Applied Mathematics and Computation 287–288 (2016) 12–27

Contents lists available at ScienceDirect

Applied Mathematics and Computation

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

Uniform convergence and order reduction of the fractional

implicit Euler method to solve singularly perturbed 2D
reaction-diffusion problems
C. Clavero a,∗, J.C. Jorge b
a
IUMA and Department of Applied Mathematics, University of Zaragoza, Zaragoza, Spain
b
Department of Computational and Mathematical Engineering and ISC, University Pública of Navarra, Pamplona, Spain

a r t i c l e i n f o a b s t r a c t
MSC: In this paper we analyze the uniform convergence of a numerical method designed to
65N05 approximate efficiently the solution of 2D parabolic singularly perturbed problems of
65N06
reaction diffusion type. The method combines a modified fractional implicit Euler method
65N10
to discretize in time, and the classical central finite difference scheme, on a special
Keywords: nonuniform mesh, to discretize in space. The resulting fully discrete scheme is uniformly
2D parabolic singularly perturbed problems convergent with respect to the diffusion parameter. The analysis of the convergence is
Fractional implicit Euler method made by using a two step technique, which discretizes first in time and later on in space.
Nonuniform special meshes We show the order reduction phenomenon associated to the fractional implicit Euler
Uniform convergence method, which typically appears if the boundary conditions are time dependent and a
Order reduction
natural evaluation of them is done. An appropriate choice for the boundary conditions is
proposed and analyzed in detail, proving that the order reduction can be removed. Some
numerical tests show the practical effects of our method; as well, we compare it with the
classical choice for the boundary data in terms of the uniform consistency and the order
of uniform convergence of the numerical scheme.
© 2016 Elsevier Inc. All rights reserved.

1. Introduction

We consider 2D reaction-diffusion parabolic problems modeled by the differential equation

∂u
− ε 2 u + ku = f (x, y, t ),
∂t
where 0 < ε ≤ 1 and the reaction term k ≡ k(x, y, t), is a smooth function such that k ≥ β 2 > 0, β > 0. We will focus
our attention on the case ε  β , for which, in general, their solutions have a multiscale character, changing rapidly in
certain narrow regions called boundary layers (see [13,17]). If standard finite difference or finite element methods are used
on uniform meshes, for sufficiently small values of ε the numerical approximations are coarse. So, uniformly convergent
methods, for which the rate of convergence and the constant’s error of the method are both independent of ε , are convenient
to obtain reliable numerical approximations for any value of the diffusion parameter.

∗
Corresponding author. Tel.: +34 976761982; fax: +34 976761886.
E-mail addresses: clavero@unizar.es (C. Clavero), jcjorge@upna.es (J.C. Jorge).

http://dx.doi.org/10.1016/j.amc.2016.04.017
0 096-30 03/© 2016 Elsevier Inc. All rights reserved.
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 13

Many numerical uniformly convergent methods are developed in last years, by using fitted operator methods (see
[14,17] and references therein) or fitted mesh methods (see [9,13,19,20] and references therein) for different types of singu-
larly perturbed problems. The case of 2D elliptic problems was analyzed, for instance, in [4,11,12]; for 2D time dependent
problems, in [2,3,6,7] it was described a two step technique to construct the fully discrete method, which discretizes firstly
in time and later on in space. Then, using the direction alternating technique (see [21]), only tridiagonal linear systems
must be solved to obtain the numerical solution. Nevertheless, in most of papers, only homogeneous boundary conditions
were considered, to avoid the order reduction phenomenon, which is specially severe if natural evaluations of the boundary
conditions are used.
Recently, in [5], the two step technique in reverse order, i.e., firstly in space and secondly in time, was considered, proving
that the uniform convergence is reached when the discretization in space uses classical finite difference schemes defined on
an appropriate piecewise uniform Shishkin mesh, together with the classical backward Euler method, defined on a uniform
mesh, to discretize in time. In that case, no order reductions occur, but the computational cost of that scheme is substan-
tially large because large pentadiagonal linear systems, one per time step, must be solved. To reduce that computational
cost, in this paper we consider a method of type alternating directions. So, the fractional implicit Euler method is used to
discretize in time, in combination with a splitting of the diffusion-reaction operator which separates the derivatives of the
spatial variables. The analysis of the uniform convergence of the fully discrete scheme follows similar ideas to those ones
in [6]; the main differences appear in the proof of the uniform consistency of the time semidiscretization stage using the
fractional implicit Euler method. It is well known (see [1] and references therein) that a classical evaluation of the boundary
conditions, when most of the one step methods are used to integrate in time, causes a reduction in the order of conver-
gence. This order reduction is specially severe when the fractional implicit Euler method is used in problems where non
homogeneous time dependent boundary conditions appear. So, here we propose a suitable modification of these evaluations
which eliminates the order reduction.
The paper is structured as follows. In Section 2, we set up the problem to be solved and we remind some results con-
cerning to the asymptotic behavior, with respect to ε , of the exact solution and its derivatives. In Section 3, we introduce
the time semidiscretization, which uses the fractional implicit Euler method, and we prove its uniform convergence. When
non homogenous boundary conditions appear, suitable evaluations of the boundary data are essential to avoid the order
reduction. As well, we study the asymptotic behavior of the exact solution of the semidiscrete problems resulting after the
time discretization, which preserve, in essence, the same ε -asymptotic behavior that the solution of the continuous prob-
lem. In Section 4, we define the spatial discretization by using a finite differences scheme on a piecewise uniform mesh
of Shishkin type, and we remind the techniques used in previous studies for proving its uniform convergence. Combining
the results of the two stages of discretization, we deduce the uniform convergence of the fully discrete scheme. Finally,
in Section 5, some numerical results are shown, which corroborate in practice the efficiency of the method, the order of
uniform of convergence and the influence of the order reduction phenomenon.
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, denoted by  · D (where D is the corresponding
domain).

2. Asymptotic behavior of the continuous problem

Let  ≡ (0, 1) × (0, 1). We consider the initial-boundary value problem

∂u
Lu ≡ + L1,ε u + L2,ε u = f1 (x, y, t ) + f2 (x, y, t ), in  × (0, T ],
∂t
(1)
u(x, y, 0 ) = ϕ (x, y ), in ,
u(x, y, t ) = g(x, y, t ), in ∂  × [0, T ],
where the spatial differential operators L1,ε , L2,ε are defined by

∂ 2u ∂ 2u
L1,ε u ≡ −ε 2 + k1 (x, y, t )u, L2,ε u ≡ −ε 2 2 + k2 (x, y, t )u, (2)
∂x 2 ∂y
with f1 + f2 = f, k = k1 + k2 , ki ≥ βi2 , βi > 0, i = 1, 2.
We assume that functions ki , fi , i = 1, 2, g and ϕ are sufficiently smooth and also that sufficient compatibility conditions
hold among them in order to u(x, y, t) ∈ C4, 2 ( × (0, T]), i.e., it has continuous derivatives up to fourth order in space and
second order in time (see [5–7]). It is well known that the exact solution of (1), in general, has boundary layers at the four
sides of the spatial domain. Moreover, the solution can be decomposed in the form


4 
4
u = u0 + up + w p,
p=1 p=1

where u0 is the regular component, u p , p = 1, 2, 3, 4, are the edge boundary layer functions associated at each one of
the four sides of the unit square and w p , p = 1, 2, 3, 4, are the corner layer functions corresponding to the corner points
14 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

c p , p = 1, 2, 3, 4, respectively, and it holds (see [7])

 
 ∂ αs +αt u0 (x, y, t ) 
 
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C,
 
 ∂ αs +αt u1 (x, y, t ) 
  −α1
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε exp (−β x/ε ),
 
 ∂ αs +αt u2 (x, y, t ) 
  −α1
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε exp (−β (1 − x )/ε ),
 
 ∂ αs +αt u3 (x, y, t ) 
  −α2
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε exp (−β y/ε ),
 
 ∂ αs +αt u4 (x, y, t ) 
  −α2
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε exp (−β (1 − y )/ε ),
 
 ∂ αs +αt w1 (x, y, t ) 
  −αs
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min {exp (−β x/ε ), exp (−β y/ε )},
 
 ∂ αs +αt w2 (x, y, t ) 
  −αs
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min {exp (−β (1 − x )/ε ), exp (−β y/ε )},
 
 ∂ αs +αt w3 (x, y, t ) 
  −αs
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min {exp (−β x/ε ), exp (−β (1 − y )/ε )},
 
 ∂ αs +αt w4 (x, y, t ) 
  −αs
 ∂ xα1 ∂ yα2 ∂ t αt  ≤ C ε min {exp (−β (1 − x )/ε ), exp (−β (1 − y )/ε )},
¯ × [0, T ], with αs = α1 + α2 , αs + 2αt ≤ 4. Jointing the preceding estimates, we deduce the corresponding
for (x, y, t ) ∈ 
bounds for u and some of its derivatives which are going to be used in next sections. Particularly, it holds
 
 ∂ αt u(x, y, t ) 
  ≤ C,
 ∂ t αt 
 
 ∂ α1 u(x, y, t ) 
  ≤ C (1 + ε −α1 (exp(−β x/ε ) + exp(−β (1 − x )/ε ))),
 ∂ xα1 
 
 ∂ α2 u(x, y, t ) 
  ≤ C (1 + ε −α2 (exp(−β y/ε ) + exp(−β (1 − y )/ε ))),
 ∂ yα2 
¯ × [0, T ], with 0 ≤ αt ≤ 2, 0 ≤ αi ≤ 4, i = 1, 2.
for (x, y, t ) ∈ 

3. Time semidiscretization: uniform convergence

3.1. The scheme

In this section we define the numerical method that we propose to discretize in time the continuous problem (1). We
consider the fractional implicit Euler method (see [6,21]), which can be written as a two half step scheme as follows. Let
τ ≡ T/M be the time step, and let us consider the uniform mesh I¯M = {tn = nτ , n = 0, 1, . . . , M}. Let un ≈ u(x, y, tn ), n =
0, 1, . . . , M, be the semidiscrete solutions to be defined here as follows:

(i ) (initialize)
u0 = ϕ (x, y ), (x, y ) ∈ .
(ii ) (first half step)
(I + τ Ln1+1
,ε )u
n+1/2
= un + τ f1n+1 , (x, y ) ∈ ,
u (x, y ) = g
n+1/2
(x, y ),
n+1/2
(x, y ) ∈ {0, 1} × [0, 1].
(iii ) (second half step)
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 15

(I + τ Ln2+1
,ε )u
n+1
= un+1/2 + τ f2n+1 , (x, y ) ∈ ,
u n+1
(x, y ) = g
n+1
(x, y ), (x, y ) ∈ [0, 1] × {0, 1},
n = 0, . . . , M − 1, (3)

with
∂ 2u
,ε u ≡ −ε
Ln1+1 + k1 (x, y, tn+1 )u,
2
∂ x2
2∂ u
2
(4)
,ε u ≡ −ε
Ln2+1 + k2 (x, y, tn+1 )u,
∂ y2
f1 = f1 (x, y, tn+1 ), f2n+1 = f2 (x, y, tn+1 ).
n+1

In the literature (see [1,15,18]), the most natural option for the boundary conditions is given by
gn+1/2 = g(x, y, tn+1 ), (x, y ) ∈ {0, 1} × [0, 1],
(5)
gn+1
= g(x, y, tn+1 ), (x, y ) ∈ [0, 1] × {0, 1}.
Nevertheless, in most of cases, this option reduces the order of consistency to zero and causes a sharp increase in the global
error of the method as well as strong difficulties for proving its uniform convergence.
Here, we propose a different choice for the boundary conditions contained in (3), which is given by

gn+1/2 = (I + τ Ln2+1
,ε
)g(x, y, tn+1 ) − τ f2n+1 , (x, y ) ∈ {0, 1} × [0, 1],
(6)
gn+1 (x, y ) = g(x, y, tn+1 ), (x, y ) ∈ [0, 1] × {0, 1}.
This proposal is different that the considered ones in previous papers concerning the order reduction of these schemes (see
[1,2]); moreover, it is a bit cheaper than these ones and requires a different, somewhat simpler, technique of analysis for its
uniform consistency and its subsequent uniform convergence.

3.2. Stability

Let us define the operators (I + τ Ln1,ε,0 )−1 (analogously (I + τ Ln2,ε,0 )−1 ), as follows: v = (I + τ Ln1,ε,0 )−1 u is the solution of
the elliptic problem
(I + τ Ln1,ε,0 )v = u(x, y ), (x, y ) ∈ ,
v(x, y ) = 0, (x, y ) ∈ {0, 1} × [0, 1].
From the classical maximum principles for one dimensional diffusion-reaction operators (see [16,17]), it is straightforward
that it holds
1
(I + τ Lni,ε,0 )−1 ¯ ≤ , i = 1, 2. (7)
1 + τ βi2

3.3. Uniform consistency

Let us denote en+1 the local truncation error of scheme (3) at the time tn+1 , i.e., en+1 = u(tn+1 ) − u˜n+1 , where u˜n+1 is the
solution of
(I + τ Ln1+1
,ε
)u˜n+1/2 = u(tn ) + τ f1n+1 , (x, y ) ∈ ,
u˜ n+1/2
(x, y ) = gn+1/2 (x, y ), (x, y ) ∈ {0, 1} × [0, 1],
(8)
(I + τ Ln2+1
,ε
)u˜n+1 = u˜n+1/2 + τ f2n+1 , (x, y ) ∈ ,
u˜ (x, y ) = gn+1 (x, y ), (x, y ) ∈ [0, 1] × {0, 1}.
n+1

Theorem 1. Under the assumptions made on data of problem (1), it holds

en ¯ ≤ C τ 2 , n = 1, . . . , M, (9)

and therefore (3) is a first order uniformly consistent method.

Proof. From the definition of u˜n+1 given in (8), it is easily deduced that
 
(I + τ Ln1+1
,ε ) (I + τ L2,ε )u
n+1
˜n+1 − τ f2n+1 = u(tn ) + τ f1n+1 . (10)

On the other hand, by using the simple Taylor expansion

∂u
u(x, y, tn ) = u(x, y, tn+1 ) − τ (x, y, tn+1 ) + O (τ 2 )
∂t
16 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

∂u
and taking into account that = f1 + f2 − L1,ε u − L2,ε u, it can be deduced that
∂t
 
(I + τ Ln1+1
,ε ) ( I + τ L n+1
2,ε ) u (t n +1 ) − τ f 2 ( x.y.t n +1 ) = u(tn ) + τ f1n+1 + O (τ 2 ). (11)

Subtracting Eqs. (11) and (10), it holds

(I + τ Ln1+1
,ε )(I + τ L2,ε )e
n+1 n+1
= O ( τ 2 ). (12)

Regarding to the boundary conditions, it is trivial that u(x, y, tn+1 ) = gn+1 (x, y, t n+1 ) in [0, 1] × {0, 1}, and it is also clear
that it holds
(I + τ Ln2+1
,ε )u (x, y, tn+1 ) − τ f 2
n+1
= gn+1/2 (x, y, tn+1 ), in {0, 1} × [0, 1].
Then, the local error can be written as the solution of a problem of the form
(I + τ Ln1+1
,ε
)en+1/2 = O (τ 2 ), (x, y ) ∈ ,
en+1/2 (x, y ) = 0, (x, y ) ∈ {0, 1} × [0, 1],
(13)
(I + τ Ln2+1
,ε
)en+1 = en+1/2 , (x, y ) ∈ ,
en+1 (x, y ) = 0, (x, y ) ∈ [0, 1] × {0, 1}.
From (13) and the stability property (7), the required result (9) follows. 

Remark 1. Note that for non homogeneous boundary data g(x, y, t), in general Ln2+1 ,ε
g − f2
= 0 in {0, 1} × [0, 1] × [0, T].
Therefore, in the first half step of (8), a term of size O(τ ) appears as the difference between the classical boundary conditions
given by u˜n+1/2 = g(x, y, tn+1 ), and those ones considered here, u˜n+1/2 = (I + τ Ln2+1
,ε
)g(x, y, tn+1 ) − τ f2n+1 . For this reason, an
order reduction occurs in the consistency, up to order 0, if the natural boundary conditions are chosen.

Remark 2. In [3,7], the authors consider a problem less general than here (therein, the boundary conditions are g = 0). In
those papers, to avoid the order reduction, a suitable splitting f = f 1 + f2 , in such way that f2 = 0 in {0, 1} × [0, 1] ×
[0, T], was necessary to complete the analysis. This restriction agrees completely with the analysis for the case which we
have studied here. Moreover, an additional advantage of the boundary data (6) which we propose here is that any smooth
splitting for f is valid.

3.4. Uniform convergence of the time semidiscretization

Now we prove the uniform convergence of this discretization stage as an additional interesting feature, although the
results of this subsection are not necessary for proving the uniform convergence of the fully discrete scheme.
Let us introduce the global error of the scheme (3) at time tn as usual, i.e., E n ≡ u(tn ) − un . Subtracting and adding u˜n ,
trivially we have
E n = en + u˜n − un = en + (I + τ Ln2,ε,0 )−1 (I + τ Ln1,ε,0 )−1 E n−1 .
Using this recurrence, combined with (7) and (9), it follows
E n ¯ ≤ C τ ,
and therefore the semidiscrete scheme (3) is uniformly convergent of first order.

3.5. Asymptotic behavior of semidiscrete solutions

In this subsection we study the behavior, with respect to the singular perturbation parameter ε , of the semidiscrete in
1
time solutions of (8) and their derivatives. In fact, we need to deduce appropriate (ε -dependent) bounds for u˜n+ 2 (x, y ),
and its x-derivatives up to fourth order, and also for u˜n+1 (x, y ) and its y-derivatives up to fourth order. Such bounds will
be used in the uniform convergence analysis of the spatial discretization stage in the following section. Although, at first
sight, the problem (8) can seem a double parameter (ε , τ ) singularly perturbed problem, we are going to prove that, under
the smoothness and compatibility assumptions made on data, such derivatives preserve in essence the same exponential
behavior as the solution of the continuous problem. The rest of this subsection contains the details for the proofs of the
following bounds:
 
 ∂ i u˜n+1/2   
 ( x, y ) ≤ C 1 + ε −i (exp(−β1 x/ε ) + exp(−β1 (1 − x )/ε ) ) ,
 ∂ xi  (14)
0 ≤ i ≤ 4,
 
 ∂ i u˜n+1   
 
 ∂ yi (x, y ) ≤ C 1 + ε (exp(−β2 y/ε ) + exp(−β2 (1 − y )/ε ) ) ,
−i
(15)
0 ≤ i ≤ 4,
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 17

1
¯ . Firstly, resorting to classical maximum principles, it is immediate to deduce the bound u˜n+ 2  ¯ ≤ C,
for (x, y ) ∈  
since we have previously assumed that u(tn ), f1n+1 and gn+1/2 (x, y ) are ε -uniformly bounded. To obtain bounds for its
1
x-derivatives, we make use of ω ≡ Ln1+1
,ε
u˜n+ 2 , which is the solution of the problem

(I + τ Ln1+1
,ε
)ω = Ln1+1
,ε
u(x, y, tn ) + τ Ln1+1
,ε 1
f n+1 ,
gn+1/2
(x, y ) − g(x, y, tn )
ω = f1n+1 − , (x, y ) ∈ {0, 1} × [0, 1];
τ
and now, taking into account that |Ln1+1 ,ε
u(x, y, tn )| ≤ C , |Ln1+1 f n+1 | ≤ C in 
,ε 1
¯ , and also that it holds |ω(x, y)| ≤ C in {0, 1} ×
[0, 1], because
gn+1/2 (x, y ) − g(x, y, tn ) g(x, y, tn+1 ) − g(x, y, tn )
f1n+1 − = f (x, y, tn+1 ) − Ln2+1
,ε g(x, y, tn+1 ) − , (16)
τ τ
1
it follows ω¯ ≤ C. Then, u˜n+ 2 is the solution of the problem

Ln1+1 u˜n+ 2 = ω,
1

,ε (17)
u˜n+1/2 (x, y ) = gn+1/2 (x, y ), (x, y ) ∈ {0, 1} × [0, 1],
and proceeding in similar way as in [8,10], for reaction-diffusion problems, we obtain
   
 ∂ i u˜n+ 12   ∂ i u˜n+ 12 
   
 ∂ xi ( 0, y ) ≤ C ε ,  ∂ xi (1, y ) ≤ C ε , y ∈ [0, 1], 0 ≤ i ≤ 2.
−i −i

Let us prove that the last bounds are also valid for i = 3, 4. Firstly, we use that ω̄ = Ln1+1
,ε
ω is the solution of the boundary
value problem
(I + τ Ln1+1
,ε
)ω̄ = (Ln1+1
,ε
)2 u(x, y, tn ) + τ (Ln1+1
,ε
)2 f1n+1,
ω (x, y ) − Ln1+1
,ε
u(x, y, tn ) (18)
ω̄ = Ln1+1 n+1
,ε f 1 − , (x, y ) ∈ {0, 1} × [0, 1],
τ
where the right hand side of the differential equation is ε -uniformly bounded. Note that the bounds for the solution of
problem (1) given in the previous section, guarantee that |(Ln1+1
,ε
)2 u(x, y, tn )| ≤ C. As well, from the boundary data given in
(18) and the smoothness and the compatibility conditions assumed for (1) , we prove that |ω̄ (x, y )| ≤ C, for (x, y) ∈ {0, 1}
× [0, 1], (from the development (16), it is easy to deduce that ω = Ln1+1,ε
u(x, y, tn+1 ) + O (τ ) in {0, 1} × [0, 1]). Resorting to
classical maximum principles, it is deduced that ω̄¯ ≤ C and using again the technique proposed in [10], it follows
   
 ∂ iω   ∂ iω 
 ( 0 , y )  ≤ C ε −i ,  ( 1 , y )  ≤ C ε −i , y ∈ [0, 1], i = 1, 2.
 ∂ xi   ∂ xi 
To end this part, we only need to take into account that

∂ 3 u˜n+ 2
1
1
 ∂ω ∂ u˜n+ 2 ∂ k1 n+ 12 
1

= 2 − + k1 + u˜ ,
∂x 3 ε ∂x ∂x ∂x
and also that
∂ 4 u˜n+ 2
1
1
 ∂ 2ω ∂ 2 u˜n+ 2 ∂ k1 ∂ u˜n+ 2 ∂ 2 k1 n+ 12 
1 1

= − + k + 2 + u˜ ,
∂ x4 ε2 ∂ x2 1
∂ x2 ∂x ∂x ∂ x2
where k1 ≡ k1 (x, y, tn+1 ). Using the bounds for the terms of the right hand side of these two equalities, which we have
deduced previously for (x, y) ∈ {0, 1} × [0, 1], it is straightforward deduce that
   
 ∂ i u˜n+ 12   ∂ i u˜n+ 12 
   
 ∂ xi ( 0, y ) ≤ C ε ,  ∂ xi (1, y ) ≤ C ε , y ∈ [0, 1], i = 3, 4.
−i −i

To obtain the desired bounds for (x, y) ∈ (0, 1) × [0, 1], we derive the first half step of (8) with respect to x, obtaining

∂ u(x, y, tn ) ∂ f n+1 ∂ k1 n+ 12
(I + τ Ln1+1
,ε )sx (x, y ) = − +τ 1 −τ u˜ ≡ h(x, y ), (19)
∂x ∂x ∂x
1
∂ uˆn+ 2
where sx (x, y ) ≡ and
∂x

|h(x, y )| ≤ C 1 + ε −1 (exp(−β1 x/ε ) + exp(−β1 (1 − x )/ε ) ) .
Defining
s1 (x, y ) = 1 + x, s2 (x, y ) = ε −1 (exp(−β1 x/ε ) + exp(−β1 (1 − x )/ε ) ),
18 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

we can find ε -independent positive constants C1 , C2 , such that

(I + τ Ln1+1
,ε )(C1 s1 + C2 s2 ± sx (x, y )) ≥ 0, (20)
C1 s1 (0, y ) + C2 s2 (0, y ) ≥ |sx (0, y )|,
C1 s1 (1, y ) + C2 s2 (1, y ) ≥ |sx (1, y )|,
which implies
 
|sx (x, y )| ≤ C1 s1 + C2 s2 ≤ C 1 + ε −1 (exp(−β1 x/ε ) + exp(−β1 (1 − x )/ε ) ) ,
1
i n+
¯ . The rest of bounds for ∂ u˜ 2 , i = 2, 3, 4, can be obtained by successive x-derivations of the first half step
for (x, y ) ∈ 
∂ xi
of (8) choosing suitable barrier functions, taking into account the available bounds for the right hand side, as well as the
1 1
i n+ i n+
bounds for the boundary conditions ∂ u˜ i 2 (0, y ) and ∂ u˜ i 2 (1, y ), which have been deduced previously.
∂x ∂x
1
i n+
The bounds for the derivatives ∂ u˜ i 2 , i = 2, 3, 4 are much simpler. Suffice it to derive the equation and the boundary
∂y
conditions of we derive the first half step of (8) with respect to y and apply a classical maximum principle for the corre-
sponding boundary value problems (where x can be considered as the only variable and y as a parameter). For instance,
n+1
sy ≡ ∂ u˜∂ y 2 is the solution of the problem

∂ u(tn ) ∂ f n+1 ∂ k1 (x, y, tn+1 ) n+ 12

(I + τ Ln1+1
,ε )sy = +τ 1 −τ u˜ , (x, y ) ∈ ,
∂y ∂y ∂y
∂ gn+1/2 (x, y )
sy = , (x, y ) ∈ {0, 1} × [0, 1].
∂y
In this way, the following bounds are proven
 
 ∂ i u˜n+ 12   
 
 ∂ yi (x, y ) ≤ C 1 + ε (exp(−β2 y/ε ) + exp(−β2 (1 − y )/ε ) ) ,
−i

(x, y ) ∈ 
¯ , 0 ≤ i ≤ 4.

From these estimates, it is straightforward to deduce that

Ln2+1 ˜n+ 2 (x, y )¯ ≤ C, (Ln2+1 (x, y )¯ ≤ C.

1 1

,ε ) u
2 n+ 2
,ε u
˜

Next, we prove the bounds (15) for u˜n+1 , by using the same technique. Renaming ω ≡ Ln2+1
,ε
u˜n+1 and ω̄ ≡ (Ln2+1
,ε
)2 u˜n+1 ,
such functions can be described as the solutions of the boundary value problems

(I + τ Ln2+1 )ω = Ln2+1 u˜n+ 2 + τ Ln2+1 f n+1 ,

1

,ε ,ε ,ε 2
1
u˜n+1 − u˜n+ 2
ω = f2n+1 − , (x, y ) ∈ [0, 1] × {0, 1},
τ
and
(I + τ Ln2+1 )ω̄ = (Ln2+1 )2 u˜n+ 2 + (τ Ln2+1 )2 f2n+1 ,
1

,ε ,ε ,ε

ω − Ln2+1
1

,ε
u˜n+ 2
ω̄ = Ln2+1 n+1
,ε f 2 − , (x, y ) ∈ [0, 1] × {0, 1},
τ
respectively, where (in both problems) the non homogeneous terms of the differential equations and the boundary data are
ε uniformly bounded. This permits us to ensure that ω(x, y) and ω̄ (x, y ) are bounded ε-uniformly in 
¯ and, resorting to the
same technique as before, it can be deduced that
   
 ∂ i u˜n+1   ∂ i u˜n+1 
 ( x, 0 )  ≤ C ε −i ,  ( x, 1 )  ≤ C ε −i , x ∈ [0, 1], 0 ≤ i ≤ 4.
 ∂ yi   ∂ yi 

Finally, reproducing the barrier function technique described in detail in (19) and (20), now for the y-derivatives of u˜n+1 ,
making use of the second half step of (8), the bounds (15) are obtained.

4. Spatial discretization: uniform convergence

In this section we discretize the 1D problems of (3) on a rectangular mesh w̄N ≡ I1,ε,N × I2,ε,N ⊂ 
¯ , having (N + 1 )2 nodes
(for simplicity, we assume that the number of mesh points is the same at both spatial directions), where

I1,ε,N = {0 = x0 < · · · < xN = 1}, I2,ε,N = {0 = y0 < · · · < yN = 1}.

 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 19

From the asymptotic behavior of the exact solution of semidiscrete problems, we know that a boundary layer appears at
the four sides of the spatial domain, which has a size O(ε ). We give the details of the construction of I1, ε, N and analogously
we proceed for I2, ε, N . We define the transition parameter
1 
σx = min , mx ε log N , (21)
4
where mx ≥ 1/β 2 and the mesh points are defined by
⎧
⎪ 4 σx N
⎪
⎪i , i = 0, . . . , ,
⎪
⎨ N
 N  2(14− 2σ ) N 3N
x i = σx + i −
x
, i= + 1, . . . , , (22)
⎪
⎪ 
4

N 4 4
⎪
⎪
⎩1 − σx + i − 3N 4σx , i = 3N + 1, . . . , N.
4 N 4
We denote by hx,i = xi − xi−1 , hy,i = yi − yi−1 , i = 1, . . . , N.
Let us denote in the form •N the discrete functions defined in w̄N , as [.]N the discrete function resulting as the operation
of restriction of a function defined in  ¯ to w̄N and by U n the discrete function which will approach to [u(x, y, tn )]N by using
N
our numerical algorithm. On the mesh w̄N × I¯M , the fully discrete scheme is given by
(i ) UN0 = [ϕ (x, y )]N ,
(ii ) (I + τ Ln1+1
,ε ,N
)UNn+1/2 = UNn + τ [ f1n+1 ]N (x, y ), (x, y ) ∈ wN ,
U n+1/2 (x, y ) = gn+1/2 (x, y ), (x, y ) ∈ {0, 1} × I2,ε,N , (23)
(iii ) (I + τ Ln2+1
,ε ,N
)UNn+1 = UNn+1/2 + τ [ f2n+1 ]N (x, y ), (x, y ) ∈ wN ,
U n+1 (x, y ) = gn+1 (x, y ), (x, y ) ∈ I1,ε,N × {0, 1}, n = 0, 1, . . . , M − 1,

where UNn ≈ u(x, y, tn ), (x, y ) ∈ w̄N , n = 0, 1, . . . , M, gn+1/2 and gn+1 are defined in (6) and Ln1+1
,ε ,N
, Ln2+1
,ε ,N
are given by

Ln1+1
,ε ,N
W ≡ li−, jW (xi−1 , y j ) + lic, jW (xi , y j ) + li+, jW (xi+1 , y j ), i, j = 1, . . . , N − 1,
(24)
Ln2+1
,ε ,N
W ≡ li, j−W (xi , y j−1 ) + li, jcW (xi , y j ) + li, j+W (xi , y j+1 ), i, j = 1, . . . , N − 1,
with
−ε
li−, j = li+, j = , l = k1 (xi , y j , tn+1 ) − li−, j − li+, j , i, j = 1, . . . , N − 1,
˜ x,i ic, j
hx,i h
(25)
−ε
li, j− = li, j+ = , l = k2 (xi , y j , tn+1 ) − li, j− − li, j+ , i, j = 1, . . . , N − 1,
˜ y, j i, jc
hy, j h

where h ˜ x,i = (hx,i + hx,i+1 )/2, i = 1, . . . , N, h

˜ y, j = (hy, j + hy, j+1 )/2, j = 1, . . . , N.
The analysis of the uniform convergence of the spatial discretization given in (24) and (25) follows the same route as
in [7]. For that, we introduce the intermediate approach U˜Nn+1 which is the result of applying one step of scheme (23), but
substituting UNn by [u(x, y, tn ]N .
To deduce the uniform convergence of the spatial discretization, we analyze the spatial local truncation error. For the
first half step, this local error is defined as
υ ≡ (I + τ Ln1+1 ˜ n+1/2 − [(I + τ Ln+1 )u˜n+1/2 ]N .
,ε ,N )UN 1,ε

Using Taylor expansions and considering the two cases xi ∈ (0, σx ] ∪ [1 − σx , 1 ) or xi ∈ (σx , 1 − σx ), taking into account the
bounds for the derivatives respect to x of u˜n+1/2 , proved in previous section, and the definition of the mesh, it is deduced
that
υω̄N ≤ CM−1 (N−1 ln N )2 .
Combining this bound with the uniform stability of the discrete operator (I + τ Ln1+1
,ε ,N
), it follows

[u˜n+1/2 ]N − U˜Nn+1/2 ω̄N ≤ C M−1 (N−1 ln N )2 .

Analogously, using the bounds for the derivatives respect to y of u˜n+1 , the definition of the mesh and the uniform sta-
bility of (I + τ Ln2+1
,ε ,N
), it is proved that
[u˜n+1 ]n − U˜Nn+1 ω̄N ≤ CM−1 (N−1 ln N )2 , (26)
and therefore the spatial semidiscretization, defined on the piecewise uniform Shishkin mesh, is uniformly convergent of
almost second order.
We want to highlight the presence of the factor M−1 in (26), which is essential for ending the analysis of the uniform
convergence of the fully discrete scheme (23). Now, we are ready to prove the main result of the paper.
20 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

Theorem 2. Assuming that the solution of the continuous problem (1) u ∈ C 4,2 (
¯ × [0, T ] ), the global error associated to the
numerical method (23), satisfies
 n   
max UN (xi , y j ) − u(xi , y j , tn ) ≤ C (N−1 ln N )2 + M−1 , (27)
0≤i, j≤N,0≤n≤M

where C is a positive constant independent of the diffusion parameter and the discretization parameters N and M.

Proof. We split the contribution to the global error of the time and spatial discretization processes by incorporating the
intermediate approaches u˜n and U˜Nn . Then, the global error at time tn , ENn ≡ [u(x, y, tn )]N − UNn , can be decomposed in the
form
ENn = [u(x, y, tn ) − u˜n ]N + ([u˜n ]n − U˜Nn ) + (U˜Nn − UNn ).
Therefore,
ENn ω̄N ≤ [en ]N ω̄N + [u˜n ]n − U˜Nn ω̄N + (I + τ Ln2+1
,ε ,N ) (I + τ L1,ε ,N ) EN ω̄N .
−1 n+1 −1 −1

From the consistency results of previous section, it is straightforward that [en ]N ω̄N ≤ CM−2 . Next, as the operators Lni,+1
ε,N ,
i = 1, 2 are inverse monotone, it holds that (I + τ Lni,+1
ε,N ) −1 
ω̄N , ≤ 1 , i = 1 , 2 . Therefore, using (26) , it holds
 
ENn ω̄N ≤ CM−1 (N−1 ln N )2 + M−1 + EN−1 ω̄N .
From this recurrence, the bound (27) follows. 

5. Numerical results

In this section we show the numerical results obtained with the algorithm proposed here to solve successfully some
problems of type (1). For the sake of simplicity, for all test examples we have chosen the same decomposition for the
reaction term: k1 (x, y, t ) = k2 (x, y, t ) = k(x, y, t )/2.
The first example that we consider is given by
ut − ε 2 u + (1 + x2 y )u = f (x, y, t ), (x, y, t ) ∈  × [0, 1],
u(x, y, t ) = g(x, y, t ), (x, y ) in ∂  × [0, 1], (28)
u(x, y, 0 ) = ϕ (x, y ), in ,
where f, g and ϕ are such that the exact solution is
u(x, y, t ) = (1 + t − et )(φ (x, ε ) + ψ (y, ε ) − 2 + xy ),
with
e−x/ε + e−(1−x )/ε e−2y/ε + e−2(1−y )/ε
φ (x, ε ) = , ψ (y, ε ) = .
1 + e−1/ε 1 + e−2/ε
Fig. 1 shows the solution at the final time t = 1; from it, we clearly see the boundary layer at the four sides of the spatial
domain.
In all tables we take mx = my = 1 to define the transition parameters of the meshes I1, ε, N and I2, ε, N respectively. In this
case, we decompose the right-hand side in the form f (x, y, t ) = f 1 (x, y, t ) + f2 (x, y, t ), where
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 ).
To show the numerical behavior of the method, we estimate firstly the local errors in time. As the exact solution is
known, the local errors at the mesh points have been estimated by
e˜N,M = max max max |U˜Nn − u(xi , y j , tn )|.
0≤n≤M 0≤i≤N 0≤ j≤N

From them, we obtain the quantities

log (e˜N,M /e˜N,2M )
p˜ = ,
log 2
and note that the corresponding numerical orders of consistency are given by p˜ − 1.
In next two tables we show these numerical local errors and the values of p˜ corresponding to the two choices of the
boundary data. To reduce the influence of the local spatial error, we take the discretization parameter N sufficiently large;
concretely we take N = 512. Table 1 displays the local errors when boundary conditions (5) are taken, and Table 2 displays
the local errors when (6) are considered. Note that the local errors are substantially smaller when the non natural option
is used; moreover, the order of consistency is one, and for classical evaluation is zero. Then, clearly the option (6) is much
better than the classical one.
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 21

Fig. 1. Solution of example (28) for ε = 10−2 .

Table 1
Local errors and values of p˜ for (28) with natural boundary conditions.

ε M=8 M = 16 M = 32 M = 64 M = 128

20 9.6560E−2 6.6436E−2 4.0565E−2 2.2529E−2 1.1774E−2

0.539 0.712 0.848 0.936
2−2 2.4869E−1 1.4057E−1 7.4335E−2 3.7883E−2 1.8939E−2
0.823 0.919 0.973 1.0 0 0
−4
2 3.5403E−1 1.8313E−1 9.2209E−2 4.5699E−2 2.2368E−2
0.951 0.990 1.013 1.031
2−6 3.6569E−1 1.8472E−1 9.1282E−2 4.4306E−2 2.1107E−2
0.985 1.017 1.043 1.070
2−8 3.5250E−1 1.7449E−1 8.3926E−2 3.9249E−2 1.7762E−2
1.014 1.056 1.096 1.144
−10
2 3.3808E−1 1.6369E−1 7.6409E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224
2−12 3.3809E−1 1.6369E−1 7.6410E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224
2−14 3.3809E−1 1.6369E−1 7.6410E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224
2−16 3.3809E−1 1.6369E−1 7.6410E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224
−18
2 3.3809E−1 1.6369E−1 7.6410E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224
2−20 3.3809E−1 1.6369E−1 7.6410E−2 3.4306E−2 1.4687E−2
1.046 1.099 1.155 1.224

Next, to show the uniform convergence of the method, global errors are displayed. As the exact solution is known, the
maximum global errors at the mesh points can be computed exactly by

eN,M = max max max |UNn − u(xi , y j , tn )|,

0≤n≤M 0≤i≤N 0≤ j≤N

and following the double mesh philosophy, the numerical orders of convergence are calculated by

log (eN,M /e2N,2M )

p= .
log 2
22 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

Table 2
Local errors and values of p˜ for (28) with improved boundary conditions.

ε M=8 M = 16 M = 32 M = 64 M = 128

20 1.2448E−2 5.8955E−3 2.1928E−3 6.7573E−4 1.8678E−4

1.078 1.427 1.698 1.855
2−2 3.7268E−2 1.1918E−2 3.3290E−3 8.6976E−4 2.2139E−4
1.861 1.935 1.969 1.985
−4
2 5.7936E−2 1.5947E−2 4.1692E−3 1.0648E−3 2.6903E−4
1.852 1.932 1.968 1.984
2−6 6.4105E−2 1.7288E−2 4.4906E−3 1.1445E−3 2.8891E−4
1.891 1.945 1.972 1.986
2−8 6.6297E−2 1.7817E−2 4.6212E−3 1.1769E−3 2.9697E−4
1.896 1.947 1.973 1.987
2−10 6.7262E−2 1.8056E−2 4.6808E−3 1.1918E−3 3.0070E−4
1.897 1.948 1.974 1.987
−12
2 6.7654E−2 1.8155E−2 4.7063E−3 1.1985E−3 3.0257E−4
1.898 1.947 1.973 1.984
2−14 6.7802E−2 1.8195E−2 4.7176E−3 1.2021E−3 3.0384E−4
1.903 1.962 1.992 2.002
2−16 6.7856E−2 1.8214E−2 4.7249E−3 1.2053E−3 3.0533E−4
1.897 1.947 1.971 1.981
−18
2 6.7890E−2 1.8230E−2 4.7331E−3 1.2093E−3 3.0732E−4
1.897 1.945 1.969 1.976
2−20 6.7921E−2 1.8247E−2 4.7419E−3 1.2138E−3 3.0954E−4
1.896 1.944 1.966 1.971

Table 3
Maximum errors and orders of convergence for (28) with natural boundary conditions (5).

ε N = 16 N = 32 N = 64 N = 128 N = 256
M=8 M = 16 M = 32 M = 64 M = 128

20 8.6671E−2 6.2381E−2 3.9401E−2 2.2406E−2 1.1910E−2

0.474 0.663 0.814 0.912
2−2 2.1393E−1 1.3021E−1 7.2402E−2 3.8135E−2 1.9550E−2
0.716 0.847 0.925 0.964
2−4 2.8856E−1 1.6606E−1 8.9976E−2 4.6944E−2 2.3992E−2
0.797 0.884 0.939 0.968
−6
2 2.7115E−1 1.5808E−1 8.8232E−2 4.7044E−2 2.4346E−2
0.778 0.841 0.907 0.950
2−8 2.5536E−1 1.4153E−1 7.9831E−2 4.4442E−2 2.3629E−2
0.851 0.826 0.845 0.911
2−10 2.5658E−1 1.4187E−1 7.9336E−2 4.3428E−2 2.3094E−2
0.855 0.838 0.869 0.911
2−12 2.5736E−1 1.4206E−1 7.9376E−2 4.3437E−2 2.3097E−2
0.857 0.840 0.870 0.911
−14
2 2.5833E−1 1.4224E−1 7.9409E−2 4.3443E−2 2.3098E−2
0.861 0.841 0.870 0.911
2−16 2.5883E−1 1.4234E−1 7.9429E−2 4.3447E−2 2.3098E−2
0.863 0.842 0.870 0.911
2−18 2.5909E−1 1.4239E−1 7.9440E−2 4.3449E−2 2.3099E−2
0.864 0.842 0.871 0.911
−20
2 2.5922E−1 1.4241E−1 7.9446E−2 4.3451E−2 2.3099E−2
0.864 0.842 0.871 0.912
emaxN,M 2.8856E−1 1.6606E−1 8.9976E−2 4.7044E−2 2.4346E−2
puni 0.797 0.884 0.936 0.950

From these values we calculate the uniform maximum errors by

emaxN,M = max eN,M ,
ε
and from them, in a usual way, the corresponding uniform orders of convergence will be given by
 
uni
log emaxN,M /emax2N,2M
p = .
log 2
Table 3 displays the results for problem (28) by using our fully discrete scheme when the classical boundary conditions
(5) are used. From it, we deduce that the method is uniformly convergent, showing uniform convergence of first order, be-
cause of the choice of the discretization parameters made here, the time errors dominate in the global error of the numerical
scheme.
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 23

Table 4
Maximum errors and orders of convergence for (28) with improved boundary conditions (6).

ε N = 16 N = 32 N = 64 N = 128 N = 256
M=8 M = 16 M = 32 M = 64 M = 128

20 1.4500E−2 8.9250E−3 5.0250E−3 2.6776E−3 1.3840E−3

0.700 0.829 0.908 0.952
2−2 6.5848E−2 3.5221E−2 1.8214E−2 9.2572E−3 4.6663E−3
0.903 0.951 0.976 0.988
−4
2 1.5830E−1 8.0282E−2 4.0374E−2 2.0236E−2 1.0129E−2
0.979 0.992 0.996 0.998
2−6 2.0499E−1 1.0228E−1 5.1071E−2 2.5510E−2 1.2748E−2
1.003 1.002 1.001 1.001
2−8 2.1899E−1 1.0903E−1 5.4333E−2 2.7131E−2 1.3556E−2
1.006 1.005 1.002 1.001
−10
2 2.2514E−1 1.1212E−1 5.5764E−2 2.7792E−2 1.3875E−2
1.006 1.008 1.005 1.002
2−12 2.3052E−1 1.1441E−1 5.6624E−2 2.8103E−2 1.4006E−2
1.011 1.015 1.011 1.005
2−14 2.3385E−1 1.1609E−1 5.7433E−2 2.8400E−2 1.4091E−2
1.010 1.015 1.016 1.011
2−16 2.3562E−1 1.1712E−1 5.8096E−2 2.8744E−2 1.4215E−2
1.008 1.012 1.015 1.016
−18
2 2.3654E−1 1.1769E−1 5.8524E−2 2.9045E−2 1.4376E−2
1.007 1.008 1.011 1.015
2−20 2.3700E−1 1.1799E−1 5.8766E−2 2.9244E−2 1.4520E−2
1.006 1.006 1.007 1.010
emaxN,M 2.3700E−1 1.1799E−1 5.8766E−2 2.9244E−2 1.4520E−2
puni 1.006 1.006 1.007 1.010

As we are interested in showing the influence of the discretization of the boundary conditions on the maximum errors
and the orders of convergence, in second place, we consider the choice the boundary data given in (6). Table 4 displays the
results for problem (28) by using this option. From it, again we deduce that the method is uniformly convergent, showing
also the first order of convergence. Moreover, the maximum errors for any value of ε are smaller than those ones given in
Table 3. Maybe, it can seem contradictory that in Table 3, apparently no order reduction occurs (note that the numerical
orders of convergence are similar). This phenomenon has been widely observed and studied in the numerical integration
of parabolic problems, via one step methods. In [15,18], the authors justify, using a summation by parts technique, that a
recovering of one unit in the global error, with respect to the order obtained for the local error, is present many times
when the order reduction phenomenon takes place. Up to our knowledge, such technique has not been used in a context
of uniform convergence analysis of singularly perturbed problems of this type; indeed, such analysis, in case of being tried,
would be much more complicated than the performed here. Thus, our proposal is better both from theoretical and practical
points of view.
The second example that we consider is given by
 
ut − ε 2 u + 1 + tx2 + xyt 2 u = f (x, y, t ), (x, y, t ) ∈  × [0, 1],
u(x, y, t ) = g(x, y, t ), (x, y ) in ∂  × [0, 1], (29)
u(x, y, 0 ) = 0, in ,
where

f (x, y, t ) = t (1 − et )(sin(π x ) + sin(π y )), g(x, y, t ) = t 3 sin(π (x + y )).

In this case the exact solution is unknown; Fig. 2 shows the numerical solution at the final time t = 1; from it, again
boundary layers can be observed.
Again we take mx = my = 1 to define the piecewise uniform Shishkin mesh, and now we decompose the source term in
a different way, taking f1 (x, y, t ) = f2 (x, y, t ) = f (x, y, t )/2.
To approximate the maximum pointwise errors, we compare the numerical solution unN with the numerical solution ob-
tained on a finest mesh of Shishkin type, taking N = 512, M = 256, which is denoted by u512, 256 . Then, if ū512,256 denotes
the piecewise bilinear interpolation of u512, 256 , the maximum errors at the mesh points of the coarse mesh are approxi-
mated by computing

dN,M = max max |UNn (xi , y j ) − ū512,256 (xi , y j , tn )|, (30)

0≤n≤M 0≤i, j≤N

and the orders of convergence are calculated by

log (dN,M /d2N,2M )
q= .
log 2
24 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

Fig. 2. Solution of example (29) for ε = 10−2 .

From errors in (30), the uniform maximum errors are given by

dmaxN,M = max dN,M ,

ε

and from them, in a usual way, the corresponding numerical uniform orders of convergence by
 
uni
log dmaxN,M /dmax2N,2M
q = .
log 2
Table 5 displays the results for problem (29) by using our method when classical boundary conditions are used. From it,
we again deduce first order of convergence for any value of ε .
In second place, we consider the no natural choice of the boundary data given in (6). Table 6 displays the results for
problem (29) by using this option. Again a first order of uniform convergence is shown in this case; this is due to the
choice which we have done for the discretization parameters M and N which make that the errors associated to the time
discretization dominate in the global error for any value of ε . Note also that the maximum errors are smaller than the
corresponding ones given in Table 5, showing again that this choice for the boundary conditions is more adequate than the
classical one.
Note that the numbers at the last columns of Tables 5 and 6 (maximum errors and orders of convergence) are better than
expected because the coarse mesh is too close to the finest mesh; this behavior is usual when the errors are approximated
by comparison to the finest mesh.
To show the influence of the spatial discretization in the global error, we calculate with different relations between N
and M; now when we duplicate N, M is multiplied by 4; moreover, we begin with a larger value of M in such way that,
in the first column of the tables, the errors associated to the time discretization have a small influence on the global error.
Table 7 displays the results obtained in this case; from it we clearly see the almost second order of uniform convergence,
according the theoretical results.
Finally, in a similar way that in the first example, we estimate the local errors for this second example. As now we do
not know the exact solution, to approximate U˜Nn we use one step of the fully discrete scheme given in (23), but replacing the
numerical approximation at time tn−1 , UNn−1 , by the numerical solution obtained on the finest mesh on the corresponding
time level, which is a sufficiently good approach to [u(tn−1 )]N for our purposes. We take again a large fixed value of the
discretization parameter N, concretely N = 128, in order to the errors in time dominate. Then, the local errors have been
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 25

Table 5
Maximum errors and orders of convergence for (29) with natural boundary conditions.

ε N = 16 N = 32 N = 64 N = 128 N = 256
M=8 M = 16 M = 32 M = 64 M = 128

20 4.2302E−1 3.2842E−1 2.0556E−1 1.0441E−1 3.8183E−2

0.365 0.676 0.977 1.451
2−2 2.1884E−1 1.5003E−1 8.4596E−2 4.0025E−2 1.4035E−2
0.545 0.827 1.080 1.512
−4
2 1.0860E−1 7.6061E−2 4.3512E−2 2.0716E−2 7.2834E−3
0.514 0.806 1.071 1.508
2−6 1.1082E−1 5.4246E−2 2.8843E−2 1.4800E−2 5.3953E−3
1.031 0.911 0.963 1.456
2−8 1.1725E−1 5.7166E−2 2.6745E−2 1.1472E−2 4.5422E−3
1.036 1.096 1.221 1.337
−10
2 1.1870E−1 5.7774E−2 2.7036E−2 1.1599E−2 5.0834E−3
1.039 1.096 1.221 1.190
2−12 1.1898E−1 5.7838E−2 2.7103E−2 1.1635E−2 5.0526E−3
1.041 1.094 1.220 1.203
2−14 1.1900E−1 5.7827E−2 2.7131E−2 1.1643E−2 5.0422E−3
1.041 1.092 1.221 1.207
2−16 1.1899E−1 5.7810E−2 2.7139E−2 1.1643E−2 5.0383E−3
1.041 1.091 1.221 1.208
−18
2 1.1898E−1 5.7800E−2 2.7137E−2 1.1640E−2 5.0367E−3
1.042 1.091 1.221 1.209
2−20 1.1896E−1 5.7790E−2 2.7135E−2 1.1638E−2 5.0359E−3
1.042 1.091 1.221 1.209
dmaxN,M 4.2302E−1 3.2842E−1 2.0556E−1 1.0441E−1 3.8183E−2
quni 0.365 0.676 0.977 1.451

Table 6
Maximum errors and orders of convergence for (29) with improved boundary condi-
tions.

ε N = 16 N = 32 N = 64 N = 128 N = 256
M=8 M = 16 M = 32 M = 64 M = 128

20 1.1474E−1 7.3307E−2 4.0583E−2 1.9195E−2 6.7458E−3

0.646 0.853 1.080 1.509
2−2 7.8082E−2 4.1773E−2 2.0498E−2 9.0143E−3 3.0445E−3
0.902 1.027 1.185 1.566
−4
2 8.9549E−2 4.4256E−2 2.0845E−2 8.9733E−3 2.9977E−3
1.017 1.086 1.216 1.582
2−6 1.1064E−1 5.4201E−2 2.5432E−2 1.0927E−2 3.6463E−3
1.029 1.092 1.219 1.583
2−8 1.1725E−1 5.7165E−2 2.6745E−2 1.1472E−2 3.8257E−3
1.036 1.096 1.221 1.584
−10
2 1.1870E−1 5.7774E−2 2.7036E−2 1.1599E−2 3.8712E−3
1.039 1.096 1.221 1.583
2−12 1.1898E−1 5.7838E−2 2.7103E−2 1.1635E−2 3.8838E−3
1.041 1.094 1.220 1.583
2−14 1.1900E−1 5.7827E−2 2.7131E−2 1.1643E−2 3.8864E−3
1.041 1.092 1.221 1.583
2−16 1.1899E−1 5.7810E−2 2.7139E−2 1.1643E−2 3.8841E−3
1.041 1.091 1.221 1.584
−18
2 1.1898E−1 5.7800E−2 2.7137E−2 1.1640E−2 3.8823E−3
1.042 1.091 1.221 1.584
2−20 1.1896E−1 5.7790E−2 2.7135E−2 1.1638E−2 3.8811E−3
1.042 1.091 1.221 1.584
dmaxN,M 1.1900E−1 7.3307E−2 4.0583E−2 1.9195E−2 6.7458E−3
quni 0.699 0.853 1.080 1.509

estimated by
d˜N,M = max max max |U˜Nn − ū512,256 (xi , y j , tn )|,
0≤n≤M 0≤i≤N 0≤ j≤N

and from them we calculate

 
log d˜N,M /d˜N,2M
q˜ = ,
log 2
and the orders of consistency are given by q˜ − 1.
26 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27

Table 7
Maximum errors and orders of convergence for (29) with improved
boundary conditions.

ε N = 16 N = 32 N = 64 N = 128
M = 32 M = 128 M = 512 M = 2048

20 2.1893E−2 6.9355E−3 1.8484E−3 4.6966E−4

1.658 1.908 1.977
−2
2 1.2528E−2 3.3669E−3 8.5653E−4 2.1513E−4
1.896 1.975 1.993
2−4 1.3242E−2 3.3980E−3 8.5488E−4 2.1408E−4
1.962 1.991 1.998
2−6 1.7369E−2 4.7581E−3 1.2206E−3 3.0720E−4
1.868 1.963 1.990
2−8 2.5243E−2 1.2195E−2 4.3154E−3 1.1022E−3
1.050 1.499 1.969
−10
2 2.5664E−2 1.2220E−2 4.6491E−3 1.6044E−3
1.071 1.394 1.535
2−12 2.5911E−2 1.2251E−2 4.6572E−3 1.6060E−3
1.081 1.395 1.536
2−14 2.6044E−2 1.2272E−2 4.6625E−3 1.6075E−3
1.086 1.396 1.536
−16
2 2.6112E−2 1.2283E−2 4.6655E−3 1.6084E−3
1.088 1.397 1.536
2−18 2.6146E−2 1.2289E−2 4.6670E−3 1.6089E−3
1.089 1.397 1.536
2−20 2.6163E−2 1.2292E−2 4.6678E−3 1.6091E−3
1.090 1.397 1.536
dmaxN,M 2.6163E−2 1.2292E−2 4.6678E−3 1.6091E−3
quni 1.090 1.397 1.536

Table 8
Local errors and values of q˜ for (29) with natural boundary conditions.

ε M=8 M = 16 M = 32 M = 64 M = 128

20 5.4758E−1 3.6680E−1 2.1084E−1 1.0130E−1 3.5436E−2

0.578 0.799 1.057 1.515
2−2 3.1390E−1 1.7295E−1 8.5556E−2 3.7010E−2 1.2007E−2
0.860 1.015 1.209 1.624
−4
2 1.8656E−1 9.3007E−2 4.2762E−2 1.7390E−2 5.2916E−3
1.004 1.121 1.298 1.716
2−6 1.3540E−1 6.3570E−2 2.7451E−2 1.0342E−2 2.8439E−3
1.091 1.211 1.408 1.863
2−8 1.0158E−1 4.4109E−2 1.7199E−2 5.6444E−3 1.2721E−3
1.203 1.359 1.607 2.150
2−10 9.0972E−2 3.8380E−2 1.4438E−2 4.5448E−3 9.8753E−4
1.245 1.410 1.668 2.202
−12
2 9.0413E−2 3.8150E−2 1.4353E−2 4.5182E−3 9.8149E−4
1.245 1.410 1.668 2.203
2−14 9.0221E−2 3.8077E−2 1.4328E−2 4.5104E−3 9.7973E−4
1.245 1.410 1.667 2.203
2−16 9.0146E−2 3.8051E−2 1.4319E−2 4.5079E−3 9.7918E−4
1.244 1.410 1.667 2.203
−18
2 9.0114E−2 3.8041E−2 1.4316E−2 4.5070E−3 9.7898E−4
1.244 1.410 1.667 2.203
2−20 9.0100E−2 3.8036E−2 1.4315E−2 4.5066E−3 9.7890E−4
1.244 1.410 1.667 2.203

In next two tables we show these numerical local errors and the values of q˜ corresponding to the two choices of the
boundary data. Table 8 displays the local errors when boundary conditions (5) are taken, and Table 9 displays the local
errors when (6) are considered. In both tables, again the last column is not very useful because the mesh is too close to the
finest mesh. Note that the local error is considerably smaller when the non natural option is used, and also that the orders
of consistency are bigger.

Acknowledgments

This research was partially supported by the project MTM2014-52859-P and the Diputación General de Aragón.
 C. Clavero, J.C. Jorge / Applied Mathematics and Computation 287–288 (2016) 12–27 27

Table 9
Local errors and values of q˜ for (29) with improved boundary conditions.

ε M=8 M = 16 M = 32 M = 64 M = 128

20 1.0795E−1 6.1073E−2 2.6859E−2 8.7963E−3 1.9041E−3

0.822 1.185 1.610 2.208
2−2 5.7374E−2 2.1965E−2 6.9096E−3 1.8860E−3 3.8478E−4
1.385 1.669 1.873 2.293
−4
2 3.9484E−2 1.2500E−2 3.7372E−3 9.8250E−4 1.9761E−4
1.659 1.742 1.927 2.314
2−6 3.9703E−2 1.0726E−2 3.0923E−3 8.2332E−4 1.7780E−4
1.888 1.794 1.909 2.211
2−8 4.0875E−2 1.0950E−2 3.0183E−3 8.8581E−4 2.3486E−4
1.900 1.859 1.769 1.915
2−10 4.1128E−2 1.10 0 0E−2 2.9083E−3 8.2207E−4 2.0449E−4
1.903 1.919 1.823 2.007
−12
2 4.1188E−2 1.1015E−2 2.8680E−3 8.1187E−4 2.0331E−4
1.903 1.941 1.821 1.998
2−14 4.1209E−2 1.1017E−2 2.8486E−3 8.0658E−4 2.0241E−4
1.903 1.951 1.820 1.995
2−16 4.1212E−2 1.1018E−2 2.8390E−3 8.0378E−4 2.0091E−4
1.903 1.956 1.821 2.0 0 0
−18
2 4.1210E−2 1.1018E−2 2.8343E−3 8.0251E−4 1.9988E−4
1.903 1.959 1.820 2.005
2−20 4.1209E−2 1.1017E−2 2.8319E−3 8.0187E−4 1.9953E−4
1.903 1.960 1.820 2.007

References

[1] I. Alonso-Mallo, B. Cano, J.C. Jorge, Spectral-fractional step Runge-Kutta discretizations for initial boundary value problems with time dependent bound-
ary conditions, Math. Comput. 73 (2004) 1801–1825.
[2] 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.
[3] C. Clavero, J.L. Gracia, J.C. Jorge, A uniformly convergent alternating direction HODIE finite difference scheme for 2D time dependent convection-diffu-
sion problems, IMA J. Numer. Anal. 26 (2006) 155–172.
[4] C. Clavero, J.L. Gracia, E. O’Riordan, A parameter robust numerical method for a two dimensional reaction-diffusion problem, Math. Comput. 74 (2005)
1743–1758.
[5] C. Clavero, J.C. Jorge, Another uniform convergence analysis technique of some numerical methods for parabolic singularly perturbed problems, Com-
put. Math. Appl. 70 (2015) 222–235.
[6] C. Clavero, J.C. Jorge, F. Lisbona, G.I. Shishkin, A fractional step method on a special mesh for the resolution of multidimensional evolutionary convec-
tion-diffusion problem, Appl. Numer. Math. 27 (1998) 211–231.
[7] 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 (20 0 0) 263–280.
[8] 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, 20 0 0.
[9] D. Herceg, Uniform fourth order difference scheme for a singular perturbation problem, Numer. Math. 56 (1990) 675–693.
[10] R.B. Kellogg, A. Tsan, Analysis of some difference approximations for a singularly perturbed problem without turning point, Math. Comput. 32 (1978)
1025–1039.
[11] T. Linss, M. Stynes, Asymptotic analysis and Shishkin-type decomposition for an elliptic convection-diffusion problem, J. Math. Anal. Appl. 261 (2001)
604–632.
[12] 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.
[13] J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems, revised ed., World Scientific, 2012.
[14] E. O’Riordan, M. Stynes, A globally convergent finite element method for a singularly perturbed elliptic problem in two dimensions, Math. Comput. 57
(1991) 47–62.
[15] A. Ostermann, M. Roche, Runge-Kutta methods for partial differential equations and fractional orders of convergence, Math. Comput. 59 (1992)
403–420.
[16] M.H. Protter, H.F. Weimberger, Maximum Principles in Differential Equations, Springer-Verlag, 1984.
[17] H.G. Roos, M. Stynes, L. Tobiska, Robust numerical methods for singularly perturbed differential equations, second ed., Springer-Verlag, 2008.
[18] J.G. Sanz-Serna, J. Verwer, W.H. Hundsdorfer, Convergence and order reduction of Runge-Kutta schemes applied to evolutionary problems in partial
differential equations, Numer. Math. 50 (1986) 405–418.
[19] G.I. Shishkin, Approximation of the solution to singularly perturbed boundary value problems with boundary layers, U.S.S.R. Comput. Maths. Math.
Phys. 29 (1989) 1–10.
[20] G.I. Shishkin, L.P. Shsihkina, Difference Methods for Singular Perturbation Problems, Chapman & Hall/CRC, 2009.
[21] N.N. Yanenko, The Method of Fractional Steps, Springer, 1971.