Sei sulla pagina 1di 14

Numer Algor

DOI 10.1007/s11075-016-0193-9

ORIGINAL PAPER

A sufficient condition for the stability of direct


quadrature methods for Volterra integral equations

Eleonora Messina1 · Antonia Vecchio2

Received: 23 February 2016 / Accepted: 9 August 2016


© Springer Science+Business Media New York 2016

Abstract Within the theoretical framework of the numerical stability analysis for the
Volterra integral equations, we consider a new class of test problems and we study the
long-time behavior of the numerical solution obtained by direct quadrature methods
as a function of the stepsize. Furthermore, we analyze how the numerical solution
responds to certain perturbations in the kernel.

Keywords Numerical stability · Volterra equation · Direct quadrature methods

Mathematics Subject Classification (2010) MSC 39A10 · MSC 65R20

1 Introduction

We study the numerical stability of a class of direct quadrature (DQ) methods for the
solution of the linear Volterra integral equations (VIEs) of the type
 t
y(t) = f (t) + k(t, s)y(s)ds, t ≥ 0, (1)
0

 Eleonora Messina
eleonora.messina@unina.it
Antonia Vecchio
antonia.vecchio@cnr.it

1 Dipartimento di Matematica e Applicazioni, Università degli Studi di Napoli “Federico II”- Via
Cintia, I-80126 Napoli, Italy

2 C.N.R. National Research Council of Italy, Institute for Computational Application


“Mauro Picone”, Via P. Castellino, 111, 80131 Napoli Italy
Numer Algor

where k, defined for 0 ≤ s ≤ t < ∞, with values in Rd×d is a Volterra kernel


(k(t, s) = 0 when s > t) and the forcing function f is defined for t ≥ 0 and takes
values in Rd . We assume that f (t) and k(t, s) are at least continuous for t ≥ 0 and
0 ≤ s ≤ t < +∞, respectively. Existence and uniqueness results can be found for
example in [9, 13]. Systems of VIEs are widely used in real life history-dependent
problems such as population dynamics, fluid dynamics, feedback control theory, and
neural fields (see e.g., [3, 4, 6, 9, 11]).
Here, we consider the numerical solution to (1) obtained by DQ−(ρ, σ ) methods
with Gregory convolution weights (see for example [2, 14, 20])

0 −1
n 
n
yn = fn + h wnj knj yj + h ωn−j knj yj , n = n0 , n0 + 1, . . . , (2)
j =0 j =n0

where yn  y(tn ), fn = f (tn ), knj = k(tn , tj ), with tn = nh for n = 0, 1, . . . ,


h > 0 is the stepsize and wnj and ωj are the weights. Here, we assume that the
weights are non-negative and that y0 = f0 , y1 . . . , yn0 −1 , n0 ≥ 1, are given starting
values.
In [16], we identified a class of test problems (1) for the stability analysis of the
DQ methods (2); here, we want to enlarge such a class by relaxing the following
hypothesis contained there, in Section 3,
 t
sup k(t, s)ds ≤ α < 1.
t≥0 0

This condition is classical in the literature about stable VIEs (see [9, Sections 9.8,
9.9], [18, Proposition 4]). However, in [8, Theorem 3] the stability of solutions to
VIEs is proved under more general hypotheses on the integral of the kernel k. More
precisely, it is assumed that there exists a t¯ < +∞ such that t¯ k(t, s)ds < 1,
t

for all t > t¯. This hypothesis on the continuous problem is indeed the starting
point for our investigation on the stability analysis of DQ methods and our aim
here is to study how the freedom before t¯ affects the solution over the entire inter-
val [0, +∞]. The answer to this question is our main result and it is contained in
Section 2.
The organization of the paper is the following. In the next section, after intro-
ducing some basic materials, we give our main result which assures the stability of
the considered methods under some hypotheses on the kernel and for a sufficiently
small stepsize. In Section 3, we refer to an analysis that the de Hoog and Ander-
ssen [12] have conducted on the effect of perturbations in the kernel on the solution
to (1). Here, we use this study as our starting point for investigating the behavior of
the numerical solutions when these perturbations occur. In Section 4, an extension
of the previous results to a Hammerstein type nonlinear equation is shown. Finally,
Sections 5 and 6 contain some numerical tests and some concluding comments,
respectively.
Numer Algor

2 Stability results

Subtracting (2) from (1), the following expression



n
en = T̄n (h) + h ωn−j knj ej , n = n0 , n0 + 1, . . . (3)
j =n0

represents the equation for the global error en = y(tn ) − yn of the DQ method (2)
(see for example [13, p.104]). Here, T̄n (h) is given by
0 −1
n
T̄n (h) = Tn (h) + h wnj knj ηj , (4)
j =0

⎛ ⎞
 tn 0 −1
n 
n
Tn (h) = k(tn , s)y(s)ds − ⎝h wnj knj y(tj ) + h ωn−j knj y(tj )⎠ , n = n0 , n0 + 1, . . . ,
0 j =0 j =n0

being the local truncation error and ηj , j = 0, .., n0 − 1 are the starting errors.
The weights wnj , n = 0, 1, . . . , j = 0, 1, . . . , n0 − 1 appearing in (4) as the
coefficients of the starting errors ηj , j = 0, .., n0 − 1 are called the starting weights
and satisfy (see [2, p.79, Th. 2.6.10])

sup wnj ≤ W < +∞, j = 0, . . . , n0 − 1. (5)


n≥0

Here, we assume that ηj , j = 0, . . . , n0 − 1 go to zero as h → 0. Since n0 is


fixed, this implies that
0 −1
n
lim ηj  = 0. (6)
h→0
j =0

If the solution to (1) is bounded and the kernel k(t, s) is sufficiently regular, by
taking into account (5) and (6), we have

T̄n (h) ≤ Ch, n = n0 , n0 + 1, . . . , (7)

with C independent of h and n. In order to assure the solvability of the discrete


problem (2) from now on we assume that h satisfies

det (I − hω0 knn )  = 0,


where I is the identity matrix of size d. Moreover, we recall the following properties
of the Gregory convolution weights ωn (see for example [2, 20]):
sup ωn =  < +∞, (8)
n

ωi = 1, for i ≥ n0 (9)
Numer Algor

(observe that (9) comes from [2, (2.6.16’), p.72] and from the consistency conditions
of the Adams methods for ODEs). Define ω̃j as

ωj if ωj ≤ 1
ω̃j = , j = 0, . . . , n0 − 1, (10)
1 if ωj > 1
we prove the following result:

Lemma 1 Assume that


(a) supt≥s k(t, s) ≤ K < +∞,
¯ such that n0 ≤ m̄ ≤ m̄
then, for all h > 0, n0 ≥ 1 and m̄, m̄ ¯


m̄  tm̄¯
h ωn−j knj  ≤ k(tn , s)ds + hA, (11)
j =m̄ tm̄−1

¯ and
where n ≥ m̄
0 −1
n  t


∂k(t, s)

A=K (ωj − ω̃j ) +  sup



(12)

∂s
ds,
j =0 t≥0 0

with  given in (8) and ω˜j , j = 0, . . . , n0 − 1, in (10).

Proof Consider
 tj  tj  tj
k(tn , s)ds = hk(tn , tj ) + k(tn , s)ds − k(tn , tj )ds
tj −1 tj −1 tj −1
 tj  tj ∂k
= hk(tn , tj ) − (tn , x)dxds.
tj −1 s ∂x
So

m̄ ¯
m̄  tj
h ωn−j knj  ≤ ωn−j k(tn , s)ds
j =m̄ j =m̄ tj −1



n tj
∂k

+h ωn−j
(tn , s)
ds. (13)

∂s

j =n0 tj −1

¯ ≥ m̄ ≥ n0 ≥ 1. For DQ methods
In the last sum, we have used the fact that n ≥ m̄
(2), we have already observed that (8) and (9) hold. Then,


m̄  tm̄¯ ¯
m̄  tj
h ωn−j knj  ≤ k(tn , s)ds + (ωn−j − ω̃n−j ) k(tn , s)ds
j =m̄ tm̄−1 j =l tj −1


tn
∂k(tn , s)

+ h


∂s
ds,
0
Numer Algor

where l = max{m̄, n − n0 + 1} and ω̃j , j = 0, . . . , n0 − 1 are defined in (10). Since


a) is true, we obtain


m̄  tj 0 −1
n
(ωn−j − ω̃n−j ) k(tn , s)ds ≤ hK (ωj − ω̃j ),
j =l tj −1 j =0

and (11) with (12) comes straightforward.

For the error defined in equation (3), the following stability result holds.

Theorem 1 Assume that:


(a) supt≥s k(t, s) ≤ K < +∞;
t
(b) there exists t¯ ≥ 0
: supt≥t¯ t¯ k(t, s)ds ≤ α < 1;
 t


∂k(t, s)

(c) sup
∂s
ds < ∞.
t>0 0
Let N > 0 be a fixed integer number such that

Kτ ≤ α < 1, (14)


with τ = N. Then, there exists a constant C̃ independent of h such that

en  ≤ C̃h (15)

for all h ∈ (0, h0 ), with h0 = min {τ, (1 − α)/A}, where A is given in (12).

Proof Choose

h < τ with α + Ah ≤ γ < 1, (16)

where A is given in (12) and A < +∞ because hypothesis (c) holds. Equation (16)
defines the new value γ . Since h < τ, there exists an integer m(h) ≥ 1 such that
m(h)h ≤ τ ≤ (m(h) + 1)h. In what follows, we omit the dependence of m on h, for
brevity.
Now we want to investigate the asymptotic behavior of the global error (3)
resulting from the application of the DQ method to (1) with a stepsize h satisfy-
ing (16). To this purpose, first consider an arbitrary integer Q = Q(h) such that
t¯ ≤ (2Nm + 1)h ≤ Qh. From the error (3), for any n such that 2Nm + 1 ≤ n ≤ Q
there results

n
en  ≤ T̄n (h) + h ωn−j knj ej , (17)
j =n0
Numer Algor

by splitting the right-hand side of (17) over the mesh and using the bound (7), there
results

m 
2m
en  ≤ Ch + h ωn−j knj ej  + h ωn−j knj ej 
j =n0 j =m+1


3m 
4m
+h ωn−j knj ej  + h ωn−j knj ej 
j =2m+1 j =3m+1

n
+ ... + h ωn−j knj ej .
j =2Nm+1

Hence,

m 
2m
en  ≤ Ch + max er h ωn−j knj  + max er h ωn−j knj 
n0 ≤r≤m m+1≤r≤2m
j =n0 j =m+1


3m
+ max er h ωn−j knj 
2m+1≤r≤3m
j =2m+1


4m
+ max er h ωn−j knj  + . . .
3m+1≤r≤4m
j =3m+1
n
+ max er h ωn−j knj . (18)
2Nm+1≤r≤Q
j =2Nm+1

The application of Lemma 1 to all but the last one summation involved in (18),
together with hypotheses (a) and (c) and (14) and (16), leads to


(i+1)m  (i+1)mh
h ωn−j knj  ≤ k(tn , s)ds + hA ≤ Kτ + Ah ≤ γ , for i = 1, . . . , 2N − 1,
j =im+1 imh

and

m  mh
h ωn−j knj  ≤ k(tn , s)ds + hA ≤ Kτ + Ah ≤ γ ,
j =n0 (n0 −1)h

where γ ≤ 1 is the constant defined in (16). Furthermore, by Lemma 1 and hypothe-


ses (a), (b), and (c), since 2Nmh ≥ N(m + 1)h ≥ t¯, one also gets that, for any n
such that 2Nm + 1 ≤ n ≤ Q,


n  tn  t
h ωn−j knj  ≤ k(tn , s)ds + hA ≤ sup k(t, s)ds + hA ≤ α + Ah ≤ γ ,
j =2Nm+1 2Nmh t≥t¯ t¯
Numer Algor

The resulting bound for en becomes



1
en  ≤ Ch + γ max er  + max er  + . . . + max er  .
1−γ n0 ≤r≤m m+1≤r≤2m (2N−1)m+1≤r≤2Nm
(19)

Now we have to prove that maxn0 ≤r≤m er  + 2N−1i=1 maxim+1≤r≤(i+1)m er  <
Dh, where D is a positive constant not depending on h. For this reason, consider
each error involved in the previous expression and proceed step by step.
Bounding maxn0 ≤r≤m er  : from the error (3), by using the bound (7) for T̄r (h),
we have

r
er  ≤ Ch + max el h ωr−j krj , n0 ≤ r ≤ m,
n0 ≤l≤m
j =n0

which, in view of Lemma 1, hypotheses (a) and (c) and (14) and (16), leads to
 tr
er  ≤ Ch+ max el  k(tr , s)ds + Ah ≤ Kτ +Ah < γ , n0 ≤ r ≤ m.
n0 ≤l≤m 0

Thus,
Ch
max er  ≤ . (20)
n0 ≤r≤m 1−γ
Bounding maxm+1≤r≤2m er  : since the bound (20) for maxn0 ≤r≤m er  is
already known, it is convenient to split the error (3) over the mesh as follows

m 
r
er = T̄r (h) + h ωr−j krj ej + h ωr−j krj ej , m + 1 ≤ r ≤ 2m,
j =n0 j =m+1

by using the bound (7) for T̄r (h), we have



m
er  ≤ Ch + max el h ωr−j krj 
n0 ≤l≤m
j =n0
r
+ max el h ωr−j krj , m + 1 ≤ r ≤ 2m.
m+1≤l≤2m
j =m+1

Once again, Lemma 1, hypotheses (a) and (c), (14) and (16), together with (20)
give
 tr
Ch
er  ≤ Ch + γ + max el  k(tr , s)ds + Ah
1−γ m+1≤l≤2m mh
Ch Ch
≤ + (Kτ + Ah) max el  ≤ + γ max el 
1−γ m+1≤l≤2m 1−γ m+1≤l≤2m
Numer Algor

and so,
Ch
max er  ≤ .
m+1≤r≤2m (1 − γ )2
By proceeding in the same way, we get,
Ch
max er  ≤ , for each i = 1, ..., 2N − 1. (21)
im+1≤r≤(i+1)m (1 − γ )i+1
using (20) and (21) into (19) we get
 
Ch 
2N−1
1
max en  ≤ 1+γ ,
2Nm+1≤n≤Q 1−γ (1 − γ )i+1
i=0
which becomes
Ch
max en  ≤ . (22)
2Nm+1≤n≤Q (1 − γ )2N+1
The desired result comes from the arbitrariness of Q and by recalling that N does
not depend on h.

Definition 1 The numerical method (2) for solving a class of Volterra integral (1) is
said to be stable for that class of equations if, using a given stepsize h > 0, the global
error en given by (3) satisfies
en  ≤ Ehp , ∀n ≥ 0,
with E constant independent of n and h and p ≥ 1.

According to this definition, we claim that Theorem 1 gives sufficient conditions


for the stability of the class of the DQ methods (2) for all h ∈ (0, h0 ), with h0 defined
in Theorem 1.
Of course, since we are dealing with linear problems, an analysis similar to the
one contained in the proof of Theorem 1 can be used to prove the boundedness of
the numerical solution obtained by applying (2) to a VIE of the type (1) with a
bounded forcing term f (t). Observe that, as already mentioned in Section 1, hypothe-
ses (a) and (b) are also sufficient for the boundedness of the analytical solution of (1)
provided that f (t) is bounded.

Remark 1 As already observed in the Introduction, Theorem 1 represents a gener-


alization of what proved in [16, Theorem 2], that we report here for the sake of
completeness:

Theorem 2 Assume that (a) and (c) hold for (1); furthermore, let
t
b∗∗ ) sup k(t, s)ds ≤ β < 1, then
t≥0 0
Ch
supn≥n0 en  ≤ , (23)
1 − (β + Ah)
for all stepsizes h satisfying β + Ah < 1, where β < 1 is the positive constant in
b∗∗ ), A is given in (12) and Ch is the bound for the local error T̄n (h) defined in (7).
Numer Algor

3 Effect of kernel perturbations

As pointed out in [12], in many practical applications (as for example rheology, risk
analysis, renewal theory) only approximations of the kernel k are known. In these
cases, in the error analysis of numerical methods one has to take into account an
additional term depending on the accuracy of these approximations.
First of all, we consider (1) and assume that a) and b) hold, let k̃ = k + δk (with
δk(t, s) = 0 when s > t) be the perturbed kernel with
a∗) K̃ = supt≥s k̃(t, s) < ∞, ∀t ≥ s ≥ 0,
 t
b∗ ) sup k̃(t, s)ds < 1.
t≥t¯ t¯

Once again, according to [8, Theorem 3], the solution ỹ of the perturbed problem
is bounded. Let δy = ỹ − y, then
 t  t
δy(t) = δk(t, s)y(s)ds + k̃(t, s)δy(s)ds. (24)
0 0
Once again, the same arguments used in [12] combined to the boundedness results
in [8] for (24) lead to
 t
δy ≤ C̄ sup y(t) δk(t, s)ds. (25)
t≥0 0
Assume now that also
 t


∂ k̃(t, s)

c∗ ) sup

ds < +∞,
t≥0 0
∂s

holds. The numerical approximation of (1) with perturbed kernel k̃ is, in fact, the
solution of the following discrete equation
0 −1
n 
n
ỹn = fn + h wnj k̃nj ỹj + h ωn−j k̃nj ỹj , n = n0 , n0 + 1, . . . , (26)
j =0 j =n0

where k̃nj = k̃(tn , tj ). We are interested in how the total error y(tn ) − ỹn evolves,
with respect to the perturbations δknj , over long time intervals.

Theorem 3 Assume that, in (1), a), b), and c) hold for k and that the forcing term
f is bounded in [0, +∞). Furthermore, assume that the perturbed kernel k̃ satisfies
a ∗ ), b∗ ), and c∗ ) then
 tn
sup y(tn ) − ỹn  ≤ C̄ sup δk(tn , s)ds sup y(tn ) + C̃h, (27)
n≥n0 n≥n0 0 n≥n0

for all sufficiently small stepsizes h.

Proof In order to obtain a bound for the total error in DQ approximation plus per-
turbation, we rewrite it as the sum of the following two terms y(tn ) − ỹn = y(tn ) −
Numer Algor

ỹ(tn ) + (ỹ(tn ) − ỹn ). Under hypotheses a), a ∗ ) and b), b∗ ), (25) holds. In order to
obtain a bound for the numerical error ỹ(tn ) − ỹn , we use the result in Theorem 1
with a), b), and c) replaced by a ∗ ), b∗ ), and c∗ ), (27) comes straightforward.

4 Extension to nonlinear equations

For nonlinear equations of the form


 t
y(t) = f (t) + k(t, s)g(y(s))ds, t ≥ 0, (28)
0

where the nonlinearity g : Rd → Rd is a continuously differentiable function, the


DQ−(ρ, σ ) method reads

0 −1
n 
n
yn = fn + h wnj knj g(yj ) + h ωn−j knj g(yj ), n = n0 , n0 + 1, . . . , (29)
j =0 j =n0

and the equation for the global error en = y(tn ) − yn is


n
en = T̄n (h) + h ωnj knj J¯(ζj )ej , (30)
j =n0

with T̄n (h) satisfies (7). Here, we have used the mean value theorem to get g(y(tn ))−
g(yn ) = J¯(ζn )(y(tn ) − yn ), and the notation J¯ for the jacobian J of g means that it is
evaluated at different mean values, all of which are internal points on the line segment
in Rd from y(tj ) to yj . Some important results have been obtained for the numerical
stability of nonlinear equations of the type (28) when the kernel is of convolution
type (see [1, 5, 7]). In case of nonconvolution kernels, the literature is much less
developed and it is mainly concerned with scalar equations (see for example [15]).
Here, by extending the theory developed for the linear case, we obtain the following
theorem.

Theorem 4 Assume that, for (28), a) and c) hold. Furthermore, let J¯ < G, with
0 ≤ G < +∞ and
t
b ) there exists t¯ ≥ 0 such that supt≥t¯ t¯ k(t, s)ds ≤ α < 1/G,

then the numerical DQ-(ρ, σ ) method is stable for all sufficiently small stepsizes h.

Here, the stability interval for h is given by (0, h0 ) with h0 = min


{τ, (1/G − α)/A}, where τ is such that τ K < 1/G and A and α are defined in (12)
and b ), respectively.
Numer Algor

5 Numerical experiments

In this section, we present some numerical examples on a theoretical problem which


is consistent with Theorem 1 but does not satisfies either the hypotheses of its coun-
terpart in [16, Theorem 2] or the ones in [19, Theorem 3.1]. To this aim, let us
consider equation (1) with

(t+s+2)
q (t+1)(s+1) 2, s≤t
k(t, s) = (31)
0 s>t

q ≥ 1 and f (t), such that y(t) = 1 + t. An analysis of problem (31) highlights that
t 2
supt≥0 0 k(t, s)ds = q 1+e e2
> 1, so k does not satisfy hypothesis b∗∗ ) in Remark
1 (hyp. b) in [16, Theorem 2]). Furthermore, k is nonnegative, and although it can
be bounded by an ultimately decreasing the function of s, this function is not inte-
grable; hence, the kernel in (31) does satisfy neither hypothesis h2) in [15] nor i),
iii) of Theorem 3.1 in [19]. On the contrary, it can be easily seen that it fulfills a)–
c) of Theorem 1 of this paper, with t¯ which varies with q. Our experiments have
been performed with q = 1. Figure 1a, b shows the numerical errors, over the long-
time interval [0, 1000], obtained after the discretization of problems (1)–(31) by the
DQ method (2) whose convolution weights are generated by the trapezoidal and the
fifth order Gregory DQ method, respectively. Here, the restrictions on the stepsize h
imposed by Theorem 1 are h < 0.17 for the Trapezoidal DQ method and for h < 0.11
in case of Gregory DQ method of order 5. Our tests have been performed with dif-
ferent values of the stepsize h (h = 0.1, h = 0.05 and h = 0.025), and from the
figures, it is clear the correct behavior of the global errors. For q >> 1, more severe
restrictions on the stepsize h are demanded, due to an increasing stiffness in the prob-
lem. For example, when q = 100, according to Theorem 1, a stable behavior of the
Trapezoidal DQ method is guaranteed only for very small stepsizes (h < 0.002). In
Fig. 2, the unstable behavior of the global error for h = 0.1 is shown.

a −3 b 6 x 10
−6

x 10
2.5 h=0.1
h=0.1 h=0.05
h=0.05 5 h=0.025
2 h=0.025
4
1.5
e e 3
1
2

0.5
1

0
0 200 400 600 800 1000 0
t 0 200 400 600 800 1000
t

Fig. 1 Problem (31) with q = 1, error behavior: a Trapezoidal DQ method and b Gregory DQ of order 5
Numer Algor

Fig. 2 Problem (31) with 38


x 10
q = 100, error behavior for 14
h = .1
12

10

8
e
6

0
0 100 200 300 400 500
t

Our second test consists in the following problem



(s + 1)3 e−s(t+1) , s ≤ t
k(t, s) = (32)
0 s>t
+cos(2t)
2
and f (t) = 4t 6t 2 +1 . Since f (t) is bounded and k satisfies a), b), and c) in
Section 2 with t¯ = 0.5 and α = 0.7, we expect a bounded numerical solution for any
h ∈ (0, h0 ), where h0 ≈ 0.15. In Fig. 3, we report the plot of the numerical solution
to problems (1)–(32) (solid line) with stepsize h = 0.1. In the same plot, the dash-
dotted line represents the numerical solution of a problem obtained by perturbing the
kernel k in (32) by
δk = 10−1 e−s , (33)
while the dotted line refers to
δk = 10−2 e−s . (34)
It is clear that the error in the numerical solution remains small after this perturba-
t
tion and that the magnitude of this error is related to supt≥t¯ t¯ δk(t, s)ds according
to the theoretical results in Section 3.
Finally, we have performed some tests on nonlinear problems with bounded
jacobian; for example, solving (28)–(32) with g(y) = ye−y produces a bounded

Fig. 3 Problem (32), h = .1 : 3 y


n
numerical solution (solid line)
~y
and perturbed numerical n
2.5 ~ yn
solutions (dashed-dotted (33)
and dotted (34) lines,
respectively) 2
y
1.5

0.5
0 10 20 30 40 50
t
Numer Algor

numerical solution for any h ∈ (0, h0 ) with h0 given in 4, and the behavior resembles
the one in Fig. 3, and hence, we do not report the plot.

Remark 2 In our tests we have observed that the restrictions on the stepsize h
imposed by Theorem 1 are quite reasonable. However, if in (12) A >> 1 and
0 < 1 − α << 1, with α defined in b), as it happens in the case of test (31) with
q = 100, the requirement on the stepsize h may be very restrictive. These are very
difficult situations, in which the performances of DQ methods are rather poor and
more robust integration should be considered.

6 Concluding remarks

Inspired by the paper [8], we have proved a generalization of our Theorem 2 in [16]
enlarging the class of problems for which the DQ-(ρ, σ ) methods present a stable
behavior for sufficiently small stepsizes. A comparison to its continuous analogue,
Theorem 3 in [8] reveals that an additional hypothesis on the derivative of the kernel
appears in our result (see (c) in Theorem 1). As it is clear in Lemma 1, this hypothesis
allows us to “link” the assumptions made on the continuous problem to the behavior
of the solution of the Volterra discrete equation (VDE) representing our method.
We want also to mention that a result similar to (15), with assumptions made
directly on the discrete problem, has been presented by D. Reynolds, that we thank
for the helpful discussions, in [10] for VDEs and in his lecture in [17] about a general
parametric VDE.

Acknowledgments We thank the anonymous reviewers whose comments helped improve and clarify
this manuscript. The research that led to the present paper was partially supported by a grant of the group
GNCS of INdAM.

References

1. Baker, C.: A perspective on the numerical treatment of Volterra equations. Numerical analysis 2000,
vol. vi, ordinary differential equations and integral equations. J. Comput. Appl. Math. 125(1–2), 217–
249 (2000). doi:10.1016/S0377-0427(00)00470-2
2. Brunner, H., van der Houwen, P.: The numerical solution of Volterra equations. North-Holland (1986)
3. Burton, T.: Volterra integral and differential equations. Elsevier B. V., Amsterdam (2005)
4. Cardone, A., Messina, E., Vecchio, A.: An adaptive method for Volterra-Fredholm integral equations
on the half line. J. Comput. Appl. Math. 228(2), 538–547 (2009). doi:10.1016/j.cam.2008.03.036
5. Eggermont, P., Lubich, C.: Uniform error estimates of operational quadrature methods for nonlinear
convolution equations on the half-line. Math. Comp 56(193), 149–176 (1991). doi:10.2307/2008535
6. Feng, Z., Xu, D., H.Z.: Epidemiological models with non-exponentially distributed dis-
ease stages and applications to disease control. Bull. Math. Biol. 69(5), 1511–1536 (2007).
doi:10.1007/s11538-006-9174-9
7. Ford, N., Baker, C.: Qualitative behaviour and stability of solutions of discretised nonlinear
Volterra integral equations of convolution type. J. Comput. Appl. Math. 66(1-2), 213–225 (1996).
doi:10.1016/0377-0427(95)00180-8
8. Gripenberg, G.: On the resolvents of nonconvolution Volterra kernels. Funkcial. Ekvac. 23(1), 83–95
(1980)
Numer Algor

9. Gripenberg, G., Londen, S.O., Staffans, O.: Volterra integral and functional equations. Cambridge
University Press, Cambridge (1990)
10. Győri, I., Reynolds, D.: On admissibility of the resolvent of discrete Volterra equations. J. Difference
Equ. Appl. 16(12), 1393–1412 (2010)
11. Hartung, N.: Efficient resolution of metastatic tumor growth models by reformulation into integral
equations. Discret. Contin. Dyn. Syst. - Ser. B 20(2), 445–467 (2015). doi:10.3934/dcdsb.2015.20.445
12. de Hoog, F., Anderssen, R.: Kernel perturbations for a class of second-kind convolution
Volterra equations with non-negative kernels. Appl. Math. Lett. 25(9), 1222–1225 (2012).
doi:10.1016/j.aml.2012.02.058
13. Linz, P.: Analytical and numerical methods for Volterra equations. SIAM, Philadelphia (1985)
14. Lubich, C.: On the stability of linear multistep methods for Volterra convolution equations. IMA J.
Numer. Anal. 3(4), 439–465 (1983). doi:10.1093/imanum/3.4.439
15. Messina, E., Vecchio, A.: Nonlinear stability of direct quadrature methods for Volterra integral
equations. Math. Comput. Simul. 110, 155–164 (2015). doi:10.1016/j.matcom.2013.04.007
16. Messina, E., Vecchio, A.: Stability and boundedness of numerical approximations to Volterra integral
equations. Submitted (2016)
17. Reynolds, D.W.: Parameter-dependent Volterra summation equations. In: The International Confer-
ence on Difference Equations and Applications. Private communication. Bialystok, Poland (2015)
18. Strauss, A.: On a perturbed Volterra integral equation. J. Math. Anal. Appl. 30, 564–575 (1970)
19. Vecchio, A.: Boundedness of the global error of some linear and nonlinear methods for Volterra
integral equations. J. Integr. Equ. Appl. 12(4), 449–465 (2000). doi:10.1216/jiea/1020282238
20. Wolkenfelt, P.H.M.: Linear multistep methods and the construction of quadrature formulae for
Volterra integral and integro-differential equations. Afdeling Numerieke Wiskunde [Department of
Numerical Mathematics] 76 (1979)

Potrebbero piacerti anche