STURM-LIOUVILLE BOUNDARY VALUE PROBLEMS

Throughout, we let [a, b] be a bounded interval in R. C

2
([a, b]) denotes
the space of functions with derivatives of second order continuous up to
the endpoints. C
2
c
([a, b]) is the subspace of functions that vanish near the
endpoints.
Let L denote a second order dierential operator of the form
Lu(x) = r(x)u

(x) + r

(x)u

(x) + q(x)u(x)
=
d
dx
_
r(x)
du
dx
_
+ q(x)u(x) .
(1)
We assume that r C
1
([a, b]) and q C
0
([a, b]) are real, and that r(x) c
for some c > 0.
The operator L is the most general second order real ODE which is formally
self-adjoint on L
2
(dx), in that
_
b
a
(Lu) v dx =
_
b
a
u(Lv) dx u, v C
2
c
([a, b]) .
The condition u, v C
2
c
([a, b]) implies that when integrating by parts the
boundary terms vanish. Since L has real coecients, conjugating v or not
does not aect the denition.
For general u, v C
2
([a, b]),
(2)
_
b
a
(Lu)v u(Lv) dx = r
_
u

v uv

b
a
and we need to impose rst order conditions on u, v at the endpoints to
make the right hand side vanish.
A boundary condition B is an expression of the form
Bu = u(a) + u(b) + u

(a) + u

(b)
for real constants , , , . We will impose two conditions B
1
u = 0 and
B
2
u = 0 where B
1
and B
2
are independent (i.e. the corresponding vectors
(, , , ) are independent), chosen to guarantee that the right hand side
of (2) vanish.
Denition 1. The boundary conditions B
1
, B
2
are self-adjoint for L if, for
all u, v C
2
([a, b]) which satisfy B
1
u = B
2
u = B
1
v = B
2
v = 0, then
_
b
a
(Lu) v dx =
_
b
a
u(Lv) dx.
In other words, the vanishing of B
j
u and B
j
v implies the right-hand side of
(2) vanishes.
1
2 556 LECTURE NOTES
Dirichlet conditions: B
1
u = u(a) , B
2
u = u(b) .
Neumann conditions: B
1
u = u

(a) , B
2
u = u

(b) .
Robin conditions: B
1
u = u

(a)u(a) , B
2
u = u

(b)+u(b) , , > 0.
The above are separated boundary conditions, in that B
1
is a condition at a
and B
2
is a condition at b. Any pair of separated conditions is self-adjoint
for general L. The most common non-separated condition is
Periodic conditions: B
1
u = u(b) u(a) , B
2
u = u

(b) u

(a) .
These are self-adjoint for L if r(b) = r(a).
Another way to state self-adjointness is to consider the subspace
C
2
B
([a, b]) =
_
u C
2
([a, b]) : B
1
u = B
2
u = 0
_
.
Then (L, B
1
, B
2
) is self-adjoint provided that

Lu, v
_
=

u, Lv
_
for u, v
C
2
B
([a, b]), where

u, v
_
=
_
b
a
u v dx.
We next x a positive weight function (x) C
2
([a, b]), so (x) c > 0 for
x [a, b], and consider the Sturm-Liouville eigenvalue problem
Lu = u, B
1
u = B
2
u = 0 .
We say that the number is an eigenvalue if there is a nonzero solution
u C
2
([a, b]) to this equation, and call u an eigenfunction.
Lemma 2. Let (L, B
1
, B
2
, ) be a self-adjoint Sturm-Liouville system.
a. The associated eigenvalues are all real numbers.
b. Eigenfunctions associated to dierent eigenvalues are orthogonal in
the inner product

u, v
_

=
_
b
a
u(x) v(x) (x) dx.
c. The dimension of each eigenspace is at most 2; if the boundary con-
ditions are separated then it is exactly 1.
Proof. Note that an eigenfunction u is an eigenvector for the operator
1
L,
i.e.
1
Lu = u, and that

1
Lu, v
_

u,
1
Lv
_

, so that
1
L is self-
adjoint on the domain C
2
B
([a, b]) with respect to the inner product

,
_

.
The proof of a and b then follow exactly as for nite dimensional operators.
For c, we note that the space of solutions to (L)u = 0 is a 2-dimensional
subspace of C
2
([a, b]). If one imposes a separated condition B
1
u = 0, this
restricts the initial conditions
_
u(a), u

(a)
_
to a 1-dimensional space, hence
there is at most a 1-dimensional space of solutions to (L )u = 0 with
the boundary conditions imposed.
Theorem 3. Given a self-adjoint Sturm-Liouville system as above, there is
an orthonormal basis for the space L
2

([a, b]) consisting of eigenfunctions for

the Sturm-Liouville problem. The eigenvalues satisfy
n
.
556 LECTURE NOTES 3
Here, L
2

([a, b]) is the space of measurable u on [a, b] such that

u
2
L
2

=
_
b
a
|u(x)|
2
(x) dx < .
Since (x) is bounded above and below, this is the same space of functions
as L
2
([a, b]), but the norm and inner product

,
_

are dierent. The map

u
1
2
u is easily seen to be a unitary map of L
2

onto L
2
:
1
2
u
L
2 = u
L
2

.
In particular, {u
j
}

j=1
is an orthonormal basis for L
2

i {
1
2
u
j
}

j=1
is an
orthonormal basis for L
2
.
We will prove Theorem 3 in the case of separated boundary conditions for
simplicity, but it holds for general self-adjoint boundary conditions. We
start the proof by reducing to the case where = 1. Consider the operator

Lu =

1
2
L(

1
2
u) =
d
dx
_
r

du
dx
_
+ qu, q =

1
2
L(

1
2
) ,
which is formally self-adjoint on L
2
(dx), and Lu = u i

L(
1
2
u) =
1
2
u.
We also dene boundary conditions

B
j
(u) = B
j
(

1
2
u); it follows easily that

B
1
,

B
2
are self-adjoint for

L.
We conclude there is orthonormal basis for L
2
(dx) of eigenfunctions for

1
L satisfying B
j
i there is an orthonormal basis for L
2
(dx) consisting of
eigenfunctions for

L satisfying

B
j
, where the bases are related by multiplying
by
1
2
.
We thus assume = 1, and consider the eigenfunction problem Lu = u,
where Lu = (ru

+qu, and we impose self-adjoint conditions B

1
u = B
2
u =
0.
Lemma 4. If is an eigenvalue for Lu = u, then C for some constant
C depending on (L, B
1
, B
2
).
Proof. Integrating by parts we have

_
b
a
|u|
2
dx =
_
b
a
(Lu) udx =
_
b
a
r|u

|
2
+ q |u|
2
dx + r(x)u(x)u

(x)

x=b
x=a
.
For Dirichlet or Neumann conditions, the boundary terms vanish, and
(3)
_
b
a
|u|
2
dx
_
b
a
q |u|
2
dx
_
max
[a,b]
q
_
_
b
a
|u|
2
dx.
For Robin conditions u

(a) = u(a) , u

(b) = u(b), we get

_
b
a
|u|
2
dx =
_
b
a
r|u|
2
+ q |u|
2
dx r(a)|u(a)|
2
r(b)|u(b)|
2
.
4 556 LECTURE NOTES
In the physically realistic case , 0, then (3) still applies. If one or both
is negative, we need more work. We bound
max
[a,b]
u
2
min
[a,b]
u
2

_
b
a

d
dx
u
2

dx = 2
_
b
a
|u

| |u| dx

_
b
a
|u

|
2
+ 2
1
_
b
a
|u|
2
dx.
Taking small, for C

suciently large we have

||r(a)|u(a)|
2
+||r(b)|u(b)|
2

_
b
a
r|u

|
2
+ C

_
b
a
|u|
2
dx
and then

_
b
a
|u|
2
dx
_
b
a
(q + C

) |u|
2
dx
_
C

+ max
[a,b]
q
_
_
b
a
|u|
2
dx.

We replace L by L
0
with
0
= 1+C, which has the same eigenfunctions,
but with eigenvalues shifted by
0
. We may thus assume all eigenvalues
satisfy 1, and in particular
(4) Lu = 0 , B
1
u = B
2
u = 0 , implies u = 0 .
We produce eigenfunctions for L by nding eigenfunctions for the operator
L
1
, which we express as an integral kernel. Thus, we seek to express the
solution to
Lu = f , B
1
u = B
2
u = 0 , where f C([a, b]) ,
in the form
(5) u(x) =
_
b
a
G(x, y) f(y) dy .
The function G(x, y) is called Greens kernel for the problem (L, B
1
, B
2
). We
will apply variation of parameters: let u
1
and u
2
be nonzero real solutions
to
Lu
1
= Lu
2
= 0 , B
1
u
1
= 0 , B
2
u
2
= 0 .
Since B
1
and B
2
are separated, then u
1
and u
2
are determined up to a con-
stant multiple. Furthermore, u
1
and u
2
are linearly independent; otherwise
they are both solutions to (4). Thus
W(x) = det
_
u
1
(x) u
2
(x)
u

1
(x) u

2
(x)
_
= 0 .
By Abels theorem, rW

+ r

W = 0, so rW = constant. The variation of

parameters method states that if
_
u
1
(x) u
2
(x)
u

1
(x) u

2
(x)
__
c

1
(x)
c

2
(x)
_
=
_
0
f/r
_
556 LECTURE NOTES 5
then u = c
1
u
1
+ c
2
u
2
solves Lu = f. This has solution
c

1
=
u
2
f
rW
, c

2
=
u
1
f
rW
.
Since B
1
is a rst order operator at a, then B
1
u = c
1
(a)B
1
u
1
+ c
2
(a)B
1
u
2
,
and this vanishes if c
2
(a) = 0. Similarly B
2
u = 0 if c
1
(b) = 0. Thus we set
c
1
(x) =
_
b
x
u
2
(y)
rW
f(y) dy , c
2
(x) =
_
x
a
u
1
(y)
rW
f(y) dy .
Then (5) holds, where
G(x, y) =
_
(rW)
1
u
1
(x) u
2
(y) , y x,
(rW)
1
u
2
(x) u
1
(y) , x y .
Note that G(x, y) = G(y, x), and that G(x, y) is continuous on [a, b] [a, b].
Furthermore, G is real since u
1
and u
2
are. The kernel G is also a left inverse
for L, in that if v C
2
B
([a, b]), then
v(x) =
_
b
a
G(x, y) (Lv)(y) dy .
This follows by uniqueness of solutions (4). In particular, if one considers
the maps
L : C
2
B
([a, b]) C([a, b]) , G : C([a, b]) C
2
B
([a, b]) ,
then these maps are respectively the inverse of each other. It follows that
G is 1-1 on the space of continuous functions, but we need a stronger result
for the diagonalization argument.
Lemma 5. Suppose that f L
2
([a, b]), and that
_
b
a
G(x, y) f(y) dy = 0 for
all x. Then f(y) = 0 a.e.
Proof. We will show that
_
f(y)(y) dy = 0 for all C
2
c
([a, b]), and the
result follows by density of C
2
c
in L
2
. We then write
_
b
a
f(y) (y) dy =
_
b
a
f(y)
__
b
a
G(y, x) (L)(x) dx
_
dy
=
_
b
a
__
b
a
G(x, y) f(y) dy
_
(L)(x) dx
using Fubinis theorem.
The operator Tf(x) =
_
b
a
G(x, y)f(y) dy is then a self-adjoint, compact op-
erator on L
2
([a, b]), and 0 is not an eigenvalue of T. There thus exists an
orthonormal basis {u
j
}

j=1
for L
2
([a, b]), where
_
b
a
G(x, y) u
j
(y) dy =
j
u
j
(x) ,
j
= 0 .
Since the left hand side is continuous in x so is u
j
(x), and thus the left hand
side belongs to C
2
B
([a, b]), and so u
j
C
2
B
([a, b]). We then have
Lu
j
=
j
u
j
,
j
=
1
j
,
6 556 LECTURE NOTES
and by Lemma 2 each eigenspace is 1-dimensional (separated boundary con-
ditions). We have arranged that
j
1, which means we can order them
so that
j
in a decreasing manner. The eigenvalues for the original
problem are
j
+
0
, which still decrease to , but there may be nitely
many positive eigenvalues, depending on q and the boundary conditions.