Chapter 4

Reaction-Diffusion Problem of
Fourth-order

4.1 Introduction

In Chapters 2 and 3, hybrid difference schemes are proposed to solve second-order

reaction-diffusion problems with Dirichlet and Robin boundary conditions, respec-
tively. In this chapter, the scope of work is extended to solve fourth-order singularly
perturbed boundary value problems of reaction-diffusion type.

As stated, many authors [154–156, 248] have considered discrete analogues of

arc-length monitor function to obtain an accurate approximate solution for reaction-
diffusion problems. However, Kopteva et al. [163] demonstrated that an arc-length
monitor function fails to produce satisfactory numerical approximation for reaction-
diffusion problems and proposed a new monitor function in [249]. The proposed
monitor function is similar to the one obtained in [245] and works well for both
convection-diffusion and reaction-diffusion problems of scalar form [250]. It is even
suitable for reaction-diffusion problems with robin type boundary conditions [173].
Das and Natesan in [251] presented a rigorous convergence analysis based on arc
length and curvature-based monitor functions. A higher-order error estimate by
a curvature-based error monitor function is proposed in [252]. Here, initial mesh
points are moved using de Boor’s algorithm [253]. Nevertheless, all of these efforts
are for second-order singularly perturbed problems. Researchers in [251, 254] pre-
sented adaptive difference schemes to solve fourth-order differential equations. These

101
4.2. The Continuous Problem 102

schemes are based on standard difference methods on suitably chosen layer adapted
grids. Nevertheless, in both cases, only first-order convergence is achieved. The
analysis of special methods for higher-order singular perturbation problems based on
equidistributed grids has seen little development and lacks due attention.

In this chapter, the proposed parameter-uniform hybrid finite difference approxi-

mation technique is applied to solve fourth-order reaction-diffusion problems on a
layer adapted grid. The proposed technique attains fourth-order uniform convergence
in discrete maximum norm. Firstly, the fourth-order equation is transformed into
a coupled system of two second-order differential equations. Later, for the coupled
system of differential equations, a highly accurate numerical approximation scheme
is generated by suitably combining fourth-order Hermite difference scheme with
the classical difference scheme on a layer adapted grid. The layer adapted grid
𝑁 is generated by equidistributing the monitor function 𝑀 (𝑠, 𝑢(𝑠)) using the
{𝑥𝑖 }𝑖=0
equidistribution principle specified in (2.1). The stability analysis is carried out for
the proposed hybrid difference scheme. The error analysis indicates that the proposed
scheme has uniform convergence of fourth-order.

4.2 The Continuous Problem

Consider the following fourth-order two-point singularly perturbed boundary value
problem

𝐿𝑢(𝑥) ≡ −𝜖𝑢 (𝑖𝑣) (𝑥) + 𝑏(𝑥)𝑢00 (𝑥) − 𝑐(𝑥)𝑢(𝑥) = − 𝑓 (𝑥), 𝑥 ∈ Ω = (0, 1),
)
(4.1)
𝑢(0) = 𝛼1 , 𝑢(1) = 𝛼2 , 𝑢00 (0) = −𝛼3 , 𝑢00 (4) = −𝛼4 .

Here, the functions 𝑏(𝑥), 𝑐(𝑥) and 𝑓 (𝑥) are sufficiently smooth and satisfy

𝜌 ∗ > 𝑏(𝑥) > 𝜌 > 0 and − 𝛾 ≤ 𝑐(𝑥) ≤ 0, 𝛾 > 0, ∀𝑥 ∈ Ω = [0, 1], (4.2)

with
0 < 𝜂 ≤ 𝜌 − 2𝛾, (4.3)

where 𝜂 is a real number. These conditions ensure that the problem (4.1) has a
unique solution. Moreover, the solution of the problem exhibits boundary layers at
outflow boundary regions of the domain. To solve the fourth-order problem (4.1), it
103 4. Reaction-Diffusion Problem of Fourth-order

is firstly transformed into the following coupled system of two second-order ordinary
differential equations

𝐿 1 u(𝑥) ≡ −𝑢001 (𝑥) − 𝑢 2 (𝑥) = 0, 𝑥 ∈ Ω,






(4.4)


Lu(𝑥) = f(𝑥) ⇔ 𝐿 2 u(𝑥) ≡ −𝜖𝑢002 (𝑥) + 𝑏(𝑥)𝑢 2 (𝑥) + 𝑐(𝑥)𝑢 1 (𝑥) = 𝑓 (𝑥),



 𝑢 1 (0) = 𝛼1 , 𝑢 1 (1) = 𝛼2 , 𝑢 2 (0) = 𝛼3 , 𝑢 2 (1) = 𝛼4 ,


where, L = (𝐿 1 , 𝐿 2 )𝑇 , u(𝑥) = (𝑢 1 (𝑥), 𝑢 2 (𝑥))𝑇 and f(𝑥) = (0, 𝑓 (𝑥))𝑇 .

So, throughout the chapter, the system of second-order equations (4.4) will repre-
sent the fourth-order problem (4.1).

4.2.1 Properties of the Exact Solution

For the coupled system of second-order equations (4.4), we begin our analysis with
the following maximum principle.

Lemma 4.2.1. Maximum Principle: Let u(𝑥) be a smooth function satisfying (4.4).
Assume that u(0) ≥ 0 and u(1) ≥ 0. Then, Lu ≥ 0 on Ω = (0, 1) implies that u ≥ 0
on Ω = [0, 1].

Proof. Let 𝑝, 𝑞 be such that 𝑢 1 ( 𝑝) = min {𝑢 1 (𝑥)} and 𝑢 2 (𝑞) = min {𝑢 2 (𝑥)}. Without
𝑥∈Ω 𝑥∈Ω
loss of generality, assume that 𝑢 1 ( 𝑝) ≤ 𝑢 2 (𝑞). Also, assume that 𝑢 1 ( 𝑝) < 0. Clearly,
𝑝 ≠ {0, 1} and 𝑢001 ( 𝑝) ≥ 0. Then, the first component of Lu( 𝑝) is

𝐿 1 u( 𝑝) ≡ −𝑢001 ( 𝑝) − 𝑢 2 ( 𝑝) = −𝑢001 ( 𝑝) + 𝑢 2 (𝑞) − 𝑢 2 ( 𝑝) − 𝑢 2 (𝑞) < 0,

which is a contradiction. Hence, it follows that u(𝑥) ≥ 0, ∀𝑥 ∈ Ω. 

An immediate consequence of the above maximum principle is the following

stability estimate.

Lemma 4.2.2. Let u be the solution of the coupled system of equations (4.4). Then,
there exists a constant 𝐶 independent of 𝜖 such that
 
kuk ≤ 𝐶 max ku(0) k , ku(1) k , max kf(𝑥) k on Ω.
𝑥∈Ω
4.2. The Continuous Problem 104

 
Proof. Let 𝜃 = 𝐶 max ku(0)k , ku(1)k , max kf(𝑥) k . For 𝑥 ∈ Ω, we define two
𝑥∈Ω
functions 𝜓 (𝑥) = 𝜓1 (𝑥), 𝜓2 (𝑥) as
± ± ±

𝑇

𝑥2
 
𝜓1± (𝑥) = 𝜃 (1 + 𝛿) 1 − ± 𝑢 1 (𝑥) and 𝜓2± (𝑥) = 𝜃 ± 𝑢 2 (𝑥), where 0 < 𝛿 << 1. (4.5)
2

Then, for a proper choice of 𝐶, 𝐿 1𝜓 ± (𝑥) ≥ 0 and 𝐿 2𝜓 ± (𝑥) ≥ 0 as follows. From (4.5),
we determine
2𝑥
𝜓1±0 (𝑥) = 𝜃 (1 + 𝛿)(− ) ± 𝑢01 (𝑥) = 𝜃 (1 + 𝛿)(−𝑥) ± 𝑢01 (𝑥),
2
𝜓1±00 (𝑥) = 𝜃 (1 + 𝛿)(−1) ± 𝑢001 (𝑥),

and
𝜓2±0 (𝑥) = ±𝑢02 (𝑥), and 𝜓2±00 (𝑥) = ±𝑢002 (𝑥).

It follows that
 𝑇
±
𝐿 1𝜓 (𝑥) = 𝐿 1 𝜓1± (𝑥), 𝜓2± (𝑥) = −𝜓1±00 (𝑥) − 𝜓2± (𝑥)
= 𝜃 (1 + 𝛿) ∓ 𝑢001 (𝑥) − (𝜃 ± 𝑢 2 (𝑥)) = 𝜃 (1 + 𝛿) ∓ 𝑢001 (𝑥) − 𝜃 ∓ 𝑢 2 (𝑥)
= 𝜃𝛿 ∓ 𝑢001 (𝑥) ∓ 𝑢 2 (𝑥) = 𝜃𝛿 ≥ 0,

and
 𝑇
𝐿 2𝜓 ± (𝑥) = 𝐿 2 𝜓1± (𝑥), 𝜓2± (𝑥) = −𝜖𝜓2±00 (𝑥) + 𝑏(𝑥)𝜓2 (𝑥) + 𝑐(𝑥)𝜓1 (𝑥)
𝑥2
   
= ∓𝜖 (𝑢 2 (𝑥)) + 𝑏(𝑥)(𝜃 ± 𝑢 2 (𝑥)) + 𝑐(𝑥) 𝜃 (1 + 𝛿) 1 −
00
± 𝑢 1 (𝑥)
2
𝑥2
 
= ∓𝜖𝑢 2 (𝑥) + 𝑏(𝑥)𝜃 ± 𝑏(𝑥)𝑢 2 (𝑥) + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
00
± 𝑐(𝑥)𝑢 1 (𝑥)
2
𝑥2
 
= ∓𝜖𝑢 2 (𝑥) ± 𝑏(𝑥)𝑢 2 (𝑥) ± 𝑐(𝑥)𝑢 1 (𝑥) + 𝑏(𝑥)𝜃 + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
00
2
2
 
𝑥
= 𝑓 (𝑥) + 𝑏(𝑥)𝜃 + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
2
2 − 𝑥2
  
≥ 0 + 𝜌𝜃 + −𝛾𝜃 (1 + 𝛿) ≥ 𝜌𝜃 − 2𝛾𝜃 ≥ 0.
2

This implies that L𝜓

𝜓 ± (𝑥) ≥ 0. Furthermore, we have 𝜓 ± (0) ≥ 0 and 𝜓 ± (1) ≥ 0. Thus,
the result follows by using the maximum principle in Lemma 4.2.1. 
105 4. Reaction-Diffusion Problem of Fourth-order

The next lemma presents apriori bounds on the solution u and its derivatives.

Lemma 4.2.3. For 𝑟 = 0, . . . , 6, the solution u = 𝑢 1 (𝑥), 𝑢 2 (𝑥) of the coupled

𝑇

system of equations (4.4) satisfies

!√︂ √︂ !!!
𝑏(0) 𝑏(1)
𝑢 1(𝑟) (𝑥) ≤ 𝐶 1+𝜖 (1− 𝑟2 )
exp − 𝑥 + exp − (1 − 𝑥) , and(4.6)
𝜖 𝜖
√︂ ! √︂ !!!
𝑏(0) 𝑏(1)
𝑢 2(𝑟) (𝑥) ≤ 𝐶 1 + 𝜖 (− 2 ) exp − 𝑥 + exp − (4.7)
𝑟
(1 − 𝑥) .
𝜖 𝜖

Proof. The proof is by induction on 𝑟. For 𝑟 = 0, the result follows from the stability
result in Lemma 4.2.2. Since kuk ≤ 𝐶 clearly implies that |𝑢 1 (𝑥)| ≤ 𝐶 and |𝑢 2 (𝑥)| ≤ 𝐶.
Assume that the result is true for all 𝑝 ≤ 𝑟. Then, for 𝑟 > 0, differentiate the differential
equation
−𝜖𝑢002 (𝑥) + 𝑏(𝑥)𝑢 2 (𝑥) + 𝑐(𝑥)𝑢 1 (𝑥) = 𝑓 (𝑥)

𝑟 times repeatedly to get

00
−𝜖 𝑢 2(𝑟) (𝑥) + 𝑏(𝑥)𝑢 2(𝑟) (𝑥) = 𝑔(𝑥),

where
𝑟−1
(𝑟) (𝑟) ©𝑟 ª ∑︁ ©𝑟 ª
(𝑠)
𝑔(𝑥) = 𝑓 (𝑥) − 𝑐 (𝑥)𝑢 1 (𝑥) − ­ ® 𝑐 (𝑟−1) (𝑥)𝑢01 (𝑥) − ­ ® 𝑏 (𝑟−𝑠) (𝑥)𝑢 2 (𝑥) +
«1¬ 𝑠=0 « 𝑠 ¬

𝑟−2
∑︁ © 𝑟 ª (𝑟−𝑠−2)
­ ®𝑐 (𝑥)𝑢 2(𝑠) (𝑥),
𝑠=0 « 𝑠 + 2¬

using 𝑢001 (𝑥) = −𝑢 2 (𝑥).

Using Lemma 4.2.2 and (4.7), for 𝑝 ≤ 𝑟, we get
√︂ ! √︂ !!!
(𝑟+2) 𝑟 +2 𝑏(0) 𝑏(1)
𝑢 2 (𝑥) ≤ 𝐶 1 + 𝜖 −( 2 ) exp − 𝑥 + exp − (1 − 𝑥) .
𝜖 𝜖

To obtain the desired bounds for 𝑝 = 𝑟 + 1, we use the following result

2 𝜇 0
G0 ≤ G + G 𝐼, (4.8)
𝐼 𝜇 𝐼 2

where 𝐼 = [ 𝜒, 𝜒 + 𝜇] is an arbitrary interval with 𝜇 > 0 and G ∈ 𝐶 2 (𝐼). Thus, by

4.2. The Continuous Problem 106

applying (4.8) with G = 𝑢 2(𝑟) , 𝐼 ⊆ [0, 1] and 𝜇 = min{1, 𝜖 }, we get

√︂ ! √︂ !!!
−( 𝑟 +1 𝑏(0) 𝑏(1)
𝑢 2(𝑟+1) (𝑥) ≤ 𝐶 1+𝜖 2 ) exp − 𝑥 + exp − (1 − 𝑥) .
𝜖 𝜖

Using 𝑢001 (𝑥) = −𝑢 2 (𝑥) and −𝜖𝑢002 (𝑥) + 𝑏(𝑥)𝑢 2 (𝑥) + 𝑐(𝑥)𝑢 1 (𝑥) = 𝑓 (𝑥), we have
√︂ ! √︂ !!!
(𝑟+1) ( 1−𝑟 𝑏(0) 𝑏(1)
𝑢 1 (𝑥) ≤ 𝐶 1 + 𝜖 2 exp − )
𝑥 + exp − (1 − 𝑥) .
𝜖 𝜖

This completes the proof. 

4.2.2 Solution Decomposition

For the analysis of numerical discretization technique, the solution u of (4.4) is

decomposed into the sum of a smooth part v and a layer part w, i.e., u = v + w. A
higher-order decomposition of the analytical solution using minimum regularity of
the coefficient functions 𝑏(𝑥), 𝑐(𝑥) and 𝑓 (𝑥) can be found in Linß[169], where the
smooth part v = (𝑣 1 , 𝑣 2 )𝑇 satisfies

Lv(𝑥) = 𝑓 (𝑥), 𝑥 ∈ Ω, 




𝑇

𝑇  ,
v(0) = 𝑢 1 (0), 𝑓 (0) − 𝑐(0)𝑢 1 (0) 𝑏(0) , v(1) = 𝑢 1 (1), 𝑓 (1) − 𝑐(1)𝑢 1 (1) 𝑏(1) 
(4.9)


and the layer part w = (𝑤 1 , 𝑤 2 )𝑇 satisfies

Lw(𝑥) = 0,

𝑥 ∈ Ω, 
(4.10)


.
w(0) = u(0) − v(0), w(1) = u(1) − v(1) 


This decomposition gives us a crude estimate of the derivative bounds on v and w
as shown in the following lemma.

Lemma 4.2.4. Let the coefficient functions 𝑏(𝑥), 𝑐(𝑥), 𝑓 (𝑥) ∈ 𝐶 4 (Ω) satisfy the as-
sumptions (4.2)–(4.3). Then, the coupled system of equations (4.4) possesses a unique
solution u(𝑥) such that u(𝑥) = v(𝑥) + w(𝑥). The smooth part v(𝑥) = 𝑣 1 (𝑥), 𝑣 2 (𝑥)

𝑇

satisfies (4.9) and

 
𝑣 1(𝑟) ≤ 𝐶, 𝑣 2(𝑟) (𝑥) ≤ 𝐶 1 + 𝜖 (2− 2 ) , 𝑟 = 0, . . . , 6, (4.11)
𝑟
107 4. Reaction-Diffusion Problem of Fourth-order

whereas the layer part w(𝑥) = 𝑤 1 (𝑥), 𝑤 2 (𝑥) satisfies (4.10) and

𝑇

√︂ ! √︂ !!
1−𝑟 𝑏(0) 𝑏(1) 
(𝑟)
𝑤 1 (𝑥) ≤ 𝐶𝜖 ( 2 ) exp − 𝑥 + exp −

(1 − 𝑥) 


𝜖 𝜖 
(4.12)


√︂ ! √︂ !!
(𝑟) 𝑏(0) 𝑏(1) 
𝑤 2 (𝑥) ≤ 𝐶𝜖 −( 2 ) exp − 𝑥 + exp − (1 − 𝑥) , 𝑟 = [0, 6]. 
𝑟 


𝜖 𝜖 

Proof. Using the matched asymptotic expansion technique, we define the smooth
component v(𝑥) as

v(𝑥) = v0 (𝑥) + 𝜖v1 (𝑥) + 𝜖 2 v∗1 (𝑥) = 𝑣 01 (𝑥) + 𝜖𝑣 11 (𝑥) + 𝜖 2 𝑣 ∗21 (𝑥), 𝑣 02 (𝑥) + 𝜖𝑣 12 (𝑥) + 𝜖 2 𝑣 ∗22 (𝑥)

𝑇

Substituting v(𝑥) in (4.4) and equating like powers of 𝜖 we get

−𝑣 0001 (𝑥) − 𝑣 02 (𝑥) = 0, −𝑣 0011 (𝑥) − 𝑣 12 (𝑥) = 0,

 
, , and
𝑏(𝑥)𝑣 02 (𝑥) + 𝑐(𝑥)𝑣 01 (𝑥) = 𝑓 (𝑥), 𝑏(𝑥)𝑣 12 (𝑥) + 𝑐(𝑥)𝑣 11 (𝑥) = 𝑣 0002 (𝑥),

−𝑣 0021 (𝑥) − 𝑣 22 (𝑥) = 0


.
𝐿 2 v∗2 (𝑥) = 𝑣 0012 (𝑥) with v∗2 (0) = v∗2 (1) = 0

Using the above decomposition and Lemma 4.2.3, it can be clearly observed that
 
𝑣 1(𝑟) (𝑥) ≤ 𝐶, 𝑣 2(𝑟) (𝑥) ≤ 𝐶 1 + 𝜖 (2− 2 ) , 𝑟 = 0, . . . , 6.
𝑟

To find the bounds on the layer part w(𝑥), we further decompose it as w(𝑥) = w− (𝑥) + w+ (𝑥)
where w− (𝑥) = 𝑤 −1 (𝑥), 𝑤 −2 (𝑥) satisfies

𝑇

Lw− (𝑥) = 0, 𝑥 ∈ Ω; w− (0) = w(0), w− (1) = 0, (4.13)

and w+ (𝑥) = 𝑤 +1 (𝑥), 𝑤 +2 (𝑥) satisfies

𝑇

Lw+ (𝑥) = 0, 𝑥 ∈ Ω; w+ (0) = 0, w+ (1) = w(1). (4.14)

Using the method of matched asymptotic expansion again, we have
2𝑝+1
∑︁
and (4.15)
𝑘
−
w (𝑥) = 𝜖 2 w−𝑘 (𝑥) + 𝜖 𝑝+1 w∗−
2𝑝+2 (𝑥),
𝑘=0
2𝑝+1
∑︁
for 𝑚 = 1, 2. (4.16)
𝑘
w+ (𝑥) = 𝜖 2 w+𝑘 (𝑥) + 𝜖 𝑝+1 w∗+
2𝑝+2 (𝑥),
𝑘=0

For the layer component on the left w− , introduce the co-ordinate transformation
4.2. The Continuous Problem 108

√ √ √
𝜉 = 𝑥/ 𝜖. Use Taylor’s series expansion of 𝑏(𝜉 𝜖) and 𝑐(𝜉 𝜖) in terms of variable 𝜉
to define
𝑑2
𝐿 = − 2 + 𝑏(0)𝐼.
b
𝑑𝜉
By equating the coefficients of like powers of 𝜖 in (4.13), it is observed that w− (𝜉)
satisfies
𝑑 2 𝑤 0,1
0,

− − =

2
𝑤 0,2 


𝑑𝜉 ,
𝑏(0)𝑤 0,2 + 𝑐(0)𝑤 0,1 = 0 with w−0 (0) = u(0) − v0 (0) and lim w−0 (𝜉) = 0,  

𝜉→∞ 
2
𝑑 𝑤 1,1
0,

− 2
− 𝑤 1,2 = 


𝑑𝜉 



𝑏(0)𝑤 1,2 + 𝜉𝑏 (0)𝑤 0,2 + 𝑐(0)𝑤 1,1 + 𝜉𝑐 (0)𝑤 0,1 = 0
0 0

,

with w−1 (0) = u(0) − v1 (0) and lim w−1 (𝜉) = 0, 



𝜉→∞ 
2
𝑑 𝑤 𝑘,1 
− − = 0,

2
𝑤 𝑘,2



𝑑𝜉 


2 𝑘   
𝑖 𝑖
, and


𝑑 𝑤 𝑘−2,2 ∑︁ 𝜉 𝑖 − 𝜉 𝑖 −
= 𝑏 (0)𝑤 𝑘−𝑖,2 (𝜉) + 𝑐 (0)𝑤 𝑘−𝑖,1 (𝜉)
𝑑𝜉 2 𝑖=0
𝑖! 𝑖! 




with w 𝑘 (0) = u(0) − v 𝑘 (0), and lim w 𝑘 (𝜉) = 0, 
− − 


𝜉→∞ 
(2 𝑝+1)
∗− −( 𝑝+1) − − 
Lw2𝑝+2 (𝑥) = −𝜖 L(w0 + · · · + 𝜖 2 w2𝑝+1 )(𝑥), 


(2 𝑝+1) .
w∗−
2𝑝+2 (0) = 0, w ∗−
2𝑝+2 (1) = −𝜖 −( 𝑝+1)
(w −
0 + · · · + 𝜖 2 w −
2𝑝+1 )(1). 


Thus, from Lemma 4.2.3, we get
√︂ !
𝑏(0)
𝑤 −1 (0) < 𝐶, 𝑤 −2 (0) < 𝐶, 𝑤 −2 (1) < 𝐶 exp − .
𝜖

This implies that

√︂ ! √︂ !
1 𝑏(0) 𝑏(0)
𝑤 −1 (1) < 𝐶𝜖 exp −𝑥
2 , (𝑤 −2 ) (𝑟) (𝑥) < 𝐶𝜖 − 𝑟2
exp −𝑥 and
𝜖 𝜖
√︂ !
− (1−𝑟 ) 𝑏(0)
(𝑤 −1 ) (𝑟) (𝑥) < 𝐶𝜖 2 exp −𝑥 .
𝜖

A similar argument can be used to establish the bounds for the right side of the
boundary layer. Hence the proof. 
109 4. Reaction-Diffusion Problem of Fourth-order

4.2.3 Grid Structure

The presence of steep boundary layers in the solution of coupled system of equations
(4.4) poses a major obstacle in approximating the solution accurately. By and large,
the layer part of the solution is responsible for the layer (i.e., abrupt) behavior of
the solution. Since, a small perturbation parameter is present only in the differential
operator 𝐿 2 , the following monitor function is considered
1
𝑀 = 𝛼 + 𝑤 002 4
(4.17)
where 𝛼 is a positive constant independent of 𝑁. The floor 𝛼 plays a vital role in
strengthening the structure of the grid.

On limiting its values away from zero, the grid inhibits the clustering of points
within the layers and ensures their presence in the outer region as well. The adaptive
grid generation process involves estimation of posteriori error bound in the numerical
solution on a fixed ordinary grid. This error bound is equidistributed in the form of
a monitor function to generate an appropriate grid. This process is iterated until a
measure indicating the convergence is attained. It has been empirically demonstrated
by Beckett and Mackenzie [245] that choosing the floor 𝛼 in a skillful manner improves
the convergence of this iterative process. Based on the equidistribution principle, we
now consider the structure of the grid with a proper choice of the floor 𝛼.

Using the leading terms in the expansions (4.15) and (4.16), an approximate value
of 𝑤 002 is given by
 √︃ 
exp −𝑥 𝑥 ∈ 0, 12 , 

 𝜆0 𝑏(0) 𝑏(0)  



 𝜖 , 𝜖 


𝑤 002 (𝑥) ≈  √︃  
𝜆1 𝑏(1)
exp −(1 − 𝑥) 𝑏(1) , 𝑥 ∈ 12 , 1 , 
  

 
 𝜖 𝜖 
 
where the constants 𝜆 0 , 𝜆1 are independent of 𝑥 and 𝜖. Thus, we get
1
∫ 1 ∫
2
∫ 1
1 1 1
|𝑤 002 (𝑥)| 4 𝑑𝑥 = |𝑤 002 (𝑥)| 4 𝑑𝑥 + |𝑤 002 (𝑥)| 4 𝑑𝑥
1
0 0 2
1
1
∫ √︂ ! 4
2 𝜆 0 𝑏(0) 𝑏(0)
= exp −𝑥 𝑑𝑥 +
0 𝜖 𝜖
1
1
∫ √︂ ! 4
𝜆 1 𝑏(1) 𝑏(1)
exp −(1 − 𝑥) 𝑑𝑥
1 𝜖 𝜖
2
4.2. The Continuous Problem 110

1
 √︃  2
 √︃  1
1
4
exp − 4𝑥 𝑏(0)
𝜖
1
4
exp − (1−𝑥)
4
𝑏(1)
𝜖
𝜆 0 𝑏(0) 𝜆 1 𝑏(1)
=  √︃  +  √︃ 
𝜖 1 𝑏(0) 𝜖 1 𝑏(1)
−4 𝜖 4 𝜖
1
0
1   √︃  1   √︃ 2
4 4
= 4 𝜆0 𝜖
𝑏(0) 1 − exp − 18 𝑏(0)
𝜖 +4 𝜆1 𝜖
𝑏(1) 1 − exp − 18 𝑏(1)
𝜖
1 1 !
1 𝜆0 4 𝜆1 4
≈ 4𝜖 4 + ≡Ψ (4.18)
𝑏(0) 𝑏(1)
 √︃   √︃ 
since exp − 81 𝑏(0)
𝜖 → 0 and exp − 18 𝑏(1)
𝜖 → 0 as 𝜖 → 0.
Now using the equidistribution principle (2.2) with the monitor function (4.17) yields
the following relation
∫ 𝑥(𝜉)  1  ∫ 1  1 
𝛼+ 𝑑𝑥 = 𝜉𝑤 002 (𝑥)
𝛼+ 4
𝑑𝑥 𝑤 002 (𝑥) 4

0 0
∫ 𝑥(𝜉) 1
∫ 1 1
00
⇒ 𝛼𝑥(𝜉) + 4
𝑤 2 (𝑥) 𝑑𝑥 = 𝛼𝜉 + 𝑤 002 (𝑥) 4 𝑑𝑥
0 0
1 !
4
∫ 𝑥(𝜉) √︂
𝜆 0 𝑏(0) 𝑥 𝑏(0)
⇒ 𝛼𝑥(𝜉) + exp − 𝑑𝑥 = 𝜉 (𝛼 + Ψ)
𝜖 0
4 𝜖
 √︃  𝑥(𝜉)
1
4 exp − 4 𝑥 𝑏(0)
𝜖
𝜆 0 𝑏(0)
⇒ 𝛼𝑥(𝜉) + √︃ = 𝜉 (𝛼 + Ψ)
− 14 𝑏(0)
𝜖
𝜖
0
1 ! !
4
√︂
𝜆0 𝜖 𝑥(𝜉)𝑏(0)
⇒ 𝛼𝑥(𝜉) − 4 exp − − 1 = 𝜉 (𝛼 + Ψ)
𝑏(0) 4 𝜖
1 !!
4
√︂
𝜆0 𝜖 𝑥(𝜉) 𝑏(0)
⇒ 𝛼𝑥(𝜉) + 4 1 − exp − = 𝜉 (𝛼 + Ψ).
𝑏(0) 4 𝜖

Dividing by Ψ throughout, we get

1 !!
4
√︂
𝛼𝑥(𝜉) 4 𝜆 0 𝜖 𝑥(𝜉) 𝑏(0) 𝛼
+ 1 − exp − = 𝜉( + 1)
Ψ Ψ 𝑏(0) 4 𝜖 Ψ
111 4. Reaction-Diffusion Problem of Fourth-order

1
𝜆0 4 √︂ !!
𝛼𝑥(𝜉) 𝑏0 𝑥(𝜉) 𝑏(0) 𝛼
⇒ + 1 − exp − = 𝜉( + 1)
Ψ 1
𝜆0 4 𝜆1
1
4 4 𝜖 Ψ
𝑏0 + 𝑏1
√︂ !!
𝛼 𝑥(𝜉) 𝑏(0) 𝛼 
⇒ 𝑥(𝜉) + 𝜇0 1 − exp − =𝜉 +1 (4.19)
Ψ 4 𝜖 Ψ
 1 . 1 1 
4 4 4
where 𝜇0 = 𝜆0
𝑏(0)
𝜆0
𝑏(0) + 𝜆1
𝑏(1) and 𝑥(𝜉) ≤ 12 .
Similarly, for 𝑥(𝜉) > 21 , we have
√︂ !!
𝛼 (1 − 𝑥(𝜉)) 𝑏(1) 𝛼 
(1 − 𝑥(𝜉)) + 𝜇1 1 − exp − = (1 − 𝜉) +1 , (4.20)
Ψ 4 𝜖 Ψ
 1 . 1 1 
4 4 4
where 𝜇1 = 𝜆1
𝑏(1)
𝜆0
𝑏(0) + 𝜆1
𝑏(1) = 1 − 𝜇0 .

Thus, the non-uniform grid 𝑥𝑖 corresponding to an equidistant grid {𝜉𝑖 = is

 𝑁 𝑖 𝑁
𝑖=0 𝑁 𝑖=0
given by
√︂ !!
𝛼 𝑥𝑖 𝑏(0) 𝑖 𝛼  1
𝑥𝑖 + 𝜇0 1 − exp − = + 1 , for 𝑥𝑖 ≤ , (4.21)
Ψ 4 𝜖 𝑁 Ψ 2

and
√︂ !!
1

𝛼 (1 − 𝑥𝑖 ) 𝑏(1) 𝑖 𝛼 
(1 − 𝑥𝑖 ) + 𝜇1 1 − exp − = 1− + 1 , for 𝑥𝑖 > . (4.22)
Ψ 4 𝜖 𝑁 Ψ 2

Clearly a simultaneous solution of the non-linear algebraic equations (4.21)–(4.22)

generates the desired adaptive grid. For a suitable choice of the floor 𝛼, we now
illustrate the distribution of nodes of the adaptive grid in the following lemma.

Lemma 4.2.5. For 𝛼 = Ψ, the grid points of the non-uniform grid (4.21)–(4.22)
satisfy
√︂ √︂
𝜖 𝜖
𝑥 𝑘𝑙 < 4 log 𝑁 < 𝑥 𝑘 𝑙 +1 and 𝑥 𝑘 𝑟 −1 < 1 − 4 log 𝑁 < 𝑥 𝑘 𝑟
𝑏(0) 𝑏(1)

where
1
  √︂ 
𝜖
𝑘𝑙 = 𝜇0 (𝑁 − 1) + 4 𝑁 log 𝑁 ,
2 𝑏(0)
4.2. The Continuous Problem 112

1
  √︂ 
𝜖
𝑘𝑟 = 𝑁 − 𝜇1 (𝑁 − 1) + 4 𝑁 log 𝑁 + 1,
2 𝑏(1)

and [.] denotes the integral part of the term. Moreover, for a generic constant 𝐶 > 1,
we have  √︃ 
exp − 𝑥4𝑖 ≤ 𝐶𝑁 −1 ,
𝑏(0)
𝜖 𝑖 ≥ 𝑘 𝑙 − 1, 𝑥𝑖 ≤ 21 , and
 √︃ 
(1−𝑥𝑖 ) 𝑏(1)
exp − 4 𝜖 ≤ 𝐶𝑁 −1 , 𝑖 ≤ 𝑘 𝑟 , 𝑥𝑖 > 12 .
√︃
Proof. Similar to Lemma 2.2.3, put 𝑥𝑖 = 4 𝑏(0)
𝜖
log 𝑁 with 𝛼 = Ψ in (4.21). Solve it
for 𝑖 to get the requisite value of 𝑘 𝑙 . A likewise treatment of (4.22) yields the value of
𝑘𝑟 . 

The precise choice of the floor 𝛼 = Ψ is truly inspired by the well known piecewise
uniform Shishkin mesh. Consequently, the equidistributed grid and the apriori
mesh have some common key features. However, the earlier grid overpowers the
latter in terms of accuracy. Moreover, a careful examination conducted by Beckett
and Mackenzie [245], and Gupta and Kaushik [248] reveal the advancement of the
equidistributed grid in several aspects.
 
𝑘 𝑙 −1
After characterizing the boundary layer region {𝑥𝑖 }𝑖=0 and {𝑥𝑖 }𝑖=𝑘
𝑁
𝑟 +1
and the
 
outer region {𝑥𝑖 }𝑖=𝑘
𝑘𝑟
𝑙
, the following lemma determines the bounds on the grid width
in the boundary layer region.

Lemma 4.2.6. For the boundary layer region, the grid spacing ℎ𝑖 satisfies
√︃
 4𝐶 𝜖
𝑖 = 1, . . . , 𝑘 𝑙 ,

𝑏(0) ,


ℎ𝑖 < √︃
 4𝐶
 𝜖
𝑏(1) , 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁,

in addition to
 𝐶ℎ𝑖2 , 𝑖 = 1, . . . , 𝑘 𝑙 − 1,



|ℎ𝑖+1 − ℎ𝑖 | ≤
 2 , 𝑖 = 𝑘 + 1, . . . , 𝑁 − 1.
 𝐶ℎ𝑖+1 𝑟

Proof. Similar to Lemma 2.2.4, firstly, we deal with the left side of the boundary
layer region and use (4.21) to define 𝑥𝑖 < 𝑥𝑖 such that
√︂ !
𝑥𝑖 𝑏(0) 2𝑖
exp − =1− .
4 𝜖 𝜇0 𝑁
113 4. Reaction-Diffusion Problem of Fourth-order

Rearrangement the terms gives

2𝑖
√︂  
𝜖
𝑥𝑖 < 𝑥𝑖 = −4 log 1 − .
𝑏(0) 𝜇0 𝑁

Using 𝑥𝑖 in (4.21) again, we define 𝑥𝑖 < 𝑥𝑖 such that

1 2𝑖 2𝑖
√︂   √︂  
𝜖 𝜖
𝑥𝑖 > 𝑥𝑖 = −4 log 1 − +4 log 1 − .
𝑏(0) 𝜇0 𝑁 𝑏(0) 𝜇0 𝑁

Thus, for 𝑖 = 1, . . . , 𝑘 𝑙 , we have

√︃  
√︂ 2+4 𝜖
𝑏(0) 𝑁 log 𝜇0 𝑁
𝜇0 𝑁−2(𝑖−1)
√︂
𝜖 𝜖
ℎ𝑖 < 𝑥 𝑖 − 𝑥𝑖−1 = 4 log ­­1 + ® < 4𝐶
© ª
.
𝑏(0) 𝜇0 𝑁 − 2𝑖 ® 𝑏(0)
« ¬
For further analysis, we approximate ℎ𝑖 in terms of the computational coordinates 𝜉𝑖
to get
 
(1)
|ℎ𝑖+1 − ℎ𝑖 | 2 𝑥 𝜉𝜉 𝜃 𝑖 (1) (2)
2
≤   2 where 𝜃 𝑖 ∈ (𝜉𝑖−1 , 𝜉𝑖+1 ) and 𝜃 𝑖 ∈ (𝜉𝑖−1 , 𝜉𝑖 ) .
𝑥 𝜉 𝜃 𝑖(2)
ℎ𝑖

Substitute 𝛼 = Ψ in (4.19) to get

√︃  √︃ 
−16𝜇0 𝑏(0) exp − 4
√︃ 𝑥(𝜃) 𝑏(0)
8 𝑏(0)
𝜖 𝜖
𝜖
𝑥 𝜉 (𝜃) = √︃  √︃  and 𝑥 𝜉𝜉 (𝜃) =  √︃  √︃  3 .
4 𝑏(0)
𝜖
+ 𝜇0 exp − 4
𝑥(𝜃) 𝑏(0)
4 𝑏(0)
𝜖
+ 𝜇0 exp − 𝑥(𝜃) 𝑏(0)
𝜖 4 𝜖

Thus,
√︃  √︃  √︃  2
𝜇0 𝑏(0)
4 𝜖 + 𝜇0 exp − 4
𝑏(0) 𝑥 𝑖−1 𝑏(0)
|ℎ𝑖+1 − ℎ𝑖 | 𝜖 𝜖
≤ ≤ 𝐶.
ℎ𝑖2
 √︃  √︃  3
4 4 𝜖 + 𝜇0 exp − 4
𝑏(0) 𝑥 𝑖+1 𝑏(0)
𝜖

A similar argument can be used to establish the bounds for the right side of the
boundary layer region. Hence the proof. 

Extending this result to the entire domain, the following lemma determines the
bounds on the grid width.
4.3. The Discretization 114

Lemma 4.2.7. For all 𝑖 = 1, . . . , 𝑁, the grid spacing ℎ𝑖 satisfies

ℎ𝑖 ≤ 𝐶𝑁 −1 .

Proof. Similar to Lemma 2.2.6, we use the solution derivatives from Lemma 4.2.4
and (4.17) to obtain
1
𝑀 = 𝛼 + 𝑤 002 4
≥ 𝛼 = Ψ.

Using the solution derivatives from Lemma 4.2.4 and (4.18), we obtain
∫ 1
𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡 ≤ 𝐶𝛼.
0

Apply the equidistribution principle (2.1) to get

1
1
∫ 𝑥𝑖 ∫
𝛼ℎ𝑖 ≤ 𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡 = 𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡 ≤ 𝐶𝛼𝑁 −1 .
𝑥 𝑖−1 𝑁 0

Hence the proof. 

These bounds will be used to analyze the stability of the discrete operator in the
following section.

4.3 The Discretization

In this section, we present a hybrid difference scheme to discretize the coupled
𝑁
system of equations (4.4) on the equidistributed grid Ω𝐸 . The hybrid difference
scheme comprises of a fourth-order Hermite difference scheme in the boundary layer
region and the central difference method in the outer layer region. Then, the discrete
problem for the system of second order equations (4.4) takes the following form:
Determine U = (𝑈1 , 𝑈2 )𝑇 such that



−
𝐿 1𝑁 U𝑖 ≡ 𝑟 1,𝑖 𝑐
𝑈1,𝑖−1 + 𝑟 1,𝑖 +
𝑈1,𝑖 + 𝑟 1,𝑖 𝑈1,𝑖+1 − 𝑞𝑖−𝑈2,𝑖−1 − 𝑞𝑖𝑐𝑈2,𝑖 − 𝑞𝑖+𝑈2,𝑖+1 = 0,


−𝑈 + − +
 𝐿 2𝑁 U𝑖 ≡ 𝑟 2,𝑖 2,𝑖−1 + 𝑟 2,𝑖 𝑈2,𝑖 + 𝑟 2,𝑖 𝑈2,𝑖+1 + 𝑞 𝑖 𝑐𝑖−1𝑈1,𝑖−1 + 𝑞 𝑖 𝑐𝑖 𝑈1,𝑖 + 𝑞 𝑖 𝑐𝑖+1𝑈1,𝑖+1
𝑐 𝑐



LN Ui = fi ⇔


 = 𝑞𝑖− 𝑓𝑖−1 + 𝑞𝑖𝑐 𝑓𝑖 + 𝑞𝑖+ 𝑓𝑖+1 ,



𝑈 = 𝛼 ,
 1,0 1 𝑈1,𝑁 = 𝛼2 , 𝑈2,0 = 𝛼3 , 𝑈2,𝑁 = 𝛼4 ,
(4.23)
115 4. Reaction-Diffusion Problem of Fourth-order

where the finite difference operator discretizing the continuous problem is given by
L𝑁 = 𝐿 1𝑁 , 𝐿 2𝑁 and fi = (0, 𝑓𝑖 )𝑇 .

𝑇

The coefficients 𝑟 𝑚,𝑖

∗ , 𝑚 = 1, 2, 𝑖 = 1, . . . , 𝑁 − 1, ∗ = −, 𝑐, + are defined by

−2 2 −2
−
𝑟 1,𝑖 = , 𝑐
𝑟 1,𝑖 = , +
𝑟 1,𝑖 = , (4.24)
ℎ𝑖 (ℎ𝑖 + ℎ𝑖+1 ) ℎ𝑖 ℎ𝑖+1 ℎ𝑖+1 (ℎ𝑖 + ℎ𝑖+1 )

− −2𝜖 2𝜖 −2𝜖
𝑟 2,𝑖 = + 𝑞𝑖− 𝑏𝑖−1 , 𝑐
𝑟 2,𝑖 = + 𝑞𝑖𝑐 𝑏𝑖 , +
𝑟 2,𝑖 = + 𝑞𝑖+ 𝑏𝑖+1 .
ℎ𝑖 (ℎ𝑖 + ℎ𝑖+1 ) ℎ𝑖 ℎ𝑖+1 ℎ𝑖+1 (ℎ𝑖 + ℎ𝑖+1 )
(4.25)
The coefficients of the fourth-order Hermite difference scheme satisfy the normaliza-
tion condition 𝑞𝑖− + 𝑞𝑖𝑐 + 𝑞𝑖+ = 1 to ensure that the method is exact for polynomials
up to degree four. Thus, on the basis of the location of grid point 𝑥𝑖 , we define the
coefficients 𝑞𝑖∗ , 𝑖 = 1, . . . , 𝑁 − 1, ∗ = −, 𝑐, +, as follows

(i) In the boundary layer region, the coefficients 𝑞𝑖∗ , 𝑖 = {1, . . . , 𝑘 𝑙 − 1} ∪ {𝑘 𝑟 +

1, . . . , 𝑁 − 1}, ∗ = −, 𝑐, + of the fourth-order Hermite difference scheme are given
by

ℎ𝑖2 + ℎ𝑖 ℎ𝑖+1 − ℎ𝑖+1

2 ℎ𝑖2 + 3ℎ𝑖 ℎ𝑖+1 + ℎ𝑖+1
2 2 +ℎ ℎ
ℎ𝑖+1 𝑖 𝑖+1 − ℎ𝑖
2
𝑞𝑖− = , 𝑞𝑖𝑐 = , 𝑞𝑖+ = .
6ℎ𝑖 (ℎ𝑖 + ℎ𝑖+1 ) 6ℎ𝑖 ℎ𝑖+1 6ℎ𝑖+1 (ℎ𝑖 + ℎ𝑖+1 )
(4.26)

(ii) Outside the layer region, the coefficients 𝑞𝑖∗ , 𝑖 = {𝑘 𝑙 , . . . , 𝑘 𝑟 }, ∗ = −, 𝑐, +, depend

on a relation between ℎ𝑚𝑎𝑥 , the perturbation parameter 𝜖, and a positive constant
k independent of 𝑁 and 𝜖 such that

(a) If kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖, a classical central difference scheme is used and the
coefficients 𝑞𝑖∗ , 𝑖 = {𝑘 𝑙 , . . . , 𝑘 𝑟 }, ∗ = −, 𝑐, +, associated with the scheme are
given by
𝑞𝑖− = 0, 𝑞𝑖𝑐 = 1, 𝑞𝑖+ = 0. (4.27)

(b) If kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖, the fourth-order Hermite difference scheme is used

whose coefficients 𝑞𝑖∗ , 𝑖 = {𝑘 𝑙 , . . . , 𝑘 𝑟 }, ∗ = −, 𝑐, +, are given by the relation
(4.26).

Using these values of the coefficients 𝑟 𝑚,𝑖

∗ and 𝑞 ∗ , 𝑖 = 1, . . . , 𝑁 − 1, we obtain the
𝑖
hybrid difference discretization technique (4.23) for the coupled system of equations
4.3. The Discretization 116

(4.4) on the adaptive grid. To improve the order of convergence of the discretization,
 
𝑘 𝑙 −1
the fourth-order Hermite difference scheme is applied to the grid points 𝑥𝑖 𝑖=0
 
and 𝑥𝑖 𝑖=𝑘 𝑟 +1 lying in the boundary layer region, and to {𝑥𝑖 }𝑖=𝑘 in the outer layer
𝑁 𝑘𝑟
𝑙

region whenever kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖. The fourth-order approximation fails to generate

non-positive off diagonal entries in the outer layer region when kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖, so
we use more stable classical central difference approximation there. Consequently,
the discrete hybrid operator L𝑁 associated with discretization technique (4.23) yields
positive diagonal and non-positive off diagonal entries for a suitable choice of the
constant k. The following lemma establishes that the coefficient matrix of the
proposed hybrid difference discretization technique is an M-matrix.
1
Lemma 4.3.1. Let k = 6 and 𝑁0 be the smallest positive integer such that

2Ψ2 k𝑏k ∞ < 3𝜁 𝑁02 where 𝜁 = max{𝜇0 , 𝜇1 }. (4.28)

Then, for every 𝑁 ≥ 𝑁0 , the coefficients of the discrete hybrid operator L𝑁 defined
in (4.23) satisfy

𝑐
𝑟 2,𝑖 > 0, −
𝑟 2,𝑖 ≤ 0, +
𝑟 2,𝑖 ≤ 0, 𝑐
𝑟 2,𝑖 −
+ 𝑟 2,𝑖 +
+ 𝑟 2,𝑖 > 0, 𝑖 = 1, . . . , 𝑁 − 1.

Proof. The coefficients of the discrete operator 𝐿 1𝑁 associated with the fourth-order
Hermite difference scheme given in (4.24) clearly satisfy

𝑐
𝑟 1,𝑖 > 0, −
𝑟 1,𝑖 ≤ 0, +
𝑟 1,𝑖 ≤ 0, 𝑐
𝑟 1,𝑖 −
+ 𝑟 1,𝑖 +
+ 𝑟 1,𝑖 > 0, 𝑖 = 1, . . . , 𝑁 − 1,

since ℎ𝑖 > 0, ∀𝑖 = 1, . . . , 𝑁. We now partition the coefficients of the discrete hybrid

operator 𝐿 2𝑁 on the basis of location of the grid point 𝑥𝑖 as follows
  𝑁−1 
𝑘 𝑙 −1
(i) Consider the boundary layer region 𝑥𝑖 𝑖=1 and 𝑥𝑖 𝑖=𝑘 𝑟 +1 where the coeffi-
cients of the fourth-order Hermite difference scheme are given by (4.25) and
(4.26). Similar to Lemma 2.2.1, we use Lemmas 4.2.5, 4.2.6 and (4.28) to get

𝑐
𝑟 2,𝑖 > 0, −
𝑟 2,𝑖 ≤ 0, +
𝑟 2,𝑖 ≤ 0, 𝑖 = {1, . . . , 𝑘 𝑙 − 1} ∪ {𝑘 𝑟 + 1, . . . , 𝑁 − 1}.

Moreover, using (4.2), a rearrangement of the row sum is given by

𝑐
𝑟 2,𝑖 −
+ 𝑟 2,𝑖 +
+ 𝑟 2,𝑖 > 𝜌 > 0.
117 4. Reaction-Diffusion Problem of Fourth-order

 
(ii) In the outer layer region , the coefficients of the hybrid difference
𝑘𝑟
𝑥𝑖 𝑖=𝑘 𝑙
scheme are defined on the basis of a relation between ℎ𝑚𝑎𝑥 and 𝜖 such that

(a) if kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖, then the coefficients of the central difference approxima-
tion given by (4.25) and (4.27) clearly satisfy

𝑐
𝑟 2,𝑖 > 0, −
𝑟 2,𝑖 ≤ 0, +
𝑟 2,𝑖 ≤ 0, 𝑐
𝑟 2,𝑖 −
+ 𝑟 2,𝑖 +
+ 𝑟 2,𝑖 > 0, 𝑖 = 𝑘𝑙 , . . . , 𝑘𝑟 .

(b) if kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖, we again use fourth-order Hermite difference scheme

whose coefficients are given by (4.25) and (4.26). Then, similar to Lemma
1
2.3.1, the application of inequality kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖 for k = 6 concludes the
result. 

We now show that the discrete hybrid operator L𝑁 satisfies the following discrete
maximum principle.

Lemma 4.3.2. For the discrete hybrid operator L𝑁 , assume that a grid func-
tion V = (𝑉1 , 𝑉2 )𝑇 satisfies V(𝑥 0 ) ≥ 0 and V𝑥(𝑁) ≥ 0. Then, L𝑁 V(𝑥𝑖 ) ≥ 0,
∀ 𝑖 = 1, . . . , 𝑁 − 1 implies that V(𝑥𝑖 ) ≥ 0, ∀ 𝑖 = 0, . . . , 𝑁.

Proof. If the conclusion of the lemma is false, then we can choose a 𝑘 such that
 
min 𝑉𝑚 (𝑥 𝑘 ) = min min 𝑉𝑚 (𝑥 𝑘 ) < 0.
 
𝑚=1,2 𝑚=1,2 𝑘

Without loss of generality, assume that 𝑉1 (𝑥 𝑘 ) ≤ 𝑉2 (𝑥 𝑘 ). Clearly, 𝑘 ∉ {0, 𝑁 }. Since,

− < 0, 𝑟 + < 0 and 𝑞 − > 0, 𝑞 𝑐 > 0, 𝑞 + > 0, therefore
𝑟 1,𝑘 1,𝑘 𝑘 𝑘 𝑘

−
𝐿 1𝑁 V(𝑥 𝑘 ) ≡ 𝑟 1,𝑘 𝑐
𝑉1,𝑘−1 + 𝑟 1,𝑘 +
𝑉1,𝑘 + 𝑟 1,𝑘 𝑉1,𝑘+1 − 𝑞 −𝑘 𝑉2,𝑘−1 − 𝑞 𝑐𝑘 𝑉2,𝑘 − 𝑞 +𝑘 𝑉2,𝑘+1 < 0,

which contradicts the hypothesis that L𝑁 V(𝑥 𝑘 ) ≥ 0. Hence, V(𝑥𝑖 ) ≥ 0, 0 ≤ 𝑖 ≤ 𝑁. 

An immediate consequence of this discrete maximum principle is the following

𝜖-uniform stability result for the discrete hybrid operator L𝑁 .

Lemma 4.3.3. If any grid function V satisfies the discrete problem (4.23), then
there is a constant 𝐶 independent of 𝑁 and 𝜖 such that

kVk Ω 𝑁 ≤ 𝐶 kL𝑁 Vk Ω 𝑁 .
𝐸 𝐸
4.4. Error Analysis 118

 
Proof. Similar to Lemma 4.2.2, let 𝜃 = 𝐶 max V(0) , V(1) , max L𝑁 V(𝑥 𝑖) .
1≤𝑖≤𝑁−1
For 𝑖 = 0, . . . , 𝑁, define two mesh functions 𝜓 (𝑥) = 𝜓1 (𝑥), 𝜓2 (𝑥) as
± ± ±

𝑇

𝑥2
 
𝜓1± (𝑥) = 𝜃 (1 + 𝛿) 1 − ± 𝑈1 (𝑥) and 𝜓2± (𝑥) = 𝜃 ± 𝑈2 (𝑥), where 0 < 𝛿 << 1.
2

Clearly, 𝜓 ± (𝑥0 ) ≥ 0 and 𝜓 ± (𝑥 𝑁 ) ≥ 0. Then, for a proper choice of 𝐶, L𝑁 𝜓 ± (𝑥) =

𝐿 1𝑁 , 𝐿 2𝑁 𝜓 ± (𝑥) ≥ 0.

𝑇

 𝑇
Since, 𝐿 1𝑁 𝜓 ± (𝑥) = 𝐿 1𝑁 𝜓1± (𝑥), 𝜓2± (𝑥) = −𝜓1±00 (𝑥) − 𝜓2± (𝑥)
= 𝜃 (1 + 𝛿) ∓ 𝑈100 (𝑥) − (𝜃 ± 𝑈2 (𝑥)) = 𝜃 (1 + 𝛿) ∓ 𝑈100 (𝑥) − 𝜃 ∓ 𝑈2 (𝑥)
= 𝜃𝛿 ∓ 𝑈100 (𝑥) ∓ 𝑈2 (𝑥) = 𝜃𝛿 ≥ 0, and
 𝑇
𝐿 2 𝜓 (𝑥) = 𝐿 2 𝜓1 (𝑥), 𝜓2 (𝑥) = −𝜖𝜓2±00 (𝑥) + 𝑏(𝑥)𝜓2 (𝑥) + 𝑐(𝑥)𝜓1 (𝑥)
𝑁 ± 𝑁 ± ±

𝑥2
   
= ∓𝜖 (𝑈2 (𝑥)) + 𝑏(𝑥)(𝜃 ± 𝑈2 (𝑥)) + 𝑐(𝑥) 𝜃 (1 + 𝛿) 1 −
00
± 𝑈1 (𝑥)
2
𝑥2
 
= ∓𝜖𝑈2 (𝑥) + 𝑏(𝑥)𝜃 ± 𝑏(𝑥)𝑈2 (𝑥) + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
00
± 𝑐(𝑥)𝑈1 (𝑥)
2
𝑥2
 
= ∓𝜖𝑈2 (𝑥) ± 𝑏(𝑥)𝑈2 (𝑥) ± 𝑐(𝑥)𝑈1 (𝑥) + 𝑏(𝑥)𝜃 + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
00
2
2
 
𝑥
= 𝐿 2𝑁 U(𝑥) + 𝑏(𝑥)𝜃 + 𝑐(𝑥)𝜃 (1 + 𝛿) 1 −
2
2−𝑥 2
  
≥ 0 + 𝜌𝜃 + −𝛾𝜃 (1 + 𝛿) ≥ 𝜌𝜃 − 2𝛾𝜃 ≥ 0.
2

This implies that L𝑁 𝜓 ± (𝑥) ≥ 0. Thus, the result follows by using the discrete
maximum principle in Lemma 4.3.2. 

4.4 Error Analysis

In this section, we analyze the truncation error associated with the numerical solution
obtained by using the hybrid discretization technique (4.23) on the equidistributed
grid. Similar to the continuous solution, we decompose the discrete approximate
solution as U = V + W. The smooth component V satisfies

L𝑁 V𝑖 = f𝑖 , 𝑖 = 1, . . . , 𝑁 − 1; V0 = v(0) and V𝑁 = v(1),

119 4. Reaction-Diffusion Problem of Fourth-order

while the layer component W satisfies

L𝑁 W𝑖 = 0, 𝑖 = 1, . . . , 𝑁 − 1; W0 = w(0) and W𝑁 = w(1).

At each grid point 𝑥𝑖 , the error in the numerical approximation 𝑈𝑖 can be estimated
by using the inequality

U𝑖 − u(𝑥𝑖 ) ≤ V𝑖 − v(𝑥𝑖 ) + W𝑖 − w(𝑥𝑖 ) . (4.29)

Thus, the consistency error estimation of the numerical solution relies on the respec-
tive error bounds of the smooth and layer parts. We begin our analysis with the
smooth component of the solution.

Lemma 4.4.1. For 𝑖 = 1, . . . , 𝑁 − 1, the smooth component V satisfies

𝐿 𝑁 (V − v)(𝑥𝑖 ) ≤ 𝐶𝑁 −4 .

Proof. Based on the location of the grid points 𝑥𝑖 , partition the domain as above to
estimate the truncation error in the smooth part V as follows

(a) For 𝑖 = 1, . . . , 𝑘 𝑙 − 1 and 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁 − 1 , the hybrid discretization

 

technique (4.23) is defined by the fourth-order Hermite difference scheme within

the boundary layer region. Then, by applying Taylor’s expansion, we obtain

𝐿 1𝑁 (V − v)(𝑥𝑖 ) = (𝐿 1 − 𝐿 1𝑁 )v(𝑥𝑖 )
≤ 𝐶 ℎ𝑖+1 − ℎ𝑖 (ℎ𝑖+1 + ℎ𝑖 ) 2 𝑣 1(𝑣) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶 (ℎ𝑖4 + ℎ𝑖+1
4
) 𝑣 1(𝑣𝑖) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
,
and 𝐿 2𝑁 (V − v)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2𝑁 )v(𝑥𝑖 )
≤ 𝐶𝜖 ℎ𝑖+1 − ℎ𝑖 (ℎ𝑖+1 + ℎ𝑖 ) 2 𝑣 2(𝑣) (𝑥) [𝑥𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶𝜖 (ℎ𝑖4 + ℎ𝑖+1
4
) 𝑣 2(𝑣𝑖) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
.

By combining the bounds of Lemmas 4.2.4, 4.2.6 and 4.2.7 with the assumption
√
𝜖 << 𝑁 −1 , it follows that

𝐿 1𝑁 (V − v)(𝑥𝑖 ) ≤ 𝐶𝑁 −4 and 𝐿 2𝑁 (V − v)(𝑥𝑖 ) ≤ 𝐶𝑁 −4 .

(b) For 𝑖 = 𝑘 𝑙 , . . . , 𝑘 𝑟 , the hybrid discretization technique in the outer layer region


depends on a relation between ℎ𝑚𝑎𝑥 and the perturbation parameter 𝜖 such that
4.4. Error Analysis 120

(i) Whenever kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖, the classical central difference scheme is used
to approximate the solution in the outer layer region. Clearly under this
condition, the central difference scheme is restricted to the singularly per-
turbed differential equation corresponding to 𝐿 2𝑁 only. Thus, using the
bounds of Lemmas 4.2.4 and 4.2.7 in the Taylor’s expansion of 𝑣 2 yields

𝐿 2𝑁 (V − v)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2𝑁 )v(𝑥𝑖 ) ≤ 𝐶𝜖 ℎ𝑖2 𝑣 2(𝑖𝑣) (𝑥) [𝑥𝑖−1 , 𝑥 𝑖+1 ]

≤ 𝐶𝑁 −4 .

This proves the result for the case kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖.
(ii) When kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖, in the outer layer region, the fourth-order Hermite
difference scheme is used again to approximate the solution. Then, an
analogous application of the Taylor’s expansion combined with the estimates
of Lemmas 4.2.4 and 4.2.7 gives

𝐿 1𝑁 (V − v)(𝑥𝑖 ) = (𝐿 1 − 𝐿 1𝑁 )v(𝑥𝑖 ) ≤ 𝐶ℎ𝑖4 𝑣 1(𝑣𝑖) (𝑥) [𝑥𝑖−1 , 𝑥𝑖+1 ]

≤ 𝐶𝑁 −4
and 𝐿 2𝑁 (V − v)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2𝑁 )v(𝑥𝑖 ) ≤ 𝐶𝜖 ℎ𝑖4 𝑣 2(𝑣𝑖) (𝑥) [𝑥𝑖−1 , 𝑥𝑖+1 ]
≤ 𝐶𝑁 −4 .

Thus, combining the obtained estimates completes the proof. 

Lemma 4.4.2. For 𝑖 = 1, . . . , 𝑁 − 1, the error associated with the layer component
W satisfies
𝐿 𝑁 (W − w)(𝑥𝑖 ) ≤ 𝐶𝑁 −4 .

Proof. As for the smooth component, the truncation error in the layer component W
can be computed by treating different regions of the domain on the basis of location
of the grid points 𝑥𝑖 . Then

(a) For 𝑖 = 1, . . . , 𝑘 𝑙 − 1 and 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁 − 1 , the fourth-order Hermite

 

difference scheme is used to define the hybrid discretization technique (4.23)

within the boundary layer region. A direct computation using Taylor’s expansion
yields

𝐿 1𝑁 (W − w)(𝑥𝑖 ) = (𝐿 1 − 𝐿 1𝑁 )w(𝑥𝑖 )
≤ 𝐶 ℎ𝑖+1 − ℎ𝑖 (ℎ𝑖+1 + ℎ𝑖 ) 2 𝑤 1(𝑣) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶 (ℎ𝑖4 4
+ ℎ𝑖+1 ) 𝑤 1(𝑣𝑖) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]

≤ 𝐶ℎ𝑖4 𝑤 1(𝑣) (𝑥) [𝑥𝑖−1 , 𝑥 𝑖+1 ]

+ 𝐶ℎ𝑖4 𝑤 1(𝑣𝑖) (𝑥) [𝑥𝑖−1 , 𝑥𝑖+1 ] .
121 4. Reaction-Diffusion Problem of Fourth-order

Using the bounds of Lemmas 4.2.4, 4.2.6 and 4.2.7 for the boundary layer region
on the left side of the domain, we obtain
! √︂ √︂ !
− 32 𝑏(0) 𝑏(0)
𝐿 1𝑁 (W − w)(𝑥𝑖 ) ≤ 𝐶ℎ𝑖4 𝜖 exp −𝑥𝑖 + 𝐶ℎ𝑖4 𝜖 −2 exp −𝑥𝑖
𝜖 𝜖
∫ 𝑥𝑖 √︂ ! !4
𝑡 𝑏(0)
≤ 𝐶𝜖 −2 exp − 𝑑𝑡
𝑥𝑖−1
4 𝜖
∫ 𝑥𝑖 !4
1
≤ 𝐶𝜖 −2 𝜖 4 𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡
𝑥𝑖−1
≤ 𝐶𝜖 −1 Ψ4 𝑁 −4
≤ 𝐶𝑁 −4 .

For the truncation error associated with 𝐿 2𝑁 , a similar calculation as above yields

𝐿 2𝑁 (W − w)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2𝑁 )w(𝑥𝑖 )
≤ 𝐶𝜖 ℎ𝑖+1 − ℎ𝑖 (ℎ𝑖+1 + ℎ𝑖 ) 2 𝑤 2(𝑣) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶𝜖 (ℎ𝑖4 4
+ ℎ𝑖+1 ) 𝑤 2(𝑣𝑖) (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]

≤ 𝐶𝜖 ℎ𝑖4 𝑤 2(𝑣) (𝑥) [𝑥𝑖−1 , 𝑥 𝑖+1 ]

+ 𝐶𝜖 ℎ𝑖4 𝑤 2(𝑣𝑖) (𝑥) [𝑥𝑖−1 , 𝑥𝑖+1 ] .

Using the equidistribution principle (2.1), Lemmas 4.2.4, 4.2.6 and 4.2.7 for the
boundary layer region on the left side of the domain, we get
√︂ ! √︂ !
3 𝑏(0) 𝑏(0)
𝐿 2𝑁 (W − w)(𝑥𝑖 ) ≤ 𝐶ℎ𝑖4 𝜖 − 2 exp −𝑥𝑖 + 𝐶ℎ𝑖4 𝜖 −2 exp −𝑥𝑖
𝜖 𝜖
∫ 𝑥𝑖 √︂ ! ! 4
𝑡 𝑏(0)
≤ 𝐶𝜖 −2
exp − 𝑑𝑡
𝑥𝑖−1
4 𝜖
∫ 𝑥 !4
𝑖
1
≤ 𝐶𝜖 −2 𝜖 4 𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡
𝑥𝑖−1
≤ 𝐶𝜖 −1 Ψ4 𝑁 −4
≤ 𝐶𝑁 −4 .

By imitating the above steps for the boundary layer region on the right side of
the domain, we have

𝐿 𝑁 (W − w)(𝑥𝑖 ) ≤ 𝐶𝑁 −4 for 𝑖 = 1, . . . , 𝑘 𝑙 − 1 and 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁 − 1 .

 

(b) For 𝑖 = 𝑘 𝑙 , . . . , 𝑘 𝑟 , the truncation error in the layer part can be estimated

4.4. Error Analysis 122

through the same calculation as above for the outer layer region where on the
basis of a relation between ℎ𝑚𝑎𝑥 and the perturbation parameter 𝜖, the hybrid
discretization technique is given by

(i) The fourth-order Hermite difference scheme whenever kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖.

Then, using Taylor’s expansion similar as above, we get

𝐿 1𝑁 (W − w)(𝑥𝑖 ) = (𝐿 1 − 𝐿 1𝑁 )w(𝑥𝑖 )
≤ 𝐶 𝑤 001 (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶ℎ𝑖4 𝑤 1(𝑣𝑖) (𝑥)
[𝑥 𝑖−1 , 𝑥 𝑖+1 ]

and 𝐿 2𝑁 (W − 𝑁
w)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2 )w(𝑥𝑖 )
≤ 𝐶𝜖 𝑤 002 (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
+ 𝐶𝜖 ℎ𝑖4 𝑤 2(𝑣𝑖) (𝑥) [𝑥𝑖−1 , 𝑥𝑖+1 ] .

The derivative bounds of 𝑤 1 and 𝑤 2 from Lemma 4.2.4 combined with the
inequality kℎ2𝑚𝑎𝑥 k𝑏k ∞ ≤ 𝜖 gives
 √︃ 
 exp −𝑥𝑖−1 𝜖 , 𝑥𝑖 ≤ 12 , 

 𝑏(0)



 
𝑁
 

𝐿 (W − w)(𝑥𝑖 ) ≤ 𝐶  √︃  .
1
 exp −(1 − 𝑥𝑖+1 ) 𝜖 ,
 𝑏(1) 

 𝑥𝑖 > 2



 
Using Lemma 4.2.5, for 𝑖 ≥ 𝑘 𝑙 and 𝑥𝑖 ≤ 12 , we have
 √︃    √︃  4
𝑥 𝑘 −1
𝐿 𝑁 (W − w)(𝑥 𝑖) ≤ 𝐶 exp −𝑥 𝑘 𝑙 −1 𝑏(0)
𝜖 = 𝐶 exp − 𝑙4 𝑏(0)
𝜖 ≤ 𝐶𝑁 −4 .

1
Similarly, the bounds for 𝑖 ≤ 𝑘 𝑟 and 𝑥𝑖 > 2 can be easily established.

(ii) When kℎ2𝑚𝑎𝑥 k𝑏k ∞ > 𝜖, the classical central difference method is applied to
approximate the solution where a repeated application of the mean value
theorem yields

𝐿 1𝑁 (W − w)(𝑥𝑖 ) = (𝐿 1 − 𝐿 1𝑁 )w(𝑥𝑖 )
≤ 𝐶 (𝛿2 𝑤 1 − 𝑤 001 )(𝑥𝑖 ) ≤ 𝐶 𝑤 001 (𝑥) [𝑥 𝑖−1 , 𝑥 𝑖+1 ]
,
and 𝐿 2𝑁 (W − w)(𝑥𝑖 ) = (𝐿 2 − 𝐿 2𝑁 )w(𝑥𝑖 )
≤ 𝐶𝜖 (𝛿2 𝑤 2 − 𝑤 002 )(𝑥𝑖 ) ≤ 𝐶𝜖 𝑤 002 (𝑥) [𝑥𝑖−1 , 𝑥 𝑖+1 ]
.

It then follows from the derivative bounds of 𝑤 1 and 𝑤 2 in Lemma 4.2.4

123 4. Reaction-Diffusion Problem of Fourth-order

that
 √︃ 
 exp −𝑥𝑖−1 𝜖 , 𝑥𝑖 ≤ 12 ,

 𝑏(0)



 

𝑁
 

𝐿 (W − w)(𝑥𝑖 ) ≤ 𝐶  √︃  .
1
 exp −(1 − 𝑥𝑖+1 ) 𝜖 , 𝑥𝑖 > 2

 𝑏(1) 

 

 
1
Now, use Lemma 4.2.5 for 𝑖 ≥ 𝑘 𝑙 and 𝑥𝑖 ≤ 2 to get
 √︃    √︃  4
𝑥 𝑘 −1
𝐿 𝑁 (W − w)(𝑥𝑖 ) ≤ 𝐶 exp −𝑥 𝑘 𝑙 −1 𝑏(0)
𝜖 = 𝐶 exp − 𝑙4 𝑏(0)
𝜖 ≤ 𝐶𝑁 −4 .

Similarly, the bounds are established for 𝑖 ≤ 𝑘 𝑟 and 𝑥𝑖 > 21 .

Thus, combining the various estimates concludes the proof. 

After estimating the errors of both the components V and W, the main conver-
gence result with respect to the computed approximate solution U is stated in the
following theorem.

Theorem 4.4.1. Let u be the solution of the problem (4.4) and U be the solution of
the discrete problem (4.23) on the equidistributed grid defined by (4.21)–(4.22) with
𝛼 = Ψ. Then, under the assumptions of Lemma 4.3.1, there exists a positive constant
𝐶 independent of 𝑁 and 𝜖 such that

u−U Ω𝐸
𝑁 ≤ 𝐶𝑁 −4 .

Proof. It is clear from Lemma 4.3.1 that the inverse of the discrete hybrid operator
L𝑁 is uniformly bounded with respect to the perturbation parameter 𝜖. The proof
thus follows as an immediate consequence of Lemmas 4.4.1, 4.4.2 and the triangle
inequality 4.29. 

Remark 4.4.1. The component 𝑢 1 of the solution u = (𝑢 1 , 𝑢 2 )𝑇 to the coupled system

of second-order differential equations (4.4) represents the solution of the fourth-order
singularly perturbed boundary value problem (4.1). Hence, the component 𝑈1 of
the numerical solution U = (𝑈1 , 𝑈2 )𝑇 to the discrete problem (4.23) is a parameter
robust numerical approximation to the solution of (4.1).
4.5. Numerical Experiments 124

4.5 Numerical Experiments

We now apply the proposed hybrid discretization technique to solve some test ex-
amples. The proposed technique requires a layer-adapted grid that is generated by
applying adaptive grid generation Algorithm 2.1 described in Section 2.5.1 of Chapter
2 with the user chosen constant 𝑄 = 1.1. To analyze the performance of the proposed
technique, we estimate the maximum pointwise errors and the order of convergence
associated with each problem obtained by using the proposed hybrid discretization
technique on the equidistributed grid.

Example 4.5.1. Consider the following fourth-order two-point problem

− 𝜖𝑢𝑖𝑣 (𝑥) + 4𝑢00 (𝑥) + 𝑢(𝑥) = − 𝑓 (𝑥), 𝑥 ∈ Ω = (0, 1) 



2 1 𝜖 2 2 𝜖
       

𝜖
𝑢(0) = − 1 + exp − √ exp − √ − −

, 𝑢(1) = −
4 𝜖 2 4 𝜖 3 4


00 00
𝑢 (0) = −2, 𝑢 (1) = −3



where 𝑓 (𝑥) is such that the exact solution is given by

2𝑥 2(1 − 𝑥) 𝑥3 𝑥2 2
       
𝜖
𝑢(𝑥) = − exp − √ + exp − √ − − 1 − exp − √ . (4.30)
4 𝜖 𝜖 6 6 𝜖

In the analysis that follows, we evaluate the maximum pointwise errors by using

𝐸 𝜖𝑁 = max 𝑈 𝑁 (𝑥𝑖 ) − 𝑢(𝑥𝑖 ) ,

0≤𝑖≤𝑁

where 𝑢(𝑥𝑖 ) is the exact solution and 𝑈 𝑁 (𝑥𝑖 ) is the numerical approximation obtained
at the grid point 𝑥𝑖 with 𝑁 number of intervals.

Example 4.5.2. Consider the following fourth-order two-point problem

− 𝜖𝑢𝑖𝑣 (𝑥) + 5 exp(1 − 𝑥)𝑢00 (𝑥) + (1 + 𝑥 3 )𝑢(𝑥) = −2 exp(𝑥), 𝑥 ∈ Ω = (0, 1),

)

𝑢(0) = 1, 𝑢(1) = 1, 𝑢00 (0) = −1, 𝑢00 (1) = −1.

Since Example 4.5.2 does not have an exact solution, we apply a variant of double
mesh principle to estimate the accuracy of the numerical solution. Let 𝑈 𝑁 be the
e2𝑁 be the numerical
computed approximate solution with 𝑁 number of intervals and 𝑈
solution on a mesh having 2𝑁 intervals that comprises of the points of the original
125 4. Reaction-Diffusion Problem of Fourth-order

mesh and its mid-points such that

(𝑥𝑖 + 𝑥𝑖+1 )
𝑥 2𝑖 = 𝑥𝑖 , 𝑖 = 0, . . . , 𝑁 and e
𝑥2𝑖+1 = , 𝑖 = 0, . . . , 𝑁 − 1.
2
e

e2𝑁 (e
Then, the maximum pointwise error is estimated by 𝐸 𝜖𝑁 = max 𝑈 𝑁 (𝑥𝑖 ) − 𝑈 𝑥 2𝑖 ) .
0≤𝑖≤𝑁
Taking the maximum over a wide range of 𝜖, the uniform errors for a fixed value
of 𝑁 are computed by 𝐸 𝑁 = max𝜖 𝐸 𝜖𝑁 . The corresponding parameter-uniform

 𝑁
convergence rates are given by 𝑝 = log2 𝐸 2𝑁 .
𝑁 𝐸𝜖
𝜖

On the adaptively generated grid, the fourth-order problems in Examples 4.5.1

and 4.5.2 are firstly converted into a coupled system of second-order differential
equations. The maximum nodal errors and the corresponding convergence rates for
the computed approximate solution of the coupled system of differential equations
are displayed in Tables 4.1 and 4.2, respectively. For a fixed value of 𝑁 the last row
of each table displays the 𝜖-uniform errors. The numerical results clearly indicate
𝜖-uniform fourth order convergence of the proposed hybrid difference scheme.

To demonstrate the efficiency of the computed approximate solution on the

equidistributed grid, we have compared the uniform errors and the associated rates
of convergence of the proposed scheme on piecewise uniform Shishkin mesh. A
significant gain in the accuracy of the numerical method on the equidistributed grid
over the piecewise uniform Shishkin mesh can be seen from the results given in Tables
4.3 and 4.4. From these results, it is evident that the uniform errors on piecewise
uniform mesh are more than that on the equidistributed grid which is exponentially
stretched in the boundary layer region. The piecewise uniform Shishkin mesh is
constructed by using an optimal value of the parameter 𝜎 = 4.

Tables ?? and 4.6 compares the 𝜖-uniform errors and convergence rates of the
proposed hybrid difference technique and the classical central difference scheme on
equidistributed grid. As depicted, the results of the classical central difference scheme
have uniform second-order convergence and are inferior than that of the proposed
hybrid difference scheme. Therefore, from the numerical results presented in these
tables, we can conclude that the proposed hybrid difference scheme yields parameter
uniform fourth-order convergence on the equidistributed grid which justifies the
theoretical estimates of Section 4.4.
4.5. Numerical Experiments 126

Table 4.1: Max-norm errors and the order of convergence for the proposed hybrid
scheme on a layer adapted grid for Example 4.5.1

Number of intervals
𝜖 = 2−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211
𝑘=4 1.4060e-05 1.0010e-06 6.2500e-08 3.9010e-09 2.4380e-10 1.5230e-11
𝑝𝑁 3.81 4.00 4.00 4.00 4.00
𝑘=8 2.0230e-05 1.2640e-06 7.9000e-08 4.9370e-09 3.0840e-10 1.9270e-11
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 12 4.1040e-05 2.6000e-06 1.6240e-07 1.0150e-08 6.3340e-10 3.9570e-11
𝑝𝑁 3.98 4.00 4.00 4.00 4.00
𝑘 = 16 5.3390e-05 3.3510e-06 2.0930e-07 1.3080e-08 8.1730e-10 5.1060e-11
𝑝𝑁 3.99 4.00 4.00 4.00 4.00
𝑘 = 20 5.2980e-05 3.3150e-06 2.0710e-07 1.2940e-08 8.0850e-10 5.0520e-11
𝑝𝑁 3.99 4.00 4.00 4.00 4.00
𝑘 = 24 5.2920e-05 3.3070e-06 2.0620e-07 1.2820e-08 8.0110e-10 5.0040e-11
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 28 5.2890e-05 3.3040e-06 2.0570e-07 1.2780e-08 7.9860e-10 4.9890e-11
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 32 5.2890e-05 3.3040e-06 2.0570e-07 1.2780e-08 7.9860e-10 4.9890e-11
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 36 5.2890e-05 3.3040e-06 2.0570e-07 1.2780e-08 7.9860e-10 4.9890e-11
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑒𝑁 5.3390e-05 3.3510e-06 2.0930e-07 1.3080e-08 8.1730e-10 5.1060e-11
127 4. Reaction-Diffusion Problem of Fourth-order

Table 4.2: Max-norm errors and the order of convergence for the proposed hybrid
scheme on a layer adapted grid for Example 4.5.2

Number of intervals
𝜖 = 2−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211
𝑘=4 3.9950e-04 2.4120e-05 1.5030e-06 9.3340e-08 5.8310e-09 3.6390e-10
𝑝𝑁 4.04 4.00 4.00 4.00 4.00
𝑘=8 9.4600e-05 5.8880e-06 3.6800e-07 2.3000e-08 1.4370e-09 8.9230e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 12 5.6210e-04 3.5120e-05 2.1920e-06 1.3690e-07 8.5130e-09 5.3110e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 16 3.3070e-04 2.0660e-05 1.2910e-06 8.0640e-08 5.0400e-09 3.1490e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 20 3.4330e-04 2.1450e-05 1.3400e-06 8.3710e-08 5.2170e-09 3.2610e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 24 3.4670e-04 2.1660e-05 1.3520e-06 8.4470e-08 5.2760e-09 3.2960e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 28 3.4750e-04 2.1710e-05 1.3560e-06 8.4710e-08 5.2940e-09 3.3070e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 32 3.4750e-04 2.1710e-05 1.3560e-06 8.4710e-08 5.2940e-09 3.3070e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑘 = 36 3.4750e-04 2.1710e-05 1.3560e-06 8.4710e-08 5.2940e-09 3.3070e-10
𝑝𝑁 4.00 4.00 4.00 4.00 4.00
𝑒𝑁 3.9950e-04 2.4120e-05 1.5030e-06 9.3340e-08 5.8310e-09 3.6390e-10
4.5. Numerical Experiments 128

Table 4.3: Uniform errors and the order of convergence of the proposed hybrid
scheme on Shishkin and Equidistributed meshes for Example 4.5.1

Number of Shishkin Mesh Equidistributed Mesh

Intervals
𝑁 𝐸𝑁 𝑝𝑁 𝐸𝑁 𝑝𝑁
26 8.638e-04 2.94 5.34e-05 3.99
27 1.121e-04 3.02 3.35e-06 4.00
28 1.379e-05 3.13 2.09e-07 4.00
29 1.565e-06 3.17 1.31e-08 4.00
210 1.732e-07 3.18 8.17e-10 4.00
211 1.911e-08 5.11e-11

Table 4.4: Uniform errors and the order of convergence of the proposed hybrid
scheme on Shishkin and Equidistributed meshes for Example 4.5.2

Number of Shishkin Mesh Equidistributed Mesh

Intervals
𝑁 𝐸𝑁 𝑝𝑁 𝐸𝑁 𝑝𝑁
26 5.144e-04 2.56 4.00e-04 4.04
27 8.686e-04 3.09 2.41e-05 4.00
28 1.019e-04 3.1 1.50e-06 4.00
29 1.182e-05 3.12 9.33e-08 4.00
210 1.354e-06 3.18 5.83e-09 4.00
211 1.493e-07 3.64e-10
129 4. Reaction-Diffusion Problem of Fourth-order

Table 4.5: Uniform errors and the order of convergence of the central difference
scheme and the proposed hybrid scheme on equidistributed mesh for Example 4.5.1

Number of Central Difference Method Hybrid Difference Scheme

Intervals
𝑁 𝐸𝑁 𝑝𝑁 𝐸𝑁 𝑝𝑁
26 9.487e-04 1.97 5.339e-05 3.99
27 2.412e-04 2.00 3.351e-06 4.00
28 6.016e-05 2.00 2.093e-07 4.00
29 1.502e-05 2.00 1.308e-08 4.00
210 3.748e-06 2.00 8.173e-10 4.00
211 9.334e-07 5.106e-11

Table 4.6: Uniform errors and the order of convergence of the central difference
scheme and the proposed hybrid scheme on equidistributed mesh for Example 4.5.2

Number of Central Difference Method Hybrid Difference Scheme

Intervals
𝑁 𝐸𝑁 𝑝𝑁 𝐸𝑁 𝑝𝑁
26 8.302e-04 2.00 3.995e-04 4.04
27 2.072e-04 2.00 2.412e-05 4.00
28 5.177e-05 2.00 1.503e-06 4.00
29 1.288e-05 2.01 9.334e-08 4.00
210 3.196e-06 2.03 5.831e-09 4.00
211 7.791e-07 6.639e-10
4.6. Conclusions 130

4.6 Conclusions
We proposed a parameter-uniform hybrid finite-difference approximation technique
for fourth-order singularly perturbed differential equations on a layer-adapted non-
uniform grid. Firstly, the fourth-order equations are transformed into a coupled
system of two second-order differential equations. A highly accurate numerical ap-
proximation is generated for the coupled system of differential equations by using a
suitable combination of the fourth-order Hermite difference scheme with the classical
central difference method on an equidistributed grid. The adaptive grid is gener-
ated by equidistributing a positive monitor function that automatically detects the
characteristics of the boundary layers present in the exact solution, thus avoiding
any prior information. The exponential stretching of the grid within the layers is
responsible for an improved convergence rate compared to the piecewise uniform
mesh. The numerical experiments support the theoretical error analysis showing
parameter uniform fourth-order convergence of the hybrid difference scheme on the
equidistributed grid.