TEMA Tend. Mat. Apl. Comput., 8, No. 2 (2007), 229-238.

c Uma Publicac

ao da Sociedade Brasileira de Matem
atica Aplicada e Computacional.

Collocation Solutions of a Weakly Singular Volterra

Integral Equation
T. DIOGO1, P. LIMA2, Centro de Matem
atica e Aplicacoes, Instituto Superior
Tecnico, Av. Rovisco Pais, 1049-001 Lisboa, Portugal.

Abstract. The discrete superconvergence properties of spline collocation solutions

for a certain Volterra integral equation with weakly singular kernel are analyzed. In
particular, the attainable convergence orders at the collocation points are examined
for certain choices of the collocation parameters.

1.

Introduction

The Volterra integral equation

Z t
y(t) =
K(t, s)p(t, s)y(s)ds + g(t),
0

t I = (0, T ],

(1.1)

s1
,
(1.2)
t
where > 0, K(t, s) is a smooth function and g is a given function, can arise,
e.g., in heat conduction problems with mixed boundary conditions ([2], [10]). The
case when K(t, s) = 1 has been considered in several papers. The following lemma
summarizes the analytical results for (1.1) in the case K(t, s) = 1.
p(t, s) :=

Lemma 1.1. (a) [12] Let > 1 in (1.2). If the function g belongs to C m [0, T ]
then the integral equation
Z t
y(t) =
p(t, s)y(s)ds + g(t), t (0, T ],
(1.3)
0

possesses a unique solution y C m [0, T ].

(b) [16] In the above case (a), the unique solution y C m [0, T ] is given by
Z t
1
y(t) = g(t) + t
s2 g(s)ds.
(1.4)
0

(c) [16] However, if 0 < 1 and g C 1 [0, T ] (with g(0) = 0 if = 1), then
(1.3) has a family of solutions in C[0, T ] of which only one has C 1 continuity.
1 [email protected]
2 [email protected]

230

Diogo and Lima

We note that equation (1.3) differs substantially from the well known case
of Abel-type equations whose kernels contain a singularity of the form q(t, s) =
(t s) , 0 < < 1 (see e.g. [6], [15] for comprehensive studies). For these
equations, a smooth forcing function leads to a solution which has typically unbounded derivatives at t = 0. As a consequence of this non-smooth behaviour, if
uniform meshes are used then the convergence order of polynomial spline collocation methods (studied in detail in [6]) is only 1 , independently of the degree
of the polynomials. In order to recover the optimal convergence orders one has to
use suitable graded meshes ([3], [8], [19]). Alternatively, one may keep the uniform
meshes but then use nonpolynomial spline approximating functions reflecting the
singularity; variable transformations followed by standard methods have also been
considered by several authors (e.g. [13], [14], [1], [18]). For an extensive list of
references on these and other approaches see [4].
Equation (1.3) has been the subject of several works for the case when > 1:
certain classes of product integration methods based on Newton-Cotes rules were
studied in [10]; Diogo et al [12] considered a fourth order Hermite-type collocation
method and Lima and Diogo [17] developed an extrapolation algorithm, based on
Eulers method. Recently, it was shown that general collocation methods on uniform
meshes based on piecewise polynomials of degree m1 yield convergence of order m.
In [11] it was also proved that if certain derivatives of the exact solution are zero at
the origin, then a higher order is attained at the mesh points by a special choice of
the collocation points, like the Radau II points. In this work we investigate the use
of the Gauss-Legendre points (that is, the zeros of the shifted Legendre polynomial
Pm (2s 1)) as collocation parameters. In this case, the interesting feature is that
superconvergence does not occur at the mesh points, that is, the superconvergence
results for ODEs when using the Gauss points as collocation parameters do not
carry over to second kind Volterra equations. However, in Section 3 of this work
we give sufficient conditions for superconvergence to be attained at the collocation
points.
The above properties are in sharp contrast with the situation for Abel-type
equations for which no superconvergence properties can be obtained ([7], [6], [4]).

2.

Collocation Methods

Let us consider equation (1.3), that is, (1.1) with K(t, s) = 1, and in future work
the results will be extended to a more general function K. We shall take > 1.
We follow the notations of [6]. Given the following partition of the interval I
{tj = jh,

0 j N; Nh = T} ,

let 0 := [t0 , t1 ], n := (tS

n , tn+1 ], 1 n N 1 and define ZN := {tn : n =
1, . . . , N 1}, Z N := ZN T . Furthermore, let m1 denote the space of polynomials of degree m 1. The exact solution of (1.3) will be approximated in the
piecewise polynomial space
1
Sm1
(ZN ) := {u : u|n =: un m1 ,

0 n N 1},

(2.1)

Collocation Solutions of a Volterra Integral Equation

231

whose elements, in general, will possess jump discontinuities at their knots ZN .

Consider the following finite subset of I
X(N ) =

N[
1

Xn ,

(2.2)

n=0

with
Xn := {tnj := tn + cj h : 0 c1 < . . . < cm 1}.

(2.3)

1
Sm1
(ZN )

The approximate solution u

will be required to satisfy the original
equation (1.3) on X(N ) (set of collocation points). We thus have the following
collocation equation
Z t
u(t) = g(t) +
p(t, s)u(s)ds, t X(N ),
(2.4)
0

which can be rewritten as

Z
un (tnj ) = g(tnj ) +

cj

(n + )1
un (tn + h)d
(n + cj )
0
n1
X Z 1 (i + )1
ui (ti + h)d,
+
(n + cj )
i=0 0

(2.5)

j = 1, . . . , m (n = 0, . . . , N 1).
Let l be the canonical Lagrange polynomials associated with the collocation parameters, defined by
m
Y
ci
l ( ) :=
.
(2.6)
cl ci
i=1
i6=l

If on each subinterval n u is given by its Lagrange formula

un (tn + h) :=

m
X

l ( )Unl ,

l=1

tn + h n ,

(2.7)

where Unl := un (tn + cl h), then (2.5) represents a sequence of N linear systems in
the unknowns (Un1 , . . . , Unm )T , 0 n N 1:
m Z cj
X
(n + )1
(
Unj = g(tnj ) +
l ( )d )Unl
(n + cj )
l=1 0
n1
m Z 1
XX
(i + )1
( )d )Uil .
(2.8)
+
(
l
0 (n + cj )
i=0
l=1

We see that the integrals in (2.8) can be evaluated analytically.

Let e := y u and denote its restriction to the subinterval n by en . We define
kek := sup{|en (t)| : t n ,

n = 0, . . . , N 1}.

We have the following global convergence result [9].

232

Diogo and Lima

(1)

Theorem 2.1. Suppose > 1 in (1.2). Let u Sm1 (ZN ) denote the collocation
approximation to the solution of the integral equation (1.3) and assume that g
C m (I). Then, for every choice of the collocation parameters {cj }, with 0 c1 <
. . . < cm 1, we have
kek = O(hm ),

3.

(as h 0, N h = T ).

(2.9)

Superconvergence

In [11] we investigated the discrete superconvergence properties of the collocation

solutions for equation (1.3). By this we mean the possibility of attaining a higher
order p > m at the mesh points, by a special choice of the collocation points. We
showed that if certain derivatives of y vanish at the origin then a higher order of
the collocation error near the origin is attained. This fact together with a judicious
choice of the collocation points leads to discrete superconvergence at the mesh
points. In particular, we obtained the following result.
(1)

Theorem 3.1. Consider > 1 in (1.3) and let u Sm1 (ZN ) denote the collocation approximation to the solution of the integral equation (1.3). Let the collocation parameters {cj } be the Radau II points for (0,1], that is, the zeros of
Pm1 (2s 1) Pm (2s 1), where Pm is the mth degree Legendre polynomial. If
g C 2m1 [0, T ] and
y (m) (0) = y (m+1) (0) = = y (2m2) (0) = 0,

m 1,

(3.1)

then we have local superconvergence of order 2m 1 at the mesh points, that is,
max |e(tn )| = O(h2m1 ),

tn Z N

(as h 0, N h = T ).

(3.2)

However, if we take as collocation parameters the Gauss-Legendre points (that

is, the zeros of the shifted Legendre polynomial Pm (2s 1)) then superconvergence
does not occur at the mesh points.
On the other hand, we now show that the Gauss points lead to superconvergence
at the collocation points which is stated in the following theorem.
(1)

Theorem 3.2. Let > 1 and u Sm1 (ZN ) denote the collocation approximation
to the solution of the integral equation (1.3), defined by (2.4), where the collocation
parameters {cj } are the Gauss points for (0,1) and let g C m+1 [0, T ]. If the m-th
order derivative of y is such that
y (m) (0) = 0,

m 1,

(3.3)

then we have superconvergence of order m + 1 at the collocation points, that is

max |e(tnj )| = O(hm+1 ),

tnj XN

(as h 0, N h = T ).

(3.4)

Collocation Solutions of a Volterra Integral Equation

233

Proof. It will be convenient to rewrite the collocation equation (2.4) in the form
Z t
p(t, s)u(s)ds (t), t I,
(3.5)
u(t) = g(t) +
0

where is a suitable function which is zero at the collocation points (tnj ). Moreover
will be smooth on each subinterval, with the degree of smoothness given by that
of g. Subtracting (3.5) from (1.3) gives the following second kind Volterra integral
equation for the error function
Z t
e(t) = (t) +
p(t, s)e(s)ds, t I,
(3.6)
0

which is of the type of (1.3) and whose solution is given by (cf. (1.4))
Z t 2
s
e(t) = (t) +
(s)ds, t I.
1
0 t

(3.7)

We now analyze (3.7) at t = tnj , using the fact that (tnj ) = 0. In the case
t = t0j , since y C m+1 [0, T ] and y (m) (0) = 0, it can be proved that (cf. [11])
max |e(t0j )| = O(hm+1 ).

1jm

(3.8)

Consider now the case when t = tnj , n 1. By making appropriate changes of

variables in (3.7) yields
e(tnj ) =

n1
X

+h

cj

(tn + h)2
(tn + h)d .
(tnj )1

i=0

(ti + h)2
(ti + h)d
(tnj )1

If i = 0, we have
Z 1
Z 1  1
Z 1
( h)2
( h)
( h)
h
( h)d =
d
d,
1
tn
h
h
0 (tn )
0
0

(3.9)

(3.10)

where we have used the fact that ( h/tn )

1. By a similar analysis as in [11],
it can be shown that
Z 1
( h)
d = O(hm ).
(3.11)
h
0
If i 1 we proceed as follows:
Z 1
Z 1
Z 1
1
(ti + h)1
(ti + h)
(ti + h)2
(ti + h)d =
(ti + h)d
d.
1
1
(t
)
(t
)
(t
+

h)
ti + h
nj
nj
i
0
0
0
Using an m-point quadrature rule based on the abscissas til , with remainder Eni ,
for each integral, we obtain
Z 1
m
X
(ti + h)
(til )
bl
d =
+ Eni = O(hm+1 ),
(3.12)
t
+

h
t
i
il
0
l=1

234

Diogo and Lima

where we have used the fact that (til ) = 0; on the other hand, since the cj satisfy
the orthogonality condition
Z

m
1Y

0 j=1

(s cj )ds = 0

(3.13)

and the integrand is in the class C m+1 , then Eni = O(hm+1 ) (see e.g. [5]).
Finally, for the last integral in (3.9), we have
Z cj
Z cj
(tn + h)
(tn + h)2
(t
+

h)d

h
d.
(3.14)
h
n
1
(tnj )
tn + h
0
0
From (3.6) we can easily obtain the relation
|(t)|

+1
sup |e(t)|,
t(tn ,tnj ]

t (tn , tnj ],

which, together with (3.14), gives

Z cj
Z cj
(tn + h)2

h
(t
+

h)d

C
h
(tn + h)d = O(hm+1 ).
n
1
(t
)
nj
0
0

(3.15)

(3.16)

In the analysis above we have assumed that tn > h, since the analysis for e(t), with
t near the origin, has been considered in (3.11). Summing up the contributions
(3.11), (3.12), (3.16) in (3.9) yields
|e(tnj )| h O(hm ) + h

n1
X
i=1

|Eni | + O(hm+1 )

O(hm+1 ), n = 1, ..., N.

(3.17)

From (3.17) and (3.8) we obtain the desired result (3.4).

Remark 1. We note that the result of Theorem 3.2 can be generalized as follows.
If the collocation points tij = ti + cj h are such that the collocation parameters
{cj } satisfy the orthogonality condition (3.13), then the result (3.4) remains valid.
A further example of such parameters are the Radau II points and a numerical
illustration is given in the next section (cf. Table 4).

4.

Numerical Results

We have considered the numerical solution of equation (1.3), with various choices
of g(t) and, consequently, of y(t), for t [0, 2].
Example 1. = 1.5 and g such that y(t) = t4 + t4.5 .
Example 2. = 1.5 and g such that y(t) = t5 + t5.4 .
Example 3. = 1.5 and g such that y(t) = t3 .

235

Collocation Solutions of a Volterra Integral Equation

Example 4. = 1.5 and g such that y(t) = t4 .

Tables 1-5 illustrate the performance of the collocation method defined by (2.8)
applied to the above examples.
1
In Table 1 we have considered collocation in Sm1
(ZN ), with m = 3, applied

to Example
1.
We
have
used
the
Gauss
points
{c
=
(5
15)/10, c2 = 1/2,c3 =
1

(5 + 15)/10} as collocation parameters and the results of the first column confirm
the discrete convergence order m = 3, that is, no superconvergence at the mesh
points. However, since y (0) = 0, that is, (3.3) is satisfied, local superconvergence
of order m + 1 = 4 takes place at the collocation points and this is shown in the
third column. Table 2 displays the results obtained with cubic collocation, that is,
1
in Sm1
(ZN ), with m = 4, applied to Example 2. Again we have used the Gauss
points, now given by

525 + 70 30
c1 =
,
p 70

35 + 525 70 30
,
c3 =
70
35

525 70 30
c2 =
,
p 70

35 + 525 + 70 30
c4 =
.
70
35

(4.1)

The results of the first column confirm the discrete convergence order m = 4 (no
superconvergence at the mesh points) and the third column shows local superconvergence of order m + 1 = 5 at the collocation points.
Table 1: Quadratic collocation using Gauss points Example 1:
y(t) = t4 + t4.5 ,
t [0, 2]


 h
max|e(ti )| order max|e(tij )| order 

 0.05

1.6D 4
1.9D 6


 0.025

2.1D

5
3.0
1.1D

7
4.1


 0.0125

2.6D

6
3.0
6.7D

9
4.1


 0.00625

3.2D

7
3.0
4.1D

10
4.0


 0.003125 4.0D 8
3.0
2.5D 11
4.0 

Table 2: Cubic collocation using Gauss points Example 2: y(t) = t5 +t5.4 , t [0, 2]


 h
max|e(ti )| order max|e(tij )| order 


 0.05
2.7D 6
1.3D 8


 0.025
1.7D 7
4.0
3.7D 10
5.1 

 0.0125
1.1D 8
4.0
1.1D 11
5.1 

 0.00625
6.8D 10
4.0
3.3D 13
5.1 

 0.003125 4.3D 11
4.0
1.0D 14
5.0 

236

Diogo and Lima

For completeness, we use Example 2 to illustrate the result of Theorem 3.1 in the
case of quadratic collocation. Since y (0) = y iv (0) = 0, that is, (3.1) is satisfied,
local superconvergence of order 2m
at the mesh points, using
1 = 5 takes place
the Radau II points {c1 = (4 6)/10, c2 = (4 + 6)/10, c3 = 1} as collocation
parameters. This is shown in the first column of Table 3. The third column displays
the errors at the collocation points indicating convergence of order m + 1 = 4, in
accordance with Remark 1.
In Tables 4, 5 we illustrate the importance of conditions (3.1), (3.3) for attaining superconvergence. The first column of Table 4 shows the errors for quadratic
collocation using the Radau II points, applied to Example 3, indicating third order
of convergence at the mesh points; we do not have superconvergence and this is not
surprising since y (0) 6= 0 (cf. (3.1)). The third column of Table 4 shows a similar
situation with collocation based on the Gauss points, since (3.3) is not satisfied;
there we only have order 3 at the collocation points instead of order m + 1 = 4 (cf.
(3.4)).
Table 5 contains the results for Example 4, obtained with cubic collocation based
on the Radau II points (first column) and on the Gauss points (third column). Here,
the function y of Example 4 is such that y iv (0) 6= 0, therefore conditions (3.1) and
(3.3) are not satisfied. Convergence of order 4 is obtained with the Radau II points
(on the mesh) and with the Gauss points (at the collocation points), instead of the
orders 2m 1 = 7 and m + 1 = 5, respectively (cf. (3.2) and (3.4)).
Table 3: Quadratic collocation using Radau II
t [0, 2]

 h
max|e(ti )| order

 0.05
2.8D 8

 0.025
8.1D
10
5.1

 0.0125
2.4D

11
5.1

 0.00625
7.3D

13
5.1

 0.003125 2.2D 14
5.0

points Example 2: y(t) = t5 + t5.4 ,

max|e(tij )|
2.4D 6
1.5D 7
9.2D 9
5.7D 10
3.6D 11


order 


4.0 
4.0 
4.0 
4.0 

Table 4: Quadratic collocation. The importance of conditions (3.1), (3.3).

max |e(ti )| and max |e(tij )| for Example 3














Radau II

Gauss

0.05
0.025
0.0125
0.00625
0.003125

3.3D 6
4.2D 7
5.2D 8
6.5D 9
8.1D 10

3.0
3.0
3.0
3.0

6.0D 6
7.5D 7
9.4D 8
1.2D 8
1.5D 9


p 



3.0 
3.0 
3.0 
3.0 

237

Collocation Solutions of a Volterra Integral Equation

Table 5: Cubic collocation.

The importance of conditions (3.1), (3.3).
max |e(ti )| and max |e(tij )| for Example 4














Radau II

Gauss

0.05
0.025
0.0125
0.00625
0.003125

3.7D 8
2.3D 9
1.4D 10
9.1D 12
5.7D 13

4.0
4.0
4.0
4.0

9.0D 8
5.6D 9
3.5D 10
4.0D 11
1.4D 12


p 



4.0 
4.0 
4.0 
4.0 

References
[1] P. Baratella, A.P. Orsi, A new approach to the numerical solution of weakly
singular volterra integral equations, J. Comput. Appl. Math., 163 (2004), 401418.
28 (1975),
[2] M.A. Bartoshevich, On a heat conduction problem, Inz.- Fiz. Z.,
340-346. (In russian)
[3] H. Brunner, The numerical solution of weakly singular Volterra integral equations by collocation on graded meshes, Math. Comp., 45 (1985), 417-437.
[4] H. Brunner, Collocation methods for Volterra integral and related functional
differential equations, Cambridge University Press, 2004.
[5] H. Brunner, On discrete superconvergence properties of spline collocation
methods for nonlinear Volterra integral equations, J. Comput. Appl. Math.,
10 (1992), 348-357.
[6] H. Brunner, P.J. van der Houwen, The Numerical Solution of Volterra Equations, North Holland, 1986.
[7] H. Brunner, S. Nrsett, Superconvergence of collocation methods for Volterra
and Abel integral equations of the second kind, Numer. Math., 36 (1981),
347-358.
[8] H. Brunner, A. Pedas, G. Vainikko, The piecewise polynomial collocation
method for nonlinear weakly singular Volterra equations, Math. Comp., 68
(1999), 1079-1095.
[9] T. Diogo, A note on collocation and iterated collocation methods for a class of
weakly singular Volterra integral equations, submitted.
[10] T. Diogo, N.B. Franco, P. Lima, High order product integration methods for
a Volterra integral equation with logarithmic singular kernel, Commun. Pure
Appl. Anal., 3 (2004), 217-235.

238

Diogo and Lima

[11] T. Diogo, P. Lima, Superconvergence properties of collocation methods for a

Volterra integral equation with weakly singular kernel, to appear in J. Comput.
Appl. Math..
[12] T. Diogo, S. McKee, T. Tang, A Hermite-type collocation method for the
solution of an integral equation with a certain weakly singular kernel, IMA J.
Numer. Anal., 11 (1991), 595-605.
[13] T. Diogo, S. Mckee, T. Tang, Collocation methods for second-kind Volterra
integral equations with weakly singular kernels, Proc. Roy. Soc. Edinburgh,
124A (1994), 199-210.
[14] E.A. Galperin, E.J. Kansa, A. Makroglou, S.A. Nelson, Variable transformations in the numerical solution of the second kind Volterra integral equations
with continous and weakly singular kernels; extension to Fredholm integral
equations, J. Comput. Appl. Math., 115 (2000), 193-211.
[15] R. Gorenflo, S. Vessella, Abel Integral Equations, a Historico-bibliographical
Survey, IAGA, Firenze, 1984.
[16] W. Han, Existence, uniqueness and smoothness results for second-kind Volterra
equations with weakly singular kernels, J. Integral Equations Appl., 6 (1994),
365-384.
[17] P.M. Lima, T. Diogo, An extrapolation method for a Volterra integral equation
with weakly singular kernel, Appl. Numer. Math. 24 (1997), 131-148.
[18] A. Pedas, G. Vainikko, Smoothing transformation and piecewise polynomial
collocation for weakly singular Volterra integral equations, Computing, 73
(2004), 271-293.
[19] T. Tang, A note on collocation methods for Volterra integro-differential equations with weakly singular kernels, IMA J. Numer. Math., 13 (1993), 93-99.