Das 2018

Computers and Mathematics with Applications ( ) –
Contents lists available at ScienceDirect
Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa
Higher-order convergence with fractional-step method for

singularly perturbed 2D parabolic convection–diffusion
problems on Shishkin mesh
Abhishek Das a , Srinivasan Natesan b, *
a
Faulty of Science and Technology, ICFAI University, Tripura Campus, Agartala, Tripura, India
b
Department of Mathematics, Indian Institute of Technology, Guwahati 781039, India
article info a b s t r a c t
Article history: In this article, we propose a second-order uniformly convergent numerical method for a
Received 24 January 2017 singularly perturbed 2D parabolic convection–diffusion initial–boundary-value problem.
Received in revised form 29 July 2017 First, we use a fractional-step method to discretize the time derivative of the continuous
Accepted 16 December 2017
problem on uniform mesh in the temporal direction, which gives a set of two 1D problems.
Available online xxxx
Then, we use the classical finite difference scheme to discretize those 1D problems on
Keywords: a special mesh, which results almost first-order convergence, i.e., O(N −1+β ln N + ∆t).
Singularly perturbed 2D parabolic To enhance the order of convergence to O(N −2+β ln2 N + ∆t 2 ), we use the Richardson
convection–diffusion problems extrapolation technique. In support of the theoretical results, numerical experiments are
Fractional-step method performed by employing the proposed technique.
Finite difference scheme © 2017 Elsevier Ltd. All rights reserved.
Piecewise-uniform Shishkin meshes
Richardson extrapolation technique
Uniform convergence
1. Introduction
In this paper, we consider the following singularly perturbed 2D parabolic convection–diffusion initial–boundary-value
problem (IBVP) posed on the domain G = D × Ωt , where D = Ix × Iy = (0, 1)2 and Ωt = (0, T ]:
⎨ut + Lε u(x, y, t) = f (x, y, t), (x, y, t) ∈ G,

⎧
u(x, y, 0) = φ (x, y), (x, y) ∈ D, (1)

u(x, y, t) = 0, (x, y, t) ∈ ∂ D × Ω t ,
⎩
where
Lε u = −ε ∆u + a(x, y).∇ u + b(x, y)u,
0 < ε ≪ 1 is the perturbation parameter. The coefficients a = (a1 , a2 ) and b are assumed to be sufficiently smooth and
bounded functions such that a1 (x, y) ≥ αx > 0, a2 (x, y) ≥ αy > 0 and b(x, y) ≥ 0, on D.
Under the sufficient smoothness and necessary compatibility conditions [1] imposed on the functions f and φ , the
parabolic IBVP (1) admits a unique solution u(x, y, t), which exhibits a regular boundary layer of width O(ε ) along the sides
x = 1 and y = 1, and a corner layer at (x, y) = (1, 1) [2].
* Corresponding author.
E-mail addresses: abhishek.das@iitg.ernet.in (A. Das), natesan@iitg.ernet.in (S. Natesan).
https://doi.org/10.1016/j.camwa.2017.12.013
0898-1221/© 2017 Elsevier Ltd. All rights reserved.
Please cite this article in press as: A. Das, S. Natesan, Higher-order convergence with fractional-step method for singularly perturbed 2D parabolic
convection–diffusion problems on Shishkin mesh, Computers and Mathematics with Applications (2018), https://doi.org/10.1016/j.camwa.2017.12.013.
2 A. Das, S. Natesan / Computers and Mathematics with Applications ( ) –
Nomenclature
u Solution of the model problem (1)

v, w Smooth and singular components of u
ûn+1 Solution of the semidiscrete problem (5)–(6) on ΩtM
ẑ n+1 Solution of the semidiscrete problem (5)–(6) on Ωt2M
ûextpt Solution of the semidiscrete problem (5)–(6) on ΩtM after time extrapolation
n+1/2 n+1/2
v̂ , ŵ Smooth and singular components of ûn+1/2 on ΩtM
n+1/2 n+1/2
ψ̂v , ψ̂w Smooth and singular components of ẑ n+1/2 on Ωt2M
n+1
Û Solution of (16)–(17) on DN with M mesh intervals in the temporal direction
n+1
Ũ Solution of (16)–(17) on D2N with 2M mesh intervals in the temporal direction
U Solution of the fully discrete problem (20)–(21) on GN ,M
n+1/2
V̂ , Ŵ n+1/2 Smooth and singular components of Û n+1/2 on DN with M mesh intervals in the temporal direction
n+1/2
V̂ , Ŵ n+1/2 Smooth and singular components of Û n+1/2 on DN with 2M mesh intervals in the temporal direction
V̂extpt , Ŵextpt Smooth and singular components after time extrapolation on DN with M mesh intervals in the temporal
direction
V̂extpt , Ŵextpt Smooth and singular components after time extrapolation on D2N with M mesh intervals in the
temporal direction
Ûextp Solution of (16)–(17) after both space and time extrapolation on DN with M mesh intervals in the temporal
direction
V̂extp , Ŵextp Smooth and singular components of Ûextp after both space and time extrapolation on DN with M mesh
intervals in the temporal direction
Uextp Final extrapolated solution of the fully discrete problem (20)–(21) on GN ,M
This type of problem occurs in several branches of engineering and applied mathematics, such as skin layers in electrical
application, edge layers in solid mechanics and boundary layers in fluid mechanics [3]. For example, consider the unsteady
incompressible viscous fluid flow problem governed by the Navier–Stokes equation:
⎨ ∂ u + u · ∇ u + ∇ p = 1 ∇ 2 u,
⎧
∂t Re
∇ · u = 0,
⎩
where u is the velocity field whose components are u1 , u2 along x and y directions, and p is the pressure. For sufficiently
large value of the Reynolds number (Re), the above equation behaves like a singular perturbation problem (SPP). One can
observe that the solution of SPP varies rapidly inside the boundary layer region and behaves smoothly in the outer region.
Due to such behavior, classical finite difference methods fail to provide satisfactory numerical result on uniform meshes and
as a remedy one needs to reduce the spatial step size with respect to ε , to obtain a stable solution, which is computationally
expensive.
Two of the most reliable numerical methods for solving such type of problems which are available in the literatures
are fitted operator methods (FOMs) and fitted mesh methods (FMMs). In FOMs, one uses an exponentially fitted scheme,
which has coefficients of exponential type adapted to the SPP. The extension of FOMs to higher-dimensional problems are
too difficult and in some cases it may not be even possible. Whereas, in FMMs, one can use the classical finite difference
schemes on the piecewise-uniform (Shishkin) mesh or any other layer-adapted nonuniform meshes. In FMMs, the meshes
are fine in the boundary layer regions and coarse in the outer region. More information about the layer-adapted nonuniform
meshes and the numerical schemes for SPPs can be found in the books of Farrell et al. [4], Miller et al. [5], and Roos et al. [2].
Due to the presence of boundary layers in SPPs, finding higher-order ε -uniformly convergent numerical solution to SPPs is
indeed a difficult task. To obtain even second-order uniform convergence for the case of convection–diffusion SPPs, one has
to devise the numerical scheme very cautiously. One cannot simply approximate the convection term by central difference
quotient, because it leads to spurious nonphysical oscillations in the numerical solution. Therefore, one has to give some
special attention to obtain second-order convergent numerical solutions for SPPs having convection term. One can surpass
such difficulty by using the hybrid scheme of [6], or by the Richardson extrapolation technique [7–9].
Since, the problem considered in this article is two-dimensional in space, it can model a physical phenomenon more
appropriately. One can find the numerical treatment of such a problem in [10–13], where the authors of [10,11] considered
the stationary case and [12,13] are dedicated for time dependent problem. The authors of [14] solved such problem by
using the backward-Euler scheme for time derivative and the upwind finite difference scheme for spatial derivatives. But
the drawback in this scheme is that, to obtain the numerical solution, one has to handle a banded pentadiagonal matrix
at each time step, which is not very efficient in computational perspective. One can overcome such difficulty by using the
fraction-step method, which converts the 2D problem into two 1D problems. Clavero et al. [13] used this method to solve
A. Das, S. Natesan / Computers and Mathematics with Applications ( ) – 3
singularly perturbed 2D convection–diffusion parabolic PDEs. But in all the cases both the spatial and temporal order of
convergence are almost first-order.
The main aim of this article is to obtain ε -uniformly convergent second-order numerical solution of the model problem
(1) by combining the fractional-step method with the Richardson extrapolation technique. In the literature, we can find
several articles, where the Richardson extrapolation technique has been used for SPPs. To understand about the Richardson
extrapolation in a better way, one can see [7,9]. In [7], the authors considered a two-point BVP for ordinary differential
equations (ODEs) and applied this technique to enhance the order of convergence, whereas Mohapatra and Natesan [15]
applied this technique for a singularly perturbed delay two-point BVPs for ODEs, and Das and Natesan [16] used the same
technique for singularly perturbed delay parabolic IBVPs. In [8], Deb and Natesan adapted this technique for system of
convection–diffusion two-point BVPs for ODEs.
In this work, first, by using the fractional-step method used in [13], we convert the 2D parabolic convection–diffusion
problem (1) to a set of 1D stationary convection–diffusion problems. Then, we discretize the spatial domains by using the
piecewise-uniform Shishkin mesh and we use the upwind finite difference scheme to discretize the spatial derivatives.
We solve those two 1D problems by the upwind scheme. After that, we use the idea of the Richardson extrapolation
technique given in [7]. To apply the extrapolation technique, we double the number of mesh intervals (by fixing the transition
parameter) and solve the same 1D problems. Finally, by combining these two solutions, we obtain a better discrete solution.
This method enhances the order of convergence from first-order to second-order.
The rest of the paper is organized as follows: In Section 2, the time semidiscretization process through fractional-step
method has been discussed. Section 3 contains the extrapolation technique of the semidiscrete solution. Section 4 deals with
the piecewise-uniform Shishkin mesh and the classical upwind scheme. The main proof of convergence has been provided
in Section 5. Numerical results are presented in Section 6 and the paper ends with Section 7, that summarizes the main
conclusions.
Throughout this article, C has been used as a generic positive constant which is independent of ε , the mesh points and
the mesh sizes throughout the paper. Standard supremum norm has been denoted by ∥·∥∞ and is defined by
∥g ∥∞ = sup |g(x, y, t)| ,

(x,y,t)∈G
for a function g defined on some domain G.
2. Time semidiscretization
Here, we describe the time semidiscretization of the singularly perturbed 2D parabolic IBVP (1).
Let us denote the differential operators in x and y directions by
∂2 ∂
Lx,ε ≡ −ε + a1 (x, y) + b1 (x, y),
∂ x2 ∂x
∂2 ∂
Ly,ε ≡ −ε 2 + a2 (x, y) + b2 (x, y),
∂y ∂y
with b(x, y) = b1 (x, y) + b2 (x, y). The operator Lx,ε can be considered as a family of one-dimensional differential operators
with one parameter y ∈ (0, 1) (similarly for Ly,ε ). We split the source term into two terms by f (x, y, t) = f1 (x, y, t) + f2 (x, y, t),
such that
f1 (x, 0, t) = f1 (x, 1, t) = f2 (0, y, t) = f2 (1, y, t) = 0. (2)
Now, we introduce a time semidiscretization process by means of the following fractional-step scheme:
u0 = φ (x, y), (x, y) ∈ D,
(I + ∆t Lx,ε )un+1/2 = un + ∆tf1 (x, y, tn+1 ), y ∈ (0, 1),

{
(3)
un+1/2 (0, y) = un+1/2 (1, y) = 0,
(I + ∆t Ly,ε )un+1 = un+1/2 + ∆tf2 (x, y, tn+1 ), x ∈ (0, 1),
{
(4)
un+1 (x, 0) = un+1 (x, 1) = 0.
The above scheme provides the approximation un (x, y) to the solution u(x, y, t) of (1) at time level tn = n∆t.
The operators (I + ∆t Li,ε ), i = x, y, satisfy the following maximum principle.
Lemma 2.1 (Maximum Principle). Let Ψ (x, y) ≥ 0 on the boundary ∂ D and (I + ∆t Li,ε )Ψ (x, y) ≥ 0, i = x, y on D. Then
Ψ (x, y) ≥ 0 on D.
Proof. First we prove the maximum principle for the operator (I + ∆t Lx,ε ). Let (x∗ , y∗ ) ∈ D be such that Ψ (x∗ , y∗ ) =
minD Ψ (x, y) < 0. It is clear that the point (x∗ , y∗ ) ̸ ∈ ∂ D, which implies (x∗ , y∗ ) ∈ D. Since Ψx (x∗ , y∗ ) = 0 and Ψxx (x∗ , y∗ ) ≥ 0,
we get
(I + ∆t Lx,ε )Ψ (x∗ , y∗ ) = Ψ (x∗ , y∗ ) − ε ∆t Ψxx (x∗ , y∗ ) + ∆ta1 (x∗ , y∗ )Ψx (x∗ , y∗ )

+ ∆tb1 (x∗ , y∗ )Ψ (x∗ , y∗ ) < 0,
which is a contradiction. Hence the required result follows for the operator (I +∆t Lx,ε ). Similarly, we can prove the maximum
principle for the operator (I + ∆t Ly,ε ). □
By using the maximum principle, we can show that (I + ∆t Li,ε )−1 ∞ ≤ 1/(1 + b̃∆t), i = x, y, where b̃ = min{b1 , b2 }.
 
Now, the stability of the semidiscrete scheme (3)–(4) can be ensured from [13].
In order to analyze the convergence of the semidiscrete scheme (3)–(4), we find the local error by ûn+1 (x, y) − u(x, y, tn+1 ),
where ûn+1 (x, y) be the solution of the following system:
un (x, y) = u(x, y, tn ), (x, y) ∈ D,
n+1/2
(I + ∆t Lx,ε )û (x, y) = un (x, y) + ∆tf1 (x, y, tn+1 ), y ∈ (0, 1),
{
n+1/2 n+1/2
(5)
û (0, y) = û (1, y) = 0,
(I + ∆t Ly,ε )ûn+1 (x, y) = ûn+1/2 (x, y) + ∆tf2 (x, y, tn+1 ), x ∈ (0, 1),
{
(6)
ûn+1 (x, 0) = ûn+1 (x, 1) = 0.
Lemma 2.2. The solution of the semidiscrete problem (5) satisfies the following bound
⏐ i n+1/2 ⏐
⏐ ∂ û
⏐ ∂ xi ⏐ ≤ C 1 + ε exp −α x (1 − x)/ε , 0 ≤ i ≤ 4.
⏐ ( −i
( ))
⏐ ⏐
Proof. One can see [13, Appendix A] for the proof. □

Now, we decompose ûn+1/2 = v̂ n+1/2 + ŵ n+1/2 , where v̂ n+1/2 is the smooth component and ŵ n+1/2 is the singular
component, for the purpose of convergence analysis to be carried out in Section 5.
Note that, since, we have the bounds of the derivatives of ûn+1/2 , one can easily obtain the following bounds for the
derivatives of the smooth and singular components:
 i n+1/2 
 ∂ v̂
≤ C 1 + ε 3−i ,
 ( )

 ∂ xi 
 (7)
∞
and
⏐ i n+1/2 ⏐
⏐ ∂ ŵ
⏐ ∂ xi ⏐ ≤ C ε exp(−α x (1 − x)/ε), 0 ≤ i ≤ 4,
⏐ −i
⏐ ⏐ (8)
by following the same idea as used in [6].
3. Extrapolation of û
This section introduces the extrapolation of the time semidiscrete solution û. The error estimate after the extrapolation
technique associated to û is given in this section.
Let ΩtM = {tn = n∆t , n = 0, . . . , M , ∆t = T /M } be the discretization of the time domain Ωt , where M is the number
of mesh points in the t-direction. For the improvement of the semidiscrete solution û(tn ), we need to solve the semidiscrete
problem (5)–(6) on the fine mesh Ωt2M , with 2M number of mesh intervals in the temporal direction. From such construction,
it is clear that ΩtM = {tn } ⊂ Ωt2M = {t̃n }. In fact, Ωt2M is obtained from ΩtM by bisecting each mesh interval of ΩtM . Therefore,
t̃n − t̃n−1 = ∆t /2, for t̃n ∈ Ωt2M . Let ẑ be the solution of the semidiscrete problem (5)–(6) on the mesh Ωt2M . Now, we know
that
û(tn ) − u(tn ) = ∆t + o(∆t), tn ∈ ΩtM . (9)

Similarly, we have
ẑ(t̃n ) − u(t̃n ) = ∆t /2 + o(∆t), t̃n ∈ Ωt2M . (10)

Now, from (9) and (10), we get
u(tn ) − (2ẑ(tn ) − û(tn )) = o(∆t), tn ∈ ΩtM .
We shall therefore use the following extrapolation formula
ûextpt (tn ) = 2ẑ(tn ) − û(tn ), tn ∈ ΩtM , (11)

to obtain a better approximate semidiscrete solution of problem (1).
3.1. Error estimate for ûextpt
To obtain the nodal error |u(x, y, tn ) − ûextpt (x, y, tn )|, first by combining (5) and (6), we get
(I + ∆t Lx,ε )((I + ∆t Ly,ε )ûn+1 (x, y) − ∆tf2 (x, y, tn+1 )) = un (x, y) + ∆tf1 (x, y, tn+1 ),
which leads to
(I + ∆t Lx,ε + ∆t Ly,ε )ûn+1 (x, y) = un (x, y) + ∆tf (x, y, tn+1 ) + (∆t)2 Lx,ε (f2 ) − (∆t)2 Lx,ε Ly,ε ûn+1 .
Now, if we express the parabolic PDE given in (1) in the semidiscrete operator form, we obtain
u(x, y, tn+1 ) − u(x, y, tn ) ∆t
+ (Lx,ε + Ly,ε )u(x, y, tn+1 ) = f (x, y, tn+1 ) − utt (tn+1 ) + O(∆t 2 ),
∆t 2
which implies that
(∆t)2
(I + ∆t Lx,ε + ∆t Ly,ε )u(x, y, tn+1 ) = un (x, y) + ∆tf (x, y, tn+1 ) − utt (tn+1 ) + O(∆t 3 ).
2
Therefore, the local truncation error is given by
(I + ∆t Lx,ε + ∆t Ly,ε )(ûn+1 (x, y) − u(x, y, tn+1 ))

(∆t)2
= −(∆t)2 Lx,ε Ly,ε ûn+1 (x, y) + (∆t)2 Lx,ε (f2 ) +
utt (tn+1 ) + O(∆t 3 ). (12)
2
Now, by following the approach of [7], define the function λ(x, y, tn+1 ), to be the solution of the following BVP:
(I + ∆t Lε ) λ(x, y, tn+1 ) = χ ∆t , in D,
{
(13)
λ(x, y, tn+1 ) = 0, (x, y) ∈ ∂ D,
where χ = −Lx,ε Ly,ε ûn+1 (x, y) + Lx,ε (f2 ) + (1/2)utt (tn+1 ).

By simple calculations, from (13) we obtain
∂ λ(x, y, tn )
( )
+ Lε λ(x, y, tn+1 ) = χ − + O(∆t),
∂t ∆t
which again can be written as a semidiscrete system similar to (5)–(6), whose solution is λ̂. Hence, by following (12), one
can write
(I + ∆t Lε )(λ̂ − λ)(x, y, tn+1 ) ≤ C (∆t)2 . (14)

Now, (12), (13) and (14) together imply that
(I + ∆t Lε )(û − u − ∆t λ̂)(x, y, tn+1 ) = O(∆t 3 ).

Next, by choosing an appropriate barrier function along with the maximum principle ( Lemma 2.1), we can obtain that
û − u − ∆t λ̂ (x, y, tn+1 ) ≤ C (∆t 3 ),
( )
which implies that

(û − u) (x, y, tn+1 ) = ∆t λ̂ + O(∆t 3 ).
For the Richardson extrapolation, we consider the finer mesh in the temporal direction, i.e., we use the temporal mesh
size as ∆t /2. Then, one can have
∆t
(ẑ − u) (x, y, tn+1 ) =λ̂ + O(∆t 3 ).
2
Therefore, by using the formula given in (11) for the Richardson extrapolation technique, we obtain that
u − ûextpt (x, y, tn+1 ) = (u − 2ẑ + û) (x, y, tn+1 )

( )
= (−2(ẑ − u) + (û − u)) (x, y, tn+1 )

= O(∆t 3 ). (15)
4. The discrete problem
In this section, we describe the layer-adapted Shishkin meshes for the spatial directions, then we discretize the spatial
derivatives of the semidiscrete scheme (5)–(6) by using the upwind finite difference scheme.
4.1. Discretization of the domain
N N N N
Let the rectangular mesh D be defined by the tensor product of the 1D Shishkin meshes, i.e., D = I x × I y , which is
constructed as follows:
First, define the transition parameters
{ }
1
ρ l = min , ρ l,0 ε ln N , l = x, y,
2
where ρ l,0 ≥ 2/α l . In the analysis, we shall assume that ρ l = ρ l,0 ε ln N . Note that, if ρ l = 1/2, l = x, y , then the mesh is
uniform and in that case the error estimates can be obtained in the classical way.
N
To obtain the piecewise-uniform Shishkin mesh, we divide the domain I x into two sub-domains [0, 1 −ρ x ] and (1 −ρ x , 1]
and each sub-domain will have N /2 uniform mesh intervals, and denote it by
N
I x = 0 = x0 , x1 , . . . , xN /2 = 1 − ρ x , . . . , xN = 1 .
{ }
N
{ }
Similarly, we define I y = 0 = y0 , y1 , . . . , yN /2 = 1 − ρ y , . . . , yN = 1 .
We denote the mesh sizes in both the spatial directions by
hx,i = xi − xi−1 , i = 1, . . . , N , ĥx,i = hx,i + hx,i+1 , i = 1, . . . , N − 1,

hy,j = yj − yj−1 , j = 1, . . . , N , ĥy,j = hy,j + hy,j+1 , j = 1, . . . , N − 1,
and let Hl = 2(1 − ρ l )/N and hl = 2ρ l /N , l = x, y, be the mesh sizes in [0, 1 − ρ l ] and [1 − ρ l , 1] respectively. Then it is
easy to see that
N −1 ≤ Hl ≤ 2N −1 , hl = 2ρ l,0 ε N −1 ln N , l = x, y.
We define the discrete domain by GN , M = IxN × IyN × ΩtM .
4.2. Numerical scheme
Before describing the numerical scheme, we introduce the difference operators for a given mesh function g(xi , y, tn ) =
gxni ,y , y ∈ IyN as follows:
gxni+1 ,y − gxni ,y gxni ,y − gxni−1 ,y

δx+ gxni ,y = , δx− gxni ,y = ,
hx,i+1 hx,i
2(δx+ gxni ,y − δx− gxni ,y )
and δx2 gxni ,y = .
ĥx,i
In the x-direction, the forward difference operator δx+ , the backward difference operator δx− are the approximations of first-
order derivative and the central difference operator δx2 is the approximation of second-order derivative. Similarly, for a given
mesh function g(x, yj , tn ) = gxn,yj , x ∈ IxN , we define the difference operators δy+ , δy− , and δy2 .
Let LNx,ε be the discretization of the differential operator Lx,ε after replacing the spatial derivatives by the upwind finite
difference scheme on IxN . Then the discretized equation can be written as follows: For y ∈ IyN ,
⎧
n+1/2
⎨(I + ∆t Lx,ε )Ûxi ,y
⎪ N
= (I + ∆t(−εδx2 + a1 (xi , y)δx− + b1 (xi , y)))Ûxni+,y1/2
= u(xi , y, tn ) + ∆tf1 (xi , y, tn+1 ), i = 1, . . . , N − 1, (16)
⎩Û n+1/2 = Û n+1/2 = 0.
⎪
0,y 1,y
Similarly, for x ∈ IxN , we have

⎧
⎨(I + ∆t Ly,ε )Ûx,yj = (I + ∆t(−εδy + a2 (x, yj )δy + b2 (x, yj )))Ûx,yj
N n+1 2 − n+1
⎪
= Ûxn,+yj1/2 + ∆tf2 (x, yj , tn+1 ), j = 1, . . . , N − 1, (17)
⎪
Ûx,0 = Ûx,1 = 0.
⎩ n+1 n+1
After rearranging the terms in (16), we obtain the following tridiagonal system of linear algebraic equations:
⎧
n+1/2 − n+1/2 0 n+1/2 + n+1/2
⎨(I + ∆t Lx,ε )Ûxi ,y ≡ ri Ûxi−1 ,y + ri Ûxi ,y +ri Ûxi+1 ,y = Fxi ,y ,
⎪ N
i = 1, . . . , N − 1, (18)
⎩Û n+1/2 = Û n+1/2 = 0,
⎪
0,y 1,y
where
⎨r − = ∆t − 2ε − a1 (xi , y) , r + = ∆t − 2ε
⎧ ( ) ( )
,
⎪
i i
ĥx,i hx,i hx , i ĥx,i hx,i+1 (19)
ri = 1 + ∆tb1 (xi , y) − ri− − ri+ , Fxi ,y = u(xi , y, tn ) + ∆tf1 (xi , y, tn+1 ).
⎩ 0
⎪
Lemma 4.1. The tridiagonal matrix given in (18) associated with the finite difference scheme (16) is an M-matrix.
Proof. It is clear from Eqs. (18) and (19) that,
ri− < 0, ri+ < 0 and ri0 > 0
along with
|r10 | − |r1+ | > 0, |rN0 −1 | − |rN−−1 | > 0 and |ri0 | − |ri+ | − |ri− | > 0, for i = 1, . . . , N − 1.
Therefore, the matrix associated with the finite difference scheme is an M-matrix. □
As a consequence (see [2]), the difference operator given in (16) satisfies the following discrete maximum principle.
Lemma 4.2 (Discrete Maximum Principle). Assume that the discrete function Ψi satisfies Ψi ≥ 0 on i = 0, N. Then
N
(I + ∆t LNx,ε )Ψi ≥ 0 on IxN implies that Ψi ≥ 0 at each point of I x .
Hence the method is uniformly stable in the supremum norm (see [13]). Similarly, we can prove the above result for the
difference operator given in (17).
4.3. Error estimate for the discrete solution
N ,M
Let us define the fully discrete scheme on G as follows:
Ui0,j = φ (xi , yj ), i, j = 0, . . . , N ,
n+1/2
= Uin,j + ∆tf1 (xi , yj , tn+1 ), 1 ≤ i ≤ N − 1,
{
(I + ∆t LNx,ε )Ui,j
n+1/2 n+1/2
(20)
U0,j = UN ,j = 0, 0 ≤ j ≤ N,
n+1/2
{
(I + ∆t LNy,ε )Uin,j+1 = Ui,j + ∆tf2 (xi , yj , tn+1 ), 1 ≤ j ≤ N − 1,
(21)
Uin,0+1 = Uin,N+1 = 0, 0 ≤ i ≤ N.
The following theorem shows that the upwind scheme defined in (20)–(21), converges ε -uniformly on GN ,M , with almost
first-order accuracy.
Theorem 4.3. Let u be the exact solution of (1) and U be the numerical solution of (20)–(21) at time level tn = n∆t. Then there
exists a positive constant C , independent of ε, N, with 0 < β < 1 such that
u(xi , yj , tn ) − U n  ≤ C (∆t + N −1+β ln N),
 
i,j ∞
for (xi , yj , tn ) ∈ GN , M .
Proof. The proof of this theorem can be found in [13]. □
One can observe from the above theorem that the temporal order of convergence is one and the spatial order of
convergence is almost one (reduced by a logarithmic factor). To enhance the order of convergence, we apply the Richardson
extrapolation technique on the discrete solution Uin,j of the fully discrete scheme (20)–(21).
5. Extrapolation of Û
In this section, we describe the Richardson extrapolation technique in the spatial directions, which is used to increase
the accuracy of the computed solutions of the basic scheme. The main result of this paper for the ε -uniform convergence of
the extrapolated numerical solution is given at the end of this section.
To apply this technique, we will solve the discrete problem (16)–(17) on the fine mesh D2N = Ix2N × Iy2N with 2N mesh
intervals in the spatial direction, where Ix2N and Iy2N are the piecewise-uniform Shishkin meshes having the same transition
point 1 − ρ x and 1 − ρ y respectively, as used in IxN and IyN . The discrete domain Ix2N (similarly Iy2N ) can be obtained through
bisecting each mesh interval of IxN . From such construction, it is clear that DN = {(xi , yj )} ⊂ D2N = {(x̃i , ỹj )}. Hence the
corresponding mesh sizes in D2N are given by
Hx /2, for x̃i ∈ Ix2N ∩ [0, 1 − ρ x ],
{
x̃i − x̃i−1 =
hx /2, for x̃i ∈ Ix2N ∩ [1 − ρ x , 1].
Let Ûxni ,yj be the numerical solution of the discrete problem (16)–(17) on the mesh DN . From [13, Theorem 2.3], we get that
the fractional-step method along with the upwind scheme on the Shishkin mesh yields the error bound as
≤ C ∆tN β−1 ln N ,
 n 
(Û − ûn )
∞
provided N −β < C ∆t, where 0 < β < 1. Therefore, if we replace ∆t in the above inequality by using the condition
N −β < C ∆t, we obtain that
≤ CN −1 ln N .
 n 
(Û − ûn )
∞
Therefore, one can express the error as
(Û n − ûn )(xi , yj ) = C N −1 ln N + o(N −1 ln N)
( )
= C1 N −1 ln N + C2 N −1 ln N + o(N −1 ln N)
( ) ( )
( ) ( )
N
= C1 N −1 (ρ x /ρ x,0 ε ) + C2 N −1 (ρ y /ρ y,0 ε ) + o(N −1 ln N), (xi , yj ) ∈ D , (22)
where C1 and C2 are some fixed constants. Similarly, if Ũx̃n ,ỹ is the solution of the discrete problem (16)–(17) on the mesh
i j
D2N , then we have
( ) ( )
2N
(Ũ n − ûn )(x̃i , ỹj ) = C1 (2N)−1 (ρ x /ρ x,0 ε ) + C2 (2N)−1 (ρ y /ρ y,0 ε ) + o(N −1 ln N), (x̃i , ỹj ) ∈ D . (23)
Now, from (22) and (23), eliminating terms of order O(N −1 ), we get
ûn − 2Ũ n − Û n (xi , yj ) = o(N −1 ln N), (xi , yj ) ∈ DN .
( ( ))
We shall therefore use the following extrapolation formula
Ûextp (xi , yj ) = 2Ũ(xi , yj ) − Û(xi , yj ), (xi , yj ) ∈ DN , (24)

to obtain a better approximate solution to the problem (5)–(6).
To proceed further, we decompose u appears in RHS of (5) into smooth and singular components as v + w , where v is the
n+1/2
smooth component and w is the singular component. In a similar way, we decompose the discrete solution Ûxi ,y of (16)
into smooth and singular components as V̂ n+1/2 + Ŵ n+1/2 in GN , M , and V̂ n+1/2 + Ŵ n+1/2 in GN , 2M .
n+1/2 n+1/2
We decompose ẑ n+1/2 as ẑ n+1/2 = ψ̂v + ψ̂w . The smooth components v̂ n+1/2 and v are further decomposed into
n+1/2 n+1/2 n+1/2 2 n+1/2
v̂ = v̂0 + ε v̂1 + ε v̂2 and v = v0 + εv1 + ε 2 v2 , so that v̂ n+1/2 and v satisfy the following problem:
−ε ∆t v̂ ′′ (x, y, tn+1/2 ) + a1 (x, y)∆t v̂ ′ (x, y, tn+1/2 )

⎧
⎪
⎪
⎪
⎨+(1 + ∆tb1 (x, y))v̂ (x, y, tn+1/2 ) = v (x, y, tn ) + ∆tf1 (x, y, tn+1 ),
⎪
y ∈ (0, 1),
2
n+1/2 n+1/2 n+1/2
⎪ ∑
v̂ , v̂ ε i v̂i (1),
⎪
⎪
⎪
⎩ (0) = 0 (1) =
i=0
and the singular components ŵ n+1/2 and w satisfy the following problem:
⎨−ε∆t ŵ (x, y, tn+1/2 ) + a1 (x, y)∆t ŵ (x, y, tn+1/2 )

′′ ′
⎧
⎪
+(1 + ∆tb1 (x, y))ŵ(x, y, tn+1/2 ) = w (x, y, tn ), y ∈ (0, 1),
⎪
⎩ n+1/2
ŵ (0) = 0, ŵ n+1/2 (1) = ûn+1/2 |x=1 − v̂ n+1/2 (1).
The smooth component V̂ n+1/2 is the solution of the following discrete problem:
n+1/2
= (I + ∆t(−εδx2 + a1 (xi , y)δx− + b1 (xi , y)))V̂ n+1/2
⎧
⎨(I + ∆t Lx,ε )V̂
N
⎪
= v (xi , y, tn ) + ∆tf1 (xi , y, tn+1 ), i = 1, . . . , N − 1,
⎩ n+1/2 n+1/2 n+1/2
= v̂ n+1/2 (1),
⎪
V̂0 = v̂ (0), V̂1
and the singular component Ŵ n+1/2 satisfies

n+1/2
= (I + ∆t(−εδx2 + a1 (xi , y)δx− + b1 (xi , y)))Ŵ n+1/2
⎧
⎨(I + ∆t Lx,ε )Ŵ
N
⎪
= w(xi , y, tn ), i = 1, . . . , N − 1, (25)
n+1/2 n+1/2 n+1/2 n+1/2
⎪
Ŵ0 = ŵ (0), Ŵ1 = ŵ (1).
⎩
At the node (xi , y, tn+1/2 ) the error can be written in the form
(Û − û)(xi , y, tn+1/2 ) = (V̂ − v̂ )(xi , y, tn+1/2 ) + (Ŵ − ŵ )(xi , y, tn+1/2 ).
5.1. Error estimate for the extrapolated smooth component
Lemma 5.1. Assume that ε ≤ N −1 . Then, the local truncation error associated to the smooth component satisfies
hx , i ∂ 2 v̂
(I + ∆t LNx,ε )(V̂ − v̂ )(xi , y, tn+1/2 ) = ∆t a1 (xi , y) (xi , y, tn+1/2 ) + O(Hx2 ).
2 ∂ x2
Proof. By using (7) and ε < N −1 < Hx along with the following Taylor’s expansion,
(I + ∆t LNx,ε )(V̂ − v̂ )(xi , y, tn+1/2 )

ε ∆t ∂ 3 v̂ ∂ 3 v̂
[ ]
= h2x,i+1 3 (ζ1 , y, tn+1/2 ) − h2x,i 3 (ζ2 , y, tn+1/2 )
3(hx,i + hx,i+1 ) ∂x ∂x
hx,i ∂ 2 v̂ h2x,i ∂ 3 v̂
+ ∆t a1 (xi , y)(xi , y, t n+ 1/2 ) − ∆ t a1 (xi , y) 3 (ζ3 , y, tn+1/2 ),
2 ∂x 2 6 ∂x
for some ζ1 ∈ (xi , xi+1 ), ζ2 , ζ3 ∈ (xi−1 , xi ), it is easy to obtain the required result. □
Let us assume that E be the solution of the following BVP:
⎨(I + ∆t L )E = ∆t a (x, y) ∂ v̂ (x, y, t

2
⎧
x,ε 1 n+1/2 ), y ∈ (0, 1),
2 ∂ x2 (26)
E(0) = E(1) = 0.
⎩
Now, we decompose E as E = η + ν , where η is the smooth component and ν is the singular component, which satisfy the
following equations for y ∈ (0, 1):
⎨(I + ∆t L )η = ∆t a (x, y) ∂ v̂ (x, y, t

2
⎧
x,ε 1 n+1/2 ), (I + ∆t Lx,ε )ν = 0,
2 ∂ x2 (27)
η(0) = 0 = ν (0), η(1) = −ν (1).
⎩
Theorem 5.2. The smooth component η defined in (27) satisfies the following bounds
∂ η ∂ η
 k   3 
 ∂ xk  ≤ C , k = 0, 1, 2, and 
 ∂ x3  ≤ C ε .
   −1
∞ ∞
Proof. First, by following [13] along with the bounds of the derivatives of v̂ n+1/2 given in (7), we get that
⏐∂ E ⏐
⏐ i ⏐
⏐ ∂ xi ⏐ ≤ C 1 + ε exp −α x (1 − x)/ε , 0 ≤ i ≤ 2,
−i
( ( ))
⏐ ⏐ (28)
where E is the solution of (26).

Now, we express the smooth component η as η = η0 + εη1 + ε 2 η2 , where ηj , j = 0, 1 are the solutions of
⎨η + ∆ta(x, y)(η ) + b(x, y)η = ∆t a (x, y) ∂ v̂ (x, y, t

2
⎧
0 0 x 0 1 n+1/2 ), y ∈ (0, 1),
2 ∂ x2 (29)
η0 (0) = 0,
⎩
and
η1 + ∆ta(x, y)(η1 )x + b(x, y)η1 = (η0 )xx , y ∈ (0, 1),
{
(30)
η1 (0) = 0,
respectively.
Lastly, the function η2 satisfies
(I + ∆t Lx,ε )η2 = ∆t(η1 )xx , y ∈ (0, 1),

{
(31)
η2 (0) = 0 = η2 (1).
We can see that the problems given in (29) and (30) are independent of ε . Therefore, for η0 and η1 , we can have the following
ε -uniform bound:
 ∂ ηj 
 k 
 ∂ xk  ≤ C , j = 0, 1.
 
∞
Now problem (31) is of the form (26). Hence, the solution η2 will have the same bound as given in (28). Thus, by summing
up the bounds for the derivatives of ηj , j = 0, 1, 2, we have
∂ η
 k 
 ∂ xk  ≤ C (1 + ε ), k = 0, 1, 2, 3.
  2−k
Hence the proof is completed. □
Lemma 5.3. Assume that ε ≤ N −1 . Then, for 1 ≤ i ≤ N − 1, the error in the smooth component satisfies
(V̂ − v̂ )(xi , y, tn+1/2 ) = hx,i η(xi ) + O(N −2 ).
Proof. By using Lemma 5.1 and (27), we have
(I + ∆t LNx,ε )(V̂ − v̂ )(xi , y, tn+1/2 ) = hx,i (I + ∆t Lx,ε )η(xi ) + O(Hx2 ). (32)
Now, by using Taylor’s expansion, it can be obtained easily that
⏐ (I + ∆t LN ) − (I + ∆t Lx,ε ) η(xi )⏐
⏐( ) ⏐
x,ε
 ∂ η  + ∆t hx,i a1 (x, y)  ∂ η 
ε (hx,i + hx,i+1 ) 
[  3   2  ]
≤ ∆t
  
3  ∂x ∞
3  2  ∂ x ∞
2
and by using Theorem 5.2 along with the inequality hx,i ≤ Hx , we obtain that
⏐hx,i (I + ∆t LN ) − (I + ∆t Lx,ε ) η(xi )⏐ ≤ C H 2 ∆t ≤ CH 2 .

⏐ ( ) ⏐ ( )
x,ε x x
(33)
Therefore, (32) and (33) together imply that
(I + ∆t LNx,ε )[(V̂ − v̂ )(xi , y, tn+1/2 ) − hx,i η] ≤ CHx2 .
From the discrete maximum principle (Lemma 4.2), we get that
(V̂ − v̂ )(xi , y, tn+1/2 ) = hx,i η(xi ) + O(N −2 ),
which is the required result. □
Similarly, in the finer mesh of temporal direction, we get that
(V̂ − ψ̂v )(xi , y, tn+1/2 ) = hx,i η(xi ) + O(N −2 ).
Therefore, we have
(V̂extpt − v̂extpt )(xi , y, tn+1/2 ) = hx,i η(xi ) + O(N −2 ),

n+1/2 n+1/2
where V̂extpt and v̂extpt are the time extrapolated solutions of V̂ n+1/2 and v̂ n+1/2 , respectively.
Lemma 5.4. Assume that ε ≤ N −1 . Then, the error after extrapolation associated to the smooth component V̂ n+1/2 satisfies
⏐ V̂extp − v̂extpt (xi , y, tn+1/2 )⏐ ≤ CN −2 .

⏐( ) ⏐
2N ,2M N ,M
Proof. Based on the construction of the meshes in G and G , we have
(Hx /2)η(xi )+ O N −2 ,
⎧ ( )
⎪
for 1 ≤ i ≤ N /2,
⎪
⎨
− v̂extpt (xi , y, tn+1/2 ) =
( )
V̂extpt
(hx /2)η(xi )+ O N −2 ,
( )
⎪
⎪
for N /2 + 1 ≤ i ≤ N − 1,
⎩
n+1/2 2N ,M
where V̂extpt is the time extrapolated solution in G .
Therefore, we have
V̂extp − v̂extpt (xi , y, tn+1/2 )
( )
2V̂extpt − V̂extpt − v̂extpt (xi , y, tn+1/2 )

(( ) )
=
= 2 V̂extpt − v̂extpt − V̂extpt − v̂extpt (xi , y, tn+1/2 )
( ( ) ( ))
= O(N −2 ),
5.2. Error estimate for the extrapolated singular component
Lemma 5.5. The singular component Ŵ n+1/2 satisfies

N ( )−1
n+1/2
∏ α x hx , j
|Ŵ |≤ 1+ , for i = 0, 1, . . . , N − 1.
ε
j=i+1
α x hx , j
( )
∏i
Proof. Let us define Si = 1+ , with S0 = 1.
j=1
ε
Therefore, we have
N ( )−1
Si ∏ α x hx , j
= 1+
SN ε
j=i+1
N
∏
exp −α x hx,j /ε
( )
≥
j=i+1(
= exp −α x (1 − xi )/ε .
)
Let Yi = CSi /SN , for i = 0, 1, . . . , N − 1. Applying the difference operator defined in (16) on Yi , we get
C
(I + ∆t LNx,ε )Yi = (I + ∆t LNx,ε )Si .
SN
Now, we have
(I + ∆t LNx,ε )Si = I + ∆t(−εδx2 + a1 (xi , y)δx− + b1 (xi , y)) Si

( )
= (1 + ∆tb1 (xi , y))Si − ∆t εδx2 Si + ∆ta1 (xi , y)δx− Si . (34)

It is easy to derive that
αx αx −αx2 hx,i
δx− Si = Si , δx+ Si = Si , and εδx2 Si = Si .
(ε + αx hx,i ) ε ĥx,i (ε + αx hx,i )
Therefore, from (34), we get
C
(I + ∆t LNx,ε )Si ≥ Si .
max{ε, hx,i }
Now, by using the bound for the singular component in (25), we obtain
(I + ∆t LNx,ε )Ŵ n+1/2 ≤ C exp −α x (1 − xi )/ε
( )
Si
≤C
SN
≤ (I + ∆t LNx,ε )Yi .
By using the discrete maximum principle ( Lemma 4.2), we obtain
Ŵ n+1/2 ≤ CYi .
With the same argument, we can bound −Ŵ n+1/2 as follows:
− Ŵ n+1/2 ≤ CYi .
Hence, we have
|Ŵ n+1/2 | ≤ CYi ,
which is the required bound. □
Lemma 5.6. Assume that ε ≤ N −1 . Then, the error after extrapolation associated to the singular component Ŵ n+1/2 satisfies
⏐ Ŵextp − ŵextpt (xi , y, tn+1/2 )⏐ ≤ CN −2 , 1 ≤ i ≤ N /2.
⏐( ) ⏐
for
Proof. By using (8), we can get
ŵn+1/2 ≤ C exp(−αx ρx /ε ).
Since, ρx ≥ (2ε/αx ) ln N, it is easy to obtain that ⏐ŵ n+1/2 ⏐ ≤ CN −2 .
⏐ ⏐
Now, by using the similar approach used in [17] on the result obtained in Lemma 5.5, we obtain ⏐Ŵ n+1/2 ⏐ ≤ CN −2 .
⏐ ⏐
Therefore, we have
⏐ Ŵ − ŵ (xi , y, tn+1/2 )⏐ ≤ CN −2 .
⏐( ) ⏐
Similarly, in the finer mesh of temporal direction, we get

⏐ Ŵ − ψ̂w (xi , y, tn+1/2 )⏐ ≤ CN −2 .
⏐( ) ⏐
So, we have
⏐ Ŵextpt − ŵextpt (xi , y, tn+1/2 )⏐ ≤ CN −2 ,
⏐( ) ⏐
n+1/2 n+1/2
where Ŵextpt and ŵextpt are the time extrapolated solutions of Ŵ n+1/2 and ŵ n+1/2 , respectively. Thus, we have
⏐ Ŵextp − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ) ⏐
= ⏐ 2Ŵextpt − Ŵextpt − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ) ⏐
= ⏐ 2 Ŵextpt − ŵextpt − Ŵextpt − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ( ) ( )) ⏐
≤ CN −2 ,
To analyze the effect of the extrapolation for xi ∈ (1 − ρx , 1], we use the Taylor series expansion, i.e.,
I + ∆t LNx,ε Ŵ − ŵ (xi , y, tn+1/2 )

( )( )
I + ∆t Lx,ε − I + ∆t LNx,ε ŵ(xi , y, tn+1/2 )

(( ) ( ))
=
ε ∂ ŵ
∆th2x 4
∂ 4 ŵ
[ ]
=− ( ξ1 , y, tn+1/2 ) + ( ξ2 , y , t n+1/2 )
4! ∂ x4 ∂ x4
hx ∆t ∂ ŵ
2
h ∆t
2
∂ 3 ŵ
+ a1 (xi , y) 2 (xi , y, tn+1/2 ) − x a1 (xi , y) 3 (ξ3 , y, tn+1/2 ),
2 ∂x 3! ∂x
where ξ1 ∈ (xi , xi+1 ) and ξ2 , ξ3 ∈ (xi−1 , xi ).
Now, by using (8), we can write
I + ∆t LNx,ε Ŵ − ŵ (xi , y, tn+1/2 )

( )( )
2ε ∆t ( −1 ∂ 2 ŵ
N ln N a1 (xi , y) 2 (xi , y, tn+1/2 ) + O(ε −1 exp(−αx (1 − xi+1 )/ε )N −2 ln2 N).
)
=
αx ∂x
We consider a BVP, when x ∈ (1 − ρx , 1),
2ε ∂ 2 ŵ
I + ∆t Lx,ε F1 = ∆ta1 (x, y) (x, y, tn+1/2 ), y ∈ (0, 1),
( )
αx ∂ x2
F1 (1 − ρx ) = F1 (1) = 0.
Since, this is an ODE, we can obtain
(Ŵ − ŵ )(xi , y, tn+1/2 ) = (N −1 ln N)F1 (xi , y) + O(N −2 ln2 N),
by following the idea of [7].

Similarly, in the finer mesh of temporal direction, we get that
(Ŵ − ψ̂w )(xi , y, tn+1/2 ) = (N −1 ln N)F1 (xi , y) + O(N −2 ln2 N).
Now, one can easily obtain that
Ŵextpt − ŵextpt (xi , y, tn+1/2 ) = (N −1 ln N)F1 (xi , y) + O(N −2 ln2 N).

( )
So,
⏐ Ŵextp − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ) ⏐
= ⏐ 2Ŵextpt − Ŵextpt − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ) ⏐
= ⏐ 2 Ŵextpt − ŵextpt − Ŵextpt − ŵextpt (xi , y, tn+1/2 )⏐

⏐( ( ) ( )) ⏐
≤ CN −2 ln2 N .
Therefore, we can state the following lemma.
Lemma 5.7. The error after extrapolation associated to the singular component Ŵ n+1/2 satisfies
⏐ Ŵextp − ŵextpt (xi , y, tn+1/2 )⏐ ≤ C N −2 ln2 N , N /2 + 1 ≤ i ≤ N − 1.

⏐( ) ⏐ ( )
for
Now, by combining the results of Lemmas 5.4, 5.6 and 5.7, we obtain that
⏐( ⏐
⏐ n+1/2 n+1/2
)
⏐ ûextpt − Ûextp (xi , y)⏐ ≤ CN −2 ln2 N . (35)
⏐
The semidiscrete problems (5) and (6) are almost of the same type except the first term in the RHS of both the equations,
i.e., the RHS of (6) contains ûn+1/2 instead of un as in (5). Therefore, to obtain the similar error estimate for the solution of
(6) and (17), first, we decompose ûn+1/2 , ûn+1 , Û n+1/2 and Û n+1 of (6) and (17) into the smooth and singular components.
Then, by using their error bounds as obtained in the previous case and proceeding in a similar way as done before, we can
obtain the following error bound:
⏐( ) ⏐
(x, yj )⏐ ≤ CN −2 ln2 N , for 1 ≤ j ≤ N − 1.
⏐ n+1 n+1
⏐ ûextpt − Ûextp (36)
⏐
The main result of this paper is stated in the following theorem, which demonstrates that, by using the idea of the
Richardson extrapolation technique [7], we are able to obtain almost second-order ε -uniform convergence in space and
time, which is optimal compared to the result obtained in [13].
Theorem 5.8 (Error after Extrapolation). Assume that ε ≤ N −1 . Let u be the solution of the continuous problem (1) and Uextp
be the solution obtained via the Richardson extrapolation technique, by solving the fully discrete scheme (20)–(21) at time level
tn = n∆t on two meshes GN ,M and G2N ,2M . Then, for 0 < β < 1, we have the following error bound associated with Uextp :
u(xi , yj , tn ) − Uextp (xi , yj , tn ) ≤ C N −2+β ln2 N + ∆t 2 , for (xi , yj , tn ) ∈ GN ,M .

  ( )
∞
Proof.
 u − Uextp (xi , yj , tn )
( ) 
∞
=  u − ûextpt + ûextpt − Ûextp + Ûextp − Uextp (xi , yj , tn )∞
( ) 
≤  u − ûextpt (xi , yj , tn )∞ +  ûextpt − Ûextp (xi , yj , tn )∞ +  Ûextp − Uextp (xi , yj , tn )∞
( )  ( )  ( ) 
≤ C ∆t + N −2 ln2 N +  Ûextp − Uextp (xi , yj , tn ) (by using (15) and (36)).

( 3 )  ( ) 
∞
Note that, if we take N −β ≤ C ∆t with 0 < β < 1, we can deduce that
 u − Uextp (xi , yj , tn ) = C ∆t ∆t 2 + N −2+β ln2 N +  Ûextp − Uextp (xi , yj , tn )∞ .

( )  ( ) ( ) 
∞
(37)
Then, by using the stability of the fully discrete scheme, it can be obtained that
 
 Ûextp − Uextp (xi , yj , tn ) ≤ u(xi , yj , tn−1 ) − Uextp nxi−,y1j  .
( )   
∞ ∞
Hence, (37) can be written as
u(xi , yj , tn ) − Uextp (xi , yj , tn ) ≤ C N −2+β ln2 N + ∆t 2 ,

  ( )
∞
which is the required error estimate. □
6. Numerical results
To verify the theoretical result proved in the previous section, in this section we show the numerical results obtained
by the proposed higher-order method for two 2D test problems. In the tables, we start with N = 32 and the time step
∆t = 0.025 and we multiply N by two and divide ∆t by two. Note that in the following examples, we decompose the source
term f (x, y, t) in the form
f2 (x, y, t) = f (x, 0, t) + y (f (x, 1, t) − f (x, 0, t)) ,

f1 (x, y, t) = f (x, y, t) − f2 (x, y, t),
so that the property (2) is satisfied under the following compatibility condition:
f (0, 0, t) = f (0, 1, t) = f (1, 0, t) = f (1, 1, t) = 0, for t ∈ [0, T ].
Example 6.1. Consider the following singularly perturbed 2D parabolic IBVP with constant coefficients:
⎨ut − ε∆u + ux + uy = f (x, y, t), (x, y, t) ∈ D × (0, 1],

⎧
u(x, y, 0) = φ (x, y), (x, y) ∈ D,

u(x, y, t) = 0, (x, y, t) ∈ ∂ D × [0, 1].
⎩
We choose the initial data φ (x, y) and the source function f (x, y, t) to fit with the exact solution
u(x, y, t) = (1 − exp(−t)) (m1 + m2 x + exp(−(1 − x)/ε )) (m1 + m2 y + exp(−(1 − y)/ε )) ,
where m1 = − exp(−1/ε ), m2 = −1 − m1 . We calculate the maximum pointwise error for each ε by
eNε ,∆t = ⏐u(xi , yj , tn ) − U(xi , yj , tn )⏐ , (before extrapolation),

⏐ ⏐
max
(xi ,yj ,tn )∈GN ,M
and
N ,∆ t ⏐u(xi , yj , tn ) − Uextp (xi , yj , tn )⏐ , (after extrapolation),
⏐ ⏐
eε, extp = max
(xi ,yj ,tn )∈GN ,M
where u(xi , yj , tn ), U(xi , yj , tn ) and Uextp (xi , yj , tn ) denote the exact solution, the numerical solution obtained before extrap-
olation and the numerical solution obtained after extrapolation in GN ,M with N mesh intervals in the spatial direction and
M mesh intervals in the temporal direction, such that ∆t = T /M is the uniform time step. We determine the corresponding
order of convergence for each ε by
eεN ,∆t
( )
pNε ,∆t = log2 2N ,∆t /2
, (before extrapolation),
eε
and
N ,∆t
( )
N ,∆t eε, extp
pε, extp = log2 2N ,∆t /2
, (after extrapolation).
eε, extp
Now, for each N and ∆t, we define the ε -uniform maximum pointwise error by
eN ,∆t = max eNε ,∆t , (before extrapolation),

ε
and
N ,∆t N ,∆t
eextp = max eε, extp , (after extrapolation),
ε
and the corresponding ε -uniform order of convergence by
e N ,∆ t
( )
pN ,∆t = log2 , (before extrapolation),
e2N ,∆t /2
and
N ,∆ t
( )
N ,∆t eextp
pextp = log2 2N ,∆t /2
, (after extrapolation).
eextp
Next, we consider an example with variable convection coefficients.
Table 1
Maximum pointwise errors of the solution and the corresponding order of convergence for Example 6.1.
ε Extrapolation Number of mesh intervals N
32 64 128 256 512
Before 2.6781e−3 1.4181e−3 7.3066e−4 3.7095e−4 1.8692e−4
0.9172 0.9567 0.9780 0.9888
100
After 1.8172e−4 5.0401e−5 1.3293e−5 3.4157e−6 8.6575e−7
1.8502 1.9227 1.9604 1.9802
Before 6.0581e−2 3.8660e−2 2.3563e−2 1.3843e−2 7.9089e−3
10−2 0.6480 0.7143 0.7674 0.8076
After 6.1159e−3 2.4952e−3 9.1768e−4 3.1554e−4 1.0285e−4
1.2934 1.4431 1.5402 1.6173
Before 6.4493e−2 4.0947e−2 2.4830e−2 1.4557e−2 8.3145e−3
−8 0.6554 0.7217 0.7703 0.8080
10
After 6.6689e−3 2.6758e−3 9.7337e−4 3.3286e−4 1.0816e−4
1.3175 1.4589 1.5481 1.6217
eN , ∆ t Before 6.4493e−2 4.0947e−2 2.4830e−2 1.4557e−2 8.3145e−3
pN , ∆ t 0.6554 0.7217 0.7703 0.8080
N , ∆t
eextp After 6.6689e−3 2.6758e−3 9.7341e−4 3.3287e−4 1.0816e−4
N , ∆t
pextp 1.3175 1.4588 1.5481 1.6217
Example 6.2. Consider the following singularly perturbed 2D parabolic IBVP:
ut − ε ∆u + (1 + x)ux + (2 − y)uy + (x2 + y2 + 1)u = f (x, y, t),

⎧
⎪
(x, y, t) ∈ D × (0, 1],
⎪
⎨
⎪
⎪ u(x, y, 0) = φ (x, y), (x, y) ∈ D,
u(x, y, t) = 0, (x, y, t) ∈ ∂ D × [0, 1].
⎩
We choose the initial data φ (x, y) and the source function f (x, y, t) to fit with the exact solution
u(x, y, t) = exp(−t)xy(q1 (x) − 1)(q2 (y) − 1),
with
q1 (x) = exp(−(3 − 2x − x2 )/(2ε )) and q2 (y) = exp(−(3 − 4y + y2 )/(2ε )).
We calculate the maximum pointwise error and the corresponding order of convergence in a similar way as discussed
earlier. The calculated maximum pointwise errors and the corresponding order of convergence for Examples 6.1 and 6.2
are presented in Tables 1 and 2, respectively, for various values of ε and N. It can be seen that, in Tables 1 and 2, the
mesh resolution condition (ε ≤ 1/N) used for the theoretical analysis is not being satisfied for certain values. From the
construction of the Shishkin mesh it is assured that in such cases the piecewise-uniform mesh will become uniform mesh.
From the results presented in Tables 1 and 2, we can observe that for fixed ε , the maximum pointwise errors computed before
and after extrapolation, decrease monotonically as N increases, which confirms that the proposed method is ε -uniformly
convergent. Moreover, one can observe that the maximum pointwise errors and the corresponding order of convergence
have been improved after the extrapolation technique. To visualize the appearance of the boundary layer and its behavior
for different ε , we have given the surface plots of the solutions of Example 6.1 for ε = 10−2 , 10−8 and N = 32 in Fig. 1.
In order to reveal the numerical order of convergence, we have plotted the maximum pointwise errors (before and after
extrapolation) in loglog scale in Figs. 2 and 3, which again ensures the usefulness of the extrapolation technique.
7. Conclusions
In this paper, we have solved the singularly perturbed 2D parabolic convection–diffusion problem of the form (1), using
the uniform mesh for the temporal domain and the piecewise-uniform Shishkin mesh for the spatial domains. To discretize
the time derivative, we have used the fractional-step method then to solve the resulting 1D stationary problems, we have
used the upwind finite difference scheme. To obtain second-order uniformly convergent numerical solution of (1), we have
applied the Richardson extrapolation technique. For that, we have obtained the numerical solution of the discrete problems
on two embedded meshes, first with N + 1 and M + 1 number of mesh-points in the spatial (both x and y) and temporal
directions, respectively, then with 2N + 1 and 2M + 1 number of mesh-points (by fixing the transition parameter in the
spatial directions), then we have calculated the extrapolated solution by using these two numerical results. Theoretically
we have proved that the extrapolation provides almost second-order (up to a logarithmic factor) ε -uniform convergence.
Numerical experiments are carried out to verify the theoretical findings. As a practical interest, for our future work, we
shall try to discretize the spatial domain with the Bakhvalov–Shishkin mesh to overcome the reduction of order, due to the
logarithmic factor.
(a) ε = 1e − 2.
(b) ε = 1e − 8.
Fig. 1. Surface plots of the numerical solutions at t = 1 and N = 32 for Example 6.1.
100
10-1
10-2
10-3
10-4
101 102 103
Fig. 2. Visualization of the order of convergence through loglog plot for Example 6.1.
100
10-1
10-2
10-3
10-4 1
10 102 103
Fig. 3. Visualization of the order of convergence through loglog plot for Example 6.2.
Table 2
Maximum pointwise errors of the solution and the corresponding order of convergence for Example 6.2.
ε Extrapolation Number of meshintervals N
32 64 128 256 512
Before 5.7732e−5 2.9180e−5 1.4673e−5 7.3567e−6 3.6835e−6
0.9844 0.9918 0.9960 0.9980
100
After 7.1978e−7 1.8985e−7 4.8886e−8 1.2410e−8 3.1270e−9
1.9227 1.9573 1.9779 1.9886
Before 1.2074e−1 8.2246e−2 5.2326e−2 3.1338e−2 1.8177e−2
−2 0.5539 0.6524 0.7396 0.7858
10
After 2.5458e−2 1.2441e−2 5.6588e−3 2.2416e−3 7.8676e−4
1.0330 1.1365 1.3359 1.5106
Before 1.3376e−1 9.0620e−2 5.7751e−2 3.4580e−2 2.0040e−2
−8 0.5618 0.6500 0.7399 0.7871
10
After 2.1302e−2 8.9105e−3 3.5751e−3 1.3057e−3 4.3800e−4
1.2574 1.3175 1.4532 1.5758
eN , ∆ t Before 1.3376e−1 9.0620e−2 5.7751e−2 3.4580e−2 2.0040e−2
pN , ∆ t 0.5618 0.6500 0.7399 0.7871
N , ∆t
eextp After 2.5458e−2 1.2441e−2 5.6588e−3 2.2416e−3 7.8676e−4
N , ∆t
pextp 1.0330 1.1365 1.3359 1.5106
Acknowledgments
The authors wish to acknowledge the referees for their valuable comments and suggestions, which helped to improve
the presentation.
References
[1] O.A. Ladyženskaja, V.A. Solonnikov, N.N. Ural´ceva, Linear and Quasilinear Equations of Parabolic Type, American Mathematical Society, Providence,
R.I., 1968.
[2] H.-G. Roos, M. Stynes, L. Tobiska, Robust Numerical Methods for Singularly Perturbed Differential Equations, Springer-Verlag, Berlin, 2008.
[3] A.H. Nayfeh, Introduction to Perturbation Techniques, John Wiley & Sons, New York, 1981.
[4] P.A. Farrell, A.F. Hegarty, J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Robust Computational Techniques for Boundary Layers, Chapman & Hall/CRC, Boca
Raton, FL, 2000.
[5] J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems, World Scientific Publishing Co. Pvt. Ltd.,
Singapore, 2012.
[6] A. Das, S. Natesan, Uniformly convergent hybrid numerical scheme for singularly perturbed delay parabolic convection–diffusion problems on Shishkin
mesh, Appl. Math. Comput. 271 (2015) 168–186.
[7] M.C. Natividad, M. Stynes, Richardson extrapolation for a convection–diffusion problem using a Shishkin mesh, Appl. Numer. Math. 45 (2–3) (2003)
315–329.
[8] B.S. Deb, S. Natesan, Richardson extrapolation method for singularly perturbed coupled system of convection–diffusion boundary-value problems,
CMES Comput. Model. Eng. Sci. 38 (2) (2008) 179–199.
[9] K. Mukherjee, S. Natesan, Richardson extrapolation technique for singularly perturbed parabolic convection–diffusion problems, Computing 92 (1)
(2011) 1–32.
[10] E. O’Riordan, M.L. Pickett, A parameter-uniform numerical method for a singularly perturbed two parameter elliptic problem, Adv. Comput. Math.
35 (1) (2011) 57–82.
[11] T. Linß, M. Stynes, Asymptotic analysis and Shishkin-type decomposition for an elliptic convection–diffusion problem, J. Math. Anal. Appl. 261 (2)
(2001) 604–632.
[12] B. Bujanda, C. Clavero, J.L. Gracia, J.C. Jorge, A high order uniformly convergent alternating direction scheme for time dependent reaction–diffusion
singularly perturbed problems, Numer. Math. 107 (1) (2007) 1–25.
[13] C. Clavero, J.C. Jorge, F. Lisbona, G.I. Shishkin, A fractional step method on a special mesh for the resolution of multidimensional evolutionary
convection–diffusion problems, Appl. Numer. Math. 27 (3) (1998) 211–231.
[14] C. Clavero, J.C. Jorge, Another uniform convergence analysis technique of some numerical methods for parabolic singularly perturbed problems,
Comput. Math. Appl. 70 (3) (2015) 222–235.
[15] J. Mohapatra, S. Natesan, Uniformly convergent second-order numerical method for singularly perturbed delay differential equations, Neural Parallel
Sci. Comput. 16 (3) (2008) 353–370.
[16] A. Das, S. Natesan, Second-order uniformly convergent numerical method for singularly perturbed delay parabolic partial differential equations, Int. J.
Comput. Math. http://dx.doi.org/10.1080/00207160.2017.1290439.
[17] M. Stynes, H.-G. Roos, The midpoint upwind scheme, Appl. Numer. Math. 23 (3) (1997) 361–374.

Das 2018

Uploaded by

Copyright:

Available Formats

Das 2018

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Das 2018

Uploaded by

Copyright:

Available Formats

Computers and Mathematics with Applications ( ) –

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

Higher-order convergence with fractional-step method for

⎨ut + Lε u(x, y, t) = f (x, y, t), (x, y, t) ∈ G,

u(x, y, 0) = φ (x, y), (x, y) ∈ D, (1)

Lε u = −ε ∆u + a(x, y).∇ u + b(x, y)u,

u Solution of the model problem (1)

∥g ∥∞ = sup |g(x, y, t)| ,

for a function g defined on some domain G.

f1 (x, 0, t) = f1 (x, 1, t) = f2 (0, y, t) = f2 (1, y, t) = 0. (2)

u0 = φ (x, y), (x, y) ∈ D,

(I + ∆t Lx,ε )un+1/2 = un + ∆tf1 (x, y, tn+1 ), y ∈ (0, 1),

(I + ∆t Lx,ε )Ψ (x∗ , y∗ ) = Ψ (x∗ , y∗ ) − ε ∆t Ψxx (x∗ , y∗ ) + ∆ta1 (x∗ , y∗ )Ψx (x∗ , y∗ )

Proof. One can see [13, Appendix A] for the proof. □

by following the same idea as used in [6].

û(tn ) − u(tn ) = ∆t + o(∆t), tn ∈ ΩtM . (9)

ẑ(t̃n ) − u(t̃n ) = ∆t /2 + o(∆t), t̃n ∈ Ωt2M . (10)

ûextpt (tn ) = 2ẑ(tn ) − û(tn ), tn ∈ ΩtM , (11)

3.1. Error estimate for ûextpt

(I + ∆t Lx,ε + ∆t Ly,ε )(ûn+1 (x, y) − u(x, y, tn+1 ))

where χ = −Lx,ε Ly,ε ûn+1 (x, y) + Lx,ε (f2 ) + (1/2)utt (tn+1 ).

(I + ∆t Lε )(λ̂ − λ)(x, y, tn+1 ) ≤ C (∆t)2 . (14)

(I + ∆t Lε )(û − u − ∆t λ̂)(x, y, tn+1 ) = O(∆t 3 ).

which implies that

u − ûextpt (x, y, tn+1 ) = (u − 2ẑ + û) (x, y, tn+1 )

= (−2(ẑ − u) + (û − u)) (x, y, tn+1 )

4. The discrete problem

4.1. Discretization of the domain

hx,i = xi − xi−1 , i = 1, . . . , N , ĥx,i = hx,i + hx,i+1 , i = 1, . . . , N − 1,

We define the discrete domain by GN , M = IxN × IyN × ΩtM .

4.2. Numerical scheme

gxni+1 ,y − gxni ,y gxni ,y − gxni−1 ,y

Similarly, for x ∈ IxN , we have

Proof. It is clear from Eqs. (18) and (19) that,

ri− < 0, ri+ < 0 and ri0 > 0

4.3. Error estimate for the discrete solution

Proof. The proof of this theorem can be found in [13]. □

We shall therefore use the following extrapolation formula

Ûextp (xi , yj ) = 2Ũ(xi , yj ) − Û(xi , yj ), (xi , yj ) ∈ DN , (24)

−ε ∆t v̂ ′′ (x, y, tn+1/2 ) + a1 (x, y)∆t v̂ ′ (x, y, tn+1/2 )

⎨−ε∆t ŵ (x, y, tn+1/2 ) + a1 (x, y)∆t ŵ (x, y, tn+1/2 )

and the singular component Ŵ n+1/2 satisfies

(Û − û)(xi , y, tn+1/2 ) = (V̂ − v̂ )(xi , y, tn+1/2 ) + (Ŵ − ŵ )(xi , y, tn+1/2 ).

5.1. Error estimate for the extrapolated smooth component

(I + ∆t LNx,ε )(V̂ − v̂ )(xi , y, tn+1/2 )

⎨(I + ∆t L )E = ∆t a (x, y) ∂ v̂ (x, y, t

⎨(I + ∆t L )η = ∆t a (x, y) ∂ v̂ (x, y, t

where E is the solution of (26).

⎨η + ∆ta(x, y)(η ) + b(x, y)η = ∆t a (x, y) ∂ v̂ (x, y, t

(I + ∆t Lx,ε )η2 = ∆t(η1 )xx , y ∈ (0, 1),

Hence the proof is completed. □

(V̂ − v̂ )(xi , y, tn+1/2 ) = hx,i η(xi ) + O(N −2 ).

Proof. By using Lemma 5.1 and (27), we have

(I + ∆t LNx,ε )(V̂ − v̂ )(xi , y, tn+1/2 ) = hx,i (I + ∆t Lx,ε )η(xi ) + O(Hx2 ). (32)

Now, by using Taylor’s expansion, it can be obtained easily that

⏐hx,i (I + ∆t LN ) − (I + ∆t Lx,ε ) η(xi )⏐ ≤ C H 2 ∆t ≤ CH 2 .

Therefore, (32) and (33) together imply that

(I + ∆t LNx,ε )[(V̂ − v̂ )(xi , y, tn+1/2 ) − hx,i η] ≤ CHx2 .