Sie sind auf Seite 1von 16

Applied Mathematics and Computation 187 (2007) 725–740

www.elsevier.com/locate/amc

Segmented Tau approximation for test neutral


functional differential equations
Luis F. Cordero a, René Escalante b,*,1

a
Departamento de Matemática y Estadı́stica, Facultad de Ciencias Económicas y Sociales,
Universidad de Carabobo, Venezuela
b
Departamento de Cómputo Cientı́fico y Estadı́stica, División de Ciencias Fı́sicas y Matemáticas,
Universidad Simón Bolı́var, Ap. 89000, Caracas 1080-A, Venezuela

Abstract

We use the segmented formulation of the Tau method to approximate the solutions of the neutral delay differential
equation
y 0 ðtÞ ¼ ayðtÞ þ byðt  sÞ þ cy 0 ðt  sÞ þ f ðtÞ; t P 0;
yðtÞ ¼ WðtÞ; t 6 0;
which represents, for different values of a, b, c and s, a family of functional differential equations that some authors have
considered as test equations in different numerical experimentations. The Tau method introduced by Lanczos is an impor-
tant example of how to get approximations of functions defined by a differential equation. In the formulation of a step by
step Tau version is expected that the error is minimized at the matching points of successive steps. Through the study of
recent papers it seems to be demonstrated that the step by step Tau method is a natural and promising strategy for the
numerical solution of functional differential equations. In preliminary experimentation significant improvements have been
obtained when compared with the numerical results obtained elsewhere.
 2006 Elsevier Inc. All rights reserved.

Keywords: Functional differential equations; Neutral delay differential equations; Polynomial approximation; Step by step Tau method
approximation

1. Introduction and preliminaries

In this paper the segmented Lanczos–Tau method is used to find numerical solutions of the linear nonho-
mogeneous functional differential problem of neutral type

*
Corresponding author. Address: CCS: 91646, P.O. Box 025323, Miami, FL 33102 5323, USA.
E-mail addresses: lfcordero@cantv.net (L.F. Cordero), rene@cesma.usb.ve (R. Escalante).
1
This author was partially supported by the Centro de Estadı́stica y Software Matemático (CESMa) and the Decanato de Investigación
y Desarrollo (DID) at USB, project DI-CB-020-05.

0096-3003/$ - see front matter  2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2006.08.085
726 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

y 0 ðtÞ ¼ ayðtÞ þ byðt  sÞ þ cy 0 ðt  sÞ þ f ðtÞ; t P 0;


ð1Þ
yðtÞ ¼ WðtÞ; t 6 0;

where s > 0, W is a continuous function in [s, 0] and f is a continuous function in [0, 1). This equation has as
particular cases: the test neutral (c 5 0 case) and test delay (c = 0 case) functional differential equations. For
example, if in (1) we do

• a = b = 1, c = 2, s = 1, W(t) = t and f(t) = 0, we obtain test equation used in [1] with an unstable dif-
ference operator (see Section 3, Experiment 2),
• a = c = 1, b = 0, s = 1, W(t) = 1 and f(t) = 0, we obtain the test equation used in [2] (see Section 3, Exper-
iment 3),
• a = 1, b = 1, c = 1/4, s = 1, W(t) = t and f(t) = sin(t), we obtain the test equation considered in [1].
Neglecting the forcing term, the resulting problem is unstable (see Section 3, Experiment 5),
• a = p  e3pp/2, b = 1, c = 0, s = 3p/2, W(t) = ept + sin(t) and f(t) = (p  e3pp/2)sin(t), we obtain the test
equation considered in [3]. When p takes large negative values the stiffness of Eq. (1) is increased substan-
tially (see Section 3, Experiment 4),

which all have available analytical solutions [4] (we hope that this fact will help us to show the ‘kindness’ of the
method here proposed).
Although the class of equations considered here is relatively simple, the method can be extended to more
general problems (e.g. nonlinear systems). See [5–7] where some extensions to nonlinear case are considered
using the step by step Tau method. In [3] the particular case c = 0 was studied applying a spectral method type
Tau [8] (our numerical results are consistent with the theoretical and practical results reported there, see Sec-
tion 3, Experiment 4). The case c = 0 is the so-called linear stability DDE test equation, and for a particular
choice of parameters we can produce a stiff delay differential equation (e.g., see Section 3, Experiment 4) or an
unstable problem (for example, if a = 5, b = 1, W(t) = 1 and f(t) = 0 [9,10]).
The theory of delay differential equations has undergone rapid development in the last fifty years. We sug-
gest to see [11,12], where the readers will be able to find a wealth of reference materials on the subject. In more
general models, the derivative y 0 (t) may depend on y and y 0 itself at some past value t  s. These type of equa-
tions are the so-called neutral differential equations. Recently these equations have received considerable atten-
tion by many authors (see some references in [12]). For a review of some basic concepts concerning neutral and
delay differential equations we refer to [13,14].
The Tau method (or the s-method), first introduced by Lanczos [15–17] and later used by Luke [18] to obtain
rational and polynomial approximations mostly for hypergeometric functions, is an important example of
how to get approximations of functions defined by a differential equation.
A convergence analysis and error bounds for Tau method was considered by Lanczos [15,19], Luke [18],
and Ortiz and Pham [20,21]. The recursive form of the Tau method, formalized by Ortiz in [22], was extended
to the case of systems of ordinary differential equations in [23], and also an error analysis was given there. The
basic philosophy of the method was extended to the numerical solution of linear and nonlinear initial value,
boundary value, and mixed problems for ordinary differential equations (see [24,21,25]), consequently applied
to the eigenvalue problems [26–28], to ‘‘stiff’’ problems [24], and to the partial differential equations [29],
among others. The Tau method has also been used as an analytic tool in the discussion of equivalence results
across numerical methods (see [30,31]). It is an important feature of the Tau method that no trial solutions,
approximate quadratures or large matrix inversions are required [24].
In the formulation of a step by step Tau version it is allowed to construct piecewise polynomial approxima-
tions of a given function which can be used to start a refining process (see [32] for details). The Ortiz’ Step by
Step Tau method (or SST method to abbreviate) was later applied efficiently to the solution of linear and non-
linear boundary value problems [24]. More recently, computational strategies for a parallel implementation of
the SST method were proposed in [33].
In recent papers [5–7] the step by step Tau method was applied to find polynomial approximations to the
solution of particular nonlinear delay differential equations. These papers seem to show that the segmented
Tau method is a natural and promising strategy in the numerical solution of functional differential equations.
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 727

Next we briefly sketch the Ortiz’s recursive formulation of the Tau method and its segmented version (the
SST method).
The notation used here is, with slight modifications, that of [22]. Let Pj be the class of polynomials of degree
less than or equal to j. Let us consider the equation defined by the differential operator L
Xm
LyðxÞ  pi ðxÞy ðiÞ ðxÞ ¼ f ðxÞ; ð2Þ
i¼0

where pi(x) and f(x) are polynomials of finite degree (on the contrary, they are polynomial approximations of
given functions), and y(i) represents the ith derivative of y(x). We assume also that the solution y(x) of (2) sat-
isfies conditions
ðfk ; yÞ ¼ sk ; k ¼ 1; . . . ; m; x 2 ½a; b; ð3Þ
where [a, b] is a compact interval, fk are linear functionals acting on y(x), and represents the supplementary
conditions (i.e., the initial, boundary or mixed conditions) to be satisfied by the required solution y(x).
In order to express the Tau method approximations as a weighted arithmetic mean of successive partial
sums of the power series solutions and obtaining approximations of high precision, Lanczos proposed in
1956 [17] the concept of the canonical polynomials Qn(x), n 2 N0  N [ f0g, associated with a linear differen-
tial operator. Afterwards, Ortiz introduced the more workable definition of Qn(x) [22], LQn(x) = xn + Rn,
where Rn(x) is a polynomial generated by {xi}, i 2 S (set of indices for which canonical polynomials remain
undefined), and is called the residual polynomial of Qn(x). Note: RS  spani2S{xi}.
Another important concept is that of generating polynomials [22], which are obtained from applying L to xn,
where n 2 N0 . From them we can find a recursive relation for the canonical polynomials.
Each differential operator L (that belongs to the class characterized by (2)) is uniquely associated with a
sequence {Qn}, n 2 N0  S, of canonical polynomials. For further details about how to generate this sequence
and precise definitions of Ortiz’ algebraic theory of the Tau method we refer to [22].
We now follow the notation used in [24]. Let v = {vi(x)} = Vx be a polynomial basis defined by a lower
^ ¼ fQ
triangular matrix V = (vij), with i; j 2PN, acting on x = (1, x, x2, . . .)t. Let Q ^ n ðxÞg, n 2 N  S such that
^ ^ ^ n
LQn ðxÞ ¼ vn ðxÞ þ Rn ðxÞ, and Qn ðxÞ ¼ j¼0 vn Qj ðxÞ, where j 2 S. We consider the perturbed equation
Ly n ðxÞ ¼ f ðxÞ þ H n ðxÞ; ð4Þ
Pm ðnÞ
where H n ðxÞ ¼ i¼0 si vni ðxÞ 2 P n ; which is called the perturbation term, is expressed
Pr in terms of the basis v.
ðnÞ
The si ’s are parameters that we wish to find. Further we assume that f ðxÞ ¼ i¼0 ai vi ðxÞ; where r 6 n. Hence,
Xm
ðnÞ ^
X
r
y n ðxÞ ¼ si Q ni ðxÞ þ
^ i ðxÞ;
ai Q ð5Þ
i¼0 i¼0

where i 62 S. Eq. (5) is called the Tau approximant of order n of y(x), and satisfies exactly (4) and (3). m Tau
parameters are chosen such that yn satisfies exactly conditions (3). Additional Tau parameters (say s) are cho-
sen such that the residuals of Lyn(x) match the components of f(x) + Hn(x) belonging to RS. If there exist t
exact polynomial solutions of (2), then s + m  t  1 = m (see [24,22] for details).
We observe that the generation of the canonical polynomials depend neither on the conditions (3) nor on
the interval in which the solution is required. These features allow us to introduce the concept of Ortiz’s seg-
mented approximations (i.e. SST method).
Let P be a partition of the interval [a, b] into subintervals Ij = [xj1, xj], j = 1, . . . , E, where
x0 = a < x1 < x2 <    < xE = b. We consider, for j = 1, . . . , E, the Tau approximants y nj ðxÞ, with x 2 Ij, which
define a piecewise Tau approximant yn(x) of order n of the solution y(x) of problem (2) and (3), if each of the
y nj ’s satisfies (2) with a polynomial perturbation term H jn ðxÞ, with x 2 Ij, and, for k = 1, . . . , m, ðfk ; y nr Þ ¼ sk
(r = 1, E), and, for j = 2, . . . , E, y ðiÞ ðiÞ
nj1 ðxj1 Þ ¼ y nj ðxj1 Þ, i = 0, . . . , m  1. The construction of a piecewise Tau
approximation yn(x) of the solution y(x) of problem (2) and (3) depends only on one matrix V and one canon-
ical sequence Q [24, Theorem 2].
There is a range of possibilities in the choice of a basis v. For instance, if v ¼ fxn gn2N , the Tau method car-
ries out the power series expansion method, where a high accuracy is to be expected near the point of expan-
sion. On the other hand, Lanczos in [17,19] suggested the choice of Chebyshev and Legendre polynomials to
728 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

obtain a better distribution of errors over the interval in which the original problem is defined and the approx-
imate solution is required. In particular, Lanczos’ remark concerning the approximations obtained using a
Legendre polynomial perturbation term, allows us to obtain accurate estimations at the end point of that
interval [19], which is the key fact to construct the step by step formulation of the Tau method in which
the error is minimized at the matching point of successive steps [32].
This paper is organized as follows. In Section 2, we find the piecewise polynomial approximation of the
involved neutral problem using the Lanczos–Tau method, and we solve the outlined test neutral differential
problem using the recursive formulation of the Tau method. In Section 3, we present preliminary numerical
experimentation to illustrate the advantages of the proposed strategy. Finally, in Section 4, we present some
concluding remarks.

2. Solving test neutral differential equation using the segmented formulation

We will use the following notation. The symbol I will denote the unit interval [0, 1] and, for each integer k,
k P 1, the interval [ks, (k + 1)s] will be denoted by Ik.
Motivated by some recent papers [6,7] we consider the application of the segmented Tau method approach
to problem (1). In particular, we apply the Tau method to each finite interval Ik of the domain and we solve the
given problem in each interval separately.
Starting with the continuous function W given on [s, 0], which will be a polynomial function (otherwise,
we will consider an accurate polynomial approximation of W), the method generates a polynomial approxi-
mation on the next interval [0, s]. In the same way it proceeds into subsequent intervals, each one with the
same length s. The initial condition is satisfied by the approximate solution in the first interval, and the initial
condition in each subsequent interval is taken to be the endpoint value of the previous solution. Thus, Eq. (1)
is defined over [0, 1) in a piecewise manner. Here, we notice that our segmented strategy will force to shift the
intervals Iks to the unit interval I.
It is clear that if t 2 [s, 1), then t 2 Ik for some k P 1. Hence, t = s(x + k), with x 2 I. Then, we can
define
y k ðxÞ ¼ yðsðx þ kÞÞ ¼ yðtÞ; x 2 I; k P 0;
y 1 ðxÞ ¼ wðsðx  1ÞÞ; x 2 I;
F k ðxÞ ¼ fk ðsðx þ kÞÞ; x 2 I;
where w is equal to W if the latter is a polynomial (on the contrary, it is polynomial approximation in [s, 0]).
Moreover, if f is a polynomial, then for each k P 0, fk is equal to f when t is restricted to lie in Ik (on the con-
trary, it is a polynomial approximation of f in Ik). We will suppose that
Xr
ð1Þ
Xrk
ðkÞ
y 1 ðxÞ ¼ aj xj and F k ðxÞ ¼ cj xj ; k P 0;
j¼0 j¼0

where r and rk are nonnegative integers. We will distinguish two cases: a 5 0 and a = 0.

Case a 5 0:
Eq. (1) leads to the sequence of differential equations
y 0k ðxÞ  asy k ðxÞ ¼ bsy k1 ðxÞ þ cy 0k1 ðxÞ þ sF k ðxÞ; x 2 I; k P 0;
X r
ð1Þ ð6Þ
y 1 ðxÞ ¼ aj xj ; x 2 I;
j¼0

subject to conditions
y k ð0Þ ¼ y k1 ð1Þ; k P 0: ð7Þ
Let n be a fixed integer greater than or equal to max{r, rk} (where r = degree(y1(x)) and rk = degree(Fk(x)),
for each k P 0). Let us note that if there exists a countable set of fk’s, then it is required that the set of all its degrees
is bounded above. From Eq. (4), for each k P 0, we will consider the following perturbed form of Eq. (6):
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 729

Y 0k ðxÞ  asY k ðxÞ ¼ bsY k1 ðxÞ þ cY 0k1 ðxÞ þ sF k ðxÞ þ H ðkÞ
n ðxÞ; x 2 I; k P 0; ð8Þ
where Y1(x) = y1(x) (x 2 I), Yk1(x), for each k P 1, is the polynomial solution from the previous interval,
and the perturbation term H ðkÞ
n ðxÞ is defined in terms of a shifted Legendre polynomial as
X
n
H ðkÞ
n ðxÞ ¼ sk C j xj ;
j¼0

where the Cjs are the coefficients of the shifted Legendre polynomial P n ðxÞ of degree n in I. The choice of
Legendre over Chebyshev polynomials is motivated by their superior endpoint accuracy (see comments at
the end of Section 1, and [19]).
The polynomials we seek are assumed to be of the form
X
n
ðkÞ
Y k ðxÞ ¼ aj xj ; k P 1;
j¼0

with the conditions


Y k ð0Þ ¼ Y k1 ð1Þ; k P 0: ð9Þ
Next, let us define the linear operator L to be
d
Lð:Þ ¼ ðÞ  asðÞ:
dx
So, the perturbed Eq. (8) becomes
X
rk
ðkÞ
X
n
LY k ðxÞ ¼ bsY k1 ðxÞ þ cY 0k1 ðxÞ þ scj xj þ sk C j xj : ð10Þ
j¼0 j¼0

Now, since
n1 
X 
ðk1Þ ðk1Þ
bsY k1 ðxÞ þ cY 0k1 ðxÞ ¼ bsaðk1Þ
n xn þ bsaj þ cðj þ 1Þajþ1 xj
j¼0

and denoting, for rk < n,


8 ðk1Þ ðk1Þ ðkÞ
> þ cðj þ 1Þajþ1 þ scj ; 0 6 j 6 rk ;
< bsaj
>
bðk1Þ
j ¼ bsaðk1Þ ðk1Þ
þ cðj þ 1Þajþ1 ; rk þ 1 6 j 6 n  1;
>
>
j
:
bsaðk1Þ
n ; j ¼ n;
and for rk = n,
( ðk1Þ ðk1Þ ðkÞ
bsaj þ cðj þ 1Þajþ1 þ scj ; 0 6 j 6 n  1;
bðk1Þ
j ¼
bsanðk1Þ þ scðkÞ
n ; j ¼ n;
Eq. (10) becomes
X
n
LY k ðxÞ ¼ ðbðk1Þ
j þ sk C j Þxj : ð11Þ
j¼0

On the other hand, in view of the linearity of L, the canonical polynomials can be obtained from the gener-
ating polynomials (see Section 1)
Lxm ¼ mxm1  asxm ¼ mLQm1 ðxÞ  asLQm ðxÞ ¼ LðmQm1 ðxÞ  asQm ðxÞÞ;
from which
xm ¼ mQm1 ðxÞ  asQm ðxÞ:
730 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

Thus, we obtain a recursive definition of the canonical polynomials


1
Qm ðxÞ ¼ ðxm þ mQm1 ðxÞÞ; m P 0
as
and by induction on m, the above relation reduces to
m! Xm
1
Qm ðxÞ ¼  xi ; m P 0: ð12Þ
as i¼0 ðasÞmi i!
Note that there are no undefined canonical polynomials. Therefore, the exact polynomial solution of the per-
turbed Eq. (11) is set to be
X
n
Y k ðxÞ ¼ ðbðk1Þ
j þ sk C j ÞQj ðxÞ
j¼0

(note that Eq. (11) can be retrieved by applying the operator L to both sides of this expression). Now, if we
substitute (12) in the above expression we have
0   1
1 X n X j bðk1Þ
j þ sk C j j!
Y k ðxÞ ¼  @ x i A:
ji
as j¼0 i¼0 ðasÞ i!
Pn Pj i
 Pn Pn 
j
Moreover, since j¼0 i¼0 Aij x ¼ j¼0 i¼j Aji x (where Aij denotes the ijth element in an array), the exact

polynomial solution of the perturbed equation (11) is reduced to


0   1
ðk1Þ
1 Xn X
n bi þ s k C i i!
Y k ðxÞ ¼  @ Ax j : ð13Þ
ij
as j¼0 i¼j ðasÞ j!

Now, for each k P 0, since the parameter sk was introduced in order to account for the initial condition im-
posed at x = 0, it is calculated using (9). Therefore, by setting x = 0 in (13) we solve for each k, the equation
Yk(0) = Yk1(1), to obtain
!1 !
Xn
C i i! Xn
biðk1Þ i!
sk ¼  i asY k1 ð1Þ þ i :
i¼0 ðasÞ i¼0 ðasÞ

Finally, from (13) we have the desired polynomial approximations to the solution of the functional differential
equation (1)
t 
yðtÞ  Y k ðxÞ ¼ Y k  k ; t 2 I k ; k P 0;
s
or also,
t h t i
yðtÞ  Y ½ t   ; t P 0;
s s s
where [w] denotes the integer part of w.

Case a = 0:
Eq. (1) leads to the sequence of differential equations:
y 0k ðxÞ ¼ bsy k1 ðxÞ þ cy 0k1 ðxÞ þ sF k ðxÞ; x 2 I; k P 0;
Xr
ð1Þ ð14Þ
y 1 ðxÞ ¼ aj xj ; x 2 I;
j¼0

subject to conditions
y k ð0Þ ¼ y k1 ð1Þ; k P 0: ð15Þ
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 731

We can summarize this case in the following theorem.


P k1 ðk1Þ j
Theorem 2.1. For all k P 0, if y k1 ¼ nj¼0 aj x , then problem (14) and (15) has an exact polynomial solution
given by
ðk1Þ
! ðkÞ
X
nk1
bsaj1 ðk1Þ bsanðk1Þ Xrk
scj jþ1
y k ðxÞ ¼ y k1 ð1Þ þ þ caj xj þ k1
xnk1 þ1 þ x ; x 2 I: ð16Þ
j¼1
j nk1 þ 1 j¼0
jþ1

Proof. It follows by induction on k. h


Finally, with the yk’s given by (16), we have the polynomial approximations to the solution of the func-
tional differential Eq. (1)
t 
yðtÞ  y k ðxÞ ¼ y k k ; t 2 I k ; k P 0;
s

which we can also express as


t h t i
yðtÞ  y ½st   ; t P 0:
s s

.Remarks

• In the case a = 0 we do not need to apply the Tau method directly (see Section 3, Experiment 7).
• In case that W and all the fk’s are polynomials, exact solutions are obtained instead of approximate
solutions.

3. Numerical experiments

In this section we compare the performance of the piecewise polynomial approximation obtained in Section
2 to solve problem (1) with the numerical results reported by other authors. The numerical examples illustrate
that feasibility and higher-order accuracy can be achieved. Moreover, the proposed method requires less com-
puting time and can be used widely.
As we observed at the end of Section 1, Lanczos demonstrated that when Chebyshev polynomials are
replaced by Legendre polynomials as a basis a significant increase in accuracy at the endpoint of the approx-
imation range of a Tau approximant is attained [19]. It follows from this result that more accurate information
is provided to each interval by its predecessor by way of supplying it with a more accurate initial value. For
this reason we preferred to use in (8) Legendre polynomials.
The results in this section were obtained using MATLAB 7.0 package and were run on an Intel Pentium 4
Processor (with 16 significant digits). Here, we will take into account some computational parameters
obtained as the processing time and the accuracy of the approximations. It already had been reported in
[34] that the original Tau method is, in many cases, comparable to the accuracy of best uniform approxima-
tions or near optimal polynomial approximations of the same degree. In each case we compare the numerical
results obtained by us with the best values attained from the direct evaluation of the analytical solution (i.e.,
see YEXAC in tables below). In the tables of numerical results we also show the nodes ‘t’ in which we evaluate
the approximate solutions represented by YSST (i.e., the approximate solutions achieved from to apply the
SST method). Likewise, the tables show the corresponding absolute errors (i.e., ERROR) and computing time
CPUTIME (except Tables 1, 3 and 7).
We use the symbol ‘ < ’ in some tables to indicate that a higher precision than 1016 was reached in any
node by a particular numerical solution (although we cannot know exactly how precise it is). In the reported
experiments we also compare our numerical results with the those obtained by other authors elsewhere.
Next we show some of the carried out numerical experiments. For our experimentation we select, for dif-
ferent values of the parameters, some particular equations of the type (1) taken from [35,4]. Let us note that
732 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

Table 1
Results of Experiment 1 showing the errors involved by the application of the SST method and spline strategies reported in [1]
t YEXACT n=3 n=7
ERROR(YSST) ERROR(YSST) ERROR(d16
S ) ERROR(d16
H)

0.2 0.2553506895400424. . . 1.42 · 1004 5.60 · 1011 1.59 · 103 3.10 · 104
0.4 0.5229561744103176. . . 1.39 · 1004 1.39 · 1012 3.89 · 103 7.22 · 104
0.6 0.8055297000976271. . . 1.78 · 1004 6.36 · 1011 5.09 · 103 9.25 · 104
0.8 1.1063852321231171. . . 9.36 · 1005 9.44 · 1012 1.60 · 103 1.33 · 103
1.0 1.4295704571147614. . . 7.01 · 1006 2.22 · 1016 4.05 · 103 4.04 · 103
1.2 1.7025852818153557. . . 1.62 · 1003 8.08 · 1010 2.57 · 103 1.10 · 103
1.4 2.0904677160858514. . . 1.51 · 1003 4.10 · 1012 5.49 · 104 2.47 · 103
1.6 2.6208949716308472. . . 1.96 · 1003 9.20 · 1010 2.90 · 103 2.44 · 103
1.8 3.3281691659926915. . . 1.06 · 1003 1.16 · 1010 2.28 · 103 3.64 · 103
2.0 4.2547941531425408. . . 1.16 · 1004 3.55 · 1015 2.29 · 104 4.64 · 103

Experiments 4 and 5 involve nonhomogeneous problems, and Experiments 1, 2, 3, 5 consider functional dif-
ferential equations of neutral type (i.e., c 5 0).
MATLAB function dde23 is a solver based on the explicit Runge–Kutta (2, 3) pair used by the ODE solver
ode23 [36]. We test the case c = 0 through the Experiments 4 and 6 and we demonstrate the reliability, pre-
cision and speed of our method as compared to dde23. We always ran DDE solver dde23 using the optional
integration argument options to set the scalar relative error tolerance RelTol in 1e8 and the vector of
absolute error tolerances AbsTol in 1e16 (in all components).
Although MATLAB does not currently possess a built-in routine to solve neutral differential equation
problems directly, we made use of dde23 in our programming software to treat numerically the neutral case
(i.e. c 5 0). This was made in two ways. In the first one, we use in the current step the derivative information,
provided by the same dde23, in the immediately previous interval Ik (see ‘‘approxder’’ at top of the neutral
case tables). In the second way, we will suppose that the analytical solution of the given problem is previously
known, so that we could calculate the derivative corresponding to the current interval Ik directly, and use this
information to compute the numerical solution in the following interval (see ‘‘exactder’’ at top of the neutral
case tables). Since in this paper we treat with a simple equation (i.e., Eq. (1)), these computing strategies do not
present bigger difficulty (nevertheless, we were monitoring any theoretical or practical difficulty that could
exist [37, Section 4.5]).
Finally, in Experiment 6, we consider the particular case a = 0, which was treated in the last part of Section
2 (see Theorem 2.1).
Experiment 1. We consider Eq. (1) with a = b = 1, c = 1/4, s = 1, W(t) = t and f(t) = 0 (see [1]). It has in
[0, 2] the analytical solution
(
 1 þ t þ 14 et ; 0 6 t 6 1;
yðtÞ ¼ 1 4 1 t 17 t1 3 t1
2
 t þ 4 e þ 16 e þ 16 te ; 1 6 t 6 2;

with a first-order discontinuity at t = k for k P 0.


Table 1 shows the numerical results obtained. The last two columns of Table 1 show the errors obtained by
Kappel and Kunisch in [1]. They correspond to the best approximations attained for this experiment using
cubic spline (d16 16
S ) and cubic Hermite spline (dH ) approximations. It is interesting to observe that it is sufficient
to use polynomial approximations of degree 3 to attain similar errors to those obtained in [1], which are also
shown in Table 1. Unfortunately, we cannot make an appropriate comparison with the results obtained by
these authors since the hardware and software platforms used by them were very different to the one used by
us.
Table 2 presents numerical results when the MATLAB function dde23 is used. Here we only show
segmented Tau approximations of degree 7. It is clear that the results obtained were favorable to our
approach.
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 733

Table 2
Results for Experiment 1
t YEXACT n=7 Approxder Exactder
ERROR(YSST) ERROR(dde23) ERROR(dde23)
0.2 0.2553506895400424. . . 5.60 · 1011 2.73 · 1010 2.73 · 1010
0.4 0.5229561744103176. . . 1.39 · 1012 1.19 · 1009 1.19 · 1009
0.6 0.8055297000976271. . . 6.36 · 1011 2.93 · 1009 2.93 · 1009
0.8 1.1063852321231171. . . 9.44 · 1012 5.72 · 1009 5.72 · 1009
1.0 1.4295704571147614. . . 2.22 · 1016 9.80 · 1009 9.80 · 1009
1.2 1.7025852818153557. . . 8.08 · 1010 2.23 · 1008 2.22 · 1008
1.4 2.0904677160858514. . . 4.10 · 1012 2.33 · 1008 2.30 · 1008
1.6 2.6208949716308472. . . 9.20 · 1010 2.36 · 1008 2.27 · 1008
1.8 3.3281691659926915. . . 1.16 · 1010 2.25 · 1008 2.06 · 1008
2.0 4.2547941531425408. . . 3.55 · 1015 1.90 · 1008 1.56 · 1008

CPUTIME (s) 0.320461 1.552232 0.680979

Experiment 2. Now we consider Eq. (1) with a = b = 1, c = 2, s = 1, W(t) = t and f(t) = 0 (see [1]). It has in
[0, 2] the analytical solution

2 þ t þ 2et 0 6 t 6 1;
yðtÞ ¼ t t1
4  t þ 2e  2ðt þ 1Þe 1 6 t 6 2;
with a first-order discontinuity at t = k for k P 0.
Table 3 shows the numerical results obtained. The observations made in the previous experiment can also
be applied here. However, in this case we have an unstable difference operator. For that reason in [1] it is
required to use the weighting function g(s) = 1  8s, s 2 [1, 0] (i.e., for an unstable difference operator they
cannot use g  1 instead of the appropriate weighting function). In our case it is also enough to use
polynomial approximations of degree 3 to attain similar errors to those obtained in [1]. Table 3 shows the
errors corresponding to segmented Tau approximations of degrees 7 and 16.
Numerical results obtained using the DDE solver dde23 are shown in Table 4. Here we also report the
step-by-step Tau approximations of degree 7. It is clear the superior accuracy obtained from our piecewise
polynomial approximations. In Fig. 1, we plot the piecewise polynomial approximations obtained.

Experiment 3. We consider Eq. (1) with a = c = 1, b = 0, s = 1, W(t) = 1 and f(t) = 0 (see [2]). It has in [0, 4]
the analytical solution
8 t
> e; 0 6 t 6 1;
>
>
< ðt  1Þet1 þ et ; 1 6 t 6 2;
yðtÞ ¼ 1 2
>
>
t2 t1
ðt  2tÞe þ ðt  1Þe þ e ; t
2 6 t 6 3;
>
: 21 3 2 t3 1 2 t2 t1 t
6
ðt  3t  3t þ 9Þe þ 2
ðt  2tÞe þ ðt  1Þe þ e ; 3 6 t 6 4;

Table 3
Results for Experiment 2 showing the errors obtained by the application of the SST method and spline strategies reported in [1]
t YEXACT n=7 n = 16
ERROR(YSST) ERROR(YSST) ERROR(d16
S ) ERROR(d16
H)
09 16 2
0.25 0.8180508333754827. . . 2.41 · 10 2.22 · 10 4.08 · 10 6.04 · 103
0.50 1.7974425414002564. . . 3.74 · 1009 <1.00 · 1016 1.77 · 102 5.00 · 105
0.75 2.9840000332253496. . . 2.66 · 1009 <1.00 · 1016 1.45 · 102 1.65 · 102
1.00 4.4365636569180911. . . <1.00 · 1016 <1.00 · 1016 5.63 · 102 2.46 · 102
1.25 3.9525715398288463. . . 1.86 · 1008 1.78 · 1015 3.75 · 102 2.23 · 102
1.50 3.2197717871754872. . . 2.91 · 1008 <1.00 · 1016 2.22 · 102 1.61 · 102
1.75 2.1157052606417484. . . 2.08 · 1008 8.88 · 1016 3.35 · 102 1.16 · 102
2.00 0.4684212271070258. . . 3.55 · 1015 8.88 · 1016 1.75 · 101 1.16 · 101
734 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

Table 4
Results for Experiment 2 showing also the errors produced by the MATLAB function
t YEXACT n=7 Approxder Exactder
ERROR(YSST) ERROR(dde23) ERROR(dde23)
0.2 0.6428055163203396. . . 4.48 · 1010 6.79 · 1010 6.79 · 1010
0.4 1.3836493952825406. . . 1.11 · 1011 3.06 · 1009 3.06 · 1009
0.6 2.2442376007810179. . . 5.09 · 1010 7.76 · 1009 7.76 · 1009
0.8 3.2510818569849356. . . 7.55 · 1011 1.56 · 1008 1.56 · 1008
1.0 4.4365636569180911. . . <1.00 · 1016 2.74 · 1008 2.74 · 1008
1.2 4.0660617095683476. . . 3.69 · 1009 7.08 · 1008 6.94 · 1008
1.4 3.5496413850112525. . . 1.66 · 1010 9.67 · 1008 9.00 · 1008
1.6 2.8310470867595825. . . 4.21 · 1009 1.32 · 1007 1.14 · 1007
1.8 1.8362657292680744. . . 4.11 · 1010 1.80 · 1007 1.40 · 1007
2.0 0.4684212271070258. . . 3.55 · 1015 2.43 · 1007 1.68 · 1007

CPUTIME (s) 0.270389 1.512174 0.711022

Fig. 1. Segmented Tau approximations corresponding to Experiment 2.

with a first-order discontinuity at t = k for k P 0. Table 5 shows the numerical results in the set of nodes con-
sidered. There we compare our approximate solution (n = 7) with the one obtained using the MATLAB func-
tion dde23. We can also compare our numerical solution at t = 4 with the one obtained using the STRIDE
(i.e. acronym of STable Runge–Kutta Integrator for Differential Equations) code, which was originally in-
tended as a robust adaptive code for solving initial value problems for ordinary differential equations. In
[38] STRIDE was adapted for solving delay differential equation, and in [35] was reported its application
to the numerical solution of a test set of delay and neutral differential equations. The reported results in
[35], using STRIDE code, show that for t = 4 the calculated solution was 150.3005956, with the corresponding
error equal to 0.2685334834 · 1006 (compare with the value shown at t = 4 corresponding to the column
ERROR(YSST) of Table 5). In this experiment it is clear the very accurate numerical solution of our approach
as compared to other numerical schemes.

Experiment 4. Let us consider Eq. (1) with a = (p  e3pp/2), b = 1, c = 0, s = 3p/2, W(t) = ept + sin(t) and
f(t) = (p  e3pp/2)sin(t), which has the analytical solution
yðtÞ ¼ ept þ sinðtÞ; t P 0:
The c = 0 case was efficiently treated by Ito et al. in [3] using Legendre–Tau approximations, where this equa-
tion was also considered. Here the parameter p is a stiffness parameter.
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 735

Table 5
Results for Experiment 3
t YEXACT n=7 Approxder Exactder
ERROR(YSST) ERROR(dde23) ERROR(dde23)
0.2 1.2214027581601699. . . 2.24 · 1010 2.48 · 1009 2.48 · 1009
0.4 1.4918246976412703. . . 5.55 · 1012 6.06 · 1009 6.06 · 1009
0.6 1.8221188003905089. . . 2.54 · 1010 1.11 · 1008 1.11 · 1008
0.8 2.2255409284924679. . . 3.77 · 1011 1.81 · 1008 1.81 · 1008
1.0 2.7182818284590455. . . <1.00 · 1016 2.75 · 1008 2.75 · 1008
1.2 3.5643974743685813. . . 2.61 · 1009 1.46 · 1007 1.43 · 1007
1.4 4.6519298459011829. . . 6.39 · 1011 1.90 · 1007 1.83 · 1007
1.6 6.0463037046294206. . . 2.98 · 1009 2.48 · 1007 2.33 · 1007
1.8 7.8300802072069207. . . 3.35 · 1010 3.24 · 1007 2.98 · 1007
2.0 10.107337927389697. . . 8.88 · 1015 4.23 · 1007 3.81 · 1007
2.2 13.277862413513219. . . 1.78 · 1008 9.14 · 1007 8.27 · 1007
2.4 17.416532189091964. . . 1.13 · 1009 1.19 · 1006 1.04 · 1006
2.6 22.809842578338476. . . 2.03 · 1008 1.56 · 1006 1.30 · 1006
2.8 29.826618046951911. . . 1.71 · 1009 2.03 · 1006 1.64 · 1006
3.0 38.941071863737534. . . 1.28 · 1013 2.65 · 1006 2.06 · 1006
3.2 51.056949586487917. . . 9.36 · 1008 4.83 · 1006 3.87 · 1006
3.4 66.922433909481157. . . 8.78 · 1009 6.30 · 1006 4.83 · 1006
3.6 87.683517042129267. . . 1.07 · 1007 8.23 · 1006 6.04 · 1006
3.8 114.83068154359199. . . 6.72 · 1009 1.08 · 1005 7.55 · 1006
4.0 150.30059582675779. . . 1.14 · 1012 1.41 · 1005 9.45 · 1006

CPUTIME (s) 0.160230 4.516494 0.711022

In Table 6, we search to compare, for different values of p, the errors corresponding to the application of
the spectral method proposed in [3], the DDE solver dde23 and our approach. In this case we prefer to use
Chebyshev polynomials to approximate the W and f functions. In this experiment, dnSPC denotes the 1-norm
absolute error of the solution components computed by the spectral method using Legendre polynomial of
degree n [3], and dnYSST denotes the absolute error obtained by applying the segmented Tau approximations of
degree n.
We ran our piecewise polynomial approximations and the dde23 function using the same software and
hardware platforms. However, it is important to observe the very long CPU time required by the DDE solver
dde23 as compared with the CPU time produced by the application of the segmented Tau method (see, for
example, the p = 2.0 case in Table 6). We also note that our numerical results are consistent with those
obtained using the scheme proposed in [3]. A slightly higher accuracy in the quality of the solution in the
spectral method case was observed as compared with our approach as well as faster computing times, which
could be explained with the intensive use of Legendre polynomials (see [3]), the FORTRAN language, and the
hardware platform (i.e. IBM 3090) used in [3]. In Fig. 2 we plot, for different values of p, the corresponding
segmented Tau approximations.

Experiment 5. Next, we consider Eq. (1) with a = b = 1, c = 1/4, s = 1, W(t) = t and f(t) = sin(t) (see [1]).
This nonhomogeneous problem has the analytical solution
8 1 3 t 1 1
<  4 þ t þ 4 e  2 cosðtÞ  2 sinðtÞ;
> 0 6 t 6 1;
yðtÞ ¼ 2  t þ 4 e þ 16 ð3t þ 1Þe þ 18 sinðt  1Þ
1 3 t 3 t1
>
:
þ 12 ðcosðt  1Þ  cosðtÞ  sinðtÞÞ; 1 6 t 6 2;
with a first-order discontinuity at t = k for k P 0.
Table 7 shows the numerical results obtained. The last two columns dNH;l denotes the absolute error
obtained in [1] for different values of parameters N and l (it is shown that the scheme proposed converges to
the true solution if, for each N, l is chosen appropriately). They correspond to the best approximations there
attained for this experiment using cubic Hermite spline approximations. For this nonhomogeneous problem
736 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

Table 6
Results for Experiment 4
t YEXACT ERROR(d16
YSST ) ERROR(dde23) ERROR(d16
SPC )
07 07
3p/4 0.7160900722076770. . . 2.72 · 10 6.52 · 10 1.3 · 1010
3p/2 0.9999193004824297. . . 1.38 · 1006 1.27 · 1008 1.1 · 1009
9p/4 0.7071075061337983. . . 2.70 · 1007 6.91 · 1007 2.1 · 1010
3p 0.0000000065124125. . . 1.37 · 1006 1.60 · 1011 1.1 · 1009
15p/4 0.7071067811280456. . . 2.77 · 1007 7.48 · 1008 2.1 · 1010
p = 2.0 CPUTIME (s) 0.230331 520.428339 0.036

ERROR(d10
YSST ) ERROR(dde23) ERROR(d10
SPC )

3p/4 0.8018870060287024. . . 5.33 · 1007 2.08 · 1006 8.3 · 1008


3p/2 0.9910167089788706. . . 1.62 · 1006 1.25 · 1006 7.6 · 1007
9p/4 0.7079582195293519. . . 3.82 · 1007 1.90 · 1006 1.5 · 1008
3p 0.0000806995175707. . . 1.38 · 1006 2.41 · 1008 4.2 · 1007
15p/4 0.7070991324681285. . . 3.14 · 1007 1.05 · 1006 2.0 · 1007
p = 1.0 CPUTIME (s) 0.190274 1.492146 0.017

ERROR(d8YSST ) ERROR(dde23) ERROR(d8SPC )


3p/4 1.4971880641243032. . . 4.54 · 1007 3.71 · 1005 2.6 · 1006
3p/2 0.3757715663514303. . . 1.43 · 1006 6.51 · 1007 7.9 · 1008
9p/4 1.2002979828898344. . . 7.78 · 1006 2.10 · 1005 1.0 · 1005
3p 0.3896611373753472. . . 6.35 · 1007 4.97 · 1006 3.1 · 1007
15p/4 0.3992428098580494. . . 1.78 · 1006 1.75 · 1007 8.4 · 1007
p = 0.1 CPUTIME (s) 0.170245 0.540778 0.009

ERROR(d16
YSST ) ERROR(dde23) ERROR(dNSPC )
13 05
3p/4 11.257830855384309. . . 5.70 · 10 4.94 · 10
3p/2 110.31777848985621. . . 3.79 · 1012 7.55 · 1004 case no reported
9p/4 1175.1902721803249. . . 4.89 · 1011 1.06 · 1002
3p 12391.647807916692. . . 2.46 · 1010 1.39 · 1001
15p/4 130740.14973918526. . . 1.18 · 1008 1.76 · 10+00
p=1 CPUTIME (s) 0.240345 0.530763

Fig. 2. Segmented Tau approximation in [0, 18p/4], for different values of p, corresponding to Experiment 4.

the application of that approach needs of a double limit process [1, Theorem 3.2] (we cannot make an
appropriate comparison since the used hardware and software platforms are very different). In Table 8 we also
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 737

Table 7
Results for Experiment 5
t YEXACT n=7 n = 16
ERROR(YSST) ERROR(YSST) ERROR(d4;16
H )
8;32
ERROR(dH )
0.25 0.3548608720332222. . . 3.74 · 1010 2.87 · 1010 2.28 · 104 8.32 · 104
0.50 0.8080369027778082. . . 1.40 · 1009 4.81 · 1010 3.13 · 103 2.84 · 103
0.75 1.3810861980109286. . . 3.66 · 1008 6.19 · 1010 7.10 · 103 7.10 · 103
1.00 2.0978247260062659. . . 5.27 · 1007 7.99 · 1010 1.76 · 102 1.60 · 102
1.25 2.8945705697206110. . . 6.68 · 1007 3.06 · 1009 2.64 · 102 2.83 · 102
1.50 4.0261139922978888. . . 8.90 · 1007 4.01 · 1009 3.76 · 102 4.17 · 102
1.75 5.5949907352998807. . . 1.39 · 1006 5.25 · 1009 4.32 · 102 5.62 · 102
2.00 7.7382967049462721. . . 5.06 · 1006 6.87 · 1009 4.27 · 102 7.20 · 102

Table 8
Results for Experiment 5 showing also the dde23 errors
t YEXACT n=7 Approxder Exactder
ERROR(YSST) ERROR(dde23) ERROR(dde23)
0.2 0.2766841143019760. . . 9.02 · 1011 3.00 · 1010 3.00 · 1010
0.4 0.6136288550751848. . . 2.40 · 1011 1.50 · 1009 1.50 · 1009
0.6 1.0216000561405245. . . 1.06 · 1010 4.15 · 1009 4.15 · 1009
0.8 1.5121242962460073. . . 1.13 · 1011 8.89 · 1009 8.89 · 1009
1.0 2.0978247260062663. . . 1.02 · 1011 1.66 · 1008 1.66 · 1008
1.2 2.7112161060136102. . . 1.73 · 1009 3.50 · 1008 3.49 · 1008
1.4 3.5274284086794170. . . 5.62 · 1011 3.62 · 1008 3.58 · 1008
1.6 4.5943895899801248. . . 1.96 · 1009 3.59 · 1008 3.46 · 1008
1.8 5.9725848104441406. . . 1.92 · 1010 3.28 · 1008 2.99 · 1008
2.0 7.7382967049462721. . . 5.29 · 1011 2.58 · 1008 2.01 · 1008

CPUTIME (s) 0.260374 1.371973 0.721037

show the use of MATLAB function dde23. From Tables 7 and 8 the numerical advantages are clear when
applying the segmented Tau approximations.

Experiment 6. Let us consider now Eq. (1) with a = c = 0, b = a, s = b, W(t) = c and f(t) = 0 (see [39]). It has
an (k + 1)-st order discontinuity at t = kb. In this case, the equation that is obtained, leads to the sequence of
differential equations
y 0k ðxÞ ¼ aby k1 ðxÞ; x 2 I; k P 0;
y 1 ðxÞ ¼ c; x 2 I;
subject to conditions yk(0) = yk1(1), k P 0. By means of the application of Theorem 2.1 can be established
that these have exact polynomial solutions given by
!
X
kþ1 X
kþ1 i
ðabÞ ðk  ði  1ÞÞ
ij
y k ðxÞ ¼ c xj ; x 2 I; k P 0:
j¼0 i¼j
j!ði  jÞ!
This solution only depends on the parameters a, b and c. From this result and the second remark at the end of
Section 2, it is followed that in this case, for all t P 0, Eq. (1) has exact solution given by
 

t t
yðtÞ ¼ y ½ t  
b b b
0 h i ij 1
X BX
½t=bþ1 ½t=bþ1 ðabÞ
i t
 ði  1Þ  
j
b C t t
¼c @ A 
j¼0 i¼j
j!ði  jÞ! b b
738 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740
0 h i ji  h ii 1
j t t t
X j ðabÞ  ðj  1Þ 
BX
½t=bþ1
b b b C
¼c @ A
j¼0 i¼0
i!ðj  iÞ!

X j
 
ji  
i !
ðabÞj X j
½t=bþ1
t t t
¼c  ðj  1Þ 
j¼0
j! i¼0
i b b b
X ðabÞ
½t=bþ1 j  
 

j
t t t
¼c  ðj  1Þ þ 
j¼0
j! b b b
X
½t=bþ1
aj
¼c ðt  ðj  1ÞbÞj ;
j¼0
j!

Table 9
Results for Experiment 6
t (Case a = b = c = 1) (Case a = e3, b = c = 1)
YEXACT ERROR(dde23) YEXACT ERROR(dde23)
0.5 1.500000000000000. . . <1.00 · 1016 11.04276846159383. . . <1.00 · 1016
1.0 2.000000000000000. . . <1.00 · 1016 21.08553692318766. . . <1.00 · 1016
1.5 2.625000000000000. . . <1.00 · 1016 81.55690457137339. . . 2.84 · 1014
2.0 3.500000000000000. . . <1.00 · 1016 242.8854705927428. . . 2.84 · 1014
2.5 4.645833333333333. . . 8.88 · 1016 673.8854834784500. . . 6.13 · 1012
3.0 6.166666666666667. . . 3.55 · 1015 2218.628185684264. . . 1.36 · 1012
3.5 8.190104166666666. . . 2.50 · 1009 6313.839670810763. . . 6.76 · 1006
4.0 10.87500000000000. . . 4.78 · 1009 19482.33326430240. . . 1.28 · 1005
4.5 14.44036458333333. . . 8.36 · 1009 58846.56276005233. . . 9.68 · 1005
5.0 19.17500000000000. . . 1.43 · 1008 175537.7414232505. . . 2.88 · 1004
5.5 25.46174045138888. . . 2.29 · 1008 535292.2690046248. . . 1.03 · 1003
6.0 33.80972222222222. . . 3.56 · 1008 1603827.056978995. . . 3.85 · 1003
6.5 44.89472811259920. . . 5.42 · 1008 4848078.964499989. . . 1.20 · 1002
7.0 59.61408730158729. . . 8.13 · 1008 14630140.34294186. . . 4.27 · 1002
7.5 79.15940144856770. . . 1.20 · 1007 44059792.33391116. . . 1.39 · 1001
8.0 105.1129216269841. . . 1.76 · 1007 133067201.6152763. . . 4.59 · 1001
8.5 139.5756650948229. . . 2.55 · 1007 401047099.6138002. . . 1.52 · 10+00
9.0 185.3375027557318. . . 3.68 · 1007 1209850918.511728. . . 4.88 · 10+00
9.5 246.1030000719612. . . 5.27 · 1007 3649443550.035652. . . 1.59 · 10+01
10.0 326.7913169642857. . . 7.48 · 1007 11004958281.89848. . . 5.11 · 10+01

CPUTIME (s) 0.711022 2.763974

Fig. 3. Segmented Tau approximation in [0, 10] corresponding to Experiment 6 with a = e3 and b = c = 1.
L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740 739
Pu P  Pu  Pj 
u
where we use the fact that j¼0 i¼j A ji vj ¼ j¼0 i
i¼0 Aij v and the Newton’s binomial formula. This last
expression coincides with the analytical solution reported in [4].
In the cases where c and the nonhomogeneous term are polynomial functions, the application of the
procedure proposed here (Theorem 2.1) will produce an exact polynomial solution (see the second remark at
the end of Section 2). But if we apply the MATLAB function dde23 the solution will not be exact (as it is
shown in Table 9). Therefore we decided not to report the errors corresponding to the application of our
approach in Table 9 because the solutions that we obtained were ‘almost’ exact (small round-off errors are
present) with CPUTIMEs of around 9 ms. Hence it is clear that the exact solution that was obtained here will
be always best that any approximate solution. In Fig. 3, we show the solution’s graph of the segmented Tau
solution in [0, 10] corresponding to the case a = e3 and b = c = 1.

4. Final remarks

This paper together with some recent papers [5–7] seems to demonstrate that the step by step Tau method is
a natural and promising strategy in the numerical solution of functional differential equations.
We observe here that significant improvements have been obtained compared with the numerical results
reported by other authors elsewhere.
We point out the potential application of parallel version of the segmented Tau approximation [33] at func-
tional-differential problems. Let us suppose that we know the exact solution (or a very accurate solution) in
consecutive nodes ks (e.g., using direct evaluations from analytical solution, or using the best values attained
from the normal recursive version). We can take this information to make use of the domain partitioning tech-
nique (it is due to the independence of data and processes). Initially we can compute successive Tau approx-
imations over each subinterval (each one corresponding to one processor), which we can use later to start a
refining process for finding best segmented approximations. Thus, in our problem, the succession of Tau
approximations is defined over consecutive subintervals Ik. In a first fine-grained stage each task is associated
with one processor. The key observation here is that the SST method leads naturally to a first parallel
approach (for parallel strategy details see [33]).

References

[1] F. Kappel, K. Kunisch, Spline approximations for neutral functional differential equations, SIAM J. Numer. Anal. 18 (1981) 1058–
1080.
[2] Z. Jackiewicz, E. Lo, The numerical integration of neutral functional-differential equations by fully implicit one-step methods,
Technical Report No. 129, Arizona State University, Mathematics Department, 1991.
[3] K. Ito, H. Tran, A. Manitius, A fully-discrete spectral method for delay differential equations, SIAM J. Numer. Anal. 28 (1991) 1121–
1140.
[4] C. Paul, A test set of functional differential equations, Technical Report No. 243, University of Manchester, Manchester Centre for
Computational Mathematics, Numerical Analysis Reports, 1994.
[5] L.F. Cordero, R. Escalante, Segmented Tau approximation for a parametric nonlinear neutral differential equation, submitted for
publication.
[6] H. Khajah, Tau method treatment of a delayed negative feedback equation, Comput. Math. Appl. 49 (2005) 1767–1772.
[7] H. Khajah, E. Ortiz, On a differential-delay equation arising in number theory, Appl. Numer. Math. 21 (1996) 431–437.
[8] D. Gottlieb, Orszag, Numerical analysis of spectral methods: theory and applications, in: CBMS-NSF Regional Conference Series in
Applied Mathematics, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1977.
[9] H. Banks, F. Kappel, Spline approximations for functional differential equations, J. Differen. Equat. 34 (1979) 496–522.
[10] K. Ito, R. Teglas, Legendre-tau approximations for functional differential equations, SIAM J. Control Optim. 24 (1986) 737–759.
[11] C. Baker, C. Paul, D. Willé, A bibliography on the numerical solution of delay differential equations, Technical Report NA No. 269,
University of Manchester, Department of Mathematics, Manchester Centre for Computational Mathematics, 1995.
[12] Y. Dib, M. Maroun, Y. Raffoul, Periodicity and stability in neutral nonlinear differential equations with functional delay, Electron. J.
Differen. Equat. 142 (2005) 1–11.
[13] A. Bellen, M. Zennaro, Numerical Methods for Delay Differential Equations, Oxford University Press, New York, 2003.
[14] O. Diekmann, S. van Gils, S.V. Lunel, H.-O. Walther, Delay Equations, Springer, New York, 1995.
[15] C. Lanczos, Trigonometric interpolation of empirical and analytical functions, J. Math. Phys. 17 (1938) 123–199.
[16] C. Lanczos, Introduction, Tables of Chebyshev polynomials, Appl. Math. Ser. U.S. Bur. Stand., 9, Government Printing Office,
Washington, 1952.
740 L.F. Cordero, R. Escalante / Applied Mathematics and Computation 187 (2007) 725–740

[17] C. Lanczos, Applied Analysis, Prentice-Hall Inc., New Jersey, 1956.


[18] Y. Luke, The Special Functions and their Approximations, vol. II, Academic Press, New York, 1969.
[19] C. Lanczos, Legendre versus Chebyshev polynomials, in: J. Miller (Ed.), Topics in Numerical Analysis, Academic Press, New York,
1973.
[20] E. Ortiz, A.P.N. Dinh, An error analysis of the Tau method for a class of singularly perturbed problems for differential equations,
Math. Methods Appl. Sci. 6 (1984) 457–466.
[21] E. Ortiz, A.P.N. Dinh, On the convergence of the Tau method for nonlinear differential equations of Riccati’s type, Nonlinear Anal.
Theory Methods Appl. 9 (1985) 53–60.
[22] E. Ortiz, The Tau method, SIAM J. Numer. Anal. 6 (1969) 480–492.
[23] J. Freilich, E. Ortiz, Numerical solution of systems of ordinary differential equations with the tau method: an error analysis, Math.
Comput. 39 (1982) 467–479.
[24] P. Onumanyi, E. Ortiz, Numerical solution of stiff and singularly perturbed boundary value problems with a segmented-adaptive
formulation of the Tau method, Math. Comput. 43 (167) (1984) 189–203.
[25] E. Ortiz, H. Samara, An operational approach to the Tau method for the numerical solution of non-linear differential equations,
Computing 27 (1981) 15–25.
[26] T. Chaves, E. Ortiz, On the numerical solution of two point boundary value problems for linear differential equations, Z. Angew.
Math. Mech. 48 (1968) 415–418.
[27] K. Liu, E. Ortiz, Tau method approximate solution of high-order differential eigenvalue problems defined in the complex plane, with
an application to Orr–Sommerfeld stability equation, Commun. Appl. Numer. Methods 3 (1987) 187–194.
[28] E. Ortiz, H. Samara, Numerical solution of differential eigenvalue problems with an operational approach to the Tau method,
Computing 31 (1983) 95–103.
[29] S. Namasivayam, E. Ortiz, Best approximation and the numerical solution of partial differential equations with the Tau method, Port.
Math. 40 (1985) 97–119.
[30] M. El-Daou, E. Ortiz, The Tau method as an analytic tool in the discussion of equivalence results across numerical methods,
Computing 60 (4) (1998) 365–376.
[31] E. Ortiz, A.P.N. Dinh, Some remarks on structural relations between the Tau method and the finite element method, Comput. Math.
Appl. 33 (4) (1997) 105–113.
[32] E. Ortiz, Step by step Tau method – Part I: Piecewise polynomial approximations, Comput. Math. Appl. 1 (1975) 381–392.
[33] R. Escalante, Parallel strategies for the step by step Tau method, Appl. Math. Comput. 137 (2003) 277–292.
[34] E. Ortiz, W. Purser, F. Rodriguez-Cañizares, Automation of the Tau method, Technical Report NAS 01-72, Imperial College, 1972
(Presented to the Conference on Numerical Analysis organized by the Royal Irish Academy, Dublin, 1972).
[35] C. Baker, J. Butcher, C. Paul, Experience of STRIDE applied to delay differential equations, Technical Report NA No. 208,
University of Manchester, Department of Mathematics, Manchester Centre for Computational Mathematics, 1992.
[36] L. Shampine, S. Thompson, Solving DDEs in MATLAB, Appl. Numer. Math. 37 (2001) 441–458.
[37] L. Shampine, I. Gladwell, S. Thompson, Solving ODEs with MATLAB, Cambridge University Press, Cambridge, UK, 2003.
[38] J. Butcher, The adaptation of STRIDE to delay differential equations, Appl. Numer. Math. 9 (1992) 415–425.
[39] L. El’sgol’ts, S. Norkin, Introduction to the Theory and Applications of Differential Equations with Deviating Arguments, Academic
Press, New York, 1973.

Das könnte Ihnen auch gefallen