Numerical Algorithms

https://doi.org/10.1007/s11075-020-01016-2
ORIGINAL PAPER

Simple numerical methods of second-

and third-order convergence for solving a fully
third-order nonlinear boundary value problem

Quang A Dang1,3 · Quang Long Dang2

Received: 14 July 2020 / Accepted: 14 September 2020 /

© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
In this paper, we consider a fully third-order nonlinear boundary value problem that
is of great interest of many researchers. First, we establish the existence and unique-
ness of solution. Next, we propose simple iterative methods on both continuous and
discrete levels. We prove that the discrete methods are of second-order and third-
order of accuracy due to the use of appropriate formulas for numerical integration
and obtain estimate for total error. Some examples demonstrate the validity of the
obtained theoretical results and the efficiency of the iterative methods.

Keywords Third-order nonlinear equation · Existence and uniqueness of solution ·

Iterative method · Third-order accuracy · Total error

Mathematics Subject Classification (2010) MSC 34B15 · MSC 65L10

1 Introduction

Boundary value problems (BVPs) for third-order nonlinear differential equations

appear in many applied fields, such as flexibility mechanics, chemical engineering,
and heat conduction. Many works are devoted to the qualitative aspects of the prob-
lems (see, e.g., [6, 7, 22, 23, 33, 35, 37]). There are also many methods concerning the
solution of third-order BVPs including analytical methods [1, 28, 32] and numerical
methods by using interpolation polynomials [3] Taylor series expansion [5], quartic

* Quang A Dang
[email protected]

Quang Long Dang

[email protected]

1 Centre for Informatics and Computing, VAST, 18 Hoang Quoc Viet, Cau Giay, Hanoi, Vietnam
2 Institute of Information Technology, VAST, 18 Hoang Quoc Viet, Cau Giay, Hanoi, Vietnam
3 Lac Hong University, 10 Huynh Van Nghe, Bien Hoa, Dong Nai, Vietnam
 Numerical Algorithms

splines [21, 31], quintic splines [26], non-polynomial splines [24, 25, 27, 34], and
wavelets [19]. A majority of the mentioned above numerical methods are devoted to
linear equations or special nonlinear third-order differential equations.
In this paper, we consider the following BVP:

u(3) (t) = f (t, u(t), u′ (t), u′′ (t)), 0 < t < 1,

(1)
u(0) = c1 , u′ (0) = c2 , u′ (1) = c3 .

Some authors have studied the existence and positivity of solution for this problem,
for example, by using the lower and upper solutions method and fixed point theorem
on cones; in [36], Yao and Feng established the existence of solution and positive
solution for the case f = f (t, u(t)). In [20], Feng and Liu obtained existence results
by the use of the lower and upper solutions method and a new maximum principle for
the case f = f (t, u(t), u′ (t)). It should be emphasized that the results of these two
works are pure existence but are not methods for finding solutions. Many researchers
are interested in numerical solution of the problem (1) without giving attention to
qualitative aspects of it, or refer to the book [2].
Below, we mention some works devoted to solution methods for the problem (1).
Namely, Al Said et al. [4] have solved a third-order two-point BVP using cubic
splines. Noor et al. [29] generated second-order method based on quartic splines.
Other authors [8, 26] generated finite difference schemes using fourth-degree B-
spline and quintic polynomial spline for this problem subject to other boundary
conditions. El-Danaf [10] constructed a new spline method based on quartic nonpoly-
nomial spline functions that has a polynomial part and a trigonometric part to develop
numerical methods for a linear differential equation with the boundary conditions as
in (1). Recently, in 2016, Pandey [30] solved the problem for the case f = f (t, u)
by the use of quartic polynomial splines. The convergence of the method of at least
O(h2 ) for the linear case f = f (t) was proved. In the next year, this author in [31]
proposed two difference schemes for the general case f = f (t, u(t), u′ (t), u′′ (t))
and also established the second-order accuracy for the linear case. In 2019, Chaura-
sia et al. [9] use exponential amalgamation of cubic spline functions to form a novel
numerical method of second-order accuracy. It should be emphasized that all above-
mentioned authors only draw attention to the construction of the discrete analogue
of the problem (1) and estimate the error of the obtained solution assuming that the
nonlinear system of algebraic equations can be solved by known iterative methods.
Thus, they did not take into account the errors arising in the last iterative methods.
Motivated by these facts, in this paper, we propose a completely different method,
specifically, an iterative method on both continuous and discrete levels for the prob-
lem (1). We give an analysis of the total error of the solution obtained. This error
includes the error of the iterative method on continuous level and the error arising in
the numerical realization of this iterative method. The obtained total error estimate
suggests how to choose a suitable grid size for discretization to get an approxi-
mate solution with a given accuracy. In order to justify the total error estimate, first
we establish some results on existence and uniqueness of solution. These results
are obtained by the method developed in [11–18]. Some examples demonstrate the
validity of the obtained theoretical results and the efficiency of the iterative method.
Numerical Algorithms

2 Existence results

For simplicity of presentation, we consider the problem (1) with homogeneous

boundary conditions, i.e., the problem:
u(3) (t) = f (t, u(t), u′ (t), u′′ (t)), 0 < t < 1,
(2)
u(0) = 0, u′ (0) = 0, u′ (1) = 0.
To investigate this problem, we associate it with an operator equation as follows.
For a function ϕ(t) ∈ C[0, 1], consider the nonlinear operator A defined by:
(Aϕ)(t) = f (t, u(t), u′ (t), u′′ (t)), (3)
where u(t) is the solution of the problem
u′′′ (t) = ϕ(t), 0 < t < 1
(4)
u(0) = 0, u′ (0) = 0, u′ (1) = 0.

Proposition 1 If the function ϕ(t) is a fixed point of the operator A, i.e., ϕ(t) is a
solution of the operator equation ϕ = Aϕ , then the function u(t) determined from
the BVP (4) solves the problem (2). Conversely, if u(t) is a solution of the BVP (2)
then the function ϕ(t) = f (t, u(t), u′ (t), u′′ (t)) is a fixed point of the operator A
defined above by (3) and (4).

Thus, the problem (2) is reduced to the fixed point problem for A.
Now, we study the properties of A. For this purpose, notice that the problem (2)
has a unique solution representable in the form:
Z 1
u(t) = G0 (t, s)ϕ(s)ds, 0 < t < 1, (5)
0

where G0 (t, s) is the Green function of the problem (4):

(
s 2
(t − 2t + s), 0 ≤ s ≤ t ≤ 1,
G0 (t, s) = 2t 2
2 (s − 1), 0 ≤ t ≤ s ≤ 1.

Differentiating both sides of (5) gives:

Z 1
u′ (t) = G1 (t, s)ϕ(s)ds, (6)
0
Z 1
′′
u (t) = G2 (t, s)ϕ(s)ds, (7)
0

where

s(t − 1), 0 ≤ s ≤ t ≤ 1,
G1 (t, s) =
t (s − 1), 0 ≤ t ≤ s ≤ 1,

s, 0 ≤ s ≤ t ≤ 1,
G2 (t, s) = (8)
s − 1, 0 ≤ t ≤ s ≤ 1.
 Numerical Algorithms

It is easily seen that G0 (t, s) ≤ 0, G1 (t, s) ≤ 0 in Q = [0, 1]2 and

Z 1 Z 1
1 1
M0 = max |G(t, s)| ds = , M1 = max |G1 (t, s)| ds = ,
0≤t ≤1 0 12 0≤t ≤1 0 8
Z 1 (9)
1
M2 = max |G2 (t, s)| ds = .
0≤t ≤1 0 2
Next, for each fixed real number M > 0, introduce the domain:
DM = {(t, x, y, z)| 0 ≤ t ≤ 1, |x| ≤ M0 M, |y| ≤ M1 M, |z| ≤ M2 M},
and as usual, by B[O, M] we denote the closed ball of radius M centered at 0 in the
space of continuous in [0, 1] functions, namely, B[O, M] = {ϕ ∈ C[0, 1]| kϕk ≤
M}, where kϕk = max0≤t ≤1 |ϕ(t)|.
By the analogous techniques as in [11]–[18], we have proved the following results.

Theorem 1 (Existence of solutions) Suppose that there exists a number M > 0 such
that the function f (t, x, y, z) is continuous and bounded by M in the domain DM ,
i.e.:
|f (t, x, y, z)| ≤ M
for any (t, x, y, z) ∈ DM .
Then, the problem (1) has a solution u(t) satisfying:
|u(t)| ≤ M0 M, |u′ (t)| ≤ M1 M, |u′′ (t)| ≤ M2 M for any 0 ≤ t ≤ 1.

Theorem 2 (Existence and uniqueness of solution) Assume that there exist numbers
M, L0 , L1 , L2 ≥ 0 such that:
|f (t, x, y, z)| ≤ M,

|f (t, x2 , y2 , z2 ) − f (t, x1, y1 , z1 )| ≤ L0 |x2 − x1 | + L1 |y2 − y1 | + L2 |z2 − z1 |

(10)
for any (t, x, y, z), (t, xi , yi , zi ) ∈ DM (i = 1, 2) and
q := L0 M0 + L1 M1 + L2 M2 < 1.
Then, the problem (2) has a unique solution u(t) such that |u(t)| ≤ M0 M, |u′ (t)| ≤
M1 M, |u′′ (t)| ≤ M2 M for any 0 ≤ t ≤ 1.

Remark The problem (1) for u(t) with non-homogeneous boundary conditions can
be reduced to the problem with homogeneous for function v(t) if setting u(t) =
v(t) + P2 (t), where P2 (t) is the second-degree polynomial satisfying the boundary
conditions P2 (0) = c1 , P2′ (0) = c2 , P2 (1) = c3 .

3 Iterative method on continuous level

Consider the following iterative method for solving the problem (2):
Numerical Algorithms

1. Given

ϕ0 (t) = f (t, 0, 0, 0). (11)

2. Knowing ϕk (t) (k = 0, 1, ...) compute

Z 1
uk (t) = G0 (t, s)ϕk (s)ds,
0
Z 1
yk (t) = G1 (t, s)ϕk (s)ds,
0
Z 1
zk (t) = G2 (t, s)ϕk (s)ds, (12)
0

3. Update

ϕk+1 (t) = f (t, uk (t), yk (t), zk (t)). (13)

Set

qk
pk = , d = kϕ1 − ϕ0 k. (14)
1−q

Theorem 3 (Convergence) Under the assumptions of Theorem 2, the above iterative

method converges and there holds the estimates:

kuk − uk ≤ M0 pk d, ku′k − u′ k ≤ M1 pk d, ku′′k − u′′ k ≤ M2 pk d,

where u is the exact solution of the problem (2) and M0 , M1 , M2 are given by (9).

This theorem follows straightforwardly from the convergence of the succes-

sive approximation method for finding fixed point of the operator A and the
representations (5)–(7) and (12).

4 Discrete iterative method 1

To numerically realize the above iterative method, we construct the correspond-

ing discrete iterative methods. For this purpose, cover the interval [0, 1] by the
uniform grid ω̄h = {ti = ih, h = 1/N, i = 0, 1, ..., N } and denote by
Φk (t), Uk (t), Yk (t), Zk (t) the grid functions, which are defined on the grid ω̄h and
approximate the functions ϕk (t), uk (t), yk (t), zk (t) on this grid, respectively.
First, consider the following discrete iterative method, named method 1:
1. Given

Φ0 (ti ) = f (ti , 0, 0, 0), i = 0, ..., N . (15)

 Numerical Algorithms

2. Knowing Φk (ti ), k = 0, 1, ...; i = 0, ..., N, compute approximately the definite

integrals (12) by the trapezoidal rule:
N
X
Uk (ti ) = hρj G0 (ti , tj )Φk (tj ),
j =0
N
X
Yk (ti ) = hρj G1 (ti , tj )Φk (tj ),
j =0
N
X
Zk (ti ) = hρj G∗2 (ti , tj )Φk (tj ), i = 0, ..., N, (16)
j =0

where ρj are the weights

(
1/2, j = 0, N
ρj =
1, j = 1, 2, ..., N − 1
and 
s,
 0 ≤ s < t ≤ 1,
∗
G2 (t, s) = s − 1/2, s = t, (17)

s − 1, 0 ≤ t < s ≤ 1.

3. Update
Φk+1 (ti ) = f (ti , Uk (ti ), Yk (ti ), Zk (ti )). (18)
In order to get the error estimates for the numerical approximate solution for u(t) and
its derivatives on the grid, we need some following auxiliary results.

Proposition 2 Assume that the function f (t, x, y, z) has all continuous par-
tial derivatives up to second order in the domain DM . Then, for the functions
uk (t), yk (t), zk (t), k = 0, 1, .., constructed by the iterative method (11)-(13), we
have zk (t) ∈ C 3 [0, 1], yk (t) ∈ C 4 [0, 1], uk (t) ∈ C 5 [0, 1].

Proof We prove the proposition by induction. For k = 0, by the assumption on the

function f , we have ϕ0 (t) ∈ C 2 [0, 1] since ϕ0 (t) = f (t, 0, 0, 0). Taking into account
the expression (8) of the function G2 (t, s), we have:
Z 1 Z t Z 1
z0 (t) = G2 (t, s)ϕ0 (s)ds = sϕ0 (s)ds − (s − 1)ϕ0 (s)ds.
0 0 t
It is easy to see that z0′ (t) = ϕ0 (t). Therefore, z0 (t) ∈ C 3 [0, 1]. This implies y0 (t) ∈
C 4 [0, 1], u0 (t) ∈ C 5 [0, 1].
Now suppose zk (t) ∈ C 3 [0, 1], yk (t) ∈ C 4 [0, 1], uk (t) ∈ C 5 [0, 1]. Then,
because ϕk+1 (t) = f (t, uk (t), yk (t), zk (t)) and the function f by the assumption
has continuous derivative in all variables up to order 2, it follows that ϕk+1 (t) ∈
C 2 [0, 1]. Repeating the same argument as for ϕ0 (t) above, we obtain that zk+1 (t) ∈
C 3 [0, 1], yk+1 (t) ∈ C 4 [0, 1], uk+1 (t) ∈ C 5 [0, 1]. Thus, the proposition is
proved.
Numerical Algorithms

Proposition 3 For any function ϕ(t) ∈ C 2 [0, 1] we have:

Z 1 N
X
Gn (ti , s)ϕ(s)ds = hρj Gn (ti , tj )ϕ(tj ) + O(h2 ), (n = 0, 1) (19)
0 j =0

Z 1 N
X
G2 (ti , s)ϕ(s)ds = hρj G∗2 (ti , tj )ϕ(tj ) + O(h2 ). (20)
0 j =0

Proof In the case n = 0, 1, since the functions Gn (ti , s) are continuous at s = ti and
are polynomials in s in the intervals [0, ti ] and [ti , 1], we have:
Z 1 Z ti Z 1
Gn (ti , s)ϕ(s)ds = Gn (ti , s)ϕ(s)ds + Gn (ti , s)ϕ(s)ds
0 0 ti

1
=h Gn (ti , t0 )ϕ(t0 ) + Gn (ti , t1 )ϕ(t1 ) + ... + Gn (ti , ti−1 )ϕ(ti−1 )
2

1
+ G2 (ti , ti )ϕ(ti )
2

1
+h Gn (ti , ti )ϕ(ti ) + Gn (ti , ti+1 )ϕ(ti+1 ) + ... + Gn (ti , tN −1 )ϕ(tN −1 )
2

1
+ Gn (ti , tN )ϕ(tN ) + O(h2 )
2
N
X
= hρj Gn (ti , tj )ϕ(tj ) + O(h2 ) (n = 0, 1).
j =0

Thus, the estimate (19) is established. The estimate (20) is obtained using the
following result, which is easily proved.

Lemma 1 Let p(t) be a function having continuous derivatives up to second order

in the interval [0, 1] except for the point ti , 0 < ti < 1, where it has a jump. Denote
limt →ti −0 p(t) = pi− , limt →ti +0 p(t) = pi+ , pi = 12 (pi− + pi+ ). Then
Z 1 N
X
p(t)dt = hρj p(j ) + O(h2 ), (21)
0 j =0

where pj = p(tj ), j 6= i.

Proposition 4 Under the assumption of Proposition 2 for any k = 0, 1, ..., we have

kΦk − ϕk k = O(h2 ), (22)

kUk − uk k = O(h2 ), kYk − yk k = O(h2 ), kZk − zk k = O(h2 ), (23)

where k.kC(ω̄h ) is the max-norm of function on the grid ω̄h .
 Numerical Algorithms

Proof We prove the proposition by induction. For k = 0, we have immediately kΦ0 −

ϕ0 k = 0. Next, by the first equation in (12) and Proposition 3, we have:
Z 1 N
X
u0 (ti ) = G0 (ti , s)ϕ0 (s)ds = hρj G0 (ti , tj )ϕ0 (tj ) + O(h2 ) (24)
0 j =0

for any i = 0, ..., N . On the other hand, in view of the first equation in (16), we have:
N
X
U0 (ti ) = hρj G0 (ti , tj )ϕ0 (tj ). (25)
j =0

Therefore, |U0 (ti ) − u0 (ti )| = O(h2 ). Consequently, kU0 − u0 k = O(h2 ).

Similarly, we have:
kY0 − y0 k = O(h2 ), kZ0 − z0 k = O(h2 ). (26)
Now suppose that (22) and (23) are valid for k ≥ 0. We shall show that these esti-
mates are valid for k + 1.
Indeed, by the Lipschitz condition of the function f and the estimates (23), it is easy
to obtain the estimate:
kΦk+1 − ϕk+1 k = O(h2 ) (27)
Now from the first equation in (12) by Proposition 3, we have:
Z 1 N
X
uk+1 (ti ) = G0 (ti , s)ϕk+1 (s)ds = hρj G0 (ti , tj )ϕk+1 (tj ) + O(h2 )
0 j =0

On the other hand by the first formula in (16), we have:

N
X
Uk+1 (ti ) = hρj G0 (ti , tj )Φk+1 (tj ).
j =0

From the above equalities, having in mind the estimate (27), we obtain the estimate:
kUk+1 − uk+1 k = O(h2 ).
Similarly, we obtain:
kYk+1 − yk+1 k = O(h2 ), kZk+1 − zk+1 k = O(h2 ).
Thus, by induction, we have proved the proposition.

Now, combining Proposition 4 and Theorem 3 results in the following theorem.

Theorem 4 For the approximate solution of the problem (2) obtained by the dis-
crete iterative method (15)–(18) on the uniform grid with grid size h, we have the
estimates:
kUk − uk ≤ M0 pk d + O(h2 ), kYk − u′ k ≤ M1 pk d + O(h2 ),
kZk − u′′ k ≤ M2 pk d + O(h2 ),
where M0 , M1 , M2 are defined by (9) and pk , d are defined by (14).
Numerical Algorithms

Remark 1 We perform the discrete iterative process (15)–(18) until kΦk+1 − Φk k ≤

T OL, where T OL is a given tolerance. From Theorem 4, it is seen that the accu-
racy of the discrete approximate solution depends on both the number q defined in
Theorem 2, which determines the number of iterations of the continuous iterative
method and the grid size h. The number q describes the nature of the BVP; there-
fore, it is necessary to choose an appropriate h consistent with q as the choice of very
small h does not increase the accuracy of the approximate discrete solution. Below,
in examples we shall see this fact.

Remark 2 As mentioned in the “Introduction,” in 2016, Pandey [30] discretized the

problem (1) by quartic splines and proved the second-order convergence only for
the linear case (when f = f (x)). In the following year, in [31], he constructed two
difference schemes for the problem and also proved the second-order convergence
for the linear case. The obtained system of difference equations is solved iteratively
by the Gauss-Seidel or Newton-Raphson method. The error arising in these iterative
methods is not considered together with the error of the discretization.

5 Discrete iterative method 2

Consider another discrete iterative method, named method 2. The steps of this
method are the same as of method 1 with an essential difference in step 2 and now
the number of grid points is even, N = 2n.
Namely, 2’. Knowing Φk (ti ), k = 0, 1, ...; i = 0, ..., N, compute approximately
the definite integrals (12) by the modified Simpson rule:

Uk (ti ) = F (G0 (ti , .)Φk (.)),

Yk (ti ) = F (G1 (ti , .)Φk (.)),
Zk (ti ) = F (G∗2 (ti , .)Φk (.)),

where

F (Gl (ti , .)Φk (.))

PN

 j =0 hρj Gl (ti , tj )Φk (tj ) if i is even
PN hρ G (t , t )Φ (t ) + h (G (t , t )Φ (t ) − 2G (t , t )Φ (t )

j =0 j l i j k j 6 l i i−1 k i−1 l i i k i
=


 +G l (ti , ti+1 )Φ k (ti+1 )) if i is odd ,

l = 0, 1; i = 0, 1, 2, ..., N .

ρj are the weights of the Simpson rule


1/3, j = 0, N

ρj = 4/3, j = 1, 3, ..., N − 1

2/3, j = 2, 4, ..., N − 2,

 Numerical Algorithms

F (G∗2 (ti , .)Φk (.)) is calculated in the same way as F (Gl (ti , .)Φk (.)) above, where
Gl is replaced by G∗2 defined by the formula (17).

Proposition 5 Assume that the function f (t, x, y, z) has all continuous par-
tial derivatives up to fourth order in the domain DM . Then, for the functions
uk (t), yk (t), zk (t), ϕk+1 (t), k = 0, 1, ..., constructed by the iterative method (11)-
(13) we have zk (t) ∈ C 5 [0, 1], yk (t) ∈ C 6 [0, 1], uk (t) ∈ C 7 [0, 1], ϕk+1(t) ∈
C 4 [0, 1].

Proposition 6 For any function ϕ(t) ∈ C 4 [0, 1], we have:

Z 1
Gl (ti , s)ϕ(s)ds = F (Gl (ti , .)ϕ(.)) + O(h3 ), (l = 0, 1) (28)
0

Z 1
G2 (ti , s)ϕ(s)ds = F (G∗2 (ti , .)ϕ(.)) + O(h3 ). (29)
0

Proof Recall that the interval [0, 1] is divided into N = 2n subintervals by the
points ti = ih, h = 1/N . In each subinterval [0, ti ] and [ti , 1] the functions Gl (ti , s)
are continuous as polynomials. Therefore, if i is an even number, i = 2m then we
represent:

Z 1 Z t2m Z 1
Gl (ti , s)ϕ(s)ds = Gl (ti , s)ϕ(s)ds + Gl (ti , s)ϕ(s)ds.
0 0 t2m

Applying the Simpson rule to the integrals in the right-hand side, we obtain:
Z 1
Gl (ti , s)ϕ(s)ds = F (Gl (ti , .)ϕ(.)) + O(h4 )
0

because by assumption ϕ(t) ∈ C 4 [0, 1].

Now consider the case when i is an odd number, i = 2m + 1. In this case, we
represent:
Z 1 Z t2m Z t2m+1
I= Gl (ti , s)ϕ(s)ds = Gl (ti , s)ϕ(s)ds + Gl (ti , s)ϕ(s)ds
0 0 t2m
Z t2m+2 Z 1
+ Gl (ti , s)ϕ(s)ds + Gl (ti , s)ϕ(s)ds. (30)
t2m+1 t2m+2

For simplicity, we denote:

fj = Gl (ti , sj )ϕ(sj )
Numerical Algorithms

Applying the Simpson rule to the first and the fourth integrals in the right-hand side
(30) and the trapezoidal rule to the second and the third integrals, we obtain:
h
I= [f0 + f2m + 4(f1 + f3 + ... + f2m−1 ) + 2(f2 + f4 + ... + f2m−2 )] + O(h4 )
3
h h
+ (f2m + f2m+1 ) + O(h3 ) + (f2m+1 + f2m+2 ) + O(h3 )
2 2
h
+ [f2m+2 + f2n + 4(f2m+3 + f2m+5 + ... + f2n−1 )
3
+2(f2m+4 + f2m+6 + ... + f2n−2 )] + O(h4 )
h
= [f0 + f2n + 4(f1 + f3 + ... + f2n−1 ) + 2(f2 + f4 + ... + f2n−2 )]
3
h
+ (f2m − 2f2m+1 + f2m+2 ) + O(h3 )
6
= F (Gl (ti , .)ϕ(.)) + O(h3 )
Thus, in the both cases of i, even or odd, we have the estimate (28).
The estimate (29) is obtained analogously as (28) if taking into account that
2G∗2 (ti , ti ) = G− +
2 (ti , ti ) + G2 (ti , ti ),
where
G±
2 (ti , ti ) = lim G2 (ti , s).
s→ti ±0

Theorem 5 Under the assumptions of Proposition 5, for the approximate solution

of the problem (2) obtained by Discrete iterative method 2 on the uniform grid with
grid size h, we have the estimates:
kUk − uk ≤ M0 pk d + O(h3 ), kYk − u′ k ≤ M1 pk d + O(h3 ),
kZk − u′′ k ≤ M2 pk d + O(h3 ).

6 Examples

Consider some examples for confirming the validity of the obtained theoretical
results and the efficiency of the proposed iterative method. For the first two exam-
ples, the exact solutions are known, and for the third example the exact solution is
not known.

Example 1 (Problem 2 in [30]) Consider the problem:

u′′′ (x) = x 4 u(x) − u2 (x) + g(x), 0 < x < 1,
(31)
u(0) = 0, u′ (0) = −1, u′ (1) = sin(1),
where g(x) = −3 sin(x) − cos(x)(x − 1) − x 4 (x − 1) sin(x) + (x − 1)2 sin2 (x). It
is possible to verify that the function
u∗ (x) = (x − 1) sin(x)
 Numerical Algorithms

is the exact solution of the problem.

By setting u(x) = v(x) + P (x), where P (x) = 12 (1 + sin(1))x 2 − x is the poly-
nomial of the second degree satisfying the boundary conditions in (31), the problem
for u(x) is reduced to the following problem for v(x):
v ′′′ (x) = x 4 v(x) − v 2 (x) − 2P (x)v(x) + x 4 P (x) − P 2 (x) + g(x), 0 < x < 1,
v(0) = 0, v ′ (0) = 0, v ′ (1) = 0.
(32)
In order to apply Theorem 2 of the existence and uniqueness of solution for the above
problem, we need to determine the number M. For the right-hand side function:
f (x, v) = −v 2 (x) + x 4 v(x) − 2P (x)v(x) + x 4 P (x) − P 2 (x) + g(x)
1
in the domain DM = {(x, v)| 0 ≤ t ≤ 1, |v| ≤ M0 M}, where M0 = 12 we have:
|f | ≤ |v|2 + |v| + 2|P (x)||v| + |x 4 P (x)| + |P (x)|2 + |g(x)|
M M
≤ ( )2 + (1 + 2 ∗ 0.2715) + 0.1 + 0.27152 + 4.12
12 12
M2 1.55M
< + + 4.3
144 12
Here, we use the estimates:
|P (x)| ≤ 0.2715, |x 4 P (x)| ≤ 0.1, x ∈ [0, 1],
that are easily obtained. Besides, for estimating |g(x)|, we use the estimates:
|(x − 1) sin(x)| ≤ 0.2401, |x 4 (x − 1) sin(x)| ≤ 0.0596, x ∈ [0, 1]
2
It is easy to verify that with M = 6 then 144 M
+ 1.55M
12 + 4.3 < M. Hence, for
this chosen M, we have |f (x, v)| ≤ M in DM . Furthermore, in this domain, the
function f (x, v) satisfies the Lipschitz condition in the variable v with the coefficient
L0 = 2.543. Therefore, q = 0.2119. Hence, all conditions of Theorem 2 are satisfied,
so the problem has a unique solution and the iterative method converges. The results
of the numerical experiments with two different tolerances are given in Tables 1, 2,
and 3.
In the above tables, N is the number of grid points, K is the number of iterations,
Errort rap , ErrorSimp are errors kUK − u∗ k in the cases of using method 1 and

Table 1 The convergence in example 1 for T OL = 10−4

N K Errortrap Order ErrorSimp Order

8 3 9.9153e−04 9.7143e−04
16 3 2.4646e−04 2.0083 1.3101e−04 2.8905
32 3 6.0906e−05 2.0167 1.6020e−05 3.0317
64 3 1.4563e−05 2.0643 1.2587e−06 3.6696
128 3 2.9796e−06 2.2891 8.8553e−07 0.5073
256 3 4.3187e−07 2.7865 8.8165e−07 0.0063
Numerical Algorithms

Table 2 The convergence in example 1 for T OL = 10−6

N K Errortrap Order ErrorSimp Order

8 4 9.99237e−04 9.7223e−04
16 4 2.4734e−04 2.0044 1.3189e−04 2.8820
32 4 6.1802e−05 2.0008 1.6915e−05 2.9629
64 4 1.5462e−05 1.9989 2.1492e−06 2.9765
128 4 3.8797e−06 1.9947 2.8688e−07 2.9053
256 4 9.8437e−07 1.9787 5.2749e−08 2.4439
512 4 2.6054e−07 1.9177 2.3446e−08 1.1698
1024 4 7.9583e−08 1.7110 1.9786e−08 0.2448

method 2, respectively, and Order is the order of convergence calculated by the

formula:
N /2
kUK − u∗ k
Order = log2 .
kUKN − u∗ k
In the above formula, the superscripts N/2 and N of UK mean that UK is com-
puted on the grid with the corresponding number of grid points. From the tables, we
observe that for each tolerance the number of iterations is constant and the errors of
the approximate solution decrease with the rate (or order) close to 2 for method 1 and
close to 3 for method 2 until they cannot improve. This can be explained as follows.
Since the total error of the actual approximate solution consists of two terms: the error
of the iterative method on continuous level and the error of numerical integration at
each iteration, when these errors are balanced, the further increase of number of grid
points N (or equivalently, the decrease of grid size h) cannot in general improve the
accuracy of approximate solution.
Notice that in [30] the author used Newton-Raphson iteration method to solve the
nonlinear system of equations arising after discretization of the differential problem.
The iteration process is continued until the maximum difference between two suc-
cessive iterations, i.e., kUk+1 − Uk k is less than 10−10 . The number of iterations for

Table 3 The convergence in example 1 for T OL = 10−10

N K Errortrap Order ErrorSimp Order

8 7 9.9235e−04 9.7222e−04
16 7 2.4732e−04 2.0045 1.3187e−04 2.8822
32 7 6.1782e−05 2.0011 1.6896e−05 2.9643
64 7 1.5443e−05 2.0003 2.1301e−06 2.9877
128 7 3.8605e−06 2.0001 2.6774e−07 2.9923
256 7 9.6511e−07 2.0000 3.3544e−08 2.9965
512 7 2.4128e−07 2.0000 4.1977e−09 2.9984
1024 7 6.0319e−08 2.0000 5.2483e−10 2.9997
 Numerical Algorithms

achieving this tolerance is not reported. The accuracy for some different N is given
in Table 4 (see [30, Table 2])
From the tables of our results and of Pandey, it is clear that our method gives much
better accuracy.

Example 2 (Problem 2 in [31]) Consider the problem:

u′′′ (x) = −xu′′ (x) − 6x 2 + 3x − 6, 0 < x < 1,
u(0) = 0, u′ (0) = 0, u′ (1) = 0.

It is easy to verify that with M = 18, L0 = L1 = 0, L2 = 1, q = 0.5 all condi-

tions of Theorem 2 are satisfied, so the problem has a unique solution. This solution
is u(x) = x 2 ( 32 − x). The results of the numerical experiments with different toler-
ances are given in Tables 5, 6, and 7.

Notice that in [31] the author used Gauss-Seidel iteration method to solve lin-
ear system of equations arisen after discretization of the differential problem. The
iteration process is continued until the maximum difference between two successive
iterations, i.e., kUk+1 − Uk k is less than 10−10 . The results for some different N are
given in Table 8.
From Tables 5, 6, and 7 of our results and Table 8 of the Pandey’ results, it is clear
that our method gives better accuracy and requires less computational work.

Example 3 Consider the problem for fully third-order differential equation:

′ 1
u′′′ (x) = −eu(x) − eu (x) − 10 (u′′ (x))2 , 0 < x < 1,
′ ′ (33)
u(0) = 0, u (0) = 0, u (1) = 0.

For the above problem, the exact solution is not known. It is easy to verify that all
conditions of Theorem 2 are satisfied with M = 3, L0 = 1.284, L1 = 1.455, L2 =
0.3, and q = 0.4389. So, the problem has a unique solution and the iterative method
for it converges. The convergence of the method is reported in Table 9.
The numerical solution of the problem is depicted in Fig. 1.
Notice that in [20] the authors could only establish the existence but not the
uniqueness of a solution to the equation u′′′ (x) = −eu(x) associated with the bound-
ary conditions as in (33), and later, in [36], Yao and Feng also could only obtain the
′
similar result for the equation u′′′ (x) = −eu(x) − eu (x).

Remarks on the convergence of the iterative method It should be remarked that

Theorem 2 in Section 2 provides only sufficient conditions for the existence and

Table 4 The results in [30] for the problem in example 1

N 8 16 32 64

Error 0.11921225e−01 0.33391170e−02 0.87742222e−03 0.23732412e−03

Numerical Algorithms

Table 5 The convergence in example 2 for T OL = 10−4

N K Errortrap Order ErrorSimp Order

8 6 0.0078 9.7662e−04
16 6 0.0020 2.0000 1.2215e−04 2.9991
32 6 4.8837e−04 1.9998 1.5345e−05 2.9929
64 6 1.2216e−04 1.9992 1.9936e−06 2.9443
128 6 3.0604e−05 1.9969 3.2471e−07 2.6181
256 6 7.7157e−06 1.9878 1.1612e−07 1.4835

Table 6 The convergence in example 2 for T OL = 10−6

N K Errortrap Order ErrorSimp Order

8 8 0.0078 9.7662e−04
16 6 0.0020 2.0000 1.2215e−04 2.9991
32 6 4.8837e−04 1.9998 1.5345e−05 2.9929
64 6 1.2216e−04 1.9992 1.9936e−06 2.9443
128 6 3.0604e−05 1.9969 3.2471e−07 2.6181
256 6 7.7157e−06 1.9878 1.1612e−07 1.4835
512 6 1.9937e−06 1.9524 9.0051e−08 0.3868
1024 6 5.6316e−07 1.8238 8.6794e−08 0.0532

Table 7 The convergence in example 2 for T OL = 10−10

N K Errortrap ErrorSimp N K Errortrap ErrorSimp

8 11 0.0078 2.0650e−13 64 11 1.2207e−04 2.5890e−13

16 11 0.0020 2.6790e−13 128 11 3.0518e−05 2.5790e−13
32 11 4.8828e-04 2.6279e−13 256 11 7.6294e−06 2.5802e−13

Table 8 The results in [31] for the problem in example 2

N 128 256 512 1024

Error 0.30696392e−4 0.61094761(−5) 0.14379621e−5 0.41723251e−6

Iter 53 5 3 4

Table 9 The convergence in

example 3 for T OL = 10−10 N 8 16 32 64 128 256
K 15 15 15 15 15 15
 Numerical Algorithms

0.25

0.2

0.15

0.1

0.05

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 1 The graph of the approximate solution in example 3

uniqueness of a solution to the problem (2) and Theorem 3 gives the convergence
rate of the iterative method for finding the solution. When these conditions are not
satisfied, the iterative method may converge or not converge. Below, we give some
examples for illustrating this statement.
First, consider the problem:
′
u′′′ (x) = −eu(x) − eu (x) − (u′′ (x))2 , 0 < x < 1,
u(0) = 0, u′ (0) = 0, u′ (1) = 0.

For this problem, the right-hand side function is f (x, u, y, z) = −eu − ey − z2 . In

the domain:

M M M
DM = {(x, u, y, z)| 0 ≤ x ≤ 1, |u| ≤ , |y| ≤ , |z| ≤ }
12 8 2

we have
!2
M/12 M/8 M
g(M) := max |f (x, u, y, z)| = e +e + .
(x,u,y,z)∈DM 2

It is easy to verify that g(M) ≥ M + 1.4019 > M for any M > 0. Hence, there
does not exist M > 0 such that |f (x, u, y, z)| ≤ M ∀(x, u, y, z) ∈ DM . There-
fore, Theorem 2 cannot guarantee the existence and uniqueness of a solution and the
convergence of the iterative method. Nevertheless, for T OL = 10−10 the iterative
method converges after 23 iterations.
Next, we give an example, when the conditions of Theorem 2 are not satisfied and
the iterative method does not converge. This is the example for the equation:
′
u′′′ (x) = −eu(x) − eu (x) − (u′′ (x))2 + 5u′′ (x) + 10, 0 < x < 1.
Numerical Algorithms

7 On some extensions of the problem

7.1 The problem on large intervals

First, consider the problem (2) on the interval [0, T ], i.e., the problem:
u(3) (t) = f (t, u(t), u′ (t), u′′ (t)), 0 < t < T ,
(34)
u(0) = 0, u′ (0) = 0, u′ (T ) = 0.
For this problem, it is easy to verify that the Green function is:
(
s t2
( − 2t + s), 0 ≤ s ≤ t ≤ T ,
G0 (t, s) = 2t 2 T s
2 ( T − 1), 0 ≤ t ≤ s ≤ T.
The first and the second derivatives of this function with respect to t are
 t
s( T − 1), 0 ≤ s ≤ t ≤ T ,
G1 (t, s) =
t ( Ts − 1), 0 ≤ t ≤ s ≤ T ,
 s
, 0 ≤ s ≤ t ≤ T,
G2 (t, s) = Ts
T − 1, 0 ≤ t ≤ s ≤ T.
It is easy to see that G0 (t, s) ≤ 0, G1 (t, s) ≤ 0 in Q = [0, 1]2 and
T3 T2
Z T Z T
M0 = max |G(t, s)| ds = , M1 = max |G1 (t, s)| ds = ,
0≤t ≤T 0 12 0≤t ≤T 0 8
Z T (35)
T
M2 = max |G2 (t, s)| ds = .
0≤t ≤T 0 2
Clearly, the numbers Mi (i = 0, 1, 2) increase with the increase of T . Therefore,
the domain DM becomes more extended. This implies that the Lipschitz coeffi-
cients L0 , L1 , L2 of the function f (t, x, y, z) with respect to x, y, z do not decrease,
and accordingly, the number q = L0 M0 + L1 M1 + L2 M2 increases. This leads
to narrowing the scope of applicability of Theorem 2 on the existence and unique-
ness of solution and Theorem 3 on the convergence of the iterative method. + For
demonstrating the above remark, we consider some examples.

Example 2a Consider the problem on [0, T ] for the equation of example 2 in the
previous section, namely, the problem:
u′′′ (x) = −xu′′ (x) − 6x 2 + 3x − 6, 0 < x < T ,
u(0) = 0, u′ (0) = 0, u′ (T ) = 0.

Below, in Table 10 there the results of convergence for the discrete iterative method
2 with n = 256 for some T :

Table 10 The convergence in

example 2a for T OL = 10−6 T 1 2 3 4 5
K 8 18 82 2009 No convergence
 Numerical Algorithms

Table 11 The convergence in

example 4 for T OL = 10−6 T 1 3 5 10 15 20 40 100
K 6 12 13 16 18 20 27 37

Here, K is the number of iterations for achieving the given tolerance T OL.
Notice that from T = 2 the conditions of Theorem 2 are not satisfied but only from
T = 5 the iterative method diverges. From Table 10, it is clear that the convergence
of the iterative method depends on the width of the interval, where the problem is
considered.

Example 4 Consider the problem:

2 ′′ 2
u′′′ (x) = − 16 e−u + e−(u ) , 0 < x < T ,
u(0) = 0, u′ (0) = 0, u′ (T ) = 0.
2
For this example, the right-hand side function is f = f (x, u, y, z) = − 16 e−u +
e −(z)2 . In any domain

T3 T2 T
DM = {(x, u, y, z)| 0 ≤ x ≤ T , |u| ≤ M, |y| ≤ M, |z| ≤ M},
12 8 2
we always have |f | ≤ 76 . Therefore, in Theorem 2, we take M = 76 . The Lipschitz
coefficients of the function f are L0 = 0.1430, L1 = 0, L2 = 0.8579. So, q =
3
0.1430 T12 + 0.8579 T2 = 0.0119 T 3 + 0.4289 T . Clearly, for large values of T not all
conditions of Theorem 2 are satisfied, and it is expected that the iterative method will
diverge for large T . But it is interesting that this does not occur. Below, in Table 11
there are the results of the convergence of the iterative method for n = 200.

The approximate solution of the problem for T = 100 is depicted in Fig. 2.

-1000

-2000

-3000
u

-4000

-5000

-6000
0 20 40 60 80 100
x

Fig. 2 The graph of the approximate solution in example 4

Numerical Algorithms

7.2 The problem for unbounded nonlinear terms

For the problem with unbounded nonlinear terms (right-hand sides) f (t, u, y, z)
caused by singular points, of course, Theorem 2 cannot work, and Theorem 3 can-
not ensure the convergence of the iterative method. But it is interesting that in some
special cases, the discrete iterative methods still converge. Below, we report some
nonlinear terms f (t, u, y, z) for which the iterative method converges:
2 u2
(i) √u + ey + 1, (ii) |t − π4 |
+ ey + z2 + 1,
|t − π4 |
u2√
(iii) + ey + z2 + 1.
|t − 42 |

Notice that in the above three functions, the singular points are irrational points;
therefore, when using the discrete methods on the grids with rational points, then
the denominators always are not zero. For this reason, the computations can be
performed.
When we use the uniform grids with the number of grid points n = 2k , k =
2
3, 4, 5, ..., the iterative methods also converge for f = q u 1 + ey + 1. This is due
|t − 3 |
to the fact that i/2k 6= 1/3 for any i and k.
Above, we only made some remarks on the problem (2) when the nonlinear term
is unbounded. In the future, we will study this issue deeply.

8 Conclusion

In this paper, we established the existence and uniqueness of solution for a bound-
ary value problem for fully third-order differential equations. Next, for finding this
solution, we proposed iterative methods at both continuous and discrete levels. The
numerical realization of the discrete iterative methods is very simple. It is based on
the popular trapezoidal rule and a modified Simpson rule for numerical integration.
One of the important results is that we obtained an estimate for the total error of
the approximate solution which is actually obtained. This total error depends on the
number of iterations performed and the discretization parameter. The validity of the
theoretical results and the efficiency of the iterative methods are illustrated in exam-
ples. In addition, we made some remarks on the iterative method for two extensions
of the problem for large intervals and when the nonlinear terms are unbounded due
to interior singular points. In the future, we will deeply study these issues.
The method for investigating the existence and uniqueness of solution and the
iterative schemes for finding solution in this paper can be applied to other third-
order nonlinear boundary value problems, and in general, for higher order nonlinear
boundary value problems.

Funding The second author, Dang Quang Long, was supported by Institute of Information Technology,
VAST under the project CS 20.01.
 Numerical Algorithms

References

1. Abushammala, M., Khuri, S.A., Sayfy, A.: A novel fixed point iteration method for the solution of
third order boundary value problems. Appl. Math. Comput. 271, 131–141 (2015)
2. Agarwal, R.P.: Boundary Value Problems for Higher Order Differential Equations. World Scientific,
Singapore (1986)
3. Al-Said, E.A.: Numerical solutions for system of third-order boundary value problems. Int. J. Comput.
Math. 78(1), 111–121 (2001)
4. Al-Said, E.A., Noor, M.A.: Cubic splines method for a system of third boundary value problems.
Appl. Math. Comput. 142, 195–204 (2003)
5. Al-Said, E.A., Noor, M.A.: Numerical solutions of third-order system of boundary value problems.
Appl. Math. Comput. 190, 332–338 (2007)
6. Bai, Z.: Existence of solutions for some third-order boundary-value problems. Electron. J. Differential
Equations 2008(25), 1–6 (2008)
7. Cabada, A.: The method of lower and upper solutions for third-order periodic boundary value
problems. J. Math. Anal. Appl. 195, 568–589 (1995)
8. Calagar, H.N., Cagalar, S.H., Twizell, E.H.: The numerical solution of third order boundary value
problems with fourth degree B-Spline. Int. J. Comput. Math. 71, 373–381 (1999)
9. Chaurasia, A., Srivastava, P.C., Gupta, Y.: Exponential Spline Approximation for the Solution
of Third-Order Boundary Value Problems. In: Balas V., Sharma N., Chakrabarti A. (eds.) Data
Management, Analytics and Innovation. Adv. Intell. Syst. Comput. 808 (2019)
10. El-Danaf, T.S.: Quartic nonpolynomial spline solutions for third order Two-Point boundary value
problem. Inter. J. Math. Comput. Sci. 2(9), 637–640 (2008)
11. Dang, Q.A., Ngo, T.K.Q.: Existence results and iterative method for solving the cantilever beam
equation with fully nonlinear term. Nonlinear Anal. Real World Appl. 36, 56–68 (2017)
12. Dang, Q.A., Dang, Q.L., Ngo, T.K.Q.: A novel efficient method for nonlinear boundary value
problems. Numer. Algorithms 76, 427–439 (2017)
13. Dang, Q.A., Ngo, T.K.Q.: New fixed point approach for a fully nonlinear fourth order boundary value
problem. Bol. Soc. Paran. Mat. 36(4), 209–223 (2018)
14. Dang, Q.A., Nguyen, T.H.: The unique solvability and approximation of BVP for a nonlinear fourth
order kirchhoff type equation. East Asian J. Appl. Math. 8(2), 323–335 (2018)
15. Dang, Q.A., Nguyen, T.H.: Existence results and iterative method for solving a nonlinear biharmonic
equation of Kirchhoff type. Comput. Math. Appl. 76, 11–22 (2018)
16. Dang, Q.A., Dang, Q.L.: A simple efficient method for solving sixth-order nonlinear boundary value
problems. Comp. Appl. Math. 37(1), 16 (2018)
17. Dang, Q.A., Nguyen, T.H.: Existence results and numerical method for a fourth order nonlinear
problem. Inter. J. Appl. Comput. Math. 4, 148 (2018)
18. Dang, Q.A., Nguyen, T.H.: Solving the Dirichlet problem for fully fourth order nonlinear differential
equation. Afr. Mat. 30, 623–641 (2019)
19. Fazal-I-Haq, I.H., Ali, A.: A Haar wavelets based numerical method for third-order boundary and
initial value problems. World Appl Sci J 13(10), 2244–2251 (2011)
20. Feng, Y., Liu, S.: Solvability of a third-order two-point boundary value problem. Appl. Math. Lett.
18, 1034–1040 (2005)
21. Gao, F., Chi, C.M.: Solving third-order obstacle problems with quartic B-splines. Appl. Math. Comp.
180(1), 270–274 (2006)
22. Grossinho, M.R., Minhos, F.: Existence result for some third order separated boundary value
problems. Nonlinear Anal. 47, 2407–2418 (2001)
23. Guo, Y., Liu, Y., Liang, Y.: Positive solutions for the third-order boundary value problems with the
second derivatives. Bound. Value Probl. 2012, 34 (2012)
24. Islam, S., Khan, M.A., Tirmizi, I.A., Twizell, E.H.: Non-polynomial splines approach to the solution
of a system of third-order boundary-value problems. Appl. Math. Comput. 168(1), 152–163 (2007)
25. Islam, S., Tirmizi, I.A., Khan, M.A.: Quartic non-polynomial spline approach to the solution of a
system of third-order boundary-value problems. J. Math. Anal. Appl. 335(2), 1095–1104 (2007)
26. Khan, A., Aziz, T.: The Numerical Solution of Third Order Boundary Value Problems using quintic
spline. Appl. Math. Comput. 137, 253–260 (2003)
27. Khan, A., Sultana, T.: Non-polynomial quintic spline solution for the system of third order boundary-
value problems. Numer. Algorithms 59, 541–559 (2012)
Numerical Algorithms

28. Lv, X., Gao, J.: Treatment for third-order nonlinear differential equations based on the Adomian
decomposition method. LMS J. Comput. Math. 20(1), 1–10 (2017)
29. Noor, M.A., Al-Said, E.A.: Quartic spline solutions of third order obstacle boundary value problems.
Appl. Math. Comput. 153, 307–316 (2004)
30. Pandey, P.K.: Solving third-order boundary value problems with quartic splines. SpringerPlus 5, 326
(2016)
31. Pandey, P.K.: A numerical method for the solution of general third order boundary value problem in
ordinary differential equations. Bull. Inter. Math. Virtual Inst. 7, 129–138 (2017)
32. Pue-on, P., Viriyapong, N.: Modified adomian decomposition method for solving particular third-
order ordinary differential equations. Appl. Math. Sci. 6(30), 1463–1469 (2012)
33. Rezaiguia, A., Kelaiaia, S.: Existence of a positive solution for a third-order three point boundary
value problem. Mat. Vesn. 68(1), 12–25 (2016)
34. Srivastava, P.K., Kumar, M.: Numerical algorithm based on quintic nonpolynomial spline for solving
third-order boundary value problems associated with draining and coating flow. Chin. Ann. Math. Ser.
B 33(6), 831–840 (2012)
35. Sun, Y., Zhao, M., Li, S.: Monotone positive solution of nonlinear third-order two-point boundary
value problem. Miskolc Math. Notes 15, 743–752 (2014)
36. Yao, Q., Feng, Y.: The existence of solutions for a third order two-point boundary value problem.
Appl. Math. Lett. 15, 227–232 (2002)
37. Zhai, C., Zhao, L., Li, S., Marasi, H.R.: On some properties of positive solutions for a third-order
three-point boundary value problem with a parameter. Adv. Difference Equ. 2017, 187 (2017)

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.