Journal of Computational and Applied Mathematics 370 (2020) 112638

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Singularly perturbed reaction–diffusion problems with

discontinuities in the initial and/or the boundary data
∗
J.L. Gracia a ,1 , , E. O’Riordan b
a
IUMA — Department of Applied Mathematics, University of Zaragoza, Spain
b
School of Mathematical Sciences, Dublin City University, Ireland

article info a b s t r a c t

Article history: Numerical approximations to the solutions of three different problem classes of sin-
Received 9 November 2018 gularly perturbed parabolic reaction–diffusion problems, each with a discontinuity in
Received in revised form 14 November 2019 the boundary-initial data, are generated. For each problem class, an analytical function
Keywords: associated with the discontinuity in the data, is identified. Parameter-uniform numerical
Singularly perturbed approximations to the difference between the analytical function and the solution of the
Discontinuous data singularly perturbed problem are generated using piecewise-uniform Shishkin meshes.
Reaction diffusion Numerical results are given to illustrate all the theoretical error bounds established in
the paper.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction

To establish theoretical error bounds for a numerical method, one requires the solution of the continuous problem
to be sufficiently regular, in order that certain partial derivatives of the solution are bounded on the closed domain.
For parabolic problems, this often requires the assumption of sufficient compatibility conditions between the initial and
boundary data, which can be viewed as solely a theoretical [1] constraint. However, accuracy can be lost in the numerical
approximations if insufficient compatibility is imposed [2,3], especially if a higher order numerical method is utilized.
Moreover, certain mathematical models (e.g., Biot’s consolidation theory of porous media [4,5]) consider mathematical
models with discontinuities between the initial and boundary data deliberately built into the problem formulation. In this
paper, we consider the effect of discontinuous boundary/initial data on the accuracy of numerical approximations, in the
context of singularly perturbed parabolic problems.
The solutions of singularly perturbed problems, with smooth data, typically exhibit boundary layers, whose widths
depend on the singular perturbation parameter. Additional interior layers can appear when the coefficients of the
differential operator are discontinuous or if the inhomogeneous term contains a point source [6–9]. In all of these problem
classes, parameter-uniform numerical methods [10] have been constructed, by using a priori information about both the
location and character of all boundary/interior layers that are present in the solution and using this analytical information
to design an appropriate piecewise-uniform Shishkin mesh [10] for the problem.
In the main, the boundary and initial data for the problem need to be sufficiently smooth, in order to prevent further
classical singularities appearing in the solution. As the smoothness of the boundary and initial data reduces, then the

∗ Corresponding author.
E-mail addresses: [email protected] (J.L. Gracia), [email protected] (E. O’Riordan).
1 The research of this author was partly supported by the Institute of Mathematics and Applications (IUMA), Spain, the project MTM2016-75139-R
and the Diputación General de Aragón, Spain (E24-17R).

https://doi.org/10.1016/j.cam.2019.112638
0377-0427/© 2019 Elsevier B.V. All rights reserved.
2 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

order of convergence can also reduce [11–14]. In the case of sufficiently smooth data satisfying second-order compatibility
conditions at the corners of the space–time domain (see Appendix A), the typical error bound [15] in the L∞ norm for
singularly perturbed parabolic problems of reaction–diffusion type, is of the form

∥Ū − u∥ ≤ C (N −1 ln N)2 + CM −1 ,
when one uses a tensor product of an appropriate piecewise-uniform Shishkin mesh in space (with N elements) and a
uniform mesh (with M elements) in time, to generate a global approximation Ū to the continuous solution u. If there
is only zero-order compatibility conditions assumed and a possible jump in the first time derivative of the boundary
data, then the same numerical method retains parameter-uniform convergence, albeit with some minor reduction in the
order [11] of convergence in space. If there is a discontinuity in the first space derivative of the initial condition, the order
of convergence can drop to O(N −1 + M −0.5 ) [12], [16, §14.2]. Nevertheless, the numerical method (based on an appropriate
piecewise-uniform mesh) retains parameter-uniform convergence of some positive order.
However, for the singularly perturbed heat equation

−ε uxx + ut = 0, (x, t) ∈ (0, 1) × (0, T ],

if the initial condition u(x, 0) is discontinuous [17] or if there is an incompatibility between the initial and boundary
conditions u(0+ , 0) ̸ = u(0, 0+ ) then rectangular meshes do not produce parameter-uniform numerical √ methods [18,19].
Note that if one incorporates a co-ordinate system aligned to the similarity transformation θ = x/(2 ε t), then one can
design a piecewise-uniform mesh [20] in this transformed co-ordinate system, to generate parameter-uniform numerical
approximations. However, we will not consider the use of such transformed co-ordinate systems here.
In this paper, we introduce a mixed analytical/numerical method, which is based on the ideas in [3]. This method first
identifies explicitly the main singular component s(x, t) associated with the singularity and uses a piecewise-uniform
Shishkin mesh to generate a parameter-uniform numerical approximation to the difference, u − s, between the exact
solution u and the main singular component s. In this way, parameter-uniform numerical approximations are created
for singularly perturbed problems with discontinuous initial conditions, problems with incompatible initial/boundary
conditions and problems with discontinuous boundary conditions.
Below we examine singularly perturbed problems of the form

ut − ε uxx + b(t)u = f (x, t), (x, t) ∈ Q := (0, 1) × (0, T ]; b(t) ≥ 0, t ≥ 0,

where the boundary/initial data will have a discontinuity at some point on the boundary ∂ Q := Q̄ \ Q . Note that the
coefficient b(t) is assumed to be independent of the space variable x. This assumption permits us to present relatively
simple proofs for all of the pointwise bounds on the derivatives of the components of the continuous solutions, presented
below. In [21], a related problem class was considered, which involved the differential equation

ε (ut − uxx ) + b(x, t)u = f (x, t), (x, t) ∈ (0, 1) × (0, 1]; b(x, t) > 0;

with incompatible boundary–initial data. Note that the coefficient b(x, t) can vary in space. However, the corresponding
proofs (establishing bounds on the derivatives of the layer components) are significantly longer and contain much more
technical detail to what is required for the three problem classes considered in the current paper. Nevertheless, in our
numerical results section, we present test examples where the coefficient b(x, t) does vary in space and we see that
(from a computational perspective) the assumption b(t) appears not necessary, in practice. In summary, the assumption
b(t) allows us to present theoretical error bounds for three problem classes in a single publication. In this way, we see
the minor modifications in the overall approach, when dealing with singularly perturbed problems with discontinuous
boundary and initial data. In addition, initial layers appeared in the problem class studied in [21], which required the use
of a Shishkin mesh in time. In the current paper, a uniform mesh in time suffices, as the time derivative of the continuous
solution has a coefficient of order one in this paper.
The paper is structured as follows: In Section 2 the asymptotic behaviour of the solution u of the three classes problems
is analysed. For each class of problems the singular component s is identified and the behaviour of u − s is revealed by using
an appropriate decomposition into regular and layer components. In Section 3 a finite difference scheme is proposed to
approximate u − s for each class of problems. Each scheme uses the backward Euler method in time and standard central
differences in space defined on appropriately constructed meshes of Shishkin type. Error estimates in the maximum norm
are established, which yield global parameter-uniform convergence for the methods. In Section 4 some numerical results
for the three classes of problems are given and they indicate that our error estimates are sharp. The paper is completed
with two technical appendices.

Notation. Throughout the paper, C denotes a generic constant that is independent of the singular perturbation parameter
ε and of all discretization parameters. The L∞ norm on the domain D will be denoted by ∥·∥D and the subscript is omitted
if the domain is Q̄ .
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 3

2. Three classes of problem

Before we define the three problem classes to be examined in this paper, we define a set of singular functions which
are associated with the singularities that are generated by discontinuous boundary/initial data in singularly perturbed
problems.
The singular function s : (−∞, ∞) × (0, ∞) → (−1, 1) is defined as
z
( ) ∫
x 2 2
s(x, t) := e−b(0)t erf √ , where erf(z) := √ e−r dr . (1)
2 εt π r =0

This function satisfies the constant coefficient quarter plane homogeneous problem

st − ε sxx + b(0)s = 0, (x, t) ∈ (0, ∞) × (0, ∞),

s(0, t) = 0, t > 0; s(x, 0) = 1, x > 0.

Observe that s ̸ ∈ C ([0, ∞) × [0, ∞)). Define the associated set of functions
0

sn (x, t) := t n s(x, t), n ≥ 0; csn (x, t) := t n (1 − s(x, t)).

Then sn , csn ∈ C n−1+γ ([0, ∞) × [0, ∞)), n ≥ 1.2 Note further that

(sn )t − ε (sn )xx + b(0)(sn ) = n(sn−1 )

and (csn )t − ε (csn )xx + b(0)(csn ) = n(csn−1 ) + b(0)t n .

Hence, for all n ≥ 1, we have the recurrence relationship

Lsn (x, t) = nsn−1 (x, t) + (b(t) − b(0))sn (x, t), (2)

where Lz := zt − ε zxx + b(t)z .

2.1. Problem Class 1: incompatible boundary–initial data

Consider the singularly perturbed parabolic problem: Find u : Q̄ → R with Q := (0, 1) × (0, T ], such that

Lu = ut − ε uxx + b(t)u = f (x, t), in Q ; b(t) ≥ β ≥ 0, ∀t ≥ 0; (3a)

u(0, t) = 0, u(1, t) = 0, t ≥ 0, u(x, 0) = φ (x), 0 < x < 1; (3b)
φ (0+ ) ̸= 0, φ (1) = 0, f , b ∈ C 4+γ (Q̄ ), φ ∈ C 4 (0, 1); (3c)
(iv )
f (1, 0) = −εφ (1 ), ′′ −
−εφ (1 ) + b(0)φ (1 ) = (ft + fxx )(1, 0);
− ′′ −
(3d)
f (0, 0) = −εφ (0 ). ′′ +
(3e)

Since this problem is linear, there is no loss in generality in assuming homogeneous boundary conditions. Observe that
there is a discontinuity in the data at the corner point (0, 0). The discontinuity in the data for this first problem class
is the same discontinuity as that examined in [21]. The assumption (3e) on the data allows us to present a simplified
version of the numerical analysis. Without this assumption, we would require more of the analysis from [21]. By assuming
the compatibility conditions (3d), we prevent any classical singularities appearing in the vicinity of the point (1, 0) (see
Appendix A).
In order to deduce the asymptotic behaviour of the solution of problem (3), it is decomposed into the sum

u = φ (0+ )s(x, t) + y. (4)

Note that |φ (0 )| is the magnitude of the jump in the boundary/initial data, at (0, 0). The remainder y, defined by (4),
+

satisfies the problem

Ly = F := f − (b(t) − b(0))φ (0+ )s, (x, t) ∈ Q ; (5a)

y(0, t) = 0, y(1, t) = −φ (0 )s(1, t), t ≥ 0; +
(5b)
y(x, 0) = φ (x) − φ (0 ), 0 < x < 1.+
(5c)

2 The space C n+γ (Q̄ ) is the set of all functions, whose derivatives of order n are Hölder continuous of degree γ > 0. That is,

∂ i+j z
{ }
γ
C n+γ (Q̄ ) := z: ∈ C (Q̄ ) , 0 ≤ i + 2j ≤ n .
∂ xi ∂ t j
4 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

Recall that φ (1) = 0 and so y(1− , 0) = y(1, 0+ ). Hence the boundary and initial data are continuous in the case of problem
(5), F ∈ C 0+γ (Q̄ ) and, using assumption (3e), we have that y ∈ C 2+γ (Q̄ ). We further decompose the solution of (5) as
follows:

y = v + wL + wR , (6a)

where the regular component v satisfies the problem

L∗ v ∗ = f ∗ − (b(t) − b(0))φ (0+ ), (x, t) ∈ Q0∗ := (−a, 1 + a) × (0, T ], (6b)

which is posed on an extended (in the spatial direction) domain Q0∗ 3 and a is an arbitrary positive parameter. The
initial/boundary values for the regular component are determined by v ∗ = v0∗ + εv1∗ , where these two subcomponents, in
turn, satisfy

bv0∗ + (v0∗ )t = f ∗ − (b(t) − b(0))φ (0+ ), t > 0; (6c)

v0 (x, 0) = y (x, 0),
∗ ∗
x ∈ (−a, 1 + a), (6d)
L v1 = (v0 )xx ,
∗ ∗ ∗
(x, t) ∈ Q0 ,
∗
v1 = 0, (x, t) ∈ ∂ Q0 .
∗ ∗
(6e)

Observe that the singular function s is not involved in the definition of the regular component. Moreover, observe that
∫ t
v0 (0, t) = f (0, s) − (b(s) − b(0))φ (0+ ) e−(b(t)−b(s)) ds, t ≥ 0.
( )
s=0

The boundary layer components wL , wR satisfy the problems

L∗ wL∗ = F ∗ − L∗ v ∗ , (x, t) ∈ Q1∗ := (0, 1 + a) × (0, T ], (6f)

wL (0, t) = −v ∗ (0, t), wL∗ (x, 0) = 0, wL∗ (1 + a, t) = 0;
∗
(6g)
LwR = 0, (x, t) ∈ Q , (6h)
wR (1, t) = −(v + wL )(1, t), wR (x, 0) = 0, wR (0, t) = 0.
∗ ∗
(6i)

By construction and by using the extended domains to avoid compatibility issues, v, wR ∈ C 4+γ (Q̄ ). Note that L∗ wL∗ (0, 0) =
0 and vt (0, 0) = f (0, 0), so the first order compatibility conditions are satisfied (see Appendix A). Hence, wL ∈ C 2+γ (Q̄ );
but, in general, wL ̸ ∈ C 4+γ (Q̄ ).

Theorem 1 (Problem Class 1). For the regular component v we have, for all 0 ≤ i + 2j ≤ 4 with 0 < µ < 1, the bounds
∂ v
 i+j 
1−i/2
 ∂ xi ∂ t j  ≤ C (1 + ε ), (7a)
 

and for the boundary layer components, we have the bounds

⏐ ∂
⏐ i+j ⏐
(1−x)
w , ⏐ ≤ C ε−i/2 e− √ε , 0 ≤ i + 2j ≤ 4; (x, t) ∈ Q̄ ;
⏐
⏐ ∂ xi ∂ t j R
⏐ (x t)⏐ (7b)

⏐ ∂
⏐ i+j
µ x
⏐
−i/2 − 2 √T ε
⏐ ∂ xi ∂ t j wL (x, t)⏐ ≤ C ε e , 0 ≤ i + 2j ≤ 2; (x, t) ∈ Q̄ .
⏐
⏐ ⏐ (7c)

In addition, for the higher derivatives, we have

⏐∂
⏐ i
− µ √x
⏐
⏐ wL (x, t)⏐ ≤ √ C e 2 Tε , i = 3, 4; , (x, t) ∈ Q ;
⏐
⏐ ∂ xi (7d)
⏐ ε ( εt) i− 2

⏐∂
⏐ 2 ⏐
C
⏐ ∂ t 2 wL (x, t)⏐ ≤ t , (x, t) ∈ Q .
⏐
⏐ ⏐ (7e)

Proof. We begin by establishing the bounds on the regular component. √ Note that v0 is bounded independently of ε.
Consider the problem (6e) transformed with the stretched variable x/ ε . Apply the a priori bounds [22] to establish
bounds on the partial derivatives of v1 in the stretched variables. Transforming back to the original variables, we deduce
the bounds (7a). The bounds on wR are obtained in the usual way [15].
We now consider the component wL . Observe that, with csn := t n (1 − s), we have

L∗ wL∗ = (b(t) − b(0) − tb′ (0))φ (0+ )cs0 + b′ (0)φ (0+ )cs1 , (x, t) ∈ Q1∗ ;

3 We use the notation f ∗ : Q̄ ∗ → R to denote the extension of any function f : Q̄ → R such that f ∗ (x, t) ≡ f (x, t), (x, t) ∈ Q̄ and Q̄ ⊂ Q̄ ∗ .
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 5

and L∗ wL∗ ∈ C 0+γ (Q̄1∗ ). Using e−z ≤ e0.25−z , it follows that

2

− √x − √x
|cs0 (x, t)| ≤ Ce 2 εT , |cs1 (x, t)| ≤ Cte 2 εT , t ≤ T;
⏐∂
⏐ ⏐ √
1
⏐ cs1 (x, t)⏐ ≤ Ce−µ 2t ε x , µ < 1.
⏐
⏐ ∂t ⏐
Hence, using a maximum principle, we can deduce that
− √x 1
|wL (x, t)| ≤ Ce 2 εT eθ t , θ> − β, (x, t) ∈ Q ;
4T
and, by applying the arguments from [15], we get that for 0 ≤ i + 2j ≤ 2,
⏐ ∂
⏐ i+j
µx
⏐
w , ⏐ ≤ C ε−i/2 e− 2√T ε , (x, t) ∈ Q .
⏐
⏐
⏐ ∂ xi ∂ t j L (x t) ⏐
To obtain bounds on the higher derivatives of wL , we introduce a further decomposition of this boundary layer function.
Consider the continuous function
∫ t
P(x, t) := B(t) − s0 (x, r) dr ,
r =0
{
1−e−b(0)t
b(0)
, if b(0) ̸ = 0,
where B(t) :=
t, if b(0) = 0.

This function has been constructed to satisfy the following problem

Pt − ε Pxx + b(0)P = 0, in Q ,
P(0, t) = B(t) t ≥ 0, P(x, 0) = 0, 0 < x < 1.

Note that
∫ t
P(x, t) = B(t) − t + cs0 (x, r) dr .
r =0

We introduce the secondary expansion

1
wL∗ (x, t) = −vt∗ (0, 0)P ∗ (x, t) + b′ (0)φ (0+ ) cs2 (x, t) + R∗ (x, t);
2
and the remainder term R∗ , defined over Q̄1∗ , satisfies the problem

L∗ R∗ = (b(t) − b(0) − b′ (0)t)φ (0+ )cs0 + (b(t) − b(0))vt∗ (0, 0)P ∗

1
− (b(t) − b(0))b′ (0)φ (0+ ) cs2 ,
2
R∗ (x, 0) = 0, 0 < x < 1 + a; R∗ (1 + a, t) = R∗ (1 + a, t), t ≥ 0;
t2
R∗ (0, t) = (t vt∗ (0, 0) − v ∗ (0, t)) + vt∗ (0, 0)(B(t) − t) − b′ (0)φ (0+ ) .
2
Note that v (0, 0) = 0 and |R∗ (0, t)| ≤ Ct 2 . Moreover, using properties of s2 (see Appendix B) one can check that
L∗ R∗ ∈ C 2+γ (Q̄1∗ ) and second level compatibility is satisfied at the point (0, 0). Hence, we have R∗ ∈ C 4+γ (Q̄1∗ ). Using
this regularity, we can deduce the bounds
∂ R 
 i+j ∗ 
−i/2
 ∂ xi ∂ t j  ≤ C (1 + ε ), 0 ≤ i + 2j ≤ 4.
 

From this, we obtain

⏐∂
⏐ 2 ⏐
w , ⏐ ≤ Ct −1 , (x, t) ∈ Q .
⏐
⏐
⏐ ∂t2 L (x t) ⏐
To obtain the desired bounds (involving the decaying exponential) on the third and fourth space derivatives of wL , we
form a problem for R∗2 := ε R∗xx by differentiating the differential equation satisfied by R∗ twice to get

L∗ R∗2 = (b(0) + b′ (0)t − b(t))φ (0+ )ε (cs0 )xx + (b(t) − b(0))vt (0, 0)ε Pxx
∗

1
− (b(t) − b(0))b′ (0)φ (0+ )ε (cs2 )xx ,
2
R∗2 (x, 0) = 0, R∗2 (1 + a, t) = R∗2 (1 + a, t); R∗2 (0, t) = (g ′ + bg − h)(t);
6 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

where g(t) := R∗ (0, t), h(t) := L∗ R∗ (0, t) are both smooth functions independent of ε . We can complete the proof (as
in [15]) by noting that
x
⏐L R (x, t)⏐ ≤ Ce−µ √4εT ;
⏐ ∗ ∗ ⏐
2
x
⏐(L R )t (x, t)⏐ + ⏐ε(L∗ R∗ )xx (x, t)⏐ ≤ Ce−µ √4εT . □
⏐ ∗ ∗ ⏐ ⏐ ⏐
2 2

2.2. Problem Class 2: discontinuous initial condition

Consider the singularly perturbed parabolic problem

Lu = f (x, t), (x, t) in Q ; u(0, t) = 0, u(1, t) = 0, t ≥ 0; (8a)

u(x, 0) = φ (x), 0 < x < 1, φ (0) = φ (1) = 0; (8b)
φ (d ) ̸= φ (d ),
− +
0 < d < 1; (8c)
(iv )
f (0, 0) = −εφ (0 ), −εφ′′ +
(0 ) + b(0)φ (0 ) = (ft + fxx )(0, 0);
+ ′′ +
(8d)
f (1, 0) = −εφ ′′ (1− ), −εφ (iv ) (1− ) + b(0)φ ′′ (1− ) = (ft + fxx )(1, 0); (8e)
f,b ∈ C 4+γ
(Q̄ ), φ ∈ C ((0, 1) \ {d}),
4
(8f)
φ (d ) = φ (d ),
′ − ′ +
φ (d ) = φ (d ).
′′ − ′′ +
(8g)

The assumption of the compatibility conditions (8b), (8d) and (8e) ensures that no classical singularity appears near the
corner points (0, 0), (1, 0). However, observe that the initial function φ (x) is discontinuous at x = d. This will cause an
interior layer to appear in the solution, near the point (d, 0). The assumption φ ′ (d− ) = φ ′ (d+ ) on the data prevents a drop
in the order of convergence in our numerical approximations, as in the case of [16, §14.2].
Decompose the solution of (8) into the following sum
[φ](d)
u(x, t) = s(x − d, t) + y(x, t), where [φ](d) := φ (d+ ) − φ (d− ). (9)
2
By the definition (1) of the discontinuous function s, we have that

⎨−1, for x < d,

⎧

s(x − d, 0) = 0, for x = d,
1, for x > d.
⎩

The component y is the solution of the problem

Ly = f − (b(t) − b(0))0.5[φ](d)s(x − d, t), (x, t) ∈ Q , (10a)

y(0, t) = −0.5[φ](d)s(−d, t), y(1, t) = −0.5[φ](d)s(1 − d, t), t ≥ 0, (10b)
y(x, 0) = φ (x) − 0.5[φ](d)s(x − d, 0), x ̸ = d; (10c)
y(d, 0) = (φ (d+ ) + φ (d− ))/2; (10d)

and y(x, 0) is continuous for all x ∈ [0, 1]. Moreover, due to (8g), we have y(x, 0) ∈ C 2 (0, 1). Using the maximum principle

∥y∥ ≤ C .
The solution y is further decomposed into the sum

y = v + wL + wR + wI ,

where the components v and wI are discontinuous functions and the components wL and wR are continuous functions.
The regular component v is constructed to satisfy the problem

L∗ v ∗ = f ∗ − (b(t) − b(0))0.5[φ](d)e−b(0)t s(x − d, 0), (x, t) ∈ Q ∗ . (11)

This problem is posed on the extended domain Q := (−a, 1 + a) × (0, T ], a > 0. The initial/boundary values for the
∗

regular component are determined by v ∗ = v0∗ + εv1∗ , where the reduced solution v0 satisfies the initial value problem

bv0∗ + (v0∗ )t = f ∗ − (b(t) − b(0))0.5[φ](d)e−b(0)t s(x − d, 0), t > 0; (12a)

v0 (x, 0) = y (x, 0), x ∈ (−a, 1 + a).
∗ ∗
(12b)

Observe that the reduced solution v0 is continuous, but in general

∗

(v0∗ )x (d+ , 0) ̸ = (v0∗ )x (d− , 0).

 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 7

The first correction v1 is defined as the multi-valued function

{ −
v1 , for x ≤ d, t ≥ 0,
v1 (x, t) :=
∗
v1 ,
+
for x ≥ d, t ≥ 0,
and the two sides of this function are the solutions of
L∗ v1− = (v0∗ )xx ,
−a < x ≤ d, t > 0;
v1 (−a, t) = 0, t ≥ 0; v1− (x, 0) = 0, x ≤ d;
−

L∗ v1+ = (v0∗ )xx , d ≤ x < 1 + a, t > 0 ;

v1 (1 + a, t) = 0, t ≥ 0; v1+ (x, 0) = 0, x ≥ d;
+

where we use Lv1− (d, t) := (v0 )xx (d− , t) and Lv1+ (d, t) := (v0 )xx (d+ , t). Since the regular component v is multi-valued we
now define the subdomains
Q − := (0, d) × (0, T ] and Q + := (d, 1) × (0, T ].
By using suitable extensions to these subdomains, we can have v1± ∈ C 4+γ (Q̄ ± ). For example,
L∗ (v1− )∗ = ((v0 )xx )∗ , (x, t) ∈ Q∗− := (−a, d + a) × (0, T ];
(v1− )∗ (−a, t) = (v1− )∗ (d + a, t) = 0, t ≥ 0, (v1− )∗ (x, 0) = 0, −a < x < d + a.

The boundary layer components wL , wR satisfy the problems

LwL = LwR = 0, (x, t) ∈ Q , (13a)
wL (0, t) = −v (0, t), wL (x, 0) = 0, wL (1, t) = 0, (13b)
wR (0, t) = 0, wR (x, 0) = 0, wR (1, t) = −v (1, t). (13c)

Theorem 2 (Problem Class 2). For the regular component v we have, for all 0 ≤ i + 2j ≤ 4, the bounds
∂ v 
 ,  ∂ v  ≤ C (1 + ε1−i/2 ).
 i+j −   i+j + 
 

 ∂ xi ∂ t j  −  ∂ xi ∂ t j  + (14a)
Q̄ Q̄

For the boundary layer components, for all 0 ≤ i + 2j ≤ 4 and (x, t) ∈ Q̄ ,

⏐ ∂ wL ⏐ ∂ wR
⏐ i+j ⏐ ⏐ i+j ⏐
x (1−x)
, ⏐ ≤ C ε−i/2 e− √ε ; , ⏐ ≤ C ε−i/2 e− √ε .
⏐ ⏐
⏐
⏐ ∂ xi ∂ t j (x t) ⏐
⏐
⏐ ∂ xi ∂ t j (x t)⏐ (14b)

Proof. Adapt appropriately the argument from the proof of Theorem 1. □

Finally the multi-valued interior layer component
{ −
wI , for 0 ≤ x ≤ d, t ≥ 0,
wI (x, t) :=
wI ,+
for d ≤ x ≤ 1, t ≥ 0;

is defined implicitly by the sum y = v + wL + wR + wI . Hence, ∥wI ∥ ≤ C . Moreover, wI satisfies the problem
LwI = R(x, t), (x, t) ∈ Q − ∪ Q + , (15a)
where
R(x, t) := (b(t) − b(0))0.5[φ](d) e−b(0)t s(x − d, 0) − s(x − d, t) ,
( )
(15b)
wI (0, t) = 0, wI (1, t) = 0, t ≥ 0; wI (x, 0) = 0, 0 ≤ x ≤ 1; (15c)
[wI ](d, t) = −ε[v1 ](d, t), [(wI )x ](d, t) = −ε[(v1 )x ](d, t). (15d)

Note that, R ∈ C 0+γ

(Q̄ ), R(d, t) ̸ = 0 for all t > 0 such that b(t) ̸ = b(0). Moreover,
(x−d)2 − |x√−d|
|R(x, t)| ≤ Cte− 4ε t ≤ Cte 2 εT , (x, t) ∈ Q̄
and
⏐∂ µ|x−d|
⏐ ⏐
⏐ R(x, t)⏐ ≤ Ce− 2√εt , µ < 1, (x, t) ∈ Q .
⏐
⏐ ∂t ⏐

Using a maximum principle, either side of x = d, we have for 0 ≤ i + 2j ≤ 2

− |x√−d|
|wI (x, t)| ≤ Ce 2 εT ; (x, t) ∈ Q̄ − ∪ Q̄ + ; (16a)
8 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

⏐ ∂
⏐ i+j
µ|x−d|
⏐
w , ⏐ ≤ C ε−i/2 e− 2√εT , µ < 1, (x, t) ∈ Q̄ − ∪ Q̄ + .
⏐
⏐
⏐ ∂ xi ∂ t j I (x t) ⏐ (16b)

For the higher derivatives, we need to repeat the argument from the proof of Theorem 1, from the last section, to establish
the additional bounds
⏐∂
⏐ i
−µ
⏐
⏐ wI (x, t)⏐ ≤ √ C
x−d|
|√
Tε , i = 3, 4, (x, t) ∈ Q − ∪ Q + ;
⏐
2
⏐ ∂ xi ⏐ ε( εt)i−2 e (16c)

⏐∂
⏐ 2 ⏐
⏐ C
⏐
⏐ ∂t2 wI (x, t) ⏐ ≤ , (x, t) ∈ Q − ∪ Q + . (16d)
⏐ t

2.3. Problem Class 3: discontinuous boundary data

Consider the singularly perturbed parabolic problem

Lu = f (x, t) in Q , u(1, t) = 0, t ≥ 0, u(x, 0) = 0, 0 < x < 1, (17a)

and the boundary condition at x = 0 is given by

φ1 (t), if 0 ≤ t ≤ d,
{
u(0, t) = φ1 (d− ) ̸= φ2 (d+ ), φ1′ (d− ) = φ2′ (d+ ). (17b)
φ2 (t), if d < t ≤ T ,
Note that there is no loss of generality in assuming a homogeneous initial condition. We assume that the following
compatibility conditions are satisfied at (0, 0) and (1, 0):

φ1 (0) = 0, f (0, 0) = φ1′ (0+ ), (ft + fxx )(0, 0) = φ1′′ (0+ ) + b(0)φ1′ (0+ ), (17c)
f (1, 0) = (ft + fxx )(1, 0) = 0, (17d)

and also the following regularity conditions

f , b ∈ C 4+γ (Q̄ ), φ1 ∈ C 2 (0, d), φ2 ∈ C 2 (d, T ). (17e)

The discontinuous boundary condition on the left, will cause a singularity to appear in the solution for t ≥ d.
Decompose the solution of (17) into the sum

u = [φ](d)H(t − d)cs(x, t − d) + y, [φ](d) := φ2 (d+ ) − φ1 (d− ), (18)

where H(·) is a unit step function defined by

0, for x < 0,
{
H(x) :=
1, for x ≥ 0.

Note that cs(x, 0) = 0. Observe that y is the solution of the parabolic problem

Ly = f + (b(d) − b(t))[φ](d)H(t − d)cs(x, t − d) in Q , (19a)

y(1, t) = −[φ](d)H(t − d)cs(1, t − d), t ≥ 0, (19b)
y(x, 0) = φ (x), 0 < x < 1, (19c)

and

y(0, t) = u(0, t) − [φ](d)H(t − d). (19d)

As in previous sections, we decompose y into three subcomponents

y = v + wR + wL ,

which are defined as the solutions of the following three parabolic problems. The regular component satisfies

L∗ v ∗ = f ∗ in Q ∗ := (−a, 1 + a) × (0, T ], (20a)

v (−a, t) = v ∗ (1 + a, t) = 0, t ≥ 0,
∗
(20b)
v ∗ (x, 0) = φ ∗ (x), −a < x < 1 + a; (20c)

where φ (x) is a smooth extension of the initial condition (17a). The right boundary layer component satisfies
∗

LwR = 0 in Q , wR (x, 0) = 0, 0 < x < 1, (20d)

wR (1, t) = y(1, t) − v (1, t), wR (0, t) = 0, t ≥ 0; (20e)
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 9

and the left boundary layer component satisfies

LwL = Ly − f in Q , wL (x, 0) = 0, 0 < x < 1, (20f)

wL (1, t) = 0, wL (0, t) = y(0, t) − v (0, t), t ≥ 0. (20g)

The regular component v ∈ C 4+γ (Q̄ ) and since all time derivatives of cs(1, t − d) are zero at t = d, we have that
wR ∈ C 4+γ (Q̄ ). In addition, wL ∈ C 2+γ (Q̄ ). Hence, the character of the function y for Problem Class 3 is the same as
for Problem Class 1. In other words, the bounds on the derivatives of the components of y given in Theorem 1 also apply
in the case of Problem Class 3. However, the character of the singular component u − y is different for the two problem
classes.

3. Numerical method

For all three problem classes we employ a classical finite difference operator (backward Euler in time and standard
central differences in space) on an appropriate mesh (which will be piecewise-uniform in space and uniform in time).
The piecewise-uniform Shishkin mesh for each of the three Problem Classes n, (n = 1, 2, 3) will be denoted by Q̄nN ,M and
∂ QnN ,M := Q̄nN ,M ∩ ∂ Q . The numerical method4 :
LN ,M Y (xi , tj ) = f (xi , tj ), (xi , tj ) ∈ QnN ,M , (21a)
Y (xi , tj ) = y(xi , tj ), (xi , tj ) ∈ ∂ QnN ,M , (21b)
where LN ,M Y (xi , tj ) := (−εδx2 + b(tj )I + D−
t )Y (xi , tj ). (21c)
N ,M
For Problem Class 1, the Shishkin mesh Q̄1 is defined via:
1 4√
{ }
[0, 1] = [0, σ ] ∪ [σ , 1 − σ ] ∪ [1 − σ , 1], σ := min , ε T ln N , µ < 1;
4 µ
and N /4, N /2, N /4 grid points are uniformly distributed in each subinterval, respectively. For Problem Class 3, the Shishkin
N ,M
mesh is Q̄3 = Q̄1N ,M .
N ,M
For Problem Class 2, the Shishkin mesh is Q̄2 , which is defined via

[0, 1] = [0, τ ] ∪ [τ , d − τ ] ∪ [d − τ , d + τ ] ∪ [d + τ , 1 − τ ] ∪ [1 − τ , 1],

with
1 4√
{ }
τ := min , ε T ln N , µ<1
8 µ
and N /8, N /4, N /4, N /4, N /8 grid points are uniformly distributed in each subinterval, respectively. Although it is
required that µ < 1 in the theoretical error analysis, in the numerical results section, we simply have taken µ = 1.

Theorem 3. Let Y be the solution of the finite difference scheme (21) and y the solution of the continuous problem. Then, the
global approximation Ȳ on Q̄ generated by the values of Y on Q̄nN ,M and bilinear interpolation, satisfies

∥y − Ȳ ∥Q̄ ≤ (CN −2 ln2 N + CM −1 ) ln M , (22)

for each of the three Problem Classes (3), (8) and (17).

Proof. For each of the three Problem Classes, the discrete solution Y is decomposed along the same lines as its continuous
counterpart y.
Let us first consider Problem Classes 1 and 3. Using the bounds on the derivatives of the components in Theorem 1,
truncation error bounds, discrete maximum principle and a suitable discrete barrier function and following the arguments
in [15], we can establish the following bounds

∥v − V ∥Q̄ N ,M , ∥wR − WR ∥Q̄ N ,M ≤ CN −2 ln2 N + CM −1 .

4 The finite difference operators are defined by:

U(xi , tj ) − U(xi−1 , tj )
x U(xi , tj ) := Dx U(xi+1 , tj );
D+ x U(xi , tj ) := ,
−
D−
hi
U(xi , tj ) − U(xi , tj−1 ) x − Dx )U(xi , tj )
(D+ −
t U(xi , tj ) :=
D− , δx2 U(xi , tj ) :=
kj h̄i
and the mesh steps are hi := xi − xi−1 , h̄i = (hi+1 + hi )/2, k = kj := tj − tj−1 .
10 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

It remains to bound the error due to the left boundary layer component. We introduce the following notation for this
error and the associated truncation error

and Ti := LN ,M Ei .
j j j
Ei := (wL − WL )(xi , tj )

Note that
 ∂ wL 
 2 
|δ wL (xi , tj )| ≤ C 
2
x

 ∂ x2  ,
(xi−1 ,xi+1 )×{tj }

 ∂wL 
 
|D−t w (x
L i j, t )| ≤ C 
 ∂t 
 .
{xi }×(tj−1 ,tj )

Hence, using the bounds (7c) on the derivatives of wL , we have that outside the left layer

|Ti j | ≤ CN −2 , xi ≥ σ , tj > 0.

Within the left layer, using the bounds (7d), (7e) on the higher derivatives of wL , we have the truncation error bounds
(N −1 ln N)2  ∂wL  (N −1 ln N)2
 
1
|Ti | ≤ C +C   ≤ C + C, xi < σ
t1 ∂ t {xi }×(0,t1 ) t1

(N −1 ln N)2  ∂ wL 
 2 
j
|Ti | ≤ C + Ck 
 ∂t2 

tj {xi }×(tj−1 ,tj )

(N −1 ln N)2 k
≤C +C , xi < σ , tj > t1 .
tj tj−1
Hence, at all time levels, we have the truncation error bound
(N −1 ln N)2 + M −1
|Ti j | ≤ C , xi < σ , tj ≥ t1 .
tj
We now mimic the argument in [13] and note that at each time level,
( )
1 1 j−1
−εδx2 Eij + b(xi , tj ) + j j
Ei = Ti + Ei , tj > 0.
k k
From this we can deduce the error bound
j j
j
∑
n
( −1 2 −1
)∑ 1
| | ≤ Ck
Ei |Ti | ≤ C (N ln N) + M
n
n=1 n=1
j
( ∫ )
ds
≤ C (N −1 ln N)2 + M −1 1 +
( )
s=1 s

≤ C ((N −1
ln N) + M 2 −1
) ln(1 + j). (23)

In the case of Problem Class 2, we have an additional interior layer component wI . The bounding of the error ∥wI − WI ∥
follows the same argument as above.
One can extend this nodal error bound to a global error bound by applying the argument in [10, pp. 56–57] and using
the modification in [21] to manage the initial singularity. □

4. Numerical results

The orders of convergence of the finite difference scheme (21) are estimated using the two-mesh principle [10]. We
denote by Y N ,M and Y 2N ,2M the computed solutions with (21) on the Shishkin meshes QnN ,M and Qn2N ,2M , respectively.
These solutions are used to compute the maximum two-mesh global differences

DNε ,M := ∥Ȳ N ,M − Ȳ 2N ,2M ∥Q N ,M ∪Q 2N ,2M

n n

N ,M 2N ,2M
where Ȳ and Ȳ denote the bilinear interpolation of the discrete solutions Y N ,M and Y 2N ,2M on the mesh QnN ,M ∪
Qn2N ,2M . Then, the orders of global convergence PεN ,M are estimated in a standard way [10]

DNε ,M
( )
PεN ,M := log2 2N ,2M
.
Dε
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 11

Table 1
Example 1 from Problem Class 1: Maximum two-mesh global differences and orders of convergence for the function y in (5) using a piecewise
uniform Shishkin mesh.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
1.295E−02 6.990E−03 3.650E−03 1.870E−03 9.453E−04
ε=2 0
0.890 0.938 0.965 0.984
3.789E−03 1.980E−03 1.013E−03 5.128E−04 2.580E−04
ε = 2−2
0.936 0.966 0.983 0.991
2.214E−03 1.081E−03 5.339E−04 2.653E−04 1.322E−04
ε = 2−4
1.035 1.017 1.009 1.005
4.216E−03 2.083E−03 1.035E−03 5.161E−04 2.577E−04
ε = 2−6
1.017 1.008 1.004 1.002
4.971E−03 2.456E−03 1.220E−03 6.084E−04 3.037E−04
ε = 2−8
1.017 1.009 1.004 1.002
5.269E−03 2.598E−03 1.290E−03 6.428E−04 3.208E−04
ε = 2−10
1.020 1.010 1.005 1.003
6.402E−03 2.668E−03 1.321E−03 6.569E−04 3.276E−04
ε = 2−12
1.263 1.015 1.008 1.004
1.092E−02 4.011E−03 1.342E−03 6.655E−04 3.313E−04
ε = 2−14
1.445 1.580 1.012 1.006
1.092E−02 4.013E−03 1.347E−03 6.685E−04 3.326E−04
ε = 2−16
1.445 1.575 1.011 1.007
. . . . . .
. . . . . .
. . . . . .
1.093E−02 4.014E−03 1.352E−03 6.707E−04 3.337E−04
ε = 2−30
1.445 1.571 1.011 1.007
DN , M 1.295E−02 6.990E−03 3.650E−03 1.870E−03 9.453E−04
P N ,M 0.890 0.938 0.965 0.984

The uniform two-mesh global differences DN ,M and their corresponding uniform orders of global convergence P N ,M are
calculated by
D N ,M
( )
DN ,M := max DNε ,M , P N ,M := log2 ,
ε∈S D2N ,2M

where S = {20 , 2−1 , . . . , 2−30 }.

In order that the temporal discretization error dominates the spatial discretization error, in all the tables of this paper,
except in Table 2, we have taken N = 24 × M.

4.1. Problem Class 1

We present numerical results for two examples from this first class of problems. In the first example the coefficient of
the reaction term depends only on the temporal variable; while in the second example, it depends on the spatial variable.
The numerical results computed with the analytical/numerical method of this paper suggest that the method is uniformly
and globally convergent in both cases.

Example 1. Consider problem (3), with the data given by

b(x, t) = 1 + t , f (x, t) = 4x(1 − x)t + t 2 , φ (x) = (1 − x)2 . (24)

The maximum two-mesh global differences associated with the component y and the orders of convergence are given in
Table 1. Observe that the numerical results show that the method is first-order globally parameter-uniformly convergent.
In Table 1 we take N = 16M, which means that the largest spatial mesh step is significantly smaller than the uniform mesh
step in time; so the temporal convergence rate dominates. In Table 2, we give the uniform two-mesh global differences
taking N = M and the computed orders of convergence illustrate that the method is almost second order convergent; in
this case the spatial discretization errors dominate the temporal discretization errors. The numerical results in Tables 1
and 2 are in agreement with our error estimates in Theorem 3.
We display in Fig. 1 the numerical approximation to the function y defined in (5), which exhibits only boundary layers.
The numerical solution to problem (3)–(24) is displayed in Fig. 2, which exhibits both boundary layers and the singularity
caused by the incompatibility between the initial and boundary conditions.
12 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

Table 2
Example 1 from Problem Class 1: Uniform two-mesh global differences and orders of convergence for the function y in (5) using a piecewise uniform
Shishkin mesh with N = M.
N = M = 64 N = M = 128 N = M = 256 N = M = 512 N = M = 1024
DN , M 4.972E−02 2.548E−02 1.117E−02 3.983E−03 1.330E−03
P N ,M 0.964 1.189 1.488 1.583

Fig. 1. Example 1 from Problem Class 1: The numerical approximation to y with ε = 2−16 and N = M = 64.

Fig. 2. Example 1 from Problem Class 1: The approximation s + Y to the solution u with ε = 2−16 and N = M = 64.

Example 2. Consider problem (3), with the data given by

b(x, t) = 1 + 10x, f (x, t) = 4x(1 − x)t + t 2 , φ (x) = (1 − x)2 . (25)
The maximum two-mesh global differences associated with the component y and the orders of convergence are given
in Table 3. Observe that the numerical results indicate that the method is globally parameter-uniformly convergent.
Comparing these orders of convergence with those in Table 1, we see that the theoretical assumption of b being
independent of the space variable appears to be not necessary in order to observe parameter-uniform convergence.

4.2. Problem Class 2

Example 3. Consider problem (8) with the data given by

b(x, t) = 1 + 10xt , f (x, t) = 4x(1 − x)t + t 2 , (26a)
−1 + (2x − 1) ,2
0 ≤ x ≤ 0.5,
{
if
φ (x) = (26b)
1 − (2x − 1)2 , if 0.5 < x ≤ 1.
Observe that in this example the coefficient b depends on the temporal and spatial variables and, moreover, φ ′ (0.5+ ) =
φ ′ (0.5− ) but φ ′′ (0.5+ ) ̸= φ ′′ (0.5− ). The schemes considered here to approximate the solution of this example are defined
N ,M
on the Shishkin mesh Q̄2 .
If the singularity is not stripped off and Example 3 is simply solved with backward Euler method and standard central
N ,M
finite differences on the Shishkin mesh Q̄2 , the method is not globally convergent for any value of ε . This is illustrated
in Table 4 where the uniform two-mesh global differences are given.
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 13

Table 3
Example 2 from Problem Class 1: Maximum two-mesh global differences and orders of convergence for the function y in (5) using a piecewise
uniform Shishkin mesh.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
4.837E−03 4.267E−03 2.321E−03 1.160E−03 5.823E−04
ε=2 0
0.181 0.878 1.000 0.995
9.114E−03 4.665E−03 2.341E−03 1.172E−03 5.863E−04
ε = 2−2
0.966 0.994 0.998 0.999
1.086E−02 5.523E−03 2.787E−03 1.400E−03 7.016E−04
ε = 2−4
0.975 0.987 0.993 0.997
1.092E−02 5.531E−03 2.784E−03 1.398E−03 7.006E−04
ε = 2−5
0.982 0.990 0.994 0.997
1.068E−02 5.387E−03 2.712E−03 1.361E−03 6.814E−04
ε = 2−6
0.988 0.990 0.995 0.998
1.047E−02 5.305E−03 2.672E−03 1.341E−03 6.717E−04
ε = 2−8
0.980 0.990 0.995 0.997
1.056E−02 5.349E−03 2.693E−03 1.351E−03 6.769E−04
ε = 2−10
0.982 0.990 0.995 0.997
1.059E−02 5.361E−03 2.698E−03 1.354E−03 6.782E−04
ε = 2−12
0.982 0.990 0.995 0.997
1.060E−02 5.365E−03 2.700E−03 1.355E−03 6.786E−04
ε = 2−14
0.982 0.991 0.995 0.997
1.061E−02 5.368E−03 2.701E−03 1.355E−03 6.787E−04
ε = 2−16
0.983 0.991 0.995 0.997
. . . . . .
. . . . . .
. . . . . .
1.062E−02 5.371E−03 2.702E−03 1.355E−03 6.788E−04
ε = 2−30
0.984 0.991 0.995 0.998
DN , M 1.092E−02 5.531E−03 2.787E−03 1.400E−03 7.016E−04
P N ,M 0.982 0.989 0.993 0.997

Table 4
Example 3 from Problem Class 2: Maximum two-mesh global differences and orders of convergence for u using a piecewise uniform Shishkin mesh,
without separating off the singularity.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
DN , M 6.698E−01 5.707E−01 4.992E−01 4.994E−01 4.996E−01
P N ,M 0.231 0.193 −0.001 −0.001

Fig. 3. Example 3 from Problem Class 2: The numerical approximation to y and the approximation s + Y to the solution u, with ε = 2−16 and
N = M = 64.

We show now the numerical results when the singularity is stripped off. The maximum two-mesh global differences
associated with the component y and the orders of convergence are given in Table 5. Observe that the numerical results
indicate that the method is globally and uniformly convergent.
Fig. 3 displays both the numerical approximation to the function y defined in (10) and the numerical solution to
problem (8) and (26) is displayed in Fig. 3. The presence of an interior layer is evident in both figures.
14 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

Table 5
Example 3 from Problem Class 2: Maximum two-mesh global differences and orders of convergence for the function y in (10) using a piecewise
uniform Shishkin mesh.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
1.683E−02 8.549E−03 4.277E−03 2.134E−03 1.066E−03
ε=2 0
0.978 0.999 1.003 1.001
6.557E−03 3.177E−03 1.563E−03 7.741E−04 3.852E−04
ε = 2−2
1.045 1.024 1.014 1.007
5.748E−03 2.992E−03 1.527E−03 7.717E−04 3.879E−04
ε = 2−4
0.942 0.970 0.985 0.992
8.330E−03 4.346E−03 2.220E−03 1.122E−03 5.642E−04
ε = 2−6
0.939 0.969 0.984 0.992
9.535E−03 4.981E−03 2.546E−03 1.287E−03 6.469E−04
ε = 2−8
0.937 0.968 0.984 0.992
1.128E−02 5.618E−03 2.804E−03 1.401E−03 6.999E−04
ε = 2−10
1.005 1.003 1.001 1.001
1.245E−02 6.212E−03 3.103E−03 1.551E−03 7.751E−04
ε = 2−12
1.003 1.001 1.001 1.000
1.880E−02 6.559E−03 3.278E−03 1.639E−03 8.194E−04
ε = 2−14
1.519 1.000 1.000 1.000
3.134E−02 1.104E−02 3.335E−03 1.667E−03 8.338E−04
ε = 2−15
1.505 1.727 1.000 1.000
2.964E−02 1.266E−02 4.588E−03 1.689E−03 8.445E−04
ε = 2−16
1.228 1.464 1.442 1.000
. . . . . .
. . . . . .
. . . . . .
2.957E−02 1.264E−02 4.584E−03 1.752E−03 8.755E−04
ε = 2−30
1.226 1.463 1.388 1.001
DN , M 3.134E−02 1.266E−02 4.588E−03 2.134E−03 1.066E−03
P N ,M 1.308 1.464 1.104 1.001

Table 6
Example 4 from Problem Class 3: Maximum two-mesh global differences and orders of convergence for u using a piecewise uniform Shishkin mesh,
without separating off the singularity.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
DN , M 2.500E−01 2.500E−01 2.500E−01 2.500E−01 2.500E−01
P N ,M 0.000 0.000 0.000 0.000

4.3. Problem Class 3

Example 4. Consider the problem (17), with the data taken to be

b(x, t) = 1 + x, f (x, t) = 4(1 + x)(1 − x)t + t 2 , (27)

and
0, if 0 ≤ t ≤ 0.25,
{
u(0, t) =
0.5, if 0.25 < t ≤ 1.
Observe that in this example the function b = b(x). The schemes considered here to approximate the solution are defined
N ,M
on the Shishkin mesh Q̄3 .

Once again, we first confirm the need to use our analytical/numerical approach to approximate the Problem Class
N ,M
3. If we use backward Euler method and standard central finite differences on the Shishkin mesh Q̄3 to approximate
Example 4 without separating off the singularity, it is not globally convergent for any value of ε . By way of illustration,
the uniform two-mesh global differences are given in Table 6.
We show now the numerical results when the singularity is stripped off. The maximum two-mesh global differences
associated with the component y and the orders of convergence are given in Table 7. Observe that the numerical results
indicate that the method is globally parameter-uniformly convergent. Fig. 4 displays the numerical approximation to the
function y defined in (19) and the approximation to the solution of problem (17) and (27). Thin boundary layer regions
near x = 0 and x = 1 are visible in both plots, while large time derivatives near t = 0.25 are visible only in the plot of
the approximation s + Y .
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 15

Table 7
Example 4 from Problem Class 3: Maximum two-mesh global differences and orders of convergence for the function y in (19) using a piecewise
uniform Shishkin mesh.
N = 256, M = 16 N = 512, M = 32 N = 1024, M = 64 N = 2048, M = 128 N = 4096, M = 256
2.765E−03 1.478E−03 7.901E−04 4.057E−04 2.058E−04
ε=2 0
0.903 0.904 0.962 0.979
4.421E−03 2.250E−03 1.136E−03 5.714E−04 2.866E−04
ε = 2−2
0.975 0.985 0.992 0.996
7.926E−03 3.909E−03 1.940E−03 9.664E−04 4.823E−04
ε = 2−4
1.020 1.010 1.005 1.003
1.022E−02 5.079E−03 2.532E−03 1.264E−03 6.315E−04
ε = 2−6
1.009 1.004 1.002 1.001
1.050E−02 5.214E−03 2.598E−03 1.297E−03 6.478E−04
ε = 2−8
1.010 1.005 1.003 1.001
1.059E−02 5.260E−03 2.621E−03 1.308E−03 6.535E−04
ε = 2−10
1.010 1.005 1.003 1.001
1.062E−02 5.275E−03 2.628E−03 1.312E−03 6.553E−04
ε = 2−12
1.010 1.005 1.003 1.001
1.063E−02 5.278E−03 2.630E−03 1.313E−03 6.558E−04
ε = 2−14
1.010 1.005 1.003 1.001
1.063E−02 5.279E−03 2.630E−03 1.313E−03 6.559E−04
ε = 2−16
1.009 1.005 1.003 1.001
. . . . . .
. . . . . .
. . . . . .
1.063E−02 5.279E−03 2.631E−03 1.313E−03 6.559E−04
ε = 2−30
1.010 1.005 1.003 1.001
DN , M 1.063E−02 5.280E−03 2.631E−03 1.313E−03 6.559E−04
P N ,M 1.010 1.005 1.003 1.001

Fig. 4. Example 4 from Problem Class 3: The numerical approximation to y and the approximation s + Y to the solution u, with ε = 2−16 and
N = M = 64.

Appendix A. Compatibility conditions

Below we place certain regularity and compatibility restrictions on the data of the problem

Lu := ut − ε uxx + b(x, t)u = f (x, t), (x, t) ∈ Q , (28a)

u(0, t) = gL (t), u(1, t) = gR (t) t ≥ 0, u(x, 0) = φ (x), 0 < x < 1, (28b)

in order that the solution u ∈ C 4+γ (Q ). Compatibility conditions at the zero-order level correspond to:

φ (0+ ) = gL (0) and φ (1− ) = gR (0). (29a)

Assuming (29a), we can write u = Φ (x, t) + z , (x, t) ∈ Q where

Φ (x, t) := φ (x) + (1 − x)(gL (t) − gL (0)) + x (gR (t) − gR (0)) ;

16 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638

Lz = f − LΦ ; and z(x, t) = 0, (x, t) ∈ ∂ Q . Note that

LΦ = (1 − x)gL′ (t) + xgR′ (t) − εφ ′′ (x) + bΦ .

From [22], if b, f , LΦ ∈ C 0+γ (Q̄ ) and the first-order compatibility conditions

(gR′ (0) − εφ ′′ (1− )) + b(1, 0)φ (1− ) = f (1, 0), (29b)

(gL (0) − εφ (0 )) + b(0, 0)φ (0 ) = f (0, 0),
′ ′′ + +
(29c)

are satisfied, then u ∈ C (Q̄ ). If b, f , LΦ ∈ C

2+γ 2+γ
(Q̄ ) and we further assume second-order compatibility (so that the
mixed derivative zxxt is well defined at (0, 0) and (1, 0)), such that

(f − LΦ )t (0, 0+ ) + (f − LΦ )xx (0+ , 0) = 0; (29d)

(f − LΦ )t (1, 0+ ) + (f − LΦ )xx (1− , 0) = 0, (29e)
then the solution of (28) satisfies u ∈ C 4+γ (Q̄ ).

Appendix B. Properties of s2 (x, t)

Recall that
( )
x
s2 (x, t) := t e
2 −b(0)t
erf √ .
2 εt
Using the inequality t p e−t ≤ Cp,µ e−µt , t ∈ [0, ∞), µ < 1, p > 0; we have the following bounds, for all (x, t) ∈ Q̄ ,

∥ s2 ∥ ≤ C ; (30a)
⏐ ∂ s2 ⏐ ∂ s2
⏐ ⏐ ⏐ 2 ⏐
xt − x2 −µ √ x
⏐ ∂ t (x, t)⏐ + ε ⏐ ∂ x2 (x, t)⏐ ≤ C √εt e 4εt ≤ Cte
⏐ ⏐
⏐ ⏐ ⏐ ⏐ 4ε T ; (30b)
√ ⏐
2 ⏐ ∂ s2
⏐ 4 −b(0)t
x 3 ⏐ − x2
⏐ ⏐
2e t ⏐ 3x
ε ⏐ 4 (x, t)⏐ = ε
⏐ ⏐
⏐ √ ⏐ − ⏐ e 4ε t
∂x 2 ε επ ⏐ 2ε t 4(ε t)2 ⏐
−µ √ x
≤ Ce 4ε T , (30c)
and
∂ 2 s2 x3 (b(0)t)2
( ) ( ) ( )
1 3x 2
− 4xεt x
(x, t) = e−b(0)t √ √ − √ e + erf √ 2 1 − 2b(0)t + e−b(0)t .
∂t2 4 π εt 2ε t ε t 2 εt 2
Hence,
⏐ ∂ s2
⏐ 2 ⏐
, ⏐ ≤ C, x > 0.
⏐
⏐
⏐ ∂t2 (x t)⏐ (30d)

The second order time derivative is bounded, but not continuous, on the closed domain.

References

[1] J. Boyd, N. Flyer, Compatibility conditions for time-dependent partial differential equations and the rate of convergence of Chebyshev and
Fourier spectral methods, Comput. Methods Appl. Mech. Engrg. 175 (3–4) (1999) 281–309.
[2] Q. Chen, Z. Qin, R. Temam, Treatment of incompatible initial and boundary data for parabolic equations in higher dimensions, Math. Comp. 80
(276) (2011) 2071–2096.
[3] N. Flyer, B. Fornberg, Accurate numerical resolution of transients in initial–boundary value problems for the heat equation, J. Comput. Phys.
184 (2) (2003) 526–539.
[4] M.A. Biot, General theory of three-dimensional consolidation, J. Appl. Phys. 12 (1941) 155–164.
[5] M.A. Biot, Theory of elasticity and consolidation for a porous anisotropic solid, J. Appl. Phys. 26 (1955) 182–185.
[6] R.K. Dunne, E. O’Riordan, Interior layers arising in linear singularly perturbed differential equations with discontinuous coefficients, in: I. Farago,
P. Vabishchevich, L. Vulkov (Eds.), Proc. 4th International Conference on Finite Difference Methods: Theory and Applications, Rousse University,
Bulgaria, 2007, pp. 29–38.
[7] J.L. Gracia, E. O’Riordan, A singularly perturbed parabolic problem with a layer in the initial condition, Appl. Math. Comput. 219 (2012) 498–510.
[8] E. O’Riordan, G.I. Shishkin, Singularly perturbed parabolic problems with non-smooth data, J. Comput. Appl. Math. 166 (2004) 233–245.
[9] G.I. Shishkin, A difference scheme for a singularly perturbed equation of parabolic type with discontinuous coefficients and concentrated factors,
Zh. Vychisl. Mat. Mat. Fiz. 29 (1989) 1277–1290.
[10] P.A. Farrell, A.F. Hegarty, J.J.H. Miller, E. O’Riordan, G.I. Shishkin, Robust Computational Techniques for Boundary Layers, CRC Press, 2000.
[11] G.I. Shishkin, Grid approximation of singularly perturbed parabolic reaction–diffusion equations with piecewise smooth initial–boundary
conditions, Math. Model. Anal. 12 (2007) 235–254.
[12] G.I. Shishkin, Grid approximation of singularly perturbed parabolic equations with piecewise continuous initial–boundary conditions, Proc.
Steklov Inst. Math. 2 (2007) S213–S230.
[13] U.Kh. Zhemukhov, Parameter-uniform error estimate for the implicit four-point scheme for a singularly perturbed heat equation with corner
singularities, Transl. Differ. Uravn. 50 (7) (2014) 923–936; Differ. Equ. 50 (7) (2014) 913–926.
 J.L. Gracia and E. O’Riordan / Journal of Computational and Applied Mathematics 370 (2020) 112638 17

[14] U.Kh. Zhemukhov, On the convergence of the numerical solution of an initial–boundary value problem for the heat equation in the presence of
a corner singularity in the derivatives of the solution, Vestnik Moskov. Univ. Ser. XV Vychisl. Mat. Kibernet. 50 (4) (2013) 9–18, (in Russian);
Transl. Mosc. Univ. Comput. Math. Cybernet. 37 (4) (2013) 162–171.
[15] J.J.H. Miller, E. O’Riordan, G.I. Shishkin, L.P. Shishkina, Fitted mesh methods for problems with parabolic boundary layers, Math. Proc. R. Ir.
Acad. 98A (1998) 173–190.
[16] G.I. Shishkin, L.P. Shishkina, Difference Methods for Singular Perturbation Problems, CRC Press, 2009.
[17] J.L. Gracia, E. O’Riordan, A singularly perturbed reaction–diffusion problem with incompatible boundary–initial data, in: I. Dimov, I. Farago, L.
Vulkov (Eds.), Numerical Analysis and its Applications: 5th International Conference, in: Lecture Notes in Computer Science, vol. 8236, Springer,
Heidelberg, 2013, pp. 303–310, Revised Selected Papers.
[18] P.W. Hemker, G.I. Shishkin, Approximation of parabolic PDEs with a discontinuous initial condition, East-West J. Numer. Math. 1 (1993) 287–302.
[19] P.W. Hemker, G.I. Shishkin, Discrete approximation of singularly perturbed parabolic PDEs with a discontinuous initial condition, Comput. Fluid
Dyn. 2 (1994) 375–392.
[20] G.I. Shishkin, A difference scheme for a singularly perturbed equation of parabolic type with a discontinuous initial condition, Dokl. Akad. Nauk
SSSR 300 (1988) 1066–1070.
[21] J.L. Gracia, E.O’. Riordan, Parameter-uniform numerical methods for singularly perturbed parabolic problems with incompatible boundary–initial
data, Appl. Numer. Math. 146 (2019) 436–451.
[22] O.A. Ladyzhenskaya, V.A. Solonnikov, N.N. Ural’tseva, Linear and Quasilinear Equations of Parabolic Type, in: Transactions of Mathematical
Monographs, vol. 23, American Mathematical Society, 1968.