Return to Article Details A numerical approach for singularly perturbed parabolic reaction-diffusion problem on a modified graded mesh

A numerical approach for
singularly perturbed parabolic reaction-diffusion
problem on a modified graded mesh

Kishun Kumar Sah and Subramaniam Gowrisankar
(Date: November 11, 2024; accepted: March 06, 2025; published online: June 30, 2025.)
Abstract.

This paper addresses the numerical approximations of solutions for one dimensional parabolic singularly perturbed problems of reaction-diffusion type. The solution of this class of problems exhibit boundary layers on both sides of the domain. The proposed numerical method involves combining the backward Euler method on a uniform mesh for temporal discretization and an upwind finite difference scheme for spatial discretization on a modified graded mesh. The numerical solutions presented here are calculated using a modified graded mesh and the error bounds are rigorously assessed within the discrete maximum norm. The primary focus of this study is to underscore the crucial importance of utilizing a modified graded mesh to enhance the order of convergence in numerical solutions. The method demonstrates uniform convergence, with first-order accuracy in time and nearly second-order accuracy in space concerning the perturbation parameter. Theoretical findings are supported by numerical results presented in the paper.

Key words and phrases:
Singular perturbation problem, Parabolic reaction-diffusion problems, Finite difference methods, Modified graded mesh, Boundary layers, Uniform convergence
2005 Mathematics Subject Classification:
65M06, 65M12, 65M15.
Department of Mathematics, National Institute of Technology Patna, Patna - 800005, India, e-mail: kishuns.phd20.ma@nitp.ac.in.
Department of Mathematics, National Institute of Technology Patna, Patna - 800005, India, e-mail: s.gowri@nitp.ac.in.

1. Introduction

Singular perturbation problem consists of small parameter associated with highest oredr derivative in the equation. Presence of the parameter will cause the distortion in the solution in the vicinity of the boundary. Such problems have boundary or interior layers in their solutions. The interior layer develops when there is a discontinuity in the provided data, and the boundary layer happens when a term containing the highest order derivative is multiplied by a singular perturbation parameter. Furthermore, the problem is stiff due to the simultaneous presence of a discontinuous coefficient and a delay parameter, and the solution shows multi-scale character as ε0. Many physical and real world problems are followed by singularly perturbed parabolic problem. In [7], its applications to many field of interests have been described. One can refer the book [2] for the application of the problem in medicine while, [11, 13] are dealing with the applications of such problems in mechanics. There are many articles in the literature that are imerged as a result of convergence of numerical schemes applied to singularly perturbed problem of parabolic type, e.g., one can refer to the articles [7, 3, 4, 5].

Finite difference method (FDM) is a sort of numerical technique that approximates the derivatives in a differential equation by replacing them with finite differences, calculated over a discrete grid of points, to solve for the unknown function or solution. There are a lots of beauties of this methods including

  • Discretization: FDM discretizes the continuous domain into a grid of points, allowing for numerical approximation.

  • Approximation of derivatives: FDM approximates derivatives using finite differences, such as forward, backward, or central differences.

  • Local accuracy: FDM has high local accuracy, meaning it’s accurate in small neighbourhood around each grid point.

  • Global convergence: FDM can converge to the exact solution as the grid size decreases.

  • Easy to implement: FDM is relatively simple to understand and implement, especially for simple geometries.

  • Computational efficiency: FDM can be computationally efficient, especially for large-scale problems.

  • Flexibility: FDM can be applied to various types of differential equations, including non-linear and time-dependent equations.

However, certain limitations to the method are grid dependence, numerical diffusion and boundary treatment. Overall, the finite difference method is a powerful tool for solving differential equations, offering a good balance between accuracy, efficiency, and simplicity.

The manuscripts that are utilizing the methods to obtain accuracy of discrtized solution include [10, 15, 18, 6]. Different variants of FDMs have been utilized to get the accuracy and convergence of the problem with different conditions and restrictions. [8] is the study of convection–reaction problem in which space direction is discretized by Shishkin mesh while Crank-Nikolson is adapted to discrtize in time direction. As a result, second order convergence in time domain and higher order convergence in space domain are gained, respectively. except this [9] is the study of the discrtization of convection-reaction problem using Euler technique in time domain and a central difference method comprises with space domain. Almost second order convergence upto logarithmic term is shown as a convergence result.

Throughing these manuscripts, we have considered the singularly perturbed parabolic initial-boundary-value problem (IBVP):

(1) {uθ(r,θ)+𝔏εu(r,θ)=f(r,θ),(r,θ)Π=Υr×Υθ,u(r,θ)=ψb(r),onΥr,u(r,θ)=ψr(θ),{1}×Λr={(1,θ),0<θ𝒯},u(r,θ)=ψl(θ),{0}×Λl={(0,θ),0<θ𝒯},

where,

(2) 𝔏εuεurr+b(r)u

0<ε1 is a small parameter, b and f are sufficiently smooth functions with b(r)δ>0 for rΥ¯r=[0,1]. Here, Π = Υr×Υθ, where Υr=(0,1) and Υθ=(0,𝒯] and also Λ=ΛlΛbΛr. The above considered singular perturbation parameter is ε(0,1] and also considered function b(r), f(r,θ), ψb(r), ψl(θ) and ψr(θ) are sufficiently smooth, bounded and also satisfy the condition b(r)0. The reduced problem corresponding to (1) is

(3) {u(r,θ)θ+b(r)u(r,θ)=f(r,θ),(r,θ)Π,u(r,θ)=ψb(r),rΛb.

As (1) is characterized by boundary layers, uniform meshes do not work properly [1], which results to the construction of non-uniform and layer adapted meshes. In recent days, layer adapted meshes are better in handling the problems possess boundary and interior layers. [3, 14, 17] are the study about the convergence of finite difference methods over Shishkin mesh and other layer adapted meshes for reaction-convection equations.

In the present study, we have adopted the modified graded mesh for spatial domain. simple upwinding are introduced for discretizing the problem and uniform grid is taken place as domain discretization in time domain. As a consequence, almost second order convergence upto logarithimic factor we have achieved in spactial direction as a convergence rate of numerical solution which is optimal. Finally, we have presented two test problems to validate our theretical conclusions.

The rest of the manuscripts are adored in the following manner: Section 2 is a brief study of the analytical behavior of the problem, including an introduction to the appropriate Hölder spaces. The discussion includes the maximum principle for the differential operator, highlighting its role in ensuring ε-uniform stability. We also present sufficient compatibility conditions on the initial and boundary data to ensure the existence, uniqueness, and appropriate regularity of the problem’s solutions. In Theorem 2, both classical and new sharper ε-uniform bounds in the maximum norm on the derivative of the solution are discussed. The latter are obtained by means of a new decomposition of the solution, which leads to a deceptively simple proof of the required results. We generated the modified graded mesh and their properties are contained in Section 3. The upwind finite difference scheme is constructed in Section 4. We have completed error analysis for the presented method in Section 5. In Section 6, the numerical results are reported, which validate the results predicted by the theory and in fact show that the numerical methods work equally well in practice for a much broader class of problems than the theory predicts. Section 7 ends with the conclusion.

Notation: 𝒞 is used as a generic constant throughout the paper and also, 𝒞 is independent of the perturbation parameter ε and the mesh point 𝒩 and we use discrete maximum norm to study convergence which is defined as,

uΠ=maxrΠ|u(r)|.

2. Analytical behaviour of the continuous problem

To explore the consistency of solutions to the time-dependent problem under consideration, we introduce certain function spaces characterized by Hölder continuity in both spatial and temporal dimensions. Specifically, consider Υ and Π be a convex domain in Υ¯×[0,T]. Again we suppose that κ and fulfilled the condition 0<κ1. Then a function y is said to be Hölder continuous in Π of degree κ if, for all (r,θ),(r,θ)Π,

|u(r,θ)u(r,θ)|𝒞(|rr|2+|θθ|)κ/2.

Notice the distinction in the metrics employed for the spatial and temporal variables. The collection of all Holder continuous functions constitutes a normed linear vector space denoted by 𝒞κ0(Π), equipped with the norm

uκ,Π=uΠ+sup(r,θ),(r,θ)Π|u(r,θ)u(r,θ)|(|rr|2+|θθ|)κ/2,

where

uΠ=sup(r,θ)Π|u(r,θ)|.

We introduce the subspace 𝒞κp(Π) of 𝒞κ0(Π) for each integer p1. These are functions within 𝒞κ0(Π) that possess Hölder continuous derivatives. and also introduced

𝒞κp(Π)={u:i+juriθj𝒞κ0(Π)for all positive integers i and j with 0i+2jp}.

The norm on 𝒞κp(Π) is taken to be

up,κ,Π=max0i+2jpi+juriθjκ,Π.

Observe once more the distinct handling of space and time derivatives. For u𝒞κp(Π) and 0qp, we also define the following semi-norms:

uq,κ,Π=maxi+2jqi+juriθjκ,Π.

From these definitions, it is clear that

up,κ,Π=max0qpuq,κ,Π

Adopting the notational convention u0,κ,Π=|u|0,κ,Π=uκ,Π, when the domain is apparent or of no specific importance, Π is typically omitted.

Furthermore, the initial functions ψb(r), are essential to meet compatibility conditions at corner points of the domain, namely at (0,0) and (1,0). These points significantly represents the boundaries or changes in the boundary conditions, satisfying compatibility conditions at these points ensures a consistency where the boundary or initial conditions may change sharply. The required compatibility conditions at the corners are stated below,

ψb(0) =ψl(0),
ψb(1) =ψr(0),
dψl(0)dθεd2ψb(0)dr2+b(0)ψb(0) =f(0,0),
dψr(0)dθεd2ψb(1)dr2+b(1)ψb(1) =f(1,0).

Under the above compatibility condition, (1) possess unique solution with possibility of layer at both the end points r=0 and r=1.

Minimum Principle. Assume that b𝒞0(Π¯) and let ψ𝒞2(Π)𝒞0(Π¯) also suppose that ψ0 on Λ then 𝔏εψ0 in Π implies that ψ0 in Π¯.

Proof.

Let us assume that for all (r,θ)Λ, we have ψ(r,θ)0, and for all (r,θ)Λ, 𝔏εψ(r,θ)0. Now, we apply the differential operator 𝔏ε on ψ. To contradict we assume that there exists a point (r0,θ0)Π such that ψ(r0,θ0)<0, and ψ(r0,θ0) is a minimum of ψ in Π. Since, we have assumed that ψ(r0,θ0) minimum and also ψ is differentiable, then by applying a second order differential operator 𝔏ε, on ψ shows that it has upward concavity, which contradicts our supposition that ψ(r0,θ0)<0. Also, if ψ attain its minimum on Π¯, then by assumption that ψ(r,θ)0 for all (r,θ)Π and hence on Π¯, ψ must also be non-negative. ∎

Ensuring the stability of 𝔏ε and establishing an ε-uniform bound for the solution (1) in the maximum norm naturally follows from this.

Theorem 1.

Consider any function 𝒮 within the domain of the differential operator 𝔏ε in (1). Then

𝒮(1+α𝒯)max{𝔏ε𝒮,𝒮Λ},

and any solution u(r,θ) of (1) has the ε-uniform upper bound

u(r,θ)(1+α𝒯)max{f,φ},

where α=maxΠ¯{0,(1b)}1/δ.

The subsequent classical theorem provides adequate conditions for guaranteeing the existence of a unique solution.

Theorem 2.

Suppose that φ=0, the data b and f 𝒞λ0(Π¯), and that the compatibility conditions

f(0,0)=f(1,0)=0

are fulfilled. then (1) has a unique solution u and u𝒞λ2(Π¯).

sectionBounds on the solution and its derivatives

The proposed method’s error estimate, discussed below, is validated assuming that the solution of (1) exhibits a higher level of regularity than guaranteed by Theorem 4. Achieving this heightened regularity necessitates the imposition of stronger compatibility conditions at the two corners of Π¯. The additional compatibility conditions are

(4) (θ+ε2r2)(f)(0,0)=0and(θ+ε2r2)(f)(1,0)=0

Note that these conditions necessitate additional smoothness of f. The subsequent theorem establishes the existence of a smooth solution for the problem with homogeneous boundary conditions.

Theorem 3.

Assume that φ=0, the data b and f𝒞λ2(Π), and that the compatibility conditions

f(0,0)=f(1,0)=0

and

(θ+ε2r2)(f)(0,0)=(θ+ε2r2)(f)(1,0)=0

are fulfilled. Then (1) has a unique solution u and u𝒞λ4(Π). Furthermore, the derivatives of the solution u satisfy, for all non-negative integers i,j, such that 0i+2j4,

i+juεriθjΠ𝒞εi/2,

where the constant 𝒞 is independent of ε.

Proof.

The first part’s proof is provided in [12]. Bounds on the derivatives are acquired through the following process: by transforming the variable r into the stretched variable r~=r/ε, problem (1) becomes problem

(5) {2u~r~2+b~u~+u~θ=f~onΠ~εu~=0,

where Π~ε=(0,1/ε)×(0,𝒯]. The differential equation in (5) is not dependent on ε. Utilizing the estimate (10.5) from [12], it is deduced that for all non-negative integers i,j such that 0i+2j4, and all 𝒩~δ in Π~ε,

i+ju~r~iθj𝒩~δ𝒞(1+u~𝒩~2δ).

Hence, we used the Theorem 1 which bound on the uε. ∎

The limits on the derivatives of the solution, as provided in Theorem 3, were initially deduced from classical findings. Nevertheless, it has been discovered that they lack sufficiency for proving the ε-uniform error estimate. To address this, more robust boundaries on these derivatives are acquired through a methodology initially presented in [16]. A pivotal maneuver involves breaking down the solution,uε, into components of smoothness and singularity.

Suppose that the solution u of equation (1) and write in the form as

(6) uε=ε+𝒮ε,

where ε and 𝒮ε are smooth and singular components of uε .

Now, again decomposed the smooth component ε in the form such that

ε=0+ε1,

where the define the component 0 and 1 in the form as

b0+0θ=finΠ,0=0onΛ¯b,
𝔏ε1=20r2inΠ,1=0onΛ.

Therefore, it is evident that 0 represents the solution to the reduced problem and moreover, ε satisfies

𝔏εε=fandε=0onΛlΛr.

With ε defined in this manner, it follows that 𝒮ε is determined and satisfies

𝔏ε𝒮ε=0inΠ𝒮ε=0onΠand𝒮ε=0onΠ¯.

Now, we can also easily to write of 𝒮ε such that

𝒮ε=𝒮l+𝒮r,

where 𝒮l and 𝒮r are define by

{𝔏ε𝒮l=0inΠ,𝔏ε𝒮r=0inΠ
{𝒮l=0onΛl,𝒮r=0onΛr,
{𝒮l=0onΛbΛr,𝒮r=0onΛlΛb

It’s evident that 𝒮l and 𝒮r denote the boundary layers on Λl and Λr , respectively. The theorem following contains the necessary non-classical bounds on ε and 𝒮ε, along with their derivatives.

Theorem 4.

Let’s consider problem (1). We assume that the data b𝒞λ2(Π¯) and f𝒞λ4(Π¯), and that the compatibility conditions of the previous theorem are met. Under these conditions, the reduced solution 0 exists and belongs to 𝒞λ4(Π¯). Additionally, if the supplementary compatibility conditions are satisfied, then

(7) 2f(0,0)r2=2f(1,0)r2=0

and

(8) f(0,0)θ=f(1,0)θ=0

are satisfied then 1 and 𝒮ε exists and 1,𝒮ε𝒞λ4(Π¯). Also, for all positive integers i,j, such that 0i+2j4

i+jεriθjΠ¯𝒞(1+ε1i/2),

and for all (r,θ)Π,

i+j𝒮l(r,θ)riθj𝒞εi/2er/ε

and

i+jr(r,θ)riθj𝒞εi/2e(1r)/ε,

where 𝒞 is a constant independent of ε.

Proof.

References for the existence and regularity results can be found in [12]. The proofs of the bounds on the functions and their derivatives are presented subsequently.

The reduced solution, denoted as 0, emerges as the solution to a first-order differential equation. Employing a classical argument, we arrive at an estimate denoted as

(9) i+j0riθjΠ¯𝒞.

Moreover, the function 1 represents the solution to a problem that conforms to the conditions specified in Theorem 3. Consequently, it follows that

(10) i+j1riθjΠ¯𝒞εi/2.

Since

i+jεriθj=i+j0riθj+εi+j1riθj,

The necessary estimates for the smooth component ε and its derivatives can be obtained by employing equations (9) and (10). Similarly, the necessary bounds on 𝒮l and 𝒮r, as well as their derivatives, are derived in a similar manner.

Therefore, to simplify the proof, we will focus solely on 𝒮l and its derivatives. To establish a bound on 𝒮l, we begin by defining

ψ±(r,θ)=𝒞er/εeαθ±𝒮l(r,θ).

Then, if 𝒞 is chosen sufficiently large and α0,

ψ±(r,0)=𝒞er/ε0,
ψ±(0,θ)=𝒞eαθ±00,
ψ±(1,θ)=𝒞e1/εeαθ0

and

𝔏εψ±(r,θ)=𝒞(b1+α)er/εeαθ0

if the value of α is chosen as in Theorem 1 to be α=maxΠ¯{0,(1b)}. It follows from the maximum principle that for all (r,θ)Π¯

|𝒮l(r,θ)|𝒞er/εeαθ𝒞er/ε

as required. ∎

3. The modified graded mesh and mesh discretization

In this section, we have constructed a modified graded mesh for the interval [0,1] to address the reaction-diffusion parabolic problem, incorporating the effects of the perturbation parameter associated with boundary layers on both sides of the domain-specifically, at r=0 (left) and r=1 (right). The modified graded mesh is generated using a piecewise-defined function within the given interval [0,1]. The subintervals [0,ε] and [1ε,ε] represent the finer regions in the spatial direction, forming a uniform grid divided into 𝒩/4 segments. In contrast, the interval [ε,1ε] constitutes the coarser region of the graded mesh, characterized by a non-uniform grid obtained through the non-linear equation (12) and subdivided into 𝒩/2 segments. The nonlinear equation (12) itself is derived using the bisection method.

Refer to caption
Figure 1. Modified graded mesh in the spatial direction for 𝒩=32 and ε=101.
(11) {η0=0,ηi=4εi/𝒩,i=1,2,,𝒩/4ηi+1=(1+σh)ηi,i=𝒩/4,,(𝒩/2)2η𝒩/2=1/2,ηi=1η𝒩i,i=(𝒩/2)+1,,𝒩.

These equations establish a piecewise function for ηi, where η0 is explicitly defined as 0, η𝒩/2 is explicitly set to 1/2, and the rest of the values are determined based on the specified conditions. This sequence of values for ηi varies according to the ranges of i, utilizing a non-linear equation involving the parameter h and satisfies the following non-linear equation such that

(12) ln(1/ε)=(𝒩/2)ln(1+σh).

This technique guarantees a precise arrangement of grid points across distinct subintervals within the range [0,1]. Within each subinterval [0,ε] and [1ε,1],𝒩/4 points are evenly distributed with a step length of 4ε/N. Here, 𝒩 denotes the total grid point count, and ε represents a small positive value indicating the width of boundary regions. The selection of h within the central subinterval [ε,1ε] is established iteratively through a non-linear equation (12). The mesh length is denoted by hi=ηiηi1 for i=1,2,,𝒩.

Remark 5.

The proposed modified graded mesh adheres to the following bounds for mesh size.

hi={4ε/𝒩fori=1,,𝒩/4ρhηi1fori=𝒩/4+1,𝒩/2+2,,3𝒩/44ε/𝒩fori=3𝒩/4,3𝒩/4+1,,𝒩.
Lemma 6.

The mesh described in equation (11) fulfills the subsequent estimates.

|hi+1hi|{𝒞h2fori=𝒩/2+1,𝒩/2+2,,3𝒩/4,0otherwise.
Proof.

Initially, we investigate the results for i=0,1,2,,𝒩/2. No further establishment is required since the mesh is uniform in this segment.

Now, we have described the following for i=𝒩/2+1,𝒩/2+2,,𝒩 such that

|hi+1hi| =|ρhηiρhηi1|,
=ρh|ηiηi1|,
=ρ2h2ηi1,
𝒞h2.

Here, we have taken 0<ρ,h<1. ∎

Lemma 7.

The parameter h for the modified graded mesh defined in Equation (11) adheres to the following bound:

h𝒞𝒩1ln(1/ε).
Proof.

In the mesh partition defined in Equation (11), let Q1 and Q3 denote the number of points where the mesh spacing remains constant. It is clear that Q1 andQ3 are bounded above by C/h. Conversely, Q2 represents the count of points within the same partition where the mesh steps vary, indicating a non-uniform mesh size. Additionally, we consider η𝒩/2+1, identified as the minimum point for which ηi>ε. Now, we aim to derive an upper limit for Q2. Under the assumption that ρh1, we have

Q2 =𝒩/2+1𝒩1=𝒩/2+1𝒩(ηi+1ηi)1ηiηi+1𝑑η,
=𝒩/2+1𝒩(hi+1)1ηiηi+1𝑑η,
=𝒩/2+1𝒩(hρηi)1ηiηi+1𝑑η,
𝒩/2+1𝒩(2/ρhηi+1)1ηjηi+1𝑑η,

because ηi+1<2ηi. For any η[ηi,ηi+1], we have

Q2 𝒩/2+1𝒩2(ρh)1ηiηi+11η𝑑η,
2(ρh)1ε11η𝑑η,
2(ρh)1ln(1/ε).

Recalling 𝒩=Q1+Q2, we have

𝒩 𝒞/ρh+2(ρh)1ln(1/ε),
𝒩 1/h(ρ𝒞+2ρln(1/ε)),
𝒩 1/h(𝒞ln(1/ε)),

Finally, we get

h𝒞𝒩1ln(1/ε),

where 𝒩 represents grid points in the r-direction. ∎

Remark 8.

From Lemma 6 and Lemma 7 it is clear that the modified graded mesh satisfies

|hi+1hi|CN2ln2(1/ε)

.

3.1. Temporal Discretization

We present mesh with uniform step length in the time direction since layer phenomenon has no effect on the temporal variable θ. The following notation and definition apply to the uniform mesh in the time direction:

Υ={θs=sΔθ,s=0,1,,M,Δθ=𝒯}.

Here represents the grid points in the temporal direction.

4. Discretization method

Now, we proceed to discretize our domain systematically. We construct a non-uniform grid Υr𝒩 on Υr, consisting of 𝒩 mesh points. This grid is formed by placing 𝒩/2 uniform mesh points within the layer part and 𝒩/2 mesh points outside the layer part. Additionally, we consider an equidistant grid Υ on [0,𝒯] with a uniform step length of Δθ. The discretized domain is then defined as:

Π𝒩=Υr𝒩×Υ,

We now discretize our problem using the aforementioned discretized mesh. We utilize an upwind difference method for spatial derivatives and employ backward Euler for temporal derivatives.

(13) {(DθεDr+Drhi^+b(i))Ui,j+1=fi,j+1,i=1,2,,𝒩1,U0,j+1=ψl(θj+1),U𝒩,j+1=ψr(θj+1),Ui,j=ψb(ri,θj)i=1,2,,𝒩1,

by arranging (13), we will get a tri-diagonal system

(14) (αi,j+1)Ui1,j+1+(βi,j+1)Ui,j+1+(γi,j+1)Ui+1,j+1=hi,j,

where

{αi,j+1=εΔθhih^i,βi,j+1=1+εΔthi+1h^i+εΔthih^i+biΔθ,γi,j+1=εΔθhi+1h^i,hi,j=fi,j+1Δθ+Ui,j,h^i=hi+1+hi2,

we defined the following parameter of finite difference scheme forward, backward,and central and the mesh function v~(ri,θj)=v~i,j such that

Dr+v~i,j=v~i+1,jv~i,jhi+1,Drv~i,j=v~i,jv~i1,jhi,
δr2v~i,j=(Dr+Dr)v~i,jh^i,Dθv~i,j=v~i,jv~i,j1hi+1.

4.1. Numerical algorithm

The following algorithm provides the grid construction and the corresponding numerical solution:

Step 1. Given the number of mesh points in the temporal and the spatial direction, and 𝒩, respectively, take uniform mesh points in the temporal direction, Υj=0 .

Step 2. For the finer part in the spatial direction (i.e.,[0,ε],[1ε,1]), we have the uniform mesh {ηi}i=0𝒩/4.

Step 3. For the coarser part (i.e.,[ε,1ε]), the graded mesh parameter h is obtained by solving the nonlinear equation (12) by the bisection method.

Step 4. Using the graded mesh parameter h, obtain the graded mesh in the interval [ε,1ε] from (11).

Step 5. Set j=1.

Step 6. For the value of j, solve the tridiagonal system (14) to obtain the solution for the time level t=j.

Step 7. j=j+1 goto Step 6.

Step 8. If j=, then stop and mark Ui,j, as the required solution.

5. Error convergence

Lemma 9.

(Discrete maximum principle) Suppose that ψ(ri,θj) is the mesh function which satisfies ψ(ri,θj)0 on Λ𝒩. If 𝔏ε𝒩ψ(ri,θj)0 on (ri,θj)Π𝒩, then ψ(ri,θj)0 on Π¯𝒩.

Proof.

Let us assume that for all (ri,θj)Λ𝒩, we have ψ(ri,θj)0,and for all (ri,θj)Π𝒩,𝔏ε𝒩ψ(ri,θj)0. To contradict we assume that there exists a point (r0,θ0)ΠN such that ψ(r0,θ0)<0, and ψ(r0,θ0) is a minimum of ψ in ΠN.

Since, we have assumed that ψ(r0,θ0) minimum and also ψ is differentiable, then by applying a second order differential operator 𝔏ε𝒩 on ψ shows that it has upward concavity, which contradicts our supposition that ψ(r0,θ0)<0. Also, if ψ attain its minimum on Π¯N, then by assumption that ψ(ri,θj)0 for all (ri,θj)Λ𝒩 and hence on Π¯N,ψ must also be non-negative. ∎

Lemma 10.

For any solution U(ri,θj) of (13), we have

(15) U(ri,θj)(1+α𝒯)max(𝔏ε𝒩,ψΠ𝒩)
Proof.

By constructing the barrier function

(16) U^±(ri,θj)=(1+α𝒯)max(𝔏ε𝒩,ψΠ𝒩)±U(ri,θj).

Now, we can obtained the result (15) with apply the Lemma 9 on the result which is defined in (16). ∎

Theorem 11.

Assume that u and U be the solution of continuous problem (1) and discretized problem (13), respectively. Furthermore, both the solutions meet the corners compatibility requirements. Then, the error estimate is given by

(17) max|(uU)(ri,θj)|𝒞[Δθ+𝒩2ln2(1/ε)],(ri,θj)Π𝒩.

Here, constant 𝒞 is free from 𝒩, Δθ and ε.

Proof.

We consider the following SPPDEs,

(18) (θε2r2+b(r))u(r,θ)=f(r,θ),
u(r,θ)=u(r,θl),rΛb
u(0,θ)=ψ0(θ),u(1,θ)=ψ1(θ),θ[0,𝒯]

To obtain the numerical solution, we discretize equation (18) employing the upwind finite difference method for spatial derivatives and the backward-Euler scheme for the time derivative.

(19) {𝔏r𝒩U(ri,θj)DθUi,jεδr2Ui,j+biUi,j=f(ri,θj),(ri,θj)Π,U(ri,θj)=U1(ri,θj),(ri,θj)Λb,U(0,θj)=ψ0(θj),U(1,θj)=ψ1(θj),

where U1(ri,θj) is the approximate solution obtained over the interval Π𝒩.

Presently, we partition the solution u of equation (1) into its regular and singular constituents, denoted as u=y^+z^. Additionally, we decompose y^ as y^=y0+εy1, where y0 represents the solution to the reduced problem.

(θ+b(r,θ))y0(r,θ)=f(r,θ),(r,θ)Π𝒩.
y0(r,θ)=u(r,θ),
y0(0,θ)=u(0,θ),

and

𝔏r𝒩y1(r,θ)=2y0(r,θ)r2
y1(r,θ)=0,(r,θ)Λb𝒩
y1(0,θ)=y1(1,θ)=0,(r,θ)ΛN.

Further y^ satisfies

𝔏r𝒩y^(r,θ)=f(r,θ),(r,θ)Π𝒩
y^(r,θ)=u(r,θ),(r,θ)Λb𝒩
y^(0,θ)=y0(0,θ),y^(1,θ)=y0(1,θ).

and the singular components z satisfies

𝔏r𝒩z^(r,θ)=0,(r,θ)Π𝒩
z^(r,θ)=0,(r,θ)Λb𝒩
z^(1,θ)=0,z^(0,θ)=ψl(θ)y0(0,θ),onΛN.

We now decompose the numerical solution U of equation (19) into U=Y+Z, where Z denotes the singular component of the decomposition, and Y represents the regular component, which is the solution to the subsequent non-homogeneous problem

𝔏r𝒩Y=f(ri,θj),(ri,θj)Π𝒩
Y(ri,θj)=U1(ri,θj),(ri,θj)Λb𝒩
Y(0,θj)=y^(0,θj),Y(1,θj)=y^(1,θj),

and the singular component Z must satisfy,

𝔏r𝒩Z=0,(ri,θj)Π𝒩
Z(ri,θj)=0,(ri,θj)Π𝒩
Z(1,θj)=0,Z(0,θj)=ψl(θj)y^(0,θj)onΛN.

Thus, the error can be expressed as:

Uu=(Yy^)+(Zz^)

We will now derive the bounds for the smooth and the layer components. To establish the bound for the smooth component, we employ a classical argument. The smooth error component can be written as:

𝔏r𝒩(Yy^) =f(ri,θj)𝔏r𝒩y^
=U(ri,θj)+𝔏r𝒩y^

thus we obtain

𝔏r𝒩(Yy^)=ε(2r2δr2)y^+(θδθ)y^.

When taking the absolute values on both sides and applying (17), the preceding inequality becomes,

|𝔏r𝒩(Yy^)|𝒞(Δθ+𝒩2ln2(1/ε))+ε|(2r2δr2)y^|+|(θδθ)y^|,

also we used the Taylor series expansion such that

𝒞(Δθ+𝒩2ln2(1/ε))+ε12(ri+1ri1)24y^r4+(θjθj1)22y^θ2.

On using Remark 8, estimates of derivatives, bounds of mesh length and discrete maximum principle, we have

(20) |(Yy^)(ri,θj)|𝒞[Δθ+𝒩2ln2(1/ε)].

Similar to the continuous variable z, the discrete counterpart Z is utilized to estimate the singular component, we have

𝔏r𝒩Z=0,θj)Π𝒩
Z(ri,θj)=0,
Z(1,θj)=0,Z(0,θj)=ψl(θj)y^(0,θj).

The singular component error can be expressed as,

𝔏r𝒩(Zz^) =𝔏r𝒩Z𝔏r𝒩z^,
=ε(2r2δr2)z^+(θδθ)z^

then the classical estimates gives

|𝔏r𝒩(Zz^)(ri,θj)|𝒞[Δθ+𝒩2ln2(1/ε)]

The discrete maximum principle is satisfied by the operator 𝔏r𝒩 and also, due to the uniform boundedness of the inverse operator, the above inequality reduces to,

(21) |(Zz^)(ri,θj)|𝒞[Δθ+𝒩2ln2(1/ε)]

Combining (20) and (21) completes the proof on prescribed domain. ∎

6. Numerical Experiments

This section contains two examples featuring boundary layers to demonstrate the discussed numerical method. Through tables and graphs, we present the efficiency of the numerical approach. All numerical computations are conducted with ρ=0.9. The numerical outcomes validate our theoretical conclusions.

Example 12.

Consider the following parabolic IBVP:

(22) {(u/θ)ε2u/r22u(r,θ)=f(r,θ),(r,θ)(0,1)×(0,1],u(r,0)=u0(r),r1u(0,θ)=exp(θ),u(1,θ)=exp((θ+1/ε),0θ1.

We select the initial data u0(r,θ) and the source function f(r,θ) to match the exact solution u(r,θ)=exp((θ+r/ε)). Moreover, we define the maximum point-wise error eε𝒩,Δθ corresponding order of convergence pεN,Δθ for each ε as:

eε𝒩,Δθ= max|(uU)(ri,θj)|,(ri,θj)Π𝒩,
pεN,Δθ= log(eN,Δθ/e2N,Δθ/2)log2.

where u and U are exact and approximate solution respectively. We compute the maximum point-wise error eε𝒩,Δθ and the order of convergence pεN,Δθ for Example 12 using a modified graded mesh and Shishkin mesh, which includes boundary layers at r=0 and r=1 on both the left and right sides of the domain. We apply the upwind finite difference scheme and summarize the results in Table 1 and Table 2. Our observation from the results in Table 1 and Table 2 is that the scheme exhibits second-order convergence. Additionally, in the log-log plot shown in Fig. 2, it is evident that the maximum point-wise error of the proposed numerical method exhibits a second-order convergence with the perturbation parameter ε in space.

Number of Intervals N/Δθ
ε 128/110 256/120 512/140 1024/180
24 1.5547E03 3.9188E04 9.8239E05 2.4586E05
1.9882 1.9960 1.9985
26 1.4792E03 3.7299E04 9.3163E05 2.3317E05
1.9876 2.0013 1.9983
28 1.4000E03 3.5012E04 8.7423E05 2.1836E05
1.9995 2.0018 2.0013
210 1.2888E03 3.2081E04 7.9953E05 1.9953E05
2.0062 2.0045 2.0025
eε𝒩,Δθ 1.2888E03 3.2081E04 7.9953E05 1.9953E05
pεN,Δθ 2.0062 2.0045 2.0025
Table 1. Maximum point-wise errors and order of convergence on a modified graded mesh for Example 12.

In order to compare our results with the well established Shishkin mesh [19], we define the the Shishkin mesh for the problem considered as follows. The Shishkin mesh is a piecewise uniform mesh, depends on two transition points which are defined by means of the transition parameter.

(23) σ=min{1/4,σ0εln𝒩},

where σ0 is a constant to be fixed later. A uniform mesh is placed in [0,σ], [σ,1σ] and [1σ,1], such that r0=0,r𝒩/4=σ, r3𝒩/4=1σ, and r𝒩=1. Therefore the mesh points are given by ri=4iσ/𝒩, for i=0,1,,𝒩/4, ri=σ+2(i𝒩)(12σ)/𝒩, for i=𝒩/4+1,,3𝒩/4, and ri=1σ+4(i3𝒩/4)σ/𝒩, for i=3𝒩/4+1,,𝒩.

Number of Intervals N/Δθ
ε 128/110 256/120 512/140 1024/180
24 1.6096E03 4.0509E04 1.0144E04 2.5371E05
1.9903 1.9975 1.9993
26 1.6420E03 4.1335E04 1.0352E04 2.5891E05
1.9900 1.9974 1.9993
28 1.7637E03 4.4458E04 1.1138E04 2.7862E05
1.9881 1.9969 1.9991
210 1.8389E03 4.8298E04 1.2641E04 3.3124E05
1.9288 1.9339 1.9321
eε𝒩,Δθ 1.8389E03 4.8298E04 1.2641E04 3.3124E05
pεN,Δθ 1.9288 1.9339 1.9321
Table 2. Maximum point-wise errors and order of convergence on a Shishkin mesh for Example 12.
Example 13.

Consider the following problem

uθεurr+((1+r2)/2)u=eθ1+sin(πr),(r,θ)(0,1)×(0,1],

with homogeneous initial and boundary conditions.

In Example 13, an exact solution is unavailable. Consequently, to assess the maximum point-wise error and determine the order of convergence for the computed solution, we employ the double mesh principle. This involves considering 2N intervals in the spatial direction and 2M intervals in the temporal direction. We obtain the numerical solution U~(ri,θj) on the mesh Π¯2N=Υ¯r2N×Υ¯2M, where i=1,2,,N. Notably, each ith point of the mesh ΥrN coincides with the 2ith point of the mesh Υ¯r2N.

The maximum point-wise error for each ε is defined by,

eεN,Δθ=max|(UU~)(ri,θj)|,(ri,θj)ΠN,

and the order of convergence is given by

pεN,Δθ=log(eN,Δθ/e2N,Δθ/2)log2.

Similarly, we compute the maximum point-wise error eε𝒩,Δθ and the order of convergence pεN,Δθ for Example 13 using a modified graded mesh and Shishkin mesh, which includes boundary layers at r=0 and r=1 on both the left and right sides of the domain. We apply the upwind finite difference scheme and summarize the results in Table 3 and Table 4. Our observation from the results in Table 3 and Table 4. is that the scheme exhibits second-order convergence with the perturbation parameter ε in space. Additionally, in the log-log plot shown in Fig. 2, it is evident that the maximum point-wise error of the proposed numerical method exhibits a second-order uniform convergence in the perturbation parameter ε in space.

Number of Intervals N/Δθ
ε 128/110 256/120 512/140 1024/180
24 2.7121E03 6.7948E04 1.6990E04 4.2469E05
1.9969 1.9997 2.0002
26 4.5714E03 1.1459E03 2.8663E04 7.1668E05
1.9962 1.9992 1.9998
28 5.7130E03 1.4323E03 3.5830E04 8.9590E05
1.9960 1.9990 1.9998
210 6.3522E03 1.5926E03 3.9840E04 9.9615E05
1.9959 1.9991 1.9998
eε𝒩,Δθ 6.3522E03 1.5926E03 3.9840E04 9.9615E05
pεN,Δθ 1.9959 1.9991 1.9998
Table 3. Maximum point-wise errors and order of convergence on a modified graded mesh for Example 13.
Number of Intervals N/Δθ
ε 128/110 256/120 512/140 1024/180
24 5.7176E03 1.6558E03 4.2963E04 1.0843E04
1.7878 1.9463 1.9862
26 1.6010E02 4.0517E03 1.0159E03 2.5417E04
1.9823 1.9957 1.9989
28 2.7172E02 6.8564E03 1.7181E03 4.2978E04
1.9865 1.9966 1.9991
210 3.3892E02 8.5641E03 2.1467E03 5.3704E04
1.9845 1.9962 1.9990
eε𝒩,Δθ 3.3892E02 8.5641E03 2.1467E03 5.3704E04
pεN,Δθ 1.9959 1.9991 1.9998
Table 4. Maximum point-wise errors and order of convergence on a Shishkin mesh for Example 13.
Results by proposed method on the modified graded mesh
Number of Intervals N/Δθ
(ε=24,26,28,210) 128/110 256/120 512/140 1024/180
eε𝒩,Δθ 2.7121E03 6.7948E04 1.6990E04 4.2469E05
pεN,Δθ 1.9969 1.9997 2.0002
Results by proposed method on the Shishkin mesh
Number of Intervals N/Δθ
(ε=24,26,28,210) 128/110 256/120 512/140 1024/180
eε𝒩,Δθ 3.3892E02 8.5641E03 2.1467E03 5.3704E04
pεN,Δθ 1.9959 1.9991 1.9998
Results by [19]
Number of Intervals N/Δθ
(ε=24,26,28,210) 128/110 256/120 512/140 1024/180
eε𝒩,Δθ 1.6600E03 5.7800E04 1.9200E04 5.9500E05
pεN,Δθ 1.5270 1.5900 1.6900
Table 5. Comparison of Maximum uniform errors and order of uniform convergence for Example 13.
Refer to caption
Refer to caption
Figure 2. Log-log plot for Example 12 in Fig. 2 and for Example 13 in Fig. 2.

7. Discussion and Conclusions

In this article, for the first time, we propose a modified graded mesh for reaction-diffusion problems that provides second-order uniform convergence with respect to the perturbation parameter. We have presented effective numerical approaches in this work that are based on a modified graded mesh. We have presented upwind finite difference schemes on modified graded mesh and Shishkin mesh. In order to verify the theoretical estimation established, we conduct numerical experiments for two test problems for various values of ε, and step sizes N and M. In order to find maximum point-wise error and corresponding order of convergence, we double the number of mesh points in the time and spatial direction for the proposed schemes on the modified graded mesh and Shishkin mesh. Through this procedure we get the second-order convergence. These can be observed from the results presented in Table 1 and Table 2 for Example 12 and Table 3, Table 4 for Example 13. From the above tables it can be confirmed that overall second order uniform convergence. Corresponding log-log plot are provided for the Example 12 and Example 13. Fig. 2 shows the overall second order of convergence for various values of ε with the upwind finite difference scheme on modified graded mesh and Shishkin mesh. Table 5 display comparison of the proposed scheme with the previous works. It is observed the proposed scheme is better than those considered in [19]. The uniform convergence of the proposed methods is shown by the numerical results obtained for the test problems. The ability to build higher-order, more time-accurate numerical schemes using the current setting is a feasible extension that may be used to improve accuracy while reducing computing costs. Finally, numerical outcomes supports the theoretical findings.

Funding: The first author Mr. Kishun Kumar Sah conveys his profound gratitude to the Department of Science and Technology, Govt. of India for providing INSPIRE fellowship (IF 190722).

Data Availability: Enquiries about data availability should be directed to the authors.

Acknowledgements.

The authors would like to thank the referees and editor for their useful remarks, which contributed to improving the quality of this article.

References