Vous êtes sur la page 1sur 8

# Journal of Computational and Applied Mathematics 147 (2002) 445452

www.elsevier.com/locate/cam
A spline method for second-order singularly perturbed
boundary-value problems

Department of Applied Mathematics, Faculty of Engineering and Technology, Aligarh Muslim University,
Aligarh 202002, India
Abstract
We consider a dierence scheme based on cubic spline in compression for second-order singularly perturbed
boundary value problem of the form
c,

= (x),

## + q(x), + r(x), ,(a) = :

0
, ,(b) = :
1
.
The method is shown to have second- and fourth-order convergence depending on the choice of parameters z
1
and z
2
involved in the method. The method is tested on an example and the results found to be in agreement
MSC: 65L10
Keywords: Cubic spline in compression; Boundary-layers; Monotone matrix; Irreducible matrix
1. Introduction
We consider a second-order singularly perturbed boundary value problem
c,

= (x),

## + q(x), + r(x), (1)

,(a) = :
0
, ,(b) = :
1
, (2)
where , q, r are smooth, bounded, real functions : R R, q : R R, r : R R and c is a
parameter such that 0 c1. It is known that problem (1)(2) exhibits boundary layers at one or
both ends of the interval depending on the choice of the function .

Corresponding author.
E-mail addresses: amt01ta@amu.up.nic.in (T. Aziz), akhan@bharatmail.com (A. Khan).
PII: S0377- 0427(02)00479- X
446 T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452
The approximate solution of boundary-value problems with a small parameter aecting highest
derivative of the dierential equation is described. It is a well-known fact that the solution of sin-
gularly perturbed boundary-value problem exhibits a multiscale character. That is, there is a thin
layer where the solution varies rapidly, while away from the layer the solution behaves regularly
and varies slowly. This class of problems has recently gained importance in the literature for two
main reasons. Firstly, they occur frequently in many areas of science and engineering, for example,
combustion, chemical reactor theory, nuclear engineering, control theory, elasticity, uid mechanics
etc. A few notable examples are boundary-layer problems, WKB Theory, the modelling of steady and
unsteady viscous ow problems with large Reynolds number and convective heat transport problems
with large Peclet number. Secondly, the occurrence of sharp boundary-layers as c, the coecient
of highest derivative, approaches zero creates diculty for most standard numerical schemes.There
exist a variety of techniques for solving singularly perturbed boundary value problems [2,5,9,10,12,
1416]. The numerical solution of two point boundary-value problems using splines has been con-
sidered by many authors; see for example, [1,3,6,7,9,11,15] and references therein.
In the present paper, we have derived a uniformly convergent uniform mesh dierence scheme
using cubic spline in compression for the solution of (1). The main idea is to use the condition
of continuity as a discretization equation for (1). The advantage of our second- and fourth-order
methods is higher accuracy with the same computational eort. Kadalbajoo and Bawa  give a
second-order method which becomes a special case of our method. The use of cubic splines for the
solution of (regular) linear two point boundary-value problems was suggested by Bickley . Later,
Fyfe  discussed the application of deferred corrections to the method suggested by Bickley, by
considering the case of (regular) linear boundary-value problems. Our scheme for the corresponding
problem (i.e. c = 1, 0) reduces to the Bickley scheme.
However, it is well known since then that the cubic spline method of Bickley gives only O(h
2
)
convergent approximations. But cubic spline itself is a fourth-order process . Hence, it is natural
to look for a fourth-order spline solution. In comparison with the nite dierence methods, spline
solution has its own advantages. For example, once the solution has been computed, the information
required for spline interpolation between mesh points is available. This is particularly important
when the solution of the boundary-value problem is required at various locations in the interval
[a, b]. Analysis of the method shows that it has second-order convergence for arbitrary z
1
, z
2
such
that z
1
+ z
2
=
1
2
and fourth-order convergence for z
1
=
1
12
, z
2
=
5
12
. In Section 2, we have given
derivation of the method and in Section 3 convergence of the method has been discussed. Section 4
contains numerical illustrations and results are compared with the method of Kadalbajoo and Bawa
 to demonstrate the eciency of the method.
2. Derivation of the method
Let x
0
= a, x
N
= b, x
i
= a + ih, h = (b a)}N.
A function S(x, t) of class C
2
[a, b] which interpolates ,(x) at the mesh point x
i
depends on a
parameter t, reduces to cubic spline in [a, b] as t 0 is termed as parametric cubic spline function.
The Spline function S(x, t) = S(x) satisfying in [x
i
, x
i+1
], the dierential equation
S

(x) + tS(x) = [S

(x
i
) + tS(x
i
)]
(x
i+1
x)
h
+ [S

(x
i+1
) + tS(x
i+1
)]
(x x
i
)
h
, (3)
where S(x
i
) = ,
i
and t 0 is termed as cubic spline in compression.
T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452 447
Solving the linear second-order dierential equation (3) and determining the arbitrary constants
from the interpolatory conditions S(x
i+1
) = ,
i+1
, S(x
i
) = ,
i
we get, after writing z = h t
1}2
S(x) =
h
2
z
2
sin z

M
i+1
sin
z(x x
i
)
h
+ M
i
sin
z(x
i+1
x)
h

+
h
2
z
2

(x x
i
)
h

M
i+1
+
z
2
h
2
,
i+1

+
(x
i+1
x)
h

M
i
+
z
2
h
2
,
i

. (4)
Dierentiating Eq. (4) and letting x tend to x
i
, we obtain
S

(x
i
+) =
,
i+1
,
i
h
+
h
z
2

1
z
sin z

M
i+1
(1 z cot z)M
i

.
Considering the interval (x
i1
, x
i
) and proceeding similarly, we obtain
S

(x
i
) =
,
i
,
i1
h
+
h
z
2

(1 z cot z)M
i

1
z
sinz

M
i1

.
Equating the left- and right-hand derivatives at x
i
, we have
,
i
,
i1
h
+
h
z
2

(1 z cot z)M
i

1
z
sin z

M
i1

=
,
i+1
,
i
h
+
h
z
2

1
z
sin z

M
i+1
(1 z cot z)M
i

. (5)
This leads to the tridiagonal system
h
2
(z
1
M
i1
+ 2z
2
M
i
+ z
1
M
i+1
) = ,
i+1
2,
i
+ ,
i1
, (6)
where
z
1
=
1
z
2

z
sin z
1

, z
2
=
1
z
2
(1 z cot z), M
i
= S

(x
i
), i = 1(1)N 1.
The condition of continuity given by (6) ensures the continuity of the rst-order derivatives of
the spline S(x, t) at interior nodes.
Substituting cM
i
=(x
i
),

i
+q(x
i
),
i
+r(x
i
), in Eq. (6) and using the following approximations for
rst derivative of ,:
,

i

=
(,
i+1
,
i1
)
2h
, (7)
,

i+1

=
(3,
i+1
4,
i
+ ,
i1
)
2h
, (8)
,

i1

=
(,
i+1
+ 4,
i
3,
i1
)
2h
, (9)
448 T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452
we arrive at the following linear system which may be solved to get the approximations ,
1
, ,
2
, . . . ,
,
N1
of the solutions ,(x) at x
1
, x
2
, . . . , x
N1
:

c + z
1
h
2
q
i1

3
2
z
1
h
i1
z
2
h
i
+
z
1
2
h
i+1

,
i1
+ (2c + 2z
2
h
2
q
i
+ 2z
1
h
i1
2z
1
h
i+1
),
i
+

c + z
1
h
2
q
i+1

z
1
2
h
i1
+ z
2
h
i
+
3
2
z
1
h
i+1

,
i+1
= h
2
(z
1
r
i1
+ 2z
2
r
i
+ z
1
r
i+1
), i = 1(1)N 1 (10)
with ,(a) = :
0
, ,(b) = :
1
, where
i
= (x
i
), q
i
= q(x
i
), r
i
= r(x
i
), i = 0(1)N.
Remark 2.1. For z
1
=
1
6
, z
2
=
1
3
, our method reduces to the Kadalbajoo and Bawas method  for
uniform mesh.
Remark 2.2. For c = 1 (regular problem), uniform mesh and 0, our method reduces to the
well-known Bickley scheme  for the regular problem ,

## =q(x),+r(x) with ,(a) =:

0
, ,(b) =:
1
.
3. Convergence of the method
Putting the tridiagonal system (10) in matrixvector form,
AY + h
2
BR = C (11)
in which A = (a
i, )
) is tridiagonal and diagonally dominant matrix of order N 1, with
a
i, i+1
= coecient of ,
i+1
in (10), i = 1(1)N 2,
a
i, i
= coecient of ,
i
in (10), i = 1(1)N 1,
a
i, i1
= coecient of ,
i1
in (10), i = 2(1)N 1
and R = ( r
1
, r
2
, . . . , r
N1
)
T
, where r
i
= h
2
(z
1
r
i1
+ 2z
2
r
i
+ z
1
r
i+1
), i = 1(1)N 1 and
Y = (,
1
, ,
2
, . . . , ,
N1
)
T
.
The tridiagonal matrix B is given by
B =

2z
2
z
1
0
z
1
2z
2
z
1
0 z
1
2z
2
z
1
0

z
1
2z
2
z
1
0 z
1
2z
2

and C = (c
1
, 0, 0.........0, c
N1
)
T
,
T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452 449
where
c
1
=

c z
1
h
2
q
0
+

3
2

z
1
h
0
+ z
2
h
1

z
1
2

h
2

:
0
h
2
z
1
r
0
,
c
N1
=

c z
1
h
2
q
N
+

z
1
2

h
N2
z
2
h
N1

3
2

z
1
h
N

:
1
h
2
z
1
r
N
.
Also, we have
A

Y + h
2
BR = 1(h) + C, (12)
where

Y =(,(x
1
), ,(x
2
), . . . , ,(x
N1
))
T
denotes the exact solution vector, and 1(h)=(1
1
(h), 1
2
(h), . . . ,
1
N1
(h))
T
is the local truncation error vector, where
1
i
(h) = (1 + 12z
1
)
ch
4
12
,
(4)
(
i
) + (1 + 30z
1
)
ch
6
360
,
(6)
(
i
), x
i1

i
x
i+1
(13)
for any choice of z
1
and z
2
whose sum is
1
2
, except z
1
=
1
12
, z
2
=
5
12
. For the choice z
1
=
1
12
, z
2
=
5
12
,
1
i
(h) =

ch
6
240

,
(6)
(
i
), x
i1

i
x
i+1
. (14)
From (11) and (12), we get
A(

Y Y) = 1(h),
AE = 1(h), (15)
where E =

Y Y = (e
1
, e
2
, . . . , e
N1
)
T
.
Clearly,
S
1
=
N1

)=1
a
1, )
= c + 2z
2
h
2
q
1
+

3
2

z
1
h
0

z
1
2

h
2
+ z
1
h
2
q
2
+ z
2
h
1
,
S
i
=
N1

)=1
a
i, )
= h
2
(z
1
q
i1
+ 2z
2
q
i
+ z
1
q
i+1
) = h
2
B
i
, i = 2(1)N 2,
S
N1
=
N1

)=1
a
N1, )
= c + z
1
h
2
q
N2
+

z
1
2

h
N2
z
2
h
N1

3
2

z
1
h
N
+ 2z
2
h
2
q
N1
.
In the following cases:
(i) (x) 0, q(x) 0,
(ii) (x) 0, q(x) 0,
(iii) (x) 0, q(x) 0,
(iv) (x) 0, q(x) 0,
(v) (x) 0, q(x) 0.
450 T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452
we can choose h suciently small so that the matrix A is irreducible and monotone . It follows
that A
1
exists and its elements are nonnegative.
Hence from Eq. (15), we have
E = A
1
1(h). (16)
Also, from the theory of matrices we have
N1

i=1
a
k, i
S
i
= 1, k = 1(1)N 1, (17)
where a
k, i
is the (k, i) element of the matrix A
1
.
Therefore,
N1

i=1
a
k, i
6
1
min
16i6N1
S
i
=
1
h
2
B
i
0
6
1
h
2
|B
i
0
|
(18)
for some i
0
between 1 and N 1.
From (13), (16) and (17) we have
e
)
=
N1

i=1
a
), i
1
i
(h), ) = 1(1)N 1 (19)
and therefore
|e
)
| 6
Kh
2
|B
i
0
|
, ) = 1(1)N 1, (20)
where K is constant independent of h.
Therefore
E = O(h
2
). (21)
However, for the choice of parameters z
1
=
1
12
, z
2
=
5
12
,
|e
)
| 6
Kh
4
|B
i
0
|
, ) = 1(1)N 1 (22)
Therefore E = O(h
4
).
We summarise the above results in the following theorem:
Theorem. The method given by Eq. (10) for solving the boundary value problem (1)(2) for
q(x) 0 and suciently small h gives a second-order convergent solution for arbitrary z
1
, z
2
with
z
1
+ z
2
=
1
2
and a fourth-order convergent solution for z
1
=
1
12
, z
2
=
5
12
.
4. Numerical illustrations and discussion
We have implemented our method on one example which supports the theoretical analysis for
second- and fourth-order convergence. The maximum error at the nodal points, max |,(x
i
) ,
i
| is
tabulated in Tables 1 and 2 for dierent values of the parameters c, N, z
1
and z
2
.
T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452 451
Table 1
Maximum absolute error for second-order method
z
1
and z
2
c = 10
5
c = 10
8
c = 10
10
N = 100 N = 200 N = 250
1}18, 4}9 1.44463E 03 6.22342E 02 6.27380E 02
1}14, 3}7 1.52823E 02 8.33647E 02 8.39115E 02
1}24, 11}24 1.00616E 02 4.50702E 02 4.55413E 02
1}30, 14}30 1.67078E 02 3.52995E 02 3.57527E 02
1}6, 1}3 1.1971E 01 2.6683E 01 2.6793E 01
Table 2
Maximum absolute error for fourth-order method
z
1
= 1}12, z
2
= 5}12 c = 10
3
c = 10
4
N = 50 1.1727E 04 8.3186E 03
N = 100 7.9662E 06 7.4138E 04
N = 200 8.1975E 07 4.7790E 05
Example 4.1 (Doolan et al. ).
c,

= , + cos
2
(x) + 2c
2
cos(2x),
,(0) = ,(1) = 0.
The exact solution is given by
,(x) = [exp((1 x)}

c) + exp(x}

c)]}[1 + exp(1}

c)] cos
2
(x),
since 0 and q 1 0, the boundary layer exists at both ends.
The numerical calculations were carried out on a ALPHA-VMS 3000}4 AS Computer at the
Computer Centre, Aligarh Muslim University, Aligarh, using double precision arithmetic in order to
reduce the round-o errors to a minimum.
Acknowledgements
The authors would like to thank the referee for the valuable suggestions and comments.
References
 E.A. Al-Said, Cubic spline method for solving two point boundary-value problem, Korean J. Comput. Appl. Math.
5 (3) (1998) 759770.
 U.M. Ascher, R.M.M. Mattheij, R.D. Russell, Numerical Solution of Boundary Value Problems for Ordinary
dierential Equations, Prentice-Hall, Englewood Clis, NJ, 1988.
 A.K. Aziz, Numerical Solution of Two Point Boundary-Value Problem, Blaisdal, New York, 1975.
 W.G. Bickley, Piecewise cubic interpolation and two point boundary value problems, Comput. J. 11 (1968) 206208.
452 T. Aziz, A. Khan / Journal of Computational and Applied Mathematics 147 (2002) 445452
 E.P. Doolan, J.J.H. Miller, W.H.A. Schilders, Uniform Numerical Methods for Problems with Initial and Boundary
Layers, Boole Press, Dublin, Ireland, 1980.
 J.E. Flaherty, R.E. O Malley, The numerical solution of boundary-value-problems for sti dierential equations,
Math. Comput. 31 (1977) 6693.
 D.J. Fyfe, The use of cubic splines in the solution of two point boundary-value problems, Comput. J. 12 (1969)
188192.
 P. Henrici, Discrete Variable Methods in Ordinary Dierential Equations, Wiley, New York, 1962.
 M.K. Kadalbajoo, R.K. Bawa, Variable mesh dierence scheme for singularly-perturbed boundary-value problems
using splines, J. Optim. Theory Appl. 90 (2) (1996) 405416.
 M.K. Kadalbajoo, Y.N. Reddy, Asymptotic and numerical analysis of singular perturbation problems: A survey,
Appl. Math. Comput. 30 (1989) 223259.
 G. Micula, S. Micula, Hand Book of Splines, Kluwer Academic Publishers, Dordrecht, London, Boston, 1999.
 S. Natesan, N. Ramanujam, A Booster method for singular perturbation problems arising in chemical reactor theory,
Appl. Math. Comput. 100 (1999) 2748.
 P.M. Prenter, Splines and Variational Methods, Wiley, New York, 1975.
 H.G. Roos, M. Stynes, L. Tobiska, Numerical Methods for Singularly-Perturbed Dierential Equations, Springer,
New York, 1996.
 M. Stojanovic, Spline dierence methods for a singular perturbation problem, Appl. Numer. Math. 21 (3) (1996)
321333.
 O.A. Taiwo, D.J. Evans, On the solution of singularly-perturbed boundary-value problem by a Chebyshev method,
Int. J. Comput. Math. 71 (1999) 521529.