Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Numerical solution for the weakly singular Fredholm

integro-differential equations using Legendre multiwavelets
Mehrdad Lakestani a , Behzad Nemati Saray a , Mehdi Dehghan b,∗
a
Faculty of Mathematical Sciences, University of Tabriz, Tabriz, Iran
b
Department of Applied Mathematics, Faculty of Mathematics and Computer Sciences, Amirkabir University of Technology, No. 424 Hafez Avenue, Tehran, Iran

article info abstract

Article history: An effective method based upon Legendre multiwavelets is proposed for the solution
Received 10 December 2009 of Fredholm weakly singular integro-differential equations. The properties of Legendre
Received in revised form 20 January 2011 multiwavelets are first given and their operational matrices of integral are constructed.
These wavelets are utilized to reduce the solution of the given integro-differential equation
Keywords: to the solution of a sparse linear system of algebraic equations. In order to save memory
Legendre multiwavelets
requirement and computational time, a threshold procedure is applied to obtain the
Weakly singular integro-differential
equation
solution to this system of algebraic equations. Through numerical examples, performance
Spectral method of the present method is investigated concerning the convergence and the sparseness of
Operational matrix of integral the resulted matrix equation.
Thresholding © 2011 Elsevier B.V. All rights reserved.
Sparse matrix

1. Introduction

The main aim of the current paper is to present high order numerical methods for solving the weakly singular Fredholm
integro-differential equations of the form
∫ b
µ 1 u′ ( t ) + µ 2 u( t ) + k(s, t )|t − s|−α u(s)ds = f (t ), a ≤ t ≤ b, (1.1)
a

with
u(a) = λ, (1.2)
where µ1 .µ2 ̸= 0, 0 < α < 1 and λ is a constant and k and f are given functions and u is the solution to be determined.
Moreover we assume that the kernel k is in L2 [a, b]2 and the unknown u and the right-hand-side f are in L2 [a, b]. For
simplicity, we restrict our attention to the interval [a, b] = [0, 1]. Also we suppose that k(s, t ) satisfies in the Lipschitz
condition,
|k(s1 , t ) − k(s2 , t )| ≤ Ls |s1 − s2 |, (1.3)
where Ls is the Lipschitz constant.
Integral equation (1.1) with µ1 = 0 often arises in practical applications such as Dirichlet problems, mathematical
problems of radiative equilibrium and radiative heat transfer problems [1–3].

∗
Corresponding author. Fax: +98 21 66497930.
E-mail addresses: [email protected] (M. Lakestani), [email protected] (B.N. Saray), [email protected], [email protected]
(M. Dehghan).

0377-0427/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2011.01.043
3292 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

As said in [4] the construction of high order methods for these equations is, however, not an easy task because of the
singularity in the ‘‘weakly singular’’ kernel |t − s|−α k(t , s); in fact, in this case the solution u is generally not differentiable
at the endpoints (i.e. t = a and t = b) [5,4,6,2,7], and due to this, to the best of the authors’ knowledge the best convergence
rate if achieved remains only at polynomial order. For example, if we set the uniform meshes with n + 1 grid points and
apply the spline methods of order m, then the convergence rate is O(n−2p ) at most [8,9,3], and it cannot be improved by
increasing m. We refer the interested reader to [4] for a nice investigation on weakly singular Fredholm integral equations
of the second kind.
In the current paper we present an alternative way of solving weakly singular Fredholm integral equations which is based
on using the Legendre multiwavelets.
Wavelet theory is relatively new and is an emerging area in mathematical research. In recent years, wavelets have found
their way into different fields of science and engineering. Wavelet analysis assumed significance due to the successful
applications in signal and image processing during the eighties. The smooth orthonormal basis obtained by the translation
and dilation of a single function in a hierarchical fashion proved very useful to develop compression algorithms for signals
and images up to a chosen threshold of relevant amplitudes [10,11].
After discretizing the given differential equation in a conventional way like the finite difference approximation [12–14],
wavelets can be used for algebraic manipulations in the obtained system of equations which may lead to better condition
number of the resulting system. Beylkin’s approach [15] to the study of partial differential equations is based on this idea
where he has shown that the differential operator is approximately diagonal in the wavelet bases.
Another approach to the study of differential equations is to use the wavelet bases in the place of other conventional
bases like Fourier, Legendre or Chebyshev bases in the spectral methods [16,17,15,18–22] and so on. Wavelets permit the
accurate representation of a variety of functions and operators. Orthogonal functions and polynomial series have been
receiving considerable attention in dealing with various problems of dynamic systems. In the current investigation, we
consider the application of wavelet bases in the solution of integro-differential equations. The conventional discretization
of the integral equations leads to a dense matrix owing to the non-local nature of the situation. It was discovered in [23] that
the representation of an integral operator by compactly supported orthonormal wavelets produces sparse matrices to some
degrees of precision. Alpert et al. [16] introduced wavelet like bases for the fast solution [24] of the second kind integral
equations in O(n log n) operations where n is the number of points in the discretization. Glowinski et al. [25] considered
wavelet based on variational methods to solve the one-dimensional linear and nonlinear ordinary differential equations.
Different variations of wavelet bases (orthogonal, biorthogonal, multiwavelets) have been presented and the design of
the corresponding wavelet and scaling functions have been addressed [26,10,11,27]. Multiwavelets are generated by more
than one scaling function [28,27]. Multiwavelets have some advantages in comparison to single wavelets. The properties
such as short support, orthogonality, symmetry and vanishing moments are known to be important in signal processing
and numerical methods. A single wavelet cannot possess all these properties at the same time. On the other hand, a
multiwavelet system can have all of them simultaneously [29]. This suggests that multiwavelets could perform better in
various applications.
In this paper, we use the Legendre multiwavelets for solving weakly singular integro-differential equations. These
multiwavelets have been constructed in [30] and also have been considered in [17,31,29]. Our method consists of reducing
the given weakly singular integro-differential equation to a set of algebraic equations by expanding the equation as Legendre
multiwavelets with unknown coefficients. The properties of these multiwavelets are then utilized to evaluate the unknown
coefficients. We refer the interested reader to [32–44] for some basic ideas on numerical solution of integral equations. Also
the reader can see [45,46] for applications of the partial integro-differential equations.
The paper is organized as follows: Section 2 is devoted to the basic formulation of the Legendre multiwavelets required
for our subsequent development. In Section 3 the proposed method is used to approximate the weakly singular integro-
differential equations. In Section 4, we report our numerical findings and demonstrate the accuracy of the proposed
numerical scheme by considering several examples.

2. Legendre multiwavelet systems

2.1. Multiresolution analysis

For functions φ m ∈ L2 (R), m = 0, . . . , r, let a reference subspace or sample space V0 be generated as the L2 -closure of
the linear span of the integer translation of φ m , namely:

V0 = closL2 φ m (. − k) : k ∈ Z , m = 0, . . . , r ,
 

and consider other subspace

Vj = closL2 φjm,k : k ∈ Z , j ∈ Z , m = 0, . . . , r ,
 

where φjm,k = φ m (2j x − k), j, k ∈ Z , m = 0, . . . , r.

 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3293

Definition 1. Functions φ m ∈ L2 (R), are said to generate a multiresolution analysis (MRA) if they generate a nested sequence
of closed subspaces Vj that satisfies

(i) ··· ⊂ V−1 ⊂  V0 ⊂ V1 ⊂, . . . ,




 
(ii) Vj = L2 (R),

closL2




 j ∈Z
(2.1)


 (iii) Vj = {0},
j∈Z


(iv) f (x) ∈ Vj ⇐⇒ f (x + 2−j ) ∈ Vj ⇐⇒ f (2x) ∈ Vj+1 ,




(v) {φ m (. − k)}k∈Z , form a Riesz basis of V0 .
If φ m generate an MRA, then φ m are called scaling functions. In the
 ∞case where the different integer translates of φ m are
orthogonal (with respect to the standard inner product ⟨f , g ⟩ = −∞ f (x)g (x)dx for two functions in L (R)), denoted by
2

φ m (. − k)⊥φ m̃ (. − k̃) for m ̸= m̃, k ̸= k̃, the scaling functions are called orthogonal scaling functions.
As the subspaces Vj are nested, there exist complementary orthogonal subspaces Wj such that

Vj +1 = Vj Wj , j ∈ Z,

here and in the following denotes orthogonal sums.
This yields an orthogonal decomposition of L2 (R), namely:

L2 (R) = Wj .
j∈Z

Definition 2. Functions ψ m ∈ L2 (R) are called wavelets, if they generate the complementary orthogonal subspaces Wj of
an MRA, i.e.,

,k , k ∈ Z ⟩,
Wj = closL2 ⟨ψjm j ∈ Z , m = 0, . . . , r ,

where ψ m
j ,k = ψ (2 x − k), j, k ∈ Z .
m j

j/2 m j̃/2 m̃
Obviously, ψjm
,k ⊥ψ for j ̸= j̃, m ̸= m̃ and k ̸= k̃ if ⟨2 ψj,k , 2 ψ ⟩δj,j̃ δk,k̃ δm,m̃ then ψ are called orthonormal wavelets.
m̃ m
j̃,k̃ j̃,k̃
Now we define Legendre scaling functions and the corresponding multiwavelets according to the above MRA.

2.2. Construction of scaling functions

Legendre multiwavelet systems with multiplicity r consist of r + 1 scaling functions and r + 1 wavelets. The rth order
Legendre scaling functions are the set of r + 1 functions φ 0 (x), . . . , φ r (x) where φ i (x) is a polynomial of ith order and all
φ ’s form orthonormal basis, that is [30,31], for i = 0, 1, . . . , r,
i
−
φ i ( x) = aik xk , for i = 0, 1, . . . , r . (2.2)
k=0

The coefficients aik are chosen such that aik ≥ 0, and

∫ 1
φ i (x)φ k (x)dx = δi,k , for i, k = 0, 1, . . . , r , (2.3)
0
where
1, i = k,

δi,k =
0, i ̸= k.
The scaling functions φ i (x) have symmetry, anti-symmetry properties for odd or even i, respectively. The two-scale
relations for Legendre scaling functions of order r, are in the form [30]:
r
− r
−
φ i ( x) = pi,j φ j (2x) + pi,r +j+1 φ j (2x − 1), i = 0, 1, . . . , r . (2.4)
j =0 j =0

The coefficients {p} are determined uniquely by substituting Eq. (2.2) into (2.4). Now we would like to mention two remarks
on the two-scale relations. These remarks are taken from the literature [47].
i. Since φ i (x) is a ith order polynomial, the right hand side of (2.4) has at most ith order scaling functions. Therefore,
pi,j = pi,r +j+1 = 0 for i < j.
ii. The two-scale relations for the Legendre scaling function of order n which is lower than r is a subset of the first n two-scale
relations for φ i for i = 0, 1, . . . , n from rth order two-scale relations.
3294 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

2.3. Construction of wavelets

The two-scale relations for the rth order Legendre multiwavelets are in the form [30]:
r
− r
−
ψ i ( x) = qi,j φ j (2x) + qi,r +j+1 φ j (2x − 1). (2.5)
j =0 j=0

As we have 2(r + 1)2 unknown coefficients {qi,j } in (2.5), we use the following 2r (r + 1) vanishing moment conditions (2.6)
and 2(r + 1) orthonormal conditions (2.7) to determine them.
1. Vanishing moments
∫ 1
ψ i (x)xj = 0, for i = 0, 1, . . . , r , j = 0, 1, . . . , i + r . (2.6)
0
2. Orthonormality
∫ 1
ψ i (x)ψ j (x) = δi,j , for i, j = 0, 1, . . . , r . (2.7)
0

For example the cubic Legendre scaling functions consist of four functions are given in the following:
φ (x) = 1 √, 0 ≤ x < 1,
 0

φ (x) = √3(2x − 1), 0 ≤ x < 1,
 1

(2.8)

 φ 2 (x) = √5(6x2 − 6x + 1), 0 ≤ x < 1,
φ (x) = 7(20x3 − 30x2 + 12x − 1), 0 ≤ x < 1.
 3

The closed form solution to the cubic Legendre multiwavelets ψ 0 (x), ψ 1 (x), ψ 2 (x) and ψ 3 (x) are given in (2.9)–(2.12)
which are determined using conditions (2.6) and (2.7). Figs. 1 and 2 show the plots of cubic Legendre multiwavelets.
 
15 1
− (224x3 − 216x2 + 56x − 3), 0≤x< ,


17 2
ψ ( x) = 
0
(2.9)
 15 (224x3 − 456x2 + 296x − 61), 1

≤ x < 1,

17 2

1 1
(1680x3 − 1320x2 + 270x − 11), 0≤x< ,



21 2
ψ ( x) = 
1
(2.10)
 1 (1680x3 − 3720x2 + 2670x − 619), 1

≤ x < 1,

21 2
 
35 1
− (256x3 − 174x2 + 30x − 1), 0≤x< ,


17 2
ψ ( x) = 
2
(2.11)
 35 (256x3 − 594x2 + 450x − 111), 1

≤ x < 1,

17 2

5 1
(420x3 − 246x2 + 36x − 1), 0≤x< ,



42 2
ψ ( x) = 
3
(2.12)
 5 (420x3 − 1014x2 + 804x − 209), 1

≤ x < 1.

42 2

2.4. Function approximation

j −1
It can be verified that Vj ⊕ Wj = Vj+1 , thus we can write Vj = V0 ⊕ (⊕i=0 Wi ) and we have two kinds of basis sets for
J ∈N
 T
ΦJ (x) = φJ0,0 (x), . . . , φJr,0 (x), |, . . . , φJ0,(2J −1) (x), . . . , φJr,(2J −1) (x) , (2.13)

ΨJ ( x ) = φ00,0 (x), . . . , φ0r ,0 (x), |ψ00,0 (x), . . . , ψ0r,0 (x)|,
T
. . . , |ψ 0
J −1,0 (x), . . . , ψ r
J − 1 ,0 (x)|, . . . , ψ 0
J −1,2J −1 −1
(x), . . . , ψ r
J −1,2J −1 −1
(x) . (2.14)
 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3295

Fig. 1. The plots of cubic Legendre multiwavelets ψ 0 (left), and ψ 1 (right).

Fig. 2. The plots of cubic Legendre multiwavelets ψ 2 (left), and ψ 3 (right).

Now any function f (x) on [0, 1] can be approximated using the scaling functions as [31,48]

2J −1
r
−−
f (x) ≈ PJ f (x) = cJ ,k φJm,k (x) = C T ΦJ (x), (2.15)
k=0 m=0

and the corresponding wavelet functions as

 
r  J −1 2j −1
− − − 
f (x) ≈ PJ f (x) = c0m,0 φ0m,0 (x) + dm ψ m
j,k j,k (x ) = DT ΨJ (x), (2.16)
m=0
 j =0 k=0


where
∫ 1
cJm,k = f (x)φJm,k (x)dx, (2.17)
0
∫ 1
dm
j ,k = f (x)ψjm
,k (x)dx, (2.18)
0

and D and C are (n × 1) vectors with n = (r + 1)2J given by

 T
D = c00,0 , . . . , c0r ,0 |d00,0 , . . . , dr0,0 | . . . |d0J −1,0 , . . . , drJ−1,0 |, . . . , d0J −1,2J −1 −1 , . . . drJ−1,2J −1 −1 , (2.19)
 T
C = cJ0,0 , . . . , cJr,0 | . . . |cJ0,2J −1 , . . . , cJr,2J −1 . (2.20)

We refer to PJ as the orthogonal projection of f onto VJ .

Theorem 2.1. Suppose that the function f : [0, 1] → R and f ∈ C r [0, 1]. Then PJ f approximates f with mean error bounded as
follows [28]:
2
‖PJ f − f ‖ ≤ 2−nr sup |f (r ) (x)|.
4r r ! x∈[0,1]
3296 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

2.5. The operational matrix of integral

The integral of vectors ΨJ (x) and ΦJ (x) can be expressed as

∫ x
ΨJ (t )dt = Iψ ΨJ (x), (2.21)
0
∫ x
ΦJ (t )dt = Iφ ΦJ (x), (2.22)
0

where Iφ and Iψ are (n × n) operational matrices of integral for Legendre scaling functions and multiwavelets, respectively.
The matrix Iψ can be obtained by the following process. Let

a0 = 1, ai = a2i−1 + 2, for i = 1, 2, . . . , r ,
1
bi = , for i = 1, 2, . . . , r ,
ai−1 ai
0 b1 0 
 0 b2 0 
 0 b3 0 
.. .. ..
 
Ar =  . . . ,
 

 
 0 b r −2 0 
 
0 b r −1
0 r ×r
 
1 ··· 0
1 . .. 
Br = .. . ,
2J
0 ··· 0 r ×r
1
Ar − ATr + 2J Br .
 
Mr =
2 J +1
Now it can be shown that
Mr Br ··· Br
 
.. .. .. 
. . .  .

Iφ = 
Mr Br
 
0 Mr n,n

The matrix Iψ can be obtained by considering

ΨJ = GΦJ +1 , (2.23)
where G is an (n × n) matrix, which can be calculated as follows.
Eq. (2.4) gives
Φk = Pk Φk+1 , (2.24)
where Pk , k = 1, 2, . . . , J is a (r2 k−1
, r2 ) matrix, and the entries of Pk are the coefficients pi,j given in Eq. (2.4). From (2.5)
k

we have
Ψk = Qk Φk+1 , (2.25)
where Qk , k = 1, 2, . . . , J is a (r2 k−1
, r2 ) matrix, and the entries of Qk are the coefficients qi,j given in Eq. (2.5). Using
k

Eqs. (2.23)–(2.25) yield

 
P1 × P2 × · · · × PJ
 Q1 × P2 × · · · × PJ

 .. 
. .
 
G=  (2.26)
QJ −2 × PJ −1 × · · · × PJ 
 
 Q ×P 
J −1 J
QJ n×n

Also using Eqs. (2.21)–(2.23) we have

∫ x ∫ x
ΨJ (t )dt = G ΦJ +1 (t )dt = GIφ ΦJ +1 (x) = GIφ G−1 ΨJ (x). (2.27)
0 0
 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3297

Comparing Eqs. (2.21) and (2.27) we get

Iψ = GIφ G−1 . (2.28)

3. Description of numerical method

In this section, we solve the one-dimensional integro-differential equation of the form (1.1) with initial value (1.2), using
the Legendre multiwavelets.
The integral term in Eq. (1.1) can be written as:
∫ 1 ∫ 1 ∫ 1
k(s, t )|t − s|−α u(s)ds = (k(s, t ) − k(t , t ))|t − s|−α u(s)ds + k(t , t )|t − s|−α u(s)ds
0 0 0
∫ 1
= (k(s, t ) − k(t , t ))|t − s|−α u(s)ds + k(t , t )
0
[∫ 1 ∫ 1 ]
× (u(s) − u(t ))|t − s| ds +
−α
u(t )|t − s| ds .
−α
0 0

Now when s −→ t, using Lipschitz condition (1.3), we get

∫ 1  ∫ 1
 (k(s, t ) − k(t , t ))|t − s|−α u(s)ds ≤ |(k(s, t ) − k(t , t ))| |t − s|−α |u(s)|ds
 
 
0 0
∫ 1
≤ Ls |t − s|1−α |u(s)|ds −→ 0. (3.1)
0

Also we have
∫ 1  ∫ 1 ∫ 1
 (u(s) − u(t ))|t − s|−α ds ≤ |(u(s) − u(t ))| |t − s|−α ds ≤ u′ (t )
 
  |t − s|1−α ds
0 0 0

|t − s|2−α
= u′ ( t ) −→ 0. (3.2)
2−α
We introduce the function h(s, t ) as

k(s, t )|t − s|−α , t ̸= s,


h(s, t ) = (3.3)
0 , t = s.
Using Eqs. (3.1)–(3.3), we get
∫ 1 ∫ 1
k(s, t )|t − s|−α u(s)ds = h(s, t )u(s)ds. (3.4)
0 0

Note that here the kernel h(s, t ) is not singular in [0, 1]. Now using Eq. (3.4) the IDE (1.1) is changed to
∫ 1
µ 1 u′ ( t ) + µ 2 u( t ) + h(s, t )u(s)ds = f (t ). (3.5)
0

Using Eq. (2.16), the function u′ (t ) can be approximated as

u′ (t ) = U T ΨJ (t ). (3.6)
Note that U is an n × 1 unknown vector which should be found.
By integrating from both sides of Eq. (3.6), and using Eq. (2.21), we get
∫ t
u(t ) − u(0) = U T ΨJ (x)dx = U T Iψ ΨJ (t ). (3.7)
0

Applying the boundary value (1.2) in Eq. (3.7), we get

u(t ) = U T IΨ ΨJ (t ) + λ. (3.8)
The functions h(s, t ) and f (t ) in Eq. (3.5) can be approximated as

h(s, t ) = ΨJT (s)H ΨJ (t ), (3.9)

f (t ) = F ΨJ (t ),
T
(3.10)
3298 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

where H is an n × n matrix and F is an n × 1 vector that can be found as

1 1
∫ [∫ ]
Hi,j = h(s, t )ΨJi,1 (t )dt ΨJj,1 (s)ds, (3.11)
0 0
∫ 1
Fi,1 = f (t )ΨJi,1 (t )dt . (3.12)
0

Applying (3.6) and (3.8)–(3.10) in Eq. (3.5), we get

∫ 1
µ1 U ΨJ (t ) + µ2 (U IΨ ΨJ (t ) + λ) +
T T
(U T Iψ ΨJ (s) + λ)ΨJT (s)H ΨJ (t )ds = F T ΨJ (t ). (3.13)
0

The constant λ in Eq. (3.13) can be approximated as

λ = ΛT ΨJ (s), (3.14)

where Λ is an (n × 1) vector with the entries

Λ = λ[1, . . . , 1, |0, . . . , 0| . . . |0, . . . , 0|]T .

Replacing Eq. (3.14) in Eq. (3.13) we obtain
1
∫ 
µ1 U T + µ2 (U T IΨ + ΛT ) ΨJ (t ) + (U T Iψ + ΛT ) ΨJ (s)
ΨJT (s)ds H ΨJ (t ) = F T ΨJ (t ).
 
(3.15)
0

The orthonormality of the Legendre multiwavelets implies

µ1 U T + µ2 U T IΨ + µ2 ΛT + U T Iψ H + ΛT H − F T ΨJ (t ) = 0.
 
(3.16)

Because of independency of entries of vector ΨJ (t ), we get

µ1 U T + µ2 U T IΨ + µ2 ΛT + U T Iψ H + ΛT H − F T = 0. (3.17)

Eq. (3.17) gives a system of linear equations with n equations and n unknowns as

(µ1 I + µ2 IΨ + Iψ H )T U = F − H T Λ − µ2 Λ. (3.18)

Eq. (3.18) can be solved to find U. So the unknown function u(t ) may be found using Eq. (3.8).

4. Computational results and test problems

In this section, some numerical examples are presented to illustrate the validity and the merits of this technique. As
mentioned before one main merit of this technique is the generation of a sparse matrix. This advantage is illustrated in
Examples 1 and 4. Suppose T = (µ1 I + µ2 IΨ + Iψ H )T and C = F − H T Λ − µ2 Λ, then Eq. (3.18) results the linear system
TU = C . We approximate T with a matrix Tε whose elements are defined by the formula

Ti,j , |Ti,j | ≥ ε,

Tεi,j = (4.1)
0 otherwise,

where the threshold ε is chosen so that a desired precision τ is maintained [16]:

‖Tε − T ‖ ≤ τ ‖T ‖. (4.2)

Here the norm ‖.‖ is the row-sum norm. The threshold ε is given by ε = τ ‖T ‖/n. In the current paper we use the threshold
parameter ε = 10−2 , ε = 10−3 and ε = 10−4 and see that by increasing the value of the threshold parameter, the sparsity
and error values are increased regarding Eqs. (4.1) and (4.2).
The percent sparsity (Sε ) of matrix Tε is then defined by [18]
N0 − Nε
Sε = × 100%,
N0
where N0 is the total number of elements Tε and Nε is the number of the nonzero elements Tε . We define the maximum
error as MaxError = max |uexact − uappr | to calculate the errors in Tables 1–5.
 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3299

Table 1
Sparsity and MaxError for r = 3.
x Threshold parameter (ε ) Sparsity (Sε ) (%) MaxError

0 0 3.5 × 10−6
J =3 10−4 62.1 1.1 × 10−5
10−3 86.3 2.9 × 10−5
10−2 96.8 6.3 × 10−4
0 0 2.2 × 10−7
J =4 10−4 71.1 2.8 × 10−6
10−3 90.4 7.4 × 10−5
10−2 98.3 3.3 × 10−5

Fig. 3. Plots of sparse matrix after thresholding with ε = 10−3 (left) ε = 10−2 (right) for Example 1.

Example 1. Consider the following integral equation [4]

1
 ∫
1
u (t ) − |t − s|−1/3 u(s)ds = f (t ),



 10 0
u(0) = 1,
27

f (t ) = t 2 (1 − t )2 − t 8/3 (54t 2 − 126t + 77) + (1 − t )8/3 (54t 2 + 18t + 5) .

  
30 800
The exact solution of this equation is

u(t ) = t 2 (1 − t )2 .

Table 1 shows the sparsity and MaxError for r = 3, J = 3, 4 for different values of the thresholding parameter, using the
method presented in the previous section. Fig. 3 shows the plot of the matrix elements for r = 3 and J = 3 after thresholding.

Example 2. Consider the following integro-differential equation

 ∫ 1
u′ (t ) + |t − s|−1/2 u(s)ds = f (t ),

0
(4.3)
u(0) = −1.


Here the forcing function f is selected such that

u(t ) = t 5 (1 − t + t 2 ) − 1,

is the exact solution of problem (4.3). Table 2 shows the sparsity and MaxError for r = 4, J = 2, 3 and different values of
thresholding parameter, using the method presented in the previous section.
3300 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

Table 2
Sparsity and MaxError for r = 4.
x Threshold parameter (ε ) Sparsity (Sε ) (%) MaxError

0 0 4.6 × 10−5
J =2 10−4 47.7 3.7 × 10−5
10−3 76.5 1.2 × 10−4
10−2 88.7 6.6 × 10−4
0 0 5.4 × 10−7
J =3 10−4 69.6 2.8 × 10−6
10−3 91.4 7.4 × 10−5
10−2 95.0 3.3 × 10−4

Table 3
Sparsity and MaxError for r = 4.
x Threshold parameter (ε ) Sparsity (Sε ) (%) MaxError

0 0 1.9 × 10−6
J =2 10−4 47.7 3.9 × 10−6
10−3 76.5 6.1 × 10−5
10−2 88.7 6.9 × 10−4
0 0 6.1 × 10−8
J =3 10−4 69.6 4.8 × 10−7
10−3 91.4 1.4 × 10−5
10−2 95.0 4.2 × 10−4

Example 3. Consider the following integro-differential equation

 ∫ 1
u (t ) +
′
|t − s|−1/2 u(s)ds = f (t ),

0
(4.4)
u(0) = 1.


Here the forcing function f is selected such that

1
u(t ) = t 2 + ,
t +1
is the exact solution. Table 3 shows the sparsity and MaxError for r = 4, J = 2, 3 and different values of thresholding
parameter, using the method presented in the previous section.

Example 4. Consider the following integro-differential equation

 ∫ 1
u′ ( t ) + s|t − s|−1/2 u(s)ds = f (t ),

0
(4.5)
u(0) = 0.


Here the forcing function f is selected so that

u(t ) = t 2 (1 − t 3 + t 2 ),

is the exact solution of problem (4.5). Table 4 shows the sparsity and MaxError for r = 3, J = 4, 5 and different values
of thresholding parameter, using the method presented in the previous section. Figs. 4 and 5 show the plots of the matrix
elements for r = 3 and J = 4 with ε = 10−5 , 10−4 , 10−3 , and 10−2 .

Example 5. Consider the following integral equation

 ∫ 1
u′ ( t ) − |t − s|−1/3 u(s)ds = f (t ),

0
(4.6)
u(0) = 1.


Here the forcing function f is selected so that

u(t ) = t 5 + 5t 3 − t + 1,

is the exact solution of problem (4.6).

 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3301

Table 4
Sparsity and MaxError for r = 3.
x Threshold parameter (ε ) Sparsity (Sε ) (%) MaxError

0 0 2.6 × 10−8
J =4 10−4 88.0 4.5 × 10−6
10−3 94.9 5.5 × 10−5
10−2 97.9 3.2 × 10−3
0 0 7.3 × 10−9
J =5 10−4 89.6 8.4 × 10−7
10−3 95.4 2.1 × 10−5
10−2 98.2 1.1 × 10−3

Fig. 4. Plots of sparse matrix after thresholding with ε = 10−5 (left) ε = 10−4 (right) for Example 4.

Fig. 5. Plots of sparse matrix after thresholding with ε = 10−3 (left) ε = 10−2 (right) for Example 4.

Table 5 shows the sparsity and MaxError for r = 3, J = 3, 4 and different values of thresholding parameter, using the
method presented in the previous section.

Example 6. Consider the following integral equation [49]

1
 ∫
u(t ) − |t − s|−1/2 u(s)ds = f (t ),

√
0
f (t ) = t − 2t 1 − t − 2t 2/3 .

3302 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303

Table 5
Sparsity and MaxError for r = 3.
x Threshold parameter (ε ) Sparsity (Sε ) (%) MaxError

0 0 1.6 × 10−5
J =3 10−4 74.4 2.5 × 10−5
10−3 88.9 4.0 × 10−4
10−2 95.7 7.0 × 10−3
0 0 6.1 × 10−7
J =4 10−4 78.3 4.8 × 10−6
10−3 83.7 1.4 × 10−5
10−2 96.2 4.2 × 10−4

Table 6
Absolute errors for Example 6.
t Our method Method presented in [49]
n = 16 n = 32 n = 20 n = 40

0.1 6.8 × 10−28 8.8 × 10−28 3.8 × 10−3 9.5 × 10−4

0.2 8.6 × 10−28 1.1 × 10−27 4.4 × 10−3 1.1 × 10−3
0.3 6.8 × 10−28 7.5 × 10−28 2.9 × 10−3 7.3 × 10−4
0.4 1.0 × 10−29 2.7 × 10−28 1.6 × 10−5 1.8 × 10−4
0.5 3.5 × 10−27 4.3 × 10−27 3.3 × 10−3 7.9 × 10−3
0.6 1.3 × 10−27 6.1 × 10−27 6.0 × 10−3 1.5 × 10−3
0.7 1.8 × 10−27 7.1 × 10−28 7.8 × 10−3 1.8 × 10−3
0.8 1.3 × 10−27 1.2 × 10−28 6.9 × 10−3 1.7 × 10−3
0.9 3.4 × 10−28 1.9 × 10−28 4.5 × 10−3 1.2 × 10−3
1.0 1.3 × 10−27 2.5 × 10−28 1.8 × 10−3 4.9 × 10−4

Table 7
MaxError.
Our method Product integration method [50] Lagrangian
interpolant [50]

n = 16 1.27 × 10−6 1.15 × 10−5 2.12 × 10−5

n = 32 2.71 × 10−8 9.39 × 10−7 1.94 × 10−6

The exact solution of this equation is

u(t ) = t .
Table 6 shows the absolute errors without thresholding for r = 3, J = 2, 3, using the method presented in the previous
section. The comparison of the obtained results with the results presented in [49] shows that our method is more accurate
than the method used in [49].

Example 7. Consider the following integral equation [50]

∫ 1
u(t ) − |t − s|−1/2 u(s)ds = f (t ), (4.7)
0

where f is chosen such that u(t ) = t . Table 7 shows the MaxError without threshold for r = 3, J = 2, 3, using the method
presented in the previous section. Also we see that our method is better than the methods presented in [50].

5. Conclusion

In this paper the Legendre multiwavelets are employed to solve the weakly singular Fredholm integro-differential
equations. Considering the properties of these wavelets and using an approach based on the operational matrices of integrals,
the solution of the weakly singular integral equations is converted to the solution of a sparse linear system of algebraic
equations. Using the spare structure of this linear system, the given problem is solved and memory times are reduced.
Several test examples are used to observe the efficiency and applicability of the new method.

Acknowledgements

The authors are very much thankful to the reviewers for carefully reading the paper, their comments and suggestions
have improved the quality of the paper. Also we would like to thank the principal editor professor Taketomo Mitsui and
 M. Lakestani et al. / Journal of Computational and Applied Mathematics 235 (2011) 3291–3303 3303

professor Masaaki Sugihara, advisory editor of the journal for their useful comments and for managing the review process
for this paper.

References

[1] Y. Ren, B. Zhang, H. Qiao, A simple Taylor-series expansion method for a class of second kind integral equations, J. Comput. Appl. Math. 110 (1999)
15–24.
[2] C. Schneider, Regularity of the solution to a class of weakly singular Fredholm integral equations of the second kind, Integral Equations Operator
Theory 2 (1979) 62–68.
[3] C. Schneider, Product integration for weakly singular integral equations, Math. Comp. 36 (1981) 207–213.
[4] T. Okayama, T. Matsuo, M. Sugihara, Sinc-collocation methods for weakly singular Fredholm integral equation of the second kind, J. Comput. Appl.
Math. 234 (2010) 1211–1227.
[5] I.G. Graham, Singularity expansions for the solutions of second kind Fredholm integral equations with weakly singular convolution kernels, J. Integral
Equations Appl. 4 (1982) 1–30.
[6] G.R. Richter, On weakly singular Fredholm integral equations with displacement kernels, J. Math. Anal. Appl. 55 (1976) 32–42.
[7] G. Vainikko, A. Pedas, The properties of solutions of weakly singular integral equations, J. Aust. Math. Soc. B 22 (1981) 419–430.
[8] A. Avudainayagam, C. Vani, Wavelet-Galerkin method for integro-differential equations, Appl. Numer. Math. 32 (2000) 247–254.
[9] I.G. Graham, Galerkin methods for second kind integral equations with singularities, Math. Comp. 39 (1982) 519–533.
[10] I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure Appl. Math. 41 (1988) 909–996.
[11] I. Daubechies, Ten Lectures on Wavelets, in: CBMS-NSF Lecture Notes, vol. 61, SIAM, 1992.
[12] M. Dehghan, Finite difference procedures for solving a problem arising in modeling and design of certain optoelectronic devices, Math. Comput.
Simulation 71 (2006) 16–30.
[13] M. Dehghan, The one-dimensional heat equation subject to a boundary integral specification, Chaos Solitons Fractals 32 (2007) 661–675.
[14] M. Dehghan, A computational study of the one-dimensional parabolic equation subject to nonclassical boundary specifications, Numer. Methods
Partial Differential Equations 22 (2006) 220–257.
[15] G. Beylkin, On wavelet based algorithms for solving differential equations, in: J.J. Benedetto, et al. (Eds.), Wavelets: Mathematics and Applications,
CRC Press, Boca Raton, FL, 1993, pp. 449–466.
[16] B. Alpert, G. Beylkin, R.R. Coifman, V. Rokhlin, Wavelet-like bases for the fast solution of second-kind integral equations, SIAM J. Sci. Stat. Comput. 14
(1) (1993) 159–184.
[17] A. Averbucha, M. Israeli, L. Vozovoi, Solution of time-dependent diffusion equations with variable coefficients using maltiwavelets, J. Comput. Phys.
150 (1999) 394–424.
[18] J.C. Goswami, A.K. Chan, C.K. Chui, On solving first-kind integral equations using wavelets on bounded interval, IEEE Trans. Antennas Propag. 43 (1995)
614–622.
[19] M. Lakestani, M. Razzaghi, M. Dehghan, Solution of nonlinear Fredholm–Hammerstein integral equations by using semiorthogonal spline wavelets,
Math. Probl. Eng. 1 (2005) 113–121.
[20] M. Lakestani, M. Dehghan, The solution of a second-order nonlinear differential equation with Neumann boundary conditions using semi-orthogonal
B-spline wavelets, Int. J. Comput. Math. 83 (2006) 685–694.
[21] M. Lakestani, M. Razzaghi, M. Dehghan, Numerical solution of the controlled duffing oscillator by semi-orthogonal spline wavelets, Phys. Scr. 74 (2006)
362–366.
[22] R.D. Nevels, J.C. Goswami, H. Tehrani, Semi-orthogonal versus orthogonal wavelet basis sets for solving integral equations, IEEE Trans. Antennas
Propag. 45 (1997) 1332–1339.
[23] G. Beylkin, R. Coifman, V. Rokhlin, Fast wavelet transforms and numerical algorithms I, Comm. Pure Appl. Math. 44 (1991) 141–183.
[24] B. Alpert, G. Beylkin, R. Coifman, R. Rokhlin, Wavelets for the fast solution of second-kind integral equations, Technical Report, Department of Computer
Science, Yale University, New Haven, CT, 1990.
[25] R. Glowinski, W. Lawton, M. Ravachol, E. Tenenbaum, Wavelet solutions of linear and nonlinear elliptic, parabolic and hyperbolic problems in one
space dimension, Comput. Methods Appl. Sci. Engrg. (1990) 55–120.
[26] A. Cohen, I. Daubechies, J. Feauveau, Biorthogonal bases of comactly supported wavelets, Comm. Pure Appl. Math. 45 (1992) 485–560.
[27] A. Herve, Multi-resolution analysis of multiplicity, application to dyadic interpolation, Comput. Harmonic Anal. 1 (1994) 299–315.
[28] B. Alpert, A class of bases in L2 for the sparse representation of integral operators, SIAM J. Math. Anal. 24 (1) (1993) 246–262.
[29] G. Strang, V. Strela, Short wavelets and matrix dilation equations, IEEE Trans. Signal Process. 43 (1995) 108–115.
[30] B. Alpert, G. Beylkin, D. Gines, L. Vozovoi, Adaptive solution of partial differential equations in multiwavelet bases, J. Comput. Phys. 182 (2002) 149–190.
[31] M. Shamsi, M. Razzaghi, Solution of Hallen’s integral equation using multiwavelets, Comput. Phys. Comm. 168 (2005) 187–197.
[32] K.E. Atkinson, A survey of numerical methods for solving nonlinear integral equations, J. Integral Equations Appl. 4 (1992) 15–46.
[33] K.E. Atkinson, The Numerical Solution of Integral Equations of the Second Kind, Cambridge University Press, Cambridge, 1997.
[34] K.E. Atkinson, A. Bogomolny, The discrete Galerkin method for integral equations, Math. Comp. 48 (1987) 595–616.
[35] K.E. Atkinson, F.A. Potra, The discrete Galerkin method for nonlinear integral equations, J. Integral Equations Appl. 1 (1988) 17–54.
[36] C.T. Baker, The Numerical Treatment of Integral Equations, Clarendon Press, Oxford, 1977.
[37] M. Dehghan, A. Saadatmandi, Chebyshev finite difference method for Fredholm integro-differential equation, Int. J. Comput. Math. 86 (2008) 123–130.
[38] M. Muhammad, A. Nurmuhammad, M. Mori, M. Sugihara, Numerical solution of integral equations by means of the sinc collocation method based on
the double exponential transformation, J. Comput. Appl. Math. 177 (2005) 269–286.
[39] D. Mirzaei, M. Dehghan, A meshless based method for solution of integral equations, Appl. Numer. Math. 60 (2010) 245–262.
[40] M. Shakourifar, M. Dehghan, On the numerical solution of nonlinear systems of Volterra integro-differential equations with delay arguments,
Computing 82 (2008) 241–260.
[41] A. Alipanah, M. Dehghan, Numerical solution of the nonlinear Fredholm integral equations by positive definite functions, Appl. Math. Comput. 190
(2007) 1754–1761.
[42] M. Dehghan, On the solution of an initial-boundary value problem that combines Neumann and integral condition for the wave equation, Numer.
Methods Partial Differential Equations 21 (2005) 24–40.
[43] S.A. Yousefi, A. Lotfi, M. Dehghan, He’s variational iteration method for the non-linear mixed Volterra–Fredholm integral equations, Comput. Math.
Appl. 58 (2009) 2172–2176.
[44] M. Dehghan, F. Shakeri, Solution of parabolic integro-differential equations arising in heat conduction in materials with memory via He’s variational
iteration technique, Int. J. Numer. Meth. Biomed. Engng. 26 (2010) 705–715.
[45] M. Dehghan, Solution of a partial integro-differential equation arising from viscoelasticity, Int. J. Comput. Math. 83 (2006) 123–129.
[46] A. Saadatmandi, M. Dehghan, Numerical solution of the higher-order linear Fredholm integro-differential–difference equation with variable
coefficients, Comput. Math. Appl. 59 (2010) 2996–3004.
[47] M. Ueda, S.K. Lodha, Wavelets, An Elementary Introduction and Some Examples of Wavelets, University of California, Santa Cruz, CA, 1995, UCSC-CRL-
94-47.
[48] S. Yousefi, M. Razzaghi, Legendre wavelets method for the nonlinear Volterra–Fredholm integral equations, Math. Comput. Simulation 70 (2005) 1–8.
[49] E.A. Galperin, E.J. Kansa, A. Makroglou, S.A. Nelson, Variable transformations in the numerical solution of second kind Volterra integral equations with
continuous and weakly singular kernels; extensions to Fredholm integral equations, J. Comput. Appl. Math. 115 (2000) 193–211.
[50] C. Allouch, P. Sablonnière, D. Sbibih, M. Tahrichi, Product integration methods based on discrete spline quasi-interpolants and application to weakly
singular integral equations, J. Comput. Appl. Math. 233 (2010) 2855–2866.