The current issue and full text archive of this journal is available on Emerald Insight at:

[Link]/[Link]

EC
36,2 A second order numerical method
for singularly perturbed delay
parabolic partial
420 differential equation
Received 3 August 2018 Lolugu Govindarao and Jugal Mohapatra
Revised 2 October 2018
24 October 2018
Department of Mathematics, National Institute of Technology, Rourkela, India
Accepted 3 November 2018

Abstract
Purpose – The purpose of this paper is to provide a robust second-order numerical scheme for singularly
perturbed delay parabolic convection–diffusion initial boundary value problem.
Design/methodology/approach – For the parabolic convection-diffusion initial boundary value
problem, the authors solve the problem numerically by discretizing the domain in the spatial direction using
the Shishkin-type meshes (standard Shishkin mesh, Bakhvalov–Shishkin mesh) and in temporal direction
using the uniform mesh. The time derivative is discretized by the implicit-trapezoidal scheme, and the spatial
derivatives are discretized by the hybrid scheme, which is a combination of the midpoint upwind scheme and
central difference scheme.
Findings – The authors find a parameter-uniform convergent scheme which is of second-order accurate
globally with respect to space and time for the singularly perturbed delay parabolic convection–diffusion
initial boundary value problem. Also, the Thomas algorithm is used which takes much less computational
time.
Originality/value – A singularly perturbed delay parabolic convection–diffusion initial boundary value
problem is considered. The solution of the problem possesses a regular boundary layer. The authors solve this
problem numerically using a hybrid scheme. The method is parameter-uniform convergent and is of second
order accurate globally with respect to space and time. Numerical results are carried out to verify the
theoretical estimates.
Keywords Boundary layer, Delay differential equation, Hybrid scheme
Paper type Research paper

1. Introduction
Differential equations with an arbitrarily small parameter « multiplying the highest-order
derivative terms are said to be singularly perturbed problems (SPPs), and normally,
boundary layers occur in their solutions. Singularly perturbed parabolic convection–
diffusion equations appear in different branches of engineering and science. Simulations of
oil extraction from underground reservoirs (Ewing, 1983), convective heat transport with
large pèclet number, these are the some of the examples for such equations. The inclusion of
time delay terms in the differential equations also involves the small parameter. These
delays can represent gestation times, incubation periods, transport delays or can simply
lump complicated biological processes together, accounting only for the time required for
these processes to occur. Such models have the advantage of combining a simple, intuitive
Engineering Computations
derivation with a wide variety of possible behavior regimes for a single system. Delay
Vol. 36 No. 2, 2019
pp. 420-444
© Emerald Publishing Limited The authors express their sincere thanks to DST, Govt. of India for supporting this work under the
0264-4401
DOI 10.1108/EC-08-2018-0337 research grant EMR/2016/005805.
models are becoming more common, appearing in many branches of biological modeling Partial
(Bansal et al., 2017; Ewing, 1983; Wang, 1963). Singularly perturbed parabolic differential differential
equations with delay in time are more practical in nature as the time delay represents the
time-lag or after effect (Kuang, 1993). These type of models cannot be modeled by
equation
conventional differential equations, and hence makes the delay parabolic equations more
peculiar.
In this article, we consider the following singularly perturbed delay parabolic
convection–diffusion initial-boundary-value problem (IBVP) with Dirichlet boundary 421
conditions:
8

>
> @
>
> þ L« ;x uðx; tÞ ¼ cðx; tÞuðx; t  t Þ þ f ðx; tÞ; ðx; tÞ 2 D;
>
> @t
>
<
uðx; tÞ ¼ u b ðx; tÞ; ðx; tÞ 2 Cb ; (1.1)
>
>
>
> uð0; tÞ ¼ u l ðtÞ; on Cl ¼ fð0; tÞ : 0 # t # Tg;
>
>
>
: uð1; tÞ ¼ u ðtÞ;
r on C ¼ fð1; tÞ : 0 # t # Tg;
r

where

L« ;x uðx; tÞ ¼ « uxx ðx; tÞ þ aðxÞux ðx; tÞ þ bðx; tÞuðx; tÞ:

Here X = (0, 1), D = X  (0, T], C = Cl | Cb | Cr. Cl and Cr are the left and the right side of
the rectangular domain D is corresponding to x = 0 and x = 1, respectively. Cb = [0, 1] 
[–t , 0]. Also, 0 < «  1 and t > 0 is given a constant. The functions a(x), b(x,t), c(x,t), f(x,t) on
D and u b(x,t), u l(t), u r(t) on C are sufficiently smooth, bounded functions that satisfy a(x) 
a > 0, b(x, t) > 0, c(x, t)  b > 0 on D. The terminal time T is assumed to satisfy the
condition T = kt for some positive integer k. Under these conditions, the solution of IBVP
(1.1) exhibits a regular boundary layer of width O(« ) along x = 1.
There are many articles dealing with the numerical methods for convection–diffusion
and associated problems. It is well known that standard finite difference methods on
uniform mesh for (1.1) are unstable and fail to give good accurate results, because of the
small parameter « involves in these equations. To refer few articles, Clavero et al. (2005,
2003) and Salama and Al-Amerya (2017) used various numerical techniques for parabolic
IBVPs with regular layers. Das (2018a, 2018b) applied adaptive grid technique for solving
various parabolic IBVPs. Ng-Stynes et al. (1988) solved the time-dependent convection–
diffusion equations. By using the hybrid method, Mukherjee and Natesan (2009) solved the
time-dependent convection–diffusion problem. One can refer the book by Shishkin and
Shishkina (2009) for more details.
But only few articles are available in literature which deals with the theoretical analysis
and numerical methods for singularly perturbed delay ordinary/partial differential
equations (PDE). To cite, Mohapatra and Natesan solved the singularly perturbed delay
two-point BVP using layer adapted mesh in Mohapatra and Natesan (2010). Ansari et al.
(2007) solved the singularly perturbed delay parabolic reaction–diffusion problem on
piecewise uniform Shishkin mesh using finite difference scheme. Gowrisankar and Natesan
(2014) solved the parabolic IBVPs on layer-adapted mesh by equidistributing a monitor
function and in Gowrisankar and Natesan (2017) solved the singularly perturbed time delay
parabolic convection–diffusion problem of type (1.1) by the upwind scheme on Shishkin
mesh. But most of the above schemes are of first-order accurate. To increase the rate of
EC convergence, Das and Natesan (2015) solved the IBVP (1.1) by the hybrid scheme for spatial
36,2 direction on Shishkin mesh and the backward Euler method for the time direction on
uniform mesh. They showed that the rate of convergence is almost first order in time
direction and almost second order up to a logarithm factor in spatial direction. To show
globally the second-order rate of convergence, they assumed square of number of the mesh
intervals in spatial direction equal to the number of mesh intervals in the temporal direction.
422 Motivated by the work done in Das and Natesan (2015), to achieve a second-order
accuracy globally both with respect to time and space, we proposed in this paper a hybrid
scheme which is a combination of the midpoint upwind scheme and central difference
scheme for spatial direction and implicit-trapezoidal scheme for temporal direction. Again,
in Das and Natesan (2015), only standard Shishkin mesh is used, whereas here we extend the
hybrid scheme for Shishkin type of meshes (standard Shishkin mesh and the Bakhvalov–
Shishkin mesh) for the problem (1.1). Here, the main aim in this work is to obtain a second-
order accuracy globally both with respect to time and space without the restriction
mentioned earlier. This paper is organized as follows. In Section 2, we study the bounds on
the solution and its derivatives. Section 3 describes the Shishkin-type meshes and
construction of finite difference scheme. In Section 4, we study the convergence analysis.
Numerical results in shape of plots and tables are presented in Section 5, and Section 6
includes the conclusion.
Throughout the paper, C, C1, C2 denote generic positive constants independent of « ,
mesh points and mesh sizes. Here, k.k denotes the standard supremum norm, which is
defined as jjf jj1 ¼ sup jf ðx; tÞj; for a function f defined on some domain D.
ðx;tÞ2D

2. Bounds on the solution and its derivatives

The existence and uniqueness of the solution of (1.1) can be guaranteed by the sufficient
smoothness on u b(x,t), u l(t) and u r(t) along with the compatibility conditions at the corner
points and the delay terms as mentioned below:

u b ð0; 0Þ ¼ u l ð0Þ; u b ð1; 0Þ ¼ u r ð0Þ;

d u l ð0Þ @ 2 u b ð0; 0Þ @ u b ð0; 0Þ
« þ að0Þ þ bð0; 0Þu b ð0; 0Þ ¼ cð0; 0Þu b ð0; t Þ þ f ð0; 0Þ;
dt @x2 @x
du r ð0Þ @ 2 u b ð1; 0Þ @ u b ð1; 0Þ
« þ að1Þ þ bð1; 0Þu b ð1; 0Þ ¼ cð1; 0Þu b ð1; t Þ þ f ð1; 0Þ:
dt @x 2 @x
 
@
The operator @t þ L« ;x in (1.1) satisfies the following maximum principle:
Lemma  2.1 (Maximum
 principle) Suppose the function c ðx; tÞ 2 C0 ðD Þ \ C2 ð DÞ,
@
satisfies @t þ L« ;x c ðx; tÞ  0 in D and c (x, t)  0 on C. Then, c (x, t)  0 for all
ðx; tÞ 2 ðD Þ.    
Proof. Let c x* ; t* 2 ðD Þ such that c x* ; t* ¼ min c ðx; tÞ and assume that
 ðx;tÞ2ðD Þ 
c (x , t ) # 0. Clearly, (x , t ) 62 C and (x , t ) [ D.
As
it attains minimum at (x ,t ), we have
c x = 0, c t = 0 and c xx  0 at (x , t ). Therefore,
  from
 (1.1), we have
 
@ @
@t þ L « ;x c x ;
* *
t # 0; which is a contradiction to @t þ L « ;x c ð x; tÞ  0: Hence,
c (x,t)  0 8ðx; tÞ 2 ðD Þ. Refer to Farrell et al. (2000) for details.
Lemma 2.2 The solution of the problem (1.1) satisfies the following bounds:
 juðx; tÞ  u b ðx; 0Þj # Ct; ðx; tÞ 2 ðD Þ; Partial
differential
where C is independent of « . equation
Proof. Define the function q(x, t) as
(
juðx; tÞ  u b ðx; 0Þj; for 0 # t # T;
qðx; tÞ ¼
juðx; tÞ  u b ðx; tÞj; for  t # t # 0; 423
 
then q(x, t) satisfies the PDE @t@ þ L« ;x qðx; tÞ ¼ hðx; tÞ, where


@
hðx; tÞ ¼ f ðx; tÞ  þ L« ;x u b ðx; 0Þ  cðx; tÞu b ðx; t Þ;
@t

¼ f ðx; tÞ þ « ðu b Þxx ðx; 0Þ  aðxÞðu b Þx ðx; 0Þ

 bðx; tÞðu b Þðx; 0Þ  cðx; tÞu b ðx; t Þ;

and on the boundary Cb: q(x,t) = u(x,t) – u b(x,t) = u b(x,t) – u b(x,t) = 0, then q(0, t) = u l(t) –
u b(0,0), and q(1, t) = u r(t) – u b(1,0), where t [ [0, T]. Now, taking the absolute values on both
the sides and by using the compatibility conditions at corner points, we get |q(0,t)| = C1t
and |q(1,t)| = C1t, t [ [0, T] for some constants C1 and C2. On the other hand, by setting:
(
Ct 0 # t # T;
rðx; tÞ ¼
0 t # t # 0;

We have
8

>
> @  
>
> þ L« ;x r ðx; t Þ ¼ C 1 þ tbðx; t Þ ;
< @t
>
> rðx; tÞ ¼ 0; for Cb ;
>
>
:
rð0; tÞ ¼ rð1; tÞ ¼ Ct; for 0 # t # T:
   
Now, j @t@ þ L« ;x qðx; tÞj # @t@ þ L« ;x rðx; tÞ on D and |q(x, t)| # r(x, t) on C. Therefore,
using the maximum principle (Lemma 2.1), we get jqðx; tÞj # rðx; tÞ; on D, which implies
that |q(x,t) # Ct|. One can refer to Ng-Stynes et al. (1988) for more details.
Lemma 2.3 The solution u(x, t) of (1.1) satisfies the following bound: |u(x,t)| # C, on D.
Proof. From Lemma 2.2, we know that |u(x,t)| # Ct, on D. Again, as t [ [0,T], |u(x,t)| #
C, on D.
Theorem 2.4 For all non-negative integers l, m, satisfying 0 # l þ m # 5, the
derivatives of the exact solution u(x, t) of IBVP (1.1) satisfy the estimate:

@l þ mu   
1 ð1  xÞ=« ; ðx; tÞ 2 D:
j j # C 1 þ « exp a
@xl @tm
EC Proof. The proof of this theorem can be found in Das and Natesan (2015) and Salama
36,2 and Al-Amerya (2017).

2.1 Decomposition of the solution

To obtain the « -uniform error estimate, we need bounds on the derivatives of the analytical
solution u(x, t) of (1.1). So, we decompose the analytical solution u as u = v þ w, where v and
424 w are the regular and singular components, respectively. We express the regular component
X4
v as vðx; tÞ ¼ « k vk ðx; tÞ; ðx; tÞ 2 D, where vk(x, t), k = 0,1,2,3, are the solutions of the
k¼0
following first-order PDEs:
8
>
>
> ðv0 Þt ðx; tÞ þ aðxÞðv0 Þx ðx; tÞ þ bðx; tÞðv0 Þðx; tÞ ¼  cðx; tÞðv0 Þðx; t  t Þ
>
>
< þ f ðx; tÞ; ðx; tÞ 2 D;
(2.1)
>
> ðv0 Þðx; tÞ ¼ u b ðx; tÞ; ðx; tÞ 2 Cb ;
>
>
>
: ðv Þð0; tÞ ¼ u ðtÞ; on 0 # t # T;
0 l

8
>
>
> ðvk Þt ðx; tÞ þ aðxÞðvk Þx ðx; tÞ þ bðx; tÞðvk Þðx; tÞ ¼  cðx; tÞðvk Þðx; t  t Þ
>
>
< þ ðvk  1 Þxx ðx; tÞ; ðx; tÞ 2 D;
> ðvk Þðx; tÞ ¼ 0; ðx; tÞ 2 Cb ; k ¼ 1; 2; 3;
>
>
>
>
: ðv Þð0; tÞ ¼ 0; on 0 # t # T;
k

(2.2)

and the final component v4 satisfies

8
>
> ðv Þ ðx; tÞ þ L« ;x ðv4 Þðx; tÞ ¼  cðx; tÞðv4 Þðx; t  t Þ þ ðv3 Þxx ðx; tÞ; ðx; tÞ 2 D;
< 4 t
ðv4 Þðx; tÞ ¼ 0; ðx; tÞ 2 Cb ;
>
>
: v ð0; tÞ ¼ 0 ¼ v ð1; tÞ; on 0 # t # T:
ð 4Þ ð 4Þ
(2.3)

The regular component v(x, t) satisfies the following problem:

8
>
> vt ðx; tÞ þ L« ;x vðx; tÞ ¼  cðx; tÞðvÞðx; t  t Þ þ f ðx; tÞ; ðx; tÞ 2 D;
>
>
>
>
> vðx; tÞ ¼ u b ðx; tÞ; ðx; tÞ 2 Cb ;
>
<
vð0; tÞ ¼ u l ðtÞ; on 0 # t # T; (2.4)
>
>
>
> X4
>
>
>
> vð 1; t Þ ¼ « k vk ð1; tÞ; on 0 # t # T;
:
k¼0

and the singular component w(x, t) satisfies the PDE:

 8
>
> wt ðx; tÞ þ L« ;x wðx; tÞ ¼  cðx; tÞwðx; t  t Þ; ðx; tÞ 2 D; Partial
>
>
>
< wðx; tÞ ¼ 0; ðx; tÞ 2 Cb ; differential
>
(2.5) equation
>
> wð0; tÞ ¼ 0; on 0 # t # T;
>
>
: wð1; tÞ ¼ uð1; tÞ  vð1; tÞ; on 0 # t # T:

The following theorem provides the bounds for the derivatives of the regular and singular 425
components.
Theorem 2.5 For all non-negative integers l, m, satisfying 0 # l þ m # 5 and with
sufficient compatibility condition at the corners, the regular component v and the singular
component w, defined in (2.4) and (2.5), respectively, satisfy the following bounds:
 
 @l þ mv   
 
 l m  # C 1 þ « 4  l ; ðx; tÞ 2 D:
 @x @t 
1

and

@l þ mw  
j j # C «  l exp  að1  xÞ=« ; ðx; tÞ 2 D:
@xl @tm 1

Proof. One may refer to Shishkin and Shishkina (2009) for the details of the proof.

3. Numerical scheme and the non-uniform mesh

This section describes the semidiscretization and Shishkin-type meshes for the spatial
discretization.

3.1 The time semidiscretization

On the time domain [0, T], we use uniform mesh with time step D t,
XM t ¼ ftn ¼ nDt; n ¼ 0; . . . ; M; tM ¼ T; Dt ¼ T=Mg; Xpt ¼ ftj ¼ jDt; j ¼ 0; . . . ; p;
tp ¼ t ; Dt ¼ t =pg; where M is number of mesh points in t-direction on the interval [0, T],
and p is the number of mesh points in [–t , 0]. The step length D t satisfies pD t = t , where p
is a positive integer tn = nDt n  –p.
To discretize the time variable for (1.1), we use the implicit-trapezoidal method, defined
by:
8
> j
> u ¼ u b ðx;  tj Þ; for j ¼ 0; . . . ; p; x 2 X;
>
>
>
>

>
> Dt Dt  n þ 1 n  p þ 1 
>
> nþ1
þ f n þ 1  cn un  p þ f n
< I þ 2 L« ;x u ¼
2
c u

 (3.1)
>
> Dt
>
> þ I  L« ;x u ; n
>
>
>
> 2
>
>
: un þ 1 ð0Þ ¼ u ðt Þ; un þ 1 ð1Þ ¼ u ðt Þ;
l nþ1 r nþ1

where f n = f(x,tn),cn = c(x, tn), un = u(x, tn) is the semidiscrete approximation to the exact
solution u(x, t) of (1.1) at the time level tn = nD t.
EC The local truncation error is defined by jjenþ1 jj ¼ juðx; tnþ1 Þ  u ^ nþ1 j; where u
^ n is the
36,2 solution obtained after one step of the semidiscrete scheme taking the exact value u(x,tn),
instead of un, as the starting data. Concretely, we have the implicit-trapezoidal method,
defined by:
8
>
> ^  j ¼ u b ðx; tj Þ; for j ¼ 0; . . . ; p; x 2 X;
u
>
>
426 >
>

>
> Dt Dt  nþ1 n  p þ 1 
>
> I þ L ^ nþ1 ¼ c u þ f n þ 1  cn u n  p þ f n
< « ;x u
2 2

 (3.2)
>
> Dt
>
>
> þ I  L« ;x uðx; tn Þ;
>
>
> 2
>
>
:u ^ nþ1
^ n þ 1 ð1Þ ¼ u r ðtn þ 1 Þ:
ð0Þ ¼ u l ðtn þ 1 Þ; u

The stability of (3.1) in the maximum norm can found in Clavero et al. (2005) and Kumar and
Sekhara Rao (2010) as mentioned below. Let Enþ1 = |u(x, tnþ1) – unþ1(x)| be the global error
associated with the scheme (3.1). Then, it can be written as Enþ1 = enþ1 þ REn, where:

1

Dt Dt
R¼ Iþ L« ;x I L« ;x ;
2 2

is the transition operator defined as follows: REn is the result obtained after one time step of
(3.1) using un = En as the starting data, null boundary condition and zero source term f.
X n
Using this, we get the following recurrence relation Enþ1 ¼ Rnk ekþ1 . Thus, if the power
k¼0
of the transition operator R preserves the uniform boundedness behavior, that is,

jjRj jj1 # C 8 j ¼ 0; 1; . . . ; n; (3.3)

2
then it immediately follows that sup jjEnþ1 jj1 # C ðDtÞ ; that is, the semidiscrete
nDt # T
scheme (3.1) is parameter uniformly convergent of second order.
Note that (3.3) is typically a stability condition. A similar result is established in Palencia
(1993) for any operator R of the form R(Lx), where R(z) is a rational A-acceptable function,
and Lx is any operator that generates an analytic semigroup etLx . R(z) is the amplification
function of the Crank–Nicolson scheme, and Lx is the Laplacian operator. To prove (3.3), we
show that Lx,« is « -uniformly sectorial. Assume that (X,[Link]) and (Y,[Link]) be two Banach
spaces, and let L(X, Y) be the space of bounded linear operators from X into Y equipped with
the natural norm jjAjjLð X;Y Þ ¼ sup jjAxjjY =jjXjjX : A linear operator A: D(A)  X ! X is
x2X;x6¼0
said to be sectorial operator in X if there exists a constant w [ R, u [ (p /2, p ) and M > 0
such that
Xthe following holds:
 : ¼ fz 2 C : z 6¼ w; jargð z  wÞj < u g  r ð AÞ, the resolvent set of A, and
u ;w X
1
 jjð zI  AÞ jj # M for z 2 :
X zw
u ;w
A detail of this argument is given in Palencia (1993). If ‘A’ is sectorial, then ‘A’ generates an Partial
analytic semigroup S(t) = etA Vt  0. The operator ‘A’ is said to be « -uniformly sectorial if differential
we choose the Banach space X equipped with the sup norm k.k1 and the constant M is
independent of « . Let X = C(X) where X  R and equipped with sup norm. Now, consider
equation
the linear operator L« ,x: C2(X)  C(X) ! C(X), defined by L« ;x ¼ « @@xu2 þ aðxÞ @u
2
@x þ bðxÞu;
which we decomposed as Lx,« : = S1 þ S2, The operator S1: C2(X)  C(X) ! C(X) is defined
by S1 u ¼ « @@xu2 , and the operator S2: C(X) ! C(X) is defined by S2 u ¼ aðxÞ @u
2
@x þ bðxÞu:
427
The corollary (3.1.9) in Lunardi (1995) with the property that sectors are invariant under the
dilations gives that S1 is « -uniformly sectorial. Using this with S2 [ L(X, X), the proposition
(2.4.1) of Lunardi (1995) concludes that L« ,x is « -uniformly sectorial, which is our desired
result.
Lemma 3.1 If j @t@ i uðx; tÞj # C; 8ðx; tÞ 2 D; for 0 # i # 3, the local error associated to
i

scheme (3.1) satisfies:

3
jjenþ1 jj # C ðDtÞ : (3.4)

Proof. Using the central difference formula, it follows that:

uðx; tnþ1 Þ  uðx; tn Þ   2

¼ ut x; tn þ Dt=2 þ O ððDtÞ Þ;
Dt
   
¼  L« ;x u x; tn þ Dt=2 þ f x; tn þ Dt=2 þ OðDtÞ2 :
       
Where f x; tn þ Dt=2 ¼ c x; tn þ Dt=2 u x; tn þ Dt=2  p þ f x; tn þ Dt=2 . Again, from f ðx;
nþ1 n
þf 2
tn þ Dt=2 Þ ¼ f 2 þ OðDtÞ , we get

  uðx; tn þ 1 Þ þ uðx; tn Þ 2
L« ;x u x; tn þ Dt=2 ¼ L« ;x þ OðDtÞ :
2

It is straightforward to show that the local error kenþ1k is the solution of the following BVP:
8

>
< I þ Dt L« ;x e
> 3
nþ1 ¼ OðDt Þ ;
2
>
>
: e ð0Þ ¼ 0; e ð1Þ ¼ 0:
nþ1 nþ1

 
Now, using the maximum principle for the operator I þ Dt2 L« ;x proves the result. One can
refer to Clavero et al. (2005) for more details on this argument.
Combining the consistency result (3.4) with the stability result (3.3), we get the second-
order convergence of scheme (3.1).
Theorem 3.2 The global error estimate En associated with implicit trapezoidal scheme
in the time direction at nth level is given by:
2
jjEn jj1 # C ðDtÞ ; 8n # T=Dt:

Proof. The proof of the theorem can be found in Clavero et al. (2005).
EC 3.2 The spatial discretization
36,2 3.2.1 Shishkin mesh (S-mesh). Shishkin mesh is a piecewise uniform mesh. What
distinguish an S-mesh from any other piecewise uniform mesh is the choice of the transition
parameter (s), which is the point (s) at which the mesh size changes suddenly. As the
problem (1.1) has a boundary layer along the side x = 1, the mesh should be dense in the
neighborhood of x = 1. To define the piecewise uniform mesh, we divide the domain [0, 1]
428 into two subdomains such that [0, 1] = [0, 1 – s ] | (1 – s , 1] and then divide each of the
subdomain into N/2 equal intervals, N  4 be an even integer. Denote the spatial grid by
XNx ¼ f0 ¼ x0 ; x1 ; . . . ; xN=2 ¼ 1  s ; . . . ; xN ¼ 1g; where:
8
>
> 2ð1  s Þ
>
<i ; i ¼ 0; . . . ; N=2;
N
xi ¼

>
> N 2s
>
: ð1  s Þ þ i  ; i ¼ ð N=2Þ þ 1; . . . ; N:
2 N

The transition point 1–s , which separates the coarse and the fine points on the mesh, is
obtained by taking s = min{1/2,s 0« ln N}, where s 0  2/a. The analysis has been done by
assuming that s = s 0« ln N, as otherwise N is exponentially large compared to « . From the
definition of xi 0 s, the spatial step size can be denoted by:

2ð1  s Þ
H :¼ hi ¼ ; i ¼ 0; . . . ; N=2;
N

2s
h :¼ hi ¼ ; i ¼ ð N=2Þ þ 1; . . . ; N:
N

H and h are the mesh widths in [0,1 – s ] and [1 – s ,1], respectively. It is clear from the above
equation that N1 # H # 2N1, h = s 0« N1 and the uniform mesh can be obtained by
choosing s = 1/2. One may refer to Linß (1999).
3.2.2 Bakhvalov Shishkin mesh (B-S-mesh). The B-S mesh is a modification of the S-mesh
described above that incorporates an idea given by Bakhvalov, who proposed a mesh
condensing in the boundary layer by effectively inverting the boundary layer term. But the
original Bakhvalov mesh requires the solution of a non-linear equation to determine the
transition point where the mesh switches from coarse to fine. Instead, we fix the transition
point as done in the S-mesh:
8
>
> 2i
< ð1  s Þ N ;
> i ¼ 0; . . . ; N=2;
xi ¼

>
> 2« N 2  2ð N  iÞð N  1Þ
>
: 1 þ ln ; i ¼ ð N=2Þ þ 1; . . . ; N;
a N2

From the definition of xi 0 s, the spatial step size can be denoted by:

2ð1  s Þ
H :¼ hi ¼ ; i ¼ 0; . . . ; N=2;
N
 h :¼ hi ¼ xðiÞ  xði  1Þ; i ¼ ð N=2Þ þ 1; . . . ; N: Partial
differential
H and h are the mesh widths in [0,1 – s ] and (1 – s ,1], respectively. One may refer to Linß equation
(1999) for more details on the B-S mesh. We define the discretized domain DN ¼ XNx  XM t ,
CNb ¼ XNx  Xpt , where Xpt denotes the set of p þ 1 uniform mesh points in [–t ,0].

3.3 Finite difference scheme

429
We consider difference approximation of (1.1) on the domain XNx . Let us denote hj = xj – xj–1.
Given a mesh function f j, the forward, backward and the central difference operators in
space are given by:

f njþ1  f nj f nj  f nj 1 0 n f njþ 1  f nj 1

Dþ
x fj ¼
n
; D f n
¼ ; Dx f j ¼ ;
hjþ1 x j
hj hjþ1 þ hj

respectively, and the second-order central difference operator is given by:

!
þ  n 2 f njþ1  f nj f nj  f nj1
Dx Dx f j ¼  :
hj þ hjþ1 hjþ1 hj
f n  f n1
Define the backward difference operator in time by D t fj ¼
j
Dt
j
; where
f j ¼ f ðxj ; tn Þ: We propose a numerical scheme to solve (1.1), which consists of the
n

implicit-trapezoidal scheme for the time derivative, and the hybrid scheme for the spatial
derivatives. The hybrid scheme, which is a proper combination of the midpoint upwind
scheme in the outer region [0, 1-s ] and the central difference scheme in the inner region
(1s , 1]. Our proposed scheme is:
8
> 1 !
>
>  cn 1 U n  p þ f n 1  cn  11 U n  p  1 þ f n 11 ; 1 # i # N=2;
>
>
1  < 2 1 1
i i i i i i
D
t Ui þ
n
Lhyb Uin þ Lhyb Uin1 ¼ 2 2 2 2 2 2
2 >
>  
>
> 1
>
:  ci Uin n  p 
þ fi  ci Ui
n n 1 n  p  1
þ fi n  1
; N=2 < i # N  1;
2

U0n ¼ u l ðtn Þ; UNn ¼ u r ðtn Þ; Ui j ¼ u b ðxi ; tj Þ; for j ¼ 0; . . . ; p; and i ¼ 1; . . . ; N ¼ 1;

multiplying with 2 on both sides to the above equation and replacing n by n þ 1, we get
8
>
> nþ1 nþ1p
þ f n þ 1  cn U n  p þ f n  Lhyb Uin ; 1 # i # N=2;
> c U
   n þ 1 < i 1 i 1 i
1
i
1
i
1
i
1
2Dt þ Lhyb Ui ¼ 2 2 2 2 2 2
>
>
>
: cn U n þ 1  p þ f n þ 1  cn U n  p þ f n  Lhyb U n ; N=2 < i # N  1;
i i i i i i i

U0n þ 1 ¼ u l ðtn þ 1 Þ; UNn þ 1 ¼ u r ðtn þ 1 Þ; Ui j

¼ u b ðxi ; tj Þ; for j ¼ 0; . . . ; p; and i ¼ 1; . . . ; N ¼ 1;

where
 8
EC >
> L U n þ 1 ¼  « Dþ  nþ1
x Dx Ui þ a 1 D nþ1
x Ui þ bn þ 1 U n þ 1 ; for 1 # i # N=2;
< mu i i
1 1
36,2 Lhyb Uin þ 1 ¼ 2
i
2
i
2
>
>
:
Lce Uin þ 1 ¼ « Dþ D  nþ1
x x iU þ a D 0 nþ1
U
i x i þ bnþ1 nþ1
i U i ; for N=2 < i # N  1;

here, Lmu is midpoint upwind scheme and Lce is central difference scheme. The above
430 equation can be rewritten as follows:
8
> Dt Dt !
>
> nþ1 nþ1
 cn þ 1 U n  p þ 1 þ f n þ 1  cn 1 U n  p þ f n 1
> U 1 þ Lmu Ui
> ¼
1 1
>
> 2 2 1 1
>
> i i i i i i i
>
> 2 2 2 2 2 2 2
>
>

>
> Dt
>
> þ Un  Lmu Uin ; for 1 # i # N=2;
>
>
>
> 1 2
>
> i
>
<
2

Dt Dt  n þ 1 n  p þ 1 
>
> 1 þ Lce Uin þ 1 ¼  ci U i þ fin þ 1  cni Uin  p þ fin
>
> 2 2
>
>

>
>
>
> Dt
>
> þ 1  Lce Uin ; for N=2 < i # N  1;
>
> 2
>
>
>
>
>
> U0n þ 1
¼ u l ðtn þ 1 Þ; UNn þ 1 ¼ u r ðtn þ 1 Þ;
>
>
>
>
: U  j ¼ u b ðxi ; tj Þ for j ¼ 0; . . . ; p; and i ¼ 1; . . . ; N  1;
i

(3.5)
bni þbni1 Uin þUi1
n
In the above equations, ai1 ¼ ai þa2 i1 , bni1 ¼ 2 , Ui n
1 ¼ 2
n
fi 1 ¼

fin þfi1
n    2  2 2 2

2 ; fin ¼ f xðiÞ; tn ; Uin ¼ U xðiÞ; tn and ai = a(xi). After rearranging the terms in
(3.5), we obtain the following system of equations:
8
>  nþ1 o nþ1
þ riþ Uinþþ11 ¼ gin ; i ¼ 1; . . . ; N  1;
> ri Ui  1 þ ri Ui
>
<
U0n þ 1 ¼ u l ðtn þ 1 Þ; UNn þ 1 ¼ u r ðtn þ 1 Þ; (3.6)
>
>
>
: U j ¼ u ðx ; t Þ; for j ¼ 0; . . . ; p; and i ¼ 1; . . . ; N  1:
i b i j

The coefficients are given by:

8 0 1
>
> a 1 b 1
>
> Dt @ 2« i i ; n þ 1 A 1
>
> ri ¼   2þ 2 þ ;
>
>
>
> 2 h1ðiÞhðiÞ hðiÞ 2 2
>
>
>
< 0 1
a 1 b 1
i ; n þ 1 (3.7)
> Dt @ 2« iþ
2þ 2 A 1
>
> ri ¼
o
þ þ ;
>
> 2 h ð i þ 1Þ h ð i Þ h ð i Þ 2 2
>
>
>
>

>
> Dt 2«
>
> þ
: ri ¼ ; for 1 # i # N=2;
2 hði þ 1Þh1ðiÞ

and
 8

>
> Dt 2« ai Partial
>
> i
r 
¼   ;
>
> 2 h1ðiÞhðiÞ h1ðiÞ differential
>
>
>
<
 equation
Dt 2«
rio ¼ þ bi;nþ1 þ 1; (3.8)
>
> 2 hði þ 1ÞhðiÞ
>
>

>
>
>
> þ Dt 2« ai
> ri ¼
:  þ ; for N=2 < i # N  1;
2 h1ðiÞhði þ 1Þ h1ðiÞ 431

where h1ðiÞ ¼ hðiÞ þ hði þ 1Þ and

8   
> 
>  Dt cn þ 1 U n  pþ1 þ cn þ 1 U n  p þ 1  cni1 U n  p þ cni U n  p
>
>
>
> 4 i1 i1 i i i1 i
>
>
n 

>
< Dt   Dt 
nþ1 nþ1
 Ui þ Ui1 n
Dt
gi ¼ þ
n
f þ fi1 þ f þ fi1 þ
n n
 Lmu Uin ; for 1 # i # N=2;
>
> 4 i 4 i 2 2
>
>
>
> Dt  
Dt

>
> þ1
>
: 2 ci Ui
n npþ1 np
 ci Ui þ fi þ fi þ 1  Lce Uin ; for N=2 < i # N  1:
nþ1 nþ1 n
2

4. Convergence analysis
In this section, the stability of the proposed scheme (3.5) is discussed, and by using the
truncation error, we obtain the « -uniform error estimate. The tri-diagonal systems
 (3.7) and

(3.8) have the following properties: Assume that s 0 jjajj1 < lnNN
and aN  jjbjj1 þ Dt1 ,
then we have:
8
>
>
> r < 0; riþ < 0; for 1 # i # N  1;
< i
jr10 j  jr1þ j  0; jri0 j  jri j  jriþ j  0; for 1 < i # N=2;
>
>
>
: jr0 j  jrþ j > 0; jr0 j  jr j  jrþ j > 0; for N=2 < i < N  1:
N1 N1 i i i

The following theorem helps to study for the « -uniform convergence of the numerical
solution which is the main result of our work. Here, we define the domain D = D1 | D2,
where D1 = X  (0,t ), D2 = X(t ,2t ). Further, we define the discretized domain
DN ¼ DN1 [ DN2 , where DN1 ¼ XNx  Xp1;t and DN2 ¼ XNx  Xp2;t ; here, Xpj;t denotes the set of
p þ 1 uniform mesh points in [j – 1]t , jt for j = 1, 2.
Theorem 4.1 Let u and U be, respectively, the continuous and numerical solution of the
problem (1.1) and (3.5) and satisfying the compatibility conditions at the corners. Then, we
have error bound on th S-mesh for (xi, tn) [ DN:
8  
>
< C N 1 ð« þ N 1 Þ þ ðDtÞ2 ; for 0 # i # N=2;
max jðu  U Þðxi ; tn Þj #   (4.1)
i;n >
: C N 2 ln2 N þ ðDtÞ2 ; for N=2 < i < N;

and error bound on the B-S-mesh for (xi, tn)[DN

EC max jðu  U Þðxi ; tn Þj # ðCN 2 þ Dt2 Þ; for 0 # i # N  1; (4.2)
i;n
36,2
where U ðxi ; tn Þ ¼ Uin .
Proof. We divide the proof into various steps for different time levels. We notice that, on
the first interval t[[0, t ], i.e. where the time discretization parameter n varies from 0 to p, the
right-hand side f(x,t) – c(x,t)u b(x,t – t ) of (1.1) is independent of « . So, for ðxi ; tn Þ 2 DN1 , using
432 the convergence results of the midpoint upwind scheme discussed in Stynes and Roos (1997)
and Theorem 3.2, we obtain:
8  
>
< C N 1 ð« þ N 1 Þ þ ðDtÞ2 ; for 0 # i # N=2;
max jðu  U Þðxi ; tn Þj #   (4.3)
i;n >
: C N 2 ln2 N þ ðDtÞ2 for N=2 < i < N;

for the S-mesh. Again using the results in Stynes and Roos (1997) and using Theorem 3.2, we
obtain:
max jðu  U Þðxi ; tn Þj # ðCN 2 þ Dt2 Þ for 0 # i # N  1: (4.4)
i;n

for the B-S-mesh. Now, for t  t , the above approach is not applicable as the term u(x, t)
depends on u(x,t–t ), which is unknown for t  t . To get the error over the interval [t , 2t ],
the following proof is provided. On the domain D2, we consider following delay IBVP:
8

>
> @
>
> þ L« uðx; tÞ ¼ cðx; tÞuðx; t  t Þ þ f ðx; tÞ; ðx; tÞ 2 D2 ;
< @t ;x

(4.5)
>
> uðx; tÞ ¼ ut ðx; tÞ; ðx; tÞ 2 D1 ¼ X  ð0; t Þ;
>
>
:
uð0; tÞ ¼ u l ðtÞ; uð1; tÞ ¼ u r ðtÞ; t 2 ½t ; 2t ;
where ut is the exact solution on D1. We apply the implicit-trapezoidal scheme for the time
derivative and the hybrid scheme for the spatial derivative to determine the numerical
solution Uin of (4.5) at ðxi ; tn Þ 2 DN2 : The discrete problem corresponding to (4.5) is given by:
8
> Dt Dt !
>
> nþ1
þ nþ1
¼ c n þ 1 n  pþ1
þ nþ1
 n np
þ n
>
> U Lmu U
1
U f
1
c 1 U f 1
>
> 1 2 i
2 1 1
>
> i i i i i i i
>
> 2 2 2 2 2 2 2
>
>

>
> Dt
>
> þU n  Lmu Uin ; for 1 # i # N=2; ðxi ; tn Þ 2 ½0; 1  s  ½t ; 2t ;
>
>
>
> 1 2
>
> i
>
<
2

Dt Dt  nþ1 npþ1 
>
> 1 þ L U nþ1
¼ c U þ f nþ1
 c n np
U þ f n
>
> 2
ce i
2 i i i i i i
>
>

>
>
>
> Dt
>
> þ 1  Lce Uin ; for N=2 < i # N  1 ðxi ; tn Þ 2 ð1  s ; 1  ½t ; 2t ;
>
> 2
>
>
>
>
>
> U0 ¼ u l ðtnþ1 Þ; UNnþ1 ¼ u r ðtnþ1 Þ; tn 2 Xp2;t ;
nþ1
>
>
>
>
: U n ¼ U1 ðxi ; tn Þ ðxi ; tn Þ 2 DN ;
i 1

(4.6)
where U1(.,.) is the numerical solution calculated on DN1 . The solution u of (4.5) is Partial
decomposed into regular and singular components as u = v þ w because D2  D; the differential
estimates given in (2.4) and (2.5) can be used for both v and w. The regular component v is
the further decomposed into v = v0þ« v1, where v0 and v1 are the solutions of the following
equation
problems:
(
ðv0 Þt ðx; tÞ þ aðxÞðv0 Þx ðx; tÞ þ bðx; tÞðv0 Þðx; tÞ ¼ cðx; tÞðv0 Þðx; t  t Þ þ f ðx; tÞ; ðx; tÞ 2 D2 ; 433
ðv0 Þðx; tÞ ¼ uðx; tÞ; ðv0 Þð0; tÞ ¼ uð0; tÞ; ðx; tÞ 2 D1 ; t 2 ½t ; 2t ;

and
8

>
> @ @ 2 v0
>
> þ L« ;x v1 ðx; tÞ ¼  cðx; tÞv1 ðx; t  t Þ þ 2 ðx; tÞ; ðx; tÞ 2 D2 ;
< @t @x
>
> v1 ðx; tÞ ¼ 0; ðx; tÞ 2 D1 ;
>
>
:
v1 ð0; tÞ ¼ v1 ð1; tÞ ¼ 0; t 2 ½t ; 2t :

Therefore, v satisfies:
8

>
< @ þL
« ;x vðx; t Þ ¼  cðx; tÞvðx; t  t Þ þ f ðx; t Þ; ðx; tÞ 2 D2 ;
@t
>
:
vðx; tÞ ¼ uðx; tÞ; vð0; tÞ ¼ v0 ð0; tÞ; vð1; tÞ ¼ v0 ð1; tÞ; ðx; tÞ 2 D1 ; t 2 ½t ; 2t :
(4.7)

The singular component w is satisfies the following:

8

>
> @
>
> þ L« ;x wðx; t Þ ¼  cðx; t Þwðx; t  t Þ; ðx; tÞ 2 D2 ;
< @t
(4.8)
>
> wðx; tÞ ¼ 0; ðx; tÞ 2 D1 ;
>
>
:
wð0; tÞ ¼ u l ðtÞ  v0 ð0; tÞ; wð1; tÞ ¼ u r ðtÞ  v0 ð1; tÞ; t 2 ½t ; 2t :

In a similar way, the solution U of (4.6) can be decomposed into the regular and singular
components V and W, respectively. Thus, U = V þW, where V is the solution of following
non-homogeneous problem:
8  n
>
> 2Dt þ Lhyb Vi ¼ F ðV ðxi ; tn ÞÞ; ðxi ; tn Þ 2 D2
N
>
<
V0n ¼ vð0; tn Þ; VNn ¼ vð1; tn Þ; tn 2 Xp2;t (4.9)
>
>
>
: n
Vi ¼ U1 ðxi ; tn Þ; ðxi ; tn Þ 2 DN1 ;

and W satisfies:
 8  n
EC > 2D
> t þ Lhyb Wi ¼ F ðW ðxi ; tn ÞÞ; ðxi ; tn Þ 2 D2
N
>
>
36,2 <
W0n ¼ u l ðtn Þ  vð0; tn Þ; WNn ¼ u r ðtn Þ  vð1; tn Þ; tn 2 Xp2;t (4.10)
>
>
>
>
: W n ¼ 0 ðxi ; tn Þ 2 DN ;
i 1

434
where, Lhyb Vin is defined as:

(
Lmu Vin ; for 1 # i # N=2;
Lhyb Vin ¼
Lce Vin ; for N=2 < i # N  1;

similarly, Lhyb Win is defined, and

8 ! ! ! !
>
>  cn V x 1 ; tnp þ f ; tn  cn1 V ; tnp1 þ f ; tn1
>
> 1 x 1 1 x 1 x 1
>
> i i i i
>
>
i
2 2
i
2 2
>
> 2 2
<
F ðV ðxi ; tn ÞÞ ¼  Lhyb V ðxi ; tn1 Þ; for 1 # i # N=2;
>
>
>
>  cn V ðx ; t Þ þ f ðx ; t Þ  cn1 V ðx ; t
>
> i np1 Þ þ f ðxi ; tn1 Þ
>
> i i np i n i
>
>
:  L V ðx ; t Þ; for N=2 < i # N  1;
hyb i n1

8 ! !
>
>  cn W x 1 ; tnp  cn1 W x 1 ; tnp1  Lhyb W ðxi ; tn1 Þ; 1 # i # N=2;
>
< 1 1
i i i i
F ðW ðxi ; tn ÞÞ ¼ 2 2 2 2
>
>
>
:  cn W ðx ; t Þ  cn1 W ðx ; t
i i np i i np1 Þ  Lhyb W ðxi ; tn1 Þ; N=2 < i # N  1:

At the node (xi, tn), the error can be written in the form:

ðU  uÞðxi ; tn Þ ¼ ðV  vÞðxi ; tn Þ þ ðW  wÞðxi ; tn Þ;

which implies:

jðU  uÞðxi ; tn Þj # jðV  vÞðxi ; tn Þj þ jðW  wÞðxi ; tn Þj: (4.11)

Now, the error for the regular and singular components can be estimated separately.
Case I. For the regular component, the truncation error at the node (xi, tn) can be written
as:
 8 ! ! !
>
>  cn V x 1 ; tnp  cn1 V x 1 ; tnp1 þ f x 1 ; tn Partial
>
> 1 1
>
>
>
>
>
i
2
i
2
i
2
i
2
i
2 differential
>
>
>
> !   ! ! equation
<þf 
  
2Dt þ Lhyb ðV  vÞ ¼ x 1 ; tn1  2Dt þ Lmu v x 1 ; tn  Lmu V x 1 ; tn  1 ; i # N=2
i i i
>
> 2 2 2
>
>
>
>
>
>  ci V ðxi ; tnp Þ þ f ðxi ; tn Þ  ci V ðxi ; tnp1 Þ
n n1
>
>
>
> 435
: þ f ðxi ; tn1 Þ þ 2D þ Lce vðxi ; tn Þ  Lce V ðxi ; tn Þ; i > N=2
>
t

(4.12)

First, we consider in outer region (i.e. i > N/2). From (4.12), we get:
    
2D ð Þ
t þ Lhyb V  v ðxi ; tn Þ ¼  ci V ðxi ; tnp Þ  vðxi ; tnp Þ  ci
n n1
V ðxi ; tnp1 Þ


 
@  
 vðxi ; tnp1 ÞÞ þ þ L« ;x  2D þ Lce vðxi ; tn Þ
@t t


@
þ þ L« ;x vðxi ; tn1 Þ  Lce V ðxi ; tn1 Þ:
@t

Using the initial conditions from equation (4.7) and (4.9), we have:

    
2D ð Þ
t þ Lhyb V  v ðxi ; tn Þ ¼ ci U1 ðxi ; tnp Þ  uðxi ; tnp Þ  ci
n n1
U1 ðxi ; tnp1 Þ

 
@  
 uðxi ; tnp1 ÞÞ þ þ L« ;x  2D þ Lce vðxi ; tn Þ
@t t


@
þ þ L« ;x vðxi ; tn1 Þ  Lce vðxi ; tn1 Þ:
@t

Therefore,

    
2D ð Þ
t þ Lhyb V  v ¼  ci U1 ðxi ; tnp Þ  uðxi ; tnp Þ  ci
n n1
U1 ðxi ; tnp1 Þ

   
 uðxi ; tnp1 ÞÞ þ L« ;x  Lce vðxi ; tn Þ þ L« ;x  Lce vðxi ; tn1 Þ


@ @vðxi ; tn1 Þ
þ  2D vðxi ; tn Þ þ : (4.13)
@t t
@t

The last term of equation (4.13) can be written as:



EC @ @vðxi ; tn1 Þ @ðvðxi ; tn Þ þ ðvðxi ; tn1 Þ
 2D vðxi ; tn Þ þ ¼  2D
t vðxi ; tn Þ;
36,2 @t t
@t @t
!
@v xi ; t 1
n
@vðvðxi ; tn1 Þ vðxi ; tn Þ  vðvðxi ; tn1 Þ 2
¼2  ;¼ 2
@t Dt @t
436
2
! !
vðxi ; tn Þ  v xi ; t 1 þ v xi ; t 1  vðvðxi ; tn1 Þ
n n
2 2
 ;
Dt
2

After Taylor series expansion at (xi, tn –1/2) and taking modulus, we get:

  
@  @vðxi ; tn1 Þ Dt2  
 @3v 
 2Dt vðxi ; tn Þ þ #  3 :
@t @t 12  @t 
1

Now, taking the modulus of both the sides to (4.13), applying the triangle inequality and
using (4.3) on the S-mesh, we get:
     
j 2D ð Þ 2 2
t þ Lhyb V  v j # C4 N ln N þ Dt
2
þ C5 N 2 ln2 N þ Dt2


  
2 3 
@2 @ Dt  @ v 
þ «  Dþ xD x v þ ai  D0 xx v þ   :
@x2 @x 12  @t3 
1

We write the above inequality as:


    @2
ð Þ þ 
j 2D 2 2
t þ Lhyb V  v j # C N ln N þ Dt
2
þ «  D xD x v
@x2


@ Dt2 @ 3 v
þ jai  D xx vj þ
0
k k ;
@x 12 @t3 1

where C = max{C4,C5}. From the Taylor series expansion, we get

2       3
      @4v   @3v   @3v 
     
j 2Dt þ Lhyb ðV  vÞj # C 4 N ln N þ Dt þ « hi  4  þ hi  3  þ Dt  3 v1 5:
2 2 2 2 2 2
 @x   @x   @t 
1 1

Using the estimates for the derivatives from Theorem 2.5, we get

   
j 2D ð Þ 2 2
t þ Lhyb V  v j # C N ln N þ Dt ; ðxi ; tn Þ 2 D2 :
2 N
  
As the operator 2D ð Þ
t þ Lhyb V  v satisfies the discrete maximum principle, and the Partial
inverse operator is uniformly bounded, the inequality given above can be reduced to: differential
  equation
jðV  vÞj # C N 2 ln2 N þ Dt2 ; ðxi ; tn Þ 2 DN2 : (4.14)

Proceeding in the same fashion, we can get the following bound on the B-S-mesh:
437
jðV  vÞj # C ðN 2 þ Dt2 Þ; ðxi ; tn Þ 2 DN2 : (4.15)

Now, we consider the inner region (i.e. i # N/2). From (4.12), we have:
   ! !!
2Dt þ Lhyb ðV  vÞ ¼  cn u x 1 ; tnp  U1 x 1 ; tnp
1
i i i
2 2 2
! !!
 cn11 u ;t  U1 ;t
x 1 np1 x 1 np1
i i i
2 2 2


@ !  
þ  2D v 1 ; tn þ L« ;x  Lmu vðx; tn Þ
@t t x
i
2
@ !  
þ v 1 ; tn  1 þ L« ;x  Lmu vðx; tn1 Þ:
@t x
i
2

Now for i > N/2 and using Lemma 2.3, we get:

2  
   @3v 
j 2D þ Lhyb ðV  vÞj # C 4ðN 1 ðe þ N 1 Þ þ Dt2 Þ þ ð« þ hiþ1 Þðhi þ hiþ1 Þ
 3

t
 @x 
1

0    1   3
 @2v   @v   @3v 
     
þ h2i @ 2  þ   A þ Dt2  3  5:
 @x   @x   @t 
1 1 1

Using the estimates given in Theorem 2.5 and because hi # 2N1, hi þ hiþ 1 # 4N1, we
have:
 
j 2D ð Þ ð 1 ðe þ N 1 Þ þ Dt2 Þ; ðxi ; tn Þ 2 DN2 :
t þ Lhyb V  v j # C N

 
The operator 2D ð Þ
t þ Lhyb V  v satisfies the discrete maximum principle and the inverse
operator is uniformly bounded, then inequality given above can be reduced to:

jðV  vÞj # C ðN 1 ðe þ N 1 Þ þ Dt2 Þ; ðxi ; tn Þ 2 DN2 : (4.16)

In the same way, we can get following bound on the B-S-mesh:

EC jðV  vÞj # C ðN 2 þ Dt2 Þ; ðxi ; tn Þ 2 DN2 : (4.17)
36,2
Case II. Here, we derive the error bound for the singular component (W–w)(.,.). From (4.8)
and the difference equation (4.10), we get the truncation error as:
8 ! !!
>
> cn w x 1 ; tnp  W x 1 ; tnp
>
> 1
>
>
438 >
>
i i
2
i
2
>
> 2
>
> ! !!
> n1
>
>
> þc w x 1 ; tnp1  W x 1 ; tnp1
>
> i 1
>
> i i
>
> 2 2 2
>
>
>
>

>
> @ !  
>
>
>
> þ  2D w x 1 ; tn þ L« ;x  Lmu wðx; tn Þ
>
> @t t
>
> i
>
> 2
>
>
>
> !
>
<þ @ w
  
2Dt þ Lhyb ðW  wÞ ¼ @t x 1 ; tn  1 þ L« ;x wðx; tn1 Þ  Lmu W ðx; tn1 Þ;
>
>
i
>
> 2
>
>
>
>
>
> for i # N=2;
>
>
>
>    
>
> cni W ðxi ; tnp Þ  wðxi ; tnp Þ  cn1 W ðxi ; tnp Þ  wðxi ; tnp1 Þ
>
> i
>
>

 
>
>
>
> @   
>þ
> þ L «  2D þ L wðxi ; tn Þ
>
> @t
;x t ce
>
>
>
>

>
> @
>
>
>
> þ þ L « ;x wðxi ; tn1 Þ  Lce W ðxi ; tn1 Þ;
>
> @t
>
>
>
: for i > N=2:

(4.18)
Using the initial conditions given in (4.8) and (4.10), we have:
8

>
> @ !  
>
>  D 
w ; tn þ L« ;x  Lmu wðx; tn Þ
>
> @t t x 1
>
>
>
>
i
2
>
>
>
>
>
> @ !
>
>
> þ w x 1 ; tn  1 þ L« ;x wðx; tn1 Þ; for i # N=2;
<
   @t
2Dt þ Lhyb ðW  wÞ ¼
i
> 2
>
>

 
>
>   
>
> @
>
> þ L« ;x  2Dt þ Lce wðxi ; tn Þ
>
> @t
>
>
>

> 
>
> @
>
> þ þ
: L « ;x wðxi ; tn1 Þ; for i > N=2:
@t
(4.19)
By fixing t, the right-hand side of the equation can be seen as the truncation error of the two
BVPs as done in Stynes and Roos (1997), we obtain for (xi, tn):
 8
  < C ðN 1 ð« þ N 1 Þ þ Dt2 Þ for i # N=2; Partial
j 2D
t þ Lhyb
ð W  wÞ j ¼   (4.20) differential
: C N 2 ln2 N þ Dt2 for i > N=2: equation
 
The operator 2D ð Þ
t þ Lhyb W  w satisfies the discrete maximum principle and inverse
operator is uniformly bounded; from (4.20) we have:
8 439
< C ðN 1 ð« þ N 1 Þ þ Dt2 Þ for i # N=2;
ð
j W w j¼ Þ   (4.21)
: C N 2 ln2 N þ Dt2 for i > N=2: on S  mesh

Similarly, we can get the following bound:

jðW  wÞj # C ðN 2 þ Dt2 Þ; ðxi ; tn Þ 2 DN2 : on B  S  mesh (4.22)

Combining (4.14), (4.16), (4.21) and using it in (4.11), we complete the proof on the S-mesh for
the interval [t , 2t ]. Again combining (4.15), (4.17), (4.22) and using it in (4.11), we complete
the proof on the B-S-mesh for the interval [t , 2t ].

5. Numerical results
In this section, we present the numerical results obtained by the proposed numerical
scheme (3.6) for two test problems on the piecewise-uniform rectangular mesh
DN ¼ XNx  XM t : In all cases, we perform the numerical experiments by taking the constant
s 0 = 4.2 and the time step D t = 0.8/N.
Example 5.1 Consider the following problem:
8
< ut  « uxx þ ð1 þ xð1  xÞÞux ¼ uðx; t  1Þ þ f ðx; tÞ; ðx; tÞ 2 ð0; 1Þ  ð0; 2 ;
: uðx; tÞ ¼ u ðx; tÞ; uð0; tÞ ¼ 0; uð1; tÞ ¼ 0; t 2 ½0; 2 ; ðx; tÞ 2 ½0; 1  ½1; 0 ;
0

(5.1)

We choose the initial data u0(x,t) and the source function f(x, t) to fit with the exact solution:
 
uðx; tÞ ¼ expðtÞfm1 þ m2 x  exp ð1  xÞ=« g;

where m1 = exp(–1/« ) and m2 = 1–m1. The maximum pointwise error is calculated by

E«N;Dt ¼ maxðxi ;tn Þ2DN juðxi ; tn Þ  U N;Dt ðxi ; tn Þj; where u(xi, tn) and UN, Dt (xi, tn) denote the
exact and numerical solutions, respectively, obtained on the mesh DN with N mesh intervals
in the spatial direction and M mesh intervals in the t-direction, such that Dt = T/M. The
corresponding order of convergence is given by:
!
E«N;Dt
P«N;Dt ¼ log2 2N;Dt=2
:
E«

Example 5.2 Consider the following singular perturbed delay parabolic IBVP:
 8
EC >
> u  « uxx þ ð2  x2 Þux þ xu ¼ uðx; t  1Þ
>
< t
36,2
þ10t2 expðtÞxð1  xÞ; ðx; tÞ 2 ð0; 1Þ  ð0; 2 ; (5.2)
>
>
>
: uðx; tÞ ¼ 0; uð0; tÞ ¼ 0; uð1; tÞ ¼ 0; t 2 ½0; 2 ; ðx; tÞ 2 ½0; 1  ½1; 0 :

As the exact solution of Example 5.2 is unknown, to obtain the pointwise errors and to
440 verify the « -uniform convergence of the proposed scheme, we use the double-mesh principle
which is described as below. Let U ~ ðxi ; tn Þ be the numerical solution obtained on the fine
mesh D ~ ¼X
2N
~ X
2N
~ 2M
with 2N mesh intervals in the spatial direction and 2M mesh
x t
intervals in the t-direction, where X ~ is piecewise-uniform Shishkin mesh as like XN with
2N
x x
the same transition parameter. Then, for each « , we calculate the maximum pointwise errors
~ N;Dt ¼ max ~
by E Þ2DN jU ðxi ; tn Þ  U ðxi ; tn Þj; and the corresponding order of convergence
«
ðxi ;tnN;Dt 
by P~ N;Dt ¼ log2 E~2N;Dt=2
«
:
« ~ E«
The numerical solution of Example 5.1 is plotted in Figures 1 (a) and (b) for « = 1e –
1 and « = 1e – 6, respectively. These figures confirm the existence of the boundary layer
near x = 1. For various values of « and N, the calculated maximum pointwise errors E«N
and rate of convergence P«N for Example 5.1 by using hybrid scheme are presented in
Tables I and II on the S-mesh and B-S-mesh, respectively. To reveal the numerical order

0.8
1

0.6 0.8

0.6
0.4
u
u

0.4

0.2 0.2

0 2
0 1
2 0 0.2 0.4 0.6
0 1 0.8 10
Figure 1. 0.5 10 x t
x t
Surface plot of the
numerical solution for (a) (b)
Example 5.1
Notes: (a) ε = 1e − 1 and N = 64; (b) ε = 1e − 6 and N = 64

No. of intervals N
« 16 32 64 128 256 512

1e2 6.9625e2 2.5245e2 8.8005e3 2.9322e3 9.4699e4 2.9975e4

1.4697 1.5203 1.5856 1.6306 1.6596
1e3 7.0668e1 0.6280e2 8.9005e3 2.9412e3 9.5291e4 3.0846e4
1.4271 1.5620 1.5975 1.6260 1.6273
Table I. 1e4 7.0761e2 2.6370e2 8.9325e3 2.9512e3 9.5369e4 3.1246e4
E«N and P N generated 1.4242 1.5618 1.5981 1.6297 1.6273
on the S-mesh for 1e6 to 7.07542e2 2.6393e2 8.9425e3 2.9532e3 9.5399e4 3.1846e4
Example 5.1 1e10 1.4227 1.5636 1.6046 1.6302 1.5829
of convergence, the maximum pointwise errors are plotted in log-log scale in Figures 2 Partial
(a) and (b) for Example 5.1 on the S-mesh and B-S mesh, respectively. The numerical differential
solution of Example 5.2 is plotted in Figures 3 (a) and (b) for « = 1e – 1 and « = 1e – 6,
respectively, which clearly indicates the layer near x = 1. The calculated maximum
equation
pointwise errors E~ N and rate of convergence P~ N for Example 5.2 on the B-S-mesh are
« «
presented in Table IV. From these results given in tables and figures, one can observe
that more accurate results are obtained using the B-S-mesh than the S-mesh. This is 441

No. of intervals N
« 16 32 64 128 256 512

1e2 1.5392e2 4.9172e3 1.4283e3 3.8331e4 9.9185e5 2.5221e5

1.6462 1.7835 1.8978 1.9503 1.9755
1e3 1.3835e2 3.7569e3 1.1276e3 3.4683e4 9.5862e5 2.5456e5
1.8806 1.7363 1.7010 1.8552 1.9130
1e4 3.8000e2 9.7171e3 2.4728e3 6.2751e4 1.5827e4 3.9757e5 Table II.
1.9674 1.9744 1.9785 1.9872 1.9931 E«N and P N generated
1e6 to 4.5671e2 1.2684e2 3.3333e3 8.5025e4 2.1366e4 5.3527e5 on the B-S-mesh for
1e10 1.8483 1.9280 1.9710 1.9926 1.9970 Example 5.1

−1
10
−1 10
1e−2 1e−2
1e−5 10
−2 1e−5
−2
10
1e−7 1e−7
−2 2
−3 −2
O(N lnN ) 10 O(N
−3
10
−4
10

−4
10 −5
10

−5 −6
10 10 1 2 3
1 2 3
10 10 10 10 10 10

(a) (b) Figure 2.

Log-log plot for
Example 5.1
Notes: (a) On S-mesh; (b) on B-S-mesh

0.6 0.6

0.4 0.4
u
u

0.2
0.2

0
0 0
0
1 1
0.6 0.8 1
2 0 0.2 0.4 0.6 0.8 1 2 0 0.2 0.4
t t x Figure 3.
x
(a) (b) Surface plot of the
numerical solution for
Example 5.2
Notes: (a) ε = 1e − 1 and N = 64; (b) ε = 1e − 6 and N = 64
EC because of the presence of the logarithm term given in the bound (4.21). One can observe
36,2 that we have obtained second-order convergence with respect to time.
To show the second-order convergence, usually researchers adapt the idea of using M =
N2 (Dt = Dx2) (Das and Natesan, 2015; Gowrisankar and Natesan, 2017). But, here in this
paper, we have developed a second order of convergence for both space and time,
without such a restriction which shows the efficiency of the proposed scheme over
442 others. To verify that the proposed method performs better than the earlier methods
available in the literature, we have compared our result with the results given in Das
and Natesan (2015) in Table III for Example 5.1 which clearly shows the advantage of
our proposed scheme. All the results and graphs shown here confirm that the present
method is « -uniform with global second-order convergence i.e. second order both in
space and time.

6. Conclusion
In this article, we propose a numerical method to solve singularly perturbed delay
parabolic convection–diffusion problem. The domain is discretized with a uniform
mesh in time and for space variable non-uniform meshes (S-mesh and B-S-mesh) are
considered. The numerical scheme consists of the implicit-trapezoidal scheme for the
time derivative and the hybrid numerical scheme which is a combination of the
midpoint-upwind scheme in the outer region and central difference scheme in the inner
region for the spatial derivatives. We have derived the « -uniform error estimate for
the proposed scheme, which is globally second order in space and time (O(Dt2 þ N2)).
The numerical results are carried out to show the efficiency and the accuracy of the
proposed method.

Results in Das and Natesan (2015)

Our results S-mesh B-S-mesh S-mesh
N 1e–4 1e–6 1e–4 1e–6 1e–4 1e–6

32 2.6370e–2 2.6393e–2 9.7171e–3 1.2684e–2 2.7273e–2 2.7434e–2

1.561 1.5636 1.9744 1.9280 1.5577 1.5563
64 8.9325e–3 8.9425e–3 2.4728e–3 3.3333e–3 9.2019e–3 9.3275e–3
Table III. 1.5981 1.6046 1.9785 1.9710 1.5685 1.5743
Comparison of 128 2.9512e–3 2.9532e–3 6.2751e–4 8.5025e–4 3.1235e–3 3.1319e–3
numerical results 1.6297 1.6302 1.9872 1.9926 1.6218 1.6260

No. of intervals N
« 16 32 64 128 256

1e–4 3.3549e–2 9.2118e–3 2.6992e–3 7.8307e–4 2.5310e–4

1.8647 1.7710 1.7853 1.7972
Table IV. 1e–5 3.3686e–2 9.2916e–3 2.7232e–3 7.8416e–4 2.2631e–4
~ N and P
E ~N 1.8581 1.7706 1.7961 1.7915
«
generated on the B-S- 1e–6 to 3.3699e–2 9.2996e–3 2.7256e–3 7.8539e–4 2.2638e–4
mesh for Example 5.2 1e–10 1.8575 1.7706 1.7951 1.7920
References Partial
Ansari, A.R., Bakr, S.A. and Shishkin, G.I. (2007), “Parameter-robust finite difference method for differential
singularly perturbed delay parabolic partial differential equations”, Journal of Computational
and Applied Mathematics, Vol. 205 No. 1, pp. 552-566.
equation
Bansal, K., Rai, P. and Sharma, K.K. (2017), “Numerical treatment for the class of time dependent
singularly perturbed parabolic problems with general shift arguments”, Differential Equations
and Dynamical Systems, Vol. 25 No. 2, pp. 327-346.
443
Clavero, C., Gracia, J.L. and Jorge, J.C. (2005), “High-order numerical methods for one-dimensional
parabolic singularly perturbed problems with regular layers”, Numerical Methods for Partial
Differential Equations, Vol. 21 No. 1, pp. 149-169.
Clavero, C., Jorge, J.C. and Lisbona, F. (2003), “A uniformly convergent scheme on a nonuniform mesh
for convection-diffusion parabolic problems”, Journal of Computational and Applied
Mathematics, Vol. 154 No. 2, pp. 415-429.
Das, A. and Natesan, S. (2015), “Uniformly convergent hybrid numerical scheme for singularly
perturbed delay parabolic convection-diffusion problems on Shishkin mesh”, Applied
Mathematics and Computation, Vol. 271, pp. 168-186.
Das, P. (2018a), “A higher order difference method for singularly perturbed parabolic partial
differential equations”, Journal of Difference Equations and Applications, Vol. 24 No. 3,
pp. 452-477.
Das, P. (2018b), “An a posteriori based convergence analysis for a nonlinear singularly perturbed
system of delay differential equations on an adaptive mesh”, Numerical Algorithms, doi: 10.1007/
s11075-018-0557-4.
Ewing, R.E. (1983), The Mathematics of Reservoir Simulation, Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA.
Farrell, P.A., Hegarty, A.F., Miller, J.J.H., O’Riordan, E. and Shishkin, G.I. (2000), Robust Computational
Techniques for Boundary Layers, Chapman and Hall/CRC Press, Boca Raton, FL.
Gowrisankar, S. and Natesan, S. (2014), “A robust numerical scheme for singularly perturbed delay
parabolic initial-boundary-value problems on equidistributed grids”, [Link]. Numer.
Anal, Vol. 41, pp. 376-395.
Gowrisankar, S. and Natesan, S. (2017), “« -uniformly convergent numerical scheme for singularly
perturbed delay parabolic partial differential equations”, International Journal of Computer
Mathematics, Vol. 94 No. 5, pp. 902-921.
Kuang, Y. (1993), Delay Differential Equations with Applications in Population Dynamics, volume 191 of
Mathematics in Science and Engineering, Academic Press, Boston, MA.
Kumar, M. and Sekhara Rao, S.C. (2010), “High order parameter-robust numerical method for time
dependent singularly perturbed reaction-diffusion problems”, Computing, Vol. 90 Nos 1/2,
pp. 15-38.
Linß, T. (1999), “An upwind difference scheme on a novel Shishkin-type mesh for a linear convection-
diffusion problem”, Journal of Computational and Applied Mathematics, Vol. 110 No. 1,
pp. 93-104.
Lunardi, A. (1995), Analytic Semigroups and Optimal Regularity in Parabolic Problems, Progress in
nonlinear differential equations and their applications, Birkhuser, Basel, p. 16.
Mohapatra, J. and Natesan, S. (2010), “Uniform convergence analysis of finite difference scheme for
singularly perturbed delay differential equation on an adaptively generated grid”, Numerical
Mathematics: Theory, Methods and Applications, Vol. 3, pp. 1-22.
Mukherjee, K. and Natesan, S. (2009), “Parameter-uniform hybrid numerical scheme for time-dependent
convection-dominated initial-boundary-value problems”, Computing, Vol. 84 Nos 3/4,
pp. 209-230.
EC Ng-Stynes, M.J., O’Riordan, E. and Stynes, M. (1988), “Numerical methods for time-dependent
convection-diffusion equations”, Journal of Computational and Applied Mathematics, Vol. 21
36,2 No. 3, pp. 289-310.
Palencia, C. (1993), “A stability result for sectorial operators in Banach spaces”, SIAM Journal on
Numerical Analysis, Vol. 30 No. 5, pp. 1373-1384.
Salama, A.A. and Al-Amerya, D.G. (2017), “A higher order uniformly convergent method for singularly
perturbed delay parabolic partial differential equations”, International Journal of Computer
444 Mathematics, Vol. 94 No. 12, pp. 2520-2546.
Shishkin, G.I. and Shishkina, L.P. (2009), Difference Methods for Singular Perturbation Problems, CRC
Press, Boca Raton.
Stynes, M. and Roos, H.-G. (1997), “The midpoint upwind scheme”, Applied Numerical Mathematics,
Vol. 23 No. 3, pp. 361-374.
Wang, P.K.C. (1963), “Asymptotic stability of a time-delayed diffusion system”, Journal of Applied
Mechanics, Vol. 30 No. 4, pp. 500-504.

Corresponding author
Jugal Mohapatra can be contacted at: jugal@[Link]

For instructions on how to order reprints of this article, please visit our website:
[Link]/licensing/[Link]
Or contact us for further details: permissions@[Link]