Applied Numerical Mathematics 51 (2004) 171–186

[Link]/locate/apnum

A parameter robust numerical method for a system

of two singularly perturbed convection–diffusion equations ✩
S. Bellew a , E. O’Riordan b,∗
a Institute of Technology Dundalk, Co., Louth, Ireland
b School of Mathematical Sciences, Dublin City University, Dublin 9, Ireland

Available online 2 July 2004

Abstract
In this paper a coupled system of two singularly perturbed convection–diffusion ordinary differential equations
is examined. A numerical method is constructed for this system which involves an appropriate piecewise-uniform
Shishkin mesh. The numerical approximations are shown to converge to the continuous solutions uniformly with
respect to the two singular perturbation parameters. Numerical results are presented which illustrate the theoretical
results.
 2004 IMACS. Published by Elsevier B.V. All rights reserved.

Keywords: System; Singularly perturbed; Convection–diffusion; Shishkin mesh

1. Introduction

The asymptotic nature of solutions to systems of singularly perturbed second order differential
equations has been examined by several authors (for example, see [4,17] and the references therein). In
many papers, the coupling is assumed to only occur through the zero-order terms in the system. Within
the literature on the asymptotic solution of singularly perturbed vector problems of second order, there are
few papers dealing with strongly coupled systems where the first derivatives are coupled. In the strongly
coupled case, the existence of solutions for a singularly perturbed quasilinear system has been studied
in [3,5] for sufficiently small values of ε (the singular perturbation parameter). Asymptotic analysis can
describe the layer structure of the solution for small values of ε. This asymptotic information may be
employed in the design of appropriate numerical methods for the problem. In this paper, we examine

✩
This research was supported by the Albert College Fellowship scheme of Dublin City University.
* Corresponding author.
E-mail addresses: [Link]@[Link] (S. Bellew), [Link]@[Link] (E. O’Riordan).

0168-9274/$30.00  2004 IMACS. Published by Elsevier B.V. All rights reserved.

doi:10.1016/[Link].2004.05.006
172 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

numerical methods that are shown to be parameter-uniform [11,6]. These are numerical methods whose
numerical approximations U N satisfy error bounds of the form




uε − U N

CN −p , p > 0

where uε is the solution of the continuous problem, N (independent of ε) is the dimension of the discrete
problem and C is a constant which is independent of both N and ε. In other words, the numerical
approximations converge for all values of ε (from the sufficiently small values of ε to values of order
one).
The design and analysis of appropriate numerical methods for singularly perturbed differential
equations is an area of current interest (see [6,7,11–13] and the references therein). The advent of
Shishkin meshes [14,15] has fuelled significant advances into the broad area of singularly perturbed
differential equations. However, there has been little work done on parameter-robust numerical methods
for systems of singularly perturbed differential equations. This paper applies the analytical tools
developed for a single singularly perturbed differential equation to a system of two singularly perturbed
ordinary differential equations.
Consider the following system of singularly perturbed ordinary differential equations:

ε1 u1 (x) + a1,1 (x)u1 (x) = f1 (x), x ∈ Ω, (1.1a)

ε2 u2 (x) + a2,2 (x)u2 (x) + a2,1 (x)u1 (x) = f2 (x), x ∈ Ω, (1.1b)
a1,1 (x)  α1 > 0, a2,2 (x)  α2 > 0, (1.1c)
u1 (0), u1 (1), u2 (0), u2 (1) are given constants. (1.1d)

The functions a1,1 , a2,1 , a2,2 , f1 , f2 ∈ C 3 (Ω), Ω = (0, 1) and 0 < ε1
1 and 0 < ε2
1. When a2,1 (x) =
0, the system (1.1) is a coupled system of convection–diffusion problems. There are two scales in this
system associated with the small parameters ε1 and ε2 . In general, there will be a boundary layer of width
O(ε1 ) near x = 0 in the solution u1 . The nature of the layers present in the solution u2 depends on the
relative values of the two small parameters and will be the main focus in this paper.
It should be noted that in order to determine an analytic approximation to the solution of the above
system, one could study the problem
 (i)   
ε2 u (x) + a(x)u (x) = f (x, ε1 ), f (x, ε1 )
C 1 + ε −(1+i) e−α1 x/ε1 .
1

However, our interest lies in examining the convergence of numerical approximations to the solutions of
the system (1.1), which are generated by discretizing the system directly. In this way, we examine the
effect the discretization error in the first equation has on the discretization error of the second equation.
We will discretize (1.1) using a numerical method, which is composed of an upwind finite difference
operator on a piecewise-uniform Shishkin mesh (see, for example, [6,11] for motivation for this type
of mesh). The mesh is designed to simultaneously resolve the layers in both u1 and u2 . The mesh is
composed of two transition points σ1 , σ2 , which partition the domain Ω into three subdomains. The
mesh points are uniformly spaced within each subdomain. The transition points are defined to be
       
ε1 ε2 σ2 ε1 ε2 1
σ1 = min 2 min , ln N, , σ2 = min 4 max , ln N, , (1.2)
γ1 γ2 2 γ1 γ2 2
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 173

where γi < αi , i = 1, 2. The piecewise-uniform mesh ΩεN1 ,ε2 is defined as


 iH1 if i
N/4,

Ωε1 ,ε2 = {xi }, xi = σ1 + (i − N/4)H2 if N/4 < i
N/2,
N
(1.3)

σ2 + (i − N/2)H3 if i > N/2,

where H1 = 4σN1 , H2 = 4(σ2N−σ1 ) , H3 = 2(1−σ

N
2)
, hi = xi − xi−1 and h̄i = (hi+1 + hi )/2. The discrete system
of equations corresponding to system (1.1) is:

ε1 δ 2 U1 (xi ) + a1,1 (xi )D + U1 (xi ) = f1 (xi ), xi ∈ ΩεN1 ,ε2 , (1.4a)

ε2 δ 2 U2 (xi ) + a2,2 (xi )D + U2 (xi ) + a2,1 (xi )D + U1 (xi ) = f2 (xi ), xi ∈ ΩεN1 ,ε2 , (1.4b)
Uj (0) = uj (0), Uj (1) = uj (1), j = 1, 2,

where the finite difference operators D + , D − , δ 2 are the standard finite difference approximations to the
first and second derivatives. That is, for any mesh function Y ,

Y (xi+1 ) − Y (xi ) Y (xi ) − Y (xi−1 )

D + Y (xi ) = , D − Yi = ,
xi+1 − xi xi − xi−1
D + Y (xi ) − D − Y (xi )
δ 2 Yi = 2 .
xi+1 − xi−1
This paper will establish an asymptotic error bound for the numerical approximations {U1 , U2 } of (1.4)
to the solutions {u1 , u2 } of (1.1) of the form



max
U i (x) − ui (x)
Ω
CN −1 (ln N)2 ,
i=1,2

where φD = maxx∈D⊂Ω |φ(x)| is the standard pointwise maximum norm, Ui is the piecewise linear
interpolant of Ui on Ω  and C denotes a generic constant which is independent of ε1 , ε2 and N . We say
that a numerical method is parameter-robust [6] if the numerical approximations converge independently
of the singular perturbation parameters.
Parameter-robust numerical methods for a system of singularly perturbed reaction–diffusion ordinary
differential equations have been examined by [16,9,10]. The analysis in these papers relied on the
availability of a standard comparison principle for the system. Let y = (y1 , y2 ) and
Ly = (−ε1 y1 +

a11 y1 + a12 y2 , −ε2 y2 + a21 y1 + a22 y2 ), aii > 0, a12 < 0, a21 < 0. Then if y(0)  0, y(1)  0 and

Ly(x)  0, x ∈ Ω then y(x)  0, x ∈ Ω. The corresponding standard comparison principle does not
apply to the system (1.1) of convection–diffusion equations. Moreover, in the system of reaction–
diffusion equations the coupling arises via the solutions of the components y1 , y2 , while in the system
of convection–diffusion equations the coupling is via a derivative of the component u1 . Within the
framework of singularly perturbed problems, coupling via the derivatives is significantly stronger than
coupling via the solutions.

Note. Throughout this paper, C denotes a generic constant (sometimes subscripted) which is always
independent of ε1 , ε2 and N . We adopt the convention that  ·  =  · Ω .
174 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

2. Decomposition of the solutions

The solution u1 of (1.1a) can be written as a sum u1 = v1 + w1 of two components v1 , w1 where the
regular component v1 and the singular component w1 are the solutions of the following problems
L1 v1 = f1 , v1 (0) = A, v1 (1) = u1 (1), (2.1a)
L1 w1 = 0, w1 (0) = u1 (0) − v1 (0), w1 (1) = 0 (2.1b)
and L1 φ ≡ ε1 φ  + a1,1 φ  . The boundary value v1 (0) can be chosen so that v1 is a regular component (i.e.,
first and second derivative of function are bounded independently of ε1 ). We have the following bounds
(see [6, Chapter 3]) on the components v1 , w1 and their first three derivatives:

(k)
   (k) 

v

C 1 + ε 2−k , w (x)
Cε −k e−α1 x/ε1 , k = 0, 1, 2, 3. (2.2)
1 1 1 1

The second equation in the system (1.1) can be rewritten as:

L2 u2 = −a2,1 u1 + f2 = −a2,1 v1 − a2,1 w1 + f2 ,
where L2 φ ≡ ε2 φ  + a2,2 φ  . Define a decomposition of u2 = v2 + w2 + z, where
L2 v2 = f2 − a2,1 v1 , v2 (0) = B, v2 (1) = u2 (1), (2.3a)
L2 w2 = 0, w2 (0) = u2 (0) − v2 (0) − z(0), w2 (1) = 0, (2.3b)

L2 z = −a2,1 w1 , z(0) = D, z(1) = 0. (2.3c)
The boundary value v2 (0) is again chosen so that v2 is a regular component. If z(0) is chosen such that
|z(0)|
C, we have the following bounds on the components v2 , w2 and their derivatives:

(k)
   (k) 

v

C 1 + ε 2−k , w (x)
Cε −k e−α2 x/ε2 , k = 0, 1, 2, 3.
2 2 2 2

The coupling component z can be further decomposed into the sum

z = z0 + ε2 z1 + ε22 z2 . (2.4a)
The components z0 , z1 and z2 are the solutions of the problems
a2,2 z0 = −a2,1 w1 , z0 (1) = 0, (2.4b)
 
a2,2 z1 = −z0 , z1 (1) = 0, (2.4c)

L2 z2 = −z1 , z2 (0) = 0, z2 (1) = 0. (2.4d)
The next lemma shows that when α1 ε2
α2 ε1 , the coupling component z has essentially an O(ε1 ) layer
near x = 0. In the subsequent analysis, we will often make use of the following inequality
 
  γi εj
Lj e−γi x/εi
γi − αi e−γi x/εi , i = 1, 2, j = 1, 2. (2.5)
εi εi

Lemma 1. Assume that α1 ε2
α2 ε1 . If z is the solution of (2.3c) and z(0) = z0 (0) + ε2 z1 (0) (see (2.4)),

then for x ∈ Ω,
γ1 x
−
(1) |z(x)|
Ca21 e ε1 , γ1 < α1 ;
(2) z(i) 
Cε1−i , i = 1, 2;
(3) z(3)
Cε1−2 ε2−1 .
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 175

Proof. Note first that |z(0)| = |z0 (0) + ε2 z1 (0)|
C, if ε2
Cε1 . When α1 ε2
α2 ε1 , we can use the
following barrier function
Ψ (x) = C1 e−γ1 x/ε1 ± z(x),
to get the desired bound on |z(x)|. Check that C1 can be chosen so that L2 Ψ
0 and Ψ (0)  0. The
following bounds for the derivatives of z0 , z1 are easily obtained from (2.4b) and (2.4c):
 (i) 
z (x)
Cε −i e−α1 x/ε1 , i = 0, . . . , 3, (2.6a)
0 1
 (i) 
z (x)
Cε −(i+1) −α
e 1 1 , i = 0, . . . , 2.
x/ε
(2.6b)
1 1

From the differential equation which defines z2 , we can establish that z2 
Cε1−2 ε2−1 and z2 

Cε1−2 ε2−2 . We have thus established the bounds on the first two derivatives of z. The bounds on the
third derivative of z follows by differentiating the differential equation (2.3c). 2

The proof in the previous lemma relied on the assumption that α1 ε2
α2 ε1 . In the case when
α2 ε1
α1 ε2 , we set z(0) = 0 and show that the coupling component has a layer of O(ε2 ) and the
maximum amplitude is proportional to ε1 /ε2 . Note that in this case the expansion (2.4) in ε2 is not
used.

Lemma 2. Assume that α2 ε1
α1 ε2 . If z is the solution of (2.3c) and z(0) = 0, then for x ∈ Ω

(1) |z(x)|
Ca21 ε1 ε2−1 e−γ2 x/ε2 , γ2 < α2 ,

(2) |z (x)|
Cε2−1 e−γ2 x/ε2 ,
(3) |z(i) (x)|
Cε2−1 (ε2−(i−1) e−γ2 x/ε2 + ε1−(i−1) e−α1 x/ε1 ), i = 2, 3.
γ2 x
−
Proof. If αα12  εε12  C2 > 0 then as in the previous lemma, using the barrier function e ε2
we deduce
from (2.5) that
 
z(x)
Ce−γ2 x/ε2
C ε1 e−γ2 x/ε2 .
ε2
Let C2 = α1
2a2,2 
. Then when ε1
ε2

α1
2a2,2 
, use the barrier function
ε1  −γ2 x/ε2 
Ψ (x) = C e − e−α1 x/ε1
ε2
to obtain the bound on z(x). Note that
α1
|L2 z|
C e−α1 x/ε1 , Ψ (0) = 0, Ψ (1)  0
ε1
and
 
ε1 γ2 −γ2 x/ε2 ε1 α1 α1 ε2
L2 Ψ = C (γ2 − a2,2 )e −C − a2,2 e−α1 x/ε1
ε2 ε2 ε2 ε1 ε1
 
α1 α1 a2,2 ε1 α1

−C α1 − e−α1 x/ε1 using

ε1 2a2,2  ε2 2a2,2 
α1 −α1 x/ε1

−C e .
ε1
176 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

Using this pointwise bound on the solution z and the differential equation (2.3c) defining z, establish that
|z (1)|
Cε1−1 e−α1 x/ε1 . Integrate the differential equation (2.3c) from t = x to t = 1 to get the desired
bound on the first derivative. Use (2.3c) again to bound the remaining derivatives of z. 2

We summarize this section with the statement of the following theorem.

Theorem 3. The solutions u1 , u2 of problem (1.1) can be decomposed into the sums
u1 = v1 + w1 , u2 = v2 + w2 + z
where for j = 1, 2 the regular and singular components vj , wj and their derivatives satisfy the bounds

(k)
   (k) 

v

C 1 + ε 2−k , w (x)
Cε −k e−αj x/εj , k = 0, 1, 2, 3.
j j j j

For an appropriate choice of z(0), the bounds on the coupling component z are of the form: if
α1 ε2
α2 ε1 , then
 
(i)

(3)

z(x)
Ce−γ1 x/ε1 , γ1 < α1 ,
z

Cε −i , i = 1, 2,
z

Cε −2 ε −1 .
1 1 2

and if, α2 ε1
α1 ε2 then

    
z(x)
Cε1 ε −1 e−γ2 x/ε2 , z (x)
Cε −1 e−γ2 x/ε2 , γ2 < α2 ,
2 2
 (i)   
z (x)
Cε −1 ε −(i−1) e−γ2 x/ε2 + ε −(i−1) e−α1 x/ε1 , i = 2, 3.
2 2 1

3. Error analysis

The piecewise-uniform mesh is designed to be a uniform mesh when the perturbation parameters ε1
and ε2 are sufficiently large relative to N . This occurs when 8εi  γi (ln N)−1 , i = 1, 2. If one of the small
parameters is relatively large, then the piecewise-uniform mesh has only two different mesh sizes. In this
paper, we confine our attention to two interesting cases when both parameters are relatively small and
the mesh has three different mesh sizes. That is, we consider the cases of
ε1 ln N ε2 ln N
α2 ε1
α1 ε2 and σ1 = 2 , σ2 = 4 (3.1a)
γ1 γ2
ε2 ln N ε1 ln N
α2 ε1  α1 ε2 and σ1 = 2 , σ2 = 4 . (3.1b)
γ2 γ1
The other cases when σ1 = 1/4 or σ2 = 1/2 or both may be dealt with using the techniques in [6] (see
also [2]).
The discrete solutions U1 and U2 are decomposed in an analogous fashion to the continuous solutions.
Let U1 = V1 + W1 , where

1 V1 (xi ) = f1 (xi ),
LN V1 (0) = v1 (0), V1 (1) = v1 (1), (3.2a)
LN
1 W1 (xi ) = 0, W1 (0) = w1 (0), W1 (1) = 0. (3.2b)
Let U2 = V2 + W2 + Z, where
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 177

+
2 V2 (xi ) = f2 (xi ) − a2,1 (xi )D V1 ,
LN V2 (0) = v2 (0), V2 (1) = v2 (1), (3.3a)
2 W2 (xi ) = 0,
LN W2 (0) = w2 (0), W2 (1) = 0, (3.3b)
L2 Z(xi ) = −a2,1 (xi )D + W1 ,
N
Z(0) = z(0), Z(1) = 0. (3.3c)

Lemma 4. If v1 , V1 are the solutions of (2.1a), (3.2a), respectively, then

 
(V1 − v1 )(xi )
CN −1 (1 − xi ), xi ∈ Ω εN,ε .
1 2

Proof. Note that |LN (V1 − v1 )(xi )|
CN −1 and use the barrier function CN −1 (1 − xi ) to finish. 2

Kopteva and Stynes [8] derived sharp estimates for the errors in the discrete derivatives in the case of a
piecewise-uniform mesh with one transition parameter. The next lemma gives analogous sharp weighted
derivative estimates in the case of a mesh with two transition points.

Lemma 5. If the transition points are such that σ1 = 2ε1γln1 N or σ2 = 4ε1 ln N

γ1
then for xi ∈ ΩεN1 ,ε2 ∪ {0}

 1
 min{ε1 ,ε2 } if xi < σ1 ,
 + 
D (V1 − v1 )(xi )
CN −1 × 1
if σ1
xi < σ2 ,

 max{ε1 ,ε2 }
1 if xi  σ2 ,
where v1 , V1 are the solutions of (2.1a), (3.2a), respectively.

Proof. The proof follows the argument in [6, Lemma 3.14], where the difference equations (3.2a) are
summed over the mesh points from xi = 1 to xi = 0. Let e1 (xi ) = (V1 − v1 )(xi ) and E1 (xj ) = D + e1 (xj ),
then
+
1 e1 (xj ) = ε1 δ e1 (xj ) + a1,1 (xi )D e1 (xj ) = τ (xj )
LN 2
(3.4)
−1
and by a standard truncation argument τ (xj ) = |(L1 − LN 1 )(v1 (xj ))|
CN . By the previous lemma
−2 −1
|e1 (xN−1 )|
CN and so |E1 (xN−1 )|
CN . Summing the difference equations (3.4) from xj = xi+1
to xj = xN−1 we get that
 
    
 N−1   N−1
 h̄j  
ε1 E1 (xi )
ε1 E1 (xN−1 ) + h̄j τ (xj ) +  a(xj ) e1 (xj +1 ) − e1 (xj ) .
 hj +1 
j =i+1 j =i+1

Using a telescoping effect for the last term, |e1 (xi )|
CN −1 and a  
C, we get that
 
ε1 D + (V1 − v1 )(xi )
CN −1 . (3.5)
One can further sharpen the weighted derivative estimates by summing difference equations (3.4) over
the mesh points from xi = 0 to xi = 1 and using (3.5).
Over the region [0, σ1 ), the result follows immediately from (3.5). Assume that σ1 = 2 ε1 αln1 N . We can
rewrite (3.4) in the form
ρj τ (xj ) a1,1 (xj )h̄j
(1 + ρj )E1 (xj ) = E1 (xj −1 ) + , ρj = . (3.6)
a1,1 (xj ) ε1
Summing these equations from j = 1 to j = N/4, we get that
178 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

    −(N/4−1)
E1 (σ1 )
E(0) (1 + ρ) + CN −1 ρ =
α1 H 1
1 + ρN/4 ε1
−1  −1 
CN CN

+ CN −1 since (1 + ρ)−N/4
CN −1
ε1 1 + 2α1 σ2 /(Nε1 )
 
CN −1 max{ε1 , ε2 }

+ CN −1
max{ε1 , ε2 } Nε1 + 2α1 σ2
CN −1

.
max{ε1 , ε2 }
Summing the equations in (3.6) from j = N/4 to j = i < N/2, we have that
    −(i−N/4−1)
CN −1
E1 (xi )
E1 (σ1 ) (1 + ρ̄) + CN −1
, ρ̄ =
α1 H 2
.
1 + ρi max{ε1 , ε2 } ε1
For j = N/2,
    −(N/4−1)
E1 (σ2 )
E1 (σ1 ) (1 + ρ̄) + CN −1 . (3.7)
1 + ρN/2
In the case when α2 ε1
α1 ε2 , then ρ̄  α2 H2 /ε2 and
 
  −1
CN −1
E1 (σ2 )
CN + CN −1
CN −1 .
max{ε1 , ε2 } 1 + α1 /(Nε1 )
For i > N/2 use (3.6) to complete the proof. On the other hand, if α2 ε1  α1 ε2 and σ1 = 2 ε2αln2 N , then
using the previous lemma we have that
  −1
E1 (σ1 )
CN .
ε1
Insert this into (3.7) to get |E1 (σ2 )|
CN −1 and use (3.6) to complete the proof. 2

The error bounds for the second regular component are easily established using the results just
obtained for the first regular component and the stability technique of Andreyev and Savin [1]. For our
purposes, we write this stability technique as follows:
Consider the finite difference operator LN associated with the difference equations:
LN Y (xi ) = εδ 2 Y (xi ) + a(xi )D + Y (xi ), i = 1, . . . , k − 1,
a(xi )  α > 0,
Y (0) = 0, Y (xk ) = B, xk
1.
Then, for any mesh function Y , with Y (0) = 0,

k−1
   
Y Ω N ∩[0,xk ]
h̄i LN Y (xi ) + C Y (xk ). (3.8)
i=1

2εj ln N
Lemma 6. If the transition points (1.2) are such that σ1 = αj
, j = 1 or 2 then
 
(V2 − v2 )(xi )
CN −1 ln N,
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 179

εN ,ε , where v2 and V2 are the solutions of (2.3a) and (3.3a), respectively.
at each mesh point xi ∈ Ω 1 2

Proof. The truncation error is

   + 
  
2 (V2 − v2 ) = L2 V2 − L2 v2 + L2 − L2 v2 = −a2,1 D V1 − v1 + L2 − L2 v2
LN N N N
     
= −a2,1 D + V1 − D + v1 − a2,1 D + v1 − v1 + L2 − LN2 v2 .
By the previous lemma and a standard truncation error estimate for the second and third terms we have
that

 min{ε1 ,ε } if xi < σ1 ,

 N  1 2
L (V2 − v2 )
CN −1 + CN −1 × 1
if σ1
xi < σ2 ,
2

 max{ε1 ,ε2 }
1 if xi  σ2 ,
and (V2 − v2 )(0) = (V2 − v2 )(1) = 0. From the stability bound (3.8) of Andreev and Savin [1] it follows
that
 
N−1

V2 − v2 ΩN
C h̄i LN 
2 (V2 − v2 )(xi )
CN
−1
ln N.
i=1
This completes the proof. 2

We now consider the error in the singular components. Note that,

 
wj (x)
CN −p , x  pεj ln N , j = 1, 2.
αj
Note also that

i 
i
e−γ1 xi /ε1 = e−γ1 hj /ε1
(1 + γ1 hj /ε1 )−1 .
j =1 j =1

We define the discrete barrier functions Bγk ,εk , k = 1, 2 to be

 i  
γk hj −1
Bγk ,εk (xi ) = 1+ . (3.9)
j =1
2εk
Note that we have a discrete counterpart to (2.5)
 
C εj γk
LN B (x
j γk ,εk i )
− a j,j Bγk ,εk (xi+1 ), j = 1, 2, k = 1, 2. (3.10)
εk εk
From this one can establish the bound
   
W1 (xi )
w1 (0)Bγ ,ε (xi ), γ1 < α1 .
1 1

Note that, for k = 1, 2,

   
γk H1 −N/4 2γk σ1 −N/4
Bγk ,εk (σ1 ) = 1 + = 1+ , (3.11a)
2εk Nεk
     
2γk σ1 −N/4 2γk (σ2 − σ1 ) −N/4 2γk σ2 −N/4
Bγk ,εk (σ2 ) = 1 + 1+
1+ . (3.11b)
Nεk Nεk Nεk
180 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

If σ1 = 2ε1 ln N/γ1 , then Bγ1 ,ε1 (σ1 )
(1 + 4 lnNN )−N/4
CN −1 . If σ2 = 4ε1 ln N/γ1 , then Bγ1 ,ε1 (σ2 )

(1 + 8 lnNN )−N/4
CN −2 . Thus, if σ1 = 2ε1 ln N/γ1 then |W1 (xi )|
CN −1 , xi  σ1 and if σ2 =
4ε1 ln N/γ1 then |W1 (xi )|
CN −2 , xi  σ2 .
2εj ln N
Lemma 7. If the transition points (1.2) are such that σ1 = γj
, j = 1 or 2 then
 
(W1 − w1 )(xi )
CN −1 (ln N),

at each mesh point xi ∈ ΩεN1 ,ε2 , where w1 and W1 are the solutions of (2.1b) and (3.2b), respectively.

Proof. First we consider the case of σ1 = 2ε2 ln N/γ2 and σ2 = 4ε1 ln N/γ1 . From above
|W1 − w1 |
|W1 | + |w1 |
CN −2 , xi  σ2 .
Hence |(W1 − w1 )(0)| = 0, |(W1 − w1 )(σ2 )|
CN −2 and on the interval (0, σ2 ), we have the truncation
error bound
 N 
L (W1 − w1 )(xi )
CN −1 σ2 ε −2 e−α1 xi−1 /ε1 , xi ∈ (0, σ2 ).
1 1

Use the barrier function

Ψi = Cσ2 ε1−1 N −1 eγ1 H2 /ε1 Bγ1 ,ε1 (xi ) + CN −2 ± (W1 − w1 )(xi )
to establish that
 
(w1 − W1 )(xi )
CN −1 ln N, xi ∈ (0, σ2 ).
An analogous argument establishes the bound in the case of σ1 = 2ε1 ln N/α1 . 2
2εj ln N
Lemma 8. If the transition points (1.2) are such that σ1 = αj
, j = 1 or 2 then
 
(W2 − w2 )(xi )
CN −1 (ln N),

at each mesh point xi ∈ ΩεN1 ,ε2 , where w2 and W2 are the solutions of (2.3b) and (3.3b), respectively.

Proof. Follows the same lines as the previous proof. 2

Lemma 9. If the transition points (1.2) are such that σ1 = 2ε1γln1 N or σ2 = 4ε1 ln N
γ1
then

 1
 min{ε1 ,ε2 } if xi < σ1 ,
 + 
D (W1 − w1 )(xi )
CN −1 ln N × 1
if σ1
xi < σ2 ,

 max{ε1 ,ε2 }
1 if xi  σ2 ,
where w1 and W1 are the solutions of (2.1b) and (3.2b), respectively.

Proof. We give the proof in the case of σ2 = 4ε1γln1 N . The proof in the other case of σ1 = 2ε1γln1 N is similar.
From the particular choice of transition points, we have that |W1 (xi )|
CN −2 , |w1 (xi )|
CN −2 for
xi  σ2 . This implies that
 + 
D (W1 − w1 )(xi )
CN −1 , xi ∈ [σ2 , 1].
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 181

1 W1 (σ2 ) = 0 yields
The finite difference equation LN
    
ε1 D W1 (σ2 − H2 ) =  ε1 + a1,1 (σ2 )h̄N/2 D + W1 (σ2 )
CN −1 .
+

Also note that |ε1 D + w1 (σ2 − H2 )| = |ε1 w1 (η)|
CN −1 . Thus

 
ε1 D + (W1 − w1 )(σ2 − H2 )
CN −1 .

Let ê1 (xi ) = (W1 − w1 )(xi ) and τ̂ (xi ) = LN 1 ê1 (xi ) then on the interval [σ1 , σ2 )
 
ε1 D + ê1 (xj ) − ε1 D + ê1 (xj −1 ) + H2 a1,1 (xj ) ê1 (xj +1 ) − e1 (xj ) = H2 τ̂ (xj ). (3.12)
Summing this set of difference equations from xj = xi > σ1 to xj = σ2 − H2 we obtain
ε1 D + ê(xi−1 ) = ε1 D + ê(σ2 − H2 ) + H2 a1,1 (σ2 − H2 )ê(σ2 ) − H2 a1,1 (xi )ê(xi )

N/2−1
  
N/2−1
− H2 a1,1 (xj ) − a1,1 (xj −1 ) êj − H2 τ̂j .
j =i j =i

which for each xi ∈ (σ1 , σ2 ) yields the bound

  
N/2−1
ε1 D + ê(xi−1 )
CN −1 ln N + Cε −2 N −2 e−α1 xj−1 /ε1
CN −1 ln N.
1
j =i

Check that |ε1 D + ê(σ1 − H1 )|
CN −1 ln N . Finally over the inner layer region [0, σ1 ) we repeat the
procedure just applied over the outer layer region [σ1 , σ2 ) to complete the proof. 2

We now consider the error in the coupling component z.

Lemma 10. If the transition points (1.2) are such that σ1 = 2ε1 ln N
γ1
or σ2 = 4ε1 ln N
γ1
then
 
(Z − z)(xi )
CN −1 (ln N)2 , xi ∈ Ω N ,
ε1 ,ε2

where z and Z are the solutions of (2.3c) and (3.3c), respectively.

Proof. Note that we can further decompose the discrete coupling component as follows

+ Z,
Z=Z  (3.13a)
where

+
2 Z = −a2,1 D w1 ,
LN
= Z(0),
Z(0)
= 0,
Z(1) (3.13b)
 +
2 Z = −a2,1 D (W1 − w1 ),
LN (0) = Z
Z (1) = 0. (3.13c)
Using (3.8) and the previous lemma we have that




N−1


Z
ΩN
C

 h̄i
D + (W1 − w1 )
ΩN
CN −1 (ln N)2 .
i=1

Thus,



Z − zΩN


Z − z
ΩN + CN −1 (ln N)2 .
182 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

The coupling component z is small outside the layers. The following bounds follow easily from the
pointwise bounds on |z(x)| given in Section 2.
  pε1 ln N
If α1 ε2
α2 ε1 , then z(x)
CN −p for x  ;
γ1
  pε2 ln N
If α2 ε1
α1 ε2 , then z(x)
CN −p for x  ;
γ2
 
If α2 ε1
α1 ε2 and ε1
CN −1 ε2 , then z(x)
CN −1 , ∀x.

We have the
By using the appropriate barrier functions, one can establish equivalent bounds on Z.
following:
 
If α1 ε2
α2 ε1 , then 
Z(xi )
CBγ1 ,ε1 (xi−1 );
  ε1
If α2 ε1
α1 ε2 , then 
Z(xi )
C Bγ2 ,ε2 (xi−1 ).
ε2
For the last inequality, when ε1
ε2

α1
2a2,2 
, use the barrier function
ε1  
C Bγ2 ,ε2 (xi−1 ) − Bγ1 ,ε1 (xi−1 ) .
ε2
Note that
      +   
LN
+ +
2 Z − z = a2,1 w1 − D w1 + a2,1 D w1 − D W1 + L2 − L2 z.
N

Using standard truncation error bounds and the bounds on the derivatives w1 , z , z , we get that
 N    
L Z
− z (xi )
C D + (W1 − w1 )(xi )
2
 h
 C εi+1
2 if α1 ε2
α2 ε1 ,
+ 1
1 −α1 xi−1 /ε1 
 Chi+1 2 + e 2 if α2 ε1
α1 ε2 .
ε ε 2 1

In the case of α1 ε2
α2 ε1 , then

    ε1


Z − z

Z  + |z|
CN −2 , xi  σ2 = 4 ln N,
γ1
and

 N   1
if xi < σ1 ,
L Z
− z (xi )
CN −1 ln N × min{ε1 ,ε2 }
2 1
max{ε1 ,ε2 }
if σ1
xi < σ2 .
Use (3.8) on the interval [0, σ2 ] to finish in the case of α1 ε2
α2 ε1 . In the other case of α2 ε1
α1 ε2 then
       ε2
 Z
− z (xi )


Z(xi ) + z(xi )
CN −2 , xi  σ2 = 4 ln N.
γ2
If ε1  CN −1 ε2 , then for xi  σ1 = 2ε1 ln N/γ1 ,
hi+1 e−α1 xi−1 /ε1 N −2 ε22 (N −1 ln N) N −1 ln N


C .
ε12 ε12 ε2
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 183

Hence

 N   1
if xi < σ1 ,
L Z
 −1
− z (xi )
CN ln N ×
min{ε1 ,ε2 }
2 1
max{ε1 ,ε2 }
if σ1
xi < σ2 .
Use (3.8) to get the result when α1 ε2  α2 ε1  CN −1 ε2 . Finally, if ε1
CN −1 ε2 , then observe that
      
 Z
− z (xi )


Z(xi ) + z(xi )
CN −1 , ∀xi

which completes the proof. 2

Theorem 11. If the transition points (1.2) are such that σ1 = 2ε1 ln N
γ1
or σ2 = 4ε1 ln N
γ1
then



max
Ui (x) − ui (x)

CN −1 (ln N)2
i=1,2
where {U1 , U2 } and {u1 , u2 } are the solutions of (1.4) and (1.1), respectively.

Proof. From the previous results we have the nodal error estimate
 
max max Ui (xj ) − ui (xj )
CN −1 (ln N)2 .
i=1,2 xj ∈Ω N

Note the following: If α1 ε2
α2 ε1 then

w2 − w̄2 [σ1 ,1]
2w2 [σ1 ,1]
CN −2 ,
 2
σ1  2
w2 − w̄2 [0,σ1 ]
C N −2
C N −1 ln N ,
ε2
z − z̄[σ2 ,1]
2z[σ2 ,1]
CN −2 ,
 2
σ2  2
z − z̄[0,σ2 ]
C N −2
C N −1 ln N .
ε1
On the other hand, if α2 ε1
α1 ε2 then
w2 − w̄2 [σ1 ,1]
2w2 [σ2 ,1]
CN −2 ,
 2
w2 − w̄2 [0,σ2 ]
C N −1 ln N ,
z − z̄[σ2 ,1]
CN −2 ,
 2 ε2 N −1  −1 2
z − z̄[σ1 ,σ2 ]
C N −1 ln N + C N ln N ,
ε1
 −1 2
z − z̄[0,σ1 ]
C N ln N .
Use the argument in [6, Theorem 3.12] to complete the proof. 2

4. Numerical results

Consider the constant coefficient problem for x ∈ Ω,

ε1 u1 (x) + 3u1 (x) = 15x 4 , (4.1a)
ε2 u2 (x) + 2u2 (x) + 2.75u1 (x) = 0.6ex , (4.1b)
184 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

Table 1
Exact errors EεN1 (u1 ), computed {ε1 , ε2 }-uniform errors E N (u1 ) and orders of convergence pN (u1 ) for the solution u1 of
(4.1a) on the piecewise uniform mesh ΩεN1 ,ε2 using the transition points (1.2)
ε1 Number of mesh points N
64 128 256 512 1024 2048
20 0.007838 0.003966 0.001996 0.001001 0.000501 0.000251
2−2 0.054983 0.029258 0.015102 0.007672 0.003867 0.001942
2−4 0.145435 0.093646 0.056039 0.031720 0.016389 0.008347
2−6 0.167567 0.108315 0.064312 0.037516 0.021132 0.011678
2−8 0.173777 0.112432 0.066790 0.038929 0.021911 0.012108
2−10 0.175373 0.113490 0.067445 0.039292 0.022112 0.012219
2−12 0.175774 0.113756 0.067609 0.039384 0.022162 0.012247
2−14 0.175875 0.113823 0.067651 0.039407 0.022175 0.012254
2−16 0.175900 0.113839 0.067661 0.039412 0.022178 0.012255
2−18 0.175906 0.113843 0.067664 0.039414 0.022179 0.012256
2−20 0.175908 0.113844 0.067664 0.039414 0.022179 0.012256
2−22 0.175908 0.113845 0.067664 0.039414 0.022179 0.012256
.. .. .. .. .. .. ..
. . . . . . .
2−30 0.175908 0.113845 0.067664 0.039414 0.022179 0.012256
2−34 0.175908 0.113843 0.067661 0.039413 0.022179 0.012256
2−38 0.175908 0.113812 0.067617 0.039399 0.022176 0.012254
2−42 0.175908 0.113184 0.067365 0.039070 0.022034 0.012177
2−46 0.154189 0.089766 0.050589 0.027318 0.014354 0.007358
2−50 0.130130 0.074975 0.042520 0.023660 0.012958 0.007018
E N (u1 ) 0.175908 0.113845 0.067664 0.039414 0.022179 0.012256
pN (u1 ) 0.628 0.751 0.778 0.830 0.856 0.875

where u1 (0) = 0, u1 (1) = 0 , u2 (0) = 0 and u2 (1) = 0. Note that u1 (x) ≈ x 5 − 1, u2 (x) ≈ 0.3(ex − e1 ) −
1.375(x 5 − 1) and, hence, u1  ≈ 1, u2  ≈ 1.
We define the exact errors EεN1 ,ε2 (uk ), EεN1 (uk ) and the computed parameter-uniform maximum
pointwise errors E N (uk ) as follows:



EεN1 ,ε2 (uk ) =
UkN − uk
Ω N , k = 1, 2,
ε1 ,ε2

EεN1 (uk ) = max EεN1 ,ε2 (uk ), k = 1, 2,

ε2

E N (uk ) = max EεN1 (uk ), k = 1, 2,

ε1

where UkN are the numerical approximations to uk (k = 1, 2) for a particular value N , ε1 and ε2 . We also
N
define the computed parameter-uniform orders of convergence to be p N = log2 EE2N . The nodal errors are
computed over the range S = {(ε1 , ε2 ): 2−50
ε1
1, 2−50
ε2
1}.
Table 1 illustrates that the approximations U1 converge {ε1 , ε2 }-uniformly to the exact solution u1 .
Similarly Table 2 exhibits the {ε1 , ε2 }-uniform convergence of the numerical approximations U2 to u2 .
 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186 185

Table 2
Exact errors EεN1 (u2 ), computed {ε1 , ε2 }-uniform errors E N (u2 ) and orders of convergence pN (u2 ) for the solution u2 of
(4.1b) on the piecewise uniform mesh ΩεN1 ,ε2 using the transition points (1.2)
ε1 Number of mesh points N
64 128 256 512 1024 2048
20 0.050392 0.029756 0.017750 0.010214 0.005748 0.003181
2−2 0.070608 0.037731 0.021373 0.012493 0.007096 0.003946
2−4 0.195129 0.126169 0.075772 0.042935 0.022203 0.011307
2−6 0.224595 0.145498 0.086910 0.050760 0.028655 0.015862
2−8 0.232548 0.150735 0.089964 0.052572 0.029660 0.016417
2−10 0.234579 0.152072 0.090748 0.053034 0.029916 0.016558
2−12 0.235090 0.152408 0.090958 0.053150 0.029981 0.016594
2−14 0.235218 0.152492 0.091010 0.053179 0.029997 0.016603
2−16 0.235250 0.152513 0.091023 0.053186 0.030001 0.016605
2−18 0.235258 0.152518 0.091026 0.053188 0.030002 0.016605
2−20 0.235260 0.152519 0.091027 0.053188 0.030002 0.016605
2−22 0.235260 0.152520 0.091028 0.053188 0.030002 0.016606
.. .. .. .. .. .. ..
. . . . . . .
2−30 0.235260 0.152520 0.091027 0.053189 0.030002 0.016606
2−34 0.235260 0.152519 0.091025 0.053188 0.030002 0.016606
2−38 0.235260 0.152513 0.091004 0.053183 0.030002 0.016606
2−42 0.235260 0.152215 0.091004 0.052953 0.029933 0.016570
2−46 0.214928 0.126649 0.071979 0.039056 0.020581 0.011216
2−50 0.172974 0.107959 0.063729 0.036550 0.020421 0.011216
E N (u2 ) 0.235260 0.152520 0.091028 0.053188 0.030002 0.016606
pN (u2 ) 0.625 0.745 0.775 0.826 0.853 0.874

Table 3
Orders of convergence pN corresponding to different theoretical error
bounds for various values of N
Number of intervals N
64 128 256 512 1024 2048
N −1 ln N 0.74 0.78 0.81 0.83 0.85 0.86
N −1 (ln N)2 0.44 0.53 0.60 0.65 0.69 0.72
(N −1 ln N)2 1.53 1.59 1.64 1.68 1.71 1.74

The boldface in each table highlights the maximum error over ε1 for each N value. The computed orders
of convergence in Tables 1 and 2 may be compared with the double mesh orders of convergence [6]
given in Table 3 generated from various theoretical error bounds. This comparison suggests that the
asymptotic error bound N −1 ln N is valid, as opposed to the error bound N −1 (ln N)2 established in
Theorem 11.
186 S. Bellew, E. O’Riordan / Applied Numerical Mathematics 51 (2004) 171–186

References

[1] V.B. Andreyev, I.A. Savin, The uniform convergence with respect to a small parameter of A.A. Samarskii’s monotone
scheme and its modification, Comput. Math. Math. Phys. 35 (1995) 581–591.
[2] S. Bellew, A Dirichlet problem for a coupled system of two singularly perturbed convection–diffusion ordinary differential
equations, [Link]. Thesis, School of Mathematical Sciences, Dublin City University, 2003.
[3] K.W. Chang, Singular perturbations of a boundary value problem for a vector second order differential equation, SIAM J.
Appl. Math. 30 (1) (1976) 42–54.
[4] K.W. Chang, F.A. Howes, Nonlinear Singular Perturbation Phenomena, Springer, New York, 1984.
[5] F.W. Dorr, S.V. Parter, L.F. Shampine, Applications of the maximum principle to singular perturbation problems, SIAM
Rev. 15 (1) (1973) 43–88.
[6] P.A. Farrell, A.F. Hegarty, J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Robust Computational Techniques for Boundary
Layers, Chapman and Hall/CRC Press, Boca Raton, FL, 2000.
[7] T. Linß, Layer-adapted meshes for convection–diffusion problems, Comput. Methods Appl. Mech. Engrg. 192 (2003)
1061–1105.
[8] N. Kopteva, M. Stynes, Approximation of derivatives in a convection–diffusion two-point boundary value problem, Appl.
Numer. Math. 39 (2001) 47–60.
[9] N. Madden, M. Stynes, A uniformly convergent numerical method for a coupled system of two singularly perturbed linear
reaction–diffusion problems, IMA J. Numer. Anal. 23 (4) (2003) 627–644.
[10] S. Matthews, E. O’Riordan, G.I. Shishkin, A numerical method for a system of singularly perturbed reaction–diffusion
equations, J. Comput. Appl. Math. 145 (2002) 151–166.
[11] J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems, World Scientific,
Singapore, 1996.
[12] H.-G. Roos, Layer-adapted grids for singular perturbation problems, Z. Angew. Math. Mech. 78 (5) (1998) 291–309.
[13] H.-G. Roos, M. Stynes, L. Tobiska, Numerical Methods for Singularly Perturbed Differential Equations. Convection–
Diffusion and Flow Problems, Springer, New York, 1996.
[14] G.I. Shishkin, A difference scheme for a singularly perturbed equation of parabolic type with a discontinuous initial
condition, Soviet Math. Dokl. 37 (1988) 792–796.
[15] G.I. Shishkin, Discrete approximation of singularly perturbed elliptic and parabolic equations, Russian Academy of
Sciences, Ural Section, Ekaterinburg, 1992 (in Russian).
[16] G.I. Shishkin, Mesh approximation of singularly perturbed boundary-value problems for systems of elliptic and parabolic
equations, Comput. Math. Math. Phys. 35 (4) (1995) 429–446.
[17] A.B. Vasil’eva, V.F. Butuzov, L.V. Kalachev, The Boundary Function Method for Singular Perturbation Problems, SIAM
Stud. in Appl. Math., vol. 16, SIAM, Philadelphia, PA, 1995.