 (x)
where c1 and c2 are arbitrary constants and {1 , 2 } is any linearly independent pair of solutions
to the differential equation. 1 can be chosen as one of the eigenfunctions. Whether or not 2
can be chosen to be an eigenfunction depends on whether or not there is a linearly independent
pair of eigenfunctions corresponding to this . That gives us exactly two possibilities:
Sturm-Liouville
1.
There is not an independent pair of eigenvectors corresponding to . This means the

eigenspace corresponding to eigenvalue is one dimensional (i.e., is a simple
eigenvalue), and every eigenfunction is a constant multiple of 1 .
2.
There is an independent pair of eigenvectors corresponding to . This means the

eigenspace corresponding to eigenvalue is two dimensional (i.e., is a double
eigenvalue), and every solution to the differential equation (with the given ) is an
eigenfunction.
52.7
Inner Products, Orthogonality and Generalized

Fourier Series
We must, briefly, forget about Sturm-Liouville problems, and develop the basic analog of the inner
product described for finite dimensional vectors at the beginning of this chapter. Throughout this
discussion, (a, b) is an interval, and w = w(x) is a function on (a, b) satisfying
w(x) > 0
whenever a < x < b
This function, w , will be called the weight function for our inner product. Often, it is simply the
constant 1 function (i.e., w(x) = 1 for every x ).
Also, throughout this discussion, let us assume that our functions are all piecewise continuous
and sufficiently integrable for the computations being described. Well discuss what suitably
integrable means if I decide it is relevant. Otherwise, well slop over issues of integrability.
Inner Products for Functions

Let f and g be two functions on the interval (a, b) . We define the inner product (with weight
function w ) of f with g denoted h f | g i by
h f | gi =
f (x) g(x)w(x) d x
For convenience, lets say that the standard inner product is simply the inner product with weight
function w = 1 ,
Z b
h f | gi =
f (x) g(x) d x .
a
For now, you can consider the inner product to be simply shorthand for some integrals that will
often arise in our work.
!Example 52.6:
Let (a, b) = (0, 2) and w(x) = x 2 . Then

h f | gi =
f (x) g(x)x 2 d x
Inner Products, Orthogonality and Generalized Fourier Series
In particular

Z 2
1
1

6x + 2i
=
(6x + 2i) x 2 d x
x
x
0
Z 2
=
(6x 2i)x d x
0
2

6x i2x d x
2
= 2x 3 i x 2 0 = 16 4i
The inner product of functions just defined is, in many ways, analogous to the inner product
defined for finite dimensional vectors at the beginning of this chapter. To see this, well verify
the following theorem, which is very similar to theorem 52.2 on page 524.
Theorem 52.11 (properties of the inner product)

Let h | i be an inner product as just defined above. Suppose and are two (possibly
complex) constants, and f , g , and h are functions on (a, b) . Then
1.
h f | g i = h g | f i ,
2.
h h | f + g i = h h | f i + h h | g i ,
3.
h f + g | h i = h f | h i + h g | h i ,
and
4.
h f | f i 0 with h f | f i = 0 if and only if f = 0 on (a, b) .
PROOF:
To Be Written (Someday) --- Take Notes in Class
Norms
Recall that the norm of any vector v in C N is related to the inner product of v with itself by
p
kvk = h v | v i .
In turn, for each inner product h | i , we define the corresponding norm of a function f by
p
kfk = h f | f i .
In general
kfk = h f | f i =
f (x) f (x)w(x) d x =
| f (x)|2 w(x) d x
Do note that, in a loose sense,

f is generally small over (a, b)
k f k is small .
!Example 52.7:
Sturm-Liouville
Let (a, b) = (0, 2) and w(x) = x 2 . Then
p
kfk = h f | f i =
s
Z
( f (x)) f (x)x 2 d x =
| f (x)|2 x 2 d x
In particular,
k5x + 6ik2 =
=
=
=
(5x + 6i) (5x + 6i)x 2 d x
(5x 6i)(5x + 6i)x 2 d x

25x 2 + 36 x 2 d x

25x 4 + 36x 2 d x
2
= 5x 5 + 12x 3 0
= 256 .
So
k5x + 6ik =
256 = 16
Orthogonality
Recall that any two vectors v and w in C N are orthogonal if and only if
hv | wi = 0
Analogously, we say that any pair of functions f and g is orthogonal (over the interval) (with
respect to the inner product, or with respect to the weight function) if and only if
h f | gi = 0
More generally, we will refer to any indexed set of nonzero functions

{ 1 , 2 , 3 , . . . }
as being orthogonal if and only if
h k | n i = 0
whenever
k 6= n
If, in addition, we have

kk k = 1
for each k
then we say the set is orthonormal. For our work, orthogonality will be important, but we wont
spend time or effort making the sets orthonormal.
!Example 52.8:
Consider the set

n

o
k
sin
x : k = 1, 2, 3, . . .
L
which is the set of sine functions (without the arbitrary constants) obtained as eigenfunctions
in example 52.2 on page 529. The interval is (0, L) . For the weight function, well use
w(x) = 1 . Observe that, if k and n are two different positive integers, then, using the
trigonometric identity
2 sin(A) sin(B) = cos(A B) cos(A + B) ,
we have
D
sin
k
x
L
Z

E
n

sin
x
=

L
k
x
L
So
sin
sin
sin
k
x
L
k
x
L

sin
sin
n
x
L
n
x
L
: k = 1, 2, 3, . . .
dx
dx = = 0
is an orthogonal set of functions on (0, L) with respect to the weight function w(x) = 1 .
Generalized Fourier Series

Now suppose {1 , 2 , 3 , . . . } is some orthogonal set of nonzero functions on (a, b) , and f
is some function that can be written as a (possibly infinite) linear combination of the k s ,
X
f (x) =
ck k (x)
for a < x < b .
k
To find each constant ck , first observe what happens when we take the inner product of both
sides of the above with one of the k s , say, 3 . Using the linearity of the inner product and the
orthogonallity of our functions, we get
D X
E

h 3 | f i = 3
c k k
k
c k h 3 | k i
ck
(
k3 k2
0
So
c3 =
if k = 3
if k 6= 3
h 3 | f i
k3 k2
= c3 k3 k2
Since there is nothing special about k = 3 , we clearly have

ck =
h k | f i
kk k2
for all k
Sturm-Liouville
More generally, whether or not f can be expressed as a linear combination of the k s ,

we define the generalized Fourier series for f (with respect to the given inner product and
orthogonal set {1 , . . . , } ) to be
X
G.F.S.[ f ]|x =
ck k (x)
k
where, for each k
h k | f i
ck =
kk k2
The ck s are called the corresponding generalized Fourier coefficients of f .2 Dont forget:
Z b
h k | f i =
k (x) f (x)w(x) d x
a
and
kk k =
|k (x)|2 w(x) d x
We will also refer to G.F.S.[ f ] as the expansion of f in terms of the k s . If the k s

just happen to be eigenfunctions from some Sturm-Liouville problem, we will even refer to
G.F.S.[ f ] as the eigenfunction expansion of f .
!Example 52.9:
From previous exercises, we know

n

o
k
{1 , 2 , 3 , . . . } = sin
x : k = 1, 2, 3, . . .
L
is an orthogonal set of functions on (0, L) with respect to the weight function w(x) = 1 .
(Recall that this is a set of eigenfunctions for the Sturm-Liouville problem from example 52.2
on page 529.) Using this set,
X
h k | f i
G.F.S.[ f ]|x =
ck k (x)
with ck =
2
kk k
becomes
G.F.S.[ f ]|x =
with
ck =
h k | f i
kk k2
G.F.S.[ f ]|x =
X
k=1
k=1
k
x
L
Z L
k
k
sin
x
f (x) dx
f (x) sin
x dx
L
L
0
=
Z L
Z L

k
k 2

x dx
sin2
x dx
sin
sin2
the above reduces to
ck sin
Since
k
ck sin
x
L
k
x
L
dx = =
with ck
L
2
2
=
L
k
f (x) sin
x
L
dx
2 Compare the formula for the generalized Fourier coefficients with formula in theorem (52.4) on page 526 for the
components of a vector with respect to any orthogonal basis. They are virtually the same!
In particular, suppose f (x) = x for 0 < x < L . Then the above formula for ck yields

Z L
2
k
ck =
x sin
x dx
L
L

#
Z L

2
k
+
cos
x dx

k 0
L
0
"

2
L #

2
2L
k
2
k
4
=
0+
cos
L +
sin
x
= (1)k
2
=
L
"
2x
k
cos
x
k
L
So, using the given interval, weight function and orthogonal set, the generalized Fourier series
for f (x) = x , is

X
k
k 4
(1)
sin
x
.
k=1
(In fact, this is the classic Fourier sine series for f (x) = x on (0, L) .)
Approximations and Completeness

Lets say we have an orthogonal set of functions {1 , 2 , 3 , . . . } and some function f on
(a, b) . Let
X
G.F.S.[ f ]|x =
ck k (x)
k=1
In practice, to avoid using the entire series, we may simply wish to approximate f using the N th
partial sum,
N
X
f (x)
ck k (x) .
k=1
The error in using this is

E N (x) = f (x)
N
X
ck k (x) ,
k=1
and the square of its norm

2
N

X

kE N k = f (x)
ck k (x)

2
k=1

2
N

b
X

ck k (x) w(x) d x
f (x)

a
k=1
give a convenient measure of how good this approximation

P N is. The above integral is sometimes
known as the (weighted) mean square error in using k=1
ck k (x) for f (x) on the interval
(a, b) .
We, of course, hope the error shrinks to zero (as measured by kE N k ) as N . If we
can be sure this happens no matter what piecewise continuous function f we start with, then we
say the orthogonal set {1 , 2 , 3 , . . . } is complete.
Sturm-Liouville
Now if {1 , 2 , 3 , . . . } is complete, then taking the limits above yield

k f G.F.S.[ f ]k = 0
Equivalently

2

X

f
(x)
(x)

w(x) d x = 0 .
k k

a
k=1
P
In practice, all of this usually means that the infinite series k ck k (x) converges to f (x) at
every x in (a, b) at which f is continuous. In any case, if the set {1 , 2 , 3 , . . . } is complete,
then we can view the corresponding generalized Fourier series for a function f as being the
same as that function, and can write
X
f (x) =
ck k (x)
for a < x < b
Z
where
ck =
h k | f i
kk k2
In other words, a complete orthogonal set of (nonzero) functions can be viewed as a basis for
the vector space of all functions of interest.
52.8
Sturm-Liouville Problems and Eigenfunction

Expansions
Suppose we have some Sturm-Liouville problem with differential equation

d
d
p(x)
+ q(x) = w
for a < x < b
dx
dx
and Sturm-Liouville appropriate boundary conditions. As usual, well let L be the operator
given by

d
d
L[] =
p(x)
+ q(x) .
dx
dx
Remember, by definition, p and q are real-valued functions and w is a positive function on

(a, b) .
Self-Adjointness and Some Immediate Results

Glance back at lemma 52.8 on page 5217. It tells us that, whenever u and v are sufficiently
differentiable functions satisfying the boundary conditions,
Z b
Z b
u L[v] d x =
L[u] v d x .
a
Using the standard inner product for functions on (a, b) ,

Z b
h f | gi =
f (x) g(x) d x
a
Sturm-Liouville Problems and Eigenfunction Expansions
we can write this equality as

h u | L[v] i = h L[u] | v i
which looks very similar to the equation characterizing self-adjointness for a matrix A in theorem
52.5 on page 528. Because of this we often say either that Sturm-Liouville problems are self
adjoint, or that the operator L is self adjoint.3 The terminology is important for communication,
but what is even more important is that functional analogs to the results obtained for self-adjoint
matrices also hold for here.
To derive two important results, let (1 , 1 ) and (2 , 2 ) be two solutions to our SturmLiouville problem (i.e., 1 and 2 are two eigenvalues, and 1 and 2 are corresponding
eigenfunctions). From a corollary to Greens formula (lemma 52.8, noted just above), we know
Z b
Z b
1 L[2 ] d x =
L[1 ] 2 d x .
a
But, from the differential equation in the problem, we also have

,
L[2 ] = 2 w2
L[1 ] = 1 w1
and thus,
L[1 ] = (1 w1 ) = 1 w1
(remember w is a positive function). So,
Z b
Z
1 L[2 ] d x =
a
1 (2 w2 ) d x =
L[1 ] 2 d x
(1 w1 )2 d x
1 2 w d x = 1
1 2 w d x
Since the integrals on both sides of the last equation are the same, we must have either
Z b
2 = 1
or
1 2 w d x = 0 .
(52.10)
a
Now, we did not necessarily assume the solutions were different. If they are the same,
(1 , 1 ) = (2 , 2 ) = (, )
and the above reduces to
or
w d x = 0
But, since w is a positive function and is necessarily nontrivial,

Z b
Z b
|(x)|2 w(x) d x > 0 (6= 0) .

w d x =
a
3 The correct terminology is that L is a self-adjoint operator on the vector space of twice-differentiable functions
satisfying the given homogeneous boundary conditions.
Sturm-Liouville
So we must have
=
which is only possible if is a real number. Thus
FACT: The eigenvalues are all real numbers.

Now suppose 1 and 2 are not the same. Then, since they are different real numbers, we
certainly do not have
2 = 1 .
Line (52.10) then tells us that we must have
Z b
1 2 w d x = 0
which we can also write as

h 1 | 2 i = 0
using the inner product with weight function w ,
Z b
h f | gi =
f (x) g(x)w(x) d x
Thus,
FACT: Eigenfunctions corresponding to different eigenvalues are orthogonal with

respect to the inner product with weight function w(x) .
By the way, since the eigenvalues are real, it is fairly easy to show that the real part and the
imaginary part of each eigenfunction is also an eigenfunction. From this it follows that we can
always choose real-valued functions as our basis for each eigenspace.
What all the above means, at least in part, is that we will be constructing generalized Fourier
series using eigenfunctions from Sturm-Liouville problems, and that the inner product used will
be based on the weight function w(x) from the differential equation

d
d
p(x)
+ q(x) = w
for a < x < b .
dx
dx
in the Sturm-Liouville problem. Accordingly, we may refer to w(x) and the corresponding inner
product on (a, b) as the natural weight function and the natural inner product corresponding to
the given Sturm-Liouville problem.
Other Results Concerning Eigenvalues

Using the Rayleigh quotient it can be shown that there is a smallest eigenvalue 0 for each
Sturm-Liouville problem normally encountered in practice. The rest of the eigenvalues are
larger. Unfortunately, verifying this is beyond our ability (unless the Sturm-Liouville problem
is sufficiently simple). Another fact you will just have to accept without proof is that, for each
Sturm-Liouville problem normally encountered in practice, the eigenvalues form an infinite
increasing sequence
0 < 1 < 2 < 3 <
with
lim k = .
Sturm-Liouville Problems and Eigenfunction Expansions
The Eigenfunctions
Let us assume that
E = { 0 , 1 , 2 , 3 , . . . }
is the set of all distinct eigenvalues for our Sturm-Liouville problem (indexed so that 0 is the
smallest and k < k+1 in general). Remember, each eigenvalue will be either a simple or a
double eigenvalue. Next, choose a set of eigenfunctions
B = { 0 , 1 , 2 , 3 , . . . }
as follows:
1.
For each simple eigenvalue, choose exactly one corresponding (real-valued) eigenfunction for B .
2.
For each double eigenvalue, choose exactly one orthogonal pair of corresponding (realvalued) eigenfunctions for B .
Remember, this set of functions will be orthogonal with respect to the weight function w from
the differential equation in the Sturm-Liouville problem. (Note: Each k is an eigenfunction
corresponding to eigenvalue k only if all the eigenvalues are simple.)
Now let f be a function on (a, b) . Since B is an orthogonal set, we can construct the
corresponding generalized Fourier series for f
G.F.S.[ f (x)] =
with
ck
h k | f i
=
=
kk k2
ck k (x)
k=0
k (x) f (x)w(x) dx
b
a
.
2
|k (x)| w(x) dx
The obvious question to now ask is Is this orthogonal set of eigenfunctions complete? That is,
can we assume
X
f =
c k k
on (a, b) ?
k=0
The answer is yes, at least for the Sturm-Liouville problems normally encountered in practice.
4
But
P you will have to trust me on this. There are some issues regarding the convergence of
k=0 ck k (x) when x is a point at which f is discontinuous or when x is an endpoint of
(a, b) and f does not satisfy the same boundary conditions as in the Sturm-Liouville problem,
but we will gloss over those issues for now.
Finally (assuming reasonable assumptions concerning the functions in the differential
equation), it can be shown that the graphs of the eigenfunctions corresponding to higher values
of the eigenvalues wiggle more than those corresponding to the lower-valued eigenvalues. To
4 One approach to proving this is to consider the problem of minimizing the Rayleigh quotient over the vector space
U of functions orthogonal to the vector space of all the eigenfunctions. You can then (I think) show that, if the
orthogonal set of eigenfunctions in not orthogonal, then U is not empty, and this minimization problem has a
solution (, ) and that this solution must also satisfy the Sturm-Liouville problem. Hence is an eigenfunction
in U , contrary to the definition of U . So U must be empty and the given set of eigenfunctions must be complete.
Sturm-Liouville
be precise, eigenfunctions corresponding to higher values of the eigenvalues must cross the X
axis (i.e., be zero) more oftern that do those corresponding to the lower-valued eigenvalues. To
see this (sort of), suppose 0 is never zero on (a, b) (so, it hardly wiggles this is typically the
case with 0 ). So 0 is either always positive or always negative on (a, b) . Since 0 will
also be an eigenfunction, we can assume weve chosen 0 to always be positive on the interval.
Now let k be an eigenfunction corresponding to another eigenvalue. If it, too, is never zero on
(a, b) , then, as with 0 , we can assume weve chosen k to always be positive on (a, b) . But
then,
Z
b
h 0 | k i =
(x)k (x)w(x) d x > 0

|0
{z
}
>0
contrary to the known orthogonality of eigenfunctions corresponding to different eigenvalues.

Thus, each k other than 0 must be zero at least at one point in (a, b) . This idea can be
extended, showing that eigenfunctions corresponding to high-valued eigenvalues cross the X
axis more often than do those corresponding to lower-valued eigenvalues, but requires developing
much more differential equation theory than we have time (or patience) for.
52.9
The Main Results Summarized (Sort of)
A Mega-Theorem
We have just gone through a general discussion of the general things that can (often) be derived
regarding the solutions to Sturm-Liouville problems. Precisely what can be proven depends
somewhat on the problem. Here is one standard theorem that can be found (usually unproven)
in many texts on partial differential equations and mathematical physics. It concerns the regular
Sturm-Liouville problems (see page 5220).
Theorem 52.12 (Mega-Theorem on Regular Sturm-Liouville problems)

Consider a regular Sturm-Liouville problem5 with differential equation

d
d
p(x)
+ q(x) = w(x)
for a < x < b
dx
dx
and homogeneous regular boundary conditions at the endpoints of the finite interval (a, b) .
Then, all of the following hold:
1.
All the eigenvalues are real.
2.
The eigenvalues form an ordered sequence

0 < 1 < 2 < 3 <
with a smallest eigenvalue (usually denoted by 0 or 1 ) and no largest eigenvalue. In

fact,
k
as k .
5 i.e., p , q and w are real valued and continuous on the closed interval [a, b] , with p being differentiable on
(a, b) , and both p and w being positive on the closed interval [a, b] .
The Main Results Summarized (Sort of)
3.
All the eigenvalues are simple.
4.
If, for each eigenvalue k , we choose a corresponding eigenfunction k , then the set
B = { 0 , 1 , 2 , 3 , . . . }
is a complete, orthonormal set of functions relative to the inner product

Z b
hu | vi =
u(x) v(x)w(x) d x .
a
Moreover, if f is any piecewise smooth function on (a, b) , then the corresponding

generalized Fourier series of f ,
G.F.S.[ f ] =
ck k (x)
k=0
with ck =
h k | f i
kk k2
converges for each x in (a, b) , and it converges to

(a)
f (x) if f is continuous at x .
(b)
the midpoint of the jump in f ,

lim
0+
1
[ f (x + ) + f (x )]
2
if f is discontinuous at x .
5.
0 , the eigenfunction in B corresponding to the smallest eigenvalue, is never zero on

(a, b) . Moreover, for each k , k+1 has exactly one more zero in (a, b) than does k .
6.
Each eigenvalue is related to any corresponding eigenfunction via the Rayleigh

quotient
b
Z b h 2
i

d
d
q ||2 d x
p
+
p
dx
dx a
a
=
.
kk2
Similar mega-theorems can be proven for other Sturm-Liouville problems. The main difference occurs when we have periodic boundary conditions. Then most of the eigenvalues are
double eigenvalues, and our complete set of eigenfunctions looks like
{ . . . , k , k , . . . }
where {k , k } is an orthogonal pair of eigenfunctions corresponding to eigenvalue k . (Typically, though, the smallest eigenvalue is still simple.)
Sturm-Liouville
Additional Exercises
52.1. Let 1 = 3 , 2 = 2 ,
" #
2
b1 =
3
and
b2 =
"
3
2
and assume (1 , b1 ) and (2 , b2 ) are eigenpairs for a matrix A .

a. Verify that {b1 , b2 } is an orthogonal set.

b. Compute b1 and b2 .
c. Express each of the following vectors in terms of b1 and b2 using the formulas in
theorem 52.4 on page 526.
" #
" #
" #
4
12
0
ii. v =
iii. w =
i. u =
6
5
1
d. Compute the following using the vectors from the previous part. Leave your answers
in terms of b1 and b2 .
i. Au
ii. Av
52.2. Let 1 = 2 , 2 = 4 , 3 = 6 ,

1
2

1
2
b = 2
,
b = 1
1
0
iii. Aw
and
1

b3 = 2
5
and assume (1 , b1 ) , (2 , b2 ) and (3 , b3 ) are eigenpairs for a matrix A .
a. Verify that {b1 , b2 , b3 } is an orthogonal set.

b. Compute b1 , b2 and b3 .
c. Express each of the following vectors in terms of b1 , b2 and b3 using the formulas
in theorem 52.4 on page 526.

4
1
0

ii. v = 2
iii. w = 1
i. u = 7
22
3
0
in terms of b1 , b2 and b3 .
i. Au
ii. Av
52.3. Let 1 = 1 , 2 = 0 , 3 = 1 ,

1
1

1
2
b = 1
,
b = 2
1
1
iii. Aw
and

1

3
b = 0
1
and assume (1 , b1 ) , (2 , b2 ) and (3 , b3 ) are eigenpairs for a matrix A .

a. Verify that {b1 , b2 , b3 } is an orthogonal set.

b. Compute b1 , b2 and b3 .
c. Express each of the following vectors in terms of b1 , b2 and b3 using the formulas
in theorem 52.4 on page 526.

5
0
1

ii. v = 3
iii. w = 0
i. u = 8
3
6
0
in terms of b1 , b2 and b3 .
ii. Av
i. Au
iii. Aw
52.4. Verify each of the claims in theorem 52.2 on page 524.

52.5. Finish verifying the claims in theorem 52.5 on page 528. In particular show that, if B
and A are N N matrices such that
h v | Bu i = h Av | u i
for every v, u in R N
then B = A .
52.6. Find the general solutions for each of the following corresponding to each real value .
Be sure the state the values of for which each general solution is valid. (Note: Some
of these are Euler equations see chapter 18.)
a. + 4 =
b. + 2 =
c. x 2 + x =
for 0 < x
d. x 2 + 3x =
for 0 < x
52.7. Rewrite each of the following equations in Sturm-Liouville form.

a. + 4 + = 0
b. + 2 =
c. x 2 + x =
for 0 < x
d. x 2 + 3x =
for 0 < x
e. 2x =
(Hermites Equation)

for 1 < x < 1
f. 1 x 2 x =
g. x + (1 x) =
for 0 < x
(Chebyshevs Equation)
(Laguerres Equation)
52.8. Solve the Sturm-Liouville problems consisting of the differential equation

+ = 0
for 0 < x < L
(where L is some finite positive length) and each of the following sets of boundary
conditons:
Sturm-Liouville
a. (0) = 0 and (L) = 0

b. (0) = 0 and (L) = 0
52.9. Solve the following Sturm-Liouville problems:
a. =
(2)
with (0) = (2) and (0) =
for 0 < x < 2
b. x 2 + x =
for 1 < x < e
with (1) = 0 and (e ) = 0
c. x 2 +3x =
for 1 < x < e
with (1) = 0 and (e ) = 0
52.10. (Preliminary Greens formula) Assume (a, b) is some finite interval, and let L be the
operator given by

d
d
p(x)
+ q(x) .
L[] =
dx
dx
where p and q are any suitably smooth and integrable functions on (a, b) . Using
integration by parts, show that
b
Z b
Z b
Z b
d f dg
dg
f L[g] d x = p(x) f (x)
p
dx +
q f g dx
dx
dx dx
whenever f and g are suitably smooth and integrable functions on (a, b) .
52.11. Let L be the differential operator

d
L[] =
dx
d
p(x)
dx
+ q(x)
where p and q are continuous functions on the closed interval [a, b] .

a. Verify that the boundary conditions
(a) = 0
and
(b) = 0
is a Sturm-Liouville appropriate set of boundary conditions.

b. Show that any pair of homogeneous regular boundary conditions at a and b is SturmLiouville appropriate.
c. Suppose our boundary conditions are periodic; that is,
(a) = (b)
and
(a) = (b) .
What must p satisfy for these boundary conditions to be Sturm-Liouville appropriate?

52.12. Compute the following, assuming the interval is (0, 3) and the weight function is
w(x) = 1 .

b. x 2 9 + i8x
a. h x | sin(2 x) i

c. 9 + i8x x 2
d. ei 2 x x
e. kxk
f. k9 + i8xk

h. ei 2 x
g. ksin(2 x)k
52.13. Compute the following, assuming the interval is (0, 1) and the weight function is
w(x) = x .

b. x 2 9 + i8x
a. h x | sin(2 x) i

c. 9 + i8x x 2
d. ei 2 x x
e. kxk
f. k4 + i8xk

h. ei 2 x
g. ksin(2 x)k
52.14. Determine all the values for so that

n
o
2
2
ei 2 x , ei 2x
is an orthogonal set on (0, 1) with weight function w(x) = x .
52.15. Let L be a positive value. Verify that each of the following sets of functions is orthogonal
on (0, L) with respect to the weight function w(x) = 1 .

o
n

k
a. cos
x : k = 1, 2, 3, . . .
b. ei 2k x/L : k = 0, 1, 2, 3, . . .
L
52.16. Consider

o
n
2
3
1
x , cos
x , cos
x , ...
cos
{1 , 2 , 3 , . . . } =
which, in the exercise above, you showed is an orthogonal set of functions on (0, L)
with respect to the weight function w(x) = 1 . Using this interval, weight function and
orthogonal set, do the following:
a. Show that, in this case, the generalized Fourier series
G.F.S.[ f ] =
ck k (x)
k=1
with ck =
h k | f i
kk k2
is given by
G.F.S.[ f ] =
X
k=1

k
ck cos
x
L
with ck
2
=
L
k
x
L
dx
if 0 < x <
L/
2
if
f (x) cos
b. Find the generalized Fourier series for

ii. f (x) =
i. f (x) = x
iii. f (x) =
x
Lx
if 0 < x <
L/
2
if
L/
2
<x<L
L/
2
<x<L
c. Show that this set of cosines is not a complete set by showing

Z L

X
2
k
k
f (x) 6=
ck cos
x
with ck =
f (x) cos
x dx
k=1
when f is the constant function f (x) = 1 on (0, L) .
Sturm-Liouville
52.17. In several exercises above, you considered either the Sturm-Liouville problem
+ = 0
for 0 < x < L
with (0) = 0 and (L) = 0
or, at least, the above differential equation. You may use the results from those exercises
to help answer the ones below:
a. What is the natural weight function w(x) and inner product h f | g i corresponding
to this Sturm-Liouville problem?
b. We know that

2
1 2
2 2
k
{0 , 1 , 2 , . . . , k , . . .} = 0,
,
, ...,
, ...
L
is the complete set of eigenvalues for this Sturm-Liouville problem. Now write out a
corresponding orthogonal set of eigenfunctions (without arbitrary constants),
{0 , 1 , 2 , . . . , k , . . .}
c. Compute the norm of each of your k s from your answer to the last part.
d. To what does each coefficient in the generalized Fourier series
G.F.S.[ f ] =
ck k (x)
k=0
with ck =
h k | f i
kk k2
reduce to using the eigenfunctions and inner products from previous parts of this
exercise?
e. Using the results from the last part, find the generalized Fourier series for the following:
i. f (x) = 1
(
iii. f (x) =
ii. f (x) = x
1
if 0 < x <
if
L/
2
L/
2
<x<L
x 2 + x =
for 1 < x < e
with (1) = 0 and (e ) = 0
to answer the ones below:
b. We know that
{1 , 2 , 2 , . . . , k , . . .} =
1, 4, 9, . . . , k 2 , . . .
{1 , 2 , 3 , . . . , k , . . .}
c. Compute the norm of each of your k s from your answer to the last part. (A simple
substitution may help.)
G.F.S.[ f ] =
with ck =
ck k (x)
k=0
h k | f i
kk k2
exercise?
iii. f (x) = x
ii. f (x) = ln |x|
i. f (x) = 1
x 2 + 3x =
for 1 < x < e
with (1) = 0 and (e ) = 0
to answer the ones below:
b. We know that
{1 , 2 , 2 , . . . , k , . . .} =
2, 5, 10, . . . , k 2 + 1, . . .
{1 , 2 , 3 , . . . , k , . . .}
c. Compute the norm of each of your k s from your answer to the last part. (A simple
substitution may help.)
G.F.S.[ f ] =
ck k (x)
k=0
with ck =
h k | f i
kk k2
exercise?
i. f (x) = 1
ii. f (x) = x
Sturm-Liouville
52.20. Consider the Sturm-Liouville problem

1 x 2 x + = 0
for 1 < x < 1
with
|(1)| <
|(1)| < .
and
This is not a regular Sturm-Liouville problem, but it can be shown that the results
claimed in mega-theorem 52.12 hold for this Sturm-Liouville problem. In particular,
it can be shown that the Chebyshev polynomials (type I),
0 (x) = 1 ,
1 (x) = x
, 2 (x) = 2x 2 1 ,
3 (x) = 4x 3 3x
...
make up a complete, orthogonal set of eigenfunctions for this Sturm-Liouville problem.

Assuming all this, do the following:
a. Using the above differential equation and formulas for 0 , 1 , 2 and 3 , find the
corresponding eigenvalues 0 , 1 , 2 and 3 .
b. Rewrite the above differential equation in Sturm-Liouville form.
c. Based on the above, what is the appropriate inner product h f | g i when using the
type I Chebyshev polynomials?
d. Using the inner product from the previous part, and the above k s , compute the
following norms6 :
i. k0 k
ii. k1 k
iii. k2 k
iv. k3 k
e. According to our theory, any reasonable function f (x) on (1, 1) can be approximated by the N th partial sum
N
X
ck k (x)
k=0
where
ck k (x)
k=0
is the generalized Fourier series for f (x) using the type I Chebyshev polynomials.
For the following, let
p
f (x) = x 1 x 2
and find the above mentioned N th partial sum when
i. N = 0
ii. N = 1
6 the substitution x = sin( ) may be useful
iii. N = 2
iv. N = 3
Sturm-Liouville
Some Answers to Some of the Exercises

WARNING! Most of the following answers were prepared hastily and late at night. They
have not been properly proofread! Errors are likely!

1b. b1 = 13 and b2 = 13
1c i. u = 2b1
1c ii. v = 3b1 2b2
3
2
1c iii. w = b1 + b2
13
13
1d i. Au = 6b1
1d ii. Av = 9b1 + 4b2
9
4
1d iii. Aw = b1 b2
1 13 2 13

2b. b = 6 , b = 5 and b3 = 30
2c i. u = 2b1 + 3b2 4b3
4
1
2c ii. v = b1 b3
3
1
3
1
5
2c iii. w = b1 b2 +
1 3
b
15
2d i. Au = 4b1 + 12b2 24b3

8
2d ii. Av = b1 2b3
3
2d iii. Aw = b1 b2 + b3
3 5 5

3b. b1 = 3 , b2 = 6 and b3 = 2
3c i. u = 2b1 3b2 + 4b3
3c ii. v = b1 + 2b2 + 3b3
1
1
1
3c iii. w = b1 + b2 b3
3
3d i. Au = 2b1 + 4b3
3d ii. Av = b1 + 3b3
1
1
3d iii. Aw = b1 b3
3
2
p
6a. If < 4 , (x) = aex + bex with = (4
+ ) ; If = 4 , 4 (x) = ax + b ;
If > 4 , (x) = a cos(x) + b sin(x) with = 4 +
6b.
If < 1 , (x) = ae(1+)x + be(1)x with = 1 ; If = 1 , 1 (x) =
aex + bxex ; If > 1 , (x) = aex cos(x) +bex sin(x) with = 1
6c. If < 0 , (x) = ax + bx with = ; If
= 0 , 0 (x) = a ln |x| + b ;
If > 0 , (x) = a cos( ln |x|) + b sin( ln |x|) with
1
1
1
6d. If < 1 , (x) = ax
+ bx
with = 1 ; If = 1 , 1 (x)
= ax ln |x| +
1
1
1
bx ; If > 1 , (x) = ax cos( ln |x|) + bx sin( ln |x|) with = 1
7a. +
4 =
7b.
7c.
7d.
7e.
7f.
d
d
e2x
= e2x
dx
dx

d
d
1
x
=
x
dx dx
d
d
x3
= x
dx dx
2 d
2
d
ex
= ex
dx
dx

p
d
d
1
2
1x
=
dx
dx
1 x2
7g.
d
dx
xex
d
dx
= ex

k
8a. (0 , 0 (x)) = (0, c0 ) and (k , k (x)) =
, ck cos
x
for k = 1, 2, 3, . . .
L

(2k + 1) 2
(2k + 1)
8b. (k , k (x)) =
, ck sin
x
for k = 1, 2, 3, . . .
2L
2L

9a. (0 , 0 (x)) = (0, c0 ) and (k , k (x))
= k 2 , ck,1 cos(kx) + ck,2 sin(kx) for k = 1, 2, 3, . . .

9b. (k , k (x)) = k 2 , ck sin(k ln |x|) for k =
1, 2, 3, . . .
9c. (k , k (x)) = k 2 + 1 , ck x 1 sin(k ln |x|) for k = 1, 2, 3, . . .
11c. p(b) = p(a)
3
12a.
2
12b. 81 + 162i
12c. 81 162i
3
12d.
i
2
12e.
3
12f. r819

k 2
L
3
2
12g.
12h.
3
1
13a.
13b.
13c.
13d.
13e.
9
4
9
4
8
5
8
i
5
1
1
+
i
2
2 2
1
2
+i
13f. 2 r6
13g.
1
2
1+
13h.
1
2
14. can be any integer except 1 .
X
k
2L
k
16b i.
(1)
1
cos
x
2
L
(k )
16b ii.
k=1
2
k
k=1
sin
k
2

k
cos
x
L
RL
h
|
i
17a. w(x)
=
1
,
f
g
=
n

0 ( f (x)) g(x)
d x
o
1
2
k
17b. 1, cos
x , cos
x , . . . , cos
x , ...
L
L
L

k

L
17c. k1k = L , cos
x =
L
2 Z
Z L

L
1
2
k
17d. c0 =
f (x) d x , ck =
f (x) cos
x d x for k > 0
L
17e i. 1
17e ii.
L
2
X
2L
k=1
(k )2
k
(1)k 1 cos
x
L
17e iii.
1
2
X
2
k=1
Sturm-Liouville
sin
1
x
k
2

k
cos
x
L
R e
18a. w(x) = , h f | g i = 1 ( f (x)) g(x) d x

x
18b. {sin(1 ln |x|) , sin(2 ln |x|) , sin(3 ln |x|) , . . . , sin(k ln |x|) , . . .}
18c. ksin(k ln |x|)k =

2
Z e
2
1
18d. ck =
f (x) sin(k ln |x|) d x
X

2
18e i.
1 (1)k sin(k ln |x|)
18e ii.
18e iii.
k=1
2
k
(1)k+1 sin(k ln |x|)
k=1
2k
( 2 + k 2 )

1 (1)k e sin(k ln |x|)
k=1
R e
h
|
i
19a. w(x)
=
x
,
f
g
=
1 ( f (x)) g(x)x d x
1

19b. x sin(1 ln |x|) , x 1 sin(2 ln |x|) , x 1 sin(3 ln |x|) , . . . , x 1 sin(k ln |x|) , . . .
q

19c. x 1 sin(k ln |x|) =

2
Z e
2
f (x) sin(k ln |x|) d x
19d. ck =
19e i.
19e ii.
2k

1 + k2
k=1
X
k=1

sin(k ln |x|)
1 (1)k e
2k

( + 1)2 + k 2
1 (1)k e(+1)
sin(k ln |x|)
20a. 0 =
0 , 1 = 1 ,2 = 4 , 3 = 9
p
d
1
d
20b.
1 x2
=
dx
dx
1 x2
Z 1
1/2
20c. h f | g i =
( f (x)) g(x) 1 x 2
dx
1
20d i. q
20d ii.
r2
20d iii.
3
q 8
20d iv.
20e i. 0
4
20e ii.
x
20e iii.
20e iv.
3
4
x
3
4
4
x x3
3
5

SL

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SL

Uploaded by

Copyright:

Available Formats

52

Chapter & Page: 522

Linear Algebraic Antecedents

Recollections From (Real) Linear Algebra

= |v1 |2 + |v2 |2 + + |v N |2 = kvk2

Linear Algebraic Antecedents

Chapter & Page: 523

All the eigenvalues of A are real.

There is an orthogonal basis for R N consisting of eigenvectors for A .

Linear Algebra with Complex Components

Vectors and Matrices with Complex-Valued Components

The natural norm (or length) of v is still given by

Chapter & Page: 524

Theorem 52.2 (properties of the inner product)

Linear Algebraic Antecedents

Chapter & Page: 525

A self-adjoint matrix is automatically square.

All the eigenvalues of A are real.

There is an orthogonal basis for C N consisting of eigenvectors for A .

Chapter & Page: 526

and can reduce the computation of Av to

= v1 Ab1 + v2 Ab2 + v3 Ab3 + + v N Ab N

Finding Vector Components

Remember, this set of bk s is orthogonal. That means

Taking this into account, we continue our computation of b1 v :

Solving for v1 gives us

Repeating this using each bk yields the next theorem:

be an orthogonal basis for C N . Then, for any vector v in R N ,

Linear Algebraic Antecedents

Chapter & Page: 527

About Normalizing a Basis

Moreover, if bk is an eigenvector for a matrix A with corresponding eigenvalue k , so is nk .

Inner Products and Adjoints

Chapter & Page: 528

So, for any N N matrix A and any two vectors v and u in C N ,

Consequently, if A is self adjoint (i.e., A = A ), then

Theorem 52.5 (characterization of adjointness)

Moreover, both of the following hold:

for every v and u in C N

for every v and u in C N

A is self adjoint if and only if

Boundary-Value Problems with Parameters

Chapter & Page: 529

Boundary-Value Problems with Parameters

where A , B and C are known functions.

with boundary conditions

where L is some positive number (think of it as a given length).

Chapter & Page: 5210

< 0 : In this case, > 0 . For convenience, let =

where c1 and c2 are arbitrary constants, and, as already stated, =

Applying the first boundary condition:

which tells us that

Combining this with the boundary condition at x = L yields

Now e L > e0 = 1 and e L < e0 = 1 . So

and the only way we can have

is for c1 = 0 . Thus, the only solution to this problem with < 0 is

the trivial solution (which we dont really care about).

Integrating this twice yields

Applying the first boundary condition gives us

Combined with the boundary condition at x = L , we then have

again, just the trivial solution.

Boundary-Value Problems with Parameters

Chapter & Page: 5211

where, again, c1 and c2 are arbitrary constants, and, as already stated, =