International Journal of Modern Research in Engineering and Technology (IJMRET)

www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

Numerical Solution of the Nonlocal Singularly Perturbed Problem

Musa Çakır, Derya Arslan

Department of Mathematics, Faculty of Science, University of Yuzuncu Yil, Van, Turkey
Ministry of National Education, Turkey

Abstract: The study is concerned with a different perspective which the numerical solution of the singularly
perturbed nonlinear boundary value problem with integral boundary condition using finite difference method on
Bakhvalov mesh. So, we show some properties of the exact solution. We establish uniformly convergent finite
difference scheme on Bakhvalov mesh. The error analysis for the difference scheme is performed. The numerical
experiment implies that the method is the first order convergent in the discrete maximum norm, independently of
𝜀- singular perturbation parameter with effective and efficient iterative algorithm. The numerical results are
shown in table and graphs.
Keywords: Singular perturbation, finite difference method, Bakhvalov mesh, uniformly convergence, integral
condition.

1. Introduction

This paper focused on the first-order nonlinear singularly perturbed differential equation.
𝜀𝑢′ + 𝑓 𝑥, 𝑢 = 0, 𝑥 ∈ 0, 𝐿 , 𝐿>0 (1)
subject to the integral boundary condition
𝐿
𝑢 0 = 𝛽𝑢 𝐿 + 𝑏 𝑠 𝑢 𝑠 𝑑𝑠 + 𝑑 (2)
0

where 𝜀 is a small perturbation parameter 0 < 𝜀 ≪ 1 . 𝛼, 𝛽, 𝑑 are constant, which are independent of 𝜀. We
assume that 𝑏 𝑥 , 𝑓(𝑥, 𝑢) are sufficiently smooth functions in the intervals (0, 𝐿] and (0, 𝐿]𝑥𝑅. Also,
𝜕𝑓
𝜕𝑢
≥ 𝛼 > 0. The solution of (1) has in general a boundary layer with O(𝜀) near to 𝑥 = 0. The first time,
nonlocal problems were studied by Bitsadze and Samarskii [16]. Differential equations with conditions which
connect the values of the unknown solution at the boundary with values in the interior are known as nonlocal
boundary value problems [16]. Existence and uniqueness of nonlocal problems can be seen in [1-3]. Singular
perturbation problems can also arise in fluid mechanics, quantum mechanics, hydromechanical problems,
chemical-reactor theory, control theory, oceanography, meteorology, electrical networks and other physical
models. The subject of nonlocal boundary value problems for singularly perturbed differential equations has
been studied by many authors. For example, Cziegis [4] studied the numerical solution of singularly perturbed
nonlocal problem. Cziegis [5] analyzed the difference schemes for problems with nonlocal conditions.
Amiraliyev and Çakır [8] applied the difference method on a Shishkin mesh to the Singularly perturbed three-
point boundary value problem. Amiraliyev and Çakır [9] studied numerical solution of the singularly perturbed
problem with nonlocal boundary condition. Çakır and Arslan [14] analyzed a numerical method for nonlinear
singularly perturbed multi-point boundary value problem. Adzic and Ovcin [11] studied nonlinear spp with
nonlocal boundary conditions and spectral approximation. Amiraliyev, Amiraliyeva and Kudu [12] applied a
numerical treatment for singularly perturbed differential equations with integral boundary condition. Herceg [6]
studied the numerical solution of a singularly perturbed nonlocal problem. Herceg [7] researched solving a
nonlocal singularly perturbed problem by splines in tension. Çakır [10] studied uniform second–order difference
method for a singularly perturbed three-point boundary value problem. Geng [15] applied a numerical algorithm
for nonlinear multi-point boundary value problems. Çakır [13] obtained a numerical study on the difference
solution of singularly perturbed semilinear problem with integral boundary condition. It is well known that
standard discretization methods for solving nonlocal singular perturbation problem are unstable and these don’t
give accurate results for 𝜀. Therefore, it is very important to find suitable numerical methods to these problem.
So, we use finite difference method in this paper. According to this method, we will analyze that this method for
the numerical solution of the nonlocal problem (1)-(2) is uniformly convergent of first order on Bakhvalov
mesh, in discrete maximum norm, indepedently of singular perturbation parameter 𝜀. In section 2, some
properties of the exact solution of problem described in (1)-(2) is investigated. According to the perturbation
parameter, by the method of integral identities with the use exponential basisfunctions and interpolating
quadrature rules with the weight and remainder terms in integral form uniformly convergent finite difference
scheme on Bakhvalov mesh is established in section 3. The error analysis for the difference scheme is
performed in section 4. In section5, numerical examples are presented to find the solution of approximation. In

www.ijmret.org Page 13
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

this paper, we use 𝑔(𝑥) ∞ for the continuous maximum norm on the corresponding interval. 𝑔 𝑥 is any
continuous function.

2. Properties of the Exact Solution

The exact solution of (1)-(2) has some properties, which are needed in later sections for analysis of the
numerical solution. Thus, we describe these properties with Lemma 2.1.

𝜕𝑓 (𝑥,𝑢)
Lemma 2.1. If ≥ 𝛼 > 0 and 𝑓 𝑥, 𝑢 ∈ (0, 𝐿]𝑥𝑅, the estimates
𝜕𝑢

𝑢(𝑥) ≤ 𝐶0 (3)
where
−𝛼𝑡 −𝛼𝑇
𝐶0 = 𝑢 0 𝑒 𝜀 + 1−𝑒 𝜀 𝛼 −1 𝐹 ∞, 𝐹 𝑥 = −𝑓(𝑥, 0)

and
1 −𝛼𝑥
𝑢′ (𝑥) ≤𝐶 1+ 𝑒 𝜀 (4)
𝜀
hold for the solution 𝑢(𝑥), where 𝐶0 and 𝐶 are constant independent of 𝜀 and h.

Proof. We use intermediate value theorem for 𝑓 𝑥, 𝑢 in Equation (1)

𝑓 𝑥, 𝑢 − 𝑓(𝑥, 0) 𝜕𝑓 𝑥, 𝑢
= , 𝑢 = 𝛾𝑢, 0<𝛾<1
𝑢 𝜕𝑢
and
𝜕𝑓 𝑥, 𝑢
𝑎 𝑥 = .
𝜕𝑢
Thus, we obtain the following linear equation,
𝜀𝑢′ + 𝑎 𝑥 𝑢 𝑥 = 𝐹 𝑥 (5)
We can write the solution of the Equation (5) as follows:
𝑥
−1 𝑥 1 −1 𝑥
𝑢 𝑥 =𝑢 0 𝑒𝜀 0 𝑎 𝜏 𝑑𝜏 + 𝐹(𝜉)𝑒 𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝜉 (6)
𝜀 0

In Equation (5), 𝑥 = 𝑇 and 𝑥 = 𝑠, we obtain Eq (7)-Eq(9)

𝑇
−1 𝑇 1 −1 𝑇
𝑢 𝑇 =𝑢 0 𝑒𝜀 0 𝑎 𝜏 𝑑𝜏 + 𝐹(𝜉)𝑒 𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝜉 (7)
𝜀 0
𝑠
−1 𝑠 1 −1 𝑠
𝑢 𝑠 =𝑢 0 𝑒𝜀 0 𝑎 𝜁 𝑑𝜁 + 𝐹(𝜉)𝑒 𝜀 𝜉 𝑎 𝜁 𝑑𝜁 𝑑𝜉 (8)
𝜀 0

and
−1 𝑇 −1 𝑇
𝛽 𝑇 𝑎 𝜏 𝑑𝜏 𝑇 1 𝑇 𝑎 𝜏 𝑑𝜏
𝑑+ 0
𝐹(𝜉)𝑒 𝜀 𝜉 𝑑𝜉 +0
𝑏(𝑠) [ 0 𝐹 𝜉 𝑒 𝜀 𝜉 𝑑𝜉 ]𝑑𝑠
𝜀 𝜀
𝑢 0 = −1 𝑇 −1 𝑠 . (9)
𝑇
1 − 𝛽𝑒 𝜀 0 𝑎 𝜏 𝑑𝜏 − 0 𝑏(𝑠)𝑒 𝜀 0 𝑎 𝜏 𝑑𝜏 𝑑𝑠
Now, we evaluate to Equation (9) for 𝑎 𝜏 ≥ 𝛼 > 0 and 𝑏 ∗ = 𝑚𝑎𝑥 𝑏(𝑥) .
First, we prove the estimate the following inequality,
−1 𝑇 𝑇 −1 𝑠 −𝛼𝑇 𝑇 −𝛼𝑠 −𝛼𝑇 −𝛼𝑇
1 − 𝛽𝑒 𝜀 0 𝑎 𝜏 𝑑𝜏 − 𝑏 𝑠 𝑒𝜀 0 𝑎 𝜏 𝑑𝜏 𝑑𝑠 ≥ 1 − 𝛽𝑒 𝜀 − 𝑏∗ 𝑒 𝜀 𝑑𝑠 ≥ 1 − 𝛽𝑒 𝜀 − 𝑏∗ 1 − 𝑒 𝜀 ≥ 𝐶0
0 0

>0
Second, we prove the estimate the following inequality,

www.ijmret.org Page 14
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.
𝑇 𝑇 𝑇
𝛽 −1 𝑇 1 −1 𝑇
𝑑+ 𝐹 𝜉 𝑒𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝜉 + 𝑏(𝑠) [ 𝐹 𝜉 𝑒𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝜉 ]𝑑𝑠
𝜀 0 𝜀 0 0
𝑇 𝑇 𝑇
1 −1 𝑇 1 −1 𝑇
≤ 𝑑 + 𝛽 𝐹 𝑒𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝝃 + 𝑏 𝑠 𝐹 𝜉 𝑒𝜀 𝜉 𝑎 𝜏 𝑑𝜏 𝑑𝜉 𝑑𝑠
∞
𝜀 0 𝜀 0 0
𝑇 𝑇
1 −𝛼
𝑇−𝜉 1 −𝛼𝑠
= 𝑑 + 𝛽 𝐹 ∞ 𝑒 𝜀 𝑑𝝃 + 𝐹 ∞ 𝑏 𝑠 1−𝑒 𝜀 𝑑𝑠
𝜀 0 𝛼 0
𝑇
1 𝜀 −𝛼𝑇 1 −𝛼𝑠
≤ 𝑑 + 𝛽 𝐹 ∞ 1−𝑒 𝜀 + 𝐹 ∞ 𝑏 𝑠 1−𝑒 𝜀 𝑑𝑠
𝜀 𝛼 𝛼 0
𝑇
1 −𝛼𝑇 1 −𝛼𝑠
= 𝑑 + 𝛽 𝐹 ∞ 1−𝑒 𝜀 + 𝐹 ∞ 𝑏 𝑠 1−𝑒 𝜀 𝑑𝑠
𝛼 𝛼 0

1 1
≤ 𝑑 + 𝛽 𝐹 ∞ + 𝐹 ∞ 𝑏 1 ≤ 𝐶1
𝛼 𝛼
where
𝑇
𝑏 𝑠 𝑑𝑠 = 𝑏 1 .
0

Consequently, we obtain the following inequality:

𝑢 0 ≤𝐶
and

−𝛼𝑥 −𝛼𝑇
𝑢 𝑥 ≤ 𝑢 0 𝑒 𝜀 + 1−𝑒 𝜀 𝛼 −1 𝐹 ∞

−𝛼𝑥 −𝛼𝑇
≤ 𝐶0 −1 𝑑 + 𝐶0 −1 𝛽 + 𝑏 1 𝛼 −1 𝐹 ∞ 𝑒 𝜀 + 1−𝑒 𝜀 𝛼 −1 𝐹 ∞

≤ [𝐶0 −1 𝑑 + 𝐶0 −1 𝛽 + 𝑏 1 𝛼 −1 𝐹 ∞ ≤ 𝐶0 .
We now prove the estimate the Equation (4) :
𝜕𝑓
𝜕𝑥
≤ 𝐶, 𝑢(𝑥) ≤ 𝐶0.

If the above inequalities replace the Equation (6), we have the following inequality
−𝛼𝑥 −𝛼𝑇
𝑢 𝑥 ≤ 𝑢 0 𝑒 𝜀 + 1−𝑒 𝜀 𝛼 −1 𝐹 ∞ (10)

where 𝑢 0 ≤ 𝐶, 𝛼 −1 𝐹 ∞ ≤ 𝐶.
If we take derivative of the Equation (10) we obtain as following:
𝐶𝛼 −𝛼𝑥 𝛼 −𝛼𝑥 1 −𝛼𝑥
𝑢′ 𝑥 ≤ 𝑢 0
𝑒 𝜀 + 𝑒 𝜀 ≤ 𝐶 1+ 𝑒 𝜀
𝜀 𝜀 𝜀
thus, the proving of the Lemma 1 is completed.

3. The Establishment of Difference Scheme

Let us consider the following any non-uniform mesh on [0, 𝐿],

𝜔𝑁 = 𝑥0 < 𝑥1 < ⋯ < 𝑥𝑁−1 < 𝑥𝑁 , 𝑕𝑖 = 𝑥𝑖 − 𝑥𝑖−1

𝜔𝑁 = 𝜔𝑁 ∪ 𝑥 = 0, 𝑥 = 𝐿 .

First, we will construct the difference scheme for the Equation (1). First, we integrate the Equation (1) over
(𝑥𝑖−1 , 𝑥𝑖 )

www.ijmret.org Page 15
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.
𝑥𝑖 𝑥𝑖
𝑕𝑖 −1 𝜀𝑢′ (𝑥))𝑑𝑥 + 𝑕𝑖 −1 𝑓 𝑥, 𝑢 𝑑𝑥 = 0, 𝑖 = 1, 𝑁 − 1 (11)
𝑥 𝑖−1 𝑥 𝑖−1

and we obtain the following inequality

𝑥𝑖 𝑥𝑖
𝑕𝑖 −1 𝜀 𝑢 𝑥𝑖 − 𝑢 𝑥𝑖−1 + 𝑕𝑖 −1 𝑓 𝑥, 𝑢 𝑑𝑥 = 𝜀𝑢𝑥 ,𝑖 + 𝑕𝑖 −1 𝑓 𝑥, 𝑢 𝑑𝑥 = 0
𝑥 𝑖−1 𝑥 𝑖−1

where

𝑥𝑖
𝑅𝑖 = −𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 𝑓 ′ (𝑥, 𝑢)𝑑𝑥 . (12)
𝑥 𝑖−1

So, from the Equation (12) the difference scheme is defined by

𝜀𝑢𝑥 ,𝑖 + 𝑓 𝑥𝑖 , 𝑢𝑖 + 𝑅𝑖 = 0, 𝑖 = 1, 𝑁 (13)

Second, we define an approximation for the boundary condition of the Equation (1).
𝑁
𝑥𝑖
𝑢0 = 𝛽𝑢𝑁 + 𝑏 𝑠 𝑢 𝑠 𝑑𝑠 + 𝑑 (14)
𝑖=1 𝑥 𝑖−1
where
𝑁 𝑁
𝑥𝑖 𝑥𝑖
𝑏 𝑠 𝑢 𝑠 𝑑𝑠 = [𝑏 𝑠 𝑢 𝑠 − 𝑏𝑖 𝑢𝑖 ]𝑑𝑠
𝑖=1 𝑥 𝑖−1 𝑖=1 𝑥 𝑖−1
𝑁 𝑁

+ 𝑏𝑖 𝑢𝑖 𝑕𝑖 = 𝑟 + 𝑏𝑖 𝑢𝑖 𝑕𝑖 (15)
𝑖=1 𝑖=1
and remainder term
𝑁
𝑥𝑖
𝑑
𝑟= 𝑥 − 𝑥𝑖−1 𝑏 𝑡 𝑢 𝑡 𝑑𝒕 . (16)
𝑥 𝑖−1 𝑑𝑡
𝑖=1
𝑁

𝑢0 = 𝛽𝑢𝑁 + 𝑕𝑖 𝑏𝑖 𝑢𝑖 + 𝑑 + 𝑟. (17)
𝑖=1
Thus, by neglecting 𝑅𝑖 and 𝑟 in the Equation (13) and the Equation (17), we suggest the following difference
scheme for approximating the Equations (1)-(2) :
𝜀𝑦𝑥 ,𝑖 + 𝑓 𝑥𝑖 , 𝑦𝑖 = 0 , 𝑖 = 1, 𝑁 (18)
𝑁

𝑦0 = 𝛽𝑦𝑁 + 𝑕𝑖 𝑏𝑖 𝑦𝑖 + 𝑑 (19)
𝑖=1

4. Bakhvalov Mesh

We use Bakhvalov mesh, because the difference scheme (18)-(19) must be 𝜀 −uniform convergent.
According to the transition point following as:
𝐿
𝜎 = 𝑚𝑖𝑛 , −𝛼 −1 𝜀𝑙𝑛𝜀
2
𝑁
the piecewise uniform mesh divides each of the interval 0, 𝜎 and 𝜎, 𝐿 into 2
equidistant subintervals.
The mesh point 𝑥𝑖 is defined as follows:
𝐿 2𝑖 𝑁
𝜎< 𝑖𝑠𝑒, 𝑥𝑖 ∈ 0, 𝜎 , 𝑥𝑖 = −𝛼 −1 𝜀 ln 1 − 1 − 𝜀 , 𝑖 = 0,1, … ,
2 𝑁 2
−𝛼𝐿
𝐿 2𝑖 𝑁
𝜎= 𝑖𝑠𝑒 , 𝑥𝑖 ∈ 0, 𝜎 , 𝑥𝑖 = −𝛼 −1 𝜀 ln 1 − 1 − 𝑒 2𝜀 , 𝑖 = 0,1, … ,
2 𝑁 2
𝑁 𝑁 2 𝐿−𝜎
𝑥𝑖 ∈ 𝜎, 𝐿 , 𝑥𝑖 = 𝜎 + 𝑖 − 2
𝑕, 𝑖 = 2
+ 1, … , 𝑁, 𝑕 = 𝑁
.

www.ijmret.org Page 16
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

5. Uniform Error Estimates

In this section, we obtain the convergence of the method. First, we give the error function
𝑧𝑖 = 𝑦𝑖 − 𝑢𝑖 , 𝑖 = 0, 𝑁, 𝑧𝑖 is the solution of the discrete problem,
𝜀𝑧𝑥 ,𝑖 + 𝑓 𝑥𝑖 , 𝑦𝑖 − 𝑓 𝑥𝑖 , 𝑢𝑖 = 𝑅𝑖 , 𝑖 = 1, 𝑁 (20)
𝑁

𝑧0 = 𝛽𝑧𝑁 + 𝑕𝑖 𝑏𝑖 𝑧𝑖 − 𝑟 (21)
𝑖=1
where 𝑅𝑖 and r are defined by Equation (12) and Equation (16).

Lemma 5.1. Let 𝑧𝑖 be the solution to Equation (20),

𝑕𝑘
𝑏∗ = max𝑖 𝑏𝑖 , 𝜌𝑘 = , 𝑘 = 1,2 and 𝛼 = max𝑖 𝑎𝑖
𝜀

and

𝑁 𝑁 𝑖
1 1
1−𝛽 − 𝑏∗ 𝑕𝑖 ≠ 0.
1 + 𝛼𝜌𝑖 1 + 𝛼𝜌𝑘
𝑖=1

Then the estimate

𝑧 ∞,𝜔 𝑁 ≤𝐶 𝑅 ∞,𝜔 𝑁 + 𝑟 (22)

holds.

Proof. We use intermediate value theorem to rewrite the Equation (20) as follows:
𝑓 𝑥𝑖 , 𝑦𝑖 − 𝑓 𝑥𝑖 , 𝑢𝑖 = 𝑓𝑢 𝑦𝑖 − 𝑢𝑖 = 𝑓𝑢 𝑧𝑖
Here we assumed that
𝑓𝑢 = 𝑎𝑖 .
Thus, we can write
𝜀𝑧𝑥 ,𝑖 + 𝑎𝑖 𝑧𝑖 = 𝑅𝑖 , 𝑖 = 1, 𝑁 − 1 (23)
𝑁

𝑧0 = 𝛽𝑧𝑁 + 𝑕𝑖 𝑏𝑖 𝑧𝑖 − 𝑟. (24)
𝑖=1

Rearranging Equation (23) gives

𝜀 𝑕𝑖 𝑅𝑖
𝑧𝑖 = 𝑧𝑖−1 +
𝜀 + 𝑎𝑖 𝑕𝑖 𝜀 + 𝑎𝑖 𝑕𝑖
and
𝑖 𝑖 𝑖

𝑧𝑖 = 𝑧0 [ 𝑞𝑘 ] + 𝜑𝑘 [ 𝑞𝑗 ]
𝑘=1 𝑘 =1 𝑗 =𝑘+1

where
𝜀
𝑞𝑘 =
𝜀 + 𝑎𝑘 𝑕𝑘

and

𝑕𝑘 𝑅𝑘
𝜑𝑘 = .
𝜀 + 𝑎𝑘 𝑕𝑘

According to the maximum principle for Equations (23)-(24), we have the following inequality
𝑧 ∞,𝜔 𝑁 ≤ 𝑧0 + 𝛼 −1 𝑅 ∞,𝜔 𝑁 . (25)

www.ijmret.org Page 17
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

Now, we evaluate the Equation (24). After a simple calculation we obtain

𝑁

𝑧0 ≤ 𝛽 𝑧𝑁 + 𝑏∗ 𝑕𝑖 𝑧𝑖 + 𝑟 (26)
𝑖=1

where

𝑏∗ = max 𝑏𝑖
𝑖

and
𝑁

𝑧𝑁 = 𝑧0 ∅𝑁 + 𝜑𝑘 ∅𝑁−𝑘 . 27
𝑘 =1

The Equation (27) write in the Equation (26), we have

𝑁 𝑁 𝑖

𝑧0 ≤ 𝛽 𝑧0 ∅𝑁 + 𝜑𝑘 ∅𝑁−𝑘 + 𝑏∗ 𝑕𝑖 𝑧0 ∅𝑖 + 𝜑𝑘 ∅𝑖−𝑘 + 𝑟
𝑘=1 𝑖=1 𝑘=1

After doing some arragement,

𝑁
𝛽 𝑘=1 𝜑𝑘 ∅𝑁−𝑘 + 𝑏∗ 𝑁
𝑖=1 𝑕𝑖
𝑖
𝑘=1 𝜑𝑘 ∅𝑖−𝑘 + 𝑟
𝑧0 ≤ 𝑁 (28)
1 − 𝛽 ∅𝑁 − 𝑏∗ 𝑖=1 𝑕𝑖 ∅𝑖

it follows that

𝜌 𝑘 𝑅𝑘 1 𝑁 𝜌 𝑘 𝑅𝑘 1 𝑖
𝑁 𝑁 𝑖
𝛽 𝑘=1 1+𝛼𝜌 + 𝑏∗ 𝑖=1 𝑕𝑖 𝑘=1 1+𝛼𝜌 + 𝑟
𝑘 1+𝛼𝜌 𝑘 𝑘 1+𝛼𝜌 𝑘
𝑧0 ≤ 𝑁 𝑖
1 𝑁 1
1−𝛽 − 𝑏∗ 𝑖=1 𝑕𝑖
1+𝛼𝜌 𝑖 1+𝛼𝜌 𝑘
≤ 𝐶( 𝑅 ∞,𝜔 𝑁 + 𝑟 ) (29)

where
𝑕𝑘
𝜌𝑘 = , 𝑘 = 1,2 and 𝛼 = max𝑖 𝑎𝑖 .
𝜀

If the Equation(29) write in the Equation (25), we have

𝑧 ∞,𝜔 𝑁 ≤𝐶 𝑅 ∞,𝜔 𝑁 + 𝑟 + 𝛼 −1 𝑅 ∞,𝜔 𝑁

≤𝐶 𝑅 ∞,𝜔 𝑁 + 𝑟 .
𝜕𝑓
Lemma 5. 2. Under the assumptions of Lemma 1 and 𝜕𝑢 ≥ 𝛼 > 0, 𝑓 𝑥, 𝑢 ∈ 𝐶 2 0, 𝐿 , the solution of (1)-(2)
satisfies the following estimates:
𝑅 ∞,𝜔 𝑁 ≤ 𝐶𝑁 −1 (30)

𝑟 ≤ 𝐶𝑁 −1 (31)

Proof. First, we consider error function 𝑅𝑖 ,

𝑥𝑖
𝑅𝑖 ≤ 𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 𝑓 ′ (𝑥, 𝑢) 𝑑𝑥
𝑥 𝑖−1
𝑥𝑖
𝜕𝑓 𝑑𝑥 𝜕𝑓 𝑥, 𝑢 𝑥 𝑑𝑢
= 𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 + 𝑑𝑥
𝑥 𝑖−1 𝜕𝑢 𝑑𝑥 𝜕𝑢 𝑑𝑥
𝑥𝑖 𝑥𝑖
≤ 𝐶𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 𝑑𝑥 + 𝐶𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 𝑢′ 𝑥 𝑑𝑥
𝑥 𝑖−1 𝑥 𝑖−1

www.ijmret.org Page 18
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.
𝑥𝑖
1 −𝛼𝑥
≤ 𝐶𝑕𝑖 + 𝐶𝑕𝑖−1 𝑥𝑖 − 𝑥𝑖−1 1 + 𝑒 𝜀 𝑑𝑥
𝑥 𝑖−1 𝜀
𝑥𝑖 −𝛼𝑥
≤ 𝐶 𝑕𝑖 + 𝑕𝑖−1 𝜀 −1 𝑥𝑖 − 𝑥𝑖−1 𝑒 𝜀 𝑑𝑥 , 1 ≤ 𝑖 ≤ 𝑁. (32)
𝑥 𝑖−1

In the first case, 𝑥𝑖 ∈ 0, 𝜎 ,

1
𝑎) 𝜎 < 2 , 𝜎 = −𝛼 −1 𝜀𝑙𝑛𝜀:

2𝑖 2(𝑖 − 1)
𝑕𝑖 = 𝑥𝑖 − 𝑥𝑖−1 = 𝛼 −1 𝜀 ln 1 − 1 − 𝜀 − ln 1 − 1 − 𝜀
𝑁 𝑁
where, according to 𝑖, we use intermediate value theorem as follows:

2(1 − 𝜀)𝑁 −1
𝑕𝑖 = 𝛼 −1 𝜀 ≤ 2𝛼 −1 1 − 𝜀 𝑁 −1 ≤ 𝐶𝑁 −1 (33)
1 − 𝑖1 2(1 − 𝜀)𝑁 −1

Thus, we obtain
𝑥𝑖 −𝛼𝑥 𝑥𝑖 −𝛼𝑥
𝑅𝑖 ≤ 𝐶 𝑕𝑖 + 𝑕𝑖−1 𝜀 −1 𝑥𝑖 − 𝑥𝑖−1 𝑒 𝜀 𝑑𝑥 ≤ 𝐶 𝑕𝑖 + 𝑕𝑖−1 𝜀 −1 𝑕𝑖 𝑒 𝜀 𝑑𝑥
𝑥 𝑖−1 𝑥 𝑖−1
𝑥𝑖 −𝛼𝑥 −𝛼 𝑥 𝑖 −𝛼 𝑥 𝑖−1
≤ 𝐶 𝑕𝑖 + 𝜀 −1 𝑒 𝜀 𝑑𝑥 = 𝐶 𝑕𝑖 + 𝛼 −1 𝑒 𝜀 −𝑒 𝜀
𝑥 𝑖−1
−𝛼 2𝑖 2 𝑖−1
−1 𝛼 −1 𝜀 ln 1− 1−𝜀 𝛼 −1 𝜀 ln 1− 1−𝜀
= 𝐶 𝑕𝑖 + 𝛼 𝑒 𝜀 𝑒 𝑁 −𝑒 𝑁 = 𝐶 𝑕𝑖 + 2𝛼 −1 1 − 𝜀 𝑁 −1
≤ 𝐶 𝐶𝑁 −1 + 2𝛼 −1 1 − 𝜀 𝑁 −1 ≤ 𝐶𝑁 −1 .
1 1
𝑏) 𝜎 = 2 , 𝜎 = 2 < −𝛼 −1 𝜀𝑙𝑛𝜀 :

−𝛼 2𝑖 −𝛼 2(𝑖 − 1)
𝑕𝑖 = 𝑥𝑖 − 𝑥𝑖−1 = 𝛼 −1 𝜀 ln 1 − 1 − 𝑒 2𝜀 − ln 1 − 1 − 𝑒 2𝜀
𝑁 𝑁
where, according to 𝑖, we use intermediate value theorem as follows:
−𝛼

−1
2(1 − 𝑒 2𝜀 )𝑁 −1
𝑕𝑖 = 𝛼 𝜀 −𝛼 ≤ 2𝛼 −1 𝑁 −1 ≤ 𝐶𝑁 −1 . (34)
1 − 𝑖1 2(1 − 𝑒 2𝜀 )𝑁 −1

Hence
−𝛼 −𝛼
2𝑖 2 𝑖−1
−𝛼 𝑥 𝑖 −𝛼 𝑥 𝑖−1 −𝛼 𝛼 −1 𝜀 ln 1− 1−𝑒 2𝜀 𝛼 −1 𝜀 ln 1− 1−𝑒 2𝜀
−1 −1 𝑁 𝑁
𝑅𝑖 ≤ 𝐶 𝑕𝑖 + 𝛼 𝑒 𝜀 −𝑒 𝜀 = 𝐶 𝑕𝑖 + 𝛼 𝑒 𝜀 𝑒 −𝑒

≤ 𝐶𝑁 −1

In the second case, 𝑥𝑖 ∈ 𝜎, 1 , we obtain

𝑥𝑖 −𝛼𝑥 1
𝑅𝑖 ≤ 𝐶 𝑕𝑖 + 𝜀 −1 𝑒 𝜀 𝑑𝑥 ≤ 𝐶 + 𝜀 −1 𝑕𝑖 ≤ 𝐶 𝑁 −1 + 𝜀 −1 𝑕−1 ≤ 𝐶 𝑁 −1 + 𝜀 −1 𝑁 −1 ≤ 𝐶𝑁 −1 .
𝑥 𝑖−1 𝑁

Consquently, we have

𝑅 ∞,𝜔 𝑁 ≤ 𝐶𝑁 −1 . (35)

Now, we evaluate error function𝑟 as follows:

𝑥 − 𝑥𝑖−1 ≤ 𝑕𝑖 , 𝑏 ∗ = 𝑚𝑎𝑥 𝑏 𝑥 ≤ 𝐶 , 𝑏 ′ 𝑥 ≤ 𝐶 ve 𝑢′ (𝑥) ≤ 𝐶0

under the above assumptions,

www.ijmret.org Page 19
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.
𝑁
𝑥𝑖
𝑟 ≤ 𝑥 − 𝑥𝑖−1 𝑏 ′ 𝑥 𝑢 𝑥 + 𝑏 𝑥 𝑢′ 𝑥 𝑑𝑥
𝑖=1 𝑥 𝑖−1
𝑁
𝑥𝑖
1 𝛼𝑥
≤ 𝑕𝑖 𝐶 + 𝐶 1 + 𝑒− 𝜀 𝑑𝑥
𝑥 𝑖−1 𝜀
𝑖=1
𝑁
𝑥𝑖
1 𝛼𝑥
≤𝐶 𝑕𝑖 1 + 𝑒 − 𝜀 𝑑𝑥
𝑥 𝑖−1 𝜀
𝑖=1
𝑁
2 𝑁
𝑥𝑖 𝑥𝑖
1 𝛼𝑥 1 𝛼𝑥
≤𝐶 𝑕𝑖 1 + 𝑒 − 𝜀 𝑑𝑥 + 𝑕𝑖 1 + 𝑒 − 𝜀 𝑑𝑥
𝑥 𝑖−1 𝜀 𝑁 𝑥 𝑖−1 𝜀
𝑖=1 𝑖= +1
2
𝜎 1
1 𝛼𝑥 1 𝛼𝑥
= 𝐶 𝑕𝑖 1 + 𝑒 − 𝜀 𝑑𝑥 + 𝑕𝑖 1 + 𝑒 − 𝜀 𝑑𝑥 ≤ 𝐶 𝑕𝑖 (36)
0 𝜀 𝜎 𝜀
In the first case, 𝑥𝑖 ∈ 0, 𝜎 :
1
𝑎) 𝜎 < , 𝜎 = −𝛼 −1 𝜀𝑙𝑛𝜀,
2

2(𝑖 − 1)
𝑥𝑖−1 = 𝛼 −1 𝜀 ln 1 − 1 − 𝜀
𝑁
and

2𝑖 2(𝑖 − 1)
𝑕𝑖 = 𝑥𝑖 − 𝑥𝑖−1 = 𝛼 −1 𝜀 ln 1 − 1 − 𝜀 − ln 1 − 1 − 𝜀
𝑁 𝑁
where, according to 𝑖, we use intermediate value theorem as follows:

2(1 − 𝜀)𝑁 −1
𝑕𝑖 = 𝛼 −1 𝜀 ≤ 2𝛼 −1 1 − 𝜀 𝑁 −1 ≤ 𝐶𝑁 −1 (37)
1 − 𝑖1 2(1 − 𝜀)𝑁 −1

Thus, we obtain

𝑟 ≤ 𝐶 𝑕𝑖 + 𝑕𝑖 ≤ 𝐶𝑁 −1 .
1 1
b) 𝜎 = , 𝜎 = < −𝛼 −1 𝜀𝑙𝑛𝜀,
2 2
After a simple calculation we obtain,
−𝛼
2(1 − 𝑒 2𝜀 )𝑁 −1
𝑕𝑖 = 𝛼 −1 𝜀 −𝛼 ≤ 2𝛼 −1 𝑁 −1 ≤ 𝐶𝑁 −1 . (38)
1 − 𝑖1 2(1 − 𝑒 2𝜀 )𝑁 −1

Hence

𝑟 ≤ 𝐶 𝑕𝑖 + 𝑕𝑖 ≤ 𝐶𝑁 −1 .

In the second case, 𝑥𝑖 ∈ 𝜎, 1 :

1 1
a) 𝜎 < 2 , −𝛼 −1 𝜀𝑙𝑛𝜀 < 2,
we use the following equation,
2 1−𝜎 2 𝑁
𝑕 = 𝑕𝑖 = 𝑕2 = ≤ , 𝑖= + 1, 𝑁
𝑁 𝑁 2
we obtain,

𝑟 ≤ 𝐶 𝑕𝑖 + 𝑕𝑖 ≤ 𝐶 𝑕1 + 𝑕2 ≤ 𝐶 𝑁 −1 + 2𝑁 −1 ≤ 𝐶𝑁 −1

and

www.ijmret.org Page 20
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

𝑟 ≤ 𝐶𝑁 −1 . (39)

Teorem 5.1. Under the assumptions of Lemma 5.1. and Lemma 5.2. Let 𝑢(𝑥), be the solution of (1)-(2) and 𝑦𝑖 ,
be the solution of (18)-(19). Then, the following uniform error estimate satisfies

𝑦−𝑢 ∞,𝜔 𝑁 ≤ 𝐶𝑁 −1 . (40)

6. Algorithm and Numerical Results

We give some numerical results to solve the nonlinear problem (18)-(19) on Bakhvalov mesh using
quasilinearization technique.

6.1. Algorithm

We present the following quasilinearization technique for the difference scheme (18)-(19):
𝜀 𝑛 𝜀 𝜕𝑓 𝑥𝑖 , 𝑦𝑖 𝜕𝑓 𝑥𝑖 , 𝑦𝑖
𝑦𝑖−1 − + 𝑦𝑖 𝑛 = 𝑓 𝑥𝑖 , 𝑦𝑖 − 𝑦𝑖 𝑛−1 , 𝑖 = 1, … , 𝑁
𝑕𝑖 𝑕𝑖 𝜕𝑦 𝜕𝑦
𝑛−1 𝜕𝑓 𝑥 𝑖 ,𝑦 𝑖 𝜀 𝑛
𝑓 𝑥𝑖 , 𝑦𝑖 − 𝑦𝑖 𝜕𝑦
− 𝑕 𝑦𝑖−1
𝑛 𝑖
𝑦𝑖 = 𝜀 𝜕𝑓 𝑥 𝑖 ,𝑦 𝑖
, 𝑖 = 1, . . , 𝑁
− +
𝑕𝑖 𝜕𝑦
𝑁
(𝑛−1) (𝑛−1)
𝑦0 𝑛 = 𝛽𝑦𝑁 + 𝑕𝑖 𝑏𝑖 𝑦𝑖 + 𝑑, 𝑛 = 1,2, …
𝑖=1

6.2. Numerical Results

We give some examples to see the effectiveness of the presented method

Example 1

𝜀𝑢′ + 2𝑢 − 𝑒 −𝑢 + 𝑥 2 = 0, 0 < 𝑥 < 1 (41)

1
1 1
𝑢 0 = 𝑢 1 + 𝑒 −𝑠 𝑢 𝑠 𝑑𝑠 + 1 (42)
2 4 0

this problem has not the exact solution. Thus, we use 𝜔 −mesh to calculate the errors and convergence rates,
respectively.

𝑒𝑁
𝑒 𝑁 = max 𝑒𝜀𝑁 , 𝑒𝜀𝑁 = max 𝑦𝑖𝑁 − 𝑦2𝑖
2𝑁
, 𝑃𝜀𝑁 = log 2
𝜀 𝑖 𝑒 2𝑁

where, 𝑦𝑖𝑁 and 𝑦2𝑖

2𝑁
are the numerical solutions for 𝑁and 2𝑁.

According to (6.1), we give the algorithm as follows:

𝜀 (𝑛) 𝜀 𝑛 −1 𝑛 −1 𝑛 −1
𝑦𝑖−1 − + 2 + 𝑒 −𝑦𝑖 𝑦𝑖 𝑛 = 2𝑦𝑖 𝑛−1 + 𝑥𝑖2 − 𝑦𝑖 𝑛−1 2 + 𝑒 −𝑦𝑖 − 𝑒 −𝑦𝑖
𝑕𝑖 𝑕𝑖

𝜀 𝑛 𝑛−1 𝑛−1 𝑛 −1 𝑛 −1

𝑕𝑖
𝑦𝑖−1 − (2𝑦𝑖 + 𝑥𝑖2 − 𝑦𝑖 2 + 𝑒 −𝑦𝑖 − 𝑒 −𝑦𝑖 )
𝑛
𝑦𝑖 = 𝑛 −1
, 𝑖 = 1, . . , 𝑁
𝜀
𝑕𝑖
+ 2 + 𝑒 −𝑦𝑖

www.ijmret.org Page 21
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.
𝑁
𝑛 1
𝑦0 = 𝑦𝑁𝑛−1 + 𝑕𝑖 𝑏𝑖 𝑦𝑖 𝑛−1 + 1, 𝑛 = 1,2, …
2
𝑖=1

the initial guess in the iteration procedure is 𝑦𝑖 0 = 0.5,

(𝑛) (𝑛−1)
The stopping criterion is taken as 𝑦𝑖 − 𝑦𝑖 ≤ 10−5 .
The numerical results obtained from the problem of the difference scheme by comparison, the error and
uniform rates of convergence were found and these are shown in Table 1. Consequently, numerical results show
that the proposed scheme is working very well.

Table 1. The computed maximum pointwise errors 𝑒𝑁 and 𝑒2𝑁 , the numerical rate of convergence 𝑝𝑁 on the
Bakhvalow mesh 𝜔𝑁 for different values of 𝑁and 𝜀.
𝜀/𝑁 8 16 32 64

2−1 0.0364705764 0.0213939341 0.0115893393 0.0051560120

p=0.76 p=0.88 p=1.05

2−2 0.0457353908 0.0263063297 0.0141504831 0.0073451227

p=0.79 p=0.89 p=0.94

2−3 0.0547795035 0.0316355902 0.0172632555 0.0090159047

p=0.79 p=0.87 p=0.93

2−4 0.0592489265 0.0354285745 0.0192899598 0.0100776157

p=0.74 p=0.87 p=0.93

2−5 0.0617440905 0.0376291853 0.0204860025 0.0107447933

p=0.71 p=0.87 p=0.93

2−6 0.0632177961 0.0389122636 0.0212015157 0.0111341923

p=0.70 p=0.87 p=0.92

2−7 0.0640959677 0.0396529864 0.0216490391 0.0113563269

p=0.69 p=0.87 p=0.93

2−8 0.0646130925 0.0400746536 0.0219009225 0.010926500

p=0.68 p=1.00
P=0.87

Figure 1. Exact solution distribution for 𝑁 = 16 and 𝜀 = 2−4 , 2−6 , 2−8

www.ijmret.org Page 22
 International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

Figure 2. Error distribution for 𝑁 = 16 and 𝜀 = 2−2 , 2−4 , 2−6 , 2−8

From the graps it is show that the error is maximum near the boundary layer and it is almost zero in
outer region in the Figure 2. When 𝜀 −values are small, graphic curves leaned more towards the coordinate axes
in Figure 1.
Consequently, the purpose of this study was to give uniform finite difference method for numerical
solution of nonlinear singularly perturbed problem with nonlocal boundary conditions. The numerical method
was constructed on Bakhvalov mesh. The method was pointed out to be convergent, uniformly in the ε-
parameter, of first order in the discrete maximum norm. The numerical example illustrated in practice the result
of convergence proved theoretically.

References

[1] N. Adzic. (2000) Spectral Approximation and Nonlocal Boundary Value Problems. Novi Sad J. Math.,
30:1-10.
[2] M. Benchohra and S. K. Ntouyas. (2000) Existence of Solutions of Nonlinear Differential Equations
with Nonlocal Conditions. J. Math. Anal. Appl., 252:477-483.
[3] Jankowski, T. (2004) Existence of Solutions of Differential Equations with Nonlinear Multi-point
Boundary Conditions. Comput. Math. Appl., 47:1095-1103.
[4] Cziegis, R. (1988) The Numerical Solution of Singularly Pertürbed Nonlocal Problem. Lietuvos
Matematikos Rinkinys, 28, 144-152. (In Russian)
[5] Cziegis, R. (1991) The Difference Schemes for Problems with Nonlocal Conditions. Informatica
(Lietuva), 2, 155-170.
[6] Herceg, D. (1990) On the Numerical Solution of a Singularly Perturbed Nonlocal Problem. Univ. u
Novom Sadu Zb. Rad. Prirod.-Mat. Fak. Ser. Mat., 20, 1-10.
[7] Herceg, D. (1991) Solving a Nonlocal Singularly Perturbed Problem by Splines in Tension. Univ. u
Novom Sadu Zb. Rad. Prirod.-Mat. Fak. Ser. Mat, 21, 119-132.
[8] Amiraliyev, G.M. and Cakir, M. (2007) Numerical Solution of a Singularly Perturbed Three-Point
Boundary Value Problem. International Journal of Applied Mathematics, 84, 1465-1481.
[9] Amiraliyev, G. M. and Çakır, M. (2002) Numerical Solution of the Singularly Perturbed Problem with
Nonlocal Boundary Condition. Applied Mathematics and Mechanics (English Edition), 23, 755-764.
[10] Çakır, M. (2010) Uniform Second-Order Difference Method for a Singularly Perturbed Three-Point
Boundary Value Problem. Hindawi Publising Corporation Advances in Difference Equations, Vol.
2010, 13 p.
[11] Adzic, N. and Ovcin, Z. (2001) Nonlinear Spp with Nonlocal Boundary Conditions and Spectral
Approximation. Novi Sad Journal of Mathematics, 31, 85-91.
[12] Amiraliyev, G.M., Amiraliyeva, I.G. and Kudu, M. (2007) A Numerical Treatment for Singularly
Perturbed Differential Equations with İntegral Boundary Condition. Applied Mathematics and
Computations, 185, 574-582.
[13] Çakır, M. (2016) A Numerical Study on the Difference Solution of Singularly Perturbed Semilinear
Problem with Integral Boundary Condition. Mathematical Modelling and Analysis Vol. 21, 644-658.

www.ijmret.org Page 23
International Journal of Modern Research in Engineering and Technology (IJMRET)
www.ijmret.org Volume 1 Issue 5 ǁ December 2016.

[14] Çakır, M. and Arslan, D. (2016) A Numerical Method for Nonlinear Singularly Perturbed Multi-Point
Boundary Value Problem. Journal of Applied Mathematics and Physics, 4, 1143-1156.
[15] Geng, F.Z. (2012) A Numerical Algorithm for Nonlinear Multi-Point Boundary Value Problems.
Journal of Computational and Applied Mathematics, 236, 1789-1794.
[16] Bitsadze, A. V. and Samarskii, A. A., (1969) On Some Simpler Generalization of Linear Elliptic
Boundary Value Problems. Doklady Akademii Nauk SSSR. 185: 739-740.

www.ijmret.org Page 24