Numer Algor (2017) 75:809–826

DOI 10.1007/s11075-016-0221-9

ORIGINAL PAPER

A fractional step method for 2D parabolic

convection-diffusion singularly perturbed problems:
uniform convergence and order reduction

C. Clavero1 · J. C. Jorge2

Received: 1 April 2016 / Accepted: 3 October 2016 / Published online: 19 October 2016
© Springer Science+Business Media New York 2016

Abstract In this work, we are concerned with the efficient resolution of two dimen-
sional parabolic singularly perturbed problems of convection-diffusion type. The
numerical method combines the fractional implicit Euler method to discretize in time
on a uniform mesh and the classical upwind finite difference scheme, defined on a
Shishkin mesh, to discretize in space. We consider general time-dependent Dirich-
let boundary conditions, and we show that classical evaluations of the boundary
conditions cause an order reduction in the consistency of the time integrator. An
appropriate correction for the evaluations of the boundary data permits to remove
such order reduction. Using this correction, we prove that the fully discrete scheme
is uniformly convergent of first order in time and of almost first order in space. Some
numerical experiments, which corroborate in practice the robustness and the effi-
ciency of the proposed numerical algorithm, are shown; from them, we bring to light
the influence in practice of the two options for the boundary data considered here,
which is in agreement with the theoretical results.

Keywords Convection-diffusion · Singularly perturbed problems · Fractional euler

method · Piecewise uniform meshes · Uniform convergence · Order reduction

Mathematics Subject Classification (2010) 65N05 · 65N06 · 65N10

 C. Clavero
clavero@unizar.es
J. C. Jorge
jcjorge@upna.es

1 Department of Applied Mathematics and IUMA, University of Zaragoza, Zaragoza, Spain

2 Department of Computational and Mathematical Engineering, Public University of Navarra,

Pamplona, Spain
810 Numer Algor (2017) 75:809–826

1 Introduction

Let us consider the initial and boundary value 2D time-dependent convection-

diffusion problem
∂u  
Lu ≡ + L1,ε (t) + L2,ε (t) u = f, in Ω × (0, T ],
∂t
u(x, y, 0) = ϕ(x, y), in Ω, (1)
u(x, y, t) = g(x, y, t), in ∂Ω × [0, T ],

where Ω ≡ (0, 1) × (0, 1), and the spatial differential operators Li,ε , i = 1, 2 are
given by

∂2 ∂
L1,ε (t) ≡ −ε 2
+ v1 (x, y, t) + k1 (x, y, t), (2)
∂x ∂x
∂2 ∂
L2,ε (t) ≡ −ε 2 + v2 (x, y, t) + k2 (x, y, t),
∂y ∂y
respectively. We assume that the diffusion parameter ε, 0 < ε ≤ 1, can be very small,
that the convective coefficients are strictly positive, i.e., vi (x, y, t) ≥ v > 0, and
that the reaction terms satisfy ki (x, y, t) ≥ 0, i = 1, 2. Henceforth, we will suppose
that enough smoothness and compatibility conditions among data hold in order to the
solution is four times derivable in space and twice in time in the domain Ω × [0, T ].
In [6], the following sufficient conditions were given:

f ∈ C 2+2α,1+α (Ω × [0, T ]), ϕ ∈ C 4 (Ω)(α > 0),

(3)
{g(0, y, t), g(1, y, t), g(x, 0, t), g(x, 1, t)} ⊂ C 4,2 ([0, 1] × [0, T ]),

g(x, y, 0) = ϕ(x, y), (x, y) ∈ ∂Ω,

∂g  
(x, y, 0) = f (x, y, 0) − L1,ε (0) + L2,ε (0) ϕ(x, y), (x, y) ∈ ∂Ω,
∂t (4)
∂ 2g ∂f  
2
(x, y, 0) = (x, y, 0) − L1,ε (0) + L2,ε (0) f (x, y, 0)
∂t ∂t 
2
+ L1,ε (0) + L2,ε (0) ϕ(x, y), (x, y) ∈ ∂Ω,
and
∂g  
(x, y, t) + L1,ε (t) + L2,ε (t)g(x, y, t) = f (x, y, t),
∂t
(x, y) ∈ {(0, 0), (0, 1), (1, 0), (1, 1)}, t ∈ [0, T ].
(5)

The numerical resolution of singularly perturbed problems is an interesting subject

in applied mathematics which has been received many attention in the last years.
To dispose of reliable numerical methods, fitted operator methods or fitted mesh
methods are used (see [9, 11, 14, 16, 17] and references therein). Specially interesting
is the case when the problem is two dimensional in space of convection-diffusion
Numer Algor (2017) 75:809–826 811

type. The stationary case is considered, for instance, in [4, 10, 12], and the time-
dependent problem is analyzed in [2, 3, 5]. In some of these works, it was described
a two-step technique to construct the fully discrete scheme.
In [6], the authors consider and analyze the uniform convergence of a first stage
of spatial semidiscretization, via an upwind finite difference method on a Shishkin
mesh for a parabolic problem of convection-diffusion type; afterwards, the Euler
implicit method completes the discretization. Using this method, the computational
cost is large because pentadiagonal linear systems, one per time step, must be solved.
This drawback is typical when classical implicit time integrators are used for solving,
combined with finite differences, 2D parabolic problems. In [7], the reverse order
was chosen, discretizing firstly in time and later on in space, for a parabolic problem
of reaction-diffusion type; then, the use of an alternating direction method (see [18]),
permits that only tridiagonal linear systems must be solved to obtain the numerical
solution, which implies a remarkable cost reduction with respect to other, more clas-
sical, integration techniques. Here, we are going to extend the technique developed in
[7] to the case of parabolic 2D convection-diffusion, but following the reverse order
in the two-step technique considered in [5]. The reasons for choosing this option are
that, firstly, the analysis of the uniform convergence of the fully discrete is simpler
and, besides, that, in contrast with previous works (see for instance [3, 5]), no ratios
between the discretization parameters are needed to prove the uniform convergence
of the scheme.
Another interesting question, which is usually posed in the convergence analy-
sis of numerical methods for parabolic problems, is related to the influence of the
time-dependent boundary data of (1). It is well known (see, for instance [1, 15] and
references therein) that, when using one step methods, a classical evaluation of the
boundary conditions causes, in general, a reduction in the numerical order of conver-
gence; this phenomenon is specially severe for most methods of alternating direction
type. So, in this paper, we consider a different and simpler modification of these
evaluations for the fractional implicit Euler method, which eliminates the loss of
consistency without increasing the computational cost of the algorithm.
The rest of the paper is structured as follows. In Section 2, we introduce the spatial
discretization of the continuous problem on a piecewise uniform mesh of Shishkin
type, and we prove that it is uniformly convergent with respect to the diffusion param-
eter of almost first order. In Section 3, we introduce the time discretization proposed,
by using the fractional implicit Euler method and proving that this time discretiza-
tion is a first order uniformly convergent method. Also, we analyze the effect of
two choices of evaluations of the data functions associated to the non-homogenous
boundary conditions, in the uniform consistency of the method. Finally, in Section 4,
some numerical results obtained for different test problems are shown; from them, we
can observe the uniform convergent behavior of the numerical algorithm proposed
here; as well, we show the effects of the classical evaluations of the boundary data
and the improvement of them which we propose here.
Henceforth, C denotes a generic positive constant independent of the diffusion
parameter ε and also of the discretization parameters N and M. Below, we always
use the pointwise maximum norm, denoted by  · D (where D is the corresponding
domain).
812 Numer Algor (2017) 75:809–826

2 Spatial semidiscretization

In this section, we define approximations of u(xi , yj , t), where (xi , yj ) are going to
be the grid points of a rectangular mesh Ω N ≡ Ix,ε,N × Iy,ε,N ⊂ Ω, which has
(N + 1)2 nodes, for simplicity, we take the same number N of subintervals in both
space directions, and t ∈ [0, T ] remains as a continuous variable. We will denote ΩN
the subgrid of Ω N obtained by removing the points which belong to ∂Ω.
Let us denote [.]N the restriction of a function defined on Ω to ΩN , uN (t), the
semidiscrete grid functions which will approach [u(x, y, t]N , [.]N , the restriction of
a function defined on Ω to Ω N , and uN (t) the semidiscrete grid functions which will
approach [u(x, y, t)]N .
Typically, uN (t) is defined as the solution of a stiff Initial Value Problem which
can be written in the form
 
uN (t) + L1,ε,N (t) + L2,ε,N (t) uN (t) = [f ]N ,
uN (t) = [g]N in Ω N \ΩN , (6)
uN (0) = [ϕ]N ,
where Li,ε,N , i = 1, 2 must be appropriate spatial discretizations of the elliptic
convection-diffusion operators Li,ε , i = 1, 2 given in (2).
Here, we consider the special mesh Ω N ≡ Ix,ε,N × Iy,ε,N , tensor product of
one-dimensional meshes,
Ix,ε,N = {0 = x0 < . . . < xN = 1}, Iy,ε,N = {0 = y0 < . . . < yN = 1},
of Shishkin type as follows. We only give the details of the construction of Ix,ε,N and
analogously we proceed for Iy,ε,N .
Let us choose N as an even number. We define the transition parameter
σx = min(1/2, mx ε ln N), (7)
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

2i(1 − σx )/N , i = 0, . . . , N/2,
xi = (8)
1 − σx + 2(i − N/2)σx /N , i = N/2 + 1, . . . , N.
Using these meshes, Li,ε,N , i = 1, 2 are the discretization of the differential
convection-diffusion operator Li,ε , i = 1, 2, via the simple upwind finite difference
scheme, which is given by
L1,ε,N (t)uN (t)(xi , yj ) ≡ li−,j uN (t)(xi−1 , yj ) + li+,j uN (t)(xi+1 , yj )
(9)
+li,j
1 u (t)(x , y ), i = 1, . . . , N − 1, j = 0, . . . , N,
N i j

where
−ε v1 (xi , yj , t) −ε
li−,j = − , li+,j = , (10)
hx,i h̃x,i hx,i hx,i+1 h̃x,i
1 = −l
li,j i−,j − li+,j + k1 (xi , yj , t),
Numer Algor (2017) 75:809–826 813

and analogously
L2,ε,N (t)uN (t)(xi , yj ) ≡ li,j − uN (t)(xi , yj −1 ) + li,j + uN (t)(xi , yj +1 )
(11)
+li,j
2 u (t)(x , y ), j = 1, . . . , N − 1, i = 0, . . . , N,
N i j

where
−ε v2 (xi , yj , t) −ε
li,j − = − , li,j + = , (12)
hy,j h̃y,j hy,j hy,j +1 h̃y,j
2 = −l
li,j i,j − − li,j + + k2 (xi , yj , t),

with hx,i = xi − xi−1 , i = 1, . . . , N, hy,j = yj − yj −1 , j = 1, . . . , N, h̃x,i =

(hx,i + hx,i+1 )/2, i = 1, . . . , N − 1, h̃y,j = (hy,j + hy,j +1 )/2, j = 1, . . . , N − 1.
Using this semidiscretization, in [6], the authors prove that the upwind scheme
is an almost first-order uniformly convergent scheme; summarizing, under the
smoothness assumptions made here, it holds that
[u(x, y, t)]N − uN (t)Ω N ≤ CN −1 ln N, ∀ t ∈ (0, T ]. (13)

3 Time integration: unconditional convergence

In this section, we introduce the numerical algorithm which we propose to integrate

successfully the continuous problem (1). After having considered the spatial semidis-
cretization stage in the previous section, we complete the discretization process by
using the fractional implicit Euler method as time integrator.
(N+1)2
Using this 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.
As a previous task, we need to establish the uniform behavior of the time
derivatives and the fractional differentials of uN (t) involved in the analysis of the
uniform consistency for the time integration process. In [6], the authors explain
briefly that, under the smoothness and compatibility assumptions made here for
2
the data of (1), it holds that uN (t) ∈ (C 2 [0, T ])(N+1) , and its second deriva-
tive is bounded independently of ε. From this property, joint to the smoothness
assumptions made on f , it is immediate to deduce that the elementary differentials
Lε,N uN (t), L2ε,N uN (t) are bounded independently of ε. These differentials are typ-
ically involved in the consistency analysis of classical one step time integrators, for
instance, Runge-Kutta methods. In this paper, we will assume a similar but extra nat-
ural condition in this context which is that the fractionated elementary differentials
Li,ε,N uN (t), Li,ε,N Lj,ε,N uN (t), i, j = 1, 2, satisfy
Li,ε,N uN (t)ΩN ≤ C, Li,ε,N Lj,ε,N uN (t)ΩN ≤ C, i, j = 1, 2. (14)
This property is easily deduced in the cases that operators Li,ε,N , i = 1, 2 com-
mute. In such cases, these fractional elementary differentials can be described as
solutions of Initial Value Problems similar to (15), where all of the data functions
814 Numer Algor (2017) 75:809–826

are ε-uniformly bounded. For example, ω ≡ L1,ε,N uN (t) can be described as the
solution of
ω (t) + (L1,ε,N (t) + L2,ε,N (t))ω(t) = L1,ε,N [f ]N ,
ω(t) = g̃N in Ω N \ΩN , (15)
ω(0) = L1,ε,N [ϕ]N ,

being g̃N = L1,ε,N [g]N for y ∈ {0, 1} and g̃N = [f ]N − [ ∂g

∂t ]N − L2,ε,N [g]N for
x ∈ {0, 1}.

3.1 The fully discrete scheme

In this section, we describe the numerical method which we propose to discretize

in time the Initial Value Problems (15). We consider the fractional implicit Euler
method (see [5, 18]), 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 mesh I M = {tm = mτ, m =
N ≈ uN (x, y, tm ), m = 0, 1, . . . , M, be the numerical solutions
0, 1, . . . , M}. Let um
defined as follows:
i) (initialize)
u0N = [ϕ(x, y)]N , in ΩN .
u0N = [g(x, y, 0)]N , in ΩN \ΩN .
ii) (first half step)
m+1/2 m+1
(I + τ L1,ε,N (tm+1 ))uN = um
N
+ τf1,N , in ΩN \{0, 1} × [0, 1],
m+1/2 m+1/2 (16)
u =g , in ΩN ∩ {0, 1} × [0, 1].
N N
iii) (second half step)
m+1/2
(I + τ L2,ε,N (tm+1 ))um+1 = u + τf m+1 , in ΩN \[0, 1] × {0, 1}, ,
N N 2,N
um+1 (x, y) = g m+1 , in ΩN ∩ [0, 1] × {0, 1},
N N
m = 0, . . . , M − 1,
being
f = f1 + f2 , f m+1 = [f1 (x, y, tm+1 )]N , f m+1 = [f2 (x, y, tm+1 )]N . (17)
1,N 2,N

Of course, as a typical advantage of Alternating Direction methods, the resolution

of the half steps involves a set of N + 1 uncoupled linear tridiagonal systems of
dimension N − 1 which can be solved in parallel. Moreover, the advance of one step
in time can be done with a number of arithmetic operations of size O(N 2 ) which is
optimal in terms of the computational complexity of the method.
In the literature (see [1, 13, 15]), the most classical option for the evaluations of
the boundary data is given by
m+1/2
g = [g(x, y, tm+1 )]N , in ΩN ∩ {0, 1} × [0, 1],
N (18)
g m+1 = [g(x, y, tm+1 )]N , in ΩN ∩ [0, 1] × {0, 1}.
N
Numer Algor (2017) 75:809–826 815

This option is natural in the sense that, classically, the contributions of the boundary
data are added to the source term [f (x, y, t)]N before the time integration process.
Nevertheless, in most of cases, this choice reduces the order of unconditional (inde-
pendent of N) 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 data, given by
m+1/2
g = (I + τ L2,ε,N (tm+1 )[g(x, y, tm+1 )]N − τf m+1 , in ΩN ∩ {0, 1} × [0, 1],
N 2,N
m+1
gN = [g(x, y, tm+1 )]N , in ΩN ∩ [0, 1] × {0, 1}.
(19)

3.2 Stability

Let us define the operators (I + τ L1,ε,N,0 (t))−1 (analogously (I + τ L2,ε,N,0 (t))−1 ),

as follows. Let vN = (I + τ Ln1,ε,N,0 (t))−1 uN be the solution of the linear system

(I + τ Ln1,ε,N (t))vN = uN , (x, y) in ΩN ,

(20)
vN = 0, in ΩN \ΩN .

From classical properties of M-matrices (see, e.g., [14]), it is straightforward to

deduce

(I + τ Li,ε,N,0 (t))−1 ∞ ≤ 1, i = 1, 2. (21)

3.3 Uniform and unconditional consistency

m+1
Let us define eN the local error in time of the scheme (16) at tm+1 as follows:

m+1
eN = uN (tm+1 ) − ũm+1
N , (22)

where ũm+1
N is the solution of the auxiliary problem

m+1/2 m+1
(I + τ L1,ε,N (tm+1 ))ũN = [u(tm )]m
N + τf1,N , in ΩN \{0, 1} × [0, 1],
m+1/2 m+1/2
ũ =g , in ΩN ∩ {0, 1} × [0, 1],
N N
(I + τ L2,ε,N (tm+1 ))ũm+1 = ũN
m+1/2 m+1
+ τf2,N , in ΩN \[0, 1] × {0, 1}, (23)
N
ũm+1
N (x, y) = g
m+1
, in ΩN ∩ [0, 1] × {0, 1},
N
m = 0, . . . , M − 1.

Lemma 1 Under the assumptions (3), (4), (5) and (14), if we choose the boundary
data given in (19), then the local error defined by (22), (23) satisfies

eN
m
ΩN ≤ Cτ 2 , m = 1, . . . , M, (24)

and therefore (16) is a first order uniformly consistent method.

816 Numer Algor (2017) 75:809–826

Proof From the definition of ũm+1

N given in (23), it is easily deduced that in ΩN it
holds
 
(I + τ L1,ε,N (tm+1 ) I + τ L2,ε,N (tm+1 ))ũm+1
N − τf m+1
2,N
m+1
= [u(tm )]N + τf1,N .
(25)
On the other hand, by using the simple Taylor expansion
uN (tm ) = uN (tm+1 ) − τ uN (tm+1 ) + O(τ 2 ),
and taking into account that uN (tm+1 ) = f1,N
m+1 m+1
+ f2,N − L1,ε,N (tm+1 )uN (tm+1 ) −
L2,ε,n (tm+1 )uN (tm+1 ), it can be deduced that
 
(I + τ L1,ε,N (tm+1 )) (I + τ L2,ε,N (tm+1 ))uN (tm+1 ) − τf2,N m+1
(26)
m+1
= uN (tm ) + τf1,N + O(τ 2 ).
Subtracting (26) and (25), it holds
m+1
(I + τ L1,ε,N (tm+1 ))(I + τ L2,ε,N (tm+1 ))eN = O(τ 2 ). (27)
Regarding to the boundary data, it is trivial that uN (tm+1 ) = g m+1 in ΩN ∩[0, 1]×
N
{0, 1}, and it is also clear that, as we have chosen
m+1/2
g = (I + τ L2,ε,N (tm+1 ))uN (tm+1 ) − τf m+1 inΩN ∩ {0, 1} × [0, 1],
N 2,N
the local error can be written as the solution of a problem of the form
m+1/2
(I + τ L1,ε,N (tm+1 ))eN = O(τ 2 ), in ΩN ,
m+1/2
e = 0, in ΩN \ΩN ,
N  (28)
m+1/2
I + τ Ln+1
2,ε,N (tm+1 ) e
m+1
=e , in ΩN ,
N N
m+1
eN (x, y) = 0, in ΩN \ΩN .
From (28) and the stability property (21), the required result (24) follows.

Remark 1 Note that for non-homogeneous boundary data g(x, y, t), in general
L2,ε,N (tm+1 )[g(x, y, tm+1 )]N − f m+1 = 0. Therefore, a term of size O(τ ) appears
2,N
in the difference between the natural boundary data given by [g(x, y, tm+1 )]N , and
m+1/2
those ones considered here (g ) for the first half step. This fact causes an order
N
reduction in the order of consistency, up to order 0 if the classical evaluations of the
boundary data are chosen.

Remark 2 In [3], the authors propose and analyze a similar algorithm for a less gen-
eral problem; therein, the boundary conditions are homogeneous. In that paper, to
avoid the order reduction, a suitable splitting f = f1 + 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 (more general) case which we have stud-
ied here. In fact, with the boundary data proposed in (19) any smooth splitting for f
can be considered without risk of order reduction.
Numer Algor (2017) 75:809–826 817

3.4 Uniform convergence of the time integration

m ≡
Let us introduce the global error of the scheme (16) at time tm as usual, i.e., EN
uN (tm ) − uN .
m

Theorem 1 Under the assumptions (3), (4), (5) and (14), if we choose the boundary
data given in (19), then it holds
EN
m
ΩN ≤ Cτ, ∀m = 1, . . . , M, (29)
i. e., the time integration process (16) is uniformly and unconditionally convergent of
first order.

Proof Subtracting and adding ũm

N , trivially we have
−1 −1 n−1
m
EN = eN
m
+ ũm
N − uN = eN + (I + τ L2,ε,N,0 (tm )) (I + τ L1,ε,N,0 (tm )) E
m m n
.
Using this recurrence, combined with (21) and (24), the result trivially follows.

3.5 Uniform convergence of the algorithm

Theorem 2 Assuming (3), (4), (5), and (14), if we use the improved boundary data
(19), then the global error given by
EN,M ≡ max [u(x, y, tm )]N − um
N ΩN ,
1≤m≤M
satisfies
EN,M ≤ C(N −1 ln N + M −1 ). (30)
Thus, the proposed method is unconditionally and uniformly convergent of first order
in time and almost first order in space, up to a typical logarithmic factor.

Proof We only need to add and subtract the term uN (tm ) and use the triangle
inequality to deduce
EN,M ≤ max [u(x, y, tm )]N − uN (tm )ΩN + max EN
m
ΩN .
1≤m≤M 1≤m≤M
Finally, from (13) and (29), the result follows.

4 Numerical experiments

In this section, we solve some test problems using our numerical algorithm. The first
example is given by
ut − εΔu + (2x + 1)ux + (2y + 1)uy + 30u = f (x, y, t), (x, y, t) ∈ Ω × [0, 1],
u(x, y, t) = g(x, y, t), in ∂Ω × [0, 1]
u(x, y, 0) = ϕ(x, y), x, y ∈ [0, 1],
(31)
 818 Numer Algor (2017) 75:809–826

where f (x, y, t), g(x, y, t) and ϕ(x, y) are chosen in such way that the exact solution
is u(x, y, t) = (1 − e−30t ) (Ψ (x)Ψ (y) − xy), with
2 z2 +z−2
e− ε − e ε
Ψ (z) ≡ z + 2
.
1 − e− ε
Figure 1 shows the solution at the final time t = 1; from it, we clearly see the
boundary layers at x = 1 and y = 1.
In all tables corresponding to example (31), we take mx = my = 1 to
define the transition parameters of the meshes I1,ε,N and I2,ε,N , respectively.
Moreover, we decompose (see [3]) the right-hand side in the form f (x, y, t) =
f1 (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))
and f1 (x, y, t) = f (x, y, t) − f2 (x, y, t).
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 , yj , tn )|,

0≤n≤M 0≤i≤N 0≤j ≤N

and therefore the numerical orders of convergence are calculated by

 
p = log eN,M /e2N,2M /log 2.

From these values, we calculate the uniform maximum errors by

emax N,M = max eN,M ,

ε
numerical solution at t=1

0
−0.1
−0.2
−0.3
−0.4
−0.5
−0.6
−0.7
0
−0.8 0.1
−0.9 0.2
0.3
−1 0.4
0 0.5
0.1
0.2 0.6
0.3
0.4 0.7
0.5
0.6 0.8
0.7
0.8 0.9
0.9
1 1
x axis
y axis

Fig. 1 Numerical solution of example (31) for ε = 10−2 , N = M = 32, at the final time t = 1
Numer Algor (2017) 75:809–826 819

and from them, in a usual way, the corresponding numerical uniform orders of
convergence are given by
 
puni = log emax N,M /emax 2N,2M /log 2.

Table 1 displays the results for problem (31) by using our fully discrete scheme
when the classical boundary conditions are used. From them, we observe a reduction
in the numerical orders of convergence.
Next, we consider the choice the boundary data given in (19). Table 2 displays the
results for problem (31) by using this option. From it, an almost first order uniformly
convergent numerical behavior is observed. Moreover, note that the maximum errors
for any value of ε are smaller than those ones given in Table 1. A similar behavior
has been observed in all of the experiments which we have performed.
Nevertheless, in the case of using natural boundary data, the global errors behave
better than expected; this phenomenon has been deeply studied in the numerical inte-
gration of canonical parabolic problems like the heat equation, where several authors
have proven that one order of convergence can be recovered many times, by using
a summation by parts reasoning. This technique, which has been successfully used
for studying the integration of simple parabolic problems (see, e.g., [1, 13, 15]),
has not been used in the context of parabolic singularly perturbed problems and

Table 1 Maximum errors and orders of convergence for (31) 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

2−6 8.9908E-1 6.3453E-1 4.0693E-1 2.5911E-1 1.6161E-1

0.503 0.641 0.651 0.681
2−8 9.3261E-1 6.5698E-1 4.2064E-1 2.6994E-1 1.6796E-1
0.505 0.643 0.640 0.685
2−10 9.4281E-1 6.6341E-1 4.2459E-1 2.7338E-1 1.7008E-1
0.507 0.644 0.635 0.685
2−12 9.4556E-1 6.6516E-1 4.2567E-1 2.7438E-1 1.7070E-1
0.507 0.644 0.634 0.685
2−14 9.4625E-1 6.6561E-1 4.2595E-1 2.7465E-1 1.7088E-1
0.508 0.644 0.633 0.685
2−16 9.4642E-1 6.6573E-1 4.2602E-1 2.7473E-1 1.7092E-1
0.508 0.644 0.633 0.685
– – – – – –
– – – –
2−26 9.4647E-1 6.6577E-1 4.2604E-1 2.7476E-1 1.7094E-1
0.508 0.644 0.633 0.685
emax N,M 9.4647E-1 6.6577E-1 4.2604E-1 2.7476E-1 1.7094E-1
p uni 0.508 0.644 0.633 0.685
820 Numer Algor (2017) 75:809–826

Table 2 Maximum errors and orders of convergence for (31) with improved boundary conditions

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

M=8 M=16 M=32 M=64 M=128

2−6 8.4881E-1 5.8678E-1 3.6410E-1 2.0838E-1 1.1328E-1
0.533 0.688 0.805 0.879
2−8 8.8872E-1 6.0980E-1 3.7657E-1 2.1503E-1 1.1683E-1
0.543 0.695 0.808 0.880
2−10 9.0122E-1 6.1653E-1 3.7993E-1 2.1676E-1 1.1776E-1
0.548 0.698 0.810 0.880
2−12 9.0470E-1 6.1830E-1 3.8080E-1 2.1720E-1 1.1799E-1
0.549 0.699 0.810 0.880
2−14 9.0562E-1 6.1874E-1 3.8102E-1 2.1731E-1 1.1805E-1
0.550 0.699 0.810 0.880
2−16 9.0585E-1 6.1885E-1 3.8107E-1 2.1734E-1 1.1806E-1
0.550 0.700 0.810 0.880
– – – – – –
– – – –
2−26 9.0592E-1 6.1889E-1 3.8109E-1 2.1735E-1 1.1807E-1
0.550 0.700 0.810 0.880
emax N,M 9.0592E-1 6.1889E-1 3.8109E-1 2.1735E-1 1.1807E-1
p uni 0.550 0.700 0.810 0.880

seems complicated to apply it. Thus, we can conclude that the new proposal provides
improvements both from theoretical and practical points of view.
In order to clarify a bit more the influence, in the numerical behavior of the
method, of the two options for the boundary data considered here as well as the
improvements provided by the non natural evaluations of the boundary conditions,
we estimate the local errors in time. As the exact solution is known, these estimates
can be approximated by
ẽN,M = max max max |ŨNm − u(xi , yj , tm )|.
0≤m≤M 0≤i≤N 0≤j ≤N

where N must be chosen large enough in order to the contribution of the spatial
discretization can be neglected. From them, we obtain the quantities
 
p̃ = log ẽN,M /ẽN,2M /log 2,
and note that the corresponding numerical orders of consistency are given by p̃ − 1.
In next tables, we show such estimated local errors and the values of p̃ corre-
sponding to the two choices of the boundary data, taking N = 512 fixed. Table 3
displays the estimated local errors when natural boundary conditions are chosen.
Here, uniform consistency of order zero is observed.
Table 4 displays the local errors obtained when the improved boundary conditions
are considered. Note that the local errors are substantially smaller than in the previous
Numer Algor (2017) 75:809–826 821

Table 3 Local errors and values of p̃ for (31) with natural boundary conditions, N = 512

ε M=8 M = 16 M = 32 M = 64 M = 128

2−6 7.7818E-1 5.5586E-1 4.2260E-1 2.8458E-1 1.7001E-1

0.485 0.395 0.570 0.743
2−8 8.0824E-1 5.7012E-1 4.3375E-1 2.9248E-1 1.7518E-1
0.504 0.394 0.569 0.739
2−10 8.1763E-1 5.7464E-1 4.3731E-1 2.9506E-1 1.7694E-1
0.509 0.394 0.568 0.738
2−12 8.2041E-1 5.7599E-1 4.3837E-1 2.9583E-1 1.7746E-1
0.510 0.394 0.567 0.737
2−14 8.2120E-1 5.7637E-1 4.3868E-1 2.9605E-1 1.7761E-1
0.511 0.394 0.567 0.737
2−16 8.2142E-1 5.7649E-1 4.3876E-1 2.9611E-1 1.7765E-1
0.511 0.394 0.567 0.737
– – – – – –
– – – –
2−26 8.2150E-1 5.7653E-1 4.3880E-1 2.9613E-1 1.7767E-1
0.511 0.394 0.567 0.737

table, where the classical boundary data are used and also that the orders of consis-
tency are higher in this case. In both Tables 3 and 4, the numbers do not show clearly
the zero and first orders of consistency, respectively, specially in the last columns;

Table 4 Local errors and values of p̃ for (31) with improved boundary conditions, N = 512

ε M=8 M = 16 M = 32 M = 64 M = 128

2−6 7.6949E-1 4.5196E-1 2.1054E-1 8.0750E-2 3.1093E-2

0.768 1.102 1.383 1.377
2−8 8.0530E-1 4.7209E-1 2.1897E-1 8.3539E-2 3.4114E-2
0.770 1.108 1.390 1.292
2−10 8.1665E-1 4.7848E-1 2.2157E-1 8.4303E-2 3.5206E-2
0.771 1.111 1.394 1.260
2−12 8.2006E-1 4.8039E-1 2.2232E-1 8.4499E-2 3.5513E-2
0.772 1.112 1.396 1.251
2−14 8.2105E-1 4.8094E-1 2.2253E-1 8.4548E-2 3.5592E-2
0.772 1.112 1.396 1.248
2−16 8.2133E-1 4.8110E-1 2.2259E-1 8.4561E-2 3.5612E-2
0.772 1.112 1.396 1.248
– – – – – –
– – – –
2−26 8.2144E-1 4.8116E-1 2.2261E-1 8.4565E-2 3.5619E-2
0.772 1.112 1.396 1.247
822 Numer Algor (2017) 75:809–826

Table 5 Local errors and values of p̃ for (31) with natural boundary conditions, ε = 2−12

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

1024 8.2025E-1 5.7784E-1 4.4430E-1 3.0196E-1 1.8190E-1

0.505 0.379 0.557 0.731
2048 8.2021E-1 5.7923E-1 4.4803E-1 3.0573E-1 1.8460E-1
0.502 0.371 0.551 0.728

these ones correspond to the largest values of M, where the fixed value of N is not
sufficiently large, and therefore, at these columns, in the estimated errors there is a
substantial influence of the errors associated to the spatial discretization. To reduce
this effect, in Tables 5 and 6, we show the numerical results obtained for a partic-
ular value of ε when the discretization parameter N is increased to N = 1024 and
N = 2048; similar results are obtained for other values of ε. From them, we observe
the corresponding numerical orders of consistency, closer to one in Table 6 as the
theory predicts.
The second example that we consider is given by

  
ut − εΔu + ux + uy + 5 + 2t 2 exp −1/((x − x 2 )(y − y 2 )) u
= 2e−5t (Ψ (x)Ψ (y) − t 2 ), (x, y, t) ∈ Ω × [0, 1],
u(x, y, t) = e−5t (x + y − 2t), in∂Ω × [0, 1]
u(x, y, 0) = x + y, x, y ∈ [0, 1],
(32)

where
e−1/ε (1 − ez/ε )
Ψ (z) ≡ z + .
1 − e−1/ε

In this case, the exact solution is unknown.

We take again mx = my = 1 to define the piecewise uniform Shishkin mesh, and
we decompose the source term by taking f1 (x, y, t) = f2 (x, y, t) = f (x, y, t)/2.

Table 6 Local errors and values of p̃ for (31) with improved boundary conditions, ε = 2−12

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

1024 8.2006E-1 4.8037E-1 2.2226E-1 8.2802E-2 2.9222E-2

0.772 1.112 1.424 1.503
2048 8.2006E-1 4.8036E-1 2.2224E-1 8.2616E-2 2.6699E-2
0.772 1.112 1.428 1.630
Numer Algor (2017) 75:809–826 823

To approximate the maximum pointwise errors, we use a variant of the two-

mesh principle (see [8, 9]). We calculate {ûN }, the numerical solution on the mesh
{(x̂i , ŷj , tˆn )} 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,
tˆ2m = tm , m = 0, . . . , M, tˆ2m+1 = (tm + tm+1 )/2, m = 0, . . . , M − 1.
Then, the maximum errors at the mesh points of the coarse mesh are approximated by
di,j,N,M = max max |uN (xi , yj , tm ) − ûN (xi , yj , tm )|, (33)
0≤m≤M 0≤i,j ≤N

and the orders of convergence are given by

 
q = log di,j,N,M /di,j,2N,2M /log 2.
From the double-mesh differences in (33), we obtain the uniform maximum errors by
d N,M = max di,j,N,M ,
ε
and from them, in a usual way, the corresponding numerical uniform orders of
convergence by  
q uni = log d N,M /d 2N,2M /log 2.
Table 7 displays the results for problem (32) by using our method when classical
boundary conditions are chosen. From them, we again observe a reduction in the
order of uniform convergence.
Next, we evaluate the boundary conditions using (19) to solve the same problem.
Table 8 displays the numerical results corresponding to this choice; from them, we
observe smaller maximum errors, in comparison to Table 7, and also we deduce a
first order uniformly convergent behavior.

Table 7 Maximum errors and orders of convergence for (32) 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

2−6 5.9524E-2 6.9820E-2 5.3599E-2 3.3555E-2 1.9020E-2

-.230 0.381 0.676 0.819
2−8 6.5906E-2 7.5892E-2 5.7369E-2 3.5686E-2 2.0030E-2
-.204 0.404 0.685 0.833
2−10 6.7763E-2 7.7925E-2 5.8878E-2 3.6511E-2 2.0450E-2
-.202 0.404 0.689 0.836
– – – – – –
– – – –
2−26 6.8412E-2 7.8680E-2 5.9490E-2 3.6912E-2 2.0677E-2
-.202 0.403 0.689 0.836
d N,M 6.8412E-2 7.8680E-2 5.9490E-2 3.6912E-2 2.0677E-2
q uni -.202 0.403 0.689 0.836
824 Numer Algor (2017) 75:809–826

Table 8 Maximum errors and orders of convergence for (32) with improved boundary conditions

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

M=8 M = 16 M = 32 M = 64 M = 128
2−6 3.4674E-2 1.9782E-2 1.0357E-2 5.2911E-3 2.6748E-3
0.810 0.934 0.969 0.984
2−8 3.6338E-2 2.0256E-2 1.0613E-2 5.4266E-3 2.7390E-3
0.843 0.932 0.968 0.986
2−10 3.6910E-2 2.0575E-2 1.0703E-2 5.4676E-3 2.7580E-3
0.843 0.943 0.969 0.987
– – – – – –
– – – –
2−26 3.7122E-2 2.0764E-2 1.0811E-2 5.4988E-3 2.7696E-3
0.838 0.942 0.975 0.989
d N,M 3.7122E-2 2.0764E-2 1.0811E-2 5.4988E-3 2.7696E-3
q uni 0.838 0.942 0.975 0.989

Now, we estimate the local errors for this second example. As the exact solution
is unknown, to approximate the errors, we proceed in a different way than before.
To approximate ŨNn , we use one step of the fully discrete scheme given in (16), but
replacing the numerical approximation at time tn−1 , UNn−1 , by the numerical solution
obtained on a very fine mesh of Shishkin type, with N = 1024, M = 512, which is
denoted by u1024,512 . If ū1024,512 is the piecewise bilinear interpolation of u1024,512 .
We take again a large fixed value for spatial discretization parameter, concretely N =
256, in order to the errors in time dominate, then the local errors are approximated by
d̃N,M = max max max |ŨNn − ū1024,512 (xi , yj , tn )|,
0≤n≤M 0≤i≤N 0≤j ≤N

and from them we calculate

 
q̃ = log d̃N,M /d̃N,2M /log 2,

and the orders of consistency are given by q̃ − 1.

Table 9 Local errors and values of q̃ for (32) with natural boundary conditions, N = 256

ε M=8 M = 16 M = 32 M = 64 M = 128

2−6 2.3118E-1 1.9211E-1 1.2116E-1 6.4267E-2 2.9005E-2

0.267 0.665 0.915 1.148
2−8 2.3903E-1 1.9833E-1 1.2524E-1 6.6707E-2 3.0369E-2
0.269 0.663 0.909 1.135
2−10 2.4156E-1 2.0036E-1 1.2659E-1 6.7566E-2 3.0924E-2
0.270 0.662 0.906 1.128
Numer Algor (2017) 75:809–826 825

Table 10 Local errors and values of q̃ for (32) with improved boundary conditions, N = 256

ε M=8 M = 16 M = 32 M = 64 M = 128

2−6 7.0714E-2 2.5038E-2 7.5868E-3 2.0084E-3 4.6152E-4

1.498 1.723 1.917 2.122
2−8 7.1447E-2 2.5566E-2 7.7610E-3 2.0499E-3 4.6930E-4
1.483 1.720 1.921 2.127
2−10 7.1701E-2 2.5793E-2 7.8290E-3 2.0662E-3 7.2746E-4
1.475 1.720 1.922 1.506

Table 9 displays the local errors when boundary conditions (18) are taken, and
Table 10 displays the local errors when (19) are considered. In order to not enlarge
the paper in excess, we only show the results for three values of ε which are suf-
ficient to understand the differences between the two evaluations of the boundary
conditions. In these tables, again the local errors for improved boundary conditions
are smaller than for classical ones; also, the orders of consistency tend to zero for
natural boundary conditions and to one for improved boundary conditions.

Acknowledgements The authors thank the referees their valuable suggestions which have helped to
improve the presentation of this paper. This research was partially supported by the project MTM2014-
52859 and the Diputación General de Aragón.

References

1. Alonso-Mallo, I., Cano, B., Jorge, J.C.: Spectral-fractional step Runge-Kutta discretizations for initial
boundary value problems with time dependent boundary conditions. Math. Comp. 73, 1801–1825
(2004)
2. Bujanda, B., Clavero, C., Gracia, J.L., Jorge, J.C.: A high order uniformly convergent alternating
direction scheme for time dependent reaction-diffusion singularly perturbed problems. Num. Math.
107, 1–25 (2007)
3. Clavero, C., Gracia, J.L., Jorge, J.C.: A uniformly convergent alternating direction HODIE finite dif-
ference scheme for 2D time dependent convection-diffusion problems. IMA J. Numer. Anal. 26, 155–
172 (2006)
4. Clavero, C., Gracia, J.L., O’Riordan, E.: A parameter robust numerical method for a two dimensional
reaction-diffusion problem. Math. Comp. 74, 1743–1758 (2005)
5. Clavero, C., Jorge, J.C., Lisbona, F., Shishkin, G.I.: A fractional step method on a special mesh for
the resolution of multidimensional evolutionary convection-diffusion problems. Appl. Num. Math.
27, 211–231 (1998)
6. Clavero, C., Jorge, J.C.: Another uniform convergence analysis technique of some numerical methods
for parabolic singularly perturbed problem. Comp. Math. Appl. 70, 222–235 (2015)
7. Clavero, C., Jorge, J.C.: Uniform convergence and order reduction of the fractional implicit Euler
method to solve singularly perturbed 2D reaction-diffusion problems. Appl. Math. Comp. 287–88,
12–27 (2016)
8. Farrell, P.A., Hegarty, A.: On the determination of the order of uniform convergence. In: Proceedings
of IMACS’91 v. 2, pp. 501–502 (1991)
9. Farrell, P.A., Hegarty, A.F., Miller, J.J.H., O’ Riordan, E., Shishkin, G.I.: Robust computational
techniques for boundary layers. Chapman and Hall (2000)
826 Numer Algor (2017) 75:809–826

10. Linss, T., Stynes, M.: Asymptotic analysis and Shishkin-type decomposition for an elliptic convection-
diffusion problem. J. Math. Anal. Appl. 261, 604–632 (2001)
11. Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: Fitted numerical methods for singular perturbatiom
problems, (revised edition). World Scientific (2012)
12. O’Riordan, E., Shishkin, G.I.: A technique to prove parameter-uniform convergence for a singularly
perturbed convection-diffusion equation. J. Comput. Appl. Math. 206, 136–145 (2007)
13. Ostermann, A., Roche, M.: Runge-Kutta methods for partial differential equations and fractional
orders of convergence. Math. Comp. 59, 403–420 (1992)
14. Roos, H.G., Stynes, M., Tobiska, L.: Robust numerical methods for singularly perturbed differential
equations, (second edition). Springer-Verlag (2008)
15. Sanz-Serna, J.G., Verwer, J., Hundsdorfer, W.H.: Convergence and order reduction of Runge-Kutta
schemes applied to evolutionary problems in partial differential equations. Numer. Math. 50, 405–418
(1986)
16. Shishkin, G.I.: Approximation of the solution to singularly perturbed boundary value problems with
boundary layers. U.S.S.R. Comput. Maths. Math. Phys. 29, 1–10 (1989)
17. Shishkin, G.I., Shsihkina, L.P.: Difference methods for singular perturbation problems. Chapman &
Hall/CRC (2009)
18. Yanenko, N.N.: The method of fractional steps. Springer (1971)