- 1.

Computational Procedure for Sturm—Liouvflle Problem

MIKHAIL D. MIKHAILOV AND NIKOLAI L. V ULCIIANOV

A p p l i e d M a t h e m a t i c s C e n t r e , S a f i n , H u l g i j r i Received September 24, I981

The approaches or DatzelT (C. K . A m i . D i ' l y , . S c i . 17- (1959), 113, 285) and Wituick
and Williams [ Q u a r t . J . M c c h . A p p ! . M c . i h . XXIV (1971). 263; I n t . J . M e c . ' i . S r i .
If, (197-1). 209) have been incorporated to develop a procedure for the automatic
computation of the eigen-values and the eigenfunctions of one-dimensional linear Sturm-
Liouville boundary value eigenproblems, bolh singular and nonsingular. The continuous
coefficients of a regular Slurm-Liouville problem have been approximated by a finite number
of step functions. In each step the resulting boundary value problem has been integrated
exactly and the solutions have then been matched to construct, the continuously
differentiable solution of the original problem and the corresponding eigencondition. Step
sizes have been chosen automatically so that the local error has been held in a
predetermined interval. Representative test examples have been computed lo illustrate the
accuracy, reliability, and efficiency or the algorithm proposed.

I. INTRODUCTION

The solution of various problems in the field of mathematical physics and engineering
is closely related to the solution closely corresponding one-dimensional linear Sturm-
Liouville boundary value cigenproblem. As particular examples, one could imagine the
one-dimensional Sehrodinger equation with space-dcpcndent potential function and
numerous problems of physico-chemical kinetics [5].
A number of methods have been developed lo solve one-dimensional Sturm- Liouville
problems, among these finite difference methods [6,7 ], imbedding techniques |8], and
shooting type procedures |9, 10] based on integrators for systems of ordinary
differential equations. Regardless of the particular method utilized, one always has
difficulties with the computation of the higher-order eigenvalues and eigenfunctions due
to their rapid oscillation character. This results in the need for finer interval or
integration resolution, thus making the computer time involved unreasonable.
Recently Hargrave (11) and Bailey [12,13] proposed lo solve linear Sturm-Liouville
eigenproblems utilising a Pruffer Transformation of the original problem. This turned
out to be helpful because the approach allowed one lo compute the eigenvalues solving
only one ordinary differential equation numerically. Moreover, in the corresponding
initial value problem the order of the eigenvalue was present explicitly, thus making the
computational procedure safe f r o m e i g e n v a l u e s .

One othcr feature of the method discussed is that the general purpose integrators are
safer when applied to the phase function instead of being applied directly on the
eigenfunction [12]. After analyzing in detail the integration methods available to solve
numerically Sturm-Liouville eigenproblems, Bailey developed a general purpose solver
based on the Prtiffer transformation that was tested on a number of examples [J2, J3].
 In [1,2] Datzeff developed a method enabling one to obtain principally an approx-
imation of the whole infinite spectrum of the one-dimensional linear space-dependent
Schrodinger equation. A stepwise approximation of the external potential function was
applied, thus replacing the original problem with a sequence of finite sets of Sturm-
Liouville problems with constant coefficients defined on subintervals of the initial
interval of integration. He also proved that the spectra of the approximate problems
tend to the spectrum of the original problem as the length of the largest subinterval
tends to zero. AnX-^Qll.gpPrO-^.^^tion~af^tTc^n^t_na_j_ pro_bjem_can be integrated
_cxactix.jn_t_erms of circular and hyperboTIcJmciiOils. The solutions defined on
consequent subintervals can then be (Tnatche^at the corresponding interfaces to
ensure continuous differentiability of the solution of the approximate problem and to
derive the eigencondition of the problem. Such an approach guarantees roughly the
same accuracy of the ground and the higher-order eigenvalues and eigenfunctions.
These papers were published at the beginning of the computer era and thus no
numerical examples were presented to illustrate the computational advantages of this
method.
[ The same approach was later developed by Canosa and Olivcira [I4j. They ) illustrated
its efficiency by many numerical examples. The method was tested again | by Canosa
(15] to show its applicability for some practically asymptotic problems. A / trial and
error procedure was used to locate the eigenvalues, based on the computation of
dct[/v(A)] for randomly chosen values of X and simply observing where the
determinant changed sign. Such a procedure, however, is generally not safe ^J"rom
missing eigenvalues in the course of the computation.
Wittrick and Williams |3, 4] developed an extremely efficient algorithm (in terms of
computer time) for ihe safe and automatic computation of the natural frequencies and
buckling loads of linear clastic skeletal structures. It permits one to estimate exactly
how many natural frequencies lie below any fixed value of the frequency parameter
without calculating them explicitly, thus being able to obtain upper and lower bounds
for any required eigcnfrcqncncy to any desired accuracy.
Purpose of the- present paper is lo combine ihe approaches of Datzeff and of
Wituick and Williams to develop a computational procedure for the solution of one-
dimensional linear Slurm-Liouville boundary value eigenproblems. The applicability
of the algorithm will be demonstrated by recalculating several problems
32 previously published.
4
 COMPUTING STURM-I.IOUVIJ.U- I'UOHl.liMS 325

2. STATEMENT ANO ANALYTICAL ASPECTS or THE PROBLEM

present discussion concerns the one-dimensional linear Sturm-Liouville . . _~

_5ry value eigenproblem

( d / d x ) ( p { x M c l x ) T ( x , A )) f- ( A r ( x ) - q ( x ) ) Y'(.v, A ) - 0, a < x <

b , ( l a ) a0 V ' ( x , A ) — P 0 p ( x ) ( c l / d x ) V'(.r, /J) = 0, .r «,

(lb) a „ V ( x , A ) + p m p ( x ) { d / d x ) V ' ( x , A ) = 0, x = b y (lc)

where the constants a0, /?„, arid (Jn arc real and independent of the parameter /<.
A linear boundary value problem described by equations such as (la)-(lc) is said to be
regular [11-13] when:

/ (a) [n, />] is finite, J : (b) p ( x ) , r(A')' and if(jc) are continuous in [a, />], / (c) />(x) is
continuously differentiable in [a, /;], and [ (d) p ( x ) and ;(.v) are positive in [a, £}.
If one or more of conditions (a)-(d) are violated, the problem is said to be singular. Let
the coefficients p { x ) , r ( . r), and q ( x ) be approximate-.! by stepwise functions as

p(x) - py , r(x~f— rK , q ( v) - (7,

for .r^_ , < .t < .t^. k = 1, 2,..., n, x0 = a, x„ — b . This choice of the approximation of
the coefficients is convenient in utilizing Ihe standard mathematical functions present
in practically all kinds of digital computers [16]. In this way even a regular problem
like Eqs. (la)-(lc) is replaced by a singular one. The values of the cc-nsianls pk,
> ' • . , and q k , k — 1,2,..., H, can be choscn in various manners; they cart be taken
as the values of the coefficients p { x ) , r(.v), and q ( x ) at the midpoints of (he
corresponding subintcrvals l k , l k = x k — or thc-y can be c&limrlcrj as the
(eventually weighted)
average of the values of the coefficients at the cndpoinls of the subinterval as in [ M,
15]. Here the division of the original interval into subintcrvah is taken Tor granted,
while in a later section a procedure will lie described to i.olve this problem.
In any of the subintcrvals l k ~ x k —k = 1, 2,.... Lq. (a) can be replaced by an equation
of the kind

{<nyk{x,A)/tb:l) + wlvk{xfX)=--Q, xf, < x < x A , (3a)

where

< o l = { * r k - q k ) / p k , k — 1, 2,..., n. (3b)

It is seen (hat the coefficients in F.q. (3a) satisfy (he regularity requirements in the
interior of the corresponding subinterval except possibly at some endpoint, a case lo
be discussed in what follows.
 M1KHAIL0V AND VULCHANOV

COMPUTING S rURM—LIOUVILLE PRODLEMS 5

If one imposes ihe restriction that A ) has a continuous first derivative in the interior of
the original interval of integration, then one has to (fnatcH)the eigenfunction and its
first derivative at the endpoints of the corresponding suHntcrvals, thus

V k ( x % X ) = x = * k * k = 2,..., n — 1, (3c) p k { d V f k { x . X ) / d x ) =

p k A l { d i f f k J k l { x t X ) / d x ) , x = x k , *= 1,2...............................................n^-.j, (3d)

while Eqs. (lb) and (lc) take the form

«o V l ( x , \ ) - p 0 p l { ( l \ t / l ( x , \ ) / d . x ) = 0 , Jf = x0, (3e)

« „ A) + P „ p „ ( d v n ( x , X ) / d x ) = 0, x = x m . (3f)

Thus the original problem, Eqs. (la)-(lc) has been replaced by the approximation, Eqs.
(3a)-(3f). The computed eigenvalues A of the latter problem will be considered in what
follows as approximations to the eigenvalues X of problem (la)-(lc). Moreover, the finer
the resolution of the original interval of integration, the better an approximation of the
eigenvalues of the original problem (Eqs. (la)-(lc)) will be obtained, as was shown by
our numerical experiments and in |14, 15|.
The solution of Eq. (3a) for c o l > 0 is __________________________.________________
—
"17
, , sin(a),.(jrt - A)) , , %
sin(o>T(JF — A*.,)) A

where xk_, < x < xk, o > J k j n y j = 1, 2,.... ir c o j : = 0, then Eq. (4a) takes the form V* ( a%

A ) = V* ( • * * _ » » X X ( x k - x ) / l k ) -I- y k { x k , X ) ( ( x - x k _ ,)//<) (4b)

If (o2k < 0, Eq. (4 a) takes the form ' . T . __ [ ^

> /^f
A) = ,. A) y " + V k ( x k , A) sinh(<M ■ (4c)

C O * = (abs(w;))''-\ k = 1,2,..., n . (4d)

In Eqs. (4a)-(4c) the integration constants were expressed by means of the values of
the eigenfunction at the endpoints of the corresponding subintcrvals.
To satisfy the external and intcrfacc boundary conditions of Eqs. (3c)-(3f), one has

((«,//?,) + • • '• (5a)

-ZJ, W. ,(/•)■ + A k
. ,) YkiA) - B k4l
vK, ,(;.) = 0. k = 1 , 2 i t - "I. (5b)

-^V/„_I(A)+(^n + (r/„/Ov/fl(A) = 0, (5c)

326
 where

(5d)

and

= Pk">Jsin(a,\/J, A k = B k cos ( c o k l k ) . <i>* > 0, (6a

B )
>

Bk = PJh* Ak col = (6b

)
B k = inh(cu*/A), A k = B k cosh(a>*/t). u^ <0. (6c
)

and &
= 1,
2,..., n.
Equati
ons
(5a)-
(7a)
(5c) .
form a
linear
syste
m of
(// + I)
homog
eneou
s
equati
ons
for the
deter
minati
on of A = A _ , + Ak> k = 1 , 2 , 3 -
k k

& 1 An=An + {aJPn\

—
and 0.

1»

2,...,
n. In
matrix
notati
COMPUTING S rURM—LIOUVILLE PRODLEMS 7

on it
has
the
form

|*(A)|{v,(A)}=0,

where

0 0- 0 0 0
Ax -B2 0- 0 0 0
0 -Bj A 2
0 0 0

0 •••
-Bk * •• 0

.(7b)

0 0

0
... 0 0 A
„

(7c) (7d) (7e)

{y

(A

)}

^(

A)

^(

A)

,..

.,
 v

^(

A)

(7

is the

transp

ose of

{^(A)}

. In

order

that a

nonzer

solutio

n for |

y/(A)}

exists,

one

has
dct([AT(A)j = 0.

(8)

The
infinit
e
numbe
r of
real
COMPUTING S rURM—LIOUVILLE PRODLEMS 9

roots
of
transc
enden
tal
equati
on (8)
arc
the
eigenv
alues
of the
appro
ximate
proble
m of
Eqs.
(3a>-
(3f).
Their
compu
tation
is the
core of
the
proced
ure to
be
descri
bed in
the
sectio
ns to
follow.

/
 M
IKHA1LOV
AND
VULCHANO
V

3.
PROC
EDUR
E FOR
THE
COMP
UTATI
ON
OF
THE
EIGE
NVAL
UES

Witlri
ck
and
Willia
ms (3,
4]
devel
oped
an
efficie
nt
proce
dure
to
comp
ute
the
eigenf
reque
ncies
and
the
buckli
ng
loads
of
clastic
linear
skelet
al
COMPUTING S rURM—LIOUVILLE PRODLEMS 11

struct
ures.
Since
Eq.
(8) is
prese
nted
in the
same
form
as
that
utilize
d by
Wittri
ck
and
Willia
ms, it
is not
difficu
lt to
adapt
their
proce
dure
to the
proble
m
consid
ered
in the
prese
nt
paper.
Willia
ms
and
Wittri
ck
have
shown
that
the
numb
er of
positi
ve
eigen
values
A'(^)
lying
betwe
en
zero
and
some
prescr
ibed
positi
ve
value
o ft—
X is

(9)
7
where
N0(l)
is the
numb
er of
positi
ve
eigen
values
not
excee
ding
when
all
comp
onent
s of
the
vector
COMPUTING S rURM—LIOUVILLE PRODLEMS 13

corres
pondi
ng
to }
*'(/[)]
arc
zero
an£
5(|
A'(2)])
denote
s the
"sign
count"
of
(A"(7.)|
.
To
find
Nn(X)
one
takes
into
accou
nt the
fact
that
since
all
comp
onent
s of
the
vector

are
zero,
the
syste
m of
equati
ons
(3a)
degen
erates
into a
decou
pled
set
of
equati
ons

I-

colvk(

xt l) =

0, x k - \

< x <

xk* k

= 1,

2,...,

n,

(10a)

subjec

t to

the

bound
 COMPUTING S rURM—LIOUVILLE PRODLEMS 15

ary

condit

ions

The
eigenc
onditio
n of
proble
m
(10a)-
(10c) is

s \ n ( ( d k l k ) = 0.

(Ha)

It has
an
explicit
solutio
n only
if < t > l
> 0,
namely
,

d ) k l k = j n , j = . 1,2

(lib)

Thus,
Ar0(A)
is
equal
to the
total
numbe
r of
eigenv
alues
of the
kind
(Eq. (1
lb))

Noa)= V int /Mi,

 ' (12)
t—IV 71 // 1

where
the
value
or the
functi
on
int(-)
is the
larges
t
intege
r not
excee
ding
the
value
of the
argum
e
n
t

328

of tlie
functi
on.'
< The
"sign
count
" is
shown
(3| to
be
equal
10 the
numb

(10b)

(10c)
er of
COMPUTING S rURM—LIOUVILLE PRODLEMS 17

negati
ve
eleme
nts
along
the
main
diago
nal of
the
matri
x
|/v*(/)l
-
which
is the
triang
ulated
fo;m
of the
matri
x
(A'(2)|
, or
equiv
alent!
)', the
numb
er of
negati
ve
eleme
nts in
(he
seque
nce
D \t°r

D
n/t
>n-n
wher
e Dky
k =
1,2
den
ote
s
the
Ath
orde
r
p
r
i
n
c (13
i )
p
a
l

m
i
n
o
r

o
f

t
h
e

m
a
t
r
i
x

[
K
(
7
.
)
\
COMPUTING S rURM—LIOUVILLE PRODLEMS 19
,
D
0
=

l
t

£
>
,
=
,
7
,
.

I
n

v
i
e
w

o
f

t
h
e

t
r
i
d
i
a
g
o
n
a
l

s
t
r
u
c
t
u
r
e

o
f

t
h
i
s

m
a
t
r
i
x
,

o
n
e

h
a
s

"

—
*

D
* =
D
k-,
(>4*
+
Ak_,)
~
D
^ I
k
= 2,
3,...,
n .
J

T
h
e

c
a
COMPUTING S rURM—LIOUVILLE PRODLEMS 21

s
e

o
f

D
i
r
i
c
h
l
e
t

b
o
u
n
d
a
r
y

c
o
n
d
i
t
i
o
n
s

a
t

=
x
0

o
r

a
t

x
n

o
r

a
t

b
o
t
h

e
j
i
d
p
o
m
t
s

c
o
r
r
e
s
p
COMPUTING S rURM—LIOUVILLE PRODLEMS 23

o
n
d
s

t
o

/
J
0

o
r
/
a
n
d

/
?
„

0
,

t
h
u
s

r
e
s
u
l
t
i
n
g

i
n

/
T
,
"
-
c
o

o
r
/
a
n
d

c
o
.

I
n

a
n
y

o
f

t
h
e
s
e

e
a
COMPUTING S rURM—LIOUVILLE PRODLEMS 25

s
e

o
n
e

s
i
m
p
l
y

n
e
g
l
c
c
t
s

t
h
e

c
o
r
r
e
s
p
o
n
d
i
n
g

r
o
v
T

a
n
d

c
o
l
u
m
n

i
s
j
h
c
y

d
o

n
o
t

i
n
f
l
u
q
a
c
e

t
h
e
 COMPUTING S rURM—LIOUVILLE PRODLEMS 27

e
l
i
m
i
n
a
t
i
o
n

p
r
o
c
e
s
s

d
e
s
c
r
i
b
e
d

b
y

E
q
.

(
1
3
)
a
n
d

i
m
p
I
y
/
^
&
r
X
)

o
r
/
a
n
d

^
f
t
?
!
)

0
.
A
s

b
o
t
h
COMPUTING S rURM—LIOUVILLE PRODLEMS 29

t
e
r
m
s

o
n

t
h
e

r
i
g
h
t
-
h
a
n
d

s
i
d
e

o
f

E
q
.

(
9
)

c
a
n

b
e

c
o
m
p
u
t
e
d
,

o
n
e

h
a
s

t
h
e

f
o
l
l
o
w
i
n
g

a
l
g
o
r
i
t
h
COMPUTING S rURM—LIOUVILLE PRODLEMS 31

t
o

c
a
l
c
u
l
a
t
e

a
n
y

p
o
s
i
t
i
v
e

e
i
g
e
n
v
a
l
u
e

o
f
t
h
e

a
p
p
r
o
x
i
m
a
t
e

p
r
o
b
l
e
m
of
Eqs.
(3a)-
(30:
S
t
e
p

I
.

P
r
e
s
c
r
i
b
e
v
COMPUTING S rURM—LIOUVILLE PRODLEMS 33

a
l
u
e
s

o
f

/
,

t
h
e

o
r
d
e
r

o
f

t
h
e

e
i
g
e
n
v
a
l
u
e

X
,
d
e
s
i
r
e
d

a
n
d

e
f

t
h
e

(
a
b
s
o
l
u
t
e
)

a
c
c
u
r
a
c
y

i
n
X
COMPUTING S rURM—LIOUVILLE PRODLEMS 35

,
;

s
e
t

l
o
w
e
r

b
o
u
n
d

f
o
r

X
,
,

A
,

0
,

u
p
p
e
r
,
b
o
u
n
d

f
o
r

X
,
,

A
„

0
,

c
u
r
r
e
n
t

v
a
l
u
e

o
f

t
h
e

A
-
COMPUTING S rURM—LIOUVILLE PRODLEMS 37

p
a
r
a
m
e
t
e
r
,

0
;

i
n

c
a
s
e

t
h
e
r
e

i
s

a
d
d
i
t
i
o
n
a
l

i
n
f
o
r
m
a
t
i
o
n

c
o
n
c
e
r
n
i
n
g

X
,
y

s
e
t

m
o
r
e

a
p
p
r
o
COMPUTING S rURM—LIOUVILLE PRODLEMS 39

p
r
i
a
t
e

v
a
l
u
e
s

f
o
r

t
h
e

p
a
r
a
m
e
t
e
r
s

l
i
s
t
e
d

t
o
s
a
v
e

c
o
m
p
u
t
e
r

t
i
m
e
.

Step

2.

Defi

ne

the

incre

ment

S in

X, 5

* 0;

set 1

= I -f

A„ =

com

pute
 COMPUTING S rURM—LIOUVILLE PRODLEMS 41

(^g^

Step

3. If

N {1)

> /,

go to

Step

5,

else

go to

Step

4.

Step

4. '

Para

mete

r I is
a

lowe

boun

d for

A,;

set

A,,=

2, go
to

Step

2.
 COMPUTING S rURM—LIOUVILLE PRODLEMS 43

S T E P 5. PARAMETER 1 IS AN UPPER BOUND FOR A,; SET A == I. . S T E P 6. COMPUTE A X =

U
ABS(A„ - A,); IF

<JA < C(, GO LO STEP 10, ELSE GO TO STEP 7. S T E P 1 . COMPUTE I = (A 4- A,)/2; A

COMPUTE(N(I)) S T E P 8. IF

> /, SET A = 1
U
AND GO TO STEP 6, ELSE GO TO STEP 9. S T E P 9. SET A, = I; GO TO STEP 6. S T E P 10. X , =

(A -F A,)/2; Y330
U
MIKMAIL.OV AND VULCHANOV

into Eq. (5a), one obtains

V/.iO*,) = («0 + 1 Pi))/B\- (Mb)

Equation (5b) can be rewritten as

VM + IO* / ) « { { A k + A k 4 ,) - B k v/4_ ,(A(.))//;k4 ,, k — I , 2................(I4c)

The last recurrence relation, Eq. (14c), enables one to compute all the components of the
vector {y(A,)| of Eq. (7f), which arc the values of the eigenfunction v,(a\ A,)
corresponding to the eigenvalue Af at the endpoints of the subintervals lkt k = 1,2,..., n .
From Eqs. (5) it can be seen that in case the exact eigenvalues of the approximate
problem (Eqs. (3)) arc known and if the cocfficicnts A k and B k , k — 1, 2,..., of Eqs. (6)
can be computed exactly, then Eq. (5c) should be satisfied identically. But to the extent
that the computation cannot be performed exactly and the eigenvalues can be computed
approximately in the sense of the algorithm described in Section 3, the left-hand side of
Eq. (5c) cannot become zero. Its value can be considered as an estimate for the global
error in the computation of the eigenfunction. For the special case of a Dirichlct
boundary, when Eq. (5c) is not considered, the accuracy estimate can be obtained by Eq.
(5b). written for k = n — 1 and bearing in mind that that = 0. '
The norm r of the eigenfunction y/((.r, A,) can be defined as

= t V'L{ x . X , ) d x , (15)
A r-I J.r,

and in view of Eqs. (4) and (6) one has

V
R v/J.Xv,;,.)^- = + vl){BlUPk-Ak)
J ~~

+ (16) where y / k — V i t { x k , A,) and the constant ( v ' k ,

A k , and D k are calculated for A = some function \ j / , ( x , X , ) that oscillates at least as

rapidly as the cigenfunction wanted, where X , is some upper bound for the eigenvalue

Af, o n the basis of two different divisions of the interval [ a , b j , the second of therri j
times finer than the first, and the corresponding values computed are equal in the sense

of some predetermined local error level, then the rougher of these two divisions of \ a , b \

could be used to compute the desired eigenvalue and the corresponding cigenfunction

with approximately the same local accuracy.

Having in mind the procedure described in the previous section for the computation of
the eigenfunction, the following principal algorithm could be used to find the division of
[a, bJ:

S t e p 1. Prescribe the following quantities: upper and lower bounds z m a x and £mln,
respectively, for the local error of the function \j/({x, A,) for an arbitrary subin- terval,
where / is the order of the eigenvalue A y and the corresponding eigenfunction y / , ( x , X , )
to be computed; A, is an upper bound of /., and y/,(x, Af) oscillates at least as rapidly as
A,); /„ is the initial guess for the first subintcrval /, — x , — .v0; is the minimal admissible
length of an arbitrary subinterval of the division of [a, b ) to be computed (the maximum
admissible length for the subintcrval is 6) itself): fix an integer j by means of which a
/times finer division of [«,/>) could be generated.

S t e p 2. Set Jt = 1 ,([)= i 0 , .v = .
S t e p 3. Set l j = /// y l S

S t e p 4. Compute + A Xj} directly from Eqs. (14).

S t e p 5. Compute utilizing successively the values + /,,Af),-
V?,(.x + 21 Jy X v , ( x + if-0b>Xt)> l -
and Ec s

S t e p 6. If v \ \ x + /, Af) and \ j / , ( x + /, X , ) are equal in the sense of the local error

J

prescribed, go to Step 9; else go to Step 7.

S t e p 7. If / < /mln, go to Step 12; else go to Step 8.
S t e p 8. Decrease (set /=/,) or increase (set I = j ! ) the guess Tor the subintcrval
considered and go to Step 3.
S t e p 9. Accuracy requirements arc satisfied; set l h = /, x = x + / , k = k + I.
S t e p 10. If x + / = x „ y go to Step 3; else go to Step 11.
S t e p 1 1 . This is the correct exit of the algorithm; the last subintcrval has to be adjusted
so that /„ =x„—x in the sense of some error tolerance. At this place one has the number
of the subintervals for the eigenvalue considered n = k + \ and the division of |o, b \ into
n subintervals, thus being able to utilize the algorithms for the cdirtfnitation of A, and
A,) described earlier.

S t e p 12. This is the error exit, indicating in general that the accuracy requirements arc
not appropriately chosen.
 COMPUTING S rURM—LIOUVILLE PRODLEMS 45
332
MJKLLA1LOV AND VULCHANOV

The estimate A,+ , for the next eigenvalue A( + , can be obtained utilizing the division by
means of which A, and v/,(.v, X , ) have been computed and A, can be obtained starting
from some initial (e.g., uniform) division of [«,/>).
A simpler approach can be used by finding an estimate A,„ for the eigenvalue X m , where
m is the order of the highest eigenvalue that has to be computed.
A check on the global error of the y function computed can be performed by making use
of Eq. (5c) or Eq. (5b) for A. = n — 1 as was described in the previous section; thus the
inequality should be satisfied

abs(-£„ _,, Af) -f- A „ i j i , { x „ , A,)) <

(17)

where for an upper bound for the global error one can assume

cr,obai ~ cn,« * ' (1)

In case Eq. (5b) has to be used instead of Eq. (5c), one sets k = n — I and obtains the
corresponding expression for the left-hand side of Eq. (17).
The choice of step sizes in our method; as seen from the above, is not directly related to
the eigenvalue under consideration, since the analytical function representing the
eigenfunction may have any number of oscillations in a single subiti- terval.

1( i i / i i x )(/;(.V)(«/Y't.V, M ) / d x ) \ (,V//(.V) - q * { x ) ) V(.Y. A/) - 0, (20b)

where

q*(x) = Cr(x) + q(x) (20c)

and thc constant C is chosen so that q * ( A) > 0, O<A<6 Thc shifted problem or Eq. (20b)
can readily be solved utilizing the algorithm described to find a sequence of
eigenvalues

0<M, <m2 < — (20d)

and through Eq. (20a) to recover the original spectrum A, /== 1, 2,....
Ince has also shown that if r ( x ) changes sign in \ a , b ) while q ( x ) > 0 there, then one
has Tor the spectrum

-oo <• - < X ~ <A"_, < ••• <;.," < 0 < A<A,+ <-•• <+co. (21a)

After computing the positive part of thc spectrum as was already described, one might
set If

X = —v, (2 lb)

which is equivalent to a reversal of ..he sign of r(A*); in this way one is able to. compute
the negative part of thc spectrum to such a problem.
Thc next case to be considered h the presence of interface resistance h k k 4 t 118, 19]
between two subintervals lk anJ lk+l of thc division of \ayb|. Now, boundary condition (3d)
takes thc form

p k { d V k { x k y X ) l d x ) = h k m k + ,(y* + % ( x k , X ) - y , k { x k . X)). (22a)

Assume that one could insert a fictitious subinterval /lJt4., between the subintervals lk
and /Af,, such that a)^ A+l = 0. Consequently, r k k + l = 0 and q k . k + t = 0 . Taking arbitrary
values for pk^k+] and /A>*+, such that

thc procedure described above can again be applied to solve the corresponding Sturm-
Liouville eigenproblem.
The last class of problems to be considered in this section is that of the singular
problems, characterized by infinite endpoint singularities. The theory developed in 120,
Part 1, Chap. V, 5.1 J considers the complex Sturm-Liouville problem and slates that
(only case b-* oo will be considered):
(i) if b ~ * co and limx „ h q ( x ) oo, the spectrum is numerable: (ii) if b -» cc and
limx _ h q ( x ) — q x t . where r/„ is finite, then the spectrum has cither a finite discrete
component and a continuous component or it is continuous:
 abs(<7(x)) " 1/2
dx

has a finite limit (a is assumed to be finite), one has the spectrum of (i); otherwise the
spectrum is continuous.
Case (iii) will not be considered in what follows inasmuch as the authors have no
evidence of the applicability of this ease to real Sturm-Liouville problems.
To discuss eases (i) and (ii), first rccall Eq. (6c). The constants A k and B k can be
rewritten as

<4* = +cxp(-2o;*/<))(! -cxp(2</4)r\ (23a)

B k = 2 A k c x p { - o ) * l k ) ( l -f cxp(-2</,))-\ (23b)

and one recalls that co\ < 0. Usually, when one solves numerically some initial or
boundary value problem defined on an infinite interval, he cither explicitly or implicitly
assumes that the solution changes significantly over some reasonable long finite
subintcrval, while over the remaining infinite subintcrval the solution is assumed to have
some asymptotic behavior. This implies that Tor the singular problems under
consideration one has to separate the infinite interval |o, 00) into a finite subinterval [a,
c) and an infinite subinterval (c, co). Over the former of these one can apply the
procedure developed in the previous sections, while in this section the infinite subin-
terval (r, co) will be paid some particular attention. It is seen from Eqs. (23) that c o
implies B t . -» 0 and A k -*■ p k t o % > 0.
334 MLKHAILOV AND VUL.CIIAN'OV

(iii) if b -» oo and lim^q ( x ) — -oo while the integral

.00
.00
For case (i), <?(.v)-» 00
implies A k
-» co and B k ~ * 0 independently of I k. Consequently, the finite subintcrval
[a, cj can be fixed from the condition that B k behaves as zero in the sense of the
machine word length. Having in mind the fact thai A k ~ * 00, the last two rows and
columns of the matrix |/C(A)], which correspond to the infinite interval (c, 00), are
eliminated from consideration, similarly lo the ones used for Dirichlct boundary
conditions at a finite endpoint of a regular problem. Consequently, for case (i) one can
apply the algorithm already discussed. Our conclusion corresponds to Bailey e l o l . |2]
in the sense that for a limit point singularity the solution approaches the singular point
regardless of the boundary condition imposed there.
. Case (ii) can be treated in a similar manner. Now B k -»0, but A k remains finite. Thus,
the finite subinterval [o, cj has lo be determined from the condition that A k behaves as a
constant as x - * b and the corresponding subintervals remain finite. After this has been
done, one fixes the division of 00) as composed of a finite number of finite subintervals
plus the infinite subintcrval (c, co). This means that if the division of |<7, c] comprises n
subinlervals, the division of \ a , 00) will comprise n + 1 subintervals. the last of which
will be the infinite one. Consequently. will have the
form
.(
(24)
having in mind that for thc case under consideration p { x ) = I. This formulation
corresponds to [12] in thc sense lhat when a limit circle singularity is present, the
solution advances towards the singular endpoint following a trajectory prescribed by
the particular boundary condition imposed at thc singutar endpoint. Moreover, as Eqs.
(23) arc valid for c o 2 k < 0, it is clcar that if q 0 < 0 (depending on r(.r). which is
supposed bounded and nonnegativc), one may have a finite number of eigenvalues on
thc negative semiaxis, while otherwise thc whole spectrum will be continuous. <

* 7. TEST CALCULATIONS

The procedure for thc computation of the eigenvalues and the eigcnfunctions
developed in this paper was programmed in APPLESOFT BASIC and a number of f test
examples were calculated on an APPLE II personal computer. The code was J tested on a
number of problems, the solutions of which had been previously published. Some
illustrative results are presented in Table I. The symbols appearing in column I of this
table are thc same as in thc original references.
First, regular problems with constant coefficients Subject to different boundary
conditions |2I] were solved. The eigenvalues of each particular problem were
calculated utilizing several different divisions of thc integration interval into subin-
tervals. Thc results for thc eigenvalues of a particular problem obtained through the
various divisions were fully coincidcnt—row 1 from Table I. Next, some discontinuous
coefficient problems on finite intervals were solved (22, 23 J, utilizing the. same
testing strategy—row 2 in Table I. Our results are in full agreement with the
corresponding references.
Wc resolved most of the problems from (14, 15]. No case of missing eigenvalues
was identified. This test also proved the applicability of our algorithm for practically
asymptotic problems. Some illustrative results arc given in row 3 of Table I.
 TABLE L
Problem
COMPUTING S rURM—LIOUVILLE PRODLEMS 49
Mikhailov \ 1 ] Bi„= I, 1 1.7070 13.4923 43.3572

Muhlholland and Cobbler |23| 277.433 1953.73 . 3404.21 '

. * • Canosa and dc Olivcira (14| q = -3119.20. -2959.43 -2801.85 '.

I600,/I= 104 . Neumann
boundary

Nunpc and GUI |241 ,Y= 1.73. 5.9S01 -0.00 304.453 — 997.89.7
>J=!00 80.S314 -263.279

MlKHAIl.OV AND VULC.'IANQV

The last test to be discusscd licrc is the recalculation of the first three eigenvalues
X f c and i — 1,2,3 from |24]—row 4 of Table 1. Again there is full agreement between
our results and the ones obtained in the original reference.
The test calculations for infinite endpoint singular Sturm-Liouville problems as
discussed in the previous section arc presently in preparation. 1 he tests performed to
verify the procedure for the computation of the cigcnfunctions and the choice of step
sizes gave more than cnccuri'.ging results. They will be discussed in a future
publication together with some algorithmic and programming details concerning the
code.

REFERENCES

_ 1. A. B. DATZtrrF, C. R . A c a d . H u l ^ . S c i 12 (2) (1955), 113. 4 2. A. B. DATtfiFF. C.

R . A c a d . H u l g . S c i . 12 (4) (1959), 285.
____* 3. W. H. WI n RICK AND F. JV. WILLIAMS. Q u a r t . J . M r c h . A p p l . M o t h . XXIV (3) (1971). 263.
4.W. M. WITTRICK AND !•*. W. WILLIAMS. I n t . J . M c c h . S c i . 16 (197-1). 7 0 9 .
5.N. N. TUNITSKH, V. A. KAMINSKII, AND S. F. TIMASIIEV, "Methods in Physico-Chcmical Kinctics." Chemistry,
Moscow. 1972. jin Russian)
6.D. G. TRUULER, J . C o m p u i . P h y s . 10 (1972). 123.
7.J. D. TALMAN, J. Compui. Phys. 37 (1950). 19.
8.E. WASSI-RSTROM. J . C o m p u t . P h y s . 9 (1972). 53,
9.D. M;C»HH, A. K. AGIIAWAI, AND J. L. TIF.T7.E. J . C o m p u i . P h y s . 15. (1974). 117.
10.P. N'Et SON, J . C o m p u t . P n y i . 37 (I9S0), 388.
11.B. A. Harr,ravc. J . C o m p u i . P h y s . 20 (1976). 381.
12.P. B. BAIL!:V. M. K. COYNES', AND L. F. SHAMI-INE. A C M T r a n s . M a t h . S o f t w a r e 4 (1978). 193.
13.P. B. BAHF.Y, SI.F.IGN: An fiigcnvaluc Eif.cnfunction Code for Sturm-Liouville Problems. SAND77 20-14,
1978-
14.J. C-ANOSA AND K. G. DE OLIVP.IRA. J . C o m p u t . P h y s . 5 (1970). 188.
15.J. CANOSA, J . C o m p u t . P h y s . 7 (1J/71). 255. . ..
16.R. G. GORDON, i n "Methods in Computational Physics" (AJlrr c t a ! . . Eds.).
Vol.'10'.~Acadcmic Press, New York. 1971.
17.F.. L. INCF.. "Ordinary Differential Equations." Kharkov. 1939.
18.E. J. Scorr L. II. TUNG, AND II. G. DRICKMRN. ./. C h e n i . P i irv. 19 (1951). 1075.
19.R- J. NUNGF. AND W. N. GII.I.. A / C h i l ./. 12 (1966). 279.
20.E. C. TITOIIMARSH. "Eigenfunction Expansions Associated with Sccond-Ordcr DifTcrcntial Equations."
 Moscow, 1960. |in Russian).
21.M. D. MIKHAILOV. "Transient Temperature Fields in Shells." Energy. Moscow. 1967. Jin Russian)
22.C. O. 1 (ORGAN AND S. NUMAT-NASSER. Z . A / I g c u - . M o t h . P h y s . 30 (1979). 77.
23.' G. 7\ MUIIMIOLLAND AND M. H. COBBLE. I n t . J . H e a t M a s s T r a n s f e r 15 (1872), 147.
L

33 r. j. Nunc.e and w. n. Gill, I n t . J . I l e a l , . M a s s T r a n s f e r 8 (1965).

6 873.4. COMPUTATION OF THE EIGENFUNCTIONS

Once the eigenvalue X , has been located, one might want to know some values
of the cigenfunction V,(-Y,A,) at (possibly) previously prescribed positions. The
computation of the cigenfunction can be carried out making use of Eqs. (5) and (6). '
The choice of ft, and p j d f d x ) satisfies
identically
Eq. (3e). Therefore, introducing .
5. CHOOSING STEP SIZES

The purpose of the procedure to be described is to construct the division of the interval
{<7, 6| into subintervals l k . k — 1, 2,..., //, in such a manner that the local error in the
values of the cigcnfunction computed be held at a prescribed level and be of comparable
magnitude for any of the subintervals necessary to approximate the original problem.
' •
Our strategy is based on the assumption that the cigcnfunction is quite sensitive in
rcspcct to the eigenvalue parameter A. Thus, if one is able to compute the values of
6. APPLICAWIUTY OF THC ALGORITHM
To discuss the applicability of this algorithm one has to recall some well-known
qualitative results from thc theory of the one-dimensional linear Sturm-Liouville
problems. Ince has shown [17, Sect. 10.61 ] that if p { x ) > 0, r { x ) > 0, and q(v) > 0 in thc
finite interval [<7, 6], and if these three functions together with thc coefficients a 0. fJ0, un,
and Pn do not depend on the eigenvalue parameter A, then the corresponding regular
eigenvalue problem has a numerable set of positive eigenvalues
0<A,<A2<-.. (19)
with +oo as thc only limit point of the sequence in Eq. (19). This is the case
discusscd in thc previous sections.
If 9(a) changes sign in [<7, b \ , the set of thc eigenvalues has the same character with
thc only exception that thc sequence in Eq. (19) is shifted towards the negative
scmiaxis and A, > min(R/(A)), a < A- < b . So let

' X=ti-C, A = M-C> (20a)

and substitute Eq. (20a) into Eq. (la) to obtain