You are on page 1of 14

Computers and Mathematics with Applications 70 (2015) 222–235

Contents lists available at ScienceDirect

Computers and Mathematics with Applications


journal homepage: www.elsevier.com/locate/camwa

Another uniform convergence analysis technique


of some numerical methods for parabolic singularly
perturbed problems
C. Clavero a,∗ , J.C. Jorge b
a
IUMA and Department of Applied Mathematics, University of Zaragoza, Zaragoza, Spain
b
Institute of Smart Cities and Departamento de Ingeniería Matemática e Informática, Universidad Pública de Navarra, Pamplona, Spain

article info abstract


Article history: In this work we analyze the uniform convergence of some monotone finite difference
Received 6 October 2014 schemes which are used to semidiscretize in space time dependent two dimensional
Received in revised form 26 February 2015 singularly perturbed parabolic problems of convection–diffusion or reaction–diffusion
Accepted 5 April 2015
type. The analysis technique combines a suitable semidiscrete maximum principle joint
Available online 27 April 2015
to some techniques used in the study of the convergence of well known numerical
schemes developed to solve elliptic singularly perturbed problems. We focus our attention
Keywords:
Singular perturbation
to standard central or upwind finite difference schemes defined on special nonuniform
Discrete maximum principles meshes of Shishkin type. Nevertheless, the technique can be easily extended to other
Special meshes classical discretization methods like, for example, fitted operator methods. We prove that
Uniform convergence the stiff initial value problems resulting of the spatial semidiscretization processes, have
a unique solution which converges uniformly with respect to the singular perturbation
parameter. To obtain an efficient numerical algorithm, such initial value problems are
discretized by using appropriate time integrators; here, we have chosen the Implicit Euler
method for doing this, obtaining an unconditional and uniformly convergent numerical
method in a simple way. Finally, some numerical results are shown in order to illustrate
the efficiency of the numerical methods, according to the theoretical results.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction

In this paper we are concerned with the efficient numerical approximation of convection–reaction–diffusion phenomena,
modeled through partial differential equations of type
∂u −

− ε 1u + − →
v (x, y, t ) · ∇ u + k(x, y, t )u = f (x, y, t ),
∂t


where ε is the positive diffusion parameter, which can take very small values, compared to the sizes of v or k.
To not enlarge the paper in excess, here we only study two typical cases; nevertheless, we can extend the use of this
technique to many other situations. Such cases are the following ones:
(1) (convection–diffusion). We consider a smooth vector function v (x, y, t ) ≡

→ (v1 (x, y, t ), v2 (x, y, t )), satisfying
vi (x, y, t ) ≥ v > 0, i = 1, 2; as well, k(x, y, t ), is a smooth non negative function.
(2) (diffusion–reaction). Now vi (x, y, t ) ≡ 0, i = 1, 2 and k(x, y, t ) ≥ k > 0.

∗ Corresponding author. Tel.: +34 976761982; fax: +34 976761886.


E-mail addresses: clavero@unizar.es (C. Clavero), jcjorge@unavarra.es (J.C. Jorge).

http://dx.doi.org/10.1016/j.camwa.2015.04.006
0898-1221/© 2015 Elsevier Ltd. All rights reserved.
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 223

It is well known that, in general, the solution of these problems has a multiscale character even for smooth data, changing
rapidly in certain narrow regions called layers (see [1–3]). Such behavior causes very inaccurate numerical approximations
if standard finite difference or finite element methods are used on uniform meshes, unless a large number (ε -dependent)
of mesh points is considered. This drawback, which appears even for the simplest singular perturbation problems (one-
dimensional and stationary), has led to the development of adaptive techniques, capable of producing good approximations,
even inside the layers, using meshes with a number of grid points which does not depend of the size of ε . In this context,
the uniform convergence is the key property, since it implies that the rate of convergence and the constant’s error of the
method are both independent of the singular perturbation parameter ε .
To construct uniformly convergent schemes, two main different ways have been used in last years, both in the contexts
of finite difference and finite element methods. On one hand, the named fitted operator methods, which have coefficients
of exponential type capable of reproducing the multiscale character of the exact solution of some singularly perturbed
problems (see [4–7,3]); on the other one, the class of fitted mesh methods (see [8–11,2,12–14]), which concentrates
appropriately the grid points in narrow regions containing the boundary layers. In these two ways, several schemes for
stationary and time dependent problems have been deeply studied.
Numerical methods for 2D singularly perturbed elliptic problems have been developed and analyzed in many papers
(see [15–17], for instance, and references therein). For 2D time dependent problems, in [18–21] the fully discrete method
is defined in two steps, discretizing first in time and later on discretizing in space the 1D stationary problems resulting of
the first step. Following this idea, the direction alternating technique (see [22]) is used and, consequently, a remarkable
reduction in computational cost is achieved with respect to classical implicit methods because only tridiagonal linear
systems must be solved. Nevertheless, in the analysis of the uniform convergence there are some drawbacks related to the
uniform stability of the time discretization process and, sometimes, a ratio between the discretization parameters is needed
to prove theoretically the uniform convergence of the scheme. Such restriction does not appear in the numerical experiments
performed with the numerical methods proposed therein. Because of this, in this paper we consider an alternative technique
of analysis, where the discretization process is studied in the inverse order, i.e., we firstly discretize in space and, later on,
we integrate in an appropriate way the resulting stiff Initial Value Problems (henceforth IVP).
Following this idea, we prove that the main results deduced for the schemes used to solve stationary problems, can be
extended successfully to semidiscrete in space parabolic problems. We focus the attention on the fact that a discretization
of a stationary problem and a spatial semidiscretization of the corresponding time dependent model, have many things in
common. More concretely, we prove that standard techniques used in the analysis of the uniform convergence, by combining
uniform consistency and stability properties, or nonuniform consistency together with the barrier function technique, are
easily adapted to the analysis of the uniform convergence for the same type of spatial semidiscretizations.
The paper is structured as follows: in Section 2, we introduce the spatial discretization of the continuous problem on a
special nonuniform mesh of Shishkin type. In Section 3 we analyze the uniform convergence of the spatial discretization,
proving almost first order for the convection–diffusion problem and almost second order for the reaction–diffusion case.
In Section 4 we define the time discretization proving also its uniform convergence. Finally, in Section 5 some numerical
results are shown, illustrating the efficiency of the method and providing numerical orders of convergence as expected by
theory.
Henceforth, C denotes a generic positive constant independent of the diffusion parameter ε and also of the discretization
parameters N and M. We always use the pointwise maximum norm and denote it by ∥ · ∥D (where D is the corresponding
domain).

2. Spatial semidiscretizations

For simplicity, let us consider the unit square Ω ≡ (0, 1)×(0, 1) as the spatial domain and the following initial–boundary
value problem

∂u
+ Lε (t )u = f , in Ω × (0, T ],
∂t (1)
u(x, y, 0) = ϕ(x, y), in Ω ,
u(x, y, t ) = g (x, y, t ), in ∂ Ω × [0, T ],

where the spatial differential operator Lε is given by




Lε (t )u ≡ −ε 1u + −

v (x, y, t ) · ∇ u + k(x, y, t )u. (2)

We assume that the data of the problem are sufficiently smooth functions and also that sufficient compatibility conditions
hold among them in order to u(x, y, t ) ∈ C 4,2 (Ω × (0, T ]), i.e., it has continuous derivatives up to fourth order in space and
second order in time (see [20,21] and references therein). For instance, typical smoothness assumptions, for the right-hand
side, the boundary and the initial condition, are

f ∈ C 2+2α,1+α (Ω̄ × [0, T ]), g ∈ C 4,2 (∂ Ω × [0, T ]), ϕ ∈ C 4 (Ω̄ ), (0 < α < 1). (3)
224 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

With respect to the compatibility conditions, we assume that

g (x, y, 0) = ϕ(x, y) in ∂ Ω ,
∂g
(x, y, 0) = −Lε (0)ϕ + f (x, y, 0), in ∂ Ω ,
∂t
∂ 2g ∂f (4)
(x, y, 0) = (Lε (0))2 ϕ − Lε (0)f (x, y, 0) + (x, y, 0), in ∂ Ω ,
∂t2 ∂t
∂g
+ Lε (g ) = f , in {0, 1} × {0, 1} × (0, T ].
∂t
A finite difference spatial semidiscretization of (1) permits to construct approximations of u(xi , yj , t ), where (xi , yj ) are
the grid points of a rectangular mesh Ω N ≡ Ix,ε,N × Iy,ε,N ⊂ Ω , which, in the cases considered here, will have (N + 1)2
nodes, and t ∈ [0, T ] remains as a continuous variable.
Let us denote [·]N the restriction of a function defined on Ω to Ω N and uN (t ) the semidiscrete approximations defined
by the spatial semidiscretizations which we are going to consider, i.e., uN (t ) approaches the exact solution u(xi , yj , t ) of (1)
for all (xi , yj ) ∈ Ω N . Typically, uN (t ) is defined as the solution of an IVP of the form

u′N (t ) + Lε,N (t )uN (t ) = [f˜ ]N (t ),


(5)
uN (0) = [ϕ]N ,
where Lε,N (t ) are appropriate spatial discretizations of the elliptic convection–reaction–diffusion operators Lε (t ) given in
(2) and
f (xi , yj , t ), ( x i , y j ) ∈ ΩN ,

[f˜ ]N (t )(xi , yj ) ≡ ∂g
g (xi , yj , t ) + (xi , yj , t ), (xi , yj ) ∈ ∂ ΩN ,
∂t
where ∂ ΩN = Ω N \ ΩN .
If we consider the special mesh technique, the key to get uniform convergence is the location of the points of the spatial
mesh, which is the tensor product Ix,ε,N × Iy,ε,N , where
Ix,ε,N = {0 = x0 < · · · < xN = 1}, Iy,ε,N = {0 = y0 < · · · < yN = 1}.
Here we use standard piecewise uniform Shishkin meshes for 2D elliptic convection–diffusion or reaction–diffusion
problems (see [16,2,12,23] and references therein). For simplicity, we take the same number N of mesh points in both space
directions. We give the details of the construction of Ix,ε,N and analogously we proceed for Iy,ε,N .
Let us choose N as a multiple of four. For the case (1) (positive convection terms), we define the transition parameter
 
1
σx = min , mx ε ln N , (6)
2
where mx ≥ 1/v ; then, the piecewise uniform mesh has N /2 + 1 points in [0, 1 − σx ] and [1 − σx , 1]. Therefore, the mesh
points are given by
2(1 − σx )

N
i

 , i = 0, . . . , ,
N  2
xi = (7)
N 2σx

N
1 − σx + i − , i = + 1, . . . , N .


2 N 2
In the case (2) (diffusion–reaction), the transition parameter is defined by
 
1 √
σx = min , mx ε log N , (8)
4

where mx ≥ 1/k; then, the piecewise uniform mesh has N /2 + 1 points in [σx , 1 − σx ] and N /4 + 1 points in the subintervals
[0, σx ] and [1 − σx , 1]. Therefore, the mesh points are given by
4 N


i , i = 0, . . . , ,
N 

 4
N 2(1 − 2σx )

N 3N


xi = σ x + i − , i = + 1, . . . , , (9)
 4 N 4 4
3N 4σx
  
3N


1 − σx + i − , i= + 1, . . . , N .


4 N 4
In both cases we denote by hx,i = xi − xi−1 , hy,i = yi − yi−1 , i = 1, . . . , N.
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 225

Using the special meshes (6)–(7) or (8)–(9), Lε,N (t ) is the discretization of the differential convection–reaction–diffusion
operator Lε (t ), using the simple upwind finite difference scheme reaction case, given by

Lε,N (t )uN (t )(xi , yj ) ≡ li,j− (t )uN (t )(xi , yj−1 ) + li−,j (t )uN (t )(xi−1 , yj ) + li+,j (t )uN (t )(xi+1 , yj )
+ li,j (t )uN (t )(xi , yj ) + li,j+ (t )uN (t )(xi , yj+1 ), i, j = 1, . . . , N − 1, (10)
Lε,N (t )uN (t )(xi , yj ) ≡ uN (xi , yj , t ) = g (xi , yj , t ), i = 0, N or j = 0, N ,
where
−ε v1 (xi , yj , t ) −ε
li−,j (t ) = − , li+,j (t ) = , (11)
hx,i h̃x,i h x ,i hx,i+1 h̃x,i
−ε v2 (xi , yj , t ) −ε
li,j− (t ) = − , li,j+ (t ) = ,
hy,j h̃y,j hy,j hy,j+1 h̃y,j
li,j (t ) = k(xi , yj , t ) − li,j− (t ) − li−,j (t ) − li+,j (t ) − li,j+ (t ),

with h̃x,i = (hx,i + hx,i+1 )/2, i = 1, . . . , N , h̃y,j = (hy,j + hy,j+1 )/2, j = 1, . . . , N.

3. Convergence analysis

In this section the approximation properties of the spatial semidiscretizations introduced in the previous section are
analyzed. We show that, in the uniform convergence analysis of these types of monotone and consistent spatial semidis-
cretization processes the presence of the time variable does add important difficulties, compared to their corresponding
stationary models. There are two key properties for proving this fact: firstly, the local truncation error of the space semidis-
cretization will not contain remainder terms related to the derivative ∂ u/∂ t, since we do not discretize at this stage the time
derivative; the second important property is stated in the following semidiscrete maximum principle.

Lemma 1. Let us suppose that Lε,N (t ) are inverse monotone and consistent operators of the form (10). If [f (x, y, t )]N ≤ 0,
[ϕ(x, y)]N ≥ 0, for any (x, y) ∈ Ω N and any t ∈ [0, T ], and g (x, y, t ) ≥ 0 for any (x, y) ∈ ∂ ΩN ≡ Ix,ε,N ×{0, 1} {0, 1}× Iy,ε,N

and any t ∈ [0, T ], then the solution of (5) achieves its maximum value at the discrete initial condition or at the discrete boundary
∂ ΩN × [0, T ], i.e., it holds
max {uN (t )(x, y)} ≤ max {ϕ(x, y), g (x, y, t )}.
(x,y,t )∈Ω N ×[0,T ] (x,y)∈Ω N , (x,y,t )∈∂ ΩN ×[0,T ]

Proof. Let us suppose that the maximum value of uN (t )(xi , yj ), which will be necessarily greater or equal to zero, is reached
at (xi , yj ) ̸∈ ∂ ΩN and t > 0. Then, it follows u′N (t )(xi , yj ) = 0 if t ∈ (0, T ) and u′N (t )(xi , yj ) ≥ 0 if t = T ; in both cases we
deduce u′N (t )(xi , yj ) ≥ 0, and therefore
Lε,N (t )uN (t )(xi , yj ) ≡ li,j (t )uN (t )(xi , yj ) + li,j− (t )uN (t )(xi , yj−1 ) + li−,j (t )uN (t )(xi−1 , yj )
+ li+,j (t )uN (t )(xi+1 , yj ) + li,j+ (t )uN (t )(xi , yj+1 )
= −u′N (xi , yj ) + f (xi , yj , t ) ≤ 0.
Note that the coefficients l•,• (t ) satisfy
li,j ≥ 0, li−,j , li,j− , li+,j , li,j+ < 0.
Besides, from the consistency of the scheme it follows that
li,j (t ) + li−,j (t ) + li,j− (t ) + li+,j (t ) + li,j+ (t ) = k(xi , yj , t ) ≥ 0.
Then, we can immediately deduce that
 
Lε,N (t )uN (t )(xi , yj ) = k(xi , yj , t )uN (t )(xi , yj ) − li,j− (t ) uN (t )(xi , yj ) − uN (t )(xi , yj−1 )
   
− li−,j (t ) uN (t )(xi , yj ) − uN (t )(xi−1 , yj ) − li+,j (t ) uN (t )(xi , yj ) − uN (t )(xi+ , yj )
 
− li,j+ (t ) uN (t )(xi , yj ) − uN (t )(xi , yj+1 ) ≤ 0,

and we conclude that the last inequality only holds if


uN (t )(xi , yj ) = uN (t )(xi , yj−1 ) = uN (t )(xi−1 , yj )
= uN (t )(xi+1 , yj ) = uN (t )(xi , yj+1 ).
Then, the maximum value for uN (t )(xi , yj ) is also reached at these four neighbor points.
226 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

Repeating the same reasoning for these four points, and inductively for neighbors of neighbors, it is straightforward to
conclude that the maximum value for uN (t ) is also reached at the points (xi , yj ) ∈ ∂ ΩN , where uN (t )(xi , yj ) = g (xi , yj , t ),
from the result follows. 

Corollary 1 (Inverse Monotonicity of the Semidiscrete Operator). If a semidiscrete function ψ defined on ΩN × [0, T ] is less or
equal to zero at the initial points (xi , yj , 0) and also on the points belonging to ∂ ΩN × [0, T ], and it satisfies that
ψ ′ (t )(x, y) + (Lε,N (t )ψ(t ))(x, y) ≤ 0, ∀ (x, y, t ) ∈ ΩN × [0, T ],
then, it holds
ψ(t )(x, y) ≤ 0 ∀ (x, y, t ) ∈ Ω N × [0, T ].

Corollary 2. There exists a constant C , depending only on v (k in the diffusion–reaction case), such that
|uN (t )|∞,N ≤ max{|[ϕ]N |∞,N , G} + C max |[f (x, y, t )]N |∞,N ,
t ∈[0,T ]

where
|ψ|∞,N ≡ max |ψ(x, y)|, G= max |g (x, y, t )|.
(x,y)∈Ω N (x,y,t )∈∂ Ωn ×[0,T ]

Proof. It suffices to apply Corollary 1 to the semidiscrete functions


x
ψ ≡ − max{|[ϕ]N |∞,N , G} − max |[f (x, y, t )]N |∞,N ± uN (t ),
v t ∈[0,T ]

or to the functions
1
ψ ≡ − max{|[ϕ]N |∞,N , G} − max |[f (x, y, t )]N |∞,N ± uN (t ). 
k t ∈[0,T ]

Corollary 2 proves the uniform stability of the semidiscretization process. Moreover, using the smoothness and
compatibility assumptions made for the data of (1), it is easy to deduce that uN (t ) ∈ (C 2 [0, T ])(N +1) and its second derivative
2

is bounded independently of ε . This property will be seen in more detail at the beginning of Section 4, where it is required for
the subsequent analysis performed there. To end the uniform convergence analysis we combine this result, or the barrier
function technique if it is required, with the suitable consistency concept in this framework, which we introduce in the
following definition.

Definition 1. For the solution u(x, y, t ) of problem (1), we define the global error and the local truncation error, at any
time t ∈ [0, T ], and we denote them by eN u(t ) and τN u(t ) respectively, being u(t ) ≡ u(x, y, t ), by means of the following
formulae:

    
d
eN u(t ) = [u(t )]N − uN (t ), τN u(t ) = + Lε (t ) u(t ) − + Lε,N (t ) [u(t )]N .
∂t N dt

As we have assumed that ∂ u/∂ t is continuous in Ω̄ × [0, T ], obviously [ ∂∂t u(t )]N = d
dt
[u(t )]N and therefore
τN u(t ) = [Lε (t )u(t )]N − Lε,N (t )[u(t )]N , (12)
and, consequently, the analysis of the consistency for this semidiscretization process is essentially identical to the analysis
which we need to perform for its corresponding stationary version, where the singularly perturbed differential operator is
given by Lε . To end the uniform convergence analysis we can also use either the uniform stability principle or the barrier
function technique if it is required, since this semidiscretization algorithm satisfies that
 
d
+ Lε,N (t ) eN u(t ) = τN u(t ).
dt
Besides, as the term τN u(t ) is essentially independent of the time variable, to get a uniform convergence result it is possible to
choose barrier functions Ψ (xi , yj , t ) which do not depend of the time variable to bound eN u(t ); thus, the natural candidates
to barrier functions are those ones which appear in the corresponding uniform convergence analysis for stationary problems
of type Lε v(x, y) = ψ(x, y).
First we analyze the simpler reaction–diffusion case. To find appropriate bounds for the local error we need to know the
asymptotic behavior of the exact solution u of the continuous problem (1). In [21] it was proven that u can be decomposed
in the form
4
 4

u = u0 + up + wp ,
p=1 p=1
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 227

where u0 is the regular component, up , p = 1, 2, 3, 4, are the edge boundary layer functions associated at each one of
the four sides of the unit square and wp , p = 1, 2, 3, 4, are the corner layer functions corresponding to the corner points
cp , p = 1, 2, 3, 4, respectively. Moreover, it holds
 αs +αt
∂ u0 (x, y, t ) 

 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ,
 
 αs +αt
∂ u1 (x, y, t ) 
   
  ≤ C ε −α1 /2 exp − k/ε x ,
 ∂ xα1 ∂ yα2 ∂ t αt 
 αs +αt
∂ u2 (x, y, t ) 
   
  ≤ C ε −α1 /2 exp − k/ε (1 − x) ,
 ∂ xα1 ∂ yα2 ∂ t αt 
 αs +αt
∂ u3 (x, y, t ) 
   
  ≤ C ε −α2 /2 exp − k/ε y ,
 ∂ xα1 ∂ yα2 ∂ t αt 
 αs +αt
∂ u4 (x, y, t ) 
   
  ≤ C ε −α2 /2 exp − k/ε (1 − y) ,
 ∂ xα1 ∂ yα2 ∂ t αt 
 αs +αt
∂ w1 (x, y, t ) 
       
−αs /2

 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min exp − k/ε x , exp − k/ε y ,
 αs +αt
∂ w2 (x, y, t ) 
       
−αs /2
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min exp − k/ε (1 − x) , exp − k/ε y ,

 αs +αt
∂ w3 (x, y, t ) 
       
−αs /2
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min exp − k/ε x , exp − k/ε (1 − y) ,

 αs +αt
∂ w4 (x, y, t ) 
       
−αs /2

 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min exp − k/ε (1 − x) , exp − k/ε (1 − y ) ,

where αs = α1 + α2 , αs + 2αt ≤ 4.
We decompose the numerical solution in a similar way, i.e.,
4
 4

uN (t ) = u0,N (t ) + up,N (t ) + wp,N (t ),
p=1 p=1

where
u′0,N (t ) + Lε,N (t )u0,N (t ) = [f ]N ,

u0,N (0) = [u0 ]N ,

u′ (t ) + L (t )u (t ) = g + ∂ g , p = 1, 2, 3, 4,
  
p,N ε,N p, N p
∂t N
up,N (0) = [up ]N , p = 1, 2, 3, 4,

w ′ (t ) + L (t )w (t ) = g̃ + ∂ g̃ , p = 1, 2, 3, 4,
  
p,N ε,N p,N p
∂t N
wp,N (0) = [wp ]N , p = 1, 2, 3, 4,

and gp contains evaluations of certain boundary conditions at the nodes located at the four sides of Ω and g̃p contains
evaluations of boundary conditions at the four corners of Ω .
To analyze the error we follow the ideas of [15], where full details can be viewed. Taking into account (12), using a Taylor
expansion, it is straightforward to prove that
 √ −1
C ε N , if xi = σx , 1 − σx , or yj = σy , 1 − σy ,
|τN u0 (t )(xi , yj )| ≤
CN −2 , otherwise.
Then, using a special barrier function, which is typical in the context of singularly perturbed problems solved by methods
defined on Shishkin meshes, and the discrete maximum principle on all domain, it follows.
|eN u0 (t )| ≤ C (N −1 ln N )2 .
Next, for the edge boundary layer function u1 (similarly for the other ones), we analyze the error depending on the
location of the mesh point (xi , yj ). First, when N /4 ≤ i ≤ N , 0 ≤ j ≤ N, we use that |eN u1 (t )(xi , yj )| ≤ |u1 (t )(xi , yj )| +
|u1,N (t )(xi , yj )| and therefore it can be proved that

|eN u1 (t )(xi , yj )| ≤ CN −2 , N /4 ≤ i ≤ N , 0 ≤ j ≤ N . (13)


228 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

In second place, when 0 < i < N /4, 0 < j < N, from Taylor expansions, it follows that the local error satisfies

ε N + (N −1 ln N )2 , j = N /4, 3N /4,
 √ −1 
C
|τN u1 (t )(xi , yj )| ≤
C (N −1 ln N )2 , otherwise.

Then, using an appropriate barrier function (see [15]) and the discrete maximum principle, now on the domain Ω1N =
{(xi , yj )|0 < i < N /4, 0 < j < N }, it can be proved that
|eN u1 (t )| ≤ C (N −1 ln N )2 ,
where the bounds (13) are used on the boundary of Ω1N .
Finally, for the corner layer function w1 (similarly for the other ones), if the mesh point is in Ω1N,2 , where Ω1N,2 =
{(xi , yj )| N /4 ≤ i, j ≤ N }, the bound for the local error and the discrete maximum principle on this domain proves
|τN w1 (t )(xi , yj )| ≤ C (N −1 ln N )2 .
Secondly, using barrier functions (see [15]), it is easy to prove that if the mesh point is in Ω N \ Ω1N,2 it holds

|eN w1 (t )(xi , yj )| ≤ CN −2 .
From the two preceding estimates it follows
|eN w1 (t )| ≤ C (N −1 ln N )2 .
Collecting all bounds obtained for the different components of u(t ), the global error satisfies
|eN u(t )| ≤ C (N −1 ln N )2 ,
proving that the space discretization is an almost second order uniformly convergent scheme.
In second place, we consider the case of having positive convection components, where regular boundary layers appear
at x = 1, y = 1 as well as a corner layer at (x, y) = (1, 1) of width O(ε). For this case, in [20] it was proven that u can be
decomposed in the form u = u0 + w , where u0 is the regular part of u and w is the singular component, which can also be
decomposed in the form w = u1 + u2 + u3 , where u1 , u2 are the regular layer functions near x = 1 and y = 1 respectively
and u3 is the corner layer function; moreover, it holds
 αs +αt
∂ u0 (x, y, t ) 

 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ,
 
 αs +αt
∂ u1 ( x , y , t ) 

−α1 −v(1−x)/ε
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε e ,
 
 αs +αt
∂ u2 ( x 1 , x 2 , t ) 

−α2 −v(1−y)/ε
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε e ,
 
 αs +αt
∂ u3 ( x , y , t ) 

 −v(1−x)/ε −v(1−y)/ε 
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min e ,e ,
  −αs

where αs = α1 + α2 , αs + 2αt ≤ 4.
Following to the continuous problem, we decompose the numerical solution of the upwind scheme in a similar way, i.e.,
3

uN (t ) = u0,N (t ) + up,N (t ),
p=1

where
u′0,N (t ) + Lε,N (t )u0,N (t ) = [f ]N ,

u0,N (0) = [u0 ]N ,

u′ (t ) + L (t )u (t ) = g + ∂ g ,
  
p,N ε,N p,N p p = 1, 2, 3,
∂t N
up,N (0) = [up ]N , p = 1, 2, 3

and g1 , g2 contain evaluations of certain boundary conditions at the two outflow sides of Ω (x = 1 and y = 1), and g̃3
contains evaluations of boundary conditions at the corner (1, 1).
The analysis of the error follows the same ideas than in the reaction–diffusion case. Here we only show the basic tricks
used to find the appropriate bounds for the global error (see [24] for full details of this). Using again (12), from Taylor
expansion it is possible to prove that
|τN u0 (t )(xi , yj )| ≤ CN −1 .
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 229

Now we define the barrier function


ψij0 = CN −1 (xi + yj ).
Then, using the discrete maximum principle on all discrete domain, it is straightforward to prove that
|eN u0 (t )| ≤ CN −1 .
Next, for the regular boundary layer function u1 (similarly for u2 ), when the mesh point (xi , yj ) is outside the layer, i.e.,
0 ≤ i ≤ N /2, 0 ≤ j ≤ N, using again that |eN u1 (t )(xi , yj )| ≤ |u1 (t )(xi , yj )| + |u1,N (t )(xi , yj )|, it is possible to prove that it
holds
|eN u1 (t )(xi , yj )| ≤ CN −1 , 0 ≤ i ≤ N /2, 0 ≤ j ≤ N .
In second place, when N /2 < i < N , 0 ≤ j ≤ N, from Taylor expansions, with the remainder terms in integral form, it is
possible to prove that
|τN u1 (t )(xi , yj )| ≤ Lε,N [ψij1 ].
Then, defining the barrier function
N
 −1
ψij1 = CN −1 ln N 1 + mx hx,k /ε ,

k=i+1

and using the maximum discrete principle on Ω1N = {(xi , yj )|N /2 < i < N /2, 0 < j < N }, it follows that

|eN u1 (t )| ≤ CN −1 ln N ,
also at these mesh points.
Finally, for the corner layer function u3 , if the mesh point is in Ω1N,2 , where Ω1N,2 = {(xi , yj )| 0 ≤ i + j ≤ 3N /2}, using
again Taylor expansions to bound the local error and taking now the barrier function
N N
 −1  −1
ψij2 = CN −1 ln N 1 + mx hx,k /ε 1 + my hy,k /ε ,
 
k=i+1 k=j+1

it is straightforward to prove that


|τN u3 (t )(xi , yj )| ≤ Lε,N [ψij2 ],
and therefore, using the discrete maximum principle on Ω1N,2 , we deduce

|eN u3 (t )| ≤ CN −1 ln N .
Then, using barrier functions (see [24]), it is easy to prove that if the mesh point is in Ω N \ Ω1N,2 , it holds

|eN u3 (t )(xi , yj )| ≤ CN −1 .
Collecting all bounds obtained for the different components of u(t ), the global error satisfies
|eN u(t )| ≤ CN −1 ln N ,
proving that the space discretization is an almost first order uniformly convergent scheme.
We wish to remark that if we chose to discretize in space by using fitted operator methods on non special meshes, the
analysis of their uniform convergence would be similar to the discretizations considered here. The differences in the analysis,
which also appear when this type of schemes are used for stationary problems, are associated to the relations between the
exponential characters of the coefficients of the numerical scheme and the exact solution. In this manuscript we have not
considered that analysis in order to shorten the paper.

4. Time integration: unconditional convergence

In this section we define the numerical method which we propose to integrate successfully the continuous problem (1).
After having considered the spatial semidiscretization stage in the previous section, we complete the discretization process
by using an appropriate time integrator. In terms of simplicity, the implicit Euler method is the best candidate, although
other time integrators will give us better results in terms of computational cost (see the numerical section later).
(N +1)2
Using the Euler method, we obtain numerical approximations um N ∈ R to uN (tm ), being tm = mτ , where τ = T /M
is the chosen time step, which we consider constant for simplicity. Thus, the fully discrete scheme is given by
m−1
um
N − uN
+ Lε,N (tm )um
N = [f ]N (tm ),
˜ m = 1, . . . , M ,
τ (14)
uN = [ϕ]N .
0
230 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

It is clear that the advance in time requires to solve a pentadiagonal linear system per time step of the form
(IN + τ Lε,N (tm ))um
N = GN ,
m
where Gm m−1
N = uN + τ [f˜ ]N (tm ).
Before studying the uniform convergence of this discretization stage, we assume that uN (t ) is twice continuously
differentiable and that its second time derivative is uniformly bounded, in the maximum norm, with respect to ε and N.
Such assumption is natural since wN ≡ u′′N (t ) is the solution of
wN′ (t ) + Lε,N (t )wN (t ) = fˆN , wN (0) = ϕ̂N ,
[ ∂∂ t 2f ]N (t ),

where fˆN = and
 
∂ f˜
ϕ̂N = (0) − Lε,N (0)([f˜ ]N (0) − Lε,N (0)[ϕ]N ).
∂t
N
Following the same principles than those ones developed in Section 3 for bounding the solution of (5), it can be deduced
that |wN (t )|∞,N ≤ C .

4.1. Unconditional stability

In this section we remark one of the main qualities of the chosen time integrator, which is focused in the fact that the
maximum principle that the semidiscrete problems (5) possess, are preserved in the fully discrete scheme (14). As a direct
consequence, it is deduced that this scheme is unconditionally stable, i.e., the stability bounds are independent of ε, N and
M. As the required techniques are classical (see for example [3]), we give without proof the following result.

Theorem 1. Operators IN + τ Lε,N (tm ) are inverse monotone and it holds

∥(IN + τ Lε,N (tm ))−1 ∥∞,N ≤ 1. (15)

From this result, we can deduce immediately the two following properties of the scheme.

Corollary 3 (Numerical Stability). If we consider the perturbed problem


m−1
ũm
N − ũN
+ Lε,N (tm )ũm
N = [f ]N (tm ) + δ ,
˜ m
m = 1, . . . , M ,
τ
ũ0N = [ϕ]N + τ δ 0 ,
then it holds that
m

N − uN ∥∞,N ≤ τ
∥ũm ∥δ i ∥∞,N .
m

i=o

Corollary 4 (Contractivity). If we consider f = 0, g = 0 in the continuous problem (1), then it holds


∥ um m−1
N ∥∞,N ≤ ∥uN ∥∞,N .

4.2. Consistency

Let us introduce the local truncation error at time tm for method (14) in the usual way, i.e.,

N = uN (tm ) − (IN + τ Lε,N (tm )) (uN (tm−1 ) + τ [f˜ ]N (tm )).


−1
em
With the assumptions posed at the beginning of this section we can prove the following result.

Theorem 2. The local truncation error satisfies

N ∥∞,N ≤ C τ ,
∥em 2
(16)
being C a constant independent of ε and N.
Proof. Using appropriate Taylor expansions, we deduce
τ2
uN (tm−1 )(xi , yj ) = uN (tm )(xi , yj ) − τ u′N (tm )(xi , yj ) + u′′N (tm − θi,j τ )(xi , yj ),
2
with θi,j ∈ (0, 1); now, taking into account that u′N (t ) = [f (x, y, t )]N − Lε,N (t )uN (t ), we deduce

N = (IN + τ Lε,N (tm )) τ ΨN ,


−1 2
em
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 231

where ΨN (xi , yj ) = − 12 u′′N (tm − θi,j τ )(xi , yj ) and, from the previous smoothness assumptions, it holds |ΨN |∞,N ≤ C . From
this fact and (15), the result follows. 

4.3. Convergence of the time integration

We define the global error in time for the scheme (14) as usual, i.e.,
ENm = uN (tm ) − um
N.

Now, combining the consistency and uniform stability results stated previously, we deduce the following result.

Theorem 3. The global error satisfies

∥ENm ∥∞,N ≤ CM −1 , (17)


being C a constant independent of ε and N.

Proof. If we add and subtract the term (IN + τ Lε,N (tm ))−1 (uN (tm−1 ) + τ [f˜ ]N (tm )) in the definition of ENm , and we take into
N = (IN + τ Lε,N (tm )) (uN + τ [f˜ ]N (tm )), we deduce
−1 m−1
account that um

N + (IN + τ Lε,N (tm )) EN .


−1 m−1
ENm = em
Using m times this recurrence, it follows that
m 
 m
ENm = (IN + τ Lε,N (tj ))−1 eiN
i=1 j=i+1

and using now (15) and (16) it is straightforward to deduce (17). 

4.4. Uniform convergence

The analysis of the uniform convergence ends by jointing the results of uniform convergence of both discretization stages.
If we define the global error at each spatial mesh point (xi , yj ) and at each time level tm as follows
Eim,j = u(xi , yj , tm ) − um
N (xi , yj ),

we can split the contributions of both discretization stages, by using that


[u(x, y, tm )]N − um
N = ([u(x, y, tm )]N − uN (tm )) + EN .
m

Now, we can combine the uniform convergence results of Section 2 and Theorem 3 to deduce the main result of the paper.

Theorem 4 (Uniform and Unconditional Convergence). We assume that the solution of the continuous problem (1) u ∈
C 4,2 (Ω̄ × [0, T ]). Then, the global error, associates to the numerical method combining the scheme (10)–(11) and the implicit
Euler method, satisfies
|Eim,j | ≤ C (N −1 ln N )2 + M −1 ,
 
max
1≤i,j≤N ,1≤m≤M

for the diffusion–reaction case and


|Eim,j | ≤ C N −1 ln N + M −1 ,
 
max
1≤i,j≤N ,1≤m≤M

for the convection–diffusion case. In both cases, C is a constant independent of ε, N and M.

5. Numerical results

In this section we show the numerical results obtained in the integration of some problems of type (1). We combine
the implicit Euler method to discretize in time and the upwind scheme to discretize in space, on the appropriate piecewise
uniform Shishkin mesh for each case.
The first example is given by
 xy   xy 
ut − ε 1u + 1 − + t ux + − 1 + (3t + 1) cos(t ) uy = f (x, y, t ), (x, y, t ) ∈ Ω × [0, 1],
2 2
u(0, y, t ) = y + e , −t
u(1, y, t ) = 1 + y + e , −t
y ∈ [0, 1], t ∈ [0, 1] (18)
u(x, 0, t ) = x + e , −t
u(x, 1, t ) = 1 + x + e , −t
x ∈ [0, 1], t ∈ [0, 1]
u(x, y, 0) = x + y + 1, x, y ∈ [0, 1],
232 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

Fig. 1. Numerical solution of example (18) for ε = 10−3 with N = M = 32.

with
f (x, y, t ) = −e−t + (3t + 1) cos(t ) + t 1 + te2−xy(1−x)(1−y) .
 

Note that the components of the convective term are both positive; then, the Shishkin mesh is defined by (6)–(7). We
take mx = 2, my = 1 to define the transition parameters of the meshes Ix,ε,N and Iy,ε,N respectively.
The exact solution of (18) is unknown; Fig. 1 shows the numerical solution at the final time t = 1 for some values of the
diffusion and the discretization parameter, where regular layers can be observed.
To approximate the maximum pointwise errors, we use a variant of the two-mesh principle (see [25,26] for a justification
of this method). Then, we calculate {ûN }, the numerical solution on the mesh {(x̂i , ŷj , t̂n )} containing the original mesh points
and its midpoints, i.e.,
x̂2i = xi , i = 0, . . . , N , x̂2i+1 = (xi + xi+1 )/2, i = 0, . . . , N − 1,
ŷ2j = yj , j = 0, . . . , N , ŷ2j+1 = (yj + yj+1 )/2, j = 0, . . . , N − 1,
t̂2m = tm , m = 0, . . . , M , t̂2m+1 = (tm + tm+1 )/2, m = 0, . . . , M − 1.
The maximum errors at the mesh points of the coarse mesh are approximated by computing the following two-mesh
differences
di,j,N ,M = max max |uN (xi , yj , tm ) − ûN (xi , yj , tm )|, (19)
0≤m≤M 0≤i,j≤N

and the orders of convergence are calculated by


log di,j,N ,M /di,j,2N ,2M
 
q= .
log 2
From the double-mesh differences computed in (19) we obtain the uniform maximum errors by
dN ,M = max di,j,N ,M ,
ε

and from them, in a usual way, the corresponding numerical uniform orders of convergence by
log dN ,M /d2N ,2M
 
quni
= .
log 2
To solve the linear system associated to the full discrete method at each time level, we use the BI-CGSTAB algorithm
with a relaxed incomplete LU-factorization (see [27]) applied to a pentadiagonal matrix, with a tolerance equal to 10−10 .
Table 1 displays the obtained results for this case; from it we deduce the almost first order of uniform convergence. So, we
can conclude that the errors related to the spatial semidiscretization stage dominate in the global error of the numerical
scheme (note that convergence in space has almost first order due to the logarithmic factor).
In the second example we consider the following diffusion–reaction problem

ut − ε 1u + (1 + x2 y2 )u = f (x, y, t ), (x, y, t ) ∈ Ω × [0, 1],


u(0, y, t ) = u(1, y, t ) = 0, y ∈ [0, 1], t ∈ [0, 1]
(20)
u(x, 0, t ) = u(x, 1, t ) = 0, x ∈ [0, 1], t ∈ [0, 1]
u(x, y, 0) = 0, x, y ∈ [0, 1],

with f (x, y, t ) = (1 − y) et sin(π x)y − (1 + x2 y2 )(1 − x) .


 
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 233

Fig. 2. Numerical solution of example (20) for ε = 10−4 with N = M = 32.

Table 1
Maximum errors and orders of convergence for example (18).
ε N = 16 N = 32 N = 64 N = 128 N = 256
M =8 M = 16 M = 32 M = 64 M = 128

2−6 1.6551E−1 1.1808E−1 7.8369E−2 4.9171E−2 2.9300E−2


0.487 0.591 0.672 0.747
2−8 1.7427E−1 1.2433E−1 8.4367E−2 5.3463E−2 3.2102E−2
0.487 0.559 0.658 0.736
2−10 1.7636E−1 1.2599E−1 8.6154E−2 5.4956E−2 3.3202E−2
0.485 0.548 0.649 0.727
2−12 1.7687E−1 1.2649E−1 8.6613E−2 5.5362E−2 3.3506E−2
0.484 0.546 0.646 0.724
2−14 1.7700E−1 1.2662E−1 8.6729E−2 5.5465E−2 3.3581E−2
0.483 0.546 0.645 0.724
2−16 1.7703E−1 1.2665E−1 8.6758E−2 5.5491E−2 3.3600E−2
0.483 0.546 0.645 0.724
dN , M 1.7703E−1 1.2665E−1 8.6758E−2 5.5491E−2 3.3600E−2
quni 0.483 0.546 0.645 0.724

In this case the Shishkin mesh is defined by (8)–(9). Now the values mx = my = 2 are used to define the mesh, and we
again solve with the BI-CGSTAB algorithm with the same tolerance as before. Again, the exact solution is unknown; Fig. 2
shows the numerical solution at the final time t = 1 for some values of the diffusion and the discretization parameter,
where parabolic layers can be observed.
Table 2 displays the corresponding results; from it, we see that the order of convergence is one; therefore, we can
conclude that the global errors are dominated by the errors corresponding to the time discretization.
To show the contribution to the global error of both discretizations, we calculate the following orders of convergence

log di,j,N ,M /di,j,2N ,4M


 
q= .
log 2

Note that the time step is divided by four, whereas the number of grid points in space is doubled at each spatial direction;
the reason for choosing such divisions is that, in this case, the order of convergence in time is one, but the order in space is
almost two; so, the contributions of spatial and time discretizations are balanced. Table 3 displays the corresponding results;
from it, we see the almost second order of convergence and therefore we can conclude that the global errors are dominated
by the errors corresponding to the space discretization.

Acknowledgments

The authors thank the referees for their valuable suggestions which have helped to improve the presentation of this paper.
This research was partially supported by the projects MEC/FEDER MTM 2010-16917, MTM 2010-21037 and the Diputación
General de Aragón.
234 C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235

Table 2
Maximum errors and orders of convergence for example (20).
ε N = 32 N = 64 N = 128 N = 256 N = 512
M =8 M = 16 M = 32 M = 64 M = 128

2−6 8.970E−3 4.729E−3 2.460E−3 1.269E−3 6.484E−4


0.924 0.943 0.955 0.969
2−8 1.330E−2 6.259E−3 3.012E−3 1.477E−3 7.320E−4
1.087 1.055 1.028 1.013
2−10 2.134E−2 1.012E−2 4.134E−3 1.807E−3 8.386E−4
1.076 1.292 1.194 1.108
2−12 2.180E−2 1.076E−2 4.842E−3 2.161E−3 9.789E−4
1.018 1.152 1.164 1.142
2−14 2.203E−2 1.089E−2 4.894E−3 2.188E−3 9.933E−4
1.016 1.154 1.161 1.140
2−16 2.214E−2 1.096E−2 4.920E−3 2.202E−3 1.001E−3
1.014 1.156 1.160 1.138
dN , M 2.214E−2 1.096E−2 4.920E−3 2.202E−3 1.001E−3
quni 1.014 1.156 1.160 1.138

Table 3
Maximum errors and orders of convergence for example (20).
ε N = 32 N = 64 N = 128 N = 256 N = 512
M =8 M = 32 M = 128 M = 512 M = 2048

2−6 8.970E−3 2.687E−3 7.288E−4 1.892E−4 4.831E−5


1.739 1.882 1.946 1.969
2−8 1.330E−2 3.930E−3 1.051E−3 2.705E−4 6.853E−5
1.758 1.902 1.959 1.981
2−10 2.134E−2 7.860E−3 2.182E−3 5.738E−4 1.469E−4
1.441 1.849 1.927 1.966
2−12 2.180E−2 8.399E−3 2.900E−3 9.438E−4 2.957E−4
1.376 1.534 1.620 1.674
2−14 2.203E−2 8.494E−3 2.918E−3 9.484E−4 2.967E−4
1.375 1.542 1.621 1.676
2−16 2.214E−2 8.542E−3 2.926E−3 9.507E−4 2.972E−4
1.374 1.545 1.622 1.677
dN , M 2.214E−2 8.542E−3 2.926E−3 9.507E−4 2.972E−4
quni 1.374 1.545 1.622 1.677

References

[1] R.B. Kellogg, S. Shih, Asymptotic analysis of a singular perturbation problem, SIAM J. Math. Anal. 18 (1987) 1467–1511.
[2] J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems, revised ed., World Scientific, 2012.
[3] H.G. Roos, M. Stynes, L. Tobiska, Robust Numerical Methods for Singularly Perturbed Differential Equations, second ed., Springer-Verlag, 2008.
[4] E.C. Gartland, Uniform high-order difference schemes for a singularly perturbed two point boundary value problem, Math. Comp. 48 (1987) 551–564.
[5] W. Guo, M. Stynes, Finite element analysis of an exponentially fitted lumped scheme for time-dependent convection–diffusion problems, Numer.
Math. 66 (1993) 347–371.
[6] H. Houston, E. Süli, Adaptive Lagrange–Galerkin methods for unsteady convection-dominated diffusion problems, Math. Comp. 70 (2001) 77–106.
[7] E. O’Riordan, M. Stynes, A globally convergent finite element method for a singularly perturbed elliptic problem in two dimensions, Math. Comp. 57
(1991) 47–62.
[8] R.G. Durán, A.L. Lombardi, M.I. Prieto, Superconvergence for finite element approximation of a convection–diffusion equation using graded meshes,
IMA J. Numer. Anal. 32 (2012) 511–533.
[9] E.C. Gartland, Graded-mesh difference schemes for singularly perturbed two-point boundary value problems, Math. Comp. 51 (1988) 631–657.
[10] D. Herceg, Uniform fourth order difference scheme for a singular perturbation problem, Numer. Math. 56 (1990) 675–693.
[11] T. Linss, Layer adapted meshes and FEM for time-dependent singularly perturbed reaction–diffusion problems, Int. J. Comput. Sci. Math. 1 (2007)
259–270.
[12] G.I. Shishkin, Approximation of the solution to singularly perturbed boundary value problems with boundary layers, USSR Comput. Math. Math. Phys.
29 (1989) 1–10.
[13] G.I. Shishkin, Grid approximation of singularly perturbed boundary value problems with convective terms, Sov. J. Numer. Anal. Math. Model. 5 (1990)
173–187.
[14] M. Stynes, L. Tobiska, The SDFEM for a convection–diffusion problem with a boundary layer: optimal error analysis and enhancement of accuracy,
SIAM J. Numer. Anal. 41 (2003) 1620–1642.
[15] C. Clavero, J.L. Gracia, E. O’Riordan, A parameter robust numerical method for a two dimensional reaction–diffusion problem, Math. Comp. 74 (2005)
1743–1758.
[16] T. Linss, M. Stynes, Asymptotic analysis and Shishkin-type decomposition for an elliptic convection–diffusion problem, J. Math. Anal. Appl. 261 (2001)
604–632.
[17] F. Liu, N. Madden, M. Stynes, A. Zhou, A two-scale sparse grid method for a singularly perturbed reaction–diffusion problem in two dimensions, IMA
J. Numer. Anal. 29 (2009) 986–1007.
C. Clavero, J.C. Jorge / Computers and Mathematics with Applications 70 (2015) 222–235 235

[18] 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 (2007) 1–25.
[19] C. Clavero, J.L. Gracia, J.C. Jorge, A uniformly convergent alternating direction HODIE finite difference ccheme for 2D time dependent
convection–diffusion problems, IMA J. Numer. Anal. 26 (2006) 155–172.
[20] 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 problem, Appl. Numer. Math. 27 (1998) 211–231.
[21] C. Clavero, J.C. Jorge, F. Lisbona, G.I. Shishkin, An alternating direction scheme on a nonuniform mesh for reaction–diffusion parabolic problems, IMA
J. Numer. Anal. 20 (2000) 263–280.
[22] N.N. Yanenko, The Method of Fractional Steps, Springer, 1971.
[23] G.I. Shishkin, L.P. Shishkina, Difference Methods for Singular Perturbation Problems, CRC Press, 2009.
[24] T. Linss, M. Stynes, A hybrid difference scheme on a Shishkin mesh for linear connvection-diffusion problem, Appl. Numer. Math. 31 (1999) 255–270.
[25] P.A. Farrell, A. Hegarty, On the determination of the order of uniform convergence, in: Proceedings of IMACS’91 Vol. 2, 1991, pp. 501–502.
[26] P.A. Farrell, A.F. Hegarty, J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Robust Computational Techniques for Boundary Layers, Chapman and Hall, 2000.
[27] H.A. Van der Vorst, BI-CGSTAB: a fast and smoothly converging variant of BI-CG for the solution of nonsymmetric linear systems, SIAM J. Sci. Stat.
Comput. 13 (2) (1992) 631–644.

You might also like