Neural, Parallel, and Scientific Computations 17 (2009) 147-168

AN ALMOST SECOND ORDER FEM FOR A WEAKLY COUPLED

SYSTEM OF TWO SINGULARLY PERTURBED DIFFERENTIAL
EQUATIONS OF REACTION-DIFFUSION TYPE WITH
DISCONTINUOUS SOURCE TERM
A. RAMESH BABU AND N. RAMANUJAM
Department of Mathematics, Bharathidasan University
Tiruchirappalli - 620024, Tamilnadu, India
ABSTRACT. In this paper, we consider a Boundary Value Problem(BVP) for a weakly coupled
system of two singularly perturbed ordinary dierential equations of reaction-diusion type with
discontinuous source term. The solution of this type of problem exhibits boundary and interior
layers. A numerical method based on nite element method on Shishkin and Bakhvalov-Shishkin
meshes is presented. We derive an error estimate of order O(N
2
ln
2
N) in the maximum norm.
Numerical experiments are also presented to support our theoretical results.
Key Words Singularly perturbed problem, Weakly coupled system, Finite element method, Dis-
continuous source term, Boundary value problem, Layer adapted mesh.
1. INTRODUCTION
Dierential equations with a small parameter (0 < 1) multiplying the high-
est order derivatives, termed as Singularly Perturbed Dierential Equations (SPDEs),
arise in diverse areas of applied mathematics, including linearized Navier - Stokes
equations of high Reynolds number, heat transfer problem with large Peclet number,
drift diusion equations of semiconductor device modelling, chemical reactor theory,
etc., In general, this type of equations exhibit boundary and/or interior layers. Stan-
dard numerical methods like nite dierence and nite element methods on uniform
mesh for solving this type of euations fail to produce good approximations to exact
solutions. Many authors [1, 2, 3, 6, 14] have developed ecient numerical methods
to resolve boundary and interior layers. A good number of articles have been ap-
pearing in the past three decades on non-classical methods which cover mostly single
second order equation. But, a few authors only have considered system of SPDEs
[9, 11, 12, 13].
In this paper, we consider the following system of singularly perturbed second or-
der ordinary dierential equations of reaction-diusion type with discontinuous source
Received June 1, 2009 1061-5369 $15.00 c _Dynamic Publishers, Inc.
148 A. R. BABU AND N. RAMANUJAM
term:
L
1
u := u

1
(x) + a
11
(x)u
1
(x) + a
12
(x)u
2
(x) = f
1
(x), x

+
,
L
2
u := u

2
(x) + a
21
(x)u
1
(x) + a
22
(x)u
2
(x) = f
2
(x), x

+
,
(1.1)
(1.2) u
1
(0) = 0, u
1
(1) = 0, u
2
(0) = 0, u
2
(1) = 0,
with conditions on coecients
a
12
(x) 0, a
21
(x) 0, (1.3)
a
11
(x) > [a
12
(x)[, a
22
(x) > [a
21
(x)[, x

, (1.4)
and for the matrix A = [a
ij
]
(1.5) A
T

T
for every = (
1
,
2
) R
2
.
Here > 0 is a small parameter, > 0, = (0, 1),

= (0, d),
+
= (d, 1), d ,
and u
1
, u
2
C
0
(

) C
1
() C
2
(

+
), u = (u
1
, u
2
)
T
. Further it is assumed that
the source terms f
1
, f
2
are suciently smooth on

d; both the functions f
1
(x)
and f
2
(x) are assumed to have a single jump discontinuity at the point d . That
is f
i
(d) ,= f
i
(d+), i = 1, 2. In general this type of discontinuity gives rise to interior
layers in the solution of the problem. Because f
i
, i = 1, 2 are discontinuous at d the
solution u of (1.1)(1.2) does not necessarily have a continuous second derivative at
the point d. That is u
1
, u
2
/ C
2
(). But the rst derivative of the solution exists
and is continuous.
Systems of this kind have applications in electro analytic chemistry when in-
vestigating diusion processes complicated by chemical reactions. The parameters
multiplying the highest derivatives characterize the diusion coecient of the sub-
stances. Other applications include equations of prey-predator population dynamics.
As mentioned above, in general, classical numerical methods fail to produce good ap-
proximations to singularly perturbed equations. Hence various methods are proposed
in the literature in order to obtain numerical solution to singularly perturbed system
of second order dierential equations subject to Dirichlet type boundary conditions
when the source terms are smooth [9]. Motivated by the works of H-G. Roos et al.[8],
in the present paper we suggest a numerical method for the above BVP. This method
is based on Finite Element Method (FEM) with layer adapted meshes like Shishkin
and Bakhvalov-Shishkin meshes. For this method we derive an error estimate of or-
der O(N
2
ln
2
N) for Shishkin mesh, O(N
2
) for Bakhvalov-Shishkin mesh, in the
maximum norm. In order to capture a boundary layer with a numerical method, it
is essential that the approximate solutions generated by the numerical method are
dened globally at each point of the domain of the exact solution. The numerical
solution obtained from a nite element method dened only at the mesh points, is
extended it to the whole domain by a simple interpolation process such as piecewise
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 149
linear interpolation. Because we want our technique to be capable of extension to
complex problems in higher dimensions, we only consider the nite element subspaces
by piecewise polynomial basis functions.
In this connection we wish to state that the authors from [13] proved almost rst
order convergence with respect to on a Shishkin mesh of the nite dierence method
with special discretization at the point d. When we compute numerical solutions, it
is not desirable to obtain error estimates in L
1
or L
2
norm, as they do not detect the
local phenomena such as boundary or interior layer. Therefore the most appropriate
norm for the study of singular perturbation problem is the maximum norm [6]. The
main signicance of this paper is that the error estimate for numerical solution is given
in terms of the maximum norm. Now we dene the maximum norm of u = (u
1
, u
2
)
as
| u |

= max| u
1
|

, | u
2
|

, | u
1
|

= max
x[0,1]
[u
1
(x)[,
| u
2
|

= max
x[0,1]
[u
2
(x)[, | u |
[x
i1
,x
i
]
= max| u
1
|
[x
i1
,x
i
]
, | u
2
|
[x
i1
,x
i
]
,
| u
1
|
[x
i1
,x
i
]
= max
x[x
i1
,x
i
]
[u
1
(x)[, | u
2
|
[x
i1
,x
i
]
= max
x[x
i1
,x
i
]
[u
2
(x)[.
Further we dene
[ u(x)[ = [(u
1
(x), u
2
(x))[ = max([u
1
(x)[, [u
2
(x)[).
Remark 1.1. Through out this paper, C denotes generic constant that is independent
of the parameter and N, the dimension of the discrete problem.
Lemma 1.2. [13] The solution u of the problem (1.1)(1.2) can be decomposed of
smooth part v and layer part w as u = v + w, v = (v
1
, v
2
), w = (w
1
, w
2
). Then, for
each k, 0 k 3, we have
[v
(k)
i
(x)[
_
_
_
C(1 +
(1k/2)
e
1
(x)), x

,
C(1 +
(1k/2)
e
2
(x)), x
+
, i = 1, 2,
[w
(k)
i
(x)[
_
_
_
C(
(k/2)
e
1
(x)), x

,
C(
(k/2)
e
2
(x)), x
+
, i = 1, 2.
where e
1
(x) = e
x

+ e
(dx)

, e
2
(x) = e
(xd)

+ e
(1x)

and =
min
x

a
11
(x) + a
12
(x), a
21
(x) + a
22
(x).
This paper is organized as follows. Section 2 presents a weak formulation of
the BVP (1.1)(1.2) and describes a nite element discretization of the problem.
Section 3 presents a role of projection operator on approximation space and error
representation. It also includes an analysis of the corresponding scheme on Shishkin
and Bakhvalov-Shishkin meshes and an interpolation error in the maximum norm. In
150 A. R. BABU AND N. RAMANUJAM
Section 4 we present a detailed error analysis of the projection operator, consistency
part and other error terms. The paper concludes with numerical examples.
2. ANALYTICAL RESULTS
A standard weak formulation of (1.1)(1.2) is: Find u = (u
1
, u
2
) (H
1
0
())
2
such
that
(2.1) B( u, v) = f

( v), v (H
1
0
())
2
,
with
B( u, v) := (B
1
( u, v), B
2
( u, v)) and f

( v) := (f

1
( v), f

2
( v)),
where
B
1
( u, v) := (u

1
, v

1
) + (a
11
u
1
+ a
12
u
2
, v
1
), (2.2)
B
2
( u, v) := (u

2
, v

2
) + (a
21
u
1
+ a
22
u
2
, v
2
), (2.3)
and
f

1
( v) = (f
1
, v
1
),
f

2
( v) = (f
2
, v
2
).
Here H
1
0
() denotes the usual Sobolev space and (, ) is the inner product on L
2
().
Now we dene a norm on (H
1
0
())
2
associated with the bilinear form B(, ), called
energy norm as
[[[ u[[[ = [([u
1
[
2
1
+[u
2
[
2
1
) + (|u
1
|
2
0
+|u
2
|
2
0
)]
1/2
,
where |u|
0
:= (u, u)
1/2
is the standard norm on L
2
(), while [u[
1
:= |u

|
0
is the
usual semi-norm on H
1
0
(). We also use the notation | u|
0
= (|u
1
|
2
0
+ |u
2
|
2
0
)
1/2
for
the norm in (L
2
())
2
. B is a bilinear functional dened on (H
1
0
())
2
. We now prove
that it is coercive with respect to [[[.[[[, that is
[ B( u, u) [
1
2
[[[ u[[[
2
,
where [B( u, u)[
2
= B
1
( u, u)
2
+ B
2
( u, u)
2
.
Lemma 2.1. A bilinear functional B satises the coercive property with respect to
[[[.[[[.
Proof. Let u (H
1
0
())
2
. Then
[B( u, u)[ =
_
B
1
( u, u)
2
+ B
2
( u, u)
2

1
2
[[B
1
( u, u)[ + [B
2
( u, u)[]
=
1
2
[(u

1
, u

1
) + (u

2
, u

2
) + (a
11
u
1
+ a
12
u
2
, u
1
) + (a
21
u
1
+ a
22
u
2
, u
2
)]
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 151

1
2
[([u
1
[
2
1
+[u
2
[
2
1
) +
_
1
0
(u
2
1
+ u
2
2
)dx]
=
1
2
[([u
1
[
2
1
+[u
2
[
2
1
) + (u
1
, u
1
) + (u
2
, u
2
)]
[B( u, u)[
1
2
[([u
1
[
2
1
+[u
2
[
2
1
) + (|u
1
|
2
0
+|u
2
|
2
0
)].
Therefore we have
[B( u, u)[
1
2
[[[ u[[[
2
.
Also we observe that B is continuous in the energy norm, that is, [ B( u, v) [
[[[ u[[[ [[[ v[[[ for some > 0. Further f

is a bounded linear functional on (H

1
0
())
2
.
By Lax-Milgram Theorem [14] we conclude that the problem (2.1) has a unique
solution.
2.1. Discretization of Weak Problem. Let

= x
0
, x
1
, . . . , x
N
to be the set of
mesh points x
i
, for some positive integer N. For i 1, 2, . . . , N we set h
i
= x
i
x
i1
to be the local mesh step size, and for i 1, 2, . . . , N let

h
i
= (h
i
+ h
i+1
)/2. We
use linear nite elements with a lumping process. That is, for discretization of (2.2)
and (2.3)
B
1h
( u, v) := (u

1
, v

1
) +
N1

i=1

h
i
a
11
(x
i
)u
1,i
v
1,i
+
N1

i=1

h
i
a
12
(x
i
)u
2,i
v
1,i
,
B
2h
( u, v) := (u

2
, v

2
) +
N1

i=1

h
i
a
21
(x
i
)u
1,i
v
2,i
+
N1

i=1

h
i
a
22
(x
i
)u
2,i
v
2,i
and f

k
( v) is replaced by

N1
i=1,i=
N
2

h
i
f
k,i
v
k,i
+
1
2
(hN
2
f
k,
N
2
1
v
k,
N
2
1
+hN
2
+1
f
k,
N
2
+1
v
k,
N
2
+1
)
for k = 1, 2. and u
k,i
= u
k
(x
i
). Then we have
B
h
( u, v) := (B
1h
( u, v), B
2h
( u, v)),
and f

h
( v) := (f

1h
( v), f

2h
( v)).
Now the discrete problem of (2.1) is: Find u
h
V
2
h
such that
(2.4) B
h
( u
h
, v
h
) = f

h
( v
h
), v
h
V
2
h
,
where V
2
h
= V
h
V
h
, V
h
is a nite dimensional subspace of H
1
0
() and the basis
functions of V
h
are given by

i
(x) =
_

_
xx
i1
h
i
, x [x
i1
, x
i
]
x
i+1
x
h
i+1
, x [x
i
, x
i+1
]
0, x / [x
i1
, x
i+1
].
152 A. R. BABU AND N. RAMANUJAM
Then

2N2
i=1
where

i
= (
i
, 0) for i = 1(1)N 1 and

i
= (0,
Ni+1
) for
i = N(1)2N 2, is a basis function of V
2
h
. Here we dene a discrete energy norm on
V
2
h
associated with a bilinear form B
h
(, ) as
[[[ u
h
[[[
V
h
= [([u
1h
[
2
1
+[u
2h
[
2
1
) + (|u
1h
|
2
0,
N

+|u
2h
|
2
0,
N

]
1/2
where |u
h
|
2
0,
N

:=

N1
k=1

h
k
u
2
h,k
and | u
h
|
0,
N

:= [|u
1h
|
2
0,
N

+ |u
2h
|
2
0,
N

]
1/2
is a dis-
crete norm on V
2
h
. B
h
is a bilinear functional dened on V
2
h
and it is coercive with
respect to [[[.[[[
V
h
. That is, [B
h
( u
h
, u
h
)[ [[[ u
h
[[[
2
V
h
, for some > 0. We can also
prove that it is continuous and f

h
is bounded linear functional on V
2
h
. By Lax-Milgram
Theorem, we conclude that the discrete problem (2.4) admits a unique solution [14].
The dierence scheme corresponding to the discrete problem (2.4) is
(2.5)

L
N
(

U
i
) :=
_

_
(L
N
1

U
i
, L
N
2

U
i
) = (

h
i
f
1,i
,

h
i
f
2,i
), i = 1(1)N 1, i ,=
N
2
,
(L
N
1

UN
2
, L
N
2

UN
2
) = (
1
2
[hN
2
f
1,
N
2
1
+ hN
2
+1
f
1,
N
2
+1
],
1
2
[hN
2
f
2,
N
2
1
+ hN
2
+1
f
2,
N
2
+1
]),

U
0
=

U
N
=

0 = (0, 0),
where
L
N
1

U
i
= (
U
1,i+1
U
1,i
h
i+1

U
1,i
U
1,i1
h
i
) +

h
i
a
11
(x
i
)U
1,i
+

h
i
a
12
(x
i
)U
2,i
L
N
2

U
i
= (
U
2,i+1
U
2,i
h
i+1

U
2,i
U
2,i1
h
i
) + +

h
i
a
21
(x
i
)U
1,i
+

h
i
a
22
(x
i
)U
2,i
and

U
i
= (U
1,i
, U
2,i
), U
1,i
= U
1
(x
i
), f
1,i
= f
1
(x
i
) and similarly for U
2,i
, f
2,i
, i =
1(1)N 1.
3. ERROR ANALYSIS - I
Now the given discrete problem is: Find u
h
V
2
h
(H
1
0
())
2
such that
(3.1) B
h
( u
h
, v
h
) = f

h
( v
h
), v
h
V
2
h
.
Since the above discrete problem admits a unique solution and some interpolant
u
I
V
2
h
of u is well dened. We dene a biorthogonal basis of V
2
h
with respect to B
h
to be the set of functions

2N2
j=1
where

j
= (
j
1
,
j
2
) for j = 1(1)2N 2, which
satises the condition
B
h
(

i
,

j
) = (
ij
,
ij
) for i, j = 1(1)2N 2. (3.2)
In otherwords
B
1h
(

i
,

j
) =
ij
for i, j = 1(1)2N 2,
B
2h
(

i
,

j
) =
ij
,
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 153
where
ij
is the Kronecker symbol. Then the components u
1h
and u
2h
can be uniquely
represented as
u
1h
=
N1

i=1
B
1h
((u
1h
, u
2h
), (
i
1
,
i
2
))
i
and u
2h
=
N1

i=1
B
2h
((u
1h
, u
2h
), (
N+i1
1
,
N+i1
2
))
i
.
Dene linear transformations P
1
, P
2
: (H
1
0
())
2
V
h
such that
P
1
u :=
N1

i=1
B
1h
((u
1
, u
2
), (
i
1
,
i
2
))
i
and P
2
u :=
N1

i=1
B
2h
((u
1
, u
2
), (
N+i1
1
,
N+i1
2
))
i
.
Let

P = (P
1
, P
2
) and u
h
V
2
h
. Then

P u
h
= (P
1
u
h
, P
2
u
h
)
=
_
N1

i=1
B
1h
((u
1h
, u
2h
), (
i
1
,
i
2
))
i
,
N1

i=1
B
2h
((u
1h
, u
2h
), (
N+i1
1
,
N+i1
2
))
i
_
= (u
1h
, u
2h
).
That is,

P u
h
= u
h
, u
h
V
2
h
.
Hence

P is a projection operator on V
2
h
. Now, the error u u
h
can be written as,
(3.3) u u
h
= u u
I
+

P( u
I
u) +

P u u
h
.
We estimate this error in the rest of this section.
3.1. Shishkin and Bakhvalov-Shishkin Meshes. For the discretization described
above we shall use meshes of the general type introduced in [5], but here adapted for
the boundary layers at x = 0 and x = 1 and the interior layers at x = d. Let N > 8
be a positive even integer and

1
= min
d
4
,
_

ln N,
2
= min
1 d
4
,
_

ln N, 2.
Our mesh will be equidistant on

S
, where

S
= (
1
, d
1
) (d +
2
, 1
2
)
and graded on

0
where

0
= (0,
1
) (d
1
, d) (d, d +
2
) (1
2
, 1).
154 A. R. BABU AND N. RAMANUJAM
We choose the transition points to be
x
N/8
=
1
, x
3N/8
= d
1
, x
N/2
= d, x
5N/8
= d +
2
, x
7N/8
= 1
2
.
Because of the specic layers, here we have to use four mesh generating functions

1
,
2
,
3
and
4
which are all continuous and piecewise continuously dierentiable,
with the following properties:
1
and
3
are monotonically increasing,
2
and
4
are
monotonically decreasing functions and

1
(0) = 0,
1
(1/8) = lnN,

2
(3/8) = ln N,
2
(1/2) = 0,

3
(1/2) = 0,
3
(5/8) = lnN,

4
(7/8) = ln N,
4
(1) = 0.
The mesh points are
x
i
=
_

_
2
_

1
(t
i
), i = 0(1)N/8,

1
+
4
N
(d 2
1
)(i N/8), i = N/8 + 1(1)3N/8,
d 2
_

2
(t
i
), i = 3N/8 + 1(1)N/2,
d + 2
_

3
(t
i
), i = N/2 + 1(1)5N/8,
d +
2
+
4
N
(1 d 2
2
)(i 5N/8), i = 5N/8 + 1(1)7N/8,
1 2
_

4
(t
i
), i = 7N/8 + 1(1)N,
where t
i
= i/N. We dene new functions
1
,
2
,
3
and
4
by

i
= ln
i
, i = 1(1)4.
There are several mesh-characterizing functions
i
in the literature, but we shall
use only those which correspond to Shishkin mesh and Bakhvalov-Shishkin mesh with
the following properties
max [

[ = C ln N for Shishkin mesh,

max [

[ = C for Bakhvalov-Shishkin mesh.

For Shishkin mesh we take

1
(t) = e
8t ln N
,
2
(t) = e
4(12t) ln N
,

3
(t) = e
4(2t1) ln N
,
4
(t) = e
8(1t) lnN
,
whereas for Bakhvalov-Shishkin mesh

1
(t) = 1 8(1 N
1
)t,
2
(t) = 1 4(1 N
1
)(1 2t),

3
(t) = 1 4(1 N
1
)(2t 1),
4
(t) = 1 8(1 N
1
)(1 t).
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 155
The set of interior mesh points is denoted by
N

x
0
, x
N/2
, x
N
. Also, for the
both meshes, on the coarse part
S
we have
h
i
CN
1
.
It is well known that on the layer part [8]
h
i
C

N
1
ln N for Shishkin mesh
and
h
i
C(

+ N
1
) for Bakhvalov-Shishkin mesh.
In the later analysis, the following estimates of e
1
(x) and e
2
(x) will be used [8]:
(3.4) e
1
(x)
_
_
_
C, x

0
CN

, x

S
.
e
2
(x)
_
_
_
C, x
+

0
CN

, x
+

S
.
3.2. Interpolation Error. To derive an error estimate, we consider the interpolation
error in the maximum norm. Let f C
2
[x
i1
, x
i
] be arbitrary and f
I
a piecewise
linear interpolant to f on . Then from the classical theory, we have
[(f
I
f)(x)[ 2
_
x
i
x
i1
[f

(t)[(t x
i1
)dt, x [x
i1
, x
i
].
Now we compute the interpolation error for u
i
, i = 1, 2 seperately.
Lemma 3.1. If

CN
1
and for the Shishkin mesh, we have
[u
i
(x) u
I
i
(x)[
_
_
_
CN
2
ln
2
N, x
0
CN
2
, x
S
and for the Bakhavalov-Shishkin mesh it holds
[u
i
(x) u
I
i
(x)[ CN
2
, x

+
, i = 1, 2.
Proof. We now give a proof for the case i = 1 for the Shishkin mesh. To prove the
estimate we use the decomposition of solution as smooth and layer components and
triangle inequality
(3.5) [(u
1
u
I
1
)(x)[ [(v
1
v
I
1
)(x)[ +[(w
1
w
I
1
)(x)[.
On Shishkin mesh, let x [x
i1
, x
i
]

S
. Then by using (3.4) we have to
compute the rst term of (3.5)
[(v
1
v
I
1
)(x)[ 2
_
x
i
x
i1
[v

1
(t)[(t x
i1
)dt
2C
_
x
i
x
i1
(t x
i1
)dt + 2C
_
x
i
x
i1
[ e
1
(t) [ (t x
i1
)dt
C
h
2
i
2
+ CN

h
2
i
2
156 A. R. BABU AND N. RAMANUJAM
[(v
1
v
I
1
)(x)[ CN
2
.
Again the second term of (3.5) will be
[(w
1
w
I
1
)(x)[ 2|w
1
(x)|
L[x
i1
,x
i
]
CN

[(w
1
w
I
1
)(x)[ CN
2
.
Now let x [x
i1
, x
i
]

0
we have
[(v
1
v
I
1
)(x)[ 2
_
x
i
x
i1
[v

1
(t)[(t x
i1
)dt
2C
_
x
i
x
i1
(t x
i1
)dt + 2C
_
x
i
x
i1
[ e
1
(t) [ (t x
i1
)dt
C
h
2
i
2
[(v
1
v
I
1
)(x)[ CN
2
,
and also the layer component will be
[(w
1
w
I
1
)(x)[ 2
_
x
i
x
i1
[w

1
(t)[(t x
i1
)dt
2C
1
_
x
i
x
i1
[ e
1
(t) [ (t x
i1
)dt
C
1
h
2
i
2
C
1
(

N
1
ln N)
2
[(w
1
w
I
1
)(x)[ CN
2
ln
2
N.
Similarly we can obtain a similar estimate for x
+
.
To prove estimates on Bakhvalov-Shishkin mesh, we follow the above procedure.
If x [x
i1
, x
i
]
S
then h
i
CN
1
and if x [x
i1
, x
i
]
0
then h
i
C(

+
N
1
). Using the fact that max [

[ = C and

CN
1
we can arrive the required
result.
Then the interpolation error of u in maximum norm is
(3.6) | u u
I
|

_
_
_
CN
2
ln
2
N, for Shishkin mesh,
CN
2
, for Bakhalov-Shishkin mesh.
4. ERROR ANALYSIS - II
Let x
k

be a mesh point. From equation (3.3), the second term at the points
of the mesh is

P( u
I
u)(x
k
) = (P
1
( u
I
u)(x
k
), P
2
( u
I
u)(x
k
)).
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 157
Each of the components of the above will be estimated seperately. We have
P
1
( u
I
u)(x
k
) = B
1h
(((u
I
1
u
1
), (u
I
2
u
2
)), (
k
1
,
k
2
))
= ((u
I
1
u
1
)

, (
k
1
)

) +
N1

i=1

h
i
a
11
(x
i
)(u
I
1
u
1
)(x
i
)
k
1,i
+
N1

i=1

h
i
a
12
(x
i
)(u
I
2
u
2
)(x
i
)
k
1,i
.
Using integration by parts and the fact that (
k
1
)

= 0, u
I
1
(x
i
) = u
1
(x
i
), u
I
2
(x
i
) =
u
2
(x
i
) for i = 1(1)N 1, we have P
1
( u
I
u)(x
k
) = 0. Similarly P
2
( u
I
u)(x
k
) = 0.
Therefore, we have
(4.1)

P( u
I
u)(x
k
) =

0.
The remaining part of this section is devoted for the estimation of the third term
of the error representation (3.3). For this representation, we rst need the following
L
1
estimates of discrete Greens functions
j
1
and
j
2
.
Lemma 4.1. On an arbitrary mesh, the discrete Greens functions (
j
1
,
j
2
) for B
h
satisfy |
j
1
|
L
1
()
C and |
j
2
|
L
1
()
C, where |
j
1
|
L
1
()
=

N
i=1
_
x
i
x
i1
[
j
1
[dx and
|
j
2
|
L
1
()
=

N
i=1
_
x
i
x
i1
[
j
2
[dx.
Proof. Following the procedure adopted in [8, 10], we can prove this theorem.
Let

K = (K
1
, K
2
) =

P u u
h
= ((P
1
uu
1h
), (P
2
uu
2h
)). That is, K
1
= P
1
uu
1h
and K
2
= P
2
u u
2h
. Now,
K
1
= P
1
u u
1h
=
N1

i=1
B
1h
((u
1
, u
2
), (
i
1
,
i
2
))
i

N1

i=1
B
1h
((u
1h
, u
2h
), (
i
1
,
i
2
))
i
=
N1

i=1
B
1h
((u
1
, u
2
), (
i
1
,
i
2
))
i
+
N1

i=1
f

1
((
i
1
,
i
2
))
i

N1

i=1
B
1
((u
1
, u
2
), (
i
1
,
i
2
))
i

N1

i=1
B
1h
((u
1h
, u
2h
), (
i
1
,
i
2
))
i
=
N1

i=1
(B
1h
B
1
)((u
1
, u
2
), (
i
1
,
i
2
))
N1

i=1
f

1h
((
i
1
,
i
2
))
i
+
N1

i=1
f

1
((
i
1
,
i
2
))
i
.
K
1
=
N1

i=1
(B
1h
B
1
)((u
1
, u
2
), (
i
1
,
i
2
)
i
+
N1

i=1
(f

1
f

1h
)((
i
1
,
i
2
))
i
,
where u
1
(x
i
) = u
1,i
, u
2
(x
i
) = u
2,i
and a
12
(x
i
) = a
12,i
. Then we have
K
1
(x
k
) = (B
1h
B
1
)((u
1
, u
2
), (
k
1
,
k
2
)) + (f

1
f

1h
)((
k
1
,
k
2
))
= B
1h
((u
1
, u
2
), (
k
1
,
k
2
))
158 A. R. BABU AND N. RAMANUJAM
B
1
((u
1
, u
2
), (
k
1
,
k
2
)) + f

1
((
k
1
,
k
2
)) f

1h
((
k
1
,
k
2
))
= (u

1
, (
k
1
)

) +
N1

i=1

h
i
a
11,i
u
1,i

k
1,i
+
N1

i=1

h
i
a
12,i
u
2,i

k
1,i
(u

1
, (
k
)

_
1
0
a
11
(x)u
1
(x)
k
(x)dx
_
1
0
a
12
(x)u
2
(x)
k
(x)dx
+
_
1
0
f
1
(x)
k
1
(x)dx
N1

i=1

h
i
f
1,i

k
1,i
K
1
(x
k
) = (
N1

i=1

h
i
a
11,i
u
1,i

k
1,i

_
1
0
a
11
(x)u
1
(x)
k
(x)dx) + (
N1

i=1

h
i
a
12,i
u
2,i

k
1,i

_
1
0
a
12
(x)u
2
(x)
k
1
(x)dx) + (
_
1
0
f
1
(x)
k
1
(x)dx
N1

i=1

h
i
f
1,i

k
1,i
).
Similarly we get
K
2
(x
k
) = (
N1

i=1

h
i
a
21,i
u
1,i

N+k1
2,i

_
1
0
a
21
(x)u
1
(x)
N+k1
2
(x)dx) + (
N1

i=1

h
i
a
22,i
u
2,i

N+k1
2,i

_
1
0
a
22
(x)u
2
(x)
N+k1
2
(x)dx) + (
_
1
0
f
2
(x)
N+k1
2
(x)dx
N1

i=1

h
i
f
2,i

N+k1
2,i
).
Now we dene
K

1
(x
k
) =
N1

i=1

h
i
a
11,i
u
1,i

k
1,i

i=1
_
x
i
x
i1
(a
11
(x)u
1
(x))
I

k
1
(x)dx +
N1

i=1

h
i
a
12,i
u
2,i

k
1,i

i=1
_
x
i
x
i1
(a
12
(x)u
2
(x))
I

k
1
(x)dx +
N

i=1
_
x
i
x
i=1
f
I
1
(x)
k
1
(x)dx
N1

i=1

h
i
f
1,i

k
1,i
.
Then we can write K
1
(x
k
) as
K
1
(x
k
) = K

1
(x
k
) +
N

i=1
_
x
i
x
i1
((a
11
(x)u
1
(x))
I
(a
11
(x)u
1
(x)))
k
1
(x)dx
+
N

i=1
_
x
i
x
i1
((a
12
(x)u
2
(x))
I
(a
12
(x)u
2
(x)))
k
1
(x)dx

i=1
_
x
i
x
i1
((f
1
(x))
I
f
1
(x))
k
1
(x)dx.
(4.2)
The later sums of K
1
(x
k
) can be bounded by
[
N

i=1
_
x
i
x
i1
((a
11
(x)u
1
(x))
I
(a
11
(x)u
1
(x)))
k
1
(x)dx [
C(|u
1
u
I
1
|

|a
11
|

+|a
I
11
a
11
|

|u
1
|

)|
k
1
|
L
1
()
C(|u
1
u
I
1
|

+ N
2
|u
1
|

)|
k
1
|
L
1
()
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 159
C(|u
1
u
I
1
|

+ N
2
)
CN
2
max [

[
2
,
[
N

i=1
_
x
i
x
i1
((a
11
(x)u
1
(x))
I
(a
11
(x)u
1
(x)))
k
1
(x)dx[ CN
2
max [

[
2
, (4.3)
[
N

i=1
_
x
i
x
i1
((a
12
(x)u
2
(x))
I
(a
12
(x)u
2
(x)))
k
1
(x)dx[ CN
2
max [

[
2
, (4.4)
and
(4.5) [
N

i=1
_
x
i
x
i1
(f
1
f
I
1
)
k
1
(x)dx [ CN
2
|
k
1
|
L
1
()
CN
2
.
If we dene K

2
(x
k
) similar to K

1
(x
k
), then we can write
K
2
(x
k
) = K

2
(x
k
) +
N

i=1
_
x
i
x
i1
((a
21
(x)u
1
(x))
I
(a
21
(x)u
1
(x)))
N+k1
2
(x)dx
+
N

i=1
_
x
i
x
i1
((a
22
(x)u
2
(x))
I
(a
22
(x)u
2
(x)))
N+k1
2
(x)dx

i=1
_
x
i
x
i1
((f
2
(x))
I
f
2
(x))
N+k1
2
(x)dx.
(4.6)
We can also estimate the later sums of K
2
(x
k
) as done for K
1
(x
k
). In the pointwise
errors, K
1
(x
k
) and K
2
(x
k
) it remains only to estimate the expressions K

1
(x
k
) and
K

2
(x
k
). First we write K

1
(x
k
) and K

2
(x
k
) in the form
K

1
(x
k
) = (a
11
u
1
)
I
,
k
1
)
h
+(a
12
u
2
)
I
,
k
1
)
h
f
I
1
,
k
1
)
h
, (4.7)
K

2
(x
k
) = (a
21
u
1
)
I
,
N+k1
2
)
h
+(a
22
u
2
)
I
,
N+k1
2
)
h
f
I
2
,
N+k1
2
)
h
, (4.8)
where
g,
k
)
h
=
N1

i=1

h
i
g(x
i
)
k
i

N

i=1
_
x
i
x
i1
g(x)
k
(x)dx,
for a piecewise linear function g, not neccessarily continuous. For integrals in the
previous formula, we use Simpsons rule
(4.9) g,
k
)
h
=
1
6
N1

i=1
(h
i
(g

i
g
+
i1
) h
i+1
(g

i+1
g
+
i
))
k
i
.
In order to estimate K

1
(x
k
) and K

2
(x
k
), we start with the decomposition of the
solution u. Hence we seperately analyze smooth part v and the layer part w. Now
the equation (4.7) can be rewritten as
(4.10) K

1
(x
k
) = (v

1
)
I
,
k
1
)
h
+(a
11
w
1
+ a
12
w
2
)
I
,
k
1
)
h
.
160 A. R. BABU AND N. RAMANUJAM
The rst term in the above expression containing the regular component v
1
can be
easily estimated. In fact,
[ (v

1
)
I
,
k
1
)
h
[

6
N1

i=1
(h
i
(v

1,i
v

+
1,i1
) h
i+1
(v

1,i+1
v

+
1,i
))
k
1,i
.
Then we have
[ (v

1
)
I
,
k
1
)
h
[

6
N1

i=1
(h
2
i
|v

1
|
L

[x
i1
,x
i
]
+ h
2
i+1
|v

1
|
L

[x
i
,x
i+1
]

k
1,i

3
(CN
1
)|v

1
|
L

()
N1

i=1
(
h
i
+ h
i+1
2
)
k
1,i
C

N
1
N1

i=1

h
i

k
1,i
C

N
1
|
k
1
|
L
1
()
,
by using (1.2), h
i
CN
1
, i = 1(1)N 1 and |
k
1
|
L
1
()
C. Finally, we get
(4.11) [ (v

1
)
I
,
k
1
)
h
[ C

N
1
.
Let us denote the coecient in < (a
11
w
1
+ a
12
w
2
)
I
,
k
1
>
h
corresponding to
k
1,i
by
m
i
. Depending on the values of index i, we consider dierent cases. In general, g

i
denotes right-limit and left-limit of a function g at a mesh point x
i
.
Case 1: When
N
8
+ 1 i
3N
8
1 or
5N
8
+ 1 i
7N
8
1. That is,
[x
i1
, x
i+1
]
S
. The coecient m
i
can be estimated by
[m
i
[ = [ h
i
(a

11,i
w

1,i
a
+
11,i1
w
+
1,i1
) h
i+1
(a

11,i+1
w

1,i+1
a
+
11,i
w
+
1,i
) [
+ [ h
i
(a

12,i
w

2,i
a
+
12,i1
w
+
2,i1
) h
i+1
(a

12,i+1
w

2,i+1
a
+
12,i
w
+
2,i
) [
C

h
i
[|w
1
|
L[x
i1
,x
i+1
]
+|w
2
|
L[x
i1
,x
i+1
]
]
C

h
i
[max
x
S
[e
1
(x)[ + max
x
S
[e
2
(x)[], from (1.2) and (3.4).
(4.12) [ m
i
[ C

h
i
N

.
Case 2: When 1 i
N
8
1 or
3N
8
+ 1 i
N
2
1 or
N
2
+ 1 i
5N
8
1 or
3N
8
+1 i N 1. That is, the subinterval [x
i1
, x
i+1
]
0
. The layer part will be
calculated by estimating m
i
. We have
m
i
= h
i
(a

11,i
w

1,i
a
+
11,i1
w
+
1,i1
) h
i+1
(a

11,i+1
w

1,i+1
a
+
11,i
w
+
1,i
)
+ h
i
(a

12,i
w

2,i
a
+
12,i1
w
+
2,i1
) h
i+1
(a

12,i+1
w

2,i+1
a
+
12,i
w
+
2,i
)
= h
i
(a
11,i+1
w
1,i+1
+ 2a
11,i
w
1,i
a
11,i1
w
1,i1
)
+ (h
i
h
i+1
)(a
11,i+1
w
1,i+1
a
11,i
w
1,i
)
+ h
i
(a
12,i+1
w
2,i+1
+ 2a
12,i
w
2,i
a
12,i1
w
2,i1
)
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 161
+ (h
i
h
i+1
)(a
12,i+1
w
2,i+1
a
12,i
w
2,i
)
= a
11,i
(h
i
(w
1,i+1
+ 2w
11,i
w
1,i1
) + (h
i
h
i+1
)(w
1,i+1
w
1,i
))
+ h
i
(a
11,i
a
11,i1
)
(w
1,i1
w
1,i
) + h
i+1
(a
11,i+1
a
11,i
)(w
1,i
w
1,i+1
)
+ w
1,i
(h
i+1
a
11,i+1
+ (h
i
+ h
i+1
)
a
11,i
h
i
a
11,i1
)a
12,i
(h
i
(w
2,i+1
+ 2w
21,i
w
2,i1
) + (h
i
h
i+1
)(w
2,i+1
w
2,i
))
+ h
i
(a
12,i
a
12,i1
)(w
2,i1
w
2,i
) + h
i+1
(a
12,i+1
a
12,i
)(w
2,i
w
2,i+1
)
+ w
2,i
(h
i+1
a
12,i+1
+ (h
i
+ h
i+1
)a
12,i
h
i
a
12,i1
).
Using the Taylors expansion for each of the terms in the previous expression yields
h
i
a
11,i
(w
1,i+1
+ 2w
1,i
w
1,i1
) = h
i
(h
i
h
i+1
)a
11,i
w

1,i

h
3
i
2
a
11,i
w

1
(
i
)
h
i
h
2
i+1
2
a
11,i
w

1
(
i+1
),
h
i
a
12,i
(w
2,i+1
+ 2w
2,i
w
2,i1
) = h
i
(h
i
h
i+1
)a
12,i
w

2,i

h
3
i
2
a
12,i
w

2
(
i
)
h
i
h
2
i+1
2
a
12,i
w

2
(
i+1
),
(h
i
h
i+1
)a
11,i
(w
1,i+1
w
1,i
) = h
i+1
(h
i
h
i+1
)a
11,i
w

1
(
i+1
),
(h
i
h
i+1
)a
12,i
(w
2,i+1
w
2,i
) = h
i+1
(h
i
h
i+1
)a
12,i
w

2
(
i+1
),
h
i
(a
11,i
a
11,i1
)(w
1,i1
w
1,i
) = h
3
i
a

11
(
i
)w

1
(
i
),
h
i
(a
12,i
a
12,i1
)(w
2,i1
w
2,i
) = h
3
i
a

12
(
i
)w

2
(
i
),
h
i+1
(a
11,i+1
a
11,i
)(w
1,i
w
1,i+1
) = h
3
i+1
a

11
(
i+1
)w

1
(
i+1
),
h
i+1
(a
12,i+1
a
12,i
)(w
2,i
w
2,i+1
) = (h
2
i
h
2
i+1
)a

12,i
w
2,i

1
2
(h
3
i
a

12
(
k
) + h
3
i+1
a

12
(
i+1
))w
2,i
,
where
i
,
i
,
i
,
i
[x
i1
, x
i
].
To derive an estimate for [m
i
[, we need the following lemma.
Lemma 4.2. For the points x
i1
, x
i
, x
i+1

0
, x
i
,= d = xN
2
of the mesh with 2
the following holds
[ (h
i
h
i+1
)(w
1,i+1
w
1,i
) [ Ch
i+1
N
2
,
[ (h
i
h
i+1
)(w
2,i+1
w
2,i
) [ Ch
i+1
N
2
,
[ (h
i
h
i+1
)w

1,i
[ CN
2
and [ (h
i
h
i+1
)w

2,i
[ CN
2
.
162 A. R. BABU AND N. RAMANUJAM
Proof. Let x
i1
, x
i
, x
i+1

and x
i
,= d = xN
2
[ h
i
h
i+1
[= 2
_

N
1
[

(
i
)

(
i+1
) [
for
i
,
i+1
(t
i1
, t
i+1
). Also [ w
1,i+1
w
1,i
[= h
i+1
[ w

1,i
(
i+1
) [,
i+1
(x
i
, x
i+1
)
[ (h
i
h
i+1
)(w
1,i+1
w
1,i
) [ C

h
i+1
N
2
[

1
(
i
) [[ w

1
(
i+1
) [
C

h
i+1
N
2
(

(x
i
)

1
(x
i
)
)
2
e
1
(
i+1
)
1
2
Ch
i+1
N
2
(
max

1
(x
i
)
)
2
e
1
(
i+1
)
Ch
i+1
N
2
(
1
(t
i+1
))
2
e
1
(
i+1
).
Using the fact that max [

1
[ = C and e
1
(
i+1
)
1
(t
i
)
2
+ N

, we have
[ (h
i
h
i+1
)(w
1,i+1
w
1,i
) [ Ch
i+1
N
2
(
1
(t
i
)
2
+ N

)(
1
(t
i+1
))
2
[ (h
i
h
i+1
)(w
1,i+1
w
1,i
) [ Ch
i+1
N
2
,
since 2. When [x
i1
, x
i+1
] [d, d +
2
] and [x
i1
, x
i+1
] [1
2
, 1], the above
estimate is also true for these intervals. From the previous analysis, we get
h
i
a
11,i
(w
1,i+1
+ 2w
1,i
w
1,i1
) Ch
i
N
2
+ Ch
i
N
2
max [

1
[,
h
i
a
12,i
(w
2,i+1
+ 2w
2,i
w
2,i1
) Ch
i
N
2
+ Ch
i
N
2
max [

1
[,
and
(h
i
h
i+1
)a
11,i
(w
1,i+1
w
1,i
) Ch
i+1
N
2
,
(h
i
h
i+1
)a
12,i
(w
2,i+1
w
2,i
) Ch
i+1
N
2
.
Applying the above Lemma 4.2 to each of the terms in m
i
of Case 2, we have
(4.13) [m
i
[ C

h
i
N
2
max [

[
2
.
Now it remains to prove the estimates at the transition points.
Case 3: When x
i
, i
N
8
,
3N
8
,
5N
8
,
7N
8
and i ,=
N
2
. At these points w
1,i
, w
1,i1
and w
2,i
, w
2,i1
are bounded by CN

. Then, using the expression for [m

i
[ given in
Case 2,
(4.14) [ m
i
[ C

h
i
N

.
Case 4: When i =
N
2
. That is, x
i
= d
m
i
= h
i
(a

11,i
w

1,i
a
+
11,i1
w
+
1,i1
) h
i+1
(a

11,i+1
w

1,i+1
a
+
11,i
w
+
1,i
)
+ h
i
(a

12,i
w

2,i
a
+
12,i1
w
+
2,i1
) h
i+1
(a

12,i+1
w

2,i+1
a
+
12,i
w
+
2,i
)
= h
i
(a
11,i+1
w
1,i+1
+ a
+
11,i
w
+
1,i
+ a

11,i
w

1,i
a
11,i1
w
1,i1
)
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 163
h
i
(a
12,i+1
w
2,i+1
+ a
+
12,i
w
+
2,i
+ a

12,i
w

2,i
a
12,i1
w
2,i1
)
[m
i
[ h
i
[(a
+
11,i
a
11,i+1
)w
+
1,i
+ (a

11,i
a
11,i1
)w

1,i
[ + h
i
[a
11+1,i
(w
+
1,i
w
1,i+1
)
+ a
11,i1
(w

1,i
w
1,i1
)[ + h
i
[(a
+
12,i
a
12,i+1
)w
+
2,i
+ (a

12,i
a
12,i1
)w

2,i
[
+ h
i
[a
12+1,i
(w
+
2,i
w
2,i+1
) + a
12,i1
(w

2,i
w
2,i1
)[
Ch
i
h
i+1
[w
+
1,i
[ + Ch
2
i
[w

1,i
[ + Ch
i
(h
i
(a
11,i1
a

11,i
) w

1,i

1
2
h
2
i+1
a

11,i
w

1
(
i
)
+
1
2
h
2
i
a
11,i1
w

1
(
i
) + R
1
) + Ch
i
h
i+1
[w
+
2,i
[ + Ch
2
i
[w

2,i
[ + Ch
i
(h
i
(a
12,i1
a

12,i
) w

2,i

1
2
h
2
i+1
a

12,i
w

2
(
i
) +
1
2
h
2
i
a
12,i1
w

2
(
i
) + R
2
),
i
[x
i1
, x
i
].
We use the asymptotic expansion of the layer components w
1
= w
1
+ R
1
and w
2
=
w
2
+R
2
, that can be derived using the technique from [4]. It can be concluded that the
leading part w

1
of w

1
and w

2
of w

2
are continuous at x = d, enabling us to use Taylors
expansions for estimating w
+
1,i
w
1,i+1
, w

1,i
w
1,i1
and w
+
2,i
w
2,i+1
, w

2,i
w
2,i1
.
Since R
1
, R
2
contain lower order terms, we have
(4.15) [ m
i
[ C

h
i

N
1
+ C

h
i

N
2
max [

[
2
+ C

h
i
N
2
max [

[
2
,
and we use the estimate of max [

[ in the above result to obtain

[ m
i
[
_
_
_
C

h
i
(

+ N
1
)N
1
ln
2
N, for Shishkin mesh,
C

h
i
(

+ N
1
)N
1
, for Bakhalov-Shishkin mesh.
Collecting estimates (4.12)(4.15) from the previously analyzed cases and using


CN
1
, we have
[< (a
11
w
1
+ a
12
w
2
)
I
,
k
1
>
h
[
1
6
N1

i=1
[m
i
[
k
1,i
C(N

+ N
2
max [

[)
N1

i=1

h
i

k
1,i
CN
2
max [

[|
k
1
|
L
1
()
CN
2
max [

[,
since 2 and |
k
1
|
L
1
()
C.
From (4.10), (4.11) and the above estimate, we have
(4.16) K

1
(x
k
) C

N
1
+ CN
2
max [

[.
A similar estimate is also hold for K

2
(x
k
), from (4.8). Therefore from equations
(4.2)(4.6), (4.16) and max [

[ = C ln N (in case of Shishkin mesh), for p = 1, 2 we

have
K
p
(x
k
) C

N
1
+ CN
2
ln
2
N.
164 A. R. BABU AND N. RAMANUJAM
Since [

K(x
i
)[ = max([K
1
(x
i
)[, [K
2
(x
i
)[), we have
[

P u(x
i
) u
h
(x
i
)[ = [

K(x
i
)[ C

N
1
+ CN
2
ln
2
N.
Therefore we conclude
Lemma 4.3. Let u and u
h
be solution of the BVP (1.1)(1.2) and (2.5) respectively.
Then for Shishkin mesh, the pointwise maximum norm of the error satises
[ u(x
i
) u
h
(x
i
)[ = [

P u(x
i
) u
h
(x
i
)[ C

N
1
+ CN
2
ln
2
N, x
i

N

.
Now, since V
h
uses linear Lagrange elements, we can easily derive a bound for
the error u
j
u
jh
on each element [x
i1
, x
i
], i = 1(1)N, j = 1, 2. For arbitrary
i 1, 2, ..., N and x [x
i1
, x
i
], the triangle inequality implies
[u
j
(x) u
jh
(x)[ [u
j
(x) u
I
j
(x)[ +[u
I
j
(x) u
jh
(x)[, j = 1, 2.
The dierence between the piecewise linear function u
I
j
and u
jh
at the point x is
estimated by
[u
I
j
(x) u
jh
(x)[ = [u
j
(x
i1
)
i1
(x) + u
j
(x
i
)
i
(x) u
jh
(x
i1
)
i1
(x) u
jh
(x
i
)
i
(x)[
[u
j
(x
i1
) u
jh
(x
i1
)[
i1
(x) +[u
j
(x
i
) u
jh
(x
i
)[
i
(x)
C

N
1
+ CN
2
ln
2
N, by Lemma 4.3 for i = 2(1)N 1,
where
i
are functions dened in Section 2.1. The same bound holds for i = 1 and
i = N. Therefore for each interval [x
i1
, x
i
] we nally obtain the error estimate
(4.17) [u
j
(x) u
jh
(x)[ | u u
I
|
[x
i1
,x
i
]
+ C

N
1
+ CN
2
ln
2
N, j = 1, 2
where | u u
I
|
[x
i1
,x
i
]
= max(|u
1
u
I
1
|
[x
i1
,x
i
]
, |u
2
u
I
2
|
[x
i1
,x
i
]
).
5. ERROR ESTIMATE
The following theorem gives us the result on the maximum norm of the error
u u
h
not just on the mesh points, but on the whole domain [0, 1].
Theorem 5.1. Let u and u
h
be solution of BVP (1.1)(1.2) and (2.5) respectively,

CN
1
and 2. Then we have
| u u
h
|

_
_
_
CN
2
ln
2
N, for Shishkin mesh,
CN
2
, for Bakhalov-Shishkin mesh.
Proof. For Shishkin mesh, the theorem follows from the inequality (4.17) and the
results on interpolation error (3.6). Similarly we can prove for Bakhalov-Shishkin
mesh.
Remark 5.2. All the results in this paper also hold in case when the functions f
1
and f
2
have more than one point of discontinuity.
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 165
6. NUMERICAL EXPERIMENTS
In this section we experimentally verify our theoretical results proved in the
previous section.
Example 6.1. Consider the BVP
u

1
(x) + 2u
1
(x) u
2
(x) = f
1
(x), x

+
,
u

2
(x) u
1
(x) + 2u
2
(x) = f
2
(x), x

+
,
(6.1)
(6.2) u
1
(0) = 0, u
1
(1) = 0, u
2
(0) = 0, u
2
(1) = 0,
where
f
1
(x) =
_
_
_
1, 0 x 0.5,
0.8, 0.5 x 1
and f
2
(x) =
_
_
_
2, 0 x 0.5,
1.8, 0.5 x 1.
Example 6.2. Consider the BVP
u

1
(x) + 2(x + 1)
2
u
1
(x) (1 + x
3
)u
2
(x) = f
1
(x), x

+
,
u

2
(x) 2cos(x/4)u
1
(x) + 2.2e
1x
u
2
(x) = f
2
(x), x

+
,
(6.3)
(6.4) u
1
(0) = 0, u
1
(1) = 0, u
2
(0) = 0, u
2
(1) = 0,
where
f
1
(x) =
_
_
_
2e
x
, 0 x 0.5,
1, 0.5 x 1
and f
2
(x) =
_
_
_
10x + 1, 0 x 0.5,
2, 0.5 x 1.
For our tests, we take = 2
18
, which is suciently small to bring out the
singularly perturbed nature of the problem. We measure the accuracy in maximum
norm and the rates of convergence r
N
are computed using the following formula:
r
N
= log
2
(
E
N
E
2N
),
where
E
N
=| u
h
u
I
2h
|

,
and u
I
h
denotes the piecewise linear interpolant of u
h
. In Table 1, we present values
of E
N
, r
N
for the solution of the BVPs (6.1)(6.2) and (6.3)(6.4) for Shishkin and
B-Shishkin meshes respectively. The Figures 1 and 2 depict the numerical solution
of the BVP (6.1)(6.2) for Shishkin mesh with N = 512. We compare the values
of E
N
, r
N
for the solution of the same BVP (6.1)(6.2) for Shishkin mesh using
the standard upwind scheme adopted [13]. From these tables, it is obvious that
the method presented in this paper performs well. The numerical results are clear
illustrations of the convergence estimates derived in the present paper for both the
type of meshes.
166 A. R. BABU AND N. RAMANUJAM
Table 1. Values of E
N
and r
N
for the solution of the BVPs (6.1)-(6.2)
and (6.3)-(6.4) respectively.
N Shishkin mesh B-Shishkin mesh Shishkin mesh B-Shishkin mesh
E
N
r
N
E
N
r
N
E
N
r
N
E
N
r
N
32 7.5324e-03 0.0542 6.8353e-02 2.4161 4.5727e-03 0.0788 1.0557e-01 2.3448
64 7.2545e-03 0.4497 1.2807e-02 2.1880 4.3295e-03 0.1207 2.0782e-02 2.1671
128 5.3117e-03 0.9314 2.8106e-03 2.0923 3.9818e-03 0.5378 4.6273e-03 2.0866
256 2.7851e-03 1.2331 6.5912e-04 2.0460 2.7426e-03 1.1084 1.0894e-03 2.0447
512 1.1848e-03 1.2217 1.5961e-04 2.0230 1.2723e-03 1.1524 2.6405e-04 2.0226
1024 5.0802e-04 - 3.9272e-05 - 5.7236e-04 - 6.4986e-05 -
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
= 2
12
= 2
18
Figure 1. Graphs of
the rst component u
1
of
the Example 6.1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
= 2
12
= 2
18
Figure 2. Graphs of
the second component u
2
of the Example 6.1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
= 2
12
= 2
18
Figure 3. Graphs of
the rst component u
1
of
the Example 6.2
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
= 2
12
= 2
18
Figure 4. Graphs of
the rst component u
2
of
the Example 6.2
AN ALMOST SECOND ORDER FINITE ELEMENT METHOD 167
REFERENCES
[1] E.P. Doolan, J.J.H.Miller, W. H. A. Schilders, Uniform numerical methods for problems with
initial and boundary layers, Boole, Dublin, 1980.
[2] A. H. Nayfeh, Introduction to Perturbation Methods, Wiley, New York, 1981.
[3] R. E. OMalley, Singular perturbation methods for ordinary dierential equations, Springer, New
York, 1990.
[4] H. Han and R. B. Kellog, Dierentiability properties of solution of the equation
2
u + ru =
f(x, y) in a square, SIAM Journal of Math. Anal.,21(2): 394408, 1990.
[5] H-G. Roos, T. Lin, Sucient conditions for uniform convergence on layer-adapted grids, Com-
puting 63: 2745, 1999.
[6] P. A. Farrell, A. F. Hegarty, J.J.H.Miller, E. ORiordan, G. I. Shishkin, Robust computational
techniques for boundary layers, Chapman Hall/ CRC, Boca Raton, 2000.
[7] S. Matthews, Parameter robust numerical methods for a system of two coupled singularly per-
turbed reaction-diusion equations, Master Thesis, School of Mathematical Sciences, Dublin City
University, 2000.
[8] H-G. Roos, Helena Zarin, A second-order scheme for singularly perturbed dierential equations
with discontinuous source term,J.Numer.Math.,Vol.10,No.4: 275-289, 2002.
[9] N. Madden, M. Stynes, A uniformly convergent numerical method for a coupled system of two
singularly perturbed linear reaction-diusion problems, IMA. J. Numer. Anal.23(4): 627644,
2003.
[10] Helena Zarin, Finite element methods for singularly perturbed problems with special emphasis
on discontinuities, Ph.D., Thesis, Department of Mathematics and Informatics, University of
Novi sad, Novi sad, 2003.
[11] T. Lin, N. Madden, An improved error estimate for a system of coupled singularly perturbed
reaction-diusion equations, Comput. Methods Appl. Math. 3: 417423, 2003.
[12] T. Lin, N. Madden, A nite element anaysis of a coupled system of singularly perturbed
reaction-diusion equations, Appl. Math. Comput. 148(3): 869880, 2004.
[13] A. Tamilselvan, N. Ramanujam, V. Shanthi, A numerical method for singularly perturbed weakly
coupled system of two second order ordinary dierential equations with discontinuous source term,
Journal of Computational and Applied Mathematics 202: 203216, 2007.
[14] H-G. Roos, M. Stynes, L. Tobiska, Numerical methods for singularly perturbed dierential equa-
tions, Volume 24 of Springer series in Computational Mathematics, (2
nd
Edition) Springer-
Verlag, Berlin, 2008.