EJ-MATH, European Journal of Mathematics and Statistics

ISSN: 2736-5484

Numerical Approximations of a Class of

Nonlinear Second-Order Boundary Value
Problems using Galerkin-Compact Finite
Difference Method
Shovan Sourav Datta Pranta and Md. Shafiqul Islam

Abstract — In this study, we examine numerical approximations for 2nd-order linear-nonlinear

differential equations with diverse boundary conditions, followed by the residual corrections of
the first approximations. We first obtain numerical results using the Galerkin weighted residual
approach with Bernstein polynomials. The generation of residuals is brought on by the fact that
our first approximation is computed using numerical methods. To minimize these residuals, we
use the compact finite difference scheme of 4th-order convergence to solve the error differential
equations in accordance with the error boundary conditions. We also introduce the formulation
of the compact finite difference method of fourth-order convergence for the nonlinear BVPs. The
improved approximations are produced by adding the error values derived from the
approximations of the error differential equation to the weighted residual values. Numerical
results are compared to the exact solutions and to the solutions available in the published
literature to validate the proposed scheme, and high accuracy is achieved in all cases.

Keywords — Compact Finite Difference Method; Galerkin Method; Nonlinear BVPs; Residual
Correction.

I. INTRODUCTION
The capacity to provide the nature of the solution to any physical occurrence, even when analytical
answers are not possible, is a key advantage of the numerical approach. In addition, a numerical method
requires simply the evaluation of standard functions and the four operations of addition, subtraction,
multiplication, and division. There has been a significant increase in interest in the study of boundary value
problems for second- and higher-order differential equations because both linear and nonlinear differential
equations have the ability to replicate a wide variety of natural processes. They are also implemented in a
variety of scientific and engineering applications.
In this study, we consider the following generic form of a second-order linear-nonlinear boundary value
problem:

d2 f  df 
= g  x, f ,  a xb (1)
dx 2  dx 

in accordance with the boundary conditions,

df
a1 f ( a ) + b1 = 1
dx x=a
(2)
df
a2 f ( b ) + b2 = 2
dx x =b

where 𝑎1 , 𝑎2 , 𝑏1 , 𝑏2 , λ1 , λ2 are some constants. If

1. 𝑏1 = 𝑏2 = 0 and 𝑎1 ≠ 0, 𝑎2 ≠ 0, (2) is referred to as Dirichlet boundary conditions.
2. 𝑎1 = 𝑎2 = 0 and 𝑏1 ≠ 0, 𝑏2 ≠ 0, (2) is referred to as Neumann boundary conditions.
3. If all of the parameters 𝑎1 , 𝑎2 , 𝑏1 , 𝑏2 are not equal to 0, the Robin boundary condition is established.

In the field of numerical analysis, there are a lot of different ways to deal with boundary value problems

Published on August 31, 2023.

S. S. Datta Pranta, Department of Applied Mathematics, University of Dhaka, Bangladesh.
(e-mail: [email protected])
Md. Shafiqul Islam, Department of Applied Mathematics, University of Dhaka, Bangladesh.
(corresponding author and e-mail: [email protected])

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 56

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

(BVPs). In their respective books, Bender et al. [1] and Collatz [2] provided a full theory and some basic
numerical treatments of boundary value problems. Kanth and Reddy [3] employed the cubic spline
technique for solving two-point boundary value problems numerically. Then using the same technique,
Kanth and Bhattacharya [4] analyzed and numerically resolved a class of nonlinear boundary value
problems appearing in physiology. Sibana et al. [5] used a new spectral-homotopy model to approximate
the solution of second-order nonlinear BVPs. A class of second-order differential equations was solved
numerically by Islam and Shirin [6] using a weighted residual technique called the Galerkin method via
Bernoulli polynomials. In certain instances, however, a significant number of Bernoulli polynomials are
necessary to achieve the needed precision, necessitating a substantial amount of computational time.
Burden and Faires [7] explored certain numerical techniques for handling boundary value problems in their
book on numerical analysis, such as the shooting method and the finite difference method. While using the
parametric difference approaches, Pandey [8] solved the two-point boundary value problems. Ramos and
Rufai [9] implemented a third-derivative, two-step block Falkner approach to solve linear and non-linear
BVPs. Some stochastic nonlinear second-order boundary value problems driven by additive noise have
been solved by Baccouch [10] using the finite difference method. Numerous physical phenomena are
modeled using strongly non-linear BVPs with specific parameter values. In particular, Bratu's problem is a
special type of nonlinear BVP of order two based on eigenvalues. Complex physical and chemical models
are frequently described using Bratu's problem in both science and engineering. This problem is employed
in a wide range of applications, such as the combustion theory's fuel ignition model, the thermal reaction
mechanism, the Chandrasekhar model of universe expansion and chemical reaction theory, radiative heat
transmission, and nanotechnology. It has been solved and analysed by different authors using different
methods [11]-[16] in the literature.
The Galerkin method is a very well-known method for numerical approximations of various types of
problems. Numerous authors come up with various types of problems and solve them numerically using
the Galerkin method [17]-[21]. In recent years, the implicit finite difference methodology, also known as
the Compact Finite Difference method, has gained popularity. Lele [22] and Mehra and Patel [23] gave a
discussion on a variety of ordering schemes in addition to a set of compact finite difference approaches.
While Malele et al. [24] used the highly precise compact finite difference approach for the solution of BVPs
with boundary conditions of Robin type.
In this study, an algorithm is used to improve the accuracy of computation for a class of second-order
linear and nonlinear BVPs. For the proposed scheme, we employ the well-known Galerkin and Compact
Finite Difference methods. We also introduce the mathematical formulation of 4th-order compact finite
difference scheme for the general nonlinear BVPs. We use the Galerkin method to find our first
approximation. Then, we use the compact finite difference method to solve the error differential equation
in order to improve the accuracy of our first approximation.

II. GALERKIN METHOD FOR BVPS

The modified Galerkin method is a remarkably potent numerical technique for approximating boundary
value problems. The fundamental method employs a trial solution that satisfies all of the problem's
boundary conditions.
Let us consider the trial approximation [25] of such a function 𝑓(𝑥) of the differential equation (1), given
by,

n
f ( x ) = N0 ( x ) +  j N j ( x ) (3)
j =1

where α𝑗 denotes the undetermined parameters and 𝑁𝑗 (𝑥) denotes the basis functions (here Bernstein
Polynomials).
The preceding is the generic form of Bernstein polynomials of degree 𝑝 over the interval [𝑥0 , 𝑥𝑛 ]:

 p  ( x − x ) ( x − x0 )
p −i i

B p ,i ( x ) =   n i = 0,1, 2,....... p (4)

( xn − x0 )
p
i

Thus, over the interval [0,1]. The Bernstein polynomials of degree 𝑝 can be represented as:

 p
B p ,i ( x ) =   (1 − x ) x i
p −i
i = 0,1, 2,., p (5)
i

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 57

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

These Bernstein polynomials of degree 𝑝 over the interval [0,1] have some special properties. One of
them is, 𝐵𝑝,𝑖 (0) = 0 = 𝐵𝑝,𝑖 (1) for 𝑖 = 1,2,3, … . , 𝑝 − 1. So, they can be used as the basis of the Galerkin
method in the case of Dirichlet boundary conditions.
Now, the Galerkin weighted residual equation becomes [25]:

d2 f  df  
a  dx2 − g  x, f , dx  Ni ( x ) dx = 0
b
(6)
 

After applying integration by parts to the second derivative part of (6) and then substituting the trial
solution (3) into it, we get:

 dN 0 n dN j  
b
dN j  dNi  n
dN 0 n  df 
  +  j
b
a  dx j =1 dx  dx 
  +  j  + g  x , N 0 +  j N j ,
dx
 Ni dx =  Ni 
dx    dx  a
(7)
 j =1 j =1

Or equivalently, they can be written as the matrix form as


j =1
j Kij = Fi or K = F (8)

where

b  dN dN j  dN j  
Kij =   i + Q1  x, N j ,  Ni dx (9)
a
 dx dx  dx  
b
 df  b  dN dN  dN  
Fi =  Ni  −   0 i
+ Q0  x, N 0 , 0  N i dx (10)
 dx  a
a
 dx dx  dx  

Solving the system of equations gives us the values of the unknowns, α. In the case of nonlinear BVPs,
iteration is required to find the value of these unknown coefficients. Then, the first approximate solution of
the BVP is obtained by substituting the values of α's into (3).

III. COMPACT FINITE DIFFERENCE METHOD FOR BVPS

Firstly, consider the generic form of the differential equation (1) in accordance with the boundary
conditions (2). The domain in which the boundary value problem is valid is [𝑎, 𝑏].
In order to simplify things further, let us divide the domain [𝑎, 𝑏] into 𝑛 sub-intervals, with the length of
each sub-interval being ℎ.
It is possible to obtain an implicit numerical approximation of the 1st derivative 𝑓 ′ (𝑥) at the mesh points
by expressing:

a1
Af i −1 + f i  + Af i +1 = ( fi +1 − fi −1 ) (11)
2h

Here, 𝐴 and 𝑎1 are arbitrarily chosen constants. These constants for 4 th-order implicit compact finite
difference estimations are what we are attempting to discover now. We obtain the following from the Taylor
series expansion:

h 2  h3  h 4 iv h5 v
fi +1 = fi + hfi  +
2!
fi +
3!
fi +
4!
fi +
5!
fi + (h )
6
(12)

h 2  h3  h 4 iv h5 v
fi −1 = fi − hfi  +
2!
fi −
3!
fi +
4!
fi −
5!
fi + (h )
6
(13)

Now, subtraction of (13) from (12) becomes:

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 58

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

h3  h5 v
fi +1 − fi −1 = 2hfi  +
3
fi +
60
fi + (h ) 7
(14)

Again, from the Taylor series expansion:

h 2  h3 iv h 4 v
f i +1 = f i  + hf i  +
2!
fi +
3!
fi +
4!
fi + (h ) 5
(15)

h 2  h3 iv h 4 v
f i −1 = f i  − hf i  +
2!
fi −
3!
fi +
4!
fi + (h ) 5
(16)

Now, adding (16) and (15) we get:

h4 v
f i +1 + f i −1 = 2 f i  + h 2 f i  +
12
fi + (h )6
(17)

Now,

a1
Af i −1 + f i  + Af i +1 − ( fi +1 − fi −1 )
2h
 h4 v  a1  h3  h5 v 
= A  2 fi  + h 2 f i  + fi + ( h ) + f
6
i

− 
 2hf i + fi + fi + ( h )
7

 12  2h  3 60 
 a   A a 
= ( 2 A + 1 − a1 ) f i  +  A − 1  h 2 f i  +  − 1  h 4 f i v + (h )
6

 6  12 120 

Setting the coefficient of the first term to zero yields a second-order scheme, whereas setting the
coefficients of the first two terms to zero yields a fourth-order scheme.
1 3
Setting the first two coefficients equal to zero we get, 𝐴 = , 𝑎1 = , then for these values (11) is a two-
4 2
parameter family of fourth-order scheme with the truncation error:

A a  1 4 ( 5)
i =  − 1  h 4 f i v+ = h f i +
 12 120  120

So, the first difference equation is

3 3 1 1
fi −1 − fi +1 + fi −1 + fi  + fi +1 = 0 i = 1, n − 1 (18)
4h 4h 4 4

and this difference equation is of convergent order 𝒪(ℎ4 ), which is equivalent to [26].
For the second difference equation, we need our given differential equation. At the interior nodes 𝑥𝑖 , (1)
becomes:

f i  = g ( x, f i , f i  ) (19)

In order to solve this problem, we must discover a 4 th-order approximation of 𝑓𝑖′′ . Now using the Taylor
series expansion, we get:

h 2  h3  h 4 iv h5 v h6 vi
f i +1 = f i + hf i  +
2
fi +
6
fi +
24
fi +
120
fi +
720
fi + (h )
7
(20)

h 2  h3  h 4 iv h5 v h 6 vi
f i −1 = f i − hf i  +
2
fi −
6
fi +
24
fi −
120
fi +
720
fi + (h )
7
(21)

Adding (20) and (21) we get:

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 59

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

h 4 iv h6 vi
f i +1 + f i −1 = 2 f i + h 2 f i  +
12
fi +
360
fi + (h )
8
(22)

Again, from the Taylor series expansion, we get:

h 2  h3 iv h 4 v h5 vi
f i +1 = f i  + hfi  +
2
fi +
6
fi +
24
fi +
120
fi + (h )
6
(23)

h 2  h3 iv h 4 v h5 vi
f i −1 = f i  − hf i  +
2
fi −
6
fi +
24
fi −
120
fi + (h )
6
(24)

Subtracting (24) from (23) we get:

h3 iv h5 vi
fi +1 − fi −1 = 2hfi  +
3
fi +
60
fi + (h )7
(25)

Firstly, multiplying (22) by 4 and (25) by ℎ and then, subtracting (25) form (22) we get:

 h 4 iv
4 ( fi +1 + fi −1 ) − h ( f i +1 − fi −1 ) = 4  2 f i + h 2 f i  + fi
 12
h6 vi   h3 iv h5 vi 
+ fi + ( h )  − h  2hf
8
i

+ fi + fi + (h )
7
(26)
360   3 60 

After simplification we get:

2h 2 f i  = 4 ( f i +1 − 2 f i + f i −1 ) − h ( f i +1 − f i −1 ) + (h ) 6
(27)

Or equivalently,
f − 2 f i + f i −1 ) − ( f i +1 − f i −1 ) + (h )
2 1
2 ( i +1
f i  = 4
(28)
h 2h

which is an approximation of the second derivative with convergent order 𝒪(ℎ4 ).

Now, substituting (28) in (19) we get:

( fi +1 − 2 fi + fi −1 ) − ( fi +1 − fi −1 ) = g ( xi , fi , fi  )
2 1
(29)
h2 2h

Or equivalently:

( fi +1 − 2 fi + fi −1 ) − ( fi +1 − fi −1 ) − g ( xi , fi , fi  ) = 0
2 1
i = 1, n − 1 (30)
h2 2h

This is our 2nd difference equation of convergence order 𝒪(ℎ4 ) for the nonlinear differential equation
(1).
We obtain (𝑛 − 1) equations for (𝑛 − 1) interior nodes from both of the difference equations (18) and
(30). This indicates that concurrently, they provide a total of (2𝑛 − 2) equations.
To satisfy the criterion for a unique solution, the number of equations and the number of unknowns must
be equal. In addition, we need two additional equations for the Dirichlet and Neumann boundary conditions
and four additional equations for the Robin boundary conditions. These extra equations are obtained from
the boundary conditions provided.
At node 𝑥 = 𝑥0 = 𝑎, we get from (19),

f 0 = g ( x0 , f 0 , f 0 ) (31)

and at node 𝑥 = 𝑥𝑛 = 𝑏, we get from (19),

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 60

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

f n = g ( xn , f n , f n ) (32)

The next step is to establish a fourth-order approximation for 𝑓0′′ and 𝑓𝑛′′ in terms of the variables 𝑓0 , 𝑓1 ,
and 𝑓2 as well as the variables 𝑓0′ , 𝑓1′ and 𝑓2′ . Hence, they can be substituted in (31) and (32) to obtain our
desired equations. In order to accomplish this goal, we will need to expand 𝑓1 , 𝑓2 , 𝑓1′ , and 𝑓2′ in Taylor's
series.
Then from Taylor series expression of these terms, we obtain:

h 2  h3  h 4 iv h5 v
f1 = f 0 + hf 0 +
2
f0 +
6
f0 +
24
f0 +
120
f0 + (h )6

8h3  16h 4 iv 32h5 v

f 2 = f 0 + 2hf 0 + 2h 2 f 0 +
6
f0 +
24
f0 +
120
f0 + (h )6

(33)
h 2  h3 iv h 4 v
1

f = f + hf + 0
2

f0 +
6
0

f0 +
24
f0 + (h )
5

8h3 iv 16h 4 v
f 2 = f 0 + 2hf 0 + 2h 2 f 0 +
6
f0 +
24
f0 + (h )
5

Solving this system for four unknowns 𝑓0′′ , 𝑓0′′′ , 𝑓0𝑖𝑣 , and 𝑓0𝑣 we get the value of 𝑓0′′ as,

−23 f 0 + 16 f1 + 7 f 2 ) − ( 6 f 0 + 8 f1 + f 2 ) + (h )
1 1
2 (
f 0 = 4
(34)
2h h

which is the 4th order estimate of the second derivative at the boundary node 𝑥 = 𝑥0 = 𝑎.
Again, from Taylor series expansion we get:

h 2  h3  h 4 iv h5 v
f n −1 = f n − hf n +
2
fn −
6
fn +
24
fn −
120
f0 + (h )6

8h3  16h 4 iv 32h5 v

f n − 2 = f n − 2hf n + 2h 2 f n −
6
fn +
24
fn −
120
fn + (h )6

(35)
h 2  h3 iv h 4 v
f 
n −1 n

= f − hf +
2
fn −n


6
fn +
24
fn + (h ) 5

8h3 iv 16h 4 v
f n− 2 = f n − 2hf n + 2h 2 f n −
6
fn +
24
fn + (h ) 5

Solving this system for four unknowns 𝑓𝑛′′ , 𝑓𝑛′′′ , 𝑓𝑛𝑖𝑣 , and 𝑓𝑛𝑣 we get the value of 𝑓𝑛′′ as.

7 f n − 2 + 16 f n −1 − 23 f n ) + ( f n− 2 + 8 f n−1 + 6 f n ) + (h )
1 1
2 (
f n = 4
(36)
2h h

which is the fourth-order estimate of the second derivative at the boundary 𝑥 = 𝑥𝑛 = 𝑏.

Substituting 𝑓0′′ in (31) we get:

( −23 f 0 + 16 f1 + 7 f 2 ) − ( 6 f 0 + 8 f1 + f 2 ) − g ( x0 , f 0 , f 0 ) = 0
1 1
(37)
2h 2 h

This is our first additional equation derived from the left boundary with convergent order 𝒪(ℎ4 ) for the
nonlinear differential equation.
Again, substituting 𝑓𝑛′′ in (32) we get:

7 f n − 2 + 16 f n −1 − 23 f n ) + ( f n− 2 + 8 f n−1 + 6 f n ) − g ( xn , f n , f n ) = 0
1 1
2 (
(38)
2h h

This is our second additional equation derived from the right boundary with convergent order 𝒪(ℎ4 ) for
the nonlinear differential equation.

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 61

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

For Dirichlet and Neumann boundary conditions, the required 2𝑛 equations come from the two difference
equations (18) and (30), and from the two additional equations (37) and (38) at the boundary of the domain.
But, for Robin boundary conditions, we required two more additional equations. So, to reduce the number
of unknowns, we have to put the boundary conditions into both the difference equations and the equations
that come from the boundary. So, we get a reduced system of 2𝑛 equations with 2𝑛 unknowns.
When we evaluate these equations simultaneously, we can find the approximate solutions to any
differential equation with different kinds of boundary conditions.

IV. RESIDUAL CORRECTION

Residual refers to the amount of error remaining after comparing an approximation to the real value of a
solution. Boundary value problems can be approximated using a variety of numerical approximation
techniques. But there are residuals for each. Residual correction is a technique for minimizing residuals by
solving the error differential equation. The residual correction methodology was first presented by Oliveira
[27] using the explicit finite difference method. Celik [28] utilizes the Chebyshev series for residual
corrections later on. For residual corrections, this section employs the implicit compact finite difference
approach.
Consider the differential equation (1) in accordance with its boundary conditions (2) again. As we intend
to apply the compact finite difference approach for residual correction, we must first discretize the domain
into subdomains. Following this, we are tasked with deriving the differential error equation as well as the
boundary conditions that correspond to it from the BVP that was provided.
If 𝑓̃ is the approximate solution and 𝑓 is the exact solution, the residual can be obtained by,

θ = 𝑓 − 𝑓̃ ⇒ 𝑓 = 𝑓̃ + θ (39)

Substituting this result into (1) we get:

 df d 
d2
dx 2
( )
f +  = g  x, f +  , +  (40)
 dx dx 

Or equivalently, this can be written as,

d 2  d  d2 f  df 
− M 1 x ,  ,  = − + M 2  x, f ,  (41)
 dx 
2 2
dx dx  dx 

which is the differential equation of the error which follows the subsequent boundary conditions:

df d
a1 f ( a ) + a1 ( a ) + b1 + b1 = 1
dx x=a
dx x=a
(42)
df d
a2 f ( b ) + a2 ( b ) + b2 + b2 = 2
dx x =b
dx x =b

But the approximate solution satisfies the given boundary conditions. So, the boundary conditions for
the error differential equations becomes:

d
a1 ( a ) + b1 =0
dx x=a
(43)
d
a2 ( b ) + b2 =0
dx x =b

We solve the BVP (41) in accordance with the boundary conditions (43) by 4 th order compact finite
difference method. Then, the updated approximate value becomes,

(Updated Approximation) = (Weighted Residual Value) + ( Error Value) (44)

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 62

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

V. CONVERGENCE ANALYSIS
In this section, we discuss the convergence of our proposed scheme for solving the linear and nonlinear
BVPs (1)-(2).
Let Ω = 𝐶 𝑙 ([𝑥0 , 𝑥𝑛 ] ∈ 𝑅) be the linear space of real-valued functions that are 𝑙 times differentiable on
𝐷 = [𝑥0 , 𝑥𝑛 ]. Suppose that,

  1 ,  2 =  w0 ( x )  1 ( x )  2 ( x ) dx (45)
D

be the 𝐿2 inner product on Ω for some sufficiently smooth weight function 𝑤0 that induces the 𝐿2 norm,

 1 =  w0 ( x )  12 ( x ) dx
2
(46)
D

for which Ω is an infinite dimensional Hilbert space. Assume 𝑉 = {𝐵_𝑖|𝑖 = 1,2,3, … … . } be a Schauder
basis of Ω, where 𝐵𝑖 's are Bernstein polynomials on 𝐷 = [𝑥0 , 𝑥𝑛 ]. Let us begin with an approximation
subspace Ω𝑁 spanned by {ψ1 , ψ2 , ψ3 … … … … . . , ψ𝑁 } that satisfies appropriate boundary conditions. The
Galerkin weighted residual equation is given by < 𝑅 (𝑥, 𝑓̃(𝑥)) , ψ𝑗 >= 0 where 𝑓̃(𝑥) = 𝑁0 (𝑥) +
∑𝑛𝑗=1 α𝑗 𝑁𝑗 (𝑥) is a trial solution of the differential equation. In particular, the residual function 𝑅 (𝑥, 𝑓̃(𝑥))
is orthogonal to each function of the basis of the approximate subspace Ω𝑁 . It is well known that Bernstein
polynomials can approximate any continuous function with arbitrary precision. If the dimension of the
subspace Ω𝑁 is infinitely large, the residual function 𝑅 (𝑥, 𝑓̃(𝑥)) is orthogonal to each Bernstein polynomial
which immediately implies that the residual 𝑅 (𝑥, 𝑓̃(𝑥)) is orthogonal to any continuous function in Ω. A
function that is orthogonal to any other functions in the space is necessarily the zero function.
For the error differential equation in accordance with the error boundary conditions, we use 4 th-order
compact finite difference method.
Definition 1: [29] Let us consider Ψℎ as the solution of the discretized equation ℒ𝒽 Ψℎ = 𝑏ℎ which
converges to the solution ψ of the given differential equation ℒψ = 𝑏 if ||ψℎ − Ψℎ || ⟶ 0 as ℎ ⟶ 0.
Moreover, if for a positive constant 𝑘 and another positive constant 𝑀0 > 0 which is independent of 𝑘
in the sense that,  h −  h  M 0 h k in this case, it is stated that the discretized equation has 𝑘-th order
precision with convergence order ℎ𝑘 .
Definition 2: [29] The discretized Equation is stable if there exists ℎ0 > 0 and δ > 0 in such a way that
for any ℎ < ℎ0 and any ϵℎ ∈ ℬ𝒽 satisfying ||ϵ|| <, the perturbed difference equation ℒ𝒽 𝑤ℎ = 𝑏ℎ + δϵℎ
has unique solution 𝑤ℎ , satisfying ||𝑤ℎ − Ψℎ || ≤ 𝑀||ϵℎ || where Ψℎ is the solution of the unperturbed
difference equation and 𝑀 > 0 is independent of ℎ.
Theorem 1: [29] If the discretized equation is stable and also consistent (with order ℎ𝑘 ) with the given
differential equation, then, the solution of the discretized equation 𝛹ℎ converges to the solution 𝜓 and
which satisfies ||𝜓ℎ − 𝛹ℎ || ≤ 𝑀𝑀1 ℎ𝑘 , where 𝑀 and 𝑀1 are certain constants. Alternatively, the order in
which the difference scheme approaches the continuous problem corresponds to the accuracy order of the
difference scheme.
Proof: [29] Since the difference scheme is consistent, it is established that ||δ𝑏ℎ || ≤ 𝑀1 ℎ𝑘 ⟶ 0 as ℎ ⟶
0. This means that a grid 𝒟𝒽 can be constructed in such way that ℎ < ℎ0 and δ𝑏ℎ < as in Definition 2, and

h h = bh +  bh (47)

Therefore, Ψℎ satisfies the conditions for stability. Due to this,  h −  h  M  bh  M ( M 1h k ) (48)

Hence, the discrete solution Ψℎ converges to the continuous solution ψ of order ℎ𝑘 .

VI. RESULTS AND DISCUSSIONS

In this section, we will apply our proposed scheme to some second-order linear and nonlinear problems
and calculate the accuracy and stability of our proposed scheme. We compute the 𝐿∞ norm by comparing
the approximate solutions with the exact solutions as follows:

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 63

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

L = max1 x  n f i − f i (49)

We also compute the Convergence Rate (𝒞ℛ) as follows [24],

 
log  1 
=  2 (50)
 h1 
log  
 h2 

where, ϵ1 and ϵ2 are the 𝐿∞ for grid size ℎ1 and ℎ2 , respectively.

Problem 1
Let us consider the linear DE in conjunction with Neumann boundary conditions [6], [30],

d2 f 1 − cos (1) 1 − cos (1)

+ f = −1 with f  (0) = , f  (1) = − (51)
dx 2 sin (1) sin (1)

with the exact solution,

1 − cos (1)
f ( x ) = cos ( x ) + sin ( x ) − 1 (52)
sin (1)

In the first stage, the BVP is solved numerically using the modified Galerkin approach with the aid of
the trial solution as (3). For residual correction, we need the error differential equation with error BCs. The
error BVP for the given differential equation is,

d 2 d2 f
2
+  = − 2 − f −1 with   ( 0 ) = 0,  (1) = 0 (53)
dx dx

The following table shows 𝐿∞ and 𝒞ℛ generated using the proposed method for different grid sizes ℎ
for residual corrections.

TABLE I: 𝐿∞ AND CONVERGENCE RATE (𝒞ℛ) FOR PROBLEM 1

h 𝐿∞ 𝒞ℛ Reference Results (𝐿∞)
0.1 2.0797 × 10−08 [6], 2011:
0.05 1.5311 × 10−09 3.7637 Bernoulli Pol. (5th degree)
Bernstein 0.025 1.3140 × 10−10 3.5425 1.173569 × 10−06
Polynomials 0.0125 9.3694 × 10−12 3.8099 Bernoulli Pol. (7th degree)
of degree 4 0.01 3.9338 × 10−12 3.8892 1.2808 × 10−09
0.005 2.5785 × 10−13 3.9313 Bernoulli Pol. (10th degree)
0.0025 1.6570 × 10−14 3.9599 1.0131 × 10−14

From Table I, we can observe that we obtain high accuracy by using Bernstein polynomials of degree 4
only in our prescribed scheme. Whereas, in [6], Islam & Shirin implemented the modified Galerkin
approach and the 𝐿∞ obtained using Bernoulli polynomials of degree 10 is 10−14 . So, we can say that, in
our proposed scheme, we can attain better accuracy by using fewer polynomials and then correcting the
residuals of the approximation.
Problem 2
Let us consider the following nonlinear DE in conjunction with Dirichlet boundary conditions [6],

d 2 f 1 df  1  43
+ f =  4 + x3  with f (1) = 17, f ( 3) = (54)
dx 2 8 dx  4  3

with the exact solution,

16
f ( x ) = x2 + (55)
x

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 64

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

Equivalent BVP on [0,1] becomes:

d 2 f 1 df 43
+ f = 16 + ( 2 x + 1) f ( 0 ) = 17, f (1) =
3
2
with (56)
dx 4 dx 3

In the first stage, the BVP is solved numerically using the modified Galerkin approach with the aid of
the trial solution as (3). For residual correction, we need the error differential equation with error BCs.
The error BVP for the given differential equation is:

d 2 1 d 1 df 1 d d 2 f 1 df
+ f +  +  = 16 + ( 2 x + 1) − 2 − f  ( 0 ) =  (1) = 0 (57)
3
2
with
dx 4 dx 4 dx 4 dx dx 4 dx

The following table shows 𝐿∞ and 𝒞ℛ generated using the proposed method for different grid sizes ℎ
for residual corrections.

TABLE II: 𝐿∞ AND CONVERGENCE RATE (𝒞ℛ) FOR PROBLEM 2

h 𝐿∞ 𝒞ℛ Reference Results (𝐿∞)

0.1 5.1458 × 10−04 [6], 2011:
0.05 5.1111 × 10−05 3.3315 Bernoulli Pol. (10th degree)
Bernstein 0.025 4.2499 × 10−06 3.5883 1.4449 × 10−05
Polynomials 0.0125 3.1017 × 10−07 3.7763 Bernoulli Pol. (15th degree)
of degree 4 0.01 7.6237 × 10−08 3.8548 6.3806 × 10−08
0.005 8.7611 × 10−09 3.9048
0.0025 5.6623 × 10−10 3.9516
From Table II, we can observe that we obtain high accuracy by using Bernstein polynomials of degree 4
only in our prescribed scheme. Whereas, in [6], Islam & Shirin implemented the modified Galerkin
approach and the 𝐿∞ obtained using Bernoulli polynomials of degree 15 is 10−08 . So, we can say that, in
our proposed scheme, we can attain better accuracy by using fewer polynomials and then correcting the
residuals of the approximation.
Problem 3
Let us consider the following nonlinear DE in conjunction with Robin boundary conditions [6],[31],

d2 f 1
= (1 + x + f ) f  ( 0 ) − f ( 0 ) = −1/ 2, f (1) + f (1) = 1
3
with (58)
dx 2 2

with the exact solution,

2
f ( x) = − x −1 (59)
2− x

In the first stage, the BVP is solved numerically using the modified Galerkin approach with the aid of
the trial solution as (3). For residual correction we need the error differential equation with error BCs. The
error BVP for the given differential equation is,

d 2  3 3  3 
−  (1 + x ) + 3 (1 + x ) f + f 2   −  (1 + x ) +
3 1 d2 f 1
( )
3
f  2 −  3 = − 2 + 1 + x + f
2
(60)
dx 2  2 2   2 2  2 dx 2

with the boundary conditions, θ′ (0) − θ(0) = −1/2, θ′ (1) + θ(1) = 1.

The following table shows 𝐿∞ and 𝒞ℛ generated using the proposed method for different grid sizes ℎ
for residual corrections.

TABLE III: 𝐿∞ AND CONVERGENCE RATE (𝒞ℛ) FOR PROBLEM 3

h 𝐿∞ 𝒞ℛ Reference Results (𝐿∞)

0.1 7.1799 × 10−06 [6], 2011: [31], 2022
0.05 2.8731 × 10−07 4.6433 Bernoulli Pol. (8th degree) Legendre Pol. (8th degree)
Bernstein 0.025 1.1124 × 10−08 4.6909 5.52 × 10−07 2.65 × 10−09
Polynomials of 0.0125 8.9334 × 10−10 3.6383 Bernoulli Pol. (10th degree) Legendre Pol. (10th degree)
degree 4 0.01 3.8393 × 10−10 3.7846 6.99 × 10−09 7.77 × 10−11
0.005 2.6363 × 10−11 3.8643 Legendre Pol. (11th degree)
0.0025 1.7256 × 10−12 3.9333 8.15 × 10−12

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 65

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

From Table III, we can observe that we obtain high accuracy by using Bernstein polynomials of degree
4 in our proposed scheme. Whereas, in [6], Islam & Shirin implemented the modified Galerkin approach
using Bernoulli polynomials of degree 10 and the obtained 𝐿∞ is 10−9 . In [31], Sohel et al. implemented
the residual correction approach and the 𝐿∞ using Legendre polynomials of degree 11 is 10−12 . So, we can
say that, we can attain better accuracy by using fewer polynomials and then correcting the residuals of the
approximation using our present approach.
Problem 4
Let us consider the following strongly nonlinear Bratu's problem [12],[15],[32]-[35]:

d2 f
+  e f ( x ) = 0, with f ( 0 ) = 0, f (1) = 0 (61)
dx 2

The term strongly non-linear is employed because the non-linearity is generated by exponential term.
This BVP has an exact solution,

  1 
 cosh  x − 2  2 
f ( x ) = −2 ln    
(62)
 cosh    
   
 4 

β
where, β satisfies, √2λ cosh ( ) = β. There are zero, one, and two solutions to the Bratu's problem when
4
λ > λ𝑐 , λ = λ𝑐 and λ < λ𝑐 , respectively, where λ𝑐 = 3.513830719.
In the first stage, the BVP is solved numerically using the modified Galerkin approach with the aid of
the trial solution as (3). For residual correction, we need the error differential equation with error BCs. The
error BVP for the given differential equation is:

d 2 d2 f
2
+  e f e = − 2 with  ( 0 ) = 0, (1) = 0 (63)
dx dx

The following table shows 𝐿∞ and 𝒞ℛ generated using the proposed method for different grid sizes ℎ
for residual corrections.
Firstly, representing the results for λ = 1.

TABLE IV: 𝐿∞ AND CONVERGENCE RATE (𝒞ℛ) FOR PROBLEM 4 FOR Λ = 1

h 𝐿∞ 𝒞ℛ Reference Results (𝐿∞)
−08
0.1 8.2287 × 10 [12], 2021: 6.187 × 10−05
Bernstein 0.05 7.6868 × 10−09 3.4202 [15], 2004: 1.348 × 10−05
Polynomials 0.025 6.2340 × 10−10 3.6242 [32], 2011: 9.048 × 10−07
of 0.01 1.8176 × 10−11 3.8580 [34], 2019: 9.90 × 10−08
degree 4 0.005 1.1832 × 10−12 3.9413 [35], 2019: 6.41 × 10−13
0.0025 7.6439 × 10−14 3.9522

From Table IV, we can observe that we obtain the accuracy of 10−14 by using Bernstein polynomials of
degree 4 in our present scheme for λ = 1. Whereas, for λ = 1, in [12], Mustafa et al. implemented the
Subdivision collocation method and in [15], Khuri implemented Laplace transformation method, and the
obtained 𝐿∞ . In both cases is 10−05 , in [32], Abbasbandy et al. implemented the Lie Group Shooting
method and the obtained 𝐿∞ is of 10−07 , in [34] Ala O et al. implemented the Quantic B-spline approach,
and the obtained 𝐿∞ is of 10−08 and in [35] Roul & Thula implemented the B-spline Collocation method
and the obtained 𝐿∞ is of 10−13 .
Then, representing the results for λ = 2,

TABLE V: 𝐿∞ AND CONVERGENCE RATE (𝒞ℛ) FOR PROBLEM 4 FOR 𝜆 = 2

h 𝐿∞ 𝒞ℛ Reference Results (𝐿∞)
0.1 9.5936 × 10−07 [33], 2000: 1.52× 10−02
Bernstein 0.05 1.1273 × 10−07 3.0892 [12], 2021: 4.939× 10−04
Polynomials 0.025 8.6011 × 10−09 3.7122 [32], 2011: 5.709× 10−06
of 0.01 2.4317 × 10−10 3.8616 [34], 2019: 1.60× 10−06
degree 4 0.005 1.5668 × 10−11 3.9561 [36], 2015: 1.98 × 10−09
0.0025 9.9537 × 10−13 3.9764

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 66

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

From Table V we can observe that we obtain the accuracy of 10−13 by using Bernstein polynomials of
degree 4 in our present scheme for λ = 2. Whereas, for λ = 2, in [12], Mustafa et al. implemented the
subdivision collocation method and the obtained 𝐿∞ is of 10−04 , in [32], Abbasbandy et al. implemented
the Lie Group Shooting method and the obtained 𝐿∞ is of 10−06 , in [33], Deeba et al. implemented the
Decomposition algorithm, and the obtained 𝐿∞ is of 10−02 , in [34], Ala O et al. implemented the Quantic
B-spline method and the obtained 𝐿∞ is of 10−06 and in [36], Farzana & Islam implemented the Chebyshev-
Legendre collocation method, and the obtained 𝐿∞ is of 10−09 . So, we can say that, in our proposed scheme,
we can attain better accuracy by using fewer polynomials and then correcting the residuals.

VII. CONCLUSION
In this research, we have obtained numerical solutions for both linear and nonlinear BVPs, followed by
residual corrections. Using the weighted residual technique, a numerical solution has been generated in the
first part. Then, using the implicit compact finite difference technique, updated approximations were
established. We have developed the formulation of the compact finite difference scheme for the nonlinear
differential equations both at the interior node and the boundary with Dirichlet, Neumann, and Robin
boundary conditions and established the convergence and stability of our numerical solutions. Our research
has shown the effectiveness of the residual correction technique in improving the accuracy of the
approximations. We have compared our results with those published in the literature and demonstrated the
superiority of our approximations in terms of accuracy. Overall, this research offers valuable insights into
the numerical solution of boundary value problems and provides a practical method for achieving high-
precision results for various applications. The proposed method can be applied to higher-order boundary
value problems as well as to partial differential equations.

ACKNOWLEDGMENT
The authors are thankful to the reviewers for their valuable comments. The first author is thankful to the
Ministry of Science and Technology, Government of the People's Republic of Bangladesh, for providing
partial financial support under the "National Science and Technology Fellowship" program during the
research time.

CONFLICT OF INTEREST
Authors declare that they do not have any conflict of interest.

REFERENCES
[1] Bender CM, Orszag S, Orszag SA. Advanced mathematical methods for scientists and engineers I: Asymptotic methods and
perturbation theory. Springer Science & Business Media; 1999.
[2] Collatz L. The numerical treatment of differential equations. Springer Science & Business Media; 2012.
[3] Kanth AR, Reddy YN. Cubic spline for a class of singular two-point boundary value problems. Applied mathematics and
computation. 2005;170(2):733-40.
[4] Kanth AR, Bhattacharya V. Cubic spline for a class of non-linear singular boundary value problems arising in physiology.
Applied Mathematics and Computation. 2006;174(1):768-74.
[5] Motsa SS, Sibanda P, Shateyi S. A new spectral-homotopy analysis method for solving a nonlinear second order BVP.
Communications in Nonlinear Science and Numerical Simulation. 2010;15(9):2293-302.
[6] Islam MD, Shirin A. Numerical solutions of a class of second order boundary value problems on using Bernoulli polynomials.
arXiv preprint arXiv:1309.6064. 2013.
[7] Burden RL, Faires JD, Burden AM. Numerical analysis. Cengage learning; 2015.
[8] Pandey PK. Solution of two-point boundary value problems, a numerical approach: parametric difference method. Applied
Mathematics and Nonlinear Sciences. 2018;3(2):649-58.
[9] Ramos H, Rufai MA. A third-derivative two-step block Falkner-type method for solving general second-order boundary-value
systems. Mathematics and Computers in Simulation. 2019;165:139-55.
[10] Baccouch M. A finite difference method for stochastic nonlinear second-order boundary-value problems driven by additive
noises. International Journal of Numerical Analysis & Modeling. 2020;17(3).
[11] Abbasbandy S, Hajiketabi M. A simple, efficient and accurate new Lie--group shooting method for solving nonlinear boundary
value problems. International Journal of Nonlinear Analysis and Applications. 2021;12(1):761-81.
[12] Mustafa G, Ejaz ST, Kouser S, Ali S, Aslam M. Subdivision collocation method for one-dimensional Bratu’s problem. Journal
of Mathematics. 2021;2021:1-8.
[13] Boyd JP. Chebyshev polynomial expansions for simultaneous approximation of two branches of a function with application to
the one-dimensional Bratu equation. Applied Mathematics and Computation. 2003;143(2-3):189-200.
[14] Caglar H, Caglar N, Özer M, Valarıstos A, Anagnostopoulos AN. B-spline method for solving Bratu's problem. International
Journal of Computer Mathematics. 2010;87(8):1885-91.
[15] Khuri SA. A new approach to Bratu’s problem. Applied mathematics and computation. 2004;147(1):131-6.
[16] Ben-Romdhane M, Temimi H, Baccouch M. An iterative finite difference method for approximating the two-branched solution
of Bratu's problem. Applied Numerical Mathematics. 2019;139:62-76.

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 67

 EJ-MATH, European Journal of Mathematics and Statistics
ISSN: 2736-5484

[17] Qin Y. Simulation based on Galerkin method for solidification of water through energy storage enclosure. Journal of Energy
Storage. 2022;50:104672.
[18] Li X, Li S. A fast element-free Galerkin method for the fractional diffusion-wave equation. Applied Mathematics Letters.
2021;122:107529.
[19] Abdelhakem M, Alaa-Eldeen T, Baleanu D, Alshehri MG, El-Kady M. Approximating real-life BVPs via Chebyshev
polynomials’ first derivative pseudo-Galerkin method. Fractal and Fractional. 2021;5(4):165.
[20] Ruman U, Islam MS. Galerkin Weighted Residual Method for Solving Fourth Order Fractional Differential and Integral
Boundary Value Problems. Journal of Applied Mathematics and Computation. 2022;6(2):410-422.
[21] Ali H, Kamrujjaman M, Islam MS. Numerical computation of Fitzhugh-Nagumo equation: a novel Galerkin finite element
approach. International Journal of Mathematical Research. 2020;9(1):20-7.
[22] Lele SK. Compact finite difference schemes with spectral-like resolution. Journal of Computational Physics. 1992;103(1):16-
42.
[23] Mehra M, Patel KS. Algorithm 986: a suite of compact finite difference schemes. ACM Transactions on Mathematical Software
(TOMS). 2017;44(2):1-31.
[24] Malele J, Dlamini P, Simelane S. Highly Accurate Compact Finite Difference Schemes for Two-Point Boundary Value Problems
with Robin Boundary Conditions. Symmetry. 2022;14(8):1720.
[25] Lewis PE, Ward JP. The finite element method: principles and applications. Wokingham: Addison-Wesley; 1991.
[26] Pettigrew MF, Rasmussen H. A compact method for second-order boundary value problems on nonuniform grids. Computers
& Mathematics with Applications. 1996;31(9):1-6.
[27] Oliveira FA. Collocation and residual correction. Numerische Mathematik. 1980;36:27-31.
[28] Çelik İ. Collocation method and residual correction using Chebyshev series. Applied Mathematics and Computation.
2006;174(2):910-20.
[29] Tyler JG. Analysis and implementation of high-order compact finite difference schemes. Brigham Young University; 2007.
[30] MÜLLENHEIM G. Solving two-point boundary value problems with spline functions. IMA Journal of Numerical Analysis.
1992;12(4):503-18.
[31] Sohel MN, Islam MS, Islam MS, Kamrujjaman M. Galerkin Method and Its Residual Correction with Modified Legendre
Polynomials. Contemporary Mathematics. 2022:188-202.
[32] Abbasbandy S, Hashemi MS, Liu CS. The Lie-group shooting method for solving the Bratu equation. Communications in
Nonlinear Science and Numerical Simulation. 2011;16(11):4238-49.
[33] Deeba E, Khuri SA, Xie S. An Algorithm for Solving Boundary Value Problems. Journal of Computational Physics.
2001;1(170):448.
[34] Ala’yed O, Batiha B, Abdelrahim R, Jawarneh AA. On the numerical solution of the nonlinear Bratu type equation via quintic
B-spline method. Journal of Interdisciplinary Mathematics. 2019;22(4):405-13.
[35] Roul P, Thula K. A fourth-order B-spline collocation method and its error analysis for Bratu-type and Lane–Emden problems.
International Journal of Computer Mathematics. 2019;96(1):85-104.
[36] Farzana H, Islam MS. Computation of some second order Sturm-Liouville BVPs using Chebyshev-Legendre collocation method.
GANIT: Journal of Bangladesh Mathematical Society. 2015;35:95-112.

DOI: http://dx.doi.org/10.24018/ejmath.2023.4.4.235 Vol 4 | Issue 4 | August 2023 68