International Journal of Computer Mathematics

Vol. 82, No. 2, February 2005, 177–192

Parameter-uniform finite difference scheme for a system of

coupled singularly perturbed convection–diffusion equations
ZHONGDI CEN*
Institute of Mathematics, Zhejiang Wanli University, Ningbo 315400, P.R. China

(Revised 21 May 2004; in final form 24 June 2004)

In this paper, we discuss the parameter-uniform finite difference method for a coupled system of
singularly perturbed convection–diffusion equations. The leading term of each equation is multiplied
by a small but different magnitude positive parameter, which leads to the overlap and interact boundary
layer. We analyze the boundary layer and construct a piecewise-uniform mesh on the variant of the
Shishkin mesh. We prove that our schemes converge almost first-order uniformly with respect to small
parameters. We present some numerical experiments to support our theoretical analysis.

Keywords: Convection–diffusion; Singular perturbation; Solution decomposition; Shishkin mesh;

Upwind difference scheme

AMS Subject Classifications: 65L10; 65L12; 65L60

C.R. Category: G.1.7

1. Introduction

Singular perturbation problems arise in diverse areas of applied mathematics, including

linearized Navier–Stokes equation at high Reynolds number, heat transport problems with
large Peclet numbers, magneto-hydrodynamics duct problems at Hartman number and drift–
diffusion equation of semiconductor device modeling. For the past two decades, an extensive
research has been made on numerical methods for the singularly perturbed differential equa-
tions; see ref. [5–11] and reference there in. Robust techniques have been developed for a
single singularly perturbed differential equation, but for system of equations, only few results
are reported in the literature.
In this paper, we treat the following system of two coupled singularly perturbed convection–
diffusion problems:

L1 u
≡ −εu1 (x) − a1 (x)u1 (x) + b11 (x)u1 (x) + b12 (x)u2 (x) = f1 (x), (1)
L2 u
≡ −µu2 (x) − a2 (x)u2 (x) + b21 (x)u1 (x) + b22 (x)u2 (x) = f2 (x), (2)

*Email: czdningbo@tom.com

International Journal of Computer Mathematics

ISSN 0020-7160 print/ISSN 1029-0265 online © 2005 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/0020716042000301798
178 Zhongdi Cen

where u
= (u1 , u2 )T , x ∈  = (0, 1) with the boundary conditions:

u1 (0) = u2 (0) = u1 (1) = u2 (1) = 0. (3)

The parameters ε and µ ∈ (0, 1] are small positive constants. Without loss of generality, we
shall assume that
0 < ε ≤ µ  1. (4)
The solution u
of equations (1)–(3) has overlapping boundary layers of width O(ε ln ε) and
O(µ ln µ) at x = 0. We shall assume that B = {bij }2i,j =1 is an L0 matrix with

min {b11 (x) + b12 (x), b21 (x) + b22 (x)} > β > 0, (5)
x∈[0,1]

i.e., B is an M-matrix whose inverse is bounded by β −1 in the maximum norm. In addition,

we assume that
a1 (x) ≥ α > 0, a2 (x) ≥ α > 0. (6)
Matthews et al. [1,2] consider a system of singularly perturbed reaction–diffusion equations
for the cases (i) 0 < ε = µ  1 and (ii) 0 < ε  µ = 1. Linß and Madden [3] also consider
a finite element method for the system of singularly perturbed reaction–diffusion problems
for the case of 0 < ε ≤ µ < 1.
This present study is devoted to a finite difference method for the two coupled convection–
diffusion equations (1)–(3) on Shishkin-type meshes. We first prove bounds for u
and its
derivatives. These bounds will enable us to construct a special piecewis-uniform mesh on
which we can prove that the upwind difference scheme is almost first-order accurate, uniformly
in ε and µ. Our result is achieved by a hybrid stability inequality which shows that L∞ error
of the numerical solution is bounded by a discrete L1 norm of the consistency error.
An outline of the paper is as follows. In the next section, we derive a decomposition of
the exact solution and give bounds of the derivatives of the solution of equations (1)–(3). The
mesh and finite difference scheme that approximates the singularly perturbed problems are
constructed in Section 3. In Section 4, we show that the numerical solution computed by this
method is almost first-order accurate in the maximum norm, uniformly in the parameters ε
and µ. Finally, numerical results are presented in Section 5.

Notation: Throughout the paper, C will denote a generic positive constant (possibly sub-
scripted) that is independent of ε, µ and of the mesh. Note that C is not necessarily the same
at each occurrence. We define the maximum norm by

ψ(x) = max |ψ(x)|,

ψ(x) = max { ψi (x) },
¯
 1≤i≤m

ψ(x) = (ψ1 (x), ψ2 (x), . . . , ψm (x)).

We also set |ψ(x)| = (|ψ1 (x)|, |ψ2 (x)|, . . . , |ψm (x)|).

ASSUMPTION 1 Throughout the paper, we shall also assume that ε ≤ CN −1 and µ ≤ CN −1

as is generally the case.

2. Properties of the exact solution

In the following, the maximum principle for equations (1)–(3) is established. Then using this
principle, a stability result is derived.
 Convection–diffusion equations 179

THEOREM 1 (maximum principle) Assume that L1 y
≥ 0, L2 y
≥ 0 in  and y
(0) ≥ 0,

then y
(x) ≥ 0
in .
y
(1) ≥ 0, ¯

Proof Let y1 (p) = min¯ y1 (x) and y2 (q) = min¯ y2 (x). Assume without loss of generality
that y1 (p) ≤ y2 (q). If that the lemma is not true. Then y1 (p) < 0. Note that p = 0, 1 and
y1 (p) = 0, y1 (p) ≥ 0.

L1 y
(p) = −εy1 (p) − a1 (p)y1 (p) + b11 (p)y1 (p) + b12 (p)y2 (p)
= −εy1 (p) + (b11 (p) + b12 (p))y1 (p) + (y2 (p) − y1 (p))b12 (p) < 0

which contradicts the hypotheses of the lemma.

An immediate consequence is the following result.

LEMMA 1 (stability result) If y
(x) is a smooth function, then


y
(x) ≤ C max y
(0) , y
(1) , max |L1 y
|, max |L2 y
| for all x ∈ D̄.
¯
x∈ ¯
x∈

Now, we give bounds on the derivatives of the exact solution u
(x) for system (1)–(3). These
bounds will be used in the error analysis in later sections.

LEMMA 2 Let u
(x) be the solution of equations (1)–(3). Then

−k −k
|u(k)
1 (x)| ≤ Cε , |u(k)
2 (x)| ≤ Cµ , k = 1, 2

and

|u −1 −2
1 (x)| ≤ Cε (ε + µ−1 ), |u −1
2 (x)| ≤ Cµ (µ
−2 ¯
+ ε −1 ) for all x ∈ .

Proof By the mean value theorem, there exists a point z ∈ (0, 1) such that

u1 (ε) − u1 (0)
u1 (z) =
ε

and so |εu1 (z)| ≤ 2 u1 . Integrating the differential equation (1), we get

 z  z
|ε(u1 (z) − u1 (0))| = b11 (t)u1 (t)dt + b12 (t)u2 (t) dt
0 0
 z  z
− a1 (t)u1 (t)dt − f1 (t) dt
0 0
 z  z
= b11 (t)u1 (t)dt + b12 (t)u2 (t)dt − [a1 (t)u1 (t)]z0
0 0
 z  z
+ a1 (t)u1 (t)dt − f1 (t) dt (7)
0 0
180 Zhongdi Cen

It follows that |εu1 (0)| ≤ f1 (x) + C( u1 (x) + u2 (x) ). Using equation (7) with z = x,
we see that
|εu1 (x)| ≤ C( f1 (x) + u1 (x) + u2 (x) ) (8)
¯
for all x ∈ .
Similarly, we can get

|µu2 (x)| ≤ C( f2 (x) + u1 (x) + u2 (x) ) ¯

for all x ∈  (9)

Hence, we get the required result for k = 1 by equations (8) and (9). Now, from the differential
equations (1) and (2), we have

εu1 (x) = −a1 (x)u1 (x) + b11 (x)u1 (x) + b12 (x)u2 (x) − f1 (x),
µu2 (x) = −a2 (x)u2 (x) + b21 (x)u1 (x) + b22 (x)u2 (x) − f2 (x),

and

εu  
1 (x) = (−a1 (x)u1 (x) + b11 (x)u1 (x) + b12 (x)u2 (x) − f1 (x)) ,

µu  
2 (x) = (−a2 (x)u2 (x) + b21 (x)u1 (x) + b22 (x)u2 (x) − f2 (x)) ,

from which we obtain successively the required bounds on the second and third derivatives.

In order to prove that the numerical method is parameters uniform, more precise information
on the behavior of the exact solution of problem (1)–(3) is needed. This is obtained by writing
the solution in the form
u
(x) = v
(x) + w(x),

¯
x ∈ , (10)
where v
(x) is the solution of

L1 v
(x) = f1 (x), L2 v
(x) = f2 (x), x ∈ , (11)

v
(1) = u
(1), v1 (0) = v1,0 (0) + εµv1,1 (0), v2 (0) = v2,0 (0) + εµv2,1 (0), (12)

and w(x) satisfy

L1 w(x) = 0,

L2 w(x) = 0, x ∈ , (13)


w(1)

= 0,

w(0) = u
(0) − v
(0). (14)

Here, v1,i are v2,i (i = 0, 1) are defined by


− a1 (x)v1,0 (x) + b11 (x)v1,0 (x) + b12 (x)v2,0 (x) = f1 (x), (15)

− a2 (x)v2,0 (x) + b21 (x)v1,0 (x) + b22 (x)v2,0 (x) = f2 (x), (16)
v1,0 (1) = u1 (1), v2,0 (1) = u2 (1), (17)

and

− εv1,0 − a1 (x)(v1,0 + εµv1,1 ) + b11 (x)(v1,0 + εµv1,1 )
+ b12 (x)(v2,0 + εµv2,1 ) = f1 (x), (18)

− µv2,0 − a2 (x)(v2,0 + εµv2,1 ) + b21 (x)(v1,0 + εµv1,1 )
+ b22 (x)(v2,0 + εµv2,1 ) = f2 (x), (19)
v1,1 (1) = v2,1 (1) = 0. (20)
 Convection–diffusion equations 181

LEMMA 3 There exists a constant C such that

v
(k) (x) ≤ C, ¯
k = 1, 2, 3, x ∈ .

Proof Note that v1,0 , v2,0 and their derivatives are independent of ε and µ, and

− εv1,0 (x) − a1 (x)(v1,0 (x) + εµv1,1 (x)) + b11 (x)(v1,0 (x) + εµv1,1 (x))
+ b12 (x)(v2,0 (x) + εµv2,1 (x)) = f1 (x)

= −a1 (x)v1,0 (x) + b11 (x)v1,0 (x) + b21 (x)v2,0 (x)

and

− µv2,0 (x) − a2 (x)(v2,0 (x) + εµv2,1 (x)) + b21 (x)(v1,0 (x) + εµv1,1 (x))
+ b22 (x)(v2,0 (x) + εµv2,1 (x)) = f2 (x)

= −a2 (x)v2,0 (x) + b21 (x)v1,0 (x) + b22 (x)v2,0 (x).

Hence,

 1 
− a1 (x)v1,1 (x) + b11 (x)v1,1 (x) + b12 (x)v2,1 (x) = v (x), (21)
µ 1,0
 1 
− a2 (x)v2,1 (x) + b21 (x)v1,1 (x) + b22 (x)v2,1 (x) = v2,0 (0) (22)
ε
Therefore, we have

(k) 1 (k) 1
|v1,1 (x)| ≤ C , |v2,1 (x)| ≤ C , k = 0, 1. (23)
µ ε
Differentiating the equations (21) and (22), we also obtain
   
(k) 1 1 (k) 1 1
|v1,1 (x)| ≤ C + , |v2,1 (x)| ≤ C + , k = 2, 3. (24)
ε µ ε µ
We can write

v1 = v1,0 + εµv1,1 + ε 2 µ2 v1,2 and v2 = v2,0 + εµv2,1 + ε 2 µ2 v2,2 ,

where v1,2 and v2,2 are the solutions of

     
− ε(v1,0 + εµv1,1 + ε 2 µ2 v1,2 ) − a1 (x)(v1,0 + εµv1,1 + ε 2 µ2 v1,2 )
+ b11 (x)(v1,0 + εµv1,1 + ε 2 µ2 v1,2 ) + b12 (x)(v2,0 + εµv2,1 + ε 2 µ2 v2,2 ) = f1 (x),
     
− µ(v2,0 + εµv2,1 + ε 2 µ2 v2,2 ) − a2 (x)(v2,0 + εµv2,1 + ε 2 µ2 v2,2 )
+ b21 (x)(v1,0 + εµv1,1 + ε 2 µ2 v1,2 ) + b22 (x)(v2,0 + εµv2,1 + ε 2 µ2 v2,2 ) = f2 (x).

Hence, we have

  1 
− v1,2 (x) − a1 (x)v1,2 (x) + b11 (x)v1,2 (x) + b12 (x)v2,2 (x) = v (x), (25)
µ 1,1
  1 
− v2,2 (x) − a2 (x)v2,2 (x) + b21 (x)v1,2 (x) + b22 (x)v2,2 (x) = v2,1 (x). (26)
ε
182 Zhongdi Cen

Then following the proof of Lemma 2 and equation (24), we can get
     
(k) 1 1 1 (k) 1 1 1
|v1,2 (k)| ≤ C 1 + + , |v2,2 (k)| ≤ C 1 + + , k = 0, 1, 2, 3.
ε ε µ ε ε µ
(27)
From equations (23), (24) and (27), we obtain the desired result.

We now find bounds on the layer part w(x) of the Shishkin decompositions of u
(x).

LEMMA 4 Let w(x)
= (w1 (x), w2 (x))T be the solution to equations (13) and (14). Then there
¯ we have
exists a constant C such that for all x ∈ ,
−αx −αx
|w1 (x)| ≤ C exp , |w2 (x)| ≤ C exp , (28)
µ µ
 
−αx −αx −αx
|w1 (x)| ≤ C ε −1 exp + µ−1 exp , |w2 (x)| ≤ Cµ−1 exp , (29)
ε µ µ
 
 −2 −αx −2 −αx −αx
|w1 (x)| ≤ C ε exp + µ exp , |w2 (x)| ≤ Cµ−2 exp , (30)
ε µ µ
 
−αx −αx
|w1 (x)| ≤ C ε −3 exp + µ−3 exp , (31)
ε µ
 
 −1 −2 −αx −2 −αx
|w2 (x)| ≤ Cµ ε exp + µ exp . (32)
ε µ

Proof By equation (14) and Lemma 3, we have w

≤ C on ∂. Define the function ψ(x) =


(C exp(−αx/µ), C exp(−αx/µ))T with C chosen sufficiently large, so that ψ(x) ≥ |w(x)|


on ∂ ; as
 
−αx α ε α2


L1 ψ(x) = C exp a11 (x) + a12 (x) + a1 (x) −

≥ 0 = L1 w(x),
µ µ µ µ
 
−αx α α2


L2 ψ(x) = C exp a21 (x) + a22 (x) + a2 (x) −

≥ 0 = L2 w(x)
µ µ µ

on , the maximum principle of Theorem 1 implies that ψ(x) ≥ |w(x)|

¯ and equation
on ,
(28) is proved.
The second equation of equation (13) is

µ(w2 (x)) + a2 (x)w2 (x) = b21 (x)w1 (x) + b22 (x)w2 (x) ≡ h2 (x). (33)

Let A2 be an infinite integral of a2 (x). Then

 x
w2 (x) = w2 (0) exp(−µ−1 (A2 (x) − A2 (0))) − µ−1 h2 (t) exp(−µ−1 (A2 (x) − A2 (t)))dt.
0
(34)
From Lemmas 2 and 3, we have
|w2 (0)| ≤ Cµ−1 . (35)
Using the inequality
−α(x − t)
exp(−µ−1 (A2 (x) − A2 (t))) ≤ exp , x≥t (36)
µ
 Convection–diffusion equations 183

and equation (28), we can obtain


−αx x
−αt −α(x − t)
|w2 (x)| = Cµ −1
exp + Cµ−1 exp exp dt. (37)
µ 0 µ µ
Hence,
−αx
|w2 (x)| ≤ Cµ−1 exp . (38)
µ
Substituting this inequality and equation (28) into equation (33) yields
−αx
|w2 (x)| ≤ Cµ−2 exp . (39)
µ
Applying the same techniques to the first equation of equation (13), we obtain
 x
−αx −αt −α(x − t)
|w1 (x)| ≤ Cε −1 exp + Cε −1 exp exp dt (40)
ε 0 µ ε
 
−αx −αx
≤ C ε−1 exp + µ−1 exp , (41)
ε µ
where we have used integration by parts. This inequality and equation (38) constitute
equation (29). Substituting this inequality into equation (13) yields
 
−αx −αx

|w1 (x)| ≤ Cε −1 −1
ε exp −1
+ µ exp ¯
for x ∈ .
ε µ

This bound is satisfactory on ∂ but is not sharp inside , if ε  µ, then exp(−αx/µ)

decays much more slowly than exp(−αx/ε). To improve it, differentiate the first equation of
equation (13), which yields

− ε(w1 (x)) − a1 (x)w1 (x) = a1 (x)w1 (x) + b11 (x)w1 (x)

+ b11 (x)w1 (x) + b12 (x)w2 (x) + b12

(x)w2 (x) ≡ h1 (x). (42)

Using the previous bounds, we can get

 
−1 −αx −1 −αx
|h1 (x)| ≤ C ε exp + µ exp . (43)
ε µ

Let A1 be an indefinite integral of a1 (x). Then


−(A1 (x) − A1 (0)) x
−(A1 (x) − A1 (t))
w1 (x) = w1 (0) exp − ε −1 h1 (t) exp dt.
ε 0 ε

From equations (36) and (43) and |w1 (x)| ≤ Cε−2 , we have
 x 
−αx −αx −αx −α(x − t)
|w1 (x)| ≤ Cε −2 exp + Cε −1 ε −1 exp + µ−1 exp exp dt
ε 0 ε µ ε
 
−2 −αx −2 −αx
≤ C ε exp + µ exp ,
ε µ
where we also have used integration by parts. This inequality and equation (39) show that
equation (30) holds true.
184 Zhongdi Cen

Differentiating the second equation of equation (13) gives

µw2 (x) = (−a1 (x)w2 (x) + b21 (x)w1 (x) + b22 (x)w2 (x)) .

The bounds (28)–(30) now imply that

 
−αx −αx
|w2 (x)| ≤ Cµ−1 ε −2 exp + µ−2 exp ¯
on .
ε µ
 
Finally, to bound |w1 (x)|, the argument resembles the one used to bound |w1 (x)|. Hence,
we obtain  
−αx −αx

|w1 (x)| ≤ C ε−3 exp + µ−3 exp ¯
on .
ε µ

3. Mesh and scheme

Shishkin meshes for a single convection–diffusion two-point boundary value problem divide
the interval [0, 1] into two subintervals, on each of them the mesh is uniform. These subintervals
correspond to the region where the solution is smooth and to the boundary layer at x = 0. When
0 < ε < µ  1, the solution to equations (1)–(3) has a overlapping boundary layer at x = 0.
This necessitates the construction of a mesh that is uniform on each of three subintervals.
On the main subinterval, where the solution is smooth, the mesh is coarse; on the other two
subintervals, it is very fine.
We define τε and τµ as



1 2µ 1 τµ 2ε
τµ = min , ln N and τε = min , , ln N .
2 α 4 2 α

A piecewise-uniform mesh  ¯ N = {xi }N

0 is constructed by dividing [0, 1] into three subintervals
[0, τε ], [τε , τµ ] and [τµ , 1]. Then subdivide [τµ , 1] into N/2 mesh intervals, and subdivide each
of the other two subintervals into N/4 mesh intervals.
When τε = τµ /2, then µ = O(ε), and the result can be easily obtained. Therefore, we only
consider the case τε < τµ /2.
The following lemma gives some estimates of the mesh sizes that will be used later.

LEMMA 5 The step sizes of the mesh N satisfy


 4τε N

 H1 = 1≤1≤ ,

 N 4

4(τµ − τε ) N N
hi = H2 = <i≤ ,

 N 4 2



H3 = 2(1 − τµ ) N
< i ≤ N.
N 2
Proof This can be easily computed.

Based on this mesh, we propose the upwind finite difference scheme for problem (1)–(3).
+ N
LN
N
1 u i = −εδ u1,i − a1,i D u1,i + b11,i u1,i + b12,i u2,i = f1,i ,
2 N N N
(44)
 Convection–diffusion equations 185

+ N
LN
N
2 u i = −µδ u2,i − a2,i D u2,i + b21,i u1,i + b22,i u2,i = f2,i ,
2 N N N
(45)

1,0 = u1,N = u2,0 = u2,N = 0,

uN N N N
(46)

where
zi+1 − zi zi − zi−1 (D + − D − )zi
D + zi = , D − zi = and δ 2 zi = .
hi+1 hi hi+1

4. Analysis of the method

We show now that the difference operator satisfies a discrete maximum principle.

THEOREM 2 (discrete maximum principle) Let {
zi }N i=0 be any vector mesh function such that
z
0 ≥ 0
and z
N ≥ 0.

Suppose that LN
1

z i ≥ 0 and L N

i ≥ 0 for 1 ≤ i ≤ N − 1. Then z
i ≥ 0

2 z
for all 0 ≤ i ≤ N .

Proof Define
γ = max{ max (−z1,i ), max (−z2,i )}.
0≤i≤N 0≤i≤N

Assume that the conclusion of the lemma is false. Then γ > 0, and we have

z1,i + γ ≥ 0, z2,i + γ ≥ 0, i = 0, 1, . . . , N.

Case (i): Let z1,i + γ = 0 for some i = k. Then 0 < LN z + γ s
)k = −εδ 2 (z1,k +
1 (

+
γ ) − a1,i D (z1,k + γ ) + b11,k (z1,k + γ ) + b12,k (z2,k + γ ) < 0, where s
= (1, 1)T . It is
contraction.
Case (ii): Let z2,k + γ = 0 for some i = k. Similarly, we get a contraction.
Hence, we complete the proof.

An immediate consequence of this discrete maximum principle is the following stability

result.

LEMMA 6 If {
zi }N
i=0 be any vector mesh function, then

z
≤ C max z
0 , z
N , |LN
i |, |LN

1 z
i | .
2 z

LEMMA 7 For i = 0, 1, . . . , N, define the mesh function

i  −1 i  −1
αhj αhj
Sε,i = 1+ , Sµ,i = 1+
j =1
ε j =1
µ

with the usual convention that if i = 0, then Sε,0 = Sµ,0 = 1. If i = 1, 2, . . . , N − 1, then

we have

CSε,i
CSε,i
LN1 Sε,i ≥ , LN2 Sε,i ≥
max{ε, hi } max{ε, hi }
and

CSµ,i
CSµ,i
1 Sµ,i ≥
LN , 2 Sµ,i ≥
LN
max{µ, hi } max{µ, hi }
for some constant C, where S
ε,i = (Sε,i , Sε,i )T and S
µ,i = (Sµ,i , Sµ,i )T .
186 Zhongdi Cen

Proof This is an easy calculation [4].

We now prove a fundamental stability result for LN = (LN N

1 , L2 ).
 
AN B N
LEMMA 8 Write M2N = for the (2N + 2) × (2N + 2) matrix of the differ-
CN DN
ence scheme LN , where the boundary conditions are handled by (AN )0,0 = (AN )N,N =
1, (DN )0,0 = (DN )N,N = 1 and (AN )0,i = (AN )N,i−1 = (DN )0,i = (DN )N,i−1 = 0 for i =
1, . . . , N and (BN )0,i = (BN )N,i = (CN )0,i = (CN )N,i = 0 for i = 0, 1, . . . , N. Then for any
(N + 1)-dimensional row vectors v1 = (0, v11 , v12 , . . . , v1N −1 , 0) and v2 = (0, v21 , v22 , . . . ,
v2N −1 , 0), we have
N −1
max{ v1 , v2 } ≤ C hj +1 (|(AN v1T + BN v2T )j | + |(CN v1T + DN v2T )j |),
j =i

where T denotes transpose.

Proof For each j ∈ {1, . . . , N − 1}, suppose that we can construct column vectors
(j +N +1) (j +N +1) (j +N +1)
y (j ) = (y10 , . . . , y1N , y20 , . . . , y2N ) and y (j +N +1) = (y10
(j ) (j ) (j ) (j )
, . . . , y1N , y20 ,...,
(j +N +1)
y2N ) such that
(j ) (j ) (j +N +1) (j +N +1)
|y1i | ≤ Chj +1 , |y2i | ≤ Chj +1 , |y1i | ≤ Chj +1 , |y2i | ≤ Chj +1 , (47)

for all i and

M2N y (j ) ≥ (0, 0, . . . , 0, 1, 0, . . . , 0)T , (48)
where the 1 is located in the j th position and

M2N y (j +N +1) ≥ (0, 0, . . . , 0, 1, 0, . . . , 0)T , (49)

−1
where the 1 is located in the j + N + 1th position. As M2N is an M-matrix, M2N ≥ 0. Hence,
using equation (47), for each i, we have
(j ) −1 −1
yki = (M2N M2N y (j ) )i+(k−1)(N +1) ≥ (M2N (0, . . . , 1, 0, 0, . . . , 0)T )i+(k−1)(N +1),j
−1
= (M2N )i+(k−1)(N +1),j (50)

and
(j +N +1) −1 −1
yki = (M2N M2N y (j +N +1) )i+(k−1)(N +1) ≥ (M2N (0, . . . , 0, 1, 0, . . . , 0)T )i+(k−1)(N +1)
−1
= (M2N )i+(k−1)(N +1),j +N +1 (51)

for k = 1, 2. If v = (0, v11 , . . . , v1,N −1 , 0, 0, v11 , . . . , v1N −1 , 0), then for each i, we have
2N +2 
 
−1  −1 
|vki | = |(M2N M2N v T )i+(k−1)(N +1) | =  (M2N )i+(k−1)(N +1),s (M2N v T )s 
 
s=1
N −1
≤C hj +1 (|(AN v1T + BN v2T )j | + |(CN v1T + DN v2T )j |), (52)
j =1

−1
where we used v10 = v1N = v20 = v2N = 0, M2N ≥ 0, and equations (47), (50) and (51).
 Convection–diffusion equations 187

Thus, it remains only to construct column vectors y (j ) and y (j +N +1) with the three properties
stated earlier.
(j ) (j ) (j ) (j )
Define y (j ) = (y10 , . . . , y1N , y20 , . . . , y2N ) by

 hj +1

 α i = 0, 1, . . . , j

(j )
yki = i  −1 (53)


hj +1
1+
αhk
i = j + 1, . . . , N

 α
k=j +1
ε

for k = 1, 2.
(j )
Clearly, we have |yki | ≤ Chj +1 for all j . To verify property (48), we first note that
(j ) (j )
(M2N y (j ) )0 = y10 > 0 and (M2N y (j ) )N = y1N > 0. Next, we divide the argument into three
cases as follows:
When 0 < i < j , clearly (M2N y (j ) )i = 0.
When j < i < N, a direct calculation yields
j  
hj +1 αhk
N

(M2N y (j )
)i = 1+ 1 Sε,i + L2 Sε,i ) > 0
(LN
α k=1
ε

by Lemma 7.
When for i = j , we get
   
ε hj +1 α µ hj +1 α
(M2N y (j ) )i = +α + +α
hj +1 α ε + αhj +1 hj +1 α ε + αhj +1
hj +1
+ (b11,j + b12,j + b21,j + b22,j ) > 1.
α

Similarly, we can define y (j +N +1) by




hj +1
i = 0, 1, . . . , j

 α
= hj +1 i  
(j +N +1)
yki
 αhk −1 (54)

 1+ i = j + 1, . . . , N
 α µ
k=j +1

for k = 1, 2.

Let r
N denote the truncation error of our scheme, i.e.,

N
r1,i = LN
− f1,i ,
1 u
N
r2,i = LN
− f2,i .
2 u

Estimates for the truncation error are provided by the following lemma.

LEMMA 9 Assume that the conclusions of Lemmas 3 and 4 hold true. Then we have
 
−αxi−1 −αxi−1
hi+1 |r1,i | ≤ C hi+1 + ε + exp
N
+ µ + exp , (55)
ε µ
 
−αxi−1
hi+1 |r2,i
N
| ≤ C hi+1 + µ + exp (56)
µ
188 Zhongdi Cen

and

 
−1 −αxi −αxi
hi+1 |r1,i
N
| ≤ C |hi+1 − hi | 1 + ε exp + µ−1 exp
ε µ
 
−αxi−1 −αxi−1
+(h2i + h2i+1 ) 1 + ε −2 exp + µ−2 exp , (57)
ε µ

 
−1 −αxi
hi+1 |r2,i | ≤ C |hi+1 − hi | 1 + µ exp
N
µ
 
−αxi−1
+(h2i + h2i+1 ) 1 + µ−2 exp , (58)
µ

for i = 1, 2, . . . , N − 1.

Proof Using the results given in Lemmas 3 and 4 and applying the arguments given in
[9, Lemma 3.3], we can get the desired results.

Now we can get our main result for our difference scheme.

THEOREM 3 Let u
(x) be the solution to equations (1)–(3) and u
N be the solution to the
discrete problem (44)–(46) on the mesh N defined in Section 3. Then there exists a constant C
such that
u
− u
N ≤ CN −1 ln N.

Proof Recall Lemma 9, it is sufficient to show that

N −1
hi+1 (|r1N | + |r2N |) ≤ CN −1 ln N.
i=1

First, we state some properties of the mesh. Lemma 5 implies that

N N N
hi ≤ Cε for i = 1, 2, . . . , and hi ≤ Cµ for i = + 1, . . . ,
4 4 2

and consequently

αhi N αhi N N
exp ≤ C for i = 1, 2, . . . , and exp ≤ C for i = + 1, . . . , .
ε 4 µ 4 2

First, we estimate hi+1 |r1N |. We apply equation (50) to bound hi+1 |r1,i
N
| for i = N/2,
N/2 + 1 and equation (51) otherwise.

N −1 N/2−1
 
−αxi −1−1 −αxi
hi+1 |r1,i
N
| ≤C (hi+1 − hi ) 1 + ε exp + µ exp
i=1 i=1
ε µ
 
−2 −αxi−1 −2 −αxi−1
+(hi + hi+1 ) 1 + ε exp
2 2
+ µ exp
ε µ
 Convection–diffusion equations 189


−αxN/2−1 −αxN/2 −αxN/2−1
+ C N −1 + ε + exp + exp + µ + exp
ε ε µ
 N −1 
−αxN/2 −αxi−1
+ exp +C N −2 1 + ε −2 exp
µ i=N/2+2
ε

−αxi−1
+ µ−2 exp (59)
µ

We bound the terms on the right-hand side separately in reverse order.

For i = N/2 + 2, . . . , N − 1, we have

−αxi−1 −αH3 −αxN/2

ε −2 exp ≤ ε−2 exp exp ≤ C,
ε ε ε
−αxi−1 −αH3 −αxN/2
µ−2 exp ≤ µ−2 exp exp ≤ C.
µ µ µ

Thus,
N −1  
−2 −2 −αxi−1 −αxi
N 1+ε exp + µ−2 exp ≤ CN −1 . (60)
i=N/2+2
ε µ

Furthermore, we have
−αxN/2−1 −αxN/2 −αxN/2−1
N −1 + ε + exp + exp + µ + exp
ε ε µ
 
−αxN/2 αhN/2 −αxN/2
+ exp ≤ CN −1 + C 1 + exp exp ≤ CN −1 . (61)
µ µ µ

Next, we bound the first sum in equation (54). Clearly, we have

N/2−1
(hi+1 − hi + h2i + h2i+1 ) ≤ CN −1 (62)
i=1

Using the inequality

−αt −αt kε
ε−k exp ≤ µ−k exp for t > , k = 1, 2, . . . , (63)
ε µ α
N/2−1  
−1 −αxi −1 −αxi
(hi+1 − hi ) ε exp + µ exp
i=1
ε µ
N/4−1  
−αxi −αxi
≤ (hi+1 − hi ) ε−1 exp + µ−1 exp
i=1
ε µ
N/2−1  
−1 −αxi −αxi
+ (hi+1 − hi ) ε exp + µ−1 exp
i=N/4
ε µ
N/4−1 N/2−1
−1 −1
≤ (hi+1 − hi )(ε +µ )+ (hi+1 − hi )2µ−1 ≤ CN −1 ln N. (64)
i=1 i=N/4
190 Zhongdi Cen

Finally, we have
N/4−1  
−2 −αxi−1 −αxi−1
(h2i+1 + h2i ) ε exp + µ−2 exp
i=1
ε µ
 N/4−1 N/4−1

−2 −αxi−1 −αxi−1
= 2H12 ε exp + µ−2 exp
i=1
ε i=1
µ
 
−21 −2 1
≤ 2H12
ε +µ
1 − exp(−αH1 /ε) 1 − exp(−αH1 /µ)
     
αH1 ε αH1 µ
≤ 2H12 ε −2 exp + µ−2 exp ≤ CN −1 ln N. (65)
ε αH1 µ αH1
Applying equation (58) and the calculation mentioned earlier also yield
N/2−1  
−αxi−1 −αxi−1
(h2i+1 + h2i ) ε−2 exp + µ−2 exp
i=N/4
ε µ
N/2−1
−αxi−1
≤ (h2i+1 + h2i )2µ−2 exp ≤ CN −1 ln N. (66)
i=N/4
µ

Substituting equations (55)–(61) into equation (54), we get

N −1
hi+1 |r1,i
N
| ≤ CN −1 ln N. (67)
i=1

Similarly, we can get the following estimate

N −1
hi+1 |r2,i
N
| ≤ CN −1 ln N. (68)
i=1

From equations (62) and (63), we complete the proof.

5. Numerical results

In this section, we present some numerical experiments. We study the test problem

− εu1  (x) − u1 (x) + 2u1 (x) − u2 (x) = f1 (x), (69)

Table 1. ε = 10−8 and µ = 10−1 .

N Error Rate Constant

64 1.4079e−1 0.931 2.1666

128 7.3838e−2 0.930 1.9479
256 3.8765e−2 0.932 1.7896
512 2.0320e−2 0.935 1.6678
1024 1.0630e−2 – 1.5703
 Convection–diffusion equations 191

Table 2. ε = 10−8 and µ = 10−4 .

N Error Rate Constant

64 3.5106e−1 1.010 5.4024

128 1.7431e−1 0.857 4.5983
256 9.6253e−2 0.837 4.4436
512 5.3897e−2 0.846 4.4235
1024 2.9989e−2 – 4.4303

Table 3. ε = 10−8 and µ = 10−8 .

N Error Rate Constant

64 1.6827e−1 0.827 2.5895

128 9.4853e−2 0.719 2.5023
256 5.7606e−2 0.783 2.6595
512 3.3476e−2 0.818 2.7478
1024 1.8985e−2 – 2.8047

− µu2  (x) − 2u2 (x) + 4u2 (x) − u1 (x) = f2 (x), (70)

u1 (0) = u1 (1) = u2 (0) = u2 (1) = 0, (71)

where f1 (x), f2 (x) are chosen such that

1 − exp(−x/ε) 1 − exp(−x/µ) π
u1 (x) = + − 2 sin x,
1 − exp(−1/ε) 1 − exp(−1/µ) 2
1 − exp(−x/µ)
u2 (x) = − x exp(x − 1).
1 − exp(−1/µ)

For our tests, we take ε = 10−8 and µ = 10−1 , 10−4 and 10−8 , respectively, which are suffi-
ciently small choices to bring out the singularly perturbed nature of the problem (Tables 1– 3).
We measure the accuracy in the discrete maximum norm · . The rates of convergence r N are
computed using the following formula:

 
u
− u
N
r N = log2 .
u
− u
2N

We also compute the constants in the error estimate, i.e., if we have the theoretical error
bound

u
− u
N ≤ Cλ(N ),

then we compute the quantity

u
− u
N
CN = .
λ(N )

The numerical results are clear illustrations of the convergence estimate of Theorem 3. They
indicate that the theoretical results are fairly sharp.
192 Zhongdi Cen

Acknowledgement

The work was supported by National Natural Science Foundation (Grant Nos 10301029 and
10241003) of China.

References
[1] Matthews, S., Miller, J.J.H., O’Riordan, E., Shishkin, G.I., 2000, In: J.J.H. Miller, G.I. Shishkin, and L. Vulkov
(Eds) Analytical and Numerical Methods for Convection-Dominated and Singularly Perturbed Problems,
(New York: Nova Science Publishers).
[2] Matthews, S., O’Riordan, R., and Shishkin, G.I., 2000, Numerical Methods for a System of Singularly Perturbed
Reaction-Diffusion Equations (Dublin City University), Preprint MS-00-06.
[3] Linß, T. and Madden, N., 2004, Applied Mathematics and Computation, 148, 869–880.
[4] Kellogg, R.B. and Tsan, A., 1978, Mathematics of Computation, 32, 1025–1039.
[5] Roos, H.-G. and Linß, T., 1999, Computing, 63, 27–45.
[6] Farrell, P.A., O’Riordan, E., Miller, J.J.H., Shishkin, G.I., 2001, Computational Mathematics and Applied
Mathematics, 1, 154–172.
[7] Stynes, M., and Roos, H.-G., 1997, Applied Numerical Mathematics, 23, 361–374.
[8] Linß, T., 1999, Journal of Computational and Applied Mathematics, 110, 93–104.
[9] Linß, T., Roos, H.-G. and Vulanovi’c, R., 2000, SIAM. Journal of Numerical Analysis, 38, 897–912.
[10] Stynes, M. and Tobiska, L., 1998, Numerical Algrorithms, 18, 337–360.
[11] Kadalbajoo, M.K., and Patidar, K.C., 2002, Applied Mathematics and Computation, 130, 457–510.