STURM-LIOUVILLE PROBLEMS

ANTON ZETTL Mathematics Department, Northern Illinois University, DeKalb,

Illinois 60115.

Dedicated to the memory of John Barrett.

ABSTRACT. Regular and singular Sturm-Liouville problems (SLP) are studied

including the continuous and differentiable dependence of eigenvalues on the prob-
lem. Also initial value problems (IVP) are considered for the SL equation and for
general first order systems.

1 INTRODUCTION

The purpose of this paper is to survey some basic properties of Sturm-Liouville

problems (SLP). These problems originated in a series of papers by these two au-
thors in 1836-1837. Many thousands of papers, by Mathematicians and by others,
have been published on this topic since then. Yet, remarkably, this subject is an
intensely active field of research today. Dozens of papers are written on SLP ev-
ery year. We have made a serious effort to try to make this article useful to both
researchers as well as students.
There are major and deep theorems in the theory of nonlinear ordinary differen-
tial equations about the continuous and differentiable dependence of solutions on
parameters. When specialized to the linear case these results are often not best
possible and, even when they are, there frequently exist simpler proofs for the linear
case.
One of the main goals of this paper is to give the linear results under minimal
hypotheses and with elementary proofs whenever possible. Following this intro-
ductory section we take up initial value problems (IVP) for first order systems of
arbitrary dimension in Section 2; for scalar second order problems in Section 3. In
Section 4 we study eigenvalues and eigenfunctions of regular SLP; in Section 5 we
take up the singular case.
Readers not particularly interested in initial value problems are ad-
vised to start in Section 4 and refer back to Sections 2 and 3 as needed.

1
2

Each section, except the introduction, ends with comments. Here we make some
historical remarks, mention the names of some of the major contributors, give
references for further reading, state some problems, etc. These problems reflect the
interests and knowledge, or lack thereof, of the author. No effort has been made to
classify these problems by difficulty, some may be routine, others intractible.
Although the subject of Sturm-Liouville problems is over 160 years old a surpris-
ing number of the results surveyed here are of recent origin, some were published
within the last couple of years and a few are not in print at the time of this writing.
Instructions for downloading the SLEIGN2 package, including a FORTRAN code
to compute eigenvalues and eigenfunctions for regular and singular SLP, from the
internet are given in Subsectiom 5.3. This package consists of 8 files and a number
of associated papers, all on SLP. The “readme” file has more detailed instructions
and some additional information. The code “sleign2” comes with a user friendly
interface and can be used by novices and experts alike. It can also be used to ap-
proximate the continuous (essential) spectrum of singular problems when combined
with some theoretical results. These can be found in the associated papers which
can also be downloaded.
I take this opportunity to thank Don Hinton and the Mathematics Department of
the University of Tennessee, Knoxville, particularly Phil Schaefer and John Conway,
for the invitation to give the Barrett Lectures of 1996. This forced me to try to
organize my notes and my thoughts on SLP. Also Don Hinton has been one of my
mathematical heroes from the time we were graduate students together until now.
Special thanks also to my pre-Ph.D. teachers John Neuberger, William Mahavier,
and the late John Barrett for stimulating and encouraging my interest and curiosity
about the wonderful subject of Mathematics. And to my chief post-Ph.D. teacher
and friend, Norrie Everitt, with whom I have been privileged to collaborate for
some twenty five years now.
Also I wish to thank my colleague Qingkai Kong for finding and correcting my
errors. His criticisms have significantly improved the final version of the paper in
several respects. But I am solely responsible for any remaining errors.
Last but not least I thank my wife Sandra for her help with the database for the
references and, especially, for helping with the hardware and software problems that
arose during the typing of these notes. Also for her tolerance and understanding
during this and many other Mathematics projects.
The world of Mathematics is full of wonders and of mysteries, at least as much
so as the physical world.

2 FIRST ORDER SYSTEMS

2.1 Introduction.

This section is devoted to the study of basic properties of first order systems.
NOTATION. An open interval is denoted by (a, b) with −∞ ≤ a < b ≤ ∞; [a, b]
always denotes the compact interval with (finite) left endpoint a and (finite) right
endpoint b. Let R denote the reals, C the complex numbers, and

N = {1, 2, 3, . . . }, N0 = {0, 1, 2, . . . }, Z = {. . . , −2, −1, 0, 1, 2, . . . }.

 3

For any interval J of the real line, open, closed, half open, bounded or unbounded,
by L(J) we denote the linear manifold of complex valued Lebesgue measurable
functions y defined on J for which

Z b Z Z
|y(t)| dt ≡ |y(t)| dt ≡ |y| < ∞.
a J J

The notation Lloc (J) is used to denote the linear manifold of functions y sat-
isfying y ∈ L([α, β]) for all compact intervals [α, β] ⊆ J. If J = [a, b], then
Lloc (J) = L(J). Also, we denote by ACloc (J) the collection of complex-valued
functions y which are absolutely continuous on all compact intervals [α, β] ⊆ J.
For a given set S, Mn,m (S) denotes the set of n×m matrices with entries from S.
If n = m we write Mn (S); also if m = 1 we sometimes write S n for Mn,1 (S). The
norm of a constant matrix as well as the norm of a matrix function P is denoted
by |P |. This may be taken as

X
|P | = |pij |.

2.2 Existence and uniqueness of solutions.

DEFINITION 2.1 (Solution). Let J be any interval, open, closed, half open,
bounded or unbounded; let n, m ∈ N, let P : J → Mn (C), F : J → Mn,m (C).
By a solution of the equation Y 0 = P Y + F on J we mean a function Y from J
into Mn,m (C) which is absolutely continuous on all compact subintervals of J and
satisfies the equation a.e. on J. A matrix function is absolutely continuous if each
of its components is absolutely continuous.

THEOREM 2.2. Let J be any interval, open, closed, half open, bounded or
unbounded; let n, m ∈ N. If

(2.1) P ∈ Mn (Lloc (J))

and

(2.2) F ∈ Mn,m (Lloc (J))

then every initial value problem (IVP)

(2.3) Y 0 = PY + F
4

(2.4) Y (u) = C, u ∈ J, C ∈ Mn,m (C)

has a unique solution defined on all of J. Furthermore, if C, P , F , are all real

valued, then the solution is also real valued.
PROOF: We give two proofs of this important theorem; the second one is the stan-
dard succesive approximations proof. As we will see later the analytic dependence
of solutions on the spectral parameter λ follows more readily from the second proof
than the first.
For both proofs we note that if Y is a solution of the IVP (2.3), (2.4) then an
integration yields

Z t
(2.5) Y (t) = C + (P Y + F ), t ∈ J.
u

Conversely, every solution of the integral equation (2.5) is also a solution of the
IVP (2.3) (2.4).
Choose c in J, c 6= u. We show that (2.3), (2.4 ) has a unique solution on [u, c]
if c > u and on [c, u] if c < u . Assume c > u. Let

B = {Y : [u, c] → Mn,m (C), Y ∈ C[u, c]}.

Following Bielecki [12] we define the norm of any function Y ∈ B to be

Rt
(2.6) kY k = sup {e−K u
|P (s)|ds
|Y (t)|, t ∈ [u, c]},

where K is a fixed positive constant K > 1. It is easy to see that with this norm B
is a Banach space. Let the operator T : B → B be defined by

Z t
(2.7) (T Y )(t) = C + (P Y + F )(s) ds, t ∈ [u, c], Y ∈ B.
u

Then for Y, Z ∈ B we have

Z t
|(T Y )(t) − (T Z)(t)| ≤ |P (s)||Y (s) − Z(s)|ds
u

and hence

Rt
Z t Rt
e−K u
|P (s)|ds
|(T Y )(t) − (T Z)(t)| ≤ kY − Zk |P (s)|e−K s
|P (r)|dr
ds
u
1
≤ kY − Zk.
K
 5

Therefore

1
kT Y − T Zk ≤ kY − Zk.
K

¿From the contraction mapping principle in Banach space it follows that the
map T has a unique fixed point and therefore the IVP (2.3), (2.4) has a unique
solution on [u, c]. The proof for the case c < u is similar; in this case the norm of
B is modified to

Rt
|P (s)|
kY k = sup{eK u ds |Y (t)|, t ∈ [c, u]}.

Since there is a unique solution on every compact subinterval [u, c] and [c, u] for c ∈
J, c 6= u it follows that there is a unique solution on J. To establish the furthermore
part take the Banach space of real-valued functions and proceed similarly. This
completes the first proof.
For the second proof we constuct a solution of (2.5) by successive approximations.
Define

Z t
(2.8) Y0 (t) = C, Yn+1 (t) = C + (P Yn + F ), t ∈ J, n = 0, 1, 2, . . .
u

Then Yn is a continuous function on J for each n ∈ N0 . We show that the sequence

{Yn : n ∈ N0 } converges to a function Y uniformly on each compact subinterval of
J and that the limit function Y is the unique solution of the integral equation and
hence also of the IVP. Choose b ∈ J, b > u and define

Z t
(2.9) p(t) = |P (s)| ds, t ∈ J; Bn (t) = maxu≤s≤t |Yn+1 (s) − Yn (s)|, u ≤ t ≤ b.
u

Then

Z t
(2.10) Yn+1 (t) − Yn (t) = P (s)[Yn (s) − Yn−1 (s)] ds, t ∈ J, n ∈ N.
u

From this we get

Z t
(2.11) |Y2 (t) − Y1 (t)| ≤ B0 (t) |P (s)| ds = B0 (t) p(t) ≤ B0 (b) p(b), u ≤ t ≤ b.
u
6

Z t Z t
|Y3 (t) − Y2 (t)| ≤ |P (s)| |Y2 (s) − Y1 (s)| ds ≤ |P (s)| B0 (s) p(s) ds
u u
Z t
p2 (t)
≤ B0 (t) |P (s)| p(s) ds ≤ B0 (b)
u 2!
2
p (b)
≤ B0 (b) , u ≤ t ≤ b.
2!

¿From this and mathematical induction we get

pn (b)
|Yn+1 (t) − Yn (t)| ≤ B0 (b) , u ≤ t ≤ b.
n!

Hence for any k ∈ N

|Yn+k+1 (t) − Yn (t)| ≤ |Yn+k+1 (t) − Yn+k (t)| + |Yn+k (t) − Yn+k−1 (t)| +
. . . + |Yn+1 (t) − Yn (t)|
pn (b) p(b) p2 (b)
≤ B0 (b) [1 + + + ...]
n! n + 1 (n + 2)(n + 1)
2
p(b) p (b)
Choose m large enough so that n+1 ≤ 12 then (n+2)(n+1) ≤ 41 , etc. when n > m
and the term in brackets is bounded above by 2. It follows that the sequence
{Yn : n ∈ N0 } converges uniformly, say to Y, on [u, b]. From this it follows that Y
satisfies the integral equation (2.5) and hence also the IVP (2.3), (2.4) on [u, b].
To show that Y is the unique solution assume Z is another one; then Z is
continuous and therefore |Y − Z| is bounded, say by M > 0 on [u, b]. Then

Z t Z t
|Y (t) − Z(t)| = P (s)[Y (s) − Z(s)] ds ≤ M P (s) ds ≤ M p(t), u ≤ t ≤ b.
u u

Now proceeding as above we get

pn (t) pn (b)
|Y (t) − Z(t)| ≤ M ≤M , u ≤ t ≤ b, n ∈ N.
n! n!

Therefore Y = Z on [u, b]. There is a similar proof for the case when b < u. This
completes the second proof. 2

It is interesting to observe that the initial approximation Y0 (t) = C can be

replaced with Y0 (t) = G(t) for any continuous function G without any essential
change in the proof.
Let J be an interval. For each P ∈ Mn (Lloc (J)), each F ∈ Mn,m (Lloc (J)),
each u ∈ J and each C ∈ Mn,m (C) there is, according to Theorem 2.2, a unique
Y ∈ Mn,m (ACloc (J)) such that Y 0 = P Y + F, Y (u) = C. We use the notation
 7

(2.12) Y = Y (·, u, C, P, F )

to indicate the dependence of the unique solution Y on these quantities. Below,

if the variation of Y with respect to some of the variables u, C, P, F is studied
while the others remain fixed we abbreviate the notation (2.12) by dropping those
quantities which remain fixed. Thus we may use Y (t) for the value of the solution
at t ∈ J when u, C, P, F are fixed or Y (·, u) to study the variation of the solution
function Y with respect to u, Y (·, P ) to study Y as a function of P, etc.

THEOREM 2.3. Let J = (a, b), and assume that P ∈ Mn (Lloc (J)). If Y is an
n × m matrix solution of

(2.13) Y 0 = P Y on J,

then we have

(2.14) rank Y (t) = rank Y (u), t, u ∈ J.

Moreover, if m = n and u ∈ J, then

Z t 
(2.15) (det Y )(t) = (det Y )(u)exp traceP (s)ds , t ∈ J.
u

PROOF: The formula (2.15) follows from the fact that y = detY satisfies the first
order scalar equation y 0 + py = 0 where p = traceP. To prove the general case let
Y (u) = C and let rank C = r. If r = 0, then Y (t) = 0 for all t by Theorem 2.2.
For r > 0 let Ci , i = 1, . . . r be linearly independent columns of C and construct
a nonsingular n × n matrix D by adding n − r appropriate constant vectors to
Ci , i = 1, . . . , r. Denote by Z the solution of (2.13) satisfying the initial condition
Z(u) = D. Then by (2.15) rank Z(t) = n, for t ∈ J. Hence the first r columns
of Z(t), Z1 (t), ..., Zr (t) are linearly independent. From this and the uniqueness
part of Theorem 2.2 the (constant) n-vectors Y1 (t), Y2 (t), . . . , Yr (t) are linearly
independent since Zj = Yj on J. Hence rank Y (t) ≥ r, for t ∈ J. Now suppose
that rank Y (c) > r for some c in J. Then by repeating the above argument with
u replaced by c we reach the conclusion that rank Y (t) > r for all t ∈ J. But this
contradicts rank Y (u) = r and concludes the proof. 2

THEOREM 2.4. Let P : J → Mn (C) and F : J → Mn,1 (C), J = (a, b), −∞ ≤

a < b ≤ ∞. If for any u ∈ J and any linearly independent constant vectors
C1 , . . . , Cn each initial value problem

Y 0 = P Y + F, Y (u) = Ci , i = 1, . . . , n,
8

has a (vector) solution Yi on J, then P ∈ Mn (Lloc (J)) and F ∈ Mn,1 (Lloc (J)).
Furthermore, if each Yi is a C 1 solution, then there exist such P and F which are
continuous.
PROOF: We first prove the special case when F = 0 on J. Let Yi be a vector
solution satisfying Yi (u) = Ci and let Y be the matrix whose i − th column is
Yi , i = 1, . . . , n. By Theorem 2.3 the solution Y is nonsingular at each point of J.
Choose

P = Y 0 Y −1 .

Let K be a compact subinterval of J. Then since Y is continuous and invertible on

J it follows that Y −1 is continuous and hence bounded on K. Also, Y 0 is integrable
on K since Y is absolutely continuous on K by virtue of the fact that it is a solution
on J. Therefore P ∈ Mn (Lloc (J)).
To establish the case when F is not identically zero on J let Y be a matrix
solution of Y 0 = P Y + F satisfying Y (u) = 0 and choose a solution Z of this
equation such that Z(u) = C and let V = Z − Y. Then V 0 = P V and V (u) = C.
Since this holds for arbitrary C we may conclude from the special case established
above that P ∈ Lloc (J). Hence F = V 0 − P V ∈ Lloc (J). The furthermore statement
is clear from the proof. 2

2.3 Variation of parameters.

Let P ∈ Mn (Lloc (J)). From Theorem 2.2 we know that for each point u of J there
is exactly one matrix solution X of (2.13) satisfying X(u) = In where In denotes
the n × n identity matrix.

DEFINITION 2.5 (The Fundamental Matrix Φ). For each fixed u ∈ J let Φ(·, u)
be the fundamental matrix of (2.13) satisfying

Φ(u, u) = In .

Note that for each fixed u in J, Φ(·, u) belongs to Mn (ACloc (J)). Furthermore, if J
is compact and P ∈ Mn (L(J)), then u can be an endpoint of J and Φ(·, u) belongs
to Mn (AC(J)). By Theorem 2.3, Φ(t, u) is invertible for each t, u ∈ J and we note
that

(2.16) Φ(t, u) = Y (t)Y −1 (u)

for any fundamental matrix Y of (2.13).

We also write
 9

(2.17) Φ = Φ(P ) = (Φrs )nr,s=1 , Φ(P )(t, u) = Φ(t, u, P ).

Observe that for any constant n × m matrix C, ΦC is also a solution of (2.13). If C

is a constant nonsingular n × n matrix then ΦC is a fundamental matrix solution
and every fundamental matrix solution has this form.
The next result is called the variation of parameters formula and is fundamental
in the theory of linear differential equations.

THEOREM 2.6 (Variation of Parameters Formula). Let J be any interval, let

P ∈ Mn (Lloc (J)) and let Φ = Φ(·, u, P ) be the fundamental matrix of (2.13)
defined above. Let F ∈ Mn,m (Lloc (J)), u ∈ J and C ∈ Mn,m (C). Then

Z t
(2.18) Y (t) = Φ(t, u, P ) C + Φ(t, s, P ) F (s) ds, t∈J
u

is the solution of (2.3), (2.4). Note that if J is compact and P ∈ Mn (L(J)),

F ∈ Mn,m (L(J)), then Y ∈ Mn,m (AC(J)), and u can be an endpoint or an interior
point of J.

PROOF: Clearly Y (u) = C. Differentiate (2.18) and substitute into the equation
(2.3). 2

2.4 The Gronwall Inequality.

Since we need a Gronwall inequality which is more general than the one usually
found in the literature we state and proof it here.

THEOREM 2.7. (The Gronwall Inequality)

(i) (The “right” Gronwall inequality) Let J = [a, b]. Assume g in L(J) with
g ≥ 0 a.e. , f real valued and continuous on J. If y is continuous, real valued, and
satisfies

Z t
(2.19) y(t) ≤ f (t) + g(s) y(s)ds , a ≤ t ≤ b,
a

then

Z t Z t 
(2.20) y(t) ≤ f (t) + f (s) g(s) exp( g(u) du ) ds , a ≤ t ≤ b.
a s

For the special case when f (t) = c, a constant, we get

10

Z t 
(2.21) y(t) ≤ c exp g(s)ds , t ∈ J.
a

For the special case when f is nondecreasing on [a, b] we get

Z t
(2.22) y(t) ≤ f (t) exp( g(s)ds), a ≤ t ≤ b.
a

(ii) (The “left” Gronwall inequality) Let J = [a, b]. Assume g in L(J) with
g ≥ 0 a.e. f real valued and continuous on J. If y is continuous, real valued, and
satisfies

Z b
(2.23) y(t) ≤ f (t) + g(s) y(s)ds , a ≤ t ≤ b,
t

then

!
Z b Z s
(2.24) y(t) ≤ f (t) + f (s) g(s) exp( g(u) du ) ds , a ≤ t ≤ b.
t t

For the special case when f (t) = c, a constant, we get

!
Z b
(2.25) y(t) ≤ c exp g(s)ds , a ≤ t ≤ b.
t

For the special case when f is nondecreasing on [a, b] we have

!
Z b
(2.26) y(t) ≤ f (t) exp g(s)ds , a ≤ t ≤ b.
t

Rt
PROOF: For part (i) let z(t) = a
g y, t ∈ J and note that

z 0 = gy ≤ gf + gz a.e.

Rt
Hence with G = exp( a g) we have

(z exp(−G))0 ≤ g f exp(−G)

and
 11

Z t
y(t) ≤ f (t) + z(t) ≤ c exp(G) + exp(G) (g f G)
a

Rb
from which the conclusion follows. For part (ii) let z(t) = t
g y and note that

z 0 + gy ≥ −g f,

then proceed as in part (i). 2

2.5 Bounds and extensions to the endpoints.

THEOREM 2.8. Let J = (a, b), −∞ ≤ a < b ≤ ∞, let n, m ∈ N . Suppose that

P ∈ Mn (L(J)); F ∈ Mn,m (L(J)). Assume that for some u ∈ J, C ∈ Mn,m (C) , we
have

(2.27) Y 0 = P Y + F on J, Y (u) = C.

Then

! !
Z b Z b
(2.28) |Y (t)| ≤ |C| + |F | exp |P | , a < t < b.
a a

PROOF: Note that (2.27) is equivalent to

Z t
(2.29) Y (t) = C + (P (s) Y (s) + F (s) ) ds, a < t < b.
u

Case 1. u ≤ t < b. From (2.29) we get

Z t Z t
|Y (t)| ≤ |C| + | (P Y + F ) | ≤ |C| + (|P | |Y | + |F | )
u u
Z b ! Z
t
≤ |C| + |F | + (|P ||Y |) , u ≤ t < b.
u u

¿From this and Gronwall’s inequality we get

! !
Z b Rt
Z b Rb
|Y (t)| ≤ |C| + |F | e( u
|P |)
≤ |C| + |F | e( u
|P |)
, u ≤ t < b.
u u
12

Case 2. a < t ≤ u. From (2.29)

Z t Z u
|Y (t)| ≤ |C| + (P Y + F ) ≤ |C| + (|P | |Y | + |F |)
u t
 Z u  Z u
≤ |C| + |F | + (|P ||Y |) , a ≤ t < u.
a t

¿From this and the “left” Gronwall inequality we get

 Z u  R  Z u  R
u u
|Y (t)| ≤ |C| + |F | e( t
|P |)
≤ |C| + |F | e( a |P |) , a < t ≤ u.
a a

Combining the two cases we conclude that (2.28) holds. 2

Below we will show that, under the conditions of Theorem 2.8, a < t < b can
be replaced with a ≤ t ≤ b in (2.28). For this Y (a) and Y (b) are defined as limits.
This holds for both finite and infinite endpoints a, b.

THEOREM 2.9. Let J = (a, b), −∞ ≤ a < b ≤ ∞. Assume that

(2.30) P ∈ Mn (Lloc (a, b)); F ∈ Mn,m (Lloc (a, b))

i) Suppose, in addition to (2.30), that

(2.31) P ∈ Mn (L(a, c)); F ∈ Mn,m (L(a, c))

for some c ∈ (a, b). For some u ∈ J and C ∈ Mn,m (C), let Y be the solution of the
IVP (2.3), (2.4) on J. Then

(2.32) Y (a) = lim Y (t)

t→a+

exists and is finite.

ii) Suppose that, in addition to (2.30), P, F satisfy

(2.33) P ∈ Mn (L(c, b)); F ∈ Mn,m (L(c, b))

for some c ∈ (a, b). For some u ∈ J and C ∈ Mn,m (C), let Y be the solution of the
IVP (2.3), (2.4) on J. Then

(2.34) Y (b) = lim− Y (t)

t→b
 13

exists and is finite.

PROOF: We establish Theorem 2.9 for b; the proof for the endpoint a is similar
and hence omitted. It follows from (2.28) that |Y | is bounded on [c, b) for c ∈ J,
say by B. Let {bi } be any strictly increasing sequence converging to b. Then for
j > i we have

Z bj Z bj
|Y (bj ) − Y (bi )| = PY ≤ B |P |.
bi bi

¿From this and the absolute continuity of the Lebesgue integral it follows that
{Y (bi ) : i ∈ N } is a Cauchy sequence and hence converges to a finite limit. 2

THEOREM 2.10. Let J = (a, b), −∞ ≤ a < b ≤ ∞. Assume that

(2.35) P ∈ Mn (Lloc (a, b)).

i) Suppose, in addition to (2.35), that

(2.36) P ∈ Mn (L(a, c))

for some c ∈ (a, b). Let, for some u ∈ J and C ∈ Mn,m (C), Y be the solution of
the IVP (2.3), (2.4) with F = 0 on J. Then

(2.37) rank Y (a) = rank Y (u)

where Y (a) is given by (2.32). Moreover, given any C ∈ Mn,m (C) there exists a
unique solution Y of the “endpoint” value problem:

(2.38) Y 0 = P Y, Y (a) = C.

ii) Suppose, in addition to (2.35), that

(2.39) P ∈ Mn (L(c, b))

for some c ∈ (a, b). Let, for some u ∈ J and C ∈ Mn,m (C), Y be the solution of
the IVP (2.3), (2.4) with F = 0 on J. Then

(2.40) rank Y (b) = rank Y (u)

where Y (b) is given by (2.34). Moreover, given any C ∈ Mn,m (C) there exists a
unique solution Y of the “endpoint” value problem:
14

(2.41) Y 0 = P Y, Y (b) = C.

Note that the endpoints a and b in Theorem 2.10 may be finite or infinite.
PROOF: Note that (2.37) or (2.40) do not follow directly from (2.14) and (2.32)
or (2.34) since the rank of a matrix is not a continuous function of the matrix. We
argue as follows: Let Y (u) = C, rank C = r. If r = 0, then Y (t) = 0 for all t ∈ J
and Y (b) = 0 by (2.34). If r > 0, let C1 , . . . , Cr be linearly independent columns
of Y (u) and construct a nonsingular n × n matrix D by adding n − r appropriate
columns to Cj , j = 1, ..., r. Let Z denote the solution of (2.13) determined by the
initial condition Z(u) = D. It follows from (2.15) that Z(t) is nonsingular for each
t ∈ J and hence Z(b) is nonsingular by (2.34) and (2.15). (Note that it does
not follow directly from (2.34) alone that Y (b) is nonsingular since the rank of a
matrix is not a continuous function of its coefficients.) Therefore Z1 (b), ..., Zr (b)
are linearly independent. From the uniqueness part of the existence-uniqueness
theorem - Theorem 2.2 - Y (t) = Z(t) for t ∈ J and hence also for t = b by (2.34).
The proof for the endpoint a is similar. This establishes (2.37) and (2.40). To
prove the moreover parts of the Theorem consider the fundamental matrix Φ , see
Definition 2.5, choose u ∈ J and determine the solution Y of (2.13) by the initial
condition Y (u) = Φ(b, u) C then Y (b) = C. Note that Φ(b, u) exists by (2.34). The
proof of (2.38) is similar. 2

2.6 Continuous dependence of solutions on the problem.

THEOREM 2.11. Let u, v ∈ J = (a, b), −∞ ≤ a < b ≤ ∞ , C, D ∈ Mn,m (C),

P, Q ∈ Mn (L(J)), F, G ∈ Mn,m (L(J)). Assume

(2.42) Y 0 = P Y + F on J, Y (u) = C; Z 0 = Q Z + G on J, Z(u) = D.

Then

Rb
(2.43) |Y (t) − Z(t)| ≤ K e( a
|Q|)
, a ≤ t ≤ b,

where

Z v Z v Z b Z b
(2.44) K = |C − D| + |F | + M |P | + |F − G| + M |P − Q|,
u u a a

and

! !
Z b Z b
(2.45) M= |C| + |F | exp |P | .
a a
 15

PROOF: For a < t < b this follows from Theorem 2.8 and the Gronwall inequality.
The case t = a and t = b then follows from Theorem 2.9. 2

THEOREM 2.12. Let J = (a, b), −∞ ≤ a < b ≤ ∞, u ∈ J, C ∈ Mn,m (C),

P ∈ Mn (L(J)), and F ∈ Mn,m (L(J)). Let Y = Y (·, u, C, P, F ) be the solution
of (2.3), (2.4) on J. Then Y is a continuous function of all its variables u, C, P, F
uniformly on the closure of J ; more precisely, for fixed P, F, u, C; given any
 > 0 there is a δ > 0 such that if v ∈ J, D ∈ Mn,m (C), Q ∈ Mn (L(J)), and
G ∈ Mn,m (L(J)) satisfy

Z b Z b
(2.46) |u − v| + |C − D| + |P − Q| + |F − G| < δ,
a a

then

(2.47) |Y (t, u, C, P, F ) − Y (t, v, D, Q, G)| <  , a ≤ t ≤ b.

¯
Note that Y (t, u, C, P, F ) is jointly continuous in u, C, P, F, uniformly for t in J.
PROOF: The absolute continuity of the Lebesgue integral and (2.46) imply that
the constant K in (2.44) can be made arbitrarily small. The conclusion then follows
from Theorem 2.11. 2

THEOREM 2.13. Let J = (a, b), −∞ ≤ a < b ≤ ∞, let Pk ∈ Mn,n (Lloc (J)),
Fk ∈ Mn,m (Lloc (J)), Ck ∈ Mn,m , uk ∈ J, k ∈ N0 = {0, 1, 2, . . . }. Assume
(i) Pk → P0 as k → ∞
locally in Lloc (J) in the sense that for each compact subinterval K of J we
have

Z
|Pk − P0 | → 0 as k → ∞;
K

(ii) Fk → F0 as k → ∞
locally in Lloc (J) in the sense that for each compact subinterval K of J we
have

Z
|Fk − F0 | → 0 as k → ∞;
K

(iii) Ck → C0 ∈ C as k → ∞;
(iv) uk → u0 ∈ J as k → ∞.
16

Then

Y (t, uk , Ck , Pk , Fk ) → Y (t, u0 , C0 , P0 , F0 ) as k → ∞

locally uniformly on J, i.e., uniformly in t on each compact subinterval of J.

Moreover, if Pk ∈ Mn (L(J)), Fk ∈ Mn,m (L(J)), and (i), (ii) hold in L(J), i.e.
with K replaced by J and (iii), (iv) hold, then

Y (t, uk , Ck , Pk , Fk ) → Y (t, u0 , C0 , P0 , F0 ) as k → ∞

uniformly on the closure of J.

PROOF: This follows from Theorem 2.12. 2

2.7 Differentiable dependence of solutions on the data including the

coefficients.

Theorem 2.12 shows that the solution of the initial value problem (2.3), (2.4) with
P ∈ Mn,n (Lloc (J)) , F ∈ Mn,m (Lloc (J)) , C ∈ Mn,m (C) depends continuously on
all the given data. In this section we show that this dependence is differentiable.

DEFINITION 2.14 (The Frechet derivative on Banach spaces). A map T from

a Banach space X into a Banach space Z, T : X → Z, is differentiable at a point
x ∈ X if there exists a bounded linear map T 0 (x) : X → Z such that

|T (x + h) − T (x) − T 0 (x) h| = o (|h|), as h → 0 in X.

That is, for each ε > 0 there is a δ > 0 such that

|T (x + h) − T (x) − T 0 (x) h| ≤ ε (|h|) f or all h ∈ X with |h| < δ.

If such a map T 0 (x) exists, it is unique and is called the (Frechet) derivative of T
at x. A map T is differentiable on a set S ⊂ X if it is differentiable at each point of
S. In this case the derivative is a map : x → T 0 (x) from S into the Banach space
L(X, Z) of all bounded linear operators from X into Z denoted by T 0 . To say that
T 0 is continuously differentiable on S or T is C 1 on S means that the map T 0 is
continuous in the operator topology of the Banach space L(X, Z).
The differentiability of the solution

Y = Y (t, u, C, P, F )
 17

with respect to t follows from the definition of solution. The differentiability of Y

with respect to u is established in the next lemma.

LEMMA 2.15. Fix t, C, P, F and consider Y as a function of u. Then Y ∈

ACloc (J).
PROOF: It follows from the representation (2.16) that the fundamental matrix
Φ(t, u) is differentiable with respect to u, since the inverse of a differentiable matrix
is differentiable. The differentiability of Y with respect to u then follows from the
representation

Z t
Y (t, u) = Φ(t, u) C + Φ(t, s) F (s) ds.
u

This concludes the proof. 2

For fixed t, u, Y is a function of C, P, F mapping Mn,m (C) × Mn (L(J)) ×

Mn,m (L(J)) into Mn,m (L(J)). By Theorem 2.12, Y is continuous in C, P, F. Is
it differentiable in C ? in P ? in F ?

THEOREM 2.16. For fixed t, u ∈ J, P ∈ Mn (L(J)) and F ∈ Mn,m (L(J)), the

solution Y = Y (t, u, C, P, F ) of (2.3), (2.4) is differentiable in C; its derivative is
given by

∂Y
(2.48) Y 0 (C) = (t, u, C, P, F ) = Φ(t, u, P ).
∂C

Thus we have

(2.49) Y 0 (C) H = Φ(t, u, P ) H, H ∈ Mn,m (C).

The derivative Y 0 (C) is constant in C and in F.

PROOF: This follows directly from the variation of parameters formula and the
definition of derivative. 2

THEOREM 2.17. Let J = [a, b]. Fix t, u ∈ J, C ∈ Mn,m (C), P ∈ Mn (L(J)),

and F ∈ Mn,m (L(J)); let Y = Y (t, u, C, P, F ). We have

(2.50)
Z t
∂Y
Y 0 (F )(H) = (t, u, C, P, F )(H) = Φ(t, s)H(s) ds, H ∈ Mn,m (L(J)).
∂F u

Here the right side of equation (2.50) defines a bounded linear operator on the space
Mn,m (L(J)). The derivative Y 0 (F ) is constant in F .
18

PROOF: From the variation of parameters formula we get

Z t
Y (t, u, C, P, F + H) − Y (t, u, C, P, F ) = Φ(t, s, P )H(s) ds.
u

The conclusion follows from this equation and the definition of derivative. 2

Before stating the next Theorem we give two lemmas. These may be of inde-
pendent interest.

LEMMA 2.18. Let J = (a, b), −∞ ≤ a < b ≤ ∞, let P ∈ Mn (Lloc (J)), let
u ∈ J. Then for any t ∈ J we have

Z t Z t Z r
Φ(t, u, P ) = I+ P+ P (r) P (s) drds
u u u
Z t Z r Z s
(2.51) + P (r) P (s) P (x) dxdsdr + · · · .
u u u

PROOF: This follows directly from the successive approximations proof of the
existence-uniqueness
Rt Theorem : Start with the first approximation Φ0 = I; then
Φ1 = I + u P, etc. 2

LEMMA 2.19. Let P, H ∈ Mn (Lloc (J)). Then for any t, u ∈ J we have

(2.52) Φ(t, u, P + H) = Φ(t, u, P ) Φ(t, u, S),

where

(2.53) S = Φ−1 (·, u, P ) H Φ(·, u, P ).

PROOF: The proof consists in showing that both sides satisfy the same initial value
problem and then using the existence-uniqueness Theorem. 2

LEMMA 2.20. If P commutes with the integral of H in the sense that

Z s Z s
(2.54) P (t) ( H) = ( H) P (t), s, t, u ∈ J,
u u

then the exponential law holds:

(2.55) Φ(t, u, P + H) = Φ(t, u, P ) Φ(t, u, H).

 19

PROOF: It follows from Lemmas 2.18, 2.19 and hypothesis (2.54) that Φ(·, u, P ) H =
H Φ(·, u, P ) and hence S = H in (2.53). 2

THEOREM 2.21. Let J = [a, b]. Fix t, u ∈ J, C ∈ Mn,m (C), F ∈ Mn,m (L(J)).
For P ∈ Mn (L(J)) let Y = Y (t, u, C, P, F ) be the unique solution of (2.3), (2.4).
Then the map P → Y (t, u, C, P, F ) from the Banach space Mn (L(J)) to Mn,m (C)
is differentiable and its derivative
20

∂Y
(2.56) Y 0 (P ) = (t, u, C, P, F )
∂P

is the bounded linear transformation from the Banach space Mn (L(J)) to the Ba-
nach space Mn,m (C) given by

Z t 
0 −1
Y (P ) H = Φ(t, u, P ) Φ (r, u, P )H(r)Φ(r, u, P ) dr C
u

(2.57)
Z t Z t 
−1
+ Φ(t, r, P ) Φ (s, u, P )H(s)Φ(s, u, P ) ds F (r) dr, H ∈ Mn (L(J)).
u r

PROOF: Fix t, u, C, F and let Y (t, P ) = Y (t, u, C, P, F ). From the variation of

parameters formula it follows that for H ∈ Mn,n (L(J)) and S defined by (2.53) we
have

Y (t, P + H) − Y (t, P )
Z t
= Φ(t, P + H) C + Φ(t, s, P + H) F (s) ds − Φ(t, P ) C
u
Z t
− Φ(t, s, P ) F (s) ds
u
Z t
= Φ(t, u, P ) [Φ(t, u, S) − I] C + Φ(t, s, P )[Φ(t, r, S) − I]F (r)dr
u
Z t Z t Z x  Z t
= Φ(t, u, P ) S+ S(x) S(y)dydx + . . . C + Φ(t, r, P )
u u u u
Z t Z t Z x 
S+ S(x) S(y)dydx + . . . F (r)dr
r r r

Hence

Z t 
Y (t, u, P + H) − Y (t, u, P ) − Φ(t, u, P ) S(r)dr C
u
Z t Z t 
− Φ(t, r, P ) S(x)dx F (r)dr
u r
Z t Z x 
= Φ(t, u, P ) S(x) S(y)dydx + . . . C
u u
Z t Z t Z x 
+ Φ(t, r, P ) S(x) S(y)dydx + . . . F (r)dr
u r r
= E(H).
 21

Noting that |S|(b−a)| ≤ |k H| for some k ∈ R, that |Φ(t, u, P )| and |Φ−1 (t, u, P )|
are bounded on J, there exists an M > 0 such that

|E(H)| ≤ M |C| [|S(b − a)|2 + |S(b − a)|3 . . . ] + M |F | [|S(b − a)|2 + |S(b − a)|3 . . . ]
≤ M |C| |kH| [|kH| + |kH|2 + . . . ] + M |F | |kH| [|kH| + |kH|2 + . . . ]

¿From this it follows that

|E(H)|
→ 0 as |H| → 0 in Mn (L(J)).
|H|

This completes the proof. 2

THEOREM 2.22. Let the hypotheses and notations of Theorem 2.21 hold and
assume, in addition, that the commutativity hypothesis (2.54) is satisfied. Then
1.
(2.58) H(t) Φ(t, u, P ) = Φ(t, u, P ) H(t), t, u ∈ J.

2. The exponential law holds, i.e.

(2.59) Φ(t, u, P + H) = Φ(t, u, P ) Φ(t, u, H), t, u ∈ J.

3. Formula (2.57) reduces to

Z t 
Y 0 (P )(H) = Φ(t, u, P ) H(s) ds C +
u
Z t Z t 
(2.60) Φ(t, r, P ) H(s) ds F (r) dr, t, u ∈ J.
u r

Note however that Y 0 (P ) is not the operator defined by the right hand side of
(2.60) since H cannot be restricted to satisfy the commutativity hypothesis (2.54)
in the definition of the derivative Y 0 (P ).
PROOF: This follows from Theorem 2.21 and Lemma 2.20. 2

REMARK 1. In the special case when P and H are constant matrices we have

Z t 
Y 0 (P )(H) = exp((t − u)P ) exp((u − r)P )H exp((r − u)P ) dr C
u
22

Z t Z t 
(2.61) + exp((t − r)P ) exp((u − s)P )H exp((s − u)P )ds F (r) dr.
u r

Note that if P and H are constant and commute, then (2.61) reduces to
Z t
0
Y (P )(H) = (t − u) exp ((t − u)P ) HC + (t − r) exp((t − r)P )HF (r) dr.
u

But this reduction does not hold, in general, for constant matrices which do not
commute.

COROLLARY 2.23. Consider the exponential map of matrices:

E(A) = eA , A ∈ Mn (C).

The Frechet derivative of E is the bounded linear operator from Mn (C) into
Mn (C) given by

Z 1
0
(2.62) E (A) H = e A
e− r A H er A dr, H ∈ Mn (C).
0

Note that (2.62) reduces to the more familiar formula E 0 (A) = E(A) for all
A ∈ Mn (C) only in the one dimentional case n = 1. When n > 1 (2.62) reduces to
E 0 (A) = E(A) only for constant multiples A = cIn , c ∈ C , of the identity since
only multiples of the identity satisfy the commutativity condition with respect to
all matrices in Mn (C).

In Corollary 2.23 Mn (C) can be replaced by Mn (R); in fact Mn (C) can be

replaced by an arbitrary Banach algebra. See [56].

PROOF: This is the special case of Theorem 2.21 when a = 0 = u, b = 1, P (t) = A

for all t ∈ [0, 1], F ≡ 0, C = I. 2

For fixed t, u, C, F replace P by P + zW and fix P and W . The next result

shows that the solution Y = Y (t, u, C, P + zW, F ) of (2.3), (2.4) with P replaced
by P + zW is an entire function of z. What is

∂Y
Y 0 (z) = ?
∂z

This question is answered by

 23

THEOREM 2.24. Let J = (a, b), −∞ ≤ a < b ≤ ∞, t, u ∈ J, C ∈ Mn,m ,

P, W ∈ Mn (L(J)), F ∈ Mn,m (L(J)); let Y = Y (t, u, C, P + zW, F ) denote the
unique solution of (2.3), (2.4) for each z ∈ C. Then Y is an entire function of z
and

Z t 
Y 0 (z) = Φ(t, u, P + zW ) Φ−1 (r, u, P + zW ) W (r) Φ(r, u, P + zW ) dr C
u
Z t Z t 
+ Φ(t, r, P + zW ) Φ−1 (s, u, P + zW ) W (s) Φ(s, u, P + zW ) ds F (r)dr
u r

PROOF:
[Y (t, u, C, P + (z + h)W, F ) − Y (t, u, C, P + zW, F )]
= [Φ(t, u, P + (z + h)W ) − Φ(t, u, P + zW )] C
Z t
(2.63) + [Φ(t, r, P + (z + h)W ) − Φ(t, r, P + zW )] F (r) dr.
u

Let
S(z) = Φ−1 (·, u, P + zW )W (·) Φ(·, u, P + zW ).

Proceeding similarly to the proof of Theorem 2.21 we get

Φ(t, u, P + (z + h)W ) − Φ(t, u, P + zW )

= Φ(t, u, P + (z + h)W )[Φ(t, u, hS(z)) − I]
Z t
= Φ(t, u, P + zW )[h S(z) + o(h)]
u
Z t
(2.64) = hΦ(t, u, P + zW ) S(z) + o(h)
u

Combining these two we get

[Y (t, u, C, P + (z + h)W, F ) − Y (t, u, C, P + zW, F )]

 Z t  Z t Z t  
= h Φ(t, u, P + zW ) S(z) C + Φ(t, r, P + zW ) S(z) F (r) dr
u u r
+o(h).

And the result follows. 2

THEOREM 2.25. Let J = (a, b), −∞ ≤ a < b ≤ ∞, t, u ∈ J, C ∈ Mn,m (C),

P ∈ Mn (L(J)), F ∈ Mn,m (L(J)); and for each z ∈ C let Y = Y (t, u, C, P +
zW, F ) denote the unique solution of (2.3), (2.4) for each W ∈ L(J). Then Y is a
differentiable function of W and
24

Z t 
Y 0 (W ) H = z Φ(t, u, P + zW ) Φ−1 (r, u, P + zW )H(r) Φ(r, u, P + zW ) dr C
u
Z t Z t 
−1
+z Φ(t, r, P + zW ) Φ (s, u, P + zW )H(s) Φ(s, u, P + zW ) ds F (r) dr,
u r

for H ∈ L(J).
PROOF: The proof is similar to that of Theorem 2.21 and hence omitted. 2

2.8 Adjoint systems.

LEMMA 2.26. Let P, Q be any k × k complex matrix functions on J. Let F, G

be k × m complex matrix functions on J. If Y 0 = P Y + F and Z 0 = QZ + G and
C ∈ E k (C), then

(2.65) (Z ∗ CY )0 = Z ∗ (Q∗ C + CP )Y + Z ∗ CF + G∗ CY.

PROOF: This follows from a straightforward computation and is therefore omitted.

2

COROLLARY 2.27. Let the assumptions and notation be as in Lemma 2.26 .

∗
If, in addition, C is invertible and Q = −C −1 P ∗ C ∗ , then

(2.66) (Z ∗ CY )0 = Z ∗ CF + G∗ CY.

PROOF: This follows from (2.65). 2

The fundamental matrices of adjoint systems are closely related to each other.
The next result gives this relationship. It plays an important role in the theory of
adjoint and, in particular, self-adjoint boundary value problems.

THEOREM 2.28 (Adjointness Lemma). Let P ∈ Mn (Lloc (J)), let E ∈ Mn (C).

Assume

(2.67) E −1 E ∗ = I or E −1 E ∗ = −I

and define

(2.68) P + = −E −1 P ∗ E.
 25

Then

(2.69) Φ(t, s, P ) = E −1 Φ∗ (s, t, P + ) E, s, t ∈ J.

PROOF: Fix s ∈ J and let

Z(t) = E −1∗ Φ∗ (t, s, P ) E ∗ Φ(t, s, P + ), t ∈ J.

Note that Z(s) = I and

Z 0 (t) = E −1∗ [P (t)Φ(t, s, P )]∗ E ∗ Φ(t, s, P + )

+E −1∗ Φ∗ (t, s, P ) E ∗ P + (t) Φ(t, s, P + )
= E −1∗ Φ∗ (t, s, P ) EE −1 P ∗ (t)E E −1 E ∗ Φ(t, s, P + )
+E −1∗ Φ∗ (t, s, P ) E ∗ P + (t) Φ(t, s, P + )
= −E −1∗ Φ∗ (t, s, P ) E ∗ P + (t) Φ(t, s, P + )
+E −1∗ Φ∗ (t, s, P ) E ∗ P + (t) Φ(t, s, P + )
= 0, t ∈ J,

using (2.67) and (2.68). Hence Z(t) = I, for t ∈ J. That this is equivalent to
(2.69) follows from the representation Φ(t, s, P + ) = Y (t) Y −1 (s), s, t ∈ J, for any
fundamental matrix Y of Y 0 = P + Y. 2

2.9 Inverse Initial Value Problems.

NOTATION. Given d n-dimensional vectors Y1 , Y2 , . . . , Yd we denote the n × d

matrix whose i − th column is Yi , i = 1, . . . , d, by

(2.70) Y = [Y1 , Y2 , . . . , Yd ].

Above we started with a coefficient matrix P and, possibly, a nonhomogeneous

term F and then studied the existence of solutions and their properties. Here we
reverse this. Given a number of functions, under what conditions are they solutions
of a first order linear system? For the sake of completeness we state the theorem
for both the direct and the inverse problems.

THEOREM 2.29. (i) Let 1 ≤ d ≤ n, P ∈ Mn (Lloc (J)). Assume that Yi ,

i = 1, . . . , d are vector solutions of

(2.71) Y 0 = PY .
26

If
(2.72) rank [Y1 , Y2 , . . . , Yd ](t) = d

for some t in J, then this is true for every t in J.

(ii) Let Yi ∈ Mn,1 (ACloc (J)), i = 1, . . . , d , 1 ≤ d ≤ n . Assume that

(2.73) rank[Y1 , . . . , Yd ](t) = d t ∈ J.

Then there exists an n×n matrix P ∈ Mn (Lloc (J)) such that Yi , i = 1, . . . , d,

are solutions of (2.71).
Furthermore, if Yi ∈ Mn,1 (C 1 (J)), i = 1, . . . , d then there exists a contin-
uous such P .
PROOF: Part (i) is contained in Theorem 2.2 so we only prove part (ii). If d = n
take P = Y 0 Y −1 . If d < n we construct an n × n matrix

M = [Y1 , Y2 , . . . , Yd , Yd+1, . . . , Yn ]

as follows. For each t1 ∈ J there is a d × d nonsingular submatrix of the n × d

matrix [Y1, . . . , Yd ](t1 ). Let its rows be numbered by r1, . . . , rd . To the right of the
first row which in not one of these place the first row of the (n − d) × (n − d) identity
matrix; to the right of the second row which is not one of these place the second
row of the (n − d) × (n − d) identity matrix, and so on. Thus each of Yi for i > d
is a constant matrix with all components zero except one which is the number 1.
For each t1 ∈ J the matrix M so constructed is nonsingular at t1 and by continuity
det M (t) 6= 0 for all t in some neighborhood Nt1 of t1 . Take

P (t) = M 0 (t) M −1 (t) , f or t ∈ Nt1 .

Any compact subinterval of J can be covered by a finite number of such neigh-

borhoods Nt1 and hence P can be defined on J. On points which are covered by
more than one such neighborhood, P is multiply defined, we just choose one def-
inition, say the one determined by the lowest numbered neighborhood. Clearly
P ∈ Mn (Lloc (J)) and Yi i = 1, . . . , d are solutions. This completes the proof of the
first part of (ii).
To prove the furthermore part we note that the constructed matrix P is piecewise
continuous by construction. Thus to get a continuous P we remove the multiply
defined aspect of the above construction as follows: On a subinterval which is
covered by two or more of the neighborhoods Nt discard all definitions of M used
above - just on this subinterval - then connect the two remaining pieces together in
such a way as to keep M nonsingular on J. Then construct a new P from the new
M as above for all t ∈ J. This results in a continuous P and completes the proof.2
 27

2.10 Comments.

Most of Section 2 is based on the paper [59] by Kong and Zettl. Below we comment
on each subsection separately.

1. The notation for matrix functions such as Mn (L(J)) is taken from [70].
2. The sufficiency of the local integrability conditions of Theorem 2.2 are well
known - see [72] or [81]; the necessity given by Theorem 2.4 is due to Everitt
and Race - see [25]. Except for the use of the Bielecki norm the first proof
of Theorem 2.2 is the standard successive approximations argument although
it is dressed in the clothes of the Contraction Mapping Theorem in Banach
space here. The advantage of the Bielecki norm is that it yields a global proof;
the sup norm would only give a local proof and then one has to patch together
the intervals of existence. The second proof is a minor variant of the usual
successive approximations argument.
The constancy of the rank of solutions given by Theorem 2.3 is known -
see [42] or [75] but we haven’t seen it stated under these general conditions.
It is surprising how many authors, including the two just mentioned assume
continuity of the coefficients when only Lebesgue integrability is needed. This
is of some consequence both theoretically and numerically when coefficients
are approximated by piece-wise constants, piece-wise linear functions, etc.
3. The variation of parameters formula given by Theorem 2.6 is standard, but
our notation is not. We use a notation which shows the dependence of the
fundamental matrix on the coefficient matrix P . This is handy for the differ-
entiation results that follow.
4. A detailed discussion of the Gronwall inequality is given here because it is a
very useful tool and we do not want any continuity assumptions on f and g.
The Gronwall inequality has many extensions: see - [11],
5. Theorem 2.8 is elementary but we have not seen it stated in this generality.
The continuous extensions of solutions given by Theorem 2.9 are a special
case of much more powerful results e.g. Levinsons asymptotic theorem -
[14]. Often the existence of limits of solutions are stated only for infinite
endpoints. We want to emphasize here that the relevant consideration is not
whether the endpoint is finite or infinite but whether the coefficient matrix
P and the inhomogeneous term F are integrable or not all the way to the
endpoint. Theorem 2.10 may be new in [59].
6. Theorems 2.11, 2.12, and 2.13 illustrate clearly that the natural space in which
to study solutions of linear ode’s is L1loc (J) in the singular case and L1 (J) in
the regular case.
7. Sections 7 and 9 were motivated to some extend by the elegant treatment of
the inverse spectral theory for regular Sturm-Liouville problems by Poeschel
and Trubowitz in [76]. Theorem 2.14 is standard - see [81] or [3]. Theorems
2.14 and 2.15 are trivial consequences of the variation of parameters formula.
Theorems 2.21, 2.24, 2.25, are not so trivial consequences of the Variation of
Parameters Formula; and may be new in [59].
8. Adjoint systems of this type were used by Atkinson [3] . They will be used
in the next chapter to provide an elegant proof of a very general Lagrange
identity due to Everitt and Neumann, see [24]. Theorem 2.27, the Adjointness
Lemma, is due to Zettl, see [82], [83], [87].
28

9. These kinds of inverse problems are discussed by Hartman [42] and Petrovski
[75] but not in this generality.

3 SCALAR INITIAL VALUE PROBLEMS (IVP)

We study initial value problems (IVP) consisting of the equation

(3.1) −(py 0 )0 + qy = f on J

together with initial conditions

(3.2) y(c) = h, (py 0 )(c) = k, c ∈ J, h, k ∈ C

where

(3.3) J = (a, b), −∞ ≤ a < b ≤ ∞, p, q, f : J → C.

3.1 Existence and uniqueness.

DEFINITION 3.1 (Solution). By a solution of equation (3.1) we mean a func-

tion y : J → C such that y and py 0 are absolutely continuous on each compact
subinterval of J and the equation is satisfied a.e. on J. Given a solution y of (3.1)
we refer to (py 0 ) as its quasi-derivative to distinguish it from the classical derivative
y0 .
Note that the classical derivative of a solution y, in general, exists only almost
everywhere but the quasi-derivative (py 0 ) is absolutely continuous on all compact
subintervals of J and thus exists and is continuous at each point of J.

THEOREM 3.2. Every initial value problem (IVP) (3.1), (3.2), (3.3) has a solu-
tion defined on J and this solution is unique if and only if

(3.4) 1/p, q, f ∈ Lloc (J).

If all the data p, q, f, h, k is real, then the solution is real on J.

PROOF: Let

     
0 1/p 0 y
(3.5) P = , F = , Y =
q 0 f py 0
 29

Then the equation (3.1) is equivalent to the first order system

(3.6) Y 0 = P Y + F on J,

in the sense that, given any scalar solution y of (3.1) the vector Y defined by
(3.5) is a solution of the system (3.6) and conversely, given any vector solution Y
of system (3.6) its top component y is a solution of (3.1). Theorem 3.2 follows from
this system representation and Theorems 2.2 and 2.4. 2

Notation. Given (3.3) and (3.4), the unique solution y of (3.1), (3.2) and its
quasi-derivative py 0 are denoted by

(3.7) y = y(·, c, h, k, 1/p, q, f ), py 0 = (py 0 )(·, c, h, k, 1/p, q, f ),

to highlight their dependence on these quantities. In the theory of boundary value

problems the spectral parameter λ and the weight function w play important roles;
thus we also study the equation

(3.8) −(py 0 )0 + qy = λ w y on J, λ ∈ C,

where

(3.9) p, q, w : J → C, 1/p, q, w ∈ Lloc (J) on J.

For this case we use the notation

(3.10) y = y(·, c, h, k, 1/p, q, w, λ), (py 0 ) = (py 0 )(·, c, h, k, 1/p, q, w, λ).

Theorem 3.2 and its proof readily extend to the case when q is replaced by q − λw.
Below, when we study the dependence of y and py 0 on one of these quantities
with all the others fixed we further abbreviate this notation by simply omitting all
the fixed variables. Thus we write y = y(·, c) when we wish to study the unique
solution as a function of c, y = y(·, q) to study the dependence of y on q, etc.

3.2 Continuous extensions to the endpoints.

DEFINITION 3.3 (Regular and singular endpoints). Let J = (a, b), −∞ ≤ a <
b ≤ ∞, and consider the equation (3.8) with conditions (3.9).
The end-point a is said to be regular if
30

(3.11) 1/p, q, w ∈ L(a, d)

for some d ∈ J; otherwise it is called singular. Similarly, the end-point b is said to

be regular if

(3.12) 1/p, q, w ∈ L(d, b)

for some d ∈ J ; otherwise it is called singular.

Note that, given (3.9), if (3.11) or (3.12) hold for some d ∈ J then they hold for
any such d.

REMARK 2. In much of the literature an infinite endpoint is automat-

ically classified as singular in contrast with Definition 3.3. We propose
this definition in view of the fact that, given (3.9), (3.12) is necessary
and sufficient for all solutions y of (3.8) and their quasi-derivatives py 0
to have a finite limit at b. See Theorem 3.4 below.
It is not the finite or infinite nature of the endpoint b but condition
(3.12) which determines whether or not all solutions of (3.8) and their
quasi-derivatives have finite limits at b. This is a natural definition of
“regular” behavior at b. Similar remarks apply at the endpoint a.

THEOREM 3.4. The limits

(3.13) y(a) = lim+ y(t), (py 0 )(a) = lim+ (py 0 )(t)

t→a t→a

both exist and are finite for the solution y of every initial value problem (3.1), (3.2),
(3.3), (3.4) if and only if

(3.14) 1/p, q, f ∈ L(a, d)

for some d ∈ (a, b); the limits

(3.15) y(b) = lim y(t), (py)(b) = lim (py 0 )(t)

t→b− t→b−

both exist and are finite for the solution y of every initial value problem (3.1), (3.2),
(3.3), (3.4) if and only if

(3.16) 1/p, q, f ∈ L(d, b)

 31

for some d ∈ (a, b).

PROOF: The sufficiency of the conditions (3.16), (3.14) follows from Theorem 2.9.
A proof of the necessity can be constructed along the lines of the proof of Theorem
2.4. 2

THEOREM 3.5. Let (3.1), (3.2), (3.3), (3.4) hold.

• Assume (3.14) and define y(a) and (py 0 )(a) by (3.13). Then each “initial value
problem ” consisting of equation (3.1) and the terminal value conditions

(3.17) y(a) = h, (py 0 )(a) = k, h ∈ C

has a unique solution on J; this solution is real if all the data is real.
• Assume (3.16) and define y(b) and (py 0 )(b) by (3.15). Then each “initial value
problem ” consisting of equation (3.1) and the terminal value conditions

(3.18) y(b) = h, (py 0 )(b) = k, h, k ∈ C

has a unique solution on J; this solution is real if all the data is real.
PROOF: This follows from the moreover part of Theorem 2.10. 2

3.3 Continuous dependence of solutions on the problem.

THEOREM 3.6. Let (3.1), (3.2), (3.3) and (3.4) hold. Using the notation (3.7)
each solution y of (3.1), (3.2) and its quasi-derivative py 0 is a jointly continuous
function of all its variables, uniformly on compact subintervals of J = (a, b). More
precisely, given cj ∈ J, hj , kj ∈ C, 1/pj , qj , fj ∈ Lloc (J), j = 1, 2, and given  > 0
and a compact subinterval K = [a1 , b1 ] of J containing c1 and c2 , there exists a
δ > 0 such that if

(3.19)
Z
|c1 − c2 | + |h1 − h2 | + |k1 − k2 | + (|1/p1 − 1/p2 | + |q1 − q2 | + |f1 − f2 |) < δ,
K

then

(3.20) |y(t, c1 , h1 , k1 , 1/p1 , q1 , f1 ) − |y(t, c2 , h2 , k2 , 1/p2 , q2 , f2 )| < 

and

(3.21) |(py 0 )(t, c1 , h1 , k1 , 1/p1 , q1 , f1 ) − (py 0 )(t, c2 , h2 , k2 , 1/p2 , q2 , f2 )| < 

32

both for all t ∈ K.

Furthermore, if 1/p, q, w ∈ L(J), and (3.19) holds with K = J, then (3.20) and
(3.21) hold on J.
PROOF: This is a consequence of Theorem 2.12. 2

3.4 Differentiable dependence of solutions on the data.

The differentiability of the solution

y(t, c, h, k, 1/p, q, w, λ)

of (3.8), (3.9) and its quasi-derivative (py 0 ) with respect to t follows from the
definition of solution; the differentiability of y and (py 0 ) with respect to c is a
consequence Lemma 2.16. The differentiability of y and of (py 0 ) with respect to the
other variables is studied in this subsection.

THEOREM 3.7. Let (3.3), (3.8), (3.9) hold. Let u, v be solutions of (3.8) deter-
mined by the initial conditions

u(c) = 0, (pu0 )(c) = 1; v(c) = 1, (pv 0 )(c) = 0, c ∈ J.

Using the notation (3.10) with the associated convention mentioned in the para-
graph below (3.10), we have that each of the following maps from C to C :

h → y(t, c, h, k, 1/p, q, w, λ), h → (py 0 )(t, c, h, k, 1/p, q, w, λ),

k → y(t, c, h, k, 1/p, q, w, λ), k → (py 0 )(t, c, h, k, 1/p, q, w, λ)

is differentiable and the derivatives are given by:

y 0 (h) = v(t) h, h ∈ C
0
(py 0 ) (h) = (pv 0 )(t) h, h ∈ C

y 0 (k) = u(t) k, k ∈ C
0
(py 0 ) (k) = (pu0 )(t) k, k ∈ C,

respectively. Note that here y 0 (h) denotes the derivative of y with respect to h, and
in (py 0 )0 (h) the outside prime denotes the derivative of the quasi-derivative (py 0 )
with respect to h. Thus the two primes in (py 0 )0 have different meanings in this
formula - the outside one is for differentiation with respect to h , the inside one
for diffentiation with respect to t - but since t is fixed here this should not cause
 33

confusion. Similar remarks apply to the formulas for differentiation with respect to
k.
Let K = [a1 , b1 ] be a compact subinterval of J. Each of the following maps from
C to the Banach space C(K):

h → y(·, c, h, k, 1/p, q, w, λ), h → (py 0 )(·, c, h, k, 1/p, q, w, λ),

k → y(·, c, h, k, 1/p, q, w, λ), k → (py 0 )(·, c, h, k, 1/p, q, w, λ)

is differentiable and its Frechet derivative is given by

y 0 (k) (g) = v g, g ∈ C(K)

(py 0 )0 (k) (g) = (pv 0 ) g, g ∈ C(K)

y 0 (k)(g) = u g, g ∈ C(K)

(py 0 )0 (k)(g) = (pu0 ) g, g ∈ C(K),

respectively.
PROOF: This is a straightforward consequence of the variation of parameters for-
mula and the definition of the Frechet derivative. See the above remarks about
notation. 2

To compute the derivatives of y and py 0 with respect to 1/p, q, w and λ we use

the fundamental matrix of the system representation of equation (3.8): Let

   
0 1/p 0 0
P = , W =
q 0 w 0

and define Φ(t, s, P, w, λ) = (φij ) to be the fundamental matrix of the system

Y 0 = (P − λW )Y determined by the initial condition Y (s) = I for each s ∈ J
where I is the identity matrix. (See Definition 2.5 for Φ.)

THEOREM 3.8. Let (3.3), (3.8), (3.9) hold, and let K be a compact subinterval
of J. Fix t, c ∈ K, h, k, λ ∈ C, 1/p, w ∈ L(K); then the maps q → y(t, q), w →
y(t, w), 1/p → y(t, 1/p) from L(K) to C as well as the map λ → y(t, λ) from C → C
are differentiable and their derivatives are given by

Z t
y 0 (t, q)(r) = − φ1,2 (t, s) y(t, s) r(s) ds, r ∈ L(K),
c

Z t
0
(py )(t, q)(r) = − φ2,2 (t, s) y(t, s) r(s) ds, r ∈ L(K),
c
34

Z t Z s 
0 0
y (t, 1/p)(r) = φ1,2 (t, s) q(s) (py )(x) r(x) dx ds
c c
Z t
+ (py 0 )(x) r(x) dx, r ∈ L(K),
c
Z t Z s 
0 0
(py )(t, 1/p)(r) = φ2,2 (t, s) q(s) (py )(x) r(x) dx ds, r ∈ L(K),
c c
Z t
y 0 (t, w)(r) = λ φ1,2 (t, s, w) y(s, w) r(s) ds, r ∈ L(K);
c
Z t
(py 0 )(t, w)(r) = λ φ2,2 (t, s, w) y(s, w) r(s) ds, r ∈ L(K);
c
Z t
y 0 (t, λ) = φ1,2 (t, s, λ) w(s) y(s, λ) ds, λ ∈ C.
c

PROOF: We prove some of these, the proofs of the others are similar. Let

−(py 0 )0 + qy = 0, y(c) = h, (py 0 )(c) = k;

−(pz 0 )0 + (q + r)z = 0, z(c) = h, (pz 0 )(c) = k.

Let x = z − y. Then

−(px0 )0 + qx = −r z, x(c) = 0, (px0 )(c) = 0.

From the variation of parameters formula it follows that

Z t
x(t) = φ1,2 (t, s)(−r(s)) z(s)ds.
c

Letting z = y + (z − y) we get

Z t Z t
z(t) − y(t) + φ1,2 (t, s)(r(s) y(s)ds = − φ1,2 (t, s)[ z(s) − y(s)] r(s) ds
c c
= o(r) as r → 0 in L(J).

The last equality follows from the fact that φ1,2 is bounded on K × K and z → y
uniformly on K by the furthermore part of Theorem 3.6. Similarly we get

Z t
0 0 0
(pz − py )(t) = (px )(t) = φ2,2 (t, s)(−r(s) z(s)ds
c
 35

and from this, proceeding as above, we obtain the formulas for y(t, q)(r) and for
(py 0 )(t, q)(r).
To derive the formulas for the derivatives with respect to 1/p we proceed as
follows. Let

1 1
= + r, r ∈ L(J)
pr p

and let y = y(t, 1/p), z = z(t, 1/pr ). Set

Z t
1
x(t) = (py 0 − pr z 0 ).
c p

Then

Z t
−(px0 )0 + qx = f, x(c) = 0, (px0 )(c) = 0, f (t) = −q(t) (pr z 0 ) r.
c

From the variation of parameters formula we get

Z t Z s 
0
x(t) = − φ1,2 (t, s) q(s) (pr z )(u) r(u)du ds,
c c

Z t Z s 
(px0 )(t) = − φ2,2 (t, s) q(s) (pr z 0 )(u) r(u)du ds.
c c

and note that

Z t
1 1
z(t) − y(t) = (pr z 0 ) − (py 0 )]
[
c pr p
Z t Z t
1 1
= [ (pr z 0 ) − (py 0 )] + (pr z 0 ) r
c p p c
Z t
= −x(t) + (pr z 0 ) r
c
Z t Z s  Z t
= φ1,2 (t, s) q(s) (pr z 0 )(u) r(u)du ds + (pr z 0 ) r.
c c c

Setting

pr z 0 = py 0 + [pr z 0 − py 0 ]

we obtain
36

Z t Z s  Z t
z(t) − y(t) − φ1,2 (t, s) q(s) (py 0 )(u) r(u)du ds + (py 0 ) r
c c c
Z t Z s  Z t
= φ1,2 (t, s) q(s) [(pr z 0 ) − py 0 ](u) r(u)du ds + [(pr z 0 ) − py 0 ] r
c c c
= o(r) as r → 0 in L(J).

The last equality follows from the boundedness of φ1,2 on K ×K, from q ∈ L(K),
and from the fact that, by Theorem 3.6, (pr z 0 ) → py 0 uniformly on K as r → 0 in
L(J). 2

There is an interesting and subtle point involved in the proof of Theorem 3.9:
1
p + r may be identically zero on a subinterval of J. Note that the solutions y
depend on p1 , not on p. Therefore p1 may be identically zero on a subinterval of J or
even on all of J. This is allowed by the existence-uniqueness Theorem 2.2 and the
subsequent theorems. However, the equation (3.8) has to be interpreted properly
in this case as Atkinson [3] has pointed out. In fact Atkinson uses the notation

1
−( y 0 )0 + qy = λ w y
p

for equation (3.8) but this notation has not been widely accepted.
For regular equations each solution y and its quasi-derivative py 0 are not only
entire functions of λ but have order at most 1/2.

THEOREM 3.9. Let (3.3), (3.9) hold. Assume that

(3.22) 1/p, q, w ∈ L(J).

Then every nontrivial solution y of (3.8) and its quasi-derivative py 0 are entire
functions of λ of order at most 1/2. More precisely, there exist positive constants
M, B, δ such that

√
|λ|
|y(t, λ)| ≤ B eM , a ≤ t ≤ b, |λ| ≥ δ
√
|(py 0 )(t, λ)| ≤ B eM |λ|
, a ≤ t ≤ b, |λ| ≥ δ

PROOF: Let v = py 0 then v 0 = (q − λw) y. Fix λ and let prime “0 ” denote differen-
tiation with respect to t. Then

[|λ| |y|2 + |v|2 ]0 = [|λ| ȳ y + v̄ v]0

1 1
= |λ| ( y v̄ + v ȳ) + v̄ (q − λw) y + v (q̄ − λ̄ w̄) ȳ.
p̄ p
 37

¿From this and the elementary inequality

|λ| |a|2 + |b|2

2 |a b| ≤ p , |λ| =
6 0
|λ|

we get

|λ| |y|2 + |v|2

 
2 2 0 1
[|λ| |y| + |v| ] ≤ p |λ| + |q| + |λ| |w|
|λ| |p|

and hence

p 1 1 p
[log (|λ| |y|2 + |v|2 )]0 ≤ |λ| + p |q| + |λ| |w|.
|p| |λ|

An integration yields

√ Rt 1 Rt
|λ| a ( |p| +|w|)+ √1 |q|
|λ| |y(t, λ)|2 + |v(t, λ)|2 ≤ C e |λ\ a

√
 
Z b √1
Rb
M |λ| 1 a
|q|
≤Be , 0<M = ( + |w|) < ∞, e |λ|
< B < ∞.
a |p|

This completes the proof. 2

3.5 Endpoint classifications: R, LC, LP, O, NO, LCNO, LCO.

Definition of regular (R), limit-circle (LC), limit-point (LP), oscillatory

(O), and nonoscillatory (NO) end-points.
Consider the equation

(3.23) −(py 0 )0 + qy = λ w y, λ ∈ C, on J,

with

(3.24) J = (a, b), −∞ ≤ a < b ≤ ∞, p, q, w : J → C, 1/p, q, w ∈ Lloc (J).

The (finite or infinite) endpoint a is regular if, in addition to (3.24),

1/p, q, w ∈ L(a, d)
38

holds for some (and hence any) d ∈ J; is limit-circle if all solutions of the equation
Rd
(3.23) are in L2w (a, d) = {f : (a, d) → C, a |f |2 w < ∞} for some (and hence any)
d ∈ (a, b); is LP if it is not LC; is O if there is a nontrivial solution with an infinite
number of zeros in any right neighborhood of a ; is NO if it is not O; is LCO if it
is both LC and O; and is LCNO if it is both LC and NO. Similar definitions are
made at b. An endpoint is called singular if it is not regular.
It is well known [81] that the LC, LP, classifications are independent of λ ∈ C
and that the LCO and LCNO classifications are independent of λ ∈ R. At an LP
endpoint the O classification, in general, depends on λ.

PROPOSITION 3.10. Let (3.23), (3.24) hold. In addition assume that

p, q, w : J → R, p > 0, a.e., λ ∈ R.

Then the zeros of every nontrivial solution y of (3.23) are isolated in the interior of
J and also at regular endpoints of J i.e. if a nontrivial solution y has a zero at a
regular endpoint of J then there is an appropriate one sided neighborhood of this
endpoint in which y has no other zero. Thus only the singular endpoints of J can
be accumulation points of zeros of y.
PROOF: Let y be a nontrivial solution of (3.23). First we show that if y has
consecutive zeros at c, d ∈ (a, b), c < d, then (py 0 )(h) = 0 for some h ∈ (c, d). We
have

Z d Z d Z d
0 1 1
0 = y(d) − y(c) = y = (py 0 ) = (py 0 )(h)
c c p c p

by the Mean Value Theorem for the Lebesgue integral. (Recall that (py 0 ) is con-
Rd
tinuous on J.) Hence either (py 0 )(h) = 0 or c p1 = 0, but the latter would imply
that p1 = 0 a.e. in (c, d) in contradiction to the hypothesis that p > 0 a.e. in (a, b).
Now to prove the main Proposition suppose there exists a sequence {tn ∈ (a, b) :
n ∈ N0 } such that tn → t0 and y(tn ) = 0, n ∈ N0 . Then y(t0 ) = 0 and from the
first part of the proof we get a sequence {sn : n ∈ N } with sn → t0 such that
(py 0 )(sn ) = 0. Since (py 0 ) is continuous in (a, b) it follows that (py 0 )(t0 ) = 0. But
y(t0 ) = 0 and (py 0 )(t0 ) = 0 implies that y is identically zero on J by the uniqueness
of initial value problems. 2

We end this subsection with an example to show that when p changes sign the
behavior of the classical and quasi-derivatives of a solution can be quite different.

EXAMPLE 3.11. Let

1
p(t) = , 0 < t ≤ 1.
cos(log(t))
 39

Then 1/p ∈ L(0, 1) and the equation

−(py 0 )0 = 0 on (0, 1)

has

Z
y= cos(log(t)) dt, v(t) = 1

as solutions. Note that the point 0 is an accumulation point of zeros of y 0 since

y 0 (tk ) = 0 where

tk = e−kπ/2 → 0, as k → ∞,

but (py 0 )(t) = 1 for t ∈ [0, 1]. The Wronskian

y v
W (y, v)(t) = (t) = −1, 0 ≤ t ≤ 1,
py 0 pv 0

but the classical Wronskian

y v
(t) = −cos(log(t)), 0 < t ≤ 1
y0 v0

is nonconstant with an accumulation point of zeros at 0.

3.6 The maximal domain and Lagrange form.

Let (3.9) hold and assume w > 0 a.e. on J; let

M y = [−(py 0 )0 + qy]

The maximal domain ∆ = ∆(M, w, J) is defined by

∆ = {y : J → C : y, py 0 ∈ ACloc (J), y, w−1 M y ∈ L2w (J)}

The Lagrange sesquilinear form is given by

[y, z] = ypz̄ 0 − z̄py 0 , y, z ∈ ∆.

40

3.7 Regularizing functions.

Here we construct a pair of functions u, v which we call “regularizing” functions

since they can be used to “regularize” singular equations with LCNO endpoints.

THEOREM 3.12. Let (3.3), (3.8), and (3.9) hold; and suppose that p, q, w are
real-valued and p > 0, w > 0 a.e. on J. Assume each endpoint is either regular
or LCNO. Let ∆ and M be defined as in subsection 6. Then there exist functions
u, v ∈ ∆ satisfying the following conditions:
1. They are real valued.
2. For some real λ = λa , u is a principal solution at a and v is a nonprincipal
solution at a.
3. For some real λ = λb , u is a principal solution at b and v is a nonprincipal
solution at b.
4. These functions u, v need not be solutions through the interior of (a, b), and,
in case λa = λb , they need not be the same solution near a and near b.
5. [u, v](a) = limt→a+ [u, v](t) = 1,
6. [u, v](b) = limt→b− [u, v](t) = 1,
7. v > 0 on J = (a, b).
PROOF: See Subsection 5.2 of Section 5 below for the definition of principal and
non-principal solution; and see Lemma 7 in Niessen and Zettl [74] for a proof. 2

We call such functions u and v “regularizing functions” of the equation (3.8), on

(a, b). The reason for this terminology will become clear in subsection 9 where we
show that with the help of such functions u, v, particularly v, one can construct a
regular equation which is “equivalent” in a natural sense to the singular equation
(3.8). We call equation (3.8) singular if at least one endpoint of the underlying
interval J is singular; the equation (3.8) is said to be regular if both endpoints are
regular.

3.8 Limit-Circle “Initial value problems”.

THEOREM 3.13. Let (3.3), (3.8), (3.9) hold. Assume p, q, w are real valued,
w > 0 a.e. and the left endpoint a is R or LC. Suppose u, v are real valued linearly
independent solutions on some interval (a, d] for some fixed real λ0 . Given any λ ∈ R
and any h, k ∈ R the singular initial value problem consisting of the equation

−(py 0 )0 + qy = λ w y on J

and the singular “initial condition”

[y, u](a) = h, [y, v](a) = k

has a unique real solution y on J. Similarly at b.

 41

PROOF: Since u, v are linearly independent solutions near a we have [u, v](a) 6= 0
and we can assume that [u, v](a) = 1. Let

 
u v
U= , Z = U −1 Y, Y 0 = (P − λW )Y, U 0 = (P − λ0 W )U,
pu0 pv 0
 
0 0
and, using the notation from (3.5) for P, Y, let with W = . Note that
w 0
U is a fundamental matrix solution for a fixed λ0 but Y is a vector solution for an
arbitrary λ ∈ R . A direct computation reveals that

Z 0 = (λ0 − λ)(U −1 W U )Z = (λ0 − λ) G Z on (a, d]

where

−v 2 w
 
−1 −uvw
G=U WU = ∈ L(a, d).
u2 w uvw

Note that G ∈ L(a, d) follows from the Cauchy-Schwarz inequality coupled with
the assumption that a is in the LC case. Hence by Theorem 2.10 all initial value
problems

Z 0 = (λ0 − λ) G Z on (a, d], Z(a) = C,

 
z1
have a unique solution. From Y = U Z, Z = we get
z2

u z1 + v z2 = y, (pu0 ) z1 + (pv 0 ) z2 = py 0 both on (a, d].

¿From Cramer’s rule we get

y(t) v(t)
z1 (t) = = [y, v](t), a < t < d
(py 0 )(t) (pv 0 )(t)

u(t) y(t)
z2 (t) = = −[y, u](t), a < t < d.
(pu0 )(t) (py 0 )(t)

By letting t → a we get z1 (a) = k, z2 (a) = −h. Since this holds for arbitrary
h, k the proof is complete. 2

REMARK 3. Note that the assumption p > 0 is not needed in Theorem 3.13,
nor have we assumed that a is LCNO but only that a is LC. The transformation
Z = U −1 Y transforms the singular scalar equation (3.8) into a first order regular
system.
42

3.9 Factorization of solutions near an LCNO endpoint.

A singular equation with an LCNO endpoint can be “regularized” using the function
v from a regularizing pair u, v of functions as defined in subsection 8.

THEOREM 3.14. Consider the equation (3.8) with J given by (3.3) and with

p, q, w : J → R, 1/p, q, w ∈ Lloc (J), p > 0, w > 0 a.e. λ ∈ C.

Assume that the left endpoint a is LCNO and let u, v be a pair of regularizing
functions at a i.e. on (a, d) for some d ∈ (a, b) as defined in section 7. Define

P = v 2 p, W = v 2 w, Q = w v M v on J,

and consider the equation

(3.25) −(P z 0 )0 + Qz = λ W z on J.

Then we have

1/P, Q, W ∈ L(a, d), a < d < b,

i.e. the equation (3.25) is regular at a and

1. If y is a solution of (3.8) on (a, d), then z = y/v is a solution of (3.25) on
(a, d). Conversely, if z is a solution of (3.25) on (a, d) then y = vz is a solution
of (3.8) on (a, d).
2. The limits

z(a) = lim+ z(t); (P z 0 )(a) = lim+ (P z 0 )(t)

t→a t→a

exist and are finite. Thus the solution z and its quasi-derivative (P z 0 ) can be
continuously extended to the (finite or infinite) endpoint a.
3. Note that v is independent of λ ∈ R but does depend on (M, w) i.e. on
1/p, q, w and on the endpoint a i.e. on some neighborhood (a, d) for d ∈ (a, b).
4. The one-to-one mappings y(t, λ) = v(t) z(t, λ) and (py 0 )(t, λ) = v(t) (P z 0 )(t, λ)
can be given more explicitly, using the notation (3.10), by

y(t, c, h, k, 1/p, q, w, λ) = v(t)z(t, c, h/v(c), kv(c) − h(pv 0 )(c), 1/P, Q, W, λ)

(py 0 )(t, c, h, k, 1/p, q, w, λ)

= (pv 0 )(t)(P z 0 )(t, c, h/v(c), kv(c) − h(pv 0 )(c), 1/P, Q, W, λ)
 43

for c ∈ J, and h, k ∈ C.
PROOF: See Niessen and Zettl [74]. Although the explicit formulas for the 1-1 map
y → vz are not given by these authors it can easily be obtained from there. 2

REMARK 4. Note that, in general, y and v do not exist at a. Thus we have

y
z(a) = (a); (P z 0 )(a) = (vpy 0 − ypv 0 )(a) = [v, y](a)
v

but neither the numerator y nor the denominator v , nor the individual terms in
(P z 0 )(a), can be evaluated separately at a. Of course there is an entirely analogous
theorem and remark for the endpoint at b. If each endpoint is either R or LCNO
then Theorem 3.14 holds on the entire interval J.

THEOREM 3.15. Let the notation of Theorem 3.14 hold. If each endpoint of
the interval J is either R or LCNO, then there exist a pair of regularizing functions
u, v defined on all of J such that the conclusions of Theorem 3.14 hold on the whole
interval J.
PROOF: See Niessen and Zettl [74]. 2

The next example illustrates Theorem 3.15.

EXAMPLE 3.16. For the classical Legendre equation

−((1 − t2 ) y 0 )0 = λ y on (−1, 1),

we have

u(t) = 1, −1 < t < 1,


−(1/2) ln((1 − t)/(1 + t)) 1/2 ≤ t < 1
v(t) = .
(1/2) ln((1 − t)/(1 + t)) −1 < t ≤ −1/2

Note that v > 0 near +1 and near -1. Every solution y can be factored as follows:

y(t, λ) = v(t) z(t, λ), −1 < t < 1, λ ∈ R

where z is continuous on the closed interval [−1, 1]. This shows, in particular, that
the asymptotic behavior of solutions of the Legendre equation satisfying a fixed
initial condition is independent of λ ∈ C .
44

3.10 Comments.

These are made for each Subsection separately.

1. In Theorem 3.2 the sufficiency of condition (3.4) is well known, the necessity
is due to Everitt and Race [25].
2. The results of Theorems 3.4, 3.5 and 3.6 are surely not new, but we don’t
know of a reference where they can be found in this generality.
3. Again, other than [59], we don’t know of a reference where the continuous
dependence of solutions of initial value problems on 1/p, q, w in the L1 norm
is established. The continuous dependence of solutions on initial conditions
i.e. on c, h, k is discussed by Hille [43].
4. The differentiable dependence of solutions and their quasi-derivatives on each
parameter as well as the formulas for the derivatives seems to be new. The
proof of the differentiable dependence of y and of py 0 on 1/p is due to Qingkai
Kong and published here for the first time with his permission.
Is y jointly differentiable in all its variables: c, h, k, 1/p, q, w ? Ditto for
py 0 . What are the derivatives ?
Theorem 3.9 is adapted from Atkinson [3]. It follows from the asymptotic
form of the eigenvalues of regular self-adjoint SLP that for p, q, w real-valued
p ≥ 0, w > 0 that the non-trivial solutions - as functions of λ - are of order
exactly 1/2.
Under what more general conditions on p, q, w are the non-trivial solutions
of exact order 1/2 ?
Under what general conditions (excluding trivial cases such as p = q =
w = 0) are the solutions of the SL equation of order zero as functions of λ
? Are there general classes of equations for which the solutions have order r,
for 0 < r < 1/2 as functions of λ ?
5. The definitions of R, LC, LP, O, NO, LCNO, LCO are standard except, as
pointed out in Remark 2, we classify an infinite endpoint as regular if 1/p, q, w
are integrable in a neighborhood of this point. This contrasts with the usual
practice. For a definition of oscillation of difference equations see [58].
The standard proof of the invariance of the L2w (J) solutions with respect
to λ is based on the variation of parameter formula [14].
The invariance of the LCO case with respect to real λ follows from the
spectral theory of ordinary differential operators, see [81]. In general for the
symmetric case i.e. p, q, w real and 1/p, q, w ∈ Lloc (a, d), a < d < b, p ≥
0, w > 0 there exists a σ0 , −∞ ≤ σ0 ≤ ∞, such that the equation (3.8) is O
at a for λ > σ0 and is NO for λ < σ0 . Examples show that for λ = σ0 the
equation can be O or NO. This “oscillation number” σ0 is also the starting
point of the essential (continuous) spectrum of every self-adjoint realization of
the equation. The case σ0 = −∞ is interpreted as meaning that the essential
spectrum is not bounded below; the case σ0 = ∞ means that the essential
spectrum is empty. The latter holds for all SLP for which each endpoint
is either R or LC since in this case the spectrum is discrete. The essential
but not discrete spectrum is the same for all self-adjoint realizations of the
equation (3.8). For proofs of these statements as well as further information
the reader is referred to [81].
 45

6. The definition of the maximal domain and of the Lagrange sesquilinear form
is standard. These play an important role in the theory of boundary value
problems.
7. This section is based on Niessen and Zettl [74].
8. The “system regularization” of LC endpoints based on the fundamental ma-
trix U is not new. It has been used by Fulton and by Fulton and Krall. It
has other applications besides the one given here to singular IVP:
(a) It can be used to prove the LC invariance with repect to λ : ¿From

y(t, λ) = u(t, λ0 ) z1 (t, λ) + v(t, λ0 ) z2 (t, λ)

and zj (t, λ) → zj (a, λ) as t → a, j = 1, 2 it follows that zj are bounded

(“nearly constant”) in a neighborhood of a. Hence y(·, λ) is in L2w (a, d)
,a < d < b, since both u, and v are; this for every λ ∈ C . This proof
is much simpler than the standard one based on variation of parameters
and seems to be new.
(b) In connection with boundary value problems the transformation Z =
U −1 Y has been used by Fulton and by Fulton and Krall.
Another application of the transformation Z = U −1 Y is to prove the
invariance of the O classification with respect to all real λ at an LC
endpoint:
Assume a is LC and O for some real λ0 and p > 0. For any real λ we
have

y(t, λ) = u(t, λ0 ) z1 (t, λ) + v(t, λ0 ) z2 (t, λ)

case 1. zj (a) > 0, j = 1, 2. It is easy to see that between any three

zeros of u there must be a zero of y : Sketch the graph of u, v keeping
in mind the Sturm Separation Theorem. Now it is easy to identify two
disjoint intervals on one of which y is positive and on the other negative.
Hence y must have a zero between these intervals. The other cases are
all established similarly. This “oscillation theory proof ” is much simpler
than the spectral theory proof and seems to be new.
9. Theorem 3.14 is taken from Niessen and Zettl [74] and Theorem 3.15 is also
adapted from this paper although phrased in terms of factoring rather than
regularizing.
Example 3.15 illustrates Theorem 3.14 by showing that the classical Le-
gendre equation, which is singular at both endpoints, is “equivalent” to a
regular equation on the same interval. The leading coefficient P and the
weight function W of the regular equation are not bounded on (−1, 1) but,
nevertheless, satisfy the regularity conditions 1/P, W ∈ L(−1, 1). The Le-
gendre equation and its regularization are equivalent in the sense that the
transformation

y(t, λ) = v(t) z(t, λ)

46

(λ0 = 0 in this case) maps solutions y of the Legendre equation into solutions
z of the regular equation in a 1-1 onto manner. All solutions z of the regular
equation are continuous on the closed interval [−1, 1]. So the singular behavior
is contained in the transformation function v. Since v is independent of λ this
shows that the singular behavior of the Legendre equation is independent of
λ ∈ C . The invariance of the LC classification with respect to λ ∈ C and
the invariance of the LCNO classifications with respect to λ ∈ R are merely
specific instances of this general invariance property.
Of course, these remarks apply to all other equations where each endpoint
is either regular or LCNO.

4 REGULAR TWO POINT BOUNDARY VALUE PROBLEMS

4.1 Introduction.

In this section we study regular Sturm-Liouville problems (SLP) with self-adjoint

and non-self-adjoint two point boundary conditions (BC). There are two basic meth-
ods available for such a study: operator theory and complex function theory. Both
are employed here, singly and in combination.

4.2 Characterization of the eigenvalues.

A regular two point SLP consists of the equation

(4.1) −(py 0 )0 + qy = λ w y on (a, b) = J, −∞ ≤ a < b ≤ ∞,

where

(4.2) 1/p, q, w : J → C, 1/p, q, w ∈ L(J), λ ∈ C,

together with boundary conditions

 
y
(4.3) AY (a) + BY (b) = 0, Y = , A, B ∈ M2 (C).
py 0

By Theorem 3.4, Y (a), Y (b) exist and are finite so that (4.3) is well defined. Let

   
0 1/p 0 0
(4.4) P = , W = .
q 0 w 0

Then the scalar equation (4.1) is equivalent with the first order system
 47

   
0 1/p y
(4.5) Y 0 = (P − λW )Y = Y, Y = .
q − λw 0 py 0

Let Φ(·, u, P, w, λ) be the matrix solution of the initial value problem

(4.6) Φ0 = (P − λW ) Φ, Φ(u) = I, u ∈ J, λ ∈ C,

and define the characteristic function ∆, see Lemma1 below, by

(4.7) ∆(λ) = det[A + B Φ(b, a, P, w, λ)], λ ∈ C.

Although the notation ∆ was used in section 3 to denote the maximal domain
there should be no confusion with its use here as the characteristic function.

LEMMA 4.1. Let (4.1) to (4.7) hold. Then the characteristic function ∆ is de-
fined and continuous at a and b for fixed P, w, λ and is an entire function of λ for
fixed a, b, P, w.
PROOF: It follows from Theorems 2.9 and 2.10 that, for fixed P, w, λ, ∆(a, b) exists
and is continuous at a and b. The entire dependence on λ follows from the second
proof using successive approximations of Theorem 2.2, the existence-uniqueness
theorem. Each successive approximation is a polynomial in λ. Since these converge
uniformly on each compact subset K of the complex plane to the solution, this
solution is analytic on K. Thus the solution is entire in λ since this holds for each
such K. 2

LEMMA 4.2. Let (4.1) to (4.7) hold. A complex number λ is an eigenvalue of

the BVP (4.1), (4.2), (4.3) if and only if ∆(λ) = 0. Furthermore the multiplicity
of the eigenvalue λ is equal to the number of linearly independent vector solutions
C = Y (a) of the linear algebra system

(4.8) [A + B Φ(b, a, λ)] C = 0.

PROOF: Suppose ∆(λ) = 0. Then (4.8) has a nontrivial vector solution for C. Let
Y (a) = C and solve the IVP

Y 0 = (P − λW )Y, Y (a) = C, on J.

Then

Y (b) = Φ(b, a, λ) Y (a) and [A + B Φ(b, a, λ)] Y (a) = 0.

48

From this it follows that the top component of Y, say, y is an eigenvector of the
BVP (4.1), (4.2), (4.3); that means λ is an eigenvalue ofthis BVP.  Conversely, if
y
λ is an eigenvalue and y an eigenvector of λ, then Y = satisfies Y (b) =
py 0
Φ(b, a, λ) Y (a) and consequently [A + B Φ(b, a, λ)] Y (a) = 0. Since Y (a) = 0 would
imply that y is the trivial solution in contradiction to it being an eigenvector, we
have that det[A+B Φ(b, a, λ)] = 0. If (4.8) has two linearly independent solutions for
C, say C1 , C2 , then solve the IVP with the initial conditions Y (a) = C1 , Y (a) = C2
to obtain solutions Y1 , Y2 . Then Y1 , Y2 are linearly independent vector solutions of
(4.5) and their top components y1 , y2 are linearly independent solutions of (4.1).
Conversely, if y1, y2 are linearly dependent solutions of (4.1) we can reverse the
steps above to obtain two linearly independent solutions of the algebraic system
(4.8).

LEMMA 4.3. For the BVP (4.1), (4.2), (4.3) exactly one of the following four
cases holds:
1. There are no eigenvalues in C .
2. Every complex number is an eigenvalue.
3. There are a nonzero finite number of eigenvalues in C .
4. There are an infinite but countable number of eigenvalues in C and these have
no finite accumulation point in C.
PROOF: This follows directly from Lemmas 4.1 and 4.2 and the fact that the zeros
of an entire function are isolated and have no accumulation point in the finite plane
C. 2

It is convenient to separate the boundary conditions (BC) (4.3) into two mutually
exclusive classes: separated and coupled. Note that, since the BC are homogeneous,
multiplication by a nonzero constant or a nonsingular matrix leads to equivalent
boundary conditions.

LEMMA 4.4 (Separated BC). Let (4.1) to (4.7) hold. Fix P, W, J and assume

   
A1 A2 0 0
(4.9) A= , B= .
0 0 B1 B2

Then

∆(λ) = A2 B1 φ11 (b, a, λ) + A2 B2 φ21 (b, a, λ) − A1 B1 φ12 (b, a, λ) − A1 B2 φ22 (b, a, λ)

for λ ∈ C.
PROOF: This follows directly from the definition of ∆. 2
 49

LEMMA 4.5 (Coupled self-adjoint BC). Let (4.1) to (4.7) hold. Fix P, W, J and
assume that

 
−1 0
(4.10) B= , A = eiα K, −π ≤ α ≤ π, K ∈ SL2 (R),
0 −1

i.e. K is a real 2 × 2 matrix with determinant 1. Let K = (kij ) and define

(4.11)
D(λ, K) = k11 φ11 (b, a, λ) + k12 φ21 (b, a, λ) + k21 φ12 (b, a, λ) + k22 φ22 (b, a, λ)

for λ ∈ C. Note that D(λ, K) does not depend on α. Then

1. the complex number λ is an eigenvalue of BVP (4.1), (4.2), (4.3), (4.10)
if and only if

(4.12) D(λ, K) = −2 cos α, −π ≤ α ≤ π;

2. for fixed K a complex number λ is an eigenvalue for α if and only if it is

an eigenvalue for −α and if u is an eigenfunction of α then its conjugate
ū is an eigenfunction of −α.
PROOF: By Theorem 2.3 det Φ(b, a, λ) = 1. This and det K = 1 together with a
straightforward but tedious computation yields that

∆(λ) = 1 + e2iα + eiα D(λ, K).

Thus ∆(λ) = 0 is equivalent to

D(λ, K) = −eiα − eiα = −2 cos α.

REMARK 5. Although the matrices (4.10) determine self-adjoint boundary con-

ditions (they are the canonical form of all coupled self-adjoint BC) , no conditions
other than (4.2) are assumed on p, q, w in Lemma 4.5. In particular no symme-
try (formal self-adjointness) or definiteness assumption is made on equation (4.1).
Thus the characterization of the eigenvalues given by (4.12) applies not only to so
called left-definite, right-definite and indefinite SLP but the coefficients p, q and the
weight function w can be complex valued.

Eberhard and Freiling [20], Mennicken and Möller [67], have established the
existence of infinitely many eigenvalues for the symmetric equation (4.1) but with
non-self-adjoint BC. We have
50

THEOREM 4.6. Assume that

p, q, w : J = [a, b] → R; −∞ < a < b < ∞, p > 0, w > 0,

p = w ∈ AC(J), p0 /p, q/w ∈ Lr (J), 1 < r ≤ ∞.

Then
1. the BVP (4.1), (4.9) with (A1 , A2 ) 6= (0, 0) 6= (B1 , B2 ) has an infinite number
of eigenvalues;
2. the BVP (4.1) together with the coupled BC

Y (b) = A Y (a)

has an infinite number of eigenvalue for any A ∈ M2 (C) satisfying either

a12 6= 0 or a12 = 0 and a11 + a22 6= 0.
PROOF: See [67], [20]. 2

These authors also establish an expansion theorem for these cases.

4.3 The Fourier equation.

We now pause to consider the simplest SLP for at least two reasons: (i) to illustrate
the results of the previous section and (ii) to indicate some of the coming attractions
of section 5. It is remarkable how many properties of SLP for the simplest SL
equation

−y 00 = λ y

hold for the general case. This is a third reason for discussing this equation here.
Consider the equation

(4.13) −y 00 = λ y on (a, b), −∞ ≤ a < b ≤ ∞, λ ∈ C.

We include the case when one or both endpoints are infinite here, even though this
is a singular problem then (p = 1 = w are not in L(J) if J is unbounded) and
singular problems are not discussed, in general, until the next section, to highlight
the interplay between regular and singular problems.
We content that to fully understand regular SLP requires a perspective which
includes the singular case.
Each infinite endpoint is in the LP case since the constant 1 is a non L2 solution
for λ = 0. Thus there is one and only one self-adjoint realization, say S, of the
equation (4.13) in the space L2 (−∞, ∞). The spectrum of S, σ(S), contains no
 51

eigenvalues and thus coincides with the essential (continuous) spectrum σe (S); we
have

σ(S) = σe (S) = [0, ∞).

In this case

   
0, 1 0, 0
P = , W = .
0, 0 1, 0

Since these are fixed for this example we will omit them in the notation for Φ. This
fundamental matrix Φ = Φ(t, u, P, W, λ) = Φ(t, u, λ) is determined as the unique
solution of the initial value problem

Φ0 = (P − λW ) Φ, Φ(u) = I, t, u ∈ R, λ ∈ C.

√
To compute Φ we choose an analytic branch of the square root function z as
follows :

√
µ= λ = s + it, s > 0, t > 0, f or λ 6= 0,

and obtain

 1 
cosh(iµ(t − u)), iµ sinh(iµ(t − u))
(4.14) Φ(t, u, λ) = ,
iµ sinh(iµ(t − u)), cosh(iµ(t − u))
√
t, u ∈ R, λ ∈ C, λ 6= 0, µ = λ,

and

 
1, t − u
(4.15) Φ(t, u, 0) = , t, u ∈ R.
0, 1

Note that for fixed t, u ∈ R the fundamental matrix Φ(t, u, λ) is analytic at λ for
each λ ∈ C including λ = 0. (This can be confirmed from the series expansions of
the hyperbolic sinh and cosh functions.)
For the convenience of the reader we now recall some definitions and properties
of hyperbolic functions which will be used below.
1. 2 sinh z = ez − e−z z ∈ C, 2 cosh z = ez + e−z , z ∈ C, tanh z = sinh z/ cosh z,
2. sinh z = −i sin iz, cosh z = cos iz, tanh z = −i tan iz
3. sinh(z + 2kπi) = sinh z, cosh(z + 2kπi) = cosh z
4. (sinh z)0 = cosh z, (cosh z)0 = sinh z
52

5. sin z = 0 if and only if z = kπ, k ∈ Z; cos z = 0 if and only if z = (2k +

1)π/2, k ∈ Z
6. The general solutions of the equations sin x = z, cos x = z, tan x = z are
given by, respectively,
7. x = (−1)k arcsin z + kπ, k ∈ Z,
8. x = ± arccos z + 2kπ, k ∈ Z,
9. x = arctan z + kπ, k ∈ Z,
10. For −1 ≤ t ≤ 1, arcsin t, arccos t are real and −π/2 ≤ arcsin t ≤ π/2, 0 ≤
arccos t ≤ π.
Now let −∞ < a < b < ∞ and consider the two point boundary condition

(4.16) A Y (a) + B Y (b) = 0, A, B ∈ M2 (C).

We now discuss coupled and separated boundary conditions separately.

• Coupled BC. Let

   
1, 0 c, 0
(4.17) A= , B= .
0, 1 0, d

From (4.7), (4.14), (4.15), (4.17) we get

√
(4.18) ∆(λ) = 1 + cd + (c + d) cosh(iµ(b − a)), µ = λ 6= 0.

In all of the examples below the case λ = 0 needs to be checked separately

since λ = 0 plays a special role in these formulas.
We now consider a number of special cases of (4.17).
1. c = −d. Then ∆(λ) = 1−c2 , a constant independent of λ. If this constant
is zero then every complex number is an eigenvalue; if this constant is not
zero, then no complex number is an eigenvalue. In particular we have
(a) For c = 1, d = −1 every complex number is an eigenvalue.
(b) For c = −1, d = 1 every complex number is an eigenvalue.
(c) For c ∈ C , c 6= 1, c 6= −1, no complex number is an eigenvalue.
2. c 6= −d. The characteristic equation for the eigenvalues is :

1 + cd
cosh((iµ(b − a)) = cos(−µ(b − a)) = − = r.
c+d

The roots of this equation are given by

−µ(b − a) = ± arccos r + 2kπ, k ∈ Z,

Z 1
dt
arccos z = = π/2 − arcsin z,
(1 − t2 )1/2
Zz z
dt
arcsin z = ,
0 (1 − t2 )1/2
 53

and both integrals must be taken along a path which does not cross the
real axis.
When r is real and −1 ≤ r ≤ 1 then the roots for µ(b − a) are real and
we get

µ(b − a) = ∓ arccos r − 2kπ, k ∈ Z

From this and µ = s > 0 we get

µ(b − a) = arccos r + 2kπ, k ∈ N0 .

We now consider some special cases of this case :

(a) c = i = d. Here r = 0 and

µ(b − a) = π/2 + 2kπ, k ∈ N0

and the eigenvalues are

(π/2 + 2nπ)2
λn = , n ∈ N0 .
(b − a)2

(b) c = −1 = d. This is the self-adjoint periodic (P) case with r = 1.

First note that λ = 0 is an eigenvalue for this case. Our search for the
other eigenvalues leads to µ(b − a) = 2kπ, k ∈ N and consequently

(2nπ)2
λP
n = , n ∈ N0 .
(b − a)2

(c) c = 1 = d. This is the self-adjoint semi-periodic (S) case and gives r =

−1. We note that λ = 0 is not an eigenvalue in this case. Proceeding
as above we get

µ(b − a) = π + 2kπ = (2k + 1)π, k ∈ N0 .

Hence the eigenvalues in this case are given by

((2n + 1)π)2
λSn = , n ∈ N0 .
(b − a)2
54

(d) c = 1/d, c ∈ R , c 6= 0. This gives r = −(1 + 1)/(c + 1/c). Since

c + 1c > 2 if c > 0, and c 6= 1 ; c + 1c < −2 if c < 0 and c 6= −1 , we
have −1 < r < 1. Let

t0 (c) = arccos(r), 0 < t0 (c) < π.

Then the roots are given by

(b − a)µ = t0 (c) + 2kπ, k ∈ N0

and so the eigenvalues are

(t0 (c) + 2nπ)2

λn (c) = , n ∈ N0
(b − a)2

(e) c = eiα = d, 0 < α < π. Here

1 + e2iα
r=− = − cos α, t0 (α) = arccos(− cos α) = π − α ∈ (0, π).
2eiα

So

(b − a)µ = π − α + 2kπ, k ∈ N0 ,

and therefore

(π − α + 2nπ)2
λn (α) = , n ∈ N0 .
(b − a)2

• Separated BC. Here we take

   
A1 , A2 0, 0
A= , B= , Aj , Bj ∈ C, j = 1, 2.
0, 0 B 1 , B2

A direct calculation gives

∆(λ) = A1 B1 Φ12 (b, a, λ) + A1 B2 Φ22 (b, a, λ) − A2 B1 Φ11 (b, a, λ)

−A2 B2 Φ21 (b, a, λ)
1
= (A1 B2 − A2 B1 ) cosh((iµ(b − a)) + ( A1 B1 − iµA2 B2 ) sinh((iµ(b − a))
µ

Note that ∆(λ) is periodic with fundamental period 2kπi, so if there is

one eigenvalue then there is a countable infinity of them. For each eigenvalue
 55

there is only one linearly independent eigenvector. We now consider some

special cases. For each case we list the characteristic equation, its roots for
µ and the corresponding eigenvalues. As in the earlier cases λ = 0 has to be
checked independently.
1. A1 B2 − A2 B1 = 0. Note that this includes both the Dirichlet and Neu-
mann boundary conditions. Here we have

1
∆(λ) = ( A1 B1 − iµA2 B2 ) sinh((iµ(b − a)).
µ

To find all eigenvalues we proceed as follows: (i) we find all eigenvalues

produced by the roots of the sinh factor, (ii) check to see if any of these
roots yield a non-zero root of the first factor, and (iii) check separately
to see if λ = 0 is an eigenvalue. Clearly the first factor can produce at
most one eigenvalue and that only in exceptional cases.

sinh((iµ(b − a)) = −i sin((−µ(b − a)) = 0

−µ(b − a) = (−1)k arcsin(0) + kπ = kπ, k ∈ Z

Since µ = s > 0 we get the following eigenvalues from the periodic factor:

(nπ)2
λn = , n ∈ N.
(b − a)2

2. A1 = 1 = B1 , A2 = 0 = B2 . For these Dirichlet BC λ = 0 is not an

eigenvalue and also the first factor does not produce an eigenvalue; hence
all the eigenvalues are given by

(nπ)2
λn = , n ∈ N.
(b − a)2

3. A1 = 0 = B1 , A2 = 1 = B2 . These are the Neuman (N) BC.

−iµ sinh((iµ(b − a)) = µ sin(−µ(b − a)) = 0

µ(b − a) = kπ, k ∈ Z.

Since λ = 0 is also an eigenvalue in this case and the first factor does not
produce an eigenvalue we get that

(nπ)2
λN
n = , n ∈ N0 .
(b − a)2
56

4. A1 = 1 = B2 , A2 = 0 = B1 .

cosh((iµ(b − a)) = cos(−µ(b − a)) = 0

−µ(b − a) = ± arccos(0) + 2kπ, k ∈ Z,

(π/2 + 2nπ)2
λDN
n = , n ∈ N0 .
(b − a)2

5. A1 = 0 = B2 , A2 = 1 = B1 .

− cosh((iµ(b − a)) = − cos((−µ)(b − a)) = 0.

Here we have the same roots and hence the same eigenvalues as in the
previous case

(π/2 + 2nπ)2
λN
n
D
= , n ∈ N0 .
(b − a)2

6. A1 B2 − A2 B1 6= 0.

( µ1 A1 B1 − iµA2 B2 )
coth(iµ(b − a)) = .
A1 B2 − A2 B1

The roots of this equation are not so easy to find explicitly since the unknown µ
appears on both sides. However, numerical approximations can be obtained from
a root finder code.
Some observations
1. Let {ak : k ∈ N } be a decreasing sequence to −∞; {bk : k ∈ N }an increasing
sequence to +∞ and let

E = {λD
n (ak , bk ) : n ∈ N0 , k ∈ N.

Then E is dense in [0, ∞). Thus every point of the essential spectrum of the
self-adjoint realization S of the Fourier equation on (−∞, ∞) is the limit of
a sequence of eigenvalues of regular problems on the intervals (ak , bk ), k ∈
N. This illustrates a general result, to be discussed in section 5, about the
approximation of the spectrum of singular SLP by eigenvalues of a sequence
of regular SLP.
2. For any n ∈ N0 the Dirichlet eigenvalue λD n is greater than or equal to λn for
any other self-adjoint boundary condition.
3. λD
n (a, b) → 0 as (b − a) → ∞ for each n ∈ N0 .
4. λD
n (a, b) → ∞ as (b − a) → 0 for each n ∈ N0 .
 57

5. λD0 (a, b, q) → ∞ as (b − a) → 0 for q(t) = c, a constant. But

6. λN0 (a, b, q) → c as (b − a) → 0 for q(t) = c, a constant. Thus, as the length of
the interval shrinks to zero the difference between the Dirichlet and Neuman
eigenvalues goes to infinity. This is a general phenomenon of SLP found
recently by Kong and Zettl, see [55], [57].

4.4 The space of regular Boundary Value Problems.

We want to show that if two SLP are “close” to each other then their eigenvalues
and eigenfunctions are also “close” to each other. To study the “closeness” of two
BVP we introduce a “boundary value problem space” with a distance function. Let

0 0 0 0
J = (a , b ), −∞ ≤ a < b ≤ ∞,

(4.19) Ω = {ω = (a, b, A, B, 1/p, q, w)}

such that

0 0
−∞ ≤ a < a < b < b ≤ ∞, A, B ∈ M2 (C), p, q, w : J → C, 1/p, q, w ∈ Lloc (J).

By an eigenvalue of ω ∈ Ω we mean an eigenvalue of the BVP determined

by ω. Note that we have changed the notation for the endpoints of the interval J
from a to a0 and b to b0 ; this is so that we can use a and b to denote endpoints of
varying subintervals of J.
We want to show that the eigenvalues and eigenfunctions depend continuously
on the problem, i.e. a small change of the problem results in a small change of each
eigenvalue and each eigenfunction. This means we have to compare the spectrum
of different problems which may be defined on different intervals. Each ω ∈ Ω
determines a unique SL problem: a, b the interval, A, B the boundary condition,
and the restrictions of p, q, w on [a, b] the equation. Observe that the values of
p, q, w outside the interval [a, b], i.e. in (a0 , b0 ) \ [a, b], do not affect the spectrum
of the problem determined by ω. To account for this and to facilitate comparisons
between eigenvalues of problems defined on different intervals we let

(4.20) Ω̃ = {ω̃ = (a, b, A, B, 1/p,

g q̃, w̃)}

where


q on [a, b]
(4.21) q̃ =
0 otherwise

and 1/p,
g w̃ are defined similarly. Now we introduce the Banach space
58

(4.22) X = R × R × M2,2 (C) × M2,2 (C) × L1 (a0 , b0 ) × L1 (a0 , b0 ) × L1 (a0 , b0 )

with its “natural” norm

Z b0  
(4.23) ||ω|| = ||ω̃|| = |a| + |b| + ||A|| + ||B|| + |1/p|
g + |q̃| + |w̃|
a0

where ||A|| is any fixed matrix norm.

This space X is a “natural” setting for the study of regular SL problems. Note
that, since 1/p, q, w are only assumed to be in Lloc (a0 , b0 ), Ω is not a subset of X
g q̃, w̃ are in L1 (a0 , b0 ). Now we identify Ω with Ω̃ as a subset of
but Ω̃ is since 1/p,
X. Then Ω inherits the norm from X, and the convergence in Ω is determined by
this norm. It is easy to see that every point in Ω is an accumulation point of Ω
with respect to this norm in X.

4.5 Continuity of Eigenvalues and Eigenfunctions.

Are the eigenvalues of a regular SLP continuous functions of the problem? The
answer is YES and NO.
NO, because for a fixed index n, the n-th eigenvalue λn of a self-adjoint SLP is,
in general, not a continuous function of the problem i.e. of ω.
YES, because every isolated eigenvalue can be embedded in a “continuous eigen-
value branch”.
In this section these statements will be made precise. Also the continuous de-
pendence of the eigenfunctions on the problem is studied here.
Consider the SLP consisting of the equation

(4.24) −(py 0 )0 + qy = λ wy, on (a, b), −∞ < a < b < ∞,

together with separated boundary conditions

(4.25) A1 y(a) + A2 (py 0 )(a) = 0, (A1 , A2 ) 6= (0, 0), A1 , A2 ∈ R,

(4.26) B1 y(b) + B2 (py 0 )(b) = 0, (B1 , B2 ) 6= (0, 0), B1 , B2 ∈ R,

and coefficients satisfying

(4.27) p, q, w : (a, b) → R, 1/p, q, w ∈ L(a, b), p > 0, w > 0, a.e. on (a, b).
 59

It is well known that this problem is self-adjoint and has an infinite but countable
number of eigenvalues {λn : n ∈ N0 }, these are all real, simple, bounded below and
can be indexed to satisfy

−∞ < λ0 < λ1 < λ2 < . . . ; and λn → ∞ as n → ∞.

THEOREM 4.7. Let (4.24) to (4.27) hold. Fix a, b, p, q, w.

• Fix B1 , B2 and let A1 = 1. Consider λn = λn (A2 ) as a function of A2 ∈ R.
Then for each n ∈ N0 , λn (A2 ) is continuous at A2 for A2 > 0 and A2 < 0 but
has a jump discontinuity at A2 = 0. More precisely we have:
1. λn (A2 ) → λn (0) as A2 → 0− , n ∈ N0 .
2. λ0 (A2 ) → −∞ as A2 → 0+ .
3. λn+1 (A2 ) → λn (0) as A2 → 0+ .
• Fix A1 , A2 and let B1 = 1. Consider λn = λn (B2 ) as a function of B2 ∈ R.
Then for each n ∈ N0 , λn (B2 ) is continuous at B2 for B2 > 0 and B2 < 0 but
has a jump discontinuity at B2 = 0. More precisely we have:
1. λn (B2 ) → λn (0) as B2 → 0+ , n ∈ N0 .
2. λ0 (B2 ) → −∞ as B2 → 0− .
3. λn+1 (B2 ) → λn (0) as B2 → 0− .
PROOF. See Everitt, Möller and Zettl [23]. 2

REMARK 6. Note that λ0 (A2 ) has an infinite jump discontinuity at A2 = 0, but

for all n ≥ 1, λn (A2 ) has a finite jump discontinuity at A2 = 0, λn (A2 ) is left but
not right continuous at 0. Similarly, λ0 (B2 ) has an infinite jump discontinuity at
B2 = 0, but for all n ≥ 1, λn (B2 ) has a finite jump discontinuity at B2 = 0; λn (B2 )
is right but not left continuous at 0. In all cases λn (0) is embedded in a continuous
branch of eigenvalues as A2 or B2 passes through zero but this branch is not given
by a fixed index n; the index “jumps” from n to n + 1 as A2 or B2 pass through
zero from the appropriate direction.

REMARK 7. This forced “index jumping” in order to stay on a continuous

branch of eigenvalues plays an important role in some of the algorithms and their
numerical implementations used in the code SLEIGN2 [9] for the numerical approx-
imation of the spectrum of regular and singular SLP.

REMARK 8. Kong and Zettl [57] have shown that each continuous eigenvalue
branch is in fact differentiable everywhere including the point A0 (or B0 ) where the
index jumps. This also follows from Möller and Zettl [69].

EXAMPLE 4.8. Consider the BVP with Dirichlet BC and the equation

−(pε y 0 )0 = λ y on (0, 1),

60

where ε ∈ [0, 1] and


−1, if 0 ≤ t ≤ ε,
pε (t) =
1, ε < t ≤ 1.

Then for ε = 0 the spectrum is bounded below but for each ε > 0 the spectrum is
unbounded below. Note that 1/pε → 1/p0 in L(0, 1).

LEMMA 4.9 (Continuity of the roots as functions of parameters). Let A be an

open set in C , F a metric space, f a continuous complex valued function on A × F
such that for each α ∈ F , the map z → f (z, α) is an analytic function on A . Let
B be an open subset of A whose closure B̄ in C is compact and contained in A, and
let α0 ∈ F be such that no zero of f (z, α0 ) is on the boundary of B. Then there
exists a neighborhood W of α0 in F such that :
1. For any α ∈ W, f (z, α) has no zero on the boundary of B;
2. for any α ∈ W, the sum of the orders of the zeros of f (z, α) contained in B is
independent of α.
PROOF. See page 248 in Dieudonné [16] . 2

THEOREM 4.10. Let ω0 = (a0 , b0 , A0 , B0 , 1/p0 , q0 , w0 ) ∈ Ω. Let λ(ω0 ) be an

isolated eigenvalue of the SL problem (4.2), (4.2), (4.3). Then, given any  > 0,
there exists a δ > 0 such that if ω = ω(a, b, A, B, 1/p, q, w) ∈ Ω satisfies

||ω − ω0 || = |a − a0 | + |b − b0 | + ||A − A0 || + ||B − B0 ||+

Z b0  
1 1
(4.28) + | − | + |q̃ − q̃0 | + |w̃ − w̃0 | < δ
a0 p̃ p˜0

then ω has an eigenvalue λ(ω) satisfying

(4.29) |λ(ω) − λ(ω0 )| < .

Furthermore, if λ(ω0 ) is simple then there is exactly one λ(ω) satisfying (4.29); if
λ(ω0 ) is a double eigenvalue, then either (4.29) holds for a double eigenvalue of ω
or for exactly two simple ones.
PROOF. See Kong and Zettl [57] Theorem 3.1. 2

REMARK 9. Besides establishing the continuity of the eigenvalues within con-

tinuous eigenvalue branches for self-adjoint and non-self-adjoint SLP, Theorem 4.10
is also an existence theorem for eigenvalues. Since any self-adjoint SLP ω0 is known
to have a countably infinite number of isolated eigenvalues, we can conclude from
 61

Theorem 4.10 that every SLP ω which is sufficiently close to ω0 , whether it is self-
adjoint or not, must have n eigenvalues close to n eigenvalues of λ(ω0 ), for any
positive integer n. Note that it does not follow directly from Theorem 4.10 that ω
must have an infinite number of eigenvalues close to those of ω0 .

REMARK 10. Below, each eigenvalue will be assumed to be embedded

in a continuous eigenvalue branch in the sense of Theorem 4.10.

By a normalized eigenfunction u of an SL problem we mean an eigenfunction u

that satisfies

Z b
(4.30) |u|2 w = 1.
a

Next we state a result for normalized eigenfunctions. Note that these are not
uniquely determined. In the case of a simple eigenvalue they are unique up to
sign, but for a double eigenvalue there are pairs of linearly independent normalized
eigenfunctions.
Below when considering SLP on two subintervals of J we extend the solutions
and their quasi-derivatives continuously to the whole of the open interval J. This
is done to facilitate comparisons between solutions and their quasi-derivatives on
different intervals.

THEOREM 4.11. Let the notation and hypotheses of Theorem 4.10 hold.
(i) Assume the eigenvalue λ(ω0 ) is simple for some ω0 ∈ Ω and let u(·, ω0 ) denote
a normalized eigenfunction of λ(ω0 ). Then there is a neighborhood M of ω0 in Ω
such that λ(ω) is simple for every ω ∈ M and there exist normalized eigenfunctions
u(·, ω) of λ(ω) for ω ∈ M such that

(4.31) u(·, ω) → u(·, ω0 ), (pu0 )(·, ω) → (pu0 )(·, ω0 ), as ω → ω0 in Ω,

both uniformly on any compact subinterval K of (a0 , b0 ).

(ii) Assume that λ(ω) is a double eigenvalue for all ω in some neighborhood M
of ω0 in Ω. Let u(·, ω0 ) be any normalized eigenfunction of λ(ω0 ). Then there exist
normalized eigenfunctions u = (·, ω) of λ(ω) such that

(4.32) u(·, ω) → u(·, ω0 ), (pu0 )(·, λ) → (pu0 )(·, ω0 ), as ω → ω0 in Ω,

both uniformly on any compact subinterval K of (a0 , b0 ). Note that in this case,
given two linearly independent normalized eigenfunctions uj of λ(ω0 ) there ex-
ist a pair uj of linearly independent normalized eigenfunctions of λ(ω) such that
uj (·, ω) → uj (·, ω0 ) as ω → ω0 in Ω for j = 1, 2.
62

PROOF. See [57], Theorem 3.2. 2

4.6 Self-adjoint problems with w > 0.

The space of self-adjoint SLP is denoted by Ωs−a i.e.

Ωs−a = {ω = (a, b, A, B, 1/p, q, w)}

where we assume that

0 0
−∞ ≤ a < a < b < b ≤ ∞, A, B ∈ M2 (C),
 
0, −1
A E A∗ ∗
= B E B , rank (A|B) = 2, E = ,
1, 0
p, q, w : J → R, 1/p, q, w ∈ Lloc (J), w > 0, J = (a0 , b0 ).

1. Note that no sign condition is placed on p.

For our purposes here it is convenient to divide the self-adjoint boundary
conditions into three mutually exclusive subclasses and to use the following
canonical representations of these subclasses:
2. Separated self-adjoint BC. These are

(4.33) A1 y(a) + A2 (py 0 )(a) = 0, A1 , A2 ∈ R, (A1 , A2 ) 6= (0, 0)

(4.34) B1 y(b) + B2 (py 0 )(b) = 0, B1 , B2 ∈ R, (B1 , B2 ) 6= (0, 0)

These separated conditions can be parameterized as follows:

(4.35) cos α y(a) − sin α (py 0 )(a) = 0, 0 ≤ α < π;

(4.36) cos β y(b) − sin β (py 0 )(b) = 0, 0 < β ≤ π.

Note the different normalization in (4.36) for β than that used for α in (4.35).
This is for convenience in stating some of the results below.
3. All real coupled self-adjoint BC. These can be formulated as follows:

   
y(b) y(a)
(4.37) =K
(py 0 )(b) (py 0 )(a)

where K ∈ SL2 (R), i.e. K satisfies

 63

 
k11 k12
(4.38) K= , kij ∈ R, det K = 1.
k21 k22

4. All complex coupled self-adjoint BC. These can be formulated as follows:

   
y(b) y(a)
(4.39) = exp(i α) K
(py 0 )(b) (py 0 )(a)

where K satisfies (4.6) and −π < α < 0, or 0 < α < π.

For the canonical form (4.3), (4.4) of the special case of separated self-adjoint
BC we use the notation

(4.40) Ωs = {ω = (a, b, α, β, 1/p, q, w)};

for the general self-adjoint coupled case we let

(4.41) Ωcc = {ω = (a, b, α, K, 1/p, q, w), −π < α < 0, or 0 < α < π} .

When α = 0 we shorten (4.41) to

(4.42) Ωrc = {ω = (a, b, K, 1/p, q, w)} .

Most of the following results are well-known. See [81] for some proofs with only
integrable coefficients; see [68] for the case when p changes sign, and see [8], [18]
for the case of complex couple BC.

THEOREM 4.12. Let ω ∈ Ωs−a ; then ω is in exactly one of the subclasses: Ωs ,

Ωcc , Ωrc .
(a) Assume that

p ≥ 0 a.e. on [a, b].

Then for ω ∈ Ωs the BVP ω has only real and simple eigenvalues; there are an
infinite but countable number of them; they are bounded below and can be ordered
to satisfy

(4.43) −∞ < λ0 < λ1 < λ2 < · · · ; λn → +∞ as n → ∞

64

If un is an eigenfunction of λn , then un is unique up to constant multiples and un

has exactly n zeros in the open interval (a, b), n ∈ N0 = {0, 1, 2, . . . }.

Notation. Let

(4.44) λn = λn (a, b, α, β, 1/p, q, w); un = un (·, a, b, α, β, 1/p, q, w), n ∈ N0 ,

to highlight the dependence on these quantities.

For ω ∈ Ωrc the BVP ω has only real eigenvalues; each of these may be simple or
double; there are an infinite but countable number of them and they can be ordered
to satisfy

(4.45) −∞ < λ0 ≤ λ1 ≤ λ2 ≤ . . . ; λn → +∞ , as n → ∞.

Notation. Let

(4.46) λn = λn (a, b, K, 1/p, q, w); un = un (·, a, b, K, 1/p, q, w), n ∈ N0 .

Note that there is some arbitrariness in the indexing of the eigenfunctions corre-
sponding to a double eigenvalue.
For ω ∈ Ωcc the BVP ω has only real and simple eigenvalues; there are an infinite
but countable number of them and they can be ordered to satisfy

(4.47) −∞ < λ0 < λ1 < λ2 < . . . λn < ...

Notation. Denote these eigenvalues by

(4.48) λn = λn (a, b, α, K, 1/p, q, w); un = un (·, a, b, α, K, 1/p, q, w), n ∈ N0 .

For any ω ∈ Ωs−a we have the following asymptotic formula:

Z b r !−2
λn w
(4.49) → c = π2 , as n → ∞.
n2 a p

If we fix all variables except α and shorten the notation to λn = λn (α), then we
have λn (−α) = λn (α),and the complex conjugate of an eigenfunction of λn (α) is
an eigenfunction of λn (−α).
(b) Assume that p changes sign in the interval [a,b], i.e. p is positive on
a subset of [a, b] of positive Lebesgue measure and p is negative on a subset of the
interval [a, b] of positive Lebesgue measure. Then
 65

Each BVP ω ∈ Ωs has only real and simple eigenvalues; there are an infinite
but countable number of them; they are unbounded below and above and can be
ordered to satisfy

. . . < λ−2 < λ−1 < λ0 < λ1 < λ2 < . . . ;

(4.50) λn → +∞ as n → ∞; λn → −∞ as n → −∞.

Each BVP ω ∈ Ωrc has only real eigenvalues; each of these may be simple or
double; there are an infinite but countable number of them; they are unbounded
below and above and can be ordered to satisfy

... ≤ λ−2 ≤ λ−1 ≤ λ0 ≤ λ1 ≤ λ2 ≤ . . . ;

(4.51) λn → +∞ n → ∞, λn → −∞ n → −∞.

The notations for eigenvalues λn and eigenfunctions un , n ∈ Z, for part (b) are the
same as those introduced in part (a) for n ∈ N0 .

PROOF. The fact that the eigenvalues are unbounded below when p changes sign
was estalished by M. Möller. This holds even if there is no subinterval on which
p is negative. The fact that the eigenvalues for any ω ∈ Ωcc are all simple follows
from Weidmann [81], although it doesn’t appear to be stated explicitly there for
the general case. See also [8]. The other results are standard. 2

Notation. In the following we denote by λn and un the n-th eigenvalue and

the n-th eigenfunction of a SL problem where n ∈ N0 if p ≥ 0 a.e. on (a, b) and
n ∈ Z if p changes sign on (a, b), respectively. When p changes sign the eigenvalues
are indexed by Z and we follow the sleign2 convention and denote by λ0 the first
non-negative eigenvalue. This determines the indexing scheme uniquely.

THEOREM 4.13. Let ω = (a, b, A, B, 1/p, q, w) ∈ Ωs−a . Fix a, b, p, q, w and let

A = eiα K, K = (kij ), B = −I, where K ∈ SL2 (R) i.e. we have the BC

Y (b) = eiα K Y (a), −π ≤ α ≤ π.

Also assume that p > 0 a.e. on J and denote the eigenvalues for this boundary
condition by λn (α, K), abbreviated to λn (K) when α = 0, for n ∈ N0 .
Suppose that either k12 < 0 or k12 = 0 and k11 + k22 > 0. Then
1. λ0 (K) is simple;
2. λ0 (K) < λ0 (−K) and
3. the following inequalities hold for −π < α < 0 and 0 < α < π :

−∞ < λ0 (K) < λ0 (α, K) < λ0 (−K) ≤ λ1 (−K) < λ1 (α, K) < λ1 (K)
≤ λ2 (K) < λ2 (α, K) < λ2 (−K) ≤ λ3 (−K) < . . .
66

Furthermore, for 0 < α < β < π we have

λ0 (α, K) < λ0 (β, K) < λ1 (β, K) < λ1 (α, K) < λ2 (α, K) < λ2 (β, K)
< λ3 (β, K) < λ3 (α, K) < . . .

Suppose that either k12 > 0 or k12 = 0 and k11 + k22 < 0. Then
1. λ0 (−K) is simple;
2. λ0 (−K) < λ0 (K) and
3. the following inequalities hold for −π < α < 0 and 0 < α < π :

−∞ < λ0 (−K) < λ0 (α, K) < λ0 (K) ≤ λ1 (K) < λ1 (α, K) < λ1 (−K)
≤ λ2 (−K) < λ2 (α, K) < λ2 (K) ≤ λ3 (K) < . . .

Furthermore, for 0 < α < β < π we have

λ0 (β, K) < λ0 (α, K) < λ1 (α, K) < λ1 (β, K) < λ2 (β, K) < λ2 (α, K)
< λ3 (α, K) < λ3 (β, K) < . . .

PROOF. This general result is due to Eastham in a private communication [18],

special cases were established in [8], [81], [19]. 2

THEOREM 4.14. Let ω = (a, b, 0, π, 1/p, q, w) ∈ Ωs and fix a, p, q, w. Assume

that

0 0
p ≥ 0 a.e. and q 2 /w ∈ Lloc (a , b ).

0
Then for any n ∈ N0 , λD
n (b) is strictly decreasing in (a, b ) and

λD +
n (b) → ∞ as b → a .

PROOF. See [55]. 2

THEOREM 4.15. Let ω = (a, b, α, π/2, 1/p, q, w) ∈ Ωs (i.e. we have an arbitrary

separated self-adjoint BC at a but a Neuman condition at b.) Assume that

0
Q = q/w ∈ ACloc [a, b ), p(b) ≥ δ > 0 f or b ∈ (a, b0 ).

Then
 67

1. λ0 (b) → Q(a) = q(a)/w(a) as b → a+ .

2. λn (b) → ∞ as b → a+ f or n = 1, 2, 3, . . .
3. If Q is decreasing in (a, b0 ) then λn (b) is decreasing in (a, b0 ) and λn (b) ≥ Q(b)
for each n ∈ N0 .
4. If Q is increasing in (a, b0 ) and Q(b) → ∞ as b → b0 the λ0 (b) is increasing in
(a, b0 ) and λ0 (b) ≤ Q(b); for n ∈ N , λn (b) has a unique extremum in (a, b0 )
and this extremum is a strict minimum.
5. If Q has a unique extremum in (a, b0 ) and this extremum is a strict minimum
and Q(b) → ∞ as b → b0 then for any n ∈ N0 , λn (b) has a unique extremum
in (a, b0 ) and this extremum is a strict minimum.
PROOF. See Theorem 4.4 in [55]. 2

4.7 Differentiability Properties of Eigenvalues.

In this section we study the differentiable dependence of the simple eigenvalues on

the problem. It turns out that the continuous eigenvalue branches studied above
in Subsection 4.5 are differentiable.

THEOREM 4.16. Let ω = (a, b, A, B, 1/p, q, w) ∈ Ωs−a .

1. Fix all the components of ω except the left endpoint a and consider λ = λ(a)
as a function of a. Assume that for some a, λ(a) is a simple eigenvalue of
ω(a). Then there is a neighborhood U of a and a neighborhood V of λ(a)
such that for every c in U the BVP ω(c) has exactly one eigenvalue in V and
it is simple. The map λ : U → V is differentiable almost everywhere in U and
we have

1
(4.52) λ0 (a) = |pu0 |2 (a) − |u|2 (a)[q(a) − λ(a)w(a)] a.e. in U.
p(a)

Furthermore, if p, q, w are continuous at a and p(a) 6= 0, then (4.52) holds at

the point a.
2. Fix all the components of ω except b. Assume that for some b, λ(b) is a simple
eigenvalue of ω(b). Then there is a neighborhood U of b and a neighborhood
V of λ(b) such that for every c in U the BVP ω(c) has exactly one eigenvalue
in V and it is simple. The map λ : U → V is differentiable almost everywhere
in U and we have

1
(4.53) λ0 (b) = − |pu0 |2 (b) + |u|2 (b) [q(b) − λ(b)w(b)].
p(b)

Furthermore, if p, q, w are continuous at b and p(b) 6= 0, then (4.53) holds at

the point b.

REMARK 11. In his well known monograph on variational methods for eigen-
value problems Hans Weinberger states, without proof or reference to a proof, that
68

the Dirichlet eigenvalues are decreasing functions of the length of the interval but
that this is not true for the Neuman eigenvalues. Theorem 4.16 sheds a great deal
of light on this.
Assume that p > 0. For Dirichlet BC u(a) = u(a, a, 0, π) = 0 and u(b, b, 0, π) = 0
hence the second term in (4.52) and in (4.53) is zero; thus it is clear from these
formulas that the Dirichlet eigenvalues are increasing functions of the left endpoint
and decreasing functions of the right endpoint. It is also clear from (4.52), (4.53)
that this is not true, in general, for any other boundary conditions. However if q/w is
bounded above, say by C, then for any boundary conditions all eigenvalues greater
than C are increasing functions of the left endpoint a and decreasing functions
of the right endpoint b. Since for any fixed regular SLP the eigenvalues λn →
∞ asymptotically as n2 it is clear that if q/w is bounded above only the lower
eigenvalues may fail to be monotonic functions of the length of the interval.

THEOREM 4.17. Let ω = ω(a, b, A, B, 1/p, q, w) ∈ Ωs−a and fix a, b, p, q, w.

(Recall that A, B are replaced by α, β for BC (4.35), (4.36); by α, K for (4.39)
and by K for (4.37)). Fix all components of ω except α and let λ = λ(α) and
u = u(·, α). Then λ is differentiable and

(4.54) λ0 (α) = −u2 (a) − (pu0 )2 (a), 0 ≤ α < π.

1. Fix all components of ω except β and let λ = λ(β) and u = u(·, β). Then λ
is differentiable and

(4.55) λ0 (β) = u2 (b) + (pu0 )2 (b), 0 < β ≤ π.

2. Fix all components of ω except α min(4.39) and let λ = λ(α) and u = u(·, α).
Then λ is differentiable at α for any α satisfying −π < α < 0 or 0 < α < π
and

(4.56) λ0 (α) = −2 Im[u(b) (pū0 )(b)],

where Im[z] denotes the imaginary part of z.

3. Fix all components of ω except K. Assume that λ = λ(K) is a simple
eigenvalue and u = u(·, K) a normalized eigenfunction of K. Then there
exists a neighborhood U of K and a neighborhood V of λ(K) such that for
every G in U the (self-adjoint or non-self-adjoint since G is not assumed to
be in SL2 (R) ) BVP ω(G) has exactly one eigenvalue in V and it is simple.
The map λ : U → V is differentiable at K and its Frechet derivative is given
by:

 
u(b)
(4.57) λ0 (K) H = [pu0 (b), −u(b)] HK −1 , H ∈ M2,2 (C).
(pu0 )(b)
 69

PROOF. See [57] for parts (1) and (2) and see [69] for part (3). 2

THEOREM 4.18. Let ω = (a, b, A, B, 1/p, q, w) ∈ Ω . Fix a, b, A, B and assume

that A, B satisfy the self-adjointness conditions:

 
∗ ∗ 0 −1
AE A = B E B , rank (A|B) = 2, E = .
1 0

Suppose that λ(ω) is an isolated simple eigenvalue of ω. Then there is a simple

closed curve Γ in C with λ(ω) in its interior and a neighborhood U of ω in Ω such
that for any ρ in U the BVP identified with ρ has exactly one simple eigenvalue
inside Γ. This map λ : U → C is differentiable with respect to
1. q with p, w fixed and

Z b
(4.58) λ0 (q) h = |u(·, q)|2 h, h ∈ L(a, b),
a

where u = u(·, q) is a normalized eigenfunction of λ(q);

2. 1/p with q, w fixed and

Z b
0
(4.59) λ (1/p) h = − |(pu0 )(·, 1/p)|2 h, h ∈ L(a, b),
a

where u = u(·, 1/p) is a normalized eigenfunction of λ(1/p);

3. w with 1/p, q fixed and

Z b
(4.60) λ0 (w) h = −λ(w) |u(·, w)|2 h, (h ∈ L(a, b)),
a

where u = u(·, w) is a normalized eigenfunction of λ(w).

PROOF. See [69]. 2

THEOREM 4.19. Let ω = (a, b, A, B, 1/p, q, w) ∈ Ωs−a and fix a, b, A, B. As-

sume that λ(ω) is a simple eigenvalue of ω. (Recall that each eigenvalue is assumed
to lie on a continuous eigenvalue branch; see Theorem 4.10 and Remark 10.)
1. Fix p, w. Suppose Q ∈ L(a, b) and Q ≥ q on [a, b]. Assume that λ(s(t)) is on
the same continuous eigenvalue branch as λ(q) for all t ∈ [0, 1], where

s(t) = q + t(Q − q).

Then λ(Q) ≥ λ(q). If Q > q on a subset of [a, b] having positive Lebesgue

measure, the λ(Q) > λ(q).
70

2. Fix q, w. Suppose P ∈ L(a, b) and P ≤ p on [a, b]. Assume that λ(s(t)) is on

the same continuous eigenvalue branch as λ(1/p) for t ∈ [0, 1], where

s(t) = 1/p + t(1/P − 1/p).

Then λ(1/P ) ≥ λ(1/p). If 1/P < 1/p on a subset of [a, b] having positive
Lebesgue measure, the λ(1/P ) < λ(1/p).
3. Fix p, q. Suppose W ∈ L(a, b) and W ≥ w > 0 on [a, b]. Assume that λ(s(t))
is on the same continuous eigenvalue branch as λ(w) for all t ∈ [0, 1], where

s(t) = w + t(W − w).

Then λ(W ) ≥ λ(w) if λ(W ) < 0 and λ(w) < 0; but λ(W ) ≤ λ(w) if λ(W ) > 0
and λ(w) > 0. Furthermore, if strict inequality holds in the hypothesis on a set
of positive Lebesgue measure, then strict inequality holds in the conclusion.
PROOF. We give the proof for (1), the proofs of (2) and (3) are similar. Define

f (t) = λ(s(t)), t ∈ [0, 1].

From the chain rule and (4.58) we have

Z b
f 0 (t) = λ0 ((s(t)) s0 (t) = |u2 (r, s(t))| (Q(r) − q(r)) dr ≥ 0, t ∈ [0, 1].
a

Hence f (1) = λ(Q) ≥ λ(q) = f (0). The strict inequality part of the theorem also
follows from this. 2

THEOREM 4.20. Fix a, b ∈ J, a < b. Let ω = (A, B, 1/p, q, w) ∈ Ω = (M2 (C))2 ×

(L(a, b))3 Assume that λ(ω) is an isolated simple eigenvalue of ω. Then there is a
simple closed curve Γ in C with λ(ω) in its interior and a neighborhood U of ω in Ω
such that for each ϕ ∈ U the BVP identified with ϕ has exactly one simple eigen-
value λ(ϕ) inside Γ. This map λ : U → C is differentiable at ω and its derivative is
given by:

Z b
(4.61) λ0 (ω) ρ = {−py 0 p̄z̄ 0 (1/r) + [g − λ(ρ)v]yz̄} + d∗ [CY (a) + DY (b)],
a

where ρ = (C, D, 1/r, g, v) ∈ Ω, y, z are biorthogonal solutions of the given and its
adjoint boundary value problems at λ(ω), i.e.

 
0 0 y
(4.62) −(py ) + qy = λ(ω) w y, AY (a) + BY (b) = 0, Y =
py 0
 71

  Z b
0 0 z
(4.63) −(p̄z ) + q̄z = λ̄(ω)w̄z, Z = , y z̄ w = 1,
p̄z 0 a

and where d ∈ C2 is such that

 
0 −1
(4.64) Z(a) = EA∗ d, Z(b) = −EB ∗ d, E = .
1 0

PROOF. See [69]. 2

REMARK 12. Note that for Theorem 4.20 no self-adjointness hypothesis is need-
ed: the coefficients p, q and the weight function w may be complex-valued; the
boundary conditions need not be self-adjoint. The existence of infinitely many
eigenvalues for non-self-adjoint SLP is well known for so called Birkhoff regular and
Stone regular SLP, see [67], [20].

4.8 Comments.

1. These are made separately for each subsection.

2. Characterizing the eigenvalues as the zeros of an entire function is a standard
technique but we don’t know of a refernce where it is done for the general
∆ function given by (4.7). Lemmas (4.1) to (4.3) are standard. The Fourier
equation in subsection 3 provides examples for cases (1), (2) and (4) but not
(3) of Lemma 4.3. Atkinson indicates in his book [3] that there are examples of
case (3) of lemma (4.3) with a leading coefficient p such that 1/p is identically
zero on subintervals. Are there such examples with p > 0 ?
Lemma 4.5 was established by Bailey, Everitt and Zettl [8] for self-adjoint,
regular and singular SLP. It was only during the writing of these notes that
the author realized that no symmetry (formal self-adjointness) or definiteness
assumption is needed on the equation.
Theorem 4.6 is only a special case of results of Mennicken and Möller and
of Eberhard and Freiling. More results of this kind and detailed proofs can
be found in the forthcoming book by the former two authors [67]
3. Coddington and Levinson [14] have examples showing that the Fourier equa-
tion with non-self-adjoint two point boundary conditions can have either no
eigenvalues or every complex number may be an eigenvalue. The writer found
it rather interesting that dozens of properties of the hyperbolic functions come
into play just to compute eigenvalues of regular problems for the Fourier
equation. This was a humbling experience. Thank goodness for the book by
Abramowitz and Stegun[1] .
4. For regular problems on an interval J it seems that the L1 (J) norm for
1/p, q, w is the “natural” norm to use.
5. The “index jumping” phenomenon in order to preserve continuity given by
Theorem 4.7 is due to Everitt, Möller and Zettl [23]. Another version of
it, which involves changing the endpoint and the boundary conditions “ in
harmony with each other ” was discovered by Bailey, Everitt and Zettl and
72

is used in their numerical code SLEIGN2 to compute eigenvalues of singular

problems with oscillatory endpoints. See the forthcoming article by Everitt,
Möller and Zettl.
The continuity result of Theorem 4.10 is basically a consequence of a the-
orem in complex variables which extends the result on the continuity of the
roots of polynomial equations as functions of the coefficients, to analytic func-
tions. Given an entire function whose coefficients from its series expansion
depend continuously on a parameter which lives in a metric space, then its
zeros are continuous functions of this parameter. The previous subsection on
the space of BVP served to provide such a metric space, actually a Banach
space for our application. The continuity result of Theorem 4.10 allows us to
work on “continuous eigenvalue branches” which, as Theorem 4.7 shows, are
not always specified by a fixed eigenvalue index.
Theorem 4.11 appears to be new in [57], see also [54].
6. M. Möller [68] showed that the eigenvalues are not bounded below when p
changes sign even if there is no subinterval on which p is negative. The other
results summarized in Theorem 4.12 are standard.
Theorem 4.13 is due to Eastham [18] in a private communication and
is presented here for the first time with his permission; special cases were
established
 in [8],
 in Weidmann’s Lecture Notes [81] for the special case
c 0
K = with c a non zero real number. For a very different proof
0 1/c
for the case c = 1 see the book of Eastham [19].
The asymptotic formula (4.49) has many extensions, see Atkinson and
Mingarelli [5], Harris [39], [52].
7. The differentiability results of Subsection 7 are due to Kong and Zettl [55],
[57]. The differentiability results with respect to the endpoints were inspired
by earlier such results of Dauge and Helffer [15]. For extensions of all these
results to higher order ode’s see Kong, Wu and Zettl [54]. For a far reaching
extension to operator theory with applications to matrix theory, ordinary
differential equations, partial differential equations, etc. see the article of
Möller and Zettl [69].

5 SINGULAR BOUNDARY VALUE PROBLEMS

5.1 Introduction.

In this section we discuss singular self-adjoint SLP. This is a field so vast that we
can only hope to give a brief introduction here. Following a review of the some
of the basic theory we will focus on two topics: (i) The behavior of eigenvalues of
regular SLP near a singular boundary and (ii) The approximation of the discrete
as well as continuous (essential) spectrum of a given singular problem with spectra
of regular problems. On these two topics we aim to bring the reader to the frontier.
To illustrate some of the basic behavior of the spectrum of SLP we discuss briefly
the 29 examples from the sleign2 package in Subsection 5.8.
 73

5.2 Principal and non-principal solutions.

Consider the symmetric equation

(5.1)M y ≡ −(py 0 )0 + qy = λ w y on J = (a, b), −∞ ≤ a < b ≤ ∞,

λ ∈ R , p, q, w : J → R, 1/p, q, w ∈ Lloc (J), p ≥ 0, w > 0 a.e. on J.

Note that for this section we assume that p ≥ 0 and w > 0 unless explicitly stated
otherwise.
Recall that, according to Proposition 3.10 in section 3 no nontrivial solution of
(5.1) can have an accumulation point of zeros in the interior of J. The zeros, if any,
of any nontrivial solution of (5.1) inside the interval J are isolated. Thus only an
endpoint of J can be an accumulation point of zeros of a nontrivial solution of (5.1)
and that can happen only at a singular endpoint.

DEFINITION 5.1 (Principal Solution). Let u, v be real solutions of (5.1). Then

• u is called a principal solution at a if
1. u(t) 6= 0 for t ∈ (a, d] and some d ∈ J,
2. every solution y of (5.1) which is not a multiple of u satisfies

(5.2) u(t) = o(y(t)) as t → a.

• v is called a non-principal solution at a if

1. v(t) 6= 0 for t ∈ (a, d] and some d ∈ J,
2. v is not a principal solution at a.
Principal and non-principal solutions at b are defined similarly.
To simplify things we state definitions and assertions only for the left endpoint
a. Similar definitions and assertions for the right endpoint b always hold and are
freely used.

LEMMA 5.2. If (5.1) has a principal solution u at a, then every non-zero real
multiple of u is also a principle solution and no other solution is a principal solution
at a.
PROOF. This follows directly from the definition. 2

REMARK 13. By Lemma (5.2) the principal solution u at an endpoint, if it

exists, is unique up to real constant multiplicative factors. Non-principal solutions
are never unique, since if v is non-principal and u is principal, both at the same
endpoint, then v +cu is also a non-principal solution for any c ∈ R. Simple examples
show that the same solution may be principal at one endpoint and non-principal at
the other. Clearly principal and non-principle solutions do not exist at an oscillatory
endpoint.
74

REMARK 14. If the equation in (5.1) is regular at a then for any solution y, y
and py 0 can be continuously extended to a and principal solutions u exist and satisfy
the initial conditions : u(a) = 0, (pu0 )(a) 6= 0. Any non-principal solution v at a
satisfies : v(a) 6= 0.

THEOREM 5.3. The equation (5.1) is non-oscillatory at a if and only if there

exists a principal solution at a.
PROOF. See p. 547 in Niessen and Zettl [74]. 2

The next result gives a characterization of principal and non-principal solutions.

This will be used below in “regularizing ” singular LCNO endpoints.

THEOREM 5.4. Assume that (5.1) is non-oscillatory at a for some λ ∈ R . Let

u, v be real solutions of (5.1) satisfying u(t) 6= 0, v(t) 6= 0 for t ∈ (a, d] and some
d ∈ J. Then
1. u is a principal solution at a if and only if

Z d
1
(5.3) = ∞;
a p u2

2. v is a non-principal solution at a if and only if

Z d
1
(5.4) < ∞;
a p v2

3. if u is a principal solution and v is a non-principal solution at a , then there

exists a c ∈ R , c 6= 0, such that

Z t
c
(5.5) u(t) = v(t) , a < t ≤ d,
a p v2

4. and

(5.6) |u(t) v(x)| < |u(x) v(t)|, f or a < t < x ≤ d.

PROOF. See p. 548 in [74]. 2

5.3 Singular Boundary Conditions.

Boundary conditions of the form (4.3) do not make sense when one endpoint, say
a, is singular since Y (a) does not exist, in general. What takes their place ? This
depends on the endpoint classification e.g. LP or LC. Before going into the details
 75

we need to set the stage first. Recall the definitions of the maximal domain ∆ and
the sesquilinear form [·, ·] from Subsection 3.6 of Section 3.

DEFINITION 5.5 (Maximal and Minimal Operators). Let (5.1) hold and let ∆
and M be defined as in
Subsection 3.6. Define
T1 f = M f, f or f ∈ ∆.

T00 f = M f, f ∈ ∆, f has compact support in J.

Then T1 is called the maximal operator of (M, w) (or of the equation (5.1)) on J,
and the minimal operator T0 of (M, w) is defined as the closure of T00 .

LEMMA 5.6. The maximal and minimal domains are dense in the Hilbert space

Z
H = L2 (J, w) = {f : J → C, |f |2 w < ∞},
J
T0 is a closed symmetric operator and T0∗ = T, T ∗ = T0 . Hence any self-adjoint
extension of T0 is also a self-adjoint restriction of T and conversely.
PROOF. See [72], [81]. 2

Any self-adjoint extension S of the minimal operator T0 satisfies

(5.7) T0 ⊂ S = S ∗ ⊂ T1 ,

and can be determined by two point boundary conditions. These, however, are
vacuous at an LP endpoint. To describe these conditions it is convenient to take
cases depending on the LP/LC classification of the endpoints. Here LC/LP will
mean that the left endpoint a is LC and the right endpoint b is LP, etc.
An operator S satisfying (5.7) is called a self-adjoint extension of T0 on J, or a
self-adjoint restriction of T1 on J, or simply a self-adjoint realization of the equation
(5.1) on J, or a self-adjoint realization of (M, w) on J.

THEOREM 5.7. Let (5.1) hold.

• Assume each endpoint is LP. Then the minimal operator T0 is itself self-
adjoint and has no proper self-adjoint extensions since T0 = S = S ∗ = T1 .
Thus there are no boundary conditions needed nor allowed in this case.
• Assume a is either R or LC and b is LP. Then there is no boundary condition
needed or allowed at b and all the self-adjoint BC at a can be characterized
as follows: Let u, v ∈ ∆ be real-valued such that [u, v](a) 6= 0, and let

(5.8) A1 [u, y](a) + A2 [v, y](a) = 0, A1 , A2 ∈ R, (A1 , A2 ) 6= (0, 0).

76

In other words, if S is the restriction of the maximal operator to functions

y ∈ ∆ satisfying (5.8) then S is a self-adjoint extension of T0 (and a self-
adjoint restriction of T1 ), and, conversely, given any self-adjoint extension of
T0 its domain consists of all y ∈ ∆ satisfying (5.8) for some such A1 , A2 .
• Assume a is LP and b is either R or LC. Then there is no boundary condition
needed or allowed at a and all the self-adjoint BC at b can be characterized
as follows: Let u, v ∈ ∆ be real valued and satisfy [u, v](b) 6= 0 and let

(5.9) B1 [u, y](b) + B2 [v, y](b) = 0, B1 , B2 ∈ R, (B1 , B2 ) 6= (0, 0).

In other words, if S is the restriction of the maximal operator to functions

y ∈ ∆ satisfying (5.9) then S is a self-adjoint extension of T0 (and a self-
adjoint restriction of T1 ), and, conversely, given any self-adjoint extension of
T0 its domain consists of all y ∈ ∆ satisfying (5.9) for some B1 , B2 .
• Assume the left endpoint a is R or LC and the right endpoint b is also in the R
or LC case. Then there are boundary conditions required at both endpoints in
order to determine a self-adjoint extension of T0 . These may be separated (i.e.
separate conditions at each endpoint) or coupled. Choose real valued u, v ∈ ∆
satisfying [u, v](a) 6= 0 and [u, v](b) 6= 0. Then all self-adjoint extensions of T0
are determined by BC which have the following form:

 
[y, u]
(5.10) AY (a) + BY (b) = 0, Y = , A, B ∈ M2 (C),
[y, v]
 
0 −1
rank(A : B) = 2, AEA∗ = BEB ∗ , E = .
1 0

In other words, every BC (5.10) determines a self-adjoint extension of T0 and

every such extension is determined by (5.10) for some matrices A, B satisfying
these conditions.
The existence of Y (a), as a finite limit, follows from Green’s formula and
the hypothesis that each endpoint is either R or LC.
PROOF. This is well known; see [6], [61], [8], [81]. 2

REMARK 15. Functions u, v needed in (5.10) can often, but not always, be
obtained by choosing linearly independent real solutions of (5.1) for some particular
real value of λ, e.g. λ = 0.

Below, an operator S satisfying (5.7) will be called a self-adjoint realization of

the equation (5.1) or of (M, w) where the expression M is defined in Subsection
3.6.
Now that we know the self-adjoint realizations S of M we next discuss their
spectrum σ(S). Let σd (S) denote the discrete spectrum i.e. the set of isolated
eigenvalues, if any, of S. Set σe = σ − σd , then σe is called the essential spectrum; in
 77

some of the literature this is also referred to as the continuous part of the spectrum.
Either one, but not both, of σd and σe may be empty.
Next we summarize some basic properties of the spectrum.

PROPOSITION 5.8. Let S be a self-adjoint realization of (M, w). Then

1. the spectrum of S is a closed subset of the reals R which is not bounded
above; it may or may not be bounded below;
2. σe (S) = σe (T0 ), and so the essential spectrum of all self-adjoint extensions is
the same; the essential spectrum is also invariant under an L1 (a, b) perturba-
tion of q, (See Hinton and Shaw [45]);
3. the operator T0 is bounded below if and only if σ(S) is bounded below for
each self-adjoint realization S; (But not uniformly for all S);
4. Each one of the sets σ(S), σe (S) is a closed subset of R. Either one of these
two sets, but not both, may be empty.

REMARK 16. We see from this Proposition that the essensial spectrum does
not depend on the boundary conditions; thus σe = σe (J, p, q, w). The eigenvalues
do depend on the boundary condition. To the left of the continuous spectrum
there may be no eigenvalues, a finite number of them, or an infinite number of
them. In the case of an infinite number of eigenvalues to the left of σ0 = inf σe
these can have no accumulation point other than, possibly, σ0 and −∞. There also
may be eigenvalues embedded in the continuous spectrum as well as in gaps of the
continuous spectrum. The essential spectrum may have no gaps, a finite number of
them, or an infinite number of them. A remarkable result of Hartmann [41] states
that any closed set of real numbers which is not bounded above is the spectrum of
a Sturm-Liouville operator ! See also Halvorsen [37]. In particular the spectrum of
an SLP can be a Cantor-like set which is not bounded above.
For specific examples of SLP which illustrate some of these features (but not
the Cantor-like behavior of the spectrum) the reader is referred to the examples
discussed in Subsection 5.8 which come with the code SLEIGN2; these, along with
the code, can be downloaded from the WWW using netscape or mosaic or lynx by
specifying the URL :
ftp://ftp.math.niu.edu./pub/papers/Zettl/Sleign2
or by accessing the web page
http://www.math.niu.edu./˜zettl/SL2/

5.4 The Friedrichs extension.

If there is one proper self-adjoint extension of the minimal operator then there are
an infinite number of such extensions. Friedrichs singled out one of these which
he called “ausgezeichnete” and which has come to be known as “the Friedrichs
extension”. He singled one out by giving a construction which constructed such
an extension while preserving the lower bound. (However having the same lower
bound does not characterize the Friedrichs extension since there are, in general,
other self-adjoint extensions which also have the same lower bound.) To get the
Friedrichs extension you must use the Friedrichs construction or some equivalent
78

version of it. This construction works for any symmetric, densely defined, bounded
below, operator in a Hilbert space. When it is applied to T0 it makes no explicit
use of boundary conditions. Thus the question arises : What boundary conditions
determine the Friedrichs extension of the minimal operator T0 ?

DEFINITION 5.9 (The Friedrichs Extension). Suppose S is a densely defined

symmetric (but not necessarily closed) operator in a Hilbert space H which is
bounded below. Let D(SF ) denote the set of all y ∈ D(S ∗ ) for which there exists
a sequence {yn : n ∈ N } such that
1. yn → y in H as n → ∞,
2. (S(yn − ym ), yn − ym ) → 0 as n, m → ∞.
Define the operator SF by

SF y = S ∗ y, f or y ∈ D(SF ).

Then SF is called the Friedrichs extension of S. According to the well known result
of Friedrichs [28], SF is a self-adjoint operator with the same lower bound as S.

Friedrichs himself [29] addressed the question of which boundary condition de-
termines the Friedrichs extension for regular SLP ? For regular SLP the answer in
general is : the Dirichlet condition, see [73], [50], [10], [78]. We take up the singular
case next.

THEOREM 5.10. If the minimal operator T0 is bounded below with lower bound
c then the equation (5.1) is NO at a and at b for any λ < c. Conversely, if the
equation (5.1) is NO at a for some real λa and at b for some real λb , then the
minimal operator T0 is bounded below.
PROOF. See Niessen and Zettl [74] Corollary 2.1 and Theorem 4.2. 2

THEOREM 5.11. If (M, w) is in the LC case at a and if M y = λwy is NO at a

for some real λ, then M y = λw y is NO at a for every real λ.
PROOF. See Theorem 4.1 in Niessen and Zettl [74]. 2

THEOREM 5.12. Let (5.1) hold. Let the maximal domain ∆, the Lagrange
form [·, ·] and the expression M be defined as in Subsection 3.6. Assume that each
endpoint is either regular or LCNO. Then the minimal operator is bounded below
and thus has a Friedrichs extension SF . Let ua be a principal solution at a for some
real λa and let ub be a principal solution at b for some real λb . Then the domain
D(SF ) of SF is given by

(5.11) D(SF ) = {y ∈ ∆ : [y, ua ](a) = 0 = [y, ub ](b)}.

 79

Note that (5.11) is independent of the principal solution chosen for λa and is
independent of λa ∈ R; similarly at b.
PROOF. See Niessen and Zettl [74], Theorems 4.2 and 4.3. 2

REMARK 17. At a regular endpoint, say a, (5.11) reduces to y(a) = 0. Thus the
BC (5.11) can be viewed as the singular analogues of the regular Dirichlet boundary
conditions. It is also shown in [74] that the conditions (5.11) are equivalent to

y(t) y(t)
lim+ = 0 = lim− ,
t→a va (t) t→b vb (t)

where va , vb are arbitrary non-principal solutions at a and b , respectively, for

arbitrary real λa , λb.

5.5 R or LCNO/ R or LCNO.

We now study the case when each endpoint is either regular or LCNO. The proper-
ties of the eigenvalues and eigenfunctions in this case are similar to the regular case.
In fact Niessen and Zettl [74] have shown that, given any SLP with endpoints which
are either regular or LCNO there exists a regular SLP which has exactly the same
spectrum as this singular problem and furthermore the eigenfunctions of the given
singular problem {yn : n ∈ N0 } are related to the eigenfunctions {zn : n ∈ N0 } of
the corresponding regular problem by the equation

(5.12) yn (t) = v(t) zn (t), t ∈ (a, b), n ∈ N0 ,

for some function v in the maximal domain of the singular problem which satisfies
v(t) > 0 for t ∈ (a, b). Since each zn is a solution to a regular problem on (a, b) z
and its quasi-derivative can be continuously extended to the endpoints by Theorem
3.4. Hence the singular behavior at each endpoint is contained in v. In particular,
this shows that at each endpoint the singular (e.g. asymptotic) behavior of all
eigenfunctions (in fact of all solutions for all real λ) is the same.

THEOREM 5.13. Let (5.1) hold and assume that each endpoint is either R or
LCNO. Let S be a self-adjoint realization of (5.1).
1. Then the spectrum of S is discrete and bounded below. It consists of a
countably infinite sequence {λn : n ∈ N0 } of real eigenvalues tending to +∞
which can be ordered to satisfy

(5.13) −∞ < λ0 ≤ λ1 ≤ λ2 ≤ λ3 ≤ . . . → ∞.

Here the eigenvalues are counted according to their multiplicity. Each eigen-
value can have multiplicity one, in which case it is called simple, or two, in
80

which case it is called double. Therefore, in (5.13), equality cannot hold for
more than two consecutive terms.
2. Let yn be a real eigenfunction of λn . Then yn has at least n − 1 and at most
n + 1 zeros in (a, b). If [yn , ua ](a) = 0, where ua is a principal solution at a,
then yn has at most n zeros in (a, b).
3. Let N (λ) denote the number of eigenvalues of S in the interval (−∞, λ]. Then

s
Z b
N (λ) 1 w(t)
(5.14) √ → dt < ∞, as λ → ∞,
λ π a p(t)

4. and

λn 1
(5.15) → q 2 , as n → ∞.
n2 π 2 R b w(t)
a p(t) dt

The finiteness of the integral in (5.14) and (5.15) is a consequence of the assump-
tion that each endpoint is R or LCNO.
PROOF. See Theorem 5.2 in [74]. 2

THEOREM 5.14. Let (5.1) hold. Assume that each endpoint is regular or LCNO.
Let u, v be real maximal domain functions such that u, v are principal and non-
principal solutions at a for some real λa and at b for some real λb , respectively,
normalized to satisfy : [u, v](a) = 1 = [u, v](b). (Such u, v exist by Lemma 5.2 and
Theorem 5.3 ). Let S with spectrum σ(S) be the self-adjoint realization determined
by the normalized separated boundary conditions

(5.16) cos(α) [y, u](a) + sin (α) [y, v](a) = 0, 0 ≤ α < π,

(5.17) cos(β) [y, u](b) + sin(β) [y, v](b) = 0, 0 < β ≤ π.

Then
1. σ(S) = {λn : n ∈ N0 } and each eigenvalue is simple.These eigenvalues can be
ordered to satisfy

(5.18) −∞ < λ0 < λ1 < λ2 < λ3 < . . . → ∞;

2. if yn is an eigenfunction corresponding to λn , then yn has exactly n zeros in

the open interval (a, b) for any n ∈ N0 ;
3. λn (α, β) is a continuous function of α, β , n ∈ N0 ;
4. λn (α, β) is strictly decreasing in α for each fixed β and strictly increasing in
β for each fixed α, n ∈ N0 ;
 81

5. with the understanding that terms involving λ−1 , λ−2 are not present we have
the following inequalities for α ∈ [0, π), β ∈ (0, π], and n ∈ N0

   
λn−1 (0, β) λn (0, β)
(5.19) λn−2 (0, π) < < λn (α, β) ≤ ≤ λn (0, π).
λn−1 (α, π) λn (α, π)

PROOF. The regular case is known, see [81]. The singular case then follows from
the regular case and the transformation employed in [74] to “regularize” singular
LCNO endpoints. 2

THEOREM 5.15 (Canonical Coupled BC). Let (5.1) hold. Assume that each
endpoint is regular or LCNO. Then the canonical form of all coupled self-adjoint
boundary conditions is

(5.20) Y (b) = eiα K Y (a)

where −π ≤ α ≤ π,

 
[y, θ]
Y = , θ, ϕ ∈ ∆, θ, ϕ real, K ∈ SL2 (R),
[y, ϕ]
(5.21) [θ, ϕ](a) = 1 = [θ, ϕ](b); Y (a) = lim+ Y (t), Y (b) = lim− Y (t).
t→a t→b

Fix p, q, w, a, b and let λn (α, K), n ∈ N0 denote the eigenvalues for BC (5.20);
when α = 0 this notation is abbreviated to λn (K). Suppose that

k12 < 0 or k12 = 0 and k11 + k22 > 0,

Then
1. λ0 (K) is simple;
2. λ0 (K) < λ0 (−K);
3. the following inequalities hold for -π < α < 0 and 0 < α < π :

−∞ < λ0 (K) < λ0 (α, K) < λ0 (−K) ≤ λ1 (−K) < λ1 (α, K) < λ1 (K)
(5.22) ≤ λ2 (K) < λ2 (α, K) < λ2 (−K) ≤ λ3 (−K) < . . .

Furthermore, for 0 < α < β < π we have

λ0 (α, K) < λ0 (β, K) < λ1 (β, K) < λ1 (α, K) < λ2 (α, K) < λ2 (β, K)
(5.23) < λ3 (β, K) < λ3 (α, K) < . . .
82

Note that K ∈ SL2 (R) implies −K ∈ SL2 (R) and therefore if the hypothesis :

k12 < 0 or k12 = 0 and k11 + k22 > 0

fails to hold for K then it holds for −K.

PROOF. In Niessen and Zettl [74] it is shown for any singular problem with each
endpoint either R or LCNO there exists a regular problem which has exactly the
same eigenvalues. This result then follows from the regular case, see Theorem 4.13.
The regular case was established by Eastham [18] using an argument similar to that
in his book [19] for the periodic and semi-periodic cases; and by Bailey, Everitt and
Zettl [8] using an argument similar to that used by Weidmann [81]. 2

The next result characterizes the eigenvalues of singular SLP consisting of the
canonical form (5.20) of the coupled self-adjoint boundary conditions but for the
general equation (5.1) with real or complex valued coefficients.

THEOREM 5.16. Consider the SLP consisting of the equation

(5.24) −(py 0 )0 + qy = λwy on J = (a, b), −∞ ≤ a < b ≤ ∞,

where

(5.25) p, q, w : J → C, 1/p, q, w ∈ Lloc (J),

with the boundary conditions (5.20), (5.21) and assume each endpoint is either R
or LC. For each λ ∈ C determine unique solutions u = u(·, λ), v = v(·, λ) by the
‘singular initial conditions’

(5.26) [u, θ](a, λ) = 0, [u, ϕ](a, λ) = 1, [v, θ](a, λ) = 1, [v, ϕ](a, λ) = 0.

Such solutions u, v exist by Theorem 3.13. Let K ∈ SL2 (R). Then for any α,
−π ≤ α ≤ π, a number λ ∈ C is an eigenvalue of the BVP (5.26), (5.25), (5.20),
(5.21), (5.24) if and only if

(5.27) D(K, λ) = 2 cos(α),

where for λ ∈ C

D(K, λ) = k11 [u(·, λ), ϕ](b) + k22 [v(·, λ), θ](b)

(5.28) −k12 [v(·, λ), ϕ](b) − k21 [u(·, λ), θ](b).
 83

PROOF. See Theorem 3.1 in [8]. Although this Theorem is stated there only for
the case when p, q, w are real valued and w > 0 the proof given there holds with no
significant changes when p, q, w are complex valued. 2

COROLLARY 5.17. For any n ∈ N0 we have λn (−α, K) = λn (α, K).

PROOF. This follows directly from (5.27) since cos(α) = cos(−α). 2

5.6 Behavior of eigenvalues near a singular boundary.

For this subsection we change the notation for the interval J from

J = (a, b)

to

J = (a0 , b0 ).

The reason for this change of notation is that we wish to consider “approximations”
of a singular SLP on an interval (a0 , b0 ) by a sequence of regular SLP on truncated
intervals (ar , br ) where

−∞ ≤ a0 < ar < br < b0 ≤ ∞

and the sequence {ar : r ∈ N } converges decreasingly to a and the sequence

{br : r ∈ N } converges increasingly to b. By S and Sr we denote self-adjoint
realizations of (M, w) on the intervals (a0 , b0 ) and (ar , br ), respectively. Thus S and
Sr are a self-adjoint operator in the Hilbert spaces H = L2 ((a0 , b0 ), w) and Hr =
L2 ((ar , br ), w), respectively. Similarly the spectrum and eigenvalues are denoted
by σ(S), σ(Sr ), λn (S), λn (Sr ), n ∈ N0 , r ∈ N. Below, the “inherited” operators
Sri are defined by “inherited” boundary conditions. These play a special role in the
approximation of the singular spectrum.

THEOREM 5.18. Let (5.1) hold. Let S, Sr be arbitrary self-adjoint realizations

of (M, w) on the intervals (a0 , b0 ) and (ar , br ), respectively, for r ∈ N.
For all endpoint classifications : If σ(S) is not bounded below, then

(5.29) λn (Sr ) → −∞, as r → ∞, f or each n ∈ N0 .

PROOF. This is an as yet unpublished result of Everitt and Zettl [26]. 2

84

DEFINITION 5.19 (Inherited BC and Operators). • Near an LP endpoint

the inherited BC is the Dirichlet condition.
• Assume a0 is LC and b0 is LP. Then there is no singular boundary condition
at b0 and all singular self-adjoint BC at a0 have the form

(5.30) A1 [y, u](a0 ) + A2 [y, v](a0 ) = 0, A1 , A2 ∈ R, (A1 , A2 ) 6= (0, 0),

where u, v are real valued maximal domain functions satisfying [u, v](a0 ) = 1.
The inherited BC on (ar , br ) are obtained by replacing a0 by ar in (5.30)
and by using the Dirichlet conditions y(br ) = 0 at br . Note that, although
u, v and their quasi-derivatives may not be defined at a0 they are well defined
at any point a0 < ar < br < b0 .
• Assume b0 is LC and a0 is LP. Then there is no singular boundary condition
at a0 and all singular self-adjoint BC at b0 have the form

B1 [y, u](b0 ) + B2 [y, v](b0 ) = 0, B1 , B2 ∈ R, (B1 , B2 ) 6= (0, 0),

where u, v are real valued maximal domain functions satisfying [u, v](b0 ) = 1.
The inherited BC on (ar , br ) are obtained by replacing b0 by br in these
conditions and by using the Dirichlet conditions y(ar ) = 0 at ar . Note that,
although u, v and their quasi-derivatives may not be defined at b0 they are
well defined at any point br , a0 < ar < br < b0 .
• Each endpoint is either regular or LC. In this case we have self-adjoint real-
izations determined by both separated and coupled BC. Let the BC on (a0 , b0 )
be determined by (5.10) but with our changed notation a0 for a and b0 for b.
To obtain the inherited BC just replace a, b in (5.10) by (ar , br ). Note that
although the same matrices A, B occur in the BC on (a0 , b0 ) and on (ar , br )
when these inherited boundary conditions are written in the usual form (4.3)
for regular BC the coefficient matrices, say A = A(ar ), B = B(br ) depend on
values of u, v and their quasi-derivatives at ar , br . In particular as the end-
point ar or br is changed, e.g. in the code sleign2, the inherited boundary
conditions change accordingly.
• The inherited operators and their spectral quantities are identified with the
superscript i : Sri , λin (ar , br ).

DEFINITION 5.20 (Start of the essential spectrum). For any operator S let

(5.31) σ0 = inf σe , −∞ ≤ σ0 ≤ ∞,

where σe denotes the essential spectrum. Here σ0 = −∞ is interpreted to mean that

the essential spectrum is not bounded below and σ0 = ∞ means that the essential
spectrum is empty i.e. the spectrum consists entirely of isolated eigenvalues.
 85

THEOREM 5.21. Let S be a self-adjoint realization of (M, w) on (a0 , b0 ), let

Sri be the inherited operator on (ar , br ). Assume that σ(S) is bounded below and
discrete. Let σ(S) = {λn (S) : n ∈ N0 }, σ(Sr ) = {λn (Sri ) : n ∈ N0 }. Then

(5.32) λn (Sri ) → λn (S), as r → ∞, f or each n ∈ N0 .

PROOF. This is an as yet unpublished result of Everitt and Zettl [26]. 2

REMARK 18. We comment on the contrast between (5.32) and (5.29). This
markedly different behavior of the eigenvalues of regular problems shows the enor-
mous influence that the spectrum of a singular problem has on the regular problems
which are “close” to the singular one; in this case by virtue of the fact that the
endpoints of the regular problem are close to the endpoints of the singular one.
To understand the behavior of the eigenvalues of regular problems one needs a
perspective which includes the singular case. This is even more interesting when
viewed in the light of the asymptotic formula (5.15) for the eigenvalues on each
fixed interval (ar , br ).

THEOREM 5.22. Assume each endpoint is either R or LC and at least one

endpoint is O. Let S and Sr be as in Theorem 5.20. Then

(5.33) σ(S) = {λn : n ∈ Z} and λn (Sri ) → −∞, as r → ∞, n ∈ N0 .

Nevertheless we have: Given any λk ∈ σ(S) there exists an increasing (index)

sequence of positive integers n(r, k), depending on r and on λk such that

(5.34) λn(r,k) (Sri ) → λk as r → ∞.

PROOF. This is contained in Theorem 4.1 of [6]; see also Remark 1 (ii) on pages
15-16. 2

In this case, i.e. when the eigenvalues are unbounded below as well as above,
we follow the SLEIGN2 convention that λ0 denotes the smallest nonnegative eigen-
value. This makes the indexing scheme unique.

THEOREM 5.23. Let S, Sri be as in Theorem 5.21. Assume that

(5.35) −∞ < σ0 < ∞.

(Hence at least one endpoint is LP.)

86

1. If S has no eigenvalue less than σ0 , then

(5.36) λn (Sri ) → σ0 , r → ∞, n ∈ N0 .

2. If S has exactly one eigenvalue below σ0 , say λ0 < σ0 , then

(5.37) λ0 (Sri ) → λ0 , and λn (Sri ) → σ0 , as r → ∞, n ≥ 1.

3. If S has exactly two eigenvalues, say λ0 and λ1 , below σ0 , then

(5.38) λ0 (Sri ) → λ0 , λ1 (Sri ) → λ1, r → ∞; λn (Sri ) → σ0 , r → ∞, n ≥ 2.

4. etc
5. If S has an infinite number of eigenvalues {λn : n ∈ N0 } to the left of σ0 ,
then
(5.39) λn (Sri ) → λn , r → ∞, n ∈ N0 .

REMARK 19. Theorem 5.23 can be used to detect the number of eigenvalues to
the left of the essential spectrum. Using one of the numerical codes such as sleign2
or the Fulton and Pruess code sledge [30] one can ascertain which of (5.36), (5.37),
(5.38) or (5.39) holds.

REMARK 20. The so called Coffee-Evans equation has attracted a good deal of
attention in the literature because it has the interesting feature that, despite the
asymptotic behavior (5.15), the lower eigenvalues occur in clusters of three which are
close together. It is clear from Theorem 5.23 how to construct examples of regular
problems with clusters of three million or three billion eigenvalues as close together
as you please. This also shows that it easy to construct examples of regular SLP
which will defeat any numerical code for the computation of eigenvalues. See Zettl
[86] for some illustrations of this. (This paper can be downloaded from the author’s
web page following the instructions for downloading sleign2 given in Subsection 3
above.)

5.7 Approximating the spectrum of a given singular problem with eigen-

values of regular problems.

THEOREM 5.24. Let S be any self-adjoint realization of (M, w). The sequence
of inherited operators {Sri : r ∈ N} is spectral included for S i.e. given any λ ∈ σ(S)
there exists an n(r, λ) ∈ N0 for each r ∈ N such that
 87

(5.40) λn(r,λ) (Sri ) → λ, as r → ∞.

PROOF. This is contained in [6]. 2

THEOREM 5.25. Assume that σ(S) is bounded below. Then the sequence of
inherited operators {Sri : r ∈ N} is spectral exact for S below σ0 (S), i.e. it is spectral
included and if the convergence (5.40) holds for some λ < σ0 , then λ ∈ σ(S).
PROOF. This is contained in [6]. 2

By Theorem 5.23 any point of the spectrum of a singular problem can, in princi-
ple, be approximated arbitrarily closely by eigenvalues from the inherited sequence
of regular problems. In practice this isn’t feasible since there are an uncountable
number of points in the spectrum of the singular problem and finding an index
sequence for each one is a hopeless task.
Nevertheless Theorems (5.23), (5.24), (5.25) together than be used to approx-
imate the spectrum of many singular SLP quite effectively. This is done by ap-
proximating, not the individual points of the spectrum, but the spectral bands and
gaps. It is remarkable that for so many singular problems the first few spectral
bands and gaps - or the absence of gaps - can be detected and approximated from
the distribution of a few thousand eigenvalues of the inherited problems which can
be computed with sleign2 or sledge. For illustrations of this scheme see Zettl [88].
This paper can be downloaded as part of the SLEIGN2 package of files from the
internet; see the instructions in Subsection 3 of this Section.

5.8 Examples.

Here we list a few examples to illustrate some of the concepts and results discussed
above. These are taken from the the SLEIGN2 package. We follow the notation
established above. We supplement the endpoint classifications given above with the
weakly regular (WR) classification used by sleign2 : The endpoint a is WR if

1/p, q, w ∈ L(a, c), a < c < b,

but at least one of 1/p, q, w is not bounded in (a, c) for any c, a < c < b, or
w(a) = 0.
1. The Legendre equation on (−1, 1).

p(t) = 1 − t2 , q(t) = 1/4, w(t) = 1,

1 1+t
u(t) = 1, v(t) = log( ), −1 < t < 1.
2 1−t
88

The maximal domain functions u, v are solutions for λ = 0, -1 is LCNO, 1 is

LCNO. The BC

[y, u](−1) = −(py 0 )(−1) = 0, [y, u](1) = −(py)0 (1) = 0,

determines the Friedrichs extension whose eigenvalues are given by

λn = n(n + 1), n ∈ N0 ;

and the eigenfunctions are the classical Legendre polynomials.

The BC

[y, v](−1) = 0, [y, v](1) = 0,

is a separated non-Friedrichs BC. Analogues of the regular periodic and semi-

periodic BC are

[y, u](−1) = [y, u](1), [y, v](−1) = [y, v](1);

[y, v](−1) = [y, v](1), [y, u](−1) = [y, u](1);

respectively.
2. The Liouville form of the Bessel equation on (0, ∞).

ν 2 − 1/4
p = 1, q(t) = , w = 1.
t2

LP at ∞ for all ν.
At 0 :
• LCNO for −1 < ν < 1 but ν 2 6= 1/4
• R for ν 2 = 1/4
• LP for ν 2 > 1/4.

u(t) = tν+1/2 , v(t) = t−ν+1/2 , f or ν 6= 0, −1/2, 1/2,

u(t) = t, v(t) = 1, f or ν = −1/2,

u(t) = t, v(t) = −1, f or ν = 1/2,

√ √
u(t) = t, v(t) = t log(t), f or ν = 0.

For ν ≥ 0, u is the principal solution; thus [y, u](0) = 0 is the Friedrichs

condition at 0; there is no BC at ∞ since this endpoint is LP. But for −1/2 <
 89

ν < 0 note that v is the principal solution and hence [y, v](0) = 0 is the
Friedrichs BC.
3. The Halvorsen equation on (0, ∞).

e−2/t
p(t) = 1, q(t) = , w(t) = 1;
t4

0 is WR; ∞is LCNO. In this case

 the BC vector
 Y has the form Y (0) =
y(0) [y, u](∞)
at 0 and Y (∞) = where u(t) = 1, v(t) = t.
(py 0 )(0) [y, v](∞)
4. The Boyd equation on (−∞, 0) and on (0, ∞).

p = 1, q(t) = −1/t, w = 1

LP at −∞ and ∞; LCNO at 0+ and 0− .

u(t) = t, v(t) = 1 − t(log |t|)

Solutions can be given in terms of Whittaker functions, see [7]; the given
u, v are maximal domain functions which are not solutions for any λ. This
equation on the interval (−1, 1) , hence with an interior singularity at 0, arose
in a model in connection with the study of eddies in the atmosphere, see Boyd
[13].
5. The regularized Boyd equation on (−∞, 0) and on (0, ∞).

p = r2 , q = −r2 (log |t|), w = r2 ,

where

r(t) = e−(t log |t|)−t

LP at −∞ and ∞; W R at 0− and 0+ .
This is a WR form of example 4; the singularity at zero has been “regular-
ized” using quasi-derivatives. There is a one-to-one correspondence between
all self-adjoint BC of this example and example 4 considered as a “two inter-
val” problem. Each self-adjoint realization of the Boyd equation on (−∞, ∞)
is unitarily equivalent to a self-adjoint realization of the regularized Boyd
equation and conversely. Thus these problems have the same spectrum. Also
the eigenfunctions are closely related (but, of course not the same since one
is singular and the other regular). For details see [4], [7], [21], [74].
6. The Sears -Titchmarsh equation on (0, ∞).

p(t) = t, q(t) = −t, w(t) = 1/t

90

LP at 0; LCO at ∞.
This equation was studied by Titchmarsh [79] and by the two named au-
thors .
For problems on [1, ∞) the spectrum is discrete but unbounded above and
below; for some numerical results see [7].

cos(t) + sin(t) cos(t) − sin(t)

u(t) = √ , v(t) = √
t t

7. The BEZ equation on (−∞, 0) and on (0, ∞).

p(t) = t, q(t) = 1/t, w(t) = 1;

LP at −∞ and ∞; LCO at 0− and 0+ . See example 5 in [7] for some numerical

results.

u(t) = cos(log |t|), v(t) = sin(log |t|).

8. The Laplace tidal wave equation on (0, ∞).

k k2
p(t) = 1/t, q(t) = 2
+ , w = 1, k ∈ R, k 6= 0.
t t

LCNO at 0 for all k; LP at ∞ for all k. This is only a very special case of
the named equation, see Homer [48] for the general equation and references
to the applied literature.
Even for this special case there are no representations of solutions in terms
of the well known special functions. Thus to determine boundary conditions
one must use maximal domain functions. Such functions are given by

u(t) = t2 , v(t) = t − 1/k.

9. The Latzko equation on (0, 1).

p(t) = 1 − t7 , q = 0, w(t) = t7 ,

WR at 0 (since w(0) = 0); LCNO at 1. The singularity at 1 requires the use

of maximal domain functions such as

u(t) = 1, v(t) = − log(1 − t)

 91

to determine the boundary condition vsctor Y (1). This example has a long
and celebrated history; see Fichera [27].
10. A weakly regular equation on (0, ∞).

√ 1
p(t) = t, q = 0, w(t) = √ .
t

LP at ∞ and weakly regular at zero due to both p and w.

11. The Plum equation on (−∞, ∞).

p(t) = 1, q(t) = 100 cos2 (t), w(t) = 1

LP at −∞ and ∞.
This is a form of the Mathieu equation. Since both endpoints are LP there
are no boundary conditions needed or allowed; there is a unique self-adjoint
realization of this equation which has no eigenvalues. It’s spectrum consists
of an infinite number of compact intervals, the spectral bands, separated by
gaps. The first few spectral band are rather thin, the first has a width on the
order of 10−4 .
The eigenvalues of this problem on the interval (−b, b), 0 < b < ∞, with
Dirichlet boundary conditions

y(−b) = 0 = y(b),

tend to “bunch up” in the spectral bands of the whole line problem, partic-
ularly the fist band. See the comments in the last paragraph of Subsection
5.7. From Theorem 5.23 we know that

λn (b) → σ0 = inf σe .

In particular, for any positive ε and any positive integer n the first n
eigenvalues differ from each other by less than ε if b is sufficiently large, e.g.
one can choose b large enough so that the first 7 million eigenvalues agree
to the first 1000 digits. Given any numerical code for the computation of
Sturm-Liouville eigenvalues, it can be defeated simply by choosing a large
enough b. All this for such a “simple” regular SLP. The singular problem on
(−∞, ∞) exerts a strong influence on the behavior of the eigenvalues of the
regular problems on (−b, b).
12. The Mathieu equation on (−∞, ∞) and on (0, ∞).

p = 1, q(t) = sin(t), w = 1.

LP at −∞ and at ∞; R at 0.
92

This classical Mathieu equation has a celebrated history and voluminous

literature. There are no eigenvalues for this problem on (−∞, ∞). On (0, ∞)
there may be one eigenvalue depending on the boundary condition at 0. The
essential spectrum is the same for the whole line problem and the half line
problem and consists of an infinite number of disjoint compact intervals sepa-
rated by gaps. All the gaps are present. The endpoints of the spectral bands
and gaps are periodic and semi-periodic eigenvalues of the problem on the
interval [0, 2π]. These can be computed with sleign2.
The above remarks, with some appropriate modifications, apply to the gen-
eral socalled Hill’s equation. This is the SL equation with periodic coefficients
p, q, w all of the with the same period.
Of special intereat is the starting point of the essential spectrum σ0 , which
is finite in these cases for p > 0, w > 0. This point σ0 is also the “oscillation
number” of the equation; this means the equation is NO for λ < σ0 and
O for λ > σ0 . It may be O or NO for λ = σ0 . For the Mathieu equation
considered here (p = 1 = w, q(t) = sin(t)) σ0 is not known explicitly but is
approximately σ0 ∼ −0.378. This can be checked by computing the lowest
eigenvalue of the periodic problem on [0, 2π] i. e. σ0 = λP 0 (0, 2π).
13. The hydrogen atom equation on (0, ∞).

k h
p(t) = 1, q(t) = + 2 , w(t) = 1, k, h ∈ R.
t t

This is the two parameter version of the classical one-dimensional equation for
quantum theory modelling of the hydrogen atom; see [49], section 10 where
most of the results reported on here can be found. A few of these results can
be found in the commentary file xamples.tex of the package of files comprising
the SLEIGN2 package. For all h, k there are no positive eigenvalues, ∞ is LP
and the essential spectrum is [0, ∞). If k = 0 the equation reduces to Bessel,
see example #2 with h = ν 2 − 1/4.
LP at ∞ for all h, k.
At 0 :
• R for h = 0 = k,
• LCNO for h = 0 and all k 6= 0
• LCNO for −1/4 ≤ h < 3/4, but h 6= 0 and all k
• LCO for h < −1/4 and all k
• LP for h ≥ 3/4 and all k.
Let

p
ρ= h + 1/4, f or h ≥ −1/4.

: (a) For h ≥ 3/4 and k ≥ 0 there is at most one negative eigenvalue and
λ = 0 may be an eigenvalue; for h ≥ 3/4 and k < 0 there are infinitely
many negative eigenvalues given by

−k 2
λn = , n ∈ N0 ,
(2n + 2ρ + 1)2
 93

and λ = 0 is not an eigenvalue.

: (b) For h = 0, u(t) = t, v(t) = 1+kt log t. For some computed eigenvalues
see [7] and [49], section 10.
: (c) −1/4 < h < 3/4, 0 < ρ < 1 but ρ 6= 1/2. All the following results hold
for the non-Friedrichs boundary condition : [y, v](0) = 0, where

k
u(t) = tρ+1/2 , v(t) = t1/2−ρ + t3/2−ρ .
1 − 2ρ

(a) k > 0, 0 < ρ < 1/2 there are no negative eigenvalues

(b) k > 0, 1/2 < ρ < 1 there is exactly one negative eigenvalue given by

−k 2
λ0 =
(2ρ − 1)2

(c) if k < 0, 0 < ρ < 1/2 there are infinitely many negative eigenvalues
given by

−k 2
λn = , n ∈ N0
(2n − 2ρ + 1)2

(d) if k < 0, 1/2 < ρ < 1 there are infinitely many negative eigenvalues
given by

−k 2
λn = , n ∈ N0
(2n − 2ρ + 3)2

The next few results refer to the BC

A1 [y, u](0) + A2 [y, v](0) = 0.

(e) if k = 0 and A1 A2 < 0 ther is exactly one negative eigenvalue given

by :

 1/ρ
−A1 Γ(1 + ρ)
λ0 = −4
A2 Γ(1 − ρ)

(f) Note that for h = −1/4, k ∈ R, the LCNO classification at 0 holds.

Here

√ √ √ √ √
u(t) = t + kt t, v(t) = 2 t + ( t + kt t) log(t).
94

For k = 0 and A1 A2 < 0 there is exactly one negative eigenvalue

given by

λ0 = c e2A1 /A2 , c = 4 e4−2γ , γ = 0.5772156649...,

is Eulers constant.
(g) h < −1/4, k ∈ R, the equation is LCO at 0. For k = 0 this equation
reduces to the Krall equation, see example 20. For k 6= 0 explicit
formulas for the eigenvalues are not available; some qualitative prop-
erties of the spectrum are :
for all k there are infinitely many negative eigenvalues going expo-
nentially to −∞
for k > 0 the point 0 is not an accumulation point of eigenvalues
for k ≤ 0 the eigenvalues also accumulate at 0.
14. The Marletta equation on (0, ∞).

3(t − 31)
p = 1, q(t) = , w = 1.
4(t + 1)(4 + t)2

This equation is R at 0 and LP at ∞. For some boundary conditions at 0

this equation deceives both codes sleign (not sleign2) and sledge into falsly
reporting λ = 0 as an eigenvalue since there is a “near” eigenfunction there.
For details see Marletta’s certification report for the code sleign [66].
15. The harmonic oscillator equation on (−∞, ∞).

p = 1, q(t) = t2 , w = 1

LP at −∞ and at ∞. Thus there is a unique self-adjoint extension; it has

discrete spectrum given by

λn = 2n + 1, n ∈ N0 .

See [79] for a classical treatment.

16. The Jacobi equation on (−1, 1).

p(t) = (1 − t)α+1 (1 + t)β+1 , q = 0, w(t) = (1 − t)α (1 + t)β

At -1 for all α :
• LP for β ≤ −1 and for β ≥ 1
• WR for −1 < β < 0
• LCNO for 0 ≤ β < 1
At +1 for all β:
 95

• LP for α ≤ −1 and for α ≥ 1

• WR for −1 < α < 0
• LCNO for 0 ≤ α < 1.
The boundary condition functions u, v can be taken as follows :
• For t < 0 :
: (i) if −1 < β < 0 then u(t) = (1 + t)−β , v(t) = 1
1+t
: (ii) if β = 0 then u(t) = 1, v(t) = log 1−t
: (iii) if 0 < β < 1 then u(t) = 1, v(t) = (1 + t)−β
• For t > 0 :
: (i) if −1 < α < 0 then u(t) = (1 − t)−α , v(t) = 1
1+t
: (ii) if α = 0 then u(t) = 1, v(t) = log 1−t
: (iii) if 0 < α < 1 then u(t) = 1, v(t) = (1 − t)−α .
To get the classical Jacobi polynomials take −1 < α, −1 < β; then note
the following endpoint classifications and required boundary conditions:
At +1:
(a) −1 < α < 0, W R, (py 0 )(1) = 0
(b) 0 ≤ α < 1, LCN O, [y, u](1) = 0
(c) 1 ≤ α, LP
At -1:
(a) −1 < β < 0, W R, (py 0 )(−1) = 0
(b) 0 ≤ β < 1, LCN O, [y, u](−1) = 0
(c) 1 ≤ β, LP.
For the classical Jacobi orthogonal polynomials the eigenvalues are given
by:

λn = n(n + α + β + 1), n ∈ N0 .

It is interesting to observe that the required boundary condition for the Jacobi
polynomials is the Friedrichs condition in the LCNO case but not in the WR
case.
17. The rotation Morse oscillator on (0, ∞).

2
p = 1, w = 1, q(t) = − 2000(2E − E 2 ), E = e−1.7(t−1.3) .
t2

LP at 0 and ∞. Hence there is a unique self-adjoint realization; its essential

spectrum is [0, ∞] and it has 26 negative eigenvalues. (Ask sleign2 to compute
the first 28 eigenvalues and note the appearance of the 26 negative eigenvalues
and the starting point of the essential spectrum at 0.)
18. The Dunsch equation on (−1, 1). See Dunford and Schwartz [17] chapter VIII,
pp. 1510-1520 for a discussion of this problem.

2α2 2β 2
p(t) = 1 − t2 , q(t) = + , w = 1, 0 ≤ α, 0 ≤ β.
1+t 1−t

At -1:
96

LP for α ≥ 1/2 and all β

LCNO for 0 ≤ α < 1/2 and all β
At +1:
LP for β ≥ 1/2 and all α
LCNO for 0 ≤ β < 1/2 and all α.
Boundary condition functions can be obtained as follows:

At − 1 : u− (t) = (1 + t)α , v− (t) = (1 + t)−α ,

At + 1 : u+ (t) = (1 + t)β , v+ (t) = (1 + t)−β .

Note that u, v are maximal domain functions but not solutions. In [17] on
p. 1519 it is claimed that the boundary value problem determined by the
boundary conditions

[y, u− ](−1) = 0 = [y, u+ ](1)

has eigenvalues given by

λn = (n + α + β + 1) (n + α + β), n ∈ N0 .

19. The Donsch equation on (−1, 1). This is a modification of example 18 which
illustrates an LCNO/LCO mix. Replace α in 18 by iγ. This changes the
singularity at -1 from LCNO to LCO. For γ > 0 and 0 < β < 1/2 we have

At − 1 : u(t) = cos(γ log(1 + t)), v(t) = sin(γ log(1 + t))

At + 1 : u(t) = (1 − t)β , v(t) = (1 − t)−β .

These u, v are maximal domain functions which are not solutions.

20. The Krall equation on (0, ∞). This example is a special case of the Bessel
equation, see example 2 above. Its solutions can be obtained in terms of
modified Bessel functions. (We have followed the sleign2 package here by
adding the constant 1 to q, this is done in sleign2 to facilitate numerical
computations.)

k 2 + 1/4
p = 1, q(t) = 1 − , w = 1, k ∈ R, k 6= 0.
t2

LCO at 0 and LP at ∞ for all k 6= 0. The essential spectrum is [1, ∞).

For the boundary condition

[y, u](0) = 0, u(t) = t1/2 cos(k log(t))

 97

there are an infinite number of eigenvalues which cluster at −∞ and at 1.

(Note that u need only be a maximal domain function on [0, d) for 0 <
d < ∞. To get a maximal domain function on (0, ∞) one can patch u at d
appropriately.) The eigenvalues approach −∞ and 1 very rapidly, see [7] or
use sleign2 for more details; also see [60] for more information.
21. The Fourier equation. See Subsection 4.3 for a description of the eigenvalues
for various boundary conditions, self-adjoint and otherwise.
22. The Laguerre equation on (0, ∞).

p(t) = tα+1 e−t , q = 0, w(t) = tα e−t , α ∈ R.

LP at ∞ for all α.
At 0 :
• LP for α ≤ −1
• WR for −1 < α < 0
• LCNO for 0 ≤ α < 1
• LP for α > 1.
This is the classical form of the celebrated equation, which for parameter
values α > 1 produces the Laguerre polynomials as eigenfunctions; for the
appropriate boundary at 0, when needed, the eigenvalues are given by

λn = n, n ∈ N0 .

Remarkably, these are independent of α, see Abramovitz and Stegun [1], chap-
ter 22, section 22.6 for more details. See the file xamples.f (this is not a typo)
of the sleign2 package for details of the boundary condition functions u, v.
The code sleign2 has only very limited success with this problem; for nu-
merical computations the Laguerre/Liouville equation, which has the same
eigenvalues (for the appropriate corresponding boundary conditions) is more
convenient, see example 23 to follow.
23. The Laguerre/Liouville equation on (0, ∞).

α2 − 1/4 α + 1 t2
p = 1, w = 1, q(t) = − + , α ∈ R.
t2 2 16

LP at ∞ for all α
LCNO for −1 < α < 1 but α2 6= 1/4
R for α2 = 1/4
LP for α ≥ 1
See the xamples.f file of the sleign2 package for details of appropriate
boundary condition functions.
24. Jacobi/Liouville form of the Jacobi equation. See the files xamples.f and
xamples.tex of the sleign2 package for details.
25. The Meissner equation on (−∞, ∞).
98


1 f or t<0
p = 1, q = 0, w =
9 f or t≥0

LP at −∞ and ∞.
This equation is well known in the applied literature in connection with
the modelling of crystals in one dimension, see [19], [47].
Periodic boundary conditions on (−1/2, 1/2). We have λ0 = 0 and

λ4n+1 = (2nπ + α)2 , λ4n+2 = (2(n + 1)π − α)2 ,

−7
λ4n+3 = λ4n+4 = (2(n + 1)π)2 , α = cos−1 ( ), n ∈ N0 .
8

Note that there are infinitely many simple and infinitely many double periodic
eigenvalues on the interval (−1/2, 1/2).
Semi-Periodic eigenvalues on (−1/2, 1/2).

λ4n = (2nπ + β)2 , λ4n+1 = (2nπ + γ)2 , λ4n+2 = (2(n + 1)π − γ)2 ,
√ √
2 −1 1 + 33 −1 1 − 33
λ4n+3 = (2(n + 1)π − β) , β = cos ( ), γ = cos ( ).
16 16

Observe that the semi-periodic eigenvalues are all simple.

26. The Lohner equation on (−∞, ∞).

p = 1, w = 1, q(t) = 1000t

LP at −∞ and ∞. Lohner in [65] computed eigenvalues of this equation on a

compact interval with regular boundary conditions using interval arithmetic.
He obtained good approximations with rigorously guarenteed upper and lower
bounds.
27. The Jörgens equation on (−∞, ∞). This is a remarkable example from Jörgens,
see [49], part II, section 10.

1 2t
p = 1, w = 1, q(t) = e − k et , k ∈ R.
4

LP at −∞ and at ∞.
The essential spectrum starts at 0; for k ≤ 1/2 there are no eigenvalues;
for

h < k − 1/2 ≤ h + 1, h = 0, 1, 2, 3, . . .

there are exactly h + 1 eigenvalues and these are all below the essential spec-
trum i.e. they are all negative. They can be given explicitly by
 99

λn = −(k − 1/2 − n)2 , n = 0, 1, 2, 3, . . . , h, h ∈ N0 .

28. The Behnke-Goerisch equation on (−∞, ∞).

p = 1, w = 1, q(t) = k cos2 (t), k ∈ R.

This is a form of the Mathieu equation previously discussed, see examples 11

and 12 above. These authors computed Neumann eigenvalues on a compact
interval using interval arithmetic and obtained good approximations with
rigorous upper and lower bounds.
29. The Whittaker equation on (0, ∞).

1 k2 − 1 1
p = 1, q(t) = + , w(t) = , k ∈ N.
4 4t2 t

LP at 0 and ∞.
This equation is studied in [49], part II, section 10. The spectrum is discrete
and can be given explicitly by

k+1
λn = n + , n ∈ N0 .
2

5.9 Comments.

Much of Section 5 is based on three papers : Bailey, Everitt, Weidmann and Zettl
[6], Niessen and Zettl [74] and the pre-print of Everitt and Zettl [26].
1. Comments are made separately for each subsection.
2. This treatment of principal and non-principal solutions is based on Niessen
and Zettl [74]. According to Hartman [42] these terms were coined by Leighton
in [64]. The principal solution is the “small” solution but using a term such
as small might lead to confusion since the same solution may be small at one
endpoint but not the other.
3. The structure of singular limit circle boundary conditions is well known, see
[81], [72], [2], [61] but it is not easy to find a clear and comprehensive treatment
of them in the literature. We hope that this paper makes a positive contribu-
tion in this area. The conditions (5.10) characterize all self-adjoint boundary
conditions for the case when each endpoint is either R or LC (either LCO or
LCNO). The canonical form (5.16) and (5.17) represents all separated self-
adjoint BC; (5.20) is the canonical form of all coupled self-adjoint BC, both
for the case R or LC/ R or LC. Clearly (5.16) is a canonical form of (5.8) and
(5.17) represents (5.9). For a different representation of singular self-adjoint
BC see the treatment of Dunford and Schwartz [17].
100

4. The characterization of the Friedrichs extension given here is based on Niessen

and Zettl [73]. This in turn is based on Rellich [77]. Also see Kaper, Kwong
and Zettl [51], Kalf [50], [78]. For a characterization of the Friedrichs ex-
tension of powers of some special Sturm-Liouville operators see Baxley [10].
Besides this paper of Baxley this writer is aware of only one other result
which characterizes the Friedrichs extension in terms of boundary conditions
for singular problems of order higher than two. This is an unpublished paper
of Zettl [88] for a very special class of problems. An illustrative example is,
with n > 1,

c
(−1)n y (2n) ± y = λ y on 0 < t < 1, c ∈ R, c 6= 0.
t

For the regular case very general results are known, see Niessen and Zettl [73],
Möller and Zettl [71].
5. Möller [68] has shown that, for regular as well as singular SLP, if either p
changes sign on the underlying interval then the spectrum is not bounded
below. This holds even if there is no subinterval of the underlying interval on
which p is negative. The corresponding result for very general higher order
problems was established by Möller and Zettl in [70].
Inequalities (5.19), (5.22), (5.23) can be found in Weidmann [81] for the
regular case with

 
c
K= 1 , c 6= 0, c ∈ R.
c

For earlier work see Jörgens [49], Rellich [77]. These were extended to the
singular case by Niessen and Zettl in [73], then further extended to more
general K by Bailey, Everitt and Zettl [8], and finally Eastham established
the general case [18].
In [8] there is also the characterization of the eigenvalues given by (5.27)
for the case when the expression M is symmetric and w > 0. It was only
during the writing of these notes that the author realized that the proof given
in [8] still holds, in the regular case, for complex valued p, q, w.
Fore an extension of the asymptotic formula (5.15) to the case when p or w
are allowed to change sign see Atkinson and Mingarelli [5]. Additional terms
in this formula can be obtained if the coefficients satisfy some smoothness
assumtions, see Harris [40], Hartman [42], and the references therein.
6. This Subsection is based on the unfinished paper of Everitt and Zettl [26]
which in turn is based, to a considerable extent, on Bailey, Everitt, Weidmann
and Zettl [6].
7. This Subsection is also based on [26].
8. These examples were taken from the sleign2 files: xamples.f, xamples.tex.
See Subsection 5.3 for instructions on how to download these files from the
internet.
 101

5.10 Comments on some topics not covered.

1. The LP/LC dichotomy. See the monograph by Kauffman, Read and Zettl [53]
for a brief introduction to LP and LC criteria. There are many sufficient con-
ditions known for LC and also for LP and even some necessary and sufficient
conditions but no necessary and sufficient conditions which can be checked in
each case. Finding such conditions is still an open problem. One obstacle to
finding such conditions is that for the LP case to hold it is enough to give
conditions on a sequence of intervals, with almost no requirements on the co-
efficients outside these intervals, whereas essentially pointwise conditions are
required for the LC case. For an entirely different approach see Zettl [85] for
a construction of all LC expressions.
2. The O/NO alternative. See Kauffman, Read and Zettl [53] for a brief intro-
duction to O/NO criteria, also see the monograph of Havorsen and Mingarelli
[38] and its references. Many conditions, both necessary and sufficient, are
known but there are no known necessary and suffcient conditions which can
be checked in each case. As in the LC/LP situation the most general sufficient
conditions for O are of “interval” type, see e.g. Kwong and Zettl [62], but
such conditions are not appropriate for the NO case to hold. For a completely
different approach i.e. a constuction of all disconjugate (NO) equations see
Zettl [84].
3. Absolutely continuous spectrum. See the contributions to these Proceedings
from Last and Jitomirskaya and from Stolz. We have only considered the
simplest division of the spectrum into its discrete and essential (continuous)
parts. See the seminal paper of Gilbert and Pearson [36] for criteria involving
subordinate solutions - an extension of principal solutions - for the absolutely
continuous spectrum; see also Hinton and Shaw [46], [44], [45], Gesztesy, Gu-
rarie, Holden, Klaus, Sadun, Simon, and Vogl [33] and the references therein.
There is an extensive literature on spectral properties of Sturm-Liouville op-
erators by Simon and his 100 co-authors.
4. There is also a vast literature on so called left definite problems, when the
weight function w is allowed to change sign. See the monograph by Mingarelli,
and its references. Here the operator theory is studied in the setting of Krein
and Pontryagen spaces rather than Hilbert space. For oscillatory properties
of the eigenfunctions when the weight function changes sign or is identically
zero on subintervals, see Everitt, Kwong and Zettl, [22].
5. Another popular topic we have not discussed is the two, or multi, parameter
theory. See Volkmer [80] and the papers by Binding and Sleeman.
6. Multi interval theory. See the contribution of Everitt, Shubin, Stolz and Zettl
to these Proceedings and the references therein for an introduction to SLP
problems on infinitely many intervals. Also see Geszteszy and Kirsch [34],
[35], [32]. If an equation has an interior singularity then, in general, the
solutions cannot be continuously moved through this singularity. One can
study this problem on two separate intervals each of wich has this singular
point on the boundary. Take the direct sum of two self-adjoint operators from
the two intervals and you have a two-interval self-adjoint operator which is
not particularly interesting because it doesn’t “connect” the two intervals
together. More interesting operators are obtained by connecting solutions
102

through the singular point, even if they blow up there, in such a way as to get
a new “two interval” self-adjoint operator. Actually this construction is also
of interest when the interior point is not singular but regular, obtaining what
is known as “point interactions”. For an early contribution see Zettl [82].
7. Discreteness criteria. See the contribution of Read to these Proceedings and
the references therin. Also see [63].
8. Inverse spectral theory. See the landmark paper of Gelfand and Levitan
[31], the elegant exposition of Pöschel and Trubowitz [76] and the references
therein.

References
[1] A.B. Abramowitz and I.A. Stegun. Handbook of Mathematical Functions with Formulas and
Mathematical Tables. Dover Publications Inc., New York, 1970.
[2] N. I. Akhiezer and I. M. Glazman. Theory of linear operators in Hilbert space, Volumes I
and II. Pitman and Scottish Academic Press, London and Edinburgh, 1980.
[3] F. V. Atkinson. Discrete and continuous boundary value problems. Academic Press, New
York /London, 1964.
[4] F. V. Atkinson, W. N. Everitt, and A. Zettl. Regularisation of a Sturm-Liouville problem with
an interior singularity using quasi-derivatives. Differential and Integral Equations 1 (1988),
213-221.
[5] F. V. Atkinson and A. B. Mingarelli. Asymptotics of the number of zeros and of the eigen-
values of general weighted Sturm-Liouville problems. J. reine angew. Math. 375/376 (1987),
380-393.
[6] P. B. Bailey, W. N. Everitt, J. Weidmann, and A. Zettl. Regular approximations of singular
Sturm-L iouville problems. Results in Mathematics, v. 23 (1993), 3-22.
[7] P. B. Bailey, W. N. Everitt, and A. Zettl. Computing eigenvalues of singular Sturm-Liouville
problems. Results in Mathematics, v. 20 (1991), 391-423.
[8] P. B. Bailey, W. N. Everitt, and A. Zettl. Regular and singular Sturm-Liouville problems
with coupled boundary conditions. Proc. Royal Soc. of Edinburgh (A) 126 (1996), 505-514.
[9] P. B. Bailey, W. N. Everitt, and A. Zettl. Sleign2: An eigenfunction - eigenvalue code for
singular Sturm-L iouville problems. pre-print.
[10] J. V. Baxley. Eigenvalues of singular differential operators by finite difference methods I. J.
Math. Anal. Appl. 37 (1972), 244-254.
[11] P. R. Beesack. Gronwall inequalities. Carleton University, Dept. of Mathematics, Carleton
Mathematical Lecture Notes, no. 11, Ottawa, 1975.
[12] A. Bielecki. Une remarque sur la methode de Banach-Cacciopoli-Tikhonov dans la therie des
equations differentielles ordinaires. Bull. Acad. Polon. Sci. Cl. III 4(1956), 261-264.
[13] J. P. Boyd. Sturm-Liouville eigenvalue problems with an interior pole. J. Math. Physics 22
(1981), 1575-1590.
[14] E. A. Coddington and N. Levinson. Theory of ordinary differential equations. McGraw-Hill,
New York /London/Toronto, 1955.
[15] M. Dauge and B. Helffer. Eigenvalues V ariation. I. Neumann Problem for Sturm-Liouville
Operators. J. Diff. Equations, v. 104 (1993), 243-262.
[16] J. Dieudonne. Foundations of Modern Analysis. Academic Press, New York/London, 1969.
[17] N. Dunford and J. T. Schwartz. Linear operators , vol.II. Wiley, New York, 1963.
[18] M. S. P. Eastham. Eigenvalue inequalities for coupled Sturm-Liouville problems. private com-
munication.
[19] M. S. P. Eastham. Limit-circle differential expressions of the second-order with an oscillating
coefficient. Quart. J. Math. Oxford (2) 24 (1973).
[20] W. Eberhard and G. Freiling. Stone-regulare Eigenwertprobleme. Math. Z. 160(1978), 139-
161.
 103

[21] W. N. Everitt, J. Gunson, and A. Zettl. Some comments on Sturm-Liouville eigenvalue prob-
lems with interior singularities. Zeitschrift fuer Angewandte Mathematik und Physik (ZAMP)
v. 38 (1987),813-838.
[22] W. N. Everitt, M. K. Kwong, and A. Zettl. Oscillation on eigenfunctions of weighted regular
Sturm-Liouville problems. J. London Math. Soc. 27(1983), 106-120.
[23] W. N. Everitt, M. Moeller, and A. Zettl. Discontinuous dependence of the n-th Sturm-
Liouville eigenvalue. to appear.
[24] W. N. Everitt and F. Neuman. A concept of adjointness and symmetry of ordinary differential
expressions based on the generalized Lagrange identity and Green’s formula. Lecture Note in
Math. 1032 , Springer Verlag, Berlin (1983), 161-169.
[25] W. N. Everitt and D. Race. On necessary and sufficient conditions for the existence of
Caratheodory solutions of ordinary differential equations. Quaestiones Math. 3, no.2, (1976),
507-512.
[26] W. N. Everitt and A. Zettl. Behavior of eigenvalues of regular Sturm-Liouville problems near
a singular boundary. pre-print.
[27] G. Fichera. Numerical and quantitative analysis. Pitman Press, London, 1978.
[28] K. O. Friedrichs. Spectraltheorie halbbeschraenkter Operatoren und Anwendungen auf die
Spektralzerlegung von Differentialoperatoren. I,II. Math. Ann.109 (1933/34), 465-487. Cor-
rections, ibid. 110 (1934/35), 777-779.
[29] K. O. Friedrichs. Ueber die ausgezeichnete Randbedingung in der Spektraltheorie der
halbbeschraenkten gewoehnlichen Differentialoperatoren zweiter Ordnung. Math. Ann. 112
(1935/36), 1-23.
[30] C.T. Fulton and S. Pruess. Mathematical software for Sturm-Liouville problems. ACM. Trans.
Math. Software 19 (1993), 360-376.
[31] I. M. Gelfand and B. M. Levitan. On the determination of a differential equation from its
spectral function. English transl., AMS transl. (2), 1 (1955), 253-304.
[32] F. Gesztesy. On the one-demensional Coulomb Hamiltonian. J. Physics A 13 (1980), 867-
875.
[33] F. Gesztesy, D. Gurarie, H. Holden, M. Klaus, L. Sadun, B. Simon, and P. Vogel. Trapping
and cascading of Eigenvalues in the Large Coupling Limit. CMP 118 (88), 597-634.
[34] F. Gesztesy and W. Kirsch. One-dimensional Schroedinger operators with interactions on a
discrete set. J. fuer Mathematik 362 (1985), 28-50.
[35] F. Gesztesy, C. Macedo, and L. Streit. An exactly solvable periodic Schroedinger operator.
J. Phys. A: Math. Gen. 18 (1985), 503-507.
[36] D. J. Gilbert and D. B. Pearson. On subordinacy and analysis of the spectrum of one-
dimensional Schroedinger operators. J. Math. Anal. Appl. 128 (1987), 130-56.
[37] S. G. Halvorsen. A function-theoretic property of solution of the equation x” + (l w-q)x=0.
Quart. J. Math. Oxford (2), 38 (1987), 73-76.
[38] S. G. Halvorsen and A. B. Mingarelli. Non-Oscillation Domains of Differential Equations
with Two Parameters. Lecture Notes in Mathematics no 1338, Springer-Verlag, Berlin, New
York, 1988.
[39] B. J. Harris. Eigenvalue problems for second-order differential equations. Proc. 1986-87 Fo-
cused Research Program on Spectral Theory and Boundry Value Problems, ANL 87-26.
[40] B. J. Harris. On the asymptotic properties of linear differential equations. Mathematika 34
(1987), 187-198.
[41] P. Hartman. On the number of L2-solutions of x” + q(t)x =l x. Amer. J. Math. 73 (1951),
635-645.
[42] P. Hartman. Ordinary differential equations. Wiley and Sons, Inc., New York /Lon-
don/Sydney, 1964.
[43] E. Hille. Lectures on ordinary differential equations. Addison-Wesley, London, 1969.
[44] D. B. Hinton and J. K. Shaw. Absolutely continuous spectra of perturbed periodic Hamil-
tonian systems. Rocky Mountain J. Math. 17, no.4, (1987), 727-748.
[45] D. B. Hinton and J. K. Shaw. Absolutely continuous spectra of second-order differential
operators with long and short range potentials. SIAM J. Math. Anal. 17 (1986), 182-196.
[46] D. B. Hinton and J. K. Shaw. On the absolutely continuous spectrum of the perturbed Hill’s
equations. Proc. London Math. Soc. (3) 50 (1985), 175-192.
[47] H. Hochstadt. A special Hill’s equation with discontinuous coefficients. Amer. Math. Monthly,
70 (1963), 18-26.
104

[48] M. S. Homer. Boundry value problems for the Laplace tidal wave equation. Proc. Royal Soc.
London Ser. A 428 (1990), 157-180.
[49] K. Joergens. Spectral theory of second order ordinary differential operators. Lectures delivered
at Aarhus Universitet 1962/63, Matimatisk Institut Aarhus 1964.
[50] H. Kalf. A characterization of the Friedrichs extension of Sturm-Liouville operators. J. Lon-
don Math. Soc. (2) 17 (1978), 511-521.
[51] H. G. Kaper, M. K.Kwong, and A. Zettl. Characterization of the Friedrichs extensions of
singular Sturm-Liouville expressions. SIAM J. Math. Anal., 7(1986), 772-777.
[52] H. G. Kaper and M. K. Kwong. Asymptotics for the Titchmarsh-Weyl m-coefficient for
integrable potentials. Proc. Royal Soc. Edinburgh 103A (1986), 347-358.
[53] R. M. Kauffman, T. T. Read, and A. Zettl. The deficiency index problem for powers of
ordinary differential expressions. Lecture Notes in Mathematics 621, Springer-Verlag.
[54] Q. Kong, H. Wu, and A. Zettl. Dependence of eigenvalues on the problem. Mathematische
Nachrichten, (to appear).
[55] Q. Kong and A. Zettl. Dependence of eigenvalues of Sturm-Liouville problems on the bound-
ary. J. Differential Equations, vol.126,no.2, (1996), 389-407.
[56] Q. Kong and A. Zettl. The derivative of the exponential map of matrices. pre-print.
[57] Q. Kong and A. Zettl. Eigenvalues of regular Sturm-Liouville problems. J. Differential Equa-
tions, vol. 131, no.1, (1996), 1-19.
[58] Q. Kong and A. Zettl. Interval oscillation conditions for difference equations. SIAM J. Math.
Analysis, 26, no.4 (1995), 1047-1060.
[59] Q. Kong and A. Zettl. Linear ordinary differential equations. WSSIAA v.3, (1994), (Special
volume dedicated to W.Walter), Inequalities and Applications, edited by R. P. Agarwal, 381-
397.
[60] A. M. Krall. Boundary value problems for an eigenvalue problem with a singular potential.
J. Diff. Equations, 45 (1982), 128-138.
[61] A. M. Krall and A. Zettl. Singular self-adjoint Sturm-Liouville problems. Differential and
Integral Equations v. 1 (1988), 423-432.
[62] M. K. Kwong and A. Zettl. Asymptotically constant functions and second-order linear oscil-
lation. J. of Math. Anal. and Applications, 93 (1983), 475-494.
[63] M. K. Kwong and A. Zettl. Discreteness conditions for the spectrum of ordinary differential
operators. J. Diff. Equations v. 40 n. 1 (1981), 53-70.
[64] W. Leighton. Principal quadratic functionals. Trans. Amer. Math. Soc. 67 (1949), 253-274.
[65] R. J. Lohner. Verified solution of eigenvalue problems in ordinary differential equations. pri-
vate communication.
[66] M. Marletta. Numerical tests of the SLEIGN software for Sturm-Liouville problems. ACM
TOMS, v17 (1991), 501-503.
[67] R. Mennicken and M. Moeller. Nonselfadjoint boundary eigenvalue problems. preprint, Lon-
don, 1997.
[68] M. Moeller. On the unboundedness below of the Sturm-Liouville operator. Differential and
Integral Equations, (to appear).
[69] M. Moeller and A. Zettl. Differentiable dependence of eigenvalues of operators in Banach
spaces. J. Operator Theory, (to appear).
[70] M. Moeller and A. Zettl. Semi-boundedness of ordinary differential operators. J. Differential
Equations, v.115, no.1, (1995), 24-49.
[71] M. Moeller and A. Zettl. Symmetric differential operators and their Friedrichs extension. J.
Differential Equations, v.115, no.1, (1995), 50-69.
[72] M. A. Naimark. Linear Differential Operators. Ungar, New York, 1968.
[73] H.-D. Niessen and A. Zettl. The Friedrichs extension of regular ordinary differential operators.
Proc. Roy. Soc. Edinburgh, 114A, (1990), 229-236.
[74] H.-D. Niessen and A. Zettl. Singular Sturm-Liouville problems: The Friedrichs extension and
comparison of eigenvalues. Proc. London Math. Soc. v.64, (1992), 545-578.
[75] I. G. Petrovski. Ordinary differential equations. Prentice Hall, Inc., London, 1966.
[76] J. Poeschel and E. Trubowitz. Inverse Spectral Theory. Academic Press, New York /Lon-
don/Sydney, 1987.
[77] F. Rellich. Halbbeschraenkte gewoehnliche Diffferentialoperatoren zweiter Ordnung. Math.
Ann. 122 (1950/51), 343-368.
 105

[78] R. Rosenberger. Charakterisierung der Friedrichsfortsetzung von halbbeschraenkten Sturm-

Louiville Opatoren. Dissertation, Technische Hochschule Darmstadt, 1984.
[79] E. C. Titchmarch. Eigenfunction Expansions Associated with Second Order Differential
Equations, Part II, (2nd ed). Clarendon Press, Oxford, 1962.
[80] Hans Volkmer. Multiparameter Eigenvalue Problems and Expansion Theorems. Lecture Notes
in Mathematics no 1356, Springer-Verlag, Berlin, New York, 1988.
[81] J. Weidmann. Spectral Theory of Ordinary Differential Operators. Lecture Notes in Mathe-
matics 1258. Springer Verlag, Berlin, 1987.
[82] A. Zettl. Adjoint and self-adjoint problems with interface conditions. SIAM J. Applied Math.
16, (1968), 851-859.
[83] A. Zettl. Adjointness in nonadjoint boundary value problems. SIAM J. Applied Math. 17,
(1969), 1268-1279.
[84] A. Zettl. An algorithm for the construction of all disconjugate operators. Proc. Royal Soc.
Edinburgh, 75A, 4, (1976).
[85] A. Zettl. An algorithm for the construction of limit circle expressions. Proc. Royal Soc.
Edinburgh, 75A, 1, (1976).
[86] A. Zettl. Computing continuous spectrum. Proc. International Symposium on Trends and
developments in ordinary differential equations, Yousef Alavi and Po-Fang Hsieh, editors,
World Scientific,(1993), 393-406.
[87] A. Zettl. Formally self-adjoint quasi-differential operators. Rocky Mountain J. Math. 5(1975),
453-474.
[88] A. Zettl. On the Friedrichs extension of singular differential operators. pre-print.