Chapter 3

Reaction-Diffusion Problem with

Robin Boundary Conditions

3.1 Introduction

In Chapter 2, a hybrid difference scheme, comprising of Hermite difference scheme

in the layer region and classical difference scheme in the outer region, on an equidis-
tributed mesh is proposed to find an approximate solution to the reaction-diffusion
problem with Dirichlet boundary conditions.

In this chapter, a hybrid spline difference scheme that uses exponential and cubic
splines on a posteriori layer-adapted grid is proposed to solve a singular perturbation
problem with robin boundary conditions. The adaptive mesh {𝑥𝑖 }𝑖=0
𝑁 is generated by

equidistributing the monitor function 𝑀 (𝑠, 𝑢(𝑠)) using the equidistribution principle
(2.1). The nonlinearity in (2.1) is solved using the adaptive mesh algorithm proposed
by Kopteva and Stynes [155]. This algorithm addresses all the fundamental issues
involved in mesh generation. For an accurate numerical solution to the problem,
the exponential splines are used at the internal nodes of the layer-adapted mesh
and cubic splines for the more general robin boundary conditions. We prove that
the proposed numerical technique is parameter-uniform second-order convergent in
discrete maximum norm. The smoothness properties of the adaptively generated
mesh are also analyzed to establish uniform second-order convergence of the proposed
technique.

71
3.2. The Continuous Problem 72

3.2 The Continuous Problem

Consider the following singularly perturbed boundary value problem

𝐿𝑢(𝑥) := −𝜖𝑢00 (𝑥) + 𝑏(𝑥)𝑢(𝑥) = 𝑓 (𝑥), 𝑥 ∈ Ω = (0, 1),

)
√ √ (3.1)
𝐵𝑙 𝑢(0) ≡ 𝛼0 𝑢(0) − 𝛽0 𝜖𝑢0 (0) = 𝛾0 , 𝐵𝑟 𝑢(1) ≡ 𝛼1 𝑢(1) + 𝛽1 𝜖𝑢0 (1) = 𝛾1 .

Here, 𝑏(𝑥) and 𝑓 (𝑥) are sufficiently smooth such that 𝑏(𝑥) ≥ 𝜌 > 0, for all 𝑥 ∈ Ω =
[0, 1] and 𝜖 is the perturbation parameter. The constants 𝛼0 , 𝛼1 , 𝛽0 , 𝛽1 , 𝛾0 , 𝛾1 satisfy
𝛼 𝑗 , 𝛽 𝑗 ≥ 0, 𝛼 𝑗 + 𝛽 𝑗 > 0, 𝑗 = 0, 1 with 𝛼0 ≠ 𝛽0 𝑏(0). These conditions guarantee the
√︁

existence of a unique solution to the problem (3.1). Moreover, the solution of the
problem exhibits exponential boundary layers at 𝑥 = 0 and 𝑥 = 1.

3.2.1 Solution Decomposition

The solution to the problem exhibits multiscale behaviour. The standard decomposi-
tion is used to divide it into a smooth and layer component. An initial insight into
the analytical behaviour of the solution is obtained through this decomposition. The
next two lemmas estimate the derivative bounds of the solution and its components
using the standard decomposition. Later, these bounds are used to determine the
error estimates of the numerical discretization.

Lemma 3.2.1. Let 𝑏(𝑥) ≥ 𝜌 > 0. Then, for 𝑘 = 0, . . . , 2𝑝 + 1, the solution

𝑢(𝑥) ∈ 𝐶 2 (Ω) ∩ 𝐶 1 (Ω) satisfies
  √︂   √︂ 
−𝑘 𝜌 −𝑘 𝜌
𝑢 (𝑥) ≤ 𝐶 1 + 𝜖 2 exp −𝑥
(𝑘)
+ 𝜖 2 exp −(1 − 𝑥) ,
𝜖 𝜖

for an arbitrary 𝑝.

Proof. The proof follows from the inductive argument given by Miller et al. in
[246]. 

Lemma 3.2.2. The solution 𝑢(𝑥) admits a decomposition 𝑢(𝑥) := 𝑣(𝑥) + 𝑤(𝑥) where
the smooth part 𝑣(𝑥) satisfies

𝑓 (0) 𝑓 (1)
𝐿𝑣(𝑥) = 𝑓 (𝑥), 𝑥 ∈ Ω = (0, 1) and 𝐵𝑙 𝑣(0) = , 𝐵𝑟 𝑣(1) = ,
𝑏(0) 𝑏(1)
73 3. Reaction-Diffusion Problem with Robin Boundary Conditions

and the layer part 𝑤(𝑥) satisfies

𝐿𝑤(𝑥) = 0, 𝑥 ∈ Ω = (0, 1) and 𝐵𝑙 𝑤(0) = 𝛾0 − 𝐵𝑙 𝑣(0), 𝐵𝑟 𝑤(1) = 𝛾1 − 𝐵𝑟 𝑣(1),
such that

|𝑣 (𝑘) (𝑥)| ≤ 𝐶, and  √︃   

√︃
−𝑘
|𝑤 (𝑥)| ≤ 𝐶𝜖 2 exp −𝑥 𝜖 + exp −(1 − 𝑥) 𝜖
(𝑘) 𝑏(0) 𝑏(1)
, 𝑘 = 0, . . . , 2𝑝 + 1
for an arbitrary 𝑝.

Proof. Define the asymptotic expansion of the smooth part 𝑣(𝑥) by

2𝑝+1
∑︁ 𝑗
𝑣(𝑥) = 𝜖 2 𝑣 𝑗 (𝑥) + 𝜖 𝑝+1 𝑣 ∗2𝑝+2 (𝑥).
𝑗=0

Now, by substituting 𝑣(𝑥) in (3.1) and equating the coefficients of like powers of 𝜖,
we get

𝑏(𝑥)𝑣 0 (𝑥) = 𝑓 (𝑥),

𝑏(𝑥)𝑣 2 𝑗 (𝑥) = 𝑣 002 𝑗−2 (𝑥) for 𝑗 = 1, . . . , 𝑝, and 𝑣 2 𝑗+1 (𝑥) = 0 for 𝑗 = 0, . . . , 𝑝,
𝐿𝑣 ∗2𝑝+2 (𝑥) = 𝑣 002𝑝 (𝑥), 𝐵𝑙 𝑣 ∗2𝑝+2 (0) = 𝐵𝑟 𝑣 ∗2𝑝+2 (1) = 0.

From Lemma 3.2.1, we get 𝑣 (𝑘) (𝑥) ≤ 𝐶, 𝑘 = 0, . . . , 2𝑝+1. Since, the smooth part 𝑣(𝑥)
fails to satisfy the boundary conditions, the layer part 𝑤(𝑥) is further decomposed as
𝑤(𝑥) = 𝑤 𝐿 (𝑥) + 𝑤 𝑅 (𝑥) such that
2𝑝+1
∑︁ 𝑗
2𝑝+2 (𝑥), and
𝜖 2 𝑤 𝐿𝑗 (𝑥) + 𝜖 𝑝+1 𝑤 ∗𝐿
𝐿
𝑤 (𝑥) =
𝑗=0
2𝑝+1
∑︁ 𝑗
𝑅
𝑤 (𝑥) = 𝜖 2 𝑤 𝑅𝑗 (𝑥) + 𝜖 𝑝+1 𝑤 ∗𝑅
2𝑝+2 (𝑥).
𝑗=0
√
Firstly, consider the left-hand boundary layer. Substitute 𝑥 = 𝜉 𝜖. Then, Taylor’s
√
expansion of 𝑏(𝜉 𝜖) leads to

𝑑2
𝐿𝑤 0𝐿 (𝜉) = 0 where b
b 𝐿 ≡ − 2 + 𝑏(0)𝐼,
𝑑𝜉
e𝑙 𝑤 (0) = 𝛾0 − 𝐵𝑙 𝑣 0 (0), and
𝐵 𝐿
0

lim 𝑤 0𝐿 (𝜉) = 0 where 𝐵

e𝑙 𝑔(0) ≡ 𝛼0 𝑔(0) − 𝛽0 𝑔0 (0).
𝜉→∞
3.2. The Continuous Problem 74

Moreover, it follows that

𝑗
    ∑︁ 𝜉 𝑖 (𝑖)
𝐿𝑤 𝑗 (𝜉) = − 𝐿 𝑤 0 , . . . , 𝑤 𝑗−1 (𝜉) where 𝐿 𝑤 0 , . . . , 𝑤 𝑗−1 =
b 𝐿 b ∗ 𝐿 𝐿 b∗ 𝐿 𝐿
𝑏 (0)𝑤 𝐿𝑗−𝑖 (𝑥),
𝑖=1
𝑖!
e𝑙 𝑤 𝐿𝑗 (0) = −𝐵𝑙 𝑣 𝑗 (0) and lim 𝑤 𝐿𝑗 (𝜉) = 0, 𝑗 = 1, . . . , 2𝑝 + 1,
𝐵
𝜉→∞
and
2 𝑝+1
 
𝐿𝑤 ∗𝐿
2𝑝+2 (𝑥) = −𝜖 −𝑝−1
𝐿 𝑤 𝐿
0 + · · · + 𝜖 2 𝑤𝐿
2𝑝+1 (𝑥),
2 𝑝+1
 
e𝑙 𝑤 ∗𝐿 (0) = 0 and 𝐵𝑟 𝑤 ∗𝐿 (1) = −𝜖 −𝑝−1 𝐵𝑟 𝑤 𝐿 + · · · + 𝜖 2 𝑤 𝐿
𝐵 2𝑝+2 2𝑝+2 0 2𝑝+1 (1).

Therefore, for 𝑤 𝐿 (𝜉), we get

√︂ ! √︂ !
𝑏(0) −𝑘 𝑏(0)
𝑤 𝐿 (0) < 𝐶, 𝑤 𝐿 (1) < 𝐶 exp − and 𝑤 𝐿 (𝑘) (𝑥) < 𝐶𝜖 2 exp −𝑥 .
𝜖 𝜖

A similar argument can be used to establish the bounds for the right-hand boundary
layer. Hence the proof. 

3.2.2 Grid Structure

To construct a problem fitted layer-adapted non-equidistant grid, the adaptive al-

gorithm starts with an arbitrary uniform grid. Using the equidistribution principle,
the algorithm develops a layer-adapted grid using an error monitor function. The
non-negative monitor function guides to improve the structure of the grid by esti-
mating an error bound on the computed approximation. For a singularly perturbed
differential equation, these estimates can be tailored towards the abrupt behaviour of
the solution instigated by its layer part. Hence, we use the following monitor function
𝑀 = 𝛼 + |𝑤 00 (𝑥)| 1/2 . (3.2)
where the floor 𝛼 > 0 is a positive constant which ensure the presence of grid points
outside the boundary layer. The adaptive algorithm generates a non-uniform grid in
an iterative manner until some measure of convergence is attained.

To generate grid, equidistribute the monitor function (3.2) using the equidistribu-
tion principle (2.2) to get
∫ 𝑥(𝜉) ∫ 1
1/2
00
(𝛼 + |𝑤 (𝑥)| )𝑑𝑥 = 𝜉 (𝛼 + |𝑤 00 (𝑥)| 1/2 )𝑑𝑥
0 0
75 3. Reaction-Diffusion Problem with Robin Boundary Conditions

∫ 𝑥(𝜉) ∫ 𝑥(𝜉) ∫ 1 ∫ 1
1/2
⇒ 𝛼𝑑𝑥 + |𝑤 (𝑥)| 𝑑𝑥 = 𝜉 00
𝛼𝑑𝑥 + 𝜉 |𝑤 00 (𝑥)| 1/2 𝑑𝑥
0 0 0 0
∫ 𝑥(𝜉) ∫ 1
|𝑤 00 (𝑥)| 1/2 𝑑𝑥 = 𝜉𝛼𝑥| 10 + 𝜉 |𝑤 00 (𝑥)| 1/2 𝑑𝑥
𝑥(𝜉)
⇒ 𝛼𝑥| 0 +
0 0
∫ 𝑥(𝜉) ∫ 1
⇒ 𝛼𝑥(𝜉) + |𝑤 00 (𝑥)| 1/2 𝑑𝑥 = 𝜉𝛼 + 𝜉 |𝑤 00 (𝑥)| 1/2 𝑑𝑥. (3.3)
0 0

We use the expansion of 𝑤(𝑥) in Lemma  √︃ 3.2.2  to approximate 𝑤 (𝑥) as

00
 1 
exp 0, 2 , 
 𝜆0 𝑏(0) 𝑏(0)
−𝑥 ∈
 

 𝜖 𝜖 , 𝑥 
(3.4)

 

𝑤 00 (𝑥) ≈  √︃ 
1  
𝜆1 𝑏(1)
exp −(1 − 𝑏(1)
∈ 2, 1 , 



 𝜖 𝑥) 𝜖 , 𝑥 

where
 

      0  .  
𝑓 𝑓 √︁
𝜆 0 = 𝛾 0 − 𝛼0 (0) − 𝛽0 (0) 𝛼0 − 𝛽0 𝑏(0) , and
𝑏 𝑏
      0  .  
𝑓 𝑓 √︁
𝜆 1 = 𝛾1 − 𝛼1 (1) − 𝛽1 (1) 𝛼1 + 𝛽1 𝑏(1) .
𝑏 𝑏
∫1
Using (3.4), the term 0
|𝑤 00 (𝑥)| 1/2 𝑑𝑥 on the right-hand side (RHS) of (3.3) becomes
1
∫ 1 1
∫
2 1
∫ 1 1
00 00
|𝑤 (𝑥)| 𝑑𝑥 = 2 |𝑤 (𝑥)| 𝑑𝑥 + 2 |𝑤 00 (𝑥)| 2 𝑑𝑥
1
0 0 2

∫ 1 √︂ ! 12 ∫ 1 √︂ ! 21
2 𝜆 0 𝑏(0) 𝑏(0) 𝜆 1 𝑏(1) 𝑏(1)
= exp −𝑥 𝑑𝑥 + exp −(1 − 𝑥) 𝑑𝑥
1
0 𝜖 𝜖 2
𝜖 𝜖
1 ∫ 1 1 ∫
𝜆 1 𝑏(1) 2 1
√︂ ! √︂ !
𝜆 0 𝑏(0) 2 2 𝑥 𝑏(0) (1 − 𝑥) 𝑏(1)
= exp − 𝑑𝑥 + exp − 𝑑𝑥
𝜖 0 2 𝜖 𝜖 1
2
2 𝜖
 √︃  12  √︃  1
(1−𝑥) 𝑏(1)
1 exp − 1 exp −
𝑥 𝑏(0)
2 2
𝜆 0 𝑏(0) 2 𝜖
𝜆 1 𝑏(1) 2 𝜖
= √︃ + √︃
− 21 𝑏(0) 1 𝑏(1)
𝜖 𝜖
𝜖 2 𝜖
1
0 2
1 √︂ ! !
𝜆 0 𝑏(0) 2 1 1 𝑏(0)
= −2
𝜖   12 exp − 4 𝜖
−1 +
𝑏(0)
𝜖
1 √︂ !!
𝜆 1 𝑏(1) 2 1 1 𝑏(1)
2  21 1 − exp − 4
𝜖 
𝑏(1) 𝜖
𝜖
3.2. The Continuous Problem 76

√︂ !! √︂ !!
1 1 𝑏(0) 1 1 𝑏(1)
= 2 |𝜆 0 | 2 1 − exp − + 2 |𝜆1 | 2 1 − exp − . (3.5)
4 𝜖 4 𝜖
 √︃   √︃ 
As 𝜖 → 0, exp − 14 𝑏(0)
𝜖 → 0 and exp − 14 𝑏(1)
𝜖 → 0. Thus, (3.5) reduces to

∫ 1 1
 1 1

|𝑤 (𝑥)| 𝑑𝑥 ≈ 2 |𝜆 0 | + |𝜆 1 |
00 2 2 2 ≡ Ψ. (3.6)
0

Substitute (3.6) in (3.3) to obtain

∫ 𝑥(𝜉)
𝛼𝑥(𝜉) + |𝑤 00 (𝑥)| 1/2 𝑑𝑥 = 𝜉 (𝛼 + Ψ). (3.7)
0

For 0 ≤ 𝑥(𝜉) ≤ 12 , substitute 𝑤 00 (𝑥) from (3.4) in left-hand side (LHS) of (3.7) to get
∫ 𝑥(𝜉)
! 21√︂
𝜆 0 𝑏(0) 𝑏(0)
𝛼𝑥 (𝜉) + exp −𝑥 𝑑𝑥 = 𝜉 (𝛼 + Ψ)
0
𝜖 𝜖
1 ∫ 𝑥(𝜉) √︂ !
𝜆0 𝑏(0) 2 𝑥 𝑏(0)
⇒ 𝛼𝑥(𝜉) + exp − 𝑑𝑥 = 𝜉 (𝛼 + Ψ)
𝜖 0
2 𝜖
 √︃  𝑥(𝜉)
1
2
exp − 2 𝑥 𝑏(0)
𝜖
𝜆0 𝑏(0)
⇒ 𝛼𝑥(𝜉) + √︃ = 𝜉 (𝛼 + Ψ)
𝜖 1 𝑏(0)
−2 𝜖
0
! !
√︂
1 𝑥(𝜉) 𝑏(0)
⇒ 𝛼𝑥(𝜉) − 2 |𝜆 0 | 2 exp − − 1 = 𝜉 (𝛼 + Ψ)
2 𝜖
√︂ !!
1 𝑥(𝜉) 𝑏(0)
⇒ 𝛼𝑥(𝜉) + 2 |𝜆 0 | 2 1 − exp − = 𝜉 (𝛼 + Ψ).
2 𝜖

Uniformly divide both the LHS and RHS by Ψ to get

1 √︂ !!
𝑥(𝜉) 2|𝜆 0 | 2 𝑥(𝜉) 𝑏(0) 𝛼 
𝛼 + 1 − exp − =𝜉 +1 . (3.8)
Ψ Ψ 2 𝜖 Ψ

Similarly, for 𝑥(𝜉) > 12 , the map reads

1 √︂ !!
(1 − 𝑥(𝜉)) 2|𝜆 1 | 2 (1 − 𝑥(𝜉)) 𝑏(1) 𝛼 
𝛼 + 1 − exp − = (1 − 𝜉) +1 . (3.9)
Ψ Ψ 2 𝜖 Ψ
77 3. Reaction-Diffusion Problem with Robin Boundary Conditions

This correlation between the non-equidistant grid {𝑥𝑖 }𝑖=0

𝑁 and the uniform grid

𝜉𝑖 = 𝑁𝑖 𝑖=0 formulates the following grid

 𝑁

√︂ !!
𝑥𝑖 𝑥𝑖 𝑏(0) 𝑖 𝛼 
𝛼 + 𝜃 0 1 − exp − = +1 , 𝑥𝑖 ≤ 1/2, (3.10)
Ψ 2 𝜖 𝑁 Ψ

and
√︂ !! 
(1 − 𝑥𝑖 ) (1 − 𝑥𝑖 ) 𝑏(1) 𝑖 𝛼 
𝛼 + 𝜃 1 1 − exp − = 1− +1 , 𝑥𝑖 > 1/2, (3.11)
Ψ 2 𝜖 𝑁 Ψ
 .   . 
where 𝜃 0 = |𝜆 0 | 1/2 |𝜆 0 | 1/2 + |𝜆 1 | 1/2 and 𝜃 1 = |𝜆1 | 1/2 |𝜆 0 | 1/2 + |𝜆1 | 1/2 .

On solving the non-linear algebraic equations (3.10) and (3.11), we can determine
the grid points implicitly. With a suitable choice of the floor 𝛼, the following lemma
provides a substantial configuration of the equidistributed grid.
Lemma 3.2.3. Let the layer-adapted grid (3.10) and (3.11) is generated for 𝛼 = Ψ.
Then
√︂ √︂
𝜖 𝜖
𝑥 𝑘𝑙 <2 log(𝑁) < 𝑥 𝑘 𝑙+1 and 𝑥 𝑘 𝑟 −1 < 1 − 2 log(𝑁) < 𝑥 𝑘 𝑟
𝑏(0) 𝑏(1)

where
 √︂ 
𝜃0 𝜖
𝑘𝑙 = (𝑁 − 1) + 𝑁 log(𝑁) ,
2 𝑏(0)
 √︂ 
𝜃1 𝜖
𝑘 𝑟 = 𝑁 − (𝑁 − 1) − 𝑁 log(𝑁) + 1,
2 𝑏(1)

and [.] denotes the integral part of the term. Moreover, for a generic constant 𝐶 > 1,
we have  √︃ 
−𝑥 𝑖
exp 2
𝑏(0)
𝜖 ≤ 𝐶
𝑁, 𝑖 ≥ 𝑘 𝑙 − 1, 𝑥𝑖 ≤ 1/2, and
 √︃ 
−(1−𝑥 𝑖 )
exp 2
𝑏(1)
𝜖 ≤ 𝐶
𝑁, 𝑖 ≤ 𝑘𝑟 , 𝑥𝑖 > 1/2.
√︃
Proof. Firstly, consider the left-hand boundary layer. Put 𝑥𝑖 = 2 𝜖
𝑏(0) log(𝑁) and
𝛼 = Ψ in equation (3.10) to get
√︃ √︃
2 𝑏(0)
𝜖
log(𝑁) 2 𝑏(0) log(𝑁)
𝜖 √︂ 
ªª 𝑖 Ψ

𝑏(0)
+ 𝜃 0 ­­1 − exp ­­− +1
© ©
Ψ ®® =
Ψ 2 𝜖 ®® 𝑁 Ψ
« « ¬¬
3.2. The Continuous Problem 78

2𝑖
√︂
𝜖
⇒ 2 log(𝑁) + 𝜃 0 1 − exp − log(𝑁) =
𝑏(0) 𝑁
√︂
𝜖 𝜃0   𝑖
⇒ log(𝑁) + 1 − 𝑁 −1 =
𝑏(0) 2 𝑁
√︂
𝜖 𝜃0
⇒ 𝑁 log(𝑁) + (𝑁 − 1) = 𝑖
𝑏(0) 2
 √︂ 
𝜃0 𝜖
⇒ 𝑘𝑙 = (𝑁 − 1) + 𝑁 log(𝑁) .
2 𝑏(0)

Using (3.10), for 𝑖 ≥ 𝑘 𝑙 − 1 and 𝑥𝑖 ≤ 21

√︂ !
1 2 (𝑘 𝑙 − 1)
 
−𝑥 𝑘 𝑙 −1 𝑏(0)
exp = 𝑥 𝑘 𝑙 −1 + 𝜃 0 −
4 𝜖 𝜃0 𝑁
1 1
 √︂  √︂ 
𝜖 𝜖
≤ 2 log 𝑁 + 𝜃 0 − 𝜃 0 (𝑁 − 1) + 2 𝑁 log 𝑁 − 2
𝜃0 𝑏(0) 𝑁 𝑏(0)
≤ 𝐶𝑁 −1 .

Similar treatment of the right-hand boundary layer results in

 √︂ 
𝜃1 𝜖
𝑘 𝑟 = 𝑁 − (𝑁 − 1) − 𝑁 log(𝑁) + 1
2 𝑏(1)

and √︂ !
−(1 − 𝑥𝑖 ) 𝑏(1) 𝐶 1
exp ≤ , 𝑖 ≤ 𝑘𝑟 , 𝑥𝑖 > . 
2 𝜖 𝑁 2

The analysis of most of the numerical discretization techniques that used standard
Shishkin mesh inspires to choose this particular value of 𝛼. Therefore, the adaptive
grid shares some of its key features with Shishkin mesh but is better in accuracy due
to exponential stretching within the layers. Both the grids have been analyzed and
compared in [245]. The next lemma presents the bounds on the grid spacing ℎ𝑖 in
the boundary layer region.

Lemma 3.2.4. Within the boundary layer region, the grid spacing ℎ𝑖 satisfies
√︂ √︂
𝜖 𝜖
ℎ𝑖 < 𝐶 , 𝑖 = 1, . . . , 𝑘 𝑙 , and ℎ𝑖 < 𝐶 , 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁.
𝑏(0) 𝑏(1)

Proof. Firstly, consider left-hand boundary layer. To determine the upper and lower
79 3. Reaction-Diffusion Problem with Robin Boundary Conditions

bounds of any point 𝑥𝑖 , consider the map in (3.10). Substitute 𝛼 = Ψ to obtain

√︂ !! √︂ !!
𝑥𝑖 𝑏(0) 2𝑖 𝑥𝑖 𝑥𝑖 𝑏(0) 2𝑖
𝑥𝑖 + 𝜃 0 1 − exp − = ⇒ + 1 − exp − =
2 𝜖 𝑁 𝜃0 2 𝜖 𝜃0 𝑁
√︂ ! √︂ !
𝑥𝑖 𝑥𝑖 𝑏(0) 2𝑖 𝑥𝑖 𝑏(0) 2𝑖 𝑥𝑖
⇒ − − 1 + exp − =− ⇒ exp − =1− + .
𝜃0 2 𝜖 𝜃0 𝑁 2 𝜖 𝜃0 𝑁 𝜃0

Define 𝑥𝑖 such that for 𝑥𝑖 < 𝑥𝑖 , we have

√︂ ! √︂
2𝑖 2𝑖
 
𝑥𝑖 𝑏(0) 𝑥𝑖 𝑏(0)
exp − =1− ⇒− = log 1 −
2 𝜖 𝜃0 𝑁 2 𝜖 𝜃0 𝑁
2𝑖
√︂  
𝜖
⇒ 𝑥𝑖 = −2 log 1 − > 𝑥𝑖 . (3.12)
𝑏(0) 𝜃0 𝑁

Using 𝑥𝑖 in (3.10), we define 𝑥𝑖 < 𝑥𝑖 such that

√︂ !! √︂ !
𝑥𝑖 𝑏(0) 2𝑖 𝑥𝑖 𝑥𝑖 𝑏(0) 2𝑖
𝑥𝑖 + 𝜃 0 1 − exp − = ⇒ + 1 − exp − =
2 𝜖 𝑁 𝜃0 2 𝜖 𝜃0 𝑁
√︂ !
𝑥𝑖 𝑏(0) 𝑥𝑖 2𝑖
⇒ exp − =1+ −
2 𝜖 𝜃0 𝜃0 𝑁
√︂ !
2𝑖 2 2𝑖
√︂  
𝑥𝑖 𝑏(0) 𝜖
⇒ exp − =1− − log 1 −
2 𝜖 𝜃 0 𝑁 𝜃 0 𝑏(0) 𝜃0 𝑁
√︂
2𝑖 2 2𝑖
 √︂  
𝑥𝑖 𝑏(0) 𝜖
⇒ − = log 1 − − log 1 −
2 𝜖 𝜃 0 𝑁 𝜃 0 𝑏(0) 𝜃0 𝑁
2𝑖 2 2𝑖
√︂  √︂  
𝜖 𝜖
⇒ 𝑥𝑖 = −2 log 1 − − log 1 − . (3.13)
𝑏(0) 𝜃 0 𝑁 𝜃 0 𝑏(0) 𝜃0 𝑁

Thus, from (3.12) and (3.13), we obtain the following bounds for ℎ𝑖 in the boundary
layer region

ℎ𝑖 = 𝑥𝑖 − 𝑥𝑖−1 ≤ 𝑥𝑖 − 𝑥𝑖−1
2𝑖 2(𝑖 − 1)
√︂    √︂ 
𝜖 𝜖
= −2 log 1 − − −2 log 1 −
𝑏(0) 𝜃0 𝑁 𝑏(0) 𝜃0 𝑁
2 2(𝑖 − 1)
√︂  
𝜖
− log 1 −
𝜃 0 𝑏(0) 𝜃0 𝑁
2(𝑖 − 1) 2 2(𝑖 − 1) 2𝑖
√︂   √︂    
𝜖 𝜖
= 2 log 1 − − log 1 − − log 1 −
𝑏(0) 𝜃0 𝑁 𝜃 0 𝑏(0) 𝜃0 𝑁 𝜃0 𝑁
3.2. The Continuous Problem 80

2(𝑖−1) 2(𝑖−1)
 
2
√︃
 ©1 − − 𝜖
log 1 −
√︂  
𝜖 𝜃0 𝑁 𝜃0 𝑏(0) 𝜃0 𝑁 ª
= 2 log ­
 ­ ®
𝑏(0) 
 1 − 𝜃2𝑖
0𝑁
®

 « ¬
1 − 2(𝑖−1)
√︃ 
√︂
© 0
𝜃 𝑁 − 2𝑖 + 2 − 2𝑁 𝜖
𝑏(0) log 𝜃0 𝑁 ª
𝜖
= 2 log ­­ ®
𝑏(0) 𝜃 0 𝑁 − 2𝑖 ®
« √︃   ¬
√︂ 2 − 2𝑁 𝑏(0) 𝜖
log 0 𝜃 0 𝑁
𝜃 𝑁−2(𝑖−1)
𝜖
2 log ­1 +
© ª
= ­ ®
𝑏(0) 𝜃 0 𝑁 − 2𝑖 ®
« √︃  ¬
√︂ 2 + 2𝑁 𝑏(0) log 𝜃 0 𝑁−2(𝑖−1) ª
𝜖 𝜃0 𝑁 √︂
𝜖 𝜖
2 log ­­1 +
©
= ®<𝐶 .
𝑏(0) 𝜃 0 𝑁 − 2𝑖 ® 𝑏(0)
« ¬
Similarly

2
√︂   √︂
𝜖 𝜖
ℎ1 < 𝑥 1 − 𝑥0 = −2 log 1 − <2 (𝜃 0 𝑁) −1 , and
𝑏(0) 𝜃0 𝑁 𝑏(0)
√︂
𝜖
ℎ𝑁 < 2 (𝜃 1 𝑁) −1 .
𝑏(1)

Similar treatment of the right-hand boundary layer completes the proof. 

Extending the result to the entire domain, the next lemma examines the width of
the adaptive grid to furnish its bounds.

Lemma 3.2.5. For 𝑖 = 1, . . . , 𝑁, the width of the generated grid spacing satisfies

ℎ𝑖 < 𝐶𝑁 −1 .

Proof. Owing to (3.6), we have

∫ 1 1
 1 1

|𝑤 00 (𝑥)| 2 𝑑𝑥 = Ψ = 2 |𝜆 0 | 2 + |𝜆 1 | 2 .
0

Then by using the monitor function defined in (3.2) and 𝛼 = Ψ, we have

∫ 1 ∫ 1 1
𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥 = (𝛼 + |𝑤 00 (𝑥)|) 2 𝑑𝑥 = 𝛼 + Ψ
0 0
∫ 1
𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥 = 𝛼 + Ψ = 2𝛼 ≤ 𝐶𝛼.
0
81 3. Reaction-Diffusion Problem with Robin Boundary Conditions

Using the equidistribution principle (2.1), we have

∫ 1
1
∫ 𝑥𝑖
𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥 = 𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥
𝑥 𝑖−1 𝑁 0
∫ 1
1
∫ 𝑥𝑖 ∫ 𝑥𝑖 ∫ 𝑥𝑖
𝐶𝛼
⇒ 𝛼𝑑𝑥 = 𝛼 𝑑𝑥 = 𝛼ℎ𝑖 ≤ 𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥 = 𝑀 (𝑥, 𝑢(𝑥))𝑑𝑥 ≤
𝑥 𝑖−1 𝑥 𝑖−1 𝑥𝑖−1 𝑁 0 𝑁
𝐶𝛼 𝐶
⇒ 𝛼ℎ𝑖 ≤ ⇒ ℎ𝑖 ≤ = 𝐶𝑁 −1 . 
𝑁 𝑁

3.3 Spline Difference Discretization

In this section, we present a spline difference discretization scheme to solve the
𝑁
problem (3.1) on the non-uniform equidistributed grid Ω𝐸 . The proposed spline dif-
ference discretization technique applies cubic splines to discretize the robin boundary
conditions and exponential splines at the internal nodes {𝑥𝑖 }𝑖=1
𝑁−1
of the layer-adapted
mesh. This section also presents the stability analysis of the discrete operator.

3.3.1 Cubic Spline Difference Scheme

To develop a cubic spline difference scheme on the non-uniform equidistributed
grid, we firstly introduce a cubic spline polynomial 𝑆(𝑥). The spacing between the
nodes of the grid is defined as ℎ𝑖 = 𝑥𝑖 − 𝑥𝑖−1 , 𝑖 = 1, . . . , 𝑁. For a given set of values
𝑢(𝑥 0 ), 𝑢(𝑥 1 ), . . . , 𝑢(𝑥 𝑁 ) of a function 𝑢(𝑥) at the nodes 𝑥 0 , 𝑥1 , . . . , 𝑥 𝑁 , the polynomial
𝑆(𝑥) satisfies

1. 𝑆(𝑥) ∈ 𝐶 2 [0, 1],

2. On each subinterval [𝑥𝑖−1 , 𝑥𝑖 ], 𝑖 = 1, . . . , 𝑁, 𝑆(𝑥) is a polynomial of degree 3,

and

3. 𝑆(𝑥𝑖 ) = 𝑢(𝑥𝑖 ), 𝑖 = 0, 1, . . . , 𝑁.

The cubic spline polynomial 𝑆(𝑥) that represents the solution of the differential
equation 𝐷 4 (𝑆(𝑥)) = 0 can be expressed as

(𝑥𝑖 −𝑥) 3 (𝑥−𝑥𝑖−1 ) 3 ℎ2𝑖 ℎ2𝑖

   
(𝑥𝑖 −𝑥) (𝑥−𝑥𝑖−1 )
𝑆(𝑥) = 6ℎ𝑖 𝑀𝑖−1 + 6ℎ𝑖 𝑀𝑖 + 𝑢(𝑥𝑖−1 ) − 6 𝑀𝑖−1 ℎ𝑖 + 𝑢(𝑥𝑖 ) − 6 𝑀𝑖 ℎ𝑖
(3.14)
3.3. Spline Difference Discretization 82

where 𝑀𝑖 = 𝑆00 (𝑥𝑖 ), 𝑖 = 0, 1, . . . , 𝑁 and 𝑥𝑖−1 ≤ 𝑥 ≤ 𝑥𝑖 , 𝑖 = 1, . . . , 𝑁. Differentiate

(3.14) with respect to 𝑥. The left-hand derivative of 𝑆(𝑥) at the node 𝑥𝑖 is given as
2 2
! ! 
3(𝑥 2 1 1

− 𝑥 ) ℎ ℎ
𝑆0 (𝑥𝑖 −) =
𝑖 𝑖−1
𝑀𝑖 + 𝑢(𝑥𝑖−1 ) − 𝑖 𝑀𝑖−1 − + 𝑢(𝑥𝑖 ) − 𝑖 𝑀𝑖
6ℎ𝑖 6 ℎ𝑖 6 ℎ𝑖
3ℎ𝑖2 𝑢(𝑥𝑖−1 ) ℎ𝑖 𝑢(𝑥𝑖 ) ℎ𝑖
= 𝑀𝑖 − + 𝑀𝑖−1 + − 𝑀𝑖
6ℎ𝑖 ℎ𝑖 6 ℎ𝑖 6
   
ℎ𝑖 ℎ𝑖 ℎ𝑖 𝑢(𝑥𝑖 ) 𝑢(𝑥𝑖−1 )
= 𝑀𝑖−1 + 𝑀𝑖 − 𝑀𝑖 + −
6 2 6 ℎ𝑖 ℎ𝑖
ℎ𝑖 ℎ𝑖 𝑈𝑖 − 𝑈𝑖−1
= 𝑀𝑖−1 + 𝑀𝑖 + (3.15)
6 3 ℎ𝑖
where 𝑈𝑖 = 𝑢(𝑥𝑖 ), 𝑖 = 0, 1, . . . , 𝑁. Similarly, the right-hand derivative of 𝑆(𝑥) at the
node 𝑥𝑖 is given as
ℎ𝑖+1 ℎ𝑖+1 𝑈𝑖+1 − 𝑈𝑖
𝑆0 (𝑥𝑖 +) = − 𝑀𝑖 − 𝑀𝑖+1 + . (3.16)
3 6 ℎ𝑖+1
To find a numerical solution to the problem (3.1), substitute 𝑀𝑖 from

−𝜖 𝑀𝑖 + 𝑏𝑖 𝑈𝑖 = 𝑓𝑖 , where 𝑏𝑖 = 𝑏(𝑥𝑖 ), 𝑓𝑖 = 𝑓 (𝑥𝑖 ) and 𝑖 = 0, 1, . . . , 𝑁,

in (3.15) and (3.16). Then, equate them to ensure the continuity of the first-order
derivative of the cubic spline polynomial at the respective endpoints. Using (3.16),
for robin boundary condition at 𝑥 0 = 0, we get
ℎ1 ℎ1 𝑈1 − 𝑈0 𝑏 0𝑈0 − 𝑓0
𝑢0 (0) = 𝑆0 (0) = − 𝑀0 − 𝑀1 + where 𝑀0 = . (3.17)
3 6 ℎ1 𝜖
From the robin boundary condition at 𝑥0 given in the problem (3.1), we get
𝛼0𝑈0 − 𝛾0
𝑢0 (0) = √ . (3.18)
𝛽0 𝜖

Equate (3.17) and (3.18), to get

   
𝛼 0 𝑈0 − 𝛾 0 ℎ1 𝑏 0𝑈0 − 𝑓0 ℎ1 𝑏 1𝑈1 − 𝑓1 𝑈1 − 𝑈0
√ =− − + .
𝛽0 𝜖 3 𝜖 6 𝜖 ℎ1

Multiply throughout by 𝛽0 𝜖 to get

√ ℎ1 𝑏 0 𝛽0𝑈0 ℎ1 𝛽0 𝑓0 ℎ1 𝑏 1 𝛽0𝑈1 ℎ1 𝛽0 𝑓1 𝛽0 𝜖𝑈1 𝛽0 𝜖𝑈0
(𝛼0𝑈0 − 𝛾0 ) 𝜖 = − + − + + −
3 3 6 6 ℎ1 ℎ1
83 3. Reaction-Diffusion Problem with Robin Boundary Conditions

√ √
   
ℎ 1 𝑏 0 𝛽0 𝛽0 𝜖 ℎ 1 𝑏 1 𝛽0 𝛽0 𝜖 ℎ 1 𝛽0 𝑓0 ℎ 1 𝛽0 𝑓1
⇒ + + 𝛼 0 𝜖 𝑈0 + − 𝑈1 = + + 𝛾0 𝜖
3 ℎ1 6 ℎ1 3 6
√ ! ! √
3𝛼0 𝜖 3𝛽0 𝜖 𝑏 1 𝛽0 3𝛽0 𝜖 𝛽0 𝑓1 3𝛾0 𝜖
⇒ + 𝑏 0 𝛽0 + 2 𝑈0 + − 2 𝑈1 = 𝛽 0 𝑓 0 + + . (3.19)
ℎ1 ℎ1 2 ℎ1 2 ℎ1

Similarly for the robin boundary condition at 𝑥 𝑁 = 1, we get

! √ ! √
𝑏 𝑁−1 𝛽1 3𝛽1 𝜖 3𝛼1 𝜖 3𝛽1 𝜖 𝛽1 𝑓 𝑁−1 3𝛾1 𝜖
− 2 𝑈𝑁−1 + + 𝑏 𝑁 𝛽1 + 2 𝑈 𝑁 = + 𝛽1 𝑓 𝑁 + .
2 ℎ𝑁 ℎ𝑁 ℎ𝑁 2 ℎ𝑁
(3.20)

3.3.2 Exponential Spline Difference Scheme

The discretization technique uses exponential splines to estimate the solution of the
problem (3.1) at the internal nodes {𝑥𝑖 }𝑖=1
𝑁−1
of the non-uniform grid Ω𝐸𝑁 . On each
subinterval [𝑥𝑖−1 , 𝑥𝑖 ], 𝑖 = 2, 3, . . . , 𝑁 − 1, an exponential spline is defined as the
solution to the differential equation
)
(𝐷 4 − 𝑝𝑖2 𝐷 2 )𝑇 = 0; 𝑇 (𝑥𝑖−1 ) = 𝑈𝑖−1 , 𝑇 (𝑥𝑖 ) = 𝑈𝑖 ,
(3.21)
𝑇 00 (𝑥𝑖−1 ) = 𝑇𝑖−1
00 , 𝑇 00 (𝑥 ) = 𝑇 00,
𝑖 𝑖

where 𝑝𝑖 s are non-negative tension parameters defined on each subinterval [𝑥𝑖−1 , 𝑥𝑖 ],

𝑖 = 2, 3, . . . , 𝑁 − 1 and 𝑇𝑖00 are yet to be determined. As the tension parameter 𝑝𝑖 → 0,
the differential equation (3.21) reduces to (𝐷 4 )𝑇 = 0, leading to a cubic spline.

A general solution of the differential equation (𝐷 4 − 𝑝𝑖2 𝐷 2 )𝑇 = 0 belongs to the

span 1, 𝑥, exp( 𝑝𝑖 𝑥), exp(−𝑝𝑖 𝑥) [174]. The continuity constraint on the first-order


derivative at the nodes is used to determine 𝑇𝑖00. Thus, the exponential spline relation
at the internal nodes 𝑥𝑖 , 𝑖 = 1, 2, . . . , 𝑁 − 1 is given as
𝑈𝑖+1 − 𝑈𝑖 𝑈𝑖 − 𝑈𝑖−1
00
𝑒𝑖 𝑇𝑖−1 + (𝑑𝑖 + 𝑑𝑖+1 )𝑇𝑖00 + 𝑒𝑖+1𝑇𝑖+1
00
= − , 𝑖 = 1, 2, . . . , 𝑁 − 1 (3.22)
ℎ𝑖+1 ℎ𝑖
where we set
𝑠𝑖 − 𝑝 𝑖 ℎ𝑖 𝑝 𝑖 ℎ𝑖 𝑐 𝑖 − 𝑠𝑖
𝑠𝑖 = sinh( 𝑝𝑖 ℎ𝑖 ), 𝑐𝑖 = cosh( 𝑝𝑖 ℎ𝑖 ), 𝑒𝑖 = , 𝑑 = ,
𝑝𝑖2 𝑠𝑖 ℎ𝑖 𝑝𝑖2 𝑠𝑖 ℎ𝑖
𝑖

𝑠𝑖+1 − 𝑝𝑖+1 ℎ𝑖+1 𝑝𝑖+1 ℎ𝑖+1 𝑐𝑖+1 − 𝑠𝑖+1

𝑒𝑖+1 = 2 𝑠 ℎ
, and 𝑑𝑖+1 = 2 𝑠 ℎ
.
𝑝𝑖+1 𝑖+1 𝑖+1 𝑝𝑖+1 𝑖+1 𝑖+1
3.3. Spline Difference Discretization 84

To approximate the solution of the problem (3.1) on the basis of exponential splines,
we rewrite (3.1) as
𝑏𝑖 𝑈𝑖 − 𝑓𝑖
− 𝜖𝑇𝑖00 + 𝑏𝑖 𝑈𝑖 = 𝑓𝑖 ⇒ 𝑇𝑖00 = . (3.23)
𝜖
Substitute (3.23) in (3.22) to get the following system of equations
      
𝑏𝑖−1𝑈𝑖−1 − 𝑓𝑖−1 𝑏𝑖 𝑈𝑖 − 𝑓𝑖 𝑏𝑖+1𝑈𝑖+1 − 𝑓𝑖+1
𝑒𝑖 + 𝑑𝑖 + 𝑑𝑖+1 + 𝑒𝑖+1 =
𝜖 𝜖 𝜖
𝑈𝑖+1 − 𝑈𝑖 𝑈𝑖 − 𝑈𝑖−1
−
ℎ𝑖+1 ℎ
 𝑖
𝑒𝑖 𝑏𝑖−1 1 1 1 1
    

𝑏𝑖 𝑒𝑖+1 𝑏𝑖+1
⇒ − 𝑈𝑖−1 + 𝑑𝑖 + 𝑑𝑖+1 + + 𝑈𝑖 + − 𝑈𝑖+1 =
𝜖 ℎ𝑖 𝜖 ℎ𝑖+1 ℎ𝑖 𝜖 ℎ𝑖+1
 
𝑒𝑖 𝑓𝑖−1 𝑑𝑖 + 𝑑𝑖+1 𝑒𝑖+1 𝑓𝑖+1
+ 𝑓𝑖 +
𝜖 𝜖 𝜖
      
𝜖
𝜖 𝜖 𝜖
⇒ 𝑒𝑖 𝑏𝑖−1 − 𝑈𝑖−1 + 𝑑𝑖 + 𝑑𝑖+1 𝑏𝑖 + + 𝑈𝑖−1 + 𝑒𝑖+1 𝑏𝑖+1 − 𝑈𝑖+1 =
ℎ𝑖 ℎ𝑖+1 ℎ𝑖 ℎ𝑖+1
𝑒𝑖 𝑓𝑖−1 + (𝑑𝑖 + 𝑑𝑖+1 ) 𝑓𝑖 + 𝑒𝑖+1 𝑓𝑖+1 . (3.24)

Using exponential spline difference scheme (3.24) to discretize the differential equa-
tion at the internal nodes of the grid 𝑖 = 1, . . . , 𝑁 − 1 and the cubic spline difference
scheme (3.19)–(3.20) to discretize the robin boundary conditions, we now define the
numerical discretization of (3.1) as

𝐿 𝑁 𝑈𝑖 ≡ −𝜖 𝛿2𝑈𝑖 + (ℓ𝑏𝑈)𝑖 = (ℓ 𝑓 )𝑖 , 𝑖 = 1, . . . , 𝑁 − 1,
)
(3.25)
𝐵𝑙𝑁 𝑈0 = 𝛾0 , 𝐵𝑟𝑁 𝑈𝑁 = 𝛾1 ,

where
1 𝑣 𝑗+1 − 𝑣 𝑗 𝑣 𝑗 − 𝑣 𝑗−1
 
2 ℎ 𝑗 + ℎ 𝑗+1
𝛿 𝑣𝑗 = − , (ℓ𝑣) 𝑗 = ℓ −𝑗 𝑣 𝑗−1 + ℓ 𝑐𝑗 𝑣 𝑗 + ℓ +𝑗 𝑣 𝑗+1 , ℏ 𝑗 = ,
ℏ𝑗 ℎ 𝑗+1 ℎ𝑗 2
! ! !
𝑠 − 𝑝 ℎ 𝑝 ℎ 𝑐 − 𝑠 𝑝 ℎ 𝑐 − 𝑠 𝑠 − 𝑝 ℎ
ℓ −𝑗 = , ℓ +𝑗 =
𝑗 𝑗 𝑗 𝑗 𝑗 𝑗 𝑗 𝑗+1 𝑗+1 𝑗+1 𝑗+1 𝑗+1 𝑗+1 𝑗+1
2
, ℓ 𝑐𝑗 = 2
+ 2 2
.
ℏ𝑗 𝑝 𝑗 𝑠𝑗 ℎ𝑗 ℏ𝑗 𝑝 𝑗 𝑠𝑗 ℎ𝑗 ℏ 𝑗 𝑝 𝑗+1 𝑠 𝑗+1 ℎ 𝑗+1 ℏ 𝑗 𝑝 𝑗+1 𝑠 𝑗+1 ℎ 𝑗+1

By rearranging the terms, the tri-diagonal form of the spline difference scheme is
given as

𝐿 𝑁 𝑈𝑖 ≡ 𝑟𝑖−𝑈𝑖−1 + 𝑟𝑖𝑐𝑈𝑖 + 𝑟𝑖+𝑈𝑖+1 = 𝑞𝑖− 𝑓𝑖−1 + 𝑞𝑖𝑐 𝑓𝑖 + 𝑞𝑖+ 𝑓𝑖+1 , 𝑖 = 1, . . . , 𝑁 − 1 (3.26)

85 3. Reaction-Diffusion Problem with Robin Boundary Conditions

where the mixed robin boundary conditions are estimated as

√
𝐵𝑙𝑁 : 𝑟 0𝑐𝑈0 + 𝑟 0+𝑈1 = 𝛾0 𝜖 + 𝑞 0𝑐 𝑓0 + 𝑞 +0 𝑓1 𝛽0

)
√ . (3.27)
𝐵𝑟𝑁 : 𝑟 𝑁− 𝑈𝑁−1 + 𝑟 𝑁𝑐 𝑈𝑁 = 𝑞 −𝑁 𝑓 𝑁−1 + 𝑞 𝑐𝑁 𝑓 𝑁 𝛽1 + 𝛾1 𝜖

The coefficients 𝑟𝑖∗ , ∗ = −, 𝑐, +, 0 ≤ 𝑖 ≤ 𝑁 of the spline difference operator 𝐿 𝑁

discretizing the continuous problem (3.1) are given by
!
− 𝜖 𝑠𝑖 − 𝑝 𝑖 ℎ𝑖 𝜖 

𝑟𝑖 = 𝑒𝑖 𝑏𝑖−1 − = 𝑏𝑖−1 −

2


ℎ𝑖 𝑝 𝑖 𝑠𝑖 ℎ𝑖 ℎ𝑖 


! 


𝑐 𝜖 𝜖 𝑝𝑖 ℎ𝑖 𝑐𝑖 − 𝑠𝑖 𝑝𝑖+1 ℎ𝑖+1 𝑐𝑖+1 − 𝑠𝑖+1 𝜖 𝜖
𝑟𝑖 = (𝑑𝑖 + 𝑑𝑖+1 ) 𝑏𝑖 + + = + + +

2 2
𝑏 𝑖


ℎ𝑖+1 ℎ𝑖 𝑝 𝑖 𝑠𝑖 ℎ𝑖 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 ℎ𝑖+1 ℎ𝑖 

! 


𝑠𝑖+1 − 𝑝𝑖+1 ℎ𝑖+1

+ 𝜖 𝜖 
𝑟𝑖 = 𝑒𝑖+1 𝑏𝑖+1 − = −

2
𝑏 𝑖+1



ℎ𝑖+1 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 ℎ𝑖+1

− 𝑠𝑖 − 𝑝 𝑖 ℎ𝑖 

𝑞 𝑖 = 𝑒𝑖 = 2 



𝑝 𝑖 𝑠𝑖 ℎ𝑖 


− −

𝑐 𝑝 𝑖 ℎ 𝑖 𝑐 𝑖 𝑠 𝑖 𝑝 𝑖+1 ℎ 𝑖+1 𝑐 𝑖+1 𝑠 𝑖+1

𝑞𝑖 = (𝑑𝑖 + 𝑑𝑖+1 ) = +


2 2


𝑝 𝑖 𝑠𝑖 ℎ𝑖 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 



+ 𝑠 𝑖+1 − 𝑝 ℎ
𝑖+1 𝑖+1


𝑞𝑖 = 𝑒𝑖+1 = 2
. 


𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 
(3.28)


From (3.19), (3.20) and (3.27), the cubic spline difference approximation of the robin
boundary conditions yield
√  √ 
−𝛽0 𝜖
𝑟 0𝑐 = 𝜖 𝛼0 + 0ℎ1 + 𝑏 0 𝛽0 𝑞 0𝑐 , 𝑟 0+ = + 𝑏 1 𝛽0 𝑞 +0
𝛽 𝜖 

ℎ1 


𝑞 0𝑐 = ℎ31 , 𝑞 +0 = ℎ1 
(3.29)


6
𝛽1 𝜖 √ 𝛽1 𝜖
𝑟 𝑁− = 𝛽1 𝑏 𝑁−1 𝑞 −𝑁 − ℎ𝑁 , 𝑟 𝑁𝑐 = 𝛼1 𝜖 + 𝛽1 𝑞 𝑐𝑁 𝑏 𝑁 + ℎ𝑁




𝑞 −𝑁 = ℎ6𝑁 , 𝑞 𝑐𝑁 = ℎ3𝑁 .



A simultaneous solution of the system of linear algebraic equations (3.26) and (3.27)
generates a numerical approximation of the reaction-diffusion problem (3.1) on the
layer-adapted grid.

A rigorous analysis shows that the cubic spline difference scheme ceases to be
uniformly stable and does not satisfy the M-matrix criteria. However, for a particular
choice of the tension parameter 𝑝 = min 𝑝𝑖 , the resulting coefficient matrix of
𝑖=1,...,𝑁−1
the proposed scheme is proved to be M-matrix and hence inverse monotone.
3.3. Spline Difference Discretization 86

To estimate the stability of the discrete operator 𝐿 𝑁 , the next lemma shows that
the coefficient matrix of the proposed spline difference technique is an M-matrix with
positive diagonal and non-positive off-diagonal entries.

Lemma 3.3.1. Let 𝑁0 be the smallest positive integer such that

2Ψ2 (k𝑏k ∞ ) < 3𝜁 𝑁02 , (3.30)

where 𝜁 = max 𝜆 0 , 𝜆1 . Then, for every 𝑁 ≥ 𝑁0 , the coefficients of the discrete



operator 𝐿 𝑁 satisfy 𝑟𝑖𝑐 > 0, 𝑟𝑖− < 0, 𝑟𝑖+ < 0 and 𝑟𝑖𝑐 + 𝑟𝑖− + 𝑟𝑖+ > 0, 𝑖 = 0, . . . , 𝑁.

Proof. For the spline discretization technique (3.26)–(3.27), the associated coefficient
matrix is partitioned on the basis of location of the nodes 𝑥𝑖 . For 𝑖 = 1, 2, . . . , 𝑁 − 1,
the application of exponential splines yields
! !
𝑠 − 𝑝 ℎ 𝜖 1 𝑏 𝑏𝑖−1
𝑟𝑖− =
𝑖 𝑖 𝑖 𝑖−1
2
𝑏𝑖−1 − = −𝜖 + 2 − ,
𝑝 𝑖 𝑠𝑖 ℎ𝑖 ℎ𝑖 ℎ𝑖 𝑝𝑖 𝑠𝑖 𝑝 𝑖
! !
𝑠 − 𝑝 ℎ 𝜖 1 𝑏 𝑏𝑖+1
𝑟𝑖+ =
𝑖+1 𝑖+1 𝑖+1 𝑖+1
2
𝑏𝑖+1 − = −𝜖 + 2 − ,
𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 ℎ𝑖+1 ℎ𝑖+1 𝑝𝑖+1 𝑝𝑖+1 𝑠𝑖+1
!
𝑝 𝑖 ℎ 𝑖 𝑐 𝑖 − 𝑠 𝑖 𝑝 𝑖+1 ℎ 𝑖+1 𝑐 𝑖+1 − 𝑠 𝑖+1 𝜖 𝜖
𝑟𝑖𝑐 = 2
+ 2
𝑏𝑖 + +
𝑝 𝑖 𝑠𝑖 ℎ𝑖 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 ℎ𝑖+1 ℎ𝑖
!  !
1 1

𝑏𝑖 𝑐𝑖 𝑐𝑖+1 𝑏𝑖
= 𝜖− 2 + + 𝑏𝑖 + 𝜖− 2 .
ℎ𝑖 𝑝𝑖 𝑝𝑖 𝑠𝑖 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖+1 𝑝𝑖+1
√︃ 
Using uniform tension parameter 𝑝 = min {𝑝𝑖 } ≥ max 𝑏𝑖
𝜖 , we obtain
𝑖=1,...,𝑁−1 𝑖=1,...,𝑁−1
𝑟𝑖− < 0, 𝑟𝑖+ < 0 and 𝑟𝑖𝑐 > 0 as

( 𝑝 𝑖 ℎ𝑖 ) 2 ( 𝑝 𝑖 ℎ𝑖 ) 4
𝑐𝑖 = cosh( 𝑝𝑖 ℎ𝑖 ) = 1 + + + . . . , and
2 24
( 𝑝 𝑖 ℎ𝑖 ) 3 ( 𝑝 𝑖 ℎ𝑖 ) 5
𝑠𝑖 = sinh( 𝑝𝑖 ℎ𝑖 ) = 𝑝𝑖 ℎ𝑖 + + +....
6 120
Moreover, for 𝑖 = 1, . . . , 𝑁 − 1, the row sum given by
! ! !
𝑝 ℎ 𝑐 − 𝑠 𝑝 ℎ 𝑐 − 𝑠 𝑠 − 𝑝 ℎ 𝑠 − 𝑝 ℎ
𝑟𝑖𝑐 + 𝑟𝑖− + 𝑟𝑖+ =
𝑖 𝑖 𝑖 𝑖 𝑖+1 𝑖+1 𝑖+1 𝑖+1 𝑖 𝑖 𝑖 𝑖+1 𝑖+1 𝑖+1
+ 𝑏𝑖 + 𝑏𝑖−1 + 𝑏𝑖+1
𝑝𝑖2 𝑠𝑖 ℎ𝑖 2 𝑠 ℎ
𝑝𝑖+1 𝑖+1 𝑖+1 𝑝𝑖2 𝑠𝑖 ℎ𝑖 2 𝑠 ℎ
𝑝𝑖+1 𝑖+1 𝑖+1
1 1
   
𝜌 𝑐𝑖 𝑐𝑖+1 𝜌
≥ + − +
ℎ𝑖 + ℎ𝑖+1 𝑝𝑖 𝑠𝑖 𝑝𝑖+1 𝑠𝑖+1 ℎ𝑖 + ℎ𝑖+1 𝑝𝑖 𝑠𝑖 𝑝𝑖+1 𝑠𝑖+1
87 3. Reaction-Diffusion Problem with Robin Boundary Conditions

𝑐𝑖 − 1 𝑐𝑖+1 − 1
 
𝜌
≥ +
ℎ𝑖 + ℎ𝑖+1 𝑝𝑖 𝑠𝑖 𝑝𝑖+1 𝑠𝑖+1
( 𝑝 𝑖 ℎ𝑖 ) 2 ( 𝑝 𝑖 ℎ𝑖 ) 4
> 0 since 𝑐𝑖 − 1 = )+ + · · · > 0.
2 24
For 𝑖 = 0, 𝑁, the associated coefficients of 𝐵𝑙𝑁 and 𝐵𝑟𝑁 discretizing the robin boundary
conditions are given by
√ 2
!
𝛽0 𝛽0 𝜖 𝛼0 𝜖 0 ℎ
𝑟 0+ = 2 −𝜖 + 1 𝑏 1 ,
𝑐 𝛽
𝑟0 = 𝑏0 + 2 + ,
3 ℎ1 ℎ1 ℎ1 6
2
! √
𝛽 1 ℎ 𝛼1 𝜖 𝛽1 𝛽1 𝜖
𝑟 𝑁− = 2 −𝜖 + 𝑁 𝑏 𝑁−1 , 𝑐
𝑟𝑁 = + 𝑏𝑁 + 2 .
ℎ𝑁 6 ℎ𝑁 3 ℎ𝑁

Using Lemma 3.2.4 and (3.30), we get 𝑟 0𝑐 > 0, 𝑟 0+ < 0, 𝑟 𝑁− < 0, and 𝑟 𝑁𝑐 > 0. Moreover,
the row sum 𝑟 0𝑐 + 𝑟 0+ > 0 and 𝑟 𝑁− + 𝑟 𝑁𝑐 > 0. Hence, the coefficient matrix (3.28)–(3.29)
generated by the spline discretization technique (3.26)–(3.27) is an M-matrix. 

From Lemma 3.3.1, it is evident that the coefficient matrix associated with the
spline discretization technique (3.26)–(3.27) is an M-matrix. As a result, the corre-
sponding operator 𝐿 𝑁 satisfies the following discrete maximum principle.

Lemma 3.3.2. The Discrete Maximum Principle: For the discrete spline operator
𝐿 𝑁 , assume that the grid function 𝐺 satisfies 𝐵𝑙𝑁 𝐺 0 ≥ 0, 𝐵𝑟𝑁 𝐺 𝑁 ≥ 0. Then, 𝐿 𝑁 𝐺 𝑖 ≥ 0
for 𝑖 = 1, . . . , 𝑁 − 1 implies that 𝐺 𝑖 ≥ 0, ∀𝑖 = 0, . . . , 𝑁.

Proof. Let 𝑘 be such that 𝐺 𝑘 = min {𝐺 𝑖 } and assume that 𝐺 𝑘 < 0. It is clear that
𝑖
𝑘 ∉ {0, 𝑁 }, 𝐺 𝑘+1 − 𝐺 𝑘 ≥ 0 and 𝐺 𝑘 − 𝐺 𝑘−1 ≤ 0. Therefore

𝐿 𝑁 𝐺 𝑘 = 𝑟 𝑘− 𝐺 𝑘−1 + 𝑟 𝑘𝑐 𝐺 𝑘 + 𝑟 𝑘+ 𝐺 𝑘+1 < 0 since 𝑟 𝑘− < 0, 𝑟 𝑘+ < 0, and 𝐺 𝑘 < 0.

which is a contradiction. It follows that 𝐺 𝑘 > 0 and thus 𝐺 𝑖 ≥ 0, ∀𝑖 = 0, . . . , 𝑁. 

An immediate consequence of this discrete maximum principle is the following

𝜖-uniform stability result.

Lemma 3.3.3. Let 𝐺 be any grid function satisfying the hypothesis of the above
lemma. Then
1 𝑁
 
k𝐺 k Ω𝐸 ≤𝐶 𝐿 𝐺 Ω𝑁
𝐸
𝑁 𝜌
3.4. Error Analysis 88

𝐸
where Ω𝑁 is the layer-adapted equidistributed grid.

Proof. Let
1 𝑁
𝑉𝑖± = 𝐿 𝐺 Ω𝑁
𝐸 ± 𝐺 𝑖 , 𝑖 = 0, . . . , 𝑁.
𝜌
Clearly, 𝑉0± ≥ 0 and 𝑉𝑁± ≥ 0. Moreover, for 𝑖 = 0, . . . , 𝑁 − 1

1 𝑁
𝐿 𝑁 𝑉𝑖± = 𝐿 𝐺 Ω𝑁
𝐸 𝑏𝑖 ± 𝐿 𝑁 𝐺 𝑖 ≥ 0
𝜌

because 𝑏𝑖 ≥ 𝜌. The discrete maximum principle then implies that 𝑉𝑖± ≥ 0, ∀𝑖 =

0, . . . , 𝑁 as required. 

3.4 Error Analysis

In this section, we analyze the convergence of the proposed spline discretization
technique by estimating the error in the computed approximation. Analogous to the
continuous solution, we decompose the numerical approximation 𝑈 (𝑥𝑖 ) = 𝑈𝑖 into
smooth and layer components as 𝑈𝑖 = 𝑉𝑖 + 𝑊𝑖 , 𝑖 = 0, . . . , 𝑁. The smooth part of the
decomposition satisfies

𝐿 𝑁 𝑉𝑖 = 𝑓 (𝑥𝑖 ), 𝐵𝑙𝑁 𝑉0 = 𝐵𝑙 𝑣(0), 𝐵𝑟𝑁 𝑉𝑁 = 𝐵𝑟 𝑣(1),

whereas the layer part satisfies

𝐿 𝑁 𝑊𝑖 = 0, 𝐵𝑙𝑁 𝑊0 = 𝐵𝑙 𝑤(0), 𝐵𝑟𝑁 𝑊𝑁 = 𝐵𝑟 𝑤(1).

Using the triangle inequality, the truncation error of the numerically approximated
solution at the nodes can be expressed as

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

To estimate the consistency of the spline discretization technique, we compute the

truncation error associated with the numerical solution by using the triangle in-
equality. In this sense, we estimate the errors associated with the smooth and layer
components of the numerically approximated solution respectively at every node of
the equidistributed grid. For any function 𝜑(𝑥), we define the truncation error at 𝑥𝑖 to
be e𝑖 = (𝐿𝜑)(𝑥𝑖 ) − 𝐿 𝑁 𝜑𝑖 . We begin our analysis with the smooth part of the solution.
𝜑
89 3. Reaction-Diffusion Problem with Robin Boundary Conditions

3.4.1 Error due to the Smooth Part

Let us partition the domain based on whether the grid point 𝑥𝑖 lies inside or outside
the boundary layer region.

Lemma 3.4.1. For 𝑖 = 1, 2, . . . , 𝑁 − 1, the error associated with the smooth part 𝑉
satisfies
e𝑉𝑖 ≤ 𝐶𝑁 −2 , 𝑖 = 1, 2, . . . , 𝑁 − 1.

Proof. The truncation error at node 𝑥𝑖 is estimated using Taylor’s expansion as

e𝑉𝑖 ≤ 𝐶𝜖 (ℎ𝑖−1 + ℎ𝑖 ) 𝑝 k𝑣 000 (𝑥) k [𝑥𝑖−1 , 𝑥𝑖+1 ] ,

√︃ 
√ 1
where 𝑝 = min {𝑝𝑖 } = max 𝑏𝑖
and 𝜖 < 𝑁. By combining the crude bounds
1≤𝑖≤𝑁−1 1≤𝑖≤𝑁−1 𝜖
of 𝑣 in Lemma 3.2.2 with the bounds of grid spacing in Lemma 3.2.5, we obtain
|e𝑉𝑖 | ≤ 𝐶𝑁 −2 . 

3.4.2 Error due to the Layer Part

The error associated with the layer part of the solution is estimated in the following
lemma.

Lemma 3.4.2. For 𝑖 = 𝑘 𝑙 , . . . , 𝑘 𝑟 , the error associated with the layer part 𝑊 satisfies

−2
e𝑊
𝑖 ≤ 𝐶𝑁 .

Proof. Use mean value theorem to obtain the truncation error of 𝑊 as

 √︃ 
 exp −𝑥𝑖−1 𝜖 , 𝑥𝑖 ≤ 12

 𝑏(0)



 

𝑊 00
 

e𝑖 ≤ 𝐶𝜖 k𝑤 (𝑥)k [𝑥𝑖−1 ,𝑥𝑖+1 ] ≤ 𝐶  √︃ 
1
 exp −(1 − 𝑥𝑖+1 ) 𝜖 , 𝑥𝑖 > 2

 𝑏(1) 

 

 
For 𝑖 ≥ 𝑘 𝑙 and 𝑥𝑖 ≤ 21 , apply Lemma 3.2.3 to get
√︂ ! √︂ !! 2
𝑏(0) −𝑥 𝑘 𝑙 −1 𝑏(0)
𝑖 ≤ 𝐶 exp −𝑥 𝑘 𝑙 −1
e𝑊 = 𝐶 exp ≤ 𝐶𝑁 −2 .
𝜖 2 𝜖
3.4. Error Analysis 90

For 𝑖 ≤ 𝑘 𝑟 and 𝑥𝑖 > 12 , the application of Lemma 3.2.3 again yields

√︂ ! √︂ !! 2
𝑏(1) −(1 − 𝑥 𝑘 𝑟 +1 ) 𝑏(1)
𝑖 ≤ 𝐶 exp − 1 − 𝑥 𝑘 𝑟 +1 = 𝐶 exp ≤ 𝐶𝑁 −2 . 


e𝑊
𝜖 2 𝜖

Lemma 3.4.3. For 𝑖 = 1, . . . , 𝑘 𝑙 − 1 and 𝑖 = 𝑘 𝑟 + 1, . . . , 𝑁 − 1, the layer part 𝑊 satisfies

−2
e𝑊
𝑖 ≤ 𝐶𝑁 .

Proof. Consider the left-hand boundary layer region. The Taylor’s expansion leads
to
2 𝑤 000 (𝜃 (1) ) − ℎ2 𝑤 000 (𝜃 (2) )
𝜖 ℎ𝑖+1 𝑖 𝑖 𝑖
e𝑊 =
𝑖
3(ℎ𝑖 + ℎ𝑖+1 )
where 𝜃 𝑖(1) ∈ (𝑥𝑖 , 𝑥𝑖+1 ) and 𝜃 𝑖(2) ∈ (𝑥𝑖−1 , 𝑥𝑖 ). From Lemma 3.2.4, we get

2
ℎ𝑖+1 𝑤 000 (𝜃 𝑖(1) ) − ℎ𝑖2 𝑤 000 (𝜃 𝑖(2) ) ≤ 2
ℎ𝑖+1 − ℎ𝑖2 𝑤 000 (𝜃 𝑖(1) ) + ℎ𝑖2 𝑤 000 (𝜃 𝑖(1) ) − 𝑤 000 (𝜃 𝑖(2) )
 
2
≤ 𝐶 ℎ𝑖+1 − ℎ𝑖2 |𝑤 000 (𝑥𝑖 )| + ℎ𝑖2 (ℎ𝑖 + ℎ𝑖+1 ) 𝑤 𝑖𝑣 (𝑥𝑖 )

where 𝜃 𝑖(1) − 𝜃 𝑖(2) < (ℎ𝑖 + ℎ𝑖+1 ). This leads to

√︂ !
√ 𝐶 𝑏(0)
e𝑊
𝑖 ≤ 𝐶 𝜖 ℎ𝑖2 |𝑤 000 (𝑥𝑖 )| ≤ ℎ𝑖2 exp −𝑥𝑖
𝜖 𝜖
! !2 2
𝐶Ψ2
√︂
𝐶 √
∫ 𝑥𝑖  ∫ 𝑥𝑖
𝐶 𝑡 𝑏(0) 𝐶
≤ exp − 𝑑𝑡 ≤ 𝜖 𝑀 (𝑡, 𝑢(𝑡))𝑑𝑡 ≤ 2
≤ 2.
𝜖 𝑥𝑖−1 2 𝜖 𝜖 𝑥 𝑖−1 𝑁 𝑁

A similar estimation can be done for the right-hand boundary layer region. Hence
the result. 

The next lemma determines the truncation error of the numerical approximation
at the two ends of the domain.

Lemma 3.4.4. The error bound in approximating the robin boundary conditions at
the two extremities 𝑥 0 = 0 and 𝑥 𝑁 = 1 is given by

𝑖 ≤ 𝐶𝑁 , where 𝑖 = 0, 𝑁.
−2
e𝑈

Proof. To estimate the error associated with 𝑈 at 𝑥 0 and 𝑥 𝑁 , we use Lemmas 3.2.1,
91 3. Reaction-Diffusion Problem with Robin Boundary Conditions

√
3.2.4, the assumption 𝜖 << 𝑁 −1 and Taylor’s expansion to get
√ 2 (𝑖𝑣)
|e𝑈
𝑖 | ≤ 𝐶 𝜖 𝛽0 𝜖 ℎ 1 𝑢 (𝑥)
(𝑥 0 , 𝑥 1 )
√︂   √︂ 
5 𝜌 𝜌
≤ 𝐶 𝛽0 𝜖 (𝜃 0 𝑁)
2 −2
(1 + 𝜖 −2
exp −𝑥 + 𝜖 exp −(1 − 𝑥)
−2
𝜖 𝜖
√
𝜖 𝐶
≤ 𝐶 2 ≤ 2.
𝑁 𝑁
The truncation error bound at the endpoint 𝑥 𝑁 = 1 can be determined in the same
way. Hence the proof. 

3.4.3 The Principal Convergence Theorem

After estimating the errors, we now establish uniform second-order convergence of
the proposed spline discretization technique (3.26)–(3.27) in the following theorem.

Theorem 3.4.1. Let 𝑈 be an approximation to the solution 𝑢 of the problem (3.1)

obtained by using the spline discretization technique (3.26)–(3.27) on a layer-adapted
grid generated by equidistributing the monitor function (3.2). Then, for a sufficiently
large value of 𝑁 satisfying (3.30), there exists a constant 𝐶 independent of 𝑁 and 𝜖
such that
|𝑢(𝑥𝑖 ) − 𝑈𝑖 | ≤ 𝐶𝑁 −2 , for 𝑖 = 0, 1, . . . 𝑁.

Proof. Lemma 3.3.1 states that the coefficient matrix of the proposed spline dis-
cretization technique (3.26)–(3.27) is an M-matrix. This immediately implies that the
associated discrete operator is inverse monotone and uniformly stable with respect
to the perturbation parameter. It then follows from the error estimates computed in
Lemmas 3.4.1–3.4.4 that

|𝑢(𝑥𝑖 ) − 𝑈𝑖 | ≤ 𝐶𝑁 −2 , for 𝑖 = 0, 1, . . . , 𝑁. 

The above result asserts that the proposed spline discretization technique is uni-
formly accurate at all the nodes of the equidistributed grid. Intending to extend the
analysis globally, we define a piecewise linear interpolant of the numerical approxima-
tion. In the following theorem, we show that this global approximation is uniformly
accurate at all the points of the domain.
3.5. Numerical Experiments 92

Theorem 3.4.2. Let 𝑈 be the piecewise linear interpolant of the approximate solution
𝑈 obtained by using the spline discretization technique. Then, for all 𝑥 ∈ [0, 1]

𝑢 − 𝑈 ≤ 𝐶𝑁 −2 .

Proof. At every node of the equidistributed grid, we use the piecewise linear inter-
polant of the exact solution 𝑢 as 𝑢 and the triangle inequality as

𝑢 − 𝑈 ≤ k𝑢 − 𝑢k + 𝑢 − 𝑈 .

The truncation error in the first term on the RHS can be determined using the
pointwise error estimation of the decomposed solution [245]. Using Theorem 3.4.1,
the bounds of the second term on the RHS can also be estimated to obtain the
result. 

3.5 Numerical Experiments

In this section, we present the numerical results obtained by applying the proposed
technique on two test examples. The layer-adapted grid is generated by equidis-
tributing the monitor function (3.2) using the adaptive grid generation Algorithm 2.1
presented in Section 2.5.1 of Chapter 2.

Example 3.5.1. Consider the following problem

−𝜖𝑢00 (𝑥) + 𝑢(𝑥) = − cos2 (𝜋𝑥) − 2𝜖 𝜋 2 cos(2𝜋𝑥), 𝑥 ∈ (0, 1),

−1
√ −1
√
√ 4𝑒 𝜖 √ 3(1 − 𝑒 𝜖 )
𝑢(0) − 2 𝜖𝑢0 (0) = 2 − −1
and 𝑢(1) + 3 𝜖𝑢0 (1) = −1
.
√ √
1+𝑒 𝜖 1+𝑒 𝜖
The exact solution to the problem is given by
−(1−𝑥) −𝑥
√ √
𝑒 𝜖 +𝑒 𝜖
𝑢(𝑥) = −1
− cos2 (𝜋𝑥).
√
1+𝑒 𝜖

Example 3.5.2. Consider the following problem

− 𝜖𝑢00 (𝑥) + (1 + 𝑥(1 − 𝑥))𝑢(𝑥) = 𝑓 (𝑥),

√ −𝑥 √ −(1−𝑥)
𝑓 (𝑥) = 1 + 𝑥(1 − 𝑥) + [2 𝜖 − 𝑥(1 − 𝑥) 2 ]𝑒 𝜖 + [2 𝜖 − 𝑥 2 (1 − 𝑥)]𝑒 𝜖 ,
√ √

√ √ −1
√ √ −1
√ √
𝑢(0) − 𝜖𝑢0 (0) = −1 − 𝜖 + 𝑒 𝜖 and 𝑢(1) + 𝜖𝑢0 (1) = 𝑒 𝜖 − 1 − 𝜖 .
93 3. Reaction-Diffusion Problem with Robin Boundary Conditions

−𝑥 −(1−𝑥)
√ √
The exact solution to the problem is given by 𝑢(𝑥) = 1 + (𝑥 − 1)𝑒 𝜖 − 𝑥𝑒 𝜖 .

The maximum pointwise errors in the numerical solution are computed as

𝐸 𝜖𝑁 = max 𝑢(𝑥𝑖 ) − 𝑈 𝑁 (𝑥𝑖 ) ,
𝑖=0,...,𝑁
where 𝑈 𝑁 (𝑥 𝑖) is an approximation of the solution 𝑢(𝑥𝑖 ) generated by the spline
discretization technique (3.26)–(3.27) at the node {𝑥𝑖 } of the layer-adapted grid. The
rate of convergence is determined as

ln 𝐸 𝜖𝑁 − ln 𝐸 𝜖2𝑁
 𝑁 
𝐸𝜖
𝑁
𝑝 = = log2 .
ln 2 𝐸 𝜖2𝑁

For every 𝑁, we define the uniform error as 𝐸 𝑁 = max 𝐸 𝜖𝑁 over a wide range of 𝜖. The
𝜖
user chosen constant 𝑄 in Algorithm 2.1 is taken to be 𝑄 = 1.3 for the construction
of layer adaptive grid. Tables 3.1 and 3.2 display the maximum norm errors and the
rates of convergence for the approximate solution computed on the layer-adapted grid
for the Examples 3.5.1 and 3.5.2, respectively. The error plots on logarithmic scale for
both the examples are shown in Figure 3.1. It can be clearly observed from the tables
and figure that the errors are uniform with respect to the perturbation parameter and
decrease monotonically with increase in 𝑁. Hence, the proposed spline discretization
technique is parameter uniform second-order convergent on layer-adapted grid.

For the purpose of comparison, we have additionally considered Bakhvalov,

Shishkin and Uniform mesh. Both the Bakhvalov and Shishkin meshes have been
constructed with the parameters 𝑞 = 1/3 and 𝜎 = 2.5. For Examples 3.5.1 and 3.5.2,
the maximum norm errors and the rates of convergence for the approximate solution
computed on Bakhvalov mesh are listed in Tables 3.3 and 3.4, and on Shishkin mesh
in Tables 3.5 and 3.6, respectively. The experimental results indicate that the layer-
adapted grid generated by the algorithm is similar to the graded Bakhvalov mesh in
the layer region and is better than Shishkin mesh. Furthermore, the maximum norm
errors obtained by applying the proposed spline discretization technique, exponential
spline difference scheme and cubic spline difference scheme on uniform mesh for
Examples 3.5.1 and 3.5.2 are given in Tables 3.7 and 3.8, respectively. The results
clearly indicate the superiority of the proposed discretization technique over the other
two techniques.
3.5. Numerical Experiments 94

Table 3.1: Max-norm errors and the order of convergence for the proposed spline
discretization scheme on layer-adapted grid for Example 3.5.1

Number of intervals
𝜖= 10−k N= 26 N= 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 2.048e-03 5.390e-04 1.355e-04 3.397e-05 8.504e-06 2.127e-06 5.319e-07

𝑝𝑁 1.93 1.99 2.00 2.00 2.00 2.00
𝑘=1 1.422e-03 3.208e-04 7.957e-05 2.005e-05 5.007e-06 1.252e-06 3.131e-07
𝑝𝑁 2.15 2.01 1.99 2.00 2.00 2.00
𝑘=2 1.622e-03 4.066e-04 1.017e-04 2.544e-05 6.359e-06 1.590e-06 3.975e-07
𝑝𝑁 2.00 2.00 2.00 2.00 2.00 2.00
𝑘=3 1.578e-03 3.998e-04 1.003e-04 2.509e-05 6.274e-06 1.569e-06 3.922e-07
𝑝𝑁 1.98 1.99 2.00 2.00 2.00 2.00
𝑘=4 1.390e-03 3.859e-04 9.938e-05 2.504e-05 6.271e-06 1.568e-06 3.922e-07
𝑝𝑁 1.85 1.96 1.99 2.00 2.00 2.00
𝑘=5 7.753e-04 2.971e-04 9.146e-05 2.448e-05 6.235e-06 1.566e-06 3.920e-07
𝑝𝑁 1.38 1.70 1.90 1.97 1.99 2.00
𝑘=6 2.877e-04 1.344e-04 5.742e-05 2.040e-05 5.909e-06 1.544e-06 3.906e-07
𝑝𝑁 1.10 1.23 1.49 1.79 1.94 1.98
𝑘=7 9.525e-05 4.675e-05 2.240e-05 1.022e-05 4.147e-06 1.365e-06 3.774e-07
𝑝𝑁 1.03 1.06 1.13 1.30 1.60 1.85
𝑘=8 3.055e-05 1.521e-05 7.512e-06 3.658e-06 1.730e-06 7.665e-07 2.919e-07
𝑝𝑁 1.01 1.02 1.04 1.08 1.17 1.39
𝑘=9 9.545e-06 4.653e-06 2.228e-06 1.105e-06 4.898e-07 2.25e-07 1.027e-07
𝑝𝑁 1.04 1.07 1.08 1.10 1.12 1.13
𝑘 = 10 9.543e-07 4.621e-07 2.218e-07 1.06e-07 5.051e-08 2.39e-08 1.112e-08
𝑝𝑁 1.05 1.06 1.07 1.07 1.08 1.09
𝐸𝑁 2.048e-03 5.390e-04 1.355e-04 3.397e-05 8.504e-06 2.127e-06 5.319e-07
95 3. Reaction-Diffusion Problem with Robin Boundary Conditions

Table 3.2: Max-norm errors and the order of convergence for the proposed spline
discretization scheme on layer-adapted grid for Example 3.5.2

Number of intervals
𝜖= 10−k N= 26 N= 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 3.52e-05 9.37e-06 2.41e-06 6.10e-07 1.54e-07 3.85e-08 9.65e-09

𝑝𝑁 1.91 1.96 1.98 1.99 2.00 2.00
𝑘=1 2.55e-04 3.43e-05 8.57e-06 2.14e-06 5.36e-07 1.34e-07 3.35e-08
𝑝𝑁 2.89 2.00 2.00 2.00 2.00 2.00
𝑘=2 2.86e-04 7.13e-05 1.78e-05 4.45e-06 1.11e-06 2.78e-07 6.96e-08
𝑝𝑁 2.00 2.00 2.00 2.00 2.00 2.00
𝑘=3 9.60e-04 2.34e-04 5.82e-05 1.45e-05 3.63e-06 9.07e-07 2.27e-07
𝑝𝑁 2.04 2.01 2.00 2.00 2.00 2.00
𝑘=4 1.52e-03 4.99e-04 1.59e-04 4.61e-05 1.16e-05 2.91e-06 7.26e-07
𝑝𝑁 1.61 1.65 1.79 1.99 2.00 2.00
𝑘=5 3.10e-04 1.20e-04 4.20e-05 1.32e-05 3.80e-06 1.12e-06 3.30e-07
𝑝𝑁 1.37 1.51 1.67 1.79 1.76 1.77
𝑘=6 3.11e-05 1.21e-05 4.24e-06 1.27e-06 3.70e-07 1.10e-07 3.23e-08
𝑝𝑁 1.36 1.51 1.74 1.78 1.75 1.77
𝑘=7 3.11e-06 1.21e-06 4.26e-07 1.25e-07 3.69e-08 1.12e-08 3.42e-09
𝑝 𝑁 1.36 1.51 1.76 1.77 1.72 1.71
𝑘=8 3.11e-07 1.21e-07 4.25e-08 1.26e-08 3.71e-09 1.12e-09 3.36e-10
𝑝𝑁 1.36 1.51 1.76 1.76 1.73 1.73
𝑘=9 3.11e-08 1.20e-08 4.21e-09 1.26e-09 3.67e-10 1.11e-10 3.37e-11
𝑝𝑁 1.37 1.51 1.75 1.77 1.73 1.72
𝑘 = 10 3.11e-09 1.20e-09 4.18e-10 1.26e-10 3.80e-11 1.14e-11 3.37e-12
𝑝𝑁 1.37 1.52 1.74 1.72 1.74 1.76
𝐸𝑁 1.52e-03 4.99e-04 1.59e-04 4.61e-05 1.16e-05 2.91e-06 7.26e-07
3.5. Numerical Experiments 96

(a)

(b)

Figure 3.1: Max-norm error plotted on loglog scale for (a) Example 3.5.1 and (b)
Example 3.5.2
97 3. Reaction-Diffusion Problem with Robin Boundary Conditions

Table 3.3: Max-norm errors of the proposed spline difference scheme on Bakhvalov
mesh for Example 3.5.1

𝜖 = 10−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 3.025e-03 7.574e-04 1.894e-04 4.736e-05 1.184e-05 2.960e-06 7.395e-07

𝑘=1 2.236e-03 5.598e-04 1.400e-04 3.501e-05 8.752e-06 2.188e-06 5.470e-07
𝑘=2 1.622e-03 4.066e-04 1.017e-04 2.544e-05 6.359e-06 1.590e-06 3.975e-07
𝑘=3 1.578e-03 3.998e-04 1.003e-04 2.509e-05 6.274e-06 1.569e-06 3.922e-07
𝑘=4 1.390e-03 3.859e-04 9.938e-05 2.504e-05 6.271e-06 1.568e-06 3.922e-07
𝑘=5 7.753e-04 2.971e-04 9.146e-05 2.448e-05 6.235e-06 1.566e-06 3.920e-07
𝑘=6 2.877e-04 1.344e-04 5.742e-05 2.040e-05 5.909e-06 1.544e-06 3.906e-07
𝑘=7 9.525e-05 4.675e-05 2.240e-05 1.022e-05 4.147e-06 1.365e-06 3.774e-07
𝑘=8 3.055e-05 1.521e-05 7.512e-06 3.658e-06 1.730e-06 7.665e-07 2.919e-07
𝑘=9 9.545e-06 4.653e-06 2.228e-06 1.105e-06 4.898e-07 2.25e-07 1.027e-07
𝑘 = 10 9.543e-07 4.621e-07 2.218e-07 1.06e-07 5.051e-08 2.39e-08 1.112e-08

Table 3.4: Max-norm errors of the proposed spline difference scheme on Bakhvalov
mesh for Example 3.5.2

𝜖 = 10−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 2.234e-05 5.586e-06 1.396e-06 3.491e-07 8.729e-08 2.187e-08 5.675e-09

𝑘=1 1.372e-04 3.429e-05 8.573e-06 2.143e-06 5.358e-07 1.340e-07 3.350e-08
𝑘=2 2.858e-04 7.126e-05 1.780e-05 4.451e-06 1.113e-06 2.782e-07 6.955e-08
𝑘=3 9.60e-04 2.34e-04 5.82e-05 1.45e-05 3.63e-06 9.07e-07 2.27e-07
𝑘=4 1.52e-03 7.161e-04 1.847e-04 4.605e-05 1.163e-05 2.905e-06 7.261e-07
𝑘=5 3.10e-04 2.89e-04 2.01e-04 1.02e-04 3.648e-05 9.230e-06 2.302e-06
𝑘=6 3.11e-05 1.21e-05 9.24e-06 5.27e-06 9.70e-07 4.10e-07 1.23e-07
𝑘=7 3.11e-06 1.21e-06 1.01e-06 9.25e-07 5.69e-07 1.12e-07 5.42e-08
𝑘=8 3.11e-07 1.24e-07 1.017e-07 9.35e-08 5.782e-08 1.134e-08 5.46e-09
𝑘=9 3.11e-08 1.20e-08 1.03e-08 9.425e-09 5.632e-09 1.172e-09 5.63e-10
𝑘 = 10 3.11e-09 1.267e-09 1.024e-09 9.436e-10 5.712e-10 1.136e-10 5.45e-11
3.5. Numerical Experiments 98

Table 3.5: Max-norm errors of the proposed spline difference scheme on Shishkin
mesh for Example 3.5.1

𝜖 = 10−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 3.025e-03 7.574e-04 1.894e-04 4.736e-05 1.184e-05 2.960e-06 7.395e-07

𝑘=1 2.236e-03 5.598e-04 1.400e-04 3.501e-05 8.752e-06 2.188e-06 5.470e-07
𝑘=2 4.419e-03 1.023e-03 2.346e-04 5.350e-05 1.213e-05 2.735e-06 6.119e-07
𝑘=3 5.793e-03 1.513e-03 3.805e-04 9.475e-05 2.354e-05 5.843e-06 1.450e-06
𝑘=4 4.151e-03 1.385e-03 3.841e-04 9.883e-05 2.488e-05 6.227e-06 1.556e-06
𝑘=5 1.729e-03 7.751e-04 2.969e-04 9.141e-05 2.447e-05 6.231e-06 1.565e-06
𝑘=6 5.892e-04 2.877e-04 1.344e-04 5.741e-05 2.039e-05 5.909e-06 1.544e-06
𝑘=7 1.906e-04 9.525e-05 4.675e-05 2.240e-05 1.022e-05 4.147e-06 1.365e-06
𝑘=8 6.070e-05 3.055e-05 1.521e-05 7.512e-06 3.658e-06 1.730e-06 7.665e-07
𝑘=9 9.645e-06 7.672e-06 4.853e-06 2.418e-06 1.199e-06 5.898e-07 2.85e-07
𝑘 = 10 9.544e-07 7.821e-07 5.218e-07 4.113e-07 3.835e-07 1.908e-07 9.441e-08

Table 3.6: Max-norm errors of the proposed spline difference scheme on Shishkin
mesh for Example 3.5.2

𝜖 = 10−k N = 26 N = 27 N = 28 N = 29 N = 210 N = 211 N = 212

𝑘=0 2.234e-05 5.586e-06 1.396e-06 3.491e-07 8.729e-08 2.187e-08 5.675e-09

𝑘=1 1.372e-04 3.429e-05 8.573e-06 2.143e-06 5.358e-07 1.340e-07 3.350e-08
𝑘=2 4.862e-04 9.807e-05 1.979e-05 4.049e-06 8.477e-07 1.839e-07 4.543e-08
𝑘=3 2.446e-04 8.010e-04 1.907e-04 4.706e-05 1.128e-05 2.702e-06 6.452e-07
𝑘=4 1.288e-03 1.121e-03 6.632e-04 1.824e-04 4.509e-05 1.120e-05 2.787e-06
𝑘=5 1.474e-04 1.212e-04 1.181e-04 1.02e-04 9.948e-05 3.588e-05 9.208e-06
𝑘=6 1.500e-05 1.33e-05 1.26e-05 1.18e-05 1.01e-05 9.79e-06 1.93e-06
𝑘=7 1.508e-06 1.235e-06 1.120e-06 1.108e-06 1.001e-06 9.22e-07 3.42e-07
𝑘=8 1.510e-07 1.233e-07 1.121e-07 1.107e-07 1.003e-07 9.422e-08 3.62e-08
𝑘=9 6.311e-08 2.256e-08 1.83e-08 1.612e-08 1.304e-08 1.172e-08 9.889e-09
𝑘 = 10 7.414e-09 3.892e-09 2.411e-09 1.320e-09 1.209e-09 1.012e-09 9.884e-10
99 3. Reaction-Diffusion Problem with Robin Boundary Conditions

Table 3.7: Max-norm errors of spline difference schemes on Uniform mesh for Exam-
ple 3.5.1

N Proposed Exponential Spline Cubic Spline

Scheme Difference Scheme Difference Scheme

𝑁 = 25 1.390e-03 3.93e-03 2.96e-03

𝑁 = 26 3.859e-04 1.22e-03 1.389e-03
𝑁 = 27 9.938e-05 3.53e-04 3.847e-04
𝑁 = 28 2.504e-05 2.96e-05 9.886e-05
𝑁 = 29 6.271e-06 3.66e-05 2.492e-05
𝑁 = 210 1.568e-06 1.98e-06 6.234e-06

Table 3.8: Max-norm errors of spline difference schemes on Uniform mesh for Exam-
ple 3.5.2

N Proposed Exponential Spline Cubic Spline

Scheme Difference Scheme Difference Scheme

𝑁 = 25 1.52e-03 2.86e-03 1.82e-02

𝑁 = 26 4.99E-04 1.03e-03 3.56e-03
𝑁 = 27 1.59E-04 3.48e-04 7.24e-04
𝑁 = 28 4.61E-05 1.14e-04 3.02e-04
𝑁 = 29 1.16E-05 3.61e-05 8.75e-05
𝑁 = 210 2.91E-06 1.11e-05 4.01e-05
3.6. Conclusions 100

3.6 Conclusions
We have presented a robust computational technique to approximate the solution of a
reaction-diffusion problem with robin boundary conditions. The proposed discretiza-
tion technique uses cubic splines at the robin boundary conditions and exponential
splines with a uniform tension parameter at the internal nodes of a layer-adapted grid.
The grid is generated by equidistributing a positive monitor function via an adaptive
grid generation algorithm. It is shown that the error bounds of the computed solution
yield parameter-uniform second-order convergence in discrete maximum norm. The
numerical results obtained for the proposed discretization technique are uniformly
accurate and confirm the theoretical error estimates. Furthermore, since the spline
functions are continuous throughout the domain, the proposed spline discretization
technique ensures a solution to the problem at any point of the domain. Hence, it is
advantageous to use a spline difference scheme compared to the standard difference
approximation techniques that can generate the solution only at the nodal points of
the grid.